Bug Summary

File:src/gnu/usr.bin/clang/libclangCodeGen/../../../llvm/clang/lib/CodeGen/CGBuiltin.cpp
Warning:line 15676, column 22
Called C++ object pointer is null

Annotated Source Code

Press '?' to see keyboard shortcuts

clang -cc1 -cc1 -triple amd64-unknown-openbsd7.0 -analyze -disable-free -disable-llvm-verifier -discard-value-names -main-file-name CGBuiltin.cpp -analyzer-store=region -analyzer-opt-analyze-nested-blocks -analyzer-checker=core -analyzer-checker=apiModeling -analyzer-checker=unix -analyzer-checker=deadcode -analyzer-checker=cplusplus -analyzer-checker=security.insecureAPI.UncheckedReturn -analyzer-checker=security.insecureAPI.getpw -analyzer-checker=security.insecureAPI.gets -analyzer-checker=security.insecureAPI.mktemp -analyzer-checker=security.insecureAPI.mkstemp -analyzer-checker=security.insecureAPI.vfork -analyzer-checker=nullability.NullPassedToNonnull -analyzer-checker=nullability.NullReturnedFromNonnull -analyzer-output plist -w -setup-static-analyzer -mrelocation-model static -mframe-pointer=all -relaxed-aliasing -fno-rounding-math -mconstructor-aliases -munwind-tables -target-cpu x86-64 -tune-cpu generic -debugger-tuning=gdb -fcoverage-compilation-dir=/usr/src/gnu/usr.bin/clang/libclangCodeGen/obj -resource-dir /usr/local/lib/clang/13.0.0 -I /usr/src/gnu/usr.bin/clang/libclangCodeGen/../../../llvm/clang/include -I /usr/src/gnu/usr.bin/clang/libclangCodeGen/../../../llvm/llvm/include -I /usr/src/gnu/usr.bin/clang/libclangCodeGen/../include -I /usr/src/gnu/usr.bin/clang/libclangCodeGen/obj -I /usr/src/gnu/usr.bin/clang/libclangCodeGen/obj/../include -D NDEBUG -D __STDC_LIMIT_MACROS -D __STDC_CONSTANT_MACROS -D __STDC_FORMAT_MACROS -D LLVM_PREFIX="/usr" -internal-isystem /usr/include/c++/v1 -internal-isystem /usr/local/lib/clang/13.0.0/include -internal-externc-isystem /usr/include -O2 -Wno-unused-parameter -Wwrite-strings -Wno-missing-field-initializers -Wno-long-long -Wno-comment -std=c++14 -fdeprecated-macro -fdebug-compilation-dir=/usr/src/gnu/usr.bin/clang/libclangCodeGen/obj -ferror-limit 19 -fvisibility-inlines-hidden -fwrapv -stack-protector 2 -fno-rtti -fgnuc-version=4.2.1 -vectorize-loops -vectorize-slp -fno-builtin-malloc -fno-builtin-calloc -fno-builtin-realloc -fno-builtin-valloc -fno-builtin-free -fno-builtin-strdup -fno-builtin-strndup -analyzer-output=html -faddrsig -D__GCC_HAVE_DWARF2_CFI_ASM=1 -o /home/ben/Projects/vmm/scan-build/2022-01-12-194120-40624-1 -x c++ /usr/src/gnu/usr.bin/clang/libclangCodeGen/../../../llvm/clang/lib/CodeGen/CGBuiltin.cpp
1//===---- CGBuiltin.cpp - Emit LLVM Code for builtins ---------------------===//
2//
3// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4// See https://llvm.org/LICENSE.txt for license information.
5// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6//
7//===----------------------------------------------------------------------===//
8//
9// This contains code to emit Builtin calls as LLVM code.
10//
11//===----------------------------------------------------------------------===//
12
13#include "CGCUDARuntime.h"
14#include "CGCXXABI.h"
15#include "CGObjCRuntime.h"
16#include "CGOpenCLRuntime.h"
17#include "CGRecordLayout.h"
18#include "CodeGenFunction.h"
19#include "CodeGenModule.h"
20#include "ConstantEmitter.h"
21#include "PatternInit.h"
22#include "TargetInfo.h"
23#include "clang/AST/ASTContext.h"
24#include "clang/AST/Attr.h"
25#include "clang/AST/Decl.h"
26#include "clang/AST/OSLog.h"
27#include "clang/Basic/TargetBuiltins.h"
28#include "clang/Basic/TargetInfo.h"
29#include "clang/CodeGen/CGFunctionInfo.h"
30#include "llvm/ADT/APFloat.h"
31#include "llvm/ADT/APInt.h"
32#include "llvm/ADT/SmallPtrSet.h"
33#include "llvm/ADT/StringExtras.h"
34#include "llvm/Analysis/ValueTracking.h"
35#include "llvm/IR/DataLayout.h"
36#include "llvm/IR/InlineAsm.h"
37#include "llvm/IR/Intrinsics.h"
38#include "llvm/IR/IntrinsicsAArch64.h"
39#include "llvm/IR/IntrinsicsAMDGPU.h"
40#include "llvm/IR/IntrinsicsARM.h"
41#include "llvm/IR/IntrinsicsBPF.h"
42#include "llvm/IR/IntrinsicsHexagon.h"
43#include "llvm/IR/IntrinsicsNVPTX.h"
44#include "llvm/IR/IntrinsicsPowerPC.h"
45#include "llvm/IR/IntrinsicsR600.h"
46#include "llvm/IR/IntrinsicsRISCV.h"
47#include "llvm/IR/IntrinsicsS390.h"
48#include "llvm/IR/IntrinsicsWebAssembly.h"
49#include "llvm/IR/IntrinsicsX86.h"
50#include "llvm/IR/MDBuilder.h"
51#include "llvm/IR/MatrixBuilder.h"
52#include "llvm/Support/ConvertUTF.h"
53#include "llvm/Support/ScopedPrinter.h"
54#include "llvm/Support/X86TargetParser.h"
55#include <sstream>
56
57using namespace clang;
58using namespace CodeGen;
59using namespace llvm;
60
61static
62int64_t clamp(int64_t Value, int64_t Low, int64_t High) {
63 return std::min(High, std::max(Low, Value));
64}
65
66static void initializeAlloca(CodeGenFunction &CGF, AllocaInst *AI, Value *Size,
67 Align AlignmentInBytes) {
68 ConstantInt *Byte;
69 switch (CGF.getLangOpts().getTrivialAutoVarInit()) {
70 case LangOptions::TrivialAutoVarInitKind::Uninitialized:
71 // Nothing to initialize.
72 return;
73 case LangOptions::TrivialAutoVarInitKind::Zero:
74 Byte = CGF.Builder.getInt8(0x00);
75 break;
76 case LangOptions::TrivialAutoVarInitKind::Pattern: {
77 llvm::Type *Int8 = llvm::IntegerType::getInt8Ty(CGF.CGM.getLLVMContext());
78 Byte = llvm::dyn_cast<llvm::ConstantInt>(
79 initializationPatternFor(CGF.CGM, Int8));
80 break;
81 }
82 }
83 if (CGF.CGM.stopAutoInit())
84 return;
85 auto *I = CGF.Builder.CreateMemSet(AI, Byte, Size, AlignmentInBytes);
86 I->addAnnotationMetadata("auto-init");
87}
88
89/// getBuiltinLibFunction - Given a builtin id for a function like
90/// "__builtin_fabsf", return a Function* for "fabsf".
91llvm::Constant *CodeGenModule::getBuiltinLibFunction(const FunctionDecl *FD,
92 unsigned BuiltinID) {
93 assert(Context.BuiltinInfo.isLibFunction(BuiltinID))((void)0);
94
95 // Get the name, skip over the __builtin_ prefix (if necessary).
96 StringRef Name;
97 GlobalDecl D(FD);
98
99 // If the builtin has been declared explicitly with an assembler label,
100 // use the mangled name. This differs from the plain label on platforms
101 // that prefix labels.
102 if (FD->hasAttr<AsmLabelAttr>())
103 Name = getMangledName(D);
104 else
105 Name = Context.BuiltinInfo.getName(BuiltinID) + 10;
106
107 llvm::FunctionType *Ty =
108 cast<llvm::FunctionType>(getTypes().ConvertType(FD->getType()));
109
110 return GetOrCreateLLVMFunction(Name, Ty, D, /*ForVTable=*/false);
111}
112
113/// Emit the conversions required to turn the given value into an
114/// integer of the given size.
115static Value *EmitToInt(CodeGenFunction &CGF, llvm::Value *V,
116 QualType T, llvm::IntegerType *IntType) {
117 V = CGF.EmitToMemory(V, T);
118
119 if (V->getType()->isPointerTy())
120 return CGF.Builder.CreatePtrToInt(V, IntType);
121
122 assert(V->getType() == IntType)((void)0);
123 return V;
124}
125
126static Value *EmitFromInt(CodeGenFunction &CGF, llvm::Value *V,
127 QualType T, llvm::Type *ResultType) {
128 V = CGF.EmitFromMemory(V, T);
129
130 if (ResultType->isPointerTy())
131 return CGF.Builder.CreateIntToPtr(V, ResultType);
132
133 assert(V->getType() == ResultType)((void)0);
134 return V;
135}
136
137/// Utility to insert an atomic instruction based on Intrinsic::ID
138/// and the expression node.
139static Value *MakeBinaryAtomicValue(
140 CodeGenFunction &CGF, llvm::AtomicRMWInst::BinOp Kind, const CallExpr *E,
141 AtomicOrdering Ordering = AtomicOrdering::SequentiallyConsistent) {
142 QualType T = E->getType();
143 assert(E->getArg(0)->getType()->isPointerType())((void)0);
144 assert(CGF.getContext().hasSameUnqualifiedType(T,((void)0)
145 E->getArg(0)->getType()->getPointeeType()))((void)0);
146 assert(CGF.getContext().hasSameUnqualifiedType(T, E->getArg(1)->getType()))((void)0);
147
148 llvm::Value *DestPtr = CGF.EmitScalarExpr(E->getArg(0));
149 unsigned AddrSpace = DestPtr->getType()->getPointerAddressSpace();
150
151 llvm::IntegerType *IntType =
152 llvm::IntegerType::get(CGF.getLLVMContext(),
153 CGF.getContext().getTypeSize(T));
154 llvm::Type *IntPtrType = IntType->getPointerTo(AddrSpace);
155
156 llvm::Value *Args[2];
157 Args[0] = CGF.Builder.CreateBitCast(DestPtr, IntPtrType);
158 Args[1] = CGF.EmitScalarExpr(E->getArg(1));
159 llvm::Type *ValueType = Args[1]->getType();
160 Args[1] = EmitToInt(CGF, Args[1], T, IntType);
161
162 llvm::Value *Result = CGF.Builder.CreateAtomicRMW(
163 Kind, Args[0], Args[1], Ordering);
164 return EmitFromInt(CGF, Result, T, ValueType);
165}
166
167static Value *EmitNontemporalStore(CodeGenFunction &CGF, const CallExpr *E) {
168 Value *Val = CGF.EmitScalarExpr(E->getArg(0));
169 Value *Address = CGF.EmitScalarExpr(E->getArg(1));
170
171 // Convert the type of the pointer to a pointer to the stored type.
172 Val = CGF.EmitToMemory(Val, E->getArg(0)->getType());
173 Value *BC = CGF.Builder.CreateBitCast(
174 Address, llvm::PointerType::getUnqual(Val->getType()), "cast");
175 LValue LV = CGF.MakeNaturalAlignAddrLValue(BC, E->getArg(0)->getType());
176 LV.setNontemporal(true);
177 CGF.EmitStoreOfScalar(Val, LV, false);
178 return nullptr;
179}
180
181static Value *EmitNontemporalLoad(CodeGenFunction &CGF, const CallExpr *E) {
182 Value *Address = CGF.EmitScalarExpr(E->getArg(0));
183
184 LValue LV = CGF.MakeNaturalAlignAddrLValue(Address, E->getType());
185 LV.setNontemporal(true);
186 return CGF.EmitLoadOfScalar(LV, E->getExprLoc());
187}
188
189static RValue EmitBinaryAtomic(CodeGenFunction &CGF,
190 llvm::AtomicRMWInst::BinOp Kind,
191 const CallExpr *E) {
192 return RValue::get(MakeBinaryAtomicValue(CGF, Kind, E));
193}
194
195/// Utility to insert an atomic instruction based Intrinsic::ID and
196/// the expression node, where the return value is the result of the
197/// operation.
198static RValue EmitBinaryAtomicPost(CodeGenFunction &CGF,
199 llvm::AtomicRMWInst::BinOp Kind,
200 const CallExpr *E,
201 Instruction::BinaryOps Op,
202 bool Invert = false) {
203 QualType T = E->getType();
204 assert(E->getArg(0)->getType()->isPointerType())((void)0);
205 assert(CGF.getContext().hasSameUnqualifiedType(T,((void)0)
206 E->getArg(0)->getType()->getPointeeType()))((void)0);
207 assert(CGF.getContext().hasSameUnqualifiedType(T, E->getArg(1)->getType()))((void)0);
208
209 llvm::Value *DestPtr = CGF.EmitScalarExpr(E->getArg(0));
210 unsigned AddrSpace = DestPtr->getType()->getPointerAddressSpace();
211
212 llvm::IntegerType *IntType =
213 llvm::IntegerType::get(CGF.getLLVMContext(),
214 CGF.getContext().getTypeSize(T));
215 llvm::Type *IntPtrType = IntType->getPointerTo(AddrSpace);
216
217 llvm::Value *Args[2];
218 Args[1] = CGF.EmitScalarExpr(E->getArg(1));
219 llvm::Type *ValueType = Args[1]->getType();
220 Args[1] = EmitToInt(CGF, Args[1], T, IntType);
221 Args[0] = CGF.Builder.CreateBitCast(DestPtr, IntPtrType);
222
223 llvm::Value *Result = CGF.Builder.CreateAtomicRMW(
224 Kind, Args[0], Args[1], llvm::AtomicOrdering::SequentiallyConsistent);
225 Result = CGF.Builder.CreateBinOp(Op, Result, Args[1]);
226 if (Invert)
227 Result =
228 CGF.Builder.CreateBinOp(llvm::Instruction::Xor, Result,
229 llvm::ConstantInt::getAllOnesValue(IntType));
230 Result = EmitFromInt(CGF, Result, T, ValueType);
231 return RValue::get(Result);
232}
233
234/// Utility to insert an atomic cmpxchg instruction.
235///
236/// @param CGF The current codegen function.
237/// @param E Builtin call expression to convert to cmpxchg.
238/// arg0 - address to operate on
239/// arg1 - value to compare with
240/// arg2 - new value
241/// @param ReturnBool Specifies whether to return success flag of
242/// cmpxchg result or the old value.
243///
244/// @returns result of cmpxchg, according to ReturnBool
245///
246/// Note: In order to lower Microsoft's _InterlockedCompareExchange* intrinsics
247/// invoke the function EmitAtomicCmpXchgForMSIntrin.
248static Value *MakeAtomicCmpXchgValue(CodeGenFunction &CGF, const CallExpr *E,
249 bool ReturnBool) {
250 QualType T = ReturnBool ? E->getArg(1)->getType() : E->getType();
251 llvm::Value *DestPtr = CGF.EmitScalarExpr(E->getArg(0));
252 unsigned AddrSpace = DestPtr->getType()->getPointerAddressSpace();
253
254 llvm::IntegerType *IntType = llvm::IntegerType::get(
255 CGF.getLLVMContext(), CGF.getContext().getTypeSize(T));
256 llvm::Type *IntPtrType = IntType->getPointerTo(AddrSpace);
257
258 Value *Args[3];
259 Args[0] = CGF.Builder.CreateBitCast(DestPtr, IntPtrType);
260 Args[1] = CGF.EmitScalarExpr(E->getArg(1));
261 llvm::Type *ValueType = Args[1]->getType();
262 Args[1] = EmitToInt(CGF, Args[1], T, IntType);
263 Args[2] = EmitToInt(CGF, CGF.EmitScalarExpr(E->getArg(2)), T, IntType);
264
265 Value *Pair = CGF.Builder.CreateAtomicCmpXchg(
266 Args[0], Args[1], Args[2], llvm::AtomicOrdering::SequentiallyConsistent,
267 llvm::AtomicOrdering::SequentiallyConsistent);
268 if (ReturnBool)
269 // Extract boolean success flag and zext it to int.
270 return CGF.Builder.CreateZExt(CGF.Builder.CreateExtractValue(Pair, 1),
271 CGF.ConvertType(E->getType()));
272 else
273 // Extract old value and emit it using the same type as compare value.
274 return EmitFromInt(CGF, CGF.Builder.CreateExtractValue(Pair, 0), T,
275 ValueType);
276}
277
278/// This function should be invoked to emit atomic cmpxchg for Microsoft's
279/// _InterlockedCompareExchange* intrinsics which have the following signature:
280/// T _InterlockedCompareExchange(T volatile *Destination,
281/// T Exchange,
282/// T Comparand);
283///
284/// Whereas the llvm 'cmpxchg' instruction has the following syntax:
285/// cmpxchg *Destination, Comparand, Exchange.
286/// So we need to swap Comparand and Exchange when invoking
287/// CreateAtomicCmpXchg. That is the reason we could not use the above utility
288/// function MakeAtomicCmpXchgValue since it expects the arguments to be
289/// already swapped.
290
291static
292Value *EmitAtomicCmpXchgForMSIntrin(CodeGenFunction &CGF, const CallExpr *E,
293 AtomicOrdering SuccessOrdering = AtomicOrdering::SequentiallyConsistent) {
294 assert(E->getArg(0)->getType()->isPointerType())((void)0);
295 assert(CGF.getContext().hasSameUnqualifiedType(((void)0)
296 E->getType(), E->getArg(0)->getType()->getPointeeType()))((void)0);
297 assert(CGF.getContext().hasSameUnqualifiedType(E->getType(),((void)0)
298 E->getArg(1)->getType()))((void)0);
299 assert(CGF.getContext().hasSameUnqualifiedType(E->getType(),((void)0)
300 E->getArg(2)->getType()))((void)0);
301
302 auto *Destination = CGF.EmitScalarExpr(E->getArg(0));
303 auto *Comparand = CGF.EmitScalarExpr(E->getArg(2));
304 auto *Exchange = CGF.EmitScalarExpr(E->getArg(1));
305
306 // For Release ordering, the failure ordering should be Monotonic.
307 auto FailureOrdering = SuccessOrdering == AtomicOrdering::Release ?
308 AtomicOrdering::Monotonic :
309 SuccessOrdering;
310
311 // The atomic instruction is marked volatile for consistency with MSVC. This
312 // blocks the few atomics optimizations that LLVM has. If we want to optimize
313 // _Interlocked* operations in the future, we will have to remove the volatile
314 // marker.
315 auto *Result = CGF.Builder.CreateAtomicCmpXchg(
316 Destination, Comparand, Exchange,
317 SuccessOrdering, FailureOrdering);
318 Result->setVolatile(true);
319 return CGF.Builder.CreateExtractValue(Result, 0);
320}
321
322// 64-bit Microsoft platforms support 128 bit cmpxchg operations. They are
323// prototyped like this:
324//
325// unsigned char _InterlockedCompareExchange128...(
326// __int64 volatile * _Destination,
327// __int64 _ExchangeHigh,
328// __int64 _ExchangeLow,
329// __int64 * _ComparandResult);
330static Value *EmitAtomicCmpXchg128ForMSIntrin(CodeGenFunction &CGF,
331 const CallExpr *E,
332 AtomicOrdering SuccessOrdering) {
333 assert(E->getNumArgs() == 4)((void)0);
334 llvm::Value *Destination = CGF.EmitScalarExpr(E->getArg(0));
335 llvm::Value *ExchangeHigh = CGF.EmitScalarExpr(E->getArg(1));
336 llvm::Value *ExchangeLow = CGF.EmitScalarExpr(E->getArg(2));
337 llvm::Value *ComparandPtr = CGF.EmitScalarExpr(E->getArg(3));
338
339 assert(Destination->getType()->isPointerTy())((void)0);
340 assert(!ExchangeHigh->getType()->isPointerTy())((void)0);
341 assert(!ExchangeLow->getType()->isPointerTy())((void)0);
342 assert(ComparandPtr->getType()->isPointerTy())((void)0);
343
344 // For Release ordering, the failure ordering should be Monotonic.
345 auto FailureOrdering = SuccessOrdering == AtomicOrdering::Release
346 ? AtomicOrdering::Monotonic
347 : SuccessOrdering;
348
349 // Convert to i128 pointers and values.
350 llvm::Type *Int128Ty = llvm::IntegerType::get(CGF.getLLVMContext(), 128);
351 llvm::Type *Int128PtrTy = Int128Ty->getPointerTo();
352 Destination = CGF.Builder.CreateBitCast(Destination, Int128PtrTy);
353 Address ComparandResult(CGF.Builder.CreateBitCast(ComparandPtr, Int128PtrTy),
354 CGF.getContext().toCharUnitsFromBits(128));
355
356 // (((i128)hi) << 64) | ((i128)lo)
357 ExchangeHigh = CGF.Builder.CreateZExt(ExchangeHigh, Int128Ty);
358 ExchangeLow = CGF.Builder.CreateZExt(ExchangeLow, Int128Ty);
359 ExchangeHigh =
360 CGF.Builder.CreateShl(ExchangeHigh, llvm::ConstantInt::get(Int128Ty, 64));
361 llvm::Value *Exchange = CGF.Builder.CreateOr(ExchangeHigh, ExchangeLow);
362
363 // Load the comparand for the instruction.
364 llvm::Value *Comparand = CGF.Builder.CreateLoad(ComparandResult);
365
366 auto *CXI = CGF.Builder.CreateAtomicCmpXchg(Destination, Comparand, Exchange,
367 SuccessOrdering, FailureOrdering);
368
369 // The atomic instruction is marked volatile for consistency with MSVC. This
370 // blocks the few atomics optimizations that LLVM has. If we want to optimize
371 // _Interlocked* operations in the future, we will have to remove the volatile
372 // marker.
373 CXI->setVolatile(true);
374
375 // Store the result as an outparameter.
376 CGF.Builder.CreateStore(CGF.Builder.CreateExtractValue(CXI, 0),
377 ComparandResult);
378
379 // Get the success boolean and zero extend it to i8.
380 Value *Success = CGF.Builder.CreateExtractValue(CXI, 1);
381 return CGF.Builder.CreateZExt(Success, CGF.Int8Ty);
382}
383
384static Value *EmitAtomicIncrementValue(CodeGenFunction &CGF, const CallExpr *E,
385 AtomicOrdering Ordering = AtomicOrdering::SequentiallyConsistent) {
386 assert(E->getArg(0)->getType()->isPointerType())((void)0);
387
388 auto *IntTy = CGF.ConvertType(E->getType());
389 auto *Result = CGF.Builder.CreateAtomicRMW(
390 AtomicRMWInst::Add,
391 CGF.EmitScalarExpr(E->getArg(0)),
392 ConstantInt::get(IntTy, 1),
393 Ordering);
394 return CGF.Builder.CreateAdd(Result, ConstantInt::get(IntTy, 1));
395}
396
397static Value *EmitAtomicDecrementValue(CodeGenFunction &CGF, const CallExpr *E,
398 AtomicOrdering Ordering = AtomicOrdering::SequentiallyConsistent) {
399 assert(E->getArg(0)->getType()->isPointerType())((void)0);
400
401 auto *IntTy = CGF.ConvertType(E->getType());
402 auto *Result = CGF.Builder.CreateAtomicRMW(
403 AtomicRMWInst::Sub,
404 CGF.EmitScalarExpr(E->getArg(0)),
405 ConstantInt::get(IntTy, 1),
406 Ordering);
407 return CGF.Builder.CreateSub(Result, ConstantInt::get(IntTy, 1));
408}
409
410// Build a plain volatile load.
411static Value *EmitISOVolatileLoad(CodeGenFunction &CGF, const CallExpr *E) {
412 Value *Ptr = CGF.EmitScalarExpr(E->getArg(0));
413 QualType ElTy = E->getArg(0)->getType()->getPointeeType();
414 CharUnits LoadSize = CGF.getContext().getTypeSizeInChars(ElTy);
415 llvm::Type *ITy =
416 llvm::IntegerType::get(CGF.getLLVMContext(), LoadSize.getQuantity() * 8);
417 Ptr = CGF.Builder.CreateBitCast(Ptr, ITy->getPointerTo());
418 llvm::LoadInst *Load = CGF.Builder.CreateAlignedLoad(ITy, Ptr, LoadSize);
419 Load->setVolatile(true);
420 return Load;
421}
422
423// Build a plain volatile store.
424static Value *EmitISOVolatileStore(CodeGenFunction &CGF, const CallExpr *E) {
425 Value *Ptr = CGF.EmitScalarExpr(E->getArg(0));
426 Value *Value = CGF.EmitScalarExpr(E->getArg(1));
427 QualType ElTy = E->getArg(0)->getType()->getPointeeType();
428 CharUnits StoreSize = CGF.getContext().getTypeSizeInChars(ElTy);
429 llvm::Type *ITy =
430 llvm::IntegerType::get(CGF.getLLVMContext(), StoreSize.getQuantity() * 8);
431 Ptr = CGF.Builder.CreateBitCast(Ptr, ITy->getPointerTo());
432 llvm::StoreInst *Store =
433 CGF.Builder.CreateAlignedStore(Value, Ptr, StoreSize);
434 Store->setVolatile(true);
435 return Store;
436}
437
438// Emit a simple mangled intrinsic that has 1 argument and a return type
439// matching the argument type. Depending on mode, this may be a constrained
440// floating-point intrinsic.
441static Value *emitUnaryMaybeConstrainedFPBuiltin(CodeGenFunction &CGF,
442 const CallExpr *E, unsigned IntrinsicID,
443 unsigned ConstrainedIntrinsicID) {
444 llvm::Value *Src0 = CGF.EmitScalarExpr(E->getArg(0));
445
446 if (CGF.Builder.getIsFPConstrained()) {
447 CodeGenFunction::CGFPOptionsRAII FPOptsRAII(CGF, E);
448 Function *F = CGF.CGM.getIntrinsic(ConstrainedIntrinsicID, Src0->getType());
449 return CGF.Builder.CreateConstrainedFPCall(F, { Src0 });
450 } else {
451 Function *F = CGF.CGM.getIntrinsic(IntrinsicID, Src0->getType());
452 return CGF.Builder.CreateCall(F, Src0);
453 }
454}
455
456// Emit an intrinsic that has 2 operands of the same type as its result.
457// Depending on mode, this may be a constrained floating-point intrinsic.
458static Value *emitBinaryMaybeConstrainedFPBuiltin(CodeGenFunction &CGF,
459 const CallExpr *E, unsigned IntrinsicID,
460 unsigned ConstrainedIntrinsicID) {
461 llvm::Value *Src0 = CGF.EmitScalarExpr(E->getArg(0));
462 llvm::Value *Src1 = CGF.EmitScalarExpr(E->getArg(1));
463
464 if (CGF.Builder.getIsFPConstrained()) {
465 CodeGenFunction::CGFPOptionsRAII FPOptsRAII(CGF, E);
466 Function *F = CGF.CGM.getIntrinsic(ConstrainedIntrinsicID, Src0->getType());
467 return CGF.Builder.CreateConstrainedFPCall(F, { Src0, Src1 });
468 } else {
469 Function *F = CGF.CGM.getIntrinsic(IntrinsicID, Src0->getType());
470 return CGF.Builder.CreateCall(F, { Src0, Src1 });
471 }
472}
473
474// Emit an intrinsic that has 3 operands of the same type as its result.
475// Depending on mode, this may be a constrained floating-point intrinsic.
476static Value *emitTernaryMaybeConstrainedFPBuiltin(CodeGenFunction &CGF,
477 const CallExpr *E, unsigned IntrinsicID,
478 unsigned ConstrainedIntrinsicID) {
479 llvm::Value *Src0 = CGF.EmitScalarExpr(E->getArg(0));
480 llvm::Value *Src1 = CGF.EmitScalarExpr(E->getArg(1));
481 llvm::Value *Src2 = CGF.EmitScalarExpr(E->getArg(2));
482
483 if (CGF.Builder.getIsFPConstrained()) {
484 CodeGenFunction::CGFPOptionsRAII FPOptsRAII(CGF, E);
485 Function *F = CGF.CGM.getIntrinsic(ConstrainedIntrinsicID, Src0->getType());
486 return CGF.Builder.CreateConstrainedFPCall(F, { Src0, Src1, Src2 });
487 } else {
488 Function *F = CGF.CGM.getIntrinsic(IntrinsicID, Src0->getType());
489 return CGF.Builder.CreateCall(F, { Src0, Src1, Src2 });
490 }
491}
492
493// Emit an intrinsic where all operands are of the same type as the result.
494// Depending on mode, this may be a constrained floating-point intrinsic.
495static Value *emitCallMaybeConstrainedFPBuiltin(CodeGenFunction &CGF,
496 unsigned IntrinsicID,
497 unsigned ConstrainedIntrinsicID,
498 llvm::Type *Ty,
499 ArrayRef<Value *> Args) {
500 Function *F;
501 if (CGF.Builder.getIsFPConstrained())
502 F = CGF.CGM.getIntrinsic(ConstrainedIntrinsicID, Ty);
503 else
504 F = CGF.CGM.getIntrinsic(IntrinsicID, Ty);
505
506 if (CGF.Builder.getIsFPConstrained())
507 return CGF.Builder.CreateConstrainedFPCall(F, Args);
508 else
509 return CGF.Builder.CreateCall(F, Args);
510}
511
512// Emit a simple mangled intrinsic that has 1 argument and a return type
513// matching the argument type.
514static Value *emitUnaryBuiltin(CodeGenFunction &CGF,
515 const CallExpr *E,
516 unsigned IntrinsicID) {
517 llvm::Value *Src0 = CGF.EmitScalarExpr(E->getArg(0));
518
519 Function *F = CGF.CGM.getIntrinsic(IntrinsicID, Src0->getType());
520 return CGF.Builder.CreateCall(F, Src0);
521}
522
523// Emit an intrinsic that has 2 operands of the same type as its result.
524static Value *emitBinaryBuiltin(CodeGenFunction &CGF,
525 const CallExpr *E,
526 unsigned IntrinsicID) {
527 llvm::Value *Src0 = CGF.EmitScalarExpr(E->getArg(0));
528 llvm::Value *Src1 = CGF.EmitScalarExpr(E->getArg(1));
529
530 Function *F = CGF.CGM.getIntrinsic(IntrinsicID, Src0->getType());
531 return CGF.Builder.CreateCall(F, { Src0, Src1 });
532}
533
534// Emit an intrinsic that has 3 operands of the same type as its result.
535static Value *emitTernaryBuiltin(CodeGenFunction &CGF,
536 const CallExpr *E,
537 unsigned IntrinsicID) {
538 llvm::Value *Src0 = CGF.EmitScalarExpr(E->getArg(0));
539 llvm::Value *Src1 = CGF.EmitScalarExpr(E->getArg(1));
540 llvm::Value *Src2 = CGF.EmitScalarExpr(E->getArg(2));
541
542 Function *F = CGF.CGM.getIntrinsic(IntrinsicID, Src0->getType());
543 return CGF.Builder.CreateCall(F, { Src0, Src1, Src2 });
544}
545
546// Emit an intrinsic that has 1 float or double operand, and 1 integer.
547static Value *emitFPIntBuiltin(CodeGenFunction &CGF,
548 const CallExpr *E,
549 unsigned IntrinsicID) {
550 llvm::Value *Src0 = CGF.EmitScalarExpr(E->getArg(0));
551 llvm::Value *Src1 = CGF.EmitScalarExpr(E->getArg(1));
552
553 Function *F = CGF.CGM.getIntrinsic(IntrinsicID, Src0->getType());
554 return CGF.Builder.CreateCall(F, {Src0, Src1});
555}
556
557// Emit an intrinsic that has overloaded integer result and fp operand.
558static Value *
559emitMaybeConstrainedFPToIntRoundBuiltin(CodeGenFunction &CGF, const CallExpr *E,
560 unsigned IntrinsicID,
561 unsigned ConstrainedIntrinsicID) {
562 llvm::Type *ResultType = CGF.ConvertType(E->getType());
563 llvm::Value *Src0 = CGF.EmitScalarExpr(E->getArg(0));
564
565 if (CGF.Builder.getIsFPConstrained()) {
566 CodeGenFunction::CGFPOptionsRAII FPOptsRAII(CGF, E);
567 Function *F = CGF.CGM.getIntrinsic(ConstrainedIntrinsicID,
568 {ResultType, Src0->getType()});
569 return CGF.Builder.CreateConstrainedFPCall(F, {Src0});
570 } else {
571 Function *F =
572 CGF.CGM.getIntrinsic(IntrinsicID, {ResultType, Src0->getType()});
573 return CGF.Builder.CreateCall(F, Src0);
574 }
575}
576
577/// EmitFAbs - Emit a call to @llvm.fabs().
578static Value *EmitFAbs(CodeGenFunction &CGF, Value *V) {
579 Function *F = CGF.CGM.getIntrinsic(Intrinsic::fabs, V->getType());
580 llvm::CallInst *Call = CGF.Builder.CreateCall(F, V);
581 Call->setDoesNotAccessMemory();
582 return Call;
583}
584
585/// Emit the computation of the sign bit for a floating point value. Returns
586/// the i1 sign bit value.
587static Value *EmitSignBit(CodeGenFunction &CGF, Value *V) {
588 LLVMContext &C = CGF.CGM.getLLVMContext();
589
590 llvm::Type *Ty = V->getType();
591 int Width = Ty->getPrimitiveSizeInBits();
592 llvm::Type *IntTy = llvm::IntegerType::get(C, Width);
593 V = CGF.Builder.CreateBitCast(V, IntTy);
594 if (Ty->isPPC_FP128Ty()) {
595 // We want the sign bit of the higher-order double. The bitcast we just
596 // did works as if the double-double was stored to memory and then
597 // read as an i128. The "store" will put the higher-order double in the
598 // lower address in both little- and big-Endian modes, but the "load"
599 // will treat those bits as a different part of the i128: the low bits in
600 // little-Endian, the high bits in big-Endian. Therefore, on big-Endian
601 // we need to shift the high bits down to the low before truncating.
602 Width >>= 1;
603 if (CGF.getTarget().isBigEndian()) {
604 Value *ShiftCst = llvm::ConstantInt::get(IntTy, Width);
605 V = CGF.Builder.CreateLShr(V, ShiftCst);
606 }
607 // We are truncating value in order to extract the higher-order
608 // double, which we will be using to extract the sign from.
609 IntTy = llvm::IntegerType::get(C, Width);
610 V = CGF.Builder.CreateTrunc(V, IntTy);
611 }
612 Value *Zero = llvm::Constant::getNullValue(IntTy);
613 return CGF.Builder.CreateICmpSLT(V, Zero);
614}
615
616static RValue emitLibraryCall(CodeGenFunction &CGF, const FunctionDecl *FD,
617 const CallExpr *E, llvm::Constant *calleeValue) {
618 CGCallee callee = CGCallee::forDirect(calleeValue, GlobalDecl(FD));
619 return CGF.EmitCall(E->getCallee()->getType(), callee, E, ReturnValueSlot());
620}
621
622/// Emit a call to llvm.{sadd,uadd,ssub,usub,smul,umul}.with.overflow.*
623/// depending on IntrinsicID.
624///
625/// \arg CGF The current codegen function.
626/// \arg IntrinsicID The ID for the Intrinsic we wish to generate.
627/// \arg X The first argument to the llvm.*.with.overflow.*.
628/// \arg Y The second argument to the llvm.*.with.overflow.*.
629/// \arg Carry The carry returned by the llvm.*.with.overflow.*.
630/// \returns The result (i.e. sum/product) returned by the intrinsic.
631static llvm::Value *EmitOverflowIntrinsic(CodeGenFunction &CGF,
632 const llvm::Intrinsic::ID IntrinsicID,
633 llvm::Value *X, llvm::Value *Y,
634 llvm::Value *&Carry) {
635 // Make sure we have integers of the same width.
636 assert(X->getType() == Y->getType() &&((void)0)
637 "Arguments must be the same type. (Did you forget to make sure both "((void)0)
638 "arguments have the same integer width?)")((void)0);
639
640 Function *Callee = CGF.CGM.getIntrinsic(IntrinsicID, X->getType());
641 llvm::Value *Tmp = CGF.Builder.CreateCall(Callee, {X, Y});
642 Carry = CGF.Builder.CreateExtractValue(Tmp, 1);
643 return CGF.Builder.CreateExtractValue(Tmp, 0);
644}
645
646static Value *emitRangedBuiltin(CodeGenFunction &CGF,
647 unsigned IntrinsicID,
648 int low, int high) {
649 llvm::MDBuilder MDHelper(CGF.getLLVMContext());
650 llvm::MDNode *RNode = MDHelper.createRange(APInt(32, low), APInt(32, high));
651 Function *F = CGF.CGM.getIntrinsic(IntrinsicID, {});
652 llvm::Instruction *Call = CGF.Builder.CreateCall(F);
653 Call->setMetadata(llvm::LLVMContext::MD_range, RNode);
654 return Call;
655}
656
657namespace {
658 struct WidthAndSignedness {
659 unsigned Width;
660 bool Signed;
661 };
662}
663
664static WidthAndSignedness
665getIntegerWidthAndSignedness(const clang::ASTContext &context,
666 const clang::QualType Type) {
667 assert(Type->isIntegerType() && "Given type is not an integer.")((void)0);
668 unsigned Width = Type->isBooleanType() ? 1
669 : Type->isExtIntType() ? context.getIntWidth(Type)
670 : context.getTypeInfo(Type).Width;
671 bool Signed = Type->isSignedIntegerType();
672 return {Width, Signed};
673}
674
675// Given one or more integer types, this function produces an integer type that
676// encompasses them: any value in one of the given types could be expressed in
677// the encompassing type.
678static struct WidthAndSignedness
679EncompassingIntegerType(ArrayRef<struct WidthAndSignedness> Types) {
680 assert(Types.size() > 0 && "Empty list of types.")((void)0);
681
682 // If any of the given types is signed, we must return a signed type.
683 bool Signed = false;
684 for (const auto &Type : Types) {
685 Signed |= Type.Signed;
686 }
687
688 // The encompassing type must have a width greater than or equal to the width
689 // of the specified types. Additionally, if the encompassing type is signed,
690 // its width must be strictly greater than the width of any unsigned types
691 // given.
692 unsigned Width = 0;
693 for (const auto &Type : Types) {
694 unsigned MinWidth = Type.Width + (Signed && !Type.Signed);
695 if (Width < MinWidth) {
696 Width = MinWidth;
697 }
698 }
699
700 return {Width, Signed};
701}
702
703Value *CodeGenFunction::EmitVAStartEnd(Value *ArgValue, bool IsStart) {
704 llvm::Type *DestType = Int8PtrTy;
705 if (ArgValue->getType() != DestType)
706 ArgValue =
707 Builder.CreateBitCast(ArgValue, DestType, ArgValue->getName().data());
708
709 Intrinsic::ID inst = IsStart ? Intrinsic::vastart : Intrinsic::vaend;
710 return Builder.CreateCall(CGM.getIntrinsic(inst), ArgValue);
711}
712
713/// Checks if using the result of __builtin_object_size(p, @p From) in place of
714/// __builtin_object_size(p, @p To) is correct
715static bool areBOSTypesCompatible(int From, int To) {
716 // Note: Our __builtin_object_size implementation currently treats Type=0 and
717 // Type=2 identically. Encoding this implementation detail here may make
718 // improving __builtin_object_size difficult in the future, so it's omitted.
719 return From == To || (From == 0 && To == 1) || (From == 3 && To == 2);
720}
721
722static llvm::Value *
723getDefaultBuiltinObjectSizeResult(unsigned Type, llvm::IntegerType *ResType) {
724 return ConstantInt::get(ResType, (Type & 2) ? 0 : -1, /*isSigned=*/true);
725}
726
727llvm::Value *
728CodeGenFunction::evaluateOrEmitBuiltinObjectSize(const Expr *E, unsigned Type,
729 llvm::IntegerType *ResType,
730 llvm::Value *EmittedE,
731 bool IsDynamic) {
732 uint64_t ObjectSize;
733 if (!E->tryEvaluateObjectSize(ObjectSize, getContext(), Type))
734 return emitBuiltinObjectSize(E, Type, ResType, EmittedE, IsDynamic);
735 return ConstantInt::get(ResType, ObjectSize, /*isSigned=*/true);
736}
737
738/// Returns a Value corresponding to the size of the given expression.
739/// This Value may be either of the following:
740/// - A llvm::Argument (if E is a param with the pass_object_size attribute on
741/// it)
742/// - A call to the @llvm.objectsize intrinsic
743///
744/// EmittedE is the result of emitting `E` as a scalar expr. If it's non-null
745/// and we wouldn't otherwise try to reference a pass_object_size parameter,
746/// we'll call @llvm.objectsize on EmittedE, rather than emitting E.
747llvm::Value *
748CodeGenFunction::emitBuiltinObjectSize(const Expr *E, unsigned Type,
749 llvm::IntegerType *ResType,
750 llvm::Value *EmittedE, bool IsDynamic) {
751 // We need to reference an argument if the pointer is a parameter with the
752 // pass_object_size attribute.
753 if (auto *D = dyn_cast<DeclRefExpr>(E->IgnoreParenImpCasts())) {
754 auto *Param = dyn_cast<ParmVarDecl>(D->getDecl());
755 auto *PS = D->getDecl()->getAttr<PassObjectSizeAttr>();
756 if (Param != nullptr && PS != nullptr &&
757 areBOSTypesCompatible(PS->getType(), Type)) {
758 auto Iter = SizeArguments.find(Param);
759 assert(Iter != SizeArguments.end())((void)0);
760
761 const ImplicitParamDecl *D = Iter->second;
762 auto DIter = LocalDeclMap.find(D);
763 assert(DIter != LocalDeclMap.end())((void)0);
764
765 return EmitLoadOfScalar(DIter->second, /*Volatile=*/false,
766 getContext().getSizeType(), E->getBeginLoc());
767 }
768 }
769
770 // LLVM can't handle Type=3 appropriately, and __builtin_object_size shouldn't
771 // evaluate E for side-effects. In either case, we shouldn't lower to
772 // @llvm.objectsize.
773 if (Type == 3 || (!EmittedE && E->HasSideEffects(getContext())))
774 return getDefaultBuiltinObjectSizeResult(Type, ResType);
775
776 Value *Ptr = EmittedE ? EmittedE : EmitScalarExpr(E);
777 assert(Ptr->getType()->isPointerTy() &&((void)0)
778 "Non-pointer passed to __builtin_object_size?")((void)0);
779
780 Function *F =
781 CGM.getIntrinsic(Intrinsic::objectsize, {ResType, Ptr->getType()});
782
783 // LLVM only supports 0 and 2, make sure that we pass along that as a boolean.
784 Value *Min = Builder.getInt1((Type & 2) != 0);
785 // For GCC compatibility, __builtin_object_size treat NULL as unknown size.
786 Value *NullIsUnknown = Builder.getTrue();
787 Value *Dynamic = Builder.getInt1(IsDynamic);
788 return Builder.CreateCall(F, {Ptr, Min, NullIsUnknown, Dynamic});
789}
790
791namespace {
792/// A struct to generically describe a bit test intrinsic.
793struct BitTest {
794 enum ActionKind : uint8_t { TestOnly, Complement, Reset, Set };
795 enum InterlockingKind : uint8_t {
796 Unlocked,
797 Sequential,
798 Acquire,
799 Release,
800 NoFence
801 };
802
803 ActionKind Action;
804 InterlockingKind Interlocking;
805 bool Is64Bit;
806
807 static BitTest decodeBitTestBuiltin(unsigned BuiltinID);
808};
809} // namespace
810
811BitTest BitTest::decodeBitTestBuiltin(unsigned BuiltinID) {
812 switch (BuiltinID) {
813 // Main portable variants.
814 case Builtin::BI_bittest:
815 return {TestOnly, Unlocked, false};
816 case Builtin::BI_bittestandcomplement:
817 return {Complement, Unlocked, false};
818 case Builtin::BI_bittestandreset:
819 return {Reset, Unlocked, false};
820 case Builtin::BI_bittestandset:
821 return {Set, Unlocked, false};
822 case Builtin::BI_interlockedbittestandreset:
823 return {Reset, Sequential, false};
824 case Builtin::BI_interlockedbittestandset:
825 return {Set, Sequential, false};
826
827 // X86-specific 64-bit variants.
828 case Builtin::BI_bittest64:
829 return {TestOnly, Unlocked, true};
830 case Builtin::BI_bittestandcomplement64:
831 return {Complement, Unlocked, true};
832 case Builtin::BI_bittestandreset64:
833 return {Reset, Unlocked, true};
834 case Builtin::BI_bittestandset64:
835 return {Set, Unlocked, true};
836 case Builtin::BI_interlockedbittestandreset64:
837 return {Reset, Sequential, true};
838 case Builtin::BI_interlockedbittestandset64:
839 return {Set, Sequential, true};
840
841 // ARM/AArch64-specific ordering variants.
842 case Builtin::BI_interlockedbittestandset_acq:
843 return {Set, Acquire, false};
844 case Builtin::BI_interlockedbittestandset_rel:
845 return {Set, Release, false};
846 case Builtin::BI_interlockedbittestandset_nf:
847 return {Set, NoFence, false};
848 case Builtin::BI_interlockedbittestandreset_acq:
849 return {Reset, Acquire, false};
850 case Builtin::BI_interlockedbittestandreset_rel:
851 return {Reset, Release, false};
852 case Builtin::BI_interlockedbittestandreset_nf:
853 return {Reset, NoFence, false};
854 }
855 llvm_unreachable("expected only bittest intrinsics")__builtin_unreachable();
856}
857
858static char bitActionToX86BTCode(BitTest::ActionKind A) {
859 switch (A) {
860 case BitTest::TestOnly: return '\0';
861 case BitTest::Complement: return 'c';
862 case BitTest::Reset: return 'r';
863 case BitTest::Set: return 's';
864 }
865 llvm_unreachable("invalid action")__builtin_unreachable();
866}
867
868static llvm::Value *EmitX86BitTestIntrinsic(CodeGenFunction &CGF,
869 BitTest BT,
870 const CallExpr *E, Value *BitBase,
871 Value *BitPos) {
872 char Action = bitActionToX86BTCode(BT.Action);
873 char SizeSuffix = BT.Is64Bit ? 'q' : 'l';
874
875 // Build the assembly.
876 SmallString<64> Asm;
877 raw_svector_ostream AsmOS(Asm);
878 if (BT.Interlocking != BitTest::Unlocked)
879 AsmOS << "lock ";
880 AsmOS << "bt";
881 if (Action)
882 AsmOS << Action;
883 AsmOS << SizeSuffix << " $2, ($1)";
884
885 // Build the constraints. FIXME: We should support immediates when possible.
886 std::string Constraints = "={@ccc},r,r,~{cc},~{memory}";
887 std::string MachineClobbers = CGF.getTarget().getClobbers();
888 if (!MachineClobbers.empty()) {
889 Constraints += ',';
890 Constraints += MachineClobbers;
891 }
892 llvm::IntegerType *IntType = llvm::IntegerType::get(
893 CGF.getLLVMContext(),
894 CGF.getContext().getTypeSize(E->getArg(1)->getType()));
895 llvm::Type *IntPtrType = IntType->getPointerTo();
896 llvm::FunctionType *FTy =
897 llvm::FunctionType::get(CGF.Int8Ty, {IntPtrType, IntType}, false);
898
899 llvm::InlineAsm *IA =
900 llvm::InlineAsm::get(FTy, Asm, Constraints, /*hasSideEffects=*/true);
901 return CGF.Builder.CreateCall(IA, {BitBase, BitPos});
902}
903
904static llvm::AtomicOrdering
905getBitTestAtomicOrdering(BitTest::InterlockingKind I) {
906 switch (I) {
907 case BitTest::Unlocked: return llvm::AtomicOrdering::NotAtomic;
908 case BitTest::Sequential: return llvm::AtomicOrdering::SequentiallyConsistent;
909 case BitTest::Acquire: return llvm::AtomicOrdering::Acquire;
910 case BitTest::Release: return llvm::AtomicOrdering::Release;
911 case BitTest::NoFence: return llvm::AtomicOrdering::Monotonic;
912 }
913 llvm_unreachable("invalid interlocking")__builtin_unreachable();
914}
915
916/// Emit a _bittest* intrinsic. These intrinsics take a pointer to an array of
917/// bits and a bit position and read and optionally modify the bit at that
918/// position. The position index can be arbitrarily large, i.e. it can be larger
919/// than 31 or 63, so we need an indexed load in the general case.
920static llvm::Value *EmitBitTestIntrinsic(CodeGenFunction &CGF,
921 unsigned BuiltinID,
922 const CallExpr *E) {
923 Value *BitBase = CGF.EmitScalarExpr(E->getArg(0));
924 Value *BitPos = CGF.EmitScalarExpr(E->getArg(1));
925
926 BitTest BT = BitTest::decodeBitTestBuiltin(BuiltinID);
927
928 // X86 has special BT, BTC, BTR, and BTS instructions that handle the array
929 // indexing operation internally. Use them if possible.
930 if (CGF.getTarget().getTriple().isX86())
931 return EmitX86BitTestIntrinsic(CGF, BT, E, BitBase, BitPos);
932
933 // Otherwise, use generic code to load one byte and test the bit. Use all but
934 // the bottom three bits as the array index, and the bottom three bits to form
935 // a mask.
936 // Bit = BitBaseI8[BitPos >> 3] & (1 << (BitPos & 0x7)) != 0;
937 Value *ByteIndex = CGF.Builder.CreateAShr(
938 BitPos, llvm::ConstantInt::get(BitPos->getType(), 3), "bittest.byteidx");
939 Value *BitBaseI8 = CGF.Builder.CreatePointerCast(BitBase, CGF.Int8PtrTy);
940 Address ByteAddr(CGF.Builder.CreateInBoundsGEP(CGF.Int8Ty, BitBaseI8,
941 ByteIndex, "bittest.byteaddr"),
942 CharUnits::One());
943 Value *PosLow =
944 CGF.Builder.CreateAnd(CGF.Builder.CreateTrunc(BitPos, CGF.Int8Ty),
945 llvm::ConstantInt::get(CGF.Int8Ty, 0x7));
946
947 // The updating instructions will need a mask.
948 Value *Mask = nullptr;
949 if (BT.Action != BitTest::TestOnly) {
950 Mask = CGF.Builder.CreateShl(llvm::ConstantInt::get(CGF.Int8Ty, 1), PosLow,
951 "bittest.mask");
952 }
953
954 // Check the action and ordering of the interlocked intrinsics.
955 llvm::AtomicOrdering Ordering = getBitTestAtomicOrdering(BT.Interlocking);
956
957 Value *OldByte = nullptr;
958 if (Ordering != llvm::AtomicOrdering::NotAtomic) {
959 // Emit a combined atomicrmw load/store operation for the interlocked
960 // intrinsics.
961 llvm::AtomicRMWInst::BinOp RMWOp = llvm::AtomicRMWInst::Or;
962 if (BT.Action == BitTest::Reset) {
963 Mask = CGF.Builder.CreateNot(Mask);
964 RMWOp = llvm::AtomicRMWInst::And;
965 }
966 OldByte = CGF.Builder.CreateAtomicRMW(RMWOp, ByteAddr.getPointer(), Mask,
967 Ordering);
968 } else {
969 // Emit a plain load for the non-interlocked intrinsics.
970 OldByte = CGF.Builder.CreateLoad(ByteAddr, "bittest.byte");
971 Value *NewByte = nullptr;
972 switch (BT.Action) {
973 case BitTest::TestOnly:
974 // Don't store anything.
975 break;
976 case BitTest::Complement:
977 NewByte = CGF.Builder.CreateXor(OldByte, Mask);
978 break;
979 case BitTest::Reset:
980 NewByte = CGF.Builder.CreateAnd(OldByte, CGF.Builder.CreateNot(Mask));
981 break;
982 case BitTest::Set:
983 NewByte = CGF.Builder.CreateOr(OldByte, Mask);
984 break;
985 }
986 if (NewByte)
987 CGF.Builder.CreateStore(NewByte, ByteAddr);
988 }
989
990 // However we loaded the old byte, either by plain load or atomicrmw, shift
991 // the bit into the low position and mask it to 0 or 1.
992 Value *ShiftedByte = CGF.Builder.CreateLShr(OldByte, PosLow, "bittest.shr");
993 return CGF.Builder.CreateAnd(
994 ShiftedByte, llvm::ConstantInt::get(CGF.Int8Ty, 1), "bittest.res");
995}
996
997static llvm::Value *emitPPCLoadReserveIntrinsic(CodeGenFunction &CGF,
998 unsigned BuiltinID,
999 const CallExpr *E) {
1000 Value *Addr = CGF.EmitScalarExpr(E->getArg(0));
1001
1002 SmallString<64> Asm;
1003 raw_svector_ostream AsmOS(Asm);
1004 llvm::IntegerType *RetType = CGF.Int32Ty;
1005
1006 switch (BuiltinID) {
1007 case clang::PPC::BI__builtin_ppc_ldarx:
1008 AsmOS << "ldarx ";
1009 RetType = CGF.Int64Ty;
1010 break;
1011 case clang::PPC::BI__builtin_ppc_lwarx:
1012 AsmOS << "lwarx ";
1013 RetType = CGF.Int32Ty;
1014 break;
1015 case clang::PPC::BI__builtin_ppc_lharx:
1016 AsmOS << "lharx ";
1017 RetType = CGF.Int16Ty;
1018 break;
1019 case clang::PPC::BI__builtin_ppc_lbarx:
1020 AsmOS << "lbarx ";
1021 RetType = CGF.Int8Ty;
1022 break;
1023 default:
1024 llvm_unreachable("Expected only PowerPC load reserve intrinsics")__builtin_unreachable();
1025 }
1026
1027 AsmOS << "$0, ${1:y}";
1028
1029 std::string Constraints = "=r,*Z,~{memory}";
1030 std::string MachineClobbers = CGF.getTarget().getClobbers();
1031 if (!MachineClobbers.empty()) {
1032 Constraints += ',';
1033 Constraints += MachineClobbers;
1034 }
1035
1036 llvm::Type *IntPtrType = RetType->getPointerTo();
1037 llvm::FunctionType *FTy =
1038 llvm::FunctionType::get(RetType, {IntPtrType}, false);
1039
1040 llvm::InlineAsm *IA =
1041 llvm::InlineAsm::get(FTy, Asm, Constraints, /*hasSideEffects=*/true);
1042 return CGF.Builder.CreateCall(IA, {Addr});
1043}
1044
1045namespace {
1046enum class MSVCSetJmpKind {
1047 _setjmpex,
1048 _setjmp3,
1049 _setjmp
1050};
1051}
1052
1053/// MSVC handles setjmp a bit differently on different platforms. On every
1054/// architecture except 32-bit x86, the frame address is passed. On x86, extra
1055/// parameters can be passed as variadic arguments, but we always pass none.
1056static RValue EmitMSVCRTSetJmp(CodeGenFunction &CGF, MSVCSetJmpKind SJKind,
1057 const CallExpr *E) {
1058 llvm::Value *Arg1 = nullptr;
1059 llvm::Type *Arg1Ty = nullptr;
1060 StringRef Name;
1061 bool IsVarArg = false;
1062 if (SJKind == MSVCSetJmpKind::_setjmp3) {
1063 Name = "_setjmp3";
1064 Arg1Ty = CGF.Int32Ty;
1065 Arg1 = llvm::ConstantInt::get(CGF.IntTy, 0);
1066 IsVarArg = true;
1067 } else {
1068 Name = SJKind == MSVCSetJmpKind::_setjmp ? "_setjmp" : "_setjmpex";
1069 Arg1Ty = CGF.Int8PtrTy;
1070 if (CGF.getTarget().getTriple().getArch() == llvm::Triple::aarch64) {
1071 Arg1 = CGF.Builder.CreateCall(
1072 CGF.CGM.getIntrinsic(Intrinsic::sponentry, CGF.AllocaInt8PtrTy));
1073 } else
1074 Arg1 = CGF.Builder.CreateCall(
1075 CGF.CGM.getIntrinsic(Intrinsic::frameaddress, CGF.AllocaInt8PtrTy),
1076 llvm::ConstantInt::get(CGF.Int32Ty, 0));
1077 }
1078
1079 // Mark the call site and declaration with ReturnsTwice.
1080 llvm::Type *ArgTypes[2] = {CGF.Int8PtrTy, Arg1Ty};
1081 llvm::AttributeList ReturnsTwiceAttr = llvm::AttributeList::get(
1082 CGF.getLLVMContext(), llvm::AttributeList::FunctionIndex,
1083 llvm::Attribute::ReturnsTwice);
1084 llvm::FunctionCallee SetJmpFn = CGF.CGM.CreateRuntimeFunction(
1085 llvm::FunctionType::get(CGF.IntTy, ArgTypes, IsVarArg), Name,
1086 ReturnsTwiceAttr, /*Local=*/true);
1087
1088 llvm::Value *Buf = CGF.Builder.CreateBitOrPointerCast(
1089 CGF.EmitScalarExpr(E->getArg(0)), CGF.Int8PtrTy);
1090 llvm::Value *Args[] = {Buf, Arg1};
1091 llvm::CallBase *CB = CGF.EmitRuntimeCallOrInvoke(SetJmpFn, Args);
1092 CB->setAttributes(ReturnsTwiceAttr);
1093 return RValue::get(CB);
1094}
1095
1096// Many of MSVC builtins are on x64, ARM and AArch64; to avoid repeating code,
1097// we handle them here.
1098enum class CodeGenFunction::MSVCIntrin {
1099 _BitScanForward,
1100 _BitScanReverse,
1101 _InterlockedAnd,
1102 _InterlockedDecrement,
1103 _InterlockedExchange,
1104 _InterlockedExchangeAdd,
1105 _InterlockedExchangeSub,
1106 _InterlockedIncrement,
1107 _InterlockedOr,
1108 _InterlockedXor,
1109 _InterlockedExchangeAdd_acq,
1110 _InterlockedExchangeAdd_rel,
1111 _InterlockedExchangeAdd_nf,
1112 _InterlockedExchange_acq,
1113 _InterlockedExchange_rel,
1114 _InterlockedExchange_nf,
1115 _InterlockedCompareExchange_acq,
1116 _InterlockedCompareExchange_rel,
1117 _InterlockedCompareExchange_nf,
1118 _InterlockedCompareExchange128,
1119 _InterlockedCompareExchange128_acq,
1120 _InterlockedCompareExchange128_rel,
1121 _InterlockedCompareExchange128_nf,
1122 _InterlockedOr_acq,
1123 _InterlockedOr_rel,
1124 _InterlockedOr_nf,
1125 _InterlockedXor_acq,
1126 _InterlockedXor_rel,
1127 _InterlockedXor_nf,
1128 _InterlockedAnd_acq,
1129 _InterlockedAnd_rel,
1130 _InterlockedAnd_nf,
1131 _InterlockedIncrement_acq,
1132 _InterlockedIncrement_rel,
1133 _InterlockedIncrement_nf,
1134 _InterlockedDecrement_acq,
1135 _InterlockedDecrement_rel,
1136 _InterlockedDecrement_nf,
1137 __fastfail,
1138};
1139
1140static Optional<CodeGenFunction::MSVCIntrin>
1141translateArmToMsvcIntrin(unsigned BuiltinID) {
1142 using MSVCIntrin = CodeGenFunction::MSVCIntrin;
1143 switch (BuiltinID) {
1144 default:
1145 return None;
1146 case ARM::BI_BitScanForward:
1147 case ARM::BI_BitScanForward64:
1148 return MSVCIntrin::_BitScanForward;
1149 case ARM::BI_BitScanReverse:
1150 case ARM::BI_BitScanReverse64:
1151 return MSVCIntrin::_BitScanReverse;
1152 case ARM::BI_InterlockedAnd64:
1153 return MSVCIntrin::_InterlockedAnd;
1154 case ARM::BI_InterlockedExchange64:
1155 return MSVCIntrin::_InterlockedExchange;
1156 case ARM::BI_InterlockedExchangeAdd64:
1157 return MSVCIntrin::_InterlockedExchangeAdd;
1158 case ARM::BI_InterlockedExchangeSub64:
1159 return MSVCIntrin::_InterlockedExchangeSub;
1160 case ARM::BI_InterlockedOr64:
1161 return MSVCIntrin::_InterlockedOr;
1162 case ARM::BI_InterlockedXor64:
1163 return MSVCIntrin::_InterlockedXor;
1164 case ARM::BI_InterlockedDecrement64:
1165 return MSVCIntrin::_InterlockedDecrement;
1166 case ARM::BI_InterlockedIncrement64:
1167 return MSVCIntrin::_InterlockedIncrement;
1168 case ARM::BI_InterlockedExchangeAdd8_acq:
1169 case ARM::BI_InterlockedExchangeAdd16_acq:
1170 case ARM::BI_InterlockedExchangeAdd_acq:
1171 case ARM::BI_InterlockedExchangeAdd64_acq:
1172 return MSVCIntrin::_InterlockedExchangeAdd_acq;
1173 case ARM::BI_InterlockedExchangeAdd8_rel:
1174 case ARM::BI_InterlockedExchangeAdd16_rel:
1175 case ARM::BI_InterlockedExchangeAdd_rel:
1176 case ARM::BI_InterlockedExchangeAdd64_rel:
1177 return MSVCIntrin::_InterlockedExchangeAdd_rel;
1178 case ARM::BI_InterlockedExchangeAdd8_nf:
1179 case ARM::BI_InterlockedExchangeAdd16_nf:
1180 case ARM::BI_InterlockedExchangeAdd_nf:
1181 case ARM::BI_InterlockedExchangeAdd64_nf:
1182 return MSVCIntrin::_InterlockedExchangeAdd_nf;
1183 case ARM::BI_InterlockedExchange8_acq:
1184 case ARM::BI_InterlockedExchange16_acq:
1185 case ARM::BI_InterlockedExchange_acq:
1186 case ARM::BI_InterlockedExchange64_acq:
1187 return MSVCIntrin::_InterlockedExchange_acq;
1188 case ARM::BI_InterlockedExchange8_rel:
1189 case ARM::BI_InterlockedExchange16_rel:
1190 case ARM::BI_InterlockedExchange_rel:
1191 case ARM::BI_InterlockedExchange64_rel:
1192 return MSVCIntrin::_InterlockedExchange_rel;
1193 case ARM::BI_InterlockedExchange8_nf:
1194 case ARM::BI_InterlockedExchange16_nf:
1195 case ARM::BI_InterlockedExchange_nf:
1196 case ARM::BI_InterlockedExchange64_nf:
1197 return MSVCIntrin::_InterlockedExchange_nf;
1198 case ARM::BI_InterlockedCompareExchange8_acq:
1199 case ARM::BI_InterlockedCompareExchange16_acq:
1200 case ARM::BI_InterlockedCompareExchange_acq:
1201 case ARM::BI_InterlockedCompareExchange64_acq:
1202 return MSVCIntrin::_InterlockedCompareExchange_acq;
1203 case ARM::BI_InterlockedCompareExchange8_rel:
1204 case ARM::BI_InterlockedCompareExchange16_rel:
1205 case ARM::BI_InterlockedCompareExchange_rel:
1206 case ARM::BI_InterlockedCompareExchange64_rel:
1207 return MSVCIntrin::_InterlockedCompareExchange_rel;
1208 case ARM::BI_InterlockedCompareExchange8_nf:
1209 case ARM::BI_InterlockedCompareExchange16_nf:
1210 case ARM::BI_InterlockedCompareExchange_nf:
1211 case ARM::BI_InterlockedCompareExchange64_nf:
1212 return MSVCIntrin::_InterlockedCompareExchange_nf;
1213 case ARM::BI_InterlockedOr8_acq:
1214 case ARM::BI_InterlockedOr16_acq:
1215 case ARM::BI_InterlockedOr_acq:
1216 case ARM::BI_InterlockedOr64_acq:
1217 return MSVCIntrin::_InterlockedOr_acq;
1218 case ARM::BI_InterlockedOr8_rel:
1219 case ARM::BI_InterlockedOr16_rel:
1220 case ARM::BI_InterlockedOr_rel:
1221 case ARM::BI_InterlockedOr64_rel:
1222 return MSVCIntrin::_InterlockedOr_rel;
1223 case ARM::BI_InterlockedOr8_nf:
1224 case ARM::BI_InterlockedOr16_nf:
1225 case ARM::BI_InterlockedOr_nf:
1226 case ARM::BI_InterlockedOr64_nf:
1227 return MSVCIntrin::_InterlockedOr_nf;
1228 case ARM::BI_InterlockedXor8_acq:
1229 case ARM::BI_InterlockedXor16_acq:
1230 case ARM::BI_InterlockedXor_acq:
1231 case ARM::BI_InterlockedXor64_acq:
1232 return MSVCIntrin::_InterlockedXor_acq;
1233 case ARM::BI_InterlockedXor8_rel:
1234 case ARM::BI_InterlockedXor16_rel:
1235 case ARM::BI_InterlockedXor_rel:
1236 case ARM::BI_InterlockedXor64_rel:
1237 return MSVCIntrin::_InterlockedXor_rel;
1238 case ARM::BI_InterlockedXor8_nf:
1239 case ARM::BI_InterlockedXor16_nf:
1240 case ARM::BI_InterlockedXor_nf:
1241 case ARM::BI_InterlockedXor64_nf:
1242 return MSVCIntrin::_InterlockedXor_nf;
1243 case ARM::BI_InterlockedAnd8_acq:
1244 case ARM::BI_InterlockedAnd16_acq:
1245 case ARM::BI_InterlockedAnd_acq:
1246 case ARM::BI_InterlockedAnd64_acq:
1247 return MSVCIntrin::_InterlockedAnd_acq;
1248 case ARM::BI_InterlockedAnd8_rel:
1249 case ARM::BI_InterlockedAnd16_rel:
1250 case ARM::BI_InterlockedAnd_rel:
1251 case ARM::BI_InterlockedAnd64_rel:
1252 return MSVCIntrin::_InterlockedAnd_rel;
1253 case ARM::BI_InterlockedAnd8_nf:
1254 case ARM::BI_InterlockedAnd16_nf:
1255 case ARM::BI_InterlockedAnd_nf:
1256 case ARM::BI_InterlockedAnd64_nf:
1257 return MSVCIntrin::_InterlockedAnd_nf;
1258 case ARM::BI_InterlockedIncrement16_acq:
1259 case ARM::BI_InterlockedIncrement_acq:
1260 case ARM::BI_InterlockedIncrement64_acq:
1261 return MSVCIntrin::_InterlockedIncrement_acq;
1262 case ARM::BI_InterlockedIncrement16_rel:
1263 case ARM::BI_InterlockedIncrement_rel:
1264 case ARM::BI_InterlockedIncrement64_rel:
1265 return MSVCIntrin::_InterlockedIncrement_rel;
1266 case ARM::BI_InterlockedIncrement16_nf:
1267 case ARM::BI_InterlockedIncrement_nf:
1268 case ARM::BI_InterlockedIncrement64_nf:
1269 return MSVCIntrin::_InterlockedIncrement_nf;
1270 case ARM::BI_InterlockedDecrement16_acq:
1271 case ARM::BI_InterlockedDecrement_acq:
1272 case ARM::BI_InterlockedDecrement64_acq:
1273 return MSVCIntrin::_InterlockedDecrement_acq;
1274 case ARM::BI_InterlockedDecrement16_rel:
1275 case ARM::BI_InterlockedDecrement_rel:
1276 case ARM::BI_InterlockedDecrement64_rel:
1277 return MSVCIntrin::_InterlockedDecrement_rel;
1278 case ARM::BI_InterlockedDecrement16_nf:
1279 case ARM::BI_InterlockedDecrement_nf:
1280 case ARM::BI_InterlockedDecrement64_nf:
1281 return MSVCIntrin::_InterlockedDecrement_nf;
1282 }
1283 llvm_unreachable("must return from switch")__builtin_unreachable();
1284}
1285
1286static Optional<CodeGenFunction::MSVCIntrin>
1287translateAarch64ToMsvcIntrin(unsigned BuiltinID) {
1288 using MSVCIntrin = CodeGenFunction::MSVCIntrin;
1289 switch (BuiltinID) {
1290 default:
1291 return None;
1292 case AArch64::BI_BitScanForward:
1293 case AArch64::BI_BitScanForward64:
1294 return MSVCIntrin::_BitScanForward;
1295 case AArch64::BI_BitScanReverse:
1296 case AArch64::BI_BitScanReverse64:
1297 return MSVCIntrin::_BitScanReverse;
1298 case AArch64::BI_InterlockedAnd64:
1299 return MSVCIntrin::_InterlockedAnd;
1300 case AArch64::BI_InterlockedExchange64:
1301 return MSVCIntrin::_InterlockedExchange;
1302 case AArch64::BI_InterlockedExchangeAdd64:
1303 return MSVCIntrin::_InterlockedExchangeAdd;
1304 case AArch64::BI_InterlockedExchangeSub64:
1305 return MSVCIntrin::_InterlockedExchangeSub;
1306 case AArch64::BI_InterlockedOr64:
1307 return MSVCIntrin::_InterlockedOr;
1308 case AArch64::BI_InterlockedXor64:
1309 return MSVCIntrin::_InterlockedXor;
1310 case AArch64::BI_InterlockedDecrement64:
1311 return MSVCIntrin::_InterlockedDecrement;
1312 case AArch64::BI_InterlockedIncrement64:
1313 return MSVCIntrin::_InterlockedIncrement;
1314 case AArch64::BI_InterlockedExchangeAdd8_acq:
1315 case AArch64::BI_InterlockedExchangeAdd16_acq:
1316 case AArch64::BI_InterlockedExchangeAdd_acq:
1317 case AArch64::BI_InterlockedExchangeAdd64_acq:
1318 return MSVCIntrin::_InterlockedExchangeAdd_acq;
1319 case AArch64::BI_InterlockedExchangeAdd8_rel:
1320 case AArch64::BI_InterlockedExchangeAdd16_rel:
1321 case AArch64::BI_InterlockedExchangeAdd_rel:
1322 case AArch64::BI_InterlockedExchangeAdd64_rel:
1323 return MSVCIntrin::_InterlockedExchangeAdd_rel;
1324 case AArch64::BI_InterlockedExchangeAdd8_nf:
1325 case AArch64::BI_InterlockedExchangeAdd16_nf:
1326 case AArch64::BI_InterlockedExchangeAdd_nf:
1327 case AArch64::BI_InterlockedExchangeAdd64_nf:
1328 return MSVCIntrin::_InterlockedExchangeAdd_nf;
1329 case AArch64::BI_InterlockedExchange8_acq:
1330 case AArch64::BI_InterlockedExchange16_acq:
1331 case AArch64::BI_InterlockedExchange_acq:
1332 case AArch64::BI_InterlockedExchange64_acq:
1333 return MSVCIntrin::_InterlockedExchange_acq;
1334 case AArch64::BI_InterlockedExchange8_rel:
1335 case AArch64::BI_InterlockedExchange16_rel:
1336 case AArch64::BI_InterlockedExchange_rel:
1337 case AArch64::BI_InterlockedExchange64_rel:
1338 return MSVCIntrin::_InterlockedExchange_rel;
1339 case AArch64::BI_InterlockedExchange8_nf:
1340 case AArch64::BI_InterlockedExchange16_nf:
1341 case AArch64::BI_InterlockedExchange_nf:
1342 case AArch64::BI_InterlockedExchange64_nf:
1343 return MSVCIntrin::_InterlockedExchange_nf;
1344 case AArch64::BI_InterlockedCompareExchange8_acq:
1345 case AArch64::BI_InterlockedCompareExchange16_acq:
1346 case AArch64::BI_InterlockedCompareExchange_acq:
1347 case AArch64::BI_InterlockedCompareExchange64_acq:
1348 return MSVCIntrin::_InterlockedCompareExchange_acq;
1349 case AArch64::BI_InterlockedCompareExchange8_rel:
1350 case AArch64::BI_InterlockedCompareExchange16_rel:
1351 case AArch64::BI_InterlockedCompareExchange_rel:
1352 case AArch64::BI_InterlockedCompareExchange64_rel:
1353 return MSVCIntrin::_InterlockedCompareExchange_rel;
1354 case AArch64::BI_InterlockedCompareExchange8_nf:
1355 case AArch64::BI_InterlockedCompareExchange16_nf:
1356 case AArch64::BI_InterlockedCompareExchange_nf:
1357 case AArch64::BI_InterlockedCompareExchange64_nf:
1358 return MSVCIntrin::_InterlockedCompareExchange_nf;
1359 case AArch64::BI_InterlockedCompareExchange128:
1360 return MSVCIntrin::_InterlockedCompareExchange128;
1361 case AArch64::BI_InterlockedCompareExchange128_acq:
1362 return MSVCIntrin::_InterlockedCompareExchange128_acq;
1363 case AArch64::BI_InterlockedCompareExchange128_nf:
1364 return MSVCIntrin::_InterlockedCompareExchange128_nf;
1365 case AArch64::BI_InterlockedCompareExchange128_rel:
1366 return MSVCIntrin::_InterlockedCompareExchange128_rel;
1367 case AArch64::BI_InterlockedOr8_acq:
1368 case AArch64::BI_InterlockedOr16_acq:
1369 case AArch64::BI_InterlockedOr_acq:
1370 case AArch64::BI_InterlockedOr64_acq:
1371 return MSVCIntrin::_InterlockedOr_acq;
1372 case AArch64::BI_InterlockedOr8_rel:
1373 case AArch64::BI_InterlockedOr16_rel:
1374 case AArch64::BI_InterlockedOr_rel:
1375 case AArch64::BI_InterlockedOr64_rel:
1376 return MSVCIntrin::_InterlockedOr_rel;
1377 case AArch64::BI_InterlockedOr8_nf:
1378 case AArch64::BI_InterlockedOr16_nf:
1379 case AArch64::BI_InterlockedOr_nf:
1380 case AArch64::BI_InterlockedOr64_nf:
1381 return MSVCIntrin::_InterlockedOr_nf;
1382 case AArch64::BI_InterlockedXor8_acq:
1383 case AArch64::BI_InterlockedXor16_acq:
1384 case AArch64::BI_InterlockedXor_acq:
1385 case AArch64::BI_InterlockedXor64_acq:
1386 return MSVCIntrin::_InterlockedXor_acq;
1387 case AArch64::BI_InterlockedXor8_rel:
1388 case AArch64::BI_InterlockedXor16_rel:
1389 case AArch64::BI_InterlockedXor_rel:
1390 case AArch64::BI_InterlockedXor64_rel:
1391 return MSVCIntrin::_InterlockedXor_rel;
1392 case AArch64::BI_InterlockedXor8_nf:
1393 case AArch64::BI_InterlockedXor16_nf:
1394 case AArch64::BI_InterlockedXor_nf:
1395 case AArch64::BI_InterlockedXor64_nf:
1396 return MSVCIntrin::_InterlockedXor_nf;
1397 case AArch64::BI_InterlockedAnd8_acq:
1398 case AArch64::BI_InterlockedAnd16_acq:
1399 case AArch64::BI_InterlockedAnd_acq:
1400 case AArch64::BI_InterlockedAnd64_acq:
1401 return MSVCIntrin::_InterlockedAnd_acq;
1402 case AArch64::BI_InterlockedAnd8_rel:
1403 case AArch64::BI_InterlockedAnd16_rel:
1404 case AArch64::BI_InterlockedAnd_rel:
1405 case AArch64::BI_InterlockedAnd64_rel:
1406 return MSVCIntrin::_InterlockedAnd_rel;
1407 case AArch64::BI_InterlockedAnd8_nf:
1408 case AArch64::BI_InterlockedAnd16_nf:
1409 case AArch64::BI_InterlockedAnd_nf:
1410 case AArch64::BI_InterlockedAnd64_nf:
1411 return MSVCIntrin::_InterlockedAnd_nf;
1412 case AArch64::BI_InterlockedIncrement16_acq:
1413 case AArch64::BI_InterlockedIncrement_acq:
1414 case AArch64::BI_InterlockedIncrement64_acq:
1415 return MSVCIntrin::_InterlockedIncrement_acq;
1416 case AArch64::BI_InterlockedIncrement16_rel:
1417 case AArch64::BI_InterlockedIncrement_rel:
1418 case AArch64::BI_InterlockedIncrement64_rel:
1419 return MSVCIntrin::_InterlockedIncrement_rel;
1420 case AArch64::BI_InterlockedIncrement16_nf:
1421 case AArch64::BI_InterlockedIncrement_nf:
1422 case AArch64::BI_InterlockedIncrement64_nf:
1423 return MSVCIntrin::_InterlockedIncrement_nf;
1424 case AArch64::BI_InterlockedDecrement16_acq:
1425 case AArch64::BI_InterlockedDecrement_acq:
1426 case AArch64::BI_InterlockedDecrement64_acq:
1427 return MSVCIntrin::_InterlockedDecrement_acq;
1428 case AArch64::BI_InterlockedDecrement16_rel:
1429 case AArch64::BI_InterlockedDecrement_rel:
1430 case AArch64::BI_InterlockedDecrement64_rel:
1431 return MSVCIntrin::_InterlockedDecrement_rel;
1432 case AArch64::BI_InterlockedDecrement16_nf:
1433 case AArch64::BI_InterlockedDecrement_nf:
1434 case AArch64::BI_InterlockedDecrement64_nf:
1435 return MSVCIntrin::_InterlockedDecrement_nf;
1436 }
1437 llvm_unreachable("must return from switch")__builtin_unreachable();
1438}
1439
1440static Optional<CodeGenFunction::MSVCIntrin>
1441translateX86ToMsvcIntrin(unsigned BuiltinID) {
1442 using MSVCIntrin = CodeGenFunction::MSVCIntrin;
1443 switch (BuiltinID) {
1444 default:
1445 return None;
1446 case clang::X86::BI_BitScanForward:
1447 case clang::X86::BI_BitScanForward64:
1448 return MSVCIntrin::_BitScanForward;
1449 case clang::X86::BI_BitScanReverse:
1450 case clang::X86::BI_BitScanReverse64:
1451 return MSVCIntrin::_BitScanReverse;
1452 case clang::X86::BI_InterlockedAnd64:
1453 return MSVCIntrin::_InterlockedAnd;
1454 case clang::X86::BI_InterlockedCompareExchange128:
1455 return MSVCIntrin::_InterlockedCompareExchange128;
1456 case clang::X86::BI_InterlockedExchange64:
1457 return MSVCIntrin::_InterlockedExchange;
1458 case clang::X86::BI_InterlockedExchangeAdd64:
1459 return MSVCIntrin::_InterlockedExchangeAdd;
1460 case clang::X86::BI_InterlockedExchangeSub64:
1461 return MSVCIntrin::_InterlockedExchangeSub;
1462 case clang::X86::BI_InterlockedOr64:
1463 return MSVCIntrin::_InterlockedOr;
1464 case clang::X86::BI_InterlockedXor64:
1465 return MSVCIntrin::_InterlockedXor;
1466 case clang::X86::BI_InterlockedDecrement64:
1467 return MSVCIntrin::_InterlockedDecrement;
1468 case clang::X86::BI_InterlockedIncrement64:
1469 return MSVCIntrin::_InterlockedIncrement;
1470 }
1471 llvm_unreachable("must return from switch")__builtin_unreachable();
1472}
1473
1474// Emit an MSVC intrinsic. Assumes that arguments have *not* been evaluated.
1475Value *CodeGenFunction::EmitMSVCBuiltinExpr(MSVCIntrin BuiltinID,
1476 const CallExpr *E) {
1477 switch (BuiltinID) {
1478 case MSVCIntrin::_BitScanForward:
1479 case MSVCIntrin::_BitScanReverse: {
1480 Address IndexAddress(EmitPointerWithAlignment(E->getArg(0)));
1481 Value *ArgValue = EmitScalarExpr(E->getArg(1));
1482
1483 llvm::Type *ArgType = ArgValue->getType();
1484 llvm::Type *IndexType =
1485 IndexAddress.getPointer()->getType()->getPointerElementType();
1486 llvm::Type *ResultType = ConvertType(E->getType());
1487
1488 Value *ArgZero = llvm::Constant::getNullValue(ArgType);
1489 Value *ResZero = llvm::Constant::getNullValue(ResultType);
1490 Value *ResOne = llvm::ConstantInt::get(ResultType, 1);
1491
1492 BasicBlock *Begin = Builder.GetInsertBlock();
1493 BasicBlock *End = createBasicBlock("bitscan_end", this->CurFn);
1494 Builder.SetInsertPoint(End);
1495 PHINode *Result = Builder.CreatePHI(ResultType, 2, "bitscan_result");
1496
1497 Builder.SetInsertPoint(Begin);
1498 Value *IsZero = Builder.CreateICmpEQ(ArgValue, ArgZero);
1499 BasicBlock *NotZero = createBasicBlock("bitscan_not_zero", this->CurFn);
1500 Builder.CreateCondBr(IsZero, End, NotZero);
1501 Result->addIncoming(ResZero, Begin);
1502
1503 Builder.SetInsertPoint(NotZero);
1504
1505 if (BuiltinID == MSVCIntrin::_BitScanForward) {
1506 Function *F = CGM.getIntrinsic(Intrinsic::cttz, ArgType);
1507 Value *ZeroCount = Builder.CreateCall(F, {ArgValue, Builder.getTrue()});
1508 ZeroCount = Builder.CreateIntCast(ZeroCount, IndexType, false);
1509 Builder.CreateStore(ZeroCount, IndexAddress, false);
1510 } else {
1511 unsigned ArgWidth = cast<llvm::IntegerType>(ArgType)->getBitWidth();
1512 Value *ArgTypeLastIndex = llvm::ConstantInt::get(IndexType, ArgWidth - 1);
1513
1514 Function *F = CGM.getIntrinsic(Intrinsic::ctlz, ArgType);
1515 Value *ZeroCount = Builder.CreateCall(F, {ArgValue, Builder.getTrue()});
1516 ZeroCount = Builder.CreateIntCast(ZeroCount, IndexType, false);
1517 Value *Index = Builder.CreateNSWSub(ArgTypeLastIndex, ZeroCount);
1518 Builder.CreateStore(Index, IndexAddress, false);
1519 }
1520 Builder.CreateBr(End);
1521 Result->addIncoming(ResOne, NotZero);
1522
1523 Builder.SetInsertPoint(End);
1524 return Result;
1525 }
1526 case MSVCIntrin::_InterlockedAnd:
1527 return MakeBinaryAtomicValue(*this, AtomicRMWInst::And, E);
1528 case MSVCIntrin::_InterlockedExchange:
1529 return MakeBinaryAtomicValue(*this, AtomicRMWInst::Xchg, E);
1530 case MSVCIntrin::_InterlockedExchangeAdd:
1531 return MakeBinaryAtomicValue(*this, AtomicRMWInst::Add, E);
1532 case MSVCIntrin::_InterlockedExchangeSub:
1533 return MakeBinaryAtomicValue(*this, AtomicRMWInst::Sub, E);
1534 case MSVCIntrin::_InterlockedOr:
1535 return MakeBinaryAtomicValue(*this, AtomicRMWInst::Or, E);
1536 case MSVCIntrin::_InterlockedXor:
1537 return MakeBinaryAtomicValue(*this, AtomicRMWInst::Xor, E);
1538 case MSVCIntrin::_InterlockedExchangeAdd_acq:
1539 return MakeBinaryAtomicValue(*this, AtomicRMWInst::Add, E,
1540 AtomicOrdering::Acquire);
1541 case MSVCIntrin::_InterlockedExchangeAdd_rel:
1542 return MakeBinaryAtomicValue(*this, AtomicRMWInst::Add, E,
1543 AtomicOrdering::Release);
1544 case MSVCIntrin::_InterlockedExchangeAdd_nf:
1545 return MakeBinaryAtomicValue(*this, AtomicRMWInst::Add, E,
1546 AtomicOrdering::Monotonic);
1547 case MSVCIntrin::_InterlockedExchange_acq:
1548 return MakeBinaryAtomicValue(*this, AtomicRMWInst::Xchg, E,
1549 AtomicOrdering::Acquire);
1550 case MSVCIntrin::_InterlockedExchange_rel:
1551 return MakeBinaryAtomicValue(*this, AtomicRMWInst::Xchg, E,
1552 AtomicOrdering::Release);
1553 case MSVCIntrin::_InterlockedExchange_nf:
1554 return MakeBinaryAtomicValue(*this, AtomicRMWInst::Xchg, E,
1555 AtomicOrdering::Monotonic);
1556 case MSVCIntrin::_InterlockedCompareExchange_acq:
1557 return EmitAtomicCmpXchgForMSIntrin(*this, E, AtomicOrdering::Acquire);
1558 case MSVCIntrin::_InterlockedCompareExchange_rel:
1559 return EmitAtomicCmpXchgForMSIntrin(*this, E, AtomicOrdering::Release);
1560 case MSVCIntrin::_InterlockedCompareExchange_nf:
1561 return EmitAtomicCmpXchgForMSIntrin(*this, E, AtomicOrdering::Monotonic);
1562 case MSVCIntrin::_InterlockedCompareExchange128:
1563 return EmitAtomicCmpXchg128ForMSIntrin(
1564 *this, E, AtomicOrdering::SequentiallyConsistent);
1565 case MSVCIntrin::_InterlockedCompareExchange128_acq:
1566 return EmitAtomicCmpXchg128ForMSIntrin(*this, E, AtomicOrdering::Acquire);
1567 case MSVCIntrin::_InterlockedCompareExchange128_rel:
1568 return EmitAtomicCmpXchg128ForMSIntrin(*this, E, AtomicOrdering::Release);
1569 case MSVCIntrin::_InterlockedCompareExchange128_nf:
1570 return EmitAtomicCmpXchg128ForMSIntrin(*this, E, AtomicOrdering::Monotonic);
1571 case MSVCIntrin::_InterlockedOr_acq:
1572 return MakeBinaryAtomicValue(*this, AtomicRMWInst::Or, E,
1573 AtomicOrdering::Acquire);
1574 case MSVCIntrin::_InterlockedOr_rel:
1575 return MakeBinaryAtomicValue(*this, AtomicRMWInst::Or, E,
1576 AtomicOrdering::Release);
1577 case MSVCIntrin::_InterlockedOr_nf:
1578 return MakeBinaryAtomicValue(*this, AtomicRMWInst::Or, E,
1579 AtomicOrdering::Monotonic);
1580 case MSVCIntrin::_InterlockedXor_acq:
1581 return MakeBinaryAtomicValue(*this, AtomicRMWInst::Xor, E,
1582 AtomicOrdering::Acquire);
1583 case MSVCIntrin::_InterlockedXor_rel:
1584 return MakeBinaryAtomicValue(*this, AtomicRMWInst::Xor, E,
1585 AtomicOrdering::Release);
1586 case MSVCIntrin::_InterlockedXor_nf:
1587 return MakeBinaryAtomicValue(*this, AtomicRMWInst::Xor, E,
1588 AtomicOrdering::Monotonic);
1589 case MSVCIntrin::_InterlockedAnd_acq:
1590 return MakeBinaryAtomicValue(*this, AtomicRMWInst::And, E,
1591 AtomicOrdering::Acquire);
1592 case MSVCIntrin::_InterlockedAnd_rel:
1593 return MakeBinaryAtomicValue(*this, AtomicRMWInst::And, E,
1594 AtomicOrdering::Release);
1595 case MSVCIntrin::_InterlockedAnd_nf:
1596 return MakeBinaryAtomicValue(*this, AtomicRMWInst::And, E,
1597 AtomicOrdering::Monotonic);
1598 case MSVCIntrin::_InterlockedIncrement_acq:
1599 return EmitAtomicIncrementValue(*this, E, AtomicOrdering::Acquire);
1600 case MSVCIntrin::_InterlockedIncrement_rel:
1601 return EmitAtomicIncrementValue(*this, E, AtomicOrdering::Release);
1602 case MSVCIntrin::_InterlockedIncrement_nf:
1603 return EmitAtomicIncrementValue(*this, E, AtomicOrdering::Monotonic);
1604 case MSVCIntrin::_InterlockedDecrement_acq:
1605 return EmitAtomicDecrementValue(*this, E, AtomicOrdering::Acquire);
1606 case MSVCIntrin::_InterlockedDecrement_rel:
1607 return EmitAtomicDecrementValue(*this, E, AtomicOrdering::Release);
1608 case MSVCIntrin::_InterlockedDecrement_nf:
1609 return EmitAtomicDecrementValue(*this, E, AtomicOrdering::Monotonic);
1610
1611 case MSVCIntrin::_InterlockedDecrement:
1612 return EmitAtomicDecrementValue(*this, E);
1613 case MSVCIntrin::_InterlockedIncrement:
1614 return EmitAtomicIncrementValue(*this, E);
1615
1616 case MSVCIntrin::__fastfail: {
1617 // Request immediate process termination from the kernel. The instruction
1618 // sequences to do this are documented on MSDN:
1619 // https://msdn.microsoft.com/en-us/library/dn774154.aspx
1620 llvm::Triple::ArchType ISA = getTarget().getTriple().getArch();
1621 StringRef Asm, Constraints;
1622 switch (ISA) {
1623 default:
1624 ErrorUnsupported(E, "__fastfail call for this architecture");
1625 break;
1626 case llvm::Triple::x86:
1627 case llvm::Triple::x86_64:
1628 Asm = "int $$0x29";
1629 Constraints = "{cx}";
1630 break;
1631 case llvm::Triple::thumb:
1632 Asm = "udf #251";
1633 Constraints = "{r0}";
1634 break;
1635 case llvm::Triple::aarch64:
1636 Asm = "brk #0xF003";
1637 Constraints = "{w0}";
1638 }
1639 llvm::FunctionType *FTy = llvm::FunctionType::get(VoidTy, {Int32Ty}, false);
1640 llvm::InlineAsm *IA =
1641 llvm::InlineAsm::get(FTy, Asm, Constraints, /*hasSideEffects=*/true);
1642 llvm::AttributeList NoReturnAttr = llvm::AttributeList::get(
1643 getLLVMContext(), llvm::AttributeList::FunctionIndex,
1644 llvm::Attribute::NoReturn);
1645 llvm::CallInst *CI = Builder.CreateCall(IA, EmitScalarExpr(E->getArg(0)));
1646 CI->setAttributes(NoReturnAttr);
1647 return CI;
1648 }
1649 }
1650 llvm_unreachable("Incorrect MSVC intrinsic!")__builtin_unreachable();
1651}
1652
1653namespace {
1654// ARC cleanup for __builtin_os_log_format
1655struct CallObjCArcUse final : EHScopeStack::Cleanup {
1656 CallObjCArcUse(llvm::Value *object) : object(object) {}
1657 llvm::Value *object;
1658
1659 void Emit(CodeGenFunction &CGF, Flags flags) override {
1660 CGF.EmitARCIntrinsicUse(object);
1661 }
1662};
1663}
1664
1665Value *CodeGenFunction::EmitCheckedArgForBuiltin(const Expr *E,
1666 BuiltinCheckKind Kind) {
1667 assert((Kind == BCK_CLZPassedZero || Kind == BCK_CTZPassedZero)((void)0)
1668 && "Unsupported builtin check kind")((void)0);
1669
1670 Value *ArgValue = EmitScalarExpr(E);
1671 if (!SanOpts.has(SanitizerKind::Builtin) || !getTarget().isCLZForZeroUndef())
1672 return ArgValue;
1673
1674 SanitizerScope SanScope(this);
1675 Value *Cond = Builder.CreateICmpNE(
1676 ArgValue, llvm::Constant::getNullValue(ArgValue->getType()));
1677 EmitCheck(std::make_pair(Cond, SanitizerKind::Builtin),
1678 SanitizerHandler::InvalidBuiltin,
1679 {EmitCheckSourceLocation(E->getExprLoc()),
1680 llvm::ConstantInt::get(Builder.getInt8Ty(), Kind)},
1681 None);
1682 return ArgValue;
1683}
1684
1685/// Get the argument type for arguments to os_log_helper.
1686static CanQualType getOSLogArgType(ASTContext &C, int Size) {
1687 QualType UnsignedTy = C.getIntTypeForBitwidth(Size * 8, /*Signed=*/false);
1688 return C.getCanonicalType(UnsignedTy);
1689}
1690
1691llvm::Function *CodeGenFunction::generateBuiltinOSLogHelperFunction(
1692 const analyze_os_log::OSLogBufferLayout &Layout,
1693 CharUnits BufferAlignment) {
1694 ASTContext &Ctx = getContext();
1695
1696 llvm::SmallString<64> Name;
1697 {
1698 raw_svector_ostream OS(Name);
1699 OS << "__os_log_helper";
1700 OS << "_" << BufferAlignment.getQuantity();
1701 OS << "_" << int(Layout.getSummaryByte());
1702 OS << "_" << int(Layout.getNumArgsByte());
1703 for (const auto &Item : Layout.Items)
1704 OS << "_" << int(Item.getSizeByte()) << "_"
1705 << int(Item.getDescriptorByte());
1706 }
1707
1708 if (llvm::Function *F = CGM.getModule().getFunction(Name))
1709 return F;
1710
1711 llvm::SmallVector<QualType, 4> ArgTys;
1712 FunctionArgList Args;
1713 Args.push_back(ImplicitParamDecl::Create(
1714 Ctx, nullptr, SourceLocation(), &Ctx.Idents.get("buffer"), Ctx.VoidPtrTy,
1715 ImplicitParamDecl::Other));
1716 ArgTys.emplace_back(Ctx.VoidPtrTy);
1717
1718 for (unsigned int I = 0, E = Layout.Items.size(); I < E; ++I) {
1719 char Size = Layout.Items[I].getSizeByte();
1720 if (!Size)
1721 continue;
1722
1723 QualType ArgTy = getOSLogArgType(Ctx, Size);
1724 Args.push_back(ImplicitParamDecl::Create(
1725 Ctx, nullptr, SourceLocation(),
1726 &Ctx.Idents.get(std::string("arg") + llvm::to_string(I)), ArgTy,
1727 ImplicitParamDecl::Other));
1728 ArgTys.emplace_back(ArgTy);
1729 }
1730
1731 QualType ReturnTy = Ctx.VoidTy;
1732
1733 // The helper function has linkonce_odr linkage to enable the linker to merge
1734 // identical functions. To ensure the merging always happens, 'noinline' is
1735 // attached to the function when compiling with -Oz.
1736 const CGFunctionInfo &FI =
1737 CGM.getTypes().arrangeBuiltinFunctionDeclaration(ReturnTy, Args);
1738 llvm::FunctionType *FuncTy = CGM.getTypes().GetFunctionType(FI);
1739 llvm::Function *Fn = llvm::Function::Create(
1740 FuncTy, llvm::GlobalValue::LinkOnceODRLinkage, Name, &CGM.getModule());
1741 Fn->setVisibility(llvm::GlobalValue::HiddenVisibility);
1742 CGM.SetLLVMFunctionAttributes(GlobalDecl(), FI, Fn, /*IsThunk=*/false);
1743 CGM.SetLLVMFunctionAttributesForDefinition(nullptr, Fn);
1744 Fn->setDoesNotThrow();
1745
1746 // Attach 'noinline' at -Oz.
1747 if (CGM.getCodeGenOpts().OptimizeSize == 2)
1748 Fn->addFnAttr(llvm::Attribute::NoInline);
1749
1750 auto NL = ApplyDebugLocation::CreateEmpty(*this);
1751 StartFunction(GlobalDecl(), ReturnTy, Fn, FI, Args);
1752
1753 // Create a scope with an artificial location for the body of this function.
1754 auto AL = ApplyDebugLocation::CreateArtificial(*this);
1755
1756 CharUnits Offset;
1757 Address BufAddr(Builder.CreateLoad(GetAddrOfLocalVar(Args[0]), "buf"),
1758 BufferAlignment);
1759 Builder.CreateStore(Builder.getInt8(Layout.getSummaryByte()),
1760 Builder.CreateConstByteGEP(BufAddr, Offset++, "summary"));
1761 Builder.CreateStore(Builder.getInt8(Layout.getNumArgsByte()),
1762 Builder.CreateConstByteGEP(BufAddr, Offset++, "numArgs"));
1763
1764 unsigned I = 1;
1765 for (const auto &Item : Layout.Items) {
1766 Builder.CreateStore(
1767 Builder.getInt8(Item.getDescriptorByte()),
1768 Builder.CreateConstByteGEP(BufAddr, Offset++, "argDescriptor"));
1769 Builder.CreateStore(
1770 Builder.getInt8(Item.getSizeByte()),
1771 Builder.CreateConstByteGEP(BufAddr, Offset++, "argSize"));
1772
1773 CharUnits Size = Item.size();
1774 if (!Size.getQuantity())
1775 continue;
1776
1777 Address Arg = GetAddrOfLocalVar(Args[I]);
1778 Address Addr = Builder.CreateConstByteGEP(BufAddr, Offset, "argData");
1779 Addr = Builder.CreateBitCast(Addr, Arg.getPointer()->getType(),
1780 "argDataCast");
1781 Builder.CreateStore(Builder.CreateLoad(Arg), Addr);
1782 Offset += Size;
1783 ++I;
1784 }
1785
1786 FinishFunction();
1787
1788 return Fn;
1789}
1790
1791RValue CodeGenFunction::emitBuiltinOSLogFormat(const CallExpr &E) {
1792 assert(E.getNumArgs() >= 2 &&((void)0)
1793 "__builtin_os_log_format takes at least 2 arguments")((void)0);
1794 ASTContext &Ctx = getContext();
1795 analyze_os_log::OSLogBufferLayout Layout;
1796 analyze_os_log::computeOSLogBufferLayout(Ctx, &E, Layout);
1797 Address BufAddr = EmitPointerWithAlignment(E.getArg(0));
1798 llvm::SmallVector<llvm::Value *, 4> RetainableOperands;
1799
1800 // Ignore argument 1, the format string. It is not currently used.
1801 CallArgList Args;
1802 Args.add(RValue::get(BufAddr.getPointer()), Ctx.VoidPtrTy);
1803
1804 for (const auto &Item : Layout.Items) {
1805 int Size = Item.getSizeByte();
1806 if (!Size)
1807 continue;
1808
1809 llvm::Value *ArgVal;
1810
1811 if (Item.getKind() == analyze_os_log::OSLogBufferItem::MaskKind) {
1812 uint64_t Val = 0;
1813 for (unsigned I = 0, E = Item.getMaskType().size(); I < E; ++I)
1814 Val |= ((uint64_t)Item.getMaskType()[I]) << I * 8;
1815 ArgVal = llvm::Constant::getIntegerValue(Int64Ty, llvm::APInt(64, Val));
1816 } else if (const Expr *TheExpr = Item.getExpr()) {
1817 ArgVal = EmitScalarExpr(TheExpr, /*Ignore*/ false);
1818
1819 // If a temporary object that requires destruction after the full
1820 // expression is passed, push a lifetime-extended cleanup to extend its
1821 // lifetime to the end of the enclosing block scope.
1822 auto LifetimeExtendObject = [&](const Expr *E) {
1823 E = E->IgnoreParenCasts();
1824 // Extend lifetimes of objects returned by function calls and message
1825 // sends.
1826
1827 // FIXME: We should do this in other cases in which temporaries are
1828 // created including arguments of non-ARC types (e.g., C++
1829 // temporaries).
1830 if (isa<CallExpr>(E) || isa<ObjCMessageExpr>(E))
1831 return true;
1832 return false;
1833 };
1834
1835 if (TheExpr->getType()->isObjCRetainableType() &&
1836 getLangOpts().ObjCAutoRefCount && LifetimeExtendObject(TheExpr)) {
1837 assert(getEvaluationKind(TheExpr->getType()) == TEK_Scalar &&((void)0)
1838 "Only scalar can be a ObjC retainable type")((void)0);
1839 if (!isa<Constant>(ArgVal)) {
1840 CleanupKind Cleanup = getARCCleanupKind();
1841 QualType Ty = TheExpr->getType();
1842 Address Alloca = Address::invalid();
1843 Address Addr = CreateMemTemp(Ty, "os.log.arg", &Alloca);
1844 ArgVal = EmitARCRetain(Ty, ArgVal);
1845 Builder.CreateStore(ArgVal, Addr);
1846 pushLifetimeExtendedDestroy(Cleanup, Alloca, Ty,
1847 CodeGenFunction::destroyARCStrongPrecise,
1848 Cleanup & EHCleanup);
1849
1850 // Push a clang.arc.use call to ensure ARC optimizer knows that the
1851 // argument has to be alive.
1852 if (CGM.getCodeGenOpts().OptimizationLevel != 0)
1853 pushCleanupAfterFullExpr<CallObjCArcUse>(Cleanup, ArgVal);
1854 }
1855 }
1856 } else {
1857 ArgVal = Builder.getInt32(Item.getConstValue().getQuantity());
1858 }
1859
1860 unsigned ArgValSize =
1861 CGM.getDataLayout().getTypeSizeInBits(ArgVal->getType());
1862 llvm::IntegerType *IntTy = llvm::Type::getIntNTy(getLLVMContext(),
1863 ArgValSize);
1864 ArgVal = Builder.CreateBitOrPointerCast(ArgVal, IntTy);
1865 CanQualType ArgTy = getOSLogArgType(Ctx, Size);
1866 // If ArgVal has type x86_fp80, zero-extend ArgVal.
1867 ArgVal = Builder.CreateZExtOrBitCast(ArgVal, ConvertType(ArgTy));
1868 Args.add(RValue::get(ArgVal), ArgTy);
1869 }
1870
1871 const CGFunctionInfo &FI =
1872 CGM.getTypes().arrangeBuiltinFunctionCall(Ctx.VoidTy, Args);
1873 llvm::Function *F = CodeGenFunction(CGM).generateBuiltinOSLogHelperFunction(
1874 Layout, BufAddr.getAlignment());
1875 EmitCall(FI, CGCallee::forDirect(F), ReturnValueSlot(), Args);
1876 return RValue::get(BufAddr.getPointer());
1877}
1878
1879static bool isSpecialUnsignedMultiplySignedResult(
1880 unsigned BuiltinID, WidthAndSignedness Op1Info, WidthAndSignedness Op2Info,
1881 WidthAndSignedness ResultInfo) {
1882 return BuiltinID == Builtin::BI__builtin_mul_overflow &&
1883 Op1Info.Width == Op2Info.Width && Op2Info.Width == ResultInfo.Width &&
1884 !Op1Info.Signed && !Op2Info.Signed && ResultInfo.Signed;
1885}
1886
1887static RValue EmitCheckedUnsignedMultiplySignedResult(
1888 CodeGenFunction &CGF, const clang::Expr *Op1, WidthAndSignedness Op1Info,
1889 const clang::Expr *Op2, WidthAndSignedness Op2Info,
1890 const clang::Expr *ResultArg, QualType ResultQTy,
1891 WidthAndSignedness ResultInfo) {
1892 assert(isSpecialUnsignedMultiplySignedResult(((void)0)
1893 Builtin::BI__builtin_mul_overflow, Op1Info, Op2Info, ResultInfo) &&((void)0)
1894 "Cannot specialize this multiply")((void)0);
1895
1896 llvm::Value *V1 = CGF.EmitScalarExpr(Op1);
1897 llvm::Value *V2 = CGF.EmitScalarExpr(Op2);
1898
1899 llvm::Value *HasOverflow;
1900 llvm::Value *Result = EmitOverflowIntrinsic(
1901 CGF, llvm::Intrinsic::umul_with_overflow, V1, V2, HasOverflow);
1902
1903 // The intrinsic call will detect overflow when the value is > UINT_MAX,
1904 // however, since the original builtin had a signed result, we need to report
1905 // an overflow when the result is greater than INT_MAX.
1906 auto IntMax = llvm::APInt::getSignedMaxValue(ResultInfo.Width);
1907 llvm::Value *IntMaxValue = llvm::ConstantInt::get(Result->getType(), IntMax);
1908
1909 llvm::Value *IntMaxOverflow = CGF.Builder.CreateICmpUGT(Result, IntMaxValue);
1910 HasOverflow = CGF.Builder.CreateOr(HasOverflow, IntMaxOverflow);
1911
1912 bool isVolatile =
1913 ResultArg->getType()->getPointeeType().isVolatileQualified();
1914 Address ResultPtr = CGF.EmitPointerWithAlignment(ResultArg);
1915 CGF.Builder.CreateStore(CGF.EmitToMemory(Result, ResultQTy), ResultPtr,
1916 isVolatile);
1917 return RValue::get(HasOverflow);
1918}
1919
1920/// Determine if a binop is a checked mixed-sign multiply we can specialize.
1921static bool isSpecialMixedSignMultiply(unsigned BuiltinID,
1922 WidthAndSignedness Op1Info,
1923 WidthAndSignedness Op2Info,
1924 WidthAndSignedness ResultInfo) {
1925 return BuiltinID == Builtin::BI__builtin_mul_overflow &&
1926 std::max(Op1Info.Width, Op2Info.Width) >= ResultInfo.Width &&
1927 Op1Info.Signed != Op2Info.Signed;
1928}
1929
1930/// Emit a checked mixed-sign multiply. This is a cheaper specialization of
1931/// the generic checked-binop irgen.
1932static RValue
1933EmitCheckedMixedSignMultiply(CodeGenFunction &CGF, const clang::Expr *Op1,
1934 WidthAndSignedness Op1Info, const clang::Expr *Op2,
1935 WidthAndSignedness Op2Info,
1936 const clang::Expr *ResultArg, QualType ResultQTy,
1937 WidthAndSignedness ResultInfo) {
1938 assert(isSpecialMixedSignMultiply(Builtin::BI__builtin_mul_overflow, Op1Info,((void)0)
1939 Op2Info, ResultInfo) &&((void)0)
1940 "Not a mixed-sign multipliction we can specialize")((void)0);
1941
1942 // Emit the signed and unsigned operands.
1943 const clang::Expr *SignedOp = Op1Info.Signed ? Op1 : Op2;
1944 const clang::Expr *UnsignedOp = Op1Info.Signed ? Op2 : Op1;
1945 llvm::Value *Signed = CGF.EmitScalarExpr(SignedOp);
1946 llvm::Value *Unsigned = CGF.EmitScalarExpr(UnsignedOp);
1947 unsigned SignedOpWidth = Op1Info.Signed ? Op1Info.Width : Op2Info.Width;
1948 unsigned UnsignedOpWidth = Op1Info.Signed ? Op2Info.Width : Op1Info.Width;
1949
1950 // One of the operands may be smaller than the other. If so, [s|z]ext it.
1951 if (SignedOpWidth < UnsignedOpWidth)
1952 Signed = CGF.Builder.CreateSExt(Signed, Unsigned->getType(), "op.sext");
1953 if (UnsignedOpWidth < SignedOpWidth)
1954 Unsigned = CGF.Builder.CreateZExt(Unsigned, Signed->getType(), "op.zext");
1955
1956 llvm::Type *OpTy = Signed->getType();
1957 llvm::Value *Zero = llvm::Constant::getNullValue(OpTy);
1958 Address ResultPtr = CGF.EmitPointerWithAlignment(ResultArg);
1959 llvm::Type *ResTy = ResultPtr.getElementType();
1960 unsigned OpWidth = std::max(Op1Info.Width, Op2Info.Width);
1961
1962 // Take the absolute value of the signed operand.
1963 llvm::Value *IsNegative = CGF.Builder.CreateICmpSLT(Signed, Zero);
1964 llvm::Value *AbsOfNegative = CGF.Builder.CreateSub(Zero, Signed);
1965 llvm::Value *AbsSigned =
1966 CGF.Builder.CreateSelect(IsNegative, AbsOfNegative, Signed);
1967
1968 // Perform a checked unsigned multiplication.
1969 llvm::Value *UnsignedOverflow;
1970 llvm::Value *UnsignedResult =
1971 EmitOverflowIntrinsic(CGF, llvm::Intrinsic::umul_with_overflow, AbsSigned,
1972 Unsigned, UnsignedOverflow);
1973
1974 llvm::Value *Overflow, *Result;
1975 if (ResultInfo.Signed) {
1976 // Signed overflow occurs if the result is greater than INT_MAX or lesser
1977 // than INT_MIN, i.e when |Result| > (INT_MAX + IsNegative).
1978 auto IntMax =
1979 llvm::APInt::getSignedMaxValue(ResultInfo.Width).zextOrSelf(OpWidth);
1980 llvm::Value *MaxResult =
1981 CGF.Builder.CreateAdd(llvm::ConstantInt::get(OpTy, IntMax),
1982 CGF.Builder.CreateZExt(IsNegative, OpTy));
1983 llvm::Value *SignedOverflow =
1984 CGF.Builder.CreateICmpUGT(UnsignedResult, MaxResult);
1985 Overflow = CGF.Builder.CreateOr(UnsignedOverflow, SignedOverflow);
1986
1987 // Prepare the signed result (possibly by negating it).
1988 llvm::Value *NegativeResult = CGF.Builder.CreateNeg(UnsignedResult);
1989 llvm::Value *SignedResult =
1990 CGF.Builder.CreateSelect(IsNegative, NegativeResult, UnsignedResult);
1991 Result = CGF.Builder.CreateTrunc(SignedResult, ResTy);
1992 } else {
1993 // Unsigned overflow occurs if the result is < 0 or greater than UINT_MAX.
1994 llvm::Value *Underflow = CGF.Builder.CreateAnd(
1995 IsNegative, CGF.Builder.CreateIsNotNull(UnsignedResult));
1996 Overflow = CGF.Builder.CreateOr(UnsignedOverflow, Underflow);
1997 if (ResultInfo.Width < OpWidth) {
1998 auto IntMax =
1999 llvm::APInt::getMaxValue(ResultInfo.Width).zext(OpWidth);
2000 llvm::Value *TruncOverflow = CGF.Builder.CreateICmpUGT(
2001 UnsignedResult, llvm::ConstantInt::get(OpTy, IntMax));
2002 Overflow = CGF.Builder.CreateOr(Overflow, TruncOverflow);
2003 }
2004
2005 // Negate the product if it would be negative in infinite precision.
2006 Result = CGF.Builder.CreateSelect(
2007 IsNegative, CGF.Builder.CreateNeg(UnsignedResult), UnsignedResult);
2008
2009 Result = CGF.Builder.CreateTrunc(Result, ResTy);
2010 }
2011 assert(Overflow && Result && "Missing overflow or result")((void)0);
2012
2013 bool isVolatile =
2014 ResultArg->getType()->getPointeeType().isVolatileQualified();
2015 CGF.Builder.CreateStore(CGF.EmitToMemory(Result, ResultQTy), ResultPtr,
2016 isVolatile);
2017 return RValue::get(Overflow);
2018}
2019
2020static llvm::Value *dumpRecord(CodeGenFunction &CGF, QualType RType,
2021 Value *&RecordPtr, CharUnits Align,
2022 llvm::FunctionCallee Func, int Lvl) {
2023 ASTContext &Context = CGF.getContext();
2024 RecordDecl *RD = RType->castAs<RecordType>()->getDecl()->getDefinition();
2025 std::string Pad = std::string(Lvl * 4, ' ');
2026
2027 Value *GString =
2028 CGF.Builder.CreateGlobalStringPtr(RType.getAsString() + " {\n");
2029 Value *Res = CGF.Builder.CreateCall(Func, {GString});
2030
2031 static llvm::DenseMap<QualType, const char *> Types;
2032 if (Types.empty()) {
2033 Types[Context.CharTy] = "%c";
2034 Types[Context.BoolTy] = "%d";
2035 Types[Context.SignedCharTy] = "%hhd";
2036 Types[Context.UnsignedCharTy] = "%hhu";
2037 Types[Context.IntTy] = "%d";
2038 Types[Context.UnsignedIntTy] = "%u";
2039 Types[Context.LongTy] = "%ld";
2040 Types[Context.UnsignedLongTy] = "%lu";
2041 Types[Context.LongLongTy] = "%lld";
2042 Types[Context.UnsignedLongLongTy] = "%llu";
2043 Types[Context.ShortTy] = "%hd";
2044 Types[Context.UnsignedShortTy] = "%hu";
2045 Types[Context.VoidPtrTy] = "%p";
2046 Types[Context.FloatTy] = "%f";
2047 Types[Context.DoubleTy] = "%f";
2048 Types[Context.LongDoubleTy] = "%Lf";
2049 Types[Context.getPointerType(Context.CharTy)] = "%s";
2050 Types[Context.getPointerType(Context.getConstType(Context.CharTy))] = "%s";
2051 }
2052
2053 for (const auto *FD : RD->fields()) {
2054 Value *FieldPtr = RecordPtr;
2055 if (RD->isUnion())
2056 FieldPtr = CGF.Builder.CreatePointerCast(
2057 FieldPtr, CGF.ConvertType(Context.getPointerType(FD->getType())));
2058 else
2059 FieldPtr = CGF.Builder.CreateStructGEP(CGF.ConvertType(RType), FieldPtr,
2060 FD->getFieldIndex());
2061
2062 GString = CGF.Builder.CreateGlobalStringPtr(
2063 llvm::Twine(Pad)
2064 .concat(FD->getType().getAsString())
2065 .concat(llvm::Twine(' '))
2066 .concat(FD->getNameAsString())
2067 .concat(" : ")
2068 .str());
2069 Value *TmpRes = CGF.Builder.CreateCall(Func, {GString});
2070 Res = CGF.Builder.CreateAdd(Res, TmpRes);
2071
2072 QualType CanonicalType =
2073 FD->getType().getUnqualifiedType().getCanonicalType();
2074
2075 // We check whether we are in a recursive type
2076 if (CanonicalType->isRecordType()) {
2077 TmpRes = dumpRecord(CGF, CanonicalType, FieldPtr, Align, Func, Lvl + 1);
2078 Res = CGF.Builder.CreateAdd(TmpRes, Res);
2079 continue;
2080 }
2081
2082 // We try to determine the best format to print the current field
2083 llvm::Twine Format = Types.find(CanonicalType) == Types.end()
2084 ? Types[Context.VoidPtrTy]
2085 : Types[CanonicalType];
2086
2087 Address FieldAddress = Address(FieldPtr, Align);
2088 FieldPtr = CGF.Builder.CreateLoad(FieldAddress);
2089
2090 // FIXME Need to handle bitfield here
2091 GString = CGF.Builder.CreateGlobalStringPtr(
2092 Format.concat(llvm::Twine('\n')).str());
2093 TmpRes = CGF.Builder.CreateCall(Func, {GString, FieldPtr});
2094 Res = CGF.Builder.CreateAdd(Res, TmpRes);
2095 }
2096
2097 GString = CGF.Builder.CreateGlobalStringPtr(Pad + "}\n");
2098 Value *TmpRes = CGF.Builder.CreateCall(Func, {GString});
2099 Res = CGF.Builder.CreateAdd(Res, TmpRes);
2100 return Res;
2101}
2102
2103static bool
2104TypeRequiresBuiltinLaunderImp(const ASTContext &Ctx, QualType Ty,
2105 llvm::SmallPtrSetImpl<const Decl *> &Seen) {
2106 if (const auto *Arr = Ctx.getAsArrayType(Ty))
2107 Ty = Ctx.getBaseElementType(Arr);
2108
2109 const auto *Record = Ty->getAsCXXRecordDecl();
2110 if (!Record)
2111 return false;
2112
2113 // We've already checked this type, or are in the process of checking it.
2114 if (!Seen.insert(Record).second)
2115 return false;
2116
2117 assert(Record->hasDefinition() &&((void)0)
2118 "Incomplete types should already be diagnosed")((void)0);
2119
2120 if (Record->isDynamicClass())
2121 return true;
2122
2123 for (FieldDecl *F : Record->fields()) {
2124 if (TypeRequiresBuiltinLaunderImp(Ctx, F->getType(), Seen))
2125 return true;
2126 }
2127 return false;
2128}
2129
2130/// Determine if the specified type requires laundering by checking if it is a
2131/// dynamic class type or contains a subobject which is a dynamic class type.
2132static bool TypeRequiresBuiltinLaunder(CodeGenModule &CGM, QualType Ty) {
2133 if (!CGM.getCodeGenOpts().StrictVTablePointers)
2134 return false;
2135 llvm::SmallPtrSet<const Decl *, 16> Seen;
2136 return TypeRequiresBuiltinLaunderImp(CGM.getContext(), Ty, Seen);
2137}
2138
2139RValue CodeGenFunction::emitRotate(const CallExpr *E, bool IsRotateRight) {
2140 llvm::Value *Src = EmitScalarExpr(E->getArg(0));
2141 llvm::Value *ShiftAmt = EmitScalarExpr(E->getArg(1));
2142
2143 // The builtin's shift arg may have a different type than the source arg and
2144 // result, but the LLVM intrinsic uses the same type for all values.
2145 llvm::Type *Ty = Src->getType();
2146 ShiftAmt = Builder.CreateIntCast(ShiftAmt, Ty, false);
2147
2148 // Rotate is a special case of LLVM funnel shift - 1st 2 args are the same.
2149 unsigned IID = IsRotateRight ? Intrinsic::fshr : Intrinsic::fshl;
2150 Function *F = CGM.getIntrinsic(IID, Ty);
2151 return RValue::get(Builder.CreateCall(F, { Src, Src, ShiftAmt }));
2152}
2153
2154// Map math builtins for long-double to f128 version.
2155static unsigned mutateLongDoubleBuiltin(unsigned BuiltinID) {
2156 switch (BuiltinID) {
2157#define MUTATE_LDBL(func) \
2158 case Builtin::BI__builtin_##func##l: \
2159 return Builtin::BI__builtin_##func##f128;
2160 MUTATE_LDBL(sqrt)
2161 MUTATE_LDBL(cbrt)
2162 MUTATE_LDBL(fabs)
2163 MUTATE_LDBL(log)
2164 MUTATE_LDBL(log2)
2165 MUTATE_LDBL(log10)
2166 MUTATE_LDBL(log1p)
2167 MUTATE_LDBL(logb)
2168 MUTATE_LDBL(exp)
2169 MUTATE_LDBL(exp2)
2170 MUTATE_LDBL(expm1)
2171 MUTATE_LDBL(fdim)
2172 MUTATE_LDBL(hypot)
2173 MUTATE_LDBL(ilogb)
2174 MUTATE_LDBL(pow)
2175 MUTATE_LDBL(fmin)
2176 MUTATE_LDBL(fmax)
2177 MUTATE_LDBL(ceil)
2178 MUTATE_LDBL(trunc)
2179 MUTATE_LDBL(rint)
2180 MUTATE_LDBL(nearbyint)
2181 MUTATE_LDBL(round)
2182 MUTATE_LDBL(floor)
2183 MUTATE_LDBL(lround)
2184 MUTATE_LDBL(llround)
2185 MUTATE_LDBL(lrint)
2186 MUTATE_LDBL(llrint)
2187 MUTATE_LDBL(fmod)
2188 MUTATE_LDBL(modf)
2189 MUTATE_LDBL(nan)
2190 MUTATE_LDBL(nans)
2191 MUTATE_LDBL(inf)
2192 MUTATE_LDBL(fma)
2193 MUTATE_LDBL(sin)
2194 MUTATE_LDBL(cos)
2195 MUTATE_LDBL(tan)
2196 MUTATE_LDBL(sinh)
2197 MUTATE_LDBL(cosh)
2198 MUTATE_LDBL(tanh)
2199 MUTATE_LDBL(asin)
2200 MUTATE_LDBL(acos)
2201 MUTATE_LDBL(atan)
2202 MUTATE_LDBL(asinh)
2203 MUTATE_LDBL(acosh)
2204 MUTATE_LDBL(atanh)
2205 MUTATE_LDBL(atan2)
2206 MUTATE_LDBL(erf)
2207 MUTATE_LDBL(erfc)
2208 MUTATE_LDBL(ldexp)
2209 MUTATE_LDBL(frexp)
2210 MUTATE_LDBL(huge_val)
2211 MUTATE_LDBL(copysign)
2212 MUTATE_LDBL(nextafter)
2213 MUTATE_LDBL(nexttoward)
2214 MUTATE_LDBL(remainder)
2215 MUTATE_LDBL(remquo)
2216 MUTATE_LDBL(scalbln)
2217 MUTATE_LDBL(scalbn)
2218 MUTATE_LDBL(tgamma)
2219 MUTATE_LDBL(lgamma)
2220#undef MUTATE_LDBL
2221 default:
2222 return BuiltinID;
2223 }
2224}
2225
2226RValue CodeGenFunction::EmitBuiltinExpr(const GlobalDecl GD, unsigned BuiltinID,
2227 const CallExpr *E,
2228 ReturnValueSlot ReturnValue) {
2229 const FunctionDecl *FD = GD.getDecl()->getAsFunction();
2230 // See if we can constant fold this builtin. If so, don't emit it at all.
2231 Expr::EvalResult Result;
2232 if (E->EvaluateAsRValue(Result, CGM.getContext()) &&
2233 !Result.hasSideEffects()) {
2234 if (Result.Val.isInt())
2235 return RValue::get(llvm::ConstantInt::get(getLLVMContext(),
2236 Result.Val.getInt()));
2237 if (Result.Val.isFloat())
2238 return RValue::get(llvm::ConstantFP::get(getLLVMContext(),
2239 Result.Val.getFloat()));
2240 }
2241
2242 // If current long-double semantics is IEEE 128-bit, replace math builtins
2243 // of long-double with f128 equivalent.
2244 // TODO: This mutation should also be applied to other targets other than PPC,
2245 // after backend supports IEEE 128-bit style libcalls.
2246 if (getTarget().getTriple().isPPC64() &&
2247 &getTarget().getLongDoubleFormat() == &llvm::APFloat::IEEEquad())
2248 BuiltinID = mutateLongDoubleBuiltin(BuiltinID);
2249
2250 // If the builtin has been declared explicitly with an assembler label,
2251 // disable the specialized emitting below. Ideally we should communicate the
2252 // rename in IR, or at least avoid generating the intrinsic calls that are
2253 // likely to get lowered to the renamed library functions.
2254 const unsigned BuiltinIDIfNoAsmLabel =
2255 FD->hasAttr<AsmLabelAttr>() ? 0 : BuiltinID;
2256
2257 // There are LLVM math intrinsics/instructions corresponding to math library
2258 // functions except the LLVM op will never set errno while the math library
2259 // might. Also, math builtins have the same semantics as their math library
2260 // twins. Thus, we can transform math library and builtin calls to their
2261 // LLVM counterparts if the call is marked 'const' (known to never set errno).
2262 if (FD->hasAttr<ConstAttr>()) {
2263 switch (BuiltinIDIfNoAsmLabel) {
2264 case Builtin::BIceil:
2265 case Builtin::BIceilf:
2266 case Builtin::BIceill:
2267 case Builtin::BI__builtin_ceil:
2268 case Builtin::BI__builtin_ceilf:
2269 case Builtin::BI__builtin_ceilf16:
2270 case Builtin::BI__builtin_ceill:
2271 case Builtin::BI__builtin_ceilf128:
2272 return RValue::get(emitUnaryMaybeConstrainedFPBuiltin(*this, E,
2273 Intrinsic::ceil,
2274 Intrinsic::experimental_constrained_ceil));
2275
2276 case Builtin::BIcopysign:
2277 case Builtin::BIcopysignf:
2278 case Builtin::BIcopysignl:
2279 case Builtin::BI__builtin_copysign:
2280 case Builtin::BI__builtin_copysignf:
2281 case Builtin::BI__builtin_copysignf16:
2282 case Builtin::BI__builtin_copysignl:
2283 case Builtin::BI__builtin_copysignf128:
2284 return RValue::get(emitBinaryBuiltin(*this, E, Intrinsic::copysign));
2285
2286 case Builtin::BIcos:
2287 case Builtin::BIcosf:
2288 case Builtin::BIcosl:
2289 case Builtin::BI__builtin_cos:
2290 case Builtin::BI__builtin_cosf:
2291 case Builtin::BI__builtin_cosf16:
2292 case Builtin::BI__builtin_cosl:
2293 case Builtin::BI__builtin_cosf128:
2294 return RValue::get(emitUnaryMaybeConstrainedFPBuiltin(*this, E,
2295 Intrinsic::cos,
2296 Intrinsic::experimental_constrained_cos));
2297
2298 case Builtin::BIexp:
2299 case Builtin::BIexpf:
2300 case Builtin::BIexpl:
2301 case Builtin::BI__builtin_exp:
2302 case Builtin::BI__builtin_expf:
2303 case Builtin::BI__builtin_expf16:
2304 case Builtin::BI__builtin_expl:
2305 case Builtin::BI__builtin_expf128:
2306 return RValue::get(emitUnaryMaybeConstrainedFPBuiltin(*this, E,
2307 Intrinsic::exp,
2308 Intrinsic::experimental_constrained_exp));
2309
2310 case Builtin::BIexp2:
2311 case Builtin::BIexp2f:
2312 case Builtin::BIexp2l:
2313 case Builtin::BI__builtin_exp2:
2314 case Builtin::BI__builtin_exp2f:
2315 case Builtin::BI__builtin_exp2f16:
2316 case Builtin::BI__builtin_exp2l:
2317 case Builtin::BI__builtin_exp2f128:
2318 return RValue::get(emitUnaryMaybeConstrainedFPBuiltin(*this, E,
2319 Intrinsic::exp2,
2320 Intrinsic::experimental_constrained_exp2));
2321
2322 case Builtin::BIfabs:
2323 case Builtin::BIfabsf:
2324 case Builtin::BIfabsl:
2325 case Builtin::BI__builtin_fabs:
2326 case Builtin::BI__builtin_fabsf:
2327 case Builtin::BI__builtin_fabsf16:
2328 case Builtin::BI__builtin_fabsl:
2329 case Builtin::BI__builtin_fabsf128:
2330 return RValue::get(emitUnaryBuiltin(*this, E, Intrinsic::fabs));
2331
2332 case Builtin::BIfloor:
2333 case Builtin::BIfloorf:
2334 case Builtin::BIfloorl:
2335 case Builtin::BI__builtin_floor:
2336 case Builtin::BI__builtin_floorf:
2337 case Builtin::BI__builtin_floorf16:
2338 case Builtin::BI__builtin_floorl:
2339 case Builtin::BI__builtin_floorf128:
2340 return RValue::get(emitUnaryMaybeConstrainedFPBuiltin(*this, E,
2341 Intrinsic::floor,
2342 Intrinsic::experimental_constrained_floor));
2343
2344 case Builtin::BIfma:
2345 case Builtin::BIfmaf:
2346 case Builtin::BIfmal:
2347 case Builtin::BI__builtin_fma:
2348 case Builtin::BI__builtin_fmaf:
2349 case Builtin::BI__builtin_fmaf16:
2350 case Builtin::BI__builtin_fmal:
2351 case Builtin::BI__builtin_fmaf128:
2352 return RValue::get(emitTernaryMaybeConstrainedFPBuiltin(*this, E,
2353 Intrinsic::fma,
2354 Intrinsic::experimental_constrained_fma));
2355
2356 case Builtin::BIfmax:
2357 case Builtin::BIfmaxf:
2358 case Builtin::BIfmaxl:
2359 case Builtin::BI__builtin_fmax:
2360 case Builtin::BI__builtin_fmaxf:
2361 case Builtin::BI__builtin_fmaxf16:
2362 case Builtin::BI__builtin_fmaxl:
2363 case Builtin::BI__builtin_fmaxf128:
2364 return RValue::get(emitBinaryMaybeConstrainedFPBuiltin(*this, E,
2365 Intrinsic::maxnum,
2366 Intrinsic::experimental_constrained_maxnum));
2367
2368 case Builtin::BIfmin:
2369 case Builtin::BIfminf:
2370 case Builtin::BIfminl:
2371 case Builtin::BI__builtin_fmin:
2372 case Builtin::BI__builtin_fminf:
2373 case Builtin::BI__builtin_fminf16:
2374 case Builtin::BI__builtin_fminl:
2375 case Builtin::BI__builtin_fminf128:
2376 return RValue::get(emitBinaryMaybeConstrainedFPBuiltin(*this, E,
2377 Intrinsic::minnum,
2378 Intrinsic::experimental_constrained_minnum));
2379
2380 // fmod() is a special-case. It maps to the frem instruction rather than an
2381 // LLVM intrinsic.
2382 case Builtin::BIfmod:
2383 case Builtin::BIfmodf:
2384 case Builtin::BIfmodl:
2385 case Builtin::BI__builtin_fmod:
2386 case Builtin::BI__builtin_fmodf:
2387 case Builtin::BI__builtin_fmodf16:
2388 case Builtin::BI__builtin_fmodl:
2389 case Builtin::BI__builtin_fmodf128: {
2390 CodeGenFunction::CGFPOptionsRAII FPOptsRAII(*this, E);
2391 Value *Arg1 = EmitScalarExpr(E->getArg(0));
2392 Value *Arg2 = EmitScalarExpr(E->getArg(1));
2393 return RValue::get(Builder.CreateFRem(Arg1, Arg2, "fmod"));
2394 }
2395
2396 case Builtin::BIlog:
2397 case Builtin::BIlogf:
2398 case Builtin::BIlogl:
2399 case Builtin::BI__builtin_log:
2400 case Builtin::BI__builtin_logf:
2401 case Builtin::BI__builtin_logf16:
2402 case Builtin::BI__builtin_logl:
2403 case Builtin::BI__builtin_logf128:
2404 return RValue::get(emitUnaryMaybeConstrainedFPBuiltin(*this, E,
2405 Intrinsic::log,
2406 Intrinsic::experimental_constrained_log));
2407
2408 case Builtin::BIlog10:
2409 case Builtin::BIlog10f:
2410 case Builtin::BIlog10l:
2411 case Builtin::BI__builtin_log10:
2412 case Builtin::BI__builtin_log10f:
2413 case Builtin::BI__builtin_log10f16:
2414 case Builtin::BI__builtin_log10l:
2415 case Builtin::BI__builtin_log10f128:
2416 return RValue::get(emitUnaryMaybeConstrainedFPBuiltin(*this, E,
2417 Intrinsic::log10,
2418 Intrinsic::experimental_constrained_log10));
2419
2420 case Builtin::BIlog2:
2421 case Builtin::BIlog2f:
2422 case Builtin::BIlog2l:
2423 case Builtin::BI__builtin_log2:
2424 case Builtin::BI__builtin_log2f:
2425 case Builtin::BI__builtin_log2f16:
2426 case Builtin::BI__builtin_log2l:
2427 case Builtin::BI__builtin_log2f128:
2428 return RValue::get(emitUnaryMaybeConstrainedFPBuiltin(*this, E,
2429 Intrinsic::log2,
2430 Intrinsic::experimental_constrained_log2));
2431
2432 case Builtin::BInearbyint:
2433 case Builtin::BInearbyintf:
2434 case Builtin::BInearbyintl:
2435 case Builtin::BI__builtin_nearbyint:
2436 case Builtin::BI__builtin_nearbyintf:
2437 case Builtin::BI__builtin_nearbyintl:
2438 case Builtin::BI__builtin_nearbyintf128:
2439 return RValue::get(emitUnaryMaybeConstrainedFPBuiltin(*this, E,
2440 Intrinsic::nearbyint,
2441 Intrinsic::experimental_constrained_nearbyint));
2442
2443 case Builtin::BIpow:
2444 case Builtin::BIpowf:
2445 case Builtin::BIpowl:
2446 case Builtin::BI__builtin_pow:
2447 case Builtin::BI__builtin_powf:
2448 case Builtin::BI__builtin_powf16:
2449 case Builtin::BI__builtin_powl:
2450 case Builtin::BI__builtin_powf128:
2451 return RValue::get(emitBinaryMaybeConstrainedFPBuiltin(*this, E,
2452 Intrinsic::pow,
2453 Intrinsic::experimental_constrained_pow));
2454
2455 case Builtin::BIrint:
2456 case Builtin::BIrintf:
2457 case Builtin::BIrintl:
2458 case Builtin::BI__builtin_rint:
2459 case Builtin::BI__builtin_rintf:
2460 case Builtin::BI__builtin_rintf16:
2461 case Builtin::BI__builtin_rintl:
2462 case Builtin::BI__builtin_rintf128:
2463 return RValue::get(emitUnaryMaybeConstrainedFPBuiltin(*this, E,
2464 Intrinsic::rint,
2465 Intrinsic::experimental_constrained_rint));
2466
2467 case Builtin::BIround:
2468 case Builtin::BIroundf:
2469 case Builtin::BIroundl:
2470 case Builtin::BI__builtin_round:
2471 case Builtin::BI__builtin_roundf:
2472 case Builtin::BI__builtin_roundf16:
2473 case Builtin::BI__builtin_roundl:
2474 case Builtin::BI__builtin_roundf128:
2475 return RValue::get(emitUnaryMaybeConstrainedFPBuiltin(*this, E,
2476 Intrinsic::round,
2477 Intrinsic::experimental_constrained_round));
2478
2479 case Builtin::BIsin:
2480 case Builtin::BIsinf:
2481 case Builtin::BIsinl:
2482 case Builtin::BI__builtin_sin:
2483 case Builtin::BI__builtin_sinf:
2484 case Builtin::BI__builtin_sinf16:
2485 case Builtin::BI__builtin_sinl:
2486 case Builtin::BI__builtin_sinf128:
2487 return RValue::get(emitUnaryMaybeConstrainedFPBuiltin(*this, E,
2488 Intrinsic::sin,
2489 Intrinsic::experimental_constrained_sin));
2490
2491 case Builtin::BIsqrt:
2492 case Builtin::BIsqrtf:
2493 case Builtin::BIsqrtl:
2494 case Builtin::BI__builtin_sqrt:
2495 case Builtin::BI__builtin_sqrtf:
2496 case Builtin::BI__builtin_sqrtf16:
2497 case Builtin::BI__builtin_sqrtl:
2498 case Builtin::BI__builtin_sqrtf128:
2499 return RValue::get(emitUnaryMaybeConstrainedFPBuiltin(*this, E,
2500 Intrinsic::sqrt,
2501 Intrinsic::experimental_constrained_sqrt));
2502
2503 case Builtin::BItrunc:
2504 case Builtin::BItruncf:
2505 case Builtin::BItruncl:
2506 case Builtin::BI__builtin_trunc:
2507 case Builtin::BI__builtin_truncf:
2508 case Builtin::BI__builtin_truncf16:
2509 case Builtin::BI__builtin_truncl:
2510 case Builtin::BI__builtin_truncf128:
2511 return RValue::get(emitUnaryMaybeConstrainedFPBuiltin(*this, E,
2512 Intrinsic::trunc,
2513 Intrinsic::experimental_constrained_trunc));
2514
2515 case Builtin::BIlround:
2516 case Builtin::BIlroundf:
2517 case Builtin::BIlroundl:
2518 case Builtin::BI__builtin_lround:
2519 case Builtin::BI__builtin_lroundf:
2520 case Builtin::BI__builtin_lroundl:
2521 case Builtin::BI__builtin_lroundf128:
2522 return RValue::get(emitMaybeConstrainedFPToIntRoundBuiltin(
2523 *this, E, Intrinsic::lround,
2524 Intrinsic::experimental_constrained_lround));
2525
2526 case Builtin::BIllround:
2527 case Builtin::BIllroundf:
2528 case Builtin::BIllroundl:
2529 case Builtin::BI__builtin_llround:
2530 case Builtin::BI__builtin_llroundf:
2531 case Builtin::BI__builtin_llroundl:
2532 case Builtin::BI__builtin_llroundf128:
2533 return RValue::get(emitMaybeConstrainedFPToIntRoundBuiltin(
2534 *this, E, Intrinsic::llround,
2535 Intrinsic::experimental_constrained_llround));
2536
2537 case Builtin::BIlrint:
2538 case Builtin::BIlrintf:
2539 case Builtin::BIlrintl:
2540 case Builtin::BI__builtin_lrint:
2541 case Builtin::BI__builtin_lrintf:
2542 case Builtin::BI__builtin_lrintl:
2543 case Builtin::BI__builtin_lrintf128:
2544 return RValue::get(emitMaybeConstrainedFPToIntRoundBuiltin(
2545 *this, E, Intrinsic::lrint,
2546 Intrinsic::experimental_constrained_lrint));
2547
2548 case Builtin::BIllrint:
2549 case Builtin::BIllrintf:
2550 case Builtin::BIllrintl:
2551 case Builtin::BI__builtin_llrint:
2552 case Builtin::BI__builtin_llrintf:
2553 case Builtin::BI__builtin_llrintl:
2554 case Builtin::BI__builtin_llrintf128:
2555 return RValue::get(emitMaybeConstrainedFPToIntRoundBuiltin(
2556 *this, E, Intrinsic::llrint,
2557 Intrinsic::experimental_constrained_llrint));
2558
2559 default:
2560 break;
2561 }
2562 }
2563
2564 switch (BuiltinIDIfNoAsmLabel) {
2565 default: break;
2566 case Builtin::BI__builtin___CFStringMakeConstantString:
2567 case Builtin::BI__builtin___NSStringMakeConstantString:
2568 return RValue::get(ConstantEmitter(*this).emitAbstract(E, E->getType()));
2569 case Builtin::BI__builtin_stdarg_start:
2570 case Builtin::BI__builtin_va_start:
2571 case Builtin::BI__va_start:
2572 case Builtin::BI__builtin_va_end:
2573 return RValue::get(
2574 EmitVAStartEnd(BuiltinID == Builtin::BI__va_start
2575 ? EmitScalarExpr(E->getArg(0))
2576 : EmitVAListRef(E->getArg(0)).getPointer(),
2577 BuiltinID != Builtin::BI__builtin_va_end));
2578 case Builtin::BI__builtin_va_copy: {
2579 Value *DstPtr = EmitVAListRef(E->getArg(0)).getPointer();
2580 Value *SrcPtr = EmitVAListRef(E->getArg(1)).getPointer();
2581
2582 llvm::Type *Type = Int8PtrTy;
2583
2584 DstPtr = Builder.CreateBitCast(DstPtr, Type);
2585 SrcPtr = Builder.CreateBitCast(SrcPtr, Type);
2586 return RValue::get(Builder.CreateCall(CGM.getIntrinsic(Intrinsic::vacopy),
2587 {DstPtr, SrcPtr}));
2588 }
2589 case Builtin::BI__builtin_abs:
2590 case Builtin::BI__builtin_labs:
2591 case Builtin::BI__builtin_llabs: {
2592 // X < 0 ? -X : X
2593 // The negation has 'nsw' because abs of INT_MIN is undefined.
2594 Value *ArgValue = EmitScalarExpr(E->getArg(0));
2595 Value *NegOp = Builder.CreateNSWNeg(ArgValue, "neg");
2596 Constant *Zero = llvm::Constant::getNullValue(ArgValue->getType());
2597 Value *CmpResult = Builder.CreateICmpSLT(ArgValue, Zero, "abscond");
2598 Value *Result = Builder.CreateSelect(CmpResult, NegOp, ArgValue, "abs");
2599 return RValue::get(Result);
2600 }
2601 case Builtin::BI__builtin_complex: {
2602 Value *Real = EmitScalarExpr(E->getArg(0));
2603 Value *Imag = EmitScalarExpr(E->getArg(1));
2604 return RValue::getComplex({Real, Imag});
2605 }
2606 case Builtin::BI__builtin_conj:
2607 case Builtin::BI__builtin_conjf:
2608 case Builtin::BI__builtin_conjl:
2609 case Builtin::BIconj:
2610 case Builtin::BIconjf:
2611 case Builtin::BIconjl: {
2612 ComplexPairTy ComplexVal = EmitComplexExpr(E->getArg(0));
2613 Value *Real = ComplexVal.first;
2614 Value *Imag = ComplexVal.second;
2615 Imag = Builder.CreateFNeg(Imag, "neg");
2616 return RValue::getComplex(std::make_pair(Real, Imag));
2617 }
2618 case Builtin::BI__builtin_creal:
2619 case Builtin::BI__builtin_crealf:
2620 case Builtin::BI__builtin_creall:
2621 case Builtin::BIcreal:
2622 case Builtin::BIcrealf:
2623 case Builtin::BIcreall: {
2624 ComplexPairTy ComplexVal = EmitComplexExpr(E->getArg(0));
2625 return RValue::get(ComplexVal.first);
2626 }
2627
2628 case Builtin::BI__builtin_dump_struct: {
2629 llvm::Type *LLVMIntTy = getTypes().ConvertType(getContext().IntTy);
2630 llvm::FunctionType *LLVMFuncType = llvm::FunctionType::get(
2631 LLVMIntTy, {llvm::Type::getInt8PtrTy(getLLVMContext())}, true);
2632
2633 Value *Func = EmitScalarExpr(E->getArg(1)->IgnoreImpCasts());
2634 CharUnits Arg0Align = EmitPointerWithAlignment(E->getArg(0)).getAlignment();
2635
2636 const Expr *Arg0 = E->getArg(0)->IgnoreImpCasts();
2637 QualType Arg0Type = Arg0->getType()->getPointeeType();
2638
2639 Value *RecordPtr = EmitScalarExpr(Arg0);
2640 Value *Res = dumpRecord(*this, Arg0Type, RecordPtr, Arg0Align,
2641 {LLVMFuncType, Func}, 0);
2642 return RValue::get(Res);
2643 }
2644
2645 case Builtin::BI__builtin_preserve_access_index: {
2646 // Only enabled preserved access index region when debuginfo
2647 // is available as debuginfo is needed to preserve user-level
2648 // access pattern.
2649 if (!getDebugInfo()) {
2650 CGM.Error(E->getExprLoc(), "using builtin_preserve_access_index() without -g");
2651 return RValue::get(EmitScalarExpr(E->getArg(0)));
2652 }
2653
2654 // Nested builtin_preserve_access_index() not supported
2655 if (IsInPreservedAIRegion) {
2656 CGM.Error(E->getExprLoc(), "nested builtin_preserve_access_index() not supported");
2657 return RValue::get(EmitScalarExpr(E->getArg(0)));
2658 }
2659
2660 IsInPreservedAIRegion = true;
2661 Value *Res = EmitScalarExpr(E->getArg(0));
2662 IsInPreservedAIRegion = false;
2663 return RValue::get(Res);
2664 }
2665
2666 case Builtin::BI__builtin_cimag:
2667 case Builtin::BI__builtin_cimagf:
2668 case Builtin::BI__builtin_cimagl:
2669 case Builtin::BIcimag:
2670 case Builtin::BIcimagf:
2671 case Builtin::BIcimagl: {
2672 ComplexPairTy ComplexVal = EmitComplexExpr(E->getArg(0));
2673 return RValue::get(ComplexVal.second);
2674 }
2675
2676 case Builtin::BI__builtin_clrsb:
2677 case Builtin::BI__builtin_clrsbl:
2678 case Builtin::BI__builtin_clrsbll: {
2679 // clrsb(x) -> clz(x < 0 ? ~x : x) - 1 or
2680 Value *ArgValue = EmitScalarExpr(E->getArg(0));
2681
2682 llvm::Type *ArgType = ArgValue->getType();
2683 Function *F = CGM.getIntrinsic(Intrinsic::ctlz, ArgType);
2684
2685 llvm::Type *ResultType = ConvertType(E->getType());
2686 Value *Zero = llvm::Constant::getNullValue(ArgType);
2687 Value *IsNeg = Builder.CreateICmpSLT(ArgValue, Zero, "isneg");
2688 Value *Inverse = Builder.CreateNot(ArgValue, "not");
2689 Value *Tmp = Builder.CreateSelect(IsNeg, Inverse, ArgValue);
2690 Value *Ctlz = Builder.CreateCall(F, {Tmp, Builder.getFalse()});
2691 Value *Result = Builder.CreateSub(Ctlz, llvm::ConstantInt::get(ArgType, 1));
2692 Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true,
2693 "cast");
2694 return RValue::get(Result);
2695 }
2696 case Builtin::BI__builtin_ctzs:
2697 case Builtin::BI__builtin_ctz:
2698 case Builtin::BI__builtin_ctzl:
2699 case Builtin::BI__builtin_ctzll: {
2700 Value *ArgValue = EmitCheckedArgForBuiltin(E->getArg(0), BCK_CTZPassedZero);
2701
2702 llvm::Type *ArgType = ArgValue->getType();
2703 Function *F = CGM.getIntrinsic(Intrinsic::cttz, ArgType);
2704
2705 llvm::Type *ResultType = ConvertType(E->getType());
2706 Value *ZeroUndef = Builder.getInt1(getTarget().isCLZForZeroUndef());
2707 Value *Result = Builder.CreateCall(F, {ArgValue, ZeroUndef});
2708 if (Result->getType() != ResultType)
2709 Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true,
2710 "cast");
2711 return RValue::get(Result);
2712 }
2713 case Builtin::BI__builtin_clzs:
2714 case Builtin::BI__builtin_clz:
2715 case Builtin::BI__builtin_clzl:
2716 case Builtin::BI__builtin_clzll: {
2717 Value *ArgValue = EmitCheckedArgForBuiltin(E->getArg(0), BCK_CLZPassedZero);
2718
2719 llvm::Type *ArgType = ArgValue->getType();
2720 Function *F = CGM.getIntrinsic(Intrinsic::ctlz, ArgType);
2721
2722 llvm::Type *ResultType = ConvertType(E->getType());
2723 Value *ZeroUndef = Builder.getInt1(getTarget().isCLZForZeroUndef());
2724 Value *Result = Builder.CreateCall(F, {ArgValue, ZeroUndef});
2725 if (Result->getType() != ResultType)
2726 Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true,
2727 "cast");
2728 return RValue::get(Result);
2729 }
2730 case Builtin::BI__builtin_ffs:
2731 case Builtin::BI__builtin_ffsl:
2732 case Builtin::BI__builtin_ffsll: {
2733 // ffs(x) -> x ? cttz(x) + 1 : 0
2734 Value *ArgValue = EmitScalarExpr(E->getArg(0));
2735
2736 llvm::Type *ArgType = ArgValue->getType();
2737 Function *F = CGM.getIntrinsic(Intrinsic::cttz, ArgType);
2738
2739 llvm::Type *ResultType = ConvertType(E->getType());
2740 Value *Tmp =
2741 Builder.CreateAdd(Builder.CreateCall(F, {ArgValue, Builder.getTrue()}),
2742 llvm::ConstantInt::get(ArgType, 1));
2743 Value *Zero = llvm::Constant::getNullValue(ArgType);
2744 Value *IsZero = Builder.CreateICmpEQ(ArgValue, Zero, "iszero");
2745 Value *Result = Builder.CreateSelect(IsZero, Zero, Tmp, "ffs");
2746 if (Result->getType() != ResultType)
2747 Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true,
2748 "cast");
2749 return RValue::get(Result);
2750 }
2751 case Builtin::BI__builtin_parity:
2752 case Builtin::BI__builtin_parityl:
2753 case Builtin::BI__builtin_parityll: {
2754 // parity(x) -> ctpop(x) & 1
2755 Value *ArgValue = EmitScalarExpr(E->getArg(0));
2756
2757 llvm::Type *ArgType = ArgValue->getType();
2758 Function *F = CGM.getIntrinsic(Intrinsic::ctpop, ArgType);
2759
2760 llvm::Type *ResultType = ConvertType(E->getType());
2761 Value *Tmp = Builder.CreateCall(F, ArgValue);
2762 Value *Result = Builder.CreateAnd(Tmp, llvm::ConstantInt::get(ArgType, 1));
2763 if (Result->getType() != ResultType)
2764 Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true,
2765 "cast");
2766 return RValue::get(Result);
2767 }
2768 case Builtin::BI__lzcnt16:
2769 case Builtin::BI__lzcnt:
2770 case Builtin::BI__lzcnt64: {
2771 Value *ArgValue = EmitScalarExpr(E->getArg(0));
2772
2773 llvm::Type *ArgType = ArgValue->getType();
2774 Function *F = CGM.getIntrinsic(Intrinsic::ctlz, ArgType);
2775
2776 llvm::Type *ResultType = ConvertType(E->getType());
2777 Value *Result = Builder.CreateCall(F, {ArgValue, Builder.getFalse()});
2778 if (Result->getType() != ResultType)
2779 Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true,
2780 "cast");
2781 return RValue::get(Result);
2782 }
2783 case Builtin::BI__popcnt16:
2784 case Builtin::BI__popcnt:
2785 case Builtin::BI__popcnt64:
2786 case Builtin::BI__builtin_popcount:
2787 case Builtin::BI__builtin_popcountl:
2788 case Builtin::BI__builtin_popcountll: {
2789 Value *ArgValue = EmitScalarExpr(E->getArg(0));
2790
2791 llvm::Type *ArgType = ArgValue->getType();
2792 Function *F = CGM.getIntrinsic(Intrinsic::ctpop, ArgType);
2793
2794 llvm::Type *ResultType = ConvertType(E->getType());
2795 Value *Result = Builder.CreateCall(F, ArgValue);
2796 if (Result->getType() != ResultType)
2797 Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true,
2798 "cast");
2799 return RValue::get(Result);
2800 }
2801 case Builtin::BI__builtin_unpredictable: {
2802 // Always return the argument of __builtin_unpredictable. LLVM does not
2803 // handle this builtin. Metadata for this builtin should be added directly
2804 // to instructions such as branches or switches that use it.
2805 return RValue::get(EmitScalarExpr(E->getArg(0)));
2806 }
2807 case Builtin::BI__builtin_expect: {
2808 Value *ArgValue = EmitScalarExpr(E->getArg(0));
2809 llvm::Type *ArgType = ArgValue->getType();
2810
2811 Value *ExpectedValue = EmitScalarExpr(E->getArg(1));
2812 // Don't generate llvm.expect on -O0 as the backend won't use it for
2813 // anything.
2814 // Note, we still IRGen ExpectedValue because it could have side-effects.
2815 if (CGM.getCodeGenOpts().OptimizationLevel == 0)
2816 return RValue::get(ArgValue);
2817
2818 Function *FnExpect = CGM.getIntrinsic(Intrinsic::expect, ArgType);
2819 Value *Result =
2820 Builder.CreateCall(FnExpect, {ArgValue, ExpectedValue}, "expval");
2821 return RValue::get(Result);
2822 }
2823 case Builtin::BI__builtin_expect_with_probability: {
2824 Value *ArgValue = EmitScalarExpr(E->getArg(0));
2825 llvm::Type *ArgType = ArgValue->getType();
2826
2827 Value *ExpectedValue = EmitScalarExpr(E->getArg(1));
2828 llvm::APFloat Probability(0.0);
2829 const Expr *ProbArg = E->getArg(2);
2830 bool EvalSucceed = ProbArg->EvaluateAsFloat(Probability, CGM.getContext());
2831 assert(EvalSucceed && "probability should be able to evaluate as float")((void)0);
2832 (void)EvalSucceed;
2833 bool LoseInfo = false;
2834 Probability.convert(llvm::APFloat::IEEEdouble(),
2835 llvm::RoundingMode::Dynamic, &LoseInfo);
2836 llvm::Type *Ty = ConvertType(ProbArg->getType());
2837 Constant *Confidence = ConstantFP::get(Ty, Probability);
2838 // Don't generate llvm.expect.with.probability on -O0 as the backend
2839 // won't use it for anything.
2840 // Note, we still IRGen ExpectedValue because it could have side-effects.
2841 if (CGM.getCodeGenOpts().OptimizationLevel == 0)
2842 return RValue::get(ArgValue);
2843
2844 Function *FnExpect =
2845 CGM.getIntrinsic(Intrinsic::expect_with_probability, ArgType);
2846 Value *Result = Builder.CreateCall(
2847 FnExpect, {ArgValue, ExpectedValue, Confidence}, "expval");
2848 return RValue::get(Result);
2849 }
2850 case Builtin::BI__builtin_assume_aligned: {
2851 const Expr *Ptr = E->getArg(0);
2852 Value *PtrValue = EmitScalarExpr(Ptr);
2853 Value *OffsetValue =
2854 (E->getNumArgs() > 2) ? EmitScalarExpr(E->getArg(2)) : nullptr;
2855
2856 Value *AlignmentValue = EmitScalarExpr(E->getArg(1));
2857 ConstantInt *AlignmentCI = cast<ConstantInt>(AlignmentValue);
2858 if (AlignmentCI->getValue().ugt(llvm::Value::MaximumAlignment))
2859 AlignmentCI = ConstantInt::get(AlignmentCI->getType(),
2860 llvm::Value::MaximumAlignment);
2861
2862 emitAlignmentAssumption(PtrValue, Ptr,
2863 /*The expr loc is sufficient.*/ SourceLocation(),
2864 AlignmentCI, OffsetValue);
2865 return RValue::get(PtrValue);
2866 }
2867 case Builtin::BI__assume:
2868 case Builtin::BI__builtin_assume: {
2869 if (E->getArg(0)->HasSideEffects(getContext()))
2870 return RValue::get(nullptr);
2871
2872 Value *ArgValue = EmitScalarExpr(E->getArg(0));
2873 Function *FnAssume = CGM.getIntrinsic(Intrinsic::assume);
2874 return RValue::get(Builder.CreateCall(FnAssume, ArgValue));
2875 }
2876 case Builtin::BI__arithmetic_fence: {
2877 // Create the builtin call if FastMath is selected, and the target
2878 // supports the builtin, otherwise just return the argument.
2879 CodeGenFunction::CGFPOptionsRAII FPOptsRAII(*this, E);
2880 llvm::FastMathFlags FMF = Builder.getFastMathFlags();
2881 bool isArithmeticFenceEnabled =
2882 FMF.allowReassoc() &&
2883 getContext().getTargetInfo().checkArithmeticFenceSupported();
2884 QualType ArgType = E->getArg(0)->getType();
2885 if (ArgType->isComplexType()) {
2886 if (isArithmeticFenceEnabled) {
2887 QualType ElementType = ArgType->castAs<ComplexType>()->getElementType();
2888 ComplexPairTy ComplexVal = EmitComplexExpr(E->getArg(0));
2889 Value *Real = Builder.CreateArithmeticFence(ComplexVal.first,
2890 ConvertType(ElementType));
2891 Value *Imag = Builder.CreateArithmeticFence(ComplexVal.second,
2892 ConvertType(ElementType));
2893 return RValue::getComplex(std::make_pair(Real, Imag));
2894 }
2895 ComplexPairTy ComplexVal = EmitComplexExpr(E->getArg(0));
2896 Value *Real = ComplexVal.first;
2897 Value *Imag = ComplexVal.second;
2898 return RValue::getComplex(std::make_pair(Real, Imag));
2899 }
2900 Value *ArgValue = EmitScalarExpr(E->getArg(0));
2901 if (isArithmeticFenceEnabled)
2902 return RValue::get(
2903 Builder.CreateArithmeticFence(ArgValue, ConvertType(ArgType)));
2904 return RValue::get(ArgValue);
2905 }
2906 case Builtin::BI__builtin_bswap16:
2907 case Builtin::BI__builtin_bswap32:
2908 case Builtin::BI__builtin_bswap64: {
2909 return RValue::get(emitUnaryBuiltin(*this, E, Intrinsic::bswap));
2910 }
2911 case Builtin::BI__builtin_bitreverse8:
2912 case Builtin::BI__builtin_bitreverse16:
2913 case Builtin::BI__builtin_bitreverse32:
2914 case Builtin::BI__builtin_bitreverse64: {
2915 return RValue::get(emitUnaryBuiltin(*this, E, Intrinsic::bitreverse));
2916 }
2917 case Builtin::BI__builtin_rotateleft8:
2918 case Builtin::BI__builtin_rotateleft16:
2919 case Builtin::BI__builtin_rotateleft32:
2920 case Builtin::BI__builtin_rotateleft64:
2921 case Builtin::BI_rotl8: // Microsoft variants of rotate left
2922 case Builtin::BI_rotl16:
2923 case Builtin::BI_rotl:
2924 case Builtin::BI_lrotl:
2925 case Builtin::BI_rotl64:
2926 return emitRotate(E, false);
2927
2928 case Builtin::BI__builtin_rotateright8:
2929 case Builtin::BI__builtin_rotateright16:
2930 case Builtin::BI__builtin_rotateright32:
2931 case Builtin::BI__builtin_rotateright64:
2932 case Builtin::BI_rotr8: // Microsoft variants of rotate right
2933 case Builtin::BI_rotr16:
2934 case Builtin::BI_rotr:
2935 case Builtin::BI_lrotr:
2936 case Builtin::BI_rotr64:
2937 return emitRotate(E, true);
2938
2939 case Builtin::BI__builtin_constant_p: {
2940 llvm::Type *ResultType = ConvertType(E->getType());
2941
2942 const Expr *Arg = E->getArg(0);
2943 QualType ArgType = Arg->getType();
2944 // FIXME: The allowance for Obj-C pointers and block pointers is historical
2945 // and likely a mistake.
2946 if (!ArgType->isIntegralOrEnumerationType() && !ArgType->isFloatingType() &&
2947 !ArgType->isObjCObjectPointerType() && !ArgType->isBlockPointerType())
2948 // Per the GCC documentation, only numeric constants are recognized after
2949 // inlining.
2950 return RValue::get(ConstantInt::get(ResultType, 0));
2951
2952 if (Arg->HasSideEffects(getContext()))
2953 // The argument is unevaluated, so be conservative if it might have
2954 // side-effects.
2955 return RValue::get(ConstantInt::get(ResultType, 0));
2956
2957 Value *ArgValue = EmitScalarExpr(Arg);
2958 if (ArgType->isObjCObjectPointerType()) {
2959 // Convert Objective-C objects to id because we cannot distinguish between
2960 // LLVM types for Obj-C classes as they are opaque.
2961 ArgType = CGM.getContext().getObjCIdType();
2962 ArgValue = Builder.CreateBitCast(ArgValue, ConvertType(ArgType));
2963 }
2964 Function *F =
2965 CGM.getIntrinsic(Intrinsic::is_constant, ConvertType(ArgType));
2966 Value *Result = Builder.CreateCall(F, ArgValue);
2967 if (Result->getType() != ResultType)
2968 Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/false);
2969 return RValue::get(Result);
2970 }
2971 case Builtin::BI__builtin_dynamic_object_size:
2972 case Builtin::BI__builtin_object_size: {
2973 unsigned Type =
2974 E->getArg(1)->EvaluateKnownConstInt(getContext()).getZExtValue();
2975 auto *ResType = cast<llvm::IntegerType>(ConvertType(E->getType()));
2976
2977 // We pass this builtin onto the optimizer so that it can figure out the
2978 // object size in more complex cases.
2979 bool IsDynamic = BuiltinID == Builtin::BI__builtin_dynamic_object_size;
2980 return RValue::get(emitBuiltinObjectSize(E->getArg(0), Type, ResType,
2981 /*EmittedE=*/nullptr, IsDynamic));
2982 }
2983 case Builtin::BI__builtin_prefetch: {
2984 Value *Locality, *RW, *Address = EmitScalarExpr(E->getArg(0));
2985 // FIXME: Technically these constants should of type 'int', yes?
2986 RW = (E->getNumArgs() > 1) ? EmitScalarExpr(E->getArg(1)) :
2987 llvm::ConstantInt::get(Int32Ty, 0);
2988 Locality = (E->getNumArgs() > 2) ? EmitScalarExpr(E->getArg(2)) :
2989 llvm::ConstantInt::get(Int32Ty, 3);
2990 Value *Data = llvm::ConstantInt::get(Int32Ty, 1);
2991 Function *F = CGM.getIntrinsic(Intrinsic::prefetch, Address->getType());
2992 return RValue::get(Builder.CreateCall(F, {Address, RW, Locality, Data}));
2993 }
2994 case Builtin::BI__builtin_readcyclecounter: {
2995 Function *F = CGM.getIntrinsic(Intrinsic::readcyclecounter);
2996 return RValue::get(Builder.CreateCall(F));
2997 }
2998 case Builtin::BI__builtin___clear_cache: {
2999 Value *Begin = EmitScalarExpr(E->getArg(0));
3000 Value *End = EmitScalarExpr(E->getArg(1));
3001 Function *F = CGM.getIntrinsic(Intrinsic::clear_cache);
3002 return RValue::get(Builder.CreateCall(F, {Begin, End}));
3003 }
3004 case Builtin::BI__builtin_trap:
3005 return RValue::get(EmitTrapCall(Intrinsic::trap));
3006 case Builtin::BI__debugbreak:
3007 return RValue::get(EmitTrapCall(Intrinsic::debugtrap));
3008 case Builtin::BI__builtin_unreachable: {
3009 EmitUnreachable(E->getExprLoc());
3010
3011 // We do need to preserve an insertion point.
3012 EmitBlock(createBasicBlock("unreachable.cont"));
3013
3014 return RValue::get(nullptr);
3015 }
3016
3017 case Builtin::BI__builtin_powi:
3018 case Builtin::BI__builtin_powif:
3019 case Builtin::BI__builtin_powil: {
3020 llvm::Value *Src0 = EmitScalarExpr(E->getArg(0));
3021 llvm::Value *Src1 = EmitScalarExpr(E->getArg(1));
3022
3023 if (Builder.getIsFPConstrained()) {
3024 CodeGenFunction::CGFPOptionsRAII FPOptsRAII(*this, E);
3025 Function *F = CGM.getIntrinsic(Intrinsic::experimental_constrained_powi,
3026 Src0->getType());
3027 return RValue::get(Builder.CreateConstrainedFPCall(F, { Src0, Src1 }));
3028 }
3029
3030 Function *F = CGM.getIntrinsic(Intrinsic::powi,
3031 { Src0->getType(), Src1->getType() });
3032 return RValue::get(Builder.CreateCall(F, { Src0, Src1 }));
3033 }
3034 case Builtin::BI__builtin_isgreater:
3035 case Builtin::BI__builtin_isgreaterequal:
3036 case Builtin::BI__builtin_isless:
3037 case Builtin::BI__builtin_islessequal:
3038 case Builtin::BI__builtin_islessgreater:
3039 case Builtin::BI__builtin_isunordered: {
3040 // Ordered comparisons: we know the arguments to these are matching scalar
3041 // floating point values.
3042 CodeGenFunction::CGFPOptionsRAII FPOptsRAII(*this, E);
3043 // FIXME: for strictfp/IEEE-754 we need to not trap on SNaN here.
3044 Value *LHS = EmitScalarExpr(E->getArg(0));
3045 Value *RHS = EmitScalarExpr(E->getArg(1));
3046
3047 switch (BuiltinID) {
3048 default: llvm_unreachable("Unknown ordered comparison")__builtin_unreachable();
3049 case Builtin::BI__builtin_isgreater:
3050 LHS = Builder.CreateFCmpOGT(LHS, RHS, "cmp");
3051 break;
3052 case Builtin::BI__builtin_isgreaterequal:
3053 LHS = Builder.CreateFCmpOGE(LHS, RHS, "cmp");
3054 break;
3055 case Builtin::BI__builtin_isless:
3056 LHS = Builder.CreateFCmpOLT(LHS, RHS, "cmp");
3057 break;
3058 case Builtin::BI__builtin_islessequal:
3059 LHS = Builder.CreateFCmpOLE(LHS, RHS, "cmp");
3060 break;
3061 case Builtin::BI__builtin_islessgreater:
3062 LHS = Builder.CreateFCmpONE(LHS, RHS, "cmp");
3063 break;
3064 case Builtin::BI__builtin_isunordered:
3065 LHS = Builder.CreateFCmpUNO(LHS, RHS, "cmp");
3066 break;
3067 }
3068 // ZExt bool to int type.
3069 return RValue::get(Builder.CreateZExt(LHS, ConvertType(E->getType())));
3070 }
3071 case Builtin::BI__builtin_isnan: {
3072 CodeGenFunction::CGFPOptionsRAII FPOptsRAII(*this, E);
3073 Value *V = EmitScalarExpr(E->getArg(0));
3074 llvm::Type *Ty = V->getType();
3075 const llvm::fltSemantics &Semantics = Ty->getFltSemantics();
3076 if (!Builder.getIsFPConstrained() ||
3077 Builder.getDefaultConstrainedExcept() == fp::ebIgnore ||
3078 !Ty->isIEEE()) {
3079 V = Builder.CreateFCmpUNO(V, V, "cmp");
3080 return RValue::get(Builder.CreateZExt(V, ConvertType(E->getType())));
3081 }
3082
3083 if (Value *Result = getTargetHooks().testFPKind(V, BuiltinID, Builder, CGM))
3084 return RValue::get(Result);
3085
3086 // NaN has all exp bits set and a non zero significand. Therefore:
3087 // isnan(V) == ((exp mask - (abs(V) & exp mask)) < 0)
3088 unsigned bitsize = Ty->getScalarSizeInBits();
3089 llvm::IntegerType *IntTy = Builder.getIntNTy(bitsize);
3090 Value *IntV = Builder.CreateBitCast(V, IntTy);
3091 APInt AndMask = APInt::getSignedMaxValue(bitsize);
3092 Value *AbsV =
3093 Builder.CreateAnd(IntV, llvm::ConstantInt::get(IntTy, AndMask));
3094 APInt ExpMask = APFloat::getInf(Semantics).bitcastToAPInt();
3095 Value *Sub =
3096 Builder.CreateSub(llvm::ConstantInt::get(IntTy, ExpMask), AbsV);
3097 // V = sign bit (Sub) <=> V = (Sub < 0)
3098 V = Builder.CreateLShr(Sub, llvm::ConstantInt::get(IntTy, bitsize - 1));
3099 if (bitsize > 32)
3100 V = Builder.CreateTrunc(V, ConvertType(E->getType()));
3101 return RValue::get(V);
3102 }
3103
3104 case Builtin::BI__builtin_matrix_transpose: {
3105 const auto *MatrixTy = E->getArg(0)->getType()->getAs<ConstantMatrixType>();
3106 Value *MatValue = EmitScalarExpr(E->getArg(0));
3107 MatrixBuilder<CGBuilderTy> MB(Builder);
3108 Value *Result = MB.CreateMatrixTranspose(MatValue, MatrixTy->getNumRows(),
3109 MatrixTy->getNumColumns());
3110 return RValue::get(Result);
3111 }
3112
3113 case Builtin::BI__builtin_matrix_column_major_load: {
3114 MatrixBuilder<CGBuilderTy> MB(Builder);
3115 // Emit everything that isn't dependent on the first parameter type
3116 Value *Stride = EmitScalarExpr(E->getArg(3));
3117 const auto *ResultTy = E->getType()->getAs<ConstantMatrixType>();
3118 auto *PtrTy = E->getArg(0)->getType()->getAs<PointerType>();
3119 assert(PtrTy && "arg0 must be of pointer type")((void)0);
3120 bool IsVolatile = PtrTy->getPointeeType().isVolatileQualified();
3121
3122 Address Src = EmitPointerWithAlignment(E->getArg(0));
3123 EmitNonNullArgCheck(RValue::get(Src.getPointer()), E->getArg(0)->getType(),
3124 E->getArg(0)->getExprLoc(), FD, 0);
3125 Value *Result = MB.CreateColumnMajorLoad(
3126 Src.getPointer(), Align(Src.getAlignment().getQuantity()), Stride,
3127 IsVolatile, ResultTy->getNumRows(), ResultTy->getNumColumns(),
3128 "matrix");
3129 return RValue::get(Result);
3130 }
3131
3132 case Builtin::BI__builtin_matrix_column_major_store: {
3133 MatrixBuilder<CGBuilderTy> MB(Builder);
3134 Value *Matrix = EmitScalarExpr(E->getArg(0));
3135 Address Dst = EmitPointerWithAlignment(E->getArg(1));
3136 Value *Stride = EmitScalarExpr(E->getArg(2));
3137
3138 const auto *MatrixTy = E->getArg(0)->getType()->getAs<ConstantMatrixType>();
3139 auto *PtrTy = E->getArg(1)->getType()->getAs<PointerType>();
3140 assert(PtrTy && "arg1 must be of pointer type")((void)0);
3141 bool IsVolatile = PtrTy->getPointeeType().isVolatileQualified();
3142
3143 EmitNonNullArgCheck(RValue::get(Dst.getPointer()), E->getArg(1)->getType(),
3144 E->getArg(1)->getExprLoc(), FD, 0);
3145 Value *Result = MB.CreateColumnMajorStore(
3146 Matrix, Dst.getPointer(), Align(Dst.getAlignment().getQuantity()),
3147 Stride, IsVolatile, MatrixTy->getNumRows(), MatrixTy->getNumColumns());
3148 return RValue::get(Result);
3149 }
3150
3151 case Builtin::BIfinite:
3152 case Builtin::BI__finite:
3153 case Builtin::BIfinitef:
3154 case Builtin::BI__finitef:
3155 case Builtin::BIfinitel:
3156 case Builtin::BI__finitel:
3157 case Builtin::BI__builtin_isinf:
3158 case Builtin::BI__builtin_isfinite: {
3159 // isinf(x) --> fabs(x) == infinity
3160 // isfinite(x) --> fabs(x) != infinity
3161 // x != NaN via the ordered compare in either case.
3162 CodeGenFunction::CGFPOptionsRAII FPOptsRAII(*this, E);
3163 Value *V = EmitScalarExpr(E->getArg(0));
3164 llvm::Type *Ty = V->getType();
3165 if (!Builder.getIsFPConstrained() ||
3166 Builder.getDefaultConstrainedExcept() == fp::ebIgnore ||
3167 !Ty->isIEEE()) {
3168 Value *Fabs = EmitFAbs(*this, V);
3169 Constant *Infinity = ConstantFP::getInfinity(V->getType());
3170 CmpInst::Predicate Pred = (BuiltinID == Builtin::BI__builtin_isinf)
3171 ? CmpInst::FCMP_OEQ
3172 : CmpInst::FCMP_ONE;
3173 Value *FCmp = Builder.CreateFCmp(Pred, Fabs, Infinity, "cmpinf");
3174 return RValue::get(Builder.CreateZExt(FCmp, ConvertType(E->getType())));
3175 }
3176
3177 if (Value *Result = getTargetHooks().testFPKind(V, BuiltinID, Builder, CGM))
3178 return RValue::get(Result);
3179
3180 // Inf values have all exp bits set and a zero significand. Therefore:
3181 // isinf(V) == ((V << 1) == ((exp mask) << 1))
3182 // isfinite(V) == ((V << 1) < ((exp mask) << 1)) using unsigned comparison
3183 unsigned bitsize = Ty->getScalarSizeInBits();
3184 llvm::IntegerType *IntTy = Builder.getIntNTy(bitsize);
3185 Value *IntV = Builder.CreateBitCast(V, IntTy);
3186 Value *Shl1 = Builder.CreateShl(IntV, 1);
3187 const llvm::fltSemantics &Semantics = Ty->getFltSemantics();
3188 APInt ExpMask = APFloat::getInf(Semantics).bitcastToAPInt();
3189 Value *ExpMaskShl1 = llvm::ConstantInt::get(IntTy, ExpMask.shl(1));
3190 if (BuiltinID == Builtin::BI__builtin_isinf)
3191 V = Builder.CreateICmpEQ(Shl1, ExpMaskShl1);
3192 else
3193 V = Builder.CreateICmpULT(Shl1, ExpMaskShl1);
3194 return RValue::get(Builder.CreateZExt(V, ConvertType(E->getType())));
3195 }
3196
3197 case Builtin::BI__builtin_isinf_sign: {
3198 // isinf_sign(x) -> fabs(x) == infinity ? (signbit(x) ? -1 : 1) : 0
3199 CodeGenFunction::CGFPOptionsRAII FPOptsRAII(*this, E);
3200 // FIXME: for strictfp/IEEE-754 we need to not trap on SNaN here.
3201 Value *Arg = EmitScalarExpr(E->getArg(0));
3202 Value *AbsArg = EmitFAbs(*this, Arg);
3203 Value *IsInf = Builder.CreateFCmpOEQ(
3204 AbsArg, ConstantFP::getInfinity(Arg->getType()), "isinf");
3205 Value *IsNeg = EmitSignBit(*this, Arg);
3206
3207 llvm::Type *IntTy = ConvertType(E->getType());
3208 Value *Zero = Constant::getNullValue(IntTy);
3209 Value *One = ConstantInt::get(IntTy, 1);
3210 Value *NegativeOne = ConstantInt::get(IntTy, -1);
3211 Value *SignResult = Builder.CreateSelect(IsNeg, NegativeOne, One);
3212 Value *Result = Builder.CreateSelect(IsInf, SignResult, Zero);
3213 return RValue::get(Result);
3214 }
3215
3216 case Builtin::BI__builtin_isnormal: {
3217 // isnormal(x) --> x == x && fabsf(x) < infinity && fabsf(x) >= float_min
3218 CodeGenFunction::CGFPOptionsRAII FPOptsRAII(*this, E);
3219 // FIXME: for strictfp/IEEE-754 we need to not trap on SNaN here.
3220 Value *V = EmitScalarExpr(E->getArg(0));
3221 Value *Eq = Builder.CreateFCmpOEQ(V, V, "iseq");
3222
3223 Value *Abs = EmitFAbs(*this, V);
3224 Value *IsLessThanInf =
3225 Builder.CreateFCmpULT(Abs, ConstantFP::getInfinity(V->getType()),"isinf");
3226 APFloat Smallest = APFloat::getSmallestNormalized(
3227 getContext().getFloatTypeSemantics(E->getArg(0)->getType()));
3228 Value *IsNormal =
3229 Builder.CreateFCmpUGE(Abs, ConstantFP::get(V->getContext(), Smallest),
3230 "isnormal");
3231 V = Builder.CreateAnd(Eq, IsLessThanInf, "and");
3232 V = Builder.CreateAnd(V, IsNormal, "and");
3233 return RValue::get(Builder.CreateZExt(V, ConvertType(E->getType())));
3234 }
3235
3236 case Builtin::BI__builtin_flt_rounds: {
3237 Function *F = CGM.getIntrinsic(Intrinsic::flt_rounds);
3238
3239 llvm::Type *ResultType = ConvertType(E->getType());
3240 Value *Result = Builder.CreateCall(F);
3241 if (Result->getType() != ResultType)
3242 Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true,
3243 "cast");
3244 return RValue::get(Result);
3245 }
3246
3247 case Builtin::BI__builtin_fpclassify: {
3248 CodeGenFunction::CGFPOptionsRAII FPOptsRAII(*this, E);
3249 // FIXME: for strictfp/IEEE-754 we need to not trap on SNaN here.
3250 Value *V = EmitScalarExpr(E->getArg(5));
3251 llvm::Type *Ty = ConvertType(E->getArg(5)->getType());
3252
3253 // Create Result
3254 BasicBlock *Begin = Builder.GetInsertBlock();
3255 BasicBlock *End = createBasicBlock("fpclassify_end", this->CurFn);
3256 Builder.SetInsertPoint(End);
3257 PHINode *Result =
3258 Builder.CreatePHI(ConvertType(E->getArg(0)->getType()), 4,
3259 "fpclassify_result");
3260
3261 // if (V==0) return FP_ZERO
3262 Builder.SetInsertPoint(Begin);
3263 Value *IsZero = Builder.CreateFCmpOEQ(V, Constant::getNullValue(Ty),
3264 "iszero");
3265 Value *ZeroLiteral = EmitScalarExpr(E->getArg(4));
3266 BasicBlock *NotZero = createBasicBlock("fpclassify_not_zero", this->CurFn);
3267 Builder.CreateCondBr(IsZero, End, NotZero);
3268 Result->addIncoming(ZeroLiteral, Begin);
3269
3270 // if (V != V) return FP_NAN
3271 Builder.SetInsertPoint(NotZero);
3272 Value *IsNan = Builder.CreateFCmpUNO(V, V, "cmp");
3273 Value *NanLiteral = EmitScalarExpr(E->getArg(0));
3274 BasicBlock *NotNan = createBasicBlock("fpclassify_not_nan", this->CurFn);
3275 Builder.CreateCondBr(IsNan, End, NotNan);
3276 Result->addIncoming(NanLiteral, NotZero);
3277
3278 // if (fabs(V) == infinity) return FP_INFINITY
3279 Builder.SetInsertPoint(NotNan);
3280 Value *VAbs = EmitFAbs(*this, V);
3281 Value *IsInf =
3282 Builder.CreateFCmpOEQ(VAbs, ConstantFP::getInfinity(V->getType()),
3283 "isinf");
3284 Value *InfLiteral = EmitScalarExpr(E->getArg(1));
3285 BasicBlock *NotInf = createBasicBlock("fpclassify_not_inf", this->CurFn);
3286 Builder.CreateCondBr(IsInf, End, NotInf);
3287 Result->addIncoming(InfLiteral, NotNan);
3288
3289 // if (fabs(V) >= MIN_NORMAL) return FP_NORMAL else FP_SUBNORMAL
3290 Builder.SetInsertPoint(NotInf);
3291 APFloat Smallest = APFloat::getSmallestNormalized(
3292 getContext().getFloatTypeSemantics(E->getArg(5)->getType()));
3293 Value *IsNormal =
3294 Builder.CreateFCmpUGE(VAbs, ConstantFP::get(V->getContext(), Smallest),
3295 "isnormal");
3296 Value *NormalResult =
3297 Builder.CreateSelect(IsNormal, EmitScalarExpr(E->getArg(2)),
3298 EmitScalarExpr(E->getArg(3)));
3299 Builder.CreateBr(End);
3300 Result->addIncoming(NormalResult, NotInf);
3301
3302 // return Result
3303 Builder.SetInsertPoint(End);
3304 return RValue::get(Result);
3305 }
3306
3307 case Builtin::BIalloca:
3308 case Builtin::BI_alloca:
3309 case Builtin::BI__builtin_alloca: {
3310 Value *Size = EmitScalarExpr(E->getArg(0));
3311 const TargetInfo &TI = getContext().getTargetInfo();
3312 // The alignment of the alloca should correspond to __BIGGEST_ALIGNMENT__.
3313 const Align SuitableAlignmentInBytes =
3314 CGM.getContext()
3315 .toCharUnitsFromBits(TI.getSuitableAlign())
3316 .getAsAlign();
3317 AllocaInst *AI = Builder.CreateAlloca(Builder.getInt8Ty(), Size);
3318 AI->setAlignment(SuitableAlignmentInBytes);
3319 initializeAlloca(*this, AI, Size, SuitableAlignmentInBytes);
3320 return RValue::get(AI);
3321 }
3322
3323 case Builtin::BI__builtin_alloca_with_align: {
3324 Value *Size = EmitScalarExpr(E->getArg(0));
3325 Value *AlignmentInBitsValue = EmitScalarExpr(E->getArg(1));
3326 auto *AlignmentInBitsCI = cast<ConstantInt>(AlignmentInBitsValue);
3327 unsigned AlignmentInBits = AlignmentInBitsCI->getZExtValue();
3328 const Align AlignmentInBytes =
3329 CGM.getContext().toCharUnitsFromBits(AlignmentInBits).getAsAlign();
3330 AllocaInst *AI = Builder.CreateAlloca(Builder.getInt8Ty(), Size);
3331 AI->setAlignment(AlignmentInBytes);
3332 initializeAlloca(*this, AI, Size, AlignmentInBytes);
3333 return RValue::get(AI);
3334 }
3335
3336 case Builtin::BIbzero:
3337 case Builtin::BI__builtin_bzero: {
3338 Address Dest = EmitPointerWithAlignment(E->getArg(0));
3339 Value *SizeVal = EmitScalarExpr(E->getArg(1));
3340 EmitNonNullArgCheck(RValue::get(Dest.getPointer()), E->getArg(0)->getType(),
3341 E->getArg(0)->getExprLoc(), FD, 0);
3342 Builder.CreateMemSet(Dest, Builder.getInt8(0), SizeVal, false);
3343 return RValue::get(nullptr);
3344 }
3345 case Builtin::BImemcpy:
3346 case Builtin::BI__builtin_memcpy:
3347 case Builtin::BImempcpy:
3348 case Builtin::BI__builtin_mempcpy: {
3349 Address Dest = EmitPointerWithAlignment(E->getArg(0));
3350 Address Src = EmitPointerWithAlignment(E->getArg(1));
3351 Value *SizeVal = EmitScalarExpr(E->getArg(2));
3352 EmitNonNullArgCheck(RValue::get(Dest.getPointer()), E->getArg(0)->getType(),
3353 E->getArg(0)->getExprLoc(), FD, 0);
3354 EmitNonNullArgCheck(RValue::get(Src.getPointer()), E->getArg(1)->getType(),
3355 E->getArg(1)->getExprLoc(), FD, 1);
3356 Builder.CreateMemCpy(Dest, Src, SizeVal, false);
3357 if (BuiltinID == Builtin::BImempcpy ||
3358 BuiltinID == Builtin::BI__builtin_mempcpy)
3359 return RValue::get(Builder.CreateInBoundsGEP(Dest.getElementType(),
3360 Dest.getPointer(), SizeVal));
3361 else
3362 return RValue::get(Dest.getPointer());
3363 }
3364
3365 case Builtin::BI__builtin_memcpy_inline: {
3366 Address Dest = EmitPointerWithAlignment(E->getArg(0));
3367 Address Src = EmitPointerWithAlignment(E->getArg(1));
3368 uint64_t Size =
3369 E->getArg(2)->EvaluateKnownConstInt(getContext()).getZExtValue();
3370 EmitNonNullArgCheck(RValue::get(Dest.getPointer()), E->getArg(0)->getType(),
3371 E->getArg(0)->getExprLoc(), FD, 0);
3372 EmitNonNullArgCheck(RValue::get(Src.getPointer()), E->getArg(1)->getType(),
3373 E->getArg(1)->getExprLoc(), FD, 1);
3374 Builder.CreateMemCpyInline(Dest, Src, Size);
3375 return RValue::get(nullptr);
3376 }
3377
3378 case Builtin::BI__builtin_char_memchr:
3379 BuiltinID = Builtin::BI__builtin_memchr;
3380 break;
3381
3382 case Builtin::BI__builtin___memcpy_chk: {
3383 // fold __builtin_memcpy_chk(x, y, cst1, cst2) to memcpy iff cst1<=cst2.
3384 Expr::EvalResult SizeResult, DstSizeResult;
3385 if (!E->getArg(2)->EvaluateAsInt(SizeResult, CGM.getContext()) ||
3386 !E->getArg(3)->EvaluateAsInt(DstSizeResult, CGM.getContext()))
3387 break;
3388 llvm::APSInt Size = SizeResult.Val.getInt();
3389 llvm::APSInt DstSize = DstSizeResult.Val.getInt();
3390 if (Size.ugt(DstSize))
3391 break;
3392 Address Dest = EmitPointerWithAlignment(E->getArg(0));
3393 Address Src = EmitPointerWithAlignment(E->getArg(1));
3394 Value *SizeVal = llvm::ConstantInt::get(Builder.getContext(), Size);
3395 Builder.CreateMemCpy(Dest, Src, SizeVal, false);
3396 return RValue::get(Dest.getPointer());
3397 }
3398
3399 case Builtin::BI__builtin_objc_memmove_collectable: {
3400 Address DestAddr = EmitPointerWithAlignment(E->getArg(0));
3401 Address SrcAddr = EmitPointerWithAlignment(E->getArg(1));
3402 Value *SizeVal = EmitScalarExpr(E->getArg(2));
3403 CGM.getObjCRuntime().EmitGCMemmoveCollectable(*this,
3404 DestAddr, SrcAddr, SizeVal);
3405 return RValue::get(DestAddr.getPointer());
3406 }
3407
3408 case Builtin::BI__builtin___memmove_chk: {
3409 // fold __builtin_memmove_chk(x, y, cst1, cst2) to memmove iff cst1<=cst2.
3410 Expr::EvalResult SizeResult, DstSizeResult;
3411 if (!E->getArg(2)->EvaluateAsInt(SizeResult, CGM.getContext()) ||
3412 !E->getArg(3)->EvaluateAsInt(DstSizeResult, CGM.getContext()))
3413 break;
3414 llvm::APSInt Size = SizeResult.Val.getInt();
3415 llvm::APSInt DstSize = DstSizeResult.Val.getInt();
3416 if (Size.ugt(DstSize))
3417 break;
3418 Address Dest = EmitPointerWithAlignment(E->getArg(0));
3419 Address Src = EmitPointerWithAlignment(E->getArg(1));
3420 Value *SizeVal = llvm::ConstantInt::get(Builder.getContext(), Size);
3421 Builder.CreateMemMove(Dest, Src, SizeVal, false);
3422 return RValue::get(Dest.getPointer());
3423 }
3424
3425 case Builtin::BImemmove:
3426 case Builtin::BI__builtin_memmove: {
3427 Address Dest = EmitPointerWithAlignment(E->getArg(0));
3428 Address Src = EmitPointerWithAlignment(E->getArg(1));
3429 Value *SizeVal = EmitScalarExpr(E->getArg(2));
3430 EmitNonNullArgCheck(RValue::get(Dest.getPointer()), E->getArg(0)->getType(),
3431 E->getArg(0)->getExprLoc(), FD, 0);
3432 EmitNonNullArgCheck(RValue::get(Src.getPointer()), E->getArg(1)->getType(),
3433 E->getArg(1)->getExprLoc(), FD, 1);
3434 Builder.CreateMemMove(Dest, Src, SizeVal, false);
3435 return RValue::get(Dest.getPointer());
3436 }
3437 case Builtin::BImemset:
3438 case Builtin::BI__builtin_memset: {
3439 Address Dest = EmitPointerWithAlignment(E->getArg(0));
3440 Value *ByteVal = Builder.CreateTrunc(EmitScalarExpr(E->getArg(1)),
3441 Builder.getInt8Ty());
3442 Value *SizeVal = EmitScalarExpr(E->getArg(2));
3443 EmitNonNullArgCheck(RValue::get(Dest.getPointer()), E->getArg(0)->getType(),
3444 E->getArg(0)->getExprLoc(), FD, 0);
3445 Builder.CreateMemSet(Dest, ByteVal, SizeVal, false);
3446 return RValue::get(Dest.getPointer());
3447 }
3448 case Builtin::BI__builtin___memset_chk: {
3449 // fold __builtin_memset_chk(x, y, cst1, cst2) to memset iff cst1<=cst2.
3450 Expr::EvalResult SizeResult, DstSizeResult;
3451 if (!E->getArg(2)->EvaluateAsInt(SizeResult, CGM.getContext()) ||
3452 !E->getArg(3)->EvaluateAsInt(DstSizeResult, CGM.getContext()))
3453 break;
3454 llvm::APSInt Size = SizeResult.Val.getInt();
3455 llvm::APSInt DstSize = DstSizeResult.Val.getInt();
3456 if (Size.ugt(DstSize))
3457 break;
3458 Address Dest = EmitPointerWithAlignment(E->getArg(0));
3459 Value *ByteVal = Builder.CreateTrunc(EmitScalarExpr(E->getArg(1)),
3460 Builder.getInt8Ty());
3461 Value *SizeVal = llvm::ConstantInt::get(Builder.getContext(), Size);
3462 Builder.CreateMemSet(Dest, ByteVal, SizeVal, false);
3463 return RValue::get(Dest.getPointer());
3464 }
3465 case Builtin::BI__builtin_wmemchr: {
3466 // The MSVC runtime library does not provide a definition of wmemchr, so we
3467 // need an inline implementation.
3468 if (!getTarget().getTriple().isOSMSVCRT())
3469 break;
3470
3471 llvm::Type *WCharTy = ConvertType(getContext().WCharTy);
3472 Value *Str = EmitScalarExpr(E->getArg(0));
3473 Value *Chr = EmitScalarExpr(E->getArg(1));
3474 Value *Size = EmitScalarExpr(E->getArg(2));
3475
3476 BasicBlock *Entry = Builder.GetInsertBlock();
3477 BasicBlock *CmpEq = createBasicBlock("wmemchr.eq");
3478 BasicBlock *Next = createBasicBlock("wmemchr.next");
3479 BasicBlock *Exit = createBasicBlock("wmemchr.exit");
3480 Value *SizeEq0 = Builder.CreateICmpEQ(Size, ConstantInt::get(SizeTy, 0));
3481 Builder.CreateCondBr(SizeEq0, Exit, CmpEq);
3482
3483 EmitBlock(CmpEq);
3484 PHINode *StrPhi = Builder.CreatePHI(Str->getType(), 2);
3485 StrPhi->addIncoming(Str, Entry);
3486 PHINode *SizePhi = Builder.CreatePHI(SizeTy, 2);
3487 SizePhi->addIncoming(Size, Entry);
3488 CharUnits WCharAlign =
3489 getContext().getTypeAlignInChars(getContext().WCharTy);
3490 Value *StrCh = Builder.CreateAlignedLoad(WCharTy, StrPhi, WCharAlign);
3491 Value *FoundChr = Builder.CreateConstInBoundsGEP1_32(WCharTy, StrPhi, 0);
3492 Value *StrEqChr = Builder.CreateICmpEQ(StrCh, Chr);
3493 Builder.CreateCondBr(StrEqChr, Exit, Next);
3494
3495 EmitBlock(Next);
3496 Value *NextStr = Builder.CreateConstInBoundsGEP1_32(WCharTy, StrPhi, 1);
3497 Value *NextSize = Builder.CreateSub(SizePhi, ConstantInt::get(SizeTy, 1));
3498 Value *NextSizeEq0 =
3499 Builder.CreateICmpEQ(NextSize, ConstantInt::get(SizeTy, 0));
3500 Builder.CreateCondBr(NextSizeEq0, Exit, CmpEq);
3501 StrPhi->addIncoming(NextStr, Next);
3502 SizePhi->addIncoming(NextSize, Next);
3503
3504 EmitBlock(Exit);
3505 PHINode *Ret = Builder.CreatePHI(Str->getType(), 3);
3506 Ret->addIncoming(llvm::Constant::getNullValue(Str->getType()), Entry);
3507 Ret->addIncoming(llvm::Constant::getNullValue(Str->getType()), Next);
3508 Ret->addIncoming(FoundChr, CmpEq);
3509 return RValue::get(Ret);
3510 }
3511 case Builtin::BI__builtin_wmemcmp: {
3512 // The MSVC runtime library does not provide a definition of wmemcmp, so we
3513 // need an inline implementation.
3514 if (!getTarget().getTriple().isOSMSVCRT())
3515 break;
3516
3517 llvm::Type *WCharTy = ConvertType(getContext().WCharTy);
3518
3519 Value *Dst = EmitScalarExpr(E->getArg(0));
3520 Value *Src = EmitScalarExpr(E->getArg(1));
3521 Value *Size = EmitScalarExpr(E->getArg(2));
3522
3523 BasicBlock *Entry = Builder.GetInsertBlock();
3524 BasicBlock *CmpGT = createBasicBlock("wmemcmp.gt");
3525 BasicBlock *CmpLT = createBasicBlock("wmemcmp.lt");
3526 BasicBlock *Next = createBasicBlock("wmemcmp.next");
3527 BasicBlock *Exit = createBasicBlock("wmemcmp.exit");
3528 Value *SizeEq0 = Builder.CreateICmpEQ(Size, ConstantInt::get(SizeTy, 0));
3529 Builder.CreateCondBr(SizeEq0, Exit, CmpGT);
3530
3531 EmitBlock(CmpGT);
3532 PHINode *DstPhi = Builder.CreatePHI(Dst->getType(), 2);
3533 DstPhi->addIncoming(Dst, Entry);
3534 PHINode *SrcPhi = Builder.CreatePHI(Src->getType(), 2);
3535 SrcPhi->addIncoming(Src, Entry);
3536 PHINode *SizePhi = Builder.CreatePHI(SizeTy, 2);
3537 SizePhi->addIncoming(Size, Entry);
3538 CharUnits WCharAlign =
3539 getContext().getTypeAlignInChars(getContext().WCharTy);
3540 Value *DstCh = Builder.CreateAlignedLoad(WCharTy, DstPhi, WCharAlign);
3541 Value *SrcCh = Builder.CreateAlignedLoad(WCharTy, SrcPhi, WCharAlign);
3542 Value *DstGtSrc = Builder.CreateICmpUGT(DstCh, SrcCh);
3543 Builder.CreateCondBr(DstGtSrc, Exit, CmpLT);
3544
3545 EmitBlock(CmpLT);
3546 Value *DstLtSrc = Builder.CreateICmpULT(DstCh, SrcCh);
3547 Builder.CreateCondBr(DstLtSrc, Exit, Next);
3548
3549 EmitBlock(Next);
3550 Value *NextDst = Builder.CreateConstInBoundsGEP1_32(WCharTy, DstPhi, 1);
3551 Value *NextSrc = Builder.CreateConstInBoundsGEP1_32(WCharTy, SrcPhi, 1);
3552 Value *NextSize = Builder.CreateSub(SizePhi, ConstantInt::get(SizeTy, 1));
3553 Value *NextSizeEq0 =
3554 Builder.CreateICmpEQ(NextSize, ConstantInt::get(SizeTy, 0));
3555 Builder.CreateCondBr(NextSizeEq0, Exit, CmpGT);
3556 DstPhi->addIncoming(NextDst, Next);
3557 SrcPhi->addIncoming(NextSrc, Next);
3558 SizePhi->addIncoming(NextSize, Next);
3559
3560 EmitBlock(Exit);
3561 PHINode *Ret = Builder.CreatePHI(IntTy, 4);
3562 Ret->addIncoming(ConstantInt::get(IntTy, 0), Entry);
3563 Ret->addIncoming(ConstantInt::get(IntTy, 1), CmpGT);
3564 Ret->addIncoming(ConstantInt::get(IntTy, -1), CmpLT);
3565 Ret->addIncoming(ConstantInt::get(IntTy, 0), Next);
3566 return RValue::get(Ret);
3567 }
3568 case Builtin::BI__builtin_dwarf_cfa: {
3569 // The offset in bytes from the first argument to the CFA.
3570 //
3571 // Why on earth is this in the frontend? Is there any reason at
3572 // all that the backend can't reasonably determine this while
3573 // lowering llvm.eh.dwarf.cfa()?
3574 //
3575 // TODO: If there's a satisfactory reason, add a target hook for
3576 // this instead of hard-coding 0, which is correct for most targets.
3577 int32_t Offset = 0;
3578
3579 Function *F = CGM.getIntrinsic(Intrinsic::eh_dwarf_cfa);
3580 return RValue::get(Builder.CreateCall(F,
3581 llvm::ConstantInt::get(Int32Ty, Offset)));
3582 }
3583 case Builtin::BI__builtin_return_address: {
3584 Value *Depth = ConstantEmitter(*this).emitAbstract(E->getArg(0),
3585 getContext().UnsignedIntTy);
3586 Function *F = CGM.getIntrinsic(Intrinsic::returnaddress);
3587 return RValue::get(Builder.CreateCall(F, Depth));
3588 }
3589 case Builtin::BI_ReturnAddress: {
3590 Function *F = CGM.getIntrinsic(Intrinsic::returnaddress);
3591 return RValue::get(Builder.CreateCall(F, Builder.getInt32(0)));
3592 }
3593 case Builtin::BI__builtin_frame_address: {
3594 Value *Depth = ConstantEmitter(*this).emitAbstract(E->getArg(0),
3595 getContext().UnsignedIntTy);
3596 Function *F = CGM.getIntrinsic(Intrinsic::frameaddress, AllocaInt8PtrTy);
3597 return RValue::get(Builder.CreateCall(F, Depth));
3598 }
3599 case Builtin::BI__builtin_extract_return_addr: {
3600 Value *Address = EmitScalarExpr(E->getArg(0));
3601 Value *Result = getTargetHooks().decodeReturnAddress(*this, Address);
3602 return RValue::get(Result);
3603 }
3604 case Builtin::BI__builtin_frob_return_addr: {
3605 Value *Address = EmitScalarExpr(E->getArg(0));
3606 Value *Result = getTargetHooks().encodeReturnAddress(*this, Address);
3607 return RValue::get(Result);
3608 }
3609 case Builtin::BI__builtin_dwarf_sp_column: {
3610 llvm::IntegerType *Ty
3611 = cast<llvm::IntegerType>(ConvertType(E->getType()));
3612 int Column = getTargetHooks().getDwarfEHStackPointer(CGM);
3613 if (Column == -1) {
3614 CGM.ErrorUnsupported(E, "__builtin_dwarf_sp_column");
3615 return RValue::get(llvm::UndefValue::get(Ty));
3616 }
3617 return RValue::get(llvm::ConstantInt::get(Ty, Column, true));
3618 }
3619 case Builtin::BI__builtin_init_dwarf_reg_size_table: {
3620 Value *Address = EmitScalarExpr(E->getArg(0));
3621 if (getTargetHooks().initDwarfEHRegSizeTable(*this, Address))
3622 CGM.ErrorUnsupported(E, "__builtin_init_dwarf_reg_size_table");
3623 return RValue::get(llvm::UndefValue::get(ConvertType(E->getType())));
3624 }
3625 case Builtin::BI__builtin_eh_return: {
3626 Value *Int = EmitScalarExpr(E->getArg(0));
3627 Value *Ptr = EmitScalarExpr(E->getArg(1));
3628
3629 llvm::IntegerType *IntTy = cast<llvm::IntegerType>(Int->getType());
3630 assert((IntTy->getBitWidth() == 32 || IntTy->getBitWidth() == 64) &&((void)0)
3631 "LLVM's __builtin_eh_return only supports 32- and 64-bit variants")((void)0);
3632 Function *F =
3633 CGM.getIntrinsic(IntTy->getBitWidth() == 32 ? Intrinsic::eh_return_i32
3634 : Intrinsic::eh_return_i64);
3635 Builder.CreateCall(F, {Int, Ptr});
3636 Builder.CreateUnreachable();
3637
3638 // We do need to preserve an insertion point.
3639 EmitBlock(createBasicBlock("builtin_eh_return.cont"));
3640
3641 return RValue::get(nullptr);
3642 }
3643 case Builtin::BI__builtin_unwind_init: {
3644 Function *F = CGM.getIntrinsic(Intrinsic::eh_unwind_init);
3645 return RValue::get(Builder.CreateCall(F));
3646 }
3647 case Builtin::BI__builtin_extend_pointer: {
3648 // Extends a pointer to the size of an _Unwind_Word, which is
3649 // uint64_t on all platforms. Generally this gets poked into a
3650 // register and eventually used as an address, so if the
3651 // addressing registers are wider than pointers and the platform
3652 // doesn't implicitly ignore high-order bits when doing
3653 // addressing, we need to make sure we zext / sext based on
3654 // the platform's expectations.
3655 //
3656 // See: http://gcc.gnu.org/ml/gcc-bugs/2002-02/msg00237.html
3657
3658 // Cast the pointer to intptr_t.
3659 Value *Ptr = EmitScalarExpr(E->getArg(0));
3660 Value *Result = Builder.CreatePtrToInt(Ptr, IntPtrTy, "extend.cast");
3661
3662 // If that's 64 bits, we're done.
3663 if (IntPtrTy->getBitWidth() == 64)
3664 return RValue::get(Result);
3665
3666 // Otherwise, ask the codegen data what to do.
3667 if (getTargetHooks().extendPointerWithSExt())
3668 return RValue::get(Builder.CreateSExt(Result, Int64Ty, "extend.sext"));
3669 else
3670 return RValue::get(Builder.CreateZExt(Result, Int64Ty, "extend.zext"));
3671 }
3672 case Builtin::BI__builtin_setjmp: {
3673 // Buffer is a void**.
3674 Address Buf = EmitPointerWithAlignment(E->getArg(0));
3675
3676 // Store the frame pointer to the setjmp buffer.
3677 Value *FrameAddr = Builder.CreateCall(
3678 CGM.getIntrinsic(Intrinsic::frameaddress, AllocaInt8PtrTy),
3679 ConstantInt::get(Int32Ty, 0));
3680 Builder.CreateStore(FrameAddr, Buf);
3681
3682 // Store the stack pointer to the setjmp buffer.
3683 Value *StackAddr =
3684 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::stacksave));
3685 Address StackSaveSlot = Builder.CreateConstInBoundsGEP(Buf, 2);
3686 Builder.CreateStore(StackAddr, StackSaveSlot);
3687
3688 // Call LLVM's EH setjmp, which is lightweight.
3689 Function *F = CGM.getIntrinsic(Intrinsic::eh_sjlj_setjmp);
3690 Buf = Builder.CreateBitCast(Buf, Int8PtrTy);
3691 return RValue::get(Builder.CreateCall(F, Buf.getPointer()));
3692 }
3693 case Builtin::BI__builtin_longjmp: {
3694 Value *Buf = EmitScalarExpr(E->getArg(0));
3695 Buf = Builder.CreateBitCast(Buf, Int8PtrTy);
3696
3697 // Call LLVM's EH longjmp, which is lightweight.
3698 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::eh_sjlj_longjmp), Buf);
3699
3700 // longjmp doesn't return; mark this as unreachable.
3701 Builder.CreateUnreachable();
3702
3703 // We do need to preserve an insertion point.
3704 EmitBlock(createBasicBlock("longjmp.cont"));
3705
3706 return RValue::get(nullptr);
3707 }
3708 case Builtin::BI__builtin_launder: {
3709 const Expr *Arg = E->getArg(0);
3710 QualType ArgTy = Arg->getType()->getPointeeType();
3711 Value *Ptr = EmitScalarExpr(Arg);
3712 if (TypeRequiresBuiltinLaunder(CGM, ArgTy))
3713 Ptr = Builder.CreateLaunderInvariantGroup(Ptr);
3714
3715 return RValue::get(Ptr);
3716 }
3717 case Builtin::BI__sync_fetch_and_add:
3718 case Builtin::BI__sync_fetch_and_sub:
3719 case Builtin::BI__sync_fetch_and_or:
3720 case Builtin::BI__sync_fetch_and_and:
3721 case Builtin::BI__sync_fetch_and_xor:
3722 case Builtin::BI__sync_fetch_and_nand:
3723 case Builtin::BI__sync_add_and_fetch:
3724 case Builtin::BI__sync_sub_and_fetch:
3725 case Builtin::BI__sync_and_and_fetch:
3726 case Builtin::BI__sync_or_and_fetch:
3727 case Builtin::BI__sync_xor_and_fetch:
3728 case Builtin::BI__sync_nand_and_fetch:
3729 case Builtin::BI__sync_val_compare_and_swap:
3730 case Builtin::BI__sync_bool_compare_and_swap:
3731 case Builtin::BI__sync_lock_test_and_set:
3732 case Builtin::BI__sync_lock_release:
3733 case Builtin::BI__sync_swap:
3734 llvm_unreachable("Shouldn't make it through sema")__builtin_unreachable();
3735 case Builtin::BI__sync_fetch_and_add_1:
3736 case Builtin::BI__sync_fetch_and_add_2:
3737 case Builtin::BI__sync_fetch_and_add_4:
3738 case Builtin::BI__sync_fetch_and_add_8:
3739 case Builtin::BI__sync_fetch_and_add_16:
3740 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Add, E);
3741 case Builtin::BI__sync_fetch_and_sub_1:
3742 case Builtin::BI__sync_fetch_and_sub_2:
3743 case Builtin::BI__sync_fetch_and_sub_4:
3744 case Builtin::BI__sync_fetch_and_sub_8:
3745 case Builtin::BI__sync_fetch_and_sub_16:
3746 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Sub, E);
3747 case Builtin::BI__sync_fetch_and_or_1:
3748 case Builtin::BI__sync_fetch_and_or_2:
3749 case Builtin::BI__sync_fetch_and_or_4:
3750 case Builtin::BI__sync_fetch_and_or_8:
3751 case Builtin::BI__sync_fetch_and_or_16:
3752 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Or, E);
3753 case Builtin::BI__sync_fetch_and_and_1:
3754 case Builtin::BI__sync_fetch_and_and_2:
3755 case Builtin::BI__sync_fetch_and_and_4:
3756 case Builtin::BI__sync_fetch_and_and_8:
3757 case Builtin::BI__sync_fetch_and_and_16:
3758 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::And, E);
3759 case Builtin::BI__sync_fetch_and_xor_1:
3760 case Builtin::BI__sync_fetch_and_xor_2:
3761 case Builtin::BI__sync_fetch_and_xor_4:
3762 case Builtin::BI__sync_fetch_and_xor_8:
3763 case Builtin::BI__sync_fetch_and_xor_16:
3764 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xor, E);
3765 case Builtin::BI__sync_fetch_and_nand_1:
3766 case Builtin::BI__sync_fetch_and_nand_2:
3767 case Builtin::BI__sync_fetch_and_nand_4:
3768 case Builtin::BI__sync_fetch_and_nand_8:
3769 case Builtin::BI__sync_fetch_and_nand_16:
3770 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Nand, E);
3771
3772 // Clang extensions: not overloaded yet.
3773 case Builtin::BI__sync_fetch_and_min:
3774 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Min, E);
3775 case Builtin::BI__sync_fetch_and_max:
3776 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Max, E);
3777 case Builtin::BI__sync_fetch_and_umin:
3778 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::UMin, E);
3779 case Builtin::BI__sync_fetch_and_umax:
3780 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::UMax, E);
3781
3782 case Builtin::BI__sync_add_and_fetch_1:
3783 case Builtin::BI__sync_add_and_fetch_2:
3784 case Builtin::BI__sync_add_and_fetch_4:
3785 case Builtin::BI__sync_add_and_fetch_8:
3786 case Builtin::BI__sync_add_and_fetch_16:
3787 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Add, E,
3788 llvm::Instruction::Add);
3789 case Builtin::BI__sync_sub_and_fetch_1:
3790 case Builtin::BI__sync_sub_and_fetch_2:
3791 case Builtin::BI__sync_sub_and_fetch_4:
3792 case Builtin::BI__sync_sub_and_fetch_8:
3793 case Builtin::BI__sync_sub_and_fetch_16:
3794 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Sub, E,
3795 llvm::Instruction::Sub);
3796 case Builtin::BI__sync_and_and_fetch_1:
3797 case Builtin::BI__sync_and_and_fetch_2:
3798 case Builtin::BI__sync_and_and_fetch_4:
3799 case Builtin::BI__sync_and_and_fetch_8:
3800 case Builtin::BI__sync_and_and_fetch_16:
3801 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::And, E,
3802 llvm::Instruction::And);
3803 case Builtin::BI__sync_or_and_fetch_1:
3804 case Builtin::BI__sync_or_and_fetch_2:
3805 case Builtin::BI__sync_or_and_fetch_4:
3806 case Builtin::BI__sync_or_and_fetch_8:
3807 case Builtin::BI__sync_or_and_fetch_16:
3808 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Or, E,
3809 llvm::Instruction::Or);
3810 case Builtin::BI__sync_xor_and_fetch_1:
3811 case Builtin::BI__sync_xor_and_fetch_2:
3812 case Builtin::BI__sync_xor_and_fetch_4:
3813 case Builtin::BI__sync_xor_and_fetch_8:
3814 case Builtin::BI__sync_xor_and_fetch_16:
3815 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Xor, E,
3816 llvm::Instruction::Xor);
3817 case Builtin::BI__sync_nand_and_fetch_1:
3818 case Builtin::BI__sync_nand_and_fetch_2:
3819 case Builtin::BI__sync_nand_and_fetch_4:
3820 case Builtin::BI__sync_nand_and_fetch_8:
3821 case Builtin::BI__sync_nand_and_fetch_16:
3822 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Nand, E,
3823 llvm::Instruction::And, true);
3824
3825 case Builtin::BI__sync_val_compare_and_swap_1:
3826 case Builtin::BI__sync_val_compare_and_swap_2:
3827 case Builtin::BI__sync_val_compare_and_swap_4:
3828 case Builtin::BI__sync_val_compare_and_swap_8:
3829 case Builtin::BI__sync_val_compare_and_swap_16:
3830 return RValue::get(MakeAtomicCmpXchgValue(*this, E, false));
3831
3832 case Builtin::BI__sync_bool_compare_and_swap_1:
3833 case Builtin::BI__sync_bool_compare_and_swap_2:
3834 case Builtin::BI__sync_bool_compare_and_swap_4:
3835 case Builtin::BI__sync_bool_compare_and_swap_8:
3836 case Builtin::BI__sync_bool_compare_and_swap_16:
3837 return RValue::get(MakeAtomicCmpXchgValue(*this, E, true));
3838
3839 case Builtin::BI__sync_swap_1:
3840 case Builtin::BI__sync_swap_2:
3841 case Builtin::BI__sync_swap_4:
3842 case Builtin::BI__sync_swap_8:
3843 case Builtin::BI__sync_swap_16:
3844 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xchg, E);
3845
3846 case Builtin::BI__sync_lock_test_and_set_1:
3847 case Builtin::BI__sync_lock_test_and_set_2:
3848 case Builtin::BI__sync_lock_test_and_set_4:
3849 case Builtin::BI__sync_lock_test_and_set_8:
3850 case Builtin::BI__sync_lock_test_and_set_16:
3851 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xchg, E);
3852
3853 case Builtin::BI__sync_lock_release_1:
3854 case Builtin::BI__sync_lock_release_2:
3855 case Builtin::BI__sync_lock_release_4:
3856 case Builtin::BI__sync_lock_release_8:
3857 case Builtin::BI__sync_lock_release_16: {
3858 Value *Ptr = EmitScalarExpr(E->getArg(0));
3859 QualType ElTy = E->getArg(0)->getType()->getPointeeType();
3860 CharUnits StoreSize = getContext().getTypeSizeInChars(ElTy);
3861 llvm::Type *ITy = llvm::IntegerType::get(getLLVMContext(),
3862 StoreSize.getQuantity() * 8);
3863 Ptr = Builder.CreateBitCast(Ptr, ITy->getPointerTo());
3864 llvm::StoreInst *Store =
3865 Builder.CreateAlignedStore(llvm::Constant::getNullValue(ITy), Ptr,
3866 StoreSize);
3867 Store->setAtomic(llvm::AtomicOrdering::Release);
3868 return RValue::get(nullptr);
3869 }
3870
3871 case Builtin::BI__sync_synchronize: {
3872 // We assume this is supposed to correspond to a C++0x-style
3873 // sequentially-consistent fence (i.e. this is only usable for
3874 // synchronization, not device I/O or anything like that). This intrinsic
3875 // is really badly designed in the sense that in theory, there isn't
3876 // any way to safely use it... but in practice, it mostly works
3877 // to use it with non-atomic loads and stores to get acquire/release
3878 // semantics.
3879 Builder.CreateFence(llvm::AtomicOrdering::SequentiallyConsistent);
3880 return RValue::get(nullptr);
3881 }
3882
3883 case Builtin::BI__builtin_nontemporal_load:
3884 return RValue::get(EmitNontemporalLoad(*this, E));
3885 case Builtin::BI__builtin_nontemporal_store:
3886 return RValue::get(EmitNontemporalStore(*this, E));
3887 case Builtin::BI__c11_atomic_is_lock_free:
3888 case Builtin::BI__atomic_is_lock_free: {
3889 // Call "bool __atomic_is_lock_free(size_t size, void *ptr)". For the
3890 // __c11 builtin, ptr is 0 (indicating a properly-aligned object), since
3891 // _Atomic(T) is always properly-aligned.
3892 const char *LibCallName = "__atomic_is_lock_free";
3893 CallArgList Args;
3894 Args.add(RValue::get(EmitScalarExpr(E->getArg(0))),
3895 getContext().getSizeType());
3896 if (BuiltinID == Builtin::BI__atomic_is_lock_free)
3897 Args.add(RValue::get(EmitScalarExpr(E->getArg(1))),
3898 getContext().VoidPtrTy);
3899 else
3900 Args.add(RValue::get(llvm::Constant::getNullValue(VoidPtrTy)),
3901 getContext().VoidPtrTy);
3902 const CGFunctionInfo &FuncInfo =
3903 CGM.getTypes().arrangeBuiltinFunctionCall(E->getType(), Args);
3904 llvm::FunctionType *FTy = CGM.getTypes().GetFunctionType(FuncInfo);
3905 llvm::FunctionCallee Func = CGM.CreateRuntimeFunction(FTy, LibCallName);
3906 return EmitCall(FuncInfo, CGCallee::forDirect(Func),
3907 ReturnValueSlot(), Args);
3908 }
3909
3910 case Builtin::BI__atomic_test_and_set: {
3911 // Look at the argument type to determine whether this is a volatile
3912 // operation. The parameter type is always volatile.
3913 QualType PtrTy = E->getArg(0)->IgnoreImpCasts()->getType();
3914 bool Volatile =
3915 PtrTy->castAs<PointerType>()->getPointeeType().isVolatileQualified();
3916
3917 Value *Ptr = EmitScalarExpr(E->getArg(0));
3918 unsigned AddrSpace = Ptr->getType()->getPointerAddressSpace();
3919 Ptr = Builder.CreateBitCast(Ptr, Int8Ty->getPointerTo(AddrSpace));
3920 Value *NewVal = Builder.getInt8(1);
3921 Value *Order = EmitScalarExpr(E->getArg(1));
3922 if (isa<llvm::ConstantInt>(Order)) {
3923 int ord = cast<llvm::ConstantInt>(Order)->getZExtValue();
3924 AtomicRMWInst *Result = nullptr;
3925 switch (ord) {
3926 case 0: // memory_order_relaxed
3927 default: // invalid order
3928 Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, Ptr, NewVal,
3929 llvm::AtomicOrdering::Monotonic);
3930 break;
3931 case 1: // memory_order_consume
3932 case 2: // memory_order_acquire
3933 Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, Ptr, NewVal,
3934 llvm::AtomicOrdering::Acquire);
3935 break;
3936 case 3: // memory_order_release
3937 Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, Ptr, NewVal,
3938 llvm::AtomicOrdering::Release);
3939 break;
3940 case 4: // memory_order_acq_rel
3941
3942 Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, Ptr, NewVal,
3943 llvm::AtomicOrdering::AcquireRelease);
3944 break;
3945 case 5: // memory_order_seq_cst
3946 Result = Builder.CreateAtomicRMW(
3947 llvm::AtomicRMWInst::Xchg, Ptr, NewVal,
3948 llvm::AtomicOrdering::SequentiallyConsistent);
3949 break;
3950 }
3951 Result->setVolatile(Volatile);
3952 return RValue::get(Builder.CreateIsNotNull(Result, "tobool"));
3953 }
3954
3955 llvm::BasicBlock *ContBB = createBasicBlock("atomic.continue", CurFn);
3956
3957 llvm::BasicBlock *BBs[5] = {
3958 createBasicBlock("monotonic", CurFn),
3959 createBasicBlock("acquire", CurFn),
3960 createBasicBlock("release", CurFn),
3961 createBasicBlock("acqrel", CurFn),
3962 createBasicBlock("seqcst", CurFn)
3963 };
3964 llvm::AtomicOrdering Orders[5] = {
3965 llvm::AtomicOrdering::Monotonic, llvm::AtomicOrdering::Acquire,
3966 llvm::AtomicOrdering::Release, llvm::AtomicOrdering::AcquireRelease,
3967 llvm::AtomicOrdering::SequentiallyConsistent};
3968
3969 Order = Builder.CreateIntCast(Order, Builder.getInt32Ty(), false);
3970 llvm::SwitchInst *SI = Builder.CreateSwitch(Order, BBs[0]);
3971
3972 Builder.SetInsertPoint(ContBB);
3973 PHINode *Result = Builder.CreatePHI(Int8Ty, 5, "was_set");
3974
3975 for (unsigned i = 0; i < 5; ++i) {
3976 Builder.SetInsertPoint(BBs[i]);
3977 AtomicRMWInst *RMW = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg,
3978 Ptr, NewVal, Orders[i]);
3979 RMW->setVolatile(Volatile);
3980 Result->addIncoming(RMW, BBs[i]);
3981 Builder.CreateBr(ContBB);
3982 }
3983
3984 SI->addCase(Builder.getInt32(0), BBs[0]);
3985 SI->addCase(Builder.getInt32(1), BBs[1]);
3986 SI->addCase(Builder.getInt32(2), BBs[1]);
3987 SI->addCase(Builder.getInt32(3), BBs[2]);
3988 SI->addCase(Builder.getInt32(4), BBs[3]);
3989 SI->addCase(Builder.getInt32(5), BBs[4]);
3990
3991 Builder.SetInsertPoint(ContBB);
3992 return RValue::get(Builder.CreateIsNotNull(Result, "tobool"));
3993 }
3994
3995 case Builtin::BI__atomic_clear: {
3996 QualType PtrTy = E->getArg(0)->IgnoreImpCasts()->getType();
3997 bool Volatile =
3998 PtrTy->castAs<PointerType>()->getPointeeType().isVolatileQualified();
3999
4000 Address Ptr = EmitPointerWithAlignment(E->getArg(0));
4001 unsigned AddrSpace = Ptr.getPointer()->getType()->getPointerAddressSpace();
4002 Ptr = Builder.CreateBitCast(Ptr, Int8Ty->getPointerTo(AddrSpace));
4003 Value *NewVal = Builder.getInt8(0);
4004 Value *Order = EmitScalarExpr(E->getArg(1));
4005 if (isa<llvm::ConstantInt>(Order)) {
4006 int ord = cast<llvm::ConstantInt>(Order)->getZExtValue();
4007 StoreInst *Store = Builder.CreateStore(NewVal, Ptr, Volatile);
4008 switch (ord) {
4009 case 0: // memory_order_relaxed
4010 default: // invalid order
4011 Store->setOrdering(llvm::AtomicOrdering::Monotonic);
4012 break;
4013 case 3: // memory_order_release
4014 Store->setOrdering(llvm::AtomicOrdering::Release);
4015 break;
4016 case 5: // memory_order_seq_cst
4017 Store->setOrdering(llvm::AtomicOrdering::SequentiallyConsistent);
4018 break;
4019 }
4020 return RValue::get(nullptr);
4021 }
4022
4023 llvm::BasicBlock *ContBB = createBasicBlock("atomic.continue", CurFn);
4024
4025 llvm::BasicBlock *BBs[3] = {
4026 createBasicBlock("monotonic", CurFn),
4027 createBasicBlock("release", CurFn),
4028 createBasicBlock("seqcst", CurFn)
4029 };
4030 llvm::AtomicOrdering Orders[3] = {
4031 llvm::AtomicOrdering::Monotonic, llvm::AtomicOrdering::Release,
4032 llvm::AtomicOrdering::SequentiallyConsistent};
4033
4034 Order = Builder.CreateIntCast(Order, Builder.getInt32Ty(), false);
4035 llvm::SwitchInst *SI = Builder.CreateSwitch(Order, BBs[0]);
4036
4037 for (unsigned i = 0; i < 3; ++i) {
4038 Builder.SetInsertPoint(BBs[i]);
4039 StoreInst *Store = Builder.CreateStore(NewVal, Ptr, Volatile);
4040 Store->setOrdering(Orders[i]);
4041 Builder.CreateBr(ContBB);
4042 }
4043
4044 SI->addCase(Builder.getInt32(0), BBs[0]);
4045 SI->addCase(Builder.getInt32(3), BBs[1]);
4046 SI->addCase(Builder.getInt32(5), BBs[2]);
4047
4048 Builder.SetInsertPoint(ContBB);
4049 return RValue::get(nullptr);
4050 }
4051
4052 case Builtin::BI__atomic_thread_fence:
4053 case Builtin::BI__atomic_signal_fence:
4054 case Builtin::BI__c11_atomic_thread_fence:
4055 case Builtin::BI__c11_atomic_signal_fence: {
4056 llvm::SyncScope::ID SSID;
4057 if (BuiltinID == Builtin::BI__atomic_signal_fence ||
4058 BuiltinID == Builtin::BI__c11_atomic_signal_fence)
4059 SSID = llvm::SyncScope::SingleThread;
4060 else
4061 SSID = llvm::SyncScope::System;
4062 Value *Order = EmitScalarExpr(E->getArg(0));
4063 if (isa<llvm::ConstantInt>(Order)) {
4064 int ord = cast<llvm::ConstantInt>(Order)->getZExtValue();
4065 switch (ord) {
4066 case 0: // memory_order_relaxed
4067 default: // invalid order
4068 break;
4069 case 1: // memory_order_consume
4070 case 2: // memory_order_acquire
4071 Builder.CreateFence(llvm::AtomicOrdering::Acquire, SSID);
4072 break;
4073 case 3: // memory_order_release
4074 Builder.CreateFence(llvm::AtomicOrdering::Release, SSID);
4075 break;
4076 case 4: // memory_order_acq_rel
4077 Builder.CreateFence(llvm::AtomicOrdering::AcquireRelease, SSID);
4078 break;
4079 case 5: // memory_order_seq_cst
4080 Builder.CreateFence(llvm::AtomicOrdering::SequentiallyConsistent, SSID);
4081 break;
4082 }
4083 return RValue::get(nullptr);
4084 }
4085
4086 llvm::BasicBlock *AcquireBB, *ReleaseBB, *AcqRelBB, *SeqCstBB;
4087 AcquireBB = createBasicBlock("acquire", CurFn);
4088 ReleaseBB = createBasicBlock("release", CurFn);
4089 AcqRelBB = createBasicBlock("acqrel", CurFn);
4090 SeqCstBB = createBasicBlock("seqcst", CurFn);
4091 llvm::BasicBlock *ContBB = createBasicBlock("atomic.continue", CurFn);
4092
4093 Order = Builder.CreateIntCast(Order, Builder.getInt32Ty(), false);
4094 llvm::SwitchInst *SI = Builder.CreateSwitch(Order, ContBB);
4095
4096 Builder.SetInsertPoint(AcquireBB);
4097 Builder.CreateFence(llvm::AtomicOrdering::Acquire, SSID);
4098 Builder.CreateBr(ContBB);
4099 SI->addCase(Builder.getInt32(1), AcquireBB);
4100 SI->addCase(Builder.getInt32(2), AcquireBB);
4101
4102 Builder.SetInsertPoint(ReleaseBB);
4103 Builder.CreateFence(llvm::AtomicOrdering::Release, SSID);
4104 Builder.CreateBr(ContBB);
4105 SI->addCase(Builder.getInt32(3), ReleaseBB);
4106
4107 Builder.SetInsertPoint(AcqRelBB);
4108 Builder.CreateFence(llvm::AtomicOrdering::AcquireRelease, SSID);
4109 Builder.CreateBr(ContBB);
4110 SI->addCase(Builder.getInt32(4), AcqRelBB);
4111
4112 Builder.SetInsertPoint(SeqCstBB);
4113 Builder.CreateFence(llvm::AtomicOrdering::SequentiallyConsistent, SSID);
4114 Builder.CreateBr(ContBB);
4115 SI->addCase(Builder.getInt32(5), SeqCstBB);
4116
4117 Builder.SetInsertPoint(ContBB);
4118 return RValue::get(nullptr);
4119 }
4120
4121 case Builtin::BI__builtin_signbit:
4122 case Builtin::BI__builtin_signbitf:
4123 case Builtin::BI__builtin_signbitl: {
4124 return RValue::get(
4125 Builder.CreateZExt(EmitSignBit(*this, EmitScalarExpr(E->getArg(0))),
4126 ConvertType(E->getType())));
4127 }
4128 case Builtin::BI__warn_memset_zero_len:
4129 return RValue::getIgnored();
4130 case Builtin::BI__annotation: {
4131 // Re-encode each wide string to UTF8 and make an MDString.
4132 SmallVector<Metadata *, 1> Strings;
4133 for (const Expr *Arg : E->arguments()) {
4134 const auto *Str = cast<StringLiteral>(Arg->IgnoreParenCasts());
4135 assert(Str->getCharByteWidth() == 2)((void)0);
4136 StringRef WideBytes = Str->getBytes();
4137 std::string StrUtf8;
4138 if (!convertUTF16ToUTF8String(
4139 makeArrayRef(WideBytes.data(), WideBytes.size()), StrUtf8)) {
4140 CGM.ErrorUnsupported(E, "non-UTF16 __annotation argument");
4141 continue;
4142 }
4143 Strings.push_back(llvm::MDString::get(getLLVMContext(), StrUtf8));
4144 }
4145
4146 // Build and MDTuple of MDStrings and emit the intrinsic call.
4147 llvm::Function *F =
4148 CGM.getIntrinsic(llvm::Intrinsic::codeview_annotation, {});
4149 MDTuple *StrTuple = MDTuple::get(getLLVMContext(), Strings);
4150 Builder.CreateCall(F, MetadataAsValue::get(getLLVMContext(), StrTuple));
4151 return RValue::getIgnored();
4152 }
4153 case Builtin::BI__builtin_annotation: {
4154 llvm::Value *AnnVal = EmitScalarExpr(E->getArg(0));
4155 llvm::Function *F = CGM.getIntrinsic(llvm::Intrinsic::annotation,
4156 AnnVal->getType());
4157
4158 // Get the annotation string, go through casts. Sema requires this to be a
4159 // non-wide string literal, potentially casted, so the cast<> is safe.
4160 const Expr *AnnotationStrExpr = E->getArg(1)->IgnoreParenCasts();
4161 StringRef Str = cast<StringLiteral>(AnnotationStrExpr)->getString();
4162 return RValue::get(
4163 EmitAnnotationCall(F, AnnVal, Str, E->getExprLoc(), nullptr));
4164 }
4165 case Builtin::BI__builtin_addcb:
4166 case Builtin::BI__builtin_addcs:
4167 case Builtin::BI__builtin_addc:
4168 case Builtin::BI__builtin_addcl:
4169 case Builtin::BI__builtin_addcll:
4170 case Builtin::BI__builtin_subcb:
4171 case Builtin::BI__builtin_subcs:
4172 case Builtin::BI__builtin_subc:
4173 case Builtin::BI__builtin_subcl:
4174 case Builtin::BI__builtin_subcll: {
4175
4176 // We translate all of these builtins from expressions of the form:
4177 // int x = ..., y = ..., carryin = ..., carryout, result;
4178 // result = __builtin_addc(x, y, carryin, &carryout);
4179 //
4180 // to LLVM IR of the form:
4181 //
4182 // %tmp1 = call {i32, i1} @llvm.uadd.with.overflow.i32(i32 %x, i32 %y)
4183 // %tmpsum1 = extractvalue {i32, i1} %tmp1, 0
4184 // %carry1 = extractvalue {i32, i1} %tmp1, 1
4185 // %tmp2 = call {i32, i1} @llvm.uadd.with.overflow.i32(i32 %tmpsum1,
4186 // i32 %carryin)
4187 // %result = extractvalue {i32, i1} %tmp2, 0
4188 // %carry2 = extractvalue {i32, i1} %tmp2, 1
4189 // %tmp3 = or i1 %carry1, %carry2
4190 // %tmp4 = zext i1 %tmp3 to i32
4191 // store i32 %tmp4, i32* %carryout
4192
4193 // Scalarize our inputs.
4194 llvm::Value *X = EmitScalarExpr(E->getArg(0));
4195 llvm::Value *Y = EmitScalarExpr(E->getArg(1));
4196 llvm::Value *Carryin = EmitScalarExpr(E->getArg(2));
4197 Address CarryOutPtr = EmitPointerWithAlignment(E->getArg(3));
4198
4199 // Decide if we are lowering to a uadd.with.overflow or usub.with.overflow.
4200 llvm::Intrinsic::ID IntrinsicId;
4201 switch (BuiltinID) {
4202 default: llvm_unreachable("Unknown multiprecision builtin id.")__builtin_unreachable();
4203 case Builtin::BI__builtin_addcb:
4204 case Builtin::BI__builtin_addcs:
4205 case Builtin::BI__builtin_addc:
4206 case Builtin::BI__builtin_addcl:
4207 case Builtin::BI__builtin_addcll:
4208 IntrinsicId = llvm::Intrinsic::uadd_with_overflow;
4209 break;
4210 case Builtin::BI__builtin_subcb:
4211 case Builtin::BI__builtin_subcs:
4212 case Builtin::BI__builtin_subc:
4213 case Builtin::BI__builtin_subcl:
4214 case Builtin::BI__builtin_subcll:
4215 IntrinsicId = llvm::Intrinsic::usub_with_overflow;
4216 break;
4217 }
4218
4219 // Construct our resulting LLVM IR expression.
4220 llvm::Value *Carry1;
4221 llvm::Value *Sum1 = EmitOverflowIntrinsic(*this, IntrinsicId,
4222 X, Y, Carry1);
4223 llvm::Value *Carry2;
4224 llvm::Value *Sum2 = EmitOverflowIntrinsic(*this, IntrinsicId,
4225 Sum1, Carryin, Carry2);
4226 llvm::Value *CarryOut = Builder.CreateZExt(Builder.CreateOr(Carry1, Carry2),
4227 X->getType());
4228 Builder.CreateStore(CarryOut, CarryOutPtr);
4229 return RValue::get(Sum2);
4230 }
4231
4232 case Builtin::BI__builtin_add_overflow:
4233 case Builtin::BI__builtin_sub_overflow:
4234 case Builtin::BI__builtin_mul_overflow: {
4235 const clang::Expr *LeftArg = E->getArg(0);
4236 const clang::Expr *RightArg = E->getArg(1);
4237 const clang::Expr *ResultArg = E->getArg(2);
4238
4239 clang::QualType ResultQTy =
4240 ResultArg->getType()->castAs<PointerType>()->getPointeeType();
4241
4242 WidthAndSignedness LeftInfo =
4243 getIntegerWidthAndSignedness(CGM.getContext(), LeftArg->getType());
4244 WidthAndSignedness RightInfo =
4245 getIntegerWidthAndSignedness(CGM.getContext(), RightArg->getType());
4246 WidthAndSignedness ResultInfo =
4247 getIntegerWidthAndSignedness(CGM.getContext(), ResultQTy);
4248
4249 // Handle mixed-sign multiplication as a special case, because adding
4250 // runtime or backend support for our generic irgen would be too expensive.
4251 if (isSpecialMixedSignMultiply(BuiltinID, LeftInfo, RightInfo, ResultInfo))
4252 return EmitCheckedMixedSignMultiply(*this, LeftArg, LeftInfo, RightArg,
4253 RightInfo, ResultArg, ResultQTy,
4254 ResultInfo);
4255
4256 if (isSpecialUnsignedMultiplySignedResult(BuiltinID, LeftInfo, RightInfo,
4257 ResultInfo))
4258 return EmitCheckedUnsignedMultiplySignedResult(
4259 *this, LeftArg, LeftInfo, RightArg, RightInfo, ResultArg, ResultQTy,
4260 ResultInfo);
4261
4262 WidthAndSignedness EncompassingInfo =
4263 EncompassingIntegerType({LeftInfo, RightInfo, ResultInfo});
4264
4265 llvm::Type *EncompassingLLVMTy =
4266 llvm::IntegerType::get(CGM.getLLVMContext(), EncompassingInfo.Width);
4267
4268 llvm::Type *ResultLLVMTy = CGM.getTypes().ConvertType(ResultQTy);
4269
4270 llvm::Intrinsic::ID IntrinsicId;
4271 switch (BuiltinID) {
4272 default:
4273 llvm_unreachable("Unknown overflow builtin id.")__builtin_unreachable();
4274 case Builtin::BI__builtin_add_overflow:
4275 IntrinsicId = EncompassingInfo.Signed
4276 ? llvm::Intrinsic::sadd_with_overflow
4277 : llvm::Intrinsic::uadd_with_overflow;
4278 break;
4279 case Builtin::BI__builtin_sub_overflow:
4280 IntrinsicId = EncompassingInfo.Signed
4281 ? llvm::Intrinsic::ssub_with_overflow
4282 : llvm::Intrinsic::usub_with_overflow;
4283 break;
4284 case Builtin::BI__builtin_mul_overflow:
4285 IntrinsicId = EncompassingInfo.Signed
4286 ? llvm::Intrinsic::smul_with_overflow
4287 : llvm::Intrinsic::umul_with_overflow;
4288 break;
4289 }
4290
4291 llvm::Value *Left = EmitScalarExpr(LeftArg);
4292 llvm::Value *Right = EmitScalarExpr(RightArg);
4293 Address ResultPtr = EmitPointerWithAlignment(ResultArg);
4294
4295 // Extend each operand to the encompassing type.
4296 Left = Builder.CreateIntCast(Left, EncompassingLLVMTy, LeftInfo.Signed);
4297 Right = Builder.CreateIntCast(Right, EncompassingLLVMTy, RightInfo.Signed);
4298
4299 // Perform the operation on the extended values.
4300 llvm::Value *Overflow, *Result;
4301 Result = EmitOverflowIntrinsic(*this, IntrinsicId, Left, Right, Overflow);
4302
4303 if (EncompassingInfo.Width > ResultInfo.Width) {
4304 // The encompassing type is wider than the result type, so we need to
4305 // truncate it.
4306 llvm::Value *ResultTrunc = Builder.CreateTrunc(Result, ResultLLVMTy);
4307
4308 // To see if the truncation caused an overflow, we will extend
4309 // the result and then compare it to the original result.
4310 llvm::Value *ResultTruncExt = Builder.CreateIntCast(
4311 ResultTrunc, EncompassingLLVMTy, ResultInfo.Signed);
4312 llvm::Value *TruncationOverflow =
4313 Builder.CreateICmpNE(Result, ResultTruncExt);
4314
4315 Overflow = Builder.CreateOr(Overflow, TruncationOverflow);
4316 Result = ResultTrunc;
4317 }
4318
4319 // Finally, store the result using the pointer.
4320 bool isVolatile =
4321 ResultArg->getType()->getPointeeType().isVolatileQualified();
4322 Builder.CreateStore(EmitToMemory(Result, ResultQTy), ResultPtr, isVolatile);
4323
4324 return RValue::get(Overflow);
4325 }
4326
4327 case Builtin::BI__builtin_uadd_overflow:
4328 case Builtin::BI__builtin_uaddl_overflow:
4329 case Builtin::BI__builtin_uaddll_overflow:
4330 case Builtin::BI__builtin_usub_overflow:
4331 case Builtin::BI__builtin_usubl_overflow:
4332 case Builtin::BI__builtin_usubll_overflow:
4333 case Builtin::BI__builtin_umul_overflow:
4334 case Builtin::BI__builtin_umull_overflow:
4335 case Builtin::BI__builtin_umulll_overflow:
4336 case Builtin::BI__builtin_sadd_overflow:
4337 case Builtin::BI__builtin_saddl_overflow:
4338 case Builtin::BI__builtin_saddll_overflow:
4339 case Builtin::BI__builtin_ssub_overflow:
4340 case Builtin::BI__builtin_ssubl_overflow:
4341 case Builtin::BI__builtin_ssubll_overflow:
4342 case Builtin::BI__builtin_smul_overflow:
4343 case Builtin::BI__builtin_smull_overflow:
4344 case Builtin::BI__builtin_smulll_overflow: {
4345
4346 // We translate all of these builtins directly to the relevant llvm IR node.
4347
4348 // Scalarize our inputs.
4349 llvm::Value *X = EmitScalarExpr(E->getArg(0));
4350 llvm::Value *Y = EmitScalarExpr(E->getArg(1));
4351 Address SumOutPtr = EmitPointerWithAlignment(E->getArg(2));
4352
4353 // Decide which of the overflow intrinsics we are lowering to:
4354 llvm::Intrinsic::ID IntrinsicId;
4355 switch (BuiltinID) {
4356 default: llvm_unreachable("Unknown overflow builtin id.")__builtin_unreachable();
4357 case Builtin::BI__builtin_uadd_overflow:
4358 case Builtin::BI__builtin_uaddl_overflow:
4359 case Builtin::BI__builtin_uaddll_overflow:
4360 IntrinsicId = llvm::Intrinsic::uadd_with_overflow;
4361 break;
4362 case Builtin::BI__builtin_usub_overflow:
4363 case Builtin::BI__builtin_usubl_overflow:
4364 case Builtin::BI__builtin_usubll_overflow:
4365 IntrinsicId = llvm::Intrinsic::usub_with_overflow;
4366 break;
4367 case Builtin::BI__builtin_umul_overflow:
4368 case Builtin::BI__builtin_umull_overflow:
4369 case Builtin::BI__builtin_umulll_overflow:
4370 IntrinsicId = llvm::Intrinsic::umul_with_overflow;
4371 break;
4372 case Builtin::BI__builtin_sadd_overflow:
4373 case Builtin::BI__builtin_saddl_overflow:
4374 case Builtin::BI__builtin_saddll_overflow:
4375 IntrinsicId = llvm::Intrinsic::sadd_with_overflow;
4376 break;
4377 case Builtin::BI__builtin_ssub_overflow:
4378 case Builtin::BI__builtin_ssubl_overflow:
4379 case Builtin::BI__builtin_ssubll_overflow:
4380 IntrinsicId = llvm::Intrinsic::ssub_with_overflow;
4381 break;
4382 case Builtin::BI__builtin_smul_overflow:
4383 case Builtin::BI__builtin_smull_overflow:
4384 case Builtin::BI__builtin_smulll_overflow:
4385 IntrinsicId = llvm::Intrinsic::smul_with_overflow;
4386 break;
4387 }
4388
4389
4390 llvm::Value *Carry;
4391 llvm::Value *Sum = EmitOverflowIntrinsic(*this, IntrinsicId, X, Y, Carry);
4392 Builder.CreateStore(Sum, SumOutPtr);
4393
4394 return RValue::get(Carry);
4395 }
4396 case Builtin::BI__builtin_addressof:
4397 return RValue::get(EmitLValue(E->getArg(0)).getPointer(*this));
4398 case Builtin::BI__builtin_operator_new:
4399 return EmitBuiltinNewDeleteCall(
4400 E->getCallee()->getType()->castAs<FunctionProtoType>(), E, false);
4401 case Builtin::BI__builtin_operator_delete:
4402 return EmitBuiltinNewDeleteCall(
4403 E->getCallee()->getType()->castAs<FunctionProtoType>(), E, true);
4404
4405 case Builtin::BI__builtin_is_aligned:
4406 return EmitBuiltinIsAligned(E);
4407 case Builtin::BI__builtin_align_up:
4408 return EmitBuiltinAlignTo(E, true);
4409 case Builtin::BI__builtin_align_down:
4410 return EmitBuiltinAlignTo(E, false);
4411
4412 case Builtin::BI__noop:
4413 // __noop always evaluates to an integer literal zero.
4414 return RValue::get(ConstantInt::get(IntTy, 0));
4415 case Builtin::BI__builtin_call_with_static_chain: {
4416 const CallExpr *Call = cast<CallExpr>(E->getArg(0));
4417 const Expr *Chain = E->getArg(1);
4418 return EmitCall(Call->getCallee()->getType(),
4419 EmitCallee(Call->getCallee()), Call, ReturnValue,
4420 EmitScalarExpr(Chain));
4421 }
4422 case Builtin::BI_InterlockedExchange8:
4423 case Builtin::BI_InterlockedExchange16:
4424 case Builtin::BI_InterlockedExchange:
4425 case Builtin::BI_InterlockedExchangePointer:
4426 return RValue::get(
4427 EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedExchange, E));
4428 case Builtin::BI_InterlockedCompareExchangePointer:
4429 case Builtin::BI_InterlockedCompareExchangePointer_nf: {
4430 llvm::Type *RTy;
4431 llvm::IntegerType *IntType =
4432 IntegerType::get(getLLVMContext(),
4433 getContext().getTypeSize(E->getType()));
4434 llvm::Type *IntPtrType = IntType->getPointerTo();
4435
4436 llvm::Value *Destination =
4437 Builder.CreateBitCast(EmitScalarExpr(E->getArg(0)), IntPtrType);
4438
4439 llvm::Value *Exchange = EmitScalarExpr(E->getArg(1));
4440 RTy = Exchange->getType();
4441 Exchange = Builder.CreatePtrToInt(Exchange, IntType);
4442
4443 llvm::Value *Comparand =
4444 Builder.CreatePtrToInt(EmitScalarExpr(E->getArg(2)), IntType);
4445
4446 auto Ordering =
4447 BuiltinID == Builtin::BI_InterlockedCompareExchangePointer_nf ?
4448 AtomicOrdering::Monotonic : AtomicOrdering::SequentiallyConsistent;
4449
4450 auto Result = Builder.CreateAtomicCmpXchg(Destination, Comparand, Exchange,
4451 Ordering, Ordering);
4452 Result->setVolatile(true);
4453
4454 return RValue::get(Builder.CreateIntToPtr(Builder.CreateExtractValue(Result,
4455 0),
4456 RTy));
4457 }
4458 case Builtin::BI_InterlockedCompareExchange8:
4459 case Builtin::BI_InterlockedCompareExchange16:
4460 case Builtin::BI_InterlockedCompareExchange:
4461 case Builtin::BI_InterlockedCompareExchange64:
4462 return RValue::get(EmitAtomicCmpXchgForMSIntrin(*this, E));
4463 case Builtin::BI_InterlockedIncrement16:
4464 case Builtin::BI_InterlockedIncrement:
4465 return RValue::get(
4466 EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedIncrement, E));
4467 case Builtin::BI_InterlockedDecrement16:
4468 case Builtin::BI_InterlockedDecrement:
4469 return RValue::get(
4470 EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedDecrement, E));
4471 case Builtin::BI_InterlockedAnd8:
4472 case Builtin::BI_InterlockedAnd16:
4473 case Builtin::BI_InterlockedAnd:
4474 return RValue::get(EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedAnd, E));
4475 case Builtin::BI_InterlockedExchangeAdd8:
4476 case Builtin::BI_InterlockedExchangeAdd16:
4477 case Builtin::BI_InterlockedExchangeAdd:
4478 return RValue::get(
4479 EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedExchangeAdd, E));
4480 case Builtin::BI_InterlockedExchangeSub8:
4481 case Builtin::BI_InterlockedExchangeSub16:
4482 case Builtin::BI_InterlockedExchangeSub:
4483 return RValue::get(
4484 EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedExchangeSub, E));
4485 case Builtin::BI_InterlockedOr8:
4486 case Builtin::BI_InterlockedOr16:
4487 case Builtin::BI_InterlockedOr:
4488 return RValue::get(EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedOr, E));
4489 case Builtin::BI_InterlockedXor8:
4490 case Builtin::BI_InterlockedXor16:
4491 case Builtin::BI_InterlockedXor:
4492 return RValue::get(EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedXor, E));
4493
4494 case Builtin::BI_bittest64:
4495 case Builtin::BI_bittest:
4496 case Builtin::BI_bittestandcomplement64:
4497 case Builtin::BI_bittestandcomplement:
4498 case Builtin::BI_bittestandreset64:
4499 case Builtin::BI_bittestandreset:
4500 case Builtin::BI_bittestandset64:
4501 case Builtin::BI_bittestandset:
4502 case Builtin::BI_interlockedbittestandreset:
4503 case Builtin::BI_interlockedbittestandreset64:
4504 case Builtin::BI_interlockedbittestandset64:
4505 case Builtin::BI_interlockedbittestandset:
4506 case Builtin::BI_interlockedbittestandset_acq:
4507 case Builtin::BI_interlockedbittestandset_rel:
4508 case Builtin::BI_interlockedbittestandset_nf:
4509 case Builtin::BI_interlockedbittestandreset_acq:
4510 case Builtin::BI_interlockedbittestandreset_rel:
4511 case Builtin::BI_interlockedbittestandreset_nf:
4512 return RValue::get(EmitBitTestIntrinsic(*this, BuiltinID, E));
4513
4514 // These builtins exist to emit regular volatile loads and stores not
4515 // affected by the -fms-volatile setting.
4516 case Builtin::BI__iso_volatile_load8:
4517 case Builtin::BI__iso_volatile_load16:
4518 case Builtin::BI__iso_volatile_load32:
4519 case Builtin::BI__iso_volatile_load64:
4520 return RValue::get(EmitISOVolatileLoad(*this, E));
4521 case Builtin::BI__iso_volatile_store8:
4522 case Builtin::BI__iso_volatile_store16:
4523 case Builtin::BI__iso_volatile_store32:
4524 case Builtin::BI__iso_volatile_store64:
4525 return RValue::get(EmitISOVolatileStore(*this, E));
4526
4527 case Builtin::BI__exception_code:
4528 case Builtin::BI_exception_code:
4529 return RValue::get(EmitSEHExceptionCode());
4530 case Builtin::BI__exception_info:
4531 case Builtin::BI_exception_info:
4532 return RValue::get(EmitSEHExceptionInfo());
4533 case Builtin::BI__abnormal_termination:
4534 case Builtin::BI_abnormal_termination:
4535 return RValue::get(EmitSEHAbnormalTermination());
4536 case Builtin::BI_setjmpex:
4537 if (getTarget().getTriple().isOSMSVCRT() && E->getNumArgs() == 1 &&
4538 E->getArg(0)->getType()->isPointerType())
4539 return EmitMSVCRTSetJmp(*this, MSVCSetJmpKind::_setjmpex, E);
4540 break;
4541 case Builtin::BI_setjmp:
4542 if (getTarget().getTriple().isOSMSVCRT() && E->getNumArgs() == 1 &&
4543 E->getArg(0)->getType()->isPointerType()) {
4544 if (getTarget().getTriple().getArch() == llvm::Triple::x86)
4545 return EmitMSVCRTSetJmp(*this, MSVCSetJmpKind::_setjmp3, E);
4546 else if (getTarget().getTriple().getArch() == llvm::Triple::aarch64)
4547 return EmitMSVCRTSetJmp(*this, MSVCSetJmpKind::_setjmpex, E);
4548 return EmitMSVCRTSetJmp(*this, MSVCSetJmpKind::_setjmp, E);
4549 }
4550 break;
4551
4552 case Builtin::BI__GetExceptionInfo: {
4553 if (llvm::GlobalVariable *GV =
4554 CGM.getCXXABI().getThrowInfo(FD->getParamDecl(0)->getType()))
4555 return RValue::get(llvm::ConstantExpr::getBitCast(GV, CGM.Int8PtrTy));
4556 break;
4557 }
4558
4559 case Builtin::BI__fastfail:
4560 return RValue::get(EmitMSVCBuiltinExpr(MSVCIntrin::__fastfail, E));
4561
4562 case Builtin::BI__builtin_coro_size: {
4563 auto & Context = getContext();
4564 auto SizeTy = Context.getSizeType();
4565 auto T = Builder.getIntNTy(Context.getTypeSize(SizeTy));
4566 Function *F = CGM.getIntrinsic(Intrinsic::coro_size, T);
4567 return RValue::get(Builder.CreateCall(F));
4568 }
4569
4570 case Builtin::BI__builtin_coro_id:
4571 return EmitCoroutineIntrinsic(E, Intrinsic::coro_id);
4572 case Builtin::BI__builtin_coro_promise:
4573 return EmitCoroutineIntrinsic(E, Intrinsic::coro_promise);
4574 case Builtin::BI__builtin_coro_resume:
4575 return EmitCoroutineIntrinsic(E, Intrinsic::coro_resume);
4576 case Builtin::BI__builtin_coro_frame:
4577 return EmitCoroutineIntrinsic(E, Intrinsic::coro_frame);
4578 case Builtin::BI__builtin_coro_noop:
4579 return EmitCoroutineIntrinsic(E, Intrinsic::coro_noop);
4580 case Builtin::BI__builtin_coro_free:
4581 return EmitCoroutineIntrinsic(E, Intrinsic::coro_free);
4582 case Builtin::BI__builtin_coro_destroy:
4583 return EmitCoroutineIntrinsic(E, Intrinsic::coro_destroy);
4584 case Builtin::BI__builtin_coro_done:
4585 return EmitCoroutineIntrinsic(E, Intrinsic::coro_done);
4586 case Builtin::BI__builtin_coro_alloc:
4587 return EmitCoroutineIntrinsic(E, Intrinsic::coro_alloc);
4588 case Builtin::BI__builtin_coro_begin:
4589 return EmitCoroutineIntrinsic(E, Intrinsic::coro_begin);
4590 case Builtin::BI__builtin_coro_end:
4591 return EmitCoroutineIntrinsic(E, Intrinsic::coro_end);
4592 case Builtin::BI__builtin_coro_suspend:
4593 return EmitCoroutineIntrinsic(E, Intrinsic::coro_suspend);
4594 case Builtin::BI__builtin_coro_param:
4595 return EmitCoroutineIntrinsic(E, Intrinsic::coro_param);
4596
4597 // OpenCL v2.0 s6.13.16.2, Built-in pipe read and write functions
4598 case Builtin::BIread_pipe:
4599 case Builtin::BIwrite_pipe: {
4600 Value *Arg0 = EmitScalarExpr(E->getArg(0)),
4601 *Arg1 = EmitScalarExpr(E->getArg(1));
4602 CGOpenCLRuntime OpenCLRT(CGM);
4603 Value *PacketSize = OpenCLRT.getPipeElemSize(E->getArg(0));
4604 Value *PacketAlign = OpenCLRT.getPipeElemAlign(E->getArg(0));
4605
4606 // Type of the generic packet parameter.
4607 unsigned GenericAS =
4608 getContext().getTargetAddressSpace(LangAS::opencl_generic);
4609 llvm::Type *I8PTy = llvm::PointerType::get(
4610 llvm::Type::getInt8Ty(getLLVMContext()), GenericAS);
4611
4612 // Testing which overloaded version we should generate the call for.
4613 if (2U == E->getNumArgs()) {
4614 const char *Name = (BuiltinID == Builtin::BIread_pipe) ? "__read_pipe_2"
4615 : "__write_pipe_2";
4616 // Creating a generic function type to be able to call with any builtin or
4617 // user defined type.
4618 llvm::Type *ArgTys[] = {Arg0->getType(), I8PTy, Int32Ty, Int32Ty};
4619 llvm::FunctionType *FTy = llvm::FunctionType::get(
4620 Int32Ty, llvm::ArrayRef<llvm::Type *>(ArgTys), false);
4621 Value *BCast = Builder.CreatePointerCast(Arg1, I8PTy);
4622 return RValue::get(
4623 EmitRuntimeCall(CGM.CreateRuntimeFunction(FTy, Name),
4624 {Arg0, BCast, PacketSize, PacketAlign}));
4625 } else {
4626 assert(4 == E->getNumArgs() &&((void)0)
4627 "Illegal number of parameters to pipe function")((void)0);
4628 const char *Name = (BuiltinID == Builtin::BIread_pipe) ? "__read_pipe_4"
4629 : "__write_pipe_4";
4630
4631 llvm::Type *ArgTys[] = {Arg0->getType(), Arg1->getType(), Int32Ty, I8PTy,
4632 Int32Ty, Int32Ty};
4633 Value *Arg2 = EmitScalarExpr(E->getArg(2)),
4634 *Arg3 = EmitScalarExpr(E->getArg(3));
4635 llvm::FunctionType *FTy = llvm::FunctionType::get(
4636 Int32Ty, llvm::ArrayRef<llvm::Type *>(ArgTys), false);
4637 Value *BCast = Builder.CreatePointerCast(Arg3, I8PTy);
4638 // We know the third argument is an integer type, but we may need to cast
4639 // it to i32.
4640 if (Arg2->getType() != Int32Ty)
4641 Arg2 = Builder.CreateZExtOrTrunc(Arg2, Int32Ty);
4642 return RValue::get(
4643 EmitRuntimeCall(CGM.CreateRuntimeFunction(FTy, Name),
4644 {Arg0, Arg1, Arg2, BCast, PacketSize, PacketAlign}));
4645 }
4646 }
4647 // OpenCL v2.0 s6.13.16 ,s9.17.3.5 - Built-in pipe reserve read and write
4648 // functions
4649 case Builtin::BIreserve_read_pipe:
4650 case Builtin::BIreserve_write_pipe:
4651 case Builtin::BIwork_group_reserve_read_pipe:
4652 case Builtin::BIwork_group_reserve_write_pipe:
4653 case Builtin::BIsub_group_reserve_read_pipe:
4654 case Builtin::BIsub_group_reserve_write_pipe: {
4655 // Composing the mangled name for the function.
4656 const char *Name;
4657 if (BuiltinID == Builtin::BIreserve_read_pipe)
4658 Name = "__reserve_read_pipe";
4659 else if (BuiltinID == Builtin::BIreserve_write_pipe)
4660 Name = "__reserve_write_pipe";
4661 else if (BuiltinID == Builtin::BIwork_group_reserve_read_pipe)
4662 Name = "__work_group_reserve_read_pipe";
4663 else if (BuiltinID == Builtin::BIwork_group_reserve_write_pipe)
4664 Name = "__work_group_reserve_write_pipe";
4665 else if (BuiltinID == Builtin::BIsub_group_reserve_read_pipe)
4666 Name = "__sub_group_reserve_read_pipe";
4667 else
4668 Name = "__sub_group_reserve_write_pipe";
4669
4670 Value *Arg0 = EmitScalarExpr(E->getArg(0)),
4671 *Arg1 = EmitScalarExpr(E->getArg(1));
4672 llvm::Type *ReservedIDTy = ConvertType(getContext().OCLReserveIDTy);
4673 CGOpenCLRuntime OpenCLRT(CGM);
4674 Value *PacketSize = OpenCLRT.getPipeElemSize(E->getArg(0));
4675 Value *PacketAlign = OpenCLRT.getPipeElemAlign(E->getArg(0));
4676
4677 // Building the generic function prototype.
4678 llvm::Type *ArgTys[] = {Arg0->getType(), Int32Ty, Int32Ty, Int32Ty};
4679 llvm::FunctionType *FTy = llvm::FunctionType::get(
4680 ReservedIDTy, llvm::ArrayRef<llvm::Type *>(ArgTys), false);
4681 // We know the second argument is an integer type, but we may need to cast
4682 // it to i32.
4683 if (Arg1->getType() != Int32Ty)
4684 Arg1 = Builder.CreateZExtOrTrunc(Arg1, Int32Ty);
4685 return RValue::get(EmitRuntimeCall(CGM.CreateRuntimeFunction(FTy, Name),
4686 {Arg0, Arg1, PacketSize, PacketAlign}));
4687 }
4688 // OpenCL v2.0 s6.13.16, s9.17.3.5 - Built-in pipe commit read and write
4689 // functions
4690 case Builtin::BIcommit_read_pipe:
4691 case Builtin::BIcommit_write_pipe:
4692 case Builtin::BIwork_group_commit_read_pipe:
4693 case Builtin::BIwork_group_commit_write_pipe:
4694 case Builtin::BIsub_group_commit_read_pipe:
4695 case Builtin::BIsub_group_commit_write_pipe: {
4696 const char *Name;
4697 if (BuiltinID == Builtin::BIcommit_read_pipe)
4698 Name = "__commit_read_pipe";
4699 else if (BuiltinID == Builtin::BIcommit_write_pipe)
4700 Name = "__commit_write_pipe";
4701 else if (BuiltinID == Builtin::BIwork_group_commit_read_pipe)
4702 Name = "__work_group_commit_read_pipe";
4703 else if (BuiltinID == Builtin::BIwork_group_commit_write_pipe)
4704 Name = "__work_group_commit_write_pipe";
4705 else if (BuiltinID == Builtin::BIsub_group_commit_read_pipe)
4706 Name = "__sub_group_commit_read_pipe";
4707 else
4708 Name = "__sub_group_commit_write_pipe";
4709
4710 Value *Arg0 = EmitScalarExpr(E->getArg(0)),
4711 *Arg1 = EmitScalarExpr(E->getArg(1));
4712 CGOpenCLRuntime OpenCLRT(CGM);
4713 Value *PacketSize = OpenCLRT.getPipeElemSize(E->getArg(0));
4714 Value *PacketAlign = OpenCLRT.getPipeElemAlign(E->getArg(0));
4715
4716 // Building the generic function prototype.
4717 llvm::Type *ArgTys[] = {Arg0->getType(), Arg1->getType(), Int32Ty, Int32Ty};
4718 llvm::FunctionType *FTy =
4719 llvm::FunctionType::get(llvm::Type::getVoidTy(getLLVMContext()),
4720 llvm::ArrayRef<llvm::Type *>(ArgTys), false);
4721
4722 return RValue::get(EmitRuntimeCall(CGM.CreateRuntimeFunction(FTy, Name),
4723 {Arg0, Arg1, PacketSize, PacketAlign}));
4724 }
4725 // OpenCL v2.0 s6.13.16.4 Built-in pipe query functions
4726 case Builtin::BIget_pipe_num_packets:
4727 case Builtin::BIget_pipe_max_packets: {
4728 const char *BaseName;
4729 const auto *PipeTy = E->getArg(0)->getType()->castAs<PipeType>();
4730 if (BuiltinID == Builtin::BIget_pipe_num_packets)
4731 BaseName = "__get_pipe_num_packets";
4732 else
4733 BaseName = "__get_pipe_max_packets";
4734 std::string Name = std::string(BaseName) +
4735 std::string(PipeTy->isReadOnly() ? "_ro" : "_wo");
4736
4737 // Building the generic function prototype.
4738 Value *Arg0 = EmitScalarExpr(E->getArg(0));
4739 CGOpenCLRuntime OpenCLRT(CGM);
4740 Value *PacketSize = OpenCLRT.getPipeElemSize(E->getArg(0));
4741 Value *PacketAlign = OpenCLRT.getPipeElemAlign(E->getArg(0));
4742 llvm::Type *ArgTys[] = {Arg0->getType(), Int32Ty, Int32Ty};
4743 llvm::FunctionType *FTy = llvm::FunctionType::get(
4744 Int32Ty, llvm::ArrayRef<llvm::Type *>(ArgTys), false);
4745
4746 return RValue::get(EmitRuntimeCall(CGM.CreateRuntimeFunction(FTy, Name),
4747 {Arg0, PacketSize, PacketAlign}));
4748 }
4749
4750 // OpenCL v2.0 s6.13.9 - Address space qualifier functions.
4751 case Builtin::BIto_global:
4752 case Builtin::BIto_local:
4753 case Builtin::BIto_private: {
4754 auto Arg0 = EmitScalarExpr(E->getArg(0));
4755 auto NewArgT = llvm::PointerType::get(Int8Ty,
4756 CGM.getContext().getTargetAddressSpace(LangAS::opencl_generic));
4757 auto NewRetT = llvm::PointerType::get(Int8Ty,
4758 CGM.getContext().getTargetAddressSpace(
4759 E->getType()->getPointeeType().getAddressSpace()));
4760 auto FTy = llvm::FunctionType::get(NewRetT, {NewArgT}, false);
4761 llvm::Value *NewArg;
4762 if (Arg0->getType()->getPointerAddressSpace() !=
4763 NewArgT->getPointerAddressSpace())
4764 NewArg = Builder.CreateAddrSpaceCast(Arg0, NewArgT);
4765 else
4766 NewArg = Builder.CreateBitOrPointerCast(Arg0, NewArgT);
4767 auto NewName = std::string("__") + E->getDirectCallee()->getName().str();
4768 auto NewCall =
4769 EmitRuntimeCall(CGM.CreateRuntimeFunction(FTy, NewName), {NewArg});
4770 return RValue::get(Builder.CreateBitOrPointerCast(NewCall,
4771 ConvertType(E->getType())));
4772 }
4773
4774 // OpenCL v2.0, s6.13.17 - Enqueue kernel function.
4775 // It contains four different overload formats specified in Table 6.13.17.1.
4776 case Builtin::BIenqueue_kernel: {
4777 StringRef Name; // Generated function call name
4778 unsigned NumArgs = E->getNumArgs();
4779
4780 llvm::Type *QueueTy = ConvertType(getContext().OCLQueueTy);
4781 llvm::Type *GenericVoidPtrTy = Builder.getInt8PtrTy(
4782 getContext().getTargetAddressSpace(LangAS::opencl_generic));
4783
4784 llvm::Value *Queue = EmitScalarExpr(E->getArg(0));
4785 llvm::Value *Flags = EmitScalarExpr(E->getArg(1));
4786 LValue NDRangeL = EmitAggExprToLValue(E->getArg(2));
4787 llvm::Value *Range = NDRangeL.getAddress(*this).getPointer();
4788 llvm::Type *RangeTy = NDRangeL.getAddress(*this).getType();
4789
4790 if (NumArgs == 4) {
4791 // The most basic form of the call with parameters:
4792 // queue_t, kernel_enqueue_flags_t, ndrange_t, block(void)
4793 Name = "__enqueue_kernel_basic";
4794 llvm::Type *ArgTys[] = {QueueTy, Int32Ty, RangeTy, GenericVoidPtrTy,
4795 GenericVoidPtrTy};
4796 llvm::FunctionType *FTy = llvm::FunctionType::get(
4797 Int32Ty, llvm::ArrayRef<llvm::Type *>(ArgTys), false);
4798
4799 auto Info =
4800 CGM.getOpenCLRuntime().emitOpenCLEnqueuedBlock(*this, E->getArg(3));
4801 llvm::Value *Kernel =
4802 Builder.CreatePointerCast(Info.Kernel, GenericVoidPtrTy);
4803 llvm::Value *Block =
4804 Builder.CreatePointerCast(Info.BlockArg, GenericVoidPtrTy);
4805
4806 AttrBuilder B;
4807 B.addByValAttr(NDRangeL.getAddress(*this).getElementType());
4808 llvm::AttributeList ByValAttrSet =
4809 llvm::AttributeList::get(CGM.getModule().getContext(), 3U, B);
4810
4811 auto RTCall =
4812 EmitRuntimeCall(CGM.CreateRuntimeFunction(FTy, Name, ByValAttrSet),
4813 {Queue, Flags, Range, Kernel, Block});
4814 RTCall->setAttributes(ByValAttrSet);
4815 return RValue::get(RTCall);
4816 }
4817 assert(NumArgs >= 5 && "Invalid enqueue_kernel signature")((void)0);
4818
4819 // Create a temporary array to hold the sizes of local pointer arguments
4820 // for the block. \p First is the position of the first size argument.
4821 auto CreateArrayForSizeVar = [=](unsigned First)
4822 -> std::tuple<llvm::Value *, llvm::Value *, llvm::Value *> {
4823 llvm::APInt ArraySize(32, NumArgs - First);
4824 QualType SizeArrayTy = getContext().getConstantArrayType(
4825 getContext().getSizeType(), ArraySize, nullptr, ArrayType::Normal,
4826 /*IndexTypeQuals=*/0);
4827 auto Tmp = CreateMemTemp(SizeArrayTy, "block_sizes");
4828 llvm::Value *TmpPtr = Tmp.getPointer();
4829 llvm::Value *TmpSize = EmitLifetimeStart(
4830 CGM.getDataLayout().getTypeAllocSize(Tmp.getElementType()), TmpPtr);
4831 llvm::Value *ElemPtr;
4832 // Each of the following arguments specifies the size of the corresponding
4833 // argument passed to the enqueued block.
4834 auto *Zero = llvm::ConstantInt::get(IntTy, 0);
4835 for (unsigned I = First; I < NumArgs; ++I) {
4836 auto *Index = llvm::ConstantInt::get(IntTy, I - First);
4837 auto *GEP = Builder.CreateGEP(Tmp.getElementType(), TmpPtr,
4838 {Zero, Index});
4839 if (I == First)
4840 ElemPtr = GEP;
4841 auto *V =
4842 Builder.CreateZExtOrTrunc(EmitScalarExpr(E->getArg(I)), SizeTy);
4843 Builder.CreateAlignedStore(
4844 V, GEP, CGM.getDataLayout().getPrefTypeAlign(SizeTy));
4845 }
4846 return std::tie(ElemPtr, TmpSize, TmpPtr);
4847 };
4848
4849 // Could have events and/or varargs.
4850 if (E->getArg(3)->getType()->isBlockPointerType()) {
4851 // No events passed, but has variadic arguments.
4852 Name = "__enqueue_kernel_varargs";
4853 auto Info =
4854 CGM.getOpenCLRuntime().emitOpenCLEnqueuedBlock(*this, E->getArg(3));
4855 llvm::Value *Kernel =
4856 Builder.CreatePointerCast(Info.Kernel, GenericVoidPtrTy);
4857 auto *Block = Builder.CreatePointerCast(Info.BlockArg, GenericVoidPtrTy);
4858 llvm::Value *ElemPtr, *TmpSize, *TmpPtr;
4859 std::tie(ElemPtr, TmpSize, TmpPtr) = CreateArrayForSizeVar(4);
4860
4861 // Create a vector of the arguments, as well as a constant value to
4862 // express to the runtime the number of variadic arguments.
4863 llvm::Value *const Args[] = {Queue, Flags,
4864 Range, Kernel,
4865 Block, ConstantInt::get(IntTy, NumArgs - 4),
4866 ElemPtr};
4867 llvm::Type *const ArgTys[] = {
4868 QueueTy, IntTy, RangeTy, GenericVoidPtrTy,
4869 GenericVoidPtrTy, IntTy, ElemPtr->getType()};
4870
4871 llvm::FunctionType *FTy = llvm::FunctionType::get(Int32Ty, ArgTys, false);
4872 auto Call = RValue::get(
4873 EmitRuntimeCall(CGM.CreateRuntimeFunction(FTy, Name), Args));
4874 if (TmpSize)
4875 EmitLifetimeEnd(TmpSize, TmpPtr);
4876 return Call;
4877 }
4878 // Any calls now have event arguments passed.
4879 if (NumArgs >= 7) {
4880 llvm::Type *EventTy = ConvertType(getContext().OCLClkEventTy);
4881 llvm::PointerType *EventPtrTy = EventTy->getPointerTo(
4882 CGM.getContext().getTargetAddressSpace(LangAS::opencl_generic));
4883
4884 llvm::Value *NumEvents =
4885 Builder.CreateZExtOrTrunc(EmitScalarExpr(E->getArg(3)), Int32Ty);
4886
4887 // Since SemaOpenCLBuiltinEnqueueKernel allows fifth and sixth arguments
4888 // to be a null pointer constant (including `0` literal), we can take it
4889 // into account and emit null pointer directly.
4890 llvm::Value *EventWaitList = nullptr;
4891 if (E->getArg(4)->isNullPointerConstant(
4892 getContext(), Expr::NPC_ValueDependentIsNotNull)) {
4893 EventWaitList = llvm::ConstantPointerNull::get(EventPtrTy);
4894 } else {
4895 EventWaitList = E->getArg(4)->getType()->isArrayType()
4896 ? EmitArrayToPointerDecay(E->getArg(4)).getPointer()
4897 : EmitScalarExpr(E->getArg(4));
4898 // Convert to generic address space.
4899 EventWaitList = Builder.CreatePointerCast(EventWaitList, EventPtrTy);
4900 }
4901 llvm::Value *EventRet = nullptr;
4902 if (E->getArg(5)->isNullPointerConstant(
4903 getContext(), Expr::NPC_ValueDependentIsNotNull)) {
4904 EventRet = llvm::ConstantPointerNull::get(EventPtrTy);
4905 } else {
4906 EventRet =
4907 Builder.CreatePointerCast(EmitScalarExpr(E->getArg(5)), EventPtrTy);
4908 }
4909
4910 auto Info =
4911 CGM.getOpenCLRuntime().emitOpenCLEnqueuedBlock(*this, E->getArg(6));
4912 llvm::Value *Kernel =
4913 Builder.CreatePointerCast(Info.Kernel, GenericVoidPtrTy);
4914 llvm::Value *Block =
4915 Builder.CreatePointerCast(Info.BlockArg, GenericVoidPtrTy);
4916
4917 std::vector<llvm::Type *> ArgTys = {
4918 QueueTy, Int32Ty, RangeTy, Int32Ty,
4919 EventPtrTy, EventPtrTy, GenericVoidPtrTy, GenericVoidPtrTy};
4920
4921 std::vector<llvm::Value *> Args = {Queue, Flags, Range,
4922 NumEvents, EventWaitList, EventRet,
4923 Kernel, Block};
4924
4925 if (NumArgs == 7) {
4926 // Has events but no variadics.
4927 Name = "__enqueue_kernel_basic_events";
4928 llvm::FunctionType *FTy = llvm::FunctionType::get(
4929 Int32Ty, llvm::ArrayRef<llvm::Type *>(ArgTys), false);
4930 return RValue::get(
4931 EmitRuntimeCall(CGM.CreateRuntimeFunction(FTy, Name),
4932 llvm::ArrayRef<llvm::Value *>(Args)));
4933 }
4934 // Has event info and variadics
4935 // Pass the number of variadics to the runtime function too.
4936 Args.push_back(ConstantInt::get(Int32Ty, NumArgs - 7));
4937 ArgTys.push_back(Int32Ty);
4938 Name = "__enqueue_kernel_events_varargs";
4939
4940 llvm::Value *ElemPtr, *TmpSize, *TmpPtr;
4941 std::tie(ElemPtr, TmpSize, TmpPtr) = CreateArrayForSizeVar(7);
4942 Args.push_back(ElemPtr);
4943 ArgTys.push_back(ElemPtr->getType());
4944
4945 llvm::FunctionType *FTy = llvm::FunctionType::get(
4946 Int32Ty, llvm::ArrayRef<llvm::Type *>(ArgTys), false);
4947 auto Call =
4948 RValue::get(EmitRuntimeCall(CGM.CreateRuntimeFunction(FTy, Name),
4949 llvm::ArrayRef<llvm::Value *>(Args)));
4950 if (TmpSize)
4951 EmitLifetimeEnd(TmpSize, TmpPtr);
4952 return Call;
4953 }
4954 LLVM_FALLTHROUGH[[gnu::fallthrough]];
4955 }
4956 // OpenCL v2.0 s6.13.17.6 - Kernel query functions need bitcast of block
4957 // parameter.
4958 case Builtin::BIget_kernel_work_group_size: {
4959 llvm::Type *GenericVoidPtrTy = Builder.getInt8PtrTy(
4960 getContext().getTargetAddressSpace(LangAS::opencl_generic));
4961 auto Info =
4962 CGM.getOpenCLRuntime().emitOpenCLEnqueuedBlock(*this, E->getArg(0));
4963 Value *Kernel = Builder.CreatePointerCast(Info.Kernel, GenericVoidPtrTy);
4964 Value *Arg = Builder.CreatePointerCast(Info.BlockArg, GenericVoidPtrTy);
4965 return RValue::get(EmitRuntimeCall(
4966 CGM.CreateRuntimeFunction(
4967 llvm::FunctionType::get(IntTy, {GenericVoidPtrTy, GenericVoidPtrTy},
4968 false),
4969 "__get_kernel_work_group_size_impl"),
4970 {Kernel, Arg}));
4971 }
4972 case Builtin::BIget_kernel_preferred_work_group_size_multiple: {
4973 llvm::Type *GenericVoidPtrTy = Builder.getInt8PtrTy(
4974 getContext().getTargetAddressSpace(LangAS::opencl_generic));
4975 auto Info =
4976 CGM.getOpenCLRuntime().emitOpenCLEnqueuedBlock(*this, E->getArg(0));
4977 Value *Kernel = Builder.CreatePointerCast(Info.Kernel, GenericVoidPtrTy);
4978 Value *Arg = Builder.CreatePointerCast(Info.BlockArg, GenericVoidPtrTy);
4979 return RValue::get(EmitRuntimeCall(
4980 CGM.CreateRuntimeFunction(
4981 llvm::FunctionType::get(IntTy, {GenericVoidPtrTy, GenericVoidPtrTy},
4982 false),
4983 "__get_kernel_preferred_work_group_size_multiple_impl"),
4984 {Kernel, Arg}));
4985 }
4986 case Builtin::BIget_kernel_max_sub_group_size_for_ndrange:
4987 case Builtin::BIget_kernel_sub_group_count_for_ndrange: {
4988 llvm::Type *GenericVoidPtrTy = Builder.getInt8PtrTy(
4989 getContext().getTargetAddressSpace(LangAS::opencl_generic));
4990 LValue NDRangeL = EmitAggExprToLValue(E->getArg(0));
4991 llvm::Value *NDRange = NDRangeL.getAddress(*this).getPointer();
4992 auto Info =
4993 CGM.getOpenCLRuntime().emitOpenCLEnqueuedBlock(*this, E->getArg(1));
4994 Value *Kernel = Builder.CreatePointerCast(Info.Kernel, GenericVoidPtrTy);
4995 Value *Block = Builder.CreatePointerCast(Info.BlockArg, GenericVoidPtrTy);
4996 const char *Name =
4997 BuiltinID == Builtin::BIget_kernel_max_sub_group_size_for_ndrange
4998 ? "__get_kernel_max_sub_group_size_for_ndrange_impl"
4999 : "__get_kernel_sub_group_count_for_ndrange_impl";
5000 return RValue::get(EmitRuntimeCall(
5001 CGM.CreateRuntimeFunction(
5002 llvm::FunctionType::get(
5003 IntTy, {NDRange->getType(), GenericVoidPtrTy, GenericVoidPtrTy},
5004 false),
5005 Name),
5006 {NDRange, Kernel, Block}));
5007 }
5008
5009 case Builtin::BI__builtin_store_half:
5010 case Builtin::BI__builtin_store_halff: {
5011 Value *Val = EmitScalarExpr(E->getArg(0));
5012 Address Address = EmitPointerWithAlignment(E->getArg(1));
5013 Value *HalfVal = Builder.CreateFPTrunc(Val, Builder.getHalfTy());
5014 return RValue::get(Builder.CreateStore(HalfVal, Address));
5015 }
5016 case Builtin::BI__builtin_load_half: {
5017 Address Address = EmitPointerWithAlignment(E->getArg(0));
5018 Value *HalfVal = Builder.CreateLoad(Address);
5019 return RValue::get(Builder.CreateFPExt(HalfVal, Builder.getDoubleTy()));
5020 }
5021 case Builtin::BI__builtin_load_halff: {
5022 Address Address = EmitPointerWithAlignment(E->getArg(0));
5023 Value *HalfVal = Builder.CreateLoad(Address);
5024 return RValue::get(Builder.CreateFPExt(HalfVal, Builder.getFloatTy()));
5025 }
5026 case Builtin::BIprintf:
5027 if (getTarget().getTriple().isNVPTX())
5028 return EmitNVPTXDevicePrintfCallExpr(E, ReturnValue);
5029 if (getTarget().getTriple().getArch() == Triple::amdgcn &&
5030 getLangOpts().HIP)
5031 return EmitAMDGPUDevicePrintfCallExpr(E, ReturnValue);
5032 break;
5033 case Builtin::BI__builtin_canonicalize:
5034 case Builtin::BI__builtin_canonicalizef:
5035 case Builtin::BI__builtin_canonicalizef16:
5036 case Builtin::BI__builtin_canonicalizel:
5037 return RValue::get(emitUnaryBuiltin(*this, E, Intrinsic::canonicalize));
5038
5039 case Builtin::BI__builtin_thread_pointer: {
5040 if (!getContext().getTargetInfo().isTLSSupported())
5041 CGM.ErrorUnsupported(E, "__builtin_thread_pointer");
5042 // Fall through - it's already mapped to the intrinsic by GCCBuiltin.
5043 break;
5044 }
5045 case Builtin::BI__builtin_os_log_format:
5046 return emitBuiltinOSLogFormat(*E);
5047
5048 case Builtin::BI__xray_customevent: {
5049 if (!ShouldXRayInstrumentFunction())
5050 return RValue::getIgnored();
5051
5052 if (!CGM.getCodeGenOpts().XRayInstrumentationBundle.has(
5053 XRayInstrKind::Custom))
5054 return RValue::getIgnored();
5055
5056 if (const auto *XRayAttr = CurFuncDecl->getAttr<XRayInstrumentAttr>())
5057 if (XRayAttr->neverXRayInstrument() && !AlwaysEmitXRayCustomEvents())
5058 return RValue::getIgnored();
5059
5060 Function *F = CGM.getIntrinsic(Intrinsic::xray_customevent);
5061 auto FTy = F->getFunctionType();
5062 auto Arg0 = E->getArg(0);
5063 auto Arg0Val = EmitScalarExpr(Arg0);
5064 auto Arg0Ty = Arg0->getType();
5065 auto PTy0 = FTy->getParamType(0);
5066 if (PTy0 != Arg0Val->getType()) {
5067 if (Arg0Ty->isArrayType())
5068 Arg0Val = EmitArrayToPointerDecay(Arg0).getPointer();
5069 else
5070 Arg0Val = Builder.CreatePointerCast(Arg0Val, PTy0);
5071 }
5072 auto Arg1 = EmitScalarExpr(E->getArg(1));
5073 auto PTy1 = FTy->getParamType(1);
5074 if (PTy1 != Arg1->getType())
5075 Arg1 = Builder.CreateTruncOrBitCast(Arg1, PTy1);
5076 return RValue::get(Builder.CreateCall(F, {Arg0Val, Arg1}));
5077 }
5078
5079 case Builtin::BI__xray_typedevent: {
5080 // TODO: There should be a way to always emit events even if the current
5081 // function is not instrumented. Losing events in a stream can cripple
5082 // a trace.
5083 if (!ShouldXRayInstrumentFunction())
5084 return RValue::getIgnored();
5085
5086 if (!CGM.getCodeGenOpts().XRayInstrumentationBundle.has(
5087 XRayInstrKind::Typed))
5088 return RValue::getIgnored();
5089
5090 if (const auto *XRayAttr = CurFuncDecl->getAttr<XRayInstrumentAttr>())
5091 if (XRayAttr->neverXRayInstrument() && !AlwaysEmitXRayTypedEvents())
5092 return RValue::getIgnored();
5093
5094 Function *F = CGM.getIntrinsic(Intrinsic::xray_typedevent);
5095 auto FTy = F->getFunctionType();
5096 auto Arg0 = EmitScalarExpr(E->getArg(0));
5097 auto PTy0 = FTy->getParamType(0);
5098 if (PTy0 != Arg0->getType())
5099 Arg0 = Builder.CreateTruncOrBitCast(Arg0, PTy0);
5100 auto Arg1 = E->getArg(1);
5101 auto Arg1Val = EmitScalarExpr(Arg1);
5102 auto Arg1Ty = Arg1->getType();
5103 auto PTy1 = FTy->getParamType(1);
5104 if (PTy1 != Arg1Val->getType()) {
5105 if (Arg1Ty->isArrayType())
5106 Arg1Val = EmitArrayToPointerDecay(Arg1).getPointer();
5107 else
5108 Arg1Val = Builder.CreatePointerCast(Arg1Val, PTy1);
5109 }
5110 auto Arg2 = EmitScalarExpr(E->getArg(2));
5111 auto PTy2 = FTy->getParamType(2);
5112 if (PTy2 != Arg2->getType())
5113 Arg2 = Builder.CreateTruncOrBitCast(Arg2, PTy2);
5114 return RValue::get(Builder.CreateCall(F, {Arg0, Arg1Val, Arg2}));
5115 }
5116
5117 case Builtin::BI__builtin_ms_va_start:
5118 case Builtin::BI__builtin_ms_va_end:
5119 return RValue::get(
5120 EmitVAStartEnd(EmitMSVAListRef(E->getArg(0)).getPointer(),
5121 BuiltinID == Builtin::BI__builtin_ms_va_start));
5122
5123 case Builtin::BI__builtin_ms_va_copy: {
5124 // Lower this manually. We can't reliably determine whether or not any
5125 // given va_copy() is for a Win64 va_list from the calling convention
5126 // alone, because it's legal to do this from a System V ABI function.
5127 // With opaque pointer types, we won't have enough information in LLVM
5128 // IR to determine this from the argument types, either. Best to do it
5129 // now, while we have enough information.
5130 Address DestAddr = EmitMSVAListRef(E->getArg(0));
5131 Address SrcAddr = EmitMSVAListRef(E->getArg(1));
5132
5133 llvm::Type *BPP = Int8PtrPtrTy;
5134
5135 DestAddr = Address(Builder.CreateBitCast(DestAddr.getPointer(), BPP, "cp"),
5136 DestAddr.getAlignment());
5137 SrcAddr = Address(Builder.CreateBitCast(SrcAddr.getPointer(), BPP, "ap"),
5138 SrcAddr.getAlignment());
5139
5140 Value *ArgPtr = Builder.CreateLoad(SrcAddr, "ap.val");
5141 return RValue::get(Builder.CreateStore(ArgPtr, DestAddr));
5142 }
5143
5144 case Builtin::BI__builtin_get_device_side_mangled_name: {
5145 auto Name = CGM.getCUDARuntime().getDeviceSideName(
5146 cast<DeclRefExpr>(E->getArg(0)->IgnoreImpCasts())->getDecl());
5147 auto Str = CGM.GetAddrOfConstantCString(Name, "");
5148 llvm::Constant *Zeros[] = {llvm::ConstantInt::get(SizeTy, 0),
5149 llvm::ConstantInt::get(SizeTy, 0)};
5150 auto *Ptr = llvm::ConstantExpr::getGetElementPtr(Str.getElementType(),
5151 Str.getPointer(), Zeros);
5152 return RValue::get(Ptr);
5153 }
5154 }
5155
5156 // If this is an alias for a lib function (e.g. __builtin_sin), emit
5157 // the call using the normal call path, but using the unmangled
5158 // version of the function name.
5159 if (getContext().BuiltinInfo.isLibFunction(BuiltinID))
5160 return emitLibraryCall(*this, FD, E,
5161 CGM.getBuiltinLibFunction(FD, BuiltinID));
5162
5163 // If this is a predefined lib function (e.g. malloc), emit the call
5164 // using exactly the normal call path.
5165 if (getContext().BuiltinInfo.isPredefinedLibFunction(BuiltinID))
5166 return emitLibraryCall(*this, FD, E,
5167 cast<llvm::Constant>(EmitScalarExpr(E->getCallee())));
5168
5169 // Check that a call to a target specific builtin has the correct target
5170 // features.
5171 // This is down here to avoid non-target specific builtins, however, if
5172 // generic builtins start to require generic target features then we
5173 // can move this up to the beginning of the function.
5174 checkTargetFeatures(E, FD);
5175
5176 if (unsigned VectorWidth = getContext().BuiltinInfo.getRequiredVectorWidth(BuiltinID))
5177 LargestVectorWidth = std::max(LargestVectorWidth, VectorWidth);
5178
5179 // See if we have a target specific intrinsic.
5180 const char *Name = getContext().BuiltinInfo.getName(BuiltinID);
5181 Intrinsic::ID IntrinsicID = Intrinsic::not_intrinsic;
5182 StringRef Prefix =
5183 llvm::Triple::getArchTypePrefix(getTarget().getTriple().getArch());
5184 if (!Prefix.empty()) {
5185 IntrinsicID = Intrinsic::getIntrinsicForGCCBuiltin(Prefix.data(), Name);
5186 // NOTE we don't need to perform a compatibility flag check here since the
5187 // intrinsics are declared in Builtins*.def via LANGBUILTIN which filter the
5188 // MS builtins via ALL_MS_LANGUAGES and are filtered earlier.
5189 if (IntrinsicID == Intrinsic::not_intrinsic)
5190 IntrinsicID = Intrinsic::getIntrinsicForMSBuiltin(Prefix.data(), Name);
5191 }
5192
5193 if (IntrinsicID != Intrinsic::not_intrinsic) {
5194 SmallVector<Value*, 16> Args;
5195
5196 // Find out if any arguments are required to be integer constant
5197 // expressions.
5198 unsigned ICEArguments = 0;
5199 ASTContext::GetBuiltinTypeError Error;
5200 getContext().GetBuiltinType(BuiltinID, Error, &ICEArguments);
5201 assert(Error == ASTContext::GE_None && "Should not codegen an error")((void)0);
5202
5203 Function *F = CGM.getIntrinsic(IntrinsicID);
5204 llvm::FunctionType *FTy = F->getFunctionType();
5205
5206 for (unsigned i = 0, e = E->getNumArgs(); i != e; ++i) {
5207 Value *ArgValue;
5208 // If this is a normal argument, just emit it as a scalar.
5209 if ((ICEArguments & (1 << i)) == 0) {
5210 ArgValue = EmitScalarExpr(E->getArg(i));
5211 } else {
5212 // If this is required to be a constant, constant fold it so that we
5213 // know that the generated intrinsic gets a ConstantInt.
5214 ArgValue = llvm::ConstantInt::get(
5215 getLLVMContext(),
5216 *E->getArg(i)->getIntegerConstantExpr(getContext()));
5217 }
5218
5219 // If the intrinsic arg type is different from the builtin arg type
5220 // we need to do a bit cast.
5221 llvm::Type *PTy = FTy->getParamType(i);
5222 if (PTy != ArgValue->getType()) {
5223 // XXX - vector of pointers?
5224 if (auto *PtrTy = dyn_cast<llvm::PointerType>(PTy)) {
5225 if (PtrTy->getAddressSpace() !=
5226 ArgValue->getType()->getPointerAddressSpace()) {
5227 ArgValue = Builder.CreateAddrSpaceCast(
5228 ArgValue,
5229 ArgValue->getType()->getPointerTo(PtrTy->getAddressSpace()));
5230 }
5231 }
5232
5233 assert(PTy->canLosslesslyBitCastTo(FTy->getParamType(i)) &&((void)0)
5234 "Must be able to losslessly bit cast to param")((void)0);
5235 ArgValue = Builder.CreateBitCast(ArgValue, PTy);
5236 }
5237
5238 Args.push_back(ArgValue);
5239 }
5240
5241 Value *V = Builder.CreateCall(F, Args);
5242 QualType BuiltinRetType = E->getType();
5243
5244 llvm::Type *RetTy = VoidTy;
5245 if (!BuiltinRetType->isVoidType())
5246 RetTy = ConvertType(BuiltinRetType);
5247
5248 if (RetTy != V->getType()) {
5249 // XXX - vector of pointers?
5250 if (auto *PtrTy = dyn_cast<llvm::PointerType>(RetTy)) {
5251 if (PtrTy->getAddressSpace() != V->getType()->getPointerAddressSpace()) {
5252 V = Builder.CreateAddrSpaceCast(
5253 V, V->getType()->getPointerTo(PtrTy->getAddressSpace()));
5254 }
5255 }
5256
5257 assert(V->getType()->canLosslesslyBitCastTo(RetTy) &&((void)0)
5258 "Must be able to losslessly bit cast result type")((void)0);
5259 V = Builder.CreateBitCast(V, RetTy);
5260 }
5261
5262 return RValue::get(V);
5263 }
5264
5265 // Some target-specific builtins can have aggregate return values, e.g.
5266 // __builtin_arm_mve_vld2q_u32. So if the result is an aggregate, force
5267 // ReturnValue to be non-null, so that the target-specific emission code can
5268 // always just emit into it.
5269 TypeEvaluationKind EvalKind = getEvaluationKind(E->getType());
5270 if (EvalKind == TEK_Aggregate && ReturnValue.isNull()) {
5271 Address DestPtr = CreateMemTemp(E->getType(), "agg.tmp");
5272 ReturnValue = ReturnValueSlot(DestPtr, false);
5273 }
5274
5275 // Now see if we can emit a target-specific builtin.
5276 if (Value *V = EmitTargetBuiltinExpr(BuiltinID, E, ReturnValue)) {
5277 switch (EvalKind) {
5278 case TEK_Scalar:
5279 return RValue::get(V);
5280 case TEK_Aggregate:
5281 return RValue::getAggregate(ReturnValue.getValue(),
5282 ReturnValue.isVolatile());
5283 case TEK_Complex:
5284 llvm_unreachable("No current target builtin returns complex")__builtin_unreachable();
5285 }
5286 llvm_unreachable("Bad evaluation kind in EmitBuiltinExpr")__builtin_unreachable();
5287 }
5288
5289 ErrorUnsupported(E, "builtin function");
5290
5291 // Unknown builtin, for now just dump it out and return undef.
5292 return GetUndefRValue(E->getType());
5293}
5294
5295static Value *EmitTargetArchBuiltinExpr(CodeGenFunction *CGF,
5296 unsigned BuiltinID, const CallExpr *E,
5297 ReturnValueSlot ReturnValue,
5298 llvm::Triple::ArchType Arch) {
5299 switch (Arch) {
5300 case llvm::Triple::arm:
5301 case llvm::Triple::armeb:
5302 case llvm::Triple::thumb:
5303 case llvm::Triple::thumbeb:
5304 return CGF->EmitARMBuiltinExpr(BuiltinID, E, ReturnValue, Arch);
5305 case llvm::Triple::aarch64:
5306 case llvm::Triple::aarch64_32:
5307 case llvm::Triple::aarch64_be:
5308 return CGF->EmitAArch64BuiltinExpr(BuiltinID, E, Arch);
5309 case llvm::Triple::bpfeb:
5310 case llvm::Triple::bpfel:
5311 return CGF->EmitBPFBuiltinExpr(BuiltinID, E);
5312 case llvm::Triple::x86:
5313 case llvm::Triple::x86_64:
5314 return CGF->EmitX86BuiltinExpr(BuiltinID, E);
5315 case llvm::Triple::ppc:
5316 case llvm::Triple::ppcle:
5317 case llvm::Triple::ppc64:
5318 case llvm::Triple::ppc64le:
5319 return CGF->EmitPPCBuiltinExpr(BuiltinID, E);
5320 case llvm::Triple::r600:
5321 case llvm::Triple::amdgcn:
5322 return CGF->EmitAMDGPUBuiltinExpr(BuiltinID, E);
5323 case llvm::Triple::systemz:
5324 return CGF->EmitSystemZBuiltinExpr(BuiltinID, E);
5325 case llvm::Triple::nvptx:
5326 case llvm::Triple::nvptx64:
5327 return CGF->EmitNVPTXBuiltinExpr(BuiltinID, E);
5328 case llvm::Triple::wasm32:
5329 case llvm::Triple::wasm64:
5330 return CGF->EmitWebAssemblyBuiltinExpr(BuiltinID, E);
5331 case llvm::Triple::hexagon:
5332 return CGF->EmitHexagonBuiltinExpr(BuiltinID, E);
5333 case llvm::Triple::riscv32:
5334 case llvm::Triple::riscv64:
5335 return CGF->EmitRISCVBuiltinExpr(BuiltinID, E, ReturnValue);
5336 default:
5337 return nullptr;
5338 }
5339}
5340
5341Value *CodeGenFunction::EmitTargetBuiltinExpr(unsigned BuiltinID,
5342 const CallExpr *E,
5343 ReturnValueSlot ReturnValue) {
5344 if (getContext().BuiltinInfo.isAuxBuiltinID(BuiltinID)) {
5345 assert(getContext().getAuxTargetInfo() && "Missing aux target info")((void)0);
5346 return EmitTargetArchBuiltinExpr(
5347 this, getContext().BuiltinInfo.getAuxBuiltinID(BuiltinID), E,
5348 ReturnValue, getContext().getAuxTargetInfo()->getTriple().getArch());
5349 }
5350
5351 return EmitTargetArchBuiltinExpr(this, BuiltinID, E, ReturnValue,
5352 getTarget().getTriple().getArch());
5353}
5354
5355static llvm::FixedVectorType *GetNeonType(CodeGenFunction *CGF,
5356 NeonTypeFlags TypeFlags,
5357 bool HasLegalHalfType = true,
5358 bool V1Ty = false,
5359 bool AllowBFloatArgsAndRet = true) {
5360 int IsQuad = TypeFlags.isQuad();
5361 switch (TypeFlags.getEltType()) {
5362 case NeonTypeFlags::Int8:
5363 case NeonTypeFlags::Poly8:
5364 return llvm::FixedVectorType::get(CGF->Int8Ty, V1Ty ? 1 : (8 << IsQuad));
5365 case NeonTypeFlags::Int16:
5366 case NeonTypeFlags::Poly16:
5367 return llvm::FixedVectorType::get(CGF->Int16Ty, V1Ty ? 1 : (4 << IsQuad));
5368 case NeonTypeFlags::BFloat16:
5369 if (AllowBFloatArgsAndRet)
5370 return llvm::FixedVectorType::get(CGF->BFloatTy, V1Ty ? 1 : (4 << IsQuad));
5371 else
5372 return llvm::FixedVectorType::get(CGF->Int16Ty, V1Ty ? 1 : (4 << IsQuad));
5373 case NeonTypeFlags::Float16:
5374 if (HasLegalHalfType)
5375 return llvm::FixedVectorType::get(CGF->HalfTy, V1Ty ? 1 : (4 << IsQuad));
5376 else
5377 return llvm::FixedVectorType::get(CGF->Int16Ty, V1Ty ? 1 : (4 << IsQuad));
5378 case NeonTypeFlags::Int32:
5379 return llvm::FixedVectorType::get(CGF->Int32Ty, V1Ty ? 1 : (2 << IsQuad));
5380 case NeonTypeFlags::Int64:
5381 case NeonTypeFlags::Poly64:
5382 return llvm::FixedVectorType::get(CGF->Int64Ty, V1Ty ? 1 : (1 << IsQuad));
5383 case NeonTypeFlags::Poly128:
5384 // FIXME: i128 and f128 doesn't get fully support in Clang and llvm.
5385 // There is a lot of i128 and f128 API missing.
5386 // so we use v16i8 to represent poly128 and get pattern matched.
5387 return llvm::FixedVectorType::get(CGF->Int8Ty, 16);
5388 case NeonTypeFlags::Float32:
5389 return llvm::FixedVectorType::get(CGF->FloatTy, V1Ty ? 1 : (2 << IsQuad));
5390 case NeonTypeFlags::Float64:
5391 return llvm::FixedVectorType::get(CGF->DoubleTy, V1Ty ? 1 : (1 << IsQuad));
5392 }
5393 llvm_unreachable("Unknown vector element type!")__builtin_unreachable();
5394}
5395
5396static llvm::VectorType *GetFloatNeonType(CodeGenFunction *CGF,
5397 NeonTypeFlags IntTypeFlags) {
5398 int IsQuad = IntTypeFlags.isQuad();
5399 switch (IntTypeFlags.getEltType()) {
5400 case NeonTypeFlags::Int16:
5401 return llvm::FixedVectorType::get(CGF->HalfTy, (4 << IsQuad));
5402 case NeonTypeFlags::Int32:
5403 return llvm::FixedVectorType::get(CGF->FloatTy, (2 << IsQuad));
5404 case NeonTypeFlags::Int64:
5405 return llvm::FixedVectorType::get(CGF->DoubleTy, (1 << IsQuad));
5406 default:
5407 llvm_unreachable("Type can't be converted to floating-point!")__builtin_unreachable();
5408 }
5409}
5410
5411Value *CodeGenFunction::EmitNeonSplat(Value *V, Constant *C,
5412 const ElementCount &Count) {
5413 Value *SV = llvm::ConstantVector::getSplat(Count, C);
5414 return Builder.CreateShuffleVector(V, V, SV, "lane");
5415}
5416
5417Value *CodeGenFunction::EmitNeonSplat(Value *V, Constant *C) {
5418 ElementCount EC = cast<llvm::VectorType>(V->getType())->getElementCount();
5419 return EmitNeonSplat(V, C, EC);
5420}
5421
5422Value *CodeGenFunction::EmitNeonCall(Function *F, SmallVectorImpl<Value*> &Ops,
5423 const char *name,
5424 unsigned shift, bool rightshift) {
5425 unsigned j = 0;
5426 for (Function::const_arg_iterator ai = F->arg_begin(), ae = F->arg_end();
5427 ai != ae; ++ai, ++j) {
5428 if (F->isConstrainedFPIntrinsic())
5429 if (ai->getType()->isMetadataTy())
5430 continue;
5431 if (shift > 0 && shift == j)
5432 Ops[j] = EmitNeonShiftVector(Ops[j], ai->getType(), rightshift);
5433 else
5434 Ops[j] = Builder.CreateBitCast(Ops[j], ai->getType(), name);
5435 }
5436
5437 if (F->isConstrainedFPIntrinsic())
5438 return Builder.CreateConstrainedFPCall(F, Ops, name);
5439 else
5440 return Builder.CreateCall(F, Ops, name);
5441}
5442
5443Value *CodeGenFunction::EmitNeonShiftVector(Value *V, llvm::Type *Ty,
5444 bool neg) {
5445 int SV = cast<ConstantInt>(V)->getSExtValue();
5446 return ConstantInt::get(Ty, neg ? -SV : SV);
5447}
5448
5449// Right-shift a vector by a constant.
5450Value *CodeGenFunction::EmitNeonRShiftImm(Value *Vec, Value *Shift,
5451 llvm::Type *Ty, bool usgn,
5452 const char *name) {
5453 llvm::VectorType *VTy = cast<llvm::VectorType>(Ty);
5454
5455 int ShiftAmt = cast<ConstantInt>(Shift)->getSExtValue();
5456 int EltSize = VTy->getScalarSizeInBits();
5457
5458 Vec = Builder.CreateBitCast(Vec, Ty);
5459
5460 // lshr/ashr are undefined when the shift amount is equal to the vector
5461 // element size.
5462 if (ShiftAmt == EltSize) {
5463 if (usgn) {
5464 // Right-shifting an unsigned value by its size yields 0.
5465 return llvm::ConstantAggregateZero::get(VTy);
5466 } else {
5467 // Right-shifting a signed value by its size is equivalent
5468 // to a shift of size-1.
5469 --ShiftAmt;
5470 Shift = ConstantInt::get(VTy->getElementType(), ShiftAmt);
5471 }
5472 }
5473
5474 Shift = EmitNeonShiftVector(Shift, Ty, false);
5475 if (usgn)
5476 return Builder.CreateLShr(Vec, Shift, name);
5477 else
5478 return Builder.CreateAShr(Vec, Shift, name);
5479}
5480
5481enum {
5482 AddRetType = (1 << 0),
5483 Add1ArgType = (1 << 1),
5484 Add2ArgTypes = (1 << 2),
5485
5486 VectorizeRetType = (1 << 3),
5487 VectorizeArgTypes = (1 << 4),
5488
5489 InventFloatType = (1 << 5),
5490 UnsignedAlts = (1 << 6),
5491
5492 Use64BitVectors = (1 << 7),
5493 Use128BitVectors = (1 << 8),
5494
5495 Vectorize1ArgType = Add1ArgType | VectorizeArgTypes,
5496 VectorRet = AddRetType | VectorizeRetType,
5497 VectorRetGetArgs01 =
5498 AddRetType | Add2ArgTypes | VectorizeRetType | VectorizeArgTypes,
5499 FpCmpzModifiers =
5500 AddRetType | VectorizeRetType | Add1ArgType | InventFloatType
5501};
5502
5503namespace {
5504struct ARMVectorIntrinsicInfo {
5505 const char *NameHint;
5506 unsigned BuiltinID;
5507 unsigned LLVMIntrinsic;
5508 unsigned AltLLVMIntrinsic;
5509 uint64_t TypeModifier;
5510
5511 bool operator<(unsigned RHSBuiltinID) const {
5512 return BuiltinID < RHSBuiltinID;
5513 }
5514 bool operator<(const ARMVectorIntrinsicInfo &TE) const {
5515 return BuiltinID < TE.BuiltinID;
5516 }
5517};
5518} // end anonymous namespace
5519
5520#define NEONMAP0(NameBase) \
5521 { #NameBase, NEON::BI__builtin_neon_ ## NameBase, 0, 0, 0 }
5522
5523#define NEONMAP1(NameBase, LLVMIntrinsic, TypeModifier) \
5524 { #NameBase, NEON:: BI__builtin_neon_ ## NameBase, \
5525 Intrinsic::LLVMIntrinsic, 0, TypeModifier }
5526
5527#define NEONMAP2(NameBase, LLVMIntrinsic, AltLLVMIntrinsic, TypeModifier) \
5528 { #NameBase, NEON:: BI__builtin_neon_ ## NameBase, \
5529 Intrinsic::LLVMIntrinsic, Intrinsic::AltLLVMIntrinsic, \
5530 TypeModifier }
5531
5532static const ARMVectorIntrinsicInfo ARMSIMDIntrinsicMap [] = {
5533 NEONMAP1(__a32_vcvt_bf16_v, arm_neon_vcvtfp2bf, 0),
5534 NEONMAP0(splat_lane_v),
5535 NEONMAP0(splat_laneq_v),
5536 NEONMAP0(splatq_lane_v),
5537 NEONMAP0(splatq_laneq_v),
5538 NEONMAP2(vabd_v, arm_neon_vabdu, arm_neon_vabds, Add1ArgType | UnsignedAlts),
5539 NEONMAP2(vabdq_v, arm_neon_vabdu, arm_neon_vabds, Add1ArgType | UnsignedAlts),
5540 NEONMAP1(vabs_v, arm_neon_vabs, 0),
5541 NEONMAP1(vabsq_v, arm_neon_vabs, 0),
5542 NEONMAP0(vadd_v),
5543 NEONMAP0(vaddhn_v),
5544 NEONMAP0(vaddq_v),
5545 NEONMAP1(vaesdq_v, arm_neon_aesd, 0),
5546 NEONMAP1(vaeseq_v, arm_neon_aese, 0),
5547 NEONMAP1(vaesimcq_v, arm_neon_aesimc, 0),
5548 NEONMAP1(vaesmcq_v, arm_neon_aesmc, 0),
5549 NEONMAP1(vbfdot_v, arm_neon_bfdot, 0),
5550 NEONMAP1(vbfdotq_v, arm_neon_bfdot, 0),
5551 NEONMAP1(vbfmlalbq_v, arm_neon_bfmlalb, 0),
5552 NEONMAP1(vbfmlaltq_v, arm_neon_bfmlalt, 0),
5553 NEONMAP1(vbfmmlaq_v, arm_neon_bfmmla, 0),
5554 NEONMAP1(vbsl_v, arm_neon_vbsl, AddRetType),
5555 NEONMAP1(vbslq_v, arm_neon_vbsl, AddRetType),
5556 NEONMAP1(vcadd_rot270_v, arm_neon_vcadd_rot270, Add1ArgType),
5557 NEONMAP1(vcadd_rot90_v, arm_neon_vcadd_rot90, Add1ArgType),
5558 NEONMAP1(vcaddq_rot270_v, arm_neon_vcadd_rot270, Add1ArgType),
5559 NEONMAP1(vcaddq_rot90_v, arm_neon_vcadd_rot90, Add1ArgType),
5560 NEONMAP1(vcage_v, arm_neon_vacge, 0),
5561 NEONMAP1(vcageq_v, arm_neon_vacge, 0),
5562 NEONMAP1(vcagt_v, arm_neon_vacgt, 0),
5563 NEONMAP1(vcagtq_v, arm_neon_vacgt, 0),
5564 NEONMAP1(vcale_v, arm_neon_vacge, 0),
5565 NEONMAP1(vcaleq_v, arm_neon_vacge, 0),
5566 NEONMAP1(vcalt_v, arm_neon_vacgt, 0),
5567 NEONMAP1(vcaltq_v, arm_neon_vacgt, 0),
5568 NEONMAP0(vceqz_v),
5569 NEONMAP0(vceqzq_v),
5570 NEONMAP0(vcgez_v),
5571 NEONMAP0(vcgezq_v),
5572 NEONMAP0(vcgtz_v),
5573 NEONMAP0(vcgtzq_v),
5574 NEONMAP0(vclez_v),
5575 NEONMAP0(vclezq_v),
5576 NEONMAP1(vcls_v, arm_neon_vcls, Add1ArgType),
5577 NEONMAP1(vclsq_v, arm_neon_vcls, Add1ArgType),
5578 NEONMAP0(vcltz_v),
5579 NEONMAP0(vcltzq_v),
5580 NEONMAP1(vclz_v, ctlz, Add1ArgType),
5581 NEONMAP1(vclzq_v, ctlz, Add1ArgType),
5582 NEONMAP1(vcnt_v, ctpop, Add1ArgType),
5583 NEONMAP1(vcntq_v, ctpop, Add1ArgType),
5584 NEONMAP1(vcvt_f16_f32, arm_neon_vcvtfp2hf, 0),
5585 NEONMAP0(vcvt_f16_v),
5586 NEONMAP1(vcvt_f32_f16, arm_neon_vcvthf2fp, 0),
5587 NEONMAP0(vcvt_f32_v),
5588 NEONMAP2(vcvt_n_f16_v, arm_neon_vcvtfxu2fp, arm_neon_vcvtfxs2fp, 0),
5589 NEONMAP2(vcvt_n_f32_v, arm_neon_vcvtfxu2fp, arm_neon_vcvtfxs2fp, 0),
5590 NEONMAP1(vcvt_n_s16_v, arm_neon_vcvtfp2fxs, 0),
5591 NEONMAP1(vcvt_n_s32_v, arm_neon_vcvtfp2fxs, 0),
5592 NEONMAP1(vcvt_n_s64_v, arm_neon_vcvtfp2fxs, 0),
5593 NEONMAP1(vcvt_n_u16_v, arm_neon_vcvtfp2fxu, 0),
5594 NEONMAP1(vcvt_n_u32_v, arm_neon_vcvtfp2fxu, 0),
5595 NEONMAP1(vcvt_n_u64_v, arm_neon_vcvtfp2fxu, 0),
5596 NEONMAP0(vcvt_s16_v),
5597 NEONMAP0(vcvt_s32_v),
5598 NEONMAP0(vcvt_s64_v),
5599 NEONMAP0(vcvt_u16_v),
5600 NEONMAP0(vcvt_u32_v),
5601 NEONMAP0(vcvt_u64_v),
5602 NEONMAP1(vcvta_s16_v, arm_neon_vcvtas, 0),
5603 NEONMAP1(vcvta_s32_v, arm_neon_vcvtas, 0),
5604 NEONMAP1(vcvta_s64_v, arm_neon_vcvtas, 0),
5605 NEONMAP1(vcvta_u16_v, arm_neon_vcvtau, 0),
5606 NEONMAP1(vcvta_u32_v, arm_neon_vcvtau, 0),
5607 NEONMAP1(vcvta_u64_v, arm_neon_vcvtau, 0),
5608 NEONMAP1(vcvtaq_s16_v, arm_neon_vcvtas, 0),
5609 NEONMAP1(vcvtaq_s32_v, arm_neon_vcvtas, 0),
5610 NEONMAP1(vcvtaq_s64_v, arm_neon_vcvtas, 0),
5611 NEONMAP1(vcvtaq_u16_v, arm_neon_vcvtau, 0),
5612 NEONMAP1(vcvtaq_u32_v, arm_neon_vcvtau, 0),
5613 NEONMAP1(vcvtaq_u64_v, arm_neon_vcvtau, 0),
5614 NEONMAP1(vcvth_bf16_f32, arm_neon_vcvtbfp2bf, 0),
5615 NEONMAP1(vcvtm_s16_v, arm_neon_vcvtms, 0),
5616 NEONMAP1(vcvtm_s32_v, arm_neon_vcvtms, 0),
5617 NEONMAP1(vcvtm_s64_v, arm_neon_vcvtms, 0),
5618 NEONMAP1(vcvtm_u16_v, arm_neon_vcvtmu, 0),
5619 NEONMAP1(vcvtm_u32_v, arm_neon_vcvtmu, 0),
5620 NEONMAP1(vcvtm_u64_v, arm_neon_vcvtmu, 0),
5621 NEONMAP1(vcvtmq_s16_v, arm_neon_vcvtms, 0),
5622 NEONMAP1(vcvtmq_s32_v, arm_neon_vcvtms, 0),
5623 NEONMAP1(vcvtmq_s64_v, arm_neon_vcvtms, 0),
5624 NEONMAP1(vcvtmq_u16_v, arm_neon_vcvtmu, 0),
5625 NEONMAP1(vcvtmq_u32_v, arm_neon_vcvtmu, 0),
5626 NEONMAP1(vcvtmq_u64_v, arm_neon_vcvtmu, 0),
5627 NEONMAP1(vcvtn_s16_v, arm_neon_vcvtns, 0),
5628 NEONMAP1(vcvtn_s32_v, arm_neon_vcvtns, 0),
5629 NEONMAP1(vcvtn_s64_v, arm_neon_vcvtns, 0),
5630 NEONMAP1(vcvtn_u16_v, arm_neon_vcvtnu, 0),
5631 NEONMAP1(vcvtn_u32_v, arm_neon_vcvtnu, 0),
5632 NEONMAP1(vcvtn_u64_v, arm_neon_vcvtnu, 0),
5633 NEONMAP1(vcvtnq_s16_v, arm_neon_vcvtns, 0),
5634 NEONMAP1(vcvtnq_s32_v, arm_neon_vcvtns, 0),
5635 NEONMAP1(vcvtnq_s64_v, arm_neon_vcvtns, 0),
5636 NEONMAP1(vcvtnq_u16_v, arm_neon_vcvtnu, 0),
5637 NEONMAP1(vcvtnq_u32_v, arm_neon_vcvtnu, 0),
5638 NEONMAP1(vcvtnq_u64_v, arm_neon_vcvtnu, 0),
5639 NEONMAP1(vcvtp_s16_v, arm_neon_vcvtps, 0),
5640 NEONMAP1(vcvtp_s32_v, arm_neon_vcvtps, 0),
5641 NEONMAP1(vcvtp_s64_v, arm_neon_vcvtps, 0),
5642 NEONMAP1(vcvtp_u16_v, arm_neon_vcvtpu, 0),
5643 NEONMAP1(vcvtp_u32_v, arm_neon_vcvtpu, 0),
5644 NEONMAP1(vcvtp_u64_v, arm_neon_vcvtpu, 0),
5645 NEONMAP1(vcvtpq_s16_v, arm_neon_vcvtps, 0),
5646 NEONMAP1(vcvtpq_s32_v, arm_neon_vcvtps, 0),
5647 NEONMAP1(vcvtpq_s64_v, arm_neon_vcvtps, 0),
5648 NEONMAP1(vcvtpq_u16_v, arm_neon_vcvtpu, 0),
5649 NEONMAP1(vcvtpq_u32_v, arm_neon_vcvtpu, 0),
5650 NEONMAP1(vcvtpq_u64_v, arm_neon_vcvtpu, 0),
5651 NEONMAP0(vcvtq_f16_v),
5652 NEONMAP0(vcvtq_f32_v),
5653 NEONMAP2(vcvtq_n_f16_v, arm_neon_vcvtfxu2fp, arm_neon_vcvtfxs2fp, 0),
5654 NEONMAP2(vcvtq_n_f32_v, arm_neon_vcvtfxu2fp, arm_neon_vcvtfxs2fp, 0),
5655 NEONMAP1(vcvtq_n_s16_v, arm_neon_vcvtfp2fxs, 0),
5656 NEONMAP1(vcvtq_n_s32_v, arm_neon_vcvtfp2fxs, 0),
5657 NEONMAP1(vcvtq_n_s64_v, arm_neon_vcvtfp2fxs, 0),
5658 NEONMAP1(vcvtq_n_u16_v, arm_neon_vcvtfp2fxu, 0),
5659 NEONMAP1(vcvtq_n_u32_v, arm_neon_vcvtfp2fxu, 0),
5660 NEONMAP1(vcvtq_n_u64_v, arm_neon_vcvtfp2fxu, 0),
5661 NEONMAP0(vcvtq_s16_v),
5662 NEONMAP0(vcvtq_s32_v),
5663 NEONMAP0(vcvtq_s64_v),
5664 NEONMAP0(vcvtq_u16_v),
5665 NEONMAP0(vcvtq_u32_v),
5666 NEONMAP0(vcvtq_u64_v),
5667 NEONMAP2(vdot_v, arm_neon_udot, arm_neon_sdot, 0),
5668 NEONMAP2(vdotq_v, arm_neon_udot, arm_neon_sdot, 0),
5669 NEONMAP0(vext_v),
5670 NEONMAP0(vextq_v),
5671 NEONMAP0(vfma_v),
5672 NEONMAP0(vfmaq_v),
5673 NEONMAP2(vhadd_v, arm_neon_vhaddu, arm_neon_vhadds, Add1ArgType | UnsignedAlts),
5674 NEONMAP2(vhaddq_v, arm_neon_vhaddu, arm_neon_vhadds, Add1ArgType | UnsignedAlts),
5675 NEONMAP2(vhsub_v, arm_neon_vhsubu, arm_neon_vhsubs, Add1ArgType | UnsignedAlts),
5676 NEONMAP2(vhsubq_v, arm_neon_vhsubu, arm_neon_vhsubs, Add1ArgType | UnsignedAlts),
5677 NEONMAP0(vld1_dup_v),
5678 NEONMAP1(vld1_v, arm_neon_vld1, 0),
5679 NEONMAP1(vld1_x2_v, arm_neon_vld1x2, 0),
5680 NEONMAP1(vld1_x3_v, arm_neon_vld1x3, 0),
5681 NEONMAP1(vld1_x4_v, arm_neon_vld1x4, 0),
5682 NEONMAP0(vld1q_dup_v),
5683 NEONMAP1(vld1q_v, arm_neon_vld1, 0),
5684 NEONMAP1(vld1q_x2_v, arm_neon_vld1x2, 0),
5685 NEONMAP1(vld1q_x3_v, arm_neon_vld1x3, 0),
5686 NEONMAP1(vld1q_x4_v, arm_neon_vld1x4, 0),
5687 NEONMAP1(vld2_dup_v, arm_neon_vld2dup, 0),
5688 NEONMAP1(vld2_lane_v, arm_neon_vld2lane, 0),
5689 NEONMAP1(vld2_v, arm_neon_vld2, 0),
5690 NEONMAP1(vld2q_dup_v, arm_neon_vld2dup, 0),
5691 NEONMAP1(vld2q_lane_v, arm_neon_vld2lane, 0),
5692 NEONMAP1(vld2q_v, arm_neon_vld2, 0),
5693 NEONMAP1(vld3_dup_v, arm_neon_vld3dup, 0),
5694 NEONMAP1(vld3_lane_v, arm_neon_vld3lane, 0),
5695 NEONMAP1(vld3_v, arm_neon_vld3, 0),
5696 NEONMAP1(vld3q_dup_v, arm_neon_vld3dup, 0),
5697 NEONMAP1(vld3q_lane_v, arm_neon_vld3lane, 0),
5698 NEONMAP1(vld3q_v, arm_neon_vld3, 0),
5699 NEONMAP1(vld4_dup_v, arm_neon_vld4dup, 0),
5700 NEONMAP1(vld4_lane_v, arm_neon_vld4lane, 0),
5701 NEONMAP1(vld4_v, arm_neon_vld4, 0),
5702 NEONMAP1(vld4q_dup_v, arm_neon_vld4dup, 0),
5703 NEONMAP1(vld4q_lane_v, arm_neon_vld4lane, 0),
5704 NEONMAP1(vld4q_v, arm_neon_vld4, 0),
5705 NEONMAP2(vmax_v, arm_neon_vmaxu, arm_neon_vmaxs, Add1ArgType | UnsignedAlts),
5706 NEONMAP1(vmaxnm_v, arm_neon_vmaxnm, Add1ArgType),
5707 NEONMAP1(vmaxnmq_v, arm_neon_vmaxnm, Add1ArgType),
5708 NEONMAP2(vmaxq_v, arm_neon_vmaxu, arm_neon_vmaxs, Add1ArgType | UnsignedAlts),
5709 NEONMAP2(vmin_v, arm_neon_vminu, arm_neon_vmins, Add1ArgType | UnsignedAlts),
5710 NEONMAP1(vminnm_v, arm_neon_vminnm, Add1ArgType),
5711 NEONMAP1(vminnmq_v, arm_neon_vminnm, Add1ArgType),
5712 NEONMAP2(vminq_v, arm_neon_vminu, arm_neon_vmins, Add1ArgType | UnsignedAlts),
5713 NEONMAP2(vmmlaq_v, arm_neon_ummla, arm_neon_smmla, 0),
5714 NEONMAP0(vmovl_v),
5715 NEONMAP0(vmovn_v),
5716 NEONMAP1(vmul_v, arm_neon_vmulp, Add1ArgType),
5717 NEONMAP0(vmull_v),
5718 NEONMAP1(vmulq_v, arm_neon_vmulp, Add1ArgType),
5719 NEONMAP2(vpadal_v, arm_neon_vpadalu, arm_neon_vpadals, UnsignedAlts),
5720 NEONMAP2(vpadalq_v, arm_neon_vpadalu, arm_neon_vpadals, UnsignedAlts),
5721 NEONMAP1(vpadd_v, arm_neon_vpadd, Add1ArgType),
5722 NEONMAP2(vpaddl_v, arm_neon_vpaddlu, arm_neon_vpaddls, UnsignedAlts),
5723 NEONMAP2(vpaddlq_v, arm_neon_vpaddlu, arm_neon_vpaddls, UnsignedAlts),
5724 NEONMAP1(vpaddq_v, arm_neon_vpadd, Add1ArgType),
5725 NEONMAP2(vpmax_v, arm_neon_vpmaxu, arm_neon_vpmaxs, Add1ArgType | UnsignedAlts),
5726 NEONMAP2(vpmin_v, arm_neon_vpminu, arm_neon_vpmins, Add1ArgType | UnsignedAlts),
5727 NEONMAP1(vqabs_v, arm_neon_vqabs, Add1ArgType),
5728 NEONMAP1(vqabsq_v, arm_neon_vqabs, Add1ArgType),
5729 NEONMAP2(vqadd_v, uadd_sat, sadd_sat, Add1ArgType | UnsignedAlts),
5730 NEONMAP2(vqaddq_v, uadd_sat, sadd_sat, Add1ArgType | UnsignedAlts),
5731 NEONMAP2(vqdmlal_v, arm_neon_vqdmull, sadd_sat, 0),
5732 NEONMAP2(vqdmlsl_v, arm_neon_vqdmull, ssub_sat, 0),
5733 NEONMAP1(vqdmulh_v, arm_neon_vqdmulh, Add1ArgType),
5734 NEONMAP1(vqdmulhq_v, arm_neon_vqdmulh, Add1ArgType),
5735 NEONMAP1(vqdmull_v, arm_neon_vqdmull, Add1ArgType),
5736 NEONMAP2(vqmovn_v, arm_neon_vqmovnu, arm_neon_vqmovns, Add1ArgType | UnsignedAlts),
5737 NEONMAP1(vqmovun_v, arm_neon_vqmovnsu, Add1ArgType),
5738 NEONMAP1(vqneg_v, arm_neon_vqneg, Add1ArgType),
5739 NEONMAP1(vqnegq_v, arm_neon_vqneg, Add1ArgType),
5740 NEONMAP1(vqrdmulh_v, arm_neon_vqrdmulh, Add1ArgType),
5741 NEONMAP1(vqrdmulhq_v, arm_neon_vqrdmulh, Add1ArgType),
5742 NEONMAP2(vqrshl_v, arm_neon_vqrshiftu, arm_neon_vqrshifts, Add1ArgType | UnsignedAlts),
5743 NEONMAP2(vqrshlq_v, arm_neon_vqrshiftu, arm_neon_vqrshifts, Add1ArgType | UnsignedAlts),
5744 NEONMAP2(vqshl_n_v, arm_neon_vqshiftu, arm_neon_vqshifts, UnsignedAlts),
5745 NEONMAP2(vqshl_v, arm_neon_vqshiftu, arm_neon_vqshifts, Add1ArgType | UnsignedAlts),
5746 NEONMAP2(vqshlq_n_v, arm_neon_vqshiftu, arm_neon_vqshifts, UnsignedAlts),
5747 NEONMAP2(vqshlq_v, arm_neon_vqshiftu, arm_neon_vqshifts, Add1ArgType | UnsignedAlts),
5748 NEONMAP1(vqshlu_n_v, arm_neon_vqshiftsu, 0),
5749 NEONMAP1(vqshluq_n_v, arm_neon_vqshiftsu, 0),
5750 NEONMAP2(vqsub_v, usub_sat, ssub_sat, Add1ArgType | UnsignedAlts),
5751 NEONMAP2(vqsubq_v, usub_sat, ssub_sat, Add1ArgType | UnsignedAlts),
5752 NEONMAP1(vraddhn_v, arm_neon_vraddhn, Add1ArgType),
5753 NEONMAP2(vrecpe_v, arm_neon_vrecpe, arm_neon_vrecpe, 0),
5754 NEONMAP2(vrecpeq_v, arm_neon_vrecpe, arm_neon_vrecpe, 0),
5755 NEONMAP1(vrecps_v, arm_neon_vrecps, Add1ArgType),
5756 NEONMAP1(vrecpsq_v, arm_neon_vrecps, Add1ArgType),
5757 NEONMAP2(vrhadd_v, arm_neon_vrhaddu, arm_neon_vrhadds, Add1ArgType | UnsignedAlts),
5758 NEONMAP2(vrhaddq_v, arm_neon_vrhaddu, arm_neon_vrhadds, Add1ArgType | UnsignedAlts),
5759 NEONMAP1(vrnd_v, arm_neon_vrintz, Add1ArgType),
5760 NEONMAP1(vrnda_v, arm_neon_vrinta, Add1ArgType),
5761 NEONMAP1(vrndaq_v, arm_neon_vrinta, Add1ArgType),
5762 NEONMAP0(vrndi_v),
5763 NEONMAP0(vrndiq_v),
5764 NEONMAP1(vrndm_v, arm_neon_vrintm, Add1ArgType),
5765 NEONMAP1(vrndmq_v, arm_neon_vrintm, Add1ArgType),
5766 NEONMAP1(vrndn_v, arm_neon_vrintn, Add1ArgType),
5767 NEONMAP1(vrndnq_v, arm_neon_vrintn, Add1ArgType),
5768 NEONMAP1(vrndp_v, arm_neon_vrintp, Add1ArgType),
5769 NEONMAP1(vrndpq_v, arm_neon_vrintp, Add1ArgType),
5770 NEONMAP1(vrndq_v, arm_neon_vrintz, Add1ArgType),
5771 NEONMAP1(vrndx_v, arm_neon_vrintx, Add1ArgType),
5772 NEONMAP1(vrndxq_v, arm_neon_vrintx, Add1ArgType),
5773 NEONMAP2(vrshl_v, arm_neon_vrshiftu, arm_neon_vrshifts, Add1ArgType | UnsignedAlts),
5774 NEONMAP2(vrshlq_v, arm_neon_vrshiftu, arm_neon_vrshifts, Add1ArgType | UnsignedAlts),
5775 NEONMAP2(vrshr_n_v, arm_neon_vrshiftu, arm_neon_vrshifts, UnsignedAlts),
5776 NEONMAP2(vrshrq_n_v, arm_neon_vrshiftu, arm_neon_vrshifts, UnsignedAlts),
5777 NEONMAP2(vrsqrte_v, arm_neon_vrsqrte, arm_neon_vrsqrte, 0),
5778 NEONMAP2(vrsqrteq_v, arm_neon_vrsqrte, arm_neon_vrsqrte, 0),
5779 NEONMAP1(vrsqrts_v, arm_neon_vrsqrts, Add1ArgType),
5780 NEONMAP1(vrsqrtsq_v, arm_neon_vrsqrts, Add1ArgType),
5781 NEONMAP1(vrsubhn_v, arm_neon_vrsubhn, Add1ArgType),
5782 NEONMAP1(vsha1su0q_v, arm_neon_sha1su0, 0),
5783 NEONMAP1(vsha1su1q_v, arm_neon_sha1su1, 0),
5784 NEONMAP1(vsha256h2q_v, arm_neon_sha256h2, 0),
5785 NEONMAP1(vsha256hq_v, arm_neon_sha256h, 0),
5786 NEONMAP1(vsha256su0q_v, arm_neon_sha256su0, 0),
5787 NEONMAP1(vsha256su1q_v, arm_neon_sha256su1, 0),
5788 NEONMAP0(vshl_n_v),
5789 NEONMAP2(vshl_v, arm_neon_vshiftu, arm_neon_vshifts, Add1ArgType | UnsignedAlts),
5790 NEONMAP0(vshll_n_v),
5791 NEONMAP0(vshlq_n_v),
5792 NEONMAP2(vshlq_v, arm_neon_vshiftu, arm_neon_vshifts, Add1ArgType | UnsignedAlts),
5793 NEONMAP0(vshr_n_v),
5794 NEONMAP0(vshrn_n_v),
5795 NEONMAP0(vshrq_n_v),
5796 NEONMAP1(vst1_v, arm_neon_vst1, 0),
5797 NEONMAP1(vst1_x2_v, arm_neon_vst1x2, 0),
5798 NEONMAP1(vst1_x3_v, arm_neon_vst1x3, 0),
5799 NEONMAP1(vst1_x4_v, arm_neon_vst1x4, 0),
5800 NEONMAP1(vst1q_v, arm_neon_vst1, 0),
5801 NEONMAP1(vst1q_x2_v, arm_neon_vst1x2, 0),
5802 NEONMAP1(vst1q_x3_v, arm_neon_vst1x3, 0),
5803 NEONMAP1(vst1q_x4_v, arm_neon_vst1x4, 0),
5804 NEONMAP1(vst2_lane_v, arm_neon_vst2lane, 0),
5805 NEONMAP1(vst2_v, arm_neon_vst2, 0),
5806 NEONMAP1(vst2q_lane_v, arm_neon_vst2lane, 0),
5807 NEONMAP1(vst2q_v, arm_neon_vst2, 0),
5808 NEONMAP1(vst3_lane_v, arm_neon_vst3lane, 0),
5809 NEONMAP1(vst3_v, arm_neon_vst3, 0),
5810 NEONMAP1(vst3q_lane_v, arm_neon_vst3lane, 0),
5811 NEONMAP1(vst3q_v, arm_neon_vst3, 0),
5812 NEONMAP1(vst4_lane_v, arm_neon_vst4lane, 0),
5813 NEONMAP1(vst4_v, arm_neon_vst4, 0),
5814 NEONMAP1(vst4q_lane_v, arm_neon_vst4lane, 0),
5815 NEONMAP1(vst4q_v, arm_neon_vst4, 0),
5816 NEONMAP0(vsubhn_v),
5817 NEONMAP0(vtrn_v),
5818 NEONMAP0(vtrnq_v),
5819 NEONMAP0(vtst_v),
5820 NEONMAP0(vtstq_v),
5821 NEONMAP1(vusdot_v, arm_neon_usdot, 0),
5822 NEONMAP1(vusdotq_v, arm_neon_usdot, 0),
5823 NEONMAP1(vusmmlaq_v, arm_neon_usmmla, 0),
5824 NEONMAP0(vuzp_v),
5825 NEONMAP0(vuzpq_v),
5826 NEONMAP0(vzip_v),
5827 NEONMAP0(vzipq_v)
5828};
5829
5830static const ARMVectorIntrinsicInfo AArch64SIMDIntrinsicMap[] = {
5831 NEONMAP1(__a64_vcvtq_low_bf16_v, aarch64_neon_bfcvtn, 0),
5832 NEONMAP0(splat_lane_v),
5833 NEONMAP0(splat_laneq_v),
5834 NEONMAP0(splatq_lane_v),
5835 NEONMAP0(splatq_laneq_v),
5836 NEONMAP1(vabs_v, aarch64_neon_abs, 0),
5837 NEONMAP1(vabsq_v, aarch64_neon_abs, 0),
5838 NEONMAP0(vadd_v),
5839 NEONMAP0(vaddhn_v),
5840 NEONMAP0(vaddq_p128),
5841 NEONMAP0(vaddq_v),
5842 NEONMAP1(vaesdq_v, aarch64_crypto_aesd, 0),
5843 NEONMAP1(vaeseq_v, aarch64_crypto_aese, 0),
5844 NEONMAP1(vaesimcq_v, aarch64_crypto_aesimc, 0),
5845 NEONMAP1(vaesmcq_v, aarch64_crypto_aesmc, 0),
5846 NEONMAP2(vbcaxq_v, aarch64_crypto_bcaxu, aarch64_crypto_bcaxs, Add1ArgType | UnsignedAlts),
5847 NEONMAP1(vbfdot_v, aarch64_neon_bfdot, 0),
5848 NEONMAP1(vbfdotq_v, aarch64_neon_bfdot, 0),
5849 NEONMAP1(vbfmlalbq_v, aarch64_neon_bfmlalb, 0),
5850 NEONMAP1(vbfmlaltq_v, aarch64_neon_bfmlalt, 0),
5851 NEONMAP1(vbfmmlaq_v, aarch64_neon_bfmmla, 0),
5852 NEONMAP1(vcadd_rot270_v, aarch64_neon_vcadd_rot270, Add1ArgType),
5853 NEONMAP1(vcadd_rot90_v, aarch64_neon_vcadd_rot90, Add1ArgType),
5854 NEONMAP1(vcaddq_rot270_v, aarch64_neon_vcadd_rot270, Add1ArgType),
5855 NEONMAP1(vcaddq_rot90_v, aarch64_neon_vcadd_rot90, Add1ArgType),
5856 NEONMAP1(vcage_v, aarch64_neon_facge, 0),
5857 NEONMAP1(vcageq_v, aarch64_neon_facge, 0),
5858 NEONMAP1(vcagt_v, aarch64_neon_facgt, 0),
5859 NEONMAP1(vcagtq_v, aarch64_neon_facgt, 0),
5860 NEONMAP1(vcale_v, aarch64_neon_facge, 0),
5861 NEONMAP1(vcaleq_v, aarch64_neon_facge, 0),
5862 NEONMAP1(vcalt_v, aarch64_neon_facgt, 0),
5863 NEONMAP1(vcaltq_v, aarch64_neon_facgt, 0),
5864 NEONMAP0(vceqz_v),
5865 NEONMAP0(vceqzq_v),
5866 NEONMAP0(vcgez_v),
5867 NEONMAP0(vcgezq_v),
5868 NEONMAP0(vcgtz_v),
5869 NEONMAP0(vcgtzq_v),
5870 NEONMAP0(vclez_v),
5871 NEONMAP0(vclezq_v),
5872 NEONMAP1(vcls_v, aarch64_neon_cls, Add1ArgType),
5873 NEONMAP1(vclsq_v, aarch64_neon_cls, Add1ArgType),
5874 NEONMAP0(vcltz_v),
5875 NEONMAP0(vcltzq_v),
5876 NEONMAP1(vclz_v, ctlz, Add1ArgType),
5877 NEONMAP1(vclzq_v, ctlz, Add1ArgType),
5878 NEONMAP1(vcmla_rot180_v, aarch64_neon_vcmla_rot180, Add1ArgType),
5879 NEONMAP1(vcmla_rot270_v, aarch64_neon_vcmla_rot270, Add1ArgType),
5880 NEONMAP1(vcmla_rot90_v, aarch64_neon_vcmla_rot90, Add1ArgType),
5881 NEONMAP1(vcmla_v, aarch64_neon_vcmla_rot0, Add1ArgType),
5882 NEONMAP1(vcmlaq_rot180_v, aarch64_neon_vcmla_rot180, Add1ArgType),
5883 NEONMAP1(vcmlaq_rot270_v, aarch64_neon_vcmla_rot270, Add1ArgType),
5884 NEONMAP1(vcmlaq_rot90_v, aarch64_neon_vcmla_rot90, Add1ArgType),
5885 NEONMAP1(vcmlaq_v, aarch64_neon_vcmla_rot0, Add1ArgType),
5886 NEONMAP1(vcnt_v, ctpop, Add1ArgType),
5887 NEONMAP1(vcntq_v, ctpop, Add1ArgType),
5888 NEONMAP1(vcvt_f16_f32, aarch64_neon_vcvtfp2hf, 0),
5889 NEONMAP0(vcvt_f16_v),
5890 NEONMAP1(vcvt_f32_f16, aarch64_neon_vcvthf2fp, 0),
5891 NEONMAP0(vcvt_f32_v),
5892 NEONMAP2(vcvt_n_f16_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0),
5893 NEONMAP2(vcvt_n_f32_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0),
5894 NEONMAP2(vcvt_n_f64_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0),
5895 NEONMAP1(vcvt_n_s16_v, aarch64_neon_vcvtfp2fxs, 0),
5896 NEONMAP1(vcvt_n_s32_v, aarch64_neon_vcvtfp2fxs, 0),
5897 NEONMAP1(vcvt_n_s64_v, aarch64_neon_vcvtfp2fxs, 0),
5898 NEONMAP1(vcvt_n_u16_v, aarch64_neon_vcvtfp2fxu, 0),
5899 NEONMAP1(vcvt_n_u32_v, aarch64_neon_vcvtfp2fxu, 0),
5900 NEONMAP1(vcvt_n_u64_v, aarch64_neon_vcvtfp2fxu, 0),
5901 NEONMAP0(vcvtq_f16_v),
5902 NEONMAP0(vcvtq_f32_v),
5903 NEONMAP1(vcvtq_high_bf16_v, aarch64_neon_bfcvtn2, 0),
5904 NEONMAP2(vcvtq_n_f16_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0),
5905 NEONMAP2(vcvtq_n_f32_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0),
5906 NEONMAP2(vcvtq_n_f64_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0),
5907 NEONMAP1(vcvtq_n_s16_v, aarch64_neon_vcvtfp2fxs, 0),
5908 NEONMAP1(vcvtq_n_s32_v, aarch64_neon_vcvtfp2fxs, 0),
5909 NEONMAP1(vcvtq_n_s64_v, aarch64_neon_vcvtfp2fxs, 0),
5910 NEONMAP1(vcvtq_n_u16_v, aarch64_neon_vcvtfp2fxu, 0),
5911 NEONMAP1(vcvtq_n_u32_v, aarch64_neon_vcvtfp2fxu, 0),
5912 NEONMAP1(vcvtq_n_u64_v, aarch64_neon_vcvtfp2fxu, 0),
5913 NEONMAP1(vcvtx_f32_v, aarch64_neon_fcvtxn, AddRetType | Add1ArgType),
5914 NEONMAP2(vdot_v, aarch64_neon_udot, aarch64_neon_sdot, 0),
5915 NEONMAP2(vdotq_v, aarch64_neon_udot, aarch64_neon_sdot, 0),
5916 NEONMAP2(veor3q_v, aarch64_crypto_eor3u, aarch64_crypto_eor3s, Add1ArgType | UnsignedAlts),
5917 NEONMAP0(vext_v),
5918 NEONMAP0(vextq_v),
5919 NEONMAP0(vfma_v),
5920 NEONMAP0(vfmaq_v),
5921 NEONMAP1(vfmlal_high_v, aarch64_neon_fmlal2, 0),
5922 NEONMAP1(vfmlal_low_v, aarch64_neon_fmlal, 0),
5923 NEONMAP1(vfmlalq_high_v, aarch64_neon_fmlal2, 0),
5924 NEONMAP1(vfmlalq_low_v, aarch64_neon_fmlal, 0),
5925 NEONMAP1(vfmlsl_high_v, aarch64_neon_fmlsl2, 0),
5926 NEONMAP1(vfmlsl_low_v, aarch64_neon_fmlsl, 0),
5927 NEONMAP1(vfmlslq_high_v, aarch64_neon_fmlsl2, 0),
5928 NEONMAP1(vfmlslq_low_v, aarch64_neon_fmlsl, 0),
5929 NEONMAP2(vhadd_v, aarch64_neon_uhadd, aarch64_neon_shadd, Add1ArgType | UnsignedAlts),
5930 NEONMAP2(vhaddq_v, aarch64_neon_uhadd, aarch64_neon_shadd, Add1ArgType | UnsignedAlts),
5931 NEONMAP2(vhsub_v, aarch64_neon_uhsub, aarch64_neon_shsub, Add1ArgType | UnsignedAlts),
5932 NEONMAP2(vhsubq_v, aarch64_neon_uhsub, aarch64_neon_shsub, Add1ArgType | UnsignedAlts),
5933 NEONMAP1(vld1_x2_v, aarch64_neon_ld1x2, 0),
5934 NEONMAP1(vld1_x3_v, aarch64_neon_ld1x3, 0),
5935 NEONMAP1(vld1_x4_v, aarch64_neon_ld1x4, 0),
5936 NEONMAP1(vld1q_x2_v, aarch64_neon_ld1x2, 0),
5937 NEONMAP1(vld1q_x3_v, aarch64_neon_ld1x3, 0),
5938 NEONMAP1(vld1q_x4_v, aarch64_neon_ld1x4, 0),
5939 NEONMAP2(vmmlaq_v, aarch64_neon_ummla, aarch64_neon_smmla, 0),
5940 NEONMAP0(vmovl_v),
5941 NEONMAP0(vmovn_v),
5942 NEONMAP1(vmul_v, aarch64_neon_pmul, Add1ArgType),
5943 NEONMAP1(vmulq_v, aarch64_neon_pmul, Add1ArgType),
5944 NEONMAP1(vpadd_v, aarch64_neon_addp, Add1ArgType),
5945 NEONMAP2(vpaddl_v, aarch64_neon_uaddlp, aarch64_neon_saddlp, UnsignedAlts),
5946 NEONMAP2(vpaddlq_v, aarch64_neon_uaddlp, aarch64_neon_saddlp, UnsignedAlts),
5947 NEONMAP1(vpaddq_v, aarch64_neon_addp, Add1ArgType),
5948 NEONMAP1(vqabs_v, aarch64_neon_sqabs, Add1ArgType),
5949 NEONMAP1(vqabsq_v, aarch64_neon_sqabs, Add1ArgType),
5950 NEONMAP2(vqadd_v, aarch64_neon_uqadd, aarch64_neon_sqadd, Add1ArgType | UnsignedAlts),
5951 NEONMAP2(vqaddq_v, aarch64_neon_uqadd, aarch64_neon_sqadd, Add1ArgType | UnsignedAlts),
5952 NEONMAP2(vqdmlal_v, aarch64_neon_sqdmull, aarch64_neon_sqadd, 0),
5953 NEONMAP2(vqdmlsl_v, aarch64_neon_sqdmull, aarch64_neon_sqsub, 0),
5954 NEONMAP1(vqdmulh_lane_v, aarch64_neon_sqdmulh_lane, 0),
5955 NEONMAP1(vqdmulh_laneq_v, aarch64_neon_sqdmulh_laneq, 0),
5956 NEONMAP1(vqdmulh_v, aarch64_neon_sqdmulh, Add1ArgType),
5957 NEONMAP1(vqdmulhq_lane_v, aarch64_neon_sqdmulh_lane, 0),
5958 NEONMAP1(vqdmulhq_laneq_v, aarch64_neon_sqdmulh_laneq, 0),
5959 NEONMAP1(vqdmulhq_v, aarch64_neon_sqdmulh, Add1ArgType),
5960 NEONMAP1(vqdmull_v, aarch64_neon_sqdmull, Add1ArgType),
5961 NEONMAP2(vqmovn_v, aarch64_neon_uqxtn, aarch64_neon_sqxtn, Add1ArgType | UnsignedAlts),
5962 NEONMAP1(vqmovun_v, aarch64_neon_sqxtun, Add1ArgType),
5963 NEONMAP1(vqneg_v, aarch64_neon_sqneg, Add1ArgType),
5964 NEONMAP1(vqnegq_v, aarch64_neon_sqneg, Add1ArgType),
5965 NEONMAP1(vqrdmulh_lane_v, aarch64_neon_sqrdmulh_lane, 0),
5966 NEONMAP1(vqrdmulh_laneq_v, aarch64_neon_sqrdmulh_laneq, 0),
5967 NEONMAP1(vqrdmulh_v, aarch64_neon_sqrdmulh, Add1ArgType),
5968 NEONMAP1(vqrdmulhq_lane_v, aarch64_neon_sqrdmulh_lane, 0),
5969 NEONMAP1(vqrdmulhq_laneq_v, aarch64_neon_sqrdmulh_laneq, 0),
5970 NEONMAP1(vqrdmulhq_v, aarch64_neon_sqrdmulh, Add1ArgType),
5971 NEONMAP2(vqrshl_v, aarch64_neon_uqrshl, aarch64_neon_sqrshl, Add1ArgType | UnsignedAlts),
5972 NEONMAP2(vqrshlq_v, aarch64_neon_uqrshl, aarch64_neon_sqrshl, Add1ArgType | UnsignedAlts),
5973 NEONMAP2(vqshl_n_v, aarch64_neon_uqshl, aarch64_neon_sqshl, UnsignedAlts),
5974 NEONMAP2(vqshl_v, aarch64_neon_uqshl, aarch64_neon_sqshl, Add1ArgType | UnsignedAlts),
5975 NEONMAP2(vqshlq_n_v, aarch64_neon_uqshl, aarch64_neon_sqshl,UnsignedAlts),
5976 NEONMAP2(vqshlq_v, aarch64_neon_uqshl, aarch64_neon_sqshl, Add1ArgType | UnsignedAlts),
5977 NEONMAP1(vqshlu_n_v, aarch64_neon_sqshlu, 0),
5978 NEONMAP1(vqshluq_n_v, aarch64_neon_sqshlu, 0),
5979 NEONMAP2(vqsub_v, aarch64_neon_uqsub, aarch64_neon_sqsub, Add1ArgType | UnsignedAlts),
5980 NEONMAP2(vqsubq_v, aarch64_neon_uqsub, aarch64_neon_sqsub, Add1ArgType | UnsignedAlts),
5981 NEONMAP1(vraddhn_v, aarch64_neon_raddhn, Add1ArgType),
5982 NEONMAP1(vrax1q_v, aarch64_crypto_rax1, 0),
5983 NEONMAP2(vrecpe_v, aarch64_neon_frecpe, aarch64_neon_urecpe, 0),
5984 NEONMAP2(vrecpeq_v, aarch64_neon_frecpe, aarch64_neon_urecpe, 0),
5985 NEONMAP1(vrecps_v, aarch64_neon_frecps, Add1ArgType),
5986 NEONMAP1(vrecpsq_v, aarch64_neon_frecps, Add1ArgType),
5987 NEONMAP2(vrhadd_v, aarch64_neon_urhadd, aarch64_neon_srhadd, Add1ArgType | UnsignedAlts),
5988 NEONMAP2(vrhaddq_v, aarch64_neon_urhadd, aarch64_neon_srhadd, Add1ArgType | UnsignedAlts),
5989 NEONMAP1(vrnd32x_v, aarch64_neon_frint32x, Add1ArgType),
5990 NEONMAP1(vrnd32xq_v, aarch64_neon_frint32x, Add1ArgType),
5991 NEONMAP1(vrnd32z_v, aarch64_neon_frint32z, Add1ArgType),
5992 NEONMAP1(vrnd32zq_v, aarch64_neon_frint32z, Add1ArgType),
5993 NEONMAP1(vrnd64x_v, aarch64_neon_frint64x, Add1ArgType),
5994 NEONMAP1(vrnd64xq_v, aarch64_neon_frint64x, Add1ArgType),
5995 NEONMAP1(vrnd64z_v, aarch64_neon_frint64z, Add1ArgType),
5996 NEONMAP1(vrnd64zq_v, aarch64_neon_frint64z, Add1ArgType),
5997 NEONMAP0(vrndi_v),
5998 NEONMAP0(vrndiq_v),
5999 NEONMAP2(vrshl_v, aarch64_neon_urshl, aarch64_neon_srshl, Add1ArgType | UnsignedAlts),
6000 NEONMAP2(vrshlq_v, aarch64_neon_urshl, aarch64_neon_srshl, Add1ArgType | UnsignedAlts),
6001 NEONMAP2(vrshr_n_v, aarch64_neon_urshl, aarch64_neon_srshl, UnsignedAlts),
6002 NEONMAP2(vrshrq_n_v, aarch64_neon_urshl, aarch64_neon_srshl, UnsignedAlts),
6003 NEONMAP2(vrsqrte_v, aarch64_neon_frsqrte, aarch64_neon_ursqrte, 0),
6004 NEONMAP2(vrsqrteq_v, aarch64_neon_frsqrte, aarch64_neon_ursqrte, 0),
6005 NEONMAP1(vrsqrts_v, aarch64_neon_frsqrts, Add1ArgType),
6006 NEONMAP1(vrsqrtsq_v, aarch64_neon_frsqrts, Add1ArgType),
6007 NEONMAP1(vrsubhn_v, aarch64_neon_rsubhn, Add1ArgType),
6008 NEONMAP1(vsha1su0q_v, aarch64_crypto_sha1su0, 0),
6009 NEONMAP1(vsha1su1q_v, aarch64_crypto_sha1su1, 0),
6010 NEONMAP1(vsha256h2q_v, aarch64_crypto_sha256h2, 0),
6011 NEONMAP1(vsha256hq_v, aarch64_crypto_sha256h, 0),
6012 NEONMAP1(vsha256su0q_v, aarch64_crypto_sha256su0, 0),
6013 NEONMAP1(vsha256su1q_v, aarch64_crypto_sha256su1, 0),
6014 NEONMAP1(vsha512h2q_v, aarch64_crypto_sha512h2, 0),
6015 NEONMAP1(vsha512hq_v, aarch64_crypto_sha512h, 0),
6016 NEONMAP1(vsha512su0q_v, aarch64_crypto_sha512su0, 0),
6017 NEONMAP1(vsha512su1q_v, aarch64_crypto_sha512su1, 0),
6018 NEONMAP0(vshl_n_v),
6019 NEONMAP2(vshl_v, aarch64_neon_ushl, aarch64_neon_sshl, Add1ArgType | UnsignedAlts),
6020 NEONMAP0(vshll_n_v),
6021 NEONMAP0(vshlq_n_v),
6022 NEONMAP2(vshlq_v, aarch64_neon_ushl, aarch64_neon_sshl, Add1ArgType | UnsignedAlts),
6023 NEONMAP0(vshr_n_v),
6024 NEONMAP0(vshrn_n_v),
6025 NEONMAP0(vshrq_n_v),
6026 NEONMAP1(vsm3partw1q_v, aarch64_crypto_sm3partw1, 0),
6027 NEONMAP1(vsm3partw2q_v, aarch64_crypto_sm3partw2, 0),
6028 NEONMAP1(vsm3ss1q_v, aarch64_crypto_sm3ss1, 0),
6029 NEONMAP1(vsm3tt1aq_v, aarch64_crypto_sm3tt1a, 0),
6030 NEONMAP1(vsm3tt1bq_v, aarch64_crypto_sm3tt1b, 0),
6031 NEONMAP1(vsm3tt2aq_v, aarch64_crypto_sm3tt2a, 0),
6032 NEONMAP1(vsm3tt2bq_v, aarch64_crypto_sm3tt2b, 0),
6033 NEONMAP1(vsm4ekeyq_v, aarch64_crypto_sm4ekey, 0),
6034 NEONMAP1(vsm4eq_v, aarch64_crypto_sm4e, 0),
6035 NEONMAP1(vst1_x2_v, aarch64_neon_st1x2, 0),
6036 NEONMAP1(vst1_x3_v, aarch64_neon_st1x3, 0),
6037 NEONMAP1(vst1_x4_v, aarch64_neon_st1x4, 0),
6038 NEONMAP1(vst1q_x2_v, aarch64_neon_st1x2, 0),
6039 NEONMAP1(vst1q_x3_v, aarch64_neon_st1x3, 0),
6040 NEONMAP1(vst1q_x4_v, aarch64_neon_st1x4, 0),
6041 NEONMAP0(vsubhn_v),
6042 NEONMAP0(vtst_v),
6043 NEONMAP0(vtstq_v),
6044 NEONMAP1(vusdot_v, aarch64_neon_usdot, 0),
6045 NEONMAP1(vusdotq_v, aarch64_neon_usdot, 0),
6046 NEONMAP1(vusmmlaq_v, aarch64_neon_usmmla, 0),
6047 NEONMAP1(vxarq_v, aarch64_crypto_xar, 0),
6048};
6049
6050static const ARMVectorIntrinsicInfo AArch64SISDIntrinsicMap[] = {
6051 NEONMAP1(vabdd_f64, aarch64_sisd_fabd, Add1ArgType),
6052 NEONMAP1(vabds_f32, aarch64_sisd_fabd, Add1ArgType),
6053 NEONMAP1(vabsd_s64, aarch64_neon_abs, Add1ArgType),
6054 NEONMAP1(vaddlv_s32, aarch64_neon_saddlv, AddRetType | Add1ArgType),
6055 NEONMAP1(vaddlv_u32, aarch64_neon_uaddlv, AddRetType | Add1ArgType),
6056 NEONMAP1(vaddlvq_s32, aarch64_neon_saddlv, AddRetType | Add1ArgType),
6057 NEONMAP1(vaddlvq_u32, aarch64_neon_uaddlv, AddRetType | Add1ArgType),
6058 NEONMAP1(vaddv_f32, aarch64_neon_faddv, AddRetType | Add1ArgType),
6059 NEONMAP1(vaddv_s32, aarch64_neon_saddv, AddRetType | Add1ArgType),
6060 NEONMAP1(vaddv_u32, aarch64_neon_uaddv, AddRetType | Add1ArgType),
6061 NEONMAP1(vaddvq_f32, aarch64_neon_faddv, AddRetType | Add1ArgType),
6062 NEONMAP1(vaddvq_f64, aarch64_neon_faddv, AddRetType | Add1ArgType),
6063 NEONMAP1(vaddvq_s32, aarch64_neon_saddv, AddRetType | Add1ArgType),
6064 NEONMAP1(vaddvq_s64, aarch64_neon_saddv, AddRetType | Add1ArgType),
6065 NEONMAP1(vaddvq_u32, aarch64_neon_uaddv, AddRetType | Add1ArgType),
6066 NEONMAP1(vaddvq_u64, aarch64_neon_uaddv, AddRetType | Add1ArgType),
6067 NEONMAP1(vcaged_f64, aarch64_neon_facge, AddRetType | Add1ArgType),
6068 NEONMAP1(vcages_f32, aarch64_neon_facge, AddRetType | Add1ArgType),
6069 NEONMAP1(vcagtd_f64, aarch64_neon_facgt, AddRetType | Add1ArgType),
6070 NEONMAP1(vcagts_f32, aarch64_neon_facgt, AddRetType | Add1ArgType),
6071 NEONMAP1(vcaled_f64, aarch64_neon_facge, AddRetType | Add1ArgType),
6072 NEONMAP1(vcales_f32, aarch64_neon_facge, AddRetType | Add1ArgType),
6073 NEONMAP1(vcaltd_f64, aarch64_neon_facgt, AddRetType | Add1ArgType),
6074 NEONMAP1(vcalts_f32, aarch64_neon_facgt, AddRetType | Add1ArgType),
6075 NEONMAP1(vcvtad_s64_f64, aarch64_neon_fcvtas, AddRetType | Add1ArgType),
6076 NEONMAP1(vcvtad_u64_f64, aarch64_neon_fcvtau, AddRetType | Add1ArgType),
6077 NEONMAP1(vcvtas_s32_f32, aarch64_neon_fcvtas, AddRetType | Add1ArgType),
6078 NEONMAP1(vcvtas_u32_f32, aarch64_neon_fcvtau, AddRetType | Add1ArgType),
6079 NEONMAP1(vcvtd_n_f64_s64, aarch64_neon_vcvtfxs2fp, AddRetType | Add1ArgType),
6080 NEONMAP1(vcvtd_n_f64_u64, aarch64_neon_vcvtfxu2fp, AddRetType | Add1ArgType),
6081 NEONMAP1(vcvtd_n_s64_f64, aarch64_neon_vcvtfp2fxs, AddRetType | Add1ArgType),
6082 NEONMAP1(vcvtd_n_u64_f64, aarch64_neon_vcvtfp2fxu, AddRetType | Add1ArgType),
6083 NEONMAP1(vcvtd_s64_f64, aarch64_neon_fcvtzs, AddRetType | Add1ArgType),
6084 NEONMAP1(vcvtd_u64_f64, aarch64_neon_fcvtzu, AddRetType | Add1ArgType),
6085 NEONMAP1(vcvth_bf16_f32, aarch64_neon_bfcvt, 0),
6086 NEONMAP1(vcvtmd_s64_f64, aarch64_neon_fcvtms, AddRetType | Add1ArgType),
6087 NEONMAP1(vcvtmd_u64_f64, aarch64_neon_fcvtmu, AddRetType | Add1ArgType),
6088 NEONMAP1(vcvtms_s32_f32, aarch64_neon_fcvtms, AddRetType | Add1ArgType),
6089 NEONMAP1(vcvtms_u32_f32, aarch64_neon_fcvtmu, AddRetType | Add1ArgType),
6090 NEONMAP1(vcvtnd_s64_f64, aarch64_neon_fcvtns, AddRetType | Add1ArgType),
6091 NEONMAP1(vcvtnd_u64_f64, aarch64_neon_fcvtnu, AddRetType | Add1ArgType),
6092 NEONMAP1(vcvtns_s32_f32, aarch64_neon_fcvtns, AddRetType | Add1ArgType),
6093 NEONMAP1(vcvtns_u32_f32, aarch64_neon_fcvtnu, AddRetType | Add1ArgType),
6094 NEONMAP1(vcvtpd_s64_f64, aarch64_neon_fcvtps, AddRetType | Add1ArgType),
6095 NEONMAP1(vcvtpd_u64_f64, aarch64_neon_fcvtpu, AddRetType | Add1ArgType),
6096 NEONMAP1(vcvtps_s32_f32, aarch64_neon_fcvtps, AddRetType | Add1ArgType),
6097 NEONMAP1(vcvtps_u32_f32, aarch64_neon_fcvtpu, AddRetType | Add1ArgType),
6098 NEONMAP1(vcvts_n_f32_s32, aarch64_neon_vcvtfxs2fp, AddRetType | Add1ArgType),
6099 NEONMAP1(vcvts_n_f32_u32, aarch64_neon_vcvtfxu2fp, AddRetType | Add1ArgType),
6100 NEONMAP1(vcvts_n_s32_f32, aarch64_neon_vcvtfp2fxs, AddRetType | Add1ArgType),
6101 NEONMAP1(vcvts_n_u32_f32, aarch64_neon_vcvtfp2fxu, AddRetType | Add1ArgType),
6102 NEONMAP1(vcvts_s32_f32, aarch64_neon_fcvtzs, AddRetType | Add1ArgType),
6103 NEONMAP1(vcvts_u32_f32, aarch64_neon_fcvtzu, AddRetType | Add1ArgType),
6104 NEONMAP1(vcvtxd_f32_f64, aarch64_sisd_fcvtxn, 0),
6105 NEONMAP1(vmaxnmv_f32, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType),
6106 NEONMAP1(vmaxnmvq_f32, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType),
6107 NEONMAP1(vmaxnmvq_f64, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType),
6108 NEONMAP1(vmaxv_f32, aarch64_neon_fmaxv, AddRetType | Add1ArgType),
6109 NEONMAP1(vmaxv_s32, aarch64_neon_smaxv, AddRetType | Add1ArgType),
6110 NEONMAP1(vmaxv_u32, aarch64_neon_umaxv, AddRetType | Add1ArgType),
6111 NEONMAP1(vmaxvq_f32, aarch64_neon_fmaxv, AddRetType | Add1ArgType),
6112 NEONMAP1(vmaxvq_f64, aarch64_neon_fmaxv, AddRetType | Add1ArgType),
6113 NEONMAP1(vmaxvq_s32, aarch64_neon_smaxv, AddRetType | Add1ArgType),
6114 NEONMAP1(vmaxvq_u32, aarch64_neon_umaxv, AddRetType | Add1ArgType),
6115 NEONMAP1(vminnmv_f32, aarch64_neon_fminnmv, AddRetType | Add1ArgType),
6116 NEONMAP1(vminnmvq_f32, aarch64_neon_fminnmv, AddRetType | Add1ArgType),
6117 NEONMAP1(vminnmvq_f64, aarch64_neon_fminnmv, AddRetType | Add1ArgType),
6118 NEONMAP1(vminv_f32, aarch64_neon_fminv, AddRetType | Add1ArgType),
6119 NEONMAP1(vminv_s32, aarch64_neon_sminv, AddRetType | Add1ArgType),
6120 NEONMAP1(vminv_u32, aarch64_neon_uminv, AddRetType | Add1ArgType),
6121 NEONMAP1(vminvq_f32, aarch64_neon_fminv, AddRetType | Add1ArgType),
6122 NEONMAP1(vminvq_f64, aarch64_neon_fminv, AddRetType | Add1ArgType),
6123 NEONMAP1(vminvq_s32, aarch64_neon_sminv, AddRetType | Add1ArgType),
6124 NEONMAP1(vminvq_u32, aarch64_neon_uminv, AddRetType | Add1ArgType),
6125 NEONMAP1(vmull_p64, aarch64_neon_pmull64, 0),
6126 NEONMAP1(vmulxd_f64, aarch64_neon_fmulx, Add1ArgType),
6127 NEONMAP1(vmulxs_f32, aarch64_neon_fmulx, Add1ArgType),
6128 NEONMAP1(vpaddd_s64, aarch64_neon_uaddv, AddRetType | Add1ArgType),
6129 NEONMAP1(vpaddd_u64, aarch64_neon_uaddv, AddRetType | Add1ArgType),
6130 NEONMAP1(vpmaxnmqd_f64, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType),
6131 NEONMAP1(vpmaxnms_f32, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType),
6132 NEONMAP1(vpmaxqd_f64, aarch64_neon_fmaxv, AddRetType | Add1ArgType),
6133 NEONMAP1(vpmaxs_f32, aarch64_neon_fmaxv, AddRetType | Add1ArgType),
6134 NEONMAP1(vpminnmqd_f64, aarch64_neon_fminnmv, AddRetType | Add1ArgType),
6135 NEONMAP1(vpminnms_f32, aarch64_neon_fminnmv, AddRetType | Add1ArgType),
6136 NEONMAP1(vpminqd_f64, aarch64_neon_fminv, AddRetType | Add1ArgType),
6137 NEONMAP1(vpmins_f32, aarch64_neon_fminv, AddRetType | Add1ArgType),
6138 NEONMAP1(vqabsb_s8, aarch64_neon_sqabs, Vectorize1ArgType | Use64BitVectors),
6139 NEONMAP1(vqabsd_s64, aarch64_neon_sqabs, Add1ArgType),
6140 NEONMAP1(vqabsh_s16, aarch64_neon_sqabs, Vectorize1ArgType | Use64BitVectors),
6141 NEONMAP1(vqabss_s32, aarch64_neon_sqabs, Add1ArgType),
6142 NEONMAP1(vqaddb_s8, aarch64_neon_sqadd, Vectorize1ArgType | Use64BitVectors),
6143 NEONMAP1(vqaddb_u8, aarch64_neon_uqadd, Vectorize1ArgType | Use64BitVectors),
6144 NEONMAP1(vqaddd_s64, aarch64_neon_sqadd, Add1ArgType),
6145 NEONMAP1(vqaddd_u64, aarch64_neon_uqadd, Add1ArgType),
6146 NEONMAP1(vqaddh_s16, aarch64_neon_sqadd, Vectorize1ArgType | Use64BitVectors),
6147 NEONMAP1(vqaddh_u16, aarch64_neon_uqadd, Vectorize1ArgType | Use64BitVectors),
6148 NEONMAP1(vqadds_s32, aarch64_neon_sqadd, Add1ArgType),
6149 NEONMAP1(vqadds_u32, aarch64_neon_uqadd, Add1ArgType),
6150 NEONMAP1(vqdmulhh_s16, aarch64_neon_sqdmulh, Vectorize1ArgType | Use64BitVectors),
6151 NEONMAP1(vqdmulhs_s32, aarch64_neon_sqdmulh, Add1ArgType),
6152 NEONMAP1(vqdmullh_s16, aarch64_neon_sqdmull, VectorRet | Use128BitVectors),
6153 NEONMAP1(vqdmulls_s32, aarch64_neon_sqdmulls_scalar, 0),
6154 NEONMAP1(vqmovnd_s64, aarch64_neon_scalar_sqxtn, AddRetType | Add1ArgType),
6155 NEONMAP1(vqmovnd_u64, aarch64_neon_scalar_uqxtn, AddRetType | Add1ArgType),
6156 NEONMAP1(vqmovnh_s16, aarch64_neon_sqxtn, VectorRet | Use64BitVectors),
6157 NEONMAP1(vqmovnh_u16, aarch64_neon_uqxtn, VectorRet | Use64BitVectors),
6158 NEONMAP1(vqmovns_s32, aarch64_neon_sqxtn, VectorRet | Use64BitVectors),
6159 NEONMAP1(vqmovns_u32, aarch64_neon_uqxtn, VectorRet | Use64BitVectors),
6160 NEONMAP1(vqmovund_s64, aarch64_neon_scalar_sqxtun, AddRetType | Add1ArgType),
6161 NEONMAP1(vqmovunh_s16, aarch64_neon_sqxtun, VectorRet | Use64BitVectors),
6162 NEONMAP1(vqmovuns_s32, aarch64_neon_sqxtun, VectorRet | Use64BitVectors),
6163 NEONMAP1(vqnegb_s8, aarch64_neon_sqneg, Vectorize1ArgType | Use64BitVectors),
6164 NEONMAP1(vqnegd_s64, aarch64_neon_sqneg, Add1ArgType),
6165 NEONMAP1(vqnegh_s16, aarch64_neon_sqneg, Vectorize1ArgType | Use64BitVectors),
6166 NEONMAP1(vqnegs_s32, aarch64_neon_sqneg, Add1ArgType),
6167 NEONMAP1(vqrdmulhh_s16, aarch64_neon_sqrdmulh, Vectorize1ArgType | Use64BitVectors),
6168 NEONMAP1(vqrdmulhs_s32, aarch64_neon_sqrdmulh, Add1ArgType),
6169 NEONMAP1(vqrshlb_s8, aarch64_neon_sqrshl, Vectorize1ArgType | Use64BitVectors),
6170 NEONMAP1(vqrshlb_u8, aarch64_neon_uqrshl, Vectorize1ArgType | Use64BitVectors),
6171 NEONMAP1(vqrshld_s64, aarch64_neon_sqrshl, Add1ArgType),
6172 NEONMAP1(vqrshld_u64, aarch64_neon_uqrshl, Add1ArgType),
6173 NEONMAP1(vqrshlh_s16, aarch64_neon_sqrshl, Vectorize1ArgType | Use64BitVectors),
6174 NEONMAP1(vqrshlh_u16, aarch64_neon_uqrshl, Vectorize1ArgType | Use64BitVectors),
6175 NEONMAP1(vqrshls_s32, aarch64_neon_sqrshl, Add1ArgType),
6176 NEONMAP1(vqrshls_u32, aarch64_neon_uqrshl, Add1ArgType),
6177 NEONMAP1(vqrshrnd_n_s64, aarch64_neon_sqrshrn, AddRetType),
6178 NEONMAP1(vqrshrnd_n_u64, aarch64_neon_uqrshrn, AddRetType),
6179 NEONMAP1(vqrshrnh_n_s16, aarch64_neon_sqrshrn, VectorRet | Use64BitVectors),
6180 NEONMAP1(vqrshrnh_n_u16, aarch64_neon_uqrshrn, VectorRet | Use64BitVectors),
6181 NEONMAP1(vqrshrns_n_s32, aarch64_neon_sqrshrn, VectorRet | Use64BitVectors),
6182 NEONMAP1(vqrshrns_n_u32, aarch64_neon_uqrshrn, VectorRet | Use64BitVectors),
6183 NEONMAP1(vqrshrund_n_s64, aarch64_neon_sqrshrun, AddRetType),
6184 NEONMAP1(vqrshrunh_n_s16, aarch64_neon_sqrshrun, VectorRet | Use64BitVectors),
6185 NEONMAP1(vqrshruns_n_s32, aarch64_neon_sqrshrun, VectorRet | Use64BitVectors),
6186 NEONMAP1(vqshlb_n_s8, aarch64_neon_sqshl, Vectorize1ArgType | Use64BitVectors),
6187 NEONMAP1(vqshlb_n_u8, aarch64_neon_uqshl, Vectorize1ArgType | Use64BitVectors),
6188 NEONMAP1(vqshlb_s8, aarch64_neon_sqshl, Vectorize1ArgType | Use64BitVectors),
6189 NEONMAP1(vqshlb_u8, aarch64_neon_uqshl, Vectorize1ArgType | Use64BitVectors),
6190 NEONMAP1(vqshld_s64, aarch64_neon_sqshl, Add1ArgType),
6191 NEONMAP1(vqshld_u64, aarch64_neon_uqshl, Add1ArgType),
6192 NEONMAP1(vqshlh_n_s16, aarch64_neon_sqshl, Vectorize1ArgType | Use64BitVectors),
6193 NEONMAP1(vqshlh_n_u16, aarch64_neon_uqshl, Vectorize1ArgType | Use64BitVectors),
6194 NEONMAP1(vqshlh_s16, aarch64_neon_sqshl, Vectorize1ArgType | Use64BitVectors),
6195 NEONMAP1(vqshlh_u16, aarch64_neon_uqshl, Vectorize1ArgType | Use64BitVectors),
6196 NEONMAP1(vqshls_n_s32, aarch64_neon_sqshl, Add1ArgType),
6197 NEONMAP1(vqshls_n_u32, aarch64_neon_uqshl, Add1ArgType),
6198 NEONMAP1(vqshls_s32, aarch64_neon_sqshl, Add1ArgType),
6199 NEONMAP1(vqshls_u32, aarch64_neon_uqshl, Add1ArgType),
6200 NEONMAP1(vqshlub_n_s8, aarch64_neon_sqshlu, Vectorize1ArgType | Use64BitVectors),
6201 NEONMAP1(vqshluh_n_s16, aarch64_neon_sqshlu, Vectorize1ArgType | Use64BitVectors),
6202 NEONMAP1(vqshlus_n_s32, aarch64_neon_sqshlu, Add1ArgType),
6203 NEONMAP1(vqshrnd_n_s64, aarch64_neon_sqshrn, AddRetType),
6204 NEONMAP1(vqshrnd_n_u64, aarch64_neon_uqshrn, AddRetType),
6205 NEONMAP1(vqshrnh_n_s16, aarch64_neon_sqshrn, VectorRet | Use64BitVectors),
6206 NEONMAP1(vqshrnh_n_u16, aarch64_neon_uqshrn, VectorRet | Use64BitVectors),
6207 NEONMAP1(vqshrns_n_s32, aarch64_neon_sqshrn, VectorRet | Use64BitVectors),
6208 NEONMAP1(vqshrns_n_u32, aarch64_neon_uqshrn, VectorRet | Use64BitVectors),
6209 NEONMAP1(vqshrund_n_s64, aarch64_neon_sqshrun, AddRetType),
6210 NEONMAP1(vqshrunh_n_s16, aarch64_neon_sqshrun, VectorRet | Use64BitVectors),
6211 NEONMAP1(vqshruns_n_s32, aarch64_neon_sqshrun, VectorRet | Use64BitVectors),
6212 NEONMAP1(vqsubb_s8, aarch64_neon_sqsub, Vectorize1ArgType | Use64BitVectors),
6213 NEONMAP1(vqsubb_u8, aarch64_neon_uqsub, Vectorize1ArgType | Use64BitVectors),
6214 NEONMAP1(vqsubd_s64, aarch64_neon_sqsub, Add1ArgType),
6215 NEONMAP1(vqsubd_u64, aarch64_neon_uqsub, Add1ArgType),
6216 NEONMAP1(vqsubh_s16, aarch64_neon_sqsub, Vectorize1ArgType | Use64BitVectors),
6217 NEONMAP1(vqsubh_u16, aarch64_neon_uqsub, Vectorize1ArgType | Use64BitVectors),
6218 NEONMAP1(vqsubs_s32, aarch64_neon_sqsub, Add1ArgType),
6219 NEONMAP1(vqsubs_u32, aarch64_neon_uqsub, Add1ArgType),
6220 NEONMAP1(vrecped_f64, aarch64_neon_frecpe, Add1ArgType),
6221 NEONMAP1(vrecpes_f32, aarch64_neon_frecpe, Add1ArgType),
6222 NEONMAP1(vrecpxd_f64, aarch64_neon_frecpx, Add1ArgType),
6223 NEONMAP1(vrecpxs_f32, aarch64_neon_frecpx, Add1ArgType),
6224 NEONMAP1(vrshld_s64, aarch64_neon_srshl, Add1ArgType),
6225 NEONMAP1(vrshld_u64, aarch64_neon_urshl, Add1ArgType),
6226 NEONMAP1(vrsqrted_f64, aarch64_neon_frsqrte, Add1ArgType),
6227 NEONMAP1(vrsqrtes_f32, aarch64_neon_frsqrte, Add1ArgType),
6228 NEONMAP1(vrsqrtsd_f64, aarch64_neon_frsqrts, Add1ArgType),
6229 NEONMAP1(vrsqrtss_f32, aarch64_neon_frsqrts, Add1ArgType),
6230 NEONMAP1(vsha1cq_u32, aarch64_crypto_sha1c, 0),
6231 NEONMAP1(vsha1h_u32, aarch64_crypto_sha1h, 0),
6232 NEONMAP1(vsha1mq_u32, aarch64_crypto_sha1m, 0),
6233 NEONMAP1(vsha1pq_u32, aarch64_crypto_sha1p, 0),
6234 NEONMAP1(vshld_s64, aarch64_neon_sshl, Add1ArgType),
6235 NEONMAP1(vshld_u64, aarch64_neon_ushl, Add1ArgType),
6236 NEONMAP1(vslid_n_s64, aarch64_neon_vsli, Vectorize1ArgType),
6237 NEONMAP1(vslid_n_u64, aarch64_neon_vsli, Vectorize1ArgType),
6238 NEONMAP1(vsqaddb_u8, aarch64_neon_usqadd, Vectorize1ArgType | Use64BitVectors),
6239 NEONMAP1(vsqaddd_u64, aarch64_neon_usqadd, Add1ArgType),
6240 NEONMAP1(vsqaddh_u16, aarch64_neon_usqadd, Vectorize1ArgType | Use64BitVectors),
6241 NEONMAP1(vsqadds_u32, aarch64_neon_usqadd, Add1ArgType),
6242 NEONMAP1(vsrid_n_s64, aarch64_neon_vsri, Vectorize1ArgType),
6243 NEONMAP1(vsrid_n_u64, aarch64_neon_vsri, Vectorize1ArgType),
6244 NEONMAP1(vuqaddb_s8, aarch64_neon_suqadd, Vectorize1ArgType | Use64BitVectors),
6245 NEONMAP1(vuqaddd_s64, aarch64_neon_suqadd, Add1ArgType),
6246 NEONMAP1(vuqaddh_s16, aarch64_neon_suqadd, Vectorize1ArgType | Use64BitVectors),
6247 NEONMAP1(vuqadds_s32, aarch64_neon_suqadd, Add1ArgType),
6248 // FP16 scalar intrinisics go here.
6249 NEONMAP1(vabdh_f16, aarch64_sisd_fabd, Add1ArgType),
6250 NEONMAP1(vcvtah_s32_f16, aarch64_neon_fcvtas, AddRetType | Add1ArgType),
6251 NEONMAP1(vcvtah_s64_f16, aarch64_neon_fcvtas, AddRetType | Add1ArgType),
6252 NEONMAP1(vcvtah_u32_f16, aarch64_neon_fcvtau, AddRetType | Add1ArgType),
6253 NEONMAP1(vcvtah_u64_f16, aarch64_neon_fcvtau, AddRetType | Add1ArgType),
6254 NEONMAP1(vcvth_n_f16_s32, aarch64_neon_vcvtfxs2fp, AddRetType | Add1ArgType),
6255 NEONMAP1(vcvth_n_f16_s64, aarch64_neon_vcvtfxs2fp, AddRetType | Add1ArgType),
6256 NEONMAP1(vcvth_n_f16_u32, aarch64_neon_vcvtfxu2fp, AddRetType | Add1ArgType),
6257 NEONMAP1(vcvth_n_f16_u64, aarch64_neon_vcvtfxu2fp, AddRetType | Add1ArgType),
6258 NEONMAP1(vcvth_n_s32_f16, aarch64_neon_vcvtfp2fxs, AddRetType | Add1ArgType),
6259 NEONMAP1(vcvth_n_s64_f16, aarch64_neon_vcvtfp2fxs, AddRetType | Add1ArgType),
6260 NEONMAP1(vcvth_n_u32_f16, aarch64_neon_vcvtfp2fxu, AddRetType | Add1ArgType),
6261 NEONMAP1(vcvth_n_u64_f16, aarch64_neon_vcvtfp2fxu, AddRetType | Add1ArgType),
6262 NEONMAP1(vcvth_s32_f16, aarch64_neon_fcvtzs, AddRetType | Add1ArgType),
6263 NEONMAP1(vcvth_s64_f16, aarch64_neon_fcvtzs, AddRetType | Add1ArgType),
6264 NEONMAP1(vcvth_u32_f16, aarch64_neon_fcvtzu, AddRetType | Add1ArgType),
6265 NEONMAP1(vcvth_u64_f16, aarch64_neon_fcvtzu, AddRetType | Add1ArgType),
6266 NEONMAP1(vcvtmh_s32_f16, aarch64_neon_fcvtms, AddRetType | Add1ArgType),
6267 NEONMAP1(vcvtmh_s64_f16, aarch64_neon_fcvtms, AddRetType | Add1ArgType),
6268 NEONMAP1(vcvtmh_u32_f16, aarch64_neon_fcvtmu, AddRetType | Add1ArgType),
6269 NEONMAP1(vcvtmh_u64_f16, aarch64_neon_fcvtmu, AddRetType | Add1ArgType),
6270 NEONMAP1(vcvtnh_s32_f16, aarch64_neon_fcvtns, AddRetType | Add1ArgType),
6271 NEONMAP1(vcvtnh_s64_f16, aarch64_neon_fcvtns, AddRetType | Add1ArgType),
6272 NEONMAP1(vcvtnh_u32_f16, aarch64_neon_fcvtnu, AddRetType | Add1ArgType),
6273 NEONMAP1(vcvtnh_u64_f16, aarch64_neon_fcvtnu, AddRetType | Add1ArgType),
6274 NEONMAP1(vcvtph_s32_f16, aarch64_neon_fcvtps, AddRetType | Add1ArgType),
6275 NEONMAP1(vcvtph_s64_f16, aarch64_neon_fcvtps, AddRetType | Add1ArgType),
6276 NEONMAP1(vcvtph_u32_f16, aarch64_neon_fcvtpu, AddRetType | Add1ArgType),
6277 NEONMAP1(vcvtph_u64_f16, aarch64_neon_fcvtpu, AddRetType | Add1ArgType),
6278 NEONMAP1(vmulxh_f16, aarch64_neon_fmulx, Add1ArgType),
6279 NEONMAP1(vrecpeh_f16, aarch64_neon_frecpe, Add1ArgType),
6280 NEONMAP1(vrecpxh_f16, aarch64_neon_frecpx, Add1ArgType),
6281 NEONMAP1(vrsqrteh_f16, aarch64_neon_frsqrte, Add1ArgType),
6282 NEONMAP1(vrsqrtsh_f16, aarch64_neon_frsqrts, Add1ArgType),
6283};
6284
6285#undef NEONMAP0
6286#undef NEONMAP1
6287#undef NEONMAP2
6288
6289#define SVEMAP1(NameBase, LLVMIntrinsic, TypeModifier) \
6290 { \
6291 #NameBase, SVE::BI__builtin_sve_##NameBase, Intrinsic::LLVMIntrinsic, 0, \
6292 TypeModifier \
6293 }
6294
6295#define SVEMAP2(NameBase, TypeModifier) \
6296 { #NameBase, SVE::BI__builtin_sve_##NameBase, 0, 0, TypeModifier }
6297static const ARMVectorIntrinsicInfo AArch64SVEIntrinsicMap[] = {
6298#define GET_SVE_LLVM_INTRINSIC_MAP
6299#include "clang/Basic/arm_sve_builtin_cg.inc"
6300#undef GET_SVE_LLVM_INTRINSIC_MAP
6301};
6302
6303#undef SVEMAP1
6304#undef SVEMAP2
6305
6306static bool NEONSIMDIntrinsicsProvenSorted = false;
6307
6308static bool AArch64SIMDIntrinsicsProvenSorted = false;
6309static bool AArch64SISDIntrinsicsProvenSorted = false;
6310static bool AArch64SVEIntrinsicsProvenSorted = false;
6311
6312static const ARMVectorIntrinsicInfo *
6313findARMVectorIntrinsicInMap(ArrayRef<ARMVectorIntrinsicInfo> IntrinsicMap,
6314 unsigned BuiltinID, bool &MapProvenSorted) {
6315
6316#ifndef NDEBUG1
6317 if (!MapProvenSorted) {
6318 assert(llvm::is_sorted(IntrinsicMap))((void)0);
6319 MapProvenSorted = true;
6320 }
6321#endif
6322
6323 const ARMVectorIntrinsicInfo *Builtin =
6324 llvm::lower_bound(IntrinsicMap, BuiltinID);
6325
6326 if (Builtin != IntrinsicMap.end() && Builtin->BuiltinID == BuiltinID)
6327 return Builtin;
6328
6329 return nullptr;
6330}
6331
6332Function *CodeGenFunction::LookupNeonLLVMIntrinsic(unsigned IntrinsicID,
6333 unsigned Modifier,
6334 llvm::Type *ArgType,
6335 const CallExpr *E) {
6336 int VectorSize = 0;
6337 if (Modifier & Use64BitVectors)
6338 VectorSize = 64;
6339 else if (Modifier & Use128BitVectors)
6340 VectorSize = 128;
6341
6342 // Return type.
6343 SmallVector<llvm::Type *, 3> Tys;
6344 if (Modifier & AddRetType) {
6345 llvm::Type *Ty = ConvertType(E->getCallReturnType(getContext()));
6346 if (Modifier & VectorizeRetType)
6347 Ty = llvm::FixedVectorType::get(
6348 Ty, VectorSize ? VectorSize / Ty->getPrimitiveSizeInBits() : 1);
6349
6350 Tys.push_back(Ty);
6351 }
6352
6353 // Arguments.
6354 if (Modifier & VectorizeArgTypes) {
6355 int Elts = VectorSize ? VectorSize / ArgType->getPrimitiveSizeInBits() : 1;
6356 ArgType = llvm::FixedVectorType::get(ArgType, Elts);
6357 }
6358
6359 if (Modifier & (Add1ArgType | Add2ArgTypes))
6360 Tys.push_back(ArgType);
6361
6362 if (Modifier & Add2ArgTypes)
6363 Tys.push_back(ArgType);
6364
6365 if (Modifier & InventFloatType)
6366 Tys.push_back(FloatTy);
6367
6368 return CGM.getIntrinsic(IntrinsicID, Tys);
6369}
6370
6371static Value *EmitCommonNeonSISDBuiltinExpr(
6372 CodeGenFunction &CGF, const ARMVectorIntrinsicInfo &SISDInfo,
6373 SmallVectorImpl<Value *> &Ops, const CallExpr *E) {
6374 unsigned BuiltinID = SISDInfo.BuiltinID;
6375 unsigned int Int = SISDInfo.LLVMIntrinsic;
6376 unsigned Modifier = SISDInfo.TypeModifier;
6377 const char *s = SISDInfo.NameHint;
6378
6379 switch (BuiltinID) {
6380 case NEON::BI__builtin_neon_vcled_s64:
6381 case NEON::BI__builtin_neon_vcled_u64:
6382 case NEON::BI__builtin_neon_vcles_f32:
6383 case NEON::BI__builtin_neon_vcled_f64:
6384 case NEON::BI__builtin_neon_vcltd_s64:
6385 case NEON::BI__builtin_neon_vcltd_u64:
6386 case NEON::BI__builtin_neon_vclts_f32:
6387 case NEON::BI__builtin_neon_vcltd_f64:
6388 case NEON::BI__builtin_neon_vcales_f32:
6389 case NEON::BI__builtin_neon_vcaled_f64:
6390 case NEON::BI__builtin_neon_vcalts_f32:
6391 case NEON::BI__builtin_neon_vcaltd_f64:
6392 // Only one direction of comparisons actually exist, cmle is actually a cmge
6393 // with swapped operands. The table gives us the right intrinsic but we
6394 // still need to do the swap.
6395 std::swap(Ops[0], Ops[1]);
6396 break;
6397 }
6398
6399 assert(Int && "Generic code assumes a valid intrinsic")((void)0);
6400
6401 // Determine the type(s) of this overloaded AArch64 intrinsic.
6402 const Expr *Arg = E->getArg(0);
6403 llvm::Type *ArgTy = CGF.ConvertType(Arg->getType());
6404 Function *F = CGF.LookupNeonLLVMIntrinsic(Int, Modifier, ArgTy, E);
6405
6406 int j = 0;
6407 ConstantInt *C0 = ConstantInt::get(CGF.SizeTy, 0);
6408 for (Function::const_arg_iterator ai = F->arg_begin(), ae = F->arg_end();
6409 ai != ae; ++ai, ++j) {
6410 llvm::Type *ArgTy = ai->getType();
6411 if (Ops[j]->getType()->getPrimitiveSizeInBits() ==
6412 ArgTy->getPrimitiveSizeInBits())
6413 continue;
6414
6415 assert(ArgTy->isVectorTy() && !Ops[j]->getType()->isVectorTy())((void)0);
6416 // The constant argument to an _n_ intrinsic always has Int32Ty, so truncate
6417 // it before inserting.
6418 Ops[j] = CGF.Builder.CreateTruncOrBitCast(
6419 Ops[j], cast<llvm::VectorType>(ArgTy)->getElementType());
6420 Ops[j] =
6421 CGF.Builder.CreateInsertElement(UndefValue::get(ArgTy), Ops[j], C0);
6422 }
6423
6424 Value *Result = CGF.EmitNeonCall(F, Ops, s);
6425 llvm::Type *ResultType = CGF.ConvertType(E->getType());
6426 if (ResultType->getPrimitiveSizeInBits().getFixedSize() <
6427 Result->getType()->getPrimitiveSizeInBits().getFixedSize())
6428 return CGF.Builder.CreateExtractElement(Result, C0);
6429
6430 return CGF.Builder.CreateBitCast(Result, ResultType, s);
6431}
6432
6433Value *CodeGenFunction::EmitCommonNeonBuiltinExpr(
6434 unsigned BuiltinID, unsigned LLVMIntrinsic, unsigned AltLLVMIntrinsic,
6435 const char *NameHint, unsigned Modifier, const CallExpr *E,
6436 SmallVectorImpl<llvm::Value *> &Ops, Address PtrOp0, Address PtrOp1,
6437 llvm::Triple::ArchType Arch) {
6438 // Get the last argument, which specifies the vector type.
6439 const Expr *Arg = E->getArg(E->getNumArgs() - 1);
6440 Optional<llvm::APSInt> NeonTypeConst =
6441 Arg->getIntegerConstantExpr(getContext());
6442 if (!NeonTypeConst)
6443 return nullptr;
6444
6445 // Determine the type of this overloaded NEON intrinsic.
6446 NeonTypeFlags Type(NeonTypeConst->getZExtValue());
6447 bool Usgn = Type.isUnsigned();
6448 bool Quad = Type.isQuad();
6449 const bool HasLegalHalfType = getTarget().hasLegalHalfType();
6450 const bool AllowBFloatArgsAndRet =
6451 getTargetHooks().getABIInfo().allowBFloatArgsAndRet();
6452
6453 llvm::FixedVectorType *VTy =
6454 GetNeonType(this, Type, HasLegalHalfType, false, AllowBFloatArgsAndRet);
6455 llvm::Type *Ty = VTy;
6456 if (!Ty)
6457 return nullptr;
6458
6459 auto getAlignmentValue32 = [&](Address addr) -> Value* {
6460 return Builder.getInt32(addr.getAlignment().getQuantity());
6461 };
6462
6463 unsigned Int = LLVMIntrinsic;
6464 if ((Modifier & UnsignedAlts) && !Usgn)
6465 Int = AltLLVMIntrinsic;
6466
6467 switch (BuiltinID) {
6468 default: break;
6469 case NEON::BI__builtin_neon_splat_lane_v:
6470 case NEON::BI__builtin_neon_splat_laneq_v:
6471 case NEON::BI__builtin_neon_splatq_lane_v:
6472 case NEON::BI__builtin_neon_splatq_laneq_v: {
6473 auto NumElements = VTy->getElementCount();
6474 if (BuiltinID == NEON::BI__builtin_neon_splatq_lane_v)
6475 NumElements = NumElements * 2;
6476 if (BuiltinID == NEON::BI__builtin_neon_splat_laneq_v)
6477 NumElements = NumElements.divideCoefficientBy(2);
6478
6479 Ops[0] = Builder.CreateBitCast(Ops[0], VTy);
6480 return EmitNeonSplat(Ops[0], cast<ConstantInt>(Ops[1]), NumElements);
6481 }
6482 case NEON::BI__builtin_neon_vpadd_v:
6483 case NEON::BI__builtin_neon_vpaddq_v:
6484 // We don't allow fp/int overloading of intrinsics.
6485 if (VTy->getElementType()->isFloatingPointTy() &&
6486 Int == Intrinsic::aarch64_neon_addp)
6487 Int = Intrinsic::aarch64_neon_faddp;
6488 break;
6489 case NEON::BI__builtin_neon_vabs_v:
6490 case NEON::BI__builtin_neon_vabsq_v:
6491 if (VTy->getElementType()->isFloatingPointTy())
6492 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::fabs, Ty), Ops, "vabs");
6493 return EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Ty), Ops, "vabs");
6494 case NEON::BI__builtin_neon_vadd_v:
6495 case NEON::BI__builtin_neon_vaddq_v: {
6496 llvm::Type *VTy = llvm::FixedVectorType::get(Int8Ty, Quad ? 16 : 8);
6497 Ops[0] = Builder.CreateBitCast(Ops[0], VTy);
6498 Ops[1] = Builder.CreateBitCast(Ops[1], VTy);
6499 Ops[0] = Builder.CreateXor(Ops[0], Ops[1]);
6500 return Builder.CreateBitCast(Ops[0], Ty);
6501 }
6502 case NEON::BI__builtin_neon_vaddhn_v: {
6503 llvm::FixedVectorType *SrcTy =
6504 llvm::FixedVectorType::getExtendedElementVectorType(VTy);
6505
6506 // %sum = add <4 x i32> %lhs, %rhs
6507 Ops[0] = Builder.CreateBitCast(Ops[0], SrcTy);
6508 Ops[1] = Builder.CreateBitCast(Ops[1], SrcTy);
6509 Ops[0] = Builder.CreateAdd(Ops[0], Ops[1], "vaddhn");
6510
6511 // %high = lshr <4 x i32> %sum, <i32 16, i32 16, i32 16, i32 16>
6512 Constant *ShiftAmt =
6513 ConstantInt::get(SrcTy, SrcTy->getScalarSizeInBits() / 2);
6514 Ops[0] = Builder.CreateLShr(Ops[0], ShiftAmt, "vaddhn");
6515
6516 // %res = trunc <4 x i32> %high to <4 x i16>
6517 return Builder.CreateTrunc(Ops[0], VTy, "vaddhn");
6518 }
6519 case NEON::BI__builtin_neon_vcale_v:
6520 case NEON::BI__builtin_neon_vcaleq_v:
6521 case NEON::BI__builtin_neon_vcalt_v:
6522 case NEON::BI__builtin_neon_vcaltq_v:
6523 std::swap(Ops[0], Ops[1]);
6524 LLVM_FALLTHROUGH[[gnu::fallthrough]];
6525 case NEON::BI__builtin_neon_vcage_v:
6526 case NEON::BI__builtin_neon_vcageq_v:
6527 case NEON::BI__builtin_neon_vcagt_v:
6528 case NEON::BI__builtin_neon_vcagtq_v: {
6529 llvm::Type *Ty;
6530 switch (VTy->getScalarSizeInBits()) {
6531 default: llvm_unreachable("unexpected type")__builtin_unreachable();
6532 case 32:
6533 Ty = FloatTy;
6534 break;
6535 case 64:
6536 Ty = DoubleTy;
6537 break;
6538 case 16:
6539 Ty = HalfTy;
6540 break;
6541 }
6542 auto *VecFlt = llvm::FixedVectorType::get(Ty, VTy->getNumElements());
6543 llvm::Type *Tys[] = { VTy, VecFlt };
6544 Function *F = CGM.getIntrinsic(LLVMIntrinsic, Tys);
6545 return EmitNeonCall(F, Ops, NameHint);
6546 }
6547 case NEON::BI__builtin_neon_vceqz_v:
6548 case NEON::BI__builtin_neon_vceqzq_v:
6549 return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OEQ,
6550 ICmpInst::ICMP_EQ, "vceqz");
6551 case NEON::BI__builtin_neon_vcgez_v:
6552 case NEON::BI__builtin_neon_vcgezq_v:
6553 return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OGE,
6554 ICmpInst::ICMP_SGE, "vcgez");
6555 case NEON::BI__builtin_neon_vclez_v:
6556 case NEON::BI__builtin_neon_vclezq_v:
6557 return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OLE,
6558 ICmpInst::ICMP_SLE, "vclez");
6559 case NEON::BI__builtin_neon_vcgtz_v:
6560 case NEON::BI__builtin_neon_vcgtzq_v:
6561 return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OGT,
6562 ICmpInst::ICMP_SGT, "vcgtz");
6563 case NEON::BI__builtin_neon_vcltz_v:
6564 case NEON::BI__builtin_neon_vcltzq_v:
6565 return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OLT,
6566 ICmpInst::ICMP_SLT, "vcltz");
6567 case NEON::BI__builtin_neon_vclz_v:
6568 case NEON::BI__builtin_neon_vclzq_v:
6569 // We generate target-independent intrinsic, which needs a second argument
6570 // for whether or not clz of zero is undefined; on ARM it isn't.
6571 Ops.push_back(Builder.getInt1(getTarget().isCLZForZeroUndef()));
6572 break;
6573 case NEON::BI__builtin_neon_vcvt_f32_v:
6574 case NEON::BI__builtin_neon_vcvtq_f32_v:
6575 Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
6576 Ty = GetNeonType(this, NeonTypeFlags(NeonTypeFlags::Float32, false, Quad),
6577 HasLegalHalfType);
6578 return Usgn ? Builder.CreateUIToFP(Ops[0], Ty, "vcvt")
6579 : Builder.CreateSIToFP(Ops[0], Ty, "vcvt");
6580 case NEON::BI__builtin_neon_vcvt_f16_v:
6581 case NEON::BI__builtin_neon_vcvtq_f16_v:
6582 Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
6583 Ty = GetNeonType(this, NeonTypeFlags(NeonTypeFlags::Float16, false, Quad),
6584 HasLegalHalfType);
6585 return Usgn ? Builder.CreateUIToFP(Ops[0], Ty, "vcvt")
6586 : Builder.CreateSIToFP(Ops[0], Ty, "vcvt");
6587 case NEON::BI__builtin_neon_vcvt_n_f16_v:
6588 case NEON::BI__builtin_neon_vcvt_n_f32_v:
6589 case NEON::BI__builtin_neon_vcvt_n_f64_v:
6590 case NEON::BI__builtin_neon_vcvtq_n_f16_v:
6591 case NEON::BI__builtin_neon_vcvtq_n_f32_v:
6592 case NEON::BI__builtin_neon_vcvtq_n_f64_v: {
6593 llvm::Type *Tys[2] = { GetFloatNeonType(this, Type), Ty };
6594 Int = Usgn ? LLVMIntrinsic : AltLLVMIntrinsic;
6595 Function *F = CGM.getIntrinsic(Int, Tys);
6596 return EmitNeonCall(F, Ops, "vcvt_n");
6597 }
6598 case NEON::BI__builtin_neon_vcvt_n_s16_v:
6599 case NEON::BI__builtin_neon_vcvt_n_s32_v:
6600 case NEON::BI__builtin_neon_vcvt_n_u16_v:
6601 case NEON::BI__builtin_neon_vcvt_n_u32_v:
6602 case NEON::BI__builtin_neon_vcvt_n_s64_v:
6603 case NEON::BI__builtin_neon_vcvt_n_u64_v:
6604 case NEON::BI__builtin_neon_vcvtq_n_s16_v:
6605 case NEON::BI__builtin_neon_vcvtq_n_s32_v:
6606 case NEON::BI__builtin_neon_vcvtq_n_u16_v:
6607 case NEON::BI__builtin_neon_vcvtq_n_u32_v:
6608 case NEON::BI__builtin_neon_vcvtq_n_s64_v:
6609 case NEON::BI__builtin_neon_vcvtq_n_u64_v: {
6610 llvm::Type *Tys[2] = { Ty, GetFloatNeonType(this, Type) };
6611 Function *F = CGM.getIntrinsic(LLVMIntrinsic, Tys);
6612 return EmitNeonCall(F, Ops, "vcvt_n");
6613 }
6614 case NEON::BI__builtin_neon_vcvt_s32_v:
6615 case NEON::BI__builtin_neon_vcvt_u32_v:
6616 case NEON::BI__builtin_neon_vcvt_s64_v:
6617 case NEON::BI__builtin_neon_vcvt_u64_v:
6618 case NEON::BI__builtin_neon_vcvt_s16_v:
6619 case NEON::BI__builtin_neon_vcvt_u16_v:
6620 case NEON::BI__builtin_neon_vcvtq_s32_v:
6621 case NEON::BI__builtin_neon_vcvtq_u32_v:
6622 case NEON::BI__builtin_neon_vcvtq_s64_v:
6623 case NEON::BI__builtin_neon_vcvtq_u64_v:
6624 case NEON::BI__builtin_neon_vcvtq_s16_v:
6625 case NEON::BI__builtin_neon_vcvtq_u16_v: {
6626 Ops[0] = Builder.CreateBitCast(Ops[0], GetFloatNeonType(this, Type));
6627 return Usgn ? Builder.CreateFPToUI(Ops[0], Ty, "vcvt")
6628 : Builder.CreateFPToSI(Ops[0], Ty, "vcvt");
6629 }
6630 case NEON::BI__builtin_neon_vcvta_s16_v:
6631 case NEON::BI__builtin_neon_vcvta_s32_v:
6632 case NEON::BI__builtin_neon_vcvta_s64_v:
6633 case NEON::BI__builtin_neon_vcvta_u16_v:
6634 case NEON::BI__builtin_neon_vcvta_u32_v:
6635 case NEON::BI__builtin_neon_vcvta_u64_v:
6636 case NEON::BI__builtin_neon_vcvtaq_s16_v:
6637 case NEON::BI__builtin_neon_vcvtaq_s32_v:
6638 case NEON::BI__builtin_neon_vcvtaq_s64_v:
6639 case NEON::BI__builtin_neon_vcvtaq_u16_v:
6640 case NEON::BI__builtin_neon_vcvtaq_u32_v:
6641 case NEON::BI__builtin_neon_vcvtaq_u64_v:
6642 case NEON::BI__builtin_neon_vcvtn_s16_v:
6643 case NEON::BI__builtin_neon_vcvtn_s32_v:
6644 case NEON::BI__builtin_neon_vcvtn_s64_v:
6645 case NEON::BI__builtin_neon_vcvtn_u16_v:
6646 case NEON::BI__builtin_neon_vcvtn_u32_v:
6647 case NEON::BI__builtin_neon_vcvtn_u64_v:
6648 case NEON::BI__builtin_neon_vcvtnq_s16_v:
6649 case NEON::BI__builtin_neon_vcvtnq_s32_v:
6650 case NEON::BI__builtin_neon_vcvtnq_s64_v:
6651 case NEON::BI__builtin_neon_vcvtnq_u16_v:
6652 case NEON::BI__builtin_neon_vcvtnq_u32_v:
6653 case NEON::BI__builtin_neon_vcvtnq_u64_v:
6654 case NEON::BI__builtin_neon_vcvtp_s16_v:
6655 case NEON::BI__builtin_neon_vcvtp_s32_v:
6656 case NEON::BI__builtin_neon_vcvtp_s64_v:
6657 case NEON::BI__builtin_neon_vcvtp_u16_v:
6658 case NEON::BI__builtin_neon_vcvtp_u32_v:
6659 case NEON::BI__builtin_neon_vcvtp_u64_v:
6660 case NEON::BI__builtin_neon_vcvtpq_s16_v:
6661 case NEON::BI__builtin_neon_vcvtpq_s32_v:
6662 case NEON::BI__builtin_neon_vcvtpq_s64_v:
6663 case NEON::BI__builtin_neon_vcvtpq_u16_v:
6664 case NEON::BI__builtin_neon_vcvtpq_u32_v:
6665 case NEON::BI__builtin_neon_vcvtpq_u64_v:
6666 case NEON::BI__builtin_neon_vcvtm_s16_v:
6667 case NEON::BI__builtin_neon_vcvtm_s32_v:
6668 case NEON::BI__builtin_neon_vcvtm_s64_v:
6669 case NEON::BI__builtin_neon_vcvtm_u16_v:
6670 case NEON::BI__builtin_neon_vcvtm_u32_v:
6671 case NEON::BI__builtin_neon_vcvtm_u64_v:
6672 case NEON::BI__builtin_neon_vcvtmq_s16_v:
6673 case NEON::BI__builtin_neon_vcvtmq_s32_v:
6674 case NEON::BI__builtin_neon_vcvtmq_s64_v:
6675 case NEON::BI__builtin_neon_vcvtmq_u16_v:
6676 case NEON::BI__builtin_neon_vcvtmq_u32_v:
6677 case NEON::BI__builtin_neon_vcvtmq_u64_v: {
6678 llvm::Type *Tys[2] = { Ty, GetFloatNeonType(this, Type) };
6679 return EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Tys), Ops, NameHint);
6680 }
6681 case NEON::BI__builtin_neon_vcvtx_f32_v: {
6682 llvm::Type *Tys[2] = { VTy->getTruncatedElementVectorType(VTy), Ty};
6683 return EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Tys), Ops, NameHint);
6684
6685 }
6686 case NEON::BI__builtin_neon_vext_v:
6687 case NEON::BI__builtin_neon_vextq_v: {
6688 int CV = cast<ConstantInt>(Ops[2])->getSExtValue();
6689 SmallVector<int, 16> Indices;
6690 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i)
6691 Indices.push_back(i+CV);
6692
6693 Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
6694 Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
6695 return Builder.CreateShuffleVector(Ops[0], Ops[1], Indices, "vext");
6696 }
6697 case NEON::BI__builtin_neon_vfma_v:
6698 case NEON::BI__builtin_neon_vfmaq_v: {
6699 Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
6700 Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
6701 Ops[2] = Builder.CreateBitCast(Ops[2], Ty);
6702
6703 // NEON intrinsic puts accumulator first, unlike the LLVM fma.
6704 return emitCallMaybeConstrainedFPBuiltin(
6705 *this, Intrinsic::fma, Intrinsic::experimental_constrained_fma, Ty,
6706 {Ops[1], Ops[2], Ops[0]});
6707 }
6708 case NEON::BI__builtin_neon_vld1_v:
6709 case NEON::BI__builtin_neon_vld1q_v: {
6710 llvm::Type *Tys[] = {Ty, Int8PtrTy};
6711 Ops.push_back(getAlignmentValue32(PtrOp0));
6712 return EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Tys), Ops, "vld1");
6713 }
6714 case NEON::BI__builtin_neon_vld1_x2_v:
6715 case NEON::BI__builtin_neon_vld1q_x2_v:
6716 case NEON::BI__builtin_neon_vld1_x3_v:
6717 case NEON::BI__builtin_neon_vld1q_x3_v:
6718 case NEON::BI__builtin_neon_vld1_x4_v:
6719 case NEON::BI__builtin_neon_vld1q_x4_v: {
6720 llvm::Type *PTy = llvm::PointerType::getUnqual(VTy->getElementType());
6721 Ops[1] = Builder.CreateBitCast(Ops[1], PTy);
6722 llvm::Type *Tys[2] = { VTy, PTy };
6723 Function *F = CGM.getIntrinsic(LLVMIntrinsic, Tys);
6724 Ops[1] = Builder.CreateCall(F, Ops[1], "vld1xN");
6725 Ty = llvm::PointerType::getUnqual(Ops[1]->getType());
6726 Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
6727 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]);
6728 }
6729 case NEON::BI__builtin_neon_vld2_v:
6730 case NEON::BI__builtin_neon_vld2q_v:
6731 case NEON::BI__builtin_neon_vld3_v:
6732 case NEON::BI__builtin_neon_vld3q_v:
6733 case NEON::BI__builtin_neon_vld4_v:
6734 case NEON::BI__builtin_neon_vld4q_v:
6735 case NEON::BI__builtin_neon_vld2_dup_v:
6736 case NEON::BI__builtin_neon_vld2q_dup_v:
6737 case NEON::BI__builtin_neon_vld3_dup_v:
6738 case NEON::BI__builtin_neon_vld3q_dup_v:
6739 case NEON::BI__builtin_neon_vld4_dup_v:
6740 case NEON::BI__builtin_neon_vld4q_dup_v: {
6741 llvm::Type *Tys[] = {Ty, Int8PtrTy};
6742 Function *F = CGM.getIntrinsic(LLVMIntrinsic, Tys);
6743 Value *Align = getAlignmentValue32(PtrOp1);
6744 Ops[1] = Builder.CreateCall(F, {Ops[1], Align}, NameHint);
6745 Ty = llvm::PointerType::getUnqual(Ops[1]->getType());
6746 Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
6747 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]);
6748 }
6749 case NEON::BI__builtin_neon_vld1_dup_v:
6750 case NEON::BI__builtin_neon_vld1q_dup_v: {
6751 Value *V = UndefValue::get(Ty);
6752 Ty = llvm::PointerType::getUnqual(VTy->getElementType());
6753 PtrOp0 = Builder.CreateBitCast(PtrOp0, Ty);
6754 LoadInst *Ld = Builder.CreateLoad(PtrOp0);
6755 llvm::Constant *CI = ConstantInt::get(SizeTy, 0);
6756 Ops[0] = Builder.CreateInsertElement(V, Ld, CI);
6757 return EmitNeonSplat(Ops[0], CI);
6758 }
6759 case NEON::BI__builtin_neon_vld2_lane_v:
6760 case NEON::BI__builtin_neon_vld2q_lane_v:
6761 case NEON::BI__builtin_neon_vld3_lane_v:
6762 case NEON::BI__builtin_neon_vld3q_lane_v:
6763 case NEON::BI__builtin_neon_vld4_lane_v:
6764 case NEON::BI__builtin_neon_vld4q_lane_v: {
6765 llvm::Type *Tys[] = {Ty, Int8PtrTy};
6766 Function *F = CGM.getIntrinsic(LLVMIntrinsic, Tys);
6767 for (unsigned I = 2; I < Ops.size() - 1; ++I)
6768 Ops[I] = Builder.CreateBitCast(Ops[I], Ty);
6769 Ops.push_back(getAlignmentValue32(PtrOp1));
6770 Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), NameHint);
6771 Ty = llvm::PointerType::getUnqual(Ops[1]->getType());
6772 Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
6773 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]);
6774 }
6775 case NEON::BI__builtin_neon_vmovl_v: {
6776 llvm::FixedVectorType *DTy =
6777 llvm::FixedVectorType::getTruncatedElementVectorType(VTy);
6778 Ops[0] = Builder.CreateBitCast(Ops[0], DTy);
6779 if (Usgn)
6780 return Builder.CreateZExt(Ops[0], Ty, "vmovl");
6781 return Builder.CreateSExt(Ops[0], Ty, "vmovl");
6782 }
6783 case NEON::BI__builtin_neon_vmovn_v: {
6784 llvm::FixedVectorType *QTy =
6785 llvm::FixedVectorType::getExtendedElementVectorType(VTy);
6786 Ops[0] = Builder.CreateBitCast(Ops[0], QTy);
6787 return Builder.CreateTrunc(Ops[0], Ty, "vmovn");
6788 }
6789 case NEON::BI__builtin_neon_vmull_v:
6790 // FIXME: the integer vmull operations could be emitted in terms of pure
6791 // LLVM IR (2 exts followed by a mul). Unfortunately LLVM has a habit of
6792 // hoisting the exts outside loops. Until global ISel comes along that can
6793 // see through such movement this leads to bad CodeGen. So we need an
6794 // intrinsic for now.
6795 Int = Usgn ? Intrinsic::arm_neon_vmullu : Intrinsic::arm_neon_vmulls;
6796 Int = Type.isPoly() ? (unsigned)Intrinsic::arm_neon_vmullp : Int;
6797 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmull");
6798 case NEON::BI__builtin_neon_vpadal_v:
6799 case NEON::BI__builtin_neon_vpadalq_v: {
6800 // The source operand type has twice as many elements of half the size.
6801 unsigned EltBits = VTy->getElementType()->getPrimitiveSizeInBits();
6802 llvm::Type *EltTy =
6803 llvm::IntegerType::get(getLLVMContext(), EltBits / 2);
6804 auto *NarrowTy =
6805 llvm::FixedVectorType::get(EltTy, VTy->getNumElements() * 2);
6806 llvm::Type *Tys[2] = { Ty, NarrowTy };
6807 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, NameHint);
6808 }
6809 case NEON::BI__builtin_neon_vpaddl_v:
6810 case NEON::BI__builtin_neon_vpaddlq_v: {
6811 // The source operand type has twice as many elements of half the size.
6812 unsigned EltBits = VTy->getElementType()->getPrimitiveSizeInBits();
6813 llvm::Type *EltTy = llvm::IntegerType::get(getLLVMContext(), EltBits / 2);
6814 auto *NarrowTy =
6815 llvm::FixedVectorType::get(EltTy, VTy->getNumElements() * 2);
6816 llvm::Type *Tys[2] = { Ty, NarrowTy };
6817 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vpaddl");
6818 }
6819 case NEON::BI__builtin_neon_vqdmlal_v:
6820 case NEON::BI__builtin_neon_vqdmlsl_v: {
6821 SmallVector<Value *, 2> MulOps(Ops.begin() + 1, Ops.end());
6822 Ops[1] =
6823 EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Ty), MulOps, "vqdmlal");
6824 Ops.resize(2);
6825 return EmitNeonCall(CGM.getIntrinsic(AltLLVMIntrinsic, Ty), Ops, NameHint);
6826 }
6827 case NEON::BI__builtin_neon_vqdmulhq_lane_v:
6828 case NEON::BI__builtin_neon_vqdmulh_lane_v:
6829 case NEON::BI__builtin_neon_vqrdmulhq_lane_v:
6830 case NEON::BI__builtin_neon_vqrdmulh_lane_v: {
6831 auto *RTy = cast<llvm::FixedVectorType>(Ty);
6832 if (BuiltinID == NEON::BI__builtin_neon_vqdmulhq_lane_v ||
6833 BuiltinID == NEON::BI__builtin_neon_vqrdmulhq_lane_v)
6834 RTy = llvm::FixedVectorType::get(RTy->getElementType(),
6835 RTy->getNumElements() * 2);
6836 llvm::Type *Tys[2] = {
6837 RTy, GetNeonType(this, NeonTypeFlags(Type.getEltType(), false,
6838 /*isQuad*/ false))};
6839 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, NameHint);
6840 }
6841 case NEON::BI__builtin_neon_vqdmulhq_laneq_v:
6842 case NEON::BI__builtin_neon_vqdmulh_laneq_v:
6843 case NEON::BI__builtin_neon_vqrdmulhq_laneq_v:
6844 case NEON::BI__builtin_neon_vqrdmulh_laneq_v: {
6845 llvm::Type *Tys[2] = {
6846 Ty, GetNeonType(this, NeonTypeFlags(Type.getEltType(), false,
6847 /*isQuad*/ true))};
6848 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, NameHint);
6849 }
6850 case NEON::BI__builtin_neon_vqshl_n_v:
6851 case NEON::BI__builtin_neon_vqshlq_n_v:
6852 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshl_n",
6853 1, false);
6854 case NEON::BI__builtin_neon_vqshlu_n_v:
6855 case NEON::BI__builtin_neon_vqshluq_n_v:
6856 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshlu_n",
6857 1, false);
6858 case NEON::BI__builtin_neon_vrecpe_v:
6859 case NEON::BI__builtin_neon_vrecpeq_v:
6860 case NEON::BI__builtin_neon_vrsqrte_v:
6861 case NEON::BI__builtin_neon_vrsqrteq_v:
6862 Int = Ty->isFPOrFPVectorTy() ? LLVMIntrinsic : AltLLVMIntrinsic;
6863 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, NameHint);
6864 case NEON::BI__builtin_neon_vrndi_v:
6865 case NEON::BI__builtin_neon_vrndiq_v:
6866 Int = Builder.getIsFPConstrained()
6867 ? Intrinsic::experimental_constrained_nearbyint
6868 : Intrinsic::nearbyint;
6869 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, NameHint);
6870 case NEON::BI__builtin_neon_vrshr_n_v:
6871 case NEON::BI__builtin_neon_vrshrq_n_v:
6872 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrshr_n",
6873 1, true);
6874 case NEON::BI__builtin_neon_vsha512hq_v:
6875 case NEON::BI__builtin_neon_vsha512h2q_v:
6876 case NEON::BI__builtin_neon_vsha512su0q_v:
6877 case NEON::BI__builtin_neon_vsha512su1q_v: {
6878 Function *F = CGM.getIntrinsic(Int);
6879 return EmitNeonCall(F, Ops, "");
6880 }
6881 case NEON::BI__builtin_neon_vshl_n_v:
6882 case NEON::BI__builtin_neon_vshlq_n_v:
6883 Ops[1] = EmitNeonShiftVector(Ops[1], Ty, false);
6884 return Builder.CreateShl(Builder.CreateBitCast(Ops[0],Ty), Ops[1],
6885 "vshl_n");
6886 case NEON::BI__builtin_neon_vshll_n_v: {
6887 llvm::FixedVectorType *SrcTy =
6888 llvm::FixedVectorType::getTruncatedElementVectorType(VTy);
6889 Ops[0] = Builder.CreateBitCast(Ops[0], SrcTy);
6890 if (Usgn)
6891 Ops[0] = Builder.CreateZExt(Ops[0], VTy);
6892 else
6893 Ops[0] = Builder.CreateSExt(Ops[0], VTy);
6894 Ops[1] = EmitNeonShiftVector(Ops[1], VTy, false);
6895 return Builder.CreateShl(Ops[0], Ops[1], "vshll_n");
6896 }
6897 case NEON::BI__builtin_neon_vshrn_n_v: {
6898 llvm::FixedVectorType *SrcTy =
6899 llvm::FixedVectorType::getExtendedElementVectorType(VTy);
6900 Ops[0] = Builder.CreateBitCast(Ops[0], SrcTy);
6901 Ops[1] = EmitNeonShiftVector(Ops[1], SrcTy, false);
6902 if (Usgn)
6903 Ops[0] = Builder.CreateLShr(Ops[0], Ops[1]);
6904 else
6905 Ops[0] = Builder.CreateAShr(Ops[0], Ops[1]);
6906 return Builder.CreateTrunc(Ops[0], Ty, "vshrn_n");
6907 }
6908 case NEON::BI__builtin_neon_vshr_n_v:
6909 case NEON::BI__builtin_neon_vshrq_n_v:
6910 return EmitNeonRShiftImm(Ops[0], Ops[1], Ty, Usgn, "vshr_n");
6911 case NEON::BI__builtin_neon_vst1_v:
6912 case NEON::BI__builtin_neon_vst1q_v:
6913 case NEON::BI__builtin_neon_vst2_v:
6914 case NEON::BI__builtin_neon_vst2q_v:
6915 case NEON::BI__builtin_neon_vst3_v:
6916 case NEON::BI__builtin_neon_vst3q_v:
6917 case NEON::BI__builtin_neon_vst4_v:
6918 case NEON::BI__builtin_neon_vst4q_v:
6919 case NEON::BI__builtin_neon_vst2_lane_v:
6920 case NEON::BI__builtin_neon_vst2q_lane_v:
6921 case NEON::BI__builtin_neon_vst3_lane_v:
6922 case NEON::BI__builtin_neon_vst3q_lane_v:
6923 case NEON::BI__builtin_neon_vst4_lane_v:
6924 case NEON::BI__builtin_neon_vst4q_lane_v: {
6925 llvm::Type *Tys[] = {Int8PtrTy, Ty};
6926 Ops.push_back(getAlignmentValue32(PtrOp0));
6927 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "");
6928 }
6929 case NEON::BI__builtin_neon_vsm3partw1q_v:
6930 case NEON::BI__builtin_neon_vsm3partw2q_v:
6931 case NEON::BI__builtin_neon_vsm3ss1q_v:
6932 case NEON::BI__builtin_neon_vsm4ekeyq_v:
6933 case NEON::BI__builtin_neon_vsm4eq_v: {
6934 Function *F = CGM.getIntrinsic(Int);
6935 return EmitNeonCall(F, Ops, "");
6936 }
6937 case NEON::BI__builtin_neon_vsm3tt1aq_v:
6938 case NEON::BI__builtin_neon_vsm3tt1bq_v:
6939 case NEON::BI__builtin_neon_vsm3tt2aq_v:
6940 case NEON::BI__builtin_neon_vsm3tt2bq_v: {
6941 Function *F = CGM.getIntrinsic(Int);
6942 Ops[3] = Builder.CreateZExt(Ops[3], Int64Ty);
6943 return EmitNeonCall(F, Ops, "");
6944 }
6945 case NEON::BI__builtin_neon_vst1_x2_v:
6946 case NEON::BI__builtin_neon_vst1q_x2_v:
6947 case NEON::BI__builtin_neon_vst1_x3_v:
6948 case NEON::BI__builtin_neon_vst1q_x3_v:
6949 case NEON::BI__builtin_neon_vst1_x4_v:
6950 case NEON::BI__builtin_neon_vst1q_x4_v: {
6951 llvm::Type *PTy = llvm::PointerType::getUnqual(VTy->getElementType());
6952 // TODO: Currently in AArch32 mode the pointer operand comes first, whereas
6953 // in AArch64 it comes last. We may want to stick to one or another.
6954 if (Arch == llvm::Triple::aarch64 || Arch == llvm::Triple::aarch64_be ||
6955 Arch == llvm::Triple::aarch64_32) {
6956 llvm::Type *Tys[2] = { VTy, PTy };
6957 std::rotate(Ops.begin(), Ops.begin() + 1, Ops.end());
6958 return EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Tys), Ops, "");
6959 }
6960 llvm::Type *Tys[2] = { PTy, VTy };
6961 return EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Tys), Ops, "");
6962 }
6963 case NEON::BI__builtin_neon_vsubhn_v: {
6964 llvm::FixedVectorType *SrcTy =
6965 llvm::FixedVectorType::getExtendedElementVectorType(VTy);
6966
6967 // %sum = add <4 x i32> %lhs, %rhs
6968 Ops[0] = Builder.CreateBitCast(Ops[0], SrcTy);
6969 Ops[1] = Builder.CreateBitCast(Ops[1], SrcTy);
6970 Ops[0] = Builder.CreateSub(Ops[0], Ops[1], "vsubhn");
6971
6972 // %high = lshr <4 x i32> %sum, <i32 16, i32 16, i32 16, i32 16>
6973 Constant *ShiftAmt =
6974 ConstantInt::get(SrcTy, SrcTy->getScalarSizeInBits() / 2);
6975 Ops[0] = Builder.CreateLShr(Ops[0], ShiftAmt, "vsubhn");
6976
6977 // %res = trunc <4 x i32> %high to <4 x i16>
6978 return Builder.CreateTrunc(Ops[0], VTy, "vsubhn");
6979 }
6980 case NEON::BI__builtin_neon_vtrn_v:
6981 case NEON::BI__builtin_neon_vtrnq_v: {
6982 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty));
6983 Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
6984 Ops[2] = Builder.CreateBitCast(Ops[2], Ty);
6985 Value *SV = nullptr;
6986
6987 for (unsigned vi = 0; vi != 2; ++vi) {
6988 SmallVector<int, 16> Indices;
6989 for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) {
6990 Indices.push_back(i+vi);
6991 Indices.push_back(i+e+vi);
6992 }
6993 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ty, Ops[0], vi);
6994 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], Indices, "vtrn");
6995 SV = Builder.CreateDefaultAlignedStore(SV, Addr);
6996 }
6997 return SV;
6998 }
6999 case NEON::BI__builtin_neon_vtst_v:
7000 case NEON::BI__builtin_neon_vtstq_v: {
7001 Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
7002 Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
7003 Ops[0] = Builder.CreateAnd(Ops[0], Ops[1]);
7004 Ops[0] = Builder.CreateICmp(ICmpInst::ICMP_NE, Ops[0],
7005 ConstantAggregateZero::get(Ty));
7006 return Builder.CreateSExt(Ops[0], Ty, "vtst");
7007 }
7008 case NEON::BI__builtin_neon_vuzp_v:
7009 case NEON::BI__builtin_neon_vuzpq_v: {
7010 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty));
7011 Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
7012 Ops[2] = Builder.CreateBitCast(Ops[2], Ty);
7013 Value *SV = nullptr;
7014
7015 for (unsigned vi = 0; vi != 2; ++vi) {
7016 SmallVector<int, 16> Indices;
7017 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i)
7018 Indices.push_back(2*i+vi);
7019
7020 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ty, Ops[0], vi);
7021 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], Indices, "vuzp");
7022 SV = Builder.CreateDefaultAlignedStore(SV, Addr);
7023 }
7024 return SV;
7025 }
7026 case NEON::BI__builtin_neon_vxarq_v: {
7027 Function *F = CGM.getIntrinsic(Int);
7028 Ops[2] = Builder.CreateZExt(Ops[2], Int64Ty);
7029 return EmitNeonCall(F, Ops, "");
7030 }
7031 case NEON::BI__builtin_neon_vzip_v:
7032 case NEON::BI__builtin_neon_vzipq_v: {
7033 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty));
7034 Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
7035 Ops[2] = Builder.CreateBitCast(Ops[2], Ty);
7036 Value *SV = nullptr;
7037
7038 for (unsigned vi = 0; vi != 2; ++vi) {
7039 SmallVector<int, 16> Indices;
7040 for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) {
7041 Indices.push_back((i + vi*e) >> 1);
7042 Indices.push_back(((i + vi*e) >> 1)+e);
7043 }
7044 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ty, Ops[0], vi);
7045 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], Indices, "vzip");
7046 SV = Builder.CreateDefaultAlignedStore(SV, Addr);
7047 }
7048 return SV;
7049 }
7050 case NEON::BI__builtin_neon_vdot_v:
7051 case NEON::BI__builtin_neon_vdotq_v: {
7052 auto *InputTy =
7053 llvm::FixedVectorType::get(Int8Ty, Ty->getPrimitiveSizeInBits() / 8);
7054 llvm::Type *Tys[2] = { Ty, InputTy };
7055 Int = Usgn ? LLVMIntrinsic : AltLLVMIntrinsic;
7056 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vdot");
7057 }
7058 case NEON::BI__builtin_neon_vfmlal_low_v:
7059 case NEON::BI__builtin_neon_vfmlalq_low_v: {
7060 auto *InputTy =
7061 llvm::FixedVectorType::get(HalfTy, Ty->getPrimitiveSizeInBits() / 16);
7062 llvm::Type *Tys[2] = { Ty, InputTy };
7063 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vfmlal_low");
7064 }
7065 case NEON::BI__builtin_neon_vfmlsl_low_v:
7066 case NEON::BI__builtin_neon_vfmlslq_low_v: {
7067 auto *InputTy =
7068 llvm::FixedVectorType::get(HalfTy, Ty->getPrimitiveSizeInBits() / 16);
7069 llvm::Type *Tys[2] = { Ty, InputTy };
7070 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vfmlsl_low");
7071 }
7072 case NEON::BI__builtin_neon_vfmlal_high_v:
7073 case NEON::BI__builtin_neon_vfmlalq_high_v: {
7074 auto *InputTy =
7075 llvm::FixedVectorType::get(HalfTy, Ty->getPrimitiveSizeInBits() / 16);
7076 llvm::Type *Tys[2] = { Ty, InputTy };
7077 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vfmlal_high");
7078 }
7079 case NEON::BI__builtin_neon_vfmlsl_high_v:
7080 case NEON::BI__builtin_neon_vfmlslq_high_v: {
7081 auto *InputTy =
7082 llvm::FixedVectorType::get(HalfTy, Ty->getPrimitiveSizeInBits() / 16);
7083 llvm::Type *Tys[2] = { Ty, InputTy };
7084 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vfmlsl_high");
7085 }
7086 case NEON::BI__builtin_neon_vmmlaq_v: {
7087 auto *InputTy =
7088 llvm::FixedVectorType::get(Int8Ty, Ty->getPrimitiveSizeInBits() / 8);
7089 llvm::Type *Tys[2] = { Ty, InputTy };
7090 Int = Usgn ? LLVMIntrinsic : AltLLVMIntrinsic;
7091 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmmla");
7092 }
7093 case NEON::BI__builtin_neon_vusmmlaq_v: {
7094 auto *InputTy =
7095 llvm::FixedVectorType::get(Int8Ty, Ty->getPrimitiveSizeInBits() / 8);
7096 llvm::Type *Tys[2] = { Ty, InputTy };
7097 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vusmmla");
7098 }
7099 case NEON::BI__builtin_neon_vusdot_v:
7100 case NEON::BI__builtin_neon_vusdotq_v: {
7101 auto *InputTy =
7102 llvm::FixedVectorType::get(Int8Ty, Ty->getPrimitiveSizeInBits() / 8);
7103 llvm::Type *Tys[2] = { Ty, InputTy };
7104 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vusdot");
7105 }
7106 case NEON::BI__builtin_neon_vbfdot_v:
7107 case NEON::BI__builtin_neon_vbfdotq_v: {
7108 llvm::Type *InputTy =
7109 llvm::FixedVectorType::get(BFloatTy, Ty->getPrimitiveSizeInBits() / 16);
7110 llvm::Type *Tys[2] = { Ty, InputTy };
7111 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vbfdot");
7112 }
7113 case NEON::BI__builtin_neon___a32_vcvt_bf16_v: {
7114 llvm::Type *Tys[1] = { Ty };
7115 Function *F = CGM.getIntrinsic(Int, Tys);
7116 return EmitNeonCall(F, Ops, "vcvtfp2bf");
7117 }
7118
7119 }
7120
7121 assert(Int && "Expected valid intrinsic number")((void)0);
7122
7123 // Determine the type(s) of this overloaded AArch64 intrinsic.
7124 Function *F = LookupNeonLLVMIntrinsic(Int, Modifier, Ty, E);
7125
7126 Value *Result = EmitNeonCall(F, Ops, NameHint);
7127 llvm::Type *ResultType = ConvertType(E->getType());
7128 // AArch64 intrinsic one-element vector type cast to
7129 // scalar type expected by the builtin
7130 return Builder.CreateBitCast(Result, ResultType, NameHint);
7131}
7132
7133Value *CodeGenFunction::EmitAArch64CompareBuiltinExpr(
7134 Value *Op, llvm::Type *Ty, const CmpInst::Predicate Fp,
7135 const CmpInst::Predicate Ip, const Twine &Name) {
7136 llvm::Type *OTy = Op->getType();
7137
7138 // FIXME: this is utterly horrific. We should not be looking at previous
7139 // codegen context to find out what needs doing. Unfortunately TableGen
7140 // currently gives us exactly the same calls for vceqz_f32 and vceqz_s32
7141 // (etc).
7142 if (BitCastInst *BI = dyn_cast<BitCastInst>(Op))
7143 OTy = BI->getOperand(0)->getType();
7144
7145 Op = Builder.CreateBitCast(Op, OTy);
7146 if (OTy->getScalarType()->isFloatingPointTy()) {
7147 Op = Builder.CreateFCmp(Fp, Op, Constant::getNullValue(OTy));
7148 } else {
7149 Op = Builder.CreateICmp(Ip, Op, Constant::getNullValue(OTy));
7150 }
7151 return Builder.CreateSExt(Op, Ty, Name);
7152}
7153
7154static Value *packTBLDVectorList(CodeGenFunction &CGF, ArrayRef<Value *> Ops,
7155 Value *ExtOp, Value *IndexOp,
7156 llvm::Type *ResTy, unsigned IntID,
7157 const char *Name) {
7158 SmallVector<Value *, 2> TblOps;
7159 if (ExtOp)
7160 TblOps.push_back(ExtOp);
7161
7162 // Build a vector containing sequential number like (0, 1, 2, ..., 15)
7163 SmallVector<int, 16> Indices;
7164 auto *TblTy = cast<llvm::FixedVectorType>(Ops[0]->getType());
7165 for (unsigned i = 0, e = TblTy->getNumElements(); i != e; ++i) {
7166 Indices.push_back(2*i);
7167 Indices.push_back(2*i+1);
7168 }
7169
7170 int PairPos = 0, End = Ops.size() - 1;
7171 while (PairPos < End) {
7172 TblOps.push_back(CGF.Builder.CreateShuffleVector(Ops[PairPos],
7173 Ops[PairPos+1], Indices,
7174 Name));
7175 PairPos += 2;
7176 }
7177
7178 // If there's an odd number of 64-bit lookup table, fill the high 64-bit
7179 // of the 128-bit lookup table with zero.
7180 if (PairPos == End) {
7181 Value *ZeroTbl = ConstantAggregateZero::get(TblTy);
7182 TblOps.push_back(CGF.Builder.CreateShuffleVector(Ops[PairPos],
7183 ZeroTbl, Indices, Name));
7184 }
7185
7186 Function *TblF;
7187 TblOps.push_back(IndexOp);
7188 TblF = CGF.CGM.getIntrinsic(IntID, ResTy);
7189
7190 return CGF.EmitNeonCall(TblF, TblOps, Name);
7191}
7192
7193Value *CodeGenFunction::GetValueForARMHint(unsigned BuiltinID) {
7194 unsigned Value;
7195 switch (BuiltinID) {
7196 default:
7197 return nullptr;
7198 case ARM::BI__builtin_arm_nop:
7199 Value = 0;
7200 break;
7201 case ARM::BI__builtin_arm_yield:
7202 case ARM::BI__yield:
7203 Value = 1;
7204 break;
7205 case ARM::BI__builtin_arm_wfe:
7206 case ARM::BI__wfe:
7207 Value = 2;
7208 break;
7209 case ARM::BI__builtin_arm_wfi:
7210 case ARM::BI__wfi:
7211 Value = 3;
7212 break;
7213 case ARM::BI__builtin_arm_sev:
7214 case ARM::BI__sev:
7215 Value = 4;
7216 break;
7217 case ARM::BI__builtin_arm_sevl:
7218 case ARM::BI__sevl:
7219 Value = 5;
7220 break;
7221 }
7222
7223 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_hint),
7224 llvm::ConstantInt::get(Int32Ty, Value));
7225}
7226
7227enum SpecialRegisterAccessKind {
7228 NormalRead,
7229 VolatileRead,
7230 Write,
7231};
7232
7233// Generates the IR for the read/write special register builtin,
7234// ValueType is the type of the value that is to be written or read,
7235// RegisterType is the type of the register being written to or read from.
7236static Value *EmitSpecialRegisterBuiltin(CodeGenFunction &CGF,
7237 const CallExpr *E,
7238 llvm::Type *RegisterType,
7239 llvm::Type *ValueType,
7240 SpecialRegisterAccessKind AccessKind,
7241 StringRef SysReg = "") {
7242 // write and register intrinsics only support 32 and 64 bit operations.
7243 assert((RegisterType->isIntegerTy(32) || RegisterType->isIntegerTy(64))((void)0)
7244 && "Unsupported size for register.")((void)0);
7245
7246 CodeGen::CGBuilderTy &Builder = CGF.Builder;
7247 CodeGen::CodeGenModule &CGM = CGF.CGM;
7248 LLVMContext &Context = CGM.getLLVMContext();
7249
7250 if (SysReg.empty()) {
7251 const Expr *SysRegStrExpr = E->getArg(0)->IgnoreParenCasts();
7252 SysReg = cast<clang::StringLiteral>(SysRegStrExpr)->getString();
7253 }
7254
7255 llvm::Metadata *Ops[] = { llvm::MDString::get(Context, SysReg) };
7256 llvm::MDNode *RegName = llvm::MDNode::get(Context, Ops);
7257 llvm::Value *Metadata = llvm::MetadataAsValue::get(Context, RegName);
7258
7259 llvm::Type *Types[] = { RegisterType };
7260
7261 bool MixedTypes = RegisterType->isIntegerTy(64) && ValueType->isIntegerTy(32);
7262 assert(!(RegisterType->isIntegerTy(32) && ValueType->isIntegerTy(64))((void)0)
7263 && "Can't fit 64-bit value in 32-bit register")((void)0);
7264
7265 if (AccessKind != Write) {
7266 assert(AccessKind == NormalRead || AccessKind == VolatileRead)((void)0);
7267 llvm::Function *F = CGM.getIntrinsic(
7268 AccessKind == VolatileRead ? llvm::Intrinsic::read_volatile_register
7269 : llvm::Intrinsic::read_register,
7270 Types);
7271 llvm::Value *Call = Builder.CreateCall(F, Metadata);
7272
7273 if (MixedTypes)
7274 // Read into 64 bit register and then truncate result to 32 bit.
7275 return Builder.CreateTrunc(Call, ValueType);
7276
7277 if (ValueType->isPointerTy())
7278 // Have i32/i64 result (Call) but want to return a VoidPtrTy (i8*).
7279 return Builder.CreateIntToPtr(Call, ValueType);
7280
7281 return Call;
7282 }
7283
7284 llvm::Function *F = CGM.getIntrinsic(llvm::Intrinsic::write_register, Types);
7285 llvm::Value *ArgValue = CGF.EmitScalarExpr(E->getArg(1));
7286 if (MixedTypes) {
7287 // Extend 32 bit write value to 64 bit to pass to write.
7288 ArgValue = Builder.CreateZExt(ArgValue, RegisterType);
7289 return Builder.CreateCall(F, { Metadata, ArgValue });
7290 }
7291
7292 if (ValueType->isPointerTy()) {
7293 // Have VoidPtrTy ArgValue but want to return an i32/i64.
7294 ArgValue = Builder.CreatePtrToInt(ArgValue, RegisterType);
7295 return Builder.CreateCall(F, { Metadata, ArgValue });
7296 }
7297
7298 return Builder.CreateCall(F, { Metadata, ArgValue });
7299}
7300
7301/// Return true if BuiltinID is an overloaded Neon intrinsic with an extra
7302/// argument that specifies the vector type.
7303static bool HasExtraNeonArgument(unsigned BuiltinID) {
7304 switch (BuiltinID) {
7305 default: break;
7306 case NEON::BI__builtin_neon_vget_lane_i8:
7307 case NEON::BI__builtin_neon_vget_lane_i16:
7308 case NEON::BI__builtin_neon_vget_lane_bf16:
7309 case NEON::BI__builtin_neon_vget_lane_i32:
7310 case NEON::BI__builtin_neon_vget_lane_i64:
7311 case NEON::BI__builtin_neon_vget_lane_f32:
7312 case NEON::BI__builtin_neon_vgetq_lane_i8:
7313 case NEON::BI__builtin_neon_vgetq_lane_i16:
7314 case NEON::BI__builtin_neon_vgetq_lane_bf16:
7315 case NEON::BI__builtin_neon_vgetq_lane_i32:
7316 case NEON::BI__builtin_neon_vgetq_lane_i64:
7317 case NEON::BI__builtin_neon_vgetq_lane_f32:
7318 case NEON::BI__builtin_neon_vduph_lane_bf16:
7319 case NEON::BI__builtin_neon_vduph_laneq_bf16:
7320 case NEON::BI__builtin_neon_vset_lane_i8:
7321 case NEON::BI__builtin_neon_vset_lane_i16:
7322 case NEON::BI__builtin_neon_vset_lane_bf16:
7323 case NEON::BI__builtin_neon_vset_lane_i32:
7324 case NEON::BI__builtin_neon_vset_lane_i64:
7325 case NEON::BI__builtin_neon_vset_lane_f32:
7326 case NEON::BI__builtin_neon_vsetq_lane_i8:
7327 case NEON::BI__builtin_neon_vsetq_lane_i16:
7328 case NEON::BI__builtin_neon_vsetq_lane_bf16:
7329 case NEON::BI__builtin_neon_vsetq_lane_i32:
7330 case NEON::BI__builtin_neon_vsetq_lane_i64:
7331 case NEON::BI__builtin_neon_vsetq_lane_f32:
7332 case NEON::BI__builtin_neon_vsha1h_u32:
7333 case NEON::BI__builtin_neon_vsha1cq_u32:
7334 case NEON::BI__builtin_neon_vsha1pq_u32:
7335 case NEON::BI__builtin_neon_vsha1mq_u32:
7336 case NEON::BI__builtin_neon_vcvth_bf16_f32:
7337 case clang::ARM::BI_MoveToCoprocessor:
7338 case clang::ARM::BI_MoveToCoprocessor2:
7339 return false;
7340 }
7341 return true;
7342}
7343
7344Value *CodeGenFunction::EmitARMBuiltinExpr(unsigned BuiltinID,
7345 const CallExpr *E,
7346 ReturnValueSlot ReturnValue,
7347 llvm::Triple::ArchType Arch) {
7348 if (auto Hint = GetValueForARMHint(BuiltinID))
7349 return Hint;
7350
7351 if (BuiltinID == ARM::BI__emit) {
7352 bool IsThumb = getTarget().getTriple().getArch() == llvm::Triple::thumb;
7353 llvm::FunctionType *FTy =
7354 llvm::FunctionType::get(VoidTy, /*Variadic=*/false);
7355
7356 Expr::EvalResult Result;
7357 if (!E->getArg(0)->EvaluateAsInt(Result, CGM.getContext()))
7358 llvm_unreachable("Sema will ensure that the parameter is constant")__builtin_unreachable();
7359
7360 llvm::APSInt Value = Result.Val.getInt();
7361 uint64_t ZExtValue = Value.zextOrTrunc(IsThumb ? 16 : 32).getZExtValue();
7362
7363 llvm::InlineAsm *Emit =
7364 IsThumb ? InlineAsm::get(FTy, ".inst.n 0x" + utohexstr(ZExtValue), "",
7365 /*hasSideEffects=*/true)
7366 : InlineAsm::get(FTy, ".inst 0x" + utohexstr(ZExtValue), "",
7367 /*hasSideEffects=*/true);
7368
7369 return Builder.CreateCall(Emit);
7370 }
7371
7372 if (BuiltinID == ARM::BI__builtin_arm_dbg) {
7373 Value *Option = EmitScalarExpr(E->getArg(0));
7374 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_dbg), Option);
7375 }
7376
7377 if (BuiltinID == ARM::BI__builtin_arm_prefetch) {
7378 Value *Address = EmitScalarExpr(E->getArg(0));
7379 Value *RW = EmitScalarExpr(E->getArg(1));
7380 Value *IsData = EmitScalarExpr(E->getArg(2));
7381
7382 // Locality is not supported on ARM target
7383 Value *Locality = llvm::ConstantInt::get(Int32Ty, 3);
7384
7385 Function *F = CGM.getIntrinsic(Intrinsic::prefetch, Address->getType());
7386 return Builder.CreateCall(F, {Address, RW, Locality, IsData});
7387 }
7388
7389 if (BuiltinID == ARM::BI__builtin_arm_rbit) {
7390 llvm::Value *Arg = EmitScalarExpr(E->getArg(0));
7391 return Builder.CreateCall(
7392 CGM.getIntrinsic(Intrinsic::bitreverse, Arg->getType()), Arg, "rbit");
7393 }
7394
7395 if (BuiltinID == ARM::BI__builtin_arm_cls) {
7396 llvm::Value *Arg = EmitScalarExpr(E->getArg(0));
7397 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_cls), Arg, "cls");
7398 }
7399 if (BuiltinID == ARM::BI__builtin_arm_cls64) {
7400 llvm::Value *Arg = EmitScalarExpr(E->getArg(0));
7401 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_cls64), Arg,
7402 "cls");
7403 }
7404
7405 if (BuiltinID == ARM::BI__clear_cache) {
7406 assert(E->getNumArgs() == 2 && "__clear_cache takes 2 arguments")((void)0);
7407 const FunctionDecl *FD = E->getDirectCallee();
7408 Value *Ops[2];
7409 for (unsigned i = 0; i < 2; i++)
7410 Ops[i] = EmitScalarExpr(E->getArg(i));
7411 llvm::Type *Ty = CGM.getTypes().ConvertType(FD->getType());
7412 llvm::FunctionType *FTy = cast<llvm::FunctionType>(Ty);
7413 StringRef Name = FD->getName();
7414 return EmitNounwindRuntimeCall(CGM.CreateRuntimeFunction(FTy, Name), Ops);
7415 }
7416
7417 if (BuiltinID == ARM::BI__builtin_arm_mcrr ||
7418 BuiltinID == ARM::BI__builtin_arm_mcrr2) {
7419 Function *F;
7420
7421 switch (BuiltinID) {
7422 default: llvm_unreachable("unexpected builtin")__builtin_unreachable();
7423 case ARM::BI__builtin_arm_mcrr:
7424 F = CGM.getIntrinsic(Intrinsic::arm_mcrr);
7425 break;
7426 case ARM::BI__builtin_arm_mcrr2:
7427 F = CGM.getIntrinsic(Intrinsic::arm_mcrr2);
7428 break;
7429 }
7430
7431 // MCRR{2} instruction has 5 operands but
7432 // the intrinsic has 4 because Rt and Rt2
7433 // are represented as a single unsigned 64
7434 // bit integer in the intrinsic definition
7435 // but internally it's represented as 2 32
7436 // bit integers.
7437
7438 Value *Coproc = EmitScalarExpr(E->getArg(0));
7439 Value *Opc1 = EmitScalarExpr(E->getArg(1));
7440 Value *RtAndRt2 = EmitScalarExpr(E->getArg(2));
7441 Value *CRm = EmitScalarExpr(E->getArg(3));
7442
7443 Value *C1 = llvm::ConstantInt::get(Int64Ty, 32);
7444 Value *Rt = Builder.CreateTruncOrBitCast(RtAndRt2, Int32Ty);
7445 Value *Rt2 = Builder.CreateLShr(RtAndRt2, C1);
7446 Rt2 = Builder.CreateTruncOrBitCast(Rt2, Int32Ty);
7447
7448 return Builder.CreateCall(F, {Coproc, Opc1, Rt, Rt2, CRm});
7449 }
7450
7451 if (BuiltinID == ARM::BI__builtin_arm_mrrc ||
7452 BuiltinID == ARM::BI__builtin_arm_mrrc2) {
7453 Function *F;
7454
7455 switch (BuiltinID) {
7456 default: llvm_unreachable("unexpected builtin")__builtin_unreachable();
7457 case ARM::BI__builtin_arm_mrrc:
7458 F = CGM.getIntrinsic(Intrinsic::arm_mrrc);
7459 break;
7460 case ARM::BI__builtin_arm_mrrc2:
7461 F = CGM.getIntrinsic(Intrinsic::arm_mrrc2);
7462 break;
7463 }
7464
7465 Value *Coproc = EmitScalarExpr(E->getArg(0));
7466 Value *Opc1 = EmitScalarExpr(E->getArg(1));
7467 Value *CRm = EmitScalarExpr(E->getArg(2));
7468 Value *RtAndRt2 = Builder.CreateCall(F, {Coproc, Opc1, CRm});
7469
7470 // Returns an unsigned 64 bit integer, represented
7471 // as two 32 bit integers.
7472
7473 Value *Rt = Builder.CreateExtractValue(RtAndRt2, 1);
7474 Value *Rt1 = Builder.CreateExtractValue(RtAndRt2, 0);
7475 Rt = Builder.CreateZExt(Rt, Int64Ty);
7476 Rt1 = Builder.CreateZExt(Rt1, Int64Ty);
7477
7478 Value *ShiftCast = llvm::ConstantInt::get(Int64Ty, 32);
7479 RtAndRt2 = Builder.CreateShl(Rt, ShiftCast, "shl", true);
7480 RtAndRt2 = Builder.CreateOr(RtAndRt2, Rt1);
7481
7482 return Builder.CreateBitCast(RtAndRt2, ConvertType(E->getType()));
7483 }
7484
7485 if (BuiltinID == ARM::BI__builtin_arm_ldrexd ||
7486 ((BuiltinID == ARM::BI__builtin_arm_ldrex ||
7487 BuiltinID == ARM::BI__builtin_arm_ldaex) &&
7488 getContext().getTypeSize(E->getType()) == 64) ||
7489 BuiltinID == ARM::BI__ldrexd) {
7490 Function *F;
7491
7492 switch (BuiltinID) {
7493 default: llvm_unreachable("unexpected builtin")__builtin_unreachable();
7494 case ARM::BI__builtin_arm_ldaex:
7495 F = CGM.getIntrinsic(Intrinsic::arm_ldaexd);
7496 break;
7497 case ARM::BI__builtin_arm_ldrexd:
7498 case ARM::BI__builtin_arm_ldrex:
7499 case ARM::BI__ldrexd:
7500 F = CGM.getIntrinsic(Intrinsic::arm_ldrexd);
7501 break;
7502 }
7503
7504 Value *LdPtr = EmitScalarExpr(E->getArg(0));
7505 Value *Val = Builder.CreateCall(F, Builder.CreateBitCast(LdPtr, Int8PtrTy),
7506 "ldrexd");
7507
7508 Value *Val0 = Builder.CreateExtractValue(Val, 1);
7509 Value *Val1 = Builder.CreateExtractValue(Val, 0);
7510 Val0 = Builder.CreateZExt(Val0, Int64Ty);
7511 Val1 = Builder.CreateZExt(Val1, Int64Ty);
7512
7513 Value *ShiftCst = llvm::ConstantInt::get(Int64Ty, 32);
7514 Val = Builder.CreateShl(Val0, ShiftCst, "shl", true /* nuw */);
7515 Val = Builder.CreateOr(Val, Val1);
7516 return Builder.CreateBitCast(Val, ConvertType(E->getType()));
7517 }
7518
7519 if (BuiltinID == ARM::BI__builtin_arm_ldrex ||
7520 BuiltinID == ARM::BI__builtin_arm_ldaex) {
7521 Value *LoadAddr = EmitScalarExpr(E->getArg(0));
7522
7523 QualType Ty = E->getType();
7524 llvm::Type *RealResTy = ConvertType(Ty);
7525 llvm::Type *PtrTy = llvm::IntegerType::get(
7526 getLLVMContext(), getContext().getTypeSize(Ty))->getPointerTo();
7527 LoadAddr = Builder.CreateBitCast(LoadAddr, PtrTy);
7528
7529 Function *F = CGM.getIntrinsic(BuiltinID == ARM::BI__builtin_arm_ldaex
7530 ? Intrinsic::arm_ldaex
7531 : Intrinsic::arm_ldrex,
7532 PtrTy);
7533 Value *Val = Builder.CreateCall(F, LoadAddr, "ldrex");
7534
7535 if (RealResTy->isPointerTy())
7536 return Builder.CreateIntToPtr(Val, RealResTy);
7537 else {
7538 llvm::Type *IntResTy = llvm::IntegerType::get(
7539 getLLVMContext(), CGM.getDataLayout().getTypeSizeInBits(RealResTy));
7540 Val = Builder.CreateTruncOrBitCast(Val, IntResTy);
7541 return Builder.CreateBitCast(Val, RealResTy);
7542 }
7543 }
7544
7545 if (BuiltinID == ARM::BI__builtin_arm_strexd ||
7546 ((BuiltinID == ARM::BI__builtin_arm_stlex ||
7547 BuiltinID == ARM::BI__builtin_arm_strex) &&
7548 getContext().getTypeSize(E->getArg(0)->getType()) == 64)) {
7549 Function *F = CGM.getIntrinsic(BuiltinID == ARM::BI__builtin_arm_stlex
7550 ? Intrinsic::arm_stlexd
7551 : Intrinsic::arm_strexd);
7552 llvm::Type *STy = llvm::StructType::get(Int32Ty, Int32Ty);
7553
7554 Address Tmp = CreateMemTemp(E->getArg(0)->getType());
7555 Value *Val = EmitScalarExpr(E->getArg(0));
7556 Builder.CreateStore(Val, Tmp);
7557
7558 Address LdPtr = Builder.CreateBitCast(Tmp,llvm::PointerType::getUnqual(STy));
7559 Val = Builder.CreateLoad(LdPtr);
7560
7561 Value *Arg0 = Builder.CreateExtractValue(Val, 0);
7562 Value *Arg1 = Builder.CreateExtractValue(Val, 1);
7563 Value *StPtr = Builder.CreateBitCast(EmitScalarExpr(E->getArg(1)), Int8PtrTy);
7564 return Builder.CreateCall(F, {Arg0, Arg1, StPtr}, "strexd");
7565 }
7566
7567 if (BuiltinID == ARM::BI__builtin_arm_strex ||
7568 BuiltinID == ARM::BI__builtin_arm_stlex) {
7569 Value *StoreVal = EmitScalarExpr(E->getArg(0));
7570 Value *StoreAddr = EmitScalarExpr(E->getArg(1));
7571
7572 QualType Ty = E->getArg(0)->getType();
7573 llvm::Type *StoreTy = llvm::IntegerType::get(getLLVMContext(),
7574 getContext().getTypeSize(Ty));
7575 StoreAddr = Builder.CreateBitCast(StoreAddr, StoreTy->getPointerTo());
7576
7577 if (StoreVal->getType()->isPointerTy())
7578 StoreVal = Builder.CreatePtrToInt(StoreVal, Int32Ty);
7579 else {
7580 llvm::Type *IntTy = llvm::IntegerType::get(
7581 getLLVMContext(),
7582 CGM.getDataLayout().getTypeSizeInBits(StoreVal->getType()));
7583 StoreVal = Builder.CreateBitCast(StoreVal, IntTy);
7584 StoreVal = Builder.CreateZExtOrBitCast(StoreVal, Int32Ty);
7585 }
7586
7587 Function *F = CGM.getIntrinsic(BuiltinID == ARM::BI__builtin_arm_stlex
7588 ? Intrinsic::arm_stlex
7589 : Intrinsic::arm_strex,
7590 StoreAddr->getType());
7591 return Builder.CreateCall(F, {StoreVal, StoreAddr}, "strex");
7592 }
7593
7594 if (BuiltinID == ARM::BI__builtin_arm_clrex) {
7595 Function *F = CGM.getIntrinsic(Intrinsic::arm_clrex);
7596 return Builder.CreateCall(F);
7597 }
7598
7599 // CRC32
7600 Intrinsic::ID CRCIntrinsicID = Intrinsic::not_intrinsic;
7601 switch (BuiltinID) {
7602 case ARM::BI__builtin_arm_crc32b:
7603 CRCIntrinsicID = Intrinsic::arm_crc32b; break;
7604 case ARM::BI__builtin_arm_crc32cb:
7605 CRCIntrinsicID = Intrinsic::arm_crc32cb; break;
7606 case ARM::BI__builtin_arm_crc32h:
7607 CRCIntrinsicID = Intrinsic::arm_crc32h; break;
7608 case ARM::BI__builtin_arm_crc32ch:
7609 CRCIntrinsicID = Intrinsic::arm_crc32ch; break;
7610 case ARM::BI__builtin_arm_crc32w:
7611 case ARM::BI__builtin_arm_crc32d:
7612 CRCIntrinsicID = Intrinsic::arm_crc32w; break;
7613 case ARM::BI__builtin_arm_crc32cw:
7614 case ARM::BI__builtin_arm_crc32cd:
7615 CRCIntrinsicID = Intrinsic::arm_crc32cw; break;
7616 }
7617
7618 if (CRCIntrinsicID != Intrinsic::not_intrinsic) {
7619 Value *Arg0 = EmitScalarExpr(E->getArg(0));
7620 Value *Arg1 = EmitScalarExpr(E->getArg(1));
7621
7622 // crc32{c,}d intrinsics are implemnted as two calls to crc32{c,}w
7623 // intrinsics, hence we need different codegen for these cases.
7624 if (BuiltinID == ARM::BI__builtin_arm_crc32d ||
7625 BuiltinID == ARM::BI__builtin_arm_crc32cd) {
7626 Value *C1 = llvm::ConstantInt::get(Int64Ty, 32);
7627 Value *Arg1a = Builder.CreateTruncOrBitCast(Arg1, Int32Ty);
7628 Value *Arg1b = Builder.CreateLShr(Arg1, C1);
7629 Arg1b = Builder.CreateTruncOrBitCast(Arg1b, Int32Ty);
7630
7631 Function *F = CGM.getIntrinsic(CRCIntrinsicID);
7632 Value *Res = Builder.CreateCall(F, {Arg0, Arg1a});
7633 return Builder.CreateCall(F, {Res, Arg1b});
7634 } else {
7635 Arg1 = Builder.CreateZExtOrBitCast(Arg1, Int32Ty);
7636
7637 Function *F = CGM.getIntrinsic(CRCIntrinsicID);
7638 return Builder.CreateCall(F, {Arg0, Arg1});
7639 }
7640 }
7641
7642 if (BuiltinID == ARM::BI__builtin_arm_rsr ||
7643 BuiltinID == ARM::BI__builtin_arm_rsr64 ||
7644 BuiltinID == ARM::BI__builtin_arm_rsrp ||
7645 BuiltinID == ARM::BI__builtin_arm_wsr ||
7646 BuiltinID == ARM::BI__builtin_arm_wsr64 ||
7647 BuiltinID == ARM::BI__builtin_arm_wsrp) {
7648
7649 SpecialRegisterAccessKind AccessKind = Write;
7650 if (BuiltinID == ARM::BI__builtin_arm_rsr ||
7651 BuiltinID == ARM::BI__builtin_arm_rsr64 ||
7652 BuiltinID == ARM::BI__builtin_arm_rsrp)
7653 AccessKind = VolatileRead;
7654
7655 bool IsPointerBuiltin = BuiltinID == ARM::BI__builtin_arm_rsrp ||
7656 BuiltinID == ARM::BI__builtin_arm_wsrp;
7657
7658 bool Is64Bit = BuiltinID == ARM::BI__builtin_arm_rsr64 ||
7659 BuiltinID == ARM::BI__builtin_arm_wsr64;
7660
7661 llvm::Type *ValueType;
7662 llvm::Type *RegisterType;
7663 if (IsPointerBuiltin) {
7664 ValueType = VoidPtrTy;
7665 RegisterType = Int32Ty;
7666 } else if (Is64Bit) {
7667 ValueType = RegisterType = Int64Ty;
7668 } else {
7669 ValueType = RegisterType = Int32Ty;
7670 }
7671
7672 return EmitSpecialRegisterBuiltin(*this, E, RegisterType, ValueType,
7673 AccessKind);
7674 }
7675
7676 // Handle MSVC intrinsics before argument evaluation to prevent double
7677 // evaluation.
7678 if (Optional<MSVCIntrin> MsvcIntId = translateArmToMsvcIntrin(BuiltinID))
7679 return EmitMSVCBuiltinExpr(*MsvcIntId, E);
7680
7681 // Deal with MVE builtins
7682 if (Value *Result = EmitARMMVEBuiltinExpr(BuiltinID, E, ReturnValue, Arch))
7683 return Result;
7684 // Handle CDE builtins
7685 if (Value *Result = EmitARMCDEBuiltinExpr(BuiltinID, E, ReturnValue, Arch))
7686 return Result;
7687
7688 // Find out if any arguments are required to be integer constant
7689 // expressions.
7690 unsigned ICEArguments = 0;
7691 ASTContext::GetBuiltinTypeError Error;
7692 getContext().GetBuiltinType(BuiltinID, Error, &ICEArguments);
7693 assert(Error == ASTContext::GE_None && "Should not codegen an error")((void)0);
7694
7695 auto getAlignmentValue32 = [&](Address addr) -> Value* {
7696 return Builder.getInt32(addr.getAlignment().getQuantity());
7697 };
7698
7699 Address PtrOp0 = Address::invalid();
7700 Address PtrOp1 = Address::invalid();
7701 SmallVector<Value*, 4> Ops;
7702 bool HasExtraArg = HasExtraNeonArgument(BuiltinID);
7703 unsigned NumArgs = E->getNumArgs() - (HasExtraArg ? 1 : 0);
7704 for (unsigned i = 0, e = NumArgs; i != e; i++) {
7705 if (i == 0) {
7706 switch (BuiltinID) {
7707 case NEON::BI__builtin_neon_vld1_v:
7708 case NEON::BI__builtin_neon_vld1q_v:
7709 case NEON::BI__builtin_neon_vld1q_lane_v:
7710 case NEON::BI__builtin_neon_vld1_lane_v:
7711 case NEON::BI__builtin_neon_vld1_dup_v:
7712 case NEON::BI__builtin_neon_vld1q_dup_v:
7713 case NEON::BI__builtin_neon_vst1_v:
7714 case NEON::BI__builtin_neon_vst1q_v:
7715 case NEON::BI__builtin_neon_vst1q_lane_v:
7716 case NEON::BI__builtin_neon_vst1_lane_v:
7717 case NEON::BI__builtin_neon_vst2_v:
7718 case NEON::BI__builtin_neon_vst2q_v:
7719 case NEON::BI__builtin_neon_vst2_lane_v:
7720 case NEON::BI__builtin_neon_vst2q_lane_v:
7721 case NEON::BI__builtin_neon_vst3_v:
7722 case NEON::BI__builtin_neon_vst3q_v:
7723 case NEON::BI__builtin_neon_vst3_lane_v:
7724 case NEON::BI__builtin_neon_vst3q_lane_v:
7725 case NEON::BI__builtin_neon_vst4_v:
7726 case NEON::BI__builtin_neon_vst4q_v:
7727 case NEON::BI__builtin_neon_vst4_lane_v:
7728 case NEON::BI__builtin_neon_vst4q_lane_v:
7729 // Get the alignment for the argument in addition to the value;
7730 // we'll use it later.
7731 PtrOp0 = EmitPointerWithAlignment(E->getArg(0));
7732 Ops.push_back(PtrOp0.getPointer());
7733 continue;
7734 }
7735 }
7736 if (i == 1) {
7737 switch (BuiltinID) {
7738 case NEON::BI__builtin_neon_vld2_v:
7739 case NEON::BI__builtin_neon_vld2q_v:
7740 case NEON::BI__builtin_neon_vld3_v:
7741 case NEON::BI__builtin_neon_vld3q_v:
7742 case NEON::BI__builtin_neon_vld4_v:
7743 case NEON::BI__builtin_neon_vld4q_v:
7744 case NEON::BI__builtin_neon_vld2_lane_v:
7745 case NEON::BI__builtin_neon_vld2q_lane_v:
7746 case NEON::BI__builtin_neon_vld3_lane_v:
7747 case NEON::BI__builtin_neon_vld3q_lane_v:
7748 case NEON::BI__builtin_neon_vld4_lane_v:
7749 case NEON::BI__builtin_neon_vld4q_lane_v:
7750 case NEON::BI__builtin_neon_vld2_dup_v:
7751 case NEON::BI__builtin_neon_vld2q_dup_v:
7752 case NEON::BI__builtin_neon_vld3_dup_v:
7753 case NEON::BI__builtin_neon_vld3q_dup_v:
7754 case NEON::BI__builtin_neon_vld4_dup_v:
7755 case NEON::BI__builtin_neon_vld4q_dup_v:
7756 // Get the alignment for the argument in addition to the value;
7757 // we'll use it later.
7758 PtrOp1 = EmitPointerWithAlignment(E->getArg(1));
7759 Ops.push_back(PtrOp1.getPointer());
7760 continue;
7761 }
7762 }
7763
7764 if ((ICEArguments & (1 << i)) == 0) {
7765 Ops.push_back(EmitScalarExpr(E->getArg(i)));
7766 } else {
7767 // If this is required to be a constant, constant fold it so that we know
7768 // that the generated intrinsic gets a ConstantInt.
7769 Ops.push_back(llvm::ConstantInt::get(
7770 getLLVMContext(),
7771 *E->getArg(i)->getIntegerConstantExpr(getContext())));
7772 }
7773 }
7774
7775 switch (BuiltinID) {
7776 default: break;
7777
7778 case NEON::BI__builtin_neon_vget_lane_i8:
7779 case NEON::BI__builtin_neon_vget_lane_i16:
7780 case NEON::BI__builtin_neon_vget_lane_i32:
7781 case NEON::BI__builtin_neon_vget_lane_i64:
7782 case NEON::BI__builtin_neon_vget_lane_bf16:
7783 case NEON::BI__builtin_neon_vget_lane_f32:
7784 case NEON::BI__builtin_neon_vgetq_lane_i8:
7785 case NEON::BI__builtin_neon_vgetq_lane_i16:
7786 case NEON::BI__builtin_neon_vgetq_lane_i32:
7787 case NEON::BI__builtin_neon_vgetq_lane_i64:
7788 case NEON::BI__builtin_neon_vgetq_lane_bf16:
7789 case NEON::BI__builtin_neon_vgetq_lane_f32:
7790 case NEON::BI__builtin_neon_vduph_lane_bf16:
7791 case NEON::BI__builtin_neon_vduph_laneq_bf16:
7792 return Builder.CreateExtractElement(Ops[0], Ops[1], "vget_lane");
7793
7794 case NEON::BI__builtin_neon_vrndns_f32: {
7795 Value *Arg = EmitScalarExpr(E->getArg(0));
7796 llvm::Type *Tys[] = {Arg->getType()};
7797 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vrintn, Tys);
7798 return Builder.CreateCall(F, {Arg}, "vrndn"); }
7799
7800 case NEON::BI__builtin_neon_vset_lane_i8:
7801 case NEON::BI__builtin_neon_vset_lane_i16:
7802 case NEON::BI__builtin_neon_vset_lane_i32:
7803 case NEON::BI__builtin_neon_vset_lane_i64:
7804 case NEON::BI__builtin_neon_vset_lane_bf16:
7805 case NEON::BI__builtin_neon_vset_lane_f32:
7806 case NEON::BI__builtin_neon_vsetq_lane_i8:
7807 case NEON::BI__builtin_neon_vsetq_lane_i16:
7808 case NEON::BI__builtin_neon_vsetq_lane_i32:
7809 case NEON::BI__builtin_neon_vsetq_lane_i64:
7810 case NEON::BI__builtin_neon_vsetq_lane_bf16:
7811 case NEON::BI__builtin_neon_vsetq_lane_f32:
7812 return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vset_lane");
7813
7814 case NEON::BI__builtin_neon_vsha1h_u32:
7815 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_sha1h), Ops,
7816 "vsha1h");
7817 case NEON::BI__builtin_neon_vsha1cq_u32:
7818 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_sha1c), Ops,
7819 "vsha1h");
7820 case NEON::BI__builtin_neon_vsha1pq_u32:
7821 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_sha1p), Ops,
7822 "vsha1h");
7823 case NEON::BI__builtin_neon_vsha1mq_u32:
7824 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_sha1m), Ops,
7825 "vsha1h");
7826
7827 case NEON::BI__builtin_neon_vcvth_bf16_f32: {
7828 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vcvtbfp2bf), Ops,
7829 "vcvtbfp2bf");
7830 }
7831
7832 // The ARM _MoveToCoprocessor builtins put the input register value as
7833 // the first argument, but the LLVM intrinsic expects it as the third one.
7834 case ARM::BI_MoveToCoprocessor:
7835 case ARM::BI_MoveToCoprocessor2: {
7836 Function *F = CGM.getIntrinsic(BuiltinID == ARM::BI_MoveToCoprocessor ?
7837 Intrinsic::arm_mcr : Intrinsic::arm_mcr2);
7838 return Builder.CreateCall(F, {Ops[1], Ops[2], Ops[0],
7839 Ops[3], Ops[4], Ops[5]});
7840 }
7841 }
7842
7843 // Get the last argument, which specifies the vector type.
7844 assert(HasExtraArg)((void)0);
7845 const Expr *Arg = E->getArg(E->getNumArgs()-1);
7846 Optional<llvm::APSInt> Result = Arg->getIntegerConstantExpr(getContext());
7847 if (!Result)
7848 return nullptr;
7849
7850 if (BuiltinID == ARM::BI__builtin_arm_vcvtr_f ||
7851 BuiltinID == ARM::BI__builtin_arm_vcvtr_d) {
7852 // Determine the overloaded type of this builtin.
7853 llvm::Type *Ty;
7854 if (BuiltinID == ARM::BI__builtin_arm_vcvtr_f)
7855 Ty = FloatTy;
7856 else
7857 Ty = DoubleTy;
7858
7859 // Determine whether this is an unsigned conversion or not.
7860 bool usgn = Result->getZExtValue() == 1;
7861 unsigned Int = usgn ? Intrinsic::arm_vcvtru : Intrinsic::arm_vcvtr;
7862
7863 // Call the appropriate intrinsic.
7864 Function *F = CGM.getIntrinsic(Int, Ty);
7865 return Builder.CreateCall(F, Ops, "vcvtr");
7866 }
7867
7868 // Determine the type of this overloaded NEON intrinsic.
7869 NeonTypeFlags Type = Result->getZExtValue();
7870 bool usgn = Type.isUnsigned();
7871 bool rightShift = false;
7872
7873 llvm::FixedVectorType *VTy =
7874 GetNeonType(this, Type, getTarget().hasLegalHalfType(), false,
7875 getTarget().hasBFloat16Type());
7876 llvm::Type *Ty = VTy;
7877 if (!Ty)
7878 return nullptr;
7879
7880 // Many NEON builtins have identical semantics and uses in ARM and
7881 // AArch64. Emit these in a single function.
7882 auto IntrinsicMap = makeArrayRef(ARMSIMDIntrinsicMap);
7883 const ARMVectorIntrinsicInfo *Builtin = findARMVectorIntrinsicInMap(
7884 IntrinsicMap, BuiltinID, NEONSIMDIntrinsicsProvenSorted);
7885 if (Builtin)
7886 return EmitCommonNeonBuiltinExpr(
7887 Builtin->BuiltinID, Builtin->LLVMIntrinsic, Builtin->AltLLVMIntrinsic,
7888 Builtin->NameHint, Builtin->TypeModifier, E, Ops, PtrOp0, PtrOp1, Arch);
7889
7890 unsigned Int;
7891 switch (BuiltinID) {
7892 default: return nullptr;
7893 case NEON::BI__builtin_neon_vld1q_lane_v:
7894 // Handle 64-bit integer elements as a special case. Use shuffles of
7895 // one-element vectors to avoid poor code for i64 in the backend.
7896 if (VTy->getElementType()->isIntegerTy(64)) {
7897 // Extract the other lane.
7898 Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
7899 int Lane = cast<ConstantInt>(Ops[2])->getZExtValue();
7900 Value *SV = llvm::ConstantVector::get(ConstantInt::get(Int32Ty, 1-Lane));
7901 Ops[1] = Builder.CreateShuffleVector(Ops[1], Ops[1], SV);
7902 // Load the value as a one-element vector.
7903 Ty = llvm::FixedVectorType::get(VTy->getElementType(), 1);
7904 llvm::Type *Tys[] = {Ty, Int8PtrTy};
7905 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vld1, Tys);
7906 Value *Align = getAlignmentValue32(PtrOp0);
7907 Value *Ld = Builder.CreateCall(F, {Ops[0], Align});
7908 // Combine them.
7909 int Indices[] = {1 - Lane, Lane};
7910 return Builder.CreateShuffleVector(Ops[1], Ld, Indices, "vld1q_lane");
7911 }
7912 LLVM_FALLTHROUGH[[gnu::fallthrough]];
7913 case NEON::BI__builtin_neon_vld1_lane_v: {
7914 Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
7915 PtrOp0 = Builder.CreateElementBitCast(PtrOp0, VTy->getElementType());
7916 Value *Ld = Builder.CreateLoad(PtrOp0);
7917 return Builder.CreateInsertElement(Ops[1], Ld, Ops[2], "vld1_lane");
7918 }
7919 case NEON::BI__builtin_neon_vqrshrn_n_v:
7920 Int =
7921 usgn ? Intrinsic::arm_neon_vqrshiftnu : Intrinsic::arm_neon_vqrshiftns;
7922 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqrshrn_n",
7923 1, true);
7924 case NEON::BI__builtin_neon_vqrshrun_n_v:
7925 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqrshiftnsu, Ty),
7926 Ops, "vqrshrun_n", 1, true);
7927 case NEON::BI__builtin_neon_vqshrn_n_v:
7928 Int = usgn ? Intrinsic::arm_neon_vqshiftnu : Intrinsic::arm_neon_vqshiftns;
7929 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshrn_n",
7930 1, true);
7931 case NEON::BI__builtin_neon_vqshrun_n_v:
7932 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqshiftnsu, Ty),
7933 Ops, "vqshrun_n", 1, true);
7934 case NEON::BI__builtin_neon_vrecpe_v:
7935 case NEON::BI__builtin_neon_vrecpeq_v:
7936 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vrecpe, Ty),
7937 Ops, "vrecpe");
7938 case NEON::BI__builtin_neon_vrshrn_n_v:
7939 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vrshiftn, Ty),
7940 Ops, "vrshrn_n", 1, true);
7941 case NEON::BI__builtin_neon_vrsra_n_v:
7942 case NEON::BI__builtin_neon_vrsraq_n_v:
7943 Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
7944 Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
7945 Ops[2] = EmitNeonShiftVector(Ops[2], Ty, true);
7946 Int = usgn ? Intrinsic::arm_neon_vrshiftu : Intrinsic::arm_neon_vrshifts;
7947 Ops[1] = Builder.CreateCall(CGM.getIntrinsic(Int, Ty), {Ops[1], Ops[2]});
7948 return Builder.CreateAdd(Ops[0], Ops[1], "vrsra_n");
7949 case NEON::BI__builtin_neon_vsri_n_v:
7950 case NEON::BI__builtin_neon_vsriq_n_v:
7951 rightShift = true;
7952 LLVM_FALLTHROUGH[[gnu::fallthrough]];
7953 case NEON::BI__builtin_neon_vsli_n_v:
7954 case NEON::BI__builtin_neon_vsliq_n_v:
7955 Ops[2] = EmitNeonShiftVector(Ops[2], Ty, rightShift);
7956 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vshiftins, Ty),
7957 Ops, "vsli_n");
7958 case NEON::BI__builtin_neon_vsra_n_v:
7959 case NEON::BI__builtin_neon_vsraq_n_v:
7960 Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
7961 Ops[1] = EmitNeonRShiftImm(Ops[1], Ops[2], Ty, usgn, "vsra_n");
7962 return Builder.CreateAdd(Ops[0], Ops[1]);
7963 case NEON::BI__builtin_neon_vst1q_lane_v:
7964 // Handle 64-bit integer elements as a special case. Use a shuffle to get
7965 // a one-element vector and avoid poor code for i64 in the backend.
7966 if (VTy->getElementType()->isIntegerTy(64)) {
7967 Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
7968 Value *SV = llvm::ConstantVector::get(cast<llvm::Constant>(Ops[2]));
7969 Ops[1] = Builder.CreateShuffleVector(Ops[1], Ops[1], SV);
7970 Ops[2] = getAlignmentValue32(PtrOp0);
7971 llvm::Type *Tys[] = {Int8PtrTy, Ops[1]->getType()};
7972 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_neon_vst1,
7973 Tys), Ops);
7974 }
7975 LLVM_FALLTHROUGH[[gnu::fallthrough]];
7976 case NEON::BI__builtin_neon_vst1_lane_v: {
7977 Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
7978 Ops[1] = Builder.CreateExtractElement(Ops[1], Ops[2]);
7979 Ty = llvm::PointerType::getUnqual(Ops[1]->getType());
7980 auto St = Builder.CreateStore(Ops[1], Builder.CreateBitCast(PtrOp0, Ty));
7981 return St;
7982 }
7983 case NEON::BI__builtin_neon_vtbl1_v:
7984 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl1),
7985 Ops, "vtbl1");
7986 case NEON::BI__builtin_neon_vtbl2_v:
7987 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl2),
7988 Ops, "vtbl2");
7989 case NEON::BI__builtin_neon_vtbl3_v:
7990 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl3),
7991 Ops, "vtbl3");
7992 case NEON::BI__builtin_neon_vtbl4_v:
7993 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl4),
7994 Ops, "vtbl4");
7995 case NEON::BI__builtin_neon_vtbx1_v:
7996 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx1),
7997 Ops, "vtbx1");
7998 case NEON::BI__builtin_neon_vtbx2_v:
7999 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx2),
8000 Ops, "vtbx2");
8001 case NEON::BI__builtin_neon_vtbx3_v:
8002 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx3),
8003 Ops, "vtbx3");
8004 case NEON::BI__builtin_neon_vtbx4_v:
8005 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx4),
8006 Ops, "vtbx4");
8007 }
8008}
8009
8010template<typename Integer>
8011static Integer GetIntegerConstantValue(const Expr *E, ASTContext &Context) {
8012 return E->getIntegerConstantExpr(Context)->getExtValue();
8013}
8014
8015static llvm::Value *SignOrZeroExtend(CGBuilderTy &Builder, llvm::Value *V,
8016 llvm::Type *T, bool Unsigned) {
8017 // Helper function called by Tablegen-constructed ARM MVE builtin codegen,
8018 // which finds it convenient to specify signed/unsigned as a boolean flag.
8019 return Unsigned ? Builder.CreateZExt(V, T) : Builder.CreateSExt(V, T);
8020}
8021
8022static llvm::Value *MVEImmediateShr(CGBuilderTy &Builder, llvm::Value *V,
8023 uint32_t Shift, bool Unsigned) {
8024 // MVE helper function for integer shift right. This must handle signed vs
8025 // unsigned, and also deal specially with the case where the shift count is
8026 // equal to the lane size. In LLVM IR, an LShr with that parameter would be
8027 // undefined behavior, but in MVE it's legal, so we must convert it to code
8028 // that is not undefined in IR.
8029 unsigned LaneBits = cast<llvm::VectorType>(V->getType())
8030 ->getElementType()
8031 ->getPrimitiveSizeInBits();
8032 if (Shift == LaneBits) {
8033 // An unsigned shift of the full lane size always generates zero, so we can
8034 // simply emit a zero vector. A signed shift of the full lane size does the
8035 // same thing as shifting by one bit fewer.
8036 if (Unsigned)
8037 return llvm::Constant::getNullValue(V->getType());
8038 else
8039 --Shift;
8040 }
8041 return Unsigned ? Builder.CreateLShr(V, Shift) : Builder.CreateAShr(V, Shift);
8042}
8043
8044static llvm::Value *ARMMVEVectorSplat(CGBuilderTy &Builder, llvm::Value *V) {
8045 // MVE-specific helper function for a vector splat, which infers the element
8046 // count of the output vector by knowing that MVE vectors are all 128 bits
8047 // wide.
8048 unsigned Elements = 128 / V->getType()->getPrimitiveSizeInBits();
8049 return Builder.CreateVectorSplat(Elements, V);
8050}
8051
8052static llvm::Value *ARMMVEVectorReinterpret(CGBuilderTy &Builder,
8053 CodeGenFunction *CGF,
8054 llvm::Value *V,
8055 llvm::Type *DestType) {
8056 // Convert one MVE vector type into another by reinterpreting its in-register
8057 // format.
8058 //
8059 // Little-endian, this is identical to a bitcast (which reinterprets the
8060 // memory format). But big-endian, they're not necessarily the same, because
8061 // the register and memory formats map to each other differently depending on
8062 // the lane size.
8063 //
8064 // We generate a bitcast whenever we can (if we're little-endian, or if the
8065 // lane sizes are the same anyway). Otherwise we fall back to an IR intrinsic
8066 // that performs the different kind of reinterpretation.
8067 if (CGF->getTarget().isBigEndian() &&
8068 V->getType()->getScalarSizeInBits() != DestType->getScalarSizeInBits()) {
8069 return Builder.CreateCall(
8070 CGF->CGM.getIntrinsic(Intrinsic::arm_mve_vreinterpretq,
8071 {DestType, V->getType()}),
8072 V);
8073 } else {
8074 return Builder.CreateBitCast(V, DestType);
8075 }
8076}
8077
8078static llvm::Value *VectorUnzip(CGBuilderTy &Builder, llvm::Value *V, bool Odd) {
8079 // Make a shufflevector that extracts every other element of a vector (evens
8080 // or odds, as desired).
8081 SmallVector<int, 16> Indices;
8082 unsigned InputElements =
8083 cast<llvm::FixedVectorType>(V->getType())->getNumElements();
8084 for (unsigned i = 0; i < InputElements; i += 2)
8085 Indices.push_back(i + Odd);
8086 return Builder.CreateShuffleVector(V, Indices);
8087}
8088
8089static llvm::Value *VectorZip(CGBuilderTy &Builder, llvm::Value *V0,
8090 llvm::Value *V1) {
8091 // Make a shufflevector that interleaves two vectors element by element.
8092 assert(V0->getType() == V1->getType() && "Can't zip different vector types")((void)0);
8093 SmallVector<int, 16> Indices;
8094 unsigned InputElements =
8095 cast<llvm::FixedVectorType>(V0->getType())->getNumElements();
8096 for (unsigned i = 0; i < InputElements; i++) {
8097 Indices.push_back(i);
8098 Indices.push_back(i + InputElements);
8099 }
8100 return Builder.CreateShuffleVector(V0, V1, Indices);
8101}
8102
8103template<unsigned HighBit, unsigned OtherBits>
8104static llvm::Value *ARMMVEConstantSplat(CGBuilderTy &Builder, llvm::Type *VT) {
8105 // MVE-specific helper function to make a vector splat of a constant such as
8106 // UINT_MAX or INT_MIN, in which all bits below the highest one are equal.
8107 llvm::Type *T = cast<llvm::VectorType>(VT)->getElementType();
8108 unsigned LaneBits = T->getPrimitiveSizeInBits();
8109 uint32_t Value = HighBit << (LaneBits - 1);
8110 if (OtherBits)
8111 Value |= (1UL << (LaneBits - 1)) - 1;
8112 llvm::Value *Lane = llvm::ConstantInt::get(T, Value);
8113 return ARMMVEVectorSplat(Builder, Lane);
8114}
8115
8116static llvm::Value *ARMMVEVectorElementReverse(CGBuilderTy &Builder,
8117 llvm::Value *V,
8118 unsigned ReverseWidth) {
8119 // MVE-specific helper function which reverses the elements of a
8120 // vector within every (ReverseWidth)-bit collection of lanes.
8121 SmallVector<int, 16> Indices;
8122 unsigned LaneSize = V->getType()->getScalarSizeInBits();
8123 unsigned Elements = 128 / LaneSize;
8124 unsigned Mask = ReverseWidth / LaneSize - 1;
8125 for (unsigned i = 0; i < Elements; i++)
8126 Indices.push_back(i ^ Mask);
8127 return Builder.CreateShuffleVector(V, Indices);
8128}
8129
8130Value *CodeGenFunction::EmitARMMVEBuiltinExpr(unsigned BuiltinID,
8131 const CallExpr *E,
8132 ReturnValueSlot ReturnValue,
8133 llvm::Triple::ArchType Arch) {
8134 enum class CustomCodeGen { VLD24, VST24 } CustomCodeGenType;
8135 Intrinsic::ID IRIntr;
8136 unsigned NumVectors;
8137
8138 // Code autogenerated by Tablegen will handle all the simple builtins.
8139 switch (BuiltinID) {
8140 #include "clang/Basic/arm_mve_builtin_cg.inc"
8141
8142 // If we didn't match an MVE builtin id at all, go back to the
8143 // main EmitARMBuiltinExpr.
8144 default:
8145 return nullptr;
8146 }
8147
8148 // Anything that breaks from that switch is an MVE builtin that
8149 // needs handwritten code to generate.
8150
8151 switch (CustomCodeGenType) {
8152
8153 case CustomCodeGen::VLD24: {
8154 llvm::SmallVector<Value *, 4> Ops;
8155 llvm::SmallVector<llvm::Type *, 4> Tys;
8156
8157 auto MvecCType = E->getType();
8158 auto MvecLType = ConvertType(MvecCType);
8159 assert(MvecLType->isStructTy() &&((void)0)
8160 "Return type for vld[24]q should be a struct")((void)0);
8161 assert(MvecLType->getStructNumElements() == 1 &&((void)0)
8162 "Return-type struct for vld[24]q should have one element")((void)0);
8163 auto MvecLTypeInner = MvecLType->getStructElementType(0);
8164 assert(MvecLTypeInner->isArrayTy() &&((void)0)
8165 "Return-type struct for vld[24]q should contain an array")((void)0);
8166 assert(MvecLTypeInner->getArrayNumElements() == NumVectors &&((void)0)
8167 "Array member of return-type struct vld[24]q has wrong length")((void)0);
8168 auto VecLType = MvecLTypeInner->getArrayElementType();
8169
8170 Tys.push_back(VecLType);
8171
8172 auto Addr = E->getArg(0);
8173 Ops.push_back(EmitScalarExpr(Addr));
8174 Tys.push_back(ConvertType(Addr->getType()));
8175
8176 Function *F = CGM.getIntrinsic(IRIntr, makeArrayRef(Tys));
8177 Value *LoadResult = Builder.CreateCall(F, Ops);
8178 Value *MvecOut = UndefValue::get(MvecLType);
8179 for (unsigned i = 0; i < NumVectors; ++i) {
8180 Value *Vec = Builder.CreateExtractValue(LoadResult, i);
8181 MvecOut = Builder.CreateInsertValue(MvecOut, Vec, {0, i});
8182 }
8183
8184 if (ReturnValue.isNull())
8185 return MvecOut;
8186 else
8187 return Builder.CreateStore(MvecOut, ReturnValue.getValue());
8188 }
8189
8190 case CustomCodeGen::VST24: {
8191 llvm::SmallVector<Value *, 4> Ops;
8192 llvm::SmallVector<llvm::Type *, 4> Tys;
8193
8194 auto Addr = E->getArg(0);
8195 Ops.push_back(EmitScalarExpr(Addr));
8196 Tys.push_back(ConvertType(Addr->getType()));
8197
8198 auto MvecCType = E->getArg(1)->getType();
8199 auto MvecLType = ConvertType(MvecCType);
8200 assert(MvecLType->isStructTy() && "Data type for vst2q should be a struct")((void)0);
8201 assert(MvecLType->getStructNumElements() == 1 &&((void)0)
8202 "Data-type struct for vst2q should have one element")((void)0);
8203 auto MvecLTypeInner = MvecLType->getStructElementType(0);
8204 assert(MvecLTypeInner->isArrayTy() &&((void)0)
8205 "Data-type struct for vst2q should contain an array")((void)0);
8206 assert(MvecLTypeInner->getArrayNumElements() == NumVectors &&((void)0)
8207 "Array member of return-type struct vld[24]q has wrong length")((void)0);
8208 auto VecLType = MvecLTypeInner->getArrayElementType();
8209
8210 Tys.push_back(VecLType);
8211
8212 AggValueSlot MvecSlot = CreateAggTemp(MvecCType);
8213 EmitAggExpr(E->getArg(1), MvecSlot);
8214 auto Mvec = Builder.CreateLoad(MvecSlot.getAddress());
8215 for (unsigned i = 0; i < NumVectors; i++)
8216 Ops.push_back(Builder.CreateExtractValue(Mvec, {0, i}));
8217
8218 Function *F = CGM.getIntrinsic(IRIntr, makeArrayRef(Tys));
8219 Value *ToReturn = nullptr;
8220 for (unsigned i = 0; i < NumVectors; i++) {
8221 Ops.push_back(llvm::ConstantInt::get(Int32Ty, i));
8222 ToReturn = Builder.CreateCall(F, Ops);
8223 Ops.pop_back();
8224 }
8225 return ToReturn;
8226 }
8227 }
8228 llvm_unreachable("unknown custom codegen type.")__builtin_unreachable();
8229}
8230
8231Value *CodeGenFunction::EmitARMCDEBuiltinExpr(unsigned BuiltinID,
8232 const CallExpr *E,
8233 ReturnValueSlot ReturnValue,
8234 llvm::Triple::ArchType Arch) {
8235 switch (BuiltinID) {
8236 default:
8237 return nullptr;
8238#include "clang/Basic/arm_cde_builtin_cg.inc"
8239 }
8240}
8241
8242static Value *EmitAArch64TblBuiltinExpr(CodeGenFunction &CGF, unsigned BuiltinID,
8243 const CallExpr *E,
8244 SmallVectorImpl<Value *> &Ops,
8245 llvm::Triple::ArchType Arch) {
8246 unsigned int Int = 0;
8247 const char *s = nullptr;
8248
8249 switch (BuiltinID) {
8250 default:
8251 return nullptr;
8252 case NEON::BI__builtin_neon_vtbl1_v:
8253 case NEON::BI__builtin_neon_vqtbl1_v:
8254 case NEON::BI__builtin_neon_vqtbl1q_v:
8255 case NEON::BI__builtin_neon_vtbl2_v:
8256 case NEON::BI__builtin_neon_vqtbl2_v:
8257 case NEON::BI__builtin_neon_vqtbl2q_v:
8258 case NEON::BI__builtin_neon_vtbl3_v:
8259 case NEON::BI__builtin_neon_vqtbl3_v:
8260 case NEON::BI__builtin_neon_vqtbl3q_v:
8261 case NEON::BI__builtin_neon_vtbl4_v:
8262 case NEON::BI__builtin_neon_vqtbl4_v:
8263 case NEON::BI__builtin_neon_vqtbl4q_v:
8264 break;
8265 case NEON::BI__builtin_neon_vtbx1_v:
8266 case NEON::BI__builtin_neon_vqtbx1_v:
8267 case NEON::BI__builtin_neon_vqtbx1q_v:
8268 case NEON::BI__builtin_neon_vtbx2_v:
8269 case NEON::BI__builtin_neon_vqtbx2_v:
8270 case NEON::BI__builtin_neon_vqtbx2q_v:
8271 case NEON::BI__builtin_neon_vtbx3_v:
8272 case NEON::BI__builtin_neon_vqtbx3_v:
8273 case NEON::BI__builtin_neon_vqtbx3q_v:
8274 case NEON::BI__builtin_neon_vtbx4_v:
8275 case NEON::BI__builtin_neon_vqtbx4_v:
8276 case NEON::BI__builtin_neon_vqtbx4q_v:
8277 break;
8278 }
8279
8280 assert(E->getNumArgs() >= 3)((void)0);
8281
8282 // Get the last argument, which specifies the vector type.
8283 const Expr *Arg = E->getArg(E->getNumArgs() - 1);
8284 Optional<llvm::APSInt> Result = Arg->getIntegerConstantExpr(CGF.getContext());
8285 if (!Result)
8286 return nullptr;
8287
8288 // Determine the type of this overloaded NEON intrinsic.
8289 NeonTypeFlags Type = Result->getZExtValue();
8290 llvm::FixedVectorType *Ty = GetNeonType(&CGF, Type);
8291 if (!Ty)
8292 return nullptr;
8293
8294 CodeGen::CGBuilderTy &Builder = CGF.Builder;
8295
8296 // AArch64 scalar builtins are not overloaded, they do not have an extra
8297 // argument that specifies the vector type, need to handle each case.
8298 switch (BuiltinID) {
8299 case NEON::BI__builtin_neon_vtbl1_v: {
8300 return packTBLDVectorList(CGF, makeArrayRef(Ops).slice(0, 1), nullptr,
8301 Ops[1], Ty, Intrinsic::aarch64_neon_tbl1,
8302 "vtbl1");
8303 }
8304 case NEON::BI__builtin_neon_vtbl2_v: {
8305 return packTBLDVectorList(CGF, makeArrayRef(Ops).slice(0, 2), nullptr,
8306 Ops[2], Ty, Intrinsic::aarch64_neon_tbl1,
8307 "vtbl1");
8308 }
8309 case NEON::BI__builtin_neon_vtbl3_v: {
8310 return packTBLDVectorList(CGF, makeArrayRef(Ops).slice(0, 3), nullptr,
8311 Ops[3], Ty, Intrinsic::aarch64_neon_tbl2,
8312 "vtbl2");
8313 }
8314 case NEON::BI__builtin_neon_vtbl4_v: {
8315 return packTBLDVectorList(CGF, makeArrayRef(Ops).slice(0, 4), nullptr,
8316 Ops[4], Ty, Intrinsic::aarch64_neon_tbl2,
8317 "vtbl2");
8318 }
8319 case NEON::BI__builtin_neon_vtbx1_v: {
8320 Value *TblRes =
8321 packTBLDVectorList(CGF, makeArrayRef(Ops).slice(1, 1), nullptr, Ops[2],
8322 Ty, Intrinsic::aarch64_neon_tbl1, "vtbl1");
8323
8324 llvm::Constant *EightV = ConstantInt::get(Ty, 8);
8325 Value *CmpRes = Builder.CreateICmp(ICmpInst::ICMP_UGE, Ops[2], EightV);
8326 CmpRes = Builder.CreateSExt(CmpRes, Ty);
8327
8328 Value *EltsFromInput = Builder.CreateAnd(CmpRes, Ops[0]);
8329 Value *EltsFromTbl = Builder.CreateAnd(Builder.CreateNot(CmpRes), TblRes);
8330 return Builder.CreateOr(EltsFromInput, EltsFromTbl, "vtbx");
8331 }
8332 case NEON::BI__builtin_neon_vtbx2_v: {
8333 return packTBLDVectorList(CGF, makeArrayRef(Ops).slice(1, 2), Ops[0],
8334 Ops[3], Ty, Intrinsic::aarch64_neon_tbx1,
8335 "vtbx1");
8336 }
8337 case NEON::BI__builtin_neon_vtbx3_v: {
8338 Value *TblRes =
8339 packTBLDVectorList(CGF, makeArrayRef(Ops).slice(1, 3), nullptr, Ops[4],
8340 Ty, Intrinsic::aarch64_neon_tbl2, "vtbl2");
8341
8342 llvm::Constant *TwentyFourV = ConstantInt::get(Ty, 24);
8343 Value *CmpRes = Builder.CreateICmp(ICmpInst::ICMP_UGE, Ops[4],
8344 TwentyFourV);
8345 CmpRes = Builder.CreateSExt(CmpRes, Ty);
8346
8347 Value *EltsFromInput = Builder.CreateAnd(CmpRes, Ops[0]);
8348 Value *EltsFromTbl = Builder.CreateAnd(Builder.CreateNot(CmpRes), TblRes);
8349 return Builder.CreateOr(EltsFromInput, EltsFromTbl, "vtbx");
8350 }
8351 case NEON::BI__builtin_neon_vtbx4_v: {
8352 return packTBLDVectorList(CGF, makeArrayRef(Ops).slice(1, 4), Ops[0],
8353 Ops[5], Ty, Intrinsic::aarch64_neon_tbx2,
8354 "vtbx2");
8355 }
8356 case NEON::BI__builtin_neon_vqtbl1_v:
8357 case NEON::BI__builtin_neon_vqtbl1q_v:
8358 Int = Intrinsic::aarch64_neon_tbl1; s = "vtbl1"; break;
8359 case NEON::BI__builtin_neon_vqtbl2_v:
8360 case NEON::BI__builtin_neon_vqtbl2q_v: {
8361 Int = Intrinsic::aarch64_neon_tbl2; s = "vtbl2"; break;
8362 case NEON::BI__builtin_neon_vqtbl3_v:
8363 case NEON::BI__builtin_neon_vqtbl3q_v:
8364 Int = Intrinsic::aarch64_neon_tbl3; s = "vtbl3"; break;
8365 case NEON::BI__builtin_neon_vqtbl4_v:
8366 case NEON::BI__builtin_neon_vqtbl4q_v:
8367 Int = Intrinsic::aarch64_neon_tbl4; s = "vtbl4"; break;
8368 case NEON::BI__builtin_neon_vqtbx1_v:
8369 case NEON::BI__builtin_neon_vqtbx1q_v:
8370 Int = Intrinsic::aarch64_neon_tbx1; s = "vtbx1"; break;
8371 case NEON::BI__builtin_neon_vqtbx2_v:
8372 case NEON::BI__builtin_neon_vqtbx2q_v:
8373 Int = Intrinsic::aarch64_neon_tbx2; s = "vtbx2"; break;
8374 case NEON::BI__builtin_neon_vqtbx3_v:
8375 case NEON::BI__builtin_neon_vqtbx3q_v:
8376 Int = Intrinsic::aarch64_neon_tbx3; s = "vtbx3"; break;
8377 case NEON::BI__builtin_neon_vqtbx4_v:
8378 case NEON::BI__builtin_neon_vqtbx4q_v:
8379 Int = Intrinsic::aarch64_neon_tbx4; s = "vtbx4"; break;
8380 }
8381 }
8382
8383 if (!Int)
8384 return nullptr;
8385
8386 Function *F = CGF.CGM.getIntrinsic(Int, Ty);
8387 return CGF.EmitNeonCall(F, Ops, s);
8388}
8389
8390Value *CodeGenFunction::vectorWrapScalar16(Value *Op) {
8391 auto *VTy = llvm::FixedVectorType::get(Int16Ty, 4);
8392 Op = Builder.CreateBitCast(Op, Int16Ty);
8393 Value *V = UndefValue::get(VTy);
8394 llvm::Constant *CI = ConstantInt::get(SizeTy, 0);
8395 Op = Builder.CreateInsertElement(V, Op, CI);
8396 return Op;
8397}
8398
8399/// SVEBuiltinMemEltTy - Returns the memory element type for this memory
8400/// access builtin. Only required if it can't be inferred from the base pointer
8401/// operand.
8402llvm::Type *CodeGenFunction::SVEBuiltinMemEltTy(SVETypeFlags TypeFlags) {
8403 switch (TypeFlags.getMemEltType()) {
8404 case SVETypeFlags::MemEltTyDefault:
8405 return getEltType(TypeFlags);
8406 case SVETypeFlags::MemEltTyInt8:
8407 return Builder.getInt8Ty();
8408 case SVETypeFlags::MemEltTyInt16:
8409 return Builder.getInt16Ty();
8410 case SVETypeFlags::MemEltTyInt32:
8411 return Builder.getInt32Ty();
8412 case SVETypeFlags::MemEltTyInt64:
8413 return Builder.getInt64Ty();
8414 }
8415 llvm_unreachable("Unknown MemEltType")__builtin_unreachable();
8416}
8417
8418llvm::Type *CodeGenFunction::getEltType(SVETypeFlags TypeFlags) {
8419 switch (TypeFlags.getEltType()) {
8420 default:
8421 llvm_unreachable("Invalid SVETypeFlag!")__builtin_unreachable();
8422
8423 case SVETypeFlags::EltTyInt8:
8424 return Builder.getInt8Ty();
8425 case SVETypeFlags::EltTyInt16:
8426 return Builder.getInt16Ty();
8427 case SVETypeFlags::EltTyInt32:
8428 return Builder.getInt32Ty();
8429 case SVETypeFlags::EltTyInt64:
8430 return Builder.getInt64Ty();
8431
8432 case SVETypeFlags::EltTyFloat16:
8433 return Builder.getHalfTy();
8434 case SVETypeFlags::EltTyFloat32:
8435 return Builder.getFloatTy();
8436 case SVETypeFlags::EltTyFloat64:
8437 return Builder.getDoubleTy();
8438
8439 case SVETypeFlags::EltTyBFloat16:
8440 return Builder.getBFloatTy();
8441
8442 case SVETypeFlags::EltTyBool8:
8443 case SVETypeFlags::EltTyBool16:
8444 case SVETypeFlags::EltTyBool32:
8445 case SVETypeFlags::EltTyBool64:
8446 return Builder.getInt1Ty();
8447 }
8448}
8449
8450// Return the llvm predicate vector type corresponding to the specified element
8451// TypeFlags.
8452llvm::ScalableVectorType *
8453CodeGenFunction::getSVEPredType(SVETypeFlags TypeFlags) {
8454 switch (TypeFlags.getEltType()) {
8455 default: llvm_unreachable("Unhandled SVETypeFlag!")__builtin_unreachable();
8456
8457 case SVETypeFlags::EltTyInt8:
8458 return llvm::ScalableVectorType::get(Builder.getInt1Ty(), 16);
8459 case SVETypeFlags::EltTyInt16:
8460 return llvm::ScalableVectorType::get(Builder.getInt1Ty(), 8);
8461 case SVETypeFlags::EltTyInt32:
8462 return llvm::ScalableVectorType::get(Builder.getInt1Ty(), 4);
8463 case SVETypeFlags::EltTyInt64:
8464 return llvm::ScalableVectorType::get(Builder.getInt1Ty(), 2);
8465
8466 case SVETypeFlags::EltTyBFloat16:
8467 return llvm::ScalableVectorType::get(Builder.getInt1Ty(), 8);
8468 case SVETypeFlags::EltTyFloat16:
8469 return llvm::ScalableVectorType::get(Builder.getInt1Ty(), 8);
8470 case SVETypeFlags::EltTyFloat32:
8471 return llvm::ScalableVectorType::get(Builder.getInt1Ty(), 4);
8472 case SVETypeFlags::EltTyFloat64:
8473 return llvm::ScalableVectorType::get(Builder.getInt1Ty(), 2);
8474
8475 case SVETypeFlags::EltTyBool8:
8476 return llvm::ScalableVectorType::get(Builder.getInt1Ty(), 16);
8477 case SVETypeFlags::EltTyBool16:
8478 return llvm::ScalableVectorType::get(Builder.getInt1Ty(), 8);
8479 case SVETypeFlags::EltTyBool32:
8480 return llvm::ScalableVectorType::get(Builder.getInt1Ty(), 4);
8481 case SVETypeFlags::EltTyBool64:
8482 return llvm::ScalableVectorType::get(Builder.getInt1Ty(), 2);
8483 }
8484}
8485
8486// Return the llvm vector type corresponding to the specified element TypeFlags.
8487llvm::ScalableVectorType *
8488CodeGenFunction::getSVEType(const SVETypeFlags &TypeFlags) {
8489 switch (TypeFlags.getEltType()) {
8490 default:
8491 llvm_unreachable("Invalid SVETypeFlag!")__builtin_unreachable();
8492
8493 case SVETypeFlags::EltTyInt8:
8494 return llvm::ScalableVectorType::get(Builder.getInt8Ty(), 16);
8495 case SVETypeFlags::EltTyInt16:
8496 return llvm::ScalableVectorType::get(Builder.getInt16Ty(), 8);
8497 case SVETypeFlags::EltTyInt32:
8498 return llvm::ScalableVectorType::get(Builder.getInt32Ty(), 4);
8499 case SVETypeFlags::EltTyInt64:
8500 return llvm::ScalableVectorType::get(Builder.getInt64Ty(), 2);
8501
8502 case SVETypeFlags::EltTyFloat16:
8503 return llvm::ScalableVectorType::get(Builder.getHalfTy(), 8);
8504 case SVETypeFlags::EltTyBFloat16:
8505 return llvm::ScalableVectorType::get(Builder.getBFloatTy(), 8);
8506 case SVETypeFlags::EltTyFloat32:
8507 return llvm::ScalableVectorType::get(Builder.getFloatTy(), 4);
8508 case SVETypeFlags::EltTyFloat64:
8509 return llvm::ScalableVectorType::get(Builder.getDoubleTy(), 2);
8510
8511 case SVETypeFlags::EltTyBool8:
8512 return llvm::ScalableVectorType::get(Builder.getInt1Ty(), 16);
8513 case SVETypeFlags::EltTyBool16:
8514 return llvm::ScalableVectorType::get(Builder.getInt1Ty(), 8);
8515 case SVETypeFlags::EltTyBool32:
8516 return llvm::ScalableVectorType::get(Builder.getInt1Ty(), 4);
8517 case SVETypeFlags::EltTyBool64:
8518 return llvm::ScalableVectorType::get(Builder.getInt1Ty(), 2);
8519 }
8520}
8521
8522llvm::Value *CodeGenFunction::EmitSVEAllTruePred(SVETypeFlags TypeFlags) {
8523 Function *Ptrue =
8524 CGM.getIntrinsic(Intrinsic::aarch64_sve_ptrue, getSVEPredType(TypeFlags));
8525 return Builder.CreateCall(Ptrue, {Builder.getInt32(/*SV_ALL*/ 31)});
8526}
8527
8528constexpr unsigned SVEBitsPerBlock = 128;
8529
8530static llvm::ScalableVectorType *getSVEVectorForElementType(llvm::Type *EltTy) {
8531 unsigned NumElts = SVEBitsPerBlock / EltTy->getScalarSizeInBits();
8532 return llvm::ScalableVectorType::get(EltTy, NumElts);
8533}
8534
8535// Reinterpret the input predicate so that it can be used to correctly isolate
8536// the elements of the specified datatype.
8537Value *CodeGenFunction::EmitSVEPredicateCast(Value *Pred,
8538 llvm::ScalableVectorType *VTy) {
8539 auto *RTy = llvm::VectorType::get(IntegerType::get(getLLVMContext(), 1), VTy);
8540 if (Pred->getType() == RTy)
8541 return Pred;
8542
8543 unsigned IntID;
8544 llvm::Type *IntrinsicTy;
8545 switch (VTy->getMinNumElements()) {
8546 default:
8547 llvm_unreachable("unsupported element count!")__builtin_unreachable();
8548 case 2:
8549 case 4:
8550 case 8:
8551 IntID = Intrinsic::aarch64_sve_convert_from_svbool;
8552 IntrinsicTy = RTy;
8553 break;
8554 case 16:
8555 IntID = Intrinsic::aarch64_sve_convert_to_svbool;
8556 IntrinsicTy = Pred->getType();
8557 break;
8558 }
8559
8560 Function *F = CGM.getIntrinsic(IntID, IntrinsicTy);
8561 Value *C = Builder.CreateCall(F, Pred);
8562 assert(C->getType() == RTy && "Unexpected return type!")((void)0);
8563 return C;
8564}
8565
8566Value *CodeGenFunction::EmitSVEGatherLoad(SVETypeFlags TypeFlags,
8567 SmallVectorImpl<Value *> &Ops,
8568 unsigned IntID) {
8569 auto *ResultTy = getSVEType(TypeFlags);
8570 auto *OverloadedTy =
8571 llvm::ScalableVectorType::get(SVEBuiltinMemEltTy(TypeFlags), ResultTy);
8572
8573 // At the ACLE level there's only one predicate type, svbool_t, which is
8574 // mapped to <n x 16 x i1>. However, this might be incompatible with the
8575 // actual type being loaded. For example, when loading doubles (i64) the
8576 // predicated should be <n x 2 x i1> instead. At the IR level the type of
8577 // the predicate and the data being loaded must match. Cast accordingly.
8578 Ops[0] = EmitSVEPredicateCast(Ops[0], OverloadedTy);
8579
8580 Function *F = nullptr;
8581 if (Ops[1]->getType()->isVectorTy())
8582 // This is the "vector base, scalar offset" case. In order to uniquely
8583 // map this built-in to an LLVM IR intrinsic, we need both the return type
8584 // and the type of the vector base.
8585 F = CGM.getIntrinsic(IntID, {OverloadedTy, Ops[1]->getType()});
8586 else
8587 // This is the "scalar base, vector offset case". The type of the offset
8588 // is encoded in the name of the intrinsic. We only need to specify the
8589 // return type in order to uniquely map this built-in to an LLVM IR
8590 // intrinsic.
8591 F = CGM.getIntrinsic(IntID, OverloadedTy);
8592
8593 // Pass 0 when the offset is missing. This can only be applied when using
8594 // the "vector base" addressing mode for which ACLE allows no offset. The
8595 // corresponding LLVM IR always requires an offset.
8596 if (Ops.size() == 2) {
8597 assert(Ops[1]->getType()->isVectorTy() && "Scalar base requires an offset")((void)0);
8598 Ops.push_back(ConstantInt::get(Int64Ty, 0));
8599 }
8600
8601 // For "vector base, scalar index" scale the index so that it becomes a
8602 // scalar offset.
8603 if (!TypeFlags.isByteIndexed() && Ops[1]->getType()->isVectorTy()) {
8604 unsigned BytesPerElt =
8605 OverloadedTy->getElementType()->getScalarSizeInBits() / 8;
8606 Value *Scale = ConstantInt::get(Int64Ty, BytesPerElt);
8607 Ops[2] = Builder.CreateMul(Ops[2], Scale);
8608 }
8609
8610 Value *Call = Builder.CreateCall(F, Ops);
8611
8612 // The following sext/zext is only needed when ResultTy != OverloadedTy. In
8613 // other cases it's folded into a nop.
8614 return TypeFlags.isZExtReturn() ? Builder.CreateZExt(Call, ResultTy)
8615 : Builder.CreateSExt(Call, ResultTy);
8616}
8617
8618Value *CodeGenFunction::EmitSVEScatterStore(SVETypeFlags TypeFlags,
8619 SmallVectorImpl<Value *> &Ops,
8620 unsigned IntID) {
8621 auto *SrcDataTy = getSVEType(TypeFlags);
8622 auto *OverloadedTy =
8623 llvm::ScalableVectorType::get(SVEBuiltinMemEltTy(TypeFlags), SrcDataTy);
8624
8625 // In ACLE the source data is passed in the last argument, whereas in LLVM IR
8626 // it's the first argument. Move it accordingly.
8627 Ops.insert(Ops.begin(), Ops.pop_back_val());
8628
8629 Function *F = nullptr;
8630 if (Ops[2]->getType()->isVectorTy())
8631 // This is the "vector base, scalar offset" case. In order to uniquely
8632 // map this built-in to an LLVM IR intrinsic, we need both the return type
8633 // and the type of the vector base.
8634 F = CGM.getIntrinsic(IntID, {OverloadedTy, Ops[2]->getType()});
8635 else
8636 // This is the "scalar base, vector offset case". The type of the offset
8637 // is encoded in the name of the intrinsic. We only need to specify the
8638 // return type in order to uniquely map this built-in to an LLVM IR
8639 // intrinsic.
8640 F = CGM.getIntrinsic(IntID, OverloadedTy);
8641
8642 // Pass 0 when the offset is missing. This can only be applied when using
8643 // the "vector base" addressing mode for which ACLE allows no offset. The
8644 // corresponding LLVM IR always requires an offset.
8645 if (Ops.size() == 3) {
8646 assert(Ops[1]->getType()->isVectorTy() && "Scalar base requires an offset")((void)0);
8647 Ops.push_back(ConstantInt::get(Int64Ty, 0));
8648 }
8649
8650 // Truncation is needed when SrcDataTy != OverloadedTy. In other cases it's
8651 // folded into a nop.
8652 Ops[0] = Builder.CreateTrunc(Ops[0], OverloadedTy);
8653
8654 // At the ACLE level there's only one predicate type, svbool_t, which is
8655 // mapped to <n x 16 x i1>. However, this might be incompatible with the
8656 // actual type being stored. For example, when storing doubles (i64) the
8657 // predicated should be <n x 2 x i1> instead. At the IR level the type of
8658 // the predicate and the data being stored must match. Cast accordingly.
8659 Ops[1] = EmitSVEPredicateCast(Ops[1], OverloadedTy);
8660
8661 // For "vector base, scalar index" scale the index so that it becomes a
8662 // scalar offset.
8663 if (!TypeFlags.isByteIndexed() && Ops[2]->getType()->isVectorTy()) {
8664 unsigned BytesPerElt =
8665 OverloadedTy->getElementType()->getScalarSizeInBits() / 8;
8666 Value *Scale = ConstantInt::get(Int64Ty, BytesPerElt);
8667 Ops[3] = Builder.CreateMul(Ops[3], Scale);
8668 }
8669
8670 return Builder.CreateCall(F, Ops);
8671}
8672
8673Value *CodeGenFunction::EmitSVEGatherPrefetch(SVETypeFlags TypeFlags,
8674 SmallVectorImpl<Value *> &Ops,
8675 unsigned IntID) {
8676 // The gather prefetches are overloaded on the vector input - this can either
8677 // be the vector of base addresses or vector of offsets.
8678 auto *OverloadedTy = dyn_cast<llvm::ScalableVectorType>(Ops[1]->getType());
8679 if (!OverloadedTy)
8680 OverloadedTy = cast<llvm::ScalableVectorType>(Ops[2]->getType());
8681
8682 // Cast the predicate from svbool_t to the right number of elements.
8683 Ops[0] = EmitSVEPredicateCast(Ops[0], OverloadedTy);
8684
8685 // vector + imm addressing modes
8686 if (Ops[1]->getType()->isVectorTy()) {
8687 if (Ops.size() == 3) {
8688 // Pass 0 for 'vector+imm' when the index is omitted.
8689 Ops.push_back(ConstantInt::get(Int64Ty, 0));
8690
8691 // The sv_prfop is the last operand in the builtin and IR intrinsic.
8692 std::swap(Ops[2], Ops[3]);
8693 } else {
8694 // Index needs to be passed as scaled offset.
8695 llvm::Type *MemEltTy = SVEBuiltinMemEltTy(TypeFlags);
8696 unsigned BytesPerElt = MemEltTy->getPrimitiveSizeInBits() / 8;
8697 Value *Scale = ConstantInt::get(Int64Ty, BytesPerElt);
8698 Ops[2] = Builder.CreateMul(Ops[2], Scale);
8699 }
8700 }
8701
8702 Function *F = CGM.getIntrinsic(IntID, OverloadedTy);
8703 return Builder.CreateCall(F, Ops);
8704}
8705
8706Value *CodeGenFunction::EmitSVEStructLoad(SVETypeFlags TypeFlags,
8707 SmallVectorImpl<Value*> &Ops,
8708 unsigned IntID) {
8709 llvm::ScalableVectorType *VTy = getSVEType(TypeFlags);
8710 auto VecPtrTy = llvm::PointerType::getUnqual(VTy);
8711 auto EltPtrTy = llvm::PointerType::getUnqual(VTy->getElementType());
8712
8713 unsigned N;
8714 switch (IntID) {
8715 case Intrinsic::aarch64_sve_ld2:
8716 N = 2;
8717 break;
8718 case Intrinsic::aarch64_sve_ld3:
8719 N = 3;
8720 break;
8721 case Intrinsic::aarch64_sve_ld4:
8722 N = 4;
8723 break;
8724 default:
8725 llvm_unreachable("unknown intrinsic!")__builtin_unreachable();
8726 }
8727 auto RetTy = llvm::VectorType::get(VTy->getElementType(),
8728 VTy->getElementCount() * N);
8729
8730 Value *Predicate = EmitSVEPredicateCast(Ops[0], VTy);
8731 Value *BasePtr= Builder.CreateBitCast(Ops[1], VecPtrTy);
8732 Value *Offset = Ops.size() > 2 ? Ops[2] : Builder.getInt32(0);
8733 BasePtr = Builder.CreateGEP(VTy, BasePtr, Offset);
8734 BasePtr = Builder.CreateBitCast(BasePtr, EltPtrTy);
8735
8736 Function *F = CGM.getIntrinsic(IntID, {RetTy, Predicate->getType()});
8737 return Builder.CreateCall(F, { Predicate, BasePtr });
8738}
8739
8740Value *CodeGenFunction::EmitSVEStructStore(SVETypeFlags TypeFlags,
8741 SmallVectorImpl<Value*> &Ops,
8742 unsigned IntID) {
8743 llvm::ScalableVectorType *VTy = getSVEType(TypeFlags);
8744 auto VecPtrTy = llvm::PointerType::getUnqual(VTy);
8745 auto EltPtrTy = llvm::PointerType::getUnqual(VTy->getElementType());
8746
8747 unsigned N;
8748 switch (IntID) {
8749 case Intrinsic::aarch64_sve_st2:
8750 N = 2;
8751 break;
8752 case Intrinsic::aarch64_sve_st3:
8753 N = 3;
8754 break;
8755 case Intrinsic::aarch64_sve_st4:
8756 N = 4;
8757 break;
8758 default:
8759 llvm_unreachable("unknown intrinsic!")__builtin_unreachable();
8760 }
8761 auto TupleTy =
8762 llvm::VectorType::get(VTy->getElementType(), VTy->getElementCount() * N);
8763
8764 Value *Predicate = EmitSVEPredicateCast(Ops[0], VTy);
8765 Value *BasePtr = Builder.CreateBitCast(Ops[1], VecPtrTy);
8766 Value *Offset = Ops.size() > 3 ? Ops[2] : Builder.getInt32(0);
8767 Value *Val = Ops.back();
8768 BasePtr = Builder.CreateGEP(VTy, BasePtr, Offset);
8769 BasePtr = Builder.CreateBitCast(BasePtr, EltPtrTy);
8770
8771 // The llvm.aarch64.sve.st2/3/4 intrinsics take legal part vectors, so we
8772 // need to break up the tuple vector.
8773 SmallVector<llvm::Value*, 5> Operands;
8774 Function *FExtr =
8775 CGM.getIntrinsic(Intrinsic::aarch64_sve_tuple_get, {VTy, TupleTy});
8776 for (unsigned I = 0; I < N; ++I)
8777 Operands.push_back(Builder.CreateCall(FExtr, {Val, Builder.getInt32(I)}));
8778 Operands.append({Predicate, BasePtr});
8779
8780 Function *F = CGM.getIntrinsic(IntID, { VTy });
8781 return Builder.CreateCall(F, Operands);
8782}
8783
8784// SVE2's svpmullb and svpmullt builtins are similar to the svpmullb_pair and
8785// svpmullt_pair intrinsics, with the exception that their results are bitcast
8786// to a wider type.
8787Value *CodeGenFunction::EmitSVEPMull(SVETypeFlags TypeFlags,
8788 SmallVectorImpl<Value *> &Ops,
8789 unsigned BuiltinID) {
8790 // Splat scalar operand to vector (intrinsics with _n infix)
8791 if (TypeFlags.hasSplatOperand()) {
8792 unsigned OpNo = TypeFlags.getSplatOperand();
8793 Ops[OpNo] = EmitSVEDupX(Ops[OpNo]);
8794 }
8795
8796 // The pair-wise function has a narrower overloaded type.
8797 Function *F = CGM.getIntrinsic(BuiltinID, Ops[0]->getType());
8798 Value *Call = Builder.CreateCall(F, {Ops[0], Ops[1]});
8799
8800 // Now bitcast to the wider result type.
8801 llvm::ScalableVectorType *Ty = getSVEType(TypeFlags);
8802 return EmitSVEReinterpret(Call, Ty);
8803}
8804
8805Value *CodeGenFunction::EmitSVEMovl(SVETypeFlags TypeFlags,
8806 ArrayRef<Value *> Ops, unsigned BuiltinID) {
8807 llvm::Type *OverloadedTy = getSVEType(TypeFlags);
8808 Function *F = CGM.getIntrinsic(BuiltinID, OverloadedTy);
8809 return Builder.CreateCall(F, {Ops[0], Builder.getInt32(0)});
8810}
8811
8812Value *CodeGenFunction::EmitSVEPrefetchLoad(SVETypeFlags TypeFlags,
8813 SmallVectorImpl<Value *> &Ops,
8814 unsigned BuiltinID) {
8815 auto *MemEltTy = SVEBuiltinMemEltTy(TypeFlags);
8816 auto *VectorTy = getSVEVectorForElementType(MemEltTy);
8817 auto *MemoryTy = llvm::ScalableVectorType::get(MemEltTy, VectorTy);
8818
8819 Value *Predicate = EmitSVEPredicateCast(Ops[0], MemoryTy);
8820 Value *BasePtr = Ops[1];
8821
8822 // Implement the index operand if not omitted.
8823 if (Ops.size() > 3) {
8824 BasePtr = Builder.CreateBitCast(BasePtr, MemoryTy->getPointerTo());
8825 BasePtr = Builder.CreateGEP(MemoryTy, BasePtr, Ops[2]);
8826 }
8827
8828 // Prefetch intriniscs always expect an i8*
8829 BasePtr = Builder.CreateBitCast(BasePtr, llvm::PointerType::getUnqual(Int8Ty));
8830 Value *PrfOp = Ops.back();
8831
8832 Function *F = CGM.getIntrinsic(BuiltinID, Predicate->getType());
8833 return Builder.CreateCall(F, {Predicate, BasePtr, PrfOp});
8834}
8835
8836Value *CodeGenFunction::EmitSVEMaskedLoad(const CallExpr *E,
8837 llvm::Type *ReturnTy,
8838 SmallVectorImpl<Value *> &Ops,
8839 unsigned BuiltinID,
8840 bool IsZExtReturn) {
8841 QualType LangPTy = E->getArg(1)->getType();
8842 llvm::Type *MemEltTy = CGM.getTypes().ConvertType(
8843 LangPTy->castAs<PointerType>()->getPointeeType());
8844
8845 // The vector type that is returned may be different from the
8846 // eventual type loaded from memory.
8847 auto VectorTy = cast<llvm::ScalableVectorType>(ReturnTy);
8848 auto MemoryTy = llvm::ScalableVectorType::get(MemEltTy, VectorTy);
8849
8850 Value *Predicate = EmitSVEPredicateCast(Ops[0], MemoryTy);
8851 Value *BasePtr = Builder.CreateBitCast(Ops[1], MemoryTy->getPointerTo());
8852 Value *Offset = Ops.size() > 2 ? Ops[2] : Builder.getInt32(0);
8853 BasePtr = Builder.CreateGEP(MemoryTy, BasePtr, Offset);
8854
8855 BasePtr = Builder.CreateBitCast(BasePtr, MemEltTy->getPointerTo());
8856 Function *F = CGM.getIntrinsic(BuiltinID, MemoryTy);
8857 Value *Load = Builder.CreateCall(F, {Predicate, BasePtr});
8858
8859 return IsZExtReturn ? Builder.CreateZExt(Load, VectorTy)
8860 : Builder.CreateSExt(Load, VectorTy);
8861}
8862
8863Value *CodeGenFunction::EmitSVEMaskedStore(const CallExpr *E,
8864 SmallVectorImpl<Value *> &Ops,
8865 unsigned BuiltinID) {
8866 QualType LangPTy = E->getArg(1)->getType();
8867 llvm::Type *MemEltTy = CGM.getTypes().ConvertType(
8868 LangPTy->castAs<PointerType>()->getPointeeType());
8869
8870 // The vector type that is stored may be different from the
8871 // eventual type stored to memory.
8872 auto VectorTy = cast<llvm::ScalableVectorType>(Ops.back()->getType());
8873 auto MemoryTy = llvm::ScalableVectorType::get(MemEltTy, VectorTy);
8874
8875 Value *Predicate = EmitSVEPredicateCast(Ops[0], MemoryTy);
8876 Value *BasePtr = Builder.CreateBitCast(Ops[1], MemoryTy->getPointerTo());
8877 Value *Offset = Ops.size() == 4 ? Ops[2] : Builder.getInt32(0);
8878 BasePtr = Builder.CreateGEP(MemoryTy, BasePtr, Offset);
8879
8880 // Last value is always the data
8881 llvm::Value *Val = Builder.CreateTrunc(Ops.back(), MemoryTy);
8882
8883 BasePtr = Builder.CreateBitCast(BasePtr, MemEltTy->getPointerTo());
8884 Function *F = CGM.getIntrinsic(BuiltinID, MemoryTy);
8885 return Builder.CreateCall(F, {Val, Predicate, BasePtr});
8886}
8887
8888// Limit the usage of scalable llvm IR generated by the ACLE by using the
8889// sve dup.x intrinsic instead of IRBuilder::CreateVectorSplat.
8890Value *CodeGenFunction::EmitSVEDupX(Value *Scalar, llvm::Type *Ty) {
8891 auto F = CGM.getIntrinsic(Intrinsic::aarch64_sve_dup_x, Ty);
8892 return Builder.CreateCall(F, Scalar);
8893}
8894
8895Value *CodeGenFunction::EmitSVEDupX(Value* Scalar) {
8896 return EmitSVEDupX(Scalar, getSVEVectorForElementType(Scalar->getType()));
8897}
8898
8899Value *CodeGenFunction::EmitSVEReinterpret(Value *Val, llvm::Type *Ty) {
8900 // FIXME: For big endian this needs an additional REV, or needs a separate
8901 // intrinsic that is code-generated as a no-op, because the LLVM bitcast
8902 // instruction is defined as 'bitwise' equivalent from memory point of
8903 // view (when storing/reloading), whereas the svreinterpret builtin
8904 // implements bitwise equivalent cast from register point of view.
8905 // LLVM CodeGen for a bitcast must add an explicit REV for big-endian.
8906 return Builder.CreateBitCast(Val, Ty);
8907}
8908
8909static void InsertExplicitZeroOperand(CGBuilderTy &Builder, llvm::Type *Ty,
8910 SmallVectorImpl<Value *> &Ops) {
8911 auto *SplatZero = Constant::getNullValue(Ty);
8912 Ops.insert(Ops.begin(), SplatZero);
8913}
8914
8915static void InsertExplicitUndefOperand(CGBuilderTy &Builder, llvm::Type *Ty,
8916 SmallVectorImpl<Value *> &Ops) {
8917 auto *SplatUndef = UndefValue::get(Ty);
8918 Ops.insert(Ops.begin(), SplatUndef);
8919}
8920
8921SmallVector<llvm::Type *, 2> CodeGenFunction::getSVEOverloadTypes(
8922 SVETypeFlags TypeFlags, llvm::Type *ResultType, ArrayRef<Value *> Ops) {
8923 if (TypeFlags.isOverloadNone())
8924 return {};
8925
8926 llvm::Type *DefaultType = getSVEType(TypeFlags);
8927
8928 if (TypeFlags.isOverloadWhile())
8929 return {DefaultType, Ops[1]->getType()};
8930
8931 if (TypeFlags.isOverloadWhileRW())
8932 return {getSVEPredType(TypeFlags), Ops[0]->getType()};
8933
8934 if (TypeFlags.isOverloadCvt() || TypeFlags.isTupleSet())
8935 return {Ops[0]->getType(), Ops.back()->getType()};
8936
8937 if (TypeFlags.isTupleCreate() || TypeFlags.isTupleGet())
8938 return {ResultType, Ops[0]->getType()};
8939
8940 assert(TypeFlags.isOverloadDefault() && "Unexpected value for overloads")((void)0);
8941 return {DefaultType};
8942}
8943
8944Value *CodeGenFunction::EmitAArch64SVEBuiltinExpr(unsigned BuiltinID,
8945 const CallExpr *E) {
8946 // Find out if any arguments are required to be integer constant expressions.
8947 unsigned ICEArguments = 0;
8948 ASTContext::GetBuiltinTypeError Error;
8949 getContext().GetBuiltinType(BuiltinID, Error, &ICEArguments);
8950 assert(Error == ASTContext::GE_None && "Should not codegen an error")((void)0);
8951
8952 llvm::Type *Ty = ConvertType(E->getType());
8953 if (BuiltinID >= SVE::BI__builtin_sve_reinterpret_s8_s8 &&
8954 BuiltinID <= SVE::BI__builtin_sve_reinterpret_f64_f64) {
8955 Value *Val = EmitScalarExpr(E->getArg(0));
8956 return EmitSVEReinterpret(Val, Ty);
8957 }
8958
8959 llvm::SmallVector<Value *, 4> Ops;
8960 for (unsigned i = 0, e = E->getNumArgs(); i != e; i++) {
8961 if ((ICEArguments & (1 << i)) == 0)
8962 Ops.push_back(EmitScalarExpr(E->getArg(i)));
8963 else {
8964 // If this is required to be a constant, constant fold it so that we know
8965 // that the generated intrinsic gets a ConstantInt.
8966 Optional<llvm::APSInt> Result =
8967 E->getArg(i)->getIntegerConstantExpr(getContext());
8968 assert(Result && "Expected argument to be a constant")((void)0);
8969
8970 // Immediates for SVE llvm intrinsics are always 32bit. We can safely
8971 // truncate because the immediate has been range checked and no valid
8972 // immediate requires more than a handful of bits.
8973 *Result = Result->extOrTrunc(32);
8974 Ops.push_back(llvm::ConstantInt::get(getLLVMContext(), *Result));
8975 }
8976 }
8977
8978 auto *Builtin = findARMVectorIntrinsicInMap(AArch64SVEIntrinsicMap, BuiltinID,
8979 AArch64SVEIntrinsicsProvenSorted);
8980 SVETypeFlags TypeFlags(Builtin->TypeModifier);
8981 if (TypeFlags.isLoad())
8982 return EmitSVEMaskedLoad(E, Ty, Ops, Builtin->LLVMIntrinsic,
8983 TypeFlags.isZExtReturn());
8984 else if (TypeFlags.isStore())
8985 return EmitSVEMaskedStore(E, Ops, Builtin->LLVMIntrinsic);
8986 else if (TypeFlags.isGatherLoad())
8987 return EmitSVEGatherLoad(TypeFlags, Ops, Builtin->LLVMIntrinsic);
8988 else if (TypeFlags.isScatterStore())
8989 return EmitSVEScatterStore(TypeFlags, Ops, Builtin->LLVMIntrinsic);
8990 else if (TypeFlags.isPrefetch())
8991 return EmitSVEPrefetchLoad(TypeFlags, Ops, Builtin->LLVMIntrinsic);
8992 else if (TypeFlags.isGatherPrefetch())
8993 return EmitSVEGatherPrefetch(TypeFlags, Ops, Builtin->LLVMIntrinsic);
8994 else if (TypeFlags.isStructLoad())
8995 return EmitSVEStructLoad(TypeFlags, Ops, Builtin->LLVMIntrinsic);
8996 else if (TypeFlags.isStructStore())
8997 return EmitSVEStructStore(TypeFlags, Ops, Builtin->LLVMIntrinsic);
8998 else if (TypeFlags.isUndef())
8999 return UndefValue::get(Ty);
9000 else if (Builtin->LLVMIntrinsic != 0) {
9001 if (TypeFlags.getMergeType() == SVETypeFlags::MergeZeroExp)
9002 InsertExplicitZeroOperand(Builder, Ty, Ops);
9003
9004 if (TypeFlags.getMergeType() == SVETypeFlags::MergeAnyExp)
9005 InsertExplicitUndefOperand(Builder, Ty, Ops);
9006
9007 // Some ACLE builtins leave out the argument to specify the predicate
9008 // pattern, which is expected to be expanded to an SV_ALL pattern.
9009 if (TypeFlags.isAppendSVALL())
9010 Ops.push_back(Builder.getInt32(/*SV_ALL*/ 31));
9011 if (TypeFlags.isInsertOp1SVALL())
9012 Ops.insert(&Ops[1], Builder.getInt32(/*SV_ALL*/ 31));
9013
9014 // Predicates must match the main datatype.
9015 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
9016 if (auto PredTy = dyn_cast<llvm::VectorType>(Ops[i]->getType()))
9017 if (PredTy->getElementType()->isIntegerTy(1))
9018 Ops[i] = EmitSVEPredicateCast(Ops[i], getSVEType(TypeFlags));
9019
9020 // Splat scalar operand to vector (intrinsics with _n infix)
9021 if (TypeFlags.hasSplatOperand()) {
9022 unsigned OpNo = TypeFlags.getSplatOperand();
9023 Ops[OpNo] = EmitSVEDupX(Ops[OpNo]);
9024 }
9025
9026 if (TypeFlags.isReverseCompare())
9027 std::swap(Ops[1], Ops[2]);
9028
9029 if (TypeFlags.isReverseUSDOT())
9030 std::swap(Ops[1], Ops[2]);
9031
9032 // Predicated intrinsics with _z suffix need a select w/ zeroinitializer.
9033 if (TypeFlags.getMergeType() == SVETypeFlags::MergeZero) {
9034 llvm::Type *OpndTy = Ops[1]->getType();
9035 auto *SplatZero = Constant::getNullValue(OpndTy);
9036 Function *Sel = CGM.getIntrinsic(Intrinsic::aarch64_sve_sel, OpndTy);
9037 Ops[1] = Builder.CreateCall(Sel, {Ops[0], Ops[1], SplatZero});
9038 }
9039
9040 Function *F = CGM.getIntrinsic(Builtin->LLVMIntrinsic,
9041 getSVEOverloadTypes(TypeFlags, Ty, Ops));
9042 Value *Call = Builder.CreateCall(F, Ops);
9043
9044 // Predicate results must be converted to svbool_t.
9045 if (auto PredTy = dyn_cast<llvm::VectorType>(Call->getType()))
9046 if (PredTy->getScalarType()->isIntegerTy(1))
9047 Call = EmitSVEPredicateCast(Call, cast<llvm::ScalableVectorType>(Ty));
9048
9049 return Call;
9050 }
9051
9052 switch (BuiltinID) {
9053 default:
9054 return nullptr;
9055
9056 case SVE::BI__builtin_sve_svmov_b_z: {
9057 // svmov_b_z(pg, op) <=> svand_b_z(pg, op, op)
9058 SVETypeFlags TypeFlags(Builtin->TypeModifier);
9059 llvm::Type* OverloadedTy = getSVEType(TypeFlags);
9060 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_sve_and_z, OverloadedTy);
9061 return Builder.CreateCall(F, {Ops[0], Ops[1], Ops[1]});
9062 }
9063
9064 case SVE::BI__builtin_sve_svnot_b_z: {
9065 // svnot_b_z(pg, op) <=> sveor_b_z(pg, op, pg)
9066 SVETypeFlags TypeFlags(Builtin->TypeModifier);
9067 llvm::Type* OverloadedTy = getSVEType(TypeFlags);
9068 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_sve_eor_z, OverloadedTy);
9069 return Builder.CreateCall(F, {Ops[0], Ops[1], Ops[0]});
9070 }
9071
9072 case SVE::BI__builtin_sve_svmovlb_u16:
9073 case SVE::BI__builtin_sve_svmovlb_u32:
9074 case SVE::BI__builtin_sve_svmovlb_u64:
9075 return EmitSVEMovl(TypeFlags, Ops, Intrinsic::aarch64_sve_ushllb);
9076
9077 case SVE::BI__builtin_sve_svmovlb_s16:
9078 case SVE::BI__builtin_sve_svmovlb_s32:
9079 case SVE::BI__builtin_sve_svmovlb_s64:
9080 return EmitSVEMovl(TypeFlags, Ops, Intrinsic::aarch64_sve_sshllb);
9081
9082 case SVE::BI__builtin_sve_svmovlt_u16:
9083 case SVE::BI__builtin_sve_svmovlt_u32:
9084 case SVE::BI__builtin_sve_svmovlt_u64:
9085 return EmitSVEMovl(TypeFlags, Ops, Intrinsic::aarch64_sve_ushllt);
9086
9087 case SVE::BI__builtin_sve_svmovlt_s16:
9088 case SVE::BI__builtin_sve_svmovlt_s32:
9089 case SVE::BI__builtin_sve_svmovlt_s64:
9090 return EmitSVEMovl(TypeFlags, Ops, Intrinsic::aarch64_sve_sshllt);
9091
9092 case SVE::BI__builtin_sve_svpmullt_u16:
9093 case SVE::BI__builtin_sve_svpmullt_u64:
9094 case SVE::BI__builtin_sve_svpmullt_n_u16:
9095 case SVE::BI__builtin_sve_svpmullt_n_u64:
9096 return EmitSVEPMull(TypeFlags, Ops, Intrinsic::aarch64_sve_pmullt_pair);
9097
9098 case SVE::BI__builtin_sve_svpmullb_u16:
9099 case SVE::BI__builtin_sve_svpmullb_u64:
9100 case SVE::BI__builtin_sve_svpmullb_n_u16:
9101 case SVE::BI__builtin_sve_svpmullb_n_u64:
9102 return EmitSVEPMull(TypeFlags, Ops, Intrinsic::aarch64_sve_pmullb_pair);
9103
9104 case SVE::BI__builtin_sve_svdup_n_b8:
9105 case SVE::BI__builtin_sve_svdup_n_b16:
9106 case SVE::BI__builtin_sve_svdup_n_b32:
9107 case SVE::BI__builtin_sve_svdup_n_b64: {
9108 Value *CmpNE =
9109 Builder.CreateICmpNE(Ops[0], Constant::getNullValue(Ops[0]->getType()));
9110 llvm::ScalableVectorType *OverloadedTy = getSVEType(TypeFlags);
9111 Value *Dup = EmitSVEDupX(CmpNE, OverloadedTy);
9112 return EmitSVEPredicateCast(Dup, cast<llvm::ScalableVectorType>(Ty));
9113 }
9114
9115 case SVE::BI__builtin_sve_svdupq_n_b8:
9116 case SVE::BI__builtin_sve_svdupq_n_b16:
9117 case SVE::BI__builtin_sve_svdupq_n_b32:
9118 case SVE::BI__builtin_sve_svdupq_n_b64:
9119 case SVE::BI__builtin_sve_svdupq_n_u8:
9120 case SVE::BI__builtin_sve_svdupq_n_s8:
9121 case SVE::BI__builtin_sve_svdupq_n_u64:
9122 case SVE::BI__builtin_sve_svdupq_n_f64:
9123 case SVE::BI__builtin_sve_svdupq_n_s64:
9124 case SVE::BI__builtin_sve_svdupq_n_u16:
9125 case SVE::BI__builtin_sve_svdupq_n_f16:
9126 case SVE::BI__builtin_sve_svdupq_n_bf16:
9127 case SVE::BI__builtin_sve_svdupq_n_s16:
9128 case SVE::BI__builtin_sve_svdupq_n_u32:
9129 case SVE::BI__builtin_sve_svdupq_n_f32:
9130 case SVE::BI__builtin_sve_svdupq_n_s32: {
9131 // These builtins are implemented by storing each element to an array and using
9132 // ld1rq to materialize a vector.
9133 unsigned NumOpnds = Ops.size();
9134
9135 bool IsBoolTy =
9136 cast<llvm::VectorType>(Ty)->getElementType()->isIntegerTy(1);
9137
9138 // For svdupq_n_b* the element type of is an integer of type 128/numelts,
9139 // so that the compare can use the width that is natural for the expected
9140 // number of predicate lanes.
9141 llvm::Type *EltTy = Ops[0]->getType();
9142 if (IsBoolTy)
9143 EltTy = IntegerType::get(getLLVMContext(), SVEBitsPerBlock / NumOpnds);
9144
9145 SmallVector<llvm::Value *, 16> VecOps;
9146 for (unsigned I = 0; I < NumOpnds; ++I)
9147 VecOps.push_back(Builder.CreateZExt(Ops[I], EltTy));
9148 Value *Vec = BuildVector(VecOps);
9149
9150 SVETypeFlags TypeFlags(Builtin->TypeModifier);
9151 Value *Pred = EmitSVEAllTruePred(TypeFlags);
9152
9153 llvm::Type *OverloadedTy = getSVEVectorForElementType(EltTy);
9154 Value *InsertSubVec = Builder.CreateInsertVector(
9155 OverloadedTy, UndefValue::get(OverloadedTy), Vec, Builder.getInt64(0));
9156
9157 Function *F =
9158 CGM.getIntrinsic(Intrinsic::aarch64_sve_dupq_lane, OverloadedTy);
9159 Value *DupQLane =
9160 Builder.CreateCall(F, {InsertSubVec, Builder.getInt64(0)});
9161
9162 if (!IsBoolTy)
9163 return DupQLane;
9164
9165 // For svdupq_n_b* we need to add an additional 'cmpne' with '0'.
9166 F = CGM.getIntrinsic(NumOpnds == 2 ? Intrinsic::aarch64_sve_cmpne
9167 : Intrinsic::aarch64_sve_cmpne_wide,
9168 OverloadedTy);
9169 Value *Call = Builder.CreateCall(
9170 F, {Pred, DupQLane, EmitSVEDupX(Builder.getInt64(0))});
9171 return EmitSVEPredicateCast(Call, cast<llvm::ScalableVectorType>(Ty));
9172 }
9173
9174 case SVE::BI__builtin_sve_svpfalse_b:
9175 return ConstantInt::getFalse(Ty);
9176
9177 case SVE::BI__builtin_sve_svlen_bf16:
9178 case SVE::BI__builtin_sve_svlen_f16:
9179 case SVE::BI__builtin_sve_svlen_f32:
9180 case SVE::BI__builtin_sve_svlen_f64:
9181 case SVE::BI__builtin_sve_svlen_s8:
9182 case SVE::BI__builtin_sve_svlen_s16:
9183 case SVE::BI__builtin_sve_svlen_s32:
9184 case SVE::BI__builtin_sve_svlen_s64:
9185 case SVE::BI__builtin_sve_svlen_u8:
9186 case SVE::BI__builtin_sve_svlen_u16:
9187 case SVE::BI__builtin_sve_svlen_u32:
9188 case SVE::BI__builtin_sve_svlen_u64: {
9189 SVETypeFlags TF(Builtin->TypeModifier);
9190 auto VTy = cast<llvm::VectorType>(getSVEType(TF));
9191 auto *NumEls =
9192 llvm::ConstantInt::get(Ty, VTy->getElementCount().getKnownMinValue());
9193
9194 Function *F = CGM.getIntrinsic(Intrinsic::vscale, Ty);
9195 return Builder.CreateMul(NumEls, Builder.CreateCall(F));
9196 }
9197
9198 case SVE::BI__builtin_sve_svtbl2_u8:
9199 case SVE::BI__builtin_sve_svtbl2_s8:
9200 case SVE::BI__builtin_sve_svtbl2_u16:
9201 case SVE::BI__builtin_sve_svtbl2_s16:
9202 case SVE::BI__builtin_sve_svtbl2_u32:
9203 case SVE::BI__builtin_sve_svtbl2_s32:
9204 case SVE::BI__builtin_sve_svtbl2_u64:
9205 case SVE::BI__builtin_sve_svtbl2_s64:
9206 case SVE::BI__builtin_sve_svtbl2_f16:
9207 case SVE::BI__builtin_sve_svtbl2_bf16:
9208 case SVE::BI__builtin_sve_svtbl2_f32:
9209 case SVE::BI__builtin_sve_svtbl2_f64: {
9210 SVETypeFlags TF(Builtin->TypeModifier);
9211 auto VTy = cast<llvm::VectorType>(getSVEType(TF));
9212 auto TupleTy = llvm::VectorType::getDoubleElementsVectorType(VTy);
9213 Function *FExtr =
9214 CGM.getIntrinsic(Intrinsic::aarch64_sve_tuple_get, {VTy, TupleTy});
9215 Value *V0 = Builder.CreateCall(FExtr, {Ops[0], Builder.getInt32(0)});
9216 Value *V1 = Builder.CreateCall(FExtr, {Ops[0], Builder.getInt32(1)});
9217 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_sve_tbl2, VTy);
9218 return Builder.CreateCall(F, {V0, V1, Ops[1]});
9219 }
9220 }
9221
9222 /// Should not happen
9223 return nullptr;
9224}
9225
9226Value *CodeGenFunction::EmitAArch64BuiltinExpr(unsigned BuiltinID,
9227 const CallExpr *E,
9228 llvm::Triple::ArchType Arch) {
9229 if (BuiltinID >= AArch64::FirstSVEBuiltin &&
9230 BuiltinID <= AArch64::LastSVEBuiltin)
9231 return EmitAArch64SVEBuiltinExpr(BuiltinID, E);
9232
9233 unsigned HintID = static_cast<unsigned>(-1);
9234 switch (BuiltinID) {
9235 default: break;
9236 case AArch64::BI__builtin_arm_nop:
9237 HintID = 0;
9238 break;
9239 case AArch64::BI__builtin_arm_yield:
9240 case AArch64::BI__yield:
9241 HintID = 1;
9242 break;
9243 case AArch64::BI__builtin_arm_wfe:
9244 case AArch64::BI__wfe:
9245 HintID = 2;
9246 break;
9247 case AArch64::BI__builtin_arm_wfi:
9248 case AArch64::BI__wfi:
9249 HintID = 3;
9250 break;
9251 case AArch64::BI__builtin_arm_sev:
9252 case AArch64::BI__sev:
9253 HintID = 4;
9254 break;
9255 case AArch64::BI__builtin_arm_sevl:
9256 case AArch64::BI__sevl:
9257 HintID = 5;
9258 break;
9259 }
9260
9261 if (HintID != static_cast<unsigned>(-1)) {
9262 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_hint);
9263 return Builder.CreateCall(F, llvm::ConstantInt::get(Int32Ty, HintID));
9264 }
9265
9266 if (BuiltinID == AArch64::BI__builtin_arm_prefetch) {
9267 Value *Address = EmitScalarExpr(E->getArg(0));
9268 Value *RW = EmitScalarExpr(E->getArg(1));
9269 Value *CacheLevel = EmitScalarExpr(E->getArg(2));
9270 Value *RetentionPolicy = EmitScalarExpr(E->getArg(3));
9271 Value *IsData = EmitScalarExpr(E->getArg(4));
9272
9273 Value *Locality = nullptr;
9274 if (cast<llvm::ConstantInt>(RetentionPolicy)->isZero()) {
9275 // Temporal fetch, needs to convert cache level to locality.
9276 Locality = llvm::ConstantInt::get(Int32Ty,
9277 -cast<llvm::ConstantInt>(CacheLevel)->getValue() + 3);
9278 } else {
9279 // Streaming fetch.
9280 Locality = llvm::ConstantInt::get(Int32Ty, 0);
9281 }
9282
9283 // FIXME: We need AArch64 specific LLVM intrinsic if we want to specify
9284 // PLDL3STRM or PLDL2STRM.
9285 Function *F = CGM.getIntrinsic(Intrinsic::prefetch, Address->getType());
9286 return Builder.CreateCall(F, {Address, RW, Locality, IsData});
9287 }
9288
9289 if (BuiltinID == AArch64::BI__builtin_arm_rbit) {
9290 assert((getContext().getTypeSize(E->getType()) == 32) &&((void)0)
9291 "rbit of unusual size!")((void)0);
9292 llvm::Value *Arg = EmitScalarExpr(E->getArg(0));
9293 return Builder.CreateCall(
9294 CGM.getIntrinsic(Intrinsic::bitreverse, Arg->getType()), Arg, "rbit");
9295 }
9296 if (BuiltinID == AArch64::BI__builtin_arm_rbit64) {
9297 assert((getContext().getTypeSize(E->getType()) == 64) &&((void)0)
9298 "rbit of unusual size!")((void)0);
9299 llvm::Value *Arg = EmitScalarExpr(E->getArg(0));
9300 return Builder.CreateCall(
9301 CGM.getIntrinsic(Intrinsic::bitreverse, Arg->getType()), Arg, "rbit");
9302 }
9303
9304 if (BuiltinID == AArch64::BI__builtin_arm_cls) {
9305 llvm::Value *Arg = EmitScalarExpr(E->getArg(0));
9306 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::aarch64_cls), Arg,
9307 "cls");
9308 }
9309 if (BuiltinID == AArch64::BI__builtin_arm_cls64) {
9310 llvm::Value *Arg = EmitScalarExpr(E->getArg(0));
9311 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::aarch64_cls64), Arg,
9312 "cls");
9313 }
9314
9315 if (BuiltinID == AArch64::BI__builtin_arm_frint32zf ||
9316 BuiltinID == AArch64::BI__builtin_arm_frint32z) {
9317 llvm::Value *Arg = EmitScalarExpr(E->getArg(0));
9318 llvm::Type *Ty = Arg->getType();
9319 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::aarch64_frint32z, Ty),
9320 Arg, "frint32z");
9321 }
9322
9323 if (BuiltinID == AArch64::BI__builtin_arm_frint64zf ||
9324 BuiltinID == AArch64::BI__builtin_arm_frint64z) {
9325 llvm::Value *Arg = EmitScalarExpr(E->getArg(0));
9326 llvm::Type *Ty = Arg->getType();
9327 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::aarch64_frint64z, Ty),
9328 Arg, "frint64z");
9329 }
9330
9331 if (BuiltinID == AArch64::BI__builtin_arm_frint32xf ||
9332 BuiltinID == AArch64::BI__builtin_arm_frint32x) {
9333 llvm::Value *Arg = EmitScalarExpr(E->getArg(0));
9334 llvm::Type *Ty = Arg->getType();
9335 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::aarch64_frint32x, Ty),
9336 Arg, "frint32x");
9337 }
9338
9339 if (BuiltinID == AArch64::BI__builtin_arm_frint64xf ||
9340 BuiltinID == AArch64::BI__builtin_arm_frint64x) {
9341 llvm::Value *Arg = EmitScalarExpr(E->getArg(0));
9342 llvm::Type *Ty = Arg->getType();
9343 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::aarch64_frint64x, Ty),
9344 Arg, "frint64x");
9345 }
9346
9347 if (BuiltinID == AArch64::BI__builtin_arm_jcvt) {
9348 assert((getContext().getTypeSize(E->getType()) == 32) &&((void)0)
9349 "__jcvt of unusual size!")((void)0);
9350 llvm::Value *Arg = EmitScalarExpr(E->getArg(0));
9351 return Builder.CreateCall(
9352 CGM.getIntrinsic(Intrinsic::aarch64_fjcvtzs), Arg);
9353 }
9354
9355 if (BuiltinID == AArch64::BI__builtin_arm_ld64b ||
9356 BuiltinID == AArch64::BI__builtin_arm_st64b ||
9357 BuiltinID == AArch64::BI__builtin_arm_st64bv ||
9358 BuiltinID == AArch64::BI__builtin_arm_st64bv0) {
9359 llvm::Value *MemAddr = EmitScalarExpr(E->getArg(0));
9360 llvm::Value *ValPtr = EmitScalarExpr(E->getArg(1));
9361
9362 if (BuiltinID == AArch64::BI__builtin_arm_ld64b) {
9363 // Load from the address via an LLVM intrinsic, receiving a
9364 // tuple of 8 i64 words, and store each one to ValPtr.
9365 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_ld64b);
9366 llvm::Value *Val = Builder.CreateCall(F, MemAddr);
9367 llvm::Value *ToRet;
9368 for (size_t i = 0; i < 8; i++) {
9369 llvm::Value *ValOffsetPtr =
9370 Builder.CreateGEP(Int64Ty, ValPtr, Builder.getInt32(i));
9371 Address Addr(ValOffsetPtr, CharUnits::fromQuantity(8));
9372 ToRet = Builder.CreateStore(Builder.CreateExtractValue(Val, i), Addr);
9373 }
9374 return ToRet;
9375 } else {
9376 // Load 8 i64 words from ValPtr, and store them to the address
9377 // via an LLVM intrinsic.
9378 SmallVector<llvm::Value *, 9> Args;
9379 Args.push_back(MemAddr);
9380 for (size_t i = 0; i < 8; i++) {
9381 llvm::Value *ValOffsetPtr =
9382 Builder.CreateGEP(Int64Ty, ValPtr, Builder.getInt32(i));
9383 Address Addr(ValOffsetPtr, CharUnits::fromQuantity(8));
9384 Args.push_back(Builder.CreateLoad(Addr));
9385 }
9386
9387 auto Intr = (BuiltinID == AArch64::BI__builtin_arm_st64b
9388 ? Intrinsic::aarch64_st64b
9389 : BuiltinID == AArch64::BI__builtin_arm_st64bv
9390 ? Intrinsic::aarch64_st64bv
9391 : Intrinsic::aarch64_st64bv0);
9392 Function *F = CGM.getIntrinsic(Intr);
9393 return Builder.CreateCall(F, Args);
9394 }
9395 }
9396
9397 if (BuiltinID == AArch64::BI__builtin_arm_rndr ||
9398 BuiltinID == AArch64::BI__builtin_arm_rndrrs) {
9399
9400 auto Intr = (BuiltinID == AArch64::BI__builtin_arm_rndr
9401 ? Intrinsic::aarch64_rndr
9402 : Intrinsic::aarch64_rndrrs);
9403 Function *F = CGM.getIntrinsic(Intr);
9404 llvm::Value *Val = Builder.CreateCall(F);
9405 Value *RandomValue = Builder.CreateExtractValue(Val, 0);
9406 Value *Status = Builder.CreateExtractValue(Val, 1);
9407
9408 Address MemAddress = EmitPointerWithAlignment(E->getArg(0));
9409 Builder.CreateStore(RandomValue, MemAddress);
9410 Status = Builder.CreateZExt(Status, Int32Ty);
9411 return Status;
9412 }
9413
9414 if (BuiltinID == AArch64::BI__clear_cache) {
9415 assert(E->getNumArgs() == 2 && "__clear_cache takes 2 arguments")((void)0);
9416 const FunctionDecl *FD = E->getDirectCallee();
9417 Value *Ops[2];
9418 for (unsigned i = 0; i < 2; i++)
9419 Ops[i] = EmitScalarExpr(E->getArg(i));
9420 llvm::Type *Ty = CGM.getTypes().ConvertType(FD->getType());
9421 llvm::FunctionType *FTy = cast<llvm::FunctionType>(Ty);
9422 StringRef Name = FD->getName();
9423 return EmitNounwindRuntimeCall(CGM.CreateRuntimeFunction(FTy, Name), Ops);
9424 }
9425
9426 if ((BuiltinID == AArch64::BI__builtin_arm_ldrex ||
9427 BuiltinID == AArch64::BI__builtin_arm_ldaex) &&
9428 getContext().getTypeSize(E->getType()) == 128) {
9429 Function *F = CGM.getIntrinsic(BuiltinID == AArch64::BI__builtin_arm_ldaex
9430 ? Intrinsic::aarch64_ldaxp
9431 : Intrinsic::aarch64_ldxp);
9432
9433 Value *LdPtr = EmitScalarExpr(E->getArg(0));
9434 Value *Val = Builder.CreateCall(F, Builder.CreateBitCast(LdPtr, Int8PtrTy),
9435 "ldxp");
9436
9437 Value *Val0 = Builder.CreateExtractValue(Val, 1);
9438 Value *Val1 = Builder.CreateExtractValue(Val, 0);
9439 llvm::Type *Int128Ty = llvm::IntegerType::get(getLLVMContext(), 128);
9440 Val0 = Builder.CreateZExt(Val0, Int128Ty);
9441 Val1 = Builder.CreateZExt(Val1, Int128Ty);
9442
9443 Value *ShiftCst = llvm::ConstantInt::get(Int128Ty, 64);
9444 Val = Builder.CreateShl(Val0, ShiftCst, "shl", true /* nuw */);
9445 Val = Builder.CreateOr(Val, Val1);
9446 return Builder.CreateBitCast(Val, ConvertType(E->getType()));
9447 } else if (BuiltinID == AArch64::BI__builtin_arm_ldrex ||
9448 BuiltinID == AArch64::BI__builtin_arm_ldaex) {
9449 Value *LoadAddr = EmitScalarExpr(E->getArg(0));
9450
9451 QualType Ty = E->getType();
9452 llvm::Type *RealResTy = ConvertType(Ty);
9453 llvm::Type *PtrTy = llvm::IntegerType::get(
9454 getLLVMContext(), getContext().getTypeSize(Ty))->getPointerTo();
9455 LoadAddr = Builder.CreateBitCast(LoadAddr, PtrTy);
9456
9457 Function *F = CGM.getIntrinsic(BuiltinID == AArch64::BI__builtin_arm_ldaex
9458 ? Intrinsic::aarch64_ldaxr
9459 : Intrinsic::aarch64_ldxr,
9460 PtrTy);
9461 Value *Val = Builder.CreateCall(F, LoadAddr, "ldxr");
9462
9463 if (RealResTy->isPointerTy())
9464 return Builder.CreateIntToPtr(Val, RealResTy);
9465
9466 llvm::Type *IntResTy = llvm::IntegerType::get(
9467 getLLVMContext(), CGM.getDataLayout().getTypeSizeInBits(RealResTy));
9468 Val = Builder.CreateTruncOrBitCast(Val, IntResTy);
9469 return Builder.CreateBitCast(Val, RealResTy);
9470 }
9471
9472 if ((BuiltinID == AArch64::BI__builtin_arm_strex ||
9473 BuiltinID == AArch64::BI__builtin_arm_stlex) &&
9474 getContext().getTypeSize(E->getArg(0)->getType()) == 128) {
9475 Function *F = CGM.getIntrinsic(BuiltinID == AArch64::BI__builtin_arm_stlex
9476 ? Intrinsic::aarch64_stlxp
9477 : Intrinsic::aarch64_stxp);
9478 llvm::Type *STy = llvm::StructType::get(Int64Ty, Int64Ty);
9479
9480 Address Tmp = CreateMemTemp(E->getArg(0)->getType());
9481 EmitAnyExprToMem(E->getArg(0), Tmp, Qualifiers(), /*init*/ true);
9482
9483 Tmp = Builder.CreateBitCast(Tmp, llvm::PointerType::getUnqual(STy));
9484 llvm::Value *Val = Builder.CreateLoad(Tmp);
9485
9486 Value *Arg0 = Builder.CreateExtractValue(Val, 0);
9487 Value *Arg1 = Builder.CreateExtractValue(Val, 1);
9488 Value *StPtr = Builder.CreateBitCast(EmitScalarExpr(E->getArg(1)),
9489 Int8PtrTy);
9490 return Builder.CreateCall(F, {Arg0, Arg1, StPtr}, "stxp");
9491 }
9492
9493 if (BuiltinID == AArch64::BI__builtin_arm_strex ||
9494 BuiltinID == AArch64::BI__builtin_arm_stlex) {
9495 Value *StoreVal = EmitScalarExpr(E->getArg(0));
9496 Value *StoreAddr = EmitScalarExpr(E->getArg(1));
9497
9498 QualType Ty = E->getArg(0)->getType();
9499 llvm::Type *StoreTy = llvm::IntegerType::get(getLLVMContext(),
9500 getContext().getTypeSize(Ty));
9501 StoreAddr = Builder.CreateBitCast(StoreAddr, StoreTy->getPointerTo());
9502
9503 if (StoreVal->getType()->isPointerTy())
9504 StoreVal = Builder.CreatePtrToInt(StoreVal, Int64Ty);
9505 else {
9506 llvm::Type *IntTy = llvm::IntegerType::get(
9507 getLLVMContext(),
9508 CGM.getDataLayout().getTypeSizeInBits(StoreVal->getType()));
9509 StoreVal = Builder.CreateBitCast(StoreVal, IntTy);
9510 StoreVal = Builder.CreateZExtOrBitCast(StoreVal, Int64Ty);
9511 }
9512
9513 Function *F = CGM.getIntrinsic(BuiltinID == AArch64::BI__builtin_arm_stlex
9514 ? Intrinsic::aarch64_stlxr
9515 : Intrinsic::aarch64_stxr,
9516 StoreAddr->getType());
9517 return Builder.CreateCall(F, {StoreVal, StoreAddr}, "stxr");
9518 }
9519
9520 if (BuiltinID == AArch64::BI__getReg) {
9521 Expr::EvalResult Result;
9522 if (!E->getArg(0)->EvaluateAsInt(Result, CGM.getContext()))
9523 llvm_unreachable("Sema will ensure that the parameter is constant")__builtin_unreachable();
9524
9525 llvm::APSInt Value = Result.Val.getInt();
9526 LLVMContext &Context = CGM.getLLVMContext();
9527 std::string Reg = Value == 31 ? "sp" : "x" + toString(Value, 10);
9528
9529 llvm::Metadata *Ops[] = {llvm::MDString::get(Context, Reg)};
9530 llvm::MDNode *RegName = llvm::MDNode::get(Context, Ops);
9531 llvm::Value *Metadata = llvm::MetadataAsValue::get(Context, RegName);
9532
9533 llvm::Function *F =
9534 CGM.getIntrinsic(llvm::Intrinsic::read_register, {Int64Ty});
9535 return Builder.CreateCall(F, Metadata);
9536 }
9537
9538 if (BuiltinID == AArch64::BI__builtin_arm_clrex) {
9539 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_clrex);
9540 return Builder.CreateCall(F);
9541 }
9542
9543 if (BuiltinID == AArch64::BI_ReadWriteBarrier)
9544 return Builder.CreateFence(llvm::AtomicOrdering::SequentiallyConsistent,
9545 llvm::SyncScope::SingleThread);
9546
9547 // CRC32
9548 Intrinsic::ID CRCIntrinsicID = Intrinsic::not_intrinsic;
9549 switch (BuiltinID) {
9550 case AArch64::BI__builtin_arm_crc32b:
9551 CRCIntrinsicID = Intrinsic::aarch64_crc32b; break;
9552 case AArch64::BI__builtin_arm_crc32cb:
9553 CRCIntrinsicID = Intrinsic::aarch64_crc32cb; break;
9554 case AArch64::BI__builtin_arm_crc32h:
9555 CRCIntrinsicID = Intrinsic::aarch64_crc32h; break;
9556 case AArch64::BI__builtin_arm_crc32ch:
9557 CRCIntrinsicID = Intrinsic::aarch64_crc32ch; break;
9558 case AArch64::BI__builtin_arm_crc32w:
9559 CRCIntrinsicID = Intrinsic::aarch64_crc32w; break;
9560 case AArch64::BI__builtin_arm_crc32cw:
9561 CRCIntrinsicID = Intrinsic::aarch64_crc32cw; break;
9562 case AArch64::BI__builtin_arm_crc32d:
9563 CRCIntrinsicID = Intrinsic::aarch64_crc32x; break;
9564 case AArch64::BI__builtin_arm_crc32cd:
9565 CRCIntrinsicID = Intrinsic::aarch64_crc32cx; break;
9566 }
9567
9568 if (CRCIntrinsicID != Intrinsic::not_intrinsic) {
9569 Value *Arg0 = EmitScalarExpr(E->getArg(0));
9570 Value *Arg1 = EmitScalarExpr(E->getArg(1));
9571 Function *F = CGM.getIntrinsic(CRCIntrinsicID);
9572
9573 llvm::Type *DataTy = F->getFunctionType()->getParamType(1);
9574 Arg1 = Builder.CreateZExtOrBitCast(Arg1, DataTy);
9575
9576 return Builder.CreateCall(F, {Arg0, Arg1});
9577 }
9578
9579 // Memory Tagging Extensions (MTE) Intrinsics
9580 Intrinsic::ID MTEIntrinsicID = Intrinsic::not_intrinsic;
9581 switch (BuiltinID) {
9582 case AArch64::BI__builtin_arm_irg:
9583 MTEIntrinsicID = Intrinsic::aarch64_irg; break;
9584 case AArch64::BI__builtin_arm_addg:
9585 MTEIntrinsicID = Intrinsic::aarch64_addg; break;
9586 case AArch64::BI__builtin_arm_gmi:
9587 MTEIntrinsicID = Intrinsic::aarch64_gmi; break;
9588 case AArch64::BI__builtin_arm_ldg:
9589 MTEIntrinsicID = Intrinsic::aarch64_ldg; break;
9590 case AArch64::BI__builtin_arm_stg:
9591 MTEIntrinsicID = Intrinsic::aarch64_stg; break;
9592 case AArch64::BI__builtin_arm_subp:
9593 MTEIntrinsicID = Intrinsic::aarch64_subp; break;
9594 }
9595
9596 if (MTEIntrinsicID != Intrinsic::not_intrinsic) {
9597 llvm::Type *T = ConvertType(E->getType());
9598
9599 if (MTEIntrinsicID == Intrinsic::aarch64_irg) {
9600 Value *Pointer = EmitScalarExpr(E->getArg(0));
9601 Value *Mask = EmitScalarExpr(E->getArg(1));
9602
9603 Pointer = Builder.CreatePointerCast(Pointer, Int8PtrTy);
9604 Mask = Builder.CreateZExt(Mask, Int64Ty);
9605 Value *RV = Builder.CreateCall(
9606 CGM.getIntrinsic(MTEIntrinsicID), {Pointer, Mask});
9607 return Builder.CreatePointerCast(RV, T);
9608 }
9609 if (MTEIntrinsicID == Intrinsic::aarch64_addg) {
9610 Value *Pointer = EmitScalarExpr(E->getArg(0));
9611 Value *TagOffset = EmitScalarExpr(E->getArg(1));
9612
9613 Pointer = Builder.CreatePointerCast(Pointer, Int8PtrTy);
9614 TagOffset = Builder.CreateZExt(TagOffset, Int64Ty);
9615 Value *RV = Builder.CreateCall(
9616 CGM.getIntrinsic(MTEIntrinsicID), {Pointer, TagOffset});
9617 return Builder.CreatePointerCast(RV, T);
9618 }
9619 if (MTEIntrinsicID == Intrinsic::aarch64_gmi) {
9620 Value *Pointer = EmitScalarExpr(E->getArg(0));
9621 Value *ExcludedMask = EmitScalarExpr(E->getArg(1));
9622
9623 ExcludedMask = Builder.CreateZExt(ExcludedMask, Int64Ty);
9624 Pointer = Builder.CreatePointerCast(Pointer, Int8PtrTy);
9625 return Builder.CreateCall(
9626 CGM.getIntrinsic(MTEIntrinsicID), {Pointer, ExcludedMask});
9627 }
9628 // Although it is possible to supply a different return
9629 // address (first arg) to this intrinsic, for now we set
9630 // return address same as input address.
9631 if (MTEIntrinsicID == Intrinsic::aarch64_ldg) {
9632 Value *TagAddress = EmitScalarExpr(E->getArg(0));
9633 TagAddress = Builder.CreatePointerCast(TagAddress, Int8PtrTy);
9634 Value *RV = Builder.CreateCall(
9635 CGM.getIntrinsic(MTEIntrinsicID), {TagAddress, TagAddress});
9636 return Builder.CreatePointerCast(RV, T);
9637 }
9638 // Although it is possible to supply a different tag (to set)
9639 // to this intrinsic (as first arg), for now we supply
9640 // the tag that is in input address arg (common use case).
9641 if (MTEIntrinsicID == Intrinsic::aarch64_stg) {
9642 Value *TagAddress = EmitScalarExpr(E->getArg(0));
9643 TagAddress = Builder.CreatePointerCast(TagAddress, Int8PtrTy);
9644 return Builder.CreateCall(
9645 CGM.getIntrinsic(MTEIntrinsicID), {TagAddress, TagAddress});
9646 }
9647 if (MTEIntrinsicID == Intrinsic::aarch64_subp) {
9648 Value *PointerA = EmitScalarExpr(E->getArg(0));
9649 Value *PointerB = EmitScalarExpr(E->getArg(1));
9650 PointerA = Builder.CreatePointerCast(PointerA, Int8PtrTy);
9651 PointerB = Builder.CreatePointerCast(PointerB, Int8PtrTy);
9652 return Builder.CreateCall(
9653 CGM.getIntrinsic(MTEIntrinsicID), {PointerA, PointerB});
9654 }
9655 }
9656
9657 if (BuiltinID == AArch64::BI__builtin_arm_rsr ||
9658 BuiltinID == AArch64::BI__builtin_arm_rsr64 ||
9659 BuiltinID == AArch64::BI__builtin_arm_rsrp ||
9660 BuiltinID == AArch64::BI__builtin_arm_wsr ||
9661 BuiltinID == AArch64::BI__builtin_arm_wsr64 ||
9662 BuiltinID == AArch64::BI__builtin_arm_wsrp) {
9663
9664 SpecialRegisterAccessKind AccessKind = Write;
9665 if (BuiltinID == AArch64::BI__builtin_arm_rsr ||
9666 BuiltinID == AArch64::BI__builtin_arm_rsr64 ||
9667 BuiltinID == AArch64::BI__builtin_arm_rsrp)
9668 AccessKind = VolatileRead;
9669
9670 bool IsPointerBuiltin = BuiltinID == AArch64::BI__builtin_arm_rsrp ||
9671 BuiltinID == AArch64::BI__builtin_arm_wsrp;
9672
9673 bool Is64Bit = BuiltinID != AArch64::BI__builtin_arm_rsr &&
9674 BuiltinID != AArch64::BI__builtin_arm_wsr;
9675
9676 llvm::Type *ValueType;
9677 llvm::Type *RegisterType = Int64Ty;
9678 if (IsPointerBuiltin) {
9679 ValueType = VoidPtrTy;
9680 } else if (Is64Bit) {
9681 ValueType = Int64Ty;
9682 } else {
9683 ValueType = Int32Ty;
9684 }
9685
9686 return EmitSpecialRegisterBuiltin(*this, E, RegisterType, ValueType,
9687 AccessKind);
9688 }
9689
9690 if (BuiltinID == AArch64::BI_ReadStatusReg ||
9691 BuiltinID == AArch64::BI_WriteStatusReg) {
9692 LLVMContext &Context = CGM.getLLVMContext();
9693
9694 unsigned SysReg =
9695 E->getArg(0)->EvaluateKnownConstInt(getContext()).getZExtValue();
9696
9697 std::string SysRegStr;
9698 llvm::raw_string_ostream(SysRegStr) <<
9699 ((1 << 1) | ((SysReg >> 14) & 1)) << ":" <<
9700 ((SysReg >> 11) & 7) << ":" <<
9701 ((SysReg >> 7) & 15) << ":" <<
9702 ((SysReg >> 3) & 15) << ":" <<
9703 ( SysReg & 7);
9704
9705 llvm::Metadata *Ops[] = { llvm::MDString::get(Context, SysRegStr) };
9706 llvm::MDNode *RegName = llvm::MDNode::get(Context, Ops);
9707 llvm::Value *Metadata = llvm::MetadataAsValue::get(Context, RegName);
9708
9709 llvm::Type *RegisterType = Int64Ty;
9710 llvm::Type *Types[] = { RegisterType };
9711
9712 if (BuiltinID == AArch64::BI_ReadStatusReg) {
9713 llvm::Function *F = CGM.getIntrinsic(llvm::Intrinsic::read_register, Types);
9714
9715 return Builder.CreateCall(F, Metadata);
9716 }
9717
9718 llvm::Function *F = CGM.getIntrinsic(llvm::Intrinsic::write_register, Types);
9719 llvm::Value *ArgValue = EmitScalarExpr(E->getArg(1));
9720
9721 return Builder.CreateCall(F, { Metadata, ArgValue });
9722 }
9723
9724 if (BuiltinID == AArch64::BI_AddressOfReturnAddress) {
9725 llvm::Function *F =
9726 CGM.getIntrinsic(Intrinsic::addressofreturnaddress, AllocaInt8PtrTy);
9727 return Builder.CreateCall(F);
9728 }
9729
9730 if (BuiltinID == AArch64::BI__builtin_sponentry) {
9731 llvm::Function *F = CGM.getIntrinsic(Intrinsic::sponentry, AllocaInt8PtrTy);
9732 return Builder.CreateCall(F);
9733 }
9734
9735 if (BuiltinID == AArch64::BI__mulh || BuiltinID == AArch64::BI__umulh) {
9736 llvm::Type *ResType = ConvertType(E->getType());
9737 llvm::Type *Int128Ty = llvm::IntegerType::get(getLLVMContext(), 128);
9738
9739 bool IsSigned = BuiltinID == AArch64::BI__mulh;
9740 Value *LHS =
9741 Builder.CreateIntCast(EmitScalarExpr(E->getArg(0)), Int128Ty, IsSigned);
9742 Value *RHS =
9743 Builder.CreateIntCast(EmitScalarExpr(E->getArg(1)), Int128Ty, IsSigned);
9744
9745 Value *MulResult, *HigherBits;
9746 if (IsSigned) {
9747 MulResult = Builder.CreateNSWMul(LHS, RHS);
9748 HigherBits = Builder.CreateAShr(MulResult, 64);
9749 } else {
9750 MulResult = Builder.CreateNUWMul(LHS, RHS);
9751 HigherBits = Builder.CreateLShr(MulResult, 64);
9752 }
9753 HigherBits = Builder.CreateIntCast(HigherBits, ResType, IsSigned);
9754
9755 return HigherBits;
9756 }
9757
9758 // Handle MSVC intrinsics before argument evaluation to prevent double
9759 // evaluation.
9760 if (Optional<MSVCIntrin> MsvcIntId = translateAarch64ToMsvcIntrin(BuiltinID))
9761 return EmitMSVCBuiltinExpr(*MsvcIntId, E);
9762
9763 // Find out if any arguments are required to be integer constant
9764 // expressions.
9765 unsigned ICEArguments = 0;
9766 ASTContext::GetBuiltinTypeError Error;
9767 getContext().GetBuiltinType(BuiltinID, Error, &ICEArguments);
9768 assert(Error == ASTContext::GE_None && "Should not codegen an error")((void)0);
9769
9770 llvm::SmallVector<Value*, 4> Ops;
9771 Address PtrOp0 = Address::invalid();
9772 for (unsigned i = 0, e = E->getNumArgs() - 1; i != e; i++) {
9773 if (i == 0) {
9774 switch (BuiltinID) {
9775 case NEON::BI__builtin_neon_vld1_v:
9776 case NEON::BI__builtin_neon_vld1q_v:
9777 case NEON::BI__builtin_neon_vld1_dup_v:
9778 case NEON::BI__builtin_neon_vld1q_dup_v:
9779 case NEON::BI__builtin_neon_vld1_lane_v:
9780 case NEON::BI__builtin_neon_vld1q_lane_v:
9781 case NEON::BI__builtin_neon_vst1_v:
9782 case NEON::BI__builtin_neon_vst1q_v:
9783 case NEON::BI__builtin_neon_vst1_lane_v:
9784 case NEON::BI__builtin_neon_vst1q_lane_v:
9785 // Get the alignment for the argument in addition to the value;
9786 // we'll use it later.
9787 PtrOp0 = EmitPointerWithAlignment(E->getArg(0));
9788 Ops.push_back(PtrOp0.getPointer());
9789 continue;
9790 }
9791 }
9792 if ((ICEArguments & (1 << i)) == 0) {
9793 Ops.push_back(EmitScalarExpr(E->getArg(i)));
9794 } else {
9795 // If this is required to be a constant, constant fold it so that we know
9796 // that the generated intrinsic gets a ConstantInt.
9797 Ops.push_back(llvm::ConstantInt::get(
9798 getLLVMContext(),
9799 *E->getArg(i)->getIntegerConstantExpr(getContext())));
9800 }
9801 }
9802
9803 auto SISDMap = makeArrayRef(AArch64SISDIntrinsicMap);
9804 const ARMVectorIntrinsicInfo *Builtin = findARMVectorIntrinsicInMap(
9805 SISDMap, BuiltinID, AArch64SISDIntrinsicsProvenSorted);
9806
9807 if (Builtin) {
9808 Ops.push_back(EmitScalarExpr(E->getArg(E->getNumArgs() - 1)));
9809 Value *Result = EmitCommonNeonSISDBuiltinExpr(*this, *Builtin, Ops, E);
9810 assert(Result && "SISD intrinsic should have been handled")((void)0);
9811 return Result;
9812 }
9813
9814 const Expr *Arg = E->getArg(E->getNumArgs()-1);
9815 NeonTypeFlags Type(0);
9816 if (Optional<llvm::APSInt> Result = Arg->getIntegerConstantExpr(getContext()))
9817 // Determine the type of this overloaded NEON intrinsic.
9818 Type = NeonTypeFlags(Result->getZExtValue());
9819
9820 bool usgn = Type.isUnsigned();
9821 bool quad = Type.isQuad();
9822
9823 // Handle non-overloaded intrinsics first.
9824 switch (BuiltinID) {
9825 default: break;
9826 case NEON::BI__builtin_neon_vabsh_f16:
9827 Ops.push_back(EmitScalarExpr(E->getArg(0)));
9828 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::fabs, HalfTy), Ops, "vabs");
9829 case NEON::BI__builtin_neon_vaddq_p128: {
9830 llvm::Type *Ty = GetNeonType(this, NeonTypeFlags::Poly128);
9831 Ops.push_back(EmitScalarExpr(E->getArg(1)));
9832 Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
9833 Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
9834 Ops[0] = Builder.CreateXor(Ops[0], Ops[1]);
9835 llvm::Type *Int128Ty = llvm::Type::getIntNTy(getLLVMContext(), 128);
9836 return Builder.CreateBitCast(Ops[0], Int128Ty);
9837 }
9838 case NEON::BI__builtin_neon_vldrq_p128: {
9839 llvm::Type *Int128Ty = llvm::Type::getIntNTy(getLLVMContext(), 128);
9840 llvm::Type *Int128PTy = llvm::PointerType::get(Int128Ty, 0);
9841 Value *Ptr = Builder.CreateBitCast(EmitScalarExpr(E->getArg(0)), Int128PTy);
9842 return Builder.CreateAlignedLoad(Int128Ty, Ptr,
9843 CharUnits::fromQuantity(16));
9844 }
9845 case NEON::BI__builtin_neon_vstrq_p128: {
9846 llvm::Type *Int128PTy = llvm::Type::getIntNPtrTy(getLLVMContext(), 128);
9847 Value *Ptr = Builder.CreateBitCast(Ops[0], Int128PTy);
9848 return Builder.CreateDefaultAlignedStore(EmitScalarExpr(E->getArg(1)), Ptr);
9849 }
9850 case NEON::BI__builtin_neon_vcvts_f32_u32:
9851 case NEON::BI__builtin_neon_vcvtd_f64_u64:
9852 usgn = true;
9853 LLVM_FALLTHROUGH[[gnu::fallthrough]];
9854 case NEON::BI__builtin_neon_vcvts_f32_s32:
9855 case NEON::BI__builtin_neon_vcvtd_f64_s64: {
9856 Ops.push_back(EmitScalarExpr(E->getArg(0)));
9857 bool Is64 = Ops[0]->getType()->getPrimitiveSizeInBits() == 64;
9858 llvm::Type *InTy = Is64 ? Int64Ty : Int32Ty;
9859 llvm::Type *FTy = Is64 ? DoubleTy : FloatTy;
9860 Ops[0] = Builder.CreateBitCast(Ops[0], InTy);
9861 if (usgn)
9862 return Builder.CreateUIToFP(Ops[0], FTy);
9863 return Builder.CreateSIToFP(Ops[0], FTy);
9864 }
9865 case NEON::BI__builtin_neon_vcvth_f16_u16:
9866 case NEON::BI__builtin_neon_vcvth_f16_u32:
9867 case NEON::BI__builtin_neon_vcvth_f16_u64:
9868 usgn = true;
9869 LLVM_FALLTHROUGH[[gnu::fallthrough]];
9870 case NEON::BI__builtin_neon_vcvth_f16_s16:
9871 case NEON::BI__builtin_neon_vcvth_f16_s32:
9872 case NEON::BI__builtin_neon_vcvth_f16_s64: {
9873 Ops.push_back(EmitScalarExpr(E->getArg(0)));
9874 llvm::Type *FTy = HalfTy;
9875 llvm::Type *InTy;
9876 if (Ops[0]->getType()->getPrimitiveSizeInBits() == 64)
9877 InTy = Int64Ty;
9878 else if (Ops[0]->getType()->getPrimitiveSizeInBits() == 32)
9879 InTy = Int32Ty;
9880 else
9881 InTy = Int16Ty;
9882 Ops[0] = Builder.CreateBitCast(Ops[0], InTy);
9883 if (usgn)
9884 return Builder.CreateUIToFP(Ops[0], FTy);
9885 return Builder.CreateSIToFP(Ops[0], FTy);
9886 }
9887 case NEON::BI__builtin_neon_vcvtah_u16_f16:
9888 case NEON::BI__builtin_neon_vcvtmh_u16_f16:
9889 case NEON::BI__builtin_neon_vcvtnh_u16_f16:
9890 case NEON::BI__builtin_neon_vcvtph_u16_f16:
9891 case NEON::BI__builtin_neon_vcvth_u16_f16:
9892 case NEON::BI__builtin_neon_vcvtah_s16_f16:
9893 case NEON::BI__builtin_neon_vcvtmh_s16_f16:
9894 case NEON::BI__builtin_neon_vcvtnh_s16_f16:
9895 case NEON::BI__builtin_neon_vcvtph_s16_f16:
9896 case NEON::BI__builtin_neon_vcvth_s16_f16: {
9897 unsigned Int;
9898 llvm::Type* InTy = Int32Ty;
9899 llvm::Type* FTy = HalfTy;
9900 llvm::Type *Tys[2] = {InTy, FTy};
9901 Ops.push_back(EmitScalarExpr(E->getArg(0)));
9902 switch (BuiltinID) {
9903 default: llvm_unreachable("missing builtin ID in switch!")__builtin_unreachable();
9904 case NEON::BI__builtin_neon_vcvtah_u16_f16:
9905 Int = Intrinsic::aarch64_neon_fcvtau; break;
9906 case NEON::BI__builtin_neon_vcvtmh_u16_f16:
9907 Int = Intrinsic::aarch64_neon_fcvtmu; break;
9908 case NEON::BI__builtin_neon_vcvtnh_u16_f16:
9909 Int = Intrinsic::aarch64_neon_fcvtnu; break;
9910 case NEON::BI__builtin_neon_vcvtph_u16_f16:
9911 Int = Intrinsic::aarch64_neon_fcvtpu; break;
9912 case NEON::BI__builtin_neon_vcvth_u16_f16:
9913 Int = Intrinsic::aarch64_neon_fcvtzu; break;
9914 case NEON::BI__builtin_neon_vcvtah_s16_f16:
9915 Int = Intrinsic::aarch64_neon_fcvtas; break;
9916 case NEON::BI__builtin_neon_vcvtmh_s16_f16:
9917 Int = Intrinsic::aarch64_neon_fcvtms; break;
9918 case NEON::BI__builtin_neon_vcvtnh_s16_f16:
9919 Int = Intrinsic::aarch64_neon_fcvtns; break;
9920 case NEON::BI__builtin_neon_vcvtph_s16_f16:
9921 Int = Intrinsic::aarch64_neon_fcvtps; break;
9922 case NEON::BI__builtin_neon_vcvth_s16_f16:
9923 Int = Intrinsic::aarch64_neon_fcvtzs; break;
9924 }
9925 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "fcvt");
9926 return Builder.CreateTrunc(Ops[0], Int16Ty);
9927 }
9928 case NEON::BI__builtin_neon_vcaleh_f16:
9929 case NEON::BI__builtin_neon_vcalth_f16:
9930 case NEON::BI__builtin_neon_vcageh_f16:
9931 case NEON::BI__builtin_neon_vcagth_f16: {
9932 unsigned Int;
9933 llvm::Type* InTy = Int32Ty;
9934 llvm::Type* FTy = HalfTy;
9935 llvm::Type *Tys[2] = {InTy, FTy};
9936 Ops.push_back(EmitScalarExpr(E->getArg(1)));
9937 switch (BuiltinID) {
9938 default: llvm_unreachable("missing builtin ID in switch!")__builtin_unreachable();
9939 case NEON::BI__builtin_neon_vcageh_f16:
9940 Int = Intrinsic::aarch64_neon_facge; break;
9941 case NEON::BI__builtin_neon_vcagth_f16:
9942 Int = Intrinsic::aarch64_neon_facgt; break;
9943 case NEON::BI__builtin_neon_vcaleh_f16:
9944 Int = Intrinsic::aarch64_neon_facge; std::swap(Ops[0], Ops[1]); break;
9945 case NEON::BI__builtin_neon_vcalth_f16:
9946 Int = Intrinsic::aarch64_neon_facgt; std::swap(Ops[0], Ops[1]); break;
9947 }
9948 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "facg");
9949 return Builder.CreateTrunc(Ops[0], Int16Ty);
9950 }
9951 case NEON::BI__builtin_neon_vcvth_n_s16_f16:
9952 case NEON::BI__builtin_neon_vcvth_n_u16_f16: {
9953 unsigned Int;
9954 llvm::Type* InTy = Int32Ty;
9955 llvm::Type* FTy = HalfTy;
9956 llvm::Type *Tys[2] = {InTy, FTy};
9957 Ops.push_back(EmitScalarExpr(E->getArg(1)));
9958 switch (BuiltinID) {
9959 default: llvm_unreachable("missing builtin ID in switch!")__builtin_unreachable();
9960 case NEON::BI__builtin_neon_vcvth_n_s16_f16:
9961 Int = Intrinsic::aarch64_neon_vcvtfp2fxs; break;
9962 case NEON::BI__builtin_neon_vcvth_n_u16_f16:
9963 Int = Intrinsic::aarch64_neon_vcvtfp2fxu; break;
9964 }
9965 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "fcvth_n");
9966 return Builder.CreateTrunc(Ops[0], Int16Ty);
9967 }
9968 case NEON::BI__builtin_neon_vcvth_n_f16_s16:
9969 case NEON::BI__builtin_neon_vcvth_n_f16_u16: {
9970 unsigned Int;
9971 llvm::Type* FTy = HalfTy;
9972 llvm::Type* InTy = Int32Ty;
9973 llvm::Type *Tys[2] = {FTy, InTy};
9974 Ops.push_back(EmitScalarExpr(E->getArg(1)));
9975 switch (BuiltinID) {
9976 default: llvm_unreachable("missing builtin ID in switch!")__builtin_unreachable();
9977 case NEON::BI__builtin_neon_vcvth_n_f16_s16:
9978 Int = Intrinsic::aarch64_neon_vcvtfxs2fp;
9979 Ops[0] = Builder.CreateSExt(Ops[0], InTy, "sext");
9980 break;
9981 case NEON::BI__builtin_neon_vcvth_n_f16_u16:
9982 Int = Intrinsic::aarch64_neon_vcvtfxu2fp;
9983 Ops[0] = Builder.CreateZExt(Ops[0], InTy);
9984 break;
9985 }
9986 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "fcvth_n");
9987 }
9988 case NEON::BI__builtin_neon_vpaddd_s64: {
9989 auto *Ty = llvm::FixedVectorType::get(Int64Ty, 2);
9990 Value *Vec = EmitScalarExpr(E->getArg(0));
9991 // The vector is v2f64, so make sure it's bitcast to that.
9992 Vec = Builder.CreateBitCast(Vec, Ty, "v2i64");
9993 llvm::Value *Idx0 = llvm::ConstantInt::get(SizeTy, 0);
9994 llvm::Value *Idx1 = llvm::ConstantInt::get(SizeTy, 1);
9995 Value *Op0 = Builder.CreateExtractElement(Vec, Idx0, "lane0");
9996 Value *Op1 = Builder.CreateExtractElement(Vec, Idx1, "lane1");
9997 // Pairwise addition of a v2f64 into a scalar f64.
9998 return Builder.CreateAdd(Op0, Op1, "vpaddd");
9999 }
10000 case NEON::BI__builtin_neon_vpaddd_f64: {
10001 auto *Ty = llvm::FixedVectorType::get(DoubleTy, 2);
10002 Value *Vec = EmitScalarExpr(E->getArg(0));
10003 // The vector is v2f64, so make sure it's bitcast to that.
10004 Vec = Builder.CreateBitCast(Vec, Ty, "v2f64");
10005 llvm::Value *Idx0 = llvm::ConstantInt::get(SizeTy, 0);
10006 llvm::Value *Idx1 = llvm::ConstantInt::get(SizeTy, 1);
10007 Value *Op0 = Builder.CreateExtractElement(Vec, Idx0, "lane0");
10008 Value *Op1 = Builder.CreateExtractElement(Vec, Idx1, "lane1");
10009 // Pairwise addition of a v2f64 into a scalar f64.
10010 return Builder.CreateFAdd(Op0, Op1, "vpaddd");
10011 }
10012 case NEON::BI__builtin_neon_vpadds_f32: {
10013 auto *Ty = llvm::FixedVectorType::get(FloatTy, 2);
10014 Value *Vec = EmitScalarExpr(E->getArg(0));
10015 // The vector is v2f32, so make sure it's bitcast to that.
10016 Vec = Builder.CreateBitCast(Vec, Ty, "v2f32");
10017 llvm::Value *Idx0 = llvm::ConstantInt::get(SizeTy, 0);
10018 llvm::Value *Idx1 = llvm::ConstantInt::get(SizeTy, 1);
10019 Value *Op0 = Builder.CreateExtractElement(Vec, Idx0, "lane0");
10020 Value *Op1 = Builder.CreateExtractElement(Vec, Idx1, "lane1");
10021 // Pairwise addition of a v2f32 into a scalar f32.
10022 return Builder.CreateFAdd(Op0, Op1, "vpaddd");
10023 }
10024 case NEON::BI__builtin_neon_vceqzd_s64:
10025 case NEON::BI__builtin_neon_vceqzd_f64:
10026 case NEON::BI__builtin_neon_vceqzs_f32:
10027 case NEON::BI__builtin_neon_vceqzh_f16:
10028 Ops.push_back(EmitScalarExpr(E->getArg(0)));
10029 return EmitAArch64CompareBuiltinExpr(
10030 Ops[0], ConvertType(E->getCallReturnType(getContext())),
10031 ICmpInst::FCMP_OEQ, ICmpInst::ICMP_EQ, "vceqz");
10032 case NEON::BI__builtin_neon_vcgezd_s64:
10033 case NEON::BI__builtin_neon_vcgezd_f64:
10034 case NEON::BI__builtin_neon_vcgezs_f32:
10035 case NEON::BI__builtin_neon_vcgezh_f16:
10036 Ops.push_back(EmitScalarExpr(E->getArg(0)));
10037 return EmitAArch64CompareBuiltinExpr(
10038 Ops[0], ConvertType(E->getCallReturnType(getContext())),
10039 ICmpInst::FCMP_OGE, ICmpInst::ICMP_SGE, "vcgez");
10040 case NEON::BI__builtin_neon_vclezd_s64:
10041 case NEON::BI__builtin_neon_vclezd_f64:
10042 case NEON::BI__builtin_neon_vclezs_f32:
10043 case NEON::BI__builtin_neon_vclezh_f16:
10044 Ops.push_back(EmitScalarExpr(E->getArg(0)));
10045 return EmitAArch64CompareBuiltinExpr(
10046 Ops[0], ConvertType(E->getCallReturnType(getContext())),
10047 ICmpInst::FCMP_OLE, ICmpInst::ICMP_SLE, "vclez");
10048 case NEON::BI__builtin_neon_vcgtzd_s64:
10049 case NEON::BI__builtin_neon_vcgtzd_f64:
10050 case NEON::BI__builtin_neon_vcgtzs_f32:
10051 case NEON::BI__builtin_neon_vcgtzh_f16:
10052 Ops.push_back(EmitScalarExpr(E->getArg(0)));
10053 return EmitAArch64CompareBuiltinExpr(
10054 Ops[0], ConvertType(E->getCallReturnType(getContext())),
10055 ICmpInst::FCMP_OGT, ICmpInst::ICMP_SGT, "vcgtz");
10056 case NEON::BI__builtin_neon_vcltzd_s64:
10057 case NEON::BI__builtin_neon_vcltzd_f64:
10058 case NEON::BI__builtin_neon_vcltzs_f32:
10059 case NEON::BI__builtin_neon_vcltzh_f16:
10060 Ops.push_back(EmitScalarExpr(E->getArg(0)));
10061 return EmitAArch64CompareBuiltinExpr(
10062 Ops[0], ConvertType(E->getCallReturnType(getContext())),
10063 ICmpInst::FCMP_OLT, ICmpInst::ICMP_SLT, "vcltz");
10064
10065 case NEON::BI__builtin_neon_vceqzd_u64: {
10066 Ops.push_back(EmitScalarExpr(E->getArg(0)));
10067 Ops[0] = Builder.CreateBitCast(Ops[0], Int64Ty);
10068 Ops[0] =
10069 Builder.CreateICmpEQ(Ops[0], llvm::Constant::getNullValue(Int64Ty));
10070 return Builder.CreateSExt(Ops[0], Int64Ty, "vceqzd");
10071 }
10072 case NEON::BI__builtin_neon_vceqd_f64:
10073 case NEON::BI__builtin_neon_vcled_f64:
10074 case NEON::BI__builtin_neon_vcltd_f64:
10075 case NEON::BI__builtin_neon_vcged_f64:
10076 case NEON::BI__builtin_neon_vcgtd_f64: {
10077 llvm::CmpInst::Predicate P;
10078 switch (BuiltinID) {
10079 default: llvm_unreachable("missing builtin ID in switch!")__builtin_unreachable();
10080 case NEON::BI__builtin_neon_vceqd_f64: P = llvm::FCmpInst::FCMP_OEQ; break;
10081 case NEON::BI__builtin_neon_vcled_f64: P = llvm::FCmpInst::FCMP_OLE; break;
10082 case NEON::BI__builtin_neon_vcltd_f64: P = llvm::FCmpInst::FCMP_OLT; break;
10083 case NEON::BI__builtin_neon_vcged_f64: P = llvm::FCmpInst::FCMP_OGE; break;
10084 case NEON::BI__builtin_neon_vcgtd_f64: P = llvm::FCmpInst::FCMP_OGT; break;
10085 }
10086 Ops.push_back(EmitScalarExpr(E->getArg(1)));
10087 Ops[0] = Builder.CreateBitCast(Ops[0], DoubleTy);
10088 Ops[1] = Builder.CreateBitCast(Ops[1], DoubleTy);
10089 Ops[0] = Builder.CreateFCmp(P, Ops[0], Ops[1]);
10090 return Builder.CreateSExt(Ops[0], Int64Ty, "vcmpd");
10091 }
10092 case NEON::BI__builtin_neon_vceqs_f32:
10093 case NEON::BI__builtin_neon_vcles_f32:
10094 case NEON::BI__builtin_neon_vclts_f32:
10095 case NEON::BI__builtin_neon_vcges_f32:
10096 case NEON::BI__builtin_neon_vcgts_f32: {
10097 llvm::CmpInst::Predicate P;
10098 switch (BuiltinID) {
10099 default: llvm_unreachable("missing builtin ID in switch!")__builtin_unreachable();
10100 case NEON::BI__builtin_neon_vceqs_f32: P = llvm::FCmpInst::FCMP_OEQ; break;
10101 case NEON::BI__builtin_neon_vcles_f32: P = llvm::FCmpInst::FCMP_OLE; break;
10102 case NEON::BI__builtin_neon_vclts_f32: P = llvm::FCmpInst::FCMP_OLT; break;
10103 case NEON::BI__builtin_neon_vcges_f32: P = llvm::FCmpInst::FCMP_OGE; break;
10104 case NEON::BI__builtin_neon_vcgts_f32: P = llvm::FCmpInst::FCMP_OGT; break;
10105 }
10106 Ops.push_back(EmitScalarExpr(E->getArg(1)));
10107 Ops[0] = Builder.CreateBitCast(Ops[0], FloatTy);
10108 Ops[1] = Builder.CreateBitCast(Ops[1], FloatTy);
10109 Ops[0] = Builder.CreateFCmp(P, Ops[0], Ops[1]);
10110 return Builder.CreateSExt(Ops[0], Int32Ty, "vcmpd");
10111 }
10112 case NEON::BI__builtin_neon_vceqh_f16:
10113 case NEON::BI__builtin_neon_vcleh_f16:
10114 case NEON::BI__builtin_neon_vclth_f16:
10115 case NEON::BI__builtin_neon_vcgeh_f16:
10116 case NEON::BI__builtin_neon_vcgth_f16: {
10117 llvm::CmpInst::Predicate P;
10118 switch (BuiltinID) {
10119 default: llvm_unreachable("missing builtin ID in switch!")__builtin_unreachable();
10120 case NEON::BI__builtin_neon_vceqh_f16: P = llvm::FCmpInst::FCMP_OEQ; break;
10121 case NEON::BI__builtin_neon_vcleh_f16: P = llvm::FCmpInst::FCMP_OLE; break;
10122 case NEON::BI__builtin_neon_vclth_f16: P = llvm::FCmpInst::FCMP_OLT; break;
10123 case NEON::BI__builtin_neon_vcgeh_f16: P = llvm::FCmpInst::FCMP_OGE; break;
10124 case NEON::BI__builtin_neon_vcgth_f16: P = llvm::FCmpInst::FCMP_OGT; break;
10125 }
10126 Ops.push_back(EmitScalarExpr(E->getArg(1)));
10127 Ops[0] = Builder.CreateBitCast(Ops[0], HalfTy);
10128 Ops[1] = Builder.CreateBitCast(Ops[1], HalfTy);
10129 Ops[0] = Builder.CreateFCmp(P, Ops[0], Ops[1]);
10130 return Builder.CreateSExt(Ops[0], Int16Ty, "vcmpd");
10131 }
10132 case NEON::BI__builtin_neon_vceqd_s64:
10133 case NEON::BI__builtin_neon_vceqd_u64:
10134 case NEON::BI__builtin_neon_vcgtd_s64:
10135 case NEON::BI__builtin_neon_vcgtd_u64:
10136 case NEON::BI__builtin_neon_vcltd_s64:
10137 case NEON::BI__builtin_neon_vcltd_u64:
10138 case NEON::BI__builtin_neon_vcged_u64:
10139 case NEON::BI__builtin_neon_vcged_s64:
10140 case NEON::BI__builtin_neon_vcled_u64:
10141 case NEON::BI__builtin_neon_vcled_s64: {
10142 llvm::CmpInst::Predicate P;
10143 switch (BuiltinID) {
10144 default: llvm_unreachable("missing builtin ID in switch!")__builtin_unreachable();
10145 case NEON::BI__builtin_neon_vceqd_s64:
10146 case NEON::BI__builtin_neon_vceqd_u64:P = llvm::ICmpInst::ICMP_EQ;break;
10147 case NEON::BI__builtin_neon_vcgtd_s64:P = llvm::ICmpInst::ICMP_SGT;break;
10148 case NEON::BI__builtin_neon_vcgtd_u64:P = llvm::ICmpInst::ICMP_UGT;break;
10149 case NEON::BI__builtin_neon_vcltd_s64:P = llvm::ICmpInst::ICMP_SLT;break;
10150 case NEON::BI__builtin_neon_vcltd_u64:P = llvm::ICmpInst::ICMP_ULT;break;
10151 case NEON::BI__builtin_neon_vcged_u64:P = llvm::ICmpInst::ICMP_UGE;break;
10152 case NEON::BI__builtin_neon_vcged_s64:P = llvm::ICmpInst::ICMP_SGE;break;
10153 case NEON::BI__builtin_neon_vcled_u64:P = llvm::ICmpInst::ICMP_ULE;break;
10154 case NEON::BI__builtin_neon_vcled_s64:P = llvm::ICmpInst::ICMP_SLE;break;
10155 }
10156 Ops.push_back(EmitScalarExpr(E->getArg(1)));
10157 Ops[0] = Builder.CreateBitCast(Ops[0], Int64Ty);
10158 Ops[1] = Builder.CreateBitCast(Ops[1], Int64Ty);
10159 Ops[0] = Builder.CreateICmp(P, Ops[0], Ops[1]);
10160 return Builder.CreateSExt(Ops[0], Int64Ty, "vceqd");
10161 }
10162 case NEON::BI__builtin_neon_vtstd_s64:
10163 case NEON::BI__builtin_neon_vtstd_u64: {
10164 Ops.push_back(EmitScalarExpr(E->getArg(1)));
10165 Ops[0] = Builder.CreateBitCast(Ops[0], Int64Ty);
10166 Ops[1] = Builder.CreateBitCast(Ops[1], Int64Ty);
10167 Ops[0] = Builder.CreateAnd(Ops[0], Ops[1]);
10168 Ops[0] = Builder.CreateICmp(ICmpInst::ICMP_NE, Ops[0],
10169 llvm::Constant::getNullValue(Int64Ty));
10170 return Builder.CreateSExt(Ops[0], Int64Ty, "vtstd");
10171 }
10172 case NEON::BI__builtin_neon_vset_lane_i8:
10173 case NEON::BI__builtin_neon_vset_lane_i16:
10174 case NEON::BI__builtin_neon_vset_lane_i32:
10175 case NEON::BI__builtin_neon_vset_lane_i64:
10176 case NEON::BI__builtin_neon_vset_lane_bf16:
10177 case NEON::BI__builtin_neon_vset_lane_f32:
10178 case NEON::BI__builtin_neon_vsetq_lane_i8:
10179 case NEON::BI__builtin_neon_vsetq_lane_i16:
10180 case NEON::BI__builtin_neon_vsetq_lane_i32:
10181 case NEON::BI__builtin_neon_vsetq_lane_i64:
10182 case NEON::BI__builtin_neon_vsetq_lane_bf16:
10183 case NEON::BI__builtin_neon_vsetq_lane_f32:
10184 Ops.push_back(EmitScalarExpr(E->getArg(2)));
10185 return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vset_lane");
10186 case NEON::BI__builtin_neon_vset_lane_f64:
10187 // The vector type needs a cast for the v1f64 variant.
10188 Ops[1] =
10189 Builder.CreateBitCast(Ops[1], llvm::FixedVectorType::get(DoubleTy, 1));
10190 Ops.push_back(EmitScalarExpr(E->getArg(2)));
10191 return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vset_lane");
10192 case NEON::BI__builtin_neon_vsetq_lane_f64:
10193 // The vector type needs a cast for the v2f64 variant.
10194 Ops[1] =
10195 Builder.CreateBitCast(Ops[1], llvm::FixedVectorType::get(DoubleTy, 2));
10196 Ops.push_back(EmitScalarExpr(E->getArg(2)));
10197 return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vset_lane");
10198
10199 case NEON::BI__builtin_neon_vget_lane_i8:
10200 case NEON::BI__builtin_neon_vdupb_lane_i8:
10201 Ops[0] =
10202 Builder.CreateBitCast(Ops[0], llvm::FixedVectorType::get(Int8Ty, 8));
10203 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)),
10204 "vget_lane");
10205 case NEON::BI__builtin_neon_vgetq_lane_i8:
10206 case NEON::BI__builtin_neon_vdupb_laneq_i8:
10207 Ops[0] =
10208 Builder.CreateBitCast(Ops[0], llvm::FixedVectorType::get(Int8Ty, 16));
10209 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)),
10210 "vgetq_lane");
10211 case NEON::BI__builtin_neon_vget_lane_i16:
10212 case NEON::BI__builtin_neon_vduph_lane_i16:
10213 Ops[0] =
10214 Builder.CreateBitCast(Ops[0], llvm::FixedVectorType::get(Int16Ty, 4));
10215 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)),
10216 "vget_lane");
10217 case NEON::BI__builtin_neon_vgetq_lane_i16:
10218 case NEON::BI__builtin_neon_vduph_laneq_i16:
10219 Ops[0] =
10220 Builder.CreateBitCast(Ops[0], llvm::FixedVectorType::get(Int16Ty, 8));
10221 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)),
10222 "vgetq_lane");
10223 case NEON::BI__builtin_neon_vget_lane_i32:
10224 case NEON::BI__builtin_neon_vdups_lane_i32:
10225 Ops[0] =
10226 Builder.CreateBitCast(Ops[0], llvm::FixedVectorType::get(Int32Ty, 2));
10227 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)),
10228 "vget_lane");
10229 case NEON::BI__builtin_neon_vdups_lane_f32:
10230 Ops[0] =
10231 Builder.CreateBitCast(Ops[0], llvm::FixedVectorType::get(FloatTy, 2));
10232 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)),
10233 "vdups_lane");
10234 case NEON::BI__builtin_neon_vgetq_lane_i32:
10235 case NEON::BI__builtin_neon_vdups_laneq_i32:
10236 Ops[0] =
10237 Builder.CreateBitCast(Ops[0], llvm::FixedVectorType::get(Int32Ty, 4));
10238 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)),
10239 "vgetq_lane");
10240 case NEON::BI__builtin_neon_vget_lane_i64:
10241 case NEON::BI__builtin_neon_vdupd_lane_i64:
10242 Ops[0] =
10243 Builder.CreateBitCast(Ops[0], llvm::FixedVectorType::get(Int64Ty, 1));
10244 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)),
10245 "vget_lane");
10246 case NEON::BI__builtin_neon_vdupd_lane_f64:
10247 Ops[0] =
10248 Builder.CreateBitCast(Ops[0], llvm::FixedVectorType::get(DoubleTy, 1));
10249 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)),
10250 "vdupd_lane");
10251 case NEON::BI__builtin_neon_vgetq_lane_i64:
10252 case NEON::BI__builtin_neon_vdupd_laneq_i64:
10253 Ops[0] =
10254 Builder.CreateBitCast(Ops[0], llvm::FixedVectorType::get(Int64Ty, 2));
10255 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)),
10256 "vgetq_lane");
10257 case NEON::BI__builtin_neon_vget_lane_f32:
10258 Ops[0] =
10259 Builder.CreateBitCast(Ops[0], llvm::FixedVectorType::get(FloatTy, 2));
10260 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)),
10261 "vget_lane");
10262 case NEON::BI__builtin_neon_vget_lane_f64:
10263 Ops[0] =
10264 Builder.CreateBitCast(Ops[0], llvm::FixedVectorType::get(DoubleTy, 1));
10265 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)),
10266 "vget_lane");
10267 case NEON::BI__builtin_neon_vgetq_lane_f32:
10268 case NEON::BI__builtin_neon_vdups_laneq_f32:
10269 Ops[0] =
10270 Builder.CreateBitCast(Ops[0], llvm::FixedVectorType::get(FloatTy, 4));
10271 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)),
10272 "vgetq_lane");
10273 case NEON::BI__builtin_neon_vgetq_lane_f64:
10274 case NEON::BI__builtin_neon_vdupd_laneq_f64:
10275 Ops[0] =
10276 Builder.CreateBitCast(Ops[0], llvm::FixedVectorType::get(DoubleTy, 2));
10277 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)),
10278 "vgetq_lane");
10279 case NEON::BI__builtin_neon_vaddh_f16:
10280 Ops.push_back(EmitScalarExpr(E->getArg(1)));
10281 return Builder.CreateFAdd(Ops[0], Ops[1], "vaddh");
10282 case NEON::BI__builtin_neon_vsubh_f16:
10283 Ops.push_back(EmitScalarExpr(E->getArg(1)));
10284 return Builder.CreateFSub(Ops[0], Ops[1], "vsubh");
10285 case NEON::BI__builtin_neon_vmulh_f16:
10286 Ops.push_back(EmitScalarExpr(E->getArg(1)));
10287 return Builder.CreateFMul(Ops[0], Ops[1], "vmulh");
10288 case NEON::BI__builtin_neon_vdivh_f16:
10289 Ops.push_back(EmitScalarExpr(E->getArg(1)));
10290 return Builder.CreateFDiv(Ops[0], Ops[1], "vdivh");
10291 case NEON::BI__builtin_neon_vfmah_f16:
10292 // NEON intrinsic puts accumulator first, unlike the LLVM fma.
10293 return emitCallMaybeConstrainedFPBuiltin(
10294 *this, Intrinsic::fma, Intrinsic::experimental_constrained_fma, HalfTy,
10295 {EmitScalarExpr(E->getArg(1)), EmitScalarExpr(E->getArg(2)), Ops[0]});
10296 case NEON::BI__builtin_neon_vfmsh_f16: {
10297 // FIXME: This should be an fneg instruction:
10298 Value *Zero = llvm::ConstantFP::getZeroValueForNegation(HalfTy);
10299 Value* Sub = Builder.CreateFSub(Zero, EmitScalarExpr(E->getArg(1)), "vsubh");
10300
10301 // NEON intrinsic puts accumulator first, unlike the LLVM fma.
10302 return emitCallMaybeConstrainedFPBuiltin(
10303 *this, Intrinsic::fma, Intrinsic::experimental_constrained_fma, HalfTy,
10304 {Sub, EmitScalarExpr(E->getArg(2)), Ops[0]});
10305 }
10306 case NEON::BI__builtin_neon_vaddd_s64:
10307 case NEON::BI__builtin_neon_vaddd_u64:
10308 return Builder.CreateAdd(Ops[0], EmitScalarExpr(E->getArg(1)), "vaddd");
10309 case NEON::BI__builtin_neon_vsubd_s64:
10310 case NEON::BI__builtin_neon_vsubd_u64:
10311 return Builder.CreateSub(Ops[0], EmitScalarExpr(E->getArg(1)), "vsubd");
10312 case NEON::BI__builtin_neon_vqdmlalh_s16:
10313 case NEON::BI__builtin_neon_vqdmlslh_s16: {
10314 SmallVector<Value *, 2> ProductOps;
10315 ProductOps.push_back(vectorWrapScalar16(Ops[1]));
10316 ProductOps.push_back(vectorWrapScalar16(EmitScalarExpr(E->getArg(2))));
10317 auto *VTy = llvm::FixedVectorType::get(Int32Ty, 4);
10318 Ops[1] = EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqdmull, VTy),
10319 ProductOps, "vqdmlXl");
10320 Constant *CI = ConstantInt::get(SizeTy, 0);
10321 Ops[1] = Builder.CreateExtractElement(Ops[1], CI, "lane0");
10322
10323 unsigned AccumInt = BuiltinID == NEON::BI__builtin_neon_vqdmlalh_s16
10324 ? Intrinsic::aarch64_neon_sqadd
10325 : Intrinsic::aarch64_neon_sqsub;
10326 return EmitNeonCall(CGM.getIntrinsic(AccumInt, Int32Ty), Ops, "vqdmlXl");
10327 }
10328 case NEON::BI__builtin_neon_vqshlud_n_s64: {
10329 Ops.push_back(EmitScalarExpr(E->getArg(1)));
10330 Ops[1] = Builder.CreateZExt(Ops[1], Int64Ty);
10331 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqshlu, Int64Ty),
10332 Ops, "vqshlu_n");
10333 }
10334 case NEON::BI__builtin_neon_vqshld_n_u64:
10335 case NEON::BI__builtin_neon_vqshld_n_s64: {
10336 unsigned Int = BuiltinID == NEON::BI__builtin_neon_vqshld_n_u64
10337 ? Intrinsic::aarch64_neon_uqshl
10338 : Intrinsic::aarch64_neon_sqshl;
10339 Ops.push_back(EmitScalarExpr(E->getArg(1)));
10340 Ops[1] = Builder.CreateZExt(Ops[1], Int64Ty);
10341 return EmitNeonCall(CGM.getIntrinsic(Int, Int64Ty), Ops, "vqshl_n");
10342 }
10343 case NEON::BI__builtin_neon_vrshrd_n_u64:
10344 case NEON::BI__builtin_neon_vrshrd_n_s64: {
10345 unsigned Int = BuiltinID == NEON::BI__builtin_neon_vrshrd_n_u64
10346 ? Intrinsic::aarch64_neon_urshl
10347 : Intrinsic::aarch64_neon_srshl;
10348 Ops.push_back(EmitScalarExpr(E->getArg(1)));
10349 int SV = cast<ConstantInt>(Ops[1])->getSExtValue();
10350 Ops[1] = ConstantInt::get(Int64Ty, -SV);
10351 return EmitNeonCall(CGM.getIntrinsic(Int, Int64Ty), Ops, "vrshr_n");
10352 }
10353 case NEON::BI__builtin_neon_vrsrad_n_u64:
10354 case NEON::BI__builtin_neon_vrsrad_n_s64: {
10355 unsigned Int = BuiltinID == NEON::BI__builtin_neon_vrsrad_n_u64
10356 ? Intrinsic::aarch64_neon_urshl
10357 : Intrinsic::aarch64_neon_srshl;
10358 Ops[1] = Builder.CreateBitCast(Ops[1], Int64Ty);
10359 Ops.push_back(Builder.CreateNeg(EmitScalarExpr(E->getArg(2))));
10360 Ops[1] = Builder.CreateCall(CGM.getIntrinsic(Int, Int64Ty),
10361 {Ops[1], Builder.CreateSExt(Ops[2], Int64Ty)});
10362 return Builder.CreateAdd(Ops[0], Builder.CreateBitCast(Ops[1], Int64Ty));
10363 }
10364 case NEON::BI__builtin_neon_vshld_n_s64:
10365 case NEON::BI__builtin_neon_vshld_n_u64: {
10366 llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(1)));
10367 return Builder.CreateShl(
10368 Ops[0], ConstantInt::get(Int64Ty, Amt->getZExtValue()), "shld_n");
10369 }
10370 case NEON::BI__builtin_neon_vshrd_n_s64: {
10371 llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(1)));
10372 return Builder.CreateAShr(
10373 Ops[0], ConstantInt::get(Int64Ty, std::min(static_cast<uint64_t>(63),
10374 Amt->getZExtValue())),
10375 "shrd_n");
10376 }
10377 case NEON::BI__builtin_neon_vshrd_n_u64: {
10378 llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(1)));
10379 uint64_t ShiftAmt = Amt->getZExtValue();
10380 // Right-shifting an unsigned value by its size yields 0.
10381 if (ShiftAmt == 64)
10382 return ConstantInt::get(Int64Ty, 0);
10383 return Builder.CreateLShr(Ops[0], ConstantInt::get(Int64Ty, ShiftAmt),
10384 "shrd_n");
10385 }
10386 case NEON::BI__builtin_neon_vsrad_n_s64: {
10387 llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(2)));
10388 Ops[1] = Builder.CreateAShr(
10389 Ops[1], ConstantInt::get(Int64Ty, std::min(static_cast<uint64_t>(63),
10390 Amt->getZExtValue())),
10391 "shrd_n");
10392 return Builder.CreateAdd(Ops[0], Ops[1]);
10393 }
10394 case NEON::BI__builtin_neon_vsrad_n_u64: {
10395 llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(2)));
10396 uint64_t ShiftAmt = Amt->getZExtValue();
10397 // Right-shifting an unsigned value by its size yields 0.
10398 // As Op + 0 = Op, return Ops[0] directly.
10399 if (ShiftAmt == 64)
10400 return Ops[0];
10401 Ops[1] = Builder.CreateLShr(Ops[1], ConstantInt::get(Int64Ty, ShiftAmt),
10402 "shrd_n");
10403 return Builder.CreateAdd(Ops[0], Ops[1]);
10404 }
10405 case NEON::BI__builtin_neon_vqdmlalh_lane_s16:
10406 case NEON::BI__builtin_neon_vqdmlalh_laneq_s16:
10407 case NEON::BI__builtin_neon_vqdmlslh_lane_s16:
10408 case NEON::BI__builtin_neon_vqdmlslh_laneq_s16: {
10409 Ops[2] = Builder.CreateExtractElement(Ops[2], EmitScalarExpr(E->getArg(3)),
10410 "lane");
10411 SmallVector<Value *, 2> ProductOps;
10412 ProductOps.push_back(vectorWrapScalar16(Ops[1]));
10413 ProductOps.push_back(vectorWrapScalar16(Ops[2]));
10414 auto *VTy = llvm::FixedVectorType::get(Int32Ty, 4);
10415 Ops[1] = EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqdmull, VTy),
10416 ProductOps, "vqdmlXl");
10417 Constant *CI = ConstantInt::get(SizeTy, 0);
10418 Ops[1] = Builder.CreateExtractElement(Ops[1], CI, "lane0");
10419 Ops.pop_back();
10420
10421 unsigned AccInt = (BuiltinID == NEON::BI__builtin_neon_vqdmlalh_lane_s16 ||
10422 BuiltinID == NEON::BI__builtin_neon_vqdmlalh_laneq_s16)
10423 ? Intrinsic::aarch64_neon_sqadd
10424 : Intrinsic::aarch64_neon_sqsub;
10425 return EmitNeonCall(CGM.getIntrinsic(AccInt, Int32Ty), Ops, "vqdmlXl");
10426 }
10427 case NEON::BI__builtin_neon_vqdmlals_s32:
10428 case NEON::BI__builtin_neon_vqdmlsls_s32: {
10429 SmallVector<Value *, 2> ProductOps;
10430 ProductOps.push_back(Ops[1]);
10431 ProductOps.push_back(EmitScalarExpr(E->getArg(2)));
10432 Ops[1] =
10433 EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqdmulls_scalar),
10434 ProductOps, "vqdmlXl");
10435
10436 unsigned AccumInt = BuiltinID == NEON::BI__builtin_neon_vqdmlals_s32
10437 ? Intrinsic::aarch64_neon_sqadd
10438 : Intrinsic::aarch64_neon_sqsub;
10439 return EmitNeonCall(CGM.getIntrinsic(AccumInt, Int64Ty), Ops, "vqdmlXl");
10440 }
10441 case NEON::BI__builtin_neon_vqdmlals_lane_s32:
10442 case NEON::BI__builtin_neon_vqdmlals_laneq_s32:
10443 case NEON::BI__builtin_neon_vqdmlsls_lane_s32:
10444 case NEON::BI__builtin_neon_vqdmlsls_laneq_s32: {
10445 Ops[2] = Builder.CreateExtractElement(Ops[2], EmitScalarExpr(E->getArg(3)),
10446 "lane");
10447 SmallVector<Value *, 2> ProductOps;
10448 ProductOps.push_back(Ops[1]);
10449 ProductOps.push_back(Ops[2]);
10450 Ops[1] =
10451 EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqdmulls_scalar),
10452 ProductOps, "vqdmlXl");
10453 Ops.pop_back();
10454
10455 unsigned AccInt = (BuiltinID == NEON::BI__builtin_neon_vqdmlals_lane_s32 ||
10456 BuiltinID == NEON::BI__builtin_neon_vqdmlals_laneq_s32)
10457 ? Intrinsic::aarch64_neon_sqadd
10458 : Intrinsic::aarch64_neon_sqsub;
10459 return EmitNeonCall(CGM.getIntrinsic(AccInt, Int64Ty), Ops, "vqdmlXl");
10460 }
10461 case NEON::BI__builtin_neon_vget_lane_bf16:
10462 case NEON::BI__builtin_neon_vduph_lane_bf16:
10463 case NEON::BI__builtin_neon_vduph_lane_f16: {
10464 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)),
10465 "vget_lane");
10466 }
10467 case NEON::BI__builtin_neon_vgetq_lane_bf16:
10468 case NEON::BI__builtin_neon_vduph_laneq_bf16:
10469 case NEON::BI__builtin_neon_vduph_laneq_f16: {
10470 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)),
10471 "vgetq_lane");
10472 }
10473
10474 case AArch64::BI_InterlockedAdd: {
10475 Value *Arg0 = EmitScalarExpr(E->getArg(0));
10476 Value *Arg1 = EmitScalarExpr(E->getArg(1));
10477 AtomicRMWInst *RMWI = Builder.CreateAtomicRMW(
10478 AtomicRMWInst::Add, Arg0, Arg1,
10479 llvm::AtomicOrdering::SequentiallyConsistent);
10480 return Builder.CreateAdd(RMWI, Arg1);
10481 }
10482 }
10483
10484 llvm::FixedVectorType *VTy = GetNeonType(this, Type);
10485 llvm::Type *Ty = VTy;
10486 if (!Ty)
10487 return nullptr;
10488
10489 // Not all intrinsics handled by the common case work for AArch64 yet, so only
10490 // defer to common code if it's been added to our special map.
10491 Builtin = findARMVectorIntrinsicInMap(AArch64SIMDIntrinsicMap, BuiltinID,
10492 AArch64SIMDIntrinsicsProvenSorted);
10493
10494 if (Builtin)
10495 return EmitCommonNeonBuiltinExpr(
10496 Builtin->BuiltinID, Builtin->LLVMIntrinsic, Builtin->AltLLVMIntrinsic,
10497 Builtin->NameHint, Builtin->TypeModifier, E, Ops,
10498 /*never use addresses*/ Address::invalid(), Address::invalid(), Arch);
10499
10500 if (Value *V = EmitAArch64TblBuiltinExpr(*this, BuiltinID, E, Ops, Arch))
10501 return V;
10502
10503 unsigned Int;
10504 switch (BuiltinID) {
10505 default: return nullptr;
10506 case NEON::BI__builtin_neon_vbsl_v:
10507 case NEON::BI__builtin_neon_vbslq_v: {
10508 llvm::Type *BitTy = llvm::VectorType::getInteger(VTy);
10509 Ops[0] = Builder.CreateBitCast(Ops[0], BitTy, "vbsl");
10510 Ops[1] = Builder.CreateBitCast(Ops[1], BitTy, "vbsl");
10511 Ops[2] = Builder.CreateBitCast(Ops[2], BitTy, "vbsl");
10512
10513 Ops[1] = Builder.CreateAnd(Ops[0], Ops[1], "vbsl");
10514 Ops[2] = Builder.CreateAnd(Builder.CreateNot(Ops[0]), Ops[2], "vbsl");
10515 Ops[0] = Builder.CreateOr(Ops[1], Ops[2], "vbsl");
10516 return Builder.CreateBitCast(Ops[0], Ty);
10517 }
10518 case NEON::BI__builtin_neon_vfma_lane_v:
10519 case NEON::BI__builtin_neon_vfmaq_lane_v: { // Only used for FP types
10520 // The ARM builtins (and instructions) have the addend as the first
10521 // operand, but the 'fma' intrinsics have it last. Swap it around here.
10522 Value *Addend = Ops[0];
10523 Value *Multiplicand = Ops[1];
10524 Value *LaneSource = Ops[2];
10525 Ops[0] = Multiplicand;
10526 Ops[1] = LaneSource;
10527 Ops[2] = Addend;
10528
10529 // Now adjust things to handle the lane access.
10530 auto *SourceTy = BuiltinID == NEON::BI__builtin_neon_vfmaq_lane_v
10531 ? llvm::FixedVectorType::get(VTy->getElementType(),
10532 VTy->getNumElements() / 2)
10533 : VTy;
10534 llvm::Constant *cst = cast<Constant>(Ops[3]);
10535 Value *SV = llvm::ConstantVector::getSplat(VTy->getElementCount(), cst);
10536 Ops[1] = Builder.CreateBitCast(Ops[1], SourceTy);
10537 Ops[1] = Builder.CreateShuffleVector(Ops[1], Ops[1], SV, "lane");
10538
10539 Ops.pop_back();
10540 Int = Builder.getIsFPConstrained() ? Intrinsic::experimental_constrained_fma
10541 : Intrinsic::fma;
10542 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "fmla");
10543 }
10544 case NEON::BI__builtin_neon_vfma_laneq_v: {
10545 auto *VTy = cast<llvm::FixedVectorType>(Ty);
10546 // v1f64 fma should be mapped to Neon scalar f64 fma
10547 if (VTy && VTy->getElementType() == DoubleTy) {
10548 Ops[0] = Builder.CreateBitCast(Ops[0], DoubleTy);
10549 Ops[1] = Builder.CreateBitCast(Ops[1], DoubleTy);
10550 llvm::FixedVectorType *VTy =
10551 GetNeonType(this, NeonTypeFlags(NeonTypeFlags::Float64, false, true));
10552 Ops[2] = Builder.CreateBitCast(Ops[2], VTy);
10553 Ops[2] = Builder.CreateExtractElement(Ops[2], Ops[3], "extract");
10554 Value *Result;
10555 Result = emitCallMaybeConstrainedFPBuiltin(
10556 *this, Intrinsic::fma, Intrinsic::experimental_constrained_fma,
10557 DoubleTy, {Ops[1], Ops[2], Ops[0]});
10558 return Builder.CreateBitCast(Result, Ty);
10559 }
10560 Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
10561 Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
10562
10563 auto *STy = llvm::FixedVectorType::get(VTy->getElementType(),
10564 VTy->getNumElements() * 2);
10565 Ops[2] = Builder.CreateBitCast(Ops[2], STy);
10566 Value *SV = llvm::ConstantVector::getSplat(VTy->getElementCount(),
10567 cast<ConstantInt>(Ops[3]));
10568 Ops[2] = Builder.CreateShuffleVector(Ops[2], Ops[2], SV, "lane");
10569
10570 return emitCallMaybeConstrainedFPBuiltin(
10571 *this, Intrinsic::fma, Intrinsic::experimental_constrained_fma, Ty,
10572 {Ops[2], Ops[1], Ops[0]});
10573 }
10574 case NEON::BI__builtin_neon_vfmaq_laneq_v: {
10575 Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
10576 Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
10577
10578 Ops[2] = Builder.CreateBitCast(Ops[2], Ty);
10579 Ops[2] = EmitNeonSplat(Ops[2], cast<ConstantInt>(Ops[3]));
10580 return emitCallMaybeConstrainedFPBuiltin(
10581 *this, Intrinsic::fma, Intrinsic::experimental_constrained_fma, Ty,
10582 {Ops[2], Ops[1], Ops[0]});
10583 }
10584 case NEON::BI__builtin_neon_vfmah_lane_f16:
10585 case NEON::BI__builtin_neon_vfmas_lane_f32:
10586 case NEON::BI__builtin_neon_vfmah_laneq_f16:
10587 case NEON::BI__builtin_neon_vfmas_laneq_f32:
10588 case NEON::BI__builtin_neon_vfmad_lane_f64:
10589 case NEON::BI__builtin_neon_vfmad_laneq_f64: {
10590 Ops.push_back(EmitScalarExpr(E->getArg(3)));
10591 llvm::Type *Ty = ConvertType(E->getCallReturnType(getContext()));
10592 Ops[2] = Builder.CreateExtractElement(Ops[2], Ops[3], "extract");
10593 return emitCallMaybeConstrainedFPBuiltin(
10594 *this, Intrinsic::fma, Intrinsic::experimental_constrained_fma, Ty,
10595 {Ops[1], Ops[2], Ops[0]});
10596 }
10597 case NEON::BI__builtin_neon_vmull_v:
10598 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics.
10599 Int = usgn ? Intrinsic::aarch64_neon_umull : Intrinsic::aarch64_neon_smull;
10600 if (Type.isPoly()) Int = Intrinsic::aarch64_neon_pmull;
10601 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmull");
10602 case NEON::BI__builtin_neon_vmax_v:
10603 case NEON::BI__builtin_neon_vmaxq_v:
10604 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics.
10605 Int = usgn ? Intrinsic::aarch64_neon_umax : Intrinsic::aarch64_neon_smax;
10606 if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fmax;
10607 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmax");
10608 case NEON::BI__builtin_neon_vmaxh_f16: {
10609 Ops.push_back(EmitScalarExpr(E->getArg(1)));
10610 Int = Intrinsic::aarch64_neon_fmax;
10611 return EmitNeonCall(CGM.getIntrinsic(Int, HalfTy), Ops, "vmax");
10612 }
10613 case NEON::BI__builtin_neon_vmin_v:
10614 case NEON::BI__builtin_neon_vminq_v:
10615 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics.
10616 Int = usgn ? Intrinsic::aarch64_neon_umin : Intrinsic::aarch64_neon_smin;
10617 if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fmin;
10618 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmin");
10619 case NEON::BI__builtin_neon_vminh_f16: {
10620 Ops.push_back(EmitScalarExpr(E->getArg(1)));
10621 Int = Intrinsic::aarch64_neon_fmin;
10622 return EmitNeonCall(CGM.getIntrinsic(Int, HalfTy), Ops, "vmin");
10623 }
10624 case NEON::BI__builtin_neon_vabd_v:
10625 case NEON::BI__builtin_neon_vabdq_v:
10626 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics.
10627 Int = usgn ? Intrinsic::aarch64_neon_uabd : Intrinsic::aarch64_neon_sabd;
10628 if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fabd;
10629 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vabd");
10630 case NEON::BI__builtin_neon_vpadal_v:
10631 case NEON::BI__builtin_neon_vpadalq_v: {
10632 unsigned ArgElts = VTy->getNumElements();
10633 llvm::IntegerType *EltTy = cast<IntegerType>(VTy->getElementType());
10634 unsigned BitWidth = EltTy->getBitWidth();
10635 auto *ArgTy = llvm::FixedVectorType::get(
10636 llvm::IntegerType::get(getLLVMContext(), BitWidth / 2), 2 * ArgElts);
10637 llvm::Type* Tys[2] = { VTy, ArgTy };
10638 Int = usgn ? Intrinsic::aarch64_neon_uaddlp : Intrinsic::aarch64_neon_saddlp;
10639 SmallVector<llvm::Value*, 1> TmpOps;
10640 TmpOps.push_back(Ops[1]);
10641 Function *F = CGM.getIntrinsic(Int, Tys);
10642 llvm::Value *tmp = EmitNeonCall(F, TmpOps, "vpadal");
10643 llvm::Value *addend = Builder.CreateBitCast(Ops[0], tmp->getType());
10644 return Builder.CreateAdd(tmp, addend);
10645 }
10646 case NEON::BI__builtin_neon_vpmin_v:
10647 case NEON::BI__builtin_neon_vpminq_v:
10648 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics.
10649 Int = usgn ? Intrinsic::aarch64_neon_uminp : Intrinsic::aarch64_neon_sminp;
10650 if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fminp;
10651 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpmin");
10652 case NEON::BI__builtin_neon_vpmax_v:
10653 case NEON::BI__builtin_neon_vpmaxq_v:
10654 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics.
10655 Int = usgn ? Intrinsic::aarch64_neon_umaxp : Intrinsic::aarch64_neon_smaxp;
10656 if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fmaxp;
10657 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpmax");
10658 case NEON::BI__builtin_neon_vminnm_v:
10659 case NEON::BI__builtin_neon_vminnmq_v:
10660 Int = Intrinsic::aarch64_neon_fminnm;
10661 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vminnm");
10662 case NEON::BI__builtin_neon_vminnmh_f16:
10663 Ops.push_back(EmitScalarExpr(E->getArg(1)));
10664 Int = Intrinsic::aarch64_neon_fminnm;
10665 return EmitNeonCall(CGM.getIntrinsic(Int, HalfTy), Ops, "vminnm");
10666 case NEON::BI__builtin_neon_vmaxnm_v:
10667 case NEON::BI__builtin_neon_vmaxnmq_v:
10668 Int = Intrinsic::aarch64_neon_fmaxnm;
10669 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmaxnm");
10670 case NEON::BI__builtin_neon_vmaxnmh_f16:
10671 Ops.push_back(EmitScalarExpr(E->getArg(1)));
10672 Int = Intrinsic::aarch64_neon_fmaxnm;
10673 return EmitNeonCall(CGM.getIntrinsic(Int, HalfTy), Ops, "vmaxnm");
10674 case NEON::BI__builtin_neon_vrecpss_f32: {
10675 Ops.push_back(EmitScalarExpr(E->getArg(1)));
10676 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_frecps, FloatTy),
10677 Ops, "vrecps");
10678 }
10679 case NEON::BI__builtin_neon_vrecpsd_f64:
10680 Ops.push_back(EmitScalarExpr(E->getArg(1)));
10681 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_frecps, DoubleTy),
10682 Ops, "vrecps");
10683 case NEON::BI__builtin_neon_vrecpsh_f16:
10684 Ops.push_back(EmitScalarExpr(E->getArg(1)));
10685 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_frecps, HalfTy),
10686 Ops, "vrecps");
10687 case NEON::BI__builtin_neon_vqshrun_n_v:
10688 Int = Intrinsic::aarch64_neon_sqshrun;
10689 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshrun_n");
10690 case NEON::BI__builtin_neon_vqrshrun_n_v:
10691 Int = Intrinsic::aarch64_neon_sqrshrun;
10692 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqrshrun_n");
10693 case NEON::BI__builtin_neon_vqshrn_n_v:
10694 Int = usgn ? Intrinsic::aarch64_neon_uqshrn : Intrinsic::aarch64_neon_sqshrn;
10695 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshrn_n");
10696 case NEON::BI__builtin_neon_vrshrn_n_v:
10697 Int = Intrinsic::aarch64_neon_rshrn;
10698 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrshrn_n");
10699 case NEON::BI__builtin_neon_vqrshrn_n_v:
10700 Int = usgn ? Intrinsic::aarch64_neon_uqrshrn : Intrinsic::aarch64_neon_sqrshrn;
10701 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqrshrn_n");
10702 case NEON::BI__builtin_neon_vrndah_f16: {
10703 Ops.push_back(EmitScalarExpr(E->getArg(0)));
10704 Int = Builder.getIsFPConstrained()
10705 ? Intrinsic::experimental_constrained_round
10706 : Intrinsic::round;
10707 return EmitNeonCall(CGM.getIntrinsic(Int, HalfTy), Ops, "vrnda");
10708 }
10709 case NEON::BI__builtin_neon_vrnda_v:
10710 case NEON::BI__builtin_neon_vrndaq_v: {
10711 Int = Builder.getIsFPConstrained()
10712 ? Intrinsic::experimental_constrained_round
10713 : Intrinsic::round;
10714 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrnda");
10715 }
10716 case NEON::BI__builtin_neon_vrndih_f16: {
10717 Ops.push_back(EmitScalarExpr(E->getArg(0)));
10718 Int = Builder.getIsFPConstrained()
10719 ? Intrinsic::experimental_constrained_nearbyint
10720 : Intrinsic::nearbyint;
10721 return EmitNeonCall(CGM.getIntrinsic(Int, HalfTy), Ops, "vrndi");
10722 }
10723 case NEON::BI__builtin_neon_vrndmh_f16: {
10724 Ops.push_back(EmitScalarExpr(E->getArg(0)));
10725 Int = Builder.getIsFPConstrained()
10726 ? Intrinsic::experimental_constrained_floor
10727 : Intrinsic::floor;
10728 return EmitNeonCall(CGM.getIntrinsic(Int, HalfTy), Ops, "vrndm");
10729 }
10730 case NEON::BI__builtin_neon_vrndm_v:
10731 case NEON::BI__builtin_neon_vrndmq_v: {
10732 Int = Builder.getIsFPConstrained()
10733 ? Intrinsic::experimental_constrained_floor
10734 : Intrinsic::floor;
10735 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndm");
10736 }
10737 case NEON::BI__builtin_neon_vrndnh_f16: {
10738 Ops.push_back(EmitScalarExpr(E->getArg(0)));
10739 Int = Builder.getIsFPConstrained()
10740 ? Intrinsic::experimental_constrained_roundeven
10741 : Intrinsic::roundeven;
10742 return EmitNeonCall(CGM.getIntrinsic(Int, HalfTy), Ops, "vrndn");
10743 }
10744 case NEON::BI__builtin_neon_vrndn_v:
10745 case NEON::BI__builtin_neon_vrndnq_v: {
10746 Int = Builder.getIsFPConstrained()
10747 ? Intrinsic::experimental_constrained_roundeven
10748 : Intrinsic::roundeven;
10749 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndn");
10750 }
10751 case NEON::BI__builtin_neon_vrndns_f32: {
10752 Ops.push_back(EmitScalarExpr(E->getArg(0)));
10753 Int = Builder.getIsFPConstrained()
10754 ? Intrinsic::experimental_constrained_roundeven
10755 : Intrinsic::roundeven;
10756 return EmitNeonCall(CGM.getIntrinsic(Int, FloatTy), Ops, "vrndn");
10757 }
10758 case NEON::BI__builtin_neon_vrndph_f16: {
10759 Ops.push_back(EmitScalarExpr(E->getArg(0)));
10760 Int = Builder.getIsFPConstrained()
10761 ? Intrinsic::experimental_constrained_ceil
10762 : Intrinsic::ceil;
10763 return EmitNeonCall(CGM.getIntrinsic(Int, HalfTy), Ops, "vrndp");
10764 }
10765 case NEON::BI__builtin_neon_vrndp_v:
10766 case NEON::BI__builtin_neon_vrndpq_v: {
10767 Int = Builder.getIsFPConstrained()
10768 ? Intrinsic::experimental_constrained_ceil
10769 : Intrinsic::ceil;
10770 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndp");
10771 }
10772 case NEON::BI__builtin_neon_vrndxh_f16: {
10773 Ops.push_back(EmitScalarExpr(E->getArg(0)));
10774 Int = Builder.getIsFPConstrained()
10775 ? Intrinsic::experimental_constrained_rint
10776 : Intrinsic::rint;
10777 return EmitNeonCall(CGM.getIntrinsic(Int, HalfTy), Ops, "vrndx");
10778 }
10779 case NEON::BI__builtin_neon_vrndx_v:
10780 case NEON::BI__builtin_neon_vrndxq_v: {
10781 Int = Builder.getIsFPConstrained()
10782 ? Intrinsic::experimental_constrained_rint
10783 : Intrinsic::rint;
10784 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndx");
10785 }
10786 case NEON::BI__builtin_neon_vrndh_f16: {
10787 Ops.push_back(EmitScalarExpr(E->getArg(0)));
10788 Int = Builder.getIsFPConstrained()
10789 ? Intrinsic::experimental_constrained_trunc
10790 : Intrinsic::trunc;
10791 return EmitNeonCall(CGM.getIntrinsic(Int, HalfTy), Ops, "vrndz");
10792 }
10793 case NEON::BI__builtin_neon_vrnd32x_v:
10794 case NEON::BI__builtin_neon_vrnd32xq_v: {
10795 Ops.push_back(EmitScalarExpr(E->getArg(0)));
10796 Int = Intrinsic::aarch64_neon_frint32x;
10797 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrnd32x");
10798 }
10799 case NEON::BI__builtin_neon_vrnd32z_v:
10800 case NEON::BI__builtin_neon_vrnd32zq_v: {
10801 Ops.push_back(EmitScalarExpr(E->getArg(0)));
10802 Int = Intrinsic::aarch64_neon_frint32z;
10803 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrnd32z");
10804 }
10805 case NEON::BI__builtin_neon_vrnd64x_v:
10806 case NEON::BI__builtin_neon_vrnd64xq_v: {
10807 Ops.push_back(EmitScalarExpr(E->getArg(0)));
10808 Int = Intrinsic::aarch64_neon_frint64x;
10809 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrnd64x");
10810 }
10811 case NEON::BI__builtin_neon_vrnd64z_v:
10812 case NEON::BI__builtin_neon_vrnd64zq_v: {
10813 Ops.push_back(EmitScalarExpr(E->getArg(0)));
10814 Int = Intrinsic::aarch64_neon_frint64z;
10815 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrnd64z");
10816 }
10817 case NEON::BI__builtin_neon_vrnd_v:
10818 case NEON::BI__builtin_neon_vrndq_v: {
10819 Int = Builder.getIsFPConstrained()
10820 ? Intrinsic::experimental_constrained_trunc
10821 : Intrinsic::trunc;
10822 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndz");
10823 }
10824 case NEON::BI__builtin_neon_vcvt_f64_v:
10825 case NEON::BI__builtin_neon_vcvtq_f64_v:
10826 Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
10827 Ty = GetNeonType(this, NeonTypeFlags(NeonTypeFlags::Float64, false, quad));
10828 return usgn ? Builder.CreateUIToFP(Ops[0], Ty, "vcvt")
10829 : Builder.CreateSIToFP(Ops[0], Ty, "vcvt");
10830 case NEON::BI__builtin_neon_vcvt_f64_f32: {
10831 assert(Type.getEltType() == NeonTypeFlags::Float64 && quad &&((void)0)
10832 "unexpected vcvt_f64_f32 builtin")((void)0);
10833 NeonTypeFlags SrcFlag = NeonTypeFlags(NeonTypeFlags::Float32, false, false);
10834 Ops[0] = Builder.CreateBitCast(Ops[0], GetNeonType(this, SrcFlag));
10835
10836 return Builder.CreateFPExt(Ops[0], Ty, "vcvt");
10837 }
10838 case NEON::BI__builtin_neon_vcvt_f32_f64: {
10839 assert(Type.getEltType() == NeonTypeFlags::Float32 &&((void)0)
10840 "unexpected vcvt_f32_f64 builtin")((void)0);
10841 NeonTypeFlags SrcFlag = NeonTypeFlags(NeonTypeFlags::Float64, false, true);
10842 Ops[0] = Builder.CreateBitCast(Ops[0], GetNeonType(this, SrcFlag));
10843
10844 return Builder.CreateFPTrunc(Ops[0], Ty, "vcvt");
10845 }
10846 case NEON::BI__builtin_neon_vcvt_s32_v:
10847 case NEON::BI__builtin_neon_vcvt_u32_v:
10848 case NEON::BI__builtin_neon_vcvt_s64_v:
10849 case NEON::BI__builtin_neon_vcvt_u64_v:
10850 case NEON::BI__builtin_neon_vcvt_s16_v:
10851 case NEON::BI__builtin_neon_vcvt_u16_v:
10852 case NEON::BI__builtin_neon_vcvtq_s32_v:
10853 case NEON::BI__builtin_neon_vcvtq_u32_v:
10854 case NEON::BI__builtin_neon_vcvtq_s64_v:
10855 case NEON::BI__builtin_neon_vcvtq_u64_v:
10856 case NEON::BI__builtin_neon_vcvtq_s16_v:
10857 case NEON::BI__builtin_neon_vcvtq_u16_v: {
10858 Int =
10859 usgn ? Intrinsic::aarch64_neon_fcvtzu : Intrinsic::aarch64_neon_fcvtzs;
10860 llvm::Type *Tys[2] = {Ty, GetFloatNeonType(this, Type)};
10861 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvtz");
10862 }
10863 case NEON::BI__builtin_neon_vcvta_s16_v:
10864 case NEON::BI__builtin_neon_vcvta_u16_v:
10865 case NEON::BI__builtin_neon_vcvta_s32_v:
10866 case NEON::BI__builtin_neon_vcvtaq_s16_v:
10867 case NEON::BI__builtin_neon_vcvtaq_s32_v:
10868 case NEON::BI__builtin_neon_vcvta_u32_v:
10869 case NEON::BI__builtin_neon_vcvtaq_u16_v:
10870 case NEON::BI__builtin_neon_vcvtaq_u32_v:
10871 case NEON::BI__builtin_neon_vcvta_s64_v:
10872 case NEON::BI__builtin_neon_vcvtaq_s64_v:
10873 case NEON::BI__builtin_neon_vcvta_u64_v:
10874 case NEON::BI__builtin_neon_vcvtaq_u64_v: {
10875 Int = usgn ? Intrinsic::aarch64_neon_fcvtau : Intrinsic::aarch64_neon_fcvtas;
10876 llvm::Type *Tys[2] = { Ty, GetFloatNeonType(this, Type) };
10877 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvta");
10878 }
10879 case NEON::BI__builtin_neon_vcvtm_s16_v:
10880 case NEON::BI__builtin_neon_vcvtm_s32_v:
10881 case NEON::BI__builtin_neon_vcvtmq_s16_v:
10882 case NEON::BI__builtin_neon_vcvtmq_s32_v:
10883 case NEON::BI__builtin_neon_vcvtm_u16_v:
10884 case NEON::BI__builtin_neon_vcvtm_u32_v:
10885 case NEON::BI__builtin_neon_vcvtmq_u16_v:
10886 case NEON::BI__builtin_neon_vcvtmq_u32_v:
10887 case NEON::BI__builtin_neon_vcvtm_s64_v:
10888 case NEON::BI__builtin_neon_vcvtmq_s64_v:
10889 case NEON::BI__builtin_neon_vcvtm_u64_v:
10890 case NEON::BI__builtin_neon_vcvtmq_u64_v: {
10891 Int = usgn ? Intrinsic::aarch64_neon_fcvtmu : Intrinsic::aarch64_neon_fcvtms;
10892 llvm::Type *Tys[2] = { Ty, GetFloatNeonType(this, Type) };
10893 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvtm");
10894 }
10895 case NEON::BI__builtin_neon_vcvtn_s16_v:
10896 case NEON::BI__builtin_neon_vcvtn_s32_v:
10897 case NEON::BI__builtin_neon_vcvtnq_s16_v:
10898 case NEON::BI__builtin_neon_vcvtnq_s32_v:
10899 case NEON::BI__builtin_neon_vcvtn_u16_v:
10900 case NEON::BI__builtin_neon_vcvtn_u32_v:
10901 case NEON::BI__builtin_neon_vcvtnq_u16_v:
10902 case NEON::BI__builtin_neon_vcvtnq_u32_v:
10903 case NEON::BI__builtin_neon_vcvtn_s64_v:
10904 case NEON::BI__builtin_neon_vcvtnq_s64_v:
10905 case NEON::BI__builtin_neon_vcvtn_u64_v:
10906 case NEON::BI__builtin_neon_vcvtnq_u64_v: {
10907 Int = usgn ? Intrinsic::aarch64_neon_fcvtnu : Intrinsic::aarch64_neon_fcvtns;
10908 llvm::Type *Tys[2] = { Ty, GetFloatNeonType(this, Type) };
10909 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvtn");
10910 }
10911 case NEON::BI__builtin_neon_vcvtp_s16_v:
10912 case NEON::BI__builtin_neon_vcvtp_s32_v:
10913 case NEON::BI__builtin_neon_vcvtpq_s16_v:
10914 case NEON::BI__builtin_neon_vcvtpq_s32_v:
10915 case NEON::BI__builtin_neon_vcvtp_u16_v:
10916 case NEON::BI__builtin_neon_vcvtp_u32_v:
10917 case NEON::BI__builtin_neon_vcvtpq_u16_v:
10918 case NEON::BI__builtin_neon_vcvtpq_u32_v:
10919 case NEON::BI__builtin_neon_vcvtp_s64_v:
10920 case NEON::BI__builtin_neon_vcvtpq_s64_v:
10921 case NEON::BI__builtin_neon_vcvtp_u64_v:
10922 case NEON::BI__builtin_neon_vcvtpq_u64_v: {
10923 Int = usgn ? Intrinsic::aarch64_neon_fcvtpu : Intrinsic::aarch64_neon_fcvtps;
10924 llvm::Type *Tys[2] = { Ty, GetFloatNeonType(this, Type) };
10925 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvtp");
10926 }
10927 case NEON::BI__builtin_neon_vmulx_v:
10928 case NEON::BI__builtin_neon_vmulxq_v: {
10929 Int = Intrinsic::aarch64_neon_fmulx;
10930 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmulx");
10931 }
10932 case NEON::BI__builtin_neon_vmulxh_lane_f16:
10933 case NEON::BI__builtin_neon_vmulxh_laneq_f16: {
10934 // vmulx_lane should be mapped to Neon scalar mulx after
10935 // extracting the scalar element
10936 Ops.push_back(EmitScalarExpr(E->getArg(2)));
10937 Ops[1] = Builder.CreateExtractElement(Ops[1], Ops[2], "extract");
10938 Ops.pop_back();
10939 Int = Intrinsic::aarch64_neon_fmulx;
10940 return EmitNeonCall(CGM.getIntrinsic(Int, HalfTy), Ops, "vmulx");
10941 }
10942 case NEON::BI__builtin_neon_vmul_lane_v:
10943 case NEON::BI__builtin_neon_vmul_laneq_v: {
10944 // v1f64 vmul_lane should be mapped to Neon scalar mul lane
10945 bool Quad = false;
10946 if (BuiltinID == NEON::BI__builtin_neon_vmul_laneq_v)
10947 Quad = true;
10948 Ops[0] = Builder.CreateBitCast(Ops[0], DoubleTy);
10949 llvm::FixedVectorType *VTy =
10950 GetNeonType(this, NeonTypeFlags(NeonTypeFlags::Float64, false, Quad));
10951 Ops[1] = Builder.CreateBitCast(Ops[1], VTy);
10952 Ops[1] = Builder.CreateExtractElement(Ops[1], Ops[2], "extract");
10953 Value *Result = Builder.CreateFMul(Ops[0], Ops[1]);
10954 return Builder.CreateBitCast(Result, Ty);
10955 }
10956 case NEON::BI__builtin_neon_vnegd_s64:
10957 return Builder.CreateNeg(EmitScalarExpr(E->getArg(0)), "vnegd");
10958 case NEON::BI__builtin_neon_vnegh_f16:
10959 return Builder.CreateFNeg(EmitScalarExpr(E->getArg(0)), "vnegh");
10960 case NEON::BI__builtin_neon_vpmaxnm_v:
10961 case NEON::BI__builtin_neon_vpmaxnmq_v: {
10962 Int = Intrinsic::aarch64_neon_fmaxnmp;
10963 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpmaxnm");
10964 }
10965 case NEON::BI__builtin_neon_vpminnm_v:
10966 case NEON::BI__builtin_neon_vpminnmq_v: {
10967 Int = Intrinsic::aarch64_neon_fminnmp;
10968 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpminnm");
10969 }
10970 case NEON::BI__builtin_neon_vsqrth_f16: {
10971 Ops.push_back(EmitScalarExpr(E->getArg(0)));
10972 Int = Builder.getIsFPConstrained()
10973 ? Intrinsic::experimental_constrained_sqrt
10974 : Intrinsic::sqrt;
10975 return EmitNeonCall(CGM.getIntrinsic(Int, HalfTy), Ops, "vsqrt");
10976 }
10977 case NEON::BI__builtin_neon_vsqrt_v:
10978 case NEON::BI__builtin_neon_vsqrtq_v: {
10979 Int = Builder.getIsFPConstrained()
10980 ? Intrinsic::experimental_constrained_sqrt
10981 : Intrinsic::sqrt;
10982 Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
10983 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vsqrt");
10984 }
10985 case NEON::BI__builtin_neon_vrbit_v:
10986 case NEON::BI__builtin_neon_vrbitq_v: {
10987 Int = Intrinsic::bitreverse;
10988 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrbit");
10989 }
10990 case NEON::BI__builtin_neon_vaddv_u8:
10991 // FIXME: These are handled by the AArch64 scalar code.
10992 usgn = true;
10993 LLVM_FALLTHROUGH[[gnu::fallthrough]];
10994 case NEON::BI__builtin_neon_vaddv_s8: {
10995 Int = usgn ? Intrinsic::aarch64_neon_uaddv : Intrinsic::aarch64_neon_saddv;
10996 Ty = Int32Ty;
10997 VTy = llvm::FixedVectorType::get(Int8Ty, 8);
10998 llvm::Type *Tys[2] = { Ty, VTy };
10999 Ops.push_back(EmitScalarExpr(E->getArg(0)));
11000 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddv");
11001 return Builder.CreateTrunc(Ops[0], Int8Ty);
11002 }
11003 case NEON::BI__builtin_neon_vaddv_u16:
11004 usgn = true;
11005 LLVM_FALLTHROUGH[[gnu::fallthrough]];
11006 case NEON::BI__builtin_neon_vaddv_s16: {
11007 Int = usgn ? Intrinsic::aarch64_neon_uaddv : Intrinsic::aarch64_neon_saddv;
11008 Ty = Int32Ty;
11009 VTy = llvm::FixedVectorType::get(Int16Ty, 4);
11010 llvm::Type *Tys[2] = { Ty, VTy };
11011 Ops.push_back(EmitScalarExpr(E->getArg(0)));
11012 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddv");
11013 return Builder.CreateTrunc(Ops[0], Int16Ty);
11014 }
11015 case NEON::BI__builtin_neon_vaddvq_u8:
11016 usgn = true;
11017 LLVM_FALLTHROUGH[[gnu::fallthrough]];
11018 case NEON::BI__builtin_neon_vaddvq_s8: {
11019 Int = usgn ? Intrinsic::aarch64_neon_uaddv : Intrinsic::aarch64_neon_saddv;
11020 Ty = Int32Ty;
11021 VTy = llvm::FixedVectorType::get(Int8Ty, 16);
11022 llvm::Type *Tys[2] = { Ty, VTy };
11023 Ops.push_back(EmitScalarExpr(E->getArg(0)));
11024 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddv");
11025 return Builder.CreateTrunc(Ops[0], Int8Ty);
11026 }
11027 case NEON::BI__builtin_neon_vaddvq_u16:
11028 usgn = true;
11029 LLVM_FALLTHROUGH[[gnu::fallthrough]];
11030 case NEON::BI__builtin_neon_vaddvq_s16: {
11031 Int = usgn ? Intrinsic::aarch64_neon_uaddv : Intrinsic::aarch64_neon_saddv;
11032 Ty = Int32Ty;
11033 VTy = llvm::FixedVectorType::get(Int16Ty, 8);
11034 llvm::Type *Tys[2] = { Ty, VTy };
11035 Ops.push_back(EmitScalarExpr(E->getArg(0)));
11036 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddv");
11037 return Builder.CreateTrunc(Ops[0], Int16Ty);
11038 }
11039 case NEON::BI__builtin_neon_vmaxv_u8: {
11040 Int = Intrinsic::aarch64_neon_umaxv;
11041 Ty = Int32Ty;
11042 VTy = llvm::FixedVectorType::get(Int8Ty, 8);
11043 llvm::Type *Tys[2] = { Ty, VTy };
11044 Ops.push_back(EmitScalarExpr(E->getArg(0)));
11045 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv");
11046 return Builder.CreateTrunc(Ops[0], Int8Ty);
11047 }
11048 case NEON::BI__builtin_neon_vmaxv_u16: {
11049 Int = Intrinsic::aarch64_neon_umaxv;
11050 Ty = Int32Ty;
11051 VTy = llvm::FixedVectorType::get(Int16Ty, 4);
11052 llvm::Type *Tys[2] = { Ty, VTy };
11053 Ops.push_back(EmitScalarExpr(E->getArg(0)));
11054 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv");
11055 return Builder.CreateTrunc(Ops[0], Int16Ty);
11056 }
11057 case NEON::BI__builtin_neon_vmaxvq_u8: {
11058 Int = Intrinsic::aarch64_neon_umaxv;
11059 Ty = Int32Ty;
11060 VTy = llvm::FixedVectorType::get(Int8Ty, 16);
11061 llvm::Type *Tys[2] = { Ty, VTy };
11062 Ops.push_back(EmitScalarExpr(E->getArg(0)));
11063 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv");
11064 return Builder.CreateTrunc(Ops[0], Int8Ty);
11065 }
11066 case NEON::BI__builtin_neon_vmaxvq_u16: {
11067 Int = Intrinsic::aarch64_neon_umaxv;
11068 Ty = Int32Ty;
11069 VTy = llvm::FixedVectorType::get(Int16Ty, 8);
11070 llvm::Type *Tys[2] = { Ty, VTy };
11071 Ops.push_back(EmitScalarExpr(E->getArg(0)));
11072 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv");
11073 return Builder.CreateTrunc(Ops[0], Int16Ty);
11074 }
11075 case NEON::BI__builtin_neon_vmaxv_s8: {
11076 Int = Intrinsic::aarch64_neon_smaxv;
11077 Ty = Int32Ty;
11078 VTy = llvm::FixedVectorType::get(Int8Ty, 8);
11079 llvm::Type *Tys[2] = { Ty, VTy };
11080 Ops.push_back(EmitScalarExpr(E->getArg(0)));
11081 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv");
11082 return Builder.CreateTrunc(Ops[0], Int8Ty);
11083 }
11084 case NEON::BI__builtin_neon_vmaxv_s16: {
11085 Int = Intrinsic::aarch64_neon_smaxv;
11086 Ty = Int32Ty;
11087 VTy = llvm::FixedVectorType::get(Int16Ty, 4);
11088 llvm::Type *Tys[2] = { Ty, VTy };
11089 Ops.push_back(EmitScalarExpr(E->getArg(0)));
11090 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv");
11091 return Builder.CreateTrunc(Ops[0], Int16Ty);
11092 }
11093 case NEON::BI__builtin_neon_vmaxvq_s8: {
11094 Int = Intrinsic::aarch64_neon_smaxv;
11095 Ty = Int32Ty;
11096 VTy = llvm::FixedVectorType::get(Int8Ty, 16);
11097 llvm::Type *Tys[2] = { Ty, VTy };
11098 Ops.push_back(EmitScalarExpr(E->getArg(0)));
11099 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv");
11100 return Builder.CreateTrunc(Ops[0], Int8Ty);
11101 }
11102 case NEON::BI__builtin_neon_vmaxvq_s16: {
11103 Int = Intrinsic::aarch64_neon_smaxv;
11104 Ty = Int32Ty;
11105 VTy = llvm::FixedVectorType::get(Int16Ty, 8);
11106 llvm::Type *Tys[2] = { Ty, VTy };
11107 Ops.push_back(EmitScalarExpr(E->getArg(0)));
11108 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv");
11109 return Builder.CreateTrunc(Ops[0], Int16Ty);
11110 }
11111 case NEON::BI__builtin_neon_vmaxv_f16: {
11112 Int = Intrinsic::aarch64_neon_fmaxv;
11113 Ty = HalfTy;
11114 VTy = llvm::FixedVectorType::get(HalfTy, 4);
11115 llvm::Type *Tys[2] = { Ty, VTy };
11116 Ops.push_back(EmitScalarExpr(E->getArg(0)));
11117 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv");
11118 return Builder.CreateTrunc(Ops[0], HalfTy);
11119 }
11120 case NEON::BI__builtin_neon_vmaxvq_f16: {
11121 Int = Intrinsic::aarch64_neon_fmaxv;
11122 Ty = HalfTy;
11123 VTy = llvm::FixedVectorType::get(HalfTy, 8);
11124 llvm::Type *Tys[2] = { Ty, VTy };
11125 Ops.push_back(EmitScalarExpr(E->getArg(0)));
11126 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv");
11127 return Builder.CreateTrunc(Ops[0], HalfTy);
11128 }
11129 case NEON::BI__builtin_neon_vminv_u8: {
11130 Int = Intrinsic::aarch64_neon_uminv;
11131 Ty = Int32Ty;
11132 VTy = llvm::FixedVectorType::get(Int8Ty, 8);
11133 llvm::Type *Tys[2] = { Ty, VTy };
11134 Ops.push_back(EmitScalarExpr(E->getArg(0)));
11135 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv");
11136 return Builder.CreateTrunc(Ops[0], Int8Ty);
11137 }
11138 case NEON::BI__builtin_neon_vminv_u16: {
11139 Int = Intrinsic::aarch64_neon_uminv;
11140 Ty = Int32Ty;
11141 VTy = llvm::FixedVectorType::get(Int16Ty, 4);
11142 llvm::Type *Tys[2] = { Ty, VTy };
11143 Ops.push_back(EmitScalarExpr(E->getArg(0)));
11144 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv");
11145 return Builder.CreateTrunc(Ops[0], Int16Ty);
11146 }
11147 case NEON::BI__builtin_neon_vminvq_u8: {
11148 Int = Intrinsic::aarch64_neon_uminv;
11149 Ty = Int32Ty;
11150 VTy = llvm::FixedVectorType::get(Int8Ty, 16);
11151 llvm::Type *Tys[2] = { Ty, VTy };
11152 Ops.push_back(EmitScalarExpr(E->getArg(0)));
11153 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv");
11154 return Builder.CreateTrunc(Ops[0], Int8Ty);
11155 }
11156 case NEON::BI__builtin_neon_vminvq_u16: {
11157 Int = Intrinsic::aarch64_neon_uminv;
11158 Ty = Int32Ty;
11159 VTy = llvm::FixedVectorType::get(Int16Ty, 8);
11160 llvm::Type *Tys[2] = { Ty, VTy };
11161 Ops.push_back(EmitScalarExpr(E->getArg(0)));
11162 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv");
11163 return Builder.CreateTrunc(Ops[0], Int16Ty);
11164 }
11165 case NEON::BI__builtin_neon_vminv_s8: {
11166 Int = Intrinsic::aarch64_neon_sminv;
11167 Ty = Int32Ty;
11168 VTy = llvm::FixedVectorType::get(Int8Ty, 8);
11169 llvm::Type *Tys[2] = { Ty, VTy };
11170 Ops.push_back(EmitScalarExpr(E->getArg(0)));
11171 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv");
11172 return Builder.CreateTrunc(Ops[0], Int8Ty);
11173 }
11174 case NEON::BI__builtin_neon_vminv_s16: {
11175 Int = Intrinsic::aarch64_neon_sminv;
11176 Ty = Int32Ty;
11177 VTy = llvm::FixedVectorType::get(Int16Ty, 4);
11178 llvm::Type *Tys[2] = { Ty, VTy };
11179 Ops.push_back(EmitScalarExpr(E->getArg(0)));
11180 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv");
11181 return Builder.CreateTrunc(Ops[0], Int16Ty);
11182 }
11183 case NEON::BI__builtin_neon_vminvq_s8: {
11184 Int = Intrinsic::aarch64_neon_sminv;
11185 Ty = Int32Ty;
11186 VTy = llvm::FixedVectorType::get(Int8Ty, 16);
11187 llvm::Type *Tys[2] = { Ty, VTy };
11188 Ops.push_back(EmitScalarExpr(E->getArg(0)));
11189 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv");
11190 return Builder.CreateTrunc(Ops[0], Int8Ty);
11191 }
11192 case NEON::BI__builtin_neon_vminvq_s16: {
11193 Int = Intrinsic::aarch64_neon_sminv;
11194 Ty = Int32Ty;
11195 VTy = llvm::FixedVectorType::get(Int16Ty, 8);
11196 llvm::Type *Tys[2] = { Ty, VTy };
11197 Ops.push_back(EmitScalarExpr(E->getArg(0)));
11198 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv");
11199 return Builder.CreateTrunc(Ops[0], Int16Ty);
11200 }
11201 case NEON::BI__builtin_neon_vminv_f16: {
11202 Int = Intrinsic::aarch64_neon_fminv;
11203 Ty = HalfTy;
11204 VTy = llvm::FixedVectorType::get(HalfTy, 4);
11205 llvm::Type *Tys[2] = { Ty, VTy };
11206 Ops.push_back(EmitScalarExpr(E->getArg(0)));
11207 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv");
11208 return Builder.CreateTrunc(Ops[0], HalfTy);
11209 }
11210 case NEON::BI__builtin_neon_vminvq_f16: {
11211 Int = Intrinsic::aarch64_neon_fminv;
11212 Ty = HalfTy;
11213 VTy = llvm::FixedVectorType::get(HalfTy, 8);
11214 llvm::Type *Tys[2] = { Ty, VTy };
11215 Ops.push_back(EmitScalarExpr(E->getArg(0)));
11216 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv");
11217 return Builder.CreateTrunc(Ops[0], HalfTy);
11218 }
11219 case NEON::BI__builtin_neon_vmaxnmv_f16: {
11220 Int = Intrinsic::aarch64_neon_fmaxnmv;
11221 Ty = HalfTy;
11222 VTy = llvm::FixedVectorType::get(HalfTy, 4);
11223 llvm::Type *Tys[2] = { Ty, VTy };
11224 Ops.push_back(EmitScalarExpr(E->getArg(0)));
11225 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxnmv");
11226 return Builder.CreateTrunc(Ops[0], HalfTy);
11227 }
11228 case NEON::BI__builtin_neon_vmaxnmvq_f16: {
11229 Int = Intrinsic::aarch64_neon_fmaxnmv;
11230 Ty = HalfTy;
11231 VTy = llvm::FixedVectorType::get(HalfTy, 8);
11232 llvm::Type *Tys[2] = { Ty, VTy };
11233 Ops.push_back(EmitScalarExpr(E->getArg(0)));
11234 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxnmv");
11235 return Builder.CreateTrunc(Ops[0], HalfTy);
11236 }
11237 case NEON::BI__builtin_neon_vminnmv_f16: {
11238 Int = Intrinsic::aarch64_neon_fminnmv;
11239 Ty = HalfTy;
11240 VTy = llvm::FixedVectorType::get(HalfTy, 4);
11241 llvm::Type *Tys[2] = { Ty, VTy };
11242 Ops.push_back(EmitScalarExpr(E->getArg(0)));
11243 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminnmv");
11244 return Builder.CreateTrunc(Ops[0], HalfTy);
11245 }
11246 case NEON::BI__builtin_neon_vminnmvq_f16: {
11247 Int = Intrinsic::aarch64_neon_fminnmv;
11248 Ty = HalfTy;
11249 VTy = llvm::FixedVectorType::get(HalfTy, 8);
11250 llvm::Type *Tys[2] = { Ty, VTy };
11251 Ops.push_back(EmitScalarExpr(E->getArg(0)));
11252 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminnmv");
11253 return Builder.CreateTrunc(Ops[0], HalfTy);
11254 }
11255 case NEON::BI__builtin_neon_vmul_n_f64: {
11256 Ops[0] = Builder.CreateBitCast(Ops[0], DoubleTy);
11257 Value *RHS = Builder.CreateBitCast(EmitScalarExpr(E->getArg(1)), DoubleTy);
11258 return Builder.CreateFMul(Ops[0], RHS);
11259 }
11260 case NEON::BI__builtin_neon_vaddlv_u8: {
11261 Int = Intrinsic::aarch64_neon_uaddlv;
11262 Ty = Int32Ty;
11263 VTy = llvm::FixedVectorType::get(Int8Ty, 8);
11264 llvm::Type *Tys[2] = { Ty, VTy };
11265 Ops.push_back(EmitScalarExpr(E->getArg(0)));
11266 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv");
11267 return Builder.CreateTrunc(Ops[0], Int16Ty);
11268 }
11269 case NEON::BI__builtin_neon_vaddlv_u16: {
11270 Int = Intrinsic::aarch64_neon_uaddlv;
11271 Ty = Int32Ty;
11272 VTy = llvm::FixedVectorType::get(Int16Ty, 4);
11273 llvm::Type *Tys[2] = { Ty, VTy };
11274 Ops.push_back(EmitScalarExpr(E->getArg(0)));
11275 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv");
11276 }
11277 case NEON::BI__builtin_neon_vaddlvq_u8: {
11278 Int = Intrinsic::aarch64_neon_uaddlv;
11279 Ty = Int32Ty;
11280 VTy = llvm::FixedVectorType::get(Int8Ty, 16);
11281 llvm::Type *Tys[2] = { Ty, VTy };
11282 Ops.push_back(EmitScalarExpr(E->getArg(0)));
11283 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv");
11284 return Builder.CreateTrunc(Ops[0], Int16Ty);
11285 }
11286 case NEON::BI__builtin_neon_vaddlvq_u16: {
11287 Int = Intrinsic::aarch64_neon_uaddlv;
11288 Ty = Int32Ty;
11289 VTy = llvm::FixedVectorType::get(Int16Ty, 8);
11290 llvm::Type *Tys[2] = { Ty, VTy };
11291 Ops.push_back(EmitScalarExpr(E->getArg(0)));
11292 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv");
11293 }
11294 case NEON::BI__builtin_neon_vaddlv_s8: {
11295 Int = Intrinsic::aarch64_neon_saddlv;
11296 Ty = Int32Ty;
11297 VTy = llvm::FixedVectorType::get(Int8Ty, 8);
11298 llvm::Type *Tys[2] = { Ty, VTy };
11299 Ops.push_back(EmitScalarExpr(E->getArg(0)));
11300 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv");
11301 return Builder.CreateTrunc(Ops[0], Int16Ty);
11302 }
11303 case NEON::BI__builtin_neon_vaddlv_s16: {
11304 Int = Intrinsic::aarch64_neon_saddlv;
11305 Ty = Int32Ty;
11306 VTy = llvm::FixedVectorType::get(Int16Ty, 4);
11307 llvm::Type *Tys[2] = { Ty, VTy };
11308 Ops.push_back(EmitScalarExpr(E->getArg(0)));
11309 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv");
11310 }
11311 case NEON::BI__builtin_neon_vaddlvq_s8: {
11312 Int = Intrinsic::aarch64_neon_saddlv;
11313 Ty = Int32Ty;
11314 VTy = llvm::FixedVectorType::get(Int8Ty, 16);
11315 llvm::Type *Tys[2] = { Ty, VTy };
11316 Ops.push_back(EmitScalarExpr(E->getArg(0)));
11317 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv");
11318 return Builder.CreateTrunc(Ops[0], Int16Ty);
11319 }
11320 case NEON::BI__builtin_neon_vaddlvq_s16: {
11321 Int = Intrinsic::aarch64_neon_saddlv;
11322 Ty = Int32Ty;
11323 VTy = llvm::FixedVectorType::get(Int16Ty, 8);
11324 llvm::Type *Tys[2] = { Ty, VTy };
11325 Ops.push_back(EmitScalarExpr(E->getArg(0)));
11326 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv");
11327 }
11328 case NEON::BI__builtin_neon_vsri_n_v:
11329 case NEON::BI__builtin_neon_vsriq_n_v: {
11330 Int = Intrinsic::aarch64_neon_vsri;
11331 llvm::Function *Intrin = CGM.getIntrinsic(Int, Ty);
11332 return EmitNeonCall(Intrin, Ops, "vsri_n");
11333 }
11334 case NEON::BI__builtin_neon_vsli_n_v:
11335 case NEON::BI__builtin_neon_vsliq_n_v: {
11336 Int = Intrinsic::aarch64_neon_vsli;
11337 llvm::Function *Intrin = CGM.getIntrinsic(Int, Ty);
11338 return EmitNeonCall(Intrin, Ops, "vsli_n");
11339 }
11340 case NEON::BI__builtin_neon_vsra_n_v:
11341 case NEON::BI__builtin_neon_vsraq_n_v:
11342 Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
11343 Ops[1] = EmitNeonRShiftImm(Ops[1], Ops[2], Ty, usgn, "vsra_n");
11344 return Builder.CreateAdd(Ops[0], Ops[1]);
11345 case NEON::BI__builtin_neon_vrsra_n_v:
11346 case NEON::BI__builtin_neon_vrsraq_n_v: {
11347 Int = usgn ? Intrinsic::aarch64_neon_urshl : Intrinsic::aarch64_neon_srshl;
11348 SmallVector<llvm::Value*,2> TmpOps;
11349 TmpOps.push_back(Ops[1]);
11350 TmpOps.push_back(Ops[2]);
11351 Function* F = CGM.getIntrinsic(Int, Ty);
11352 llvm::Value *tmp = EmitNeonCall(F, TmpOps, "vrshr_n", 1, true);
11353 Ops[0] = Builder.CreateBitCast(Ops[0], VTy);
11354 return Builder.CreateAdd(Ops[0], tmp);
11355 }
11356 case NEON::BI__builtin_neon_vld1_v:
11357 case NEON::BI__builtin_neon_vld1q_v: {
11358 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(VTy));
11359 return Builder.CreateAlignedLoad(VTy, Ops[0], PtrOp0.getAlignment());
11360 }
11361 case NEON::BI__builtin_neon_vst1_v:
11362 case NEON::BI__builtin_neon_vst1q_v:
11363 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(VTy));
11364 Ops[1] = Builder.CreateBitCast(Ops[1], VTy);
11365 return Builder.CreateAlignedStore(Ops[1], Ops[0], PtrOp0.getAlignment());
11366 case NEON::BI__builtin_neon_vld1_lane_v:
11367 case NEON::BI__builtin_neon_vld1q_lane_v: {
11368 Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
11369 Ty = llvm::PointerType::getUnqual(VTy->getElementType());
11370 Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
11371 Ops[0] = Builder.CreateAlignedLoad(VTy->getElementType(), Ops[0],
11372 PtrOp0.getAlignment());
11373 return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vld1_lane");
11374 }
11375 case NEON::BI__builtin_neon_vld1_dup_v:
11376 case NEON::BI__builtin_neon_vld1q_dup_v: {
11377 Value *V = UndefValue::get(Ty);
11378 Ty = llvm::PointerType::getUnqual(VTy->getElementType());
11379 Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
11380 Ops[0] = Builder.CreateAlignedLoad(VTy->getElementType(), Ops[0],
11381 PtrOp0.getAlignment());
11382 llvm::Constant *CI = ConstantInt::get(Int32Ty, 0);
11383 Ops[0] = Builder.CreateInsertElement(V, Ops[0], CI);
11384 return EmitNeonSplat(Ops[0], CI);
11385 }
11386 case NEON::BI__builtin_neon_vst1_lane_v:
11387 case NEON::BI__builtin_neon_vst1q_lane_v:
11388 Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
11389 Ops[1] = Builder.CreateExtractElement(Ops[1], Ops[2]);
11390 Ty = llvm::PointerType::getUnqual(Ops[1]->getType());
11391 return Builder.CreateAlignedStore(Ops[1], Builder.CreateBitCast(Ops[0], Ty),
11392 PtrOp0.getAlignment());
11393 case NEON::BI__builtin_neon_vld2_v:
11394 case NEON::BI__builtin_neon_vld2q_v: {
11395 llvm::Type *PTy = llvm::PointerType::getUnqual(VTy);
11396 Ops[1] = Builder.CreateBitCast(Ops[1], PTy);
11397 llvm::Type *Tys[2] = { VTy, PTy };
11398 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld2, Tys);
11399 Ops[1] = Builder.CreateCall(F, Ops[1], "vld2");
11400 Ops[0] = Builder.CreateBitCast(Ops[0],
11401 llvm::PointerType::getUnqual(Ops[1]->getType()));
11402 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]);
11403 }
11404 case NEON::BI__builtin_neon_vld3_v:
11405 case NEON::BI__builtin_neon_vld3q_v: {
11406 llvm::Type *PTy = llvm::PointerType::getUnqual(VTy);
11407 Ops[1] = Builder.CreateBitCast(Ops[1], PTy);
11408 llvm::Type *Tys[2] = { VTy, PTy };
11409 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld3, Tys);
11410 Ops[1] = Builder.CreateCall(F, Ops[1], "vld3");
11411 Ops[0] = Builder.CreateBitCast(Ops[0],
11412 llvm::PointerType::getUnqual(Ops[1]->getType()));
11413 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]);
11414 }
11415 case NEON::BI__builtin_neon_vld4_v:
11416 case NEON::BI__builtin_neon_vld4q_v: {
11417 llvm::Type *PTy = llvm::PointerType::getUnqual(VTy);
11418 Ops[1] = Builder.CreateBitCast(Ops[1], PTy);
11419 llvm::Type *Tys[2] = { VTy, PTy };
11420 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld4, Tys);
11421 Ops[1] = Builder.CreateCall(F, Ops[1], "vld4");
11422 Ops[0] = Builder.CreateBitCast(Ops[0],
11423 llvm::PointerType::getUnqual(Ops[1]->getType()));
11424 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]);
11425 }
11426 case NEON::BI__builtin_neon_vld2_dup_v:
11427 case NEON::BI__builtin_neon_vld2q_dup_v: {
11428 llvm::Type *PTy =
11429 llvm::PointerType::getUnqual(VTy->getElementType());
11430 Ops[1] = Builder.CreateBitCast(Ops[1], PTy);
11431 llvm::Type *Tys[2] = { VTy, PTy };
11432 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld2r, Tys);
11433 Ops[1] = Builder.CreateCall(F, Ops[1], "vld2");
11434 Ops[0] = Builder.CreateBitCast(Ops[0],
11435 llvm::PointerType::getUnqual(Ops[1]->getType()));
11436 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]);
11437 }
11438 case NEON::BI__builtin_neon_vld3_dup_v:
11439 case NEON::BI__builtin_neon_vld3q_dup_v: {
11440 llvm::Type *PTy =
11441 llvm::PointerType::getUnqual(VTy->getElementType());
11442 Ops[1] = Builder.CreateBitCast(Ops[1], PTy);
11443 llvm::Type *Tys[2] = { VTy, PTy };
11444 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld3r, Tys);
11445 Ops[1] = Builder.CreateCall(F, Ops[1], "vld3");
11446 Ops[0] = Builder.CreateBitCast(Ops[0],
11447 llvm::PointerType::getUnqual(Ops[1]->getType()));
11448 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]);
11449 }
11450 case NEON::BI__builtin_neon_vld4_dup_v:
11451 case NEON::BI__builtin_neon_vld4q_dup_v: {
11452 llvm::Type *PTy =
11453 llvm::PointerType::getUnqual(VTy->getElementType());
11454 Ops[1] = Builder.CreateBitCast(Ops[1], PTy);
11455 llvm::Type *Tys[2] = { VTy, PTy };
11456 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld4r, Tys);
11457 Ops[1] = Builder.CreateCall(F, Ops[1], "vld4");
11458 Ops[0] = Builder.CreateBitCast(Ops[0],
11459 llvm::PointerType::getUnqual(Ops[1]->getType()));
11460 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]);
11461 }
11462 case NEON::BI__builtin_neon_vld2_lane_v:
11463 case NEON::BI__builtin_neon_vld2q_lane_v: {
11464 llvm::Type *Tys[2] = { VTy, Ops[1]->getType() };
11465 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld2lane, Tys);
11466 std::rotate(Ops.begin() + 1, Ops.begin() + 2, Ops.end());
11467 Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
11468 Ops[2] = Builder.CreateBitCast(Ops[2], Ty);
11469 Ops[3] = Builder.CreateZExt(Ops[3], Int64Ty);
11470 Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), "vld2_lane");
11471 Ty = llvm::PointerType::getUnqual(Ops[1]->getType());
11472 Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
11473 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]);
11474 }
11475 case NEON::BI__builtin_neon_vld3_lane_v:
11476 case NEON::BI__builtin_neon_vld3q_lane_v: {
11477 llvm::Type *Tys[2] = { VTy, Ops[1]->getType() };
11478 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld3lane, Tys);
11479 std::rotate(Ops.begin() + 1, Ops.begin() + 2, Ops.end());
11480 Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
11481 Ops[2] = Builder.CreateBitCast(Ops[2], Ty);
11482 Ops[3] = Builder.CreateBitCast(Ops[3], Ty);
11483 Ops[4] = Builder.CreateZExt(Ops[4], Int64Ty);
11484 Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), "vld3_lane");
11485 Ty = llvm::PointerType::getUnqual(Ops[1]->getType());
11486 Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
11487 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]);
11488 }
11489 case NEON::BI__builtin_neon_vld4_lane_v:
11490 case NEON::BI__builtin_neon_vld4q_lane_v: {
11491 llvm::Type *Tys[2] = { VTy, Ops[1]->getType() };
11492 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld4lane, Tys);
11493 std::rotate(Ops.begin() + 1, Ops.begin() + 2, Ops.end());
11494 Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
11495 Ops[2] = Builder.CreateBitCast(Ops[2], Ty);
11496 Ops[3] = Builder.CreateBitCast(Ops[3], Ty);
11497 Ops[4] = Builder.CreateBitCast(Ops[4], Ty);
11498 Ops[5] = Builder.CreateZExt(Ops[5], Int64Ty);
11499 Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), "vld4_lane");
11500 Ty = llvm::PointerType::getUnqual(Ops[1]->getType());
11501 Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
11502 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]);
11503 }
11504 case NEON::BI__builtin_neon_vst2_v:
11505 case NEON::BI__builtin_neon_vst2q_v: {
11506 std::rotate(Ops.begin(), Ops.begin() + 1, Ops.end());
11507 llvm::Type *Tys[2] = { VTy, Ops[2]->getType() };
11508 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st2, Tys),
11509 Ops, "");
11510 }
11511 case NEON::BI__builtin_neon_vst2_lane_v:
11512 case NEON::BI__builtin_neon_vst2q_lane_v: {
11513 std::rotate(Ops.begin(), Ops.begin() + 1, Ops.end());
11514 Ops[2] = Builder.CreateZExt(Ops[2], Int64Ty);
11515 llvm::Type *Tys[2] = { VTy, Ops[3]->getType() };
11516 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st2lane, Tys),
11517 Ops, "");
11518 }
11519 case NEON::BI__builtin_neon_vst3_v:
11520 case NEON::BI__builtin_neon_vst3q_v: {
11521 std::rotate(Ops.begin(), Ops.begin() + 1, Ops.end());
11522 llvm::Type *Tys[2] = { VTy, Ops[3]->getType() };
11523 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st3, Tys),
11524 Ops, "");
11525 }
11526 case NEON::BI__builtin_neon_vst3_lane_v:
11527 case NEON::BI__builtin_neon_vst3q_lane_v: {
11528 std::rotate(Ops.begin(), Ops.begin() + 1, Ops.end());
11529 Ops[3] = Builder.CreateZExt(Ops[3], Int64Ty);
11530 llvm::Type *Tys[2] = { VTy, Ops[4]->getType() };
11531 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st3lane, Tys),
11532 Ops, "");
11533 }
11534 case NEON::BI__builtin_neon_vst4_v:
11535 case NEON::BI__builtin_neon_vst4q_v: {
11536 std::rotate(Ops.begin(), Ops.begin() + 1, Ops.end());
11537 llvm::Type *Tys[2] = { VTy, Ops[4]->getType() };
11538 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st4, Tys),
11539 Ops, "");
11540 }
11541 case NEON::BI__builtin_neon_vst4_lane_v:
11542 case NEON::BI__builtin_neon_vst4q_lane_v: {
11543 std::rotate(Ops.begin(), Ops.begin() + 1, Ops.end());
11544 Ops[4] = Builder.CreateZExt(Ops[4], Int64Ty);
11545 llvm::Type *Tys[2] = { VTy, Ops[5]->getType() };
11546 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st4lane, Tys),
11547 Ops, "");
11548 }
11549 case NEON::BI__builtin_neon_vtrn_v:
11550 case NEON::BI__builtin_neon_vtrnq_v: {
11551 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty));
11552 Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
11553 Ops[2] = Builder.CreateBitCast(Ops[2], Ty);
11554 Value *SV = nullptr;
11555
11556 for (unsigned vi = 0; vi != 2; ++vi) {
11557 SmallVector<int, 16> Indices;
11558 for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) {
11559 Indices.push_back(i+vi);
11560 Indices.push_back(i+e+vi);
11561 }
11562 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ty, Ops[0], vi);
11563 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], Indices, "vtrn");
11564 SV = Builder.CreateDefaultAlignedStore(SV, Addr);
11565 }
11566 return SV;
11567 }
11568 case NEON::BI__builtin_neon_vuzp_v:
11569 case NEON::BI__builtin_neon_vuzpq_v: {
11570 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty));
11571 Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
11572 Ops[2] = Builder.CreateBitCast(Ops[2], Ty);
11573 Value *SV = nullptr;
11574
11575 for (unsigned vi = 0; vi != 2; ++vi) {
11576 SmallVector<int, 16> Indices;
11577 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i)
11578 Indices.push_back(2*i+vi);
11579
11580 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ty, Ops[0], vi);
11581 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], Indices, "vuzp");
11582 SV = Builder.CreateDefaultAlignedStore(SV, Addr);
11583 }
11584 return SV;
11585 }
11586 case NEON::BI__builtin_neon_vzip_v:
11587 case NEON::BI__builtin_neon_vzipq_v: {
11588 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty));
11589 Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
11590 Ops[2] = Builder.CreateBitCast(Ops[2], Ty);
11591 Value *SV = nullptr;
11592
11593 for (unsigned vi = 0; vi != 2; ++vi) {
11594 SmallVector<int, 16> Indices;
11595 for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) {
11596 Indices.push_back((i + vi*e) >> 1);
11597 Indices.push_back(((i + vi*e) >> 1)+e);
11598 }
11599 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ty, Ops[0], vi);
11600 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], Indices, "vzip");
11601 SV = Builder.CreateDefaultAlignedStore(SV, Addr);
11602 }
11603 return SV;
11604 }
11605 case NEON::BI__builtin_neon_vqtbl1q_v: {
11606 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbl1, Ty),
11607 Ops, "vtbl1");
11608 }
11609 case NEON::BI__builtin_neon_vqtbl2q_v: {
11610 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbl2, Ty),
11611 Ops, "vtbl2");
11612 }
11613 case NEON::BI__builtin_neon_vqtbl3q_v: {
11614 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbl3, Ty),
11615 Ops, "vtbl3");
11616 }
11617 case NEON::BI__builtin_neon_vqtbl4q_v: {
11618 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbl4, Ty),
11619 Ops, "vtbl4");
11620 }
11621 case NEON::BI__builtin_neon_vqtbx1q_v: {
11622 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbx1, Ty),
11623 Ops, "vtbx1");
11624 }
11625 case NEON::BI__builtin_neon_vqtbx2q_v: {
11626 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbx2, Ty),
11627 Ops, "vtbx2");
11628 }
11629 case NEON::BI__builtin_neon_vqtbx3q_v: {
11630 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbx3, Ty),
11631 Ops, "vtbx3");
11632 }
11633 case NEON::BI__builtin_neon_vqtbx4q_v: {
11634 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbx4, Ty),
11635 Ops, "vtbx4");
11636 }
11637 case NEON::BI__builtin_neon_vsqadd_v:
11638 case NEON::BI__builtin_neon_vsqaddq_v: {
11639 Int = Intrinsic::aarch64_neon_usqadd;
11640 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vsqadd");
11641 }
11642 case NEON::BI__builtin_neon_vuqadd_v:
11643 case NEON::BI__builtin_neon_vuqaddq_v: {
11644 Int = Intrinsic::aarch64_neon_suqadd;
11645 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vuqadd");
11646 }
11647 }
11648}
11649
11650Value *CodeGenFunction::EmitBPFBuiltinExpr(unsigned BuiltinID,
11651 const CallExpr *E) {
11652 assert((BuiltinID == BPF::BI__builtin_preserve_field_info ||((void)0)
11653 BuiltinID == BPF::BI__builtin_btf_type_id ||((void)0)
11654 BuiltinID == BPF::BI__builtin_preserve_type_info ||((void)0)
11655 BuiltinID == BPF::BI__builtin_preserve_enum_value) &&((void)0)
11656 "unexpected BPF builtin")((void)0);
11657
11658 // A sequence number, injected into IR builtin functions, to
11659 // prevent CSE given the only difference of the funciton
11660 // may just be the debuginfo metadata.
11661 static uint32_t BuiltinSeqNum;
11662
11663 switch (BuiltinID) {
11664 default:
11665 llvm_unreachable("Unexpected BPF builtin")__builtin_unreachable();
11666 case BPF::BI__builtin_preserve_field_info: {
11667 const Expr *Arg = E->getArg(0);
11668 bool IsBitField = Arg->IgnoreParens()->getObjectKind() == OK_BitField;
11669
11670 if (!getDebugInfo()) {
11671 CGM.Error(E->getExprLoc(),
11672 "using __builtin_preserve_field_info() without -g");
11673 return IsBitField ? EmitLValue(Arg).getBitFieldPointer()
11674 : EmitLValue(Arg).getPointer(*this);
11675 }
11676
11677 // Enable underlying preserve_*_access_index() generation.
11678 bool OldIsInPreservedAIRegion = IsInPreservedAIRegion;
11679 IsInPreservedAIRegion = true;
11680 Value *FieldAddr = IsBitField ? EmitLValue(Arg).getBitFieldPointer()
11681 : EmitLValue(Arg).getPointer(*this);
11682 IsInPreservedAIRegion = OldIsInPreservedAIRegion;
11683
11684 ConstantInt *C = cast<ConstantInt>(EmitScalarExpr(E->getArg(1)));
11685 Value *InfoKind = ConstantInt::get(Int64Ty, C->getSExtValue());
11686
11687 // Built the IR for the preserve_field_info intrinsic.
11688 llvm::Function *FnGetFieldInfo = llvm::Intrinsic::getDeclaration(
11689 &CGM.getModule(), llvm::Intrinsic::bpf_preserve_field_info,
11690 {FieldAddr->getType()});
11691 return Builder.CreateCall(FnGetFieldInfo, {FieldAddr, InfoKind});
11692 }
11693 case BPF::BI__builtin_btf_type_id:
11694 case BPF::BI__builtin_preserve_type_info: {
11695 if (!getDebugInfo()) {
11696 CGM.Error(E->getExprLoc(), "using builtin function without -g");
11697 return nullptr;
11698 }
11699
11700 const Expr *Arg0 = E->getArg(0);
11701 llvm::DIType *DbgInfo = getDebugInfo()->getOrCreateStandaloneType(
11702 Arg0->getType(), Arg0->getExprLoc());
11703
11704 ConstantInt *Flag = cast<ConstantInt>(EmitScalarExpr(E->getArg(1)));
11705 Value *FlagValue = ConstantInt::get(Int64Ty, Flag->getSExtValue());
11706 Value *SeqNumVal = ConstantInt::get(Int32Ty, BuiltinSeqNum++);
11707
11708 llvm::Function *FnDecl;
11709 if (BuiltinID == BPF::BI__builtin_btf_type_id)
11710 FnDecl = llvm::Intrinsic::getDeclaration(
11711 &CGM.getModule(), llvm::Intrinsic::bpf_btf_type_id, {});
11712 else
11713 FnDecl = llvm::Intrinsic::getDeclaration(
11714 &CGM.getModule(), llvm::Intrinsic::bpf_preserve_type_info, {});
11715 CallInst *Fn = Builder.CreateCall(FnDecl, {SeqNumVal, FlagValue});
11716 Fn->setMetadata(LLVMContext::MD_preserve_access_index, DbgInfo);
11717 return Fn;
11718 }
11719 case BPF::BI__builtin_preserve_enum_value: {
11720 if (!getDebugInfo()) {
11721 CGM.Error(E->getExprLoc(), "using builtin function without -g");
11722 return nullptr;
11723 }
11724
11725 const Expr *Arg0 = E->getArg(0);
11726 llvm::DIType *DbgInfo = getDebugInfo()->getOrCreateStandaloneType(
11727 Arg0->getType(), Arg0->getExprLoc());
11728
11729 // Find enumerator
11730 const auto *UO = cast<UnaryOperator>(Arg0->IgnoreParens());
11731 const auto *CE = cast<CStyleCastExpr>(UO->getSubExpr());
11732 const auto *DR = cast<DeclRefExpr>(CE->getSubExpr());
11733 const auto *Enumerator = cast<EnumConstantDecl>(DR->getDecl());
11734
11735 auto &InitVal = Enumerator->getInitVal();
11736 std::string InitValStr;
11737 if (InitVal.isNegative() || InitVal > uint64_t(INT64_MAX0x7fffffffffffffffLL))
11738 InitValStr = std::to_string(InitVal.getSExtValue());
11739 else
11740 InitValStr = std::to_string(InitVal.getZExtValue());
11741 std::string EnumStr = Enumerator->getNameAsString() + ":" + InitValStr;
11742 Value *EnumStrVal = Builder.CreateGlobalStringPtr(EnumStr);
11743
11744 ConstantInt *Flag = cast<ConstantInt>(EmitScalarExpr(E->getArg(1)));
11745 Value *FlagValue = ConstantInt::get(Int64Ty, Flag->getSExtValue());
11746 Value *SeqNumVal = ConstantInt::get(Int32Ty, BuiltinSeqNum++);
11747
11748 llvm::Function *IntrinsicFn = llvm::Intrinsic::getDeclaration(
11749 &CGM.getModule(), llvm::Intrinsic::bpf_preserve_enum_value, {});
11750 CallInst *Fn =
11751 Builder.CreateCall(IntrinsicFn, {SeqNumVal, EnumStrVal, FlagValue});
11752 Fn->setMetadata(LLVMContext::MD_preserve_access_index, DbgInfo);
11753 return Fn;
11754 }
11755 }
11756}
11757
11758llvm::Value *CodeGenFunction::
11759BuildVector(ArrayRef<llvm::Value*> Ops) {
11760 assert((Ops.size() & (Ops.size() - 1)) == 0 &&((void)0)
11761 "Not a power-of-two sized vector!")((void)0);
11762 bool AllConstants = true;
11763 for (unsigned i = 0, e = Ops.size(); i != e && AllConstants; ++i)
11764 AllConstants &= isa<Constant>(Ops[i]);
11765
11766 // If this is a constant vector, create a ConstantVector.
11767 if (AllConstants) {
11768 SmallVector<llvm::Constant*, 16> CstOps;
11769 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
11770 CstOps.push_back(cast<Constant>(Ops[i]));
11771 return llvm::ConstantVector::get(CstOps);
11772 }
11773
11774 // Otherwise, insertelement the values to build the vector.
11775 Value *Result = llvm::UndefValue::get(
11776 llvm::FixedVectorType::get(Ops[0]->getType(), Ops.size()));
11777
11778 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
11779 Result = Builder.CreateInsertElement(Result, Ops[i], Builder.getInt32(i));
11780
11781 return Result;
11782}
11783
11784// Convert the mask from an integer type to a vector of i1.
11785static Value *getMaskVecValue(CodeGenFunction &CGF, Value *Mask,
11786 unsigned NumElts) {
11787
11788 auto *MaskTy = llvm::FixedVectorType::get(
11789 CGF.Builder.getInt1Ty(),
11790 cast<IntegerType>(Mask->getType())->getBitWidth());
11791 Value *MaskVec = CGF.Builder.CreateBitCast(Mask, MaskTy);
11792
11793 // If we have less than 8 elements, then the starting mask was an i8 and
11794 // we need to extract down to the right number of elements.
11795 if (NumElts < 8) {
11796 int Indices[4];
11797 for (unsigned i = 0; i != NumElts; ++i)
11798 Indices[i] = i;
11799 MaskVec = CGF.Builder.CreateShuffleVector(MaskVec, MaskVec,
11800 makeArrayRef(Indices, NumElts),
11801 "extract");
11802 }
11803 return MaskVec;
11804}
11805
11806static Value *EmitX86MaskedStore(CodeGenFunction &CGF, ArrayRef<Value *> Ops,
11807 Align Alignment) {
11808 // Cast the pointer to right type.
11809 Value *Ptr = CGF.Builder.CreateBitCast(Ops[0],
11810 llvm::PointerType::getUnqual(Ops[1]->getType()));
11811
11812 Value *MaskVec = getMaskVecValue(
11813 CGF, Ops[2],
11814 cast<llvm::FixedVectorType>(Ops[1]->getType())->getNumElements());
11815
11816 return CGF.Builder.CreateMaskedStore(Ops[1], Ptr, Alignment, MaskVec);
11817}
11818
11819static Value *EmitX86MaskedLoad(CodeGenFunction &CGF, ArrayRef<Value *> Ops,
11820 Align Alignment) {
11821 // Cast the pointer to right type.
11822 llvm::Type *Ty = Ops[1]->getType();
11823 Value *Ptr =
11824 CGF.Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty));
11825
11826 Value *MaskVec = getMaskVecValue(
11827 CGF, Ops[2], cast<llvm::FixedVectorType>(Ty)->getNumElements());
11828
11829 return CGF.Builder.CreateMaskedLoad(Ty, Ptr, Alignment, MaskVec, Ops[1]);
11830}
11831
11832static Value *EmitX86ExpandLoad(CodeGenFunction &CGF,
11833 ArrayRef<Value *> Ops) {
11834 auto *ResultTy = cast<llvm::VectorType>(Ops[1]->getType());
11835 llvm::Type *PtrTy = ResultTy->getElementType();
11836
11837 // Cast the pointer to element type.
11838 Value *Ptr = CGF.Builder.CreateBitCast(Ops[0],
11839 llvm::PointerType::getUnqual(PtrTy));
11840
11841 Value *MaskVec = getMaskVecValue(
11842 CGF, Ops[2], cast<FixedVectorType>(ResultTy)->getNumElements());
11843
11844 llvm::Function *F = CGF.CGM.getIntrinsic(Intrinsic::masked_expandload,
11845 ResultTy);
11846 return CGF.Builder.CreateCall(F, { Ptr, MaskVec, Ops[1] });
11847}
11848
11849static Value *EmitX86CompressExpand(CodeGenFunction &CGF,
11850 ArrayRef<Value *> Ops,
11851 bool IsCompress) {
11852 auto *ResultTy = cast<llvm::FixedVectorType>(Ops[1]->getType());
11853
11854 Value *MaskVec = getMaskVecValue(CGF, Ops[2], ResultTy->getNumElements());
11855
11856 Intrinsic::ID IID = IsCompress ? Intrinsic::x86_avx512_mask_compress
11857 : Intrinsic::x86_avx512_mask_expand;
11858 llvm::Function *F = CGF.CGM.getIntrinsic(IID, ResultTy);
11859 return CGF.Builder.CreateCall(F, { Ops[0], Ops[1], MaskVec });
11860}
11861
11862static Value *EmitX86CompressStore(CodeGenFunction &CGF,
11863 ArrayRef<Value *> Ops) {
11864 auto *ResultTy = cast<llvm::FixedVectorType>(Ops[1]->getType());
11865 llvm::Type *PtrTy = ResultTy->getElementType();
11866
11867 // Cast the pointer to element type.
11868 Value *Ptr = CGF.Builder.CreateBitCast(Ops[0],
11869 llvm::PointerType::getUnqual(PtrTy));
11870
11871 Value *MaskVec = getMaskVecValue(CGF, Ops[2], ResultTy->getNumElements());
11872
11873 llvm::Function *F = CGF.CGM.getIntrinsic(Intrinsic::masked_compressstore,
11874 ResultTy);
11875 return CGF.Builder.CreateCall(F, { Ops[1], Ptr, MaskVec });
11876}
11877
11878static Value *EmitX86MaskLogic(CodeGenFunction &CGF, Instruction::BinaryOps Opc,
11879 ArrayRef<Value *> Ops,
11880 bool InvertLHS = false) {
11881 unsigned NumElts = Ops[0]->getType()->getIntegerBitWidth();
11882 Value *LHS = getMaskVecValue(CGF, Ops[0], NumElts);
11883 Value *RHS = getMaskVecValue(CGF, Ops[1], NumElts);
11884
11885 if (InvertLHS)
11886 LHS = CGF.Builder.CreateNot(LHS);
11887
11888 return CGF.Builder.CreateBitCast(CGF.Builder.CreateBinOp(Opc, LHS, RHS),
11889 Ops[0]->getType());
11890}
11891
11892static Value *EmitX86FunnelShift(CodeGenFunction &CGF, Value *Op0, Value *Op1,
11893 Value *Amt, bool IsRight) {
11894 llvm::Type *Ty = Op0->getType();
11895
11896 // Amount may be scalar immediate, in which case create a splat vector.
11897 // Funnel shifts amounts are treated as modulo and types are all power-of-2 so
11898 // we only care about the lowest log2 bits anyway.
11899 if (Amt->getType() != Ty) {
11900 unsigned NumElts = cast<llvm::FixedVectorType>(Ty)->getNumElements();
11901 Amt = CGF.Builder.CreateIntCast(Amt, Ty->getScalarType(), false);
11902 Amt = CGF.Builder.CreateVectorSplat(NumElts, Amt);
11903 }
11904
11905 unsigned IID = IsRight ? Intrinsic::fshr : Intrinsic::fshl;
11906 Function *F = CGF.CGM.getIntrinsic(IID, Ty);
11907 return CGF.Builder.CreateCall(F, {Op0, Op1, Amt});
11908}
11909
11910static Value *EmitX86vpcom(CodeGenFunction &CGF, ArrayRef<Value *> Ops,
11911 bool IsSigned) {
11912 Value *Op0 = Ops[0];
11913 Value *Op1 = Ops[1];
11914 llvm::Type *Ty = Op0->getType();
11915 uint64_t Imm = cast<llvm::ConstantInt>(Ops[2])->getZExtValue() & 0x7;
11916
11917 CmpInst::Predicate Pred;
11918 switch (Imm) {
11919 case 0x0:
11920 Pred = IsSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT;
11921 break;
11922 case 0x1:
11923 Pred = IsSigned ? ICmpInst::ICMP_SLE : ICmpInst::ICMP_ULE;
11924 break;
11925 case 0x2:
11926 Pred = IsSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
11927 break;
11928 case 0x3:
11929 Pred = IsSigned ? ICmpInst::ICMP_SGE : ICmpInst::ICMP_UGE;
11930 break;
11931 case 0x4:
11932 Pred = ICmpInst::ICMP_EQ;
11933 break;
11934 case 0x5:
11935 Pred = ICmpInst::ICMP_NE;
11936 break;
11937 case 0x6:
11938 return llvm::Constant::getNullValue(Ty); // FALSE
11939 case 0x7:
11940 return llvm::Constant::getAllOnesValue(Ty); // TRUE
11941 default:
11942 llvm_unreachable("Unexpected XOP vpcom/vpcomu predicate")__builtin_unreachable();
11943 }
11944
11945 Value *Cmp = CGF.Builder.CreateICmp(Pred, Op0, Op1);
11946 Value *Res = CGF.Builder.CreateSExt(Cmp, Ty);
11947 return Res;
11948}
11949
11950static Value *EmitX86Select(CodeGenFunction &CGF,
11951 Value *Mask, Value *Op0, Value *Op1) {
11952
11953 // If the mask is all ones just return first argument.
11954 if (const auto *C = dyn_cast<Constant>(Mask))
11955 if (C->isAllOnesValue())
11956 return Op0;
11957
11958 Mask = getMaskVecValue(
11959 CGF, Mask, cast<llvm::FixedVectorType>(Op0->getType())->getNumElements());
11960
11961 return CGF.Builder.CreateSelect(Mask, Op0, Op1);
11962}
11963
11964static Value *EmitX86ScalarSelect(CodeGenFunction &CGF,
11965 Value *Mask, Value *Op0, Value *Op1) {
11966 // If the mask is all ones just return first argument.
11967 if (const auto *C = dyn_cast<Constant>(Mask))
11968 if (C->isAllOnesValue())
11969 return Op0;
11970
11971 auto *MaskTy = llvm::FixedVectorType::get(
11972 CGF.Builder.getInt1Ty(), Mask->getType()->getIntegerBitWidth());
11973 Mask = CGF.Builder.CreateBitCast(Mask, MaskTy);
11974 Mask = CGF.Builder.CreateExtractElement(Mask, (uint64_t)0);
11975 return CGF.Builder.CreateSelect(Mask, Op0, Op1);
11976}
11977
11978static Value *EmitX86MaskedCompareResult(CodeGenFunction &CGF, Value *Cmp,
11979 unsigned NumElts, Value *MaskIn) {
11980 if (MaskIn) {
11981 const auto *C = dyn_cast<Constant>(MaskIn);
11982 if (!C || !C->isAllOnesValue())
11983 Cmp = CGF.Builder.CreateAnd(Cmp, getMaskVecValue(CGF, MaskIn, NumElts));
11984 }
11985
11986 if (NumElts < 8) {
11987 int Indices[8];
11988 for (unsigned i = 0; i != NumElts; ++i)
11989 Indices[i] = i;
11990 for (unsigned i = NumElts; i != 8; ++i)
11991 Indices[i] = i % NumElts + NumElts;
11992 Cmp = CGF.Builder.CreateShuffleVector(
11993 Cmp, llvm::Constant::getNullValue(Cmp->getType()), Indices);
11994 }
11995
11996 return CGF.Builder.CreateBitCast(Cmp,
11997 IntegerType::get(CGF.getLLVMContext(),
11998 std::max(NumElts, 8U)));
11999}
12000
12001static Value *EmitX86MaskedCompare(CodeGenFunction &CGF, unsigned CC,
12002 bool Signed, ArrayRef<Value *> Ops) {
12003 assert((Ops.size() == 2 || Ops.size() == 4) &&((void)0)
12004 "Unexpected number of arguments")((void)0);
12005 unsigned NumElts =
12006 cast<llvm::FixedVectorType>(Ops[0]->getType())->getNumElements();
12007 Value *Cmp;
12008
12009 if (CC == 3) {
12010 Cmp = Constant::getNullValue(
12011 llvm::FixedVectorType::get(CGF.Builder.getInt1Ty(), NumElts));
12012 } else if (CC == 7) {
12013 Cmp = Constant::getAllOnesValue(
12014 llvm::FixedVectorType::get(CGF.Builder.getInt1Ty(), NumElts));
12015 } else {
12016 ICmpInst::Predicate Pred;
12017 switch (CC) {
12018 default: llvm_unreachable("Unknown condition code")__builtin_unreachable();
12019 case 0: Pred = ICmpInst::ICMP_EQ; break;
12020 case 1: Pred = Signed ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT; break;
12021 case 2: Pred = Signed ? ICmpInst::ICMP_SLE : ICmpInst::ICMP_ULE; break;
12022 case 4: Pred = ICmpInst::ICMP_NE; break;
12023 case 5: Pred = Signed ? ICmpInst::ICMP_SGE : ICmpInst::ICMP_UGE; break;
12024 case 6: Pred = Signed ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; break;
12025 }
12026 Cmp = CGF.Builder.CreateICmp(Pred, Ops[0], Ops[1]);
12027 }
12028
12029 Value *MaskIn = nullptr;
12030 if (Ops.size() == 4)
12031 MaskIn = Ops[3];
12032
12033 return EmitX86MaskedCompareResult(CGF, Cmp, NumElts, MaskIn);
12034}
12035
12036static Value *EmitX86ConvertToMask(CodeGenFunction &CGF, Value *In) {
12037 Value *Zero = Constant::getNullValue(In->getType());
12038 return EmitX86MaskedCompare(CGF, 1, true, { In, Zero });
12039}
12040
12041static Value *EmitX86ConvertIntToFp(CodeGenFunction &CGF, const CallExpr *E,
12042 ArrayRef<Value *> Ops, bool IsSigned) {
12043 unsigned Rnd = cast<llvm::ConstantInt>(Ops[3])->getZExtValue();
12044 llvm::Type *Ty = Ops[1]->getType();
12045
12046 Value *Res;
12047 if (Rnd != 4) {
12048 Intrinsic::ID IID = IsSigned ? Intrinsic::x86_avx512_sitofp_round
12049 : Intrinsic::x86_avx512_uitofp_round;
12050 Function *F = CGF.CGM.getIntrinsic(IID, { Ty, Ops[0]->getType() });
12051 Res = CGF.Builder.CreateCall(F, { Ops[0], Ops[3] });
12052 } else {
12053 CodeGenFunction::CGFPOptionsRAII FPOptsRAII(CGF, E);
12054 Res = IsSigned ? CGF.Builder.CreateSIToFP(Ops[0], Ty)
12055 : CGF.Builder.CreateUIToFP(Ops[0], Ty);
12056 }
12057
12058 return EmitX86Select(CGF, Ops[2], Res, Ops[1]);
12059}
12060
12061// Lowers X86 FMA intrinsics to IR.
12062static Value *EmitX86FMAExpr(CodeGenFunction &CGF, const CallExpr *E,
12063 ArrayRef<Value *> Ops, unsigned BuiltinID,
12064 bool IsAddSub) {
12065
12066 bool Subtract = false;
12067 Intrinsic::ID IID = Intrinsic::not_intrinsic;
12068 switch (BuiltinID) {
12069 default: break;
12070 case clang::X86::BI__builtin_ia32_vfmsubps512_mask3:
12071 Subtract = true;
12072 LLVM_FALLTHROUGH[[gnu::fallthrough]];
12073 case clang::X86::BI__builtin_ia32_vfmaddps512_mask:
12074 case clang::X86::BI__builtin_ia32_vfmaddps512_maskz:
12075 case clang::X86::BI__builtin_ia32_vfmaddps512_mask3:
12076 IID = llvm::Intrinsic::x86_avx512_vfmadd_ps_512; break;
12077 case clang::X86::BI__builtin_ia32_vfmsubpd512_mask3:
12078 Subtract = true;
12079 LLVM_FALLTHROUGH[[gnu::fallthrough]];
12080 case clang::X86::BI__builtin_ia32_vfmaddpd512_mask:
12081 case clang::X86::BI__builtin_ia32_vfmaddpd512_maskz:
12082 case clang::X86::BI__builtin_ia32_vfmaddpd512_mask3:
12083 IID = llvm::Intrinsic::x86_avx512_vfmadd_pd_512; break;
12084 case clang::X86::BI__builtin_ia32_vfmsubaddps512_mask3:
12085 Subtract = true;
12086 LLVM_FALLTHROUGH[[gnu::fallthrough]];
12087 case clang::X86::BI__builtin_ia32_vfmaddsubps512_mask:
12088 case clang::X86::BI__builtin_ia32_vfmaddsubps512_maskz:
12089 case clang::X86::BI__builtin_ia32_vfmaddsubps512_mask3:
12090 IID = llvm::Intrinsic::x86_avx512_vfmaddsub_ps_512;
12091 break;
12092 case clang::X86::BI__builtin_ia32_vfmsubaddpd512_mask3:
12093 Subtract = true;
12094 LLVM_FALLTHROUGH[[gnu::fallthrough]];
12095 case clang::X86::BI__builtin_ia32_vfmaddsubpd512_mask:
12096 case clang::X86::BI__builtin_ia32_vfmaddsubpd512_maskz:
12097 case clang::X86::BI__builtin_ia32_vfmaddsubpd512_mask3:
12098 IID = llvm::Intrinsic::x86_avx512_vfmaddsub_pd_512;
12099 break;
12100 }
12101
12102 Value *A = Ops[0];
12103 Value *B = Ops[1];
12104 Value *C = Ops[2];
12105
12106 if (Subtract)
12107 C = CGF.Builder.CreateFNeg(C);
12108
12109 Value *Res;
12110
12111 // Only handle in case of _MM_FROUND_CUR_DIRECTION/4 (no rounding).
12112 if (IID != Intrinsic::not_intrinsic &&
12113 (cast<llvm::ConstantInt>(Ops.back())->getZExtValue() != (uint64_t)4 ||
12114 IsAddSub)) {
12115 Function *Intr = CGF.CGM.getIntrinsic(IID);
12116 Res = CGF.Builder.CreateCall(Intr, {A, B, C, Ops.back() });
12117 } else {
12118 llvm::Type *Ty = A->getType();
12119 Function *FMA;
12120 if (CGF.Builder.getIsFPConstrained()) {
12121 CodeGenFunction::CGFPOptionsRAII FPOptsRAII(CGF, E);
12122 FMA = CGF.CGM.getIntrinsic(Intrinsic::experimental_constrained_fma, Ty);
12123 Res = CGF.Builder.CreateConstrainedFPCall(FMA, {A, B, C});
12124 } else {
12125 FMA = CGF.CGM.getIntrinsic(Intrinsic::fma, Ty);
12126 Res = CGF.Builder.CreateCall(FMA, {A, B, C});
12127 }
12128 }
12129
12130 // Handle any required masking.
12131 Value *MaskFalseVal = nullptr;
12132 switch (BuiltinID) {
12133 case clang::X86::BI__builtin_ia32_vfmaddps512_mask:
12134 case clang::X86::BI__builtin_ia32_vfmaddpd512_mask:
12135 case clang::X86::BI__builtin_ia32_vfmaddsubps512_mask:
12136 case clang::X86::BI__builtin_ia32_vfmaddsubpd512_mask:
12137 MaskFalseVal = Ops[0];
12138 break;
12139 case clang::X86::BI__builtin_ia32_vfmaddps512_maskz:
12140 case clang::X86::BI__builtin_ia32_vfmaddpd512_maskz:
12141 case clang::X86::BI__builtin_ia32_vfmaddsubps512_maskz:
12142 case clang::X86::BI__builtin_ia32_vfmaddsubpd512_maskz:
12143 MaskFalseVal = Constant::getNullValue(Ops[0]->getType());
12144 break;
12145 case clang::X86::BI__builtin_ia32_vfmsubps512_mask3:
12146 case clang::X86::BI__builtin_ia32_vfmaddps512_mask3:
12147 case clang::X86::BI__builtin_ia32_vfmsubpd512_mask3:
12148 case clang::X86::BI__builtin_ia32_vfmaddpd512_mask3:
12149 case clang::X86::BI__builtin_ia32_vfmsubaddps512_mask3:
12150 case clang::X86::BI__builtin_ia32_vfmaddsubps512_mask3:
12151 case clang::X86::BI__builtin_ia32_vfmsubaddpd512_mask3:
12152 case clang::X86::BI__builtin_ia32_vfmaddsubpd512_mask3:
12153 MaskFalseVal = Ops[2];
12154 break;
12155 }
12156
12157 if (MaskFalseVal)
12158 return EmitX86Select(CGF, Ops[3], Res, MaskFalseVal);
12159
12160 return Res;
12161}
12162
12163static Value *EmitScalarFMAExpr(CodeGenFunction &CGF, const CallExpr *E,
12164 MutableArrayRef<Value *> Ops, Value *Upper,
12165 bool ZeroMask = false, unsigned PTIdx = 0,
12166 bool NegAcc = false) {
12167 unsigned Rnd = 4;
12168 if (Ops.size() > 4)
12169 Rnd = cast<llvm::ConstantInt>(Ops[4])->getZExtValue();
12170
12171 if (NegAcc)
12172 Ops[2] = CGF.Builder.CreateFNeg(Ops[2]);
12173
12174 Ops[0] = CGF.Builder.CreateExtractElement(Ops[0], (uint64_t)0);
12175 Ops[1] = CGF.Builder.CreateExtractElement(Ops[1], (uint64_t)0);
12176 Ops[2] = CGF.Builder.CreateExtractElement(Ops[2], (uint64_t)0);
12177 Value *Res;
12178 if (Rnd != 4) {
12179 Intrinsic::ID IID = Ops[0]->getType()->getPrimitiveSizeInBits() == 32 ?
12180 Intrinsic::x86_avx512_vfmadd_f32 :
12181 Intrinsic::x86_avx512_vfmadd_f64;
12182 Res = CGF.Builder.CreateCall(CGF.CGM.getIntrinsic(IID),
12183 {Ops[0], Ops[1], Ops[2], Ops[4]});
12184 } else if (CGF.Builder.getIsFPConstrained()) {
12185 CodeGenFunction::CGFPOptionsRAII FPOptsRAII(CGF, E);
12186 Function *FMA = CGF.CGM.getIntrinsic(
12187 Intrinsic::experimental_constrained_fma, Ops[0]->getType());
12188 Res = CGF.Builder.CreateConstrainedFPCall(FMA, Ops.slice(0, 3));
12189 } else {
12190 Function *FMA = CGF.CGM.getIntrinsic(Intrinsic::fma, Ops[0]->getType());
12191 Res = CGF.Builder.CreateCall(FMA, Ops.slice(0, 3));
12192 }
12193 // If we have more than 3 arguments, we need to do masking.
12194 if (Ops.size() > 3) {
12195 Value *PassThru = ZeroMask ? Constant::getNullValue(Res->getType())
12196 : Ops[PTIdx];
12197
12198 // If we negated the accumulator and the its the PassThru value we need to
12199 // bypass the negate. Conveniently Upper should be the same thing in this
12200 // case.
12201 if (NegAcc && PTIdx == 2)
12202 PassThru = CGF.Builder.CreateExtractElement(Upper, (uint64_t)0);
12203
12204 Res = EmitX86ScalarSelect(CGF, Ops[3], Res, PassThru);
12205 }
12206 return CGF.Builder.CreateInsertElement(Upper, Res, (uint64_t)0);
12207}
12208
12209static Value *EmitX86Muldq(CodeGenFunction &CGF, bool IsSigned,
12210 ArrayRef<Value *> Ops) {
12211 llvm::Type *Ty = Ops[0]->getType();
12212 // Arguments have a vXi32 type so cast to vXi64.
12213 Ty = llvm::FixedVectorType::get(CGF.Int64Ty,
12214 Ty->getPrimitiveSizeInBits() / 64);
12215 Value *LHS = CGF.Builder.CreateBitCast(Ops[0], Ty);
12216 Value *RHS = CGF.Builder.CreateBitCast(Ops[1], Ty);
12217
12218 if (IsSigned) {
12219 // Shift left then arithmetic shift right.
12220 Constant *ShiftAmt = ConstantInt::get(Ty, 32);
12221 LHS = CGF.Builder.CreateShl(LHS, ShiftAmt);
12222 LHS = CGF.Builder.CreateAShr(LHS, ShiftAmt);
12223 RHS = CGF.Builder.CreateShl(RHS, ShiftAmt);
12224 RHS = CGF.Builder.CreateAShr(RHS, ShiftAmt);
12225 } else {
12226 // Clear the upper bits.
12227 Constant *Mask = ConstantInt::get(Ty, 0xffffffff);
12228 LHS = CGF.Builder.CreateAnd(LHS, Mask);
12229 RHS = CGF.Builder.CreateAnd(RHS, Mask);
12230 }
12231
12232 return CGF.Builder.CreateMul(LHS, RHS);
12233}
12234
12235// Emit a masked pternlog intrinsic. This only exists because the header has to
12236// use a macro and we aren't able to pass the input argument to a pternlog
12237// builtin and a select builtin without evaluating it twice.
12238static Value *EmitX86Ternlog(CodeGenFunction &CGF, bool ZeroMask,
12239 ArrayRef<Value *> Ops) {
12240 llvm::Type *Ty = Ops[0]->getType();
12241
12242 unsigned VecWidth = Ty->getPrimitiveSizeInBits();
12243 unsigned EltWidth = Ty->getScalarSizeInBits();
12244 Intrinsic::ID IID;
12245 if (VecWidth == 128 && EltWidth == 32)
12246 IID = Intrinsic::x86_avx512_pternlog_d_128;
12247 else if (VecWidth == 256 && EltWidth == 32)
12248 IID = Intrinsic::x86_avx512_pternlog_d_256;
12249 else if (VecWidth == 512 && EltWidth == 32)
12250 IID = Intrinsic::x86_avx512_pternlog_d_512;
12251 else if (VecWidth == 128 && EltWidth == 64)
12252 IID = Intrinsic::x86_avx512_pternlog_q_128;
12253 else if (VecWidth == 256 && EltWidth == 64)
12254 IID = Intrinsic::x86_avx512_pternlog_q_256;
12255 else if (VecWidth == 512 && EltWidth == 64)
12256 IID = Intrinsic::x86_avx512_pternlog_q_512;
12257 else
12258 llvm_unreachable("Unexpected intrinsic")__builtin_unreachable();
12259
12260 Value *Ternlog = CGF.Builder.CreateCall(CGF.CGM.getIntrinsic(IID),
12261 Ops.drop_back());
12262 Value *PassThru = ZeroMask ? ConstantAggregateZero::get(Ty) : Ops[0];
12263 return EmitX86Select(CGF, Ops[4], Ternlog, PassThru);
12264}
12265
12266static Value *EmitX86SExtMask(CodeGenFunction &CGF, Value *Op,
12267 llvm::Type *DstTy) {
12268 unsigned NumberOfElements =
12269 cast<llvm::FixedVectorType>(DstTy)->getNumElements();
12270 Value *Mask = getMaskVecValue(CGF, Op, NumberOfElements);
12271 return CGF.Builder.CreateSExt(Mask, DstTy, "vpmovm2");
12272}
12273
12274// Emit binary intrinsic with the same type used in result/args.
12275static Value *EmitX86BinaryIntrinsic(CodeGenFunction &CGF,
12276 ArrayRef<Value *> Ops, Intrinsic::ID IID) {
12277 llvm::Function *F = CGF.CGM.getIntrinsic(IID, Ops[0]->getType());
12278 return CGF.Builder.CreateCall(F, {Ops[0], Ops[1]});
12279}
12280
12281Value *CodeGenFunction::EmitX86CpuIs(const CallExpr *E) {
12282 const Expr *CPUExpr = E->getArg(0)->IgnoreParenCasts();
12283 StringRef CPUStr = cast<clang::StringLiteral>(CPUExpr)->getString();
12284 return EmitX86CpuIs(CPUStr);
12285}
12286
12287// Convert F16 halfs to floats.
12288static Value *EmitX86CvtF16ToFloatExpr(CodeGenFunction &CGF,
12289 ArrayRef<Value *> Ops,
12290 llvm::Type *DstTy) {
12291 assert((Ops.size() == 1 || Ops.size() == 3 || Ops.size() == 4) &&((void)0)
12292 "Unknown cvtph2ps intrinsic")((void)0);
12293
12294 // If the SAE intrinsic doesn't use default rounding then we can't upgrade.
12295 if (Ops.size() == 4 && cast<llvm::ConstantInt>(Ops[3])->getZExtValue() != 4) {
12296 Function *F =
12297 CGF.CGM.getIntrinsic(Intrinsic::x86_avx512_mask_vcvtph2ps_512);
12298 return CGF.Builder.CreateCall(F, {Ops[0], Ops[1], Ops[2], Ops[3]});
12299 }
12300
12301 unsigned NumDstElts = cast<llvm::FixedVectorType>(DstTy)->getNumElements();
12302 Value *Src = Ops[0];
12303
12304 // Extract the subvector.
12305 if (NumDstElts !=
12306 cast<llvm::FixedVectorType>(Src->getType())->getNumElements()) {
12307 assert(NumDstElts == 4 && "Unexpected vector size")((void)0);
12308 Src = CGF.Builder.CreateShuffleVector(Src, ArrayRef<int>{0, 1, 2, 3});
12309 }
12310
12311 // Bitcast from vXi16 to vXf16.
12312 auto *HalfTy = llvm::FixedVectorType::get(
12313 llvm::Type::getHalfTy(CGF.getLLVMContext()), NumDstElts);
12314 Src = CGF.Builder.CreateBitCast(Src, HalfTy);
12315
12316 // Perform the fp-extension.
12317 Value *Res = CGF.Builder.CreateFPExt(Src, DstTy, "cvtph2ps");
12318
12319 if (Ops.size() >= 3)
12320 Res = EmitX86Select(CGF, Ops[2], Res, Ops[1]);
12321 return Res;
12322}
12323
12324// Convert a BF16 to a float.
12325static Value *EmitX86CvtBF16ToFloatExpr(CodeGenFunction &CGF,
12326 const CallExpr *E,
12327 ArrayRef<Value *> Ops) {
12328 llvm::Type *Int32Ty = CGF.Builder.getInt32Ty();
12329 Value *ZeroExt = CGF.Builder.CreateZExt(Ops[0], Int32Ty);
12330 Value *Shl = CGF.Builder.CreateShl(ZeroExt, 16);
12331 llvm::Type *ResultType = CGF.ConvertType(E->getType());
12332 Value *BitCast = CGF.Builder.CreateBitCast(Shl, ResultType);
12333 return BitCast;
12334}
12335
12336Value *CodeGenFunction::EmitX86CpuIs(StringRef CPUStr) {
12337
12338 llvm::Type *Int32Ty = Builder.getInt32Ty();
12339
12340 // Matching the struct layout from the compiler-rt/libgcc structure that is
12341 // filled in:
12342 // unsigned int __cpu_vendor;
12343 // unsigned int __cpu_type;
12344 // unsigned int __cpu_subtype;
12345 // unsigned int __cpu_features[1];
12346 llvm::Type *STy = llvm::StructType::get(Int32Ty, Int32Ty, Int32Ty,
12347 llvm::ArrayType::get(Int32Ty, 1));
12348
12349 // Grab the global __cpu_model.
12350 llvm::Constant *CpuModel = CGM.CreateRuntimeVariable(STy, "__cpu_model");
12351 cast<llvm::GlobalValue>(CpuModel)->setDSOLocal(true);
12352
12353 // Calculate the index needed to access the correct field based on the
12354 // range. Also adjust the expected value.
12355 unsigned Index;
12356 unsigned Value;
12357 std::tie(Index, Value) = StringSwitch<std::pair<unsigned, unsigned>>(CPUStr)
12358#define X86_VENDOR(ENUM, STRING) \
12359 .Case(STRING, {0u, static_cast<unsigned>(llvm::X86::ENUM)})
12360#define X86_CPU_TYPE_ALIAS(ENUM, ALIAS) \
12361 .Case(ALIAS, {1u, static_cast<unsigned>(llvm::X86::ENUM)})
12362#define X86_CPU_TYPE(ENUM, STR) \
12363 .Case(STR, {1u, static_cast<unsigned>(llvm::X86::ENUM)})
12364#define X86_CPU_SUBTYPE(ENUM, STR) \
12365 .Case(STR, {2u, static_cast<unsigned>(llvm::X86::ENUM)})
12366#include "llvm/Support/X86TargetParser.def"
12367 .Default({0, 0});
12368 assert(Value != 0 && "Invalid CPUStr passed to CpuIs")((void)0);
12369
12370 // Grab the appropriate field from __cpu_model.
12371 llvm::Value *Idxs[] = {ConstantInt::get(Int32Ty, 0),
12372 ConstantInt::get(Int32Ty, Index)};
12373 llvm::Value *CpuValue = Builder.CreateGEP(STy, CpuModel, Idxs);
12374 CpuValue = Builder.CreateAlignedLoad(Int32Ty, CpuValue,
12375 CharUnits::fromQuantity(4));
12376
12377 // Check the value of the field against the requested value.
12378 return Builder.CreateICmpEQ(CpuValue,
12379 llvm::ConstantInt::get(Int32Ty, Value));
12380}
12381
12382Value *CodeGenFunction::EmitX86CpuSupports(const CallExpr *E) {
12383 const Expr *FeatureExpr = E->getArg(0)->IgnoreParenCasts();
12384 StringRef FeatureStr = cast<StringLiteral>(FeatureExpr)->getString();
12385 return EmitX86CpuSupports(FeatureStr);
12386}
12387
12388uint64_t
12389CodeGenFunction::GetX86CpuSupportsMask(ArrayRef<StringRef> FeatureStrs) {
12390 // Processor features and mapping to processor feature value.
12391 uint64_t FeaturesMask = 0;
12392 for (const StringRef &FeatureStr : FeatureStrs) {
12393 unsigned Feature =
12394 StringSwitch<unsigned>(FeatureStr)
12395#define X86_FEATURE_COMPAT(ENUM, STR) .Case(STR, llvm::X86::FEATURE_##ENUM)
12396#include "llvm/Support/X86TargetParser.def"
12397 ;
12398 FeaturesMask |= (1ULL << Feature);
12399 }
12400 return FeaturesMask;
12401}
12402
12403Value *CodeGenFunction::EmitX86CpuSupports(ArrayRef<StringRef> FeatureStrs) {
12404 return EmitX86CpuSupports(GetX86CpuSupportsMask(FeatureStrs));
12405}
12406
12407llvm::Value *CodeGenFunction::EmitX86CpuSupports(uint64_t FeaturesMask) {
12408 uint32_t Features1 = Lo_32(FeaturesMask);
12409 uint32_t Features2 = Hi_32(FeaturesMask);
12410
12411 Value *Result = Builder.getTrue();
12412
12413 if (Features1 != 0) {
12414 // Matching the struct layout from the compiler-rt/libgcc structure that is
12415 // filled in:
12416 // unsigned int __cpu_vendor;
12417 // unsigned int __cpu_type;
12418 // unsigned int __cpu_subtype;
12419 // unsigned int __cpu_features[1];
12420 llvm::Type *STy = llvm::StructType::get(Int32Ty, Int32Ty, Int32Ty,
12421 llvm::ArrayType::get(Int32Ty, 1));
12422
12423 // Grab the global __cpu_model.
12424 llvm::Constant *CpuModel = CGM.CreateRuntimeVariable(STy, "__cpu_model");
12425 cast<llvm::GlobalValue>(CpuModel)->setDSOLocal(true);
12426
12427 // Grab the first (0th) element from the field __cpu_features off of the
12428 // global in the struct STy.
12429 Value *Idxs[] = {Builder.getInt32(0), Builder.getInt32(3),
12430 Builder.getInt32(0)};
12431 Value *CpuFeatures = Builder.CreateGEP(STy, CpuModel, Idxs);
12432 Value *Features = Builder.CreateAlignedLoad(Int32Ty, CpuFeatures,
12433 CharUnits::fromQuantity(4));
12434
12435 // Check the value of the bit corresponding to the feature requested.
12436 Value *Mask = Builder.getInt32(Features1);
12437 Value *Bitset = Builder.CreateAnd(Features, Mask);
12438 Value *Cmp = Builder.CreateICmpEQ(Bitset, Mask);
12439 Result = Builder.CreateAnd(Result, Cmp);
12440 }
12441
12442 if (Features2 != 0) {
12443 llvm::Constant *CpuFeatures2 = CGM.CreateRuntimeVariable(Int32Ty,
12444 "__cpu_features2");
12445 cast<llvm::GlobalValue>(CpuFeatures2)->setDSOLocal(true);
12446
12447 Value *Features = Builder.CreateAlignedLoad(Int32Ty, CpuFeatures2,
12448 CharUnits::fromQuantity(4));
12449
12450 // Check the value of the bit corresponding to the feature requested.
12451 Value *Mask = Builder.getInt32(Features2);
12452 Value *Bitset = Builder.CreateAnd(Features, Mask);
12453 Value *Cmp = Builder.CreateICmpEQ(Bitset, Mask);
12454 Result = Builder.CreateAnd(Result, Cmp);
12455 }
12456
12457 return Result;
12458}
12459
12460Value *CodeGenFunction::EmitX86CpuInit() {
12461 llvm::FunctionType *FTy = llvm::FunctionType::get(VoidTy,
12462 /*Variadic*/ false);
12463 llvm::FunctionCallee Func =
12464 CGM.CreateRuntimeFunction(FTy, "__cpu_indicator_init");
12465 cast<llvm::GlobalValue>(Func.getCallee())->setDSOLocal(true);
12466 cast<llvm::GlobalValue>(Func.getCallee())
12467 ->setDLLStorageClass(llvm::GlobalValue::DefaultStorageClass);
12468 return Builder.CreateCall(Func);
12469}
12470
12471Value *CodeGenFunction::EmitX86BuiltinExpr(unsigned BuiltinID,
12472 const CallExpr *E) {
12473 if (BuiltinID == X86::BI__builtin_cpu_is)
12474 return EmitX86CpuIs(E);
12475 if (BuiltinID == X86::BI__builtin_cpu_supports)
12476 return EmitX86CpuSupports(E);
12477 if (BuiltinID == X86::BI__builtin_cpu_init)
12478 return EmitX86CpuInit();
12479
12480 // Handle MSVC intrinsics before argument evaluation to prevent double
12481 // evaluation.
12482 if (Optional<MSVCIntrin> MsvcIntId = translateX86ToMsvcIntrin(BuiltinID))
12483 return EmitMSVCBuiltinExpr(*MsvcIntId, E);
12484
12485 SmallVector<Value*, 4> Ops;
12486 bool IsMaskFCmp = false;
12487
12488 // Find out if any arguments are required to be integer constant expressions.
12489 unsigned ICEArguments = 0;
12490 ASTContext::GetBuiltinTypeError Error;
12491 getContext().GetBuiltinType(BuiltinID, Error, &ICEArguments);
12492 assert(Error == ASTContext::GE_None && "Should not codegen an error")((void)0);
12493
12494 for (unsigned i = 0, e = E->getNumArgs(); i != e; i++) {
12495 // If this is a normal argument, just emit it as a scalar.
12496 if ((ICEArguments & (1 << i)) == 0) {
12497 Ops.push_back(EmitScalarExpr(E->getArg(i)));
12498 continue;
12499 }
12500
12501 // If this is required to be a constant, constant fold it so that we know
12502 // that the generated intrinsic gets a ConstantInt.
12503 Ops.push_back(llvm::ConstantInt::get(
12504 getLLVMContext(), *E->getArg(i)->getIntegerConstantExpr(getContext())));
12505 }
12506
12507 // These exist so that the builtin that takes an immediate can be bounds
12508 // checked by clang to avoid passing bad immediates to the backend. Since
12509 // AVX has a larger immediate than SSE we would need separate builtins to
12510 // do the different bounds checking. Rather than create a clang specific
12511 // SSE only builtin, this implements eight separate builtins to match gcc
12512 // implementation.
12513 auto getCmpIntrinsicCall = [this, &Ops](Intrinsic::ID ID, unsigned Imm) {
12514 Ops.push_back(llvm::ConstantInt::get(Int8Ty, Imm));
12515 llvm::Function *F = CGM.getIntrinsic(ID);
12516 return Builder.CreateCall(F, Ops);
12517 };
12518
12519 // For the vector forms of FP comparisons, translate the builtins directly to
12520 // IR.
12521 // TODO: The builtins could be removed if the SSE header files used vector
12522 // extension comparisons directly (vector ordered/unordered may need
12523 // additional support via __builtin_isnan()).
12524 auto getVectorFCmpIR = [this, &Ops, E](CmpInst::Predicate Pred,
12525 bool IsSignaling) {
12526 CodeGenFunction::CGFPOptionsRAII FPOptsRAII(*this, E);
12527 Value *Cmp;
12528 if (IsSignaling)
12529 Cmp = Builder.CreateFCmpS(Pred, Ops[0], Ops[1]);
12530 else
12531 Cmp = Builder.CreateFCmp(Pred, Ops[0], Ops[1]);
12532 llvm::VectorType *FPVecTy = cast<llvm::VectorType>(Ops[0]->getType());
12533 llvm::VectorType *IntVecTy = llvm::VectorType::getInteger(FPVecTy);
12534 Value *Sext = Builder.CreateSExt(Cmp, IntVecTy);
12535 return Builder.CreateBitCast(Sext, FPVecTy);
12536 };
12537
12538 switch (BuiltinID) {
12539 default: return nullptr;
12540 case X86::BI_mm_prefetch: {
12541 Value *Address = Ops[0];
12542 ConstantInt *C = cast<ConstantInt>(Ops[1]);
12543 Value *RW = ConstantInt::get(Int32Ty, (C->getZExtValue() >> 2) & 0x1);
12544 Value *Locality = ConstantInt::get(Int32Ty, C->getZExtValue() & 0x3);
12545 Value *Data = ConstantInt::get(Int32Ty, 1);
12546 Function *F = CGM.getIntrinsic(Intrinsic::prefetch, Address->getType());
12547 return Builder.CreateCall(F, {Address, RW, Locality, Data});
12548 }
12549 case X86::BI_mm_clflush: {
12550 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_sse2_clflush),
12551 Ops[0]);
12552 }
12553 case X86::BI_mm_lfence: {
12554 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_sse2_lfence));
12555 }
12556 case X86::BI_mm_mfence: {
12557 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_sse2_mfence));
12558 }
12559 case X86::BI_mm_sfence: {
12560 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_sse_sfence));
12561 }
12562 case X86::BI_mm_pause: {
12563 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_sse2_pause));
12564 }
12565 case X86::BI__rdtsc: {
12566 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_rdtsc));
12567 }
12568 case X86::BI__builtin_ia32_rdtscp: {
12569 Value *Call = Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_rdtscp));
12570 Builder.CreateDefaultAlignedStore(Builder.CreateExtractValue(Call, 1),
12571 Ops[0]);
12572 return Builder.CreateExtractValue(Call, 0);
12573 }
12574 case X86::BI__builtin_ia32_lzcnt_u16:
12575 case X86::BI__builtin_ia32_lzcnt_u32:
12576 case X86::BI__builtin_ia32_lzcnt_u64: {
12577 Function *F = CGM.getIntrinsic(Intrinsic::ctlz, Ops[0]->getType());
12578 return Builder.CreateCall(F, {Ops[0], Builder.getInt1(false)});
12579 }
12580 case X86::BI__builtin_ia32_tzcnt_u16:
12581 case X86::BI__builtin_ia32_tzcnt_u32:
12582 case X86::BI__builtin_ia32_tzcnt_u64: {
12583 Function *F = CGM.getIntrinsic(Intrinsic::cttz, Ops[0]->getType());
12584 return Builder.CreateCall(F, {Ops[0], Builder.getInt1(false)});
12585 }
12586 case X86::BI__builtin_ia32_undef128:
12587 case X86::BI__builtin_ia32_undef256:
12588 case X86::BI__builtin_ia32_undef512:
12589 // The x86 definition of "undef" is not the same as the LLVM definition
12590 // (PR32176). We leave optimizing away an unnecessary zero constant to the
12591 // IR optimizer and backend.
12592 // TODO: If we had a "freeze" IR instruction to generate a fixed undef
12593 // value, we should use that here instead of a zero.
12594 return llvm::Constant::getNullValue(ConvertType(E->getType()));
12595 case X86::BI__builtin_ia32_vec_init_v8qi:
12596 case X86::BI__builtin_ia32_vec_init_v4hi:
12597 case X86::BI__builtin_ia32_vec_init_v2si:
12598 return Builder.CreateBitCast(BuildVector(Ops),
12599 llvm::Type::getX86_MMXTy(getLLVMContext()));
12600 case X86::BI__builtin_ia32_vec_ext_v2si:
12601 case X86::BI__builtin_ia32_vec_ext_v16qi:
12602 case X86::BI__builtin_ia32_vec_ext_v8hi:
12603 case X86::BI__builtin_ia32_vec_ext_v4si:
12604 case X86::BI__builtin_ia32_vec_ext_v4sf:
12605 case X86::BI__builtin_ia32_vec_ext_v2di:
12606 case X86::BI__builtin_ia32_vec_ext_v32qi:
12607 case X86::BI__builtin_ia32_vec_ext_v16hi:
12608 case X86::BI__builtin_ia32_vec_ext_v8si:
12609 case X86::BI__builtin_ia32_vec_ext_v4di: {
12610 unsigned NumElts =
12611 cast<llvm::FixedVectorType>(Ops[0]->getType())->getNumElements();
12612 uint64_t Index = cast<ConstantInt>(Ops[1])->getZExtValue();
12613 Index &= NumElts - 1;
12614 // These builtins exist so we can ensure the index is an ICE and in range.
12615 // Otherwise we could just do this in the header file.
12616 return Builder.CreateExtractElement(Ops[0], Index);
12617 }
12618 case X86::BI__builtin_ia32_vec_set_v16qi:
12619 case X86::BI__builtin_ia32_vec_set_v8hi:
12620 case X86::BI__builtin_ia32_vec_set_v4si:
12621 case X86::BI__builtin_ia32_vec_set_v2di:
12622 case X86::BI__builtin_ia32_vec_set_v32qi:
12623 case X86::BI__builtin_ia32_vec_set_v16hi:
12624 case X86::BI__builtin_ia32_vec_set_v8si:
12625 case X86::BI__builtin_ia32_vec_set_v4di: {
12626 unsigned NumElts =
12627 cast<llvm::FixedVectorType>(Ops[0]->getType())->getNumElements();
12628 unsigned Index = cast<ConstantInt>(Ops[2])->getZExtValue();
12629 Index &= NumElts - 1;
12630 // These builtins exist so we can ensure the index is an ICE and in range.
12631 // Otherwise we could just do this in the header file.
12632 return Builder.CreateInsertElement(Ops[0], Ops[1], Index);
12633 }
12634 case X86::BI_mm_setcsr:
12635 case X86::BI__builtin_ia32_ldmxcsr: {
12636 Address Tmp = CreateMemTemp(E->getArg(0)->getType());
12637 Builder.CreateStore(Ops[0], Tmp);
12638 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_sse_ldmxcsr),
12639 Builder.CreateBitCast(Tmp.getPointer(), Int8PtrTy));
12640 }
12641 case X86::BI_mm_getcsr:
12642 case X86::BI__builtin_ia32_stmxcsr: {
12643 Address Tmp = CreateMemTemp(E->getType());
12644 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_sse_stmxcsr),
12645 Builder.CreateBitCast(Tmp.getPointer(), Int8PtrTy));
12646 return Builder.CreateLoad(Tmp, "stmxcsr");
12647 }
12648 case X86::BI__builtin_ia32_xsave:
12649 case X86::BI__builtin_ia32_xsave64:
12650 case X86::BI__builtin_ia32_xrstor:
12651 case X86::BI__builtin_ia32_xrstor64:
12652 case X86::BI__builtin_ia32_xsaveopt:
12653 case X86::BI__builtin_ia32_xsaveopt64:
12654 case X86::BI__builtin_ia32_xrstors:
12655 case X86::BI__builtin_ia32_xrstors64:
12656 case X86::BI__builtin_ia32_xsavec:
12657 case X86::BI__builtin_ia32_xsavec64:
12658 case X86::BI__builtin_ia32_xsaves:
12659 case X86::BI__builtin_ia32_xsaves64:
12660 case X86::BI__builtin_ia32_xsetbv:
12661 case X86::BI_xsetbv: {
12662 Intrinsic::ID ID;
12663#define INTRINSIC_X86_XSAVE_ID(NAME) \
12664 case X86::BI__builtin_ia32_##NAME: \
12665 ID = Intrinsic::x86_##NAME; \
12666 break
12667 switch (BuiltinID) {
12668 default: llvm_unreachable("Unsupported intrinsic!")__builtin_unreachable();
12669 INTRINSIC_X86_XSAVE_ID(xsave);
12670 INTRINSIC_X86_XSAVE_ID(xsave64);
12671 INTRINSIC_X86_XSAVE_ID(xrstor);
12672 INTRINSIC_X86_XSAVE_ID(xrstor64);
12673 INTRINSIC_X86_XSAVE_ID(xsaveopt);
12674 INTRINSIC_X86_XSAVE_ID(xsaveopt64);
12675 INTRINSIC_X86_XSAVE_ID(xrstors);
12676 INTRINSIC_X86_XSAVE_ID(xrstors64);
12677 INTRINSIC_X86_XSAVE_ID(xsavec);
12678 INTRINSIC_X86_XSAVE_ID(xsavec64);
12679 INTRINSIC_X86_XSAVE_ID(xsaves);
12680 INTRINSIC_X86_XSAVE_ID(xsaves64);
12681 INTRINSIC_X86_XSAVE_ID(xsetbv);
12682 case X86::BI_xsetbv:
12683 ID = Intrinsic::x86_xsetbv;
12684 break;
12685 }
12686#undef INTRINSIC_X86_XSAVE_ID
12687 Value *Mhi = Builder.CreateTrunc(
12688 Builder.CreateLShr(Ops[1], ConstantInt::get(Int64Ty, 32)), Int32Ty);
12689 Value *Mlo = Builder.CreateTrunc(Ops[1], Int32Ty);
12690 Ops[1] = Mhi;
12691 Ops.push_back(Mlo);
12692 return Builder.CreateCall(CGM.getIntrinsic(ID), Ops);
12693 }
12694 case X86::BI__builtin_ia32_xgetbv:
12695 case X86::BI_xgetbv:
12696 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_xgetbv), Ops);
12697 case X86::BI__builtin_ia32_storedqudi128_mask:
12698 case X86::BI__builtin_ia32_storedqusi128_mask:
12699 case X86::BI__builtin_ia32_storedquhi128_mask:
12700 case X86::BI__builtin_ia32_storedquqi128_mask:
12701 case X86::BI__builtin_ia32_storeupd128_mask:
12702 case X86::BI__builtin_ia32_storeups128_mask:
12703 case X86::BI__builtin_ia32_storedqudi256_mask:
12704 case X86::BI__builtin_ia32_storedqusi256_mask:
12705 case X86::BI__builtin_ia32_storedquhi256_mask:
12706 case X86::BI__builtin_ia32_storedquqi256_mask:
12707 case X86::BI__builtin_ia32_storeupd256_mask:
12708 case X86::BI__builtin_ia32_storeups256_mask:
12709 case X86::BI__builtin_ia32_storedqudi512_mask:
12710 case X86::BI__builtin_ia32_storedqusi512_mask:
12711 case X86::BI__builtin_ia32_storedquhi512_mask:
12712 case X86::BI__builtin_ia32_storedquqi512_mask:
12713 case X86::BI__builtin_ia32_storeupd512_mask:
12714 case X86::BI__builtin_ia32_storeups512_mask:
12715 return EmitX86MaskedStore(*this, Ops, Align(1));
12716
12717 case X86::BI__builtin_ia32_storess128_mask:
12718 case X86::BI__builtin_ia32_storesd128_mask:
12719 return EmitX86MaskedStore(*this, Ops, Align(1));
12720
12721 case X86::BI__builtin_ia32_vpopcntb_128:
12722 case X86::BI__builtin_ia32_vpopcntd_128:
12723 case X86::BI__builtin_ia32_vpopcntq_128:
12724 case X86::BI__builtin_ia32_vpopcntw_128:
12725 case X86::BI__builtin_ia32_vpopcntb_256:
12726 case X86::BI__builtin_ia32_vpopcntd_256:
12727 case X86::BI__builtin_ia32_vpopcntq_256:
12728 case X86::BI__builtin_ia32_vpopcntw_256:
12729 case X86::BI__builtin_ia32_vpopcntb_512:
12730 case X86::BI__builtin_ia32_vpopcntd_512:
12731 case X86::BI__builtin_ia32_vpopcntq_512:
12732 case X86::BI__builtin_ia32_vpopcntw_512: {
12733 llvm::Type *ResultType = ConvertType(E->getType());
12734 llvm::Function *F = CGM.getIntrinsic(Intrinsic::ctpop, ResultType);
12735 return Builder.CreateCall(F, Ops);
12736 }
12737 case X86::BI__builtin_ia32_cvtmask2b128:
12738 case X86::BI__builtin_ia32_cvtmask2b256:
12739 case X86::BI__builtin_ia32_cvtmask2b512:
12740 case X86::BI__builtin_ia32_cvtmask2w128:
12741 case X86::BI__builtin_ia32_cvtmask2w256:
12742 case X86::BI__builtin_ia32_cvtmask2w512:
12743 case X86::BI__builtin_ia32_cvtmask2d128:
12744 case X86::BI__builtin_ia32_cvtmask2d256:
12745 case X86::BI__builtin_ia32_cvtmask2d512:
12746 case X86::BI__builtin_ia32_cvtmask2q128:
12747 case X86::BI__builtin_ia32_cvtmask2q256:
12748 case X86::BI__builtin_ia32_cvtmask2q512:
12749 return EmitX86SExtMask(*this, Ops[0], ConvertType(E->getType()));
12750
12751 case X86::BI__builtin_ia32_cvtb2mask128:
12752 case X86::BI__builtin_ia32_cvtb2mask256:
12753 case X86::BI__builtin_ia32_cvtb2mask512:
12754 case X86::BI__builtin_ia32_cvtw2mask128:
12755 case X86::BI__builtin_ia32_cvtw2mask256:
12756 case X86::BI__builtin_ia32_cvtw2mask512:
12757 case X86::BI__builtin_ia32_cvtd2mask128:
12758 case X86::BI__builtin_ia32_cvtd2mask256:
12759 case X86::BI__builtin_ia32_cvtd2mask512:
12760 case X86::BI__builtin_ia32_cvtq2mask128:
12761 case X86::BI__builtin_ia32_cvtq2mask256:
12762 case X86::BI__builtin_ia32_cvtq2mask512:
12763 return EmitX86ConvertToMask(*this, Ops[0]);
12764
12765 case X86::BI__builtin_ia32_cvtdq2ps512_mask:
12766 case X86::BI__builtin_ia32_cvtqq2ps512_mask:
12767 case X86::BI__builtin_ia32_cvtqq2pd512_mask:
12768 return EmitX86ConvertIntToFp(*this, E, Ops, /*IsSigned*/ true);
12769 case X86::BI__builtin_ia32_cvtudq2ps512_mask:
12770 case X86::BI__builtin_ia32_cvtuqq2ps512_mask:
12771 case X86::BI__builtin_ia32_cvtuqq2pd512_mask:
12772 return EmitX86ConvertIntToFp(*this, E, Ops, /*IsSigned*/ false);
12773
12774 case X86::BI__builtin_ia32_vfmaddss3:
12775 case X86::BI__builtin_ia32_vfmaddsd3:
12776 case X86::BI__builtin_ia32_vfmaddss3_mask:
12777 case X86::BI__builtin_ia32_vfmaddsd3_mask:
12778 return EmitScalarFMAExpr(*this, E, Ops, Ops[0]);
12779 case X86::BI__builtin_ia32_vfmaddss:
12780 case X86::BI__builtin_ia32_vfmaddsd:
12781 return EmitScalarFMAExpr(*this, E, Ops,
12782 Constant::getNullValue(Ops[0]->getType()));
12783 case X86::BI__builtin_ia32_vfmaddss3_maskz:
12784 case X86::BI__builtin_ia32_vfmaddsd3_maskz:
12785 return EmitScalarFMAExpr(*this, E, Ops, Ops[0], /*ZeroMask*/ true);
12786 case X86::BI__builtin_ia32_vfmaddss3_mask3:
12787 case X86::BI__builtin_ia32_vfmaddsd3_mask3:
12788 return EmitScalarFMAExpr(*this, E, Ops, Ops[2], /*ZeroMask*/ false, 2);
12789 case X86::BI__builtin_ia32_vfmsubss3_mask3:
12790 case X86::BI__builtin_ia32_vfmsubsd3_mask3:
12791 return EmitScalarFMAExpr(*this, E, Ops, Ops[2], /*ZeroMask*/ false, 2,
12792 /*NegAcc*/ true);
12793 case X86::BI__builtin_ia32_vfmaddps:
12794 case X86::BI__builtin_ia32_vfmaddpd:
12795 case X86::BI__builtin_ia32_vfmaddps256:
12796 case X86::BI__builtin_ia32_vfmaddpd256:
12797 case X86::BI__builtin_ia32_vfmaddps512_mask:
12798 case X86::BI__builtin_ia32_vfmaddps512_maskz:
12799 case X86::BI__builtin_ia32_vfmaddps512_mask3:
12800 case X86::BI__builtin_ia32_vfmsubps512_mask3:
12801 case X86::BI__builtin_ia32_vfmaddpd512_mask:
12802 case X86::BI__builtin_ia32_vfmaddpd512_maskz:
12803 case X86::BI__builtin_ia32_vfmaddpd512_mask3:
12804 case X86::BI__builtin_ia32_vfmsubpd512_mask3:
12805 return EmitX86FMAExpr(*this, E, Ops, BuiltinID, /*IsAddSub*/ false);
12806 case X86::BI__builtin_ia32_vfmaddsubps512_mask:
12807 case X86::BI__builtin_ia32_vfmaddsubps512_maskz:
12808 case X86::BI__builtin_ia32_vfmaddsubps512_mask3:
12809 case X86::BI__builtin_ia32_vfmsubaddps512_mask3:
12810 case X86::BI__builtin_ia32_vfmaddsubpd512_mask:
12811 case X86::BI__builtin_ia32_vfmaddsubpd512_maskz:
12812 case X86::BI__builtin_ia32_vfmaddsubpd512_mask3:
12813 case X86::BI__builtin_ia32_vfmsubaddpd512_mask3:
12814 return EmitX86FMAExpr(*this, E, Ops, BuiltinID, /*IsAddSub*/ true);
12815
12816 case X86::BI__builtin_ia32_movdqa32store128_mask:
12817 case X86::BI__builtin_ia32_movdqa64store128_mask:
12818 case X86::BI__builtin_ia32_storeaps128_mask:
12819 case X86::BI__builtin_ia32_storeapd128_mask:
12820 case X86::BI__builtin_ia32_movdqa32store256_mask:
12821 case X86::BI__builtin_ia32_movdqa64store256_mask:
12822 case X86::BI__builtin_ia32_storeaps256_mask:
12823 case X86::BI__builtin_ia32_storeapd256_mask:
12824 case X86::BI__builtin_ia32_movdqa32store512_mask:
12825 case X86::BI__builtin_ia32_movdqa64store512_mask:
12826 case X86::BI__builtin_ia32_storeaps512_mask:
12827 case X86::BI__builtin_ia32_storeapd512_mask:
12828 return EmitX86MaskedStore(
12829 *this, Ops,
12830 getContext().getTypeAlignInChars(E->getArg(1)->getType()).getAsAlign());
12831
12832 case X86::BI__builtin_ia32_loadups128_mask:
12833 case X86::BI__builtin_ia32_loadups256_mask:
12834 case X86::BI__builtin_ia32_loadups512_mask:
12835 case X86::BI__builtin_ia32_loadupd128_mask:
12836 case X86::BI__builtin_ia32_loadupd256_mask:
12837 case X86::BI__builtin_ia32_loadupd512_mask:
12838 case X86::BI__builtin_ia32_loaddquqi128_mask:
12839 case X86::BI__builtin_ia32_loaddquqi256_mask:
12840 case X86::BI__builtin_ia32_loaddquqi512_mask:
12841 case X86::BI__builtin_ia32_loaddquhi128_mask:
12842 case X86::BI__builtin_ia32_loaddquhi256_mask:
12843 case X86::BI__builtin_ia32_loaddquhi512_mask:
12844 case X86::BI__builtin_ia32_loaddqusi128_mask:
12845 case X86::BI__builtin_ia32_loaddqusi256_mask:
12846 case X86::BI__builtin_ia32_loaddqusi512_mask:
12847 case X86::BI__builtin_ia32_loaddqudi128_mask:
12848 case X86::BI__builtin_ia32_loaddqudi256_mask:
12849 case X86::BI__builtin_ia32_loaddqudi512_mask:
12850 return EmitX86MaskedLoad(*this, Ops, Align(1));
12851
12852 case X86::BI__builtin_ia32_loadss128_mask:
12853 case X86::BI__builtin_ia32_loadsd128_mask:
12854 return EmitX86MaskedLoad(*this, Ops, Align(1));
12855
12856 case X86::BI__builtin_ia32_loadaps128_mask:
12857 case X86::BI__builtin_ia32_loadaps256_mask:
12858 case X86::BI__builtin_ia32_loadaps512_mask:
12859 case X86::BI__builtin_ia32_loadapd128_mask:
12860 case X86::BI__builtin_ia32_loadapd256_mask:
12861 case X86::BI__builtin_ia32_loadapd512_mask:
12862 case X86::BI__builtin_ia32_movdqa32load128_mask:
12863 case X86::BI__builtin_ia32_movdqa32load256_mask:
12864 case X86::BI__builtin_ia32_movdqa32load512_mask:
12865 case X86::BI__builtin_ia32_movdqa64load128_mask:
12866 case X86::BI__builtin_ia32_movdqa64load256_mask:
12867 case X86::BI__builtin_ia32_movdqa64load512_mask:
12868 return EmitX86MaskedLoad(
12869 *this, Ops,
12870 getContext().getTypeAlignInChars(E->getArg(1)->getType()).getAsAlign());
12871
12872 case X86::BI__builtin_ia32_expandloaddf128_mask:
12873 case X86::BI__builtin_ia32_expandloaddf256_mask:
12874 case X86::BI__builtin_ia32_expandloaddf512_mask:
12875 case X86::BI__builtin_ia32_expandloadsf128_mask:
12876 case X86::BI__builtin_ia32_expandloadsf256_mask:
12877 case X86::BI__builtin_ia32_expandloadsf512_mask:
12878 case X86::BI__builtin_ia32_expandloaddi128_mask:
12879 case X86::BI__builtin_ia32_expandloaddi256_mask:
12880 case X86::BI__builtin_ia32_expandloaddi512_mask:
12881 case X86::BI__builtin_ia32_expandloadsi128_mask:
12882 case X86::BI__builtin_ia32_expandloadsi256_mask:
12883 case X86::BI__builtin_ia32_expandloadsi512_mask:
12884 case X86::BI__builtin_ia32_expandloadhi128_mask:
12885 case X86::BI__builtin_ia32_expandloadhi256_mask:
12886 case X86::BI__builtin_ia32_expandloadhi512_mask:
12887 case X86::BI__builtin_ia32_expandloadqi128_mask:
12888 case X86::BI__builtin_ia32_expandloadqi256_mask:
12889 case X86::BI__builtin_ia32_expandloadqi512_mask:
12890 return EmitX86ExpandLoad(*this, Ops);
12891
12892 case X86::BI__builtin_ia32_compressstoredf128_mask:
12893 case X86::BI__builtin_ia32_compressstoredf256_mask:
12894 case X86::BI__builtin_ia32_compressstoredf512_mask:
12895 case X86::BI__builtin_ia32_compressstoresf128_mask:
12896 case X86::BI__builtin_ia32_compressstoresf256_mask:
12897 case X86::BI__builtin_ia32_compressstoresf512_mask:
12898 case X86::BI__builtin_ia32_compressstoredi128_mask:
12899 case X86::BI__builtin_ia32_compressstoredi256_mask:
12900 case X86::BI__builtin_ia32_compressstoredi512_mask:
12901 case X86::BI__builtin_ia32_compressstoresi128_mask:
12902 case X86::BI__builtin_ia32_compressstoresi256_mask:
12903 case X86::BI__builtin_ia32_compressstoresi512_mask:
12904 case X86::BI__builtin_ia32_compressstorehi128_mask:
12905 case X86::BI__builtin_ia32_compressstorehi256_mask:
12906 case X86::BI__builtin_ia32_compressstorehi512_mask:
12907 case X86::BI__builtin_ia32_compressstoreqi128_mask:
12908 case X86::BI__builtin_ia32_compressstoreqi256_mask:
12909 case X86::BI__builtin_ia32_compressstoreqi512_mask:
12910 return EmitX86CompressStore(*this, Ops);
12911
12912 case X86::BI__builtin_ia32_expanddf128_mask:
12913 case X86::BI__builtin_ia32_expanddf256_mask:
12914 case X86::BI__builtin_ia32_expanddf512_mask:
12915 case X86::BI__builtin_ia32_expandsf128_mask:
12916 case X86::BI__builtin_ia32_expandsf256_mask:
12917 case X86::BI__builtin_ia32_expandsf512_mask:
12918 case X86::BI__builtin_ia32_expanddi128_mask:
12919 case X86::BI__builtin_ia32_expanddi256_mask:
12920 case X86::BI__builtin_ia32_expanddi512_mask:
12921 case X86::BI__builtin_ia32_expandsi128_mask:
12922 case X86::BI__builtin_ia32_expandsi256_mask:
12923 case X86::BI__builtin_ia32_expandsi512_mask:
12924 case X86::BI__builtin_ia32_expandhi128_mask:
12925 case X86::BI__builtin_ia32_expandhi256_mask:
12926 case X86::BI__builtin_ia32_expandhi512_mask:
12927 case X86::BI__builtin_ia32_expandqi128_mask:
12928 case X86::BI__builtin_ia32_expandqi256_mask:
12929 case X86::BI__builtin_ia32_expandqi512_mask:
12930 return EmitX86CompressExpand(*this, Ops, /*IsCompress*/false);
12931
12932 case X86::BI__builtin_ia32_compressdf128_mask:
12933 case X86::BI__builtin_ia32_compressdf256_mask:
12934 case X86::BI__builtin_ia32_compressdf512_mask:
12935 case X86::BI__builtin_ia32_compresssf128_mask:
12936 case X86::BI__builtin_ia32_compresssf256_mask:
12937 case X86::BI__builtin_ia32_compresssf512_mask:
12938 case X86::BI__builtin_ia32_compressdi128_mask:
12939 case X86::BI__builtin_ia32_compressdi256_mask:
12940 case X86::BI__builtin_ia32_compressdi512_mask:
12941 case X86::BI__builtin_ia32_compresssi128_mask:
12942 case X86::BI__builtin_ia32_compresssi256_mask:
12943 case X86::BI__builtin_ia32_compresssi512_mask:
12944 case X86::BI__builtin_ia32_compresshi128_mask:
12945 case X86::BI__builtin_ia32_compresshi256_mask:
12946 case X86::BI__builtin_ia32_compresshi512_mask:
12947 case X86::BI__builtin_ia32_compressqi128_mask:
12948 case X86::BI__builtin_ia32_compressqi256_mask:
12949 case X86::BI__builtin_ia32_compressqi512_mask:
12950 return EmitX86CompressExpand(*this, Ops, /*IsCompress*/true);
12951
12952 case X86::BI__builtin_ia32_gather3div2df:
12953 case X86::BI__builtin_ia32_gather3div2di:
12954 case X86::BI__builtin_ia32_gather3div4df:
12955 case X86::BI__builtin_ia32_gather3div4di:
12956 case X86::BI__builtin_ia32_gather3div4sf:
12957 case X86::BI__builtin_ia32_gather3div4si:
12958 case X86::BI__builtin_ia32_gather3div8sf:
12959 case X86::BI__builtin_ia32_gather3div8si:
12960 case X86::BI__builtin_ia32_gather3siv2df:
12961 case X86::BI__builtin_ia32_gather3siv2di:
12962 case X86::BI__builtin_ia32_gather3siv4df:
12963 case X86::BI__builtin_ia32_gather3siv4di:
12964 case X86::BI__builtin_ia32_gather3siv4sf:
12965 case X86::BI__builtin_ia32_gather3siv4si:
12966 case X86::BI__builtin_ia32_gather3siv8sf:
12967 case X86::BI__builtin_ia32_gather3siv8si:
12968 case X86::BI__builtin_ia32_gathersiv8df:
12969 case X86::BI__builtin_ia32_gathersiv16sf:
12970 case X86::BI__builtin_ia32_gatherdiv8df:
12971 case X86::BI__builtin_ia32_gatherdiv16sf:
12972 case X86::BI__builtin_ia32_gathersiv8di:
12973 case X86::BI__builtin_ia32_gathersiv16si:
12974 case X86::BI__builtin_ia32_gatherdiv8di:
12975 case X86::BI__builtin_ia32_gatherdiv16si: {
12976 Intrinsic::ID IID;
12977 switch (BuiltinID) {
12978 default: llvm_unreachable("Unexpected builtin")__builtin_unreachable();
12979 case X86::BI__builtin_ia32_gather3div2df:
12980 IID = Intrinsic::x86_avx512_mask_gather3div2_df;
12981 break;
12982 case X86::BI__builtin_ia32_gather3div2di:
12983 IID = Intrinsic::x86_avx512_mask_gather3div2_di;
12984 break;
12985 case X86::BI__builtin_ia32_gather3div4df:
12986 IID = Intrinsic::x86_avx512_mask_gather3div4_df;
12987 break;
12988 case X86::BI__builtin_ia32_gather3div4di:
12989 IID = Intrinsic::x86_avx512_mask_gather3div4_di;
12990 break;
12991 case X86::BI__builtin_ia32_gather3div4sf:
12992 IID = Intrinsic::x86_avx512_mask_gather3div4_sf;
12993 break;
12994 case X86::BI__builtin_ia32_gather3div4si:
12995 IID = Intrinsic::x86_avx512_mask_gather3div4_si;
12996 break;
12997 case X86::BI__builtin_ia32_gather3div8sf:
12998 IID = Intrinsic::x86_avx512_mask_gather3div8_sf;
12999 break;
13000 case X86::BI__builtin_ia32_gather3div8si:
13001 IID = Intrinsic::x86_avx512_mask_gather3div8_si;
13002 break;
13003 case X86::BI__builtin_ia32_gather3siv2df:
13004 IID = Intrinsic::x86_avx512_mask_gather3siv2_df;
13005 break;
13006 case X86::BI__builtin_ia32_gather3siv2di:
13007 IID = Intrinsic::x86_avx512_mask_gather3siv2_di;
13008 break;
13009 case X86::BI__builtin_ia32_gather3siv4df:
13010 IID = Intrinsic::x86_avx512_mask_gather3siv4_df;
13011 break;
13012 case X86::BI__builtin_ia32_gather3siv4di:
13013 IID = Intrinsic::x86_avx512_mask_gather3siv4_di;
13014 break;
13015 case X86::BI__builtin_ia32_gather3siv4sf:
13016 IID = Intrinsic::x86_avx512_mask_gather3siv4_sf;
13017 break;
13018 case X86::BI__builtin_ia32_gather3siv4si:
13019 IID = Intrinsic::x86_avx512_mask_gather3siv4_si;
13020 break;
13021 case X86::BI__builtin_ia32_gather3siv8sf:
13022 IID = Intrinsic::x86_avx512_mask_gather3siv8_sf;
13023 break;
13024 case X86::BI__builtin_ia32_gather3siv8si:
13025 IID = Intrinsic::x86_avx512_mask_gather3siv8_si;
13026 break;
13027 case X86::BI__builtin_ia32_gathersiv8df:
13028 IID = Intrinsic::x86_avx512_mask_gather_dpd_512;
13029 break;
13030 case X86::BI__builtin_ia32_gathersiv16sf:
13031 IID = Intrinsic::x86_avx512_mask_gather_dps_512;
13032 break;
13033 case X86::BI__builtin_ia32_gatherdiv8df:
13034 IID = Intrinsic::x86_avx512_mask_gather_qpd_512;
13035 break;
13036 case X86::BI__builtin_ia32_gatherdiv16sf:
13037 IID = Intrinsic::x86_avx512_mask_gather_qps_512;
13038 break;
13039 case X86::BI__builtin_ia32_gathersiv8di:
13040 IID = Intrinsic::x86_avx512_mask_gather_dpq_512;
13041 break;
13042 case X86::BI__builtin_ia32_gathersiv16si:
13043 IID = Intrinsic::x86_avx512_mask_gather_dpi_512;
13044 break;
13045 case X86::BI__builtin_ia32_gatherdiv8di:
13046 IID = Intrinsic::x86_avx512_mask_gather_qpq_512;
13047 break;
13048 case X86::BI__builtin_ia32_gatherdiv16si:
13049 IID = Intrinsic::x86_avx512_mask_gather_qpi_512;
13050 break;
13051 }
13052
13053 unsigned MinElts = std::min(
13054 cast<llvm::FixedVectorType>(Ops[0]->getType())->getNumElements(),
13055 cast<llvm::FixedVectorType>(Ops[2]->getType())->getNumElements());
13056 Ops[3] = getMaskVecValue(*this, Ops[3], MinElts);
13057 Function *Intr = CGM.getIntrinsic(IID);
13058 return Builder.CreateCall(Intr, Ops);
13059 }
13060
13061 case X86::BI__builtin_ia32_scattersiv8df:
13062 case X86::BI__builtin_ia32_scattersiv16sf:
13063 case X86::BI__builtin_ia32_scatterdiv8df:
13064 case X86::BI__builtin_ia32_scatterdiv16sf:
13065 case X86::BI__builtin_ia32_scattersiv8di:
13066 case X86::BI__builtin_ia32_scattersiv16si:
13067 case X86::BI__builtin_ia32_scatterdiv8di:
13068 case X86::BI__builtin_ia32_scatterdiv16si:
13069 case X86::BI__builtin_ia32_scatterdiv2df:
13070 case X86::BI__builtin_ia32_scatterdiv2di:
13071 case X86::BI__builtin_ia32_scatterdiv4df:
13072 case X86::BI__builtin_ia32_scatterdiv4di:
13073 case X86::BI__builtin_ia32_scatterdiv4sf:
13074 case X86::BI__builtin_ia32_scatterdiv4si:
13075 case X86::BI__builtin_ia32_scatterdiv8sf:
13076 case X86::BI__builtin_ia32_scatterdiv8si:
13077 case X86::BI__builtin_ia32_scattersiv2df:
13078 case X86::BI__builtin_ia32_scattersiv2di:
13079 case X86::BI__builtin_ia32_scattersiv4df:
13080 case X86::BI__builtin_ia32_scattersiv4di:
13081 case X86::BI__builtin_ia32_scattersiv4sf:
13082 case X86::BI__builtin_ia32_scattersiv4si:
13083 case X86::BI__builtin_ia32_scattersiv8sf:
13084 case X86::BI__builtin_ia32_scattersiv8si: {
13085 Intrinsic::ID IID;
13086 switch (BuiltinID) {
13087 default: llvm_unreachable("Unexpected builtin")__builtin_unreachable();
13088 case X86::BI__builtin_ia32_scattersiv8df:
13089 IID = Intrinsic::x86_avx512_mask_scatter_dpd_512;
13090 break;
13091 case X86::BI__builtin_ia32_scattersiv16sf:
13092 IID = Intrinsic::x86_avx512_mask_scatter_dps_512;
13093 break;
13094 case X86::BI__builtin_ia32_scatterdiv8df:
13095 IID = Intrinsic::x86_avx512_mask_scatter_qpd_512;
13096 break;
13097 case X86::BI__builtin_ia32_scatterdiv16sf:
13098 IID = Intrinsic::x86_avx512_mask_scatter_qps_512;
13099 break;
13100 case X86::BI__builtin_ia32_scattersiv8di:
13101 IID = Intrinsic::x86_avx512_mask_scatter_dpq_512;
13102 break;
13103 case X86::BI__builtin_ia32_scattersiv16si:
13104 IID = Intrinsic::x86_avx512_mask_scatter_dpi_512;
13105 break;
13106 case X86::BI__builtin_ia32_scatterdiv8di:
13107 IID = Intrinsic::x86_avx512_mask_scatter_qpq_512;
13108 break;
13109 case X86::BI__builtin_ia32_scatterdiv16si:
13110 IID = Intrinsic::x86_avx512_mask_scatter_qpi_512;
13111 break;
13112 case X86::BI__builtin_ia32_scatterdiv2df:
13113 IID = Intrinsic::x86_avx512_mask_scatterdiv2_df;
13114 break;
13115 case X86::BI__builtin_ia32_scatterdiv2di:
13116 IID = Intrinsic::x86_avx512_mask_scatterdiv2_di;
13117 break;
13118 case X86::BI__builtin_ia32_scatterdiv4df:
13119 IID = Intrinsic::x86_avx512_mask_scatterdiv4_df;
13120 break;
13121 case X86::BI__builtin_ia32_scatterdiv4di:
13122 IID = Intrinsic::x86_avx512_mask_scatterdiv4_di;
13123 break;
13124 case X86::BI__builtin_ia32_scatterdiv4sf:
13125 IID = Intrinsic::x86_avx512_mask_scatterdiv4_sf;
13126 break;
13127 case X86::BI__builtin_ia32_scatterdiv4si:
13128 IID = Intrinsic::x86_avx512_mask_scatterdiv4_si;
13129 break;
13130 case X86::BI__builtin_ia32_scatterdiv8sf:
13131 IID = Intrinsic::x86_avx512_mask_scatterdiv8_sf;
13132 break;
13133 case X86::BI__builtin_ia32_scatterdiv8si:
13134 IID = Intrinsic::x86_avx512_mask_scatterdiv8_si;
13135 break;
13136 case X86::BI__builtin_ia32_scattersiv2df:
13137 IID = Intrinsic::x86_avx512_mask_scattersiv2_df;
13138 break;
13139 case X86::BI__builtin_ia32_scattersiv2di:
13140 IID = Intrinsic::x86_avx512_mask_scattersiv2_di;
13141 break;
13142 case X86::BI__builtin_ia32_scattersiv4df:
13143 IID = Intrinsic::x86_avx512_mask_scattersiv4_df;
13144 break;
13145 case X86::BI__builtin_ia32_scattersiv4di:
13146 IID = Intrinsic::x86_avx512_mask_scattersiv4_di;
13147 break;
13148 case X86::BI__builtin_ia32_scattersiv4sf:
13149 IID = Intrinsic::x86_avx512_mask_scattersiv4_sf;
13150 break;
13151 case X86::BI__builtin_ia32_scattersiv4si:
13152 IID = Intrinsic::x86_avx512_mask_scattersiv4_si;
13153 break;
13154 case X86::BI__builtin_ia32_scattersiv8sf:
13155 IID = Intrinsic::x86_avx512_mask_scattersiv8_sf;
13156 break;
13157 case X86::BI__builtin_ia32_scattersiv8si:
13158 IID = Intrinsic::x86_avx512_mask_scattersiv8_si;
13159 break;
13160 }
13161
13162 unsigned MinElts = std::min(
13163 cast<llvm::FixedVectorType>(Ops[2]->getType())->getNumElements(),
13164 cast<llvm::FixedVectorType>(Ops[3]->getType())->getNumElements());
13165 Ops[1] = getMaskVecValue(*this, Ops[1], MinElts);
13166 Function *Intr = CGM.getIntrinsic(IID);
13167 return Builder.CreateCall(Intr, Ops);
13168 }
13169
13170 case X86::BI__builtin_ia32_vextractf128_pd256:
13171 case X86::BI__builtin_ia32_vextractf128_ps256:
13172 case X86::BI__builtin_ia32_vextractf128_si256:
13173 case X86::BI__builtin_ia32_extract128i256:
13174 case X86::BI__builtin_ia32_extractf64x4_mask:
13175 case X86::BI__builtin_ia32_extractf32x4_mask:
13176 case X86::BI__builtin_ia32_extracti64x4_mask:
13177 case X86::BI__builtin_ia32_extracti32x4_mask:
13178 case X86::BI__builtin_ia32_extractf32x8_mask:
13179 case X86::BI__builtin_ia32_extracti32x8_mask:
13180 case X86::BI__builtin_ia32_extractf32x4_256_mask:
13181 case X86::BI__builtin_ia32_extracti32x4_256_mask:
13182 case X86::BI__builtin_ia32_extractf64x2_256_mask:
13183 case X86::BI__builtin_ia32_extracti64x2_256_mask:
13184 case X86::BI__builtin_ia32_extractf64x2_512_mask:
13185 case X86::BI__builtin_ia32_extracti64x2_512_mask: {
13186 auto *DstTy = cast<llvm::FixedVectorType>(ConvertType(E->getType()));
13187 unsigned NumElts = DstTy->getNumElements();
13188 unsigned SrcNumElts =
13189 cast<llvm::FixedVectorType>(Ops[0]->getType())->getNumElements();
13190 unsigned SubVectors = SrcNumElts / NumElts;
13191 unsigned Index = cast<ConstantInt>(Ops[1])->getZExtValue();
13192 assert(llvm::isPowerOf2_32(SubVectors) && "Expected power of 2 subvectors")((void)0);
13193 Index &= SubVectors - 1; // Remove any extra bits.
13194 Index *= NumElts;
13195
13196 int Indices[16];
13197 for (unsigned i = 0; i != NumElts; ++i)
13198 Indices[i] = i + Index;
13199
13200 Value *Res = Builder.CreateShuffleVector(Ops[0],
13201 makeArrayRef(Indices, NumElts),
13202 "extract");
13203
13204 if (Ops.size() == 4)
13205 Res = EmitX86Select(*this, Ops[3], Res, Ops[2]);
13206
13207 return Res;
13208 }
13209 case X86::BI__builtin_ia32_vinsertf128_pd256:
13210 case X86::BI__builtin_ia32_vinsertf128_ps256:
13211 case X86::BI__builtin_ia32_vinsertf128_si256:
13212 case X86::BI__builtin_ia32_insert128i256:
13213 case X86::BI__builtin_ia32_insertf64x4:
13214 case X86::BI__builtin_ia32_insertf32x4:
13215 case X86::BI__builtin_ia32_inserti64x4:
13216 case X86::BI__builtin_ia32_inserti32x4:
13217 case X86::BI__builtin_ia32_insertf32x8:
13218 case X86::BI__builtin_ia32_inserti32x8:
13219 case X86::BI__builtin_ia32_insertf32x4_256:
13220 case X86::BI__builtin_ia32_inserti32x4_256:
13221 case X86::BI__builtin_ia32_insertf64x2_256:
13222 case X86::BI__builtin_ia32_inserti64x2_256:
13223 case X86::BI__builtin_ia32_insertf64x2_512:
13224 case X86::BI__builtin_ia32_inserti64x2_512: {
13225 unsigned DstNumElts =
13226 cast<llvm::FixedVectorType>(Ops[0]->getType())->getNumElements();
13227 unsigned SrcNumElts =
13228 cast<llvm::FixedVectorType>(Ops[1]->getType())->getNumElements();
13229 unsigned SubVectors = DstNumElts / SrcNumElts;
13230 unsigned Index = cast<ConstantInt>(Ops[2])->getZExtValue();
13231 assert(llvm::isPowerOf2_32(SubVectors) && "Expected power of 2 subvectors")((void)0);
13232 Index &= SubVectors - 1; // Remove any extra bits.
13233 Index *= SrcNumElts;
13234
13235 int Indices[16];
13236 for (unsigned i = 0; i != DstNumElts; ++i)
13237 Indices[i] = (i >= SrcNumElts) ? SrcNumElts + (i % SrcNumElts) : i;
13238
13239 Value *Op1 = Builder.CreateShuffleVector(Ops[1],
13240 makeArrayRef(Indices, DstNumElts),
13241 "widen");
13242
13243 for (unsigned i = 0; i != DstNumElts; ++i) {
13244 if (i >= Index && i < (Index + SrcNumElts))
13245 Indices[i] = (i - Index) + DstNumElts;
13246 else
13247 Indices[i] = i;
13248 }
13249
13250 return Builder.CreateShuffleVector(Ops[0], Op1,
13251 makeArrayRef(Indices, DstNumElts),
13252 "insert");
13253 }
13254 case X86::BI__builtin_ia32_pmovqd512_mask:
13255 case X86::BI__builtin_ia32_pmovwb512_mask: {
13256 Value *Res = Builder.CreateTrunc(Ops[0], Ops[1]->getType());
13257 return EmitX86Select(*this, Ops[2], Res, Ops[1]);
13258 }
13259 case X86::BI__builtin_ia32_pmovdb512_mask:
13260 case X86::BI__builtin_ia32_pmovdw512_mask:
13261 case X86::BI__builtin_ia32_pmovqw512_mask: {
13262 if (const auto *C = dyn_cast<Constant>(Ops[2]))
13263 if (C->isAllOnesValue())
13264 return Builder.CreateTrunc(Ops[0], Ops[1]->getType());
13265
13266 Intrinsic::ID IID;
13267 switch (BuiltinID) {
13268 default: llvm_unreachable("Unsupported intrinsic!")__builtin_unreachable();
13269 case X86::BI__builtin_ia32_pmovdb512_mask:
13270 IID = Intrinsic::x86_avx512_mask_pmov_db_512;
13271 break;
13272 case X86::BI__builtin_ia32_pmovdw512_mask:
13273 IID = Intrinsic::x86_avx512_mask_pmov_dw_512;
13274 break;
13275 case X86::BI__builtin_ia32_pmovqw512_mask:
13276 IID = Intrinsic::x86_avx512_mask_pmov_qw_512;
13277 break;
13278 }
13279
13280 Function *Intr = CGM.getIntrinsic(IID);
13281 return Builder.CreateCall(Intr, Ops);
13282 }
13283 case X86::BI__builtin_ia32_pblendw128:
13284 case X86::BI__builtin_ia32_blendpd:
13285 case X86::BI__builtin_ia32_blendps:
13286 case X86::BI__builtin_ia32_blendpd256:
13287 case X86::BI__builtin_ia32_blendps256:
13288 case X86::BI__builtin_ia32_pblendw256:
13289 case X86::BI__builtin_ia32_pblendd128:
13290 case X86::BI__builtin_ia32_pblendd256: {
13291 unsigned NumElts =
13292 cast<llvm::FixedVectorType>(Ops[0]->getType())->getNumElements();
13293 unsigned Imm = cast<llvm::ConstantInt>(Ops[2])->getZExtValue();
13294
13295 int Indices[16];
13296 // If there are more than 8 elements, the immediate is used twice so make
13297 // sure we handle that.
13298 for (unsigned i = 0; i != NumElts; ++i)
13299 Indices[i] = ((Imm >> (i % 8)) & 0x1) ? NumElts + i : i;
13300
13301 return Builder.CreateShuffleVector(Ops[0], Ops[1],
13302 makeArrayRef(Indices, NumElts),
13303 "blend");
13304 }
13305 case X86::BI__builtin_ia32_pshuflw:
13306 case X86::BI__builtin_ia32_pshuflw256:
13307 case X86::BI__builtin_ia32_pshuflw512: {
13308 uint32_t Imm = cast<llvm::ConstantInt>(Ops[1])->getZExtValue();
13309 auto *Ty = cast<llvm::FixedVectorType>(Ops[0]->getType());
13310 unsigned NumElts = Ty->getNumElements();
13311
13312 // Splat the 8-bits of immediate 4 times to help the loop wrap around.
13313 Imm = (Imm & 0xff) * 0x01010101;
13314
13315 int Indices[32];
13316 for (unsigned l = 0; l != NumElts; l += 8) {
13317 for (unsigned i = 0; i != 4; ++i) {
13318 Indices[l + i] = l + (Imm & 3);
13319 Imm >>= 2;
13320 }
13321 for (unsigned i = 4; i != 8; ++i)
13322 Indices[l + i] = l + i;
13323 }
13324
13325 return Builder.CreateShuffleVector(Ops[0], makeArrayRef(Indices, NumElts),
13326 "pshuflw");
13327 }
13328 case X86::BI__builtin_ia32_pshufhw:
13329 case X86::BI__builtin_ia32_pshufhw256:
13330 case X86::BI__builtin_ia32_pshufhw512: {
13331 uint32_t Imm = cast<llvm::ConstantInt>(Ops[1])->getZExtValue();
13332 auto *Ty = cast<llvm::FixedVectorType>(Ops[0]->getType());
13333 unsigned NumElts = Ty->getNumElements();
13334
13335 // Splat the 8-bits of immediate 4 times to help the loop wrap around.
13336 Imm = (Imm & 0xff) * 0x01010101;
13337
13338 int Indices[32];
13339 for (unsigned l = 0; l != NumElts; l += 8) {
13340 for (unsigned i = 0; i != 4; ++i)
13341 Indices[l + i] = l + i;
13342 for (unsigned i = 4; i != 8; ++i) {
13343 Indices[l + i] = l + 4 + (Imm & 3);
13344 Imm >>= 2;
13345 }
13346 }
13347
13348 return Builder.CreateShuffleVector(Ops[0], makeArrayRef(Indices, NumElts),
13349 "pshufhw");
13350 }
13351 case X86::BI__builtin_ia32_pshufd:
13352 case X86::BI__builtin_ia32_pshufd256:
13353 case X86::BI__builtin_ia32_pshufd512:
13354 case X86::BI__builtin_ia32_vpermilpd:
13355 case X86::BI__builtin_ia32_vpermilps:
13356 case X86::BI__builtin_ia32_vpermilpd256:
13357 case X86::BI__builtin_ia32_vpermilps256:
13358 case X86::BI__builtin_ia32_vpermilpd512:
13359 case X86::BI__builtin_ia32_vpermilps512: {
13360 uint32_t Imm = cast<llvm::ConstantInt>(Ops[1])->getZExtValue();
13361 auto *Ty = cast<llvm::FixedVectorType>(Ops[0]->getType());
13362 unsigned NumElts = Ty->getNumElements();
13363 unsigned NumLanes = Ty->getPrimitiveSizeInBits() / 128;
13364 unsigned NumLaneElts = NumElts / NumLanes;
13365
13366 // Splat the 8-bits of immediate 4 times to help the loop wrap around.
13367 Imm = (Imm & 0xff) * 0x01010101;
13368
13369 int Indices[16];
13370 for (unsigned l = 0; l != NumElts; l += NumLaneElts) {
13371 for (unsigned i = 0; i != NumLaneElts; ++i) {
13372 Indices[i + l] = (Imm % NumLaneElts) + l;
13373 Imm /= NumLaneElts;
13374 }
13375 }
13376
13377 return Builder.CreateShuffleVector(Ops[0], makeArrayRef(Indices, NumElts),
13378 "permil");
13379 }
13380 case X86::BI__builtin_ia32_shufpd:
13381 case X86::BI__builtin_ia32_shufpd256:
13382 case X86::BI__builtin_ia32_shufpd512:
13383 case X86::BI__builtin_ia32_shufps:
13384 case X86::BI__builtin_ia32_shufps256:
13385 case X86::BI__builtin_ia32_shufps512: {
13386 uint32_t Imm = cast<llvm::ConstantInt>(Ops[2])->getZExtValue();
13387 auto *Ty = cast<llvm::FixedVectorType>(Ops[0]->getType());
13388 unsigned NumElts = Ty->getNumElements();
13389 unsigned NumLanes = Ty->getPrimitiveSizeInBits() / 128;
13390 unsigned NumLaneElts = NumElts / NumLanes;
13391
13392 // Splat the 8-bits of immediate 4 times to help the loop wrap around.
13393 Imm = (Imm & 0xff) * 0x01010101;
13394
13395 int Indices[16];
13396 for (unsigned l = 0; l != NumElts; l += NumLaneElts) {
13397 for (unsigned i = 0; i != NumLaneElts; ++i) {
13398 unsigned Index = Imm % NumLaneElts;
13399 Imm /= NumLaneElts;
13400 if (i >= (NumLaneElts / 2))
13401 Index += NumElts;
13402 Indices[l + i] = l + Index;
13403 }
13404 }
13405
13406 return Builder.CreateShuffleVector(Ops[0], Ops[1],
13407 makeArrayRef(Indices, NumElts),
13408 "shufp");
13409 }
13410 case X86::BI__builtin_ia32_permdi256:
13411 case X86::BI__builtin_ia32_permdf256:
13412 case X86::BI__builtin_ia32_permdi512:
13413 case X86::BI__builtin_ia32_permdf512: {
13414 unsigned Imm = cast<llvm::ConstantInt>(Ops[1])->getZExtValue();
13415 auto *Ty = cast<llvm::FixedVectorType>(Ops[0]->getType());
13416 unsigned NumElts = Ty->getNumElements();
13417
13418 // These intrinsics operate on 256-bit lanes of four 64-bit elements.
13419 int Indices[8];
13420 for (unsigned l = 0; l != NumElts; l += 4)
13421 for (unsigned i = 0; i != 4; ++i)
13422 Indices[l + i] = l + ((Imm >> (2 * i)) & 0x3);
13423
13424 return Builder.CreateShuffleVector(Ops[0], makeArrayRef(Indices, NumElts),
13425 "perm");
13426 }
13427 case X86::BI__builtin_ia32_palignr128:
13428 case X86::BI__builtin_ia32_palignr256:
13429 case X86::BI__builtin_ia32_palignr512: {
13430 unsigned ShiftVal = cast<llvm::ConstantInt>(Ops[2])->getZExtValue() & 0xff;
13431
13432 unsigned NumElts =
13433 cast<llvm::FixedVectorType>(Ops[0]->getType())->getNumElements();
13434 assert(NumElts % 16 == 0)((void)0);
13435
13436 // If palignr is shifting the pair of vectors more than the size of two
13437 // lanes, emit zero.
13438 if (ShiftVal >= 32)
13439 return llvm::Constant::getNullValue(ConvertType(E->getType()));
13440
13441 // If palignr is shifting the pair of input vectors more than one lane,
13442 // but less than two lanes, convert to shifting in zeroes.
13443 if (ShiftVal > 16) {
13444 ShiftVal -= 16;
13445 Ops[1] = Ops[0];
13446 Ops[0] = llvm::Constant::getNullValue(Ops[0]->getType());
13447 }
13448
13449 int Indices[64];
13450 // 256-bit palignr operates on 128-bit lanes so we need to handle that
13451 for (unsigned l = 0; l != NumElts; l += 16) {
13452 for (unsigned i = 0; i != 16; ++i) {
13453 unsigned Idx = ShiftVal + i;
13454 if (Idx >= 16)
13455 Idx += NumElts - 16; // End of lane, switch operand.
13456 Indices[l + i] = Idx + l;
13457 }
13458 }
13459
13460 return Builder.CreateShuffleVector(Ops[1], Ops[0],
13461 makeArrayRef(Indices, NumElts),
13462 "palignr");
13463 }
13464 case X86::BI__builtin_ia32_alignd128:
13465 case X86::BI__builtin_ia32_alignd256:
13466 case X86::BI__builtin_ia32_alignd512:
13467 case X86::BI__builtin_ia32_alignq128:
13468 case X86::BI__builtin_ia32_alignq256:
13469 case X86::BI__builtin_ia32_alignq512: {
13470 unsigned NumElts =
13471 cast<llvm::FixedVectorType>(Ops[0]->getType())->getNumElements();
13472 unsigned ShiftVal = cast<llvm::ConstantInt>(Ops[2])->getZExtValue() & 0xff;
13473
13474 // Mask the shift amount to width of a vector.
13475 ShiftVal &= NumElts - 1;
13476
13477 int Indices[16];
13478 for (unsigned i = 0; i != NumElts; ++i)
13479 Indices[i] = i + ShiftVal;
13480
13481 return Builder.CreateShuffleVector(Ops[1], Ops[0],
13482 makeArrayRef(Indices, NumElts),
13483 "valign");
13484 }
13485 case X86::BI__builtin_ia32_shuf_f32x4_256:
13486 case X86::BI__builtin_ia32_shuf_f64x2_256:
13487 case X86::BI__builtin_ia32_shuf_i32x4_256:
13488 case X86::BI__builtin_ia32_shuf_i64x2_256:
13489 case X86::BI__builtin_ia32_shuf_f32x4:
13490 case X86::BI__builtin_ia32_shuf_f64x2:
13491 case X86::BI__builtin_ia32_shuf_i32x4:
13492 case X86::BI__builtin_ia32_shuf_i64x2: {
13493 unsigned Imm = cast<llvm::ConstantInt>(Ops[2])->getZExtValue();
13494 auto *Ty = cast<llvm::FixedVectorType>(Ops[0]->getType());
13495 unsigned NumElts = Ty->getNumElements();
13496 unsigned NumLanes = Ty->getPrimitiveSizeInBits() == 512 ? 4 : 2;
13497 unsigned NumLaneElts = NumElts / NumLanes;
13498
13499 int Indices[16];
13500 for (unsigned l = 0; l != NumElts; l += NumLaneElts) {
13501 unsigned Index = (Imm % NumLanes) * NumLaneElts;
13502 Imm /= NumLanes; // Discard the bits we just used.
13503 if (l >= (NumElts / 2))
13504 Index += NumElts; // Switch to other source.
13505 for (unsigned i = 0; i != NumLaneElts; ++i) {
13506 Indices[l + i] = Index + i;
13507 }
13508 }
13509
13510 return Builder.CreateShuffleVector(Ops[0], Ops[1],
13511 makeArrayRef(Indices, NumElts),
13512 "shuf");
13513 }
13514
13515 case X86::BI__builtin_ia32_vperm2f128_pd256:
13516 case X86::BI__builtin_ia32_vperm2f128_ps256:
13517 case X86::BI__builtin_ia32_vperm2f128_si256:
13518 case X86::BI__builtin_ia32_permti256: {
13519 unsigned Imm = cast<llvm::ConstantInt>(Ops[2])->getZExtValue();
13520 unsigned NumElts =
13521 cast<llvm::FixedVectorType>(Ops[0]->getType())->getNumElements();
13522
13523 // This takes a very simple approach since there are two lanes and a
13524 // shuffle can have 2 inputs. So we reserve the first input for the first
13525 // lane and the second input for the second lane. This may result in
13526 // duplicate sources, but this can be dealt with in the backend.
13527
13528 Value *OutOps[2];
13529 int Indices[8];
13530 for (unsigned l = 0; l != 2; ++l) {
13531 // Determine the source for this lane.
13532 if (Imm & (1 << ((l * 4) + 3)))
13533 OutOps[l] = llvm::ConstantAggregateZero::get(Ops[0]->getType());
13534 else if (Imm & (1 << ((l * 4) + 1)))
13535 OutOps[l] = Ops[1];
13536 else
13537 OutOps[l] = Ops[0];
13538
13539 for (unsigned i = 0; i != NumElts/2; ++i) {
13540 // Start with ith element of the source for this lane.
13541 unsigned Idx = (l * NumElts) + i;
13542 // If bit 0 of the immediate half is set, switch to the high half of
13543 // the source.
13544 if (Imm & (1 << (l * 4)))
13545 Idx += NumElts/2;
13546 Indices[(l * (NumElts/2)) + i] = Idx;
13547 }
13548 }
13549
13550 return Builder.CreateShuffleVector(OutOps[0], OutOps[1],
13551 makeArrayRef(Indices, NumElts),
13552 "vperm");
13553 }
13554
13555 case X86::BI__builtin_ia32_pslldqi128_byteshift:
13556 case X86::BI__builtin_ia32_pslldqi256_byteshift:
13557 case X86::BI__builtin_ia32_pslldqi512_byteshift: {
13558 unsigned ShiftVal = cast<llvm::ConstantInt>(Ops[1])->getZExtValue() & 0xff;
13559 auto *ResultType = cast<llvm::FixedVectorType>(Ops[0]->getType());
13560 // Builtin type is vXi64 so multiply by 8 to get bytes.
13561 unsigned NumElts = ResultType->getNumElements() * 8;
13562
13563 // If pslldq is shifting the vector more than 15 bytes, emit zero.
13564 if (ShiftVal >= 16)
13565 return llvm::Constant::getNullValue(ResultType);
13566
13567 int Indices[64];
13568 // 256/512-bit pslldq operates on 128-bit lanes so we need to handle that
13569 for (unsigned l = 0; l != NumElts; l += 16) {
13570 for (unsigned i = 0; i != 16; ++i) {
13571 unsigned Idx = NumElts + i - ShiftVal;
13572 if (Idx < NumElts) Idx -= NumElts - 16; // end of lane, switch operand.
13573 Indices[l + i] = Idx + l;
13574 }
13575 }
13576
13577 auto *VecTy = llvm::FixedVectorType::get(Int8Ty, NumElts);
13578 Value *Cast = Builder.CreateBitCast(Ops[0], VecTy, "cast");
13579 Value *Zero = llvm::Constant::getNullValue(VecTy);
13580 Value *SV = Builder.CreateShuffleVector(Zero, Cast,
13581 makeArrayRef(Indices, NumElts),
13582 "pslldq");
13583 return Builder.CreateBitCast(SV, Ops[0]->getType(), "cast");
13584 }
13585 case X86::BI__builtin_ia32_psrldqi128_byteshift:
13586 case X86::BI__builtin_ia32_psrldqi256_byteshift:
13587 case X86::BI__builtin_ia32_psrldqi512_byteshift: {
13588 unsigned ShiftVal = cast<llvm::ConstantInt>(Ops[1])->getZExtValue() & 0xff;
13589 auto *ResultType = cast<llvm::FixedVectorType>(Ops[0]->getType());
13590 // Builtin type is vXi64 so multiply by 8 to get bytes.
13591 unsigned NumElts = ResultType->getNumElements() * 8;
13592
13593 // If psrldq is shifting the vector more than 15 bytes, emit zero.
13594 if (ShiftVal >= 16)
13595 return llvm::Constant::getNullValue(ResultType);
13596
13597 int Indices[64];
13598 // 256/512-bit psrldq operates on 128-bit lanes so we need to handle that
13599 for (unsigned l = 0; l != NumElts; l += 16) {
13600 for (unsigned i = 0; i != 16; ++i) {
13601 unsigned Idx = i + ShiftVal;
13602 if (Idx >= 16) Idx += NumElts - 16; // end of lane, switch operand.
13603 Indices[l + i] = Idx + l;
13604 }
13605 }
13606
13607 auto *VecTy = llvm::FixedVectorType::get(Int8Ty, NumElts);
13608 Value *Cast = Builder.CreateBitCast(Ops[0], VecTy, "cast");
13609 Value *Zero = llvm::Constant::getNullValue(VecTy);
13610 Value *SV = Builder.CreateShuffleVector(Cast, Zero,
13611 makeArrayRef(Indices, NumElts),
13612 "psrldq");
13613 return Builder.CreateBitCast(SV, ResultType, "cast");
13614 }
13615 case X86::BI__builtin_ia32_kshiftliqi:
13616 case X86::BI__builtin_ia32_kshiftlihi:
13617 case X86::BI__builtin_ia32_kshiftlisi:
13618 case X86::BI__builtin_ia32_kshiftlidi: {
13619 unsigned ShiftVal = cast<llvm::ConstantInt>(Ops[1])->getZExtValue() & 0xff;
13620 unsigned NumElts = Ops[0]->getType()->getIntegerBitWidth();
13621
13622 if (ShiftVal >= NumElts)
13623 return llvm::Constant::getNullValue(Ops[0]->getType());
13624
13625 Value *In = getMaskVecValue(*this, Ops[0], NumElts);
13626
13627 int Indices[64];
13628 for (unsigned i = 0; i != NumElts; ++i)
13629 Indices[i] = NumElts + i - ShiftVal;
13630
13631 Value *Zero = llvm::Constant::getNullValue(In->getType());
13632 Value *SV = Builder.CreateShuffleVector(Zero, In,
13633 makeArrayRef(Indices, NumElts),
13634 "kshiftl");
13635 return Builder.CreateBitCast(SV, Ops[0]->getType());
13636 }
13637 case X86::BI__builtin_ia32_kshiftriqi:
13638 case X86::BI__builtin_ia32_kshiftrihi:
13639 case X86::BI__builtin_ia32_kshiftrisi:
13640 case X86::BI__builtin_ia32_kshiftridi: {
13641 unsigned ShiftVal = cast<llvm::ConstantInt>(Ops[1])->getZExtValue() & 0xff;
13642 unsigned NumElts = Ops[0]->getType()->getIntegerBitWidth();
13643
13644 if (ShiftVal >= NumElts)
13645 return llvm::Constant::getNullValue(Ops[0]->getType());
13646
13647 Value *In = getMaskVecValue(*this, Ops[0], NumElts);
13648
13649 int Indices[64];
13650 for (unsigned i = 0; i != NumElts; ++i)
13651 Indices[i] = i + ShiftVal;
13652
13653 Value *Zero = llvm::Constant::getNullValue(In->getType());
13654 Value *SV = Builder.CreateShuffleVector(In, Zero,
13655 makeArrayRef(Indices, NumElts),
13656 "kshiftr");
13657 return Builder.CreateBitCast(SV, Ops[0]->getType());
13658 }
13659 case X86::BI__builtin_ia32_movnti:
13660 case X86::BI__builtin_ia32_movnti64:
13661 case X86::BI__builtin_ia32_movntsd:
13662 case X86::BI__builtin_ia32_movntss: {
13663 llvm::MDNode *Node = llvm::MDNode::get(
13664 getLLVMContext(), llvm::ConstantAsMetadata::get(Builder.getInt32(1)));
13665
13666 Value *Ptr = Ops[0];
13667 Value *Src = Ops[1];
13668
13669 // Extract the 0'th element of the source vector.
13670 if (BuiltinID == X86::BI__builtin_ia32_movntsd ||
13671 BuiltinID == X86::BI__builtin_ia32_movntss)
13672 Src = Builder.CreateExtractElement(Src, (uint64_t)0, "extract");
13673
13674 // Convert the type of the pointer to a pointer to the stored type.
13675 Value *BC = Builder.CreateBitCast(
13676 Ptr, llvm::PointerType::getUnqual(Src->getType()), "cast");
13677
13678 // Unaligned nontemporal store of the scalar value.
13679 StoreInst *SI = Builder.CreateDefaultAlignedStore(Src, BC);
13680 SI->setMetadata(CGM.getModule().getMDKindID("nontemporal"), Node);
13681 SI->setAlignment(llvm::Align(1));
13682 return SI;
13683 }
13684 // Rotate is a special case of funnel shift - 1st 2 args are the same.
13685 case X86::BI__builtin_ia32_vprotb:
13686 case X86::BI__builtin_ia32_vprotw:
13687 case X86::BI__builtin_ia32_vprotd:
13688 case X86::BI__builtin_ia32_vprotq:
13689 case X86::BI__builtin_ia32_vprotbi:
13690 case X86::BI__builtin_ia32_vprotwi:
13691 case X86::BI__builtin_ia32_vprotdi:
13692 case X86::BI__builtin_ia32_vprotqi:
13693 case X86::BI__builtin_ia32_prold128:
13694 case X86::BI__builtin_ia32_prold256:
13695 case X86::BI__builtin_ia32_prold512:
13696 case X86::BI__builtin_ia32_prolq128:
13697 case X86::BI__builtin_ia32_prolq256:
13698 case X86::BI__builtin_ia32_prolq512:
13699 case X86::BI__builtin_ia32_prolvd128:
13700 case X86::BI__builtin_ia32_prolvd256:
13701 case X86::BI__builtin_ia32_prolvd512:
13702 case X86::BI__builtin_ia32_prolvq128:
13703 case X86::BI__builtin_ia32_prolvq256:
13704 case X86::BI__builtin_ia32_prolvq512:
13705 return EmitX86FunnelShift(*this, Ops[0], Ops[0], Ops[1], false);
13706 case X86::BI__builtin_ia32_prord128:
13707 case X86::BI__builtin_ia32_prord256:
13708 case X86::BI__builtin_ia32_prord512:
13709 case X86::BI__builtin_ia32_prorq128:
13710 case X86::BI__builtin_ia32_prorq256:
13711 case X86::BI__builtin_ia32_prorq512:
13712 case X86::BI__builtin_ia32_prorvd128:
13713 case X86::BI__builtin_ia32_prorvd256:
13714 case X86::BI__builtin_ia32_prorvd512:
13715 case X86::BI__builtin_ia32_prorvq128:
13716 case X86::BI__builtin_ia32_prorvq256:
13717 case X86::BI__builtin_ia32_prorvq512:
13718 return EmitX86FunnelShift(*this, Ops[0], Ops[0], Ops[1], true);
13719 case X86::BI__builtin_ia32_selectb_128:
13720 case X86::BI__builtin_ia32_selectb_256:
13721 case X86::BI__builtin_ia32_selectb_512:
13722 case X86::BI__builtin_ia32_selectw_128:
13723 case X86::BI__builtin_ia32_selectw_256:
13724 case X86::BI__builtin_ia32_selectw_512:
13725 case X86::BI__builtin_ia32_selectd_128:
13726 case X86::BI__builtin_ia32_selectd_256:
13727 case X86::BI__builtin_ia32_selectd_512:
13728 case X86::BI__builtin_ia32_selectq_128:
13729 case X86::BI__builtin_ia32_selectq_256:
13730 case X86::BI__builtin_ia32_selectq_512:
13731 case X86::BI__builtin_ia32_selectps_128:
13732 case X86::BI__builtin_ia32_selectps_256:
13733 case X86::BI__builtin_ia32_selectps_512:
13734 case X86::BI__builtin_ia32_selectpd_128:
13735 case X86::BI__builtin_ia32_selectpd_256:
13736 case X86::BI__builtin_ia32_selectpd_512:
13737 return EmitX86Select(*this, Ops[0], Ops[1], Ops[2]);
13738 case X86::BI__builtin_ia32_selectss_128:
13739 case X86::BI__builtin_ia32_selectsd_128: {
13740 Value *A = Builder.CreateExtractElement(Ops[1], (uint64_t)0);
13741 Value *B = Builder.CreateExtractElement(Ops[2], (uint64_t)0);
13742 A = EmitX86ScalarSelect(*this, Ops[0], A, B);
13743 return Builder.CreateInsertElement(Ops[1], A, (uint64_t)0);
13744 }
13745 case X86::BI__builtin_ia32_cmpb128_mask:
13746 case X86::BI__builtin_ia32_cmpb256_mask:
13747 case X86::BI__builtin_ia32_cmpb512_mask:
13748 case X86::BI__builtin_ia32_cmpw128_mask:
13749 case X86::BI__builtin_ia32_cmpw256_mask:
13750 case X86::BI__builtin_ia32_cmpw512_mask:
13751 case X86::BI__builtin_ia32_cmpd128_mask:
13752 case X86::BI__builtin_ia32_cmpd256_mask:
13753 case X86::BI__builtin_ia32_cmpd512_mask:
13754 case X86::BI__builtin_ia32_cmpq128_mask:
13755 case X86::BI__builtin_ia32_cmpq256_mask:
13756 case X86::BI__builtin_ia32_cmpq512_mask: {
13757 unsigned CC = cast<llvm::ConstantInt>(Ops[2])->getZExtValue() & 0x7;
13758 return EmitX86MaskedCompare(*this, CC, true, Ops);
13759 }
13760 case X86::BI__builtin_ia32_ucmpb128_mask:
13761 case X86::BI__builtin_ia32_ucmpb256_mask:
13762 case X86::BI__builtin_ia32_ucmpb512_mask:
13763 case X86::BI__builtin_ia32_ucmpw128_mask:
13764 case X86::BI__builtin_ia32_ucmpw256_mask:
13765 case X86::BI__builtin_ia32_ucmpw512_mask:
13766 case X86::BI__builtin_ia32_ucmpd128_mask:
13767 case X86::BI__builtin_ia32_ucmpd256_mask:
13768 case X86::BI__builtin_ia32_ucmpd512_mask:
13769 case X86::BI__builtin_ia32_ucmpq128_mask:
13770 case X86::BI__builtin_ia32_ucmpq256_mask:
13771 case X86::BI__builtin_ia32_ucmpq512_mask: {
13772 unsigned CC = cast<llvm::ConstantInt>(Ops[2])->getZExtValue() & 0x7;
13773 return EmitX86MaskedCompare(*this, CC, false, Ops);
13774 }
13775 case X86::BI__builtin_ia32_vpcomb:
13776 case X86::BI__builtin_ia32_vpcomw:
13777 case X86::BI__builtin_ia32_vpcomd:
13778 case X86::BI__builtin_ia32_vpcomq:
13779 return EmitX86vpcom(*this, Ops, true);
13780 case X86::BI__builtin_ia32_vpcomub:
13781 case X86::BI__builtin_ia32_vpcomuw:
13782 case X86::BI__builtin_ia32_vpcomud:
13783 case X86::BI__builtin_ia32_vpcomuq:
13784 return EmitX86vpcom(*this, Ops, false);
13785
13786 case X86::BI__builtin_ia32_kortestcqi:
13787 case X86::BI__builtin_ia32_kortestchi:
13788 case X86::BI__builtin_ia32_kortestcsi:
13789 case X86::BI__builtin_ia32_kortestcdi: {
13790 Value *Or = EmitX86MaskLogic(*this, Instruction::Or, Ops);
13791 Value *C = llvm::Constant::getAllOnesValue(Ops[0]->getType());
13792 Value *Cmp = Builder.CreateICmpEQ(Or, C);
13793 return Builder.CreateZExt(Cmp, ConvertType(E->getType()));
13794 }
13795 case X86::BI__builtin_ia32_kortestzqi:
13796 case X86::BI__builtin_ia32_kortestzhi:
13797 case X86::BI__builtin_ia32_kortestzsi:
13798 case X86::BI__builtin_ia32_kortestzdi: {
13799 Value *Or = EmitX86MaskLogic(*this, Instruction::Or, Ops);
13800 Value *C = llvm::Constant::getNullValue(Ops[0]->getType());
13801 Value *Cmp = Builder.CreateICmpEQ(Or, C);
13802 return Builder.CreateZExt(Cmp, ConvertType(E->getType()));
13803 }
13804
13805 case X86::BI__builtin_ia32_ktestcqi:
13806 case X86::BI__builtin_ia32_ktestzqi:
13807 case X86::BI__builtin_ia32_ktestchi:
13808 case X86::BI__builtin_ia32_ktestzhi:
13809 case X86::BI__builtin_ia32_ktestcsi:
13810 case X86::BI__builtin_ia32_ktestzsi:
13811 case X86::BI__builtin_ia32_ktestcdi:
13812 case X86::BI__builtin_ia32_ktestzdi: {
13813 Intrinsic::ID IID;
13814 switch (BuiltinID) {
13815 default: llvm_unreachable("Unsupported intrinsic!")__builtin_unreachable();
13816 case X86::BI__builtin_ia32_ktestcqi:
13817 IID = Intrinsic::x86_avx512_ktestc_b;
13818 break;
13819 case X86::BI__builtin_ia32_ktestzqi:
13820 IID = Intrinsic::x86_avx512_ktestz_b;
13821 break;
13822 case X86::BI__builtin_ia32_ktestchi:
13823 IID = Intrinsic::x86_avx512_ktestc_w;
13824 break;
13825 case X86::BI__builtin_ia32_ktestzhi:
13826 IID = Intrinsic::x86_avx512_ktestz_w;
13827 break;
13828 case X86::BI__builtin_ia32_ktestcsi:
13829 IID = Intrinsic::x86_avx512_ktestc_d;
13830 break;
13831 case X86::BI__builtin_ia32_ktestzsi:
13832 IID = Intrinsic::x86_avx512_ktestz_d;
13833 break;
13834 case X86::BI__builtin_ia32_ktestcdi:
13835 IID = Intrinsic::x86_avx512_ktestc_q;
13836 break;
13837 case X86::BI__builtin_ia32_ktestzdi:
13838 IID = Intrinsic::x86_avx512_ktestz_q;
13839 break;
13840 }
13841
13842 unsigned NumElts = Ops[0]->getType()->getIntegerBitWidth();
13843 Value *LHS = getMaskVecValue(*this, Ops[0], NumElts);
13844 Value *RHS = getMaskVecValue(*this, Ops[1], NumElts);
13845 Function *Intr = CGM.getIntrinsic(IID);
13846 return Builder.CreateCall(Intr, {LHS, RHS});
13847 }
13848
13849 case X86::BI__builtin_ia32_kaddqi:
13850 case X86::BI__builtin_ia32_kaddhi:
13851 case X86::BI__builtin_ia32_kaddsi:
13852 case X86::BI__builtin_ia32_kadddi: {
13853 Intrinsic::ID IID;
13854 switch (BuiltinID) {
13855 default: llvm_unreachable("Unsupported intrinsic!")__builtin_unreachable();
13856 case X86::BI__builtin_ia32_kaddqi:
13857 IID = Intrinsic::x86_avx512_kadd_b;
13858 break;
13859 case X86::BI__builtin_ia32_kaddhi:
13860 IID = Intrinsic::x86_avx512_kadd_w;
13861 break;
13862 case X86::BI__builtin_ia32_kaddsi:
13863 IID = Intrinsic::x86_avx512_kadd_d;
13864 break;
13865 case X86::BI__builtin_ia32_kadddi:
13866 IID = Intrinsic::x86_avx512_kadd_q;
13867 break;
13868 }
13869
13870 unsigned NumElts = Ops[0]->getType()->getIntegerBitWidth();
13871 Value *LHS = getMaskVecValue(*this, Ops[0], NumElts);
13872 Value *RHS = getMaskVecValue(*this, Ops[1], NumElts);
13873 Function *Intr = CGM.getIntrinsic(IID);
13874 Value *Res = Builder.CreateCall(Intr, {LHS, RHS});
13875 return Builder.CreateBitCast(Res, Ops[0]->getType());
13876 }
13877 case X86::BI__builtin_ia32_kandqi:
13878 case X86::BI__builtin_ia32_kandhi:
13879 case X86::BI__builtin_ia32_kandsi:
13880 case X86::BI__builtin_ia32_kanddi:
13881 return EmitX86MaskLogic(*this, Instruction::And, Ops);
13882 case X86::BI__builtin_ia32_kandnqi:
13883 case X86::BI__builtin_ia32_kandnhi:
13884 case X86::BI__builtin_ia32_kandnsi:
13885 case X86::BI__builtin_ia32_kandndi:
13886 return EmitX86MaskLogic(*this, Instruction::And, Ops, true);
13887 case X86::BI__builtin_ia32_korqi:
13888 case X86::BI__builtin_ia32_korhi:
13889 case X86::BI__builtin_ia32_korsi:
13890 case X86::BI__builtin_ia32_kordi:
13891 return EmitX86MaskLogic(*this, Instruction::Or, Ops);
13892 case X86::BI__builtin_ia32_kxnorqi:
13893 case X86::BI__builtin_ia32_kxnorhi:
13894 case X86::BI__builtin_ia32_kxnorsi:
13895 case X86::BI__builtin_ia32_kxnordi:
13896 return EmitX86MaskLogic(*this, Instruction::Xor, Ops, true);
13897 case X86::BI__builtin_ia32_kxorqi:
13898 case X86::BI__builtin_ia32_kxorhi:
13899 case X86::BI__builtin_ia32_kxorsi:
13900 case X86::BI__builtin_ia32_kxordi:
13901 return EmitX86MaskLogic(*this, Instruction::Xor, Ops);
13902 case X86::BI__builtin_ia32_knotqi:
13903 case X86::BI__builtin_ia32_knothi:
13904 case X86::BI__builtin_ia32_knotsi:
13905 case X86::BI__builtin_ia32_knotdi: {
13906 unsigned NumElts = Ops[0]->getType()->getIntegerBitWidth();
13907 Value *Res = getMaskVecValue(*this, Ops[0], NumElts);
13908 return Builder.CreateBitCast(Builder.CreateNot(Res),
13909 Ops[0]->getType());
13910 }
13911 case X86::BI__builtin_ia32_kmovb:
13912 case X86::BI__builtin_ia32_kmovw:
13913 case X86::BI__builtin_ia32_kmovd:
13914 case X86::BI__builtin_ia32_kmovq: {
13915 // Bitcast to vXi1 type and then back to integer. This gets the mask
13916 // register type into the IR, but might be optimized out depending on
13917 // what's around it.
13918 unsigned NumElts = Ops[0]->getType()->getIntegerBitWidth();
13919 Value *Res = getMaskVecValue(*this, Ops[0], NumElts);
13920 return Builder.CreateBitCast(Res, Ops[0]->getType());
13921 }
13922
13923 case X86::BI__builtin_ia32_kunpckdi:
13924 case X86::BI__builtin_ia32_kunpcksi:
13925 case X86::BI__builtin_ia32_kunpckhi: {
13926 unsigned NumElts = Ops[0]->getType()->getIntegerBitWidth();
13927 Value *LHS = getMaskVecValue(*this, Ops[0], NumElts);
13928 Value *RHS = getMaskVecValue(*this, Ops[1], NumElts);
13929 int Indices[64];
13930 for (unsigned i = 0; i != NumElts; ++i)
13931 Indices[i] = i;
13932
13933 // First extract half of each vector. This gives better codegen than
13934 // doing it in a single shuffle.
13935 LHS = Builder.CreateShuffleVector(LHS, LHS,
13936 makeArrayRef(Indices, NumElts / 2));
13937 RHS = Builder.CreateShuffleVector(RHS, RHS,
13938 makeArrayRef(Indices, NumElts / 2));
13939 // Concat the vectors.
13940 // NOTE: Operands are swapped to match the intrinsic definition.
13941 Value *Res = Builder.CreateShuffleVector(RHS, LHS,
13942 makeArrayRef(Indices, NumElts));
13943 return Builder.CreateBitCast(Res, Ops[0]->getType());
13944 }
13945
13946 case X86::BI__builtin_ia32_vplzcntd_128:
13947 case X86::BI__builtin_ia32_vplzcntd_256:
13948 case X86::BI__builtin_ia32_vplzcntd_512:
13949 case X86::BI__builtin_ia32_vplzcntq_128:
13950 case X86::BI__builtin_ia32_vplzcntq_256:
13951 case X86::BI__builtin_ia32_vplzcntq_512: {
13952 Function *F = CGM.getIntrinsic(Intrinsic::ctlz, Ops[0]->getType());
13953 return Builder.CreateCall(F, {Ops[0],Builder.getInt1(false)});
13954 }
13955 case X86::BI__builtin_ia32_sqrtss:
13956 case X86::BI__builtin_ia32_sqrtsd: {
13957 Value *A = Builder.CreateExtractElement(Ops[0], (uint64_t)0);
13958 Function *F;
13959 if (Builder.getIsFPConstrained()) {
13960 CodeGenFunction::CGFPOptionsRAII FPOptsRAII(*this, E);
13961 F = CGM.getIntrinsic(Intrinsic::experimental_constrained_sqrt,
13962 A->getType());
13963 A = Builder.CreateConstrainedFPCall(F, {A});
13964 } else {
13965 F = CGM.getIntrinsic(Intrinsic::sqrt, A->getType());
13966 A = Builder.CreateCall(F, {A});
13967 }
13968 return Builder.CreateInsertElement(Ops[0], A, (uint64_t)0);
13969 }
13970 case X86::BI__builtin_ia32_sqrtsd_round_mask:
13971 case X86::BI__builtin_ia32_sqrtss_round_mask: {
13972 unsigned CC = cast<llvm::ConstantInt>(Ops[4])->getZExtValue();
13973 // Support only if the rounding mode is 4 (AKA CUR_DIRECTION),
13974 // otherwise keep the intrinsic.
13975 if (CC != 4) {
13976 Intrinsic::ID IID = BuiltinID == X86::BI__builtin_ia32_sqrtsd_round_mask ?
13977 Intrinsic::x86_avx512_mask_sqrt_sd :
13978 Intrinsic::x86_avx512_mask_sqrt_ss;
13979 return Builder.CreateCall(CGM.getIntrinsic(IID), Ops);
13980 }
13981 Value *A = Builder.CreateExtractElement(Ops[1], (uint64_t)0);
13982 Function *F;
13983 if (Builder.getIsFPConstrained()) {
13984 CodeGenFunction::CGFPOptionsRAII FPOptsRAII(*this, E);
13985 F = CGM.getIntrinsic(Intrinsic::experimental_constrained_sqrt,
13986 A->getType());
13987 A = Builder.CreateConstrainedFPCall(F, A);
13988 } else {
13989 F = CGM.getIntrinsic(Intrinsic::sqrt, A->getType());
13990 A = Builder.CreateCall(F, A);
13991 }
13992 Value *Src = Builder.CreateExtractElement(Ops[2], (uint64_t)0);
13993 A = EmitX86ScalarSelect(*this, Ops[3], A, Src);
13994 return Builder.CreateInsertElement(Ops[0], A, (uint64_t)0);
13995 }
13996 case X86::BI__builtin_ia32_sqrtpd256:
13997 case X86::BI__builtin_ia32_sqrtpd:
13998 case X86::BI__builtin_ia32_sqrtps256:
13999 case X86::BI__builtin_ia32_sqrtps:
14000 case X86::BI__builtin_ia32_sqrtps512:
14001 case X86::BI__builtin_ia32_sqrtpd512: {
14002 if (Ops.size() == 2) {
14003 unsigned CC = cast<llvm::ConstantInt>(Ops[1])->getZExtValue();
14004 // Support only if the rounding mode is 4 (AKA CUR_DIRECTION),
14005 // otherwise keep the intrinsic.
14006 if (CC != 4) {
14007 Intrinsic::ID IID = BuiltinID == X86::BI__builtin_ia32_sqrtps512 ?
14008 Intrinsic::x86_avx512_sqrt_ps_512 :
14009 Intrinsic::x86_avx512_sqrt_pd_512;
14010 return Builder.CreateCall(CGM.getIntrinsic(IID), Ops);
14011 }
14012 }
14013 if (Builder.getIsFPConstrained()) {
14014 CodeGenFunction::CGFPOptionsRAII FPOptsRAII(*this, E);
14015 Function *F = CGM.getIntrinsic(Intrinsic::experimental_constrained_sqrt,
14016 Ops[0]->getType());
14017 return Builder.CreateConstrainedFPCall(F, Ops[0]);
14018 } else {
14019 Function *F = CGM.getIntrinsic(Intrinsic::sqrt, Ops[0]->getType());
14020 return Builder.CreateCall(F, Ops[0]);
14021 }
14022 }
14023 case X86::BI__builtin_ia32_pabsb128:
14024 case X86::BI__builtin_ia32_pabsw128:
14025 case X86::BI__builtin_ia32_pabsd128:
14026 case X86::BI__builtin_ia32_pabsb256:
14027 case X86::BI__builtin_ia32_pabsw256:
14028 case X86::BI__builtin_ia32_pabsd256:
14029 case X86::BI__builtin_ia32_pabsq128:
14030 case X86::BI__builtin_ia32_pabsq256:
14031 case X86::BI__builtin_ia32_pabsb512:
14032 case X86::BI__builtin_ia32_pabsw512:
14033 case X86::BI__builtin_ia32_pabsd512:
14034 case X86::BI__builtin_ia32_pabsq512: {
14035 Function *F = CGM.getIntrinsic(Intrinsic::abs, Ops[0]->getType());
14036 return Builder.CreateCall(F, {Ops[0], Builder.getInt1(false)});
14037 }
14038 case X86::BI__builtin_ia32_pmaxsb128:
14039 case X86::BI__builtin_ia32_pmaxsw128:
14040 case X86::BI__builtin_ia32_pmaxsd128:
14041 case X86::BI__builtin_ia32_pmaxsq128:
14042 case X86::BI__builtin_ia32_pmaxsb256:
14043 case X86::BI__builtin_ia32_pmaxsw256:
14044 case X86::BI__builtin_ia32_pmaxsd256:
14045 case X86::BI__builtin_ia32_pmaxsq256:
14046 case X86::BI__builtin_ia32_pmaxsb512:
14047 case X86::BI__builtin_ia32_pmaxsw512:
14048 case X86::BI__builtin_ia32_pmaxsd512:
14049 case X86::BI__builtin_ia32_pmaxsq512:
14050 return EmitX86BinaryIntrinsic(*this, Ops, Intrinsic::smax);
14051 case X86::BI__builtin_ia32_pmaxub128:
14052 case X86::BI__builtin_ia32_pmaxuw128:
14053 case X86::BI__builtin_ia32_pmaxud128:
14054 case X86::BI__builtin_ia32_pmaxuq128:
14055 case X86::BI__builtin_ia32_pmaxub256:
14056 case X86::BI__builtin_ia32_pmaxuw256:
14057 case X86::BI__builtin_ia32_pmaxud256:
14058 case X86::BI__builtin_ia32_pmaxuq256:
14059 case X86::BI__builtin_ia32_pmaxub512:
14060 case X86::BI__builtin_ia32_pmaxuw512:
14061 case X86::BI__builtin_ia32_pmaxud512:
14062 case X86::BI__builtin_ia32_pmaxuq512:
14063 return EmitX86BinaryIntrinsic(*this, Ops, Intrinsic::umax);
14064 case X86::BI__builtin_ia32_pminsb128:
14065 case X86::BI__builtin_ia32_pminsw128:
14066 case X86::BI__builtin_ia32_pminsd128:
14067 case X86::BI__builtin_ia32_pminsq128:
14068 case X86::BI__builtin_ia32_pminsb256:
14069 case X86::BI__builtin_ia32_pminsw256:
14070 case X86::BI__builtin_ia32_pminsd256:
14071 case X86::BI__builtin_ia32_pminsq256:
14072 case X86::BI__builtin_ia32_pminsb512:
14073 case X86::BI__builtin_ia32_pminsw512:
14074 case X86::BI__builtin_ia32_pminsd512:
14075 case X86::BI__builtin_ia32_pminsq512:
14076 return EmitX86BinaryIntrinsic(*this, Ops, Intrinsic::smin);
14077 case X86::BI__builtin_ia32_pminub128:
14078 case X86::BI__builtin_ia32_pminuw128:
14079 case X86::BI__builtin_ia32_pminud128:
14080 case X86::BI__builtin_ia32_pminuq128:
14081 case X86::BI__builtin_ia32_pminub256:
14082 case X86::BI__builtin_ia32_pminuw256:
14083 case X86::BI__builtin_ia32_pminud256:
14084 case X86::BI__builtin_ia32_pminuq256:
14085 case X86::BI__builtin_ia32_pminub512:
14086 case X86::BI__builtin_ia32_pminuw512:
14087 case X86::BI__builtin_ia32_pminud512:
14088 case X86::BI__builtin_ia32_pminuq512:
14089 return EmitX86BinaryIntrinsic(*this, Ops, Intrinsic::umin);
14090
14091 case X86::BI__builtin_ia32_pmuludq128:
14092 case X86::BI__builtin_ia32_pmuludq256:
14093 case X86::BI__builtin_ia32_pmuludq512:
14094 return EmitX86Muldq(*this, /*IsSigned*/false, Ops);
14095
14096 case X86::BI__builtin_ia32_pmuldq128:
14097 case X86::BI__builtin_ia32_pmuldq256:
14098 case X86::BI__builtin_ia32_pmuldq512:
14099 return EmitX86Muldq(*this, /*IsSigned*/true, Ops);
14100
14101 case X86::BI__builtin_ia32_pternlogd512_mask:
14102 case X86::BI__builtin_ia32_pternlogq512_mask:
14103 case X86::BI__builtin_ia32_pternlogd128_mask:
14104 case X86::BI__builtin_ia32_pternlogd256_mask:
14105 case X86::BI__builtin_ia32_pternlogq128_mask:
14106 case X86::BI__builtin_ia32_pternlogq256_mask:
14107 return EmitX86Ternlog(*this, /*ZeroMask*/false, Ops);
14108
14109 case X86::BI__builtin_ia32_pternlogd512_maskz:
14110 case X86::BI__builtin_ia32_pternlogq512_maskz:
14111 case X86::BI__builtin_ia32_pternlogd128_maskz:
14112 case X86::BI__builtin_ia32_pternlogd256_maskz:
14113 case X86::BI__builtin_ia32_pternlogq128_maskz:
14114 case X86::BI__builtin_ia32_pternlogq256_maskz:
14115 return EmitX86Ternlog(*this, /*ZeroMask*/true, Ops);
14116
14117 case X86::BI__builtin_ia32_vpshldd128:
14118 case X86::BI__builtin_ia32_vpshldd256:
14119 case X86::BI__builtin_ia32_vpshldd512:
14120 case X86::BI__builtin_ia32_vpshldq128:
14121 case X86::BI__builtin_ia32_vpshldq256:
14122 case X86::BI__builtin_ia32_vpshldq512:
14123 case X86::BI__builtin_ia32_vpshldw128:
14124 case X86::BI__builtin_ia32_vpshldw256:
14125 case X86::BI__builtin_ia32_vpshldw512:
14126 return EmitX86FunnelShift(*this, Ops[0], Ops[1], Ops[2], false);
14127
14128 case X86::BI__builtin_ia32_vpshrdd128:
14129 case X86::BI__builtin_ia32_vpshrdd256:
14130 case X86::BI__builtin_ia32_vpshrdd512:
14131 case X86::BI__builtin_ia32_vpshrdq128:
14132 case X86::BI__builtin_ia32_vpshrdq256:
14133 case X86::BI__builtin_ia32_vpshrdq512:
14134 case X86::BI__builtin_ia32_vpshrdw128:
14135 case X86::BI__builtin_ia32_vpshrdw256:
14136 case X86::BI__builtin_ia32_vpshrdw512:
14137 // Ops 0 and 1 are swapped.
14138 return EmitX86FunnelShift(*this, Ops[1], Ops[0], Ops[2], true);
14139
14140 case X86::BI__builtin_ia32_vpshldvd128:
14141 case X86::BI__builtin_ia32_vpshldvd256:
14142 case X86::BI__builtin_ia32_vpshldvd512:
14143 case X86::BI__builtin_ia32_vpshldvq128:
14144 case X86::BI__builtin_ia32_vpshldvq256:
14145 case X86::BI__builtin_ia32_vpshldvq512:
14146 case X86::BI__builtin_ia32_vpshldvw128:
14147 case X86::BI__builtin_ia32_vpshldvw256:
14148 case X86::BI__builtin_ia32_vpshldvw512:
14149 return EmitX86FunnelShift(*this, Ops[0], Ops[1], Ops[2], false);
14150
14151 case X86::BI__builtin_ia32_vpshrdvd128:
14152 case X86::BI__builtin_ia32_vpshrdvd256:
14153 case X86::BI__builtin_ia32_vpshrdvd512:
14154 case X86::BI__builtin_ia32_vpshrdvq128:
14155 case X86::BI__builtin_ia32_vpshrdvq256:
14156 case X86::BI__builtin_ia32_vpshrdvq512:
14157 case X86::BI__builtin_ia32_vpshrdvw128:
14158 case X86::BI__builtin_ia32_vpshrdvw256:
14159 case X86::BI__builtin_ia32_vpshrdvw512:
14160 // Ops 0 and 1 are swapped.
14161 return EmitX86FunnelShift(*this, Ops[1], Ops[0], Ops[2], true);
14162
14163 // Reductions
14164 case X86::BI__builtin_ia32_reduce_add_d512:
14165 case X86::BI__builtin_ia32_reduce_add_q512: {
14166 Function *F =
14167 CGM.getIntrinsic(Intrinsic::vector_reduce_add, Ops[0]->getType());
14168 return Builder.CreateCall(F, {Ops[0]});
14169 }
14170 case X86::BI__builtin_ia32_reduce_and_d512:
14171 case X86::BI__builtin_ia32_reduce_and_q512: {
14172 Function *F =
14173 CGM.getIntrinsic(Intrinsic::vector_reduce_and, Ops[0]->getType());
14174 return Builder.CreateCall(F, {Ops[0]});
14175 }
14176 case X86::BI__builtin_ia32_reduce_fadd_pd512:
14177 case X86::BI__builtin_ia32_reduce_fadd_ps512: {
14178 Function *F =
14179 CGM.getIntrinsic(Intrinsic::vector_reduce_fadd, Ops[1]->getType());
14180 Builder.getFastMathFlags().setAllowReassoc();
14181 return Builder.CreateCall(F, {Ops[0], Ops[1]});
14182 }
14183 case X86::BI__builtin_ia32_reduce_fmul_pd512:
14184 case X86::BI__builtin_ia32_reduce_fmul_ps512: {
14185 Function *F =
14186 CGM.getIntrinsic(Intrinsic::vector_reduce_fmul, Ops[1]->getType());
14187 Builder.getFastMathFlags().setAllowReassoc();
14188 return Builder.CreateCall(F, {Ops[0], Ops[1]});
14189 }
14190 case X86::BI__builtin_ia32_reduce_fmax_pd512:
14191 case X86::BI__builtin_ia32_reduce_fmax_ps512: {
14192 Function *F =
14193 CGM.getIntrinsic(Intrinsic::vector_reduce_fmax, Ops[0]->getType());
14194 Builder.getFastMathFlags().setNoNaNs();
14195 return Builder.CreateCall(F, {Ops[0]});
14196 }
14197 case X86::BI__builtin_ia32_reduce_fmin_pd512:
14198 case X86::BI__builtin_ia32_reduce_fmin_ps512: {
14199 Function *F =
14200 CGM.getIntrinsic(Intrinsic::vector_reduce_fmin, Ops[0]->getType());
14201 Builder.getFastMathFlags().setNoNaNs();
14202 return Builder.CreateCall(F, {Ops[0]});
14203 }
14204 case X86::BI__builtin_ia32_reduce_mul_d512:
14205 case X86::BI__builtin_ia32_reduce_mul_q512: {
14206 Function *F =
14207 CGM.getIntrinsic(Intrinsic::vector_reduce_mul, Ops[0]->getType());
14208 return Builder.CreateCall(F, {Ops[0]});
14209 }
14210 case X86::BI__builtin_ia32_reduce_or_d512:
14211 case X86::BI__builtin_ia32_reduce_or_q512: {
14212 Function *F =
14213 CGM.getIntrinsic(Intrinsic::vector_reduce_or, Ops[0]->getType());
14214 return Builder.CreateCall(F, {Ops[0]});
14215 }
14216 case X86::BI__builtin_ia32_reduce_smax_d512:
14217 case X86::BI__builtin_ia32_reduce_smax_q512: {
14218 Function *F =
14219 CGM.getIntrinsic(Intrinsic::vector_reduce_smax, Ops[0]->getType());
14220 return Builder.CreateCall(F, {Ops[0]});
14221 }
14222 case X86::BI__builtin_ia32_reduce_smin_d512:
14223 case X86::BI__builtin_ia32_reduce_smin_q512: {
14224 Function *F =
14225 CGM.getIntrinsic(Intrinsic::vector_reduce_smin, Ops[0]->getType());
14226 return Builder.CreateCall(F, {Ops[0]});
14227 }
14228 case X86::BI__builtin_ia32_reduce_umax_d512:
14229 case X86::BI__builtin_ia32_reduce_umax_q512: {
14230 Function *F =
14231 CGM.getIntrinsic(Intrinsic::vector_reduce_umax, Ops[0]->getType());
14232 return Builder.CreateCall(F, {Ops[0]});
14233 }
14234 case X86::BI__builtin_ia32_reduce_umin_d512:
14235 case X86::BI__builtin_ia32_reduce_umin_q512: {
14236 Function *F =
14237 CGM.getIntrinsic(Intrinsic::vector_reduce_umin, Ops[0]->getType());
14238 return Builder.CreateCall(F, {Ops[0]});
14239 }
14240
14241 // 3DNow!
14242 case X86::BI__builtin_ia32_pswapdsf:
14243 case X86::BI__builtin_ia32_pswapdsi: {
14244 llvm::Type *MMXTy = llvm::Type::getX86_MMXTy(getLLVMContext());
14245 Ops[0] = Builder.CreateBitCast(Ops[0], MMXTy, "cast");
14246 llvm::Function *F = CGM.getIntrinsic(Intrinsic::x86_3dnowa_pswapd);
14247 return Builder.CreateCall(F, Ops, "pswapd");
14248 }
14249 case X86::BI__builtin_ia32_rdrand16_step:
14250 case X86::BI__builtin_ia32_rdrand32_step:
14251 case X86::BI__builtin_ia32_rdrand64_step:
14252 case X86::BI__builtin_ia32_rdseed16_step:
14253 case X86::BI__builtin_ia32_rdseed32_step:
14254 case X86::BI__builtin_ia32_rdseed64_step: {
14255 Intrinsic::ID ID;
14256 switch (BuiltinID) {
14257 default: llvm_unreachable("Unsupported intrinsic!")__builtin_unreachable();
14258 case X86::BI__builtin_ia32_rdrand16_step:
14259 ID = Intrinsic::x86_rdrand_16;
14260 break;
14261 case X86::BI__builtin_ia32_rdrand32_step:
14262 ID = Intrinsic::x86_rdrand_32;
14263 break;
14264 case X86::BI__builtin_ia32_rdrand64_step:
14265 ID = Intrinsic::x86_rdrand_64;
14266 break;
14267 case X86::BI__builtin_ia32_rdseed16_step:
14268 ID = Intrinsic::x86_rdseed_16;
14269 break;
14270 case X86::BI__builtin_ia32_rdseed32_step:
14271 ID = Intrinsic::x86_rdseed_32;
14272 break;
14273 case X86::BI__builtin_ia32_rdseed64_step:
14274 ID = Intrinsic::x86_rdseed_64;
14275 break;
14276 }
14277
14278 Value *Call = Builder.CreateCall(CGM.getIntrinsic(ID));
14279 Builder.CreateDefaultAlignedStore(Builder.CreateExtractValue(Call, 0),
14280 Ops[0]);
14281 return Builder.CreateExtractValue(Call, 1);
14282 }
14283 case X86::BI__builtin_ia32_addcarryx_u32:
14284 case X86::BI__builtin_ia32_addcarryx_u64:
14285 case X86::BI__builtin_ia32_subborrow_u32:
14286 case X86::BI__builtin_ia32_subborrow_u64: {
14287 Intrinsic::ID IID;
14288 switch (BuiltinID) {
14289 default: llvm_unreachable("Unsupported intrinsic!")__builtin_unreachable();
14290 case X86::BI__builtin_ia32_addcarryx_u32:
14291 IID = Intrinsic::x86_addcarry_32;
14292 break;
14293 case X86::BI__builtin_ia32_addcarryx_u64:
14294 IID = Intrinsic::x86_addcarry_64;
14295 break;
14296 case X86::BI__builtin_ia32_subborrow_u32:
14297 IID = Intrinsic::x86_subborrow_32;
14298 break;
14299 case X86::BI__builtin_ia32_subborrow_u64:
14300 IID = Intrinsic::x86_subborrow_64;
14301 break;
14302 }
14303
14304 Value *Call = Builder.CreateCall(CGM.getIntrinsic(IID),
14305 { Ops[0], Ops[1], Ops[2] });
14306 Builder.CreateDefaultAlignedStore(Builder.CreateExtractValue(Call, 1),
14307 Ops[3]);
14308 return Builder.CreateExtractValue(Call, 0);
14309 }
14310
14311 case X86::BI__builtin_ia32_fpclassps128_mask:
14312 case X86::BI__builtin_ia32_fpclassps256_mask:
14313 case X86::BI__builtin_ia32_fpclassps512_mask:
14314 case X86::BI__builtin_ia32_fpclasspd128_mask:
14315 case X86::BI__builtin_ia32_fpclasspd256_mask:
14316 case X86::BI__builtin_ia32_fpclasspd512_mask: {
14317 unsigned NumElts =
14318 cast<llvm::FixedVectorType>(Ops[0]->getType())->getNumElements();
14319 Value *MaskIn = Ops[2];
14320 Ops.erase(&Ops[2]);
14321
14322 Intrinsic::ID ID;
14323 switch (BuiltinID) {
14324 default: llvm_unreachable("Unsupported intrinsic!")__builtin_unreachable();
14325 case X86::BI__builtin_ia32_fpclassps128_mask:
14326 ID = Intrinsic::x86_avx512_fpclass_ps_128;
14327 break;
14328 case X86::BI__builtin_ia32_fpclassps256_mask:
14329 ID = Intrinsic::x86_avx512_fpclass_ps_256;
14330 break;
14331 case X86::BI__builtin_ia32_fpclassps512_mask:
14332 ID = Intrinsic::x86_avx512_fpclass_ps_512;
14333 break;
14334 case X86::BI__builtin_ia32_fpclasspd128_mask:
14335 ID = Intrinsic::x86_avx512_fpclass_pd_128;
14336 break;
14337 case X86::BI__builtin_ia32_fpclasspd256_mask:
14338 ID = Intrinsic::x86_avx512_fpclass_pd_256;
14339 break;
14340 case X86::BI__builtin_ia32_fpclasspd512_mask:
14341 ID = Intrinsic::x86_avx512_fpclass_pd_512;
14342 break;
14343 }
14344
14345 Value *Fpclass = Builder.CreateCall(CGM.getIntrinsic(ID), Ops);
14346 return EmitX86MaskedCompareResult(*this, Fpclass, NumElts, MaskIn);
14347 }
14348
14349 case X86::BI__builtin_ia32_vp2intersect_q_512:
14350 case X86::BI__builtin_ia32_vp2intersect_q_256:
14351 case X86::BI__builtin_ia32_vp2intersect_q_128:
14352 case X86::BI__builtin_ia32_vp2intersect_d_512:
14353 case X86::BI__builtin_ia32_vp2intersect_d_256:
14354 case X86::BI__builtin_ia32_vp2intersect_d_128: {
14355 unsigned NumElts =
14356 cast<llvm::FixedVectorType>(Ops[0]->getType())->getNumElements();
14357 Intrinsic::ID ID;
14358
14359 switch (BuiltinID) {
14360 default: llvm_unreachable("Unsupported intrinsic!")__builtin_unreachable();
14361 case X86::BI__builtin_ia32_vp2intersect_q_512:
14362 ID = Intrinsic::x86_avx512_vp2intersect_q_512;
14363 break;
14364 case X86::BI__builtin_ia32_vp2intersect_q_256:
14365 ID = Intrinsic::x86_avx512_vp2intersect_q_256;
14366 break;
14367 case X86::BI__builtin_ia32_vp2intersect_q_128:
14368 ID = Intrinsic::x86_avx512_vp2intersect_q_128;
14369 break;
14370 case X86::BI__builtin_ia32_vp2intersect_d_512:
14371 ID = Intrinsic::x86_avx512_vp2intersect_d_512;
14372 break;
14373 case X86::BI__builtin_ia32_vp2intersect_d_256:
14374 ID = Intrinsic::x86_avx512_vp2intersect_d_256;
14375 break;
14376 case X86::BI__builtin_ia32_vp2intersect_d_128:
14377 ID = Intrinsic::x86_avx512_vp2intersect_d_128;
14378 break;
14379 }
14380
14381 Value *Call = Builder.CreateCall(CGM.getIntrinsic(ID), {Ops[0], Ops[1]});
14382 Value *Result = Builder.CreateExtractValue(Call, 0);
14383 Result = EmitX86MaskedCompareResult(*this, Result, NumElts, nullptr);
14384 Builder.CreateDefaultAlignedStore(Result, Ops[2]);
14385
14386 Result = Builder.CreateExtractValue(Call, 1);
14387 Result = EmitX86MaskedCompareResult(*this, Result, NumElts, nullptr);
14388 return Builder.CreateDefaultAlignedStore(Result, Ops[3]);
14389 }
14390
14391 case X86::BI__builtin_ia32_vpmultishiftqb128:
14392 case X86::BI__builtin_ia32_vpmultishiftqb256:
14393 case X86::BI__builtin_ia32_vpmultishiftqb512: {
14394 Intrinsic::ID ID;
14395 switch (BuiltinID) {
14396 default: llvm_unreachable("Unsupported intrinsic!")__builtin_unreachable();
14397 case X86::BI__builtin_ia32_vpmultishiftqb128:
14398 ID = Intrinsic::x86_avx512_pmultishift_qb_128;
14399 break;
14400 case X86::BI__builtin_ia32_vpmultishiftqb256:
14401 ID = Intrinsic::x86_avx512_pmultishift_qb_256;
14402 break;
14403 case X86::BI__builtin_ia32_vpmultishiftqb512:
14404 ID = Intrinsic::x86_avx512_pmultishift_qb_512;
14405 break;
14406 }
14407
14408 return Builder.CreateCall(CGM.getIntrinsic(ID), Ops);
14409 }
14410
14411 case X86::BI__builtin_ia32_vpshufbitqmb128_mask:
14412 case X86::BI__builtin_ia32_vpshufbitqmb256_mask:
14413 case X86::BI__builtin_ia32_vpshufbitqmb512_mask: {
14414 unsigned NumElts =
14415 cast<llvm::FixedVectorType>(Ops[0]->getType())->getNumElements();
14416 Value *MaskIn = Ops[2];
14417 Ops.erase(&Ops[2]);
14418
14419 Intrinsic::ID ID;
14420 switch (BuiltinID) {
14421 default: llvm_unreachable("Unsupported intrinsic!")__builtin_unreachable();
14422 case X86::BI__builtin_ia32_vpshufbitqmb128_mask:
14423 ID = Intrinsic::x86_avx512_vpshufbitqmb_128;
14424 break;
14425 case X86::BI__builtin_ia32_vpshufbitqmb256_mask:
14426 ID = Intrinsic::x86_avx512_vpshufbitqmb_256;
14427 break;
14428 case X86::BI__builtin_ia32_vpshufbitqmb512_mask:
14429 ID = Intrinsic::x86_avx512_vpshufbitqmb_512;
14430 break;
14431 }
14432
14433 Value *Shufbit = Builder.CreateCall(CGM.getIntrinsic(ID), Ops);
14434 return EmitX86MaskedCompareResult(*this, Shufbit, NumElts, MaskIn);
14435 }
14436
14437 // packed comparison intrinsics
14438 case X86::BI__builtin_ia32_cmpeqps:
14439 case X86::BI__builtin_ia32_cmpeqpd:
14440 return getVectorFCmpIR(CmpInst::FCMP_OEQ, /*IsSignaling*/false);
14441 case X86::BI__builtin_ia32_cmpltps:
14442 case X86::BI__builtin_ia32_cmpltpd:
14443 return getVectorFCmpIR(CmpInst::FCMP_OLT, /*IsSignaling*/true);
14444 case X86::BI__builtin_ia32_cmpleps:
14445 case X86::BI__builtin_ia32_cmplepd:
14446 return getVectorFCmpIR(CmpInst::FCMP_OLE, /*IsSignaling*/true);
14447 case X86::BI__builtin_ia32_cmpunordps:
14448 case X86::BI__builtin_ia32_cmpunordpd:
14449 return getVectorFCmpIR(CmpInst::FCMP_UNO, /*IsSignaling*/false);
14450 case X86::BI__builtin_ia32_cmpneqps:
14451 case X86::BI__builtin_ia32_cmpneqpd:
14452 return getVectorFCmpIR(CmpInst::FCMP_UNE, /*IsSignaling*/false);
14453 case X86::BI__builtin_ia32_cmpnltps:
14454 case X86::BI__builtin_ia32_cmpnltpd:
14455 return getVectorFCmpIR(CmpInst::FCMP_UGE, /*IsSignaling*/true);
14456 case X86::BI__builtin_ia32_cmpnleps:
14457 case X86::BI__builtin_ia32_cmpnlepd:
14458 return getVectorFCmpIR(CmpInst::FCMP_UGT, /*IsSignaling*/true);
14459 case X86::BI__builtin_ia32_cmpordps:
14460 case X86::BI__builtin_ia32_cmpordpd:
14461 return getVectorFCmpIR(CmpInst::FCMP_ORD, /*IsSignaling*/false);
14462 case X86::BI__builtin_ia32_cmpps128_mask:
14463 case X86::BI__builtin_ia32_cmpps256_mask:
14464 case X86::BI__builtin_ia32_cmpps512_mask:
14465 case X86::BI__builtin_ia32_cmppd128_mask:
14466 case X86::BI__builtin_ia32_cmppd256_mask:
14467 case X86::BI__builtin_ia32_cmppd512_mask:
14468 IsMaskFCmp = true;
14469 LLVM_FALLTHROUGH[[gnu::fallthrough]];
14470 case X86::BI__builtin_ia32_cmpps:
14471 case X86::BI__builtin_ia32_cmpps256:
14472 case X86::BI__builtin_ia32_cmppd:
14473 case X86::BI__builtin_ia32_cmppd256: {
14474 // Lowering vector comparisons to fcmp instructions, while
14475 // ignoring signalling behaviour requested
14476 // ignoring rounding mode requested
14477 // This is only possible if fp-model is not strict and FENV_ACCESS is off.
14478
14479 // The third argument is the comparison condition, and integer in the
14480 // range [0, 31]
14481 unsigned CC = cast<llvm::ConstantInt>(Ops[2])->getZExtValue() & 0x1f;
14482
14483 // Lowering to IR fcmp instruction.
14484 // Ignoring requested signaling behaviour,
14485 // e.g. both _CMP_GT_OS & _CMP_GT_OQ are translated to FCMP_OGT.
14486 FCmpInst::Predicate Pred;
14487 bool IsSignaling;
14488 // Predicates for 16-31 repeat the 0-15 predicates. Only the signalling
14489 // behavior is inverted. We'll handle that after the switch.
14490 switch (CC & 0xf) {
14491 case 0x00: Pred = FCmpInst::FCMP_OEQ; IsSignaling = false; break;
14492 case 0x01: Pred = FCmpInst::FCMP_OLT; IsSignaling = true; break;
14493 case 0x02: Pred = FCmpInst::FCMP_OLE; IsSignaling = true; break;
14494 case 0x03: Pred = FCmpInst::FCMP_UNO; IsSignaling = false; break;
14495 case 0x04: Pred = FCmpInst::FCMP_UNE; IsSignaling = false; break;
14496 case 0x05: Pred = FCmpInst::FCMP_UGE; IsSignaling = true; break;
14497 case 0x06: Pred = FCmpInst::FCMP_UGT; IsSignaling = true; break;
14498 case 0x07: Pred = FCmpInst::FCMP_ORD; IsSignaling = false; break;
14499 case 0x08: Pred = FCmpInst::FCMP_UEQ; IsSignaling = false; break;
14500 case 0x09: Pred = FCmpInst::FCMP_ULT; IsSignaling = true; break;
14501 case 0x0a: Pred = FCmpInst::FCMP_ULE; IsSignaling = true; break;
14502 case 0x0b: Pred = FCmpInst::FCMP_FALSE; IsSignaling = false; break;
14503 case 0x0c: Pred = FCmpInst::FCMP_ONE; IsSignaling = false; break;
14504 case 0x0d: Pred = FCmpInst::FCMP_OGE; IsSignaling = true; break;
14505 case 0x0e: Pred = FCmpInst::FCMP_OGT; IsSignaling = true; break;
14506 case 0x0f: Pred = FCmpInst::FCMP_TRUE; IsSignaling = false; break;
14507 default: llvm_unreachable("Unhandled CC")__builtin_unreachable();
14508 }
14509
14510 // Invert the signalling behavior for 16-31.
14511 if (CC & 0x10)
14512 IsSignaling = !IsSignaling;
14513
14514 // If the predicate is true or false and we're using constrained intrinsics,
14515 // we don't have a compare intrinsic we can use. Just use the legacy X86
14516 // specific intrinsic.
14517 // If the intrinsic is mask enabled and we're using constrained intrinsics,
14518 // use the legacy X86 specific intrinsic.
14519 if (Builder.getIsFPConstrained() &&
14520 (Pred == FCmpInst::FCMP_TRUE || Pred == FCmpInst::FCMP_FALSE ||
14521 IsMaskFCmp)) {
14522
14523 Intrinsic::ID IID;
14524 switch (BuiltinID) {
14525 default: llvm_unreachable("Unexpected builtin")__builtin_unreachable();
14526 case X86::BI__builtin_ia32_cmpps:
14527 IID = Intrinsic::x86_sse_cmp_ps;
14528 break;
14529 case X86::BI__builtin_ia32_cmpps256:
14530 IID = Intrinsic::x86_avx_cmp_ps_256;
14531 break;
14532 case X86::BI__builtin_ia32_cmppd:
14533 IID = Intrinsic::x86_sse2_cmp_pd;
14534 break;
14535 case X86::BI__builtin_ia32_cmppd256:
14536 IID = Intrinsic::x86_avx_cmp_pd_256;
14537 break;
14538 case X86::BI__builtin_ia32_cmpps512_mask:
14539 IID = Intrinsic::x86_avx512_mask_cmp_ps_512;
14540 break;
14541 case X86::BI__builtin_ia32_cmppd512_mask:
14542 IID = Intrinsic::x86_avx512_mask_cmp_pd_512;
14543 break;
14544 case X86::BI__builtin_ia32_cmpps128_mask:
14545 IID = Intrinsic::x86_avx512_mask_cmp_ps_128;
14546 break;
14547 case X86::BI__builtin_ia32_cmpps256_mask:
14548 IID = Intrinsic::x86_avx512_mask_cmp_ps_256;
14549 break;
14550 case X86::BI__builtin_ia32_cmppd128_mask:
14551 IID = Intrinsic::x86_avx512_mask_cmp_pd_128;
14552 break;
14553 case X86::BI__builtin_ia32_cmppd256_mask:
14554 IID = Intrinsic::x86_avx512_mask_cmp_pd_256;
14555 break;
14556 }
14557
14558 Function *Intr = CGM.getIntrinsic(IID);
14559 if (IsMaskFCmp) {
14560 unsigned NumElts =
14561 cast<llvm::FixedVectorType>(Ops[0]->getType())->getNumElements();
14562 Ops[3] = getMaskVecValue(*this, Ops[3], NumElts);
14563 Value *Cmp = Builder.CreateCall(Intr, Ops);
14564 return EmitX86MaskedCompareResult(*this, Cmp, NumElts, nullptr);
14565 }
14566
14567 return Builder.CreateCall(Intr, Ops);
14568 }
14569
14570 // Builtins without the _mask suffix return a vector of integers
14571 // of the same width as the input vectors
14572 if (IsMaskFCmp) {
14573 // We ignore SAE if strict FP is disabled. We only keep precise
14574 // exception behavior under strict FP.
14575 // NOTE: If strict FP does ever go through here a CGFPOptionsRAII
14576 // object will be required.
14577 unsigned NumElts =
14578 cast<llvm::FixedVectorType>(Ops[0]->getType())->getNumElements();
14579 Value *Cmp;
14580 if (IsSignaling)
14581 Cmp = Builder.CreateFCmpS(Pred, Ops[0], Ops[1]);
14582 else
14583 Cmp = Builder.CreateFCmp(Pred, Ops[0], Ops[1]);
14584 return EmitX86MaskedCompareResult(*this, Cmp, NumElts, Ops[3]);
14585 }
14586
14587 return getVectorFCmpIR(Pred, IsSignaling);
14588 }
14589
14590 // SSE scalar comparison intrinsics
14591 case X86::BI__builtin_ia32_cmpeqss:
14592 return getCmpIntrinsicCall(Intrinsic::x86_sse_cmp_ss, 0);
14593 case X86::BI__builtin_ia32_cmpltss:
14594 return getCmpIntrinsicCall(Intrinsic::x86_sse_cmp_ss, 1);
14595 case X86::BI__builtin_ia32_cmpless:
14596 return getCmpIntrinsicCall(Intrinsic::x86_sse_cmp_ss, 2);
14597 case X86::BI__builtin_ia32_cmpunordss:
14598 return getCmpIntrinsicCall(Intrinsic::x86_sse_cmp_ss, 3);
14599 case X86::BI__builtin_ia32_cmpneqss:
14600 return getCmpIntrinsicCall(Intrinsic::x86_sse_cmp_ss, 4);
14601 case X86::BI__builtin_ia32_cmpnltss:
14602 return getCmpIntrinsicCall(Intrinsic::x86_sse_cmp_ss, 5);
14603 case X86::BI__builtin_ia32_cmpnless:
14604 return getCmpIntrinsicCall(Intrinsic::x86_sse_cmp_ss, 6);
14605 case X86::BI__builtin_ia32_cmpordss:
14606 return getCmpIntrinsicCall(Intrinsic::x86_sse_cmp_ss, 7);
14607 case X86::BI__builtin_ia32_cmpeqsd:
14608 return getCmpIntrinsicCall(Intrinsic::x86_sse2_cmp_sd, 0);
14609 case X86::BI__builtin_ia32_cmpltsd:
14610 return getCmpIntrinsicCall(Intrinsic::x86_sse2_cmp_sd, 1);
14611 case X86::BI__builtin_ia32_cmplesd:
14612 return getCmpIntrinsicCall(Intrinsic::x86_sse2_cmp_sd, 2);
14613 case X86::BI__builtin_ia32_cmpunordsd:
14614 return getCmpIntrinsicCall(Intrinsic::x86_sse2_cmp_sd, 3);
14615 case X86::BI__builtin_ia32_cmpneqsd:
14616 return getCmpIntrinsicCall(Intrinsic::x86_sse2_cmp_sd, 4);
14617 case X86::BI__builtin_ia32_cmpnltsd:
14618 return getCmpIntrinsicCall(Intrinsic::x86_sse2_cmp_sd, 5);
14619 case X86::BI__builtin_ia32_cmpnlesd:
14620 return getCmpIntrinsicCall(Intrinsic::x86_sse2_cmp_sd, 6);
14621 case X86::BI__builtin_ia32_cmpordsd:
14622 return getCmpIntrinsicCall(Intrinsic::x86_sse2_cmp_sd, 7);
14623
14624 // f16c half2float intrinsics
14625 case X86::BI__builtin_ia32_vcvtph2ps:
14626 case X86::BI__builtin_ia32_vcvtph2ps256:
14627 case X86::BI__builtin_ia32_vcvtph2ps_mask:
14628 case X86::BI__builtin_ia32_vcvtph2ps256_mask:
14629 case X86::BI__builtin_ia32_vcvtph2ps512_mask: {
14630 CodeGenFunction::CGFPOptionsRAII FPOptsRAII(*this, E);
14631 return EmitX86CvtF16ToFloatExpr(*this, Ops, ConvertType(E->getType()));
14632 }
14633
14634// AVX512 bf16 intrinsics
14635 case X86::BI__builtin_ia32_cvtneps2bf16_128_mask: {
14636 Ops[2] = getMaskVecValue(
14637 *this, Ops[2],
14638 cast<llvm::FixedVectorType>(Ops[0]->getType())->getNumElements());
14639 Intrinsic::ID IID = Intrinsic::x86_avx512bf16_mask_cvtneps2bf16_128;
14640 return Builder.CreateCall(CGM.getIntrinsic(IID), Ops);
14641 }
14642 case X86::BI__builtin_ia32_cvtsbf162ss_32:
14643 return EmitX86CvtBF16ToFloatExpr(*this, E, Ops);
14644
14645 case X86::BI__builtin_ia32_cvtneps2bf16_256_mask:
14646 case X86::BI__builtin_ia32_cvtneps2bf16_512_mask: {
14647 Intrinsic::ID IID;
14648 switch (BuiltinID) {
14649 default: llvm_unreachable("Unsupported intrinsic!")__builtin_unreachable();
14650 case X86::BI__builtin_ia32_cvtneps2bf16_256_mask:
14651 IID = Intrinsic::x86_avx512bf16_cvtneps2bf16_256;
14652 break;
14653 case X86::BI__builtin_ia32_cvtneps2bf16_512_mask:
14654 IID = Intrinsic::x86_avx512bf16_cvtneps2bf16_512;
14655 break;
14656 }
14657 Value *Res = Builder.CreateCall(CGM.getIntrinsic(IID), Ops[0]);
14658 return EmitX86Select(*this, Ops[2], Res, Ops[1]);
14659 }
14660
14661 case X86::BI__emul:
14662 case X86::BI__emulu: {
14663 llvm::Type *Int64Ty = llvm::IntegerType::get(getLLVMContext(), 64);
14664 bool isSigned = (BuiltinID == X86::BI__emul);
14665 Value *LHS = Builder.CreateIntCast(Ops[0], Int64Ty, isSigned);
14666 Value *RHS = Builder.CreateIntCast(Ops[1], Int64Ty, isSigned);
14667 return Builder.CreateMul(LHS, RHS, "", !isSigned, isSigned);
14668 }
14669 case X86::BI__mulh:
14670 case X86::BI__umulh:
14671 case X86::BI_mul128:
14672 case X86::BI_umul128: {
14673 llvm::Type *ResType = ConvertType(E->getType());
14674 llvm::Type *Int128Ty = llvm::IntegerType::get(getLLVMContext(), 128);
14675
14676 bool IsSigned = (BuiltinID == X86::BI__mulh || BuiltinID == X86::BI_mul128);
14677 Value *LHS = Builder.CreateIntCast(Ops[0], Int128Ty, IsSigned);
14678 Value *RHS = Builder.CreateIntCast(Ops[1], Int128Ty, IsSigned);
14679
14680 Value *MulResult, *HigherBits;
14681 if (IsSigned) {
14682 MulResult = Builder.CreateNSWMul(LHS, RHS);
14683 HigherBits = Builder.CreateAShr(MulResult, 64);
14684 } else {
14685 MulResult = Builder.CreateNUWMul(LHS, RHS);
14686 HigherBits = Builder.CreateLShr(MulResult, 64);
14687 }
14688 HigherBits = Builder.CreateIntCast(HigherBits, ResType, IsSigned);
14689
14690 if (BuiltinID == X86::BI__mulh || BuiltinID == X86::BI__umulh)
14691 return HigherBits;
14692
14693 Address HighBitsAddress = EmitPointerWithAlignment(E->getArg(2));
14694 Builder.CreateStore(HigherBits, HighBitsAddress);
14695 return Builder.CreateIntCast(MulResult, ResType, IsSigned);
14696 }
14697
14698 case X86::BI__faststorefence: {
14699 return Builder.CreateFence(llvm::AtomicOrdering::SequentiallyConsistent,
14700 llvm::SyncScope::System);
14701 }
14702 case X86::BI__shiftleft128:
14703 case X86::BI__shiftright128: {
14704 llvm::Function *F = CGM.getIntrinsic(
14705 BuiltinID == X86::BI__shiftleft128 ? Intrinsic::fshl : Intrinsic::fshr,
14706 Int64Ty);
14707 // Flip low/high ops and zero-extend amount to matching type.
14708 // shiftleft128(Low, High, Amt) -> fshl(High, Low, Amt)
14709 // shiftright128(Low, High, Amt) -> fshr(High, Low, Amt)
14710 std::swap(Ops[0], Ops[1]);
14711 Ops[2] = Builder.CreateZExt(Ops[2], Int64Ty);
14712 return Builder.CreateCall(F, Ops);
14713 }
14714 case X86::BI_ReadWriteBarrier:
14715 case X86::BI_ReadBarrier:
14716 case X86::BI_WriteBarrier: {
14717 return Builder.CreateFence(llvm::AtomicOrdering::SequentiallyConsistent,
14718 llvm::SyncScope::SingleThread);
14719 }
14720
14721 case X86::BI_AddressOfReturnAddress: {
14722 Function *F =
14723 CGM.getIntrinsic(Intrinsic::addressofreturnaddress, AllocaInt8PtrTy);
14724 return Builder.CreateCall(F);
14725 }
14726 case X86::BI__stosb: {
14727 // We treat __stosb as a volatile memset - it may not generate "rep stosb"
14728 // instruction, but it will create a memset that won't be optimized away.
14729 return Builder.CreateMemSet(Ops[0], Ops[1], Ops[2], Align(1), true);
14730 }
14731 case X86::BI__ud2:
14732 // llvm.trap makes a ud2a instruction on x86.
14733 return EmitTrapCall(Intrinsic::trap);
14734 case X86::BI__int2c: {
14735 // This syscall signals a driver assertion failure in x86 NT kernels.
14736 llvm::FunctionType *FTy = llvm::FunctionType::get(VoidTy, false);
14737 llvm::InlineAsm *IA =
14738 llvm::InlineAsm::get(FTy, "int $$0x2c", "", /*hasSideEffects=*/true);
14739 llvm::AttributeList NoReturnAttr = llvm::AttributeList::get(
14740 getLLVMContext(), llvm::AttributeList::FunctionIndex,
14741 llvm::Attribute::NoReturn);
14742 llvm::CallInst *CI = Builder.CreateCall(IA);
14743 CI->setAttributes(NoReturnAttr);
14744 return CI;
14745 }
14746 case X86::BI__readfsbyte:
14747 case X86::BI__readfsword:
14748 case X86::BI__readfsdword:
14749 case X86::BI__readfsqword: {
14750 llvm::Type *IntTy = ConvertType(E->getType());
14751 Value *Ptr =
14752 Builder.CreateIntToPtr(Ops[0], llvm::PointerType::get(IntTy, 257));
14753 LoadInst *Load = Builder.CreateAlignedLoad(
14754 IntTy, Ptr, getContext().getTypeAlignInChars(E->getType()));
14755 Load->setVolatile(true);
14756 return Load;
14757 }
14758 case X86::BI__readgsbyte:
14759 case X86::BI__readgsword:
14760 case X86::BI__readgsdword:
14761 case X86::BI__readgsqword: {
14762 llvm::Type *IntTy = ConvertType(E->getType());
14763 Value *Ptr =
14764 Builder.CreateIntToPtr(Ops[0], llvm::PointerType::get(IntTy, 256));
14765 LoadInst *Load = Builder.CreateAlignedLoad(
14766 IntTy, Ptr, getContext().getTypeAlignInChars(E->getType()));
14767 Load->setVolatile(true);
14768 return Load;
14769 }
14770 case X86::BI__builtin_ia32_paddsb512:
14771 case X86::BI__builtin_ia32_paddsw512:
14772 case X86::BI__builtin_ia32_paddsb256:
14773 case X86::BI__builtin_ia32_paddsw256:
14774 case X86::BI__builtin_ia32_paddsb128:
14775 case X86::BI__builtin_ia32_paddsw128:
14776 return EmitX86BinaryIntrinsic(*this, Ops, Intrinsic::sadd_sat);
14777 case X86::BI__builtin_ia32_paddusb512:
14778 case X86::BI__builtin_ia32_paddusw512:
14779 case X86::BI__builtin_ia32_paddusb256:
14780 case X86::BI__builtin_ia32_paddusw256:
14781 case X86::BI__builtin_ia32_paddusb128:
14782 case X86::BI__builtin_ia32_paddusw128:
14783 return EmitX86BinaryIntrinsic(*this, Ops, Intrinsic::uadd_sat);
14784 case X86::BI__builtin_ia32_psubsb512:
14785 case X86::BI__builtin_ia32_psubsw512:
14786 case X86::BI__builtin_ia32_psubsb256:
14787 case X86::BI__builtin_ia32_psubsw256:
14788 case X86::BI__builtin_ia32_psubsb128:
14789 case X86::BI__builtin_ia32_psubsw128:
14790 return EmitX86BinaryIntrinsic(*this, Ops, Intrinsic::ssub_sat);
14791 case X86::BI__builtin_ia32_psubusb512:
14792 case X86::BI__builtin_ia32_psubusw512:
14793 case X86::BI__builtin_ia32_psubusb256:
14794 case X86::BI__builtin_ia32_psubusw256:
14795 case X86::BI__builtin_ia32_psubusb128:
14796 case X86::BI__builtin_ia32_psubusw128:
14797 return EmitX86BinaryIntrinsic(*this, Ops, Intrinsic::usub_sat);
14798 case X86::BI__builtin_ia32_encodekey128_u32: {
14799 Intrinsic::ID IID = Intrinsic::x86_encodekey128;
14800
14801 Value *Call = Builder.CreateCall(CGM.getIntrinsic(IID), {Ops[0], Ops[1]});
14802
14803 for (int i = 0; i < 6; ++i) {
14804 Value *Extract = Builder.CreateExtractValue(Call, i + 1);
14805 Value *Ptr = Builder.CreateConstGEP1_32(Int8Ty, Ops[2], i * 16);
14806 Ptr = Builder.CreateBitCast(
14807 Ptr, llvm::PointerType::getUnqual(Extract->getType()));
14808 Builder.CreateAlignedStore(Extract, Ptr, Align(1));
14809 }
14810
14811 return Builder.CreateExtractValue(Call, 0);
14812 }
14813 case X86::BI__builtin_ia32_encodekey256_u32: {
14814 Intrinsic::ID IID = Intrinsic::x86_encodekey256;
14815
14816 Value *Call =
14817 Builder.CreateCall(CGM.getIntrinsic(IID), {Ops[0], Ops[1], Ops[2]});
14818
14819 for (int i = 0; i < 7; ++i) {
14820 Value *Extract = Builder.CreateExtractValue(Call, i + 1);
14821 Value *Ptr = Builder.CreateConstGEP1_32(Int8Ty, Ops[3], i * 16);
14822 Ptr = Builder.CreateBitCast(
14823 Ptr, llvm::PointerType::getUnqual(Extract->getType()));
14824 Builder.CreateAlignedStore(Extract, Ptr, Align(1));
14825 }
14826
14827 return Builder.CreateExtractValue(Call, 0);
14828 }
14829 case X86::BI__builtin_ia32_aesenc128kl_u8:
14830 case X86::BI__builtin_ia32_aesdec128kl_u8:
14831 case X86::BI__builtin_ia32_aesenc256kl_u8:
14832 case X86::BI__builtin_ia32_aesdec256kl_u8: {
14833 Intrinsic::ID IID;
14834 StringRef BlockName;
14835 switch (BuiltinID) {
14836 default:
14837 llvm_unreachable("Unexpected builtin")__builtin_unreachable();
14838 case X86::BI__builtin_ia32_aesenc128kl_u8:
14839 IID = Intrinsic::x86_aesenc128kl;
14840 BlockName = "aesenc128kl";
14841 break;
14842 case X86::BI__builtin_ia32_aesdec128kl_u8:
14843 IID = Intrinsic::x86_aesdec128kl;
14844 BlockName = "aesdec128kl";
14845 break;
14846 case X86::BI__builtin_ia32_aesenc256kl_u8:
14847 IID = Intrinsic::x86_aesenc256kl;
14848 BlockName = "aesenc256kl";
14849 break;
14850 case X86::BI__builtin_ia32_aesdec256kl_u8:
14851 IID = Intrinsic::x86_aesdec256kl;
14852 BlockName = "aesdec256kl";
14853 break;
14854 }
14855
14856 Value *Call = Builder.CreateCall(CGM.getIntrinsic(IID), {Ops[1], Ops[2]});
14857
14858 BasicBlock *NoError =
14859 createBasicBlock(BlockName + "_no_error", this->CurFn);
14860 BasicBlock *Error = createBasicBlock(BlockName + "_error", this->CurFn);
14861 BasicBlock *End = createBasicBlock(BlockName + "_end", this->CurFn);
14862
14863 Value *Ret = Builder.CreateExtractValue(Call, 0);
14864 Value *Succ = Builder.CreateTrunc(Ret, Builder.getInt1Ty());
14865 Value *Out = Builder.CreateExtractValue(Call, 1);
14866 Builder.CreateCondBr(Succ, NoError, Error);
14867
14868 Builder.SetInsertPoint(NoError);
14869 Builder.CreateDefaultAlignedStore(Out, Ops[0]);
14870 Builder.CreateBr(End);
14871
14872 Builder.SetInsertPoint(Error);
14873 Constant *Zero = llvm::Constant::getNullValue(Out->getType());
14874 Builder.CreateDefaultAlignedStore(Zero, Ops[0]);
14875 Builder.CreateBr(End);
14876
14877 Builder.SetInsertPoint(End);
14878 return Builder.CreateExtractValue(Call, 0);
14879 }
14880 case X86::BI__builtin_ia32_aesencwide128kl_u8:
14881 case X86::BI__builtin_ia32_aesdecwide128kl_u8:
14882 case X86::BI__builtin_ia32_aesencwide256kl_u8:
14883 case X86::BI__builtin_ia32_aesdecwide256kl_u8: {
14884 Intrinsic::ID IID;
14885 StringRef BlockName;
14886 switch (BuiltinID) {
14887 case X86::BI__builtin_ia32_aesencwide128kl_u8:
14888 IID = Intrinsic::x86_aesencwide128kl;
14889 BlockName = "aesencwide128kl";
14890 break;
14891 case X86::BI__builtin_ia32_aesdecwide128kl_u8:
14892 IID = Intrinsic::x86_aesdecwide128kl;
14893 BlockName = "aesdecwide128kl";
14894 break;
14895 case X86::BI__builtin_ia32_aesencwide256kl_u8:
14896 IID = Intrinsic::x86_aesencwide256kl;
14897 BlockName = "aesencwide256kl";
14898 break;
14899 case X86::BI__builtin_ia32_aesdecwide256kl_u8:
14900 IID = Intrinsic::x86_aesdecwide256kl;
14901 BlockName = "aesdecwide256kl";
14902 break;
14903 }
14904
14905 llvm::Type *Ty = FixedVectorType::get(Builder.getInt64Ty(), 2);
14906 Value *InOps[9];
14907 InOps[0] = Ops[2];
14908 for (int i = 0; i != 8; ++i) {
14909 Value *Ptr = Builder.CreateConstGEP1_32(Ty, Ops[1], i);
14910 InOps[i + 1] = Builder.CreateAlignedLoad(Ty, Ptr, Align(16));
14911 }
14912
14913 Value *Call = Builder.CreateCall(CGM.getIntrinsic(IID), InOps);
14914
14915 BasicBlock *NoError =
14916 createBasicBlock(BlockName + "_no_error", this->CurFn);
14917 BasicBlock *Error = createBasicBlock(BlockName + "_error", this->CurFn);
14918 BasicBlock *End = createBasicBlock(BlockName + "_end", this->CurFn);
14919
14920 Value *Ret = Builder.CreateExtractValue(Call, 0);
14921 Value *Succ = Builder.CreateTrunc(Ret, Builder.getInt1Ty());
14922 Builder.CreateCondBr(Succ, NoError, Error);
14923
14924 Builder.SetInsertPoint(NoError);
14925 for (int i = 0; i != 8; ++i) {
14926 Value *Extract = Builder.CreateExtractValue(Call, i + 1);
14927 Value *Ptr = Builder.CreateConstGEP1_32(Extract->getType(), Ops[0], i);
14928 Builder.CreateAlignedStore(Extract, Ptr, Align(16));
14929 }
14930 Builder.CreateBr(End);
14931
14932 Builder.SetInsertPoint(Error);
14933 for (int i = 0; i != 8; ++i) {
14934 Value *Out = Builder.CreateExtractValue(Call, i + 1);
14935 Constant *Zero = llvm::Constant::getNullValue(Out->getType());
14936 Value *Ptr = Builder.CreateConstGEP1_32(Out->getType(), Ops[0], i);
14937 Builder.CreateAlignedStore(Zero, Ptr, Align(16));
14938 }
14939 Builder.CreateBr(End);
14940
14941 Builder.SetInsertPoint(End);
14942 return Builder.CreateExtractValue(Call, 0);
14943 }
14944 }
14945}
14946
14947Value *CodeGenFunction::EmitPPCBuiltinExpr(unsigned BuiltinID,
14948 const CallExpr *E) {
14949 SmallVector<Value*, 4> Ops;
14950
14951 for (unsigned i = 0, e = E->getNumArgs(); i != e; i++)
1
Assuming 'i' is equal to 'e'
2
Loop condition is false. Execution continues on line 14954
14952 Ops.push_back(EmitScalarExpr(E->getArg(i)));
14953
14954 Intrinsic::ID ID = Intrinsic::not_intrinsic;
14955
14956 switch (BuiltinID) {
3
Control jumps to 'case BI__builtin_vsx_xxsldwi:' at line 15673
14957 default: return nullptr;
14958
14959 // __builtin_ppc_get_timebase is GCC 4.8+'s PowerPC-specific name for what we
14960 // call __builtin_readcyclecounter.
14961 case PPC::BI__builtin_ppc_get_timebase:
14962 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::readcyclecounter));
14963
14964 // vec_ld, vec_xl_be, vec_lvsl, vec_lvsr
14965 case PPC::BI__builtin_altivec_lvx:
14966 case PPC::BI__builtin_altivec_lvxl:
14967 case PPC::BI__builtin_altivec_lvebx:
14968 case PPC::BI__builtin_altivec_lvehx:
14969 case PPC::BI__builtin_altivec_lvewx:
14970 case PPC::BI__builtin_altivec_lvsl:
14971 case PPC::BI__builtin_altivec_lvsr:
14972 case PPC::BI__builtin_vsx_lxvd2x:
14973 case PPC::BI__builtin_vsx_lxvw4x:
14974 case PPC::BI__builtin_vsx_lxvd2x_be:
14975 case PPC::BI__builtin_vsx_lxvw4x_be:
14976 case PPC::BI__builtin_vsx_lxvl:
14977 case PPC::BI__builtin_vsx_lxvll:
14978 {
14979 if(BuiltinID == PPC::BI__builtin_vsx_lxvl ||
14980 BuiltinID == PPC::BI__builtin_vsx_lxvll){
14981 Ops[0] = Builder.CreateBitCast(Ops[0], Int8PtrTy);
14982 }else {
14983 Ops[1] = Builder.CreateBitCast(Ops[1], Int8PtrTy);
14984 Ops[0] = Builder.CreateGEP(Int8Ty, Ops[1], Ops[0]);
14985 Ops.pop_back();
14986 }
14987
14988 switch (BuiltinID) {
14989 default: llvm_unreachable("Unsupported ld/lvsl/lvsr intrinsic!")__builtin_unreachable();
14990 case PPC::BI__builtin_altivec_lvx:
14991 ID = Intrinsic::ppc_altivec_lvx;
14992 break;
14993 case PPC::BI__builtin_altivec_lvxl:
14994 ID = Intrinsic::ppc_altivec_lvxl;
14995 break;
14996 case PPC::BI__builtin_altivec_lvebx:
14997 ID = Intrinsic::ppc_altivec_lvebx;
14998 break;
14999 case PPC::BI__builtin_altivec_lvehx:
15000 ID = Intrinsic::ppc_altivec_lvehx;
15001 break;
15002 case PPC::BI__builtin_altivec_lvewx:
15003 ID = Intrinsic::ppc_altivec_lvewx;
15004 break;
15005 case PPC::BI__builtin_altivec_lvsl:
15006 ID = Intrinsic::ppc_altivec_lvsl;
15007 break;
15008 case PPC::BI__builtin_altivec_lvsr:
15009 ID = Intrinsic::ppc_altivec_lvsr;
15010 break;
15011 case PPC::BI__builtin_vsx_lxvd2x:
15012 ID = Intrinsic::ppc_vsx_lxvd2x;
15013 break;
15014 case PPC::BI__builtin_vsx_lxvw4x:
15015 ID = Intrinsic::ppc_vsx_lxvw4x;
15016 break;
15017 case PPC::BI__builtin_vsx_lxvd2x_be:
15018 ID = Intrinsic::ppc_vsx_lxvd2x_be;
15019 break;
15020 case PPC::BI__builtin_vsx_lxvw4x_be:
15021 ID = Intrinsic::ppc_vsx_lxvw4x_be;
15022 break;
15023 case PPC::BI__builtin_vsx_lxvl:
15024 ID = Intrinsic::ppc_vsx_lxvl;
15025 break;
15026 case PPC::BI__builtin_vsx_lxvll:
15027 ID = Intrinsic::ppc_vsx_lxvll;
15028 break;
15029 }
15030 llvm::Function *F = CGM.getIntrinsic(ID);
15031 return Builder.CreateCall(F, Ops, "");
15032 }
15033
15034 // vec_st, vec_xst_be
15035 case PPC::BI__builtin_altivec_stvx:
15036 case PPC::BI__builtin_altivec_stvxl:
15037 case PPC::BI__builtin_altivec_stvebx:
15038 case PPC::BI__builtin_altivec_stvehx:
15039 case PPC::BI__builtin_altivec_stvewx:
15040 case PPC::BI__builtin_vsx_stxvd2x:
15041 case PPC::BI__builtin_vsx_stxvw4x:
15042 case PPC::BI__builtin_vsx_stxvd2x_be:
15043 case PPC::BI__builtin_vsx_stxvw4x_be:
15044 case PPC::BI__builtin_vsx_stxvl:
15045 case PPC::BI__builtin_vsx_stxvll:
15046 {
15047 if(BuiltinID == PPC::BI__builtin_vsx_stxvl ||
15048 BuiltinID == PPC::BI__builtin_vsx_stxvll ){
15049 Ops[1] = Builder.CreateBitCast(Ops[1], Int8PtrTy);
15050 }else {
15051 Ops[2] = Builder.CreateBitCast(Ops[2], Int8PtrTy);
15052 Ops[1] = Builder.CreateGEP(Int8Ty, Ops[2], Ops[1]);
15053 Ops.pop_back();
15054 }
15055
15056 switch (BuiltinID) {
15057 default: llvm_unreachable("Unsupported st intrinsic!")__builtin_unreachable();
15058 case PPC::BI__builtin_altivec_stvx:
15059 ID = Intrinsic::ppc_altivec_stvx;
15060 break;
15061 case PPC::BI__builtin_altivec_stvxl:
15062 ID = Intrinsic::ppc_altivec_stvxl;
15063 break;
15064 case PPC::BI__builtin_altivec_stvebx:
15065 ID = Intrinsic::ppc_altivec_stvebx;
15066 break;
15067 case PPC::BI__builtin_altivec_stvehx:
15068 ID = Intrinsic::ppc_altivec_stvehx;
15069 break;
15070 case PPC::BI__builtin_altivec_stvewx:
15071 ID = Intrinsic::ppc_altivec_stvewx;
15072 break;
15073 case PPC::BI__builtin_vsx_stxvd2x:
15074 ID = Intrinsic::ppc_vsx_stxvd2x;
15075 break;
15076 case PPC::BI__builtin_vsx_stxvw4x:
15077 ID = Intrinsic::ppc_vsx_stxvw4x;
15078 break;
15079 case PPC::BI__builtin_vsx_stxvd2x_be:
15080 ID = Intrinsic::ppc_vsx_stxvd2x_be;
15081 break;
15082 case PPC::BI__builtin_vsx_stxvw4x_be:
15083 ID = Intrinsic::ppc_vsx_stxvw4x_be;
15084 break;
15085 case PPC::BI__builtin_vsx_stxvl:
15086 ID = Intrinsic::ppc_vsx_stxvl;
15087 break;
15088 case PPC::BI__builtin_vsx_stxvll:
15089 ID = Intrinsic::ppc_vsx_stxvll;
15090 break;
15091 }
15092 llvm::Function *F = CGM.getIntrinsic(ID);
15093 return Builder.CreateCall(F, Ops, "");
15094 }
15095 case PPC::BI__builtin_vsx_ldrmb: {
15096 // Essentially boils down to performing an unaligned VMX load sequence so
15097 // as to avoid crossing a page boundary and then shuffling the elements
15098 // into the right side of the vector register.
15099 int64_t NumBytes = cast<ConstantInt>(Ops[1])->getZExtValue();
15100 llvm::Type *ResTy = ConvertType(E->getType());
15101 bool IsLE = getTarget().isLittleEndian();
15102
15103 // If the user wants the entire vector, just load the entire vector.
15104 if (NumBytes == 16) {
15105 Value *BC = Builder.CreateBitCast(Ops[0], ResTy->getPointerTo());
15106 Value *LD = Builder.CreateLoad(Address(BC, CharUnits::fromQuantity(1)));
15107 if (!IsLE)
15108 return LD;
15109
15110 // Reverse the bytes on LE.
15111 SmallVector<int, 16> RevMask;
15112 for (int Idx = 0; Idx < 16; Idx++)
15113 RevMask.push_back(15 - Idx);
15114 return Builder.CreateShuffleVector(LD, LD, RevMask);
15115 }
15116
15117 llvm::Function *Lvx = CGM.getIntrinsic(Intrinsic::ppc_altivec_lvx);
15118 llvm::Function *Lvs = CGM.getIntrinsic(IsLE ? Intrinsic::ppc_altivec_lvsr
15119 : Intrinsic::ppc_altivec_lvsl);
15120 llvm::Function *Vperm = CGM.getIntrinsic(Intrinsic::ppc_altivec_vperm);
15121 Value *HiMem = Builder.CreateGEP(
15122 Int8Ty, Ops[0], ConstantInt::get(Ops[1]->getType(), NumBytes - 1));
15123 Value *LoLd = Builder.CreateCall(Lvx, Ops[0], "ld.lo");
15124 Value *HiLd = Builder.CreateCall(Lvx, HiMem, "ld.hi");
15125 Value *Mask1 = Builder.CreateCall(Lvs, Ops[0], "mask1");
15126
15127 Ops.clear();
15128 Ops.push_back(IsLE ? HiLd : LoLd);
15129 Ops.push_back(IsLE ? LoLd : HiLd);
15130 Ops.push_back(Mask1);
15131 Value *AllElts = Builder.CreateCall(Vperm, Ops, "shuffle1");
15132 Constant *Zero = llvm::Constant::getNullValue(IsLE ? ResTy : AllElts->getType());
15133
15134 if (IsLE) {
15135 SmallVector<int, 16> Consts;
15136 for (int Idx = 0; Idx < 16; Idx++) {
15137 int Val = (NumBytes - Idx - 1 >= 0) ? (NumBytes - Idx - 1)
15138 : 16 - (NumBytes - Idx);
15139 Consts.push_back(Val);
15140 }
15141 return Builder.CreateShuffleVector(Builder.CreateBitCast(AllElts, ResTy),
15142 Zero, Consts);
15143 }
15144 SmallVector<Constant *, 16> Consts;
15145 for (int Idx = 0; Idx < 16; Idx++)
15146 Consts.push_back(Builder.getInt8(NumBytes + Idx));
15147 Value *Mask2 = ConstantVector::get(Consts);
15148 return Builder.CreateBitCast(
15149 Builder.CreateCall(Vperm, {Zero, AllElts, Mask2}, "shuffle2"), ResTy);
15150 }
15151 case PPC::BI__builtin_vsx_strmb: {
15152 int64_t NumBytes = cast<ConstantInt>(Ops[1])->getZExtValue();
15153 bool IsLE = getTarget().isLittleEndian();
15154 auto StoreSubVec = [&](unsigned Width, unsigned Offset, unsigned EltNo) {
15155 // Storing the whole vector, simply store it on BE and reverse bytes and
15156 // store on LE.
15157 if (Width == 16) {
15158 Value *BC =
15159 Builder.CreateBitCast(Ops[0], Ops[2]->getType()->getPointerTo());
15160 Value *StVec = Ops[2];
15161 if (IsLE) {
15162 SmallVector<int, 16> RevMask;
15163 for (int Idx = 0; Idx < 16; Idx++)
15164 RevMask.push_back(15 - Idx);
15165 StVec = Builder.CreateShuffleVector(Ops[2], Ops[2], RevMask);
15166 }
15167 return Builder.CreateStore(StVec,
15168 Address(BC, CharUnits::fromQuantity(1)));
15169 }
15170 auto *ConvTy = Int64Ty;
15171 unsigned NumElts = 0;
15172 switch (Width) {
15173 default:
15174 llvm_unreachable("width for stores must be a power of 2")__builtin_unreachable();
15175 case 8:
15176 ConvTy = Int64Ty;
15177 NumElts = 2;
15178 break;
15179 case 4:
15180 ConvTy = Int32Ty;
15181 NumElts = 4;
15182 break;
15183 case 2:
15184 ConvTy = Int16Ty;
15185 NumElts = 8;
15186 break;
15187 case 1:
15188 ConvTy = Int8Ty;
15189 NumElts = 16;
15190 break;
15191 }
15192 Value *Vec = Builder.CreateBitCast(
15193 Ops[2], llvm::FixedVectorType::get(ConvTy, NumElts));
15194 Value *Ptr = Builder.CreateGEP(Int8Ty, Ops[0],
15195 ConstantInt::get(Int64Ty, Offset));
15196 Value *PtrBC = Builder.CreateBitCast(Ptr, ConvTy->getPointerTo());
15197 Value *Elt = Builder.CreateExtractElement(Vec, EltNo);
15198 if (IsLE && Width > 1) {
15199 Function *F = CGM.getIntrinsic(Intrinsic::bswap, ConvTy);
15200 Elt = Builder.CreateCall(F, Elt);
15201 }
15202 return Builder.CreateStore(Elt,
15203 Address(PtrBC, CharUnits::fromQuantity(1)));
15204 };
15205 unsigned Stored = 0;
15206 unsigned RemainingBytes = NumBytes;
15207 Value *Result;
15208 if (NumBytes == 16)
15209 return StoreSubVec(16, 0, 0);
15210 if (NumBytes >= 8) {
15211 Result = StoreSubVec(8, NumBytes - 8, IsLE ? 0 : 1);
15212 RemainingBytes -= 8;
15213 Stored += 8;
15214 }
15215 if (RemainingBytes >= 4) {
15216 Result = StoreSubVec(4, NumBytes - Stored - 4,
15217 IsLE ? (Stored >> 2) : 3 - (Stored >> 2));
15218 RemainingBytes -= 4;
15219 Stored += 4;
15220 }
15221 if (RemainingBytes >= 2) {
15222 Result = StoreSubVec(2, NumBytes - Stored - 2,
15223 IsLE ? (Stored >> 1) : 7 - (Stored >> 1));
15224 RemainingBytes -= 2;
15225 Stored += 2;
15226 }
15227 if (RemainingBytes)
15228 Result =
15229 StoreSubVec(1, NumBytes - Stored - 1, IsLE ? Stored : 15 - Stored);
15230 return Result;
15231 }
15232 // Square root
15233 case PPC::BI__builtin_vsx_xvsqrtsp:
15234 case PPC::BI__builtin_vsx_xvsqrtdp: {
15235 llvm::Type *ResultType = ConvertType(E->getType());
15236 Value *X = EmitScalarExpr(E->getArg(0));
15237 if (Builder.getIsFPConstrained()) {
15238 llvm::Function *F = CGM.getIntrinsic(
15239 Intrinsic::experimental_constrained_sqrt, ResultType);
15240 return Builder.CreateConstrainedFPCall(F, X);
15241 } else {
15242 llvm::Function *F = CGM.getIntrinsic(Intrinsic::sqrt, ResultType);
15243 return Builder.CreateCall(F, X);
15244 }
15245 }
15246 // Count leading zeros
15247 case PPC::BI__builtin_altivec_vclzb:
15248 case PPC::BI__builtin_altivec_vclzh:
15249 case PPC::BI__builtin_altivec_vclzw:
15250 case PPC::BI__builtin_altivec_vclzd: {
15251 llvm::Type *ResultType = ConvertType(E->getType());
15252 Value *X = EmitScalarExpr(E->getArg(0));
15253 Value *Undef = ConstantInt::get(Builder.getInt1Ty(), false);
15254 Function *F = CGM.getIntrinsic(Intrinsic::ctlz, ResultType);
15255 return Builder.CreateCall(F, {X, Undef});
15256 }
15257 case PPC::BI__builtin_altivec_vctzb:
15258 case PPC::BI__builtin_altivec_vctzh:
15259 case PPC::BI__builtin_altivec_vctzw:
15260 case PPC::BI__builtin_altivec_vctzd: {
15261 llvm::Type *ResultType = ConvertType(E->getType());
15262 Value *X = EmitScalarExpr(E->getArg(0));
15263 Value *Undef = ConstantInt::get(Builder.getInt1Ty(), false);
15264 Function *F = CGM.getIntrinsic(Intrinsic::cttz, ResultType);
15265 return Builder.CreateCall(F, {X, Undef});
15266 }
15267 case PPC::BI__builtin_altivec_vec_replace_elt:
15268 case PPC::BI__builtin_altivec_vec_replace_unaligned: {
15269 // The third argument of vec_replace_elt and vec_replace_unaligned must
15270 // be a compile time constant and will be emitted either to the vinsw
15271 // or vinsd instruction.
15272 ConstantInt *ArgCI = dyn_cast<ConstantInt>(Ops[2]);
15273 assert(ArgCI &&((void)0)
15274 "Third Arg to vinsw/vinsd intrinsic must be a constant integer!")((void)0);
15275 llvm::Type *ResultType = ConvertType(E->getType());
15276 llvm::Function *F = nullptr;
15277 Value *Call = nullptr;
15278 int64_t ConstArg = ArgCI->getSExtValue();
15279 unsigned ArgWidth = Ops[1]->getType()->getPrimitiveSizeInBits();
15280 bool Is32Bit = false;
15281 assert((ArgWidth == 32 || ArgWidth == 64) && "Invalid argument width")((void)0);
15282 // The input to vec_replace_elt is an element index, not a byte index.
15283 if (BuiltinID == PPC::BI__builtin_altivec_vec_replace_elt)
15284 ConstArg *= ArgWidth / 8;
15285 if (ArgWidth == 32) {
15286 Is32Bit = true;
15287 // When the second argument is 32 bits, it can either be an integer or
15288 // a float. The vinsw intrinsic is used in this case.
15289 F = CGM.getIntrinsic(Intrinsic::ppc_altivec_vinsw);
15290 // Fix the constant according to endianess.
15291 if (getTarget().isLittleEndian())
15292 ConstArg = 12 - ConstArg;
15293 } else {
15294 // When the second argument is 64 bits, it can either be a long long or
15295 // a double. The vinsd intrinsic is used in this case.
15296 F = CGM.getIntrinsic(Intrinsic::ppc_altivec_vinsd);
15297 // Fix the constant for little endian.
15298 if (getTarget().isLittleEndian())
15299 ConstArg = 8 - ConstArg;
15300 }
15301 Ops[2] = ConstantInt::getSigned(Int32Ty, ConstArg);
15302 // Depending on ArgWidth, the input vector could be a float or a double.
15303 // If the input vector is a float type, bitcast the inputs to integers. Or,
15304 // if the input vector is a double, bitcast the inputs to 64-bit integers.
15305 if (!Ops[1]->getType()->isIntegerTy(ArgWidth)) {
15306 Ops[0] = Builder.CreateBitCast(
15307 Ops[0], Is32Bit ? llvm::FixedVectorType::get(Int32Ty, 4)
15308 : llvm::FixedVectorType::get(Int64Ty, 2));
15309 Ops[1] = Builder.CreateBitCast(Ops[1], Is32Bit ? Int32Ty : Int64Ty);
15310 }
15311 // Emit the call to vinsw or vinsd.
15312 Call = Builder.CreateCall(F, Ops);
15313 // Depending on the builtin, bitcast to the approriate result type.
15314 if (BuiltinID == PPC::BI__builtin_altivec_vec_replace_elt &&
15315 !Ops[1]->getType()->isIntegerTy())
15316 return Builder.CreateBitCast(Call, ResultType);
15317 else if (BuiltinID == PPC::BI__builtin_altivec_vec_replace_elt &&
15318 Ops[1]->getType()->isIntegerTy())
15319 return Call;
15320 else
15321 return Builder.CreateBitCast(Call,
15322 llvm::FixedVectorType::get(Int8Ty, 16));
15323 }
15324 case PPC::BI__builtin_altivec_vpopcntb:
15325 case PPC::BI__builtin_altivec_vpopcnth:
15326 case PPC::BI__builtin_altivec_vpopcntw:
15327 case PPC::BI__builtin_altivec_vpopcntd: {
15328 llvm::Type *ResultType = ConvertType(E->getType());
15329 Value *X = EmitScalarExpr(E->getArg(0));
15330 llvm::Function *F = CGM.getIntrinsic(Intrinsic::ctpop, ResultType);
15331 return Builder.CreateCall(F, X);
15332 }
15333 case PPC::BI__builtin_altivec_vadduqm:
15334 case PPC::BI__builtin_altivec_vsubuqm: {
15335 llvm::Type *Int128Ty = llvm::IntegerType::get(getLLVMContext(), 128);
15336 Ops[0] =
15337 Builder.CreateBitCast(Ops[0], llvm::FixedVectorType::get(Int128Ty, 1));
15338 Ops[1] =
15339 Builder.CreateBitCast(Ops[1], llvm::FixedVectorType::get(Int128Ty, 1));
15340 if (BuiltinID == PPC::BI__builtin_altivec_vadduqm)
15341 return Builder.CreateAdd(Ops[0], Ops[1], "vadduqm");
15342 else
15343 return Builder.CreateSub(Ops[0], Ops[1], "vsubuqm");
15344 }
15345 // Rotate and insert under mask operation.
15346 // __rldimi(rs, is, shift, mask)
15347 // (rotl64(rs, shift) & mask) | (is & ~mask)
15348 // __rlwimi(rs, is, shift, mask)
15349 // (rotl(rs, shift) & mask) | (is & ~mask)
15350 case PPC::BI__builtin_ppc_rldimi:
15351 case PPC::BI__builtin_ppc_rlwimi: {
15352 llvm::Type *Ty = Ops[0]->getType();
15353 Function *F = CGM.getIntrinsic(Intrinsic::fshl, Ty);
15354 if (BuiltinID == PPC::BI__builtin_ppc_rldimi)
15355 Ops[2] = Builder.CreateZExt(Ops[2], Int64Ty);
15356 Value *Shift = Builder.CreateCall(F, {Ops[0], Ops[0], Ops[2]});
15357 Value *X = Builder.CreateAnd(Shift, Ops[3]);
15358 Value *Y = Builder.CreateAnd(Ops[1], Builder.CreateNot(Ops[3]));
15359 return Builder.CreateOr(X, Y);
15360 }
15361 // Rotate and insert under mask operation.
15362 // __rlwnm(rs, shift, mask)
15363 // rotl(rs, shift) & mask
15364 case PPC::BI__builtin_ppc_rlwnm: {
15365 llvm::Type *Ty = Ops[0]->getType();
15366 Function *F = CGM.getIntrinsic(Intrinsic::fshl, Ty);
15367 Value *Shift = Builder.CreateCall(F, {Ops[0], Ops[0], Ops[1]});
15368 return Builder.CreateAnd(Shift, Ops[2]);
15369 }
15370 case PPC::BI__builtin_ppc_poppar4:
15371 case PPC::BI__builtin_ppc_poppar8: {
15372 llvm::Type *ArgType = Ops[0]->getType();
15373 Function *F = CGM.getIntrinsic(Intrinsic::ctpop, ArgType);
15374 Value *Tmp = Builder.CreateCall(F, Ops[0]);
15375
15376 llvm::Type *ResultType = ConvertType(E->getType());
15377 Value *Result = Builder.CreateAnd(Tmp, llvm::ConstantInt::get(ArgType, 1));
15378 if (Result->getType() != ResultType)
15379 Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true,
15380 "cast");
15381 return Result;
15382 }
15383 case PPC::BI__builtin_ppc_cmpb: {
15384 if (getTarget().getTriple().isPPC64()) {
15385 Function *F =
15386 CGM.getIntrinsic(Intrinsic::ppc_cmpb, {Int64Ty, Int64Ty, Int64Ty});
15387 return Builder.CreateCall(F, Ops, "cmpb");
15388 }
15389 // For 32 bit, emit the code as below:
15390 // %conv = trunc i64 %a to i32
15391 // %conv1 = trunc i64 %b to i32
15392 // %shr = lshr i64 %a, 32
15393 // %conv2 = trunc i64 %shr to i32
15394 // %shr3 = lshr i64 %b, 32
15395 // %conv4 = trunc i64 %shr3 to i32
15396 // %0 = tail call i32 @llvm.ppc.cmpb32(i32 %conv, i32 %conv1)
15397 // %conv5 = zext i32 %0 to i64
15398 // %1 = tail call i32 @llvm.ppc.cmpb32(i32 %conv2, i32 %conv4)
15399 // %conv614 = zext i32 %1 to i64
15400 // %shl = shl nuw i64 %conv614, 32
15401 // %or = or i64 %shl, %conv5
15402 // ret i64 %or
15403 Function *F =
15404 CGM.getIntrinsic(Intrinsic::ppc_cmpb, {Int32Ty, Int32Ty, Int32Ty});
15405 Value *ArgOneLo = Builder.CreateTrunc(Ops[0], Int32Ty);
15406 Value *ArgTwoLo = Builder.CreateTrunc(Ops[1], Int32Ty);
15407 Constant *ShiftAmt = ConstantInt::get(Int64Ty, 32);
15408 Value *ArgOneHi =
15409 Builder.CreateTrunc(Builder.CreateLShr(Ops[0], ShiftAmt), Int32Ty);
15410 Value *ArgTwoHi =
15411 Builder.CreateTrunc(Builder.CreateLShr(Ops[1], ShiftAmt), Int32Ty);
15412 Value *ResLo = Builder.CreateZExt(
15413 Builder.CreateCall(F, {ArgOneLo, ArgTwoLo}, "cmpb"), Int64Ty);
15414 Value *ResHiShift = Builder.CreateZExt(
15415 Builder.CreateCall(F, {ArgOneHi, ArgTwoHi}, "cmpb"), Int64Ty);
15416 Value *ResHi = Builder.CreateShl(ResHiShift, ShiftAmt);
15417 return Builder.CreateOr(ResLo, ResHi);
15418 }
15419 // Copy sign
15420 case PPC::BI__builtin_vsx_xvcpsgnsp:
15421 case PPC::BI__builtin_vsx_xvcpsgndp: {
15422 llvm::Type *ResultType = ConvertType(E->getType());
15423 Value *X = EmitScalarExpr(E->getArg(0));
15424 Value *Y = EmitScalarExpr(E->getArg(1));
15425 ID = Intrinsic::copysign;
15426 llvm::Function *F = CGM.getIntrinsic(ID, ResultType);
15427 return Builder.CreateCall(F, {X, Y});
15428 }
15429 // Rounding/truncation
15430 case PPC::BI__builtin_vsx_xvrspip:
15431 case PPC::BI__builtin_vsx_xvrdpip:
15432 case PPC::BI__builtin_vsx_xvrdpim:
15433 case PPC::BI__builtin_vsx_xvrspim:
15434 case PPC::BI__builtin_vsx_xvrdpi:
15435 case PPC::BI__builtin_vsx_xvrspi:
15436 case PPC::BI__builtin_vsx_xvrdpic:
15437 case PPC::BI__builtin_vsx_xvrspic:
15438 case PPC::BI__builtin_vsx_xvrdpiz:
15439 case PPC::BI__builtin_vsx_xvrspiz: {
15440 llvm::Type *ResultType = ConvertType(E->getType());
15441 Value *X = EmitScalarExpr(E->getArg(0));
15442 if (BuiltinID == PPC::BI__builtin_vsx_xvrdpim ||
15443 BuiltinID == PPC::BI__builtin_vsx_xvrspim)
15444 ID = Builder.getIsFPConstrained()
15445 ? Intrinsic::experimental_constrained_floor
15446 : Intrinsic::floor;
15447 else if (BuiltinID == PPC::BI__builtin_vsx_xvrdpi ||
15448 BuiltinID == PPC::BI__builtin_vsx_xvrspi)
15449 ID = Builder.getIsFPConstrained()
15450 ? Intrinsic::experimental_constrained_round
15451 : Intrinsic::round;
15452 else if (BuiltinID == PPC::BI__builtin_vsx_xvrdpic ||
15453 BuiltinID == PPC::BI__builtin_vsx_xvrspic)
15454 ID = Builder.getIsFPConstrained()
15455 ? Intrinsic::experimental_constrained_rint
15456 : Intrinsic::rint;
15457 else if (BuiltinID == PPC::BI__builtin_vsx_xvrdpip ||
15458 BuiltinID == PPC::BI__builtin_vsx_xvrspip)
15459 ID = Builder.getIsFPConstrained()
15460 ? Intrinsic::experimental_constrained_ceil
15461 : Intrinsic::ceil;
15462 else if (BuiltinID == PPC::BI__builtin_vsx_xvrdpiz ||
15463 BuiltinID == PPC::BI__builtin_vsx_xvrspiz)
15464 ID = Builder.getIsFPConstrained()
15465 ? Intrinsic::experimental_constrained_trunc
15466 : Intrinsic::trunc;
15467 llvm::Function *F = CGM.getIntrinsic(ID, ResultType);
15468 return Builder.getIsFPConstrained() ? Builder.CreateConstrainedFPCall(F, X)
15469 : Builder.CreateCall(F, X);
15470 }
15471
15472 // Absolute value
15473 case PPC::BI__builtin_vsx_xvabsdp:
15474 case PPC::BI__builtin_vsx_xvabssp: {
15475 llvm::Type *ResultType = ConvertType(E->getType());
15476 Value *X = EmitScalarExpr(E->getArg(0));
15477 llvm::Function *F = CGM.getIntrinsic(Intrinsic::fabs, ResultType);
15478 return Builder.CreateCall(F, X);
15479 }
15480
15481 // Fastmath by default
15482 case PPC::BI__builtin_ppc_recipdivf:
15483 case PPC::BI__builtin_ppc_recipdivd:
15484 case PPC::BI__builtin_ppc_rsqrtf:
15485 case PPC::BI__builtin_ppc_rsqrtd: {
15486 FastMathFlags FMF = Builder.getFastMathFlags();
15487 Builder.getFastMathFlags().setFast();
15488 llvm::Type *ResultType = ConvertType(E->getType());
15489 Value *X = EmitScalarExpr(E->getArg(0));
15490
15491 if (BuiltinID == PPC::BI__builtin_ppc_recipdivf ||
15492 BuiltinID == PPC::BI__builtin_ppc_recipdivd) {
15493 Value *Y = EmitScalarExpr(E->getArg(1));
15494 Value *FDiv = Builder.CreateFDiv(X, Y, "recipdiv");
15495 Builder.getFastMathFlags() &= (FMF);
15496 return FDiv;
15497 }
15498 auto *One = ConstantFP::get(ResultType, 1.0);
15499 llvm::Function *F = CGM.getIntrinsic(Intrinsic::sqrt, ResultType);
15500 Value *FDiv = Builder.CreateFDiv(One, Builder.CreateCall(F, X), "rsqrt");
15501 Builder.getFastMathFlags() &= (FMF);
15502 return FDiv;
15503 }
15504 case PPC::BI__builtin_ppc_alignx: {
15505 ConstantInt *AlignmentCI = cast<ConstantInt>(Ops[0]);
15506 if (AlignmentCI->getValue().ugt(llvm::Value::MaximumAlignment))
15507 AlignmentCI = ConstantInt::get(AlignmentCI->getType(),
15508 llvm::Value::MaximumAlignment);
15509
15510 emitAlignmentAssumption(Ops[1], E->getArg(1),
15511 /*The expr loc is sufficient.*/ SourceLocation(),
15512 AlignmentCI, nullptr);
15513 return Ops[1];
15514 }
15515 case PPC::BI__builtin_ppc_rdlam: {
15516 llvm::Type *Ty = Ops[0]->getType();
15517 Value *ShiftAmt = Builder.CreateIntCast(Ops[1], Ty, false);
15518 Function *F = CGM.getIntrinsic(Intrinsic::fshl, Ty);
15519 Value *Rotate = Builder.CreateCall(F, {Ops[0], Ops[0], ShiftAmt});
15520 return Builder.CreateAnd(Rotate, Ops[2]);
15521 }
15522 // FMA variations
15523 case PPC::BI__builtin_vsx_xvmaddadp:
15524 case PPC::BI__builtin_vsx_xvmaddasp:
15525 case PPC::BI__builtin_vsx_xvnmaddadp:
15526 case PPC::BI__builtin_vsx_xvnmaddasp:
15527 case PPC::BI__builtin_vsx_xvmsubadp:
15528 case PPC::BI__builtin_vsx_xvmsubasp:
15529 case PPC::BI__builtin_vsx_xvnmsubadp:
15530 case PPC::BI__builtin_vsx_xvnmsubasp: {
15531 llvm::Type *ResultType = ConvertType(E->getType());
15532 Value *X = EmitScalarExpr(E->getArg(0));
15533 Value *Y = EmitScalarExpr(E->getArg(1));
15534 Value *Z = EmitScalarExpr(E->getArg(2));
15535 llvm::Function *F;
15536 if (Builder.getIsFPConstrained())
15537 F = CGM.getIntrinsic(Intrinsic::experimental_constrained_fma, ResultType);
15538 else
15539 F = CGM.getIntrinsic(Intrinsic::fma, ResultType);
15540 switch (BuiltinID) {
15541 case PPC::BI__builtin_vsx_xvmaddadp:
15542 case PPC::BI__builtin_vsx_xvmaddasp:
15543 if (Builder.getIsFPConstrained())
15544 return Builder.CreateConstrainedFPCall(F, {X, Y, Z});
15545 else
15546 return Builder.CreateCall(F, {X, Y, Z});
15547 case PPC::BI__builtin_vsx_xvnmaddadp:
15548 case PPC::BI__builtin_vsx_xvnmaddasp:
15549 if (Builder.getIsFPConstrained())
15550 return Builder.CreateFNeg(
15551 Builder.CreateConstrainedFPCall(F, {X, Y, Z}), "neg");
15552 else
15553 return Builder.CreateFNeg(Builder.CreateCall(F, {X, Y, Z}), "neg");
15554 case PPC::BI__builtin_vsx_xvmsubadp:
15555 case PPC::BI__builtin_vsx_xvmsubasp:
15556 if (Builder.getIsFPConstrained())
15557 return Builder.CreateConstrainedFPCall(
15558 F, {X, Y, Builder.CreateFNeg(Z, "neg")});
15559 else
15560 return Builder.CreateCall(F, {X, Y, Builder.CreateFNeg(Z, "neg")});
15561 case PPC::BI__builtin_vsx_xvnmsubadp:
15562 case PPC::BI__builtin_vsx_xvnmsubasp:
15563 if (Builder.getIsFPConstrained())
15564 return Builder.CreateFNeg(
15565 Builder.CreateConstrainedFPCall(
15566 F, {X, Y, Builder.CreateFNeg(Z, "neg")}),
15567 "neg");
15568 else
15569 return Builder.CreateFNeg(
15570 Builder.CreateCall(F, {X, Y, Builder.CreateFNeg(Z, "neg")}),
15571 "neg");
15572 }
15573 llvm_unreachable("Unknown FMA operation")__builtin_unreachable();
15574 return nullptr; // Suppress no-return warning
15575 }
15576
15577 case PPC::BI__builtin_vsx_insertword: {
15578 llvm::Function *F = CGM.getIntrinsic(Intrinsic::ppc_vsx_xxinsertw);
15579
15580 // Third argument is a compile time constant int. It must be clamped to
15581 // to the range [0, 12].
15582 ConstantInt *ArgCI = dyn_cast<ConstantInt>(Ops[2]);
15583 assert(ArgCI &&((void)0)
15584 "Third arg to xxinsertw intrinsic must be constant integer")((void)0);
15585 const int64_t MaxIndex = 12;
15586 int64_t Index = clamp(ArgCI->getSExtValue(), 0, MaxIndex);
15587
15588 // The builtin semantics don't exactly match the xxinsertw instructions
15589 // semantics (which ppc_vsx_xxinsertw follows). The builtin extracts the
15590 // word from the first argument, and inserts it in the second argument. The
15591 // instruction extracts the word from its second input register and inserts
15592 // it into its first input register, so swap the first and second arguments.
15593 std::swap(Ops[0], Ops[1]);
15594
15595 // Need to cast the second argument from a vector of unsigned int to a
15596 // vector of long long.
15597 Ops[1] =
15598 Builder.CreateBitCast(Ops[1], llvm::FixedVectorType::get(Int64Ty, 2));
15599
15600 if (getTarget().isLittleEndian()) {
15601 // Reverse the double words in the vector we will extract from.
15602 Ops[0] =
15603 Builder.CreateBitCast(Ops[0], llvm::FixedVectorType::get(Int64Ty, 2));
15604 Ops[0] = Builder.CreateShuffleVector(Ops[0], Ops[0], ArrayRef<int>{1, 0});
15605
15606 // Reverse the index.
15607 Index = MaxIndex - Index;
15608 }
15609
15610 // Intrinsic expects the first arg to be a vector of int.
15611 Ops[0] =
15612 Builder.CreateBitCast(Ops[0], llvm::FixedVectorType::get(Int32Ty, 4));
15613 Ops[2] = ConstantInt::getSigned(Int32Ty, Index);
15614 return Builder.CreateCall(F, Ops);
15615 }
15616
15617 case PPC::BI__builtin_vsx_extractuword: {
15618 llvm::Function *F = CGM.getIntrinsic(Intrinsic::ppc_vsx_xxextractuw);
15619
15620 // Intrinsic expects the first argument to be a vector of doublewords.
15621 Ops[0] =
15622 Builder.CreateBitCast(Ops[0], llvm::FixedVectorType::get(Int64Ty, 2));
15623
15624 // The second argument is a compile time constant int that needs to
15625 // be clamped to the range [0, 12].
15626 ConstantInt *ArgCI = dyn_cast<ConstantInt>(Ops[1]);
15627 assert(ArgCI &&((void)0)
15628 "Second Arg to xxextractuw intrinsic must be a constant integer!")((void)0);
15629 const int64_t MaxIndex = 12;
15630 int64_t Index = clamp(ArgCI->getSExtValue(), 0, MaxIndex);
15631
15632 if (getTarget().isLittleEndian()) {
15633 // Reverse the index.
15634 Index = MaxIndex - Index;
15635 Ops[1] = ConstantInt::getSigned(Int32Ty, Index);
15636
15637 // Emit the call, then reverse the double words of the results vector.
15638 Value *Call = Builder.CreateCall(F, Ops);
15639
15640 Value *ShuffleCall =
15641 Builder.CreateShuffleVector(Call, Call, ArrayRef<int>{1, 0});
15642 return ShuffleCall;
15643 } else {
15644 Ops[1] = ConstantInt::getSigned(Int32Ty, Index);
15645 return Builder.CreateCall(F, Ops);
15646 }
15647 }
15648
15649 case PPC::BI__builtin_vsx_xxpermdi: {
15650 ConstantInt *ArgCI = dyn_cast<ConstantInt>(Ops[2]);
15651 assert(ArgCI && "Third arg must be constant integer!")((void)0);
15652
15653 unsigned Index = ArgCI->getZExtValue();
15654 Ops[0] =
15655 Builder.CreateBitCast(Ops[0], llvm::FixedVectorType::get(Int64Ty, 2));
15656 Ops[1] =
15657 Builder.CreateBitCast(Ops[1], llvm::FixedVectorType::get(Int64Ty, 2));
15658
15659 // Account for endianness by treating this as just a shuffle. So we use the
15660 // same indices for both LE and BE in order to produce expected results in
15661 // both cases.
15662 int ElemIdx0 = (Index & 2) >> 1;
15663 int ElemIdx1 = 2 + (Index & 1);
15664
15665 int ShuffleElts[2] = {ElemIdx0, ElemIdx1};
15666 Value *ShuffleCall =
15667 Builder.CreateShuffleVector(Ops[0], Ops[1], ShuffleElts);
15668 QualType BIRetType = E->getType();
15669 auto RetTy = ConvertType(BIRetType);
15670 return Builder.CreateBitCast(ShuffleCall, RetTy);
15671 }
15672
15673 case PPC::BI__builtin_vsx_xxsldwi: {
15674 ConstantInt *ArgCI = dyn_cast<ConstantInt>(Ops[2]);
4
Assuming the object is not a 'ConstantInt'
5
'ArgCI' initialized to a null pointer value
15675 assert(ArgCI && "Third argument must be a compile time constant")((void)0);
15676 unsigned Index = ArgCI->getZExtValue() & 0x3;
6
Called C++ object pointer is null
15677 Ops[0] =
15678 Builder.CreateBitCast(Ops[0], llvm::FixedVectorType::get(Int32Ty, 4));
15679 Ops[1] =
15680 Builder.CreateBitCast(Ops[1], llvm::FixedVectorType::get(Int32Ty, 4));
15681
15682 // Create a shuffle mask
15683 int ElemIdx0;
15684 int ElemIdx1;
15685 int ElemIdx2;
15686 int ElemIdx3;
15687 if (getTarget().isLittleEndian()) {
15688 // Little endian element N comes from element 8+N-Index of the
15689 // concatenated wide vector (of course, using modulo arithmetic on
15690 // the total number of elements).
15691 ElemIdx0 = (8 - Index) % 8;
15692 ElemIdx1 = (9 - Index) % 8;
15693 ElemIdx2 = (10 - Index) % 8;
15694 ElemIdx3 = (11 - Index) % 8;
15695 } else {
15696 // Big endian ElemIdx<N> = Index + N
15697 ElemIdx0 = Index;
15698 ElemIdx1 = Index + 1;
15699 ElemIdx2 = Index + 2;
15700 ElemIdx3 = Index + 3;
15701 }
15702
15703 int ShuffleElts[4] = {ElemIdx0, ElemIdx1, ElemIdx2, ElemIdx3};
15704 Value *ShuffleCall =
15705 Builder.CreateShuffleVector(Ops[0], Ops[1], ShuffleElts);
15706 QualType BIRetType = E->getType();
15707 auto RetTy = ConvertType(BIRetType);
15708 return Builder.CreateBitCast(ShuffleCall, RetTy);
15709 }
15710
15711 case PPC::BI__builtin_pack_vector_int128: {
15712 bool isLittleEndian = getTarget().isLittleEndian();
15713 Value *UndefValue =
15714 llvm::UndefValue::get(llvm::FixedVectorType::get(Ops[0]->getType(), 2));
15715 Value *Res = Builder.CreateInsertElement(
15716 UndefValue, Ops[0], (uint64_t)(isLittleEndian ? 1 : 0));
15717 Res = Builder.CreateInsertElement(Res, Ops[1],
15718 (uint64_t)(isLittleEndian ? 0 : 1));
15719 return Builder.CreateBitCast(Res, ConvertType(E->getType()));
15720 }
15721
15722 case PPC::BI__builtin_unpack_vector_int128: {
15723 ConstantInt *Index = cast<ConstantInt>(Ops[1]);
15724 Value *Unpacked = Builder.CreateBitCast(
15725 Ops[0], llvm::FixedVectorType::get(ConvertType(E->getType()), 2));
15726
15727 if (getTarget().isLittleEndian())
15728 Index = ConstantInt::get(Index->getType(), 1 - Index->getZExtValue());
15729
15730 return Builder.CreateExtractElement(Unpacked, Index);
15731 }
15732
15733 case PPC::BI__builtin_ppc_sthcx: {
15734 llvm::Function *F = CGM.getIntrinsic(Intrinsic::ppc_sthcx);
15735 Ops[0] = Builder.CreateBitCast(Ops[0], Int8PtrTy);
15736 Ops[1] = Builder.CreateSExt(Ops[1], Int32Ty);
15737 return Builder.CreateCall(F, Ops);
15738 }
15739
15740 // The PPC MMA builtins take a pointer to a __vector_quad as an argument.
15741 // Some of the MMA instructions accumulate their result into an existing
15742 // accumulator whereas the others generate a new accumulator. So we need to
15743 // use custom code generation to expand a builtin call with a pointer to a
15744 // load (if the corresponding instruction accumulates its result) followed by
15745 // the call to the intrinsic and a store of the result.
15746#define CUSTOM_BUILTIN(Name, Intr, Types, Accumulate) \
15747 case PPC::BI__builtin_##Name:
15748#include "clang/Basic/BuiltinsPPC.def"
15749 {
15750 // The first argument of these two builtins is a pointer used to store their
15751 // result. However, the llvm intrinsics return their result in multiple
15752 // return values. So, here we emit code extracting these values from the
15753 // intrinsic results and storing them using that pointer.
15754 if (BuiltinID == PPC::BI__builtin_mma_disassemble_acc ||
15755 BuiltinID == PPC::BI__builtin_vsx_disassemble_pair ||
15756 BuiltinID == PPC::BI__builtin_mma_disassemble_pair) {
15757 unsigned NumVecs = 2;
15758 auto Intrinsic = Intrinsic::ppc_vsx_disassemble_pair;
15759 if (BuiltinID == PPC::BI__builtin_mma_disassemble_acc) {
15760 NumVecs = 4;
15761 Intrinsic = Intrinsic::ppc_mma_disassemble_acc;
15762 }
15763 llvm::Function *F = CGM.getIntrinsic(Intrinsic);
15764 Address Addr = EmitPointerWithAlignment(E->getArg(1));
15765 Value *Vec = Builder.CreateLoad(Addr);
15766 Value *Call = Builder.CreateCall(F, {Vec});
15767 llvm::Type *VTy = llvm::FixedVectorType::get(Int8Ty, 16);
15768 Value *Ptr = Builder.CreateBitCast(Ops[0], VTy->getPointerTo());
15769 for (unsigned i=0; i<NumVecs; i++) {
15770 Value *Vec = Builder.CreateExtractValue(Call, i);
15771 llvm::ConstantInt* Index = llvm::ConstantInt::get(IntTy, i);
15772 Value *GEP = Builder.CreateInBoundsGEP(VTy, Ptr, Index);
15773 Builder.CreateAlignedStore(Vec, GEP, MaybeAlign(16));
15774 }
15775 return Call;
15776 }
15777 bool Accumulate;
15778 switch (BuiltinID) {
15779 #define CUSTOM_BUILTIN(Name, Intr, Types, Acc) \
15780 case PPC::BI__builtin_##Name: \
15781 ID = Intrinsic::ppc_##Intr; \
15782 Accumulate = Acc; \
15783 break;
15784 #include "clang/Basic/BuiltinsPPC.def"
15785 }
15786 if (BuiltinID == PPC::BI__builtin_vsx_lxvp ||
15787 BuiltinID == PPC::BI__builtin_vsx_stxvp ||
15788 BuiltinID == PPC::BI__builtin_mma_lxvp ||
15789 BuiltinID == PPC::BI__builtin_mma_stxvp) {
15790 if (BuiltinID == PPC::BI__builtin_vsx_lxvp ||
15791 BuiltinID == PPC::BI__builtin_mma_lxvp) {
15792 Ops[1] = Builder.CreateBitCast(Ops[1], Int8PtrTy);
15793 Ops[0] = Builder.CreateGEP(Int8Ty, Ops[1], Ops[0]);
15794 } else {
15795 Ops[2] = Builder.CreateBitCast(Ops[2], Int8PtrTy);
15796 Ops[1] = Builder.CreateGEP(Int8Ty, Ops[2], Ops[1]);
15797 }
15798 Ops.pop_back();
15799 llvm::Function *F = CGM.getIntrinsic(ID);
15800 return Builder.CreateCall(F, Ops, "");
15801 }
15802 SmallVector<Value*, 4> CallOps;
15803 if (Accumulate) {
15804 Address Addr = EmitPointerWithAlignment(E->getArg(0));
15805 Value *Acc = Builder.CreateLoad(Addr);
15806 CallOps.push_back(Acc);
15807 }
15808 for (unsigned i=1; i<Ops.size(); i++)
15809 CallOps.push_back(Ops[i]);
15810 llvm::Function *F = CGM.getIntrinsic(ID);
15811 Value *Call = Builder.CreateCall(F, CallOps);
15812 return Builder.CreateAlignedStore(Call, Ops[0], MaybeAlign(64));
15813 }
15814
15815 case PPC::BI__builtin_ppc_compare_and_swap:
15816 case PPC::BI__builtin_ppc_compare_and_swaplp: {
15817 Address Addr = EmitPointerWithAlignment(E->getArg(0));
15818 Address OldValAddr = EmitPointerWithAlignment(E->getArg(1));
15819 Value *OldVal = Builder.CreateLoad(OldValAddr);
15820 QualType AtomicTy = E->getArg(0)->getType()->getPointeeType();
15821 LValue LV = MakeAddrLValue(Addr, AtomicTy);
15822 auto Pair = EmitAtomicCompareExchange(
15823 LV, RValue::get(OldVal), RValue::get(Ops[2]), E->getExprLoc(),
15824 llvm::AtomicOrdering::Monotonic, llvm::AtomicOrdering::Monotonic, true);
15825 // Unlike c11's atomic_compare_exchange, accroding to
15826 // https://www.ibm.com/docs/en/xl-c-and-cpp-aix/16.1?topic=functions-compare-swap-compare-swaplp
15827 // > In either case, the contents of the memory location specified by addr
15828 // > are copied into the memory location specified by old_val_addr.
15829 // But it hasn't specified storing to OldValAddr is atomic or not and
15830 // which order to use. Now following XL's codegen, treat it as a normal
15831 // store.
15832 Value *LoadedVal = Pair.first.getScalarVal();
15833 Builder.CreateStore(LoadedVal, OldValAddr);
15834 return Pair.second;
15835 }
15836 case PPC::BI__builtin_ppc_fetch_and_add:
15837 case PPC::BI__builtin_ppc_fetch_and_addlp: {
15838 return MakeBinaryAtomicValue(*this, AtomicRMWInst::Add, E,
15839 llvm::AtomicOrdering::Monotonic);
15840 }
15841 case PPC::BI__builtin_ppc_fetch_and_and:
15842 case PPC::BI__builtin_ppc_fetch_and_andlp: {
15843 return MakeBinaryAtomicValue(*this, AtomicRMWInst::And, E,
15844 llvm::AtomicOrdering::Monotonic);
15845 }
15846
15847 case PPC::BI__builtin_ppc_fetch_and_or:
15848 case PPC::BI__builtin_ppc_fetch_and_orlp: {
15849 return MakeBinaryAtomicValue(*this, AtomicRMWInst::Or, E,
15850 llvm::AtomicOrdering::Monotonic);
15851 }
15852 case PPC::BI__builtin_ppc_fetch_and_swap:
15853 case PPC::BI__builtin_ppc_fetch_and_swaplp: {
15854 return MakeBinaryAtomicValue(*this, AtomicRMWInst::Xchg, E,
15855 llvm::AtomicOrdering::Monotonic);
15856 }
15857 case PPC::BI__builtin_ppc_ldarx:
15858 case PPC::BI__builtin_ppc_lwarx:
15859 case PPC::BI__builtin_ppc_lharx:
15860 case PPC::BI__builtin_ppc_lbarx:
15861 return emitPPCLoadReserveIntrinsic(*this, BuiltinID, E);
15862 case PPC::BI__builtin_ppc_mfspr: {
15863 llvm::Type *RetType = CGM.getDataLayout().getTypeSizeInBits(VoidPtrTy) == 32
15864 ? Int32Ty
15865 : Int64Ty;
15866 Function *F = CGM.getIntrinsic(Intrinsic::ppc_mfspr, RetType);
15867 return Builder.CreateCall(F, Ops);
15868 }
15869 case PPC::BI__builtin_ppc_mtspr: {
15870 llvm::Type *RetType = CGM.getDataLayout().getTypeSizeInBits(VoidPtrTy) == 32
15871 ? Int32Ty
15872 : Int64Ty;
15873 Function *F = CGM.getIntrinsic(Intrinsic::ppc_mtspr, RetType);
15874 return Builder.CreateCall(F, Ops);
15875 }
15876 case PPC::BI__builtin_ppc_popcntb: {
15877 Value *ArgValue = EmitScalarExpr(E->getArg(0));
15878 llvm::Type *ArgType = ArgValue->getType();
15879 Function *F = CGM.getIntrinsic(Intrinsic::ppc_popcntb, {ArgType, ArgType});
15880 return Builder.CreateCall(F, Ops, "popcntb");
15881 }
15882 case PPC::BI__builtin_ppc_mtfsf: {
15883 // The builtin takes a uint32 that needs to be cast to an
15884 // f64 to be passed to the intrinsic.
15885 Value *Cast = Builder.CreateUIToFP(Ops[1], DoubleTy);
15886 llvm::Function *F = CGM.getIntrinsic(Intrinsic::ppc_mtfsf);
15887 return Builder.CreateCall(F, {Ops[0], Cast}, "");
15888 }
15889
15890 case PPC::BI__builtin_ppc_swdiv_nochk:
15891 case PPC::BI__builtin_ppc_swdivs_nochk: {
15892 FastMathFlags FMF = Builder.getFastMathFlags();
15893 Builder.getFastMathFlags().setFast();
15894 Value *FDiv = Builder.CreateFDiv(Ops[0], Ops[1], "swdiv_nochk");
15895 Builder.getFastMathFlags() &= (FMF);
15896 return FDiv;
15897 }
15898 case PPC::BI__builtin_ppc_fric:
15899 return RValue::get(emitUnaryMaybeConstrainedFPBuiltin(
15900 *this, E, Intrinsic::rint,
15901 Intrinsic::experimental_constrained_rint))
15902 .getScalarVal();
15903 case PPC::BI__builtin_ppc_frim:
15904 case PPC::BI__builtin_ppc_frims:
15905 return RValue::get(emitUnaryMaybeConstrainedFPBuiltin(
15906 *this, E, Intrinsic::floor,
15907 Intrinsic::experimental_constrained_floor))
15908 .getScalarVal();
15909 case PPC::BI__builtin_ppc_frin:
15910 case PPC::BI__builtin_ppc_frins:
15911 return RValue::get(emitUnaryMaybeConstrainedFPBuiltin(
15912 *this, E, Intrinsic::round,
15913 Intrinsic::experimental_constrained_round))
15914 .getScalarVal();
15915 case PPC::BI__builtin_ppc_frip:
15916 case PPC::BI__builtin_ppc_frips:
15917 return RValue::get(emitUnaryMaybeConstrainedFPBuiltin(
15918 *this, E, Intrinsic::ceil,
15919 Intrinsic::experimental_constrained_ceil))
15920 .getScalarVal();
15921 case PPC::BI__builtin_ppc_friz:
15922 case PPC::BI__builtin_ppc_frizs:
15923 return RValue::get(emitUnaryMaybeConstrainedFPBuiltin(
15924 *this, E, Intrinsic::trunc,
15925 Intrinsic::experimental_constrained_trunc))
15926 .getScalarVal();
15927 case PPC::BI__builtin_ppc_fsqrt:
15928 case PPC::BI__builtin_ppc_fsqrts:
15929 return RValue::get(emitUnaryMaybeConstrainedFPBuiltin(
15930 *this, E, Intrinsic::sqrt,
15931 Intrinsic::experimental_constrained_sqrt))
15932 .getScalarVal();
15933 }
15934}
15935
15936namespace {
15937// If \p E is not null pointer, insert address space cast to match return
15938// type of \p E if necessary.
15939Value *EmitAMDGPUDispatchPtr(CodeGenFunction &CGF,
15940 const CallExpr *E = nullptr) {
15941 auto *F = CGF.CGM.getIntrinsic(Intrinsic::amdgcn_dispatch_ptr);
15942 auto *Call = CGF.Builder.CreateCall(F);
15943 Call->addAttribute(
15944 AttributeList::ReturnIndex,
15945 Attribute::getWithDereferenceableBytes(Call->getContext(), 64));
15946 Call->addAttribute(AttributeList::ReturnIndex,
15947 Attribute::getWithAlignment(Call->getContext(), Align(4)));
15948 if (!E)
15949 return Call;
15950 QualType BuiltinRetType = E->getType();
15951 auto *RetTy = cast<llvm::PointerType>(CGF.ConvertType(BuiltinRetType));
15952 if (RetTy == Call->getType())
15953 return Call;
15954 return CGF.Builder.CreateAddrSpaceCast(Call, RetTy);
15955}
15956
15957// \p Index is 0, 1, and 2 for x, y, and z dimension, respectively.
15958Value *EmitAMDGPUWorkGroupSize(CodeGenFunction &CGF, unsigned Index) {
15959 const unsigned XOffset = 4;
15960 auto *DP = EmitAMDGPUDispatchPtr(CGF);
15961 // Indexing the HSA kernel_dispatch_packet struct.
15962 auto *Offset = llvm::ConstantInt::get(CGF.Int32Ty, XOffset + Index * 2);
15963 auto *GEP = CGF.Builder.CreateGEP(CGF.Int8Ty, DP, Offset);
15964 auto *DstTy =
15965 CGF.Int16Ty->getPointerTo(GEP->getType()->getPointerAddressSpace());
15966 auto *Cast = CGF.Builder.CreateBitCast(GEP, DstTy);
15967 auto *LD = CGF.Builder.CreateLoad(Address(Cast, CharUnits::fromQuantity(2)));
15968 llvm::MDBuilder MDHelper(CGF.getLLVMContext());
15969 llvm::MDNode *RNode = MDHelper.createRange(APInt(16, 1),
15970 APInt(16, CGF.getTarget().getMaxOpenCLWorkGroupSize() + 1));
15971 LD->setMetadata(llvm::LLVMContext::MD_range, RNode);
15972 LD->setMetadata(llvm::LLVMContext::MD_invariant_load,
15973 llvm::MDNode::get(CGF.getLLVMContext(), None));
15974 return LD;
15975}
15976
15977// \p Index is 0, 1, and 2 for x, y, and z dimension, respectively.
15978Value *EmitAMDGPUGridSize(CodeGenFunction &CGF, unsigned Index) {
15979 const unsigned XOffset = 12;
15980 auto *DP = EmitAMDGPUDispatchPtr(CGF);
15981 // Indexing the HSA kernel_dispatch_packet struct.
15982 auto *Offset = llvm::ConstantInt::get(CGF.Int32Ty, XOffset + Index * 4);
15983 auto *GEP = CGF.Builder.CreateGEP(CGF.Int8Ty, DP, Offset);
15984 auto *DstTy =
15985 CGF.Int32Ty->getPointerTo(GEP->getType()->getPointerAddressSpace());
15986 auto *Cast = CGF.Builder.CreateBitCast(GEP, DstTy);
15987 auto *LD = CGF.Builder.CreateLoad(Address(Cast, CharUnits::fromQuantity(4)));
15988 LD->setMetadata(llvm::LLVMContext::MD_invariant_load,
15989 llvm::MDNode::get(CGF.getLLVMContext(), None));
15990 return LD;
15991}
15992} // namespace
15993
15994// For processing memory ordering and memory scope arguments of various
15995// amdgcn builtins.
15996// \p Order takes a C++11 comptabile memory-ordering specifier and converts
15997// it into LLVM's memory ordering specifier using atomic C ABI, and writes
15998// to \p AO. \p Scope takes a const char * and converts it into AMDGCN
15999// specific SyncScopeID and writes it to \p SSID.
16000bool CodeGenFunction::ProcessOrderScopeAMDGCN(Value *Order, Value *Scope,
16001 llvm::AtomicOrdering &AO,
16002 llvm::SyncScope::ID &SSID) {
16003 if (isa<llvm::ConstantInt>(Order)) {
16004 int ord = cast<llvm::ConstantInt>(Order)->getZExtValue();
16005
16006 // Map C11/C++11 memory ordering to LLVM memory ordering
16007 assert(llvm::isValidAtomicOrderingCABI(ord))((void)0);
16008 switch (static_cast<llvm::AtomicOrderingCABI>(ord)) {
16009 case llvm::AtomicOrderingCABI::acquire:
16010 case llvm::AtomicOrderingCABI::consume:
16011 AO = llvm::AtomicOrdering::Acquire;
16012 break;
16013 case llvm::AtomicOrderingCABI::release:
16014 AO = llvm::AtomicOrdering::Release;
16015 break;
16016 case llvm::AtomicOrderingCABI::acq_rel:
16017 AO = llvm::AtomicOrdering::AcquireRelease;
16018 break;
16019 case llvm::AtomicOrderingCABI::seq_cst:
16020 AO = llvm::AtomicOrdering::SequentiallyConsistent;
16021 break;
16022 case llvm::AtomicOrderingCABI::relaxed:
16023 AO = llvm::AtomicOrdering::Monotonic;
16024 break;
16025 }
16026
16027 StringRef scp;
16028 llvm::getConstantStringInfo(Scope, scp);
16029 SSID = getLLVMContext().getOrInsertSyncScopeID(scp);
16030 return true;
16031 }
16032 return false;
16033}
16034
16035Value *CodeGenFunction::EmitAMDGPUBuiltinExpr(unsigned BuiltinID,
16036 const CallExpr *E) {
16037 llvm::AtomicOrdering AO = llvm::AtomicOrdering::SequentiallyConsistent;
16038 llvm::SyncScope::ID SSID;
16039 switch (BuiltinID) {
16040 case AMDGPU::BI__builtin_amdgcn_div_scale:
16041 case AMDGPU::BI__builtin_amdgcn_div_scalef: {
16042 // Translate from the intrinsics's struct return to the builtin's out
16043 // argument.
16044
16045 Address FlagOutPtr = EmitPointerWithAlignment(E->getArg(3));
16046
16047 llvm::Value *X = EmitScalarExpr(E->getArg(0));
16048 llvm::Value *Y = EmitScalarExpr(E->getArg(1));
16049 llvm::Value *Z = EmitScalarExpr(E->getArg(2));
16050
16051 llvm::Function *Callee = CGM.getIntrinsic(Intrinsic::amdgcn_div_scale,
16052 X->getType());
16053
16054 llvm::Value *Tmp = Builder.CreateCall(Callee, {X, Y, Z});
16055
16056 llvm::Value *Result = Builder.CreateExtractValue(Tmp, 0);
16057 llvm::Value *Flag = Builder.CreateExtractValue(Tmp, 1);
16058
16059 llvm::Type *RealFlagType
16060 = FlagOutPtr.getPointer()->getType()->getPointerElementType();
16061
16062 llvm::Value *FlagExt = Builder.CreateZExt(Flag, RealFlagType);
16063 Builder.CreateStore(FlagExt, FlagOutPtr);
16064 return Result;
16065 }
16066 case AMDGPU::BI__builtin_amdgcn_div_fmas:
16067 case AMDGPU::BI__builtin_amdgcn_div_fmasf: {
16068 llvm::Value *Src0 = EmitScalarExpr(E->getArg(0));
16069 llvm::Value *Src1 = EmitScalarExpr(E->getArg(1));
16070 llvm::Value *Src2 = EmitScalarExpr(E->getArg(2));
16071 llvm::Value *Src3 = EmitScalarExpr(E->getArg(3));
16072
16073 llvm::Function *F = CGM.getIntrinsic(Intrinsic::amdgcn_div_fmas,
16074 Src0->getType());
16075 llvm::Value *Src3ToBool = Builder.CreateIsNotNull(Src3);
16076 return Builder.CreateCall(F, {Src0, Src1, Src2, Src3ToBool});
16077 }
16078
16079 case AMDGPU::BI__builtin_amdgcn_ds_swizzle:
16080 return emitBinaryBuiltin(*this, E, Intrinsic::amdgcn_ds_swizzle);
16081 case AMDGPU::BI__builtin_amdgcn_mov_dpp8:
16082 return emitBinaryBuiltin(*this, E, Intrinsic::amdgcn_mov_dpp8);
16083 case AMDGPU::BI__builtin_amdgcn_mov_dpp:
16084 case AMDGPU::BI__builtin_amdgcn_update_dpp: {
16085 llvm::SmallVector<llvm::Value *, 6> Args;
16086 for (unsigned I = 0; I != E->getNumArgs(); ++I)
16087 Args.push_back(EmitScalarExpr(E->getArg(I)));
16088 assert(Args.size() == 5 || Args.size() == 6)((void)0);
16089 if (Args.size() == 5)
16090 Args.insert(Args.begin(), llvm::UndefValue::get(Args[0]->getType()));
16091 Function *F =
16092 CGM.getIntrinsic(Intrinsic::amdgcn_update_dpp, Args[0]->getType());
16093 return Builder.CreateCall(F, Args);
16094 }
16095 case AMDGPU::BI__builtin_amdgcn_div_fixup:
16096 case AMDGPU::BI__builtin_amdgcn_div_fixupf:
16097 case AMDGPU::BI__builtin_amdgcn_div_fixuph:
16098 return emitTernaryBuiltin(*this, E, Intrinsic::amdgcn_div_fixup);
16099 case AMDGPU::BI__builtin_amdgcn_trig_preop:
16100 case AMDGPU::BI__builtin_amdgcn_trig_preopf:
16101 return emitFPIntBuiltin(*this, E, Intrinsic::amdgcn_trig_preop);
16102 case AMDGPU::BI__builtin_amdgcn_rcp:
16103 case AMDGPU::BI__builtin_amdgcn_rcpf:
16104 case AMDGPU::BI__builtin_amdgcn_rcph:
16105 return emitUnaryBuiltin(*this, E, Intrinsic::amdgcn_rcp);
16106 case AMDGPU::BI__builtin_amdgcn_sqrt:
16107 case AMDGPU::BI__builtin_amdgcn_sqrtf:
16108 case AMDGPU::BI__builtin_amdgcn_sqrth:
16109 return emitUnaryBuiltin(*this, E, Intrinsic::amdgcn_sqrt);
16110 case AMDGPU::BI__builtin_amdgcn_rsq:
16111 case AMDGPU::BI__builtin_amdgcn_rsqf:
16112 case AMDGPU::BI__builtin_amdgcn_rsqh:
16113 return emitUnaryBuiltin(*this, E, Intrinsic::amdgcn_rsq);
16114 case AMDGPU::BI__builtin_amdgcn_rsq_clamp:
16115 case AMDGPU::BI__builtin_amdgcn_rsq_clampf:
16116 return emitUnaryBuiltin(*this, E, Intrinsic::amdgcn_rsq_clamp);
16117 case AMDGPU::BI__builtin_amdgcn_sinf:
16118 case AMDGPU::BI__builtin_amdgcn_sinh:
16119 return emitUnaryBuiltin(*this, E, Intrinsic::amdgcn_sin);
16120 case AMDGPU::BI__builtin_amdgcn_cosf:
16121 case AMDGPU::BI__builtin_amdgcn_cosh:
16122 return emitUnaryBuiltin(*this, E, Intrinsic::amdgcn_cos);
16123 case AMDGPU::BI__builtin_amdgcn_dispatch_ptr:
16124 return EmitAMDGPUDispatchPtr(*this, E);
16125 case AMDGPU::BI__builtin_amdgcn_log_clampf:
16126 return emitUnaryBuiltin(*this, E, Intrinsic::amdgcn_log_clamp);
16127 case AMDGPU::BI__builtin_amdgcn_ldexp:
16128 case AMDGPU::BI__builtin_amdgcn_ldexpf:
16129 case AMDGPU::BI__builtin_amdgcn_ldexph:
16130 return emitFPIntBuiltin(*this, E, Intrinsic::amdgcn_ldexp);
16131 case AMDGPU::BI__builtin_amdgcn_frexp_mant:
16132 case AMDGPU::BI__builtin_amdgcn_frexp_mantf:
16133 case AMDGPU::BI__builtin_amdgcn_frexp_manth:
16134 return emitUnaryBuiltin(*this, E, Intrinsic::amdgcn_frexp_mant);
16135 case AMDGPU::BI__builtin_amdgcn_frexp_exp:
16136 case AMDGPU::BI__builtin_amdgcn_frexp_expf: {
16137 Value *Src0 = EmitScalarExpr(E->getArg(0));
16138 Function *F = CGM.getIntrinsic(Intrinsic::amdgcn_frexp_exp,
16139 { Builder.getInt32Ty(), Src0->getType() });
16140 return Builder.CreateCall(F, Src0);
16141 }
16142 case AMDGPU::BI__builtin_amdgcn_frexp_exph: {
16143 Value *Src0 = EmitScalarExpr(E->getArg(0));
16144 Function *F = CGM.getIntrinsic(Intrinsic::amdgcn_frexp_exp,
16145 { Builder.getInt16Ty(), Src0->getType() });
16146 return Builder.CreateCall(F, Src0);
16147 }
16148 case AMDGPU::BI__builtin_amdgcn_fract:
16149 case AMDGPU::BI__builtin_amdgcn_fractf:
16150 case AMDGPU::BI__builtin_amdgcn_fracth:
16151 return emitUnaryBuiltin(*this, E, Intrinsic::amdgcn_fract);
16152 case AMDGPU::BI__builtin_amdgcn_lerp:
16153 return emitTernaryBuiltin(*this, E, Intrinsic::amdgcn_lerp);
16154 case AMDGPU::BI__builtin_amdgcn_ubfe:
16155 return emitTernaryBuiltin(*this, E, Intrinsic::amdgcn_ubfe);
16156 case AMDGPU::BI__builtin_amdgcn_sbfe:
16157 return emitTernaryBuiltin(*this, E, Intrinsic::amdgcn_sbfe);
16158 case AMDGPU::BI__builtin_amdgcn_uicmp:
16159 case AMDGPU::BI__builtin_amdgcn_uicmpl:
16160 case AMDGPU::BI__builtin_amdgcn_sicmp:
16161 case AMDGPU::BI__builtin_amdgcn_sicmpl: {
16162 llvm::Value *Src0 = EmitScalarExpr(E->getArg(0));
16163 llvm::Value *Src1 = EmitScalarExpr(E->getArg(1));
16164 llvm::Value *Src2 = EmitScalarExpr(E->getArg(2));
16165
16166 // FIXME-GFX10: How should 32 bit mask be handled?
16167 Function *F = CGM.getIntrinsic(Intrinsic::amdgcn_icmp,
16168 { Builder.getInt64Ty(), Src0->getType() });
16169 return Builder.CreateCall(F, { Src0, Src1, Src2 });
16170 }
16171 case AMDGPU::BI__builtin_amdgcn_fcmp:
16172 case AMDGPU::BI__builtin_amdgcn_fcmpf: {
16173 llvm::Value *Src0 = EmitScalarExpr(E->getArg(0));
16174 llvm::Value *Src1 = EmitScalarExpr(E->getArg(1));
16175 llvm::Value *Src2 = EmitScalarExpr(E->getArg(2));
16176
16177 // FIXME-GFX10: How should 32 bit mask be handled?
16178 Function *F = CGM.getIntrinsic(Intrinsic::amdgcn_fcmp,
16179 { Builder.getInt64Ty(), Src0->getType() });
16180 return Builder.CreateCall(F, { Src0, Src1, Src2 });
16181 }
16182 case AMDGPU::BI__builtin_amdgcn_class:
16183 case AMDGPU::BI__builtin_amdgcn_classf:
16184 case AMDGPU::BI__builtin_amdgcn_classh:
16185 return emitFPIntBuiltin(*this, E, Intrinsic::amdgcn_class);
16186 case AMDGPU::BI__builtin_amdgcn_fmed3f:
16187 case AMDGPU::BI__builtin_amdgcn_fmed3h:
16188 return emitTernaryBuiltin(*this, E, Intrinsic::amdgcn_fmed3);
16189 case AMDGPU::BI__builtin_amdgcn_ds_append:
16190 case AMDGPU::BI__builtin_amdgcn_ds_consume: {
16191 Intrinsic::ID Intrin = BuiltinID == AMDGPU::BI__builtin_amdgcn_ds_append ?
16192 Intrinsic::amdgcn_ds_append : Intrinsic::amdgcn_ds_consume;
16193 Value *Src0 = EmitScalarExpr(E->getArg(0));
16194 Function *F = CGM.getIntrinsic(Intrin, { Src0->getType() });
16195 return Builder.CreateCall(F, { Src0, Builder.getFalse() });
16196 }
16197 case AMDGPU::BI__builtin_amdgcn_ds_faddf:
16198 case AMDGPU::BI__builtin_amdgcn_ds_fminf:
16199 case AMDGPU::BI__builtin_amdgcn_ds_fmaxf: {
16200 Intrinsic::ID Intrin;
16201 switch (BuiltinID) {
16202 case AMDGPU::BI__builtin_amdgcn_ds_faddf:
16203 Intrin = Intrinsic::amdgcn_ds_fadd;
16204 break;
16205 case AMDGPU::BI__builtin_amdgcn_ds_fminf:
16206 Intrin = Intrinsic::amdgcn_ds_fmin;
16207 break;
16208 case AMDGPU::BI__builtin_amdgcn_ds_fmaxf:
16209 Intrin = Intrinsic::amdgcn_ds_fmax;
16210 break;
16211 }
16212 llvm::Value *Src0 = EmitScalarExpr(E->getArg(0));
16213 llvm::Value *Src1 = EmitScalarExpr(E->getArg(1));
16214 llvm::Value *Src2 = EmitScalarExpr(E->getArg(2));
16215 llvm::Value *Src3 = EmitScalarExpr(E->getArg(3));
16216 llvm::Value *Src4 = EmitScalarExpr(E->getArg(4));
16217 llvm::Function *F = CGM.getIntrinsic(Intrin, { Src1->getType() });
16218 llvm::FunctionType *FTy = F->getFunctionType();
16219 llvm::Type *PTy = FTy->getParamType(0);
16220 Src0 = Builder.CreatePointerBitCastOrAddrSpaceCast(Src0, PTy);
16221 return Builder.CreateCall(F, { Src0, Src1, Src2, Src3, Src4 });
16222 }
16223 case AMDGPU::BI__builtin_amdgcn_read_exec: {
16224 CallInst *CI = cast<CallInst>(
16225 EmitSpecialRegisterBuiltin(*this, E, Int64Ty, Int64Ty, NormalRead, "exec"));
16226 CI->setConvergent();
16227 return CI;
16228 }
16229 case AMDGPU::BI__builtin_amdgcn_read_exec_lo:
16230 case AMDGPU::BI__builtin_amdgcn_read_exec_hi: {
16231 StringRef RegName = BuiltinID == AMDGPU::BI__builtin_amdgcn_read_exec_lo ?
16232 "exec_lo" : "exec_hi";
16233 CallInst *CI = cast<CallInst>(
16234 EmitSpecialRegisterBuiltin(*this, E, Int32Ty, Int32Ty, NormalRead, RegName));
16235 CI->setConvergent();
16236 return CI;
16237 }
16238 case AMDGPU::BI__builtin_amdgcn_image_bvh_intersect_ray:
16239 case AMDGPU::BI__builtin_amdgcn_image_bvh_intersect_ray_h:
16240 case AMDGPU::BI__builtin_amdgcn_image_bvh_intersect_ray_l:
16241 case AMDGPU::BI__builtin_amdgcn_image_bvh_intersect_ray_lh: {
16242 llvm::Value *NodePtr = EmitScalarExpr(E->getArg(0));
16243 llvm::Value *RayExtent = EmitScalarExpr(E->getArg(1));
16244 llvm::Value *RayOrigin = EmitScalarExpr(E->getArg(2));
16245 llvm::Value *RayDir = EmitScalarExpr(E->getArg(3));
16246 llvm::Value *RayInverseDir = EmitScalarExpr(E->getArg(4));
16247 llvm::Value *TextureDescr = EmitScalarExpr(E->getArg(5));
16248
16249 Function *F = CGM.getIntrinsic(Intrinsic::amdgcn_image_bvh_intersect_ray,
16250 {NodePtr->getType(), RayDir->getType()});
16251 return Builder.CreateCall(F, {NodePtr, RayExtent, RayOrigin, RayDir,
16252 RayInverseDir, TextureDescr});
16253 }
16254
16255 // amdgcn workitem
16256 case AMDGPU::BI__builtin_amdgcn_workitem_id_x:
16257 return emitRangedBuiltin(*this, Intrinsic::amdgcn_workitem_id_x, 0, 1024);
16258 case AMDGPU::BI__builtin_amdgcn_workitem_id_y:
16259 return emitRangedBuiltin(*this, Intrinsic::amdgcn_workitem_id_y, 0, 1024);
16260 case AMDGPU::BI__builtin_amdgcn_workitem_id_z:
16261 return emitRangedBuiltin(*this, Intrinsic::amdgcn_workitem_id_z, 0, 1024);
16262
16263 // amdgcn workgroup size
16264 case AMDGPU::BI__builtin_amdgcn_workgroup_size_x:
16265 return EmitAMDGPUWorkGroupSize(*this, 0);
16266 case AMDGPU::BI__builtin_amdgcn_workgroup_size_y:
16267 return EmitAMDGPUWorkGroupSize(*this, 1);
16268 case AMDGPU::BI__builtin_amdgcn_workgroup_size_z:
16269 return EmitAMDGPUWorkGroupSize(*this, 2);
16270
16271 // amdgcn grid size
16272 case AMDGPU::BI__builtin_amdgcn_grid_size_x:
16273 return EmitAMDGPUGridSize(*this, 0);
16274 case AMDGPU::BI__builtin_amdgcn_grid_size_y:
16275 return EmitAMDGPUGridSize(*this, 1);
16276 case AMDGPU::BI__builtin_amdgcn_grid_size_z:
16277 return EmitAMDGPUGridSize(*this, 2);
16278
16279 // r600 intrinsics
16280 case AMDGPU::BI__builtin_r600_recipsqrt_ieee:
16281 case AMDGPU::BI__builtin_r600_recipsqrt_ieeef:
16282 return emitUnaryBuiltin(*this, E, Intrinsic::r600_recipsqrt_ieee);
16283 case AMDGPU::BI__builtin_r600_read_tidig_x:
16284 return emitRangedBuiltin(*this, Intrinsic::r600_read_tidig_x, 0, 1024);
16285 case AMDGPU::BI__builtin_r600_read_tidig_y:
16286 return emitRangedBuiltin(*this, Intrinsic::r600_read_tidig_y, 0, 1024);
16287 case AMDGPU::BI__builtin_r600_read_tidig_z:
16288 return emitRangedBuiltin(*this, Intrinsic::r600_read_tidig_z, 0, 1024);
16289 case AMDGPU::BI__builtin_amdgcn_alignbit: {
16290 llvm::Value *Src0 = EmitScalarExpr(E->getArg(0));
16291 llvm::Value *Src1 = EmitScalarExpr(E->getArg(1));
16292 llvm::Value *Src2 = EmitScalarExpr(E->getArg(2));
16293 Function *F = CGM.getIntrinsic(Intrinsic::fshr, Src0->getType());
16294 return Builder.CreateCall(F, { Src0, Src1, Src2 });
16295 }
16296
16297 case AMDGPU::BI__builtin_amdgcn_fence: {
16298 if (ProcessOrderScopeAMDGCN(EmitScalarExpr(E->getArg(0)),
16299 EmitScalarExpr(E->getArg(1)), AO, SSID))
16300 return Builder.CreateFence(AO, SSID);
16301 LLVM_FALLTHROUGH[[gnu::fallthrough]];
16302 }
16303 case AMDGPU::BI__builtin_amdgcn_atomic_inc32:
16304 case AMDGPU::BI__builtin_amdgcn_atomic_inc64:
16305 case AMDGPU::BI__builtin_amdgcn_atomic_dec32:
16306 case AMDGPU::BI__builtin_amdgcn_atomic_dec64: {
16307 unsigned BuiltinAtomicOp;
16308 llvm::Type *ResultType = ConvertType(E->getType());
16309
16310 switch (BuiltinID) {
16311 case AMDGPU::BI__builtin_amdgcn_atomic_inc32:
16312 case AMDGPU::BI__builtin_amdgcn_atomic_inc64:
16313 BuiltinAtomicOp = Intrinsic::amdgcn_atomic_inc;
16314 break;
16315 case AMDGPU::BI__builtin_amdgcn_atomic_dec32:
16316 case AMDGPU::BI__builtin_amdgcn_atomic_dec64:
16317 BuiltinAtomicOp = Intrinsic::amdgcn_atomic_dec;
16318 break;
16319 }
16320
16321 Value *Ptr = EmitScalarExpr(E->getArg(0));
16322 Value *Val = EmitScalarExpr(E->getArg(1));
16323
16324 llvm::Function *F =
16325 CGM.getIntrinsic(BuiltinAtomicOp, {ResultType, Ptr->getType()});
16326
16327 if (ProcessOrderScopeAMDGCN(EmitScalarExpr(E->getArg(2)),
16328 EmitScalarExpr(E->getArg(3)), AO, SSID)) {
16329
16330 // llvm.amdgcn.atomic.inc and llvm.amdgcn.atomic.dec expects ordering and
16331 // scope as unsigned values
16332 Value *MemOrder = Builder.getInt32(static_cast<int>(AO));
16333 Value *MemScope = Builder.getInt32(static_cast<int>(SSID));
16334
16335 QualType PtrTy = E->getArg(0)->IgnoreImpCasts()->getType();
16336 bool Volatile =
16337 PtrTy->castAs<PointerType>()->getPointeeType().isVolatileQualified();
16338 Value *IsVolatile = Builder.getInt1(static_cast<bool>(Volatile));
16339
16340 return Builder.CreateCall(F, {Ptr, Val, MemOrder, MemScope, IsVolatile});
16341 }
16342 LLVM_FALLTHROUGH[[gnu::fallthrough]];
16343 }
16344 default:
16345 return nullptr;
16346 }
16347}
16348
16349/// Handle a SystemZ function in which the final argument is a pointer
16350/// to an int that receives the post-instruction CC value. At the LLVM level
16351/// this is represented as a function that returns a {result, cc} pair.
16352static Value *EmitSystemZIntrinsicWithCC(CodeGenFunction &CGF,
16353 unsigned IntrinsicID,
16354 const CallExpr *E) {
16355 unsigned NumArgs = E->getNumArgs() - 1;
16356 SmallVector<Value *, 8> Args(NumArgs);
16357 for (unsigned I = 0; I < NumArgs; ++I)
16358 Args[I] = CGF.EmitScalarExpr(E->getArg(I));
16359 Address CCPtr = CGF.EmitPointerWithAlignment(E->getArg(NumArgs));
16360 Function *F = CGF.CGM.getIntrinsic(IntrinsicID);
16361 Value *Call = CGF.Builder.CreateCall(F, Args);
16362 Value *CC = CGF.Builder.CreateExtractValue(Call, 1);
16363 CGF.Builder.CreateStore(CC, CCPtr);
16364 return CGF.Builder.CreateExtractValue(Call, 0);
16365}
16366
16367Value *CodeGenFunction::EmitSystemZBuiltinExpr(unsigned BuiltinID,
16368 const CallExpr *E) {
16369 switch (BuiltinID) {
16370 case SystemZ::BI__builtin_tbegin: {
16371 Value *TDB = EmitScalarExpr(E->getArg(0));
16372 Value *Control = llvm::ConstantInt::get(Int32Ty, 0xff0c);
16373 Function *F = CGM.getIntrinsic(Intrinsic::s390_tbegin);
16374 return Builder.CreateCall(F, {TDB, Control});
16375 }
16376 case SystemZ::BI__builtin_tbegin_nofloat: {
16377 Value *TDB = EmitScalarExpr(E->getArg(0));
16378 Value *Control = llvm::ConstantInt::get(Int32Ty, 0xff0c);
16379 Function *F = CGM.getIntrinsic(Intrinsic::s390_tbegin_nofloat);
16380 return Builder.CreateCall(F, {TDB, Control});
16381 }
16382 case SystemZ::BI__builtin_tbeginc: {
16383 Value *TDB = llvm::ConstantPointerNull::get(Int8PtrTy);
16384 Value *Control = llvm::ConstantInt::get(Int32Ty, 0xff08);
16385 Function *F = CGM.getIntrinsic(Intrinsic::s390_tbeginc);
16386 return Builder.CreateCall(F, {TDB, Control});
16387 }
16388 case SystemZ::BI__builtin_tabort: {
16389 Value *Data = EmitScalarExpr(E->getArg(0));
16390 Function *F = CGM.getIntrinsic(Intrinsic::s390_tabort);
16391 return Builder.CreateCall(F, Builder.CreateSExt(Data, Int64Ty, "tabort"));
16392 }
16393 case SystemZ::BI__builtin_non_tx_store: {
16394 Value *Address = EmitScalarExpr(E->getArg(0));
16395 Value *Data = EmitScalarExpr(E->getArg(1));
16396 Function *F = CGM.getIntrinsic(Intrinsic::s390_ntstg);
16397 return Builder.CreateCall(F, {Data, Address});
16398 }
16399
16400 // Vector builtins. Note that most vector builtins are mapped automatically
16401 // to target-specific LLVM intrinsics. The ones handled specially here can
16402 // be represented via standard LLVM IR, which is preferable to enable common
16403 // LLVM optimizations.
16404
16405 case SystemZ::BI__builtin_s390_vpopctb:
16406 case SystemZ::BI__builtin_s390_vpopcth:
16407 case SystemZ::BI__builtin_s390_vpopctf:
16408 case SystemZ::BI__builtin_s390_vpopctg: {
16409 llvm::Type *ResultType = ConvertType(E->getType());
16410 Value *X = EmitScalarExpr(E->getArg(0));
16411 Function *F = CGM.getIntrinsic(Intrinsic::ctpop, ResultType);
16412 return Builder.CreateCall(F, X);
16413 }
16414
16415 case SystemZ::BI__builtin_s390_vclzb:
16416 case SystemZ::BI__builtin_s390_vclzh:
16417 case SystemZ::BI__builtin_s390_vclzf:
16418 case SystemZ::BI__builtin_s390_vclzg: {
16419 llvm::Type *ResultType = ConvertType(E->getType());
16420 Value *X = EmitScalarExpr(E->getArg(0));
16421 Value *Undef = ConstantInt::get(Builder.getInt1Ty(), false);
16422 Function *F = CGM.getIntrinsic(Intrinsic::ctlz, ResultType);
16423 return Builder.CreateCall(F, {X, Undef});
16424 }
16425
16426 case SystemZ::BI__builtin_s390_vctzb:
16427 case SystemZ::BI__builtin_s390_vctzh:
16428 case SystemZ::BI__builtin_s390_vctzf:
16429 case SystemZ::BI__builtin_s390_vctzg: {
16430 llvm::Type *ResultType = ConvertType(E->getType());
16431 Value *X = EmitScalarExpr(E->getArg(0));
16432 Value *Undef = ConstantInt::get(Builder.getInt1Ty(), false);
16433 Function *F = CGM.getIntrinsic(Intrinsic::cttz, ResultType);
16434 return Builder.CreateCall(F, {X, Undef});
16435 }
16436
16437 case SystemZ::BI__builtin_s390_vfsqsb:
16438 case SystemZ::BI__builtin_s390_vfsqdb: {
16439 llvm::Type *ResultType = ConvertType(E->getType());
16440 Value *X = EmitScalarExpr(E->getArg(0));
16441 if (Builder.getIsFPConstrained()) {
16442 Function *F = CGM.getIntrinsic(Intrinsic::experimental_constrained_sqrt, ResultType);
16443 return Builder.CreateConstrainedFPCall(F, { X });
16444 } else {
16445 Function *F = CGM.getIntrinsic(Intrinsic::sqrt, ResultType);
16446 return Builder.CreateCall(F, X);
16447 }
16448 }
16449 case SystemZ::BI__builtin_s390_vfmasb:
16450 case SystemZ::BI__builtin_s390_vfmadb: {
16451 llvm::Type *ResultType = ConvertType(E->getType());
16452 Value *X = EmitScalarExpr(E->getArg(0));
16453 Value *Y = EmitScalarExpr(E->getArg(1));
16454 Value *Z = EmitScalarExpr(E->getArg(2));
16455 if (Builder.getIsFPConstrained()) {
16456 Function *F = CGM.getIntrinsic(Intrinsic::experimental_constrained_fma, ResultType);
16457 return Builder.CreateConstrainedFPCall(F, {X, Y, Z});
16458 } else {
16459 Function *F = CGM.getIntrinsic(Intrinsic::fma, ResultType);
16460 return Builder.CreateCall(F, {X, Y, Z});
16461 }
16462 }
16463 case SystemZ::BI__builtin_s390_vfmssb:
16464 case SystemZ::BI__builtin_s390_vfmsdb: {
16465 llvm::Type *ResultType = ConvertType(E->getType());
16466 Value *X = EmitScalarExpr(E->getArg(0));
16467 Value *Y = EmitScalarExpr(E->getArg(1));
16468 Value *Z = EmitScalarExpr(E->getArg(2));
16469 if (Builder.getIsFPConstrained()) {
16470 Function *F = CGM.getIntrinsic(Intrinsic::experimental_constrained_fma, ResultType);
16471 return Builder.CreateConstrainedFPCall(F, {X, Y, Builder.CreateFNeg(Z, "neg")});
16472 } else {
16473 Function *F = CGM.getIntrinsic(Intrinsic::fma, ResultType);
16474 return Builder.CreateCall(F, {X, Y, Builder.CreateFNeg(Z, "neg")});
16475 }
16476 }
16477 case SystemZ::BI__builtin_s390_vfnmasb:
16478 case SystemZ::BI__builtin_s390_vfnmadb: {
16479 llvm::Type *ResultType = ConvertType(E->getType());
16480 Value *X = EmitScalarExpr(E->getArg(0));
16481 Value *Y = EmitScalarExpr(E->getArg(1));
16482 Value *Z = EmitScalarExpr(E->getArg(2));
16483 if (Builder.getIsFPConstrained()) {
16484 Function *F = CGM.getIntrinsic(Intrinsic::experimental_constrained_fma, ResultType);
16485 return Builder.CreateFNeg(Builder.CreateConstrainedFPCall(F, {X, Y, Z}), "neg");
16486 } else {
16487 Function *F = CGM.getIntrinsic(Intrinsic::fma, ResultType);
16488 return Builder.CreateFNeg(Builder.CreateCall(F, {X, Y, Z}), "neg");
16489 }
16490 }
16491 case SystemZ::BI__builtin_s390_vfnmssb:
16492 case SystemZ::BI__builtin_s390_vfnmsdb: {
16493 llvm::Type *ResultType = ConvertType(E->getType());
16494 Value *X = EmitScalarExpr(E->getArg(0));
16495 Value *Y = EmitScalarExpr(E->getArg(1));
16496 Value *Z = EmitScalarExpr(E->getArg(2));
16497 if (Builder.getIsFPConstrained()) {
16498 Function *F = CGM.getIntrinsic(Intrinsic::experimental_constrained_fma, ResultType);
16499 Value *NegZ = Builder.CreateFNeg(Z, "sub");
16500 return Builder.CreateFNeg(Builder.CreateConstrainedFPCall(F, {X, Y, NegZ}));
16501 } else {
16502 Function *F = CGM.getIntrinsic(Intrinsic::fma, ResultType);
16503 Value *NegZ = Builder.CreateFNeg(Z, "neg");
16504 return Builder.CreateFNeg(Builder.CreateCall(F, {X, Y, NegZ}));
16505 }
16506 }
16507 case SystemZ::BI__builtin_s390_vflpsb:
16508 case SystemZ::BI__builtin_s390_vflpdb: {
16509 llvm::Type *ResultType = ConvertType(E->getType());
16510 Value *X = EmitScalarExpr(E->getArg(0));
16511 Function *F = CGM.getIntrinsic(Intrinsic::fabs, ResultType);
16512 return Builder.CreateCall(F, X);
16513 }
16514 case SystemZ::BI__builtin_s390_vflnsb:
16515 case SystemZ::BI__builtin_s390_vflndb: {
16516 llvm::Type *ResultType = ConvertType(E->getType());
16517 Value *X = EmitScalarExpr(E->getArg(0));
16518 Function *F = CGM.getIntrinsic(Intrinsic::fabs, ResultType);
16519 return Builder.CreateFNeg(Builder.CreateCall(F, X), "neg");
16520 }
16521 case SystemZ::BI__builtin_s390_vfisb:
16522 case SystemZ::BI__builtin_s390_vfidb: {
16523 llvm::Type *ResultType = ConvertType(E->getType());
16524 Value *X = EmitScalarExpr(E->getArg(0));
16525 // Constant-fold the M4 and M5 mask arguments.
16526 llvm::APSInt M4 = *E->getArg(1)->getIntegerConstantExpr(getContext());
16527 llvm::APSInt M5 = *E->getArg(2)->getIntegerConstantExpr(getContext());
16528 // Check whether this instance can be represented via a LLVM standard
16529 // intrinsic. We only support some combinations of M4 and M5.
16530 Intrinsic::ID ID = Intrinsic::not_intrinsic;
16531 Intrinsic::ID CI;
16532 switch (M4.getZExtValue()) {
16533 default: break;
16534 case 0: // IEEE-inexact exception allowed
16535 switch (M5.getZExtValue()) {
16536 default: break;
16537 case 0: ID = Intrinsic::rint;
16538 CI = Intrinsic::experimental_constrained_rint; break;
16539 }
16540 break;
16541 case 4: // IEEE-inexact exception suppressed
16542 switch (M5.getZExtValue()) {
16543 default: break;
16544 case 0: ID = Intrinsic::nearbyint;
16545 CI = Intrinsic::experimental_constrained_nearbyint; break;
16546 case 1: ID = Intrinsic::round;
16547 CI = Intrinsic::experimental_constrained_round; break;
16548 case 5: ID = Intrinsic::trunc;
16549 CI = Intrinsic::experimental_constrained_trunc; break;
16550 case 6: ID = Intrinsic::ceil;
16551 CI = Intrinsic::experimental_constrained_ceil; break;
16552 case 7: ID = Intrinsic::floor;
16553 CI = Intrinsic::experimental_constrained_floor; break;
16554 }
16555 break;
16556 }
16557 if (ID != Intrinsic::not_intrinsic) {
16558 if (Builder.getIsFPConstrained()) {
16559 Function *F = CGM.getIntrinsic(CI, ResultType);
16560 return Builder.CreateConstrainedFPCall(F, X);
16561 } else {
16562 Function *F = CGM.getIntrinsic(ID, ResultType);
16563 return Builder.CreateCall(F, X);
16564 }
16565 }
16566 switch (BuiltinID) { // FIXME: constrained version?
16567 case SystemZ::BI__builtin_s390_vfisb: ID = Intrinsic::s390_vfisb; break;
16568 case SystemZ::BI__builtin_s390_vfidb: ID = Intrinsic::s390_vfidb; break;
16569 default: llvm_unreachable("Unknown BuiltinID")__builtin_unreachable();
16570 }
16571 Function *F = CGM.getIntrinsic(ID);
16572 Value *M4Value = llvm::ConstantInt::get(getLLVMContext(), M4);
16573 Value *M5Value = llvm::ConstantInt::get(getLLVMContext(), M5);
16574 return Builder.CreateCall(F, {X, M4Value, M5Value});
16575 }
16576 case SystemZ::BI__builtin_s390_vfmaxsb:
16577 case SystemZ::BI__builtin_s390_vfmaxdb: {
16578 llvm::Type *ResultType = ConvertType(E->getType());
16579 Value *X = EmitScalarExpr(E->getArg(0));
16580 Value *Y = EmitScalarExpr(E->getArg(1));
16581 // Constant-fold the M4 mask argument.
16582 llvm::APSInt M4 = *E->getArg(2)->getIntegerConstantExpr(getContext());
16583 // Check whether this instance can be represented via a LLVM standard
16584 // intrinsic. We only support some values of M4.
16585 Intrinsic::ID ID = Intrinsic::not_intrinsic;
16586 Intrinsic::ID CI;
16587 switch (M4.getZExtValue()) {
16588 default: break;
16589 case 4: ID = Intrinsic::maxnum;
16590 CI = Intrinsic::experimental_constrained_maxnum; break;
16591 }
16592 if (ID != Intrinsic::not_intrinsic) {
16593 if (Builder.getIsFPConstrained()) {
16594 Function *F = CGM.getIntrinsic(CI, ResultType);
16595 return Builder.CreateConstrainedFPCall(F, {X, Y});
16596 } else {
16597 Function *F = CGM.getIntrinsic(ID, ResultType);
16598 return Builder.CreateCall(F, {X, Y});
16599 }
16600 }
16601 switch (BuiltinID) {
16602 case SystemZ::BI__builtin_s390_vfmaxsb: ID = Intrinsic::s390_vfmaxsb; break;
16603 case SystemZ::BI__builtin_s390_vfmaxdb: ID = Intrinsic::s390_vfmaxdb; break;
16604 default: llvm_unreachable("Unknown BuiltinID")__builtin_unreachable();
16605 }
16606 Function *F = CGM.getIntrinsic(ID);
16607 Value *M4Value = llvm::ConstantInt::get(getLLVMContext(), M4);
16608 return Builder.CreateCall(F, {X, Y, M4Value});
16609 }
16610 case SystemZ::BI__builtin_s390_vfminsb:
16611 case SystemZ::BI__builtin_s390_vfmindb: {
16612 llvm::Type *ResultType = ConvertType(E->getType());
16613 Value *X = EmitScalarExpr(E->getArg(0));
16614 Value *Y = EmitScalarExpr(E->getArg(1));
16615 // Constant-fold the M4 mask argument.
16616 llvm::APSInt M4 = *E->getArg(2)->getIntegerConstantExpr(getContext());
16617 // Check whether this instance can be represented via a LLVM standard
16618 // intrinsic. We only support some values of M4.
16619 Intrinsic::ID ID = Intrinsic::not_intrinsic;
16620 Intrinsic::ID CI;
16621 switch (M4.getZExtValue()) {
16622 default: break;
16623 case 4: ID = Intrinsic::minnum;
16624 CI = Intrinsic::experimental_constrained_minnum; break;
16625 }
16626 if (ID != Intrinsic::not_intrinsic) {
16627 if (Builder.getIsFPConstrained()) {
16628 Function *F = CGM.getIntrinsic(CI, ResultType);
16629 return Builder.CreateConstrainedFPCall(F, {X, Y});
16630 } else {
16631 Function *F = CGM.getIntrinsic(ID, ResultType);
16632 return Builder.CreateCall(F, {X, Y});
16633 }
16634 }
16635 switch (BuiltinID) {
16636 case SystemZ::BI__builtin_s390_vfminsb: ID = Intrinsic::s390_vfminsb; break;
16637 case SystemZ::BI__builtin_s390_vfmindb: ID = Intrinsic::s390_vfmindb; break;
16638 default: llvm_unreachable("Unknown BuiltinID")__builtin_unreachable();
16639 }
16640 Function *F = CGM.getIntrinsic(ID);
16641 Value *M4Value = llvm::ConstantInt::get(getLLVMContext(), M4);
16642 return Builder.CreateCall(F, {X, Y, M4Value});
16643 }
16644
16645 case SystemZ::BI__builtin_s390_vlbrh:
16646 case SystemZ::BI__builtin_s390_vlbrf:
16647 case SystemZ::BI__builtin_s390_vlbrg: {
16648 llvm::Type *ResultType = ConvertType(E->getType());
16649 Value *X = EmitScalarExpr(E->getArg(0));
16650 Function *F = CGM.getIntrinsic(Intrinsic::bswap, ResultType);
16651 return Builder.CreateCall(F, X);
16652 }
16653
16654 // Vector intrinsics that output the post-instruction CC value.
16655
16656#define INTRINSIC_WITH_CC(NAME) \
16657 case SystemZ::BI__builtin_##NAME: \
16658 return EmitSystemZIntrinsicWithCC(*this, Intrinsic::NAME, E)
16659
16660 INTRINSIC_WITH_CC(s390_vpkshs);
16661 INTRINSIC_WITH_CC(s390_vpksfs);
16662 INTRINSIC_WITH_CC(s390_vpksgs);
16663
16664 INTRINSIC_WITH_CC(s390_vpklshs);
16665 INTRINSIC_WITH_CC(s390_vpklsfs);
16666 INTRINSIC_WITH_CC(s390_vpklsgs);
16667
16668 INTRINSIC_WITH_CC(s390_vceqbs);
16669 INTRINSIC_WITH_CC(s390_vceqhs);
16670 INTRINSIC_WITH_CC(s390_vceqfs);
16671 INTRINSIC_WITH_CC(s390_vceqgs);
16672
16673 INTRINSIC_WITH_CC(s390_vchbs);
16674 INTRINSIC_WITH_CC(s390_vchhs);
16675 INTRINSIC_WITH_CC(s390_vchfs);
16676 INTRINSIC_WITH_CC(s390_vchgs);
16677
16678 INTRINSIC_WITH_CC(s390_vchlbs);
16679 INTRINSIC_WITH_CC(s390_vchlhs);
16680 INTRINSIC_WITH_CC(s390_vchlfs);
16681 INTRINSIC_WITH_CC(s390_vchlgs);
16682
16683 INTRINSIC_WITH_CC(s390_vfaebs);
16684 INTRINSIC_WITH_CC(s390_vfaehs);
16685 INTRINSIC_WITH_CC(s390_vfaefs);
16686
16687 INTRINSIC_WITH_CC(s390_vfaezbs);
16688 INTRINSIC_WITH_CC(s390_vfaezhs);
16689 INTRINSIC_WITH_CC(s390_vfaezfs);
16690
16691 INTRINSIC_WITH_CC(s390_vfeebs);
16692 INTRINSIC_WITH_CC(s390_vfeehs);
16693 INTRINSIC_WITH_CC(s390_vfeefs);
16694
16695 INTRINSIC_WITH_CC(s390_vfeezbs);
16696 INTRINSIC_WITH_CC(s390_vfeezhs);
16697 INTRINSIC_WITH_CC(s390_vfeezfs);
16698
16699 INTRINSIC_WITH_CC(s390_vfenebs);
16700 INTRINSIC_WITH_CC(s390_vfenehs);
16701 INTRINSIC_WITH_CC(s390_vfenefs);
16702
16703 INTRINSIC_WITH_CC(s390_vfenezbs);
16704 INTRINSIC_WITH_CC(s390_vfenezhs);
16705 INTRINSIC_WITH_CC(s390_vfenezfs);
16706
16707 INTRINSIC_WITH_CC(s390_vistrbs);
16708 INTRINSIC_WITH_CC(s390_vistrhs);
16709 INTRINSIC_WITH_CC(s390_vistrfs);
16710
16711 INTRINSIC_WITH_CC(s390_vstrcbs);
16712 INTRINSIC_WITH_CC(s390_vstrchs);
16713 INTRINSIC_WITH_CC(s390_vstrcfs);
16714
16715 INTRINSIC_WITH_CC(s390_vstrczbs);
16716 INTRINSIC_WITH_CC(s390_vstrczhs);
16717 INTRINSIC_WITH_CC(s390_vstrczfs);
16718
16719 INTRINSIC_WITH_CC(s390_vfcesbs);
16720 INTRINSIC_WITH_CC(s390_vfcedbs);
16721 INTRINSIC_WITH_CC(s390_vfchsbs);
16722 INTRINSIC_WITH_CC(s390_vfchdbs);
16723 INTRINSIC_WITH_CC(s390_vfchesbs);
16724 INTRINSIC_WITH_CC(s390_vfchedbs);
16725
16726 INTRINSIC_WITH_CC(s390_vftcisb);
16727 INTRINSIC_WITH_CC(s390_vftcidb);
16728
16729 INTRINSIC_WITH_CC(s390_vstrsb);
16730 INTRINSIC_WITH_CC(s390_vstrsh);
16731 INTRINSIC_WITH_CC(s390_vstrsf);
16732
16733 INTRINSIC_WITH_CC(s390_vstrszb);
16734 INTRINSIC_WITH_CC(s390_vstrszh);
16735 INTRINSIC_WITH_CC(s390_vstrszf);
16736
16737#undef INTRINSIC_WITH_CC
16738
16739 default:
16740 return nullptr;
16741 }
16742}
16743
16744namespace {
16745// Helper classes for mapping MMA builtins to particular LLVM intrinsic variant.
16746struct NVPTXMmaLdstInfo {
16747 unsigned NumResults; // Number of elements to load/store
16748 // Intrinsic IDs for row/col variants. 0 if particular layout is unsupported.
16749 unsigned IID_col;
16750 unsigned IID_row;
16751};
16752
16753#define MMA_INTR(geom_op_type, layout) \
16754 Intrinsic::nvvm_wmma_##geom_op_type##_##layout##_stride
16755#define MMA_LDST(n, geom_op_type) \
16756 { n, MMA_INTR(geom_op_type, col), MMA_INTR(geom_op_type, row) }
16757
16758static NVPTXMmaLdstInfo getNVPTXMmaLdstInfo(unsigned BuiltinID) {
16759 switch (BuiltinID) {
16760 // FP MMA loads
16761 case NVPTX::BI__hmma_m16n16k16_ld_a:
16762 return MMA_LDST(8, m16n16k16_load_a_f16);
16763 case NVPTX::BI__hmma_m16n16k16_ld_b:
16764 return MMA_LDST(8, m16n16k16_load_b_f16);
16765 case NVPTX::BI__hmma_m16n16k16_ld_c_f16:
16766 return MMA_LDST(4, m16n16k16_load_c_f16);
16767 case NVPTX::BI__hmma_m16n16k16_ld_c_f32:
16768 return MMA_LDST(8, m16n16k16_load_c_f32);
16769 case NVPTX::BI__hmma_m32n8k16_ld_a:
16770 return MMA_LDST(8, m32n8k16_load_a_f16);
16771 case NVPTX::BI__hmma_m32n8k16_ld_b:
16772 return MMA_LDST(8, m32n8k16_load_b_f16);
16773 case NVPTX::BI__hmma_m32n8k16_ld_c_f16:
16774 return MMA_LDST(4, m32n8k16_load_c_f16);
16775 case NVPTX::BI__hmma_m32n8k16_ld_c_f32:
16776 return MMA_LDST(8, m32n8k16_load_c_f32);
16777 case NVPTX::BI__hmma_m8n32k16_ld_a:
16778 return MMA_LDST(8, m8n32k16_load_a_f16);
16779 case NVPTX::BI__hmma_m8n32k16_ld_b:
16780 return MMA_LDST(8, m8n32k16_load_b_f16);
16781 case NVPTX::BI__hmma_m8n32k16_ld_c_f16:
16782 return MMA_LDST(4, m8n32k16_load_c_f16);
16783 case NVPTX::BI__hmma_m8n32k16_ld_c_f32:
16784 return MMA_LDST(8, m8n32k16_load_c_f32);
16785
16786 // Integer MMA loads
16787 case NVPTX::BI__imma_m16n16k16_ld_a_s8:
16788 return MMA_LDST(2, m16n16k16_load_a_s8);
16789 case NVPTX::BI__imma_m16n16k16_ld_a_u8:
16790 return MMA_LDST(2, m16n16k16_load_a_u8);
16791 case NVPTX::BI__imma_m16n16k16_ld_b_s8:
16792 return MMA_LDST(2, m16n16k16_load_b_s8);
16793 case NVPTX::BI__imma_m16n16k16_ld_b_u8:
16794 return MMA_LDST(2, m16n16k16_load_b_u8);
16795 case NVPTX::BI__imma_m16n16k16_ld_c:
16796 return MMA_LDST(8, m16n16k16_load_c_s32);
16797 case NVPTX::BI__imma_m32n8k16_ld_a_s8:
16798 return MMA_LDST(4, m32n8k16_load_a_s8);
16799 case NVPTX::BI__imma_m32n8k16_ld_a_u8:
16800 return MMA_LDST(4, m32n8k16_load_a_u8);
16801 case NVPTX::BI__imma_m32n8k16_ld_b_s8:
16802 return MMA_LDST(1, m32n8k16_load_b_s8);
16803 case NVPTX::BI__imma_m32n8k16_ld_b_u8:
16804 return MMA_LDST(1, m32n8k16_load_b_u8);
16805 case NVPTX::BI__imma_m32n8k16_ld_c:
16806 return MMA_LDST(8, m32n8k16_load_c_s32);
16807 case NVPTX::BI__imma_m8n32k16_ld_a_s8:
16808 return MMA_LDST(1, m8n32k16_load_a_s8);
16809 case NVPTX::BI__imma_m8n32k16_ld_a_u8:
16810 return MMA_LDST(1, m8n32k16_load_a_u8);
16811 case NVPTX::BI__imma_m8n32k16_ld_b_s8:
16812 return MMA_LDST(4, m8n32k16_load_b_s8);
16813 case NVPTX::BI__imma_m8n32k16_ld_b_u8:
16814 return MMA_LDST(4, m8n32k16_load_b_u8);
16815 case NVPTX::BI__imma_m8n32k16_ld_c:
16816 return MMA_LDST(8, m8n32k16_load_c_s32);
16817
16818 // Sub-integer MMA loads.
16819 // Only row/col layout is supported by A/B fragments.
16820 case NVPTX::BI__imma_m8n8k32_ld_a_s4:
16821 return {1, 0, MMA_INTR(m8n8k32_load_a_s4, row)};
16822 case NVPTX::BI__imma_m8n8k32_ld_a_u4:
16823 return {1, 0, MMA_INTR(m8n8k32_load_a_u4, row)};
16824 case NVPTX::BI__imma_m8n8k32_ld_b_s4:
16825 return {1, MMA_INTR(m8n8k32_load_b_s4, col), 0};
16826 case NVPTX::BI__imma_m8n8k32_ld_b_u4:
16827 return {1, MMA_INTR(m8n8k32_load_b_u4, col), 0};
16828 case NVPTX::BI__imma_m8n8k32_ld_c:
16829 return MMA_LDST(2, m8n8k32_load_c_s32);
16830 case NVPTX::BI__bmma_m8n8k128_ld_a_b1:
16831 return {1, 0, MMA_INTR(m8n8k128_load_a_b1, row)};
16832 case NVPTX::BI__bmma_m8n8k128_ld_b_b1:
16833 return {1, MMA_INTR(m8n8k128_load_b_b1, col), 0};
16834 case NVPTX::BI__bmma_m8n8k128_ld_c:
16835 return MMA_LDST(2, m8n8k128_load_c_s32);
16836
16837 // Double MMA loads
16838 case NVPTX::BI__dmma_m8n8k4_ld_a:
16839 return MMA_LDST(1, m8n8k4_load_a_f64);
16840 case NVPTX::BI__dmma_m8n8k4_ld_b:
16841 return MMA_LDST(1, m8n8k4_load_b_f64);
16842 case NVPTX::BI__dmma_m8n8k4_ld_c:
16843 return MMA_LDST(2, m8n8k4_load_c_f64);
16844
16845 // Alternate float MMA loads
16846 case NVPTX::BI__mma_bf16_m16n16k16_ld_a:
16847 return MMA_LDST(4, m16n16k16_load_a_bf16);
16848 case NVPTX::BI__mma_bf16_m16n16k16_ld_b:
16849 return MMA_LDST(4, m16n16k16_load_b_bf16);
16850 case NVPTX::BI__mma_bf16_m8n32k16_ld_a:
16851 return MMA_LDST(2, m8n32k16_load_a_bf16);
16852 case NVPTX::BI__mma_bf16_m8n32k16_ld_b:
16853 return MMA_LDST(8, m8n32k16_load_b_bf16);
16854 case NVPTX::BI__mma_bf16_m32n8k16_ld_a:
16855 return MMA_LDST(8, m32n8k16_load_a_bf16);
16856 case NVPTX::BI__mma_bf16_m32n8k16_ld_b:
16857 return MMA_LDST(2, m32n8k16_load_b_bf16);
16858 case NVPTX::BI__mma_tf32_m16n16k8_ld_a:
16859 return MMA_LDST(4, m16n16k8_load_a_tf32);
16860 case NVPTX::BI__mma_tf32_m16n16k8_ld_b:
16861 return MMA_LDST(2, m16n16k8_load_b_tf32);
16862 case NVPTX::BI__mma_tf32_m16n16k8_ld_c:
16863 return MMA_LDST(8, m16n16k8_load_c_f32);
16864
16865 // NOTE: We need to follow inconsitent naming scheme used by NVCC. Unlike
16866 // PTX and LLVM IR where stores always use fragment D, NVCC builtins always
16867 // use fragment C for both loads and stores.
16868 // FP MMA stores.
16869 case NVPTX::BI__hmma_m16n16k16_st_c_f16:
16870 return MMA_LDST(4, m16n16k16_store_d_f16);
16871 case NVPTX::BI__hmma_m16n16k16_st_c_f32:
16872 return MMA_LDST(8, m16n16k16_store_d_f32);
16873 case NVPTX::BI__hmma_m32n8k16_st_c_f16:
16874 return MMA_LDST(4, m32n8k16_store_d_f16);
16875 case NVPTX::BI__hmma_m32n8k16_st_c_f32:
16876 return MMA_LDST(8, m32n8k16_store_d_f32);
16877 case NVPTX::BI__hmma_m8n32k16_st_c_f16:
16878 return MMA_LDST(4, m8n32k16_store_d_f16);
16879 case NVPTX::BI__hmma_m8n32k16_st_c_f32:
16880 return MMA_LDST(8, m8n32k16_store_d_f32);
16881
16882 // Integer and sub-integer MMA stores.
16883 // Another naming quirk. Unlike other MMA builtins that use PTX types in the
16884 // name, integer loads/stores use LLVM's i32.
16885 case NVPTX::BI__imma_m16n16k16_st_c_i32:
16886 return MMA_LDST(8, m16n16k16_store_d_s32);
16887 case NVPTX::BI__imma_m32n8k16_st_c_i32:
16888 return MMA_LDST(8, m32n8k16_store_d_s32);
16889 case NVPTX::BI__imma_m8n32k16_st_c_i32:
16890 return MMA_LDST(8, m8n32k16_store_d_s32);
16891 case NVPTX::BI__imma_m8n8k32_st_c_i32:
16892 return MMA_LDST(2, m8n8k32_store_d_s32);
16893 case NVPTX::BI__bmma_m8n8k128_st_c_i32:
16894 return MMA_LDST(2, m8n8k128_store_d_s32);
16895
16896 // Double MMA store
16897 case NVPTX::BI__dmma_m8n8k4_st_c_f64:
16898 return MMA_LDST(2, m8n8k4_store_d_f64);
16899
16900 // Alternate float MMA store
16901 case NVPTX::BI__mma_m16n16k8_st_c_f32:
16902 return MMA_LDST(8, m16n16k8_store_d_f32);
16903
16904 default:
16905 llvm_unreachable("Unknown MMA builtin")__builtin_unreachable();
16906 }
16907}
16908#undef MMA_LDST
16909#undef MMA_INTR
16910
16911
16912struct NVPTXMmaInfo {
16913 unsigned NumEltsA;
16914 unsigned NumEltsB;
16915 unsigned NumEltsC;
16916 unsigned NumEltsD;
16917
16918 // Variants are ordered by layout-A/layout-B/satf, where 'row' has priority
16919 // over 'col' for layout. The index of non-satf variants is expected to match
16920 // the undocumented layout constants used by CUDA's mma.hpp.
16921 std::array<unsigned, 8> Variants;
16922
16923 unsigned getMMAIntrinsic(int Layout, bool Satf) {
16924 unsigned Index = Layout + 4 * Satf;
16925 if (Index >= Variants.size())
16926 return 0;
16927 return Variants[Index];
16928 }
16929};
16930
16931 // Returns an intrinsic that matches Layout and Satf for valid combinations of
16932 // Layout and Satf, 0 otherwise.
16933static NVPTXMmaInfo getNVPTXMmaInfo(unsigned BuiltinID) {
16934 // clang-format off
16935#define MMA_VARIANTS(geom, type) \
16936 Intrinsic::nvvm_wmma_##geom##_mma_row_row_##type, \
16937 Intrinsic::nvvm_wmma_##geom##_mma_row_col_##type, \
16938 Intrinsic::nvvm_wmma_##geom##_mma_col_row_##type, \
16939 Intrinsic::nvvm_wmma_##geom##_mma_col_col_##type
16940#define MMA_SATF_VARIANTS(geom, type) \
16941 MMA_VARIANTS(geom, type), \
16942 Intrinsic::nvvm_wmma_##geom##_mma_row_row_##type##_satfinite, \
16943 Intrinsic::nvvm_wmma_##geom##_mma_row_col_##type##_satfinite, \
16944 Intrinsic::nvvm_wmma_##geom##_mma_col_row_##type##_satfinite, \
16945 Intrinsic::nvvm_wmma_##geom##_mma_col_col_##type##_satfinite
16946// Sub-integer MMA only supports row.col layout.
16947#define MMA_VARIANTS_I4(geom, type) \
16948 0, \
16949 Intrinsic::nvvm_wmma_##geom##_mma_row_col_##type, \
16950 0, \
16951 0, \
16952 0, \
16953 Intrinsic::nvvm_wmma_##geom##_mma_row_col_##type##_satfinite, \
16954 0, \
16955 0
16956// b1 MMA does not support .satfinite.
16957#define MMA_VARIANTS_B1_XOR(geom, type) \
16958 0, \
16959 Intrinsic::nvvm_wmma_##geom##_mma_xor_popc_row_col_##type, \
16960 0, \
16961 0, \
16962 0, \
16963 0, \
16964 0, \
16965 0
16966#define MMA_VARIANTS_B1_AND(geom, type) \
16967 0, \
16968 Intrinsic::nvvm_wmma_##geom##_mma_and_popc_row_col_##type, \
16969 0, \
16970 0, \
16971 0, \
16972 0, \
16973 0, \
16974 0
16975 // clang-format on
16976 switch (BuiltinID) {
16977 // FP MMA
16978 // Note that 'type' argument of MMA_SATF_VARIANTS uses D_C notation, while
16979 // NumEltsN of return value are ordered as A,B,C,D.
16980 case NVPTX::BI__hmma_m16n16k16_mma_f16f16:
16981 return {8, 8, 4, 4, {{MMA_SATF_VARIANTS(m16n16k16, f16_f16)}}};
16982 case NVPTX::BI__hmma_m16n16k16_mma_f32f16:
16983 return {8, 8, 4, 8, {{MMA_SATF_VARIANTS(m16n16k16, f32_f16)}}};
16984 case NVPTX::BI__hmma_m16n16k16_mma_f16f32:
16985 return {8, 8, 8, 4, {{MMA_SATF_VARIANTS(m16n16k16, f16_f32)}}};
16986 case NVPTX::BI__hmma_m16n16k16_mma_f32f32:
16987 return {8, 8, 8, 8, {{MMA_SATF_VARIANTS(m16n16k16, f32_f32)}}};
16988 case NVPTX::BI__hmma_m32n8k16_mma_f16f16:
16989 return {8, 8, 4, 4, {{MMA_SATF_VARIANTS(m32n8k16, f16_f16)}}};
16990 case NVPTX::BI__hmma_m32n8k16_mma_f32f16:
16991 return {8, 8, 4, 8, {{MMA_SATF_VARIANTS(m32n8k16, f32_f16)}}};
16992 case NVPTX::BI__hmma_m32n8k16_mma_f16f32:
16993 return {8, 8, 8, 4, {{MMA_SATF_VARIANTS(m32n8k16, f16_f32)}}};
16994 case NVPTX::BI__hmma_m32n8k16_mma_f32f32:
16995 return {8, 8, 8, 8, {{MMA_SATF_VARIANTS(m32n8k16, f32_f32)}}};
16996 case NVPTX::BI__hmma_m8n32k16_mma_f16f16:
16997 return {8, 8, 4, 4, {{MMA_SATF_VARIANTS(m8n32k16, f16_f16)}}};
16998 case NVPTX::BI__hmma_m8n32k16_mma_f32f16:
16999 return {8, 8, 4, 8, {{MMA_SATF_VARIANTS(m8n32k16, f32_f16)}}};
17000 case NVPTX::BI__hmma_m8n32k16_mma_f16f32:
17001 return {8, 8, 8, 4, {{MMA_SATF_VARIANTS(m8n32k16, f16_f32)}}};
17002 case NVPTX::BI__hmma_m8n32k16_mma_f32f32:
17003 return {8, 8, 8, 8, {{MMA_SATF_VARIANTS(m8n32k16, f32_f32)}}};
17004
17005 // Integer MMA
17006 case NVPTX::BI__imma_m16n16k16_mma_s8:
17007 return {2, 2, 8, 8, {{MMA_SATF_VARIANTS(m16n16k16, s8)}}};
17008 case NVPTX::BI__imma_m16n16k16_mma_u8:
17009 return {2, 2, 8, 8, {{MMA_SATF_VARIANTS(m16n16k16, u8)}}};
17010 case NVPTX::BI__imma_m32n8k16_mma_s8:
17011 return {4, 1, 8, 8, {{MMA_SATF_VARIANTS(m32n8k16, s8)}}};
17012 case NVPTX::BI__imma_m32n8k16_mma_u8:
17013 return {4, 1, 8, 8, {{MMA_SATF_VARIANTS(m32n8k16, u8)}}};
17014 case NVPTX::BI__imma_m8n32k16_mma_s8:
17015 return {1, 4, 8, 8, {{MMA_SATF_VARIANTS(m8n32k16, s8)}}};
17016 case NVPTX::BI__imma_m8n32k16_mma_u8:
17017 return {1, 4, 8, 8, {{MMA_SATF_VARIANTS(m8n32k16, u8)}}};
17018
17019 // Sub-integer MMA
17020 case NVPTX::BI__imma_m8n8k32_mma_s4:
17021 return {1, 1, 2, 2, {{MMA_VARIANTS_I4(m8n8k32, s4)}}};
17022 case NVPTX::BI__imma_m8n8k32_mma_u4:
17023 return {1, 1, 2, 2, {{MMA_VARIANTS_I4(m8n8k32, u4)}}};
17024 case NVPTX::BI__bmma_m8n8k128_mma_xor_popc_b1:
17025 return {1, 1, 2, 2, {{MMA_VARIANTS_B1_XOR(m8n8k128, b1)}}};
17026 case NVPTX::BI__bmma_m8n8k128_mma_and_popc_b1:
17027 return {1, 1, 2, 2, {{MMA_VARIANTS_B1_AND(m8n8k128, b1)}}};
17028
17029 // Double MMA
17030 case NVPTX::BI__dmma_m8n8k4_mma_f64:
17031 return {1, 1, 2, 2, {{MMA_VARIANTS(m8n8k4, f64)}}};
17032
17033 // Alternate FP MMA
17034 case NVPTX::BI__mma_bf16_m16n16k16_mma_f32:
17035 return {4, 4, 8, 8, {{MMA_VARIANTS(m16n16k16, bf16)}}};
17036 case NVPTX::BI__mma_bf16_m8n32k16_mma_f32:
17037 return {2, 8, 8, 8, {{MMA_VARIANTS(m8n32k16, bf16)}}};
17038 case NVPTX::BI__mma_bf16_m32n8k16_mma_f32:
17039 return {8, 2, 8, 8, {{MMA_VARIANTS(m32n8k16, bf16)}}};
17040 case NVPTX::BI__mma_tf32_m16n16k8_mma_f32:
17041 return {4, 4, 8, 8, {{MMA_VARIANTS(m16n16k8, tf32)}}};
17042 default:
17043 llvm_unreachable("Unexpected builtin ID.")__builtin_unreachable();
17044 }
17045#undef MMA_VARIANTS
17046#undef MMA_SATF_VARIANTS
17047#undef MMA_VARIANTS_I4
17048#undef MMA_VARIANTS_B1_AND
17049#undef MMA_VARIANTS_B1_XOR
17050}
17051
17052} // namespace
17053
17054Value *
17055CodeGenFunction::EmitNVPTXBuiltinExpr(unsigned BuiltinID, const CallExpr *E) {
17056 auto MakeLdg = [&](unsigned IntrinsicID) {
17057 Value *Ptr = EmitScalarExpr(E->getArg(0));
17058 clang::CharUnits Align =
17059 CGM.getNaturalPointeeTypeAlignment(E->getArg(0)->getType());
17060 return Builder.CreateCall(
17061 CGM.getIntrinsic(IntrinsicID, {Ptr->getType()->getPointerElementType(),
17062 Ptr->getType()}),
17063 {Ptr, ConstantInt::get(Builder.getInt32Ty(), Align.getQuantity())});
17064 };
17065 auto MakeScopedAtomic = [&](unsigned IntrinsicID) {
17066 Value *Ptr = EmitScalarExpr(E->getArg(0));
17067 return Builder.CreateCall(
17068 CGM.getIntrinsic(IntrinsicID, {Ptr->getType()->getPointerElementType(),
17069 Ptr->getType()}),
17070 {Ptr, EmitScalarExpr(E->getArg(1))});
17071 };
17072 switch (BuiltinID) {
17073 case NVPTX::BI__nvvm_atom_add_gen_i:
17074 case NVPTX::BI__nvvm_atom_add_gen_l:
17075 case NVPTX::BI__nvvm_atom_add_gen_ll:
17076 return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::Add, E);
17077
17078 case NVPTX::BI__nvvm_atom_sub_gen_i:
17079 case NVPTX::BI__nvvm_atom_sub_gen_l:
17080 case NVPTX::BI__nvvm_atom_sub_gen_ll:
17081 return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::Sub, E);
17082
17083 case NVPTX::BI__nvvm_atom_and_gen_i:
17084 case NVPTX::BI__nvvm_atom_and_gen_l:
17085 case NVPTX::BI__nvvm_atom_and_gen_ll:
17086 return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::And, E);
17087
17088 case NVPTX::BI__nvvm_atom_or_gen_i:
17089 case NVPTX::BI__nvvm_atom_or_gen_l:
17090 case NVPTX::BI__nvvm_atom_or_gen_ll:
17091 return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::Or, E);
17092
17093 case NVPTX::BI__nvvm_atom_xor_gen_i:
17094 case NVPTX::BI__nvvm_atom_xor_gen_l:
17095 case NVPTX::BI__nvvm_atom_xor_gen_ll:
17096 return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::Xor, E);
17097
17098 case NVPTX::BI__nvvm_atom_xchg_gen_i:
17099 case NVPTX::BI__nvvm_atom_xchg_gen_l:
17100 case NVPTX::BI__nvvm_atom_xchg_gen_ll:
17101 return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::Xchg, E);
17102
17103 case NVPTX::BI__nvvm_atom_max_gen_i:
17104 case NVPTX::BI__nvvm_atom_max_gen_l:
17105 case NVPTX::BI__nvvm_atom_max_gen_ll:
17106 return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::Max, E);
17107
17108 case NVPTX::BI__nvvm_atom_max_gen_ui:
17109 case NVPTX::BI__nvvm_atom_max_gen_ul:
17110 case NVPTX::BI__nvvm_atom_max_gen_ull:
17111 return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::UMax, E);
17112
17113 case NVPTX::BI__nvvm_atom_min_gen_i:
17114 case NVPTX::BI__nvvm_atom_min_gen_l:
17115 case NVPTX::BI__nvvm_atom_min_gen_ll:
17116 return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::Min, E);
17117
17118 case NVPTX::BI__nvvm_atom_min_gen_ui:
17119 case NVPTX::BI__nvvm_atom_min_gen_ul:
17120 case NVPTX::BI__nvvm_atom_min_gen_ull:
17121 return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::UMin, E);
17122
17123 case NVPTX::BI__nvvm_atom_cas_gen_i:
17124 case NVPTX::BI__nvvm_atom_cas_gen_l:
17125 case NVPTX::BI__nvvm_atom_cas_gen_ll:
17126 // __nvvm_atom_cas_gen_* should return the old value rather than the
17127 // success flag.
17128 return MakeAtomicCmpXchgValue(*this, E, /*ReturnBool=*/false);
17129
17130 case NVPTX::BI__nvvm_atom_add_gen_f:
17131 case NVPTX::BI__nvvm_atom_add_gen_d: {
17132 Value *Ptr = EmitScalarExpr(E->getArg(0));
17133 Value *Val = EmitScalarExpr(E->getArg(1));
17134 return Builder.CreateAtomicRMW(llvm::AtomicRMWInst::FAdd, Ptr, Val,
17135 AtomicOrdering::SequentiallyConsistent);
17136 }
17137
17138 case NVPTX::BI__nvvm_atom_inc_gen_ui: {
17139 Value *Ptr = EmitScalarExpr(E->getArg(0));
17140 Value *Val = EmitScalarExpr(E->getArg(1));
17141 Function *FnALI32 =
17142 CGM.getIntrinsic(Intrinsic::nvvm_atomic_load_inc_32, Ptr->getType());
17143 return Builder.CreateCall(FnALI32, {Ptr, Val});
17144 }
17145
17146 case NVPTX::BI__nvvm_atom_dec_gen_ui: {
17147 Value *Ptr = EmitScalarExpr(E->getArg(0));
17148 Value *Val = EmitScalarExpr(E->getArg(1));
17149 Function *FnALD32 =
17150 CGM.getIntrinsic(Intrinsic::nvvm_atomic_load_dec_32, Ptr->getType());
17151 return Builder.CreateCall(FnALD32, {Ptr, Val});
17152 }
17153
17154 case NVPTX::BI__nvvm_ldg_c:
17155 case NVPTX::BI__nvvm_ldg_c2:
17156 case NVPTX::BI__nvvm_ldg_c4:
17157 case NVPTX::BI__nvvm_ldg_s:
17158 case NVPTX::BI__nvvm_ldg_s2:
17159 case NVPTX::BI__nvvm_ldg_s4:
17160 case NVPTX::BI__nvvm_ldg_i:
17161 case NVPTX::BI__nvvm_ldg_i2:
17162 case NVPTX::BI__nvvm_ldg_i4:
17163 case NVPTX::BI__nvvm_ldg_l:
17164 case NVPTX::BI__nvvm_ldg_ll:
17165 case NVPTX::BI__nvvm_ldg_ll2:
17166 case NVPTX::BI__nvvm_ldg_uc:
17167 case NVPTX::BI__nvvm_ldg_uc2:
17168 case NVPTX::BI__nvvm_ldg_uc4:
17169 case NVPTX::BI__nvvm_ldg_us:
17170 case NVPTX::BI__nvvm_ldg_us2:
17171 case NVPTX::BI__nvvm_ldg_us4:
17172 case NVPTX::BI__nvvm_ldg_ui:
17173 case NVPTX::BI__nvvm_ldg_ui2:
17174 case NVPTX::BI__nvvm_ldg_ui4:
17175 case NVPTX::BI__nvvm_ldg_ul:
17176 case NVPTX::BI__nvvm_ldg_ull:
17177 case NVPTX::BI__nvvm_ldg_ull2:
17178 // PTX Interoperability section 2.2: "For a vector with an even number of
17179 // elements, its alignment is set to number of elements times the alignment
17180 // of its member: n*alignof(t)."
17181 return MakeLdg(Intrinsic::nvvm_ldg_global_i);
17182 case NVPTX::BI__nvvm_ldg_f:
17183 case NVPTX::BI__nvvm_ldg_f2:
17184 case NVPTX::BI__nvvm_ldg_f4:
17185 case NVPTX::BI__nvvm_ldg_d:
17186 case NVPTX::BI__nvvm_ldg_d2:
17187 return MakeLdg(Intrinsic::nvvm_ldg_global_f);
17188
17189 case NVPTX::BI__nvvm_atom_cta_add_gen_i:
17190 case NVPTX::BI__nvvm_atom_cta_add_gen_l:
17191 case NVPTX::BI__nvvm_atom_cta_add_gen_ll:
17192 return MakeScopedAtomic(Intrinsic::nvvm_atomic_add_gen_i_cta);
17193 case NVPTX::BI__nvvm_atom_sys_add_gen_i:
17194 case NVPTX::BI__nvvm_atom_sys_add_gen_l:
17195 case NVPTX::BI__nvvm_atom_sys_add_gen_ll:
17196 return MakeScopedAtomic(Intrinsic::nvvm_atomic_add_gen_i_sys);
17197 case NVPTX::BI__nvvm_atom_cta_add_gen_f:
17198 case NVPTX::BI__nvvm_atom_cta_add_gen_d:
17199 return MakeScopedAtomic(Intrinsic::nvvm_atomic_add_gen_f_cta);
17200 case NVPTX::BI__nvvm_atom_sys_add_gen_f:
17201 case NVPTX::BI__nvvm_atom_sys_add_gen_d:
17202 return MakeScopedAtomic(Intrinsic::nvvm_atomic_add_gen_f_sys);
17203 case NVPTX::BI__nvvm_atom_cta_xchg_gen_i:
17204 case NVPTX::BI__nvvm_atom_cta_xchg_gen_l:
17205 case NVPTX::BI__nvvm_atom_cta_xchg_gen_ll:
17206 return MakeScopedAtomic(Intrinsic::nvvm_atomic_exch_gen_i_cta);
17207 case NVPTX::BI__nvvm_atom_sys_xchg_gen_i:
17208 case NVPTX::BI__nvvm_atom_sys_xchg_gen_l:
17209 case NVPTX::BI__nvvm_atom_sys_xchg_gen_ll:
17210 return MakeScopedAtomic(Intrinsic::nvvm_atomic_exch_gen_i_sys);
17211 case NVPTX::BI__nvvm_atom_cta_max_gen_i:
17212 case NVPTX::BI__nvvm_atom_cta_max_gen_ui:
17213 case NVPTX::BI__nvvm_atom_cta_max_gen_l:
17214 case NVPTX::BI__nvvm_atom_cta_max_gen_ul:
17215 case NVPTX::BI__nvvm_atom_cta_max_gen_ll:
17216 case NVPTX::BI__nvvm_atom_cta_max_gen_ull:
17217 return MakeScopedAtomic(Intrinsic::nvvm_atomic_max_gen_i_cta);
17218 case NVPTX::BI__nvvm_atom_sys_max_gen_i:
17219 case NVPTX::BI__nvvm_atom_sys_max_gen_ui:
17220 case NVPTX::BI__nvvm_atom_sys_max_gen_l:
17221 case NVPTX::BI__nvvm_atom_sys_max_gen_ul:
17222 case NVPTX::BI__nvvm_atom_sys_max_gen_ll:
17223 case NVPTX::BI__nvvm_atom_sys_max_gen_ull:
17224 return MakeScopedAtomic(Intrinsic::nvvm_atomic_max_gen_i_sys);
17225 case NVPTX::BI__nvvm_atom_cta_min_gen_i:
17226 case NVPTX::BI__nvvm_atom_cta_min_gen_ui:
17227 case NVPTX::BI__nvvm_atom_cta_min_gen_l:
17228 case NVPTX::BI__nvvm_atom_cta_min_gen_ul:
17229 case NVPTX::BI__nvvm_atom_cta_min_gen_ll:
17230 case NVPTX::BI__nvvm_atom_cta_min_gen_ull:
17231 return MakeScopedAtomic(Intrinsic::nvvm_atomic_min_gen_i_cta);
17232 case NVPTX::BI__nvvm_atom_sys_min_gen_i:
17233 case NVPTX::BI__nvvm_atom_sys_min_gen_ui:
17234 case NVPTX::BI__nvvm_atom_sys_min_gen_l:
17235 case NVPTX::BI__nvvm_atom_sys_min_gen_ul:
17236 case NVPTX::BI__nvvm_atom_sys_min_gen_ll:
17237 case NVPTX::BI__nvvm_atom_sys_min_gen_ull:
17238 return MakeScopedAtomic(Intrinsic::nvvm_atomic_min_gen_i_sys);
17239 case NVPTX::BI__nvvm_atom_cta_inc_gen_ui:
17240 return MakeScopedAtomic(Intrinsic::nvvm_atomic_inc_gen_i_cta);
17241 case NVPTX::BI__nvvm_atom_cta_dec_gen_ui:
17242 return MakeScopedAtomic(Intrinsic::nvvm_atomic_dec_gen_i_cta);
17243 case NVPTX::BI__nvvm_atom_sys_inc_gen_ui:
17244 return MakeScopedAtomic(Intrinsic::nvvm_atomic_inc_gen_i_sys);
17245 case NVPTX::BI__nvvm_atom_sys_dec_gen_ui:
17246 return MakeScopedAtomic(Intrinsic::nvvm_atomic_dec_gen_i_sys);
17247 case NVPTX::BI__nvvm_atom_cta_and_gen_i:
17248 case NVPTX::BI__nvvm_atom_cta_and_gen_l:
17249 case NVPTX::BI__nvvm_atom_cta_and_gen_ll:
17250 return MakeScopedAtomic(Intrinsic::nvvm_atomic_and_gen_i_cta);
17251 case NVPTX::BI__nvvm_atom_sys_and_gen_i:
17252 case NVPTX::BI__nvvm_atom_sys_and_gen_l:
17253 case NVPTX::BI__nvvm_atom_sys_and_gen_ll:
17254 return MakeScopedAtomic(Intrinsic::nvvm_atomic_and_gen_i_sys);
17255 case NVPTX::BI__nvvm_atom_cta_or_gen_i:
17256 case NVPTX::BI__nvvm_atom_cta_or_gen_l:
17257 case NVPTX::BI__nvvm_atom_cta_or_gen_ll:
17258 return MakeScopedAtomic(Intrinsic::nvvm_atomic_or_gen_i_cta);
17259 case NVPTX::BI__nvvm_atom_sys_or_gen_i:
17260 case NVPTX::BI__nvvm_atom_sys_or_gen_l:
17261 case NVPTX::BI__nvvm_atom_sys_or_gen_ll:
17262 return MakeScopedAtomic(Intrinsic::nvvm_atomic_or_gen_i_sys);
17263 case NVPTX::BI__nvvm_atom_cta_xor_gen_i:
17264 case NVPTX::BI__nvvm_atom_cta_xor_gen_l:
17265 case NVPTX::BI__nvvm_atom_cta_xor_gen_ll:
17266 return MakeScopedAtomic(Intrinsic::nvvm_atomic_xor_gen_i_cta);
17267 case NVPTX::BI__nvvm_atom_sys_xor_gen_i:
17268 case NVPTX::BI__nvvm_atom_sys_xor_gen_l:
17269 case NVPTX::BI__nvvm_atom_sys_xor_gen_ll:
17270 return MakeScopedAtomic(Intrinsic::nvvm_atomic_xor_gen_i_sys);
17271 case NVPTX::BI__nvvm_atom_cta_cas_gen_i:
17272 case NVPTX::BI__nvvm_atom_cta_cas_gen_l:
17273 case NVPTX::BI__nvvm_atom_cta_cas_gen_ll: {
17274 Value *Ptr = EmitScalarExpr(E->getArg(0));
17275 return Builder.CreateCall(
17276 CGM.getIntrinsic(
17277 Intrinsic::nvvm_atomic_cas_gen_i_cta,
17278 {Ptr->getType()->getPointerElementType(), Ptr->getType()}),
17279 {Ptr, EmitScalarExpr(E->getArg(1)), EmitScalarExpr(E->getArg(2))});
17280 }
17281 case NVPTX::BI__nvvm_atom_sys_cas_gen_i:
17282 case NVPTX::BI__nvvm_atom_sys_cas_gen_l:
17283 case NVPTX::BI__nvvm_atom_sys_cas_gen_ll: {
17284 Value *Ptr = EmitScalarExpr(E->getArg(0));
17285 return Builder.CreateCall(
17286 CGM.getIntrinsic(
17287 Intrinsic::nvvm_atomic_cas_gen_i_sys,
17288 {Ptr->getType()->getPointerElementType(), Ptr->getType()}),
17289 {Ptr, EmitScalarExpr(E->getArg(1)), EmitScalarExpr(E->getArg(2))});
17290 }
17291 case NVPTX::BI__nvvm_match_all_sync_i32p:
17292 case NVPTX::BI__nvvm_match_all_sync_i64p: {
17293 Value *Mask = EmitScalarExpr(E->getArg(0));
17294 Value *Val = EmitScalarExpr(E->getArg(1));
17295 Address PredOutPtr = EmitPointerWithAlignment(E->getArg(2));
17296 Value *ResultPair = Builder.CreateCall(
17297 CGM.getIntrinsic(BuiltinID == NVPTX::BI__nvvm_match_all_sync_i32p
17298 ? Intrinsic::nvvm_match_all_sync_i32p
17299 : Intrinsic::nvvm_match_all_sync_i64p),
17300 {Mask, Val});
17301 Value *Pred = Builder.CreateZExt(Builder.CreateExtractValue(ResultPair, 1),
17302 PredOutPtr.getElementType());
17303 Builder.CreateStore(Pred, PredOutPtr);
17304 return Builder.CreateExtractValue(ResultPair, 0);
17305 }
17306
17307 // FP MMA loads
17308 case NVPTX::BI__hmma_m16n16k16_ld_a:
17309 case NVPTX::BI__hmma_m16n16k16_ld_b:
17310 case NVPTX::BI__hmma_m16n16k16_ld_c_f16:
17311 case NVPTX::BI__hmma_m16n16k16_ld_c_f32:
17312 case NVPTX::BI__hmma_m32n8k16_ld_a:
17313 case NVPTX::BI__hmma_m32n8k16_ld_b:
17314 case NVPTX::BI__hmma_m32n8k16_ld_c_f16:
17315 case NVPTX::BI__hmma_m32n8k16_ld_c_f32:
17316 case NVPTX::BI__hmma_m8n32k16_ld_a:
17317 case NVPTX::BI__hmma_m8n32k16_ld_b:
17318 case NVPTX::BI__hmma_m8n32k16_ld_c_f16:
17319 case NVPTX::BI__hmma_m8n32k16_ld_c_f32:
17320 // Integer MMA loads.
17321 case NVPTX::BI__imma_m16n16k16_ld_a_s8:
17322 case NVPTX::BI__imma_m16n16k16_ld_a_u8:
17323 case NVPTX::BI__imma_m16n16k16_ld_b_s8:
17324 case NVPTX::BI__imma_m16n16k16_ld_b_u8:
17325 case NVPTX::BI__imma_m16n16k16_ld_c:
17326 case NVPTX::BI__imma_m32n8k16_ld_a_s8:
17327 case NVPTX::BI__imma_m32n8k16_ld_a_u8:
17328 case NVPTX::BI__imma_m32n8k16_ld_b_s8:
17329 case NVPTX::BI__imma_m32n8k16_ld_b_u8:
17330 case NVPTX::BI__imma_m32n8k16_ld_c:
17331 case NVPTX::BI__imma_m8n32k16_ld_a_s8:
17332 case NVPTX::BI__imma_m8n32k16_ld_a_u8:
17333 case NVPTX::BI__imma_m8n32k16_ld_b_s8:
17334 case NVPTX::BI__imma_m8n32k16_ld_b_u8:
17335 case NVPTX::BI__imma_m8n32k16_ld_c:
17336 // Sub-integer MMA loads.
17337 case NVPTX::BI__imma_m8n8k32_ld_a_s4:
17338 case NVPTX::BI__imma_m8n8k32_ld_a_u4:
17339 case NVPTX::BI__imma_m8n8k32_ld_b_s4:
17340 case NVPTX::BI__imma_m8n8k32_ld_b_u4:
17341 case NVPTX::BI__imma_m8n8k32_ld_c:
17342 case NVPTX::BI__bmma_m8n8k128_ld_a_b1:
17343 case NVPTX::BI__bmma_m8n8k128_ld_b_b1:
17344 case NVPTX::BI__bmma_m8n8k128_ld_c:
17345 // Double MMA loads.
17346 case NVPTX::BI__dmma_m8n8k4_ld_a:
17347 case NVPTX::BI__dmma_m8n8k4_ld_b:
17348 case NVPTX::BI__dmma_m8n8k4_ld_c:
17349 // Alternate float MMA loads.
17350 case NVPTX::BI__mma_bf16_m16n16k16_ld_a:
17351 case NVPTX::BI__mma_bf16_m16n16k16_ld_b:
17352 case NVPTX::BI__mma_bf16_m8n32k16_ld_a:
17353 case NVPTX::BI__mma_bf16_m8n32k16_ld_b:
17354 case NVPTX::BI__mma_bf16_m32n8k16_ld_a:
17355 case NVPTX::BI__mma_bf16_m32n8k16_ld_b:
17356 case NVPTX::BI__mma_tf32_m16n16k8_ld_a:
17357 case NVPTX::BI__mma_tf32_m16n16k8_ld_b:
17358 case NVPTX::BI__mma_tf32_m16n16k8_ld_c: {
17359 Address Dst = EmitPointerWithAlignment(E->getArg(0));
17360 Value *Src = EmitScalarExpr(E->getArg(1));
17361 Value *Ldm = EmitScalarExpr(E->getArg(2));
17362 Optional<llvm::APSInt> isColMajorArg =
17363 E->getArg(3)->getIntegerConstantExpr(getContext());
17364 if (!isColMajorArg)
17365 return nullptr;
17366 bool isColMajor = isColMajorArg->getSExtValue();
17367 NVPTXMmaLdstInfo II = getNVPTXMmaLdstInfo(BuiltinID);
17368 unsigned IID = isColMajor ? II.IID_col : II.IID_row;
17369 if (IID == 0)
17370 return nullptr;
17371
17372 Value *Result =
17373 Builder.CreateCall(CGM.getIntrinsic(IID, Src->getType()), {Src, Ldm});
17374
17375 // Save returned values.
17376 assert(II.NumResults)((void)0);
17377 if (II.NumResults == 1) {
17378 Builder.CreateAlignedStore(Result, Dst.getPointer(),
17379 CharUnits::fromQuantity(4));
17380 } else {
17381 for (unsigned i = 0; i < II.NumResults; ++i) {
17382 Builder.CreateAlignedStore(
17383 Builder.CreateBitCast(Builder.CreateExtractValue(Result, i),
17384 Dst.getElementType()),
17385 Builder.CreateGEP(Dst.getElementType(), Dst.getPointer(),
17386 llvm::ConstantInt::get(IntTy, i)),
17387 CharUnits::fromQuantity(4));
17388 }
17389 }
17390 return Result;
17391 }
17392
17393 case NVPTX::BI__hmma_m16n16k16_st_c_f16:
17394 case NVPTX::BI__hmma_m16n16k16_st_c_f32:
17395 case NVPTX::BI__hmma_m32n8k16_st_c_f16:
17396 case NVPTX::BI__hmma_m32n8k16_st_c_f32:
17397 case NVPTX::BI__hmma_m8n32k16_st_c_f16:
17398 case NVPTX::BI__hmma_m8n32k16_st_c_f32:
17399 case NVPTX::BI__imma_m16n16k16_st_c_i32:
17400 case NVPTX::BI__imma_m32n8k16_st_c_i32:
17401 case NVPTX::BI__imma_m8n32k16_st_c_i32:
17402 case NVPTX::BI__imma_m8n8k32_st_c_i32:
17403 case NVPTX::BI__bmma_m8n8k128_st_c_i32:
17404 case NVPTX::BI__dmma_m8n8k4_st_c_f64:
17405 case NVPTX::BI__mma_m16n16k8_st_c_f32: {
17406 Value *Dst = EmitScalarExpr(E->getArg(0));
17407 Address Src = EmitPointerWithAlignment(E->getArg(1));
17408 Value *Ldm = EmitScalarExpr(E->getArg(2));
17409 Optional<llvm::APSInt> isColMajorArg =
17410 E->getArg(3)->getIntegerConstantExpr(getContext());
17411 if (!isColMajorArg)
17412 return nullptr;
17413 bool isColMajor = isColMajorArg->getSExtValue();
17414 NVPTXMmaLdstInfo II = getNVPTXMmaLdstInfo(BuiltinID);
17415 unsigned IID = isColMajor ? II.IID_col : II.IID_row;
17416 if (IID == 0)
17417 return nullptr;
17418 Function *Intrinsic =
17419 CGM.getIntrinsic(IID, Dst->getType());
17420 llvm::Type *ParamType = Intrinsic->getFunctionType()->getParamType(1);
17421 SmallVector<Value *, 10> Values = {Dst};
17422 for (unsigned i = 0; i < II.NumResults; ++i) {
17423 Value *V = Builder.CreateAlignedLoad(
17424 Src.getElementType(),
17425 Builder.CreateGEP(Src.getElementType(), Src.getPointer(),
17426 llvm::ConstantInt::get(IntTy, i)),
17427 CharUnits::fromQuantity(4));
17428 Values.push_back(Builder.CreateBitCast(V, ParamType));
17429 }
17430 Values.push_back(Ldm);
17431 Value *Result = Builder.CreateCall(Intrinsic, Values);
17432 return Result;
17433 }
17434
17435 // BI__hmma_m16n16k16_mma_<Dtype><CType>(d, a, b, c, layout, satf) -->
17436 // Intrinsic::nvvm_wmma_m16n16k16_mma_sync<layout A,B><DType><CType><Satf>
17437 case NVPTX::BI__hmma_m16n16k16_mma_f16f16:
17438 case NVPTX::BI__hmma_m16n16k16_mma_f32f16:
17439 case NVPTX::BI__hmma_m16n16k16_mma_f32f32:
17440 case NVPTX::BI__hmma_m16n16k16_mma_f16f32:
17441 case NVPTX::BI__hmma_m32n8k16_mma_f16f16:
17442 case NVPTX::BI__hmma_m32n8k16_mma_f32f16:
17443 case NVPTX::BI__hmma_m32n8k16_mma_f32f32:
17444 case NVPTX::BI__hmma_m32n8k16_mma_f16f32:
17445 case NVPTX::BI__hmma_m8n32k16_mma_f16f16:
17446 case NVPTX::BI__hmma_m8n32k16_mma_f32f16:
17447 case NVPTX::BI__hmma_m8n32k16_mma_f32f32:
17448 case NVPTX::BI__hmma_m8n32k16_mma_f16f32:
17449 case NVPTX::BI__imma_m16n16k16_mma_s8:
17450 case NVPTX::BI__imma_m16n16k16_mma_u8:
17451 case NVPTX::BI__imma_m32n8k16_mma_s8:
17452 case NVPTX::BI__imma_m32n8k16_mma_u8:
17453 case NVPTX::BI__imma_m8n32k16_mma_s8:
17454 case NVPTX::BI__imma_m8n32k16_mma_u8:
17455 case NVPTX::BI__imma_m8n8k32_mma_s4:
17456 case NVPTX::BI__imma_m8n8k32_mma_u4:
17457 case NVPTX::BI__bmma_m8n8k128_mma_xor_popc_b1:
17458 case NVPTX::BI__bmma_m8n8k128_mma_and_popc_b1:
17459 case NVPTX::BI__dmma_m8n8k4_mma_f64:
17460 case NVPTX::BI__mma_bf16_m16n16k16_mma_f32:
17461 case NVPTX::BI__mma_bf16_m8n32k16_mma_f32:
17462 case NVPTX::BI__mma_bf16_m32n8k16_mma_f32:
17463 case NVPTX::BI__mma_tf32_m16n16k8_mma_f32: {
17464 Address Dst = EmitPointerWithAlignment(E->getArg(0));
17465 Address SrcA = EmitPointerWithAlignment(E->getArg(1));
17466 Address SrcB = EmitPointerWithAlignment(E->getArg(2));
17467 Address SrcC = EmitPointerWithAlignment(E->getArg(3));
17468 Optional<llvm::APSInt> LayoutArg =
17469 E->getArg(4)->getIntegerConstantExpr(getContext());
17470 if (!LayoutArg)
17471 return nullptr;
17472 int Layout = LayoutArg->getSExtValue();
17473 if (Layout < 0 || Layout > 3)
17474 return nullptr;
17475 llvm::APSInt SatfArg;
17476 if (BuiltinID == NVPTX::BI__bmma_m8n8k128_mma_xor_popc_b1 ||
17477 BuiltinID == NVPTX::BI__bmma_m8n8k128_mma_and_popc_b1)
17478 SatfArg = 0; // .b1 does not have satf argument.
17479 else if (Optional<llvm::APSInt> OptSatfArg =
17480 E->getArg(5)->getIntegerConstantExpr(getContext()))
17481 SatfArg = *OptSatfArg;
17482 else
17483 return nullptr;
17484 bool Satf = SatfArg.getSExtValue();
17485 NVPTXMmaInfo MI = getNVPTXMmaInfo(BuiltinID);
17486 unsigned IID = MI.getMMAIntrinsic(Layout, Satf);
17487 if (IID == 0) // Unsupported combination of Layout/Satf.
17488 return nullptr;
17489
17490 SmallVector<Value *, 24> Values;
17491 Function *Intrinsic = CGM.getIntrinsic(IID);
17492 llvm::Type *AType = Intrinsic->getFunctionType()->getParamType(0);
17493 // Load A
17494 for (unsigned i = 0; i < MI.NumEltsA; ++i) {
17495 Value *V = Builder.CreateAlignedLoad(
17496 SrcA.getElementType(),
17497 Builder.CreateGEP(SrcA.getElementType(), SrcA.getPointer(),
17498 llvm::ConstantInt::get(IntTy, i)),
17499 CharUnits::fromQuantity(4));
17500 Values.push_back(Builder.CreateBitCast(V, AType));
17501 }
17502 // Load B
17503 llvm::Type *BType = Intrinsic->getFunctionType()->getParamType(MI.NumEltsA);
17504 for (unsigned i = 0; i < MI.NumEltsB; ++i) {
17505 Value *V = Builder.CreateAlignedLoad(
17506 SrcB.getElementType(),
17507 Builder.CreateGEP(SrcB.getElementType(), SrcB.getPointer(),
17508 llvm::ConstantInt::get(IntTy, i)),
17509 CharUnits::fromQuantity(4));
17510 Values.push_back(Builder.CreateBitCast(V, BType));
17511 }
17512 // Load C
17513 llvm::Type *CType =
17514 Intrinsic->getFunctionType()->getParamType(MI.NumEltsA + MI.NumEltsB);
17515 for (unsigned i = 0; i < MI.NumEltsC; ++i) {
17516 Value *V = Builder.CreateAlignedLoad(
17517 SrcC.getElementType(),
17518 Builder.CreateGEP(SrcC.getElementType(), SrcC.getPointer(),
17519 llvm::ConstantInt::get(IntTy, i)),
17520 CharUnits::fromQuantity(4));
17521 Values.push_back(Builder.CreateBitCast(V, CType));
17522 }
17523 Value *Result = Builder.CreateCall(Intrinsic, Values);
17524 llvm::Type *DType = Dst.getElementType();
17525 for (unsigned i = 0; i < MI.NumEltsD; ++i)
17526 Builder.CreateAlignedStore(
17527 Builder.CreateBitCast(Builder.CreateExtractValue(Result, i), DType),
17528 Builder.CreateGEP(Dst.getElementType(), Dst.getPointer(),
17529 llvm::ConstantInt::get(IntTy, i)),
17530 CharUnits::fromQuantity(4));
17531 return Result;
17532 }
17533 default:
17534 return nullptr;
17535 }
17536}
17537
17538namespace {
17539struct BuiltinAlignArgs {
17540 llvm::Value *Src = nullptr;
17541 llvm::Type *SrcType = nullptr;
17542 llvm::Value *Alignment = nullptr;
17543 llvm::Value *Mask = nullptr;
17544 llvm::IntegerType *IntType = nullptr;
17545
17546 BuiltinAlignArgs(const CallExpr *E, CodeGenFunction &CGF) {
17547 QualType AstType = E->getArg(0)->getType();
17548 if (AstType->isArrayType())
17549 Src = CGF.EmitArrayToPointerDecay(E->getArg(0)).getPointer();
17550 else
17551 Src = CGF.EmitScalarExpr(E->getArg(0));
17552 SrcType = Src->getType();
17553 if (SrcType->isPointerTy()) {
17554 IntType = IntegerType::get(
17555 CGF.getLLVMContext(),
17556 CGF.CGM.getDataLayout().getIndexTypeSizeInBits(SrcType));
17557 } else {
17558 assert(SrcType->isIntegerTy())((void)0);
17559 IntType = cast<llvm::IntegerType>(SrcType);
17560 }
17561 Alignment = CGF.EmitScalarExpr(E->getArg(1));
17562 Alignment = CGF.Builder.CreateZExtOrTrunc(Alignment, IntType, "alignment");
17563 auto *One = llvm::ConstantInt::get(IntType, 1);
17564 Mask = CGF.Builder.CreateSub(Alignment, One, "mask");
17565 }
17566};
17567} // namespace
17568
17569/// Generate (x & (y-1)) == 0.
17570RValue CodeGenFunction::EmitBuiltinIsAligned(const CallExpr *E) {
17571 BuiltinAlignArgs Args(E, *this);
17572 llvm::Value *SrcAddress = Args.Src;
17573 if (Args.SrcType->isPointerTy())
17574 SrcAddress =
17575 Builder.CreateBitOrPointerCast(Args.Src, Args.IntType, "src_addr");
17576 return RValue::get(Builder.CreateICmpEQ(
17577 Builder.CreateAnd(SrcAddress, Args.Mask, "set_bits"),
17578 llvm::Constant::getNullValue(Args.IntType), "is_aligned"));
17579}
17580
17581/// Generate (x & ~(y-1)) to align down or ((x+(y-1)) & ~(y-1)) to align up.
17582/// Note: For pointer types we can avoid ptrtoint/inttoptr pairs by using the
17583/// llvm.ptrmask instrinsic (with a GEP before in the align_up case).
17584/// TODO: actually use ptrmask once most optimization passes know about it.
17585RValue CodeGenFunction::EmitBuiltinAlignTo(const CallExpr *E, bool AlignUp) {
17586 BuiltinAlignArgs Args(E, *this);
17587 llvm::Value *SrcAddr = Args.Src;
17588 if (Args.Src->getType()->isPointerTy())
17589 SrcAddr = Builder.CreatePtrToInt(Args.Src, Args.IntType, "intptr");
17590 llvm::Value *SrcForMask = SrcAddr;
17591 if (AlignUp) {
17592 // When aligning up we have to first add the mask to ensure we go over the
17593 // next alignment value and then align down to the next valid multiple.
17594 // By adding the mask, we ensure that align_up on an already aligned
17595 // value will not change the value.
17596 SrcForMask = Builder.CreateAdd(SrcForMask, Args.Mask, "over_boundary");
17597 }
17598 // Invert the mask to only clear the lower bits.
17599 llvm::Value *InvertedMask = Builder.CreateNot(Args.Mask, "inverted_mask");
17600 llvm::Value *Result =
17601 Builder.CreateAnd(SrcForMask, InvertedMask, "aligned_result");
17602 if (Args.Src->getType()->isPointerTy()) {
17603 /// TODO: Use ptrmask instead of ptrtoint+gep once it is optimized well.
17604 // Result = Builder.CreateIntrinsic(
17605 // Intrinsic::ptrmask, {Args.SrcType, SrcForMask->getType(), Args.IntType},
17606 // {SrcForMask, NegatedMask}, nullptr, "aligned_result");
17607 Result->setName("aligned_intptr");
17608 llvm::Value *Difference = Builder.CreateSub(Result, SrcAddr, "diff");
17609 // The result must point to the same underlying allocation. This means we
17610 // can use an inbounds GEP to enable better optimization.
17611 Value *Base = EmitCastToVoidPtr(Args.Src);
17612 if (getLangOpts().isSignedOverflowDefined())
17613 Result = Builder.CreateGEP(Int8Ty, Base, Difference, "aligned_result");
17614 else
17615 Result = EmitCheckedInBoundsGEP(Base, Difference,
17616 /*SignedIndices=*/true,
17617 /*isSubtraction=*/!AlignUp,
17618 E->getExprLoc(), "aligned_result");
17619 Result = Builder.CreatePointerCast(Result, Args.SrcType);
17620 // Emit an alignment assumption to ensure that the new alignment is
17621 // propagated to loads/stores, etc.
17622 emitAlignmentAssumption(Result, E, E->getExprLoc(), Args.Alignment);
17623 }
17624 assert(Result->getType() == Args.SrcType)((void)0);
17625 return RValue::get(Result);
17626}
17627
17628Value *CodeGenFunction::EmitWebAssemblyBuiltinExpr(unsigned BuiltinID,
17629 const CallExpr *E) {
17630 switch (BuiltinID) {
17631 case WebAssembly::BI__builtin_wasm_memory_size: {
17632 llvm::Type *ResultType = ConvertType(E->getType());
17633 Value *I = EmitScalarExpr(E->getArg(0));
17634 Function *Callee =
17635 CGM.getIntrinsic(Intrinsic::wasm_memory_size, ResultType);
17636 return Builder.CreateCall(Callee, I);
17637 }
17638 case WebAssembly::BI__builtin_wasm_memory_grow: {
17639 llvm::Type *ResultType = ConvertType(E->getType());
17640 Value *Args[] = {EmitScalarExpr(E->getArg(0)),
17641 EmitScalarExpr(E->getArg(1))};
17642 Function *Callee =
17643 CGM.getIntrinsic(Intrinsic::wasm_memory_grow, ResultType);
17644 return Builder.CreateCall(Callee, Args);
17645 }
17646 case WebAssembly::BI__builtin_wasm_tls_size: {
17647 llvm::Type *ResultType = ConvertType(E->getType());
17648 Function *Callee = CGM.getIntrinsic(Intrinsic::wasm_tls_size, ResultType);
17649 return Builder.CreateCall(Callee);
17650 }
17651 case WebAssembly::BI__builtin_wasm_tls_align: {
17652 llvm::Type *ResultType = ConvertType(E->getType());
17653 Function *Callee = CGM.getIntrinsic(Intrinsic::wasm_tls_align, ResultType);
17654 return Builder.CreateCall(Callee);
17655 }
17656 case WebAssembly::BI__builtin_wasm_tls_base: {
17657 Function *Callee = CGM.getIntrinsic(Intrinsic::wasm_tls_base);
17658 return Builder.CreateCall(Callee);
17659 }
17660 case WebAssembly::BI__builtin_wasm_throw: {
17661 Value *Tag = EmitScalarExpr(E->getArg(0));
17662 Value *Obj = EmitScalarExpr(E->getArg(1));
17663 Function *Callee = CGM.getIntrinsic(Intrinsic::wasm_throw);
17664 return Builder.CreateCall(Callee, {Tag, Obj});
17665 }
17666 case WebAssembly::BI__builtin_wasm_rethrow: {
17667 Function *Callee = CGM.getIntrinsic(Intrinsic::wasm_rethrow);
17668 return Builder.CreateCall(Callee);
17669 }
17670 case WebAssembly::BI__builtin_wasm_memory_atomic_wait32: {
17671 Value *Addr = EmitScalarExpr(E->getArg(0));
17672 Value *Expected = EmitScalarExpr(E->getArg(1));
17673 Value *Timeout = EmitScalarExpr(E->getArg(2));
17674 Function *Callee = CGM.getIntrinsic(Intrinsic::wasm_memory_atomic_wait32);
17675 return Builder.CreateCall(Callee, {Addr, Expected, Timeout});
17676 }
17677 case WebAssembly::BI__builtin_wasm_memory_atomic_wait64: {
17678 Value *Addr = EmitScalarExpr(E->getArg(0));
17679 Value *Expected = EmitScalarExpr(E->getArg(1));
17680 Value *Timeout = EmitScalarExpr(E->getArg(2));
17681 Function *Callee = CGM.getIntrinsic(Intrinsic::wasm_memory_atomic_wait64);
17682 return Builder.CreateCall(Callee, {Addr, Expected, Timeout});
17683 }
17684 case WebAssembly::BI__builtin_wasm_memory_atomic_notify: {
17685 Value *Addr = EmitScalarExpr(E->getArg(0));
17686 Value *Count = EmitScalarExpr(E->getArg(1));
17687 Function *Callee = CGM.getIntrinsic(Intrinsic::wasm_memory_atomic_notify);
17688 return Builder.CreateCall(Callee, {Addr, Count});
17689 }
17690 case WebAssembly::BI__builtin_wasm_trunc_s_i32_f32:
17691 case WebAssembly::BI__builtin_wasm_trunc_s_i32_f64:
17692 case WebAssembly::BI__builtin_wasm_trunc_s_i64_f32:
17693 case WebAssembly::BI__builtin_wasm_trunc_s_i64_f64: {
17694 Value *Src = EmitScalarExpr(E->getArg(0));
17695 llvm::Type *ResT = ConvertType(E->getType());
17696 Function *Callee =
17697 CGM.getIntrinsic(Intrinsic::wasm_trunc_signed, {ResT, Src->getType()});
17698 return Builder.CreateCall(Callee, {Src});
17699 }
17700 case WebAssembly::BI__builtin_wasm_trunc_u_i32_f32:
17701 case WebAssembly::BI__builtin_wasm_trunc_u_i32_f64:
17702 case WebAssembly::BI__builtin_wasm_trunc_u_i64_f32:
17703 case WebAssembly::BI__builtin_wasm_trunc_u_i64_f64: {
17704 Value *Src = EmitScalarExpr(E->getArg(0));
17705 llvm::Type *ResT = ConvertType(E->getType());
17706 Function *Callee = CGM.getIntrinsic(Intrinsic::wasm_trunc_unsigned,
17707 {ResT, Src->getType()});
17708 return Builder.CreateCall(Callee, {Src});
17709 }
17710 case WebAssembly::BI__builtin_wasm_trunc_saturate_s_i32_f32:
17711 case WebAssembly::BI__builtin_wasm_trunc_saturate_s_i32_f64:
17712 case WebAssembly::BI__builtin_wasm_trunc_saturate_s_i64_f32:
17713 case WebAssembly::BI__builtin_wasm_trunc_saturate_s_i64_f64:
17714 case WebAssembly::BI__builtin_wasm_trunc_saturate_s_i32x4_f32x4: {
17715 Value *Src = EmitScalarExpr(E->getArg(0));
17716 llvm::Type *ResT = ConvertType(E->getType());
17717 Function *Callee =
17718 CGM.getIntrinsic(Intrinsic::fptosi_sat, {ResT, Src->getType()});
17719 return Builder.CreateCall(Callee, {Src});
17720 }
17721 case WebAssembly::BI__builtin_wasm_trunc_saturate_u_i32_f32:
17722 case WebAssembly::BI__builtin_wasm_trunc_saturate_u_i32_f64:
17723 case WebAssembly::BI__builtin_wasm_trunc_saturate_u_i64_f32:
17724 case WebAssembly::BI__builtin_wasm_trunc_saturate_u_i64_f64:
17725 case WebAssembly::BI__builtin_wasm_trunc_saturate_u_i32x4_f32x4: {
17726 Value *Src = EmitScalarExpr(E->getArg(0));
17727 llvm::Type *ResT = ConvertType(E->getType());
17728 Function *Callee =
17729 CGM.getIntrinsic(Intrinsic::fptoui_sat, {ResT, Src->getType()});
17730 return Builder.CreateCall(Callee, {Src});
17731 }
17732 case WebAssembly::BI__builtin_wasm_min_f32:
17733 case WebAssembly::BI__builtin_wasm_min_f64:
17734 case WebAssembly::BI__builtin_wasm_min_f32x4:
17735 case WebAssembly::BI__builtin_wasm_min_f64x2: {
17736 Value *LHS = EmitScalarExpr(E->getArg(0));
17737 Value *RHS = EmitScalarExpr(E->getArg(1));
17738 Function *Callee =
17739 CGM.getIntrinsic(Intrinsic::minimum, ConvertType(E->getType()));
17740 return Builder.CreateCall(Callee, {LHS, RHS});
17741 }
17742 case WebAssembly::BI__builtin_wasm_max_f32:
17743 case WebAssembly::BI__builtin_wasm_max_f64:
17744 case WebAssembly::BI__builtin_wasm_max_f32x4:
17745 case WebAssembly::BI__builtin_wasm_max_f64x2: {
17746 Value *LHS = EmitScalarExpr(E->getArg(0));
17747 Value *RHS = EmitScalarExpr(E->getArg(1));
17748 Function *Callee =
17749 CGM.getIntrinsic(Intrinsic::maximum, ConvertType(E->getType()));
17750 return Builder.CreateCall(Callee, {LHS, RHS});
17751 }
17752 case WebAssembly::BI__builtin_wasm_ceil_f32x4:
17753 case WebAssembly::BI__builtin_wasm_floor_f32x4:
17754 case WebAssembly::BI__builtin_wasm_trunc_f32x4:
17755 case WebAssembly::BI__builtin_wasm_nearest_f32x4:
17756 case WebAssembly::BI__builtin_wasm_ceil_f64x2:
17757 case WebAssembly::BI__builtin_wasm_floor_f64x2:
17758 case WebAssembly::BI__builtin_wasm_trunc_f64x2:
17759 case WebAssembly::BI__builtin_wasm_nearest_f64x2: {
17760 unsigned IntNo;
17761 switch (BuiltinID) {
17762 case WebAssembly::BI__builtin_wasm_ceil_f32x4:
17763 case WebAssembly::BI__builtin_wasm_ceil_f64x2:
17764 IntNo = Intrinsic::ceil;
17765 break;
17766 case WebAssembly::BI__builtin_wasm_floor_f32x4:
17767 case WebAssembly::BI__builtin_wasm_floor_f64x2:
17768 IntNo = Intrinsic::floor;
17769 break;
17770 case WebAssembly::BI__builtin_wasm_trunc_f32x4:
17771 case WebAssembly::BI__builtin_wasm_trunc_f64x2:
17772 IntNo = Intrinsic::trunc;
17773 break;
17774 case WebAssembly::BI__builtin_wasm_nearest_f32x4:
17775 case WebAssembly::BI__builtin_wasm_nearest_f64x2:
17776 IntNo = Intrinsic::nearbyint;
17777 break;
17778 default:
17779 llvm_unreachable("unexpected builtin ID")__builtin_unreachable();
17780 }
17781 Value *Value = EmitScalarExpr(E->getArg(0));
17782 Function *Callee = CGM.getIntrinsic(IntNo, ConvertType(E->getType()));
17783 return Builder.CreateCall(Callee, Value);
17784 }
17785 case WebAssembly::BI__builtin_wasm_swizzle_i8x16: {
17786 Value *Src = EmitScalarExpr(E->getArg(0));
17787 Value *Indices = EmitScalarExpr(E->getArg(1));
17788 Function *Callee = CGM.getIntrinsic(Intrinsic::wasm_swizzle);
17789 return Builder.CreateCall(Callee, {Src, Indices});
17790 }
17791 case WebAssembly::BI__builtin_wasm_add_sat_s_i8x16:
17792 case WebAssembly::BI__builtin_wasm_add_sat_u_i8x16:
17793 case WebAssembly::BI__builtin_wasm_add_sat_s_i16x8:
17794 case WebAssembly::BI__builtin_wasm_add_sat_u_i16x8:
17795 case WebAssembly::BI__builtin_wasm_sub_sat_s_i8x16:
17796 case WebAssembly::BI__builtin_wasm_sub_sat_u_i8x16:
17797 case WebAssembly::BI__builtin_wasm_sub_sat_s_i16x8:
17798 case WebAssembly::BI__builtin_wasm_sub_sat_u_i16x8: {
17799 unsigned IntNo;
17800 switch (BuiltinID) {
17801 case WebAssembly::BI__builtin_wasm_add_sat_s_i8x16:
17802 case WebAssembly::BI__builtin_wasm_add_sat_s_i16x8:
17803 IntNo = Intrinsic::sadd_sat;
17804 break;
17805 case WebAssembly::BI__builtin_wasm_add_sat_u_i8x16:
17806 case WebAssembly::BI__builtin_wasm_add_sat_u_i16x8:
17807 IntNo = Intrinsic::uadd_sat;
17808 break;
17809 case WebAssembly::BI__builtin_wasm_sub_sat_s_i8x16:
17810 case WebAssembly::BI__builtin_wasm_sub_sat_s_i16x8:
17811 IntNo = Intrinsic::wasm_sub_sat_signed;
17812 break;
17813 case WebAssembly::BI__builtin_wasm_sub_sat_u_i8x16:
17814 case WebAssembly::BI__builtin_wasm_sub_sat_u_i16x8:
17815 IntNo = Intrinsic::wasm_sub_sat_unsigned;
17816 break;
17817 default:
17818 llvm_unreachable("unexpected builtin ID")__builtin_unreachable();
17819 }
17820 Value *LHS = EmitScalarExpr(E->getArg(0));
17821 Value *RHS = EmitScalarExpr(E->getArg(1));
17822 Function *Callee = CGM.getIntrinsic(IntNo, ConvertType(E->getType()));
17823 return Builder.CreateCall(Callee, {LHS, RHS});
17824 }
17825 case WebAssembly::BI__builtin_wasm_abs_i8x16:
17826 case WebAssembly::BI__builtin_wasm_abs_i16x8:
17827 case WebAssembly::BI__builtin_wasm_abs_i32x4:
17828 case WebAssembly::BI__builtin_wasm_abs_i64x2: {
17829 Value *Vec = EmitScalarExpr(E->getArg(0));
17830 Value *Neg = Builder.CreateNeg(Vec, "neg");
17831 Constant *Zero = llvm::Constant::getNullValue(Vec->getType());
17832 Value *ICmp = Builder.CreateICmpSLT(Vec, Zero, "abscond");
17833 return Builder.CreateSelect(ICmp, Neg, Vec, "abs");
17834 }
17835 case WebAssembly::BI__builtin_wasm_min_s_i8x16:
17836 case WebAssembly::BI__builtin_wasm_min_u_i8x16:
17837 case WebAssembly::BI__builtin_wasm_max_s_i8x16:
17838 case WebAssembly::BI__builtin_wasm_max_u_i8x16:
17839 case WebAssembly::BI__builtin_wasm_min_s_i16x8:
17840 case WebAssembly::BI__builtin_wasm_min_u_i16x8:
17841 case WebAssembly::BI__builtin_wasm_max_s_i16x8:
17842 case WebAssembly::BI__builtin_wasm_max_u_i16x8:
17843 case WebAssembly::BI__builtin_wasm_min_s_i32x4:
17844 case WebAssembly::BI__builtin_wasm_min_u_i32x4:
17845 case WebAssembly::BI__builtin_wasm_max_s_i32x4:
17846 case WebAssembly::BI__builtin_wasm_max_u_i32x4: {
17847 Value *LHS = EmitScalarExpr(E->getArg(0));
17848 Value *RHS = EmitScalarExpr(E->getArg(1));
17849 Value *ICmp;
17850 switch (BuiltinID) {
17851 case WebAssembly::BI__builtin_wasm_min_s_i8x16:
17852 case WebAssembly::BI__builtin_wasm_min_s_i16x8:
17853 case WebAssembly::BI__builtin_wasm_min_s_i32x4:
17854 ICmp = Builder.CreateICmpSLT(LHS, RHS);
17855 break;
17856 case WebAssembly::BI__builtin_wasm_min_u_i8x16:
17857 case WebAssembly::BI__builtin_wasm_min_u_i16x8:
17858 case WebAssembly::BI__builtin_wasm_min_u_i32x4:
17859 ICmp = Builder.CreateICmpULT(LHS, RHS);
17860 break;
17861 case WebAssembly::BI__builtin_wasm_max_s_i8x16:
17862 case WebAssembly::BI__builtin_wasm_max_s_i16x8:
17863 case WebAssembly::BI__builtin_wasm_max_s_i32x4:
17864 ICmp = Builder.CreateICmpSGT(LHS, RHS);
17865 break;
17866 case WebAssembly::BI__builtin_wasm_max_u_i8x16:
17867 case WebAssembly::BI__builtin_wasm_max_u_i16x8:
17868 case WebAssembly::BI__builtin_wasm_max_u_i32x4:
17869 ICmp = Builder.CreateICmpUGT(LHS, RHS);
17870 break;
17871 default:
17872 llvm_unreachable("unexpected builtin ID")__builtin_unreachable();
17873 }
17874 return Builder.CreateSelect(ICmp, LHS, RHS);
17875 }
17876 case WebAssembly::BI__builtin_wasm_avgr_u_i8x16:
17877 case WebAssembly::BI__builtin_wasm_avgr_u_i16x8: {
17878 Value *LHS = EmitScalarExpr(E->getArg(0));
17879 Value *RHS = EmitScalarExpr(E->getArg(1));
17880 Function *Callee = CGM.getIntrinsic(Intrinsic::wasm_avgr_unsigned,
17881 ConvertType(E->getType()));
17882 return Builder.CreateCall(Callee, {LHS, RHS});
17883 }
17884 case WebAssembly::BI__builtin_wasm_q15mulr_sat_s_i16x8: {
17885 Value *LHS = EmitScalarExpr(E->getArg(0));
17886 Value *RHS = EmitScalarExpr(E->getArg(1));
17887 Function *Callee = CGM.getIntrinsic(Intrinsic::wasm_q15mulr_sat_signed);
17888 return Builder.CreateCall(Callee, {LHS, RHS});
17889 }
17890 case WebAssembly::BI__builtin_wasm_extadd_pairwise_i8x16_s_i16x8:
17891 case WebAssembly::BI__builtin_wasm_extadd_pairwise_i8x16_u_i16x8:
17892 case WebAssembly::BI__builtin_wasm_extadd_pairwise_i16x8_s_i32x4:
17893 case WebAssembly::BI__builtin_wasm_extadd_pairwise_i16x8_u_i32x4: {
17894 Value *Vec = EmitScalarExpr(E->getArg(0));
17895 unsigned IntNo;
17896 switch (BuiltinID) {
17897 case WebAssembly::BI__builtin_wasm_extadd_pairwise_i8x16_s_i16x8:
17898 case WebAssembly::BI__builtin_wasm_extadd_pairwise_i16x8_s_i32x4:
17899 IntNo = Intrinsic::wasm_extadd_pairwise_signed;
17900 break;
17901 case WebAssembly::BI__builtin_wasm_extadd_pairwise_i8x16_u_i16x8:
17902 case WebAssembly::BI__builtin_wasm_extadd_pairwise_i16x8_u_i32x4:
17903 IntNo = Intrinsic::wasm_extadd_pairwise_unsigned;
17904 break;
17905 default:
17906 llvm_unreachable("unexptected builtin ID")__builtin_unreachable();
17907 }
17908
17909 Function *Callee = CGM.getIntrinsic(IntNo, ConvertType(E->getType()));
17910 return Builder.CreateCall(Callee, Vec);
17911 }
17912 case WebAssembly::BI__builtin_wasm_bitselect: {
17913 Value *V1 = EmitScalarExpr(E->getArg(0));
17914 Value *V2 = EmitScalarExpr(E->getArg(1));
17915 Value *C = EmitScalarExpr(E->getArg(2));
17916 Function *Callee =
17917 CGM.getIntrinsic(Intrinsic::wasm_bitselect, ConvertType(E->getType()));
17918 return Builder.CreateCall(Callee, {V1, V2, C});
17919 }
17920 case WebAssembly::BI__builtin_wasm_dot_s_i32x4_i16x8: {
17921 Value *LHS = EmitScalarExpr(E->getArg(0));
17922 Value *RHS = EmitScalarExpr(E->getArg(1));
17923 Function *Callee = CGM.getIntrinsic(Intrinsic::wasm_dot);
17924 return Builder.CreateCall(Callee, {LHS, RHS});
17925 }
17926 case WebAssembly::BI__builtin_wasm_popcnt_i8x16: {
17927 Value *Vec = EmitScalarExpr(E->getArg(0));
17928 Function *Callee =
17929 CGM.getIntrinsic(Intrinsic::ctpop, ConvertType(E->getType()));
17930 return Builder.CreateCall(Callee, {Vec});
17931 }
17932 case WebAssembly::BI__builtin_wasm_any_true_v128:
17933 case WebAssembly::BI__builtin_wasm_all_true_i8x16:
17934 case WebAssembly::BI__builtin_wasm_all_true_i16x8:
17935 case WebAssembly::BI__builtin_wasm_all_true_i32x4:
17936 case WebAssembly::BI__builtin_wasm_all_true_i64x2: {
17937 unsigned IntNo;
17938 switch (BuiltinID) {
17939 case WebAssembly::BI__builtin_wasm_any_true_v128:
17940 IntNo = Intrinsic::wasm_anytrue;
17941 break;
17942 case WebAssembly::BI__builtin_wasm_all_true_i8x16:
17943 case WebAssembly::BI__builtin_wasm_all_true_i16x8:
17944 case WebAssembly::BI__builtin_wasm_all_true_i32x4:
17945 case WebAssembly::BI__builtin_wasm_all_true_i64x2:
17946 IntNo = Intrinsic::wasm_alltrue;
17947 break;
17948 default:
17949 llvm_unreachable("unexpected builtin ID")__builtin_unreachable();
17950 }
17951 Value *Vec = EmitScalarExpr(E->getArg(0));
17952 Function *Callee = CGM.getIntrinsic(IntNo, Vec->getType());
17953 return Builder.CreateCall(Callee, {Vec});
17954 }
17955 case WebAssembly::BI__builtin_wasm_bitmask_i8x16:
17956 case WebAssembly::BI__builtin_wasm_bitmask_i16x8:
17957 case WebAssembly::BI__builtin_wasm_bitmask_i32x4:
17958 case WebAssembly::BI__builtin_wasm_bitmask_i64x2: {
17959 Value *Vec = EmitScalarExpr(E->getArg(0));
17960 Function *Callee =
17961 CGM.getIntrinsic(Intrinsic::wasm_bitmask, Vec->getType());
17962 return Builder.CreateCall(Callee, {Vec});
17963 }
17964 case WebAssembly::BI__builtin_wasm_abs_f32x4:
17965 case WebAssembly::BI__builtin_wasm_abs_f64x2: {
17966 Value *Vec = EmitScalarExpr(E->getArg(0));
17967 Function *Callee = CGM.getIntrinsic(Intrinsic::fabs, Vec->getType());
17968 return Builder.CreateCall(Callee, {Vec});
17969 }
17970 case WebAssembly::BI__builtin_wasm_sqrt_f32x4:
17971 case WebAssembly::BI__builtin_wasm_sqrt_f64x2: {
17972 Value *Vec = EmitScalarExpr(E->getArg(0));
17973 Function *Callee = CGM.getIntrinsic(Intrinsic::sqrt, Vec->getType());
17974 return Builder.CreateCall(Callee, {Vec});
17975 }
17976 case WebAssembly::BI__builtin_wasm_narrow_s_i8x16_i16x8:
17977 case WebAssembly::BI__builtin_wasm_narrow_u_i8x16_i16x8:
17978 case WebAssembly::BI__builtin_wasm_narrow_s_i16x8_i32x4:
17979 case WebAssembly::BI__builtin_wasm_narrow_u_i16x8_i32x4: {
17980 Value *Low = EmitScalarExpr(E->getArg(0));
17981 Value *High = EmitScalarExpr(E->getArg(1));
17982 unsigned IntNo;
17983 switch (BuiltinID) {
17984 case WebAssembly::BI__builtin_wasm_narrow_s_i8x16_i16x8:
17985 case WebAssembly::BI__builtin_wasm_narrow_s_i16x8_i32x4:
17986 IntNo = Intrinsic::wasm_narrow_signed;
17987 break;
17988 case WebAssembly::BI__builtin_wasm_narrow_u_i8x16_i16x8:
17989 case WebAssembly::BI__builtin_wasm_narrow_u_i16x8_i32x4:
17990 IntNo = Intrinsic::wasm_narrow_unsigned;
17991 break;
17992 default:
17993 llvm_unreachable("unexpected builtin ID")__builtin_unreachable();
17994 }
17995 Function *Callee =
17996 CGM.getIntrinsic(IntNo, {ConvertType(E->getType()), Low->getType()});
17997 return Builder.CreateCall(Callee, {Low, High});
17998 }
17999 case WebAssembly::BI__builtin_wasm_trunc_sat_zero_s_f64x2_i32x4:
18000 case WebAssembly::BI__builtin_wasm_trunc_sat_zero_u_f64x2_i32x4: {
18001 Value *Vec = EmitScalarExpr(E->getArg(0));
18002 unsigned IntNo;
18003 switch (BuiltinID) {
18004 case WebAssembly::BI__builtin_wasm_trunc_sat_zero_s_f64x2_i32x4:
18005 IntNo = Intrinsic::fptosi_sat;
18006 break;
18007 case WebAssembly::BI__builtin_wasm_trunc_sat_zero_u_f64x2_i32x4:
18008 IntNo = Intrinsic::fptoui_sat;
18009 break;
18010 default:
18011 llvm_unreachable("unexpected builtin ID")__builtin_unreachable();
18012 }
18013 llvm::Type *SrcT = Vec->getType();
18014 llvm::Type *TruncT =
18015 SrcT->getWithNewType(llvm::IntegerType::get(getLLVMContext(), 32));
18016 Function *Callee = CGM.getIntrinsic(IntNo, {TruncT, SrcT});
18017 Value *Trunc = Builder.CreateCall(Callee, Vec);
18018 Value *Splat = Builder.CreateVectorSplat(2, Builder.getInt32(0));
18019 Value *ConcatMask =
18020 llvm::ConstantVector::get({Builder.getInt32(0), Builder.getInt32(1),
18021 Builder.getInt32(2), Builder.getInt32(3)});
18022 return Builder.CreateShuffleVector(Trunc, Splat, ConcatMask);
18023 }
18024 case WebAssembly::BI__builtin_wasm_shuffle_i8x16: {
18025 Value *Ops[18];
18026 size_t OpIdx = 0;
18027 Ops[OpIdx++] = EmitScalarExpr(E->getArg(0));
18028 Ops[OpIdx++] = EmitScalarExpr(E->getArg(1));
18029 while (OpIdx < 18) {
18030 Optional<llvm::APSInt> LaneConst =
18031 E->getArg(OpIdx)->getIntegerConstantExpr(getContext());
18032 assert(LaneConst && "Constant arg isn't actually constant?")((void)0);
18033 Ops[OpIdx++] = llvm::ConstantInt::get(getLLVMContext(), *LaneConst);
18034 }
18035 Function *Callee = CGM.getIntrinsic(Intrinsic::wasm_shuffle);
18036 return Builder.CreateCall(Callee, Ops);
18037 }
18038 default:
18039 return nullptr;
18040 }
18041}
18042
18043static std::pair<Intrinsic::ID, unsigned>
18044getIntrinsicForHexagonNonGCCBuiltin(unsigned BuiltinID) {
18045 struct Info {
18046 unsigned BuiltinID;
18047 Intrinsic::ID IntrinsicID;
18048 unsigned VecLen;
18049 };
18050 Info Infos[] = {
18051#define CUSTOM_BUILTIN_MAPPING(x,s) \
18052 { Hexagon::BI__builtin_HEXAGON_##x, Intrinsic::hexagon_##x, s },
18053 CUSTOM_BUILTIN_MAPPING(L2_loadrub_pci, 0)
18054 CUSTOM_BUILTIN_MAPPING(L2_loadrb_pci, 0)
18055 CUSTOM_BUILTIN_MAPPING(L2_loadruh_pci, 0)
18056 CUSTOM_BUILTIN_MAPPING(L2_loadrh_pci, 0)
18057 CUSTOM_BUILTIN_MAPPING(L2_loadri_pci, 0)
18058 CUSTOM_BUILTIN_MAPPING(L2_loadrd_pci, 0)
18059 CUSTOM_BUILTIN_MAPPING(L2_loadrub_pcr, 0)
18060 CUSTOM_BUILTIN_MAPPING(L2_loadrb_pcr, 0)
18061 CUSTOM_BUILTIN_MAPPING(L2_loadruh_pcr, 0)
18062 CUSTOM_BUILTIN_MAPPING(L2_loadrh_pcr, 0)
18063 CUSTOM_BUILTIN_MAPPING(L2_loadri_pcr, 0)
18064 CUSTOM_BUILTIN_MAPPING(L2_loadrd_pcr, 0)
18065 CUSTOM_BUILTIN_MAPPING(S2_storerb_pci, 0)
18066 CUSTOM_BUILTIN_MAPPING(S2_storerh_pci, 0)
18067 CUSTOM_BUILTIN_MAPPING(S2_storerf_pci, 0)
18068 CUSTOM_BUILTIN_MAPPING(S2_storeri_pci, 0)
18069 CUSTOM_BUILTIN_MAPPING(S2_storerd_pci, 0)
18070 CUSTOM_BUILTIN_MAPPING(S2_storerb_pcr, 0)
18071 CUSTOM_BUILTIN_MAPPING(S2_storerh_pcr, 0)
18072 CUSTOM_BUILTIN_MAPPING(S2_storerf_pcr, 0)
18073 CUSTOM_BUILTIN_MAPPING(S2_storeri_pcr, 0)
18074 CUSTOM_BUILTIN_MAPPING(S2_storerd_pcr, 0)
18075 CUSTOM_BUILTIN_MAPPING(V6_vmaskedstoreq, 64)
18076 CUSTOM_BUILTIN_MAPPING(V6_vmaskedstorenq, 64)
18077 CUSTOM_BUILTIN_MAPPING(V6_vmaskedstorentq, 64)
18078 CUSTOM_BUILTIN_MAPPING(V6_vmaskedstorentnq, 64)
18079 CUSTOM_BUILTIN_MAPPING(V6_vmaskedstoreq_128B, 128)
18080 CUSTOM_BUILTIN_MAPPING(V6_vmaskedstorenq_128B, 128)
18081 CUSTOM_BUILTIN_MAPPING(V6_vmaskedstorentq_128B, 128)
18082 CUSTOM_BUILTIN_MAPPING(V6_vmaskedstorentnq_128B, 128)
18083#include "clang/Basic/BuiltinsHexagonMapCustomDep.def"
18084#undef CUSTOM_BUILTIN_MAPPING
18085 };
18086
18087 auto CmpInfo = [] (Info A, Info B) { return A.BuiltinID < B.BuiltinID; };
18088 static const bool SortOnce = (llvm::sort(Infos, CmpInfo), true);
18089 (void)SortOnce;
18090
18091 const Info *F = std::lower_bound(std::begin(Infos), std::end(Infos),
18092 Info{BuiltinID, 0, 0}, CmpInfo);
18093 if (F == std::end(Infos) || F->BuiltinID != BuiltinID)
18094 return {Intrinsic::not_intrinsic, 0};
18095
18096 return {F->IntrinsicID, F->VecLen};
18097}
18098
18099Value *CodeGenFunction::EmitHexagonBuiltinExpr(unsigned BuiltinID,
18100 const CallExpr *E) {
18101 Intrinsic::ID ID;
18102 unsigned VecLen;
18103 std::tie(ID, VecLen) = getIntrinsicForHexagonNonGCCBuiltin(BuiltinID);
18104
18105 auto MakeCircOp = [this, E](unsigned IntID, bool IsLoad) {
18106 // The base pointer is passed by address, so it needs to be loaded.
18107 Address A = EmitPointerWithAlignment(E->getArg(0));
18108 Address BP = Address(
18109 Builder.CreateBitCast(A.getPointer(), Int8PtrPtrTy), A.getAlignment());
18110 llvm::Value *Base = Builder.CreateLoad(BP);
18111 // The treatment of both loads and stores is the same: the arguments for
18112 // the builtin are the same as the arguments for the intrinsic.
18113 // Load:
18114 // builtin(Base, Inc, Mod, Start) -> intr(Base, Inc, Mod, Start)
18115 // builtin(Base, Mod, Start) -> intr(Base, Mod, Start)
18116 // Store:
18117 // builtin(Base, Inc, Mod, Val, Start) -> intr(Base, Inc, Mod, Val, Start)
18118 // builtin(Base, Mod, Val, Start) -> intr(Base, Mod, Val, Start)
18119 SmallVector<llvm::Value*,5> Ops = { Base };
18120 for (unsigned i = 1, e = E->getNumArgs(); i != e; ++i)
18121 Ops.push_back(EmitScalarExpr(E->getArg(i)));
18122
18123 llvm::Value *Result = Builder.CreateCall(CGM.getIntrinsic(IntID), Ops);
18124 // The load intrinsics generate two results (Value, NewBase), stores
18125 // generate one (NewBase). The new base address needs to be stored.
18126 llvm::Value *NewBase = IsLoad ? Builder.CreateExtractValue(Result, 1)
18127 : Result;
18128 llvm::Value *LV = Builder.CreateBitCast(
18129 EmitScalarExpr(E->getArg(0)), NewBase->getType()->getPointerTo());
18130 Address Dest = EmitPointerWithAlignment(E->getArg(0));
18131 llvm::Value *RetVal =
18132 Builder.CreateAlignedStore(NewBase, LV, Dest.getAlignment());
18133 if (IsLoad)
18134 RetVal = Builder.CreateExtractValue(Result, 0);
18135 return RetVal;
18136 };
18137
18138 // Handle the conversion of bit-reverse load intrinsics to bit code.
18139 // The intrinsic call after this function only reads from memory and the
18140 // write to memory is dealt by the store instruction.
18141 auto MakeBrevLd = [this, E](unsigned IntID, llvm::Type *DestTy) {
18142 // The intrinsic generates one result, which is the new value for the base
18143 // pointer. It needs to be returned. The result of the load instruction is
18144 // passed to intrinsic by address, so the value needs to be stored.
18145 llvm::Value *BaseAddress =
18146 Builder.CreateBitCast(EmitScalarExpr(E->getArg(0)), Int8PtrTy);
18147
18148 // Expressions like &(*pt++) will be incremented per evaluation.
18149 // EmitPointerWithAlignment and EmitScalarExpr evaluates the expression
18150 // per call.
18151 Address DestAddr = EmitPointerWithAlignment(E->getArg(1));
18152 DestAddr = Address(Builder.CreateBitCast(DestAddr.getPointer(), Int8PtrTy),
18153 DestAddr.getAlignment());
18154 llvm::Value *DestAddress = DestAddr.getPointer();
18155
18156 // Operands are Base, Dest, Modifier.
18157 // The intrinsic format in LLVM IR is defined as
18158 // { ValueType, i8* } (i8*, i32).
18159 llvm::Value *Result = Builder.CreateCall(
18160 CGM.getIntrinsic(IntID), {BaseAddress, EmitScalarExpr(E->getArg(2))});
18161
18162 // The value needs to be stored as the variable is passed by reference.
18163 llvm::Value *DestVal = Builder.CreateExtractValue(Result, 0);
18164
18165 // The store needs to be truncated to fit the destination type.
18166 // While i32 and i64 are natively supported on Hexagon, i8 and i16 needs
18167 // to be handled with stores of respective destination type.
18168 DestVal = Builder.CreateTrunc(DestVal, DestTy);
18169
18170 llvm::Value *DestForStore =
18171 Builder.CreateBitCast(DestAddress, DestVal->getType()->getPointerTo());
18172 Builder.CreateAlignedStore(DestVal, DestForStore, DestAddr.getAlignment());
18173 // The updated value of the base pointer is returned.
18174 return Builder.CreateExtractValue(Result, 1);
18175 };
18176
18177 auto V2Q = [this, VecLen] (llvm::Value *Vec) {
18178 Intrinsic::ID ID = VecLen == 128 ? Intrinsic::hexagon_V6_vandvrt_128B
18179 : Intrinsic::hexagon_V6_vandvrt;
18180 return Builder.CreateCall(CGM.getIntrinsic(ID),
18181 {Vec, Builder.getInt32(-1)});
18182 };
18183 auto Q2V = [this, VecLen] (llvm::Value *Pred) {
18184 Intrinsic::ID ID = VecLen == 128 ? Intrinsic::hexagon_V6_vandqrt_128B
18185 : Intrinsic::hexagon_V6_vandqrt;
18186 return Builder.CreateCall(CGM.getIntrinsic(ID),
18187 {Pred, Builder.getInt32(-1)});
18188 };
18189
18190 switch (BuiltinID) {
18191 // These intrinsics return a tuple {Vector, VectorPred} in LLVM IR,
18192 // and the corresponding C/C++ builtins use loads/stores to update
18193 // the predicate.
18194 case Hexagon::BI__builtin_HEXAGON_V6_vaddcarry:
18195 case Hexagon::BI__builtin_HEXAGON_V6_vaddcarry_128B:
18196 case Hexagon::BI__builtin_HEXAGON_V6_vsubcarry:
18197 case Hexagon::BI__builtin_HEXAGON_V6_vsubcarry_128B: {
18198 // Get the type from the 0-th argument.
18199 llvm::Type *VecType = ConvertType(E->getArg(0)->getType());
18200 Address PredAddr = Builder.CreateBitCast(
18201 EmitPointerWithAlignment(E->getArg(2)), VecType->getPointerTo(0));
18202 llvm::Value *PredIn = V2Q(Builder.CreateLoad(PredAddr));
18203 llvm::Value *Result = Builder.CreateCall(CGM.getIntrinsic(ID),
18204 {EmitScalarExpr(E->getArg(0)), EmitScalarExpr(E->getArg(1)), PredIn});
18205
18206 llvm::Value *PredOut = Builder.CreateExtractValue(Result, 1);
18207 Builder.CreateAlignedStore(Q2V(PredOut), PredAddr.getPointer(),
18208 PredAddr.getAlignment());
18209 return Builder.CreateExtractValue(Result, 0);
18210 }
18211
18212 case Hexagon::BI__builtin_HEXAGON_L2_loadrub_pci:
18213 case Hexagon::BI__builtin_HEXAGON_L2_loadrb_pci:
18214 case Hexagon::BI__builtin_HEXAGON_L2_loadruh_pci:
18215 case Hexagon::BI__builtin_HEXAGON_L2_loadrh_pci:
18216 case Hexagon::BI__builtin_HEXAGON_L2_loadri_pci:
18217 case Hexagon::BI__builtin_HEXAGON_L2_loadrd_pci:
18218 case Hexagon::BI__builtin_HEXAGON_L2_loadrub_pcr:
18219 case Hexagon::BI__builtin_HEXAGON_L2_loadrb_pcr:
18220 case Hexagon::BI__builtin_HEXAGON_L2_loadruh_pcr:
18221 case Hexagon::BI__builtin_HEXAGON_L2_loadrh_pcr:
18222 case Hexagon::BI__builtin_HEXAGON_L2_loadri_pcr:
18223 case Hexagon::BI__builtin_HEXAGON_L2_loadrd_pcr:
18224 return MakeCircOp(ID, /*IsLoad=*/true);
18225 case Hexagon::BI__builtin_HEXAGON_S2_storerb_pci:
18226 case Hexagon::BI__builtin_HEXAGON_S2_storerh_pci:
18227 case Hexagon::BI__builtin_HEXAGON_S2_storerf_pci:
18228 case Hexagon::BI__builtin_HEXAGON_S2_storeri_pci:
18229 case Hexagon::BI__builtin_HEXAGON_S2_storerd_pci:
18230 case Hexagon::BI__builtin_HEXAGON_S2_storerb_pcr:
18231 case Hexagon::BI__builtin_HEXAGON_S2_storerh_pcr:
18232 case Hexagon::BI__builtin_HEXAGON_S2_storerf_pcr:
18233 case Hexagon::BI__builtin_HEXAGON_S2_storeri_pcr:
18234 case Hexagon::BI__builtin_HEXAGON_S2_storerd_pcr:
18235 return MakeCircOp(ID, /*IsLoad=*/false);
18236 case Hexagon::BI__builtin_brev_ldub:
18237 return MakeBrevLd(Intrinsic::hexagon_L2_loadrub_pbr, Int8Ty);
18238 case Hexagon::BI__builtin_brev_ldb:
18239 return MakeBrevLd(Intrinsic::hexagon_L2_loadrb_pbr, Int8Ty);
18240 case Hexagon::BI__builtin_brev_lduh:
18241 return MakeBrevLd(Intrinsic::hexagon_L2_loadruh_pbr, Int16Ty);
18242 case Hexagon::BI__builtin_brev_ldh:
18243 return MakeBrevLd(Intrinsic::hexagon_L2_loadrh_pbr, Int16Ty);
18244 case Hexagon::BI__builtin_brev_ldw:
18245 return MakeBrevLd(Intrinsic::hexagon_L2_loadri_pbr, Int32Ty);
18246 case Hexagon::BI__builtin_brev_ldd:
18247 return MakeBrevLd(Intrinsic::hexagon_L2_loadrd_pbr, Int64Ty);
18248
18249 default: {
18250 if (ID == Intrinsic::not_intrinsic)
18251 return nullptr;
18252
18253 auto IsVectorPredTy = [](llvm::Type *T) {
18254 return T->isVectorTy() &&
18255 cast<llvm::VectorType>(T)->getElementType()->isIntegerTy(1);
18256 };
18257
18258 llvm::Function *IntrFn = CGM.getIntrinsic(ID);
18259 llvm::FunctionType *IntrTy = IntrFn->getFunctionType();
18260 SmallVector<llvm::Value*,4> Ops;
18261 for (unsigned i = 0, e = IntrTy->getNumParams(); i != e; ++i) {
18262 llvm::Type *T = IntrTy->getParamType(i);
18263 const Expr *A = E->getArg(i);
18264 if (IsVectorPredTy(T)) {
18265 // There will be an implicit cast to a boolean vector. Strip it.
18266 if (auto *Cast = dyn_cast<ImplicitCastExpr>(A)) {
18267 if (Cast->getCastKind() == CK_BitCast)
18268 A = Cast->getSubExpr();
18269 }
18270 Ops.push_back(V2Q(EmitScalarExpr(A)));
18271 } else {
18272 Ops.push_back(EmitScalarExpr(A));
18273 }
18274 }
18275
18276 llvm::Value *Call = Builder.CreateCall(IntrFn, Ops);
18277 if (IsVectorPredTy(IntrTy->getReturnType()))
18278 Call = Q2V(Call);
18279
18280 return Call;
18281 } // default
18282 } // switch
18283
18284 return nullptr;
18285}
18286
18287Value *CodeGenFunction::EmitRISCVBuiltinExpr(unsigned BuiltinID,
18288 const CallExpr *E,
18289 ReturnValueSlot ReturnValue) {
18290 SmallVector<Value *, 4> Ops;
18291 llvm::Type *ResultType = ConvertType(E->getType());
18292
18293 for (unsigned i = 0, e = E->getNumArgs(); i != e; i++)
18294 Ops.push_back(EmitScalarExpr(E->getArg(i)));
18295
18296 Intrinsic::ID ID = Intrinsic::not_intrinsic;
18297 unsigned NF = 1;
18298
18299 // Required for overloaded intrinsics.
18300 llvm::SmallVector<llvm::Type *, 2> IntrinsicTypes;
18301 switch (BuiltinID) {
18302 default: llvm_unreachable("unexpected builtin ID")__builtin_unreachable();
18303 case RISCV::BI__builtin_riscv_orc_b_32:
18304 case RISCV::BI__builtin_riscv_orc_b_64:
18305 case RISCV::BI__builtin_riscv_clmul:
18306 case RISCV::BI__builtin_riscv_clmulh:
18307 case RISCV::BI__builtin_riscv_clmulr:
18308 case RISCV::BI__builtin_riscv_bcompress_32:
18309 case RISCV::BI__builtin_riscv_bcompress_64:
18310 case RISCV::BI__builtin_riscv_bdecompress_32:
18311 case RISCV::BI__builtin_riscv_bdecompress_64:
18312 case RISCV::BI__builtin_riscv_grev_32:
18313 case RISCV::BI__builtin_riscv_grev_64:
18314 case RISCV::BI__builtin_riscv_gorc_32:
18315 case RISCV::BI__builtin_riscv_gorc_64:
18316 case RISCV::BI__builtin_riscv_shfl_32:
18317 case RISCV::BI__builtin_riscv_shfl_64:
18318 case RISCV::BI__builtin_riscv_unshfl_32:
18319 case RISCV::BI__builtin_riscv_unshfl_64:
18320 case RISCV::BI__builtin_riscv_xperm_n:
18321 case RISCV::BI__builtin_riscv_xperm_b:
18322 case RISCV::BI__builtin_riscv_xperm_h:
18323 case RISCV::BI__builtin_riscv_xperm_w:
18324 case RISCV::BI__builtin_riscv_crc32_b:
18325 case RISCV::BI__builtin_riscv_crc32_h:
18326 case RISCV::BI__builtin_riscv_crc32_w:
18327 case RISCV::BI__builtin_riscv_crc32_d:
18328 case RISCV::BI__builtin_riscv_crc32c_b:
18329 case RISCV::BI__builtin_riscv_crc32c_h:
18330 case RISCV::BI__builtin_riscv_crc32c_w:
18331 case RISCV::BI__builtin_riscv_crc32c_d: {
18332 switch (BuiltinID) {
18333 default: llvm_unreachable("unexpected builtin ID")__builtin_unreachable();
18334 // Zbb
18335 case RISCV::BI__builtin_riscv_orc_b_32:
18336 case RISCV::BI__builtin_riscv_orc_b_64:
18337 ID = Intrinsic::riscv_orc_b;
18338 break;
18339
18340 // Zbc
18341 case RISCV::BI__builtin_riscv_clmul:
18342 ID = Intrinsic::riscv_clmul;
18343 break;
18344 case RISCV::BI__builtin_riscv_clmulh:
18345 ID = Intrinsic::riscv_clmulh;
18346 break;
18347 case RISCV::BI__builtin_riscv_clmulr:
18348 ID = Intrinsic::riscv_clmulr;
18349 break;
18350
18351 // Zbe
18352 case RISCV::BI__builtin_riscv_bcompress_32:
18353 case RISCV::BI__builtin_riscv_bcompress_64:
18354 ID = Intrinsic::riscv_bcompress;
18355 break;
18356 case RISCV::BI__builtin_riscv_bdecompress_32:
18357 case RISCV::BI__builtin_riscv_bdecompress_64:
18358 ID = Intrinsic::riscv_bdecompress;
18359 break;
18360
18361 // Zbp
18362 case RISCV::BI__builtin_riscv_grev_32:
18363 case RISCV::BI__builtin_riscv_grev_64:
18364 ID = Intrinsic::riscv_grev;
18365 break;
18366 case RISCV::BI__builtin_riscv_gorc_32:
18367 case RISCV::BI__builtin_riscv_gorc_64:
18368 ID = Intrinsic::riscv_gorc;
18369 break;
18370 case RISCV::BI__builtin_riscv_shfl_32:
18371 case RISCV::BI__builtin_riscv_shfl_64:
18372 ID = Intrinsic::riscv_shfl;
18373 break;
18374 case RISCV::BI__builtin_riscv_unshfl_32:
18375 case RISCV::BI__builtin_riscv_unshfl_64:
18376 ID = Intrinsic::riscv_unshfl;
18377 break;
18378 case RISCV::BI__builtin_riscv_xperm_n:
18379 ID = Intrinsic::riscv_xperm_n;
18380 break;
18381 case RISCV::BI__builtin_riscv_xperm_b:
18382 ID = Intrinsic::riscv_xperm_b;
18383 break;
18384 case RISCV::BI__builtin_riscv_xperm_h:
18385 ID = Intrinsic::riscv_xperm_h;
18386 break;
18387 case RISCV::BI__builtin_riscv_xperm_w:
18388 ID = Intrinsic::riscv_xperm_w;
18389 break;
18390
18391 // Zbr
18392 case RISCV::BI__builtin_riscv_crc32_b:
18393 ID = Intrinsic::riscv_crc32_b;
18394 break;
18395 case RISCV::BI__builtin_riscv_crc32_h:
18396 ID = Intrinsic::riscv_crc32_h;
18397 break;
18398 case RISCV::BI__builtin_riscv_crc32_w:
18399 ID = Intrinsic::riscv_crc32_w;
18400 break;
18401 case RISCV::BI__builtin_riscv_crc32_d:
18402 ID = Intrinsic::riscv_crc32_d;
18403 break;
18404 case RISCV::BI__builtin_riscv_crc32c_b:
18405 ID = Intrinsic::riscv_crc32c_b;
18406 break;
18407 case RISCV::BI__builtin_riscv_crc32c_h:
18408 ID = Intrinsic::riscv_crc32c_h;
18409 break;
18410 case RISCV::BI__builtin_riscv_crc32c_w:
18411 ID = Intrinsic::riscv_crc32c_w;
18412 break;
18413 case RISCV::BI__builtin_riscv_crc32c_d:
18414 ID = Intrinsic::riscv_crc32c_d;
18415 break;
18416 }
18417
18418 IntrinsicTypes = {ResultType};
18419 break;
18420 }
18421 // Vector builtins are handled from here.
18422#include "clang/Basic/riscv_vector_builtin_cg.inc"
18423 }
18424
18425 assert(ID != Intrinsic::not_intrinsic)((void)0);
18426
18427 llvm::Function *F = CGM.getIntrinsic(ID, IntrinsicTypes);
18428 return Builder.CreateCall(F, Ops, "");
18429}