File: | src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/lib/Transforms/Scalar/LICM.cpp |
Warning: | line 1163, column 11 Called C++ object pointer is null |
Press '?' to see keyboard shortcuts
Keyboard shortcuts:
1 | //===-- LICM.cpp - Loop Invariant Code Motion Pass ------------------------===// | ||||||||
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 pass performs loop invariant code motion, attempting to remove as much | ||||||||
10 | // code from the body of a loop as possible. It does this by either hoisting | ||||||||
11 | // code into the preheader block, or by sinking code to the exit blocks if it is | ||||||||
12 | // safe. This pass also promotes must-aliased memory locations in the loop to | ||||||||
13 | // live in registers, thus hoisting and sinking "invariant" loads and stores. | ||||||||
14 | // | ||||||||
15 | // Hoisting operations out of loops is a canonicalization transform. It | ||||||||
16 | // enables and simplifies subsequent optimizations in the middle-end. | ||||||||
17 | // Rematerialization of hoisted instructions to reduce register pressure is the | ||||||||
18 | // responsibility of the back-end, which has more accurate information about | ||||||||
19 | // register pressure and also handles other optimizations than LICM that | ||||||||
20 | // increase live-ranges. | ||||||||
21 | // | ||||||||
22 | // This pass uses alias analysis for two purposes: | ||||||||
23 | // | ||||||||
24 | // 1. Moving loop invariant loads and calls out of loops. If we can determine | ||||||||
25 | // that a load or call inside of a loop never aliases anything stored to, | ||||||||
26 | // we can hoist it or sink it like any other instruction. | ||||||||
27 | // 2. Scalar Promotion of Memory - If there is a store instruction inside of | ||||||||
28 | // the loop, we try to move the store to happen AFTER the loop instead of | ||||||||
29 | // inside of the loop. This can only happen if a few conditions are true: | ||||||||
30 | // A. The pointer stored through is loop invariant | ||||||||
31 | // B. There are no stores or loads in the loop which _may_ alias the | ||||||||
32 | // pointer. There are no calls in the loop which mod/ref the pointer. | ||||||||
33 | // If these conditions are true, we can promote the loads and stores in the | ||||||||
34 | // loop of the pointer to use a temporary alloca'd variable. We then use | ||||||||
35 | // the SSAUpdater to construct the appropriate SSA form for the value. | ||||||||
36 | // | ||||||||
37 | //===----------------------------------------------------------------------===// | ||||||||
38 | |||||||||
39 | #include "llvm/Transforms/Scalar/LICM.h" | ||||||||
40 | #include "llvm/ADT/SetOperations.h" | ||||||||
41 | #include "llvm/ADT/Statistic.h" | ||||||||
42 | #include "llvm/Analysis/AliasAnalysis.h" | ||||||||
43 | #include "llvm/Analysis/AliasSetTracker.h" | ||||||||
44 | #include "llvm/Analysis/BasicAliasAnalysis.h" | ||||||||
45 | #include "llvm/Analysis/BlockFrequencyInfo.h" | ||||||||
46 | #include "llvm/Analysis/CaptureTracking.h" | ||||||||
47 | #include "llvm/Analysis/ConstantFolding.h" | ||||||||
48 | #include "llvm/Analysis/GlobalsModRef.h" | ||||||||
49 | #include "llvm/Analysis/GuardUtils.h" | ||||||||
50 | #include "llvm/Analysis/LazyBlockFrequencyInfo.h" | ||||||||
51 | #include "llvm/Analysis/Loads.h" | ||||||||
52 | #include "llvm/Analysis/LoopInfo.h" | ||||||||
53 | #include "llvm/Analysis/LoopIterator.h" | ||||||||
54 | #include "llvm/Analysis/LoopPass.h" | ||||||||
55 | #include "llvm/Analysis/MemoryBuiltins.h" | ||||||||
56 | #include "llvm/Analysis/MemorySSA.h" | ||||||||
57 | #include "llvm/Analysis/MemorySSAUpdater.h" | ||||||||
58 | #include "llvm/Analysis/MustExecute.h" | ||||||||
59 | #include "llvm/Analysis/OptimizationRemarkEmitter.h" | ||||||||
60 | #include "llvm/Analysis/ScalarEvolution.h" | ||||||||
61 | #include "llvm/Analysis/ScalarEvolutionAliasAnalysis.h" | ||||||||
62 | #include "llvm/Analysis/TargetLibraryInfo.h" | ||||||||
63 | #include "llvm/Analysis/ValueTracking.h" | ||||||||
64 | #include "llvm/IR/CFG.h" | ||||||||
65 | #include "llvm/IR/Constants.h" | ||||||||
66 | #include "llvm/IR/DataLayout.h" | ||||||||
67 | #include "llvm/IR/DebugInfoMetadata.h" | ||||||||
68 | #include "llvm/IR/DerivedTypes.h" | ||||||||
69 | #include "llvm/IR/Dominators.h" | ||||||||
70 | #include "llvm/IR/Instructions.h" | ||||||||
71 | #include "llvm/IR/IntrinsicInst.h" | ||||||||
72 | #include "llvm/IR/LLVMContext.h" | ||||||||
73 | #include "llvm/IR/Metadata.h" | ||||||||
74 | #include "llvm/IR/PatternMatch.h" | ||||||||
75 | #include "llvm/IR/PredIteratorCache.h" | ||||||||
76 | #include "llvm/InitializePasses.h" | ||||||||
77 | #include "llvm/Support/CommandLine.h" | ||||||||
78 | #include "llvm/Support/Debug.h" | ||||||||
79 | #include "llvm/Support/raw_ostream.h" | ||||||||
80 | #include "llvm/Transforms/Scalar.h" | ||||||||
81 | #include "llvm/Transforms/Scalar/LoopPassManager.h" | ||||||||
82 | #include "llvm/Transforms/Utils/AssumeBundleBuilder.h" | ||||||||
83 | #include "llvm/Transforms/Utils/BasicBlockUtils.h" | ||||||||
84 | #include "llvm/Transforms/Utils/Local.h" | ||||||||
85 | #include "llvm/Transforms/Utils/LoopUtils.h" | ||||||||
86 | #include "llvm/Transforms/Utils/SSAUpdater.h" | ||||||||
87 | #include <algorithm> | ||||||||
88 | #include <utility> | ||||||||
89 | using namespace llvm; | ||||||||
90 | |||||||||
91 | #define DEBUG_TYPE"licm" "licm" | ||||||||
92 | |||||||||
93 | STATISTIC(NumCreatedBlocks, "Number of blocks created")static llvm::Statistic NumCreatedBlocks = {"licm", "NumCreatedBlocks" , "Number of blocks created"}; | ||||||||
94 | STATISTIC(NumClonedBranches, "Number of branches cloned")static llvm::Statistic NumClonedBranches = {"licm", "NumClonedBranches" , "Number of branches cloned"}; | ||||||||
95 | STATISTIC(NumSunk, "Number of instructions sunk out of loop")static llvm::Statistic NumSunk = {"licm", "NumSunk", "Number of instructions sunk out of loop" }; | ||||||||
96 | STATISTIC(NumHoisted, "Number of instructions hoisted out of loop")static llvm::Statistic NumHoisted = {"licm", "NumHoisted", "Number of instructions hoisted out of loop" }; | ||||||||
97 | STATISTIC(NumMovedLoads, "Number of load insts hoisted or sunk")static llvm::Statistic NumMovedLoads = {"licm", "NumMovedLoads" , "Number of load insts hoisted or sunk"}; | ||||||||
98 | STATISTIC(NumMovedCalls, "Number of call insts hoisted or sunk")static llvm::Statistic NumMovedCalls = {"licm", "NumMovedCalls" , "Number of call insts hoisted or sunk"}; | ||||||||
99 | STATISTIC(NumPromoted, "Number of memory locations promoted to registers")static llvm::Statistic NumPromoted = {"licm", "NumPromoted", "Number of memory locations promoted to registers" }; | ||||||||
100 | |||||||||
101 | /// Memory promotion is enabled by default. | ||||||||
102 | static cl::opt<bool> | ||||||||
103 | DisablePromotion("disable-licm-promotion", cl::Hidden, cl::init(false), | ||||||||
104 | cl::desc("Disable memory promotion in LICM pass")); | ||||||||
105 | |||||||||
106 | static cl::opt<bool> ControlFlowHoisting( | ||||||||
107 | "licm-control-flow-hoisting", cl::Hidden, cl::init(false), | ||||||||
108 | cl::desc("Enable control flow (and PHI) hoisting in LICM")); | ||||||||
109 | |||||||||
110 | static cl::opt<unsigned> HoistSinkColdnessThreshold( | ||||||||
111 | "licm-coldness-threshold", cl::Hidden, cl::init(4), | ||||||||
112 | cl::desc("Relative coldness Threshold of hoisting/sinking destination " | ||||||||
113 | "block for LICM to be considered beneficial")); | ||||||||
114 | |||||||||
115 | static cl::opt<uint32_t> MaxNumUsesTraversed( | ||||||||
116 | "licm-max-num-uses-traversed", cl::Hidden, cl::init(8), | ||||||||
117 | cl::desc("Max num uses visited for identifying load " | ||||||||
118 | "invariance in loop using invariant start (default = 8)")); | ||||||||
119 | |||||||||
120 | // Default value of zero implies we use the regular alias set tracker mechanism | ||||||||
121 | // instead of the cross product using AA to identify aliasing of the memory | ||||||||
122 | // location we are interested in. | ||||||||
123 | static cl::opt<int> | ||||||||
124 | LICMN2Theshold("licm-n2-threshold", cl::Hidden, cl::init(0), | ||||||||
125 | cl::desc("How many instruction to cross product using AA")); | ||||||||
126 | |||||||||
127 | // Experimental option to allow imprecision in LICM in pathological cases, in | ||||||||
128 | // exchange for faster compile. This is to be removed if MemorySSA starts to | ||||||||
129 | // address the same issue. This flag applies only when LICM uses MemorySSA | ||||||||
130 | // instead on AliasSetTracker. LICM calls MemorySSAWalker's | ||||||||
131 | // getClobberingMemoryAccess, up to the value of the Cap, getting perfect | ||||||||
132 | // accuracy. Afterwards, LICM will call into MemorySSA's getDefiningAccess, | ||||||||
133 | // which may not be precise, since optimizeUses is capped. The result is | ||||||||
134 | // correct, but we may not get as "far up" as possible to get which access is | ||||||||
135 | // clobbering the one queried. | ||||||||
136 | cl::opt<unsigned> llvm::SetLicmMssaOptCap( | ||||||||
137 | "licm-mssa-optimization-cap", cl::init(100), cl::Hidden, | ||||||||
138 | cl::desc("Enable imprecision in LICM in pathological cases, in exchange " | ||||||||
139 | "for faster compile. Caps the MemorySSA clobbering calls.")); | ||||||||
140 | |||||||||
141 | // Experimentally, memory promotion carries less importance than sinking and | ||||||||
142 | // hoisting. Limit when we do promotion when using MemorySSA, in order to save | ||||||||
143 | // compile time. | ||||||||
144 | cl::opt<unsigned> llvm::SetLicmMssaNoAccForPromotionCap( | ||||||||
145 | "licm-mssa-max-acc-promotion", cl::init(250), cl::Hidden, | ||||||||
146 | cl::desc("[LICM & MemorySSA] When MSSA in LICM is disabled, this has no " | ||||||||
147 | "effect. When MSSA in LICM is enabled, then this is the maximum " | ||||||||
148 | "number of accesses allowed to be present in a loop in order to " | ||||||||
149 | "enable memory promotion.")); | ||||||||
150 | |||||||||
151 | static bool inSubLoop(BasicBlock *BB, Loop *CurLoop, LoopInfo *LI); | ||||||||
152 | static bool isNotUsedOrFreeInLoop(const Instruction &I, const Loop *CurLoop, | ||||||||
153 | const LoopSafetyInfo *SafetyInfo, | ||||||||
154 | TargetTransformInfo *TTI, bool &FreeInLoop); | ||||||||
155 | static void hoist(Instruction &I, const DominatorTree *DT, const Loop *CurLoop, | ||||||||
156 | BasicBlock *Dest, ICFLoopSafetyInfo *SafetyInfo, | ||||||||
157 | MemorySSAUpdater *MSSAU, ScalarEvolution *SE, | ||||||||
158 | OptimizationRemarkEmitter *ORE); | ||||||||
159 | static bool sink(Instruction &I, LoopInfo *LI, DominatorTree *DT, | ||||||||
160 | BlockFrequencyInfo *BFI, const Loop *CurLoop, | ||||||||
161 | ICFLoopSafetyInfo *SafetyInfo, MemorySSAUpdater *MSSAU, | ||||||||
162 | OptimizationRemarkEmitter *ORE); | ||||||||
163 | static bool isSafeToExecuteUnconditionally(Instruction &Inst, | ||||||||
164 | const DominatorTree *DT, | ||||||||
165 | const TargetLibraryInfo *TLI, | ||||||||
166 | const Loop *CurLoop, | ||||||||
167 | const LoopSafetyInfo *SafetyInfo, | ||||||||
168 | OptimizationRemarkEmitter *ORE, | ||||||||
169 | const Instruction *CtxI = nullptr); | ||||||||
170 | static bool pointerInvalidatedByLoop(MemoryLocation MemLoc, | ||||||||
171 | AliasSetTracker *CurAST, Loop *CurLoop, | ||||||||
172 | AAResults *AA); | ||||||||
173 | static bool pointerInvalidatedByLoopWithMSSA(MemorySSA *MSSA, MemoryUse *MU, | ||||||||
174 | Loop *CurLoop, Instruction &I, | ||||||||
175 | SinkAndHoistLICMFlags &Flags); | ||||||||
176 | static bool pointerInvalidatedByBlockWithMSSA(BasicBlock &BB, MemorySSA &MSSA, | ||||||||
177 | MemoryUse &MU); | ||||||||
178 | static Instruction *cloneInstructionInExitBlock( | ||||||||
179 | Instruction &I, BasicBlock &ExitBlock, PHINode &PN, const LoopInfo *LI, | ||||||||
180 | const LoopSafetyInfo *SafetyInfo, MemorySSAUpdater *MSSAU); | ||||||||
181 | |||||||||
182 | static void eraseInstruction(Instruction &I, ICFLoopSafetyInfo &SafetyInfo, | ||||||||
183 | AliasSetTracker *AST, MemorySSAUpdater *MSSAU); | ||||||||
184 | |||||||||
185 | static void moveInstructionBefore(Instruction &I, Instruction &Dest, | ||||||||
186 | ICFLoopSafetyInfo &SafetyInfo, | ||||||||
187 | MemorySSAUpdater *MSSAU, ScalarEvolution *SE); | ||||||||
188 | |||||||||
189 | static void foreachMemoryAccess(MemorySSA *MSSA, Loop *L, | ||||||||
190 | function_ref<void(Instruction *)> Fn); | ||||||||
191 | static SmallVector<SmallSetVector<Value *, 8>, 0> | ||||||||
192 | collectPromotionCandidates(MemorySSA *MSSA, AliasAnalysis *AA, Loop *L); | ||||||||
193 | |||||||||
194 | namespace { | ||||||||
195 | struct LoopInvariantCodeMotion { | ||||||||
196 | bool runOnLoop(Loop *L, AAResults *AA, LoopInfo *LI, DominatorTree *DT, | ||||||||
197 | BlockFrequencyInfo *BFI, TargetLibraryInfo *TLI, | ||||||||
198 | TargetTransformInfo *TTI, ScalarEvolution *SE, MemorySSA *MSSA, | ||||||||
199 | OptimizationRemarkEmitter *ORE, bool LoopNestMode = false); | ||||||||
200 | |||||||||
201 | LoopInvariantCodeMotion(unsigned LicmMssaOptCap, | ||||||||
202 | unsigned LicmMssaNoAccForPromotionCap) | ||||||||
203 | : LicmMssaOptCap(LicmMssaOptCap), | ||||||||
204 | LicmMssaNoAccForPromotionCap(LicmMssaNoAccForPromotionCap) {} | ||||||||
205 | |||||||||
206 | private: | ||||||||
207 | unsigned LicmMssaOptCap; | ||||||||
208 | unsigned LicmMssaNoAccForPromotionCap; | ||||||||
209 | |||||||||
210 | std::unique_ptr<AliasSetTracker> | ||||||||
211 | collectAliasInfoForLoop(Loop *L, LoopInfo *LI, AAResults *AA); | ||||||||
212 | }; | ||||||||
213 | |||||||||
214 | struct LegacyLICMPass : public LoopPass { | ||||||||
215 | static char ID; // Pass identification, replacement for typeid | ||||||||
216 | LegacyLICMPass( | ||||||||
217 | unsigned LicmMssaOptCap = SetLicmMssaOptCap, | ||||||||
218 | unsigned LicmMssaNoAccForPromotionCap = SetLicmMssaNoAccForPromotionCap) | ||||||||
219 | : LoopPass(ID), LICM(LicmMssaOptCap, LicmMssaNoAccForPromotionCap) { | ||||||||
220 | initializeLegacyLICMPassPass(*PassRegistry::getPassRegistry()); | ||||||||
221 | } | ||||||||
222 | |||||||||
223 | bool runOnLoop(Loop *L, LPPassManager &LPM) override { | ||||||||
224 | if (skipLoop(L)) | ||||||||
225 | return false; | ||||||||
226 | |||||||||
227 | LLVM_DEBUG(dbgs() << "Perform LICM on Loop with header at block "do { } while (false) | ||||||||
228 | << L->getHeader()->getNameOrAsOperand() << "\n")do { } while (false); | ||||||||
229 | |||||||||
230 | auto *SE = getAnalysisIfAvailable<ScalarEvolutionWrapperPass>(); | ||||||||
231 | MemorySSA *MSSA = EnableMSSALoopDependency | ||||||||
232 | ? (&getAnalysis<MemorySSAWrapperPass>().getMSSA()) | ||||||||
233 | : nullptr; | ||||||||
234 | bool hasProfileData = L->getHeader()->getParent()->hasProfileData(); | ||||||||
235 | BlockFrequencyInfo *BFI = | ||||||||
236 | hasProfileData ? &getAnalysis<LazyBlockFrequencyInfoPass>().getBFI() | ||||||||
237 | : nullptr; | ||||||||
238 | // For the old PM, we can't use OptimizationRemarkEmitter as an analysis | ||||||||
239 | // pass. Function analyses need to be preserved across loop transformations | ||||||||
240 | // but ORE cannot be preserved (see comment before the pass definition). | ||||||||
241 | OptimizationRemarkEmitter ORE(L->getHeader()->getParent()); | ||||||||
242 | return LICM.runOnLoop( | ||||||||
243 | L, &getAnalysis<AAResultsWrapperPass>().getAAResults(), | ||||||||
244 | &getAnalysis<LoopInfoWrapperPass>().getLoopInfo(), | ||||||||
245 | &getAnalysis<DominatorTreeWrapperPass>().getDomTree(), BFI, | ||||||||
246 | &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI( | ||||||||
247 | *L->getHeader()->getParent()), | ||||||||
248 | &getAnalysis<TargetTransformInfoWrapperPass>().getTTI( | ||||||||
249 | *L->getHeader()->getParent()), | ||||||||
250 | SE ? &SE->getSE() : nullptr, MSSA, &ORE); | ||||||||
251 | } | ||||||||
252 | |||||||||
253 | /// This transformation requires natural loop information & requires that | ||||||||
254 | /// loop preheaders be inserted into the CFG... | ||||||||
255 | /// | ||||||||
256 | void getAnalysisUsage(AnalysisUsage &AU) const override { | ||||||||
257 | AU.addPreserved<DominatorTreeWrapperPass>(); | ||||||||
258 | AU.addPreserved<LoopInfoWrapperPass>(); | ||||||||
259 | AU.addRequired<TargetLibraryInfoWrapperPass>(); | ||||||||
260 | if (EnableMSSALoopDependency) { | ||||||||
261 | AU.addRequired<MemorySSAWrapperPass>(); | ||||||||
262 | AU.addPreserved<MemorySSAWrapperPass>(); | ||||||||
263 | } | ||||||||
264 | AU.addRequired<TargetTransformInfoWrapperPass>(); | ||||||||
265 | getLoopAnalysisUsage(AU); | ||||||||
266 | LazyBlockFrequencyInfoPass::getLazyBFIAnalysisUsage(AU); | ||||||||
267 | AU.addPreserved<LazyBlockFrequencyInfoPass>(); | ||||||||
268 | AU.addPreserved<LazyBranchProbabilityInfoPass>(); | ||||||||
269 | } | ||||||||
270 | |||||||||
271 | private: | ||||||||
272 | LoopInvariantCodeMotion LICM; | ||||||||
273 | }; | ||||||||
274 | } // namespace | ||||||||
275 | |||||||||
276 | PreservedAnalyses LICMPass::run(Loop &L, LoopAnalysisManager &AM, | ||||||||
277 | LoopStandardAnalysisResults &AR, LPMUpdater &) { | ||||||||
278 | // For the new PM, we also can't use OptimizationRemarkEmitter as an analysis | ||||||||
279 | // pass. Function analyses need to be preserved across loop transformations | ||||||||
280 | // but ORE cannot be preserved (see comment before the pass definition). | ||||||||
281 | OptimizationRemarkEmitter ORE(L.getHeader()->getParent()); | ||||||||
282 | |||||||||
283 | LoopInvariantCodeMotion LICM(LicmMssaOptCap, LicmMssaNoAccForPromotionCap); | ||||||||
284 | if (!LICM.runOnLoop(&L, &AR.AA, &AR.LI, &AR.DT, AR.BFI, &AR.TLI, &AR.TTI, | ||||||||
285 | &AR.SE, AR.MSSA, &ORE)) | ||||||||
286 | return PreservedAnalyses::all(); | ||||||||
287 | |||||||||
288 | auto PA = getLoopPassPreservedAnalyses(); | ||||||||
289 | |||||||||
290 | PA.preserve<DominatorTreeAnalysis>(); | ||||||||
291 | PA.preserve<LoopAnalysis>(); | ||||||||
292 | if (AR.MSSA) | ||||||||
293 | PA.preserve<MemorySSAAnalysis>(); | ||||||||
294 | |||||||||
295 | return PA; | ||||||||
296 | } | ||||||||
297 | |||||||||
298 | PreservedAnalyses LNICMPass::run(LoopNest &LN, LoopAnalysisManager &AM, | ||||||||
299 | LoopStandardAnalysisResults &AR, | ||||||||
300 | LPMUpdater &) { | ||||||||
301 | // For the new PM, we also can't use OptimizationRemarkEmitter as an analysis | ||||||||
302 | // pass. Function analyses need to be preserved across loop transformations | ||||||||
303 | // but ORE cannot be preserved (see comment before the pass definition). | ||||||||
304 | OptimizationRemarkEmitter ORE(LN.getParent()); | ||||||||
305 | |||||||||
306 | LoopInvariantCodeMotion LICM(LicmMssaOptCap, LicmMssaNoAccForPromotionCap); | ||||||||
307 | |||||||||
308 | Loop &OutermostLoop = LN.getOutermostLoop(); | ||||||||
309 | bool Changed = LICM.runOnLoop(&OutermostLoop, &AR.AA, &AR.LI, &AR.DT, AR.BFI, | ||||||||
310 | &AR.TLI, &AR.TTI, &AR.SE, AR.MSSA, &ORE, true); | ||||||||
311 | |||||||||
312 | if (!Changed) | ||||||||
313 | return PreservedAnalyses::all(); | ||||||||
314 | |||||||||
315 | auto PA = getLoopPassPreservedAnalyses(); | ||||||||
316 | |||||||||
317 | PA.preserve<DominatorTreeAnalysis>(); | ||||||||
318 | PA.preserve<LoopAnalysis>(); | ||||||||
319 | if (AR.MSSA) | ||||||||
320 | PA.preserve<MemorySSAAnalysis>(); | ||||||||
321 | |||||||||
322 | return PA; | ||||||||
323 | } | ||||||||
324 | |||||||||
325 | char LegacyLICMPass::ID = 0; | ||||||||
326 | INITIALIZE_PASS_BEGIN(LegacyLICMPass, "licm", "Loop Invariant Code Motion",static void *initializeLegacyLICMPassPassOnce(PassRegistry & Registry) { | ||||||||
327 | false, false)static void *initializeLegacyLICMPassPassOnce(PassRegistry & Registry) { | ||||||||
328 | INITIALIZE_PASS_DEPENDENCY(LoopPass)initializeLoopPassPass(Registry); | ||||||||
329 | INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)initializeTargetLibraryInfoWrapperPassPass(Registry); | ||||||||
330 | INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)initializeTargetTransformInfoWrapperPassPass(Registry); | ||||||||
331 | INITIALIZE_PASS_DEPENDENCY(MemorySSAWrapperPass)initializeMemorySSAWrapperPassPass(Registry); | ||||||||
332 | INITIALIZE_PASS_DEPENDENCY(LazyBFIPass)initializeLazyBFIPassPass(Registry); | ||||||||
333 | INITIALIZE_PASS_END(LegacyLICMPass, "licm", "Loop Invariant Code Motion", false,PassInfo *PI = new PassInfo( "Loop Invariant Code Motion", "licm" , &LegacyLICMPass::ID, PassInfo::NormalCtor_t(callDefaultCtor <LegacyLICMPass>), false, false); Registry.registerPass (*PI, true); return PI; } static llvm::once_flag InitializeLegacyLICMPassPassFlag ; void llvm::initializeLegacyLICMPassPass(PassRegistry &Registry ) { llvm::call_once(InitializeLegacyLICMPassPassFlag, initializeLegacyLICMPassPassOnce , std::ref(Registry)); } | ||||||||
334 | false)PassInfo *PI = new PassInfo( "Loop Invariant Code Motion", "licm" , &LegacyLICMPass::ID, PassInfo::NormalCtor_t(callDefaultCtor <LegacyLICMPass>), false, false); Registry.registerPass (*PI, true); return PI; } static llvm::once_flag InitializeLegacyLICMPassPassFlag ; void llvm::initializeLegacyLICMPassPass(PassRegistry &Registry ) { llvm::call_once(InitializeLegacyLICMPassPassFlag, initializeLegacyLICMPassPassOnce , std::ref(Registry)); } | ||||||||
335 | |||||||||
336 | Pass *llvm::createLICMPass() { return new LegacyLICMPass(); } | ||||||||
337 | Pass *llvm::createLICMPass(unsigned LicmMssaOptCap, | ||||||||
338 | unsigned LicmMssaNoAccForPromotionCap) { | ||||||||
339 | return new LegacyLICMPass(LicmMssaOptCap, LicmMssaNoAccForPromotionCap); | ||||||||
340 | } | ||||||||
341 | |||||||||
342 | llvm::SinkAndHoistLICMFlags::SinkAndHoistLICMFlags(bool IsSink, Loop *L, | ||||||||
343 | MemorySSA *MSSA) | ||||||||
344 | : SinkAndHoistLICMFlags(SetLicmMssaOptCap, SetLicmMssaNoAccForPromotionCap, | ||||||||
345 | IsSink, L, MSSA) {} | ||||||||
346 | |||||||||
347 | llvm::SinkAndHoistLICMFlags::SinkAndHoistLICMFlags( | ||||||||
348 | unsigned LicmMssaOptCap, unsigned LicmMssaNoAccForPromotionCap, bool IsSink, | ||||||||
349 | Loop *L, MemorySSA *MSSA) | ||||||||
350 | : LicmMssaOptCap(LicmMssaOptCap), | ||||||||
351 | LicmMssaNoAccForPromotionCap(LicmMssaNoAccForPromotionCap), | ||||||||
352 | IsSink(IsSink) { | ||||||||
353 | assert(((L != nullptr) == (MSSA != nullptr)) &&((void)0) | ||||||||
354 | "Unexpected values for SinkAndHoistLICMFlags")((void)0); | ||||||||
355 | if (!MSSA) | ||||||||
356 | return; | ||||||||
357 | |||||||||
358 | unsigned AccessCapCount = 0; | ||||||||
359 | for (auto *BB : L->getBlocks()) | ||||||||
360 | if (const auto *Accesses = MSSA->getBlockAccesses(BB)) | ||||||||
361 | for (const auto &MA : *Accesses) { | ||||||||
362 | (void)MA; | ||||||||
363 | ++AccessCapCount; | ||||||||
364 | if (AccessCapCount > LicmMssaNoAccForPromotionCap) { | ||||||||
365 | NoOfMemAccTooLarge = true; | ||||||||
366 | return; | ||||||||
367 | } | ||||||||
368 | } | ||||||||
369 | } | ||||||||
370 | |||||||||
371 | /// Hoist expressions out of the specified loop. Note, alias info for inner | ||||||||
372 | /// loop is not preserved so it is not a good idea to run LICM multiple | ||||||||
373 | /// times on one loop. | ||||||||
374 | bool LoopInvariantCodeMotion::runOnLoop( | ||||||||
375 | Loop *L, AAResults *AA, LoopInfo *LI, DominatorTree *DT, | ||||||||
376 | BlockFrequencyInfo *BFI, TargetLibraryInfo *TLI, TargetTransformInfo *TTI, | ||||||||
377 | ScalarEvolution *SE, MemorySSA *MSSA, OptimizationRemarkEmitter *ORE, | ||||||||
378 | bool LoopNestMode) { | ||||||||
379 | bool Changed = false; | ||||||||
380 | |||||||||
381 | assert(L->isLCSSAForm(*DT) && "Loop is not in LCSSA form.")((void)0); | ||||||||
382 | |||||||||
383 | // If this loop has metadata indicating that LICM is not to be performed then | ||||||||
384 | // just exit. | ||||||||
385 | if (hasDisableLICMTransformsHint(L)) { | ||||||||
386 | return false; | ||||||||
387 | } | ||||||||
388 | |||||||||
389 | std::unique_ptr<AliasSetTracker> CurAST; | ||||||||
390 | std::unique_ptr<MemorySSAUpdater> MSSAU; | ||||||||
391 | std::unique_ptr<SinkAndHoistLICMFlags> Flags; | ||||||||
392 | |||||||||
393 | // Don't sink stores from loops with coroutine suspend instructions. | ||||||||
394 | // LICM would sink instructions into the default destination of | ||||||||
395 | // the coroutine switch. The default destination of the switch is to | ||||||||
396 | // handle the case where the coroutine is suspended, by which point the | ||||||||
397 | // coroutine frame may have been destroyed. No instruction can be sunk there. | ||||||||
398 | // FIXME: This would unfortunately hurt the performance of coroutines, however | ||||||||
399 | // there is currently no general solution for this. Similar issues could also | ||||||||
400 | // potentially happen in other passes where instructions are being moved | ||||||||
401 | // across that edge. | ||||||||
402 | bool HasCoroSuspendInst = llvm::any_of(L->getBlocks(), [](BasicBlock *BB) { | ||||||||
403 | return llvm::any_of(*BB, [](Instruction &I) { | ||||||||
404 | IntrinsicInst *II = dyn_cast<IntrinsicInst>(&I); | ||||||||
405 | return II && II->getIntrinsicID() == Intrinsic::coro_suspend; | ||||||||
406 | }); | ||||||||
407 | }); | ||||||||
408 | |||||||||
409 | if (!MSSA) { | ||||||||
410 | LLVM_DEBUG(dbgs() << "LICM: Using Alias Set Tracker.\n")do { } while (false); | ||||||||
411 | CurAST = collectAliasInfoForLoop(L, LI, AA); | ||||||||
412 | Flags = std::make_unique<SinkAndHoistLICMFlags>( | ||||||||
413 | LicmMssaOptCap, LicmMssaNoAccForPromotionCap, /*IsSink=*/true); | ||||||||
414 | } else { | ||||||||
415 | LLVM_DEBUG(dbgs() << "LICM: Using MemorySSA.\n")do { } while (false); | ||||||||
416 | MSSAU = std::make_unique<MemorySSAUpdater>(MSSA); | ||||||||
417 | Flags = std::make_unique<SinkAndHoistLICMFlags>( | ||||||||
418 | LicmMssaOptCap, LicmMssaNoAccForPromotionCap, /*IsSink=*/true, L, MSSA); | ||||||||
419 | } | ||||||||
420 | |||||||||
421 | // Get the preheader block to move instructions into... | ||||||||
422 | BasicBlock *Preheader = L->getLoopPreheader(); | ||||||||
423 | |||||||||
424 | // Compute loop safety information. | ||||||||
425 | ICFLoopSafetyInfo SafetyInfo; | ||||||||
426 | SafetyInfo.computeLoopSafetyInfo(L); | ||||||||
427 | |||||||||
428 | // We want to visit all of the instructions in this loop... that are not parts | ||||||||
429 | // of our subloops (they have already had their invariants hoisted out of | ||||||||
430 | // their loop, into this loop, so there is no need to process the BODIES of | ||||||||
431 | // the subloops). | ||||||||
432 | // | ||||||||
433 | // Traverse the body of the loop in depth first order on the dominator tree so | ||||||||
434 | // that we are guaranteed to see definitions before we see uses. This allows | ||||||||
435 | // us to sink instructions in one pass, without iteration. After sinking | ||||||||
436 | // instructions, we perform another pass to hoist them out of the loop. | ||||||||
437 | if (L->hasDedicatedExits()) | ||||||||
438 | Changed |= | ||||||||
439 | sinkRegion(DT->getNode(L->getHeader()), AA, LI, DT, BFI, TLI, TTI, L, | ||||||||
440 | CurAST.get(), MSSAU.get(), &SafetyInfo, *Flags.get(), ORE); | ||||||||
441 | Flags->setIsSink(false); | ||||||||
442 | if (Preheader) | ||||||||
443 | Changed |= hoistRegion(DT->getNode(L->getHeader()), AA, LI, DT, BFI, TLI, L, | ||||||||
444 | CurAST.get(), MSSAU.get(), SE, &SafetyInfo, | ||||||||
445 | *Flags.get(), ORE, LoopNestMode); | ||||||||
446 | |||||||||
447 | // Now that all loop invariants have been removed from the loop, promote any | ||||||||
448 | // memory references to scalars that we can. | ||||||||
449 | // Don't sink stores from loops without dedicated block exits. Exits | ||||||||
450 | // containing indirect branches are not transformed by loop simplify, | ||||||||
451 | // make sure we catch that. An additional load may be generated in the | ||||||||
452 | // preheader for SSA updater, so also avoid sinking when no preheader | ||||||||
453 | // is available. | ||||||||
454 | if (!DisablePromotion && Preheader && L->hasDedicatedExits() && | ||||||||
455 | !Flags->tooManyMemoryAccesses() && !HasCoroSuspendInst) { | ||||||||
456 | // Figure out the loop exits and their insertion points | ||||||||
457 | SmallVector<BasicBlock *, 8> ExitBlocks; | ||||||||
458 | L->getUniqueExitBlocks(ExitBlocks); | ||||||||
459 | |||||||||
460 | // We can't insert into a catchswitch. | ||||||||
461 | bool HasCatchSwitch = llvm::any_of(ExitBlocks, [](BasicBlock *Exit) { | ||||||||
462 | return isa<CatchSwitchInst>(Exit->getTerminator()); | ||||||||
463 | }); | ||||||||
464 | |||||||||
465 | if (!HasCatchSwitch) { | ||||||||
466 | SmallVector<Instruction *, 8> InsertPts; | ||||||||
467 | SmallVector<MemoryAccess *, 8> MSSAInsertPts; | ||||||||
468 | InsertPts.reserve(ExitBlocks.size()); | ||||||||
469 | if (MSSAU) | ||||||||
470 | MSSAInsertPts.reserve(ExitBlocks.size()); | ||||||||
471 | for (BasicBlock *ExitBlock : ExitBlocks) { | ||||||||
472 | InsertPts.push_back(&*ExitBlock->getFirstInsertionPt()); | ||||||||
473 | if (MSSAU) | ||||||||
474 | MSSAInsertPts.push_back(nullptr); | ||||||||
475 | } | ||||||||
476 | |||||||||
477 | PredIteratorCache PIC; | ||||||||
478 | |||||||||
479 | bool Promoted = false; | ||||||||
480 | if (CurAST.get()) { | ||||||||
481 | // Loop over all of the alias sets in the tracker object. | ||||||||
482 | for (AliasSet &AS : *CurAST) { | ||||||||
483 | // We can promote this alias set if it has a store, if it is a "Must" | ||||||||
484 | // alias set, if the pointer is loop invariant, and if we are not | ||||||||
485 | // eliminating any volatile loads or stores. | ||||||||
486 | if (AS.isForwardingAliasSet() || !AS.isMod() || !AS.isMustAlias() || | ||||||||
487 | !L->isLoopInvariant(AS.begin()->getValue())) | ||||||||
488 | continue; | ||||||||
489 | |||||||||
490 | assert(((void)0) | ||||||||
491 | !AS.empty() &&((void)0) | ||||||||
492 | "Must alias set should have at least one pointer element in it!")((void)0); | ||||||||
493 | |||||||||
494 | SmallSetVector<Value *, 8> PointerMustAliases; | ||||||||
495 | for (const auto &ASI : AS) | ||||||||
496 | PointerMustAliases.insert(ASI.getValue()); | ||||||||
497 | |||||||||
498 | Promoted |= promoteLoopAccessesToScalars( | ||||||||
499 | PointerMustAliases, ExitBlocks, InsertPts, MSSAInsertPts, PIC, LI, | ||||||||
500 | DT, TLI, L, CurAST.get(), MSSAU.get(), &SafetyInfo, ORE); | ||||||||
501 | } | ||||||||
502 | } else { | ||||||||
503 | // Promoting one set of accesses may make the pointers for another set | ||||||||
504 | // loop invariant, so run this in a loop (with the MaybePromotable set | ||||||||
505 | // decreasing in size over time). | ||||||||
506 | bool LocalPromoted; | ||||||||
507 | do { | ||||||||
508 | LocalPromoted = false; | ||||||||
509 | for (const SmallSetVector<Value *, 8> &PointerMustAliases : | ||||||||
510 | collectPromotionCandidates(MSSA, AA, L)) { | ||||||||
511 | LocalPromoted |= promoteLoopAccessesToScalars( | ||||||||
512 | PointerMustAliases, ExitBlocks, InsertPts, MSSAInsertPts, PIC, | ||||||||
513 | LI, DT, TLI, L, /*AST*/nullptr, MSSAU.get(), &SafetyInfo, ORE); | ||||||||
514 | } | ||||||||
515 | Promoted |= LocalPromoted; | ||||||||
516 | } while (LocalPromoted); | ||||||||
517 | } | ||||||||
518 | |||||||||
519 | // Once we have promoted values across the loop body we have to | ||||||||
520 | // recursively reform LCSSA as any nested loop may now have values defined | ||||||||
521 | // within the loop used in the outer loop. | ||||||||
522 | // FIXME: This is really heavy handed. It would be a bit better to use an | ||||||||
523 | // SSAUpdater strategy during promotion that was LCSSA aware and reformed | ||||||||
524 | // it as it went. | ||||||||
525 | if (Promoted) | ||||||||
526 | formLCSSARecursively(*L, *DT, LI, SE); | ||||||||
527 | |||||||||
528 | Changed |= Promoted; | ||||||||
529 | } | ||||||||
530 | } | ||||||||
531 | |||||||||
532 | // Check that neither this loop nor its parent have had LCSSA broken. LICM is | ||||||||
533 | // specifically moving instructions across the loop boundary and so it is | ||||||||
534 | // especially in need of sanity checking here. | ||||||||
535 | assert(L->isLCSSAForm(*DT) && "Loop not left in LCSSA form after LICM!")((void)0); | ||||||||
536 | assert((L->isOutermost() || L->getParentLoop()->isLCSSAForm(*DT)) &&((void)0) | ||||||||
537 | "Parent loop not left in LCSSA form after LICM!")((void)0); | ||||||||
538 | |||||||||
539 | if (MSSAU.get() && VerifyMemorySSA) | ||||||||
540 | MSSAU->getMemorySSA()->verifyMemorySSA(); | ||||||||
541 | |||||||||
542 | if (Changed && SE) | ||||||||
543 | SE->forgetLoopDispositions(L); | ||||||||
544 | return Changed; | ||||||||
545 | } | ||||||||
546 | |||||||||
547 | /// Walk the specified region of the CFG (defined by all blocks dominated by | ||||||||
548 | /// the specified block, and that are in the current loop) in reverse depth | ||||||||
549 | /// first order w.r.t the DominatorTree. This allows us to visit uses before | ||||||||
550 | /// definitions, allowing us to sink a loop body in one pass without iteration. | ||||||||
551 | /// | ||||||||
552 | bool llvm::sinkRegion(DomTreeNode *N, AAResults *AA, LoopInfo *LI, | ||||||||
553 | DominatorTree *DT, BlockFrequencyInfo *BFI, | ||||||||
554 | TargetLibraryInfo *TLI, TargetTransformInfo *TTI, | ||||||||
555 | Loop *CurLoop, AliasSetTracker *CurAST, | ||||||||
556 | MemorySSAUpdater *MSSAU, ICFLoopSafetyInfo *SafetyInfo, | ||||||||
557 | SinkAndHoistLICMFlags &Flags, | ||||||||
558 | OptimizationRemarkEmitter *ORE) { | ||||||||
559 | |||||||||
560 | // Verify inputs. | ||||||||
561 | assert(N != nullptr && AA != nullptr && LI != nullptr && DT != nullptr &&((void)0) | ||||||||
562 | CurLoop != nullptr && SafetyInfo != nullptr &&((void)0) | ||||||||
563 | "Unexpected input to sinkRegion.")((void)0); | ||||||||
564 | assert(((CurAST != nullptr) ^ (MSSAU != nullptr)) &&((void)0) | ||||||||
565 | "Either AliasSetTracker or MemorySSA should be initialized.")((void)0); | ||||||||
566 | |||||||||
567 | // We want to visit children before parents. We will enque all the parents | ||||||||
568 | // before their children in the worklist and process the worklist in reverse | ||||||||
569 | // order. | ||||||||
570 | SmallVector<DomTreeNode *, 16> Worklist = collectChildrenInLoop(N, CurLoop); | ||||||||
571 | |||||||||
572 | bool Changed = false; | ||||||||
573 | for (DomTreeNode *DTN : reverse(Worklist)) { | ||||||||
574 | BasicBlock *BB = DTN->getBlock(); | ||||||||
575 | // Only need to process the contents of this block if it is not part of a | ||||||||
576 | // subloop (which would already have been processed). | ||||||||
577 | if (inSubLoop(BB, CurLoop, LI)) | ||||||||
578 | continue; | ||||||||
579 | |||||||||
580 | for (BasicBlock::iterator II = BB->end(); II != BB->begin();) { | ||||||||
581 | Instruction &I = *--II; | ||||||||
582 | |||||||||
583 | // The instruction is not used in the loop if it is dead. In this case, | ||||||||
584 | // we just delete it instead of sinking it. | ||||||||
585 | if (isInstructionTriviallyDead(&I, TLI)) { | ||||||||
586 | LLVM_DEBUG(dbgs() << "LICM deleting dead inst: " << I << '\n')do { } while (false); | ||||||||
587 | salvageKnowledge(&I); | ||||||||
588 | salvageDebugInfo(I); | ||||||||
589 | ++II; | ||||||||
590 | eraseInstruction(I, *SafetyInfo, CurAST, MSSAU); | ||||||||
591 | Changed = true; | ||||||||
592 | continue; | ||||||||
593 | } | ||||||||
594 | |||||||||
595 | // Check to see if we can sink this instruction to the exit blocks | ||||||||
596 | // of the loop. We can do this if the all users of the instruction are | ||||||||
597 | // outside of the loop. In this case, it doesn't even matter if the | ||||||||
598 | // operands of the instruction are loop invariant. | ||||||||
599 | // | ||||||||
600 | bool FreeInLoop = false; | ||||||||
601 | if (!I.mayHaveSideEffects() && | ||||||||
602 | isNotUsedOrFreeInLoop(I, CurLoop, SafetyInfo, TTI, FreeInLoop) && | ||||||||
603 | canSinkOrHoistInst(I, AA, DT, CurLoop, CurAST, MSSAU, true, &Flags, | ||||||||
604 | ORE)) { | ||||||||
605 | if (sink(I, LI, DT, BFI, CurLoop, SafetyInfo, MSSAU, ORE)) { | ||||||||
606 | if (!FreeInLoop) { | ||||||||
607 | ++II; | ||||||||
608 | salvageDebugInfo(I); | ||||||||
609 | eraseInstruction(I, *SafetyInfo, CurAST, MSSAU); | ||||||||
610 | } | ||||||||
611 | Changed = true; | ||||||||
612 | } | ||||||||
613 | } | ||||||||
614 | } | ||||||||
615 | } | ||||||||
616 | if (MSSAU && VerifyMemorySSA) | ||||||||
617 | MSSAU->getMemorySSA()->verifyMemorySSA(); | ||||||||
618 | return Changed; | ||||||||
619 | } | ||||||||
620 | |||||||||
621 | namespace { | ||||||||
622 | // This is a helper class for hoistRegion to make it able to hoist control flow | ||||||||
623 | // in order to be able to hoist phis. The way this works is that we initially | ||||||||
624 | // start hoisting to the loop preheader, and when we see a loop invariant branch | ||||||||
625 | // we make note of this. When we then come to hoist an instruction that's | ||||||||
626 | // conditional on such a branch we duplicate the branch and the relevant control | ||||||||
627 | // flow, then hoist the instruction into the block corresponding to its original | ||||||||
628 | // block in the duplicated control flow. | ||||||||
629 | class ControlFlowHoister { | ||||||||
630 | private: | ||||||||
631 | // Information about the loop we are hoisting from | ||||||||
632 | LoopInfo *LI; | ||||||||
633 | DominatorTree *DT; | ||||||||
634 | Loop *CurLoop; | ||||||||
635 | MemorySSAUpdater *MSSAU; | ||||||||
636 | |||||||||
637 | // A map of blocks in the loop to the block their instructions will be hoisted | ||||||||
638 | // to. | ||||||||
639 | DenseMap<BasicBlock *, BasicBlock *> HoistDestinationMap; | ||||||||
640 | |||||||||
641 | // The branches that we can hoist, mapped to the block that marks a | ||||||||
642 | // convergence point of their control flow. | ||||||||
643 | DenseMap<BranchInst *, BasicBlock *> HoistableBranches; | ||||||||
644 | |||||||||
645 | public: | ||||||||
646 | ControlFlowHoister(LoopInfo *LI, DominatorTree *DT, Loop *CurLoop, | ||||||||
647 | MemorySSAUpdater *MSSAU) | ||||||||
648 | : LI(LI), DT(DT), CurLoop(CurLoop), MSSAU(MSSAU) {} | ||||||||
649 | |||||||||
650 | void registerPossiblyHoistableBranch(BranchInst *BI) { | ||||||||
651 | // We can only hoist conditional branches with loop invariant operands. | ||||||||
652 | if (!ControlFlowHoisting || !BI->isConditional() || | ||||||||
653 | !CurLoop->hasLoopInvariantOperands(BI)) | ||||||||
654 | return; | ||||||||
655 | |||||||||
656 | // The branch destinations need to be in the loop, and we don't gain | ||||||||
657 | // anything by duplicating conditional branches with duplicate successors, | ||||||||
658 | // as it's essentially the same as an unconditional branch. | ||||||||
659 | BasicBlock *TrueDest = BI->getSuccessor(0); | ||||||||
660 | BasicBlock *FalseDest = BI->getSuccessor(1); | ||||||||
661 | if (!CurLoop->contains(TrueDest) || !CurLoop->contains(FalseDest) || | ||||||||
662 | TrueDest == FalseDest) | ||||||||
663 | return; | ||||||||
664 | |||||||||
665 | // We can hoist BI if one branch destination is the successor of the other, | ||||||||
666 | // or both have common successor which we check by seeing if the | ||||||||
667 | // intersection of their successors is non-empty. | ||||||||
668 | // TODO: This could be expanded to allowing branches where both ends | ||||||||
669 | // eventually converge to a single block. | ||||||||
670 | SmallPtrSet<BasicBlock *, 4> TrueDestSucc, FalseDestSucc; | ||||||||
671 | TrueDestSucc.insert(succ_begin(TrueDest), succ_end(TrueDest)); | ||||||||
672 | FalseDestSucc.insert(succ_begin(FalseDest), succ_end(FalseDest)); | ||||||||
673 | BasicBlock *CommonSucc = nullptr; | ||||||||
674 | if (TrueDestSucc.count(FalseDest)) { | ||||||||
675 | CommonSucc = FalseDest; | ||||||||
676 | } else if (FalseDestSucc.count(TrueDest)) { | ||||||||
677 | CommonSucc = TrueDest; | ||||||||
678 | } else { | ||||||||
679 | set_intersect(TrueDestSucc, FalseDestSucc); | ||||||||
680 | // If there's one common successor use that. | ||||||||
681 | if (TrueDestSucc.size() == 1) | ||||||||
682 | CommonSucc = *TrueDestSucc.begin(); | ||||||||
683 | // If there's more than one pick whichever appears first in the block list | ||||||||
684 | // (we can't use the value returned by TrueDestSucc.begin() as it's | ||||||||
685 | // unpredicatable which element gets returned). | ||||||||
686 | else if (!TrueDestSucc.empty()) { | ||||||||
687 | Function *F = TrueDest->getParent(); | ||||||||
688 | auto IsSucc = [&](BasicBlock &BB) { return TrueDestSucc.count(&BB); }; | ||||||||
689 | auto It = llvm::find_if(*F, IsSucc); | ||||||||
690 | assert(It != F->end() && "Could not find successor in function")((void)0); | ||||||||
691 | CommonSucc = &*It; | ||||||||
692 | } | ||||||||
693 | } | ||||||||
694 | // The common successor has to be dominated by the branch, as otherwise | ||||||||
695 | // there will be some other path to the successor that will not be | ||||||||
696 | // controlled by this branch so any phi we hoist would be controlled by the | ||||||||
697 | // wrong condition. This also takes care of avoiding hoisting of loop back | ||||||||
698 | // edges. | ||||||||
699 | // TODO: In some cases this could be relaxed if the successor is dominated | ||||||||
700 | // by another block that's been hoisted and we can guarantee that the | ||||||||
701 | // control flow has been replicated exactly. | ||||||||
702 | if (CommonSucc && DT->dominates(BI, CommonSucc)) | ||||||||
703 | HoistableBranches[BI] = CommonSucc; | ||||||||
704 | } | ||||||||
705 | |||||||||
706 | bool canHoistPHI(PHINode *PN) { | ||||||||
707 | // The phi must have loop invariant operands. | ||||||||
708 | if (!ControlFlowHoisting || !CurLoop->hasLoopInvariantOperands(PN)) | ||||||||
709 | return false; | ||||||||
710 | // We can hoist phis if the block they are in is the target of hoistable | ||||||||
711 | // branches which cover all of the predecessors of the block. | ||||||||
712 | SmallPtrSet<BasicBlock *, 8> PredecessorBlocks; | ||||||||
713 | BasicBlock *BB = PN->getParent(); | ||||||||
714 | for (BasicBlock *PredBB : predecessors(BB)) | ||||||||
715 | PredecessorBlocks.insert(PredBB); | ||||||||
716 | // If we have less predecessor blocks than predecessors then the phi will | ||||||||
717 | // have more than one incoming value for the same block which we can't | ||||||||
718 | // handle. | ||||||||
719 | // TODO: This could be handled be erasing some of the duplicate incoming | ||||||||
720 | // values. | ||||||||
721 | if (PredecessorBlocks.size() != pred_size(BB)) | ||||||||
722 | return false; | ||||||||
723 | for (auto &Pair : HoistableBranches) { | ||||||||
724 | if (Pair.second == BB) { | ||||||||
725 | // Which blocks are predecessors via this branch depends on if the | ||||||||
726 | // branch is triangle-like or diamond-like. | ||||||||
727 | if (Pair.first->getSuccessor(0) == BB) { | ||||||||
728 | PredecessorBlocks.erase(Pair.first->getParent()); | ||||||||
729 | PredecessorBlocks.erase(Pair.first->getSuccessor(1)); | ||||||||
730 | } else if (Pair.first->getSuccessor(1) == BB) { | ||||||||
731 | PredecessorBlocks.erase(Pair.first->getParent()); | ||||||||
732 | PredecessorBlocks.erase(Pair.first->getSuccessor(0)); | ||||||||
733 | } else { | ||||||||
734 | PredecessorBlocks.erase(Pair.first->getSuccessor(0)); | ||||||||
735 | PredecessorBlocks.erase(Pair.first->getSuccessor(1)); | ||||||||
736 | } | ||||||||
737 | } | ||||||||
738 | } | ||||||||
739 | // PredecessorBlocks will now be empty if for every predecessor of BB we | ||||||||
740 | // found a hoistable branch source. | ||||||||
741 | return PredecessorBlocks.empty(); | ||||||||
742 | } | ||||||||
743 | |||||||||
744 | BasicBlock *getOrCreateHoistedBlock(BasicBlock *BB) { | ||||||||
745 | if (!ControlFlowHoisting) | ||||||||
746 | return CurLoop->getLoopPreheader(); | ||||||||
747 | // If BB has already been hoisted, return that | ||||||||
748 | if (HoistDestinationMap.count(BB)) | ||||||||
749 | return HoistDestinationMap[BB]; | ||||||||
750 | |||||||||
751 | // Check if this block is conditional based on a pending branch | ||||||||
752 | auto HasBBAsSuccessor = | ||||||||
753 | [&](DenseMap<BranchInst *, BasicBlock *>::value_type &Pair) { | ||||||||
754 | return BB != Pair.second && (Pair.first->getSuccessor(0) == BB || | ||||||||
755 | Pair.first->getSuccessor(1) == BB); | ||||||||
756 | }; | ||||||||
757 | auto It = llvm::find_if(HoistableBranches, HasBBAsSuccessor); | ||||||||
758 | |||||||||
759 | // If not involved in a pending branch, hoist to preheader | ||||||||
760 | BasicBlock *InitialPreheader = CurLoop->getLoopPreheader(); | ||||||||
761 | if (It == HoistableBranches.end()) { | ||||||||
762 | LLVM_DEBUG(dbgs() << "LICM using "do { } while (false) | ||||||||
763 | << InitialPreheader->getNameOrAsOperand()do { } while (false) | ||||||||
764 | << " as hoist destination for "do { } while (false) | ||||||||
765 | << BB->getNameOrAsOperand() << "\n")do { } while (false); | ||||||||
766 | HoistDestinationMap[BB] = InitialPreheader; | ||||||||
767 | return InitialPreheader; | ||||||||
768 | } | ||||||||
769 | BranchInst *BI = It->first; | ||||||||
770 | assert(std::find_if(++It, HoistableBranches.end(), HasBBAsSuccessor) ==((void)0) | ||||||||
771 | HoistableBranches.end() &&((void)0) | ||||||||
772 | "BB is expected to be the target of at most one branch")((void)0); | ||||||||
773 | |||||||||
774 | LLVMContext &C = BB->getContext(); | ||||||||
775 | BasicBlock *TrueDest = BI->getSuccessor(0); | ||||||||
776 | BasicBlock *FalseDest = BI->getSuccessor(1); | ||||||||
777 | BasicBlock *CommonSucc = HoistableBranches[BI]; | ||||||||
778 | BasicBlock *HoistTarget = getOrCreateHoistedBlock(BI->getParent()); | ||||||||
779 | |||||||||
780 | // Create hoisted versions of blocks that currently don't have them | ||||||||
781 | auto CreateHoistedBlock = [&](BasicBlock *Orig) { | ||||||||
782 | if (HoistDestinationMap.count(Orig)) | ||||||||
783 | return HoistDestinationMap[Orig]; | ||||||||
784 | BasicBlock *New = | ||||||||
785 | BasicBlock::Create(C, Orig->getName() + ".licm", Orig->getParent()); | ||||||||
786 | HoistDestinationMap[Orig] = New; | ||||||||
787 | DT->addNewBlock(New, HoistTarget); | ||||||||
788 | if (CurLoop->getParentLoop()) | ||||||||
789 | CurLoop->getParentLoop()->addBasicBlockToLoop(New, *LI); | ||||||||
790 | ++NumCreatedBlocks; | ||||||||
791 | LLVM_DEBUG(dbgs() << "LICM created " << New->getName()do { } while (false) | ||||||||
792 | << " as hoist destination for " << Orig->getName()do { } while (false) | ||||||||
793 | << "\n")do { } while (false); | ||||||||
794 | return New; | ||||||||
795 | }; | ||||||||
796 | BasicBlock *HoistTrueDest = CreateHoistedBlock(TrueDest); | ||||||||
797 | BasicBlock *HoistFalseDest = CreateHoistedBlock(FalseDest); | ||||||||
798 | BasicBlock *HoistCommonSucc = CreateHoistedBlock(CommonSucc); | ||||||||
799 | |||||||||
800 | // Link up these blocks with branches. | ||||||||
801 | if (!HoistCommonSucc->getTerminator()) { | ||||||||
802 | // The new common successor we've generated will branch to whatever that | ||||||||
803 | // hoist target branched to. | ||||||||
804 | BasicBlock *TargetSucc = HoistTarget->getSingleSuccessor(); | ||||||||
805 | assert(TargetSucc && "Expected hoist target to have a single successor")((void)0); | ||||||||
806 | HoistCommonSucc->moveBefore(TargetSucc); | ||||||||
807 | BranchInst::Create(TargetSucc, HoistCommonSucc); | ||||||||
808 | } | ||||||||
809 | if (!HoistTrueDest->getTerminator()) { | ||||||||
810 | HoistTrueDest->moveBefore(HoistCommonSucc); | ||||||||
811 | BranchInst::Create(HoistCommonSucc, HoistTrueDest); | ||||||||
812 | } | ||||||||
813 | if (!HoistFalseDest->getTerminator()) { | ||||||||
814 | HoistFalseDest->moveBefore(HoistCommonSucc); | ||||||||
815 | BranchInst::Create(HoistCommonSucc, HoistFalseDest); | ||||||||
816 | } | ||||||||
817 | |||||||||
818 | // If BI is being cloned to what was originally the preheader then | ||||||||
819 | // HoistCommonSucc will now be the new preheader. | ||||||||
820 | if (HoistTarget == InitialPreheader) { | ||||||||
821 | // Phis in the loop header now need to use the new preheader. | ||||||||
822 | InitialPreheader->replaceSuccessorsPhiUsesWith(HoistCommonSucc); | ||||||||
823 | if (MSSAU) | ||||||||
824 | MSSAU->wireOldPredecessorsToNewImmediatePredecessor( | ||||||||
825 | HoistTarget->getSingleSuccessor(), HoistCommonSucc, {HoistTarget}); | ||||||||
826 | // The new preheader dominates the loop header. | ||||||||
827 | DomTreeNode *PreheaderNode = DT->getNode(HoistCommonSucc); | ||||||||
828 | DomTreeNode *HeaderNode = DT->getNode(CurLoop->getHeader()); | ||||||||
829 | DT->changeImmediateDominator(HeaderNode, PreheaderNode); | ||||||||
830 | // The preheader hoist destination is now the new preheader, with the | ||||||||
831 | // exception of the hoist destination of this branch. | ||||||||
832 | for (auto &Pair : HoistDestinationMap) | ||||||||
833 | if (Pair.second == InitialPreheader && Pair.first != BI->getParent()) | ||||||||
834 | Pair.second = HoistCommonSucc; | ||||||||
835 | } | ||||||||
836 | |||||||||
837 | // Now finally clone BI. | ||||||||
838 | ReplaceInstWithInst( | ||||||||
839 | HoistTarget->getTerminator(), | ||||||||
840 | BranchInst::Create(HoistTrueDest, HoistFalseDest, BI->getCondition())); | ||||||||
841 | ++NumClonedBranches; | ||||||||
842 | |||||||||
843 | assert(CurLoop->getLoopPreheader() &&((void)0) | ||||||||
844 | "Hoisting blocks should not have destroyed preheader")((void)0); | ||||||||
845 | return HoistDestinationMap[BB]; | ||||||||
846 | } | ||||||||
847 | }; | ||||||||
848 | } // namespace | ||||||||
849 | |||||||||
850 | // Hoisting/sinking instruction out of a loop isn't always beneficial. It's only | ||||||||
851 | // only worthwhile if the destination block is actually colder than current | ||||||||
852 | // block. | ||||||||
853 | static bool worthSinkOrHoistInst(Instruction &I, BasicBlock *DstBlock, | ||||||||
854 | OptimizationRemarkEmitter *ORE, | ||||||||
855 | BlockFrequencyInfo *BFI) { | ||||||||
856 | // Check block frequency only when runtime profile is available | ||||||||
857 | // to avoid pathological cases. With static profile, lean towards | ||||||||
858 | // hosting because it helps canonicalize the loop for vectorizer. | ||||||||
859 | if (!DstBlock->getParent()->hasProfileData()) | ||||||||
860 | return true; | ||||||||
861 | |||||||||
862 | if (!HoistSinkColdnessThreshold || !BFI) | ||||||||
863 | return true; | ||||||||
864 | |||||||||
865 | BasicBlock *SrcBlock = I.getParent(); | ||||||||
866 | if (BFI->getBlockFreq(DstBlock).getFrequency() / HoistSinkColdnessThreshold > | ||||||||
867 | BFI->getBlockFreq(SrcBlock).getFrequency()) { | ||||||||
868 | ORE->emit([&]() { | ||||||||
869 | return OptimizationRemarkMissed(DEBUG_TYPE"licm", "SinkHoistInst", &I) | ||||||||
870 | << "failed to sink or hoist instruction because containing block " | ||||||||
871 | "has lower frequency than destination block"; | ||||||||
872 | }); | ||||||||
873 | return false; | ||||||||
874 | } | ||||||||
875 | |||||||||
876 | return true; | ||||||||
877 | } | ||||||||
878 | |||||||||
879 | /// Walk the specified region of the CFG (defined by all blocks dominated by | ||||||||
880 | /// the specified block, and that are in the current loop) in depth first | ||||||||
881 | /// order w.r.t the DominatorTree. This allows us to visit definitions before | ||||||||
882 | /// uses, allowing us to hoist a loop body in one pass without iteration. | ||||||||
883 | /// | ||||||||
884 | bool llvm::hoistRegion(DomTreeNode *N, AAResults *AA, LoopInfo *LI, | ||||||||
885 | DominatorTree *DT, BlockFrequencyInfo *BFI, | ||||||||
886 | TargetLibraryInfo *TLI, Loop *CurLoop, | ||||||||
887 | AliasSetTracker *CurAST, MemorySSAUpdater *MSSAU, | ||||||||
888 | ScalarEvolution *SE, ICFLoopSafetyInfo *SafetyInfo, | ||||||||
889 | SinkAndHoistLICMFlags &Flags, | ||||||||
890 | OptimizationRemarkEmitter *ORE, bool LoopNestMode) { | ||||||||
891 | // Verify inputs. | ||||||||
892 | assert(N != nullptr && AA != nullptr && LI != nullptr && DT != nullptr &&((void)0) | ||||||||
893 | CurLoop != nullptr && SafetyInfo != nullptr &&((void)0) | ||||||||
894 | "Unexpected input to hoistRegion.")((void)0); | ||||||||
895 | assert(((CurAST != nullptr) ^ (MSSAU != nullptr)) &&((void)0) | ||||||||
896 | "Either AliasSetTracker or MemorySSA should be initialized.")((void)0); | ||||||||
897 | |||||||||
898 | ControlFlowHoister CFH(LI, DT, CurLoop, MSSAU); | ||||||||
899 | |||||||||
900 | // Keep track of instructions that have been hoisted, as they may need to be | ||||||||
901 | // re-hoisted if they end up not dominating all of their uses. | ||||||||
902 | SmallVector<Instruction *, 16> HoistedInstructions; | ||||||||
903 | |||||||||
904 | // For PHI hoisting to work we need to hoist blocks before their successors. | ||||||||
905 | // We can do this by iterating through the blocks in the loop in reverse | ||||||||
906 | // post-order. | ||||||||
907 | LoopBlocksRPO Worklist(CurLoop); | ||||||||
908 | Worklist.perform(LI); | ||||||||
909 | bool Changed = false; | ||||||||
910 | for (BasicBlock *BB : Worklist) { | ||||||||
911 | // Only need to process the contents of this block if it is not part of a | ||||||||
912 | // subloop (which would already have been processed). | ||||||||
913 | if (!LoopNestMode && inSubLoop(BB, CurLoop, LI)) | ||||||||
914 | continue; | ||||||||
915 | |||||||||
916 | for (BasicBlock::iterator II = BB->begin(), E = BB->end(); II != E;) { | ||||||||
917 | Instruction &I = *II++; | ||||||||
918 | // Try constant folding this instruction. If all the operands are | ||||||||
919 | // constants, it is technically hoistable, but it would be better to | ||||||||
920 | // just fold it. | ||||||||
921 | if (Constant *C = ConstantFoldInstruction( | ||||||||
922 | &I, I.getModule()->getDataLayout(), TLI)) { | ||||||||
923 | LLVM_DEBUG(dbgs() << "LICM folding inst: " << I << " --> " << *Cdo { } while (false) | ||||||||
924 | << '\n')do { } while (false); | ||||||||
925 | if (CurAST) | ||||||||
926 | CurAST->copyValue(&I, C); | ||||||||
927 | // FIXME MSSA: Such replacements may make accesses unoptimized (D51960). | ||||||||
928 | I.replaceAllUsesWith(C); | ||||||||
929 | if (isInstructionTriviallyDead(&I, TLI)) | ||||||||
930 | eraseInstruction(I, *SafetyInfo, CurAST, MSSAU); | ||||||||
931 | Changed = true; | ||||||||
932 | continue; | ||||||||
933 | } | ||||||||
934 | |||||||||
935 | // Try hoisting the instruction out to the preheader. We can only do | ||||||||
936 | // this if all of the operands of the instruction are loop invariant and | ||||||||
937 | // if it is safe to hoist the instruction. We also check block frequency | ||||||||
938 | // to make sure instruction only gets hoisted into colder blocks. | ||||||||
939 | // TODO: It may be safe to hoist if we are hoisting to a conditional block | ||||||||
940 | // and we have accurately duplicated the control flow from the loop header | ||||||||
941 | // to that block. | ||||||||
942 | if (CurLoop->hasLoopInvariantOperands(&I) && | ||||||||
943 | canSinkOrHoistInst(I, AA, DT, CurLoop, CurAST, MSSAU, true, &Flags, | ||||||||
944 | ORE) && | ||||||||
945 | worthSinkOrHoistInst(I, CurLoop->getLoopPreheader(), ORE, BFI) && | ||||||||
946 | isSafeToExecuteUnconditionally( | ||||||||
947 | I, DT, TLI, CurLoop, SafetyInfo, ORE, | ||||||||
948 | CurLoop->getLoopPreheader()->getTerminator())) { | ||||||||
949 | hoist(I, DT, CurLoop, CFH.getOrCreateHoistedBlock(BB), SafetyInfo, | ||||||||
950 | MSSAU, SE, ORE); | ||||||||
951 | HoistedInstructions.push_back(&I); | ||||||||
952 | Changed = true; | ||||||||
953 | continue; | ||||||||
954 | } | ||||||||
955 | |||||||||
956 | // Attempt to remove floating point division out of the loop by | ||||||||
957 | // converting it to a reciprocal multiplication. | ||||||||
958 | if (I.getOpcode() == Instruction::FDiv && I.hasAllowReciprocal() && | ||||||||
959 | CurLoop->isLoopInvariant(I.getOperand(1))) { | ||||||||
960 | auto Divisor = I.getOperand(1); | ||||||||
961 | auto One = llvm::ConstantFP::get(Divisor->getType(), 1.0); | ||||||||
962 | auto ReciprocalDivisor = BinaryOperator::CreateFDiv(One, Divisor); | ||||||||
963 | ReciprocalDivisor->setFastMathFlags(I.getFastMathFlags()); | ||||||||
964 | SafetyInfo->insertInstructionTo(ReciprocalDivisor, I.getParent()); | ||||||||
965 | ReciprocalDivisor->insertBefore(&I); | ||||||||
966 | |||||||||
967 | auto Product = | ||||||||
968 | BinaryOperator::CreateFMul(I.getOperand(0), ReciprocalDivisor); | ||||||||
969 | Product->setFastMathFlags(I.getFastMathFlags()); | ||||||||
970 | SafetyInfo->insertInstructionTo(Product, I.getParent()); | ||||||||
971 | Product->insertAfter(&I); | ||||||||
972 | I.replaceAllUsesWith(Product); | ||||||||
973 | eraseInstruction(I, *SafetyInfo, CurAST, MSSAU); | ||||||||
974 | |||||||||
975 | hoist(*ReciprocalDivisor, DT, CurLoop, CFH.getOrCreateHoistedBlock(BB), | ||||||||
976 | SafetyInfo, MSSAU, SE, ORE); | ||||||||
977 | HoistedInstructions.push_back(ReciprocalDivisor); | ||||||||
978 | Changed = true; | ||||||||
979 | continue; | ||||||||
980 | } | ||||||||
981 | |||||||||
982 | auto IsInvariantStart = [&](Instruction &I) { | ||||||||
983 | using namespace PatternMatch; | ||||||||
984 | return I.use_empty() && | ||||||||
985 | match(&I, m_Intrinsic<Intrinsic::invariant_start>()); | ||||||||
986 | }; | ||||||||
987 | auto MustExecuteWithoutWritesBefore = [&](Instruction &I) { | ||||||||
988 | return SafetyInfo->isGuaranteedToExecute(I, DT, CurLoop) && | ||||||||
989 | SafetyInfo->doesNotWriteMemoryBefore(I, CurLoop); | ||||||||
990 | }; | ||||||||
991 | if ((IsInvariantStart(I) || isGuard(&I)) && | ||||||||
992 | CurLoop->hasLoopInvariantOperands(&I) && | ||||||||
993 | MustExecuteWithoutWritesBefore(I)) { | ||||||||
994 | hoist(I, DT, CurLoop, CFH.getOrCreateHoistedBlock(BB), SafetyInfo, | ||||||||
995 | MSSAU, SE, ORE); | ||||||||
996 | HoistedInstructions.push_back(&I); | ||||||||
997 | Changed = true; | ||||||||
998 | continue; | ||||||||
999 | } | ||||||||
1000 | |||||||||
1001 | if (PHINode *PN = dyn_cast<PHINode>(&I)) { | ||||||||
1002 | if (CFH.canHoistPHI(PN)) { | ||||||||
1003 | // Redirect incoming blocks first to ensure that we create hoisted | ||||||||
1004 | // versions of those blocks before we hoist the phi. | ||||||||
1005 | for (unsigned int i = 0; i < PN->getNumIncomingValues(); ++i) | ||||||||
1006 | PN->setIncomingBlock( | ||||||||
1007 | i, CFH.getOrCreateHoistedBlock(PN->getIncomingBlock(i))); | ||||||||
1008 | hoist(*PN, DT, CurLoop, CFH.getOrCreateHoistedBlock(BB), SafetyInfo, | ||||||||
1009 | MSSAU, SE, ORE); | ||||||||
1010 | assert(DT->dominates(PN, BB) && "Conditional PHIs not expected")((void)0); | ||||||||
1011 | Changed = true; | ||||||||
1012 | continue; | ||||||||
1013 | } | ||||||||
1014 | } | ||||||||
1015 | |||||||||
1016 | // Remember possibly hoistable branches so we can actually hoist them | ||||||||
1017 | // later if needed. | ||||||||
1018 | if (BranchInst *BI = dyn_cast<BranchInst>(&I)) | ||||||||
1019 | CFH.registerPossiblyHoistableBranch(BI); | ||||||||
1020 | } | ||||||||
1021 | } | ||||||||
1022 | |||||||||
1023 | // If we hoisted instructions to a conditional block they may not dominate | ||||||||
1024 | // their uses that weren't hoisted (such as phis where some operands are not | ||||||||
1025 | // loop invariant). If so make them unconditional by moving them to their | ||||||||
1026 | // immediate dominator. We iterate through the instructions in reverse order | ||||||||
1027 | // which ensures that when we rehoist an instruction we rehoist its operands, | ||||||||
1028 | // and also keep track of where in the block we are rehoisting to to make sure | ||||||||
1029 | // that we rehoist instructions before the instructions that use them. | ||||||||
1030 | Instruction *HoistPoint = nullptr; | ||||||||
1031 | if (ControlFlowHoisting) { | ||||||||
1032 | for (Instruction *I : reverse(HoistedInstructions)) { | ||||||||
1033 | if (!llvm::all_of(I->uses(), | ||||||||
1034 | [&](Use &U) { return DT->dominates(I, U); })) { | ||||||||
1035 | BasicBlock *Dominator = | ||||||||
1036 | DT->getNode(I->getParent())->getIDom()->getBlock(); | ||||||||
1037 | if (!HoistPoint || !DT->dominates(HoistPoint->getParent(), Dominator)) { | ||||||||
1038 | if (HoistPoint) | ||||||||
1039 | assert(DT->dominates(Dominator, HoistPoint->getParent()) &&((void)0) | ||||||||
1040 | "New hoist point expected to dominate old hoist point")((void)0); | ||||||||
1041 | HoistPoint = Dominator->getTerminator(); | ||||||||
1042 | } | ||||||||
1043 | LLVM_DEBUG(dbgs() << "LICM rehoisting to "do { } while (false) | ||||||||
1044 | << HoistPoint->getParent()->getNameOrAsOperand()do { } while (false) | ||||||||
1045 | << ": " << *I << "\n")do { } while (false); | ||||||||
1046 | moveInstructionBefore(*I, *HoistPoint, *SafetyInfo, MSSAU, SE); | ||||||||
1047 | HoistPoint = I; | ||||||||
1048 | Changed = true; | ||||||||
1049 | } | ||||||||
1050 | } | ||||||||
1051 | } | ||||||||
1052 | if (MSSAU && VerifyMemorySSA) | ||||||||
1053 | MSSAU->getMemorySSA()->verifyMemorySSA(); | ||||||||
1054 | |||||||||
1055 | // Now that we've finished hoisting make sure that LI and DT are still | ||||||||
1056 | // valid. | ||||||||
1057 | #ifdef EXPENSIVE_CHECKS | ||||||||
1058 | if (Changed) { | ||||||||
1059 | assert(DT->verify(DominatorTree::VerificationLevel::Fast) &&((void)0) | ||||||||
1060 | "Dominator tree verification failed")((void)0); | ||||||||
1061 | LI->verify(*DT); | ||||||||
1062 | } | ||||||||
1063 | #endif | ||||||||
1064 | |||||||||
1065 | return Changed; | ||||||||
1066 | } | ||||||||
1067 | |||||||||
1068 | // Return true if LI is invariant within scope of the loop. LI is invariant if | ||||||||
1069 | // CurLoop is dominated by an invariant.start representing the same memory | ||||||||
1070 | // location and size as the memory location LI loads from, and also the | ||||||||
1071 | // invariant.start has no uses. | ||||||||
1072 | static bool isLoadInvariantInLoop(LoadInst *LI, DominatorTree *DT, | ||||||||
1073 | Loop *CurLoop) { | ||||||||
1074 | Value *Addr = LI->getOperand(0); | ||||||||
1075 | const DataLayout &DL = LI->getModule()->getDataLayout(); | ||||||||
1076 | const TypeSize LocSizeInBits = DL.getTypeSizeInBits(LI->getType()); | ||||||||
1077 | |||||||||
1078 | // It is not currently possible for clang to generate an invariant.start | ||||||||
1079 | // intrinsic with scalable vector types because we don't support thread local | ||||||||
1080 | // sizeless types and we don't permit sizeless types in structs or classes. | ||||||||
1081 | // Furthermore, even if support is added for this in future the intrinsic | ||||||||
1082 | // itself is defined to have a size of -1 for variable sized objects. This | ||||||||
1083 | // makes it impossible to verify if the intrinsic envelops our region of | ||||||||
1084 | // interest. For example, both <vscale x 32 x i8> and <vscale x 16 x i8> | ||||||||
1085 | // types would have a -1 parameter, but the former is clearly double the size | ||||||||
1086 | // of the latter. | ||||||||
1087 | if (LocSizeInBits.isScalable()) | ||||||||
1088 | return false; | ||||||||
1089 | |||||||||
1090 | // if the type is i8 addrspace(x)*, we know this is the type of | ||||||||
1091 | // llvm.invariant.start operand | ||||||||
1092 | auto *PtrInt8Ty = PointerType::get(Type::getInt8Ty(LI->getContext()), | ||||||||
1093 | LI->getPointerAddressSpace()); | ||||||||
1094 | unsigned BitcastsVisited = 0; | ||||||||
1095 | // Look through bitcasts until we reach the i8* type (this is invariant.start | ||||||||
1096 | // operand type). | ||||||||
1097 | while (Addr->getType() != PtrInt8Ty) { | ||||||||
1098 | auto *BC = dyn_cast<BitCastInst>(Addr); | ||||||||
1099 | // Avoid traversing high number of bitcast uses. | ||||||||
1100 | if (++BitcastsVisited > MaxNumUsesTraversed || !BC) | ||||||||
1101 | return false; | ||||||||
1102 | Addr = BC->getOperand(0); | ||||||||
1103 | } | ||||||||
1104 | |||||||||
1105 | unsigned UsesVisited = 0; | ||||||||
1106 | // Traverse all uses of the load operand value, to see if invariant.start is | ||||||||
1107 | // one of the uses, and whether it dominates the load instruction. | ||||||||
1108 | for (auto *U : Addr->users()) { | ||||||||
1109 | // Avoid traversing for Load operand with high number of users. | ||||||||
1110 | if (++UsesVisited > MaxNumUsesTraversed) | ||||||||
1111 | return false; | ||||||||
1112 | IntrinsicInst *II = dyn_cast<IntrinsicInst>(U); | ||||||||
1113 | // If there are escaping uses of invariant.start instruction, the load maybe | ||||||||
1114 | // non-invariant. | ||||||||
1115 | if (!II || II->getIntrinsicID() != Intrinsic::invariant_start || | ||||||||
1116 | !II->use_empty()) | ||||||||
1117 | continue; | ||||||||
1118 | ConstantInt *InvariantSize = cast<ConstantInt>(II->getArgOperand(0)); | ||||||||
1119 | // The intrinsic supports having a -1 argument for variable sized objects | ||||||||
1120 | // so we should check for that here. | ||||||||
1121 | if (InvariantSize->isNegative()) | ||||||||
1122 | continue; | ||||||||
1123 | uint64_t InvariantSizeInBits = InvariantSize->getSExtValue() * 8; | ||||||||
1124 | // Confirm the invariant.start location size contains the load operand size | ||||||||
1125 | // in bits. Also, the invariant.start should dominate the load, and we | ||||||||
1126 | // should not hoist the load out of a loop that contains this dominating | ||||||||
1127 | // invariant.start. | ||||||||
1128 | if (LocSizeInBits.getFixedSize() <= InvariantSizeInBits && | ||||||||
1129 | DT->properlyDominates(II->getParent(), CurLoop->getHeader())) | ||||||||
1130 | return true; | ||||||||
1131 | } | ||||||||
1132 | |||||||||
1133 | return false; | ||||||||
1134 | } | ||||||||
1135 | |||||||||
1136 | namespace { | ||||||||
1137 | /// Return true if-and-only-if we know how to (mechanically) both hoist and | ||||||||
1138 | /// sink a given instruction out of a loop. Does not address legality | ||||||||
1139 | /// concerns such as aliasing or speculation safety. | ||||||||
1140 | bool isHoistableAndSinkableInst(Instruction &I) { | ||||||||
1141 | // Only these instructions are hoistable/sinkable. | ||||||||
1142 | return (isa<LoadInst>(I) || isa<StoreInst>(I) || isa<CallInst>(I) || | ||||||||
1143 | isa<FenceInst>(I) || isa<CastInst>(I) || isa<UnaryOperator>(I) || | ||||||||
1144 | isa<BinaryOperator>(I) || isa<SelectInst>(I) || | ||||||||
1145 | isa<GetElementPtrInst>(I) || isa<CmpInst>(I) || | ||||||||
1146 | isa<InsertElementInst>(I) || isa<ExtractElementInst>(I) || | ||||||||
1147 | isa<ShuffleVectorInst>(I) || isa<ExtractValueInst>(I) || | ||||||||
1148 | isa<InsertValueInst>(I) || isa<FreezeInst>(I)); | ||||||||
1149 | } | ||||||||
1150 | /// Return true if all of the alias sets within this AST are known not to | ||||||||
1151 | /// contain a Mod, or if MSSA knows there are no MemoryDefs in the loop. | ||||||||
1152 | bool isReadOnly(AliasSetTracker *CurAST, const MemorySSAUpdater *MSSAU, | ||||||||
1153 | const Loop *L) { | ||||||||
1154 | if (CurAST) { | ||||||||
1155 | for (AliasSet &AS : *CurAST) { | ||||||||
1156 | if (!AS.isForwardingAliasSet() && AS.isMod()) { | ||||||||
1157 | return false; | ||||||||
1158 | } | ||||||||
1159 | } | ||||||||
1160 | return true; | ||||||||
1161 | } else { /*MSSAU*/ | ||||||||
1162 | for (auto *BB : L->getBlocks()) | ||||||||
1163 | if (MSSAU->getMemorySSA()->getBlockDefs(BB)) | ||||||||
| |||||||||
1164 | return false; | ||||||||
1165 | return true; | ||||||||
1166 | } | ||||||||
1167 | } | ||||||||
1168 | |||||||||
1169 | /// Return true if I is the only Instruction with a MemoryAccess in L. | ||||||||
1170 | bool isOnlyMemoryAccess(const Instruction *I, const Loop *L, | ||||||||
1171 | const MemorySSAUpdater *MSSAU) { | ||||||||
1172 | for (auto *BB : L->getBlocks()) | ||||||||
1173 | if (auto *Accs = MSSAU->getMemorySSA()->getBlockAccesses(BB)) { | ||||||||
1174 | int NotAPhi = 0; | ||||||||
1175 | for (const auto &Acc : *Accs) { | ||||||||
1176 | if (isa<MemoryPhi>(&Acc)) | ||||||||
1177 | continue; | ||||||||
1178 | const auto *MUD = cast<MemoryUseOrDef>(&Acc); | ||||||||
1179 | if (MUD->getMemoryInst() != I || NotAPhi++ == 1) | ||||||||
1180 | return false; | ||||||||
1181 | } | ||||||||
1182 | } | ||||||||
1183 | return true; | ||||||||
1184 | } | ||||||||
1185 | } | ||||||||
1186 | |||||||||
1187 | bool llvm::canSinkOrHoistInst(Instruction &I, AAResults *AA, DominatorTree *DT, | ||||||||
1188 | Loop *CurLoop, AliasSetTracker *CurAST, | ||||||||
1189 | MemorySSAUpdater *MSSAU, | ||||||||
1190 | bool TargetExecutesOncePerLoop, | ||||||||
1191 | SinkAndHoistLICMFlags *Flags, | ||||||||
1192 | OptimizationRemarkEmitter *ORE) { | ||||||||
1193 | assert(((CurAST != nullptr) ^ (MSSAU != nullptr)) &&((void)0) | ||||||||
1194 | "Either AliasSetTracker or MemorySSA should be initialized.")((void)0); | ||||||||
1195 | |||||||||
1196 | // If we don't understand the instruction, bail early. | ||||||||
1197 | if (!isHoistableAndSinkableInst(I)) | ||||||||
| |||||||||
1198 | return false; | ||||||||
1199 | |||||||||
1200 | MemorySSA *MSSA = MSSAU ? MSSAU->getMemorySSA() : nullptr; | ||||||||
1201 | if (MSSA
| ||||||||
1202 | assert(Flags != nullptr && "Flags cannot be null.")((void)0); | ||||||||
1203 | |||||||||
1204 | // Loads have extra constraints we have to verify before we can hoist them. | ||||||||
1205 | if (LoadInst *LI
| ||||||||
1206 | if (!LI->isUnordered()) | ||||||||
1207 | return false; // Don't sink/hoist volatile or ordered atomic loads! | ||||||||
1208 | |||||||||
1209 | // Loads from constant memory are always safe to move, even if they end up | ||||||||
1210 | // in the same alias set as something that ends up being modified. | ||||||||
1211 | if (AA->pointsToConstantMemory(LI->getOperand(0))) | ||||||||
1212 | return true; | ||||||||
1213 | if (LI->hasMetadata(LLVMContext::MD_invariant_load)) | ||||||||
1214 | return true; | ||||||||
1215 | |||||||||
1216 | if (LI->isAtomic() && !TargetExecutesOncePerLoop) | ||||||||
1217 | return false; // Don't risk duplicating unordered loads | ||||||||
1218 | |||||||||
1219 | // This checks for an invariant.start dominating the load. | ||||||||
1220 | if (isLoadInvariantInLoop(LI, DT, CurLoop)) | ||||||||
1221 | return true; | ||||||||
1222 | |||||||||
1223 | bool Invalidated; | ||||||||
1224 | if (CurAST) | ||||||||
1225 | Invalidated = pointerInvalidatedByLoop(MemoryLocation::get(LI), CurAST, | ||||||||
1226 | CurLoop, AA); | ||||||||
1227 | else | ||||||||
1228 | Invalidated = pointerInvalidatedByLoopWithMSSA( | ||||||||
1229 | MSSA, cast<MemoryUse>(MSSA->getMemoryAccess(LI)), CurLoop, I, *Flags); | ||||||||
1230 | // Check loop-invariant address because this may also be a sinkable load | ||||||||
1231 | // whose address is not necessarily loop-invariant. | ||||||||
1232 | if (ORE && Invalidated && CurLoop->isLoopInvariant(LI->getPointerOperand())) | ||||||||
1233 | ORE->emit([&]() { | ||||||||
1234 | return OptimizationRemarkMissed( | ||||||||
1235 | DEBUG_TYPE"licm", "LoadWithLoopInvariantAddressInvalidated", LI) | ||||||||
1236 | << "failed to move load with loop-invariant address " | ||||||||
1237 | "because the loop may invalidate its value"; | ||||||||
1238 | }); | ||||||||
1239 | |||||||||
1240 | return !Invalidated; | ||||||||
1241 | } else if (CallInst *CI
| ||||||||
1242 | // Don't sink or hoist dbg info; it's legal, but not useful. | ||||||||
1243 | if (isa<DbgInfoIntrinsic>(I)) | ||||||||
1244 | return false; | ||||||||
1245 | |||||||||
1246 | // Don't sink calls which can throw. | ||||||||
1247 | if (CI->mayThrow()) | ||||||||
1248 | return false; | ||||||||
1249 | |||||||||
1250 | // Convergent attribute has been used on operations that involve | ||||||||
1251 | // inter-thread communication which results are implicitly affected by the | ||||||||
1252 | // enclosing control flows. It is not safe to hoist or sink such operations | ||||||||
1253 | // across control flow. | ||||||||
1254 | if (CI->isConvergent()) | ||||||||
1255 | return false; | ||||||||
1256 | |||||||||
1257 | using namespace PatternMatch; | ||||||||
1258 | if (match(CI, m_Intrinsic<Intrinsic::assume>())) | ||||||||
1259 | // Assumes don't actually alias anything or throw | ||||||||
1260 | return true; | ||||||||
1261 | |||||||||
1262 | if (match(CI, m_Intrinsic<Intrinsic::experimental_widenable_condition>())) | ||||||||
1263 | // Widenable conditions don't actually alias anything or throw | ||||||||
1264 | return true; | ||||||||
1265 | |||||||||
1266 | // Handle simple cases by querying alias analysis. | ||||||||
1267 | FunctionModRefBehavior Behavior = AA->getModRefBehavior(CI); | ||||||||
1268 | if (Behavior == FMRB_DoesNotAccessMemory) | ||||||||
1269 | return true; | ||||||||
1270 | if (AAResults::onlyReadsMemory(Behavior)) { | ||||||||
1271 | // A readonly argmemonly function only reads from memory pointed to by | ||||||||
1272 | // it's arguments with arbitrary offsets. If we can prove there are no | ||||||||
1273 | // writes to this memory in the loop, we can hoist or sink. | ||||||||
1274 | if (AAResults::onlyAccessesArgPointees(Behavior)) { | ||||||||
1275 | // TODO: expand to writeable arguments | ||||||||
1276 | for (Value *Op : CI->arg_operands()) | ||||||||
1277 | if (Op->getType()->isPointerTy()) { | ||||||||
1278 | bool Invalidated; | ||||||||
1279 | if (CurAST) | ||||||||
1280 | Invalidated = pointerInvalidatedByLoop( | ||||||||
1281 | MemoryLocation::getBeforeOrAfter(Op), CurAST, CurLoop, AA); | ||||||||
1282 | else | ||||||||
1283 | Invalidated = pointerInvalidatedByLoopWithMSSA( | ||||||||
1284 | MSSA, cast<MemoryUse>(MSSA->getMemoryAccess(CI)), CurLoop, I, | ||||||||
1285 | *Flags); | ||||||||
1286 | if (Invalidated) | ||||||||
1287 | return false; | ||||||||
1288 | } | ||||||||
1289 | return true; | ||||||||
1290 | } | ||||||||
1291 | |||||||||
1292 | // If this call only reads from memory and there are no writes to memory | ||||||||
1293 | // in the loop, we can hoist or sink the call as appropriate. | ||||||||
1294 | if (isReadOnly(CurAST, MSSAU, CurLoop)) | ||||||||
1295 | return true; | ||||||||
1296 | } | ||||||||
1297 | |||||||||
1298 | // FIXME: This should use mod/ref information to see if we can hoist or | ||||||||
1299 | // sink the call. | ||||||||
1300 | |||||||||
1301 | return false; | ||||||||
1302 | } else if (auto *FI = dyn_cast<FenceInst>(&I)) { | ||||||||
1303 | // Fences alias (most) everything to provide ordering. For the moment, | ||||||||
1304 | // just give up if there are any other memory operations in the loop. | ||||||||
1305 | if (CurAST) { | ||||||||
1306 | auto Begin = CurAST->begin(); | ||||||||
1307 | assert(Begin != CurAST->end() && "must contain FI")((void)0); | ||||||||
1308 | if (std::next(Begin) != CurAST->end()) | ||||||||
1309 | // constant memory for instance, TODO: handle better | ||||||||
1310 | return false; | ||||||||
1311 | auto *UniqueI = Begin->getUniqueInstruction(); | ||||||||
1312 | if (!UniqueI) | ||||||||
1313 | // other memory op, give up | ||||||||
1314 | return false; | ||||||||
1315 | (void)FI; // suppress unused variable warning | ||||||||
1316 | assert(UniqueI == FI && "AS must contain FI")((void)0); | ||||||||
1317 | return true; | ||||||||
1318 | } else // MSSAU | ||||||||
1319 | return isOnlyMemoryAccess(FI, CurLoop, MSSAU); | ||||||||
1320 | } else if (auto *SI = dyn_cast<StoreInst>(&I)) { | ||||||||
1321 | if (!SI->isUnordered()) | ||||||||
1322 | return false; // Don't sink/hoist volatile or ordered atomic store! | ||||||||
1323 | |||||||||
1324 | // We can only hoist a store that we can prove writes a value which is not | ||||||||
1325 | // read or overwritten within the loop. For those cases, we fallback to | ||||||||
1326 | // load store promotion instead. TODO: We can extend this to cases where | ||||||||
1327 | // there is exactly one write to the location and that write dominates an | ||||||||
1328 | // arbitrary number of reads in the loop. | ||||||||
1329 | if (CurAST) { | ||||||||
1330 | auto &AS = CurAST->getAliasSetFor(MemoryLocation::get(SI)); | ||||||||
1331 | |||||||||
1332 | if (AS.isRef() || !AS.isMustAlias()) | ||||||||
1333 | // Quick exit test, handled by the full path below as well. | ||||||||
1334 | return false; | ||||||||
1335 | auto *UniqueI = AS.getUniqueInstruction(); | ||||||||
1336 | if (!UniqueI) | ||||||||
1337 | // other memory op, give up | ||||||||
1338 | return false; | ||||||||
1339 | assert(UniqueI == SI && "AS must contain SI")((void)0); | ||||||||
1340 | return true; | ||||||||
1341 | } else { // MSSAU | ||||||||
1342 | if (isOnlyMemoryAccess(SI, CurLoop, MSSAU)) | ||||||||
1343 | return true; | ||||||||
1344 | // If there are more accesses than the Promotion cap or no "quota" to | ||||||||
1345 | // check clobber, then give up as we're not walking a list that long. | ||||||||
1346 | if (Flags->tooManyMemoryAccesses() || Flags->tooManyClobberingCalls()) | ||||||||
1347 | return false; | ||||||||
1348 | // If there are interfering Uses (i.e. their defining access is in the | ||||||||
1349 | // loop), or ordered loads (stored as Defs!), don't move this store. | ||||||||
1350 | // Could do better here, but this is conservatively correct. | ||||||||
1351 | // TODO: Cache set of Uses on the first walk in runOnLoop, update when | ||||||||
1352 | // moving accesses. Can also extend to dominating uses. | ||||||||
1353 | auto *SIMD = MSSA->getMemoryAccess(SI); | ||||||||
1354 | for (auto *BB : CurLoop->getBlocks()) | ||||||||
1355 | if (auto *Accesses = MSSA->getBlockAccesses(BB)) { | ||||||||
1356 | for (const auto &MA : *Accesses) | ||||||||
1357 | if (const auto *MU = dyn_cast<MemoryUse>(&MA)) { | ||||||||
1358 | auto *MD = MU->getDefiningAccess(); | ||||||||
1359 | if (!MSSA->isLiveOnEntryDef(MD) && | ||||||||
1360 | CurLoop->contains(MD->getBlock())) | ||||||||
1361 | return false; | ||||||||
1362 | // Disable hoisting past potentially interfering loads. Optimized | ||||||||
1363 | // Uses may point to an access outside the loop, as getClobbering | ||||||||
1364 | // checks the previous iteration when walking the backedge. | ||||||||
1365 | // FIXME: More precise: no Uses that alias SI. | ||||||||
1366 | if (!Flags->getIsSink() && !MSSA->dominates(SIMD, MU)) | ||||||||
1367 | return false; | ||||||||
1368 | } else if (const auto *MD = dyn_cast<MemoryDef>(&MA)) { | ||||||||
1369 | if (auto *LI = dyn_cast<LoadInst>(MD->getMemoryInst())) { | ||||||||
1370 | (void)LI; // Silence warning. | ||||||||
1371 | assert(!LI->isUnordered() && "Expected unordered load")((void)0); | ||||||||
1372 | return false; | ||||||||
1373 | } | ||||||||
1374 | // Any call, while it may not be clobbering SI, it may be a use. | ||||||||
1375 | if (auto *CI = dyn_cast<CallInst>(MD->getMemoryInst())) { | ||||||||
1376 | // Check if the call may read from the memory location written | ||||||||
1377 | // to by SI. Check CI's attributes and arguments; the number of | ||||||||
1378 | // such checks performed is limited above by NoOfMemAccTooLarge. | ||||||||
1379 | ModRefInfo MRI = AA->getModRefInfo(CI, MemoryLocation::get(SI)); | ||||||||
1380 | if (isModOrRefSet(MRI)) | ||||||||
1381 | return false; | ||||||||
1382 | } | ||||||||
1383 | } | ||||||||
1384 | } | ||||||||
1385 | auto *Source = MSSA->getSkipSelfWalker()->getClobberingMemoryAccess(SI); | ||||||||
1386 | Flags->incrementClobberingCalls(); | ||||||||
1387 | // If there are no clobbering Defs in the loop, store is safe to hoist. | ||||||||
1388 | return MSSA->isLiveOnEntryDef(Source) || | ||||||||
1389 | !CurLoop->contains(Source->getBlock()); | ||||||||
1390 | } | ||||||||
1391 | } | ||||||||
1392 | |||||||||
1393 | assert(!I.mayReadOrWriteMemory() && "unhandled aliasing")((void)0); | ||||||||
1394 | |||||||||
1395 | // We've established mechanical ability and aliasing, it's up to the caller | ||||||||
1396 | // to check fault safety | ||||||||
1397 | return true; | ||||||||
1398 | } | ||||||||
1399 | |||||||||
1400 | /// Returns true if a PHINode is a trivially replaceable with an | ||||||||
1401 | /// Instruction. | ||||||||
1402 | /// This is true when all incoming values are that instruction. | ||||||||
1403 | /// This pattern occurs most often with LCSSA PHI nodes. | ||||||||
1404 | /// | ||||||||
1405 | static bool isTriviallyReplaceablePHI(const PHINode &PN, const Instruction &I) { | ||||||||
1406 | for (const Value *IncValue : PN.incoming_values()) | ||||||||
1407 | if (IncValue != &I) | ||||||||
1408 | return false; | ||||||||
1409 | |||||||||
1410 | return true; | ||||||||
1411 | } | ||||||||
1412 | |||||||||
1413 | /// Return true if the instruction is free in the loop. | ||||||||
1414 | static bool isFreeInLoop(const Instruction &I, const Loop *CurLoop, | ||||||||
1415 | const TargetTransformInfo *TTI) { | ||||||||
1416 | |||||||||
1417 | if (const GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(&I)) { | ||||||||
1418 | if (TTI->getUserCost(GEP, TargetTransformInfo::TCK_SizeAndLatency) != | ||||||||
1419 | TargetTransformInfo::TCC_Free) | ||||||||
1420 | return false; | ||||||||
1421 | // For a GEP, we cannot simply use getUserCost because currently it | ||||||||
1422 | // optimistically assume that a GEP will fold into addressing mode | ||||||||
1423 | // regardless of its users. | ||||||||
1424 | const BasicBlock *BB = GEP->getParent(); | ||||||||
1425 | for (const User *U : GEP->users()) { | ||||||||
1426 | const Instruction *UI = cast<Instruction>(U); | ||||||||
1427 | if (CurLoop->contains(UI) && | ||||||||
1428 | (BB != UI->getParent() || | ||||||||
1429 | (!isa<StoreInst>(UI) && !isa<LoadInst>(UI)))) | ||||||||
1430 | return false; | ||||||||
1431 | } | ||||||||
1432 | return true; | ||||||||
1433 | } else | ||||||||
1434 | return TTI->getUserCost(&I, TargetTransformInfo::TCK_SizeAndLatency) == | ||||||||
1435 | TargetTransformInfo::TCC_Free; | ||||||||
1436 | } | ||||||||
1437 | |||||||||
1438 | /// Return true if the only users of this instruction are outside of | ||||||||
1439 | /// the loop. If this is true, we can sink the instruction to the exit | ||||||||
1440 | /// blocks of the loop. | ||||||||
1441 | /// | ||||||||
1442 | /// We also return true if the instruction could be folded away in lowering. | ||||||||
1443 | /// (e.g., a GEP can be folded into a load as an addressing mode in the loop). | ||||||||
1444 | static bool isNotUsedOrFreeInLoop(const Instruction &I, const Loop *CurLoop, | ||||||||
1445 | const LoopSafetyInfo *SafetyInfo, | ||||||||
1446 | TargetTransformInfo *TTI, bool &FreeInLoop) { | ||||||||
1447 | const auto &BlockColors = SafetyInfo->getBlockColors(); | ||||||||
1448 | bool IsFree = isFreeInLoop(I, CurLoop, TTI); | ||||||||
1449 | for (const User *U : I.users()) { | ||||||||
1450 | const Instruction *UI = cast<Instruction>(U); | ||||||||
1451 | if (const PHINode *PN = dyn_cast<PHINode>(UI)) { | ||||||||
1452 | const BasicBlock *BB = PN->getParent(); | ||||||||
1453 | // We cannot sink uses in catchswitches. | ||||||||
1454 | if (isa<CatchSwitchInst>(BB->getTerminator())) | ||||||||
1455 | return false; | ||||||||
1456 | |||||||||
1457 | // We need to sink a callsite to a unique funclet. Avoid sinking if the | ||||||||
1458 | // phi use is too muddled. | ||||||||
1459 | if (isa<CallInst>(I)) | ||||||||
1460 | if (!BlockColors.empty() && | ||||||||
1461 | BlockColors.find(const_cast<BasicBlock *>(BB))->second.size() != 1) | ||||||||
1462 | return false; | ||||||||
1463 | } | ||||||||
1464 | |||||||||
1465 | if (CurLoop->contains(UI)) { | ||||||||
1466 | if (IsFree) { | ||||||||
1467 | FreeInLoop = true; | ||||||||
1468 | continue; | ||||||||
1469 | } | ||||||||
1470 | return false; | ||||||||
1471 | } | ||||||||
1472 | } | ||||||||
1473 | return true; | ||||||||
1474 | } | ||||||||
1475 | |||||||||
1476 | static Instruction *cloneInstructionInExitBlock( | ||||||||
1477 | Instruction &I, BasicBlock &ExitBlock, PHINode &PN, const LoopInfo *LI, | ||||||||
1478 | const LoopSafetyInfo *SafetyInfo, MemorySSAUpdater *MSSAU) { | ||||||||
1479 | Instruction *New; | ||||||||
1480 | if (auto *CI = dyn_cast<CallInst>(&I)) { | ||||||||
1481 | const auto &BlockColors = SafetyInfo->getBlockColors(); | ||||||||
1482 | |||||||||
1483 | // Sinking call-sites need to be handled differently from other | ||||||||
1484 | // instructions. The cloned call-site needs a funclet bundle operand | ||||||||
1485 | // appropriate for its location in the CFG. | ||||||||
1486 | SmallVector<OperandBundleDef, 1> OpBundles; | ||||||||
1487 | for (unsigned BundleIdx = 0, BundleEnd = CI->getNumOperandBundles(); | ||||||||
1488 | BundleIdx != BundleEnd; ++BundleIdx) { | ||||||||
1489 | OperandBundleUse Bundle = CI->getOperandBundleAt(BundleIdx); | ||||||||
1490 | if (Bundle.getTagID() == LLVMContext::OB_funclet) | ||||||||
1491 | continue; | ||||||||
1492 | |||||||||
1493 | OpBundles.emplace_back(Bundle); | ||||||||
1494 | } | ||||||||
1495 | |||||||||
1496 | if (!BlockColors.empty()) { | ||||||||
1497 | const ColorVector &CV = BlockColors.find(&ExitBlock)->second; | ||||||||
1498 | assert(CV.size() == 1 && "non-unique color for exit block!")((void)0); | ||||||||
1499 | BasicBlock *BBColor = CV.front(); | ||||||||
1500 | Instruction *EHPad = BBColor->getFirstNonPHI(); | ||||||||
1501 | if (EHPad->isEHPad()) | ||||||||
1502 | OpBundles.emplace_back("funclet", EHPad); | ||||||||
1503 | } | ||||||||
1504 | |||||||||
1505 | New = CallInst::Create(CI, OpBundles); | ||||||||
1506 | } else { | ||||||||
1507 | New = I.clone(); | ||||||||
1508 | } | ||||||||
1509 | |||||||||
1510 | ExitBlock.getInstList().insert(ExitBlock.getFirstInsertionPt(), New); | ||||||||
1511 | if (!I.getName().empty()) | ||||||||
1512 | New->setName(I.getName() + ".le"); | ||||||||
1513 | |||||||||
1514 | if (MSSAU && MSSAU->getMemorySSA()->getMemoryAccess(&I)) { | ||||||||
1515 | // Create a new MemoryAccess and let MemorySSA set its defining access. | ||||||||
1516 | MemoryAccess *NewMemAcc = MSSAU->createMemoryAccessInBB( | ||||||||
1517 | New, nullptr, New->getParent(), MemorySSA::Beginning); | ||||||||
1518 | if (NewMemAcc) { | ||||||||
1519 | if (auto *MemDef = dyn_cast<MemoryDef>(NewMemAcc)) | ||||||||
1520 | MSSAU->insertDef(MemDef, /*RenameUses=*/true); | ||||||||
1521 | else { | ||||||||
1522 | auto *MemUse = cast<MemoryUse>(NewMemAcc); | ||||||||
1523 | MSSAU->insertUse(MemUse, /*RenameUses=*/true); | ||||||||
1524 | } | ||||||||
1525 | } | ||||||||
1526 | } | ||||||||
1527 | |||||||||
1528 | // Build LCSSA PHI nodes for any in-loop operands (if legal). Note that | ||||||||
1529 | // this is particularly cheap because we can rip off the PHI node that we're | ||||||||
1530 | // replacing for the number and blocks of the predecessors. | ||||||||
1531 | // OPT: If this shows up in a profile, we can instead finish sinking all | ||||||||
1532 | // invariant instructions, and then walk their operands to re-establish | ||||||||
1533 | // LCSSA. That will eliminate creating PHI nodes just to nuke them when | ||||||||
1534 | // sinking bottom-up. | ||||||||
1535 | for (Use &Op : New->operands()) | ||||||||
1536 | if (LI->wouldBeOutOfLoopUseRequiringLCSSA(Op.get(), PN.getParent())) { | ||||||||
1537 | auto *OInst = cast<Instruction>(Op.get()); | ||||||||
1538 | PHINode *OpPN = | ||||||||
1539 | PHINode::Create(OInst->getType(), PN.getNumIncomingValues(), | ||||||||
1540 | OInst->getName() + ".lcssa", &ExitBlock.front()); | ||||||||
1541 | for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i) | ||||||||
1542 | OpPN->addIncoming(OInst, PN.getIncomingBlock(i)); | ||||||||
1543 | Op = OpPN; | ||||||||
1544 | } | ||||||||
1545 | return New; | ||||||||
1546 | } | ||||||||
1547 | |||||||||
1548 | static void eraseInstruction(Instruction &I, ICFLoopSafetyInfo &SafetyInfo, | ||||||||
1549 | AliasSetTracker *AST, MemorySSAUpdater *MSSAU) { | ||||||||
1550 | if (AST) | ||||||||
1551 | AST->deleteValue(&I); | ||||||||
1552 | if (MSSAU) | ||||||||
1553 | MSSAU->removeMemoryAccess(&I); | ||||||||
1554 | SafetyInfo.removeInstruction(&I); | ||||||||
1555 | I.eraseFromParent(); | ||||||||
1556 | } | ||||||||
1557 | |||||||||
1558 | static void moveInstructionBefore(Instruction &I, Instruction &Dest, | ||||||||
1559 | ICFLoopSafetyInfo &SafetyInfo, | ||||||||
1560 | MemorySSAUpdater *MSSAU, | ||||||||
1561 | ScalarEvolution *SE) { | ||||||||
1562 | SafetyInfo.removeInstruction(&I); | ||||||||
1563 | SafetyInfo.insertInstructionTo(&I, Dest.getParent()); | ||||||||
1564 | I.moveBefore(&Dest); | ||||||||
1565 | if (MSSAU) | ||||||||
1566 | if (MemoryUseOrDef *OldMemAcc = cast_or_null<MemoryUseOrDef>( | ||||||||
1567 | MSSAU->getMemorySSA()->getMemoryAccess(&I))) | ||||||||
1568 | MSSAU->moveToPlace(OldMemAcc, Dest.getParent(), | ||||||||
1569 | MemorySSA::BeforeTerminator); | ||||||||
1570 | if (SE) | ||||||||
1571 | SE->forgetValue(&I); | ||||||||
1572 | } | ||||||||
1573 | |||||||||
1574 | static Instruction *sinkThroughTriviallyReplaceablePHI( | ||||||||
1575 | PHINode *TPN, Instruction *I, LoopInfo *LI, | ||||||||
1576 | SmallDenseMap<BasicBlock *, Instruction *, 32> &SunkCopies, | ||||||||
1577 | const LoopSafetyInfo *SafetyInfo, const Loop *CurLoop, | ||||||||
1578 | MemorySSAUpdater *MSSAU) { | ||||||||
1579 | assert(isTriviallyReplaceablePHI(*TPN, *I) &&((void)0) | ||||||||
1580 | "Expect only trivially replaceable PHI")((void)0); | ||||||||
1581 | BasicBlock *ExitBlock = TPN->getParent(); | ||||||||
1582 | Instruction *New; | ||||||||
1583 | auto It = SunkCopies.find(ExitBlock); | ||||||||
1584 | if (It != SunkCopies.end()) | ||||||||
1585 | New = It->second; | ||||||||
1586 | else | ||||||||
1587 | New = SunkCopies[ExitBlock] = cloneInstructionInExitBlock( | ||||||||
1588 | *I, *ExitBlock, *TPN, LI, SafetyInfo, MSSAU); | ||||||||
1589 | return New; | ||||||||
1590 | } | ||||||||
1591 | |||||||||
1592 | static bool canSplitPredecessors(PHINode *PN, LoopSafetyInfo *SafetyInfo) { | ||||||||
1593 | BasicBlock *BB = PN->getParent(); | ||||||||
1594 | if (!BB->canSplitPredecessors()) | ||||||||
1595 | return false; | ||||||||
1596 | // It's not impossible to split EHPad blocks, but if BlockColors already exist | ||||||||
1597 | // it require updating BlockColors for all offspring blocks accordingly. By | ||||||||
1598 | // skipping such corner case, we can make updating BlockColors after splitting | ||||||||
1599 | // predecessor fairly simple. | ||||||||
1600 | if (!SafetyInfo->getBlockColors().empty() && BB->getFirstNonPHI()->isEHPad()) | ||||||||
1601 | return false; | ||||||||
1602 | for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) { | ||||||||
1603 | BasicBlock *BBPred = *PI; | ||||||||
1604 | if (isa<IndirectBrInst>(BBPred->getTerminator()) || | ||||||||
1605 | isa<CallBrInst>(BBPred->getTerminator())) | ||||||||
1606 | return false; | ||||||||
1607 | } | ||||||||
1608 | return true; | ||||||||
1609 | } | ||||||||
1610 | |||||||||
1611 | static void splitPredecessorsOfLoopExit(PHINode *PN, DominatorTree *DT, | ||||||||
1612 | LoopInfo *LI, const Loop *CurLoop, | ||||||||
1613 | LoopSafetyInfo *SafetyInfo, | ||||||||
1614 | MemorySSAUpdater *MSSAU) { | ||||||||
1615 | #ifndef NDEBUG1 | ||||||||
1616 | SmallVector<BasicBlock *, 32> ExitBlocks; | ||||||||
1617 | CurLoop->getUniqueExitBlocks(ExitBlocks); | ||||||||
1618 | SmallPtrSet<BasicBlock *, 32> ExitBlockSet(ExitBlocks.begin(), | ||||||||
1619 | ExitBlocks.end()); | ||||||||
1620 | #endif | ||||||||
1621 | BasicBlock *ExitBB = PN->getParent(); | ||||||||
1622 | assert(ExitBlockSet.count(ExitBB) && "Expect the PHI is in an exit block.")((void)0); | ||||||||
1623 | |||||||||
1624 | // Split predecessors of the loop exit to make instructions in the loop are | ||||||||
1625 | // exposed to exit blocks through trivially replaceable PHIs while keeping the | ||||||||
1626 | // loop in the canonical form where each predecessor of each exit block should | ||||||||
1627 | // be contained within the loop. For example, this will convert the loop below | ||||||||
1628 | // from | ||||||||
1629 | // | ||||||||
1630 | // LB1: | ||||||||
1631 | // %v1 = | ||||||||
1632 | // br %LE, %LB2 | ||||||||
1633 | // LB2: | ||||||||
1634 | // %v2 = | ||||||||
1635 | // br %LE, %LB1 | ||||||||
1636 | // LE: | ||||||||
1637 | // %p = phi [%v1, %LB1], [%v2, %LB2] <-- non-trivially replaceable | ||||||||
1638 | // | ||||||||
1639 | // to | ||||||||
1640 | // | ||||||||
1641 | // LB1: | ||||||||
1642 | // %v1 = | ||||||||
1643 | // br %LE.split, %LB2 | ||||||||
1644 | // LB2: | ||||||||
1645 | // %v2 = | ||||||||
1646 | // br %LE.split2, %LB1 | ||||||||
1647 | // LE.split: | ||||||||
1648 | // %p1 = phi [%v1, %LB1] <-- trivially replaceable | ||||||||
1649 | // br %LE | ||||||||
1650 | // LE.split2: | ||||||||
1651 | // %p2 = phi [%v2, %LB2] <-- trivially replaceable | ||||||||
1652 | // br %LE | ||||||||
1653 | // LE: | ||||||||
1654 | // %p = phi [%p1, %LE.split], [%p2, %LE.split2] | ||||||||
1655 | // | ||||||||
1656 | const auto &BlockColors = SafetyInfo->getBlockColors(); | ||||||||
1657 | SmallSetVector<BasicBlock *, 8> PredBBs(pred_begin(ExitBB), pred_end(ExitBB)); | ||||||||
1658 | while (!PredBBs.empty()) { | ||||||||
1659 | BasicBlock *PredBB = *PredBBs.begin(); | ||||||||
1660 | assert(CurLoop->contains(PredBB) &&((void)0) | ||||||||
1661 | "Expect all predecessors are in the loop")((void)0); | ||||||||
1662 | if (PN->getBasicBlockIndex(PredBB) >= 0) { | ||||||||
1663 | BasicBlock *NewPred = SplitBlockPredecessors( | ||||||||
1664 | ExitBB, PredBB, ".split.loop.exit", DT, LI, MSSAU, true); | ||||||||
1665 | // Since we do not allow splitting EH-block with BlockColors in | ||||||||
1666 | // canSplitPredecessors(), we can simply assign predecessor's color to | ||||||||
1667 | // the new block. | ||||||||
1668 | if (!BlockColors.empty()) | ||||||||
1669 | // Grab a reference to the ColorVector to be inserted before getting the | ||||||||
1670 | // reference to the vector we are copying because inserting the new | ||||||||
1671 | // element in BlockColors might cause the map to be reallocated. | ||||||||
1672 | SafetyInfo->copyColors(NewPred, PredBB); | ||||||||
1673 | } | ||||||||
1674 | PredBBs.remove(PredBB); | ||||||||
1675 | } | ||||||||
1676 | } | ||||||||
1677 | |||||||||
1678 | /// When an instruction is found to only be used outside of the loop, this | ||||||||
1679 | /// function moves it to the exit blocks and patches up SSA form as needed. | ||||||||
1680 | /// This method is guaranteed to remove the original instruction from its | ||||||||
1681 | /// position, and may either delete it or move it to outside of the loop. | ||||||||
1682 | /// | ||||||||
1683 | static bool sink(Instruction &I, LoopInfo *LI, DominatorTree *DT, | ||||||||
1684 | BlockFrequencyInfo *BFI, const Loop *CurLoop, | ||||||||
1685 | ICFLoopSafetyInfo *SafetyInfo, MemorySSAUpdater *MSSAU, | ||||||||
1686 | OptimizationRemarkEmitter *ORE) { | ||||||||
1687 | bool Changed = false; | ||||||||
1688 | LLVM_DEBUG(dbgs() << "LICM sinking instruction: " << I << "\n")do { } while (false); | ||||||||
1689 | |||||||||
1690 | // Iterate over users to be ready for actual sinking. Replace users via | ||||||||
1691 | // unreachable blocks with undef and make all user PHIs trivially replaceable. | ||||||||
1692 | SmallPtrSet<Instruction *, 8> VisitedUsers; | ||||||||
1693 | for (Value::user_iterator UI = I.user_begin(), UE = I.user_end(); UI != UE;) { | ||||||||
1694 | auto *User = cast<Instruction>(*UI); | ||||||||
1695 | Use &U = UI.getUse(); | ||||||||
1696 | ++UI; | ||||||||
1697 | |||||||||
1698 | if (VisitedUsers.count(User) || CurLoop->contains(User)) | ||||||||
1699 | continue; | ||||||||
1700 | |||||||||
1701 | if (!DT->isReachableFromEntry(User->getParent())) { | ||||||||
1702 | U = UndefValue::get(I.getType()); | ||||||||
1703 | Changed = true; | ||||||||
1704 | continue; | ||||||||
1705 | } | ||||||||
1706 | |||||||||
1707 | // The user must be a PHI node. | ||||||||
1708 | PHINode *PN = cast<PHINode>(User); | ||||||||
1709 | |||||||||
1710 | // Surprisingly, instructions can be used outside of loops without any | ||||||||
1711 | // exits. This can only happen in PHI nodes if the incoming block is | ||||||||
1712 | // unreachable. | ||||||||
1713 | BasicBlock *BB = PN->getIncomingBlock(U); | ||||||||
1714 | if (!DT->isReachableFromEntry(BB)) { | ||||||||
1715 | U = UndefValue::get(I.getType()); | ||||||||
1716 | Changed = true; | ||||||||
1717 | continue; | ||||||||
1718 | } | ||||||||
1719 | |||||||||
1720 | VisitedUsers.insert(PN); | ||||||||
1721 | if (isTriviallyReplaceablePHI(*PN, I)) | ||||||||
1722 | continue; | ||||||||
1723 | |||||||||
1724 | if (!canSplitPredecessors(PN, SafetyInfo)) | ||||||||
1725 | return Changed; | ||||||||
1726 | |||||||||
1727 | // Split predecessors of the PHI so that we can make users trivially | ||||||||
1728 | // replaceable. | ||||||||
1729 | splitPredecessorsOfLoopExit(PN, DT, LI, CurLoop, SafetyInfo, MSSAU); | ||||||||
1730 | |||||||||
1731 | // Should rebuild the iterators, as they may be invalidated by | ||||||||
1732 | // splitPredecessorsOfLoopExit(). | ||||||||
1733 | UI = I.user_begin(); | ||||||||
1734 | UE = I.user_end(); | ||||||||
1735 | } | ||||||||
1736 | |||||||||
1737 | if (VisitedUsers.empty()) | ||||||||
1738 | return Changed; | ||||||||
1739 | |||||||||
1740 | ORE->emit([&]() { | ||||||||
1741 | return OptimizationRemark(DEBUG_TYPE"licm", "InstSunk", &I) | ||||||||
1742 | << "sinking " << ore::NV("Inst", &I); | ||||||||
1743 | }); | ||||||||
1744 | if (isa<LoadInst>(I)) | ||||||||
1745 | ++NumMovedLoads; | ||||||||
1746 | else if (isa<CallInst>(I)) | ||||||||
1747 | ++NumMovedCalls; | ||||||||
1748 | ++NumSunk; | ||||||||
1749 | |||||||||
1750 | #ifndef NDEBUG1 | ||||||||
1751 | SmallVector<BasicBlock *, 32> ExitBlocks; | ||||||||
1752 | CurLoop->getUniqueExitBlocks(ExitBlocks); | ||||||||
1753 | SmallPtrSet<BasicBlock *, 32> ExitBlockSet(ExitBlocks.begin(), | ||||||||
1754 | ExitBlocks.end()); | ||||||||
1755 | #endif | ||||||||
1756 | |||||||||
1757 | // Clones of this instruction. Don't create more than one per exit block! | ||||||||
1758 | SmallDenseMap<BasicBlock *, Instruction *, 32> SunkCopies; | ||||||||
1759 | |||||||||
1760 | // If this instruction is only used outside of the loop, then all users are | ||||||||
1761 | // PHI nodes in exit blocks due to LCSSA form. Just RAUW them with clones of | ||||||||
1762 | // the instruction. | ||||||||
1763 | // First check if I is worth sinking for all uses. Sink only when it is worth | ||||||||
1764 | // across all uses. | ||||||||
1765 | SmallSetVector<User*, 8> Users(I.user_begin(), I.user_end()); | ||||||||
1766 | SmallVector<PHINode *, 8> ExitPNs; | ||||||||
1767 | for (auto *UI : Users) { | ||||||||
1768 | auto *User = cast<Instruction>(UI); | ||||||||
1769 | |||||||||
1770 | if (CurLoop->contains(User)) | ||||||||
1771 | continue; | ||||||||
1772 | |||||||||
1773 | PHINode *PN = cast<PHINode>(User); | ||||||||
1774 | assert(ExitBlockSet.count(PN->getParent()) &&((void)0) | ||||||||
1775 | "The LCSSA PHI is not in an exit block!")((void)0); | ||||||||
1776 | if (!worthSinkOrHoistInst(I, PN->getParent(), ORE, BFI)) { | ||||||||
1777 | return Changed; | ||||||||
1778 | } | ||||||||
1779 | |||||||||
1780 | ExitPNs.push_back(PN); | ||||||||
1781 | } | ||||||||
1782 | |||||||||
1783 | for (auto *PN : ExitPNs) { | ||||||||
1784 | |||||||||
1785 | // The PHI must be trivially replaceable. | ||||||||
1786 | Instruction *New = sinkThroughTriviallyReplaceablePHI( | ||||||||
1787 | PN, &I, LI, SunkCopies, SafetyInfo, CurLoop, MSSAU); | ||||||||
1788 | PN->replaceAllUsesWith(New); | ||||||||
1789 | eraseInstruction(*PN, *SafetyInfo, nullptr, nullptr); | ||||||||
1790 | Changed = true; | ||||||||
1791 | } | ||||||||
1792 | return Changed; | ||||||||
1793 | } | ||||||||
1794 | |||||||||
1795 | /// When an instruction is found to only use loop invariant operands that | ||||||||
1796 | /// is safe to hoist, this instruction is called to do the dirty work. | ||||||||
1797 | /// | ||||||||
1798 | static void hoist(Instruction &I, const DominatorTree *DT, const Loop *CurLoop, | ||||||||
1799 | BasicBlock *Dest, ICFLoopSafetyInfo *SafetyInfo, | ||||||||
1800 | MemorySSAUpdater *MSSAU, ScalarEvolution *SE, | ||||||||
1801 | OptimizationRemarkEmitter *ORE) { | ||||||||
1802 | LLVM_DEBUG(dbgs() << "LICM hoisting to " << Dest->getNameOrAsOperand() << ": "do { } while (false) | ||||||||
1803 | << I << "\n")do { } while (false); | ||||||||
1804 | ORE->emit([&]() { | ||||||||
1805 | return OptimizationRemark(DEBUG_TYPE"licm", "Hoisted", &I) << "hoisting " | ||||||||
1806 | << ore::NV("Inst", &I); | ||||||||
1807 | }); | ||||||||
1808 | |||||||||
1809 | // Metadata can be dependent on conditions we are hoisting above. | ||||||||
1810 | // Conservatively strip all metadata on the instruction unless we were | ||||||||
1811 | // guaranteed to execute I if we entered the loop, in which case the metadata | ||||||||
1812 | // is valid in the loop preheader. | ||||||||
1813 | // Similarly, If I is a call and it is not guaranteed to execute in the loop, | ||||||||
1814 | // then moving to the preheader means we should strip attributes on the call | ||||||||
1815 | // that can cause UB since we may be hoisting above conditions that allowed | ||||||||
1816 | // inferring those attributes. They may not be valid at the preheader. | ||||||||
1817 | if ((I.hasMetadataOtherThanDebugLoc() || isa<CallInst>(I)) && | ||||||||
1818 | // The check on hasMetadataOtherThanDebugLoc is to prevent us from burning | ||||||||
1819 | // time in isGuaranteedToExecute if we don't actually have anything to | ||||||||
1820 | // drop. It is a compile time optimization, not required for correctness. | ||||||||
1821 | !SafetyInfo->isGuaranteedToExecute(I, DT, CurLoop)) | ||||||||
1822 | I.dropUndefImplyingAttrsAndUnknownMetadata(); | ||||||||
1823 | |||||||||
1824 | if (isa<PHINode>(I)) | ||||||||
1825 | // Move the new node to the end of the phi list in the destination block. | ||||||||
1826 | moveInstructionBefore(I, *Dest->getFirstNonPHI(), *SafetyInfo, MSSAU, SE); | ||||||||
1827 | else | ||||||||
1828 | // Move the new node to the destination block, before its terminator. | ||||||||
1829 | moveInstructionBefore(I, *Dest->getTerminator(), *SafetyInfo, MSSAU, SE); | ||||||||
1830 | |||||||||
1831 | I.updateLocationAfterHoist(); | ||||||||
1832 | |||||||||
1833 | if (isa<LoadInst>(I)) | ||||||||
1834 | ++NumMovedLoads; | ||||||||
1835 | else if (isa<CallInst>(I)) | ||||||||
1836 | ++NumMovedCalls; | ||||||||
1837 | ++NumHoisted; | ||||||||
1838 | } | ||||||||
1839 | |||||||||
1840 | /// Only sink or hoist an instruction if it is not a trapping instruction, | ||||||||
1841 | /// or if the instruction is known not to trap when moved to the preheader. | ||||||||
1842 | /// or if it is a trapping instruction and is guaranteed to execute. | ||||||||
1843 | static bool isSafeToExecuteUnconditionally(Instruction &Inst, | ||||||||
1844 | const DominatorTree *DT, | ||||||||
1845 | const TargetLibraryInfo *TLI, | ||||||||
1846 | const Loop *CurLoop, | ||||||||
1847 | const LoopSafetyInfo *SafetyInfo, | ||||||||
1848 | OptimizationRemarkEmitter *ORE, | ||||||||
1849 | const Instruction *CtxI) { | ||||||||
1850 | if (isSafeToSpeculativelyExecute(&Inst, CtxI, DT, TLI)) | ||||||||
1851 | return true; | ||||||||
1852 | |||||||||
1853 | bool GuaranteedToExecute = | ||||||||
1854 | SafetyInfo->isGuaranteedToExecute(Inst, DT, CurLoop); | ||||||||
1855 | |||||||||
1856 | if (!GuaranteedToExecute) { | ||||||||
1857 | auto *LI = dyn_cast<LoadInst>(&Inst); | ||||||||
1858 | if (LI && CurLoop->isLoopInvariant(LI->getPointerOperand())) | ||||||||
1859 | ORE->emit([&]() { | ||||||||
1860 | return OptimizationRemarkMissed( | ||||||||
1861 | DEBUG_TYPE"licm", "LoadWithLoopInvariantAddressCondExecuted", LI) | ||||||||
1862 | << "failed to hoist load with loop-invariant address " | ||||||||
1863 | "because load is conditionally executed"; | ||||||||
1864 | }); | ||||||||
1865 | } | ||||||||
1866 | |||||||||
1867 | return GuaranteedToExecute; | ||||||||
1868 | } | ||||||||
1869 | |||||||||
1870 | namespace { | ||||||||
1871 | class LoopPromoter : public LoadAndStorePromoter { | ||||||||
1872 | Value *SomePtr; // Designated pointer to store to. | ||||||||
1873 | const SmallSetVector<Value *, 8> &PointerMustAliases; | ||||||||
1874 | SmallVectorImpl<BasicBlock *> &LoopExitBlocks; | ||||||||
1875 | SmallVectorImpl<Instruction *> &LoopInsertPts; | ||||||||
1876 | SmallVectorImpl<MemoryAccess *> &MSSAInsertPts; | ||||||||
1877 | PredIteratorCache &PredCache; | ||||||||
1878 | AliasSetTracker *AST; | ||||||||
1879 | MemorySSAUpdater *MSSAU; | ||||||||
1880 | LoopInfo &LI; | ||||||||
1881 | DebugLoc DL; | ||||||||
1882 | int Alignment; | ||||||||
1883 | bool UnorderedAtomic; | ||||||||
1884 | AAMDNodes AATags; | ||||||||
1885 | ICFLoopSafetyInfo &SafetyInfo; | ||||||||
1886 | |||||||||
1887 | // We're about to add a use of V in a loop exit block. Insert an LCSSA phi | ||||||||
1888 | // (if legal) if doing so would add an out-of-loop use to an instruction | ||||||||
1889 | // defined in-loop. | ||||||||
1890 | Value *maybeInsertLCSSAPHI(Value *V, BasicBlock *BB) const { | ||||||||
1891 | if (!LI.wouldBeOutOfLoopUseRequiringLCSSA(V, BB)) | ||||||||
1892 | return V; | ||||||||
1893 | |||||||||
1894 | Instruction *I = cast<Instruction>(V); | ||||||||
1895 | // We need to create an LCSSA PHI node for the incoming value and | ||||||||
1896 | // store that. | ||||||||
1897 | PHINode *PN = PHINode::Create(I->getType(), PredCache.size(BB), | ||||||||
1898 | I->getName() + ".lcssa", &BB->front()); | ||||||||
1899 | for (BasicBlock *Pred : PredCache.get(BB)) | ||||||||
1900 | PN->addIncoming(I, Pred); | ||||||||
1901 | return PN; | ||||||||
1902 | } | ||||||||
1903 | |||||||||
1904 | public: | ||||||||
1905 | LoopPromoter(Value *SP, ArrayRef<const Instruction *> Insts, SSAUpdater &S, | ||||||||
1906 | const SmallSetVector<Value *, 8> &PMA, | ||||||||
1907 | SmallVectorImpl<BasicBlock *> &LEB, | ||||||||
1908 | SmallVectorImpl<Instruction *> &LIP, | ||||||||
1909 | SmallVectorImpl<MemoryAccess *> &MSSAIP, PredIteratorCache &PIC, | ||||||||
1910 | AliasSetTracker *ast, MemorySSAUpdater *MSSAU, LoopInfo &li, | ||||||||
1911 | DebugLoc dl, int alignment, bool UnorderedAtomic, | ||||||||
1912 | const AAMDNodes &AATags, ICFLoopSafetyInfo &SafetyInfo) | ||||||||
1913 | : LoadAndStorePromoter(Insts, S), SomePtr(SP), PointerMustAliases(PMA), | ||||||||
1914 | LoopExitBlocks(LEB), LoopInsertPts(LIP), MSSAInsertPts(MSSAIP), | ||||||||
1915 | PredCache(PIC), AST(ast), MSSAU(MSSAU), LI(li), DL(std::move(dl)), | ||||||||
1916 | Alignment(alignment), UnorderedAtomic(UnorderedAtomic), AATags(AATags), | ||||||||
1917 | SafetyInfo(SafetyInfo) {} | ||||||||
1918 | |||||||||
1919 | bool isInstInList(Instruction *I, | ||||||||
1920 | const SmallVectorImpl<Instruction *> &) const override { | ||||||||
1921 | Value *Ptr; | ||||||||
1922 | if (LoadInst *LI = dyn_cast<LoadInst>(I)) | ||||||||
1923 | Ptr = LI->getOperand(0); | ||||||||
1924 | else | ||||||||
1925 | Ptr = cast<StoreInst>(I)->getPointerOperand(); | ||||||||
1926 | return PointerMustAliases.count(Ptr); | ||||||||
1927 | } | ||||||||
1928 | |||||||||
1929 | void doExtraRewritesBeforeFinalDeletion() override { | ||||||||
1930 | // Insert stores after in the loop exit blocks. Each exit block gets a | ||||||||
1931 | // store of the live-out values that feed them. Since we've already told | ||||||||
1932 | // the SSA updater about the defs in the loop and the preheader | ||||||||
1933 | // definition, it is all set and we can start using it. | ||||||||
1934 | for (unsigned i = 0, e = LoopExitBlocks.size(); i != e; ++i) { | ||||||||
1935 | BasicBlock *ExitBlock = LoopExitBlocks[i]; | ||||||||
1936 | Value *LiveInValue = SSA.GetValueInMiddleOfBlock(ExitBlock); | ||||||||
1937 | LiveInValue = maybeInsertLCSSAPHI(LiveInValue, ExitBlock); | ||||||||
1938 | Value *Ptr = maybeInsertLCSSAPHI(SomePtr, ExitBlock); | ||||||||
1939 | Instruction *InsertPos = LoopInsertPts[i]; | ||||||||
1940 | StoreInst *NewSI = new StoreInst(LiveInValue, Ptr, InsertPos); | ||||||||
1941 | if (UnorderedAtomic) | ||||||||
1942 | NewSI->setOrdering(AtomicOrdering::Unordered); | ||||||||
1943 | NewSI->setAlignment(Align(Alignment)); | ||||||||
1944 | NewSI->setDebugLoc(DL); | ||||||||
1945 | if (AATags) | ||||||||
1946 | NewSI->setAAMetadata(AATags); | ||||||||
1947 | |||||||||
1948 | if (MSSAU) { | ||||||||
1949 | MemoryAccess *MSSAInsertPoint = MSSAInsertPts[i]; | ||||||||
1950 | MemoryAccess *NewMemAcc; | ||||||||
1951 | if (!MSSAInsertPoint) { | ||||||||
1952 | NewMemAcc = MSSAU->createMemoryAccessInBB( | ||||||||
1953 | NewSI, nullptr, NewSI->getParent(), MemorySSA::Beginning); | ||||||||
1954 | } else { | ||||||||
1955 | NewMemAcc = | ||||||||
1956 | MSSAU->createMemoryAccessAfter(NewSI, nullptr, MSSAInsertPoint); | ||||||||
1957 | } | ||||||||
1958 | MSSAInsertPts[i] = NewMemAcc; | ||||||||
1959 | MSSAU->insertDef(cast<MemoryDef>(NewMemAcc), true); | ||||||||
1960 | // FIXME: true for safety, false may still be correct. | ||||||||
1961 | } | ||||||||
1962 | } | ||||||||
1963 | } | ||||||||
1964 | |||||||||
1965 | void replaceLoadWithValue(LoadInst *LI, Value *V) const override { | ||||||||
1966 | // Update alias analysis. | ||||||||
1967 | if (AST) | ||||||||
1968 | AST->copyValue(LI, V); | ||||||||
1969 | } | ||||||||
1970 | void instructionDeleted(Instruction *I) const override { | ||||||||
1971 | SafetyInfo.removeInstruction(I); | ||||||||
1972 | if (AST) | ||||||||
1973 | AST->deleteValue(I); | ||||||||
1974 | if (MSSAU) | ||||||||
1975 | MSSAU->removeMemoryAccess(I); | ||||||||
1976 | } | ||||||||
1977 | }; | ||||||||
1978 | |||||||||
1979 | bool isNotCapturedBeforeOrInLoop(const Value *V, const Loop *L, | ||||||||
1980 | DominatorTree *DT) { | ||||||||
1981 | // We can perform the captured-before check against any instruction in the | ||||||||
1982 | // loop header, as the loop header is reachable from any instruction inside | ||||||||
1983 | // the loop. | ||||||||
1984 | // TODO: ReturnCaptures=true shouldn't be necessary here. | ||||||||
1985 | return !PointerMayBeCapturedBefore(V, /* ReturnCaptures */ true, | ||||||||
1986 | /* StoreCaptures */ true, | ||||||||
1987 | L->getHeader()->getTerminator(), DT); | ||||||||
1988 | } | ||||||||
1989 | |||||||||
1990 | /// Return true iff we can prove that a caller of this function can not inspect | ||||||||
1991 | /// the contents of the provided object in a well defined program. | ||||||||
1992 | bool isKnownNonEscaping(Value *Object, const Loop *L, | ||||||||
1993 | const TargetLibraryInfo *TLI, DominatorTree *DT) { | ||||||||
1994 | if (isa<AllocaInst>(Object)) | ||||||||
1995 | // Since the alloca goes out of scope, we know the caller can't retain a | ||||||||
1996 | // reference to it and be well defined. Thus, we don't need to check for | ||||||||
1997 | // capture. | ||||||||
1998 | return true; | ||||||||
1999 | |||||||||
2000 | // For all other objects we need to know that the caller can't possibly | ||||||||
2001 | // have gotten a reference to the object. There are two components of | ||||||||
2002 | // that: | ||||||||
2003 | // 1) Object can't be escaped by this function. This is what | ||||||||
2004 | // PointerMayBeCaptured checks. | ||||||||
2005 | // 2) Object can't have been captured at definition site. For this, we | ||||||||
2006 | // need to know the return value is noalias. At the moment, we use a | ||||||||
2007 | // weaker condition and handle only AllocLikeFunctions (which are | ||||||||
2008 | // known to be noalias). TODO | ||||||||
2009 | return isAllocLikeFn(Object, TLI) && | ||||||||
2010 | isNotCapturedBeforeOrInLoop(Object, L, DT); | ||||||||
2011 | } | ||||||||
2012 | |||||||||
2013 | } // namespace | ||||||||
2014 | |||||||||
2015 | /// Try to promote memory values to scalars by sinking stores out of the | ||||||||
2016 | /// loop and moving loads to before the loop. We do this by looping over | ||||||||
2017 | /// the stores in the loop, looking for stores to Must pointers which are | ||||||||
2018 | /// loop invariant. | ||||||||
2019 | /// | ||||||||
2020 | bool llvm::promoteLoopAccessesToScalars( | ||||||||
2021 | const SmallSetVector<Value *, 8> &PointerMustAliases, | ||||||||
2022 | SmallVectorImpl<BasicBlock *> &ExitBlocks, | ||||||||
2023 | SmallVectorImpl<Instruction *> &InsertPts, | ||||||||
2024 | SmallVectorImpl<MemoryAccess *> &MSSAInsertPts, PredIteratorCache &PIC, | ||||||||
2025 | LoopInfo *LI, DominatorTree *DT, const TargetLibraryInfo *TLI, | ||||||||
2026 | Loop *CurLoop, AliasSetTracker *CurAST, MemorySSAUpdater *MSSAU, | ||||||||
2027 | ICFLoopSafetyInfo *SafetyInfo, OptimizationRemarkEmitter *ORE) { | ||||||||
2028 | // Verify inputs. | ||||||||
2029 | assert(LI != nullptr && DT != nullptr && CurLoop != nullptr &&((void)0) | ||||||||
2030 | SafetyInfo != nullptr &&((void)0) | ||||||||
2031 | "Unexpected Input to promoteLoopAccessesToScalars")((void)0); | ||||||||
2032 | |||||||||
2033 | Value *SomePtr = *PointerMustAliases.begin(); | ||||||||
2034 | BasicBlock *Preheader = CurLoop->getLoopPreheader(); | ||||||||
2035 | |||||||||
2036 | // It is not safe to promote a load/store from the loop if the load/store is | ||||||||
2037 | // conditional. For example, turning: | ||||||||
2038 | // | ||||||||
2039 | // for () { if (c) *P += 1; } | ||||||||
2040 | // | ||||||||
2041 | // into: | ||||||||
2042 | // | ||||||||
2043 | // tmp = *P; for () { if (c) tmp +=1; } *P = tmp; | ||||||||
2044 | // | ||||||||
2045 | // is not safe, because *P may only be valid to access if 'c' is true. | ||||||||
2046 | // | ||||||||
2047 | // The safety property divides into two parts: | ||||||||
2048 | // p1) The memory may not be dereferenceable on entry to the loop. In this | ||||||||
2049 | // case, we can't insert the required load in the preheader. | ||||||||
2050 | // p2) The memory model does not allow us to insert a store along any dynamic | ||||||||
2051 | // path which did not originally have one. | ||||||||
2052 | // | ||||||||
2053 | // If at least one store is guaranteed to execute, both properties are | ||||||||
2054 | // satisfied, and promotion is legal. | ||||||||
2055 | // | ||||||||
2056 | // This, however, is not a necessary condition. Even if no store/load is | ||||||||
2057 | // guaranteed to execute, we can still establish these properties. | ||||||||
2058 | // We can establish (p1) by proving that hoisting the load into the preheader | ||||||||
2059 | // is safe (i.e. proving dereferenceability on all paths through the loop). We | ||||||||
2060 | // can use any access within the alias set to prove dereferenceability, | ||||||||
2061 | // since they're all must alias. | ||||||||
2062 | // | ||||||||
2063 | // There are two ways establish (p2): | ||||||||
2064 | // a) Prove the location is thread-local. In this case the memory model | ||||||||
2065 | // requirement does not apply, and stores are safe to insert. | ||||||||
2066 | // b) Prove a store dominates every exit block. In this case, if an exit | ||||||||
2067 | // blocks is reached, the original dynamic path would have taken us through | ||||||||
2068 | // the store, so inserting a store into the exit block is safe. Note that this | ||||||||
2069 | // is different from the store being guaranteed to execute. For instance, | ||||||||
2070 | // if an exception is thrown on the first iteration of the loop, the original | ||||||||
2071 | // store is never executed, but the exit blocks are not executed either. | ||||||||
2072 | |||||||||
2073 | bool DereferenceableInPH = false; | ||||||||
2074 | bool SafeToInsertStore = false; | ||||||||
2075 | |||||||||
2076 | SmallVector<Instruction *, 64> LoopUses; | ||||||||
2077 | |||||||||
2078 | // We start with an alignment of one and try to find instructions that allow | ||||||||
2079 | // us to prove better alignment. | ||||||||
2080 | Align Alignment; | ||||||||
2081 | // Keep track of which types of access we see | ||||||||
2082 | bool SawUnorderedAtomic = false; | ||||||||
2083 | bool SawNotAtomic = false; | ||||||||
2084 | AAMDNodes AATags; | ||||||||
2085 | |||||||||
2086 | const DataLayout &MDL = Preheader->getModule()->getDataLayout(); | ||||||||
2087 | |||||||||
2088 | bool IsKnownThreadLocalObject = false; | ||||||||
2089 | if (SafetyInfo->anyBlockMayThrow()) { | ||||||||
2090 | // If a loop can throw, we have to insert a store along each unwind edge. | ||||||||
2091 | // That said, we can't actually make the unwind edge explicit. Therefore, | ||||||||
2092 | // we have to prove that the store is dead along the unwind edge. We do | ||||||||
2093 | // this by proving that the caller can't have a reference to the object | ||||||||
2094 | // after return and thus can't possibly load from the object. | ||||||||
2095 | Value *Object = getUnderlyingObject(SomePtr); | ||||||||
2096 | if (!isKnownNonEscaping(Object, CurLoop, TLI, DT)) | ||||||||
2097 | return false; | ||||||||
2098 | // Subtlety: Alloca's aren't visible to callers, but *are* potentially | ||||||||
2099 | // visible to other threads if captured and used during their lifetimes. | ||||||||
2100 | IsKnownThreadLocalObject = !isa<AllocaInst>(Object); | ||||||||
2101 | } | ||||||||
2102 | |||||||||
2103 | // Check that all of the pointers in the alias set have the same type. We | ||||||||
2104 | // cannot (yet) promote a memory location that is loaded and stored in | ||||||||
2105 | // different sizes. While we are at it, collect alignment and AA info. | ||||||||
2106 | for (Value *ASIV : PointerMustAliases) { | ||||||||
2107 | // Check that all of the pointers in the alias set have the same type. We | ||||||||
2108 | // cannot (yet) promote a memory location that is loaded and stored in | ||||||||
2109 | // different sizes. | ||||||||
2110 | if (SomePtr->getType() != ASIV->getType()) | ||||||||
2111 | return false; | ||||||||
2112 | |||||||||
2113 | for (User *U : ASIV->users()) { | ||||||||
2114 | // Ignore instructions that are outside the loop. | ||||||||
2115 | Instruction *UI = dyn_cast<Instruction>(U); | ||||||||
2116 | if (!UI || !CurLoop->contains(UI)) | ||||||||
2117 | continue; | ||||||||
2118 | |||||||||
2119 | // If there is an non-load/store instruction in the loop, we can't promote | ||||||||
2120 | // it. | ||||||||
2121 | if (LoadInst *Load = dyn_cast<LoadInst>(UI)) { | ||||||||
2122 | if (!Load->isUnordered()) | ||||||||
2123 | return false; | ||||||||
2124 | |||||||||
2125 | SawUnorderedAtomic |= Load->isAtomic(); | ||||||||
2126 | SawNotAtomic |= !Load->isAtomic(); | ||||||||
2127 | |||||||||
2128 | Align InstAlignment = Load->getAlign(); | ||||||||
2129 | |||||||||
2130 | // Note that proving a load safe to speculate requires proving | ||||||||
2131 | // sufficient alignment at the target location. Proving it guaranteed | ||||||||
2132 | // to execute does as well. Thus we can increase our guaranteed | ||||||||
2133 | // alignment as well. | ||||||||
2134 | if (!DereferenceableInPH || (InstAlignment > Alignment)) | ||||||||
2135 | if (isSafeToExecuteUnconditionally(*Load, DT, TLI, CurLoop, | ||||||||
2136 | SafetyInfo, ORE, | ||||||||
2137 | Preheader->getTerminator())) { | ||||||||
2138 | DereferenceableInPH = true; | ||||||||
2139 | Alignment = std::max(Alignment, InstAlignment); | ||||||||
2140 | } | ||||||||
2141 | } else if (const StoreInst *Store = dyn_cast<StoreInst>(UI)) { | ||||||||
2142 | // Stores *of* the pointer are not interesting, only stores *to* the | ||||||||
2143 | // pointer. | ||||||||
2144 | if (UI->getOperand(1) != ASIV) | ||||||||
2145 | continue; | ||||||||
2146 | if (!Store->isUnordered()) | ||||||||
2147 | return false; | ||||||||
2148 | |||||||||
2149 | SawUnorderedAtomic |= Store->isAtomic(); | ||||||||
2150 | SawNotAtomic |= !Store->isAtomic(); | ||||||||
2151 | |||||||||
2152 | // If the store is guaranteed to execute, both properties are satisfied. | ||||||||
2153 | // We may want to check if a store is guaranteed to execute even if we | ||||||||
2154 | // already know that promotion is safe, since it may have higher | ||||||||
2155 | // alignment than any other guaranteed stores, in which case we can | ||||||||
2156 | // raise the alignment on the promoted store. | ||||||||
2157 | Align InstAlignment = Store->getAlign(); | ||||||||
2158 | |||||||||
2159 | if (!DereferenceableInPH || !SafeToInsertStore || | ||||||||
2160 | (InstAlignment > Alignment)) { | ||||||||
2161 | if (SafetyInfo->isGuaranteedToExecute(*UI, DT, CurLoop)) { | ||||||||
2162 | DereferenceableInPH = true; | ||||||||
2163 | SafeToInsertStore = true; | ||||||||
2164 | Alignment = std::max(Alignment, InstAlignment); | ||||||||
2165 | } | ||||||||
2166 | } | ||||||||
2167 | |||||||||
2168 | // If a store dominates all exit blocks, it is safe to sink. | ||||||||
2169 | // As explained above, if an exit block was executed, a dominating | ||||||||
2170 | // store must have been executed at least once, so we are not | ||||||||
2171 | // introducing stores on paths that did not have them. | ||||||||
2172 | // Note that this only looks at explicit exit blocks. If we ever | ||||||||
2173 | // start sinking stores into unwind edges (see above), this will break. | ||||||||
2174 | if (!SafeToInsertStore) | ||||||||
2175 | SafeToInsertStore = llvm::all_of(ExitBlocks, [&](BasicBlock *Exit) { | ||||||||
2176 | return DT->dominates(Store->getParent(), Exit); | ||||||||
2177 | }); | ||||||||
2178 | |||||||||
2179 | // If the store is not guaranteed to execute, we may still get | ||||||||
2180 | // deref info through it. | ||||||||
2181 | if (!DereferenceableInPH) { | ||||||||
2182 | DereferenceableInPH = isDereferenceableAndAlignedPointer( | ||||||||
2183 | Store->getPointerOperand(), Store->getValueOperand()->getType(), | ||||||||
2184 | Store->getAlign(), MDL, Preheader->getTerminator(), DT, TLI); | ||||||||
2185 | } | ||||||||
2186 | } else | ||||||||
2187 | return false; // Not a load or store. | ||||||||
2188 | |||||||||
2189 | // Merge the AA tags. | ||||||||
2190 | if (LoopUses.empty()) { | ||||||||
2191 | // On the first load/store, just take its AA tags. | ||||||||
2192 | UI->getAAMetadata(AATags); | ||||||||
2193 | } else if (AATags) { | ||||||||
2194 | UI->getAAMetadata(AATags, /* Merge = */ true); | ||||||||
2195 | } | ||||||||
2196 | |||||||||
2197 | LoopUses.push_back(UI); | ||||||||
2198 | } | ||||||||
2199 | } | ||||||||
2200 | |||||||||
2201 | // If we found both an unordered atomic instruction and a non-atomic memory | ||||||||
2202 | // access, bail. We can't blindly promote non-atomic to atomic since we | ||||||||
2203 | // might not be able to lower the result. We can't downgrade since that | ||||||||
2204 | // would violate memory model. Also, align 0 is an error for atomics. | ||||||||
2205 | if (SawUnorderedAtomic && SawNotAtomic) | ||||||||
2206 | return false; | ||||||||
2207 | |||||||||
2208 | // If we're inserting an atomic load in the preheader, we must be able to | ||||||||
2209 | // lower it. We're only guaranteed to be able to lower naturally aligned | ||||||||
2210 | // atomics. | ||||||||
2211 | auto *SomePtrElemType = SomePtr->getType()->getPointerElementType(); | ||||||||
2212 | if (SawUnorderedAtomic && | ||||||||
2213 | Alignment < MDL.getTypeStoreSize(SomePtrElemType)) | ||||||||
2214 | return false; | ||||||||
2215 | |||||||||
2216 | // If we couldn't prove we can hoist the load, bail. | ||||||||
2217 | if (!DereferenceableInPH) | ||||||||
2218 | return false; | ||||||||
2219 | |||||||||
2220 | // We know we can hoist the load, but don't have a guaranteed store. | ||||||||
2221 | // Check whether the location is thread-local. If it is, then we can insert | ||||||||
2222 | // stores along paths which originally didn't have them without violating the | ||||||||
2223 | // memory model. | ||||||||
2224 | if (!SafeToInsertStore) { | ||||||||
2225 | if (IsKnownThreadLocalObject) | ||||||||
2226 | SafeToInsertStore = true; | ||||||||
2227 | else { | ||||||||
2228 | Value *Object = getUnderlyingObject(SomePtr); | ||||||||
2229 | SafeToInsertStore = | ||||||||
2230 | (isAllocLikeFn(Object, TLI) || isa<AllocaInst>(Object)) && | ||||||||
2231 | isNotCapturedBeforeOrInLoop(Object, CurLoop, DT); | ||||||||
2232 | } | ||||||||
2233 | } | ||||||||
2234 | |||||||||
2235 | // If we've still failed to prove we can sink the store, give up. | ||||||||
2236 | if (!SafeToInsertStore) | ||||||||
2237 | return false; | ||||||||
2238 | |||||||||
2239 | // Otherwise, this is safe to promote, lets do it! | ||||||||
2240 | LLVM_DEBUG(dbgs() << "LICM: Promoting value stored to in loop: " << *SomePtrdo { } while (false) | ||||||||
2241 | << '\n')do { } while (false); | ||||||||
2242 | ORE->emit([&]() { | ||||||||
2243 | return OptimizationRemark(DEBUG_TYPE"licm", "PromoteLoopAccessesToScalar", | ||||||||
2244 | LoopUses[0]) | ||||||||
2245 | << "Moving accesses to memory location out of the loop"; | ||||||||
2246 | }); | ||||||||
2247 | ++NumPromoted; | ||||||||
2248 | |||||||||
2249 | // Look at all the loop uses, and try to merge their locations. | ||||||||
2250 | std::vector<const DILocation *> LoopUsesLocs; | ||||||||
2251 | for (auto U : LoopUses) | ||||||||
2252 | LoopUsesLocs.push_back(U->getDebugLoc().get()); | ||||||||
2253 | auto DL = DebugLoc(DILocation::getMergedLocations(LoopUsesLocs)); | ||||||||
2254 | |||||||||
2255 | // We use the SSAUpdater interface to insert phi nodes as required. | ||||||||
2256 | SmallVector<PHINode *, 16> NewPHIs; | ||||||||
2257 | SSAUpdater SSA(&NewPHIs); | ||||||||
2258 | LoopPromoter Promoter(SomePtr, LoopUses, SSA, PointerMustAliases, ExitBlocks, | ||||||||
2259 | InsertPts, MSSAInsertPts, PIC, CurAST, MSSAU, *LI, DL, | ||||||||
2260 | Alignment.value(), SawUnorderedAtomic, AATags, | ||||||||
2261 | *SafetyInfo); | ||||||||
2262 | |||||||||
2263 | // Set up the preheader to have a definition of the value. It is the live-out | ||||||||
2264 | // value from the preheader that uses in the loop will use. | ||||||||
2265 | LoadInst *PreheaderLoad = new LoadInst( | ||||||||
2266 | SomePtr->getType()->getPointerElementType(), SomePtr, | ||||||||
2267 | SomePtr->getName() + ".promoted", Preheader->getTerminator()); | ||||||||
2268 | if (SawUnorderedAtomic) | ||||||||
2269 | PreheaderLoad->setOrdering(AtomicOrdering::Unordered); | ||||||||
2270 | PreheaderLoad->setAlignment(Alignment); | ||||||||
2271 | PreheaderLoad->setDebugLoc(DebugLoc()); | ||||||||
2272 | if (AATags) | ||||||||
2273 | PreheaderLoad->setAAMetadata(AATags); | ||||||||
2274 | SSA.AddAvailableValue(Preheader, PreheaderLoad); | ||||||||
2275 | |||||||||
2276 | if (MSSAU) { | ||||||||
2277 | MemoryAccess *PreheaderLoadMemoryAccess = MSSAU->createMemoryAccessInBB( | ||||||||
2278 | PreheaderLoad, nullptr, PreheaderLoad->getParent(), MemorySSA::End); | ||||||||
2279 | MemoryUse *NewMemUse = cast<MemoryUse>(PreheaderLoadMemoryAccess); | ||||||||
2280 | MSSAU->insertUse(NewMemUse, /*RenameUses=*/true); | ||||||||
2281 | } | ||||||||
2282 | |||||||||
2283 | if (MSSAU && VerifyMemorySSA) | ||||||||
2284 | MSSAU->getMemorySSA()->verifyMemorySSA(); | ||||||||
2285 | // Rewrite all the loads in the loop and remember all the definitions from | ||||||||
2286 | // stores in the loop. | ||||||||
2287 | Promoter.run(LoopUses); | ||||||||
2288 | |||||||||
2289 | if (MSSAU && VerifyMemorySSA) | ||||||||
2290 | MSSAU->getMemorySSA()->verifyMemorySSA(); | ||||||||
2291 | // If the SSAUpdater didn't use the load in the preheader, just zap it now. | ||||||||
2292 | if (PreheaderLoad->use_empty()) | ||||||||
2293 | eraseInstruction(*PreheaderLoad, *SafetyInfo, CurAST, MSSAU); | ||||||||
2294 | |||||||||
2295 | return true; | ||||||||
2296 | } | ||||||||
2297 | |||||||||
2298 | static void foreachMemoryAccess(MemorySSA *MSSA, Loop *L, | ||||||||
2299 | function_ref<void(Instruction *)> Fn) { | ||||||||
2300 | for (const BasicBlock *BB : L->blocks()) | ||||||||
2301 | if (const auto *Accesses = MSSA->getBlockAccesses(BB)) | ||||||||
2302 | for (const auto &Access : *Accesses) | ||||||||
2303 | if (const auto *MUD = dyn_cast<MemoryUseOrDef>(&Access)) | ||||||||
2304 | Fn(MUD->getMemoryInst()); | ||||||||
2305 | } | ||||||||
2306 | |||||||||
2307 | static SmallVector<SmallSetVector<Value *, 8>, 0> | ||||||||
2308 | collectPromotionCandidates(MemorySSA *MSSA, AliasAnalysis *AA, Loop *L) { | ||||||||
2309 | AliasSetTracker AST(*AA); | ||||||||
2310 | |||||||||
2311 | auto IsPotentiallyPromotable = [L](const Instruction *I) { | ||||||||
2312 | if (const auto *SI = dyn_cast<StoreInst>(I)) | ||||||||
2313 | return L->isLoopInvariant(SI->getPointerOperand()); | ||||||||
2314 | if (const auto *LI = dyn_cast<LoadInst>(I)) | ||||||||
2315 | return L->isLoopInvariant(LI->getPointerOperand()); | ||||||||
2316 | return false; | ||||||||
2317 | }; | ||||||||
2318 | |||||||||
2319 | // Populate AST with potentially promotable accesses and remove them from | ||||||||
2320 | // MaybePromotable, so they will not be checked again on the next iteration. | ||||||||
2321 | SmallPtrSet<Value *, 16> AttemptingPromotion; | ||||||||
2322 | foreachMemoryAccess(MSSA, L, [&](Instruction *I) { | ||||||||
2323 | if (IsPotentiallyPromotable(I)) { | ||||||||
2324 | AttemptingPromotion.insert(I); | ||||||||
2325 | AST.add(I); | ||||||||
2326 | } | ||||||||
2327 | }); | ||||||||
2328 | |||||||||
2329 | // We're only interested in must-alias sets that contain a mod. | ||||||||
2330 | SmallVector<const AliasSet *, 8> Sets; | ||||||||
2331 | for (AliasSet &AS : AST) | ||||||||
2332 | if (!AS.isForwardingAliasSet() && AS.isMod() && AS.isMustAlias()) | ||||||||
2333 | Sets.push_back(&AS); | ||||||||
2334 | |||||||||
2335 | if (Sets.empty()) | ||||||||
2336 | return {}; // Nothing to promote... | ||||||||
2337 | |||||||||
2338 | // Discard any sets for which there is an aliasing non-promotable access. | ||||||||
2339 | foreachMemoryAccess(MSSA, L, [&](Instruction *I) { | ||||||||
2340 | if (AttemptingPromotion.contains(I)) | ||||||||
2341 | return; | ||||||||
2342 | |||||||||
2343 | llvm::erase_if(Sets, [&](const AliasSet *AS) { | ||||||||
2344 | return AS->aliasesUnknownInst(I, *AA); | ||||||||
2345 | }); | ||||||||
2346 | }); | ||||||||
2347 | |||||||||
2348 | SmallVector<SmallSetVector<Value *, 8>, 0> Result; | ||||||||
2349 | for (const AliasSet *Set : Sets) { | ||||||||
2350 | SmallSetVector<Value *, 8> PointerMustAliases; | ||||||||
2351 | for (const auto &ASI : *Set) | ||||||||
2352 | PointerMustAliases.insert(ASI.getValue()); | ||||||||
2353 | Result.push_back(std::move(PointerMustAliases)); | ||||||||
2354 | } | ||||||||
2355 | |||||||||
2356 | return Result; | ||||||||
2357 | } | ||||||||
2358 | |||||||||
2359 | /// Returns an owning pointer to an alias set which incorporates aliasing info | ||||||||
2360 | /// from L and all subloops of L. | ||||||||
2361 | std::unique_ptr<AliasSetTracker> | ||||||||
2362 | LoopInvariantCodeMotion::collectAliasInfoForLoop(Loop *L, LoopInfo *LI, | ||||||||
2363 | AAResults *AA) { | ||||||||
2364 | auto CurAST = std::make_unique<AliasSetTracker>(*AA); | ||||||||
2365 | |||||||||
2366 | // Add everything from all the sub loops. | ||||||||
2367 | for (Loop *InnerL : L->getSubLoops()) | ||||||||
2368 | for (BasicBlock *BB : InnerL->blocks()) | ||||||||
2369 | CurAST->add(*BB); | ||||||||
2370 | |||||||||
2371 | // And merge in this loop (without anything from inner loops). | ||||||||
2372 | for (BasicBlock *BB : L->blocks()) | ||||||||
2373 | if (LI->getLoopFor(BB) == L) | ||||||||
2374 | CurAST->add(*BB); | ||||||||
2375 | |||||||||
2376 | return CurAST; | ||||||||
2377 | } | ||||||||
2378 | |||||||||
2379 | static bool pointerInvalidatedByLoop(MemoryLocation MemLoc, | ||||||||
2380 | AliasSetTracker *CurAST, Loop *CurLoop, | ||||||||
2381 | AAResults *AA) { | ||||||||
2382 | // First check to see if any of the basic blocks in CurLoop invalidate *V. | ||||||||
2383 | bool isInvalidatedAccordingToAST = CurAST->getAliasSetFor(MemLoc).isMod(); | ||||||||
2384 | |||||||||
2385 | if (!isInvalidatedAccordingToAST || !LICMN2Theshold) | ||||||||
2386 | return isInvalidatedAccordingToAST; | ||||||||
2387 | |||||||||
2388 | // Check with a diagnostic analysis if we can refine the information above. | ||||||||
2389 | // This is to identify the limitations of using the AST. | ||||||||
2390 | // The alias set mechanism used by LICM has a major weakness in that it | ||||||||
2391 | // combines all things which may alias into a single set *before* asking | ||||||||
2392 | // modref questions. As a result, a single readonly call within a loop will | ||||||||
2393 | // collapse all loads and stores into a single alias set and report | ||||||||
2394 | // invalidation if the loop contains any store. For example, readonly calls | ||||||||
2395 | // with deopt states have this form and create a general alias set with all | ||||||||
2396 | // loads and stores. In order to get any LICM in loops containing possible | ||||||||
2397 | // deopt states we need a more precise invalidation of checking the mod ref | ||||||||
2398 | // info of each instruction within the loop and LI. This has a complexity of | ||||||||
2399 | // O(N^2), so currently, it is used only as a diagnostic tool since the | ||||||||
2400 | // default value of LICMN2Threshold is zero. | ||||||||
2401 | |||||||||
2402 | // Don't look at nested loops. | ||||||||
2403 | if (CurLoop->begin() != CurLoop->end()) | ||||||||
2404 | return true; | ||||||||
2405 | |||||||||
2406 | int N = 0; | ||||||||
2407 | for (BasicBlock *BB : CurLoop->getBlocks()) | ||||||||
2408 | for (Instruction &I : *BB) { | ||||||||
2409 | if (N >= LICMN2Theshold) { | ||||||||
2410 | LLVM_DEBUG(dbgs() << "Alasing N2 threshold exhausted for "do { } while (false) | ||||||||
2411 | << *(MemLoc.Ptr) << "\n")do { } while (false); | ||||||||
2412 | return true; | ||||||||
2413 | } | ||||||||
2414 | N++; | ||||||||
2415 | auto Res = AA->getModRefInfo(&I, MemLoc); | ||||||||
2416 | if (isModSet(Res)) { | ||||||||
2417 | LLVM_DEBUG(dbgs() << "Aliasing failed on " << I << " for "do { } while (false) | ||||||||
2418 | << *(MemLoc.Ptr) << "\n")do { } while (false); | ||||||||
2419 | return true; | ||||||||
2420 | } | ||||||||
2421 | } | ||||||||
2422 | LLVM_DEBUG(dbgs() << "Aliasing okay for " << *(MemLoc.Ptr) << "\n")do { } while (false); | ||||||||
2423 | return false; | ||||||||
2424 | } | ||||||||
2425 | |||||||||
2426 | bool pointerInvalidatedByLoopWithMSSA(MemorySSA *MSSA, MemoryUse *MU, | ||||||||
2427 | Loop *CurLoop, Instruction &I, | ||||||||
2428 | SinkAndHoistLICMFlags &Flags) { | ||||||||
2429 | // For hoisting, use the walker to determine safety | ||||||||
2430 | if (!Flags.getIsSink()) { | ||||||||
2431 | MemoryAccess *Source; | ||||||||
2432 | // See declaration of SetLicmMssaOptCap for usage details. | ||||||||
2433 | if (Flags.tooManyClobberingCalls()) | ||||||||
2434 | Source = MU->getDefiningAccess(); | ||||||||
2435 | else { | ||||||||
2436 | Source = MSSA->getSkipSelfWalker()->getClobberingMemoryAccess(MU); | ||||||||
2437 | Flags.incrementClobberingCalls(); | ||||||||
2438 | } | ||||||||
2439 | return !MSSA->isLiveOnEntryDef(Source) && | ||||||||
2440 | CurLoop->contains(Source->getBlock()); | ||||||||
2441 | } | ||||||||
2442 | |||||||||
2443 | // For sinking, we'd need to check all Defs below this use. The getClobbering | ||||||||
2444 | // call will look on the backedge of the loop, but will check aliasing with | ||||||||
2445 | // the instructions on the previous iteration. | ||||||||
2446 | // For example: | ||||||||
2447 | // for (i ... ) | ||||||||
2448 | // load a[i] ( Use (LoE) | ||||||||
2449 | // store a[i] ( 1 = Def (2), with 2 = Phi for the loop. | ||||||||
2450 | // i++; | ||||||||
2451 | // The load sees no clobbering inside the loop, as the backedge alias check | ||||||||
2452 | // does phi translation, and will check aliasing against store a[i-1]. | ||||||||
2453 | // However sinking the load outside the loop, below the store is incorrect. | ||||||||
2454 | |||||||||
2455 | // For now, only sink if there are no Defs in the loop, and the existing ones | ||||||||
2456 | // precede the use and are in the same block. | ||||||||
2457 | // FIXME: Increase precision: Safe to sink if Use post dominates the Def; | ||||||||
2458 | // needs PostDominatorTreeAnalysis. | ||||||||
2459 | // FIXME: More precise: no Defs that alias this Use. | ||||||||
2460 | if (Flags.tooManyMemoryAccesses()) | ||||||||
2461 | return true; | ||||||||
2462 | for (auto *BB : CurLoop->getBlocks()) | ||||||||
2463 | if (pointerInvalidatedByBlockWithMSSA(*BB, *MSSA, *MU)) | ||||||||
2464 | return true; | ||||||||
2465 | // When sinking, the source block may not be part of the loop so check it. | ||||||||
2466 | if (!CurLoop->contains(&I)) | ||||||||
2467 | return pointerInvalidatedByBlockWithMSSA(*I.getParent(), *MSSA, *MU); | ||||||||
2468 | |||||||||
2469 | return false; | ||||||||
2470 | } | ||||||||
2471 | |||||||||
2472 | bool pointerInvalidatedByBlockWithMSSA(BasicBlock &BB, MemorySSA &MSSA, | ||||||||
2473 | MemoryUse &MU) { | ||||||||
2474 | if (const auto *Accesses = MSSA.getBlockDefs(&BB)) | ||||||||
2475 | for (const auto &MA : *Accesses) | ||||||||
2476 | if (const auto *MD = dyn_cast<MemoryDef>(&MA)) | ||||||||
2477 | if (MU.getBlock() != MD->getBlock() || !MSSA.locallyDominates(MD, &MU)) | ||||||||
2478 | return true; | ||||||||
2479 | return false; | ||||||||
2480 | } | ||||||||
2481 | |||||||||
2482 | /// Little predicate that returns true if the specified basic block is in | ||||||||
2483 | /// a subloop of the current one, not the current one itself. | ||||||||
2484 | /// | ||||||||
2485 | static bool inSubLoop(BasicBlock *BB, Loop *CurLoop, LoopInfo *LI) { | ||||||||
2486 | assert(CurLoop->contains(BB) && "Only valid if BB is IN the loop")((void)0); | ||||||||
2487 | return LI->getLoopFor(BB) != CurLoop; | ||||||||
2488 | } |
1 | //===- llvm/InstrTypes.h - Important Instruction subclasses -----*- C++ -*-===// |
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 file defines various meta classes of instructions that exist in the VM |
10 | // representation. Specific concrete subclasses of these may be found in the |
11 | // i*.h files... |
12 | // |
13 | //===----------------------------------------------------------------------===// |
14 | |
15 | #ifndef LLVM_IR_INSTRTYPES_H |
16 | #define LLVM_IR_INSTRTYPES_H |
17 | |
18 | #include "llvm/ADT/ArrayRef.h" |
19 | #include "llvm/ADT/None.h" |
20 | #include "llvm/ADT/Optional.h" |
21 | #include "llvm/ADT/STLExtras.h" |
22 | #include "llvm/ADT/StringMap.h" |
23 | #include "llvm/ADT/StringRef.h" |
24 | #include "llvm/ADT/Twine.h" |
25 | #include "llvm/ADT/iterator_range.h" |
26 | #include "llvm/IR/Attributes.h" |
27 | #include "llvm/IR/CallingConv.h" |
28 | #include "llvm/IR/Constants.h" |
29 | #include "llvm/IR/DerivedTypes.h" |
30 | #include "llvm/IR/Function.h" |
31 | #include "llvm/IR/Instruction.h" |
32 | #include "llvm/IR/LLVMContext.h" |
33 | #include "llvm/IR/OperandTraits.h" |
34 | #include "llvm/IR/Type.h" |
35 | #include "llvm/IR/User.h" |
36 | #include "llvm/IR/Value.h" |
37 | #include "llvm/Support/Casting.h" |
38 | #include "llvm/Support/ErrorHandling.h" |
39 | #include <algorithm> |
40 | #include <cassert> |
41 | #include <cstddef> |
42 | #include <cstdint> |
43 | #include <iterator> |
44 | #include <string> |
45 | #include <vector> |
46 | |
47 | namespace llvm { |
48 | |
49 | namespace Intrinsic { |
50 | typedef unsigned ID; |
51 | } |
52 | |
53 | //===----------------------------------------------------------------------===// |
54 | // UnaryInstruction Class |
55 | //===----------------------------------------------------------------------===// |
56 | |
57 | class UnaryInstruction : public Instruction { |
58 | protected: |
59 | UnaryInstruction(Type *Ty, unsigned iType, Value *V, |
60 | Instruction *IB = nullptr) |
61 | : Instruction(Ty, iType, &Op<0>(), 1, IB) { |
62 | Op<0>() = V; |
63 | } |
64 | UnaryInstruction(Type *Ty, unsigned iType, Value *V, BasicBlock *IAE) |
65 | : Instruction(Ty, iType, &Op<0>(), 1, IAE) { |
66 | Op<0>() = V; |
67 | } |
68 | |
69 | public: |
70 | // allocate space for exactly one operand |
71 | void *operator new(size_t S) { return User::operator new(S, 1); } |
72 | void operator delete(void *Ptr) { User::operator delete(Ptr); } |
73 | |
74 | /// Transparently provide more efficient getOperand methods. |
75 | DECLARE_TRANSPARENT_OPERAND_ACCESSORS(Value)public: inline Value *getOperand(unsigned) const; inline void setOperand(unsigned, Value*); inline op_iterator op_begin(); inline const_op_iterator op_begin() const; inline op_iterator op_end(); inline const_op_iterator op_end() const; protected : template <int> inline Use &Op(); template <int > inline const Use &Op() const; public: inline unsigned getNumOperands() const; |
76 | |
77 | // Methods for support type inquiry through isa, cast, and dyn_cast: |
78 | static bool classof(const Instruction *I) { |
79 | return I->isUnaryOp() || |
80 | I->getOpcode() == Instruction::Alloca || |
81 | I->getOpcode() == Instruction::Load || |
82 | I->getOpcode() == Instruction::VAArg || |
83 | I->getOpcode() == Instruction::ExtractValue || |
84 | (I->getOpcode() >= CastOpsBegin && I->getOpcode() < CastOpsEnd); |
85 | } |
86 | static bool classof(const Value *V) { |
87 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
88 | } |
89 | }; |
90 | |
91 | template <> |
92 | struct OperandTraits<UnaryInstruction> : |
93 | public FixedNumOperandTraits<UnaryInstruction, 1> { |
94 | }; |
95 | |
96 | DEFINE_TRANSPARENT_OPERAND_ACCESSORS(UnaryInstruction, Value)UnaryInstruction::op_iterator UnaryInstruction::op_begin() { return OperandTraits<UnaryInstruction>::op_begin(this); } UnaryInstruction ::const_op_iterator UnaryInstruction::op_begin() const { return OperandTraits<UnaryInstruction>::op_begin(const_cast< UnaryInstruction*>(this)); } UnaryInstruction::op_iterator UnaryInstruction::op_end() { return OperandTraits<UnaryInstruction >::op_end(this); } UnaryInstruction::const_op_iterator UnaryInstruction ::op_end() const { return OperandTraits<UnaryInstruction> ::op_end(const_cast<UnaryInstruction*>(this)); } Value * UnaryInstruction::getOperand(unsigned i_nocapture) const { (( void)0); return cast_or_null<Value>( OperandTraits<UnaryInstruction >::op_begin(const_cast<UnaryInstruction*>(this))[i_nocapture ].get()); } void UnaryInstruction::setOperand(unsigned i_nocapture , Value *Val_nocapture) { ((void)0); OperandTraits<UnaryInstruction >::op_begin(this)[i_nocapture] = Val_nocapture; } unsigned UnaryInstruction::getNumOperands() const { return OperandTraits <UnaryInstruction>::operands(this); } template <int Idx_nocapture > Use &UnaryInstruction::Op() { return this->OpFrom <Idx_nocapture>(this); } template <int Idx_nocapture > const Use &UnaryInstruction::Op() const { return this ->OpFrom<Idx_nocapture>(this); } |
97 | |
98 | //===----------------------------------------------------------------------===// |
99 | // UnaryOperator Class |
100 | //===----------------------------------------------------------------------===// |
101 | |
102 | class UnaryOperator : public UnaryInstruction { |
103 | void AssertOK(); |
104 | |
105 | protected: |
106 | UnaryOperator(UnaryOps iType, Value *S, Type *Ty, |
107 | const Twine &Name, Instruction *InsertBefore); |
108 | UnaryOperator(UnaryOps iType, Value *S, Type *Ty, |
109 | const Twine &Name, BasicBlock *InsertAtEnd); |
110 | |
111 | // Note: Instruction needs to be a friend here to call cloneImpl. |
112 | friend class Instruction; |
113 | |
114 | UnaryOperator *cloneImpl() const; |
115 | |
116 | public: |
117 | |
118 | /// Construct a unary instruction, given the opcode and an operand. |
119 | /// Optionally (if InstBefore is specified) insert the instruction |
120 | /// into a BasicBlock right before the specified instruction. The specified |
121 | /// Instruction is allowed to be a dereferenced end iterator. |
122 | /// |
123 | static UnaryOperator *Create(UnaryOps Op, Value *S, |
124 | const Twine &Name = Twine(), |
125 | Instruction *InsertBefore = nullptr); |
126 | |
127 | /// Construct a unary instruction, given the opcode and an operand. |
128 | /// Also automatically insert this instruction to the end of the |
129 | /// BasicBlock specified. |
130 | /// |
131 | static UnaryOperator *Create(UnaryOps Op, Value *S, |
132 | const Twine &Name, |
133 | BasicBlock *InsertAtEnd); |
134 | |
135 | /// These methods just forward to Create, and are useful when you |
136 | /// statically know what type of instruction you're going to create. These |
137 | /// helpers just save some typing. |
138 | #define HANDLE_UNARY_INST(N, OPC, CLASS) \ |
139 | static UnaryOperator *Create##OPC(Value *V, const Twine &Name = "") {\ |
140 | return Create(Instruction::OPC, V, Name);\ |
141 | } |
142 | #include "llvm/IR/Instruction.def" |
143 | #define HANDLE_UNARY_INST(N, OPC, CLASS) \ |
144 | static UnaryOperator *Create##OPC(Value *V, const Twine &Name, \ |
145 | BasicBlock *BB) {\ |
146 | return Create(Instruction::OPC, V, Name, BB);\ |
147 | } |
148 | #include "llvm/IR/Instruction.def" |
149 | #define HANDLE_UNARY_INST(N, OPC, CLASS) \ |
150 | static UnaryOperator *Create##OPC(Value *V, const Twine &Name, \ |
151 | Instruction *I) {\ |
152 | return Create(Instruction::OPC, V, Name, I);\ |
153 | } |
154 | #include "llvm/IR/Instruction.def" |
155 | |
156 | static UnaryOperator * |
157 | CreateWithCopiedFlags(UnaryOps Opc, Value *V, Instruction *CopyO, |
158 | const Twine &Name = "", |
159 | Instruction *InsertBefore = nullptr) { |
160 | UnaryOperator *UO = Create(Opc, V, Name, InsertBefore); |
161 | UO->copyIRFlags(CopyO); |
162 | return UO; |
163 | } |
164 | |
165 | static UnaryOperator *CreateFNegFMF(Value *Op, Instruction *FMFSource, |
166 | const Twine &Name = "", |
167 | Instruction *InsertBefore = nullptr) { |
168 | return CreateWithCopiedFlags(Instruction::FNeg, Op, FMFSource, Name, |
169 | InsertBefore); |
170 | } |
171 | |
172 | UnaryOps getOpcode() const { |
173 | return static_cast<UnaryOps>(Instruction::getOpcode()); |
174 | } |
175 | |
176 | // Methods for support type inquiry through isa, cast, and dyn_cast: |
177 | static bool classof(const Instruction *I) { |
178 | return I->isUnaryOp(); |
179 | } |
180 | static bool classof(const Value *V) { |
181 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
182 | } |
183 | }; |
184 | |
185 | //===----------------------------------------------------------------------===// |
186 | // BinaryOperator Class |
187 | //===----------------------------------------------------------------------===// |
188 | |
189 | class BinaryOperator : public Instruction { |
190 | void AssertOK(); |
191 | |
192 | protected: |
193 | BinaryOperator(BinaryOps iType, Value *S1, Value *S2, Type *Ty, |
194 | const Twine &Name, Instruction *InsertBefore); |
195 | BinaryOperator(BinaryOps iType, Value *S1, Value *S2, Type *Ty, |
196 | const Twine &Name, BasicBlock *InsertAtEnd); |
197 | |
198 | // Note: Instruction needs to be a friend here to call cloneImpl. |
199 | friend class Instruction; |
200 | |
201 | BinaryOperator *cloneImpl() const; |
202 | |
203 | public: |
204 | // allocate space for exactly two operands |
205 | void *operator new(size_t S) { return User::operator new(S, 2); } |
206 | void operator delete(void *Ptr) { User::operator delete(Ptr); } |
207 | |
208 | /// Transparently provide more efficient getOperand methods. |
209 | DECLARE_TRANSPARENT_OPERAND_ACCESSORS(Value)public: inline Value *getOperand(unsigned) const; inline void setOperand(unsigned, Value*); inline op_iterator op_begin(); inline const_op_iterator op_begin() const; inline op_iterator op_end(); inline const_op_iterator op_end() const; protected : template <int> inline Use &Op(); template <int > inline const Use &Op() const; public: inline unsigned getNumOperands() const; |
210 | |
211 | /// Construct a binary instruction, given the opcode and the two |
212 | /// operands. Optionally (if InstBefore is specified) insert the instruction |
213 | /// into a BasicBlock right before the specified instruction. The specified |
214 | /// Instruction is allowed to be a dereferenced end iterator. |
215 | /// |
216 | static BinaryOperator *Create(BinaryOps Op, Value *S1, Value *S2, |
217 | const Twine &Name = Twine(), |
218 | Instruction *InsertBefore = nullptr); |
219 | |
220 | /// Construct a binary instruction, given the opcode and the two |
221 | /// operands. Also automatically insert this instruction to the end of the |
222 | /// BasicBlock specified. |
223 | /// |
224 | static BinaryOperator *Create(BinaryOps Op, Value *S1, Value *S2, |
225 | const Twine &Name, BasicBlock *InsertAtEnd); |
226 | |
227 | /// These methods just forward to Create, and are useful when you |
228 | /// statically know what type of instruction you're going to create. These |
229 | /// helpers just save some typing. |
230 | #define HANDLE_BINARY_INST(N, OPC, CLASS) \ |
231 | static BinaryOperator *Create##OPC(Value *V1, Value *V2, \ |
232 | const Twine &Name = "") {\ |
233 | return Create(Instruction::OPC, V1, V2, Name);\ |
234 | } |
235 | #include "llvm/IR/Instruction.def" |
236 | #define HANDLE_BINARY_INST(N, OPC, CLASS) \ |
237 | static BinaryOperator *Create##OPC(Value *V1, Value *V2, \ |
238 | const Twine &Name, BasicBlock *BB) {\ |
239 | return Create(Instruction::OPC, V1, V2, Name, BB);\ |
240 | } |
241 | #include "llvm/IR/Instruction.def" |
242 | #define HANDLE_BINARY_INST(N, OPC, CLASS) \ |
243 | static BinaryOperator *Create##OPC(Value *V1, Value *V2, \ |
244 | const Twine &Name, Instruction *I) {\ |
245 | return Create(Instruction::OPC, V1, V2, Name, I);\ |
246 | } |
247 | #include "llvm/IR/Instruction.def" |
248 | |
249 | static BinaryOperator * |
250 | CreateWithCopiedFlags(BinaryOps Opc, Value *V1, Value *V2, Instruction *CopyO, |
251 | const Twine &Name = "", |
252 | Instruction *InsertBefore = nullptr) { |
253 | BinaryOperator *BO = Create(Opc, V1, V2, Name, InsertBefore); |
254 | BO->copyIRFlags(CopyO); |
255 | return BO; |
256 | } |
257 | |
258 | static BinaryOperator *CreateFAddFMF(Value *V1, Value *V2, |
259 | Instruction *FMFSource, |
260 | const Twine &Name = "") { |
261 | return CreateWithCopiedFlags(Instruction::FAdd, V1, V2, FMFSource, Name); |
262 | } |
263 | static BinaryOperator *CreateFSubFMF(Value *V1, Value *V2, |
264 | Instruction *FMFSource, |
265 | const Twine &Name = "") { |
266 | return CreateWithCopiedFlags(Instruction::FSub, V1, V2, FMFSource, Name); |
267 | } |
268 | static BinaryOperator *CreateFMulFMF(Value *V1, Value *V2, |
269 | Instruction *FMFSource, |
270 | const Twine &Name = "") { |
271 | return CreateWithCopiedFlags(Instruction::FMul, V1, V2, FMFSource, Name); |
272 | } |
273 | static BinaryOperator *CreateFDivFMF(Value *V1, Value *V2, |
274 | Instruction *FMFSource, |
275 | const Twine &Name = "") { |
276 | return CreateWithCopiedFlags(Instruction::FDiv, V1, V2, FMFSource, Name); |
277 | } |
278 | static BinaryOperator *CreateFRemFMF(Value *V1, Value *V2, |
279 | Instruction *FMFSource, |
280 | const Twine &Name = "") { |
281 | return CreateWithCopiedFlags(Instruction::FRem, V1, V2, FMFSource, Name); |
282 | } |
283 | |
284 | static BinaryOperator *CreateNSW(BinaryOps Opc, Value *V1, Value *V2, |
285 | const Twine &Name = "") { |
286 | BinaryOperator *BO = Create(Opc, V1, V2, Name); |
287 | BO->setHasNoSignedWrap(true); |
288 | return BO; |
289 | } |
290 | static BinaryOperator *CreateNSW(BinaryOps Opc, Value *V1, Value *V2, |
291 | const Twine &Name, BasicBlock *BB) { |
292 | BinaryOperator *BO = Create(Opc, V1, V2, Name, BB); |
293 | BO->setHasNoSignedWrap(true); |
294 | return BO; |
295 | } |
296 | static BinaryOperator *CreateNSW(BinaryOps Opc, Value *V1, Value *V2, |
297 | const Twine &Name, Instruction *I) { |
298 | BinaryOperator *BO = Create(Opc, V1, V2, Name, I); |
299 | BO->setHasNoSignedWrap(true); |
300 | return BO; |
301 | } |
302 | |
303 | static BinaryOperator *CreateNUW(BinaryOps Opc, Value *V1, Value *V2, |
304 | const Twine &Name = "") { |
305 | BinaryOperator *BO = Create(Opc, V1, V2, Name); |
306 | BO->setHasNoUnsignedWrap(true); |
307 | return BO; |
308 | } |
309 | static BinaryOperator *CreateNUW(BinaryOps Opc, Value *V1, Value *V2, |
310 | const Twine &Name, BasicBlock *BB) { |
311 | BinaryOperator *BO = Create(Opc, V1, V2, Name, BB); |
312 | BO->setHasNoUnsignedWrap(true); |
313 | return BO; |
314 | } |
315 | static BinaryOperator *CreateNUW(BinaryOps Opc, Value *V1, Value *V2, |
316 | const Twine &Name, Instruction *I) { |
317 | BinaryOperator *BO = Create(Opc, V1, V2, Name, I); |
318 | BO->setHasNoUnsignedWrap(true); |
319 | return BO; |
320 | } |
321 | |
322 | static BinaryOperator *CreateExact(BinaryOps Opc, Value *V1, Value *V2, |
323 | const Twine &Name = "") { |
324 | BinaryOperator *BO = Create(Opc, V1, V2, Name); |
325 | BO->setIsExact(true); |
326 | return BO; |
327 | } |
328 | static BinaryOperator *CreateExact(BinaryOps Opc, Value *V1, Value *V2, |
329 | const Twine &Name, BasicBlock *BB) { |
330 | BinaryOperator *BO = Create(Opc, V1, V2, Name, BB); |
331 | BO->setIsExact(true); |
332 | return BO; |
333 | } |
334 | static BinaryOperator *CreateExact(BinaryOps Opc, Value *V1, Value *V2, |
335 | const Twine &Name, Instruction *I) { |
336 | BinaryOperator *BO = Create(Opc, V1, V2, Name, I); |
337 | BO->setIsExact(true); |
338 | return BO; |
339 | } |
340 | |
341 | #define DEFINE_HELPERS(OPC, NUWNSWEXACT) \ |
342 | static BinaryOperator *Create##NUWNSWEXACT##OPC(Value *V1, Value *V2, \ |
343 | const Twine &Name = "") { \ |
344 | return Create##NUWNSWEXACT(Instruction::OPC, V1, V2, Name); \ |
345 | } \ |
346 | static BinaryOperator *Create##NUWNSWEXACT##OPC( \ |
347 | Value *V1, Value *V2, const Twine &Name, BasicBlock *BB) { \ |
348 | return Create##NUWNSWEXACT(Instruction::OPC, V1, V2, Name, BB); \ |
349 | } \ |
350 | static BinaryOperator *Create##NUWNSWEXACT##OPC( \ |
351 | Value *V1, Value *V2, const Twine &Name, Instruction *I) { \ |
352 | return Create##NUWNSWEXACT(Instruction::OPC, V1, V2, Name, I); \ |
353 | } |
354 | |
355 | DEFINE_HELPERS(Add, NSW) // CreateNSWAdd |
356 | DEFINE_HELPERS(Add, NUW) // CreateNUWAdd |
357 | DEFINE_HELPERS(Sub, NSW) // CreateNSWSub |
358 | DEFINE_HELPERS(Sub, NUW) // CreateNUWSub |
359 | DEFINE_HELPERS(Mul, NSW) // CreateNSWMul |
360 | DEFINE_HELPERS(Mul, NUW) // CreateNUWMul |
361 | DEFINE_HELPERS(Shl, NSW) // CreateNSWShl |
362 | DEFINE_HELPERS(Shl, NUW) // CreateNUWShl |
363 | |
364 | DEFINE_HELPERS(SDiv, Exact) // CreateExactSDiv |
365 | DEFINE_HELPERS(UDiv, Exact) // CreateExactUDiv |
366 | DEFINE_HELPERS(AShr, Exact) // CreateExactAShr |
367 | DEFINE_HELPERS(LShr, Exact) // CreateExactLShr |
368 | |
369 | #undef DEFINE_HELPERS |
370 | |
371 | /// Helper functions to construct and inspect unary operations (NEG and NOT) |
372 | /// via binary operators SUB and XOR: |
373 | /// |
374 | /// Create the NEG and NOT instructions out of SUB and XOR instructions. |
375 | /// |
376 | static BinaryOperator *CreateNeg(Value *Op, const Twine &Name = "", |
377 | Instruction *InsertBefore = nullptr); |
378 | static BinaryOperator *CreateNeg(Value *Op, const Twine &Name, |
379 | BasicBlock *InsertAtEnd); |
380 | static BinaryOperator *CreateNSWNeg(Value *Op, const Twine &Name = "", |
381 | Instruction *InsertBefore = nullptr); |
382 | static BinaryOperator *CreateNSWNeg(Value *Op, const Twine &Name, |
383 | BasicBlock *InsertAtEnd); |
384 | static BinaryOperator *CreateNUWNeg(Value *Op, const Twine &Name = "", |
385 | Instruction *InsertBefore = nullptr); |
386 | static BinaryOperator *CreateNUWNeg(Value *Op, const Twine &Name, |
387 | BasicBlock *InsertAtEnd); |
388 | static BinaryOperator *CreateNot(Value *Op, const Twine &Name = "", |
389 | Instruction *InsertBefore = nullptr); |
390 | static BinaryOperator *CreateNot(Value *Op, const Twine &Name, |
391 | BasicBlock *InsertAtEnd); |
392 | |
393 | BinaryOps getOpcode() const { |
394 | return static_cast<BinaryOps>(Instruction::getOpcode()); |
395 | } |
396 | |
397 | /// Exchange the two operands to this instruction. |
398 | /// This instruction is safe to use on any binary instruction and |
399 | /// does not modify the semantics of the instruction. If the instruction |
400 | /// cannot be reversed (ie, it's a Div), then return true. |
401 | /// |
402 | bool swapOperands(); |
403 | |
404 | // Methods for support type inquiry through isa, cast, and dyn_cast: |
405 | static bool classof(const Instruction *I) { |
406 | return I->isBinaryOp(); |
407 | } |
408 | static bool classof(const Value *V) { |
409 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
410 | } |
411 | }; |
412 | |
413 | template <> |
414 | struct OperandTraits<BinaryOperator> : |
415 | public FixedNumOperandTraits<BinaryOperator, 2> { |
416 | }; |
417 | |
418 | DEFINE_TRANSPARENT_OPERAND_ACCESSORS(BinaryOperator, Value)BinaryOperator::op_iterator BinaryOperator::op_begin() { return OperandTraits<BinaryOperator>::op_begin(this); } BinaryOperator ::const_op_iterator BinaryOperator::op_begin() const { return OperandTraits<BinaryOperator>::op_begin(const_cast< BinaryOperator*>(this)); } BinaryOperator::op_iterator BinaryOperator ::op_end() { return OperandTraits<BinaryOperator>::op_end (this); } BinaryOperator::const_op_iterator BinaryOperator::op_end () const { return OperandTraits<BinaryOperator>::op_end (const_cast<BinaryOperator*>(this)); } Value *BinaryOperator ::getOperand(unsigned i_nocapture) const { ((void)0); return cast_or_null <Value>( OperandTraits<BinaryOperator>::op_begin( const_cast<BinaryOperator*>(this))[i_nocapture].get()); } void BinaryOperator::setOperand(unsigned i_nocapture, Value *Val_nocapture) { ((void)0); OperandTraits<BinaryOperator >::op_begin(this)[i_nocapture] = Val_nocapture; } unsigned BinaryOperator::getNumOperands() const { return OperandTraits <BinaryOperator>::operands(this); } template <int Idx_nocapture > Use &BinaryOperator::Op() { return this->OpFrom< Idx_nocapture>(this); } template <int Idx_nocapture> const Use &BinaryOperator::Op() const { return this-> OpFrom<Idx_nocapture>(this); } |
419 | |
420 | //===----------------------------------------------------------------------===// |
421 | // CastInst Class |
422 | //===----------------------------------------------------------------------===// |
423 | |
424 | /// This is the base class for all instructions that perform data |
425 | /// casts. It is simply provided so that instruction category testing |
426 | /// can be performed with code like: |
427 | /// |
428 | /// if (isa<CastInst>(Instr)) { ... } |
429 | /// Base class of casting instructions. |
430 | class CastInst : public UnaryInstruction { |
431 | protected: |
432 | /// Constructor with insert-before-instruction semantics for subclasses |
433 | CastInst(Type *Ty, unsigned iType, Value *S, |
434 | const Twine &NameStr = "", Instruction *InsertBefore = nullptr) |
435 | : UnaryInstruction(Ty, iType, S, InsertBefore) { |
436 | setName(NameStr); |
437 | } |
438 | /// Constructor with insert-at-end-of-block semantics for subclasses |
439 | CastInst(Type *Ty, unsigned iType, Value *S, |
440 | const Twine &NameStr, BasicBlock *InsertAtEnd) |
441 | : UnaryInstruction(Ty, iType, S, InsertAtEnd) { |
442 | setName(NameStr); |
443 | } |
444 | |
445 | public: |
446 | /// Provides a way to construct any of the CastInst subclasses using an |
447 | /// opcode instead of the subclass's constructor. The opcode must be in the |
448 | /// CastOps category (Instruction::isCast(opcode) returns true). This |
449 | /// constructor has insert-before-instruction semantics to automatically |
450 | /// insert the new CastInst before InsertBefore (if it is non-null). |
451 | /// Construct any of the CastInst subclasses |
452 | static CastInst *Create( |
453 | Instruction::CastOps, ///< The opcode of the cast instruction |
454 | Value *S, ///< The value to be casted (operand 0) |
455 | Type *Ty, ///< The type to which cast should be made |
456 | const Twine &Name = "", ///< Name for the instruction |
457 | Instruction *InsertBefore = nullptr ///< Place to insert the instruction |
458 | ); |
459 | /// Provides a way to construct any of the CastInst subclasses using an |
460 | /// opcode instead of the subclass's constructor. The opcode must be in the |
461 | /// CastOps category. This constructor has insert-at-end-of-block semantics |
462 | /// to automatically insert the new CastInst at the end of InsertAtEnd (if |
463 | /// its non-null). |
464 | /// Construct any of the CastInst subclasses |
465 | static CastInst *Create( |
466 | Instruction::CastOps, ///< The opcode for the cast instruction |
467 | Value *S, ///< The value to be casted (operand 0) |
468 | Type *Ty, ///< The type to which operand is casted |
469 | const Twine &Name, ///< The name for the instruction |
470 | BasicBlock *InsertAtEnd ///< The block to insert the instruction into |
471 | ); |
472 | |
473 | /// Create a ZExt or BitCast cast instruction |
474 | static CastInst *CreateZExtOrBitCast( |
475 | Value *S, ///< The value to be casted (operand 0) |
476 | Type *Ty, ///< The type to which cast should be made |
477 | const Twine &Name = "", ///< Name for the instruction |
478 | Instruction *InsertBefore = nullptr ///< Place to insert the instruction |
479 | ); |
480 | |
481 | /// Create a ZExt or BitCast cast instruction |
482 | static CastInst *CreateZExtOrBitCast( |
483 | Value *S, ///< The value to be casted (operand 0) |
484 | Type *Ty, ///< The type to which operand is casted |
485 | const Twine &Name, ///< The name for the instruction |
486 | BasicBlock *InsertAtEnd ///< The block to insert the instruction into |
487 | ); |
488 | |
489 | /// Create a SExt or BitCast cast instruction |
490 | static CastInst *CreateSExtOrBitCast( |
491 | Value *S, ///< The value to be casted (operand 0) |
492 | Type *Ty, ///< The type to which cast should be made |
493 | const Twine &Name = "", ///< Name for the instruction |
494 | Instruction *InsertBefore = nullptr ///< Place to insert the instruction |
495 | ); |
496 | |
497 | /// Create a SExt or BitCast cast instruction |
498 | static CastInst *CreateSExtOrBitCast( |
499 | Value *S, ///< The value to be casted (operand 0) |
500 | Type *Ty, ///< The type to which operand is casted |
501 | const Twine &Name, ///< The name for the instruction |
502 | BasicBlock *InsertAtEnd ///< The block to insert the instruction into |
503 | ); |
504 | |
505 | /// Create a BitCast AddrSpaceCast, or a PtrToInt cast instruction. |
506 | static CastInst *CreatePointerCast( |
507 | Value *S, ///< The pointer value to be casted (operand 0) |
508 | Type *Ty, ///< The type to which operand is casted |
509 | const Twine &Name, ///< The name for the instruction |
510 | BasicBlock *InsertAtEnd ///< The block to insert the instruction into |
511 | ); |
512 | |
513 | /// Create a BitCast, AddrSpaceCast or a PtrToInt cast instruction. |
514 | static CastInst *CreatePointerCast( |
515 | Value *S, ///< The pointer value to be casted (operand 0) |
516 | Type *Ty, ///< The type to which cast should be made |
517 | const Twine &Name = "", ///< Name for the instruction |
518 | Instruction *InsertBefore = nullptr ///< Place to insert the instruction |
519 | ); |
520 | |
521 | /// Create a BitCast or an AddrSpaceCast cast instruction. |
522 | static CastInst *CreatePointerBitCastOrAddrSpaceCast( |
523 | Value *S, ///< The pointer value to be casted (operand 0) |
524 | Type *Ty, ///< The type to which operand is casted |
525 | const Twine &Name, ///< The name for the instruction |
526 | BasicBlock *InsertAtEnd ///< The block to insert the instruction into |
527 | ); |
528 | |
529 | /// Create a BitCast or an AddrSpaceCast cast instruction. |
530 | static CastInst *CreatePointerBitCastOrAddrSpaceCast( |
531 | Value *S, ///< The pointer value to be casted (operand 0) |
532 | Type *Ty, ///< The type to which cast should be made |
533 | const Twine &Name = "", ///< Name for the instruction |
534 | Instruction *InsertBefore = nullptr ///< Place to insert the instruction |
535 | ); |
536 | |
537 | /// Create a BitCast, a PtrToInt, or an IntToPTr cast instruction. |
538 | /// |
539 | /// If the value is a pointer type and the destination an integer type, |
540 | /// creates a PtrToInt cast. If the value is an integer type and the |
541 | /// destination a pointer type, creates an IntToPtr cast. Otherwise, creates |
542 | /// a bitcast. |
543 | static CastInst *CreateBitOrPointerCast( |
544 | Value *S, ///< The pointer value to be casted (operand 0) |
545 | Type *Ty, ///< The type to which cast should be made |
546 | const Twine &Name = "", ///< Name for the instruction |
547 | Instruction *InsertBefore = nullptr ///< Place to insert the instruction |
548 | ); |
549 | |
550 | /// Create a ZExt, BitCast, or Trunc for int -> int casts. |
551 | static CastInst *CreateIntegerCast( |
552 | Value *S, ///< The pointer value to be casted (operand 0) |
553 | Type *Ty, ///< The type to which cast should be made |
554 | bool isSigned, ///< Whether to regard S as signed or not |
555 | const Twine &Name = "", ///< Name for the instruction |
556 | Instruction *InsertBefore = nullptr ///< Place to insert the instruction |
557 | ); |
558 | |
559 | /// Create a ZExt, BitCast, or Trunc for int -> int casts. |
560 | static CastInst *CreateIntegerCast( |
561 | Value *S, ///< The integer value to be casted (operand 0) |
562 | Type *Ty, ///< The integer type to which operand is casted |
563 | bool isSigned, ///< Whether to regard S as signed or not |
564 | const Twine &Name, ///< The name for the instruction |
565 | BasicBlock *InsertAtEnd ///< The block to insert the instruction into |
566 | ); |
567 | |
568 | /// Create an FPExt, BitCast, or FPTrunc for fp -> fp casts |
569 | static CastInst *CreateFPCast( |
570 | Value *S, ///< The floating point value to be casted |
571 | Type *Ty, ///< The floating point type to cast to |
572 | const Twine &Name = "", ///< Name for the instruction |
573 | Instruction *InsertBefore = nullptr ///< Place to insert the instruction |
574 | ); |
575 | |
576 | /// Create an FPExt, BitCast, or FPTrunc for fp -> fp casts |
577 | static CastInst *CreateFPCast( |
578 | Value *S, ///< The floating point value to be casted |
579 | Type *Ty, ///< The floating point type to cast to |
580 | const Twine &Name, ///< The name for the instruction |
581 | BasicBlock *InsertAtEnd ///< The block to insert the instruction into |
582 | ); |
583 | |
584 | /// Create a Trunc or BitCast cast instruction |
585 | static CastInst *CreateTruncOrBitCast( |
586 | Value *S, ///< The value to be casted (operand 0) |
587 | Type *Ty, ///< The type to which cast should be made |
588 | const Twine &Name = "", ///< Name for the instruction |
589 | Instruction *InsertBefore = nullptr ///< Place to insert the instruction |
590 | ); |
591 | |
592 | /// Create a Trunc or BitCast cast instruction |
593 | static CastInst *CreateTruncOrBitCast( |
594 | Value *S, ///< The value to be casted (operand 0) |
595 | Type *Ty, ///< The type to which operand is casted |
596 | const Twine &Name, ///< The name for the instruction |
597 | BasicBlock *InsertAtEnd ///< The block to insert the instruction into |
598 | ); |
599 | |
600 | /// Check whether a bitcast between these types is valid |
601 | static bool isBitCastable( |
602 | Type *SrcTy, ///< The Type from which the value should be cast. |
603 | Type *DestTy ///< The Type to which the value should be cast. |
604 | ); |
605 | |
606 | /// Check whether a bitcast, inttoptr, or ptrtoint cast between these |
607 | /// types is valid and a no-op. |
608 | /// |
609 | /// This ensures that any pointer<->integer cast has enough bits in the |
610 | /// integer and any other cast is a bitcast. |
611 | static bool isBitOrNoopPointerCastable( |
612 | Type *SrcTy, ///< The Type from which the value should be cast. |
613 | Type *DestTy, ///< The Type to which the value should be cast. |
614 | const DataLayout &DL); |
615 | |
616 | /// Returns the opcode necessary to cast Val into Ty using usual casting |
617 | /// rules. |
618 | /// Infer the opcode for cast operand and type |
619 | static Instruction::CastOps getCastOpcode( |
620 | const Value *Val, ///< The value to cast |
621 | bool SrcIsSigned, ///< Whether to treat the source as signed |
622 | Type *Ty, ///< The Type to which the value should be casted |
623 | bool DstIsSigned ///< Whether to treate the dest. as signed |
624 | ); |
625 | |
626 | /// There are several places where we need to know if a cast instruction |
627 | /// only deals with integer source and destination types. To simplify that |
628 | /// logic, this method is provided. |
629 | /// @returns true iff the cast has only integral typed operand and dest type. |
630 | /// Determine if this is an integer-only cast. |
631 | bool isIntegerCast() const; |
632 | |
633 | /// A lossless cast is one that does not alter the basic value. It implies |
634 | /// a no-op cast but is more stringent, preventing things like int->float, |
635 | /// long->double, or int->ptr. |
636 | /// @returns true iff the cast is lossless. |
637 | /// Determine if this is a lossless cast. |
638 | bool isLosslessCast() const; |
639 | |
640 | /// A no-op cast is one that can be effected without changing any bits. |
641 | /// It implies that the source and destination types are the same size. The |
642 | /// DataLayout argument is to determine the pointer size when examining casts |
643 | /// involving Integer and Pointer types. They are no-op casts if the integer |
644 | /// is the same size as the pointer. However, pointer size varies with |
645 | /// platform. Note that a precondition of this method is that the cast is |
646 | /// legal - i.e. the instruction formed with these operands would verify. |
647 | static bool isNoopCast( |
648 | Instruction::CastOps Opcode, ///< Opcode of cast |
649 | Type *SrcTy, ///< SrcTy of cast |
650 | Type *DstTy, ///< DstTy of cast |
651 | const DataLayout &DL ///< DataLayout to get the Int Ptr type from. |
652 | ); |
653 | |
654 | /// Determine if this cast is a no-op cast. |
655 | /// |
656 | /// \param DL is the DataLayout to determine pointer size. |
657 | bool isNoopCast(const DataLayout &DL) const; |
658 | |
659 | /// Determine how a pair of casts can be eliminated, if they can be at all. |
660 | /// This is a helper function for both CastInst and ConstantExpr. |
661 | /// @returns 0 if the CastInst pair can't be eliminated, otherwise |
662 | /// returns Instruction::CastOps value for a cast that can replace |
663 | /// the pair, casting SrcTy to DstTy. |
664 | /// Determine if a cast pair is eliminable |
665 | static unsigned isEliminableCastPair( |
666 | Instruction::CastOps firstOpcode, ///< Opcode of first cast |
667 | Instruction::CastOps secondOpcode, ///< Opcode of second cast |
668 | Type *SrcTy, ///< SrcTy of 1st cast |
669 | Type *MidTy, ///< DstTy of 1st cast & SrcTy of 2nd cast |
670 | Type *DstTy, ///< DstTy of 2nd cast |
671 | Type *SrcIntPtrTy, ///< Integer type corresponding to Ptr SrcTy, or null |
672 | Type *MidIntPtrTy, ///< Integer type corresponding to Ptr MidTy, or null |
673 | Type *DstIntPtrTy ///< Integer type corresponding to Ptr DstTy, or null |
674 | ); |
675 | |
676 | /// Return the opcode of this CastInst |
677 | Instruction::CastOps getOpcode() const { |
678 | return Instruction::CastOps(Instruction::getOpcode()); |
679 | } |
680 | |
681 | /// Return the source type, as a convenience |
682 | Type* getSrcTy() const { return getOperand(0)->getType(); } |
683 | /// Return the destination type, as a convenience |
684 | Type* getDestTy() const { return getType(); } |
685 | |
686 | /// This method can be used to determine if a cast from SrcTy to DstTy using |
687 | /// Opcode op is valid or not. |
688 | /// @returns true iff the proposed cast is valid. |
689 | /// Determine if a cast is valid without creating one. |
690 | static bool castIsValid(Instruction::CastOps op, Type *SrcTy, Type *DstTy); |
691 | static bool castIsValid(Instruction::CastOps op, Value *S, Type *DstTy) { |
692 | return castIsValid(op, S->getType(), DstTy); |
693 | } |
694 | |
695 | /// Methods for support type inquiry through isa, cast, and dyn_cast: |
696 | static bool classof(const Instruction *I) { |
697 | return I->isCast(); |
698 | } |
699 | static bool classof(const Value *V) { |
700 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
701 | } |
702 | }; |
703 | |
704 | //===----------------------------------------------------------------------===// |
705 | // CmpInst Class |
706 | //===----------------------------------------------------------------------===// |
707 | |
708 | /// This class is the base class for the comparison instructions. |
709 | /// Abstract base class of comparison instructions. |
710 | class CmpInst : public Instruction { |
711 | public: |
712 | /// This enumeration lists the possible predicates for CmpInst subclasses. |
713 | /// Values in the range 0-31 are reserved for FCmpInst, while values in the |
714 | /// range 32-64 are reserved for ICmpInst. This is necessary to ensure the |
715 | /// predicate values are not overlapping between the classes. |
716 | /// |
717 | /// Some passes (e.g. InstCombine) depend on the bit-wise characteristics of |
718 | /// FCMP_* values. Changing the bit patterns requires a potential change to |
719 | /// those passes. |
720 | enum Predicate : unsigned { |
721 | // Opcode U L G E Intuitive operation |
722 | FCMP_FALSE = 0, ///< 0 0 0 0 Always false (always folded) |
723 | FCMP_OEQ = 1, ///< 0 0 0 1 True if ordered and equal |
724 | FCMP_OGT = 2, ///< 0 0 1 0 True if ordered and greater than |
725 | FCMP_OGE = 3, ///< 0 0 1 1 True if ordered and greater than or equal |
726 | FCMP_OLT = 4, ///< 0 1 0 0 True if ordered and less than |
727 | FCMP_OLE = 5, ///< 0 1 0 1 True if ordered and less than or equal |
728 | FCMP_ONE = 6, ///< 0 1 1 0 True if ordered and operands are unequal |
729 | FCMP_ORD = 7, ///< 0 1 1 1 True if ordered (no nans) |
730 | FCMP_UNO = 8, ///< 1 0 0 0 True if unordered: isnan(X) | isnan(Y) |
731 | FCMP_UEQ = 9, ///< 1 0 0 1 True if unordered or equal |
732 | FCMP_UGT = 10, ///< 1 0 1 0 True if unordered or greater than |
733 | FCMP_UGE = 11, ///< 1 0 1 1 True if unordered, greater than, or equal |
734 | FCMP_ULT = 12, ///< 1 1 0 0 True if unordered or less than |
735 | FCMP_ULE = 13, ///< 1 1 0 1 True if unordered, less than, or equal |
736 | FCMP_UNE = 14, ///< 1 1 1 0 True if unordered or not equal |
737 | FCMP_TRUE = 15, ///< 1 1 1 1 Always true (always folded) |
738 | FIRST_FCMP_PREDICATE = FCMP_FALSE, |
739 | LAST_FCMP_PREDICATE = FCMP_TRUE, |
740 | BAD_FCMP_PREDICATE = FCMP_TRUE + 1, |
741 | ICMP_EQ = 32, ///< equal |
742 | ICMP_NE = 33, ///< not equal |
743 | ICMP_UGT = 34, ///< unsigned greater than |
744 | ICMP_UGE = 35, ///< unsigned greater or equal |
745 | ICMP_ULT = 36, ///< unsigned less than |
746 | ICMP_ULE = 37, ///< unsigned less or equal |
747 | ICMP_SGT = 38, ///< signed greater than |
748 | ICMP_SGE = 39, ///< signed greater or equal |
749 | ICMP_SLT = 40, ///< signed less than |
750 | ICMP_SLE = 41, ///< signed less or equal |
751 | FIRST_ICMP_PREDICATE = ICMP_EQ, |
752 | LAST_ICMP_PREDICATE = ICMP_SLE, |
753 | BAD_ICMP_PREDICATE = ICMP_SLE + 1 |
754 | }; |
755 | using PredicateField = |
756 | Bitfield::Element<Predicate, 0, 6, LAST_ICMP_PREDICATE>; |
757 | |
758 | protected: |
759 | CmpInst(Type *ty, Instruction::OtherOps op, Predicate pred, |
760 | Value *LHS, Value *RHS, const Twine &Name = "", |
761 | Instruction *InsertBefore = nullptr, |
762 | Instruction *FlagsSource = nullptr); |
763 | |
764 | CmpInst(Type *ty, Instruction::OtherOps op, Predicate pred, |
765 | Value *LHS, Value *RHS, const Twine &Name, |
766 | BasicBlock *InsertAtEnd); |
767 | |
768 | public: |
769 | // allocate space for exactly two operands |
770 | void *operator new(size_t S) { return User::operator new(S, 2); } |
771 | void operator delete(void *Ptr) { User::operator delete(Ptr); } |
772 | |
773 | /// Construct a compare instruction, given the opcode, the predicate and |
774 | /// the two operands. Optionally (if InstBefore is specified) insert the |
775 | /// instruction into a BasicBlock right before the specified instruction. |
776 | /// The specified Instruction is allowed to be a dereferenced end iterator. |
777 | /// Create a CmpInst |
778 | static CmpInst *Create(OtherOps Op, |
779 | Predicate predicate, Value *S1, |
780 | Value *S2, const Twine &Name = "", |
781 | Instruction *InsertBefore = nullptr); |
782 | |
783 | /// Construct a compare instruction, given the opcode, the predicate and the |
784 | /// two operands. Also automatically insert this instruction to the end of |
785 | /// the BasicBlock specified. |
786 | /// Create a CmpInst |
787 | static CmpInst *Create(OtherOps Op, Predicate predicate, Value *S1, |
788 | Value *S2, const Twine &Name, BasicBlock *InsertAtEnd); |
789 | |
790 | /// Get the opcode casted to the right type |
791 | OtherOps getOpcode() const { |
792 | return static_cast<OtherOps>(Instruction::getOpcode()); |
793 | } |
794 | |
795 | /// Return the predicate for this instruction. |
796 | Predicate getPredicate() const { return getSubclassData<PredicateField>(); } |
797 | |
798 | /// Set the predicate for this instruction to the specified value. |
799 | void setPredicate(Predicate P) { setSubclassData<PredicateField>(P); } |
800 | |
801 | static bool isFPPredicate(Predicate P) { |
802 | static_assert(FIRST_FCMP_PREDICATE == 0, |
803 | "FIRST_FCMP_PREDICATE is required to be 0"); |
804 | return P <= LAST_FCMP_PREDICATE; |
805 | } |
806 | |
807 | static bool isIntPredicate(Predicate P) { |
808 | return P >= FIRST_ICMP_PREDICATE && P <= LAST_ICMP_PREDICATE; |
809 | } |
810 | |
811 | static StringRef getPredicateName(Predicate P); |
812 | |
813 | bool isFPPredicate() const { return isFPPredicate(getPredicate()); } |
814 | bool isIntPredicate() const { return isIntPredicate(getPredicate()); } |
815 | |
816 | /// For example, EQ -> NE, UGT -> ULE, SLT -> SGE, |
817 | /// OEQ -> UNE, UGT -> OLE, OLT -> UGE, etc. |
818 | /// @returns the inverse predicate for the instruction's current predicate. |
819 | /// Return the inverse of the instruction's predicate. |
820 | Predicate getInversePredicate() const { |
821 | return getInversePredicate(getPredicate()); |
822 | } |
823 | |
824 | /// For example, EQ -> NE, UGT -> ULE, SLT -> SGE, |
825 | /// OEQ -> UNE, UGT -> OLE, OLT -> UGE, etc. |
826 | /// @returns the inverse predicate for predicate provided in \p pred. |
827 | /// Return the inverse of a given predicate |
828 | static Predicate getInversePredicate(Predicate pred); |
829 | |
830 | /// For example, EQ->EQ, SLE->SGE, ULT->UGT, |
831 | /// OEQ->OEQ, ULE->UGE, OLT->OGT, etc. |
832 | /// @returns the predicate that would be the result of exchanging the two |
833 | /// operands of the CmpInst instruction without changing the result |
834 | /// produced. |
835 | /// Return the predicate as if the operands were swapped |
836 | Predicate getSwappedPredicate() const { |
837 | return getSwappedPredicate(getPredicate()); |
838 | } |
839 | |
840 | /// This is a static version that you can use without an instruction |
841 | /// available. |
842 | /// Return the predicate as if the operands were swapped. |
843 | static Predicate getSwappedPredicate(Predicate pred); |
844 | |
845 | /// This is a static version that you can use without an instruction |
846 | /// available. |
847 | /// @returns true if the comparison predicate is strict, false otherwise. |
848 | static bool isStrictPredicate(Predicate predicate); |
849 | |
850 | /// @returns true if the comparison predicate is strict, false otherwise. |
851 | /// Determine if this instruction is using an strict comparison predicate. |
852 | bool isStrictPredicate() const { return isStrictPredicate(getPredicate()); } |
853 | |
854 | /// This is a static version that you can use without an instruction |
855 | /// available. |
856 | /// @returns true if the comparison predicate is non-strict, false otherwise. |
857 | static bool isNonStrictPredicate(Predicate predicate); |
858 | |
859 | /// @returns true if the comparison predicate is non-strict, false otherwise. |
860 | /// Determine if this instruction is using an non-strict comparison predicate. |
861 | bool isNonStrictPredicate() const { |
862 | return isNonStrictPredicate(getPredicate()); |
863 | } |
864 | |
865 | /// For example, SGE -> SGT, SLE -> SLT, ULE -> ULT, UGE -> UGT. |
866 | /// Returns the strict version of non-strict comparisons. |
867 | Predicate getStrictPredicate() const { |
868 | return getStrictPredicate(getPredicate()); |
869 | } |
870 | |
871 | /// This is a static version that you can use without an instruction |
872 | /// available. |
873 | /// @returns the strict version of comparison provided in \p pred. |
874 | /// If \p pred is not a strict comparison predicate, returns \p pred. |
875 | /// Returns the strict version of non-strict comparisons. |
876 | static Predicate getStrictPredicate(Predicate pred); |
877 | |
878 | /// For example, SGT -> SGE, SLT -> SLE, ULT -> ULE, UGT -> UGE. |
879 | /// Returns the non-strict version of strict comparisons. |
880 | Predicate getNonStrictPredicate() const { |
881 | return getNonStrictPredicate(getPredicate()); |
882 | } |
883 | |
884 | /// This is a static version that you can use without an instruction |
885 | /// available. |
886 | /// @returns the non-strict version of comparison provided in \p pred. |
887 | /// If \p pred is not a strict comparison predicate, returns \p pred. |
888 | /// Returns the non-strict version of strict comparisons. |
889 | static Predicate getNonStrictPredicate(Predicate pred); |
890 | |
891 | /// This is a static version that you can use without an instruction |
892 | /// available. |
893 | /// Return the flipped strictness of predicate |
894 | static Predicate getFlippedStrictnessPredicate(Predicate pred); |
895 | |
896 | /// For predicate of kind "is X or equal to 0" returns the predicate "is X". |
897 | /// For predicate of kind "is X" returns the predicate "is X or equal to 0". |
898 | /// does not support other kind of predicates. |
899 | /// @returns the predicate that does not contains is equal to zero if |
900 | /// it had and vice versa. |
901 | /// Return the flipped strictness of predicate |
902 | Predicate getFlippedStrictnessPredicate() const { |
903 | return getFlippedStrictnessPredicate(getPredicate()); |
904 | } |
905 | |
906 | /// Provide more efficient getOperand methods. |
907 | DECLARE_TRANSPARENT_OPERAND_ACCESSORS(Value)public: inline Value *getOperand(unsigned) const; inline void setOperand(unsigned, Value*); inline op_iterator op_begin(); inline const_op_iterator op_begin() const; inline op_iterator op_end(); inline const_op_iterator op_end() const; protected : template <int> inline Use &Op(); template <int > inline const Use &Op() const; public: inline unsigned getNumOperands() const; |
908 | |
909 | /// This is just a convenience that dispatches to the subclasses. |
910 | /// Swap the operands and adjust predicate accordingly to retain |
911 | /// the same comparison. |
912 | void swapOperands(); |
913 | |
914 | /// This is just a convenience that dispatches to the subclasses. |
915 | /// Determine if this CmpInst is commutative. |
916 | bool isCommutative() const; |
917 | |
918 | /// Determine if this is an equals/not equals predicate. |
919 | /// This is a static version that you can use without an instruction |
920 | /// available. |
921 | static bool isEquality(Predicate pred); |
922 | |
923 | /// Determine if this is an equals/not equals predicate. |
924 | bool isEquality() const { return isEquality(getPredicate()); } |
925 | |
926 | /// Return true if the predicate is relational (not EQ or NE). |
927 | static bool isRelational(Predicate P) { return !isEquality(P); } |
928 | |
929 | /// Return true if the predicate is relational (not EQ or NE). |
930 | bool isRelational() const { return !isEquality(); } |
931 | |
932 | /// @returns true if the comparison is signed, false otherwise. |
933 | /// Determine if this instruction is using a signed comparison. |
934 | bool isSigned() const { |
935 | return isSigned(getPredicate()); |
936 | } |
937 | |
938 | /// @returns true if the comparison is unsigned, false otherwise. |
939 | /// Determine if this instruction is using an unsigned comparison. |
940 | bool isUnsigned() const { |
941 | return isUnsigned(getPredicate()); |
942 | } |
943 | |
944 | /// For example, ULT->SLT, ULE->SLE, UGT->SGT, UGE->SGE, SLT->Failed assert |
945 | /// @returns the signed version of the unsigned predicate pred. |
946 | /// return the signed version of a predicate |
947 | static Predicate getSignedPredicate(Predicate pred); |
948 | |
949 | /// For example, ULT->SLT, ULE->SLE, UGT->SGT, UGE->SGE, SLT->Failed assert |
950 | /// @returns the signed version of the predicate for this instruction (which |
951 | /// has to be an unsigned predicate). |
952 | /// return the signed version of a predicate |
953 | Predicate getSignedPredicate() { |
954 | return getSignedPredicate(getPredicate()); |
955 | } |
956 | |
957 | /// For example, SLT->ULT, SLE->ULE, SGT->UGT, SGE->UGE, ULT->Failed assert |
958 | /// @returns the unsigned version of the signed predicate pred. |
959 | static Predicate getUnsignedPredicate(Predicate pred); |
960 | |
961 | /// For example, SLT->ULT, SLE->ULE, SGT->UGT, SGE->UGE, ULT->Failed assert |
962 | /// @returns the unsigned version of the predicate for this instruction (which |
963 | /// has to be an signed predicate). |
964 | /// return the unsigned version of a predicate |
965 | Predicate getUnsignedPredicate() { |
966 | return getUnsignedPredicate(getPredicate()); |
967 | } |
968 | |
969 | /// For example, SLT->ULT, ULT->SLT, SLE->ULE, ULE->SLE, EQ->Failed assert |
970 | /// @returns the unsigned version of the signed predicate pred or |
971 | /// the signed version of the signed predicate pred. |
972 | static Predicate getFlippedSignednessPredicate(Predicate pred); |
973 | |
974 | /// For example, SLT->ULT, ULT->SLT, SLE->ULE, ULE->SLE, EQ->Failed assert |
975 | /// @returns the unsigned version of the signed predicate pred or |
976 | /// the signed version of the signed predicate pred. |
977 | Predicate getFlippedSignednessPredicate() { |
978 | return getFlippedSignednessPredicate(getPredicate()); |
979 | } |
980 | |
981 | /// This is just a convenience. |
982 | /// Determine if this is true when both operands are the same. |
983 | bool isTrueWhenEqual() const { |
984 | return isTrueWhenEqual(getPredicate()); |
985 | } |
986 | |
987 | /// This is just a convenience. |
988 | /// Determine if this is false when both operands are the same. |
989 | bool isFalseWhenEqual() const { |
990 | return isFalseWhenEqual(getPredicate()); |
991 | } |
992 | |
993 | /// @returns true if the predicate is unsigned, false otherwise. |
994 | /// Determine if the predicate is an unsigned operation. |
995 | static bool isUnsigned(Predicate predicate); |
996 | |
997 | /// @returns true if the predicate is signed, false otherwise. |
998 | /// Determine if the predicate is an signed operation. |
999 | static bool isSigned(Predicate predicate); |
1000 | |
1001 | /// Determine if the predicate is an ordered operation. |
1002 | static bool isOrdered(Predicate predicate); |
1003 | |
1004 | /// Determine if the predicate is an unordered operation. |
1005 | static bool isUnordered(Predicate predicate); |
1006 | |
1007 | /// Determine if the predicate is true when comparing a value with itself. |
1008 | static bool isTrueWhenEqual(Predicate predicate); |
1009 | |
1010 | /// Determine if the predicate is false when comparing a value with itself. |
1011 | static bool isFalseWhenEqual(Predicate predicate); |
1012 | |
1013 | /// Determine if Pred1 implies Pred2 is true when two compares have matching |
1014 | /// operands. |
1015 | static bool isImpliedTrueByMatchingCmp(Predicate Pred1, Predicate Pred2); |
1016 | |
1017 | /// Determine if Pred1 implies Pred2 is false when two compares have matching |
1018 | /// operands. |
1019 | static bool isImpliedFalseByMatchingCmp(Predicate Pred1, Predicate Pred2); |
1020 | |
1021 | /// Methods for support type inquiry through isa, cast, and dyn_cast: |
1022 | static bool classof(const Instruction *I) { |
1023 | return I->getOpcode() == Instruction::ICmp || |
1024 | I->getOpcode() == Instruction::FCmp; |
1025 | } |
1026 | static bool classof(const Value *V) { |
1027 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
1028 | } |
1029 | |
1030 | /// Create a result type for fcmp/icmp |
1031 | static Type* makeCmpResultType(Type* opnd_type) { |
1032 | if (VectorType* vt = dyn_cast<VectorType>(opnd_type)) { |
1033 | return VectorType::get(Type::getInt1Ty(opnd_type->getContext()), |
1034 | vt->getElementCount()); |
1035 | } |
1036 | return Type::getInt1Ty(opnd_type->getContext()); |
1037 | } |
1038 | |
1039 | private: |
1040 | // Shadow Value::setValueSubclassData with a private forwarding method so that |
1041 | // subclasses cannot accidentally use it. |
1042 | void setValueSubclassData(unsigned short D) { |
1043 | Value::setValueSubclassData(D); |
1044 | } |
1045 | }; |
1046 | |
1047 | // FIXME: these are redundant if CmpInst < BinaryOperator |
1048 | template <> |
1049 | struct OperandTraits<CmpInst> : public FixedNumOperandTraits<CmpInst, 2> { |
1050 | }; |
1051 | |
1052 | DEFINE_TRANSPARENT_OPERAND_ACCESSORS(CmpInst, Value)CmpInst::op_iterator CmpInst::op_begin() { return OperandTraits <CmpInst>::op_begin(this); } CmpInst::const_op_iterator CmpInst::op_begin() const { return OperandTraits<CmpInst> ::op_begin(const_cast<CmpInst*>(this)); } CmpInst::op_iterator CmpInst::op_end() { return OperandTraits<CmpInst>::op_end (this); } CmpInst::const_op_iterator CmpInst::op_end() const { return OperandTraits<CmpInst>::op_end(const_cast<CmpInst *>(this)); } Value *CmpInst::getOperand(unsigned i_nocapture ) const { ((void)0); return cast_or_null<Value>( OperandTraits <CmpInst>::op_begin(const_cast<CmpInst*>(this))[i_nocapture ].get()); } void CmpInst::setOperand(unsigned i_nocapture, Value *Val_nocapture) { ((void)0); OperandTraits<CmpInst>::op_begin (this)[i_nocapture] = Val_nocapture; } unsigned CmpInst::getNumOperands () const { return OperandTraits<CmpInst>::operands(this ); } template <int Idx_nocapture> Use &CmpInst::Op( ) { return this->OpFrom<Idx_nocapture>(this); } template <int Idx_nocapture> const Use &CmpInst::Op() const { return this->OpFrom<Idx_nocapture>(this); } |
1053 | |
1054 | /// A lightweight accessor for an operand bundle meant to be passed |
1055 | /// around by value. |
1056 | struct OperandBundleUse { |
1057 | ArrayRef<Use> Inputs; |
1058 | |
1059 | OperandBundleUse() = default; |
1060 | explicit OperandBundleUse(StringMapEntry<uint32_t> *Tag, ArrayRef<Use> Inputs) |
1061 | : Inputs(Inputs), Tag(Tag) {} |
1062 | |
1063 | /// Return true if the operand at index \p Idx in this operand bundle |
1064 | /// has the attribute A. |
1065 | bool operandHasAttr(unsigned Idx, Attribute::AttrKind A) const { |
1066 | if (isDeoptOperandBundle()) |
1067 | if (A == Attribute::ReadOnly || A == Attribute::NoCapture) |
1068 | return Inputs[Idx]->getType()->isPointerTy(); |
1069 | |
1070 | // Conservative answer: no operands have any attributes. |
1071 | return false; |
1072 | } |
1073 | |
1074 | /// Return the tag of this operand bundle as a string. |
1075 | StringRef getTagName() const { |
1076 | return Tag->getKey(); |
1077 | } |
1078 | |
1079 | /// Return the tag of this operand bundle as an integer. |
1080 | /// |
1081 | /// Operand bundle tags are interned by LLVMContextImpl::getOrInsertBundleTag, |
1082 | /// and this function returns the unique integer getOrInsertBundleTag |
1083 | /// associated the tag of this operand bundle to. |
1084 | uint32_t getTagID() const { |
1085 | return Tag->getValue(); |
1086 | } |
1087 | |
1088 | /// Return true if this is a "deopt" operand bundle. |
1089 | bool isDeoptOperandBundle() const { |
1090 | return getTagID() == LLVMContext::OB_deopt; |
1091 | } |
1092 | |
1093 | /// Return true if this is a "funclet" operand bundle. |
1094 | bool isFuncletOperandBundle() const { |
1095 | return getTagID() == LLVMContext::OB_funclet; |
1096 | } |
1097 | |
1098 | /// Return true if this is a "cfguardtarget" operand bundle. |
1099 | bool isCFGuardTargetOperandBundle() const { |
1100 | return getTagID() == LLVMContext::OB_cfguardtarget; |
1101 | } |
1102 | |
1103 | private: |
1104 | /// Pointer to an entry in LLVMContextImpl::getOrInsertBundleTag. |
1105 | StringMapEntry<uint32_t> *Tag; |
1106 | }; |
1107 | |
1108 | /// A container for an operand bundle being viewed as a set of values |
1109 | /// rather than a set of uses. |
1110 | /// |
1111 | /// Unlike OperandBundleUse, OperandBundleDefT owns the memory it carries, and |
1112 | /// so it is possible to create and pass around "self-contained" instances of |
1113 | /// OperandBundleDef and ConstOperandBundleDef. |
1114 | template <typename InputTy> class OperandBundleDefT { |
1115 | std::string Tag; |
1116 | std::vector<InputTy> Inputs; |
1117 | |
1118 | public: |
1119 | explicit OperandBundleDefT(std::string Tag, std::vector<InputTy> Inputs) |
1120 | : Tag(std::move(Tag)), Inputs(std::move(Inputs)) {} |
1121 | explicit OperandBundleDefT(std::string Tag, ArrayRef<InputTy> Inputs) |
1122 | : Tag(std::move(Tag)), Inputs(Inputs) {} |
1123 | |
1124 | explicit OperandBundleDefT(const OperandBundleUse &OBU) { |
1125 | Tag = std::string(OBU.getTagName()); |
1126 | llvm::append_range(Inputs, OBU.Inputs); |
1127 | } |
1128 | |
1129 | ArrayRef<InputTy> inputs() const { return Inputs; } |
1130 | |
1131 | using input_iterator = typename std::vector<InputTy>::const_iterator; |
1132 | |
1133 | size_t input_size() const { return Inputs.size(); } |
1134 | input_iterator input_begin() const { return Inputs.begin(); } |
1135 | input_iterator input_end() const { return Inputs.end(); } |
1136 | |
1137 | StringRef getTag() const { return Tag; } |
1138 | }; |
1139 | |
1140 | using OperandBundleDef = OperandBundleDefT<Value *>; |
1141 | using ConstOperandBundleDef = OperandBundleDefT<const Value *>; |
1142 | |
1143 | //===----------------------------------------------------------------------===// |
1144 | // CallBase Class |
1145 | //===----------------------------------------------------------------------===// |
1146 | |
1147 | /// Base class for all callable instructions (InvokeInst and CallInst) |
1148 | /// Holds everything related to calling a function. |
1149 | /// |
1150 | /// All call-like instructions are required to use a common operand layout: |
1151 | /// - Zero or more arguments to the call, |
1152 | /// - Zero or more operand bundles with zero or more operand inputs each |
1153 | /// bundle, |
1154 | /// - Zero or more subclass controlled operands |
1155 | /// - The called function. |
1156 | /// |
1157 | /// This allows this base class to easily access the called function and the |
1158 | /// start of the arguments without knowing how many other operands a particular |
1159 | /// subclass requires. Note that accessing the end of the argument list isn't |
1160 | /// as cheap as most other operations on the base class. |
1161 | class CallBase : public Instruction { |
1162 | protected: |
1163 | // The first two bits are reserved by CallInst for fast retrieval, |
1164 | using CallInstReservedField = Bitfield::Element<unsigned, 0, 2>; |
1165 | using CallingConvField = |
1166 | Bitfield::Element<CallingConv::ID, CallInstReservedField::NextBit, 10, |
1167 | CallingConv::MaxID>; |
1168 | static_assert( |
1169 | Bitfield::areContiguous<CallInstReservedField, CallingConvField>(), |
1170 | "Bitfields must be contiguous"); |
1171 | |
1172 | /// The last operand is the called operand. |
1173 | static constexpr int CalledOperandOpEndIdx = -1; |
1174 | |
1175 | AttributeList Attrs; ///< parameter attributes for callable |
1176 | FunctionType *FTy; |
1177 | |
1178 | template <class... ArgsTy> |
1179 | CallBase(AttributeList const &A, FunctionType *FT, ArgsTy &&... Args) |
1180 | : Instruction(std::forward<ArgsTy>(Args)...), Attrs(A), FTy(FT) {} |
1181 | |
1182 | using Instruction::Instruction; |
1183 | |
1184 | bool hasDescriptor() const { return Value::HasDescriptor; } |
1185 | |
1186 | unsigned getNumSubclassExtraOperands() const { |
1187 | switch (getOpcode()) { |
1188 | case Instruction::Call: |
1189 | return 0; |
1190 | case Instruction::Invoke: |
1191 | return 2; |
1192 | case Instruction::CallBr: |
1193 | return getNumSubclassExtraOperandsDynamic(); |
1194 | } |
1195 | llvm_unreachable("Invalid opcode!")__builtin_unreachable(); |
1196 | } |
1197 | |
1198 | /// Get the number of extra operands for instructions that don't have a fixed |
1199 | /// number of extra operands. |
1200 | unsigned getNumSubclassExtraOperandsDynamic() const; |
1201 | |
1202 | public: |
1203 | using Instruction::getContext; |
1204 | |
1205 | /// Create a clone of \p CB with a different set of operand bundles and |
1206 | /// insert it before \p InsertPt. |
1207 | /// |
1208 | /// The returned call instruction is identical \p CB in every way except that |
1209 | /// the operand bundles for the new instruction are set to the operand bundles |
1210 | /// in \p Bundles. |
1211 | static CallBase *Create(CallBase *CB, ArrayRef<OperandBundleDef> Bundles, |
1212 | Instruction *InsertPt = nullptr); |
1213 | |
1214 | /// Create a clone of \p CB with the operand bundle with the tag matching |
1215 | /// \p Bundle's tag replaced with Bundle, and insert it before \p InsertPt. |
1216 | /// |
1217 | /// The returned call instruction is identical \p CI in every way except that |
1218 | /// the specified operand bundle has been replaced. |
1219 | static CallBase *Create(CallBase *CB, |
1220 | OperandBundleDef Bundle, |
1221 | Instruction *InsertPt = nullptr); |
1222 | |
1223 | /// Create a clone of \p CB with operand bundle \p OB added. |
1224 | static CallBase *addOperandBundle(CallBase *CB, uint32_t ID, |
1225 | OperandBundleDef OB, |
1226 | Instruction *InsertPt = nullptr); |
1227 | |
1228 | /// Create a clone of \p CB with operand bundle \p ID removed. |
1229 | static CallBase *removeOperandBundle(CallBase *CB, uint32_t ID, |
1230 | Instruction *InsertPt = nullptr); |
1231 | |
1232 | static bool classof(const Instruction *I) { |
1233 | return I->getOpcode() == Instruction::Call || |
1234 | I->getOpcode() == Instruction::Invoke || |
1235 | I->getOpcode() == Instruction::CallBr; |
1236 | } |
1237 | static bool classof(const Value *V) { |
1238 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
1239 | } |
1240 | |
1241 | FunctionType *getFunctionType() const { return FTy; } |
1242 | |
1243 | void mutateFunctionType(FunctionType *FTy) { |
1244 | Value::mutateType(FTy->getReturnType()); |
1245 | this->FTy = FTy; |
1246 | } |
1247 | |
1248 | DECLARE_TRANSPARENT_OPERAND_ACCESSORS(Value)public: inline Value *getOperand(unsigned) const; inline void setOperand(unsigned, Value*); inline op_iterator op_begin(); inline const_op_iterator op_begin() const; inline op_iterator op_end(); inline const_op_iterator op_end() const; protected : template <int> inline Use &Op(); template <int > inline const Use &Op() const; public: inline unsigned getNumOperands() const; |
1249 | |
1250 | /// data_operands_begin/data_operands_end - Return iterators iterating over |
1251 | /// the call / invoke argument list and bundle operands. For invokes, this is |
1252 | /// the set of instruction operands except the invoke target and the two |
1253 | /// successor blocks; and for calls this is the set of instruction operands |
1254 | /// except the call target. |
1255 | User::op_iterator data_operands_begin() { return op_begin(); } |
1256 | User::const_op_iterator data_operands_begin() const { |
1257 | return const_cast<CallBase *>(this)->data_operands_begin(); |
1258 | } |
1259 | User::op_iterator data_operands_end() { |
1260 | // Walk from the end of the operands over the called operand and any |
1261 | // subclass operands. |
1262 | return op_end() - getNumSubclassExtraOperands() - 1; |
1263 | } |
1264 | User::const_op_iterator data_operands_end() const { |
1265 | return const_cast<CallBase *>(this)->data_operands_end(); |
1266 | } |
1267 | iterator_range<User::op_iterator> data_ops() { |
1268 | return make_range(data_operands_begin(), data_operands_end()); |
1269 | } |
1270 | iterator_range<User::const_op_iterator> data_ops() const { |
1271 | return make_range(data_operands_begin(), data_operands_end()); |
1272 | } |
1273 | bool data_operands_empty() const { |
1274 | return data_operands_end() == data_operands_begin(); |
1275 | } |
1276 | unsigned data_operands_size() const { |
1277 | return std::distance(data_operands_begin(), data_operands_end()); |
1278 | } |
1279 | |
1280 | bool isDataOperand(const Use *U) const { |
1281 | assert(this == U->getUser() &&((void)0) |
1282 | "Only valid to query with a use of this instruction!")((void)0); |
1283 | return data_operands_begin() <= U && U < data_operands_end(); |
1284 | } |
1285 | bool isDataOperand(Value::const_user_iterator UI) const { |
1286 | return isDataOperand(&UI.getUse()); |
1287 | } |
1288 | |
1289 | /// Given a value use iterator, return the data operand corresponding to it. |
1290 | /// Iterator must actually correspond to a data operand. |
1291 | unsigned getDataOperandNo(Value::const_user_iterator UI) const { |
1292 | return getDataOperandNo(&UI.getUse()); |
1293 | } |
1294 | |
1295 | /// Given a use for a data operand, get the data operand number that |
1296 | /// corresponds to it. |
1297 | unsigned getDataOperandNo(const Use *U) const { |
1298 | assert(isDataOperand(U) && "Data operand # out of range!")((void)0); |
1299 | return U - data_operands_begin(); |
1300 | } |
1301 | |
1302 | /// Return the iterator pointing to the beginning of the argument list. |
1303 | User::op_iterator arg_begin() { return op_begin(); } |
1304 | User::const_op_iterator arg_begin() const { |
1305 | return const_cast<CallBase *>(this)->arg_begin(); |
1306 | } |
1307 | |
1308 | /// Return the iterator pointing to the end of the argument list. |
1309 | User::op_iterator arg_end() { |
1310 | // From the end of the data operands, walk backwards past the bundle |
1311 | // operands. |
1312 | return data_operands_end() - getNumTotalBundleOperands(); |
1313 | } |
1314 | User::const_op_iterator arg_end() const { |
1315 | return const_cast<CallBase *>(this)->arg_end(); |
1316 | } |
1317 | |
1318 | /// Iteration adapter for range-for loops. |
1319 | iterator_range<User::op_iterator> args() { |
1320 | return make_range(arg_begin(), arg_end()); |
1321 | } |
1322 | iterator_range<User::const_op_iterator> args() const { |
1323 | return make_range(arg_begin(), arg_end()); |
1324 | } |
1325 | bool arg_empty() const { return arg_end() == arg_begin(); } |
1326 | unsigned arg_size() const { return arg_end() - arg_begin(); } |
1327 | |
1328 | // Legacy API names that duplicate the above and will be removed once users |
1329 | // are migrated. |
1330 | iterator_range<User::op_iterator> arg_operands() { |
1331 | return make_range(arg_begin(), arg_end()); |
1332 | } |
1333 | iterator_range<User::const_op_iterator> arg_operands() const { |
1334 | return make_range(arg_begin(), arg_end()); |
1335 | } |
1336 | unsigned getNumArgOperands() const { return arg_size(); } |
1337 | |
1338 | Value *getArgOperand(unsigned i) const { |
1339 | assert(i < getNumArgOperands() && "Out of bounds!")((void)0); |
1340 | return getOperand(i); |
1341 | } |
1342 | |
1343 | void setArgOperand(unsigned i, Value *v) { |
1344 | assert(i < getNumArgOperands() && "Out of bounds!")((void)0); |
1345 | setOperand(i, v); |
1346 | } |
1347 | |
1348 | /// Wrappers for getting the \c Use of a call argument. |
1349 | const Use &getArgOperandUse(unsigned i) const { |
1350 | assert(i < getNumArgOperands() && "Out of bounds!")((void)0); |
1351 | return User::getOperandUse(i); |
1352 | } |
1353 | Use &getArgOperandUse(unsigned i) { |
1354 | assert(i < getNumArgOperands() && "Out of bounds!")((void)0); |
1355 | return User::getOperandUse(i); |
1356 | } |
1357 | |
1358 | bool isArgOperand(const Use *U) const { |
1359 | assert(this == U->getUser() &&((void)0) |
1360 | "Only valid to query with a use of this instruction!")((void)0); |
1361 | return arg_begin() <= U && U < arg_end(); |
1362 | } |
1363 | bool isArgOperand(Value::const_user_iterator UI) const { |
1364 | return isArgOperand(&UI.getUse()); |
1365 | } |
1366 | |
1367 | /// Given a use for a arg operand, get the arg operand number that |
1368 | /// corresponds to it. |
1369 | unsigned getArgOperandNo(const Use *U) const { |
1370 | assert(isArgOperand(U) && "Arg operand # out of range!")((void)0); |
1371 | return U - arg_begin(); |
1372 | } |
1373 | |
1374 | /// Given a value use iterator, return the arg operand number corresponding to |
1375 | /// it. Iterator must actually correspond to a data operand. |
1376 | unsigned getArgOperandNo(Value::const_user_iterator UI) const { |
1377 | return getArgOperandNo(&UI.getUse()); |
1378 | } |
1379 | |
1380 | /// Returns true if this CallSite passes the given Value* as an argument to |
1381 | /// the called function. |
1382 | bool hasArgument(const Value *V) const { |
1383 | return llvm::is_contained(args(), V); |
1384 | } |
1385 | |
1386 | Value *getCalledOperand() const { return Op<CalledOperandOpEndIdx>(); } |
1387 | |
1388 | const Use &getCalledOperandUse() const { return Op<CalledOperandOpEndIdx>(); } |
1389 | Use &getCalledOperandUse() { return Op<CalledOperandOpEndIdx>(); } |
1390 | |
1391 | /// Returns the function called, or null if this is an |
1392 | /// indirect function invocation. |
1393 | Function *getCalledFunction() const { |
1394 | return dyn_cast_or_null<Function>(getCalledOperand()); |
1395 | } |
1396 | |
1397 | /// Return true if the callsite is an indirect call. |
1398 | bool isIndirectCall() const; |
1399 | |
1400 | /// Determine whether the passed iterator points to the callee operand's Use. |
1401 | bool isCallee(Value::const_user_iterator UI) const { |
1402 | return isCallee(&UI.getUse()); |
1403 | } |
1404 | |
1405 | /// Determine whether this Use is the callee operand's Use. |
1406 | bool isCallee(const Use *U) const { return &getCalledOperandUse() == U; } |
1407 | |
1408 | /// Helper to get the caller (the parent function). |
1409 | Function *getCaller(); |
1410 | const Function *getCaller() const { |
1411 | return const_cast<CallBase *>(this)->getCaller(); |
1412 | } |
1413 | |
1414 | /// Tests if this call site must be tail call optimized. Only a CallInst can |
1415 | /// be tail call optimized. |
1416 | bool isMustTailCall() const; |
1417 | |
1418 | /// Tests if this call site is marked as a tail call. |
1419 | bool isTailCall() const; |
1420 | |
1421 | /// Returns the intrinsic ID of the intrinsic called or |
1422 | /// Intrinsic::not_intrinsic if the called function is not an intrinsic, or if |
1423 | /// this is an indirect call. |
1424 | Intrinsic::ID getIntrinsicID() const; |
1425 | |
1426 | void setCalledOperand(Value *V) { Op<CalledOperandOpEndIdx>() = V; } |
1427 | |
1428 | /// Sets the function called, including updating the function type. |
1429 | void setCalledFunction(Function *Fn) { |
1430 | setCalledFunction(Fn->getFunctionType(), Fn); |
1431 | } |
1432 | |
1433 | /// Sets the function called, including updating the function type. |
1434 | void setCalledFunction(FunctionCallee Fn) { |
1435 | setCalledFunction(Fn.getFunctionType(), Fn.getCallee()); |
1436 | } |
1437 | |
1438 | /// Sets the function called, including updating to the specified function |
1439 | /// type. |
1440 | void setCalledFunction(FunctionType *FTy, Value *Fn) { |
1441 | this->FTy = FTy; |
1442 | assert(cast<PointerType>(Fn->getType())->isOpaqueOrPointeeTypeMatches(FTy))((void)0); |
1443 | // This function doesn't mutate the return type, only the function |
1444 | // type. Seems broken, but I'm just gonna stick an assert in for now. |
1445 | assert(getType() == FTy->getReturnType())((void)0); |
1446 | setCalledOperand(Fn); |
1447 | } |
1448 | |
1449 | CallingConv::ID getCallingConv() const { |
1450 | return getSubclassData<CallingConvField>(); |
1451 | } |
1452 | |
1453 | void setCallingConv(CallingConv::ID CC) { |
1454 | setSubclassData<CallingConvField>(CC); |
1455 | } |
1456 | |
1457 | /// Check if this call is an inline asm statement. |
1458 | bool isInlineAsm() const { return isa<InlineAsm>(getCalledOperand()); } |
1459 | |
1460 | /// \name Attribute API |
1461 | /// |
1462 | /// These methods access and modify attributes on this call (including |
1463 | /// looking through to the attributes on the called function when necessary). |
1464 | ///@{ |
1465 | |
1466 | /// Return the parameter attributes for this call. |
1467 | /// |
1468 | AttributeList getAttributes() const { return Attrs; } |
1469 | |
1470 | /// Set the parameter attributes for this call. |
1471 | /// |
1472 | void setAttributes(AttributeList A) { Attrs = A; } |
1473 | |
1474 | /// Determine whether this call has the given attribute. If it does not |
1475 | /// then determine if the called function has the attribute, but only if |
1476 | /// the attribute is allowed for the call. |
1477 | bool hasFnAttr(Attribute::AttrKind Kind) const { |
1478 | assert(Kind != Attribute::NoBuiltin &&((void)0) |
1479 | "Use CallBase::isNoBuiltin() to check for Attribute::NoBuiltin")((void)0); |
1480 | return hasFnAttrImpl(Kind); |
1481 | } |
1482 | |
1483 | /// Determine whether this call has the given attribute. If it does not |
1484 | /// then determine if the called function has the attribute, but only if |
1485 | /// the attribute is allowed for the call. |
1486 | bool hasFnAttr(StringRef Kind) const { return hasFnAttrImpl(Kind); } |
1487 | |
1488 | /// adds the attribute to the list of attributes. |
1489 | void addAttribute(unsigned i, Attribute::AttrKind Kind) { |
1490 | AttributeList PAL = getAttributes(); |
1491 | PAL = PAL.addAttribute(getContext(), i, Kind); |
1492 | setAttributes(PAL); |
1493 | } |
1494 | |
1495 | /// adds the attribute to the list of attributes. |
1496 | void addAttribute(unsigned i, Attribute Attr) { |
1497 | AttributeList PAL = getAttributes(); |
1498 | PAL = PAL.addAttribute(getContext(), i, Attr); |
1499 | setAttributes(PAL); |
1500 | } |
1501 | |
1502 | /// Adds the attribute to the indicated argument |
1503 | void addParamAttr(unsigned ArgNo, Attribute::AttrKind Kind) { |
1504 | assert(ArgNo < getNumArgOperands() && "Out of bounds")((void)0); |
1505 | AttributeList PAL = getAttributes(); |
1506 | PAL = PAL.addParamAttribute(getContext(), ArgNo, Kind); |
1507 | setAttributes(PAL); |
1508 | } |
1509 | |
1510 | /// Adds the attribute to the indicated argument |
1511 | void addParamAttr(unsigned ArgNo, Attribute Attr) { |
1512 | assert(ArgNo < getNumArgOperands() && "Out of bounds")((void)0); |
1513 | AttributeList PAL = getAttributes(); |
1514 | PAL = PAL.addParamAttribute(getContext(), ArgNo, Attr); |
1515 | setAttributes(PAL); |
1516 | } |
1517 | |
1518 | /// removes the attribute from the list of attributes. |
1519 | void removeAttribute(unsigned i, Attribute::AttrKind Kind) { |
1520 | AttributeList PAL = getAttributes(); |
1521 | PAL = PAL.removeAttribute(getContext(), i, Kind); |
1522 | setAttributes(PAL); |
1523 | } |
1524 | |
1525 | /// removes the attribute from the list of attributes. |
1526 | void removeAttribute(unsigned i, StringRef Kind) { |
1527 | AttributeList PAL = getAttributes(); |
1528 | PAL = PAL.removeAttribute(getContext(), i, Kind); |
1529 | setAttributes(PAL); |
1530 | } |
1531 | |
1532 | void removeAttributes(unsigned i, const AttrBuilder &Attrs) { |
1533 | AttributeList PAL = getAttributes(); |
1534 | PAL = PAL.removeAttributes(getContext(), i, Attrs); |
1535 | setAttributes(PAL); |
1536 | } |
1537 | |
1538 | /// Removes the attribute from the given argument |
1539 | void removeParamAttr(unsigned ArgNo, Attribute::AttrKind Kind) { |
1540 | assert(ArgNo < getNumArgOperands() && "Out of bounds")((void)0); |
1541 | AttributeList PAL = getAttributes(); |
1542 | PAL = PAL.removeParamAttribute(getContext(), ArgNo, Kind); |
1543 | setAttributes(PAL); |
1544 | } |
1545 | |
1546 | /// Removes the attribute from the given argument |
1547 | void removeParamAttr(unsigned ArgNo, StringRef Kind) { |
1548 | assert(ArgNo < getNumArgOperands() && "Out of bounds")((void)0); |
1549 | AttributeList PAL = getAttributes(); |
1550 | PAL = PAL.removeParamAttribute(getContext(), ArgNo, Kind); |
1551 | setAttributes(PAL); |
1552 | } |
1553 | |
1554 | /// Removes the attributes from the given argument |
1555 | void removeParamAttrs(unsigned ArgNo, const AttrBuilder &Attrs) { |
1556 | AttributeList PAL = getAttributes(); |
1557 | PAL = PAL.removeParamAttributes(getContext(), ArgNo, Attrs); |
1558 | setAttributes(PAL); |
1559 | } |
1560 | |
1561 | /// adds the dereferenceable attribute to the list of attributes. |
1562 | void addDereferenceableAttr(unsigned i, uint64_t Bytes) { |
1563 | AttributeList PAL = getAttributes(); |
1564 | PAL = PAL.addDereferenceableAttr(getContext(), i, Bytes); |
1565 | setAttributes(PAL); |
1566 | } |
1567 | |
1568 | /// adds the dereferenceable_or_null attribute to the list of |
1569 | /// attributes. |
1570 | void addDereferenceableOrNullAttr(unsigned i, uint64_t Bytes) { |
1571 | AttributeList PAL = getAttributes(); |
1572 | PAL = PAL.addDereferenceableOrNullAttr(getContext(), i, Bytes); |
1573 | setAttributes(PAL); |
1574 | } |
1575 | |
1576 | /// Determine whether the return value has the given attribute. |
1577 | bool hasRetAttr(Attribute::AttrKind Kind) const { |
1578 | return hasRetAttrImpl(Kind); |
1579 | } |
1580 | /// Determine whether the return value has the given attribute. |
1581 | bool hasRetAttr(StringRef Kind) const { return hasRetAttrImpl(Kind); } |
1582 | |
1583 | /// Determine whether the argument or parameter has the given attribute. |
1584 | bool paramHasAttr(unsigned ArgNo, Attribute::AttrKind Kind) const; |
1585 | |
1586 | /// Get the attribute of a given kind at a position. |
1587 | Attribute getAttribute(unsigned i, Attribute::AttrKind Kind) const { |
1588 | return getAttributes().getAttribute(i, Kind); |
1589 | } |
1590 | |
1591 | /// Get the attribute of a given kind at a position. |
1592 | Attribute getAttribute(unsigned i, StringRef Kind) const { |
1593 | return getAttributes().getAttribute(i, Kind); |
1594 | } |
1595 | |
1596 | /// Get the attribute of a given kind from a given arg |
1597 | Attribute getParamAttr(unsigned ArgNo, Attribute::AttrKind Kind) const { |
1598 | assert(ArgNo < getNumArgOperands() && "Out of bounds")((void)0); |
1599 | return getAttributes().getParamAttr(ArgNo, Kind); |
1600 | } |
1601 | |
1602 | /// Get the attribute of a given kind from a given arg |
1603 | Attribute getParamAttr(unsigned ArgNo, StringRef Kind) const { |
1604 | assert(ArgNo < getNumArgOperands() && "Out of bounds")((void)0); |
1605 | return getAttributes().getParamAttr(ArgNo, Kind); |
1606 | } |
1607 | |
1608 | /// Return true if the data operand at index \p i has the attribute \p |
1609 | /// A. |
1610 | /// |
1611 | /// Data operands include call arguments and values used in operand bundles, |
1612 | /// but does not include the callee operand. This routine dispatches to the |
1613 | /// underlying AttributeList or the OperandBundleUser as appropriate. |
1614 | /// |
1615 | /// The index \p i is interpreted as |
1616 | /// |
1617 | /// \p i == Attribute::ReturnIndex -> the return value |
1618 | /// \p i in [1, arg_size + 1) -> argument number (\p i - 1) |
1619 | /// \p i in [arg_size + 1, data_operand_size + 1) -> bundle operand at index |
1620 | /// (\p i - 1) in the operand list. |
1621 | bool dataOperandHasImpliedAttr(unsigned i, Attribute::AttrKind Kind) const { |
1622 | // Note that we have to add one because `i` isn't zero-indexed. |
1623 | assert(i < (getNumArgOperands() + getNumTotalBundleOperands() + 1) &&((void)0) |
1624 | "Data operand index out of bounds!")((void)0); |
1625 | |
1626 | // The attribute A can either be directly specified, if the operand in |
1627 | // question is a call argument; or be indirectly implied by the kind of its |
1628 | // containing operand bundle, if the operand is a bundle operand. |
1629 | |
1630 | if (i == AttributeList::ReturnIndex) |
1631 | return hasRetAttr(Kind); |
1632 | |
1633 | // FIXME: Avoid these i - 1 calculations and update the API to use |
1634 | // zero-based indices. |
1635 | if (i < (getNumArgOperands() + 1)) |
1636 | return paramHasAttr(i - 1, Kind); |
1637 | |
1638 | assert(hasOperandBundles() && i >= (getBundleOperandsStartIndex() + 1) &&((void)0) |
1639 | "Must be either a call argument or an operand bundle!")((void)0); |
1640 | return bundleOperandHasAttr(i - 1, Kind); |
1641 | } |
1642 | |
1643 | /// Determine whether this data operand is not captured. |
1644 | // FIXME: Once this API is no longer duplicated in `CallSite`, rename this to |
1645 | // better indicate that this may return a conservative answer. |
1646 | bool doesNotCapture(unsigned OpNo) const { |
1647 | return dataOperandHasImpliedAttr(OpNo + 1, Attribute::NoCapture); |
1648 | } |
1649 | |
1650 | /// Determine whether this argument is passed by value. |
1651 | bool isByValArgument(unsigned ArgNo) const { |
1652 | return paramHasAttr(ArgNo, Attribute::ByVal); |
1653 | } |
1654 | |
1655 | /// Determine whether this argument is passed in an alloca. |
1656 | bool isInAllocaArgument(unsigned ArgNo) const { |
1657 | return paramHasAttr(ArgNo, Attribute::InAlloca); |
1658 | } |
1659 | |
1660 | /// Determine whether this argument is passed by value, in an alloca, or is |
1661 | /// preallocated. |
1662 | bool isPassPointeeByValueArgument(unsigned ArgNo) const { |
1663 | return paramHasAttr(ArgNo, Attribute::ByVal) || |
1664 | paramHasAttr(ArgNo, Attribute::InAlloca) || |
1665 | paramHasAttr(ArgNo, Attribute::Preallocated); |
1666 | } |
1667 | |
1668 | /// Determine whether passing undef to this argument is undefined behavior. |
1669 | /// If passing undef to this argument is UB, passing poison is UB as well |
1670 | /// because poison is more undefined than undef. |
1671 | bool isPassingUndefUB(unsigned ArgNo) const { |
1672 | return paramHasAttr(ArgNo, Attribute::NoUndef) || |
1673 | // dereferenceable implies noundef. |
1674 | paramHasAttr(ArgNo, Attribute::Dereferenceable) || |
1675 | // dereferenceable implies noundef, and null is a well-defined value. |
1676 | paramHasAttr(ArgNo, Attribute::DereferenceableOrNull); |
1677 | } |
1678 | |
1679 | /// Determine if there are is an inalloca argument. Only the last argument can |
1680 | /// have the inalloca attribute. |
1681 | bool hasInAllocaArgument() const { |
1682 | return !arg_empty() && paramHasAttr(arg_size() - 1, Attribute::InAlloca); |
1683 | } |
1684 | |
1685 | // FIXME: Once this API is no longer duplicated in `CallSite`, rename this to |
1686 | // better indicate that this may return a conservative answer. |
1687 | bool doesNotAccessMemory(unsigned OpNo) const { |
1688 | return dataOperandHasImpliedAttr(OpNo + 1, Attribute::ReadNone); |
1689 | } |
1690 | |
1691 | // FIXME: Once this API is no longer duplicated in `CallSite`, rename this to |
1692 | // better indicate that this may return a conservative answer. |
1693 | bool onlyReadsMemory(unsigned OpNo) const { |
1694 | return dataOperandHasImpliedAttr(OpNo + 1, Attribute::ReadOnly) || |
1695 | dataOperandHasImpliedAttr(OpNo + 1, Attribute::ReadNone); |
1696 | } |
1697 | |
1698 | // FIXME: Once this API is no longer duplicated in `CallSite`, rename this to |
1699 | // better indicate that this may return a conservative answer. |
1700 | bool doesNotReadMemory(unsigned OpNo) const { |
1701 | return dataOperandHasImpliedAttr(OpNo + 1, Attribute::WriteOnly) || |
1702 | dataOperandHasImpliedAttr(OpNo + 1, Attribute::ReadNone); |
1703 | } |
1704 | |
1705 | /// Extract the alignment of the return value. |
1706 | MaybeAlign getRetAlign() const { return Attrs.getRetAlignment(); } |
1707 | |
1708 | /// Extract the alignment for a call or parameter (0=unknown). |
1709 | MaybeAlign getParamAlign(unsigned ArgNo) const { |
1710 | return Attrs.getParamAlignment(ArgNo); |
1711 | } |
1712 | |
1713 | MaybeAlign getParamStackAlign(unsigned ArgNo) const { |
1714 | return Attrs.getParamStackAlignment(ArgNo); |
1715 | } |
1716 | |
1717 | /// Extract the byval type for a call or parameter. |
1718 | Type *getParamByValType(unsigned ArgNo) const { |
1719 | if (auto *Ty = Attrs.getParamByValType(ArgNo)) |
1720 | return Ty; |
1721 | if (const Function *F = getCalledFunction()) |
1722 | return F->getAttributes().getParamByValType(ArgNo); |
1723 | return nullptr; |
1724 | } |
1725 | |
1726 | /// Extract the preallocated type for a call or parameter. |
1727 | Type *getParamPreallocatedType(unsigned ArgNo) const { |
1728 | if (auto *Ty = Attrs.getParamPreallocatedType(ArgNo)) |
1729 | return Ty; |
1730 | if (const Function *F = getCalledFunction()) |
1731 | return F->getAttributes().getParamPreallocatedType(ArgNo); |
1732 | return nullptr; |
1733 | } |
1734 | |
1735 | /// Extract the preallocated type for a call or parameter. |
1736 | Type *getParamInAllocaType(unsigned ArgNo) const { |
1737 | if (auto *Ty = Attrs.getParamInAllocaType(ArgNo)) |
1738 | return Ty; |
1739 | if (const Function *F = getCalledFunction()) |
1740 | return F->getAttributes().getParamInAllocaType(ArgNo); |
1741 | return nullptr; |
1742 | } |
1743 | |
1744 | /// Extract the number of dereferenceable bytes for a call or |
1745 | /// parameter (0=unknown). |
1746 | uint64_t getDereferenceableBytes(unsigned i) const { |
1747 | return Attrs.getDereferenceableBytes(i); |
1748 | } |
1749 | |
1750 | /// Extract the number of dereferenceable_or_null bytes for a call or |
1751 | /// parameter (0=unknown). |
1752 | uint64_t getDereferenceableOrNullBytes(unsigned i) const { |
1753 | return Attrs.getDereferenceableOrNullBytes(i); |
1754 | } |
1755 | |
1756 | /// Return true if the return value is known to be not null. |
1757 | /// This may be because it has the nonnull attribute, or because at least |
1758 | /// one byte is dereferenceable and the pointer is in addrspace(0). |
1759 | bool isReturnNonNull() const; |
1760 | |
1761 | /// Determine if the return value is marked with NoAlias attribute. |
1762 | bool returnDoesNotAlias() const { |
1763 | return Attrs.hasAttribute(AttributeList::ReturnIndex, Attribute::NoAlias); |
1764 | } |
1765 | |
1766 | /// If one of the arguments has the 'returned' attribute, returns its |
1767 | /// operand value. Otherwise, return nullptr. |
1768 | Value *getReturnedArgOperand() const; |
1769 | |
1770 | /// Return true if the call should not be treated as a call to a |
1771 | /// builtin. |
1772 | bool isNoBuiltin() const { |
1773 | return hasFnAttrImpl(Attribute::NoBuiltin) && |
1774 | !hasFnAttrImpl(Attribute::Builtin); |
1775 | } |
1776 | |
1777 | /// Determine if the call requires strict floating point semantics. |
1778 | bool isStrictFP() const { return hasFnAttr(Attribute::StrictFP); } |
1779 | |
1780 | /// Return true if the call should not be inlined. |
1781 | bool isNoInline() const { return hasFnAttr(Attribute::NoInline); } |
1782 | void setIsNoInline() { |
1783 | addAttribute(AttributeList::FunctionIndex, Attribute::NoInline); |
1784 | } |
1785 | /// Determine if the call does not access memory. |
1786 | bool doesNotAccessMemory() const { return hasFnAttr(Attribute::ReadNone); } |
1787 | void setDoesNotAccessMemory() { |
1788 | addAttribute(AttributeList::FunctionIndex, Attribute::ReadNone); |
1789 | } |
1790 | |
1791 | /// Determine if the call does not access or only reads memory. |
1792 | bool onlyReadsMemory() const { |
1793 | return doesNotAccessMemory() || hasFnAttr(Attribute::ReadOnly); |
1794 | } |
1795 | |
1796 | void setOnlyReadsMemory() { |
1797 | addAttribute(AttributeList::FunctionIndex, Attribute::ReadOnly); |
1798 | } |
1799 | |
1800 | /// Determine if the call does not access or only writes memory. |
1801 | bool doesNotReadMemory() const { |
1802 | return doesNotAccessMemory() || hasFnAttr(Attribute::WriteOnly); |
1803 | } |
1804 | void setDoesNotReadMemory() { |
1805 | addAttribute(AttributeList::FunctionIndex, Attribute::WriteOnly); |
1806 | } |
1807 | |
1808 | /// Determine if the call can access memmory only using pointers based |
1809 | /// on its arguments. |
1810 | bool onlyAccessesArgMemory() const { |
1811 | return hasFnAttr(Attribute::ArgMemOnly); |
1812 | } |
1813 | void setOnlyAccessesArgMemory() { |
1814 | addAttribute(AttributeList::FunctionIndex, Attribute::ArgMemOnly); |
1815 | } |
1816 | |
1817 | /// Determine if the function may only access memory that is |
1818 | /// inaccessible from the IR. |
1819 | bool onlyAccessesInaccessibleMemory() const { |
1820 | return hasFnAttr(Attribute::InaccessibleMemOnly); |
1821 | } |
1822 | void setOnlyAccessesInaccessibleMemory() { |
1823 | addAttribute(AttributeList::FunctionIndex, Attribute::InaccessibleMemOnly); |
1824 | } |
1825 | |
1826 | /// Determine if the function may only access memory that is |
1827 | /// either inaccessible from the IR or pointed to by its arguments. |
1828 | bool onlyAccessesInaccessibleMemOrArgMem() const { |
1829 | return hasFnAttr(Attribute::InaccessibleMemOrArgMemOnly); |
1830 | } |
1831 | void setOnlyAccessesInaccessibleMemOrArgMem() { |
1832 | addAttribute(AttributeList::FunctionIndex, |
1833 | Attribute::InaccessibleMemOrArgMemOnly); |
1834 | } |
1835 | /// Determine if the call cannot return. |
1836 | bool doesNotReturn() const { return hasFnAttr(Attribute::NoReturn); } |
1837 | void setDoesNotReturn() { |
1838 | addAttribute(AttributeList::FunctionIndex, Attribute::NoReturn); |
1839 | } |
1840 | |
1841 | /// Determine if the call should not perform indirect branch tracking. |
1842 | bool doesNoCfCheck() const { return hasFnAttr(Attribute::NoCfCheck); } |
1843 | |
1844 | /// Determine if the call cannot unwind. |
1845 | bool doesNotThrow() const { return hasFnAttr(Attribute::NoUnwind); } |
1846 | void setDoesNotThrow() { |
1847 | addAttribute(AttributeList::FunctionIndex, Attribute::NoUnwind); |
1848 | } |
1849 | |
1850 | /// Determine if the invoke cannot be duplicated. |
1851 | bool cannotDuplicate() const { return hasFnAttr(Attribute::NoDuplicate); } |
1852 | void setCannotDuplicate() { |
1853 | addAttribute(AttributeList::FunctionIndex, Attribute::NoDuplicate); |
1854 | } |
1855 | |
1856 | /// Determine if the call cannot be tail merged. |
1857 | bool cannotMerge() const { return hasFnAttr(Attribute::NoMerge); } |
1858 | void setCannotMerge() { |
1859 | addAttribute(AttributeList::FunctionIndex, Attribute::NoMerge); |
1860 | } |
1861 | |
1862 | /// Determine if the invoke is convergent |
1863 | bool isConvergent() const { return hasFnAttr(Attribute::Convergent); } |
1864 | void setConvergent() { |
1865 | addAttribute(AttributeList::FunctionIndex, Attribute::Convergent); |
1866 | } |
1867 | void setNotConvergent() { |
1868 | removeAttribute(AttributeList::FunctionIndex, Attribute::Convergent); |
1869 | } |
1870 | |
1871 | /// Determine if the call returns a structure through first |
1872 | /// pointer argument. |
1873 | bool hasStructRetAttr() const { |
1874 | if (getNumArgOperands() == 0) |
1875 | return false; |
1876 | |
1877 | // Be friendly and also check the callee. |
1878 | return paramHasAttr(0, Attribute::StructRet); |
1879 | } |
1880 | |
1881 | /// Determine if any call argument is an aggregate passed by value. |
1882 | bool hasByValArgument() const { |
1883 | return Attrs.hasAttrSomewhere(Attribute::ByVal); |
1884 | } |
1885 | |
1886 | ///@{ |
1887 | // End of attribute API. |
1888 | |
1889 | /// \name Operand Bundle API |
1890 | /// |
1891 | /// This group of methods provides the API to access and manipulate operand |
1892 | /// bundles on this call. |
1893 | /// @{ |
1894 | |
1895 | /// Return the number of operand bundles associated with this User. |
1896 | unsigned getNumOperandBundles() const { |
1897 | return std::distance(bundle_op_info_begin(), bundle_op_info_end()); |
1898 | } |
1899 | |
1900 | /// Return true if this User has any operand bundles. |
1901 | bool hasOperandBundles() const { return getNumOperandBundles() != 0; } |
1902 | |
1903 | /// Return the index of the first bundle operand in the Use array. |
1904 | unsigned getBundleOperandsStartIndex() const { |
1905 | assert(hasOperandBundles() && "Don't call otherwise!")((void)0); |
1906 | return bundle_op_info_begin()->Begin; |
1907 | } |
1908 | |
1909 | /// Return the index of the last bundle operand in the Use array. |
1910 | unsigned getBundleOperandsEndIndex() const { |
1911 | assert(hasOperandBundles() && "Don't call otherwise!")((void)0); |
1912 | return bundle_op_info_end()[-1].End; |
1913 | } |
1914 | |
1915 | /// Return true if the operand at index \p Idx is a bundle operand. |
1916 | bool isBundleOperand(unsigned Idx) const { |
1917 | return hasOperandBundles() && Idx >= getBundleOperandsStartIndex() && |
1918 | Idx < getBundleOperandsEndIndex(); |
1919 | } |
1920 | |
1921 | /// Returns true if the use is a bundle operand. |
1922 | bool isBundleOperand(const Use *U) const { |
1923 | assert(this == U->getUser() &&((void)0) |
1924 | "Only valid to query with a use of this instruction!")((void)0); |
1925 | return hasOperandBundles() && isBundleOperand(U - op_begin()); |
1926 | } |
1927 | bool isBundleOperand(Value::const_user_iterator UI) const { |
1928 | return isBundleOperand(&UI.getUse()); |
1929 | } |
1930 | |
1931 | /// Return the total number operands (not operand bundles) used by |
1932 | /// every operand bundle in this OperandBundleUser. |
1933 | unsigned getNumTotalBundleOperands() const { |
1934 | if (!hasOperandBundles()) |
1935 | return 0; |
1936 | |
1937 | unsigned Begin = getBundleOperandsStartIndex(); |
1938 | unsigned End = getBundleOperandsEndIndex(); |
1939 | |
1940 | assert(Begin <= End && "Should be!")((void)0); |
1941 | return End - Begin; |
1942 | } |
1943 | |
1944 | /// Return the operand bundle at a specific index. |
1945 | OperandBundleUse getOperandBundleAt(unsigned Index) const { |
1946 | assert(Index < getNumOperandBundles() && "Index out of bounds!")((void)0); |
1947 | return operandBundleFromBundleOpInfo(*(bundle_op_info_begin() + Index)); |
1948 | } |
1949 | |
1950 | /// Return the number of operand bundles with the tag Name attached to |
1951 | /// this instruction. |
1952 | unsigned countOperandBundlesOfType(StringRef Name) const { |
1953 | unsigned Count = 0; |
1954 | for (unsigned i = 0, e = getNumOperandBundles(); i != e; ++i) |
1955 | if (getOperandBundleAt(i).getTagName() == Name) |
1956 | Count++; |
1957 | |
1958 | return Count; |
1959 | } |
1960 | |
1961 | /// Return the number of operand bundles with the tag ID attached to |
1962 | /// this instruction. |
1963 | unsigned countOperandBundlesOfType(uint32_t ID) const { |
1964 | unsigned Count = 0; |
1965 | for (unsigned i = 0, e = getNumOperandBundles(); i != e; ++i) |
1966 | if (getOperandBundleAt(i).getTagID() == ID) |
1967 | Count++; |
1968 | |
1969 | return Count; |
1970 | } |
1971 | |
1972 | /// Return an operand bundle by name, if present. |
1973 | /// |
1974 | /// It is an error to call this for operand bundle types that may have |
1975 | /// multiple instances of them on the same instruction. |
1976 | Optional<OperandBundleUse> getOperandBundle(StringRef Name) const { |
1977 | assert(countOperandBundlesOfType(Name) < 2 && "Precondition violated!")((void)0); |
1978 | |
1979 | for (unsigned i = 0, e = getNumOperandBundles(); i != e; ++i) { |
1980 | OperandBundleUse U = getOperandBundleAt(i); |
1981 | if (U.getTagName() == Name) |
1982 | return U; |
1983 | } |
1984 | |
1985 | return None; |
1986 | } |
1987 | |
1988 | /// Return an operand bundle by tag ID, if present. |
1989 | /// |
1990 | /// It is an error to call this for operand bundle types that may have |
1991 | /// multiple instances of them on the same instruction. |
1992 | Optional<OperandBundleUse> getOperandBundle(uint32_t ID) const { |
1993 | assert(countOperandBundlesOfType(ID) < 2 && "Precondition violated!")((void)0); |
1994 | |
1995 | for (unsigned i = 0, e = getNumOperandBundles(); i != e; ++i) { |
1996 | OperandBundleUse U = getOperandBundleAt(i); |
1997 | if (U.getTagID() == ID) |
1998 | return U; |
1999 | } |
2000 | |
2001 | return None; |
2002 | } |
2003 | |
2004 | /// Return the list of operand bundles attached to this instruction as |
2005 | /// a vector of OperandBundleDefs. |
2006 | /// |
2007 | /// This function copies the OperandBundeUse instances associated with this |
2008 | /// OperandBundleUser to a vector of OperandBundleDefs. Note: |
2009 | /// OperandBundeUses and OperandBundleDefs are non-trivially *different* |
2010 | /// representations of operand bundles (see documentation above). |
2011 | void getOperandBundlesAsDefs(SmallVectorImpl<OperandBundleDef> &Defs) const; |
2012 | |
2013 | /// Return the operand bundle for the operand at index OpIdx. |
2014 | /// |
2015 | /// It is an error to call this with an OpIdx that does not correspond to an |
2016 | /// bundle operand. |
2017 | OperandBundleUse getOperandBundleForOperand(unsigned OpIdx) const { |
2018 | return operandBundleFromBundleOpInfo(getBundleOpInfoForOperand(OpIdx)); |
2019 | } |
2020 | |
2021 | /// Return true if this operand bundle user has operand bundles that |
2022 | /// may read from the heap. |
2023 | bool hasReadingOperandBundles() const; |
2024 | |
2025 | /// Return true if this operand bundle user has operand bundles that |
2026 | /// may write to the heap. |
2027 | bool hasClobberingOperandBundles() const { |
2028 | for (auto &BOI : bundle_op_infos()) { |
2029 | if (BOI.Tag->second == LLVMContext::OB_deopt || |
2030 | BOI.Tag->second == LLVMContext::OB_funclet) |
2031 | continue; |
2032 | |
2033 | // This instruction has an operand bundle that is not known to us. |
2034 | // Assume the worst. |
2035 | return true; |
2036 | } |
2037 | |
2038 | return false; |
2039 | } |
2040 | |
2041 | /// Return true if the bundle operand at index \p OpIdx has the |
2042 | /// attribute \p A. |
2043 | bool bundleOperandHasAttr(unsigned OpIdx, Attribute::AttrKind A) const { |
2044 | auto &BOI = getBundleOpInfoForOperand(OpIdx); |
2045 | auto OBU = operandBundleFromBundleOpInfo(BOI); |
2046 | return OBU.operandHasAttr(OpIdx - BOI.Begin, A); |
2047 | } |
2048 | |
2049 | /// Return true if \p Other has the same sequence of operand bundle |
2050 | /// tags with the same number of operands on each one of them as this |
2051 | /// OperandBundleUser. |
2052 | bool hasIdenticalOperandBundleSchema(const CallBase &Other) const { |
2053 | if (getNumOperandBundles() != Other.getNumOperandBundles()) |
2054 | return false; |
2055 | |
2056 | return std::equal(bundle_op_info_begin(), bundle_op_info_end(), |
2057 | Other.bundle_op_info_begin()); |
2058 | } |
2059 | |
2060 | /// Return true if this operand bundle user contains operand bundles |
2061 | /// with tags other than those specified in \p IDs. |
2062 | bool hasOperandBundlesOtherThan(ArrayRef<uint32_t> IDs) const { |
2063 | for (unsigned i = 0, e = getNumOperandBundles(); i != e; ++i) { |
2064 | uint32_t ID = getOperandBundleAt(i).getTagID(); |
2065 | if (!is_contained(IDs, ID)) |
2066 | return true; |
2067 | } |
2068 | return false; |
2069 | } |
2070 | |
2071 | /// Is the function attribute S disallowed by some operand bundle on |
2072 | /// this operand bundle user? |
2073 | bool isFnAttrDisallowedByOpBundle(StringRef S) const { |
2074 | // Operand bundles only possibly disallow readnone, readonly and argmemonly |
2075 | // attributes. All String attributes are fine. |
2076 | return false; |
2077 | } |
2078 | |
2079 | /// Is the function attribute A disallowed by some operand bundle on |
2080 | /// this operand bundle user? |
2081 | bool isFnAttrDisallowedByOpBundle(Attribute::AttrKind A) const { |
2082 | switch (A) { |
2083 | default: |
2084 | return false; |
2085 | |
2086 | case Attribute::InaccessibleMemOrArgMemOnly: |
2087 | return hasReadingOperandBundles(); |
2088 | |
2089 | case Attribute::InaccessibleMemOnly: |
2090 | return hasReadingOperandBundles(); |
2091 | |
2092 | case Attribute::ArgMemOnly: |
2093 | return hasReadingOperandBundles(); |
2094 | |
2095 | case Attribute::ReadNone: |
2096 | return hasReadingOperandBundles(); |
2097 | |
2098 | case Attribute::ReadOnly: |
2099 | return hasClobberingOperandBundles(); |
2100 | } |
2101 | |
2102 | llvm_unreachable("switch has a default case!")__builtin_unreachable(); |
2103 | } |
2104 | |
2105 | /// Used to keep track of an operand bundle. See the main comment on |
2106 | /// OperandBundleUser above. |
2107 | struct BundleOpInfo { |
2108 | /// The operand bundle tag, interned by |
2109 | /// LLVMContextImpl::getOrInsertBundleTag. |
2110 | StringMapEntry<uint32_t> *Tag; |
2111 | |
2112 | /// The index in the Use& vector where operands for this operand |
2113 | /// bundle starts. |
2114 | uint32_t Begin; |
2115 | |
2116 | /// The index in the Use& vector where operands for this operand |
2117 | /// bundle ends. |
2118 | uint32_t End; |
2119 | |
2120 | bool operator==(const BundleOpInfo &Other) const { |
2121 | return Tag == Other.Tag && Begin == Other.Begin && End == Other.End; |
2122 | } |
2123 | }; |
2124 | |
2125 | /// Simple helper function to map a BundleOpInfo to an |
2126 | /// OperandBundleUse. |
2127 | OperandBundleUse |
2128 | operandBundleFromBundleOpInfo(const BundleOpInfo &BOI) const { |
2129 | auto begin = op_begin(); |
2130 | ArrayRef<Use> Inputs(begin + BOI.Begin, begin + BOI.End); |
2131 | return OperandBundleUse(BOI.Tag, Inputs); |
2132 | } |
2133 | |
2134 | using bundle_op_iterator = BundleOpInfo *; |
2135 | using const_bundle_op_iterator = const BundleOpInfo *; |
2136 | |
2137 | /// Return the start of the list of BundleOpInfo instances associated |
2138 | /// with this OperandBundleUser. |
2139 | /// |
2140 | /// OperandBundleUser uses the descriptor area co-allocated with the host User |
2141 | /// to store some meta information about which operands are "normal" operands, |
2142 | /// and which ones belong to some operand bundle. |
2143 | /// |
2144 | /// The layout of an operand bundle user is |
2145 | /// |
2146 | /// +-----------uint32_t End-------------------------------------+ |
2147 | /// | | |
2148 | /// | +--------uint32_t Begin--------------------+ | |
2149 | /// | | | | |
2150 | /// ^ ^ v v |
2151 | /// |------|------|----|----|----|----|----|---------|----|---------|----|----- |
2152 | /// | BOI0 | BOI1 | .. | DU | U0 | U1 | .. | BOI0_U0 | .. | BOI1_U0 | .. | Un |
2153 | /// |------|------|----|----|----|----|----|---------|----|---------|----|----- |
2154 | /// v v ^ ^ |
2155 | /// | | | | |
2156 | /// | +--------uint32_t Begin------------+ | |
2157 | /// | | |
2158 | /// +-----------uint32_t End-----------------------------+ |
2159 | /// |
2160 | /// |
2161 | /// BOI0, BOI1 ... are descriptions of operand bundles in this User's use |
2162 | /// list. These descriptions are installed and managed by this class, and |
2163 | /// they're all instances of OperandBundleUser<T>::BundleOpInfo. |
2164 | /// |
2165 | /// DU is an additional descriptor installed by User's 'operator new' to keep |
2166 | /// track of the 'BOI0 ... BOIN' co-allocation. OperandBundleUser does not |
2167 | /// access or modify DU in any way, it's an implementation detail private to |
2168 | /// User. |
2169 | /// |
2170 | /// The regular Use& vector for the User starts at U0. The operand bundle |
2171 | /// uses are part of the Use& vector, just like normal uses. In the diagram |
2172 | /// above, the operand bundle uses start at BOI0_U0. Each instance of |
2173 | /// BundleOpInfo has information about a contiguous set of uses constituting |
2174 | /// an operand bundle, and the total set of operand bundle uses themselves |
2175 | /// form a contiguous set of uses (i.e. there are no gaps between uses |
2176 | /// corresponding to individual operand bundles). |
2177 | /// |
2178 | /// This class does not know the location of the set of operand bundle uses |
2179 | /// within the use list -- that is decided by the User using this class via |
2180 | /// the BeginIdx argument in populateBundleOperandInfos. |
2181 | /// |
2182 | /// Currently operand bundle users with hung-off operands are not supported. |
2183 | bundle_op_iterator bundle_op_info_begin() { |
2184 | if (!hasDescriptor()) |
2185 | return nullptr; |
2186 | |
2187 | uint8_t *BytesBegin = getDescriptor().begin(); |
2188 | return reinterpret_cast<bundle_op_iterator>(BytesBegin); |
2189 | } |
2190 | |
2191 | /// Return the start of the list of BundleOpInfo instances associated |
2192 | /// with this OperandBundleUser. |
2193 | const_bundle_op_iterator bundle_op_info_begin() const { |
2194 | auto *NonConstThis = const_cast<CallBase *>(this); |
2195 | return NonConstThis->bundle_op_info_begin(); |
2196 | } |
2197 | |
2198 | /// Return the end of the list of BundleOpInfo instances associated |
2199 | /// with this OperandBundleUser. |
2200 | bundle_op_iterator bundle_op_info_end() { |
2201 | if (!hasDescriptor()) |
2202 | return nullptr; |
2203 | |
2204 | uint8_t *BytesEnd = getDescriptor().end(); |
2205 | return reinterpret_cast<bundle_op_iterator>(BytesEnd); |
2206 | } |
2207 | |
2208 | /// Return the end of the list of BundleOpInfo instances associated |
2209 | /// with this OperandBundleUser. |
2210 | const_bundle_op_iterator bundle_op_info_end() const { |
2211 | auto *NonConstThis = const_cast<CallBase *>(this); |
2212 | return NonConstThis->bundle_op_info_end(); |
2213 | } |
2214 | |
2215 | /// Return the range [\p bundle_op_info_begin, \p bundle_op_info_end). |
2216 | iterator_range<bundle_op_iterator> bundle_op_infos() { |
2217 | return make_range(bundle_op_info_begin(), bundle_op_info_end()); |
2218 | } |
2219 | |
2220 | /// Return the range [\p bundle_op_info_begin, \p bundle_op_info_end). |
2221 | iterator_range<const_bundle_op_iterator> bundle_op_infos() const { |
2222 | return make_range(bundle_op_info_begin(), bundle_op_info_end()); |
2223 | } |
2224 | |
2225 | /// Populate the BundleOpInfo instances and the Use& vector from \p |
2226 | /// Bundles. Return the op_iterator pointing to the Use& one past the last |
2227 | /// last bundle operand use. |
2228 | /// |
2229 | /// Each \p OperandBundleDef instance is tracked by a OperandBundleInfo |
2230 | /// instance allocated in this User's descriptor. |
2231 | op_iterator populateBundleOperandInfos(ArrayRef<OperandBundleDef> Bundles, |
2232 | const unsigned BeginIndex); |
2233 | |
2234 | public: |
2235 | /// Return the BundleOpInfo for the operand at index OpIdx. |
2236 | /// |
2237 | /// It is an error to call this with an OpIdx that does not correspond to an |
2238 | /// bundle operand. |
2239 | BundleOpInfo &getBundleOpInfoForOperand(unsigned OpIdx); |
2240 | const BundleOpInfo &getBundleOpInfoForOperand(unsigned OpIdx) const { |
2241 | return const_cast<CallBase *>(this)->getBundleOpInfoForOperand(OpIdx); |
2242 | } |
2243 | |
2244 | protected: |
2245 | /// Return the total number of values used in \p Bundles. |
2246 | static unsigned CountBundleInputs(ArrayRef<OperandBundleDef> Bundles) { |
2247 | unsigned Total = 0; |
2248 | for (auto &B : Bundles) |
2249 | Total += B.input_size(); |
2250 | return Total; |
2251 | } |
2252 | |
2253 | /// @} |
2254 | // End of operand bundle API. |
2255 | |
2256 | private: |
2257 | bool hasFnAttrOnCalledFunction(Attribute::AttrKind Kind) const; |
2258 | bool hasFnAttrOnCalledFunction(StringRef Kind) const; |
2259 | |
2260 | template <typename AttrKind> bool hasFnAttrImpl(AttrKind Kind) const { |
2261 | if (Attrs.hasFnAttribute(Kind)) |
2262 | return true; |
2263 | |
2264 | // Operand bundles override attributes on the called function, but don't |
2265 | // override attributes directly present on the call instruction. |
2266 | if (isFnAttrDisallowedByOpBundle(Kind)) |
2267 | return false; |
2268 | |
2269 | return hasFnAttrOnCalledFunction(Kind); |
2270 | } |
2271 | |
2272 | /// Determine whether the return value has the given attribute. Supports |
2273 | /// Attribute::AttrKind and StringRef as \p AttrKind types. |
2274 | template <typename AttrKind> bool hasRetAttrImpl(AttrKind Kind) const { |
2275 | if (Attrs.hasAttribute(AttributeList::ReturnIndex, Kind)) |
2276 | return true; |
2277 | |
2278 | // Look at the callee, if available. |
2279 | if (const Function *F = getCalledFunction()) |
2280 | return F->getAttributes().hasAttribute(AttributeList::ReturnIndex, Kind); |
2281 | return false; |
2282 | } |
2283 | }; |
2284 | |
2285 | template <> |
2286 | struct OperandTraits<CallBase> : public VariadicOperandTraits<CallBase, 1> {}; |
2287 | |
2288 | DEFINE_TRANSPARENT_OPERAND_ACCESSORS(CallBase, Value)CallBase::op_iterator CallBase::op_begin() { return OperandTraits <CallBase>::op_begin(this); } CallBase::const_op_iterator CallBase::op_begin() const { return OperandTraits<CallBase >::op_begin(const_cast<CallBase*>(this)); } CallBase ::op_iterator CallBase::op_end() { return OperandTraits<CallBase >::op_end(this); } CallBase::const_op_iterator CallBase::op_end () const { return OperandTraits<CallBase>::op_end(const_cast <CallBase*>(this)); } Value *CallBase::getOperand(unsigned i_nocapture) const { ((void)0); return cast_or_null<Value >( OperandTraits<CallBase>::op_begin(const_cast<CallBase *>(this))[i_nocapture].get()); } void CallBase::setOperand (unsigned i_nocapture, Value *Val_nocapture) { ((void)0); OperandTraits <CallBase>::op_begin(this)[i_nocapture] = Val_nocapture ; } unsigned CallBase::getNumOperands() const { return OperandTraits <CallBase>::operands(this); } template <int Idx_nocapture > Use &CallBase::Op() { return this->OpFrom<Idx_nocapture >(this); } template <int Idx_nocapture> const Use & CallBase::Op() const { return this->OpFrom<Idx_nocapture >(this); } |
2289 | |
2290 | //===----------------------------------------------------------------------===// |
2291 | // FuncletPadInst Class |
2292 | //===----------------------------------------------------------------------===// |
2293 | class FuncletPadInst : public Instruction { |
2294 | private: |
2295 | FuncletPadInst(const FuncletPadInst &CPI); |
2296 | |
2297 | explicit FuncletPadInst(Instruction::FuncletPadOps Op, Value *ParentPad, |
2298 | ArrayRef<Value *> Args, unsigned Values, |
2299 | const Twine &NameStr, Instruction *InsertBefore); |
2300 | explicit FuncletPadInst(Instruction::FuncletPadOps Op, Value *ParentPad, |
2301 | ArrayRef<Value *> Args, unsigned Values, |
2302 | const Twine &NameStr, BasicBlock *InsertAtEnd); |
2303 | |
2304 | void init(Value *ParentPad, ArrayRef<Value *> Args, const Twine &NameStr); |
2305 | |
2306 | protected: |
2307 | // Note: Instruction needs to be a friend here to call cloneImpl. |
2308 | friend class Instruction; |
2309 | friend class CatchPadInst; |
2310 | friend class CleanupPadInst; |
2311 | |
2312 | FuncletPadInst *cloneImpl() const; |
2313 | |
2314 | public: |
2315 | /// Provide fast operand accessors |
2316 | DECLARE_TRANSPARENT_OPERAND_ACCESSORS(Value)public: inline Value *getOperand(unsigned) const; inline void setOperand(unsigned, Value*); inline op_iterator op_begin(); inline const_op_iterator op_begin() const; inline op_iterator op_end(); inline const_op_iterator op_end() const; protected : template <int> inline Use &Op(); template <int > inline const Use &Op() const; public: inline unsigned getNumOperands() const; |
2317 | |
2318 | /// getNumArgOperands - Return the number of funcletpad arguments. |
2319 | /// |
2320 | unsigned getNumArgOperands() const { return getNumOperands() - 1; } |
2321 | |
2322 | /// Convenience accessors |
2323 | |
2324 | /// Return the outer EH-pad this funclet is nested within. |
2325 | /// |
2326 | /// Note: This returns the associated CatchSwitchInst if this FuncletPadInst |
2327 | /// is a CatchPadInst. |
2328 | Value *getParentPad() const { return Op<-1>(); } |
2329 | void setParentPad(Value *ParentPad) { |
2330 | assert(ParentPad)((void)0); |
2331 | Op<-1>() = ParentPad; |
2332 | } |
2333 | |
2334 | /// getArgOperand/setArgOperand - Return/set the i-th funcletpad argument. |
2335 | /// |
2336 | Value *getArgOperand(unsigned i) const { return getOperand(i); } |
2337 | void setArgOperand(unsigned i, Value *v) { setOperand(i, v); } |
2338 | |
2339 | /// arg_operands - iteration adapter for range-for loops. |
2340 | op_range arg_operands() { return op_range(op_begin(), op_end() - 1); } |
2341 | |
2342 | /// arg_operands - iteration adapter for range-for loops. |
2343 | const_op_range arg_operands() const { |
2344 | return const_op_range(op_begin(), op_end() - 1); |
2345 | } |
2346 | |
2347 | // Methods for support type inquiry through isa, cast, and dyn_cast: |
2348 | static bool classof(const Instruction *I) { return I->isFuncletPad(); } |
2349 | static bool classof(const Value *V) { |
2350 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
2351 | } |
2352 | }; |
2353 | |
2354 | template <> |
2355 | struct OperandTraits<FuncletPadInst> |
2356 | : public VariadicOperandTraits<FuncletPadInst, /*MINARITY=*/1> {}; |
2357 | |
2358 | DEFINE_TRANSPARENT_OPERAND_ACCESSORS(FuncletPadInst, Value)FuncletPadInst::op_iterator FuncletPadInst::op_begin() { return OperandTraits<FuncletPadInst>::op_begin(this); } FuncletPadInst ::const_op_iterator FuncletPadInst::op_begin() const { return OperandTraits<FuncletPadInst>::op_begin(const_cast< FuncletPadInst*>(this)); } FuncletPadInst::op_iterator FuncletPadInst ::op_end() { return OperandTraits<FuncletPadInst>::op_end (this); } FuncletPadInst::const_op_iterator FuncletPadInst::op_end () const { return OperandTraits<FuncletPadInst>::op_end (const_cast<FuncletPadInst*>(this)); } Value *FuncletPadInst ::getOperand(unsigned i_nocapture) const { ((void)0); return cast_or_null <Value>( OperandTraits<FuncletPadInst>::op_begin( const_cast<FuncletPadInst*>(this))[i_nocapture].get()); } void FuncletPadInst::setOperand(unsigned i_nocapture, Value *Val_nocapture) { ((void)0); OperandTraits<FuncletPadInst >::op_begin(this)[i_nocapture] = Val_nocapture; } unsigned FuncletPadInst::getNumOperands() const { return OperandTraits <FuncletPadInst>::operands(this); } template <int Idx_nocapture > Use &FuncletPadInst::Op() { return this->OpFrom< Idx_nocapture>(this); } template <int Idx_nocapture> const Use &FuncletPadInst::Op() const { return this-> OpFrom<Idx_nocapture>(this); } |
2359 | |
2360 | } // end namespace llvm |
2361 | |
2362 | #endif // LLVM_IR_INSTRTYPES_H |
1 | //===- PatternMatch.h - Match on the LLVM IR --------------------*- C++ -*-===// | ||||||||||||||||
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 file provides a simple and efficient mechanism for performing general | ||||||||||||||||
10 | // tree-based pattern matches on the LLVM IR. The power of these routines is | ||||||||||||||||
11 | // that it allows you to write concise patterns that are expressive and easy to | ||||||||||||||||
12 | // understand. The other major advantage of this is that it allows you to | ||||||||||||||||
13 | // trivially capture/bind elements in the pattern to variables. For example, | ||||||||||||||||
14 | // you can do something like this: | ||||||||||||||||
15 | // | ||||||||||||||||
16 | // Value *Exp = ... | ||||||||||||||||
17 | // Value *X, *Y; ConstantInt *C1, *C2; // (X & C1) | (Y & C2) | ||||||||||||||||
18 | // if (match(Exp, m_Or(m_And(m_Value(X), m_ConstantInt(C1)), | ||||||||||||||||
19 | // m_And(m_Value(Y), m_ConstantInt(C2))))) { | ||||||||||||||||
20 | // ... Pattern is matched and variables are bound ... | ||||||||||||||||
21 | // } | ||||||||||||||||
22 | // | ||||||||||||||||
23 | // This is primarily useful to things like the instruction combiner, but can | ||||||||||||||||
24 | // also be useful for static analysis tools or code generators. | ||||||||||||||||
25 | // | ||||||||||||||||
26 | //===----------------------------------------------------------------------===// | ||||||||||||||||
27 | |||||||||||||||||
28 | #ifndef LLVM_IR_PATTERNMATCH_H | ||||||||||||||||
29 | #define LLVM_IR_PATTERNMATCH_H | ||||||||||||||||
30 | |||||||||||||||||
31 | #include "llvm/ADT/APFloat.h" | ||||||||||||||||
32 | #include "llvm/ADT/APInt.h" | ||||||||||||||||
33 | #include "llvm/IR/Constant.h" | ||||||||||||||||
34 | #include "llvm/IR/Constants.h" | ||||||||||||||||
35 | #include "llvm/IR/DataLayout.h" | ||||||||||||||||
36 | #include "llvm/IR/InstrTypes.h" | ||||||||||||||||
37 | #include "llvm/IR/Instruction.h" | ||||||||||||||||
38 | #include "llvm/IR/Instructions.h" | ||||||||||||||||
39 | #include "llvm/IR/IntrinsicInst.h" | ||||||||||||||||
40 | #include "llvm/IR/Intrinsics.h" | ||||||||||||||||
41 | #include "llvm/IR/Operator.h" | ||||||||||||||||
42 | #include "llvm/IR/Value.h" | ||||||||||||||||
43 | #include "llvm/Support/Casting.h" | ||||||||||||||||
44 | #include <cstdint> | ||||||||||||||||
45 | |||||||||||||||||
46 | namespace llvm { | ||||||||||||||||
47 | namespace PatternMatch { | ||||||||||||||||
48 | |||||||||||||||||
49 | template <typename Val, typename Pattern> bool match(Val *V, const Pattern &P) { | ||||||||||||||||
50 | return const_cast<Pattern &>(P).match(V); | ||||||||||||||||
51 | } | ||||||||||||||||
52 | |||||||||||||||||
53 | template <typename Pattern> bool match(ArrayRef<int> Mask, const Pattern &P) { | ||||||||||||||||
54 | return const_cast<Pattern &>(P).match(Mask); | ||||||||||||||||
55 | } | ||||||||||||||||
56 | |||||||||||||||||
57 | template <typename SubPattern_t> struct OneUse_match { | ||||||||||||||||
58 | SubPattern_t SubPattern; | ||||||||||||||||
59 | |||||||||||||||||
60 | OneUse_match(const SubPattern_t &SP) : SubPattern(SP) {} | ||||||||||||||||
61 | |||||||||||||||||
62 | template <typename OpTy> bool match(OpTy *V) { | ||||||||||||||||
63 | return V->hasOneUse() && SubPattern.match(V); | ||||||||||||||||
64 | } | ||||||||||||||||
65 | }; | ||||||||||||||||
66 | |||||||||||||||||
67 | template <typename T> inline OneUse_match<T> m_OneUse(const T &SubPattern) { | ||||||||||||||||
68 | return SubPattern; | ||||||||||||||||
69 | } | ||||||||||||||||
70 | |||||||||||||||||
71 | template <typename Class> struct class_match { | ||||||||||||||||
72 | template <typename ITy> bool match(ITy *V) { return isa<Class>(V); } | ||||||||||||||||
73 | }; | ||||||||||||||||
74 | |||||||||||||||||
75 | /// Match an arbitrary value and ignore it. | ||||||||||||||||
76 | inline class_match<Value> m_Value() { return class_match<Value>(); } | ||||||||||||||||
77 | |||||||||||||||||
78 | /// Match an arbitrary unary operation and ignore it. | ||||||||||||||||
79 | inline class_match<UnaryOperator> m_UnOp() { | ||||||||||||||||
80 | return class_match<UnaryOperator>(); | ||||||||||||||||
81 | } | ||||||||||||||||
82 | |||||||||||||||||
83 | /// Match an arbitrary binary operation and ignore it. | ||||||||||||||||
84 | inline class_match<BinaryOperator> m_BinOp() { | ||||||||||||||||
85 | return class_match<BinaryOperator>(); | ||||||||||||||||
86 | } | ||||||||||||||||
87 | |||||||||||||||||
88 | /// Matches any compare instruction and ignore it. | ||||||||||||||||
89 | inline class_match<CmpInst> m_Cmp() { return class_match<CmpInst>(); } | ||||||||||||||||
90 | |||||||||||||||||
91 | struct undef_match { | ||||||||||||||||
92 | static bool check(const Value *V) { | ||||||||||||||||
93 | if (isa<UndefValue>(V)) | ||||||||||||||||
94 | return true; | ||||||||||||||||
95 | |||||||||||||||||
96 | const auto *CA = dyn_cast<ConstantAggregate>(V); | ||||||||||||||||
97 | if (!CA) | ||||||||||||||||
98 | return false; | ||||||||||||||||
99 | |||||||||||||||||
100 | SmallPtrSet<const ConstantAggregate *, 8> Seen; | ||||||||||||||||
101 | SmallVector<const ConstantAggregate *, 8> Worklist; | ||||||||||||||||
102 | |||||||||||||||||
103 | // Either UndefValue, PoisonValue, or an aggregate that only contains | ||||||||||||||||
104 | // these is accepted by matcher. | ||||||||||||||||
105 | // CheckValue returns false if CA cannot satisfy this constraint. | ||||||||||||||||
106 | auto CheckValue = [&](const ConstantAggregate *CA) { | ||||||||||||||||
107 | for (const Value *Op : CA->operand_values()) { | ||||||||||||||||
108 | if (isa<UndefValue>(Op)) | ||||||||||||||||
109 | continue; | ||||||||||||||||
110 | |||||||||||||||||
111 | const auto *CA = dyn_cast<ConstantAggregate>(Op); | ||||||||||||||||
112 | if (!CA) | ||||||||||||||||
113 | return false; | ||||||||||||||||
114 | if (Seen.insert(CA).second) | ||||||||||||||||
115 | Worklist.emplace_back(CA); | ||||||||||||||||
116 | } | ||||||||||||||||
117 | |||||||||||||||||
118 | return true; | ||||||||||||||||
119 | }; | ||||||||||||||||
120 | |||||||||||||||||
121 | if (!CheckValue(CA)) | ||||||||||||||||
122 | return false; | ||||||||||||||||
123 | |||||||||||||||||
124 | while (!Worklist.empty()) { | ||||||||||||||||
125 | if (!CheckValue(Worklist.pop_back_val())) | ||||||||||||||||
126 | return false; | ||||||||||||||||
127 | } | ||||||||||||||||
128 | return true; | ||||||||||||||||
129 | } | ||||||||||||||||
130 | template <typename ITy> bool match(ITy *V) { return check(V); } | ||||||||||||||||
131 | }; | ||||||||||||||||
132 | |||||||||||||||||
133 | /// Match an arbitrary undef constant. This matches poison as well. | ||||||||||||||||
134 | /// If this is an aggregate and contains a non-aggregate element that is | ||||||||||||||||
135 | /// neither undef nor poison, the aggregate is not matched. | ||||||||||||||||
136 | inline auto m_Undef() { return undef_match(); } | ||||||||||||||||
137 | |||||||||||||||||
138 | /// Match an arbitrary poison constant. | ||||||||||||||||
139 | inline class_match<PoisonValue> m_Poison() { return class_match<PoisonValue>(); } | ||||||||||||||||
140 | |||||||||||||||||
141 | /// Match an arbitrary Constant and ignore it. | ||||||||||||||||
142 | inline class_match<Constant> m_Constant() { return class_match<Constant>(); } | ||||||||||||||||
143 | |||||||||||||||||
144 | /// Match an arbitrary ConstantInt and ignore it. | ||||||||||||||||
145 | inline class_match<ConstantInt> m_ConstantInt() { | ||||||||||||||||
146 | return class_match<ConstantInt>(); | ||||||||||||||||
147 | } | ||||||||||||||||
148 | |||||||||||||||||
149 | /// Match an arbitrary ConstantFP and ignore it. | ||||||||||||||||
150 | inline class_match<ConstantFP> m_ConstantFP() { | ||||||||||||||||
151 | return class_match<ConstantFP>(); | ||||||||||||||||
152 | } | ||||||||||||||||
153 | |||||||||||||||||
154 | /// Match an arbitrary ConstantExpr and ignore it. | ||||||||||||||||
155 | inline class_match<ConstantExpr> m_ConstantExpr() { | ||||||||||||||||
156 | return class_match<ConstantExpr>(); | ||||||||||||||||
157 | } | ||||||||||||||||
158 | |||||||||||||||||
159 | /// Match an arbitrary basic block value and ignore it. | ||||||||||||||||
160 | inline class_match<BasicBlock> m_BasicBlock() { | ||||||||||||||||
161 | return class_match<BasicBlock>(); | ||||||||||||||||
162 | } | ||||||||||||||||
163 | |||||||||||||||||
164 | /// Inverting matcher | ||||||||||||||||
165 | template <typename Ty> struct match_unless { | ||||||||||||||||
166 | Ty M; | ||||||||||||||||
167 | |||||||||||||||||
168 | match_unless(const Ty &Matcher) : M(Matcher) {} | ||||||||||||||||
169 | |||||||||||||||||
170 | template <typename ITy> bool match(ITy *V) { return !M.match(V); } | ||||||||||||||||
171 | }; | ||||||||||||||||
172 | |||||||||||||||||
173 | /// Match if the inner matcher does *NOT* match. | ||||||||||||||||
174 | template <typename Ty> inline match_unless<Ty> m_Unless(const Ty &M) { | ||||||||||||||||
175 | return match_unless<Ty>(M); | ||||||||||||||||
176 | } | ||||||||||||||||
177 | |||||||||||||||||
178 | /// Matching combinators | ||||||||||||||||
179 | template <typename LTy, typename RTy> struct match_combine_or { | ||||||||||||||||
180 | LTy L; | ||||||||||||||||
181 | RTy R; | ||||||||||||||||
182 | |||||||||||||||||
183 | match_combine_or(const LTy &Left, const RTy &Right) : L(Left), R(Right) {} | ||||||||||||||||
184 | |||||||||||||||||
185 | template <typename ITy> bool match(ITy *V) { | ||||||||||||||||
186 | if (L.match(V)) | ||||||||||||||||
187 | return true; | ||||||||||||||||
188 | if (R.match(V)) | ||||||||||||||||
189 | return true; | ||||||||||||||||
190 | return false; | ||||||||||||||||
191 | } | ||||||||||||||||
192 | }; | ||||||||||||||||
193 | |||||||||||||||||
194 | template <typename LTy, typename RTy> struct match_combine_and { | ||||||||||||||||
195 | LTy L; | ||||||||||||||||
196 | RTy R; | ||||||||||||||||
197 | |||||||||||||||||
198 | match_combine_and(const LTy &Left, const RTy &Right) : L(Left), R(Right) {} | ||||||||||||||||
199 | |||||||||||||||||
200 | template <typename ITy> bool match(ITy *V) { | ||||||||||||||||
201 | if (L.match(V)) | ||||||||||||||||
202 | if (R.match(V)) | ||||||||||||||||
203 | return true; | ||||||||||||||||
204 | return false; | ||||||||||||||||
205 | } | ||||||||||||||||
206 | }; | ||||||||||||||||
207 | |||||||||||||||||
208 | /// Combine two pattern matchers matching L || R | ||||||||||||||||
209 | template <typename LTy, typename RTy> | ||||||||||||||||
210 | inline match_combine_or<LTy, RTy> m_CombineOr(const LTy &L, const RTy &R) { | ||||||||||||||||
211 | return match_combine_or<LTy, RTy>(L, R); | ||||||||||||||||
212 | } | ||||||||||||||||
213 | |||||||||||||||||
214 | /// Combine two pattern matchers matching L && R | ||||||||||||||||
215 | template <typename LTy, typename RTy> | ||||||||||||||||
216 | inline match_combine_and<LTy, RTy> m_CombineAnd(const LTy &L, const RTy &R) { | ||||||||||||||||
217 | return match_combine_and<LTy, RTy>(L, R); | ||||||||||||||||
218 | } | ||||||||||||||||
219 | |||||||||||||||||
220 | struct apint_match { | ||||||||||||||||
221 | const APInt *&Res; | ||||||||||||||||
222 | bool AllowUndef; | ||||||||||||||||
223 | |||||||||||||||||
224 | apint_match(const APInt *&Res, bool AllowUndef) | ||||||||||||||||
225 | : Res(Res), AllowUndef(AllowUndef) {} | ||||||||||||||||
226 | |||||||||||||||||
227 | template <typename ITy> bool match(ITy *V) { | ||||||||||||||||
228 | if (auto *CI = dyn_cast<ConstantInt>(V)) { | ||||||||||||||||
229 | Res = &CI->getValue(); | ||||||||||||||||
230 | return true; | ||||||||||||||||
231 | } | ||||||||||||||||
232 | if (V->getType()->isVectorTy()) | ||||||||||||||||
233 | if (const auto *C = dyn_cast<Constant>(V)) | ||||||||||||||||
234 | if (auto *CI = dyn_cast_or_null<ConstantInt>( | ||||||||||||||||
235 | C->getSplatValue(AllowUndef))) { | ||||||||||||||||
236 | Res = &CI->getValue(); | ||||||||||||||||
237 | return true; | ||||||||||||||||
238 | } | ||||||||||||||||
239 | return false; | ||||||||||||||||
240 | } | ||||||||||||||||
241 | }; | ||||||||||||||||
242 | // Either constexpr if or renaming ConstantFP::getValueAPF to | ||||||||||||||||
243 | // ConstantFP::getValue is needed to do it via single template | ||||||||||||||||
244 | // function for both apint/apfloat. | ||||||||||||||||
245 | struct apfloat_match { | ||||||||||||||||
246 | const APFloat *&Res; | ||||||||||||||||
247 | bool AllowUndef; | ||||||||||||||||
248 | |||||||||||||||||
249 | apfloat_match(const APFloat *&Res, bool AllowUndef) | ||||||||||||||||
250 | : Res(Res), AllowUndef(AllowUndef) {} | ||||||||||||||||
251 | |||||||||||||||||
252 | template <typename ITy> bool match(ITy *V) { | ||||||||||||||||
253 | if (auto *CI = dyn_cast<ConstantFP>(V)) { | ||||||||||||||||
254 | Res = &CI->getValueAPF(); | ||||||||||||||||
255 | return true; | ||||||||||||||||
256 | } | ||||||||||||||||
257 | if (V->getType()->isVectorTy()) | ||||||||||||||||
258 | if (const auto *C = dyn_cast<Constant>(V)) | ||||||||||||||||
259 | if (auto *CI = dyn_cast_or_null<ConstantFP>( | ||||||||||||||||
260 | C->getSplatValue(AllowUndef))) { | ||||||||||||||||
261 | Res = &CI->getValueAPF(); | ||||||||||||||||
262 | return true; | ||||||||||||||||
263 | } | ||||||||||||||||
264 | return false; | ||||||||||||||||
265 | } | ||||||||||||||||
266 | }; | ||||||||||||||||
267 | |||||||||||||||||
268 | /// Match a ConstantInt or splatted ConstantVector, binding the | ||||||||||||||||
269 | /// specified pointer to the contained APInt. | ||||||||||||||||
270 | inline apint_match m_APInt(const APInt *&Res) { | ||||||||||||||||
271 | // Forbid undefs by default to maintain previous behavior. | ||||||||||||||||
272 | return apint_match(Res, /* AllowUndef */ false); | ||||||||||||||||
273 | } | ||||||||||||||||
274 | |||||||||||||||||
275 | /// Match APInt while allowing undefs in splat vector constants. | ||||||||||||||||
276 | inline apint_match m_APIntAllowUndef(const APInt *&Res) { | ||||||||||||||||
277 | return apint_match(Res, /* AllowUndef */ true); | ||||||||||||||||
278 | } | ||||||||||||||||
279 | |||||||||||||||||
280 | /// Match APInt while forbidding undefs in splat vector constants. | ||||||||||||||||
281 | inline apint_match m_APIntForbidUndef(const APInt *&Res) { | ||||||||||||||||
282 | return apint_match(Res, /* AllowUndef */ false); | ||||||||||||||||
283 | } | ||||||||||||||||
284 | |||||||||||||||||
285 | /// Match a ConstantFP or splatted ConstantVector, binding the | ||||||||||||||||
286 | /// specified pointer to the contained APFloat. | ||||||||||||||||
287 | inline apfloat_match m_APFloat(const APFloat *&Res) { | ||||||||||||||||
288 | // Forbid undefs by default to maintain previous behavior. | ||||||||||||||||
289 | return apfloat_match(Res, /* AllowUndef */ false); | ||||||||||||||||
290 | } | ||||||||||||||||
291 | |||||||||||||||||
292 | /// Match APFloat while allowing undefs in splat vector constants. | ||||||||||||||||
293 | inline apfloat_match m_APFloatAllowUndef(const APFloat *&Res) { | ||||||||||||||||
294 | return apfloat_match(Res, /* AllowUndef */ true); | ||||||||||||||||
295 | } | ||||||||||||||||
296 | |||||||||||||||||
297 | /// Match APFloat while forbidding undefs in splat vector constants. | ||||||||||||||||
298 | inline apfloat_match m_APFloatForbidUndef(const APFloat *&Res) { | ||||||||||||||||
299 | return apfloat_match(Res, /* AllowUndef */ false); | ||||||||||||||||
300 | } | ||||||||||||||||
301 | |||||||||||||||||
302 | template <int64_t Val> struct constantint_match { | ||||||||||||||||
303 | template <typename ITy> bool match(ITy *V) { | ||||||||||||||||
304 | if (const auto *CI = dyn_cast<ConstantInt>(V)) { | ||||||||||||||||
305 | const APInt &CIV = CI->getValue(); | ||||||||||||||||
306 | if (Val >= 0) | ||||||||||||||||
307 | return CIV == static_cast<uint64_t>(Val); | ||||||||||||||||
308 | // If Val is negative, and CI is shorter than it, truncate to the right | ||||||||||||||||
309 | // number of bits. If it is larger, then we have to sign extend. Just | ||||||||||||||||
310 | // compare their negated values. | ||||||||||||||||
311 | return -CIV == -Val; | ||||||||||||||||
312 | } | ||||||||||||||||
313 | return false; | ||||||||||||||||
314 | } | ||||||||||||||||
315 | }; | ||||||||||||||||
316 | |||||||||||||||||
317 | /// Match a ConstantInt with a specific value. | ||||||||||||||||
318 | template <int64_t Val> inline constantint_match<Val> m_ConstantInt() { | ||||||||||||||||
319 | return constantint_match<Val>(); | ||||||||||||||||
320 | } | ||||||||||||||||
321 | |||||||||||||||||
322 | /// This helper class is used to match constant scalars, vector splats, | ||||||||||||||||
323 | /// and fixed width vectors that satisfy a specified predicate. | ||||||||||||||||
324 | /// For fixed width vector constants, undefined elements are ignored. | ||||||||||||||||
325 | template <typename Predicate, typename ConstantVal> | ||||||||||||||||
326 | struct cstval_pred_ty : public Predicate { | ||||||||||||||||
327 | template <typename ITy> bool match(ITy *V) { | ||||||||||||||||
328 | if (const auto *CV = dyn_cast<ConstantVal>(V)) | ||||||||||||||||
329 | return this->isValue(CV->getValue()); | ||||||||||||||||
330 | if (const auto *VTy = dyn_cast<VectorType>(V->getType())) { | ||||||||||||||||
331 | if (const auto *C = dyn_cast<Constant>(V)) { | ||||||||||||||||
332 | if (const auto *CV = dyn_cast_or_null<ConstantVal>(C->getSplatValue())) | ||||||||||||||||
333 | return this->isValue(CV->getValue()); | ||||||||||||||||
334 | |||||||||||||||||
335 | // Number of elements of a scalable vector unknown at compile time | ||||||||||||||||
336 | auto *FVTy = dyn_cast<FixedVectorType>(VTy); | ||||||||||||||||
337 | if (!FVTy) | ||||||||||||||||
338 | return false; | ||||||||||||||||
339 | |||||||||||||||||
340 | // Non-splat vector constant: check each element for a match. | ||||||||||||||||
341 | unsigned NumElts = FVTy->getNumElements(); | ||||||||||||||||
342 | assert(NumElts != 0 && "Constant vector with no elements?")((void)0); | ||||||||||||||||
343 | bool HasNonUndefElements = false; | ||||||||||||||||
344 | for (unsigned i = 0; i != NumElts; ++i) { | ||||||||||||||||
345 | Constant *Elt = C->getAggregateElement(i); | ||||||||||||||||
346 | if (!Elt) | ||||||||||||||||
347 | return false; | ||||||||||||||||
348 | if (isa<UndefValue>(Elt)) | ||||||||||||||||
349 | continue; | ||||||||||||||||
350 | auto *CV = dyn_cast<ConstantVal>(Elt); | ||||||||||||||||
351 | if (!CV || !this->isValue(CV->getValue())) | ||||||||||||||||
352 | return false; | ||||||||||||||||
353 | HasNonUndefElements = true; | ||||||||||||||||
354 | } | ||||||||||||||||
355 | return HasNonUndefElements; | ||||||||||||||||
356 | } | ||||||||||||||||
357 | } | ||||||||||||||||
358 | return false; | ||||||||||||||||
359 | } | ||||||||||||||||
360 | }; | ||||||||||||||||
361 | |||||||||||||||||
362 | /// specialization of cstval_pred_ty for ConstantInt | ||||||||||||||||
363 | template <typename Predicate> | ||||||||||||||||
364 | using cst_pred_ty = cstval_pred_ty<Predicate, ConstantInt>; | ||||||||||||||||
365 | |||||||||||||||||
366 | /// specialization of cstval_pred_ty for ConstantFP | ||||||||||||||||
367 | template <typename Predicate> | ||||||||||||||||
368 | using cstfp_pred_ty = cstval_pred_ty<Predicate, ConstantFP>; | ||||||||||||||||
369 | |||||||||||||||||
370 | /// This helper class is used to match scalar and vector constants that | ||||||||||||||||
371 | /// satisfy a specified predicate, and bind them to an APInt. | ||||||||||||||||
372 | template <typename Predicate> struct api_pred_ty : public Predicate { | ||||||||||||||||
373 | const APInt *&Res; | ||||||||||||||||
374 | |||||||||||||||||
375 | api_pred_ty(const APInt *&R) : Res(R) {} | ||||||||||||||||
376 | |||||||||||||||||
377 | template <typename ITy> bool match(ITy *V) { | ||||||||||||||||
378 | if (const auto *CI = dyn_cast<ConstantInt>(V)) | ||||||||||||||||
379 | if (this->isValue(CI->getValue())) { | ||||||||||||||||
380 | Res = &CI->getValue(); | ||||||||||||||||
381 | return true; | ||||||||||||||||
382 | } | ||||||||||||||||
383 | if (V->getType()->isVectorTy()) | ||||||||||||||||
384 | if (const auto *C = dyn_cast<Constant>(V)) | ||||||||||||||||
385 | if (auto *CI = dyn_cast_or_null<ConstantInt>(C->getSplatValue())) | ||||||||||||||||
386 | if (this->isValue(CI->getValue())) { | ||||||||||||||||
387 | Res = &CI->getValue(); | ||||||||||||||||
388 | return true; | ||||||||||||||||
389 | } | ||||||||||||||||
390 | |||||||||||||||||
391 | return false; | ||||||||||||||||
392 | } | ||||||||||||||||
393 | }; | ||||||||||||||||
394 | |||||||||||||||||
395 | /// This helper class is used to match scalar and vector constants that | ||||||||||||||||
396 | /// satisfy a specified predicate, and bind them to an APFloat. | ||||||||||||||||
397 | /// Undefs are allowed in splat vector constants. | ||||||||||||||||
398 | template <typename Predicate> struct apf_pred_ty : public Predicate { | ||||||||||||||||
399 | const APFloat *&Res; | ||||||||||||||||
400 | |||||||||||||||||
401 | apf_pred_ty(const APFloat *&R) : Res(R) {} | ||||||||||||||||
402 | |||||||||||||||||
403 | template <typename ITy> bool match(ITy *V) { | ||||||||||||||||
404 | if (const auto *CI = dyn_cast<ConstantFP>(V)) | ||||||||||||||||
405 | if (this->isValue(CI->getValue())) { | ||||||||||||||||
406 | Res = &CI->getValue(); | ||||||||||||||||
407 | return true; | ||||||||||||||||
408 | } | ||||||||||||||||
409 | if (V->getType()->isVectorTy()) | ||||||||||||||||
410 | if (const auto *C = dyn_cast<Constant>(V)) | ||||||||||||||||
411 | if (auto *CI = dyn_cast_or_null<ConstantFP>( | ||||||||||||||||
412 | C->getSplatValue(/* AllowUndef */ true))) | ||||||||||||||||
413 | if (this->isValue(CI->getValue())) { | ||||||||||||||||
414 | Res = &CI->getValue(); | ||||||||||||||||
415 | return true; | ||||||||||||||||
416 | } | ||||||||||||||||
417 | |||||||||||||||||
418 | return false; | ||||||||||||||||
419 | } | ||||||||||||||||
420 | }; | ||||||||||||||||
421 | |||||||||||||||||
422 | /////////////////////////////////////////////////////////////////////////////// | ||||||||||||||||
423 | // | ||||||||||||||||
424 | // Encapsulate constant value queries for use in templated predicate matchers. | ||||||||||||||||
425 | // This allows checking if constants match using compound predicates and works | ||||||||||||||||
426 | // with vector constants, possibly with relaxed constraints. For example, ignore | ||||||||||||||||
427 | // undef values. | ||||||||||||||||
428 | // | ||||||||||||||||
429 | /////////////////////////////////////////////////////////////////////////////// | ||||||||||||||||
430 | |||||||||||||||||
431 | struct is_any_apint { | ||||||||||||||||
432 | bool isValue(const APInt &C) { return true; } | ||||||||||||||||
433 | }; | ||||||||||||||||
434 | /// Match an integer or vector with any integral constant. | ||||||||||||||||
435 | /// For vectors, this includes constants with undefined elements. | ||||||||||||||||
436 | inline cst_pred_ty<is_any_apint> m_AnyIntegralConstant() { | ||||||||||||||||
437 | return cst_pred_ty<is_any_apint>(); | ||||||||||||||||
438 | } | ||||||||||||||||
439 | |||||||||||||||||
440 | struct is_all_ones { | ||||||||||||||||
441 | bool isValue(const APInt &C) { return C.isAllOnesValue(); } | ||||||||||||||||
442 | }; | ||||||||||||||||
443 | /// Match an integer or vector with all bits set. | ||||||||||||||||
444 | /// For vectors, this includes constants with undefined elements. | ||||||||||||||||
445 | inline cst_pred_ty<is_all_ones> m_AllOnes() { | ||||||||||||||||
446 | return cst_pred_ty<is_all_ones>(); | ||||||||||||||||
447 | } | ||||||||||||||||
448 | |||||||||||||||||
449 | struct is_maxsignedvalue { | ||||||||||||||||
450 | bool isValue(const APInt &C) { return C.isMaxSignedValue(); } | ||||||||||||||||
451 | }; | ||||||||||||||||
452 | /// Match an integer or vector with values having all bits except for the high | ||||||||||||||||
453 | /// bit set (0x7f...). | ||||||||||||||||
454 | /// For vectors, this includes constants with undefined elements. | ||||||||||||||||
455 | inline cst_pred_ty<is_maxsignedvalue> m_MaxSignedValue() { | ||||||||||||||||
456 | return cst_pred_ty<is_maxsignedvalue>(); | ||||||||||||||||
457 | } | ||||||||||||||||
458 | inline api_pred_ty<is_maxsignedvalue> m_MaxSignedValue(const APInt *&V) { | ||||||||||||||||
459 | return V; | ||||||||||||||||
460 | } | ||||||||||||||||
461 | |||||||||||||||||
462 | struct is_negative { | ||||||||||||||||
463 | bool isValue(const APInt &C) { return C.isNegative(); } | ||||||||||||||||
464 | }; | ||||||||||||||||
465 | /// Match an integer or vector of negative values. | ||||||||||||||||
466 | /// For vectors, this includes constants with undefined elements. | ||||||||||||||||
467 | inline cst_pred_ty<is_negative> m_Negative() { | ||||||||||||||||
468 | return cst_pred_ty<is_negative>(); | ||||||||||||||||
469 | } | ||||||||||||||||
470 | inline api_pred_ty<is_negative> m_Negative(const APInt *&V) { | ||||||||||||||||
471 | return V; | ||||||||||||||||
472 | } | ||||||||||||||||
473 | |||||||||||||||||
474 | struct is_nonnegative { | ||||||||||||||||
475 | bool isValue(const APInt &C) { return C.isNonNegative(); } | ||||||||||||||||
476 | }; | ||||||||||||||||
477 | /// Match an integer or vector of non-negative values. | ||||||||||||||||
478 | /// For vectors, this includes constants with undefined elements. | ||||||||||||||||
479 | inline cst_pred_ty<is_nonnegative> m_NonNegative() { | ||||||||||||||||
480 | return cst_pred_ty<is_nonnegative>(); | ||||||||||||||||
481 | } | ||||||||||||||||
482 | inline api_pred_ty<is_nonnegative> m_NonNegative(const APInt *&V) { | ||||||||||||||||
483 | return V; | ||||||||||||||||
484 | } | ||||||||||||||||
485 | |||||||||||||||||
486 | struct is_strictlypositive { | ||||||||||||||||
487 | bool isValue(const APInt &C) { return C.isStrictlyPositive(); } | ||||||||||||||||
488 | }; | ||||||||||||||||
489 | /// Match an integer or vector of strictly positive values. | ||||||||||||||||
490 | /// For vectors, this includes constants with undefined elements. | ||||||||||||||||
491 | inline cst_pred_ty<is_strictlypositive> m_StrictlyPositive() { | ||||||||||||||||
492 | return cst_pred_ty<is_strictlypositive>(); | ||||||||||||||||
493 | } | ||||||||||||||||
494 | inline api_pred_ty<is_strictlypositive> m_StrictlyPositive(const APInt *&V) { | ||||||||||||||||
495 | return V; | ||||||||||||||||
496 | } | ||||||||||||||||
497 | |||||||||||||||||
498 | struct is_nonpositive { | ||||||||||||||||
499 | bool isValue(const APInt &C) { return C.isNonPositive(); } | ||||||||||||||||
500 | }; | ||||||||||||||||
501 | /// Match an integer or vector of non-positive values. | ||||||||||||||||
502 | /// For vectors, this includes constants with undefined elements. | ||||||||||||||||
503 | inline cst_pred_ty<is_nonpositive> m_NonPositive() { | ||||||||||||||||
504 | return cst_pred_ty<is_nonpositive>(); | ||||||||||||||||
505 | } | ||||||||||||||||
506 | inline api_pred_ty<is_nonpositive> m_NonPositive(const APInt *&V) { return V; } | ||||||||||||||||
507 | |||||||||||||||||
508 | struct is_one { | ||||||||||||||||
509 | bool isValue(const APInt &C) { return C.isOneValue(); } | ||||||||||||||||
510 | }; | ||||||||||||||||
511 | /// Match an integer 1 or a vector with all elements equal to 1. | ||||||||||||||||
512 | /// For vectors, this includes constants with undefined elements. | ||||||||||||||||
513 | inline cst_pred_ty<is_one> m_One() { | ||||||||||||||||
514 | return cst_pred_ty<is_one>(); | ||||||||||||||||
515 | } | ||||||||||||||||
516 | |||||||||||||||||
517 | struct is_zero_int { | ||||||||||||||||
518 | bool isValue(const APInt &C) { return C.isNullValue(); } | ||||||||||||||||
519 | }; | ||||||||||||||||
520 | /// Match an integer 0 or a vector with all elements equal to 0. | ||||||||||||||||
521 | /// For vectors, this includes constants with undefined elements. | ||||||||||||||||
522 | inline cst_pred_ty<is_zero_int> m_ZeroInt() { | ||||||||||||||||
523 | return cst_pred_ty<is_zero_int>(); | ||||||||||||||||
524 | } | ||||||||||||||||
525 | |||||||||||||||||
526 | struct is_zero { | ||||||||||||||||
527 | template <typename ITy> bool match(ITy *V) { | ||||||||||||||||
528 | auto *C = dyn_cast<Constant>(V); | ||||||||||||||||
529 | // FIXME: this should be able to do something for scalable vectors | ||||||||||||||||
530 | return C && (C->isNullValue() || cst_pred_ty<is_zero_int>().match(C)); | ||||||||||||||||
531 | } | ||||||||||||||||
532 | }; | ||||||||||||||||
533 | /// Match any null constant or a vector with all elements equal to 0. | ||||||||||||||||
534 | /// For vectors, this includes constants with undefined elements. | ||||||||||||||||
535 | inline is_zero m_Zero() { | ||||||||||||||||
536 | return is_zero(); | ||||||||||||||||
537 | } | ||||||||||||||||
538 | |||||||||||||||||
539 | struct is_power2 { | ||||||||||||||||
540 | bool isValue(const APInt &C) { return C.isPowerOf2(); } | ||||||||||||||||
541 | }; | ||||||||||||||||
542 | /// Match an integer or vector power-of-2. | ||||||||||||||||
543 | /// For vectors, this includes constants with undefined elements. | ||||||||||||||||
544 | inline cst_pred_ty<is_power2> m_Power2() { | ||||||||||||||||
545 | return cst_pred_ty<is_power2>(); | ||||||||||||||||
546 | } | ||||||||||||||||
547 | inline api_pred_ty<is_power2> m_Power2(const APInt *&V) { | ||||||||||||||||
548 | return V; | ||||||||||||||||
549 | } | ||||||||||||||||
550 | |||||||||||||||||
551 | struct is_negated_power2 { | ||||||||||||||||
552 | bool isValue(const APInt &C) { return (-C).isPowerOf2(); } | ||||||||||||||||
553 | }; | ||||||||||||||||
554 | /// Match a integer or vector negated power-of-2. | ||||||||||||||||
555 | /// For vectors, this includes constants with undefined elements. | ||||||||||||||||
556 | inline cst_pred_ty<is_negated_power2> m_NegatedPower2() { | ||||||||||||||||
557 | return cst_pred_ty<is_negated_power2>(); | ||||||||||||||||
558 | } | ||||||||||||||||
559 | inline api_pred_ty<is_negated_power2> m_NegatedPower2(const APInt *&V) { | ||||||||||||||||
560 | return V; | ||||||||||||||||
561 | } | ||||||||||||||||
562 | |||||||||||||||||
563 | struct is_power2_or_zero { | ||||||||||||||||
564 | bool isValue(const APInt &C) { return !C || C.isPowerOf2(); } | ||||||||||||||||
565 | }; | ||||||||||||||||
566 | /// Match an integer or vector of 0 or power-of-2 values. | ||||||||||||||||
567 | /// For vectors, this includes constants with undefined elements. | ||||||||||||||||
568 | inline cst_pred_ty<is_power2_or_zero> m_Power2OrZero() { | ||||||||||||||||
569 | return cst_pred_ty<is_power2_or_zero>(); | ||||||||||||||||
570 | } | ||||||||||||||||
571 | inline api_pred_ty<is_power2_or_zero> m_Power2OrZero(const APInt *&V) { | ||||||||||||||||
572 | return V; | ||||||||||||||||
573 | } | ||||||||||||||||
574 | |||||||||||||||||
575 | struct is_sign_mask { | ||||||||||||||||
576 | bool isValue(const APInt &C) { return C.isSignMask(); } | ||||||||||||||||
577 | }; | ||||||||||||||||
578 | /// Match an integer or vector with only the sign bit(s) set. | ||||||||||||||||
579 | /// For vectors, this includes constants with undefined elements. | ||||||||||||||||
580 | inline cst_pred_ty<is_sign_mask> m_SignMask() { | ||||||||||||||||
581 | return cst_pred_ty<is_sign_mask>(); | ||||||||||||||||
582 | } | ||||||||||||||||
583 | |||||||||||||||||
584 | struct is_lowbit_mask { | ||||||||||||||||
585 | bool isValue(const APInt &C) { return C.isMask(); } | ||||||||||||||||
586 | }; | ||||||||||||||||
587 | /// Match an integer or vector with only the low bit(s) set. | ||||||||||||||||
588 | /// For vectors, this includes constants with undefined elements. | ||||||||||||||||
589 | inline cst_pred_ty<is_lowbit_mask> m_LowBitMask() { | ||||||||||||||||
590 | return cst_pred_ty<is_lowbit_mask>(); | ||||||||||||||||
591 | } | ||||||||||||||||
592 | |||||||||||||||||
593 | struct icmp_pred_with_threshold { | ||||||||||||||||
594 | ICmpInst::Predicate Pred; | ||||||||||||||||
595 | const APInt *Thr; | ||||||||||||||||
596 | bool isValue(const APInt &C) { | ||||||||||||||||
597 | switch (Pred) { | ||||||||||||||||
598 | case ICmpInst::Predicate::ICMP_EQ: | ||||||||||||||||
599 | return C.eq(*Thr); | ||||||||||||||||
600 | case ICmpInst::Predicate::ICMP_NE: | ||||||||||||||||
601 | return C.ne(*Thr); | ||||||||||||||||
602 | case ICmpInst::Predicate::ICMP_UGT: | ||||||||||||||||
603 | return C.ugt(*Thr); | ||||||||||||||||
604 | case ICmpInst::Predicate::ICMP_UGE: | ||||||||||||||||
605 | return C.uge(*Thr); | ||||||||||||||||
606 | case ICmpInst::Predicate::ICMP_ULT: | ||||||||||||||||
607 | return C.ult(*Thr); | ||||||||||||||||
608 | case ICmpInst::Predicate::ICMP_ULE: | ||||||||||||||||
609 | return C.ule(*Thr); | ||||||||||||||||
610 | case ICmpInst::Predicate::ICMP_SGT: | ||||||||||||||||
611 | return C.sgt(*Thr); | ||||||||||||||||
612 | case ICmpInst::Predicate::ICMP_SGE: | ||||||||||||||||
613 | return C.sge(*Thr); | ||||||||||||||||
614 | case ICmpInst::Predicate::ICMP_SLT: | ||||||||||||||||
615 | return C.slt(*Thr); | ||||||||||||||||
616 | case ICmpInst::Predicate::ICMP_SLE: | ||||||||||||||||
617 | return C.sle(*Thr); | ||||||||||||||||
618 | default: | ||||||||||||||||
619 | llvm_unreachable("Unhandled ICmp predicate")__builtin_unreachable(); | ||||||||||||||||
620 | } | ||||||||||||||||
621 | } | ||||||||||||||||
622 | }; | ||||||||||||||||
623 | /// Match an integer or vector with every element comparing 'pred' (eg/ne/...) | ||||||||||||||||
624 | /// to Threshold. For vectors, this includes constants with undefined elements. | ||||||||||||||||
625 | inline cst_pred_ty<icmp_pred_with_threshold> | ||||||||||||||||
626 | m_SpecificInt_ICMP(ICmpInst::Predicate Predicate, const APInt &Threshold) { | ||||||||||||||||
627 | cst_pred_ty<icmp_pred_with_threshold> P; | ||||||||||||||||
628 | P.Pred = Predicate; | ||||||||||||||||
629 | P.Thr = &Threshold; | ||||||||||||||||
630 | return P; | ||||||||||||||||
631 | } | ||||||||||||||||
632 | |||||||||||||||||
633 | struct is_nan { | ||||||||||||||||
634 | bool isValue(const APFloat &C) { return C.isNaN(); } | ||||||||||||||||
635 | }; | ||||||||||||||||
636 | /// Match an arbitrary NaN constant. This includes quiet and signalling nans. | ||||||||||||||||
637 | /// For vectors, this includes constants with undefined elements. | ||||||||||||||||
638 | inline cstfp_pred_ty<is_nan> m_NaN() { | ||||||||||||||||
639 | return cstfp_pred_ty<is_nan>(); | ||||||||||||||||
640 | } | ||||||||||||||||
641 | |||||||||||||||||
642 | struct is_nonnan { | ||||||||||||||||
643 | bool isValue(const APFloat &C) { return !C.isNaN(); } | ||||||||||||||||
644 | }; | ||||||||||||||||
645 | /// Match a non-NaN FP constant. | ||||||||||||||||
646 | /// For vectors, this includes constants with undefined elements. | ||||||||||||||||
647 | inline cstfp_pred_ty<is_nonnan> m_NonNaN() { | ||||||||||||||||
648 | return cstfp_pred_ty<is_nonnan>(); | ||||||||||||||||
649 | } | ||||||||||||||||
650 | |||||||||||||||||
651 | struct is_inf { | ||||||||||||||||
652 | bool isValue(const APFloat &C) { return C.isInfinity(); } | ||||||||||||||||
653 | }; | ||||||||||||||||
654 | /// Match a positive or negative infinity FP constant. | ||||||||||||||||
655 | /// For vectors, this includes constants with undefined elements. | ||||||||||||||||
656 | inline cstfp_pred_ty<is_inf> m_Inf() { | ||||||||||||||||
657 | return cstfp_pred_ty<is_inf>(); | ||||||||||||||||
658 | } | ||||||||||||||||
659 | |||||||||||||||||
660 | struct is_noninf { | ||||||||||||||||
661 | bool isValue(const APFloat &C) { return !C.isInfinity(); } | ||||||||||||||||
662 | }; | ||||||||||||||||
663 | /// Match a non-infinity FP constant, i.e. finite or NaN. | ||||||||||||||||
664 | /// For vectors, this includes constants with undefined elements. | ||||||||||||||||
665 | inline cstfp_pred_ty<is_noninf> m_NonInf() { | ||||||||||||||||
666 | return cstfp_pred_ty<is_noninf>(); | ||||||||||||||||
667 | } | ||||||||||||||||
668 | |||||||||||||||||
669 | struct is_finite { | ||||||||||||||||
670 | bool isValue(const APFloat &C) { return C.isFinite(); } | ||||||||||||||||
671 | }; | ||||||||||||||||
672 | /// Match a finite FP constant, i.e. not infinity or NaN. | ||||||||||||||||
673 | /// For vectors, this includes constants with undefined elements. | ||||||||||||||||
674 | inline cstfp_pred_ty<is_finite> m_Finite() { | ||||||||||||||||
675 | return cstfp_pred_ty<is_finite>(); | ||||||||||||||||
676 | } | ||||||||||||||||
677 | inline apf_pred_ty<is_finite> m_Finite(const APFloat *&V) { return V; } | ||||||||||||||||
678 | |||||||||||||||||
679 | struct is_finitenonzero { | ||||||||||||||||
680 | bool isValue(const APFloat &C) { return C.isFiniteNonZero(); } | ||||||||||||||||
681 | }; | ||||||||||||||||
682 | /// Match a finite non-zero FP constant. | ||||||||||||||||
683 | /// For vectors, this includes constants with undefined elements. | ||||||||||||||||
684 | inline cstfp_pred_ty<is_finitenonzero> m_FiniteNonZero() { | ||||||||||||||||
685 | return cstfp_pred_ty<is_finitenonzero>(); | ||||||||||||||||
686 | } | ||||||||||||||||
687 | inline apf_pred_ty<is_finitenonzero> m_FiniteNonZero(const APFloat *&V) { | ||||||||||||||||
688 | return V; | ||||||||||||||||
689 | } | ||||||||||||||||
690 | |||||||||||||||||
691 | struct is_any_zero_fp { | ||||||||||||||||
692 | bool isValue(const APFloat &C) { return C.isZero(); } | ||||||||||||||||
693 | }; | ||||||||||||||||
694 | /// Match a floating-point negative zero or positive zero. | ||||||||||||||||
695 | /// For vectors, this includes constants with undefined elements. | ||||||||||||||||
696 | inline cstfp_pred_ty<is_any_zero_fp> m_AnyZeroFP() { | ||||||||||||||||
697 | return cstfp_pred_ty<is_any_zero_fp>(); | ||||||||||||||||
698 | } | ||||||||||||||||
699 | |||||||||||||||||
700 | struct is_pos_zero_fp { | ||||||||||||||||
701 | bool isValue(const APFloat &C) { return C.isPosZero(); } | ||||||||||||||||
702 | }; | ||||||||||||||||
703 | /// Match a floating-point positive zero. | ||||||||||||||||
704 | /// For vectors, this includes constants with undefined elements. | ||||||||||||||||
705 | inline cstfp_pred_ty<is_pos_zero_fp> m_PosZeroFP() { | ||||||||||||||||
706 | return cstfp_pred_ty<is_pos_zero_fp>(); | ||||||||||||||||
707 | } | ||||||||||||||||
708 | |||||||||||||||||
709 | struct is_neg_zero_fp { | ||||||||||||||||
710 | bool isValue(const APFloat &C) { return C.isNegZero(); } | ||||||||||||||||
711 | }; | ||||||||||||||||
712 | /// Match a floating-point negative zero. | ||||||||||||||||
713 | /// For vectors, this includes constants with undefined elements. | ||||||||||||||||
714 | inline cstfp_pred_ty<is_neg_zero_fp> m_NegZeroFP() { | ||||||||||||||||
715 | return cstfp_pred_ty<is_neg_zero_fp>(); | ||||||||||||||||
716 | } | ||||||||||||||||
717 | |||||||||||||||||
718 | struct is_non_zero_fp { | ||||||||||||||||
719 | bool isValue(const APFloat &C) { return C.isNonZero(); } | ||||||||||||||||
720 | }; | ||||||||||||||||
721 | /// Match a floating-point non-zero. | ||||||||||||||||
722 | /// For vectors, this includes constants with undefined elements. | ||||||||||||||||
723 | inline cstfp_pred_ty<is_non_zero_fp> m_NonZeroFP() { | ||||||||||||||||
724 | return cstfp_pred_ty<is_non_zero_fp>(); | ||||||||||||||||
725 | } | ||||||||||||||||
726 | |||||||||||||||||
727 | /////////////////////////////////////////////////////////////////////////////// | ||||||||||||||||
728 | |||||||||||||||||
729 | template <typename Class> struct bind_ty { | ||||||||||||||||
730 | Class *&VR; | ||||||||||||||||
731 | |||||||||||||||||
732 | bind_ty(Class *&V) : VR(V) {} | ||||||||||||||||
733 | |||||||||||||||||
734 | template <typename ITy> bool match(ITy *V) { | ||||||||||||||||
735 | if (auto *CV = dyn_cast<Class>(V)) { | ||||||||||||||||
736 | VR = CV; | ||||||||||||||||
737 | return true; | ||||||||||||||||
738 | } | ||||||||||||||||
739 | return false; | ||||||||||||||||
740 | } | ||||||||||||||||
741 | }; | ||||||||||||||||
742 | |||||||||||||||||
743 | /// Match a value, capturing it if we match. | ||||||||||||||||
744 | inline bind_ty<Value> m_Value(Value *&V) { return V; } | ||||||||||||||||
745 | inline bind_ty<const Value> m_Value(const Value *&V) { return V; } | ||||||||||||||||
746 | |||||||||||||||||
747 | /// Match an instruction, capturing it if we match. | ||||||||||||||||
748 | inline bind_ty<Instruction> m_Instruction(Instruction *&I) { return I; } | ||||||||||||||||
749 | /// Match a unary operator, capturing it if we match. | ||||||||||||||||
750 | inline bind_ty<UnaryOperator> m_UnOp(UnaryOperator *&I) { return I; } | ||||||||||||||||
751 | /// Match a binary operator, capturing it if we match. | ||||||||||||||||
752 | inline bind_ty<BinaryOperator> m_BinOp(BinaryOperator *&I) { return I; } | ||||||||||||||||
753 | /// Match a with overflow intrinsic, capturing it if we match. | ||||||||||||||||
754 | inline bind_ty<WithOverflowInst> m_WithOverflowInst(WithOverflowInst *&I) { return I; } | ||||||||||||||||
755 | inline bind_ty<const WithOverflowInst> | ||||||||||||||||
756 | m_WithOverflowInst(const WithOverflowInst *&I) { | ||||||||||||||||
757 | return I; | ||||||||||||||||
758 | } | ||||||||||||||||
759 | |||||||||||||||||
760 | /// Match a Constant, capturing the value if we match. | ||||||||||||||||
761 | inline bind_ty<Constant> m_Constant(Constant *&C) { return C; } | ||||||||||||||||
762 | |||||||||||||||||
763 | /// Match a ConstantInt, capturing the value if we match. | ||||||||||||||||
764 | inline bind_ty<ConstantInt> m_ConstantInt(ConstantInt *&CI) { return CI; } | ||||||||||||||||
765 | |||||||||||||||||
766 | /// Match a ConstantFP, capturing the value if we match. | ||||||||||||||||
767 | inline bind_ty<ConstantFP> m_ConstantFP(ConstantFP *&C) { return C; } | ||||||||||||||||
768 | |||||||||||||||||
769 | /// Match a ConstantExpr, capturing the value if we match. | ||||||||||||||||
770 | inline bind_ty<ConstantExpr> m_ConstantExpr(ConstantExpr *&C) { return C; } | ||||||||||||||||
771 | |||||||||||||||||
772 | /// Match a basic block value, capturing it if we match. | ||||||||||||||||
773 | inline bind_ty<BasicBlock> m_BasicBlock(BasicBlock *&V) { return V; } | ||||||||||||||||
774 | inline bind_ty<const BasicBlock> m_BasicBlock(const BasicBlock *&V) { | ||||||||||||||||
775 | return V; | ||||||||||||||||
776 | } | ||||||||||||||||
777 | |||||||||||||||||
778 | /// Match an arbitrary immediate Constant and ignore it. | ||||||||||||||||
779 | inline match_combine_and<class_match<Constant>, | ||||||||||||||||
780 | match_unless<class_match<ConstantExpr>>> | ||||||||||||||||
781 | m_ImmConstant() { | ||||||||||||||||
782 | return m_CombineAnd(m_Constant(), m_Unless(m_ConstantExpr())); | ||||||||||||||||
783 | } | ||||||||||||||||
784 | |||||||||||||||||
785 | /// Match an immediate Constant, capturing the value if we match. | ||||||||||||||||
786 | inline match_combine_and<bind_ty<Constant>, | ||||||||||||||||
787 | match_unless<class_match<ConstantExpr>>> | ||||||||||||||||
788 | m_ImmConstant(Constant *&C) { | ||||||||||||||||
789 | return m_CombineAnd(m_Constant(C), m_Unless(m_ConstantExpr())); | ||||||||||||||||
790 | } | ||||||||||||||||
791 | |||||||||||||||||
792 | /// Match a specified Value*. | ||||||||||||||||
793 | struct specificval_ty { | ||||||||||||||||
794 | const Value *Val; | ||||||||||||||||
795 | |||||||||||||||||
796 | specificval_ty(const Value *V) : Val(V) {} | ||||||||||||||||
797 | |||||||||||||||||
798 | template <typename ITy> bool match(ITy *V) { return V == Val; } | ||||||||||||||||
799 | }; | ||||||||||||||||
800 | |||||||||||||||||
801 | /// Match if we have a specific specified value. | ||||||||||||||||
802 | inline specificval_ty m_Specific(const Value *V) { return V; } | ||||||||||||||||
803 | |||||||||||||||||
804 | /// Stores a reference to the Value *, not the Value * itself, | ||||||||||||||||
805 | /// thus can be used in commutative matchers. | ||||||||||||||||
806 | template <typename Class> struct deferredval_ty { | ||||||||||||||||
807 | Class *const &Val; | ||||||||||||||||
808 | |||||||||||||||||
809 | deferredval_ty(Class *const &V) : Val(V) {} | ||||||||||||||||
810 | |||||||||||||||||
811 | template <typename ITy> bool match(ITy *const V) { return V == Val; } | ||||||||||||||||
812 | }; | ||||||||||||||||
813 | |||||||||||||||||
814 | /// Like m_Specific(), but works if the specific value to match is determined | ||||||||||||||||
815 | /// as part of the same match() expression. For example: | ||||||||||||||||
816 | /// m_Add(m_Value(X), m_Specific(X)) is incorrect, because m_Specific() will | ||||||||||||||||
817 | /// bind X before the pattern match starts. | ||||||||||||||||
818 | /// m_Add(m_Value(X), m_Deferred(X)) is correct, and will check against | ||||||||||||||||
819 | /// whichever value m_Value(X) populated. | ||||||||||||||||
820 | inline deferredval_ty<Value> m_Deferred(Value *const &V) { return V; } | ||||||||||||||||
821 | inline deferredval_ty<const Value> m_Deferred(const Value *const &V) { | ||||||||||||||||
822 | return V; | ||||||||||||||||
823 | } | ||||||||||||||||
824 | |||||||||||||||||
825 | /// Match a specified floating point value or vector of all elements of | ||||||||||||||||
826 | /// that value. | ||||||||||||||||
827 | struct specific_fpval { | ||||||||||||||||
828 | double Val; | ||||||||||||||||
829 | |||||||||||||||||
830 | specific_fpval(double V) : Val(V) {} | ||||||||||||||||
831 | |||||||||||||||||
832 | template <typename ITy> bool match(ITy *V) { | ||||||||||||||||
833 | if (const auto *CFP = dyn_cast<ConstantFP>(V)) | ||||||||||||||||
834 | return CFP->isExactlyValue(Val); | ||||||||||||||||
835 | if (V->getType()->isVectorTy()) | ||||||||||||||||
836 | if (const auto *C = dyn_cast<Constant>(V)) | ||||||||||||||||
837 | if (auto *CFP = dyn_cast_or_null<ConstantFP>(C->getSplatValue())) | ||||||||||||||||
838 | return CFP->isExactlyValue(Val); | ||||||||||||||||
839 | return false; | ||||||||||||||||
840 | } | ||||||||||||||||
841 | }; | ||||||||||||||||
842 | |||||||||||||||||
843 | /// Match a specific floating point value or vector with all elements | ||||||||||||||||
844 | /// equal to the value. | ||||||||||||||||
845 | inline specific_fpval m_SpecificFP(double V) { return specific_fpval(V); } | ||||||||||||||||
846 | |||||||||||||||||
847 | /// Match a float 1.0 or vector with all elements equal to 1.0. | ||||||||||||||||
848 | inline specific_fpval m_FPOne() { return m_SpecificFP(1.0); } | ||||||||||||||||
849 | |||||||||||||||||
850 | struct bind_const_intval_ty { | ||||||||||||||||
851 | uint64_t &VR; | ||||||||||||||||
852 | |||||||||||||||||
853 | bind_const_intval_ty(uint64_t &V) : VR(V) {} | ||||||||||||||||
854 | |||||||||||||||||
855 | template <typename ITy> bool match(ITy *V) { | ||||||||||||||||
856 | if (const auto *CV = dyn_cast<ConstantInt>(V)) | ||||||||||||||||
857 | if (CV->getValue().ule(UINT64_MAX0xffffffffffffffffULL)) { | ||||||||||||||||
858 | VR = CV->getZExtValue(); | ||||||||||||||||
859 | return true; | ||||||||||||||||
860 | } | ||||||||||||||||
861 | return false; | ||||||||||||||||
862 | } | ||||||||||||||||
863 | }; | ||||||||||||||||
864 | |||||||||||||||||
865 | /// Match a specified integer value or vector of all elements of that | ||||||||||||||||
866 | /// value. | ||||||||||||||||
867 | template <bool AllowUndefs> | ||||||||||||||||
868 | struct specific_intval { | ||||||||||||||||
869 | APInt Val; | ||||||||||||||||
870 | |||||||||||||||||
871 | specific_intval(APInt V) : Val(std::move(V)) {} | ||||||||||||||||
872 | |||||||||||||||||
873 | template <typename ITy> bool match(ITy *V) { | ||||||||||||||||
874 | const auto *CI = dyn_cast<ConstantInt>(V); | ||||||||||||||||
875 | if (!CI && V->getType()->isVectorTy()) | ||||||||||||||||
876 | if (const auto *C = dyn_cast<Constant>(V)) | ||||||||||||||||
877 | CI = dyn_cast_or_null<ConstantInt>(C->getSplatValue(AllowUndefs)); | ||||||||||||||||
878 | |||||||||||||||||
879 | return CI && APInt::isSameValue(CI->getValue(), Val); | ||||||||||||||||
880 | } | ||||||||||||||||
881 | }; | ||||||||||||||||
882 | |||||||||||||||||
883 | /// Match a specific integer value or vector with all elements equal to | ||||||||||||||||
884 | /// the value. | ||||||||||||||||
885 | inline specific_intval<false> m_SpecificInt(APInt V) { | ||||||||||||||||
886 | return specific_intval<false>(std::move(V)); | ||||||||||||||||
887 | } | ||||||||||||||||
888 | |||||||||||||||||
889 | inline specific_intval<false> m_SpecificInt(uint64_t V) { | ||||||||||||||||
890 | return m_SpecificInt(APInt(64, V)); | ||||||||||||||||
891 | } | ||||||||||||||||
892 | |||||||||||||||||
893 | inline specific_intval<true> m_SpecificIntAllowUndef(APInt V) { | ||||||||||||||||
894 | return specific_intval<true>(std::move(V)); | ||||||||||||||||
895 | } | ||||||||||||||||
896 | |||||||||||||||||
897 | inline specific_intval<true> m_SpecificIntAllowUndef(uint64_t V) { | ||||||||||||||||
898 | return m_SpecificIntAllowUndef(APInt(64, V)); | ||||||||||||||||
899 | } | ||||||||||||||||
900 | |||||||||||||||||
901 | /// Match a ConstantInt and bind to its value. This does not match | ||||||||||||||||
902 | /// ConstantInts wider than 64-bits. | ||||||||||||||||
903 | inline bind_const_intval_ty m_ConstantInt(uint64_t &V) { return V; } | ||||||||||||||||
904 | |||||||||||||||||
905 | /// Match a specified basic block value. | ||||||||||||||||
906 | struct specific_bbval { | ||||||||||||||||
907 | BasicBlock *Val; | ||||||||||||||||
908 | |||||||||||||||||
909 | specific_bbval(BasicBlock *Val) : Val(Val) {} | ||||||||||||||||
910 | |||||||||||||||||
911 | template <typename ITy> bool match(ITy *V) { | ||||||||||||||||
912 | const auto *BB = dyn_cast<BasicBlock>(V); | ||||||||||||||||
913 | return BB && BB == Val; | ||||||||||||||||
914 | } | ||||||||||||||||
915 | }; | ||||||||||||||||
916 | |||||||||||||||||
917 | /// Match a specific basic block value. | ||||||||||||||||
918 | inline specific_bbval m_SpecificBB(BasicBlock *BB) { | ||||||||||||||||
919 | return specific_bbval(BB); | ||||||||||||||||
920 | } | ||||||||||||||||
921 | |||||||||||||||||
922 | /// A commutative-friendly version of m_Specific(). | ||||||||||||||||
923 | inline deferredval_ty<BasicBlock> m_Deferred(BasicBlock *const &BB) { | ||||||||||||||||
924 | return BB; | ||||||||||||||||
925 | } | ||||||||||||||||
926 | inline deferredval_ty<const BasicBlock> | ||||||||||||||||
927 | m_Deferred(const BasicBlock *const &BB) { | ||||||||||||||||
928 | return BB; | ||||||||||||||||
929 | } | ||||||||||||||||
930 | |||||||||||||||||
931 | //===----------------------------------------------------------------------===// | ||||||||||||||||
932 | // Matcher for any binary operator. | ||||||||||||||||
933 | // | ||||||||||||||||
934 | template <typename LHS_t, typename RHS_t, bool Commutable = false> | ||||||||||||||||
935 | struct AnyBinaryOp_match { | ||||||||||||||||
936 | LHS_t L; | ||||||||||||||||
937 | RHS_t R; | ||||||||||||||||
938 | |||||||||||||||||
939 | // The evaluation order is always stable, regardless of Commutability. | ||||||||||||||||
940 | // The LHS is always matched first. | ||||||||||||||||
941 | AnyBinaryOp_match(const LHS_t &LHS, const RHS_t &RHS) : L(LHS), R(RHS) {} | ||||||||||||||||
942 | |||||||||||||||||
943 | template <typename OpTy> bool match(OpTy *V) { | ||||||||||||||||
944 | if (auto *I = dyn_cast<BinaryOperator>(V)) | ||||||||||||||||
945 | return (L.match(I->getOperand(0)) && R.match(I->getOperand(1))) || | ||||||||||||||||
946 | (Commutable && L.match(I->getOperand(1)) && | ||||||||||||||||
947 | R.match(I->getOperand(0))); | ||||||||||||||||
948 | return false; | ||||||||||||||||
949 | } | ||||||||||||||||
950 | }; | ||||||||||||||||
951 | |||||||||||||||||
952 | template <typename LHS, typename RHS> | ||||||||||||||||
953 | inline AnyBinaryOp_match<LHS, RHS> m_BinOp(const LHS &L, const RHS &R) { | ||||||||||||||||
954 | return AnyBinaryOp_match<LHS, RHS>(L, R); | ||||||||||||||||
955 | } | ||||||||||||||||
956 | |||||||||||||||||
957 | //===----------------------------------------------------------------------===// | ||||||||||||||||
958 | // Matcher for any unary operator. | ||||||||||||||||
959 | // TODO fuse unary, binary matcher into n-ary matcher | ||||||||||||||||
960 | // | ||||||||||||||||
961 | template <typename OP_t> struct AnyUnaryOp_match { | ||||||||||||||||
962 | OP_t X; | ||||||||||||||||
963 | |||||||||||||||||
964 | AnyUnaryOp_match(const OP_t &X) : X(X) {} | ||||||||||||||||
965 | |||||||||||||||||
966 | template <typename OpTy> bool match(OpTy *V) { | ||||||||||||||||
967 | if (auto *I = dyn_cast<UnaryOperator>(V)) | ||||||||||||||||
968 | return X.match(I->getOperand(0)); | ||||||||||||||||
969 | return false; | ||||||||||||||||
970 | } | ||||||||||||||||
971 | }; | ||||||||||||||||
972 | |||||||||||||||||
973 | template <typename OP_t> inline AnyUnaryOp_match<OP_t> m_UnOp(const OP_t &X) { | ||||||||||||||||
974 | return AnyUnaryOp_match<OP_t>(X); | ||||||||||||||||
975 | } | ||||||||||||||||
976 | |||||||||||||||||
977 | //===----------------------------------------------------------------------===// | ||||||||||||||||
978 | // Matchers for specific binary operators. | ||||||||||||||||
979 | // | ||||||||||||||||
980 | |||||||||||||||||
981 | template <typename LHS_t, typename RHS_t, unsigned Opcode, | ||||||||||||||||
982 | bool Commutable = false> | ||||||||||||||||
983 | struct BinaryOp_match { | ||||||||||||||||
984 | LHS_t L; | ||||||||||||||||
985 | RHS_t R; | ||||||||||||||||
986 | |||||||||||||||||
987 | // The evaluation order is always stable, regardless of Commutability. | ||||||||||||||||
988 | // The LHS is always matched first. | ||||||||||||||||
989 | BinaryOp_match(const LHS_t &LHS, const RHS_t &RHS) : L(LHS), R(RHS) {} | ||||||||||||||||
990 | |||||||||||||||||
991 | template <typename OpTy> bool match(OpTy *V) { | ||||||||||||||||
992 | if (V->getValueID() == Value::InstructionVal + Opcode) { | ||||||||||||||||
993 | auto *I = cast<BinaryOperator>(V); | ||||||||||||||||
994 | return (L.match(I->getOperand(0)) && R.match(I->getOperand(1))) || | ||||||||||||||||
995 | (Commutable && L.match(I->getOperand(1)) && | ||||||||||||||||
996 | R.match(I->getOperand(0))); | ||||||||||||||||
997 | } | ||||||||||||||||
998 | if (auto *CE = dyn_cast<ConstantExpr>(V)) | ||||||||||||||||
999 | return CE->getOpcode() == Opcode && | ||||||||||||||||
1000 | ((L.match(CE->getOperand(0)) && R.match(CE->getOperand(1))) || | ||||||||||||||||
1001 | (Commutable && L.match(CE->getOperand(1)) && | ||||||||||||||||
1002 | R.match(CE->getOperand(0)))); | ||||||||||||||||
1003 | return false; | ||||||||||||||||
1004 | } | ||||||||||||||||
1005 | }; | ||||||||||||||||
1006 | |||||||||||||||||
1007 | template <typename LHS, typename RHS> | ||||||||||||||||
1008 | inline BinaryOp_match<LHS, RHS, Instruction::Add> m_Add(const LHS &L, | ||||||||||||||||
1009 | const RHS &R) { | ||||||||||||||||
1010 | return BinaryOp_match<LHS, RHS, Instruction::Add>(L, R); | ||||||||||||||||
1011 | } | ||||||||||||||||
1012 | |||||||||||||||||
1013 | template <typename LHS, typename RHS> | ||||||||||||||||
1014 | inline BinaryOp_match<LHS, RHS, Instruction::FAdd> m_FAdd(const LHS &L, | ||||||||||||||||
1015 | const RHS &R) { | ||||||||||||||||
1016 | return BinaryOp_match<LHS, RHS, Instruction::FAdd>(L, R); | ||||||||||||||||
1017 | } | ||||||||||||||||
1018 | |||||||||||||||||
1019 | template <typename LHS, typename RHS> | ||||||||||||||||
1020 | inline BinaryOp_match<LHS, RHS, Instruction::Sub> m_Sub(const LHS &L, | ||||||||||||||||
1021 | const RHS &R) { | ||||||||||||||||
1022 | return BinaryOp_match<LHS, RHS, Instruction::Sub>(L, R); | ||||||||||||||||
1023 | } | ||||||||||||||||
1024 | |||||||||||||||||
1025 | template <typename LHS, typename RHS> | ||||||||||||||||
1026 | inline BinaryOp_match<LHS, RHS, Instruction::FSub> m_FSub(const LHS &L, | ||||||||||||||||
1027 | const RHS &R) { | ||||||||||||||||
1028 | return BinaryOp_match<LHS, RHS, Instruction::FSub>(L, R); | ||||||||||||||||
1029 | } | ||||||||||||||||
1030 | |||||||||||||||||
1031 | template <typename Op_t> struct FNeg_match { | ||||||||||||||||
1032 | Op_t X; | ||||||||||||||||
1033 | |||||||||||||||||
1034 | FNeg_match(const Op_t &Op) : X(Op) {} | ||||||||||||||||
1035 | template <typename OpTy> bool match(OpTy *V) { | ||||||||||||||||
1036 | auto *FPMO = dyn_cast<FPMathOperator>(V); | ||||||||||||||||
1037 | if (!FPMO) return false; | ||||||||||||||||
1038 | |||||||||||||||||
1039 | if (FPMO->getOpcode() == Instruction::FNeg) | ||||||||||||||||
1040 | return X.match(FPMO->getOperand(0)); | ||||||||||||||||
1041 | |||||||||||||||||
1042 | if (FPMO->getOpcode() == Instruction::FSub) { | ||||||||||||||||
1043 | if (FPMO->hasNoSignedZeros()) { | ||||||||||||||||
1044 | // With 'nsz', any zero goes. | ||||||||||||||||
1045 | if (!cstfp_pred_ty<is_any_zero_fp>().match(FPMO->getOperand(0))) | ||||||||||||||||
1046 | return false; | ||||||||||||||||
1047 | } else { | ||||||||||||||||
1048 | // Without 'nsz', we need fsub -0.0, X exactly. | ||||||||||||||||
1049 | if (!cstfp_pred_ty<is_neg_zero_fp>().match(FPMO->getOperand(0))) | ||||||||||||||||
1050 | return false; | ||||||||||||||||
1051 | } | ||||||||||||||||
1052 | |||||||||||||||||
1053 | return X.match(FPMO->getOperand(1)); | ||||||||||||||||
1054 | } | ||||||||||||||||
1055 | |||||||||||||||||
1056 | return false; | ||||||||||||||||
1057 | } | ||||||||||||||||
1058 | }; | ||||||||||||||||
1059 | |||||||||||||||||
1060 | /// Match 'fneg X' as 'fsub -0.0, X'. | ||||||||||||||||
1061 | template <typename OpTy> | ||||||||||||||||
1062 | inline FNeg_match<OpTy> | ||||||||||||||||
1063 | m_FNeg(const OpTy &X) { | ||||||||||||||||
1064 | return FNeg_match<OpTy>(X); | ||||||||||||||||
1065 | } | ||||||||||||||||
1066 | |||||||||||||||||
1067 | /// Match 'fneg X' as 'fsub +-0.0, X'. | ||||||||||||||||
1068 | template <typename RHS> | ||||||||||||||||
1069 | inline BinaryOp_match<cstfp_pred_ty<is_any_zero_fp>, RHS, Instruction::FSub> | ||||||||||||||||
1070 | m_FNegNSZ(const RHS &X) { | ||||||||||||||||
1071 | return m_FSub(m_AnyZeroFP(), X); | ||||||||||||||||
1072 | } | ||||||||||||||||
1073 | |||||||||||||||||
1074 | template <typename LHS, typename RHS> | ||||||||||||||||
1075 | inline BinaryOp_match<LHS, RHS, Instruction::Mul> m_Mul(const LHS &L, | ||||||||||||||||
1076 | const RHS &R) { | ||||||||||||||||
1077 | return BinaryOp_match<LHS, RHS, Instruction::Mul>(L, R); | ||||||||||||||||
1078 | } | ||||||||||||||||
1079 | |||||||||||||||||
1080 | template <typename LHS, typename RHS> | ||||||||||||||||
1081 | inline BinaryOp_match<LHS, RHS, Instruction::FMul> m_FMul(const LHS &L, | ||||||||||||||||
1082 | const RHS &R) { | ||||||||||||||||
1083 | return BinaryOp_match<LHS, RHS, Instruction::FMul>(L, R); | ||||||||||||||||
1084 | } | ||||||||||||||||
1085 | |||||||||||||||||
1086 | template <typename LHS, typename RHS> | ||||||||||||||||
1087 | inline BinaryOp_match<LHS, RHS, Instruction::UDiv> m_UDiv(const LHS &L, | ||||||||||||||||
1088 | const RHS &R) { | ||||||||||||||||
1089 | return BinaryOp_match<LHS, RHS, Instruction::UDiv>(L, R); | ||||||||||||||||
1090 | } | ||||||||||||||||
1091 | |||||||||||||||||
1092 | template <typename LHS, typename RHS> | ||||||||||||||||
1093 | inline BinaryOp_match<LHS, RHS, Instruction::SDiv> m_SDiv(const LHS &L, | ||||||||||||||||
1094 | const RHS &R) { | ||||||||||||||||
1095 | return BinaryOp_match<LHS, RHS, Instruction::SDiv>(L, R); | ||||||||||||||||
1096 | } | ||||||||||||||||
1097 | |||||||||||||||||
1098 | template <typename LHS, typename RHS> | ||||||||||||||||
1099 | inline BinaryOp_match<LHS, RHS, Instruction::FDiv> m_FDiv(const LHS &L, | ||||||||||||||||
1100 | const RHS &R) { | ||||||||||||||||
1101 | return BinaryOp_match<LHS, RHS, Instruction::FDiv>(L, R); | ||||||||||||||||
1102 | } | ||||||||||||||||
1103 | |||||||||||||||||
1104 | template <typename LHS, typename RHS> | ||||||||||||||||
1105 | inline BinaryOp_match<LHS, RHS, Instruction::URem> m_URem(const LHS &L, | ||||||||||||||||
1106 | const RHS &R) { | ||||||||||||||||
1107 | return BinaryOp_match<LHS, RHS, Instruction::URem>(L, R); | ||||||||||||||||
1108 | } | ||||||||||||||||
1109 | |||||||||||||||||
1110 | template <typename LHS, typename RHS> | ||||||||||||||||
1111 | inline BinaryOp_match<LHS, RHS, Instruction::SRem> m_SRem(const LHS &L, | ||||||||||||||||
1112 | const RHS &R) { | ||||||||||||||||
1113 | return BinaryOp_match<LHS, RHS, Instruction::SRem>(L, R); | ||||||||||||||||
1114 | } | ||||||||||||||||
1115 | |||||||||||||||||
1116 | template <typename LHS, typename RHS> | ||||||||||||||||
1117 | inline BinaryOp_match<LHS, RHS, Instruction::FRem> m_FRem(const LHS &L, | ||||||||||||||||
1118 | const RHS &R) { | ||||||||||||||||
1119 | return BinaryOp_match<LHS, RHS, Instruction::FRem>(L, R); | ||||||||||||||||
1120 | } | ||||||||||||||||
1121 | |||||||||||||||||
1122 | template <typename LHS, typename RHS> | ||||||||||||||||
1123 | inline BinaryOp_match<LHS, RHS, Instruction::And> m_And(const LHS &L, | ||||||||||||||||
1124 | const RHS &R) { | ||||||||||||||||
1125 | return BinaryOp_match<LHS, RHS, Instruction::And>(L, R); | ||||||||||||||||
1126 | } | ||||||||||||||||
1127 | |||||||||||||||||
1128 | template <typename LHS, typename RHS> | ||||||||||||||||
1129 | inline BinaryOp_match<LHS, RHS, Instruction::Or> m_Or(const LHS &L, | ||||||||||||||||
1130 | const RHS &R) { | ||||||||||||||||
1131 | return BinaryOp_match<LHS, RHS, Instruction::Or>(L, R); | ||||||||||||||||
1132 | } | ||||||||||||||||
1133 | |||||||||||||||||
1134 | template <typename LHS, typename RHS> | ||||||||||||||||
1135 | inline BinaryOp_match<LHS, RHS, Instruction::Xor> m_Xor(const LHS &L, | ||||||||||||||||
1136 | const RHS &R) { | ||||||||||||||||
1137 | return BinaryOp_match<LHS, RHS, Instruction::Xor>(L, R); | ||||||||||||||||
1138 | } | ||||||||||||||||
1139 | |||||||||||||||||
1140 | template <typename LHS, typename RHS> | ||||||||||||||||
1141 | inline BinaryOp_match<LHS, RHS, Instruction::Shl> m_Shl(const LHS &L, | ||||||||||||||||
1142 | const RHS &R) { | ||||||||||||||||
1143 | return BinaryOp_match<LHS, RHS, Instruction::Shl>(L, R); | ||||||||||||||||
1144 | } | ||||||||||||||||
1145 | |||||||||||||||||
1146 | template <typename LHS, typename RHS> | ||||||||||||||||
1147 | inline BinaryOp_match<LHS, RHS, Instruction::LShr> m_LShr(const LHS &L, | ||||||||||||||||
1148 | const RHS &R) { | ||||||||||||||||
1149 | return BinaryOp_match<LHS, RHS, Instruction::LShr>(L, R); | ||||||||||||||||
1150 | } | ||||||||||||||||
1151 | |||||||||||||||||
1152 | template <typename LHS, typename RHS> | ||||||||||||||||
1153 | inline BinaryOp_match<LHS, RHS, Instruction::AShr> m_AShr(const LHS &L, | ||||||||||||||||
1154 | const RHS &R) { | ||||||||||||||||
1155 | return BinaryOp_match<LHS, RHS, Instruction::AShr>(L, R); | ||||||||||||||||
1156 | } | ||||||||||||||||
1157 | |||||||||||||||||
1158 | template <typename LHS_t, typename RHS_t, unsigned Opcode, | ||||||||||||||||
1159 | unsigned WrapFlags = 0> | ||||||||||||||||
1160 | struct OverflowingBinaryOp_match { | ||||||||||||||||
1161 | LHS_t L; | ||||||||||||||||
1162 | RHS_t R; | ||||||||||||||||
1163 | |||||||||||||||||
1164 | OverflowingBinaryOp_match(const LHS_t &LHS, const RHS_t &RHS) | ||||||||||||||||
1165 | : L(LHS), R(RHS) {} | ||||||||||||||||
1166 | |||||||||||||||||
1167 | template <typename OpTy> bool match(OpTy *V) { | ||||||||||||||||
1168 | if (auto *Op = dyn_cast<OverflowingBinaryOperator>(V)) { | ||||||||||||||||
1169 | if (Op->getOpcode() != Opcode) | ||||||||||||||||
1170 | return false; | ||||||||||||||||
1171 | if ((WrapFlags & OverflowingBinaryOperator::NoUnsignedWrap) && | ||||||||||||||||
1172 | !Op->hasNoUnsignedWrap()) | ||||||||||||||||
1173 | return false; | ||||||||||||||||
1174 | if ((WrapFlags & OverflowingBinaryOperator::NoSignedWrap) && | ||||||||||||||||
1175 | !Op->hasNoSignedWrap()) | ||||||||||||||||
1176 | return false; | ||||||||||||||||
1177 | return L.match(Op->getOperand(0)) && R.match(Op->getOperand(1)); | ||||||||||||||||
1178 | } | ||||||||||||||||
1179 | return false; | ||||||||||||||||
1180 | } | ||||||||||||||||
1181 | }; | ||||||||||||||||
1182 | |||||||||||||||||
1183 | template <typename LHS, typename RHS> | ||||||||||||||||
1184 | inline OverflowingBinaryOp_match<LHS, RHS, Instruction::Add, | ||||||||||||||||
1185 | OverflowingBinaryOperator::NoSignedWrap> | ||||||||||||||||
1186 | m_NSWAdd(const LHS &L, const RHS &R) { | ||||||||||||||||
1187 | return OverflowingBinaryOp_match<LHS, RHS, Instruction::Add, | ||||||||||||||||
1188 | OverflowingBinaryOperator::NoSignedWrap>( | ||||||||||||||||
1189 | L, R); | ||||||||||||||||
1190 | } | ||||||||||||||||
1191 | template <typename LHS, typename RHS> | ||||||||||||||||
1192 | inline OverflowingBinaryOp_match<LHS, RHS, Instruction::Sub, | ||||||||||||||||
1193 | OverflowingBinaryOperator::NoSignedWrap> | ||||||||||||||||
1194 | m_NSWSub(const LHS &L, const RHS &R) { | ||||||||||||||||
1195 | return OverflowingBinaryOp_match<LHS, RHS, Instruction::Sub, | ||||||||||||||||
1196 | OverflowingBinaryOperator::NoSignedWrap>( | ||||||||||||||||
1197 | L, R); | ||||||||||||||||
1198 | } | ||||||||||||||||
1199 | template <typename LHS, typename RHS> | ||||||||||||||||
1200 | inline OverflowingBinaryOp_match<LHS, RHS, Instruction::Mul, | ||||||||||||||||
1201 | OverflowingBinaryOperator::NoSignedWrap> | ||||||||||||||||
1202 | m_NSWMul(const LHS &L, const RHS &R) { | ||||||||||||||||
1203 | return OverflowingBinaryOp_match<LHS, RHS, Instruction::Mul, | ||||||||||||||||
1204 | OverflowingBinaryOperator::NoSignedWrap>( | ||||||||||||||||
1205 | L, R); | ||||||||||||||||
1206 | } | ||||||||||||||||
1207 | template <typename LHS, typename RHS> | ||||||||||||||||
1208 | inline OverflowingBinaryOp_match<LHS, RHS, Instruction::Shl, | ||||||||||||||||
1209 | OverflowingBinaryOperator::NoSignedWrap> | ||||||||||||||||
1210 | m_NSWShl(const LHS &L, const RHS &R) { | ||||||||||||||||
1211 | return OverflowingBinaryOp_match<LHS, RHS, Instruction::Shl, | ||||||||||||||||
1212 | OverflowingBinaryOperator::NoSignedWrap>( | ||||||||||||||||
1213 | L, R); | ||||||||||||||||
1214 | } | ||||||||||||||||
1215 | |||||||||||||||||
1216 | template <typename LHS, typename RHS> | ||||||||||||||||
1217 | inline OverflowingBinaryOp_match<LHS, RHS, Instruction::Add, | ||||||||||||||||
1218 | OverflowingBinaryOperator::NoUnsignedWrap> | ||||||||||||||||
1219 | m_NUWAdd(const LHS &L, const RHS &R) { | ||||||||||||||||
1220 | return OverflowingBinaryOp_match<LHS, RHS, Instruction::Add, | ||||||||||||||||
1221 | OverflowingBinaryOperator::NoUnsignedWrap>( | ||||||||||||||||
1222 | L, R); | ||||||||||||||||
1223 | } | ||||||||||||||||
1224 | template <typename LHS, typename RHS> | ||||||||||||||||
1225 | inline OverflowingBinaryOp_match<LHS, RHS, Instruction::Sub, | ||||||||||||||||
1226 | OverflowingBinaryOperator::NoUnsignedWrap> | ||||||||||||||||
1227 | m_NUWSub(const LHS &L, const RHS &R) { | ||||||||||||||||
1228 | return OverflowingBinaryOp_match<LHS, RHS, Instruction::Sub, | ||||||||||||||||
1229 | OverflowingBinaryOperator::NoUnsignedWrap>( | ||||||||||||||||
1230 | L, R); | ||||||||||||||||
1231 | } | ||||||||||||||||
1232 | template <typename LHS, typename RHS> | ||||||||||||||||
1233 | inline OverflowingBinaryOp_match<LHS, RHS, Instruction::Mul, | ||||||||||||||||
1234 | OverflowingBinaryOperator::NoUnsignedWrap> | ||||||||||||||||
1235 | m_NUWMul(const LHS &L, const RHS &R) { | ||||||||||||||||
1236 | return OverflowingBinaryOp_match<LHS, RHS, Instruction::Mul, | ||||||||||||||||
1237 | OverflowingBinaryOperator::NoUnsignedWrap>( | ||||||||||||||||
1238 | L, R); | ||||||||||||||||
1239 | } | ||||||||||||||||
1240 | template <typename LHS, typename RHS> | ||||||||||||||||
1241 | inline OverflowingBinaryOp_match<LHS, RHS, Instruction::Shl, | ||||||||||||||||
1242 | OverflowingBinaryOperator::NoUnsignedWrap> | ||||||||||||||||
1243 | m_NUWShl(const LHS &L, const RHS &R) { | ||||||||||||||||
1244 | return OverflowingBinaryOp_match<LHS, RHS, Instruction::Shl, | ||||||||||||||||
1245 | OverflowingBinaryOperator::NoUnsignedWrap>( | ||||||||||||||||
1246 | L, R); | ||||||||||||||||
1247 | } | ||||||||||||||||
1248 | |||||||||||||||||
1249 | //===----------------------------------------------------------------------===// | ||||||||||||||||
1250 | // Class that matches a group of binary opcodes. | ||||||||||||||||
1251 | // | ||||||||||||||||
1252 | template <typename LHS_t, typename RHS_t, typename Predicate> | ||||||||||||||||
1253 | struct BinOpPred_match : Predicate { | ||||||||||||||||
1254 | LHS_t L; | ||||||||||||||||
1255 | RHS_t R; | ||||||||||||||||
1256 | |||||||||||||||||
1257 | BinOpPred_match(const LHS_t &LHS, const RHS_t &RHS) : L(LHS), R(RHS) {} | ||||||||||||||||
1258 | |||||||||||||||||
1259 | template <typename OpTy> bool match(OpTy *V) { | ||||||||||||||||
1260 | if (auto *I = dyn_cast<Instruction>(V)) | ||||||||||||||||
1261 | return this->isOpType(I->getOpcode()) && L.match(I->getOperand(0)) && | ||||||||||||||||
1262 | R.match(I->getOperand(1)); | ||||||||||||||||
1263 | if (auto *CE = dyn_cast<ConstantExpr>(V)) | ||||||||||||||||
1264 | return this->isOpType(CE->getOpcode()) && L.match(CE->getOperand(0)) && | ||||||||||||||||
1265 | R.match(CE->getOperand(1)); | ||||||||||||||||
1266 | return false; | ||||||||||||||||
1267 | } | ||||||||||||||||
1268 | }; | ||||||||||||||||
1269 | |||||||||||||||||
1270 | struct is_shift_op { | ||||||||||||||||
1271 | bool isOpType(unsigned Opcode) { return Instruction::isShift(Opcode); } | ||||||||||||||||
1272 | }; | ||||||||||||||||
1273 | |||||||||||||||||
1274 | struct is_right_shift_op { | ||||||||||||||||
1275 | bool isOpType(unsigned Opcode) { | ||||||||||||||||
1276 | return Opcode == Instruction::LShr || Opcode == Instruction::AShr; | ||||||||||||||||
1277 | } | ||||||||||||||||
1278 | }; | ||||||||||||||||
1279 | |||||||||||||||||
1280 | struct is_logical_shift_op { | ||||||||||||||||
1281 | bool isOpType(unsigned Opcode) { | ||||||||||||||||
1282 | return Opcode == Instruction::LShr || Opcode == Instruction::Shl; | ||||||||||||||||
1283 | } | ||||||||||||||||
1284 | }; | ||||||||||||||||
1285 | |||||||||||||||||
1286 | struct is_bitwiselogic_op { | ||||||||||||||||
1287 | bool isOpType(unsigned Opcode) { | ||||||||||||||||
1288 | return Instruction::isBitwiseLogicOp(Opcode); | ||||||||||||||||
1289 | } | ||||||||||||||||
1290 | }; | ||||||||||||||||
1291 | |||||||||||||||||
1292 | struct is_idiv_op { | ||||||||||||||||
1293 | bool isOpType(unsigned Opcode) { | ||||||||||||||||
1294 | return Opcode == Instruction::SDiv || Opcode == Instruction::UDiv; | ||||||||||||||||
1295 | } | ||||||||||||||||
1296 | }; | ||||||||||||||||
1297 | |||||||||||||||||
1298 | struct is_irem_op { | ||||||||||||||||
1299 | bool isOpType(unsigned Opcode) { | ||||||||||||||||
1300 | return Opcode == Instruction::SRem || Opcode == Instruction::URem; | ||||||||||||||||
1301 | } | ||||||||||||||||
1302 | }; | ||||||||||||||||
1303 | |||||||||||||||||
1304 | /// Matches shift operations. | ||||||||||||||||
1305 | template <typename LHS, typename RHS> | ||||||||||||||||
1306 | inline BinOpPred_match<LHS, RHS, is_shift_op> m_Shift(const LHS &L, | ||||||||||||||||
1307 | const RHS &R) { | ||||||||||||||||
1308 | return BinOpPred_match<LHS, RHS, is_shift_op>(L, R); | ||||||||||||||||
1309 | } | ||||||||||||||||
1310 | |||||||||||||||||
1311 | /// Matches logical shift operations. | ||||||||||||||||
1312 | template <typename LHS, typename RHS> | ||||||||||||||||
1313 | inline BinOpPred_match<LHS, RHS, is_right_shift_op> m_Shr(const LHS &L, | ||||||||||||||||
1314 | const RHS &R) { | ||||||||||||||||
1315 | return BinOpPred_match<LHS, RHS, is_right_shift_op>(L, R); | ||||||||||||||||
1316 | } | ||||||||||||||||
1317 | |||||||||||||||||
1318 | /// Matches logical shift operations. | ||||||||||||||||
1319 | template <typename LHS, typename RHS> | ||||||||||||||||
1320 | inline BinOpPred_match<LHS, RHS, is_logical_shift_op> | ||||||||||||||||
1321 | m_LogicalShift(const LHS &L, const RHS &R) { | ||||||||||||||||
1322 | return BinOpPred_match<LHS, RHS, is_logical_shift_op>(L, R); | ||||||||||||||||
1323 | } | ||||||||||||||||
1324 | |||||||||||||||||
1325 | /// Matches bitwise logic operations. | ||||||||||||||||
1326 | template <typename LHS, typename RHS> | ||||||||||||||||
1327 | inline BinOpPred_match<LHS, RHS, is_bitwiselogic_op> | ||||||||||||||||
1328 | m_BitwiseLogic(const LHS &L, const RHS &R) { | ||||||||||||||||
1329 | return BinOpPred_match<LHS, RHS, is_bitwiselogic_op>(L, R); | ||||||||||||||||
1330 | } | ||||||||||||||||
1331 | |||||||||||||||||
1332 | /// Matches integer division operations. | ||||||||||||||||
1333 | template <typename LHS, typename RHS> | ||||||||||||||||
1334 | inline BinOpPred_match<LHS, RHS, is_idiv_op> m_IDiv(const LHS &L, | ||||||||||||||||
1335 | const RHS &R) { | ||||||||||||||||
1336 | return BinOpPred_match<LHS, RHS, is_idiv_op>(L, R); | ||||||||||||||||
1337 | } | ||||||||||||||||
1338 | |||||||||||||||||
1339 | /// Matches integer remainder operations. | ||||||||||||||||
1340 | template <typename LHS, typename RHS> | ||||||||||||||||
1341 | inline BinOpPred_match<LHS, RHS, is_irem_op> m_IRem(const LHS &L, | ||||||||||||||||
1342 | const RHS &R) { | ||||||||||||||||
1343 | return BinOpPred_match<LHS, RHS, is_irem_op>(L, R); | ||||||||||||||||
1344 | } | ||||||||||||||||
1345 | |||||||||||||||||
1346 | //===----------------------------------------------------------------------===// | ||||||||||||||||
1347 | // Class that matches exact binary ops. | ||||||||||||||||
1348 | // | ||||||||||||||||
1349 | template <typename SubPattern_t> struct Exact_match { | ||||||||||||||||
1350 | SubPattern_t SubPattern; | ||||||||||||||||
1351 | |||||||||||||||||
1352 | Exact_match(const SubPattern_t &SP) : SubPattern(SP) {} | ||||||||||||||||
1353 | |||||||||||||||||
1354 | template <typename OpTy> bool match(OpTy *V) { | ||||||||||||||||
1355 | if (auto *PEO = dyn_cast<PossiblyExactOperator>(V)) | ||||||||||||||||
1356 | return PEO->isExact() && SubPattern.match(V); | ||||||||||||||||
1357 | return false; | ||||||||||||||||
1358 | } | ||||||||||||||||
1359 | }; | ||||||||||||||||
1360 | |||||||||||||||||
1361 | template <typename T> inline Exact_match<T> m_Exact(const T &SubPattern) { | ||||||||||||||||
1362 | return SubPattern; | ||||||||||||||||
1363 | } | ||||||||||||||||
1364 | |||||||||||||||||
1365 | //===----------------------------------------------------------------------===// | ||||||||||||||||
1366 | // Matchers for CmpInst classes | ||||||||||||||||
1367 | // | ||||||||||||||||
1368 | |||||||||||||||||
1369 | template <typename LHS_t, typename RHS_t, typename Class, typename PredicateTy, | ||||||||||||||||
1370 | bool Commutable = false> | ||||||||||||||||
1371 | struct CmpClass_match { | ||||||||||||||||
1372 | PredicateTy &Predicate; | ||||||||||||||||
1373 | LHS_t L; | ||||||||||||||||
1374 | RHS_t R; | ||||||||||||||||
1375 | |||||||||||||||||
1376 | // The evaluation order is always stable, regardless of Commutability. | ||||||||||||||||
1377 | // The LHS is always matched first. | ||||||||||||||||
1378 | CmpClass_match(PredicateTy &Pred, const LHS_t &LHS, const RHS_t &RHS) | ||||||||||||||||
1379 | : Predicate(Pred), L(LHS), R(RHS) {} | ||||||||||||||||
1380 | |||||||||||||||||
1381 | template <typename OpTy> bool match(OpTy *V) { | ||||||||||||||||
1382 | if (auto *I = dyn_cast<Class>(V)) { | ||||||||||||||||
1383 | if (L.match(I->getOperand(0)) && R.match(I->getOperand(1))) { | ||||||||||||||||
1384 | Predicate = I->getPredicate(); | ||||||||||||||||
1385 | return true; | ||||||||||||||||
1386 | } else if (Commutable && L.match(I->getOperand(1)) && | ||||||||||||||||
1387 | R.match(I->getOperand(0))) { | ||||||||||||||||
1388 | Predicate = I->getSwappedPredicate(); | ||||||||||||||||
1389 | return true; | ||||||||||||||||
1390 | } | ||||||||||||||||
1391 | } | ||||||||||||||||
1392 | return false; | ||||||||||||||||
1393 | } | ||||||||||||||||
1394 | }; | ||||||||||||||||
1395 | |||||||||||||||||
1396 | template <typename LHS, typename RHS> | ||||||||||||||||
1397 | inline CmpClass_match<LHS, RHS, CmpInst, CmpInst::Predicate> | ||||||||||||||||
1398 | m_Cmp(CmpInst::Predicate &Pred, const LHS &L, const RHS &R) { | ||||||||||||||||
1399 | return CmpClass_match<LHS, RHS, CmpInst, CmpInst::Predicate>(Pred, L, R); | ||||||||||||||||
1400 | } | ||||||||||||||||
1401 | |||||||||||||||||
1402 | template <typename LHS, typename RHS> | ||||||||||||||||
1403 | inline CmpClass_match<LHS, RHS, ICmpInst, ICmpInst::Predicate> | ||||||||||||||||
1404 | m_ICmp(ICmpInst::Predicate &Pred, const LHS &L, const RHS &R) { | ||||||||||||||||
1405 | return CmpClass_match<LHS, RHS, ICmpInst, ICmpInst::Predicate>(Pred, L, R); | ||||||||||||||||
1406 | } | ||||||||||||||||
1407 | |||||||||||||||||
1408 | template <typename LHS, typename RHS> | ||||||||||||||||
1409 | inline CmpClass_match<LHS, RHS, FCmpInst, FCmpInst::Predicate> | ||||||||||||||||
1410 | m_FCmp(FCmpInst::Predicate &Pred, const LHS &L, const RHS &R) { | ||||||||||||||||
1411 | return CmpClass_match<LHS, RHS, FCmpInst, FCmpInst::Predicate>(Pred, L, R); | ||||||||||||||||
1412 | } | ||||||||||||||||
1413 | |||||||||||||||||
1414 | //===----------------------------------------------------------------------===// | ||||||||||||||||
1415 | // Matchers for instructions with a given opcode and number of operands. | ||||||||||||||||
1416 | // | ||||||||||||||||
1417 | |||||||||||||||||
1418 | /// Matches instructions with Opcode and three operands. | ||||||||||||||||
1419 | template <typename T0, unsigned Opcode> struct OneOps_match { | ||||||||||||||||
1420 | T0 Op1; | ||||||||||||||||
1421 | |||||||||||||||||
1422 | OneOps_match(const T0 &Op1) : Op1(Op1) {} | ||||||||||||||||
1423 | |||||||||||||||||
1424 | template <typename OpTy> bool match(OpTy *V) { | ||||||||||||||||
1425 | if (V->getValueID() == Value::InstructionVal + Opcode) { | ||||||||||||||||
1426 | auto *I = cast<Instruction>(V); | ||||||||||||||||
1427 | return Op1.match(I->getOperand(0)); | ||||||||||||||||
1428 | } | ||||||||||||||||
1429 | return false; | ||||||||||||||||
1430 | } | ||||||||||||||||
1431 | }; | ||||||||||||||||
1432 | |||||||||||||||||
1433 | /// Matches instructions with Opcode and three operands. | ||||||||||||||||
1434 | template <typename T0, typename T1, unsigned Opcode> struct TwoOps_match { | ||||||||||||||||
1435 | T0 Op1; | ||||||||||||||||
1436 | T1 Op2; | ||||||||||||||||
1437 | |||||||||||||||||
1438 | TwoOps_match(const T0 &Op1, const T1 &Op2) : Op1(Op1), Op2(Op2) {} | ||||||||||||||||
1439 | |||||||||||||||||
1440 | template <typename OpTy> bool match(OpTy *V) { | ||||||||||||||||
1441 | if (V->getValueID() == Value::InstructionVal + Opcode) { | ||||||||||||||||
1442 | auto *I = cast<Instruction>(V); | ||||||||||||||||
1443 | return Op1.match(I->getOperand(0)) && Op2.match(I->getOperand(1)); | ||||||||||||||||
1444 | } | ||||||||||||||||
1445 | return false; | ||||||||||||||||
1446 | } | ||||||||||||||||
1447 | }; | ||||||||||||||||
1448 | |||||||||||||||||
1449 | /// Matches instructions with Opcode and three operands. | ||||||||||||||||
1450 | template <typename T0, typename T1, typename T2, unsigned Opcode> | ||||||||||||||||
1451 | struct ThreeOps_match { | ||||||||||||||||
1452 | T0 Op1; | ||||||||||||||||
1453 | T1 Op2; | ||||||||||||||||
1454 | T2 Op3; | ||||||||||||||||
1455 | |||||||||||||||||
1456 | ThreeOps_match(const T0 &Op1, const T1 &Op2, const T2 &Op3) | ||||||||||||||||
1457 | : Op1(Op1), Op2(Op2), Op3(Op3) {} | ||||||||||||||||
1458 | |||||||||||||||||
1459 | template <typename OpTy> bool match(OpTy *V) { | ||||||||||||||||
1460 | if (V->getValueID() == Value::InstructionVal + Opcode) { | ||||||||||||||||
1461 | auto *I = cast<Instruction>(V); | ||||||||||||||||
1462 | return Op1.match(I->getOperand(0)) && Op2.match(I->getOperand(1)) && | ||||||||||||||||
1463 | Op3.match(I->getOperand(2)); | ||||||||||||||||
1464 | } | ||||||||||||||||
1465 | return false; | ||||||||||||||||
1466 | } | ||||||||||||||||
1467 | }; | ||||||||||||||||
1468 | |||||||||||||||||
1469 | /// Matches SelectInst. | ||||||||||||||||
1470 | template <typename Cond, typename LHS, typename RHS> | ||||||||||||||||
1471 | inline ThreeOps_match<Cond, LHS, RHS, Instruction::Select> | ||||||||||||||||
1472 | m_Select(const Cond &C, const LHS &L, const RHS &R) { | ||||||||||||||||
1473 | return ThreeOps_match<Cond, LHS, RHS, Instruction::Select>(C, L, R); | ||||||||||||||||
1474 | } | ||||||||||||||||
1475 | |||||||||||||||||
1476 | /// This matches a select of two constants, e.g.: | ||||||||||||||||
1477 | /// m_SelectCst<-1, 0>(m_Value(V)) | ||||||||||||||||
1478 | template <int64_t L, int64_t R, typename Cond> | ||||||||||||||||
1479 | inline ThreeOps_match<Cond, constantint_match<L>, constantint_match<R>, | ||||||||||||||||
1480 | Instruction::Select> | ||||||||||||||||
1481 | m_SelectCst(const Cond &C) { | ||||||||||||||||
1482 | return m_Select(C, m_ConstantInt<L>(), m_ConstantInt<R>()); | ||||||||||||||||
1483 | } | ||||||||||||||||
1484 | |||||||||||||||||
1485 | /// Matches FreezeInst. | ||||||||||||||||
1486 | template <typename OpTy> | ||||||||||||||||
1487 | inline OneOps_match<OpTy, Instruction::Freeze> m_Freeze(const OpTy &Op) { | ||||||||||||||||
1488 | return OneOps_match<OpTy, Instruction::Freeze>(Op); | ||||||||||||||||
1489 | } | ||||||||||||||||
1490 | |||||||||||||||||
1491 | /// Matches InsertElementInst. | ||||||||||||||||
1492 | template <typename Val_t, typename Elt_t, typename Idx_t> | ||||||||||||||||
1493 | inline ThreeOps_match<Val_t, Elt_t, Idx_t, Instruction::InsertElement> | ||||||||||||||||
1494 | m_InsertElt(const Val_t &Val, const Elt_t &Elt, const Idx_t &Idx) { | ||||||||||||||||
1495 | return ThreeOps_match<Val_t, Elt_t, Idx_t, Instruction::InsertElement>( | ||||||||||||||||
1496 | Val, Elt, Idx); | ||||||||||||||||
1497 | } | ||||||||||||||||
1498 | |||||||||||||||||
1499 | /// Matches ExtractElementInst. | ||||||||||||||||
1500 | template <typename Val_t, typename Idx_t> | ||||||||||||||||
1501 | inline TwoOps_match<Val_t, Idx_t, Instruction::ExtractElement> | ||||||||||||||||
1502 | m_ExtractElt(const Val_t &Val, const Idx_t &Idx) { | ||||||||||||||||
1503 | return TwoOps_match<Val_t, Idx_t, Instruction::ExtractElement>(Val, Idx); | ||||||||||||||||
1504 | } | ||||||||||||||||
1505 | |||||||||||||||||
1506 | /// Matches shuffle. | ||||||||||||||||
1507 | template <typename T0, typename T1, typename T2> struct Shuffle_match { | ||||||||||||||||
1508 | T0 Op1; | ||||||||||||||||
1509 | T1 Op2; | ||||||||||||||||
1510 | T2 Mask; | ||||||||||||||||
1511 | |||||||||||||||||
1512 | Shuffle_match(const T0 &Op1, const T1 &Op2, const T2 &Mask) | ||||||||||||||||
1513 | : Op1(Op1), Op2(Op2), Mask(Mask) {} | ||||||||||||||||
1514 | |||||||||||||||||
1515 | template <typename OpTy> bool match(OpTy *V) { | ||||||||||||||||
1516 | if (auto *I = dyn_cast<ShuffleVectorInst>(V)) { | ||||||||||||||||
1517 | return Op1.match(I->getOperand(0)) && Op2.match(I->getOperand(1)) && | ||||||||||||||||
1518 | Mask.match(I->getShuffleMask()); | ||||||||||||||||
1519 | } | ||||||||||||||||
1520 | return false; | ||||||||||||||||
1521 | } | ||||||||||||||||
1522 | }; | ||||||||||||||||
1523 | |||||||||||||||||
1524 | struct m_Mask { | ||||||||||||||||
1525 | ArrayRef<int> &MaskRef; | ||||||||||||||||
1526 | m_Mask(ArrayRef<int> &MaskRef) : MaskRef(MaskRef) {} | ||||||||||||||||
1527 | bool match(ArrayRef<int> Mask) { | ||||||||||||||||
1528 | MaskRef = Mask; | ||||||||||||||||
1529 | return true; | ||||||||||||||||
1530 | } | ||||||||||||||||
1531 | }; | ||||||||||||||||
1532 | |||||||||||||||||
1533 | struct m_ZeroMask { | ||||||||||||||||
1534 | bool match(ArrayRef<int> Mask) { | ||||||||||||||||
1535 | return all_of(Mask, [](int Elem) { return Elem == 0 || Elem == -1; }); | ||||||||||||||||
1536 | } | ||||||||||||||||
1537 | }; | ||||||||||||||||
1538 | |||||||||||||||||
1539 | struct m_SpecificMask { | ||||||||||||||||
1540 | ArrayRef<int> &MaskRef; | ||||||||||||||||
1541 | m_SpecificMask(ArrayRef<int> &MaskRef) : MaskRef(MaskRef) {} | ||||||||||||||||
1542 | bool match(ArrayRef<int> Mask) { return MaskRef == Mask; } | ||||||||||||||||
1543 | }; | ||||||||||||||||
1544 | |||||||||||||||||
1545 | struct m_SplatOrUndefMask { | ||||||||||||||||
1546 | int &SplatIndex; | ||||||||||||||||
1547 | m_SplatOrUndefMask(int &SplatIndex) : SplatIndex(SplatIndex) {} | ||||||||||||||||
1548 | bool match(ArrayRef<int> Mask) { | ||||||||||||||||
1549 | auto First = find_if(Mask, [](int Elem) { return Elem != -1; }); | ||||||||||||||||
1550 | if (First == Mask.end()) | ||||||||||||||||
1551 | return false; | ||||||||||||||||
1552 | SplatIndex = *First; | ||||||||||||||||
1553 | return all_of(Mask, | ||||||||||||||||
1554 | [First](int Elem) { return Elem == *First || Elem == -1; }); | ||||||||||||||||
1555 | } | ||||||||||||||||
1556 | }; | ||||||||||||||||
1557 | |||||||||||||||||
1558 | /// Matches ShuffleVectorInst independently of mask value. | ||||||||||||||||
1559 | template <typename V1_t, typename V2_t> | ||||||||||||||||
1560 | inline TwoOps_match<V1_t, V2_t, Instruction::ShuffleVector> | ||||||||||||||||
1561 | m_Shuffle(const V1_t &v1, const V2_t &v2) { | ||||||||||||||||
1562 | return TwoOps_match<V1_t, V2_t, Instruction::ShuffleVector>(v1, v2); | ||||||||||||||||
1563 | } | ||||||||||||||||
1564 | |||||||||||||||||
1565 | template <typename V1_t, typename V2_t, typename Mask_t> | ||||||||||||||||
1566 | inline Shuffle_match<V1_t, V2_t, Mask_t> | ||||||||||||||||
1567 | m_Shuffle(const V1_t &v1, const V2_t &v2, const Mask_t &mask) { | ||||||||||||||||
1568 | return Shuffle_match<V1_t, V2_t, Mask_t>(v1, v2, mask); | ||||||||||||||||
1569 | } | ||||||||||||||||
1570 | |||||||||||||||||
1571 | /// Matches LoadInst. | ||||||||||||||||
1572 | template <typename OpTy> | ||||||||||||||||
1573 | inline OneOps_match<OpTy, Instruction::Load> m_Load(const OpTy &Op) { | ||||||||||||||||
1574 | return OneOps_match<OpTy, Instruction::Load>(Op); | ||||||||||||||||
1575 | } | ||||||||||||||||
1576 | |||||||||||||||||
1577 | /// Matches StoreInst. | ||||||||||||||||
1578 | template <typename ValueOpTy, typename PointerOpTy> | ||||||||||||||||
1579 | inline TwoOps_match<ValueOpTy, PointerOpTy, Instruction::Store> | ||||||||||||||||
1580 | m_Store(const ValueOpTy &ValueOp, const PointerOpTy &PointerOp) { | ||||||||||||||||
1581 | return TwoOps_match<ValueOpTy, PointerOpTy, Instruction::Store>(ValueOp, | ||||||||||||||||
1582 | PointerOp); | ||||||||||||||||
1583 | } | ||||||||||||||||
1584 | |||||||||||||||||
1585 | //===----------------------------------------------------------------------===// | ||||||||||||||||
1586 | // Matchers for CastInst classes | ||||||||||||||||
1587 | // | ||||||||||||||||
1588 | |||||||||||||||||
1589 | template <typename Op_t, unsigned Opcode> struct CastClass_match { | ||||||||||||||||
1590 | Op_t Op; | ||||||||||||||||
1591 | |||||||||||||||||
1592 | CastClass_match(const Op_t &OpMatch) : Op(OpMatch) {} | ||||||||||||||||
1593 | |||||||||||||||||
1594 | template <typename OpTy> bool match(OpTy *V) { | ||||||||||||||||
1595 | if (auto *O = dyn_cast<Operator>(V)) | ||||||||||||||||
1596 | return O->getOpcode() == Opcode && Op.match(O->getOperand(0)); | ||||||||||||||||
1597 | return false; | ||||||||||||||||
1598 | } | ||||||||||||||||
1599 | }; | ||||||||||||||||
1600 | |||||||||||||||||
1601 | /// Matches BitCast. | ||||||||||||||||
1602 | template <typename OpTy> | ||||||||||||||||
1603 | inline CastClass_match<OpTy, Instruction::BitCast> m_BitCast(const OpTy &Op) { | ||||||||||||||||
1604 | return CastClass_match<OpTy, Instruction::BitCast>(Op); | ||||||||||||||||
1605 | } | ||||||||||||||||
1606 | |||||||||||||||||
1607 | /// Matches PtrToInt. | ||||||||||||||||
1608 | template <typename OpTy> | ||||||||||||||||
1609 | inline CastClass_match<OpTy, Instruction::PtrToInt> m_PtrToInt(const OpTy &Op) { | ||||||||||||||||
1610 | return CastClass_match<OpTy, Instruction::PtrToInt>(Op); | ||||||||||||||||
1611 | } | ||||||||||||||||
1612 | |||||||||||||||||
1613 | /// Matches IntToPtr. | ||||||||||||||||
1614 | template <typename OpTy> | ||||||||||||||||
1615 | inline CastClass_match<OpTy, Instruction::IntToPtr> m_IntToPtr(const OpTy &Op) { | ||||||||||||||||
1616 | return CastClass_match<OpTy, Instruction::IntToPtr>(Op); | ||||||||||||||||
1617 | } | ||||||||||||||||
1618 | |||||||||||||||||
1619 | /// Matches Trunc. | ||||||||||||||||
1620 | template <typename OpTy> | ||||||||||||||||
1621 | inline CastClass_match<OpTy, Instruction::Trunc> m_Trunc(const OpTy &Op) { | ||||||||||||||||
1622 | return CastClass_match<OpTy, Instruction::Trunc>(Op); | ||||||||||||||||
1623 | } | ||||||||||||||||
1624 | |||||||||||||||||
1625 | template <typename OpTy> | ||||||||||||||||
1626 | inline match_combine_or<CastClass_match<OpTy, Instruction::Trunc>, OpTy> | ||||||||||||||||
1627 | m_TruncOrSelf(const OpTy &Op) { | ||||||||||||||||
1628 | return m_CombineOr(m_Trunc(Op), Op); | ||||||||||||||||
1629 | } | ||||||||||||||||
1630 | |||||||||||||||||
1631 | /// Matches SExt. | ||||||||||||||||
1632 | template <typename OpTy> | ||||||||||||||||
1633 | inline CastClass_match<OpTy, Instruction::SExt> m_SExt(const OpTy &Op) { | ||||||||||||||||
1634 | return CastClass_match<OpTy, Instruction::SExt>(Op); | ||||||||||||||||
1635 | } | ||||||||||||||||
1636 | |||||||||||||||||
1637 | /// Matches ZExt. | ||||||||||||||||
1638 | template <typename OpTy> | ||||||||||||||||
1639 | inline CastClass_match<OpTy, Instruction::ZExt> m_ZExt(const OpTy &Op) { | ||||||||||||||||
1640 | return CastClass_match<OpTy, Instruction::ZExt>(Op); | ||||||||||||||||
1641 | } | ||||||||||||||||
1642 | |||||||||||||||||
1643 | template <typename OpTy> | ||||||||||||||||
1644 | inline match_combine_or<CastClass_match<OpTy, Instruction::ZExt>, OpTy> | ||||||||||||||||
1645 | m_ZExtOrSelf(const OpTy &Op) { | ||||||||||||||||
1646 | return m_CombineOr(m_ZExt(Op), Op); | ||||||||||||||||
1647 | } | ||||||||||||||||
1648 | |||||||||||||||||
1649 | template <typename OpTy> | ||||||||||||||||
1650 | inline match_combine_or<CastClass_match<OpTy, Instruction::SExt>, OpTy> | ||||||||||||||||
1651 | m_SExtOrSelf(const OpTy &Op) { | ||||||||||||||||
1652 | return m_CombineOr(m_SExt(Op), Op); | ||||||||||||||||
1653 | } | ||||||||||||||||
1654 | |||||||||||||||||
1655 | template <typename OpTy> | ||||||||||||||||
1656 | inline match_combine_or<CastClass_match<OpTy, Instruction::ZExt>, | ||||||||||||||||
1657 | CastClass_match<OpTy, Instruction::SExt>> | ||||||||||||||||
1658 | m_ZExtOrSExt(const OpTy &Op) { | ||||||||||||||||
1659 | return m_CombineOr(m_ZExt(Op), m_SExt(Op)); | ||||||||||||||||
1660 | } | ||||||||||||||||
1661 | |||||||||||||||||
1662 | template <typename OpTy> | ||||||||||||||||
1663 | inline match_combine_or< | ||||||||||||||||
1664 | match_combine_or<CastClass_match<OpTy, Instruction::ZExt>, | ||||||||||||||||
1665 | CastClass_match<OpTy, Instruction::SExt>>, | ||||||||||||||||
1666 | OpTy> | ||||||||||||||||
1667 | m_ZExtOrSExtOrSelf(const OpTy &Op) { | ||||||||||||||||
1668 | return m_CombineOr(m_ZExtOrSExt(Op), Op); | ||||||||||||||||
1669 | } | ||||||||||||||||
1670 | |||||||||||||||||
1671 | template <typename OpTy> | ||||||||||||||||
1672 | inline CastClass_match<OpTy, Instruction::UIToFP> m_UIToFP(const OpTy &Op) { | ||||||||||||||||
1673 | return CastClass_match<OpTy, Instruction::UIToFP>(Op); | ||||||||||||||||
1674 | } | ||||||||||||||||
1675 | |||||||||||||||||
1676 | template <typename OpTy> | ||||||||||||||||
1677 | inline CastClass_match<OpTy, Instruction::SIToFP> m_SIToFP(const OpTy &Op) { | ||||||||||||||||
1678 | return CastClass_match<OpTy, Instruction::SIToFP>(Op); | ||||||||||||||||
1679 | } | ||||||||||||||||
1680 | |||||||||||||||||
1681 | template <typename OpTy> | ||||||||||||||||
1682 | inline CastClass_match<OpTy, Instruction::FPToUI> m_FPToUI(const OpTy &Op) { | ||||||||||||||||
1683 | return CastClass_match<OpTy, Instruction::FPToUI>(Op); | ||||||||||||||||
1684 | } | ||||||||||||||||
1685 | |||||||||||||||||
1686 | template <typename OpTy> | ||||||||||||||||
1687 | inline CastClass_match<OpTy, Instruction::FPToSI> m_FPToSI(const OpTy &Op) { | ||||||||||||||||
1688 | return CastClass_match<OpTy, Instruction::FPToSI>(Op); | ||||||||||||||||
1689 | } | ||||||||||||||||
1690 | |||||||||||||||||
1691 | template <typename OpTy> | ||||||||||||||||
1692 | inline CastClass_match<OpTy, Instruction::FPTrunc> m_FPTrunc(const OpTy &Op) { | ||||||||||||||||
1693 | return CastClass_match<OpTy, Instruction::FPTrunc>(Op); | ||||||||||||||||
1694 | } | ||||||||||||||||
1695 | |||||||||||||||||
1696 | template <typename OpTy> | ||||||||||||||||
1697 | inline CastClass_match<OpTy, Instruction::FPExt> m_FPExt(const OpTy &Op) { | ||||||||||||||||
1698 | return CastClass_match<OpTy, Instruction::FPExt>(Op); | ||||||||||||||||
1699 | } | ||||||||||||||||
1700 | |||||||||||||||||
1701 | //===----------------------------------------------------------------------===// | ||||||||||||||||
1702 | // Matchers for control flow. | ||||||||||||||||
1703 | // | ||||||||||||||||
1704 | |||||||||||||||||
1705 | struct br_match { | ||||||||||||||||
1706 | BasicBlock *&Succ; | ||||||||||||||||
1707 | |||||||||||||||||
1708 | br_match(BasicBlock *&Succ) : Succ(Succ) {} | ||||||||||||||||
1709 | |||||||||||||||||
1710 | template <typename OpTy> bool match(OpTy *V) { | ||||||||||||||||
1711 | if (auto *BI = dyn_cast<BranchInst>(V)) | ||||||||||||||||
1712 | if (BI->isUnconditional()) { | ||||||||||||||||
1713 | Succ = BI->getSuccessor(0); | ||||||||||||||||
1714 | return true; | ||||||||||||||||
1715 | } | ||||||||||||||||
1716 | return false; | ||||||||||||||||
1717 | } | ||||||||||||||||
1718 | }; | ||||||||||||||||
1719 | |||||||||||||||||
1720 | inline br_match m_UnconditionalBr(BasicBlock *&Succ) { return br_match(Succ); } | ||||||||||||||||
1721 | |||||||||||||||||
1722 | template <typename Cond_t, typename TrueBlock_t, typename FalseBlock_t> | ||||||||||||||||
1723 | struct brc_match { | ||||||||||||||||
1724 | Cond_t Cond; | ||||||||||||||||
1725 | TrueBlock_t T; | ||||||||||||||||
1726 | FalseBlock_t F; | ||||||||||||||||
1727 | |||||||||||||||||
1728 | brc_match(const Cond_t &C, const TrueBlock_t &t, const FalseBlock_t &f) | ||||||||||||||||
1729 | : Cond(C), T(t), F(f) {} | ||||||||||||||||
1730 | |||||||||||||||||
1731 | template <typename OpTy> bool match(OpTy *V) { | ||||||||||||||||
1732 | if (auto *BI = dyn_cast<BranchInst>(V)) | ||||||||||||||||
1733 | if (BI->isConditional() && Cond.match(BI->getCondition())) | ||||||||||||||||
1734 | return T.match(BI->getSuccessor(0)) && F.match(BI->getSuccessor(1)); | ||||||||||||||||
1735 | return false; | ||||||||||||||||
1736 | } | ||||||||||||||||
1737 | }; | ||||||||||||||||
1738 | |||||||||||||||||
1739 | template <typename Cond_t> | ||||||||||||||||
1740 | inline brc_match<Cond_t, bind_ty<BasicBlock>, bind_ty<BasicBlock>> | ||||||||||||||||
1741 | m_Br(const Cond_t &C, BasicBlock *&T, BasicBlock *&F) { | ||||||||||||||||
1742 | return brc_match<Cond_t, bind_ty<BasicBlock>, bind_ty<BasicBlock>>( | ||||||||||||||||
1743 | C, m_BasicBlock(T), m_BasicBlock(F)); | ||||||||||||||||
1744 | } | ||||||||||||||||
1745 | |||||||||||||||||
1746 | template <typename Cond_t, typename TrueBlock_t, typename FalseBlock_t> | ||||||||||||||||
1747 | inline brc_match<Cond_t, TrueBlock_t, FalseBlock_t> | ||||||||||||||||
1748 | m_Br(const Cond_t &C, const TrueBlock_t &T, const FalseBlock_t &F) { | ||||||||||||||||
1749 | return brc_match<Cond_t, TrueBlock_t, FalseBlock_t>(C, T, F); | ||||||||||||||||
1750 | } | ||||||||||||||||
1751 | |||||||||||||||||
1752 | //===----------------------------------------------------------------------===// | ||||||||||||||||
1753 | // Matchers for max/min idioms, eg: "select (sgt x, y), x, y" -> smax(x,y). | ||||||||||||||||
1754 | // | ||||||||||||||||
1755 | |||||||||||||||||
1756 | template <typename CmpInst_t, typename LHS_t, typename RHS_t, typename Pred_t, | ||||||||||||||||
1757 | bool Commutable = false> | ||||||||||||||||
1758 | struct MaxMin_match { | ||||||||||||||||
1759 | using PredType = Pred_t; | ||||||||||||||||
1760 | LHS_t L; | ||||||||||||||||
1761 | RHS_t R; | ||||||||||||||||
1762 | |||||||||||||||||
1763 | // The evaluation order is always stable, regardless of Commutability. | ||||||||||||||||
1764 | // The LHS is always matched first. | ||||||||||||||||
1765 | MaxMin_match(const LHS_t &LHS, const RHS_t &RHS) : L(LHS), R(RHS) {} | ||||||||||||||||
1766 | |||||||||||||||||
1767 | template <typename OpTy> bool match(OpTy *V) { | ||||||||||||||||
1768 | if (auto *II = dyn_cast<IntrinsicInst>(V)) { | ||||||||||||||||
1769 | Intrinsic::ID IID = II->getIntrinsicID(); | ||||||||||||||||
1770 | if ((IID == Intrinsic::smax && Pred_t::match(ICmpInst::ICMP_SGT)) || | ||||||||||||||||
1771 | (IID == Intrinsic::smin && Pred_t::match(ICmpInst::ICMP_SLT)) || | ||||||||||||||||
1772 | (IID == Intrinsic::umax && Pred_t::match(ICmpInst::ICMP_UGT)) || | ||||||||||||||||
1773 | (IID == Intrinsic::umin && Pred_t::match(ICmpInst::ICMP_ULT))) { | ||||||||||||||||
1774 | Value *LHS = II->getOperand(0), *RHS = II->getOperand(1); | ||||||||||||||||
1775 | return (L.match(LHS) && R.match(RHS)) || | ||||||||||||||||
1776 | (Commutable && L.match(RHS) && R.match(LHS)); | ||||||||||||||||
1777 | } | ||||||||||||||||
1778 | } | ||||||||||||||||
1779 | // Look for "(x pred y) ? x : y" or "(x pred y) ? y : x". | ||||||||||||||||
1780 | auto *SI = dyn_cast<SelectInst>(V); | ||||||||||||||||
1781 | if (!SI) | ||||||||||||||||
1782 | return false; | ||||||||||||||||
1783 | auto *Cmp = dyn_cast<CmpInst_t>(SI->getCondition()); | ||||||||||||||||
1784 | if (!Cmp) | ||||||||||||||||
1785 | return false; | ||||||||||||||||
1786 | // At this point we have a select conditioned on a comparison. Check that | ||||||||||||||||
1787 | // it is the values returned by the select that are being compared. | ||||||||||||||||
1788 | auto *TrueVal = SI->getTrueValue(); | ||||||||||||||||
1789 | auto *FalseVal = SI->getFalseValue(); | ||||||||||||||||
1790 | auto *LHS = Cmp->getOperand(0); | ||||||||||||||||
1791 | auto *RHS = Cmp->getOperand(1); | ||||||||||||||||
1792 | if ((TrueVal != LHS || FalseVal != RHS) && | ||||||||||||||||
1793 | (TrueVal != RHS || FalseVal != LHS)) | ||||||||||||||||
1794 | return false; | ||||||||||||||||
1795 | typename CmpInst_t::Predicate Pred = | ||||||||||||||||
1796 | LHS == TrueVal ? Cmp->getPredicate() : Cmp->getInversePredicate(); | ||||||||||||||||
1797 | // Does "(x pred y) ? x : y" represent the desired max/min operation? | ||||||||||||||||
1798 | if (!Pred_t::match(Pred)) | ||||||||||||||||
1799 | return false; | ||||||||||||||||
1800 | // It does! Bind the operands. | ||||||||||||||||
1801 | return (L.match(LHS) && R.match(RHS)) || | ||||||||||||||||
1802 | (Commutable && L.match(RHS) && R.match(LHS)); | ||||||||||||||||
1803 | } | ||||||||||||||||
1804 | }; | ||||||||||||||||
1805 | |||||||||||||||||
1806 | /// Helper class for identifying signed max predicates. | ||||||||||||||||
1807 | struct smax_pred_ty { | ||||||||||||||||
1808 | static bool match(ICmpInst::Predicate Pred) { | ||||||||||||||||
1809 | return Pred == CmpInst::ICMP_SGT || Pred == CmpInst::ICMP_SGE; | ||||||||||||||||
1810 | } | ||||||||||||||||
1811 | }; | ||||||||||||||||
1812 | |||||||||||||||||
1813 | /// Helper class for identifying signed min predicates. | ||||||||||||||||
1814 | struct smin_pred_ty { | ||||||||||||||||
1815 | static bool match(ICmpInst::Predicate Pred) { | ||||||||||||||||
1816 | return Pred == CmpInst::ICMP_SLT || Pred == CmpInst::ICMP_SLE; | ||||||||||||||||
1817 | } | ||||||||||||||||
1818 | }; | ||||||||||||||||
1819 | |||||||||||||||||
1820 | /// Helper class for identifying unsigned max predicates. | ||||||||||||||||
1821 | struct umax_pred_ty { | ||||||||||||||||
1822 | static bool match(ICmpInst::Predicate Pred) { | ||||||||||||||||
1823 | return Pred == CmpInst::ICMP_UGT || Pred == CmpInst::ICMP_UGE; | ||||||||||||||||
1824 | } | ||||||||||||||||
1825 | }; | ||||||||||||||||
1826 | |||||||||||||||||
1827 | /// Helper class for identifying unsigned min predicates. | ||||||||||||||||
1828 | struct umin_pred_ty { | ||||||||||||||||
1829 | static bool match(ICmpInst::Predicate Pred) { | ||||||||||||||||
1830 | return Pred == CmpInst::ICMP_ULT || Pred == CmpInst::ICMP_ULE; | ||||||||||||||||
1831 | } | ||||||||||||||||
1832 | }; | ||||||||||||||||
1833 | |||||||||||||||||
1834 | /// Helper class for identifying ordered max predicates. | ||||||||||||||||
1835 | struct ofmax_pred_ty { | ||||||||||||||||
1836 | static bool match(FCmpInst::Predicate Pred) { | ||||||||||||||||
1837 | return Pred == CmpInst::FCMP_OGT || Pred == CmpInst::FCMP_OGE; | ||||||||||||||||
1838 | } | ||||||||||||||||
1839 | }; | ||||||||||||||||
1840 | |||||||||||||||||
1841 | /// Helper class for identifying ordered min predicates. | ||||||||||||||||
1842 | struct ofmin_pred_ty { | ||||||||||||||||
1843 | static bool match(FCmpInst::Predicate Pred) { | ||||||||||||||||
1844 | return Pred == CmpInst::FCMP_OLT || Pred == CmpInst::FCMP_OLE; | ||||||||||||||||
1845 | } | ||||||||||||||||
1846 | }; | ||||||||||||||||
1847 | |||||||||||||||||
1848 | /// Helper class for identifying unordered max predicates. | ||||||||||||||||
1849 | struct ufmax_pred_ty { | ||||||||||||||||
1850 | static bool match(FCmpInst::Predicate Pred) { | ||||||||||||||||
1851 | return Pred == CmpInst::FCMP_UGT || Pred == CmpInst::FCMP_UGE; | ||||||||||||||||
1852 | } | ||||||||||||||||
1853 | }; | ||||||||||||||||
1854 | |||||||||||||||||
1855 | /// Helper class for identifying unordered min predicates. | ||||||||||||||||
1856 | struct ufmin_pred_ty { | ||||||||||||||||
1857 | static bool match(FCmpInst::Predicate Pred) { | ||||||||||||||||
1858 | return Pred == CmpInst::FCMP_ULT || Pred == CmpInst::FCMP_ULE; | ||||||||||||||||
1859 | } | ||||||||||||||||
1860 | }; | ||||||||||||||||
1861 | |||||||||||||||||
1862 | template <typename LHS, typename RHS> | ||||||||||||||||
1863 | inline MaxMin_match<ICmpInst, LHS, RHS, smax_pred_ty> m_SMax(const LHS &L, | ||||||||||||||||
1864 | const RHS &R) { | ||||||||||||||||
1865 | return MaxMin_match<ICmpInst, LHS, RHS, smax_pred_ty>(L, R); | ||||||||||||||||
1866 | } | ||||||||||||||||
1867 | |||||||||||||||||
1868 | template <typename LHS, typename RHS> | ||||||||||||||||
1869 | inline MaxMin_match<ICmpInst, LHS, RHS, smin_pred_ty> m_SMin(const LHS &L, | ||||||||||||||||
1870 | const RHS &R) { | ||||||||||||||||
1871 | return MaxMin_match<ICmpInst, LHS, RHS, smin_pred_ty>(L, R); | ||||||||||||||||
1872 | } | ||||||||||||||||
1873 | |||||||||||||||||
1874 | template <typename LHS, typename RHS> | ||||||||||||||||
1875 | inline MaxMin_match<ICmpInst, LHS, RHS, umax_pred_ty> m_UMax(const LHS &L, | ||||||||||||||||
1876 | const RHS &R) { | ||||||||||||||||
1877 | return MaxMin_match<ICmpInst, LHS, RHS, umax_pred_ty>(L, R); | ||||||||||||||||
1878 | } | ||||||||||||||||
1879 | |||||||||||||||||
1880 | template <typename LHS, typename RHS> | ||||||||||||||||
1881 | inline MaxMin_match<ICmpInst, LHS, RHS, umin_pred_ty> m_UMin(const LHS &L, | ||||||||||||||||
1882 | const RHS &R) { | ||||||||||||||||
1883 | return MaxMin_match<ICmpInst, LHS, RHS, umin_pred_ty>(L, R); | ||||||||||||||||
1884 | } | ||||||||||||||||
1885 | |||||||||||||||||
1886 | template <typename LHS, typename RHS> | ||||||||||||||||
1887 | inline match_combine_or< | ||||||||||||||||
1888 | match_combine_or<MaxMin_match<ICmpInst, LHS, RHS, smax_pred_ty>, | ||||||||||||||||
1889 | MaxMin_match<ICmpInst, LHS, RHS, smin_pred_ty>>, | ||||||||||||||||
1890 | match_combine_or<MaxMin_match<ICmpInst, LHS, RHS, umax_pred_ty>, | ||||||||||||||||
1891 | MaxMin_match<ICmpInst, LHS, RHS, umin_pred_ty>>> | ||||||||||||||||
1892 | m_MaxOrMin(const LHS &L, const RHS &R) { | ||||||||||||||||
1893 | return m_CombineOr(m_CombineOr(m_SMax(L, R), m_SMin(L, R)), | ||||||||||||||||
1894 | m_CombineOr(m_UMax(L, R), m_UMin(L, R))); | ||||||||||||||||
1895 | } | ||||||||||||||||
1896 | |||||||||||||||||
1897 | /// Match an 'ordered' floating point maximum function. | ||||||||||||||||
1898 | /// Floating point has one special value 'NaN'. Therefore, there is no total | ||||||||||||||||
1899 | /// order. However, if we can ignore the 'NaN' value (for example, because of a | ||||||||||||||||
1900 | /// 'no-nans-float-math' flag) a combination of a fcmp and select has 'maximum' | ||||||||||||||||
1901 | /// semantics. In the presence of 'NaN' we have to preserve the original | ||||||||||||||||
1902 | /// select(fcmp(ogt/ge, L, R), L, R) semantics matched by this predicate. | ||||||||||||||||
1903 | /// | ||||||||||||||||
1904 | /// max(L, R) iff L and R are not NaN | ||||||||||||||||
1905 | /// m_OrdFMax(L, R) = R iff L or R are NaN | ||||||||||||||||
1906 | template <typename LHS, typename RHS> | ||||||||||||||||
1907 | inline MaxMin_match<FCmpInst, LHS, RHS, ofmax_pred_ty> m_OrdFMax(const LHS &L, | ||||||||||||||||
1908 | const RHS &R) { | ||||||||||||||||
1909 | return MaxMin_match<FCmpInst, LHS, RHS, ofmax_pred_ty>(L, R); | ||||||||||||||||
1910 | } | ||||||||||||||||
1911 | |||||||||||||||||
1912 | /// Match an 'ordered' floating point minimum function. | ||||||||||||||||
1913 | /// Floating point has one special value 'NaN'. Therefore, there is no total | ||||||||||||||||
1914 | /// order. However, if we can ignore the 'NaN' value (for example, because of a | ||||||||||||||||
1915 | /// 'no-nans-float-math' flag) a combination of a fcmp and select has 'minimum' | ||||||||||||||||
1916 | /// semantics. In the presence of 'NaN' we have to preserve the original | ||||||||||||||||
1917 | /// select(fcmp(olt/le, L, R), L, R) semantics matched by this predicate. | ||||||||||||||||
1918 | /// | ||||||||||||||||
1919 | /// min(L, R) iff L and R are not NaN | ||||||||||||||||
1920 | /// m_OrdFMin(L, R) = R iff L or R are NaN | ||||||||||||||||
1921 | template <typename LHS, typename RHS> | ||||||||||||||||
1922 | inline MaxMin_match<FCmpInst, LHS, RHS, ofmin_pred_ty> m_OrdFMin(const LHS &L, | ||||||||||||||||
1923 | const RHS &R) { | ||||||||||||||||
1924 | return MaxMin_match<FCmpInst, LHS, RHS, ofmin_pred_ty>(L, R); | ||||||||||||||||
1925 | } | ||||||||||||||||
1926 | |||||||||||||||||
1927 | /// Match an 'unordered' floating point maximum function. | ||||||||||||||||
1928 | /// Floating point has one special value 'NaN'. Therefore, there is no total | ||||||||||||||||
1929 | /// order. However, if we can ignore the 'NaN' value (for example, because of a | ||||||||||||||||
1930 | /// 'no-nans-float-math' flag) a combination of a fcmp and select has 'maximum' | ||||||||||||||||
1931 | /// semantics. In the presence of 'NaN' we have to preserve the original | ||||||||||||||||
1932 | /// select(fcmp(ugt/ge, L, R), L, R) semantics matched by this predicate. | ||||||||||||||||
1933 | /// | ||||||||||||||||
1934 | /// max(L, R) iff L and R are not NaN | ||||||||||||||||
1935 | /// m_UnordFMax(L, R) = L iff L or R are NaN | ||||||||||||||||
1936 | template <typename LHS, typename RHS> | ||||||||||||||||
1937 | inline MaxMin_match<FCmpInst, LHS, RHS, ufmax_pred_ty> | ||||||||||||||||
1938 | m_UnordFMax(const LHS &L, const RHS &R) { | ||||||||||||||||
1939 | return MaxMin_match<FCmpInst, LHS, RHS, ufmax_pred_ty>(L, R); | ||||||||||||||||
1940 | } | ||||||||||||||||
1941 | |||||||||||||||||
1942 | /// Match an 'unordered' floating point minimum function. | ||||||||||||||||
1943 | /// Floating point has one special value 'NaN'. Therefore, there is no total | ||||||||||||||||
1944 | /// order. However, if we can ignore the 'NaN' value (for example, because of a | ||||||||||||||||
1945 | /// 'no-nans-float-math' flag) a combination of a fcmp and select has 'minimum' | ||||||||||||||||
1946 | /// semantics. In the presence of 'NaN' we have to preserve the original | ||||||||||||||||
1947 | /// select(fcmp(ult/le, L, R), L, R) semantics matched by this predicate. | ||||||||||||||||
1948 | /// | ||||||||||||||||
1949 | /// min(L, R) iff L and R are not NaN | ||||||||||||||||
1950 | /// m_UnordFMin(L, R) = L iff L or R are NaN | ||||||||||||||||
1951 | template <typename LHS, typename RHS> | ||||||||||||||||
1952 | inline MaxMin_match<FCmpInst, LHS, RHS, ufmin_pred_ty> | ||||||||||||||||
1953 | m_UnordFMin(const LHS &L, const RHS &R) { | ||||||||||||||||
1954 | return MaxMin_match<FCmpInst, LHS, RHS, ufmin_pred_ty>(L, R); | ||||||||||||||||
1955 | } | ||||||||||||||||
1956 | |||||||||||||||||
1957 | //===----------------------------------------------------------------------===// | ||||||||||||||||
1958 | // Matchers for overflow check patterns: e.g. (a + b) u< a, (a ^ -1) <u b | ||||||||||||||||
1959 | // Note that S might be matched to other instructions than AddInst. | ||||||||||||||||
1960 | // | ||||||||||||||||
1961 | |||||||||||||||||
1962 | template <typename LHS_t, typename RHS_t, typename Sum_t> | ||||||||||||||||
1963 | struct UAddWithOverflow_match { | ||||||||||||||||
1964 | LHS_t L; | ||||||||||||||||
1965 | RHS_t R; | ||||||||||||||||
1966 | Sum_t S; | ||||||||||||||||
1967 | |||||||||||||||||
1968 | UAddWithOverflow_match(const LHS_t &L, const RHS_t &R, const Sum_t &S) | ||||||||||||||||
1969 | : L(L), R(R), S(S) {} | ||||||||||||||||
1970 | |||||||||||||||||
1971 | template <typename OpTy> bool match(OpTy *V) { | ||||||||||||||||
1972 | Value *ICmpLHS, *ICmpRHS; | ||||||||||||||||
1973 | ICmpInst::Predicate Pred; | ||||||||||||||||
1974 | if (!m_ICmp(Pred, m_Value(ICmpLHS), m_Value(ICmpRHS)).match(V)) | ||||||||||||||||
1975 | return false; | ||||||||||||||||
1976 | |||||||||||||||||
1977 | Value *AddLHS, *AddRHS; | ||||||||||||||||
1978 | auto AddExpr = m_Add(m_Value(AddLHS), m_Value(AddRHS)); | ||||||||||||||||
1979 | |||||||||||||||||
1980 | // (a + b) u< a, (a + b) u< b | ||||||||||||||||
1981 | if (Pred == ICmpInst::ICMP_ULT) | ||||||||||||||||
1982 | if (AddExpr.match(ICmpLHS) && (ICmpRHS == AddLHS || ICmpRHS == AddRHS)) | ||||||||||||||||
1983 | return L.match(AddLHS) && R.match(AddRHS) && S.match(ICmpLHS); | ||||||||||||||||
1984 | |||||||||||||||||
1985 | // a >u (a + b), b >u (a + b) | ||||||||||||||||
1986 | if (Pred == ICmpInst::ICMP_UGT) | ||||||||||||||||
1987 | if (AddExpr.match(ICmpRHS) && (ICmpLHS == AddLHS || ICmpLHS == AddRHS)) | ||||||||||||||||
1988 | return L.match(AddLHS) && R.match(AddRHS) && S.match(ICmpRHS); | ||||||||||||||||
1989 | |||||||||||||||||
1990 | Value *Op1; | ||||||||||||||||
1991 | auto XorExpr = m_OneUse(m_Xor(m_Value(Op1), m_AllOnes())); | ||||||||||||||||
1992 | // (a ^ -1) <u b | ||||||||||||||||
1993 | if (Pred == ICmpInst::ICMP_ULT) { | ||||||||||||||||
1994 | if (XorExpr.match(ICmpLHS)) | ||||||||||||||||
1995 | return L.match(Op1) && R.match(ICmpRHS) && S.match(ICmpLHS); | ||||||||||||||||
1996 | } | ||||||||||||||||
1997 | // b > u (a ^ -1) | ||||||||||||||||
1998 | if (Pred == ICmpInst::ICMP_UGT) { | ||||||||||||||||
1999 | if (XorExpr.match(ICmpRHS)) | ||||||||||||||||
2000 | return L.match(Op1) && R.match(ICmpLHS) && S.match(ICmpRHS); | ||||||||||||||||
2001 | } | ||||||||||||||||
2002 | |||||||||||||||||
2003 | // Match special-case for increment-by-1. | ||||||||||||||||
2004 | if (Pred == ICmpInst::ICMP_EQ) { | ||||||||||||||||
2005 | // (a + 1) == 0 | ||||||||||||||||
2006 | // (1 + a) == 0 | ||||||||||||||||
2007 | if (AddExpr.match(ICmpLHS) && m_ZeroInt().match(ICmpRHS) && | ||||||||||||||||
2008 | (m_One().match(AddLHS) || m_One().match(AddRHS))) | ||||||||||||||||
2009 | return L.match(AddLHS) && R.match(AddRHS) && S.match(ICmpLHS); | ||||||||||||||||
2010 | // 0 == (a + 1) | ||||||||||||||||
2011 | // 0 == (1 + a) | ||||||||||||||||
2012 | if (m_ZeroInt().match(ICmpLHS) && AddExpr.match(ICmpRHS) && | ||||||||||||||||
2013 | (m_One().match(AddLHS) || m_One().match(AddRHS))) | ||||||||||||||||
2014 | return L.match(AddLHS) && R.match(AddRHS) && S.match(ICmpRHS); | ||||||||||||||||
2015 | } | ||||||||||||||||
2016 | |||||||||||||||||
2017 | return false; | ||||||||||||||||
2018 | } | ||||||||||||||||
2019 | }; | ||||||||||||||||
2020 | |||||||||||||||||
2021 | /// Match an icmp instruction checking for unsigned overflow on addition. | ||||||||||||||||
2022 | /// | ||||||||||||||||
2023 | /// S is matched to the addition whose result is being checked for overflow, and | ||||||||||||||||
2024 | /// L and R are matched to the LHS and RHS of S. | ||||||||||||||||
2025 | template <typename LHS_t, typename RHS_t, typename Sum_t> | ||||||||||||||||
2026 | UAddWithOverflow_match<LHS_t, RHS_t, Sum_t> | ||||||||||||||||
2027 | m_UAddWithOverflow(const LHS_t &L, const RHS_t &R, const Sum_t &S) { | ||||||||||||||||
2028 | return UAddWithOverflow_match<LHS_t, RHS_t, Sum_t>(L, R, S); | ||||||||||||||||
2029 | } | ||||||||||||||||
2030 | |||||||||||||||||
2031 | template <typename Opnd_t> struct Argument_match { | ||||||||||||||||
2032 | unsigned OpI; | ||||||||||||||||
2033 | Opnd_t Val; | ||||||||||||||||
2034 | |||||||||||||||||
2035 | Argument_match(unsigned OpIdx, const Opnd_t &V) : OpI(OpIdx), Val(V) {} | ||||||||||||||||
2036 | |||||||||||||||||
2037 | template <typename OpTy> bool match(OpTy *V) { | ||||||||||||||||
2038 | // FIXME: Should likely be switched to use `CallBase`. | ||||||||||||||||
2039 | if (const auto *CI = dyn_cast<CallInst>(V)) | ||||||||||||||||
2040 | return Val.match(CI->getArgOperand(OpI)); | ||||||||||||||||
2041 | return false; | ||||||||||||||||
2042 | } | ||||||||||||||||
2043 | }; | ||||||||||||||||
2044 | |||||||||||||||||
2045 | /// Match an argument. | ||||||||||||||||
2046 | template <unsigned OpI, typename Opnd_t> | ||||||||||||||||
2047 | inline Argument_match<Opnd_t> m_Argument(const Opnd_t &Op) { | ||||||||||||||||
2048 | return Argument_match<Opnd_t>(OpI, Op); | ||||||||||||||||
2049 | } | ||||||||||||||||
2050 | |||||||||||||||||
2051 | /// Intrinsic matchers. | ||||||||||||||||
2052 | struct IntrinsicID_match { | ||||||||||||||||
2053 | unsigned ID; | ||||||||||||||||
2054 | |||||||||||||||||
2055 | IntrinsicID_match(Intrinsic::ID IntrID) : ID(IntrID) {} | ||||||||||||||||
2056 | |||||||||||||||||
2057 | template <typename OpTy> bool match(OpTy *V) { | ||||||||||||||||
2058 | if (const auto *CI
| ||||||||||||||||
2059 | if (const auto *F = CI->getCalledFunction()) | ||||||||||||||||
2060 | return F->getIntrinsicID() == ID; | ||||||||||||||||
2061 | return false; | ||||||||||||||||
2062 | } | ||||||||||||||||
2063 | }; | ||||||||||||||||
2064 | |||||||||||||||||
2065 | /// Intrinsic matches are combinations of ID matchers, and argument | ||||||||||||||||
2066 | /// matchers. Higher arity matcher are defined recursively in terms of and-ing | ||||||||||||||||
2067 | /// them with lower arity matchers. Here's some convenient typedefs for up to | ||||||||||||||||
2068 | /// several arguments, and more can be added as needed | ||||||||||||||||
2069 | template <typename T0 = void, typename T1 = void, typename T2 = void, | ||||||||||||||||
2070 | typename T3 = void, typename T4 = void, typename T5 = void, | ||||||||||||||||
2071 | typename T6 = void, typename T7 = void, typename T8 = void, | ||||||||||||||||
2072 | typename T9 = void, typename T10 = void> | ||||||||||||||||
2073 | struct m_Intrinsic_Ty; | ||||||||||||||||
2074 | template <typename T0> struct m_Intrinsic_Ty<T0> { | ||||||||||||||||
2075 | using Ty = match_combine_and<IntrinsicID_match, Argument_match<T0>>; | ||||||||||||||||
2076 | }; | ||||||||||||||||
2077 | template <typename T0, typename T1> struct m_Intrinsic_Ty<T0, T1> { | ||||||||||||||||
2078 | using Ty = | ||||||||||||||||
2079 | match_combine_and<typename m_Intrinsic_Ty<T0>::Ty, Argument_match<T1>>; | ||||||||||||||||
2080 | }; | ||||||||||||||||
2081 | template <typename T0, typename T1, typename T2> | ||||||||||||||||
2082 | struct m_Intrinsic_Ty<T0, T1, T2> { | ||||||||||||||||
2083 | using Ty = | ||||||||||||||||
2084 | match_combine_and<typename m_Intrinsic_Ty<T0, T1>::Ty, | ||||||||||||||||
2085 | Argument_match<T2>>; | ||||||||||||||||
2086 | }; | ||||||||||||||||
2087 | template <typename T0, typename T1, typename T2, typename T3> | ||||||||||||||||
2088 | struct m_Intrinsic_Ty<T0, T1, T2, T3> { | ||||||||||||||||
2089 | using Ty = | ||||||||||||||||
2090 | match_combine_and<typename m_Intrinsic_Ty<T0, T1, T2>::Ty, | ||||||||||||||||
2091 | Argument_match<T3>>; | ||||||||||||||||
2092 | }; | ||||||||||||||||
2093 | |||||||||||||||||
2094 | template <typename T0, typename T1, typename T2, typename T3, typename T4> | ||||||||||||||||
2095 | struct m_Intrinsic_Ty<T0, T1, T2, T3, T4> { | ||||||||||||||||
2096 | using Ty = match_combine_and<typename m_Intrinsic_Ty<T0, T1, T2, T3>::Ty, | ||||||||||||||||
2097 | Argument_match<T4>>; | ||||||||||||||||
2098 | }; | ||||||||||||||||
2099 | |||||||||||||||||
2100 | template <typename T0, typename T1, typename T2, typename T3, typename T4, typename T5> | ||||||||||||||||
2101 | struct m_Intrinsic_Ty<T0, T1, T2, T3, T4, T5> { | ||||||||||||||||
2102 | using Ty = match_combine_and<typename m_Intrinsic_Ty<T0, T1, T2, T3, T4>::Ty, | ||||||||||||||||
2103 | Argument_match<T5>>; | ||||||||||||||||
2104 | }; | ||||||||||||||||
2105 | |||||||||||||||||
2106 | /// Match intrinsic calls like this: | ||||||||||||||||
2107 | /// m_Intrinsic<Intrinsic::fabs>(m_Value(X)) | ||||||||||||||||
2108 | template <Intrinsic::ID IntrID> inline IntrinsicID_match m_Intrinsic() { | ||||||||||||||||
2109 | return IntrinsicID_match(IntrID); | ||||||||||||||||
2110 | } | ||||||||||||||||
2111 | |||||||||||||||||
2112 | /// Matches MaskedLoad Intrinsic. | ||||||||||||||||
2113 | template <typename Opnd0, typename Opnd1, typename Opnd2, typename Opnd3> | ||||||||||||||||
2114 | inline typename m_Intrinsic_Ty<Opnd0, Opnd1, Opnd2, Opnd3>::Ty | ||||||||||||||||
2115 | m_MaskedLoad(const Opnd0 &Op0, const Opnd1 &Op1, const Opnd2 &Op2, | ||||||||||||||||
2116 | const Opnd3 &Op3) { | ||||||||||||||||
2117 | return m_Intrinsic<Intrinsic::masked_load>(Op0, Op1, Op2, Op3); | ||||||||||||||||
2118 | } | ||||||||||||||||
2119 | |||||||||||||||||
2120 | template <Intrinsic::ID IntrID, typename T0> | ||||||||||||||||
2121 | inline typename m_Intrinsic_Ty<T0>::Ty m_Intrinsic(const T0 &Op0) { | ||||||||||||||||
2122 | return m_CombineAnd(m_Intrinsic<IntrID>(), m_Argument<0>(Op0)); | ||||||||||||||||
2123 | } | ||||||||||||||||
2124 | |||||||||||||||||
2125 | template <Intrinsic::ID IntrID, typename T0, typename T1> | ||||||||||||||||
2126 | inline typename m_Intrinsic_Ty<T0, T1>::Ty m_Intrinsic(const T0 &Op0, | ||||||||||||||||
2127 | const T1 &Op1) { | ||||||||||||||||
2128 | return m_CombineAnd(m_Intrinsic<IntrID>(Op0), m_Argument<1>(Op1)); | ||||||||||||||||
2129 | } | ||||||||||||||||
2130 | |||||||||||||||||
2131 | template <Intrinsic::ID IntrID, typename T0, typename T1, typename T2> | ||||||||||||||||
2132 | inline typename m_Intrinsic_Ty<T0, T1, T2>::Ty | ||||||||||||||||
2133 | m_Intrinsic(const T0 &Op0, const T1 &Op1, const T2 &Op2) { | ||||||||||||||||
2134 | return m_CombineAnd(m_Intrinsic<IntrID>(Op0, Op1), m_Argument<2>(Op2)); | ||||||||||||||||
2135 | } | ||||||||||||||||
2136 | |||||||||||||||||
2137 | template <Intrinsic::ID IntrID, typename T0, typename T1, typename T2, | ||||||||||||||||
2138 | typename T3> | ||||||||||||||||
2139 | inline typename m_Intrinsic_Ty<T0, T1, T2, T3>::Ty | ||||||||||||||||
2140 | m_Intrinsic(const T0 &Op0, const T1 &Op1, const T2 &Op2, const T3 &Op3) { | ||||||||||||||||
2141 | return m_CombineAnd(m_Intrinsic<IntrID>(Op0, Op1, Op2), m_Argument<3>(Op3)); | ||||||||||||||||
2142 | } | ||||||||||||||||
2143 | |||||||||||||||||
2144 | template <Intrinsic::ID IntrID, typename T0, typename T1, typename T2, | ||||||||||||||||
2145 | typename T3, typename T4> | ||||||||||||||||
2146 | inline typename m_Intrinsic_Ty<T0, T1, T2, T3, T4>::Ty | ||||||||||||||||
2147 | m_Intrinsic(const T0 &Op0, const T1 &Op1, const T2 &Op2, const T3 &Op3, | ||||||||||||||||
2148 | const T4 &Op4) { | ||||||||||||||||
2149 | return m_CombineAnd(m_Intrinsic<IntrID>(Op0, Op1, Op2, Op3), | ||||||||||||||||
2150 | m_Argument<4>(Op4)); | ||||||||||||||||
2151 | } | ||||||||||||||||
2152 | |||||||||||||||||
2153 | template <Intrinsic::ID IntrID, typename T0, typename T1, typename T2, | ||||||||||||||||
2154 | typename T3, typename T4, typename T5> | ||||||||||||||||
2155 | inline typename m_Intrinsic_Ty<T0, T1, T2, T3, T4, T5>::Ty | ||||||||||||||||
2156 | m_Intrinsic(const T0 &Op0, const T1 &Op1, const T2 &Op2, const T3 &Op3, | ||||||||||||||||
2157 | const T4 &Op4, const T5 &Op5) { | ||||||||||||||||
2158 | return m_CombineAnd(m_Intrinsic<IntrID>(Op0, Op1, Op2, Op3, Op4), | ||||||||||||||||
2159 | m_Argument<5>(Op5)); | ||||||||||||||||
2160 | } | ||||||||||||||||
2161 | |||||||||||||||||
2162 | // Helper intrinsic matching specializations. | ||||||||||||||||
2163 | template <typename Opnd0> | ||||||||||||||||
2164 | inline typename m_Intrinsic_Ty<Opnd0>::Ty m_BitReverse(const Opnd0 &Op0) { | ||||||||||||||||
2165 | return m_Intrinsic<Intrinsic::bitreverse>(Op0); | ||||||||||||||||
2166 | } | ||||||||||||||||
2167 | |||||||||||||||||
2168 | template <typename Opnd0> | ||||||||||||||||
2169 | inline typename m_Intrinsic_Ty<Opnd0>::Ty m_BSwap(const Opnd0 &Op0) { | ||||||||||||||||
2170 | return m_Intrinsic<Intrinsic::bswap>(Op0); | ||||||||||||||||
2171 | } | ||||||||||||||||
2172 | |||||||||||||||||
2173 | template <typename Opnd0> | ||||||||||||||||
2174 | inline typename m_Intrinsic_Ty<Opnd0>::Ty m_FAbs(const Opnd0 &Op0) { | ||||||||||||||||
2175 | return m_Intrinsic<Intrinsic::fabs>(Op0); | ||||||||||||||||
2176 | } | ||||||||||||||||
2177 | |||||||||||||||||
2178 | template <typename Opnd0> | ||||||||||||||||
2179 | inline typename m_Intrinsic_Ty<Opnd0>::Ty m_FCanonicalize(const Opnd0 &Op0) { | ||||||||||||||||
2180 | return m_Intrinsic<Intrinsic::canonicalize>(Op0); | ||||||||||||||||
2181 | } | ||||||||||||||||
2182 | |||||||||||||||||
2183 | template <typename Opnd0, typename Opnd1> | ||||||||||||||||
2184 | inline typename m_Intrinsic_Ty<Opnd0, Opnd1>::Ty m_FMin(const Opnd0 &Op0, | ||||||||||||||||
2185 | const Opnd1 &Op1) { | ||||||||||||||||
2186 | return m_Intrinsic<Intrinsic::minnum>(Op0, Op1); | ||||||||||||||||
2187 | } | ||||||||||||||||
2188 | |||||||||||||||||
2189 | template <typename Opnd0, typename Opnd1> | ||||||||||||||||
2190 | inline typename m_Intrinsic_Ty<Opnd0, Opnd1>::Ty m_FMax(const Opnd0 &Op0, | ||||||||||||||||
2191 | const Opnd1 &Op1) { | ||||||||||||||||
2192 | return m_Intrinsic<Intrinsic::maxnum>(Op0, Op1); | ||||||||||||||||
2193 | } | ||||||||||||||||
2194 | |||||||||||||||||
2195 | template <typename Opnd0, typename Opnd1, typename Opnd2> | ||||||||||||||||
2196 | inline typename m_Intrinsic_Ty<Opnd0, Opnd1, Opnd2>::Ty | ||||||||||||||||
2197 | m_FShl(const Opnd0 &Op0, const Opnd1 &Op1, const Opnd2 &Op2) { | ||||||||||||||||
2198 | return m_Intrinsic<Intrinsic::fshl>(Op0, Op1, Op2); | ||||||||||||||||
2199 | } | ||||||||||||||||
2200 | |||||||||||||||||
2201 | template <typename Opnd0, typename Opnd1, typename Opnd2> | ||||||||||||||||
2202 | inline typename m_Intrinsic_Ty<Opnd0, Opnd1, Opnd2>::Ty | ||||||||||||||||
2203 | m_FShr(const Opnd0 &Op0, const Opnd1 &Op1, const Opnd2 &Op2) { | ||||||||||||||||
2204 | return m_Intrinsic<Intrinsic::fshr>(Op0, Op1, Op2); | ||||||||||||||||
2205 | } | ||||||||||||||||
2206 | |||||||||||||||||
2207 | //===----------------------------------------------------------------------===// | ||||||||||||||||
2208 | // Matchers for two-operands operators with the operators in either order | ||||||||||||||||
2209 | // | ||||||||||||||||
2210 | |||||||||||||||||
2211 | /// Matches a BinaryOperator with LHS and RHS in either order. | ||||||||||||||||
2212 | template <typename LHS, typename RHS> | ||||||||||||||||
2213 | inline AnyBinaryOp_match<LHS, RHS, true> m_c_BinOp(const LHS &L, const RHS &R) { | ||||||||||||||||
2214 | return AnyBinaryOp_match<LHS, RHS, true>(L, R); | ||||||||||||||||
2215 | } | ||||||||||||||||
2216 | |||||||||||||||||
2217 | /// Matches an ICmp with a predicate over LHS and RHS in either order. | ||||||||||||||||
2218 | /// Swaps the predicate if operands are commuted. | ||||||||||||||||
2219 | template <typename LHS, typename RHS> | ||||||||||||||||
2220 | inline CmpClass_match<LHS, RHS, ICmpInst, ICmpInst::Predicate, true> | ||||||||||||||||
2221 | m_c_ICmp(ICmpInst::Predicate &Pred, const LHS &L, const RHS &R) { | ||||||||||||||||
2222 | return CmpClass_match<LHS, RHS, ICmpInst, ICmpInst::Predicate, true>(Pred, L, | ||||||||||||||||
2223 | R); | ||||||||||||||||
2224 | } | ||||||||||||||||
2225 | |||||||||||||||||
2226 | /// Matches a Add with LHS and RHS in either order. | ||||||||||||||||
2227 | template <typename LHS, typename RHS> | ||||||||||||||||
2228 | inline BinaryOp_match<LHS, RHS, Instruction::Add, true> m_c_Add(const LHS &L, | ||||||||||||||||
2229 | const RHS &R) { | ||||||||||||||||
2230 | return BinaryOp_match<LHS, RHS, Instruction::Add, true>(L, R); | ||||||||||||||||
2231 | } | ||||||||||||||||
2232 | |||||||||||||||||
2233 | /// Matches a Mul with LHS and RHS in either order. | ||||||||||||||||
2234 | template <typename LHS, typename RHS> | ||||||||||||||||
2235 | inline BinaryOp_match<LHS, RHS, Instruction::Mul, true> m_c_Mul(const LHS &L, | ||||||||||||||||
2236 | const RHS &R) { | ||||||||||||||||
2237 | return BinaryOp_match<LHS, RHS, Instruction::Mul, true>(L, R); | ||||||||||||||||
2238 | } | ||||||||||||||||
2239 | |||||||||||||||||
2240 | /// Matches an And with LHS and RHS in either order. | ||||||||||||||||
2241 | template <typename LHS, typename RHS> | ||||||||||||||||
2242 | inline BinaryOp_match<LHS, RHS, Instruction::And, true> m_c_And(const LHS &L, | ||||||||||||||||
2243 | const RHS &R) { | ||||||||||||||||
2244 | return BinaryOp_match<LHS, RHS, Instruction::And, true>(L, R); | ||||||||||||||||
2245 | } | ||||||||||||||||
2246 | |||||||||||||||||
2247 | /// Matches an Or with LHS and RHS in either order. | ||||||||||||||||
2248 | template <typename LHS, typename RHS> | ||||||||||||||||
2249 | inline BinaryOp_match<LHS, RHS, Instruction::Or, true> m_c_Or(const LHS &L, | ||||||||||||||||
2250 | const RHS &R) { | ||||||||||||||||
2251 | return BinaryOp_match<LHS, RHS, Instruction::Or, true>(L, R); | ||||||||||||||||
2252 | } | ||||||||||||||||
2253 | |||||||||||||||||
2254 | /// Matches an Xor with LHS and RHS in either order. | ||||||||||||||||
2255 | template <typename LHS, typename RHS> | ||||||||||||||||
2256 | inline BinaryOp_match<LHS, RHS, Instruction::Xor, true> m_c_Xor(const LHS &L, | ||||||||||||||||
2257 | const RHS &R) { | ||||||||||||||||
2258 | return BinaryOp_match<LHS, RHS, Instruction::Xor, true>(L, R); | ||||||||||||||||
2259 | } | ||||||||||||||||
2260 | |||||||||||||||||
2261 | /// Matches a 'Neg' as 'sub 0, V'. | ||||||||||||||||
2262 | template <typename ValTy> | ||||||||||||||||
2263 | inline BinaryOp_match<cst_pred_ty<is_zero_int>, ValTy, Instruction::Sub> | ||||||||||||||||
2264 | m_Neg(const ValTy &V) { | ||||||||||||||||
2265 | return m_Sub(m_ZeroInt(), V); | ||||||||||||||||
2266 | } | ||||||||||||||||
2267 | |||||||||||||||||
2268 | /// Matches a 'Neg' as 'sub nsw 0, V'. | ||||||||||||||||
2269 | template <typename ValTy> | ||||||||||||||||
2270 | inline OverflowingBinaryOp_match<cst_pred_ty<is_zero_int>, ValTy, | ||||||||||||||||
2271 | Instruction::Sub, | ||||||||||||||||
2272 | OverflowingBinaryOperator::NoSignedWrap> | ||||||||||||||||
2273 | m_NSWNeg(const ValTy &V) { | ||||||||||||||||
2274 | return m_NSWSub(m_ZeroInt(), V); | ||||||||||||||||
2275 | } | ||||||||||||||||
2276 | |||||||||||||||||
2277 | /// Matches a 'Not' as 'xor V, -1' or 'xor -1, V'. | ||||||||||||||||
2278 | template <typename ValTy> | ||||||||||||||||
2279 | inline BinaryOp_match<ValTy, cst_pred_ty<is_all_ones>, Instruction::Xor, true> | ||||||||||||||||
2280 | m_Not(const ValTy &V) { | ||||||||||||||||
2281 | return m_c_Xor(V, m_AllOnes()); | ||||||||||||||||
2282 | } | ||||||||||||||||
2283 | |||||||||||||||||
2284 | /// Matches an SMin with LHS and RHS in either order. | ||||||||||||||||
2285 | template <typename LHS, typename RHS> | ||||||||||||||||
2286 | inline MaxMin_match<ICmpInst, LHS, RHS, smin_pred_ty, true> | ||||||||||||||||
2287 | m_c_SMin(const LHS &L, const RHS &R) { | ||||||||||||||||
2288 | return MaxMin_match<ICmpInst, LHS, RHS, smin_pred_ty, true>(L, R); | ||||||||||||||||
2289 | } | ||||||||||||||||
2290 | /// Matches an SMax with LHS and RHS in either order. | ||||||||||||||||
2291 | template <typename LHS, typename RHS> | ||||||||||||||||
2292 | inline MaxMin_match<ICmpInst, LHS, RHS, smax_pred_ty, true> | ||||||||||||||||
2293 | m_c_SMax(const LHS &L, const RHS &R) { | ||||||||||||||||
2294 | return MaxMin_match<ICmpInst, LHS, RHS, smax_pred_ty, true>(L, R); | ||||||||||||||||
2295 | } | ||||||||||||||||
2296 | /// Matches a UMin with LHS and RHS in either order. | ||||||||||||||||
2297 | template <typename LHS, typename RHS> | ||||||||||||||||
2298 | inline MaxMin_match<ICmpInst, LHS, RHS, umin_pred_ty, true> | ||||||||||||||||
2299 | m_c_UMin(const LHS &L, const RHS &R) { | ||||||||||||||||
2300 | return MaxMin_match<ICmpInst, LHS, RHS, umin_pred_ty, true>(L, R); | ||||||||||||||||
2301 | } | ||||||||||||||||
2302 | /// Matches a UMax with LHS and RHS in either order. | ||||||||||||||||
2303 | template <typename LHS, typename RHS> | ||||||||||||||||
2304 | inline MaxMin_match<ICmpInst, LHS, RHS, umax_pred_ty, true> | ||||||||||||||||
2305 | m_c_UMax(const LHS &L, const RHS &R) { | ||||||||||||||||
2306 | return MaxMin_match<ICmpInst, LHS, RHS, umax_pred_ty, true>(L, R); | ||||||||||||||||
2307 | } | ||||||||||||||||
2308 | |||||||||||||||||
2309 | template <typename LHS, typename RHS> | ||||||||||||||||
2310 | inline match_combine_or< | ||||||||||||||||
2311 | match_combine_or<MaxMin_match<ICmpInst, LHS, RHS, smax_pred_ty, true>, | ||||||||||||||||
2312 | MaxMin_match<ICmpInst, LHS, RHS, smin_pred_ty, true>>, | ||||||||||||||||
2313 | match_combine_or<MaxMin_match<ICmpInst, LHS, RHS, umax_pred_ty, true>, | ||||||||||||||||
2314 | MaxMin_match<ICmpInst, LHS, RHS, umin_pred_ty, true>>> | ||||||||||||||||
2315 | m_c_MaxOrMin(const LHS &L, const RHS &R) { | ||||||||||||||||
2316 | return m_CombineOr(m_CombineOr(m_c_SMax(L, R), m_c_SMin(L, R)), | ||||||||||||||||
2317 | m_CombineOr(m_c_UMax(L, R), m_c_UMin(L, R))); | ||||||||||||||||
2318 | } | ||||||||||||||||
2319 | |||||||||||||||||
2320 | /// Matches FAdd with LHS and RHS in either order. | ||||||||||||||||
2321 | template <typename LHS, typename RHS> | ||||||||||||||||
2322 | inline BinaryOp_match<LHS, RHS, Instruction::FAdd, true> | ||||||||||||||||
2323 | m_c_FAdd(const LHS &L, const RHS &R) { | ||||||||||||||||
2324 | return BinaryOp_match<LHS, RHS, Instruction::FAdd, true>(L, R); | ||||||||||||||||
2325 | } | ||||||||||||||||
2326 | |||||||||||||||||
2327 | /// Matches FMul with LHS and RHS in either order. | ||||||||||||||||
2328 | template <typename LHS, typename RHS> | ||||||||||||||||
2329 | inline BinaryOp_match<LHS, RHS, Instruction::FMul, true> | ||||||||||||||||
2330 | m_c_FMul(const LHS &L, const RHS &R) { | ||||||||||||||||
2331 | return BinaryOp_match<LHS, RHS, Instruction::FMul, true>(L, R); | ||||||||||||||||
2332 | } | ||||||||||||||||
2333 | |||||||||||||||||
2334 | template <typename Opnd_t> struct Signum_match { | ||||||||||||||||
2335 | Opnd_t Val; | ||||||||||||||||
2336 | Signum_match(const Opnd_t &V) : Val(V) {} | ||||||||||||||||
2337 | |||||||||||||||||
2338 | template <typename OpTy> bool match(OpTy *V) { | ||||||||||||||||
2339 | unsigned TypeSize = V->getType()->getScalarSizeInBits(); | ||||||||||||||||
2340 | if (TypeSize == 0) | ||||||||||||||||
2341 | return false; | ||||||||||||||||
2342 | |||||||||||||||||
2343 | unsigned ShiftWidth = TypeSize - 1; | ||||||||||||||||
2344 | Value *OpL = nullptr, *OpR = nullptr; | ||||||||||||||||
2345 | |||||||||||||||||
2346 | // This is the representation of signum we match: | ||||||||||||||||
2347 | // | ||||||||||||||||
2348 | // signum(x) == (x >> 63) | (-x >>u 63) | ||||||||||||||||
2349 | // | ||||||||||||||||
2350 | // An i1 value is its own signum, so it's correct to match | ||||||||||||||||
2351 | // | ||||||||||||||||
2352 | // signum(x) == (x >> 0) | (-x >>u 0) | ||||||||||||||||
2353 | // | ||||||||||||||||
2354 | // for i1 values. | ||||||||||||||||
2355 | |||||||||||||||||
2356 | auto LHS = m_AShr(m_Value(OpL), m_SpecificInt(ShiftWidth)); | ||||||||||||||||
2357 | auto RHS = m_LShr(m_Neg(m_Value(OpR)), m_SpecificInt(ShiftWidth)); | ||||||||||||||||
2358 | auto Signum = m_Or(LHS, RHS); | ||||||||||||||||
2359 | |||||||||||||||||
2360 | return Signum.match(V) && OpL == OpR && Val.match(OpL); | ||||||||||||||||
2361 | } | ||||||||||||||||
2362 | }; | ||||||||||||||||
2363 | |||||||||||||||||
2364 | /// Matches a signum pattern. | ||||||||||||||||
2365 | /// | ||||||||||||||||
2366 | /// signum(x) = | ||||||||||||||||
2367 | /// x > 0 -> 1 | ||||||||||||||||
2368 | /// x == 0 -> 0 | ||||||||||||||||
2369 | /// x < 0 -> -1 | ||||||||||||||||
2370 | template <typename Val_t> inline Signum_match<Val_t> m_Signum(const Val_t &V) { | ||||||||||||||||
2371 | return Signum_match<Val_t>(V); | ||||||||||||||||
2372 | } | ||||||||||||||||
2373 | |||||||||||||||||
2374 | template <int Ind, typename Opnd_t> struct ExtractValue_match { | ||||||||||||||||
2375 | Opnd_t Val; | ||||||||||||||||
2376 | ExtractValue_match(const Opnd_t &V) : Val(V) {} | ||||||||||||||||
2377 | |||||||||||||||||
2378 | template <typename OpTy> bool match(OpTy *V) { | ||||||||||||||||
2379 | if (auto *I = dyn_cast<ExtractValueInst>(V)) { | ||||||||||||||||
2380 | // If Ind is -1, don't inspect indices | ||||||||||||||||
2381 | if (Ind != -1 && | ||||||||||||||||
2382 | !(I->getNumIndices() == 1 && I->getIndices()[0] == (unsigned)Ind)) | ||||||||||||||||
2383 | return false; | ||||||||||||||||
2384 | return Val.match(I->getAggregateOperand()); | ||||||||||||||||
2385 | } | ||||||||||||||||
2386 | return false; | ||||||||||||||||
2387 | } | ||||||||||||||||
2388 | }; | ||||||||||||||||
2389 | |||||||||||||||||
2390 | /// Match a single index ExtractValue instruction. | ||||||||||||||||
2391 | /// For example m_ExtractValue<1>(...) | ||||||||||||||||
2392 | template <int Ind, typename Val_t> | ||||||||||||||||
2393 | inline ExtractValue_match<Ind, Val_t> m_ExtractValue(const Val_t &V) { | ||||||||||||||||
2394 | return ExtractValue_match<Ind, Val_t>(V); | ||||||||||||||||
2395 | } | ||||||||||||||||
2396 | |||||||||||||||||
2397 | /// Match an ExtractValue instruction with any index. | ||||||||||||||||
2398 | /// For example m_ExtractValue(...) | ||||||||||||||||
2399 | template <typename Val_t> | ||||||||||||||||
2400 | inline ExtractValue_match<-1, Val_t> m_ExtractValue(const Val_t &V) { | ||||||||||||||||
2401 | return ExtractValue_match<-1, Val_t>(V); | ||||||||||||||||
2402 | } | ||||||||||||||||
2403 | |||||||||||||||||
2404 | /// Matcher for a single index InsertValue instruction. | ||||||||||||||||
2405 | template <int Ind, typename T0, typename T1> struct InsertValue_match { | ||||||||||||||||
2406 | T0 Op0; | ||||||||||||||||
2407 | T1 Op1; | ||||||||||||||||
2408 | |||||||||||||||||
2409 | InsertValue_match(const T0 &Op0, const T1 &Op1) : Op0(Op0), Op1(Op1) {} | ||||||||||||||||
2410 | |||||||||||||||||
2411 | template <typename OpTy> bool match(OpTy *V) { | ||||||||||||||||
2412 | if (auto *I = dyn_cast<InsertValueInst>(V)) { | ||||||||||||||||
2413 | return Op0.match(I->getOperand(0)) && Op1.match(I->getOperand(1)) && | ||||||||||||||||
2414 | I->getNumIndices() == 1 && Ind == I->getIndices()[0]; | ||||||||||||||||
2415 | } | ||||||||||||||||
2416 | return false; | ||||||||||||||||
2417 | } | ||||||||||||||||
2418 | }; | ||||||||||||||||
2419 | |||||||||||||||||
2420 | /// Matches a single index InsertValue instruction. | ||||||||||||||||
2421 | template <int Ind, typename Val_t, typename Elt_t> | ||||||||||||||||
2422 | inline InsertValue_match<Ind, Val_t, Elt_t> m_InsertValue(const Val_t &Val, | ||||||||||||||||
2423 | const Elt_t &Elt) { | ||||||||||||||||
2424 | return InsertValue_match<Ind, Val_t, Elt_t>(Val, Elt); | ||||||||||||||||
2425 | } | ||||||||||||||||
2426 | |||||||||||||||||
2427 | /// Matches patterns for `vscale`. This can either be a call to `llvm.vscale` or | ||||||||||||||||
2428 | /// the constant expression | ||||||||||||||||
2429 | /// `ptrtoint(gep <vscale x 1 x i8>, <vscale x 1 x i8>* null, i32 1>` | ||||||||||||||||
2430 | /// under the right conditions determined by DataLayout. | ||||||||||||||||
2431 | struct VScaleVal_match { | ||||||||||||||||
2432 | const DataLayout &DL; | ||||||||||||||||
2433 | VScaleVal_match(const DataLayout &DL) : DL(DL) {} | ||||||||||||||||
2434 | |||||||||||||||||
2435 | template <typename ITy> bool match(ITy *V) { | ||||||||||||||||
2436 | if (m_Intrinsic<Intrinsic::vscale>().match(V)) | ||||||||||||||||
2437 | return true; | ||||||||||||||||
2438 | |||||||||||||||||
2439 | Value *Ptr; | ||||||||||||||||
2440 | if (m_PtrToInt(m_Value(Ptr)).match(V)) { | ||||||||||||||||
2441 | if (auto *GEP = dyn_cast<GEPOperator>(Ptr)) { | ||||||||||||||||
2442 | auto *DerefTy = GEP->getSourceElementType(); | ||||||||||||||||
2443 | if (GEP->getNumIndices() == 1 && isa<ScalableVectorType>(DerefTy) && | ||||||||||||||||
2444 | m_Zero().match(GEP->getPointerOperand()) && | ||||||||||||||||
2445 | m_SpecificInt(1).match(GEP->idx_begin()->get()) && | ||||||||||||||||
2446 | DL.getTypeAllocSizeInBits(DerefTy).getKnownMinSize() == 8) | ||||||||||||||||
2447 | return true; | ||||||||||||||||
2448 | } | ||||||||||||||||
2449 | } | ||||||||||||||||
2450 | |||||||||||||||||
2451 | return false; | ||||||||||||||||
2452 | } | ||||||||||||||||
2453 | }; | ||||||||||||||||
2454 | |||||||||||||||||
2455 | inline VScaleVal_match m_VScale(const DataLayout &DL) { | ||||||||||||||||
2456 | return VScaleVal_match(DL); | ||||||||||||||||
2457 | } | ||||||||||||||||
2458 | |||||||||||||||||
2459 | template <typename LHS, typename RHS, unsigned Opcode> | ||||||||||||||||
2460 | struct LogicalOp_match { | ||||||||||||||||
2461 | LHS L; | ||||||||||||||||
2462 | RHS R; | ||||||||||||||||
2463 | |||||||||||||||||
2464 | LogicalOp_match(const LHS &L, const RHS &R) : L(L), R(R) {} | ||||||||||||||||
2465 | |||||||||||||||||
2466 | template <typename T> bool match(T *V) { | ||||||||||||||||
2467 | if (auto *I = dyn_cast<Instruction>(V)) { | ||||||||||||||||
2468 | if (!I->getType()->isIntOrIntVectorTy(1)) | ||||||||||||||||
2469 | return false; | ||||||||||||||||
2470 | |||||||||||||||||
2471 | if (I->getOpcode() == Opcode && L.match(I->getOperand(0)) && | ||||||||||||||||
2472 | R.match(I->getOperand(1))) | ||||||||||||||||
2473 | return true; | ||||||||||||||||
2474 | |||||||||||||||||
2475 | if (auto *SI = dyn_cast<SelectInst>(I)) { | ||||||||||||||||
2476 | if (Opcode == Instruction::And) { | ||||||||||||||||
2477 | if (const auto *C = dyn_cast<Constant>(SI->getFalseValue())) | ||||||||||||||||
2478 | if (C->isNullValue() && L.match(SI->getCondition()) && | ||||||||||||||||
2479 | R.match(SI->getTrueValue())) | ||||||||||||||||
2480 | return true; | ||||||||||||||||
2481 | } else { | ||||||||||||||||
2482 | assert(Opcode == Instruction::Or)((void)0); | ||||||||||||||||
2483 | if (const auto *C = dyn_cast<Constant>(SI->getTrueValue())) | ||||||||||||||||
2484 | if (C->isOneValue() && L.match(SI->getCondition()) && | ||||||||||||||||
2485 | R.match(SI->getFalseValue())) | ||||||||||||||||
2486 | return true; | ||||||||||||||||
2487 | } | ||||||||||||||||
2488 | } | ||||||||||||||||
2489 | } | ||||||||||||||||
2490 | |||||||||||||||||
2491 | return false; | ||||||||||||||||
2492 | } | ||||||||||||||||
2493 | }; | ||||||||||||||||
2494 | |||||||||||||||||
2495 | /// Matches L && R either in the form of L & R or L ? R : false. | ||||||||||||||||
2496 | /// Note that the latter form is poison-blocking. | ||||||||||||||||
2497 | template <typename LHS, typename RHS> | ||||||||||||||||
2498 | inline LogicalOp_match<LHS, RHS, Instruction::And> | ||||||||||||||||
2499 | m_LogicalAnd(const LHS &L, const RHS &R) { | ||||||||||||||||
2500 | return LogicalOp_match<LHS, RHS, Instruction::And>(L, R); | ||||||||||||||||
2501 | } | ||||||||||||||||
2502 | |||||||||||||||||
2503 | /// Matches L && R where L and R are arbitrary values. | ||||||||||||||||
2504 | inline auto m_LogicalAnd() { return m_LogicalAnd(m_Value(), m_Value()); } | ||||||||||||||||
2505 | |||||||||||||||||
2506 | /// Matches L || R either in the form of L | R or L ? true : R. | ||||||||||||||||
2507 | /// Note that the latter form is poison-blocking. | ||||||||||||||||
2508 | template <typename LHS, typename RHS> | ||||||||||||||||
2509 | inline LogicalOp_match<LHS, RHS, Instruction::Or> | ||||||||||||||||
2510 | m_LogicalOr(const LHS &L, const RHS &R) { | ||||||||||||||||
2511 | return LogicalOp_match<LHS, RHS, Instruction::Or>(L, R); | ||||||||||||||||
2512 | } | ||||||||||||||||
2513 | |||||||||||||||||
2514 | /// Matches L || R where L and R are arbitrary values. | ||||||||||||||||
2515 | inline auto m_LogicalOr() { | ||||||||||||||||
2516 | return m_LogicalOr(m_Value(), m_Value()); | ||||||||||||||||
2517 | } | ||||||||||||||||
2518 | |||||||||||||||||
2519 | } // end namespace PatternMatch | ||||||||||||||||
2520 | } // end namespace llvm | ||||||||||||||||
2521 | |||||||||||||||||
2522 | #endif // LLVM_IR_PATTERNMATCH_H |
1 | //===- llvm/Analysis/AliasAnalysis.h - Alias Analysis Interface -*- C++ -*-===// |
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 file defines the generic AliasAnalysis interface, which is used as the |
10 | // common interface used by all clients of alias analysis information, and |
11 | // implemented by all alias analysis implementations. Mod/Ref information is |
12 | // also captured by this interface. |
13 | // |
14 | // Implementations of this interface must implement the various virtual methods, |
15 | // which automatically provides functionality for the entire suite of client |
16 | // APIs. |
17 | // |
18 | // This API identifies memory regions with the MemoryLocation class. The pointer |
19 | // component specifies the base memory address of the region. The Size specifies |
20 | // the maximum size (in address units) of the memory region, or |
21 | // MemoryLocation::UnknownSize if the size is not known. The TBAA tag |
22 | // identifies the "type" of the memory reference; see the |
23 | // TypeBasedAliasAnalysis class for details. |
24 | // |
25 | // Some non-obvious details include: |
26 | // - Pointers that point to two completely different objects in memory never |
27 | // alias, regardless of the value of the Size component. |
28 | // - NoAlias doesn't imply inequal pointers. The most obvious example of this |
29 | // is two pointers to constant memory. Even if they are equal, constant |
30 | // memory is never stored to, so there will never be any dependencies. |
31 | // In this and other situations, the pointers may be both NoAlias and |
32 | // MustAlias at the same time. The current API can only return one result, |
33 | // though this is rarely a problem in practice. |
34 | // |
35 | //===----------------------------------------------------------------------===// |
36 | |
37 | #ifndef LLVM_ANALYSIS_ALIASANALYSIS_H |
38 | #define LLVM_ANALYSIS_ALIASANALYSIS_H |
39 | |
40 | #include "llvm/ADT/DenseMap.h" |
41 | #include "llvm/ADT/None.h" |
42 | #include "llvm/ADT/Optional.h" |
43 | #include "llvm/ADT/SmallVector.h" |
44 | #include "llvm/Analysis/MemoryLocation.h" |
45 | #include "llvm/IR/PassManager.h" |
46 | #include "llvm/Pass.h" |
47 | #include <cstdint> |
48 | #include <functional> |
49 | #include <memory> |
50 | #include <vector> |
51 | |
52 | namespace llvm { |
53 | |
54 | class AnalysisUsage; |
55 | class AtomicCmpXchgInst; |
56 | class BasicAAResult; |
57 | class BasicBlock; |
58 | class CatchPadInst; |
59 | class CatchReturnInst; |
60 | class DominatorTree; |
61 | class FenceInst; |
62 | class Function; |
63 | class InvokeInst; |
64 | class PreservedAnalyses; |
65 | class TargetLibraryInfo; |
66 | class Value; |
67 | |
68 | /// The possible results of an alias query. |
69 | /// |
70 | /// These results are always computed between two MemoryLocation objects as |
71 | /// a query to some alias analysis. |
72 | /// |
73 | /// Note that these are unscoped enumerations because we would like to support |
74 | /// implicitly testing a result for the existence of any possible aliasing with |
75 | /// a conversion to bool, but an "enum class" doesn't support this. The |
76 | /// canonical names from the literature are suffixed and unique anyways, and so |
77 | /// they serve as global constants in LLVM for these results. |
78 | /// |
79 | /// See docs/AliasAnalysis.html for more information on the specific meanings |
80 | /// of these values. |
81 | class AliasResult { |
82 | private: |
83 | static const int OffsetBits = 23; |
84 | static const int AliasBits = 8; |
85 | static_assert(AliasBits + 1 + OffsetBits <= 32, |
86 | "AliasResult size is intended to be 4 bytes!"); |
87 | |
88 | unsigned int Alias : AliasBits; |
89 | unsigned int HasOffset : 1; |
90 | signed int Offset : OffsetBits; |
91 | |
92 | public: |
93 | enum Kind : uint8_t { |
94 | /// The two locations do not alias at all. |
95 | /// |
96 | /// This value is arranged to convert to false, while all other values |
97 | /// convert to true. This allows a boolean context to convert the result to |
98 | /// a binary flag indicating whether there is the possibility of aliasing. |
99 | NoAlias = 0, |
100 | /// The two locations may or may not alias. This is the least precise |
101 | /// result. |
102 | MayAlias, |
103 | /// The two locations alias, but only due to a partial overlap. |
104 | PartialAlias, |
105 | /// The two locations precisely alias each other. |
106 | MustAlias, |
107 | }; |
108 | static_assert(MustAlias < (1 << AliasBits), |
109 | "Not enough bit field size for the enum!"); |
110 | |
111 | explicit AliasResult() = delete; |
112 | constexpr AliasResult(const Kind &Alias) |
113 | : Alias(Alias), HasOffset(false), Offset(0) {} |
114 | |
115 | operator Kind() const { return static_cast<Kind>(Alias); } |
116 | |
117 | constexpr bool hasOffset() const { return HasOffset; } |
118 | constexpr int32_t getOffset() const { |
119 | assert(HasOffset && "No offset!")((void)0); |
120 | return Offset; |
121 | } |
122 | void setOffset(int32_t NewOffset) { |
123 | if (isInt<OffsetBits>(NewOffset)) { |
124 | HasOffset = true; |
125 | Offset = NewOffset; |
126 | } |
127 | } |
128 | |
129 | /// Helper for processing AliasResult for swapped memory location pairs. |
130 | void swap(bool DoSwap = true) { |
131 | if (DoSwap && hasOffset()) |
132 | setOffset(-getOffset()); |
133 | } |
134 | }; |
135 | |
136 | static_assert(sizeof(AliasResult) == 4, |
137 | "AliasResult size is intended to be 4 bytes!"); |
138 | |
139 | /// << operator for AliasResult. |
140 | raw_ostream &operator<<(raw_ostream &OS, AliasResult AR); |
141 | |
142 | /// Flags indicating whether a memory access modifies or references memory. |
143 | /// |
144 | /// This is no access at all, a modification, a reference, or both |
145 | /// a modification and a reference. These are specifically structured such that |
146 | /// they form a three bit matrix and bit-tests for 'mod' or 'ref' or 'must' |
147 | /// work with any of the possible values. |
148 | enum class ModRefInfo : uint8_t { |
149 | /// Must is provided for completeness, but no routines will return only |
150 | /// Must today. See definition of Must below. |
151 | Must = 0, |
152 | /// The access may reference the value stored in memory, |
153 | /// a mustAlias relation was found, and no mayAlias or partialAlias found. |
154 | MustRef = 1, |
155 | /// The access may modify the value stored in memory, |
156 | /// a mustAlias relation was found, and no mayAlias or partialAlias found. |
157 | MustMod = 2, |
158 | /// The access may reference, modify or both the value stored in memory, |
159 | /// a mustAlias relation was found, and no mayAlias or partialAlias found. |
160 | MustModRef = MustRef | MustMod, |
161 | /// The access neither references nor modifies the value stored in memory. |
162 | NoModRef = 4, |
163 | /// The access may reference the value stored in memory. |
164 | Ref = NoModRef | MustRef, |
165 | /// The access may modify the value stored in memory. |
166 | Mod = NoModRef | MustMod, |
167 | /// The access may reference and may modify the value stored in memory. |
168 | ModRef = Ref | Mod, |
169 | |
170 | /// About Must: |
171 | /// Must is set in a best effort manner. |
172 | /// We usually do not try our best to infer Must, instead it is merely |
173 | /// another piece of "free" information that is presented when available. |
174 | /// Must set means there was certainly a MustAlias found. For calls, |
175 | /// where multiple arguments are checked (argmemonly), this translates to |
176 | /// only MustAlias or NoAlias was found. |
177 | /// Must is not set for RAR accesses, even if the two locations must |
178 | /// alias. The reason is that two read accesses translate to an early return |
179 | /// of NoModRef. An additional alias check to set Must may be |
180 | /// expensive. Other cases may also not set Must(e.g. callCapturesBefore). |
181 | /// We refer to Must being *set* when the most significant bit is *cleared*. |
182 | /// Conversely we *clear* Must information by *setting* the Must bit to 1. |
183 | }; |
184 | |
185 | LLVM_NODISCARD[[clang::warn_unused_result]] inline bool isNoModRef(const ModRefInfo MRI) { |
186 | return (static_cast<int>(MRI) & static_cast<int>(ModRefInfo::MustModRef)) == |
187 | static_cast<int>(ModRefInfo::Must); |
188 | } |
189 | LLVM_NODISCARD[[clang::warn_unused_result]] inline bool isModOrRefSet(const ModRefInfo MRI) { |
190 | return static_cast<int>(MRI) & static_cast<int>(ModRefInfo::MustModRef); |
191 | } |
192 | LLVM_NODISCARD[[clang::warn_unused_result]] inline bool isModAndRefSet(const ModRefInfo MRI) { |
193 | return (static_cast<int>(MRI) & static_cast<int>(ModRefInfo::MustModRef)) == |
194 | static_cast<int>(ModRefInfo::MustModRef); |
195 | } |
196 | LLVM_NODISCARD[[clang::warn_unused_result]] inline bool isModSet(const ModRefInfo MRI) { |
197 | return static_cast<int>(MRI) & static_cast<int>(ModRefInfo::MustMod); |
198 | } |
199 | LLVM_NODISCARD[[clang::warn_unused_result]] inline bool isRefSet(const ModRefInfo MRI) { |
200 | return static_cast<int>(MRI) & static_cast<int>(ModRefInfo::MustRef); |
201 | } |
202 | LLVM_NODISCARD[[clang::warn_unused_result]] inline bool isMustSet(const ModRefInfo MRI) { |
203 | return !(static_cast<int>(MRI) & static_cast<int>(ModRefInfo::NoModRef)); |
204 | } |
205 | |
206 | LLVM_NODISCARD[[clang::warn_unused_result]] inline ModRefInfo setMod(const ModRefInfo MRI) { |
207 | return ModRefInfo(static_cast<int>(MRI) | |
208 | static_cast<int>(ModRefInfo::MustMod)); |
209 | } |
210 | LLVM_NODISCARD[[clang::warn_unused_result]] inline ModRefInfo setRef(const ModRefInfo MRI) { |
211 | return ModRefInfo(static_cast<int>(MRI) | |
212 | static_cast<int>(ModRefInfo::MustRef)); |
213 | } |
214 | LLVM_NODISCARD[[clang::warn_unused_result]] inline ModRefInfo setMust(const ModRefInfo MRI) { |
215 | return ModRefInfo(static_cast<int>(MRI) & |
216 | static_cast<int>(ModRefInfo::MustModRef)); |
217 | } |
218 | LLVM_NODISCARD[[clang::warn_unused_result]] inline ModRefInfo setModAndRef(const ModRefInfo MRI) { |
219 | return ModRefInfo(static_cast<int>(MRI) | |
220 | static_cast<int>(ModRefInfo::MustModRef)); |
221 | } |
222 | LLVM_NODISCARD[[clang::warn_unused_result]] inline ModRefInfo clearMod(const ModRefInfo MRI) { |
223 | return ModRefInfo(static_cast<int>(MRI) & static_cast<int>(ModRefInfo::Ref)); |
224 | } |
225 | LLVM_NODISCARD[[clang::warn_unused_result]] inline ModRefInfo clearRef(const ModRefInfo MRI) { |
226 | return ModRefInfo(static_cast<int>(MRI) & static_cast<int>(ModRefInfo::Mod)); |
227 | } |
228 | LLVM_NODISCARD[[clang::warn_unused_result]] inline ModRefInfo clearMust(const ModRefInfo MRI) { |
229 | return ModRefInfo(static_cast<int>(MRI) | |
230 | static_cast<int>(ModRefInfo::NoModRef)); |
231 | } |
232 | LLVM_NODISCARD[[clang::warn_unused_result]] inline ModRefInfo unionModRef(const ModRefInfo MRI1, |
233 | const ModRefInfo MRI2) { |
234 | return ModRefInfo(static_cast<int>(MRI1) | static_cast<int>(MRI2)); |
235 | } |
236 | LLVM_NODISCARD[[clang::warn_unused_result]] inline ModRefInfo intersectModRef(const ModRefInfo MRI1, |
237 | const ModRefInfo MRI2) { |
238 | return ModRefInfo(static_cast<int>(MRI1) & static_cast<int>(MRI2)); |
239 | } |
240 | |
241 | /// The locations at which a function might access memory. |
242 | /// |
243 | /// These are primarily used in conjunction with the \c AccessKind bits to |
244 | /// describe both the nature of access and the locations of access for a |
245 | /// function call. |
246 | enum FunctionModRefLocation { |
247 | /// Base case is no access to memory. |
248 | FMRL_Nowhere = 0, |
249 | /// Access to memory via argument pointers. |
250 | FMRL_ArgumentPointees = 8, |
251 | /// Memory that is inaccessible via LLVM IR. |
252 | FMRL_InaccessibleMem = 16, |
253 | /// Access to any memory. |
254 | FMRL_Anywhere = 32 | FMRL_InaccessibleMem | FMRL_ArgumentPointees |
255 | }; |
256 | |
257 | /// Summary of how a function affects memory in the program. |
258 | /// |
259 | /// Loads from constant globals are not considered memory accesses for this |
260 | /// interface. Also, functions may freely modify stack space local to their |
261 | /// invocation without having to report it through these interfaces. |
262 | enum FunctionModRefBehavior { |
263 | /// This function does not perform any non-local loads or stores to memory. |
264 | /// |
265 | /// This property corresponds to the GCC 'const' attribute. |
266 | /// This property corresponds to the LLVM IR 'readnone' attribute. |
267 | /// This property corresponds to the IntrNoMem LLVM intrinsic flag. |
268 | FMRB_DoesNotAccessMemory = |
269 | FMRL_Nowhere | static_cast<int>(ModRefInfo::NoModRef), |
270 | |
271 | /// The only memory references in this function (if it has any) are |
272 | /// non-volatile loads from objects pointed to by its pointer-typed |
273 | /// arguments, with arbitrary offsets. |
274 | /// |
275 | /// This property corresponds to the combination of the IntrReadMem |
276 | /// and IntrArgMemOnly LLVM intrinsic flags. |
277 | FMRB_OnlyReadsArgumentPointees = |
278 | FMRL_ArgumentPointees | static_cast<int>(ModRefInfo::Ref), |
279 | |
280 | /// The only memory references in this function (if it has any) are |
281 | /// non-volatile stores from objects pointed to by its pointer-typed |
282 | /// arguments, with arbitrary offsets. |
283 | /// |
284 | /// This property corresponds to the combination of the IntrWriteMem |
285 | /// and IntrArgMemOnly LLVM intrinsic flags. |
286 | FMRB_OnlyWritesArgumentPointees = |
287 | FMRL_ArgumentPointees | static_cast<int>(ModRefInfo::Mod), |
288 | |
289 | /// The only memory references in this function (if it has any) are |
290 | /// non-volatile loads and stores from objects pointed to by its |
291 | /// pointer-typed arguments, with arbitrary offsets. |
292 | /// |
293 | /// This property corresponds to the IntrArgMemOnly LLVM intrinsic flag. |
294 | FMRB_OnlyAccessesArgumentPointees = |
295 | FMRL_ArgumentPointees | static_cast<int>(ModRefInfo::ModRef), |
296 | |
297 | /// The only memory references in this function (if it has any) are |
298 | /// reads of memory that is otherwise inaccessible via LLVM IR. |
299 | /// |
300 | /// This property corresponds to the LLVM IR inaccessiblememonly attribute. |
301 | FMRB_OnlyReadsInaccessibleMem = |
302 | FMRL_InaccessibleMem | static_cast<int>(ModRefInfo::Ref), |
303 | |
304 | /// The only memory references in this function (if it has any) are |
305 | /// writes to memory that is otherwise inaccessible via LLVM IR. |
306 | /// |
307 | /// This property corresponds to the LLVM IR inaccessiblememonly attribute. |
308 | FMRB_OnlyWritesInaccessibleMem = |
309 | FMRL_InaccessibleMem | static_cast<int>(ModRefInfo::Mod), |
310 | |
311 | /// The only memory references in this function (if it has any) are |
312 | /// references of memory that is otherwise inaccessible via LLVM IR. |
313 | /// |
314 | /// This property corresponds to the LLVM IR inaccessiblememonly attribute. |
315 | FMRB_OnlyAccessesInaccessibleMem = |
316 | FMRL_InaccessibleMem | static_cast<int>(ModRefInfo::ModRef), |
317 | |
318 | /// The function may perform non-volatile loads from objects pointed |
319 | /// to by its pointer-typed arguments, with arbitrary offsets, and |
320 | /// it may also perform loads of memory that is otherwise |
321 | /// inaccessible via LLVM IR. |
322 | /// |
323 | /// This property corresponds to the LLVM IR |
324 | /// inaccessiblemem_or_argmemonly attribute. |
325 | FMRB_OnlyReadsInaccessibleOrArgMem = FMRL_InaccessibleMem | |
326 | FMRL_ArgumentPointees | |
327 | static_cast<int>(ModRefInfo::Ref), |
328 | |
329 | /// The function may perform non-volatile stores to objects pointed |
330 | /// to by its pointer-typed arguments, with arbitrary offsets, and |
331 | /// it may also perform stores of memory that is otherwise |
332 | /// inaccessible via LLVM IR. |
333 | /// |
334 | /// This property corresponds to the LLVM IR |
335 | /// inaccessiblemem_or_argmemonly attribute. |
336 | FMRB_OnlyWritesInaccessibleOrArgMem = FMRL_InaccessibleMem | |
337 | FMRL_ArgumentPointees | |
338 | static_cast<int>(ModRefInfo::Mod), |
339 | |
340 | /// The function may perform non-volatile loads and stores of objects |
341 | /// pointed to by its pointer-typed arguments, with arbitrary offsets, and |
342 | /// it may also perform loads and stores of memory that is otherwise |
343 | /// inaccessible via LLVM IR. |
344 | /// |
345 | /// This property corresponds to the LLVM IR |
346 | /// inaccessiblemem_or_argmemonly attribute. |
347 | FMRB_OnlyAccessesInaccessibleOrArgMem = FMRL_InaccessibleMem | |
348 | FMRL_ArgumentPointees | |
349 | static_cast<int>(ModRefInfo::ModRef), |
350 | |
351 | /// This function does not perform any non-local stores or volatile loads, |
352 | /// but may read from any memory location. |
353 | /// |
354 | /// This property corresponds to the GCC 'pure' attribute. |
355 | /// This property corresponds to the LLVM IR 'readonly' attribute. |
356 | /// This property corresponds to the IntrReadMem LLVM intrinsic flag. |
357 | FMRB_OnlyReadsMemory = FMRL_Anywhere | static_cast<int>(ModRefInfo::Ref), |
358 | |
359 | // This function does not read from memory anywhere, but may write to any |
360 | // memory location. |
361 | // |
362 | // This property corresponds to the LLVM IR 'writeonly' attribute. |
363 | // This property corresponds to the IntrWriteMem LLVM intrinsic flag. |
364 | FMRB_OnlyWritesMemory = FMRL_Anywhere | static_cast<int>(ModRefInfo::Mod), |
365 | |
366 | /// This indicates that the function could not be classified into one of the |
367 | /// behaviors above. |
368 | FMRB_UnknownModRefBehavior = |
369 | FMRL_Anywhere | static_cast<int>(ModRefInfo::ModRef) |
370 | }; |
371 | |
372 | // Wrapper method strips bits significant only in FunctionModRefBehavior, |
373 | // to obtain a valid ModRefInfo. The benefit of using the wrapper is that if |
374 | // ModRefInfo enum changes, the wrapper can be updated to & with the new enum |
375 | // entry with all bits set to 1. |
376 | LLVM_NODISCARD[[clang::warn_unused_result]] inline ModRefInfo |
377 | createModRefInfo(const FunctionModRefBehavior FMRB) { |
378 | return ModRefInfo(FMRB & static_cast<int>(ModRefInfo::ModRef)); |
379 | } |
380 | |
381 | /// Reduced version of MemoryLocation that only stores a pointer and size. |
382 | /// Used for caching AATags independent BasicAA results. |
383 | struct AACacheLoc { |
384 | const Value *Ptr; |
385 | LocationSize Size; |
386 | }; |
387 | |
388 | template <> struct DenseMapInfo<AACacheLoc> { |
389 | static inline AACacheLoc getEmptyKey() { |
390 | return {DenseMapInfo<const Value *>::getEmptyKey(), |
391 | DenseMapInfo<LocationSize>::getEmptyKey()}; |
392 | } |
393 | static inline AACacheLoc getTombstoneKey() { |
394 | return {DenseMapInfo<const Value *>::getTombstoneKey(), |
395 | DenseMapInfo<LocationSize>::getTombstoneKey()}; |
396 | } |
397 | static unsigned getHashValue(const AACacheLoc &Val) { |
398 | return DenseMapInfo<const Value *>::getHashValue(Val.Ptr) ^ |
399 | DenseMapInfo<LocationSize>::getHashValue(Val.Size); |
400 | } |
401 | static bool isEqual(const AACacheLoc &LHS, const AACacheLoc &RHS) { |
402 | return LHS.Ptr == RHS.Ptr && LHS.Size == RHS.Size; |
403 | } |
404 | }; |
405 | |
406 | /// This class stores info we want to provide to or retain within an alias |
407 | /// query. By default, the root query is stateless and starts with a freshly |
408 | /// constructed info object. Specific alias analyses can use this query info to |
409 | /// store per-query state that is important for recursive or nested queries to |
410 | /// avoid recomputing. To enable preserving this state across multiple queries |
411 | /// where safe (due to the IR not changing), use a `BatchAAResults` wrapper. |
412 | /// The information stored in an `AAQueryInfo` is currently limitted to the |
413 | /// caches used by BasicAA, but can further be extended to fit other AA needs. |
414 | class AAQueryInfo { |
415 | public: |
416 | using LocPair = std::pair<AACacheLoc, AACacheLoc>; |
417 | struct CacheEntry { |
418 | AliasResult Result; |
419 | /// Number of times a NoAlias assumption has been used. |
420 | /// 0 for assumptions that have not been used, -1 for definitive results. |
421 | int NumAssumptionUses; |
422 | /// Whether this is a definitive (non-assumption) result. |
423 | bool isDefinitive() const { return NumAssumptionUses < 0; } |
424 | }; |
425 | using AliasCacheT = SmallDenseMap<LocPair, CacheEntry, 8>; |
426 | AliasCacheT AliasCache; |
427 | |
428 | using IsCapturedCacheT = SmallDenseMap<const Value *, bool, 8>; |
429 | IsCapturedCacheT IsCapturedCache; |
430 | |
431 | /// Query depth used to distinguish recursive queries. |
432 | unsigned Depth = 0; |
433 | |
434 | /// How many active NoAlias assumption uses there are. |
435 | int NumAssumptionUses = 0; |
436 | |
437 | /// Location pairs for which an assumption based result is currently stored. |
438 | /// Used to remove all potentially incorrect results from the cache if an |
439 | /// assumption is disproven. |
440 | SmallVector<AAQueryInfo::LocPair, 4> AssumptionBasedResults; |
441 | |
442 | AAQueryInfo() : AliasCache(), IsCapturedCache() {} |
443 | |
444 | /// Create a new AAQueryInfo based on this one, but with the cache cleared. |
445 | /// This is used for recursive queries across phis, where cache results may |
446 | /// not be valid. |
447 | AAQueryInfo withEmptyCache() { |
448 | AAQueryInfo NewAAQI; |
449 | NewAAQI.Depth = Depth; |
450 | return NewAAQI; |
451 | } |
452 | }; |
453 | |
454 | class BatchAAResults; |
455 | |
456 | class AAResults { |
457 | public: |
458 | // Make these results default constructable and movable. We have to spell |
459 | // these out because MSVC won't synthesize them. |
460 | AAResults(const TargetLibraryInfo &TLI) : TLI(TLI) {} |
461 | AAResults(AAResults &&Arg); |
462 | ~AAResults(); |
463 | |
464 | /// Register a specific AA result. |
465 | template <typename AAResultT> void addAAResult(AAResultT &AAResult) { |
466 | // FIXME: We should use a much lighter weight system than the usual |
467 | // polymorphic pattern because we don't own AAResult. It should |
468 | // ideally involve two pointers and no separate allocation. |
469 | AAs.emplace_back(new Model<AAResultT>(AAResult, *this)); |
470 | } |
471 | |
472 | /// Register a function analysis ID that the results aggregation depends on. |
473 | /// |
474 | /// This is used in the new pass manager to implement the invalidation logic |
475 | /// where we must invalidate the results aggregation if any of our component |
476 | /// analyses become invalid. |
477 | void addAADependencyID(AnalysisKey *ID) { AADeps.push_back(ID); } |
478 | |
479 | /// Handle invalidation events in the new pass manager. |
480 | /// |
481 | /// The aggregation is invalidated if any of the underlying analyses is |
482 | /// invalidated. |
483 | bool invalidate(Function &F, const PreservedAnalyses &PA, |
484 | FunctionAnalysisManager::Invalidator &Inv); |
485 | |
486 | //===--------------------------------------------------------------------===// |
487 | /// \name Alias Queries |
488 | /// @{ |
489 | |
490 | /// The main low level interface to the alias analysis implementation. |
491 | /// Returns an AliasResult indicating whether the two pointers are aliased to |
492 | /// each other. This is the interface that must be implemented by specific |
493 | /// alias analysis implementations. |
494 | AliasResult alias(const MemoryLocation &LocA, const MemoryLocation &LocB); |
495 | |
496 | /// A convenience wrapper around the primary \c alias interface. |
497 | AliasResult alias(const Value *V1, LocationSize V1Size, const Value *V2, |
498 | LocationSize V2Size) { |
499 | return alias(MemoryLocation(V1, V1Size), MemoryLocation(V2, V2Size)); |
500 | } |
501 | |
502 | /// A convenience wrapper around the primary \c alias interface. |
503 | AliasResult alias(const Value *V1, const Value *V2) { |
504 | return alias(MemoryLocation::getBeforeOrAfter(V1), |
505 | MemoryLocation::getBeforeOrAfter(V2)); |
506 | } |
507 | |
508 | /// A trivial helper function to check to see if the specified pointers are |
509 | /// no-alias. |
510 | bool isNoAlias(const MemoryLocation &LocA, const MemoryLocation &LocB) { |
511 | return alias(LocA, LocB) == AliasResult::NoAlias; |
512 | } |
513 | |
514 | /// A convenience wrapper around the \c isNoAlias helper interface. |
515 | bool isNoAlias(const Value *V1, LocationSize V1Size, const Value *V2, |
516 | LocationSize V2Size) { |
517 | return isNoAlias(MemoryLocation(V1, V1Size), MemoryLocation(V2, V2Size)); |
518 | } |
519 | |
520 | /// A convenience wrapper around the \c isNoAlias helper interface. |
521 | bool isNoAlias(const Value *V1, const Value *V2) { |
522 | return isNoAlias(MemoryLocation::getBeforeOrAfter(V1), |
523 | MemoryLocation::getBeforeOrAfter(V2)); |
524 | } |
525 | |
526 | /// A trivial helper function to check to see if the specified pointers are |
527 | /// must-alias. |
528 | bool isMustAlias(const MemoryLocation &LocA, const MemoryLocation &LocB) { |
529 | return alias(LocA, LocB) == AliasResult::MustAlias; |
530 | } |
531 | |
532 | /// A convenience wrapper around the \c isMustAlias helper interface. |
533 | bool isMustAlias(const Value *V1, const Value *V2) { |
534 | return alias(V1, LocationSize::precise(1), V2, LocationSize::precise(1)) == |
535 | AliasResult::MustAlias; |
536 | } |
537 | |
538 | /// Checks whether the given location points to constant memory, or if |
539 | /// \p OrLocal is true whether it points to a local alloca. |
540 | bool pointsToConstantMemory(const MemoryLocation &Loc, bool OrLocal = false); |
541 | |
542 | /// A convenience wrapper around the primary \c pointsToConstantMemory |
543 | /// interface. |
544 | bool pointsToConstantMemory(const Value *P, bool OrLocal = false) { |
545 | return pointsToConstantMemory(MemoryLocation::getBeforeOrAfter(P), OrLocal); |
546 | } |
547 | |
548 | /// @} |
549 | //===--------------------------------------------------------------------===// |
550 | /// \name Simple mod/ref information |
551 | /// @{ |
552 | |
553 | /// Get the ModRef info associated with a pointer argument of a call. The |
554 | /// result's bits are set to indicate the allowed aliasing ModRef kinds. Note |
555 | /// that these bits do not necessarily account for the overall behavior of |
556 | /// the function, but rather only provide additional per-argument |
557 | /// information. This never sets ModRefInfo::Must. |
558 | ModRefInfo getArgModRefInfo(const CallBase *Call, unsigned ArgIdx); |
559 | |
560 | /// Return the behavior of the given call site. |
561 | FunctionModRefBehavior getModRefBehavior(const CallBase *Call); |
562 | |
563 | /// Return the behavior when calling the given function. |
564 | FunctionModRefBehavior getModRefBehavior(const Function *F); |
565 | |
566 | /// Checks if the specified call is known to never read or write memory. |
567 | /// |
568 | /// Note that if the call only reads from known-constant memory, it is also |
569 | /// legal to return true. Also, calls that unwind the stack are legal for |
570 | /// this predicate. |
571 | /// |
572 | /// Many optimizations (such as CSE and LICM) can be performed on such calls |
573 | /// without worrying about aliasing properties, and many calls have this |
574 | /// property (e.g. calls to 'sin' and 'cos'). |
575 | /// |
576 | /// This property corresponds to the GCC 'const' attribute. |
577 | bool doesNotAccessMemory(const CallBase *Call) { |
578 | return getModRefBehavior(Call) == FMRB_DoesNotAccessMemory; |
579 | } |
580 | |
581 | /// Checks if the specified function is known to never read or write memory. |
582 | /// |
583 | /// Note that if the function only reads from known-constant memory, it is |
584 | /// also legal to return true. Also, function that unwind the stack are legal |
585 | /// for this predicate. |
586 | /// |
587 | /// Many optimizations (such as CSE and LICM) can be performed on such calls |
588 | /// to such functions without worrying about aliasing properties, and many |
589 | /// functions have this property (e.g. 'sin' and 'cos'). |
590 | /// |
591 | /// This property corresponds to the GCC 'const' attribute. |
592 | bool doesNotAccessMemory(const Function *F) { |
593 | return getModRefBehavior(F) == FMRB_DoesNotAccessMemory; |
594 | } |
595 | |
596 | /// Checks if the specified call is known to only read from non-volatile |
597 | /// memory (or not access memory at all). |
598 | /// |
599 | /// Calls that unwind the stack are legal for this predicate. |
600 | /// |
601 | /// This property allows many common optimizations to be performed in the |
602 | /// absence of interfering store instructions, such as CSE of strlen calls. |
603 | /// |
604 | /// This property corresponds to the GCC 'pure' attribute. |
605 | bool onlyReadsMemory(const CallBase *Call) { |
606 | return onlyReadsMemory(getModRefBehavior(Call)); |
607 | } |
608 | |
609 | /// Checks if the specified function is known to only read from non-volatile |
610 | /// memory (or not access memory at all). |
611 | /// |
612 | /// Functions that unwind the stack are legal for this predicate. |
613 | /// |
614 | /// This property allows many common optimizations to be performed in the |
615 | /// absence of interfering store instructions, such as CSE of strlen calls. |
616 | /// |
617 | /// This property corresponds to the GCC 'pure' attribute. |
618 | bool onlyReadsMemory(const Function *F) { |
619 | return onlyReadsMemory(getModRefBehavior(F)); |
620 | } |
621 | |
622 | /// Checks if functions with the specified behavior are known to only read |
623 | /// from non-volatile memory (or not access memory at all). |
624 | static bool onlyReadsMemory(FunctionModRefBehavior MRB) { |
625 | return !isModSet(createModRefInfo(MRB)); |
626 | } |
627 | |
628 | /// Checks if functions with the specified behavior are known to only write |
629 | /// memory (or not access memory at all). |
630 | static bool doesNotReadMemory(FunctionModRefBehavior MRB) { |
631 | return !isRefSet(createModRefInfo(MRB)); |
632 | } |
633 | |
634 | /// Checks if functions with the specified behavior are known to read and |
635 | /// write at most from objects pointed to by their pointer-typed arguments |
636 | /// (with arbitrary offsets). |
637 | static bool onlyAccessesArgPointees(FunctionModRefBehavior MRB) { |
638 | return !((unsigned)MRB & FMRL_Anywhere & ~FMRL_ArgumentPointees); |
639 | } |
640 | |
641 | /// Checks if functions with the specified behavior are known to potentially |
642 | /// read or write from objects pointed to be their pointer-typed arguments |
643 | /// (with arbitrary offsets). |
644 | static bool doesAccessArgPointees(FunctionModRefBehavior MRB) { |
645 | return isModOrRefSet(createModRefInfo(MRB)) && |
646 | ((unsigned)MRB & FMRL_ArgumentPointees); |
647 | } |
648 | |
649 | /// Checks if functions with the specified behavior are known to read and |
650 | /// write at most from memory that is inaccessible from LLVM IR. |
651 | static bool onlyAccessesInaccessibleMem(FunctionModRefBehavior MRB) { |
652 | return !((unsigned)MRB & FMRL_Anywhere & ~FMRL_InaccessibleMem); |
653 | } |
654 | |
655 | /// Checks if functions with the specified behavior are known to potentially |
656 | /// read or write from memory that is inaccessible from LLVM IR. |
657 | static bool doesAccessInaccessibleMem(FunctionModRefBehavior MRB) { |
658 | return isModOrRefSet(createModRefInfo(MRB)) && |
659 | ((unsigned)MRB & FMRL_InaccessibleMem); |
660 | } |
661 | |
662 | /// Checks if functions with the specified behavior are known to read and |
663 | /// write at most from memory that is inaccessible from LLVM IR or objects |
664 | /// pointed to by their pointer-typed arguments (with arbitrary offsets). |
665 | static bool onlyAccessesInaccessibleOrArgMem(FunctionModRefBehavior MRB) { |
666 | return !((unsigned)MRB & FMRL_Anywhere & |
667 | ~(FMRL_InaccessibleMem | FMRL_ArgumentPointees)); |
668 | } |
669 | |
670 | /// getModRefInfo (for call sites) - Return information about whether |
671 | /// a particular call site modifies or reads the specified memory location. |
672 | ModRefInfo getModRefInfo(const CallBase *Call, const MemoryLocation &Loc); |
673 | |
674 | /// getModRefInfo (for call sites) - A convenience wrapper. |
675 | ModRefInfo getModRefInfo(const CallBase *Call, const Value *P, |
676 | LocationSize Size) { |
677 | return getModRefInfo(Call, MemoryLocation(P, Size)); |
678 | } |
679 | |
680 | /// getModRefInfo (for loads) - Return information about whether |
681 | /// a particular load modifies or reads the specified memory location. |
682 | ModRefInfo getModRefInfo(const LoadInst *L, const MemoryLocation &Loc); |
683 | |
684 | /// getModRefInfo (for loads) - A convenience wrapper. |
685 | ModRefInfo getModRefInfo(const LoadInst *L, const Value *P, |
686 | LocationSize Size) { |
687 | return getModRefInfo(L, MemoryLocation(P, Size)); |
688 | } |
689 | |
690 | /// getModRefInfo (for stores) - Return information about whether |
691 | /// a particular store modifies or reads the specified memory location. |
692 | ModRefInfo getModRefInfo(const StoreInst *S, const MemoryLocation &Loc); |
693 | |
694 | /// getModRefInfo (for stores) - A convenience wrapper. |
695 | ModRefInfo getModRefInfo(const StoreInst *S, const Value *P, |
696 | LocationSize Size) { |
697 | return getModRefInfo(S, MemoryLocation(P, Size)); |
698 | } |
699 | |
700 | /// getModRefInfo (for fences) - Return information about whether |
701 | /// a particular store modifies or reads the specified memory location. |
702 | ModRefInfo getModRefInfo(const FenceInst *S, const MemoryLocation &Loc); |
703 | |
704 | /// getModRefInfo (for fences) - A convenience wrapper. |
705 | ModRefInfo getModRefInfo(const FenceInst *S, const Value *P, |
706 | LocationSize Size) { |
707 | return getModRefInfo(S, MemoryLocation(P, Size)); |
708 | } |
709 | |
710 | /// getModRefInfo (for cmpxchges) - Return information about whether |
711 | /// a particular cmpxchg modifies or reads the specified memory location. |
712 | ModRefInfo getModRefInfo(const AtomicCmpXchgInst *CX, |
713 | const MemoryLocation &Loc); |
714 | |
715 | /// getModRefInfo (for cmpxchges) - A convenience wrapper. |
716 | ModRefInfo getModRefInfo(const AtomicCmpXchgInst *CX, const Value *P, |
717 | LocationSize Size) { |
718 | return getModRefInfo(CX, MemoryLocation(P, Size)); |
719 | } |
720 | |
721 | /// getModRefInfo (for atomicrmws) - Return information about whether |
722 | /// a particular atomicrmw modifies or reads the specified memory location. |
723 | ModRefInfo getModRefInfo(const AtomicRMWInst *RMW, const MemoryLocation &Loc); |
724 | |
725 | /// getModRefInfo (for atomicrmws) - A convenience wrapper. |
726 | ModRefInfo getModRefInfo(const AtomicRMWInst *RMW, const Value *P, |
727 | LocationSize Size) { |
728 | return getModRefInfo(RMW, MemoryLocation(P, Size)); |
729 | } |
730 | |
731 | /// getModRefInfo (for va_args) - Return information about whether |
732 | /// a particular va_arg modifies or reads the specified memory location. |
733 | ModRefInfo getModRefInfo(const VAArgInst *I, const MemoryLocation &Loc); |
734 | |
735 | /// getModRefInfo (for va_args) - A convenience wrapper. |
736 | ModRefInfo getModRefInfo(const VAArgInst *I, const Value *P, |
737 | LocationSize Size) { |
738 | return getModRefInfo(I, MemoryLocation(P, Size)); |
739 | } |
740 | |
741 | /// getModRefInfo (for catchpads) - Return information about whether |
742 | /// a particular catchpad modifies or reads the specified memory location. |
743 | ModRefInfo getModRefInfo(const CatchPadInst *I, const MemoryLocation &Loc); |
744 | |
745 | /// getModRefInfo (for catchpads) - A convenience wrapper. |
746 | ModRefInfo getModRefInfo(const CatchPadInst *I, const Value *P, |
747 | LocationSize Size) { |
748 | return getModRefInfo(I, MemoryLocation(P, Size)); |
749 | } |
750 | |
751 | /// getModRefInfo (for catchrets) - Return information about whether |
752 | /// a particular catchret modifies or reads the specified memory location. |
753 | ModRefInfo getModRefInfo(const CatchReturnInst *I, const MemoryLocation &Loc); |
754 | |
755 | /// getModRefInfo (for catchrets) - A convenience wrapper. |
756 | ModRefInfo getModRefInfo(const CatchReturnInst *I, const Value *P, |
757 | LocationSize Size) { |
758 | return getModRefInfo(I, MemoryLocation(P, Size)); |
759 | } |
760 | |
761 | /// Check whether or not an instruction may read or write the optionally |
762 | /// specified memory location. |
763 | /// |
764 | /// |
765 | /// An instruction that doesn't read or write memory may be trivially LICM'd |
766 | /// for example. |
767 | /// |
768 | /// For function calls, this delegates to the alias-analysis specific |
769 | /// call-site mod-ref behavior queries. Otherwise it delegates to the specific |
770 | /// helpers above. |
771 | ModRefInfo getModRefInfo(const Instruction *I, |
772 | const Optional<MemoryLocation> &OptLoc) { |
773 | AAQueryInfo AAQIP; |
774 | return getModRefInfo(I, OptLoc, AAQIP); |
775 | } |
776 | |
777 | /// A convenience wrapper for constructing the memory location. |
778 | ModRefInfo getModRefInfo(const Instruction *I, const Value *P, |
779 | LocationSize Size) { |
780 | return getModRefInfo(I, MemoryLocation(P, Size)); |
781 | } |
782 | |
783 | /// Return information about whether a call and an instruction may refer to |
784 | /// the same memory locations. |
785 | ModRefInfo getModRefInfo(Instruction *I, const CallBase *Call); |
786 | |
787 | /// Return information about whether two call sites may refer to the same set |
788 | /// of memory locations. See the AA documentation for details: |
789 | /// http://llvm.org/docs/AliasAnalysis.html#ModRefInfo |
790 | ModRefInfo getModRefInfo(const CallBase *Call1, const CallBase *Call2); |
791 | |
792 | /// Return information about whether a particular call site modifies |
793 | /// or reads the specified memory location \p MemLoc before instruction \p I |
794 | /// in a BasicBlock. |
795 | /// Early exits in callCapturesBefore may lead to ModRefInfo::Must not being |
796 | /// set. |
797 | ModRefInfo callCapturesBefore(const Instruction *I, |
798 | const MemoryLocation &MemLoc, |
799 | DominatorTree *DT) { |
800 | AAQueryInfo AAQIP; |
801 | return callCapturesBefore(I, MemLoc, DT, AAQIP); |
802 | } |
803 | |
804 | /// A convenience wrapper to synthesize a memory location. |
805 | ModRefInfo callCapturesBefore(const Instruction *I, const Value *P, |
806 | LocationSize Size, DominatorTree *DT) { |
807 | return callCapturesBefore(I, MemoryLocation(P, Size), DT); |
808 | } |
809 | |
810 | /// @} |
811 | //===--------------------------------------------------------------------===// |
812 | /// \name Higher level methods for querying mod/ref information. |
813 | /// @{ |
814 | |
815 | /// Check if it is possible for execution of the specified basic block to |
816 | /// modify the location Loc. |
817 | bool canBasicBlockModify(const BasicBlock &BB, const MemoryLocation &Loc); |
818 | |
819 | /// A convenience wrapper synthesizing a memory location. |
820 | bool canBasicBlockModify(const BasicBlock &BB, const Value *P, |
821 | LocationSize Size) { |
822 | return canBasicBlockModify(BB, MemoryLocation(P, Size)); |
823 | } |
824 | |
825 | /// Check if it is possible for the execution of the specified instructions |
826 | /// to mod\ref (according to the mode) the location Loc. |
827 | /// |
828 | /// The instructions to consider are all of the instructions in the range of |
829 | /// [I1,I2] INCLUSIVE. I1 and I2 must be in the same basic block. |
830 | bool canInstructionRangeModRef(const Instruction &I1, const Instruction &I2, |
831 | const MemoryLocation &Loc, |
832 | const ModRefInfo Mode); |
833 | |
834 | /// A convenience wrapper synthesizing a memory location. |
835 | bool canInstructionRangeModRef(const Instruction &I1, const Instruction &I2, |
836 | const Value *Ptr, LocationSize Size, |
837 | const ModRefInfo Mode) { |
838 | return canInstructionRangeModRef(I1, I2, MemoryLocation(Ptr, Size), Mode); |
839 | } |
840 | |
841 | private: |
842 | AliasResult alias(const MemoryLocation &LocA, const MemoryLocation &LocB, |
843 | AAQueryInfo &AAQI); |
844 | bool pointsToConstantMemory(const MemoryLocation &Loc, AAQueryInfo &AAQI, |
845 | bool OrLocal = false); |
846 | ModRefInfo getModRefInfo(Instruction *I, const CallBase *Call2, |
847 | AAQueryInfo &AAQIP); |
848 | ModRefInfo getModRefInfo(const CallBase *Call, const MemoryLocation &Loc, |
849 | AAQueryInfo &AAQI); |
850 | ModRefInfo getModRefInfo(const CallBase *Call1, const CallBase *Call2, |
851 | AAQueryInfo &AAQI); |
852 | ModRefInfo getModRefInfo(const VAArgInst *V, const MemoryLocation &Loc, |
853 | AAQueryInfo &AAQI); |
854 | ModRefInfo getModRefInfo(const LoadInst *L, const MemoryLocation &Loc, |
855 | AAQueryInfo &AAQI); |
856 | ModRefInfo getModRefInfo(const StoreInst *S, const MemoryLocation &Loc, |
857 | AAQueryInfo &AAQI); |
858 | ModRefInfo getModRefInfo(const FenceInst *S, const MemoryLocation &Loc, |
859 | AAQueryInfo &AAQI); |
860 | ModRefInfo getModRefInfo(const AtomicCmpXchgInst *CX, |
861 | const MemoryLocation &Loc, AAQueryInfo &AAQI); |
862 | ModRefInfo getModRefInfo(const AtomicRMWInst *RMW, const MemoryLocation &Loc, |
863 | AAQueryInfo &AAQI); |
864 | ModRefInfo getModRefInfo(const CatchPadInst *I, const MemoryLocation &Loc, |
865 | AAQueryInfo &AAQI); |
866 | ModRefInfo getModRefInfo(const CatchReturnInst *I, const MemoryLocation &Loc, |
867 | AAQueryInfo &AAQI); |
868 | ModRefInfo getModRefInfo(const Instruction *I, |
869 | const Optional<MemoryLocation> &OptLoc, |
870 | AAQueryInfo &AAQIP); |
871 | ModRefInfo callCapturesBefore(const Instruction *I, |
872 | const MemoryLocation &MemLoc, DominatorTree *DT, |
873 | AAQueryInfo &AAQIP); |
874 | |
875 | class Concept; |
876 | |
877 | template <typename T> class Model; |
878 | |
879 | template <typename T> friend class AAResultBase; |
880 | |
881 | const TargetLibraryInfo &TLI; |
882 | |
883 | std::vector<std::unique_ptr<Concept>> AAs; |
884 | |
885 | std::vector<AnalysisKey *> AADeps; |
886 | |
887 | friend class BatchAAResults; |
888 | }; |
889 | |
890 | /// This class is a wrapper over an AAResults, and it is intended to be used |
891 | /// only when there are no IR changes inbetween queries. BatchAAResults is |
892 | /// reusing the same `AAQueryInfo` to preserve the state across queries, |
893 | /// esentially making AA work in "batch mode". The internal state cannot be |
894 | /// cleared, so to go "out-of-batch-mode", the user must either use AAResults, |
895 | /// or create a new BatchAAResults. |
896 | class BatchAAResults { |
897 | AAResults &AA; |
898 | AAQueryInfo AAQI; |
899 | |
900 | public: |
901 | BatchAAResults(AAResults &AAR) : AA(AAR), AAQI() {} |
902 | AliasResult alias(const MemoryLocation &LocA, const MemoryLocation &LocB) { |
903 | return AA.alias(LocA, LocB, AAQI); |
904 | } |
905 | bool pointsToConstantMemory(const MemoryLocation &Loc, bool OrLocal = false) { |
906 | return AA.pointsToConstantMemory(Loc, AAQI, OrLocal); |
907 | } |
908 | ModRefInfo getModRefInfo(const CallBase *Call, const MemoryLocation &Loc) { |
909 | return AA.getModRefInfo(Call, Loc, AAQI); |
910 | } |
911 | ModRefInfo getModRefInfo(const CallBase *Call1, const CallBase *Call2) { |
912 | return AA.getModRefInfo(Call1, Call2, AAQI); |
913 | } |
914 | ModRefInfo getModRefInfo(const Instruction *I, |
915 | const Optional<MemoryLocation> &OptLoc) { |
916 | return AA.getModRefInfo(I, OptLoc, AAQI); |
917 | } |
918 | ModRefInfo getModRefInfo(Instruction *I, const CallBase *Call2) { |
919 | return AA.getModRefInfo(I, Call2, AAQI); |
920 | } |
921 | ModRefInfo getArgModRefInfo(const CallBase *Call, unsigned ArgIdx) { |
922 | return AA.getArgModRefInfo(Call, ArgIdx); |
923 | } |
924 | FunctionModRefBehavior getModRefBehavior(const CallBase *Call) { |
925 | return AA.getModRefBehavior(Call); |
926 | } |
927 | bool isMustAlias(const MemoryLocation &LocA, const MemoryLocation &LocB) { |
928 | return alias(LocA, LocB) == AliasResult::MustAlias; |
929 | } |
930 | bool isMustAlias(const Value *V1, const Value *V2) { |
931 | return alias(MemoryLocation(V1, LocationSize::precise(1)), |
932 | MemoryLocation(V2, LocationSize::precise(1))) == |
933 | AliasResult::MustAlias; |
934 | } |
935 | ModRefInfo callCapturesBefore(const Instruction *I, |
936 | const MemoryLocation &MemLoc, |
937 | DominatorTree *DT) { |
938 | return AA.callCapturesBefore(I, MemLoc, DT, AAQI); |
939 | } |
940 | }; |
941 | |
942 | /// Temporary typedef for legacy code that uses a generic \c AliasAnalysis |
943 | /// pointer or reference. |
944 | using AliasAnalysis = AAResults; |
945 | |
946 | /// A private abstract base class describing the concept of an individual alias |
947 | /// analysis implementation. |
948 | /// |
949 | /// This interface is implemented by any \c Model instantiation. It is also the |
950 | /// interface which a type used to instantiate the model must provide. |
951 | /// |
952 | /// All of these methods model methods by the same name in the \c |
953 | /// AAResults class. Only differences and specifics to how the |
954 | /// implementations are called are documented here. |
955 | class AAResults::Concept { |
956 | public: |
957 | virtual ~Concept() = 0; |
958 | |
959 | /// An update API used internally by the AAResults to provide |
960 | /// a handle back to the top level aggregation. |
961 | virtual void setAAResults(AAResults *NewAAR) = 0; |
962 | |
963 | //===--------------------------------------------------------------------===// |
964 | /// \name Alias Queries |
965 | /// @{ |
966 | |
967 | /// The main low level interface to the alias analysis implementation. |
968 | /// Returns an AliasResult indicating whether the two pointers are aliased to |
969 | /// each other. This is the interface that must be implemented by specific |
970 | /// alias analysis implementations. |
971 | virtual AliasResult alias(const MemoryLocation &LocA, |
972 | const MemoryLocation &LocB, AAQueryInfo &AAQI) = 0; |
973 | |
974 | /// Checks whether the given location points to constant memory, or if |
975 | /// \p OrLocal is true whether it points to a local alloca. |
976 | virtual bool pointsToConstantMemory(const MemoryLocation &Loc, |
977 | AAQueryInfo &AAQI, bool OrLocal) = 0; |
978 | |
979 | /// @} |
980 | //===--------------------------------------------------------------------===// |
981 | /// \name Simple mod/ref information |
982 | /// @{ |
983 | |
984 | /// Get the ModRef info associated with a pointer argument of a callsite. The |
985 | /// result's bits are set to indicate the allowed aliasing ModRef kinds. Note |
986 | /// that these bits do not necessarily account for the overall behavior of |
987 | /// the function, but rather only provide additional per-argument |
988 | /// information. |
989 | virtual ModRefInfo getArgModRefInfo(const CallBase *Call, |
990 | unsigned ArgIdx) = 0; |
991 | |
992 | /// Return the behavior of the given call site. |
993 | virtual FunctionModRefBehavior getModRefBehavior(const CallBase *Call) = 0; |
994 | |
995 | /// Return the behavior when calling the given function. |
996 | virtual FunctionModRefBehavior getModRefBehavior(const Function *F) = 0; |
997 | |
998 | /// getModRefInfo (for call sites) - Return information about whether |
999 | /// a particular call site modifies or reads the specified memory location. |
1000 | virtual ModRefInfo getModRefInfo(const CallBase *Call, |
1001 | const MemoryLocation &Loc, |
1002 | AAQueryInfo &AAQI) = 0; |
1003 | |
1004 | /// Return information about whether two call sites may refer to the same set |
1005 | /// of memory locations. See the AA documentation for details: |
1006 | /// http://llvm.org/docs/AliasAnalysis.html#ModRefInfo |
1007 | virtual ModRefInfo getModRefInfo(const CallBase *Call1, const CallBase *Call2, |
1008 | AAQueryInfo &AAQI) = 0; |
1009 | |
1010 | /// @} |
1011 | }; |
1012 | |
1013 | /// A private class template which derives from \c Concept and wraps some other |
1014 | /// type. |
1015 | /// |
1016 | /// This models the concept by directly forwarding each interface point to the |
1017 | /// wrapped type which must implement a compatible interface. This provides |
1018 | /// a type erased binding. |
1019 | template <typename AAResultT> class AAResults::Model final : public Concept { |
1020 | AAResultT &Result; |
1021 | |
1022 | public: |
1023 | explicit Model(AAResultT &Result, AAResults &AAR) : Result(Result) { |
1024 | Result.setAAResults(&AAR); |
1025 | } |
1026 | ~Model() override = default; |
1027 | |
1028 | void setAAResults(AAResults *NewAAR) override { Result.setAAResults(NewAAR); } |
1029 | |
1030 | AliasResult alias(const MemoryLocation &LocA, const MemoryLocation &LocB, |
1031 | AAQueryInfo &AAQI) override { |
1032 | return Result.alias(LocA, LocB, AAQI); |
1033 | } |
1034 | |
1035 | bool pointsToConstantMemory(const MemoryLocation &Loc, AAQueryInfo &AAQI, |
1036 | bool OrLocal) override { |
1037 | return Result.pointsToConstantMemory(Loc, AAQI, OrLocal); |
1038 | } |
1039 | |
1040 | ModRefInfo getArgModRefInfo(const CallBase *Call, unsigned ArgIdx) override { |
1041 | return Result.getArgModRefInfo(Call, ArgIdx); |
1042 | } |
1043 | |
1044 | FunctionModRefBehavior getModRefBehavior(const CallBase *Call) override { |
1045 | return Result.getModRefBehavior(Call); |
1046 | } |
1047 | |
1048 | FunctionModRefBehavior getModRefBehavior(const Function *F) override { |
1049 | return Result.getModRefBehavior(F); |
1050 | } |
1051 | |
1052 | ModRefInfo getModRefInfo(const CallBase *Call, const MemoryLocation &Loc, |
1053 | AAQueryInfo &AAQI) override { |
1054 | return Result.getModRefInfo(Call, Loc, AAQI); |
1055 | } |
1056 | |
1057 | ModRefInfo getModRefInfo(const CallBase *Call1, const CallBase *Call2, |
1058 | AAQueryInfo &AAQI) override { |
1059 | return Result.getModRefInfo(Call1, Call2, AAQI); |
1060 | } |
1061 | }; |
1062 | |
1063 | /// A CRTP-driven "mixin" base class to help implement the function alias |
1064 | /// analysis results concept. |
1065 | /// |
1066 | /// Because of the nature of many alias analysis implementations, they often |
1067 | /// only implement a subset of the interface. This base class will attempt to |
1068 | /// implement the remaining portions of the interface in terms of simpler forms |
1069 | /// of the interface where possible, and otherwise provide conservatively |
1070 | /// correct fallback implementations. |
1071 | /// |
1072 | /// Implementors of an alias analysis should derive from this CRTP, and then |
1073 | /// override specific methods that they wish to customize. There is no need to |
1074 | /// use virtual anywhere, the CRTP base class does static dispatch to the |
1075 | /// derived type passed into it. |
1076 | template <typename DerivedT> class AAResultBase { |
1077 | // Expose some parts of the interface only to the AAResults::Model |
1078 | // for wrapping. Specifically, this allows the model to call our |
1079 | // setAAResults method without exposing it as a fully public API. |
1080 | friend class AAResults::Model<DerivedT>; |
1081 | |
1082 | /// A pointer to the AAResults object that this AAResult is |
1083 | /// aggregated within. May be null if not aggregated. |
1084 | AAResults *AAR = nullptr; |
1085 | |
1086 | /// Helper to dispatch calls back through the derived type. |
1087 | DerivedT &derived() { return static_cast<DerivedT &>(*this); } |
1088 | |
1089 | /// A setter for the AAResults pointer, which is used to satisfy the |
1090 | /// AAResults::Model contract. |
1091 | void setAAResults(AAResults *NewAAR) { AAR = NewAAR; } |
1092 | |
1093 | protected: |
1094 | /// This proxy class models a common pattern where we delegate to either the |
1095 | /// top-level \c AAResults aggregation if one is registered, or to the |
1096 | /// current result if none are registered. |
1097 | class AAResultsProxy { |
1098 | AAResults *AAR; |
1099 | DerivedT &CurrentResult; |
1100 | |
1101 | public: |
1102 | AAResultsProxy(AAResults *AAR, DerivedT &CurrentResult) |
1103 | : AAR(AAR), CurrentResult(CurrentResult) {} |
1104 | |
1105 | AliasResult alias(const MemoryLocation &LocA, const MemoryLocation &LocB, |
1106 | AAQueryInfo &AAQI) { |
1107 | return AAR ? AAR->alias(LocA, LocB, AAQI) |
1108 | : CurrentResult.alias(LocA, LocB, AAQI); |
1109 | } |
1110 | |
1111 | bool pointsToConstantMemory(const MemoryLocation &Loc, AAQueryInfo &AAQI, |
1112 | bool OrLocal) { |
1113 | return AAR ? AAR->pointsToConstantMemory(Loc, AAQI, OrLocal) |
1114 | : CurrentResult.pointsToConstantMemory(Loc, AAQI, OrLocal); |
1115 | } |
1116 | |
1117 | ModRefInfo getArgModRefInfo(const CallBase *Call, unsigned ArgIdx) { |
1118 | return AAR ? AAR->getArgModRefInfo(Call, ArgIdx) |
1119 | : CurrentResult.getArgModRefInfo(Call, ArgIdx); |
1120 | } |
1121 | |
1122 | FunctionModRefBehavior getModRefBehavior(const CallBase *Call) { |
1123 | return AAR ? AAR->getModRefBehavior(Call) |
1124 | : CurrentResult.getModRefBehavior(Call); |
1125 | } |
1126 | |
1127 | FunctionModRefBehavior getModRefBehavior(const Function *F) { |
1128 | return AAR ? AAR->getModRefBehavior(F) : CurrentResult.getModRefBehavior(F); |
1129 | } |
1130 | |
1131 | ModRefInfo getModRefInfo(const CallBase *Call, const MemoryLocation &Loc, |
1132 | AAQueryInfo &AAQI) { |
1133 | return AAR ? AAR->getModRefInfo(Call, Loc, AAQI) |
1134 | : CurrentResult.getModRefInfo(Call, Loc, AAQI); |
1135 | } |
1136 | |
1137 | ModRefInfo getModRefInfo(const CallBase *Call1, const CallBase *Call2, |
1138 | AAQueryInfo &AAQI) { |
1139 | return AAR ? AAR->getModRefInfo(Call1, Call2, AAQI) |
1140 | : CurrentResult.getModRefInfo(Call1, Call2, AAQI); |
1141 | } |
1142 | }; |
1143 | |
1144 | explicit AAResultBase() = default; |
1145 | |
1146 | // Provide all the copy and move constructors so that derived types aren't |
1147 | // constrained. |
1148 | AAResultBase(const AAResultBase &Arg) {} |
1149 | AAResultBase(AAResultBase &&Arg) {} |
1150 | |
1151 | /// Get a proxy for the best AA result set to query at this time. |
1152 | /// |
1153 | /// When this result is part of a larger aggregation, this will proxy to that |
1154 | /// aggregation. When this result is used in isolation, it will just delegate |
1155 | /// back to the derived class's implementation. |
1156 | /// |
1157 | /// Note that callers of this need to take considerable care to not cause |
1158 | /// performance problems when they use this routine, in the case of a large |
1159 | /// number of alias analyses being aggregated, it can be expensive to walk |
1160 | /// back across the chain. |
1161 | AAResultsProxy getBestAAResults() { return AAResultsProxy(AAR, derived()); } |
1162 | |
1163 | public: |
1164 | AliasResult alias(const MemoryLocation &LocA, const MemoryLocation &LocB, |
1165 | AAQueryInfo &AAQI) { |
1166 | return AliasResult::MayAlias; |
1167 | } |
1168 | |
1169 | bool pointsToConstantMemory(const MemoryLocation &Loc, AAQueryInfo &AAQI, |
1170 | bool OrLocal) { |
1171 | return false; |
1172 | } |
1173 | |
1174 | ModRefInfo getArgModRefInfo(const CallBase *Call, unsigned ArgIdx) { |
1175 | return ModRefInfo::ModRef; |
1176 | } |
1177 | |
1178 | FunctionModRefBehavior getModRefBehavior(const CallBase *Call) { |
1179 | return FMRB_UnknownModRefBehavior; |
1180 | } |
1181 | |
1182 | FunctionModRefBehavior getModRefBehavior(const Function *F) { |
1183 | return FMRB_UnknownModRefBehavior; |
1184 | } |
1185 | |
1186 | ModRefInfo getModRefInfo(const CallBase *Call, const MemoryLocation &Loc, |
1187 | AAQueryInfo &AAQI) { |
1188 | return ModRefInfo::ModRef; |
1189 | } |
1190 | |
1191 | ModRefInfo getModRefInfo(const CallBase *Call1, const CallBase *Call2, |
1192 | AAQueryInfo &AAQI) { |
1193 | return ModRefInfo::ModRef; |
1194 | } |
1195 | }; |
1196 | |
1197 | /// Return true if this pointer is returned by a noalias function. |
1198 | bool isNoAliasCall(const Value *V); |
1199 | |
1200 | /// Return true if this pointer refers to a distinct and identifiable object. |
1201 | /// This returns true for: |
1202 | /// Global Variables and Functions (but not Global Aliases) |
1203 | /// Allocas |
1204 | /// ByVal and NoAlias Arguments |
1205 | /// NoAlias returns (e.g. calls to malloc) |
1206 | /// |
1207 | bool isIdentifiedObject(const Value *V); |
1208 | |
1209 | /// Return true if V is umabigously identified at the function-level. |
1210 | /// Different IdentifiedFunctionLocals can't alias. |
1211 | /// Further, an IdentifiedFunctionLocal can not alias with any function |
1212 | /// arguments other than itself, which is not necessarily true for |
1213 | /// IdentifiedObjects. |
1214 | bool isIdentifiedFunctionLocal(const Value *V); |
1215 | |
1216 | /// A manager for alias analyses. |
1217 | /// |
1218 | /// This class can have analyses registered with it and when run, it will run |
1219 | /// all of them and aggregate their results into single AA results interface |
1220 | /// that dispatches across all of the alias analysis results available. |
1221 | /// |
1222 | /// Note that the order in which analyses are registered is very significant. |
1223 | /// That is the order in which the results will be aggregated and queried. |
1224 | /// |
1225 | /// This manager effectively wraps the AnalysisManager for registering alias |
1226 | /// analyses. When you register your alias analysis with this manager, it will |
1227 | /// ensure the analysis itself is registered with its AnalysisManager. |
1228 | /// |
1229 | /// The result of this analysis is only invalidated if one of the particular |
1230 | /// aggregated AA results end up being invalidated. This removes the need to |
1231 | /// explicitly preserve the results of `AAManager`. Note that analyses should no |
1232 | /// longer be registered once the `AAManager` is run. |
1233 | class AAManager : public AnalysisInfoMixin<AAManager> { |
1234 | public: |
1235 | using Result = AAResults; |
1236 | |
1237 | /// Register a specific AA result. |
1238 | template <typename AnalysisT> void registerFunctionAnalysis() { |
1239 | ResultGetters.push_back(&getFunctionAAResultImpl<AnalysisT>); |
1240 | } |
1241 | |
1242 | /// Register a specific AA result. |
1243 | template <typename AnalysisT> void registerModuleAnalysis() { |
1244 | ResultGetters.push_back(&getModuleAAResultImpl<AnalysisT>); |
1245 | } |
1246 | |
1247 | Result run(Function &F, FunctionAnalysisManager &AM); |
1248 | |
1249 | private: |
1250 | friend AnalysisInfoMixin<AAManager>; |
1251 | |
1252 | static AnalysisKey Key; |
1253 | |
1254 | SmallVector<void (*)(Function &F, FunctionAnalysisManager &AM, |
1255 | AAResults &AAResults), |
1256 | 4> ResultGetters; |
1257 | |
1258 | template <typename AnalysisT> |
1259 | static void getFunctionAAResultImpl(Function &F, |
1260 | FunctionAnalysisManager &AM, |
1261 | AAResults &AAResults) { |
1262 | AAResults.addAAResult(AM.template getResult<AnalysisT>(F)); |
1263 | AAResults.addAADependencyID(AnalysisT::ID()); |
1264 | } |
1265 | |
1266 | template <typename AnalysisT> |
1267 | static void getModuleAAResultImpl(Function &F, FunctionAnalysisManager &AM, |
1268 | AAResults &AAResults) { |
1269 | auto &MAMProxy = AM.getResult<ModuleAnalysisManagerFunctionProxy>(F); |
1270 | if (auto *R = |
1271 | MAMProxy.template getCachedResult<AnalysisT>(*F.getParent())) { |
1272 | AAResults.addAAResult(*R); |
1273 | MAMProxy |
1274 | .template registerOuterAnalysisInvalidation<AnalysisT, AAManager>(); |
1275 | } |
1276 | } |
1277 | }; |
1278 | |
1279 | /// A wrapper pass to provide the legacy pass manager access to a suitably |
1280 | /// prepared AAResults object. |
1281 | class AAResultsWrapperPass : public FunctionPass { |
1282 | std::unique_ptr<AAResults> AAR; |
1283 | |
1284 | public: |
1285 | static char ID; |
1286 | |
1287 | AAResultsWrapperPass(); |
1288 | |
1289 | AAResults &getAAResults() { return *AAR; } |
1290 | const AAResults &getAAResults() const { return *AAR; } |
1291 | |
1292 | bool runOnFunction(Function &F) override; |
1293 | |
1294 | void getAnalysisUsage(AnalysisUsage &AU) const override; |
1295 | }; |
1296 | |
1297 | /// A wrapper pass for external alias analyses. This just squirrels away the |
1298 | /// callback used to run any analyses and register their results. |
1299 | struct ExternalAAWrapperPass : ImmutablePass { |
1300 | using CallbackT = std::function<void(Pass &, Function &, AAResults &)>; |
1301 | |
1302 | CallbackT CB; |
1303 | |
1304 | static char ID; |
1305 | |
1306 | ExternalAAWrapperPass(); |
1307 | |
1308 | explicit ExternalAAWrapperPass(CallbackT CB); |
1309 | |
1310 | void getAnalysisUsage(AnalysisUsage &AU) const override { |
1311 | AU.setPreservesAll(); |
1312 | } |
1313 | }; |
1314 | |
1315 | FunctionPass *createAAResultsWrapperPass(); |
1316 | |
1317 | /// A wrapper pass around a callback which can be used to populate the |
1318 | /// AAResults in the AAResultsWrapperPass from an external AA. |
1319 | /// |
1320 | /// The callback provided here will be used each time we prepare an AAResults |
1321 | /// object, and will receive a reference to the function wrapper pass, the |
1322 | /// function, and the AAResults object to populate. This should be used when |
1323 | /// setting up a custom pass pipeline to inject a hook into the AA results. |
1324 | ImmutablePass *createExternalAAWrapperPass( |
1325 | std::function<void(Pass &, Function &, AAResults &)> Callback); |
1326 | |
1327 | /// A helper for the legacy pass manager to create a \c AAResults |
1328 | /// object populated to the best of our ability for a particular function when |
1329 | /// inside of a \c ModulePass or a \c CallGraphSCCPass. |
1330 | /// |
1331 | /// If a \c ModulePass or a \c CallGraphSCCPass calls \p |
1332 | /// createLegacyPMAAResults, it also needs to call \p addUsedAAAnalyses in \p |
1333 | /// getAnalysisUsage. |
1334 | AAResults createLegacyPMAAResults(Pass &P, Function &F, BasicAAResult &BAR); |
1335 | |
1336 | /// A helper for the legacy pass manager to populate \p AU to add uses to make |
1337 | /// sure the analyses required by \p createLegacyPMAAResults are available. |
1338 | void getAAResultsAnalysisUsage(AnalysisUsage &AU); |
1339 | |
1340 | } // end namespace llvm |
1341 | |
1342 | #endif // LLVM_ANALYSIS_ALIASANALYSIS_H |