| File: | src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/lib/Analysis/LazyValueInfo.cpp |
| Warning: | line 1322, column 34 Called C++ object pointer is null |
Press '?' to see keyboard shortcuts
Keyboard shortcuts:
| 1 | //===- LazyValueInfo.cpp - Value constraint analysis ------------*- 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 interface for lazy computation of value constraint | ||||||||
| 10 | // information. | ||||||||
| 11 | // | ||||||||
| 12 | //===----------------------------------------------------------------------===// | ||||||||
| 13 | |||||||||
| 14 | #include "llvm/Analysis/LazyValueInfo.h" | ||||||||
| 15 | #include "llvm/ADT/DenseSet.h" | ||||||||
| 16 | #include "llvm/ADT/Optional.h" | ||||||||
| 17 | #include "llvm/ADT/STLExtras.h" | ||||||||
| 18 | #include "llvm/Analysis/AssumptionCache.h" | ||||||||
| 19 | #include "llvm/Analysis/ConstantFolding.h" | ||||||||
| 20 | #include "llvm/Analysis/InstructionSimplify.h" | ||||||||
| 21 | #include "llvm/Analysis/TargetLibraryInfo.h" | ||||||||
| 22 | #include "llvm/Analysis/ValueLattice.h" | ||||||||
| 23 | #include "llvm/Analysis/ValueTracking.h" | ||||||||
| 24 | #include "llvm/IR/AssemblyAnnotationWriter.h" | ||||||||
| 25 | #include "llvm/IR/CFG.h" | ||||||||
| 26 | #include "llvm/IR/ConstantRange.h" | ||||||||
| 27 | #include "llvm/IR/Constants.h" | ||||||||
| 28 | #include "llvm/IR/DataLayout.h" | ||||||||
| 29 | #include "llvm/IR/Dominators.h" | ||||||||
| 30 | #include "llvm/IR/Instructions.h" | ||||||||
| 31 | #include "llvm/IR/IntrinsicInst.h" | ||||||||
| 32 | #include "llvm/IR/Intrinsics.h" | ||||||||
| 33 | #include "llvm/IR/LLVMContext.h" | ||||||||
| 34 | #include "llvm/IR/PatternMatch.h" | ||||||||
| 35 | #include "llvm/IR/ValueHandle.h" | ||||||||
| 36 | #include "llvm/InitializePasses.h" | ||||||||
| 37 | #include "llvm/Support/Debug.h" | ||||||||
| 38 | #include "llvm/Support/FormattedStream.h" | ||||||||
| 39 | #include "llvm/Support/KnownBits.h" | ||||||||
| 40 | #include "llvm/Support/raw_ostream.h" | ||||||||
| 41 | #include <map> | ||||||||
| 42 | using namespace llvm; | ||||||||
| 43 | using namespace PatternMatch; | ||||||||
| 44 | |||||||||
| 45 | #define DEBUG_TYPE"lazy-value-info" "lazy-value-info" | ||||||||
| 46 | |||||||||
| 47 | // This is the number of worklist items we will process to try to discover an | ||||||||
| 48 | // answer for a given value. | ||||||||
| 49 | static const unsigned MaxProcessedPerValue = 500; | ||||||||
| 50 | |||||||||
| 51 | char LazyValueInfoWrapperPass::ID = 0; | ||||||||
| 52 | LazyValueInfoWrapperPass::LazyValueInfoWrapperPass() : FunctionPass(ID) { | ||||||||
| 53 | initializeLazyValueInfoWrapperPassPass(*PassRegistry::getPassRegistry()); | ||||||||
| 54 | } | ||||||||
| 55 | INITIALIZE_PASS_BEGIN(LazyValueInfoWrapperPass, "lazy-value-info",static void *initializeLazyValueInfoWrapperPassPassOnce(PassRegistry &Registry) { | ||||||||
| 56 | "Lazy Value Information Analysis", false, true)static void *initializeLazyValueInfoWrapperPassPassOnce(PassRegistry &Registry) { | ||||||||
| 57 | INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)initializeAssumptionCacheTrackerPass(Registry); | ||||||||
| 58 | INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)initializeTargetLibraryInfoWrapperPassPass(Registry); | ||||||||
| 59 | INITIALIZE_PASS_END(LazyValueInfoWrapperPass, "lazy-value-info",PassInfo *PI = new PassInfo( "Lazy Value Information Analysis" , "lazy-value-info", &LazyValueInfoWrapperPass::ID, PassInfo ::NormalCtor_t(callDefaultCtor<LazyValueInfoWrapperPass> ), false, true); Registry.registerPass(*PI, true); return PI; } static llvm::once_flag InitializeLazyValueInfoWrapperPassPassFlag ; void llvm::initializeLazyValueInfoWrapperPassPass(PassRegistry &Registry) { llvm::call_once(InitializeLazyValueInfoWrapperPassPassFlag , initializeLazyValueInfoWrapperPassPassOnce, std::ref(Registry )); } | ||||||||
| 60 | "Lazy Value Information Analysis", false, true)PassInfo *PI = new PassInfo( "Lazy Value Information Analysis" , "lazy-value-info", &LazyValueInfoWrapperPass::ID, PassInfo ::NormalCtor_t(callDefaultCtor<LazyValueInfoWrapperPass> ), false, true); Registry.registerPass(*PI, true); return PI; } static llvm::once_flag InitializeLazyValueInfoWrapperPassPassFlag ; void llvm::initializeLazyValueInfoWrapperPassPass(PassRegistry &Registry) { llvm::call_once(InitializeLazyValueInfoWrapperPassPassFlag , initializeLazyValueInfoWrapperPassPassOnce, std::ref(Registry )); } | ||||||||
| 61 | |||||||||
| 62 | namespace llvm { | ||||||||
| 63 | FunctionPass *createLazyValueInfoPass() { return new LazyValueInfoWrapperPass(); } | ||||||||
| 64 | } | ||||||||
| 65 | |||||||||
| 66 | AnalysisKey LazyValueAnalysis::Key; | ||||||||
| 67 | |||||||||
| 68 | /// Returns true if this lattice value represents at most one possible value. | ||||||||
| 69 | /// This is as precise as any lattice value can get while still representing | ||||||||
| 70 | /// reachable code. | ||||||||
| 71 | static bool hasSingleValue(const ValueLatticeElement &Val) { | ||||||||
| 72 | if (Val.isConstantRange() && | ||||||||
| 73 | Val.getConstantRange().isSingleElement()) | ||||||||
| 74 | // Integer constants are single element ranges | ||||||||
| 75 | return true; | ||||||||
| 76 | if (Val.isConstant()) | ||||||||
| 77 | // Non integer constants | ||||||||
| 78 | return true; | ||||||||
| 79 | return false; | ||||||||
| 80 | } | ||||||||
| 81 | |||||||||
| 82 | /// Combine two sets of facts about the same value into a single set of | ||||||||
| 83 | /// facts. Note that this method is not suitable for merging facts along | ||||||||
| 84 | /// different paths in a CFG; that's what the mergeIn function is for. This | ||||||||
| 85 | /// is for merging facts gathered about the same value at the same location | ||||||||
| 86 | /// through two independent means. | ||||||||
| 87 | /// Notes: | ||||||||
| 88 | /// * This method does not promise to return the most precise possible lattice | ||||||||
| 89 | /// value implied by A and B. It is allowed to return any lattice element | ||||||||
| 90 | /// which is at least as strong as *either* A or B (unless our facts | ||||||||
| 91 | /// conflict, see below). | ||||||||
| 92 | /// * Due to unreachable code, the intersection of two lattice values could be | ||||||||
| 93 | /// contradictory. If this happens, we return some valid lattice value so as | ||||||||
| 94 | /// not confuse the rest of LVI. Ideally, we'd always return Undefined, but | ||||||||
| 95 | /// we do not make this guarantee. TODO: This would be a useful enhancement. | ||||||||
| 96 | static ValueLatticeElement intersect(const ValueLatticeElement &A, | ||||||||
| 97 | const ValueLatticeElement &B) { | ||||||||
| 98 | // Undefined is the strongest state. It means the value is known to be along | ||||||||
| 99 | // an unreachable path. | ||||||||
| 100 | if (A.isUnknown()) | ||||||||
| 101 | return A; | ||||||||
| 102 | if (B.isUnknown()) | ||||||||
| 103 | return B; | ||||||||
| 104 | |||||||||
| 105 | // If we gave up for one, but got a useable fact from the other, use it. | ||||||||
| 106 | if (A.isOverdefined()) | ||||||||
| 107 | return B; | ||||||||
| 108 | if (B.isOverdefined()) | ||||||||
| 109 | return A; | ||||||||
| 110 | |||||||||
| 111 | // Can't get any more precise than constants. | ||||||||
| 112 | if (hasSingleValue(A)) | ||||||||
| 113 | return A; | ||||||||
| 114 | if (hasSingleValue(B)) | ||||||||
| 115 | return B; | ||||||||
| 116 | |||||||||
| 117 | // Could be either constant range or not constant here. | ||||||||
| 118 | if (!A.isConstantRange() || !B.isConstantRange()) { | ||||||||
| 119 | // TODO: Arbitrary choice, could be improved | ||||||||
| 120 | return A; | ||||||||
| 121 | } | ||||||||
| 122 | |||||||||
| 123 | // Intersect two constant ranges | ||||||||
| 124 | ConstantRange Range = | ||||||||
| 125 | A.getConstantRange().intersectWith(B.getConstantRange()); | ||||||||
| 126 | // Note: An empty range is implicitly converted to unknown or undef depending | ||||||||
| 127 | // on MayIncludeUndef internally. | ||||||||
| 128 | return ValueLatticeElement::getRange( | ||||||||
| 129 | std::move(Range), /*MayIncludeUndef=*/A.isConstantRangeIncludingUndef() | | ||||||||
| 130 | B.isConstantRangeIncludingUndef()); | ||||||||
| 131 | } | ||||||||
| 132 | |||||||||
| 133 | //===----------------------------------------------------------------------===// | ||||||||
| 134 | // LazyValueInfoCache Decl | ||||||||
| 135 | //===----------------------------------------------------------------------===// | ||||||||
| 136 | |||||||||
| 137 | namespace { | ||||||||
| 138 | /// A callback value handle updates the cache when values are erased. | ||||||||
| 139 | class LazyValueInfoCache; | ||||||||
| 140 | struct LVIValueHandle final : public CallbackVH { | ||||||||
| 141 | LazyValueInfoCache *Parent; | ||||||||
| 142 | |||||||||
| 143 | LVIValueHandle(Value *V, LazyValueInfoCache *P = nullptr) | ||||||||
| 144 | : CallbackVH(V), Parent(P) { } | ||||||||
| 145 | |||||||||
| 146 | void deleted() override; | ||||||||
| 147 | void allUsesReplacedWith(Value *V) override { | ||||||||
| 148 | deleted(); | ||||||||
| 149 | } | ||||||||
| 150 | }; | ||||||||
| 151 | } // end anonymous namespace | ||||||||
| 152 | |||||||||
| 153 | namespace { | ||||||||
| 154 | using NonNullPointerSet = SmallDenseSet<AssertingVH<Value>, 2>; | ||||||||
| 155 | |||||||||
| 156 | /// This is the cache kept by LazyValueInfo which | ||||||||
| 157 | /// maintains information about queries across the clients' queries. | ||||||||
| 158 | class LazyValueInfoCache { | ||||||||
| 159 | /// This is all of the cached information for one basic block. It contains | ||||||||
| 160 | /// the per-value lattice elements, as well as a separate set for | ||||||||
| 161 | /// overdefined values to reduce memory usage. Additionally pointers | ||||||||
| 162 | /// dereferenced in the block are cached for nullability queries. | ||||||||
| 163 | struct BlockCacheEntry { | ||||||||
| 164 | SmallDenseMap<AssertingVH<Value>, ValueLatticeElement, 4> LatticeElements; | ||||||||
| 165 | SmallDenseSet<AssertingVH<Value>, 4> OverDefined; | ||||||||
| 166 | // None indicates that the nonnull pointers for this basic block | ||||||||
| 167 | // block have not been computed yet. | ||||||||
| 168 | Optional<NonNullPointerSet> NonNullPointers; | ||||||||
| 169 | }; | ||||||||
| 170 | |||||||||
| 171 | /// Cached information per basic block. | ||||||||
| 172 | DenseMap<PoisoningVH<BasicBlock>, std::unique_ptr<BlockCacheEntry>> | ||||||||
| 173 | BlockCache; | ||||||||
| 174 | /// Set of value handles used to erase values from the cache on deletion. | ||||||||
| 175 | DenseSet<LVIValueHandle, DenseMapInfo<Value *>> ValueHandles; | ||||||||
| 176 | |||||||||
| 177 | const BlockCacheEntry *getBlockEntry(BasicBlock *BB) const { | ||||||||
| 178 | auto It = BlockCache.find_as(BB); | ||||||||
| 179 | if (It == BlockCache.end()) | ||||||||
| 180 | return nullptr; | ||||||||
| 181 | return It->second.get(); | ||||||||
| 182 | } | ||||||||
| 183 | |||||||||
| 184 | BlockCacheEntry *getOrCreateBlockEntry(BasicBlock *BB) { | ||||||||
| 185 | auto It = BlockCache.find_as(BB); | ||||||||
| 186 | if (It == BlockCache.end()) | ||||||||
| 187 | It = BlockCache.insert({ BB, std::make_unique<BlockCacheEntry>() }) | ||||||||
| 188 | .first; | ||||||||
| 189 | |||||||||
| 190 | return It->second.get(); | ||||||||
| 191 | } | ||||||||
| 192 | |||||||||
| 193 | void addValueHandle(Value *Val) { | ||||||||
| 194 | auto HandleIt = ValueHandles.find_as(Val); | ||||||||
| 195 | if (HandleIt == ValueHandles.end()) | ||||||||
| 196 | ValueHandles.insert({ Val, this }); | ||||||||
| 197 | } | ||||||||
| 198 | |||||||||
| 199 | public: | ||||||||
| 200 | void insertResult(Value *Val, BasicBlock *BB, | ||||||||
| 201 | const ValueLatticeElement &Result) { | ||||||||
| 202 | BlockCacheEntry *Entry = getOrCreateBlockEntry(BB); | ||||||||
| 203 | |||||||||
| 204 | // Insert over-defined values into their own cache to reduce memory | ||||||||
| 205 | // overhead. | ||||||||
| 206 | if (Result.isOverdefined()) | ||||||||
| 207 | Entry->OverDefined.insert(Val); | ||||||||
| 208 | else | ||||||||
| 209 | Entry->LatticeElements.insert({ Val, Result }); | ||||||||
| 210 | |||||||||
| 211 | addValueHandle(Val); | ||||||||
| 212 | } | ||||||||
| 213 | |||||||||
| 214 | Optional<ValueLatticeElement> getCachedValueInfo(Value *V, | ||||||||
| 215 | BasicBlock *BB) const { | ||||||||
| 216 | const BlockCacheEntry *Entry = getBlockEntry(BB); | ||||||||
| 217 | if (!Entry) | ||||||||
| 218 | return None; | ||||||||
| 219 | |||||||||
| 220 | if (Entry->OverDefined.count(V)) | ||||||||
| 221 | return ValueLatticeElement::getOverdefined(); | ||||||||
| 222 | |||||||||
| 223 | auto LatticeIt = Entry->LatticeElements.find_as(V); | ||||||||
| 224 | if (LatticeIt == Entry->LatticeElements.end()) | ||||||||
| 225 | return None; | ||||||||
| 226 | |||||||||
| 227 | return LatticeIt->second; | ||||||||
| 228 | } | ||||||||
| 229 | |||||||||
| 230 | bool isNonNullAtEndOfBlock( | ||||||||
| 231 | Value *V, BasicBlock *BB, | ||||||||
| 232 | function_ref<NonNullPointerSet(BasicBlock *)> InitFn) { | ||||||||
| 233 | BlockCacheEntry *Entry = getOrCreateBlockEntry(BB); | ||||||||
| 234 | if (!Entry->NonNullPointers) { | ||||||||
| 235 | Entry->NonNullPointers = InitFn(BB); | ||||||||
| 236 | for (Value *V : *Entry->NonNullPointers) | ||||||||
| 237 | addValueHandle(V); | ||||||||
| 238 | } | ||||||||
| 239 | |||||||||
| 240 | return Entry->NonNullPointers->count(V); | ||||||||
| 241 | } | ||||||||
| 242 | |||||||||
| 243 | /// clear - Empty the cache. | ||||||||
| 244 | void clear() { | ||||||||
| 245 | BlockCache.clear(); | ||||||||
| 246 | ValueHandles.clear(); | ||||||||
| 247 | } | ||||||||
| 248 | |||||||||
| 249 | /// Inform the cache that a given value has been deleted. | ||||||||
| 250 | void eraseValue(Value *V); | ||||||||
| 251 | |||||||||
| 252 | /// This is part of the update interface to inform the cache | ||||||||
| 253 | /// that a block has been deleted. | ||||||||
| 254 | void eraseBlock(BasicBlock *BB); | ||||||||
| 255 | |||||||||
| 256 | /// Updates the cache to remove any influence an overdefined value in | ||||||||
| 257 | /// OldSucc might have (unless also overdefined in NewSucc). This just | ||||||||
| 258 | /// flushes elements from the cache and does not add any. | ||||||||
| 259 | void threadEdgeImpl(BasicBlock *OldSucc,BasicBlock *NewSucc); | ||||||||
| 260 | }; | ||||||||
| 261 | } | ||||||||
| 262 | |||||||||
| 263 | void LazyValueInfoCache::eraseValue(Value *V) { | ||||||||
| 264 | for (auto &Pair : BlockCache) { | ||||||||
| 265 | Pair.second->LatticeElements.erase(V); | ||||||||
| 266 | Pair.second->OverDefined.erase(V); | ||||||||
| 267 | if (Pair.second->NonNullPointers) | ||||||||
| 268 | Pair.second->NonNullPointers->erase(V); | ||||||||
| 269 | } | ||||||||
| 270 | |||||||||
| 271 | auto HandleIt = ValueHandles.find_as(V); | ||||||||
| 272 | if (HandleIt != ValueHandles.end()) | ||||||||
| 273 | ValueHandles.erase(HandleIt); | ||||||||
| 274 | } | ||||||||
| 275 | |||||||||
| 276 | void LVIValueHandle::deleted() { | ||||||||
| 277 | // This erasure deallocates *this, so it MUST happen after we're done | ||||||||
| 278 | // using any and all members of *this. | ||||||||
| 279 | Parent->eraseValue(*this); | ||||||||
| 280 | } | ||||||||
| 281 | |||||||||
| 282 | void LazyValueInfoCache::eraseBlock(BasicBlock *BB) { | ||||||||
| 283 | BlockCache.erase(BB); | ||||||||
| 284 | } | ||||||||
| 285 | |||||||||
| 286 | void LazyValueInfoCache::threadEdgeImpl(BasicBlock *OldSucc, | ||||||||
| 287 | BasicBlock *NewSucc) { | ||||||||
| 288 | // When an edge in the graph has been threaded, values that we could not | ||||||||
| 289 | // determine a value for before (i.e. were marked overdefined) may be | ||||||||
| 290 | // possible to solve now. We do NOT try to proactively update these values. | ||||||||
| 291 | // Instead, we clear their entries from the cache, and allow lazy updating to | ||||||||
| 292 | // recompute them when needed. | ||||||||
| 293 | |||||||||
| 294 | // The updating process is fairly simple: we need to drop cached info | ||||||||
| 295 | // for all values that were marked overdefined in OldSucc, and for those same | ||||||||
| 296 | // values in any successor of OldSucc (except NewSucc) in which they were | ||||||||
| 297 | // also marked overdefined. | ||||||||
| 298 | std::vector<BasicBlock*> worklist; | ||||||||
| 299 | worklist.push_back(OldSucc); | ||||||||
| 300 | |||||||||
| 301 | const BlockCacheEntry *Entry = getBlockEntry(OldSucc); | ||||||||
| 302 | if (!Entry || Entry->OverDefined.empty()) | ||||||||
| 303 | return; // Nothing to process here. | ||||||||
| 304 | SmallVector<Value *, 4> ValsToClear(Entry->OverDefined.begin(), | ||||||||
| 305 | Entry->OverDefined.end()); | ||||||||
| 306 | |||||||||
| 307 | // Use a worklist to perform a depth-first search of OldSucc's successors. | ||||||||
| 308 | // NOTE: We do not need a visited list since any blocks we have already | ||||||||
| 309 | // visited will have had their overdefined markers cleared already, and we | ||||||||
| 310 | // thus won't loop to their successors. | ||||||||
| 311 | while (!worklist.empty()) { | ||||||||
| 312 | BasicBlock *ToUpdate = worklist.back(); | ||||||||
| 313 | worklist.pop_back(); | ||||||||
| 314 | |||||||||
| 315 | // Skip blocks only accessible through NewSucc. | ||||||||
| 316 | if (ToUpdate == NewSucc) continue; | ||||||||
| 317 | |||||||||
| 318 | // If a value was marked overdefined in OldSucc, and is here too... | ||||||||
| 319 | auto OI = BlockCache.find_as(ToUpdate); | ||||||||
| 320 | if (OI == BlockCache.end() || OI->second->OverDefined.empty()) | ||||||||
| 321 | continue; | ||||||||
| 322 | auto &ValueSet = OI->second->OverDefined; | ||||||||
| 323 | |||||||||
| 324 | bool changed = false; | ||||||||
| 325 | for (Value *V : ValsToClear) { | ||||||||
| 326 | if (!ValueSet.erase(V)) | ||||||||
| 327 | continue; | ||||||||
| 328 | |||||||||
| 329 | // If we removed anything, then we potentially need to update | ||||||||
| 330 | // blocks successors too. | ||||||||
| 331 | changed = true; | ||||||||
| 332 | } | ||||||||
| 333 | |||||||||
| 334 | if (!changed) continue; | ||||||||
| 335 | |||||||||
| 336 | llvm::append_range(worklist, successors(ToUpdate)); | ||||||||
| 337 | } | ||||||||
| 338 | } | ||||||||
| 339 | |||||||||
| 340 | |||||||||
| 341 | namespace { | ||||||||
| 342 | /// An assembly annotator class to print LazyValueCache information in | ||||||||
| 343 | /// comments. | ||||||||
| 344 | class LazyValueInfoImpl; | ||||||||
| 345 | class LazyValueInfoAnnotatedWriter : public AssemblyAnnotationWriter { | ||||||||
| 346 | LazyValueInfoImpl *LVIImpl; | ||||||||
| 347 | // While analyzing which blocks we can solve values for, we need the dominator | ||||||||
| 348 | // information. | ||||||||
| 349 | DominatorTree &DT; | ||||||||
| 350 | |||||||||
| 351 | public: | ||||||||
| 352 | LazyValueInfoAnnotatedWriter(LazyValueInfoImpl *L, DominatorTree &DTree) | ||||||||
| 353 | : LVIImpl(L), DT(DTree) {} | ||||||||
| 354 | |||||||||
| 355 | void emitBasicBlockStartAnnot(const BasicBlock *BB, | ||||||||
| 356 | formatted_raw_ostream &OS) override; | ||||||||
| 357 | |||||||||
| 358 | void emitInstructionAnnot(const Instruction *I, | ||||||||
| 359 | formatted_raw_ostream &OS) override; | ||||||||
| 360 | }; | ||||||||
| 361 | } | ||||||||
| 362 | namespace { | ||||||||
| 363 | // The actual implementation of the lazy analysis and update. Note that the | ||||||||
| 364 | // inheritance from LazyValueInfoCache is intended to be temporary while | ||||||||
| 365 | // splitting the code and then transitioning to a has-a relationship. | ||||||||
| 366 | class LazyValueInfoImpl { | ||||||||
| 367 | |||||||||
| 368 | /// Cached results from previous queries | ||||||||
| 369 | LazyValueInfoCache TheCache; | ||||||||
| 370 | |||||||||
| 371 | /// This stack holds the state of the value solver during a query. | ||||||||
| 372 | /// It basically emulates the callstack of the naive | ||||||||
| 373 | /// recursive value lookup process. | ||||||||
| 374 | SmallVector<std::pair<BasicBlock*, Value*>, 8> BlockValueStack; | ||||||||
| 375 | |||||||||
| 376 | /// Keeps track of which block-value pairs are in BlockValueStack. | ||||||||
| 377 | DenseSet<std::pair<BasicBlock*, Value*> > BlockValueSet; | ||||||||
| 378 | |||||||||
| 379 | /// Push BV onto BlockValueStack unless it's already in there. | ||||||||
| 380 | /// Returns true on success. | ||||||||
| 381 | bool pushBlockValue(const std::pair<BasicBlock *, Value *> &BV) { | ||||||||
| 382 | if (!BlockValueSet.insert(BV).second) | ||||||||
| 383 | return false; // It's already in the stack. | ||||||||
| 384 | |||||||||
| 385 | LLVM_DEBUG(dbgs() << "PUSH: " << *BV.second << " in "do { } while (false) | ||||||||
| 386 | << BV.first->getName() << "\n")do { } while (false); | ||||||||
| 387 | BlockValueStack.push_back(BV); | ||||||||
| 388 | return true; | ||||||||
| 389 | } | ||||||||
| 390 | |||||||||
| 391 | AssumptionCache *AC; ///< A pointer to the cache of @llvm.assume calls. | ||||||||
| 392 | const DataLayout &DL; ///< A mandatory DataLayout | ||||||||
| 393 | |||||||||
| 394 | /// Declaration of the llvm.experimental.guard() intrinsic, | ||||||||
| 395 | /// if it exists in the module. | ||||||||
| 396 | Function *GuardDecl; | ||||||||
| 397 | |||||||||
| 398 | Optional<ValueLatticeElement> getBlockValue(Value *Val, BasicBlock *BB); | ||||||||
| 399 | Optional<ValueLatticeElement> getEdgeValue(Value *V, BasicBlock *F, | ||||||||
| 400 | BasicBlock *T, Instruction *CxtI = nullptr); | ||||||||
| 401 | |||||||||
| 402 | // These methods process one work item and may add more. A false value | ||||||||
| 403 | // returned means that the work item was not completely processed and must | ||||||||
| 404 | // be revisited after going through the new items. | ||||||||
| 405 | bool solveBlockValue(Value *Val, BasicBlock *BB); | ||||||||
| 406 | Optional<ValueLatticeElement> solveBlockValueImpl(Value *Val, BasicBlock *BB); | ||||||||
| 407 | Optional<ValueLatticeElement> solveBlockValueNonLocal(Value *Val, | ||||||||
| 408 | BasicBlock *BB); | ||||||||
| 409 | Optional<ValueLatticeElement> solveBlockValuePHINode(PHINode *PN, | ||||||||
| 410 | BasicBlock *BB); | ||||||||
| 411 | Optional<ValueLatticeElement> solveBlockValueSelect(SelectInst *S, | ||||||||
| 412 | BasicBlock *BB); | ||||||||
| 413 | Optional<ConstantRange> getRangeFor(Value *V, Instruction *CxtI, | ||||||||
| 414 | BasicBlock *BB); | ||||||||
| 415 | Optional<ValueLatticeElement> solveBlockValueBinaryOpImpl( | ||||||||
| 416 | Instruction *I, BasicBlock *BB, | ||||||||
| 417 | std::function<ConstantRange(const ConstantRange &, | ||||||||
| 418 | const ConstantRange &)> OpFn); | ||||||||
| 419 | Optional<ValueLatticeElement> solveBlockValueBinaryOp(BinaryOperator *BBI, | ||||||||
| 420 | BasicBlock *BB); | ||||||||
| 421 | Optional<ValueLatticeElement> solveBlockValueCast(CastInst *CI, | ||||||||
| 422 | BasicBlock *BB); | ||||||||
| 423 | Optional<ValueLatticeElement> solveBlockValueOverflowIntrinsic( | ||||||||
| 424 | WithOverflowInst *WO, BasicBlock *BB); | ||||||||
| 425 | Optional<ValueLatticeElement> solveBlockValueIntrinsic(IntrinsicInst *II, | ||||||||
| 426 | BasicBlock *BB); | ||||||||
| 427 | Optional<ValueLatticeElement> solveBlockValueExtractValue( | ||||||||
| 428 | ExtractValueInst *EVI, BasicBlock *BB); | ||||||||
| 429 | bool isNonNullAtEndOfBlock(Value *Val, BasicBlock *BB); | ||||||||
| 430 | void intersectAssumeOrGuardBlockValueConstantRange(Value *Val, | ||||||||
| 431 | ValueLatticeElement &BBLV, | ||||||||
| 432 | Instruction *BBI); | ||||||||
| 433 | |||||||||
| 434 | void solve(); | ||||||||
| 435 | |||||||||
| 436 | public: | ||||||||
| 437 | /// This is the query interface to determine the lattice value for the | ||||||||
| 438 | /// specified Value* at the context instruction (if specified) or at the | ||||||||
| 439 | /// start of the block. | ||||||||
| 440 | ValueLatticeElement getValueInBlock(Value *V, BasicBlock *BB, | ||||||||
| 441 | Instruction *CxtI = nullptr); | ||||||||
| 442 | |||||||||
| 443 | /// This is the query interface to determine the lattice value for the | ||||||||
| 444 | /// specified Value* at the specified instruction using only information | ||||||||
| 445 | /// from assumes/guards and range metadata. Unlike getValueInBlock(), no | ||||||||
| 446 | /// recursive query is performed. | ||||||||
| 447 | ValueLatticeElement getValueAt(Value *V, Instruction *CxtI); | ||||||||
| 448 | |||||||||
| 449 | /// This is the query interface to determine the lattice | ||||||||
| 450 | /// value for the specified Value* that is true on the specified edge. | ||||||||
| 451 | ValueLatticeElement getValueOnEdge(Value *V, BasicBlock *FromBB, | ||||||||
| 452 | BasicBlock *ToBB, | ||||||||
| 453 | Instruction *CxtI = nullptr); | ||||||||
| 454 | |||||||||
| 455 | /// Complete flush all previously computed values | ||||||||
| 456 | void clear() { | ||||||||
| 457 | TheCache.clear(); | ||||||||
| 458 | } | ||||||||
| 459 | |||||||||
| 460 | /// Printing the LazyValueInfo Analysis. | ||||||||
| 461 | void printLVI(Function &F, DominatorTree &DTree, raw_ostream &OS) { | ||||||||
| 462 | LazyValueInfoAnnotatedWriter Writer(this, DTree); | ||||||||
| 463 | F.print(OS, &Writer); | ||||||||
| 464 | } | ||||||||
| 465 | |||||||||
| 466 | /// This is part of the update interface to inform the cache | ||||||||
| 467 | /// that a block has been deleted. | ||||||||
| 468 | void eraseBlock(BasicBlock *BB) { | ||||||||
| 469 | TheCache.eraseBlock(BB); | ||||||||
| 470 | } | ||||||||
| 471 | |||||||||
| 472 | /// This is the update interface to inform the cache that an edge from | ||||||||
| 473 | /// PredBB to OldSucc has been threaded to be from PredBB to NewSucc. | ||||||||
| 474 | void threadEdge(BasicBlock *PredBB,BasicBlock *OldSucc,BasicBlock *NewSucc); | ||||||||
| 475 | |||||||||
| 476 | LazyValueInfoImpl(AssumptionCache *AC, const DataLayout &DL, | ||||||||
| 477 | Function *GuardDecl) | ||||||||
| 478 | : AC(AC), DL(DL), GuardDecl(GuardDecl) {} | ||||||||
| 479 | }; | ||||||||
| 480 | } // end anonymous namespace | ||||||||
| 481 | |||||||||
| 482 | |||||||||
| 483 | void LazyValueInfoImpl::solve() { | ||||||||
| 484 | SmallVector<std::pair<BasicBlock *, Value *>, 8> StartingStack( | ||||||||
| 485 | BlockValueStack.begin(), BlockValueStack.end()); | ||||||||
| 486 | |||||||||
| 487 | unsigned processedCount = 0; | ||||||||
| 488 | while (!BlockValueStack.empty()) { | ||||||||
| 489 | processedCount++; | ||||||||
| 490 | // Abort if we have to process too many values to get a result for this one. | ||||||||
| 491 | // Because of the design of the overdefined cache currently being per-block | ||||||||
| 492 | // to avoid naming-related issues (IE it wants to try to give different | ||||||||
| 493 | // results for the same name in different blocks), overdefined results don't | ||||||||
| 494 | // get cached globally, which in turn means we will often try to rediscover | ||||||||
| 495 | // the same overdefined result again and again. Once something like | ||||||||
| 496 | // PredicateInfo is used in LVI or CVP, we should be able to make the | ||||||||
| 497 | // overdefined cache global, and remove this throttle. | ||||||||
| 498 | if (processedCount > MaxProcessedPerValue) { | ||||||||
| 499 | LLVM_DEBUG(do { } while (false) | ||||||||
| 500 | dbgs() << "Giving up on stack because we are getting too deep\n")do { } while (false); | ||||||||
| 501 | // Fill in the original values | ||||||||
| 502 | while (!StartingStack.empty()) { | ||||||||
| 503 | std::pair<BasicBlock *, Value *> &e = StartingStack.back(); | ||||||||
| 504 | TheCache.insertResult(e.second, e.first, | ||||||||
| 505 | ValueLatticeElement::getOverdefined()); | ||||||||
| 506 | StartingStack.pop_back(); | ||||||||
| 507 | } | ||||||||
| 508 | BlockValueSet.clear(); | ||||||||
| 509 | BlockValueStack.clear(); | ||||||||
| 510 | return; | ||||||||
| 511 | } | ||||||||
| 512 | std::pair<BasicBlock *, Value *> e = BlockValueStack.back(); | ||||||||
| 513 | assert(BlockValueSet.count(e) && "Stack value should be in BlockValueSet!")((void)0); | ||||||||
| 514 | |||||||||
| 515 | if (solveBlockValue(e.second, e.first)) { | ||||||||
| 516 | // The work item was completely processed. | ||||||||
| 517 | assert(BlockValueStack.back() == e && "Nothing should have been pushed!")((void)0); | ||||||||
| 518 | #ifndef NDEBUG1 | ||||||||
| 519 | Optional<ValueLatticeElement> BBLV = | ||||||||
| 520 | TheCache.getCachedValueInfo(e.second, e.first); | ||||||||
| 521 | assert(BBLV && "Result should be in cache!")((void)0); | ||||||||
| 522 | LLVM_DEBUG(do { } while (false) | ||||||||
| 523 | dbgs() << "POP " << *e.second << " in " << e.first->getName() << " = "do { } while (false) | ||||||||
| 524 | << *BBLV << "\n")do { } while (false); | ||||||||
| 525 | #endif | ||||||||
| 526 | |||||||||
| 527 | BlockValueStack.pop_back(); | ||||||||
| 528 | BlockValueSet.erase(e); | ||||||||
| 529 | } else { | ||||||||
| 530 | // More work needs to be done before revisiting. | ||||||||
| 531 | assert(BlockValueStack.back() != e && "Stack should have been pushed!")((void)0); | ||||||||
| 532 | } | ||||||||
| 533 | } | ||||||||
| 534 | } | ||||||||
| 535 | |||||||||
| 536 | Optional<ValueLatticeElement> LazyValueInfoImpl::getBlockValue(Value *Val, | ||||||||
| 537 | BasicBlock *BB) { | ||||||||
| 538 | // If already a constant, there is nothing to compute. | ||||||||
| 539 | if (Constant *VC = dyn_cast<Constant>(Val)) | ||||||||
| 540 | return ValueLatticeElement::get(VC); | ||||||||
| 541 | |||||||||
| 542 | if (Optional<ValueLatticeElement> OptLatticeVal = | ||||||||
| 543 | TheCache.getCachedValueInfo(Val, BB)) | ||||||||
| 544 | return OptLatticeVal; | ||||||||
| 545 | |||||||||
| 546 | // We have hit a cycle, assume overdefined. | ||||||||
| 547 | if (!pushBlockValue({ BB, Val })) | ||||||||
| 548 | return ValueLatticeElement::getOverdefined(); | ||||||||
| 549 | |||||||||
| 550 | // Yet to be resolved. | ||||||||
| 551 | return None; | ||||||||
| 552 | } | ||||||||
| 553 | |||||||||
| 554 | static ValueLatticeElement getFromRangeMetadata(Instruction *BBI) { | ||||||||
| 555 | switch (BBI->getOpcode()) { | ||||||||
| 556 | default: break; | ||||||||
| 557 | case Instruction::Load: | ||||||||
| 558 | case Instruction::Call: | ||||||||
| 559 | case Instruction::Invoke: | ||||||||
| 560 | if (MDNode *Ranges = BBI->getMetadata(LLVMContext::MD_range)) | ||||||||
| 561 | if (isa<IntegerType>(BBI->getType())) { | ||||||||
| 562 | return ValueLatticeElement::getRange( | ||||||||
| 563 | getConstantRangeFromMetadata(*Ranges)); | ||||||||
| 564 | } | ||||||||
| 565 | break; | ||||||||
| 566 | }; | ||||||||
| 567 | // Nothing known - will be intersected with other facts | ||||||||
| 568 | return ValueLatticeElement::getOverdefined(); | ||||||||
| 569 | } | ||||||||
| 570 | |||||||||
| 571 | bool LazyValueInfoImpl::solveBlockValue(Value *Val, BasicBlock *BB) { | ||||||||
| 572 | assert(!isa<Constant>(Val) && "Value should not be constant")((void)0); | ||||||||
| 573 | assert(!TheCache.getCachedValueInfo(Val, BB) &&((void)0) | ||||||||
| 574 | "Value should not be in cache")((void)0); | ||||||||
| 575 | |||||||||
| 576 | // Hold off inserting this value into the Cache in case we have to return | ||||||||
| 577 | // false and come back later. | ||||||||
| 578 | Optional<ValueLatticeElement> Res = solveBlockValueImpl(Val, BB); | ||||||||
| 579 | if (!Res) | ||||||||
| 580 | // Work pushed, will revisit | ||||||||
| 581 | return false; | ||||||||
| 582 | |||||||||
| 583 | TheCache.insertResult(Val, BB, *Res); | ||||||||
| 584 | return true; | ||||||||
| 585 | } | ||||||||
| 586 | |||||||||
| 587 | Optional<ValueLatticeElement> LazyValueInfoImpl::solveBlockValueImpl( | ||||||||
| 588 | Value *Val, BasicBlock *BB) { | ||||||||
| 589 | Instruction *BBI = dyn_cast<Instruction>(Val); | ||||||||
| 590 | if (!BBI || BBI->getParent() != BB) | ||||||||
| 591 | return solveBlockValueNonLocal(Val, BB); | ||||||||
| 592 | |||||||||
| 593 | if (PHINode *PN = dyn_cast<PHINode>(BBI)) | ||||||||
| 594 | return solveBlockValuePHINode(PN, BB); | ||||||||
| 595 | |||||||||
| 596 | if (auto *SI = dyn_cast<SelectInst>(BBI)) | ||||||||
| 597 | return solveBlockValueSelect(SI, BB); | ||||||||
| 598 | |||||||||
| 599 | // If this value is a nonnull pointer, record it's range and bailout. Note | ||||||||
| 600 | // that for all other pointer typed values, we terminate the search at the | ||||||||
| 601 | // definition. We could easily extend this to look through geps, bitcasts, | ||||||||
| 602 | // and the like to prove non-nullness, but it's not clear that's worth it | ||||||||
| 603 | // compile time wise. The context-insensitive value walk done inside | ||||||||
| 604 | // isKnownNonZero gets most of the profitable cases at much less expense. | ||||||||
| 605 | // This does mean that we have a sensitivity to where the defining | ||||||||
| 606 | // instruction is placed, even if it could legally be hoisted much higher. | ||||||||
| 607 | // That is unfortunate. | ||||||||
| 608 | PointerType *PT = dyn_cast<PointerType>(BBI->getType()); | ||||||||
| 609 | if (PT && isKnownNonZero(BBI, DL)) | ||||||||
| 610 | return ValueLatticeElement::getNot(ConstantPointerNull::get(PT)); | ||||||||
| 611 | |||||||||
| 612 | if (BBI->getType()->isIntegerTy()) { | ||||||||
| 613 | if (auto *CI = dyn_cast<CastInst>(BBI)) | ||||||||
| 614 | return solveBlockValueCast(CI, BB); | ||||||||
| 615 | |||||||||
| 616 | if (BinaryOperator *BO = dyn_cast<BinaryOperator>(BBI)) | ||||||||
| 617 | return solveBlockValueBinaryOp(BO, BB); | ||||||||
| 618 | |||||||||
| 619 | if (auto *EVI = dyn_cast<ExtractValueInst>(BBI)) | ||||||||
| 620 | return solveBlockValueExtractValue(EVI, BB); | ||||||||
| 621 | |||||||||
| 622 | if (auto *II = dyn_cast<IntrinsicInst>(BBI)) | ||||||||
| 623 | return solveBlockValueIntrinsic(II, BB); | ||||||||
| 624 | } | ||||||||
| 625 | |||||||||
| 626 | LLVM_DEBUG(dbgs() << " compute BB '" << BB->getName()do { } while (false) | ||||||||
| 627 | << "' - unknown inst def found.\n")do { } while (false); | ||||||||
| 628 | return getFromRangeMetadata(BBI); | ||||||||
| 629 | } | ||||||||
| 630 | |||||||||
| 631 | static void AddNonNullPointer(Value *Ptr, NonNullPointerSet &PtrSet) { | ||||||||
| 632 | // TODO: Use NullPointerIsDefined instead. | ||||||||
| 633 | if (Ptr->getType()->getPointerAddressSpace() == 0) | ||||||||
| 634 | PtrSet.insert(getUnderlyingObject(Ptr)); | ||||||||
| 635 | } | ||||||||
| 636 | |||||||||
| 637 | static void AddNonNullPointersByInstruction( | ||||||||
| 638 | Instruction *I, NonNullPointerSet &PtrSet) { | ||||||||
| 639 | if (LoadInst *L = dyn_cast<LoadInst>(I)) { | ||||||||
| 640 | AddNonNullPointer(L->getPointerOperand(), PtrSet); | ||||||||
| 641 | } else if (StoreInst *S = dyn_cast<StoreInst>(I)) { | ||||||||
| 642 | AddNonNullPointer(S->getPointerOperand(), PtrSet); | ||||||||
| 643 | } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(I)) { | ||||||||
| 644 | if (MI->isVolatile()) return; | ||||||||
| 645 | |||||||||
| 646 | // FIXME: check whether it has a valuerange that excludes zero? | ||||||||
| 647 | ConstantInt *Len = dyn_cast<ConstantInt>(MI->getLength()); | ||||||||
| 648 | if (!Len || Len->isZero()) return; | ||||||||
| 649 | |||||||||
| 650 | AddNonNullPointer(MI->getRawDest(), PtrSet); | ||||||||
| 651 | if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) | ||||||||
| 652 | AddNonNullPointer(MTI->getRawSource(), PtrSet); | ||||||||
| 653 | } | ||||||||
| 654 | } | ||||||||
| 655 | |||||||||
| 656 | bool LazyValueInfoImpl::isNonNullAtEndOfBlock(Value *Val, BasicBlock *BB) { | ||||||||
| 657 | if (NullPointerIsDefined(BB->getParent(), | ||||||||
| 658 | Val->getType()->getPointerAddressSpace())) | ||||||||
| 659 | return false; | ||||||||
| 660 | |||||||||
| 661 | Val = Val->stripInBoundsOffsets(); | ||||||||
| 662 | return TheCache.isNonNullAtEndOfBlock(Val, BB, [](BasicBlock *BB) { | ||||||||
| 663 | NonNullPointerSet NonNullPointers; | ||||||||
| 664 | for (Instruction &I : *BB) | ||||||||
| 665 | AddNonNullPointersByInstruction(&I, NonNullPointers); | ||||||||
| 666 | return NonNullPointers; | ||||||||
| 667 | }); | ||||||||
| 668 | } | ||||||||
| 669 | |||||||||
| 670 | Optional<ValueLatticeElement> LazyValueInfoImpl::solveBlockValueNonLocal( | ||||||||
| 671 | Value *Val, BasicBlock *BB) { | ||||||||
| 672 | ValueLatticeElement Result; // Start Undefined. | ||||||||
| 673 | |||||||||
| 674 | // If this is the entry block, we must be asking about an argument. The | ||||||||
| 675 | // value is overdefined. | ||||||||
| 676 | if (BB->isEntryBlock()) { | ||||||||
| 677 | assert(isa<Argument>(Val) && "Unknown live-in to the entry block")((void)0); | ||||||||
| 678 | return ValueLatticeElement::getOverdefined(); | ||||||||
| 679 | } | ||||||||
| 680 | |||||||||
| 681 | // Loop over all of our predecessors, merging what we know from them into | ||||||||
| 682 | // result. If we encounter an unexplored predecessor, we eagerly explore it | ||||||||
| 683 | // in a depth first manner. In practice, this has the effect of discovering | ||||||||
| 684 | // paths we can't analyze eagerly without spending compile times analyzing | ||||||||
| 685 | // other paths. This heuristic benefits from the fact that predecessors are | ||||||||
| 686 | // frequently arranged such that dominating ones come first and we quickly | ||||||||
| 687 | // find a path to function entry. TODO: We should consider explicitly | ||||||||
| 688 | // canonicalizing to make this true rather than relying on this happy | ||||||||
| 689 | // accident. | ||||||||
| 690 | for (BasicBlock *Pred : predecessors(BB)) { | ||||||||
| 691 | Optional<ValueLatticeElement> EdgeResult = getEdgeValue(Val, Pred, BB); | ||||||||
| 692 | if (!EdgeResult) | ||||||||
| 693 | // Explore that input, then return here | ||||||||
| 694 | return None; | ||||||||
| 695 | |||||||||
| 696 | Result.mergeIn(*EdgeResult); | ||||||||
| 697 | |||||||||
| 698 | // If we hit overdefined, exit early. The BlockVals entry is already set | ||||||||
| 699 | // to overdefined. | ||||||||
| 700 | if (Result.isOverdefined()) { | ||||||||
| 701 | LLVM_DEBUG(dbgs() << " compute BB '" << BB->getName()do { } while (false) | ||||||||
| 702 | << "' - overdefined because of pred (non local).\n")do { } while (false); | ||||||||
| 703 | return Result; | ||||||||
| 704 | } | ||||||||
| 705 | } | ||||||||
| 706 | |||||||||
| 707 | // Return the merged value, which is more precise than 'overdefined'. | ||||||||
| 708 | assert(!Result.isOverdefined())((void)0); | ||||||||
| 709 | return Result; | ||||||||
| 710 | } | ||||||||
| 711 | |||||||||
| 712 | Optional<ValueLatticeElement> LazyValueInfoImpl::solveBlockValuePHINode( | ||||||||
| 713 | PHINode *PN, BasicBlock *BB) { | ||||||||
| 714 | ValueLatticeElement Result; // Start Undefined. | ||||||||
| 715 | |||||||||
| 716 | // Loop over all of our predecessors, merging what we know from them into | ||||||||
| 717 | // result. See the comment about the chosen traversal order in | ||||||||
| 718 | // solveBlockValueNonLocal; the same reasoning applies here. | ||||||||
| 719 | for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { | ||||||||
| 720 | BasicBlock *PhiBB = PN->getIncomingBlock(i); | ||||||||
| 721 | Value *PhiVal = PN->getIncomingValue(i); | ||||||||
| 722 | // Note that we can provide PN as the context value to getEdgeValue, even | ||||||||
| 723 | // though the results will be cached, because PN is the value being used as | ||||||||
| 724 | // the cache key in the caller. | ||||||||
| 725 | Optional<ValueLatticeElement> EdgeResult = | ||||||||
| 726 | getEdgeValue(PhiVal, PhiBB, BB, PN); | ||||||||
| 727 | if (!EdgeResult) | ||||||||
| 728 | // Explore that input, then return here | ||||||||
| 729 | return None; | ||||||||
| 730 | |||||||||
| 731 | Result.mergeIn(*EdgeResult); | ||||||||
| 732 | |||||||||
| 733 | // If we hit overdefined, exit early. The BlockVals entry is already set | ||||||||
| 734 | // to overdefined. | ||||||||
| 735 | if (Result.isOverdefined()) { | ||||||||
| 736 | LLVM_DEBUG(dbgs() << " compute BB '" << BB->getName()do { } while (false) | ||||||||
| 737 | << "' - overdefined because of pred (local).\n")do { } while (false); | ||||||||
| 738 | |||||||||
| 739 | return Result; | ||||||||
| 740 | } | ||||||||
| 741 | } | ||||||||
| 742 | |||||||||
| 743 | // Return the merged value, which is more precise than 'overdefined'. | ||||||||
| 744 | assert(!Result.isOverdefined() && "Possible PHI in entry block?")((void)0); | ||||||||
| 745 | return Result; | ||||||||
| 746 | } | ||||||||
| 747 | |||||||||
| 748 | static ValueLatticeElement getValueFromCondition(Value *Val, Value *Cond, | ||||||||
| 749 | bool isTrueDest = true); | ||||||||
| 750 | |||||||||
| 751 | // If we can determine a constraint on the value given conditions assumed by | ||||||||
| 752 | // the program, intersect those constraints with BBLV | ||||||||
| 753 | void LazyValueInfoImpl::intersectAssumeOrGuardBlockValueConstantRange( | ||||||||
| 754 | Value *Val, ValueLatticeElement &BBLV, Instruction *BBI) { | ||||||||
| 755 | BBI = BBI ? BBI : dyn_cast<Instruction>(Val); | ||||||||
| 756 | if (!BBI) | ||||||||
| 757 | return; | ||||||||
| 758 | |||||||||
| 759 | BasicBlock *BB = BBI->getParent(); | ||||||||
| 760 | for (auto &AssumeVH : AC->assumptionsFor(Val)) { | ||||||||
| 761 | if (!AssumeVH) | ||||||||
| 762 | continue; | ||||||||
| 763 | |||||||||
| 764 | // Only check assumes in the block of the context instruction. Other | ||||||||
| 765 | // assumes will have already been taken into account when the value was | ||||||||
| 766 | // propagated from predecessor blocks. | ||||||||
| 767 | auto *I = cast<CallInst>(AssumeVH); | ||||||||
| 768 | if (I->getParent() != BB || !isValidAssumeForContext(I, BBI)) | ||||||||
| 769 | continue; | ||||||||
| 770 | |||||||||
| 771 | BBLV = intersect(BBLV, getValueFromCondition(Val, I->getArgOperand(0))); | ||||||||
| 772 | } | ||||||||
| 773 | |||||||||
| 774 | // If guards are not used in the module, don't spend time looking for them | ||||||||
| 775 | if (GuardDecl && !GuardDecl->use_empty() && | ||||||||
| 776 | BBI->getIterator() != BB->begin()) { | ||||||||
| 777 | for (Instruction &I : make_range(std::next(BBI->getIterator().getReverse()), | ||||||||
| 778 | BB->rend())) { | ||||||||
| 779 | Value *Cond = nullptr; | ||||||||
| 780 | if (match(&I, m_Intrinsic<Intrinsic::experimental_guard>(m_Value(Cond)))) | ||||||||
| 781 | BBLV = intersect(BBLV, getValueFromCondition(Val, Cond)); | ||||||||
| 782 | } | ||||||||
| 783 | } | ||||||||
| 784 | |||||||||
| 785 | if (BBLV.isOverdefined()) { | ||||||||
| 786 | // Check whether we're checking at the terminator, and the pointer has | ||||||||
| 787 | // been dereferenced in this block. | ||||||||
| 788 | PointerType *PTy = dyn_cast<PointerType>(Val->getType()); | ||||||||
| 789 | if (PTy && BB->getTerminator() == BBI && | ||||||||
| 790 | isNonNullAtEndOfBlock(Val, BB)) | ||||||||
| 791 | BBLV = ValueLatticeElement::getNot(ConstantPointerNull::get(PTy)); | ||||||||
| 792 | } | ||||||||
| 793 | } | ||||||||
| 794 | |||||||||
| 795 | Optional<ValueLatticeElement> LazyValueInfoImpl::solveBlockValueSelect( | ||||||||
| 796 | SelectInst *SI, BasicBlock *BB) { | ||||||||
| 797 | // Recurse on our inputs if needed | ||||||||
| 798 | Optional<ValueLatticeElement> OptTrueVal = | ||||||||
| 799 | getBlockValue(SI->getTrueValue(), BB); | ||||||||
| 800 | if (!OptTrueVal) | ||||||||
| 801 | return None; | ||||||||
| 802 | ValueLatticeElement &TrueVal = *OptTrueVal; | ||||||||
| 803 | |||||||||
| 804 | Optional<ValueLatticeElement> OptFalseVal = | ||||||||
| 805 | getBlockValue(SI->getFalseValue(), BB); | ||||||||
| 806 | if (!OptFalseVal) | ||||||||
| 807 | return None; | ||||||||
| 808 | ValueLatticeElement &FalseVal = *OptFalseVal; | ||||||||
| 809 | |||||||||
| 810 | if (TrueVal.isConstantRange() && FalseVal.isConstantRange()) { | ||||||||
| 811 | const ConstantRange &TrueCR = TrueVal.getConstantRange(); | ||||||||
| 812 | const ConstantRange &FalseCR = FalseVal.getConstantRange(); | ||||||||
| 813 | Value *LHS = nullptr; | ||||||||
| 814 | Value *RHS = nullptr; | ||||||||
| 815 | SelectPatternResult SPR = matchSelectPattern(SI, LHS, RHS); | ||||||||
| 816 | // Is this a min specifically of our two inputs? (Avoid the risk of | ||||||||
| 817 | // ValueTracking getting smarter looking back past our immediate inputs.) | ||||||||
| 818 | if (SelectPatternResult::isMinOrMax(SPR.Flavor) && | ||||||||
| 819 | LHS == SI->getTrueValue() && RHS == SI->getFalseValue()) { | ||||||||
| 820 | ConstantRange ResultCR = [&]() { | ||||||||
| 821 | switch (SPR.Flavor) { | ||||||||
| 822 | default: | ||||||||
| 823 | llvm_unreachable("unexpected minmax type!")__builtin_unreachable(); | ||||||||
| 824 | case SPF_SMIN: /// Signed minimum | ||||||||
| 825 | return TrueCR.smin(FalseCR); | ||||||||
| 826 | case SPF_UMIN: /// Unsigned minimum | ||||||||
| 827 | return TrueCR.umin(FalseCR); | ||||||||
| 828 | case SPF_SMAX: /// Signed maximum | ||||||||
| 829 | return TrueCR.smax(FalseCR); | ||||||||
| 830 | case SPF_UMAX: /// Unsigned maximum | ||||||||
| 831 | return TrueCR.umax(FalseCR); | ||||||||
| 832 | }; | ||||||||
| 833 | }(); | ||||||||
| 834 | return ValueLatticeElement::getRange( | ||||||||
| 835 | ResultCR, TrueVal.isConstantRangeIncludingUndef() | | ||||||||
| 836 | FalseVal.isConstantRangeIncludingUndef()); | ||||||||
| 837 | } | ||||||||
| 838 | |||||||||
| 839 | if (SPR.Flavor == SPF_ABS) { | ||||||||
| 840 | if (LHS == SI->getTrueValue()) | ||||||||
| 841 | return ValueLatticeElement::getRange( | ||||||||
| 842 | TrueCR.abs(), TrueVal.isConstantRangeIncludingUndef()); | ||||||||
| 843 | if (LHS == SI->getFalseValue()) | ||||||||
| 844 | return ValueLatticeElement::getRange( | ||||||||
| 845 | FalseCR.abs(), FalseVal.isConstantRangeIncludingUndef()); | ||||||||
| 846 | } | ||||||||
| 847 | |||||||||
| 848 | if (SPR.Flavor == SPF_NABS) { | ||||||||
| 849 | ConstantRange Zero(APInt::getNullValue(TrueCR.getBitWidth())); | ||||||||
| 850 | if (LHS == SI->getTrueValue()) | ||||||||
| 851 | return ValueLatticeElement::getRange( | ||||||||
| 852 | Zero.sub(TrueCR.abs()), FalseVal.isConstantRangeIncludingUndef()); | ||||||||
| 853 | if (LHS == SI->getFalseValue()) | ||||||||
| 854 | return ValueLatticeElement::getRange( | ||||||||
| 855 | Zero.sub(FalseCR.abs()), FalseVal.isConstantRangeIncludingUndef()); | ||||||||
| 856 | } | ||||||||
| 857 | } | ||||||||
| 858 | |||||||||
| 859 | // Can we constrain the facts about the true and false values by using the | ||||||||
| 860 | // condition itself? This shows up with idioms like e.g. select(a > 5, a, 5). | ||||||||
| 861 | // TODO: We could potentially refine an overdefined true value above. | ||||||||
| 862 | Value *Cond = SI->getCondition(); | ||||||||
| 863 | TrueVal = intersect(TrueVal, | ||||||||
| 864 | getValueFromCondition(SI->getTrueValue(), Cond, true)); | ||||||||
| 865 | FalseVal = intersect(FalseVal, | ||||||||
| 866 | getValueFromCondition(SI->getFalseValue(), Cond, false)); | ||||||||
| 867 | |||||||||
| 868 | ValueLatticeElement Result = TrueVal; | ||||||||
| 869 | Result.mergeIn(FalseVal); | ||||||||
| 870 | return Result; | ||||||||
| 871 | } | ||||||||
| 872 | |||||||||
| 873 | Optional<ConstantRange> LazyValueInfoImpl::getRangeFor(Value *V, | ||||||||
| 874 | Instruction *CxtI, | ||||||||
| 875 | BasicBlock *BB) { | ||||||||
| 876 | Optional<ValueLatticeElement> OptVal = getBlockValue(V, BB); | ||||||||
| 877 | if (!OptVal) | ||||||||
| 878 | return None; | ||||||||
| 879 | |||||||||
| 880 | ValueLatticeElement &Val = *OptVal; | ||||||||
| 881 | intersectAssumeOrGuardBlockValueConstantRange(V, Val, CxtI); | ||||||||
| 882 | if (Val.isConstantRange()) | ||||||||
| 883 | return Val.getConstantRange(); | ||||||||
| 884 | |||||||||
| 885 | const unsigned OperandBitWidth = DL.getTypeSizeInBits(V->getType()); | ||||||||
| 886 | return ConstantRange::getFull(OperandBitWidth); | ||||||||
| 887 | } | ||||||||
| 888 | |||||||||
| 889 | Optional<ValueLatticeElement> LazyValueInfoImpl::solveBlockValueCast( | ||||||||
| 890 | CastInst *CI, BasicBlock *BB) { | ||||||||
| 891 | // Without knowing how wide the input is, we can't analyze it in any useful | ||||||||
| 892 | // way. | ||||||||
| 893 | if (!CI->getOperand(0)->getType()->isSized()) | ||||||||
| 894 | return ValueLatticeElement::getOverdefined(); | ||||||||
| 895 | |||||||||
| 896 | // Filter out casts we don't know how to reason about before attempting to | ||||||||
| 897 | // recurse on our operand. This can cut a long search short if we know we're | ||||||||
| 898 | // not going to be able to get any useful information anways. | ||||||||
| 899 | switch (CI->getOpcode()) { | ||||||||
| 900 | case Instruction::Trunc: | ||||||||
| 901 | case Instruction::SExt: | ||||||||
| 902 | case Instruction::ZExt: | ||||||||
| 903 | case Instruction::BitCast: | ||||||||
| 904 | break; | ||||||||
| 905 | default: | ||||||||
| 906 | // Unhandled instructions are overdefined. | ||||||||
| 907 | LLVM_DEBUG(dbgs() << " compute BB '" << BB->getName()do { } while (false) | ||||||||
| 908 | << "' - overdefined (unknown cast).\n")do { } while (false); | ||||||||
| 909 | return ValueLatticeElement::getOverdefined(); | ||||||||
| 910 | } | ||||||||
| 911 | |||||||||
| 912 | // Figure out the range of the LHS. If that fails, we still apply the | ||||||||
| 913 | // transfer rule on the full set since we may be able to locally infer | ||||||||
| 914 | // interesting facts. | ||||||||
| 915 | Optional<ConstantRange> LHSRes = getRangeFor(CI->getOperand(0), CI, BB); | ||||||||
| 916 | if (!LHSRes.hasValue()) | ||||||||
| 917 | // More work to do before applying this transfer rule. | ||||||||
| 918 | return None; | ||||||||
| 919 | const ConstantRange &LHSRange = LHSRes.getValue(); | ||||||||
| 920 | |||||||||
| 921 | const unsigned ResultBitWidth = CI->getType()->getIntegerBitWidth(); | ||||||||
| 922 | |||||||||
| 923 | // NOTE: We're currently limited by the set of operations that ConstantRange | ||||||||
| 924 | // can evaluate symbolically. Enhancing that set will allows us to analyze | ||||||||
| 925 | // more definitions. | ||||||||
| 926 | return ValueLatticeElement::getRange(LHSRange.castOp(CI->getOpcode(), | ||||||||
| 927 | ResultBitWidth)); | ||||||||
| 928 | } | ||||||||
| 929 | |||||||||
| 930 | Optional<ValueLatticeElement> LazyValueInfoImpl::solveBlockValueBinaryOpImpl( | ||||||||
| 931 | Instruction *I, BasicBlock *BB, | ||||||||
| 932 | std::function<ConstantRange(const ConstantRange &, | ||||||||
| 933 | const ConstantRange &)> OpFn) { | ||||||||
| 934 | // Figure out the ranges of the operands. If that fails, use a | ||||||||
| 935 | // conservative range, but apply the transfer rule anyways. This | ||||||||
| 936 | // lets us pick up facts from expressions like "and i32 (call i32 | ||||||||
| 937 | // @foo()), 32" | ||||||||
| 938 | Optional<ConstantRange> LHSRes = getRangeFor(I->getOperand(0), I, BB); | ||||||||
| 939 | Optional<ConstantRange> RHSRes = getRangeFor(I->getOperand(1), I, BB); | ||||||||
| 940 | if (!LHSRes.hasValue() || !RHSRes.hasValue()) | ||||||||
| 941 | // More work to do before applying this transfer rule. | ||||||||
| 942 | return None; | ||||||||
| 943 | |||||||||
| 944 | const ConstantRange &LHSRange = LHSRes.getValue(); | ||||||||
| 945 | const ConstantRange &RHSRange = RHSRes.getValue(); | ||||||||
| 946 | return ValueLatticeElement::getRange(OpFn(LHSRange, RHSRange)); | ||||||||
| 947 | } | ||||||||
| 948 | |||||||||
| 949 | Optional<ValueLatticeElement> LazyValueInfoImpl::solveBlockValueBinaryOp( | ||||||||
| 950 | BinaryOperator *BO, BasicBlock *BB) { | ||||||||
| 951 | assert(BO->getOperand(0)->getType()->isSized() &&((void)0) | ||||||||
| 952 | "all operands to binary operators are sized")((void)0); | ||||||||
| 953 | if (BO->getOpcode() == Instruction::Xor) { | ||||||||
| 954 | // Xor is the only operation not supported by ConstantRange::binaryOp(). | ||||||||
| 955 | LLVM_DEBUG(dbgs() << " compute BB '" << BB->getName()do { } while (false) | ||||||||
| 956 | << "' - overdefined (unknown binary operator).\n")do { } while (false); | ||||||||
| 957 | return ValueLatticeElement::getOverdefined(); | ||||||||
| 958 | } | ||||||||
| 959 | |||||||||
| 960 | if (auto *OBO = dyn_cast<OverflowingBinaryOperator>(BO)) { | ||||||||
| 961 | unsigned NoWrapKind = 0; | ||||||||
| 962 | if (OBO->hasNoUnsignedWrap()) | ||||||||
| 963 | NoWrapKind |= OverflowingBinaryOperator::NoUnsignedWrap; | ||||||||
| 964 | if (OBO->hasNoSignedWrap()) | ||||||||
| 965 | NoWrapKind |= OverflowingBinaryOperator::NoSignedWrap; | ||||||||
| 966 | |||||||||
| 967 | return solveBlockValueBinaryOpImpl( | ||||||||
| 968 | BO, BB, | ||||||||
| 969 | [BO, NoWrapKind](const ConstantRange &CR1, const ConstantRange &CR2) { | ||||||||
| 970 | return CR1.overflowingBinaryOp(BO->getOpcode(), CR2, NoWrapKind); | ||||||||
| 971 | }); | ||||||||
| 972 | } | ||||||||
| 973 | |||||||||
| 974 | return solveBlockValueBinaryOpImpl( | ||||||||
| 975 | BO, BB, [BO](const ConstantRange &CR1, const ConstantRange &CR2) { | ||||||||
| 976 | return CR1.binaryOp(BO->getOpcode(), CR2); | ||||||||
| 977 | }); | ||||||||
| 978 | } | ||||||||
| 979 | |||||||||
| 980 | Optional<ValueLatticeElement> | ||||||||
| 981 | LazyValueInfoImpl::solveBlockValueOverflowIntrinsic(WithOverflowInst *WO, | ||||||||
| 982 | BasicBlock *BB) { | ||||||||
| 983 | return solveBlockValueBinaryOpImpl( | ||||||||
| 984 | WO, BB, [WO](const ConstantRange &CR1, const ConstantRange &CR2) { | ||||||||
| 985 | return CR1.binaryOp(WO->getBinaryOp(), CR2); | ||||||||
| 986 | }); | ||||||||
| 987 | } | ||||||||
| 988 | |||||||||
| 989 | Optional<ValueLatticeElement> LazyValueInfoImpl::solveBlockValueIntrinsic( | ||||||||
| 990 | IntrinsicInst *II, BasicBlock *BB) { | ||||||||
| 991 | if (!ConstantRange::isIntrinsicSupported(II->getIntrinsicID())) { | ||||||||
| 992 | LLVM_DEBUG(dbgs() << " compute BB '" << BB->getName()do { } while (false) | ||||||||
| 993 | << "' - unknown intrinsic.\n")do { } while (false); | ||||||||
| 994 | return getFromRangeMetadata(II); | ||||||||
| 995 | } | ||||||||
| 996 | |||||||||
| 997 | SmallVector<ConstantRange, 2> OpRanges; | ||||||||
| 998 | for (Value *Op : II->args()) { | ||||||||
| 999 | Optional<ConstantRange> Range = getRangeFor(Op, II, BB); | ||||||||
| 1000 | if (!Range) | ||||||||
| 1001 | return None; | ||||||||
| 1002 | OpRanges.push_back(*Range); | ||||||||
| 1003 | } | ||||||||
| 1004 | |||||||||
| 1005 | return ValueLatticeElement::getRange( | ||||||||
| 1006 | ConstantRange::intrinsic(II->getIntrinsicID(), OpRanges)); | ||||||||
| 1007 | } | ||||||||
| 1008 | |||||||||
| 1009 | Optional<ValueLatticeElement> LazyValueInfoImpl::solveBlockValueExtractValue( | ||||||||
| 1010 | ExtractValueInst *EVI, BasicBlock *BB) { | ||||||||
| 1011 | if (auto *WO = dyn_cast<WithOverflowInst>(EVI->getAggregateOperand())) | ||||||||
| 1012 | if (EVI->getNumIndices() == 1 && *EVI->idx_begin() == 0) | ||||||||
| 1013 | return solveBlockValueOverflowIntrinsic(WO, BB); | ||||||||
| 1014 | |||||||||
| 1015 | // Handle extractvalue of insertvalue to allow further simplification | ||||||||
| 1016 | // based on replaced with.overflow intrinsics. | ||||||||
| 1017 | if (Value *V = SimplifyExtractValueInst( | ||||||||
| 1018 | EVI->getAggregateOperand(), EVI->getIndices(), | ||||||||
| 1019 | EVI->getModule()->getDataLayout())) | ||||||||
| 1020 | return getBlockValue(V, BB); | ||||||||
| 1021 | |||||||||
| 1022 | LLVM_DEBUG(dbgs() << " compute BB '" << BB->getName()do { } while (false) | ||||||||
| 1023 | << "' - overdefined (unknown extractvalue).\n")do { } while (false); | ||||||||
| 1024 | return ValueLatticeElement::getOverdefined(); | ||||||||
| 1025 | } | ||||||||
| 1026 | |||||||||
| 1027 | static bool matchICmpOperand(APInt &Offset, Value *LHS, Value *Val, | ||||||||
| 1028 | ICmpInst::Predicate Pred) { | ||||||||
| 1029 | if (LHS == Val) | ||||||||
| 1030 | return true; | ||||||||
| 1031 | |||||||||
| 1032 | // Handle range checking idiom produced by InstCombine. We will subtract the | ||||||||
| 1033 | // offset from the allowed range for RHS in this case. | ||||||||
| 1034 | const APInt *C; | ||||||||
| 1035 | if (match(LHS, m_Add(m_Specific(Val), m_APInt(C)))) { | ||||||||
| 1036 | Offset = *C; | ||||||||
| 1037 | return true; | ||||||||
| 1038 | } | ||||||||
| 1039 | |||||||||
| 1040 | // Handle the symmetric case. This appears in saturation patterns like | ||||||||
| 1041 | // (x == 16) ? 16 : (x + 1). | ||||||||
| 1042 | if (match(Val, m_Add(m_Specific(LHS), m_APInt(C)))) { | ||||||||
| 1043 | Offset = -*C; | ||||||||
| 1044 | return true; | ||||||||
| 1045 | } | ||||||||
| 1046 | |||||||||
| 1047 | // If (x | y) < C, then (x < C) && (y < C). | ||||||||
| 1048 | if (match(LHS, m_c_Or(m_Specific(Val), m_Value())) && | ||||||||
| 1049 | (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_ULE)) | ||||||||
| 1050 | return true; | ||||||||
| 1051 | |||||||||
| 1052 | // If (x & y) > C, then (x > C) && (y > C). | ||||||||
| 1053 | if (match(LHS, m_c_And(m_Specific(Val), m_Value())) && | ||||||||
| 1054 | (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_UGE)) | ||||||||
| 1055 | return true; | ||||||||
| 1056 | |||||||||
| 1057 | return false; | ||||||||
| 1058 | } | ||||||||
| 1059 | |||||||||
| 1060 | /// Get value range for a "(Val + Offset) Pred RHS" condition. | ||||||||
| 1061 | static ValueLatticeElement getValueFromSimpleICmpCondition( | ||||||||
| 1062 | CmpInst::Predicate Pred, Value *RHS, const APInt &Offset) { | ||||||||
| 1063 | ConstantRange RHSRange(RHS->getType()->getIntegerBitWidth(), | ||||||||
| 1064 | /*isFullSet=*/true); | ||||||||
| 1065 | if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) | ||||||||
| 1066 | RHSRange = ConstantRange(CI->getValue()); | ||||||||
| 1067 | else if (Instruction *I = dyn_cast<Instruction>(RHS)) | ||||||||
| 1068 | if (auto *Ranges = I->getMetadata(LLVMContext::MD_range)) | ||||||||
| 1069 | RHSRange = getConstantRangeFromMetadata(*Ranges); | ||||||||
| 1070 | |||||||||
| 1071 | ConstantRange TrueValues = | ||||||||
| 1072 | ConstantRange::makeAllowedICmpRegion(Pred, RHSRange); | ||||||||
| 1073 | return ValueLatticeElement::getRange(TrueValues.subtract(Offset)); | ||||||||
| 1074 | } | ||||||||
| 1075 | |||||||||
| 1076 | static ValueLatticeElement getValueFromICmpCondition(Value *Val, ICmpInst *ICI, | ||||||||
| 1077 | bool isTrueDest) { | ||||||||
| 1078 | Value *LHS = ICI->getOperand(0); | ||||||||
| 1079 | Value *RHS = ICI->getOperand(1); | ||||||||
| 1080 | |||||||||
| 1081 | // Get the predicate that must hold along the considered edge. | ||||||||
| 1082 | CmpInst::Predicate EdgePred = | ||||||||
| 1083 | isTrueDest ? ICI->getPredicate() : ICI->getInversePredicate(); | ||||||||
| 1084 | |||||||||
| 1085 | if (isa<Constant>(RHS)) { | ||||||||
| 1086 | if (ICI->isEquality() && LHS == Val) { | ||||||||
| 1087 | if (EdgePred == ICmpInst::ICMP_EQ) | ||||||||
| 1088 | return ValueLatticeElement::get(cast<Constant>(RHS)); | ||||||||
| 1089 | else if (!isa<UndefValue>(RHS)) | ||||||||
| 1090 | return ValueLatticeElement::getNot(cast<Constant>(RHS)); | ||||||||
| 1091 | } | ||||||||
| 1092 | } | ||||||||
| 1093 | |||||||||
| 1094 | Type *Ty = Val->getType(); | ||||||||
| 1095 | if (!Ty->isIntegerTy()) | ||||||||
| 1096 | return ValueLatticeElement::getOverdefined(); | ||||||||
| 1097 | |||||||||
| 1098 | APInt Offset(Ty->getScalarSizeInBits(), 0); | ||||||||
| 1099 | if (matchICmpOperand(Offset, LHS, Val, EdgePred)) | ||||||||
| 1100 | return getValueFromSimpleICmpCondition(EdgePred, RHS, Offset); | ||||||||
| 1101 | |||||||||
| 1102 | CmpInst::Predicate SwappedPred = CmpInst::getSwappedPredicate(EdgePred); | ||||||||
| 1103 | if (matchICmpOperand(Offset, RHS, Val, SwappedPred)) | ||||||||
| 1104 | return getValueFromSimpleICmpCondition(SwappedPred, LHS, Offset); | ||||||||
| 1105 | |||||||||
| 1106 | const APInt *Mask, *C; | ||||||||
| 1107 | if (match(LHS, m_And(m_Specific(Val), m_APInt(Mask))) && | ||||||||
| 1108 | match(RHS, m_APInt(C))) { | ||||||||
| 1109 | // If (Val & Mask) == C then all the masked bits are known and we can | ||||||||
| 1110 | // compute a value range based on that. | ||||||||
| 1111 | if (EdgePred == ICmpInst::ICMP_EQ) { | ||||||||
| 1112 | KnownBits Known; | ||||||||
| 1113 | Known.Zero = ~*C & *Mask; | ||||||||
| 1114 | Known.One = *C & *Mask; | ||||||||
| 1115 | return ValueLatticeElement::getRange( | ||||||||
| 1116 | ConstantRange::fromKnownBits(Known, /*IsSigned*/ false)); | ||||||||
| 1117 | } | ||||||||
| 1118 | // If (Val & Mask) != 0 then the value must be larger than the lowest set | ||||||||
| 1119 | // bit of Mask. | ||||||||
| 1120 | if (EdgePred == ICmpInst::ICMP_NE && !Mask->isNullValue() && | ||||||||
| 1121 | C->isNullValue()) { | ||||||||
| 1122 | unsigned BitWidth = Ty->getIntegerBitWidth(); | ||||||||
| 1123 | return ValueLatticeElement::getRange(ConstantRange::getNonEmpty( | ||||||||
| 1124 | APInt::getOneBitSet(BitWidth, Mask->countTrailingZeros()), | ||||||||
| 1125 | APInt::getNullValue(BitWidth))); | ||||||||
| 1126 | } | ||||||||
| 1127 | } | ||||||||
| 1128 | |||||||||
| 1129 | return ValueLatticeElement::getOverdefined(); | ||||||||
| 1130 | } | ||||||||
| 1131 | |||||||||
| 1132 | // Handle conditions of the form | ||||||||
| 1133 | // extractvalue(op.with.overflow(%x, C), 1). | ||||||||
| 1134 | static ValueLatticeElement getValueFromOverflowCondition( | ||||||||
| 1135 | Value *Val, WithOverflowInst *WO, bool IsTrueDest) { | ||||||||
| 1136 | // TODO: This only works with a constant RHS for now. We could also compute | ||||||||
| 1137 | // the range of the RHS, but this doesn't fit into the current structure of | ||||||||
| 1138 | // the edge value calculation. | ||||||||
| 1139 | const APInt *C; | ||||||||
| 1140 | if (WO->getLHS() != Val || !match(WO->getRHS(), m_APInt(C))) | ||||||||
| 1141 | return ValueLatticeElement::getOverdefined(); | ||||||||
| 1142 | |||||||||
| 1143 | // Calculate the possible values of %x for which no overflow occurs. | ||||||||
| 1144 | ConstantRange NWR = ConstantRange::makeExactNoWrapRegion( | ||||||||
| 1145 | WO->getBinaryOp(), *C, WO->getNoWrapKind()); | ||||||||
| 1146 | |||||||||
| 1147 | // If overflow is false, %x is constrained to NWR. If overflow is true, %x is | ||||||||
| 1148 | // constrained to it's inverse (all values that might cause overflow). | ||||||||
| 1149 | if (IsTrueDest) | ||||||||
| 1150 | NWR = NWR.inverse(); | ||||||||
| 1151 | return ValueLatticeElement::getRange(NWR); | ||||||||
| 1152 | } | ||||||||
| 1153 | |||||||||
| 1154 | static Optional<ValueLatticeElement> | ||||||||
| 1155 | getValueFromConditionImpl(Value *Val, Value *Cond, bool isTrueDest, | ||||||||
| 1156 | bool isRevisit, | ||||||||
| 1157 | SmallDenseMap<Value *, ValueLatticeElement> &Visited, | ||||||||
| 1158 | SmallVectorImpl<Value *> &Worklist) { | ||||||||
| 1159 | if (!isRevisit) { | ||||||||
| 1160 | if (ICmpInst *ICI = dyn_cast<ICmpInst>(Cond)) | ||||||||
| 1161 | return getValueFromICmpCondition(Val, ICI, isTrueDest); | ||||||||
| 1162 | |||||||||
| 1163 | if (auto *EVI = dyn_cast<ExtractValueInst>(Cond)) | ||||||||
| 1164 | if (auto *WO = dyn_cast<WithOverflowInst>(EVI->getAggregateOperand())) | ||||||||
| 1165 | if (EVI->getNumIndices() == 1 && *EVI->idx_begin() == 1) | ||||||||
| 1166 | return getValueFromOverflowCondition(Val, WO, isTrueDest); | ||||||||
| 1167 | } | ||||||||
| 1168 | |||||||||
| 1169 | Value *L, *R; | ||||||||
| 1170 | bool IsAnd; | ||||||||
| 1171 | if (match(Cond, m_LogicalAnd(m_Value(L), m_Value(R)))) | ||||||||
| 1172 | IsAnd = true; | ||||||||
| 1173 | else if (match(Cond, m_LogicalOr(m_Value(L), m_Value(R)))) | ||||||||
| 1174 | IsAnd = false; | ||||||||
| 1175 | else | ||||||||
| 1176 | return ValueLatticeElement::getOverdefined(); | ||||||||
| 1177 | |||||||||
| 1178 | auto LV = Visited.find(L); | ||||||||
| 1179 | auto RV = Visited.find(R); | ||||||||
| 1180 | |||||||||
| 1181 | // if (L && R) -> intersect L and R | ||||||||
| 1182 | // if (!(L || R)) -> intersect L and R | ||||||||
| 1183 | // if (L || R) -> union L and R | ||||||||
| 1184 | // if (!(L && R)) -> union L and R | ||||||||
| 1185 | if ((isTrueDest ^ IsAnd) && (LV != Visited.end())) { | ||||||||
| 1186 | ValueLatticeElement V = LV->second; | ||||||||
| 1187 | if (V.isOverdefined()) | ||||||||
| 1188 | return V; | ||||||||
| 1189 | if (RV != Visited.end()) { | ||||||||
| 1190 | V.mergeIn(RV->second); | ||||||||
| 1191 | return V; | ||||||||
| 1192 | } | ||||||||
| 1193 | } | ||||||||
| 1194 | |||||||||
| 1195 | if (LV == Visited.end() || RV == Visited.end()) { | ||||||||
| 1196 | assert(!isRevisit)((void)0); | ||||||||
| 1197 | if (LV == Visited.end()) | ||||||||
| 1198 | Worklist.push_back(L); | ||||||||
| 1199 | if (RV == Visited.end()) | ||||||||
| 1200 | Worklist.push_back(R); | ||||||||
| 1201 | return None; | ||||||||
| 1202 | } | ||||||||
| 1203 | |||||||||
| 1204 | return intersect(LV->second, RV->second); | ||||||||
| 1205 | } | ||||||||
| 1206 | |||||||||
| 1207 | ValueLatticeElement getValueFromCondition(Value *Val, Value *Cond, | ||||||||
| 1208 | bool isTrueDest) { | ||||||||
| 1209 | assert(Cond && "precondition")((void)0); | ||||||||
| 1210 | SmallDenseMap<Value*, ValueLatticeElement> Visited; | ||||||||
| 1211 | SmallVector<Value *> Worklist; | ||||||||
| 1212 | |||||||||
| 1213 | Worklist.push_back(Cond); | ||||||||
| 1214 | do { | ||||||||
| 1215 | Value *CurrentCond = Worklist.back(); | ||||||||
| 1216 | // Insert an Overdefined placeholder into the set to prevent | ||||||||
| 1217 | // infinite recursion if there exists IRs that use not | ||||||||
| 1218 | // dominated by its def as in this example: | ||||||||
| 1219 | // "%tmp3 = or i1 undef, %tmp4" | ||||||||
| 1220 | // "%tmp4 = or i1 undef, %tmp3" | ||||||||
| 1221 | auto Iter = | ||||||||
| 1222 | Visited.try_emplace(CurrentCond, ValueLatticeElement::getOverdefined()); | ||||||||
| 1223 | bool isRevisit = !Iter.second; | ||||||||
| 1224 | Optional<ValueLatticeElement> Result = getValueFromConditionImpl( | ||||||||
| 1225 | Val, CurrentCond, isTrueDest, isRevisit, Visited, Worklist); | ||||||||
| 1226 | if (Result) { | ||||||||
| 1227 | Visited[CurrentCond] = *Result; | ||||||||
| 1228 | Worklist.pop_back(); | ||||||||
| 1229 | } | ||||||||
| 1230 | } while (!Worklist.empty()); | ||||||||
| 1231 | |||||||||
| 1232 | auto Result = Visited.find(Cond); | ||||||||
| 1233 | assert(Result != Visited.end())((void)0); | ||||||||
| 1234 | return Result->second; | ||||||||
| 1235 | } | ||||||||
| 1236 | |||||||||
| 1237 | // Return true if Usr has Op as an operand, otherwise false. | ||||||||
| 1238 | static bool usesOperand(User *Usr, Value *Op) { | ||||||||
| 1239 | return is_contained(Usr->operands(), Op); | ||||||||
| 1240 | } | ||||||||
| 1241 | |||||||||
| 1242 | // Return true if the instruction type of Val is supported by | ||||||||
| 1243 | // constantFoldUser(). Currently CastInst, BinaryOperator and FreezeInst only. | ||||||||
| 1244 | // Call this before calling constantFoldUser() to find out if it's even worth | ||||||||
| 1245 | // attempting to call it. | ||||||||
| 1246 | static bool isOperationFoldable(User *Usr) { | ||||||||
| 1247 | return isa<CastInst>(Usr) || isa<BinaryOperator>(Usr) || isa<FreezeInst>(Usr); | ||||||||
| 1248 | } | ||||||||
| 1249 | |||||||||
| 1250 | // Check if Usr can be simplified to an integer constant when the value of one | ||||||||
| 1251 | // of its operands Op is an integer constant OpConstVal. If so, return it as an | ||||||||
| 1252 | // lattice value range with a single element or otherwise return an overdefined | ||||||||
| 1253 | // lattice value. | ||||||||
| 1254 | static ValueLatticeElement constantFoldUser(User *Usr, Value *Op, | ||||||||
| 1255 | const APInt &OpConstVal, | ||||||||
| 1256 | const DataLayout &DL) { | ||||||||
| 1257 | assert(isOperationFoldable(Usr) && "Precondition")((void)0); | ||||||||
| 1258 | Constant* OpConst = Constant::getIntegerValue(Op->getType(), OpConstVal); | ||||||||
| 1259 | // Check if Usr can be simplified to a constant. | ||||||||
| 1260 | if (auto *CI = dyn_cast<CastInst>(Usr)) { | ||||||||
| 1261 | assert(CI->getOperand(0) == Op && "Operand 0 isn't Op")((void)0); | ||||||||
| 1262 | if (auto *C = dyn_cast_or_null<ConstantInt>( | ||||||||
| 1263 | SimplifyCastInst(CI->getOpcode(), OpConst, | ||||||||
| 1264 | CI->getDestTy(), DL))) { | ||||||||
| 1265 | return ValueLatticeElement::getRange(ConstantRange(C->getValue())); | ||||||||
| 1266 | } | ||||||||
| 1267 | } else if (auto *BO = dyn_cast<BinaryOperator>(Usr)) { | ||||||||
| 1268 | bool Op0Match = BO->getOperand(0) == Op; | ||||||||
| 1269 | bool Op1Match = BO->getOperand(1) == Op; | ||||||||
| 1270 | assert((Op0Match || Op1Match) &&((void)0) | ||||||||
| 1271 | "Operand 0 nor Operand 1 isn't a match")((void)0); | ||||||||
| 1272 | Value *LHS = Op0Match ? OpConst : BO->getOperand(0); | ||||||||
| 1273 | Value *RHS = Op1Match ? OpConst : BO->getOperand(1); | ||||||||
| 1274 | if (auto *C = dyn_cast_or_null<ConstantInt>( | ||||||||
| 1275 | SimplifyBinOp(BO->getOpcode(), LHS, RHS, DL))) { | ||||||||
| 1276 | return ValueLatticeElement::getRange(ConstantRange(C->getValue())); | ||||||||
| 1277 | } | ||||||||
| 1278 | } else if (isa<FreezeInst>(Usr)) { | ||||||||
| 1279 | assert(cast<FreezeInst>(Usr)->getOperand(0) == Op && "Operand 0 isn't Op")((void)0); | ||||||||
| 1280 | return ValueLatticeElement::getRange(ConstantRange(OpConstVal)); | ||||||||
| 1281 | } | ||||||||
| 1282 | return ValueLatticeElement::getOverdefined(); | ||||||||
| 1283 | } | ||||||||
| 1284 | |||||||||
| 1285 | /// Compute the value of Val on the edge BBFrom -> BBTo. Returns false if | ||||||||
| 1286 | /// Val is not constrained on the edge. Result is unspecified if return value | ||||||||
| 1287 | /// is false. | ||||||||
| 1288 | static Optional<ValueLatticeElement> getEdgeValueLocal(Value *Val, | ||||||||
| 1289 | BasicBlock *BBFrom, | ||||||||
| 1290 | BasicBlock *BBTo) { | ||||||||
| 1291 | // TODO: Handle more complex conditionals. If (v == 0 || v2 < 1) is false, we | ||||||||
| 1292 | // know that v != 0. | ||||||||
| 1293 | if (BranchInst *BI
| ||||||||
| 1294 | // If this is a conditional branch and only one successor goes to BBTo, then | ||||||||
| 1295 | // we may be able to infer something from the condition. | ||||||||
| 1296 | if (BI->isConditional() && | ||||||||
| 1297 | BI->getSuccessor(0) != BI->getSuccessor(1)) { | ||||||||
| 1298 | bool isTrueDest = BI->getSuccessor(0) == BBTo; | ||||||||
| 1299 | assert(BI->getSuccessor(!isTrueDest) == BBTo &&((void)0) | ||||||||
| 1300 | "BBTo isn't a successor of BBFrom")((void)0); | ||||||||
| 1301 | Value *Condition = BI->getCondition(); | ||||||||
| 1302 | |||||||||
| 1303 | // If V is the condition of the branch itself, then we know exactly what | ||||||||
| 1304 | // it is. | ||||||||
| 1305 | if (Condition == Val) | ||||||||
| 1306 | return ValueLatticeElement::get(ConstantInt::get( | ||||||||
| 1307 | Type::getInt1Ty(Val->getContext()), isTrueDest)); | ||||||||
| 1308 | |||||||||
| 1309 | // If the condition of the branch is an equality comparison, we may be | ||||||||
| 1310 | // able to infer the value. | ||||||||
| 1311 | ValueLatticeElement Result = getValueFromCondition(Val, Condition, | ||||||||
| 1312 | isTrueDest); | ||||||||
| 1313 | if (!Result.isOverdefined()) | ||||||||
| 1314 | return Result; | ||||||||
| 1315 | |||||||||
| 1316 | if (User *Usr
| ||||||||
| 1317 | assert(Result.isOverdefined() && "Result isn't overdefined")((void)0); | ||||||||
| 1318 | // Check with isOperationFoldable() first to avoid linearly iterating | ||||||||
| 1319 | // over the operands unnecessarily which can be expensive for | ||||||||
| 1320 | // instructions with many operands. | ||||||||
| 1321 | if (isa<IntegerType>(Usr->getType()) && isOperationFoldable(Usr)) { | ||||||||
| 1322 | const DataLayout &DL = BBTo->getModule()->getDataLayout(); | ||||||||
| |||||||||
| 1323 | if (usesOperand(Usr, Condition)) { | ||||||||
| 1324 | // If Val has Condition as an operand and Val can be folded into a | ||||||||
| 1325 | // constant with either Condition == true or Condition == false, | ||||||||
| 1326 | // propagate the constant. | ||||||||
| 1327 | // eg. | ||||||||
| 1328 | // ; %Val is true on the edge to %then. | ||||||||
| 1329 | // %Val = and i1 %Condition, true. | ||||||||
| 1330 | // br %Condition, label %then, label %else | ||||||||
| 1331 | APInt ConditionVal(1, isTrueDest ? 1 : 0); | ||||||||
| 1332 | Result = constantFoldUser(Usr, Condition, ConditionVal, DL); | ||||||||
| 1333 | } else { | ||||||||
| 1334 | // If one of Val's operand has an inferred value, we may be able to | ||||||||
| 1335 | // infer the value of Val. | ||||||||
| 1336 | // eg. | ||||||||
| 1337 | // ; %Val is 94 on the edge to %then. | ||||||||
| 1338 | // %Val = add i8 %Op, 1 | ||||||||
| 1339 | // %Condition = icmp eq i8 %Op, 93 | ||||||||
| 1340 | // br i1 %Condition, label %then, label %else | ||||||||
| 1341 | for (unsigned i = 0; i < Usr->getNumOperands(); ++i) { | ||||||||
| 1342 | Value *Op = Usr->getOperand(i); | ||||||||
| 1343 | ValueLatticeElement OpLatticeVal = | ||||||||
| 1344 | getValueFromCondition(Op, Condition, isTrueDest); | ||||||||
| 1345 | if (Optional<APInt> OpConst = OpLatticeVal.asConstantInteger()) { | ||||||||
| 1346 | Result = constantFoldUser(Usr, Op, OpConst.getValue(), DL); | ||||||||
| 1347 | break; | ||||||||
| 1348 | } | ||||||||
| 1349 | } | ||||||||
| 1350 | } | ||||||||
| 1351 | } | ||||||||
| 1352 | } | ||||||||
| 1353 | if (!Result.isOverdefined()) | ||||||||
| 1354 | return Result; | ||||||||
| 1355 | } | ||||||||
| 1356 | } | ||||||||
| 1357 | |||||||||
| 1358 | // If the edge was formed by a switch on the value, then we may know exactly | ||||||||
| 1359 | // what it is. | ||||||||
| 1360 | if (SwitchInst *SI = dyn_cast<SwitchInst>(BBFrom->getTerminator())) { | ||||||||
| 1361 | Value *Condition = SI->getCondition(); | ||||||||
| 1362 | if (!isa<IntegerType>(Val->getType())) | ||||||||
| 1363 | return None; | ||||||||
| 1364 | bool ValUsesConditionAndMayBeFoldable = false; | ||||||||
| 1365 | if (Condition != Val) { | ||||||||
| 1366 | // Check if Val has Condition as an operand. | ||||||||
| 1367 | if (User *Usr = dyn_cast<User>(Val)) | ||||||||
| 1368 | ValUsesConditionAndMayBeFoldable = isOperationFoldable(Usr) && | ||||||||
| 1369 | usesOperand(Usr, Condition); | ||||||||
| 1370 | if (!ValUsesConditionAndMayBeFoldable) | ||||||||
| 1371 | return None; | ||||||||
| 1372 | } | ||||||||
| 1373 | assert((Condition == Val || ValUsesConditionAndMayBeFoldable) &&((void)0) | ||||||||
| 1374 | "Condition != Val nor Val doesn't use Condition")((void)0); | ||||||||
| 1375 | |||||||||
| 1376 | bool DefaultCase = SI->getDefaultDest() == BBTo; | ||||||||
| 1377 | unsigned BitWidth = Val->getType()->getIntegerBitWidth(); | ||||||||
| 1378 | ConstantRange EdgesVals(BitWidth, DefaultCase/*isFullSet*/); | ||||||||
| 1379 | |||||||||
| 1380 | for (auto Case : SI->cases()) { | ||||||||
| 1381 | APInt CaseValue = Case.getCaseValue()->getValue(); | ||||||||
| 1382 | ConstantRange EdgeVal(CaseValue); | ||||||||
| 1383 | if (ValUsesConditionAndMayBeFoldable) { | ||||||||
| 1384 | User *Usr = cast<User>(Val); | ||||||||
| 1385 | const DataLayout &DL = BBTo->getModule()->getDataLayout(); | ||||||||
| 1386 | ValueLatticeElement EdgeLatticeVal = | ||||||||
| 1387 | constantFoldUser(Usr, Condition, CaseValue, DL); | ||||||||
| 1388 | if (EdgeLatticeVal.isOverdefined()) | ||||||||
| 1389 | return None; | ||||||||
| 1390 | EdgeVal = EdgeLatticeVal.getConstantRange(); | ||||||||
| 1391 | } | ||||||||
| 1392 | if (DefaultCase) { | ||||||||
| 1393 | // It is possible that the default destination is the destination of | ||||||||
| 1394 | // some cases. We cannot perform difference for those cases. | ||||||||
| 1395 | // We know Condition != CaseValue in BBTo. In some cases we can use | ||||||||
| 1396 | // this to infer Val == f(Condition) is != f(CaseValue). For now, we | ||||||||
| 1397 | // only do this when f is identity (i.e. Val == Condition), but we | ||||||||
| 1398 | // should be able to do this for any injective f. | ||||||||
| 1399 | if (Case.getCaseSuccessor() != BBTo && Condition == Val) | ||||||||
| 1400 | EdgesVals = EdgesVals.difference(EdgeVal); | ||||||||
| 1401 | } else if (Case.getCaseSuccessor() == BBTo) | ||||||||
| 1402 | EdgesVals = EdgesVals.unionWith(EdgeVal); | ||||||||
| 1403 | } | ||||||||
| 1404 | return ValueLatticeElement::getRange(std::move(EdgesVals)); | ||||||||
| 1405 | } | ||||||||
| 1406 | return None; | ||||||||
| 1407 | } | ||||||||
| 1408 | |||||||||
| 1409 | /// Compute the value of Val on the edge BBFrom -> BBTo or the value at | ||||||||
| 1410 | /// the basic block if the edge does not constrain Val. | ||||||||
| 1411 | Optional<ValueLatticeElement> LazyValueInfoImpl::getEdgeValue( | ||||||||
| 1412 | Value *Val, BasicBlock *BBFrom, BasicBlock *BBTo, Instruction *CxtI) { | ||||||||
| 1413 | // If already a constant, there is nothing to compute. | ||||||||
| 1414 | if (Constant *VC
| ||||||||
| 1415 | return ValueLatticeElement::get(VC); | ||||||||
| 1416 | |||||||||
| 1417 | ValueLatticeElement LocalResult = getEdgeValueLocal(Val, BBFrom, BBTo) | ||||||||
| 1418 | .getValueOr(ValueLatticeElement::getOverdefined()); | ||||||||
| 1419 | if (hasSingleValue(LocalResult)) | ||||||||
| 1420 | // Can't get any more precise here | ||||||||
| 1421 | return LocalResult; | ||||||||
| 1422 | |||||||||
| 1423 | Optional<ValueLatticeElement> OptInBlock = getBlockValue(Val, BBFrom); | ||||||||
| 1424 | if (!OptInBlock) | ||||||||
| 1425 | return None; | ||||||||
| 1426 | ValueLatticeElement &InBlock = *OptInBlock; | ||||||||
| 1427 | |||||||||
| 1428 | // Try to intersect ranges of the BB and the constraint on the edge. | ||||||||
| 1429 | intersectAssumeOrGuardBlockValueConstantRange(Val, InBlock, | ||||||||
| 1430 | BBFrom->getTerminator()); | ||||||||
| 1431 | // We can use the context instruction (generically the ultimate instruction | ||||||||
| 1432 | // the calling pass is trying to simplify) here, even though the result of | ||||||||
| 1433 | // this function is generally cached when called from the solve* functions | ||||||||
| 1434 | // (and that cached result might be used with queries using a different | ||||||||
| 1435 | // context instruction), because when this function is called from the solve* | ||||||||
| 1436 | // functions, the context instruction is not provided. When called from | ||||||||
| 1437 | // LazyValueInfoImpl::getValueOnEdge, the context instruction is provided, | ||||||||
| 1438 | // but then the result is not cached. | ||||||||
| 1439 | intersectAssumeOrGuardBlockValueConstantRange(Val, InBlock, CxtI); | ||||||||
| 1440 | |||||||||
| 1441 | return intersect(LocalResult, InBlock); | ||||||||
| 1442 | } | ||||||||
| 1443 | |||||||||
| 1444 | ValueLatticeElement LazyValueInfoImpl::getValueInBlock(Value *V, BasicBlock *BB, | ||||||||
| 1445 | Instruction *CxtI) { | ||||||||
| 1446 | LLVM_DEBUG(dbgs() << "LVI Getting block end value " << *V << " at '"do { } while (false) | ||||||||
| 1447 | << BB->getName() << "'\n")do { } while (false); | ||||||||
| 1448 | |||||||||
| 1449 | assert(BlockValueStack.empty() && BlockValueSet.empty())((void)0); | ||||||||
| 1450 | Optional<ValueLatticeElement> OptResult = getBlockValue(V, BB); | ||||||||
| 1451 | if (!OptResult) { | ||||||||
| 1452 | solve(); | ||||||||
| 1453 | OptResult = getBlockValue(V, BB); | ||||||||
| 1454 | assert(OptResult && "Value not available after solving")((void)0); | ||||||||
| 1455 | } | ||||||||
| 1456 | ValueLatticeElement Result = *OptResult; | ||||||||
| 1457 | intersectAssumeOrGuardBlockValueConstantRange(V, Result, CxtI); | ||||||||
| 1458 | |||||||||
| 1459 | LLVM_DEBUG(dbgs() << " Result = " << Result << "\n")do { } while (false); | ||||||||
| 1460 | return Result; | ||||||||
| 1461 | } | ||||||||
| 1462 | |||||||||
| 1463 | ValueLatticeElement LazyValueInfoImpl::getValueAt(Value *V, Instruction *CxtI) { | ||||||||
| 1464 | LLVM_DEBUG(dbgs() << "LVI Getting value " << *V << " at '" << CxtI->getName()do { } while (false) | ||||||||
| 1465 | << "'\n")do { } while (false); | ||||||||
| 1466 | |||||||||
| 1467 | if (auto *C
| ||||||||
| 1468 | return ValueLatticeElement::get(C); | ||||||||
| 1469 | |||||||||
| 1470 | ValueLatticeElement Result = ValueLatticeElement::getOverdefined(); | ||||||||
| 1471 | if (auto *I
| ||||||||
| 1472 | Result = getFromRangeMetadata(I); | ||||||||
| 1473 | intersectAssumeOrGuardBlockValueConstantRange(V, Result, CxtI); | ||||||||
| 1474 | |||||||||
| 1475 | LLVM_DEBUG(dbgs() << " Result = " << Result << "\n")do { } while (false); | ||||||||
| 1476 | return Result; | ||||||||
| 1477 | } | ||||||||
| 1478 | |||||||||
| 1479 | ValueLatticeElement LazyValueInfoImpl:: | ||||||||
| 1480 | getValueOnEdge(Value *V, BasicBlock *FromBB, BasicBlock *ToBB, | ||||||||
| 1481 | Instruction *CxtI) { | ||||||||
| 1482 | LLVM_DEBUG(dbgs() << "LVI Getting edge value " << *V << " from '"do { } while (false) | ||||||||
| 1483 | << FromBB->getName() << "' to '" << ToBB->getName()do { } while (false) | ||||||||
| 1484 | << "'\n")do { } while (false); | ||||||||
| 1485 | |||||||||
| 1486 | Optional<ValueLatticeElement> Result = getEdgeValue(V, FromBB, ToBB, CxtI); | ||||||||
| 1487 | if (!Result) { | ||||||||
| 1488 | solve(); | ||||||||
| 1489 | Result = getEdgeValue(V, FromBB, ToBB, CxtI); | ||||||||
| 1490 | assert(Result && "More work to do after problem solved?")((void)0); | ||||||||
| 1491 | } | ||||||||
| 1492 | |||||||||
| 1493 | LLVM_DEBUG(dbgs() << " Result = " << *Result << "\n")do { } while (false); | ||||||||
| 1494 | return *Result; | ||||||||
| 1495 | } | ||||||||
| 1496 | |||||||||
| 1497 | void LazyValueInfoImpl::threadEdge(BasicBlock *PredBB, BasicBlock *OldSucc, | ||||||||
| 1498 | BasicBlock *NewSucc) { | ||||||||
| 1499 | TheCache.threadEdgeImpl(OldSucc, NewSucc); | ||||||||
| 1500 | } | ||||||||
| 1501 | |||||||||
| 1502 | //===----------------------------------------------------------------------===// | ||||||||
| 1503 | // LazyValueInfo Impl | ||||||||
| 1504 | //===----------------------------------------------------------------------===// | ||||||||
| 1505 | |||||||||
| 1506 | /// This lazily constructs the LazyValueInfoImpl. | ||||||||
| 1507 | static LazyValueInfoImpl &getImpl(void *&PImpl, AssumptionCache *AC, | ||||||||
| 1508 | const Module *M) { | ||||||||
| 1509 | if (!PImpl) { | ||||||||
| 1510 | assert(M && "getCache() called with a null Module")((void)0); | ||||||||
| 1511 | const DataLayout &DL = M->getDataLayout(); | ||||||||
| 1512 | Function *GuardDecl = M->getFunction( | ||||||||
| 1513 | Intrinsic::getName(Intrinsic::experimental_guard)); | ||||||||
| 1514 | PImpl = new LazyValueInfoImpl(AC, DL, GuardDecl); | ||||||||
| 1515 | } | ||||||||
| 1516 | return *static_cast<LazyValueInfoImpl*>(PImpl); | ||||||||
| 1517 | } | ||||||||
| 1518 | |||||||||
| 1519 | bool LazyValueInfoWrapperPass::runOnFunction(Function &F) { | ||||||||
| 1520 | Info.AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F); | ||||||||
| 1521 | Info.TLI = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(F); | ||||||||
| 1522 | |||||||||
| 1523 | if (Info.PImpl) | ||||||||
| 1524 | getImpl(Info.PImpl, Info.AC, F.getParent()).clear(); | ||||||||
| 1525 | |||||||||
| 1526 | // Fully lazy. | ||||||||
| 1527 | return false; | ||||||||
| 1528 | } | ||||||||
| 1529 | |||||||||
| 1530 | void LazyValueInfoWrapperPass::getAnalysisUsage(AnalysisUsage &AU) const { | ||||||||
| 1531 | AU.setPreservesAll(); | ||||||||
| 1532 | AU.addRequired<AssumptionCacheTracker>(); | ||||||||
| 1533 | AU.addRequired<TargetLibraryInfoWrapperPass>(); | ||||||||
| 1534 | } | ||||||||
| 1535 | |||||||||
| 1536 | LazyValueInfo &LazyValueInfoWrapperPass::getLVI() { return Info; } | ||||||||
| 1537 | |||||||||
| 1538 | LazyValueInfo::~LazyValueInfo() { releaseMemory(); } | ||||||||
| 1539 | |||||||||
| 1540 | void LazyValueInfo::releaseMemory() { | ||||||||
| 1541 | // If the cache was allocated, free it. | ||||||||
| 1542 | if (PImpl) { | ||||||||
| 1543 | delete &getImpl(PImpl, AC, nullptr); | ||||||||
| 1544 | PImpl = nullptr; | ||||||||
| 1545 | } | ||||||||
| 1546 | } | ||||||||
| 1547 | |||||||||
| 1548 | bool LazyValueInfo::invalidate(Function &F, const PreservedAnalyses &PA, | ||||||||
| 1549 | FunctionAnalysisManager::Invalidator &Inv) { | ||||||||
| 1550 | // We need to invalidate if we have either failed to preserve this analyses | ||||||||
| 1551 | // result directly or if any of its dependencies have been invalidated. | ||||||||
| 1552 | auto PAC = PA.getChecker<LazyValueAnalysis>(); | ||||||||
| 1553 | if (!(PAC.preserved() || PAC.preservedSet<AllAnalysesOn<Function>>())) | ||||||||
| 1554 | return true; | ||||||||
| 1555 | |||||||||
| 1556 | return false; | ||||||||
| 1557 | } | ||||||||
| 1558 | |||||||||
| 1559 | void LazyValueInfoWrapperPass::releaseMemory() { Info.releaseMemory(); } | ||||||||
| 1560 | |||||||||
| 1561 | LazyValueInfo LazyValueAnalysis::run(Function &F, | ||||||||
| 1562 | FunctionAnalysisManager &FAM) { | ||||||||
| 1563 | auto &AC = FAM.getResult<AssumptionAnalysis>(F); | ||||||||
| 1564 | auto &TLI = FAM.getResult<TargetLibraryAnalysis>(F); | ||||||||
| 1565 | |||||||||
| 1566 | return LazyValueInfo(&AC, &F.getParent()->getDataLayout(), &TLI); | ||||||||
| 1567 | } | ||||||||
| 1568 | |||||||||
| 1569 | /// Returns true if we can statically tell that this value will never be a | ||||||||
| 1570 | /// "useful" constant. In practice, this means we've got something like an | ||||||||
| 1571 | /// alloca or a malloc call for which a comparison against a constant can | ||||||||
| 1572 | /// only be guarding dead code. Note that we are potentially giving up some | ||||||||
| 1573 | /// precision in dead code (a constant result) in favour of avoiding a | ||||||||
| 1574 | /// expensive search for a easily answered common query. | ||||||||
| 1575 | static bool isKnownNonConstant(Value *V) { | ||||||||
| 1576 | V = V->stripPointerCasts(); | ||||||||
| 1577 | // The return val of alloc cannot be a Constant. | ||||||||
| 1578 | if (isa<AllocaInst>(V)) | ||||||||
| 1579 | return true; | ||||||||
| 1580 | return false; | ||||||||
| 1581 | } | ||||||||
| 1582 | |||||||||
| 1583 | Constant *LazyValueInfo::getConstant(Value *V, Instruction *CxtI) { | ||||||||
| 1584 | // Bail out early if V is known not to be a Constant. | ||||||||
| 1585 | if (isKnownNonConstant(V)) | ||||||||
| 1586 | return nullptr; | ||||||||
| 1587 | |||||||||
| 1588 | BasicBlock *BB = CxtI->getParent(); | ||||||||
| 1589 | ValueLatticeElement Result = | ||||||||
| 1590 | getImpl(PImpl, AC, BB->getModule()).getValueInBlock(V, BB, CxtI); | ||||||||
| 1591 | |||||||||
| 1592 | if (Result.isConstant()) | ||||||||
| 1593 | return Result.getConstant(); | ||||||||
| 1594 | if (Result.isConstantRange()) { | ||||||||
| 1595 | const ConstantRange &CR = Result.getConstantRange(); | ||||||||
| 1596 | if (const APInt *SingleVal = CR.getSingleElement()) | ||||||||
| 1597 | return ConstantInt::get(V->getContext(), *SingleVal); | ||||||||
| 1598 | } | ||||||||
| 1599 | return nullptr; | ||||||||
| 1600 | } | ||||||||
| 1601 | |||||||||
| 1602 | ConstantRange LazyValueInfo::getConstantRange(Value *V, Instruction *CxtI, | ||||||||
| 1603 | bool UndefAllowed) { | ||||||||
| 1604 | assert(V->getType()->isIntegerTy())((void)0); | ||||||||
| 1605 | unsigned Width = V->getType()->getIntegerBitWidth(); | ||||||||
| 1606 | BasicBlock *BB = CxtI->getParent(); | ||||||||
| 1607 | ValueLatticeElement Result = | ||||||||
| 1608 | getImpl(PImpl, AC, BB->getModule()).getValueInBlock(V, BB, CxtI); | ||||||||
| 1609 | if (Result.isUnknown()) | ||||||||
| 1610 | return ConstantRange::getEmpty(Width); | ||||||||
| 1611 | if (Result.isConstantRange(UndefAllowed)) | ||||||||
| 1612 | return Result.getConstantRange(UndefAllowed); | ||||||||
| 1613 | // We represent ConstantInt constants as constant ranges but other kinds | ||||||||
| 1614 | // of integer constants, i.e. ConstantExpr will be tagged as constants | ||||||||
| 1615 | assert(!(Result.isConstant() && isa<ConstantInt>(Result.getConstant())) &&((void)0) | ||||||||
| 1616 | "ConstantInt value must be represented as constantrange")((void)0); | ||||||||
| 1617 | return ConstantRange::getFull(Width); | ||||||||
| 1618 | } | ||||||||
| 1619 | |||||||||
| 1620 | /// Determine whether the specified value is known to be a | ||||||||
| 1621 | /// constant on the specified edge. Return null if not. | ||||||||
| 1622 | Constant *LazyValueInfo::getConstantOnEdge(Value *V, BasicBlock *FromBB, | ||||||||
| 1623 | BasicBlock *ToBB, | ||||||||
| 1624 | Instruction *CxtI) { | ||||||||
| 1625 | Module *M = FromBB->getModule(); | ||||||||
| 1626 | ValueLatticeElement Result = | ||||||||
| 1627 | getImpl(PImpl, AC, M).getValueOnEdge(V, FromBB, ToBB, CxtI); | ||||||||
| 1628 | |||||||||
| 1629 | if (Result.isConstant()) | ||||||||
| 1630 | return Result.getConstant(); | ||||||||
| 1631 | if (Result.isConstantRange()) { | ||||||||
| 1632 | const ConstantRange &CR = Result.getConstantRange(); | ||||||||
| 1633 | if (const APInt *SingleVal = CR.getSingleElement()) | ||||||||
| 1634 | return ConstantInt::get(V->getContext(), *SingleVal); | ||||||||
| 1635 | } | ||||||||
| 1636 | return nullptr; | ||||||||
| 1637 | } | ||||||||
| 1638 | |||||||||
| 1639 | ConstantRange LazyValueInfo::getConstantRangeOnEdge(Value *V, | ||||||||
| 1640 | BasicBlock *FromBB, | ||||||||
| 1641 | BasicBlock *ToBB, | ||||||||
| 1642 | Instruction *CxtI) { | ||||||||
| 1643 | unsigned Width = V->getType()->getIntegerBitWidth(); | ||||||||
| 1644 | Module *M = FromBB->getModule(); | ||||||||
| 1645 | ValueLatticeElement Result = | ||||||||
| 1646 | getImpl(PImpl, AC, M).getValueOnEdge(V, FromBB, ToBB, CxtI); | ||||||||
| 1647 | |||||||||
| 1648 | if (Result.isUnknown()) | ||||||||
| 1649 | return ConstantRange::getEmpty(Width); | ||||||||
| 1650 | if (Result.isConstantRange()) | ||||||||
| 1651 | return Result.getConstantRange(); | ||||||||
| 1652 | // We represent ConstantInt constants as constant ranges but other kinds | ||||||||
| 1653 | // of integer constants, i.e. ConstantExpr will be tagged as constants | ||||||||
| 1654 | assert(!(Result.isConstant() && isa<ConstantInt>(Result.getConstant())) &&((void)0) | ||||||||
| 1655 | "ConstantInt value must be represented as constantrange")((void)0); | ||||||||
| 1656 | return ConstantRange::getFull(Width); | ||||||||
| 1657 | } | ||||||||
| 1658 | |||||||||
| 1659 | static LazyValueInfo::Tristate | ||||||||
| 1660 | getPredicateResult(unsigned Pred, Constant *C, const ValueLatticeElement &Val, | ||||||||
| 1661 | const DataLayout &DL, TargetLibraryInfo *TLI) { | ||||||||
| 1662 | // If we know the value is a constant, evaluate the conditional. | ||||||||
| 1663 | Constant *Res = nullptr; | ||||||||
| 1664 | if (Val.isConstant()) { | ||||||||
| 1665 | Res = ConstantFoldCompareInstOperands(Pred, Val.getConstant(), C, DL, TLI); | ||||||||
| 1666 | if (ConstantInt *ResCI = dyn_cast<ConstantInt>(Res)) | ||||||||
| 1667 | return ResCI->isZero() ? LazyValueInfo::False : LazyValueInfo::True; | ||||||||
| 1668 | return LazyValueInfo::Unknown; | ||||||||
| 1669 | } | ||||||||
| 1670 | |||||||||
| 1671 | if (Val.isConstantRange()) { | ||||||||
| 1672 | ConstantInt *CI = dyn_cast<ConstantInt>(C); | ||||||||
| 1673 | if (!CI) return LazyValueInfo::Unknown; | ||||||||
| 1674 | |||||||||
| 1675 | const ConstantRange &CR = Val.getConstantRange(); | ||||||||
| 1676 | if (Pred == ICmpInst::ICMP_EQ) { | ||||||||
| 1677 | if (!CR.contains(CI->getValue())) | ||||||||
| 1678 | return LazyValueInfo::False; | ||||||||
| 1679 | |||||||||
| 1680 | if (CR.isSingleElement()) | ||||||||
| 1681 | return LazyValueInfo::True; | ||||||||
| 1682 | } else if (Pred == ICmpInst::ICMP_NE) { | ||||||||
| 1683 | if (!CR.contains(CI->getValue())) | ||||||||
| 1684 | return LazyValueInfo::True; | ||||||||
| 1685 | |||||||||
| 1686 | if (CR.isSingleElement()) | ||||||||
| 1687 | return LazyValueInfo::False; | ||||||||
| 1688 | } else { | ||||||||
| 1689 | // Handle more complex predicates. | ||||||||
| 1690 | ConstantRange TrueValues = ConstantRange::makeExactICmpRegion( | ||||||||
| 1691 | (ICmpInst::Predicate)Pred, CI->getValue()); | ||||||||
| 1692 | if (TrueValues.contains(CR)) | ||||||||
| 1693 | return LazyValueInfo::True; | ||||||||
| 1694 | if (TrueValues.inverse().contains(CR)) | ||||||||
| 1695 | return LazyValueInfo::False; | ||||||||
| 1696 | } | ||||||||
| 1697 | return LazyValueInfo::Unknown; | ||||||||
| 1698 | } | ||||||||
| 1699 | |||||||||
| 1700 | if (Val.isNotConstant()) { | ||||||||
| 1701 | // If this is an equality comparison, we can try to fold it knowing that | ||||||||
| 1702 | // "V != C1". | ||||||||
| 1703 | if (Pred == ICmpInst::ICMP_EQ) { | ||||||||
| 1704 | // !C1 == C -> false iff C1 == C. | ||||||||
| 1705 | Res = ConstantFoldCompareInstOperands(ICmpInst::ICMP_NE, | ||||||||
| 1706 | Val.getNotConstant(), C, DL, | ||||||||
| 1707 | TLI); | ||||||||
| 1708 | if (Res->isNullValue()) | ||||||||
| 1709 | return LazyValueInfo::False; | ||||||||
| 1710 | } else if (Pred == ICmpInst::ICMP_NE) { | ||||||||
| 1711 | // !C1 != C -> true iff C1 == C. | ||||||||
| 1712 | Res = ConstantFoldCompareInstOperands(ICmpInst::ICMP_NE, | ||||||||
| 1713 | Val.getNotConstant(), C, DL, | ||||||||
| 1714 | TLI); | ||||||||
| 1715 | if (Res->isNullValue()) | ||||||||
| 1716 | return LazyValueInfo::True; | ||||||||
| 1717 | } | ||||||||
| 1718 | return LazyValueInfo::Unknown; | ||||||||
| 1719 | } | ||||||||
| 1720 | |||||||||
| 1721 | return LazyValueInfo::Unknown; | ||||||||
| 1722 | } | ||||||||
| 1723 | |||||||||
| 1724 | /// Determine whether the specified value comparison with a constant is known to | ||||||||
| 1725 | /// be true or false on the specified CFG edge. Pred is a CmpInst predicate. | ||||||||
| 1726 | LazyValueInfo::Tristate | ||||||||
| 1727 | LazyValueInfo::getPredicateOnEdge(unsigned Pred, Value *V, Constant *C, | ||||||||
| 1728 | BasicBlock *FromBB, BasicBlock *ToBB, | ||||||||
| 1729 | Instruction *CxtI) { | ||||||||
| 1730 | Module *M = FromBB->getModule(); | ||||||||
| 1731 | ValueLatticeElement Result = | ||||||||
| 1732 | getImpl(PImpl, AC, M).getValueOnEdge(V, FromBB, ToBB, CxtI); | ||||||||
| 1733 | |||||||||
| 1734 | return getPredicateResult(Pred, C, Result, M->getDataLayout(), TLI); | ||||||||
| 1735 | } | ||||||||
| 1736 | |||||||||
| 1737 | LazyValueInfo::Tristate | ||||||||
| 1738 | LazyValueInfo::getPredicateAt(unsigned Pred, Value *V, Constant *C, | ||||||||
| 1739 | Instruction *CxtI, bool UseBlockValue) { | ||||||||
| 1740 | // Is or is not NonNull are common predicates being queried. If | ||||||||
| 1741 | // isKnownNonZero can tell us the result of the predicate, we can | ||||||||
| 1742 | // return it quickly. But this is only a fastpath, and falling | ||||||||
| 1743 | // through would still be correct. | ||||||||
| 1744 | Module *M = CxtI->getModule(); | ||||||||
| 1745 | const DataLayout &DL = M->getDataLayout(); | ||||||||
| 1746 | if (V->getType()->isPointerTy() && C->isNullValue() && | ||||||||
| 1747 | isKnownNonZero(V->stripPointerCastsSameRepresentation(), DL)) { | ||||||||
| 1748 | if (Pred == ICmpInst::ICMP_EQ) | ||||||||
| 1749 | return LazyValueInfo::False; | ||||||||
| 1750 | else if (Pred == ICmpInst::ICMP_NE) | ||||||||
| 1751 | return LazyValueInfo::True; | ||||||||
| 1752 | } | ||||||||
| 1753 | |||||||||
| 1754 | ValueLatticeElement Result = UseBlockValue | ||||||||
| 1755 | ? getImpl(PImpl, AC, M).getValueInBlock(V, CxtI->getParent(), CxtI) | ||||||||
| 1756 | : getImpl(PImpl, AC, M).getValueAt(V, CxtI); | ||||||||
| 1757 | Tristate Ret = getPredicateResult(Pred, C, Result, DL, TLI); | ||||||||
| 1758 | if (Ret
| ||||||||
| 1759 | return Ret; | ||||||||
| 1760 | |||||||||
| 1761 | // Note: The following bit of code is somewhat distinct from the rest of LVI; | ||||||||
| 1762 | // LVI as a whole tries to compute a lattice value which is conservatively | ||||||||
| 1763 | // correct at a given location. In this case, we have a predicate which we | ||||||||
| 1764 | // weren't able to prove about the merged result, and we're pushing that | ||||||||
| 1765 | // predicate back along each incoming edge to see if we can prove it | ||||||||
| 1766 | // separately for each input. As a motivating example, consider: | ||||||||
| 1767 | // bb1: | ||||||||
| 1768 | // %v1 = ... ; constantrange<1, 5> | ||||||||
| 1769 | // br label %merge | ||||||||
| 1770 | // bb2: | ||||||||
| 1771 | // %v2 = ... ; constantrange<10, 20> | ||||||||
| 1772 | // br label %merge | ||||||||
| 1773 | // merge: | ||||||||
| 1774 | // %phi = phi [%v1, %v2] ; constantrange<1,20> | ||||||||
| 1775 | // %pred = icmp eq i32 %phi, 8 | ||||||||
| 1776 | // We can't tell from the lattice value for '%phi' that '%pred' is false | ||||||||
| 1777 | // along each path, but by checking the predicate over each input separately, | ||||||||
| 1778 | // we can. | ||||||||
| 1779 | // We limit the search to one step backwards from the current BB and value. | ||||||||
| 1780 | // We could consider extending this to search further backwards through the | ||||||||
| 1781 | // CFG and/or value graph, but there are non-obvious compile time vs quality | ||||||||
| 1782 | // tradeoffs. | ||||||||
| 1783 | if (CxtI
| ||||||||
| 1784 | BasicBlock *BB = CxtI->getParent(); | ||||||||
| 1785 | |||||||||
| 1786 | // Function entry or an unreachable block. Bail to avoid confusing | ||||||||
| 1787 | // analysis below. | ||||||||
| 1788 | pred_iterator PI = pred_begin(BB), PE = pred_end(BB); | ||||||||
| 1789 | if (PI == PE) | ||||||||
| 1790 | return Unknown; | ||||||||
| 1791 | |||||||||
| 1792 | // If V is a PHI node in the same block as the context, we need to ask | ||||||||
| 1793 | // questions about the predicate as applied to the incoming value along | ||||||||
| 1794 | // each edge. This is useful for eliminating cases where the predicate is | ||||||||
| 1795 | // known along all incoming edges. | ||||||||
| 1796 | if (auto *PHI
| ||||||||
| 1797 | if (PHI->getParent() == BB) { | ||||||||
| 1798 | Tristate Baseline = Unknown; | ||||||||
| 1799 | for (unsigned i = 0, e = PHI->getNumIncomingValues(); i < e; i++) { | ||||||||
| 1800 | Value *Incoming = PHI->getIncomingValue(i); | ||||||||
| 1801 | BasicBlock *PredBB = PHI->getIncomingBlock(i); | ||||||||
| 1802 | // Note that PredBB may be BB itself. | ||||||||
| 1803 | Tristate Result = getPredicateOnEdge(Pred, Incoming, C, PredBB, BB, | ||||||||
| 1804 | CxtI); | ||||||||
| 1805 | |||||||||
| 1806 | // Keep going as long as we've seen a consistent known result for | ||||||||
| 1807 | // all inputs. | ||||||||
| 1808 | Baseline = (i == 0) ? Result /* First iteration */ | ||||||||
| 1809 | : (Baseline == Result ? Baseline : Unknown); /* All others */ | ||||||||
| 1810 | if (Baseline == Unknown) | ||||||||
| 1811 | break; | ||||||||
| 1812 | } | ||||||||
| 1813 | if (Baseline != Unknown) | ||||||||
| 1814 | return Baseline; | ||||||||
| 1815 | } | ||||||||
| 1816 | |||||||||
| 1817 | // For a comparison where the V is outside this block, it's possible | ||||||||
| 1818 | // that we've branched on it before. Look to see if the value is known | ||||||||
| 1819 | // on all incoming edges. | ||||||||
| 1820 | if (!isa<Instruction>(V) || | ||||||||
| 1821 | cast<Instruction>(V)->getParent() != BB) { | ||||||||
| 1822 | // For predecessor edge, determine if the comparison is true or false | ||||||||
| 1823 | // on that edge. If they're all true or all false, we can conclude | ||||||||
| 1824 | // the value of the comparison in this block. | ||||||||
| 1825 | Tristate Baseline = getPredicateOnEdge(Pred, V, C, *PI, BB, CxtI); | ||||||||
| 1826 | if (Baseline != Unknown) { | ||||||||
| 1827 | // Check that all remaining incoming values match the first one. | ||||||||
| 1828 | while (++PI != PE) { | ||||||||
| 1829 | Tristate Ret = getPredicateOnEdge(Pred, V, C, *PI, BB, CxtI); | ||||||||
| 1830 | if (Ret != Baseline) break; | ||||||||
| 1831 | } | ||||||||
| 1832 | // If we terminated early, then one of the values didn't match. | ||||||||
| 1833 | if (PI == PE) { | ||||||||
| 1834 | return Baseline; | ||||||||
| 1835 | } | ||||||||
| 1836 | } | ||||||||
| 1837 | } | ||||||||
| 1838 | } | ||||||||
| 1839 | return Unknown; | ||||||||
| 1840 | } | ||||||||
| 1841 | |||||||||
| 1842 | LazyValueInfo::Tristate LazyValueInfo::getPredicateAt(unsigned P, Value *LHS, | ||||||||
| 1843 | Value *RHS, | ||||||||
| 1844 | Instruction *CxtI, | ||||||||
| 1845 | bool UseBlockValue) { | ||||||||
| 1846 | CmpInst::Predicate Pred = (CmpInst::Predicate)P; | ||||||||
| 1847 | |||||||||
| 1848 | if (auto *C
| ||||||||
| |||||||||
| 1849 | return getPredicateAt(P, LHS, C, CxtI, UseBlockValue); | ||||||||
| 1850 | if (auto *C = dyn_cast<Constant>(LHS)) | ||||||||
| 1851 | return getPredicateAt(CmpInst::getSwappedPredicate(Pred), RHS, C, CxtI, | ||||||||
| 1852 | UseBlockValue); | ||||||||
| 1853 | |||||||||
| 1854 | // Got two non-Constant values. While we could handle them somewhat, | ||||||||
| 1855 | // by getting their constant ranges, and applying ConstantRange::icmp(), | ||||||||
| 1856 | // so far it did not appear to be profitable. | ||||||||
| 1857 | return LazyValueInfo::Unknown; | ||||||||
| 1858 | } | ||||||||
| 1859 | |||||||||
| 1860 | void LazyValueInfo::threadEdge(BasicBlock *PredBB, BasicBlock *OldSucc, | ||||||||
| 1861 | BasicBlock *NewSucc) { | ||||||||
| 1862 | if (PImpl) { | ||||||||
| 1863 | getImpl(PImpl, AC, PredBB->getModule()) | ||||||||
| 1864 | .threadEdge(PredBB, OldSucc, NewSucc); | ||||||||
| 1865 | } | ||||||||
| 1866 | } | ||||||||
| 1867 | |||||||||
| 1868 | void LazyValueInfo::eraseBlock(BasicBlock *BB) { | ||||||||
| 1869 | if (PImpl) { | ||||||||
| 1870 | getImpl(PImpl, AC, BB->getModule()).eraseBlock(BB); | ||||||||
| 1871 | } | ||||||||
| 1872 | } | ||||||||
| 1873 | |||||||||
| 1874 | |||||||||
| 1875 | void LazyValueInfo::printLVI(Function &F, DominatorTree &DTree, raw_ostream &OS) { | ||||||||
| 1876 | if (PImpl) { | ||||||||
| 1877 | getImpl(PImpl, AC, F.getParent()).printLVI(F, DTree, OS); | ||||||||
| 1878 | } | ||||||||
| 1879 | } | ||||||||
| 1880 | |||||||||
| 1881 | // Print the LVI for the function arguments at the start of each basic block. | ||||||||
| 1882 | void LazyValueInfoAnnotatedWriter::emitBasicBlockStartAnnot( | ||||||||
| 1883 | const BasicBlock *BB, formatted_raw_ostream &OS) { | ||||||||
| 1884 | // Find if there are latticevalues defined for arguments of the function. | ||||||||
| 1885 | auto *F = BB->getParent(); | ||||||||
| 1886 | for (auto &Arg : F->args()) { | ||||||||
| 1887 | ValueLatticeElement Result = LVIImpl->getValueInBlock( | ||||||||
| 1888 | const_cast<Argument *>(&Arg), const_cast<BasicBlock *>(BB)); | ||||||||
| 1889 | if (Result.isUnknown()) | ||||||||
| 1890 | continue; | ||||||||
| 1891 | OS << "; LatticeVal for: '" << Arg << "' is: " << Result << "\n"; | ||||||||
| 1892 | } | ||||||||
| 1893 | } | ||||||||
| 1894 | |||||||||
| 1895 | // This function prints the LVI analysis for the instruction I at the beginning | ||||||||
| 1896 | // of various basic blocks. It relies on calculated values that are stored in | ||||||||
| 1897 | // the LazyValueInfoCache, and in the absence of cached values, recalculate the | ||||||||
| 1898 | // LazyValueInfo for `I`, and print that info. | ||||||||
| 1899 | void LazyValueInfoAnnotatedWriter::emitInstructionAnnot( | ||||||||
| 1900 | const Instruction *I, formatted_raw_ostream &OS) { | ||||||||
| 1901 | |||||||||
| 1902 | auto *ParentBB = I->getParent(); | ||||||||
| 1903 | SmallPtrSet<const BasicBlock*, 16> BlocksContainingLVI; | ||||||||
| 1904 | // We can generate (solve) LVI values only for blocks that are dominated by | ||||||||
| 1905 | // the I's parent. However, to avoid generating LVI for all dominating blocks, | ||||||||
| 1906 | // that contain redundant/uninteresting information, we print LVI for | ||||||||
| 1907 | // blocks that may use this LVI information (such as immediate successor | ||||||||
| 1908 | // blocks, and blocks that contain uses of `I`). | ||||||||
| 1909 | auto printResult = [&](const BasicBlock *BB) { | ||||||||
| 1910 | if (!BlocksContainingLVI.insert(BB).second) | ||||||||
| 1911 | return; | ||||||||
| 1912 | ValueLatticeElement Result = LVIImpl->getValueInBlock( | ||||||||
| 1913 | const_cast<Instruction *>(I), const_cast<BasicBlock *>(BB)); | ||||||||
| 1914 | OS << "; LatticeVal for: '" << *I << "' in BB: '"; | ||||||||
| 1915 | BB->printAsOperand(OS, false); | ||||||||
| 1916 | OS << "' is: " << Result << "\n"; | ||||||||
| 1917 | }; | ||||||||
| 1918 | |||||||||
| 1919 | printResult(ParentBB); | ||||||||
| 1920 | // Print the LVI analysis results for the immediate successor blocks, that | ||||||||
| 1921 | // are dominated by `ParentBB`. | ||||||||
| 1922 | for (auto *BBSucc : successors(ParentBB)) | ||||||||
| 1923 | if (DT.dominates(ParentBB, BBSucc)) | ||||||||
| 1924 | printResult(BBSucc); | ||||||||
| 1925 | |||||||||
| 1926 | // Print LVI in blocks where `I` is used. | ||||||||
| 1927 | for (auto *U : I->users()) | ||||||||
| 1928 | if (auto *UseI = dyn_cast<Instruction>(U)) | ||||||||
| 1929 | if (!isa<PHINode>(UseI) || DT.dominates(ParentBB, UseI->getParent())) | ||||||||
| 1930 | printResult(UseI->getParent()); | ||||||||
| 1931 | |||||||||
| 1932 | } | ||||||||
| 1933 | |||||||||
| 1934 | namespace { | ||||||||
| 1935 | // Printer class for LazyValueInfo results. | ||||||||
| 1936 | class LazyValueInfoPrinter : public FunctionPass { | ||||||||
| 1937 | public: | ||||||||
| 1938 | static char ID; // Pass identification, replacement for typeid | ||||||||
| 1939 | LazyValueInfoPrinter() : FunctionPass(ID) { | ||||||||
| 1940 | initializeLazyValueInfoPrinterPass(*PassRegistry::getPassRegistry()); | ||||||||
| 1941 | } | ||||||||
| 1942 | |||||||||
| 1943 | void getAnalysisUsage(AnalysisUsage &AU) const override { | ||||||||
| 1944 | AU.setPreservesAll(); | ||||||||
| 1945 | AU.addRequired<LazyValueInfoWrapperPass>(); | ||||||||
| 1946 | AU.addRequired<DominatorTreeWrapperPass>(); | ||||||||
| 1947 | } | ||||||||
| 1948 | |||||||||
| 1949 | // Get the mandatory dominator tree analysis and pass this in to the | ||||||||
| 1950 | // LVIPrinter. We cannot rely on the LVI's DT, since it's optional. | ||||||||
| 1951 | bool runOnFunction(Function &F) override { | ||||||||
| 1952 | dbgs() << "LVI for function '" << F.getName() << "':\n"; | ||||||||
| 1953 | auto &LVI = getAnalysis<LazyValueInfoWrapperPass>().getLVI(); | ||||||||
| 1954 | auto &DTree = getAnalysis<DominatorTreeWrapperPass>().getDomTree(); | ||||||||
| 1955 | LVI.printLVI(F, DTree, dbgs()); | ||||||||
| 1956 | return false; | ||||||||
| 1957 | } | ||||||||
| 1958 | }; | ||||||||
| 1959 | } | ||||||||
| 1960 | |||||||||
| 1961 | char LazyValueInfoPrinter::ID = 0; | ||||||||
| 1962 | INITIALIZE_PASS_BEGIN(LazyValueInfoPrinter, "print-lazy-value-info",static void *initializeLazyValueInfoPrinterPassOnce(PassRegistry &Registry) { | ||||||||
| 1963 | "Lazy Value Info Printer Pass", false, false)static void *initializeLazyValueInfoPrinterPassOnce(PassRegistry &Registry) { | ||||||||
| 1964 | INITIALIZE_PASS_DEPENDENCY(LazyValueInfoWrapperPass)initializeLazyValueInfoWrapperPassPass(Registry); | ||||||||
| 1965 | INITIALIZE_PASS_END(LazyValueInfoPrinter, "print-lazy-value-info",PassInfo *PI = new PassInfo( "Lazy Value Info Printer Pass", "print-lazy-value-info" , &LazyValueInfoPrinter::ID, PassInfo::NormalCtor_t(callDefaultCtor <LazyValueInfoPrinter>), false, false); Registry.registerPass (*PI, true); return PI; } static llvm::once_flag InitializeLazyValueInfoPrinterPassFlag ; void llvm::initializeLazyValueInfoPrinterPass(PassRegistry & Registry) { llvm::call_once(InitializeLazyValueInfoPrinterPassFlag , initializeLazyValueInfoPrinterPassOnce, std::ref(Registry)) ; } | ||||||||
| 1966 | "Lazy Value Info Printer Pass", false, false)PassInfo *PI = new PassInfo( "Lazy Value Info Printer Pass", "print-lazy-value-info" , &LazyValueInfoPrinter::ID, PassInfo::NormalCtor_t(callDefaultCtor <LazyValueInfoPrinter>), false, false); Registry.registerPass (*PI, true); return PI; } static llvm::once_flag InitializeLazyValueInfoPrinterPassFlag ; void llvm::initializeLazyValueInfoPrinterPass(PassRegistry & Registry) { llvm::call_once(InitializeLazyValueInfoPrinterPassFlag , initializeLazyValueInfoPrinterPassOnce, std::ref(Registry)) ; } |
| 1 | //===- llvm/Type.h - Classes for handling data types ------------*- 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 contains the declaration of the Type class. For more "Type" |
| 10 | // stuff, look in DerivedTypes.h. |
| 11 | // |
| 12 | //===----------------------------------------------------------------------===// |
| 13 | |
| 14 | #ifndef LLVM_IR_TYPE_H |
| 15 | #define LLVM_IR_TYPE_H |
| 16 | |
| 17 | #include "llvm/ADT/APFloat.h" |
| 18 | #include "llvm/ADT/ArrayRef.h" |
| 19 | #include "llvm/ADT/SmallPtrSet.h" |
| 20 | #include "llvm/Support/CBindingWrapping.h" |
| 21 | #include "llvm/Support/Casting.h" |
| 22 | #include "llvm/Support/Compiler.h" |
| 23 | #include "llvm/Support/ErrorHandling.h" |
| 24 | #include "llvm/Support/TypeSize.h" |
| 25 | #include <cassert> |
| 26 | #include <cstdint> |
| 27 | #include <iterator> |
| 28 | |
| 29 | namespace llvm { |
| 30 | |
| 31 | class IntegerType; |
| 32 | class LLVMContext; |
| 33 | class PointerType; |
| 34 | class raw_ostream; |
| 35 | class StringRef; |
| 36 | |
| 37 | /// The instances of the Type class are immutable: once they are created, |
| 38 | /// they are never changed. Also note that only one instance of a particular |
| 39 | /// type is ever created. Thus seeing if two types are equal is a matter of |
| 40 | /// doing a trivial pointer comparison. To enforce that no two equal instances |
| 41 | /// are created, Type instances can only be created via static factory methods |
| 42 | /// in class Type and in derived classes. Once allocated, Types are never |
| 43 | /// free'd. |
| 44 | /// |
| 45 | class Type { |
| 46 | public: |
| 47 | //===--------------------------------------------------------------------===// |
| 48 | /// Definitions of all of the base types for the Type system. Based on this |
| 49 | /// value, you can cast to a class defined in DerivedTypes.h. |
| 50 | /// Note: If you add an element to this, you need to add an element to the |
| 51 | /// Type::getPrimitiveType function, or else things will break! |
| 52 | /// Also update LLVMTypeKind and LLVMGetTypeKind () in the C binding. |
| 53 | /// |
| 54 | enum TypeID { |
| 55 | // PrimitiveTypes |
| 56 | HalfTyID = 0, ///< 16-bit floating point type |
| 57 | BFloatTyID, ///< 16-bit floating point type (7-bit significand) |
| 58 | FloatTyID, ///< 32-bit floating point type |
| 59 | DoubleTyID, ///< 64-bit floating point type |
| 60 | X86_FP80TyID, ///< 80-bit floating point type (X87) |
| 61 | FP128TyID, ///< 128-bit floating point type (112-bit significand) |
| 62 | PPC_FP128TyID, ///< 128-bit floating point type (two 64-bits, PowerPC) |
| 63 | VoidTyID, ///< type with no size |
| 64 | LabelTyID, ///< Labels |
| 65 | MetadataTyID, ///< Metadata |
| 66 | X86_MMXTyID, ///< MMX vectors (64 bits, X86 specific) |
| 67 | X86_AMXTyID, ///< AMX vectors (8192 bits, X86 specific) |
| 68 | TokenTyID, ///< Tokens |
| 69 | |
| 70 | // Derived types... see DerivedTypes.h file. |
| 71 | IntegerTyID, ///< Arbitrary bit width integers |
| 72 | FunctionTyID, ///< Functions |
| 73 | PointerTyID, ///< Pointers |
| 74 | StructTyID, ///< Structures |
| 75 | ArrayTyID, ///< Arrays |
| 76 | FixedVectorTyID, ///< Fixed width SIMD vector type |
| 77 | ScalableVectorTyID ///< Scalable SIMD vector type |
| 78 | }; |
| 79 | |
| 80 | private: |
| 81 | /// This refers to the LLVMContext in which this type was uniqued. |
| 82 | LLVMContext &Context; |
| 83 | |
| 84 | TypeID ID : 8; // The current base type of this type. |
| 85 | unsigned SubclassData : 24; // Space for subclasses to store data. |
| 86 | // Note that this should be synchronized with |
| 87 | // MAX_INT_BITS value in IntegerType class. |
| 88 | |
| 89 | protected: |
| 90 | friend class LLVMContextImpl; |
| 91 | |
| 92 | explicit Type(LLVMContext &C, TypeID tid) |
| 93 | : Context(C), ID(tid), SubclassData(0) {} |
| 94 | ~Type() = default; |
| 95 | |
| 96 | unsigned getSubclassData() const { return SubclassData; } |
| 97 | |
| 98 | void setSubclassData(unsigned val) { |
| 99 | SubclassData = val; |
| 100 | // Ensure we don't have any accidental truncation. |
| 101 | assert(getSubclassData() == val && "Subclass data too large for field")((void)0); |
| 102 | } |
| 103 | |
| 104 | /// Keeps track of how many Type*'s there are in the ContainedTys list. |
| 105 | unsigned NumContainedTys = 0; |
| 106 | |
| 107 | /// A pointer to the array of Types contained by this Type. For example, this |
| 108 | /// includes the arguments of a function type, the elements of a structure, |
| 109 | /// the pointee of a pointer, the element type of an array, etc. This pointer |
| 110 | /// may be 0 for types that don't contain other types (Integer, Double, |
| 111 | /// Float). |
| 112 | Type * const *ContainedTys = nullptr; |
| 113 | |
| 114 | public: |
| 115 | /// Print the current type. |
| 116 | /// Omit the type details if \p NoDetails == true. |
| 117 | /// E.g., let %st = type { i32, i16 } |
| 118 | /// When \p NoDetails is true, we only print %st. |
| 119 | /// Put differently, \p NoDetails prints the type as if |
| 120 | /// inlined with the operands when printing an instruction. |
| 121 | void print(raw_ostream &O, bool IsForDebug = false, |
| 122 | bool NoDetails = false) const; |
| 123 | |
| 124 | void dump() const; |
| 125 | |
| 126 | /// Return the LLVMContext in which this type was uniqued. |
| 127 | LLVMContext &getContext() const { return Context; } |
| 128 | |
| 129 | //===--------------------------------------------------------------------===// |
| 130 | // Accessors for working with types. |
| 131 | // |
| 132 | |
| 133 | /// Return the type id for the type. This will return one of the TypeID enum |
| 134 | /// elements defined above. |
| 135 | TypeID getTypeID() const { return ID; } |
| 136 | |
| 137 | /// Return true if this is 'void'. |
| 138 | bool isVoidTy() const { return getTypeID() == VoidTyID; } |
| 139 | |
| 140 | /// Return true if this is 'half', a 16-bit IEEE fp type. |
| 141 | bool isHalfTy() const { return getTypeID() == HalfTyID; } |
| 142 | |
| 143 | /// Return true if this is 'bfloat', a 16-bit bfloat type. |
| 144 | bool isBFloatTy() const { return getTypeID() == BFloatTyID; } |
| 145 | |
| 146 | /// Return true if this is 'float', a 32-bit IEEE fp type. |
| 147 | bool isFloatTy() const { return getTypeID() == FloatTyID; } |
| 148 | |
| 149 | /// Return true if this is 'double', a 64-bit IEEE fp type. |
| 150 | bool isDoubleTy() const { return getTypeID() == DoubleTyID; } |
| 151 | |
| 152 | /// Return true if this is x86 long double. |
| 153 | bool isX86_FP80Ty() const { return getTypeID() == X86_FP80TyID; } |
| 154 | |
| 155 | /// Return true if this is 'fp128'. |
| 156 | bool isFP128Ty() const { return getTypeID() == FP128TyID; } |
| 157 | |
| 158 | /// Return true if this is powerpc long double. |
| 159 | bool isPPC_FP128Ty() const { return getTypeID() == PPC_FP128TyID; } |
| 160 | |
| 161 | /// Return true if this is one of the six floating-point types |
| 162 | bool isFloatingPointTy() const { |
| 163 | return getTypeID() == HalfTyID || getTypeID() == BFloatTyID || |
| 164 | getTypeID() == FloatTyID || getTypeID() == DoubleTyID || |
| 165 | getTypeID() == X86_FP80TyID || getTypeID() == FP128TyID || |
| 166 | getTypeID() == PPC_FP128TyID; |
| 167 | } |
| 168 | |
| 169 | const fltSemantics &getFltSemantics() const { |
| 170 | switch (getTypeID()) { |
| 171 | case HalfTyID: return APFloat::IEEEhalf(); |
| 172 | case BFloatTyID: return APFloat::BFloat(); |
| 173 | case FloatTyID: return APFloat::IEEEsingle(); |
| 174 | case DoubleTyID: return APFloat::IEEEdouble(); |
| 175 | case X86_FP80TyID: return APFloat::x87DoubleExtended(); |
| 176 | case FP128TyID: return APFloat::IEEEquad(); |
| 177 | case PPC_FP128TyID: return APFloat::PPCDoubleDouble(); |
| 178 | default: llvm_unreachable("Invalid floating type")__builtin_unreachable(); |
| 179 | } |
| 180 | } |
| 181 | |
| 182 | /// Return true if this is X86 MMX. |
| 183 | bool isX86_MMXTy() const { return getTypeID() == X86_MMXTyID; } |
| 184 | |
| 185 | /// Return true if this is X86 AMX. |
| 186 | bool isX86_AMXTy() const { return getTypeID() == X86_AMXTyID; } |
| 187 | |
| 188 | /// Return true if this is a FP type or a vector of FP. |
| 189 | bool isFPOrFPVectorTy() const { return getScalarType()->isFloatingPointTy(); } |
| 190 | |
| 191 | /// Return true if this is 'label'. |
| 192 | bool isLabelTy() const { return getTypeID() == LabelTyID; } |
| 193 | |
| 194 | /// Return true if this is 'metadata'. |
| 195 | bool isMetadataTy() const { return getTypeID() == MetadataTyID; } |
| 196 | |
| 197 | /// Return true if this is 'token'. |
| 198 | bool isTokenTy() const { return getTypeID() == TokenTyID; } |
| 199 | |
| 200 | /// True if this is an instance of IntegerType. |
| 201 | bool isIntegerTy() const { return getTypeID() == IntegerTyID; } |
| 202 | |
| 203 | /// Return true if this is an IntegerType of the given width. |
| 204 | bool isIntegerTy(unsigned Bitwidth) const; |
| 205 | |
| 206 | /// Return true if this is an integer type or a vector of integer types. |
| 207 | bool isIntOrIntVectorTy() const { return getScalarType()->isIntegerTy(); } |
| 208 | |
| 209 | /// Return true if this is an integer type or a vector of integer types of |
| 210 | /// the given width. |
| 211 | bool isIntOrIntVectorTy(unsigned BitWidth) const { |
| 212 | return getScalarType()->isIntegerTy(BitWidth); |
| 213 | } |
| 214 | |
| 215 | /// Return true if this is an integer type or a pointer type. |
| 216 | bool isIntOrPtrTy() const { return isIntegerTy() || isPointerTy(); } |
| 217 | |
| 218 | /// True if this is an instance of FunctionType. |
| 219 | bool isFunctionTy() const { return getTypeID() == FunctionTyID; } |
| 220 | |
| 221 | /// True if this is an instance of StructType. |
| 222 | bool isStructTy() const { return getTypeID() == StructTyID; } |
| 223 | |
| 224 | /// True if this is an instance of ArrayType. |
| 225 | bool isArrayTy() const { return getTypeID() == ArrayTyID; } |
| 226 | |
| 227 | /// True if this is an instance of PointerType. |
| 228 | bool isPointerTy() const { return getTypeID() == PointerTyID; } |
| 229 | |
| 230 | /// True if this is an instance of an opaque PointerType. |
| 231 | bool isOpaquePointerTy() const; |
| 232 | |
| 233 | /// Return true if this is a pointer type or a vector of pointer types. |
| 234 | bool isPtrOrPtrVectorTy() const { return getScalarType()->isPointerTy(); } |
| 235 | |
| 236 | /// True if this is an instance of VectorType. |
| 237 | inline bool isVectorTy() const { |
| 238 | return getTypeID() == ScalableVectorTyID || getTypeID() == FixedVectorTyID; |
| 239 | } |
| 240 | |
| 241 | /// Return true if this type could be converted with a lossless BitCast to |
| 242 | /// type 'Ty'. For example, i8* to i32*. BitCasts are valid for types of the |
| 243 | /// same size only where no re-interpretation of the bits is done. |
| 244 | /// Determine if this type could be losslessly bitcast to Ty |
| 245 | bool canLosslesslyBitCastTo(Type *Ty) const; |
| 246 | |
| 247 | /// Return true if this type is empty, that is, it has no elements or all of |
| 248 | /// its elements are empty. |
| 249 | bool isEmptyTy() const; |
| 250 | |
| 251 | /// Return true if the type is "first class", meaning it is a valid type for a |
| 252 | /// Value. |
| 253 | bool isFirstClassType() const { |
| 254 | return getTypeID() != FunctionTyID && getTypeID() != VoidTyID; |
| 255 | } |
| 256 | |
| 257 | /// Return true if the type is a valid type for a register in codegen. This |
| 258 | /// includes all first-class types except struct and array types. |
| 259 | bool isSingleValueType() const { |
| 260 | return isFloatingPointTy() || isX86_MMXTy() || isIntegerTy() || |
| 261 | isPointerTy() || isVectorTy() || isX86_AMXTy(); |
| 262 | } |
| 263 | |
| 264 | /// Return true if the type is an aggregate type. This means it is valid as |
| 265 | /// the first operand of an insertvalue or extractvalue instruction. This |
| 266 | /// includes struct and array types, but does not include vector types. |
| 267 | bool isAggregateType() const { |
| 268 | return getTypeID() == StructTyID || getTypeID() == ArrayTyID; |
| 269 | } |
| 270 | |
| 271 | /// Return true if it makes sense to take the size of this type. To get the |
| 272 | /// actual size for a particular target, it is reasonable to use the |
| 273 | /// DataLayout subsystem to do this. |
| 274 | bool isSized(SmallPtrSetImpl<Type*> *Visited = nullptr) const { |
| 275 | // If it's a primitive, it is always sized. |
| 276 | if (getTypeID() == IntegerTyID || isFloatingPointTy() || |
| 277 | getTypeID() == PointerTyID || getTypeID() == X86_MMXTyID || |
| 278 | getTypeID() == X86_AMXTyID) |
| 279 | return true; |
| 280 | // If it is not something that can have a size (e.g. a function or label), |
| 281 | // it doesn't have a size. |
| 282 | if (getTypeID() != StructTyID && getTypeID() != ArrayTyID && !isVectorTy()) |
| 283 | return false; |
| 284 | // Otherwise we have to try harder to decide. |
| 285 | return isSizedDerivedType(Visited); |
| 286 | } |
| 287 | |
| 288 | /// Return the basic size of this type if it is a primitive type. These are |
| 289 | /// fixed by LLVM and are not target-dependent. |
| 290 | /// This will return zero if the type does not have a size or is not a |
| 291 | /// primitive type. |
| 292 | /// |
| 293 | /// If this is a scalable vector type, the scalable property will be set and |
| 294 | /// the runtime size will be a positive integer multiple of the base size. |
| 295 | /// |
| 296 | /// Note that this may not reflect the size of memory allocated for an |
| 297 | /// instance of the type or the number of bytes that are written when an |
| 298 | /// instance of the type is stored to memory. The DataLayout class provides |
| 299 | /// additional query functions to provide this information. |
| 300 | /// |
| 301 | TypeSize getPrimitiveSizeInBits() const LLVM_READONLY__attribute__((__pure__)); |
| 302 | |
| 303 | /// If this is a vector type, return the getPrimitiveSizeInBits value for the |
| 304 | /// element type. Otherwise return the getPrimitiveSizeInBits value for this |
| 305 | /// type. |
| 306 | unsigned getScalarSizeInBits() const LLVM_READONLY__attribute__((__pure__)); |
| 307 | |
| 308 | /// Return the width of the mantissa of this type. This is only valid on |
| 309 | /// floating-point types. If the FP type does not have a stable mantissa (e.g. |
| 310 | /// ppc long double), this method returns -1. |
| 311 | int getFPMantissaWidth() const; |
| 312 | |
| 313 | /// Return whether the type is IEEE compatible, as defined by the eponymous |
| 314 | /// method in APFloat. |
| 315 | bool isIEEE() const { return APFloat::getZero(getFltSemantics()).isIEEE(); } |
| 316 | |
| 317 | /// If this is a vector type, return the element type, otherwise return |
| 318 | /// 'this'. |
| 319 | inline Type *getScalarType() const { |
| 320 | if (isVectorTy()) |
| 321 | return getContainedType(0); |
| 322 | return const_cast<Type *>(this); |
| 323 | } |
| 324 | |
| 325 | //===--------------------------------------------------------------------===// |
| 326 | // Type Iteration support. |
| 327 | // |
| 328 | using subtype_iterator = Type * const *; |
| 329 | |
| 330 | subtype_iterator subtype_begin() const { return ContainedTys; } |
| 331 | subtype_iterator subtype_end() const { return &ContainedTys[NumContainedTys];} |
| 332 | ArrayRef<Type*> subtypes() const { |
| 333 | return makeArrayRef(subtype_begin(), subtype_end()); |
| 334 | } |
| 335 | |
| 336 | using subtype_reverse_iterator = std::reverse_iterator<subtype_iterator>; |
| 337 | |
| 338 | subtype_reverse_iterator subtype_rbegin() const { |
| 339 | return subtype_reverse_iterator(subtype_end()); |
| 340 | } |
| 341 | subtype_reverse_iterator subtype_rend() const { |
| 342 | return subtype_reverse_iterator(subtype_begin()); |
| 343 | } |
| 344 | |
| 345 | /// This method is used to implement the type iterator (defined at the end of |
| 346 | /// the file). For derived types, this returns the types 'contained' in the |
| 347 | /// derived type. |
| 348 | Type *getContainedType(unsigned i) const { |
| 349 | assert(i < NumContainedTys && "Index out of range!")((void)0); |
| 350 | return ContainedTys[i]; |
| 351 | } |
| 352 | |
| 353 | /// Return the number of types in the derived type. |
| 354 | unsigned getNumContainedTypes() const { return NumContainedTys; } |
| 355 | |
| 356 | //===--------------------------------------------------------------------===// |
| 357 | // Helper methods corresponding to subclass methods. This forces a cast to |
| 358 | // the specified subclass and calls its accessor. "getArrayNumElements" (for |
| 359 | // example) is shorthand for cast<ArrayType>(Ty)->getNumElements(). This is |
| 360 | // only intended to cover the core methods that are frequently used, helper |
| 361 | // methods should not be added here. |
| 362 | |
| 363 | inline unsigned getIntegerBitWidth() const; |
| 364 | |
| 365 | inline Type *getFunctionParamType(unsigned i) const; |
| 366 | inline unsigned getFunctionNumParams() const; |
| 367 | inline bool isFunctionVarArg() const; |
| 368 | |
| 369 | inline StringRef getStructName() const; |
| 370 | inline unsigned getStructNumElements() const; |
| 371 | inline Type *getStructElementType(unsigned N) const; |
| 372 | |
| 373 | inline uint64_t getArrayNumElements() const; |
| 374 | |
| 375 | Type *getArrayElementType() const { |
| 376 | assert(getTypeID() == ArrayTyID)((void)0); |
| 377 | return ContainedTys[0]; |
| 378 | } |
| 379 | |
| 380 | Type *getPointerElementType() const { |
| 381 | assert(getTypeID() == PointerTyID)((void)0); |
| 382 | return ContainedTys[0]; |
| 383 | } |
| 384 | |
| 385 | /// Given vector type, change the element type, |
| 386 | /// whilst keeping the old number of elements. |
| 387 | /// For non-vectors simply returns \p EltTy. |
| 388 | inline Type *getWithNewType(Type *EltTy) const; |
| 389 | |
| 390 | /// Given an integer or vector type, change the lane bitwidth to NewBitwidth, |
| 391 | /// whilst keeping the old number of lanes. |
| 392 | inline Type *getWithNewBitWidth(unsigned NewBitWidth) const; |
| 393 | |
| 394 | /// Given scalar/vector integer type, returns a type with elements twice as |
| 395 | /// wide as in the original type. For vectors, preserves element count. |
| 396 | inline Type *getExtendedType() const; |
| 397 | |
| 398 | /// Get the address space of this pointer or pointer vector type. |
| 399 | inline unsigned getPointerAddressSpace() const; |
| 400 | |
| 401 | //===--------------------------------------------------------------------===// |
| 402 | // Static members exported by the Type class itself. Useful for getting |
| 403 | // instances of Type. |
| 404 | // |
| 405 | |
| 406 | /// Return a type based on an identifier. |
| 407 | static Type *getPrimitiveType(LLVMContext &C, TypeID IDNumber); |
| 408 | |
| 409 | //===--------------------------------------------------------------------===// |
| 410 | // These are the builtin types that are always available. |
| 411 | // |
| 412 | static Type *getVoidTy(LLVMContext &C); |
| 413 | static Type *getLabelTy(LLVMContext &C); |
| 414 | static Type *getHalfTy(LLVMContext &C); |
| 415 | static Type *getBFloatTy(LLVMContext &C); |
| 416 | static Type *getFloatTy(LLVMContext &C); |
| 417 | static Type *getDoubleTy(LLVMContext &C); |
| 418 | static Type *getMetadataTy(LLVMContext &C); |
| 419 | static Type *getX86_FP80Ty(LLVMContext &C); |
| 420 | static Type *getFP128Ty(LLVMContext &C); |
| 421 | static Type *getPPC_FP128Ty(LLVMContext &C); |
| 422 | static Type *getX86_MMXTy(LLVMContext &C); |
| 423 | static Type *getX86_AMXTy(LLVMContext &C); |
| 424 | static Type *getTokenTy(LLVMContext &C); |
| 425 | static IntegerType *getIntNTy(LLVMContext &C, unsigned N); |
| 426 | static IntegerType *getInt1Ty(LLVMContext &C); |
| 427 | static IntegerType *getInt8Ty(LLVMContext &C); |
| 428 | static IntegerType *getInt16Ty(LLVMContext &C); |
| 429 | static IntegerType *getInt32Ty(LLVMContext &C); |
| 430 | static IntegerType *getInt64Ty(LLVMContext &C); |
| 431 | static IntegerType *getInt128Ty(LLVMContext &C); |
| 432 | template <typename ScalarTy> static Type *getScalarTy(LLVMContext &C) { |
| 433 | int noOfBits = sizeof(ScalarTy) * CHAR_BIT8; |
| 434 | if (std::is_integral<ScalarTy>::value) { |
| 435 | return (Type*) Type::getIntNTy(C, noOfBits); |
| 436 | } else if (std::is_floating_point<ScalarTy>::value) { |
| 437 | switch (noOfBits) { |
| 438 | case 32: |
| 439 | return Type::getFloatTy(C); |
| 440 | case 64: |
| 441 | return Type::getDoubleTy(C); |
| 442 | } |
| 443 | } |
| 444 | llvm_unreachable("Unsupported type in Type::getScalarTy")__builtin_unreachable(); |
| 445 | } |
| 446 | static Type *getFloatingPointTy(LLVMContext &C, const fltSemantics &S) { |
| 447 | Type *Ty; |
| 448 | if (&S == &APFloat::IEEEhalf()) |
| 449 | Ty = Type::getHalfTy(C); |
| 450 | else if (&S == &APFloat::BFloat()) |
| 451 | Ty = Type::getBFloatTy(C); |
| 452 | else if (&S == &APFloat::IEEEsingle()) |
| 453 | Ty = Type::getFloatTy(C); |
| 454 | else if (&S == &APFloat::IEEEdouble()) |
| 455 | Ty = Type::getDoubleTy(C); |
| 456 | else if (&S == &APFloat::x87DoubleExtended()) |
| 457 | Ty = Type::getX86_FP80Ty(C); |
| 458 | else if (&S == &APFloat::IEEEquad()) |
| 459 | Ty = Type::getFP128Ty(C); |
| 460 | else { |
| 461 | assert(&S == &APFloat::PPCDoubleDouble() && "Unknown FP format")((void)0); |
| 462 | Ty = Type::getPPC_FP128Ty(C); |
| 463 | } |
| 464 | return Ty; |
| 465 | } |
| 466 | |
| 467 | //===--------------------------------------------------------------------===// |
| 468 | // Convenience methods for getting pointer types with one of the above builtin |
| 469 | // types as pointee. |
| 470 | // |
| 471 | static PointerType *getHalfPtrTy(LLVMContext &C, unsigned AS = 0); |
| 472 | static PointerType *getBFloatPtrTy(LLVMContext &C, unsigned AS = 0); |
| 473 | static PointerType *getFloatPtrTy(LLVMContext &C, unsigned AS = 0); |
| 474 | static PointerType *getDoublePtrTy(LLVMContext &C, unsigned AS = 0); |
| 475 | static PointerType *getX86_FP80PtrTy(LLVMContext &C, unsigned AS = 0); |
| 476 | static PointerType *getFP128PtrTy(LLVMContext &C, unsigned AS = 0); |
| 477 | static PointerType *getPPC_FP128PtrTy(LLVMContext &C, unsigned AS = 0); |
| 478 | static PointerType *getX86_MMXPtrTy(LLVMContext &C, unsigned AS = 0); |
| 479 | static PointerType *getX86_AMXPtrTy(LLVMContext &C, unsigned AS = 0); |
| 480 | static PointerType *getIntNPtrTy(LLVMContext &C, unsigned N, unsigned AS = 0); |
| 481 | static PointerType *getInt1PtrTy(LLVMContext &C, unsigned AS = 0); |
| 482 | static PointerType *getInt8PtrTy(LLVMContext &C, unsigned AS = 0); |
| 483 | static PointerType *getInt16PtrTy(LLVMContext &C, unsigned AS = 0); |
| 484 | static PointerType *getInt32PtrTy(LLVMContext &C, unsigned AS = 0); |
| 485 | static PointerType *getInt64PtrTy(LLVMContext &C, unsigned AS = 0); |
| 486 | |
| 487 | /// Return a pointer to the current type. This is equivalent to |
| 488 | /// PointerType::get(Foo, AddrSpace). |
| 489 | /// TODO: Remove this after opaque pointer transition is complete. |
| 490 | PointerType *getPointerTo(unsigned AddrSpace = 0) const; |
| 491 | |
| 492 | private: |
| 493 | /// Derived types like structures and arrays are sized iff all of the members |
| 494 | /// of the type are sized as well. Since asking for their size is relatively |
| 495 | /// uncommon, move this operation out-of-line. |
| 496 | bool isSizedDerivedType(SmallPtrSetImpl<Type*> *Visited = nullptr) const; |
| 497 | }; |
| 498 | |
| 499 | // Printing of types. |
| 500 | inline raw_ostream &operator<<(raw_ostream &OS, const Type &T) { |
| 501 | T.print(OS); |
| 502 | return OS; |
| 503 | } |
| 504 | |
| 505 | // allow isa<PointerType>(x) to work without DerivedTypes.h included. |
| 506 | template <> struct isa_impl<PointerType, Type> { |
| 507 | static inline bool doit(const Type &Ty) { |
| 508 | return Ty.getTypeID() == Type::PointerTyID; |
| 509 | } |
| 510 | }; |
| 511 | |
| 512 | // Create wrappers for C Binding types (see CBindingWrapping.h). |
| 513 | DEFINE_ISA_CONVERSION_FUNCTIONS(Type, LLVMTypeRef)inline Type *unwrap(LLVMTypeRef P) { return reinterpret_cast< Type*>(P); } inline LLVMTypeRef wrap(const Type *P) { return reinterpret_cast<LLVMTypeRef>(const_cast<Type*>( P)); } template<typename T> inline T *unwrap(LLVMTypeRef P) { return cast<T>(unwrap(P)); } |
| 514 | |
| 515 | /* Specialized opaque type conversions. |
| 516 | */ |
| 517 | inline Type **unwrap(LLVMTypeRef* Tys) { |
| 518 | return reinterpret_cast<Type**>(Tys); |
| 519 | } |
| 520 | |
| 521 | inline LLVMTypeRef *wrap(Type **Tys) { |
| 522 | return reinterpret_cast<LLVMTypeRef*>(const_cast<Type**>(Tys)); |
| 523 | } |
| 524 | |
| 525 | } // end namespace llvm |
| 526 | |
| 527 | #endif // LLVM_IR_TYPE_H |
| 1 | //===- llvm/Instructions.h - Instruction subclass definitions ---*- 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 exposes the class definitions of all of the subclasses of the |
| 10 | // Instruction class. This is meant to be an easy way to get access to all |
| 11 | // instruction subclasses. |
| 12 | // |
| 13 | //===----------------------------------------------------------------------===// |
| 14 | |
| 15 | #ifndef LLVM_IR_INSTRUCTIONS_H |
| 16 | #define LLVM_IR_INSTRUCTIONS_H |
| 17 | |
| 18 | #include "llvm/ADT/ArrayRef.h" |
| 19 | #include "llvm/ADT/Bitfields.h" |
| 20 | #include "llvm/ADT/MapVector.h" |
| 21 | #include "llvm/ADT/None.h" |
| 22 | #include "llvm/ADT/STLExtras.h" |
| 23 | #include "llvm/ADT/SmallVector.h" |
| 24 | #include "llvm/ADT/StringRef.h" |
| 25 | #include "llvm/ADT/Twine.h" |
| 26 | #include "llvm/ADT/iterator.h" |
| 27 | #include "llvm/ADT/iterator_range.h" |
| 28 | #include "llvm/IR/Attributes.h" |
| 29 | #include "llvm/IR/BasicBlock.h" |
| 30 | #include "llvm/IR/CallingConv.h" |
| 31 | #include "llvm/IR/CFG.h" |
| 32 | #include "llvm/IR/Constant.h" |
| 33 | #include "llvm/IR/DerivedTypes.h" |
| 34 | #include "llvm/IR/Function.h" |
| 35 | #include "llvm/IR/InstrTypes.h" |
| 36 | #include "llvm/IR/Instruction.h" |
| 37 | #include "llvm/IR/OperandTraits.h" |
| 38 | #include "llvm/IR/Type.h" |
| 39 | #include "llvm/IR/Use.h" |
| 40 | #include "llvm/IR/User.h" |
| 41 | #include "llvm/IR/Value.h" |
| 42 | #include "llvm/Support/AtomicOrdering.h" |
| 43 | #include "llvm/Support/Casting.h" |
| 44 | #include "llvm/Support/ErrorHandling.h" |
| 45 | #include <cassert> |
| 46 | #include <cstddef> |
| 47 | #include <cstdint> |
| 48 | #include <iterator> |
| 49 | |
| 50 | namespace llvm { |
| 51 | |
| 52 | class APInt; |
| 53 | class ConstantInt; |
| 54 | class DataLayout; |
| 55 | class LLVMContext; |
| 56 | |
| 57 | //===----------------------------------------------------------------------===// |
| 58 | // AllocaInst Class |
| 59 | //===----------------------------------------------------------------------===// |
| 60 | |
| 61 | /// an instruction to allocate memory on the stack |
| 62 | class AllocaInst : public UnaryInstruction { |
| 63 | Type *AllocatedType; |
| 64 | |
| 65 | using AlignmentField = AlignmentBitfieldElementT<0>; |
| 66 | using UsedWithInAllocaField = BoolBitfieldElementT<AlignmentField::NextBit>; |
| 67 | using SwiftErrorField = BoolBitfieldElementT<UsedWithInAllocaField::NextBit>; |
| 68 | static_assert(Bitfield::areContiguous<AlignmentField, UsedWithInAllocaField, |
| 69 | SwiftErrorField>(), |
| 70 | "Bitfields must be contiguous"); |
| 71 | |
| 72 | protected: |
| 73 | // Note: Instruction needs to be a friend here to call cloneImpl. |
| 74 | friend class Instruction; |
| 75 | |
| 76 | AllocaInst *cloneImpl() const; |
| 77 | |
| 78 | public: |
| 79 | explicit AllocaInst(Type *Ty, unsigned AddrSpace, Value *ArraySize, |
| 80 | const Twine &Name, Instruction *InsertBefore); |
| 81 | AllocaInst(Type *Ty, unsigned AddrSpace, Value *ArraySize, |
| 82 | const Twine &Name, BasicBlock *InsertAtEnd); |
| 83 | |
| 84 | AllocaInst(Type *Ty, unsigned AddrSpace, const Twine &Name, |
| 85 | Instruction *InsertBefore); |
| 86 | AllocaInst(Type *Ty, unsigned AddrSpace, |
| 87 | const Twine &Name, BasicBlock *InsertAtEnd); |
| 88 | |
| 89 | AllocaInst(Type *Ty, unsigned AddrSpace, Value *ArraySize, Align Align, |
| 90 | const Twine &Name = "", Instruction *InsertBefore = nullptr); |
| 91 | AllocaInst(Type *Ty, unsigned AddrSpace, Value *ArraySize, Align Align, |
| 92 | const Twine &Name, BasicBlock *InsertAtEnd); |
| 93 | |
| 94 | /// Return true if there is an allocation size parameter to the allocation |
| 95 | /// instruction that is not 1. |
| 96 | bool isArrayAllocation() const; |
| 97 | |
| 98 | /// Get the number of elements allocated. For a simple allocation of a single |
| 99 | /// element, this will return a constant 1 value. |
| 100 | const Value *getArraySize() const { return getOperand(0); } |
| 101 | Value *getArraySize() { return getOperand(0); } |
| 102 | |
| 103 | /// Overload to return most specific pointer type. |
| 104 | PointerType *getType() const { |
| 105 | return cast<PointerType>(Instruction::getType()); |
| 106 | } |
| 107 | |
| 108 | /// Get allocation size in bits. Returns None if size can't be determined, |
| 109 | /// e.g. in case of a VLA. |
| 110 | Optional<TypeSize> getAllocationSizeInBits(const DataLayout &DL) const; |
| 111 | |
| 112 | /// Return the type that is being allocated by the instruction. |
| 113 | Type *getAllocatedType() const { return AllocatedType; } |
| 114 | /// for use only in special circumstances that need to generically |
| 115 | /// transform a whole instruction (eg: IR linking and vectorization). |
| 116 | void setAllocatedType(Type *Ty) { AllocatedType = Ty; } |
| 117 | |
| 118 | /// Return the alignment of the memory that is being allocated by the |
| 119 | /// instruction. |
| 120 | Align getAlign() const { |
| 121 | return Align(1ULL << getSubclassData<AlignmentField>()); |
| 122 | } |
| 123 | |
| 124 | void setAlignment(Align Align) { |
| 125 | setSubclassData<AlignmentField>(Log2(Align)); |
| 126 | } |
| 127 | |
| 128 | // FIXME: Remove this one transition to Align is over. |
| 129 | unsigned getAlignment() const { return getAlign().value(); } |
| 130 | |
| 131 | /// Return true if this alloca is in the entry block of the function and is a |
| 132 | /// constant size. If so, the code generator will fold it into the |
| 133 | /// prolog/epilog code, so it is basically free. |
| 134 | bool isStaticAlloca() const; |
| 135 | |
| 136 | /// Return true if this alloca is used as an inalloca argument to a call. Such |
| 137 | /// allocas are never considered static even if they are in the entry block. |
| 138 | bool isUsedWithInAlloca() const { |
| 139 | return getSubclassData<UsedWithInAllocaField>(); |
| 140 | } |
| 141 | |
| 142 | /// Specify whether this alloca is used to represent the arguments to a call. |
| 143 | void setUsedWithInAlloca(bool V) { |
| 144 | setSubclassData<UsedWithInAllocaField>(V); |
| 145 | } |
| 146 | |
| 147 | /// Return true if this alloca is used as a swifterror argument to a call. |
| 148 | bool isSwiftError() const { return getSubclassData<SwiftErrorField>(); } |
| 149 | /// Specify whether this alloca is used to represent a swifterror. |
| 150 | void setSwiftError(bool V) { setSubclassData<SwiftErrorField>(V); } |
| 151 | |
| 152 | // Methods for support type inquiry through isa, cast, and dyn_cast: |
| 153 | static bool classof(const Instruction *I) { |
| 154 | return (I->getOpcode() == Instruction::Alloca); |
| 155 | } |
| 156 | static bool classof(const Value *V) { |
| 157 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
| 158 | } |
| 159 | |
| 160 | private: |
| 161 | // Shadow Instruction::setInstructionSubclassData with a private forwarding |
| 162 | // method so that subclasses cannot accidentally use it. |
| 163 | template <typename Bitfield> |
| 164 | void setSubclassData(typename Bitfield::Type Value) { |
| 165 | Instruction::setSubclassData<Bitfield>(Value); |
| 166 | } |
| 167 | }; |
| 168 | |
| 169 | //===----------------------------------------------------------------------===// |
| 170 | // LoadInst Class |
| 171 | //===----------------------------------------------------------------------===// |
| 172 | |
| 173 | /// An instruction for reading from memory. This uses the SubclassData field in |
| 174 | /// Value to store whether or not the load is volatile. |
| 175 | class LoadInst : public UnaryInstruction { |
| 176 | using VolatileField = BoolBitfieldElementT<0>; |
| 177 | using AlignmentField = AlignmentBitfieldElementT<VolatileField::NextBit>; |
| 178 | using OrderingField = AtomicOrderingBitfieldElementT<AlignmentField::NextBit>; |
| 179 | static_assert( |
| 180 | Bitfield::areContiguous<VolatileField, AlignmentField, OrderingField>(), |
| 181 | "Bitfields must be contiguous"); |
| 182 | |
| 183 | void AssertOK(); |
| 184 | |
| 185 | protected: |
| 186 | // Note: Instruction needs to be a friend here to call cloneImpl. |
| 187 | friend class Instruction; |
| 188 | |
| 189 | LoadInst *cloneImpl() const; |
| 190 | |
| 191 | public: |
| 192 | LoadInst(Type *Ty, Value *Ptr, const Twine &NameStr, |
| 193 | Instruction *InsertBefore); |
| 194 | LoadInst(Type *Ty, Value *Ptr, const Twine &NameStr, BasicBlock *InsertAtEnd); |
| 195 | LoadInst(Type *Ty, Value *Ptr, const Twine &NameStr, bool isVolatile, |
| 196 | Instruction *InsertBefore); |
| 197 | LoadInst(Type *Ty, Value *Ptr, const Twine &NameStr, bool isVolatile, |
| 198 | BasicBlock *InsertAtEnd); |
| 199 | LoadInst(Type *Ty, Value *Ptr, const Twine &NameStr, bool isVolatile, |
| 200 | Align Align, Instruction *InsertBefore = nullptr); |
| 201 | LoadInst(Type *Ty, Value *Ptr, const Twine &NameStr, bool isVolatile, |
| 202 | Align Align, BasicBlock *InsertAtEnd); |
| 203 | LoadInst(Type *Ty, Value *Ptr, const Twine &NameStr, bool isVolatile, |
| 204 | Align Align, AtomicOrdering Order, |
| 205 | SyncScope::ID SSID = SyncScope::System, |
| 206 | Instruction *InsertBefore = nullptr); |
| 207 | LoadInst(Type *Ty, Value *Ptr, const Twine &NameStr, bool isVolatile, |
| 208 | Align Align, AtomicOrdering Order, SyncScope::ID SSID, |
| 209 | BasicBlock *InsertAtEnd); |
| 210 | |
| 211 | /// Return true if this is a load from a volatile memory location. |
| 212 | bool isVolatile() const { return getSubclassData<VolatileField>(); } |
| 213 | |
| 214 | /// Specify whether this is a volatile load or not. |
| 215 | void setVolatile(bool V) { setSubclassData<VolatileField>(V); } |
| 216 | |
| 217 | /// Return the alignment of the access that is being performed. |
| 218 | /// FIXME: Remove this function once transition to Align is over. |
| 219 | /// Use getAlign() instead. |
| 220 | unsigned getAlignment() const { return getAlign().value(); } |
| 221 | |
| 222 | /// Return the alignment of the access that is being performed. |
| 223 | Align getAlign() const { |
| 224 | return Align(1ULL << (getSubclassData<AlignmentField>())); |
| 225 | } |
| 226 | |
| 227 | void setAlignment(Align Align) { |
| 228 | setSubclassData<AlignmentField>(Log2(Align)); |
| 229 | } |
| 230 | |
| 231 | /// Returns the ordering constraint of this load instruction. |
| 232 | AtomicOrdering getOrdering() const { |
| 233 | return getSubclassData<OrderingField>(); |
| 234 | } |
| 235 | /// Sets the ordering constraint of this load instruction. May not be Release |
| 236 | /// or AcquireRelease. |
| 237 | void setOrdering(AtomicOrdering Ordering) { |
| 238 | setSubclassData<OrderingField>(Ordering); |
| 239 | } |
| 240 | |
| 241 | /// Returns the synchronization scope ID of this load instruction. |
| 242 | SyncScope::ID getSyncScopeID() const { |
| 243 | return SSID; |
| 244 | } |
| 245 | |
| 246 | /// Sets the synchronization scope ID of this load instruction. |
| 247 | void setSyncScopeID(SyncScope::ID SSID) { |
| 248 | this->SSID = SSID; |
| 249 | } |
| 250 | |
| 251 | /// Sets the ordering constraint and the synchronization scope ID of this load |
| 252 | /// instruction. |
| 253 | void setAtomic(AtomicOrdering Ordering, |
| 254 | SyncScope::ID SSID = SyncScope::System) { |
| 255 | setOrdering(Ordering); |
| 256 | setSyncScopeID(SSID); |
| 257 | } |
| 258 | |
| 259 | bool isSimple() const { return !isAtomic() && !isVolatile(); } |
| 260 | |
| 261 | bool isUnordered() const { |
| 262 | return (getOrdering() == AtomicOrdering::NotAtomic || |
| 263 | getOrdering() == AtomicOrdering::Unordered) && |
| 264 | !isVolatile(); |
| 265 | } |
| 266 | |
| 267 | Value *getPointerOperand() { return getOperand(0); } |
| 268 | const Value *getPointerOperand() const { return getOperand(0); } |
| 269 | static unsigned getPointerOperandIndex() { return 0U; } |
| 270 | Type *getPointerOperandType() const { return getPointerOperand()->getType(); } |
| 271 | |
| 272 | /// Returns the address space of the pointer operand. |
| 273 | unsigned getPointerAddressSpace() const { |
| 274 | return getPointerOperandType()->getPointerAddressSpace(); |
| 275 | } |
| 276 | |
| 277 | // Methods for support type inquiry through isa, cast, and dyn_cast: |
| 278 | static bool classof(const Instruction *I) { |
| 279 | return I->getOpcode() == Instruction::Load; |
| 280 | } |
| 281 | static bool classof(const Value *V) { |
| 282 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
| 283 | } |
| 284 | |
| 285 | private: |
| 286 | // Shadow Instruction::setInstructionSubclassData with a private forwarding |
| 287 | // method so that subclasses cannot accidentally use it. |
| 288 | template <typename Bitfield> |
| 289 | void setSubclassData(typename Bitfield::Type Value) { |
| 290 | Instruction::setSubclassData<Bitfield>(Value); |
| 291 | } |
| 292 | |
| 293 | /// The synchronization scope ID of this load instruction. Not quite enough |
| 294 | /// room in SubClassData for everything, so synchronization scope ID gets its |
| 295 | /// own field. |
| 296 | SyncScope::ID SSID; |
| 297 | }; |
| 298 | |
| 299 | //===----------------------------------------------------------------------===// |
| 300 | // StoreInst Class |
| 301 | //===----------------------------------------------------------------------===// |
| 302 | |
| 303 | /// An instruction for storing to memory. |
| 304 | class StoreInst : public Instruction { |
| 305 | using VolatileField = BoolBitfieldElementT<0>; |
| 306 | using AlignmentField = AlignmentBitfieldElementT<VolatileField::NextBit>; |
| 307 | using OrderingField = AtomicOrderingBitfieldElementT<AlignmentField::NextBit>; |
| 308 | static_assert( |
| 309 | Bitfield::areContiguous<VolatileField, AlignmentField, OrderingField>(), |
| 310 | "Bitfields must be contiguous"); |
| 311 | |
| 312 | void AssertOK(); |
| 313 | |
| 314 | protected: |
| 315 | // Note: Instruction needs to be a friend here to call cloneImpl. |
| 316 | friend class Instruction; |
| 317 | |
| 318 | StoreInst *cloneImpl() const; |
| 319 | |
| 320 | public: |
| 321 | StoreInst(Value *Val, Value *Ptr, Instruction *InsertBefore); |
| 322 | StoreInst(Value *Val, Value *Ptr, BasicBlock *InsertAtEnd); |
| 323 | StoreInst(Value *Val, Value *Ptr, bool isVolatile, Instruction *InsertBefore); |
| 324 | StoreInst(Value *Val, Value *Ptr, bool isVolatile, BasicBlock *InsertAtEnd); |
| 325 | StoreInst(Value *Val, Value *Ptr, bool isVolatile, Align Align, |
| 326 | Instruction *InsertBefore = nullptr); |
| 327 | StoreInst(Value *Val, Value *Ptr, bool isVolatile, Align Align, |
| 328 | BasicBlock *InsertAtEnd); |
| 329 | StoreInst(Value *Val, Value *Ptr, bool isVolatile, Align Align, |
| 330 | AtomicOrdering Order, SyncScope::ID SSID = SyncScope::System, |
| 331 | Instruction *InsertBefore = nullptr); |
| 332 | StoreInst(Value *Val, Value *Ptr, bool isVolatile, Align Align, |
| 333 | AtomicOrdering Order, SyncScope::ID SSID, BasicBlock *InsertAtEnd); |
| 334 | |
| 335 | // allocate space for exactly two operands |
| 336 | void *operator new(size_t S) { return User::operator new(S, 2); } |
| 337 | void operator delete(void *Ptr) { User::operator delete(Ptr); } |
| 338 | |
| 339 | /// Return true if this is a store to a volatile memory location. |
| 340 | bool isVolatile() const { return getSubclassData<VolatileField>(); } |
| 341 | |
| 342 | /// Specify whether this is a volatile store or not. |
| 343 | void setVolatile(bool V) { setSubclassData<VolatileField>(V); } |
| 344 | |
| 345 | /// Transparently provide more efficient getOperand methods. |
| 346 | 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; |
| 347 | |
| 348 | /// Return the alignment of the access that is being performed |
| 349 | /// FIXME: Remove this function once transition to Align is over. |
| 350 | /// Use getAlign() instead. |
| 351 | unsigned getAlignment() const { return getAlign().value(); } |
| 352 | |
| 353 | Align getAlign() const { |
| 354 | return Align(1ULL << (getSubclassData<AlignmentField>())); |
| 355 | } |
| 356 | |
| 357 | void setAlignment(Align Align) { |
| 358 | setSubclassData<AlignmentField>(Log2(Align)); |
| 359 | } |
| 360 | |
| 361 | /// Returns the ordering constraint of this store instruction. |
| 362 | AtomicOrdering getOrdering() const { |
| 363 | return getSubclassData<OrderingField>(); |
| 364 | } |
| 365 | |
| 366 | /// Sets the ordering constraint of this store instruction. May not be |
| 367 | /// Acquire or AcquireRelease. |
| 368 | void setOrdering(AtomicOrdering Ordering) { |
| 369 | setSubclassData<OrderingField>(Ordering); |
| 370 | } |
| 371 | |
| 372 | /// Returns the synchronization scope ID of this store instruction. |
| 373 | SyncScope::ID getSyncScopeID() const { |
| 374 | return SSID; |
| 375 | } |
| 376 | |
| 377 | /// Sets the synchronization scope ID of this store instruction. |
| 378 | void setSyncScopeID(SyncScope::ID SSID) { |
| 379 | this->SSID = SSID; |
| 380 | } |
| 381 | |
| 382 | /// Sets the ordering constraint and the synchronization scope ID of this |
| 383 | /// store instruction. |
| 384 | void setAtomic(AtomicOrdering Ordering, |
| 385 | SyncScope::ID SSID = SyncScope::System) { |
| 386 | setOrdering(Ordering); |
| 387 | setSyncScopeID(SSID); |
| 388 | } |
| 389 | |
| 390 | bool isSimple() const { return !isAtomic() && !isVolatile(); } |
| 391 | |
| 392 | bool isUnordered() const { |
| 393 | return (getOrdering() == AtomicOrdering::NotAtomic || |
| 394 | getOrdering() == AtomicOrdering::Unordered) && |
| 395 | !isVolatile(); |
| 396 | } |
| 397 | |
| 398 | Value *getValueOperand() { return getOperand(0); } |
| 399 | const Value *getValueOperand() const { return getOperand(0); } |
| 400 | |
| 401 | Value *getPointerOperand() { return getOperand(1); } |
| 402 | const Value *getPointerOperand() const { return getOperand(1); } |
| 403 | static unsigned getPointerOperandIndex() { return 1U; } |
| 404 | Type *getPointerOperandType() const { return getPointerOperand()->getType(); } |
| 405 | |
| 406 | /// Returns the address space of the pointer operand. |
| 407 | unsigned getPointerAddressSpace() const { |
| 408 | return getPointerOperandType()->getPointerAddressSpace(); |
| 409 | } |
| 410 | |
| 411 | // Methods for support type inquiry through isa, cast, and dyn_cast: |
| 412 | static bool classof(const Instruction *I) { |
| 413 | return I->getOpcode() == Instruction::Store; |
| 414 | } |
| 415 | static bool classof(const Value *V) { |
| 416 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
| 417 | } |
| 418 | |
| 419 | private: |
| 420 | // Shadow Instruction::setInstructionSubclassData with a private forwarding |
| 421 | // method so that subclasses cannot accidentally use it. |
| 422 | template <typename Bitfield> |
| 423 | void setSubclassData(typename Bitfield::Type Value) { |
| 424 | Instruction::setSubclassData<Bitfield>(Value); |
| 425 | } |
| 426 | |
| 427 | /// The synchronization scope ID of this store instruction. Not quite enough |
| 428 | /// room in SubClassData for everything, so synchronization scope ID gets its |
| 429 | /// own field. |
| 430 | SyncScope::ID SSID; |
| 431 | }; |
| 432 | |
| 433 | template <> |
| 434 | struct OperandTraits<StoreInst> : public FixedNumOperandTraits<StoreInst, 2> { |
| 435 | }; |
| 436 | |
| 437 | DEFINE_TRANSPARENT_OPERAND_ACCESSORS(StoreInst, Value)StoreInst::op_iterator StoreInst::op_begin() { return OperandTraits <StoreInst>::op_begin(this); } StoreInst::const_op_iterator StoreInst::op_begin() const { return OperandTraits<StoreInst >::op_begin(const_cast<StoreInst*>(this)); } StoreInst ::op_iterator StoreInst::op_end() { return OperandTraits<StoreInst >::op_end(this); } StoreInst::const_op_iterator StoreInst:: op_end() const { return OperandTraits<StoreInst>::op_end (const_cast<StoreInst*>(this)); } Value *StoreInst::getOperand (unsigned i_nocapture) const { ((void)0); return cast_or_null <Value>( OperandTraits<StoreInst>::op_begin(const_cast <StoreInst*>(this))[i_nocapture].get()); } void StoreInst ::setOperand(unsigned i_nocapture, Value *Val_nocapture) { (( void)0); OperandTraits<StoreInst>::op_begin(this)[i_nocapture ] = Val_nocapture; } unsigned StoreInst::getNumOperands() const { return OperandTraits<StoreInst>::operands(this); } template <int Idx_nocapture> Use &StoreInst::Op() { return this ->OpFrom<Idx_nocapture>(this); } template <int Idx_nocapture > const Use &StoreInst::Op() const { return this->OpFrom <Idx_nocapture>(this); } |
| 438 | |
| 439 | //===----------------------------------------------------------------------===// |
| 440 | // FenceInst Class |
| 441 | //===----------------------------------------------------------------------===// |
| 442 | |
| 443 | /// An instruction for ordering other memory operations. |
| 444 | class FenceInst : public Instruction { |
| 445 | using OrderingField = AtomicOrderingBitfieldElementT<0>; |
| 446 | |
| 447 | void Init(AtomicOrdering Ordering, SyncScope::ID SSID); |
| 448 | |
| 449 | protected: |
| 450 | // Note: Instruction needs to be a friend here to call cloneImpl. |
| 451 | friend class Instruction; |
| 452 | |
| 453 | FenceInst *cloneImpl() const; |
| 454 | |
| 455 | public: |
| 456 | // Ordering may only be Acquire, Release, AcquireRelease, or |
| 457 | // SequentiallyConsistent. |
| 458 | FenceInst(LLVMContext &C, AtomicOrdering Ordering, |
| 459 | SyncScope::ID SSID = SyncScope::System, |
| 460 | Instruction *InsertBefore = nullptr); |
| 461 | FenceInst(LLVMContext &C, AtomicOrdering Ordering, SyncScope::ID SSID, |
| 462 | BasicBlock *InsertAtEnd); |
| 463 | |
| 464 | // allocate space for exactly zero operands |
| 465 | void *operator new(size_t S) { return User::operator new(S, 0); } |
| 466 | void operator delete(void *Ptr) { User::operator delete(Ptr); } |
| 467 | |
| 468 | /// Returns the ordering constraint of this fence instruction. |
| 469 | AtomicOrdering getOrdering() const { |
| 470 | return getSubclassData<OrderingField>(); |
| 471 | } |
| 472 | |
| 473 | /// Sets the ordering constraint of this fence instruction. May only be |
| 474 | /// Acquire, Release, AcquireRelease, or SequentiallyConsistent. |
| 475 | void setOrdering(AtomicOrdering Ordering) { |
| 476 | setSubclassData<OrderingField>(Ordering); |
| 477 | } |
| 478 | |
| 479 | /// Returns the synchronization scope ID of this fence instruction. |
| 480 | SyncScope::ID getSyncScopeID() const { |
| 481 | return SSID; |
| 482 | } |
| 483 | |
| 484 | /// Sets the synchronization scope ID of this fence instruction. |
| 485 | void setSyncScopeID(SyncScope::ID SSID) { |
| 486 | this->SSID = SSID; |
| 487 | } |
| 488 | |
| 489 | // Methods for support type inquiry through isa, cast, and dyn_cast: |
| 490 | static bool classof(const Instruction *I) { |
| 491 | return I->getOpcode() == Instruction::Fence; |
| 492 | } |
| 493 | static bool classof(const Value *V) { |
| 494 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
| 495 | } |
| 496 | |
| 497 | private: |
| 498 | // Shadow Instruction::setInstructionSubclassData with a private forwarding |
| 499 | // method so that subclasses cannot accidentally use it. |
| 500 | template <typename Bitfield> |
| 501 | void setSubclassData(typename Bitfield::Type Value) { |
| 502 | Instruction::setSubclassData<Bitfield>(Value); |
| 503 | } |
| 504 | |
| 505 | /// The synchronization scope ID of this fence instruction. Not quite enough |
| 506 | /// room in SubClassData for everything, so synchronization scope ID gets its |
| 507 | /// own field. |
| 508 | SyncScope::ID SSID; |
| 509 | }; |
| 510 | |
| 511 | //===----------------------------------------------------------------------===// |
| 512 | // AtomicCmpXchgInst Class |
| 513 | //===----------------------------------------------------------------------===// |
| 514 | |
| 515 | /// An instruction that atomically checks whether a |
| 516 | /// specified value is in a memory location, and, if it is, stores a new value |
| 517 | /// there. The value returned by this instruction is a pair containing the |
| 518 | /// original value as first element, and an i1 indicating success (true) or |
| 519 | /// failure (false) as second element. |
| 520 | /// |
| 521 | class AtomicCmpXchgInst : public Instruction { |
| 522 | void Init(Value *Ptr, Value *Cmp, Value *NewVal, Align Align, |
| 523 | AtomicOrdering SuccessOrdering, AtomicOrdering FailureOrdering, |
| 524 | SyncScope::ID SSID); |
| 525 | |
| 526 | template <unsigned Offset> |
| 527 | using AtomicOrderingBitfieldElement = |
| 528 | typename Bitfield::Element<AtomicOrdering, Offset, 3, |
| 529 | AtomicOrdering::LAST>; |
| 530 | |
| 531 | protected: |
| 532 | // Note: Instruction needs to be a friend here to call cloneImpl. |
| 533 | friend class Instruction; |
| 534 | |
| 535 | AtomicCmpXchgInst *cloneImpl() const; |
| 536 | |
| 537 | public: |
| 538 | AtomicCmpXchgInst(Value *Ptr, Value *Cmp, Value *NewVal, Align Alignment, |
| 539 | AtomicOrdering SuccessOrdering, |
| 540 | AtomicOrdering FailureOrdering, SyncScope::ID SSID, |
| 541 | Instruction *InsertBefore = nullptr); |
| 542 | AtomicCmpXchgInst(Value *Ptr, Value *Cmp, Value *NewVal, Align Alignment, |
| 543 | AtomicOrdering SuccessOrdering, |
| 544 | AtomicOrdering FailureOrdering, SyncScope::ID SSID, |
| 545 | BasicBlock *InsertAtEnd); |
| 546 | |
| 547 | // allocate space for exactly three operands |
| 548 | void *operator new(size_t S) { return User::operator new(S, 3); } |
| 549 | void operator delete(void *Ptr) { User::operator delete(Ptr); } |
| 550 | |
| 551 | using VolatileField = BoolBitfieldElementT<0>; |
| 552 | using WeakField = BoolBitfieldElementT<VolatileField::NextBit>; |
| 553 | using SuccessOrderingField = |
| 554 | AtomicOrderingBitfieldElementT<WeakField::NextBit>; |
| 555 | using FailureOrderingField = |
| 556 | AtomicOrderingBitfieldElementT<SuccessOrderingField::NextBit>; |
| 557 | using AlignmentField = |
| 558 | AlignmentBitfieldElementT<FailureOrderingField::NextBit>; |
| 559 | static_assert( |
| 560 | Bitfield::areContiguous<VolatileField, WeakField, SuccessOrderingField, |
| 561 | FailureOrderingField, AlignmentField>(), |
| 562 | "Bitfields must be contiguous"); |
| 563 | |
| 564 | /// Return the alignment of the memory that is being allocated by the |
| 565 | /// instruction. |
| 566 | Align getAlign() const { |
| 567 | return Align(1ULL << getSubclassData<AlignmentField>()); |
| 568 | } |
| 569 | |
| 570 | void setAlignment(Align Align) { |
| 571 | setSubclassData<AlignmentField>(Log2(Align)); |
| 572 | } |
| 573 | |
| 574 | /// Return true if this is a cmpxchg from a volatile memory |
| 575 | /// location. |
| 576 | /// |
| 577 | bool isVolatile() const { return getSubclassData<VolatileField>(); } |
| 578 | |
| 579 | /// Specify whether this is a volatile cmpxchg. |
| 580 | /// |
| 581 | void setVolatile(bool V) { setSubclassData<VolatileField>(V); } |
| 582 | |
| 583 | /// Return true if this cmpxchg may spuriously fail. |
| 584 | bool isWeak() const { return getSubclassData<WeakField>(); } |
| 585 | |
| 586 | void setWeak(bool IsWeak) { setSubclassData<WeakField>(IsWeak); } |
| 587 | |
| 588 | /// Transparently provide more efficient getOperand methods. |
| 589 | 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; |
| 590 | |
| 591 | static bool isValidSuccessOrdering(AtomicOrdering Ordering) { |
| 592 | return Ordering != AtomicOrdering::NotAtomic && |
| 593 | Ordering != AtomicOrdering::Unordered; |
| 594 | } |
| 595 | |
| 596 | static bool isValidFailureOrdering(AtomicOrdering Ordering) { |
| 597 | return Ordering != AtomicOrdering::NotAtomic && |
| 598 | Ordering != AtomicOrdering::Unordered && |
| 599 | Ordering != AtomicOrdering::AcquireRelease && |
| 600 | Ordering != AtomicOrdering::Release; |
| 601 | } |
| 602 | |
| 603 | /// Returns the success ordering constraint of this cmpxchg instruction. |
| 604 | AtomicOrdering getSuccessOrdering() const { |
| 605 | return getSubclassData<SuccessOrderingField>(); |
| 606 | } |
| 607 | |
| 608 | /// Sets the success ordering constraint of this cmpxchg instruction. |
| 609 | void setSuccessOrdering(AtomicOrdering Ordering) { |
| 610 | assert(isValidSuccessOrdering(Ordering) &&((void)0) |
| 611 | "invalid CmpXchg success ordering")((void)0); |
| 612 | setSubclassData<SuccessOrderingField>(Ordering); |
| 613 | } |
| 614 | |
| 615 | /// Returns the failure ordering constraint of this cmpxchg instruction. |
| 616 | AtomicOrdering getFailureOrdering() const { |
| 617 | return getSubclassData<FailureOrderingField>(); |
| 618 | } |
| 619 | |
| 620 | /// Sets the failure ordering constraint of this cmpxchg instruction. |
| 621 | void setFailureOrdering(AtomicOrdering Ordering) { |
| 622 | assert(isValidFailureOrdering(Ordering) &&((void)0) |
| 623 | "invalid CmpXchg failure ordering")((void)0); |
| 624 | setSubclassData<FailureOrderingField>(Ordering); |
| 625 | } |
| 626 | |
| 627 | /// Returns a single ordering which is at least as strong as both the |
| 628 | /// success and failure orderings for this cmpxchg. |
| 629 | AtomicOrdering getMergedOrdering() const { |
| 630 | if (getFailureOrdering() == AtomicOrdering::SequentiallyConsistent) |
| 631 | return AtomicOrdering::SequentiallyConsistent; |
| 632 | if (getFailureOrdering() == AtomicOrdering::Acquire) { |
| 633 | if (getSuccessOrdering() == AtomicOrdering::Monotonic) |
| 634 | return AtomicOrdering::Acquire; |
| 635 | if (getSuccessOrdering() == AtomicOrdering::Release) |
| 636 | return AtomicOrdering::AcquireRelease; |
| 637 | } |
| 638 | return getSuccessOrdering(); |
| 639 | } |
| 640 | |
| 641 | /// Returns the synchronization scope ID of this cmpxchg instruction. |
| 642 | SyncScope::ID getSyncScopeID() const { |
| 643 | return SSID; |
| 644 | } |
| 645 | |
| 646 | /// Sets the synchronization scope ID of this cmpxchg instruction. |
| 647 | void setSyncScopeID(SyncScope::ID SSID) { |
| 648 | this->SSID = SSID; |
| 649 | } |
| 650 | |
| 651 | Value *getPointerOperand() { return getOperand(0); } |
| 652 | const Value *getPointerOperand() const { return getOperand(0); } |
| 653 | static unsigned getPointerOperandIndex() { return 0U; } |
| 654 | |
| 655 | Value *getCompareOperand() { return getOperand(1); } |
| 656 | const Value *getCompareOperand() const { return getOperand(1); } |
| 657 | |
| 658 | Value *getNewValOperand() { return getOperand(2); } |
| 659 | const Value *getNewValOperand() const { return getOperand(2); } |
| 660 | |
| 661 | /// Returns the address space of the pointer operand. |
| 662 | unsigned getPointerAddressSpace() const { |
| 663 | return getPointerOperand()->getType()->getPointerAddressSpace(); |
| 664 | } |
| 665 | |
| 666 | /// Returns the strongest permitted ordering on failure, given the |
| 667 | /// desired ordering on success. |
| 668 | /// |
| 669 | /// If the comparison in a cmpxchg operation fails, there is no atomic store |
| 670 | /// so release semantics cannot be provided. So this function drops explicit |
| 671 | /// Release requests from the AtomicOrdering. A SequentiallyConsistent |
| 672 | /// operation would remain SequentiallyConsistent. |
| 673 | static AtomicOrdering |
| 674 | getStrongestFailureOrdering(AtomicOrdering SuccessOrdering) { |
| 675 | switch (SuccessOrdering) { |
| 676 | default: |
| 677 | llvm_unreachable("invalid cmpxchg success ordering")__builtin_unreachable(); |
| 678 | case AtomicOrdering::Release: |
| 679 | case AtomicOrdering::Monotonic: |
| 680 | return AtomicOrdering::Monotonic; |
| 681 | case AtomicOrdering::AcquireRelease: |
| 682 | case AtomicOrdering::Acquire: |
| 683 | return AtomicOrdering::Acquire; |
| 684 | case AtomicOrdering::SequentiallyConsistent: |
| 685 | return AtomicOrdering::SequentiallyConsistent; |
| 686 | } |
| 687 | } |
| 688 | |
| 689 | // Methods for support type inquiry through isa, cast, and dyn_cast: |
| 690 | static bool classof(const Instruction *I) { |
| 691 | return I->getOpcode() == Instruction::AtomicCmpXchg; |
| 692 | } |
| 693 | static bool classof(const Value *V) { |
| 694 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
| 695 | } |
| 696 | |
| 697 | private: |
| 698 | // Shadow Instruction::setInstructionSubclassData with a private forwarding |
| 699 | // method so that subclasses cannot accidentally use it. |
| 700 | template <typename Bitfield> |
| 701 | void setSubclassData(typename Bitfield::Type Value) { |
| 702 | Instruction::setSubclassData<Bitfield>(Value); |
| 703 | } |
| 704 | |
| 705 | /// The synchronization scope ID of this cmpxchg instruction. Not quite |
| 706 | /// enough room in SubClassData for everything, so synchronization scope ID |
| 707 | /// gets its own field. |
| 708 | SyncScope::ID SSID; |
| 709 | }; |
| 710 | |
| 711 | template <> |
| 712 | struct OperandTraits<AtomicCmpXchgInst> : |
| 713 | public FixedNumOperandTraits<AtomicCmpXchgInst, 3> { |
| 714 | }; |
| 715 | |
| 716 | DEFINE_TRANSPARENT_OPERAND_ACCESSORS(AtomicCmpXchgInst, Value)AtomicCmpXchgInst::op_iterator AtomicCmpXchgInst::op_begin() { return OperandTraits<AtomicCmpXchgInst>::op_begin(this ); } AtomicCmpXchgInst::const_op_iterator AtomicCmpXchgInst:: op_begin() const { return OperandTraits<AtomicCmpXchgInst> ::op_begin(const_cast<AtomicCmpXchgInst*>(this)); } AtomicCmpXchgInst ::op_iterator AtomicCmpXchgInst::op_end() { return OperandTraits <AtomicCmpXchgInst>::op_end(this); } AtomicCmpXchgInst:: const_op_iterator AtomicCmpXchgInst::op_end() const { return OperandTraits <AtomicCmpXchgInst>::op_end(const_cast<AtomicCmpXchgInst *>(this)); } Value *AtomicCmpXchgInst::getOperand(unsigned i_nocapture) const { ((void)0); return cast_or_null<Value >( OperandTraits<AtomicCmpXchgInst>::op_begin(const_cast <AtomicCmpXchgInst*>(this))[i_nocapture].get()); } void AtomicCmpXchgInst::setOperand(unsigned i_nocapture, Value *Val_nocapture ) { ((void)0); OperandTraits<AtomicCmpXchgInst>::op_begin (this)[i_nocapture] = Val_nocapture; } unsigned AtomicCmpXchgInst ::getNumOperands() const { return OperandTraits<AtomicCmpXchgInst >::operands(this); } template <int Idx_nocapture> Use &AtomicCmpXchgInst::Op() { return this->OpFrom<Idx_nocapture >(this); } template <int Idx_nocapture> const Use & AtomicCmpXchgInst::Op() const { return this->OpFrom<Idx_nocapture >(this); } |
| 717 | |
| 718 | //===----------------------------------------------------------------------===// |
| 719 | // AtomicRMWInst Class |
| 720 | //===----------------------------------------------------------------------===// |
| 721 | |
| 722 | /// an instruction that atomically reads a memory location, |
| 723 | /// combines it with another value, and then stores the result back. Returns |
| 724 | /// the old value. |
| 725 | /// |
| 726 | class AtomicRMWInst : public Instruction { |
| 727 | protected: |
| 728 | // Note: Instruction needs to be a friend here to call cloneImpl. |
| 729 | friend class Instruction; |
| 730 | |
| 731 | AtomicRMWInst *cloneImpl() const; |
| 732 | |
| 733 | public: |
| 734 | /// This enumeration lists the possible modifications atomicrmw can make. In |
| 735 | /// the descriptions, 'p' is the pointer to the instruction's memory location, |
| 736 | /// 'old' is the initial value of *p, and 'v' is the other value passed to the |
| 737 | /// instruction. These instructions always return 'old'. |
| 738 | enum BinOp : unsigned { |
| 739 | /// *p = v |
| 740 | Xchg, |
| 741 | /// *p = old + v |
| 742 | Add, |
| 743 | /// *p = old - v |
| 744 | Sub, |
| 745 | /// *p = old & v |
| 746 | And, |
| 747 | /// *p = ~(old & v) |
| 748 | Nand, |
| 749 | /// *p = old | v |
| 750 | Or, |
| 751 | /// *p = old ^ v |
| 752 | Xor, |
| 753 | /// *p = old >signed v ? old : v |
| 754 | Max, |
| 755 | /// *p = old <signed v ? old : v |
| 756 | Min, |
| 757 | /// *p = old >unsigned v ? old : v |
| 758 | UMax, |
| 759 | /// *p = old <unsigned v ? old : v |
| 760 | UMin, |
| 761 | |
| 762 | /// *p = old + v |
| 763 | FAdd, |
| 764 | |
| 765 | /// *p = old - v |
| 766 | FSub, |
| 767 | |
| 768 | FIRST_BINOP = Xchg, |
| 769 | LAST_BINOP = FSub, |
| 770 | BAD_BINOP |
| 771 | }; |
| 772 | |
| 773 | private: |
| 774 | template <unsigned Offset> |
| 775 | using AtomicOrderingBitfieldElement = |
| 776 | typename Bitfield::Element<AtomicOrdering, Offset, 3, |
| 777 | AtomicOrdering::LAST>; |
| 778 | |
| 779 | template <unsigned Offset> |
| 780 | using BinOpBitfieldElement = |
| 781 | typename Bitfield::Element<BinOp, Offset, 4, BinOp::LAST_BINOP>; |
| 782 | |
| 783 | public: |
| 784 | AtomicRMWInst(BinOp Operation, Value *Ptr, Value *Val, Align Alignment, |
| 785 | AtomicOrdering Ordering, SyncScope::ID SSID, |
| 786 | Instruction *InsertBefore = nullptr); |
| 787 | AtomicRMWInst(BinOp Operation, Value *Ptr, Value *Val, Align Alignment, |
| 788 | AtomicOrdering Ordering, SyncScope::ID SSID, |
| 789 | BasicBlock *InsertAtEnd); |
| 790 | |
| 791 | // allocate space for exactly two operands |
| 792 | void *operator new(size_t S) { return User::operator new(S, 2); } |
| 793 | void operator delete(void *Ptr) { User::operator delete(Ptr); } |
| 794 | |
| 795 | using VolatileField = BoolBitfieldElementT<0>; |
| 796 | using AtomicOrderingField = |
| 797 | AtomicOrderingBitfieldElementT<VolatileField::NextBit>; |
| 798 | using OperationField = BinOpBitfieldElement<AtomicOrderingField::NextBit>; |
| 799 | using AlignmentField = AlignmentBitfieldElementT<OperationField::NextBit>; |
| 800 | static_assert(Bitfield::areContiguous<VolatileField, AtomicOrderingField, |
| 801 | OperationField, AlignmentField>(), |
| 802 | "Bitfields must be contiguous"); |
| 803 | |
| 804 | BinOp getOperation() const { return getSubclassData<OperationField>(); } |
| 805 | |
| 806 | static StringRef getOperationName(BinOp Op); |
| 807 | |
| 808 | static bool isFPOperation(BinOp Op) { |
| 809 | switch (Op) { |
| 810 | case AtomicRMWInst::FAdd: |
| 811 | case AtomicRMWInst::FSub: |
| 812 | return true; |
| 813 | default: |
| 814 | return false; |
| 815 | } |
| 816 | } |
| 817 | |
| 818 | void setOperation(BinOp Operation) { |
| 819 | setSubclassData<OperationField>(Operation); |
| 820 | } |
| 821 | |
| 822 | /// Return the alignment of the memory that is being allocated by the |
| 823 | /// instruction. |
| 824 | Align getAlign() const { |
| 825 | return Align(1ULL << getSubclassData<AlignmentField>()); |
| 826 | } |
| 827 | |
| 828 | void setAlignment(Align Align) { |
| 829 | setSubclassData<AlignmentField>(Log2(Align)); |
| 830 | } |
| 831 | |
| 832 | /// Return true if this is a RMW on a volatile memory location. |
| 833 | /// |
| 834 | bool isVolatile() const { return getSubclassData<VolatileField>(); } |
| 835 | |
| 836 | /// Specify whether this is a volatile RMW or not. |
| 837 | /// |
| 838 | void setVolatile(bool V) { setSubclassData<VolatileField>(V); } |
| 839 | |
| 840 | /// Transparently provide more efficient getOperand methods. |
| 841 | 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; |
| 842 | |
| 843 | /// Returns the ordering constraint of this rmw instruction. |
| 844 | AtomicOrdering getOrdering() const { |
| 845 | return getSubclassData<AtomicOrderingField>(); |
| 846 | } |
| 847 | |
| 848 | /// Sets the ordering constraint of this rmw instruction. |
| 849 | void setOrdering(AtomicOrdering Ordering) { |
| 850 | assert(Ordering != AtomicOrdering::NotAtomic &&((void)0) |
| 851 | "atomicrmw instructions can only be atomic.")((void)0); |
| 852 | setSubclassData<AtomicOrderingField>(Ordering); |
| 853 | } |
| 854 | |
| 855 | /// Returns the synchronization scope ID of this rmw instruction. |
| 856 | SyncScope::ID getSyncScopeID() const { |
| 857 | return SSID; |
| 858 | } |
| 859 | |
| 860 | /// Sets the synchronization scope ID of this rmw instruction. |
| 861 | void setSyncScopeID(SyncScope::ID SSID) { |
| 862 | this->SSID = SSID; |
| 863 | } |
| 864 | |
| 865 | Value *getPointerOperand() { return getOperand(0); } |
| 866 | const Value *getPointerOperand() const { return getOperand(0); } |
| 867 | static unsigned getPointerOperandIndex() { return 0U; } |
| 868 | |
| 869 | Value *getValOperand() { return getOperand(1); } |
| 870 | const Value *getValOperand() const { return getOperand(1); } |
| 871 | |
| 872 | /// Returns the address space of the pointer operand. |
| 873 | unsigned getPointerAddressSpace() const { |
| 874 | return getPointerOperand()->getType()->getPointerAddressSpace(); |
| 875 | } |
| 876 | |
| 877 | bool isFloatingPointOperation() const { |
| 878 | return isFPOperation(getOperation()); |
| 879 | } |
| 880 | |
| 881 | // Methods for support type inquiry through isa, cast, and dyn_cast: |
| 882 | static bool classof(const Instruction *I) { |
| 883 | return I->getOpcode() == Instruction::AtomicRMW; |
| 884 | } |
| 885 | static bool classof(const Value *V) { |
| 886 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
| 887 | } |
| 888 | |
| 889 | private: |
| 890 | void Init(BinOp Operation, Value *Ptr, Value *Val, Align Align, |
| 891 | AtomicOrdering Ordering, SyncScope::ID SSID); |
| 892 | |
| 893 | // Shadow Instruction::setInstructionSubclassData with a private forwarding |
| 894 | // method so that subclasses cannot accidentally use it. |
| 895 | template <typename Bitfield> |
| 896 | void setSubclassData(typename Bitfield::Type Value) { |
| 897 | Instruction::setSubclassData<Bitfield>(Value); |
| 898 | } |
| 899 | |
| 900 | /// The synchronization scope ID of this rmw instruction. Not quite enough |
| 901 | /// room in SubClassData for everything, so synchronization scope ID gets its |
| 902 | /// own field. |
| 903 | SyncScope::ID SSID; |
| 904 | }; |
| 905 | |
| 906 | template <> |
| 907 | struct OperandTraits<AtomicRMWInst> |
| 908 | : public FixedNumOperandTraits<AtomicRMWInst,2> { |
| 909 | }; |
| 910 | |
| 911 | DEFINE_TRANSPARENT_OPERAND_ACCESSORS(AtomicRMWInst, Value)AtomicRMWInst::op_iterator AtomicRMWInst::op_begin() { return OperandTraits<AtomicRMWInst>::op_begin(this); } AtomicRMWInst ::const_op_iterator AtomicRMWInst::op_begin() const { return OperandTraits <AtomicRMWInst>::op_begin(const_cast<AtomicRMWInst*> (this)); } AtomicRMWInst::op_iterator AtomicRMWInst::op_end() { return OperandTraits<AtomicRMWInst>::op_end(this); } AtomicRMWInst::const_op_iterator AtomicRMWInst::op_end() const { return OperandTraits<AtomicRMWInst>::op_end(const_cast <AtomicRMWInst*>(this)); } Value *AtomicRMWInst::getOperand (unsigned i_nocapture) const { ((void)0); return cast_or_null <Value>( OperandTraits<AtomicRMWInst>::op_begin(const_cast <AtomicRMWInst*>(this))[i_nocapture].get()); } void AtomicRMWInst ::setOperand(unsigned i_nocapture, Value *Val_nocapture) { (( void)0); OperandTraits<AtomicRMWInst>::op_begin(this)[i_nocapture ] = Val_nocapture; } unsigned AtomicRMWInst::getNumOperands() const { return OperandTraits<AtomicRMWInst>::operands( this); } template <int Idx_nocapture> Use &AtomicRMWInst ::Op() { return this->OpFrom<Idx_nocapture>(this); } template <int Idx_nocapture> const Use &AtomicRMWInst ::Op() const { return this->OpFrom<Idx_nocapture>(this ); } |
| 912 | |
| 913 | //===----------------------------------------------------------------------===// |
| 914 | // GetElementPtrInst Class |
| 915 | //===----------------------------------------------------------------------===// |
| 916 | |
| 917 | // checkGEPType - Simple wrapper function to give a better assertion failure |
| 918 | // message on bad indexes for a gep instruction. |
| 919 | // |
| 920 | inline Type *checkGEPType(Type *Ty) { |
| 921 | assert(Ty && "Invalid GetElementPtrInst indices for type!")((void)0); |
| 922 | return Ty; |
| 923 | } |
| 924 | |
| 925 | /// an instruction for type-safe pointer arithmetic to |
| 926 | /// access elements of arrays and structs |
| 927 | /// |
| 928 | class GetElementPtrInst : public Instruction { |
| 929 | Type *SourceElementType; |
| 930 | Type *ResultElementType; |
| 931 | |
| 932 | GetElementPtrInst(const GetElementPtrInst &GEPI); |
| 933 | |
| 934 | /// Constructors - Create a getelementptr instruction with a base pointer an |
| 935 | /// list of indices. The first ctor can optionally insert before an existing |
| 936 | /// instruction, the second appends the new instruction to the specified |
| 937 | /// BasicBlock. |
| 938 | inline GetElementPtrInst(Type *PointeeType, Value *Ptr, |
| 939 | ArrayRef<Value *> IdxList, unsigned Values, |
| 940 | const Twine &NameStr, Instruction *InsertBefore); |
| 941 | inline GetElementPtrInst(Type *PointeeType, Value *Ptr, |
| 942 | ArrayRef<Value *> IdxList, unsigned Values, |
| 943 | const Twine &NameStr, BasicBlock *InsertAtEnd); |
| 944 | |
| 945 | void init(Value *Ptr, ArrayRef<Value *> IdxList, const Twine &NameStr); |
| 946 | |
| 947 | protected: |
| 948 | // Note: Instruction needs to be a friend here to call cloneImpl. |
| 949 | friend class Instruction; |
| 950 | |
| 951 | GetElementPtrInst *cloneImpl() const; |
| 952 | |
| 953 | public: |
| 954 | static GetElementPtrInst *Create(Type *PointeeType, Value *Ptr, |
| 955 | ArrayRef<Value *> IdxList, |
| 956 | const Twine &NameStr = "", |
| 957 | Instruction *InsertBefore = nullptr) { |
| 958 | unsigned Values = 1 + unsigned(IdxList.size()); |
| 959 | assert(PointeeType && "Must specify element type")((void)0); |
| 960 | assert(cast<PointerType>(Ptr->getType()->getScalarType())((void)0) |
| 961 | ->isOpaqueOrPointeeTypeMatches(PointeeType))((void)0); |
| 962 | return new (Values) GetElementPtrInst(PointeeType, Ptr, IdxList, Values, |
| 963 | NameStr, InsertBefore); |
| 964 | } |
| 965 | |
| 966 | static GetElementPtrInst *Create(Type *PointeeType, Value *Ptr, |
| 967 | ArrayRef<Value *> IdxList, |
| 968 | const Twine &NameStr, |
| 969 | BasicBlock *InsertAtEnd) { |
| 970 | unsigned Values = 1 + unsigned(IdxList.size()); |
| 971 | assert(PointeeType && "Must specify element type")((void)0); |
| 972 | assert(cast<PointerType>(Ptr->getType()->getScalarType())((void)0) |
| 973 | ->isOpaqueOrPointeeTypeMatches(PointeeType))((void)0); |
| 974 | return new (Values) GetElementPtrInst(PointeeType, Ptr, IdxList, Values, |
| 975 | NameStr, InsertAtEnd); |
| 976 | } |
| 977 | |
| 978 | LLVM_ATTRIBUTE_DEPRECATED(static GetElementPtrInst *CreateInBounds([[deprecated("Use the version with explicit element type instead" )]] static GetElementPtrInst *CreateInBounds( Value *Ptr, ArrayRef <Value *> IdxList, const Twine &NameStr = "", Instruction *InsertBefore = nullptr) |
| 979 | Value *Ptr, ArrayRef<Value *> IdxList, const Twine &NameStr = "",[[deprecated("Use the version with explicit element type instead" )]] static GetElementPtrInst *CreateInBounds( Value *Ptr, ArrayRef <Value *> IdxList, const Twine &NameStr = "", Instruction *InsertBefore = nullptr) |
| 980 | Instruction *InsertBefore = nullptr),[[deprecated("Use the version with explicit element type instead" )]] static GetElementPtrInst *CreateInBounds( Value *Ptr, ArrayRef <Value *> IdxList, const Twine &NameStr = "", Instruction *InsertBefore = nullptr) |
| 981 | "Use the version with explicit element type instead")[[deprecated("Use the version with explicit element type instead" )]] static GetElementPtrInst *CreateInBounds( Value *Ptr, ArrayRef <Value *> IdxList, const Twine &NameStr = "", Instruction *InsertBefore = nullptr) { |
| 982 | return CreateInBounds( |
| 983 | Ptr->getType()->getScalarType()->getPointerElementType(), Ptr, IdxList, |
| 984 | NameStr, InsertBefore); |
| 985 | } |
| 986 | |
| 987 | /// Create an "inbounds" getelementptr. See the documentation for the |
| 988 | /// "inbounds" flag in LangRef.html for details. |
| 989 | static GetElementPtrInst * |
| 990 | CreateInBounds(Type *PointeeType, Value *Ptr, ArrayRef<Value *> IdxList, |
| 991 | const Twine &NameStr = "", |
| 992 | Instruction *InsertBefore = nullptr) { |
| 993 | GetElementPtrInst *GEP = |
| 994 | Create(PointeeType, Ptr, IdxList, NameStr, InsertBefore); |
| 995 | GEP->setIsInBounds(true); |
| 996 | return GEP; |
| 997 | } |
| 998 | |
| 999 | LLVM_ATTRIBUTE_DEPRECATED(static GetElementPtrInst *CreateInBounds([[deprecated("Use the version with explicit element type instead" )]] static GetElementPtrInst *CreateInBounds( Value *Ptr, ArrayRef <Value *> IdxList, const Twine &NameStr, BasicBlock *InsertAtEnd) |
| 1000 | Value *Ptr, ArrayRef<Value *> IdxList, const Twine &NameStr,[[deprecated("Use the version with explicit element type instead" )]] static GetElementPtrInst *CreateInBounds( Value *Ptr, ArrayRef <Value *> IdxList, const Twine &NameStr, BasicBlock *InsertAtEnd) |
| 1001 | BasicBlock *InsertAtEnd),[[deprecated("Use the version with explicit element type instead" )]] static GetElementPtrInst *CreateInBounds( Value *Ptr, ArrayRef <Value *> IdxList, const Twine &NameStr, BasicBlock *InsertAtEnd) |
| 1002 | "Use the version with explicit element type instead")[[deprecated("Use the version with explicit element type instead" )]] static GetElementPtrInst *CreateInBounds( Value *Ptr, ArrayRef <Value *> IdxList, const Twine &NameStr, BasicBlock *InsertAtEnd) { |
| 1003 | return CreateInBounds( |
| 1004 | Ptr->getType()->getScalarType()->getPointerElementType(), Ptr, IdxList, |
| 1005 | NameStr, InsertAtEnd); |
| 1006 | } |
| 1007 | |
| 1008 | static GetElementPtrInst *CreateInBounds(Type *PointeeType, Value *Ptr, |
| 1009 | ArrayRef<Value *> IdxList, |
| 1010 | const Twine &NameStr, |
| 1011 | BasicBlock *InsertAtEnd) { |
| 1012 | GetElementPtrInst *GEP = |
| 1013 | Create(PointeeType, Ptr, IdxList, NameStr, InsertAtEnd); |
| 1014 | GEP->setIsInBounds(true); |
| 1015 | return GEP; |
| 1016 | } |
| 1017 | |
| 1018 | /// Transparently provide more efficient getOperand methods. |
| 1019 | 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; |
| 1020 | |
| 1021 | Type *getSourceElementType() const { return SourceElementType; } |
| 1022 | |
| 1023 | void setSourceElementType(Type *Ty) { SourceElementType = Ty; } |
| 1024 | void setResultElementType(Type *Ty) { ResultElementType = Ty; } |
| 1025 | |
| 1026 | Type *getResultElementType() const { |
| 1027 | assert(cast<PointerType>(getType()->getScalarType())((void)0) |
| 1028 | ->isOpaqueOrPointeeTypeMatches(ResultElementType))((void)0); |
| 1029 | return ResultElementType; |
| 1030 | } |
| 1031 | |
| 1032 | /// Returns the address space of this instruction's pointer type. |
| 1033 | unsigned getAddressSpace() const { |
| 1034 | // Note that this is always the same as the pointer operand's address space |
| 1035 | // and that is cheaper to compute, so cheat here. |
| 1036 | return getPointerAddressSpace(); |
| 1037 | } |
| 1038 | |
| 1039 | /// Returns the result type of a getelementptr with the given source |
| 1040 | /// element type and indexes. |
| 1041 | /// |
| 1042 | /// Null is returned if the indices are invalid for the specified |
| 1043 | /// source element type. |
| 1044 | static Type *getIndexedType(Type *Ty, ArrayRef<Value *> IdxList); |
| 1045 | static Type *getIndexedType(Type *Ty, ArrayRef<Constant *> IdxList); |
| 1046 | static Type *getIndexedType(Type *Ty, ArrayRef<uint64_t> IdxList); |
| 1047 | |
| 1048 | /// Return the type of the element at the given index of an indexable |
| 1049 | /// type. This is equivalent to "getIndexedType(Agg, {Zero, Idx})". |
| 1050 | /// |
| 1051 | /// Returns null if the type can't be indexed, or the given index is not |
| 1052 | /// legal for the given type. |
| 1053 | static Type *getTypeAtIndex(Type *Ty, Value *Idx); |
| 1054 | static Type *getTypeAtIndex(Type *Ty, uint64_t Idx); |
| 1055 | |
| 1056 | inline op_iterator idx_begin() { return op_begin()+1; } |
| 1057 | inline const_op_iterator idx_begin() const { return op_begin()+1; } |
| 1058 | inline op_iterator idx_end() { return op_end(); } |
| 1059 | inline const_op_iterator idx_end() const { return op_end(); } |
| 1060 | |
| 1061 | inline iterator_range<op_iterator> indices() { |
| 1062 | return make_range(idx_begin(), idx_end()); |
| 1063 | } |
| 1064 | |
| 1065 | inline iterator_range<const_op_iterator> indices() const { |
| 1066 | return make_range(idx_begin(), idx_end()); |
| 1067 | } |
| 1068 | |
| 1069 | Value *getPointerOperand() { |
| 1070 | return getOperand(0); |
| 1071 | } |
| 1072 | const Value *getPointerOperand() const { |
| 1073 | return getOperand(0); |
| 1074 | } |
| 1075 | static unsigned getPointerOperandIndex() { |
| 1076 | return 0U; // get index for modifying correct operand. |
| 1077 | } |
| 1078 | |
| 1079 | /// Method to return the pointer operand as a |
| 1080 | /// PointerType. |
| 1081 | Type *getPointerOperandType() const { |
| 1082 | return getPointerOperand()->getType(); |
| 1083 | } |
| 1084 | |
| 1085 | /// Returns the address space of the pointer operand. |
| 1086 | unsigned getPointerAddressSpace() const { |
| 1087 | return getPointerOperandType()->getPointerAddressSpace(); |
| 1088 | } |
| 1089 | |
| 1090 | /// Returns the pointer type returned by the GEP |
| 1091 | /// instruction, which may be a vector of pointers. |
| 1092 | static Type *getGEPReturnType(Type *ElTy, Value *Ptr, |
| 1093 | ArrayRef<Value *> IdxList) { |
| 1094 | PointerType *OrigPtrTy = cast<PointerType>(Ptr->getType()->getScalarType()); |
| 1095 | unsigned AddrSpace = OrigPtrTy->getAddressSpace(); |
| 1096 | Type *ResultElemTy = checkGEPType(getIndexedType(ElTy, IdxList)); |
| 1097 | Type *PtrTy = OrigPtrTy->isOpaque() |
| 1098 | ? PointerType::get(OrigPtrTy->getContext(), AddrSpace) |
| 1099 | : PointerType::get(ResultElemTy, AddrSpace); |
| 1100 | // Vector GEP |
| 1101 | if (auto *PtrVTy = dyn_cast<VectorType>(Ptr->getType())) { |
| 1102 | ElementCount EltCount = PtrVTy->getElementCount(); |
| 1103 | return VectorType::get(PtrTy, EltCount); |
| 1104 | } |
| 1105 | for (Value *Index : IdxList) |
| 1106 | if (auto *IndexVTy = dyn_cast<VectorType>(Index->getType())) { |
| 1107 | ElementCount EltCount = IndexVTy->getElementCount(); |
| 1108 | return VectorType::get(PtrTy, EltCount); |
| 1109 | } |
| 1110 | // Scalar GEP |
| 1111 | return PtrTy; |
| 1112 | } |
| 1113 | |
| 1114 | unsigned getNumIndices() const { // Note: always non-negative |
| 1115 | return getNumOperands() - 1; |
| 1116 | } |
| 1117 | |
| 1118 | bool hasIndices() const { |
| 1119 | return getNumOperands() > 1; |
| 1120 | } |
| 1121 | |
| 1122 | /// Return true if all of the indices of this GEP are |
| 1123 | /// zeros. If so, the result pointer and the first operand have the same |
| 1124 | /// value, just potentially different types. |
| 1125 | bool hasAllZeroIndices() const; |
| 1126 | |
| 1127 | /// Return true if all of the indices of this GEP are |
| 1128 | /// constant integers. If so, the result pointer and the first operand have |
| 1129 | /// a constant offset between them. |
| 1130 | bool hasAllConstantIndices() const; |
| 1131 | |
| 1132 | /// Set or clear the inbounds flag on this GEP instruction. |
| 1133 | /// See LangRef.html for the meaning of inbounds on a getelementptr. |
| 1134 | void setIsInBounds(bool b = true); |
| 1135 | |
| 1136 | /// Determine whether the GEP has the inbounds flag. |
| 1137 | bool isInBounds() const; |
| 1138 | |
| 1139 | /// Accumulate the constant address offset of this GEP if possible. |
| 1140 | /// |
| 1141 | /// This routine accepts an APInt into which it will accumulate the constant |
| 1142 | /// offset of this GEP if the GEP is in fact constant. If the GEP is not |
| 1143 | /// all-constant, it returns false and the value of the offset APInt is |
| 1144 | /// undefined (it is *not* preserved!). The APInt passed into this routine |
| 1145 | /// must be at least as wide as the IntPtr type for the address space of |
| 1146 | /// the base GEP pointer. |
| 1147 | bool accumulateConstantOffset(const DataLayout &DL, APInt &Offset) const; |
| 1148 | bool collectOffset(const DataLayout &DL, unsigned BitWidth, |
| 1149 | MapVector<Value *, APInt> &VariableOffsets, |
| 1150 | APInt &ConstantOffset) const; |
| 1151 | // Methods for support type inquiry through isa, cast, and dyn_cast: |
| 1152 | static bool classof(const Instruction *I) { |
| 1153 | return (I->getOpcode() == Instruction::GetElementPtr); |
| 1154 | } |
| 1155 | static bool classof(const Value *V) { |
| 1156 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
| 1157 | } |
| 1158 | }; |
| 1159 | |
| 1160 | template <> |
| 1161 | struct OperandTraits<GetElementPtrInst> : |
| 1162 | public VariadicOperandTraits<GetElementPtrInst, 1> { |
| 1163 | }; |
| 1164 | |
| 1165 | GetElementPtrInst::GetElementPtrInst(Type *PointeeType, Value *Ptr, |
| 1166 | ArrayRef<Value *> IdxList, unsigned Values, |
| 1167 | const Twine &NameStr, |
| 1168 | Instruction *InsertBefore) |
| 1169 | : Instruction(getGEPReturnType(PointeeType, Ptr, IdxList), GetElementPtr, |
| 1170 | OperandTraits<GetElementPtrInst>::op_end(this) - Values, |
| 1171 | Values, InsertBefore), |
| 1172 | SourceElementType(PointeeType), |
| 1173 | ResultElementType(getIndexedType(PointeeType, IdxList)) { |
| 1174 | assert(cast<PointerType>(getType()->getScalarType())((void)0) |
| 1175 | ->isOpaqueOrPointeeTypeMatches(ResultElementType))((void)0); |
| 1176 | init(Ptr, IdxList, NameStr); |
| 1177 | } |
| 1178 | |
| 1179 | GetElementPtrInst::GetElementPtrInst(Type *PointeeType, Value *Ptr, |
| 1180 | ArrayRef<Value *> IdxList, unsigned Values, |
| 1181 | const Twine &NameStr, |
| 1182 | BasicBlock *InsertAtEnd) |
| 1183 | : Instruction(getGEPReturnType(PointeeType, Ptr, IdxList), GetElementPtr, |
| 1184 | OperandTraits<GetElementPtrInst>::op_end(this) - Values, |
| 1185 | Values, InsertAtEnd), |
| 1186 | SourceElementType(PointeeType), |
| 1187 | ResultElementType(getIndexedType(PointeeType, IdxList)) { |
| 1188 | assert(cast<PointerType>(getType()->getScalarType())((void)0) |
| 1189 | ->isOpaqueOrPointeeTypeMatches(ResultElementType))((void)0); |
| 1190 | init(Ptr, IdxList, NameStr); |
| 1191 | } |
| 1192 | |
| 1193 | DEFINE_TRANSPARENT_OPERAND_ACCESSORS(GetElementPtrInst, Value)GetElementPtrInst::op_iterator GetElementPtrInst::op_begin() { return OperandTraits<GetElementPtrInst>::op_begin(this ); } GetElementPtrInst::const_op_iterator GetElementPtrInst:: op_begin() const { return OperandTraits<GetElementPtrInst> ::op_begin(const_cast<GetElementPtrInst*>(this)); } GetElementPtrInst ::op_iterator GetElementPtrInst::op_end() { return OperandTraits <GetElementPtrInst>::op_end(this); } GetElementPtrInst:: const_op_iterator GetElementPtrInst::op_end() const { return OperandTraits <GetElementPtrInst>::op_end(const_cast<GetElementPtrInst *>(this)); } Value *GetElementPtrInst::getOperand(unsigned i_nocapture) const { ((void)0); return cast_or_null<Value >( OperandTraits<GetElementPtrInst>::op_begin(const_cast <GetElementPtrInst*>(this))[i_nocapture].get()); } void GetElementPtrInst::setOperand(unsigned i_nocapture, Value *Val_nocapture ) { ((void)0); OperandTraits<GetElementPtrInst>::op_begin (this)[i_nocapture] = Val_nocapture; } unsigned GetElementPtrInst ::getNumOperands() const { return OperandTraits<GetElementPtrInst >::operands(this); } template <int Idx_nocapture> Use &GetElementPtrInst::Op() { return this->OpFrom<Idx_nocapture >(this); } template <int Idx_nocapture> const Use & GetElementPtrInst::Op() const { return this->OpFrom<Idx_nocapture >(this); } |
| 1194 | |
| 1195 | //===----------------------------------------------------------------------===// |
| 1196 | // ICmpInst Class |
| 1197 | //===----------------------------------------------------------------------===// |
| 1198 | |
| 1199 | /// This instruction compares its operands according to the predicate given |
| 1200 | /// to the constructor. It only operates on integers or pointers. The operands |
| 1201 | /// must be identical types. |
| 1202 | /// Represent an integer comparison operator. |
| 1203 | class ICmpInst: public CmpInst { |
| 1204 | void AssertOK() { |
| 1205 | assert(isIntPredicate() &&((void)0) |
| 1206 | "Invalid ICmp predicate value")((void)0); |
| 1207 | assert(getOperand(0)->getType() == getOperand(1)->getType() &&((void)0) |
| 1208 | "Both operands to ICmp instruction are not of the same type!")((void)0); |
| 1209 | // Check that the operands are the right type |
| 1210 | assert((getOperand(0)->getType()->isIntOrIntVectorTy() ||((void)0) |
| 1211 | getOperand(0)->getType()->isPtrOrPtrVectorTy()) &&((void)0) |
| 1212 | "Invalid operand types for ICmp instruction")((void)0); |
| 1213 | } |
| 1214 | |
| 1215 | protected: |
| 1216 | // Note: Instruction needs to be a friend here to call cloneImpl. |
| 1217 | friend class Instruction; |
| 1218 | |
| 1219 | /// Clone an identical ICmpInst |
| 1220 | ICmpInst *cloneImpl() const; |
| 1221 | |
| 1222 | public: |
| 1223 | /// Constructor with insert-before-instruction semantics. |
| 1224 | ICmpInst( |
| 1225 | Instruction *InsertBefore, ///< Where to insert |
| 1226 | Predicate pred, ///< The predicate to use for the comparison |
| 1227 | Value *LHS, ///< The left-hand-side of the expression |
| 1228 | Value *RHS, ///< The right-hand-side of the expression |
| 1229 | const Twine &NameStr = "" ///< Name of the instruction |
| 1230 | ) : CmpInst(makeCmpResultType(LHS->getType()), |
| 1231 | Instruction::ICmp, pred, LHS, RHS, NameStr, |
| 1232 | InsertBefore) { |
| 1233 | #ifndef NDEBUG1 |
| 1234 | AssertOK(); |
| 1235 | #endif |
| 1236 | } |
| 1237 | |
| 1238 | /// Constructor with insert-at-end semantics. |
| 1239 | ICmpInst( |
| 1240 | BasicBlock &InsertAtEnd, ///< Block to insert into. |
| 1241 | Predicate pred, ///< The predicate to use for the comparison |
| 1242 | Value *LHS, ///< The left-hand-side of the expression |
| 1243 | Value *RHS, ///< The right-hand-side of the expression |
| 1244 | const Twine &NameStr = "" ///< Name of the instruction |
| 1245 | ) : CmpInst(makeCmpResultType(LHS->getType()), |
| 1246 | Instruction::ICmp, pred, LHS, RHS, NameStr, |
| 1247 | &InsertAtEnd) { |
| 1248 | #ifndef NDEBUG1 |
| 1249 | AssertOK(); |
| 1250 | #endif |
| 1251 | } |
| 1252 | |
| 1253 | /// Constructor with no-insertion semantics |
| 1254 | ICmpInst( |
| 1255 | Predicate pred, ///< The predicate to use for the comparison |
| 1256 | Value *LHS, ///< The left-hand-side of the expression |
| 1257 | Value *RHS, ///< The right-hand-side of the expression |
| 1258 | const Twine &NameStr = "" ///< Name of the instruction |
| 1259 | ) : CmpInst(makeCmpResultType(LHS->getType()), |
| 1260 | Instruction::ICmp, pred, LHS, RHS, NameStr) { |
| 1261 | #ifndef NDEBUG1 |
| 1262 | AssertOK(); |
| 1263 | #endif |
| 1264 | } |
| 1265 | |
| 1266 | /// For example, EQ->EQ, SLE->SLE, UGT->SGT, etc. |
| 1267 | /// @returns the predicate that would be the result if the operand were |
| 1268 | /// regarded as signed. |
| 1269 | /// Return the signed version of the predicate |
| 1270 | Predicate getSignedPredicate() const { |
| 1271 | return getSignedPredicate(getPredicate()); |
| 1272 | } |
| 1273 | |
| 1274 | /// This is a static version that you can use without an instruction. |
| 1275 | /// Return the signed version of the predicate. |
| 1276 | static Predicate getSignedPredicate(Predicate pred); |
| 1277 | |
| 1278 | /// For example, EQ->EQ, SLE->ULE, UGT->UGT, etc. |
| 1279 | /// @returns the predicate that would be the result if the operand were |
| 1280 | /// regarded as unsigned. |
| 1281 | /// Return the unsigned version of the predicate |
| 1282 | Predicate getUnsignedPredicate() const { |
| 1283 | return getUnsignedPredicate(getPredicate()); |
| 1284 | } |
| 1285 | |
| 1286 | /// This is a static version that you can use without an instruction. |
| 1287 | /// Return the unsigned version of the predicate. |
| 1288 | static Predicate getUnsignedPredicate(Predicate pred); |
| 1289 | |
| 1290 | /// Return true if this predicate is either EQ or NE. This also |
| 1291 | /// tests for commutativity. |
| 1292 | static bool isEquality(Predicate P) { |
| 1293 | return P == ICMP_EQ || P == ICMP_NE; |
| 1294 | } |
| 1295 | |
| 1296 | /// Return true if this predicate is either EQ or NE. This also |
| 1297 | /// tests for commutativity. |
| 1298 | bool isEquality() const { |
| 1299 | return isEquality(getPredicate()); |
| 1300 | } |
| 1301 | |
| 1302 | /// @returns true if the predicate of this ICmpInst is commutative |
| 1303 | /// Determine if this relation is commutative. |
| 1304 | bool isCommutative() const { return isEquality(); } |
| 1305 | |
| 1306 | /// Return true if the predicate is relational (not EQ or NE). |
| 1307 | /// |
| 1308 | bool isRelational() const { |
| 1309 | return !isEquality(); |
| 1310 | } |
| 1311 | |
| 1312 | /// Return true if the predicate is relational (not EQ or NE). |
| 1313 | /// |
| 1314 | static bool isRelational(Predicate P) { |
| 1315 | return !isEquality(P); |
| 1316 | } |
| 1317 | |
| 1318 | /// Return true if the predicate is SGT or UGT. |
| 1319 | /// |
| 1320 | static bool isGT(Predicate P) { |
| 1321 | return P == ICMP_SGT || P == ICMP_UGT; |
| 1322 | } |
| 1323 | |
| 1324 | /// Return true if the predicate is SLT or ULT. |
| 1325 | /// |
| 1326 | static bool isLT(Predicate P) { |
| 1327 | return P == ICMP_SLT || P == ICMP_ULT; |
| 1328 | } |
| 1329 | |
| 1330 | /// Return true if the predicate is SGE or UGE. |
| 1331 | /// |
| 1332 | static bool isGE(Predicate P) { |
| 1333 | return P == ICMP_SGE || P == ICMP_UGE; |
| 1334 | } |
| 1335 | |
| 1336 | /// Return true if the predicate is SLE or ULE. |
| 1337 | /// |
| 1338 | static bool isLE(Predicate P) { |
| 1339 | return P == ICMP_SLE || P == ICMP_ULE; |
| 1340 | } |
| 1341 | |
| 1342 | /// Exchange the two operands to this instruction in such a way that it does |
| 1343 | /// not modify the semantics of the instruction. The predicate value may be |
| 1344 | /// changed to retain the same result if the predicate is order dependent |
| 1345 | /// (e.g. ult). |
| 1346 | /// Swap operands and adjust predicate. |
| 1347 | void swapOperands() { |
| 1348 | setPredicate(getSwappedPredicate()); |
| 1349 | Op<0>().swap(Op<1>()); |
| 1350 | } |
| 1351 | |
| 1352 | // Methods for support type inquiry through isa, cast, and dyn_cast: |
| 1353 | static bool classof(const Instruction *I) { |
| 1354 | return I->getOpcode() == Instruction::ICmp; |
| 1355 | } |
| 1356 | static bool classof(const Value *V) { |
| 1357 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
| 1358 | } |
| 1359 | }; |
| 1360 | |
| 1361 | //===----------------------------------------------------------------------===// |
| 1362 | // FCmpInst Class |
| 1363 | //===----------------------------------------------------------------------===// |
| 1364 | |
| 1365 | /// This instruction compares its operands according to the predicate given |
| 1366 | /// to the constructor. It only operates on floating point values or packed |
| 1367 | /// vectors of floating point values. The operands must be identical types. |
| 1368 | /// Represents a floating point comparison operator. |
| 1369 | class FCmpInst: public CmpInst { |
| 1370 | void AssertOK() { |
| 1371 | assert(isFPPredicate() && "Invalid FCmp predicate value")((void)0); |
| 1372 | assert(getOperand(0)->getType() == getOperand(1)->getType() &&((void)0) |
| 1373 | "Both operands to FCmp instruction are not of the same type!")((void)0); |
| 1374 | // Check that the operands are the right type |
| 1375 | assert(getOperand(0)->getType()->isFPOrFPVectorTy() &&((void)0) |
| 1376 | "Invalid operand types for FCmp instruction")((void)0); |
| 1377 | } |
| 1378 | |
| 1379 | protected: |
| 1380 | // Note: Instruction needs to be a friend here to call cloneImpl. |
| 1381 | friend class Instruction; |
| 1382 | |
| 1383 | /// Clone an identical FCmpInst |
| 1384 | FCmpInst *cloneImpl() const; |
| 1385 | |
| 1386 | public: |
| 1387 | /// Constructor with insert-before-instruction semantics. |
| 1388 | FCmpInst( |
| 1389 | Instruction *InsertBefore, ///< Where to insert |
| 1390 | Predicate pred, ///< The predicate to use for the comparison |
| 1391 | Value *LHS, ///< The left-hand-side of the expression |
| 1392 | Value *RHS, ///< The right-hand-side of the expression |
| 1393 | const Twine &NameStr = "" ///< Name of the instruction |
| 1394 | ) : CmpInst(makeCmpResultType(LHS->getType()), |
| 1395 | Instruction::FCmp, pred, LHS, RHS, NameStr, |
| 1396 | InsertBefore) { |
| 1397 | AssertOK(); |
| 1398 | } |
| 1399 | |
| 1400 | /// Constructor with insert-at-end semantics. |
| 1401 | FCmpInst( |
| 1402 | BasicBlock &InsertAtEnd, ///< Block to insert into. |
| 1403 | Predicate pred, ///< The predicate to use for the comparison |
| 1404 | Value *LHS, ///< The left-hand-side of the expression |
| 1405 | Value *RHS, ///< The right-hand-side of the expression |
| 1406 | const Twine &NameStr = "" ///< Name of the instruction |
| 1407 | ) : CmpInst(makeCmpResultType(LHS->getType()), |
| 1408 | Instruction::FCmp, pred, LHS, RHS, NameStr, |
| 1409 | &InsertAtEnd) { |
| 1410 | AssertOK(); |
| 1411 | } |
| 1412 | |
| 1413 | /// Constructor with no-insertion semantics |
| 1414 | FCmpInst( |
| 1415 | Predicate Pred, ///< The predicate to use for the comparison |
| 1416 | Value *LHS, ///< The left-hand-side of the expression |
| 1417 | Value *RHS, ///< The right-hand-side of the expression |
| 1418 | const Twine &NameStr = "", ///< Name of the instruction |
| 1419 | Instruction *FlagsSource = nullptr |
| 1420 | ) : CmpInst(makeCmpResultType(LHS->getType()), Instruction::FCmp, Pred, LHS, |
| 1421 | RHS, NameStr, nullptr, FlagsSource) { |
| 1422 | AssertOK(); |
| 1423 | } |
| 1424 | |
| 1425 | /// @returns true if the predicate of this instruction is EQ or NE. |
| 1426 | /// Determine if this is an equality predicate. |
| 1427 | static bool isEquality(Predicate Pred) { |
| 1428 | return Pred == FCMP_OEQ || Pred == FCMP_ONE || Pred == FCMP_UEQ || |
| 1429 | Pred == FCMP_UNE; |
| 1430 | } |
| 1431 | |
| 1432 | /// @returns true if the predicate of this instruction is EQ or NE. |
| 1433 | /// Determine if this is an equality predicate. |
| 1434 | bool isEquality() const { return isEquality(getPredicate()); } |
| 1435 | |
| 1436 | /// @returns true if the predicate of this instruction is commutative. |
| 1437 | /// Determine if this is a commutative predicate. |
| 1438 | bool isCommutative() const { |
| 1439 | return isEquality() || |
| 1440 | getPredicate() == FCMP_FALSE || |
| 1441 | getPredicate() == FCMP_TRUE || |
| 1442 | getPredicate() == FCMP_ORD || |
| 1443 | getPredicate() == FCMP_UNO; |
| 1444 | } |
| 1445 | |
| 1446 | /// @returns true if the predicate is relational (not EQ or NE). |
| 1447 | /// Determine if this a relational predicate. |
| 1448 | bool isRelational() const { return !isEquality(); } |
| 1449 | |
| 1450 | /// Exchange the two operands to this instruction in such a way that it does |
| 1451 | /// not modify the semantics of the instruction. The predicate value may be |
| 1452 | /// changed to retain the same result if the predicate is order dependent |
| 1453 | /// (e.g. ult). |
| 1454 | /// Swap operands and adjust predicate. |
| 1455 | void swapOperands() { |
| 1456 | setPredicate(getSwappedPredicate()); |
| 1457 | Op<0>().swap(Op<1>()); |
| 1458 | } |
| 1459 | |
| 1460 | /// Methods for support type inquiry through isa, cast, and dyn_cast: |
| 1461 | static bool classof(const Instruction *I) { |
| 1462 | return I->getOpcode() == Instruction::FCmp; |
| 1463 | } |
| 1464 | static bool classof(const Value *V) { |
| 1465 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
| 1466 | } |
| 1467 | }; |
| 1468 | |
| 1469 | //===----------------------------------------------------------------------===// |
| 1470 | /// This class represents a function call, abstracting a target |
| 1471 | /// machine's calling convention. This class uses low bit of the SubClassData |
| 1472 | /// field to indicate whether or not this is a tail call. The rest of the bits |
| 1473 | /// hold the calling convention of the call. |
| 1474 | /// |
| 1475 | class CallInst : public CallBase { |
| 1476 | CallInst(const CallInst &CI); |
| 1477 | |
| 1478 | /// Construct a CallInst given a range of arguments. |
| 1479 | /// Construct a CallInst from a range of arguments |
| 1480 | inline CallInst(FunctionType *Ty, Value *Func, ArrayRef<Value *> Args, |
| 1481 | ArrayRef<OperandBundleDef> Bundles, const Twine &NameStr, |
| 1482 | Instruction *InsertBefore); |
| 1483 | |
| 1484 | inline CallInst(FunctionType *Ty, Value *Func, ArrayRef<Value *> Args, |
| 1485 | const Twine &NameStr, Instruction *InsertBefore) |
| 1486 | : CallInst(Ty, Func, Args, None, NameStr, InsertBefore) {} |
| 1487 | |
| 1488 | /// Construct a CallInst given a range of arguments. |
| 1489 | /// Construct a CallInst from a range of arguments |
| 1490 | inline CallInst(FunctionType *Ty, Value *Func, ArrayRef<Value *> Args, |
| 1491 | ArrayRef<OperandBundleDef> Bundles, const Twine &NameStr, |
| 1492 | BasicBlock *InsertAtEnd); |
| 1493 | |
| 1494 | explicit CallInst(FunctionType *Ty, Value *F, const Twine &NameStr, |
| 1495 | Instruction *InsertBefore); |
| 1496 | |
| 1497 | CallInst(FunctionType *ty, Value *F, const Twine &NameStr, |
| 1498 | BasicBlock *InsertAtEnd); |
| 1499 | |
| 1500 | void init(FunctionType *FTy, Value *Func, ArrayRef<Value *> Args, |
| 1501 | ArrayRef<OperandBundleDef> Bundles, const Twine &NameStr); |
| 1502 | void init(FunctionType *FTy, Value *Func, const Twine &NameStr); |
| 1503 | |
| 1504 | /// Compute the number of operands to allocate. |
| 1505 | static int ComputeNumOperands(int NumArgs, int NumBundleInputs = 0) { |
| 1506 | // We need one operand for the called function, plus the input operand |
| 1507 | // counts provided. |
| 1508 | return 1 + NumArgs + NumBundleInputs; |
| 1509 | } |
| 1510 | |
| 1511 | protected: |
| 1512 | // Note: Instruction needs to be a friend here to call cloneImpl. |
| 1513 | friend class Instruction; |
| 1514 | |
| 1515 | CallInst *cloneImpl() const; |
| 1516 | |
| 1517 | public: |
| 1518 | static CallInst *Create(FunctionType *Ty, Value *F, const Twine &NameStr = "", |
| 1519 | Instruction *InsertBefore = nullptr) { |
| 1520 | return new (ComputeNumOperands(0)) CallInst(Ty, F, NameStr, InsertBefore); |
| 1521 | } |
| 1522 | |
| 1523 | static CallInst *Create(FunctionType *Ty, Value *Func, ArrayRef<Value *> Args, |
| 1524 | const Twine &NameStr, |
| 1525 | Instruction *InsertBefore = nullptr) { |
| 1526 | return new (ComputeNumOperands(Args.size())) |
| 1527 | CallInst(Ty, Func, Args, None, NameStr, InsertBefore); |
| 1528 | } |
| 1529 | |
| 1530 | static CallInst *Create(FunctionType *Ty, Value *Func, ArrayRef<Value *> Args, |
| 1531 | ArrayRef<OperandBundleDef> Bundles = None, |
| 1532 | const Twine &NameStr = "", |
| 1533 | Instruction *InsertBefore = nullptr) { |
| 1534 | const int NumOperands = |
| 1535 | ComputeNumOperands(Args.size(), CountBundleInputs(Bundles)); |
| 1536 | const unsigned DescriptorBytes = Bundles.size() * sizeof(BundleOpInfo); |
| 1537 | |
| 1538 | return new (NumOperands, DescriptorBytes) |
| 1539 | CallInst(Ty, Func, Args, Bundles, NameStr, InsertBefore); |
| 1540 | } |
| 1541 | |
| 1542 | static CallInst *Create(FunctionType *Ty, Value *F, const Twine &NameStr, |
| 1543 | BasicBlock *InsertAtEnd) { |
| 1544 | return new (ComputeNumOperands(0)) CallInst(Ty, F, NameStr, InsertAtEnd); |
| 1545 | } |
| 1546 | |
| 1547 | static CallInst *Create(FunctionType *Ty, Value *Func, ArrayRef<Value *> Args, |
| 1548 | const Twine &NameStr, BasicBlock *InsertAtEnd) { |
| 1549 | return new (ComputeNumOperands(Args.size())) |
| 1550 | CallInst(Ty, Func, Args, None, NameStr, InsertAtEnd); |
| 1551 | } |
| 1552 | |
| 1553 | static CallInst *Create(FunctionType *Ty, Value *Func, ArrayRef<Value *> Args, |
| 1554 | ArrayRef<OperandBundleDef> Bundles, |
| 1555 | const Twine &NameStr, BasicBlock *InsertAtEnd) { |
| 1556 | const int NumOperands = |
| 1557 | ComputeNumOperands(Args.size(), CountBundleInputs(Bundles)); |
| 1558 | const unsigned DescriptorBytes = Bundles.size() * sizeof(BundleOpInfo); |
| 1559 | |
| 1560 | return new (NumOperands, DescriptorBytes) |
| 1561 | CallInst(Ty, Func, Args, Bundles, NameStr, InsertAtEnd); |
| 1562 | } |
| 1563 | |
| 1564 | static CallInst *Create(FunctionCallee Func, const Twine &NameStr = "", |
| 1565 | Instruction *InsertBefore = nullptr) { |
| 1566 | return Create(Func.getFunctionType(), Func.getCallee(), NameStr, |
| 1567 | InsertBefore); |
| 1568 | } |
| 1569 | |
| 1570 | static CallInst *Create(FunctionCallee Func, ArrayRef<Value *> Args, |
| 1571 | ArrayRef<OperandBundleDef> Bundles = None, |
| 1572 | const Twine &NameStr = "", |
| 1573 | Instruction *InsertBefore = nullptr) { |
| 1574 | return Create(Func.getFunctionType(), Func.getCallee(), Args, Bundles, |
| 1575 | NameStr, InsertBefore); |
| 1576 | } |
| 1577 | |
| 1578 | static CallInst *Create(FunctionCallee Func, ArrayRef<Value *> Args, |
| 1579 | const Twine &NameStr, |
| 1580 | Instruction *InsertBefore = nullptr) { |
| 1581 | return Create(Func.getFunctionType(), Func.getCallee(), Args, NameStr, |
| 1582 | InsertBefore); |
| 1583 | } |
| 1584 | |
| 1585 | static CallInst *Create(FunctionCallee Func, const Twine &NameStr, |
| 1586 | BasicBlock *InsertAtEnd) { |
| 1587 | return Create(Func.getFunctionType(), Func.getCallee(), NameStr, |
| 1588 | InsertAtEnd); |
| 1589 | } |
| 1590 | |
| 1591 | static CallInst *Create(FunctionCallee Func, ArrayRef<Value *> Args, |
| 1592 | const Twine &NameStr, BasicBlock *InsertAtEnd) { |
| 1593 | return Create(Func.getFunctionType(), Func.getCallee(), Args, NameStr, |
| 1594 | InsertAtEnd); |
| 1595 | } |
| 1596 | |
| 1597 | static CallInst *Create(FunctionCallee Func, ArrayRef<Value *> Args, |
| 1598 | ArrayRef<OperandBundleDef> Bundles, |
| 1599 | const Twine &NameStr, BasicBlock *InsertAtEnd) { |
| 1600 | return Create(Func.getFunctionType(), Func.getCallee(), Args, Bundles, |
| 1601 | NameStr, InsertAtEnd); |
| 1602 | } |
| 1603 | |
| 1604 | /// Create a clone of \p CI with a different set of operand bundles and |
| 1605 | /// insert it before \p InsertPt. |
| 1606 | /// |
| 1607 | /// The returned call instruction is identical \p CI in every way except that |
| 1608 | /// the operand bundles for the new instruction are set to the operand bundles |
| 1609 | /// in \p Bundles. |
| 1610 | static CallInst *Create(CallInst *CI, ArrayRef<OperandBundleDef> Bundles, |
| 1611 | Instruction *InsertPt = nullptr); |
| 1612 | |
| 1613 | /// Generate the IR for a call to malloc: |
| 1614 | /// 1. Compute the malloc call's argument as the specified type's size, |
| 1615 | /// possibly multiplied by the array size if the array size is not |
| 1616 | /// constant 1. |
| 1617 | /// 2. Call malloc with that argument. |
| 1618 | /// 3. Bitcast the result of the malloc call to the specified type. |
| 1619 | static Instruction *CreateMalloc(Instruction *InsertBefore, Type *IntPtrTy, |
| 1620 | Type *AllocTy, Value *AllocSize, |
| 1621 | Value *ArraySize = nullptr, |
| 1622 | Function *MallocF = nullptr, |
| 1623 | const Twine &Name = ""); |
| 1624 | static Instruction *CreateMalloc(BasicBlock *InsertAtEnd, Type *IntPtrTy, |
| 1625 | Type *AllocTy, Value *AllocSize, |
| 1626 | Value *ArraySize = nullptr, |
| 1627 | Function *MallocF = nullptr, |
| 1628 | const Twine &Name = ""); |
| 1629 | static Instruction *CreateMalloc(Instruction *InsertBefore, Type *IntPtrTy, |
| 1630 | Type *AllocTy, Value *AllocSize, |
| 1631 | Value *ArraySize = nullptr, |
| 1632 | ArrayRef<OperandBundleDef> Bundles = None, |
| 1633 | Function *MallocF = nullptr, |
| 1634 | const Twine &Name = ""); |
| 1635 | static Instruction *CreateMalloc(BasicBlock *InsertAtEnd, Type *IntPtrTy, |
| 1636 | Type *AllocTy, Value *AllocSize, |
| 1637 | Value *ArraySize = nullptr, |
| 1638 | ArrayRef<OperandBundleDef> Bundles = None, |
| 1639 | Function *MallocF = nullptr, |
| 1640 | const Twine &Name = ""); |
| 1641 | /// Generate the IR for a call to the builtin free function. |
| 1642 | static Instruction *CreateFree(Value *Source, Instruction *InsertBefore); |
| 1643 | static Instruction *CreateFree(Value *Source, BasicBlock *InsertAtEnd); |
| 1644 | static Instruction *CreateFree(Value *Source, |
| 1645 | ArrayRef<OperandBundleDef> Bundles, |
| 1646 | Instruction *InsertBefore); |
| 1647 | static Instruction *CreateFree(Value *Source, |
| 1648 | ArrayRef<OperandBundleDef> Bundles, |
| 1649 | BasicBlock *InsertAtEnd); |
| 1650 | |
| 1651 | // Note that 'musttail' implies 'tail'. |
| 1652 | enum TailCallKind : unsigned { |
| 1653 | TCK_None = 0, |
| 1654 | TCK_Tail = 1, |
| 1655 | TCK_MustTail = 2, |
| 1656 | TCK_NoTail = 3, |
| 1657 | TCK_LAST = TCK_NoTail |
| 1658 | }; |
| 1659 | |
| 1660 | using TailCallKindField = Bitfield::Element<TailCallKind, 0, 2, TCK_LAST>; |
| 1661 | static_assert( |
| 1662 | Bitfield::areContiguous<TailCallKindField, CallBase::CallingConvField>(), |
| 1663 | "Bitfields must be contiguous"); |
| 1664 | |
| 1665 | TailCallKind getTailCallKind() const { |
| 1666 | return getSubclassData<TailCallKindField>(); |
| 1667 | } |
| 1668 | |
| 1669 | bool isTailCall() const { |
| 1670 | TailCallKind Kind = getTailCallKind(); |
| 1671 | return Kind == TCK_Tail || Kind == TCK_MustTail; |
| 1672 | } |
| 1673 | |
| 1674 | bool isMustTailCall() const { return getTailCallKind() == TCK_MustTail; } |
| 1675 | |
| 1676 | bool isNoTailCall() const { return getTailCallKind() == TCK_NoTail; } |
| 1677 | |
| 1678 | void setTailCallKind(TailCallKind TCK) { |
| 1679 | setSubclassData<TailCallKindField>(TCK); |
| 1680 | } |
| 1681 | |
| 1682 | void setTailCall(bool IsTc = true) { |
| 1683 | setTailCallKind(IsTc ? TCK_Tail : TCK_None); |
| 1684 | } |
| 1685 | |
| 1686 | /// Return true if the call can return twice |
| 1687 | bool canReturnTwice() const { return hasFnAttr(Attribute::ReturnsTwice); } |
| 1688 | void setCanReturnTwice() { |
| 1689 | addAttribute(AttributeList::FunctionIndex, Attribute::ReturnsTwice); |
| 1690 | } |
| 1691 | |
| 1692 | // Methods for support type inquiry through isa, cast, and dyn_cast: |
| 1693 | static bool classof(const Instruction *I) { |
| 1694 | return I->getOpcode() == Instruction::Call; |
| 1695 | } |
| 1696 | static bool classof(const Value *V) { |
| 1697 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
| 1698 | } |
| 1699 | |
| 1700 | /// Updates profile metadata by scaling it by \p S / \p T. |
| 1701 | void updateProfWeight(uint64_t S, uint64_t T); |
| 1702 | |
| 1703 | private: |
| 1704 | // Shadow Instruction::setInstructionSubclassData with a private forwarding |
| 1705 | // method so that subclasses cannot accidentally use it. |
| 1706 | template <typename Bitfield> |
| 1707 | void setSubclassData(typename Bitfield::Type Value) { |
| 1708 | Instruction::setSubclassData<Bitfield>(Value); |
| 1709 | } |
| 1710 | }; |
| 1711 | |
| 1712 | CallInst::CallInst(FunctionType *Ty, Value *Func, ArrayRef<Value *> Args, |
| 1713 | ArrayRef<OperandBundleDef> Bundles, const Twine &NameStr, |
| 1714 | BasicBlock *InsertAtEnd) |
| 1715 | : CallBase(Ty->getReturnType(), Instruction::Call, |
| 1716 | OperandTraits<CallBase>::op_end(this) - |
| 1717 | (Args.size() + CountBundleInputs(Bundles) + 1), |
| 1718 | unsigned(Args.size() + CountBundleInputs(Bundles) + 1), |
| 1719 | InsertAtEnd) { |
| 1720 | init(Ty, Func, Args, Bundles, NameStr); |
| 1721 | } |
| 1722 | |
| 1723 | CallInst::CallInst(FunctionType *Ty, Value *Func, ArrayRef<Value *> Args, |
| 1724 | ArrayRef<OperandBundleDef> Bundles, const Twine &NameStr, |
| 1725 | Instruction *InsertBefore) |
| 1726 | : CallBase(Ty->getReturnType(), Instruction::Call, |
| 1727 | OperandTraits<CallBase>::op_end(this) - |
| 1728 | (Args.size() + CountBundleInputs(Bundles) + 1), |
| 1729 | unsigned(Args.size() + CountBundleInputs(Bundles) + 1), |
| 1730 | InsertBefore) { |
| 1731 | init(Ty, Func, Args, Bundles, NameStr); |
| 1732 | } |
| 1733 | |
| 1734 | //===----------------------------------------------------------------------===// |
| 1735 | // SelectInst Class |
| 1736 | //===----------------------------------------------------------------------===// |
| 1737 | |
| 1738 | /// This class represents the LLVM 'select' instruction. |
| 1739 | /// |
| 1740 | class SelectInst : public Instruction { |
| 1741 | SelectInst(Value *C, Value *S1, Value *S2, const Twine &NameStr, |
| 1742 | Instruction *InsertBefore) |
| 1743 | : Instruction(S1->getType(), Instruction::Select, |
| 1744 | &Op<0>(), 3, InsertBefore) { |
| 1745 | init(C, S1, S2); |
| 1746 | setName(NameStr); |
| 1747 | } |
| 1748 | |
| 1749 | SelectInst(Value *C, Value *S1, Value *S2, const Twine &NameStr, |
| 1750 | BasicBlock *InsertAtEnd) |
| 1751 | : Instruction(S1->getType(), Instruction::Select, |
| 1752 | &Op<0>(), 3, InsertAtEnd) { |
| 1753 | init(C, S1, S2); |
| 1754 | setName(NameStr); |
| 1755 | } |
| 1756 | |
| 1757 | void init(Value *C, Value *S1, Value *S2) { |
| 1758 | assert(!areInvalidOperands(C, S1, S2) && "Invalid operands for select")((void)0); |
| 1759 | Op<0>() = C; |
| 1760 | Op<1>() = S1; |
| 1761 | Op<2>() = S2; |
| 1762 | } |
| 1763 | |
| 1764 | protected: |
| 1765 | // Note: Instruction needs to be a friend here to call cloneImpl. |
| 1766 | friend class Instruction; |
| 1767 | |
| 1768 | SelectInst *cloneImpl() const; |
| 1769 | |
| 1770 | public: |
| 1771 | static SelectInst *Create(Value *C, Value *S1, Value *S2, |
| 1772 | const Twine &NameStr = "", |
| 1773 | Instruction *InsertBefore = nullptr, |
| 1774 | Instruction *MDFrom = nullptr) { |
| 1775 | SelectInst *Sel = new(3) SelectInst(C, S1, S2, NameStr, InsertBefore); |
| 1776 | if (MDFrom) |
| 1777 | Sel->copyMetadata(*MDFrom); |
| 1778 | return Sel; |
| 1779 | } |
| 1780 | |
| 1781 | static SelectInst *Create(Value *C, Value *S1, Value *S2, |
| 1782 | const Twine &NameStr, |
| 1783 | BasicBlock *InsertAtEnd) { |
| 1784 | return new(3) SelectInst(C, S1, S2, NameStr, InsertAtEnd); |
| 1785 | } |
| 1786 | |
| 1787 | const Value *getCondition() const { return Op<0>(); } |
| 1788 | const Value *getTrueValue() const { return Op<1>(); } |
| 1789 | const Value *getFalseValue() const { return Op<2>(); } |
| 1790 | Value *getCondition() { return Op<0>(); } |
| 1791 | Value *getTrueValue() { return Op<1>(); } |
| 1792 | Value *getFalseValue() { return Op<2>(); } |
| 1793 | |
| 1794 | void setCondition(Value *V) { Op<0>() = V; } |
| 1795 | void setTrueValue(Value *V) { Op<1>() = V; } |
| 1796 | void setFalseValue(Value *V) { Op<2>() = V; } |
| 1797 | |
| 1798 | /// Swap the true and false values of the select instruction. |
| 1799 | /// This doesn't swap prof metadata. |
| 1800 | void swapValues() { Op<1>().swap(Op<2>()); } |
| 1801 | |
| 1802 | /// Return a string if the specified operands are invalid |
| 1803 | /// for a select operation, otherwise return null. |
| 1804 | static const char *areInvalidOperands(Value *Cond, Value *True, Value *False); |
| 1805 | |
| 1806 | /// Transparently provide more efficient getOperand methods. |
| 1807 | 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; |
| 1808 | |
| 1809 | OtherOps getOpcode() const { |
| 1810 | return static_cast<OtherOps>(Instruction::getOpcode()); |
| 1811 | } |
| 1812 | |
| 1813 | // Methods for support type inquiry through isa, cast, and dyn_cast: |
| 1814 | static bool classof(const Instruction *I) { |
| 1815 | return I->getOpcode() == Instruction::Select; |
| 1816 | } |
| 1817 | static bool classof(const Value *V) { |
| 1818 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
| 1819 | } |
| 1820 | }; |
| 1821 | |
| 1822 | template <> |
| 1823 | struct OperandTraits<SelectInst> : public FixedNumOperandTraits<SelectInst, 3> { |
| 1824 | }; |
| 1825 | |
| 1826 | DEFINE_TRANSPARENT_OPERAND_ACCESSORS(SelectInst, Value)SelectInst::op_iterator SelectInst::op_begin() { return OperandTraits <SelectInst>::op_begin(this); } SelectInst::const_op_iterator SelectInst::op_begin() const { return OperandTraits<SelectInst >::op_begin(const_cast<SelectInst*>(this)); } SelectInst ::op_iterator SelectInst::op_end() { return OperandTraits< SelectInst>::op_end(this); } SelectInst::const_op_iterator SelectInst::op_end() const { return OperandTraits<SelectInst >::op_end(const_cast<SelectInst*>(this)); } Value *SelectInst ::getOperand(unsigned i_nocapture) const { ((void)0); return cast_or_null <Value>( OperandTraits<SelectInst>::op_begin(const_cast <SelectInst*>(this))[i_nocapture].get()); } void SelectInst ::setOperand(unsigned i_nocapture, Value *Val_nocapture) { (( void)0); OperandTraits<SelectInst>::op_begin(this)[i_nocapture ] = Val_nocapture; } unsigned SelectInst::getNumOperands() const { return OperandTraits<SelectInst>::operands(this); } template <int Idx_nocapture> Use &SelectInst::Op() { return this->OpFrom<Idx_nocapture>(this); } template <int Idx_nocapture> const Use &SelectInst::Op() const { return this->OpFrom<Idx_nocapture>(this); } |
| 1827 | |
| 1828 | //===----------------------------------------------------------------------===// |
| 1829 | // VAArgInst Class |
| 1830 | //===----------------------------------------------------------------------===// |
| 1831 | |
| 1832 | /// This class represents the va_arg llvm instruction, which returns |
| 1833 | /// an argument of the specified type given a va_list and increments that list |
| 1834 | /// |
| 1835 | class VAArgInst : public UnaryInstruction { |
| 1836 | protected: |
| 1837 | // Note: Instruction needs to be a friend here to call cloneImpl. |
| 1838 | friend class Instruction; |
| 1839 | |
| 1840 | VAArgInst *cloneImpl() const; |
| 1841 | |
| 1842 | public: |
| 1843 | VAArgInst(Value *List, Type *Ty, const Twine &NameStr = "", |
| 1844 | Instruction *InsertBefore = nullptr) |
| 1845 | : UnaryInstruction(Ty, VAArg, List, InsertBefore) { |
| 1846 | setName(NameStr); |
| 1847 | } |
| 1848 | |
| 1849 | VAArgInst(Value *List, Type *Ty, const Twine &NameStr, |
| 1850 | BasicBlock *InsertAtEnd) |
| 1851 | : UnaryInstruction(Ty, VAArg, List, InsertAtEnd) { |
| 1852 | setName(NameStr); |
| 1853 | } |
| 1854 | |
| 1855 | Value *getPointerOperand() { return getOperand(0); } |
| 1856 | const Value *getPointerOperand() const { return getOperand(0); } |
| 1857 | static unsigned getPointerOperandIndex() { return 0U; } |
| 1858 | |
| 1859 | // Methods for support type inquiry through isa, cast, and dyn_cast: |
| 1860 | static bool classof(const Instruction *I) { |
| 1861 | return I->getOpcode() == VAArg; |
| 1862 | } |
| 1863 | static bool classof(const Value *V) { |
| 1864 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
| 1865 | } |
| 1866 | }; |
| 1867 | |
| 1868 | //===----------------------------------------------------------------------===// |
| 1869 | // ExtractElementInst Class |
| 1870 | //===----------------------------------------------------------------------===// |
| 1871 | |
| 1872 | /// This instruction extracts a single (scalar) |
| 1873 | /// element from a VectorType value |
| 1874 | /// |
| 1875 | class ExtractElementInst : public Instruction { |
| 1876 | ExtractElementInst(Value *Vec, Value *Idx, const Twine &NameStr = "", |
| 1877 | Instruction *InsertBefore = nullptr); |
| 1878 | ExtractElementInst(Value *Vec, Value *Idx, const Twine &NameStr, |
| 1879 | BasicBlock *InsertAtEnd); |
| 1880 | |
| 1881 | protected: |
| 1882 | // Note: Instruction needs to be a friend here to call cloneImpl. |
| 1883 | friend class Instruction; |
| 1884 | |
| 1885 | ExtractElementInst *cloneImpl() const; |
| 1886 | |
| 1887 | public: |
| 1888 | static ExtractElementInst *Create(Value *Vec, Value *Idx, |
| 1889 | const Twine &NameStr = "", |
| 1890 | Instruction *InsertBefore = nullptr) { |
| 1891 | return new(2) ExtractElementInst(Vec, Idx, NameStr, InsertBefore); |
| 1892 | } |
| 1893 | |
| 1894 | static ExtractElementInst *Create(Value *Vec, Value *Idx, |
| 1895 | const Twine &NameStr, |
| 1896 | BasicBlock *InsertAtEnd) { |
| 1897 | return new(2) ExtractElementInst(Vec, Idx, NameStr, InsertAtEnd); |
| 1898 | } |
| 1899 | |
| 1900 | /// Return true if an extractelement instruction can be |
| 1901 | /// formed with the specified operands. |
| 1902 | static bool isValidOperands(const Value *Vec, const Value *Idx); |
| 1903 | |
| 1904 | Value *getVectorOperand() { return Op<0>(); } |
| 1905 | Value *getIndexOperand() { return Op<1>(); } |
| 1906 | const Value *getVectorOperand() const { return Op<0>(); } |
| 1907 | const Value *getIndexOperand() const { return Op<1>(); } |
| 1908 | |
| 1909 | VectorType *getVectorOperandType() const { |
| 1910 | return cast<VectorType>(getVectorOperand()->getType()); |
| 1911 | } |
| 1912 | |
| 1913 | /// Transparently provide more efficient getOperand methods. |
| 1914 | 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; |
| 1915 | |
| 1916 | // Methods for support type inquiry through isa, cast, and dyn_cast: |
| 1917 | static bool classof(const Instruction *I) { |
| 1918 | return I->getOpcode() == Instruction::ExtractElement; |
| 1919 | } |
| 1920 | static bool classof(const Value *V) { |
| 1921 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
| 1922 | } |
| 1923 | }; |
| 1924 | |
| 1925 | template <> |
| 1926 | struct OperandTraits<ExtractElementInst> : |
| 1927 | public FixedNumOperandTraits<ExtractElementInst, 2> { |
| 1928 | }; |
| 1929 | |
| 1930 | DEFINE_TRANSPARENT_OPERAND_ACCESSORS(ExtractElementInst, Value)ExtractElementInst::op_iterator ExtractElementInst::op_begin( ) { return OperandTraits<ExtractElementInst>::op_begin( this); } ExtractElementInst::const_op_iterator ExtractElementInst ::op_begin() const { return OperandTraits<ExtractElementInst >::op_begin(const_cast<ExtractElementInst*>(this)); } ExtractElementInst::op_iterator ExtractElementInst::op_end() { return OperandTraits<ExtractElementInst>::op_end(this ); } ExtractElementInst::const_op_iterator ExtractElementInst ::op_end() const { return OperandTraits<ExtractElementInst >::op_end(const_cast<ExtractElementInst*>(this)); } Value *ExtractElementInst::getOperand(unsigned i_nocapture) const { ((void)0); return cast_or_null<Value>( OperandTraits< ExtractElementInst>::op_begin(const_cast<ExtractElementInst *>(this))[i_nocapture].get()); } void ExtractElementInst:: setOperand(unsigned i_nocapture, Value *Val_nocapture) { ((void )0); OperandTraits<ExtractElementInst>::op_begin(this)[ i_nocapture] = Val_nocapture; } unsigned ExtractElementInst:: getNumOperands() const { return OperandTraits<ExtractElementInst >::operands(this); } template <int Idx_nocapture> Use &ExtractElementInst::Op() { return this->OpFrom<Idx_nocapture >(this); } template <int Idx_nocapture> const Use & ExtractElementInst::Op() const { return this->OpFrom<Idx_nocapture >(this); } |
| 1931 | |
| 1932 | //===----------------------------------------------------------------------===// |
| 1933 | // InsertElementInst Class |
| 1934 | //===----------------------------------------------------------------------===// |
| 1935 | |
| 1936 | /// This instruction inserts a single (scalar) |
| 1937 | /// element into a VectorType value |
| 1938 | /// |
| 1939 | class InsertElementInst : public Instruction { |
| 1940 | InsertElementInst(Value *Vec, Value *NewElt, Value *Idx, |
| 1941 | const Twine &NameStr = "", |
| 1942 | Instruction *InsertBefore = nullptr); |
| 1943 | InsertElementInst(Value *Vec, Value *NewElt, Value *Idx, const Twine &NameStr, |
| 1944 | BasicBlock *InsertAtEnd); |
| 1945 | |
| 1946 | protected: |
| 1947 | // Note: Instruction needs to be a friend here to call cloneImpl. |
| 1948 | friend class Instruction; |
| 1949 | |
| 1950 | InsertElementInst *cloneImpl() const; |
| 1951 | |
| 1952 | public: |
| 1953 | static InsertElementInst *Create(Value *Vec, Value *NewElt, Value *Idx, |
| 1954 | const Twine &NameStr = "", |
| 1955 | Instruction *InsertBefore = nullptr) { |
| 1956 | return new(3) InsertElementInst(Vec, NewElt, Idx, NameStr, InsertBefore); |
| 1957 | } |
| 1958 | |
| 1959 | static InsertElementInst *Create(Value *Vec, Value *NewElt, Value *Idx, |
| 1960 | const Twine &NameStr, |
| 1961 | BasicBlock *InsertAtEnd) { |
| 1962 | return new(3) InsertElementInst(Vec, NewElt, Idx, NameStr, InsertAtEnd); |
| 1963 | } |
| 1964 | |
| 1965 | /// Return true if an insertelement instruction can be |
| 1966 | /// formed with the specified operands. |
| 1967 | static bool isValidOperands(const Value *Vec, const Value *NewElt, |
| 1968 | const Value *Idx); |
| 1969 | |
| 1970 | /// Overload to return most specific vector type. |
| 1971 | /// |
| 1972 | VectorType *getType() const { |
| 1973 | return cast<VectorType>(Instruction::getType()); |
| 1974 | } |
| 1975 | |
| 1976 | /// Transparently provide more efficient getOperand methods. |
| 1977 | 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; |
| 1978 | |
| 1979 | // Methods for support type inquiry through isa, cast, and dyn_cast: |
| 1980 | static bool classof(const Instruction *I) { |
| 1981 | return I->getOpcode() == Instruction::InsertElement; |
| 1982 | } |
| 1983 | static bool classof(const Value *V) { |
| 1984 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
| 1985 | } |
| 1986 | }; |
| 1987 | |
| 1988 | template <> |
| 1989 | struct OperandTraits<InsertElementInst> : |
| 1990 | public FixedNumOperandTraits<InsertElementInst, 3> { |
| 1991 | }; |
| 1992 | |
| 1993 | DEFINE_TRANSPARENT_OPERAND_ACCESSORS(InsertElementInst, Value)InsertElementInst::op_iterator InsertElementInst::op_begin() { return OperandTraits<InsertElementInst>::op_begin(this ); } InsertElementInst::const_op_iterator InsertElementInst:: op_begin() const { return OperandTraits<InsertElementInst> ::op_begin(const_cast<InsertElementInst*>(this)); } InsertElementInst ::op_iterator InsertElementInst::op_end() { return OperandTraits <InsertElementInst>::op_end(this); } InsertElementInst:: const_op_iterator InsertElementInst::op_end() const { return OperandTraits <InsertElementInst>::op_end(const_cast<InsertElementInst *>(this)); } Value *InsertElementInst::getOperand(unsigned i_nocapture) const { ((void)0); return cast_or_null<Value >( OperandTraits<InsertElementInst>::op_begin(const_cast <InsertElementInst*>(this))[i_nocapture].get()); } void InsertElementInst::setOperand(unsigned i_nocapture, Value *Val_nocapture ) { ((void)0); OperandTraits<InsertElementInst>::op_begin (this)[i_nocapture] = Val_nocapture; } unsigned InsertElementInst ::getNumOperands() const { return OperandTraits<InsertElementInst >::operands(this); } template <int Idx_nocapture> Use &InsertElementInst::Op() { return this->OpFrom<Idx_nocapture >(this); } template <int Idx_nocapture> const Use & InsertElementInst::Op() const { return this->OpFrom<Idx_nocapture >(this); } |
| 1994 | |
| 1995 | //===----------------------------------------------------------------------===// |
| 1996 | // ShuffleVectorInst Class |
| 1997 | //===----------------------------------------------------------------------===// |
| 1998 | |
| 1999 | constexpr int UndefMaskElem = -1; |
| 2000 | |
| 2001 | /// This instruction constructs a fixed permutation of two |
| 2002 | /// input vectors. |
| 2003 | /// |
| 2004 | /// For each element of the result vector, the shuffle mask selects an element |
| 2005 | /// from one of the input vectors to copy to the result. Non-negative elements |
| 2006 | /// in the mask represent an index into the concatenated pair of input vectors. |
| 2007 | /// UndefMaskElem (-1) specifies that the result element is undefined. |
| 2008 | /// |
| 2009 | /// For scalable vectors, all the elements of the mask must be 0 or -1. This |
| 2010 | /// requirement may be relaxed in the future. |
| 2011 | class ShuffleVectorInst : public Instruction { |
| 2012 | SmallVector<int, 4> ShuffleMask; |
| 2013 | Constant *ShuffleMaskForBitcode; |
| 2014 | |
| 2015 | protected: |
| 2016 | // Note: Instruction needs to be a friend here to call cloneImpl. |
| 2017 | friend class Instruction; |
| 2018 | |
| 2019 | ShuffleVectorInst *cloneImpl() const; |
| 2020 | |
| 2021 | public: |
| 2022 | ShuffleVectorInst(Value *V1, Value *V2, Value *Mask, |
| 2023 | const Twine &NameStr = "", |
| 2024 | Instruction *InsertBefor = nullptr); |
| 2025 | ShuffleVectorInst(Value *V1, Value *V2, Value *Mask, |
| 2026 | const Twine &NameStr, BasicBlock *InsertAtEnd); |
| 2027 | ShuffleVectorInst(Value *V1, Value *V2, ArrayRef<int> Mask, |
| 2028 | const Twine &NameStr = "", |
| 2029 | Instruction *InsertBefor = nullptr); |
| 2030 | ShuffleVectorInst(Value *V1, Value *V2, ArrayRef<int> Mask, |
| 2031 | const Twine &NameStr, BasicBlock *InsertAtEnd); |
| 2032 | |
| 2033 | void *operator new(size_t S) { return User::operator new(S, 2); } |
| 2034 | void operator delete(void *Ptr) { return User::operator delete(Ptr); } |
| 2035 | |
| 2036 | /// Swap the operands and adjust the mask to preserve the semantics |
| 2037 | /// of the instruction. |
| 2038 | void commute(); |
| 2039 | |
| 2040 | /// Return true if a shufflevector instruction can be |
| 2041 | /// formed with the specified operands. |
| 2042 | static bool isValidOperands(const Value *V1, const Value *V2, |
| 2043 | const Value *Mask); |
| 2044 | static bool isValidOperands(const Value *V1, const Value *V2, |
| 2045 | ArrayRef<int> Mask); |
| 2046 | |
| 2047 | /// Overload to return most specific vector type. |
| 2048 | /// |
| 2049 | VectorType *getType() const { |
| 2050 | return cast<VectorType>(Instruction::getType()); |
| 2051 | } |
| 2052 | |
| 2053 | /// Transparently provide more efficient getOperand methods. |
| 2054 | 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; |
| 2055 | |
| 2056 | /// Return the shuffle mask value of this instruction for the given element |
| 2057 | /// index. Return UndefMaskElem if the element is undef. |
| 2058 | int getMaskValue(unsigned Elt) const { return ShuffleMask[Elt]; } |
| 2059 | |
| 2060 | /// Convert the input shuffle mask operand to a vector of integers. Undefined |
| 2061 | /// elements of the mask are returned as UndefMaskElem. |
| 2062 | static void getShuffleMask(const Constant *Mask, |
| 2063 | SmallVectorImpl<int> &Result); |
| 2064 | |
| 2065 | /// Return the mask for this instruction as a vector of integers. Undefined |
| 2066 | /// elements of the mask are returned as UndefMaskElem. |
| 2067 | void getShuffleMask(SmallVectorImpl<int> &Result) const { |
| 2068 | Result.assign(ShuffleMask.begin(), ShuffleMask.end()); |
| 2069 | } |
| 2070 | |
| 2071 | /// Return the mask for this instruction, for use in bitcode. |
| 2072 | /// |
| 2073 | /// TODO: This is temporary until we decide a new bitcode encoding for |
| 2074 | /// shufflevector. |
| 2075 | Constant *getShuffleMaskForBitcode() const { return ShuffleMaskForBitcode; } |
| 2076 | |
| 2077 | static Constant *convertShuffleMaskForBitcode(ArrayRef<int> Mask, |
| 2078 | Type *ResultTy); |
| 2079 | |
| 2080 | void setShuffleMask(ArrayRef<int> Mask); |
| 2081 | |
| 2082 | ArrayRef<int> getShuffleMask() const { return ShuffleMask; } |
| 2083 | |
| 2084 | /// Return true if this shuffle returns a vector with a different number of |
| 2085 | /// elements than its source vectors. |
| 2086 | /// Examples: shufflevector <4 x n> A, <4 x n> B, <1,2,3> |
| 2087 | /// shufflevector <4 x n> A, <4 x n> B, <1,2,3,4,5> |
| 2088 | bool changesLength() const { |
| 2089 | unsigned NumSourceElts = cast<VectorType>(Op<0>()->getType()) |
| 2090 | ->getElementCount() |
| 2091 | .getKnownMinValue(); |
| 2092 | unsigned NumMaskElts = ShuffleMask.size(); |
| 2093 | return NumSourceElts != NumMaskElts; |
| 2094 | } |
| 2095 | |
| 2096 | /// Return true if this shuffle returns a vector with a greater number of |
| 2097 | /// elements than its source vectors. |
| 2098 | /// Example: shufflevector <2 x n> A, <2 x n> B, <1,2,3> |
| 2099 | bool increasesLength() const { |
| 2100 | unsigned NumSourceElts = cast<VectorType>(Op<0>()->getType()) |
| 2101 | ->getElementCount() |
| 2102 | .getKnownMinValue(); |
| 2103 | unsigned NumMaskElts = ShuffleMask.size(); |
| 2104 | return NumSourceElts < NumMaskElts; |
| 2105 | } |
| 2106 | |
| 2107 | /// Return true if this shuffle mask chooses elements from exactly one source |
| 2108 | /// vector. |
| 2109 | /// Example: <7,5,undef,7> |
| 2110 | /// This assumes that vector operands are the same length as the mask. |
| 2111 | static bool isSingleSourceMask(ArrayRef<int> Mask); |
| 2112 | static bool isSingleSourceMask(const Constant *Mask) { |
| 2113 | assert(Mask->getType()->isVectorTy() && "Shuffle needs vector constant.")((void)0); |
| 2114 | SmallVector<int, 16> MaskAsInts; |
| 2115 | getShuffleMask(Mask, MaskAsInts); |
| 2116 | return isSingleSourceMask(MaskAsInts); |
| 2117 | } |
| 2118 | |
| 2119 | /// Return true if this shuffle chooses elements from exactly one source |
| 2120 | /// vector without changing the length of that vector. |
| 2121 | /// Example: shufflevector <4 x n> A, <4 x n> B, <3,0,undef,3> |
| 2122 | /// TODO: Optionally allow length-changing shuffles. |
| 2123 | bool isSingleSource() const { |
| 2124 | return !changesLength() && isSingleSourceMask(ShuffleMask); |
| 2125 | } |
| 2126 | |
| 2127 | /// Return true if this shuffle mask chooses elements from exactly one source |
| 2128 | /// vector without lane crossings. A shuffle using this mask is not |
| 2129 | /// necessarily a no-op because it may change the number of elements from its |
| 2130 | /// input vectors or it may provide demanded bits knowledge via undef lanes. |
| 2131 | /// Example: <undef,undef,2,3> |
| 2132 | static bool isIdentityMask(ArrayRef<int> Mask); |
| 2133 | static bool isIdentityMask(const Constant *Mask) { |
| 2134 | assert(Mask->getType()->isVectorTy() && "Shuffle needs vector constant.")((void)0); |
| 2135 | SmallVector<int, 16> MaskAsInts; |
| 2136 | getShuffleMask(Mask, MaskAsInts); |
| 2137 | return isIdentityMask(MaskAsInts); |
| 2138 | } |
| 2139 | |
| 2140 | /// Return true if this shuffle chooses elements from exactly one source |
| 2141 | /// vector without lane crossings and does not change the number of elements |
| 2142 | /// from its input vectors. |
| 2143 | /// Example: shufflevector <4 x n> A, <4 x n> B, <4,undef,6,undef> |
| 2144 | bool isIdentity() const { |
| 2145 | return !changesLength() && isIdentityMask(ShuffleMask); |
| 2146 | } |
| 2147 | |
| 2148 | /// Return true if this shuffle lengthens exactly one source vector with |
| 2149 | /// undefs in the high elements. |
| 2150 | bool isIdentityWithPadding() const; |
| 2151 | |
| 2152 | /// Return true if this shuffle extracts the first N elements of exactly one |
| 2153 | /// source vector. |
| 2154 | bool isIdentityWithExtract() const; |
| 2155 | |
| 2156 | /// Return true if this shuffle concatenates its 2 source vectors. This |
| 2157 | /// returns false if either input is undefined. In that case, the shuffle is |
| 2158 | /// is better classified as an identity with padding operation. |
| 2159 | bool isConcat() const; |
| 2160 | |
| 2161 | /// Return true if this shuffle mask chooses elements from its source vectors |
| 2162 | /// without lane crossings. A shuffle using this mask would be |
| 2163 | /// equivalent to a vector select with a constant condition operand. |
| 2164 | /// Example: <4,1,6,undef> |
| 2165 | /// This returns false if the mask does not choose from both input vectors. |
| 2166 | /// In that case, the shuffle is better classified as an identity shuffle. |
| 2167 | /// This assumes that vector operands are the same length as the mask |
| 2168 | /// (a length-changing shuffle can never be equivalent to a vector select). |
| 2169 | static bool isSelectMask(ArrayRef<int> Mask); |
| 2170 | static bool isSelectMask(const Constant *Mask) { |
| 2171 | assert(Mask->getType()->isVectorTy() && "Shuffle needs vector constant.")((void)0); |
| 2172 | SmallVector<int, 16> MaskAsInts; |
| 2173 | getShuffleMask(Mask, MaskAsInts); |
| 2174 | return isSelectMask(MaskAsInts); |
| 2175 | } |
| 2176 | |
| 2177 | /// Return true if this shuffle chooses elements from its source vectors |
| 2178 | /// without lane crossings and all operands have the same number of elements. |
| 2179 | /// In other words, this shuffle is equivalent to a vector select with a |
| 2180 | /// constant condition operand. |
| 2181 | /// Example: shufflevector <4 x n> A, <4 x n> B, <undef,1,6,3> |
| 2182 | /// This returns false if the mask does not choose from both input vectors. |
| 2183 | /// In that case, the shuffle is better classified as an identity shuffle. |
| 2184 | /// TODO: Optionally allow length-changing shuffles. |
| 2185 | bool isSelect() const { |
| 2186 | return !changesLength() && isSelectMask(ShuffleMask); |
| 2187 | } |
| 2188 | |
| 2189 | /// Return true if this shuffle mask swaps the order of elements from exactly |
| 2190 | /// one source vector. |
| 2191 | /// Example: <7,6,undef,4> |
| 2192 | /// This assumes that vector operands are the same length as the mask. |
| 2193 | static bool isReverseMask(ArrayRef<int> Mask); |
| 2194 | static bool isReverseMask(const Constant *Mask) { |
| 2195 | assert(Mask->getType()->isVectorTy() && "Shuffle needs vector constant.")((void)0); |
| 2196 | SmallVector<int, 16> MaskAsInts; |
| 2197 | getShuffleMask(Mask, MaskAsInts); |
| 2198 | return isReverseMask(MaskAsInts); |
| 2199 | } |
| 2200 | |
| 2201 | /// Return true if this shuffle swaps the order of elements from exactly |
| 2202 | /// one source vector. |
| 2203 | /// Example: shufflevector <4 x n> A, <4 x n> B, <3,undef,1,undef> |
| 2204 | /// TODO: Optionally allow length-changing shuffles. |
| 2205 | bool isReverse() const { |
| 2206 | return !changesLength() && isReverseMask(ShuffleMask); |
| 2207 | } |
| 2208 | |
| 2209 | /// Return true if this shuffle mask chooses all elements with the same value |
| 2210 | /// as the first element of exactly one source vector. |
| 2211 | /// Example: <4,undef,undef,4> |
| 2212 | /// This assumes that vector operands are the same length as the mask. |
| 2213 | static bool isZeroEltSplatMask(ArrayRef<int> Mask); |
| 2214 | static bool isZeroEltSplatMask(const Constant *Mask) { |
| 2215 | assert(Mask->getType()->isVectorTy() && "Shuffle needs vector constant.")((void)0); |
| 2216 | SmallVector<int, 16> MaskAsInts; |
| 2217 | getShuffleMask(Mask, MaskAsInts); |
| 2218 | return isZeroEltSplatMask(MaskAsInts); |
| 2219 | } |
| 2220 | |
| 2221 | /// Return true if all elements of this shuffle are the same value as the |
| 2222 | /// first element of exactly one source vector without changing the length |
| 2223 | /// of that vector. |
| 2224 | /// Example: shufflevector <4 x n> A, <4 x n> B, <undef,0,undef,0> |
| 2225 | /// TODO: Optionally allow length-changing shuffles. |
| 2226 | /// TODO: Optionally allow splats from other elements. |
| 2227 | bool isZeroEltSplat() const { |
| 2228 | return !changesLength() && isZeroEltSplatMask(ShuffleMask); |
| 2229 | } |
| 2230 | |
| 2231 | /// Return true if this shuffle mask is a transpose mask. |
| 2232 | /// Transpose vector masks transpose a 2xn matrix. They read corresponding |
| 2233 | /// even- or odd-numbered vector elements from two n-dimensional source |
| 2234 | /// vectors and write each result into consecutive elements of an |
| 2235 | /// n-dimensional destination vector. Two shuffles are necessary to complete |
| 2236 | /// the transpose, one for the even elements and another for the odd elements. |
| 2237 | /// This description closely follows how the TRN1 and TRN2 AArch64 |
| 2238 | /// instructions operate. |
| 2239 | /// |
| 2240 | /// For example, a simple 2x2 matrix can be transposed with: |
| 2241 | /// |
| 2242 | /// ; Original matrix |
| 2243 | /// m0 = < a, b > |
| 2244 | /// m1 = < c, d > |
| 2245 | /// |
| 2246 | /// ; Transposed matrix |
| 2247 | /// t0 = < a, c > = shufflevector m0, m1, < 0, 2 > |
| 2248 | /// t1 = < b, d > = shufflevector m0, m1, < 1, 3 > |
| 2249 | /// |
| 2250 | /// For matrices having greater than n columns, the resulting nx2 transposed |
| 2251 | /// matrix is stored in two result vectors such that one vector contains |
| 2252 | /// interleaved elements from all the even-numbered rows and the other vector |
| 2253 | /// contains interleaved elements from all the odd-numbered rows. For example, |
| 2254 | /// a 2x4 matrix can be transposed with: |
| 2255 | /// |
| 2256 | /// ; Original matrix |
| 2257 | /// m0 = < a, b, c, d > |
| 2258 | /// m1 = < e, f, g, h > |
| 2259 | /// |
| 2260 | /// ; Transposed matrix |
| 2261 | /// t0 = < a, e, c, g > = shufflevector m0, m1 < 0, 4, 2, 6 > |
| 2262 | /// t1 = < b, f, d, h > = shufflevector m0, m1 < 1, 5, 3, 7 > |
| 2263 | static bool isTransposeMask(ArrayRef<int> Mask); |
| 2264 | static bool isTransposeMask(const Constant *Mask) { |
| 2265 | assert(Mask->getType()->isVectorTy() && "Shuffle needs vector constant.")((void)0); |
| 2266 | SmallVector<int, 16> MaskAsInts; |
| 2267 | getShuffleMask(Mask, MaskAsInts); |
| 2268 | return isTransposeMask(MaskAsInts); |
| 2269 | } |
| 2270 | |
| 2271 | /// Return true if this shuffle transposes the elements of its inputs without |
| 2272 | /// changing the length of the vectors. This operation may also be known as a |
| 2273 | /// merge or interleave. See the description for isTransposeMask() for the |
| 2274 | /// exact specification. |
| 2275 | /// Example: shufflevector <4 x n> A, <4 x n> B, <0,4,2,6> |
| 2276 | bool isTranspose() const { |
| 2277 | return !changesLength() && isTransposeMask(ShuffleMask); |
| 2278 | } |
| 2279 | |
| 2280 | /// Return true if this shuffle mask is an extract subvector mask. |
| 2281 | /// A valid extract subvector mask returns a smaller vector from a single |
| 2282 | /// source operand. The base extraction index is returned as well. |
| 2283 | static bool isExtractSubvectorMask(ArrayRef<int> Mask, int NumSrcElts, |
| 2284 | int &Index); |
| 2285 | static bool isExtractSubvectorMask(const Constant *Mask, int NumSrcElts, |
| 2286 | int &Index) { |
| 2287 | assert(Mask->getType()->isVectorTy() && "Shuffle needs vector constant.")((void)0); |
| 2288 | // Not possible to express a shuffle mask for a scalable vector for this |
| 2289 | // case. |
| 2290 | if (isa<ScalableVectorType>(Mask->getType())) |
| 2291 | return false; |
| 2292 | SmallVector<int, 16> MaskAsInts; |
| 2293 | getShuffleMask(Mask, MaskAsInts); |
| 2294 | return isExtractSubvectorMask(MaskAsInts, NumSrcElts, Index); |
| 2295 | } |
| 2296 | |
| 2297 | /// Return true if this shuffle mask is an extract subvector mask. |
| 2298 | bool isExtractSubvectorMask(int &Index) const { |
| 2299 | // Not possible to express a shuffle mask for a scalable vector for this |
| 2300 | // case. |
| 2301 | if (isa<ScalableVectorType>(getType())) |
| 2302 | return false; |
| 2303 | |
| 2304 | int NumSrcElts = |
| 2305 | cast<FixedVectorType>(Op<0>()->getType())->getNumElements(); |
| 2306 | return isExtractSubvectorMask(ShuffleMask, NumSrcElts, Index); |
| 2307 | } |
| 2308 | |
| 2309 | /// Change values in a shuffle permute mask assuming the two vector operands |
| 2310 | /// of length InVecNumElts have swapped position. |
| 2311 | static void commuteShuffleMask(MutableArrayRef<int> Mask, |
| 2312 | unsigned InVecNumElts) { |
| 2313 | for (int &Idx : Mask) { |
| 2314 | if (Idx == -1) |
| 2315 | continue; |
| 2316 | Idx = Idx < (int)InVecNumElts ? Idx + InVecNumElts : Idx - InVecNumElts; |
| 2317 | assert(Idx >= 0 && Idx < (int)InVecNumElts * 2 &&((void)0) |
| 2318 | "shufflevector mask index out of range")((void)0); |
| 2319 | } |
| 2320 | } |
| 2321 | |
| 2322 | // Methods for support type inquiry through isa, cast, and dyn_cast: |
| 2323 | static bool classof(const Instruction *I) { |
| 2324 | return I->getOpcode() == Instruction::ShuffleVector; |
| 2325 | } |
| 2326 | static bool classof(const Value *V) { |
| 2327 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
| 2328 | } |
| 2329 | }; |
| 2330 | |
| 2331 | template <> |
| 2332 | struct OperandTraits<ShuffleVectorInst> |
| 2333 | : public FixedNumOperandTraits<ShuffleVectorInst, 2> {}; |
| 2334 | |
| 2335 | DEFINE_TRANSPARENT_OPERAND_ACCESSORS(ShuffleVectorInst, Value)ShuffleVectorInst::op_iterator ShuffleVectorInst::op_begin() { return OperandTraits<ShuffleVectorInst>::op_begin(this ); } ShuffleVectorInst::const_op_iterator ShuffleVectorInst:: op_begin() const { return OperandTraits<ShuffleVectorInst> ::op_begin(const_cast<ShuffleVectorInst*>(this)); } ShuffleVectorInst ::op_iterator ShuffleVectorInst::op_end() { return OperandTraits <ShuffleVectorInst>::op_end(this); } ShuffleVectorInst:: const_op_iterator ShuffleVectorInst::op_end() const { return OperandTraits <ShuffleVectorInst>::op_end(const_cast<ShuffleVectorInst *>(this)); } Value *ShuffleVectorInst::getOperand(unsigned i_nocapture) const { ((void)0); return cast_or_null<Value >( OperandTraits<ShuffleVectorInst>::op_begin(const_cast <ShuffleVectorInst*>(this))[i_nocapture].get()); } void ShuffleVectorInst::setOperand(unsigned i_nocapture, Value *Val_nocapture ) { ((void)0); OperandTraits<ShuffleVectorInst>::op_begin (this)[i_nocapture] = Val_nocapture; } unsigned ShuffleVectorInst ::getNumOperands() const { return OperandTraits<ShuffleVectorInst >::operands(this); } template <int Idx_nocapture> Use &ShuffleVectorInst::Op() { return this->OpFrom<Idx_nocapture >(this); } template <int Idx_nocapture> const Use & ShuffleVectorInst::Op() const { return this->OpFrom<Idx_nocapture >(this); } |
| 2336 | |
| 2337 | //===----------------------------------------------------------------------===// |
| 2338 | // ExtractValueInst Class |
| 2339 | //===----------------------------------------------------------------------===// |
| 2340 | |
| 2341 | /// This instruction extracts a struct member or array |
| 2342 | /// element value from an aggregate value. |
| 2343 | /// |
| 2344 | class ExtractValueInst : public UnaryInstruction { |
| 2345 | SmallVector<unsigned, 4> Indices; |
| 2346 | |
| 2347 | ExtractValueInst(const ExtractValueInst &EVI); |
| 2348 | |
| 2349 | /// Constructors - Create a extractvalue instruction with a base aggregate |
| 2350 | /// value and a list of indices. The first ctor can optionally insert before |
| 2351 | /// an existing instruction, the second appends the new instruction to the |
| 2352 | /// specified BasicBlock. |
| 2353 | inline ExtractValueInst(Value *Agg, |
| 2354 | ArrayRef<unsigned> Idxs, |
| 2355 | const Twine &NameStr, |
| 2356 | Instruction *InsertBefore); |
| 2357 | inline ExtractValueInst(Value *Agg, |
| 2358 | ArrayRef<unsigned> Idxs, |
| 2359 | const Twine &NameStr, BasicBlock *InsertAtEnd); |
| 2360 | |
| 2361 | void init(ArrayRef<unsigned> Idxs, const Twine &NameStr); |
| 2362 | |
| 2363 | protected: |
| 2364 | // Note: Instruction needs to be a friend here to call cloneImpl. |
| 2365 | friend class Instruction; |
| 2366 | |
| 2367 | ExtractValueInst *cloneImpl() const; |
| 2368 | |
| 2369 | public: |
| 2370 | static ExtractValueInst *Create(Value *Agg, |
| 2371 | ArrayRef<unsigned> Idxs, |
| 2372 | const Twine &NameStr = "", |
| 2373 | Instruction *InsertBefore = nullptr) { |
| 2374 | return new |
| 2375 | ExtractValueInst(Agg, Idxs, NameStr, InsertBefore); |
| 2376 | } |
| 2377 | |
| 2378 | static ExtractValueInst *Create(Value *Agg, |
| 2379 | ArrayRef<unsigned> Idxs, |
| 2380 | const Twine &NameStr, |
| 2381 | BasicBlock *InsertAtEnd) { |
| 2382 | return new ExtractValueInst(Agg, Idxs, NameStr, InsertAtEnd); |
| 2383 | } |
| 2384 | |
| 2385 | /// Returns the type of the element that would be extracted |
| 2386 | /// with an extractvalue instruction with the specified parameters. |
| 2387 | /// |
| 2388 | /// Null is returned if the indices are invalid for the specified type. |
| 2389 | static Type *getIndexedType(Type *Agg, ArrayRef<unsigned> Idxs); |
| 2390 | |
| 2391 | using idx_iterator = const unsigned*; |
| 2392 | |
| 2393 | inline idx_iterator idx_begin() const { return Indices.begin(); } |
| 2394 | inline idx_iterator idx_end() const { return Indices.end(); } |
| 2395 | inline iterator_range<idx_iterator> indices() const { |
| 2396 | return make_range(idx_begin(), idx_end()); |
| 2397 | } |
| 2398 | |
| 2399 | Value *getAggregateOperand() { |
| 2400 | return getOperand(0); |
| 2401 | } |
| 2402 | const Value *getAggregateOperand() const { |
| 2403 | return getOperand(0); |
| 2404 | } |
| 2405 | static unsigned getAggregateOperandIndex() { |
| 2406 | return 0U; // get index for modifying correct operand |
| 2407 | } |
| 2408 | |
| 2409 | ArrayRef<unsigned> getIndices() const { |
| 2410 | return Indices; |
| 2411 | } |
| 2412 | |
| 2413 | unsigned getNumIndices() const { |
| 2414 | return (unsigned)Indices.size(); |
| 2415 | } |
| 2416 | |
| 2417 | bool hasIndices() const { |
| 2418 | return true; |
| 2419 | } |
| 2420 | |
| 2421 | // Methods for support type inquiry through isa, cast, and dyn_cast: |
| 2422 | static bool classof(const Instruction *I) { |
| 2423 | return I->getOpcode() == Instruction::ExtractValue; |
| 2424 | } |
| 2425 | static bool classof(const Value *V) { |
| 2426 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
| 2427 | } |
| 2428 | }; |
| 2429 | |
| 2430 | ExtractValueInst::ExtractValueInst(Value *Agg, |
| 2431 | ArrayRef<unsigned> Idxs, |
| 2432 | const Twine &NameStr, |
| 2433 | Instruction *InsertBefore) |
| 2434 | : UnaryInstruction(checkGEPType(getIndexedType(Agg->getType(), Idxs)), |
| 2435 | ExtractValue, Agg, InsertBefore) { |
| 2436 | init(Idxs, NameStr); |
| 2437 | } |
| 2438 | |
| 2439 | ExtractValueInst::ExtractValueInst(Value *Agg, |
| 2440 | ArrayRef<unsigned> Idxs, |
| 2441 | const Twine &NameStr, |
| 2442 | BasicBlock *InsertAtEnd) |
| 2443 | : UnaryInstruction(checkGEPType(getIndexedType(Agg->getType(), Idxs)), |
| 2444 | ExtractValue, Agg, InsertAtEnd) { |
| 2445 | init(Idxs, NameStr); |
| 2446 | } |
| 2447 | |
| 2448 | //===----------------------------------------------------------------------===// |
| 2449 | // InsertValueInst Class |
| 2450 | //===----------------------------------------------------------------------===// |
| 2451 | |
| 2452 | /// This instruction inserts a struct field of array element |
| 2453 | /// value into an aggregate value. |
| 2454 | /// |
| 2455 | class InsertValueInst : public Instruction { |
| 2456 | SmallVector<unsigned, 4> Indices; |
| 2457 | |
| 2458 | InsertValueInst(const InsertValueInst &IVI); |
| 2459 | |
| 2460 | /// Constructors - Create a insertvalue instruction with a base aggregate |
| 2461 | /// value, a value to insert, and a list of indices. The first ctor can |
| 2462 | /// optionally insert before an existing instruction, the second appends |
| 2463 | /// the new instruction to the specified BasicBlock. |
| 2464 | inline InsertValueInst(Value *Agg, Value *Val, |
| 2465 | ArrayRef<unsigned> Idxs, |
| 2466 | const Twine &NameStr, |
| 2467 | Instruction *InsertBefore); |
| 2468 | inline InsertValueInst(Value *Agg, Value *Val, |
| 2469 | ArrayRef<unsigned> Idxs, |
| 2470 | const Twine &NameStr, BasicBlock *InsertAtEnd); |
| 2471 | |
| 2472 | /// Constructors - These two constructors are convenience methods because one |
| 2473 | /// and two index insertvalue instructions are so common. |
| 2474 | InsertValueInst(Value *Agg, Value *Val, unsigned Idx, |
| 2475 | const Twine &NameStr = "", |
| 2476 | Instruction *InsertBefore = nullptr); |
| 2477 | InsertValueInst(Value *Agg, Value *Val, unsigned Idx, const Twine &NameStr, |
| 2478 | BasicBlock *InsertAtEnd); |
| 2479 | |
| 2480 | void init(Value *Agg, Value *Val, ArrayRef<unsigned> Idxs, |
| 2481 | const Twine &NameStr); |
| 2482 | |
| 2483 | protected: |
| 2484 | // Note: Instruction needs to be a friend here to call cloneImpl. |
| 2485 | friend class Instruction; |
| 2486 | |
| 2487 | InsertValueInst *cloneImpl() const; |
| 2488 | |
| 2489 | public: |
| 2490 | // allocate space for exactly two operands |
| 2491 | void *operator new(size_t S) { return User::operator new(S, 2); } |
| 2492 | void operator delete(void *Ptr) { User::operator delete(Ptr); } |
| 2493 | |
| 2494 | static InsertValueInst *Create(Value *Agg, Value *Val, |
| 2495 | ArrayRef<unsigned> Idxs, |
| 2496 | const Twine &NameStr = "", |
| 2497 | Instruction *InsertBefore = nullptr) { |
| 2498 | return new InsertValueInst(Agg, Val, Idxs, NameStr, InsertBefore); |
| 2499 | } |
| 2500 | |
| 2501 | static InsertValueInst *Create(Value *Agg, Value *Val, |
| 2502 | ArrayRef<unsigned> Idxs, |
| 2503 | const Twine &NameStr, |
| 2504 | BasicBlock *InsertAtEnd) { |
| 2505 | return new InsertValueInst(Agg, Val, Idxs, NameStr, InsertAtEnd); |
| 2506 | } |
| 2507 | |
| 2508 | /// Transparently provide more efficient getOperand methods. |
| 2509 | 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; |
| 2510 | |
| 2511 | using idx_iterator = const unsigned*; |
| 2512 | |
| 2513 | inline idx_iterator idx_begin() const { return Indices.begin(); } |
| 2514 | inline idx_iterator idx_end() const { return Indices.end(); } |
| 2515 | inline iterator_range<idx_iterator> indices() const { |
| 2516 | return make_range(idx_begin(), idx_end()); |
| 2517 | } |
| 2518 | |
| 2519 | Value *getAggregateOperand() { |
| 2520 | return getOperand(0); |
| 2521 | } |
| 2522 | const Value *getAggregateOperand() const { |
| 2523 | return getOperand(0); |
| 2524 | } |
| 2525 | static unsigned getAggregateOperandIndex() { |
| 2526 | return 0U; // get index for modifying correct operand |
| 2527 | } |
| 2528 | |
| 2529 | Value *getInsertedValueOperand() { |
| 2530 | return getOperand(1); |
| 2531 | } |
| 2532 | const Value *getInsertedValueOperand() const { |
| 2533 | return getOperand(1); |
| 2534 | } |
| 2535 | static unsigned getInsertedValueOperandIndex() { |
| 2536 | return 1U; // get index for modifying correct operand |
| 2537 | } |
| 2538 | |
| 2539 | ArrayRef<unsigned> getIndices() const { |
| 2540 | return Indices; |
| 2541 | } |
| 2542 | |
| 2543 | unsigned getNumIndices() const { |
| 2544 | return (unsigned)Indices.size(); |
| 2545 | } |
| 2546 | |
| 2547 | bool hasIndices() const { |
| 2548 | return true; |
| 2549 | } |
| 2550 | |
| 2551 | // Methods for support type inquiry through isa, cast, and dyn_cast: |
| 2552 | static bool classof(const Instruction *I) { |
| 2553 | return I->getOpcode() == Instruction::InsertValue; |
| 2554 | } |
| 2555 | static bool classof(const Value *V) { |
| 2556 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
| 2557 | } |
| 2558 | }; |
| 2559 | |
| 2560 | template <> |
| 2561 | struct OperandTraits<InsertValueInst> : |
| 2562 | public FixedNumOperandTraits<InsertValueInst, 2> { |
| 2563 | }; |
| 2564 | |
| 2565 | InsertValueInst::InsertValueInst(Value *Agg, |
| 2566 | Value *Val, |
| 2567 | ArrayRef<unsigned> Idxs, |
| 2568 | const Twine &NameStr, |
| 2569 | Instruction *InsertBefore) |
| 2570 | : Instruction(Agg->getType(), InsertValue, |
| 2571 | OperandTraits<InsertValueInst>::op_begin(this), |
| 2572 | 2, InsertBefore) { |
| 2573 | init(Agg, Val, Idxs, NameStr); |
| 2574 | } |
| 2575 | |
| 2576 | InsertValueInst::InsertValueInst(Value *Agg, |
| 2577 | Value *Val, |
| 2578 | ArrayRef<unsigned> Idxs, |
| 2579 | const Twine &NameStr, |
| 2580 | BasicBlock *InsertAtEnd) |
| 2581 | : Instruction(Agg->getType(), InsertValue, |
| 2582 | OperandTraits<InsertValueInst>::op_begin(this), |
| 2583 | 2, InsertAtEnd) { |
| 2584 | init(Agg, Val, Idxs, NameStr); |
| 2585 | } |
| 2586 | |
| 2587 | DEFINE_TRANSPARENT_OPERAND_ACCESSORS(InsertValueInst, Value)InsertValueInst::op_iterator InsertValueInst::op_begin() { return OperandTraits<InsertValueInst>::op_begin(this); } InsertValueInst ::const_op_iterator InsertValueInst::op_begin() const { return OperandTraits<InsertValueInst>::op_begin(const_cast< InsertValueInst*>(this)); } InsertValueInst::op_iterator InsertValueInst ::op_end() { return OperandTraits<InsertValueInst>::op_end (this); } InsertValueInst::const_op_iterator InsertValueInst:: op_end() const { return OperandTraits<InsertValueInst>:: op_end(const_cast<InsertValueInst*>(this)); } Value *InsertValueInst ::getOperand(unsigned i_nocapture) const { ((void)0); return cast_or_null <Value>( OperandTraits<InsertValueInst>::op_begin (const_cast<InsertValueInst*>(this))[i_nocapture].get() ); } void InsertValueInst::setOperand(unsigned i_nocapture, Value *Val_nocapture) { ((void)0); OperandTraits<InsertValueInst >::op_begin(this)[i_nocapture] = Val_nocapture; } unsigned InsertValueInst::getNumOperands() const { return OperandTraits <InsertValueInst>::operands(this); } template <int Idx_nocapture > Use &InsertValueInst::Op() { return this->OpFrom< Idx_nocapture>(this); } template <int Idx_nocapture> const Use &InsertValueInst::Op() const { return this-> OpFrom<Idx_nocapture>(this); } |
| 2588 | |
| 2589 | //===----------------------------------------------------------------------===// |
| 2590 | // PHINode Class |
| 2591 | //===----------------------------------------------------------------------===// |
| 2592 | |
| 2593 | // PHINode - The PHINode class is used to represent the magical mystical PHI |
| 2594 | // node, that can not exist in nature, but can be synthesized in a computer |
| 2595 | // scientist's overactive imagination. |
| 2596 | // |
| 2597 | class PHINode : public Instruction { |
| 2598 | /// The number of operands actually allocated. NumOperands is |
| 2599 | /// the number actually in use. |
| 2600 | unsigned ReservedSpace; |
| 2601 | |
| 2602 | PHINode(const PHINode &PN); |
| 2603 | |
| 2604 | explicit PHINode(Type *Ty, unsigned NumReservedValues, |
| 2605 | const Twine &NameStr = "", |
| 2606 | Instruction *InsertBefore = nullptr) |
| 2607 | : Instruction(Ty, Instruction::PHI, nullptr, 0, InsertBefore), |
| 2608 | ReservedSpace(NumReservedValues) { |
| 2609 | assert(!Ty->isTokenTy() && "PHI nodes cannot have token type!")((void)0); |
| 2610 | setName(NameStr); |
| 2611 | allocHungoffUses(ReservedSpace); |
| 2612 | } |
| 2613 | |
| 2614 | PHINode(Type *Ty, unsigned NumReservedValues, const Twine &NameStr, |
| 2615 | BasicBlock *InsertAtEnd) |
| 2616 | : Instruction(Ty, Instruction::PHI, nullptr, 0, InsertAtEnd), |
| 2617 | ReservedSpace(NumReservedValues) { |
| 2618 | assert(!Ty->isTokenTy() && "PHI nodes cannot have token type!")((void)0); |
| 2619 | setName(NameStr); |
| 2620 | allocHungoffUses(ReservedSpace); |
| 2621 | } |
| 2622 | |
| 2623 | protected: |
| 2624 | // Note: Instruction needs to be a friend here to call cloneImpl. |
| 2625 | friend class Instruction; |
| 2626 | |
| 2627 | PHINode *cloneImpl() const; |
| 2628 | |
| 2629 | // allocHungoffUses - this is more complicated than the generic |
| 2630 | // User::allocHungoffUses, because we have to allocate Uses for the incoming |
| 2631 | // values and pointers to the incoming blocks, all in one allocation. |
| 2632 | void allocHungoffUses(unsigned N) { |
| 2633 | User::allocHungoffUses(N, /* IsPhi */ true); |
| 2634 | } |
| 2635 | |
| 2636 | public: |
| 2637 | /// Constructors - NumReservedValues is a hint for the number of incoming |
| 2638 | /// edges that this phi node will have (use 0 if you really have no idea). |
| 2639 | static PHINode *Create(Type *Ty, unsigned NumReservedValues, |
| 2640 | const Twine &NameStr = "", |
| 2641 | Instruction *InsertBefore = nullptr) { |
| 2642 | return new PHINode(Ty, NumReservedValues, NameStr, InsertBefore); |
| 2643 | } |
| 2644 | |
| 2645 | static PHINode *Create(Type *Ty, unsigned NumReservedValues, |
| 2646 | const Twine &NameStr, BasicBlock *InsertAtEnd) { |
| 2647 | return new PHINode(Ty, NumReservedValues, NameStr, InsertAtEnd); |
| 2648 | } |
| 2649 | |
| 2650 | /// Provide fast operand accessors |
| 2651 | 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; |
| 2652 | |
| 2653 | // Block iterator interface. This provides access to the list of incoming |
| 2654 | // basic blocks, which parallels the list of incoming values. |
| 2655 | |
| 2656 | using block_iterator = BasicBlock **; |
| 2657 | using const_block_iterator = BasicBlock * const *; |
| 2658 | |
| 2659 | block_iterator block_begin() { |
| 2660 | return reinterpret_cast<block_iterator>(op_begin() + ReservedSpace); |
| 2661 | } |
| 2662 | |
| 2663 | const_block_iterator block_begin() const { |
| 2664 | return reinterpret_cast<const_block_iterator>(op_begin() + ReservedSpace); |
| 2665 | } |
| 2666 | |
| 2667 | block_iterator block_end() { |
| 2668 | return block_begin() + getNumOperands(); |
| 2669 | } |
| 2670 | |
| 2671 | const_block_iterator block_end() const { |
| 2672 | return block_begin() + getNumOperands(); |
| 2673 | } |
| 2674 | |
| 2675 | iterator_range<block_iterator> blocks() { |
| 2676 | return make_range(block_begin(), block_end()); |
| 2677 | } |
| 2678 | |
| 2679 | iterator_range<const_block_iterator> blocks() const { |
| 2680 | return make_range(block_begin(), block_end()); |
| 2681 | } |
| 2682 | |
| 2683 | op_range incoming_values() { return operands(); } |
| 2684 | |
| 2685 | const_op_range incoming_values() const { return operands(); } |
| 2686 | |
| 2687 | /// Return the number of incoming edges |
| 2688 | /// |
| 2689 | unsigned getNumIncomingValues() const { return getNumOperands(); } |
| 2690 | |
| 2691 | /// Return incoming value number x |
| 2692 | /// |
| 2693 | Value *getIncomingValue(unsigned i) const { |
| 2694 | return getOperand(i); |
| 2695 | } |
| 2696 | void setIncomingValue(unsigned i, Value *V) { |
| 2697 | assert(V && "PHI node got a null value!")((void)0); |
| 2698 | assert(getType() == V->getType() &&((void)0) |
| 2699 | "All operands to PHI node must be the same type as the PHI node!")((void)0); |
| 2700 | setOperand(i, V); |
| 2701 | } |
| 2702 | |
| 2703 | static unsigned getOperandNumForIncomingValue(unsigned i) { |
| 2704 | return i; |
| 2705 | } |
| 2706 | |
| 2707 | static unsigned getIncomingValueNumForOperand(unsigned i) { |
| 2708 | return i; |
| 2709 | } |
| 2710 | |
| 2711 | /// Return incoming basic block number @p i. |
| 2712 | /// |
| 2713 | BasicBlock *getIncomingBlock(unsigned i) const { |
| 2714 | return block_begin()[i]; |
| 2715 | } |
| 2716 | |
| 2717 | /// Return incoming basic block corresponding |
| 2718 | /// to an operand of the PHI. |
| 2719 | /// |
| 2720 | BasicBlock *getIncomingBlock(const Use &U) const { |
| 2721 | assert(this == U.getUser() && "Iterator doesn't point to PHI's Uses?")((void)0); |
| 2722 | return getIncomingBlock(unsigned(&U - op_begin())); |
| 2723 | } |
| 2724 | |
| 2725 | /// Return incoming basic block corresponding |
| 2726 | /// to value use iterator. |
| 2727 | /// |
| 2728 | BasicBlock *getIncomingBlock(Value::const_user_iterator I) const { |
| 2729 | return getIncomingBlock(I.getUse()); |
| 2730 | } |
| 2731 | |
| 2732 | void setIncomingBlock(unsigned i, BasicBlock *BB) { |
| 2733 | assert(BB && "PHI node got a null basic block!")((void)0); |
| 2734 | block_begin()[i] = BB; |
| 2735 | } |
| 2736 | |
| 2737 | /// Replace every incoming basic block \p Old to basic block \p New. |
| 2738 | void replaceIncomingBlockWith(const BasicBlock *Old, BasicBlock *New) { |
| 2739 | assert(New && Old && "PHI node got a null basic block!")((void)0); |
| 2740 | for (unsigned Op = 0, NumOps = getNumOperands(); Op != NumOps; ++Op) |
| 2741 | if (getIncomingBlock(Op) == Old) |
| 2742 | setIncomingBlock(Op, New); |
| 2743 | } |
| 2744 | |
| 2745 | /// Add an incoming value to the end of the PHI list |
| 2746 | /// |
| 2747 | void addIncoming(Value *V, BasicBlock *BB) { |
| 2748 | if (getNumOperands() == ReservedSpace) |
| 2749 | growOperands(); // Get more space! |
| 2750 | // Initialize some new operands. |
| 2751 | setNumHungOffUseOperands(getNumOperands() + 1); |
| 2752 | setIncomingValue(getNumOperands() - 1, V); |
| 2753 | setIncomingBlock(getNumOperands() - 1, BB); |
| 2754 | } |
| 2755 | |
| 2756 | /// Remove an incoming value. This is useful if a |
| 2757 | /// predecessor basic block is deleted. The value removed is returned. |
| 2758 | /// |
| 2759 | /// If the last incoming value for a PHI node is removed (and DeletePHIIfEmpty |
| 2760 | /// is true), the PHI node is destroyed and any uses of it are replaced with |
| 2761 | /// dummy values. The only time there should be zero incoming values to a PHI |
| 2762 | /// node is when the block is dead, so this strategy is sound. |
| 2763 | /// |
| 2764 | Value *removeIncomingValue(unsigned Idx, bool DeletePHIIfEmpty = true); |
| 2765 | |
| 2766 | Value *removeIncomingValue(const BasicBlock *BB, bool DeletePHIIfEmpty=true) { |
| 2767 | int Idx = getBasicBlockIndex(BB); |
| 2768 | assert(Idx >= 0 && "Invalid basic block argument to remove!")((void)0); |
| 2769 | return removeIncomingValue(Idx, DeletePHIIfEmpty); |
| 2770 | } |
| 2771 | |
| 2772 | /// Return the first index of the specified basic |
| 2773 | /// block in the value list for this PHI. Returns -1 if no instance. |
| 2774 | /// |
| 2775 | int getBasicBlockIndex(const BasicBlock *BB) const { |
| 2776 | for (unsigned i = 0, e = getNumOperands(); i != e; ++i) |
| 2777 | if (block_begin()[i] == BB) |
| 2778 | return i; |
| 2779 | return -1; |
| 2780 | } |
| 2781 | |
| 2782 | Value *getIncomingValueForBlock(const BasicBlock *BB) const { |
| 2783 | int Idx = getBasicBlockIndex(BB); |
| 2784 | assert(Idx >= 0 && "Invalid basic block argument!")((void)0); |
| 2785 | return getIncomingValue(Idx); |
| 2786 | } |
| 2787 | |
| 2788 | /// Set every incoming value(s) for block \p BB to \p V. |
| 2789 | void setIncomingValueForBlock(const BasicBlock *BB, Value *V) { |
| 2790 | assert(BB && "PHI node got a null basic block!")((void)0); |
| 2791 | bool Found = false; |
| 2792 | for (unsigned Op = 0, NumOps = getNumOperands(); Op != NumOps; ++Op) |
| 2793 | if (getIncomingBlock(Op) == BB) { |
| 2794 | Found = true; |
| 2795 | setIncomingValue(Op, V); |
| 2796 | } |
| 2797 | (void)Found; |
| 2798 | assert(Found && "Invalid basic block argument to set!")((void)0); |
| 2799 | } |
| 2800 | |
| 2801 | /// If the specified PHI node always merges together the |
| 2802 | /// same value, return the value, otherwise return null. |
| 2803 | Value *hasConstantValue() const; |
| 2804 | |
| 2805 | /// Whether the specified PHI node always merges |
| 2806 | /// together the same value, assuming undefs are equal to a unique |
| 2807 | /// non-undef value. |
| 2808 | bool hasConstantOrUndefValue() const; |
| 2809 | |
| 2810 | /// If the PHI node is complete which means all of its parent's predecessors |
| 2811 | /// have incoming value in this PHI, return true, otherwise return false. |
| 2812 | bool isComplete() const { |
| 2813 | return llvm::all_of(predecessors(getParent()), |
| 2814 | [this](const BasicBlock *Pred) { |
| 2815 | return getBasicBlockIndex(Pred) >= 0; |
| 2816 | }); |
| 2817 | } |
| 2818 | |
| 2819 | /// Methods for support type inquiry through isa, cast, and dyn_cast: |
| 2820 | static bool classof(const Instruction *I) { |
| 2821 | return I->getOpcode() == Instruction::PHI; |
| 2822 | } |
| 2823 | static bool classof(const Value *V) { |
| 2824 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
| 2825 | } |
| 2826 | |
| 2827 | private: |
| 2828 | void growOperands(); |
| 2829 | }; |
| 2830 | |
| 2831 | template <> |
| 2832 | struct OperandTraits<PHINode> : public HungoffOperandTraits<2> { |
| 2833 | }; |
| 2834 | |
| 2835 | DEFINE_TRANSPARENT_OPERAND_ACCESSORS(PHINode, Value)PHINode::op_iterator PHINode::op_begin() { return OperandTraits <PHINode>::op_begin(this); } PHINode::const_op_iterator PHINode::op_begin() const { return OperandTraits<PHINode> ::op_begin(const_cast<PHINode*>(this)); } PHINode::op_iterator PHINode::op_end() { return OperandTraits<PHINode>::op_end (this); } PHINode::const_op_iterator PHINode::op_end() const { return OperandTraits<PHINode>::op_end(const_cast<PHINode *>(this)); } Value *PHINode::getOperand(unsigned i_nocapture ) const { ((void)0); return cast_or_null<Value>( OperandTraits <PHINode>::op_begin(const_cast<PHINode*>(this))[i_nocapture ].get()); } void PHINode::setOperand(unsigned i_nocapture, Value *Val_nocapture) { ((void)0); OperandTraits<PHINode>::op_begin (this)[i_nocapture] = Val_nocapture; } unsigned PHINode::getNumOperands () const { return OperandTraits<PHINode>::operands(this ); } template <int Idx_nocapture> Use &PHINode::Op( ) { return this->OpFrom<Idx_nocapture>(this); } template <int Idx_nocapture> const Use &PHINode::Op() const { return this->OpFrom<Idx_nocapture>(this); } |
| 2836 | |
| 2837 | //===----------------------------------------------------------------------===// |
| 2838 | // LandingPadInst Class |
| 2839 | //===----------------------------------------------------------------------===// |
| 2840 | |
| 2841 | //===--------------------------------------------------------------------------- |
| 2842 | /// The landingpad instruction holds all of the information |
| 2843 | /// necessary to generate correct exception handling. The landingpad instruction |
| 2844 | /// cannot be moved from the top of a landing pad block, which itself is |
| 2845 | /// accessible only from the 'unwind' edge of an invoke. This uses the |
| 2846 | /// SubclassData field in Value to store whether or not the landingpad is a |
| 2847 | /// cleanup. |
| 2848 | /// |
| 2849 | class LandingPadInst : public Instruction { |
| 2850 | using CleanupField = BoolBitfieldElementT<0>; |
| 2851 | |
| 2852 | /// The number of operands actually allocated. NumOperands is |
| 2853 | /// the number actually in use. |
| 2854 | unsigned ReservedSpace; |
| 2855 | |
| 2856 | LandingPadInst(const LandingPadInst &LP); |
| 2857 | |
| 2858 | public: |
| 2859 | enum ClauseType { Catch, Filter }; |
| 2860 | |
| 2861 | private: |
| 2862 | explicit LandingPadInst(Type *RetTy, unsigned NumReservedValues, |
| 2863 | const Twine &NameStr, Instruction *InsertBefore); |
| 2864 | explicit LandingPadInst(Type *RetTy, unsigned NumReservedValues, |
| 2865 | const Twine &NameStr, BasicBlock *InsertAtEnd); |
| 2866 | |
| 2867 | // Allocate space for exactly zero operands. |
| 2868 | void *operator new(size_t S) { return User::operator new(S); } |
| 2869 | |
| 2870 | void growOperands(unsigned Size); |
| 2871 | void init(unsigned NumReservedValues, const Twine &NameStr); |
| 2872 | |
| 2873 | protected: |
| 2874 | // Note: Instruction needs to be a friend here to call cloneImpl. |
| 2875 | friend class Instruction; |
| 2876 | |
| 2877 | LandingPadInst *cloneImpl() const; |
| 2878 | |
| 2879 | public: |
| 2880 | void operator delete(void *Ptr) { User::operator delete(Ptr); } |
| 2881 | |
| 2882 | /// Constructors - NumReservedClauses is a hint for the number of incoming |
| 2883 | /// clauses that this landingpad will have (use 0 if you really have no idea). |
| 2884 | static LandingPadInst *Create(Type *RetTy, unsigned NumReservedClauses, |
| 2885 | const Twine &NameStr = "", |
| 2886 | Instruction *InsertBefore = nullptr); |
| 2887 | static LandingPadInst *Create(Type *RetTy, unsigned NumReservedClauses, |
| 2888 | const Twine &NameStr, BasicBlock *InsertAtEnd); |
| 2889 | |
| 2890 | /// Provide fast operand accessors |
| 2891 | 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; |
| 2892 | |
| 2893 | /// Return 'true' if this landingpad instruction is a |
| 2894 | /// cleanup. I.e., it should be run when unwinding even if its landing pad |
| 2895 | /// doesn't catch the exception. |
| 2896 | bool isCleanup() const { return getSubclassData<CleanupField>(); } |
| 2897 | |
| 2898 | /// Indicate that this landingpad instruction is a cleanup. |
| 2899 | void setCleanup(bool V) { setSubclassData<CleanupField>(V); } |
| 2900 | |
| 2901 | /// Add a catch or filter clause to the landing pad. |
| 2902 | void addClause(Constant *ClauseVal); |
| 2903 | |
| 2904 | /// Get the value of the clause at index Idx. Use isCatch/isFilter to |
| 2905 | /// determine what type of clause this is. |
| 2906 | Constant *getClause(unsigned Idx) const { |
| 2907 | return cast<Constant>(getOperandList()[Idx]); |
| 2908 | } |
| 2909 | |
| 2910 | /// Return 'true' if the clause and index Idx is a catch clause. |
| 2911 | bool isCatch(unsigned Idx) const { |
| 2912 | return !isa<ArrayType>(getOperandList()[Idx]->getType()); |
| 2913 | } |
| 2914 | |
| 2915 | /// Return 'true' if the clause and index Idx is a filter clause. |
| 2916 | bool isFilter(unsigned Idx) const { |
| 2917 | return isa<ArrayType>(getOperandList()[Idx]->getType()); |
| 2918 | } |
| 2919 | |
| 2920 | /// Get the number of clauses for this landing pad. |
| 2921 | unsigned getNumClauses() const { return getNumOperands(); } |
| 2922 | |
| 2923 | /// Grow the size of the operand list to accommodate the new |
| 2924 | /// number of clauses. |
| 2925 | void reserveClauses(unsigned Size) { growOperands(Size); } |
| 2926 | |
| 2927 | // Methods for support type inquiry through isa, cast, and dyn_cast: |
| 2928 | static bool classof(const Instruction *I) { |
| 2929 | return I->getOpcode() == Instruction::LandingPad; |
| 2930 | } |
| 2931 | static bool classof(const Value *V) { |
| 2932 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
| 2933 | } |
| 2934 | }; |
| 2935 | |
| 2936 | template <> |
| 2937 | struct OperandTraits<LandingPadInst> : public HungoffOperandTraits<1> { |
| 2938 | }; |
| 2939 | |
| 2940 | DEFINE_TRANSPARENT_OPERAND_ACCESSORS(LandingPadInst, Value)LandingPadInst::op_iterator LandingPadInst::op_begin() { return OperandTraits<LandingPadInst>::op_begin(this); } LandingPadInst ::const_op_iterator LandingPadInst::op_begin() const { return OperandTraits<LandingPadInst>::op_begin(const_cast< LandingPadInst*>(this)); } LandingPadInst::op_iterator LandingPadInst ::op_end() { return OperandTraits<LandingPadInst>::op_end (this); } LandingPadInst::const_op_iterator LandingPadInst::op_end () const { return OperandTraits<LandingPadInst>::op_end (const_cast<LandingPadInst*>(this)); } Value *LandingPadInst ::getOperand(unsigned i_nocapture) const { ((void)0); return cast_or_null <Value>( OperandTraits<LandingPadInst>::op_begin( const_cast<LandingPadInst*>(this))[i_nocapture].get()); } void LandingPadInst::setOperand(unsigned i_nocapture, Value *Val_nocapture) { ((void)0); OperandTraits<LandingPadInst >::op_begin(this)[i_nocapture] = Val_nocapture; } unsigned LandingPadInst::getNumOperands() const { return OperandTraits <LandingPadInst>::operands(this); } template <int Idx_nocapture > Use &LandingPadInst::Op() { return this->OpFrom< Idx_nocapture>(this); } template <int Idx_nocapture> const Use &LandingPadInst::Op() const { return this-> OpFrom<Idx_nocapture>(this); } |
| 2941 | |
| 2942 | //===----------------------------------------------------------------------===// |
| 2943 | // ReturnInst Class |
| 2944 | //===----------------------------------------------------------------------===// |
| 2945 | |
| 2946 | //===--------------------------------------------------------------------------- |
| 2947 | /// Return a value (possibly void), from a function. Execution |
| 2948 | /// does not continue in this function any longer. |
| 2949 | /// |
| 2950 | class ReturnInst : public Instruction { |
| 2951 | ReturnInst(const ReturnInst &RI); |
| 2952 | |
| 2953 | private: |
| 2954 | // ReturnInst constructors: |
| 2955 | // ReturnInst() - 'ret void' instruction |
| 2956 | // ReturnInst( null) - 'ret void' instruction |
| 2957 | // ReturnInst(Value* X) - 'ret X' instruction |
| 2958 | // ReturnInst( null, Inst *I) - 'ret void' instruction, insert before I |
| 2959 | // ReturnInst(Value* X, Inst *I) - 'ret X' instruction, insert before I |
| 2960 | // ReturnInst( null, BB *B) - 'ret void' instruction, insert @ end of B |
| 2961 | // ReturnInst(Value* X, BB *B) - 'ret X' instruction, insert @ end of B |
| 2962 | // |
| 2963 | // NOTE: If the Value* passed is of type void then the constructor behaves as |
| 2964 | // if it was passed NULL. |
| 2965 | explicit ReturnInst(LLVMContext &C, Value *retVal = nullptr, |
| 2966 | Instruction *InsertBefore = nullptr); |
| 2967 | ReturnInst(LLVMContext &C, Value *retVal, BasicBlock *InsertAtEnd); |
| 2968 | explicit ReturnInst(LLVMContext &C, BasicBlock *InsertAtEnd); |
| 2969 | |
| 2970 | protected: |
| 2971 | // Note: Instruction needs to be a friend here to call cloneImpl. |
| 2972 | friend class Instruction; |
| 2973 | |
| 2974 | ReturnInst *cloneImpl() const; |
| 2975 | |
| 2976 | public: |
| 2977 | static ReturnInst* Create(LLVMContext &C, Value *retVal = nullptr, |
| 2978 | Instruction *InsertBefore = nullptr) { |
| 2979 | return new(!!retVal) ReturnInst(C, retVal, InsertBefore); |
| 2980 | } |
| 2981 | |
| 2982 | static ReturnInst* Create(LLVMContext &C, Value *retVal, |
| 2983 | BasicBlock *InsertAtEnd) { |
| 2984 | return new(!!retVal) ReturnInst(C, retVal, InsertAtEnd); |
| 2985 | } |
| 2986 | |
| 2987 | static ReturnInst* Create(LLVMContext &C, BasicBlock *InsertAtEnd) { |
| 2988 | return new(0) ReturnInst(C, InsertAtEnd); |
| 2989 | } |
| 2990 | |
| 2991 | /// Provide fast operand accessors |
| 2992 | 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; |
| 2993 | |
| 2994 | /// Convenience accessor. Returns null if there is no return value. |
| 2995 | Value *getReturnValue() const { |
| 2996 | return getNumOperands() != 0 ? getOperand(0) : nullptr; |
| 2997 | } |
| 2998 | |
| 2999 | unsigned getNumSuccessors() const { return 0; } |
| 3000 | |
| 3001 | // Methods for support type inquiry through isa, cast, and dyn_cast: |
| 3002 | static bool classof(const Instruction *I) { |
| 3003 | return (I->getOpcode() == Instruction::Ret); |
| 3004 | } |
| 3005 | static bool classof(const Value *V) { |
| 3006 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); |
| 3007 | } |
| 3008 | |
| 3009 | private: |
| 3010 | BasicBlock *getSuccessor(unsigned idx) const { |
| 3011 | llvm_unreachable("ReturnInst has no successors!")__builtin_unreachable(); |
| 3012 | } |
| 3013 | |
| 3014 | void setSuccessor(unsigned idx, BasicBlock *B) { |
| 3015 | llvm_unreachable("ReturnInst has no successors!")__builtin_unreachable(); |
| 3016 | } |
| 3017 | }; |
| 3018 | |
| 3019 | template <> |
| 3020 | struct OperandTraits<ReturnInst> : public VariadicOperandTraits<ReturnInst> { |
| 3021 | }; |
| 3022 | |
| 3023 | DEFINE_TRANSPARENT_OPERAND_ACCESSORS(ReturnInst, Value)ReturnInst::op_iterator ReturnInst::op_begin() { return OperandTraits <ReturnInst>::op_begin(this); } ReturnInst::const_op_iterator ReturnInst::op_begin() const { return OperandTraits<ReturnInst >::op_begin(const_cast<ReturnInst*>(this)); } ReturnInst ::op_iterator ReturnInst::op_end() { return OperandTraits< ReturnInst>::op_end(this); } ReturnInst::const_op_iterator ReturnInst::op_end() const { return OperandTraits<ReturnInst >::op_end(const_cast<ReturnInst*>(this)); } Value *ReturnInst ::getOperand(unsigned i_nocapture) const { ((void)0); return cast_or_null <Value>( OperandTraits<ReturnInst>::op_begin(const_cast <ReturnInst*>(this))[i_nocapture].get()); } void ReturnInst ::setOperand(unsigned i_nocapture, Value *Val_nocapture) { (( void)0); OperandTraits<ReturnInst>::op_begin(this)[i_nocapture ] = Val_nocapture; } unsigned ReturnInst::getNumOperands() const { return OperandTraits<ReturnInst>::operands(this); } template <int Idx_nocapture> Use &ReturnInst::Op() { return this->OpFrom<Idx_nocapture>(this); } template <int Idx_nocapture> const Use &ReturnInst::Op() const { return this->OpFrom<Idx_nocapture>(this); } |
| 3024 | |
| 3025 | //===----------------------------------------------------------------------===// |
| 3026 | // BranchInst Class |
| 3027 | //===----------------------------------------------------------------------===// |
| 3028 | |
| 3029 | //===--------------------------------------------------------------------------- |
| 3030 | /// Conditional or Unconditional Branch instruction. |
| 3031 | /// |
| 3032 | class BranchInst : public Instruction { |
| 3033 | /// Ops list - Branches are strange. The operands are ordered: |
| 3034 | /// [Cond, FalseDest,] TrueDest. This makes some accessors faster because |
| 3035 | /// they don't have to check for cond/uncond branchness. These are mostly |
| 3036 | /// accessed relative from op_end(). |
| 3037 | BranchInst(const BranchInst &BI); |
| 3038 | // BranchInst constructors (where {B, T, F} are blocks, and C is a condition): |
| 3039 | // BranchInst(BB *B) - 'br B' |
| 3040 | // BranchInst(BB* T, BB *F, Value *C) - 'br C, T, F' |
| 3041 | // BranchInst(BB* B, Inst *I) - 'br B' insert before I |
| 3042 | // BranchInst(BB* T, BB *F, Value *C, Inst *I) - 'br C, T, F', insert before I |
| 3043 | // BranchInst(BB* B, BB *I) - 'br B' insert at end |
| 3044 | // BranchInst(BB* T, BB *F, Value *C, BB *I) - 'br C, T, F', insert at end |
| 3045 | explicit BranchInst(BasicBlock *IfTrue, Instruction *InsertBefore = nullptr); |
| 3046 | BranchInst(BasicBlock *IfTrue, BasicBlock *IfFalse, Value *Cond, |
| 3047 | Instruction *InsertBefore = nullptr); |
| 3048 | BranchInst(BasicBlock *IfTrue, BasicBlock *InsertAtEnd); |
| 3049 | BranchInst(BasicBlock *IfTrue, BasicBlock *IfFalse, Value *Cond, |
| 3050 | BasicBlock *InsertAtEnd); |
| 3051 | |
| 3052 | void AssertOK(); |
| 3053 | |
| 3054 | protected: |
| 3055 | // Note: Instruction needs to be a friend here to call cloneImpl. |
| 3056 | friend class Instruction; |
| 3057 | |
| 3058 | BranchInst *cloneImpl() const; |
| 3059 | |
| 3060 | public: |
| 3061 | /// Iterator type that casts an operand to a basic block. |
| 3062 | /// |
| 3063 | /// This only makes sense because the successors are stored as adjacent |
| 3064 | /// operands for branch instructions. |
| 3065 | struct succ_op_iterator |
| 3066 | : iterator_adaptor_base<succ_op_iterator, value_op_iterator, |
| 3067 | std::random_access_iterator_tag, BasicBlock *, |
| 3068 | ptrdiff_t, BasicBlock *, BasicBlock *> { |
| 3069 | explicit succ_op_iterator(value_op_iterator I) : iterator_adaptor_base(I) {} |
| 3070 | |
| 3071 | BasicBlock *operator*() const { return cast<BasicBlock>(*I); } |
| 3072 | BasicBlock *operator->() const { return operator*(); } |
| 3073 | }; |
| 3074 | |
| 3075 | /// The const version of `succ_op_iterator`. |
| 3076 | struct const_succ_op_iterator |
| 3077 | : iterator_adaptor_base<const_succ_op_iterator, const_value_op_iterator, |
| 3078 | std::random_access_iterator_tag, |
| 3079 | const BasicBlock *, ptrdiff_t, const BasicBlock *, |
| 3080 | const BasicBlock *> { |
| 3081 | explicit const_succ_op_iterator(const_value_op_iterator I) |
| 3082 | : iterator_adaptor_base(I) {} |
| 3083 | |
| 3084 | const BasicBlock *operator*() const { return cast<BasicBlock>(*I); } |
| 3085 | const BasicBlock *operator->() const { return operator*(); } |
| 3086 | }; |
| 3087 | |
| 3088 | static BranchInst *Create(BasicBlock *IfTrue, |
| 3089 | Instruction *InsertBefore = nullptr) { |
| 3090 | return new(1) BranchInst(IfTrue, InsertBefore); |
| 3091 | } |
| 3092 | |
| 3093 | static BranchInst *Create(BasicBlock *IfTrue, BasicBlock *IfFalse, |
| 3094 | Value *Cond, Instruction *InsertBefore = nullptr) { |
| 3095 | return new(3) BranchInst(IfTrue, IfFalse, Cond, InsertBefore); |
| 3096 | } |
| 3097 | |
| 3098 | static BranchInst *Create(BasicBlock *IfTrue, BasicBlock *InsertAtEnd) { |
| 3099 | return new(1) BranchInst(IfTrue, InsertAtEnd); |
| 3100 | } |
| 3101 | |
| 3102 | static BranchInst *Create(BasicBlock *IfTrue, BasicBlock *IfFalse, |
| 3103 | Value *Cond, BasicBlock *InsertAtEnd) { |
| 3104 | return new(3) BranchInst(IfTrue, IfFalse, Cond, InsertAtEnd); |
| 3105 | } |
| 3106 | |
| 3107 | /// Transparently provide more efficient getOperand methods. |
| 3108 | 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; |
| 3109 | |
| 3110 | bool isUnconditional() const { return getNumOperands() == 1; } |
| 3111 | bool isConditional() const { return getNumOperands() == 3; } |
| 3112 | |
| 3113 | Value *getCondition() const { |
| 3114 | assert(isConditional() && "Cannot get condition of an uncond branch!")((void)0); |
| 3115 | return Op<-3>(); |
| 3116 | } |
| 3117 | |
| 3118 | void setCondition(Value *V) { |
| 3119 | assert(isConditional() && "Cannot set condition of unconditional branch!")((void)0); |
| 3120 | Op<-3>() = V; |
| 3121 | } |
| 3122 | |
| 3123 | unsigned getNumSuccessors() const { return 1+isConditional(); } |
| 3124 | |
| 3125 | BasicBlock *getSuccessor(unsigned i) const { |
| 3126 | assert(i < getNumSuccessors() && "Successor # out of range for Branch!")((void)0); |
| 3127 | return cast_or_null<BasicBlock>((&Op<-1>() - i)->get()); |
| 3128 | } |
| 3129 | |
| 3130 | void setSuccessor(unsigned idx, BasicBlock *NewSucc) { |
| 3131 | assert(idx < getNumSuccessors() && "Successor # out of range for Branch!")((void)0); |
| 3132 | *(&Op<-1>() - idx) = NewSucc; |
| 3133 | } |
| 3134 | |
| 3135 | /// Swap the successors of this branch instruction. |
| 3136 | /// |
| 3137 | /// Swaps the successors of the branch instruction. This also swaps any |
| 3138 | /// branch weight metadata associated with the instruction so that it |
| 3139 | /// continues to map correctly to each operand. |