| 1 | //===- InstCombineLoadStoreAlloca.cpp -------------------------------------===// |
| 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 implements the visit functions for load, store and alloca. |
| 10 | // |
| 11 | //===----------------------------------------------------------------------===// |
| 12 | |
| 13 | #include "InstCombineInternal.h" |
| 14 | #include "llvm/ADT/MapVector.h" |
| 15 | #include "llvm/ADT/SmallString.h" |
| 16 | #include "llvm/ADT/Statistic.h" |
| 17 | #include "llvm/Analysis/AliasAnalysis.h" |
| 18 | #include "llvm/Analysis/Loads.h" |
| 19 | #include "llvm/IR/DataLayout.h" |
| 20 | #include "llvm/IR/IntrinsicInst.h" |
| 21 | #include "llvm/IR/LLVMContext.h" |
| 22 | #include "llvm/IR/PatternMatch.h" |
| 23 | #include "llvm/Transforms/InstCombine/InstCombiner.h" |
| 24 | #include "llvm/Transforms/Utils/Local.h" |
| 25 | using namespace llvm; |
| 26 | using namespace PatternMatch; |
| 27 | |
| 28 | #define DEBUG_TYPE "instcombine" |
| 29 | |
| 30 | namespace llvm { |
| 31 | extern cl::opt<bool> ProfcheckDisableMetadataFixes; |
| 32 | } |
| 33 | |
| 34 | STATISTIC(NumDeadStore, "Number of dead stores eliminated" ); |
| 35 | STATISTIC(NumGlobalCopies, "Number of allocas copied from constant global" ); |
| 36 | |
| 37 | static cl::opt<unsigned> MaxCopiedFromConstantUsers( |
| 38 | "instcombine-max-copied-from-constant-users" , cl::init(Val: 300), |
| 39 | cl::desc("Maximum users to visit in copy from constant transform" ), |
| 40 | cl::Hidden); |
| 41 | |
| 42 | /// isOnlyCopiedFromConstantMemory - Recursively walk the uses of a (derived) |
| 43 | /// pointer to an alloca. Ignore any reads of the pointer, return false if we |
| 44 | /// see any stores or other unknown uses. If we see pointer arithmetic, keep |
| 45 | /// track of whether it moves the pointer (with IsOffset) but otherwise traverse |
| 46 | /// the uses. If we see a memcpy/memmove that targets an unoffseted pointer to |
| 47 | /// the alloca, and if the source pointer is a pointer to a constant memory |
| 48 | /// location, we can optimize this. |
| 49 | static bool |
| 50 | isOnlyCopiedFromConstantMemory(AAResults *AA, AllocaInst *V, |
| 51 | MemTransferInst *&TheCopy, |
| 52 | SmallVectorImpl<Instruction *> &ToDelete) { |
| 53 | // We track lifetime intrinsics as we encounter them. If we decide to go |
| 54 | // ahead and replace the value with the memory location, this lets the caller |
| 55 | // quickly eliminate the markers. |
| 56 | |
| 57 | using ValueAndIsOffset = PointerIntPair<Value *, 1, bool>; |
| 58 | SmallVector<ValueAndIsOffset, 32> Worklist; |
| 59 | SmallPtrSet<ValueAndIsOffset, 32> Visited; |
| 60 | Worklist.emplace_back(Args&: V, Args: false); |
| 61 | while (!Worklist.empty()) { |
| 62 | ValueAndIsOffset Elem = Worklist.pop_back_val(); |
| 63 | if (!Visited.insert(Ptr: Elem).second) |
| 64 | continue; |
| 65 | if (Visited.size() > MaxCopiedFromConstantUsers) |
| 66 | return false; |
| 67 | |
| 68 | const auto [Value, IsOffset] = Elem; |
| 69 | for (auto &U : Value->uses()) { |
| 70 | auto *I = cast<Instruction>(Val: U.getUser()); |
| 71 | |
| 72 | if (auto *LI = dyn_cast<LoadInst>(Val: I)) { |
| 73 | // Ignore non-volatile loads, they are always ok. |
| 74 | if (!LI->isSimple()) return false; |
| 75 | continue; |
| 76 | } |
| 77 | |
| 78 | if (isa<PHINode, SelectInst>(Val: I)) { |
| 79 | // We set IsOffset=true, to forbid the memcpy from occurring after the |
| 80 | // phi: If one of the phi operands is not based on the alloca, we |
| 81 | // would incorrectly omit a write. |
| 82 | Worklist.emplace_back(Args&: I, Args: true); |
| 83 | continue; |
| 84 | } |
| 85 | if (isa<BitCastInst, AddrSpaceCastInst>(Val: I)) { |
| 86 | // If uses of the bitcast are ok, we are ok. |
| 87 | Worklist.emplace_back(Args&: I, Args: IsOffset); |
| 88 | continue; |
| 89 | } |
| 90 | if (auto *GEP = dyn_cast<GetElementPtrInst>(Val: I)) { |
| 91 | // If the GEP has all zero indices, it doesn't offset the pointer. If it |
| 92 | // doesn't, it does. |
| 93 | Worklist.emplace_back(Args&: I, Args: IsOffset || !GEP->hasAllZeroIndices()); |
| 94 | continue; |
| 95 | } |
| 96 | |
| 97 | if (auto *Call = dyn_cast<CallBase>(Val: I)) { |
| 98 | // If this is the function being called then we treat it like a load and |
| 99 | // ignore it. |
| 100 | if (Call->isCallee(U: &U)) |
| 101 | continue; |
| 102 | |
| 103 | unsigned DataOpNo = Call->getDataOperandNo(U: &U); |
| 104 | bool IsArgOperand = Call->isArgOperand(U: &U); |
| 105 | |
| 106 | // Inalloca arguments are clobbered by the call. |
| 107 | if (IsArgOperand && Call->isInAllocaArgument(ArgNo: DataOpNo)) |
| 108 | return false; |
| 109 | |
| 110 | // If this call site doesn't modify the memory, then we know it is just |
| 111 | // a load (but one that potentially returns the value itself), so we can |
| 112 | // ignore it if we know that the value isn't captured. |
| 113 | bool NoCapture = Call->doesNotCapture(OpNo: DataOpNo); |
| 114 | if (NoCapture && |
| 115 | (Call->onlyReadsMemory() || Call->onlyReadsMemory(OpNo: DataOpNo))) |
| 116 | continue; |
| 117 | } |
| 118 | |
| 119 | // Lifetime intrinsics can be handled by the caller. |
| 120 | if (I->isLifetimeStartOrEnd()) { |
| 121 | assert(I->use_empty() && "Lifetime markers have no result to use!" ); |
| 122 | ToDelete.push_back(Elt: I); |
| 123 | continue; |
| 124 | } |
| 125 | |
| 126 | // If this is isn't our memcpy/memmove, reject it as something we can't |
| 127 | // handle. |
| 128 | MemTransferInst *MI = dyn_cast<MemTransferInst>(Val: I); |
| 129 | if (!MI) |
| 130 | return false; |
| 131 | |
| 132 | // If the transfer is volatile, reject it. |
| 133 | if (MI->isVolatile()) |
| 134 | return false; |
| 135 | |
| 136 | // If the transfer is using the alloca as a source of the transfer, then |
| 137 | // ignore it since it is a load (unless the transfer is volatile). |
| 138 | if (U.getOperandNo() == 1) |
| 139 | continue; |
| 140 | |
| 141 | // If we already have seen a copy, reject the second one. |
| 142 | if (TheCopy) return false; |
| 143 | |
| 144 | // If the pointer has been offset from the start of the alloca, we can't |
| 145 | // safely handle this. |
| 146 | if (IsOffset) return false; |
| 147 | |
| 148 | // If the memintrinsic isn't using the alloca as the dest, reject it. |
| 149 | if (U.getOperandNo() != 0) return false; |
| 150 | |
| 151 | // If the source of the memcpy/move is not constant, reject it. |
| 152 | if (isModSet(MRI: AA->getModRefInfoMask(P: MI->getSource()))) |
| 153 | return false; |
| 154 | |
| 155 | // Otherwise, the transform is safe. Remember the copy instruction. |
| 156 | TheCopy = MI; |
| 157 | } |
| 158 | } |
| 159 | return true; |
| 160 | } |
| 161 | |
| 162 | /// isOnlyCopiedFromConstantMemory - Return true if the specified alloca is only |
| 163 | /// modified by a copy from a constant memory location. If we can prove this, we |
| 164 | /// can replace any uses of the alloca with uses of the memory location |
| 165 | /// directly. |
| 166 | static MemTransferInst * |
| 167 | isOnlyCopiedFromConstantMemory(AAResults *AA, |
| 168 | AllocaInst *AI, |
| 169 | SmallVectorImpl<Instruction *> &ToDelete) { |
| 170 | MemTransferInst *TheCopy = nullptr; |
| 171 | if (isOnlyCopiedFromConstantMemory(AA, V: AI, TheCopy, ToDelete)) |
| 172 | return TheCopy; |
| 173 | return nullptr; |
| 174 | } |
| 175 | |
| 176 | /// Returns true if V is dereferenceable for size of alloca. |
| 177 | static bool isDereferenceableForAllocaSize(const Value *V, const AllocaInst *AI, |
| 178 | const DataLayout &DL) { |
| 179 | std::optional<TypeSize> AllocaSize = AI->getAllocationSize(DL); |
| 180 | if (!AllocaSize || AllocaSize->isScalable()) |
| 181 | return false; |
| 182 | return isDereferenceableAndAlignedPointer(V, Alignment: AI->getAlign(), |
| 183 | Size: APInt(64, *AllocaSize), DL); |
| 184 | } |
| 185 | |
| 186 | static Instruction *simplifyAllocaArraySize(InstCombinerImpl &IC, |
| 187 | AllocaInst &AI, DominatorTree &DT) { |
| 188 | // Check for array size of 1 (scalar allocation). |
| 189 | if (!AI.isArrayAllocation()) { |
| 190 | // i32 1 is the canonical array size for scalar allocations. |
| 191 | if (AI.getArraySize()->getType()->isIntegerTy(Bitwidth: 32)) |
| 192 | return nullptr; |
| 193 | |
| 194 | // Canonicalize it. |
| 195 | return IC.replaceOperand(I&: AI, OpNum: 0, V: IC.Builder.getInt32(C: 1)); |
| 196 | } |
| 197 | |
| 198 | // Convert: alloca Ty, C - where C is a constant != 1 into: alloca [C x Ty], 1 |
| 199 | if (const ConstantInt *C = dyn_cast<ConstantInt>(Val: AI.getArraySize())) { |
| 200 | if (C->getValue().getActiveBits() <= 64) { |
| 201 | Type *NewTy = ArrayType::get(ElementType: AI.getAllocatedType(), NumElements: C->getZExtValue()); |
| 202 | AllocaInst *New = IC.Builder.CreateAlloca(Ty: NewTy, AddrSpace: AI.getAddressSpace(), |
| 203 | ArraySize: nullptr, Name: AI.getName()); |
| 204 | New->setAlignment(AI.getAlign()); |
| 205 | New->setUsedWithInAlloca(AI.isUsedWithInAlloca()); |
| 206 | |
| 207 | replaceAllDbgUsesWith(From&: AI, To&: *New, DomPoint&: *New, DT); |
| 208 | return IC.replaceInstUsesWith(I&: AI, V: New); |
| 209 | } |
| 210 | } |
| 211 | |
| 212 | if (isa<UndefValue>(Val: AI.getArraySize())) |
| 213 | return IC.replaceInstUsesWith(I&: AI, V: Constant::getNullValue(Ty: AI.getType())); |
| 214 | |
| 215 | // Ensure that the alloca array size argument has type equal to the offset |
| 216 | // size of the alloca() pointer, which, in the tyical case, is intptr_t, |
| 217 | // so that any casting is exposed early. |
| 218 | Type *PtrIdxTy = IC.getDataLayout().getIndexType(PtrTy: AI.getType()); |
| 219 | if (AI.getArraySize()->getType() != PtrIdxTy) { |
| 220 | Value *V = IC.Builder.CreateIntCast(V: AI.getArraySize(), DestTy: PtrIdxTy, isSigned: false); |
| 221 | return IC.replaceOperand(I&: AI, OpNum: 0, V); |
| 222 | } |
| 223 | |
| 224 | return nullptr; |
| 225 | } |
| 226 | |
| 227 | namespace { |
| 228 | // If I and V are pointers in different address space, it is not allowed to |
| 229 | // use replaceAllUsesWith since I and V have different types. A |
| 230 | // non-target-specific transformation should not use addrspacecast on V since |
| 231 | // the two address space may be disjoint depending on target. |
| 232 | // |
| 233 | // This class chases down uses of the old pointer until reaching the load |
| 234 | // instructions, then replaces the old pointer in the load instructions with |
| 235 | // the new pointer. If during the chasing it sees bitcast or GEP, it will |
| 236 | // create new bitcast or GEP with the new pointer and use them in the load |
| 237 | // instruction. |
| 238 | class PointerReplacer { |
| 239 | public: |
| 240 | PointerReplacer(InstCombinerImpl &IC, Instruction &Root, unsigned SrcAS) |
| 241 | : IC(IC), Root(Root), FromAS(SrcAS) {} |
| 242 | |
| 243 | bool collectUsers(); |
| 244 | void replacePointer(Value *V); |
| 245 | |
| 246 | private: |
| 247 | void replace(Instruction *I); |
| 248 | Value *getReplacement(Value *V) const { return WorkMap.lookup(Key: V); } |
| 249 | bool isAvailable(Instruction *I) const { |
| 250 | return I == &Root || UsersToReplace.contains(key: I); |
| 251 | } |
| 252 | |
| 253 | bool isEqualOrValidAddrSpaceCast(const Instruction *I, |
| 254 | unsigned FromAS) const { |
| 255 | const auto *ASC = dyn_cast<AddrSpaceCastInst>(Val: I); |
| 256 | if (!ASC) |
| 257 | return false; |
| 258 | unsigned ToAS = ASC->getDestAddressSpace(); |
| 259 | return (FromAS == ToAS) || IC.isValidAddrSpaceCast(FromAS, ToAS); |
| 260 | } |
| 261 | |
| 262 | SmallSetVector<Instruction *, 32> UsersToReplace; |
| 263 | MapVector<Value *, Value *> WorkMap; |
| 264 | InstCombinerImpl &IC; |
| 265 | Instruction &Root; |
| 266 | unsigned FromAS; |
| 267 | }; |
| 268 | } // end anonymous namespace |
| 269 | |
| 270 | bool PointerReplacer::collectUsers() { |
| 271 | SmallVector<Instruction *> Worklist; |
| 272 | SmallSetVector<Instruction *, 32> ValuesToRevisit; |
| 273 | |
| 274 | auto PushUsersToWorklist = [&](Instruction *Inst) { |
| 275 | for (auto *U : Inst->users()) |
| 276 | if (auto *I = dyn_cast<Instruction>(Val: U)) |
| 277 | if (!isAvailable(I) && !ValuesToRevisit.contains(key: I)) |
| 278 | Worklist.emplace_back(Args&: I); |
| 279 | }; |
| 280 | |
| 281 | auto TryPushInstOperand = [&](Instruction *InstOp) { |
| 282 | if (!UsersToReplace.contains(key: InstOp)) { |
| 283 | if (!ValuesToRevisit.insert(X: InstOp)) |
| 284 | return false; |
| 285 | Worklist.emplace_back(Args&: InstOp); |
| 286 | } |
| 287 | return true; |
| 288 | }; |
| 289 | |
| 290 | PushUsersToWorklist(&Root); |
| 291 | while (!Worklist.empty()) { |
| 292 | Instruction *Inst = Worklist.pop_back_val(); |
| 293 | if (auto *Load = dyn_cast<LoadInst>(Val: Inst)) { |
| 294 | if (Load->isVolatile()) |
| 295 | return false; |
| 296 | UsersToReplace.insert(X: Load); |
| 297 | } else if (auto *PHI = dyn_cast<PHINode>(Val: Inst)) { |
| 298 | /// TODO: Handle poison and null pointers for PHI and select. |
| 299 | // If all incoming values are available, mark this PHI as |
| 300 | // replacable and push it's users into the worklist. |
| 301 | bool IsReplaceable = all_of(Range: PHI->incoming_values(), |
| 302 | P: [](Value *V) { return isa<Instruction>(Val: V); }); |
| 303 | if (IsReplaceable && all_of(Range: PHI->incoming_values(), P: [&](Value *V) { |
| 304 | return isAvailable(I: cast<Instruction>(Val: V)); |
| 305 | })) { |
| 306 | UsersToReplace.insert(X: PHI); |
| 307 | PushUsersToWorklist(PHI); |
| 308 | continue; |
| 309 | } |
| 310 | |
| 311 | // Either an incoming value is not an instruction or not all |
| 312 | // incoming values are available. If this PHI was already |
| 313 | // visited prior to this iteration, return false. |
| 314 | if (!IsReplaceable || !ValuesToRevisit.insert(X: PHI)) |
| 315 | return false; |
| 316 | |
| 317 | // Push PHI back into the stack, followed by unavailable |
| 318 | // incoming values. |
| 319 | Worklist.emplace_back(Args&: PHI); |
| 320 | for (unsigned Idx = 0; Idx < PHI->getNumIncomingValues(); ++Idx) { |
| 321 | if (!TryPushInstOperand(cast<Instruction>(Val: PHI->getIncomingValue(i: Idx)))) |
| 322 | return false; |
| 323 | } |
| 324 | } else if (auto *SI = dyn_cast<SelectInst>(Val: Inst)) { |
| 325 | auto *TrueInst = dyn_cast<Instruction>(Val: SI->getTrueValue()); |
| 326 | auto *FalseInst = dyn_cast<Instruction>(Val: SI->getFalseValue()); |
| 327 | if (!TrueInst || !FalseInst) |
| 328 | return false; |
| 329 | |
| 330 | if (isAvailable(I: TrueInst) && isAvailable(I: FalseInst)) { |
| 331 | UsersToReplace.insert(X: SI); |
| 332 | PushUsersToWorklist(SI); |
| 333 | continue; |
| 334 | } |
| 335 | |
| 336 | // Push select back onto the stack, followed by unavailable true/false |
| 337 | // value. |
| 338 | Worklist.emplace_back(Args&: SI); |
| 339 | if (!TryPushInstOperand(TrueInst) || !TryPushInstOperand(FalseInst)) |
| 340 | return false; |
| 341 | } else if (auto *GEP = dyn_cast<GetElementPtrInst>(Val: Inst)) { |
| 342 | auto *PtrOp = dyn_cast<Instruction>(Val: GEP->getPointerOperand()); |
| 343 | if (!PtrOp) |
| 344 | return false; |
| 345 | if (isAvailable(I: PtrOp)) { |
| 346 | UsersToReplace.insert(X: GEP); |
| 347 | PushUsersToWorklist(GEP); |
| 348 | continue; |
| 349 | } |
| 350 | |
| 351 | Worklist.emplace_back(Args&: GEP); |
| 352 | if (!TryPushInstOperand(PtrOp)) |
| 353 | return false; |
| 354 | } else if (auto *MI = dyn_cast<MemTransferInst>(Val: Inst)) { |
| 355 | if (MI->isVolatile()) |
| 356 | return false; |
| 357 | UsersToReplace.insert(X: Inst); |
| 358 | } else if (isEqualOrValidAddrSpaceCast(I: Inst, FromAS)) { |
| 359 | UsersToReplace.insert(X: Inst); |
| 360 | PushUsersToWorklist(Inst); |
| 361 | } else if (Inst->isLifetimeStartOrEnd()) { |
| 362 | continue; |
| 363 | } else { |
| 364 | // TODO: For arbitrary uses with address space mismatches, should we check |
| 365 | // if we can introduce a valid addrspacecast? |
| 366 | LLVM_DEBUG(dbgs() << "Cannot handle pointer user: " << *Inst << '\n'); |
| 367 | return false; |
| 368 | } |
| 369 | } |
| 370 | |
| 371 | return true; |
| 372 | } |
| 373 | |
| 374 | void PointerReplacer::replacePointer(Value *V) { |
| 375 | assert(cast<PointerType>(Root.getType()) != cast<PointerType>(V->getType()) && |
| 376 | "Invalid usage" ); |
| 377 | WorkMap[&Root] = V; |
| 378 | SmallVector<Instruction *> Worklist; |
| 379 | SetVector<Instruction *> PostOrderWorklist; |
| 380 | SmallPtrSet<Instruction *, 32> Visited; |
| 381 | |
| 382 | // Perform a postorder traversal of the users of Root. |
| 383 | Worklist.push_back(Elt: &Root); |
| 384 | while (!Worklist.empty()) { |
| 385 | Instruction *I = Worklist.back(); |
| 386 | |
| 387 | // If I has not been processed before, push each of its |
| 388 | // replacable users into the worklist. |
| 389 | if (Visited.insert(Ptr: I).second) { |
| 390 | for (auto *U : I->users()) { |
| 391 | auto *UserInst = cast<Instruction>(Val: U); |
| 392 | if (UsersToReplace.contains(key: UserInst) && !Visited.contains(Ptr: UserInst)) |
| 393 | Worklist.push_back(Elt: UserInst); |
| 394 | } |
| 395 | // Otherwise, users of I have already been pushed into |
| 396 | // the PostOrderWorklist. Push I as well. |
| 397 | } else { |
| 398 | PostOrderWorklist.insert(X: I); |
| 399 | Worklist.pop_back(); |
| 400 | } |
| 401 | } |
| 402 | |
| 403 | // Replace pointers in reverse-postorder. |
| 404 | for (Instruction *I : reverse(C&: PostOrderWorklist)) |
| 405 | replace(I); |
| 406 | } |
| 407 | |
| 408 | void PointerReplacer::replace(Instruction *I) { |
| 409 | if (getReplacement(V: I)) |
| 410 | return; |
| 411 | |
| 412 | if (auto *LT = dyn_cast<LoadInst>(Val: I)) { |
| 413 | auto *V = getReplacement(V: LT->getPointerOperand()); |
| 414 | assert(V && "Operand not replaced" ); |
| 415 | auto *NewI = new LoadInst(LT->getType(), V, "" , LT->isVolatile(), |
| 416 | LT->getAlign(), LT->getOrdering(), |
| 417 | LT->getSyncScopeID()); |
| 418 | NewI->takeName(V: LT); |
| 419 | copyMetadataForLoad(Dest&: *NewI, Source: *LT); |
| 420 | |
| 421 | IC.InsertNewInstWith(New: NewI, Old: LT->getIterator()); |
| 422 | IC.replaceInstUsesWith(I&: *LT, V: NewI); |
| 423 | // LT has actually been replaced by NewI. It is useless to insert LT into |
| 424 | // the map. Instead, we insert NewI into the map to indicate this is the |
| 425 | // replacement (new value). |
| 426 | WorkMap[NewI] = NewI; |
| 427 | } else if (auto *PHI = dyn_cast<PHINode>(Val: I)) { |
| 428 | // Create a new PHI by replacing any incoming value that is a user of the |
| 429 | // root pointer and has a replacement. |
| 430 | Value *V = WorkMap.lookup(Key: PHI->getIncomingValue(i: 0)); |
| 431 | PHI->mutateType(Ty: V ? V->getType() : PHI->getIncomingValue(i: 0)->getType()); |
| 432 | for (unsigned int I = 0; I < PHI->getNumIncomingValues(); ++I) { |
| 433 | Value *V = WorkMap.lookup(Key: PHI->getIncomingValue(i: I)); |
| 434 | PHI->setIncomingValue(i: I, V: V ? V : PHI->getIncomingValue(i: I)); |
| 435 | } |
| 436 | WorkMap[PHI] = PHI; |
| 437 | } else if (auto *GEP = dyn_cast<GetElementPtrInst>(Val: I)) { |
| 438 | auto *V = getReplacement(V: GEP->getPointerOperand()); |
| 439 | assert(V && "Operand not replaced" ); |
| 440 | SmallVector<Value *, 8> Indices(GEP->indices()); |
| 441 | auto *NewI = |
| 442 | GetElementPtrInst::Create(PointeeType: GEP->getSourceElementType(), Ptr: V, IdxList: Indices); |
| 443 | IC.InsertNewInstWith(New: NewI, Old: GEP->getIterator()); |
| 444 | NewI->takeName(V: GEP); |
| 445 | NewI->setNoWrapFlags(GEP->getNoWrapFlags()); |
| 446 | WorkMap[GEP] = NewI; |
| 447 | } else if (auto *SI = dyn_cast<SelectInst>(Val: I)) { |
| 448 | Value *TrueValue = SI->getTrueValue(); |
| 449 | Value *FalseValue = SI->getFalseValue(); |
| 450 | if (Value *Replacement = getReplacement(V: TrueValue)) |
| 451 | TrueValue = Replacement; |
| 452 | if (Value *Replacement = getReplacement(V: FalseValue)) |
| 453 | FalseValue = Replacement; |
| 454 | auto *NewSI = SelectInst::Create(C: SI->getCondition(), S1: TrueValue, S2: FalseValue, |
| 455 | NameStr: SI->getName(), InsertBefore: nullptr, MDFrom: SI); |
| 456 | IC.InsertNewInstWith(New: NewSI, Old: SI->getIterator()); |
| 457 | NewSI->takeName(V: SI); |
| 458 | WorkMap[SI] = NewSI; |
| 459 | } else if (auto *MemCpy = dyn_cast<MemTransferInst>(Val: I)) { |
| 460 | auto *DestV = MemCpy->getRawDest(); |
| 461 | auto *SrcV = MemCpy->getRawSource(); |
| 462 | |
| 463 | if (auto *DestReplace = getReplacement(V: DestV)) |
| 464 | DestV = DestReplace; |
| 465 | if (auto *SrcReplace = getReplacement(V: SrcV)) |
| 466 | SrcV = SrcReplace; |
| 467 | |
| 468 | IC.Builder.SetInsertPoint(MemCpy); |
| 469 | auto *NewI = IC.Builder.CreateMemTransferInst( |
| 470 | IntrID: MemCpy->getIntrinsicID(), Dst: DestV, DstAlign: MemCpy->getDestAlign(), Src: SrcV, |
| 471 | SrcAlign: MemCpy->getSourceAlign(), Size: MemCpy->getLength(), isVolatile: MemCpy->isVolatile()); |
| 472 | AAMDNodes AAMD = MemCpy->getAAMetadata(); |
| 473 | if (AAMD) |
| 474 | NewI->setAAMetadata(AAMD); |
| 475 | |
| 476 | IC.eraseInstFromFunction(I&: *MemCpy); |
| 477 | WorkMap[MemCpy] = NewI; |
| 478 | } else if (auto *ASC = dyn_cast<AddrSpaceCastInst>(Val: I)) { |
| 479 | auto *V = getReplacement(V: ASC->getPointerOperand()); |
| 480 | assert(V && "Operand not replaced" ); |
| 481 | assert(isEqualOrValidAddrSpaceCast( |
| 482 | ASC, V->getType()->getPointerAddressSpace()) && |
| 483 | "Invalid address space cast!" ); |
| 484 | |
| 485 | if (V->getType()->getPointerAddressSpace() != |
| 486 | ASC->getType()->getPointerAddressSpace()) { |
| 487 | auto *NewI = new AddrSpaceCastInst(V, ASC->getType(), "" ); |
| 488 | NewI->takeName(V: ASC); |
| 489 | IC.InsertNewInstWith(New: NewI, Old: ASC->getIterator()); |
| 490 | WorkMap[ASC] = NewI; |
| 491 | } else { |
| 492 | WorkMap[ASC] = V; |
| 493 | } |
| 494 | |
| 495 | } else { |
| 496 | llvm_unreachable("should never reach here" ); |
| 497 | } |
| 498 | } |
| 499 | |
| 500 | Instruction *InstCombinerImpl::visitAllocaInst(AllocaInst &AI) { |
| 501 | if (auto *I = simplifyAllocaArraySize(IC&: *this, AI, DT)) |
| 502 | return I; |
| 503 | |
| 504 | // Move all alloca's of zero byte objects to the entry block and merge them |
| 505 | // together. Note that we only do this for alloca's, because malloc should |
| 506 | // allocate and return a unique pointer, even for a zero byte allocation. |
| 507 | std::optional<TypeSize> Size = AI.getAllocationSize(DL); |
| 508 | if (Size && Size->isZero()) { |
| 509 | // For a zero sized alloca there is no point in doing an array allocation. |
| 510 | // This is helpful if the array size is a complicated expression not used |
| 511 | // elsewhere. |
| 512 | if (AI.isArrayAllocation()) |
| 513 | return replaceOperand(I&: AI, OpNum: 0, |
| 514 | V: ConstantInt::get(Ty: AI.getArraySize()->getType(), V: 1)); |
| 515 | |
| 516 | // Get the first instruction in the entry block. |
| 517 | BasicBlock &EntryBlock = AI.getParent()->getParent()->getEntryBlock(); |
| 518 | BasicBlock::iterator FirstInst = EntryBlock.getFirstNonPHIOrDbg(); |
| 519 | if (&*FirstInst != &AI) { |
| 520 | // If the entry block doesn't start with a zero-size alloca then move |
| 521 | // this one to the start of the entry block. There is no problem with |
| 522 | // dominance as the array size was forced to a constant earlier already. |
| 523 | AllocaInst *EntryAI = dyn_cast<AllocaInst>(Val&: FirstInst); |
| 524 | std::optional<TypeSize> EntryAISize = |
| 525 | EntryAI ? EntryAI->getAllocationSize(DL) : std::nullopt; |
| 526 | if (!EntryAISize || !EntryAISize->isZero()) { |
| 527 | AI.moveBefore(InsertPos: FirstInst); |
| 528 | return &AI; |
| 529 | } |
| 530 | |
| 531 | // Replace this zero-sized alloca with the one at the start of the entry |
| 532 | // block after ensuring that the address will be aligned enough for both |
| 533 | // types. |
| 534 | const Align MaxAlign = std::max(a: EntryAI->getAlign(), b: AI.getAlign()); |
| 535 | EntryAI->setAlignment(MaxAlign); |
| 536 | return replaceInstUsesWith(I&: AI, V: EntryAI); |
| 537 | } |
| 538 | } |
| 539 | |
| 540 | // Check to see if this allocation is only modified by a memcpy/memmove from |
| 541 | // a memory location whose alignment is equal to or exceeds that of the |
| 542 | // allocation. If this is the case, we can change all users to use the |
| 543 | // constant memory location instead. This is commonly produced by the CFE by |
| 544 | // constructs like "void foo() { int A[] = {1,2,3,4,5,6,7,8,9...}; }" if 'A' |
| 545 | // is only subsequently read. |
| 546 | SmallVector<Instruction *, 4> ToDelete; |
| 547 | if (MemTransferInst *Copy = isOnlyCopiedFromConstantMemory(AA, AI: &AI, ToDelete)) { |
| 548 | Value *TheSrc = Copy->getSource(); |
| 549 | Align AllocaAlign = AI.getAlign(); |
| 550 | Align SourceAlign = getOrEnforceKnownAlignment( |
| 551 | V: TheSrc, PrefAlign: AllocaAlign, DL, CxtI: &AI, AC: &AC, DT: &DT); |
| 552 | if (AllocaAlign <= SourceAlign && |
| 553 | isDereferenceableForAllocaSize(V: TheSrc, AI: &AI, DL) && |
| 554 | !isa<Instruction>(Val: TheSrc)) { |
| 555 | // FIXME: Can we sink instructions without violating dominance when TheSrc |
| 556 | // is an instruction instead of a constant or argument? |
| 557 | LLVM_DEBUG(dbgs() << "Found alloca equal to global: " << AI << '\n'); |
| 558 | LLVM_DEBUG(dbgs() << " memcpy = " << *Copy << '\n'); |
| 559 | unsigned SrcAddrSpace = TheSrc->getType()->getPointerAddressSpace(); |
| 560 | if (AI.getAddressSpace() == SrcAddrSpace) { |
| 561 | for (Instruction *Delete : ToDelete) |
| 562 | eraseInstFromFunction(I&: *Delete); |
| 563 | |
| 564 | Instruction *NewI = replaceInstUsesWith(I&: AI, V: TheSrc); |
| 565 | eraseInstFromFunction(I&: *Copy); |
| 566 | ++NumGlobalCopies; |
| 567 | return NewI; |
| 568 | } |
| 569 | |
| 570 | PointerReplacer PtrReplacer(*this, AI, SrcAddrSpace); |
| 571 | if (PtrReplacer.collectUsers()) { |
| 572 | for (Instruction *Delete : ToDelete) |
| 573 | eraseInstFromFunction(I&: *Delete); |
| 574 | |
| 575 | PtrReplacer.replacePointer(V: TheSrc); |
| 576 | ++NumGlobalCopies; |
| 577 | } |
| 578 | } |
| 579 | } |
| 580 | |
| 581 | // At last, use the generic allocation site handler to aggressively remove |
| 582 | // unused allocas. |
| 583 | return visitAllocSite(FI&: AI); |
| 584 | } |
| 585 | |
| 586 | // Are we allowed to form a atomic load or store of this type? |
| 587 | static bool isSupportedAtomicType(Type *Ty) { |
| 588 | return Ty->isIntOrPtrTy() || Ty->isFloatingPointTy(); |
| 589 | } |
| 590 | |
| 591 | /// Helper to combine a load to a new type. |
| 592 | /// |
| 593 | /// This just does the work of combining a load to a new type. It handles |
| 594 | /// metadata, etc., and returns the new instruction. The \c NewTy should be the |
| 595 | /// loaded *value* type. This will convert it to a pointer, cast the operand to |
| 596 | /// that pointer type, load it, etc. |
| 597 | /// |
| 598 | /// Note that this will create all of the instructions with whatever insert |
| 599 | /// point the \c InstCombinerImpl currently is using. |
| 600 | LoadInst *InstCombinerImpl::combineLoadToNewType(LoadInst &LI, Type *NewTy, |
| 601 | const Twine &Suffix) { |
| 602 | assert((!LI.isAtomic() || isSupportedAtomicType(NewTy)) && |
| 603 | "can't fold an atomic load to requested type" ); |
| 604 | |
| 605 | LoadInst *NewLoad = |
| 606 | Builder.CreateAlignedLoad(Ty: NewTy, Ptr: LI.getPointerOperand(), Align: LI.getAlign(), |
| 607 | isVolatile: LI.isVolatile(), Name: LI.getName() + Suffix); |
| 608 | NewLoad->setAtomic(Ordering: LI.getOrdering(), SSID: LI.getSyncScopeID()); |
| 609 | copyMetadataForLoad(Dest&: *NewLoad, Source: LI); |
| 610 | return NewLoad; |
| 611 | } |
| 612 | |
| 613 | /// Combine a store to a new type. |
| 614 | /// |
| 615 | /// Returns the newly created store instruction. |
| 616 | static StoreInst *combineStoreToNewValue(InstCombinerImpl &IC, StoreInst &SI, |
| 617 | Value *V) { |
| 618 | assert((!SI.isAtomic() || isSupportedAtomicType(V->getType())) && |
| 619 | "can't fold an atomic store of requested type" ); |
| 620 | |
| 621 | Value *Ptr = SI.getPointerOperand(); |
| 622 | SmallVector<std::pair<unsigned, MDNode *>, 8> MD; |
| 623 | SI.getAllMetadata(MDs&: MD); |
| 624 | |
| 625 | StoreInst *NewStore = |
| 626 | IC.Builder.CreateAlignedStore(Val: V, Ptr, Align: SI.getAlign(), isVolatile: SI.isVolatile()); |
| 627 | NewStore->setAtomic(Ordering: SI.getOrdering(), SSID: SI.getSyncScopeID()); |
| 628 | for (const auto &MDPair : MD) { |
| 629 | unsigned ID = MDPair.first; |
| 630 | MDNode *N = MDPair.second; |
| 631 | // Note, essentially every kind of metadata should be preserved here! This |
| 632 | // routine is supposed to clone a store instruction changing *only its |
| 633 | // type*. The only metadata it makes sense to drop is metadata which is |
| 634 | // invalidated when the pointer type changes. This should essentially |
| 635 | // never be the case in LLVM, but we explicitly switch over only known |
| 636 | // metadata to be conservatively correct. If you are adding metadata to |
| 637 | // LLVM which pertains to stores, you almost certainly want to add it |
| 638 | // here. |
| 639 | switch (ID) { |
| 640 | case LLVMContext::MD_dbg: |
| 641 | case LLVMContext::MD_DIAssignID: |
| 642 | case LLVMContext::MD_tbaa: |
| 643 | case LLVMContext::MD_prof: |
| 644 | case LLVMContext::MD_fpmath: |
| 645 | case LLVMContext::MD_tbaa_struct: |
| 646 | case LLVMContext::MD_alias_scope: |
| 647 | case LLVMContext::MD_noalias: |
| 648 | case LLVMContext::MD_nontemporal: |
| 649 | case LLVMContext::MD_mem_parallel_loop_access: |
| 650 | case LLVMContext::MD_access_group: |
| 651 | // All of these directly apply. |
| 652 | NewStore->setMetadata(KindID: ID, Node: N); |
| 653 | break; |
| 654 | case LLVMContext::MD_invariant_load: |
| 655 | case LLVMContext::MD_nonnull: |
| 656 | case LLVMContext::MD_noundef: |
| 657 | case LLVMContext::MD_range: |
| 658 | case LLVMContext::MD_align: |
| 659 | case LLVMContext::MD_dereferenceable: |
| 660 | case LLVMContext::MD_dereferenceable_or_null: |
| 661 | // These don't apply for stores. |
| 662 | break; |
| 663 | } |
| 664 | } |
| 665 | |
| 666 | return NewStore; |
| 667 | } |
| 668 | |
| 669 | /// Combine loads to match the type of their uses' value after looking |
| 670 | /// through intervening bitcasts. |
| 671 | /// |
| 672 | /// The core idea here is that if the result of a load is used in an operation, |
| 673 | /// we should load the type most conducive to that operation. For example, when |
| 674 | /// loading an integer and converting that immediately to a pointer, we should |
| 675 | /// instead directly load a pointer. |
| 676 | /// |
| 677 | /// However, this routine must never change the width of a load or the number of |
| 678 | /// loads as that would introduce a semantic change. This combine is expected to |
| 679 | /// be a semantic no-op which just allows loads to more closely model the types |
| 680 | /// of their consuming operations. |
| 681 | /// |
| 682 | /// Currently, we also refuse to change the precise type used for an atomic load |
| 683 | /// or a volatile load. This is debatable, and might be reasonable to change |
| 684 | /// later. However, it is risky in case some backend or other part of LLVM is |
| 685 | /// relying on the exact type loaded to select appropriate atomic operations. |
| 686 | static Instruction *combineLoadToOperationType(InstCombinerImpl &IC, |
| 687 | LoadInst &Load) { |
| 688 | // FIXME: We could probably with some care handle both volatile and ordered |
| 689 | // atomic loads here but it isn't clear that this is important. |
| 690 | if (!Load.isUnordered()) |
| 691 | return nullptr; |
| 692 | |
| 693 | if (Load.use_empty()) |
| 694 | return nullptr; |
| 695 | |
| 696 | // swifterror values can't be bitcasted. |
| 697 | if (Load.getPointerOperand()->isSwiftError()) |
| 698 | return nullptr; |
| 699 | |
| 700 | // Fold away bit casts of the loaded value by loading the desired type. |
| 701 | // Note that we should not do this for pointer<->integer casts, |
| 702 | // because that would result in type punning. |
| 703 | if (Load.hasOneUse()) { |
| 704 | // Don't transform when the type is x86_amx, it makes the pass that lower |
| 705 | // x86_amx type happy. |
| 706 | Type *LoadTy = Load.getType(); |
| 707 | if (auto *BC = dyn_cast<BitCastInst>(Val: Load.user_back())) { |
| 708 | assert(!LoadTy->isX86_AMXTy() && "Load from x86_amx* should not happen!" ); |
| 709 | if (BC->getType()->isX86_AMXTy()) |
| 710 | return nullptr; |
| 711 | } |
| 712 | |
| 713 | if (auto *CastUser = dyn_cast<CastInst>(Val: Load.user_back())) { |
| 714 | Type *DestTy = CastUser->getDestTy(); |
| 715 | if (CastUser->isNoopCast(DL: IC.getDataLayout()) && |
| 716 | LoadTy->isPtrOrPtrVectorTy() == DestTy->isPtrOrPtrVectorTy() && |
| 717 | (!Load.isAtomic() || isSupportedAtomicType(Ty: DestTy))) { |
| 718 | LoadInst *NewLoad = IC.combineLoadToNewType(LI&: Load, NewTy: DestTy); |
| 719 | CastUser->replaceAllUsesWith(V: NewLoad); |
| 720 | IC.eraseInstFromFunction(I&: *CastUser); |
| 721 | return &Load; |
| 722 | } |
| 723 | } |
| 724 | } |
| 725 | |
| 726 | // FIXME: We should also canonicalize loads of vectors when their elements are |
| 727 | // cast to other types. |
| 728 | return nullptr; |
| 729 | } |
| 730 | |
| 731 | static Instruction *unpackLoadToAggregate(InstCombinerImpl &IC, LoadInst &LI) { |
| 732 | // FIXME: We could probably with some care handle both volatile and atomic |
| 733 | // stores here but it isn't clear that this is important. |
| 734 | if (!LI.isSimple()) |
| 735 | return nullptr; |
| 736 | |
| 737 | Type *T = LI.getType(); |
| 738 | if (!T->isAggregateType()) |
| 739 | return nullptr; |
| 740 | |
| 741 | StringRef Name = LI.getName(); |
| 742 | |
| 743 | if (auto *ST = dyn_cast<StructType>(Val: T)) { |
| 744 | // If the struct only have one element, we unpack. |
| 745 | auto NumElements = ST->getNumElements(); |
| 746 | if (NumElements == 1) { |
| 747 | LoadInst *NewLoad = IC.combineLoadToNewType(LI, NewTy: ST->getTypeAtIndex(N: 0U), |
| 748 | Suffix: ".unpack" ); |
| 749 | NewLoad->setAAMetadata(LI.getAAMetadata()); |
| 750 | // Copy invariant metadata from parent load. |
| 751 | NewLoad->copyMetadata(SrcInst: LI, WL: LLVMContext::MD_invariant_load); |
| 752 | return IC.replaceInstUsesWith(I&: LI, V: IC.Builder.CreateInsertValue( |
| 753 | Agg: PoisonValue::get(T), Val: NewLoad, Idxs: 0, Name)); |
| 754 | } |
| 755 | |
| 756 | // We don't want to break loads with padding here as we'd loose |
| 757 | // the knowledge that padding exists for the rest of the pipeline. |
| 758 | const DataLayout &DL = IC.getDataLayout(); |
| 759 | auto *SL = DL.getStructLayout(Ty: ST); |
| 760 | |
| 761 | if (SL->hasPadding()) |
| 762 | return nullptr; |
| 763 | |
| 764 | const auto Align = LI.getAlign(); |
| 765 | auto *Addr = LI.getPointerOperand(); |
| 766 | auto *IdxType = DL.getIndexType(PtrTy: Addr->getType()); |
| 767 | |
| 768 | Value *V = PoisonValue::get(T); |
| 769 | for (unsigned i = 0; i < NumElements; i++) { |
| 770 | auto *Ptr = IC.Builder.CreateInBoundsPtrAdd( |
| 771 | Ptr: Addr, Offset: IC.Builder.CreateTypeSize(Ty: IdxType, Size: SL->getElementOffset(Idx: i)), |
| 772 | Name: Name + ".elt" ); |
| 773 | auto *L = IC.Builder.CreateAlignedLoad( |
| 774 | Ty: ST->getElementType(N: i), Ptr, |
| 775 | Align: commonAlignment(A: Align, Offset: SL->getElementOffset(Idx: i).getKnownMinValue()), |
| 776 | Name: Name + ".unpack" ); |
| 777 | // Propagate AA metadata. It'll still be valid on the narrowed load. |
| 778 | L->setAAMetadata(LI.getAAMetadata()); |
| 779 | // Copy invariant metadata from parent load. |
| 780 | L->copyMetadata(SrcInst: LI, WL: LLVMContext::MD_invariant_load); |
| 781 | V = IC.Builder.CreateInsertValue(Agg: V, Val: L, Idxs: i); |
| 782 | } |
| 783 | |
| 784 | V->setName(Name); |
| 785 | return IC.replaceInstUsesWith(I&: LI, V); |
| 786 | } |
| 787 | |
| 788 | if (auto *AT = dyn_cast<ArrayType>(Val: T)) { |
| 789 | auto *ET = AT->getElementType(); |
| 790 | auto NumElements = AT->getNumElements(); |
| 791 | if (NumElements == 1) { |
| 792 | LoadInst *NewLoad = IC.combineLoadToNewType(LI, NewTy: ET, Suffix: ".unpack" ); |
| 793 | NewLoad->setAAMetadata(LI.getAAMetadata()); |
| 794 | return IC.replaceInstUsesWith(I&: LI, V: IC.Builder.CreateInsertValue( |
| 795 | Agg: PoisonValue::get(T), Val: NewLoad, Idxs: 0, Name)); |
| 796 | } |
| 797 | |
| 798 | // Bail out if the array is too large. Ideally we would like to optimize |
| 799 | // arrays of arbitrary size but this has a terrible impact on compile time. |
| 800 | // The threshold here is chosen arbitrarily, maybe needs a little bit of |
| 801 | // tuning. |
| 802 | if (NumElements > IC.MaxArraySizeForCombine) |
| 803 | return nullptr; |
| 804 | |
| 805 | const DataLayout &DL = IC.getDataLayout(); |
| 806 | TypeSize EltSize = DL.getTypeAllocSize(Ty: ET); |
| 807 | const auto Align = LI.getAlign(); |
| 808 | |
| 809 | auto *Addr = LI.getPointerOperand(); |
| 810 | auto *IdxType = Type::getInt64Ty(C&: T->getContext()); |
| 811 | auto *Zero = ConstantInt::get(Ty: IdxType, V: 0); |
| 812 | |
| 813 | Value *V = PoisonValue::get(T); |
| 814 | TypeSize Offset = TypeSize::getZero(); |
| 815 | for (uint64_t i = 0; i < NumElements; i++) { |
| 816 | Value *Indices[2] = { |
| 817 | Zero, |
| 818 | ConstantInt::get(Ty: IdxType, V: i), |
| 819 | }; |
| 820 | auto *Ptr = IC.Builder.CreateInBoundsGEP(Ty: AT, Ptr: Addr, IdxList: ArrayRef(Indices), |
| 821 | Name: Name + ".elt" ); |
| 822 | auto EltAlign = commonAlignment(A: Align, Offset: Offset.getKnownMinValue()); |
| 823 | auto *L = IC.Builder.CreateAlignedLoad(Ty: AT->getElementType(), Ptr, |
| 824 | Align: EltAlign, Name: Name + ".unpack" ); |
| 825 | L->setAAMetadata(LI.getAAMetadata()); |
| 826 | V = IC.Builder.CreateInsertValue(Agg: V, Val: L, Idxs: i); |
| 827 | Offset += EltSize; |
| 828 | } |
| 829 | |
| 830 | V->setName(Name); |
| 831 | return IC.replaceInstUsesWith(I&: LI, V); |
| 832 | } |
| 833 | |
| 834 | return nullptr; |
| 835 | } |
| 836 | |
| 837 | // If we can determine that all possible objects pointed to by the provided |
| 838 | // pointer value are, not only dereferenceable, but also definitively less than |
| 839 | // or equal to the provided maximum size, then return true. Otherwise, return |
| 840 | // false (constant global values and allocas fall into this category). |
| 841 | // |
| 842 | // FIXME: This should probably live in ValueTracking (or similar). |
| 843 | static bool isObjectSizeLessThanOrEq(Value *V, uint64_t MaxSize, |
| 844 | const DataLayout &DL) { |
| 845 | SmallPtrSet<Value *, 4> Visited; |
| 846 | SmallVector<Value *, 4> Worklist(1, V); |
| 847 | |
| 848 | do { |
| 849 | Value *P = Worklist.pop_back_val(); |
| 850 | P = P->stripPointerCasts(); |
| 851 | |
| 852 | if (!Visited.insert(Ptr: P).second) |
| 853 | continue; |
| 854 | |
| 855 | if (SelectInst *SI = dyn_cast<SelectInst>(Val: P)) { |
| 856 | Worklist.push_back(Elt: SI->getTrueValue()); |
| 857 | Worklist.push_back(Elt: SI->getFalseValue()); |
| 858 | continue; |
| 859 | } |
| 860 | |
| 861 | if (PHINode *PN = dyn_cast<PHINode>(Val: P)) { |
| 862 | append_range(C&: Worklist, R: PN->incoming_values()); |
| 863 | continue; |
| 864 | } |
| 865 | |
| 866 | if (GlobalAlias *GA = dyn_cast<GlobalAlias>(Val: P)) { |
| 867 | if (GA->isInterposable()) |
| 868 | return false; |
| 869 | Worklist.push_back(Elt: GA->getAliasee()); |
| 870 | continue; |
| 871 | } |
| 872 | |
| 873 | // If we know how big this object is, and it is less than MaxSize, continue |
| 874 | // searching. Otherwise, return false. |
| 875 | if (AllocaInst *AI = dyn_cast<AllocaInst>(Val: P)) { |
| 876 | std::optional<TypeSize> AllocSize = AI->getAllocationSize(DL); |
| 877 | if (!AllocSize || AllocSize->isScalable() || |
| 878 | AllocSize->getFixedValue() > MaxSize) |
| 879 | return false; |
| 880 | continue; |
| 881 | } |
| 882 | |
| 883 | if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Val: P)) { |
| 884 | if (!GV->hasDefinitiveInitializer() || !GV->isConstant()) |
| 885 | return false; |
| 886 | |
| 887 | uint64_t InitSize = GV->getGlobalSize(DL); |
| 888 | if (InitSize > MaxSize) |
| 889 | return false; |
| 890 | continue; |
| 891 | } |
| 892 | |
| 893 | return false; |
| 894 | } while (!Worklist.empty()); |
| 895 | |
| 896 | return true; |
| 897 | } |
| 898 | |
| 899 | // If we're indexing into an object of a known size, and the outer index is |
| 900 | // not a constant, but having any value but zero would lead to undefined |
| 901 | // behavior, replace it with zero. |
| 902 | // |
| 903 | // For example, if we have: |
| 904 | // @f.a = private unnamed_addr constant [1 x i32] [i32 12], align 4 |
| 905 | // ... |
| 906 | // %arrayidx = getelementptr inbounds [1 x i32]* @f.a, i64 0, i64 %x |
| 907 | // ... = load i32* %arrayidx, align 4 |
| 908 | // Then we know that we can replace %x in the GEP with i64 0. |
| 909 | // |
| 910 | // FIXME: We could fold any GEP index to zero that would cause UB if it were |
| 911 | // not zero. Currently, we only handle the first such index. Also, we could |
| 912 | // also search through non-zero constant indices if we kept track of the |
| 913 | // offsets those indices implied. |
| 914 | static bool canReplaceGEPIdxWithZero(InstCombinerImpl &IC, |
| 915 | GetElementPtrInst *GEPI, Instruction *MemI, |
| 916 | unsigned &Idx) { |
| 917 | if (GEPI->getNumOperands() < 2) |
| 918 | return false; |
| 919 | |
| 920 | // Find the first non-zero index of a GEP. If all indices are zero, return |
| 921 | // one past the last index. |
| 922 | auto FirstNZIdx = [](const GetElementPtrInst *GEPI) { |
| 923 | unsigned I = 1; |
| 924 | for (unsigned IE = GEPI->getNumOperands(); I != IE; ++I) { |
| 925 | Value *V = GEPI->getOperand(i_nocapture: I); |
| 926 | if (const ConstantInt *CI = dyn_cast<ConstantInt>(Val: V)) |
| 927 | if (CI->isZero()) |
| 928 | continue; |
| 929 | |
| 930 | break; |
| 931 | } |
| 932 | |
| 933 | return I; |
| 934 | }; |
| 935 | |
| 936 | // Skip through initial 'zero' indices, and find the corresponding pointer |
| 937 | // type. See if the next index is not a constant. |
| 938 | Idx = FirstNZIdx(GEPI); |
| 939 | if (Idx == GEPI->getNumOperands()) |
| 940 | return false; |
| 941 | if (isa<Constant>(Val: GEPI->getOperand(i_nocapture: Idx))) |
| 942 | return false; |
| 943 | |
| 944 | SmallVector<Value *, 4> Ops(GEPI->idx_begin(), GEPI->idx_begin() + Idx); |
| 945 | Type *SourceElementType = GEPI->getSourceElementType(); |
| 946 | // Size information about scalable vectors is not available, so we cannot |
| 947 | // deduce whether indexing at n is undefined behaviour or not. Bail out. |
| 948 | if (SourceElementType->isScalableTy()) |
| 949 | return false; |
| 950 | |
| 951 | Type *AllocTy = GetElementPtrInst::getIndexedType(Ty: SourceElementType, IdxList: Ops); |
| 952 | if (!AllocTy || !AllocTy->isSized()) |
| 953 | return false; |
| 954 | const DataLayout &DL = IC.getDataLayout(); |
| 955 | uint64_t TyAllocSize = DL.getTypeAllocSize(Ty: AllocTy).getFixedValue(); |
| 956 | |
| 957 | // If there are more indices after the one we might replace with a zero, make |
| 958 | // sure they're all non-negative. If any of them are negative, the overall |
| 959 | // address being computed might be before the base address determined by the |
| 960 | // first non-zero index. |
| 961 | auto IsAllNonNegative = [&]() { |
| 962 | for (unsigned i = Idx+1, e = GEPI->getNumOperands(); i != e; ++i) { |
| 963 | KnownBits Known = IC.computeKnownBits(V: GEPI->getOperand(i_nocapture: i), CxtI: MemI); |
| 964 | if (Known.isNonNegative()) |
| 965 | continue; |
| 966 | return false; |
| 967 | } |
| 968 | |
| 969 | return true; |
| 970 | }; |
| 971 | |
| 972 | // FIXME: If the GEP is not inbounds, and there are extra indices after the |
| 973 | // one we'll replace, those could cause the address computation to wrap |
| 974 | // (rendering the IsAllNonNegative() check below insufficient). We can do |
| 975 | // better, ignoring zero indices (and other indices we can prove small |
| 976 | // enough not to wrap). |
| 977 | if (Idx+1 != GEPI->getNumOperands() && !GEPI->isInBounds()) |
| 978 | return false; |
| 979 | |
| 980 | // Note that isObjectSizeLessThanOrEq will return true only if the pointer is |
| 981 | // also known to be dereferenceable. |
| 982 | return isObjectSizeLessThanOrEq(V: GEPI->getOperand(i_nocapture: 0), MaxSize: TyAllocSize, DL) && |
| 983 | IsAllNonNegative(); |
| 984 | } |
| 985 | |
| 986 | // If we're indexing into an object with a variable index for the memory |
| 987 | // access, but the object has only one element, we can assume that the index |
| 988 | // will always be zero. If we replace the GEP, return it. |
| 989 | static Instruction *replaceGEPIdxWithZero(InstCombinerImpl &IC, Value *Ptr, |
| 990 | Instruction &MemI) { |
| 991 | if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Val: Ptr)) { |
| 992 | unsigned Idx; |
| 993 | if (canReplaceGEPIdxWithZero(IC, GEPI, MemI: &MemI, Idx)) { |
| 994 | Instruction *NewGEPI = GEPI->clone(); |
| 995 | NewGEPI->setOperand(i: Idx, |
| 996 | Val: ConstantInt::get(Ty: GEPI->getOperand(i_nocapture: Idx)->getType(), V: 0)); |
| 997 | IC.InsertNewInstBefore(New: NewGEPI, Old: GEPI->getIterator()); |
| 998 | return NewGEPI; |
| 999 | } |
| 1000 | } |
| 1001 | |
| 1002 | return nullptr; |
| 1003 | } |
| 1004 | |
| 1005 | static bool canSimplifyNullStoreOrGEP(StoreInst &SI) { |
| 1006 | if (NullPointerIsDefined(F: SI.getFunction(), AS: SI.getPointerAddressSpace())) |
| 1007 | return false; |
| 1008 | |
| 1009 | auto *Ptr = SI.getPointerOperand(); |
| 1010 | if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Val: Ptr)) |
| 1011 | Ptr = GEPI->getOperand(i_nocapture: 0); |
| 1012 | return (isa<ConstantPointerNull>(Val: Ptr) && |
| 1013 | !NullPointerIsDefined(F: SI.getFunction(), AS: SI.getPointerAddressSpace())); |
| 1014 | } |
| 1015 | |
| 1016 | static bool canSimplifyNullLoadOrGEP(LoadInst &LI, Value *Op) { |
| 1017 | if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Val: Op)) { |
| 1018 | const Value *GEPI0 = GEPI->getOperand(i_nocapture: 0); |
| 1019 | if (isa<ConstantPointerNull>(Val: GEPI0) && |
| 1020 | !NullPointerIsDefined(F: LI.getFunction(), AS: GEPI->getPointerAddressSpace())) |
| 1021 | return true; |
| 1022 | } |
| 1023 | if (isa<UndefValue>(Val: Op) || |
| 1024 | (isa<ConstantPointerNull>(Val: Op) && |
| 1025 | !NullPointerIsDefined(F: LI.getFunction(), AS: LI.getPointerAddressSpace()))) |
| 1026 | return true; |
| 1027 | return false; |
| 1028 | } |
| 1029 | |
| 1030 | Value *InstCombinerImpl::simplifyNonNullOperand(Value *V, |
| 1031 | bool HasDereferenceable, |
| 1032 | unsigned Depth) { |
| 1033 | if (auto *Sel = dyn_cast<SelectInst>(Val: V)) { |
| 1034 | if (isa<ConstantPointerNull>(Val: Sel->getOperand(i_nocapture: 1))) |
| 1035 | return Sel->getOperand(i_nocapture: 2); |
| 1036 | |
| 1037 | if (isa<ConstantPointerNull>(Val: Sel->getOperand(i_nocapture: 2))) |
| 1038 | return Sel->getOperand(i_nocapture: 1); |
| 1039 | } |
| 1040 | |
| 1041 | if (!V->hasOneUse()) |
| 1042 | return nullptr; |
| 1043 | |
| 1044 | constexpr unsigned RecursionLimit = 3; |
| 1045 | if (Depth == RecursionLimit) |
| 1046 | return nullptr; |
| 1047 | |
| 1048 | if (auto *GEP = dyn_cast<GetElementPtrInst>(Val: V)) { |
| 1049 | if (HasDereferenceable || GEP->isInBounds()) { |
| 1050 | if (auto *Res = simplifyNonNullOperand(V: GEP->getPointerOperand(), |
| 1051 | HasDereferenceable, Depth: Depth + 1)) { |
| 1052 | replaceOperand(I&: *GEP, OpNum: 0, V: Res); |
| 1053 | addToWorklist(I: GEP); |
| 1054 | return nullptr; |
| 1055 | } |
| 1056 | } |
| 1057 | } |
| 1058 | |
| 1059 | if (auto *PHI = dyn_cast<PHINode>(Val: V)) { |
| 1060 | bool Changed = false; |
| 1061 | for (Use &U : PHI->incoming_values()) { |
| 1062 | // We set Depth to RecursionLimit to avoid expensive recursion. |
| 1063 | if (auto *Res = simplifyNonNullOperand(V: U.get(), HasDereferenceable, |
| 1064 | Depth: RecursionLimit)) { |
| 1065 | replaceUse(U, NewValue: Res); |
| 1066 | Changed = true; |
| 1067 | } |
| 1068 | } |
| 1069 | if (Changed) |
| 1070 | addToWorklist(I: PHI); |
| 1071 | return nullptr; |
| 1072 | } |
| 1073 | |
| 1074 | return nullptr; |
| 1075 | } |
| 1076 | |
| 1077 | Instruction *InstCombinerImpl::visitLoadInst(LoadInst &LI) { |
| 1078 | Value *Op = LI.getOperand(i_nocapture: 0); |
| 1079 | if (Value *Res = simplifyLoadInst(LI: &LI, PtrOp: Op, Q: SQ.getWithInstruction(I: &LI))) |
| 1080 | return replaceInstUsesWith(I&: LI, V: Res); |
| 1081 | |
| 1082 | // Try to canonicalize the loaded type. |
| 1083 | if (Instruction *Res = combineLoadToOperationType(IC&: *this, Load&: LI)) |
| 1084 | return Res; |
| 1085 | |
| 1086 | // Replace GEP indices if possible. |
| 1087 | if (Instruction *NewGEPI = replaceGEPIdxWithZero(IC&: *this, Ptr: Op, MemI&: LI)) |
| 1088 | return replaceOperand(I&: LI, OpNum: 0, V: NewGEPI); |
| 1089 | |
| 1090 | if (Instruction *Res = unpackLoadToAggregate(IC&: *this, LI)) |
| 1091 | return Res; |
| 1092 | |
| 1093 | // Do really simple store-to-load forwarding and load CSE, to catch cases |
| 1094 | // where there are several consecutive memory accesses to the same location, |
| 1095 | // separated by a few arithmetic operations. |
| 1096 | bool IsLoadCSE = false; |
| 1097 | BatchAAResults BatchAA(*AA); |
| 1098 | if (Value *AvailableVal = FindAvailableLoadedValue(Load: &LI, AA&: BatchAA, IsLoadCSE: &IsLoadCSE)) { |
| 1099 | if (IsLoadCSE) |
| 1100 | combineMetadataForCSE(K: cast<LoadInst>(Val: AvailableVal), J: &LI, DoesKMove: false); |
| 1101 | |
| 1102 | return replaceInstUsesWith( |
| 1103 | I&: LI, V: Builder.CreateBitOrPointerCast(V: AvailableVal, DestTy: LI.getType(), |
| 1104 | Name: LI.getName() + ".cast" )); |
| 1105 | } |
| 1106 | |
| 1107 | // None of the following transforms are legal for volatile/ordered atomic |
| 1108 | // loads. Most of them do apply for unordered atomics. |
| 1109 | if (!LI.isUnordered()) return nullptr; |
| 1110 | |
| 1111 | // load(gep null, ...) -> unreachable |
| 1112 | // load null/undef -> unreachable |
| 1113 | // TODO: Consider a target hook for valid address spaces for this xforms. |
| 1114 | if (canSimplifyNullLoadOrGEP(LI, Op)) { |
| 1115 | CreateNonTerminatorUnreachable(InsertAt: &LI); |
| 1116 | return replaceInstUsesWith(I&: LI, V: PoisonValue::get(T: LI.getType())); |
| 1117 | } |
| 1118 | |
| 1119 | if (Op->hasOneUse()) { |
| 1120 | // Change select and PHI nodes to select values instead of addresses: this |
| 1121 | // helps alias analysis out a lot, allows many others simplifications, and |
| 1122 | // exposes redundancy in the code. |
| 1123 | // |
| 1124 | // Note that we cannot do the transformation unless we know that the |
| 1125 | // introduced loads cannot trap! Something like this is valid as long as |
| 1126 | // the condition is always false: load (select bool %C, int* null, int* %G), |
| 1127 | // but it would not be valid if we transformed it to load from null |
| 1128 | // unconditionally. |
| 1129 | // |
| 1130 | |
| 1131 | AddrSpaceCastInst *ASC = dyn_cast<AddrSpaceCastInst>(Val: Op); |
| 1132 | Value *SelectOp = Op; |
| 1133 | if (ASC && ASC->getOperand(i_nocapture: 0)->hasOneUse()) |
| 1134 | SelectOp = ASC->getOperand(i_nocapture: 0); |
| 1135 | if (SelectInst *SI = dyn_cast<SelectInst>(Val: SelectOp)) { |
| 1136 | // load (select (Cond, &V1, &V2)) --> select(Cond, load &V1, load &V2). |
| 1137 | // or |
| 1138 | // load (addrspacecast(select (Cond, &V1, &V2))) --> |
| 1139 | // select(Cond, load (addrspacecast(&V1)), load (addrspacecast(&V2))). |
| 1140 | Align Alignment = LI.getAlign(); |
| 1141 | if (isSafeToLoadUnconditionally(V: SI->getOperand(i_nocapture: 1), Ty: LI.getType(), |
| 1142 | Alignment, DL, ScanFrom: SI) && |
| 1143 | isSafeToLoadUnconditionally(V: SI->getOperand(i_nocapture: 2), Ty: LI.getType(), |
| 1144 | Alignment, DL, ScanFrom: SI)) { |
| 1145 | |
| 1146 | auto MaybeCastedLoadOperand = [&](Value *Op) { |
| 1147 | if (ASC) |
| 1148 | return Builder.CreateAddrSpaceCast(V: Op, DestTy: ASC->getType(), |
| 1149 | Name: Op->getName() + ".cast" ); |
| 1150 | return Op; |
| 1151 | }; |
| 1152 | Value *LoadOp1 = MaybeCastedLoadOperand(SI->getOperand(i_nocapture: 1)); |
| 1153 | LoadInst *V1 = Builder.CreateLoad(Ty: LI.getType(), Ptr: LoadOp1, |
| 1154 | Name: LoadOp1->getName() + ".val" ); |
| 1155 | |
| 1156 | Value *LoadOp2 = MaybeCastedLoadOperand(SI->getOperand(i_nocapture: 2)); |
| 1157 | LoadInst *V2 = Builder.CreateLoad(Ty: LI.getType(), Ptr: LoadOp2, |
| 1158 | Name: LoadOp2->getName() + ".val" ); |
| 1159 | assert(LI.isUnordered() && "implied by above" ); |
| 1160 | V1->setAlignment(Alignment); |
| 1161 | V1->setAtomic(Ordering: LI.getOrdering(), SSID: LI.getSyncScopeID()); |
| 1162 | V2->setAlignment(Alignment); |
| 1163 | V2->setAtomic(Ordering: LI.getOrdering(), SSID: LI.getSyncScopeID()); |
| 1164 | // It is safe to copy any metadata that does not trigger UB. Copy any |
| 1165 | // poison-generating metadata. |
| 1166 | V1->copyMetadata(SrcInst: LI, WL: Metadata::PoisonGeneratingIDs); |
| 1167 | V2->copyMetadata(SrcInst: LI, WL: Metadata::PoisonGeneratingIDs); |
| 1168 | return SelectInst::Create(C: SI->getCondition(), S1: V1, S2: V2, NameStr: "" , InsertBefore: nullptr, |
| 1169 | MDFrom: ProfcheckDisableMetadataFixes ? nullptr : SI); |
| 1170 | } |
| 1171 | } |
| 1172 | } |
| 1173 | |
| 1174 | if (!NullPointerIsDefined(F: LI.getFunction(), AS: LI.getPointerAddressSpace())) |
| 1175 | if (Value *V = simplifyNonNullOperand(V: Op, /*HasDereferenceable=*/true)) |
| 1176 | return replaceOperand(I&: LI, OpNum: 0, V); |
| 1177 | |
| 1178 | return nullptr; |
| 1179 | } |
| 1180 | |
| 1181 | /// Look for extractelement/insertvalue sequence that acts like a bitcast. |
| 1182 | /// |
| 1183 | /// \returns underlying value that was "cast", or nullptr otherwise. |
| 1184 | /// |
| 1185 | /// For example, if we have: |
| 1186 | /// |
| 1187 | /// %E0 = extractelement <2 x double> %U, i32 0 |
| 1188 | /// %V0 = insertvalue [2 x double] undef, double %E0, 0 |
| 1189 | /// %E1 = extractelement <2 x double> %U, i32 1 |
| 1190 | /// %V1 = insertvalue [2 x double] %V0, double %E1, 1 |
| 1191 | /// |
| 1192 | /// and the layout of a <2 x double> is isomorphic to a [2 x double], |
| 1193 | /// then %V1 can be safely approximated by a conceptual "bitcast" of %U. |
| 1194 | /// Note that %U may contain non-undef values where %V1 has undef. |
| 1195 | static Value *likeBitCastFromVector(InstCombinerImpl &IC, Value *V) { |
| 1196 | Value *U = nullptr; |
| 1197 | while (auto *IV = dyn_cast<InsertValueInst>(Val: V)) { |
| 1198 | auto *E = dyn_cast<ExtractElementInst>(Val: IV->getInsertedValueOperand()); |
| 1199 | if (!E) |
| 1200 | return nullptr; |
| 1201 | auto *W = E->getVectorOperand(); |
| 1202 | if (!U) |
| 1203 | U = W; |
| 1204 | else if (U != W) |
| 1205 | return nullptr; |
| 1206 | auto *CI = dyn_cast<ConstantInt>(Val: E->getIndexOperand()); |
| 1207 | if (!CI || IV->getNumIndices() != 1 || CI->getZExtValue() != *IV->idx_begin()) |
| 1208 | return nullptr; |
| 1209 | V = IV->getAggregateOperand(); |
| 1210 | } |
| 1211 | if (!match(V, P: m_Undef()) || !U) |
| 1212 | return nullptr; |
| 1213 | |
| 1214 | auto *UT = cast<VectorType>(Val: U->getType()); |
| 1215 | auto *VT = V->getType(); |
| 1216 | // Check that types UT and VT are bitwise isomorphic. |
| 1217 | const auto &DL = IC.getDataLayout(); |
| 1218 | if (DL.getTypeStoreSizeInBits(Ty: UT) != DL.getTypeStoreSizeInBits(Ty: VT)) { |
| 1219 | return nullptr; |
| 1220 | } |
| 1221 | if (auto *AT = dyn_cast<ArrayType>(Val: VT)) { |
| 1222 | if (AT->getNumElements() != cast<FixedVectorType>(Val: UT)->getNumElements()) |
| 1223 | return nullptr; |
| 1224 | } else { |
| 1225 | auto *ST = cast<StructType>(Val: VT); |
| 1226 | if (ST->getNumElements() != cast<FixedVectorType>(Val: UT)->getNumElements()) |
| 1227 | return nullptr; |
| 1228 | for (const auto *EltT : ST->elements()) { |
| 1229 | if (EltT != UT->getElementType()) |
| 1230 | return nullptr; |
| 1231 | } |
| 1232 | } |
| 1233 | return U; |
| 1234 | } |
| 1235 | |
| 1236 | /// Combine stores to match the type of value being stored. |
| 1237 | /// |
| 1238 | /// The core idea here is that the memory does not have any intrinsic type and |
| 1239 | /// where we can we should match the type of a store to the type of value being |
| 1240 | /// stored. |
| 1241 | /// |
| 1242 | /// However, this routine must never change the width of a store or the number of |
| 1243 | /// stores as that would introduce a semantic change. This combine is expected to |
| 1244 | /// be a semantic no-op which just allows stores to more closely model the types |
| 1245 | /// of their incoming values. |
| 1246 | /// |
| 1247 | /// Currently, we also refuse to change the precise type used for an atomic or |
| 1248 | /// volatile store. This is debatable, and might be reasonable to change later. |
| 1249 | /// However, it is risky in case some backend or other part of LLVM is relying |
| 1250 | /// on the exact type stored to select appropriate atomic operations. |
| 1251 | /// |
| 1252 | /// \returns true if the store was successfully combined away. This indicates |
| 1253 | /// the caller must erase the store instruction. We have to let the caller erase |
| 1254 | /// the store instruction as otherwise there is no way to signal whether it was |
| 1255 | /// combined or not: IC.EraseInstFromFunction returns a null pointer. |
| 1256 | static bool combineStoreToValueType(InstCombinerImpl &IC, StoreInst &SI) { |
| 1257 | // FIXME: We could probably with some care handle both volatile and ordered |
| 1258 | // atomic stores here but it isn't clear that this is important. |
| 1259 | if (!SI.isUnordered()) |
| 1260 | return false; |
| 1261 | |
| 1262 | // swifterror values can't be bitcasted. |
| 1263 | if (SI.getPointerOperand()->isSwiftError()) |
| 1264 | return false; |
| 1265 | |
| 1266 | Value *V = SI.getValueOperand(); |
| 1267 | |
| 1268 | // Fold away bit casts of the stored value by storing the original type. |
| 1269 | if (auto *BC = dyn_cast<BitCastInst>(Val: V)) { |
| 1270 | assert(!BC->getType()->isX86_AMXTy() && |
| 1271 | "store to x86_amx* should not happen!" ); |
| 1272 | V = BC->getOperand(i_nocapture: 0); |
| 1273 | // Don't transform when the type is x86_amx, it makes the pass that lower |
| 1274 | // x86_amx type happy. |
| 1275 | if (V->getType()->isX86_AMXTy()) |
| 1276 | return false; |
| 1277 | if (!SI.isAtomic() || isSupportedAtomicType(Ty: V->getType())) { |
| 1278 | combineStoreToNewValue(IC, SI, V); |
| 1279 | return true; |
| 1280 | } |
| 1281 | } |
| 1282 | |
| 1283 | if (Value *U = likeBitCastFromVector(IC, V)) |
| 1284 | if (!SI.isAtomic() || isSupportedAtomicType(Ty: U->getType())) { |
| 1285 | combineStoreToNewValue(IC, SI, V: U); |
| 1286 | return true; |
| 1287 | } |
| 1288 | |
| 1289 | // FIXME: We should also canonicalize stores of vectors when their elements |
| 1290 | // are cast to other types. |
| 1291 | return false; |
| 1292 | } |
| 1293 | |
| 1294 | static bool unpackStoreToAggregate(InstCombinerImpl &IC, StoreInst &SI) { |
| 1295 | // FIXME: We could probably with some care handle both volatile and atomic |
| 1296 | // stores here but it isn't clear that this is important. |
| 1297 | if (!SI.isSimple()) |
| 1298 | return false; |
| 1299 | |
| 1300 | Value *V = SI.getValueOperand(); |
| 1301 | Type *T = V->getType(); |
| 1302 | |
| 1303 | if (!T->isAggregateType()) |
| 1304 | return false; |
| 1305 | |
| 1306 | if (auto *ST = dyn_cast<StructType>(Val: T)) { |
| 1307 | // If the struct only have one element, we unpack. |
| 1308 | unsigned Count = ST->getNumElements(); |
| 1309 | if (Count == 1) { |
| 1310 | V = IC.Builder.CreateExtractValue(Agg: V, Idxs: 0); |
| 1311 | combineStoreToNewValue(IC, SI, V); |
| 1312 | return true; |
| 1313 | } |
| 1314 | |
| 1315 | // We don't want to break loads with padding here as we'd loose |
| 1316 | // the knowledge that padding exists for the rest of the pipeline. |
| 1317 | const DataLayout &DL = IC.getDataLayout(); |
| 1318 | auto *SL = DL.getStructLayout(Ty: ST); |
| 1319 | |
| 1320 | if (SL->hasPadding()) |
| 1321 | return false; |
| 1322 | |
| 1323 | const auto Align = SI.getAlign(); |
| 1324 | |
| 1325 | SmallString<16> EltName = V->getName(); |
| 1326 | EltName += ".elt" ; |
| 1327 | auto *Addr = SI.getPointerOperand(); |
| 1328 | SmallString<16> AddrName = Addr->getName(); |
| 1329 | AddrName += ".repack" ; |
| 1330 | |
| 1331 | auto *IdxType = DL.getIndexType(PtrTy: Addr->getType()); |
| 1332 | for (unsigned i = 0; i < Count; i++) { |
| 1333 | auto *Ptr = IC.Builder.CreateInBoundsPtrAdd( |
| 1334 | Ptr: Addr, Offset: IC.Builder.CreateTypeSize(Ty: IdxType, Size: SL->getElementOffset(Idx: i)), |
| 1335 | Name: AddrName); |
| 1336 | auto *Val = IC.Builder.CreateExtractValue(Agg: V, Idxs: i, Name: EltName); |
| 1337 | auto EltAlign = |
| 1338 | commonAlignment(A: Align, Offset: SL->getElementOffset(Idx: i).getKnownMinValue()); |
| 1339 | llvm::Instruction *NS = IC.Builder.CreateAlignedStore(Val, Ptr, Align: EltAlign); |
| 1340 | NS->setAAMetadata(SI.getAAMetadata()); |
| 1341 | } |
| 1342 | |
| 1343 | return true; |
| 1344 | } |
| 1345 | |
| 1346 | if (auto *AT = dyn_cast<ArrayType>(Val: T)) { |
| 1347 | // If the array only have one element, we unpack. |
| 1348 | auto NumElements = AT->getNumElements(); |
| 1349 | if (NumElements == 1) { |
| 1350 | V = IC.Builder.CreateExtractValue(Agg: V, Idxs: 0); |
| 1351 | combineStoreToNewValue(IC, SI, V); |
| 1352 | return true; |
| 1353 | } |
| 1354 | |
| 1355 | // Bail out if the array is too large. Ideally we would like to optimize |
| 1356 | // arrays of arbitrary size but this has a terrible impact on compile time. |
| 1357 | // The threshold here is chosen arbitrarily, maybe needs a little bit of |
| 1358 | // tuning. |
| 1359 | if (NumElements > IC.MaxArraySizeForCombine) |
| 1360 | return false; |
| 1361 | |
| 1362 | const DataLayout &DL = IC.getDataLayout(); |
| 1363 | TypeSize EltSize = DL.getTypeAllocSize(Ty: AT->getElementType()); |
| 1364 | const auto Align = SI.getAlign(); |
| 1365 | |
| 1366 | SmallString<16> EltName = V->getName(); |
| 1367 | EltName += ".elt" ; |
| 1368 | auto *Addr = SI.getPointerOperand(); |
| 1369 | SmallString<16> AddrName = Addr->getName(); |
| 1370 | AddrName += ".repack" ; |
| 1371 | |
| 1372 | auto *IdxType = Type::getInt64Ty(C&: T->getContext()); |
| 1373 | auto *Zero = ConstantInt::get(Ty: IdxType, V: 0); |
| 1374 | |
| 1375 | TypeSize Offset = TypeSize::getZero(); |
| 1376 | for (uint64_t i = 0; i < NumElements; i++) { |
| 1377 | Value *Indices[2] = { |
| 1378 | Zero, |
| 1379 | ConstantInt::get(Ty: IdxType, V: i), |
| 1380 | }; |
| 1381 | auto *Ptr = |
| 1382 | IC.Builder.CreateInBoundsGEP(Ty: AT, Ptr: Addr, IdxList: ArrayRef(Indices), Name: AddrName); |
| 1383 | auto *Val = IC.Builder.CreateExtractValue(Agg: V, Idxs: i, Name: EltName); |
| 1384 | auto EltAlign = commonAlignment(A: Align, Offset: Offset.getKnownMinValue()); |
| 1385 | Instruction *NS = IC.Builder.CreateAlignedStore(Val, Ptr, Align: EltAlign); |
| 1386 | NS->setAAMetadata(SI.getAAMetadata()); |
| 1387 | Offset += EltSize; |
| 1388 | } |
| 1389 | |
| 1390 | return true; |
| 1391 | } |
| 1392 | |
| 1393 | return false; |
| 1394 | } |
| 1395 | |
| 1396 | /// equivalentAddressValues - Test if A and B will obviously have the same |
| 1397 | /// value. This includes recognizing that %t0 and %t1 will have the same |
| 1398 | /// value in code like this: |
| 1399 | /// %t0 = getelementptr \@a, 0, 3 |
| 1400 | /// store i32 0, i32* %t0 |
| 1401 | /// %t1 = getelementptr \@a, 0, 3 |
| 1402 | /// %t2 = load i32* %t1 |
| 1403 | /// |
| 1404 | static bool equivalentAddressValues(Value *A, Value *B) { |
| 1405 | // Test if the values are trivially equivalent. |
| 1406 | if (A == B) return true; |
| 1407 | |
| 1408 | // Test if the values come form identical arithmetic instructions. |
| 1409 | // This uses isIdenticalToWhenDefined instead of isIdenticalTo because |
| 1410 | // its only used to compare two uses within the same basic block, which |
| 1411 | // means that they'll always either have the same value or one of them |
| 1412 | // will have an undefined value. |
| 1413 | if (isa<BinaryOperator>(Val: A) || |
| 1414 | isa<CastInst>(Val: A) || |
| 1415 | isa<PHINode>(Val: A) || |
| 1416 | isa<GetElementPtrInst>(Val: A)) |
| 1417 | if (Instruction *BI = dyn_cast<Instruction>(Val: B)) |
| 1418 | if (cast<Instruction>(Val: A)->isIdenticalToWhenDefined(I: BI)) |
| 1419 | return true; |
| 1420 | |
| 1421 | // Otherwise they may not be equivalent. |
| 1422 | return false; |
| 1423 | } |
| 1424 | |
| 1425 | Instruction *InstCombinerImpl::visitStoreInst(StoreInst &SI) { |
| 1426 | Value *Val = SI.getOperand(i_nocapture: 0); |
| 1427 | Value *Ptr = SI.getOperand(i_nocapture: 1); |
| 1428 | |
| 1429 | // Try to canonicalize the stored type. |
| 1430 | if (combineStoreToValueType(IC&: *this, SI)) |
| 1431 | return eraseInstFromFunction(I&: SI); |
| 1432 | |
| 1433 | // Try to canonicalize the stored type. |
| 1434 | if (unpackStoreToAggregate(IC&: *this, SI)) |
| 1435 | return eraseInstFromFunction(I&: SI); |
| 1436 | |
| 1437 | // Replace GEP indices if possible. |
| 1438 | if (Instruction *NewGEPI = replaceGEPIdxWithZero(IC&: *this, Ptr, MemI&: SI)) |
| 1439 | return replaceOperand(I&: SI, OpNum: 1, V: NewGEPI); |
| 1440 | |
| 1441 | // Don't hack volatile/ordered stores. |
| 1442 | // FIXME: Some bits are legal for ordered atomic stores; needs refactoring. |
| 1443 | if (!SI.isUnordered()) return nullptr; |
| 1444 | |
| 1445 | // If the RHS is an alloca with a single use, zapify the store, making the |
| 1446 | // alloca dead. |
| 1447 | if (Ptr->hasOneUse()) { |
| 1448 | if (isa<AllocaInst>(Val: Ptr)) |
| 1449 | return eraseInstFromFunction(I&: SI); |
| 1450 | if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Val: Ptr)) { |
| 1451 | if (isa<AllocaInst>(Val: GEP->getOperand(i_nocapture: 0))) { |
| 1452 | if (GEP->getOperand(i_nocapture: 0)->hasOneUse()) |
| 1453 | return eraseInstFromFunction(I&: SI); |
| 1454 | } |
| 1455 | } |
| 1456 | } |
| 1457 | |
| 1458 | // If we have a store to a location which is known constant, we can conclude |
| 1459 | // that the store must be storing the constant value (else the memory |
| 1460 | // wouldn't be constant), and this must be a noop. |
| 1461 | if (!isModSet(MRI: AA->getModRefInfoMask(P: Ptr))) |
| 1462 | return eraseInstFromFunction(I&: SI); |
| 1463 | |
| 1464 | // Do really simple DSE, to catch cases where there are several consecutive |
| 1465 | // stores to the same location, separated by a few arithmetic operations. This |
| 1466 | // situation often occurs with bitfield accesses. |
| 1467 | BasicBlock::iterator BBI(SI); |
| 1468 | for (unsigned ScanInsts = 6; BBI != SI.getParent()->begin() && ScanInsts; |
| 1469 | --ScanInsts) { |
| 1470 | --BBI; |
| 1471 | // Don't count debug info directives, lest they affect codegen, |
| 1472 | // and we skip pointer-to-pointer bitcasts, which are NOPs. |
| 1473 | if (BBI->isDebugOrPseudoInst()) { |
| 1474 | ScanInsts++; |
| 1475 | continue; |
| 1476 | } |
| 1477 | |
| 1478 | if (StoreInst *PrevSI = dyn_cast<StoreInst>(Val&: BBI)) { |
| 1479 | // Prev store isn't volatile, and stores to the same location? |
| 1480 | if (PrevSI->isUnordered() && |
| 1481 | equivalentAddressValues(A: PrevSI->getOperand(i_nocapture: 1), B: SI.getOperand(i_nocapture: 1)) && |
| 1482 | PrevSI->getValueOperand()->getType() == |
| 1483 | SI.getValueOperand()->getType()) { |
| 1484 | ++NumDeadStore; |
| 1485 | // Manually add back the original store to the worklist now, so it will |
| 1486 | // be processed after the operands of the removed store, as this may |
| 1487 | // expose additional DSE opportunities. |
| 1488 | Worklist.push(I: &SI); |
| 1489 | eraseInstFromFunction(I&: *PrevSI); |
| 1490 | return nullptr; |
| 1491 | } |
| 1492 | break; |
| 1493 | } |
| 1494 | |
| 1495 | // If this is a load, we have to stop. However, if the loaded value is from |
| 1496 | // the pointer we're loading and is producing the pointer we're storing, |
| 1497 | // then *this* store is dead (X = load P; store X -> P). |
| 1498 | if (LoadInst *LI = dyn_cast<LoadInst>(Val&: BBI)) { |
| 1499 | if (LI == Val && equivalentAddressValues(A: LI->getOperand(i_nocapture: 0), B: Ptr)) { |
| 1500 | assert(SI.isUnordered() && "can't eliminate ordering operation" ); |
| 1501 | return eraseInstFromFunction(I&: SI); |
| 1502 | } |
| 1503 | |
| 1504 | // Otherwise, this is a load from some other location. Stores before it |
| 1505 | // may not be dead. |
| 1506 | break; |
| 1507 | } |
| 1508 | |
| 1509 | // Don't skip over loads, throws or things that can modify memory. |
| 1510 | if (BBI->mayWriteToMemory() || BBI->mayReadFromMemory() || BBI->mayThrow()) |
| 1511 | break; |
| 1512 | } |
| 1513 | |
| 1514 | // store X, null -> turns into 'unreachable' in SimplifyCFG |
| 1515 | // store X, GEP(null, Y) -> turns into 'unreachable' in SimplifyCFG |
| 1516 | if (canSimplifyNullStoreOrGEP(SI)) { |
| 1517 | if (!isa<PoisonValue>(Val)) |
| 1518 | return replaceOperand(I&: SI, OpNum: 0, V: PoisonValue::get(T: Val->getType())); |
| 1519 | return nullptr; // Do not modify these! |
| 1520 | } |
| 1521 | |
| 1522 | // This is a non-terminator unreachable marker. Don't remove it. |
| 1523 | if (isa<UndefValue>(Val: Ptr)) { |
| 1524 | // Remove guaranteed-to-transfer instructions before the marker. |
| 1525 | removeInstructionsBeforeUnreachable(I&: SI); |
| 1526 | |
| 1527 | // Remove all instructions after the marker and handle dead blocks this |
| 1528 | // implies. |
| 1529 | SmallVector<BasicBlock *> Worklist; |
| 1530 | handleUnreachableFrom(I: SI.getNextNode(), Worklist); |
| 1531 | handlePotentiallyDeadBlocks(Worklist); |
| 1532 | return nullptr; |
| 1533 | } |
| 1534 | |
| 1535 | // store undef, Ptr -> noop |
| 1536 | // FIXME: This is technically incorrect because it might overwrite a poison |
| 1537 | // value. Change to PoisonValue once #52930 is resolved. |
| 1538 | if (isa<UndefValue>(Val)) |
| 1539 | return eraseInstFromFunction(I&: SI); |
| 1540 | |
| 1541 | if (!NullPointerIsDefined(F: SI.getFunction(), AS: SI.getPointerAddressSpace())) |
| 1542 | if (Value *V = simplifyNonNullOperand(V: Ptr, /*HasDereferenceable=*/true)) |
| 1543 | return replaceOperand(I&: SI, OpNum: 1, V); |
| 1544 | |
| 1545 | return nullptr; |
| 1546 | } |
| 1547 | |
| 1548 | /// Try to transform: |
| 1549 | /// if () { *P = v1; } else { *P = v2 } |
| 1550 | /// or: |
| 1551 | /// *P = v1; if () { *P = v2; } |
| 1552 | /// into a phi node with a store in the successor. |
| 1553 | bool InstCombinerImpl::mergeStoreIntoSuccessor(StoreInst &SI) { |
| 1554 | if (!SI.isUnordered()) |
| 1555 | return false; // This code has not been audited for volatile/ordered case. |
| 1556 | |
| 1557 | // Check if the successor block has exactly 2 incoming edges. |
| 1558 | BasicBlock *StoreBB = SI.getParent(); |
| 1559 | BasicBlock *DestBB = StoreBB->getTerminator()->getSuccessor(Idx: 0); |
| 1560 | if (!DestBB->hasNPredecessors(N: 2)) |
| 1561 | return false; |
| 1562 | |
| 1563 | // Capture the other block (the block that doesn't contain our store). |
| 1564 | pred_iterator PredIter = pred_begin(BB: DestBB); |
| 1565 | if (*PredIter == StoreBB) |
| 1566 | ++PredIter; |
| 1567 | BasicBlock *OtherBB = *PredIter; |
| 1568 | |
| 1569 | // Bail out if all of the relevant blocks aren't distinct. This can happen, |
| 1570 | // for example, if SI is in an infinite loop. |
| 1571 | if (StoreBB == DestBB || OtherBB == DestBB) |
| 1572 | return false; |
| 1573 | |
| 1574 | // Verify that the other block ends in a branch and is not otherwise empty. |
| 1575 | BasicBlock::iterator BBI(OtherBB->getTerminator()); |
| 1576 | BranchInst *OtherBr = dyn_cast<BranchInst>(Val&: BBI); |
| 1577 | if (!OtherBr || BBI == OtherBB->begin()) |
| 1578 | return false; |
| 1579 | |
| 1580 | auto OtherStoreIsMergeable = [&](StoreInst *OtherStore) -> bool { |
| 1581 | if (!OtherStore || |
| 1582 | OtherStore->getPointerOperand() != SI.getPointerOperand()) |
| 1583 | return false; |
| 1584 | |
| 1585 | auto *SIVTy = SI.getValueOperand()->getType(); |
| 1586 | auto *OSVTy = OtherStore->getValueOperand()->getType(); |
| 1587 | return CastInst::isBitOrNoopPointerCastable(SrcTy: OSVTy, DestTy: SIVTy, DL) && |
| 1588 | SI.hasSameSpecialState(I2: OtherStore); |
| 1589 | }; |
| 1590 | |
| 1591 | // If the other block ends in an unconditional branch, check for the 'if then |
| 1592 | // else' case. There is an instruction before the branch. |
| 1593 | StoreInst *OtherStore = nullptr; |
| 1594 | if (OtherBr->isUnconditional()) { |
| 1595 | --BBI; |
| 1596 | // Skip over debugging info and pseudo probes. |
| 1597 | while (BBI->isDebugOrPseudoInst()) { |
| 1598 | if (BBI==OtherBB->begin()) |
| 1599 | return false; |
| 1600 | --BBI; |
| 1601 | } |
| 1602 | // If this isn't a store, isn't a store to the same location, or is not the |
| 1603 | // right kind of store, bail out. |
| 1604 | OtherStore = dyn_cast<StoreInst>(Val&: BBI); |
| 1605 | if (!OtherStoreIsMergeable(OtherStore)) |
| 1606 | return false; |
| 1607 | } else { |
| 1608 | // Otherwise, the other block ended with a conditional branch. If one of the |
| 1609 | // destinations is StoreBB, then we have the if/then case. |
| 1610 | if (OtherBr->getSuccessor(i: 0) != StoreBB && |
| 1611 | OtherBr->getSuccessor(i: 1) != StoreBB) |
| 1612 | return false; |
| 1613 | |
| 1614 | // Okay, we know that OtherBr now goes to Dest and StoreBB, so this is an |
| 1615 | // if/then triangle. See if there is a store to the same ptr as SI that |
| 1616 | // lives in OtherBB. |
| 1617 | for (;; --BBI) { |
| 1618 | // Check to see if we find the matching store. |
| 1619 | OtherStore = dyn_cast<StoreInst>(Val&: BBI); |
| 1620 | if (OtherStoreIsMergeable(OtherStore)) |
| 1621 | break; |
| 1622 | |
| 1623 | // If we find something that may be using or overwriting the stored |
| 1624 | // value, or if we run out of instructions, we can't do the transform. |
| 1625 | if (BBI->mayReadFromMemory() || BBI->mayThrow() || |
| 1626 | BBI->mayWriteToMemory() || BBI == OtherBB->begin()) |
| 1627 | return false; |
| 1628 | } |
| 1629 | |
| 1630 | // In order to eliminate the store in OtherBr, we have to make sure nothing |
| 1631 | // reads or overwrites the stored value in StoreBB. |
| 1632 | for (BasicBlock::iterator I = StoreBB->begin(); &*I != &SI; ++I) { |
| 1633 | // FIXME: This should really be AA driven. |
| 1634 | if (I->mayReadFromMemory() || I->mayThrow() || I->mayWriteToMemory()) |
| 1635 | return false; |
| 1636 | } |
| 1637 | } |
| 1638 | |
| 1639 | // Insert a PHI node now if we need it. |
| 1640 | Value *MergedVal = OtherStore->getValueOperand(); |
| 1641 | // The debug locations of the original instructions might differ. Merge them. |
| 1642 | DebugLoc MergedLoc = |
| 1643 | DebugLoc::getMergedLocation(LocA: SI.getDebugLoc(), LocB: OtherStore->getDebugLoc()); |
| 1644 | if (MergedVal != SI.getValueOperand()) { |
| 1645 | PHINode *PN = |
| 1646 | PHINode::Create(Ty: SI.getValueOperand()->getType(), NumReservedValues: 2, NameStr: "storemerge" ); |
| 1647 | PN->addIncoming(V: SI.getValueOperand(), BB: SI.getParent()); |
| 1648 | Builder.SetInsertPoint(OtherStore); |
| 1649 | PN->addIncoming(V: Builder.CreateBitOrPointerCast(V: MergedVal, DestTy: PN->getType()), |
| 1650 | BB: OtherBB); |
| 1651 | MergedVal = InsertNewInstBefore(New: PN, Old: DestBB->begin()); |
| 1652 | PN->setDebugLoc(MergedLoc); |
| 1653 | } |
| 1654 | |
| 1655 | // Advance to a place where it is safe to insert the new store and insert it. |
| 1656 | BBI = DestBB->getFirstInsertionPt(); |
| 1657 | StoreInst *NewSI = |
| 1658 | new StoreInst(MergedVal, SI.getOperand(i_nocapture: 1), SI.isVolatile(), SI.getAlign(), |
| 1659 | SI.getOrdering(), SI.getSyncScopeID()); |
| 1660 | InsertNewInstBefore(New: NewSI, Old: BBI); |
| 1661 | NewSI->setDebugLoc(MergedLoc); |
| 1662 | NewSI->mergeDIAssignID(SourceInstructions: {&SI, OtherStore}); |
| 1663 | |
| 1664 | // If the two stores had AA tags, merge them. |
| 1665 | AAMDNodes AATags = SI.getAAMetadata(); |
| 1666 | if (AATags) |
| 1667 | NewSI->setAAMetadata(AATags.merge(Other: OtherStore->getAAMetadata())); |
| 1668 | |
| 1669 | // Nuke the old stores. |
| 1670 | eraseInstFromFunction(I&: SI); |
| 1671 | eraseInstFromFunction(I&: *OtherStore); |
| 1672 | return true; |
| 1673 | } |
| 1674 | |