| 1 | //===- LoadStoreVectorizer.cpp - GPU Load & Store Vectorizer --------------===// |
| 2 | // |
| 3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
| 4 | // See https://llvm.org/LICENSE.txt for license information. |
| 5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
| 6 | // |
| 7 | //===----------------------------------------------------------------------===// |
| 8 | // |
| 9 | // This pass merges loads/stores to/from sequential memory addresses into vector |
| 10 | // loads/stores. Although there's nothing GPU-specific in here, this pass is |
| 11 | // motivated by the microarchitectural quirks of nVidia and AMD GPUs. |
| 12 | // |
| 13 | // (For simplicity below we talk about loads only, but everything also applies |
| 14 | // to stores.) |
| 15 | // |
| 16 | // This pass is intended to be run late in the pipeline, after other |
| 17 | // vectorization opportunities have been exploited. So the assumption here is |
| 18 | // that immediately following our new vector load we'll need to extract out the |
| 19 | // individual elements of the load, so we can operate on them individually. |
| 20 | // |
| 21 | // On CPUs this transformation is usually not beneficial, because extracting the |
| 22 | // elements of a vector register is expensive on most architectures. It's |
| 23 | // usually better just to load each element individually into its own scalar |
| 24 | // register. |
| 25 | // |
| 26 | // However, nVidia and AMD GPUs don't have proper vector registers. Instead, a |
| 27 | // "vector load" loads directly into a series of scalar registers. In effect, |
| 28 | // extracting the elements of the vector is free. It's therefore always |
| 29 | // beneficial to vectorize a sequence of loads on these architectures. |
| 30 | // |
| 31 | // Vectorizing (perhaps a better name might be "coalescing") loads can have |
| 32 | // large performance impacts on GPU kernels, and opportunities for vectorizing |
| 33 | // are common in GPU code. This pass tries very hard to find such |
| 34 | // opportunities; its runtime is quadratic in the number of loads in a BB. |
| 35 | // |
| 36 | // Some CPU architectures, such as ARM, have instructions that load into |
| 37 | // multiple scalar registers, similar to a GPU vectorized load. In theory ARM |
| 38 | // could use this pass (with some modifications), but currently it implements |
| 39 | // its own pass to do something similar to what we do here. |
| 40 | // |
| 41 | // Overview of the algorithm and terminology in this pass: |
| 42 | // |
| 43 | // - Break up each basic block into pseudo-BBs, composed of instructions which |
| 44 | // are guaranteed to transfer control to their successors. |
| 45 | // - Within a single pseudo-BB, find all loads, and group them into |
| 46 | // "equivalence classes" according to getUnderlyingObject() and loaded |
| 47 | // element size. Do the same for stores. |
| 48 | // - For each equivalence class, greedily build "chains". Each chain has a |
| 49 | // leader instruction, and every other member of the chain has a known |
| 50 | // constant offset from the first instr in the chain. |
| 51 | // - Break up chains so that they contain only contiguous accesses of legal |
| 52 | // size with no intervening may-alias instrs. |
| 53 | // - Convert each chain to vector instructions. |
| 54 | // |
| 55 | // The O(n^2) behavior of this pass comes from initially building the chains. |
| 56 | // In the worst case we have to compare each new instruction to all of those |
| 57 | // that came before. To limit this, we only calculate the offset to the leaders |
| 58 | // of the N most recently-used chains. |
| 59 | |
| 60 | #include "llvm/Transforms/Vectorize/LoadStoreVectorizer.h" |
| 61 | #include "llvm/ADT/APInt.h" |
| 62 | #include "llvm/ADT/ArrayRef.h" |
| 63 | #include "llvm/ADT/DenseMap.h" |
| 64 | #include "llvm/ADT/MapVector.h" |
| 65 | #include "llvm/ADT/PostOrderIterator.h" |
| 66 | #include "llvm/ADT/STLExtras.h" |
| 67 | #include "llvm/ADT/Sequence.h" |
| 68 | #include "llvm/ADT/SmallPtrSet.h" |
| 69 | #include "llvm/ADT/SmallVector.h" |
| 70 | #include "llvm/ADT/Statistic.h" |
| 71 | #include "llvm/ADT/iterator_range.h" |
| 72 | #include "llvm/Analysis/AliasAnalysis.h" |
| 73 | #include "llvm/Analysis/AssumptionCache.h" |
| 74 | #include "llvm/Analysis/MemoryLocation.h" |
| 75 | #include "llvm/Analysis/ScalarEvolution.h" |
| 76 | #include "llvm/Analysis/TargetTransformInfo.h" |
| 77 | #include "llvm/Analysis/ValueTracking.h" |
| 78 | #include "llvm/Analysis/VectorUtils.h" |
| 79 | #include "llvm/IR/Attributes.h" |
| 80 | #include "llvm/IR/BasicBlock.h" |
| 81 | #include "llvm/IR/ConstantRange.h" |
| 82 | #include "llvm/IR/Constants.h" |
| 83 | #include "llvm/IR/DataLayout.h" |
| 84 | #include "llvm/IR/DerivedTypes.h" |
| 85 | #include "llvm/IR/Dominators.h" |
| 86 | #include "llvm/IR/Function.h" |
| 87 | #include "llvm/IR/GetElementPtrTypeIterator.h" |
| 88 | #include "llvm/IR/IRBuilder.h" |
| 89 | #include "llvm/IR/InstrTypes.h" |
| 90 | #include "llvm/IR/Instruction.h" |
| 91 | #include "llvm/IR/Instructions.h" |
| 92 | #include "llvm/IR/LLVMContext.h" |
| 93 | #include "llvm/IR/Module.h" |
| 94 | #include "llvm/IR/Type.h" |
| 95 | #include "llvm/IR/Value.h" |
| 96 | #include "llvm/InitializePasses.h" |
| 97 | #include "llvm/Pass.h" |
| 98 | #include "llvm/Support/Alignment.h" |
| 99 | #include "llvm/Support/Casting.h" |
| 100 | #include "llvm/Support/Debug.h" |
| 101 | #include "llvm/Support/KnownBits.h" |
| 102 | #include "llvm/Support/MathExtras.h" |
| 103 | #include "llvm/Support/ModRef.h" |
| 104 | #include "llvm/Support/raw_ostream.h" |
| 105 | #include "llvm/Transforms/Utils/Local.h" |
| 106 | #include <algorithm> |
| 107 | #include <cassert> |
| 108 | #include <cstdint> |
| 109 | #include <cstdlib> |
| 110 | #include <iterator> |
| 111 | #include <numeric> |
| 112 | #include <optional> |
| 113 | #include <tuple> |
| 114 | #include <type_traits> |
| 115 | #include <utility> |
| 116 | #include <vector> |
| 117 | |
| 118 | using namespace llvm; |
| 119 | |
| 120 | #define DEBUG_TYPE "load-store-vectorizer" |
| 121 | |
| 122 | STATISTIC(NumVectorInstructions, "Number of vector accesses generated" ); |
| 123 | STATISTIC(NumScalarsVectorized, "Number of scalar accesses vectorized" ); |
| 124 | |
| 125 | namespace { |
| 126 | |
| 127 | // Equivalence class key, the initial tuple by which we group loads/stores. |
| 128 | // Loads/stores with different EqClassKeys are never merged. |
| 129 | // |
| 130 | // (We could in theory remove element-size from the this tuple. We'd just need |
| 131 | // to fix up the vector packing/unpacking code.) |
| 132 | using EqClassKey = |
| 133 | std::tuple<const Value * /* result of getUnderlyingObject() */, |
| 134 | unsigned /* AddrSpace */, |
| 135 | unsigned /* Load/Store element size bits */, |
| 136 | char /* IsLoad; char b/c bool can't be a DenseMap key */ |
| 137 | >; |
| 138 | [[maybe_unused]] llvm::raw_ostream &operator<<(llvm::raw_ostream &OS, |
| 139 | const EqClassKey &K) { |
| 140 | const auto &[UnderlyingObject, AddrSpace, ElementSize, IsLoad] = K; |
| 141 | return OS << (IsLoad ? "load" : "store" ) << " of " << *UnderlyingObject |
| 142 | << " of element size " << ElementSize << " bits in addrspace " |
| 143 | << AddrSpace; |
| 144 | } |
| 145 | |
| 146 | // A Chain is a set of instructions such that: |
| 147 | // - All instructions have the same equivalence class, so in particular all are |
| 148 | // loads, or all are stores. |
| 149 | // - We know the address accessed by the i'th chain elem relative to the |
| 150 | // chain's leader instruction, which is the first instr of the chain in BB |
| 151 | // order. |
| 152 | // |
| 153 | // Chains have two canonical orderings: |
| 154 | // - BB order, sorted by Instr->comesBefore. |
| 155 | // - Offset order, sorted by OffsetFromLeader. |
| 156 | // This pass switches back and forth between these orders. |
| 157 | struct ChainElem { |
| 158 | Instruction *Inst; |
| 159 | APInt OffsetFromLeader; |
| 160 | ChainElem(Instruction *Inst, APInt OffsetFromLeader) |
| 161 | : Inst(std::move(Inst)), OffsetFromLeader(std::move(OffsetFromLeader)) {} |
| 162 | }; |
| 163 | using Chain = SmallVector<ChainElem, 1>; |
| 164 | |
| 165 | void sortChainInBBOrder(Chain &C) { |
| 166 | sort(C, Comp: [](auto &A, auto &B) { return A.Inst->comesBefore(B.Inst); }); |
| 167 | } |
| 168 | |
| 169 | void sortChainInOffsetOrder(Chain &C) { |
| 170 | sort(C, Comp: [](const auto &A, const auto &B) { |
| 171 | if (A.OffsetFromLeader != B.OffsetFromLeader) |
| 172 | return A.OffsetFromLeader.slt(B.OffsetFromLeader); |
| 173 | return A.Inst->comesBefore(B.Inst); // stable tiebreaker |
| 174 | }); |
| 175 | } |
| 176 | |
| 177 | [[maybe_unused]] void dumpChain(ArrayRef<ChainElem> C) { |
| 178 | for (const auto &E : C) { |
| 179 | dbgs() << " " << *E.Inst << " (offset " << E.OffsetFromLeader << ")\n" ; |
| 180 | } |
| 181 | } |
| 182 | |
| 183 | using EquivalenceClassMap = |
| 184 | MapVector<EqClassKey, SmallVector<Instruction *, 8>>; |
| 185 | |
| 186 | // FIXME: Assuming stack alignment of 4 is always good enough |
| 187 | constexpr unsigned StackAdjustedAlignment = 4; |
| 188 | |
| 189 | Instruction *propagateMetadata(Instruction *I, const Chain &C) { |
| 190 | SmallVector<Value *, 8> Values; |
| 191 | for (const ChainElem &E : C) |
| 192 | Values.emplace_back(Args: E.Inst); |
| 193 | return propagateMetadata(I, VL: Values); |
| 194 | } |
| 195 | |
| 196 | bool isInvariantLoad(const Instruction *I) { |
| 197 | const LoadInst *LI = dyn_cast<LoadInst>(Val: I); |
| 198 | return LI != nullptr && LI->hasMetadata(KindID: LLVMContext::MD_invariant_load); |
| 199 | } |
| 200 | |
| 201 | /// Reorders the instructions that I depends on (the instructions defining its |
| 202 | /// operands), to ensure they dominate I. |
| 203 | void reorder(Instruction *I) { |
| 204 | SmallPtrSet<Instruction *, 16> InstructionsToMove; |
| 205 | SmallVector<Instruction *, 16> Worklist; |
| 206 | |
| 207 | Worklist.emplace_back(Args&: I); |
| 208 | while (!Worklist.empty()) { |
| 209 | Instruction *IW = Worklist.pop_back_val(); |
| 210 | int NumOperands = IW->getNumOperands(); |
| 211 | for (int Idx = 0; Idx < NumOperands; Idx++) { |
| 212 | Instruction *IM = dyn_cast<Instruction>(Val: IW->getOperand(i: Idx)); |
| 213 | if (!IM || IM->getOpcode() == Instruction::PHI) |
| 214 | continue; |
| 215 | |
| 216 | // If IM is in another BB, no need to move it, because this pass only |
| 217 | // vectorizes instructions within one BB. |
| 218 | if (IM->getParent() != I->getParent()) |
| 219 | continue; |
| 220 | |
| 221 | assert(IM != I && "Unexpected cycle while re-ordering instructions" ); |
| 222 | |
| 223 | if (!IM->comesBefore(Other: I)) { |
| 224 | InstructionsToMove.insert(Ptr: IM); |
| 225 | Worklist.emplace_back(Args&: IM); |
| 226 | } |
| 227 | } |
| 228 | } |
| 229 | |
| 230 | // All instructions to move should follow I. Start from I, not from begin(). |
| 231 | for (auto BBI = I->getIterator(), E = I->getParent()->end(); BBI != E;) { |
| 232 | Instruction *IM = &*(BBI++); |
| 233 | if (!InstructionsToMove.contains(Ptr: IM)) |
| 234 | continue; |
| 235 | IM->moveBefore(InsertPos: I->getIterator()); |
| 236 | } |
| 237 | } |
| 238 | |
| 239 | class Vectorizer { |
| 240 | Function &F; |
| 241 | AliasAnalysis &AA; |
| 242 | AssumptionCache &AC; |
| 243 | DominatorTree &DT; |
| 244 | ScalarEvolution &SE; |
| 245 | TargetTransformInfo &TTI; |
| 246 | const DataLayout &DL; |
| 247 | IRBuilder<> Builder; |
| 248 | |
| 249 | /// We could erase instrs right after vectorizing them, but that can mess up |
| 250 | /// our BB iterators, and also can make the equivalence class keys point to |
| 251 | /// freed memory. This is fixable, but it's simpler just to wait until we're |
| 252 | /// done with the BB and erase all at once. |
| 253 | SmallVector<Instruction *, 128> ToErase; |
| 254 | |
| 255 | /// We insert load/store instructions and GEPs to fill gaps and extend chains |
| 256 | /// to enable vectorization. Keep track and delete them later. |
| 257 | DenseSet<Instruction *> ; |
| 258 | |
| 259 | public: |
| 260 | Vectorizer(Function &F, AliasAnalysis &AA, AssumptionCache &AC, |
| 261 | DominatorTree &DT, ScalarEvolution &SE, TargetTransformInfo &TTI) |
| 262 | : F(F), AA(AA), AC(AC), DT(DT), SE(SE), TTI(TTI), |
| 263 | DL(F.getDataLayout()), Builder(SE.getContext()) {} |
| 264 | |
| 265 | bool run(); |
| 266 | |
| 267 | private: |
| 268 | static const unsigned MaxDepth = 3; |
| 269 | |
| 270 | /// Runs the vectorizer on a "pseudo basic block", which is a range of |
| 271 | /// instructions [Begin, End) within one BB all of which have |
| 272 | /// isGuaranteedToTransferExecutionToSuccessor(I) == true. |
| 273 | bool runOnPseudoBB(BasicBlock::iterator Begin, BasicBlock::iterator End); |
| 274 | |
| 275 | /// Runs the vectorizer on one equivalence class, i.e. one set of loads/stores |
| 276 | /// in the same BB with the same value for getUnderlyingObject() etc. |
| 277 | bool runOnEquivalenceClass(const EqClassKey &EqClassKey, |
| 278 | ArrayRef<Instruction *> EqClass); |
| 279 | |
| 280 | /// Runs the vectorizer on one chain, i.e. a subset of an equivalence class |
| 281 | /// where all instructions access a known, constant offset from the first |
| 282 | /// instruction. |
| 283 | bool runOnChain(Chain &C); |
| 284 | |
| 285 | /// Splits the chain into subchains of instructions which read/write a |
| 286 | /// contiguous block of memory. Discards any length-1 subchains (because |
| 287 | /// there's nothing to vectorize in there). Also attempts to fill gaps with |
| 288 | /// "extra" elements to artificially make chains contiguous in some cases. |
| 289 | std::vector<Chain> splitChainByContiguity(Chain &C); |
| 290 | |
| 291 | /// Splits the chain into subchains where it's safe to hoist loads up to the |
| 292 | /// beginning of the sub-chain and it's safe to sink loads up to the end of |
| 293 | /// the sub-chain. Discards any length-1 subchains. Also attempts to extend |
| 294 | /// non-power-of-two chains by adding "extra" elements in some cases. |
| 295 | std::vector<Chain> splitChainByMayAliasInstrs(Chain &C); |
| 296 | |
| 297 | /// Splits the chain into subchains that make legal, aligned accesses. |
| 298 | /// Discards any length-1 subchains. |
| 299 | std::vector<Chain> splitChainByAlignment(Chain &C); |
| 300 | |
| 301 | /// Converts the instrs in the chain into a single vectorized load or store. |
| 302 | /// Adds the old scalar loads/stores to ToErase. |
| 303 | bool vectorizeChain(Chain &C); |
| 304 | |
| 305 | /// Tries to compute the offset in bytes PtrB - PtrA. |
| 306 | std::optional<APInt> getConstantOffset(Value *PtrA, Value *PtrB, |
| 307 | Instruction *ContextInst, |
| 308 | unsigned Depth = 0); |
| 309 | std::optional<APInt> getConstantOffsetComplexAddrs(Value *PtrA, Value *PtrB, |
| 310 | Instruction *ContextInst, |
| 311 | unsigned Depth); |
| 312 | std::optional<APInt> getConstantOffsetSelects(Value *PtrA, Value *PtrB, |
| 313 | Instruction *ContextInst, |
| 314 | unsigned Depth); |
| 315 | |
| 316 | /// Gets the element type of the vector that the chain will load or store. |
| 317 | /// This is nontrivial because the chain may contain elements of different |
| 318 | /// types; e.g. it's legal to have a chain that contains both i32 and float. |
| 319 | Type *getChainElemTy(const Chain &C); |
| 320 | |
| 321 | /// Determines whether ChainElem can be moved up (if IsLoad) or down (if |
| 322 | /// !IsLoad) to ChainBegin -- i.e. there are no intervening may-alias |
| 323 | /// instructions. |
| 324 | /// |
| 325 | /// The map ChainElemOffsets must contain all of the elements in |
| 326 | /// [ChainBegin, ChainElem] and their offsets from some arbitrary base |
| 327 | /// address. It's ok if it contains additional entries. |
| 328 | template <bool IsLoadChain> |
| 329 | bool isSafeToMove( |
| 330 | Instruction *ChainElem, Instruction *ChainBegin, |
| 331 | const DenseMap<Instruction *, APInt /*OffsetFromLeader*/> &ChainOffsets, |
| 332 | BatchAAResults &BatchAA); |
| 333 | |
| 334 | /// Merges the equivalence classes if they have underlying objects that differ |
| 335 | /// by one level of indirection (i.e., one is a getelementptr and the other is |
| 336 | /// the base pointer in that getelementptr). |
| 337 | void mergeEquivalenceClasses(EquivalenceClassMap &EQClasses) const; |
| 338 | |
| 339 | /// Collects loads and stores grouped by "equivalence class", where: |
| 340 | /// - all elements in an eq class are a load or all are a store, |
| 341 | /// - they all load/store the same element size (it's OK to have e.g. i8 and |
| 342 | /// <4 x i8> in the same class, but not i32 and <4 x i8>), and |
| 343 | /// - they all have the same value for getUnderlyingObject(). |
| 344 | EquivalenceClassMap collectEquivalenceClasses(BasicBlock::iterator Begin, |
| 345 | BasicBlock::iterator End); |
| 346 | |
| 347 | /// Partitions Instrs into "chains" where every instruction has a known |
| 348 | /// constant offset from the first instr in the chain. |
| 349 | /// |
| 350 | /// Postcondition: For all i, ret[i][0].second == 0, because the first instr |
| 351 | /// in the chain is the leader, and an instr touches distance 0 from itself. |
| 352 | std::vector<Chain> gatherChains(ArrayRef<Instruction *> Instrs); |
| 353 | |
| 354 | /// Checks if a potential vector load/store with a given alignment is allowed |
| 355 | /// and fast. Aligned accesses are always allowed and fast, while misaligned |
| 356 | /// accesses depend on TTI checks to determine whether they can and should be |
| 357 | /// vectorized or kept as element-wise accesses. |
| 358 | bool accessIsAllowedAndFast(unsigned SizeBytes, unsigned AS, Align Alignment, |
| 359 | unsigned VecElemBits) const; |
| 360 | |
| 361 | /// Create a new GEP and a new Load/Store instruction such that the GEP |
| 362 | /// is pointing at PrevElem + Offset. In the case of stores, store poison. |
| 363 | /// Extra elements will either be combined into a masked load/store or |
| 364 | /// deleted before the end of the pass. |
| 365 | ChainElem createExtraElementAfter(const ChainElem &PrevElem, Type *Ty, |
| 366 | APInt Offset, StringRef Prefix, |
| 367 | Align Alignment = Align()); |
| 368 | |
| 369 | /// Create a mask that masks off the extra elements in the chain, to be used |
| 370 | /// for the creation of a masked load/store vector. |
| 371 | Value *createMaskForExtraElements(const ArrayRef<ChainElem> C, |
| 372 | FixedVectorType *VecTy); |
| 373 | |
| 374 | /// Delete dead GEPs and extra Load/Store instructions created by |
| 375 | /// createExtraElementAfter |
| 376 | void deleteExtraElements(); |
| 377 | }; |
| 378 | |
| 379 | class LoadStoreVectorizerLegacyPass : public FunctionPass { |
| 380 | public: |
| 381 | static char ID; |
| 382 | |
| 383 | LoadStoreVectorizerLegacyPass() : FunctionPass(ID) {} |
| 384 | |
| 385 | bool runOnFunction(Function &F) override; |
| 386 | |
| 387 | StringRef getPassName() const override { |
| 388 | return "GPU Load and Store Vectorizer" ; |
| 389 | } |
| 390 | |
| 391 | void getAnalysisUsage(AnalysisUsage &AU) const override { |
| 392 | AU.addRequired<AAResultsWrapperPass>(); |
| 393 | AU.addRequired<AssumptionCacheTracker>(); |
| 394 | AU.addRequired<ScalarEvolutionWrapperPass>(); |
| 395 | AU.addRequired<DominatorTreeWrapperPass>(); |
| 396 | AU.addRequired<TargetTransformInfoWrapperPass>(); |
| 397 | AU.setPreservesCFG(); |
| 398 | } |
| 399 | }; |
| 400 | |
| 401 | } // end anonymous namespace |
| 402 | |
| 403 | char LoadStoreVectorizerLegacyPass::ID = 0; |
| 404 | |
| 405 | INITIALIZE_PASS_BEGIN(LoadStoreVectorizerLegacyPass, DEBUG_TYPE, |
| 406 | "Vectorize load and Store instructions" , false, false) |
| 407 | INITIALIZE_PASS_DEPENDENCY(SCEVAAWrapperPass) |
| 408 | INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker); |
| 409 | INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass) |
| 410 | INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass) |
| 411 | INITIALIZE_PASS_DEPENDENCY(GlobalsAAWrapperPass) |
| 412 | INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass) |
| 413 | INITIALIZE_PASS_END(LoadStoreVectorizerLegacyPass, DEBUG_TYPE, |
| 414 | "Vectorize load and store instructions" , false, false) |
| 415 | |
| 416 | Pass *llvm::createLoadStoreVectorizerPass() { |
| 417 | return new LoadStoreVectorizerLegacyPass(); |
| 418 | } |
| 419 | |
| 420 | bool LoadStoreVectorizerLegacyPass::runOnFunction(Function &F) { |
| 421 | // Don't vectorize when the attribute NoImplicitFloat is used. |
| 422 | if (skipFunction(F) || F.hasFnAttribute(Kind: Attribute::NoImplicitFloat)) |
| 423 | return false; |
| 424 | |
| 425 | AliasAnalysis &AA = getAnalysis<AAResultsWrapperPass>().getAAResults(); |
| 426 | DominatorTree &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree(); |
| 427 | ScalarEvolution &SE = getAnalysis<ScalarEvolutionWrapperPass>().getSE(); |
| 428 | TargetTransformInfo &TTI = |
| 429 | getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F); |
| 430 | |
| 431 | AssumptionCache &AC = |
| 432 | getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F); |
| 433 | |
| 434 | return Vectorizer(F, AA, AC, DT, SE, TTI).run(); |
| 435 | } |
| 436 | |
| 437 | PreservedAnalyses LoadStoreVectorizerPass::run(Function &F, |
| 438 | FunctionAnalysisManager &AM) { |
| 439 | // Don't vectorize when the attribute NoImplicitFloat is used. |
| 440 | if (F.hasFnAttribute(Kind: Attribute::NoImplicitFloat)) |
| 441 | return PreservedAnalyses::all(); |
| 442 | |
| 443 | AliasAnalysis &AA = AM.getResult<AAManager>(IR&: F); |
| 444 | DominatorTree &DT = AM.getResult<DominatorTreeAnalysis>(IR&: F); |
| 445 | ScalarEvolution &SE = AM.getResult<ScalarEvolutionAnalysis>(IR&: F); |
| 446 | TargetTransformInfo &TTI = AM.getResult<TargetIRAnalysis>(IR&: F); |
| 447 | AssumptionCache &AC = AM.getResult<AssumptionAnalysis>(IR&: F); |
| 448 | |
| 449 | bool Changed = Vectorizer(F, AA, AC, DT, SE, TTI).run(); |
| 450 | PreservedAnalyses PA; |
| 451 | PA.preserveSet<CFGAnalyses>(); |
| 452 | return Changed ? PA : PreservedAnalyses::all(); |
| 453 | } |
| 454 | |
| 455 | bool Vectorizer::run() { |
| 456 | bool Changed = false; |
| 457 | // Break up the BB if there are any instrs which aren't guaranteed to transfer |
| 458 | // execution to their successor. |
| 459 | // |
| 460 | // Consider, for example: |
| 461 | // |
| 462 | // def assert_arr_len(int n) { if (n < 2) exit(); } |
| 463 | // |
| 464 | // load arr[0] |
| 465 | // call assert_array_len(arr.length) |
| 466 | // load arr[1] |
| 467 | // |
| 468 | // Even though assert_arr_len does not read or write any memory, we can't |
| 469 | // speculate the second load before the call. More info at |
| 470 | // https://github.com/llvm/llvm-project/issues/52950. |
| 471 | for (BasicBlock *BB : post_order(G: &F)) { |
| 472 | // BB must at least have a terminator. |
| 473 | assert(!BB->empty()); |
| 474 | |
| 475 | SmallVector<BasicBlock::iterator, 8> Barriers; |
| 476 | Barriers.emplace_back(Args: BB->begin()); |
| 477 | for (Instruction &I : *BB) |
| 478 | if (!isGuaranteedToTransferExecutionToSuccessor(I: &I)) |
| 479 | Barriers.emplace_back(Args: I.getIterator()); |
| 480 | Barriers.emplace_back(Args: BB->end()); |
| 481 | |
| 482 | for (auto It = Barriers.begin(), End = std::prev(x: Barriers.end()); It != End; |
| 483 | ++It) |
| 484 | Changed |= runOnPseudoBB(Begin: *It, End: *std::next(x: It)); |
| 485 | |
| 486 | for (Instruction *I : ToErase) { |
| 487 | // These will get deleted in deleteExtraElements. |
| 488 | // This is because ExtraElements will include both extra elements |
| 489 | // that *were* vectorized and extra elements that *were not* |
| 490 | // vectorized. ToErase will only include extra elements that *were* |
| 491 | // vectorized, so in order to avoid double deletion we skip them here and |
| 492 | // handle them in deleteExtraElements. |
| 493 | if (ExtraElements.contains(V: I)) |
| 494 | continue; |
| 495 | auto *PtrOperand = getLoadStorePointerOperand(V: I); |
| 496 | if (I->use_empty()) |
| 497 | I->eraseFromParent(); |
| 498 | RecursivelyDeleteTriviallyDeadInstructions(V: PtrOperand); |
| 499 | } |
| 500 | ToErase.clear(); |
| 501 | deleteExtraElements(); |
| 502 | } |
| 503 | |
| 504 | return Changed; |
| 505 | } |
| 506 | |
| 507 | bool Vectorizer::runOnPseudoBB(BasicBlock::iterator Begin, |
| 508 | BasicBlock::iterator End) { |
| 509 | LLVM_DEBUG({ |
| 510 | dbgs() << "LSV: Running on pseudo-BB [" << *Begin << " ... " ; |
| 511 | if (End != Begin->getParent()->end()) |
| 512 | dbgs() << *End; |
| 513 | else |
| 514 | dbgs() << "<BB end>" ; |
| 515 | dbgs() << ")\n" ; |
| 516 | }); |
| 517 | |
| 518 | bool Changed = false; |
| 519 | for (const auto &[EqClassKey, EqClass] : |
| 520 | collectEquivalenceClasses(Begin, End)) |
| 521 | Changed |= runOnEquivalenceClass(EqClassKey, EqClass); |
| 522 | |
| 523 | return Changed; |
| 524 | } |
| 525 | |
| 526 | bool Vectorizer::runOnEquivalenceClass(const EqClassKey &EqClassKey, |
| 527 | ArrayRef<Instruction *> EqClass) { |
| 528 | bool Changed = false; |
| 529 | |
| 530 | LLVM_DEBUG({ |
| 531 | dbgs() << "LSV: Running on equivalence class of size " << EqClass.size() |
| 532 | << " keyed on " << EqClassKey << ":\n" ; |
| 533 | for (Instruction *I : EqClass) |
| 534 | dbgs() << " " << *I << "\n" ; |
| 535 | }); |
| 536 | |
| 537 | std::vector<Chain> Chains = gatherChains(Instrs: EqClass); |
| 538 | LLVM_DEBUG(dbgs() << "LSV: Got " << Chains.size() |
| 539 | << " nontrivial chains.\n" ;); |
| 540 | for (Chain &C : Chains) |
| 541 | Changed |= runOnChain(C); |
| 542 | return Changed; |
| 543 | } |
| 544 | |
| 545 | bool Vectorizer::runOnChain(Chain &C) { |
| 546 | LLVM_DEBUG({ |
| 547 | dbgs() << "LSV: Running on chain with " << C.size() << " instructions:\n" ; |
| 548 | dumpChain(C); |
| 549 | }); |
| 550 | |
| 551 | // Split up the chain into increasingly smaller chains, until we can finally |
| 552 | // vectorize the chains. |
| 553 | // |
| 554 | // (Don't be scared by the depth of the loop nest here. These operations are |
| 555 | // all at worst O(n lg n) in the number of instructions, and splitting chains |
| 556 | // doesn't change the number of instrs. So the whole loop nest is O(n lg n).) |
| 557 | bool Changed = false; |
| 558 | for (auto &C : splitChainByMayAliasInstrs(C)) |
| 559 | for (auto &C : splitChainByContiguity(C)) |
| 560 | for (auto &C : splitChainByAlignment(C)) |
| 561 | Changed |= vectorizeChain(C); |
| 562 | return Changed; |
| 563 | } |
| 564 | |
| 565 | std::vector<Chain> Vectorizer::splitChainByMayAliasInstrs(Chain &C) { |
| 566 | if (C.empty()) |
| 567 | return {}; |
| 568 | |
| 569 | sortChainInBBOrder(C); |
| 570 | |
| 571 | LLVM_DEBUG({ |
| 572 | dbgs() << "LSV: splitChainByMayAliasInstrs considering chain:\n" ; |
| 573 | dumpChain(C); |
| 574 | }); |
| 575 | |
| 576 | // We know that elements in the chain with nonverlapping offsets can't |
| 577 | // alias, but AA may not be smart enough to figure this out. Use a |
| 578 | // hashtable. |
| 579 | DenseMap<Instruction *, APInt /*OffsetFromLeader*/> ChainOffsets; |
| 580 | for (const auto &E : C) |
| 581 | ChainOffsets.insert(KV: {&*E.Inst, E.OffsetFromLeader}); |
| 582 | |
| 583 | // Across a single invocation of this function the IR is not changing, so |
| 584 | // using a batched Alias Analysis is safe and can reduce compile time. |
| 585 | BatchAAResults BatchAA(AA); |
| 586 | |
| 587 | // Loads get hoisted up to the first load in the chain. Stores get sunk |
| 588 | // down to the last store in the chain. Our algorithm for loads is: |
| 589 | // |
| 590 | // - Take the first element of the chain. This is the start of a new chain. |
| 591 | // - Take the next element of `Chain` and check for may-alias instructions |
| 592 | // up to the start of NewChain. If no may-alias instrs, add it to |
| 593 | // NewChain. Otherwise, start a new NewChain. |
| 594 | // |
| 595 | // For stores it's the same except in the reverse direction. |
| 596 | // |
| 597 | // We expect IsLoad to be an std::bool_constant. |
| 598 | auto Impl = [&](auto IsLoad) { |
| 599 | // MSVC is unhappy if IsLoad is a capture, so pass it as an arg. |
| 600 | auto [ChainBegin, ChainEnd] = [&](auto IsLoad) { |
| 601 | if constexpr (IsLoad()) |
| 602 | return std::make_pair(x: C.begin(), y: C.end()); |
| 603 | else |
| 604 | return std::make_pair(x: C.rbegin(), y: C.rend()); |
| 605 | }(IsLoad); |
| 606 | assert(ChainBegin != ChainEnd); |
| 607 | |
| 608 | std::vector<Chain> Chains; |
| 609 | SmallVector<ChainElem, 1> NewChain; |
| 610 | NewChain.emplace_back(*ChainBegin); |
| 611 | for (auto ChainIt = std::next(ChainBegin); ChainIt != ChainEnd; ++ChainIt) { |
| 612 | if (isSafeToMove<IsLoad>(ChainIt->Inst, NewChain.front().Inst, |
| 613 | ChainOffsets, BatchAA)) { |
| 614 | LLVM_DEBUG(dbgs() << "LSV: No intervening may-alias instrs; can merge " |
| 615 | << *ChainIt->Inst << " into " << *ChainBegin->Inst |
| 616 | << "\n" ); |
| 617 | NewChain.emplace_back(*ChainIt); |
| 618 | } else { |
| 619 | LLVM_DEBUG( |
| 620 | dbgs() << "LSV: Found intervening may-alias instrs; cannot merge " |
| 621 | << *ChainIt->Inst << " into " << *ChainBegin->Inst << "\n" ); |
| 622 | if (NewChain.size() > 1) { |
| 623 | LLVM_DEBUG({ |
| 624 | dbgs() << "LSV: got nontrivial chain without aliasing instrs:\n" ; |
| 625 | dumpChain(NewChain); |
| 626 | }); |
| 627 | Chains.emplace_back(args: std::move(NewChain)); |
| 628 | } |
| 629 | |
| 630 | // Start a new chain. |
| 631 | NewChain = SmallVector<ChainElem, 1>({*ChainIt}); |
| 632 | } |
| 633 | } |
| 634 | if (NewChain.size() > 1) { |
| 635 | LLVM_DEBUG({ |
| 636 | dbgs() << "LSV: got nontrivial chain without aliasing instrs:\n" ; |
| 637 | dumpChain(NewChain); |
| 638 | }); |
| 639 | Chains.emplace_back(args: std::move(NewChain)); |
| 640 | } |
| 641 | return Chains; |
| 642 | }; |
| 643 | |
| 644 | if (isa<LoadInst>(Val: C[0].Inst)) |
| 645 | return Impl(/*IsLoad=*/std::bool_constant<true>()); |
| 646 | |
| 647 | assert(isa<StoreInst>(C[0].Inst)); |
| 648 | return Impl(/*IsLoad=*/std::bool_constant<false>()); |
| 649 | } |
| 650 | |
| 651 | std::vector<Chain> Vectorizer::splitChainByContiguity(Chain &C) { |
| 652 | if (C.empty()) |
| 653 | return {}; |
| 654 | |
| 655 | sortChainInOffsetOrder(C); |
| 656 | |
| 657 | LLVM_DEBUG({ |
| 658 | dbgs() << "LSV: splitChainByContiguity considering chain:\n" ; |
| 659 | dumpChain(C); |
| 660 | }); |
| 661 | |
| 662 | // If the chain is not contiguous, we try to fill the gap with "extra" |
| 663 | // elements to artificially make it contiguous, to try to enable |
| 664 | // vectorization. We only fill gaps if there is potential to end up with a |
| 665 | // legal masked load/store given the target, address space, and element type. |
| 666 | // At this point, when querying the TTI, optimistically assume max alignment |
| 667 | // and max vector size, as splitChainByAlignment will ensure the final vector |
| 668 | // shape passes the legalization check. |
| 669 | unsigned AS = getLoadStoreAddressSpace(I: C[0].Inst); |
| 670 | Type *ElementType = getLoadStoreType(I: C[0].Inst)->getScalarType(); |
| 671 | unsigned MaxVecRegBits = TTI.getLoadStoreVecRegBitWidth(AddrSpace: AS); |
| 672 | Align OptimisticAlign = Align(MaxVecRegBits / 8); |
| 673 | unsigned int MaxVectorNumElems = |
| 674 | MaxVecRegBits / DL.getTypeSizeInBits(Ty: ElementType); |
| 675 | // Note: This check decides whether to try to fill gaps based on the masked |
| 676 | // legality of the target's maximum vector size (getLoadStoreVecRegBitWidth). |
| 677 | // If a target *does not* support a masked load/store with this max vector |
| 678 | // size, but *does* support a masked load/store with a *smaller* vector size, |
| 679 | // that optimization will be missed. This does not occur in any of the targets |
| 680 | // that currently support this API. |
| 681 | FixedVectorType *OptimisticVectorType = |
| 682 | FixedVectorType::get(ElementType, NumElts: MaxVectorNumElems); |
| 683 | bool TryFillGaps = |
| 684 | isa<LoadInst>(Val: C[0].Inst) |
| 685 | ? TTI.isLegalMaskedLoad(DataType: OptimisticVectorType, Alignment: OptimisticAlign, AddressSpace: AS, |
| 686 | MaskKind: TTI::MaskKind::ConstantMask) |
| 687 | : TTI.isLegalMaskedStore(DataType: OptimisticVectorType, Alignment: OptimisticAlign, AddressSpace: AS, |
| 688 | MaskKind: TTI::MaskKind::ConstantMask); |
| 689 | |
| 690 | // Cache the best aligned element in the chain for use when creating extra |
| 691 | // elements. |
| 692 | Align BestAlignedElemAlign = getLoadStoreAlignment(I: C[0].Inst); |
| 693 | APInt OffsetOfBestAlignedElemFromLeader = C[0].OffsetFromLeader; |
| 694 | for (const auto &E : C) { |
| 695 | Align ElementAlignment = getLoadStoreAlignment(I: E.Inst); |
| 696 | if (ElementAlignment > BestAlignedElemAlign) { |
| 697 | BestAlignedElemAlign = ElementAlignment; |
| 698 | OffsetOfBestAlignedElemFromLeader = E.OffsetFromLeader; |
| 699 | } |
| 700 | } |
| 701 | |
| 702 | auto DeriveAlignFromBestAlignedElem = [&](APInt NewElemOffsetFromLeader) { |
| 703 | return commonAlignment( |
| 704 | A: BestAlignedElemAlign, |
| 705 | Offset: (NewElemOffsetFromLeader - OffsetOfBestAlignedElemFromLeader) |
| 706 | .abs() |
| 707 | .getLimitedValue()); |
| 708 | }; |
| 709 | |
| 710 | unsigned ASPtrBits = DL.getIndexSizeInBits(AS); |
| 711 | |
| 712 | std::vector<Chain> Ret; |
| 713 | Ret.push_back(x: {C.front()}); |
| 714 | |
| 715 | unsigned ChainElemTyBits = DL.getTypeSizeInBits(Ty: getChainElemTy(C)); |
| 716 | ChainElem &Prev = C[0]; |
| 717 | for (auto It = std::next(x: C.begin()), End = C.end(); It != End; ++It) { |
| 718 | auto &CurChain = Ret.back(); |
| 719 | |
| 720 | APInt PrevSzBytes = |
| 721 | APInt(ASPtrBits, DL.getTypeStoreSize(Ty: getLoadStoreType(I: Prev.Inst))); |
| 722 | APInt PrevReadEnd = Prev.OffsetFromLeader + PrevSzBytes; |
| 723 | unsigned SzBytes = DL.getTypeStoreSize(Ty: getLoadStoreType(I: It->Inst)); |
| 724 | |
| 725 | // Add this instruction to the end of the current chain, or start a new one. |
| 726 | assert( |
| 727 | 8 * SzBytes % ChainElemTyBits == 0 && |
| 728 | "Every chain-element size must be a multiple of the element size after " |
| 729 | "vectorization." ); |
| 730 | APInt ReadEnd = It->OffsetFromLeader + SzBytes; |
| 731 | // Allow redundancy: partial or full overlap counts as contiguous. |
| 732 | bool AreContiguous = false; |
| 733 | if (It->OffsetFromLeader.sle(RHS: PrevReadEnd)) { |
| 734 | // Check overlap is a multiple of the element size after vectorization. |
| 735 | uint64_t Overlap = (PrevReadEnd - It->OffsetFromLeader).getZExtValue(); |
| 736 | if (8 * Overlap % ChainElemTyBits == 0) |
| 737 | AreContiguous = true; |
| 738 | } |
| 739 | |
| 740 | LLVM_DEBUG(dbgs() << "LSV: Instruction is " |
| 741 | << (AreContiguous ? "contiguous" : "chain-breaker" ) |
| 742 | << *It->Inst << " (starts at offset " |
| 743 | << It->OffsetFromLeader << ")\n" ); |
| 744 | |
| 745 | // If the chain is not contiguous, try to fill in gaps between Prev and |
| 746 | // Curr. For now, we aren't filling gaps between load/stores of different |
| 747 | // sizes. Additionally, as a conservative heuristic, we only fill gaps of |
| 748 | // 1-2 elements. Generating loads/stores with too many unused bytes has a |
| 749 | // side effect of increasing register pressure (on NVIDIA targets at least), |
| 750 | // which could cancel out the benefits of reducing number of load/stores. |
| 751 | bool GapFilled = false; |
| 752 | if (!AreContiguous && TryFillGaps && PrevSzBytes == SzBytes) { |
| 753 | APInt GapSzBytes = It->OffsetFromLeader - PrevReadEnd; |
| 754 | if (GapSzBytes == PrevSzBytes) { |
| 755 | // There is a single gap between Prev and Curr, create one extra element |
| 756 | ChainElem NewElem = createExtraElementAfter( |
| 757 | PrevElem: Prev, Ty: getLoadStoreType(I: Prev.Inst), Offset: PrevSzBytes, Prefix: "GapFill" , |
| 758 | Alignment: DeriveAlignFromBestAlignedElem(PrevReadEnd)); |
| 759 | CurChain.push_back(Elt: NewElem); |
| 760 | GapFilled = true; |
| 761 | } |
| 762 | // There are two gaps between Prev and Curr, only create two extra |
| 763 | // elements if Prev is the first element in a sequence of four. |
| 764 | // This has the highest chance of resulting in a beneficial vectorization. |
| 765 | if ((GapSzBytes == 2 * PrevSzBytes) && (CurChain.size() % 4 == 1)) { |
| 766 | ChainElem NewElem1 = createExtraElementAfter( |
| 767 | PrevElem: Prev, Ty: getLoadStoreType(I: Prev.Inst), Offset: PrevSzBytes, Prefix: "GapFill" , |
| 768 | Alignment: DeriveAlignFromBestAlignedElem(PrevReadEnd)); |
| 769 | ChainElem NewElem2 = createExtraElementAfter( |
| 770 | PrevElem: NewElem1, Ty: getLoadStoreType(I: Prev.Inst), Offset: PrevSzBytes, Prefix: "GapFill" , |
| 771 | Alignment: DeriveAlignFromBestAlignedElem(PrevReadEnd + PrevSzBytes)); |
| 772 | CurChain.push_back(Elt: NewElem1); |
| 773 | CurChain.push_back(Elt: NewElem2); |
| 774 | GapFilled = true; |
| 775 | } |
| 776 | } |
| 777 | |
| 778 | if (AreContiguous || GapFilled) |
| 779 | CurChain.push_back(Elt: *It); |
| 780 | else |
| 781 | Ret.push_back(x: {*It}); |
| 782 | // In certain cases when handling redundant elements with partial overlaps, |
| 783 | // the previous element may still extend beyond the current element. Only |
| 784 | // update Prev if the current element is the new end of the chain. |
| 785 | if (ReadEnd.sge(RHS: PrevReadEnd)) |
| 786 | Prev = *It; |
| 787 | } |
| 788 | |
| 789 | // Filter out length-1 chains, these are uninteresting. |
| 790 | llvm::erase_if(C&: Ret, P: [](const auto &Chain) { return Chain.size() <= 1; }); |
| 791 | return Ret; |
| 792 | } |
| 793 | |
| 794 | Type *Vectorizer::getChainElemTy(const Chain &C) { |
| 795 | assert(!C.empty()); |
| 796 | // The rules are: |
| 797 | // - If there are any pointer types in the chain, use an integer type. |
| 798 | // - Prefer an integer type if it appears in the chain. |
| 799 | // - Otherwise, use the first type in the chain. |
| 800 | // |
| 801 | // The rule about pointer types is a simplification when we merge e.g. a load |
| 802 | // of a ptr and a double. There's no direct conversion from a ptr to a |
| 803 | // double; it requires a ptrtoint followed by a bitcast. |
| 804 | // |
| 805 | // It's unclear to me if the other rules have any practical effect, but we do |
| 806 | // it to match this pass's previous behavior. |
| 807 | if (any_of(Range: C, P: [](const ChainElem &E) { |
| 808 | return getLoadStoreType(I: E.Inst)->getScalarType()->isPointerTy(); |
| 809 | })) { |
| 810 | return Type::getIntNTy( |
| 811 | C&: F.getContext(), |
| 812 | N: DL.getTypeSizeInBits(Ty: getLoadStoreType(I: C[0].Inst)->getScalarType())); |
| 813 | } |
| 814 | |
| 815 | for (const ChainElem &E : C) |
| 816 | if (Type *T = getLoadStoreType(I: E.Inst)->getScalarType(); T->isIntegerTy()) |
| 817 | return T; |
| 818 | return getLoadStoreType(I: C[0].Inst)->getScalarType(); |
| 819 | } |
| 820 | |
| 821 | std::vector<Chain> Vectorizer::splitChainByAlignment(Chain &C) { |
| 822 | // We use a simple greedy algorithm. |
| 823 | // - Given a chain of length N, find all prefixes that |
| 824 | // (a) are not longer than the max register length, and |
| 825 | // (b) are a power of 2. |
| 826 | // - Starting from the longest prefix, try to create a vector of that length. |
| 827 | // - If one of them works, great. Repeat the algorithm on any remaining |
| 828 | // elements in the chain. |
| 829 | // - If none of them work, discard the first element and repeat on a chain |
| 830 | // of length N-1. |
| 831 | if (C.empty()) |
| 832 | return {}; |
| 833 | |
| 834 | sortChainInOffsetOrder(C); |
| 835 | |
| 836 | LLVM_DEBUG({ |
| 837 | dbgs() << "LSV: splitChainByAlignment considering chain:\n" ; |
| 838 | dumpChain(C); |
| 839 | }); |
| 840 | |
| 841 | bool IsLoadChain = isa<LoadInst>(Val: C[0].Inst); |
| 842 | auto GetVectorFactor = [&](unsigned VF, unsigned LoadStoreSize, |
| 843 | unsigned ChainSizeBytes, VectorType *VecTy) { |
| 844 | return IsLoadChain ? TTI.getLoadVectorFactor(VF, LoadSize: LoadStoreSize, |
| 845 | ChainSizeInBytes: ChainSizeBytes, VecTy) |
| 846 | : TTI.getStoreVectorFactor(VF, StoreSize: LoadStoreSize, |
| 847 | ChainSizeInBytes: ChainSizeBytes, VecTy); |
| 848 | }; |
| 849 | |
| 850 | #ifndef NDEBUG |
| 851 | for (const auto &E : C) { |
| 852 | Type *Ty = getLoadStoreType(E.Inst)->getScalarType(); |
| 853 | assert(isPowerOf2_32(DL.getTypeSizeInBits(Ty)) && |
| 854 | "Should have filtered out non-power-of-two elements in " |
| 855 | "collectEquivalenceClasses." ); |
| 856 | } |
| 857 | #endif |
| 858 | |
| 859 | unsigned AS = getLoadStoreAddressSpace(I: C[0].Inst); |
| 860 | unsigned VecRegBytes = TTI.getLoadStoreVecRegBitWidth(AddrSpace: AS) / 8; |
| 861 | |
| 862 | // For compile time reasons, we cache whether or not the superset |
| 863 | // of all candidate chains contains any extra loads/stores from earlier gap |
| 864 | // filling. |
| 865 | bool CandidateChainsMayContainExtraLoadsStores = any_of( |
| 866 | Range&: C, P: [this](const ChainElem &E) { return ExtraElements.contains(V: E.Inst); }); |
| 867 | |
| 868 | std::vector<Chain> Ret; |
| 869 | for (unsigned CBegin = 0; CBegin < C.size(); ++CBegin) { |
| 870 | // Find candidate chains of size not greater than the largest vector reg. |
| 871 | // These chains are over the closed interval [CBegin, CEnd]. |
| 872 | SmallVector<std::pair<unsigned /*CEnd*/, unsigned /*SizeBytes*/>, 8> |
| 873 | CandidateChains; |
| 874 | // Need to compute the size of every candidate chain from its beginning |
| 875 | // because of possible overlapping among chain elements. |
| 876 | unsigned Sz = DL.getTypeStoreSize(Ty: getLoadStoreType(I: C[CBegin].Inst)); |
| 877 | APInt PrevReadEnd = C[CBegin].OffsetFromLeader + Sz; |
| 878 | for (unsigned CEnd = CBegin + 1, Size = C.size(); CEnd < Size; ++CEnd) { |
| 879 | APInt ReadEnd = C[CEnd].OffsetFromLeader + |
| 880 | DL.getTypeStoreSize(Ty: getLoadStoreType(I: C[CEnd].Inst)); |
| 881 | unsigned BytesAdded = |
| 882 | PrevReadEnd.sle(RHS: ReadEnd) ? (ReadEnd - PrevReadEnd).getSExtValue() : 0; |
| 883 | Sz += BytesAdded; |
| 884 | if (Sz > VecRegBytes) |
| 885 | break; |
| 886 | CandidateChains.emplace_back(Args&: CEnd, Args&: Sz); |
| 887 | PrevReadEnd = APIntOps::smax(A: PrevReadEnd, B: ReadEnd); |
| 888 | } |
| 889 | |
| 890 | // Consider the longest chain first. |
| 891 | for (auto It = CandidateChains.rbegin(), End = CandidateChains.rend(); |
| 892 | It != End; ++It) { |
| 893 | auto [CEnd, SizeBytes] = *It; |
| 894 | LLVM_DEBUG( |
| 895 | dbgs() << "LSV: splitChainByAlignment considering candidate chain [" |
| 896 | << *C[CBegin].Inst << " ... " << *C[CEnd].Inst << "]\n" ); |
| 897 | |
| 898 | Type *VecElemTy = getChainElemTy(C); |
| 899 | // Note, VecElemTy is a power of 2, but might be less than one byte. For |
| 900 | // example, we can vectorize 2 x <2 x i4> to <4 x i4>, and in this case |
| 901 | // VecElemTy would be i4. |
| 902 | unsigned VecElemBits = DL.getTypeSizeInBits(Ty: VecElemTy); |
| 903 | |
| 904 | // SizeBytes and VecElemBits are powers of 2, so they divide evenly. |
| 905 | assert((8 * SizeBytes) % VecElemBits == 0); |
| 906 | unsigned NumVecElems = 8 * SizeBytes / VecElemBits; |
| 907 | FixedVectorType *VecTy = FixedVectorType::get(ElementType: VecElemTy, NumElts: NumVecElems); |
| 908 | unsigned VF = 8 * VecRegBytes / VecElemBits; |
| 909 | |
| 910 | // Check that TTI is happy with this vectorization factor. |
| 911 | unsigned TargetVF = GetVectorFactor(VF, VecElemBits, |
| 912 | VecElemBits * NumVecElems / 8, VecTy); |
| 913 | if (TargetVF != VF && TargetVF < NumVecElems) { |
| 914 | LLVM_DEBUG( |
| 915 | dbgs() << "LSV: splitChainByAlignment discarding candidate chain " |
| 916 | "because TargetVF=" |
| 917 | << TargetVF << " != VF=" << VF |
| 918 | << " and TargetVF < NumVecElems=" << NumVecElems << "\n" ); |
| 919 | continue; |
| 920 | } |
| 921 | |
| 922 | // If we're loading/storing from an alloca, align it if possible. |
| 923 | // |
| 924 | // FIXME: We eagerly upgrade the alignment, regardless of whether TTI |
| 925 | // tells us this is beneficial. This feels a bit odd, but it matches |
| 926 | // existing tests. This isn't *so* bad, because at most we align to 4 |
| 927 | // bytes (current value of StackAdjustedAlignment). |
| 928 | // |
| 929 | // FIXME: We will upgrade the alignment of the alloca even if it turns out |
| 930 | // we can't vectorize for some other reason. |
| 931 | Value *PtrOperand = getLoadStorePointerOperand(V: C[CBegin].Inst); |
| 932 | bool IsAllocaAccess = AS == DL.getAllocaAddrSpace() && |
| 933 | isa<AllocaInst>(Val: PtrOperand->stripPointerCasts()); |
| 934 | Align Alignment = getLoadStoreAlignment(I: C[CBegin].Inst); |
| 935 | Align PrefAlign = Align(StackAdjustedAlignment); |
| 936 | if (IsAllocaAccess && Alignment.value() % SizeBytes != 0 && |
| 937 | accessIsAllowedAndFast(SizeBytes, AS, Alignment: PrefAlign, VecElemBits)) { |
| 938 | Align NewAlign = getOrEnforceKnownAlignment( |
| 939 | V: PtrOperand, PrefAlign, DL, CxtI: C[CBegin].Inst, AC: nullptr, DT: &DT); |
| 940 | if (NewAlign >= Alignment) { |
| 941 | LLVM_DEBUG(dbgs() |
| 942 | << "LSV: splitByChain upgrading alloca alignment from " |
| 943 | << Alignment.value() << " to " << NewAlign.value() |
| 944 | << "\n" ); |
| 945 | Alignment = NewAlign; |
| 946 | } |
| 947 | } |
| 948 | |
| 949 | Chain ExtendingLoadsStores; |
| 950 | if (!accessIsAllowedAndFast(SizeBytes, AS, Alignment, VecElemBits)) { |
| 951 | // If we have a non-power-of-2 element count, attempt to extend the |
| 952 | // chain to the next power-of-2 if it makes the access allowed and |
| 953 | // fast. |
| 954 | bool AllowedAndFast = false; |
| 955 | if (NumVecElems < TargetVF && !isPowerOf2_32(Value: NumVecElems) && |
| 956 | VecElemBits >= 8) { |
| 957 | // TargetVF may be a lot higher than NumVecElems, |
| 958 | // so only extend to the next power of 2. |
| 959 | assert(VecElemBits % 8 == 0); |
| 960 | unsigned VecElemBytes = VecElemBits / 8; |
| 961 | unsigned NewNumVecElems = PowerOf2Ceil(A: NumVecElems); |
| 962 | unsigned NewSizeBytes = VecElemBytes * NewNumVecElems; |
| 963 | |
| 964 | assert(isPowerOf2_32(TargetVF) && |
| 965 | "TargetVF expected to be a power of 2" ); |
| 966 | assert(NewNumVecElems <= TargetVF && |
| 967 | "Should not extend past TargetVF" ); |
| 968 | |
| 969 | LLVM_DEBUG(dbgs() |
| 970 | << "LSV: attempting to extend chain of " << NumVecElems |
| 971 | << " " << (IsLoadChain ? "loads" : "stores" ) << " to " |
| 972 | << NewNumVecElems << " elements\n" ); |
| 973 | bool IsLegalToExtend = |
| 974 | IsLoadChain ? TTI.isLegalMaskedLoad( |
| 975 | DataType: FixedVectorType::get(ElementType: VecElemTy, NumElts: NewNumVecElems), |
| 976 | Alignment, AddressSpace: AS, MaskKind: TTI::MaskKind::ConstantMask) |
| 977 | : TTI.isLegalMaskedStore( |
| 978 | DataType: FixedVectorType::get(ElementType: VecElemTy, NumElts: NewNumVecElems), |
| 979 | Alignment, AddressSpace: AS, MaskKind: TTI::MaskKind::ConstantMask); |
| 980 | // Only artificially increase the chain if it would be AllowedAndFast |
| 981 | // and if the resulting masked load/store will be legal for the |
| 982 | // target. |
| 983 | if (IsLegalToExtend && |
| 984 | accessIsAllowedAndFast(SizeBytes: NewSizeBytes, AS, Alignment, |
| 985 | VecElemBits)) { |
| 986 | LLVM_DEBUG(dbgs() |
| 987 | << "LSV: extending " << (IsLoadChain ? "load" : "store" ) |
| 988 | << " chain of " << NumVecElems << " " |
| 989 | << (IsLoadChain ? "loads" : "stores" ) |
| 990 | << " with total byte size of " << SizeBytes << " to " |
| 991 | << NewNumVecElems << " " |
| 992 | << (IsLoadChain ? "loads" : "stores" ) |
| 993 | << " with total byte size of " << NewSizeBytes |
| 994 | << ", TargetVF=" << TargetVF << " \n" ); |
| 995 | |
| 996 | // Create (NewNumVecElems - NumVecElems) extra elements. |
| 997 | // We are basing each extra element on CBegin, which means the |
| 998 | // offsets should be based on SizeBytes, which represents the offset |
| 999 | // from CBegin to the current end of the chain. |
| 1000 | unsigned ASPtrBits = DL.getIndexSizeInBits(AS); |
| 1001 | for (unsigned I = 0; I < (NewNumVecElems - NumVecElems); I++) { |
| 1002 | ChainElem NewElem = createExtraElementAfter( |
| 1003 | PrevElem: C[CBegin], Ty: VecElemTy, |
| 1004 | Offset: APInt(ASPtrBits, SizeBytes + I * VecElemBytes), Prefix: "Extend" ); |
| 1005 | ExtendingLoadsStores.push_back(Elt: NewElem); |
| 1006 | } |
| 1007 | |
| 1008 | // Update the size and number of elements for upcoming checks. |
| 1009 | SizeBytes = NewSizeBytes; |
| 1010 | NumVecElems = NewNumVecElems; |
| 1011 | AllowedAndFast = true; |
| 1012 | } |
| 1013 | } |
| 1014 | if (!AllowedAndFast) { |
| 1015 | // We were not able to achieve legality by extending the chain. |
| 1016 | LLVM_DEBUG(dbgs() |
| 1017 | << "LSV: splitChainByAlignment discarding candidate chain " |
| 1018 | "because its alignment is not AllowedAndFast: " |
| 1019 | << Alignment.value() << "\n" ); |
| 1020 | continue; |
| 1021 | } |
| 1022 | } |
| 1023 | |
| 1024 | if ((IsLoadChain && |
| 1025 | !TTI.isLegalToVectorizeLoadChain(ChainSizeInBytes: SizeBytes, Alignment, AddrSpace: AS)) || |
| 1026 | (!IsLoadChain && |
| 1027 | !TTI.isLegalToVectorizeStoreChain(ChainSizeInBytes: SizeBytes, Alignment, AddrSpace: AS))) { |
| 1028 | LLVM_DEBUG( |
| 1029 | dbgs() << "LSV: splitChainByAlignment discarding candidate chain " |
| 1030 | "because !isLegalToVectorizeLoad/StoreChain." ); |
| 1031 | continue; |
| 1032 | } |
| 1033 | |
| 1034 | if (CandidateChainsMayContainExtraLoadsStores) { |
| 1035 | // If the candidate chain contains extra loads/stores from an earlier |
| 1036 | // optimization, confirm legality now. This filter is essential because |
| 1037 | // when filling gaps in splitChainByContiguity, we queried the API to |
| 1038 | // check that (for a given element type and address space) there *may* |
| 1039 | // have been a legal masked load/store we could possibly create. Now, we |
| 1040 | // need to check if the actual chain we ended up with is legal to turn |
| 1041 | // into a masked load/store. This is relevant for NVPTX, for example, |
| 1042 | // where a masked store is only legal if we have ended up with a 256-bit |
| 1043 | // vector. |
| 1044 | bool CurrCandContainsExtraLoadsStores = llvm::any_of( |
| 1045 | Range: ArrayRef<ChainElem>(C).slice(N: CBegin, M: CEnd - CBegin + 1), |
| 1046 | P: [this](const ChainElem &E) { |
| 1047 | return ExtraElements.contains(V: E.Inst); |
| 1048 | }); |
| 1049 | |
| 1050 | if (CurrCandContainsExtraLoadsStores && |
| 1051 | (IsLoadChain ? !TTI.isLegalMaskedLoad( |
| 1052 | DataType: FixedVectorType::get(ElementType: VecElemTy, NumElts: NumVecElems), |
| 1053 | Alignment, AddressSpace: AS, MaskKind: TTI::MaskKind::ConstantMask) |
| 1054 | : !TTI.isLegalMaskedStore( |
| 1055 | DataType: FixedVectorType::get(ElementType: VecElemTy, NumElts: NumVecElems), |
| 1056 | Alignment, AddressSpace: AS, MaskKind: TTI::MaskKind::ConstantMask))) { |
| 1057 | LLVM_DEBUG(dbgs() |
| 1058 | << "LSV: splitChainByAlignment discarding candidate chain " |
| 1059 | "because it contains extra loads/stores that we cannot " |
| 1060 | "legally vectorize into a masked load/store \n" ); |
| 1061 | continue; |
| 1062 | } |
| 1063 | } |
| 1064 | |
| 1065 | // Hooray, we can vectorize this chain! |
| 1066 | Chain &NewChain = Ret.emplace_back(); |
| 1067 | for (unsigned I = CBegin; I <= CEnd; ++I) |
| 1068 | NewChain.emplace_back(Args&: C[I]); |
| 1069 | for (ChainElem E : ExtendingLoadsStores) |
| 1070 | NewChain.emplace_back(Args&: E); |
| 1071 | CBegin = CEnd; // Skip over the instructions we've added to the chain. |
| 1072 | break; |
| 1073 | } |
| 1074 | } |
| 1075 | return Ret; |
| 1076 | } |
| 1077 | |
| 1078 | bool Vectorizer::vectorizeChain(Chain &C) { |
| 1079 | if (C.size() < 2) |
| 1080 | return false; |
| 1081 | |
| 1082 | bool = llvm::any_of( |
| 1083 | Range&: C, P: [this](const ChainElem &E) { return ExtraElements.contains(V: E.Inst); }); |
| 1084 | |
| 1085 | // If we are left with a two-element chain, and one of the elements is an |
| 1086 | // extra element, we don't want to vectorize |
| 1087 | if (C.size() == 2 && ChainContainsExtraLoadsStores) |
| 1088 | return false; |
| 1089 | |
| 1090 | sortChainInOffsetOrder(C); |
| 1091 | |
| 1092 | LLVM_DEBUG({ |
| 1093 | dbgs() << "LSV: Vectorizing chain of " << C.size() << " instructions:\n" ; |
| 1094 | dumpChain(C); |
| 1095 | }); |
| 1096 | |
| 1097 | Type *VecElemTy = getChainElemTy(C); |
| 1098 | bool IsLoadChain = isa<LoadInst>(Val: C[0].Inst); |
| 1099 | unsigned AS = getLoadStoreAddressSpace(I: C[0].Inst); |
| 1100 | unsigned BytesAdded = DL.getTypeStoreSize(Ty: getLoadStoreType(I: &*C[0].Inst)); |
| 1101 | APInt PrevReadEnd = C[0].OffsetFromLeader + BytesAdded; |
| 1102 | unsigned ChainBytes = BytesAdded; |
| 1103 | for (auto It = std::next(x: C.begin()), End = C.end(); It != End; ++It) { |
| 1104 | unsigned SzBytes = DL.getTypeStoreSize(Ty: getLoadStoreType(I: &*It->Inst)); |
| 1105 | APInt ReadEnd = It->OffsetFromLeader + SzBytes; |
| 1106 | // Update ChainBytes considering possible overlap. |
| 1107 | BytesAdded = |
| 1108 | PrevReadEnd.sle(RHS: ReadEnd) ? (ReadEnd - PrevReadEnd).getSExtValue() : 0; |
| 1109 | ChainBytes += BytesAdded; |
| 1110 | PrevReadEnd = APIntOps::smax(A: PrevReadEnd, B: ReadEnd); |
| 1111 | } |
| 1112 | |
| 1113 | assert(8 * ChainBytes % DL.getTypeSizeInBits(VecElemTy) == 0); |
| 1114 | // VecTy is a power of 2 and 1 byte at smallest, but VecElemTy may be smaller |
| 1115 | // than 1 byte (e.g. VecTy == <32 x i1>). |
| 1116 | unsigned NumElem = 8 * ChainBytes / DL.getTypeSizeInBits(Ty: VecElemTy); |
| 1117 | Type *VecTy = FixedVectorType::get(ElementType: VecElemTy, NumElts: NumElem); |
| 1118 | |
| 1119 | Align Alignment = getLoadStoreAlignment(I: C[0].Inst); |
| 1120 | // If this is a load/store of an alloca, we might have upgraded the alloca's |
| 1121 | // alignment earlier. Get the new alignment. |
| 1122 | if (AS == DL.getAllocaAddrSpace()) { |
| 1123 | Alignment = std::max( |
| 1124 | a: Alignment, |
| 1125 | b: getOrEnforceKnownAlignment(V: getLoadStorePointerOperand(V: C[0].Inst), |
| 1126 | PrefAlign: MaybeAlign(), DL, CxtI: C[0].Inst, AC: nullptr, DT: &DT)); |
| 1127 | } |
| 1128 | |
| 1129 | // All elements of the chain must have the same scalar-type size. |
| 1130 | #ifndef NDEBUG |
| 1131 | for (const ChainElem &E : C) |
| 1132 | assert(DL.getTypeStoreSize(getLoadStoreType(E.Inst)->getScalarType()) == |
| 1133 | DL.getTypeStoreSize(VecElemTy)); |
| 1134 | #endif |
| 1135 | |
| 1136 | Instruction *VecInst; |
| 1137 | if (IsLoadChain) { |
| 1138 | // Loads get hoisted to the location of the first load in the chain. We may |
| 1139 | // also need to hoist the (transitive) operands of the loads. |
| 1140 | Builder.SetInsertPoint( |
| 1141 | llvm::min_element(Range&: C, C: [](const auto &A, const auto &B) { |
| 1142 | return A.Inst->comesBefore(B.Inst); |
| 1143 | })->Inst); |
| 1144 | |
| 1145 | // If the chain contains extra loads, we need to vectorize into a |
| 1146 | // masked load. |
| 1147 | if (ChainContainsExtraLoadsStores) { |
| 1148 | assert(TTI.isLegalMaskedLoad(VecTy, Alignment, AS, |
| 1149 | TTI::MaskKind::ConstantMask)); |
| 1150 | Value *Mask = createMaskForExtraElements(C, VecTy: cast<FixedVectorType>(Val: VecTy)); |
| 1151 | VecInst = Builder.CreateMaskedLoad( |
| 1152 | Ty: VecTy, Ptr: getLoadStorePointerOperand(V: C[0].Inst), Alignment, Mask); |
| 1153 | } else { |
| 1154 | // This can happen due to a chain of redundant loads. |
| 1155 | // In this case, just use the element-type, and avoid ExtractElement. |
| 1156 | if (NumElem == 1) |
| 1157 | VecTy = VecElemTy; |
| 1158 | // Chain is in offset order, so C[0] is the instr with the lowest offset, |
| 1159 | // i.e. the root of the vector. |
| 1160 | VecInst = Builder.CreateAlignedLoad( |
| 1161 | Ty: VecTy, Ptr: getLoadStorePointerOperand(V: C[0].Inst), Align: Alignment); |
| 1162 | } |
| 1163 | |
| 1164 | for (const ChainElem &E : C) { |
| 1165 | Instruction *I = E.Inst; |
| 1166 | Value *V; |
| 1167 | Type *T = getLoadStoreType(I); |
| 1168 | unsigned EOffset = |
| 1169 | (E.OffsetFromLeader - C[0].OffsetFromLeader).getZExtValue(); |
| 1170 | unsigned VecIdx = 8 * EOffset / DL.getTypeSizeInBits(Ty: VecElemTy); |
| 1171 | if (!VecTy->isVectorTy()) { |
| 1172 | V = VecInst; |
| 1173 | } else if (auto *VT = dyn_cast<FixedVectorType>(Val: T)) { |
| 1174 | auto Mask = llvm::to_vector<8>( |
| 1175 | Range: llvm::seq<int>(Begin: VecIdx, End: VecIdx + VT->getNumElements())); |
| 1176 | V = Builder.CreateShuffleVector(V: VecInst, Mask, Name: I->getName()); |
| 1177 | } else { |
| 1178 | V = Builder.CreateExtractElement(Vec: VecInst, Idx: Builder.getInt32(C: VecIdx), |
| 1179 | Name: I->getName()); |
| 1180 | } |
| 1181 | if (V->getType() != I->getType()) |
| 1182 | V = Builder.CreateBitOrPointerCast(V, DestTy: I->getType()); |
| 1183 | I->replaceAllUsesWith(V); |
| 1184 | } |
| 1185 | |
| 1186 | // Finally, we need to reorder the instrs in the BB so that the (transitive) |
| 1187 | // operands of VecInst appear before it. To see why, suppose we have |
| 1188 | // vectorized the following code: |
| 1189 | // |
| 1190 | // ptr1 = gep a, 1 |
| 1191 | // load1 = load i32 ptr1 |
| 1192 | // ptr0 = gep a, 0 |
| 1193 | // load0 = load i32 ptr0 |
| 1194 | // |
| 1195 | // We will put the vectorized load at the location of the earliest load in |
| 1196 | // the BB, i.e. load1. We get: |
| 1197 | // |
| 1198 | // ptr1 = gep a, 1 |
| 1199 | // loadv = load <2 x i32> ptr0 |
| 1200 | // load0 = extractelement loadv, 0 |
| 1201 | // load1 = extractelement loadv, 1 |
| 1202 | // ptr0 = gep a, 0 |
| 1203 | // |
| 1204 | // Notice that loadv uses ptr0, which is defined *after* it! |
| 1205 | reorder(I: VecInst); |
| 1206 | } else { |
| 1207 | // Stores get sunk to the location of the last store in the chain. |
| 1208 | Builder.SetInsertPoint(llvm::max_element(Range&: C, C: [](auto &A, auto &B) { |
| 1209 | return A.Inst->comesBefore(B.Inst); |
| 1210 | })->Inst); |
| 1211 | |
| 1212 | // Build the vector to store. |
| 1213 | Value *Vec = PoisonValue::get(T: VecTy); |
| 1214 | auto InsertElem = [&](Value *V, unsigned VecIdx) { |
| 1215 | if (V->getType() != VecElemTy) |
| 1216 | V = Builder.CreateBitOrPointerCast(V, DestTy: VecElemTy); |
| 1217 | Vec = Builder.CreateInsertElement(Vec, NewElt: V, Idx: Builder.getInt32(C: VecIdx)); |
| 1218 | }; |
| 1219 | for (const ChainElem &E : C) { |
| 1220 | auto *I = cast<StoreInst>(Val: E.Inst); |
| 1221 | unsigned EOffset = |
| 1222 | (E.OffsetFromLeader - C[0].OffsetFromLeader).getZExtValue(); |
| 1223 | unsigned VecIdx = 8 * EOffset / DL.getTypeSizeInBits(Ty: VecElemTy); |
| 1224 | if (FixedVectorType *VT = |
| 1225 | dyn_cast<FixedVectorType>(Val: getLoadStoreType(I))) { |
| 1226 | for (int J = 0, JE = VT->getNumElements(); J < JE; ++J) { |
| 1227 | InsertElem(Builder.CreateExtractElement(Vec: I->getValueOperand(), |
| 1228 | Idx: Builder.getInt32(C: J)), |
| 1229 | VecIdx++); |
| 1230 | } |
| 1231 | } else { |
| 1232 | InsertElem(I->getValueOperand(), VecIdx); |
| 1233 | } |
| 1234 | } |
| 1235 | |
| 1236 | // If the chain originates from extra stores, we need to vectorize into a |
| 1237 | // masked store. |
| 1238 | if (ChainContainsExtraLoadsStores) { |
| 1239 | assert(TTI.isLegalMaskedStore(Vec->getType(), Alignment, AS, |
| 1240 | TTI::MaskKind::ConstantMask)); |
| 1241 | Value *Mask = |
| 1242 | createMaskForExtraElements(C, VecTy: cast<FixedVectorType>(Val: Vec->getType())); |
| 1243 | VecInst = Builder.CreateMaskedStore( |
| 1244 | Val: Vec, Ptr: getLoadStorePointerOperand(V: C[0].Inst), Alignment, Mask); |
| 1245 | } else { |
| 1246 | // Chain is in offset order, so C[0] is the instr with the lowest offset, |
| 1247 | // i.e. the root of the vector. |
| 1248 | VecInst = Builder.CreateAlignedStore( |
| 1249 | Val: Vec, Ptr: getLoadStorePointerOperand(V: C[0].Inst), Align: Alignment); |
| 1250 | } |
| 1251 | } |
| 1252 | |
| 1253 | propagateMetadata(I: VecInst, C); |
| 1254 | |
| 1255 | for (const ChainElem &E : C) |
| 1256 | ToErase.emplace_back(Args: E.Inst); |
| 1257 | |
| 1258 | ++NumVectorInstructions; |
| 1259 | NumScalarsVectorized += C.size(); |
| 1260 | return true; |
| 1261 | } |
| 1262 | |
| 1263 | template <bool IsLoadChain> |
| 1264 | bool Vectorizer::isSafeToMove( |
| 1265 | Instruction *ChainElem, Instruction *ChainBegin, |
| 1266 | const DenseMap<Instruction *, APInt /*OffsetFromLeader*/> &ChainOffsets, |
| 1267 | BatchAAResults &BatchAA) { |
| 1268 | LLVM_DEBUG(dbgs() << "LSV: isSafeToMove(" << *ChainElem << " -> " |
| 1269 | << *ChainBegin << ")\n" ); |
| 1270 | |
| 1271 | assert(isa<LoadInst>(ChainElem) == IsLoadChain); |
| 1272 | if (ChainElem == ChainBegin) |
| 1273 | return true; |
| 1274 | |
| 1275 | // Invariant loads can always be reordered; by definition they are not |
| 1276 | // clobbered by stores. |
| 1277 | if (isInvariantLoad(I: ChainElem)) |
| 1278 | return true; |
| 1279 | |
| 1280 | auto BBIt = std::next([&] { |
| 1281 | if constexpr (IsLoadChain) |
| 1282 | return BasicBlock::reverse_iterator(ChainElem); |
| 1283 | else |
| 1284 | return BasicBlock::iterator(ChainElem); |
| 1285 | }()); |
| 1286 | auto BBItEnd = std::next([&] { |
| 1287 | if constexpr (IsLoadChain) |
| 1288 | return BasicBlock::reverse_iterator(ChainBegin); |
| 1289 | else |
| 1290 | return BasicBlock::iterator(ChainBegin); |
| 1291 | }()); |
| 1292 | |
| 1293 | const APInt &ChainElemOffset = ChainOffsets.at(Val: ChainElem); |
| 1294 | const unsigned ChainElemSize = |
| 1295 | DL.getTypeStoreSize(Ty: getLoadStoreType(I: ChainElem)); |
| 1296 | |
| 1297 | for (; BBIt != BBItEnd; ++BBIt) { |
| 1298 | Instruction *I = &*BBIt; |
| 1299 | |
| 1300 | if (!I->mayReadOrWriteMemory()) |
| 1301 | continue; |
| 1302 | |
| 1303 | // Loads can be reordered with other loads. |
| 1304 | if (IsLoadChain && isa<LoadInst>(Val: I)) |
| 1305 | continue; |
| 1306 | |
| 1307 | // Stores can be sunk below invariant loads. |
| 1308 | if (!IsLoadChain && isInvariantLoad(I)) |
| 1309 | continue; |
| 1310 | |
| 1311 | // If I is in the chain, we can tell whether it aliases ChainIt by checking |
| 1312 | // what offset ChainIt accesses. This may be better than AA is able to do. |
| 1313 | // |
| 1314 | // We should really only have duplicate offsets for stores (the duplicate |
| 1315 | // loads should be CSE'ed), but in case we have a duplicate load, we'll |
| 1316 | // split the chain so we don't have to handle this case specially. |
| 1317 | if (auto OffsetIt = ChainOffsets.find(Val: I); OffsetIt != ChainOffsets.end()) { |
| 1318 | // I and ChainElem overlap if: |
| 1319 | // - I and ChainElem have the same offset, OR |
| 1320 | // - I's offset is less than ChainElem's, but I touches past the |
| 1321 | // beginning of ChainElem, OR |
| 1322 | // - ChainElem's offset is less than I's, but ChainElem touches past the |
| 1323 | // beginning of I. |
| 1324 | const APInt &IOffset = OffsetIt->second; |
| 1325 | unsigned IElemSize = DL.getTypeStoreSize(Ty: getLoadStoreType(I)); |
| 1326 | if (IOffset == ChainElemOffset || |
| 1327 | (IOffset.sle(RHS: ChainElemOffset) && |
| 1328 | (IOffset + IElemSize).sgt(RHS: ChainElemOffset)) || |
| 1329 | (ChainElemOffset.sle(RHS: IOffset) && |
| 1330 | (ChainElemOffset + ChainElemSize).sgt(RHS: OffsetIt->second))) { |
| 1331 | LLVM_DEBUG({ |
| 1332 | // Double check that AA also sees this alias. If not, we probably |
| 1333 | // have a bug. |
| 1334 | ModRefInfo MR = |
| 1335 | BatchAA.getModRefInfo(I, MemoryLocation::get(ChainElem)); |
| 1336 | assert(IsLoadChain ? isModSet(MR) : isModOrRefSet(MR)); |
| 1337 | dbgs() << "LSV: Found alias in chain: " << *I << "\n" ; |
| 1338 | }); |
| 1339 | return false; // We found an aliasing instruction; bail. |
| 1340 | } |
| 1341 | |
| 1342 | continue; // We're confident there's no alias. |
| 1343 | } |
| 1344 | |
| 1345 | LLVM_DEBUG(dbgs() << "LSV: Querying AA for " << *I << "\n" ); |
| 1346 | ModRefInfo MR = BatchAA.getModRefInfo(I, OptLoc: MemoryLocation::get(Inst: ChainElem)); |
| 1347 | if (IsLoadChain ? isModSet(MRI: MR) : isModOrRefSet(MRI: MR)) { |
| 1348 | LLVM_DEBUG(dbgs() << "LSV: Found alias in chain:\n" |
| 1349 | << " Aliasing instruction:\n" |
| 1350 | << " " << *I << '\n' |
| 1351 | << " Aliased instruction and pointer:\n" |
| 1352 | << " " << *ChainElem << '\n' |
| 1353 | << " " << *getLoadStorePointerOperand(ChainElem) |
| 1354 | << '\n'); |
| 1355 | |
| 1356 | return false; |
| 1357 | } |
| 1358 | } |
| 1359 | return true; |
| 1360 | } |
| 1361 | |
| 1362 | static bool checkNoWrapFlags(Instruction *I, bool Signed) { |
| 1363 | BinaryOperator *BinOpI = cast<BinaryOperator>(Val: I); |
| 1364 | return (Signed && BinOpI->hasNoSignedWrap()) || |
| 1365 | (!Signed && BinOpI->hasNoUnsignedWrap()); |
| 1366 | } |
| 1367 | |
| 1368 | static bool checkIfSafeAddSequence(const APInt &IdxDiff, Instruction *AddOpA, |
| 1369 | unsigned MatchingOpIdxA, Instruction *AddOpB, |
| 1370 | unsigned MatchingOpIdxB, bool Signed) { |
| 1371 | LLVM_DEBUG(dbgs() << "LSV: checkIfSafeAddSequence IdxDiff=" << IdxDiff |
| 1372 | << ", AddOpA=" << *AddOpA << ", MatchingOpIdxA=" |
| 1373 | << MatchingOpIdxA << ", AddOpB=" << *AddOpB |
| 1374 | << ", MatchingOpIdxB=" << MatchingOpIdxB |
| 1375 | << ", Signed=" << Signed << "\n" ); |
| 1376 | // If both OpA and OpB are adds with NSW/NUW and with one of the operands |
| 1377 | // being the same, we can guarantee that the transformation is safe if we can |
| 1378 | // prove that OpA won't overflow when Ret added to the other operand of OpA. |
| 1379 | // For example: |
| 1380 | // %tmp7 = add nsw i32 %tmp2, %v0 |
| 1381 | // %tmp8 = sext i32 %tmp7 to i64 |
| 1382 | // ... |
| 1383 | // %tmp11 = add nsw i32 %v0, 1 |
| 1384 | // %tmp12 = add nsw i32 %tmp2, %tmp11 |
| 1385 | // %tmp13 = sext i32 %tmp12 to i64 |
| 1386 | // |
| 1387 | // Both %tmp7 and %tmp12 have the nsw flag and the first operand is %tmp2. |
| 1388 | // It's guaranteed that adding 1 to %tmp7 won't overflow because %tmp11 adds |
| 1389 | // 1 to %v0 and both %tmp11 and %tmp12 have the nsw flag. |
| 1390 | assert(AddOpA->getOpcode() == Instruction::Add && |
| 1391 | AddOpB->getOpcode() == Instruction::Add && |
| 1392 | checkNoWrapFlags(AddOpA, Signed) && checkNoWrapFlags(AddOpB, Signed)); |
| 1393 | if (AddOpA->getOperand(i: MatchingOpIdxA) == |
| 1394 | AddOpB->getOperand(i: MatchingOpIdxB)) { |
| 1395 | Value *OtherOperandA = AddOpA->getOperand(i: MatchingOpIdxA == 1 ? 0 : 1); |
| 1396 | Value *OtherOperandB = AddOpB->getOperand(i: MatchingOpIdxB == 1 ? 0 : 1); |
| 1397 | Instruction *OtherInstrA = dyn_cast<Instruction>(Val: OtherOperandA); |
| 1398 | Instruction *OtherInstrB = dyn_cast<Instruction>(Val: OtherOperandB); |
| 1399 | // Match `x +nsw/nuw y` and `x +nsw/nuw (y +nsw/nuw IdxDiff)`. |
| 1400 | if (OtherInstrB && OtherInstrB->getOpcode() == Instruction::Add && |
| 1401 | checkNoWrapFlags(I: OtherInstrB, Signed) && |
| 1402 | isa<ConstantInt>(Val: OtherInstrB->getOperand(i: 1))) { |
| 1403 | int64_t CstVal = |
| 1404 | cast<ConstantInt>(Val: OtherInstrB->getOperand(i: 1))->getSExtValue(); |
| 1405 | if (OtherInstrB->getOperand(i: 0) == OtherOperandA && |
| 1406 | IdxDiff.getSExtValue() == CstVal) |
| 1407 | return true; |
| 1408 | } |
| 1409 | // Match `x +nsw/nuw (y +nsw/nuw -Idx)` and `x +nsw/nuw (y +nsw/nuw x)`. |
| 1410 | if (OtherInstrA && OtherInstrA->getOpcode() == Instruction::Add && |
| 1411 | checkNoWrapFlags(I: OtherInstrA, Signed) && |
| 1412 | isa<ConstantInt>(Val: OtherInstrA->getOperand(i: 1))) { |
| 1413 | int64_t CstVal = |
| 1414 | cast<ConstantInt>(Val: OtherInstrA->getOperand(i: 1))->getSExtValue(); |
| 1415 | if (OtherInstrA->getOperand(i: 0) == OtherOperandB && |
| 1416 | IdxDiff.getSExtValue() == -CstVal) |
| 1417 | return true; |
| 1418 | } |
| 1419 | // Match `x +nsw/nuw (y +nsw/nuw c)` and |
| 1420 | // `x +nsw/nuw (y +nsw/nuw (c + IdxDiff))`. |
| 1421 | if (OtherInstrA && OtherInstrB && |
| 1422 | OtherInstrA->getOpcode() == Instruction::Add && |
| 1423 | OtherInstrB->getOpcode() == Instruction::Add && |
| 1424 | checkNoWrapFlags(I: OtherInstrA, Signed) && |
| 1425 | checkNoWrapFlags(I: OtherInstrB, Signed) && |
| 1426 | isa<ConstantInt>(Val: OtherInstrA->getOperand(i: 1)) && |
| 1427 | isa<ConstantInt>(Val: OtherInstrB->getOperand(i: 1))) { |
| 1428 | int64_t CstValA = |
| 1429 | cast<ConstantInt>(Val: OtherInstrA->getOperand(i: 1))->getSExtValue(); |
| 1430 | int64_t CstValB = |
| 1431 | cast<ConstantInt>(Val: OtherInstrB->getOperand(i: 1))->getSExtValue(); |
| 1432 | if (OtherInstrA->getOperand(i: 0) == OtherInstrB->getOperand(i: 0) && |
| 1433 | IdxDiff.getSExtValue() == (CstValB - CstValA)) |
| 1434 | return true; |
| 1435 | } |
| 1436 | } |
| 1437 | return false; |
| 1438 | } |
| 1439 | |
| 1440 | std::optional<APInt> Vectorizer::getConstantOffsetComplexAddrs( |
| 1441 | Value *PtrA, Value *PtrB, Instruction *ContextInst, unsigned Depth) { |
| 1442 | LLVM_DEBUG(dbgs() << "LSV: getConstantOffsetComplexAddrs PtrA=" << *PtrA |
| 1443 | << " PtrB=" << *PtrB << " ContextInst=" << *ContextInst |
| 1444 | << " Depth=" << Depth << "\n" ); |
| 1445 | auto *GEPA = dyn_cast<GetElementPtrInst>(Val: PtrA); |
| 1446 | auto *GEPB = dyn_cast<GetElementPtrInst>(Val: PtrB); |
| 1447 | if (!GEPA || !GEPB) |
| 1448 | return getConstantOffsetSelects(PtrA, PtrB, ContextInst, Depth); |
| 1449 | |
| 1450 | // Look through GEPs after checking they're the same except for the last |
| 1451 | // index. |
| 1452 | if (GEPA->getNumOperands() != GEPB->getNumOperands() || |
| 1453 | GEPA->getPointerOperand() != GEPB->getPointerOperand()) |
| 1454 | return std::nullopt; |
| 1455 | gep_type_iterator GTIA = gep_type_begin(GEP: GEPA); |
| 1456 | gep_type_iterator GTIB = gep_type_begin(GEP: GEPB); |
| 1457 | for (unsigned I = 0, E = GEPA->getNumIndices() - 1; I < E; ++I) { |
| 1458 | if (GTIA.getOperand() != GTIB.getOperand()) |
| 1459 | return std::nullopt; |
| 1460 | ++GTIA; |
| 1461 | ++GTIB; |
| 1462 | } |
| 1463 | |
| 1464 | Instruction *OpA = dyn_cast<Instruction>(Val: GTIA.getOperand()); |
| 1465 | Instruction *OpB = dyn_cast<Instruction>(Val: GTIB.getOperand()); |
| 1466 | if (!OpA || !OpB || OpA->getOpcode() != OpB->getOpcode() || |
| 1467 | OpA->getType() != OpB->getType()) |
| 1468 | return std::nullopt; |
| 1469 | |
| 1470 | uint64_t Stride = GTIA.getSequentialElementStride(DL); |
| 1471 | |
| 1472 | // Only look through a ZExt/SExt. |
| 1473 | if (!isa<SExtInst>(Val: OpA) && !isa<ZExtInst>(Val: OpA)) |
| 1474 | return std::nullopt; |
| 1475 | |
| 1476 | bool Signed = isa<SExtInst>(Val: OpA); |
| 1477 | |
| 1478 | // At this point A could be a function parameter, i.e. not an instruction |
| 1479 | Value *ValA = OpA->getOperand(i: 0); |
| 1480 | OpB = dyn_cast<Instruction>(Val: OpB->getOperand(i: 0)); |
| 1481 | if (!OpB || ValA->getType() != OpB->getType()) |
| 1482 | return std::nullopt; |
| 1483 | |
| 1484 | const SCEV *OffsetSCEVA = SE.getSCEV(V: ValA); |
| 1485 | const SCEV *OffsetSCEVB = SE.getSCEV(V: OpB); |
| 1486 | const SCEV *IdxDiffSCEV = SE.getMinusSCEV(LHS: OffsetSCEVB, RHS: OffsetSCEVA); |
| 1487 | if (IdxDiffSCEV == SE.getCouldNotCompute()) |
| 1488 | return std::nullopt; |
| 1489 | |
| 1490 | ConstantRange IdxDiffRange = SE.getSignedRange(S: IdxDiffSCEV); |
| 1491 | if (!IdxDiffRange.isSingleElement()) |
| 1492 | return std::nullopt; |
| 1493 | APInt IdxDiff = *IdxDiffRange.getSingleElement(); |
| 1494 | |
| 1495 | LLVM_DEBUG(dbgs() << "LSV: getConstantOffsetComplexAddrs IdxDiff=" << IdxDiff |
| 1496 | << "\n" ); |
| 1497 | |
| 1498 | // Now we need to prove that adding IdxDiff to ValA won't overflow. |
| 1499 | bool Safe = false; |
| 1500 | |
| 1501 | // First attempt: if OpB is an add with NSW/NUW, and OpB is IdxDiff added to |
| 1502 | // ValA, we're okay. |
| 1503 | if (OpB->getOpcode() == Instruction::Add && |
| 1504 | isa<ConstantInt>(Val: OpB->getOperand(i: 1)) && |
| 1505 | IdxDiff.sle(RHS: cast<ConstantInt>(Val: OpB->getOperand(i: 1))->getSExtValue()) && |
| 1506 | checkNoWrapFlags(I: OpB, Signed)) |
| 1507 | Safe = true; |
| 1508 | |
| 1509 | // Second attempt: check if we have eligible add NSW/NUW instruction |
| 1510 | // sequences. |
| 1511 | OpA = dyn_cast<Instruction>(Val: ValA); |
| 1512 | if (!Safe && OpA && OpA->getOpcode() == Instruction::Add && |
| 1513 | OpB->getOpcode() == Instruction::Add && checkNoWrapFlags(I: OpA, Signed) && |
| 1514 | checkNoWrapFlags(I: OpB, Signed)) { |
| 1515 | // In the checks below a matching operand in OpA and OpB is an operand which |
| 1516 | // is the same in those two instructions. Below we account for possible |
| 1517 | // orders of the operands of these add instructions. |
| 1518 | for (unsigned MatchingOpIdxA : {0, 1}) |
| 1519 | for (unsigned MatchingOpIdxB : {0, 1}) |
| 1520 | if (!Safe) |
| 1521 | Safe = checkIfSafeAddSequence(IdxDiff, AddOpA: OpA, MatchingOpIdxA, AddOpB: OpB, |
| 1522 | MatchingOpIdxB, Signed); |
| 1523 | } |
| 1524 | |
| 1525 | unsigned BitWidth = ValA->getType()->getScalarSizeInBits(); |
| 1526 | |
| 1527 | // Third attempt: |
| 1528 | // |
| 1529 | // Assuming IdxDiff is positive: If all set bits of IdxDiff or any higher |
| 1530 | // order bit other than the sign bit are known to be zero in ValA, we can add |
| 1531 | // Diff to it while guaranteeing no overflow of any sort. |
| 1532 | // |
| 1533 | // If IdxDiff is negative, do the same, but swap ValA and ValB. |
| 1534 | if (!Safe) { |
| 1535 | // When computing known bits, use the GEPs as context instructions, since |
| 1536 | // they likely are in the same BB as the load/store. |
| 1537 | KnownBits Known(BitWidth); |
| 1538 | computeKnownBits(V: (IdxDiff.sge(RHS: 0) ? ValA : OpB), Known, DL, AC: &AC, CxtI: ContextInst, |
| 1539 | DT: &DT); |
| 1540 | APInt BitsAllowedToBeSet = Known.Zero.zext(width: IdxDiff.getBitWidth()); |
| 1541 | if (Signed) |
| 1542 | BitsAllowedToBeSet.clearBit(BitPosition: BitWidth - 1); |
| 1543 | Safe = BitsAllowedToBeSet.uge(RHS: IdxDiff.abs()); |
| 1544 | } |
| 1545 | |
| 1546 | if (Safe) |
| 1547 | return IdxDiff * Stride; |
| 1548 | return std::nullopt; |
| 1549 | } |
| 1550 | |
| 1551 | std::optional<APInt> Vectorizer::getConstantOffsetSelects( |
| 1552 | Value *PtrA, Value *PtrB, Instruction *ContextInst, unsigned Depth) { |
| 1553 | if (Depth++ == MaxDepth) |
| 1554 | return std::nullopt; |
| 1555 | |
| 1556 | if (auto *SelectA = dyn_cast<SelectInst>(Val: PtrA)) { |
| 1557 | if (auto *SelectB = dyn_cast<SelectInst>(Val: PtrB)) { |
| 1558 | if (SelectA->getCondition() != SelectB->getCondition()) |
| 1559 | return std::nullopt; |
| 1560 | LLVM_DEBUG(dbgs() << "LSV: getConstantOffsetSelects, PtrA=" << *PtrA |
| 1561 | << ", PtrB=" << *PtrB << ", ContextInst=" |
| 1562 | << *ContextInst << ", Depth=" << Depth << "\n" ); |
| 1563 | std::optional<APInt> TrueDiff = getConstantOffset( |
| 1564 | PtrA: SelectA->getTrueValue(), PtrB: SelectB->getTrueValue(), ContextInst, Depth); |
| 1565 | if (!TrueDiff) |
| 1566 | return std::nullopt; |
| 1567 | std::optional<APInt> FalseDiff = |
| 1568 | getConstantOffset(PtrA: SelectA->getFalseValue(), PtrB: SelectB->getFalseValue(), |
| 1569 | ContextInst, Depth); |
| 1570 | if (TrueDiff == FalseDiff) |
| 1571 | return TrueDiff; |
| 1572 | } |
| 1573 | } |
| 1574 | return std::nullopt; |
| 1575 | } |
| 1576 | |
| 1577 | void Vectorizer::mergeEquivalenceClasses(EquivalenceClassMap &EQClasses) const { |
| 1578 | if (EQClasses.size() < 2) // There is nothing to merge. |
| 1579 | return; |
| 1580 | |
| 1581 | // The reduced key has all elements of the ECClassKey except the underlying |
| 1582 | // object. Check that EqClassKey has 4 elements and define the reduced key. |
| 1583 | static_assert(std::tuple_size_v<EqClassKey> == 4, |
| 1584 | "EqClassKey has changed - EqClassReducedKey needs changes too" ); |
| 1585 | using EqClassReducedKey = |
| 1586 | std::tuple<std::tuple_element_t<1, EqClassKey> /* AddrSpace */, |
| 1587 | std::tuple_element_t<2, EqClassKey> /* Element size */, |
| 1588 | std::tuple_element_t<3, EqClassKey> /* IsLoad; */>; |
| 1589 | using ECReducedKeyToUnderlyingObjectMap = |
| 1590 | MapVector<EqClassReducedKey, |
| 1591 | SmallPtrSet<std::tuple_element_t<0, EqClassKey>, 4>>; |
| 1592 | |
| 1593 | // Form a map from the reduced key (without the underlying object) to the |
| 1594 | // underlying objects: 1 reduced key to many underlying objects, to form |
| 1595 | // groups of potentially merge-able equivalence classes. |
| 1596 | ECReducedKeyToUnderlyingObjectMap RedKeyToUOMap; |
| 1597 | bool FoundPotentiallyOptimizableEC = false; |
| 1598 | for (const auto &EC : EQClasses) { |
| 1599 | const auto &Key = EC.first; |
| 1600 | EqClassReducedKey RedKey{std::get<1>(t: Key), std::get<2>(t: Key), |
| 1601 | std::get<3>(t: Key)}; |
| 1602 | auto &UOMap = RedKeyToUOMap[RedKey]; |
| 1603 | UOMap.insert(Ptr: std::get<0>(t: Key)); |
| 1604 | if (UOMap.size() > 1) |
| 1605 | FoundPotentiallyOptimizableEC = true; |
| 1606 | } |
| 1607 | if (!FoundPotentiallyOptimizableEC) |
| 1608 | return; |
| 1609 | |
| 1610 | LLVM_DEBUG({ |
| 1611 | dbgs() << "LSV: mergeEquivalenceClasses: before merging:\n" ; |
| 1612 | for (const auto &EC : EQClasses) { |
| 1613 | dbgs() << " Key: {" << EC.first << "}\n" ; |
| 1614 | for (const auto &Inst : EC.second) |
| 1615 | dbgs() << " Inst: " << *Inst << '\n'; |
| 1616 | } |
| 1617 | }); |
| 1618 | LLVM_DEBUG({ |
| 1619 | dbgs() << "LSV: mergeEquivalenceClasses: RedKeyToUOMap:\n" ; |
| 1620 | for (const auto &RedKeyToUO : RedKeyToUOMap) { |
| 1621 | dbgs() << " Reduced key: {" << std::get<0>(RedKeyToUO.first) << ", " |
| 1622 | << std::get<1>(RedKeyToUO.first) << ", " |
| 1623 | << static_cast<int>(std::get<2>(RedKeyToUO.first)) << "} --> " |
| 1624 | << RedKeyToUO.second.size() << " underlying objects:\n" ; |
| 1625 | for (auto UObject : RedKeyToUO.second) |
| 1626 | dbgs() << " " << *UObject << '\n'; |
| 1627 | } |
| 1628 | }); |
| 1629 | |
| 1630 | using UObjectToUObjectMap = DenseMap<const Value *, const Value *>; |
| 1631 | |
| 1632 | // Compute the ultimate targets for a set of underlying objects. |
| 1633 | auto GetUltimateTargets = |
| 1634 | [](SmallPtrSetImpl<const Value *> &UObjects) -> UObjectToUObjectMap { |
| 1635 | UObjectToUObjectMap IndirectionMap; |
| 1636 | for (const auto *UObject : UObjects) { |
| 1637 | const unsigned MaxLookupDepth = 1; // look for 1-level indirections only |
| 1638 | const auto *UltimateTarget = getUnderlyingObject(V: UObject, MaxLookup: MaxLookupDepth); |
| 1639 | if (UltimateTarget != UObject) |
| 1640 | IndirectionMap[UObject] = UltimateTarget; |
| 1641 | } |
| 1642 | UObjectToUObjectMap UltimateTargetsMap; |
| 1643 | for (const auto *UObject : UObjects) { |
| 1644 | auto Target = UObject; |
| 1645 | auto It = IndirectionMap.find(Val: Target); |
| 1646 | for (; It != IndirectionMap.end(); It = IndirectionMap.find(Val: Target)) |
| 1647 | Target = It->second; |
| 1648 | UltimateTargetsMap[UObject] = Target; |
| 1649 | } |
| 1650 | return UltimateTargetsMap; |
| 1651 | }; |
| 1652 | |
| 1653 | // For each item in RedKeyToUOMap, if it has more than one underlying object, |
| 1654 | // try to merge the equivalence classes. |
| 1655 | for (auto &[RedKey, UObjects] : RedKeyToUOMap) { |
| 1656 | if (UObjects.size() < 2) |
| 1657 | continue; |
| 1658 | auto UTMap = GetUltimateTargets(UObjects); |
| 1659 | for (const auto &[UObject, UltimateTarget] : UTMap) { |
| 1660 | if (UObject == UltimateTarget) |
| 1661 | continue; |
| 1662 | |
| 1663 | EqClassKey KeyFrom{UObject, std::get<0>(t&: RedKey), std::get<1>(t&: RedKey), |
| 1664 | std::get<2>(t&: RedKey)}; |
| 1665 | EqClassKey KeyTo{UltimateTarget, std::get<0>(t&: RedKey), std::get<1>(t&: RedKey), |
| 1666 | std::get<2>(t&: RedKey)}; |
| 1667 | // The entry for KeyFrom is guarantted to exist, unlike KeyTo. Thus, |
| 1668 | // request the reference to the instructions vector for KeyTo first. |
| 1669 | const auto &VecTo = EQClasses[KeyTo]; |
| 1670 | const auto &VecFrom = EQClasses[KeyFrom]; |
| 1671 | SmallVector<Instruction *, 8> MergedVec; |
| 1672 | std::merge(first1: VecFrom.begin(), last1: VecFrom.end(), first2: VecTo.begin(), last2: VecTo.end(), |
| 1673 | result: std::back_inserter(x&: MergedVec), |
| 1674 | comp: [](Instruction *A, Instruction *B) { |
| 1675 | return A && B && A->comesBefore(Other: B); |
| 1676 | }); |
| 1677 | EQClasses[KeyTo] = std::move(MergedVec); |
| 1678 | EQClasses.erase(Key: KeyFrom); |
| 1679 | } |
| 1680 | } |
| 1681 | LLVM_DEBUG({ |
| 1682 | dbgs() << "LSV: mergeEquivalenceClasses: after merging:\n" ; |
| 1683 | for (const auto &EC : EQClasses) { |
| 1684 | dbgs() << " Key: {" << EC.first << "}\n" ; |
| 1685 | for (const auto &Inst : EC.second) |
| 1686 | dbgs() << " Inst: " << *Inst << '\n'; |
| 1687 | } |
| 1688 | }); |
| 1689 | } |
| 1690 | |
| 1691 | EquivalenceClassMap |
| 1692 | Vectorizer::collectEquivalenceClasses(BasicBlock::iterator Begin, |
| 1693 | BasicBlock::iterator End) { |
| 1694 | EquivalenceClassMap Ret; |
| 1695 | |
| 1696 | auto GetUnderlyingObject = [](const Value *Ptr) -> const Value * { |
| 1697 | const Value *ObjPtr = llvm::getUnderlyingObject(V: Ptr); |
| 1698 | if (const auto *Sel = dyn_cast<SelectInst>(Val: ObjPtr)) { |
| 1699 | // The select's themselves are distinct instructions even if they share |
| 1700 | // the same condition and evaluate to consecutive pointers for true and |
| 1701 | // false values of the condition. Therefore using the select's themselves |
| 1702 | // for grouping instructions would put consecutive accesses into different |
| 1703 | // lists and they won't be even checked for being consecutive, and won't |
| 1704 | // be vectorized. |
| 1705 | return Sel->getCondition(); |
| 1706 | } |
| 1707 | return ObjPtr; |
| 1708 | }; |
| 1709 | |
| 1710 | for (Instruction &I : make_range(x: Begin, y: End)) { |
| 1711 | auto *LI = dyn_cast<LoadInst>(Val: &I); |
| 1712 | auto *SI = dyn_cast<StoreInst>(Val: &I); |
| 1713 | if (!LI && !SI) |
| 1714 | continue; |
| 1715 | |
| 1716 | if ((LI && !LI->isSimple()) || (SI && !SI->isSimple())) |
| 1717 | continue; |
| 1718 | |
| 1719 | if ((LI && !TTI.isLegalToVectorizeLoad(LI)) || |
| 1720 | (SI && !TTI.isLegalToVectorizeStore(SI))) |
| 1721 | continue; |
| 1722 | |
| 1723 | Type *Ty = getLoadStoreType(I: &I); |
| 1724 | if (!VectorType::isValidElementType(ElemTy: Ty->getScalarType())) |
| 1725 | continue; |
| 1726 | |
| 1727 | // Skip weird non-byte sizes. They probably aren't worth the effort of |
| 1728 | // handling correctly. |
| 1729 | unsigned TySize = DL.getTypeSizeInBits(Ty); |
| 1730 | if ((TySize % 8) != 0) |
| 1731 | continue; |
| 1732 | |
| 1733 | // Skip vectors of pointers. The vectorizeLoadChain/vectorizeStoreChain |
| 1734 | // functions are currently using an integer type for the vectorized |
| 1735 | // load/store, and does not support casting between the integer type and a |
| 1736 | // vector of pointers (e.g. i64 to <2 x i16*>) |
| 1737 | if (Ty->isVectorTy() && Ty->isPtrOrPtrVectorTy()) |
| 1738 | continue; |
| 1739 | |
| 1740 | Value *Ptr = getLoadStorePointerOperand(V: &I); |
| 1741 | unsigned AS = Ptr->getType()->getPointerAddressSpace(); |
| 1742 | unsigned VecRegSize = TTI.getLoadStoreVecRegBitWidth(AddrSpace: AS); |
| 1743 | |
| 1744 | unsigned VF = VecRegSize / TySize; |
| 1745 | VectorType *VecTy = dyn_cast<VectorType>(Val: Ty); |
| 1746 | |
| 1747 | // Only handle power-of-two sized elements. |
| 1748 | if ((!VecTy && !isPowerOf2_32(Value: DL.getTypeSizeInBits(Ty))) || |
| 1749 | (VecTy && !isPowerOf2_32(Value: DL.getTypeSizeInBits(Ty: VecTy->getScalarType())))) |
| 1750 | continue; |
| 1751 | |
| 1752 | // No point in looking at these if they're too big to vectorize. |
| 1753 | if (TySize > VecRegSize / 2 || |
| 1754 | (VecTy && TTI.getLoadVectorFactor(VF, LoadSize: TySize, ChainSizeInBytes: TySize / 8, VecTy) == 0)) |
| 1755 | continue; |
| 1756 | |
| 1757 | Ret[{GetUnderlyingObject(Ptr), AS, |
| 1758 | DL.getTypeSizeInBits(Ty: getLoadStoreType(I: &I)->getScalarType()), |
| 1759 | /*IsLoad=*/LI != nullptr}] |
| 1760 | .emplace_back(Args: &I); |
| 1761 | } |
| 1762 | |
| 1763 | mergeEquivalenceClasses(EQClasses&: Ret); |
| 1764 | return Ret; |
| 1765 | } |
| 1766 | |
| 1767 | std::vector<Chain> Vectorizer::gatherChains(ArrayRef<Instruction *> Instrs) { |
| 1768 | if (Instrs.empty()) |
| 1769 | return {}; |
| 1770 | |
| 1771 | unsigned AS = getLoadStoreAddressSpace(I: Instrs[0]); |
| 1772 | unsigned ASPtrBits = DL.getIndexSizeInBits(AS); |
| 1773 | |
| 1774 | #ifndef NDEBUG |
| 1775 | // Check that Instrs is in BB order and all have the same addr space. |
| 1776 | for (size_t I = 1; I < Instrs.size(); ++I) { |
| 1777 | assert(Instrs[I - 1]->comesBefore(Instrs[I])); |
| 1778 | assert(getLoadStoreAddressSpace(Instrs[I]) == AS); |
| 1779 | } |
| 1780 | #endif |
| 1781 | |
| 1782 | // Machinery to build an MRU-hashtable of Chains. |
| 1783 | // |
| 1784 | // (Ideally this could be done with MapVector, but as currently implemented, |
| 1785 | // moving an element to the front of a MapVector is O(n).) |
| 1786 | struct InstrListElem : ilist_node<InstrListElem>, |
| 1787 | std::pair<Instruction *, Chain> { |
| 1788 | explicit InstrListElem(Instruction *I) |
| 1789 | : std::pair<Instruction *, Chain>(I, {}) {} |
| 1790 | }; |
| 1791 | struct InstrListElemDenseMapInfo { |
| 1792 | using PtrInfo = DenseMapInfo<InstrListElem *>; |
| 1793 | using IInfo = DenseMapInfo<Instruction *>; |
| 1794 | static InstrListElem *getEmptyKey() { return PtrInfo::getEmptyKey(); } |
| 1795 | static InstrListElem *getTombstoneKey() { |
| 1796 | return PtrInfo::getTombstoneKey(); |
| 1797 | } |
| 1798 | static unsigned getHashValue(const InstrListElem *E) { |
| 1799 | return IInfo::getHashValue(PtrVal: E->first); |
| 1800 | } |
| 1801 | static bool isEqual(const InstrListElem *A, const InstrListElem *B) { |
| 1802 | if (A == getEmptyKey() || B == getEmptyKey()) |
| 1803 | return A == getEmptyKey() && B == getEmptyKey(); |
| 1804 | if (A == getTombstoneKey() || B == getTombstoneKey()) |
| 1805 | return A == getTombstoneKey() && B == getTombstoneKey(); |
| 1806 | return IInfo::isEqual(LHS: A->first, RHS: B->first); |
| 1807 | } |
| 1808 | }; |
| 1809 | SpecificBumpPtrAllocator<InstrListElem> Allocator; |
| 1810 | simple_ilist<InstrListElem> MRU; |
| 1811 | DenseSet<InstrListElem *, InstrListElemDenseMapInfo> Chains; |
| 1812 | |
| 1813 | // Compare each instruction in `instrs` to leader of the N most recently-used |
| 1814 | // chains. This limits the O(n^2) behavior of this pass while also allowing |
| 1815 | // us to build arbitrarily long chains. |
| 1816 | for (Instruction *I : Instrs) { |
| 1817 | constexpr int MaxChainsToTry = 64; |
| 1818 | |
| 1819 | bool MatchFound = false; |
| 1820 | auto ChainIter = MRU.begin(); |
| 1821 | for (size_t J = 0; J < MaxChainsToTry && ChainIter != MRU.end(); |
| 1822 | ++J, ++ChainIter) { |
| 1823 | if (std::optional<APInt> Offset = getConstantOffset( |
| 1824 | PtrA: getLoadStorePointerOperand(V: ChainIter->first), |
| 1825 | PtrB: getLoadStorePointerOperand(V: I), |
| 1826 | /*ContextInst=*/ |
| 1827 | (ChainIter->first->comesBefore(Other: I) ? I : ChainIter->first))) { |
| 1828 | // `Offset` might not have the expected number of bits, if e.g. AS has a |
| 1829 | // different number of bits than opaque pointers. |
| 1830 | ChainIter->second.emplace_back(Args&: I, Args&: Offset.value()); |
| 1831 | // Move ChainIter to the front of the MRU list. |
| 1832 | MRU.remove(N&: *ChainIter); |
| 1833 | MRU.push_front(Node&: *ChainIter); |
| 1834 | MatchFound = true; |
| 1835 | break; |
| 1836 | } |
| 1837 | } |
| 1838 | |
| 1839 | if (!MatchFound) { |
| 1840 | APInt ZeroOffset(ASPtrBits, 0); |
| 1841 | InstrListElem *E = new (Allocator.Allocate()) InstrListElem(I); |
| 1842 | E->second.emplace_back(Args&: I, Args&: ZeroOffset); |
| 1843 | MRU.push_front(Node&: *E); |
| 1844 | Chains.insert(V: E); |
| 1845 | } |
| 1846 | } |
| 1847 | |
| 1848 | std::vector<Chain> Ret; |
| 1849 | Ret.reserve(n: Chains.size()); |
| 1850 | // Iterate over MRU rather than Chains so the order is deterministic. |
| 1851 | for (auto &E : MRU) |
| 1852 | if (E.second.size() > 1) |
| 1853 | Ret.emplace_back(args: std::move(E.second)); |
| 1854 | return Ret; |
| 1855 | } |
| 1856 | |
| 1857 | std::optional<APInt> Vectorizer::getConstantOffset(Value *PtrA, Value *PtrB, |
| 1858 | Instruction *ContextInst, |
| 1859 | unsigned Depth) { |
| 1860 | LLVM_DEBUG(dbgs() << "LSV: getConstantOffset, PtrA=" << *PtrA |
| 1861 | << ", PtrB=" << *PtrB << ", ContextInst= " << *ContextInst |
| 1862 | << ", Depth=" << Depth << "\n" ); |
| 1863 | // We'll ultimately return a value of this bit width, even if computations |
| 1864 | // happen in a different width. |
| 1865 | unsigned OrigBitWidth = DL.getIndexTypeSizeInBits(Ty: PtrA->getType()); |
| 1866 | APInt OffsetA(OrigBitWidth, 0); |
| 1867 | APInt OffsetB(OrigBitWidth, 0); |
| 1868 | PtrA = PtrA->stripAndAccumulateInBoundsConstantOffsets(DL, Offset&: OffsetA); |
| 1869 | PtrB = PtrB->stripAndAccumulateInBoundsConstantOffsets(DL, Offset&: OffsetB); |
| 1870 | unsigned NewPtrBitWidth = DL.getTypeStoreSizeInBits(Ty: PtrA->getType()); |
| 1871 | if (NewPtrBitWidth != DL.getTypeStoreSizeInBits(Ty: PtrB->getType())) |
| 1872 | return std::nullopt; |
| 1873 | |
| 1874 | // If we have to shrink the pointer, stripAndAccumulateInBoundsConstantOffsets |
| 1875 | // should properly handle a possible overflow and the value should fit into |
| 1876 | // the smallest data type used in the cast/gep chain. |
| 1877 | assert(OffsetA.getSignificantBits() <= NewPtrBitWidth && |
| 1878 | OffsetB.getSignificantBits() <= NewPtrBitWidth); |
| 1879 | |
| 1880 | OffsetA = OffsetA.sextOrTrunc(width: NewPtrBitWidth); |
| 1881 | OffsetB = OffsetB.sextOrTrunc(width: NewPtrBitWidth); |
| 1882 | if (PtrA == PtrB) |
| 1883 | return (OffsetB - OffsetA).sextOrTrunc(width: OrigBitWidth); |
| 1884 | |
| 1885 | // Try to compute B - A. |
| 1886 | const SCEV *DistScev = SE.getMinusSCEV(LHS: SE.getSCEV(V: PtrB), RHS: SE.getSCEV(V: PtrA)); |
| 1887 | if (DistScev != SE.getCouldNotCompute()) { |
| 1888 | LLVM_DEBUG(dbgs() << "LSV: SCEV PtrB - PtrA =" << *DistScev << "\n" ); |
| 1889 | ConstantRange DistRange = SE.getSignedRange(S: DistScev); |
| 1890 | if (DistRange.isSingleElement()) { |
| 1891 | // Handle index width (the width of Dist) != pointer width (the width of |
| 1892 | // the Offset*s at this point). |
| 1893 | APInt Dist = DistRange.getSingleElement()->sextOrTrunc(width: NewPtrBitWidth); |
| 1894 | return (OffsetB - OffsetA + Dist).sextOrTrunc(width: OrigBitWidth); |
| 1895 | } |
| 1896 | } |
| 1897 | if (std::optional<APInt> Diff = |
| 1898 | getConstantOffsetComplexAddrs(PtrA, PtrB, ContextInst, Depth)) |
| 1899 | return (OffsetB - OffsetA + Diff->sext(width: OffsetB.getBitWidth())) |
| 1900 | .sextOrTrunc(width: OrigBitWidth); |
| 1901 | return std::nullopt; |
| 1902 | } |
| 1903 | |
| 1904 | bool Vectorizer::accessIsAllowedAndFast(unsigned SizeBytes, unsigned AS, |
| 1905 | Align Alignment, |
| 1906 | unsigned VecElemBits) const { |
| 1907 | // Aligned vector accesses are ALWAYS faster than element-wise accesses. |
| 1908 | if (Alignment.value() % SizeBytes == 0) |
| 1909 | return true; |
| 1910 | |
| 1911 | // Ask TTI whether misaligned accesses are faster as vector or element-wise. |
| 1912 | unsigned VectorizedSpeed = 0; |
| 1913 | bool AllowsMisaligned = TTI.allowsMisalignedMemoryAccesses( |
| 1914 | Context&: F.getContext(), BitWidth: SizeBytes * 8, AddressSpace: AS, Alignment, Fast: &VectorizedSpeed); |
| 1915 | if (!AllowsMisaligned) { |
| 1916 | LLVM_DEBUG( |
| 1917 | dbgs() << "LSV: Access of " << SizeBytes << "B in addrspace " << AS |
| 1918 | << " with alignment " << Alignment.value() |
| 1919 | << " is misaligned, and therefore can't be vectorized.\n" ); |
| 1920 | return false; |
| 1921 | } |
| 1922 | |
| 1923 | unsigned ElementwiseSpeed = 0; |
| 1924 | (TTI).allowsMisalignedMemoryAccesses(Context&: (F).getContext(), BitWidth: VecElemBits, AddressSpace: AS, |
| 1925 | Alignment, Fast: &ElementwiseSpeed); |
| 1926 | if (VectorizedSpeed < ElementwiseSpeed) { |
| 1927 | LLVM_DEBUG(dbgs() << "LSV: Access of " << SizeBytes << "B in addrspace " |
| 1928 | << AS << " with alignment " << Alignment.value() |
| 1929 | << " has relative speed " << VectorizedSpeed |
| 1930 | << ", which is lower than the elementwise speed of " |
| 1931 | << ElementwiseSpeed |
| 1932 | << ". Therefore this access won't be vectorized.\n" ); |
| 1933 | return false; |
| 1934 | } |
| 1935 | return true; |
| 1936 | } |
| 1937 | |
| 1938 | ChainElem Vectorizer::(const ChainElem &Prev, Type *Ty, |
| 1939 | APInt Offset, StringRef Prefix, |
| 1940 | Align Alignment) { |
| 1941 | Instruction *NewElement = nullptr; |
| 1942 | Builder.SetInsertPoint(Prev.Inst->getNextNode()); |
| 1943 | if (LoadInst *PrevLoad = dyn_cast<LoadInst>(Val: Prev.Inst)) { |
| 1944 | Value *NewGep = Builder.CreatePtrAdd( |
| 1945 | Ptr: PrevLoad->getPointerOperand(), Offset: Builder.getInt(AI: Offset), Name: Prefix + "GEP" ); |
| 1946 | LLVM_DEBUG(dbgs() << "LSV: Extra GEP Created: \n" << *NewGep << "\n" ); |
| 1947 | NewElement = Builder.CreateAlignedLoad(Ty, Ptr: NewGep, Align: Alignment, Name: Prefix); |
| 1948 | } else { |
| 1949 | StoreInst *PrevStore = cast<StoreInst>(Val: Prev.Inst); |
| 1950 | |
| 1951 | Value *NewGep = Builder.CreatePtrAdd( |
| 1952 | Ptr: PrevStore->getPointerOperand(), Offset: Builder.getInt(AI: Offset), Name: Prefix + "GEP" ); |
| 1953 | LLVM_DEBUG(dbgs() << "LSV: Extra GEP Created: \n" << *NewGep << "\n" ); |
| 1954 | NewElement = |
| 1955 | Builder.CreateAlignedStore(Val: PoisonValue::get(T: Ty), Ptr: NewGep, Align: Alignment); |
| 1956 | } |
| 1957 | |
| 1958 | // Attach all metadata to the new element. |
| 1959 | // propagateMetadata will fold it into the final vector when applicable. |
| 1960 | NewElement->copyMetadata(SrcInst: *Prev.Inst); |
| 1961 | |
| 1962 | // Cache created elements for tracking and cleanup |
| 1963 | ExtraElements.insert(V: NewElement); |
| 1964 | |
| 1965 | APInt NewOffsetFromLeader = Prev.OffsetFromLeader + Offset; |
| 1966 | LLVM_DEBUG(dbgs() << "LSV: Extra Element Created: \n" |
| 1967 | << *NewElement |
| 1968 | << " OffsetFromLeader: " << NewOffsetFromLeader << "\n" ); |
| 1969 | return ChainElem{NewElement, NewOffsetFromLeader}; |
| 1970 | } |
| 1971 | |
| 1972 | Value *Vectorizer::(const ArrayRef<ChainElem> C, |
| 1973 | FixedVectorType *VecTy) { |
| 1974 | // Start each mask element as false |
| 1975 | SmallVector<Constant *, 64> MaskElts(VecTy->getNumElements(), |
| 1976 | Builder.getInt1(V: false)); |
| 1977 | // Iterate over the chain and set the corresponding mask element to true for |
| 1978 | // each element that is not an extra element. |
| 1979 | for (const ChainElem &E : C) { |
| 1980 | if (ExtraElements.contains(V: E.Inst)) |
| 1981 | continue; |
| 1982 | unsigned EOffset = |
| 1983 | (E.OffsetFromLeader - C[0].OffsetFromLeader).getZExtValue(); |
| 1984 | unsigned VecIdx = |
| 1985 | 8 * EOffset / DL.getTypeSizeInBits(Ty: VecTy->getScalarType()); |
| 1986 | if (FixedVectorType *VT = |
| 1987 | dyn_cast<FixedVectorType>(Val: getLoadStoreType(I: E.Inst))) |
| 1988 | for (unsigned J = 0; J < VT->getNumElements(); ++J) |
| 1989 | MaskElts[VecIdx + J] = Builder.getInt1(V: true); |
| 1990 | else |
| 1991 | MaskElts[VecIdx] = Builder.getInt1(V: true); |
| 1992 | } |
| 1993 | return ConstantVector::get(V: MaskElts); |
| 1994 | } |
| 1995 | |
| 1996 | void Vectorizer::() { |
| 1997 | for (auto * : ExtraElements) { |
| 1998 | if (isa<LoadInst>(Val: ExtraElement)) { |
| 1999 | [[maybe_unused]] bool Deleted = |
| 2000 | RecursivelyDeleteTriviallyDeadInstructions(V: ExtraElement); |
| 2001 | assert(Deleted && "Extra Load should always be trivially dead" ); |
| 2002 | } else { |
| 2003 | // Unlike Extra Loads, Extra Stores won't be "dead", but should all be |
| 2004 | // deleted regardless. They will have either been combined into a masked |
| 2005 | // store, or will be left behind and need to be cleaned up. |
| 2006 | auto *PtrOperand = getLoadStorePointerOperand(V: ExtraElement); |
| 2007 | ExtraElement->eraseFromParent(); |
| 2008 | RecursivelyDeleteTriviallyDeadInstructions(V: PtrOperand); |
| 2009 | } |
| 2010 | } |
| 2011 | |
| 2012 | ExtraElements.clear(); |
| 2013 | } |
| 2014 | |