| 1 | //===- SLPMemoryUtils.cpp - SLP pointer/stride helpers --------------------===// |
| 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 | #include "SLPMemoryUtils.h" |
| 10 | #include "SLPCompatibilityAnalysis.h" |
| 11 | #include "SLPCostAnalysis.h" |
| 12 | #include "SLPTypeUtils.h" |
| 13 | #include "SLPUtils.h" |
| 14 | |
| 15 | #include "llvm/ADT/APInt.h" |
| 16 | #include "llvm/ADT/MapVector.h" |
| 17 | #include "llvm/ADT/STLExtras.h" |
| 18 | #include "llvm/ADT/Sequence.h" |
| 19 | #include "llvm/ADT/SmallPtrSet.h" |
| 20 | #include "llvm/Analysis/Loads.h" |
| 21 | #include "llvm/Analysis/LoopAccessAnalysis.h" |
| 22 | #include "llvm/Analysis/ScalarEvolution.h" |
| 23 | #include "llvm/Analysis/ScalarEvolutionExpressions.h" |
| 24 | #include "llvm/Analysis/TargetTransformInfo.h" |
| 25 | #include "llvm/Analysis/ValueTracking.h" |
| 26 | #include "llvm/IR/DataLayout.h" |
| 27 | #include "llvm/IR/DerivedTypes.h" |
| 28 | #include "llvm/IR/Instructions.h" |
| 29 | #include "llvm/IR/Intrinsics.h" |
| 30 | #include "llvm/Support/InstructionCost.h" |
| 31 | |
| 32 | #include <algorithm> |
| 33 | #include <limits> |
| 34 | #include <optional> |
| 35 | #include <set> |
| 36 | #include <tuple> |
| 37 | #include <utility> |
| 38 | |
| 39 | using namespace llvm; |
| 40 | |
| 41 | namespace llvm::slpvectorizer { |
| 42 | |
| 43 | bool arePointersCompatible(Value *Ptr1, Value *Ptr2, |
| 44 | const TargetLibraryInfo &TLI, unsigned MaxDepth, |
| 45 | bool CompareOpcodes) { |
| 46 | if (getUnderlyingObject(V: Ptr1, MaxLookup: MaxDepth) != |
| 47 | getUnderlyingObject(V: Ptr2, MaxLookup: MaxDepth)) |
| 48 | return false; |
| 49 | auto *GEP1 = dyn_cast<GetElementPtrInst>(Val: Ptr1); |
| 50 | auto *GEP2 = dyn_cast<GetElementPtrInst>(Val: Ptr2); |
| 51 | return (!GEP1 || GEP1->getNumOperands() == 2) && |
| 52 | (!GEP2 || GEP2->getNumOperands() == 2) && |
| 53 | (((!GEP1 || isConstant(V: GEP1->getOperand(i_nocapture: 1))) && |
| 54 | (!GEP2 || isConstant(V: GEP2->getOperand(i_nocapture: 1)))) || |
| 55 | !CompareOpcodes || |
| 56 | (GEP1 && GEP2 && |
| 57 | getSameOpcode(VL: {GEP1->getOperand(i_nocapture: 1), GEP2->getOperand(i_nocapture: 1)}, TLI))); |
| 58 | } |
| 59 | |
| 60 | /// Calculates minimal alignment as a common alignment. |
| 61 | template <typename T> Align computeCommonAlignment(ArrayRef<Value *> VL) { |
| 62 | Align CommonAlignment = cast<T>(VL.consume_front())->getAlign(); |
| 63 | for (Value *V : VL) |
| 64 | CommonAlignment = std::min(CommonAlignment, cast<T>(V)->getAlign()); |
| 65 | return CommonAlignment; |
| 66 | } |
| 67 | |
| 68 | template Align computeCommonAlignment<LoadInst>(ArrayRef<Value *>); |
| 69 | template Align computeCommonAlignment<StoreInst>(ArrayRef<Value *>); |
| 70 | |
| 71 | const SCEV *calculateRtStride(ArrayRef<Value *> PointerOps, Type *ElemTy, |
| 72 | const DataLayout &DL, ScalarEvolution &SE, |
| 73 | SmallVectorImpl<unsigned> &SortedIndices) { |
| 74 | SmallVector<const SCEV *> SCEVs; |
| 75 | const SCEV *PtrSCEVLowest = nullptr; |
| 76 | const SCEV *PtrSCEVHighest = nullptr; |
| 77 | // Find lower/upper pointers from the PointerOps (i.e. with lowest and highest |
| 78 | // addresses). |
| 79 | for (Value *Ptr : PointerOps) { |
| 80 | const SCEV *PtrSCEV = SE.getSCEV(V: Ptr); |
| 81 | if (!PtrSCEV) |
| 82 | return nullptr; |
| 83 | SCEVs.push_back(Elt: PtrSCEV); |
| 84 | if (!PtrSCEVLowest && !PtrSCEVHighest) { |
| 85 | PtrSCEVLowest = PtrSCEVHighest = PtrSCEV; |
| 86 | continue; |
| 87 | } |
| 88 | const SCEV *Diff = SE.getMinusSCEV(LHS: PtrSCEV, RHS: PtrSCEVLowest); |
| 89 | if (isa<SCEVCouldNotCompute>(Val: Diff)) |
| 90 | return nullptr; |
| 91 | if (Diff->isNonConstantNegative()) { |
| 92 | PtrSCEVLowest = PtrSCEV; |
| 93 | continue; |
| 94 | } |
| 95 | const SCEV *Diff1 = SE.getMinusSCEV(LHS: PtrSCEVHighest, RHS: PtrSCEV); |
| 96 | if (isa<SCEVCouldNotCompute>(Val: Diff1)) |
| 97 | return nullptr; |
| 98 | if (Diff1->isNonConstantNegative()) { |
| 99 | PtrSCEVHighest = PtrSCEV; |
| 100 | continue; |
| 101 | } |
| 102 | } |
| 103 | // Dist = PtrSCEVHighest - PtrSCEVLowest; |
| 104 | const SCEV *Dist = SE.getMinusSCEV(LHS: PtrSCEVHighest, RHS: PtrSCEVLowest); |
| 105 | if (isa<SCEVCouldNotCompute>(Val: Dist)) |
| 106 | return nullptr; |
| 107 | int Size = DL.getTypeStoreSize(Ty: ElemTy); |
| 108 | auto TryGetStride = [&](const SCEV *Dist, |
| 109 | const SCEV *Multiplier) -> const SCEV * { |
| 110 | if (const auto *M = dyn_cast<SCEVMulExpr>(Val: Dist)) { |
| 111 | if (M->getOperand(i: 0) == Multiplier) |
| 112 | return M->getOperand(i: 1); |
| 113 | if (M->getOperand(i: 1) == Multiplier) |
| 114 | return M->getOperand(i: 0); |
| 115 | return nullptr; |
| 116 | } |
| 117 | if (Multiplier == Dist) |
| 118 | return SE.getConstant(Ty: Dist->getType(), V: 1); |
| 119 | return SE.getUDivExactExpr(LHS: Dist, RHS: Multiplier); |
| 120 | }; |
| 121 | // Stride_in_elements = Dist / element_size * (num_elems - 1). |
| 122 | const SCEV *Stride = nullptr; |
| 123 | if (Size != 1 || SCEVs.size() > 1) { |
| 124 | const SCEV *Sz = SE.getConstant(Ty: Dist->getType(), V: Size * (SCEVs.size() - 1)); |
| 125 | Stride = TryGetStride(Dist, Sz); |
| 126 | if (!Stride) |
| 127 | return nullptr; |
| 128 | } |
| 129 | if (!Stride || isa<SCEVConstant>(Val: Stride)) |
| 130 | return nullptr; |
| 131 | // Iterate through all pointers and check if all distances are |
| 132 | // unique multiple of Stride. |
| 133 | using DistOrdPair = std::pair<int64_t, int>; |
| 134 | auto Compare = llvm::less_first(); |
| 135 | std::set<DistOrdPair, decltype(Compare)> Offsets(Compare); |
| 136 | bool IsConsecutive = true; |
| 137 | for (const auto [Idx, PtrSCEV] : enumerate(First&: SCEVs)) { |
| 138 | unsigned Dist = 0; |
| 139 | if (PtrSCEV != PtrSCEVLowest) { |
| 140 | const SCEV *Diff = SE.getMinusSCEV(LHS: PtrSCEV, RHS: PtrSCEVLowest); |
| 141 | const SCEV *Coeff = TryGetStride(Diff, Stride); |
| 142 | if (!Coeff) |
| 143 | return nullptr; |
| 144 | const auto *SC = dyn_cast<SCEVConstant>(Val: Coeff); |
| 145 | if (!SC || isa<SCEVCouldNotCompute>(Val: SC)) |
| 146 | return nullptr; |
| 147 | if (!SE.getMinusSCEV(LHS: PtrSCEV, RHS: SE.getAddExpr(LHS: PtrSCEVLowest, |
| 148 | RHS: SE.getMulExpr(LHS: Stride, RHS: SC))) |
| 149 | ->isZero()) |
| 150 | return nullptr; |
| 151 | Dist = SC->getAPInt().getZExtValue(); |
| 152 | } |
| 153 | // If the strides are not the same or repeated, we can't vectorize. |
| 154 | if ((Dist / Size) * Size != Dist || (Dist / Size) >= SCEVs.size()) |
| 155 | return nullptr; |
| 156 | auto Res = Offsets.emplace(args&: Dist, args&: Idx); |
| 157 | if (!Res.second) |
| 158 | return nullptr; |
| 159 | // Consecutive order if the inserted element is the last one. |
| 160 | IsConsecutive = IsConsecutive && std::next(x: Res.first) == Offsets.end(); |
| 161 | } |
| 162 | SortedIndices.clear(); |
| 163 | if (!IsConsecutive) { |
| 164 | // Fill SortedIndices array only if it is non-consecutive. |
| 165 | SortedIndices.resize(N: PointerOps.size()); |
| 166 | for (const auto [Idx, Pair] : enumerate(First&: Offsets)) |
| 167 | SortedIndices[Idx] = Pair.second; |
| 168 | } |
| 169 | return Stride; |
| 170 | } |
| 171 | |
| 172 | /// Builds compress-like mask for shuffles for the given \p PointerOps, ordered |
| 173 | /// with \p Order. |
| 174 | /// \return true if the mask represents strided access, false - otherwise. |
| 175 | static bool buildCompressMask(ArrayRef<Value *> PointerOps, |
| 176 | ArrayRef<unsigned> Order, Type *ScalarTy, |
| 177 | const DataLayout &DL, ScalarEvolution &SE, |
| 178 | SmallVectorImpl<int> &CompressMask) { |
| 179 | const unsigned Sz = PointerOps.size(); |
| 180 | CompressMask.assign(NumElts: Sz, Elt: PoisonMaskElem); |
| 181 | // The first element always set. |
| 182 | CompressMask[0] = 0; |
| 183 | // Check if the mask represents strided access. |
| 184 | std::optional<unsigned> Stride = 0; |
| 185 | Value *Ptr0 = Order.empty() ? PointerOps.front() : PointerOps[Order.front()]; |
| 186 | for (unsigned I : seq<unsigned>(Begin: 1, End: Sz)) { |
| 187 | Value *Ptr = Order.empty() ? PointerOps[I] : PointerOps[Order[I]]; |
| 188 | std::optional<int64_t> OptPos = |
| 189 | getPointersDiff(ElemTyA: ScalarTy, PtrA: Ptr0, ElemTyB: ScalarTy, PtrB: Ptr, DL, SE); |
| 190 | if (!OptPos || OptPos > std::numeric_limits<unsigned>::max()) |
| 191 | return false; |
| 192 | unsigned Pos = static_cast<unsigned>(*OptPos); |
| 193 | CompressMask[I] = Pos; |
| 194 | if (!Stride) |
| 195 | continue; |
| 196 | if (*Stride == 0) { |
| 197 | *Stride = Pos; |
| 198 | continue; |
| 199 | } |
| 200 | if (Pos != *Stride * I) |
| 201 | Stride.reset(); |
| 202 | } |
| 203 | return Stride.has_value(); |
| 204 | } |
| 205 | |
| 206 | /// Checks if the \p VL can be transformed to a (masked)load + compress or |
| 207 | /// (masked) interleaved load. |
| 208 | bool isMaskedLoadCompress( |
| 209 | ArrayRef<Value *> VL, ArrayRef<Value *> PointerOps, |
| 210 | ArrayRef<unsigned> Order, const TargetTransformInfo &TTI, |
| 211 | const DataLayout &DL, ScalarEvolution &SE, AssumptionCache &AC, |
| 212 | const DominatorTree &DT, const TargetLibraryInfo &TLI, |
| 213 | const TargetTransformInfo::TargetCostKind CostKind, |
| 214 | const function_ref<bool(Value *)> AreAllUsersVectorized, bool ReVec, |
| 215 | bool &IsMasked, unsigned &InterleaveFactor, |
| 216 | SmallVectorImpl<int> &CompressMask, VectorType *&LoadVecTy) { |
| 217 | InterleaveFactor = 0; |
| 218 | Type *ScalarTy = VL.front()->getType(); |
| 219 | const size_t Sz = VL.size(); |
| 220 | auto *VecTy = cast<VectorType>(Val: getWidenedType(ScalarTy, VF: Sz)); |
| 221 | SmallVector<int> Mask; |
| 222 | if (!Order.empty()) |
| 223 | inversePermutation(Indices: Order, Mask); |
| 224 | // Check external uses. |
| 225 | for (const auto [I, V] : enumerate(First&: VL)) { |
| 226 | if (AreAllUsersVectorized(V)) |
| 227 | continue; |
| 228 | InstructionCost = |
| 229 | TTI.getVectorInstrCost(Opcode: Instruction::ExtractElement, Val: VecTy, CostKind, |
| 230 | Index: Mask.empty() ? I : Mask[I]); |
| 231 | InstructionCost ScalarCost = |
| 232 | TTI.getInstructionCost(U: cast<Instruction>(Val: V), CostKind); |
| 233 | if (ExtractCost <= ScalarCost) |
| 234 | return false; |
| 235 | } |
| 236 | Value *Ptr0; |
| 237 | Value *PtrN; |
| 238 | if (Order.empty()) { |
| 239 | Ptr0 = PointerOps.front(); |
| 240 | PtrN = PointerOps.back(); |
| 241 | } else { |
| 242 | Ptr0 = PointerOps[Order.front()]; |
| 243 | PtrN = PointerOps[Order.back()]; |
| 244 | } |
| 245 | std::optional<int64_t> Diff = |
| 246 | getPointersDiff(ElemTyA: ScalarTy, PtrA: Ptr0, ElemTyB: ScalarTy, PtrB: PtrN, DL, SE); |
| 247 | if (!Diff) |
| 248 | return false; |
| 249 | const size_t MaxRegSize = |
| 250 | TTI.getRegisterBitWidth(K: TargetTransformInfo::RGK_FixedWidthVector) |
| 251 | .getFixedValue(); |
| 252 | // Check for very large distances between elements. |
| 253 | if (*Diff / Sz >= MaxRegSize / 8) |
| 254 | return false; |
| 255 | LoadVecTy = cast<FixedVectorType>(Val: getWidenedType(ScalarTy, VF: *Diff + 1)); |
| 256 | auto *LI = cast<LoadInst>(Val: Order.empty() ? VL.front() : VL[Order.front()]); |
| 257 | Align CommonAlignment = LI->getAlign(); |
| 258 | SimplifyQuery SQ( |
| 259 | DL, &TLI, &DT, &AC, |
| 260 | cast<LoadInst>(Val: Order.empty() ? VL.back() : VL[Order.back()])); |
| 261 | IsMasked = !isSafeToLoadUnconditionally(V: Ptr0, Ty: LoadVecTy, Alignment: CommonAlignment, SQ); |
| 262 | if (IsMasked && !TTI.isLegalMaskedLoad(DataType: LoadVecTy, Alignment: CommonAlignment, |
| 263 | AddressSpace: LI->getPointerAddressSpace())) |
| 264 | return false; |
| 265 | // TODO: perform the analysis of each scalar load for better |
| 266 | // safe-load-unconditionally analysis. |
| 267 | bool IsStrided = |
| 268 | buildCompressMask(PointerOps, Order, ScalarTy, DL, SE, CompressMask); |
| 269 | assert(CompressMask.size() >= 2 && "At least two elements are required" ); |
| 270 | SmallVector<Value *> OrderedPointerOps(PointerOps); |
| 271 | if (!Order.empty()) |
| 272 | reorderScalars(Scalars&: OrderedPointerOps, Mask); |
| 273 | auto [ScalarGEPCost, VectorGEPCost] = |
| 274 | getGEPCosts(TTI, Ptrs: OrderedPointerOps, BasePtr: OrderedPointerOps.front(), |
| 275 | Opcode: Instruction::Load, CostKind, ScalarTy, VecTy: LoadVecTy); |
| 276 | // The cost of scalar loads. |
| 277 | InstructionCost ScalarLoadsCost = |
| 278 | accumulate(Range&: VL, Init: InstructionCost(), |
| 279 | Op: [&](InstructionCost C, Value *V) { |
| 280 | return C + TTI.getInstructionCost(U: cast<Instruction>(Val: V), |
| 281 | CostKind); |
| 282 | }) + |
| 283 | ScalarGEPCost; |
| 284 | APInt DemandedElts = APInt::getAllOnes(numBits: Sz); |
| 285 | InstructionCost GatherCost = |
| 286 | getScalarizationOverhead(TTI, ReVec, ScalarTy, Ty: VecTy, DemandedElts, |
| 287 | /*Insert=*/true, |
| 288 | /*Extract=*/false, CostKind) + |
| 289 | ScalarLoadsCost; |
| 290 | InstructionCost LoadCost = 0; |
| 291 | if (IsMasked) { |
| 292 | LoadCost = TTI.getMemIntrinsicInstrCost( |
| 293 | MICA: MemIntrinsicCostAttributes(Intrinsic::masked_load, LoadVecTy, |
| 294 | CommonAlignment, |
| 295 | LI->getPointerAddressSpace()), |
| 296 | CostKind); |
| 297 | } else { |
| 298 | LoadCost = |
| 299 | TTI.getMemoryOpCost(Opcode: Instruction::Load, Src: LoadVecTy, Alignment: CommonAlignment, |
| 300 | AddressSpace: LI->getPointerAddressSpace(), CostKind); |
| 301 | } |
| 302 | if (IsStrided && !IsMasked && Order.empty()) { |
| 303 | // Check for potential segmented(interleaved) loads. |
| 304 | VectorType *AlignedLoadVecTy = cast<VectorType>(Val: getWidenedType( |
| 305 | ScalarTy, |
| 306 | VF: getFullVectorNumberOfElements(TTI, Ty: ScalarTy, Sz: *Diff + 1, ReVec))); |
| 307 | SimplifyQuery SQ(DL, &TLI, &DT, &AC, cast<LoadInst>(Val: VL.back())); |
| 308 | if (!isSafeToLoadUnconditionally(V: Ptr0, Ty: AlignedLoadVecTy, Alignment: CommonAlignment, |
| 309 | SQ)) |
| 310 | AlignedLoadVecTy = LoadVecTy; |
| 311 | if (TTI.isLegalInterleavedAccessType(VTy: AlignedLoadVecTy, Factor: CompressMask[1], |
| 312 | Alignment: CommonAlignment, |
| 313 | AddrSpace: LI->getPointerAddressSpace())) { |
| 314 | InstructionCost InterleavedCost = |
| 315 | VectorGEPCost + TTI.getInterleavedMemoryOpCost( |
| 316 | Opcode: Instruction::Load, VecTy: AlignedLoadVecTy, |
| 317 | Factor: CompressMask[1], Indices: {}, Alignment: CommonAlignment, |
| 318 | AddressSpace: LI->getPointerAddressSpace(), CostKind, UseMaskForCond: IsMasked); |
| 319 | if (InterleavedCost < GatherCost) { |
| 320 | InterleaveFactor = CompressMask[1]; |
| 321 | LoadVecTy = AlignedLoadVecTy; |
| 322 | return true; |
| 323 | } |
| 324 | } |
| 325 | } |
| 326 | // Estimating the compression shuffle cost below can be extremely expensive |
| 327 | // for a very wide LoadVecTy, which is split into a large number of vector |
| 328 | // registers (see processShuffleMasks). The shuffle cost is always |
| 329 | // non-negative, so if the load cost alone already reaches the gather cost the |
| 330 | // masked-load-compress cannot be profitable. Bail out before the costly |
| 331 | // shuffle cost estimation in that case. |
| 332 | if (VectorGEPCost + LoadCost >= GatherCost) |
| 333 | return false; |
| 334 | InstructionCost CompressCost = getShuffleCost( |
| 335 | TTI, Kind: TTI::SK_PermuteSingleSrc, Tp: LoadVecTy, CostKind, Mask: CompressMask); |
| 336 | if (!Order.empty()) { |
| 337 | SmallVector<int> NewMask(Sz, PoisonMaskElem); |
| 338 | for (unsigned I : seq<unsigned>(Size: Sz)) { |
| 339 | NewMask[I] = CompressMask[Mask[I]]; |
| 340 | } |
| 341 | CompressMask.swap(RHS&: NewMask); |
| 342 | } |
| 343 | InstructionCost TotalVecCost = VectorGEPCost + LoadCost + CompressCost; |
| 344 | return TotalVecCost < GatherCost; |
| 345 | } |
| 346 | |
| 347 | /// Checks if the \p VL can be transformed to a (masked)load + compress or |
| 348 | /// (masked) interleaved load. |
| 349 | bool isMaskedLoadCompress( |
| 350 | ArrayRef<Value *> VL, ArrayRef<Value *> PointerOps, |
| 351 | ArrayRef<unsigned> Order, const TargetTransformInfo &TTI, |
| 352 | const DataLayout &DL, ScalarEvolution &SE, AssumptionCache &AC, |
| 353 | const DominatorTree &DT, const TargetLibraryInfo &TLI, |
| 354 | const TargetTransformInfo::TargetCostKind CostKind, |
| 355 | const function_ref<bool(Value *)> AreAllUsersVectorized, bool ReVec) { |
| 356 | bool IsMasked; |
| 357 | unsigned InterleaveFactor; |
| 358 | SmallVector<int> CompressMask; |
| 359 | VectorType *LoadVecTy; |
| 360 | return isMaskedLoadCompress(VL, PointerOps, Order, TTI, DL, SE, AC, DT, TLI, |
| 361 | CostKind, AreAllUsersVectorized, ReVec, IsMasked, |
| 362 | InterleaveFactor, CompressMask, LoadVecTy); |
| 363 | } |
| 364 | |
| 365 | /// Checks if the stores \p VL with pointers \p PointerOps can be lowered as a |
| 366 | /// single masked store. On success \p StoreVecTy is the widened store type and |
| 367 | /// \p ReuseShuffleIndices is the expand mask that places each stored value at |
| 368 | /// its element offset from the base (poison in the gaps). |
| 369 | bool isMaskedStoreCompress(ArrayRef<Value *> VL, ArrayRef<Value *> PointerOps, |
| 370 | ArrayRef<unsigned> Order, |
| 371 | const TargetTransformInfo &TTI, const DataLayout &DL, |
| 372 | ScalarEvolution &SE, Align CommonAlignment, |
| 373 | SmallVectorImpl<int> &ReuseShuffleIndices, |
| 374 | FixedVectorType *&StoreVecTy) { |
| 375 | Type *ScalarTy = cast<StoreInst>(Val: VL.front())->getValueOperand()->getType(); |
| 376 | const size_t Sz = VL.size(); |
| 377 | // Only simple scalar element types are supported. |
| 378 | if (Sz < 2 || (!ScalarTy->isIntOrPtrTy() && !ScalarTy->isFloatingPointTy())) |
| 379 | return false; |
| 380 | Value *Ptr0 = Order.empty() ? PointerOps.front() : PointerOps[Order.front()]; |
| 381 | Value *PtrN = Order.empty() ? PointerOps.back() : PointerOps[Order.back()]; |
| 382 | std::optional<int64_t> Diff = |
| 383 | getPointersDiff(ElemTyA: ScalarTy, PtrA: Ptr0, ElemTyB: ScalarTy, PtrB: PtrN, DL, SE); |
| 384 | if (!Diff || *Diff <= 0) |
| 385 | return false; |
| 386 | // Avoid widened vectors with very large gaps between the stored elements. |
| 387 | const unsigned MaxRegSize = |
| 388 | TTI.getRegisterBitWidth(K: TargetTransformInfo::RGK_FixedWidthVector) |
| 389 | .getFixedValue(); |
| 390 | const unsigned ScalarBits = DL.getTypeSizeInBits(Ty: ScalarTy).getFixedValue(); |
| 391 | if (ScalarBits == 0 || |
| 392 | static_cast<uint64_t>(*Diff) / Sz >= MaxRegSize / ScalarBits) |
| 393 | return false; |
| 394 | StoreVecTy = cast<FixedVectorType>(Val: getWidenedType(ScalarTy, VF: *Diff + 1)); |
| 395 | unsigned AS = cast<StoreInst>(Val: VL.front())->getPointerAddressSpace(); |
| 396 | if (!TTI.isLegalMaskedStore(DataType: StoreVecTy, Alignment: CommonAlignment, AddressSpace: AS, |
| 397 | MaskKind: TTI::ConstantMask)) |
| 398 | return false; |
| 399 | // Build the expand mask: store I (in address-sorted order) is placed at its |
| 400 | // element offset from the base, other widened lanes are poison. |
| 401 | ReuseShuffleIndices.assign(NumElts: *Diff + 1, Elt: PoisonMaskElem); |
| 402 | int64_t Prev = -1; |
| 403 | for (unsigned I : seq<unsigned>(Size: Sz)) { |
| 404 | Value *Ptr = Order.empty() ? PointerOps[I] : PointerOps[Order[I]]; |
| 405 | std::optional<int64_t> Off = |
| 406 | getPointersDiff(ElemTyA: ScalarTy, PtrA: Ptr0, ElemTyB: ScalarTy, PtrB: Ptr, DL, SE); |
| 407 | if (!Off || *Off <= Prev || *Off > *Diff) |
| 408 | return false; |
| 409 | ReuseShuffleIndices[*Off] = static_cast<int>(I); |
| 410 | Prev = *Off; |
| 411 | } |
| 412 | return true; |
| 413 | } |
| 414 | |
| 415 | bool clusterSortPtrAccesses(ArrayRef<Value *> VL, ArrayRef<BasicBlock *> BBs, |
| 416 | Type *ElemTy, const DataLayout &DL, |
| 417 | ScalarEvolution &SE, unsigned MaxDepth, |
| 418 | SmallVectorImpl<unsigned> &SortedIndices) { |
| 419 | assert( |
| 420 | all_of(VL, [](const Value *V) { return V->getType()->isPointerTy(); }) && |
| 421 | "Expected list of pointer operands." ); |
| 422 | // Map from bases to a vector of (Ptr, Offset, OrigIdx), which we insert each |
| 423 | // Ptr into, sort and return the sorted indices with values next to one |
| 424 | // another. |
| 425 | SmallMapVector< |
| 426 | std::pair<BasicBlock *, Value *>, |
| 427 | SmallVector<SmallVector<std::tuple<Value *, int64_t, unsigned>>>, 8> |
| 428 | Bases; |
| 429 | Bases |
| 430 | .try_emplace(Key: std::make_pair(x: BBs.front(), |
| 431 | y: getUnderlyingObject(V: VL.front(), MaxLookup: MaxDepth))) |
| 432 | .first->second.emplace_back() |
| 433 | .emplace_back(Args: VL.front(), Args: 0U, Args: 0U); |
| 434 | |
| 435 | SortedIndices.clear(); |
| 436 | for (auto [Cnt, Ptr] : enumerate(First: VL.drop_front())) { |
| 437 | auto Key = std::make_pair(x: BBs[Cnt + 1], y: getUnderlyingObject(V: Ptr, MaxLookup: MaxDepth)); |
| 438 | bool Found = any_of(Range&: Bases.try_emplace(Key).first->second, |
| 439 | P: [&, &Cnt = Cnt, &Ptr = Ptr](auto &Base) { |
| 440 | std::optional<int64_t> Diff = |
| 441 | getPointersDiff(ElemTy, std::get<0>(Base.front()), |
| 442 | ElemTy, Ptr, DL, SE, |
| 443 | /*StrictCheck=*/true); |
| 444 | if (!Diff) |
| 445 | return false; |
| 446 | |
| 447 | Base.emplace_back(Ptr, *Diff, Cnt + 1); |
| 448 | return true; |
| 449 | }); |
| 450 | |
| 451 | if (!Found) { |
| 452 | // If we haven't found enough to usefully cluster, return early. |
| 453 | if (Bases.size() > VL.size() / 2 - 1) |
| 454 | return false; |
| 455 | |
| 456 | // Not found already - add a new Base |
| 457 | Bases.find(Key)->second.emplace_back().emplace_back(Args: Ptr, Args: 0, Args: Cnt + 1); |
| 458 | } |
| 459 | } |
| 460 | |
| 461 | if (Bases.size() == VL.size()) |
| 462 | return false; |
| 463 | |
| 464 | if (Bases.size() == 1 && (Bases.front().second.size() == 1 || |
| 465 | Bases.front().second.size() == VL.size())) |
| 466 | return false; |
| 467 | |
| 468 | // For each of the bases sort the pointers by Offset and check if any of the |
| 469 | // base become consecutively allocated. |
| 470 | auto ComparePointers = [MaxDepth](Value *Ptr1, Value *Ptr2) { |
| 471 | SmallPtrSet<Value *, 13> FirstPointers; |
| 472 | SmallPtrSet<Value *, 13> SecondPointers; |
| 473 | Value *P1 = Ptr1; |
| 474 | Value *P2 = Ptr2; |
| 475 | unsigned Depth = 0; |
| 476 | while (!FirstPointers.contains(Ptr: P2) && !SecondPointers.contains(Ptr: P1)) { |
| 477 | if (P1 == P2 || Depth > MaxDepth) |
| 478 | return false; |
| 479 | FirstPointers.insert(Ptr: P1); |
| 480 | SecondPointers.insert(Ptr: P2); |
| 481 | P1 = getUnderlyingObject(V: P1, /*MaxLookup=*/1); |
| 482 | P2 = getUnderlyingObject(V: P2, /*MaxLookup=*/1); |
| 483 | ++Depth; |
| 484 | } |
| 485 | assert((FirstPointers.contains(P2) || SecondPointers.contains(P1)) && |
| 486 | "Unable to find matching root." ); |
| 487 | return FirstPointers.contains(Ptr: P2) && !SecondPointers.contains(Ptr: P1); |
| 488 | }; |
| 489 | for (auto &Base : Bases) { |
| 490 | for (auto &Vec : Base.second) { |
| 491 | if (Vec.size() > 1) { |
| 492 | stable_sort(Range&: Vec, C: llvm::less_second()); |
| 493 | int64_t InitialOffset = std::get<1>(t&: Vec[0]); |
| 494 | bool AnyConsecutive = |
| 495 | all_of(Range: enumerate(First&: Vec), P: [InitialOffset](const auto &P) { |
| 496 | return std::get<1>(P.value()) == |
| 497 | int64_t(P.index()) + InitialOffset; |
| 498 | }); |
| 499 | // Fill SortedIndices array only if it looks worth-while to sort the |
| 500 | // ptrs. |
| 501 | if (!AnyConsecutive) |
| 502 | return false; |
| 503 | } |
| 504 | } |
| 505 | stable_sort(Range&: Base.second, C: [&](const auto &V1, const auto &V2) { |
| 506 | return ComparePointers(std::get<0>(V1.front()), std::get<0>(V2.front())); |
| 507 | }); |
| 508 | } |
| 509 | |
| 510 | for (auto &T : Bases) |
| 511 | for (const auto &Vec : T.second) |
| 512 | for (const auto &P : Vec) |
| 513 | SortedIndices.push_back(Elt: std::get<2>(t: P)); |
| 514 | |
| 515 | assert(SortedIndices.size() == VL.size() && |
| 516 | "Expected SortedIndices to be the size of VL" ); |
| 517 | return true; |
| 518 | } |
| 519 | |
| 520 | } // namespace llvm::slpvectorizer |
| 521 | |