| 1 | //===- VPlanUtils.cpp - VPlan-related utilities ---------------------------===// |
| 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 "VPlanUtils.h" |
| 10 | #include "LoopVectorizationPlanner.h" |
| 11 | #include "VPlanAnalysis.h" |
| 12 | #include "VPlanCFG.h" |
| 13 | #include "VPlanDominatorTree.h" |
| 14 | #include "VPlanPatternMatch.h" |
| 15 | #include "llvm/ADT/SetVector.h" |
| 16 | #include "llvm/ADT/TypeSwitch.h" |
| 17 | #include "llvm/Analysis/InstSimplifyFolder.h" |
| 18 | #include "llvm/Analysis/MemoryLocation.h" |
| 19 | #include "llvm/Analysis/ScalarEvolutionExpressions.h" |
| 20 | #include "llvm/Analysis/ScalarEvolutionPatternMatch.h" |
| 21 | #include "llvm/IR/Dominators.h" |
| 22 | #include "llvm/Transforms/Utils/ScalarEvolutionExpander.h" |
| 23 | |
| 24 | using namespace llvm; |
| 25 | using namespace llvm::VPlanPatternMatch; |
| 26 | using namespace llvm::SCEVPatternMatch; |
| 27 | |
| 28 | bool vputils::onlyFirstLaneUsed(const VPValue *Def) { |
| 29 | return all_of(Range: Def->users(), |
| 30 | P: [Def](const VPUser *U) { return U->usesFirstLaneOnly(Op: Def); }); |
| 31 | } |
| 32 | |
| 33 | bool vputils::onlyFirstPartUsed(const VPValue *Def) { |
| 34 | return all_of(Range: Def->users(), |
| 35 | P: [Def](const VPUser *U) { return U->usesFirstPartOnly(Op: Def); }); |
| 36 | } |
| 37 | |
| 38 | bool vputils::onlyScalarValuesUsed(const VPValue *Def) { |
| 39 | return all_of(Range: Def->users(), |
| 40 | P: [Def](const VPUser *U) { return U->usesScalars(Op: Def); }); |
| 41 | } |
| 42 | |
| 43 | VPValue *vputils::getOrCreateVPValueForSCEVExpr(VPlan &Plan, const SCEV *Expr) { |
| 44 | if (auto *E = dyn_cast<SCEVConstant>(Val: Expr)) |
| 45 | return Plan.getOrAddLiveIn(V: E->getValue()); |
| 46 | // Skip SCEV expansion if Expr is a SCEVUnknown wrapping a non-instruction |
| 47 | // value. Otherwise the value may be defined in a loop and using it directly |
| 48 | // will break LCSSA form. The SCEV expansion takes care of preserving LCSSA |
| 49 | // form. |
| 50 | auto *U = dyn_cast<SCEVUnknown>(Val: Expr); |
| 51 | if (U && !isa<Instruction>(Val: U->getValue())) |
| 52 | return Plan.getOrAddLiveIn(V: U->getValue()); |
| 53 | auto *Expanded = new VPExpandSCEVRecipe(Expr); |
| 54 | VPBasicBlock *EntryVPBB = Plan.getEntry(); |
| 55 | auto Iter = EntryVPBB->getFirstNonPhi(); |
| 56 | while (Iter != EntryVPBB->end() && isa<VPIRInstruction>(Val: *Iter)) |
| 57 | ++Iter; |
| 58 | EntryVPBB->insert(Recipe: Expanded, InsertPt: Iter); |
| 59 | return Expanded; |
| 60 | } |
| 61 | |
| 62 | /// Returns true if \p R propagates poison from any operand to its result. |
| 63 | static bool propagatesPoisonFromRecipeOp(const VPRecipeBase *R) { |
| 64 | return TypeSwitch<const VPRecipeBase *, bool>(R) |
| 65 | .Case<VPWidenGEPRecipe, VPWidenCastRecipe>( |
| 66 | caseFn: [](const VPRecipeBase *) { return true; }) |
| 67 | .Case(caseFn: [](const VPReplicateRecipe *Rep) { |
| 68 | // GEP and casts propagate poison from all operands. |
| 69 | unsigned Opcode = Rep->getOpcode(); |
| 70 | return Opcode == Instruction::GetElementPtr || |
| 71 | Instruction::isCast(Opcode); |
| 72 | }) |
| 73 | .Default(defaultFn: [](const VPRecipeBase *) { return false; }); |
| 74 | } |
| 75 | |
| 76 | /// Returns true if \p V being poison is guaranteed to trigger UB because it |
| 77 | /// propagates to the address of a memory recipe. |
| 78 | static bool poisonGuaranteesUB(const VPValue *V) { |
| 79 | SmallPtrSet<const VPValue *, 8> Visited; |
| 80 | SmallVector<const VPValue *, 16> Worklist; |
| 81 | |
| 82 | Worklist.push_back(Elt: V); |
| 83 | |
| 84 | while (!Worklist.empty()) { |
| 85 | const VPValue *Current = Worklist.pop_back_val(); |
| 86 | if (!Visited.insert(Ptr: Current).second) |
| 87 | continue; |
| 88 | |
| 89 | for (VPUser *U : Current->users()) { |
| 90 | // Check if Current is used as an address operand for load/store. |
| 91 | if (auto *MemR = dyn_cast<VPWidenMemoryRecipe>(Val: cast<VPRecipeBase>(Val: U))) { |
| 92 | if (MemR->getAddr() == Current) |
| 93 | return true; |
| 94 | continue; |
| 95 | } |
| 96 | if (auto *Rep = dyn_cast<VPReplicateRecipe>(Val: U)) { |
| 97 | unsigned Opcode = Rep->getOpcode(); |
| 98 | if ((Opcode == Instruction::Load && Rep->getOperand(N: 0) == Current) || |
| 99 | (Opcode == Instruction::Store && Rep->getOperand(N: 1) == Current)) |
| 100 | return true; |
| 101 | } |
| 102 | |
| 103 | // Check if poison propagates through this recipe to any of its users. |
| 104 | auto *R = cast<VPRecipeBase>(Val: U); |
| 105 | for (const VPValue *Op : R->operands()) { |
| 106 | if (Op == Current && propagatesPoisonFromRecipeOp(R)) { |
| 107 | Worklist.push_back(Elt: R->getVPSingleValue()); |
| 108 | break; |
| 109 | } |
| 110 | } |
| 111 | } |
| 112 | } |
| 113 | |
| 114 | return false; |
| 115 | } |
| 116 | |
| 117 | GEPNoWrapFlags vputils::getGEPFlagsForPtr(VPValue *Ptr) { |
| 118 | // Like IR stripPointerCasts, look through GEPs with all-zero indices and |
| 119 | // casts to find a root GEP VPInstruction. |
| 120 | while (auto *PtrVPI = dyn_cast<VPInstruction>(Val: Ptr)) { |
| 121 | unsigned Opcode = PtrVPI->getOpcode(); |
| 122 | if (Opcode == Instruction::GetElementPtr) { |
| 123 | if (!all_of(Range: drop_begin(RangeOrContainer: PtrVPI->operands()), P: match_fn(P: m_ZeroInt()))) |
| 124 | return PtrVPI->getGEPNoWrapFlags(); |
| 125 | Ptr = PtrVPI->getOperand(N: 0); |
| 126 | continue; |
| 127 | } |
| 128 | if (Opcode != Instruction::BitCast && Opcode != Instruction::AddrSpaceCast) |
| 129 | break; |
| 130 | Ptr = PtrVPI->getOperand(N: 0); |
| 131 | } |
| 132 | return GEPNoWrapFlags::none(); |
| 133 | } |
| 134 | |
| 135 | const SCEV *vputils::getSCEVExprForVPValue(const VPValue *V, |
| 136 | PredicatedScalarEvolution &PSE, |
| 137 | const Loop *L) { |
| 138 | ScalarEvolution &SE = *PSE.getSE(); |
| 139 | if (auto *RV = dyn_cast<VPRegionValue>(Val: V)) { |
| 140 | assert(RV == RV->getDefiningRegion()->getCanonicalIV() && |
| 141 | "RegionValue must be canonical IV" ); |
| 142 | if (!L) |
| 143 | return SE.getCouldNotCompute(); |
| 144 | return SE.getAddRecExpr(Start: SE.getZero(Ty: RV->getType()), Step: SE.getOne(Ty: RV->getType()), |
| 145 | L, Flags: SCEV::FlagAnyWrap); |
| 146 | } |
| 147 | |
| 148 | if (isa<VPIRValue, VPSymbolicValue>(Val: V)) { |
| 149 | Value *LiveIn = V->getUnderlyingValue(); |
| 150 | if (LiveIn && SE.isSCEVable(Ty: LiveIn->getType())) |
| 151 | return SE.getSCEV(V: LiveIn); |
| 152 | return SE.getCouldNotCompute(); |
| 153 | } |
| 154 | |
| 155 | // Helper to create SCEVs for binary and unary operations. |
| 156 | auto CreateSCEV = [&](ArrayRef<VPValue *> Ops, |
| 157 | function_ref<const SCEV *(ArrayRef<SCEVUse>)> CreateFn) |
| 158 | -> const SCEV * { |
| 159 | SmallVector<SCEVUse, 2> SCEVOps; |
| 160 | for (VPValue *Op : Ops) { |
| 161 | const SCEV *S = getSCEVExprForVPValue(V: Op, PSE, L); |
| 162 | if (isa<SCEVCouldNotCompute>(Val: S)) |
| 163 | return SE.getCouldNotCompute(); |
| 164 | SCEVOps.push_back(Elt: S); |
| 165 | } |
| 166 | return PSE.getPredicatedSCEV(Expr: CreateFn(SCEVOps)); |
| 167 | }; |
| 168 | |
| 169 | VPValue *LHSVal, *RHSVal; |
| 170 | if (match(V, P: m_Add(Op0: m_VPValue(V&: LHSVal), Op1: m_VPValue(V&: RHSVal)))) |
| 171 | return CreateSCEV({LHSVal, RHSVal}, [&](ArrayRef<SCEVUse> Ops) { |
| 172 | return SE.getAddExpr(LHS: Ops[0], RHS: Ops[1], Flags: SCEV::FlagAnyWrap, Depth: 0); |
| 173 | }); |
| 174 | if (match(V, P: m_Sub(Op0: m_VPValue(V&: LHSVal), Op1: m_VPValue(V&: RHSVal)))) |
| 175 | return CreateSCEV({LHSVal, RHSVal}, [&](ArrayRef<SCEVUse> Ops) { |
| 176 | return SE.getMinusSCEV(LHS: Ops[0], RHS: Ops[1], Flags: SCEV::FlagAnyWrap, Depth: 0); |
| 177 | }); |
| 178 | if (match(V, P: m_Not(Op0: m_VPValue(V&: LHSVal)))) { |
| 179 | // not X = xor X, -1 = -1 - X |
| 180 | return CreateSCEV({LHSVal}, [&](ArrayRef<SCEVUse> Ops) { |
| 181 | return SE.getMinusSCEV(LHS: SE.getMinusOne(Ty: Ops[0]->getType()), RHS: Ops[0]); |
| 182 | }); |
| 183 | } |
| 184 | if (match(V, P: m_Mul(Op0: m_VPValue(V&: LHSVal), Op1: m_VPValue(V&: RHSVal)))) |
| 185 | return CreateSCEV({LHSVal, RHSVal}, [&](ArrayRef<SCEVUse> Ops) { |
| 186 | return SE.getMulExpr(LHS: Ops[0], RHS: Ops[1], Flags: SCEV::FlagAnyWrap, Depth: 0); |
| 187 | }); |
| 188 | // Handle shl by constant: x << c is equivalent to x * (1 << c). A shift |
| 189 | // amount >= the bit width produces poison; do not rewrite it, as |
| 190 | // getPowerOfTwo requires the power to be in range. |
| 191 | uint64_t ShiftAmt; |
| 192 | if (match(V, P: m_Shl(Op0: m_VPValue(V&: LHSVal), Op1: m_ConstantInt(C&: ShiftAmt))) && |
| 193 | ShiftAmt < LHSVal->getScalarType()->getScalarSizeInBits()) |
| 194 | return CreateSCEV(LHSVal, [&](ArrayRef<SCEVUse> Ops) { |
| 195 | return SE.getMulExpr(LHS: Ops[0], |
| 196 | RHS: SE.getPowerOfTwo(Ty: Ops[0]->getType(), Power: ShiftAmt)); |
| 197 | }); |
| 198 | if (match(V, P: m_LShr(Op0: m_VPValue(V&: LHSVal), Op1: m_ConstantInt(C&: ShiftAmt)))) { |
| 199 | Type *Ty = V->getScalarType(); |
| 200 | if (ShiftAmt < SE.getTypeSizeInBits(Ty)) |
| 201 | return CreateSCEV(LHSVal, [&](ArrayRef<SCEVUse> Ops) { |
| 202 | return SE.getUDivExpr(LHS: Ops[0], RHS: SE.getPowerOfTwo(Ty, Power: ShiftAmt)); |
| 203 | }); |
| 204 | } |
| 205 | if (match(V, P: m_UDiv(Op0: m_VPValue(V&: LHSVal), Op1: m_VPValue(V&: RHSVal)))) |
| 206 | return CreateSCEV({LHSVal, RHSVal}, [&](ArrayRef<SCEVUse> Ops) { |
| 207 | return SE.getUDivExpr(LHS: Ops[0], RHS: Ops[1]); |
| 208 | }); |
| 209 | if (match(V, P: m_URem(Op0: m_VPValue(V&: LHSVal), Op1: m_VPValue(V&: RHSVal)))) |
| 210 | return CreateSCEV({LHSVal, RHSVal}, [&](ArrayRef<SCEVUse> Ops) { |
| 211 | return SE.getURemExpr(LHS: Ops[0], RHS: Ops[1]); |
| 212 | }); |
| 213 | // A SRem with non-negative operands is equivalent to an URem. |
| 214 | if (match(V, P: m_SRem(Op0: m_VPValue(V&: LHSVal), Op1: m_VPValue(V&: RHSVal)))) { |
| 215 | return CreateSCEV({LHSVal, RHSVal}, [&](ArrayRef<SCEVUse> Ops) { |
| 216 | if (!SE.isKnownNonNegative(S: Ops[0]) || !SE.isKnownNonNegative(S: Ops[1])) |
| 217 | return SE.getCouldNotCompute(); |
| 218 | return SE.getURemExpr(LHS: Ops[0], RHS: Ops[1]); |
| 219 | }); |
| 220 | } |
| 221 | // Handle AND with constant mask: x & (2^n - 1) can be represented as x % 2^n. |
| 222 | const APInt *Mask; |
| 223 | if (match(V, P: m_c_BinaryAnd(Op0: m_VPValue(V&: LHSVal), Op1: m_APInt(C&: Mask))) && |
| 224 | (*Mask + 1).isPowerOf2()) |
| 225 | return CreateSCEV({LHSVal}, [&](ArrayRef<SCEVUse> Ops) { |
| 226 | return SE.getURemExpr(LHS: Ops[0], RHS: SE.getConstant(Val: *Mask + 1)); |
| 227 | }); |
| 228 | if (match(V, P: m_Trunc(Op0: m_VPValue(V&: LHSVal)))) { |
| 229 | Type *DestTy = V->getScalarType(); |
| 230 | return CreateSCEV({LHSVal}, [&](ArrayRef<SCEVUse> Ops) { |
| 231 | return SE.getTruncateExpr(Op: Ops[0], Ty: DestTy); |
| 232 | }); |
| 233 | } |
| 234 | if (match(V, P: m_ZExt(Op0: m_VPValue(V&: LHSVal)))) { |
| 235 | Type *DestTy = V->getScalarType(); |
| 236 | return CreateSCEV({LHSVal}, [&](ArrayRef<SCEVUse> Ops) { |
| 237 | return SE.getZeroExtendExpr(Op: Ops[0], Ty: DestTy); |
| 238 | }); |
| 239 | } |
| 240 | if (match(V, P: m_SExt(Op0: m_VPValue(V&: LHSVal)))) { |
| 241 | Type *DestTy = V->getScalarType(); |
| 242 | |
| 243 | // Mirror SCEV's createSCEV handling for sext(sub nsw): push sign extension |
| 244 | // onto the operands before computing the subtraction. |
| 245 | VPValue *SubLHS, *SubRHS; |
| 246 | auto *SubR = dyn_cast<VPRecipeWithIRFlags>(Val: LHSVal); |
| 247 | if (match(V: LHSVal, P: m_Sub(Op0: m_VPValue(V&: SubLHS), Op1: m_VPValue(V&: SubRHS))) && SubR && |
| 248 | SubR->hasNoSignedWrap() && poisonGuaranteesUB(V: LHSVal)) { |
| 249 | const SCEV *V1 = getSCEVExprForVPValue(V: SubLHS, PSE, L); |
| 250 | const SCEV *V2 = getSCEVExprForVPValue(V: SubRHS, PSE, L); |
| 251 | if (!isa<SCEVCouldNotCompute>(Val: V1) && !isa<SCEVCouldNotCompute>(Val: V2)) |
| 252 | return SE.getMinusSCEV(LHS: SE.getSignExtendExpr(Op: V1, Ty: DestTy), |
| 253 | RHS: SE.getSignExtendExpr(Op: V2, Ty: DestTy), Flags: SCEV::FlagNSW); |
| 254 | } |
| 255 | |
| 256 | return CreateSCEV({LHSVal}, [&](ArrayRef<SCEVUse> Ops) { |
| 257 | return SE.getSignExtendExpr(Op: Ops[0], Ty: DestTy); |
| 258 | }); |
| 259 | } |
| 260 | if (match(V, |
| 261 | P: m_Intrinsic<Intrinsic::umax>(Ops: m_VPValue(V&: LHSVal), Ops: m_VPValue(V&: RHSVal)))) |
| 262 | return CreateSCEV({LHSVal, RHSVal}, [&](ArrayRef<SCEVUse> Ops) { |
| 263 | return SE.getUMaxExpr(LHS: Ops[0], RHS: Ops[1]); |
| 264 | }); |
| 265 | if (match(V, |
| 266 | P: m_Intrinsic<Intrinsic::smax>(Ops: m_VPValue(V&: LHSVal), Ops: m_VPValue(V&: RHSVal)))) |
| 267 | return CreateSCEV({LHSVal, RHSVal}, [&](ArrayRef<SCEVUse> Ops) { |
| 268 | return SE.getSMaxExpr(LHS: Ops[0], RHS: Ops[1]); |
| 269 | }); |
| 270 | if (match(V, |
| 271 | P: m_Intrinsic<Intrinsic::umin>(Ops: m_VPValue(V&: LHSVal), Ops: m_VPValue(V&: RHSVal)))) |
| 272 | return CreateSCEV({LHSVal, RHSVal}, [&](ArrayRef<SCEVUse> Ops) { |
| 273 | return SE.getUMinExpr(LHS: Ops[0], RHS: Ops[1]); |
| 274 | }); |
| 275 | if (match(V, |
| 276 | P: m_Intrinsic<Intrinsic::smin>(Ops: m_VPValue(V&: LHSVal), Ops: m_VPValue(V&: RHSVal)))) |
| 277 | return CreateSCEV({LHSVal, RHSVal}, [&](ArrayRef<SCEVUse> Ops) { |
| 278 | return SE.getSMinExpr(LHS: Ops[0], RHS: Ops[1]); |
| 279 | }); |
| 280 | if (match(V, P: m_Intrinsic<Intrinsic::abs>(Ops: m_VPValue(V&: LHSVal), Ops: m_VPValue()))) |
| 281 | return CreateSCEV({LHSVal}, [&](ArrayRef<SCEVUse> Ops) { |
| 282 | // is_int_min_poison is local to this intrinsic: poison on INT_MIN is |
| 283 | // not proof that the input is never INT_MIN, nor that poison reaches |
| 284 | // UB. Do not translate it to SCEV's global IsNSW flag. |
| 285 | return SE.getAbsExpr(Op: Ops[0], /*IsNSW=*/false); |
| 286 | }); |
| 287 | |
| 288 | ArrayRef<VPValue *> Ops; |
| 289 | Type *SourceElementType; |
| 290 | if (match(V, P: m_GetElementPtr(SourceElementType, Operands&: Ops))) { |
| 291 | return CreateSCEV(Ops, [&](ArrayRef<SCEVUse> Ops) { |
| 292 | return SE.getGEPExpr(BaseExpr: Ops.front(), IndexExprs: Ops.drop_front(), SrcElementTy: SourceElementType); |
| 293 | }); |
| 294 | } |
| 295 | |
| 296 | // TODO: Support constructing SCEVs for more recipes as needed. |
| 297 | const VPRecipeBase *DefR = V->getDefiningRecipe(); |
| 298 | const SCEV *Expr = |
| 299 | TypeSwitch<const VPRecipeBase *, const SCEV *>(DefR) |
| 300 | .Case(caseFn: [](const VPExpandSCEVRecipe *R) { return R->getSCEV(); }) |
| 301 | .Case(caseFn: [&SE, &PSE, L](const VPWidenIntOrFpInductionRecipe *R) { |
| 302 | const SCEV *Step = getSCEVExprForVPValue(V: R->getStepValue(), PSE, L); |
| 303 | if (!L || isa<SCEVCouldNotCompute>(Val: Step)) |
| 304 | return SE.getCouldNotCompute(); |
| 305 | const SCEV *Start = |
| 306 | getSCEVExprForVPValue(V: R->getStartValue(), PSE, L); |
| 307 | const SCEV *AddRec = |
| 308 | SE.getAddRecExpr(Start, Step, L, Flags: SCEV::FlagAnyWrap); |
| 309 | if (R->getTruncInst()) |
| 310 | return SE.getTruncateExpr(Op: AddRec, Ty: R->getScalarType()); |
| 311 | return AddRec; |
| 312 | }) |
| 313 | .Case(caseFn: [&SE, &PSE, L](const VPWidenPointerInductionRecipe *R) { |
| 314 | const SCEV *Start = |
| 315 | getSCEVExprForVPValue(V: R->getStartValue(), PSE, L); |
| 316 | if (!L || isa<SCEVCouldNotCompute>(Val: Start)) |
| 317 | return SE.getCouldNotCompute(); |
| 318 | const SCEV *Step = getSCEVExprForVPValue(V: R->getStepValue(), PSE, L); |
| 319 | if (isa<SCEVCouldNotCompute>(Val: Step)) |
| 320 | return SE.getCouldNotCompute(); |
| 321 | return SE.getAddRecExpr(Start, Step, L, Flags: SCEV::FlagAnyWrap); |
| 322 | }) |
| 323 | .Case(caseFn: [&SE, &PSE, L](const VPDerivedIVRecipe *R) { |
| 324 | const SCEV *Start = getSCEVExprForVPValue(V: R->getOperand(N: 0), PSE, L); |
| 325 | const SCEV *IV = getSCEVExprForVPValue(V: R->getOperand(N: 1), PSE, L); |
| 326 | const SCEV *Scale = getSCEVExprForVPValue(V: R->getOperand(N: 2), PSE, L); |
| 327 | if (any_of(Range: ArrayRef({Start, IV, Scale}), |
| 328 | P: IsaPred<SCEVCouldNotCompute>)) |
| 329 | return SE.getCouldNotCompute(); |
| 330 | |
| 331 | return SE.getAddExpr( |
| 332 | LHS: SE.getTruncateOrSignExtend(V: Start, Ty: IV->getType()), |
| 333 | RHS: SE.getMulExpr( |
| 334 | LHS: IV, RHS: SE.getTruncateOrSignExtend(V: Scale, Ty: IV->getType()))); |
| 335 | }) |
| 336 | .Case(caseFn: [&SE, &PSE, L](const VPScalarIVStepsRecipe *R) { |
| 337 | const SCEV *IV = getSCEVExprForVPValue(V: R->getOperand(N: 0), PSE, L); |
| 338 | const SCEV *Step = getSCEVExprForVPValue(V: R->getOperand(N: 1), PSE, L); |
| 339 | if (isa<SCEVCouldNotCompute>(Val: IV) || !isa<SCEVConstant>(Val: Step)) |
| 340 | return SE.getCouldNotCompute(); |
| 341 | return SE.getTruncateOrSignExtend(V: IV, Ty: Step->getType()); |
| 342 | }) |
| 343 | .Default( |
| 344 | defaultFn: [&SE](const VPRecipeBase *) { return SE.getCouldNotCompute(); }); |
| 345 | |
| 346 | return PSE.getPredicatedSCEV(Expr); |
| 347 | } |
| 348 | |
| 349 | bool vputils::isAddressSCEVForCost(const SCEV *Addr, ScalarEvolution &SE, |
| 350 | const Loop *L) { |
| 351 | // If address is an SCEVAddExpr, we require that all operands must be either |
| 352 | // be invariant or a (possibly sign-extend) affine AddRec. |
| 353 | if (auto *PtrAdd = dyn_cast<SCEVAddExpr>(Val: Addr)) { |
| 354 | return all_of(Range: PtrAdd->operands(), P: [&SE, L](const SCEV *Op) { |
| 355 | return SE.isLoopInvariant(S: Op, L) || |
| 356 | match(S: Op, P: m_scev_SExt(Op0: m_scev_AffineAddRec(Op0: m_SCEV(), Op1: m_SCEV()))) || |
| 357 | match(S: Op, P: m_scev_AffineAddRec(Op0: m_SCEV(), Op1: m_SCEV())); |
| 358 | }); |
| 359 | } |
| 360 | |
| 361 | // Otherwise, check if address is loop invariant or an affine add recurrence. |
| 362 | return SE.isLoopInvariant(S: Addr, L) || |
| 363 | match(S: Addr, P: m_scev_AffineAddRec(Op0: m_SCEV(), Op1: m_SCEV())); |
| 364 | } |
| 365 | |
| 366 | unsigned vputils::getOpcode(const VPValue *V) { |
| 367 | return TypeSwitch<const VPValue *, unsigned>(V) |
| 368 | .Case<VPInstruction, VPWidenRecipe, VPWidenCastRecipe, VPWidenGEPRecipe, |
| 369 | VPReplicateRecipe, VPWidenPHIRecipe>( |
| 370 | caseFn: [](auto *I) { return I->getOpcode(); }) |
| 371 | .Case<VPVectorPointerRecipe, VPPredInstPHIRecipe, VPScalarIVStepsRecipe>( |
| 372 | caseFn: [](auto *I) { |
| 373 | // For recipes that do not directly map to LLVM IR instructions, |
| 374 | // assign opcodes after the last VPInstruction opcode (which is also |
| 375 | // after the last IR Instruction opcode), based on the VPRecipeID. |
| 376 | return VPInstruction::OpsEnd + 1 + I->getVPRecipeID(); |
| 377 | }) |
| 378 | .Default(defaultFn: [](auto *) { return 0; }); |
| 379 | } |
| 380 | |
| 381 | std::optional<std::pair<bool, unsigned>> |
| 382 | vputils::getOpcodeOrIntrinsicID(const VPValue *V) { |
| 383 | if (Intrinsic::ID IID = vputils::getIntrinsicID(R: V)) |
| 384 | return std::make_pair(x: true, y&: IID); |
| 385 | if (unsigned Opcode = vputils::getOpcode(V)) |
| 386 | return std::make_pair(x: false, y&: Opcode); |
| 387 | return {}; |
| 388 | } |
| 389 | |
| 390 | /// Returns true if \p Opcode preserves uniformity, i.e., if all operands are |
| 391 | /// uniform, the result will also be uniform. |
| 392 | static bool preservesUniformity(unsigned Opcode) { |
| 393 | if (Instruction::isBinaryOp(Opcode) || Instruction::isCast(Opcode)) |
| 394 | return true; |
| 395 | switch (Opcode) { |
| 396 | case Instruction::Freeze: |
| 397 | case Instruction::GetElementPtr: |
| 398 | case Instruction::ICmp: |
| 399 | case Instruction::FCmp: |
| 400 | case Instruction::Select: |
| 401 | case VPInstruction::Not: |
| 402 | case VPInstruction::Broadcast: |
| 403 | case VPInstruction::MaskedCond: |
| 404 | case VPInstruction::PtrAdd: |
| 405 | return true; |
| 406 | default: |
| 407 | return false; |
| 408 | } |
| 409 | } |
| 410 | |
| 411 | bool vputils::isElementwise(const VPValue *V) { |
| 412 | // TODO: Handle more opcodes and recipes. |
| 413 | if (!isa<VPInstruction, VPWidenRecipe>(Val: V)) |
| 414 | return false; |
| 415 | unsigned Opcode = getOpcode(V); |
| 416 | return Instruction::isUnaryOp(Opcode) || Instruction::isBinaryOp(Opcode); |
| 417 | } |
| 418 | |
| 419 | bool vputils::isSingleScalar(const VPValue *VPV) { |
| 420 | // Live-in, symbolic and canonical-IV region values are single-scalar. |
| 421 | if (auto *RV = dyn_cast<VPRegionValue>(Val: VPV)) |
| 422 | return RV == RV->getDefiningRegion()->getCanonicalIV(); |
| 423 | if (isa<VPIRValue, VPSymbolicValue>(Val: VPV)) |
| 424 | return true; |
| 425 | |
| 426 | if (auto *Rep = dyn_cast<VPReplicateRecipe>(Val: VPV)) { |
| 427 | const VPRegionBlock *RegionOfR = Rep->getRegion(); |
| 428 | // Don't consider recipes in replicate regions as uniform yet; their first |
| 429 | // lane cannot be accessed when executing the replicate region for other |
| 430 | // lanes. |
| 431 | if (RegionOfR && RegionOfR->isReplicator()) |
| 432 | return false; |
| 433 | return Rep->isSingleScalar() || (preservesUniformity(Opcode: Rep->getOpcode()) && |
| 434 | all_of(Range: Rep->operands(), P: isSingleScalar)); |
| 435 | } |
| 436 | if (isa<VPWidenGEPRecipe, VPBlendRecipe>(Val: VPV)) |
| 437 | return all_of(Range: VPV->getDefiningRecipe()->operands(), P: isSingleScalar); |
| 438 | if (auto *WidenR = dyn_cast<VPWidenRecipe>(Val: VPV)) { |
| 439 | return preservesUniformity(Opcode: WidenR->getOpcode()) && |
| 440 | all_of(Range: WidenR->operands(), P: isSingleScalar); |
| 441 | } |
| 442 | if (auto *VPI = dyn_cast<VPInstruction>(Val: VPV)) |
| 443 | return VPI->isSingleScalar() || VPI->isVectorToScalar() || |
| 444 | (preservesUniformity(Opcode: VPI->getOpcode()) && |
| 445 | all_of(Range: VPI->operands(), P: isSingleScalar)); |
| 446 | if (auto *RR = dyn_cast<VPReductionRecipe>(Val: VPV)) |
| 447 | return !RR->isPartialReduction(); |
| 448 | if (isa<VPVectorPointerRecipe, VPVectorEndPointerRecipe, VPDerivedIVRecipe>( |
| 449 | Val: VPV)) |
| 450 | return true; |
| 451 | if (auto *Expr = dyn_cast<VPExpressionRecipe>(Val: VPV)) |
| 452 | return Expr->isVectorToScalar(); |
| 453 | |
| 454 | // VPExpandSCEVRecipes must be placed in the entry and are always uniform. |
| 455 | return isa<VPExpandSCEVRecipe>(Val: VPV); |
| 456 | } |
| 457 | |
| 458 | bool vputils::isUniformAcrossVFsAndUFs(const VPValue *V) { |
| 459 | // Live-ins, symbolic and canonical-IV region values are uniform. |
| 460 | if (auto *RV = dyn_cast<VPRegionValue>(Val: V)) |
| 461 | return RV == RV->getDefiningRegion()->getCanonicalIV(); |
| 462 | if (isa<VPIRValue, VPSymbolicValue>(Val: V)) |
| 463 | return true; |
| 464 | |
| 465 | const VPRecipeBase *R = V->getDefiningRecipe(); |
| 466 | const VPBasicBlock *VPBB = R ? R->getParent() : nullptr; |
| 467 | const VPlan *Plan = VPBB ? VPBB->getPlan() : nullptr; |
| 468 | if (VPBB) { |
| 469 | if ((VPBB == Plan->getVectorPreheader() || VPBB == Plan->getEntry())) { |
| 470 | if (match(V: V->getDefiningRecipe(), |
| 471 | P: m_VPInstruction<VPInstruction::CanonicalIVIncrementForPart>())) |
| 472 | return false; |
| 473 | return all_of(Range: R->operands(), P: isUniformAcrossVFsAndUFs); |
| 474 | } |
| 475 | } |
| 476 | |
| 477 | return TypeSwitch<const VPRecipeBase *, bool>(R) |
| 478 | .Case(caseFn: [](const VPDerivedIVRecipe *R) { return true; }) |
| 479 | .Case(caseFn: [](const VPReplicateRecipe *R) { |
| 480 | // Be conservative about side-effects, except for the |
| 481 | // known-side-effecting assumes and stores, which we know will be |
| 482 | // uniform. |
| 483 | return R->isSingleScalar() && |
| 484 | (!R->mayHaveSideEffects() || |
| 485 | isa<AssumeInst, StoreInst>(Val: R->getUnderlyingInstr())) && |
| 486 | all_of(Range: R->operands(), P: isUniformAcrossVFsAndUFs); |
| 487 | }) |
| 488 | .Case(caseFn: [](const VPWidenRecipe *R) { |
| 489 | return preservesUniformity(Opcode: R->getOpcode()) && |
| 490 | all_of(Range: R->operands(), P: isUniformAcrossVFsAndUFs); |
| 491 | }) |
| 492 | .Case(caseFn: [](const VPPhi *) { |
| 493 | // Bail out on VPPhi, as we can end up in infinite cycles. |
| 494 | return false; |
| 495 | }) |
| 496 | .Case(caseFn: [](const VPInstruction *VPI) { |
| 497 | return (VPI->isSingleScalar() || VPI->isVectorToScalar() || |
| 498 | preservesUniformity(Opcode: VPI->getOpcode())) && |
| 499 | all_of(Range: VPI->operands(), P: isUniformAcrossVFsAndUFs); |
| 500 | }) |
| 501 | .Case(caseFn: [](const VPWidenCastRecipe *R) { |
| 502 | // A cast is uniform according to its operand. |
| 503 | return isUniformAcrossVFsAndUFs(V: R->getOperand(N: 0)); |
| 504 | }) |
| 505 | .Default(defaultFn: [](const VPRecipeBase *) { // A value is considered non-uniform |
| 506 | // unless proven otherwise. |
| 507 | return false; |
| 508 | }); |
| 509 | } |
| 510 | |
| 511 | bool vputils::doesGeneratePerAllLanes(const VPRecipeBase *R) { |
| 512 | if (auto *RepR = dyn_cast<VPReplicateRecipe>(Val: R)) |
| 513 | return RepR->doesGeneratePerAllLanes(); |
| 514 | if (auto *VPI = dyn_cast<VPInstruction>(Val: R)) |
| 515 | return VPI->doesGeneratePerAllLanes(); |
| 516 | if (auto *SIVSteps = dyn_cast<VPScalarIVStepsRecipe>(Val: R)) |
| 517 | return SIVSteps->doesGeneratePerAllLanes(); |
| 518 | return false; |
| 519 | } |
| 520 | |
| 521 | VPBasicBlock *vputils::(VPlan &Plan, VPDominatorTree &VPDT) { |
| 522 | auto DepthFirst = vp_depth_first_shallow(G: Plan.getEntry()); |
| 523 | auto I = find_if(Range&: DepthFirst, P: [&VPDT](VPBlockBase *VPB) { |
| 524 | return VPBlockUtils::isHeader(VPB, VPDT); |
| 525 | }); |
| 526 | return I == DepthFirst.end() ? nullptr : cast<VPBasicBlock>(Val: *I); |
| 527 | } |
| 528 | |
| 529 | unsigned vputils::getVFScaleFactor(VPRecipeBase *R) { |
| 530 | if (!R) |
| 531 | return 1; |
| 532 | if (auto *RR = dyn_cast<VPReductionPHIRecipe>(Val: R)) |
| 533 | return RR->getVFScaleFactor(); |
| 534 | if (auto *RR = dyn_cast<VPReductionRecipe>(Val: R)) |
| 535 | return RR->getVFScaleFactor(); |
| 536 | if (auto *ER = dyn_cast<VPExpressionRecipe>(Val: R)) |
| 537 | return ER->getVFScaleFactor(); |
| 538 | assert( |
| 539 | (!isa<VPInstruction>(R) || cast<VPInstruction>(R)->getOpcode() != |
| 540 | VPInstruction::ReductionStartVector) && |
| 541 | "getting scaling factor of reduction-start-vector not implemented yet" ); |
| 542 | return 1; |
| 543 | } |
| 544 | |
| 545 | bool vputils::cannotHoistOrSinkRecipe(const VPRecipeBase &R, bool Sinking) { |
| 546 | // Assumes don't alias anything or throw; as long as they're guaranteed to |
| 547 | // execute, they're safe to hoist. They should however not be sunk, as it |
| 548 | // would destroy information. |
| 549 | if (match(V: &R, P: m_Intrinsic<Intrinsic::assume>())) |
| 550 | return Sinking; |
| 551 | if (R.mayHaveSideEffects() || R.mayReadFromMemory() || R.isPhi()) |
| 552 | return true; |
| 553 | // Allocas cannot be hoisted. |
| 554 | auto *RepR = dyn_cast<VPReplicateRecipe>(Val: &R); |
| 555 | return RepR && RepR->getOpcode() == Instruction::Alloca; |
| 556 | } |
| 557 | |
| 558 | SmallVector<VPBasicBlock *> |
| 559 | VPBlockUtils::blocksInSingleSuccessorChainBetween(VPBasicBlock *FirstBB, |
| 560 | VPBasicBlock *LastBB) { |
| 561 | assert(FirstBB->getParent() == LastBB->getParent() && |
| 562 | "FirstBB and LastBB from different regions" ); |
| 563 | #ifndef NDEBUG |
| 564 | bool InSingleSuccChain = false; |
| 565 | for (VPBlockBase *Succ = FirstBB; Succ; Succ = Succ->getSingleSuccessor()) |
| 566 | InSingleSuccChain |= (Succ == LastBB); |
| 567 | assert(InSingleSuccChain && |
| 568 | "LastBB unreachable from FirstBB in single-successor chain" ); |
| 569 | #endif |
| 570 | auto Blocks = to_vector( |
| 571 | Range: VPBlockUtils::blocksOnly<VPBasicBlock>(Range: vp_depth_first_deep(G: FirstBB))); |
| 572 | auto *LastIt = find(Range&: Blocks, Val: LastBB); |
| 573 | assert(LastIt != Blocks.end() && |
| 574 | "LastBB unreachable from FirstBB in depth-first traversal" ); |
| 575 | Blocks.erase(CS: std::next(x: LastIt), CE: Blocks.end()); |
| 576 | return Blocks; |
| 577 | } |
| 578 | |
| 579 | VPValue *vputils::findIncomingAliasMask(const VPlan &Plan) { |
| 580 | for (VPRecipeBase &R : *Plan.getVectorPreheader()) |
| 581 | if (match(V: &R, P: m_VPInstruction<VPInstruction::IncomingAliasMask>())) |
| 582 | return cast<VPInstruction>(Val: &R); |
| 583 | return nullptr; |
| 584 | } |
| 585 | |
| 586 | bool VPBlockUtils::(const VPBlockBase *VPB, |
| 587 | const VPDominatorTree &VPDT) { |
| 588 | auto *VPBB = dyn_cast<VPBasicBlock>(Val: VPB); |
| 589 | if (!VPBB) |
| 590 | return false; |
| 591 | |
| 592 | // If VPBB is in a region R, VPBB is a loop header if R is a loop region with |
| 593 | // VPBB as its entry, i.e., free of predecessors. |
| 594 | if (auto *R = VPBB->getParent()) |
| 595 | return !R->isReplicator() && !VPBB->hasPredecessors(); |
| 596 | |
| 597 | // A header dominates its second predecessor (the latch), with the other |
| 598 | // predecessor being the preheader |
| 599 | return VPB->getPredecessors().size() == 2 && |
| 600 | VPDT.dominates(A: VPB, B: VPB->getPredecessors()[1]); |
| 601 | } |
| 602 | |
| 603 | bool VPBlockUtils::isLatch(const VPBlockBase *VPB, |
| 604 | const VPDominatorTree &VPDT) { |
| 605 | // A latch has a header as its last successor, with its other successors |
| 606 | // leaving the loop. A preheader OTOH has a header as its first (and only) |
| 607 | // successor. |
| 608 | return VPB->getNumSuccessors() >= 2 && |
| 609 | VPBlockUtils::isHeader(VPB: VPB->getSuccessors().back(), VPDT); |
| 610 | } |
| 611 | |
| 612 | std::pair<VPBasicBlock *, VPBasicBlock *> |
| 613 | VPBlockUtils::getPlainCFGHeaderAndLatch(const VPlan &Plan) { |
| 614 | auto * = cast<VPBasicBlock>( |
| 615 | Val: Plan.getEntry()->getSuccessors()[1]->getSingleSuccessor()); |
| 616 | auto *Latch = cast<VPBasicBlock>(Val: Header->getPredecessors()[1]); |
| 617 | return {Header, Latch}; |
| 618 | } |
| 619 | |
| 620 | VPBasicBlock *VPBlockUtils::getPlainCFGMiddleBlock(const VPlan &Plan) { |
| 621 | return cast<VPBasicBlock>(Val: Plan.getScalarPreheader()->getPredecessors()[0]); |
| 622 | } |
| 623 | |
| 624 | std::optional<MemoryLocation> |
| 625 | vputils::getMemoryLocation(const VPRecipeBase &R) { |
| 626 | auto *M = dyn_cast<VPIRMetadata>(Val: &R); |
| 627 | if (!M) |
| 628 | return std::nullopt; |
| 629 | MemoryLocation Loc; |
| 630 | // Populate noalias metadata from VPIRMetadata. |
| 631 | if (MDNode *NoAliasMD = M->getMetadata(Kind: LLVMContext::MD_noalias)) |
| 632 | Loc.AATags.NoAlias = NoAliasMD; |
| 633 | if (MDNode *AliasScopeMD = M->getMetadata(Kind: LLVMContext::MD_alias_scope)) |
| 634 | Loc.AATags.Scope = AliasScopeMD; |
| 635 | return Loc; |
| 636 | } |
| 637 | |
| 638 | VPInstruction *vputils::findCanonicalIVIncrement(VPlan &Plan) { |
| 639 | VPRegionBlock *LoopRegion = Plan.getVectorLoopRegion(); |
| 640 | VPRegionValue *CanIV = LoopRegion->getCanonicalIV(); |
| 641 | assert(CanIV && "Expected loop region to have a canonical IV" ); |
| 642 | |
| 643 | VPSymbolicValue &VFxUF = Plan.getVFxUF(); |
| 644 | |
| 645 | // Check if \p Step matches the expected increment step, accounting for |
| 646 | // materialization of VFxUF and UF. |
| 647 | auto IsIncrementStep = [&](VPValue *Step) -> bool { |
| 648 | if (!VFxUF.isMaterialized()) |
| 649 | return Step == &VFxUF; |
| 650 | |
| 651 | VPSymbolicValue &UF = Plan.getUF(); |
| 652 | if (!UF.isMaterialized()) |
| 653 | return Step == &UF || |
| 654 | match(V: Step, P: m_c_Mul(Op0: m_Specific(VPV: &Plan.getUF()), Op1: m_VScale())); |
| 655 | |
| 656 | // Alias masking: step is number of active lanes of a dependence mask. |
| 657 | if (match(V: Step, P: m_ZExtOrTruncOrSelf( |
| 658 | Op0: m_VPInstruction<VPInstruction::NumActiveLanes>()))) |
| 659 | return true; |
| 660 | |
| 661 | unsigned ConcreteUF = Plan.getConcreteUF(); |
| 662 | // Fixed VF: step is just the concrete UF. |
| 663 | if (match(V: Step, P: m_SpecificInt(V: ConcreteUF))) |
| 664 | return true; |
| 665 | |
| 666 | // Scalable VF: step involves VScale. |
| 667 | if (ConcreteUF == 1) |
| 668 | return match(V: Step, P: m_VScale()); |
| 669 | if (match(V: Step, P: m_c_Mul(Op0: m_SpecificInt(V: ConcreteUF), Op1: m_VScale()))) |
| 670 | return true; |
| 671 | // mul(VScale, ConcreteUF) may have been simplified to |
| 672 | // shl(VScale, log2(ConcreteUF)) when ConcreteUF is a power of 2. |
| 673 | return isPowerOf2_32(Value: ConcreteUF) && |
| 674 | match(V: Step, P: m_Shl(Op0: m_VScale(), Op1: m_SpecificInt(V: Log2_32(Value: ConcreteUF)))); |
| 675 | }; |
| 676 | |
| 677 | VPInstruction *Increment = nullptr; |
| 678 | for (VPUser *U : CanIV->users()) { |
| 679 | VPValue *Step; |
| 680 | if (isa<VPInstruction>(Val: U) && |
| 681 | match(U, P: m_c_Add(Op0: m_Specific(VPV: CanIV), Op1: m_VPValue(V&: Step))) && |
| 682 | IsIncrementStep(Step)) { |
| 683 | assert(!Increment && "There must be a unique increment" ); |
| 684 | Increment = cast<VPInstruction>(Val: U); |
| 685 | } |
| 686 | } |
| 687 | |
| 688 | assert((!VFxUF.isMaterialized() || Increment) && |
| 689 | "After materializing VFxUF, an increment must exist" ); |
| 690 | assert((!Increment || |
| 691 | LoopRegion->hasCanonicalIVNUW() == Increment->hasNoUnsignedWrap()) && |
| 692 | "NUW flag in region and increment must match" ); |
| 693 | return Increment; |
| 694 | } |
| 695 | |
| 696 | /// Find the ComputeReductionResult recipe for \p PhiR, looking through selects |
| 697 | /// inserted for predicated reductions or tail folding. |
| 698 | VPInstruction *vputils::findComputeReductionResult(VPReductionPHIRecipe *PhiR) { |
| 699 | VPValue *BackedgeVal = PhiR->getBackedgeValue(); |
| 700 | if (auto *Res = |
| 701 | findUserOf<VPInstruction::ComputeReductionResult>(V: BackedgeVal)) |
| 702 | return Res; |
| 703 | |
| 704 | // Look through selects inserted for tail folding or predicated reductions. |
| 705 | VPRecipeBase *SelR = |
| 706 | findUserOf(V: BackedgeVal, P: m_Select(Op0: m_VPValue(), Op1: m_VPValue(), Op2: m_VPValue())); |
| 707 | if (!SelR) |
| 708 | return nullptr; |
| 709 | return findUserOf<VPInstruction::ComputeReductionResult>( |
| 710 | V: cast<VPSingleDefRecipe>(Val: SelR)); |
| 711 | } |
| 712 | |
| 713 | bool vputils::isUsedByLoadStoreAddress(const VPValue *V) { |
| 714 | SmallPtrSet<const VPValue *, 4> Seen; |
| 715 | SmallVector<const VPValue *> WorkList = {V}; |
| 716 | |
| 717 | while (!WorkList.empty()) { |
| 718 | const VPValue *Cur = WorkList.pop_back_val(); |
| 719 | if (!Seen.insert(Ptr: Cur).second) |
| 720 | continue; |
| 721 | |
| 722 | auto *Blend = dyn_cast<VPBlendRecipe>(Val: Cur); |
| 723 | // Skip blends that use V only through a compare by checking if any incoming |
| 724 | // value was already visited. |
| 725 | if (Blend && none_of(Range: seq<unsigned>(Begin: 0, End: Blend->getNumIncomingValues()), |
| 726 | P: [&](unsigned I) { |
| 727 | return Seen.contains(Ptr: Blend->getIncomingValue(Idx: I)); |
| 728 | })) |
| 729 | continue; |
| 730 | |
| 731 | for (VPUser *U : Cur->users()) { |
| 732 | if (auto *InterleaveR = dyn_cast<VPInterleaveBase>(Val: U)) |
| 733 | if (InterleaveR->getAddr() == Cur) |
| 734 | return true; |
| 735 | // Cur is used as the pointer of a (possibly masked) load (operand 0) or |
| 736 | // store (operand 1). |
| 737 | if (match(U, P: m_CombineOr(Ps: m_Unary<Instruction::Load>(Op0: m_Specific(VPV: Cur)), |
| 738 | Ps: m_Binary<Instruction::Store>(Op0: m_VPValue(), |
| 739 | Op1: m_Specific(VPV: Cur))))) |
| 740 | return true; |
| 741 | if (auto *MemR = dyn_cast<VPWidenMemoryRecipe>(Val: cast<VPRecipeBase>(Val: U))) { |
| 742 | if (MemR->getAddr() == Cur && MemR->isConsecutive()) |
| 743 | return true; |
| 744 | } |
| 745 | } |
| 746 | |
| 747 | // The legacy cost model only supports scalarization loads/stores with phi |
| 748 | // addresses, if the phi is directly used as load/store address. Don't |
| 749 | // traverse further for Blends. |
| 750 | if (Blend) |
| 751 | continue; |
| 752 | |
| 753 | // Only traverse further through users that also define a value (and can |
| 754 | // thus have their own users walked). Skip when Cur is only used as mask , |
| 755 | // as well as loads: a loaded value does not depend on the load's operand. |
| 756 | for (VPUser *U : Cur->users()) { |
| 757 | auto *VPI = dyn_cast<VPInstruction>(Val: U); |
| 758 | if (VPI && VPI->getMask() == Cur && |
| 759 | none_of(Range: VPI->operandsWithoutMask(), |
| 760 | P: [Cur](VPValue *Op) { return Op == Cur; })) |
| 761 | continue; |
| 762 | if (match(U, P: m_VPInstruction<Instruction::Load>())) |
| 763 | continue; |
| 764 | if (auto *SDR = dyn_cast<VPSingleDefRecipe>(Val: U)) |
| 765 | WorkList.push_back(Elt: SDR); |
| 766 | } |
| 767 | } |
| 768 | return false; |
| 769 | } |
| 770 | |
| 771 | /// Try to find a loop-invariant IR value for \p S in the plan's entry block |
| 772 | /// that can be reused. Returns the corresponding live-in VPValue, or nullptr |
| 773 | /// if no reusable IR value is found. |
| 774 | VPValue *VPSCEVExpander::tryToReuseIRValue(const SCEV *S) { |
| 775 | if (isa<SCEVConstant, SCEVUnknown>(Val: S)) |
| 776 | return nullptr; |
| 777 | VPlan &Plan = Builder.getPlan(); |
| 778 | BasicBlock *PH = cast<VPIRBasicBlock>(Val: Plan.getEntry())->getIRBasicBlock(); |
| 779 | for (Value *V : SE.getSCEVValues(S)) { |
| 780 | // Only reuse instructions in the plan's entry block, or, when a |
| 781 | // DominatorTree is available, any instruction that dominates it. |
| 782 | // Instructions in sibling branches may not dominate the entry block. |
| 783 | auto *I = dyn_cast<Instruction>(Val: V); |
| 784 | if (!I) |
| 785 | return Plan.getOrAddLiveIn(V); |
| 786 | if (!SE.DT.dominates(A: I->getParent(), B: PH)) |
| 787 | continue; |
| 788 | SmallVector<Instruction *> DropPoisonGeneratingInsts; |
| 789 | if (!SE.canReuseInstruction(S, I, DropPoisonGeneratingInsts)) |
| 790 | continue; |
| 791 | for (Instruction *DropI : DropPoisonGeneratingInsts) |
| 792 | SCEVExpander::dropPoisonGeneratingAnnotationsAndReinfer(SE, I: DropI); |
| 793 | return Plan.getOrAddLiveIn(V); |
| 794 | } |
| 795 | return nullptr; |
| 796 | } |
| 797 | |
| 798 | VPValue *VPSCEVExpander::tryToExpand(const SCEV *S) { |
| 799 | if (VPValue *V = tryToReuseIRValue(S)) |
| 800 | return V; |
| 801 | |
| 802 | switch (S->getSCEVType()) { |
| 803 | case scConstant: |
| 804 | return Builder.getPlan().getOrAddLiveIn(V: cast<SCEVConstant>(Val: S)->getValue()); |
| 805 | case scUnknown: |
| 806 | return Builder.getPlan().getOrAddLiveIn(V: cast<SCEVUnknown>(Val: S)->getValue()); |
| 807 | case scVScale: |
| 808 | return Builder.createVScale(ResultTy: S->getType(), DL); |
| 809 | case scAddExpr: |
| 810 | case scMulExpr: { |
| 811 | auto *NAry = cast<SCEVNAryExpr>(Val: S); |
| 812 | VPIRFlags::WrapFlagsTy WrapFlags(NAry->hasNoUnsignedWrap(), |
| 813 | NAry->hasNoSignedWrap()); |
| 814 | |
| 815 | // Expanded poiner SCEVAddExpr as a ptradd of the pointer base and the |
| 816 | // integer offset, matching SCEVExpander. |
| 817 | if (S->getType()->isPointerTy()) { |
| 818 | VPValue *Base = tryToExpand(S: SE.getPointerBase(V: S)); |
| 819 | if (!Base) |
| 820 | return nullptr; |
| 821 | VPValue *Offset = tryToExpand(S: SE.removePointerBase(S)); |
| 822 | if (!Offset) |
| 823 | return nullptr; |
| 824 | GEPNoWrapFlags GEPFlags = WrapFlags.HasNUW |
| 825 | ? GEPNoWrapFlags::noUnsignedWrap() |
| 826 | : GEPNoWrapFlags::none(); |
| 827 | return Builder.createNoWrapPtrAdd(Ptr: Base, Offset, GEPFlags, DL); |
| 828 | } |
| 829 | |
| 830 | unsigned Opcode = |
| 831 | S->getSCEVType() == scAddExpr ? Instruction::Add : Instruction::Mul; |
| 832 | // Iterate in reverse so that constants are emitted last. |
| 833 | SmallVector<VPValue *, 2> Ops; |
| 834 | for (const SCEVUse &Op : reverse(C: NAry->operands())) { |
| 835 | VPValue *OpV = tryToExpand(S: Op); |
| 836 | if (!OpV) |
| 837 | return nullptr; |
| 838 | Ops.push_back(Elt: OpV); |
| 839 | } |
| 840 | VPValue *Result = Ops.front(); |
| 841 | for (VPValue *Op : drop_begin(RangeOrContainer&: Ops)) |
| 842 | Result = Builder.createOverflowingOp(Opcode, Operands: {Result, Op}, WrapFlags, DL); |
| 843 | return Result; |
| 844 | } |
| 845 | case scUDivExpr: { |
| 846 | auto *UDiv = cast<SCEVUDivExpr>(Val: S); |
| 847 | VPValue *LHS = tryToExpand(S: UDiv->getLHS()); |
| 848 | if (!LHS) |
| 849 | return nullptr; |
| 850 | VPValue *RHS = tryToExpand(S: UDiv->getRHS()); |
| 851 | if (!RHS) |
| 852 | return nullptr; |
| 853 | return Builder.createNaryOp(Opcode: Instruction::UDiv, Operands: {LHS, RHS}, |
| 854 | Flags: VPIRFlags::getDefaultFlags(Opcode: Instruction::UDiv), |
| 855 | DL); |
| 856 | } |
| 857 | case scTruncate: |
| 858 | case scZeroExtend: |
| 859 | case scSignExtend: |
| 860 | case scPtrToInt: |
| 861 | case scPtrToAddr: { |
| 862 | auto *Cast = cast<SCEVCastExpr>(Val: S); |
| 863 | VPValue *Op = tryToExpand(S: Cast->getOperand()); |
| 864 | if (!Op) |
| 865 | return nullptr; |
| 866 | Instruction::CastOps Opcode; |
| 867 | switch (S->getSCEVType()) { |
| 868 | case scTruncate: |
| 869 | Opcode = Instruction::Trunc; |
| 870 | break; |
| 871 | case scZeroExtend: |
| 872 | Opcode = Instruction::ZExt; |
| 873 | break; |
| 874 | case scSignExtend: |
| 875 | Opcode = Instruction::SExt; |
| 876 | break; |
| 877 | case scPtrToInt: |
| 878 | Opcode = Instruction::PtrToInt; |
| 879 | break; |
| 880 | case scPtrToAddr: |
| 881 | Opcode = Instruction::PtrToAddr; |
| 882 | break; |
| 883 | default: |
| 884 | llvm_unreachable("Unhandled cast SCEV" ); |
| 885 | } |
| 886 | return Builder.createScalarCast(Opcode, Op, ResultTy: S->getType(), DL); |
| 887 | } |
| 888 | case scUMaxExpr: |
| 889 | case scSMaxExpr: |
| 890 | case scUMinExpr: |
| 891 | case scSMinExpr: { |
| 892 | auto *MinMax = cast<SCEVMinMaxExpr>(Val: S); |
| 893 | Intrinsic::ID IntrinsicID; |
| 894 | switch (S->getSCEVType()) { |
| 895 | case scUMaxExpr: |
| 896 | IntrinsicID = Intrinsic::umax; |
| 897 | break; |
| 898 | case scSMaxExpr: |
| 899 | IntrinsicID = Intrinsic::smax; |
| 900 | break; |
| 901 | case scUMinExpr: |
| 902 | IntrinsicID = Intrinsic::umin; |
| 903 | break; |
| 904 | case scSMinExpr: |
| 905 | IntrinsicID = Intrinsic::smin; |
| 906 | break; |
| 907 | default: |
| 908 | llvm_unreachable("Unexpected min/max SCEV type" ); |
| 909 | } |
| 910 | // Chain operands in reverse order matching SCEVExpander's expansion of |
| 911 | // min/max expressions. |
| 912 | SmallVector<VPValue *, 2> Ops; |
| 913 | for (const SCEVUse &Op : reverse(C: MinMax->operands())) { |
| 914 | VPValue *OpV = tryToExpand(S: Op); |
| 915 | if (!OpV) |
| 916 | return nullptr; |
| 917 | Ops.push_back(Elt: OpV); |
| 918 | } |
| 919 | Type *ResultTy = MinMax->getType(); |
| 920 | VPValue *Result = Ops.front(); |
| 921 | for (VPValue *Op : drop_begin(RangeOrContainer&: Ops)) |
| 922 | Result = Builder.createScalarIntrinsic(IntrinsicID, Operands: {Result, Op}, |
| 923 | ResultTy, DL); |
| 924 | return Result; |
| 925 | } |
| 926 | default: |
| 927 | return nullptr; |
| 928 | } |
| 929 | } |
| 930 | |
| 931 | bool vputils::isDeadRecipe(VPRecipeBase &R) { |
| 932 | // Do remove conditional assume instructions as their conditions may be |
| 933 | // flattened. |
| 934 | auto *RepR = dyn_cast<VPReplicateRecipe>(Val: &R); |
| 935 | bool IsConditionalAssume = RepR && RepR->isPredicated() && |
| 936 | match(V: RepR, P: m_Intrinsic<Intrinsic::assume>()); |
| 937 | if (IsConditionalAssume) |
| 938 | return true; |
| 939 | |
| 940 | if (R.mayHaveSideEffects()) |
| 941 | return false; |
| 942 | |
| 943 | // Recipe is dead if no user keeps the recipe alive. |
| 944 | return all_of(Range: R.definedValues(), P: [](VPValue *V) { return V->user_empty(); }); |
| 945 | } |
| 946 | |
| 947 | void vputils::recursivelyDeleteDeadRecipes(VPValue *V) { |
| 948 | SmallVector<VPValue *> WorkList; |
| 949 | SmallPtrSet<VPValue *, 8> Seen; |
| 950 | WorkList.push_back(Elt: V); |
| 951 | |
| 952 | while (!WorkList.empty()) { |
| 953 | VPValue *Cur = WorkList.pop_back_val(); |
| 954 | if (!Seen.insert(Ptr: Cur).second) |
| 955 | continue; |
| 956 | VPRecipeBase *R = Cur->getDefiningRecipe(); |
| 957 | if (!R) |
| 958 | continue; |
| 959 | if (!isDeadRecipe(R&: *R)) |
| 960 | continue; |
| 961 | append_range(C&: WorkList, R: R->operands()); |
| 962 | R->eraseFromParent(); |
| 963 | } |
| 964 | } |
| 965 | |
| 966 | SmallVector<VPUser *> vputils::collectUsersRecursively(VPValue *V) { |
| 967 | SetVector<VPUser *> Users(llvm::from_range, V->users()); |
| 968 | for (unsigned I = 0; I != Users.size(); ++I) { |
| 969 | VPRecipeBase *Cur = cast<VPRecipeBase>(Val: Users[I]); |
| 970 | for (VPValue *V : Cur->definedValues()) |
| 971 | Users.insert_range(R: V->users()); |
| 972 | } |
| 973 | return Users.takeVector(); |
| 974 | } |
| 975 | |
| 976 | VPIRValue *vputils::tryToFoldLiveIns(VPSingleDefRecipe &R, |
| 977 | ArrayRef<VPValue *> Operands, |
| 978 | const DataLayout &DL) { |
| 979 | auto OpcodeOrIID = getOpcodeOrIntrinsicID(V: &R); |
| 980 | if (!OpcodeOrIID) |
| 981 | return nullptr; |
| 982 | |
| 983 | SmallVector<Value *, 4> Ops; |
| 984 | for (VPValue *Op : Operands) { |
| 985 | VPValue *Candidate = Op; |
| 986 | match(V: Op, P: m_Broadcast(Op0: m_VPValue(V&: Candidate))); |
| 987 | if (!match(V: Candidate, P: m_LiveIn())) |
| 988 | return nullptr; |
| 989 | Value *V = Candidate->getUnderlyingValue(); |
| 990 | if (!V) |
| 991 | return nullptr; |
| 992 | Ops.push_back(Elt: V); |
| 993 | } |
| 994 | |
| 995 | VPlan &Plan = *R.getParent()->getPlan(); |
| 996 | auto FoldToIRValue = [&]() -> Value * { |
| 997 | InstSimplifyFolder Folder(DL); |
| 998 | if (OpcodeOrIID->first) { |
| 999 | // VPInstructions store the called intrinsic as last operand. |
| 1000 | if (isa<VPInstruction>(Val: R)) |
| 1001 | Ops.pop_back(); |
| 1002 | |
| 1003 | auto *RFlags = dyn_cast<VPRecipeWithIRFlags>(Val: &R); |
| 1004 | return Folder.FoldIntrinsic(ID: OpcodeOrIID->second, Ops, Ty: R.getScalarType(), |
| 1005 | FMF: RFlags ? RFlags->getFastMathFlagsOrNone() |
| 1006 | : FastMathFlags()); |
| 1007 | } |
| 1008 | unsigned Opcode = OpcodeOrIID->second; |
| 1009 | if (Instruction::isBinaryOp(Opcode)) |
| 1010 | return Folder.FoldBinOp(Opc: static_cast<Instruction::BinaryOps>(Opcode), |
| 1011 | LHS: Ops[0], RHS: Ops[1]); |
| 1012 | if (Instruction::isCast(Opcode)) |
| 1013 | return Folder.FoldCast(Op: static_cast<Instruction::CastOps>(Opcode), V: Ops[0], |
| 1014 | DestTy: R.getVPSingleValue()->getScalarType()); |
| 1015 | switch (Opcode) { |
| 1016 | case VPInstruction::Not: |
| 1017 | return Folder.FoldBinOp(Opc: Instruction::BinaryOps::Xor, LHS: Ops[0], |
| 1018 | RHS: Constant::getAllOnesValue(Ty: Ops[0]->getType())); |
| 1019 | case Instruction::Select: |
| 1020 | return Folder.FoldSelect(C: Ops[0], True: Ops[1], False: Ops[2]); |
| 1021 | case Instruction::ICmp: |
| 1022 | case Instruction::FCmp: |
| 1023 | return Folder.FoldCmp(P: cast<VPRecipeWithIRFlags>(Val&: R).getPredicate(), LHS: Ops[0], |
| 1024 | RHS: Ops[1]); |
| 1025 | case Instruction::GetElementPtr: { |
| 1026 | auto &RFlags = cast<VPRecipeWithIRFlags>(Val&: R); |
| 1027 | auto *GEP = cast<GetElementPtrInst>(Val: RFlags.getUnderlyingInstr()); |
| 1028 | return Folder.FoldGEP(Ty: GEP->getSourceElementType(), Ptr: Ops[0], |
| 1029 | IdxList: drop_begin(RangeOrContainer&: Ops), NW: RFlags.getGEPNoWrapFlags()); |
| 1030 | } |
| 1031 | case VPInstruction::PtrAdd: |
| 1032 | case VPInstruction::WidePtrAdd: |
| 1033 | return Folder.FoldGEP(Ty: IntegerType::getInt8Ty(C&: Plan.getContext()), Ptr: Ops[0], |
| 1034 | IdxList: Ops[1], |
| 1035 | NW: cast<VPRecipeWithIRFlags>(Val&: R).getGEPNoWrapFlags()); |
| 1036 | // An extract of a live-in is an extract of a broadcast, so return the |
| 1037 | // broadcasted element. |
| 1038 | case Instruction::ExtractElement: |
| 1039 | assert(!Ops[0]->getType()->isVectorTy() && "Live-ins should be scalar" ); |
| 1040 | return Ops[0]; |
| 1041 | } |
| 1042 | return nullptr; |
| 1043 | }; |
| 1044 | |
| 1045 | if (Value *V = FoldToIRValue()) |
| 1046 | return Plan.getOrAddLiveIn(V); |
| 1047 | return nullptr; |
| 1048 | } |
| 1049 | |
| 1050 | void vputils::detail::pullOutPermutationsImpl( |
| 1051 | VPlan &Plan, function_ref<VPValue *(VPValue *Op)> MatchPerm, |
| 1052 | function_ref<VPSingleDefRecipe *(VPSingleDefRecipe *X)> BuildPerm) { |
| 1053 | for (VPBasicBlock *VPBB : VPBlockUtils::blocksOnly<VPBasicBlock>( |
| 1054 | Range: vp_depth_first_deep(G: Plan.getEntry()))) { |
| 1055 | for (VPRecipeBase &R : make_early_inc_range(Range&: *VPBB)) { |
| 1056 | auto *Def = dyn_cast<VPSingleDefRecipe>(Val: &R); |
| 1057 | if (!Def || !isElementwise(V: Def)) |
| 1058 | continue; |
| 1059 | |
| 1060 | // At least one of the ops must be a permutation. |
| 1061 | if (none_of(Range: Def->operands(), |
| 1062 | P: [&MatchPerm](VPValue *Op) { return MatchPerm(Op); })) |
| 1063 | continue; |
| 1064 | |
| 1065 | // All operands must be a single-use permutation or a live in (splat). |
| 1066 | if (!all_of(Range: Def->operands(), P: [&MatchPerm](VPValue *Op) { |
| 1067 | return (Op->hasOneUse() && MatchPerm(Op)) || match(V: Op, P: m_LiveIn()); |
| 1068 | })) |
| 1069 | continue; |
| 1070 | |
| 1071 | // Remove the inner permutations. |
| 1072 | for (unsigned I = 0, E = Def->getNumOperands(); I != E; ++I) |
| 1073 | if (VPValue *X = MatchPerm(Def->getOperand(N: I))) |
| 1074 | Def->setOperand(I, New: X); |
| 1075 | |
| 1076 | VPSingleDefRecipe *Res = BuildPerm(Def); |
| 1077 | Res->insertAfter(InsertPos: Def); |
| 1078 | Def->replaceUsesWithIf( |
| 1079 | New: Res, ShouldReplace: [&Res](VPUser &U, unsigned _) { return &U != Res; }); |
| 1080 | } |
| 1081 | } |
| 1082 | } |
| 1083 | |