| 1 | //===- ExpandReductions.cpp - Expand reduction intrinsics -----------------===// |
| 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 implements IR expansion for reduction intrinsics, allowing targets |
| 10 | // to enable the intrinsics until just before codegen. |
| 11 | // |
| 12 | //===----------------------------------------------------------------------===// |
| 13 | |
| 14 | #include "llvm/CodeGen/ExpandReductions.h" |
| 15 | #include "llvm/Analysis/LoopInfo.h" |
| 16 | #include "llvm/Analysis/TargetTransformInfo.h" |
| 17 | #include "llvm/CodeGen/Passes.h" |
| 18 | #include "llvm/IR/Dominators.h" |
| 19 | #include "llvm/IR/IRBuilder.h" |
| 20 | #include "llvm/IR/InstIterator.h" |
| 21 | #include "llvm/IR/IntrinsicInst.h" |
| 22 | #include "llvm/IR/Intrinsics.h" |
| 23 | #include "llvm/InitializePasses.h" |
| 24 | #include "llvm/Pass.h" |
| 25 | #include "llvm/Transforms/Utils/LoopUtils.h" |
| 26 | |
| 27 | using namespace llvm; |
| 28 | |
| 29 | namespace { |
| 30 | |
| 31 | bool expandReductions(Function &F, const TargetTransformInfo *TTI, |
| 32 | DominatorTree *DT, LoopInfo *LI) { |
| 33 | bool Changed = false; |
| 34 | SmallVector<IntrinsicInst *, 4> Worklist; |
| 35 | for (auto &I : instructions(F)) { |
| 36 | if (auto *II = dyn_cast<IntrinsicInst>(Val: &I)) { |
| 37 | switch (II->getIntrinsicID()) { |
| 38 | default: break; |
| 39 | case Intrinsic::vector_reduce_fadd: |
| 40 | case Intrinsic::vector_reduce_fmul: |
| 41 | case Intrinsic::vector_reduce_add: |
| 42 | case Intrinsic::vector_reduce_mul: |
| 43 | case Intrinsic::vector_reduce_and: |
| 44 | case Intrinsic::vector_reduce_or: |
| 45 | case Intrinsic::vector_reduce_xor: |
| 46 | case Intrinsic::vector_reduce_smax: |
| 47 | case Intrinsic::vector_reduce_smin: |
| 48 | case Intrinsic::vector_reduce_umax: |
| 49 | case Intrinsic::vector_reduce_umin: |
| 50 | case Intrinsic::vector_reduce_fmax: |
| 51 | case Intrinsic::vector_reduce_fmin: |
| 52 | if (TTI->shouldExpandReduction(II)) |
| 53 | Worklist.push_back(Elt: II); |
| 54 | |
| 55 | break; |
| 56 | } |
| 57 | } |
| 58 | } |
| 59 | |
| 60 | for (auto *II : Worklist) { |
| 61 | FastMathFlags FMF = II->getFastMathFlagsOrNone(); |
| 62 | Intrinsic::ID ID = II->getIntrinsicID(); |
| 63 | RecurKind RK = getMinMaxReductionRecurKind(RdxID: ID); |
| 64 | TargetTransformInfo::ReductionShuffle RS = |
| 65 | TTI->getPreferredExpandedReductionShuffle(II); |
| 66 | |
| 67 | Value *Rdx = nullptr; |
| 68 | IRBuilder<> Builder(II); |
| 69 | IRBuilder<>::FastMathFlagGuard FMFGuard(Builder); |
| 70 | Builder.setFastMathFlags(FMF); |
| 71 | switch (ID) { |
| 72 | default: llvm_unreachable("Unexpected intrinsic!" ); |
| 73 | case Intrinsic::vector_reduce_fadd: |
| 74 | case Intrinsic::vector_reduce_fmul: { |
| 75 | // FMFs must be attached to the call, otherwise it's an ordered reduction |
| 76 | // and it can't be handled by generating a shuffle sequence. |
| 77 | Value *Acc = II->getArgOperand(i: 0); |
| 78 | Value *Vec = II->getArgOperand(i: 1); |
| 79 | unsigned RdxOpcode = getArithmeticReductionInstruction(RdxID: ID); |
| 80 | if (isa<ScalableVectorType>(Val: Vec->getType())) { |
| 81 | Rdx = expandReductionViaLoop(Builder, Vec, RdxOpcode, Acc, DT, LI); |
| 82 | break; |
| 83 | } |
| 84 | if (!FMF.allowReassoc()) |
| 85 | Rdx = getOrderedReduction(Builder, Acc, Src: Vec, Op: RdxOpcode, MinMaxKind: RK); |
| 86 | else { |
| 87 | if (!isPowerOf2_32( |
| 88 | Value: cast<FixedVectorType>(Val: Vec->getType())->getNumElements())) |
| 89 | continue; |
| 90 | Rdx = getShuffleReduction(Builder, Src: Vec, Op: RdxOpcode, RS, MinMaxKind: RK); |
| 91 | Rdx = Builder.CreateBinOp(Opc: (Instruction::BinaryOps)RdxOpcode, LHS: Acc, RHS: Rdx, |
| 92 | Name: "bin.rdx" ); |
| 93 | } |
| 94 | break; |
| 95 | } |
| 96 | case Intrinsic::vector_reduce_and: |
| 97 | case Intrinsic::vector_reduce_or: { |
| 98 | // Canonicalize logical or/and reductions: |
| 99 | // Or reduction for i1 is represented as: |
| 100 | // %val = bitcast <ReduxWidth x i1> to iReduxWidth |
| 101 | // %res = cmp ne iReduxWidth %val, 0 |
| 102 | // And reduction for i1 is represented as: |
| 103 | // %val = bitcast <ReduxWidth x i1> to iReduxWidth |
| 104 | // %res = cmp eq iReduxWidth %val, 11111 |
| 105 | Value *Vec = II->getArgOperand(i: 0); |
| 106 | auto *FTy = cast<FixedVectorType>(Val: Vec->getType()); |
| 107 | unsigned NumElts = FTy->getNumElements(); |
| 108 | if (!isPowerOf2_32(Value: NumElts)) |
| 109 | continue; |
| 110 | |
| 111 | if (FTy->getElementType() == Builder.getInt1Ty()) { |
| 112 | Rdx = Builder.CreateBitCast(V: Vec, DestTy: Builder.getIntNTy(N: NumElts)); |
| 113 | if (ID == Intrinsic::vector_reduce_and) { |
| 114 | Rdx = Builder.CreateICmpEQ( |
| 115 | LHS: Rdx, RHS: ConstantInt::getAllOnesValue(Ty: Rdx->getType())); |
| 116 | } else { |
| 117 | assert(ID == Intrinsic::vector_reduce_or && "Expected or reduction." ); |
| 118 | Rdx = Builder.CreateIsNotNull(Arg: Rdx); |
| 119 | } |
| 120 | break; |
| 121 | } |
| 122 | unsigned RdxOpcode = getArithmeticReductionInstruction(RdxID: ID); |
| 123 | Rdx = getShuffleReduction(Builder, Src: Vec, Op: RdxOpcode, RS, MinMaxKind: RK); |
| 124 | break; |
| 125 | } |
| 126 | case Intrinsic::vector_reduce_add: |
| 127 | case Intrinsic::vector_reduce_mul: |
| 128 | case Intrinsic::vector_reduce_xor: |
| 129 | case Intrinsic::vector_reduce_smax: |
| 130 | case Intrinsic::vector_reduce_smin: |
| 131 | case Intrinsic::vector_reduce_umax: |
| 132 | case Intrinsic::vector_reduce_umin: { |
| 133 | Value *Vec = II->getArgOperand(i: 0); |
| 134 | unsigned RdxOpcode = getArithmeticReductionInstruction(RdxID: ID); |
| 135 | if (isa<ScalableVectorType>(Val: Vec->getType())) { |
| 136 | Type *EltTy = Vec->getType()->getScalarType(); |
| 137 | Value *Ident = getReductionIdentity(RdxID: ID, Ty: EltTy, FMF); |
| 138 | Rdx = expandReductionViaLoop(Builder, Vec, RdxOpcode, Acc: Ident, DT, LI); |
| 139 | break; |
| 140 | } |
| 141 | if (!isPowerOf2_32( |
| 142 | Value: cast<FixedVectorType>(Val: Vec->getType())->getNumElements())) |
| 143 | continue; |
| 144 | Rdx = getShuffleReduction(Builder, Src: Vec, Op: RdxOpcode, RS, MinMaxKind: RK); |
| 145 | break; |
| 146 | } |
| 147 | case Intrinsic::vector_reduce_fmax: |
| 148 | case Intrinsic::vector_reduce_fmin: { |
| 149 | // We require "nnan" to use a shuffle reduction; "nsz" is implied by the |
| 150 | // semantics of the reduction. |
| 151 | Value *Vec = II->getArgOperand(i: 0); |
| 152 | if (!isPowerOf2_32( |
| 153 | Value: cast<FixedVectorType>(Val: Vec->getType())->getNumElements()) || |
| 154 | !FMF.noNaNs()) |
| 155 | continue; |
| 156 | unsigned RdxOpcode = getArithmeticReductionInstruction(RdxID: ID); |
| 157 | Rdx = getShuffleReduction(Builder, Src: Vec, Op: RdxOpcode, RS, MinMaxKind: RK); |
| 158 | break; |
| 159 | } |
| 160 | } |
| 161 | II->replaceAllUsesWith(V: Rdx); |
| 162 | II->eraseFromParent(); |
| 163 | Changed = true; |
| 164 | } |
| 165 | return Changed; |
| 166 | } |
| 167 | |
| 168 | class ExpandReductions : public FunctionPass { |
| 169 | public: |
| 170 | static char ID; |
| 171 | ExpandReductions() : FunctionPass(ID) {} |
| 172 | |
| 173 | bool runOnFunction(Function &F) override { |
| 174 | const auto *TTI =&getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F); |
| 175 | auto *DTWP = getAnalysisIfAvailable<DominatorTreeWrapperPass>(); |
| 176 | auto *LIWP = getAnalysisIfAvailable<LoopInfoWrapperPass>(); |
| 177 | auto *DT = DTWP ? &DTWP->getDomTree() : nullptr; |
| 178 | auto *LI = LIWP ? &LIWP->getLoopInfo() : nullptr; |
| 179 | return expandReductions(F, TTI, DT, LI); |
| 180 | } |
| 181 | |
| 182 | void getAnalysisUsage(AnalysisUsage &AU) const override { |
| 183 | AU.addRequired<TargetTransformInfoWrapperPass>(); |
| 184 | AU.addPreserved<DominatorTreeWrapperPass>(); |
| 185 | AU.addPreserved<LoopInfoWrapperPass>(); |
| 186 | } |
| 187 | }; |
| 188 | } |
| 189 | |
| 190 | char ExpandReductions::ID; |
| 191 | INITIALIZE_PASS_BEGIN(ExpandReductions, "expand-reductions" , |
| 192 | "Expand reduction intrinsics" , false, false) |
| 193 | INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass) |
| 194 | INITIALIZE_PASS_END(ExpandReductions, "expand-reductions" , |
| 195 | "Expand reduction intrinsics" , false, false) |
| 196 | |
| 197 | FunctionPass *llvm::createExpandReductionsPass() { |
| 198 | return new ExpandReductions(); |
| 199 | } |
| 200 | |
| 201 | PreservedAnalyses ExpandReductionsPass::run(Function &F, |
| 202 | FunctionAnalysisManager &AM) { |
| 203 | const auto &TTI = AM.getResult<TargetIRAnalysis>(IR&: F); |
| 204 | auto *DT = AM.getCachedResult<DominatorTreeAnalysis>(IR&: F); |
| 205 | auto *LI = AM.getCachedResult<LoopAnalysis>(IR&: F); |
| 206 | if (!expandReductions(F, TTI: &TTI, DT, LI)) |
| 207 | return PreservedAnalyses::all(); |
| 208 | PreservedAnalyses PA; |
| 209 | PA.preserve<DominatorTreeAnalysis>(); |
| 210 | PA.preserve<LoopAnalysis>(); |
| 211 | return PA; |
| 212 | } |
| 213 | |