From f3f5a1e6f77a842ccb24cc81766437da5197d712 Mon Sep 17 00:00:00 2001 From: Hal Finkel Date: Sat, 14 Apr 2012 07:32:43 +0000 Subject: [PATCH] Enhance BBVectorize to more-properly handle pointer values and vectorize GEPs. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@154734 91177308-0d34-0410-b5e6-96231b3b80d8 --- include/llvm/Transforms/Vectorize.h | 6 ++ lib/Transforms/Vectorize/BBVectorize.cpp | 29 ++++++- .../BBVectorize/simple-ldstr-ptrs.ll | 81 +++++++++++++++++++ 3 files changed, 114 insertions(+), 2 deletions(-) create mode 100644 test/Transforms/BBVectorize/simple-ldstr-ptrs.ll diff --git a/include/llvm/Transforms/Vectorize.h b/include/llvm/Transforms/Vectorize.h index 776c005da6b..652916c26c2 100644 --- a/include/llvm/Transforms/Vectorize.h +++ b/include/llvm/Transforms/Vectorize.h @@ -34,6 +34,9 @@ struct VectorizeConfig { /// @brief Vectorize floating-point values. bool VectorizeFloats; + /// @brief Vectorize pointer values. + bool VectorizePointers; + /// @brief Vectorize casting (conversion) operations. bool VectorizeCasts; @@ -46,6 +49,9 @@ struct VectorizeConfig { /// @brief Vectorize select instructions. bool VectorizeSelect; + /// @brief Vectorize getelementptr instructions. + bool VectorizeGEP; + /// @brief Vectorize loads and stores. bool VectorizeMemOps; diff --git a/lib/Transforms/Vectorize/BBVectorize.cpp b/lib/Transforms/Vectorize/BBVectorize.cpp index b2f8c8f3285..ba033183758 100644 --- a/lib/Transforms/Vectorize/BBVectorize.cpp +++ b/lib/Transforms/Vectorize/BBVectorize.cpp @@ -83,6 +83,10 @@ static cl::opt NoFloats("bb-vectorize-no-floats", cl::init(false), cl::Hidden, cl::desc("Don't try to vectorize floating-point values")); +static cl::opt +NoPointers("bb-vectorize-no-pointers", cl::init(false), cl::Hidden, + cl::desc("Don't try to vectorize pointer values")); + static cl::opt NoCasts("bb-vectorize-no-casts", cl::init(false), cl::Hidden, cl::desc("Don't try to vectorize casting (conversion) operations")); @@ -99,6 +103,10 @@ static cl::opt NoSelect("bb-vectorize-no-select", cl::init(false), cl::Hidden, cl::desc("Don't try to vectorize select instructions")); +static cl::opt +NoGEP("bb-vectorize-no-gep", cl::init(false), cl::Hidden, + cl::desc("Don't try to vectorize getelementptr instructions")); + static cl::opt NoMemOps("bb-vectorize-no-mem-ops", cl::init(false), cl::Hidden, cl::desc("Don't try to vectorize loads and stores")); @@ -550,15 +558,22 @@ namespace { return false; Type *SrcTy = C->getSrcTy(); - if (!SrcTy->isSingleValueType() || SrcTy->isPointerTy()) + if (!SrcTy->isSingleValueType()) return false; Type *DestTy = C->getDestTy(); - if (!DestTy->isSingleValueType() || DestTy->isPointerTy()) + if (!DestTy->isSingleValueType()) return false; } else if (isa(I)) { if (!Config.VectorizeSelect) return false; + } else if (GetElementPtrInst *G = dyn_cast(I)) { + if (!Config.VectorizeGEP) + return false; + + // Currently, vector GEPs exist only with one index. + if (G->getNumIndices() != 1) + return false; } else if (!(I->isBinaryOp() || isa(I) || isa(I) || isa(I))) { return false; @@ -597,6 +612,14 @@ namespace { && (T1->isFPOrFPVectorTy() || T2->isFPOrFPVectorTy())) return false; + if ((!Config.VectorizePointers || TD == 0) + && ((T1->isPointerTy() || + (T1->isVectorTy() && T1->getScalarType()->isPointerTy())) || + (T2->isPointerTy() || + (T2->isVectorTy() && T2->getScalarType()->isPointerTy())) + )) + return false; + if (T1->getPrimitiveSizeInBits() > Config.VectorBits/2 || T2->getPrimitiveSizeInBits() > Config.VectorBits/2) return false; @@ -1898,10 +1921,12 @@ VectorizeConfig::VectorizeConfig() { VectorBits = ::VectorBits; VectorizeInts = !::NoInts; VectorizeFloats = !::NoFloats; + VectorizePointers = !::NoPointers; VectorizeCasts = !::NoCasts; VectorizeMath = !::NoMath; VectorizeFMA = !::NoFMA; VectorizeSelect = !::NoSelect; + VectorizeGEP = !::NoGEP; VectorizeMemOps = !::NoMemOps; AlignedOnly = ::AlignedOnly; ReqChainDepth= ::ReqChainDepth; diff --git a/test/Transforms/BBVectorize/simple-ldstr-ptrs.ll b/test/Transforms/BBVectorize/simple-ldstr-ptrs.ll new file mode 100644 index 00000000000..f992d415477 --- /dev/null +++ b/test/Transforms/BBVectorize/simple-ldstr-ptrs.ll @@ -0,0 +1,81 @@ +target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128-n8:16:32:64-S128" +; RUN: opt < %s -bb-vectorize -bb-vectorize-req-chain-depth=3 -instcombine -gvn -S | FileCheck %s +; RUN: opt < %s -bb-vectorize -bb-vectorize-req-chain-depth=3 -bb-vectorize-aligned-only -instcombine -gvn -S | FileCheck %s -check-prefix=CHECK-AO + +; Simple 3-pair chain also with loads and stores (using ptrs and gep) +define double @test1(i64* %a, i64* %b, i64* %c) nounwind uwtable readonly { +entry: + %i0 = load i64* %a, align 8 + %i1 = load i64* %b, align 8 + %mul = mul i64 %i0, %i1 + %arrayidx3 = getelementptr inbounds i64* %a, i64 1 + %i3 = load i64* %arrayidx3, align 8 + %arrayidx4 = getelementptr inbounds i64* %b, i64 1 + %i4 = load i64* %arrayidx4, align 8 + %mul5 = mul i64 %i3, %i4 + %ptr = inttoptr i64 %mul to double* + %ptr5 = inttoptr i64 %mul5 to double* + %aptr = getelementptr inbounds double* %ptr, i64 2 + %aptr5 = getelementptr inbounds double* %ptr5, i64 3 + %av = load double* %aptr, align 16 + %av5 = load double* %aptr5, align 16 + %r = fmul double %av, %av5 + store i64 %mul, i64* %c, align 8 + %arrayidx5 = getelementptr inbounds i64* %c, i64 1 + store i64 %mul5, i64* %arrayidx5, align 8 + ret double %r +; CHECK: @test1 +; CHECK: %i0.v.i0 = bitcast i64* %a to <2 x i64>* +; CHECK: %i1.v.i0 = bitcast i64* %b to <2 x i64>* +; CHECK: %i0 = load <2 x i64>* %i0.v.i0, align 8 +; CHECK: %i1 = load <2 x i64>* %i1.v.i0, align 8 +; CHECK: %mul = mul <2 x i64> %i0, %i1 +; CHECK: %ptr = inttoptr <2 x i64> %mul to <2 x double*> +; CHECK: %aptr = getelementptr inbounds <2 x double*> %ptr, <2 x i64> +; CHECK: %aptr.v.r1 = extractelement <2 x double*> %aptr, i32 0 +; CHECK: %aptr.v.r2 = extractelement <2 x double*> %aptr, i32 1 +; CHECK: %av = load double* %aptr.v.r1, align 16 +; CHECK: %av5 = load double* %aptr.v.r2, align 16 +; CHECK: %r = fmul double %av, %av5 +; CHECK: %0 = bitcast i64* %c to <2 x i64>* +; CHECK: store <2 x i64> %mul, <2 x i64>* %0, align 8 +; CHECK: ret double %r +; CHECK-AO: @test1 +; CHECK-AO-NOT: load <2 x +} + +; Simple 3-pair chain with loads and stores (using ptrs and gep) +define void @test2(i64** %a, i64** %b, i64** %c) nounwind uwtable readonly { +entry: + %i0 = load i64** %a, align 8 + %i1 = load i64** %b, align 8 + %arrayidx3 = getelementptr inbounds i64** %a, i64 1 + %i3 = load i64** %arrayidx3, align 8 + %arrayidx4 = getelementptr inbounds i64** %b, i64 1 + %i4 = load i64** %arrayidx4, align 8 + %o1 = load i64* %i1, align 8 + %o4 = load i64* %i4, align 8 + %ptr0 = getelementptr inbounds i64* %i0, i64 %o1 + %ptr3 = getelementptr inbounds i64* %i3, i64 %o4 + store i64* %ptr0, i64** %c, align 8 + %arrayidx5 = getelementptr inbounds i64** %c, i64 1 + store i64* %ptr3, i64** %arrayidx5, align 8 + ret void +; CHECK: @test2 +; CHECK: %i0.v.i0 = bitcast i64** %a to <2 x i64*>* +; CHECK: %i1 = load i64** %b, align 8 +; CHECK: %i0 = load <2 x i64*>* %i0.v.i0, align 8 +; CHECK: %arrayidx4 = getelementptr inbounds i64** %b, i64 1 +; CHECK: %i4 = load i64** %arrayidx4, align 8 +; CHECK: %o1 = load i64* %i1, align 8 +; CHECK: %o4 = load i64* %i4, align 8 +; CHECK: %ptr0.v.i1.1 = insertelement <2 x i64> undef, i64 %o1, i32 0 +; CHECK: %ptr0.v.i1.2 = insertelement <2 x i64> %ptr0.v.i1.1, i64 %o4, i32 1 +; CHECK: %ptr0 = getelementptr inbounds <2 x i64*> %i0, <2 x i64> %ptr0.v.i1.2 +; CHECK: %0 = bitcast i64** %c to <2 x i64*>* +; CHECK: store <2 x i64*> %ptr0, <2 x i64*>* %0, align 8 +; CHECK: ret void +; CHECK-AO: @test2 +; CHECK-AO-NOT: <2 x +} +