From eba6d384489be4e56718186aa7ed7e484df24613 Mon Sep 17 00:00:00 2001 From: Matt Arsenault Date: Mon, 4 Nov 2013 20:36:06 +0000 Subject: [PATCH] Scalarize select vector arguments when extracted. When the elements are extracted from a select on vectors or a vector select, do the select on the extracted scalars from the input if there is only one use. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@194013 91177308-0d34-0410-b5e6-96231b3b80d8 --- .../InstCombine/InstCombineVectorOps.cpp | 32 ++++++ .../InstCombine/select-extractelement.ll | 102 ++++++++++++++++++ .../LoopVectorize/minmax_reduction.ll | 58 +++++----- 3 files changed, 163 insertions(+), 29 deletions(-) create mode 100644 test/Transforms/InstCombine/select-extractelement.ll diff --git a/lib/Transforms/InstCombine/InstCombineVectorOps.cpp b/lib/Transforms/InstCombine/InstCombineVectorOps.cpp index 805c5d24430..1e724106991 100644 --- a/lib/Transforms/InstCombine/InstCombineVectorOps.cpp +++ b/lib/Transforms/InstCombine/InstCombineVectorOps.cpp @@ -282,6 +282,38 @@ Instruction *InstCombiner::visitExtractElementInst(ExtractElementInst &EI) { Worklist.AddValue(EE); return CastInst::Create(CI->getOpcode(), EE, EI.getType()); } + } else if (SelectInst *SI = dyn_cast(I)) { + if (SI->hasOneUse()) { + // TODO: For a select on vectors, it might be useful to do this if it + // has multiple extractelement uses. For vector select, that seems to + // fight the vectorizer. + + // If we are extracting an element from a vector select or a select on + // vectors, a select on the scalars extracted from the vector arguments. + Value *TrueVal = SI->getTrueValue(); + Value *FalseVal = SI->getFalseValue(); + + Value *Cond = SI->getCondition(); + if (Cond->getType()->isVectorTy()) { + Cond = Builder->CreateExtractElement(Cond, + EI.getIndexOperand(), + Cond->getName() + ".elt"); + } + + Value *V1Elem + = Builder->CreateExtractElement(TrueVal, + EI.getIndexOperand(), + TrueVal->getName() + ".elt"); + + Value *V2Elem + = Builder->CreateExtractElement(FalseVal, + EI.getIndexOperand(), + FalseVal->getName() + ".elt"); + return SelectInst::Create(Cond, + V1Elem, + V2Elem, + SI->getName() + ".elt"); + } } } return 0; diff --git a/test/Transforms/InstCombine/select-extractelement.ll b/test/Transforms/InstCombine/select-extractelement.ll new file mode 100644 index 00000000000..e7ea851d921 --- /dev/null +++ b/test/Transforms/InstCombine/select-extractelement.ll @@ -0,0 +1,102 @@ +; RUN: opt -S -instcombine < %s | FileCheck %s + +declare void @v4float_user(<4 x float>) #0 + + + +define float @extract_one_select(<4 x float> %a, <4 x float> %b, i32 %c) #0 { +; CHECK-LABEL: @extract_one_select( +; CHECK-NOT: select i1 {{.*}}, <4 x float> + %cmp = icmp ne i32 %c, 0 + %sel = select i1 %cmp, <4 x float> %a, <4 x float> %b + %extract = extractelement <4 x float> %sel, i32 2 + ret float %extract +} + +; Multiple extractelements +define <2 x float> @extract_two_select(<4 x float> %a, <4 x float> %b, i32 %c) #0 { +; CHECK-LABEL: @extract_two_select( +; CHECK: select i1 {{.*}}, <4 x float> + %cmp = icmp ne i32 %c, 0 + %sel = select i1 %cmp, <4 x float> %a, <4 x float> %b + %extract1 = extractelement <4 x float> %sel, i32 1 + %extract2 = extractelement <4 x float> %sel, i32 2 + %build1 = insertelement <2 x float> undef, float %extract1, i32 0 + %build2 = insertelement <2 x float> %build1, float %extract2, i32 1 + ret <2 x float> %build2 +} + +; Select has an extra non-extractelement user, don't change it +define float @extract_one_select_user(<4 x float> %a, <4 x float> %b, i32 %c) #0 { +; CHECK-LABEL: @extract_one_select_user( +; CHECK: select i1 {{.*}}, <4 x float> + %cmp = icmp ne i32 %c, 0 + %sel = select i1 %cmp, <4 x float> %a, <4 x float> %b + %extract = extractelement <4 x float> %sel, i32 2 + call void @v4float_user(<4 x float> %sel) + ret float %extract +} + +define float @extract_one_vselect_user(<4 x float> %a, <4 x float> %b, <4 x i32> %c) #0 { +; CHECK-LABEL: @extract_one_vselect_user( +; CHECK: select <4 x i1> {{.*}}, <4 x float> + %cmp = icmp ne <4 x i32> %c, zeroinitializer + %sel = select <4 x i1> %cmp, <4 x float> %a, <4 x float> %b + %extract = extractelement <4 x float> %sel, i32 2 + call void @v4float_user(<4 x float> %sel) + ret float %extract +} + +; Extract from a vector select +define float @extract_one_vselect(<4 x float> %a, <4 x float> %b, <4 x i32> %c) #0 { +; CHECK-LABEL: @extract_one_vselect( +; CHECK-NOT: select <4 x i1> + %cmp = icmp ne <4 x i32> %c, zeroinitializer + %select = select <4 x i1> %cmp, <4 x float> %a, <4 x float> %b + %extract = extractelement <4 x float> %select, i32 0 + ret float %extract +} + +; Multiple extractelements from a vector select +define <2 x float> @extract_two_vselect(<4 x float> %a, <4 x float> %b, <4 x i32> %c) #0 { +; CHECK-LABEL: @extract_two_vselect( +; CHECK-NOT: select i1 {{.*}}, <4 x float> + %cmp = icmp ne <4 x i32> %c, zeroinitializer + %sel = select <4 x i1> %cmp, <4 x float> %a, <4 x float> %b + %extract1 = extractelement <4 x float> %sel, i32 1 + %extract2 = extractelement <4 x float> %sel, i32 2 + %build1 = insertelement <2 x float> undef, float %extract1, i32 0 + %build2 = insertelement <2 x float> %build1, float %extract2, i32 1 + ret <2 x float> %build2 +} + +; All the vector selects should be decomposed into scalar selects +; Test multiple extractelements +define <4 x float> @simple_vector_select(<4 x float> %a, <4 x float> %b, <4 x i32> %c) #0 { +; CHECK-LABEL: @simple_vector_select( +; CHECK-NOT: select i1 {{.*}}, <4 x float> +entry: + %0 = extractelement <4 x i32> %c, i32 0 + %tobool = icmp ne i32 %0, 0 + %a.sink = select i1 %tobool, <4 x float> %a, <4 x float> %b + %1 = extractelement <4 x float> %a.sink, i32 0 + %2 = insertelement <4 x float> undef, float %1, i32 0 + %3 = extractelement <4 x i32> %c, i32 1 + %tobool1 = icmp ne i32 %3, 0 + %a.sink1 = select i1 %tobool1, <4 x float> %a, <4 x float> %b + %4 = extractelement <4 x float> %a.sink1, i32 1 + %5 = insertelement <4 x float> %2, float %4, i32 1 + %6 = extractelement <4 x i32> %c, i32 2 + %tobool6 = icmp ne i32 %6, 0 + %a.sink2 = select i1 %tobool6, <4 x float> %a, <4 x float> %b + %7 = extractelement <4 x float> %a.sink2, i32 2 + %8 = insertelement <4 x float> %5, float %7, i32 2 + %9 = extractelement <4 x i32> %c, i32 3 + %tobool11 = icmp ne i32 %9, 0 + %a.sink3 = select i1 %tobool11, <4 x float> %a, <4 x float> %b + %10 = extractelement <4 x float> %a.sink3, i32 3 + %11 = insertelement <4 x float> %8, float %10, i32 3 + ret <4 x float> %11 +} + +attributes #0 = { nounwind ssp uwtable "less-precise-fpmad"="false" "no-frame-pointer-elim"="true" "no-frame-pointer-elim-non-leaf"="true" "no-infs-fp-math"="false" "no-nans-fp-math"="false" "stack-protector-buffer-size"="8" "unsafe-fp-math"="false" "use-soft-float"="false" } diff --git a/test/Transforms/LoopVectorize/minmax_reduction.ll b/test/Transforms/LoopVectorize/minmax_reduction.ll index 22dd7833931..0e47260984f 100644 --- a/test/Transforms/LoopVectorize/minmax_reduction.ll +++ b/test/Transforms/LoopVectorize/minmax_reduction.ll @@ -17,7 +17,7 @@ target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f3 ; CHECK: select <2 x i1> ; CHECK: middle.block ; CHECK: icmp sgt <2 x i32> -; CHECK: select <2 x i1> +; CHECK: select i1 define i32 @max_red(i32 %max) { entry: @@ -46,7 +46,7 @@ for.end: ; CHECK: select <2 x i1> ; CHECK: middle.block ; CHECK: icmp sgt <2 x i32> -; CHECK: select <2 x i1> +; CHECK: select i1 define i32 @max_red_inverse_select(i32 %max) { entry: @@ -74,7 +74,7 @@ for.end: ; CHECK: select <2 x i1> ; CHECK: middle.block ; CHECK: icmp slt <2 x i32> -; CHECK: select <2 x i1> +; CHECK: select i1 define i32 @min_red(i32 %max) { entry: @@ -103,7 +103,7 @@ for.end: ; CHECK: select <2 x i1> ; CHECK: middle.block ; CHECK: icmp slt <2 x i32> -; CHECK: select <2 x i1> +; CHECK: select i1 define i32 @min_red_inverse_select(i32 %max) { entry: @@ -133,7 +133,7 @@ for.end: ; CHECK: select <2 x i1> ; CHECK: middle.block ; CHECK: icmp ugt <2 x i32> -; CHECK: select <2 x i1> +; CHECK: select i1 define i32 @umax_red(i32 %max) { entry: @@ -162,7 +162,7 @@ for.end: ; CHECK: select <2 x i1> ; CHECK: middle.block ; CHECK: icmp ugt <2 x i32> -; CHECK: select <2 x i1> +; CHECK: select i1 define i32 @umax_red_inverse_select(i32 %max) { entry: @@ -190,7 +190,7 @@ for.end: ; CHECK: select <2 x i1> ; CHECK: middle.block ; CHECK: icmp ult <2 x i32> -; CHECK: select <2 x i1> +; CHECK: select i1 define i32 @umin_red(i32 %max) { entry: @@ -219,7 +219,7 @@ for.end: ; CHECK: select <2 x i1> ; CHECK: middle.block ; CHECK: icmp ult <2 x i32> -; CHECK: select <2 x i1> +; CHECK: select i1 define i32 @umin_red_inverse_select(i32 %max) { entry: @@ -248,7 +248,7 @@ for.end: ; CHECK: select <2 x i1> ; CHECK: middle.block ; CHECK: icmp slt <2 x i32> -; CHECK: select <2 x i1> +; CHECK: select i1 define i32 @sge_min_red(i32 %max) { entry: @@ -277,7 +277,7 @@ for.end: ; CHECK: select <2 x i1> ; CHECK: middle.block ; CHECK: icmp sgt <2 x i32> -; CHECK: select <2 x i1> +; CHECK: select i1 define i32 @sle_min_red(i32 %max) { entry: @@ -306,7 +306,7 @@ for.end: ; CHECK: select <2 x i1> ; CHECK: middle.block ; CHECK: icmp ult <2 x i32> -; CHECK: select <2 x i1> +; CHECK: select i1 define i32 @uge_min_red(i32 %max) { entry: @@ -335,7 +335,7 @@ for.end: ; CHECK: select <2 x i1> ; CHECK: middle.block ; CHECK: icmp ugt <2 x i32> -; CHECK: select <2 x i1> +; CHECK: select i1 define i32 @ule_min_red(i32 %max) { entry: @@ -416,7 +416,7 @@ for.end: ; CHECK: select <2 x i1> ; CHECK: middle.block ; CHECK: fcmp ogt <2 x float> -; CHECK: select <2 x i1> +; CHECK: select i1 define float @max_red_float(float %max) #0 { entry: @@ -442,7 +442,7 @@ for.end: ; CHECK: select <2 x i1> ; CHECK: middle.block ; CHECK: fcmp ogt <2 x float> -; CHECK: select <2 x i1> +; CHECK: select i1 define float @max_red_float_ge(float %max) #0 { entry: @@ -468,7 +468,7 @@ for.end: ; CHECK: select <2 x i1> ; CHECK: middle.block ; CHECK: fcmp ogt <2 x float> -; CHECK: select <2 x i1> +; CHECK: select i1 define float @inverted_max_red_float(float %max) #0 { entry: @@ -494,7 +494,7 @@ for.end: ; CHECK: select <2 x i1> ; CHECK: middle.block ; CHECK: fcmp ogt <2 x float> -; CHECK: select <2 x i1> +; CHECK: select i1 define float @inverted_max_red_float_le(float %max) #0 { entry: @@ -520,7 +520,7 @@ for.end: ; CHECK: select <2 x i1> ; CHECK: middle.block ; CHECK: fcmp ogt <2 x float> -; CHECK: select <2 x i1> +; CHECK: select i1 define float @unordered_max_red_float(float %max) #0 { entry: @@ -546,7 +546,7 @@ for.end: ; CHECK: select <2 x i1> ; CHECK: middle.block ; CHECK: fcmp ogt <2 x float> -; CHECK: select <2 x i1> +; CHECK: select i1 define float @unordered_max_red_float_ge(float %max) #0 { entry: @@ -572,7 +572,7 @@ for.end: ; CHECK: select <2 x i1> ; CHECK: middle.block ; CHECK: fcmp ogt <2 x float> -; CHECK: select <2 x i1> +; CHECK: select i1 define float @inverted_unordered_max_red_float(float %max) #0 { entry: @@ -598,7 +598,7 @@ for.end: ; CHECK: select <2 x i1> ; CHECK: middle.block ; CHECK: fcmp ogt <2 x float> -; CHECK: select <2 x i1> +; CHECK: select i1 define float @inverted_unordered_max_red_float_le(float %max) #0 { entry: @@ -627,7 +627,7 @@ for.end: ; CHECK: select <2 x i1> ; CHECK: middle.block ; CHECK: fcmp olt <2 x float> -; CHECK: select <2 x i1> +; CHECK: select i1 define float @min_red_float(float %min) #0 { entry: @@ -653,7 +653,7 @@ for.end: ; CHECK: select <2 x i1> ; CHECK: middle.block ; CHECK: fcmp olt <2 x float> -; CHECK: select <2 x i1> +; CHECK: select i1 define float @min_red_float_le(float %min) #0 { entry: @@ -679,7 +679,7 @@ for.end: ; CHECK: select <2 x i1> ; CHECK: middle.block ; CHECK: fcmp olt <2 x float> -; CHECK: select <2 x i1> +; CHECK: select i1 define float @inverted_min_red_float(float %min) #0 { entry: @@ -705,7 +705,7 @@ for.end: ; CHECK: select <2 x i1> ; CHECK: middle.block ; CHECK: fcmp olt <2 x float> -; CHECK: select <2 x i1> +; CHECK: select i1 define float @inverted_min_red_float_ge(float %min) #0 { entry: @@ -731,7 +731,7 @@ for.end: ; CHECK: select <2 x i1> ; CHECK: middle.block ; CHECK: fcmp olt <2 x float> -; CHECK: select <2 x i1> +; CHECK: select i1 define float @unordered_min_red_float(float %min) #0 { entry: @@ -757,7 +757,7 @@ for.end: ; CHECK: select <2 x i1> ; CHECK: middle.block ; CHECK: fcmp olt <2 x float> -; CHECK: select <2 x i1> +; CHECK: select i1 define float @unordered_min_red_float_le(float %min) #0 { entry: @@ -783,7 +783,7 @@ for.end: ; CHECK: select <2 x i1> ; CHECK: middle.block ; CHECK: fcmp olt <2 x float> -; CHECK: select <2 x i1> +; CHECK: select i1 define float @inverted_unordered_min_red_float(float %min) #0 { entry: @@ -809,7 +809,7 @@ for.end: ; CHECK: select <2 x i1> ; CHECK: middle.block ; CHECK: fcmp olt <2 x float> -; CHECK: select <2 x i1> +; CHECK: select i1 define float @inverted_unordered_min_red_float_ge(float %min) #0 { entry: @@ -836,7 +836,7 @@ for.end: ; CHECK: select <2 x i1> ; CHECK: middle.block ; CHECK: fcmp olt <2 x double> -; CHECK: select <2 x i1> +; CHECK: select i1 define double @min_red_double(double %min) #0 { entry: -- 2.34.1