173 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py2; RUN: opt < %s -passes=slp-vectorizer,dce -slp-vectorize-non-power-of-2 -S -mtriple=x86_64-apple-macosx10.8.0 -mcpu=corei7-avx | FileCheck --check-prefixes=CHECK,NON-POW2 %s3; RUN: opt < %s -passes=slp-vectorizer,dce -slp-vectorize-non-power-of-2=false -S -mtriple=x86_64-apple-macosx10.8.0 -mcpu=corei7-avx | FileCheck --check-prefixes=CHECK,POW2-ONLY %s4 5;int foo(char * restrict A, ptr restrict B, float T) {6; A[0] = (T * B[10] + 4.0);7; A[1] = (T * B[11] + 5.0);8; A[2] = (T * B[12] + 6.0);9;}10 11define i32 @foo(ptr noalias nocapture %A, ptr noalias nocapture %B, float %T) {12; CHECK-LABEL: @foo(13; CHECK-NEXT: [[TMP1:%.*]] = getelementptr inbounds float, ptr [[B:%.*]], i64 1014; CHECK-NEXT: [[TMP2:%.*]] = load float, ptr [[TMP1]], align 415; CHECK-NEXT: [[TMP3:%.*]] = fmul float [[TMP2]], [[T:%.*]]16; CHECK-NEXT: [[TMP4:%.*]] = fpext float [[TMP3]] to double17; CHECK-NEXT: [[TMP5:%.*]] = fadd double [[TMP4]], 4.000000e+0018; CHECK-NEXT: [[TMP6:%.*]] = fptosi double [[TMP5]] to i819; CHECK-NEXT: store i8 [[TMP6]], ptr [[A:%.*]], align 120; CHECK-NEXT: [[TMP7:%.*]] = getelementptr inbounds float, ptr [[B]], i64 1121; CHECK-NEXT: [[TMP8:%.*]] = load float, ptr [[TMP7]], align 422; CHECK-NEXT: [[TMP9:%.*]] = fmul float [[TMP8]], [[T]]23; CHECK-NEXT: [[TMP10:%.*]] = fpext float [[TMP9]] to double24; CHECK-NEXT: [[TMP11:%.*]] = fadd double [[TMP10]], 5.000000e+0025; CHECK-NEXT: [[TMP12:%.*]] = fptosi double [[TMP11]] to i826; CHECK-NEXT: [[TMP13:%.*]] = getelementptr inbounds i8, ptr [[A]], i64 127; CHECK-NEXT: store i8 [[TMP12]], ptr [[TMP13]], align 128; CHECK-NEXT: [[TMP14:%.*]] = getelementptr inbounds float, ptr [[B]], i64 1229; CHECK-NEXT: [[TMP15:%.*]] = load float, ptr [[TMP14]], align 430; CHECK-NEXT: [[TMP16:%.*]] = fmul float [[TMP15]], [[T]]31; CHECK-NEXT: [[TMP17:%.*]] = fpext float [[TMP16]] to double32; CHECK-NEXT: [[TMP18:%.*]] = fadd double [[TMP17]], 6.000000e+0033; CHECK-NEXT: [[TMP19:%.*]] = fptosi double [[TMP18]] to i834; CHECK-NEXT: [[TMP20:%.*]] = getelementptr inbounds i8, ptr [[A]], i64 235; CHECK-NEXT: store i8 [[TMP19]], ptr [[TMP20]], align 136; CHECK-NEXT: ret i32 undef37;38 %1 = getelementptr inbounds float, ptr %B, i64 1039 %2 = load float, ptr %1, align 440 %3 = fmul float %2, %T41 %4 = fpext float %3 to double42 %5 = fadd double %4, 4.000000e+0043 %6 = fptosi double %5 to i844 store i8 %6, ptr %A, align 145 %7 = getelementptr inbounds float, ptr %B, i64 1146 %8 = load float, ptr %7, align 447 %9 = fmul float %8, %T48 %10 = fpext float %9 to double49 %11 = fadd double %10, 5.000000e+0050 %12 = fptosi double %11 to i851 %13 = getelementptr inbounds i8, ptr %A, i64 152 store i8 %12, ptr %13, align 153 %14 = getelementptr inbounds float, ptr %B, i64 1254 %15 = load float, ptr %14, align 455 %16 = fmul float %15, %T56 %17 = fpext float %16 to double57 %18 = fadd double %17, 6.000000e+0058 %19 = fptosi double %18 to i859 %20 = getelementptr inbounds i8, ptr %A, i64 260 store i8 %19, ptr %20, align 161 ret i32 undef62}63 64; PR4189265define void @test_v4f32_v2f32_store(<4 x float> %f, ptr %p){66; CHECK-LABEL: @test_v4f32_v2f32_store(67; CHECK-NEXT: [[TMP1:%.*]] = shufflevector <4 x float> [[F:%.*]], <4 x float> poison, <2 x i32> <i32 0, i32 1>68; CHECK-NEXT: store <2 x float> [[TMP1]], ptr [[P:%.*]], align 469; CHECK-NEXT: ret void70;71 %x0 = extractelement <4 x float> %f, i64 072 %x1 = extractelement <4 x float> %f, i64 173 %p1 = getelementptr inbounds float, ptr %p, i64 174 store float %x0, ptr %p, align 475 store float %x1, ptr %p1, align 476 ret void77}78 79define void @test_v4f32_v2f32_splat_store(<4 x float> %f, ptr %p){80; CHECK-LABEL: @test_v4f32_v2f32_splat_store(81; CHECK-NEXT: [[X0:%.*]] = extractelement <4 x float> [[F:%.*]], i64 082; CHECK-NEXT: [[P1:%.*]] = getelementptr inbounds float, ptr [[P:%.*]], i64 183; CHECK-NEXT: store float [[X0]], ptr [[P]], align 484; CHECK-NEXT: store float [[X0]], ptr [[P1]], align 485; CHECK-NEXT: ret void86;87 %x0 = extractelement <4 x float> %f, i64 088 %p1 = getelementptr inbounds float, ptr %p, i64 189 store float %x0, ptr %p, align 490 store float %x0, ptr %p1, align 491 ret void92}93 94define void @test_v4f32_v3f32_store(<4 x float> %f, ptr %p){95; NON-POW2-LABEL: @test_v4f32_v3f32_store(96; NON-POW2-NEXT: [[TMP1:%.*]] = shufflevector <4 x float> [[F:%.*]], <4 x float> poison, <3 x i32> <i32 0, i32 1, i32 2>97; NON-POW2-NEXT: store <3 x float> [[TMP1]], ptr [[P:%.*]], align 498; NON-POW2-NEXT: ret void99;100; POW2-ONLY-LABEL: @test_v4f32_v3f32_store(101; POW2-ONLY-NEXT: [[X2:%.*]] = extractelement <4 x float> [[F:%.*]], i64 2102; POW2-ONLY-NEXT: [[P2:%.*]] = getelementptr inbounds float, ptr [[P:%.*]], i64 2103; POW2-ONLY-NEXT: [[TMP1:%.*]] = shufflevector <4 x float> [[F]], <4 x float> poison, <2 x i32> <i32 0, i32 1>104; POW2-ONLY-NEXT: store <2 x float> [[TMP1]], ptr [[P]], align 4105; POW2-ONLY-NEXT: store float [[X2]], ptr [[P2]], align 4106; POW2-ONLY-NEXT: ret void107;108 %x0 = extractelement <4 x float> %f, i64 0109 %x1 = extractelement <4 x float> %f, i64 1110 %x2 = extractelement <4 x float> %f, i64 2111 %p1 = getelementptr inbounds float, ptr %p, i64 1112 %p2 = getelementptr inbounds float, ptr %p, i64 2113 store float %x0, ptr %p, align 4114 store float %x1, ptr %p1, align 4115 store float %x2, ptr %p2, align 4116 ret void117}118 119define void @test_v4f32_v3f32_splat_store(<4 x float> %f, ptr %p){120; CHECK-LABEL: @test_v4f32_v3f32_splat_store(121; CHECK-NEXT: [[X0:%.*]] = extractelement <4 x float> [[F:%.*]], i64 0122; CHECK-NEXT: [[P1:%.*]] = getelementptr inbounds float, ptr [[P:%.*]], i64 1123; CHECK-NEXT: [[P2:%.*]] = getelementptr inbounds float, ptr [[P]], i64 2124; CHECK-NEXT: store float [[X0]], ptr [[P]], align 4125; CHECK-NEXT: store float [[X0]], ptr [[P1]], align 4126; CHECK-NEXT: store float [[X0]], ptr [[P2]], align 4127; CHECK-NEXT: ret void128;129 %x0 = extractelement <4 x float> %f, i64 0130 %p1 = getelementptr inbounds float, ptr %p, i64 1131 %p2 = getelementptr inbounds float, ptr %p, i64 2132 store float %x0, ptr %p, align 4133 store float %x0, ptr %p1, align 4134 store float %x0, ptr %p2, align 4135 ret void136}137 138define void @test_v4f32_v4f32_store(<4 x float> %f, ptr %p){139; CHECK-LABEL: @test_v4f32_v4f32_store(140; CHECK-NEXT: store <4 x float> [[F:%.*]], ptr [[P:%.*]], align 4141; CHECK-NEXT: ret void142;143 %x0 = extractelement <4 x float> %f, i64 0144 %x1 = extractelement <4 x float> %f, i64 1145 %x2 = extractelement <4 x float> %f, i64 2146 %x3 = extractelement <4 x float> %f, i64 3147 %p1 = getelementptr inbounds float, ptr %p, i64 1148 %p2 = getelementptr inbounds float, ptr %p, i64 2149 %p3 = getelementptr inbounds float, ptr %p, i64 3150 store float %x0, ptr %p, align 4151 store float %x1, ptr %p1, align 4152 store float %x2, ptr %p2, align 4153 store float %x3, ptr %p3, align 4154 ret void155}156 157define void @test_v4f32_v4f32_splat_store(<4 x float> %f, ptr %p){158; CHECK-LABEL: @test_v4f32_v4f32_splat_store(159; CHECK-NEXT: [[TMP1:%.*]] = shufflevector <4 x float> [[F:%.*]], <4 x float> poison, <4 x i32> zeroinitializer160; CHECK-NEXT: store <4 x float> [[TMP1]], ptr [[P:%.*]], align 4161; CHECK-NEXT: ret void162;163 %x0 = extractelement <4 x float> %f, i64 0164 %p1 = getelementptr inbounds float, ptr %p, i64 1165 %p2 = getelementptr inbounds float, ptr %p, i64 2166 %p3 = getelementptr inbounds float, ptr %p, i64 3167 store float %x0, ptr %p, align 4168 store float %x0, ptr %p1, align 4169 store float %x0, ptr %p2, align 4170 store float %x0, ptr %p3, align 4171 ret void172}173