1; RUN: llc < %s -march=x86 -mcpu=pentium4 -mattr=+sse2 | FileCheck %s -check-prefix=SSE2 2; RUN: llc < %s -march=x86 -mcpu=pentium4 -mattr=+sse3 | FileCheck %s -check-prefix=SSE3 3; RUN: llc < %s -march=x86-64 -mattr=+avx | FileCheck %s -check-prefix=AVX 4 5define void @test_v4sf(<4 x float>* %P, <4 x float>* %Q, float %X) nounwind { 6 %tmp = insertelement <4 x float> zeroinitializer, float %X, i32 0 ; <<4 x float>> [#uses=1] 7 %tmp2 = insertelement <4 x float> %tmp, float %X, i32 1 ; <<4 x float>> [#uses=1] 8 %tmp4 = insertelement <4 x float> %tmp2, float %X, i32 2 ; <<4 x float>> [#uses=1] 9 %tmp6 = insertelement <4 x float> %tmp4, float %X, i32 3 ; <<4 x float>> [#uses=1] 10 %tmp8 = load <4 x float>* %Q ; <<4 x float>> [#uses=1] 11 %tmp10 = fmul <4 x float> %tmp8, %tmp6 ; <<4 x float>> [#uses=1] 12 store <4 x float> %tmp10, <4 x float>* %P 13 ret void 14 15; SSE2-LABEL: test_v4sf: 16; SSE2: pshufd $0 17 18; SSE3-LABEL: test_v4sf: 19; SSE3: pshufd $0 20} 21 22define void @test_v2sd(<2 x double>* %P, <2 x double>* %Q, double %X) nounwind { 23 %tmp = insertelement <2 x double> zeroinitializer, double %X, i32 0 ; <<2 x double>> [#uses=1] 24 %tmp2 = insertelement <2 x double> %tmp, double %X, i32 1 ; <<2 x double>> [#uses=1] 25 %tmp4 = load <2 x double>* %Q ; <<2 x double>> [#uses=1] 26 %tmp6 = fmul <2 x double> %tmp4, %tmp2 ; <<2 x double>> [#uses=1] 27 store <2 x double> %tmp6, <2 x double>* %P 28 ret void 29 30; SSE2-LABEL: test_v2sd: 31; SSE2: shufpd $0 32 33; SSE3-LABEL: test_v2sd: 34; SSE3: movddup 35} 36 37; Fold extract of a load into the load's address computation. This avoids spilling to the stack. 38define <4 x float> @load_extract_splat(<4 x float>* nocapture readonly %ptr, i64 %i, i64 %j) nounwind { 39 %1 = getelementptr inbounds <4 x float>* %ptr, i64 %i 40 %2 = load <4 x float>* %1, align 16 41 %3 = trunc i64 %j to i32 42 %4 = extractelement <4 x float> %2, i32 %3 43 %5 = insertelement <4 x float> undef, float %4, i32 0 44 %6 = insertelement <4 x float> %5, float %4, i32 1 45 %7 = insertelement <4 x float> %6, float %4, i32 2 46 %8 = insertelement <4 x float> %7, float %4, i32 3 47 ret <4 x float> %8 48 49; AVX-LABEL: load_extract_splat 50; AVX-NOT: rsp 51; AVX: vbroadcastss 52} 53 54; Fold extract of a load into the load's address computation. This avoids spilling to the stack. 55define <4 x float> @load_extract_splat1(<4 x float>* nocapture readonly %ptr, i64 %i, i64 %j) nounwind { 56 %1 = getelementptr inbounds <4 x float>* %ptr, i64 %i 57 %2 = load <4 x float>* %1, align 16 58 %3 = extractelement <4 x float> %2, i64 %j 59 %4 = insertelement <4 x float> undef, float %3, i32 0 60 %5 = insertelement <4 x float> %4, float %3, i32 1 61 %6 = insertelement <4 x float> %5, float %3, i32 2 62 %7 = insertelement <4 x float> %6, float %3, i32 3 63 ret <4 x float> %7 64 65; AVX-LABEL: load_extract_splat1 66; AVX-NOT: movs 67; AVX: vbroadcastss 68} 69