160 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py2; RUN: llc < %s -mtriple=x86_64-- -mattr=sse2 | FileCheck %s --check-prefix=SSE3; RUN: llc < %s -mtriple=x86_64-- -mattr=avx | FileCheck %s --check-prefix=AVX4 5define <2 x double> @splat_fdiv_v2f64(<2 x double> %x, double %y) {6; SSE-LABEL: splat_fdiv_v2f64:7; SSE: # %bb.0:8; SSE-NEXT: movsd {{.*#+}} xmm2 = [1.0E+0,0.0E+0]9; SSE-NEXT: divsd %xmm1, %xmm210; SSE-NEXT: unpcklpd {{.*#+}} xmm2 = xmm2[0,0]11; SSE-NEXT: mulpd %xmm2, %xmm012; SSE-NEXT: retq13;14; AVX-LABEL: splat_fdiv_v2f64:15; AVX: # %bb.0:16; AVX-NEXT: vmovsd {{.*#+}} xmm2 = [1.0E+0,0.0E+0]17; AVX-NEXT: vdivsd %xmm1, %xmm2, %xmm118; AVX-NEXT: vmovddup {{.*#+}} xmm1 = xmm1[0,0]19; AVX-NEXT: vmulpd %xmm1, %xmm0, %xmm020; AVX-NEXT: retq21 %vy = insertelement <2 x double> undef, double %y, i32 022 %splaty = shufflevector <2 x double> %vy, <2 x double> undef, <2 x i32> zeroinitializer23 %r = fdiv fast <2 x double> %x, %splaty24 ret <2 x double> %r25}26 27define <4 x double> @splat_fdiv_v4f64(<4 x double> %x, double %y) {28; SSE-LABEL: splat_fdiv_v4f64:29; SSE: # %bb.0:30; SSE-NEXT: movsd {{.*#+}} xmm3 = [1.0E+0,0.0E+0]31; SSE-NEXT: divsd %xmm2, %xmm332; SSE-NEXT: unpcklpd {{.*#+}} xmm3 = xmm3[0,0]33; SSE-NEXT: mulpd %xmm3, %xmm034; SSE-NEXT: mulpd %xmm3, %xmm135; SSE-NEXT: retq36;37; AVX-LABEL: splat_fdiv_v4f64:38; AVX: # %bb.0:39; AVX-NEXT: vmovsd {{.*#+}} xmm2 = [1.0E+0,0.0E+0]40; AVX-NEXT: vdivsd %xmm1, %xmm2, %xmm141; AVX-NEXT: vmovddup {{.*#+}} xmm1 = xmm1[0,0]42; AVX-NEXT: vinsertf128 $1, %xmm1, %ymm1, %ymm143; AVX-NEXT: vmulpd %ymm1, %ymm0, %ymm044; AVX-NEXT: retq45 %vy = insertelement <4 x double> undef, double %y, i32 046 %splaty = shufflevector <4 x double> %vy, <4 x double> undef, <4 x i32> zeroinitializer47 %r = fdiv arcp <4 x double> %x, %splaty48 ret <4 x double> %r49}50 51define <4 x float> @splat_fdiv_v4f32(<4 x float> %x, float %y) {52; SSE-LABEL: splat_fdiv_v4f32:53; SSE: # %bb.0:54; SSE-NEXT: movss {{.*#+}} xmm2 = [1.0E+0,0.0E+0,0.0E+0,0.0E+0]55; SSE-NEXT: divss %xmm1, %xmm256; SSE-NEXT: shufps {{.*#+}} xmm2 = xmm2[0,0,0,0]57; SSE-NEXT: mulps %xmm2, %xmm058; SSE-NEXT: retq59;60; AVX-LABEL: splat_fdiv_v4f32:61; AVX: # %bb.0:62; AVX-NEXT: vmovss {{.*#+}} xmm2 = [1.0E+0,0.0E+0,0.0E+0,0.0E+0]63; AVX-NEXT: vdivss %xmm1, %xmm2, %xmm164; AVX-NEXT: vshufps {{.*#+}} xmm1 = xmm1[0,0,0,0]65; AVX-NEXT: vmulps %xmm1, %xmm0, %xmm066; AVX-NEXT: retq67 %vy = insertelement <4 x float> undef, float %y, i32 068 %splaty = shufflevector <4 x float> %vy, <4 x float> undef, <4 x i32> zeroinitializer69 %r = fdiv arcp reassoc <4 x float> %x, %splaty70 ret <4 x float> %r71}72 73define <8 x float> @splat_fdiv_v8f32(<8 x float> %x, float %y) {74; SSE-LABEL: splat_fdiv_v8f32:75; SSE: # %bb.0:76; SSE-NEXT: movss {{.*#+}} xmm3 = [1.0E+0,0.0E+0,0.0E+0,0.0E+0]77; SSE-NEXT: divss %xmm2, %xmm378; SSE-NEXT: shufps {{.*#+}} xmm3 = xmm3[0,0,0,0]79; SSE-NEXT: mulps %xmm3, %xmm080; SSE-NEXT: mulps %xmm3, %xmm181; SSE-NEXT: retq82;83; AVX-LABEL: splat_fdiv_v8f32:84; AVX: # %bb.0:85; AVX-NEXT: vmovss {{.*#+}} xmm2 = [1.0E+0,0.0E+0,0.0E+0,0.0E+0]86; AVX-NEXT: vdivss %xmm1, %xmm2, %xmm187; AVX-NEXT: vshufps {{.*#+}} xmm1 = xmm1[0,0,0,0]88; AVX-NEXT: vinsertf128 $1, %xmm1, %ymm1, %ymm189; AVX-NEXT: vmulps %ymm1, %ymm0, %ymm090; AVX-NEXT: retq91 %vy = insertelement <8 x float> undef, float %y, i32 092 %splaty = shufflevector <8 x float> %vy, <8 x float> undef, <8 x i32> zeroinitializer93 %r = fdiv fast <8 x float> %x, %splaty94 ret <8 x float> %r95}96 97define <4 x float> @splat_fdiv_v4f32_estimate(<4 x float> %x, float %y) #0 {98; SSE-LABEL: splat_fdiv_v4f32_estimate:99; SSE: # %bb.0:100; SSE-NEXT: rcpss %xmm1, %xmm2101; SSE-NEXT: mulss %xmm2, %xmm1102; SSE-NEXT: movss {{.*#+}} xmm3 = [1.0E+0,0.0E+0,0.0E+0,0.0E+0]103; SSE-NEXT: subss %xmm1, %xmm3104; SSE-NEXT: mulss %xmm2, %xmm3105; SSE-NEXT: addss %xmm2, %xmm3106; SSE-NEXT: shufps {{.*#+}} xmm3 = xmm3[0,0,0,0]107; SSE-NEXT: mulps %xmm3, %xmm0108; SSE-NEXT: retq109;110; AVX-LABEL: splat_fdiv_v4f32_estimate:111; AVX: # %bb.0:112; AVX-NEXT: vrcpss %xmm1, %xmm1, %xmm2113; AVX-NEXT: vmulss %xmm2, %xmm1, %xmm1114; AVX-NEXT: vmovss {{.*#+}} xmm3 = [1.0E+0,0.0E+0,0.0E+0,0.0E+0]115; AVX-NEXT: vsubss %xmm1, %xmm3, %xmm1116; AVX-NEXT: vmulss %xmm1, %xmm2, %xmm1117; AVX-NEXT: vaddss %xmm1, %xmm2, %xmm1118; AVX-NEXT: vshufps {{.*#+}} xmm1 = xmm1[0,0,0,0]119; AVX-NEXT: vmulps %xmm1, %xmm0, %xmm0120; AVX-NEXT: retq121 %vy = insertelement <4 x float> undef, float %y, i32 0122 %splaty = shufflevector <4 x float> %vy, <4 x float> undef, <4 x i32> zeroinitializer123 %r = fdiv arcp reassoc ninf <4 x float> %x, %splaty124 ret <4 x float> %r125}126 127define <8 x float> @splat_fdiv_v8f32_estimate(<8 x float> %x, float %y) #0 {128; SSE-LABEL: splat_fdiv_v8f32_estimate:129; SSE: # %bb.0:130; SSE-NEXT: rcpss %xmm2, %xmm3131; SSE-NEXT: mulss %xmm3, %xmm2132; SSE-NEXT: movss {{.*#+}} xmm4 = [1.0E+0,0.0E+0,0.0E+0,0.0E+0]133; SSE-NEXT: subss %xmm2, %xmm4134; SSE-NEXT: mulss %xmm3, %xmm4135; SSE-NEXT: addss %xmm3, %xmm4136; SSE-NEXT: shufps {{.*#+}} xmm4 = xmm4[0,0,0,0]137; SSE-NEXT: mulps %xmm4, %xmm0138; SSE-NEXT: mulps %xmm4, %xmm1139; SSE-NEXT: retq140;141; AVX-LABEL: splat_fdiv_v8f32_estimate:142; AVX: # %bb.0:143; AVX-NEXT: vrcpss %xmm1, %xmm1, %xmm2144; AVX-NEXT: vmulss %xmm2, %xmm1, %xmm1145; AVX-NEXT: vmovss {{.*#+}} xmm3 = [1.0E+0,0.0E+0,0.0E+0,0.0E+0]146; AVX-NEXT: vsubss %xmm1, %xmm3, %xmm1147; AVX-NEXT: vmulss %xmm1, %xmm2, %xmm1148; AVX-NEXT: vaddss %xmm1, %xmm2, %xmm1149; AVX-NEXT: vshufps {{.*#+}} xmm1 = xmm1[0,0,0,0]150; AVX-NEXT: vinsertf128 $1, %xmm1, %ymm1, %ymm1151; AVX-NEXT: vmulps %ymm1, %ymm0, %ymm0152; AVX-NEXT: retq153 %vy = insertelement <8 x float> undef, float %y, i32 0154 %splaty = shufflevector <8 x float> %vy, <8 x float> undef, <8 x i32> zeroinitializer155 %r = fdiv fast <8 x float> %x, %splaty156 ret <8 x float> %r157}158 159attributes #0 = { "reciprocal-estimates"="divf,vec-divf" }160