214 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py2; RUN: llc < %s -mtriple=x86_64-unknown-linux-gnu -mattr=+avx,fma | FileCheck %s --check-prefix=FMA33; RUN: llc < %s -mtriple=x86_64-unknown-linux-gnu -mattr=+avx,fma4 | FileCheck %s --check-prefix=FMA44 5define float @test_fneg_fma_subx_y_negz_f32(float %w, float %x, float %y, float %z) {6; FMA3-LABEL: test_fneg_fma_subx_y_negz_f32:7; FMA3: # %bb.0: # %entry8; FMA3-NEXT: vsubss %xmm1, %xmm0, %xmm09; FMA3-NEXT: vfnmadd213ss {{.*#+}} xmm0 = -(xmm2 * xmm0) + xmm310; FMA3-NEXT: retq11;12; FMA4-LABEL: test_fneg_fma_subx_y_negz_f32:13; FMA4: # %bb.0: # %entry14; FMA4-NEXT: vsubss %xmm1, %xmm0, %xmm015; FMA4-NEXT: vfnmaddss {{.*#+}} xmm0 = -(xmm0 * xmm2) + xmm316; FMA4-NEXT: retq17entry:18 %subx = fsub nsz float %w, %x19 %negz = fsub float -0.000000e+00, %z20 %0 = tail call nsz float @llvm.fma.f32(float %subx, float %y, float %negz)21 %1 = fsub float -0.000000e+00, %022 ret float %123}24 25define float @test_fneg_fma_x_suby_negz_f32(float %w, float %x, float %y, float %z) {26; FMA3-LABEL: test_fneg_fma_x_suby_negz_f32:27; FMA3: # %bb.0: # %entry28; FMA3-NEXT: vsubss %xmm2, %xmm0, %xmm029; FMA3-NEXT: vfnmadd213ss {{.*#+}} xmm0 = -(xmm1 * xmm0) + xmm330; FMA3-NEXT: retq31;32; FMA4-LABEL: test_fneg_fma_x_suby_negz_f32:33; FMA4: # %bb.0: # %entry34; FMA4-NEXT: vsubss %xmm2, %xmm0, %xmm035; FMA4-NEXT: vfnmaddss {{.*#+}} xmm0 = -(xmm1 * xmm0) + xmm336; FMA4-NEXT: retq37entry:38 %suby = fsub nsz float %w, %y39 %negz = fsub float -0.000000e+00, %z40 %0 = tail call nsz float @llvm.fma.f32(float %x, float %suby, float %negz)41 %1 = fsub float -0.000000e+00, %042 ret float %143}44 45define float @test_fneg_fma_subx_suby_negz_f32(float %w, float %x, float %y, float %z) {46; FMA3-LABEL: test_fneg_fma_subx_suby_negz_f32:47; FMA3: # %bb.0: # %entry48; FMA3-NEXT: vsubss %xmm1, %xmm0, %xmm149; FMA3-NEXT: vsubss %xmm2, %xmm0, %xmm050; FMA3-NEXT: vfnmadd213ss {{.*#+}} xmm0 = -(xmm1 * xmm0) + xmm351; FMA3-NEXT: retq52;53; FMA4-LABEL: test_fneg_fma_subx_suby_negz_f32:54; FMA4: # %bb.0: # %entry55; FMA4-NEXT: vsubss %xmm1, %xmm0, %xmm156; FMA4-NEXT: vsubss %xmm2, %xmm0, %xmm057; FMA4-NEXT: vfnmaddss {{.*#+}} xmm0 = -(xmm1 * xmm0) + xmm358; FMA4-NEXT: retq59entry:60 %subx = fsub nsz float %w, %x61 %suby = fsub nsz float %w, %y62 %negz = fsub float -0.000000e+00, %z63 %0 = tail call nsz float @llvm.fma.f32(float %subx, float %suby, float %negz)64 %1 = fsub float -0.000000e+00, %065 ret float %166}67 68define float @test_fneg_fma_subx_negy_negz_f32(float %w, float %x, float %y, float %z) {69; FMA3-LABEL: test_fneg_fma_subx_negy_negz_f32:70; FMA3: # %bb.0: # %entry71; FMA3-NEXT: vsubss %xmm1, %xmm0, %xmm072; FMA3-NEXT: vfmadd213ss {{.*#+}} xmm0 = (xmm2 * xmm0) + xmm373; FMA3-NEXT: retq74;75; FMA4-LABEL: test_fneg_fma_subx_negy_negz_f32:76; FMA4: # %bb.0: # %entry77; FMA4-NEXT: vsubss %xmm1, %xmm0, %xmm078; FMA4-NEXT: vfmaddss {{.*#+}} xmm0 = (xmm0 * xmm2) + xmm379; FMA4-NEXT: retq80entry:81 %subx = fsub nsz float %w, %x82 %negy = fsub float -0.000000e+00, %y83 %negz = fsub float -0.000000e+00, %z84 %0 = tail call nsz float @llvm.fma.f32(float %subx, float %negy, float %negz)85 %1 = fsub float -0.000000e+00, %086 ret float %187}88 89define <4 x float> @test_fma_rcp_fneg_v4f32(<4 x float> %x, <4 x float> %y, <4 x float> %z) {90; FMA3-LABEL: test_fma_rcp_fneg_v4f32:91; FMA3: # %bb.0: # %entry92; FMA3-NEXT: vrcpps %xmm2, %xmm293; FMA3-NEXT: vfmsub213ps {{.*#+}} xmm0 = (xmm1 * xmm0) - xmm294; FMA3-NEXT: retq95;96; FMA4-LABEL: test_fma_rcp_fneg_v4f32:97; FMA4: # %bb.0: # %entry98; FMA4-NEXT: vrcpps %xmm2, %xmm299; FMA4-NEXT: vfmsubps {{.*#+}} xmm0 = (xmm0 * xmm1) - xmm2100; FMA4-NEXT: retq101entry:102 %0 = fneg <4 x float> %z103 %1 = tail call <4 x float> @llvm.x86.sse.rcp.ps(<4 x float> %0)104 %2 = tail call nsz <4 x float> @llvm.fma.v4f32(<4 x float> %x, <4 x float> %y, <4 x float> %1)105 ret <4 x float> %2106}107declare <4 x float> @llvm.x86.sse.rcp.ps(<4 x float>)108 109; This would crash while trying getNegatedExpression().110 111define float @negated_constant(float %x) {112; FMA3-LABEL: negated_constant:113; FMA3: # %bb.0:114; FMA3-NEXT: vmulss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm1115; FMA3-NEXT: vfnmsub132ss {{.*#+}} xmm0 = -(xmm0 * mem) - xmm1116; FMA3-NEXT: retq117;118; FMA4-LABEL: negated_constant:119; FMA4: # %bb.0:120; FMA4-NEXT: vmulss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm1121; FMA4-NEXT: vfnmsubss {{.*#+}} xmm0 = -(xmm0 * mem) - xmm1122; FMA4-NEXT: retq123 %m = fmul float %x, 42.0124 %fma = call nsz float @llvm.fma.f32(float %x, float -42.0, float %m)125 %nfma = fneg float %fma126 ret float %nfma127}128 129define <4 x double> @negated_constant_v4f64(<4 x double> %a) {130; FMA3-LABEL: negated_constant_v4f64:131; FMA3: # %bb.0:132; FMA3-NEXT: vmovapd {{.*#+}} ymm1 = [-5.0E-1,-2.5E-1,-1.25E-1,-6.25E-2]133; FMA3-NEXT: vfnmadd213pd {{.*#+}} ymm0 = -(ymm1 * ymm0) + ymm1134; FMA3-NEXT: retq135;136; FMA4-LABEL: negated_constant_v4f64:137; FMA4: # %bb.0:138; FMA4-NEXT: vmovapd {{.*#+}} ymm1 = [-5.0E-1,-2.5E-1,-1.25E-1,-6.25E-2]139; FMA4-NEXT: vfnmaddpd {{.*#+}} ymm0 = -(ymm0 * ymm1) + ymm1140; FMA4-NEXT: retq141 %t = tail call <4 x double> @llvm.fma.v4f64(<4 x double> %a, <4 x double> <double 5.000000e-01, double 2.5000000e-01, double 1.25000000e-01, double 0.62500000e-01>, <4 x double> <double -5.000000e-01, double -2.5000000e-01, double -1.25000000e-01, double -0.62500000e-01>)142 ret <4 x double> %t143}144 145define <4 x double> @negated_constant_v4f64_2fmas(<4 x double> %a, <4 x double> %b) {146; FMA3-LABEL: negated_constant_v4f64_2fmas:147; FMA3: # %bb.0:148; FMA3-NEXT: vmovapd {{.*#+}} ymm2 = [-5.0E-1,u,-2.5E+0,-4.5E+0]149; FMA3-NEXT: vmovapd %ymm2, %ymm3150; FMA3-NEXT: vfmadd213pd {{.*#+}} ymm3 = (ymm0 * ymm3) + ymm1151; FMA3-NEXT: vfnmadd213pd {{.*#+}} ymm2 = -(ymm0 * ymm2) + ymm1152; FMA3-NEXT: vaddpd %ymm2, %ymm3, %ymm0153; FMA3-NEXT: retq154;155; FMA4-LABEL: negated_constant_v4f64_2fmas:156; FMA4: # %bb.0:157; FMA4-NEXT: vmovapd {{.*#+}} ymm2 = [-5.0E-1,u,-2.5E+0,-4.5E+0]158; FMA4-NEXT: vfmaddpd {{.*#+}} ymm3 = (ymm0 * ymm2) + ymm1159; FMA4-NEXT: vfnmaddpd {{.*#+}} ymm0 = -(ymm0 * ymm2) + ymm1160; FMA4-NEXT: vaddpd %ymm0, %ymm3, %ymm0161; FMA4-NEXT: retq162 %t0 = tail call <4 x double> @llvm.fma.v4f64(<4 x double> %a, <4 x double> <double -5.000000e-01, double undef, double -25.000000e-01, double -45.000000e-01>, <4 x double> %b)163 %t1 = tail call <4 x double> @llvm.fma.v4f64(<4 x double> %a, <4 x double> <double 5.000000e-01, double undef, double 25.000000e-01, double 45.000000e-01>, <4 x double> %b)164 %t2 = fadd <4 x double> %t0, %t1165 ret <4 x double> %t2166}167 168define <4 x double> @negated_constant_v4f64_fadd(<4 x double> %a) {169; FMA3-LABEL: negated_constant_v4f64_fadd:170; FMA3: # %bb.0:171; FMA3-NEXT: vbroadcastf128 {{.*#+}} ymm1 = [1.5E+0,1.25E-1,1.5E+0,1.25E-1]172; FMA3-NEXT: # ymm1 = mem[0,1,0,1]173; FMA3-NEXT: vaddpd %ymm1, %ymm0, %ymm0174; FMA3-NEXT: vfmsub213pd {{.*#+}} ymm0 = (ymm1 * ymm0) - ymm1175; FMA3-NEXT: retq176;177; FMA4-LABEL: negated_constant_v4f64_fadd:178; FMA4: # %bb.0:179; FMA4-NEXT: vbroadcastf128 {{.*#+}} ymm1 = [1.5E+0,1.25E-1,1.5E+0,1.25E-1]180; FMA4-NEXT: # ymm1 = mem[0,1,0,1]181; FMA4-NEXT: vaddpd %ymm1, %ymm0, %ymm0182; FMA4-NEXT: vfmsubpd {{.*#+}} ymm0 = (ymm0 * ymm1) - ymm1183; FMA4-NEXT: retq184 %t0 = fadd <4 x double> %a, <double 15.000000e-01, double 1.25000000e-01, double 15.000000e-01, double 1.25000000e-01>185 %t1 = tail call <4 x double> @llvm.fma.v4f64(<4 x double> %t0, <4 x double> <double 15.000000e-01, double 1.25000000e-01, double 15.000000e-01, double 1.25000000e-01>, <4 x double> <double -15.000000e-01, double -1.25000000e-01, double -15.000000e-01, double -1.25000000e-01>)186 ret <4 x double> %t1187}188 189define <4 x double> @negated_constant_v4f64_2fma_undefs(<4 x double> %a, <4 x double> %b) {190; FMA3-LABEL: negated_constant_v4f64_2fma_undefs:191; FMA3: # %bb.0:192; FMA3-NEXT: vbroadcastsd {{.*#+}} ymm2 = [-5.0E-1,-5.0E-1,-5.0E-1,-5.0E-1]193; FMA3-NEXT: vfnmadd213pd {{.*#+}} ymm0 = -(ymm2 * ymm0) + mem194; FMA3-NEXT: vfmadd132pd {{.*#+}} ymm1 = (ymm1 * mem) + ymm2195; FMA3-NEXT: vaddpd %ymm1, %ymm0, %ymm0196; FMA3-NEXT: retq197;198; FMA4-LABEL: negated_constant_v4f64_2fma_undefs:199; FMA4: # %bb.0:200; FMA4-NEXT: vbroadcastsd {{.*#+}} ymm2 = [-5.0E-1,-5.0E-1,-5.0E-1,-5.0E-1]201; FMA4-NEXT: vfnmaddpd {{.*#+}} ymm0 = -(ymm0 * ymm2) + mem202; FMA4-NEXT: vfmaddpd {{.*#+}} ymm1 = (ymm1 * mem) + ymm2203; FMA4-NEXT: vaddpd %ymm1, %ymm0, %ymm0204; FMA4-NEXT: retq205 %t0 = tail call <4 x double> @llvm.fma.v4f64(<4 x double> %a, <4 x double> <double 5.000000e-01, double 5.000000e-01, double 5.000000e-01, double 5.000000e-01>, <4 x double> <double -5.000000e-01, double undef, double -5.000000e-01, double -5.000000e-01>)206 %t1 = tail call <4 x double> @llvm.fma.v4f64(<4 x double> %b, <4 x double> <double undef, double 5.000000e-01, double 5.000000e-01, double 5.000000e-01>, <4 x double> <double -5.000000e-01, double -5.000000e-01, double -5.000000e-01, double -5.000000e-01>)207 %t2 = fadd <4 x double> %t0, %t1208 ret <4 x double> %t2209}210 211declare float @llvm.fma.f32(float, float, float)212declare <4 x float> @llvm.fma.v4f32(<4 x float>, <4 x float>, <4 x float>)213declare <4 x double> @llvm.fma.v4f64(<4 x double>, <4 x double>, <4 x double>)214