1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py 2; RUN: llc -O3 -mtriple=x86_64-pc-linux -mattr=+fma < %s | FileCheck %s 3 4define <1 x float> @constrained_vector_fma_v1f32() #0 { 5; CHECK-LABEL: constrained_vector_fma_v1f32: 6; CHECK: # %bb.0: # %entry 7; CHECK-NEXT: vmovss {{.*#+}} xmm1 = mem[0],zero,zero,zero 8; CHECK-NEXT: vmovss {{.*#+}} xmm0 = mem[0],zero,zero,zero 9; CHECK-NEXT: vfmadd213ss {{.*#+}} xmm0 = (xmm1 * xmm0) + mem 10; CHECK-NEXT: retq 11entry: 12 %fma = call <1 x float> @llvm.experimental.constrained.fma.v1f32( 13 <1 x float> <float 0.5>, 14 <1 x float> <float 2.5>, 15 <1 x float> <float 4.5>, 16 metadata !"round.dynamic", 17 metadata !"fpexcept.strict") #0 18 ret <1 x float> %fma 19} 20 21define <2 x double> @constrained_vector_fma_v2f64() #0 { 22; CHECK-LABEL: constrained_vector_fma_v2f64: 23; CHECK: # %bb.0: # %entry 24; CHECK-NEXT: vmovapd {{.*#+}} xmm1 = [1.5E+0,5.0E-1] 25; CHECK-NEXT: vmovapd {{.*#+}} xmm0 = [3.5E+0,2.5E+0] 26; CHECK-NEXT: vfmadd213pd {{.*#+}} xmm0 = (xmm1 * xmm0) + mem 27; CHECK-NEXT: retq 28entry: 29 %fma = call <2 x double> @llvm.experimental.constrained.fma.v2f64( 30 <2 x double> <double 1.5, double 0.5>, 31 <2 x double> <double 3.5, double 2.5>, 32 <2 x double> <double 5.5, double 4.5>, 33 metadata !"round.dynamic", 34 metadata !"fpexcept.strict") #0 35 ret <2 x double> %fma 36} 37 38define <3 x float> @constrained_vector_fma_v3f32() #0 { 39; CHECK-LABEL: constrained_vector_fma_v3f32: 40; CHECK: # %bb.0: # %entry 41; CHECK-NEXT: vmovss {{.*#+}} xmm0 = mem[0],zero,zero,zero 42; CHECK-NEXT: vmovss {{.*#+}} xmm1 = mem[0],zero,zero,zero 43; CHECK-NEXT: vfmadd213ss {{.*#+}} xmm1 = (xmm0 * xmm1) + mem 44; CHECK-NEXT: vmovss {{.*#+}} xmm0 = mem[0],zero,zero,zero 45; CHECK-NEXT: vmovss {{.*#+}} xmm2 = mem[0],zero,zero,zero 46; CHECK-NEXT: vfmadd213ss {{.*#+}} xmm2 = (xmm0 * xmm2) + mem 47; CHECK-NEXT: vmovss {{.*#+}} xmm0 = mem[0],zero,zero,zero 48; CHECK-NEXT: vmovss {{.*#+}} xmm3 = mem[0],zero,zero,zero 49; CHECK-NEXT: vfmadd213ss {{.*#+}} xmm3 = (xmm0 * xmm3) + mem 50; CHECK-NEXT: vinsertps {{.*#+}} xmm0 = xmm2[0],xmm3[0],xmm2[2,3] 51; CHECK-NEXT: vinsertps {{.*#+}} xmm0 = xmm0[0,1],xmm1[0],xmm0[3] 52; CHECK-NEXT: retq 53entry: 54 %fma = call <3 x float> @llvm.experimental.constrained.fma.v3f32( 55 <3 x float> <float 2.5, float 1.5, float 0.5>, 56 <3 x float> <float 5.5, float 4.5, float 3.5>, 57 <3 x float> <float 8.5, float 7.5, float 6.5>, 58 metadata !"round.dynamic", 59 metadata !"fpexcept.strict") #0 60 ret <3 x float> %fma 61} 62 63define <3 x double> @constrained_vector_fma_v3f64() #0 { 64; CHECK-LABEL: constrained_vector_fma_v3f64: 65; CHECK: # %bb.0: # %entry 66; CHECK-NEXT: vmovsd {{.*#+}} xmm0 = mem[0],zero 67; CHECK-NEXT: vmovsd {{.*#+}} xmm1 = mem[0],zero 68; CHECK-NEXT: vfmadd213sd {{.*#+}} xmm1 = (xmm0 * xmm1) + mem 69; CHECK-NEXT: vmovapd {{.*#+}} xmm0 = [2.5E+0,1.5E+0] 70; CHECK-NEXT: vmovapd {{.*#+}} xmm2 = [5.5E+0,4.5E+0] 71; CHECK-NEXT: vfmadd213pd {{.*#+}} xmm2 = (xmm0 * xmm2) + mem 72; CHECK-NEXT: vinsertf128 $1, %xmm1, %ymm2, %ymm0 73; CHECK-NEXT: retq 74entry: 75 %fma = call <3 x double> @llvm.experimental.constrained.fma.v3f64( 76 <3 x double> <double 2.5, double 1.5, double 0.5>, 77 <3 x double> <double 5.5, double 4.5, double 3.5>, 78 <3 x double> <double 8.5, double 7.5, double 6.5>, 79 metadata !"round.dynamic", 80 metadata !"fpexcept.strict") #0 81 ret <3 x double> %fma 82} 83 84define <4 x double> @constrained_vector_fma_v4f64() #0 { 85; CHECK-LABEL: constrained_vector_fma_v4f64: 86; CHECK: # %bb.0: # %entry 87; CHECK-NEXT: vmovapd {{.*#+}} ymm1 = [3.5E+0,2.5E+0,1.5E+0,5.0E-1] 88; CHECK-NEXT: vmovapd {{.*#+}} ymm0 = [7.5E+0,6.5E+0,5.5E+0,4.5E+0] 89; CHECK-NEXT: vfmadd213pd {{.*#+}} ymm0 = (ymm1 * ymm0) + mem 90; CHECK-NEXT: retq 91entry: 92 %fma = call <4 x double> @llvm.experimental.constrained.fma.v4f64( 93 <4 x double> <double 3.5, double 2.5, double 1.5, double 0.5>, 94 <4 x double> <double 7.5, double 6.5, double 5.5, double 4.5>, 95 <4 x double> <double 11.5, double 10.5, double 9.5, double 8.5>, 96 metadata !"round.dynamic", 97 metadata !"fpexcept.strict") #0 98 ret <4 x double> %fma 99} 100 101define <4 x float> @constrained_vector_fma_v4f32() #0 { 102; CHECK-LABEL: constrained_vector_fma_v4f32: 103; CHECK: # %bb.0: # %entry 104; CHECK-NEXT: vmovaps {{.*#+}} xmm1 = [3.5E+0,2.5E+0,1.5E+0,5.0E-1] 105; CHECK-NEXT: vmovaps {{.*#+}} xmm0 = [7.5E+0,6.5E+0,5.5E+0,4.5E+0] 106; CHECK-NEXT: vfmadd213ps {{.*#+}} xmm0 = (xmm1 * xmm0) + mem 107; CHECK-NEXT: retq 108entry: 109 %fma = call <4 x float> @llvm.experimental.constrained.fma.v4f32( 110 <4 x float> <float 3.5, float 2.5, float 1.5, float 0.5>, 111 <4 x float> <float 7.5, float 6.5, float 5.5, float 4.5>, 112 <4 x float> <float 11.5, float 10.5, float 9.5, float 8.5>, 113 metadata !"round.dynamic", 114 metadata !"fpexcept.strict") #0 115 ret <4 x float> %fma 116} 117 118define <8 x float> @constrained_vector_fma_v8f32() #0 { 119; CHECK-LABEL: constrained_vector_fma_v8f32: 120; CHECK: # %bb.0: # %entry 121; CHECK-NEXT: vmovaps {{.*#+}} ymm1 = [3.5E+0,2.5E+0,1.5E+0,5.0E-1,7.5E+0,6.5E+0,5.5E+0,4.5E+0] 122; CHECK-NEXT: vmovaps {{.*#+}} ymm0 = [7.5E+0,6.5E+0,5.5E+0,4.5E+0,1.15E+1,1.05E+1,9.5E+0,8.5E+0] 123; CHECK-NEXT: vfmadd213ps {{.*#+}} ymm0 = (ymm1 * ymm0) + mem 124; CHECK-NEXT: retq 125entry: 126 %fma = call <8 x float> @llvm.experimental.constrained.fma.v8f32( 127 <8 x float> <float 3.5, float 2.5, float 1.5, float 0.5, 128 float 7.5, float 6.5, float 5.5, float 4.5>, 129 <8 x float> <float 7.5, float 6.5, float 5.5, float 4.5, 130 float 11.5, float 10.5, float 9.5, float 8.5>, 131 <8 x float> <float 11.5, float 10.5, float 9.5, float 8.5, 132 float 15.5, float 14.5, float 13.5, float 12.5>, 133 metadata !"round.dynamic", 134 metadata !"fpexcept.strict") #0 135 ret <8 x float> %fma 136} 137 138attributes #0 = { strictfp } 139 140; Single width declarations 141declare <2 x double> @llvm.experimental.constrained.fma.v2f64(<2 x double>, <2 x double>, <2 x double>, metadata, metadata) 142declare <4 x float> @llvm.experimental.constrained.fma.v4f32(<4 x float>, <4 x float>, <4 x float>, metadata, metadata) 143 144; Scalar width declarations 145declare <1 x float> @llvm.experimental.constrained.fma.v1f32(<1 x float>, <1 x float>, <1 x float>, metadata, metadata) 146 147; Illegal width declarations 148declare <3 x float> @llvm.experimental.constrained.fma.v3f32(<3 x float>, <3 x float>, <3 x float>, metadata, metadata) 149declare <3 x double> @llvm.experimental.constrained.fma.v3f64(<3 x double>, <3 x double>, <3 x double>, metadata, metadata) 150 151; Double width declarations 152declare <4 x double> @llvm.experimental.constrained.fma.v4f64(<4 x double>, <4 x double>, <4 x double>, metadata, metadata) 153declare <8 x float> @llvm.experimental.constrained.fma.v8f32(<8 x float>, <8 x float>, <8 x float>, metadata, metadata) 154