1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py 2; RUN: llc -mtriple=x86_64-unknown-unknown -mattr=+sse2 < %s | FileCheck %s --check-prefix=SSE 3; RUN: llc -mtriple=x86_64-unknown-unknown -mattr=+avx < %s | FileCheck %s --check-prefix=AVX 4; RUN: llc -mtriple=x86_64-unknown-unknown -mattr=+avx512f < %s | FileCheck %s --check-prefix=AVX 5 6; Verify we fold loads into unary sse intrinsics only when optimizing for size 7 8define float @rcpss(float* %a) { 9; SSE-LABEL: rcpss: 10; SSE: # %bb.0: 11; SSE-NEXT: movss {{.*#+}} xmm0 = mem[0],zero,zero,zero 12; SSE-NEXT: rcpss %xmm0, %xmm0 13; SSE-NEXT: retq 14; 15; AVX-LABEL: rcpss: 16; AVX: # %bb.0: 17; AVX-NEXT: vmovss {{.*#+}} xmm0 = mem[0],zero,zero,zero 18; AVX-NEXT: vrcpss %xmm0, %xmm0, %xmm0 19; AVX-NEXT: retq 20 %ld = load float, float* %a 21 %ins = insertelement <4 x float> undef, float %ld, i32 0 22 %res = tail call <4 x float> @llvm.x86.sse.rcp.ss(<4 x float> %ins) 23 %ext = extractelement <4 x float> %res, i32 0 24 ret float %ext 25} 26 27define float @rsqrtss(float* %a) { 28; SSE-LABEL: rsqrtss: 29; SSE: # %bb.0: 30; SSE-NEXT: movss {{.*#+}} xmm0 = mem[0],zero,zero,zero 31; SSE-NEXT: rsqrtss %xmm0, %xmm0 32; SSE-NEXT: retq 33; 34; AVX-LABEL: rsqrtss: 35; AVX: # %bb.0: 36; AVX-NEXT: vmovss {{.*#+}} xmm0 = mem[0],zero,zero,zero 37; AVX-NEXT: vrsqrtss %xmm0, %xmm0, %xmm0 38; AVX-NEXT: retq 39 %ld = load float, float* %a 40 %ins = insertelement <4 x float> undef, float %ld, i32 0 41 %res = tail call <4 x float> @llvm.x86.sse.rsqrt.ss(<4 x float> %ins) 42 %ext = extractelement <4 x float> %res, i32 0 43 ret float %ext 44} 45 46define float @sqrtss(float* %a) { 47; SSE-LABEL: sqrtss: 48; SSE: # %bb.0: 49; SSE-NEXT: movss {{.*#+}} xmm0 = mem[0],zero,zero,zero 50; SSE-NEXT: sqrtss %xmm0, %xmm0 51; SSE-NEXT: retq 52; 53; AVX-LABEL: sqrtss: 54; AVX: # %bb.0: 55; AVX-NEXT: vmovss {{.*#+}} xmm0 = mem[0],zero,zero,zero 56; AVX-NEXT: vsqrtss %xmm0, %xmm0, %xmm0 57; AVX-NEXT: retq 58 %ld = load float, float* %a 59 %ins = insertelement <4 x float> undef, float %ld, i32 0 60 %res = tail call <4 x float> @llvm.x86.sse.sqrt.ss(<4 x float> %ins) 61 %ext = extractelement <4 x float> %res, i32 0 62 ret float %ext 63} 64 65define double @sqrtsd(double* %a) { 66; SSE-LABEL: sqrtsd: 67; SSE: # %bb.0: 68; SSE-NEXT: movsd {{.*#+}} xmm0 = mem[0],zero 69; SSE-NEXT: sqrtsd %xmm0, %xmm0 70; SSE-NEXT: retq 71; 72; AVX-LABEL: sqrtsd: 73; AVX: # %bb.0: 74; AVX-NEXT: vmovsd {{.*#+}} xmm0 = mem[0],zero 75; AVX-NEXT: vsqrtsd %xmm0, %xmm0, %xmm0 76; AVX-NEXT: retq 77 %ld = load double, double* %a 78 %ins = insertelement <2 x double> undef, double %ld, i32 0 79 %res = tail call <2 x double> @llvm.x86.sse2.sqrt.sd(<2 x double> %ins) 80 %ext = extractelement <2 x double> %res, i32 0 81 ret double %ext 82} 83 84define float @rcpss_size(float* %a) optsize { 85; SSE-LABEL: rcpss_size: 86; SSE: # %bb.0: 87; SSE-NEXT: rcpss (%rdi), %xmm0 88; SSE-NEXT: retq 89; 90; AVX-LABEL: rcpss_size: 91; AVX: # %bb.0: 92; AVX-NEXT: vrcpss (%rdi), %xmm0, %xmm0 93; AVX-NEXT: retq 94 %ld = load float, float* %a 95 %ins = insertelement <4 x float> undef, float %ld, i32 0 96 %res = tail call <4 x float> @llvm.x86.sse.rcp.ss(<4 x float> %ins) 97 %ext = extractelement <4 x float> %res, i32 0 98 ret float %ext 99} 100 101define <4 x float> @rcpss_full_size(<4 x float>* %a) optsize { 102; SSE-LABEL: rcpss_full_size: 103; SSE: # %bb.0: 104; SSE-NEXT: rcpss (%rdi), %xmm0 105; SSE-NEXT: retq 106; 107; AVX-LABEL: rcpss_full_size: 108; AVX: # %bb.0: 109; AVX-NEXT: vrcpss (%rdi), %xmm0, %xmm0 110; AVX-NEXT: retq 111 %ld = load <4 x float>, <4 x float>* %a 112 %res = tail call <4 x float> @llvm.x86.sse.rcp.ss(<4 x float> %ld) 113 ret <4 x float> %res 114} 115 116define float @rsqrtss_size(float* %a) optsize { 117; SSE-LABEL: rsqrtss_size: 118; SSE: # %bb.0: 119; SSE-NEXT: rsqrtss (%rdi), %xmm0 120; SSE-NEXT: retq 121; 122; AVX-LABEL: rsqrtss_size: 123; AVX: # %bb.0: 124; AVX-NEXT: vrsqrtss (%rdi), %xmm0, %xmm0 125; AVX-NEXT: retq 126 %ld = load float, float* %a 127 %ins = insertelement <4 x float> undef, float %ld, i32 0 128 %res = tail call <4 x float> @llvm.x86.sse.rsqrt.ss(<4 x float> %ins) 129 %ext = extractelement <4 x float> %res, i32 0 130 ret float %ext 131} 132 133define <4 x float> @rsqrtss_full_size(<4 x float>* %a) optsize { 134; SSE-LABEL: rsqrtss_full_size: 135; SSE: # %bb.0: 136; SSE-NEXT: rsqrtss (%rdi), %xmm0 137; SSE-NEXT: retq 138; 139; AVX-LABEL: rsqrtss_full_size: 140; AVX: # %bb.0: 141; AVX-NEXT: vrsqrtss (%rdi), %xmm0, %xmm0 142; AVX-NEXT: retq 143 %ld = load <4 x float>, <4 x float>* %a 144 %res = tail call <4 x float> @llvm.x86.sse.rsqrt.ss(<4 x float> %ld) 145 ret <4 x float> %res 146} 147 148define float @sqrtss_size(float* %a) optsize{ 149; SSE-LABEL: sqrtss_size: 150; SSE: # %bb.0: 151; SSE-NEXT: sqrtss (%rdi), %xmm0 152; SSE-NEXT: retq 153; 154; AVX-LABEL: sqrtss_size: 155; AVX: # %bb.0: 156; AVX-NEXT: vsqrtss (%rdi), %xmm0, %xmm0 157; AVX-NEXT: retq 158 %ld = load float, float* %a 159 %ins = insertelement <4 x float> undef, float %ld, i32 0 160 %res = tail call <4 x float> @llvm.x86.sse.sqrt.ss(<4 x float> %ins) 161 %ext = extractelement <4 x float> %res, i32 0 162 ret float %ext 163} 164 165define <4 x float> @sqrtss_full_size(<4 x float>* %a) optsize{ 166; SSE-LABEL: sqrtss_full_size: 167; SSE: # %bb.0: 168; SSE-NEXT: movaps (%rdi), %xmm0 169; SSE-NEXT: sqrtss %xmm0, %xmm0 170; SSE-NEXT: retq 171; 172; AVX-LABEL: sqrtss_full_size: 173; AVX: # %bb.0: 174; AVX-NEXT: vmovaps (%rdi), %xmm0 175; AVX-NEXT: vsqrtss %xmm0, %xmm0, %xmm0 176; AVX-NEXT: retq 177 %ld = load <4 x float>, <4 x float>* %a 178 %res = tail call <4 x float> @llvm.x86.sse.sqrt.ss(<4 x float> %ld) 179 ret <4 x float> %res 180} 181 182define double @sqrtsd_size(double* %a) optsize { 183; SSE-LABEL: sqrtsd_size: 184; SSE: # %bb.0: 185; SSE-NEXT: sqrtsd (%rdi), %xmm0 186; SSE-NEXT: retq 187; 188; AVX-LABEL: sqrtsd_size: 189; AVX: # %bb.0: 190; AVX-NEXT: vsqrtsd (%rdi), %xmm0, %xmm0 191; AVX-NEXT: retq 192 %ld = load double, double* %a 193 %ins = insertelement <2 x double> undef, double %ld, i32 0 194 %res = tail call <2 x double> @llvm.x86.sse2.sqrt.sd(<2 x double> %ins) 195 %ext = extractelement <2 x double> %res, i32 0 196 ret double %ext 197} 198 199define <2 x double> @sqrtsd_full_size(<2 x double>* %a) optsize { 200; SSE-LABEL: sqrtsd_full_size: 201; SSE: # %bb.0: 202; SSE-NEXT: movapd (%rdi), %xmm0 203; SSE-NEXT: sqrtsd %xmm0, %xmm0 204; SSE-NEXT: retq 205; 206; AVX-LABEL: sqrtsd_full_size: 207; AVX: # %bb.0: 208; AVX-NEXT: vmovapd (%rdi), %xmm0 209; AVX-NEXT: vsqrtsd %xmm0, %xmm0, %xmm0 210; AVX-NEXT: retq 211 %ld = load <2 x double>, <2 x double>* %a 212 %res = tail call <2 x double> @llvm.x86.sse2.sqrt.sd(<2 x double> %ld) 213 ret <2 x double> %res 214} 215 216declare <4 x float> @llvm.x86.sse.rcp.ss(<4 x float>) nounwind readnone 217declare <4 x float> @llvm.x86.sse.rsqrt.ss(<4 x float>) nounwind readnone 218declare <4 x float> @llvm.x86.sse.sqrt.ss(<4 x float>) nounwind readnone 219declare <2 x double> @llvm.x86.sse2.sqrt.sd(<2 x double>) nounwind readnone 220