1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py 2; RUN: llc < %s -mtriple=x86_64-unknown -mattr=+sse2 | FileCheck %s --check-prefix=SSE2 3; RUN: llc < %s -mtriple=x86_64-unknown -mattr=+ssse3 | FileCheck %s --check-prefixes=SSSE3,SSSE3-SLOW 4; RUN: llc < %s -mtriple=x86_64-unknown -mattr=+ssse3,fast-hops | FileCheck %s --check-prefixes=SSSE3,SSSE3-FAST 5; RUN: llc < %s -mtriple=x86_64-unknown -mattr=+avx | FileCheck %s --check-prefixes=AVX1,AVX1-SLOW 6; RUN: llc < %s -mtriple=x86_64-unknown -mattr=+avx,fast-hops | FileCheck %s --check-prefixes=AVX1,AVX1-FAST 7; RUN: llc < %s -mtriple=x86_64-unknown -mattr=+avx2 | FileCheck %s --check-prefix=AVX2 8 9define float @pr26491(<4 x float> %a0) { 10; SSE2-LABEL: pr26491: 11; SSE2: # %bb.0: 12; SSE2-NEXT: movaps %xmm0, %xmm1 13; SSE2-NEXT: shufps {{.*#+}} xmm1 = xmm1[1,1],xmm0[3,3] 14; SSE2-NEXT: addps %xmm0, %xmm1 15; SSE2-NEXT: movaps %xmm1, %xmm0 16; SSE2-NEXT: unpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1] 17; SSE2-NEXT: addss %xmm1, %xmm0 18; SSE2-NEXT: retq 19; 20; SSSE3-SLOW-LABEL: pr26491: 21; SSSE3-SLOW: # %bb.0: 22; SSSE3-SLOW-NEXT: movshdup {{.*#+}} xmm1 = xmm0[1,1,3,3] 23; SSSE3-SLOW-NEXT: addps %xmm0, %xmm1 24; SSSE3-SLOW-NEXT: movaps %xmm1, %xmm0 25; SSSE3-SLOW-NEXT: unpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1] 26; SSSE3-SLOW-NEXT: addss %xmm1, %xmm0 27; SSSE3-SLOW-NEXT: retq 28; 29; SSSE3-FAST-LABEL: pr26491: 30; SSSE3-FAST: # %bb.0: 31; SSSE3-FAST-NEXT: haddps %xmm0, %xmm0 32; SSSE3-FAST-NEXT: movshdup {{.*#+}} xmm1 = xmm0[1,1,3,3] 33; SSSE3-FAST-NEXT: addss %xmm1, %xmm0 34; SSSE3-FAST-NEXT: retq 35; 36; AVX1-SLOW-LABEL: pr26491: 37; AVX1-SLOW: # %bb.0: 38; AVX1-SLOW-NEXT: vmovshdup {{.*#+}} xmm1 = xmm0[1,1,3,3] 39; AVX1-SLOW-NEXT: vaddps %xmm0, %xmm1, %xmm0 40; AVX1-SLOW-NEXT: vpermilpd {{.*#+}} xmm1 = xmm0[1,0] 41; AVX1-SLOW-NEXT: vaddss %xmm0, %xmm1, %xmm0 42; AVX1-SLOW-NEXT: retq 43; 44; AVX1-FAST-LABEL: pr26491: 45; AVX1-FAST: # %bb.0: 46; AVX1-FAST-NEXT: vhaddps %xmm0, %xmm0, %xmm0 47; AVX1-FAST-NEXT: vmovshdup {{.*#+}} xmm1 = xmm0[1,1,3,3] 48; AVX1-FAST-NEXT: vaddss %xmm0, %xmm1, %xmm0 49; AVX1-FAST-NEXT: retq 50; 51; AVX2-LABEL: pr26491: 52; AVX2: # %bb.0: 53; AVX2-NEXT: vmovshdup {{.*#+}} xmm1 = xmm0[1,1,3,3] 54; AVX2-NEXT: vaddps %xmm0, %xmm1, %xmm0 55; AVX2-NEXT: vpermilpd {{.*#+}} xmm1 = xmm0[1,0] 56; AVX2-NEXT: vaddss %xmm0, %xmm1, %xmm0 57; AVX2-NEXT: retq 58 %1 = shufflevector <4 x float> %a0, <4 x float> undef, <4 x i32> <i32 1, i32 1, i32 3, i32 3> 59 %2 = fadd <4 x float> %1, %a0 60 %3 = extractelement <4 x float> %2, i32 2 61 %4 = extractelement <4 x float> %2, i32 0 62 %5 = fadd float %3, %4 63 ret float %5 64} 65 66; When simplifying away a splat (broadcast), the hop type must match the shuffle type. 67 68define <4 x double> @PR41414(i64 %x, <4 x double> %y) { 69; SSE2-LABEL: PR41414: 70; SSE2: # %bb.0: 71; SSE2-NEXT: movq %rdi, %xmm2 72; SSE2-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],mem[0],xmm2[1],mem[1] 73; SSE2-NEXT: subpd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm2 74; SSE2-NEXT: movapd %xmm2, %xmm3 75; SSE2-NEXT: unpcklpd {{.*#+}} xmm3 = xmm3[0],xmm2[0] 76; SSE2-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1,1] 77; SSE2-NEXT: addpd %xmm3, %xmm2 78; SSE2-NEXT: divpd %xmm2, %xmm1 79; SSE2-NEXT: divpd %xmm2, %xmm0 80; SSE2-NEXT: xorpd %xmm2, %xmm2 81; SSE2-NEXT: addpd %xmm2, %xmm0 82; SSE2-NEXT: addpd %xmm2, %xmm1 83; SSE2-NEXT: retq 84; 85; SSSE3-LABEL: PR41414: 86; SSSE3: # %bb.0: 87; SSSE3-NEXT: movq %rdi, %xmm2 88; SSSE3-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],mem[0],xmm2[1],mem[1] 89; SSSE3-NEXT: subpd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm2 90; SSSE3-NEXT: haddpd %xmm2, %xmm2 91; SSSE3-NEXT: divpd %xmm2, %xmm1 92; SSSE3-NEXT: divpd %xmm2, %xmm0 93; SSSE3-NEXT: xorpd %xmm2, %xmm2 94; SSSE3-NEXT: addpd %xmm2, %xmm0 95; SSSE3-NEXT: addpd %xmm2, %xmm1 96; SSSE3-NEXT: retq 97; 98; AVX1-LABEL: PR41414: 99; AVX1: # %bb.0: 100; AVX1-NEXT: vmovq %rdi, %xmm1 101; AVX1-NEXT: vpunpckldq {{.*#+}} xmm1 = xmm1[0],mem[0],xmm1[1],mem[1] 102; AVX1-NEXT: vsubpd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1, %xmm1 103; AVX1-NEXT: vhaddpd %xmm1, %xmm1, %xmm1 104; AVX1-NEXT: vinsertf128 $1, %xmm1, %ymm1, %ymm1 105; AVX1-NEXT: vdivpd %ymm1, %ymm0, %ymm0 106; AVX1-NEXT: vxorpd %xmm1, %xmm1, %xmm1 107; AVX1-NEXT: vaddpd %ymm1, %ymm0, %ymm0 108; AVX1-NEXT: retq 109; 110; AVX2-LABEL: PR41414: 111; AVX2: # %bb.0: 112; AVX2-NEXT: vmovq %rdi, %xmm1 113; AVX2-NEXT: vpunpckldq {{.*#+}} xmm1 = xmm1[0],mem[0],xmm1[1],mem[1] 114; AVX2-NEXT: vsubpd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1, %xmm1 115; AVX2-NEXT: vpermilpd {{.*#+}} xmm2 = xmm1[1,0] 116; AVX2-NEXT: vaddsd %xmm1, %xmm2, %xmm1 117; AVX2-NEXT: vbroadcastsd %xmm1, %ymm1 118; AVX2-NEXT: vdivpd %ymm1, %ymm0, %ymm0 119; AVX2-NEXT: vxorpd %xmm1, %xmm1, %xmm1 120; AVX2-NEXT: vaddpd %ymm1, %ymm0, %ymm0 121; AVX2-NEXT: retq 122 %conv = uitofp i64 %x to double 123 %t0 = insertelement <4 x double> undef, double %conv, i32 0 124 %t1 = shufflevector <4 x double> %t0, <4 x double> undef, <4 x i32> zeroinitializer 125 %t2 = fdiv <4 x double> %y, %t1 126 %t3 = fadd <4 x double> zeroinitializer, %t2 127 ret <4 x double> %t3 128} 129 130define <4 x float> @PR48823(<4 x float> %0, <4 x float> %1) { 131; SSE2-LABEL: PR48823: 132; SSE2: # %bb.0: 133; SSE2-NEXT: movaps %xmm0, %xmm2 134; SSE2-NEXT: shufps {{.*#+}} xmm2 = xmm2[1,1],xmm1[2,3] 135; SSE2-NEXT: shufps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,2] 136; SSE2-NEXT: subps %xmm2, %xmm0 137; SSE2-NEXT: retq 138; 139; SSSE3-SLOW-LABEL: PR48823: 140; SSSE3-SLOW: # %bb.0: 141; SSSE3-SLOW-NEXT: movaps %xmm0, %xmm2 142; SSSE3-SLOW-NEXT: shufps {{.*#+}} xmm2 = xmm2[1,1],xmm1[2,3] 143; SSSE3-SLOW-NEXT: shufps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,2] 144; SSSE3-SLOW-NEXT: subps %xmm2, %xmm0 145; SSSE3-SLOW-NEXT: retq 146; 147; SSSE3-FAST-LABEL: PR48823: 148; SSSE3-FAST: # %bb.0: 149; SSSE3-FAST-NEXT: hsubps %xmm1, %xmm0 150; SSSE3-FAST-NEXT: retq 151; 152; AVX1-SLOW-LABEL: PR48823: 153; AVX1-SLOW: # %bb.0: 154; AVX1-SLOW-NEXT: vshufps {{.*#+}} xmm2 = xmm0[1,1],xmm1[2,3] 155; AVX1-SLOW-NEXT: vshufps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,2] 156; AVX1-SLOW-NEXT: vsubps %xmm2, %xmm0, %xmm0 157; AVX1-SLOW-NEXT: retq 158; 159; AVX1-FAST-LABEL: PR48823: 160; AVX1-FAST: # %bb.0: 161; AVX1-FAST-NEXT: vhsubps %xmm1, %xmm0, %xmm0 162; AVX1-FAST-NEXT: retq 163; 164; AVX2-LABEL: PR48823: 165; AVX2: # %bb.0: 166; AVX2-NEXT: vshufps {{.*#+}} xmm2 = xmm0[1,1],xmm1[2,3] 167; AVX2-NEXT: vshufps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,2] 168; AVX2-NEXT: vsubps %xmm2, %xmm0, %xmm0 169; AVX2-NEXT: retq 170 %3 = shufflevector <4 x float> %0, <4 x float> poison, <4 x i32> <i32 1, i32 undef, i32 undef, i32 undef> 171 %4 = fsub <4 x float> %0, %3 172 %5 = shufflevector <4 x float> %1, <4 x float> poison, <4 x i32> <i32 undef, i32 undef, i32 undef, i32 2> 173 %6 = fsub <4 x float> %5, %1 174 %7 = shufflevector <4 x float> %4, <4 x float> %6, <4 x i32> <i32 0, i32 undef, i32 undef, i32 7> 175 ret <4 x float> %7 176} 177