407 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py2; RUN: llc < %s -mtriple=x86_64-unknown -mattr=+ssse3 | FileCheck %s --check-prefix=SSE3; RUN: llc < %s -mtriple=x86_64-unknown -mattr=+ssse3,fast-hops | FileCheck %s --check-prefix=SSE4; RUN: llc < %s -mtriple=x86_64-unknown -mattr=+avx | FileCheck %s --check-prefixes=AVX,AVX15; RUN: llc < %s -mtriple=x86_64-unknown -mattr=+avx,fast-hops | FileCheck %s --check-prefixes=AVX,AVX16; RUN: llc < %s -mtriple=x86_64-unknown -mattr=+avx2 | FileCheck %s --check-prefixes=AVX,AVX27; RUN: llc < %s -mtriple=x86_64-unknown -mattr=+avx2,fast-hops | FileCheck %s --check-prefixes=AVX,AVX28 9define <8 x i16> @hadd_reverse_v8i16(<8 x i16> %a0, <8 x i16> %a1) nounwind {10; SSE-LABEL: hadd_reverse_v8i16:11; SSE: # %bb.0:12; SSE-NEXT: phaddw %xmm1, %xmm013; SSE-NEXT: pshuflw {{.*#+}} xmm0 = xmm0[3,2,1,0,4,5,6,7]14; SSE-NEXT: pshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,7,6,5,4]15; SSE-NEXT: retq16;17; AVX-LABEL: hadd_reverse_v8i16:18; AVX: # %bb.0:19; AVX-NEXT: vphaddw %xmm1, %xmm0, %xmm020; AVX-NEXT: vpshuflw {{.*#+}} xmm0 = xmm0[3,2,1,0,4,5,6,7]21; AVX-NEXT: vpshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,7,6,5,4]22; AVX-NEXT: retq23 %lhs = shufflevector <8 x i16> %a0, <8 x i16> %a1, <8 x i32> <i32 7, i32 5, i32 3, i32 1, i32 15, i32 13, i32 11, i32 9>24 %rhs = shufflevector <8 x i16> %a0, <8 x i16> %a1, <8 x i32> <i32 6, i32 4, i32 2, i32 0, i32 14, i32 12, i32 10, i32 8>25 %add = add <8 x i16> %lhs, %rhs26 ret <8 x i16> %add27}28 29define <8 x i16> @hadd_reverse2_v8i16(<8 x i16> %a0, <8 x i16> %a1) nounwind {30; SSE-LABEL: hadd_reverse2_v8i16:31; SSE: # %bb.0:32; SSE-NEXT: movdqa {{.*#+}} xmm2 = [14,15,12,13,10,11,8,9,6,7,4,5,2,3,0,1]33; SSE-NEXT: pshufb %xmm2, %xmm034; SSE-NEXT: pshufb %xmm2, %xmm135; SSE-NEXT: phaddw %xmm1, %xmm036; SSE-NEXT: retq37;38; AVX-LABEL: hadd_reverse2_v8i16:39; AVX: # %bb.0:40; AVX-NEXT: vmovdqa {{.*#+}} xmm2 = [14,15,12,13,10,11,8,9,6,7,4,5,2,3,0,1]41; AVX-NEXT: vpshufb %xmm2, %xmm0, %xmm042; AVX-NEXT: vpshufb %xmm2, %xmm1, %xmm143; AVX-NEXT: vphaddw %xmm1, %xmm0, %xmm044; AVX-NEXT: retq45 %shuf0 = shufflevector <8 x i16> %a0, <8 x i16> undef, <8 x i32> <i32 7, i32 6, i32 5, i32 4, i32 3, i32 2, i32 1, i32 0>46 %shuf1 = shufflevector <8 x i16> %a1, <8 x i16> undef, <8 x i32> <i32 7, i32 6, i32 5, i32 4, i32 3, i32 2, i32 1, i32 0>47 %lhs = shufflevector <8 x i16> %shuf0, <8 x i16> %shuf1, <8 x i32> <i32 0, i32 2, i32 4, i32 6, i32 8, i32 10, i32 12, i32 14>48 %rhs = shufflevector <8 x i16> %shuf0, <8 x i16> %shuf1, <8 x i32> <i32 1, i32 3, i32 5, i32 7, i32 9, i32 11, i32 13, i32 15>49 %add = add <8 x i16> %lhs, %rhs50 ret <8 x i16> %add51}52 53define <8 x float> @hadd_reverse_v8f32(<8 x float> %a0, <8 x float> %a1) {54; SSE-LABEL: hadd_reverse_v8f32:55; SSE: # %bb.0:56; SSE-NEXT: movaps %xmm0, %xmm457; SSE-NEXT: haddps %xmm3, %xmm158; SSE-NEXT: shufps {{.*#+}} xmm1 = xmm1[1,0,3,2]59; SSE-NEXT: haddps %xmm2, %xmm460; SSE-NEXT: shufps {{.*#+}} xmm4 = xmm4[1,0,3,2]61; SSE-NEXT: movaps %xmm1, %xmm062; SSE-NEXT: movaps %xmm4, %xmm163; SSE-NEXT: retq64;65; AVX1-LABEL: hadd_reverse_v8f32:66; AVX1: # %bb.0:67; AVX1-NEXT: vhaddps %ymm1, %ymm0, %ymm068; AVX1-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3,0,1]69; AVX1-NEXT: vshufps {{.*#+}} ymm0 = ymm0[1,0,3,2,5,4,7,6]70; AVX1-NEXT: retq71;72; AVX2-LABEL: hadd_reverse_v8f32:73; AVX2: # %bb.0:74; AVX2-NEXT: vhaddps %ymm1, %ymm0, %ymm075; AVX2-NEXT: vshufps {{.*#+}} ymm0 = ymm0[1,0,3,2,5,4,7,6]76; AVX2-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[2,3,0,1]77; AVX2-NEXT: retq78 %lhs = shufflevector <8 x float> %a0, <8 x float> %a1, <8 x i32> <i32 7, i32 5, i32 15, i32 13, i32 3, i32 1, i32 11, i32 9>79 %rhs = shufflevector <8 x float> %a0, <8 x float> %a1, <8 x i32> <i32 6, i32 4, i32 14, i32 12, i32 2, i32 0, i32 10, i32 8>80 %add = fadd <8 x float> %lhs, %rhs81 ret <8 x float> %add82}83 84define <8 x float> @hadd_reverse2_v8f32(<8 x float> %a0, <8 x float> %a1) {85; SSE-LABEL: hadd_reverse2_v8f32:86; SSE: # %bb.0:87; SSE-NEXT: movaps %xmm0, %xmm488; SSE-NEXT: haddps %xmm3, %xmm189; SSE-NEXT: shufps {{.*#+}} xmm1 = xmm1[1,0,3,2]90; SSE-NEXT: haddps %xmm2, %xmm491; SSE-NEXT: shufps {{.*#+}} xmm4 = xmm4[1,0,3,2]92; SSE-NEXT: movaps %xmm1, %xmm093; SSE-NEXT: movaps %xmm4, %xmm194; SSE-NEXT: retq95;96; AVX1-LABEL: hadd_reverse2_v8f32:97; AVX1: # %bb.0:98; AVX1-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3,0,1]99; AVX1-NEXT: vshufps {{.*#+}} ymm0 = ymm0[3,2,1,0,7,6,5,4]100; AVX1-NEXT: vperm2f128 {{.*#+}} ymm1 = ymm1[2,3,0,1]101; AVX1-NEXT: vshufps {{.*#+}} ymm1 = ymm1[3,2,1,0,7,6,5,4]102; AVX1-NEXT: vhaddps %ymm1, %ymm0, %ymm0103; AVX1-NEXT: retq104;105; AVX2-LABEL: hadd_reverse2_v8f32:106; AVX2: # %bb.0:107; AVX2-NEXT: vshufps {{.*#+}} ymm0 = ymm0[3,2,1,0,7,6,5,4]108; AVX2-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[2,3,0,1]109; AVX2-NEXT: vshufps {{.*#+}} ymm1 = ymm1[3,2,1,0,7,6,5,4]110; AVX2-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,3,0,1]111; AVX2-NEXT: vhaddps %ymm1, %ymm0, %ymm0112; AVX2-NEXT: retq113 %shuf0 = shufflevector <8 x float> %a0, <8 x float> undef, <8 x i32> <i32 7, i32 6, i32 5, i32 4, i32 3, i32 2, i32 1, i32 0>114 %shuf1 = shufflevector <8 x float> %a1, <8 x float> undef, <8 x i32> <i32 7, i32 6, i32 5, i32 4, i32 3, i32 2, i32 1, i32 0>115 %lhs = shufflevector <8 x float> %shuf0, <8 x float> %shuf1, <8 x i32> <i32 0, i32 2, i32 8, i32 10, i32 4, i32 6, i32 12, i32 14>116 %rhs = shufflevector <8 x float> %shuf0, <8 x float> %shuf1, <8 x i32> <i32 1, i32 3, i32 9, i32 11, i32 5, i32 7, i32 13, i32 15>117 %add = fadd <8 x float> %lhs, %rhs118 ret <8 x float> %add119}120 121define <8 x float> @hadd_reverse3_v8f32(<8 x float> %a0, <8 x float> %a1) {122; SSE-LABEL: hadd_reverse3_v8f32:123; SSE: # %bb.0:124; SSE-NEXT: haddps %xmm1, %xmm3125; SSE-NEXT: shufps {{.*#+}} xmm3 = xmm3[1,0,3,2]126; SSE-NEXT: haddps %xmm0, %xmm2127; SSE-NEXT: shufps {{.*#+}} xmm2 = xmm2[1,0,3,2]128; SSE-NEXT: movaps %xmm3, %xmm0129; SSE-NEXT: movaps %xmm2, %xmm1130; SSE-NEXT: retq131;132; AVX1-LABEL: hadd_reverse3_v8f32:133; AVX1: # %bb.0:134; AVX1-NEXT: vhaddps %ymm0, %ymm1, %ymm0135; AVX1-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3,0,1]136; AVX1-NEXT: vshufps {{.*#+}} ymm0 = ymm0[1,0,3,2,5,4,7,6]137; AVX1-NEXT: retq138;139; AVX2-LABEL: hadd_reverse3_v8f32:140; AVX2: # %bb.0:141; AVX2-NEXT: vhaddps %ymm0, %ymm1, %ymm0142; AVX2-NEXT: vshufps {{.*#+}} ymm0 = ymm0[1,0,3,2,5,4,7,6]143; AVX2-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[2,3,0,1]144; AVX2-NEXT: retq145 %shuf0 = shufflevector <8 x float> %a0, <8 x float> %a1, <8 x i32> <i32 0, i32 2, i32 8, i32 10, i32 4, i32 6, i32 12, i32 14>146 %shuf1 = shufflevector <8 x float> %a0, <8 x float> %a1, <8 x i32> <i32 1, i32 3, i32 9, i32 11, i32 5, i32 7, i32 13, i32 15>147 %add = fadd <8 x float> %shuf0, %shuf1148 %shuf2 = shufflevector <8 x float> %add, <8 x float> poison, <8 x i32> <i32 7, i32 6, i32 5, i32 4, i32 3, i32 2, i32 1, i32 0>149 ret <8 x float> %shuf2150}151 152define <16 x i16> @hadd_reverse_v16i16(<16 x i16> %a0, <16 x i16> %a1) nounwind {153; SSE-LABEL: hadd_reverse_v16i16:154; SSE: # %bb.0:155; SSE-NEXT: phaddw %xmm3, %xmm1156; SSE-NEXT: pshuflw {{.*#+}} xmm1 = xmm1[3,2,1,0,4,5,6,7]157; SSE-NEXT: pshufhw {{.*#+}} xmm3 = xmm1[0,1,2,3,7,6,5,4]158; SSE-NEXT: phaddw %xmm2, %xmm0159; SSE-NEXT: pshuflw {{.*#+}} xmm0 = xmm0[3,2,1,0,4,5,6,7]160; SSE-NEXT: pshufhw {{.*#+}} xmm1 = xmm0[0,1,2,3,7,6,5,4]161; SSE-NEXT: movdqa %xmm3, %xmm0162; SSE-NEXT: retq163;164; AVX1-LABEL: hadd_reverse_v16i16:165; AVX1: # %bb.0:166; AVX1-NEXT: vextractf128 $1, %ymm1, %xmm2167; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm3168; AVX1-NEXT: vphaddw %xmm2, %xmm3, %xmm2169; AVX1-NEXT: vpshuflw {{.*#+}} xmm2 = xmm2[3,2,1,0,4,5,6,7]170; AVX1-NEXT: vpshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,7,6,5,4]171; AVX1-NEXT: vphaddw %xmm1, %xmm0, %xmm0172; AVX1-NEXT: vpshuflw {{.*#+}} xmm0 = xmm0[3,2,1,0,4,5,6,7]173; AVX1-NEXT: vpshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,7,6,5,4]174; AVX1-NEXT: vinsertf128 $1, %xmm0, %ymm2, %ymm0175; AVX1-NEXT: retq176;177; AVX2-LABEL: hadd_reverse_v16i16:178; AVX2: # %bb.0:179; AVX2-NEXT: vphaddw %ymm1, %ymm0, %ymm0180; AVX2-NEXT: vpshuflw {{.*#+}} ymm0 = ymm0[3,2,1,0,4,5,6,7,11,10,9,8,12,13,14,15]181; AVX2-NEXT: vpshufhw {{.*#+}} ymm0 = ymm0[0,1,2,3,7,6,5,4,8,9,10,11,15,14,13,12]182; AVX2-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,3,0,1]183; AVX2-NEXT: retq184 %lhs = shufflevector <16 x i16> %a0, <16 x i16> %a1, <16 x i32> <i32 15, i32 13, i32 11, i32 9, i32 31, i32 29, i32 27, i32 25, i32 7, i32 5, i32 3, i32 1, i32 23, i32 21, i32 19, i32 17>185 %rhs = shufflevector <16 x i16> %a0, <16 x i16> %a1, <16 x i32> <i32 14, i32 12, i32 10, i32 8, i32 30, i32 28, i32 26, i32 24, i32 6, i32 4, i32 2, i32 0, i32 22, i32 20, i32 18, i32 16>186 %add = add <16 x i16> %lhs, %rhs187 ret <16 x i16> %add188}189 190define <16 x i16> @hadd_reverse2_v16i16(<16 x i16> %a0, <16 x i16> %a1) nounwind {191; SSE-LABEL: hadd_reverse2_v16i16:192; SSE: # %bb.0:193; SSE-NEXT: movdqa %xmm0, %xmm4194; SSE-NEXT: movdqa {{.*#+}} xmm0 = [14,15,12,13,10,11,8,9,6,7,4,5,2,3,0,1]195; SSE-NEXT: pshufb %xmm0, %xmm4196; SSE-NEXT: pshufb %xmm0, %xmm1197; SSE-NEXT: pshufb %xmm0, %xmm2198; SSE-NEXT: phaddw %xmm2, %xmm4199; SSE-NEXT: pshufb %xmm0, %xmm3200; SSE-NEXT: phaddw %xmm3, %xmm1201; SSE-NEXT: movdqa %xmm1, %xmm0202; SSE-NEXT: movdqa %xmm4, %xmm1203; SSE-NEXT: retq204;205; AVX1-LABEL: hadd_reverse2_v16i16:206; AVX1: # %bb.0:207; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm2208; AVX1-NEXT: vmovdqa {{.*#+}} xmm3 = [14,15,12,13,10,11,8,9,6,7,4,5,2,3,0,1]209; AVX1-NEXT: vpshufb %xmm3, %xmm2, %xmm2210; AVX1-NEXT: vpshufb %xmm3, %xmm0, %xmm0211; AVX1-NEXT: vextractf128 $1, %ymm1, %xmm4212; AVX1-NEXT: vpshufb %xmm3, %xmm4, %xmm4213; AVX1-NEXT: vphaddw %xmm4, %xmm2, %xmm2214; AVX1-NEXT: vpshufb %xmm3, %xmm1, %xmm1215; AVX1-NEXT: vphaddw %xmm1, %xmm0, %xmm0216; AVX1-NEXT: vinsertf128 $1, %xmm0, %ymm2, %ymm0217; AVX1-NEXT: retq218;219; AVX2-LABEL: hadd_reverse2_v16i16:220; AVX2: # %bb.0:221; AVX2-NEXT: vmovdqa {{.*#+}} ymm2 = [14,15,12,13,10,11,8,9,6,7,4,5,2,3,0,1,30,31,28,29,26,27,24,25,22,23,20,21,18,19,16,17]222; AVX2-NEXT: vpshufb %ymm2, %ymm0, %ymm0223; AVX2-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,3,0,1]224; AVX2-NEXT: vpshufb %ymm2, %ymm1, %ymm1225; AVX2-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,3,0,1]226; AVX2-NEXT: vphaddw %ymm1, %ymm0, %ymm0227; AVX2-NEXT: retq228 %shuf0 = shufflevector <16 x i16> %a0, <16 x i16> undef, <16 x i32> <i32 15, i32 14, i32 13, i32 12, i32 11, i32 10, i32 9, i32 8, i32 7, i32 6, i32 5, i32 4, i32 3, i32 2, i32 1, i32 0>229 %shuf1 = shufflevector <16 x i16> %a1, <16 x i16> undef, <16 x i32> <i32 15, i32 14, i32 13, i32 12, i32 11, i32 10, i32 9, i32 8, i32 7, i32 6, i32 5, i32 4, i32 3, i32 2, i32 1, i32 0>230 %lhs = shufflevector <16 x i16> %shuf0, <16 x i16> %shuf1, <16 x i32> <i32 0, i32 2, i32 4, i32 6, i32 16, i32 18, i32 20, i32 22, i32 8, i32 10, i32 12, i32 14, i32 24, i32 26, i32 28, i32 30>231 %rhs = shufflevector <16 x i16> %shuf0, <16 x i16> %shuf1, <16 x i32> <i32 1, i32 3, i32 5, i32 7, i32 17, i32 19, i32 21, i32 23, i32 9, i32 11, i32 13, i32 15, i32 25, i32 27, i32 29, i32 31>232 %add = add <16 x i16> %lhs, %rhs233 ret <16 x i16> %add234}235 236define <8 x double> @hadd_reverse_v8f64(<8 x double> %a0, <8 x double> %a1) nounwind {237; SSE-LABEL: hadd_reverse_v8f64:238; SSE: # %bb.0:239; SSE-NEXT: movapd %xmm1, %xmm8240; SSE-NEXT: movapd %xmm0, %xmm9241; SSE-NEXT: haddpd %xmm7, %xmm3242; SSE-NEXT: haddpd %xmm6, %xmm2243; SSE-NEXT: haddpd %xmm5, %xmm8244; SSE-NEXT: haddpd %xmm4, %xmm9245; SSE-NEXT: movapd %xmm3, %xmm0246; SSE-NEXT: movapd %xmm2, %xmm1247; SSE-NEXT: movapd %xmm8, %xmm2248; SSE-NEXT: movapd %xmm9, %xmm3249; SSE-NEXT: retq250;251; AVX1-LABEL: hadd_reverse_v8f64:252; AVX1: # %bb.0:253; AVX1-NEXT: vhaddpd %ymm3, %ymm1, %ymm1254; AVX1-NEXT: vperm2f128 {{.*#+}} ymm3 = ymm1[2,3,0,1]255; AVX1-NEXT: vhaddpd %ymm2, %ymm0, %ymm0256; AVX1-NEXT: vperm2f128 {{.*#+}} ymm1 = ymm0[2,3,0,1]257; AVX1-NEXT: vmovapd %ymm3, %ymm0258; AVX1-NEXT: retq259;260; AVX2-LABEL: hadd_reverse_v8f64:261; AVX2: # %bb.0:262; AVX2-NEXT: vhaddpd %ymm3, %ymm1, %ymm1263; AVX2-NEXT: vpermpd {{.*#+}} ymm3 = ymm1[2,3,0,1]264; AVX2-NEXT: vhaddpd %ymm2, %ymm0, %ymm0265; AVX2-NEXT: vpermpd {{.*#+}} ymm1 = ymm0[2,3,0,1]266; AVX2-NEXT: vmovapd %ymm3, %ymm0267; AVX2-NEXT: retq268 %lhs = shufflevector <8 x double> %a0, <8 x double> %a1, <8 x i32> <i32 7, i32 15, i32 5, i32 13, i32 3, i32 11, i32 1, i32 9>269 %rhs = shufflevector <8 x double> %a0, <8 x double> %a1, <8 x i32> <i32 6, i32 14, i32 4, i32 12, i32 2, i32 10, i32 0, i32 8>270 %fadd = fadd <8 x double> %lhs, %rhs271 ret <8 x double> %fadd272}273 274define <8 x double> @hadd_reverse2_v8f64(<8 x double> %a0, <8 x double> %a1) nounwind {275; SSE-LABEL: hadd_reverse2_v8f64:276; SSE: # %bb.0:277; SSE-NEXT: movapd %xmm1, %xmm8278; SSE-NEXT: movapd %xmm0, %xmm9279; SSE-NEXT: haddpd %xmm7, %xmm3280; SSE-NEXT: haddpd %xmm6, %xmm2281; SSE-NEXT: haddpd %xmm5, %xmm8282; SSE-NEXT: haddpd %xmm4, %xmm9283; SSE-NEXT: movapd %xmm3, %xmm0284; SSE-NEXT: movapd %xmm2, %xmm1285; SSE-NEXT: movapd %xmm8, %xmm2286; SSE-NEXT: movapd %xmm9, %xmm3287; SSE-NEXT: retq288;289; AVX1-LABEL: hadd_reverse2_v8f64:290; AVX1: # %bb.0:291; AVX1-NEXT: vhaddpd %ymm3, %ymm1, %ymm1292; AVX1-NEXT: vperm2f128 {{.*#+}} ymm3 = ymm1[2,3,0,1]293; AVX1-NEXT: vhaddpd %ymm2, %ymm0, %ymm0294; AVX1-NEXT: vperm2f128 {{.*#+}} ymm1 = ymm0[2,3,0,1]295; AVX1-NEXT: vmovapd %ymm3, %ymm0296; AVX1-NEXT: retq297;298; AVX2-LABEL: hadd_reverse2_v8f64:299; AVX2: # %bb.0:300; AVX2-NEXT: vhaddpd %ymm3, %ymm1, %ymm1301; AVX2-NEXT: vpermpd {{.*#+}} ymm3 = ymm1[2,3,0,1]302; AVX2-NEXT: vhaddpd %ymm2, %ymm0, %ymm0303; AVX2-NEXT: vpermpd {{.*#+}} ymm1 = ymm0[2,3,0,1]304; AVX2-NEXT: vmovapd %ymm3, %ymm0305; AVX2-NEXT: retq306 %shuf0 = shufflevector <8 x double> %a0, <8 x double> undef, <8 x i32> <i32 7, i32 6, i32 5, i32 4, i32 3, i32 2, i32 1, i32 0>307 %shuf1 = shufflevector <8 x double> %a1, <8 x double> undef, <8 x i32> <i32 7, i32 6, i32 5, i32 4, i32 3, i32 2, i32 1, i32 0>308 %lhs = shufflevector <8 x double> %shuf0, <8 x double> %shuf1, <8 x i32> <i32 0, i32 8, i32 2, i32 10, i32 4, i32 12, i32 6, i32 14>309 %rhs = shufflevector <8 x double> %shuf0, <8 x double> %shuf1, <8 x i32> <i32 1, i32 9, i32 3, i32 11, i32 5, i32 13, i32 7, i32 15>310 %fadd = fadd <8 x double> %lhs, %rhs311 ret <8 x double> %fadd312}313 314define <16 x float> @hadd_reverse_v16f32(<16 x float> %a0, <16 x float> %a1) nounwind {315; SSE-LABEL: hadd_reverse_v16f32:316; SSE: # %bb.0:317; SSE-NEXT: movaps %xmm0, %xmm8318; SSE-NEXT: haddps %xmm3, %xmm2319; SSE-NEXT: shufps {{.*#+}} xmm2 = xmm2[3,2,1,0]320; SSE-NEXT: haddps %xmm7, %xmm6321; SSE-NEXT: shufps {{.*#+}} xmm6 = xmm6[3,2,1,0]322; SSE-NEXT: haddps %xmm1, %xmm8323; SSE-NEXT: shufps {{.*#+}} xmm8 = xmm8[3,2,1,0]324; SSE-NEXT: haddps %xmm5, %xmm4325; SSE-NEXT: shufps {{.*#+}} xmm4 = xmm4[3,2,1,0]326; SSE-NEXT: movaps %xmm2, %xmm0327; SSE-NEXT: movaps %xmm6, %xmm1328; SSE-NEXT: movaps %xmm8, %xmm2329; SSE-NEXT: movaps %xmm4, %xmm3330; SSE-NEXT: retq331;332; AVX1-LABEL: hadd_reverse_v16f32:333; AVX1: # %bb.0:334; AVX1-NEXT: vperm2f128 {{.*#+}} ymm4 = ymm0[2,3],ymm2[2,3]335; AVX1-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm0336; AVX1-NEXT: vhaddps %ymm0, %ymm4, %ymm2337; AVX1-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm3[2,3]338; AVX1-NEXT: vinsertf128 $1, %xmm3, %ymm1, %ymm1339; AVX1-NEXT: vhaddps %ymm1, %ymm0, %ymm0340; AVX1-NEXT: vshufps {{.*#+}} ymm0 = ymm0[1,0,3,2,5,4,7,6]341; AVX1-NEXT: vshufps {{.*#+}} ymm1 = ymm2[1,0,3,2,5,4,7,6]342; AVX1-NEXT: retq343;344; AVX2-LABEL: hadd_reverse_v16f32:345; AVX2: # %bb.0:346; AVX2-NEXT: vhaddps %ymm3, %ymm1, %ymm1347; AVX2-NEXT: vshufps {{.*#+}} ymm1 = ymm1[1,0,3,2,5,4,7,6]348; AVX2-NEXT: vpermpd {{.*#+}} ymm3 = ymm1[2,0,3,1]349; AVX2-NEXT: vhaddps %ymm2, %ymm0, %ymm0350; AVX2-NEXT: vshufps {{.*#+}} ymm0 = ymm0[1,0,3,2,5,4,7,6]351; AVX2-NEXT: vpermpd {{.*#+}} ymm1 = ymm0[2,0,3,1]352; AVX2-NEXT: vmovaps %ymm3, %ymm0353; AVX2-NEXT: retq354 %lhs = shufflevector <16 x float> %a0, <16 x float> %a1, <16 x i32> <i32 15, i32 13, i32 11, i32 9, i32 31, i32 29, i32 27, i32 25, i32 7, i32 5, i32 3, i32 1, i32 23, i32 21, i32 19, i32 17>355 %rhs = shufflevector <16 x float> %a0, <16 x float> %a1, <16 x i32> <i32 14, i32 12, i32 10, i32 8, i32 30, i32 28, i32 26, i32 24, i32 6, i32 4, i32 2, i32 0, i32 22, i32 20, i32 18, i32 16>356 %fadd = fadd <16 x float> %lhs, %rhs357 ret <16 x float> %fadd358}359 360define <16 x float> @hadd_reverse2_v16f32(<16 x float> %a0, <16 x float> %a1) nounwind {361; SSE-LABEL: hadd_reverse2_v16f32:362; SSE: # %bb.0:363; SSE-NEXT: movaps %xmm1, %xmm8364; SSE-NEXT: movaps %xmm0, %xmm9365; SSE-NEXT: haddps %xmm7, %xmm3366; SSE-NEXT: shufps {{.*#+}} xmm3 = xmm3[1,0,3,2]367; SSE-NEXT: haddps %xmm6, %xmm2368; SSE-NEXT: shufps {{.*#+}} xmm2 = xmm2[1,0,3,2]369; SSE-NEXT: haddps %xmm5, %xmm8370; SSE-NEXT: shufps {{.*#+}} xmm8 = xmm8[1,0,3,2]371; SSE-NEXT: haddps %xmm4, %xmm9372; SSE-NEXT: shufps {{.*#+}} xmm9 = xmm9[1,0,3,2]373; SSE-NEXT: movaps %xmm3, %xmm0374; SSE-NEXT: movaps %xmm2, %xmm1375; SSE-NEXT: movaps %xmm8, %xmm2376; SSE-NEXT: movaps %xmm9, %xmm3377; SSE-NEXT: retq378;379; AVX1-LABEL: hadd_reverse2_v16f32:380; AVX1: # %bb.0:381; AVX1-NEXT: vhaddps %ymm3, %ymm1, %ymm1382; AVX1-NEXT: vperm2f128 {{.*#+}} ymm1 = ymm1[2,3,0,1]383; AVX1-NEXT: vshufps {{.*#+}} ymm3 = ymm1[1,0,3,2,5,4,7,6]384; AVX1-NEXT: vhaddps %ymm2, %ymm0, %ymm0385; AVX1-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3,0,1]386; AVX1-NEXT: vshufps {{.*#+}} ymm1 = ymm0[1,0,3,2,5,4,7,6]387; AVX1-NEXT: vmovaps %ymm3, %ymm0388; AVX1-NEXT: retq389;390; AVX2-LABEL: hadd_reverse2_v16f32:391; AVX2: # %bb.0:392; AVX2-NEXT: vhaddps %ymm3, %ymm1, %ymm1393; AVX2-NEXT: vshufps {{.*#+}} ymm1 = ymm1[1,0,3,2,5,4,7,6]394; AVX2-NEXT: vpermpd {{.*#+}} ymm3 = ymm1[2,3,0,1]395; AVX2-NEXT: vhaddps %ymm2, %ymm0, %ymm0396; AVX2-NEXT: vshufps {{.*#+}} ymm0 = ymm0[1,0,3,2,5,4,7,6]397; AVX2-NEXT: vpermpd {{.*#+}} ymm1 = ymm0[2,3,0,1]398; AVX2-NEXT: vmovaps %ymm3, %ymm0399; AVX2-NEXT: retq400 %shuf0 = shufflevector <16 x float> %a0, <16 x float> undef, <16 x i32> <i32 15, i32 14, i32 13, i32 12, i32 11, i32 10, i32 9, i32 8, i32 7, i32 6, i32 5, i32 4, i32 3, i32 2, i32 1, i32 0>401 %shuf1 = shufflevector <16 x float> %a1, <16 x float> undef, <16 x i32> <i32 15, i32 14, i32 13, i32 12, i32 11, i32 10, i32 9, i32 8, i32 7, i32 6, i32 5, i32 4, i32 3, i32 2, i32 1, i32 0>402 %lhs = shufflevector <16 x float> %shuf0, <16 x float> %shuf1, <16 x i32> <i32 0, i32 2, i32 16, i32 18, i32 4, i32 6, i32 20, i32 22, i32 8, i32 10, i32 24, i32 26, i32 12, i32 14, i32 28, i32 30>403 %rhs = shufflevector <16 x float> %shuf0, <16 x float> %shuf1, <16 x i32> <i32 1, i32 3, i32 17, i32 19, i32 5, i32 7, i32 21, i32 23, i32 9, i32 11, i32 25, i32 27, i32 13, i32 15, i32 29, i32 31>404 %fadd = fadd <16 x float> %lhs, %rhs405 ret <16 x float> %fadd406}407