535 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py2; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+sse2 | FileCheck %s --check-prefix=SSE23; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+sse4.1 | FileCheck %s --check-prefix=SSE414; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx | FileCheck %s --check-prefixes=AVX,AVX15; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx2 | FileCheck %s --check-prefixes=AVX,AVX26 7; If we have a cmp and a sel with different-sized operands followed by a size-changing cast,8; we may want to pull the cast ahead of the select operands to create a select with matching op sizes:9; ext (sel (cmp a, b), c, d) --> sel (cmp a, b), (ext c), (ext d)10 11define <8 x i32> @sext(<8 x float> %a, <8 x float> %b, <8 x i16> %c, <8 x i16> %d) {12; SSE2-LABEL: sext:13; SSE2: # %bb.0:14; SSE2-NEXT: cmpltps %xmm3, %xmm115; SSE2-NEXT: cmpltps %xmm2, %xmm016; SSE2-NEXT: packssdw %xmm1, %xmm017; SSE2-NEXT: pand %xmm0, %xmm418; SSE2-NEXT: pandn %xmm5, %xmm019; SSE2-NEXT: por %xmm4, %xmm020; SSE2-NEXT: punpcklwd {{.*#+}} xmm2 = xmm2[0],xmm0[0],xmm2[1],xmm0[1],xmm2[2],xmm0[2],xmm2[3],xmm0[3]21; SSE2-NEXT: psrad $16, %xmm222; SSE2-NEXT: punpckhwd {{.*#+}} xmm1 = xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]23; SSE2-NEXT: psrad $16, %xmm124; SSE2-NEXT: movdqa %xmm2, %xmm025; SSE2-NEXT: retq26;27; SSE41-LABEL: sext:28; SSE41: # %bb.0:29; SSE41-NEXT: cmpltps %xmm3, %xmm130; SSE41-NEXT: cmpltps %xmm2, %xmm031; SSE41-NEXT: packssdw %xmm1, %xmm032; SSE41-NEXT: pblendvb %xmm0, %xmm4, %xmm533; SSE41-NEXT: pmovsxwd %xmm5, %xmm034; SSE41-NEXT: pshufd {{.*#+}} xmm1 = xmm5[2,3,2,3]35; SSE41-NEXT: pmovsxwd %xmm1, %xmm136; SSE41-NEXT: retq37;38; AVX1-LABEL: sext:39; AVX1: # %bb.0:40; AVX1-NEXT: vcmpltps %ymm1, %ymm0, %ymm041; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm142; AVX1-NEXT: vpshufd {{.*#+}} xmm4 = xmm2[2,3,2,3]43; AVX1-NEXT: vpmovsxwd %xmm4, %xmm444; AVX1-NEXT: vpshufd {{.*#+}} xmm5 = xmm3[2,3,2,3]45; AVX1-NEXT: vpmovsxwd %xmm5, %xmm546; AVX1-NEXT: vblendvps %xmm1, %xmm4, %xmm5, %xmm147; AVX1-NEXT: vpmovsxwd %xmm2, %xmm248; AVX1-NEXT: vpmovsxwd %xmm3, %xmm349; AVX1-NEXT: vblendvps %xmm0, %xmm2, %xmm3, %xmm050; AVX1-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm051; AVX1-NEXT: retq52;53; AVX2-LABEL: sext:54; AVX2: # %bb.0:55; AVX2-NEXT: vcmpltps %ymm1, %ymm0, %ymm056; AVX2-NEXT: vpmovsxwd %xmm2, %ymm157; AVX2-NEXT: vpmovsxwd %xmm3, %ymm258; AVX2-NEXT: vblendvps %ymm0, %ymm1, %ymm2, %ymm059; AVX2-NEXT: retq60 %cmp = fcmp olt <8 x float> %a, %b61 %sel = select <8 x i1> %cmp, <8 x i16> %c, <8 x i16> %d62 %ext = sext <8 x i16> %sel to <8 x i32>63 ret <8 x i32> %ext64}65 66define <8 x i32> @zext(<8 x float> %a, <8 x float> %b, <8 x i16> %c, <8 x i16> %d) {67; SSE2-LABEL: zext:68; SSE2: # %bb.0:69; SSE2-NEXT: movaps %xmm0, %xmm670; SSE2-NEXT: cmpltps %xmm3, %xmm171; SSE2-NEXT: cmpltps %xmm2, %xmm672; SSE2-NEXT: packssdw %xmm1, %xmm673; SSE2-NEXT: pand %xmm6, %xmm474; SSE2-NEXT: pandn %xmm5, %xmm675; SSE2-NEXT: por %xmm4, %xmm676; SSE2-NEXT: pxor %xmm1, %xmm177; SSE2-NEXT: movdqa %xmm6, %xmm078; SSE2-NEXT: punpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]79; SSE2-NEXT: punpckhwd {{.*#+}} xmm6 = xmm6[4],xmm1[4],xmm6[5],xmm1[5],xmm6[6],xmm1[6],xmm6[7],xmm1[7]80; SSE2-NEXT: movdqa %xmm6, %xmm181; SSE2-NEXT: retq82;83; SSE41-LABEL: zext:84; SSE41: # %bb.0:85; SSE41-NEXT: cmpltps %xmm3, %xmm186; SSE41-NEXT: cmpltps %xmm2, %xmm087; SSE41-NEXT: packssdw %xmm1, %xmm088; SSE41-NEXT: pblendvb %xmm0, %xmm4, %xmm589; SSE41-NEXT: pxor %xmm1, %xmm190; SSE41-NEXT: pmovzxwd {{.*#+}} xmm0 = xmm5[0],zero,xmm5[1],zero,xmm5[2],zero,xmm5[3],zero91; SSE41-NEXT: punpckhwd {{.*#+}} xmm5 = xmm5[4],xmm1[4],xmm5[5],xmm1[5],xmm5[6],xmm1[6],xmm5[7],xmm1[7]92; SSE41-NEXT: movdqa %xmm5, %xmm193; SSE41-NEXT: retq94;95; AVX1-LABEL: zext:96; AVX1: # %bb.0:97; AVX1-NEXT: vcmpltps %ymm1, %ymm0, %ymm098; AVX1-NEXT: vxorps %xmm1, %xmm1, %xmm199; AVX1-NEXT: vpunpckhwd {{.*#+}} xmm4 = xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]100; AVX1-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm3[4],xmm1[4],xmm3[5],xmm1[5],xmm3[6],xmm1[6],xmm3[7],xmm1[7]101; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm5102; AVX1-NEXT: vblendvps %xmm5, %xmm4, %xmm1, %xmm1103; AVX1-NEXT: vpmovzxwd {{.*#+}} xmm2 = xmm2[0],zero,xmm2[1],zero,xmm2[2],zero,xmm2[3],zero104; AVX1-NEXT: vpmovzxwd {{.*#+}} xmm3 = xmm3[0],zero,xmm3[1],zero,xmm3[2],zero,xmm3[3],zero105; AVX1-NEXT: vblendvps %xmm0, %xmm2, %xmm3, %xmm0106; AVX1-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0107; AVX1-NEXT: retq108;109; AVX2-LABEL: zext:110; AVX2: # %bb.0:111; AVX2-NEXT: vcmpltps %ymm1, %ymm0, %ymm0112; AVX2-NEXT: vpmovzxwd {{.*#+}} ymm1 = xmm2[0],zero,xmm2[1],zero,xmm2[2],zero,xmm2[3],zero,xmm2[4],zero,xmm2[5],zero,xmm2[6],zero,xmm2[7],zero113; AVX2-NEXT: vpmovzxwd {{.*#+}} ymm2 = xmm3[0],zero,xmm3[1],zero,xmm3[2],zero,xmm3[3],zero,xmm3[4],zero,xmm3[5],zero,xmm3[6],zero,xmm3[7],zero114; AVX2-NEXT: vblendvps %ymm0, %ymm1, %ymm2, %ymm0115; AVX2-NEXT: retq116 %cmp = fcmp olt <8 x float> %a, %b117 %sel = select <8 x i1> %cmp, <8 x i16> %c, <8 x i16> %d118 %ext = zext <8 x i16> %sel to <8 x i32>119 ret <8 x i32> %ext120}121 122define <4 x double> @fpext(<4 x double> %a, <4 x double> %b, <4 x float> %c, <4 x float> %d) {123; SSE2-LABEL: fpext:124; SSE2: # %bb.0:125; SSE2-NEXT: cmpltpd %xmm3, %xmm1126; SSE2-NEXT: cmpltpd %xmm2, %xmm0127; SSE2-NEXT: shufps {{.*#+}} xmm0 = xmm0[0,2],xmm1[0,2]128; SSE2-NEXT: andps %xmm0, %xmm4129; SSE2-NEXT: andnps %xmm5, %xmm0130; SSE2-NEXT: orps %xmm4, %xmm0131; SSE2-NEXT: cvtps2pd %xmm0, %xmm2132; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm0[2,3,2,3]133; SSE2-NEXT: cvtps2pd %xmm0, %xmm1134; SSE2-NEXT: movaps %xmm2, %xmm0135; SSE2-NEXT: retq136;137; SSE41-LABEL: fpext:138; SSE41: # %bb.0:139; SSE41-NEXT: cmpltpd %xmm3, %xmm1140; SSE41-NEXT: cmpltpd %xmm2, %xmm0141; SSE41-NEXT: shufps {{.*#+}} xmm0 = xmm0[0,2],xmm1[0,2]142; SSE41-NEXT: blendvps %xmm0, %xmm4, %xmm5143; SSE41-NEXT: cvtps2pd %xmm5, %xmm0144; SSE41-NEXT: movhlps {{.*#+}} xmm5 = xmm5[1,1]145; SSE41-NEXT: cvtps2pd %xmm5, %xmm1146; SSE41-NEXT: retq147;148; AVX-LABEL: fpext:149; AVX: # %bb.0:150; AVX-NEXT: vcmpltpd %ymm1, %ymm0, %ymm0151; AVX-NEXT: vcvtps2pd %xmm2, %ymm1152; AVX-NEXT: vcvtps2pd %xmm3, %ymm2153; AVX-NEXT: vblendvpd %ymm0, %ymm1, %ymm2, %ymm0154; AVX-NEXT: retq155 %cmp = fcmp olt <4 x double> %a, %b156 %sel = select <4 x i1> %cmp, <4 x float> %c, <4 x float> %d157 %ext = fpext <4 x float> %sel to <4 x double>158 ret <4 x double> %ext159}160 161define <8 x i16> @trunc(<8 x i16> %a, <8 x i16> %b, <8 x i32> %c, <8 x i32> %d) {162; SSE2-LABEL: trunc:163; SSE2: # %bb.0:164; SSE2-NEXT: pcmpeqw %xmm1, %xmm0165; SSE2-NEXT: pslld $16, %xmm5166; SSE2-NEXT: psrad $16, %xmm5167; SSE2-NEXT: pslld $16, %xmm4168; SSE2-NEXT: psrad $16, %xmm4169; SSE2-NEXT: packssdw %xmm5, %xmm4170; SSE2-NEXT: pslld $16, %xmm3171; SSE2-NEXT: psrad $16, %xmm3172; SSE2-NEXT: pslld $16, %xmm2173; SSE2-NEXT: psrad $16, %xmm2174; SSE2-NEXT: packssdw %xmm3, %xmm2175; SSE2-NEXT: pand %xmm0, %xmm2176; SSE2-NEXT: pandn %xmm4, %xmm0177; SSE2-NEXT: por %xmm2, %xmm0178; SSE2-NEXT: retq179;180; SSE41-LABEL: trunc:181; SSE41: # %bb.0:182; SSE41-NEXT: pcmpeqw %xmm1, %xmm0183; SSE41-NEXT: pxor %xmm1, %xmm1184; SSE41-NEXT: pblendw {{.*#+}} xmm3 = xmm3[0],xmm1[1],xmm3[2],xmm1[3],xmm3[4],xmm1[5],xmm3[6],xmm1[7]185; SSE41-NEXT: pblendw {{.*#+}} xmm2 = xmm2[0],xmm1[1],xmm2[2],xmm1[3],xmm2[4],xmm1[5],xmm2[6],xmm1[7]186; SSE41-NEXT: packusdw %xmm3, %xmm2187; SSE41-NEXT: pblendw {{.*#+}} xmm5 = xmm5[0],xmm1[1],xmm5[2],xmm1[3],xmm5[4],xmm1[5],xmm5[6],xmm1[7]188; SSE41-NEXT: pblendw {{.*#+}} xmm1 = xmm4[0],xmm1[1],xmm4[2],xmm1[3],xmm4[4],xmm1[5],xmm4[6],xmm1[7]189; SSE41-NEXT: packusdw %xmm5, %xmm1190; SSE41-NEXT: pblendvb %xmm0, %xmm2, %xmm1191; SSE41-NEXT: movdqa %xmm1, %xmm0192; SSE41-NEXT: retq193;194; AVX1-LABEL: trunc:195; AVX1: # %bb.0:196; AVX1-NEXT: vpcmpeqw %xmm1, %xmm0, %xmm0197; AVX1-NEXT: vbroadcastss {{.*#+}} ymm1 = [65535,65535,65535,65535,65535,65535,65535,65535]198; AVX1-NEXT: vandps %ymm1, %ymm2, %ymm2199; AVX1-NEXT: vextractf128 $1, %ymm2, %xmm4200; AVX1-NEXT: vpackusdw %xmm4, %xmm2, %xmm2201; AVX1-NEXT: vandps %ymm1, %ymm3, %ymm1202; AVX1-NEXT: vextractf128 $1, %ymm1, %xmm3203; AVX1-NEXT: vpackusdw %xmm3, %xmm1, %xmm1204; AVX1-NEXT: vpblendvb %xmm0, %xmm2, %xmm1, %xmm0205; AVX1-NEXT: vzeroupper206; AVX1-NEXT: retq207;208; AVX2-LABEL: trunc:209; AVX2: # %bb.0:210; AVX2-NEXT: vpcmpeqw %xmm1, %xmm0, %xmm0211; AVX2-NEXT: vmovdqa {{.*#+}} ymm1 = [0,1,4,5,8,9,12,13,8,9,12,13,12,13,14,15,16,17,20,21,24,25,28,29,24,25,28,29,28,29,30,31]212; AVX2-NEXT: vpshufb %ymm1, %ymm2, %ymm2213; AVX2-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,2,2,3]214; AVX2-NEXT: vpshufb %ymm1, %ymm3, %ymm1215; AVX2-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,2,2,3]216; AVX2-NEXT: vpblendvb %xmm0, %xmm2, %xmm1, %xmm0217; AVX2-NEXT: vzeroupper218; AVX2-NEXT: retq219 %cmp = icmp eq <8 x i16> %a, %b220 %sel = select <8 x i1> %cmp, <8 x i32> %c, <8 x i32> %d221 %tr = trunc <8 x i32> %sel to <8 x i16>222 ret <8 x i16> %tr223}224 225define <4 x float> @fptrunc(<4 x float> %a, <4 x float> %b, <4 x double> %c, <4 x double> %d) {226; SSE2-LABEL: fptrunc:227; SSE2: # %bb.0:228; SSE2-NEXT: cmpltps %xmm1, %xmm0229; SSE2-NEXT: cvtpd2ps %xmm5, %xmm1230; SSE2-NEXT: cvtpd2ps %xmm4, %xmm4231; SSE2-NEXT: unpcklpd {{.*#+}} xmm4 = xmm4[0],xmm1[0]232; SSE2-NEXT: cvtpd2ps %xmm3, %xmm1233; SSE2-NEXT: cvtpd2ps %xmm2, %xmm2234; SSE2-NEXT: unpcklpd {{.*#+}} xmm2 = xmm2[0],xmm1[0]235; SSE2-NEXT: andpd %xmm0, %xmm2236; SSE2-NEXT: andnpd %xmm4, %xmm0237; SSE2-NEXT: orpd %xmm2, %xmm0238; SSE2-NEXT: retq239;240; SSE41-LABEL: fptrunc:241; SSE41: # %bb.0:242; SSE41-NEXT: cmpltps %xmm1, %xmm0243; SSE41-NEXT: cvtpd2ps %xmm3, %xmm1244; SSE41-NEXT: cvtpd2ps %xmm2, %xmm2245; SSE41-NEXT: unpcklpd {{.*#+}} xmm2 = xmm2[0],xmm1[0]246; SSE41-NEXT: cvtpd2ps %xmm5, %xmm3247; SSE41-NEXT: cvtpd2ps %xmm4, %xmm1248; SSE41-NEXT: unpcklpd {{.*#+}} xmm1 = xmm1[0],xmm3[0]249; SSE41-NEXT: blendvps %xmm0, %xmm2, %xmm1250; SSE41-NEXT: movaps %xmm1, %xmm0251; SSE41-NEXT: retq252;253; AVX-LABEL: fptrunc:254; AVX: # %bb.0:255; AVX-NEXT: vcmpltps %xmm1, %xmm0, %xmm0256; AVX-NEXT: vcvtpd2ps %ymm2, %xmm1257; AVX-NEXT: vcvtpd2ps %ymm3, %xmm2258; AVX-NEXT: vblendvps %xmm0, %xmm1, %xmm2, %xmm0259; AVX-NEXT: vzeroupper260; AVX-NEXT: retq261 %cmp = fcmp olt <4 x float> %a, %b262 %sel = select <4 x i1> %cmp, <4 x double> %c, <4 x double> %d263 %tr = fptrunc <4 x double> %sel to <4 x float>264 ret <4 x float> %tr265}266 267; PR14657 - avoid truncation/extension of comparison results268; These tests demonstrate the same issue as the simpler cases above,269; but also include multi-BB to show potentially larger transforms/codegen issues.270 271@da = common dso_local global [1024 x float] zeroinitializer, align 32272@db = common dso_local global [1024 x float] zeroinitializer, align 32273@dc = common dso_local global [1024 x float] zeroinitializer, align 32274@dd = common dso_local global [1024 x float] zeroinitializer, align 32275@dj = common dso_local global [1024 x i32] zeroinitializer, align 32276 277define dso_local void @example25() nounwind {278; SSE2-LABEL: example25:279; SSE2: # %bb.0: # %vector.ph280; SSE2-NEXT: movq $-4096, %rax # imm = 0xF000281; SSE2-NEXT: .p2align 4282; SSE2-NEXT: .LBB5_1: # %vector.body283; SSE2-NEXT: # =>This Inner Loop Header: Depth=1284; SSE2-NEXT: movaps da+4112(%rax), %xmm0285; SSE2-NEXT: movaps da+4096(%rax), %xmm1286; SSE2-NEXT: cmpltps db+4096(%rax), %xmm1287; SSE2-NEXT: cmpltps db+4112(%rax), %xmm0288; SSE2-NEXT: movaps dc+4112(%rax), %xmm2289; SSE2-NEXT: movaps dc+4096(%rax), %xmm3290; SSE2-NEXT: cmpltps dd+4096(%rax), %xmm3291; SSE2-NEXT: andps %xmm1, %xmm3292; SSE2-NEXT: cmpltps dd+4112(%rax), %xmm2293; SSE2-NEXT: andps %xmm0, %xmm2294; SSE2-NEXT: psrld $31, %xmm3295; SSE2-NEXT: psrld $31, %xmm2296; SSE2-NEXT: movdqa %xmm2, dj+4112(%rax)297; SSE2-NEXT: movdqa %xmm3, dj+4096(%rax)298; SSE2-NEXT: addq $32, %rax299; SSE2-NEXT: jne .LBB5_1300; SSE2-NEXT: # %bb.2: # %for.end301; SSE2-NEXT: retq302;303; SSE41-LABEL: example25:304; SSE41: # %bb.0: # %vector.ph305; SSE41-NEXT: movq $-4096, %rax # imm = 0xF000306; SSE41-NEXT: .p2align 4307; SSE41-NEXT: .LBB5_1: # %vector.body308; SSE41-NEXT: # =>This Inner Loop Header: Depth=1309; SSE41-NEXT: movaps da+4112(%rax), %xmm0310; SSE41-NEXT: movaps da+4096(%rax), %xmm1311; SSE41-NEXT: cmpltps db+4096(%rax), %xmm1312; SSE41-NEXT: cmpltps db+4112(%rax), %xmm0313; SSE41-NEXT: movaps dc+4112(%rax), %xmm2314; SSE41-NEXT: movaps dc+4096(%rax), %xmm3315; SSE41-NEXT: cmpltps dd+4096(%rax), %xmm3316; SSE41-NEXT: andps %xmm1, %xmm3317; SSE41-NEXT: cmpltps dd+4112(%rax), %xmm2318; SSE41-NEXT: andps %xmm0, %xmm2319; SSE41-NEXT: psrld $31, %xmm3320; SSE41-NEXT: psrld $31, %xmm2321; SSE41-NEXT: movdqa %xmm2, dj+4112(%rax)322; SSE41-NEXT: movdqa %xmm3, dj+4096(%rax)323; SSE41-NEXT: addq $32, %rax324; SSE41-NEXT: jne .LBB5_1325; SSE41-NEXT: # %bb.2: # %for.end326; SSE41-NEXT: retq327;328; AVX1-LABEL: example25:329; AVX1: # %bb.0: # %vector.ph330; AVX1-NEXT: movq $-4096, %rax # imm = 0xF000331; AVX1-NEXT: vbroadcastss {{.*#+}} ymm0 = [1,1,1,1,1,1,1,1]332; AVX1-NEXT: .p2align 4333; AVX1-NEXT: .LBB5_1: # %vector.body334; AVX1-NEXT: # =>This Inner Loop Header: Depth=1335; AVX1-NEXT: vmovups da+4096(%rax), %ymm1336; AVX1-NEXT: vcmpltps db+4096(%rax), %ymm1, %ymm1337; AVX1-NEXT: vmovups dc+4096(%rax), %ymm2338; AVX1-NEXT: vcmpltps dd+4096(%rax), %ymm2, %ymm2339; AVX1-NEXT: vandps %ymm2, %ymm1, %ymm1340; AVX1-NEXT: vandps %ymm0, %ymm1, %ymm1341; AVX1-NEXT: vmovups %ymm1, dj+4096(%rax)342; AVX1-NEXT: addq $32, %rax343; AVX1-NEXT: jne .LBB5_1344; AVX1-NEXT: # %bb.2: # %for.end345; AVX1-NEXT: vzeroupper346; AVX1-NEXT: retq347;348; AVX2-LABEL: example25:349; AVX2: # %bb.0: # %vector.ph350; AVX2-NEXT: movq $-4096, %rax # imm = 0xF000351; AVX2-NEXT: .p2align 4352; AVX2-NEXT: .LBB5_1: # %vector.body353; AVX2-NEXT: # =>This Inner Loop Header: Depth=1354; AVX2-NEXT: vmovups da+4096(%rax), %ymm0355; AVX2-NEXT: vcmpltps db+4096(%rax), %ymm0, %ymm0356; AVX2-NEXT: vmovups dc+4096(%rax), %ymm1357; AVX2-NEXT: vcmpltps dd+4096(%rax), %ymm1, %ymm1358; AVX2-NEXT: vandps %ymm1, %ymm0, %ymm0359; AVX2-NEXT: vpsrld $31, %ymm0, %ymm0360; AVX2-NEXT: vmovdqu %ymm0, dj+4096(%rax)361; AVX2-NEXT: addq $32, %rax362; AVX2-NEXT: jne .LBB5_1363; AVX2-NEXT: # %bb.2: # %for.end364; AVX2-NEXT: vzeroupper365; AVX2-NEXT: retq366vector.ph:367 br label %vector.body368 369vector.body:370 %index = phi i64 [ 0, %vector.ph ], [ %index.next, %vector.body ]371 %0 = getelementptr inbounds [1024 x float], ptr @da, i64 0, i64 %index372 %1 = load <8 x float>, ptr %0, align 16373 %2 = getelementptr inbounds [1024 x float], ptr @db, i64 0, i64 %index374 %3 = load <8 x float>, ptr %2, align 16375 %4 = fcmp olt <8 x float> %1, %3376 %5 = getelementptr inbounds [1024 x float], ptr @dc, i64 0, i64 %index377 %6 = load <8 x float>, ptr %5, align 16378 %7 = getelementptr inbounds [1024 x float], ptr @dd, i64 0, i64 %index379 %8 = load <8 x float>, ptr %7, align 16380 %9 = fcmp olt <8 x float> %6, %8381 %10 = and <8 x i1> %4, %9382 %11 = zext <8 x i1> %10 to <8 x i32>383 %12 = getelementptr inbounds [1024 x i32], ptr @dj, i64 0, i64 %index384 store <8 x i32> %11, ptr %12, align 16385 %index.next = add i64 %index, 8386 %13 = icmp eq i64 %index.next, 1024387 br i1 %13, label %for.end, label %vector.body388 389for.end:390 ret void391}392 393; TODO: AVX1 could have used 256-bit ops for a likely improvement.394 395define dso_local void @example24(i16 signext %x, i16 signext %y) nounwind {396; SSE2-LABEL: example24:397; SSE2: # %bb.0: # %vector.ph398; SSE2-NEXT: movd %edi, %xmm0399; SSE2-NEXT: pshuflw {{.*#+}} xmm0 = xmm0[0,0,0,0,4,5,6,7]400; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,1,0,1]401; SSE2-NEXT: movd %esi, %xmm1402; SSE2-NEXT: pshuflw {{.*#+}} xmm1 = xmm1[0,0,0,0,4,5,6,7]403; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm1[0,1,0,1]404; SSE2-NEXT: movq $-4096, %rax # imm = 0xF000405; SSE2-NEXT: .p2align 4406; SSE2-NEXT: .LBB6_1: # %vector.body407; SSE2-NEXT: # =>This Inner Loop Header: Depth=1408; SSE2-NEXT: movaps da+4096(%rax), %xmm2409; SSE2-NEXT: movaps da+4112(%rax), %xmm3410; SSE2-NEXT: cmpltps db+4112(%rax), %xmm3411; SSE2-NEXT: cmpltps db+4096(%rax), %xmm2412; SSE2-NEXT: packssdw %xmm3, %xmm2413; SSE2-NEXT: movdqa %xmm0, %xmm3414; SSE2-NEXT: pand %xmm2, %xmm3415; SSE2-NEXT: pandn %xmm1, %xmm2416; SSE2-NEXT: por %xmm3, %xmm2417; SSE2-NEXT: punpcklwd {{.*#+}} xmm3 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]418; SSE2-NEXT: psrad $16, %xmm3419; SSE2-NEXT: punpckhwd {{.*#+}} xmm2 = xmm2[4,4,5,5,6,6,7,7]420; SSE2-NEXT: psrad $16, %xmm2421; SSE2-NEXT: movdqa %xmm2, dj+4112(%rax)422; SSE2-NEXT: movdqa %xmm3, dj+4096(%rax)423; SSE2-NEXT: addq $32, %rax424; SSE2-NEXT: jne .LBB6_1425; SSE2-NEXT: # %bb.2: # %for.end426; SSE2-NEXT: retq427;428; SSE41-LABEL: example24:429; SSE41: # %bb.0: # %vector.ph430; SSE41-NEXT: movd %edi, %xmm0431; SSE41-NEXT: pshuflw {{.*#+}} xmm0 = xmm0[0,0,0,0,4,5,6,7]432; SSE41-NEXT: pshufd {{.*#+}} xmm1 = xmm0[0,1,0,1]433; SSE41-NEXT: movd %esi, %xmm0434; SSE41-NEXT: pshuflw {{.*#+}} xmm0 = xmm0[0,0,0,0,4,5,6,7]435; SSE41-NEXT: pshufd {{.*#+}} xmm2 = xmm0[0,1,0,1]436; SSE41-NEXT: movq $-4096, %rax # imm = 0xF000437; SSE41-NEXT: .p2align 4438; SSE41-NEXT: .LBB6_1: # %vector.body439; SSE41-NEXT: # =>This Inner Loop Header: Depth=1440; SSE41-NEXT: movaps da+4096(%rax), %xmm0441; SSE41-NEXT: movaps da+4112(%rax), %xmm3442; SSE41-NEXT: cmpltps db+4112(%rax), %xmm3443; SSE41-NEXT: cmpltps db+4096(%rax), %xmm0444; SSE41-NEXT: packssdw %xmm3, %xmm0445; SSE41-NEXT: movdqa %xmm2, %xmm3446; SSE41-NEXT: pblendvb %xmm0, %xmm1, %xmm3447; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm3[2,3,2,3]448; SSE41-NEXT: pmovsxwd %xmm0, %xmm0449; SSE41-NEXT: pmovsxwd %xmm3, %xmm3450; SSE41-NEXT: movdqa %xmm3, dj+4096(%rax)451; SSE41-NEXT: movdqa %xmm0, dj+4112(%rax)452; SSE41-NEXT: addq $32, %rax453; SSE41-NEXT: jne .LBB6_1454; SSE41-NEXT: # %bb.2: # %for.end455; SSE41-NEXT: retq456;457; AVX1-LABEL: example24:458; AVX1: # %bb.0: # %vector.ph459; AVX1-NEXT: vmovd %edi, %xmm0460; AVX1-NEXT: vpshuflw {{.*#+}} xmm0 = xmm0[0,0,0,0,4,5,6,7]461; AVX1-NEXT: vpshufd {{.*#+}} xmm2 = xmm0[0,1,0,1]462; AVX1-NEXT: vmovd %esi, %xmm0463; AVX1-NEXT: vpshuflw {{.*#+}} xmm0 = xmm0[0,0,0,0,4,5,6,7]464; AVX1-NEXT: vpshufd {{.*#+}} xmm3 = xmm0[0,1,0,1]465; AVX1-NEXT: movq $-4096, %rax # imm = 0xF000466; AVX1-NEXT: vpmovsxwd %xmm2, %xmm0467; AVX1-NEXT: vpmovsxwd %xmm3, %xmm1468; AVX1-NEXT: vpshufd {{.*#+}} xmm2 = xmm2[2,3,2,3]469; AVX1-NEXT: vpmovsxwd %xmm2, %xmm2470; AVX1-NEXT: vpshufd {{.*#+}} xmm3 = xmm3[2,3,2,3]471; AVX1-NEXT: vpmovsxwd %xmm3, %xmm3472; AVX1-NEXT: .p2align 4473; AVX1-NEXT: .LBB6_1: # %vector.body474; AVX1-NEXT: # =>This Inner Loop Header: Depth=1475; AVX1-NEXT: vmovups da+4096(%rax), %ymm4476; AVX1-NEXT: vcmpltps db+4096(%rax), %ymm4, %ymm4477; AVX1-NEXT: vblendvps %xmm4, %xmm0, %xmm1, %xmm5478; AVX1-NEXT: vextractf128 $1, %ymm4, %xmm4479; AVX1-NEXT: vblendvps %xmm4, %xmm2, %xmm3, %xmm4480; AVX1-NEXT: vmovaps %xmm4, dj+4112(%rax)481; AVX1-NEXT: vmovaps %xmm5, dj+4096(%rax)482; AVX1-NEXT: addq $32, %rax483; AVX1-NEXT: jne .LBB6_1484; AVX1-NEXT: # %bb.2: # %for.end485; AVX1-NEXT: vzeroupper486; AVX1-NEXT: retq487;488; AVX2-LABEL: example24:489; AVX2: # %bb.0: # %vector.ph490; AVX2-NEXT: vmovd %edi, %xmm0491; AVX2-NEXT: vpbroadcastw %xmm0, %xmm0492; AVX2-NEXT: vmovd %esi, %xmm1493; AVX2-NEXT: vpbroadcastw %xmm1, %xmm1494; AVX2-NEXT: movq $-4096, %rax # imm = 0xF000495; AVX2-NEXT: vpmovsxwd %xmm0, %ymm0496; AVX2-NEXT: vpmovsxwd %xmm1, %ymm1497; AVX2-NEXT: .p2align 4498; AVX2-NEXT: .LBB6_1: # %vector.body499; AVX2-NEXT: # =>This Inner Loop Header: Depth=1500; AVX2-NEXT: vmovups da+4096(%rax), %ymm2501; AVX2-NEXT: vcmpltps db+4096(%rax), %ymm2, %ymm2502; AVX2-NEXT: vblendvps %ymm2, %ymm0, %ymm1, %ymm2503; AVX2-NEXT: vmovups %ymm2, dj+4096(%rax)504; AVX2-NEXT: addq $32, %rax505; AVX2-NEXT: jne .LBB6_1506; AVX2-NEXT: # %bb.2: # %for.end507; AVX2-NEXT: vzeroupper508; AVX2-NEXT: retq509vector.ph:510 %0 = insertelement <8 x i16> undef, i16 %x, i32 0511 %broadcast11 = shufflevector <8 x i16> %0, <8 x i16> undef, <8 x i32> zeroinitializer512 %1 = insertelement <8 x i16> undef, i16 %y, i32 0513 %broadcast12 = shufflevector <8 x i16> %1, <8 x i16> undef, <8 x i32> zeroinitializer514 br label %vector.body515 516vector.body:517 %index = phi i64 [ 0, %vector.ph ], [ %index.next, %vector.body ]518 %2 = getelementptr inbounds [1024 x float], ptr @da, i64 0, i64 %index519 %3 = load <8 x float>, ptr %2, align 16520 %4 = getelementptr inbounds [1024 x float], ptr @db, i64 0, i64 %index521 %5 = load <8 x float>, ptr %4, align 16522 %6 = fcmp olt <8 x float> %3, %5523 %7 = select <8 x i1> %6, <8 x i16> %broadcast11, <8 x i16> %broadcast12524 %8 = sext <8 x i16> %7 to <8 x i32>525 %9 = getelementptr inbounds [1024 x i32], ptr @dj, i64 0, i64 %index526 store <8 x i32> %8, ptr %9, align 16527 %index.next = add i64 %index, 8528 %10 = icmp eq i64 %index.next, 1024529 br i1 %10, label %for.end, label %vector.body530 531for.end:532 ret void533}534 535