233 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py2; RUN: llc < %s -mtriple=i686-unknown-unknown -mattr=+avx2 | FileCheck %s --check-prefixes=CHECK,X86,X86-SLOW3; RUN: llc < %s -mtriple=i686-unknown-unknown -mattr=+avx2,+fast-variable-crosslane-shuffle,+fast-variable-perlane-shuffle | FileCheck %s --check-prefixes=CHECK,X86,X86-FAST-ALL4; RUN: llc < %s -mtriple=i686-unknown-unknown -mattr=+avx2,+fast-variable-perlane-shuffle | FileCheck %s --check-prefixes=CHECK,X86,X86-FAST-PERLANE5; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx2 | FileCheck %s --check-prefixes=CHECK,X64,X64-SLOW6; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx2,+fast-variable-crosslane-shuffle,+fast-variable-perlane-shuffle | FileCheck %s --check-prefixes=CHECK,X64,X64-FAST-ALL7; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx2,+fast-variable-perlane-shuffle | FileCheck %s --check-prefixes=CHECK,X64,X64-FAST-PERLANE8 9define <4 x i32> @trunc4(<4 x i64> %A) nounwind {10; X86-SLOW-LABEL: trunc4:11; X86-SLOW: # %bb.0:12; X86-SLOW-NEXT: vextractf128 $1, %ymm0, %xmm113; X86-SLOW-NEXT: vshufps {{.*#+}} xmm0 = xmm0[0,2],xmm1[0,2]14; X86-SLOW-NEXT: vzeroupper15; X86-SLOW-NEXT: retl16;17; X86-FAST-ALL-LABEL: trunc4:18; X86-FAST-ALL: # %bb.0:19; X86-FAST-ALL-NEXT: vbroadcastf128 {{.*#+}} ymm1 = [0,2,4,6,0,2,4,6]20; X86-FAST-ALL-NEXT: # ymm1 = mem[0,1,0,1]21; X86-FAST-ALL-NEXT: vpermps %ymm0, %ymm1, %ymm022; X86-FAST-ALL-NEXT: # kill: def $xmm0 killed $xmm0 killed $ymm023; X86-FAST-ALL-NEXT: vzeroupper24; X86-FAST-ALL-NEXT: retl25;26; X86-FAST-PERLANE-LABEL: trunc4:27; X86-FAST-PERLANE: # %bb.0:28; X86-FAST-PERLANE-NEXT: vextractf128 $1, %ymm0, %xmm129; X86-FAST-PERLANE-NEXT: vshufps {{.*#+}} xmm0 = xmm0[0,2],xmm1[0,2]30; X86-FAST-PERLANE-NEXT: vzeroupper31; X86-FAST-PERLANE-NEXT: retl32;33; X64-SLOW-LABEL: trunc4:34; X64-SLOW: # %bb.0:35; X64-SLOW-NEXT: vextractf128 $1, %ymm0, %xmm136; X64-SLOW-NEXT: vshufps {{.*#+}} xmm0 = xmm0[0,2],xmm1[0,2]37; X64-SLOW-NEXT: vzeroupper38; X64-SLOW-NEXT: retq39;40; X64-FAST-ALL-LABEL: trunc4:41; X64-FAST-ALL: # %bb.0:42; X64-FAST-ALL-NEXT: vbroadcastf128 {{.*#+}} ymm1 = [0,2,4,6,0,2,4,6]43; X64-FAST-ALL-NEXT: # ymm1 = mem[0,1,0,1]44; X64-FAST-ALL-NEXT: vpermps %ymm0, %ymm1, %ymm045; X64-FAST-ALL-NEXT: # kill: def $xmm0 killed $xmm0 killed $ymm046; X64-FAST-ALL-NEXT: vzeroupper47; X64-FAST-ALL-NEXT: retq48;49; X64-FAST-PERLANE-LABEL: trunc4:50; X64-FAST-PERLANE: # %bb.0:51; X64-FAST-PERLANE-NEXT: vextractf128 $1, %ymm0, %xmm152; X64-FAST-PERLANE-NEXT: vshufps {{.*#+}} xmm0 = xmm0[0,2],xmm1[0,2]53; X64-FAST-PERLANE-NEXT: vzeroupper54; X64-FAST-PERLANE-NEXT: retq55 %B = trunc <4 x i64> %A to <4 x i32>56 ret <4 x i32>%B57}58 59define <8 x i16> @trunc8(<8 x i32> %A) nounwind {60; CHECK-LABEL: trunc8:61; CHECK: # %bb.0:62; CHECK-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,1,4,5,8,9,12,13,u,u,u,u,u,u,u,u,16,17,20,21,24,25,28,29,u,u,u,u,u,u,u,u]63; CHECK-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,2,2,3]64; CHECK-NEXT: # kill: def $xmm0 killed $xmm0 killed $ymm065; CHECK-NEXT: vzeroupper66; CHECK-NEXT: ret{{[l|q]}}67 %B = trunc <8 x i32> %A to <8 x i16>68 ret <8 x i16>%B69}70 71define <4 x i64> @sext4(<4 x i32> %A) nounwind {72; CHECK-LABEL: sext4:73; CHECK: # %bb.0:74; CHECK-NEXT: vpmovsxdq %xmm0, %ymm075; CHECK-NEXT: ret{{[l|q]}}76 %B = sext <4 x i32> %A to <4 x i64>77 ret <4 x i64>%B78}79 80define <8 x i32> @sext8(<8 x i16> %A) nounwind {81; CHECK-LABEL: sext8:82; CHECK: # %bb.0:83; CHECK-NEXT: vpmovsxwd %xmm0, %ymm084; CHECK-NEXT: ret{{[l|q]}}85 %B = sext <8 x i16> %A to <8 x i32>86 ret <8 x i32>%B87}88 89define <4 x i64> @zext4(<4 x i32> %A) nounwind {90; CHECK-LABEL: zext4:91; CHECK: # %bb.0:92; CHECK-NEXT: vpmovzxdq {{.*#+}} ymm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero93; CHECK-NEXT: ret{{[l|q]}}94 %B = zext <4 x i32> %A to <4 x i64>95 ret <4 x i64>%B96}97 98define <8 x i32> @zext8(<8 x i16> %A) nounwind {99; CHECK-LABEL: zext8:100; CHECK: # %bb.0:101; CHECK-NEXT: vpmovzxwd {{.*#+}} ymm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero102; CHECK-NEXT: ret{{[l|q]}}103 %B = zext <8 x i16> %A to <8 x i32>104 ret <8 x i32>%B105}106 107define <8 x i32> @zext_8i8_8i32(<8 x i8> %A) nounwind {108; CHECK-LABEL: zext_8i8_8i32:109; CHECK: # %bb.0:110; CHECK-NEXT: vpmovzxbd {{.*#+}} ymm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],zero,zero,zero,xmm0[3],zero,zero,zero,xmm0[4],zero,zero,zero,xmm0[5],zero,zero,zero,xmm0[6],zero,zero,zero,xmm0[7],zero,zero,zero111; CHECK-NEXT: ret{{[l|q]}}112 %B = zext <8 x i8> %A to <8 x i32>113 ret <8 x i32>%B114}115 116define <16 x i16> @zext_16i8_16i16(<16 x i8> %z) {117; CHECK-LABEL: zext_16i8_16i16:118; CHECK: # %bb.0:119; CHECK-NEXT: vpmovzxbw {{.*#+}} ymm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero,xmm0[8],zero,xmm0[9],zero,xmm0[10],zero,xmm0[11],zero,xmm0[12],zero,xmm0[13],zero,xmm0[14],zero,xmm0[15],zero120; CHECK-NEXT: ret{{[l|q]}}121 %t = zext <16 x i8> %z to <16 x i16>122 ret <16 x i16> %t123}124 125define <16 x i16> @sext_16i8_16i16(<16 x i8> %z) {126; CHECK-LABEL: sext_16i8_16i16:127; CHECK: # %bb.0:128; CHECK-NEXT: vpmovsxbw %xmm0, %ymm0129; CHECK-NEXT: ret{{[l|q]}}130 %t = sext <16 x i8> %z to <16 x i16>131 ret <16 x i16> %t132}133 134define <16 x i8> @trunc_16i16_16i8(<16 x i16> %z) {135; X86-LABEL: trunc_16i16_16i8:136; X86: # %bb.0:137; X86-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}, %ymm0, %ymm0138; X86-NEXT: vextracti128 $1, %ymm0, %xmm1139; X86-NEXT: vpackuswb %xmm1, %xmm0, %xmm0140; X86-NEXT: vzeroupper141; X86-NEXT: retl142;143; X64-LABEL: trunc_16i16_16i8:144; X64: # %bb.0:145; X64-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm0, %ymm0146; X64-NEXT: vextracti128 $1, %ymm0, %xmm1147; X64-NEXT: vpackuswb %xmm1, %xmm0, %xmm0148; X64-NEXT: vzeroupper149; X64-NEXT: retq150 %t = trunc <16 x i16> %z to <16 x i8>151 ret <16 x i8> %t152}153 154define <4 x i64> @load_sext_test1(ptr%ptr) {155; X86-LABEL: load_sext_test1:156; X86: # %bb.0:157; X86-NEXT: movl {{[0-9]+}}(%esp), %eax158; X86-NEXT: vpmovsxdq (%eax), %ymm0159; X86-NEXT: retl160;161; X64-LABEL: load_sext_test1:162; X64: # %bb.0:163; X64-NEXT: vpmovsxdq (%rdi), %ymm0164; X64-NEXT: retq165 %X = load <4 x i32>, ptr %ptr166 %Y = sext <4 x i32> %X to <4 x i64>167 ret <4 x i64>%Y168}169 170define <4 x i64> @load_sext_test2(ptr%ptr) {171; X86-LABEL: load_sext_test2:172; X86: # %bb.0:173; X86-NEXT: movl {{[0-9]+}}(%esp), %eax174; X86-NEXT: vpmovsxbq (%eax), %ymm0175; X86-NEXT: retl176;177; X64-LABEL: load_sext_test2:178; X64: # %bb.0:179; X64-NEXT: vpmovsxbq (%rdi), %ymm0180; X64-NEXT: retq181 %X = load <4 x i8>, ptr %ptr182 %Y = sext <4 x i8> %X to <4 x i64>183 ret <4 x i64>%Y184}185 186define <4 x i64> @load_sext_test3(ptr%ptr) {187; X86-LABEL: load_sext_test3:188; X86: # %bb.0:189; X86-NEXT: movl {{[0-9]+}}(%esp), %eax190; X86-NEXT: vpmovsxwq (%eax), %ymm0191; X86-NEXT: retl192;193; X64-LABEL: load_sext_test3:194; X64: # %bb.0:195; X64-NEXT: vpmovsxwq (%rdi), %ymm0196; X64-NEXT: retq197 %X = load <4 x i16>, ptr %ptr198 %Y = sext <4 x i16> %X to <4 x i64>199 ret <4 x i64>%Y200}201 202define <8 x i32> @load_sext_test4(ptr%ptr) {203; X86-LABEL: load_sext_test4:204; X86: # %bb.0:205; X86-NEXT: movl {{[0-9]+}}(%esp), %eax206; X86-NEXT: vpmovsxwd (%eax), %ymm0207; X86-NEXT: retl208;209; X64-LABEL: load_sext_test4:210; X64: # %bb.0:211; X64-NEXT: vpmovsxwd (%rdi), %ymm0212; X64-NEXT: retq213 %X = load <8 x i16>, ptr %ptr214 %Y = sext <8 x i16> %X to <8 x i32>215 ret <8 x i32>%Y216}217 218define <8 x i32> @load_sext_test5(ptr%ptr) {219; X86-LABEL: load_sext_test5:220; X86: # %bb.0:221; X86-NEXT: movl {{[0-9]+}}(%esp), %eax222; X86-NEXT: vpmovsxbd (%eax), %ymm0223; X86-NEXT: retl224;225; X64-LABEL: load_sext_test5:226; X64: # %bb.0:227; X64-NEXT: vpmovsxbd (%rdi), %ymm0228; X64-NEXT: retq229 %X = load <8 x i8>, ptr %ptr230 %Y = sext <8 x i8> %X to <8 x i32>231 ret <8 x i32>%Y232}233