332 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py2; RUN: llc < %s -disable-peephole -mtriple=i686-unknown-unknown -mattr=+avx,+xop | FileCheck %s --check-prefixes=CHECK,X863; RUN: llc < %s -disable-peephole -mtriple=x86_64-unknown-unknown -mattr=+avx,+xop | FileCheck %s --check-prefixes=CHECK,X644 5define <16 x i8> @commute_fold_vpcomb(ptr %a0, <16 x i8> %a1) {6; X86-LABEL: commute_fold_vpcomb:7; X86: # %bb.0:8; X86-NEXT: movl {{[0-9]+}}(%esp), %eax9; X86-NEXT: vpcomgtb (%eax), %xmm0, %xmm010; X86-NEXT: retl11;12; X64-LABEL: commute_fold_vpcomb:13; X64: # %bb.0:14; X64-NEXT: vpcomgtb (%rdi), %xmm0, %xmm015; X64-NEXT: retq16 %1 = load <16 x i8>, ptr %a017 %2 = call <16 x i8> @llvm.x86.xop.vpcomb(<16 x i8> %1, <16 x i8> %a1, i8 0) ; vpcomltb18 ret <16 x i8> %219}20declare <16 x i8> @llvm.x86.xop.vpcomb(<16 x i8>, <16 x i8>, i8) nounwind readnone21 22define <4 x i32> @commute_fold_vpcomd(ptr %a0, <4 x i32> %a1) {23; X86-LABEL: commute_fold_vpcomd:24; X86: # %bb.0:25; X86-NEXT: movl {{[0-9]+}}(%esp), %eax26; X86-NEXT: vpcomged (%eax), %xmm0, %xmm027; X86-NEXT: retl28;29; X64-LABEL: commute_fold_vpcomd:30; X64: # %bb.0:31; X64-NEXT: vpcomged (%rdi), %xmm0, %xmm032; X64-NEXT: retq33 %1 = load <4 x i32>, ptr %a034 %2 = call <4 x i32> @llvm.x86.xop.vpcomd(<4 x i32> %1, <4 x i32> %a1, i8 1) ; vpcomled35 ret <4 x i32> %236}37declare <4 x i32> @llvm.x86.xop.vpcomd(<4 x i32>, <4 x i32>, i8) nounwind readnone38 39define <2 x i64> @commute_fold_vpcomq(ptr %a0, <2 x i64> %a1) {40; X86-LABEL: commute_fold_vpcomq:41; X86: # %bb.0:42; X86-NEXT: movl {{[0-9]+}}(%esp), %eax43; X86-NEXT: vpcomltq (%eax), %xmm0, %xmm044; X86-NEXT: retl45;46; X64-LABEL: commute_fold_vpcomq:47; X64: # %bb.0:48; X64-NEXT: vpcomltq (%rdi), %xmm0, %xmm049; X64-NEXT: retq50 %1 = load <2 x i64>, ptr %a051 %2 = call <2 x i64> @llvm.x86.xop.vpcomq(<2 x i64> %1, <2 x i64> %a1, i8 2) ; vpcomgtq52 ret <2 x i64> %253}54declare <2 x i64> @llvm.x86.xop.vpcomq(<2 x i64>, <2 x i64>, i8) nounwind readnone55 56define <16 x i8> @commute_fold_vpcomub(ptr %a0, <16 x i8> %a1) {57; X86-LABEL: commute_fold_vpcomub:58; X86: # %bb.0:59; X86-NEXT: movl {{[0-9]+}}(%esp), %eax60; X86-NEXT: vpcomleub (%eax), %xmm0, %xmm061; X86-NEXT: retl62;63; X64-LABEL: commute_fold_vpcomub:64; X64: # %bb.0:65; X64-NEXT: vpcomleub (%rdi), %xmm0, %xmm066; X64-NEXT: retq67 %1 = load <16 x i8>, ptr %a068 %2 = call <16 x i8> @llvm.x86.xop.vpcomub(<16 x i8> %1, <16 x i8> %a1, i8 3) ; vpcomgeub69 ret <16 x i8> %270}71declare <16 x i8> @llvm.x86.xop.vpcomub(<16 x i8>, <16 x i8>, i8) nounwind readnone72 73define <4 x i32> @commute_fold_vpcomud(ptr %a0, <4 x i32> %a1) {74; X86-LABEL: commute_fold_vpcomud:75; X86: # %bb.0:76; X86-NEXT: movl {{[0-9]+}}(%esp), %eax77; X86-NEXT: vpcomeqd (%eax), %xmm0, %xmm078; X86-NEXT: retl79;80; X64-LABEL: commute_fold_vpcomud:81; X64: # %bb.0:82; X64-NEXT: vpcomeqd (%rdi), %xmm0, %xmm083; X64-NEXT: retq84 %1 = load <4 x i32>, ptr %a085 %2 = call <4 x i32> @llvm.x86.xop.vpcomud(<4 x i32> %1, <4 x i32> %a1, i8 4) ; vpcomequd86 ret <4 x i32> %287}88declare <4 x i32> @llvm.x86.xop.vpcomud(<4 x i32>, <4 x i32>, i8) nounwind readnone89 90define <2 x i64> @commute_fold_vpcomuq(ptr %a0, <2 x i64> %a1) {91; X86-LABEL: commute_fold_vpcomuq:92; X86: # %bb.0:93; X86-NEXT: movl {{[0-9]+}}(%esp), %eax94; X86-NEXT: vpcomneqq (%eax), %xmm0, %xmm095; X86-NEXT: retl96;97; X64-LABEL: commute_fold_vpcomuq:98; X64: # %bb.0:99; X64-NEXT: vpcomneqq (%rdi), %xmm0, %xmm0100; X64-NEXT: retq101 %1 = load <2 x i64>, ptr %a0102 %2 = call <2 x i64> @llvm.x86.xop.vpcomuq(<2 x i64> %1, <2 x i64> %a1, i8 5) ; vpcomnequq103 ret <2 x i64> %2104}105declare <2 x i64> @llvm.x86.xop.vpcomuq(<2 x i64>, <2 x i64>, i8) nounwind readnone106 107define <8 x i16> @commute_fold_vpcomuw(ptr %a0, <8 x i16> %a1) {108; CHECK-LABEL: commute_fold_vpcomuw:109; CHECK: # %bb.0:110; CHECK-NEXT: vxorps %xmm0, %xmm0, %xmm0111; CHECK-NEXT: ret{{[l|q]}}112 %1 = load <8 x i16>, ptr %a0113 %2 = call <8 x i16> @llvm.x86.xop.vpcomuw(<8 x i16> %1, <8 x i16> %a1, i8 6) ; vpcomfalseuw114 ret <8 x i16> %2115}116declare <8 x i16> @llvm.x86.xop.vpcomuw(<8 x i16>, <8 x i16>, i8) nounwind readnone117 118define <8 x i16> @commute_fold_vpcomw(ptr %a0, <8 x i16> %a1) {119; CHECK-LABEL: commute_fold_vpcomw:120; CHECK: # %bb.0:121; CHECK-NEXT: vpcmpeqd %xmm0, %xmm0, %xmm0122; CHECK-NEXT: ret{{[l|q]}}123 %1 = load <8 x i16>, ptr %a0124 %2 = call <8 x i16> @llvm.x86.xop.vpcomw(<8 x i16> %1, <8 x i16> %a1, i8 7) ; vpcomtruew125 ret <8 x i16> %2126}127declare <8 x i16> @llvm.x86.xop.vpcomw(<8 x i16>, <8 x i16>, i8) nounwind readnone128 129define <4 x i32> @commute_fold_vpmacsdd(ptr %a0, <4 x i32> %a1, <4 x i32> %a2) {130; X86-LABEL: commute_fold_vpmacsdd:131; X86: # %bb.0:132; X86-NEXT: movl {{[0-9]+}}(%esp), %eax133; X86-NEXT: vpmacsdd %xmm1, (%eax), %xmm0, %xmm0134; X86-NEXT: retl135;136; X64-LABEL: commute_fold_vpmacsdd:137; X64: # %bb.0:138; X64-NEXT: vpmacsdd %xmm1, (%rdi), %xmm0, %xmm0139; X64-NEXT: retq140 %1 = load <4 x i32>, ptr %a0141 %2 = call <4 x i32> @llvm.x86.xop.vpmacsdd(<4 x i32> %1, <4 x i32> %a1, <4 x i32> %a2)142 ret <4 x i32> %2143}144declare <4 x i32> @llvm.x86.xop.vpmacsdd(<4 x i32>, <4 x i32>, <4 x i32>) nounwind readnone145 146define <2 x i64> @commute_fold_vpmacsdqh(ptr %a0, <4 x i32> %a1, <2 x i64> %a2) {147; X86-LABEL: commute_fold_vpmacsdqh:148; X86: # %bb.0:149; X86-NEXT: movl {{[0-9]+}}(%esp), %eax150; X86-NEXT: vpmacsdqh %xmm1, (%eax), %xmm0, %xmm0151; X86-NEXT: retl152;153; X64-LABEL: commute_fold_vpmacsdqh:154; X64: # %bb.0:155; X64-NEXT: vpmacsdqh %xmm1, (%rdi), %xmm0, %xmm0156; X64-NEXT: retq157 %1 = load <4 x i32>, ptr %a0158 %2 = call <2 x i64> @llvm.x86.xop.vpmacsdqh(<4 x i32> %1, <4 x i32> %a1, <2 x i64> %a2)159 ret <2 x i64> %2160}161declare <2 x i64> @llvm.x86.xop.vpmacsdqh(<4 x i32>, <4 x i32>, <2 x i64>) nounwind readnone162 163define <2 x i64> @commute_fold_vpmacsdql(ptr %a0, <4 x i32> %a1, <2 x i64> %a2) {164; X86-LABEL: commute_fold_vpmacsdql:165; X86: # %bb.0:166; X86-NEXT: movl {{[0-9]+}}(%esp), %eax167; X86-NEXT: vpmacsdql %xmm1, (%eax), %xmm0, %xmm0168; X86-NEXT: retl169;170; X64-LABEL: commute_fold_vpmacsdql:171; X64: # %bb.0:172; X64-NEXT: vpmacsdql %xmm1, (%rdi), %xmm0, %xmm0173; X64-NEXT: retq174 %1 = load <4 x i32>, ptr %a0175 %2 = call <2 x i64> @llvm.x86.xop.vpmacsdql(<4 x i32> %1, <4 x i32> %a1, <2 x i64> %a2)176 ret <2 x i64> %2177}178declare <2 x i64> @llvm.x86.xop.vpmacsdql(<4 x i32>, <4 x i32>, <2 x i64>) nounwind readnone179 180define <4 x i32> @commute_fold_vpmacssdd(ptr %a0, <4 x i32> %a1, <4 x i32> %a2) {181; X86-LABEL: commute_fold_vpmacssdd:182; X86: # %bb.0:183; X86-NEXT: movl {{[0-9]+}}(%esp), %eax184; X86-NEXT: vpmacssdd %xmm1, (%eax), %xmm0, %xmm0185; X86-NEXT: retl186;187; X64-LABEL: commute_fold_vpmacssdd:188; X64: # %bb.0:189; X64-NEXT: vpmacssdd %xmm1, (%rdi), %xmm0, %xmm0190; X64-NEXT: retq191 %1 = load <4 x i32>, ptr %a0192 %2 = call <4 x i32> @llvm.x86.xop.vpmacssdd(<4 x i32> %1, <4 x i32> %a1, <4 x i32> %a2)193 ret <4 x i32> %2194}195declare <4 x i32> @llvm.x86.xop.vpmacssdd(<4 x i32>, <4 x i32>, <4 x i32>) nounwind readnone196 197define <2 x i64> @commute_fold_vpmacssdqh(ptr %a0, <4 x i32> %a1, <2 x i64> %a2) {198; X86-LABEL: commute_fold_vpmacssdqh:199; X86: # %bb.0:200; X86-NEXT: movl {{[0-9]+}}(%esp), %eax201; X86-NEXT: vpmacssdqh %xmm1, (%eax), %xmm0, %xmm0202; X86-NEXT: retl203;204; X64-LABEL: commute_fold_vpmacssdqh:205; X64: # %bb.0:206; X64-NEXT: vpmacssdqh %xmm1, (%rdi), %xmm0, %xmm0207; X64-NEXT: retq208 %1 = load <4 x i32>, ptr %a0209 %2 = call <2 x i64> @llvm.x86.xop.vpmacssdqh(<4 x i32> %1, <4 x i32> %a1, <2 x i64> %a2)210 ret <2 x i64> %2211}212declare <2 x i64> @llvm.x86.xop.vpmacssdqh(<4 x i32>, <4 x i32>, <2 x i64>) nounwind readnone213 214define <2 x i64> @commute_fold_vpmacssdql(ptr %a0, <4 x i32> %a1, <2 x i64> %a2) {215; X86-LABEL: commute_fold_vpmacssdql:216; X86: # %bb.0:217; X86-NEXT: movl {{[0-9]+}}(%esp), %eax218; X86-NEXT: vpmacssdql %xmm1, (%eax), %xmm0, %xmm0219; X86-NEXT: retl220;221; X64-LABEL: commute_fold_vpmacssdql:222; X64: # %bb.0:223; X64-NEXT: vpmacssdql %xmm1, (%rdi), %xmm0, %xmm0224; X64-NEXT: retq225 %1 = load <4 x i32>, ptr %a0226 %2 = call <2 x i64> @llvm.x86.xop.vpmacssdql(<4 x i32> %1, <4 x i32> %a1, <2 x i64> %a2)227 ret <2 x i64> %2228}229declare <2 x i64> @llvm.x86.xop.vpmacssdql(<4 x i32>, <4 x i32>, <2 x i64>) nounwind readnone230 231define <4 x i32> @commute_fold_vpmacsswd(ptr %a0, <8 x i16> %a1, <4 x i32> %a2) {232; X86-LABEL: commute_fold_vpmacsswd:233; X86: # %bb.0:234; X86-NEXT: movl {{[0-9]+}}(%esp), %eax235; X86-NEXT: vpmacsswd %xmm1, (%eax), %xmm0, %xmm0236; X86-NEXT: retl237;238; X64-LABEL: commute_fold_vpmacsswd:239; X64: # %bb.0:240; X64-NEXT: vpmacsswd %xmm1, (%rdi), %xmm0, %xmm0241; X64-NEXT: retq242 %1 = load <8 x i16>, ptr %a0243 %2 = call <4 x i32> @llvm.x86.xop.vpmacsswd(<8 x i16> %1, <8 x i16> %a1, <4 x i32> %a2)244 ret <4 x i32> %2245}246declare <4 x i32> @llvm.x86.xop.vpmacsswd(<8 x i16>, <8 x i16>, <4 x i32>) nounwind readnone247 248define <8 x i16> @commute_fold_vpmacssww(ptr %a0, <8 x i16> %a1, <8 x i16> %a2) {249; X86-LABEL: commute_fold_vpmacssww:250; X86: # %bb.0:251; X86-NEXT: movl {{[0-9]+}}(%esp), %eax252; X86-NEXT: vpmacssww %xmm1, (%eax), %xmm0, %xmm0253; X86-NEXT: retl254;255; X64-LABEL: commute_fold_vpmacssww:256; X64: # %bb.0:257; X64-NEXT: vpmacssww %xmm1, (%rdi), %xmm0, %xmm0258; X64-NEXT: retq259 %1 = load <8 x i16>, ptr %a0260 %2 = call <8 x i16> @llvm.x86.xop.vpmacssww(<8 x i16> %1, <8 x i16> %a1, <8 x i16> %a2)261 ret <8 x i16> %2262}263declare <8 x i16> @llvm.x86.xop.vpmacssww(<8 x i16>, <8 x i16>, <8 x i16>) nounwind readnone264 265define <4 x i32> @commute_fold_vpmacswd(ptr %a0, <8 x i16> %a1, <4 x i32> %a2) {266; X86-LABEL: commute_fold_vpmacswd:267; X86: # %bb.0:268; X86-NEXT: movl {{[0-9]+}}(%esp), %eax269; X86-NEXT: vpmacswd %xmm1, (%eax), %xmm0, %xmm0270; X86-NEXT: retl271;272; X64-LABEL: commute_fold_vpmacswd:273; X64: # %bb.0:274; X64-NEXT: vpmacswd %xmm1, (%rdi), %xmm0, %xmm0275; X64-NEXT: retq276 %1 = load <8 x i16>, ptr %a0277 %2 = call <4 x i32> @llvm.x86.xop.vpmacswd(<8 x i16> %1, <8 x i16> %a1, <4 x i32> %a2)278 ret <4 x i32> %2279}280declare <4 x i32> @llvm.x86.xop.vpmacswd(<8 x i16>, <8 x i16>, <4 x i32>) nounwind readnone281 282define <8 x i16> @commute_fold_vpmacsww(ptr %a0, <8 x i16> %a1, <8 x i16> %a2) {283; X86-LABEL: commute_fold_vpmacsww:284; X86: # %bb.0:285; X86-NEXT: movl {{[0-9]+}}(%esp), %eax286; X86-NEXT: vpmacsww %xmm1, (%eax), %xmm0, %xmm0287; X86-NEXT: retl288;289; X64-LABEL: commute_fold_vpmacsww:290; X64: # %bb.0:291; X64-NEXT: vpmacsww %xmm1, (%rdi), %xmm0, %xmm0292; X64-NEXT: retq293 %1 = load <8 x i16>, ptr %a0294 %2 = call <8 x i16> @llvm.x86.xop.vpmacsww(<8 x i16> %1, <8 x i16> %a1, <8 x i16> %a2)295 ret <8 x i16> %2296}297declare <8 x i16> @llvm.x86.xop.vpmacsww(<8 x i16>, <8 x i16>, <8 x i16>) nounwind readnone298 299define <4 x i32> @commute_fold_vpmadcsswd(ptr %a0, <8 x i16> %a1, <4 x i32> %a2) {300; X86-LABEL: commute_fold_vpmadcsswd:301; X86: # %bb.0:302; X86-NEXT: movl {{[0-9]+}}(%esp), %eax303; X86-NEXT: vpmadcsswd %xmm1, (%eax), %xmm0, %xmm0304; X86-NEXT: retl305;306; X64-LABEL: commute_fold_vpmadcsswd:307; X64: # %bb.0:308; X64-NEXT: vpmadcsswd %xmm1, (%rdi), %xmm0, %xmm0309; X64-NEXT: retq310 %1 = load <8 x i16>, ptr %a0311 %2 = call <4 x i32> @llvm.x86.xop.vpmadcsswd(<8 x i16> %1, <8 x i16> %a1, <4 x i32> %a2)312 ret <4 x i32> %2313}314declare <4 x i32> @llvm.x86.xop.vpmadcsswd(<8 x i16>, <8 x i16>, <4 x i32>) nounwind readnone315 316define <4 x i32> @commute_fold_vpmadcswd(ptr %a0, <8 x i16> %a1, <4 x i32> %a2) {317; X86-LABEL: commute_fold_vpmadcswd:318; X86: # %bb.0:319; X86-NEXT: movl {{[0-9]+}}(%esp), %eax320; X86-NEXT: vpmadcswd %xmm1, (%eax), %xmm0, %xmm0321; X86-NEXT: retl322;323; X64-LABEL: commute_fold_vpmadcswd:324; X64: # %bb.0:325; X64-NEXT: vpmadcswd %xmm1, (%rdi), %xmm0, %xmm0326; X64-NEXT: retq327 %1 = load <8 x i16>, ptr %a0328 %2 = call <4 x i32> @llvm.x86.xop.vpmadcswd(<8 x i16> %1, <8 x i16> %a1, <4 x i32> %a2)329 ret <4 x i32> %2330}331declare <4 x i32> @llvm.x86.xop.vpmadcswd(<8 x i16>, <8 x i16>, <4 x i32>) nounwind readnone332