116 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py2; RUN: llc < %s -mtriple=x86_64-apple-darwin -mcpu=skx | FileCheck %s3 4declare void @f()5define <4 x i1> @test_4i1(<4 x i32> %a, <4 x i32> %b) {6; CHECK-LABEL: test_4i1:7; CHECK: ## %bb.0:8; CHECK-NEXT: pushq %rax9; CHECK-NEXT: .cfi_def_cfa_offset 1610; CHECK-NEXT: vpcmpnleud %xmm1, %xmm0, %k011; CHECK-NEXT: vpcmpgtd %xmm1, %xmm0, %k112; CHECK-NEXT: korw %k1, %k0, %k013; CHECK-NEXT: kmovw %k0, {{[-0-9]+}}(%r{{[sb]}}p) ## 2-byte Spill14; CHECK-NEXT: callq _f15; CHECK-NEXT: kmovw {{[-0-9]+}}(%r{{[sb]}}p), %k0 ## 2-byte Reload16; CHECK-NEXT: vpmovm2d %k0, %xmm017; CHECK-NEXT: popq %rax18; CHECK-NEXT: retq19 20 %cmp_res = icmp ugt <4 x i32> %a, %b21 %cmp_res2 = icmp sgt <4 x i32> %a, %b22 call void @f()23 %res = or <4 x i1> %cmp_res, %cmp_res224 ret <4 x i1> %res25}26 27define <8 x i1> @test_8i1(<8 x i32> %a, <8 x i32> %b) {28; CHECK-LABEL: test_8i1:29; CHECK: ## %bb.0:30; CHECK-NEXT: pushq %rax31; CHECK-NEXT: .cfi_def_cfa_offset 1632; CHECK-NEXT: vpcmpnleud %ymm1, %ymm0, %k033; CHECK-NEXT: vpcmpgtd %ymm1, %ymm0, %k134; CHECK-NEXT: korb %k1, %k0, %k035; CHECK-NEXT: kmovw %k0, {{[-0-9]+}}(%r{{[sb]}}p) ## 2-byte Spill36; CHECK-NEXT: vzeroupper37; CHECK-NEXT: callq _f38; CHECK-NEXT: kmovw {{[-0-9]+}}(%r{{[sb]}}p), %k0 ## 2-byte Reload39; CHECK-NEXT: vpmovm2w %k0, %xmm040; CHECK-NEXT: popq %rax41; CHECK-NEXT: retq42 43 %cmp_res = icmp ugt <8 x i32> %a, %b44 %cmp_res2 = icmp sgt <8 x i32> %a, %b45 call void @f()46 %res = or <8 x i1> %cmp_res, %cmp_res247 ret <8 x i1> %res48}49 50define <16 x i1> @test_16i1(<16 x i32> %a, <16 x i32> %b) {51; CHECK-LABEL: test_16i1:52; CHECK: ## %bb.0:53; CHECK-NEXT: pushq %rax54; CHECK-NEXT: .cfi_def_cfa_offset 1655; CHECK-NEXT: vpcmpnleud %zmm1, %zmm0, %k056; CHECK-NEXT: vpcmpgtd %zmm1, %zmm0, %k157; CHECK-NEXT: korw %k1, %k0, %k058; CHECK-NEXT: kmovw %k0, {{[-0-9]+}}(%r{{[sb]}}p) ## 2-byte Spill59; CHECK-NEXT: vzeroupper60; CHECK-NEXT: callq _f61; CHECK-NEXT: kmovw {{[-0-9]+}}(%r{{[sb]}}p), %k0 ## 2-byte Reload62; CHECK-NEXT: vpmovm2b %k0, %xmm063; CHECK-NEXT: popq %rax64; CHECK-NEXT: retq65 %cmp_res = icmp ugt <16 x i32> %a, %b66 %cmp_res2 = icmp sgt <16 x i32> %a, %b67 call void @f()68 %res = or <16 x i1> %cmp_res, %cmp_res269 ret <16 x i1> %res70}71 72define <32 x i1> @test_32i1(<32 x i16> %a, <32 x i16> %b) {73; CHECK-LABEL: test_32i1:74; CHECK: ## %bb.0:75; CHECK-NEXT: pushq %rax76; CHECK-NEXT: .cfi_def_cfa_offset 1677; CHECK-NEXT: vpcmpnleuw %zmm1, %zmm0, %k078; CHECK-NEXT: vpcmpgtw %zmm1, %zmm0, %k179; CHECK-NEXT: kord %k1, %k0, %k080; CHECK-NEXT: kmovd %k0, {{[-0-9]+}}(%r{{[sb]}}p) ## 4-byte Spill81; CHECK-NEXT: vzeroupper82; CHECK-NEXT: callq _f83; CHECK-NEXT: kmovd {{[-0-9]+}}(%r{{[sb]}}p), %k0 ## 4-byte Reload84; CHECK-NEXT: vpmovm2b %k0, %ymm085; CHECK-NEXT: popq %rax86; CHECK-NEXT: retq87 %cmp_res = icmp ugt <32 x i16> %a, %b88 %cmp_res2 = icmp sgt <32 x i16> %a, %b89 call void @f()90 %res = or <32 x i1> %cmp_res, %cmp_res291 ret <32 x i1> %res92}93 94define <64 x i1> @test_64i1(<64 x i8> %a, <64 x i8> %b) {95; CHECK-LABEL: test_64i1:96; CHECK: ## %bb.0:97; CHECK-NEXT: pushq %rax98; CHECK-NEXT: .cfi_def_cfa_offset 1699; CHECK-NEXT: vpcmpnleub %zmm1, %zmm0, %k0100; CHECK-NEXT: vpcmpgtb %zmm1, %zmm0, %k1101; CHECK-NEXT: korq %k1, %k0, %k0102; CHECK-NEXT: kmovq %k0, (%rsp) ## 8-byte Spill103; CHECK-NEXT: vzeroupper104; CHECK-NEXT: callq _f105; CHECK-NEXT: kmovq (%rsp), %k0 ## 8-byte Reload106; CHECK-NEXT: vpmovm2b %k0, %zmm0107; CHECK-NEXT: popq %rax108; CHECK-NEXT: retq109 110 %cmp_res = icmp ugt <64 x i8> %a, %b111 %cmp_res2 = icmp sgt <64 x i8> %a, %b112 call void @f()113 %res = or <64 x i1> %cmp_res, %cmp_res2114 ret <64 x i1> %res115}116