170 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py2; RUN: llc -O3 -disable-peephole -mtriple=x86_64-unknown-unknown -mattr=+avx512vp2intersect,+avx512vl < %s | FileCheck %s3 4target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"5target triple = "x86_64-unknown-unknown"6 7define void @stack_fold_vp2intersectd(ptr %a, <16 x i32> %b, ptr nocapture %m0, ptr nocapture %m1) {8; CHECK-LABEL: stack_fold_vp2intersectd:9; CHECK: # %bb.0:10; CHECK-NEXT: vmovups %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11; CHECK-NEXT: #APP12; CHECK-NEXT: nop13; CHECK-NEXT: #NO_APP14; CHECK-NEXT: vmovdqa64 (%rdi), %zmm015; CHECK-NEXT: vp2intersectd {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %k0 # 64-byte Folded Reload16; CHECK-NEXT: kmovw %k0, (%rsi)17; CHECK-NEXT: kmovw %k1, (%rdx)18; CHECK-NEXT: vzeroupper19; CHECK-NEXT: retq20 %1 = tail call <2 x i64> asm sideeffect "nop", "=x,~{xmm1},~{xmm2},~{xmm3},~{xmm4},~{xmm5},~{xmm6},~{xmm7},~{xmm8},~{xmm9},~{xmm10},~{xmm11},~{xmm12},~{xmm13},~{xmm14},~{xmm15},~{xmm16},~{xmm17},~{xmm18},~{xmm19},~{xmm20},~{xmm21},~{xmm22},~{xmm23},~{xmm24},~{xmm25},~{xmm26},~{xmm27},~{xmm28},~{xmm29},~{xmm30},~{xmm31},~{flags}"()21 %2 = load <16 x i32>, ptr %a22 %3 = tail call { <16 x i1>, <16 x i1> } @llvm.x86.avx512.vp2intersect.d.512(<16 x i32> %2, <16 x i32> %b)23 %4 = extractvalue { <16 x i1>, <16 x i1> } %3, 024 store <16 x i1> %4, ptr %m025 %5 = extractvalue { <16 x i1>, <16 x i1> } %3, 126 store <16 x i1> %5, ptr %m127 ret void28}29declare { <16 x i1>, <16 x i1> } @llvm.x86.avx512.vp2intersect.d.512(<16 x i32>, <16 x i32>)30 31define void @stack_fold_vp2intersectq(ptr %a, <8 x i64> %b, ptr nocapture %m0, ptr nocapture %m1) {32; CHECK-LABEL: stack_fold_vp2intersectq:33; CHECK: # %bb.0:34; CHECK-NEXT: vmovups %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill35; CHECK-NEXT: #APP36; CHECK-NEXT: nop37; CHECK-NEXT: #NO_APP38; CHECK-NEXT: vmovdqa64 (%rdi), %zmm039; CHECK-NEXT: vp2intersectq {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %k0 # 64-byte Folded Reload40; CHECK-NEXT: kmovw %k1, %eax41; CHECK-NEXT: kmovw %k0, %ecx42; CHECK-NEXT: movb %cl, (%rsi)43; CHECK-NEXT: movb %al, (%rdx)44; CHECK-NEXT: vzeroupper45; CHECK-NEXT: retq46 %1 = tail call <2 x i64> asm sideeffect "nop", "=x,~{xmm1},~{xmm2},~{xmm3},~{xmm4},~{xmm5},~{xmm6},~{xmm7},~{xmm8},~{xmm9},~{xmm10},~{xmm11},~{xmm12},~{xmm13},~{xmm14},~{xmm15},~{xmm16},~{xmm17},~{xmm18},~{xmm19},~{xmm20},~{xmm21},~{xmm22},~{xmm23},~{xmm24},~{xmm25},~{xmm26},~{xmm27},~{xmm28},~{xmm29},~{xmm30},~{xmm31},~{flags}"()47 %2 = load <8 x i64>, ptr %a48 %3 = tail call { <8 x i1>, <8 x i1> } @llvm.x86.avx512.vp2intersect.q.512(<8 x i64> %2, <8 x i64> %b)49 %4 = extractvalue { <8 x i1>, <8 x i1> } %3, 050 store <8 x i1> %4, ptr %m051 %5 = extractvalue { <8 x i1>, <8 x i1> } %3, 152 store <8 x i1> %5, ptr %m153 ret void54}55declare { <8 x i1>, <8 x i1> } @llvm.x86.avx512.vp2intersect.q.512(<8 x i64>, <8 x i64>)56 57define void @stack_fold_vp2intersectd_256(ptr %a, <8 x i32> %b, ptr nocapture %m0, ptr nocapture %m1) {58; CHECK-LABEL: stack_fold_vp2intersectd_256:59; CHECK: # %bb.0:60; CHECK-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill61; CHECK-NEXT: #APP62; CHECK-NEXT: nop63; CHECK-NEXT: #NO_APP64; CHECK-NEXT: vmovdqa (%rdi), %ymm065; CHECK-NEXT: vp2intersectd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %k0 # 32-byte Folded Reload66; CHECK-NEXT: kmovw %k1, %eax67; CHECK-NEXT: kmovw %k0, %ecx68; CHECK-NEXT: movb %cl, (%rsi)69; CHECK-NEXT: movb %al, (%rdx)70; CHECK-NEXT: vzeroupper71; CHECK-NEXT: retq72 %1 = tail call <2 x i64> asm sideeffect "nop", "=x,~{xmm1},~{xmm2},~{xmm3},~{xmm4},~{xmm5},~{xmm6},~{xmm7},~{xmm8},~{xmm9},~{xmm10},~{xmm11},~{xmm12},~{xmm13},~{xmm14},~{xmm15},~{xmm16},~{xmm17},~{xmm18},~{xmm19},~{xmm20},~{xmm21},~{xmm22},~{xmm23},~{xmm24},~{xmm25},~{xmm26},~{xmm27},~{xmm28},~{xmm29},~{xmm30},~{xmm31},~{flags}"()73 %2 = load <8 x i32>, ptr %a74 %3 = tail call { <8 x i1>, <8 x i1> } @llvm.x86.avx512.vp2intersect.d.256(<8 x i32> %2, <8 x i32> %b)75 %4 = extractvalue { <8 x i1>, <8 x i1> } %3, 076 store <8 x i1> %4, ptr %m077 %5 = extractvalue { <8 x i1>, <8 x i1> } %3, 178 store <8 x i1> %5, ptr %m179 ret void80}81declare { <8 x i1>, <8 x i1> } @llvm.x86.avx512.vp2intersect.d.256(<8 x i32>, <8 x i32>)82 83define void @stack_fold_vp2intersectq_256(ptr %a, <4 x i64> %b, ptr nocapture %m0, ptr nocapture %m1) {84; CHECK-LABEL: stack_fold_vp2intersectq_256:85; CHECK: # %bb.0:86; CHECK-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill87; CHECK-NEXT: #APP88; CHECK-NEXT: nop89; CHECK-NEXT: #NO_APP90; CHECK-NEXT: vmovdqa (%rdi), %ymm091; CHECK-NEXT: vp2intersectq {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %k0 # 32-byte Folded Reload92; CHECK-NEXT: kshiftlw $12, %k0, %k293; CHECK-NEXT: kshiftrw $12, %k2, %k294; CHECK-NEXT: kmovw %k2, %eax95; CHECK-NEXT: movb %al, (%rsi)96; CHECK-NEXT: kshiftlw $12, %k1, %k097; CHECK-NEXT: kshiftrw $12, %k0, %k098; CHECK-NEXT: kmovw %k0, %eax99; CHECK-NEXT: movb %al, (%rdx)100; CHECK-NEXT: vzeroupper101; CHECK-NEXT: retq102 %1 = tail call <2 x i64> asm sideeffect "nop", "=x,~{xmm1},~{xmm2},~{xmm3},~{xmm4},~{xmm5},~{xmm6},~{xmm7},~{xmm8},~{xmm9},~{xmm10},~{xmm11},~{xmm12},~{xmm13},~{xmm14},~{xmm15},~{xmm16},~{xmm17},~{xmm18},~{xmm19},~{xmm20},~{xmm21},~{xmm22},~{xmm23},~{xmm24},~{xmm25},~{xmm26},~{xmm27},~{xmm28},~{xmm29},~{xmm30},~{xmm31},~{flags}"()103 %2 = load <4 x i64>, ptr %a104 %3 = tail call { <4 x i1>, <4 x i1> } @llvm.x86.avx512.vp2intersect.q.256(<4 x i64> %2, <4 x i64> %b)105 %4 = extractvalue { <4 x i1>, <4 x i1> } %3, 0106 store <4 x i1> %4, ptr %m0107 %5 = extractvalue { <4 x i1>, <4 x i1> } %3, 1108 store <4 x i1> %5, ptr %m1109 ret void110}111declare { <4 x i1>, <4 x i1> } @llvm.x86.avx512.vp2intersect.q.256(<4 x i64>, <4 x i64>)112 113define void @stack_fold_vp2intersectd_128(ptr %a, <4 x i32> %b, ptr nocapture %m0, ptr nocapture %m1) {114; CHECK-LABEL: stack_fold_vp2intersectd_128:115; CHECK: # %bb.0:116; CHECK-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill117; CHECK-NEXT: #APP118; CHECK-NEXT: nop119; CHECK-NEXT: #NO_APP120; CHECK-NEXT: vmovdqa (%rdi), %xmm0121; CHECK-NEXT: vp2intersectd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %k0 # 16-byte Folded Reload122; CHECK-NEXT: kshiftlw $12, %k0, %k2123; CHECK-NEXT: kshiftrw $12, %k2, %k2124; CHECK-NEXT: kmovw %k2, %eax125; CHECK-NEXT: movb %al, (%rsi)126; CHECK-NEXT: kshiftlw $12, %k1, %k0127; CHECK-NEXT: kshiftrw $12, %k0, %k0128; CHECK-NEXT: kmovw %k0, %eax129; CHECK-NEXT: movb %al, (%rdx)130; CHECK-NEXT: retq131 %1 = tail call <2 x i64> asm sideeffect "nop", "=x,~{xmm1},~{xmm2},~{xmm3},~{xmm4},~{xmm5},~{xmm6},~{xmm7},~{xmm8},~{xmm9},~{xmm10},~{xmm11},~{xmm12},~{xmm13},~{xmm14},~{xmm15},~{xmm16},~{xmm17},~{xmm18},~{xmm19},~{xmm20},~{xmm21},~{xmm22},~{xmm23},~{xmm24},~{xmm25},~{xmm26},~{xmm27},~{xmm28},~{xmm29},~{xmm30},~{xmm31},~{flags}"()132 %2 = load <4 x i32>, ptr %a133 %3 = tail call { <4 x i1>, <4 x i1> } @llvm.x86.avx512.vp2intersect.d.128(<4 x i32> %2, <4 x i32> %b)134 %4 = extractvalue { <4 x i1>, <4 x i1> } %3, 0135 store <4 x i1> %4, ptr %m0136 %5 = extractvalue { <4 x i1>, <4 x i1> } %3, 1137 store <4 x i1> %5, ptr %m1138 ret void139}140declare { <4 x i1>, <4 x i1> } @llvm.x86.avx512.vp2intersect.d.128(<4 x i32>, <4 x i32>)141 142define void @stack_fold_vp2intersectq_128(ptr %a, <2 x i64> %b, ptr nocapture %m0, ptr nocapture %m1) {143; CHECK-LABEL: stack_fold_vp2intersectq_128:144; CHECK: # %bb.0:145; CHECK-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill146; CHECK-NEXT: #APP147; CHECK-NEXT: nop148; CHECK-NEXT: #NO_APP149; CHECK-NEXT: vmovdqa (%rdi), %xmm0150; CHECK-NEXT: vp2intersectq {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %k0 # 16-byte Folded Reload151; CHECK-NEXT: kshiftlw $14, %k0, %k2152; CHECK-NEXT: kshiftrw $14, %k2, %k2153; CHECK-NEXT: kmovw %k2, %eax154; CHECK-NEXT: movb %al, (%rsi)155; CHECK-NEXT: kshiftlw $14, %k1, %k0156; CHECK-NEXT: kshiftrw $14, %k0, %k0157; CHECK-NEXT: kmovw %k0, %eax158; CHECK-NEXT: movb %al, (%rdx)159; CHECK-NEXT: retq160 %1 = tail call <2 x i64> asm sideeffect "nop", "=x,~{xmm1},~{xmm2},~{xmm3},~{xmm4},~{xmm5},~{xmm6},~{xmm7},~{xmm8},~{xmm9},~{xmm10},~{xmm11},~{xmm12},~{xmm13},~{xmm14},~{xmm15},~{xmm16},~{xmm17},~{xmm18},~{xmm19},~{xmm20},~{xmm21},~{xmm22},~{xmm23},~{xmm24},~{xmm25},~{xmm26},~{xmm27},~{xmm28},~{xmm29},~{xmm30},~{xmm31},~{flags}"()161 %2 = load <2 x i64>, ptr %a162 %3 = tail call { <2 x i1>, <2 x i1> } @llvm.x86.avx512.vp2intersect.q.128(<2 x i64> %2, <2 x i64> %b)163 %4 = extractvalue { <2 x i1>, <2 x i1> } %3, 0164 store <2 x i1> %4, ptr %m0165 %5 = extractvalue { <2 x i1>, <2 x i1> } %3, 1166 store <2 x i1> %5, ptr %m1167 ret void168}169declare { <2 x i1>, <2 x i1> } @llvm.x86.avx512.vp2intersect.q.128(<2 x i64>, <2 x i64>)170