1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py 2; RUN: llc < %s -mtriple=i686-unknown -mattr=+avx512cd,+avx512vl | FileCheck %s --check-prefixes=CHECK,X86 3; RUN: llc < %s -mtriple=x86_64-unknown -mattr=+avx512cd,+avx512vl | FileCheck %s --check-prefixes=CHECK,X64 4 5declare <4 x i32> @llvm.x86.avx512.mask.lzcnt.d.128(<4 x i32>, <4 x i32>, i8) 6 7define <4 x i32>@test_int_x86_avx512_mask_vplzcnt_d_128(<4 x i32> %x0, <4 x i32> %x1, i8 %x2) { 8; X86-LABEL: test_int_x86_avx512_mask_vplzcnt_d_128: 9; X86: # %bb.0: 10; X86-NEXT: vplzcntd %xmm0, %xmm2 11; X86-NEXT: movzbl {{[0-9]+}}(%esp), %eax 12; X86-NEXT: kmovw %eax, %k1 13; X86-NEXT: vplzcntd %xmm0, %xmm1 {%k1} 14; X86-NEXT: vplzcntd %xmm0, %xmm0 {%k1} {z} 15; X86-NEXT: vpaddd %xmm0, %xmm2, %xmm0 16; X86-NEXT: vpaddd %xmm0, %xmm1, %xmm0 17; X86-NEXT: retl 18; 19; X64-LABEL: test_int_x86_avx512_mask_vplzcnt_d_128: 20; X64: # %bb.0: 21; X64-NEXT: vplzcntd %xmm0, %xmm2 22; X64-NEXT: kmovw %edi, %k1 23; X64-NEXT: vplzcntd %xmm0, %xmm1 {%k1} 24; X64-NEXT: vplzcntd %xmm0, %xmm0 {%k1} {z} 25; X64-NEXT: vpaddd %xmm0, %xmm2, %xmm0 26; X64-NEXT: vpaddd %xmm0, %xmm1, %xmm0 27; X64-NEXT: retq 28 %res = call <4 x i32> @llvm.x86.avx512.mask.lzcnt.d.128(<4 x i32> %x0, <4 x i32> %x1, i8 %x2) 29 %res1 = call <4 x i32> @llvm.x86.avx512.mask.lzcnt.d.128(<4 x i32> %x0, <4 x i32> %x1, i8 -1) 30 %res3 = call <4 x i32> @llvm.x86.avx512.mask.lzcnt.d.128(<4 x i32> %x0, <4 x i32> zeroinitializer, i8 %x2) 31 %res2 = add <4 x i32> %res, %res1 32 %res4 = add <4 x i32> %res2, %res3 33 ret <4 x i32> %res4 34} 35 36declare <8 x i32> @llvm.x86.avx512.mask.lzcnt.d.256(<8 x i32>, <8 x i32>, i8) 37 38define <8 x i32>@test_int_x86_avx512_mask_vplzcnt_d_256(<8 x i32> %x0, <8 x i32> %x1, i8 %x2) { 39; X86-LABEL: test_int_x86_avx512_mask_vplzcnt_d_256: 40; X86: # %bb.0: 41; X86-NEXT: vplzcntd %ymm0, %ymm2 42; X86-NEXT: movzbl {{[0-9]+}}(%esp), %eax 43; X86-NEXT: kmovw %eax, %k1 44; X86-NEXT: vplzcntd %ymm0, %ymm1 {%k1} 45; X86-NEXT: vpaddd %ymm2, %ymm1, %ymm0 46; X86-NEXT: retl 47; 48; X64-LABEL: test_int_x86_avx512_mask_vplzcnt_d_256: 49; X64: # %bb.0: 50; X64-NEXT: vplzcntd %ymm0, %ymm2 51; X64-NEXT: kmovw %edi, %k1 52; X64-NEXT: vplzcntd %ymm0, %ymm1 {%k1} 53; X64-NEXT: vpaddd %ymm2, %ymm1, %ymm0 54; X64-NEXT: retq 55 %res = call <8 x i32> @llvm.x86.avx512.mask.lzcnt.d.256(<8 x i32> %x0, <8 x i32> %x1, i8 %x2) 56 %res1 = call <8 x i32> @llvm.x86.avx512.mask.lzcnt.d.256(<8 x i32> %x0, <8 x i32> %x1, i8 -1) 57 %res2 = add <8 x i32> %res, %res1 58 ret <8 x i32> %res2 59} 60 61declare <2 x i64> @llvm.x86.avx512.mask.lzcnt.q.128(<2 x i64>, <2 x i64>, i8) 62 63define <2 x i64>@test_int_x86_avx512_mask_vplzcnt_q_128(<2 x i64> %x0, <2 x i64> %x1, i8 %x2) { 64; X86-LABEL: test_int_x86_avx512_mask_vplzcnt_q_128: 65; X86: # %bb.0: 66; X86-NEXT: vplzcntq %xmm0, %xmm2 67; X86-NEXT: movzbl {{[0-9]+}}(%esp), %eax 68; X86-NEXT: kmovw %eax, %k1 69; X86-NEXT: vplzcntq %xmm0, %xmm1 {%k1} 70; X86-NEXT: vpaddq %xmm2, %xmm1, %xmm0 71; X86-NEXT: retl 72; 73; X64-LABEL: test_int_x86_avx512_mask_vplzcnt_q_128: 74; X64: # %bb.0: 75; X64-NEXT: vplzcntq %xmm0, %xmm2 76; X64-NEXT: kmovw %edi, %k1 77; X64-NEXT: vplzcntq %xmm0, %xmm1 {%k1} 78; X64-NEXT: vpaddq %xmm2, %xmm1, %xmm0 79; X64-NEXT: retq 80 %res = call <2 x i64> @llvm.x86.avx512.mask.lzcnt.q.128(<2 x i64> %x0, <2 x i64> %x1, i8 %x2) 81 %res1 = call <2 x i64> @llvm.x86.avx512.mask.lzcnt.q.128(<2 x i64> %x0, <2 x i64> %x1, i8 -1) 82 %res2 = add <2 x i64> %res, %res1 83 ret <2 x i64> %res2 84} 85 86declare <4 x i64> @llvm.x86.avx512.mask.lzcnt.q.256(<4 x i64>, <4 x i64>, i8) 87 88define <4 x i64>@test_int_x86_avx512_mask_vplzcnt_q_256(<4 x i64> %x0, <4 x i64> %x1, i8 %x2) { 89; X86-LABEL: test_int_x86_avx512_mask_vplzcnt_q_256: 90; X86: # %bb.0: 91; X86-NEXT: vplzcntq %ymm0, %ymm2 92; X86-NEXT: movzbl {{[0-9]+}}(%esp), %eax 93; X86-NEXT: kmovw %eax, %k1 94; X86-NEXT: vplzcntq %ymm0, %ymm1 {%k1} 95; X86-NEXT: vpaddq %ymm2, %ymm1, %ymm0 96; X86-NEXT: retl 97; 98; X64-LABEL: test_int_x86_avx512_mask_vplzcnt_q_256: 99; X64: # %bb.0: 100; X64-NEXT: vplzcntq %ymm0, %ymm2 101; X64-NEXT: kmovw %edi, %k1 102; X64-NEXT: vplzcntq %ymm0, %ymm1 {%k1} 103; X64-NEXT: vpaddq %ymm2, %ymm1, %ymm0 104; X64-NEXT: retq 105 %res = call <4 x i64> @llvm.x86.avx512.mask.lzcnt.q.256(<4 x i64> %x0, <4 x i64> %x1, i8 %x2) 106 %res1 = call <4 x i64> @llvm.x86.avx512.mask.lzcnt.q.256(<4 x i64> %x0, <4 x i64> %x1, i8 -1) 107 %res2 = add <4 x i64> %res, %res1 108 ret <4 x i64> %res2 109} 110 111define <8 x i32> @test_x86_vbroadcastmw_256(i16 %a0) { 112; X86-LABEL: test_x86_vbroadcastmw_256: 113; X86: # %bb.0: 114; X86-NEXT: movzwl {{[0-9]+}}(%esp), %eax 115; X86-NEXT: vpbroadcastd %eax, %ymm0 116; X86-NEXT: retl 117; 118; X64-LABEL: test_x86_vbroadcastmw_256: 119; X64: # %bb.0: 120; X64-NEXT: movzwl %di, %eax 121; X64-NEXT: vpbroadcastd %eax, %ymm0 122; X64-NEXT: retq 123 %res = call <8 x i32> @llvm.x86.avx512.broadcastmw.256(i16 %a0) ; 124 ret <8 x i32> %res 125} 126declare <8 x i32> @llvm.x86.avx512.broadcastmw.256(i16) 127 128define <4 x i32> @test_x86_vbroadcastmw_128(i16 %a0) { 129; X86-LABEL: test_x86_vbroadcastmw_128: 130; X86: # %bb.0: 131; X86-NEXT: movzwl {{[0-9]+}}(%esp), %eax 132; X86-NEXT: vpbroadcastd %eax, %xmm0 133; X86-NEXT: retl 134; 135; X64-LABEL: test_x86_vbroadcastmw_128: 136; X64: # %bb.0: 137; X64-NEXT: movzwl %di, %eax 138; X64-NEXT: vpbroadcastd %eax, %xmm0 139; X64-NEXT: retq 140 %res = call <4 x i32> @llvm.x86.avx512.broadcastmw.128(i16 %a0) ; 141 ret <4 x i32> %res 142} 143declare <4 x i32> @llvm.x86.avx512.broadcastmw.128(i16) 144 145define <4 x i64> @test_x86_broadcastmb_256(i8 %a0) { 146; X86-LABEL: test_x86_broadcastmb_256: 147; X86: # %bb.0: 148; X86-NEXT: movzbl {{[0-9]+}}(%esp), %eax 149; X86-NEXT: vmovd %eax, %xmm0 150; X86-NEXT: vpbroadcastq %xmm0, %ymm0 151; X86-NEXT: retl 152; 153; X64-LABEL: test_x86_broadcastmb_256: 154; X64: # %bb.0: 155; X64-NEXT: # kill: def $edi killed $edi def $rdi 156; X64-NEXT: movzbl %dil, %eax 157; X64-NEXT: vpbroadcastq %rax, %ymm0 158; X64-NEXT: retq 159 %res = call <4 x i64> @llvm.x86.avx512.broadcastmb.256(i8 %a0) ; 160 ret <4 x i64> %res 161} 162declare <4 x i64> @llvm.x86.avx512.broadcastmb.256(i8) 163 164define <2 x i64> @test_x86_broadcastmb_128(i8 %a0) { 165; X86-LABEL: test_x86_broadcastmb_128: 166; X86: # %bb.0: 167; X86-NEXT: movzbl {{[0-9]+}}(%esp), %eax 168; X86-NEXT: vmovd %eax, %xmm0 169; X86-NEXT: vpbroadcastq %xmm0, %xmm0 170; X86-NEXT: retl 171; 172; X64-LABEL: test_x86_broadcastmb_128: 173; X64: # %bb.0: 174; X64-NEXT: # kill: def $edi killed $edi def $rdi 175; X64-NEXT: movzbl %dil, %eax 176; X64-NEXT: vpbroadcastq %rax, %xmm0 177; X64-NEXT: retq 178 %res = call <2 x i64> @llvm.x86.avx512.broadcastmb.128(i8 %a0) ; 179 ret <2 x i64> %res 180} 181declare <2 x i64> @llvm.x86.avx512.broadcastmb.128(i8) 182 183