xref: /llvm-project/llvm/test/CodeGen/X86/avx512-broadcast-arith.ll (revision 55c6bda01ef5a166a69b43956775272d9d67bda5)
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
2; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512f           | FileCheck %s --check-prefixes=AVX512F
3; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512f,+avx512bw | FileCheck %s --check-prefixes=AVX512BW
4
5; PR34666
6define <64 x i8> @add_v64i8_broadcasts(<64 x i8> %a0, i64 %a1, i8 %a2) {
7; AVX512F-LABEL: add_v64i8_broadcasts:
8; AVX512F:       # %bb.0:
9; AVX512F-NEXT:    movq %rdi, %rax
10; AVX512F-NEXT:    movl %edi, %ecx
11; AVX512F-NEXT:    kmovw %edi, %k1
12; AVX512F-NEXT:    shrq $32, %rdi
13; AVX512F-NEXT:    shrq $48, %rax
14; AVX512F-NEXT:    shrl $16, %ecx
15; AVX512F-NEXT:    vmovd %esi, %xmm1
16; AVX512F-NEXT:    vpbroadcastb %xmm1, %ymm1
17; AVX512F-NEXT:    vinserti64x4 $1, %ymm1, %zmm1, %zmm1
18; AVX512F-NEXT:    kmovw %ecx, %k2
19; AVX512F-NEXT:    kmovw %eax, %k3
20; AVX512F-NEXT:    kmovw %edi, %k4
21; AVX512F-NEXT:    vpternlogd $255, %zmm2, %zmm2, %zmm2 {%k4} {z}
22; AVX512F-NEXT:    vpmovdb %zmm2, %xmm2
23; AVX512F-NEXT:    vpternlogd $255, %zmm3, %zmm3, %zmm3 {%k3} {z}
24; AVX512F-NEXT:    vpmovdb %zmm3, %xmm3
25; AVX512F-NEXT:    vinserti128 $1, %xmm3, %ymm2, %ymm2
26; AVX512F-NEXT:    vpternlogd $255, %zmm3, %zmm3, %zmm3 {%k1} {z}
27; AVX512F-NEXT:    vpmovdb %zmm3, %xmm3
28; AVX512F-NEXT:    vpternlogd $255, %zmm4, %zmm4, %zmm4 {%k2} {z}
29; AVX512F-NEXT:    vpmovdb %zmm4, %xmm4
30; AVX512F-NEXT:    vinserti128 $1, %xmm4, %ymm3, %ymm3
31; AVX512F-NEXT:    vinserti64x4 $1, %ymm2, %zmm3, %zmm2
32; AVX512F-NEXT:    vpternlogq $216, %zmm2, %zmm1, %zmm0
33; AVX512F-NEXT:    vextracti64x4 $1, %zmm0, %ymm3
34; AVX512F-NEXT:    vpaddb %ymm1, %ymm3, %ymm3
35; AVX512F-NEXT:    vpaddb %ymm1, %ymm0, %ymm0
36; AVX512F-NEXT:    vinserti64x4 $1, %ymm3, %zmm0, %zmm4
37; AVX512F-NEXT:    vpaddb %ymm1, %ymm3, %ymm3
38; AVX512F-NEXT:    vpaddb %ymm1, %ymm0, %ymm0
39; AVX512F-NEXT:    vinserti64x4 $1, %ymm3, %zmm0, %zmm0
40; AVX512F-NEXT:    vpternlogq $226, %zmm4, %zmm2, %zmm0
41; AVX512F-NEXT:    retq
42;
43; AVX512BW-LABEL: add_v64i8_broadcasts:
44; AVX512BW:       # %bb.0:
45; AVX512BW-NEXT:    vpbroadcastb %esi, %zmm1
46; AVX512BW-NEXT:    kmovq %rdi, %k1
47; AVX512BW-NEXT:    vpbroadcastb %esi, %zmm0 {%k1}
48; AVX512BW-NEXT:    vpaddb %zmm1, %zmm0, %zmm0
49; AVX512BW-NEXT:    vpaddb %zmm1, %zmm0, %zmm0 {%k1}
50; AVX512BW-NEXT:    retq
51  %v = insertelement <64 x i8> undef, i8 %a2, i64 0
52  %broadcast = shufflevector <64 x i8> %v, <64 x i8> poison, <64 x i32> zeroinitializer
53  %mask = bitcast i64 %a1 to <64 x i1>
54  %broadcast_mask = select <64 x i1> %mask, <64 x i8> %broadcast, <64 x i8> %a0
55  %broadcast_maskz = select <64 x i1> %mask, <64 x i8> %broadcast, <64 x i8> zeroinitializer
56  %add0 = add <64 x i8> %broadcast_mask, %broadcast
57  %add1 = add <64 x i8> %add0, %broadcast_maskz
58  ret <64 x i8> %add1
59}
60