1 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
2 ; RUN: llc < %s -mtriple=i686-unknown-unknown -mattr=avx2 | FileCheck %s --check-prefixes=X86,X86_AVX256
3 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=avx2 | FileCheck %s --check-prefixes=X64,X64_AVX256
4 ; RUN: llc < %s -mtriple=i686-unknown-unknown -mattr=avx512f | FileCheck %s --check-prefixes=X86,X86_AVX512
5 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=avx512f | FileCheck %s --check-prefixes=X64,X64_AVX512
7 define <8 x float> @insert_subvector_256(i16 %x0, i16 %x1, <8 x float> %v) nounwind {
8 ; X86-LABEL: insert_subvector_256:
10 ; X86-NEXT: vmovd {{.*#+}} xmm1 = mem[0],zero,zero,zero
11 ; X86-NEXT: vpinsrw $1, {{[0-9]+}}(%esp), %xmm1, %xmm1
12 ; X86-NEXT: vpbroadcastd %xmm1, %xmm1
13 ; X86-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0],ymm1[1],ymm0[2,3,4,5,6,7]
16 ; X64-LABEL: insert_subvector_256:
18 ; X64-NEXT: vmovd %edi, %xmm1
19 ; X64-NEXT: vpinsrw $1, %esi, %xmm1, %xmm1
20 ; X64-NEXT: vpbroadcastd %xmm1, %xmm1
21 ; X64-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0],ymm1[1],ymm0[2,3,4,5,6,7]
23 %ins1 = insertelement <2 x i16> undef, i16 %x0, i32 0
24 %ins2 = insertelement <2 x i16> %ins1, i16 %x1, i32 1
25 %bc = bitcast <2 x i16> %ins2 to float
26 %ins3 = insertelement <8 x float> %v, float %bc, i32 1
30 define <8 x i64> @insert_subvector_512(i32 %x0, i32 %x1, <8 x i64> %v) nounwind {
31 ; X86_AVX256-LABEL: insert_subvector_512:
32 ; X86_AVX256: # %bb.0:
33 ; X86_AVX256-NEXT: vbroadcastss {{[0-9]+}}(%esp), %ymm2
34 ; X86_AVX256-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4],ymm0[5,6,7]
35 ; X86_AVX256-NEXT: vbroadcastss {{[0-9]+}}(%esp), %ymm2
36 ; X86_AVX256-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm2[5],ymm0[6,7]
37 ; X86_AVX256-NEXT: retl
39 ; X64_AVX256-LABEL: insert_subvector_512:
40 ; X64_AVX256: # %bb.0:
41 ; X64_AVX256-NEXT: vmovd %edi, %xmm2
42 ; X64_AVX256-NEXT: vpinsrd $1, %esi, %xmm2, %xmm2
43 ; X64_AVX256-NEXT: vpbroadcastq %xmm2, %ymm2
44 ; X64_AVX256-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5],ymm0[6,7]
45 ; X64_AVX256-NEXT: retq
47 ; X86_AVX512-LABEL: insert_subvector_512:
48 ; X86_AVX512: # %bb.0:
49 ; X86_AVX512-NEXT: vmovq {{.*#+}} xmm1 = mem[0],zero
50 ; X86_AVX512-NEXT: vpmovsxbq {{.*#+}} zmm2 = [0,1,8,3,4,5,6,7]
51 ; X86_AVX512-NEXT: vpermt2q %zmm1, %zmm2, %zmm0
52 ; X86_AVX512-NEXT: retl
54 ; X64_AVX512-LABEL: insert_subvector_512:
55 ; X64_AVX512: # %bb.0:
56 ; X64_AVX512-NEXT: vmovd %edi, %xmm1
57 ; X64_AVX512-NEXT: vpinsrd $1, %esi, %xmm1, %xmm1
58 ; X64_AVX512-NEXT: vpmovsxbq {{.*#+}} zmm2 = [0,1,8,3,4,5,6,7]
59 ; X64_AVX512-NEXT: vpermt2q %zmm1, %zmm2, %zmm0
60 ; X64_AVX512-NEXT: retq
61 %ins1 = insertelement <2 x i32> undef, i32 %x0, i32 0
62 %ins2 = insertelement <2 x i32> %ins1, i32 %x1, i32 1
63 %bc = bitcast <2 x i32> %ins2 to i64
64 %ins3 = insertelement <8 x i64> %v, i64 %bc, i32 2
68 ; PR34716 - https://bugs.llvm.org/show_bug.cgi?id=34716
69 ; Special case: if we're inserting into an undef vector, we can optimize more.
71 define <8 x i64> @insert_subvector_into_undef(i32 %x0, i32 %x1) nounwind {
72 ; X86_AVX256-LABEL: insert_subvector_into_undef:
73 ; X86_AVX256: # %bb.0:
74 ; X86_AVX256-NEXT: vbroadcastsd {{[0-9]+}}(%esp), %ymm0
75 ; X86_AVX256-NEXT: vmovaps %ymm0, %ymm1
76 ; X86_AVX256-NEXT: retl
78 ; X64_AVX256-LABEL: insert_subvector_into_undef:
79 ; X64_AVX256: # %bb.0:
80 ; X64_AVX256-NEXT: vmovd %edi, %xmm0
81 ; X64_AVX256-NEXT: vpinsrd $1, %esi, %xmm0, %xmm0
82 ; X64_AVX256-NEXT: vpbroadcastq %xmm0, %ymm0
83 ; X64_AVX256-NEXT: vmovdqa %ymm0, %ymm1
84 ; X64_AVX256-NEXT: retq
86 ; X86_AVX512-LABEL: insert_subvector_into_undef:
87 ; X86_AVX512: # %bb.0:
88 ; X86_AVX512-NEXT: vbroadcastsd {{[0-9]+}}(%esp), %zmm0
89 ; X86_AVX512-NEXT: retl
91 ; X64_AVX512-LABEL: insert_subvector_into_undef:
92 ; X64_AVX512: # %bb.0:
93 ; X64_AVX512-NEXT: vmovd %edi, %xmm0
94 ; X64_AVX512-NEXT: vpinsrd $1, %esi, %xmm0, %xmm0
95 ; X64_AVX512-NEXT: vpbroadcastq %xmm0, %zmm0
96 ; X64_AVX512-NEXT: retq
97 %ins1 = insertelement <2 x i32> undef, i32 %x0, i32 0
98 %ins2 = insertelement <2 x i32> %ins1, i32 %x1, i32 1
99 %bc = bitcast <2 x i32> %ins2 to i64
100 %ins3 = insertelement <8 x i64> undef, i64 %bc, i32 0
101 %splat = shufflevector <8 x i64> %ins3, <8 x i64> undef, <8 x i32> zeroinitializer