1 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
2 ; RUN: llc < %s -mtriple=i686-unknown-unknown -mattr=avx | FileCheck %s --check-prefix=X32
3 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=avx | FileCheck %s --check-prefix=X64
5 define void @big_nonzero_16_bytes(i32* nocapture %a) {
6 ; X32-LABEL: big_nonzero_16_bytes:
8 ; X32-NEXT: movl {{[0-9]+}}(%esp), %eax
9 ; X32-NEXT: vmovaps {{.*#+}} xmm0 = [1,2,3,4]
10 ; X32-NEXT: vmovups %xmm0, (%eax)
13 ; X64-LABEL: big_nonzero_16_bytes:
15 ; X64-NEXT: vmovaps {{.*#+}} xmm0 = [1,2,3,4]
16 ; X64-NEXT: vmovups %xmm0, (%rdi)
18 %arrayidx1 = getelementptr inbounds i32, i32* %a, i64 1
19 %arrayidx2 = getelementptr inbounds i32, i32* %a, i64 2
20 %arrayidx3 = getelementptr inbounds i32, i32* %a, i64 3
22 store i32 1, i32* %a, align 4
23 store i32 2, i32* %arrayidx1, align 4
24 store i32 3, i32* %arrayidx2, align 4
25 store i32 4, i32* %arrayidx3, align 4
29 ; TODO: We assumed that two 64-bit stores were better than 1 vector load and 1 vector store.
30 ; But if the 64-bit constants can't be represented as sign-extended 32-bit constants, then
31 ; it takes extra instructions to do this in scalar.
33 define void @big_nonzero_16_bytes_big64bit_constants(i64* nocapture %a) {
34 ; X32-LABEL: big_nonzero_16_bytes_big64bit_constants:
36 ; X32-NEXT: movl {{[0-9]+}}(%esp), %eax
37 ; X32-NEXT: vmovaps {{.*#+}} xmm0 = [1,1,1,3]
38 ; X32-NEXT: vmovups %xmm0, (%eax)
41 ; X64-LABEL: big_nonzero_16_bytes_big64bit_constants:
43 ; X64-NEXT: movabsq $4294967297, %rax # imm = 0x100000001
44 ; X64-NEXT: movq %rax, (%rdi)
45 ; X64-NEXT: movabsq $12884901889, %rax # imm = 0x300000001
46 ; X64-NEXT: movq %rax, 8(%rdi)
48 %arrayidx1 = getelementptr inbounds i64, i64* %a, i64 1
50 store i64 4294967297, i64* %a
51 store i64 12884901889, i64* %arrayidx1
55 ; Splats may be an opportunity to use a broadcast op.
57 define void @big_nonzero_32_bytes_splat(i32* nocapture %a) {
58 ; X32-LABEL: big_nonzero_32_bytes_splat:
60 ; X32-NEXT: movl {{[0-9]+}}(%esp), %eax
61 ; X32-NEXT: vmovaps {{.*#+}} ymm0 = [42,42,42,42,42,42,42,42]
62 ; X32-NEXT: vmovups %ymm0, (%eax)
63 ; X32-NEXT: vzeroupper
66 ; X64-LABEL: big_nonzero_32_bytes_splat:
68 ; X64-NEXT: vmovaps {{.*#+}} ymm0 = [42,42,42,42,42,42,42,42]
69 ; X64-NEXT: vmovups %ymm0, (%rdi)
70 ; X64-NEXT: vzeroupper
72 %arrayidx1 = getelementptr inbounds i32, i32* %a, i64 1
73 %arrayidx2 = getelementptr inbounds i32, i32* %a, i64 2
74 %arrayidx3 = getelementptr inbounds i32, i32* %a, i64 3
75 %arrayidx4 = getelementptr inbounds i32, i32* %a, i64 4
76 %arrayidx5 = getelementptr inbounds i32, i32* %a, i64 5
77 %arrayidx6 = getelementptr inbounds i32, i32* %a, i64 6
78 %arrayidx7 = getelementptr inbounds i32, i32* %a, i64 7
80 store i32 42, i32* %a, align 4
81 store i32 42, i32* %arrayidx1, align 4
82 store i32 42, i32* %arrayidx2, align 4
83 store i32 42, i32* %arrayidx3, align 4
84 store i32 42, i32* %arrayidx4, align 4
85 store i32 42, i32* %arrayidx5, align 4
86 store i32 42, i32* %arrayidx6, align 4
87 store i32 42, i32* %arrayidx7, align 4
91 ; Verify that we choose the best-sized store(s) for each chunk.
93 define void @big_nonzero_63_bytes(i8* nocapture %a) {
94 ; X32-LABEL: big_nonzero_63_bytes:
96 ; X32-NEXT: movl {{[0-9]+}}(%esp), %eax
97 ; X32-NEXT: vmovaps {{.*#+}} ymm0 = [1,0,2,0,3,0,4,0]
98 ; X32-NEXT: vmovups %ymm0, (%eax)
99 ; X32-NEXT: vmovaps {{.*#+}} xmm0 = [5,0,6,0]
100 ; X32-NEXT: vmovups %xmm0, 32(%eax)
101 ; X32-NEXT: movl $0, 52(%eax)
102 ; X32-NEXT: movl $7, 48(%eax)
103 ; X32-NEXT: movl $8, 56(%eax)
104 ; X32-NEXT: movw $9, 60(%eax)
105 ; X32-NEXT: movb $10, 62(%eax)
106 ; X32-NEXT: vzeroupper
109 ; X64-LABEL: big_nonzero_63_bytes:
111 ; X64-NEXT: vmovaps {{.*#+}} ymm0 = [1,2,3,4]
112 ; X64-NEXT: vmovups %ymm0, (%rdi)
113 ; X64-NEXT: movq $5, 32(%rdi)
114 ; X64-NEXT: movq $6, 40(%rdi)
115 ; X64-NEXT: movq $7, 48(%rdi)
116 ; X64-NEXT: movl $8, 56(%rdi)
117 ; X64-NEXT: movw $9, 60(%rdi)
118 ; X64-NEXT: movb $10, 62(%rdi)
119 ; X64-NEXT: vzeroupper
121 %a8 = bitcast i8* %a to i64*
122 %arrayidx8 = getelementptr inbounds i64, i64* %a8, i64 1
123 %arrayidx16 = getelementptr inbounds i64, i64* %a8, i64 2
124 %arrayidx24 = getelementptr inbounds i64, i64* %a8, i64 3
125 %arrayidx32 = getelementptr inbounds i64, i64* %a8, i64 4
126 %arrayidx40 = getelementptr inbounds i64, i64* %a8, i64 5
127 %arrayidx48 = getelementptr inbounds i64, i64* %a8, i64 6
128 %a4 = bitcast i8* %a to i32*
129 %arrayidx56 = getelementptr inbounds i32, i32* %a4, i64 14
130 %a2 = bitcast i8* %a to i16*
131 %arrayidx60 = getelementptr inbounds i16, i16* %a2, i64 30
132 %arrayidx62 = getelementptr inbounds i8, i8* %a, i64 62
134 store i64 1, i64* %a8
135 store i64 2, i64* %arrayidx8
136 store i64 3, i64* %arrayidx16
137 store i64 4, i64* %arrayidx24
138 store i64 5, i64* %arrayidx32
139 store i64 6, i64* %arrayidx40
140 store i64 7, i64* %arrayidx48
141 store i32 8, i32* %arrayidx56
142 store i16 9, i16* %arrayidx60
143 store i8 10, i8* %arrayidx62