xref: /aosp_15_r20/external/llvm/test/CodeGen/X86/memset-nonzero.ll (revision 9880d6810fe72a1726cb53787c6711e909410d58)
1*9880d681SAndroid Build Coastguard Worker; NOTE: Assertions have been autogenerated by update_test_checks.py
2*9880d681SAndroid Build Coastguard Worker; RUN: llc -mtriple=x86_64-unknown-unknown < %s -mattr=sse | FileCheck %s --check-prefix=SSE
3*9880d681SAndroid Build Coastguard Worker; RUN: llc -mtriple=x86_64-unknown-unknown < %s -mattr=sse2 | FileCheck %s --check-prefix=SSE
4*9880d681SAndroid Build Coastguard Worker; RUN: llc -mtriple=x86_64-unknown-unknown < %s -mattr=sse2,-slow-unaligned-mem-16 | FileCheck %s --check-prefix=SSE2FAST
5*9880d681SAndroid Build Coastguard Worker; RUN: llc -mtriple=x86_64-unknown-unknown < %s -mattr=avx | FileCheck %s --check-prefix=AVX --check-prefix=AVX1
6*9880d681SAndroid Build Coastguard Worker; RUN: llc -mtriple=x86_64-unknown-unknown < %s -mattr=avx2 | FileCheck %s --check-prefix=AVX --check-prefix=AVX2
7*9880d681SAndroid Build Coastguard Worker
8*9880d681SAndroid Build Coastguard Worker; https://llvm.org/bugs/show_bug.cgi?id=27100
9*9880d681SAndroid Build Coastguard Worker
10*9880d681SAndroid Build Coastguard Workerdefine void @memset_16_nonzero_bytes(i8* %x) {
11*9880d681SAndroid Build Coastguard Worker; SSE-LABEL: memset_16_nonzero_bytes:
12*9880d681SAndroid Build Coastguard Worker; SSE:       # BB#0:
13*9880d681SAndroid Build Coastguard Worker; SSE-NEXT:    movabsq $3038287259199220266, %rax # imm = 0x2A2A2A2A2A2A2A2A
14*9880d681SAndroid Build Coastguard Worker; SSE-NEXT:    movq %rax, 8(%rdi)
15*9880d681SAndroid Build Coastguard Worker; SSE-NEXT:    movq %rax, (%rdi)
16*9880d681SAndroid Build Coastguard Worker; SSE-NEXT:    retq
17*9880d681SAndroid Build Coastguard Worker;
18*9880d681SAndroid Build Coastguard Worker; SSE2FAST-LABEL: memset_16_nonzero_bytes:
19*9880d681SAndroid Build Coastguard Worker; SSE2FAST:       # BB#0:
20*9880d681SAndroid Build Coastguard Worker; SSE2FAST-NEXT:    movaps {{.*#+}} xmm0 = [42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42]
21*9880d681SAndroid Build Coastguard Worker; SSE2FAST-NEXT:    movups %xmm0, (%rdi)
22*9880d681SAndroid Build Coastguard Worker; SSE2FAST-NEXT:    retq
23*9880d681SAndroid Build Coastguard Worker;
24*9880d681SAndroid Build Coastguard Worker; AVX-LABEL: memset_16_nonzero_bytes:
25*9880d681SAndroid Build Coastguard Worker; AVX:       # BB#0:
26*9880d681SAndroid Build Coastguard Worker; AVX-NEXT:    vmovaps {{.*#+}} xmm0 = [42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42]
27*9880d681SAndroid Build Coastguard Worker; AVX-NEXT:    vmovups %xmm0, (%rdi)
28*9880d681SAndroid Build Coastguard Worker; AVX-NEXT:    retq
29*9880d681SAndroid Build Coastguard Worker;
30*9880d681SAndroid Build Coastguard Worker  %call = tail call i8* @__memset_chk(i8* %x, i32 42, i64 16, i64 -1)
31*9880d681SAndroid Build Coastguard Worker  ret void
32*9880d681SAndroid Build Coastguard Worker}
33*9880d681SAndroid Build Coastguard Worker
34*9880d681SAndroid Build Coastguard Workerdefine void @memset_32_nonzero_bytes(i8* %x) {
35*9880d681SAndroid Build Coastguard Worker; SSE-LABEL: memset_32_nonzero_bytes:
36*9880d681SAndroid Build Coastguard Worker; SSE:       # BB#0:
37*9880d681SAndroid Build Coastguard Worker; SSE-NEXT:    movabsq $3038287259199220266, %rax # imm = 0x2A2A2A2A2A2A2A2A
38*9880d681SAndroid Build Coastguard Worker; SSE-NEXT:    movq %rax, 24(%rdi)
39*9880d681SAndroid Build Coastguard Worker; SSE-NEXT:    movq %rax, 16(%rdi)
40*9880d681SAndroid Build Coastguard Worker; SSE-NEXT:    movq %rax, 8(%rdi)
41*9880d681SAndroid Build Coastguard Worker; SSE-NEXT:    movq %rax, (%rdi)
42*9880d681SAndroid Build Coastguard Worker; SSE-NEXT:    retq
43*9880d681SAndroid Build Coastguard Worker;
44*9880d681SAndroid Build Coastguard Worker; SSE2FAST-LABEL: memset_32_nonzero_bytes:
45*9880d681SAndroid Build Coastguard Worker; SSE2FAST:       # BB#0:
46*9880d681SAndroid Build Coastguard Worker; SSE2FAST-NEXT:    movaps {{.*#+}} xmm0 = [42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42]
47*9880d681SAndroid Build Coastguard Worker; SSE2FAST-NEXT:    movups %xmm0, 16(%rdi)
48*9880d681SAndroid Build Coastguard Worker; SSE2FAST-NEXT:    movups %xmm0, (%rdi)
49*9880d681SAndroid Build Coastguard Worker; SSE2FAST-NEXT:    retq
50*9880d681SAndroid Build Coastguard Worker;
51*9880d681SAndroid Build Coastguard Worker; AVX-LABEL: memset_32_nonzero_bytes:
52*9880d681SAndroid Build Coastguard Worker; AVX:       # BB#0:
53*9880d681SAndroid Build Coastguard Worker; AVX-NEXT:    vmovaps {{.*#+}} ymm0 = [42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42]
54*9880d681SAndroid Build Coastguard Worker; AVX-NEXT:    vmovups %ymm0, (%rdi)
55*9880d681SAndroid Build Coastguard Worker; AVX-NEXT:    vzeroupper
56*9880d681SAndroid Build Coastguard Worker; AVX-NEXT:    retq
57*9880d681SAndroid Build Coastguard Worker;
58*9880d681SAndroid Build Coastguard Worker  %call = tail call i8* @__memset_chk(i8* %x, i32 42, i64 32, i64 -1)
59*9880d681SAndroid Build Coastguard Worker  ret void
60*9880d681SAndroid Build Coastguard Worker}
61*9880d681SAndroid Build Coastguard Worker
62*9880d681SAndroid Build Coastguard Workerdefine void @memset_64_nonzero_bytes(i8* %x) {
63*9880d681SAndroid Build Coastguard Worker; SSE-LABEL: memset_64_nonzero_bytes:
64*9880d681SAndroid Build Coastguard Worker; SSE:       # BB#0:
65*9880d681SAndroid Build Coastguard Worker; SSE-NEXT:    movabsq $3038287259199220266, %rax # imm = 0x2A2A2A2A2A2A2A2A
66*9880d681SAndroid Build Coastguard Worker; SSE-NEXT:    movq %rax, 56(%rdi)
67*9880d681SAndroid Build Coastguard Worker; SSE-NEXT:    movq %rax, 48(%rdi)
68*9880d681SAndroid Build Coastguard Worker; SSE-NEXT:    movq %rax, 40(%rdi)
69*9880d681SAndroid Build Coastguard Worker; SSE-NEXT:    movq %rax, 32(%rdi)
70*9880d681SAndroid Build Coastguard Worker; SSE-NEXT:    movq %rax, 24(%rdi)
71*9880d681SAndroid Build Coastguard Worker; SSE-NEXT:    movq %rax, 16(%rdi)
72*9880d681SAndroid Build Coastguard Worker; SSE-NEXT:    movq %rax, 8(%rdi)
73*9880d681SAndroid Build Coastguard Worker; SSE-NEXT:    movq %rax, (%rdi)
74*9880d681SAndroid Build Coastguard Worker; SSE-NEXT:    retq
75*9880d681SAndroid Build Coastguard Worker;
76*9880d681SAndroid Build Coastguard Worker; SSE2FAST-LABEL: memset_64_nonzero_bytes:
77*9880d681SAndroid Build Coastguard Worker; SSE2FAST:       # BB#0:
78*9880d681SAndroid Build Coastguard Worker; SSE2FAST-NEXT:    movaps {{.*#+}} xmm0 = [42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42]
79*9880d681SAndroid Build Coastguard Worker; SSE2FAST-NEXT:    movups %xmm0, 48(%rdi)
80*9880d681SAndroid Build Coastguard Worker; SSE2FAST-NEXT:    movups %xmm0, 32(%rdi)
81*9880d681SAndroid Build Coastguard Worker; SSE2FAST-NEXT:    movups %xmm0, 16(%rdi)
82*9880d681SAndroid Build Coastguard Worker; SSE2FAST-NEXT:    movups %xmm0, (%rdi)
83*9880d681SAndroid Build Coastguard Worker; SSE2FAST-NEXT:    retq
84*9880d681SAndroid Build Coastguard Worker;
85*9880d681SAndroid Build Coastguard Worker; AVX-LABEL: memset_64_nonzero_bytes:
86*9880d681SAndroid Build Coastguard Worker; AVX:       # BB#0:
87*9880d681SAndroid Build Coastguard Worker; AVX-NEXT:    vmovaps {{.*#+}} ymm0 = [42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42]
88*9880d681SAndroid Build Coastguard Worker; AVX-NEXT:    vmovups %ymm0, 32(%rdi)
89*9880d681SAndroid Build Coastguard Worker; AVX-NEXT:    vmovups %ymm0, (%rdi)
90*9880d681SAndroid Build Coastguard Worker; AVX-NEXT:    vzeroupper
91*9880d681SAndroid Build Coastguard Worker; AVX-NEXT:    retq
92*9880d681SAndroid Build Coastguard Worker;
93*9880d681SAndroid Build Coastguard Worker  %call = tail call i8* @__memset_chk(i8* %x, i32 42, i64 64, i64 -1)
94*9880d681SAndroid Build Coastguard Worker  ret void
95*9880d681SAndroid Build Coastguard Worker}
96*9880d681SAndroid Build Coastguard Worker
97*9880d681SAndroid Build Coastguard Workerdefine void @memset_128_nonzero_bytes(i8* %x) {
98*9880d681SAndroid Build Coastguard Worker; SSE-LABEL: memset_128_nonzero_bytes:
99*9880d681SAndroid Build Coastguard Worker; SSE:       # BB#0:
100*9880d681SAndroid Build Coastguard Worker; SSE-NEXT:    movabsq $3038287259199220266, %rax # imm = 0x2A2A2A2A2A2A2A2A
101*9880d681SAndroid Build Coastguard Worker; SSE-NEXT:    movq %rax, 120(%rdi)
102*9880d681SAndroid Build Coastguard Worker; SSE-NEXT:    movq %rax, 112(%rdi)
103*9880d681SAndroid Build Coastguard Worker; SSE-NEXT:    movq %rax, 104(%rdi)
104*9880d681SAndroid Build Coastguard Worker; SSE-NEXT:    movq %rax, 96(%rdi)
105*9880d681SAndroid Build Coastguard Worker; SSE-NEXT:    movq %rax, 88(%rdi)
106*9880d681SAndroid Build Coastguard Worker; SSE-NEXT:    movq %rax, 80(%rdi)
107*9880d681SAndroid Build Coastguard Worker; SSE-NEXT:    movq %rax, 72(%rdi)
108*9880d681SAndroid Build Coastguard Worker; SSE-NEXT:    movq %rax, 64(%rdi)
109*9880d681SAndroid Build Coastguard Worker; SSE-NEXT:    movq %rax, 56(%rdi)
110*9880d681SAndroid Build Coastguard Worker; SSE-NEXT:    movq %rax, 48(%rdi)
111*9880d681SAndroid Build Coastguard Worker; SSE-NEXT:    movq %rax, 40(%rdi)
112*9880d681SAndroid Build Coastguard Worker; SSE-NEXT:    movq %rax, 32(%rdi)
113*9880d681SAndroid Build Coastguard Worker; SSE-NEXT:    movq %rax, 24(%rdi)
114*9880d681SAndroid Build Coastguard Worker; SSE-NEXT:    movq %rax, 16(%rdi)
115*9880d681SAndroid Build Coastguard Worker; SSE-NEXT:    movq %rax, 8(%rdi)
116*9880d681SAndroid Build Coastguard Worker; SSE-NEXT:    movq %rax, (%rdi)
117*9880d681SAndroid Build Coastguard Worker; SSE-NEXT:    retq
118*9880d681SAndroid Build Coastguard Worker;
119*9880d681SAndroid Build Coastguard Worker; SSE2FAST-LABEL: memset_128_nonzero_bytes:
120*9880d681SAndroid Build Coastguard Worker; SSE2FAST:       # BB#0:
121*9880d681SAndroid Build Coastguard Worker; SSE2FAST-NEXT:    movaps {{.*#+}} xmm0 = [42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42]
122*9880d681SAndroid Build Coastguard Worker; SSE2FAST-NEXT:    movups %xmm0, 112(%rdi)
123*9880d681SAndroid Build Coastguard Worker; SSE2FAST-NEXT:    movups %xmm0, 96(%rdi)
124*9880d681SAndroid Build Coastguard Worker; SSE2FAST-NEXT:    movups %xmm0, 80(%rdi)
125*9880d681SAndroid Build Coastguard Worker; SSE2FAST-NEXT:    movups %xmm0, 64(%rdi)
126*9880d681SAndroid Build Coastguard Worker; SSE2FAST-NEXT:    movups %xmm0, 48(%rdi)
127*9880d681SAndroid Build Coastguard Worker; SSE2FAST-NEXT:    movups %xmm0, 32(%rdi)
128*9880d681SAndroid Build Coastguard Worker; SSE2FAST-NEXT:    movups %xmm0, 16(%rdi)
129*9880d681SAndroid Build Coastguard Worker; SSE2FAST-NEXT:    movups %xmm0, (%rdi)
130*9880d681SAndroid Build Coastguard Worker; SSE2FAST-NEXT:    retq
131*9880d681SAndroid Build Coastguard Worker;
132*9880d681SAndroid Build Coastguard Worker; AVX-LABEL: memset_128_nonzero_bytes:
133*9880d681SAndroid Build Coastguard Worker; AVX:       # BB#0:
134*9880d681SAndroid Build Coastguard Worker; AVX-NEXT:    vmovaps {{.*#+}} ymm0 = [42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42]
135*9880d681SAndroid Build Coastguard Worker; AVX-NEXT:    vmovups %ymm0, 96(%rdi)
136*9880d681SAndroid Build Coastguard Worker; AVX-NEXT:    vmovups %ymm0, 64(%rdi)
137*9880d681SAndroid Build Coastguard Worker; AVX-NEXT:    vmovups %ymm0, 32(%rdi)
138*9880d681SAndroid Build Coastguard Worker; AVX-NEXT:    vmovups %ymm0, (%rdi)
139*9880d681SAndroid Build Coastguard Worker; AVX-NEXT:    vzeroupper
140*9880d681SAndroid Build Coastguard Worker; AVX-NEXT:    retq
141*9880d681SAndroid Build Coastguard Worker;
142*9880d681SAndroid Build Coastguard Worker  %call = tail call i8* @__memset_chk(i8* %x, i32 42, i64 128, i64 -1)
143*9880d681SAndroid Build Coastguard Worker  ret void
144*9880d681SAndroid Build Coastguard Worker}
145*9880d681SAndroid Build Coastguard Worker
146*9880d681SAndroid Build Coastguard Workerdefine void @memset_256_nonzero_bytes(i8* %x) {
147*9880d681SAndroid Build Coastguard Worker; SSE-LABEL: memset_256_nonzero_bytes:
148*9880d681SAndroid Build Coastguard Worker; SSE:       # BB#0:
149*9880d681SAndroid Build Coastguard Worker; SSE-NEXT:    pushq %rax
150*9880d681SAndroid Build Coastguard Worker; SSE-NEXT:  .Ltmp0:
151*9880d681SAndroid Build Coastguard Worker; SSE-NEXT:    .cfi_def_cfa_offset 16
152*9880d681SAndroid Build Coastguard Worker; SSE-NEXT:    movl $42, %esi
153*9880d681SAndroid Build Coastguard Worker; SSE-NEXT:    movl $256, %edx # imm = 0x100
154*9880d681SAndroid Build Coastguard Worker; SSE-NEXT:    callq memset
155*9880d681SAndroid Build Coastguard Worker; SSE-NEXT:    popq %rax
156*9880d681SAndroid Build Coastguard Worker; SSE-NEXT:    retq
157*9880d681SAndroid Build Coastguard Worker;
158*9880d681SAndroid Build Coastguard Worker; SSE2FAST-LABEL: memset_256_nonzero_bytes:
159*9880d681SAndroid Build Coastguard Worker; SSE2FAST:       # BB#0:
160*9880d681SAndroid Build Coastguard Worker; SSE2FAST-NEXT:    movaps {{.*#+}} xmm0 = [42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42]
161*9880d681SAndroid Build Coastguard Worker; SSE2FAST-NEXT:    movups %xmm0, 240(%rdi)
162*9880d681SAndroid Build Coastguard Worker; SSE2FAST-NEXT:    movups %xmm0, 224(%rdi)
163*9880d681SAndroid Build Coastguard Worker; SSE2FAST-NEXT:    movups %xmm0, 208(%rdi)
164*9880d681SAndroid Build Coastguard Worker; SSE2FAST-NEXT:    movups %xmm0, 192(%rdi)
165*9880d681SAndroid Build Coastguard Worker; SSE2FAST-NEXT:    movups %xmm0, 176(%rdi)
166*9880d681SAndroid Build Coastguard Worker; SSE2FAST-NEXT:    movups %xmm0, 160(%rdi)
167*9880d681SAndroid Build Coastguard Worker; SSE2FAST-NEXT:    movups %xmm0, 144(%rdi)
168*9880d681SAndroid Build Coastguard Worker; SSE2FAST-NEXT:    movups %xmm0, 128(%rdi)
169*9880d681SAndroid Build Coastguard Worker; SSE2FAST-NEXT:    movups %xmm0, 112(%rdi)
170*9880d681SAndroid Build Coastguard Worker; SSE2FAST-NEXT:    movups %xmm0, 96(%rdi)
171*9880d681SAndroid Build Coastguard Worker; SSE2FAST-NEXT:    movups %xmm0, 80(%rdi)
172*9880d681SAndroid Build Coastguard Worker; SSE2FAST-NEXT:    movups %xmm0, 64(%rdi)
173*9880d681SAndroid Build Coastguard Worker; SSE2FAST-NEXT:    movups %xmm0, 48(%rdi)
174*9880d681SAndroid Build Coastguard Worker; SSE2FAST-NEXT:    movups %xmm0, 32(%rdi)
175*9880d681SAndroid Build Coastguard Worker; SSE2FAST-NEXT:    movups %xmm0, 16(%rdi)
176*9880d681SAndroid Build Coastguard Worker; SSE2FAST-NEXT:    movups %xmm0, (%rdi)
177*9880d681SAndroid Build Coastguard Worker; SSE2FAST-NEXT:    retq
178*9880d681SAndroid Build Coastguard Worker;
179*9880d681SAndroid Build Coastguard Worker; AVX-LABEL: memset_256_nonzero_bytes:
180*9880d681SAndroid Build Coastguard Worker; AVX:       # BB#0:
181*9880d681SAndroid Build Coastguard Worker; AVX-NEXT:    vmovaps {{.*#+}} ymm0 = [42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42,42]
182*9880d681SAndroid Build Coastguard Worker; AVX-NEXT:    vmovups %ymm0, 224(%rdi)
183*9880d681SAndroid Build Coastguard Worker; AVX-NEXT:    vmovups %ymm0, 192(%rdi)
184*9880d681SAndroid Build Coastguard Worker; AVX-NEXT:    vmovups %ymm0, 160(%rdi)
185*9880d681SAndroid Build Coastguard Worker; AVX-NEXT:    vmovups %ymm0, 128(%rdi)
186*9880d681SAndroid Build Coastguard Worker; AVX-NEXT:    vmovups %ymm0, 96(%rdi)
187*9880d681SAndroid Build Coastguard Worker; AVX-NEXT:    vmovups %ymm0, 64(%rdi)
188*9880d681SAndroid Build Coastguard Worker; AVX-NEXT:    vmovups %ymm0, 32(%rdi)
189*9880d681SAndroid Build Coastguard Worker; AVX-NEXT:    vmovups %ymm0, (%rdi)
190*9880d681SAndroid Build Coastguard Worker; AVX-NEXT:    vzeroupper
191*9880d681SAndroid Build Coastguard Worker; AVX-NEXT:    retq
192*9880d681SAndroid Build Coastguard Worker;
193*9880d681SAndroid Build Coastguard Worker  %call = tail call i8* @__memset_chk(i8* %x, i32 42, i64 256, i64 -1)
194*9880d681SAndroid Build Coastguard Worker  ret void
195*9880d681SAndroid Build Coastguard Worker}
196*9880d681SAndroid Build Coastguard Worker
197*9880d681SAndroid Build Coastguard Workerdeclare i8* @__memset_chk(i8*, i32, i64, i64)
198*9880d681SAndroid Build Coastguard Worker
199*9880d681SAndroid Build Coastguard Worker; Repeat with a non-constant value for the stores.
200*9880d681SAndroid Build Coastguard Worker
201*9880d681SAndroid Build Coastguard Workerdefine void @memset_16_nonconst_bytes(i8* %x, i8 %c) {
202*9880d681SAndroid Build Coastguard Worker; SSE-LABEL: memset_16_nonconst_bytes:
203*9880d681SAndroid Build Coastguard Worker; SSE:       # BB#0:
204*9880d681SAndroid Build Coastguard Worker; SSE-NEXT:    movzbl %sil, %eax
205*9880d681SAndroid Build Coastguard Worker; SSE-NEXT:    movabsq $72340172838076673, %rcx # imm = 0x101010101010101
206*9880d681SAndroid Build Coastguard Worker; SSE-NEXT:    imulq %rax, %rcx
207*9880d681SAndroid Build Coastguard Worker; SSE-NEXT:    movq %rcx, 8(%rdi)
208*9880d681SAndroid Build Coastguard Worker; SSE-NEXT:    movq %rcx, (%rdi)
209*9880d681SAndroid Build Coastguard Worker; SSE-NEXT:    retq
210*9880d681SAndroid Build Coastguard Worker;
211*9880d681SAndroid Build Coastguard Worker; SSE2FAST-LABEL: memset_16_nonconst_bytes:
212*9880d681SAndroid Build Coastguard Worker; SSE2FAST:       # BB#0:
213*9880d681SAndroid Build Coastguard Worker; SSE2FAST-NEXT:    movd %esi, %xmm0
214*9880d681SAndroid Build Coastguard Worker; SSE2FAST-NEXT:    punpcklbw {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
215*9880d681SAndroid Build Coastguard Worker; SSE2FAST-NEXT:    pshuflw {{.*#+}} xmm0 = xmm0[0,0,0,0,4,5,6,7]
216*9880d681SAndroid Build Coastguard Worker; SSE2FAST-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,1,1]
217*9880d681SAndroid Build Coastguard Worker; SSE2FAST-NEXT:    movdqu %xmm0, (%rdi)
218*9880d681SAndroid Build Coastguard Worker; SSE2FAST-NEXT:    retq
219*9880d681SAndroid Build Coastguard Worker;
220*9880d681SAndroid Build Coastguard Worker; AVX1-LABEL: memset_16_nonconst_bytes:
221*9880d681SAndroid Build Coastguard Worker; AVX1:       # BB#0:
222*9880d681SAndroid Build Coastguard Worker; AVX1-NEXT:    vmovd %esi, %xmm0
223*9880d681SAndroid Build Coastguard Worker; AVX1-NEXT:    vpxor %xmm1, %xmm1, %xmm1
224*9880d681SAndroid Build Coastguard Worker; AVX1-NEXT:    vpshufb %xmm1, %xmm0, %xmm0
225*9880d681SAndroid Build Coastguard Worker; AVX1-NEXT:    vmovdqu %xmm0, (%rdi)
226*9880d681SAndroid Build Coastguard Worker; AVX1-NEXT:    retq
227*9880d681SAndroid Build Coastguard Worker;
228*9880d681SAndroid Build Coastguard Worker; AVX2-LABEL: memset_16_nonconst_bytes:
229*9880d681SAndroid Build Coastguard Worker; AVX2:       # BB#0:
230*9880d681SAndroid Build Coastguard Worker; AVX2-NEXT:    vmovd %esi, %xmm0
231*9880d681SAndroid Build Coastguard Worker; AVX2-NEXT:    vpbroadcastb %xmm0, %xmm0
232*9880d681SAndroid Build Coastguard Worker; AVX2-NEXT:    vmovdqu %xmm0, (%rdi)
233*9880d681SAndroid Build Coastguard Worker; AVX2-NEXT:    retq
234*9880d681SAndroid Build Coastguard Worker;
235*9880d681SAndroid Build Coastguard Worker  tail call void @llvm.memset.p0i8.i64(i8* %x, i8 %c, i64 16, i32 1, i1 false)
236*9880d681SAndroid Build Coastguard Worker  ret void
237*9880d681SAndroid Build Coastguard Worker}
238*9880d681SAndroid Build Coastguard Worker
239*9880d681SAndroid Build Coastguard Workerdefine void @memset_32_nonconst_bytes(i8* %x, i8 %c) {
240*9880d681SAndroid Build Coastguard Worker; SSE-LABEL: memset_32_nonconst_bytes:
241*9880d681SAndroid Build Coastguard Worker; SSE:       # BB#0:
242*9880d681SAndroid Build Coastguard Worker; SSE-NEXT:    movzbl %sil, %eax
243*9880d681SAndroid Build Coastguard Worker; SSE-NEXT:    movabsq $72340172838076673, %rcx # imm = 0x101010101010101
244*9880d681SAndroid Build Coastguard Worker; SSE-NEXT:    imulq %rax, %rcx
245*9880d681SAndroid Build Coastguard Worker; SSE-NEXT:    movq %rcx, 24(%rdi)
246*9880d681SAndroid Build Coastguard Worker; SSE-NEXT:    movq %rcx, 16(%rdi)
247*9880d681SAndroid Build Coastguard Worker; SSE-NEXT:    movq %rcx, 8(%rdi)
248*9880d681SAndroid Build Coastguard Worker; SSE-NEXT:    movq %rcx, (%rdi)
249*9880d681SAndroid Build Coastguard Worker; SSE-NEXT:    retq
250*9880d681SAndroid Build Coastguard Worker;
251*9880d681SAndroid Build Coastguard Worker; SSE2FAST-LABEL: memset_32_nonconst_bytes:
252*9880d681SAndroid Build Coastguard Worker; SSE2FAST:       # BB#0:
253*9880d681SAndroid Build Coastguard Worker; SSE2FAST-NEXT:    movd %esi, %xmm0
254*9880d681SAndroid Build Coastguard Worker; SSE2FAST-NEXT:    punpcklbw {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
255*9880d681SAndroid Build Coastguard Worker; SSE2FAST-NEXT:    pshuflw {{.*#+}} xmm0 = xmm0[0,0,0,0,4,5,6,7]
256*9880d681SAndroid Build Coastguard Worker; SSE2FAST-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,1,1]
257*9880d681SAndroid Build Coastguard Worker; SSE2FAST-NEXT:    movdqu %xmm0, 16(%rdi)
258*9880d681SAndroid Build Coastguard Worker; SSE2FAST-NEXT:    movdqu %xmm0, (%rdi)
259*9880d681SAndroid Build Coastguard Worker; SSE2FAST-NEXT:    retq
260*9880d681SAndroid Build Coastguard Worker;
261*9880d681SAndroid Build Coastguard Worker; AVX1-LABEL: memset_32_nonconst_bytes:
262*9880d681SAndroid Build Coastguard Worker; AVX1:       # BB#0:
263*9880d681SAndroid Build Coastguard Worker; AVX1-NEXT:    vmovd %esi, %xmm0
264*9880d681SAndroid Build Coastguard Worker; AVX1-NEXT:    vpxor %xmm1, %xmm1, %xmm1
265*9880d681SAndroid Build Coastguard Worker; AVX1-NEXT:    vpshufb %xmm1, %xmm0, %xmm0
266*9880d681SAndroid Build Coastguard Worker; AVX1-NEXT:    vinsertf128 $1, %xmm0, %ymm0, %ymm0
267*9880d681SAndroid Build Coastguard Worker; AVX1-NEXT:    vmovups %ymm0, (%rdi)
268*9880d681SAndroid Build Coastguard Worker; AVX1-NEXT:    vzeroupper
269*9880d681SAndroid Build Coastguard Worker; AVX1-NEXT:    retq
270*9880d681SAndroid Build Coastguard Worker;
271*9880d681SAndroid Build Coastguard Worker; AVX2-LABEL: memset_32_nonconst_bytes:
272*9880d681SAndroid Build Coastguard Worker; AVX2:       # BB#0:
273*9880d681SAndroid Build Coastguard Worker; AVX2-NEXT:    vmovd %esi, %xmm0
274*9880d681SAndroid Build Coastguard Worker; AVX2-NEXT:    vpbroadcastb %xmm0, %ymm0
275*9880d681SAndroid Build Coastguard Worker; AVX2-NEXT:    vmovdqu %ymm0, (%rdi)
276*9880d681SAndroid Build Coastguard Worker; AVX2-NEXT:    vzeroupper
277*9880d681SAndroid Build Coastguard Worker; AVX2-NEXT:    retq
278*9880d681SAndroid Build Coastguard Worker;
279*9880d681SAndroid Build Coastguard Worker  tail call void @llvm.memset.p0i8.i64(i8* %x, i8 %c, i64 32, i32 1, i1 false)
280*9880d681SAndroid Build Coastguard Worker  ret void
281*9880d681SAndroid Build Coastguard Worker}
282*9880d681SAndroid Build Coastguard Worker
283*9880d681SAndroid Build Coastguard Workerdefine void @memset_64_nonconst_bytes(i8* %x, i8 %c) {
284*9880d681SAndroid Build Coastguard Worker; SSE-LABEL: memset_64_nonconst_bytes:
285*9880d681SAndroid Build Coastguard Worker; SSE:       # BB#0:
286*9880d681SAndroid Build Coastguard Worker; SSE-NEXT:    movzbl %sil, %eax
287*9880d681SAndroid Build Coastguard Worker; SSE-NEXT:    movabsq $72340172838076673, %rcx # imm = 0x101010101010101
288*9880d681SAndroid Build Coastguard Worker; SSE-NEXT:    imulq %rax, %rcx
289*9880d681SAndroid Build Coastguard Worker; SSE-NEXT:    movq %rcx, 56(%rdi)
290*9880d681SAndroid Build Coastguard Worker; SSE-NEXT:    movq %rcx, 48(%rdi)
291*9880d681SAndroid Build Coastguard Worker; SSE-NEXT:    movq %rcx, 40(%rdi)
292*9880d681SAndroid Build Coastguard Worker; SSE-NEXT:    movq %rcx, 32(%rdi)
293*9880d681SAndroid Build Coastguard Worker; SSE-NEXT:    movq %rcx, 24(%rdi)
294*9880d681SAndroid Build Coastguard Worker; SSE-NEXT:    movq %rcx, 16(%rdi)
295*9880d681SAndroid Build Coastguard Worker; SSE-NEXT:    movq %rcx, 8(%rdi)
296*9880d681SAndroid Build Coastguard Worker; SSE-NEXT:    movq %rcx, (%rdi)
297*9880d681SAndroid Build Coastguard Worker; SSE-NEXT:    retq
298*9880d681SAndroid Build Coastguard Worker;
299*9880d681SAndroid Build Coastguard Worker; SSE2FAST-LABEL: memset_64_nonconst_bytes:
300*9880d681SAndroid Build Coastguard Worker; SSE2FAST:       # BB#0:
301*9880d681SAndroid Build Coastguard Worker; SSE2FAST-NEXT:    movd %esi, %xmm0
302*9880d681SAndroid Build Coastguard Worker; SSE2FAST-NEXT:    punpcklbw {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
303*9880d681SAndroid Build Coastguard Worker; SSE2FAST-NEXT:    pshuflw {{.*#+}} xmm0 = xmm0[0,0,0,0,4,5,6,7]
304*9880d681SAndroid Build Coastguard Worker; SSE2FAST-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,1,1]
305*9880d681SAndroid Build Coastguard Worker; SSE2FAST-NEXT:    movdqu %xmm0, 48(%rdi)
306*9880d681SAndroid Build Coastguard Worker; SSE2FAST-NEXT:    movdqu %xmm0, 32(%rdi)
307*9880d681SAndroid Build Coastguard Worker; SSE2FAST-NEXT:    movdqu %xmm0, 16(%rdi)
308*9880d681SAndroid Build Coastguard Worker; SSE2FAST-NEXT:    movdqu %xmm0, (%rdi)
309*9880d681SAndroid Build Coastguard Worker; SSE2FAST-NEXT:    retq
310*9880d681SAndroid Build Coastguard Worker;
311*9880d681SAndroid Build Coastguard Worker; AVX1-LABEL: memset_64_nonconst_bytes:
312*9880d681SAndroid Build Coastguard Worker; AVX1:       # BB#0:
313*9880d681SAndroid Build Coastguard Worker; AVX1-NEXT:    vmovd %esi, %xmm0
314*9880d681SAndroid Build Coastguard Worker; AVX1-NEXT:    vpxor %xmm1, %xmm1, %xmm1
315*9880d681SAndroid Build Coastguard Worker; AVX1-NEXT:    vpshufb %xmm1, %xmm0, %xmm0
316*9880d681SAndroid Build Coastguard Worker; AVX1-NEXT:    vinsertf128 $1, %xmm0, %ymm0, %ymm0
317*9880d681SAndroid Build Coastguard Worker; AVX1-NEXT:    vmovups %ymm0, 32(%rdi)
318*9880d681SAndroid Build Coastguard Worker; AVX1-NEXT:    vmovups %ymm0, (%rdi)
319*9880d681SAndroid Build Coastguard Worker; AVX1-NEXT:    vzeroupper
320*9880d681SAndroid Build Coastguard Worker; AVX1-NEXT:    retq
321*9880d681SAndroid Build Coastguard Worker;
322*9880d681SAndroid Build Coastguard Worker; AVX2-LABEL: memset_64_nonconst_bytes:
323*9880d681SAndroid Build Coastguard Worker; AVX2:       # BB#0:
324*9880d681SAndroid Build Coastguard Worker; AVX2-NEXT:    vmovd %esi, %xmm0
325*9880d681SAndroid Build Coastguard Worker; AVX2-NEXT:    vpbroadcastb %xmm0, %ymm0
326*9880d681SAndroid Build Coastguard Worker; AVX2-NEXT:    vmovdqu %ymm0, 32(%rdi)
327*9880d681SAndroid Build Coastguard Worker; AVX2-NEXT:    vmovdqu %ymm0, (%rdi)
328*9880d681SAndroid Build Coastguard Worker; AVX2-NEXT:    vzeroupper
329*9880d681SAndroid Build Coastguard Worker; AVX2-NEXT:    retq
330*9880d681SAndroid Build Coastguard Worker;
331*9880d681SAndroid Build Coastguard Worker  tail call void @llvm.memset.p0i8.i64(i8* %x, i8 %c, i64 64, i32 1, i1 false)
332*9880d681SAndroid Build Coastguard Worker  ret void
333*9880d681SAndroid Build Coastguard Worker}
334*9880d681SAndroid Build Coastguard Worker
335*9880d681SAndroid Build Coastguard Workerdefine void @memset_128_nonconst_bytes(i8* %x, i8 %c) {
336*9880d681SAndroid Build Coastguard Worker; SSE-LABEL: memset_128_nonconst_bytes:
337*9880d681SAndroid Build Coastguard Worker; SSE:       # BB#0:
338*9880d681SAndroid Build Coastguard Worker; SSE-NEXT:    movzbl %sil, %eax
339*9880d681SAndroid Build Coastguard Worker; SSE-NEXT:    movabsq $72340172838076673, %rcx # imm = 0x101010101010101
340*9880d681SAndroid Build Coastguard Worker; SSE-NEXT:    imulq %rax, %rcx
341*9880d681SAndroid Build Coastguard Worker; SSE-NEXT:    movq %rcx, 120(%rdi)
342*9880d681SAndroid Build Coastguard Worker; SSE-NEXT:    movq %rcx, 112(%rdi)
343*9880d681SAndroid Build Coastguard Worker; SSE-NEXT:    movq %rcx, 104(%rdi)
344*9880d681SAndroid Build Coastguard Worker; SSE-NEXT:    movq %rcx, 96(%rdi)
345*9880d681SAndroid Build Coastguard Worker; SSE-NEXT:    movq %rcx, 88(%rdi)
346*9880d681SAndroid Build Coastguard Worker; SSE-NEXT:    movq %rcx, 80(%rdi)
347*9880d681SAndroid Build Coastguard Worker; SSE-NEXT:    movq %rcx, 72(%rdi)
348*9880d681SAndroid Build Coastguard Worker; SSE-NEXT:    movq %rcx, 64(%rdi)
349*9880d681SAndroid Build Coastguard Worker; SSE-NEXT:    movq %rcx, 56(%rdi)
350*9880d681SAndroid Build Coastguard Worker; SSE-NEXT:    movq %rcx, 48(%rdi)
351*9880d681SAndroid Build Coastguard Worker; SSE-NEXT:    movq %rcx, 40(%rdi)
352*9880d681SAndroid Build Coastguard Worker; SSE-NEXT:    movq %rcx, 32(%rdi)
353*9880d681SAndroid Build Coastguard Worker; SSE-NEXT:    movq %rcx, 24(%rdi)
354*9880d681SAndroid Build Coastguard Worker; SSE-NEXT:    movq %rcx, 16(%rdi)
355*9880d681SAndroid Build Coastguard Worker; SSE-NEXT:    movq %rcx, 8(%rdi)
356*9880d681SAndroid Build Coastguard Worker; SSE-NEXT:    movq %rcx, (%rdi)
357*9880d681SAndroid Build Coastguard Worker; SSE-NEXT:    retq
358*9880d681SAndroid Build Coastguard Worker;
359*9880d681SAndroid Build Coastguard Worker; SSE2FAST-LABEL: memset_128_nonconst_bytes:
360*9880d681SAndroid Build Coastguard Worker; SSE2FAST:       # BB#0:
361*9880d681SAndroid Build Coastguard Worker; SSE2FAST-NEXT:    movd %esi, %xmm0
362*9880d681SAndroid Build Coastguard Worker; SSE2FAST-NEXT:    punpcklbw {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
363*9880d681SAndroid Build Coastguard Worker; SSE2FAST-NEXT:    pshuflw {{.*#+}} xmm0 = xmm0[0,0,0,0,4,5,6,7]
364*9880d681SAndroid Build Coastguard Worker; SSE2FAST-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,1,1]
365*9880d681SAndroid Build Coastguard Worker; SSE2FAST-NEXT:    movdqu %xmm0, 112(%rdi)
366*9880d681SAndroid Build Coastguard Worker; SSE2FAST-NEXT:    movdqu %xmm0, 96(%rdi)
367*9880d681SAndroid Build Coastguard Worker; SSE2FAST-NEXT:    movdqu %xmm0, 80(%rdi)
368*9880d681SAndroid Build Coastguard Worker; SSE2FAST-NEXT:    movdqu %xmm0, 64(%rdi)
369*9880d681SAndroid Build Coastguard Worker; SSE2FAST-NEXT:    movdqu %xmm0, 48(%rdi)
370*9880d681SAndroid Build Coastguard Worker; SSE2FAST-NEXT:    movdqu %xmm0, 32(%rdi)
371*9880d681SAndroid Build Coastguard Worker; SSE2FAST-NEXT:    movdqu %xmm0, 16(%rdi)
372*9880d681SAndroid Build Coastguard Worker; SSE2FAST-NEXT:    movdqu %xmm0, (%rdi)
373*9880d681SAndroid Build Coastguard Worker; SSE2FAST-NEXT:    retq
374*9880d681SAndroid Build Coastguard Worker;
375*9880d681SAndroid Build Coastguard Worker; AVX1-LABEL: memset_128_nonconst_bytes:
376*9880d681SAndroid Build Coastguard Worker; AVX1:       # BB#0:
377*9880d681SAndroid Build Coastguard Worker; AVX1-NEXT:    vmovd %esi, %xmm0
378*9880d681SAndroid Build Coastguard Worker; AVX1-NEXT:    vpxor %xmm1, %xmm1, %xmm1
379*9880d681SAndroid Build Coastguard Worker; AVX1-NEXT:    vpshufb %xmm1, %xmm0, %xmm0
380*9880d681SAndroid Build Coastguard Worker; AVX1-NEXT:    vinsertf128 $1, %xmm0, %ymm0, %ymm0
381*9880d681SAndroid Build Coastguard Worker; AVX1-NEXT:    vmovups %ymm0, 96(%rdi)
382*9880d681SAndroid Build Coastguard Worker; AVX1-NEXT:    vmovups %ymm0, 64(%rdi)
383*9880d681SAndroid Build Coastguard Worker; AVX1-NEXT:    vmovups %ymm0, 32(%rdi)
384*9880d681SAndroid Build Coastguard Worker; AVX1-NEXT:    vmovups %ymm0, (%rdi)
385*9880d681SAndroid Build Coastguard Worker; AVX1-NEXT:    vzeroupper
386*9880d681SAndroid Build Coastguard Worker; AVX1-NEXT:    retq
387*9880d681SAndroid Build Coastguard Worker;
388*9880d681SAndroid Build Coastguard Worker; AVX2-LABEL: memset_128_nonconst_bytes:
389*9880d681SAndroid Build Coastguard Worker; AVX2:       # BB#0:
390*9880d681SAndroid Build Coastguard Worker; AVX2-NEXT:    vmovd %esi, %xmm0
391*9880d681SAndroid Build Coastguard Worker; AVX2-NEXT:    vpbroadcastb %xmm0, %ymm0
392*9880d681SAndroid Build Coastguard Worker; AVX2-NEXT:    vmovdqu %ymm0, 96(%rdi)
393*9880d681SAndroid Build Coastguard Worker; AVX2-NEXT:    vmovdqu %ymm0, 64(%rdi)
394*9880d681SAndroid Build Coastguard Worker; AVX2-NEXT:    vmovdqu %ymm0, 32(%rdi)
395*9880d681SAndroid Build Coastguard Worker; AVX2-NEXT:    vmovdqu %ymm0, (%rdi)
396*9880d681SAndroid Build Coastguard Worker; AVX2-NEXT:    vzeroupper
397*9880d681SAndroid Build Coastguard Worker; AVX2-NEXT:    retq
398*9880d681SAndroid Build Coastguard Worker;
399*9880d681SAndroid Build Coastguard Worker  tail call void @llvm.memset.p0i8.i64(i8* %x, i8 %c, i64 128, i32 1, i1 false)
400*9880d681SAndroid Build Coastguard Worker  ret void
401*9880d681SAndroid Build Coastguard Worker}
402*9880d681SAndroid Build Coastguard Worker
403*9880d681SAndroid Build Coastguard Workerdefine void @memset_256_nonconst_bytes(i8* %x, i8 %c) {
404*9880d681SAndroid Build Coastguard Worker; SSE-LABEL: memset_256_nonconst_bytes:
405*9880d681SAndroid Build Coastguard Worker; SSE:       # BB#0:
406*9880d681SAndroid Build Coastguard Worker; SSE-NEXT:    movl $256, %edx # imm = 0x100
407*9880d681SAndroid Build Coastguard Worker; SSE-NEXT:    jmp memset # TAILCALL
408*9880d681SAndroid Build Coastguard Worker;
409*9880d681SAndroid Build Coastguard Worker; SSE2FAST-LABEL: memset_256_nonconst_bytes:
410*9880d681SAndroid Build Coastguard Worker; SSE2FAST:       # BB#0:
411*9880d681SAndroid Build Coastguard Worker; SSE2FAST-NEXT:    movd %esi, %xmm0
412*9880d681SAndroid Build Coastguard Worker; SSE2FAST-NEXT:    punpcklbw {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
413*9880d681SAndroid Build Coastguard Worker; SSE2FAST-NEXT:    pshuflw {{.*#+}} xmm0 = xmm0[0,0,0,0,4,5,6,7]
414*9880d681SAndroid Build Coastguard Worker; SSE2FAST-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,1,1]
415*9880d681SAndroid Build Coastguard Worker; SSE2FAST-NEXT:    movdqu %xmm0, 240(%rdi)
416*9880d681SAndroid Build Coastguard Worker; SSE2FAST-NEXT:    movdqu %xmm0, 224(%rdi)
417*9880d681SAndroid Build Coastguard Worker; SSE2FAST-NEXT:    movdqu %xmm0, 208(%rdi)
418*9880d681SAndroid Build Coastguard Worker; SSE2FAST-NEXT:    movdqu %xmm0, 192(%rdi)
419*9880d681SAndroid Build Coastguard Worker; SSE2FAST-NEXT:    movdqu %xmm0, 176(%rdi)
420*9880d681SAndroid Build Coastguard Worker; SSE2FAST-NEXT:    movdqu %xmm0, 160(%rdi)
421*9880d681SAndroid Build Coastguard Worker; SSE2FAST-NEXT:    movdqu %xmm0, 144(%rdi)
422*9880d681SAndroid Build Coastguard Worker; SSE2FAST-NEXT:    movdqu %xmm0, 128(%rdi)
423*9880d681SAndroid Build Coastguard Worker; SSE2FAST-NEXT:    movdqu %xmm0, 112(%rdi)
424*9880d681SAndroid Build Coastguard Worker; SSE2FAST-NEXT:    movdqu %xmm0, 96(%rdi)
425*9880d681SAndroid Build Coastguard Worker; SSE2FAST-NEXT:    movdqu %xmm0, 80(%rdi)
426*9880d681SAndroid Build Coastguard Worker; SSE2FAST-NEXT:    movdqu %xmm0, 64(%rdi)
427*9880d681SAndroid Build Coastguard Worker; SSE2FAST-NEXT:    movdqu %xmm0, 48(%rdi)
428*9880d681SAndroid Build Coastguard Worker; SSE2FAST-NEXT:    movdqu %xmm0, 32(%rdi)
429*9880d681SAndroid Build Coastguard Worker; SSE2FAST-NEXT:    movdqu %xmm0, 16(%rdi)
430*9880d681SAndroid Build Coastguard Worker; SSE2FAST-NEXT:    movdqu %xmm0, (%rdi)
431*9880d681SAndroid Build Coastguard Worker; SSE2FAST-NEXT:    retq
432*9880d681SAndroid Build Coastguard Worker;
433*9880d681SAndroid Build Coastguard Worker; AVX1-LABEL: memset_256_nonconst_bytes:
434*9880d681SAndroid Build Coastguard Worker; AVX1:       # BB#0:
435*9880d681SAndroid Build Coastguard Worker; AVX1-NEXT:    vmovd %esi, %xmm0
436*9880d681SAndroid Build Coastguard Worker; AVX1-NEXT:    vpxor %xmm1, %xmm1, %xmm1
437*9880d681SAndroid Build Coastguard Worker; AVX1-NEXT:    vpshufb %xmm1, %xmm0, %xmm0
438*9880d681SAndroid Build Coastguard Worker; AVX1-NEXT:    vinsertf128 $1, %xmm0, %ymm0, %ymm0
439*9880d681SAndroid Build Coastguard Worker; AVX1-NEXT:    vmovups %ymm0, 224(%rdi)
440*9880d681SAndroid Build Coastguard Worker; AVX1-NEXT:    vmovups %ymm0, 192(%rdi)
441*9880d681SAndroid Build Coastguard Worker; AVX1-NEXT:    vmovups %ymm0, 160(%rdi)
442*9880d681SAndroid Build Coastguard Worker; AVX1-NEXT:    vmovups %ymm0, 128(%rdi)
443*9880d681SAndroid Build Coastguard Worker; AVX1-NEXT:    vmovups %ymm0, 96(%rdi)
444*9880d681SAndroid Build Coastguard Worker; AVX1-NEXT:    vmovups %ymm0, 64(%rdi)
445*9880d681SAndroid Build Coastguard Worker; AVX1-NEXT:    vmovups %ymm0, 32(%rdi)
446*9880d681SAndroid Build Coastguard Worker; AVX1-NEXT:    vmovups %ymm0, (%rdi)
447*9880d681SAndroid Build Coastguard Worker; AVX1-NEXT:    vzeroupper
448*9880d681SAndroid Build Coastguard Worker; AVX1-NEXT:    retq
449*9880d681SAndroid Build Coastguard Worker;
450*9880d681SAndroid Build Coastguard Worker; AVX2-LABEL: memset_256_nonconst_bytes:
451*9880d681SAndroid Build Coastguard Worker; AVX2:       # BB#0:
452*9880d681SAndroid Build Coastguard Worker; AVX2-NEXT:    vmovd %esi, %xmm0
453*9880d681SAndroid Build Coastguard Worker; AVX2-NEXT:    vpbroadcastb %xmm0, %ymm0
454*9880d681SAndroid Build Coastguard Worker; AVX2-NEXT:    vmovdqu %ymm0, 224(%rdi)
455*9880d681SAndroid Build Coastguard Worker; AVX2-NEXT:    vmovdqu %ymm0, 192(%rdi)
456*9880d681SAndroid Build Coastguard Worker; AVX2-NEXT:    vmovdqu %ymm0, 160(%rdi)
457*9880d681SAndroid Build Coastguard Worker; AVX2-NEXT:    vmovdqu %ymm0, 128(%rdi)
458*9880d681SAndroid Build Coastguard Worker; AVX2-NEXT:    vmovdqu %ymm0, 96(%rdi)
459*9880d681SAndroid Build Coastguard Worker; AVX2-NEXT:    vmovdqu %ymm0, 64(%rdi)
460*9880d681SAndroid Build Coastguard Worker; AVX2-NEXT:    vmovdqu %ymm0, 32(%rdi)
461*9880d681SAndroid Build Coastguard Worker; AVX2-NEXT:    vmovdqu %ymm0, (%rdi)
462*9880d681SAndroid Build Coastguard Worker; AVX2-NEXT:    vzeroupper
463*9880d681SAndroid Build Coastguard Worker; AVX2-NEXT:    retq
464*9880d681SAndroid Build Coastguard Worker;
465*9880d681SAndroid Build Coastguard Worker  tail call void @llvm.memset.p0i8.i64(i8* %x, i8 %c, i64 256, i32 1, i1 false)
466*9880d681SAndroid Build Coastguard Worker  ret void
467*9880d681SAndroid Build Coastguard Worker}
468*9880d681SAndroid Build Coastguard Worker
469*9880d681SAndroid Build Coastguard Workerdeclare void @llvm.memset.p0i8.i64(i8* nocapture, i8, i64, i32, i1) #1
470*9880d681SAndroid Build Coastguard Worker
471