xref: /llvm-project/llvm/test/Instrumentation/MemorySanitizer/AArch64/neon_vst.ll (revision 38fffa630ee80163dc65e759392ad29798905679) 
1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 4
2; Test memory sanitizer instrumentation for Arm NEON VST instructions.
3;
4; RUN: opt < %s -passes=msan -S | FileCheck %s
5;
6; Forked from llvm/test/CodeGen/AArch64/arm64-st1.ll
7
8target datalayout = "e-m:e-i8:8:32-i16:16:32-i64:64-i128:128-n32:64-S128"
9target triple = "aarch64--linux-android9001"
10
11; -----------------------------------------------------------------------------------------------------------------------------------------------
12
13define void @st2_8b(<8 x i8> %A, <8 x i8> %B, ptr %P) nounwind sanitize_memory {
14;
15;
16; CHECK-LABEL: define void @st2_8b(
17; CHECK-SAME: <8 x i8> [[A:%.*]], <8 x i8> [[B:%.*]], ptr [[P:%.*]]) #[[ATTR0:[0-9]+]] {
18; CHECK-NEXT:    [[TMP1:%.*]] = load i64, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 16) to ptr), align 8
19; CHECK-NEXT:    [[TMP2:%.*]] = load <8 x i8>, ptr @__msan_param_tls, align 8
20; CHECK-NEXT:    [[TMP3:%.*]] = load <8 x i8>, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 8) to ptr), align 8
21; CHECK-NEXT:    call void @llvm.donothing()
22; CHECK-NEXT:    [[TMP4:%.*]] = ptrtoint ptr [[P]] to i64
23; CHECK-NEXT:    [[TMP5:%.*]] = xor i64 [[TMP4]], 193514046488576
24; CHECK-NEXT:    [[TMP6:%.*]] = inttoptr i64 [[TMP5]] to ptr
25; CHECK-NEXT:    call void @llvm.aarch64.neon.st2.v8i8.p0(<8 x i8> [[TMP2]], <8 x i8> [[TMP3]], ptr [[TMP6]])
26; CHECK-NEXT:    [[_MSCMP:%.*]] = icmp ne i64 [[TMP1]], 0
27; CHECK-NEXT:    br i1 [[_MSCMP]], label [[TMP7:%.*]], label [[TMP8:%.*]], !prof [[PROF1:![0-9]+]]
28; CHECK:       7:
29; CHECK-NEXT:    call void @__msan_warning_noreturn() #[[ATTR4:[0-9]+]]
30; CHECK-NEXT:    unreachable
31; CHECK:       8:
32; CHECK-NEXT:    call void @llvm.aarch64.neon.st2.v8i8.p0(<8 x i8> [[A]], <8 x i8> [[B]], ptr [[P]])
33; CHECK-NEXT:    ret void
34;
35  call void @llvm.aarch64.neon.st2.v8i8.p0(<8 x i8> %A, <8 x i8> %B, ptr %P)
36  ret void
37}
38
39define void @st2_8b_undefA(<8 x i8> %A, <8 x i8> %B, ptr %P) nounwind sanitize_memory {
40;
41;
42; CHECK-LABEL: define void @st2_8b_undefA(
43; CHECK-SAME: <8 x i8> [[A:%.*]], <8 x i8> [[B:%.*]], ptr [[P:%.*]]) #[[ATTR0]] {
44; CHECK-NEXT:    [[TMP1:%.*]] = load i64, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 16) to ptr), align 8
45; CHECK-NEXT:    [[TMP2:%.*]] = load <8 x i8>, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 8) to ptr), align 8
46; CHECK-NEXT:    call void @llvm.donothing()
47; CHECK-NEXT:    [[TMP3:%.*]] = ptrtoint ptr [[P]] to i64
48; CHECK-NEXT:    [[TMP4:%.*]] = xor i64 [[TMP3]], 193514046488576
49; CHECK-NEXT:    [[TMP5:%.*]] = inttoptr i64 [[TMP4]] to ptr
50; CHECK-NEXT:    call void @llvm.aarch64.neon.st2.v8i8.p0(<8 x i8> splat (i8 -1), <8 x i8> [[TMP2]], ptr [[TMP5]])
51; CHECK-NEXT:    [[_MSCMP:%.*]] = icmp ne i64 [[TMP1]], 0
52; CHECK-NEXT:    br i1 [[_MSCMP]], label [[TMP6:%.*]], label [[TMP7:%.*]], !prof [[PROF1]]
53; CHECK:       6:
54; CHECK-NEXT:    call void @__msan_warning_noreturn() #[[ATTR4]]
55; CHECK-NEXT:    unreachable
56; CHECK:       7:
57; CHECK-NEXT:    call void @llvm.aarch64.neon.st2.v8i8.p0(<8 x i8> undef, <8 x i8> [[B]], ptr [[P]])
58; CHECK-NEXT:    ret void
59;
60  call void @llvm.aarch64.neon.st2.v8i8.p0(<8 x i8> undef, <8 x i8> %B, ptr %P)
61  ret void
62}
63
64define void @st2_8b_undefB(<8 x i8> %A, <8 x i8> %B, ptr %P) nounwind sanitize_memory {
65;
66;
67; CHECK-LABEL: define void @st2_8b_undefB(
68; CHECK-SAME: <8 x i8> [[A:%.*]], <8 x i8> [[B:%.*]], ptr [[P:%.*]]) #[[ATTR0]] {
69; CHECK-NEXT:    [[TMP1:%.*]] = load i64, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 16) to ptr), align 8
70; CHECK-NEXT:    [[TMP2:%.*]] = load <8 x i8>, ptr @__msan_param_tls, align 8
71; CHECK-NEXT:    call void @llvm.donothing()
72; CHECK-NEXT:    [[TMP3:%.*]] = ptrtoint ptr [[P]] to i64
73; CHECK-NEXT:    [[TMP4:%.*]] = xor i64 [[TMP3]], 193514046488576
74; CHECK-NEXT:    [[TMP5:%.*]] = inttoptr i64 [[TMP4]] to ptr
75; CHECK-NEXT:    call void @llvm.aarch64.neon.st2.v8i8.p0(<8 x i8> [[TMP2]], <8 x i8> splat (i8 -1), ptr [[TMP5]])
76; CHECK-NEXT:    [[_MSCMP:%.*]] = icmp ne i64 [[TMP1]], 0
77; CHECK-NEXT:    br i1 [[_MSCMP]], label [[TMP6:%.*]], label [[TMP7:%.*]], !prof [[PROF1]]
78; CHECK:       6:
79; CHECK-NEXT:    call void @__msan_warning_noreturn() #[[ATTR4]]
80; CHECK-NEXT:    unreachable
81; CHECK:       7:
82; CHECK-NEXT:    call void @llvm.aarch64.neon.st2.v8i8.p0(<8 x i8> [[A]], <8 x i8> undef, ptr [[P]])
83; CHECK-NEXT:    ret void
84;
85  call void @llvm.aarch64.neon.st2.v8i8.p0(<8 x i8> %A, <8 x i8> undef, ptr %P)
86  ret void
87}
88
89define void @st2_8b_undefAB(<8 x i8> %A, <8 x i8> %B, ptr %P) nounwind sanitize_memory {
90;
91;
92; CHECK-LABEL: define void @st2_8b_undefAB(
93; CHECK-SAME: <8 x i8> [[A:%.*]], <8 x i8> [[B:%.*]], ptr [[P:%.*]]) #[[ATTR0]] {
94; CHECK-NEXT:    [[TMP1:%.*]] = load i64, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 16) to ptr), align 8
95; CHECK-NEXT:    call void @llvm.donothing()
96; CHECK-NEXT:    [[TMP2:%.*]] = ptrtoint ptr [[P]] to i64
97; CHECK-NEXT:    [[TMP3:%.*]] = xor i64 [[TMP2]], 193514046488576
98; CHECK-NEXT:    [[TMP4:%.*]] = inttoptr i64 [[TMP3]] to ptr
99; CHECK-NEXT:    call void @llvm.aarch64.neon.st2.v8i8.p0(<8 x i8> splat (i8 -1), <8 x i8> splat (i8 -1), ptr [[TMP4]])
100; CHECK-NEXT:    [[_MSCMP:%.*]] = icmp ne i64 [[TMP1]], 0
101; CHECK-NEXT:    br i1 [[_MSCMP]], label [[TMP5:%.*]], label [[TMP6:%.*]], !prof [[PROF1]]
102; CHECK:       5:
103; CHECK-NEXT:    call void @__msan_warning_noreturn() #[[ATTR4]]
104; CHECK-NEXT:    unreachable
105; CHECK:       6:
106; CHECK-NEXT:    call void @llvm.aarch64.neon.st2.v8i8.p0(<8 x i8> undef, <8 x i8> undef, ptr [[P]])
107; CHECK-NEXT:    ret void
108;
109  call void @llvm.aarch64.neon.st2.v8i8.p0(<8 x i8> undef, <8 x i8> undef, ptr %P)
110  ret void
111}
112
113define void @st3_8b(<8 x i8> %A, <8 x i8> %B, <8 x i8> %C, ptr %P) nounwind sanitize_memory {
114;
115;
116; CHECK-LABEL: define void @st3_8b(
117; CHECK-SAME: <8 x i8> [[A:%.*]], <8 x i8> [[B:%.*]], <8 x i8> [[C:%.*]], ptr [[P:%.*]]) #[[ATTR0]] {
118; CHECK-NEXT:    [[TMP1:%.*]] = load i64, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 24) to ptr), align 8
119; CHECK-NEXT:    [[TMP2:%.*]] = load <8 x i8>, ptr @__msan_param_tls, align 8
120; CHECK-NEXT:    [[TMP3:%.*]] = load <8 x i8>, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 8) to ptr), align 8
121; CHECK-NEXT:    [[TMP4:%.*]] = load <8 x i8>, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 16) to ptr), align 8
122; CHECK-NEXT:    call void @llvm.donothing()
123; CHECK-NEXT:    [[TMP5:%.*]] = ptrtoint ptr [[P]] to i64
124; CHECK-NEXT:    [[TMP6:%.*]] = xor i64 [[TMP5]], 193514046488576
125; CHECK-NEXT:    [[TMP7:%.*]] = inttoptr i64 [[TMP6]] to ptr
126; CHECK-NEXT:    call void @llvm.aarch64.neon.st3.v8i8.p0(<8 x i8> [[TMP2]], <8 x i8> [[TMP3]], <8 x i8> [[TMP4]], ptr [[TMP7]])
127; CHECK-NEXT:    [[_MSCMP:%.*]] = icmp ne i64 [[TMP1]], 0
128; CHECK-NEXT:    br i1 [[_MSCMP]], label [[TMP8:%.*]], label [[TMP9:%.*]], !prof [[PROF1]]
129; CHECK:       8:
130; CHECK-NEXT:    call void @__msan_warning_noreturn() #[[ATTR4]]
131; CHECK-NEXT:    unreachable
132; CHECK:       9:
133; CHECK-NEXT:    call void @llvm.aarch64.neon.st3.v8i8.p0(<8 x i8> [[A]], <8 x i8> [[B]], <8 x i8> [[C]], ptr [[P]])
134; CHECK-NEXT:    ret void
135;
136  call void @llvm.aarch64.neon.st3.v8i8.p0(<8 x i8> %A, <8 x i8> %B, <8 x i8> %C, ptr %P)
137  ret void
138}
139
140define void @st3_8b_undefA(<8 x i8> %A, <8 x i8> %B, <8 x i8> %C, ptr %P) nounwind sanitize_memory {
141;
142;
143; CHECK-LABEL: define void @st3_8b_undefA(
144; CHECK-SAME: <8 x i8> [[A:%.*]], <8 x i8> [[B:%.*]], <8 x i8> [[C:%.*]], ptr [[P:%.*]]) #[[ATTR0]] {
145; CHECK-NEXT:    [[TMP1:%.*]] = load i64, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 24) to ptr), align 8
146; CHECK-NEXT:    [[TMP2:%.*]] = load <8 x i8>, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 8) to ptr), align 8
147; CHECK-NEXT:    [[TMP3:%.*]] = load <8 x i8>, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 16) to ptr), align 8
148; CHECK-NEXT:    call void @llvm.donothing()
149; CHECK-NEXT:    [[TMP4:%.*]] = ptrtoint ptr [[P]] to i64
150; CHECK-NEXT:    [[TMP5:%.*]] = xor i64 [[TMP4]], 193514046488576
151; CHECK-NEXT:    [[TMP6:%.*]] = inttoptr i64 [[TMP5]] to ptr
152; CHECK-NEXT:    call void @llvm.aarch64.neon.st3.v8i8.p0(<8 x i8> splat (i8 -1), <8 x i8> [[TMP2]], <8 x i8> [[TMP3]], ptr [[TMP6]])
153; CHECK-NEXT:    [[_MSCMP:%.*]] = icmp ne i64 [[TMP1]], 0
154; CHECK-NEXT:    br i1 [[_MSCMP]], label [[TMP7:%.*]], label [[TMP8:%.*]], !prof [[PROF1]]
155; CHECK:       7:
156; CHECK-NEXT:    call void @__msan_warning_noreturn() #[[ATTR4]]
157; CHECK-NEXT:    unreachable
158; CHECK:       8:
159; CHECK-NEXT:    call void @llvm.aarch64.neon.st3.v8i8.p0(<8 x i8> undef, <8 x i8> [[B]], <8 x i8> [[C]], ptr [[P]])
160; CHECK-NEXT:    ret void
161;
162  call void @llvm.aarch64.neon.st3.v8i8.p0(<8 x i8> undef, <8 x i8> %B, <8 x i8> %C, ptr %P)
163  ret void
164}
165
166define void @st3_8b_undefB(<8 x i8> %A, <8 x i8> %B, <8 x i8> %C, ptr %P) nounwind sanitize_memory {
167;
168;
169; CHECK-LABEL: define void @st3_8b_undefB(
170; CHECK-SAME: <8 x i8> [[A:%.*]], <8 x i8> [[B:%.*]], <8 x i8> [[C:%.*]], ptr [[P:%.*]]) #[[ATTR0]] {
171; CHECK-NEXT:    [[TMP1:%.*]] = load i64, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 24) to ptr), align 8
172; CHECK-NEXT:    [[TMP2:%.*]] = load <8 x i8>, ptr @__msan_param_tls, align 8
173; CHECK-NEXT:    [[TMP3:%.*]] = load <8 x i8>, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 16) to ptr), align 8
174; CHECK-NEXT:    call void @llvm.donothing()
175; CHECK-NEXT:    [[TMP4:%.*]] = ptrtoint ptr [[P]] to i64
176; CHECK-NEXT:    [[TMP5:%.*]] = xor i64 [[TMP4]], 193514046488576
177; CHECK-NEXT:    [[TMP6:%.*]] = inttoptr i64 [[TMP5]] to ptr
178; CHECK-NEXT:    call void @llvm.aarch64.neon.st3.v8i8.p0(<8 x i8> [[TMP2]], <8 x i8> splat (i8 -1), <8 x i8> [[TMP3]], ptr [[TMP6]])
179; CHECK-NEXT:    [[_MSCMP:%.*]] = icmp ne i64 [[TMP1]], 0
180; CHECK-NEXT:    br i1 [[_MSCMP]], label [[TMP7:%.*]], label [[TMP8:%.*]], !prof [[PROF1]]
181; CHECK:       7:
182; CHECK-NEXT:    call void @__msan_warning_noreturn() #[[ATTR4]]
183; CHECK-NEXT:    unreachable
184; CHECK:       8:
185; CHECK-NEXT:    call void @llvm.aarch64.neon.st3.v8i8.p0(<8 x i8> [[A]], <8 x i8> undef, <8 x i8> [[C]], ptr [[P]])
186; CHECK-NEXT:    ret void
187;
188  call void @llvm.aarch64.neon.st3.v8i8.p0(<8 x i8> %A, <8 x i8> undef, <8 x i8> %C, ptr %P)
189  ret void
190}
191
192define void @st3_8b_undefC(<8 x i8> %A, <8 x i8> %B, <8 x i8> %C, ptr %P) nounwind sanitize_memory {
193;
194;
195; CHECK-LABEL: define void @st3_8b_undefC(
196; CHECK-SAME: <8 x i8> [[A:%.*]], <8 x i8> [[B:%.*]], <8 x i8> [[C:%.*]], ptr [[P:%.*]]) #[[ATTR0]] {
197; CHECK-NEXT:    [[TMP1:%.*]] = load i64, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 24) to ptr), align 8
198; CHECK-NEXT:    [[TMP2:%.*]] = load <8 x i8>, ptr @__msan_param_tls, align 8
199; CHECK-NEXT:    [[TMP3:%.*]] = load <8 x i8>, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 8) to ptr), align 8
200; CHECK-NEXT:    call void @llvm.donothing()
201; CHECK-NEXT:    [[TMP4:%.*]] = ptrtoint ptr [[P]] to i64
202; CHECK-NEXT:    [[TMP5:%.*]] = xor i64 [[TMP4]], 193514046488576
203; CHECK-NEXT:    [[TMP6:%.*]] = inttoptr i64 [[TMP5]] to ptr
204; CHECK-NEXT:    call void @llvm.aarch64.neon.st3.v8i8.p0(<8 x i8> [[TMP2]], <8 x i8> [[TMP3]], <8 x i8> splat (i8 -1), ptr [[TMP6]])
205; CHECK-NEXT:    [[_MSCMP:%.*]] = icmp ne i64 [[TMP1]], 0
206; CHECK-NEXT:    br i1 [[_MSCMP]], label [[TMP7:%.*]], label [[TMP8:%.*]], !prof [[PROF1]]
207; CHECK:       7:
208; CHECK-NEXT:    call void @__msan_warning_noreturn() #[[ATTR4]]
209; CHECK-NEXT:    unreachable
210; CHECK:       8:
211; CHECK-NEXT:    call void @llvm.aarch64.neon.st3.v8i8.p0(<8 x i8> [[A]], <8 x i8> [[B]], <8 x i8> undef, ptr [[P]])
212; CHECK-NEXT:    ret void
213;
214  call void @llvm.aarch64.neon.st3.v8i8.p0(<8 x i8> %A, <8 x i8> %B, <8 x i8> undef, ptr %P)
215  ret void
216}
217
218define void @st3_8b_undefAB(<8 x i8> %A, <8 x i8> %B, <8 x i8> %C, ptr %P) nounwind sanitize_memory {
219;
220;
221; CHECK-LABEL: define void @st3_8b_undefAB(
222; CHECK-SAME: <8 x i8> [[A:%.*]], <8 x i8> [[B:%.*]], <8 x i8> [[C:%.*]], ptr [[P:%.*]]) #[[ATTR0]] {
223; CHECK-NEXT:    [[TMP1:%.*]] = load i64, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 24) to ptr), align 8
224; CHECK-NEXT:    [[TMP2:%.*]] = load <8 x i8>, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 16) to ptr), align 8
225; CHECK-NEXT:    call void @llvm.donothing()
226; CHECK-NEXT:    [[TMP3:%.*]] = ptrtoint ptr [[P]] to i64
227; CHECK-NEXT:    [[TMP4:%.*]] = xor i64 [[TMP3]], 193514046488576
228; CHECK-NEXT:    [[TMP5:%.*]] = inttoptr i64 [[TMP4]] to ptr
229; CHECK-NEXT:    call void @llvm.aarch64.neon.st3.v8i8.p0(<8 x i8> splat (i8 -1), <8 x i8> splat (i8 -1), <8 x i8> [[TMP2]], ptr [[TMP5]])
230; CHECK-NEXT:    [[_MSCMP:%.*]] = icmp ne i64 [[TMP1]], 0
231; CHECK-NEXT:    br i1 [[_MSCMP]], label [[TMP6:%.*]], label [[TMP7:%.*]], !prof [[PROF1]]
232; CHECK:       6:
233; CHECK-NEXT:    call void @__msan_warning_noreturn() #[[ATTR4]]
234; CHECK-NEXT:    unreachable
235; CHECK:       7:
236; CHECK-NEXT:    call void @llvm.aarch64.neon.st3.v8i8.p0(<8 x i8> undef, <8 x i8> undef, <8 x i8> [[C]], ptr [[P]])
237; CHECK-NEXT:    ret void
238;
239  call void @llvm.aarch64.neon.st3.v8i8.p0(<8 x i8> undef, <8 x i8> undef, <8 x i8> %C, ptr %P)
240  ret void
241}
242
243define void @st3_8b_undefAC(<8 x i8> %A, <8 x i8> %B, <8 x i8> %C, ptr %P) nounwind sanitize_memory {
244;
245;
246; CHECK-LABEL: define void @st3_8b_undefAC(
247; CHECK-SAME: <8 x i8> [[A:%.*]], <8 x i8> [[B:%.*]], <8 x i8> [[C:%.*]], ptr [[P:%.*]]) #[[ATTR0]] {
248; CHECK-NEXT:    [[TMP1:%.*]] = load i64, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 24) to ptr), align 8
249; CHECK-NEXT:    [[TMP2:%.*]] = load <8 x i8>, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 8) to ptr), align 8
250; CHECK-NEXT:    call void @llvm.donothing()
251; CHECK-NEXT:    [[TMP3:%.*]] = ptrtoint ptr [[P]] to i64
252; CHECK-NEXT:    [[TMP4:%.*]] = xor i64 [[TMP3]], 193514046488576
253; CHECK-NEXT:    [[TMP5:%.*]] = inttoptr i64 [[TMP4]] to ptr
254; CHECK-NEXT:    call void @llvm.aarch64.neon.st3.v8i8.p0(<8 x i8> splat (i8 -1), <8 x i8> [[TMP2]], <8 x i8> splat (i8 -1), ptr [[TMP5]])
255; CHECK-NEXT:    [[_MSCMP:%.*]] = icmp ne i64 [[TMP1]], 0
256; CHECK-NEXT:    br i1 [[_MSCMP]], label [[TMP6:%.*]], label [[TMP7:%.*]], !prof [[PROF1]]
257; CHECK:       6:
258; CHECK-NEXT:    call void @__msan_warning_noreturn() #[[ATTR4]]
259; CHECK-NEXT:    unreachable
260; CHECK:       7:
261; CHECK-NEXT:    call void @llvm.aarch64.neon.st3.v8i8.p0(<8 x i8> undef, <8 x i8> [[B]], <8 x i8> undef, ptr [[P]])
262; CHECK-NEXT:    ret void
263;
264  call void @llvm.aarch64.neon.st3.v8i8.p0(<8 x i8> undef, <8 x i8> %B, <8 x i8> undef, ptr %P)
265  ret void
266}
267
268define void @st3_8b_undefBC(<8 x i8> %A, <8 x i8> %B, <8 x i8> %C, ptr %P) nounwind sanitize_memory {
269;
270;
271; CHECK-LABEL: define void @st3_8b_undefBC(
272; CHECK-SAME: <8 x i8> [[A:%.*]], <8 x i8> [[B:%.*]], <8 x i8> [[C:%.*]], ptr [[P:%.*]]) #[[ATTR0]] {
273; CHECK-NEXT:    [[TMP1:%.*]] = load i64, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 24) to ptr), align 8
274; CHECK-NEXT:    [[TMP2:%.*]] = load <8 x i8>, ptr @__msan_param_tls, align 8
275; CHECK-NEXT:    call void @llvm.donothing()
276; CHECK-NEXT:    [[TMP3:%.*]] = ptrtoint ptr [[P]] to i64
277; CHECK-NEXT:    [[TMP4:%.*]] = xor i64 [[TMP3]], 193514046488576
278; CHECK-NEXT:    [[TMP5:%.*]] = inttoptr i64 [[TMP4]] to ptr
279; CHECK-NEXT:    call void @llvm.aarch64.neon.st3.v8i8.p0(<8 x i8> [[TMP2]], <8 x i8> splat (i8 -1), <8 x i8> splat (i8 -1), ptr [[TMP5]])
280; CHECK-NEXT:    [[_MSCMP:%.*]] = icmp ne i64 [[TMP1]], 0
281; CHECK-NEXT:    br i1 [[_MSCMP]], label [[TMP6:%.*]], label [[TMP7:%.*]], !prof [[PROF1]]
282; CHECK:       6:
283; CHECK-NEXT:    call void @__msan_warning_noreturn() #[[ATTR4]]
284; CHECK-NEXT:    unreachable
285; CHECK:       7:
286; CHECK-NEXT:    call void @llvm.aarch64.neon.st3.v8i8.p0(<8 x i8> [[A]], <8 x i8> undef, <8 x i8> undef, ptr [[P]])
287; CHECK-NEXT:    ret void
288;
289  call void @llvm.aarch64.neon.st3.v8i8.p0(<8 x i8> %A, <8 x i8> undef, <8 x i8> undef, ptr %P)
290  ret void
291}
292
293define void @st3_8b_undefABC(<8 x i8> %A, <8 x i8> %B, <8 x i8> %C, ptr %P) nounwind sanitize_memory {
294;
295;
296; CHECK-LABEL: define void @st3_8b_undefABC(
297; CHECK-SAME: <8 x i8> [[A:%.*]], <8 x i8> [[B:%.*]], <8 x i8> [[C:%.*]], ptr [[P:%.*]]) #[[ATTR0]] {
298; CHECK-NEXT:    [[TMP1:%.*]] = load i64, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 24) to ptr), align 8
299; CHECK-NEXT:    call void @llvm.donothing()
300; CHECK-NEXT:    [[TMP2:%.*]] = ptrtoint ptr [[P]] to i64
301; CHECK-NEXT:    [[TMP3:%.*]] = xor i64 [[TMP2]], 193514046488576
302; CHECK-NEXT:    [[TMP4:%.*]] = inttoptr i64 [[TMP3]] to ptr
303; CHECK-NEXT:    call void @llvm.aarch64.neon.st3.v8i8.p0(<8 x i8> splat (i8 -1), <8 x i8> splat (i8 -1), <8 x i8> splat (i8 -1), ptr [[TMP4]])
304; CHECK-NEXT:    [[_MSCMP:%.*]] = icmp ne i64 [[TMP1]], 0
305; CHECK-NEXT:    br i1 [[_MSCMP]], label [[TMP5:%.*]], label [[TMP6:%.*]], !prof [[PROF1]]
306; CHECK:       5:
307; CHECK-NEXT:    call void @__msan_warning_noreturn() #[[ATTR4]]
308; CHECK-NEXT:    unreachable
309; CHECK:       6:
310; CHECK-NEXT:    call void @llvm.aarch64.neon.st3.v8i8.p0(<8 x i8> undef, <8 x i8> undef, <8 x i8> undef, ptr [[P]])
311; CHECK-NEXT:    ret void
312;
313  call void @llvm.aarch64.neon.st3.v8i8.p0(<8 x i8> undef, <8 x i8> undef, <8 x i8> undef, ptr %P)
314  ret void
315}
316
317define void @st4_8b(<8 x i8> %A, <8 x i8> %B, <8 x i8> %C, <8 x i8> %D, ptr %P) nounwind sanitize_memory {
318;
319;
320; CHECK-LABEL: define void @st4_8b(
321; CHECK-SAME: <8 x i8> [[A:%.*]], <8 x i8> [[B:%.*]], <8 x i8> [[C:%.*]], <8 x i8> [[D:%.*]], ptr [[P:%.*]]) #[[ATTR0]] {
322; CHECK-NEXT:    [[TMP1:%.*]] = load i64, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 32) to ptr), align 8
323; CHECK-NEXT:    [[TMP2:%.*]] = load <8 x i8>, ptr @__msan_param_tls, align 8
324; CHECK-NEXT:    [[TMP3:%.*]] = load <8 x i8>, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 8) to ptr), align 8
325; CHECK-NEXT:    [[TMP4:%.*]] = load <8 x i8>, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 16) to ptr), align 8
326; CHECK-NEXT:    [[TMP5:%.*]] = load <8 x i8>, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 24) to ptr), align 8
327; CHECK-NEXT:    call void @llvm.donothing()
328; CHECK-NEXT:    [[TMP6:%.*]] = ptrtoint ptr [[P]] to i64
329; CHECK-NEXT:    [[TMP7:%.*]] = xor i64 [[TMP6]], 193514046488576
330; CHECK-NEXT:    [[TMP8:%.*]] = inttoptr i64 [[TMP7]] to ptr
331; CHECK-NEXT:    call void @llvm.aarch64.neon.st4.v8i8.p0(<8 x i8> [[TMP2]], <8 x i8> [[TMP3]], <8 x i8> [[TMP4]], <8 x i8> [[TMP5]], ptr [[TMP8]])
332; CHECK-NEXT:    [[_MSCMP:%.*]] = icmp ne i64 [[TMP1]], 0
333; CHECK-NEXT:    br i1 [[_MSCMP]], label [[TMP9:%.*]], label [[TMP10:%.*]], !prof [[PROF1]]
334; CHECK:       9:
335; CHECK-NEXT:    call void @__msan_warning_noreturn() #[[ATTR4]]
336; CHECK-NEXT:    unreachable
337; CHECK:       10:
338; CHECK-NEXT:    call void @llvm.aarch64.neon.st4.v8i8.p0(<8 x i8> [[A]], <8 x i8> [[B]], <8 x i8> [[C]], <8 x i8> [[D]], ptr [[P]])
339; CHECK-NEXT:    ret void
340;
341  call void @llvm.aarch64.neon.st4.v8i8.p0(<8 x i8> %A, <8 x i8> %B, <8 x i8> %C, <8 x i8> %D, ptr %P)
342  ret void
343}
344
345define void @st4_8b_undefA(<8 x i8> %A, <8 x i8> %B, <8 x i8> %C, <8 x i8> %D, ptr %P) nounwind sanitize_memory {
346;
347;
348; CHECK-LABEL: define void @st4_8b_undefA(
349; CHECK-SAME: <8 x i8> [[A:%.*]], <8 x i8> [[B:%.*]], <8 x i8> [[C:%.*]], <8 x i8> [[D:%.*]], ptr [[P:%.*]]) #[[ATTR0]] {
350; CHECK-NEXT:    [[TMP1:%.*]] = load i64, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 32) to ptr), align 8
351; CHECK-NEXT:    [[TMP2:%.*]] = load <8 x i8>, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 8) to ptr), align 8
352; CHECK-NEXT:    [[TMP3:%.*]] = load <8 x i8>, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 16) to ptr), align 8
353; CHECK-NEXT:    [[TMP4:%.*]] = load <8 x i8>, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 24) to ptr), align 8
354; CHECK-NEXT:    call void @llvm.donothing()
355; CHECK-NEXT:    [[TMP5:%.*]] = ptrtoint ptr [[P]] to i64
356; CHECK-NEXT:    [[TMP6:%.*]] = xor i64 [[TMP5]], 193514046488576
357; CHECK-NEXT:    [[TMP7:%.*]] = inttoptr i64 [[TMP6]] to ptr
358; CHECK-NEXT:    call void @llvm.aarch64.neon.st4.v8i8.p0(<8 x i8> splat (i8 -1), <8 x i8> [[TMP2]], <8 x i8> [[TMP3]], <8 x i8> [[TMP4]], ptr [[TMP7]])
359; CHECK-NEXT:    [[_MSCMP:%.*]] = icmp ne i64 [[TMP1]], 0
360; CHECK-NEXT:    br i1 [[_MSCMP]], label [[TMP8:%.*]], label [[TMP9:%.*]], !prof [[PROF1]]
361; CHECK:       8:
362; CHECK-NEXT:    call void @__msan_warning_noreturn() #[[ATTR4]]
363; CHECK-NEXT:    unreachable
364; CHECK:       9:
365; CHECK-NEXT:    call void @llvm.aarch64.neon.st4.v8i8.p0(<8 x i8> undef, <8 x i8> [[B]], <8 x i8> [[C]], <8 x i8> [[D]], ptr [[P]])
366; CHECK-NEXT:    ret void
367;
368  call void @llvm.aarch64.neon.st4.v8i8.p0(<8 x i8> undef, <8 x i8> %B, <8 x i8> %C, <8 x i8> %D, ptr %P)
369  ret void
370}
371
372define void @st4_8b_undefB(<8 x i8> %A, <8 x i8> %B, <8 x i8> %C, <8 x i8> %D, ptr %P) nounwind sanitize_memory {
373;
374;
375; CHECK-LABEL: define void @st4_8b_undefB(
376; CHECK-SAME: <8 x i8> [[A:%.*]], <8 x i8> [[B:%.*]], <8 x i8> [[C:%.*]], <8 x i8> [[D:%.*]], ptr [[P:%.*]]) #[[ATTR0]] {
377; CHECK-NEXT:    [[TMP1:%.*]] = load i64, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 32) to ptr), align 8
378; CHECK-NEXT:    [[TMP2:%.*]] = load <8 x i8>, ptr @__msan_param_tls, align 8
379; CHECK-NEXT:    [[TMP3:%.*]] = load <8 x i8>, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 16) to ptr), align 8
380; CHECK-NEXT:    [[TMP4:%.*]] = load <8 x i8>, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 24) to ptr), align 8
381; CHECK-NEXT:    call void @llvm.donothing()
382; CHECK-NEXT:    [[TMP5:%.*]] = ptrtoint ptr [[P]] to i64
383; CHECK-NEXT:    [[TMP6:%.*]] = xor i64 [[TMP5]], 193514046488576
384; CHECK-NEXT:    [[TMP7:%.*]] = inttoptr i64 [[TMP6]] to ptr
385; CHECK-NEXT:    call void @llvm.aarch64.neon.st4.v8i8.p0(<8 x i8> [[TMP2]], <8 x i8> splat (i8 -1), <8 x i8> [[TMP3]], <8 x i8> [[TMP4]], ptr [[TMP7]])
386; CHECK-NEXT:    [[_MSCMP:%.*]] = icmp ne i64 [[TMP1]], 0
387; CHECK-NEXT:    br i1 [[_MSCMP]], label [[TMP8:%.*]], label [[TMP9:%.*]], !prof [[PROF1]]
388; CHECK:       8:
389; CHECK-NEXT:    call void @__msan_warning_noreturn() #[[ATTR4]]
390; CHECK-NEXT:    unreachable
391; CHECK:       9:
392; CHECK-NEXT:    call void @llvm.aarch64.neon.st4.v8i8.p0(<8 x i8> [[A]], <8 x i8> undef, <8 x i8> [[C]], <8 x i8> [[D]], ptr [[P]])
393; CHECK-NEXT:    ret void
394;
395  call void @llvm.aarch64.neon.st4.v8i8.p0(<8 x i8> %A, <8 x i8> undef, <8 x i8> %C, <8 x i8> %D, ptr %P)
396  ret void
397}
398
399define void @st4_8b_undefC(<8 x i8> %A, <8 x i8> %B, <8 x i8> %C, <8 x i8> %D, ptr %P) nounwind sanitize_memory {
400;
401;
402; CHECK-LABEL: define void @st4_8b_undefC(
403; CHECK-SAME: <8 x i8> [[A:%.*]], <8 x i8> [[B:%.*]], <8 x i8> [[C:%.*]], <8 x i8> [[D:%.*]], ptr [[P:%.*]]) #[[ATTR0]] {
404; CHECK-NEXT:    [[TMP1:%.*]] = load i64, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 32) to ptr), align 8
405; CHECK-NEXT:    [[TMP2:%.*]] = load <8 x i8>, ptr @__msan_param_tls, align 8
406; CHECK-NEXT:    [[TMP3:%.*]] = load <8 x i8>, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 8) to ptr), align 8
407; CHECK-NEXT:    [[TMP4:%.*]] = load <8 x i8>, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 24) to ptr), align 8
408; CHECK-NEXT:    call void @llvm.donothing()
409; CHECK-NEXT:    [[TMP5:%.*]] = ptrtoint ptr [[P]] to i64
410; CHECK-NEXT:    [[TMP6:%.*]] = xor i64 [[TMP5]], 193514046488576
411; CHECK-NEXT:    [[TMP7:%.*]] = inttoptr i64 [[TMP6]] to ptr
412; CHECK-NEXT:    call void @llvm.aarch64.neon.st4.v8i8.p0(<8 x i8> [[TMP2]], <8 x i8> [[TMP3]], <8 x i8> splat (i8 -1), <8 x i8> [[TMP4]], ptr [[TMP7]])
413; CHECK-NEXT:    [[_MSCMP:%.*]] = icmp ne i64 [[TMP1]], 0
414; CHECK-NEXT:    br i1 [[_MSCMP]], label [[TMP8:%.*]], label [[TMP9:%.*]], !prof [[PROF1]]
415; CHECK:       8:
416; CHECK-NEXT:    call void @__msan_warning_noreturn() #[[ATTR4]]
417; CHECK-NEXT:    unreachable
418; CHECK:       9:
419; CHECK-NEXT:    call void @llvm.aarch64.neon.st4.v8i8.p0(<8 x i8> [[A]], <8 x i8> [[B]], <8 x i8> undef, <8 x i8> [[D]], ptr [[P]])
420; CHECK-NEXT:    ret void
421;
422  call void @llvm.aarch64.neon.st4.v8i8.p0(<8 x i8> %A, <8 x i8> %B, <8 x i8> undef, <8 x i8> %D, ptr %P)
423  ret void
424}
425
426define void @st4_8b_undefD(<8 x i8> %A, <8 x i8> %B, <8 x i8> %C, <8 x i8> %D, ptr %P) nounwind sanitize_memory {
427;
428;
429; CHECK-LABEL: define void @st4_8b_undefD(
430; CHECK-SAME: <8 x i8> [[A:%.*]], <8 x i8> [[B:%.*]], <8 x i8> [[C:%.*]], <8 x i8> [[D:%.*]], ptr [[P:%.*]]) #[[ATTR0]] {
431; CHECK-NEXT:    [[TMP1:%.*]] = load i64, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 32) to ptr), align 8
432; CHECK-NEXT:    [[TMP2:%.*]] = load <8 x i8>, ptr @__msan_param_tls, align 8
433; CHECK-NEXT:    [[TMP3:%.*]] = load <8 x i8>, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 8) to ptr), align 8
434; CHECK-NEXT:    [[TMP4:%.*]] = load <8 x i8>, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 16) to ptr), align 8
435; CHECK-NEXT:    call void @llvm.donothing()
436; CHECK-NEXT:    [[TMP5:%.*]] = ptrtoint ptr [[P]] to i64
437; CHECK-NEXT:    [[TMP6:%.*]] = xor i64 [[TMP5]], 193514046488576
438; CHECK-NEXT:    [[TMP7:%.*]] = inttoptr i64 [[TMP6]] to ptr
439; CHECK-NEXT:    call void @llvm.aarch64.neon.st4.v8i8.p0(<8 x i8> [[TMP2]], <8 x i8> [[TMP3]], <8 x i8> [[TMP4]], <8 x i8> splat (i8 -1), ptr [[TMP7]])
440; CHECK-NEXT:    [[_MSCMP:%.*]] = icmp ne i64 [[TMP1]], 0
441; CHECK-NEXT:    br i1 [[_MSCMP]], label [[TMP8:%.*]], label [[TMP9:%.*]], !prof [[PROF1]]
442; CHECK:       8:
443; CHECK-NEXT:    call void @__msan_warning_noreturn() #[[ATTR4]]
444; CHECK-NEXT:    unreachable
445; CHECK:       9:
446; CHECK-NEXT:    call void @llvm.aarch64.neon.st4.v8i8.p0(<8 x i8> [[A]], <8 x i8> [[B]], <8 x i8> [[C]], <8 x i8> undef, ptr [[P]])
447; CHECK-NEXT:    ret void
448;
449  call void @llvm.aarch64.neon.st4.v8i8.p0(<8 x i8> %A, <8 x i8> %B, <8 x i8> %C, <8 x i8> undef, ptr %P)
450  ret void
451}
452
453define void @st4_8b_undefAB(<8 x i8> %A, <8 x i8> %B, <8 x i8> %C, <8 x i8> %D, ptr %P) nounwind sanitize_memory {
454;
455;
456; CHECK-LABEL: define void @st4_8b_undefAB(
457; CHECK-SAME: <8 x i8> [[A:%.*]], <8 x i8> [[B:%.*]], <8 x i8> [[C:%.*]], <8 x i8> [[D:%.*]], ptr [[P:%.*]]) #[[ATTR0]] {
458; CHECK-NEXT:    [[TMP1:%.*]] = load i64, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 32) to ptr), align 8
459; CHECK-NEXT:    [[TMP2:%.*]] = load <8 x i8>, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 16) to ptr), align 8
460; CHECK-NEXT:    [[TMP3:%.*]] = load <8 x i8>, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 24) to ptr), align 8
461; CHECK-NEXT:    call void @llvm.donothing()
462; CHECK-NEXT:    [[TMP4:%.*]] = ptrtoint ptr [[P]] to i64
463; CHECK-NEXT:    [[TMP5:%.*]] = xor i64 [[TMP4]], 193514046488576
464; CHECK-NEXT:    [[TMP6:%.*]] = inttoptr i64 [[TMP5]] to ptr
465; CHECK-NEXT:    call void @llvm.aarch64.neon.st4.v8i8.p0(<8 x i8> splat (i8 -1), <8 x i8> splat (i8 -1), <8 x i8> [[TMP2]], <8 x i8> [[TMP3]], ptr [[TMP6]])
466; CHECK-NEXT:    [[_MSCMP:%.*]] = icmp ne i64 [[TMP1]], 0
467; CHECK-NEXT:    br i1 [[_MSCMP]], label [[TMP7:%.*]], label [[TMP8:%.*]], !prof [[PROF1]]
468; CHECK:       7:
469; CHECK-NEXT:    call void @__msan_warning_noreturn() #[[ATTR4]]
470; CHECK-NEXT:    unreachable
471; CHECK:       8:
472; CHECK-NEXT:    call void @llvm.aarch64.neon.st4.v8i8.p0(<8 x i8> undef, <8 x i8> undef, <8 x i8> [[C]], <8 x i8> [[D]], ptr [[P]])
473; CHECK-NEXT:    ret void
474;
475  call void @llvm.aarch64.neon.st4.v8i8.p0(<8 x i8> undef, <8 x i8> undef, <8 x i8> %C, <8 x i8> %D, ptr %P)
476  ret void
477}
478
479define void @st4_8b_undefAC(<8 x i8> %A, <8 x i8> %B, <8 x i8> %C, <8 x i8> %D, ptr %P) nounwind sanitize_memory {
480;
481;
482; CHECK-LABEL: define void @st4_8b_undefAC(
483; CHECK-SAME: <8 x i8> [[A:%.*]], <8 x i8> [[B:%.*]], <8 x i8> [[C:%.*]], <8 x i8> [[D:%.*]], ptr [[P:%.*]]) #[[ATTR0]] {
484; CHECK-NEXT:    [[TMP1:%.*]] = load i64, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 32) to ptr), align 8
485; CHECK-NEXT:    [[TMP2:%.*]] = load <8 x i8>, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 8) to ptr), align 8
486; CHECK-NEXT:    [[TMP3:%.*]] = load <8 x i8>, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 24) to ptr), align 8
487; CHECK-NEXT:    call void @llvm.donothing()
488; CHECK-NEXT:    [[TMP4:%.*]] = ptrtoint ptr [[P]] to i64
489; CHECK-NEXT:    [[TMP5:%.*]] = xor i64 [[TMP4]], 193514046488576
490; CHECK-NEXT:    [[TMP6:%.*]] = inttoptr i64 [[TMP5]] to ptr
491; CHECK-NEXT:    call void @llvm.aarch64.neon.st4.v8i8.p0(<8 x i8> splat (i8 -1), <8 x i8> [[TMP2]], <8 x i8> splat (i8 -1), <8 x i8> [[TMP3]], ptr [[TMP6]])
492; CHECK-NEXT:    [[_MSCMP:%.*]] = icmp ne i64 [[TMP1]], 0
493; CHECK-NEXT:    br i1 [[_MSCMP]], label [[TMP7:%.*]], label [[TMP8:%.*]], !prof [[PROF1]]
494; CHECK:       7:
495; CHECK-NEXT:    call void @__msan_warning_noreturn() #[[ATTR4]]
496; CHECK-NEXT:    unreachable
497; CHECK:       8:
498; CHECK-NEXT:    call void @llvm.aarch64.neon.st4.v8i8.p0(<8 x i8> undef, <8 x i8> [[B]], <8 x i8> undef, <8 x i8> [[D]], ptr [[P]])
499; CHECK-NEXT:    ret void
500;
501  call void @llvm.aarch64.neon.st4.v8i8.p0(<8 x i8> undef, <8 x i8> %B, <8 x i8> undef, <8 x i8> %D, ptr %P)
502  ret void
503}
504
505define void @st4_8b_undefBC(<8 x i8> %A, <8 x i8> %B, <8 x i8> %C, <8 x i8> %D, ptr %P) nounwind sanitize_memory {
506;
507;
508; CHECK-LABEL: define void @st4_8b_undefBC(
509; CHECK-SAME: <8 x i8> [[A:%.*]], <8 x i8> [[B:%.*]], <8 x i8> [[C:%.*]], <8 x i8> [[D:%.*]], ptr [[P:%.*]]) #[[ATTR0]] {
510; CHECK-NEXT:    [[TMP1:%.*]] = load i64, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 32) to ptr), align 8
511; CHECK-NEXT:    [[TMP2:%.*]] = load <8 x i8>, ptr @__msan_param_tls, align 8
512; CHECK-NEXT:    [[TMP3:%.*]] = load <8 x i8>, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 24) to ptr), align 8
513; CHECK-NEXT:    call void @llvm.donothing()
514; CHECK-NEXT:    [[TMP4:%.*]] = ptrtoint ptr [[P]] to i64
515; CHECK-NEXT:    [[TMP5:%.*]] = xor i64 [[TMP4]], 193514046488576
516; CHECK-NEXT:    [[TMP6:%.*]] = inttoptr i64 [[TMP5]] to ptr
517; CHECK-NEXT:    call void @llvm.aarch64.neon.st4.v8i8.p0(<8 x i8> [[TMP2]], <8 x i8> splat (i8 -1), <8 x i8> splat (i8 -1), <8 x i8> [[TMP3]], ptr [[TMP6]])
518; CHECK-NEXT:    [[_MSCMP:%.*]] = icmp ne i64 [[TMP1]], 0
519; CHECK-NEXT:    br i1 [[_MSCMP]], label [[TMP7:%.*]], label [[TMP8:%.*]], !prof [[PROF1]]
520; CHECK:       7:
521; CHECK-NEXT:    call void @__msan_warning_noreturn() #[[ATTR4]]
522; CHECK-NEXT:    unreachable
523; CHECK:       8:
524; CHECK-NEXT:    call void @llvm.aarch64.neon.st4.v8i8.p0(<8 x i8> [[A]], <8 x i8> undef, <8 x i8> undef, <8 x i8> [[D]], ptr [[P]])
525; CHECK-NEXT:    ret void
526;
527  call void @llvm.aarch64.neon.st4.v8i8.p0(<8 x i8> %A, <8 x i8> undef, <8 x i8> undef, <8 x i8> %D, ptr %P)
528  ret void
529}
530
531define void @st4_8b_undefBD(<8 x i8> %A, <8 x i8> %B, <8 x i8> %C, <8 x i8> %D, ptr %P) nounwind sanitize_memory {
532;
533;
534; CHECK-LABEL: define void @st4_8b_undefBD(
535; CHECK-SAME: <8 x i8> [[A:%.*]], <8 x i8> [[B:%.*]], <8 x i8> [[C:%.*]], <8 x i8> [[D:%.*]], ptr [[P:%.*]]) #[[ATTR0]] {
536; CHECK-NEXT:    [[TMP1:%.*]] = load i64, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 32) to ptr), align 8
537; CHECK-NEXT:    [[TMP2:%.*]] = load <8 x i8>, ptr @__msan_param_tls, align 8
538; CHECK-NEXT:    [[TMP3:%.*]] = load <8 x i8>, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 16) to ptr), align 8
539; CHECK-NEXT:    call void @llvm.donothing()
540; CHECK-NEXT:    [[TMP4:%.*]] = ptrtoint ptr [[P]] to i64
541; CHECK-NEXT:    [[TMP5:%.*]] = xor i64 [[TMP4]], 193514046488576
542; CHECK-NEXT:    [[TMP6:%.*]] = inttoptr i64 [[TMP5]] to ptr
543; CHECK-NEXT:    call void @llvm.aarch64.neon.st4.v8i8.p0(<8 x i8> [[TMP2]], <8 x i8> splat (i8 -1), <8 x i8> [[TMP3]], <8 x i8> splat (i8 -1), ptr [[TMP6]])
544; CHECK-NEXT:    [[_MSCMP:%.*]] = icmp ne i64 [[TMP1]], 0
545; CHECK-NEXT:    br i1 [[_MSCMP]], label [[TMP7:%.*]], label [[TMP8:%.*]], !prof [[PROF1]]
546; CHECK:       7:
547; CHECK-NEXT:    call void @__msan_warning_noreturn() #[[ATTR4]]
548; CHECK-NEXT:    unreachable
549; CHECK:       8:
550; CHECK-NEXT:    call void @llvm.aarch64.neon.st4.v8i8.p0(<8 x i8> [[A]], <8 x i8> undef, <8 x i8> [[C]], <8 x i8> undef, ptr [[P]])
551; CHECK-NEXT:    ret void
552;
553  call void @llvm.aarch64.neon.st4.v8i8.p0(<8 x i8> %A, <8 x i8> undef, <8 x i8> %C, <8 x i8> undef, ptr %P)
554  ret void
555}
556
557define void @st4_8b_undefABC(<8 x i8> %A, <8 x i8> %B, <8 x i8> %C, <8 x i8> %D, ptr %P) nounwind sanitize_memory {
558;
559;
560; CHECK-LABEL: define void @st4_8b_undefABC(
561; CHECK-SAME: <8 x i8> [[A:%.*]], <8 x i8> [[B:%.*]], <8 x i8> [[C:%.*]], <8 x i8> [[D:%.*]], ptr [[P:%.*]]) #[[ATTR0]] {
562; CHECK-NEXT:    [[TMP1:%.*]] = load i64, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 32) to ptr), align 8
563; CHECK-NEXT:    [[TMP2:%.*]] = load <8 x i8>, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 24) to ptr), align 8
564; CHECK-NEXT:    call void @llvm.donothing()
565; CHECK-NEXT:    [[TMP3:%.*]] = ptrtoint ptr [[P]] to i64
566; CHECK-NEXT:    [[TMP4:%.*]] = xor i64 [[TMP3]], 193514046488576
567; CHECK-NEXT:    [[TMP5:%.*]] = inttoptr i64 [[TMP4]] to ptr
568; CHECK-NEXT:    call void @llvm.aarch64.neon.st4.v8i8.p0(<8 x i8> splat (i8 -1), <8 x i8> splat (i8 -1), <8 x i8> splat (i8 -1), <8 x i8> [[TMP2]], ptr [[TMP5]])
569; CHECK-NEXT:    [[_MSCMP:%.*]] = icmp ne i64 [[TMP1]], 0
570; CHECK-NEXT:    br i1 [[_MSCMP]], label [[TMP6:%.*]], label [[TMP7:%.*]], !prof [[PROF1]]
571; CHECK:       6:
572; CHECK-NEXT:    call void @__msan_warning_noreturn() #[[ATTR4]]
573; CHECK-NEXT:    unreachable
574; CHECK:       7:
575; CHECK-NEXT:    call void @llvm.aarch64.neon.st4.v8i8.p0(<8 x i8> undef, <8 x i8> undef, <8 x i8> undef, <8 x i8> [[D]], ptr [[P]])
576; CHECK-NEXT:    ret void
577;
578  call void @llvm.aarch64.neon.st4.v8i8.p0(<8 x i8> undef, <8 x i8> undef, <8 x i8> undef, <8 x i8> %D, ptr %P)
579  ret void
580}
581
582define void @st4_8b_undefABD(<8 x i8> %A, <8 x i8> %B, <8 x i8> %C, <8 x i8> %D, ptr %P) nounwind sanitize_memory {
583;
584;
585; CHECK-LABEL: define void @st4_8b_undefABD(
586; CHECK-SAME: <8 x i8> [[A:%.*]], <8 x i8> [[B:%.*]], <8 x i8> [[C:%.*]], <8 x i8> [[D:%.*]], ptr [[P:%.*]]) #[[ATTR0]] {
587; CHECK-NEXT:    [[TMP1:%.*]] = load i64, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 32) to ptr), align 8
588; CHECK-NEXT:    [[TMP2:%.*]] = load <8 x i8>, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 16) to ptr), align 8
589; CHECK-NEXT:    call void @llvm.donothing()
590; CHECK-NEXT:    [[TMP3:%.*]] = ptrtoint ptr [[P]] to i64
591; CHECK-NEXT:    [[TMP4:%.*]] = xor i64 [[TMP3]], 193514046488576
592; CHECK-NEXT:    [[TMP5:%.*]] = inttoptr i64 [[TMP4]] to ptr
593; CHECK-NEXT:    call void @llvm.aarch64.neon.st4.v8i8.p0(<8 x i8> splat (i8 -1), <8 x i8> splat (i8 -1), <8 x i8> [[TMP2]], <8 x i8> splat (i8 -1), ptr [[TMP5]])
594; CHECK-NEXT:    [[_MSCMP:%.*]] = icmp ne i64 [[TMP1]], 0
595; CHECK-NEXT:    br i1 [[_MSCMP]], label [[TMP6:%.*]], label [[TMP7:%.*]], !prof [[PROF1]]
596; CHECK:       6:
597; CHECK-NEXT:    call void @__msan_warning_noreturn() #[[ATTR4]]
598; CHECK-NEXT:    unreachable
599; CHECK:       7:
600; CHECK-NEXT:    call void @llvm.aarch64.neon.st4.v8i8.p0(<8 x i8> undef, <8 x i8> undef, <8 x i8> [[C]], <8 x i8> undef, ptr [[P]])
601; CHECK-NEXT:    ret void
602;
603  call void @llvm.aarch64.neon.st4.v8i8.p0(<8 x i8> undef, <8 x i8> undef, <8 x i8> %C, <8 x i8> undef, ptr %P)
604  ret void
605}
606
607define void @st4_8b_undefACD(<8 x i8> %A, <8 x i8> %B, <8 x i8> %C, <8 x i8> %D, ptr %P) nounwind sanitize_memory {
608;
609;
610; CHECK-LABEL: define void @st4_8b_undefACD(
611; CHECK-SAME: <8 x i8> [[A:%.*]], <8 x i8> [[B:%.*]], <8 x i8> [[C:%.*]], <8 x i8> [[D:%.*]], ptr [[P:%.*]]) #[[ATTR0]] {
612; CHECK-NEXT:    [[TMP1:%.*]] = load i64, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 32) to ptr), align 8
613; CHECK-NEXT:    [[TMP2:%.*]] = load <8 x i8>, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 8) to ptr), align 8
614; CHECK-NEXT:    call void @llvm.donothing()
615; CHECK-NEXT:    [[TMP3:%.*]] = ptrtoint ptr [[P]] to i64
616; CHECK-NEXT:    [[TMP4:%.*]] = xor i64 [[TMP3]], 193514046488576
617; CHECK-NEXT:    [[TMP5:%.*]] = inttoptr i64 [[TMP4]] to ptr
618; CHECK-NEXT:    call void @llvm.aarch64.neon.st4.v8i8.p0(<8 x i8> splat (i8 -1), <8 x i8> [[TMP2]], <8 x i8> splat (i8 -1), <8 x i8> splat (i8 -1), ptr [[TMP5]])
619; CHECK-NEXT:    [[_MSCMP:%.*]] = icmp ne i64 [[TMP1]], 0
620; CHECK-NEXT:    br i1 [[_MSCMP]], label [[TMP6:%.*]], label [[TMP7:%.*]], !prof [[PROF1]]
621; CHECK:       6:
622; CHECK-NEXT:    call void @__msan_warning_noreturn() #[[ATTR4]]
623; CHECK-NEXT:    unreachable
624; CHECK:       7:
625; CHECK-NEXT:    call void @llvm.aarch64.neon.st4.v8i8.p0(<8 x i8> undef, <8 x i8> [[B]], <8 x i8> undef, <8 x i8> undef, ptr [[P]])
626; CHECK-NEXT:    ret void
627;
628  call void @llvm.aarch64.neon.st4.v8i8.p0(<8 x i8> undef, <8 x i8> %B, <8 x i8> undef, <8 x i8> undef, ptr %P)
629  ret void
630}
631
632define void @st4_8b_undefBCD(<8 x i8> %A, <8 x i8> %B, <8 x i8> %C, <8 x i8> %D, ptr %P) nounwind sanitize_memory {
633;
634;
635; CHECK-LABEL: define void @st4_8b_undefBCD(
636; CHECK-SAME: <8 x i8> [[A:%.*]], <8 x i8> [[B:%.*]], <8 x i8> [[C:%.*]], <8 x i8> [[D:%.*]], ptr [[P:%.*]]) #[[ATTR0]] {
637; CHECK-NEXT:    [[TMP1:%.*]] = load i64, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 32) to ptr), align 8
638; CHECK-NEXT:    [[TMP2:%.*]] = load <8 x i8>, ptr @__msan_param_tls, align 8
639; CHECK-NEXT:    call void @llvm.donothing()
640; CHECK-NEXT:    [[TMP3:%.*]] = ptrtoint ptr [[P]] to i64
641; CHECK-NEXT:    [[TMP4:%.*]] = xor i64 [[TMP3]], 193514046488576
642; CHECK-NEXT:    [[TMP5:%.*]] = inttoptr i64 [[TMP4]] to ptr
643; CHECK-NEXT:    call void @llvm.aarch64.neon.st4.v8i8.p0(<8 x i8> [[TMP2]], <8 x i8> splat (i8 -1), <8 x i8> splat (i8 -1), <8 x i8> splat (i8 -1), ptr [[TMP5]])
644; CHECK-NEXT:    [[_MSCMP:%.*]] = icmp ne i64 [[TMP1]], 0
645; CHECK-NEXT:    br i1 [[_MSCMP]], label [[TMP6:%.*]], label [[TMP7:%.*]], !prof [[PROF1]]
646; CHECK:       6:
647; CHECK-NEXT:    call void @__msan_warning_noreturn() #[[ATTR4]]
648; CHECK-NEXT:    unreachable
649; CHECK:       7:
650; CHECK-NEXT:    call void @llvm.aarch64.neon.st4.v8i8.p0(<8 x i8> [[A]], <8 x i8> undef, <8 x i8> undef, <8 x i8> undef, ptr [[P]])
651; CHECK-NEXT:    ret void
652;
653  call void @llvm.aarch64.neon.st4.v8i8.p0(<8 x i8> %A, <8 x i8> undef, <8 x i8> undef, <8 x i8> undef, ptr %P)
654  ret void
655}
656
657define void @st4_8b_undefABCD(<8 x i8> %A, <8 x i8> %B, <8 x i8> %C, <8 x i8> %D, ptr %P) nounwind sanitize_memory {
658;
659;
660; CHECK-LABEL: define void @st4_8b_undefABCD(
661; CHECK-SAME: <8 x i8> [[A:%.*]], <8 x i8> [[B:%.*]], <8 x i8> [[C:%.*]], <8 x i8> [[D:%.*]], ptr [[P:%.*]]) #[[ATTR0]] {
662; CHECK-NEXT:    [[TMP1:%.*]] = load i64, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 32) to ptr), align 8
663; CHECK-NEXT:    call void @llvm.donothing()
664; CHECK-NEXT:    [[TMP2:%.*]] = ptrtoint ptr [[P]] to i64
665; CHECK-NEXT:    [[TMP3:%.*]] = xor i64 [[TMP2]], 193514046488576
666; CHECK-NEXT:    [[TMP4:%.*]] = inttoptr i64 [[TMP3]] to ptr
667; CHECK-NEXT:    call void @llvm.aarch64.neon.st4.v8i8.p0(<8 x i8> splat (i8 -1), <8 x i8> splat (i8 -1), <8 x i8> splat (i8 -1), <8 x i8> splat (i8 -1), ptr [[TMP4]])
668; CHECK-NEXT:    [[_MSCMP:%.*]] = icmp ne i64 [[TMP1]], 0
669; CHECK-NEXT:    br i1 [[_MSCMP]], label [[TMP5:%.*]], label [[TMP6:%.*]], !prof [[PROF1]]
670; CHECK:       5:
671; CHECK-NEXT:    call void @__msan_warning_noreturn() #[[ATTR4]]
672; CHECK-NEXT:    unreachable
673; CHECK:       6:
674; CHECK-NEXT:    call void @llvm.aarch64.neon.st4.v8i8.p0(<8 x i8> undef, <8 x i8> undef, <8 x i8> undef, <8 x i8> undef, ptr [[P]])
675; CHECK-NEXT:    ret void
676;
677  call void @llvm.aarch64.neon.st4.v8i8.p0(<8 x i8> undef, <8 x i8> undef, <8 x i8> undef, <8 x i8> undef, ptr %P)
678  ret void
679}
680
681; -----------------------------------------------------------------------------------------------------------------------------------------------
682
683declare void @llvm.aarch64.neon.st2.v8i8.p0(<8 x i8>, <8 x i8>, ptr) nounwind sanitize_memory readonly
684declare void @llvm.aarch64.neon.st3.v8i8.p0(<8 x i8>, <8 x i8>, <8 x i8>, ptr) nounwind sanitize_memory readonly
685declare void @llvm.aarch64.neon.st4.v8i8.p0(<8 x i8>, <8 x i8>, <8 x i8>, <8 x i8>, ptr) nounwind sanitize_memory readonly
686
687define void @st2_16b(<16 x i8> %A, <16 x i8> %B, ptr %P) nounwind sanitize_memory {
688;
689;
690; CHECK-LABEL: define void @st2_16b(
691; CHECK-SAME: <16 x i8> [[A:%.*]], <16 x i8> [[B:%.*]], ptr [[P:%.*]]) #[[ATTR0]] {
692; CHECK-NEXT:    [[TMP1:%.*]] = load i64, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 32) to ptr), align 8
693; CHECK-NEXT:    [[TMP2:%.*]] = load <16 x i8>, ptr @__msan_param_tls, align 8
694; CHECK-NEXT:    [[TMP3:%.*]] = load <16 x i8>, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 16) to ptr), align 8
695; CHECK-NEXT:    call void @llvm.donothing()
696; CHECK-NEXT:    [[TMP4:%.*]] = ptrtoint ptr [[P]] to i64
697; CHECK-NEXT:    [[TMP5:%.*]] = xor i64 [[TMP4]], 193514046488576
698; CHECK-NEXT:    [[TMP6:%.*]] = inttoptr i64 [[TMP5]] to ptr
699; CHECK-NEXT:    call void @llvm.aarch64.neon.st2.v16i8.p0(<16 x i8> [[TMP2]], <16 x i8> [[TMP3]], ptr [[TMP6]])
700; CHECK-NEXT:    [[_MSCMP:%.*]] = icmp ne i64 [[TMP1]], 0
701; CHECK-NEXT:    br i1 [[_MSCMP]], label [[TMP7:%.*]], label [[TMP8:%.*]], !prof [[PROF1]]
702; CHECK:       7:
703; CHECK-NEXT:    call void @__msan_warning_noreturn() #[[ATTR4]]
704; CHECK-NEXT:    unreachable
705; CHECK:       8:
706; CHECK-NEXT:    call void @llvm.aarch64.neon.st2.v16i8.p0(<16 x i8> [[A]], <16 x i8> [[B]], ptr [[P]])
707; CHECK-NEXT:    ret void
708;
709  call void @llvm.aarch64.neon.st2.v16i8.p0(<16 x i8> %A, <16 x i8> %B, ptr %P)
710  ret void
711}
712
713define void @st3_16b(<16 x i8> %A, <16 x i8> %B, <16 x i8> %C, ptr %P) nounwind sanitize_memory {
714;
715;
716; CHECK-LABEL: define void @st3_16b(
717; CHECK-SAME: <16 x i8> [[A:%.*]], <16 x i8> [[B:%.*]], <16 x i8> [[C:%.*]], ptr [[P:%.*]]) #[[ATTR0]] {
718; CHECK-NEXT:    [[TMP1:%.*]] = load i64, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 48) to ptr), align 8
719; CHECK-NEXT:    [[TMP2:%.*]] = load <16 x i8>, ptr @__msan_param_tls, align 8
720; CHECK-NEXT:    [[TMP3:%.*]] = load <16 x i8>, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 16) to ptr), align 8
721; CHECK-NEXT:    [[TMP4:%.*]] = load <16 x i8>, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 32) to ptr), align 8
722; CHECK-NEXT:    call void @llvm.donothing()
723; CHECK-NEXT:    [[TMP5:%.*]] = ptrtoint ptr [[P]] to i64
724; CHECK-NEXT:    [[TMP6:%.*]] = xor i64 [[TMP5]], 193514046488576
725; CHECK-NEXT:    [[TMP7:%.*]] = inttoptr i64 [[TMP6]] to ptr
726; CHECK-NEXT:    call void @llvm.aarch64.neon.st3.v16i8.p0(<16 x i8> [[TMP2]], <16 x i8> [[TMP3]], <16 x i8> [[TMP4]], ptr [[TMP7]])
727; CHECK-NEXT:    [[_MSCMP:%.*]] = icmp ne i64 [[TMP1]], 0
728; CHECK-NEXT:    br i1 [[_MSCMP]], label [[TMP8:%.*]], label [[TMP9:%.*]], !prof [[PROF1]]
729; CHECK:       8:
730; CHECK-NEXT:    call void @__msan_warning_noreturn() #[[ATTR4]]
731; CHECK-NEXT:    unreachable
732; CHECK:       9:
733; CHECK-NEXT:    call void @llvm.aarch64.neon.st3.v16i8.p0(<16 x i8> [[A]], <16 x i8> [[B]], <16 x i8> [[C]], ptr [[P]])
734; CHECK-NEXT:    ret void
735;
736  call void @llvm.aarch64.neon.st3.v16i8.p0(<16 x i8> %A, <16 x i8> %B, <16 x i8> %C, ptr %P)
737  ret void
738}
739
740define void @st4_16b(<16 x i8> %A, <16 x i8> %B, <16 x i8> %C, <16 x i8> %D, ptr %P) nounwind sanitize_memory {
741;
742;
743; CHECK-LABEL: define void @st4_16b(
744; CHECK-SAME: <16 x i8> [[A:%.*]], <16 x i8> [[B:%.*]], <16 x i8> [[C:%.*]], <16 x i8> [[D:%.*]], ptr [[P:%.*]]) #[[ATTR0]] {
745; CHECK-NEXT:    [[TMP1:%.*]] = load i64, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 64) to ptr), align 8
746; CHECK-NEXT:    [[TMP2:%.*]] = load <16 x i8>, ptr @__msan_param_tls, align 8
747; CHECK-NEXT:    [[TMP3:%.*]] = load <16 x i8>, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 16) to ptr), align 8
748; CHECK-NEXT:    [[TMP4:%.*]] = load <16 x i8>, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 32) to ptr), align 8
749; CHECK-NEXT:    [[TMP5:%.*]] = load <16 x i8>, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 48) to ptr), align 8
750; CHECK-NEXT:    call void @llvm.donothing()
751; CHECK-NEXT:    [[TMP6:%.*]] = ptrtoint ptr [[P]] to i64
752; CHECK-NEXT:    [[TMP7:%.*]] = xor i64 [[TMP6]], 193514046488576
753; CHECK-NEXT:    [[TMP8:%.*]] = inttoptr i64 [[TMP7]] to ptr
754; CHECK-NEXT:    call void @llvm.aarch64.neon.st4.v16i8.p0(<16 x i8> [[TMP2]], <16 x i8> [[TMP3]], <16 x i8> [[TMP4]], <16 x i8> [[TMP5]], ptr [[TMP8]])
755; CHECK-NEXT:    [[_MSCMP:%.*]] = icmp ne i64 [[TMP1]], 0
756; CHECK-NEXT:    br i1 [[_MSCMP]], label [[TMP9:%.*]], label [[TMP10:%.*]], !prof [[PROF1]]
757; CHECK:       9:
758; CHECK-NEXT:    call void @__msan_warning_noreturn() #[[ATTR4]]
759; CHECK-NEXT:    unreachable
760; CHECK:       10:
761; CHECK-NEXT:    call void @llvm.aarch64.neon.st4.v16i8.p0(<16 x i8> [[A]], <16 x i8> [[B]], <16 x i8> [[C]], <16 x i8> [[D]], ptr [[P]])
762; CHECK-NEXT:    ret void
763;
764  call void @llvm.aarch64.neon.st4.v16i8.p0(<16 x i8> %A, <16 x i8> %B, <16 x i8> %C, <16 x i8> %D, ptr %P)
765  ret void
766}
767
768; -----------------------------------------------------------------------------------------------------------------------------------------------
769
770declare void @llvm.aarch64.neon.st2.v16i8.p0(<16 x i8>, <16 x i8>, ptr) nounwind sanitize_memory readonly
771declare void @llvm.aarch64.neon.st3.v16i8.p0(<16 x i8>, <16 x i8>, <16 x i8>, ptr) nounwind sanitize_memory readonly
772declare void @llvm.aarch64.neon.st4.v16i8.p0(<16 x i8>, <16 x i8>, <16 x i8>, <16 x i8>, ptr) nounwind sanitize_memory readonly
773
774define void @st2_4h(<4 x i16> %A, <4 x i16> %B, ptr %P) nounwind sanitize_memory {
775;
776;
777; CHECK-LABEL: define void @st2_4h(
778; CHECK-SAME: <4 x i16> [[A:%.*]], <4 x i16> [[B:%.*]], ptr [[P:%.*]]) #[[ATTR0]] {
779; CHECK-NEXT:    [[TMP1:%.*]] = load i64, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 16) to ptr), align 8
780; CHECK-NEXT:    [[TMP2:%.*]] = load <4 x i16>, ptr @__msan_param_tls, align 8
781; CHECK-NEXT:    [[TMP3:%.*]] = load <4 x i16>, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 8) to ptr), align 8
782; CHECK-NEXT:    call void @llvm.donothing()
783; CHECK-NEXT:    [[TMP4:%.*]] = ptrtoint ptr [[P]] to i64
784; CHECK-NEXT:    [[TMP5:%.*]] = xor i64 [[TMP4]], 193514046488576
785; CHECK-NEXT:    [[TMP6:%.*]] = inttoptr i64 [[TMP5]] to ptr
786; CHECK-NEXT:    call void @llvm.aarch64.neon.st2.v4i16.p0(<4 x i16> [[TMP2]], <4 x i16> [[TMP3]], ptr [[TMP6]])
787; CHECK-NEXT:    [[_MSCMP:%.*]] = icmp ne i64 [[TMP1]], 0
788; CHECK-NEXT:    br i1 [[_MSCMP]], label [[TMP7:%.*]], label [[TMP8:%.*]], !prof [[PROF1]]
789; CHECK:       7:
790; CHECK-NEXT:    call void @__msan_warning_noreturn() #[[ATTR4]]
791; CHECK-NEXT:    unreachable
792; CHECK:       8:
793; CHECK-NEXT:    call void @llvm.aarch64.neon.st2.v4i16.p0(<4 x i16> [[A]], <4 x i16> [[B]], ptr [[P]])
794; CHECK-NEXT:    ret void
795;
796  call void @llvm.aarch64.neon.st2.v4i16.p0(<4 x i16> %A, <4 x i16> %B, ptr %P)
797  ret void
798}
799
800define void @st3_4h(<4 x i16> %A, <4 x i16> %B, <4 x i16> %C, ptr %P) nounwind sanitize_memory {
801;
802;
803; CHECK-LABEL: define void @st3_4h(
804; CHECK-SAME: <4 x i16> [[A:%.*]], <4 x i16> [[B:%.*]], <4 x i16> [[C:%.*]], ptr [[P:%.*]]) #[[ATTR0]] {
805; CHECK-NEXT:    [[TMP1:%.*]] = load i64, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 24) to ptr), align 8
806; CHECK-NEXT:    [[TMP2:%.*]] = load <4 x i16>, ptr @__msan_param_tls, align 8
807; CHECK-NEXT:    [[TMP3:%.*]] = load <4 x i16>, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 8) to ptr), align 8
808; CHECK-NEXT:    [[TMP4:%.*]] = load <4 x i16>, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 16) to ptr), align 8
809; CHECK-NEXT:    call void @llvm.donothing()
810; CHECK-NEXT:    [[TMP5:%.*]] = ptrtoint ptr [[P]] to i64
811; CHECK-NEXT:    [[TMP6:%.*]] = xor i64 [[TMP5]], 193514046488576
812; CHECK-NEXT:    [[TMP7:%.*]] = inttoptr i64 [[TMP6]] to ptr
813; CHECK-NEXT:    call void @llvm.aarch64.neon.st3.v4i16.p0(<4 x i16> [[TMP2]], <4 x i16> [[TMP3]], <4 x i16> [[TMP4]], ptr [[TMP7]])
814; CHECK-NEXT:    [[_MSCMP:%.*]] = icmp ne i64 [[TMP1]], 0
815; CHECK-NEXT:    br i1 [[_MSCMP]], label [[TMP8:%.*]], label [[TMP9:%.*]], !prof [[PROF1]]
816; CHECK:       8:
817; CHECK-NEXT:    call void @__msan_warning_noreturn() #[[ATTR4]]
818; CHECK-NEXT:    unreachable
819; CHECK:       9:
820; CHECK-NEXT:    call void @llvm.aarch64.neon.st3.v4i16.p0(<4 x i16> [[A]], <4 x i16> [[B]], <4 x i16> [[C]], ptr [[P]])
821; CHECK-NEXT:    ret void
822;
823  call void @llvm.aarch64.neon.st3.v4i16.p0(<4 x i16> %A, <4 x i16> %B, <4 x i16> %C, ptr %P)
824  ret void
825}
826
827define void @st4_4h(<4 x i16> %A, <4 x i16> %B, <4 x i16> %C, <4 x i16> %D, ptr %P) nounwind sanitize_memory {
828;
829;
830; CHECK-LABEL: define void @st4_4h(
831; CHECK-SAME: <4 x i16> [[A:%.*]], <4 x i16> [[B:%.*]], <4 x i16> [[C:%.*]], <4 x i16> [[D:%.*]], ptr [[P:%.*]]) #[[ATTR0]] {
832; CHECK-NEXT:    [[TMP1:%.*]] = load i64, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 32) to ptr), align 8
833; CHECK-NEXT:    [[TMP2:%.*]] = load <4 x i16>, ptr @__msan_param_tls, align 8
834; CHECK-NEXT:    [[TMP3:%.*]] = load <4 x i16>, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 8) to ptr), align 8
835; CHECK-NEXT:    [[TMP4:%.*]] = load <4 x i16>, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 16) to ptr), align 8
836; CHECK-NEXT:    [[TMP5:%.*]] = load <4 x i16>, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 24) to ptr), align 8
837; CHECK-NEXT:    call void @llvm.donothing()
838; CHECK-NEXT:    [[TMP6:%.*]] = ptrtoint ptr [[P]] to i64
839; CHECK-NEXT:    [[TMP7:%.*]] = xor i64 [[TMP6]], 193514046488576
840; CHECK-NEXT:    [[TMP8:%.*]] = inttoptr i64 [[TMP7]] to ptr
841; CHECK-NEXT:    call void @llvm.aarch64.neon.st4.v4i16.p0(<4 x i16> [[TMP2]], <4 x i16> [[TMP3]], <4 x i16> [[TMP4]], <4 x i16> [[TMP5]], ptr [[TMP8]])
842; CHECK-NEXT:    [[_MSCMP:%.*]] = icmp ne i64 [[TMP1]], 0
843; CHECK-NEXT:    br i1 [[_MSCMP]], label [[TMP9:%.*]], label [[TMP10:%.*]], !prof [[PROF1]]
844; CHECK:       9:
845; CHECK-NEXT:    call void @__msan_warning_noreturn() #[[ATTR4]]
846; CHECK-NEXT:    unreachable
847; CHECK:       10:
848; CHECK-NEXT:    call void @llvm.aarch64.neon.st4.v4i16.p0(<4 x i16> [[A]], <4 x i16> [[B]], <4 x i16> [[C]], <4 x i16> [[D]], ptr [[P]])
849; CHECK-NEXT:    ret void
850;
851  call void @llvm.aarch64.neon.st4.v4i16.p0(<4 x i16> %A, <4 x i16> %B, <4 x i16> %C, <4 x i16> %D, ptr %P)
852  ret void
853}
854
855; -----------------------------------------------------------------------------------------------------------------------------------------------
856
857declare void @llvm.aarch64.neon.st2.v4i16.p0(<4 x i16>, <4 x i16>, ptr) nounwind sanitize_memory readonly
858declare void @llvm.aarch64.neon.st3.v4i16.p0(<4 x i16>, <4 x i16>, <4 x i16>, ptr) nounwind sanitize_memory readonly
859declare void @llvm.aarch64.neon.st4.v4i16.p0(<4 x i16>, <4 x i16>, <4 x i16>, <4 x i16>, ptr) nounwind sanitize_memory readonly
860
861define void @st2_8h(<8 x i16> %A, <8 x i16> %B, ptr %P) nounwind sanitize_memory {
862;
863;
864; CHECK-LABEL: define void @st2_8h(
865; CHECK-SAME: <8 x i16> [[A:%.*]], <8 x i16> [[B:%.*]], ptr [[P:%.*]]) #[[ATTR0]] {
866; CHECK-NEXT:    [[TMP1:%.*]] = load i64, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 32) to ptr), align 8
867; CHECK-NEXT:    [[TMP2:%.*]] = load <8 x i16>, ptr @__msan_param_tls, align 8
868; CHECK-NEXT:    [[TMP3:%.*]] = load <8 x i16>, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 16) to ptr), align 8
869; CHECK-NEXT:    call void @llvm.donothing()
870; CHECK-NEXT:    [[TMP4:%.*]] = ptrtoint ptr [[P]] to i64
871; CHECK-NEXT:    [[TMP5:%.*]] = xor i64 [[TMP4]], 193514046488576
872; CHECK-NEXT:    [[TMP6:%.*]] = inttoptr i64 [[TMP5]] to ptr
873; CHECK-NEXT:    call void @llvm.aarch64.neon.st2.v8i16.p0(<8 x i16> [[TMP2]], <8 x i16> [[TMP3]], ptr [[TMP6]])
874; CHECK-NEXT:    [[_MSCMP:%.*]] = icmp ne i64 [[TMP1]], 0
875; CHECK-NEXT:    br i1 [[_MSCMP]], label [[TMP7:%.*]], label [[TMP8:%.*]], !prof [[PROF1]]
876; CHECK:       7:
877; CHECK-NEXT:    call void @__msan_warning_noreturn() #[[ATTR4]]
878; CHECK-NEXT:    unreachable
879; CHECK:       8:
880; CHECK-NEXT:    call void @llvm.aarch64.neon.st2.v8i16.p0(<8 x i16> [[A]], <8 x i16> [[B]], ptr [[P]])
881; CHECK-NEXT:    ret void
882;
883  call void @llvm.aarch64.neon.st2.v8i16.p0(<8 x i16> %A, <8 x i16> %B, ptr %P)
884  ret void
885}
886
887define void @st3_8h(<8 x i16> %A, <8 x i16> %B, <8 x i16> %C, ptr %P) nounwind sanitize_memory {
888;
889;
890; CHECK-LABEL: define void @st3_8h(
891; CHECK-SAME: <8 x i16> [[A:%.*]], <8 x i16> [[B:%.*]], <8 x i16> [[C:%.*]], ptr [[P:%.*]]) #[[ATTR0]] {
892; CHECK-NEXT:    [[TMP1:%.*]] = load i64, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 48) to ptr), align 8
893; CHECK-NEXT:    [[TMP2:%.*]] = load <8 x i16>, ptr @__msan_param_tls, align 8
894; CHECK-NEXT:    [[TMP3:%.*]] = load <8 x i16>, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 16) to ptr), align 8
895; CHECK-NEXT:    [[TMP4:%.*]] = load <8 x i16>, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 32) to ptr), align 8
896; CHECK-NEXT:    call void @llvm.donothing()
897; CHECK-NEXT:    [[TMP5:%.*]] = ptrtoint ptr [[P]] to i64
898; CHECK-NEXT:    [[TMP6:%.*]] = xor i64 [[TMP5]], 193514046488576
899; CHECK-NEXT:    [[TMP7:%.*]] = inttoptr i64 [[TMP6]] to ptr
900; CHECK-NEXT:    call void @llvm.aarch64.neon.st3.v8i16.p0(<8 x i16> [[TMP2]], <8 x i16> [[TMP3]], <8 x i16> [[TMP4]], ptr [[TMP7]])
901; CHECK-NEXT:    [[_MSCMP:%.*]] = icmp ne i64 [[TMP1]], 0
902; CHECK-NEXT:    br i1 [[_MSCMP]], label [[TMP8:%.*]], label [[TMP9:%.*]], !prof [[PROF1]]
903; CHECK:       8:
904; CHECK-NEXT:    call void @__msan_warning_noreturn() #[[ATTR4]]
905; CHECK-NEXT:    unreachable
906; CHECK:       9:
907; CHECK-NEXT:    call void @llvm.aarch64.neon.st3.v8i16.p0(<8 x i16> [[A]], <8 x i16> [[B]], <8 x i16> [[C]], ptr [[P]])
908; CHECK-NEXT:    ret void
909;
910  call void @llvm.aarch64.neon.st3.v8i16.p0(<8 x i16> %A, <8 x i16> %B, <8 x i16> %C, ptr %P)
911  ret void
912}
913
914define void @st4_8h(<8 x i16> %A, <8 x i16> %B, <8 x i16> %C, <8 x i16> %D, ptr %P) nounwind sanitize_memory {
915;
916;
917; CHECK-LABEL: define void @st4_8h(
918; CHECK-SAME: <8 x i16> [[A:%.*]], <8 x i16> [[B:%.*]], <8 x i16> [[C:%.*]], <8 x i16> [[D:%.*]], ptr [[P:%.*]]) #[[ATTR0]] {
919; CHECK-NEXT:    [[TMP1:%.*]] = load i64, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 64) to ptr), align 8
920; CHECK-NEXT:    [[TMP2:%.*]] = load <8 x i16>, ptr @__msan_param_tls, align 8
921; CHECK-NEXT:    [[TMP3:%.*]] = load <8 x i16>, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 16) to ptr), align 8
922; CHECK-NEXT:    [[TMP4:%.*]] = load <8 x i16>, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 32) to ptr), align 8
923; CHECK-NEXT:    [[TMP5:%.*]] = load <8 x i16>, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 48) to ptr), align 8
924; CHECK-NEXT:    call void @llvm.donothing()
925; CHECK-NEXT:    [[TMP6:%.*]] = ptrtoint ptr [[P]] to i64
926; CHECK-NEXT:    [[TMP7:%.*]] = xor i64 [[TMP6]], 193514046488576
927; CHECK-NEXT:    [[TMP8:%.*]] = inttoptr i64 [[TMP7]] to ptr
928; CHECK-NEXT:    call void @llvm.aarch64.neon.st4.v8i16.p0(<8 x i16> [[TMP2]], <8 x i16> [[TMP3]], <8 x i16> [[TMP4]], <8 x i16> [[TMP5]], ptr [[TMP8]])
929; CHECK-NEXT:    [[_MSCMP:%.*]] = icmp ne i64 [[TMP1]], 0
930; CHECK-NEXT:    br i1 [[_MSCMP]], label [[TMP9:%.*]], label [[TMP10:%.*]], !prof [[PROF1]]
931; CHECK:       9:
932; CHECK-NEXT:    call void @__msan_warning_noreturn() #[[ATTR4]]
933; CHECK-NEXT:    unreachable
934; CHECK:       10:
935; CHECK-NEXT:    call void @llvm.aarch64.neon.st4.v8i16.p0(<8 x i16> [[A]], <8 x i16> [[B]], <8 x i16> [[C]], <8 x i16> [[D]], ptr [[P]])
936; CHECK-NEXT:    ret void
937;
938  call void @llvm.aarch64.neon.st4.v8i16.p0(<8 x i16> %A, <8 x i16> %B, <8 x i16> %C, <8 x i16> %D, ptr %P)
939  ret void
940}
941
942; -----------------------------------------------------------------------------------------------------------------------------------------------
943
944declare void @llvm.aarch64.neon.st2.v8i16.p0(<8 x i16>, <8 x i16>, ptr) nounwind sanitize_memory readonly
945declare void @llvm.aarch64.neon.st3.v8i16.p0(<8 x i16>, <8 x i16>, <8 x i16>, ptr) nounwind sanitize_memory readonly
946declare void @llvm.aarch64.neon.st4.v8i16.p0(<8 x i16>, <8 x i16>, <8 x i16>, <8 x i16>, ptr) nounwind sanitize_memory readonly
947
948define void @st2_2s(<2 x i32> %A, <2 x i32> %B, ptr %P) nounwind sanitize_memory {
949;
950;
951; CHECK-LABEL: define void @st2_2s(
952; CHECK-SAME: <2 x i32> [[A:%.*]], <2 x i32> [[B:%.*]], ptr [[P:%.*]]) #[[ATTR0]] {
953; CHECK-NEXT:    [[TMP1:%.*]] = load i64, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 16) to ptr), align 8
954; CHECK-NEXT:    [[TMP2:%.*]] = load <2 x i32>, ptr @__msan_param_tls, align 8
955; CHECK-NEXT:    [[TMP3:%.*]] = load <2 x i32>, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 8) to ptr), align 8
956; CHECK-NEXT:    call void @llvm.donothing()
957; CHECK-NEXT:    [[TMP4:%.*]] = ptrtoint ptr [[P]] to i64
958; CHECK-NEXT:    [[TMP5:%.*]] = xor i64 [[TMP4]], 193514046488576
959; CHECK-NEXT:    [[TMP6:%.*]] = inttoptr i64 [[TMP5]] to ptr
960; CHECK-NEXT:    call void @llvm.aarch64.neon.st2.v2i32.p0(<2 x i32> [[TMP2]], <2 x i32> [[TMP3]], ptr [[TMP6]])
961; CHECK-NEXT:    [[_MSCMP:%.*]] = icmp ne i64 [[TMP1]], 0
962; CHECK-NEXT:    br i1 [[_MSCMP]], label [[TMP7:%.*]], label [[TMP8:%.*]], !prof [[PROF1]]
963; CHECK:       7:
964; CHECK-NEXT:    call void @__msan_warning_noreturn() #[[ATTR4]]
965; CHECK-NEXT:    unreachable
966; CHECK:       8:
967; CHECK-NEXT:    call void @llvm.aarch64.neon.st2.v2i32.p0(<2 x i32> [[A]], <2 x i32> [[B]], ptr [[P]])
968; CHECK-NEXT:    ret void
969;
970  call void @llvm.aarch64.neon.st2.v2i32.p0(<2 x i32> %A, <2 x i32> %B, ptr %P)
971  ret void
972}
973
974define void @st3_2s(<2 x i32> %A, <2 x i32> %B, <2 x i32> %C, ptr %P) nounwind sanitize_memory {
975;
976;
977; CHECK-LABEL: define void @st3_2s(
978; CHECK-SAME: <2 x i32> [[A:%.*]], <2 x i32> [[B:%.*]], <2 x i32> [[C:%.*]], ptr [[P:%.*]]) #[[ATTR0]] {
979; CHECK-NEXT:    [[TMP1:%.*]] = load i64, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 24) to ptr), align 8
980; CHECK-NEXT:    [[TMP2:%.*]] = load <2 x i32>, ptr @__msan_param_tls, align 8
981; CHECK-NEXT:    [[TMP3:%.*]] = load <2 x i32>, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 8) to ptr), align 8
982; CHECK-NEXT:    [[TMP4:%.*]] = load <2 x i32>, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 16) to ptr), align 8
983; CHECK-NEXT:    call void @llvm.donothing()
984; CHECK-NEXT:    [[TMP5:%.*]] = ptrtoint ptr [[P]] to i64
985; CHECK-NEXT:    [[TMP6:%.*]] = xor i64 [[TMP5]], 193514046488576
986; CHECK-NEXT:    [[TMP7:%.*]] = inttoptr i64 [[TMP6]] to ptr
987; CHECK-NEXT:    call void @llvm.aarch64.neon.st3.v2i32.p0(<2 x i32> [[TMP2]], <2 x i32> [[TMP3]], <2 x i32> [[TMP4]], ptr [[TMP7]])
988; CHECK-NEXT:    [[_MSCMP:%.*]] = icmp ne i64 [[TMP1]], 0
989; CHECK-NEXT:    br i1 [[_MSCMP]], label [[TMP8:%.*]], label [[TMP9:%.*]], !prof [[PROF1]]
990; CHECK:       8:
991; CHECK-NEXT:    call void @__msan_warning_noreturn() #[[ATTR4]]
992; CHECK-NEXT:    unreachable
993; CHECK:       9:
994; CHECK-NEXT:    call void @llvm.aarch64.neon.st3.v2i32.p0(<2 x i32> [[A]], <2 x i32> [[B]], <2 x i32> [[C]], ptr [[P]])
995; CHECK-NEXT:    ret void
996;
997  call void @llvm.aarch64.neon.st3.v2i32.p0(<2 x i32> %A, <2 x i32> %B, <2 x i32> %C, ptr %P)
998  ret void
999}
1000
1001define void @st4_2s(<2 x i32> %A, <2 x i32> %B, <2 x i32> %C, <2 x i32> %D, ptr %P) nounwind sanitize_memory {
1002;
1003;
1004; CHECK-LABEL: define void @st4_2s(
1005; CHECK-SAME: <2 x i32> [[A:%.*]], <2 x i32> [[B:%.*]], <2 x i32> [[C:%.*]], <2 x i32> [[D:%.*]], ptr [[P:%.*]]) #[[ATTR0]] {
1006; CHECK-NEXT:    [[TMP1:%.*]] = load i64, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 32) to ptr), align 8
1007; CHECK-NEXT:    [[TMP2:%.*]] = load <2 x i32>, ptr @__msan_param_tls, align 8
1008; CHECK-NEXT:    [[TMP3:%.*]] = load <2 x i32>, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 8) to ptr), align 8
1009; CHECK-NEXT:    [[TMP4:%.*]] = load <2 x i32>, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 16) to ptr), align 8
1010; CHECK-NEXT:    [[TMP5:%.*]] = load <2 x i32>, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 24) to ptr), align 8
1011; CHECK-NEXT:    call void @llvm.donothing()
1012; CHECK-NEXT:    [[TMP6:%.*]] = ptrtoint ptr [[P]] to i64
1013; CHECK-NEXT:    [[TMP7:%.*]] = xor i64 [[TMP6]], 193514046488576
1014; CHECK-NEXT:    [[TMP8:%.*]] = inttoptr i64 [[TMP7]] to ptr
1015; CHECK-NEXT:    call void @llvm.aarch64.neon.st4.v2i32.p0(<2 x i32> [[TMP2]], <2 x i32> [[TMP3]], <2 x i32> [[TMP4]], <2 x i32> [[TMP5]], ptr [[TMP8]])
1016; CHECK-NEXT:    [[_MSCMP:%.*]] = icmp ne i64 [[TMP1]], 0
1017; CHECK-NEXT:    br i1 [[_MSCMP]], label [[TMP9:%.*]], label [[TMP10:%.*]], !prof [[PROF1]]
1018; CHECK:       9:
1019; CHECK-NEXT:    call void @__msan_warning_noreturn() #[[ATTR4]]
1020; CHECK-NEXT:    unreachable
1021; CHECK:       10:
1022; CHECK-NEXT:    call void @llvm.aarch64.neon.st4.v2i32.p0(<2 x i32> [[A]], <2 x i32> [[B]], <2 x i32> [[C]], <2 x i32> [[D]], ptr [[P]])
1023; CHECK-NEXT:    ret void
1024;
1025  call void @llvm.aarch64.neon.st4.v2i32.p0(<2 x i32> %A, <2 x i32> %B, <2 x i32> %C, <2 x i32> %D, ptr %P)
1026  ret void
1027}
1028
1029declare void @llvm.aarch64.neon.st2.v2i32.p0(<2 x i32>, <2 x i32>, ptr) nounwind sanitize_memory readonly
1030declare void @llvm.aarch64.neon.st3.v2i32.p0(<2 x i32>, <2 x i32>, <2 x i32>, ptr) nounwind sanitize_memory readonly
1031declare void @llvm.aarch64.neon.st4.v2i32.p0(<2 x i32>, <2 x i32>, <2 x i32>, <2 x i32>, ptr) nounwind sanitize_memory readonly
1032
1033define void @st2_4s(<4 x i32> %A, <4 x i32> %B, ptr %P) nounwind sanitize_memory {
1034;
1035;
1036; CHECK-LABEL: define void @st2_4s(
1037; CHECK-SAME: <4 x i32> [[A:%.*]], <4 x i32> [[B:%.*]], ptr [[P:%.*]]) #[[ATTR0]] {
1038; CHECK-NEXT:    [[TMP1:%.*]] = load i64, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 32) to ptr), align 8
1039; CHECK-NEXT:    [[TMP2:%.*]] = load <4 x i32>, ptr @__msan_param_tls, align 8
1040; CHECK-NEXT:    [[TMP3:%.*]] = load <4 x i32>, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 16) to ptr), align 8
1041; CHECK-NEXT:    call void @llvm.donothing()
1042; CHECK-NEXT:    [[TMP4:%.*]] = ptrtoint ptr [[P]] to i64
1043; CHECK-NEXT:    [[TMP5:%.*]] = xor i64 [[TMP4]], 193514046488576
1044; CHECK-NEXT:    [[TMP6:%.*]] = inttoptr i64 [[TMP5]] to ptr
1045; CHECK-NEXT:    call void @llvm.aarch64.neon.st2.v4i32.p0(<4 x i32> [[TMP2]], <4 x i32> [[TMP3]], ptr [[TMP6]])
1046; CHECK-NEXT:    [[_MSCMP:%.*]] = icmp ne i64 [[TMP1]], 0
1047; CHECK-NEXT:    br i1 [[_MSCMP]], label [[TMP7:%.*]], label [[TMP8:%.*]], !prof [[PROF1]]
1048; CHECK:       7:
1049; CHECK-NEXT:    call void @__msan_warning_noreturn() #[[ATTR4]]
1050; CHECK-NEXT:    unreachable
1051; CHECK:       8:
1052; CHECK-NEXT:    call void @llvm.aarch64.neon.st2.v4i32.p0(<4 x i32> [[A]], <4 x i32> [[B]], ptr [[P]])
1053; CHECK-NEXT:    ret void
1054;
1055  call void @llvm.aarch64.neon.st2.v4i32.p0(<4 x i32> %A, <4 x i32> %B, ptr %P)
1056  ret void
1057}
1058
1059define void @st3_4s(<4 x i32> %A, <4 x i32> %B, <4 x i32> %C, ptr %P) nounwind sanitize_memory {
1060;
1061;
1062; CHECK-LABEL: define void @st3_4s(
1063; CHECK-SAME: <4 x i32> [[A:%.*]], <4 x i32> [[B:%.*]], <4 x i32> [[C:%.*]], ptr [[P:%.*]]) #[[ATTR0]] {
1064; CHECK-NEXT:    [[TMP1:%.*]] = load i64, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 48) to ptr), align 8
1065; CHECK-NEXT:    [[TMP2:%.*]] = load <4 x i32>, ptr @__msan_param_tls, align 8
1066; CHECK-NEXT:    [[TMP3:%.*]] = load <4 x i32>, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 16) to ptr), align 8
1067; CHECK-NEXT:    [[TMP4:%.*]] = load <4 x i32>, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 32) to ptr), align 8
1068; CHECK-NEXT:    call void @llvm.donothing()
1069; CHECK-NEXT:    [[TMP5:%.*]] = ptrtoint ptr [[P]] to i64
1070; CHECK-NEXT:    [[TMP6:%.*]] = xor i64 [[TMP5]], 193514046488576
1071; CHECK-NEXT:    [[TMP7:%.*]] = inttoptr i64 [[TMP6]] to ptr
1072; CHECK-NEXT:    call void @llvm.aarch64.neon.st3.v4i32.p0(<4 x i32> [[TMP2]], <4 x i32> [[TMP3]], <4 x i32> [[TMP4]], ptr [[TMP7]])
1073; CHECK-NEXT:    [[_MSCMP:%.*]] = icmp ne i64 [[TMP1]], 0
1074; CHECK-NEXT:    br i1 [[_MSCMP]], label [[TMP8:%.*]], label [[TMP9:%.*]], !prof [[PROF1]]
1075; CHECK:       8:
1076; CHECK-NEXT:    call void @__msan_warning_noreturn() #[[ATTR4]]
1077; CHECK-NEXT:    unreachable
1078; CHECK:       9:
1079; CHECK-NEXT:    call void @llvm.aarch64.neon.st3.v4i32.p0(<4 x i32> [[A]], <4 x i32> [[B]], <4 x i32> [[C]], ptr [[P]])
1080; CHECK-NEXT:    ret void
1081;
1082  call void @llvm.aarch64.neon.st3.v4i32.p0(<4 x i32> %A, <4 x i32> %B, <4 x i32> %C, ptr %P)
1083  ret void
1084}
1085
1086define void @st4_4s(<4 x i32> %A, <4 x i32> %B, <4 x i32> %C, <4 x i32> %D, ptr %P) nounwind sanitize_memory {
1087;
1088;
1089; CHECK-LABEL: define void @st4_4s(
1090; CHECK-SAME: <4 x i32> [[A:%.*]], <4 x i32> [[B:%.*]], <4 x i32> [[C:%.*]], <4 x i32> [[D:%.*]], ptr [[P:%.*]]) #[[ATTR0]] {
1091; CHECK-NEXT:    [[TMP1:%.*]] = load i64, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 64) to ptr), align 8
1092; CHECK-NEXT:    [[TMP2:%.*]] = load <4 x i32>, ptr @__msan_param_tls, align 8
1093; CHECK-NEXT:    [[TMP3:%.*]] = load <4 x i32>, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 16) to ptr), align 8
1094; CHECK-NEXT:    [[TMP4:%.*]] = load <4 x i32>, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 32) to ptr), align 8
1095; CHECK-NEXT:    [[TMP5:%.*]] = load <4 x i32>, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 48) to ptr), align 8
1096; CHECK-NEXT:    call void @llvm.donothing()
1097; CHECK-NEXT:    [[TMP6:%.*]] = ptrtoint ptr [[P]] to i64
1098; CHECK-NEXT:    [[TMP7:%.*]] = xor i64 [[TMP6]], 193514046488576
1099; CHECK-NEXT:    [[TMP8:%.*]] = inttoptr i64 [[TMP7]] to ptr
1100; CHECK-NEXT:    call void @llvm.aarch64.neon.st4.v4i32.p0(<4 x i32> [[TMP2]], <4 x i32> [[TMP3]], <4 x i32> [[TMP4]], <4 x i32> [[TMP5]], ptr [[TMP8]])
1101; CHECK-NEXT:    [[_MSCMP:%.*]] = icmp ne i64 [[TMP1]], 0
1102; CHECK-NEXT:    br i1 [[_MSCMP]], label [[TMP9:%.*]], label [[TMP10:%.*]], !prof [[PROF1]]
1103; CHECK:       9:
1104; CHECK-NEXT:    call void @__msan_warning_noreturn() #[[ATTR4]]
1105; CHECK-NEXT:    unreachable
1106; CHECK:       10:
1107; CHECK-NEXT:    call void @llvm.aarch64.neon.st4.v4i32.p0(<4 x i32> [[A]], <4 x i32> [[B]], <4 x i32> [[C]], <4 x i32> [[D]], ptr [[P]])
1108; CHECK-NEXT:    ret void
1109;
1110  call void @llvm.aarch64.neon.st4.v4i32.p0(<4 x i32> %A, <4 x i32> %B, <4 x i32> %C, <4 x i32> %D, ptr %P)
1111  ret void
1112}
1113
1114; -----------------------------------------------------------------------------------------------------------------------------------------------
1115
1116declare void @llvm.aarch64.neon.st2.v4i32.p0(<4 x i32>, <4 x i32>, ptr) nounwind sanitize_memory readonly
1117declare void @llvm.aarch64.neon.st3.v4i32.p0(<4 x i32>, <4 x i32>, <4 x i32>, ptr) nounwind sanitize_memory readonly
1118declare void @llvm.aarch64.neon.st4.v4i32.p0(<4 x i32>, <4 x i32>, <4 x i32>, <4 x i32>, ptr) nounwind sanitize_memory readonly
1119
1120; If there's only one element, st2/3/4 don't make much sense, stick to st1.
1121define void @st2_1d(<1 x i64> %A, <1 x i64> %B, ptr %P) nounwind sanitize_memory {
1122;
1123;
1124; CHECK-LABEL: define void @st2_1d(
1125; CHECK-SAME: <1 x i64> [[A:%.*]], <1 x i64> [[B:%.*]], ptr [[P:%.*]]) #[[ATTR0]] {
1126; CHECK-NEXT:    [[TMP1:%.*]] = load i64, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 16) to ptr), align 8
1127; CHECK-NEXT:    [[TMP2:%.*]] = load <1 x i64>, ptr @__msan_param_tls, align 8
1128; CHECK-NEXT:    [[TMP3:%.*]] = load <1 x i64>, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 8) to ptr), align 8
1129; CHECK-NEXT:    call void @llvm.donothing()
1130; CHECK-NEXT:    [[TMP4:%.*]] = ptrtoint ptr [[P]] to i64
1131; CHECK-NEXT:    [[TMP5:%.*]] = xor i64 [[TMP4]], 193514046488576
1132; CHECK-NEXT:    [[TMP6:%.*]] = inttoptr i64 [[TMP5]] to ptr
1133; CHECK-NEXT:    call void @llvm.aarch64.neon.st2.v1i64.p0(<1 x i64> [[TMP2]], <1 x i64> [[TMP3]], ptr [[TMP6]])
1134; CHECK-NEXT:    [[_MSCMP:%.*]] = icmp ne i64 [[TMP1]], 0
1135; CHECK-NEXT:    br i1 [[_MSCMP]], label [[TMP7:%.*]], label [[TMP8:%.*]], !prof [[PROF1]]
1136; CHECK:       7:
1137; CHECK-NEXT:    call void @__msan_warning_noreturn() #[[ATTR4]]
1138; CHECK-NEXT:    unreachable
1139; CHECK:       8:
1140; CHECK-NEXT:    call void @llvm.aarch64.neon.st2.v1i64.p0(<1 x i64> [[A]], <1 x i64> [[B]], ptr [[P]])
1141; CHECK-NEXT:    ret void
1142;
1143  call void @llvm.aarch64.neon.st2.v1i64.p0(<1 x i64> %A, <1 x i64> %B, ptr %P)
1144  ret void
1145}
1146
1147define void @st3_1d(<1 x i64> %A, <1 x i64> %B, <1 x i64> %C, ptr %P) nounwind sanitize_memory {
1148;
1149;
1150; CHECK-LABEL: define void @st3_1d(
1151; CHECK-SAME: <1 x i64> [[A:%.*]], <1 x i64> [[B:%.*]], <1 x i64> [[C:%.*]], ptr [[P:%.*]]) #[[ATTR0]] {
1152; CHECK-NEXT:    [[TMP1:%.*]] = load i64, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 24) to ptr), align 8
1153; CHECK-NEXT:    [[TMP2:%.*]] = load <1 x i64>, ptr @__msan_param_tls, align 8
1154; CHECK-NEXT:    [[TMP3:%.*]] = load <1 x i64>, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 8) to ptr), align 8
1155; CHECK-NEXT:    [[TMP4:%.*]] = load <1 x i64>, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 16) to ptr), align 8
1156; CHECK-NEXT:    call void @llvm.donothing()
1157; CHECK-NEXT:    [[TMP5:%.*]] = ptrtoint ptr [[P]] to i64
1158; CHECK-NEXT:    [[TMP6:%.*]] = xor i64 [[TMP5]], 193514046488576
1159; CHECK-NEXT:    [[TMP7:%.*]] = inttoptr i64 [[TMP6]] to ptr
1160; CHECK-NEXT:    call void @llvm.aarch64.neon.st3.v1i64.p0(<1 x i64> [[TMP2]], <1 x i64> [[TMP3]], <1 x i64> [[TMP4]], ptr [[TMP7]])
1161; CHECK-NEXT:    [[_MSCMP:%.*]] = icmp ne i64 [[TMP1]], 0
1162; CHECK-NEXT:    br i1 [[_MSCMP]], label [[TMP8:%.*]], label [[TMP9:%.*]], !prof [[PROF1]]
1163; CHECK:       8:
1164; CHECK-NEXT:    call void @__msan_warning_noreturn() #[[ATTR4]]
1165; CHECK-NEXT:    unreachable
1166; CHECK:       9:
1167; CHECK-NEXT:    call void @llvm.aarch64.neon.st3.v1i64.p0(<1 x i64> [[A]], <1 x i64> [[B]], <1 x i64> [[C]], ptr [[P]])
1168; CHECK-NEXT:    ret void
1169;
1170  call void @llvm.aarch64.neon.st3.v1i64.p0(<1 x i64> %A, <1 x i64> %B, <1 x i64> %C, ptr %P)
1171  ret void
1172}
1173
1174define void @st4_1d(<1 x i64> %A, <1 x i64> %B, <1 x i64> %C, <1 x i64> %D, ptr %P) nounwind sanitize_memory {
1175;
1176;
1177; CHECK-LABEL: define void @st4_1d(
1178; CHECK-SAME: <1 x i64> [[A:%.*]], <1 x i64> [[B:%.*]], <1 x i64> [[C:%.*]], <1 x i64> [[D:%.*]], ptr [[P:%.*]]) #[[ATTR0]] {
1179; CHECK-NEXT:    [[TMP1:%.*]] = load i64, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 32) to ptr), align 8
1180; CHECK-NEXT:    [[TMP2:%.*]] = load <1 x i64>, ptr @__msan_param_tls, align 8
1181; CHECK-NEXT:    [[TMP3:%.*]] = load <1 x i64>, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 8) to ptr), align 8
1182; CHECK-NEXT:    [[TMP4:%.*]] = load <1 x i64>, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 16) to ptr), align 8
1183; CHECK-NEXT:    [[TMP5:%.*]] = load <1 x i64>, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 24) to ptr), align 8
1184; CHECK-NEXT:    call void @llvm.donothing()
1185; CHECK-NEXT:    [[TMP6:%.*]] = ptrtoint ptr [[P]] to i64
1186; CHECK-NEXT:    [[TMP7:%.*]] = xor i64 [[TMP6]], 193514046488576
1187; CHECK-NEXT:    [[TMP8:%.*]] = inttoptr i64 [[TMP7]] to ptr
1188; CHECK-NEXT:    call void @llvm.aarch64.neon.st4.v1i64.p0(<1 x i64> [[TMP2]], <1 x i64> [[TMP3]], <1 x i64> [[TMP4]], <1 x i64> [[TMP5]], ptr [[TMP8]])
1189; CHECK-NEXT:    [[_MSCMP:%.*]] = icmp ne i64 [[TMP1]], 0
1190; CHECK-NEXT:    br i1 [[_MSCMP]], label [[TMP9:%.*]], label [[TMP10:%.*]], !prof [[PROF1]]
1191; CHECK:       9:
1192; CHECK-NEXT:    call void @__msan_warning_noreturn() #[[ATTR4]]
1193; CHECK-NEXT:    unreachable
1194; CHECK:       10:
1195; CHECK-NEXT:    call void @llvm.aarch64.neon.st4.v1i64.p0(<1 x i64> [[A]], <1 x i64> [[B]], <1 x i64> [[C]], <1 x i64> [[D]], ptr [[P]])
1196; CHECK-NEXT:    ret void
1197;
1198  call void @llvm.aarch64.neon.st4.v1i64.p0(<1 x i64> %A, <1 x i64> %B, <1 x i64> %C, <1 x i64> %D, ptr %P)
1199  ret void
1200}
1201
1202; -----------------------------------------------------------------------------------------------------------------------------------------------
1203
1204declare void @llvm.aarch64.neon.st2.v1i64.p0(<1 x i64>, <1 x i64>, ptr) nounwind sanitize_memory readonly
1205declare void @llvm.aarch64.neon.st3.v1i64.p0(<1 x i64>, <1 x i64>, <1 x i64>, ptr) nounwind sanitize_memory readonly
1206declare void @llvm.aarch64.neon.st4.v1i64.p0(<1 x i64>, <1 x i64>, <1 x i64>, <1 x i64>, ptr) nounwind sanitize_memory readonly
1207
1208define void @st2_2d(<2 x i64> %A, <2 x i64> %B, ptr %P) nounwind sanitize_memory {
1209;
1210;
1211; CHECK-LABEL: define void @st2_2d(
1212; CHECK-SAME: <2 x i64> [[A:%.*]], <2 x i64> [[B:%.*]], ptr [[P:%.*]]) #[[ATTR0]] {
1213; CHECK-NEXT:    [[TMP1:%.*]] = load i64, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 32) to ptr), align 8
1214; CHECK-NEXT:    [[TMP2:%.*]] = load <2 x i64>, ptr @__msan_param_tls, align 8
1215; CHECK-NEXT:    [[TMP3:%.*]] = load <2 x i64>, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 16) to ptr), align 8
1216; CHECK-NEXT:    call void @llvm.donothing()
1217; CHECK-NEXT:    [[TMP4:%.*]] = ptrtoint ptr [[P]] to i64
1218; CHECK-NEXT:    [[TMP5:%.*]] = xor i64 [[TMP4]], 193514046488576
1219; CHECK-NEXT:    [[TMP6:%.*]] = inttoptr i64 [[TMP5]] to ptr
1220; CHECK-NEXT:    call void @llvm.aarch64.neon.st2.v2i64.p0(<2 x i64> [[TMP2]], <2 x i64> [[TMP3]], ptr [[TMP6]])
1221; CHECK-NEXT:    [[_MSCMP:%.*]] = icmp ne i64 [[TMP1]], 0
1222; CHECK-NEXT:    br i1 [[_MSCMP]], label [[TMP7:%.*]], label [[TMP8:%.*]], !prof [[PROF1]]
1223; CHECK:       7:
1224; CHECK-NEXT:    call void @__msan_warning_noreturn() #[[ATTR4]]
1225; CHECK-NEXT:    unreachable
1226; CHECK:       8:
1227; CHECK-NEXT:    call void @llvm.aarch64.neon.st2.v2i64.p0(<2 x i64> [[A]], <2 x i64> [[B]], ptr [[P]])
1228; CHECK-NEXT:    ret void
1229;
1230  call void @llvm.aarch64.neon.st2.v2i64.p0(<2 x i64> %A, <2 x i64> %B, ptr %P)
1231  ret void
1232}
1233
1234define void @st2_2d_undefA(<2 x i64> %A, <2 x i64> %B, ptr %P) nounwind sanitize_memory {
1235;
1236;
1237; CHECK-LABEL: define void @st2_2d_undefA(
1238; CHECK-SAME: <2 x i64> [[A:%.*]], <2 x i64> [[B:%.*]], ptr [[P:%.*]]) #[[ATTR0]] {
1239; CHECK-NEXT:    [[TMP1:%.*]] = load i64, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 32) to ptr), align 8
1240; CHECK-NEXT:    [[TMP2:%.*]] = load <2 x i64>, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 16) to ptr), align 8
1241; CHECK-NEXT:    call void @llvm.donothing()
1242; CHECK-NEXT:    [[TMP3:%.*]] = ptrtoint ptr [[P]] to i64
1243; CHECK-NEXT:    [[TMP4:%.*]] = xor i64 [[TMP3]], 193514046488576
1244; CHECK-NEXT:    [[TMP5:%.*]] = inttoptr i64 [[TMP4]] to ptr
1245; CHECK-NEXT:    call void @llvm.aarch64.neon.st2.v2i64.p0(<2 x i64> splat (i64 -1), <2 x i64> [[TMP2]], ptr [[TMP5]])
1246; CHECK-NEXT:    [[_MSCMP:%.*]] = icmp ne i64 [[TMP1]], 0
1247; CHECK-NEXT:    br i1 [[_MSCMP]], label [[TMP6:%.*]], label [[TMP7:%.*]], !prof [[PROF1]]
1248; CHECK:       6:
1249; CHECK-NEXT:    call void @__msan_warning_noreturn() #[[ATTR4]]
1250; CHECK-NEXT:    unreachable
1251; CHECK:       7:
1252; CHECK-NEXT:    call void @llvm.aarch64.neon.st2.v2i64.p0(<2 x i64> undef, <2 x i64> [[B]], ptr [[P]])
1253; CHECK-NEXT:    ret void
1254;
1255  call void @llvm.aarch64.neon.st2.v2i64.p0(<2 x i64> undef, <2 x i64> %B, ptr %P)
1256  ret void
1257}
1258
1259define void @st2_2d_undefB(<2 x i64> %A, <2 x i64> %B, ptr %P) nounwind sanitize_memory {
1260;
1261;
1262; CHECK-LABEL: define void @st2_2d_undefB(
1263; CHECK-SAME: <2 x i64> [[A:%.*]], <2 x i64> [[B:%.*]], ptr [[P:%.*]]) #[[ATTR0]] {
1264; CHECK-NEXT:    [[TMP1:%.*]] = load i64, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 32) to ptr), align 8
1265; CHECK-NEXT:    [[TMP2:%.*]] = load <2 x i64>, ptr @__msan_param_tls, align 8
1266; CHECK-NEXT:    call void @llvm.donothing()
1267; CHECK-NEXT:    [[TMP3:%.*]] = ptrtoint ptr [[P]] to i64
1268; CHECK-NEXT:    [[TMP4:%.*]] = xor i64 [[TMP3]], 193514046488576
1269; CHECK-NEXT:    [[TMP5:%.*]] = inttoptr i64 [[TMP4]] to ptr
1270; CHECK-NEXT:    call void @llvm.aarch64.neon.st2.v2i64.p0(<2 x i64> [[TMP2]], <2 x i64> splat (i64 -1), ptr [[TMP5]])
1271; CHECK-NEXT:    [[_MSCMP:%.*]] = icmp ne i64 [[TMP1]], 0
1272; CHECK-NEXT:    br i1 [[_MSCMP]], label [[TMP6:%.*]], label [[TMP7:%.*]], !prof [[PROF1]]
1273; CHECK:       6:
1274; CHECK-NEXT:    call void @__msan_warning_noreturn() #[[ATTR4]]
1275; CHECK-NEXT:    unreachable
1276; CHECK:       7:
1277; CHECK-NEXT:    call void @llvm.aarch64.neon.st2.v2i64.p0(<2 x i64> [[A]], <2 x i64> undef, ptr [[P]])
1278; CHECK-NEXT:    ret void
1279;
1280  call void @llvm.aarch64.neon.st2.v2i64.p0(<2 x i64> %A, <2 x i64> undef, ptr %P)
1281  ret void
1282}
1283
1284define void @st2_2d_undefAB(<2 x i64> %A, <2 x i64> %B, ptr %P) nounwind sanitize_memory {
1285;
1286;
1287; CHECK-LABEL: define void @st2_2d_undefAB(
1288; CHECK-SAME: <2 x i64> [[A:%.*]], <2 x i64> [[B:%.*]], ptr [[P:%.*]]) #[[ATTR0]] {
1289; CHECK-NEXT:    [[TMP1:%.*]] = load i64, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 32) to ptr), align 8
1290; CHECK-NEXT:    call void @llvm.donothing()
1291; CHECK-NEXT:    [[TMP2:%.*]] = ptrtoint ptr [[P]] to i64
1292; CHECK-NEXT:    [[TMP3:%.*]] = xor i64 [[TMP2]], 193514046488576
1293; CHECK-NEXT:    [[TMP4:%.*]] = inttoptr i64 [[TMP3]] to ptr
1294; CHECK-NEXT:    call void @llvm.aarch64.neon.st2.v2i64.p0(<2 x i64> splat (i64 -1), <2 x i64> splat (i64 -1), ptr [[TMP4]])
1295; CHECK-NEXT:    [[_MSCMP:%.*]] = icmp ne i64 [[TMP1]], 0
1296; CHECK-NEXT:    br i1 [[_MSCMP]], label [[TMP5:%.*]], label [[TMP6:%.*]], !prof [[PROF1]]
1297; CHECK:       5:
1298; CHECK-NEXT:    call void @__msan_warning_noreturn() #[[ATTR4]]
1299; CHECK-NEXT:    unreachable
1300; CHECK:       6:
1301; CHECK-NEXT:    call void @llvm.aarch64.neon.st2.v2i64.p0(<2 x i64> undef, <2 x i64> undef, ptr [[P]])
1302; CHECK-NEXT:    ret void
1303;
1304  call void @llvm.aarch64.neon.st2.v2i64.p0(<2 x i64> undef, <2 x i64> undef, ptr %P)
1305  ret void
1306}
1307
1308define void @st3_2d(<2 x i64> %A, <2 x i64> %B, <2 x i64> %C, ptr %P) nounwind sanitize_memory {
1309;
1310;
1311; CHECK-LABEL: define void @st3_2d(
1312; CHECK-SAME: <2 x i64> [[A:%.*]], <2 x i64> [[B:%.*]], <2 x i64> [[C:%.*]], ptr [[P:%.*]]) #[[ATTR0]] {
1313; CHECK-NEXT:    [[TMP1:%.*]] = load i64, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 48) to ptr), align 8
1314; CHECK-NEXT:    [[TMP2:%.*]] = load <2 x i64>, ptr @__msan_param_tls, align 8
1315; CHECK-NEXT:    [[TMP3:%.*]] = load <2 x i64>, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 16) to ptr), align 8
1316; CHECK-NEXT:    [[TMP4:%.*]] = load <2 x i64>, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 32) to ptr), align 8
1317; CHECK-NEXT:    call void @llvm.donothing()
1318; CHECK-NEXT:    [[TMP5:%.*]] = ptrtoint ptr [[P]] to i64
1319; CHECK-NEXT:    [[TMP6:%.*]] = xor i64 [[TMP5]], 193514046488576
1320; CHECK-NEXT:    [[TMP7:%.*]] = inttoptr i64 [[TMP6]] to ptr
1321; CHECK-NEXT:    call void @llvm.aarch64.neon.st3.v2i64.p0(<2 x i64> [[TMP2]], <2 x i64> [[TMP3]], <2 x i64> [[TMP4]], ptr [[TMP7]])
1322; CHECK-NEXT:    [[_MSCMP:%.*]] = icmp ne i64 [[TMP1]], 0
1323; CHECK-NEXT:    br i1 [[_MSCMP]], label [[TMP8:%.*]], label [[TMP9:%.*]], !prof [[PROF1]]
1324; CHECK:       8:
1325; CHECK-NEXT:    call void @__msan_warning_noreturn() #[[ATTR4]]
1326; CHECK-NEXT:    unreachable
1327; CHECK:       9:
1328; CHECK-NEXT:    call void @llvm.aarch64.neon.st3.v2i64.p0(<2 x i64> [[A]], <2 x i64> [[B]], <2 x i64> [[C]], ptr [[P]])
1329; CHECK-NEXT:    ret void
1330;
1331  call void @llvm.aarch64.neon.st3.v2i64.p0(<2 x i64> %A, <2 x i64> %B, <2 x i64> %C, ptr %P)
1332  ret void
1333}
1334
1335define void @st3_2d_undefA(<2 x i64> %A, <2 x i64> %B, <2 x i64> %C, ptr %P) nounwind sanitize_memory {
1336;
1337;
1338; CHECK-LABEL: define void @st3_2d_undefA(
1339; CHECK-SAME: <2 x i64> [[A:%.*]], <2 x i64> [[B:%.*]], <2 x i64> [[C:%.*]], ptr [[P:%.*]]) #[[ATTR0]] {
1340; CHECK-NEXT:    [[TMP1:%.*]] = load i64, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 48) to ptr), align 8
1341; CHECK-NEXT:    [[TMP2:%.*]] = load <2 x i64>, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 16) to ptr), align 8
1342; CHECK-NEXT:    [[TMP3:%.*]] = load <2 x i64>, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 32) to ptr), align 8
1343; CHECK-NEXT:    call void @llvm.donothing()
1344; CHECK-NEXT:    [[TMP4:%.*]] = ptrtoint ptr [[P]] to i64
1345; CHECK-NEXT:    [[TMP5:%.*]] = xor i64 [[TMP4]], 193514046488576
1346; CHECK-NEXT:    [[TMP6:%.*]] = inttoptr i64 [[TMP5]] to ptr
1347; CHECK-NEXT:    call void @llvm.aarch64.neon.st3.v2i64.p0(<2 x i64> splat (i64 -1), <2 x i64> [[TMP2]], <2 x i64> [[TMP3]], ptr [[TMP6]])
1348; CHECK-NEXT:    [[_MSCMP:%.*]] = icmp ne i64 [[TMP1]], 0
1349; CHECK-NEXT:    br i1 [[_MSCMP]], label [[TMP7:%.*]], label [[TMP8:%.*]], !prof [[PROF1]]
1350; CHECK:       7:
1351; CHECK-NEXT:    call void @__msan_warning_noreturn() #[[ATTR4]]
1352; CHECK-NEXT:    unreachable
1353; CHECK:       8:
1354; CHECK-NEXT:    call void @llvm.aarch64.neon.st3.v2i64.p0(<2 x i64> undef, <2 x i64> [[B]], <2 x i64> [[C]], ptr [[P]])
1355; CHECK-NEXT:    ret void
1356;
1357  call void @llvm.aarch64.neon.st3.v2i64.p0(<2 x i64> undef, <2 x i64> %B, <2 x i64> %C, ptr %P)
1358  ret void
1359}
1360
1361define void @st3_2d_undefB(<2 x i64> %A, <2 x i64> %B, <2 x i64> %C, ptr %P) nounwind sanitize_memory {
1362;
1363;
1364; CHECK-LABEL: define void @st3_2d_undefB(
1365; CHECK-SAME: <2 x i64> [[A:%.*]], <2 x i64> [[B:%.*]], <2 x i64> [[C:%.*]], ptr [[P:%.*]]) #[[ATTR0]] {
1366; CHECK-NEXT:    [[TMP1:%.*]] = load i64, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 48) to ptr), align 8
1367; CHECK-NEXT:    [[TMP2:%.*]] = load <2 x i64>, ptr @__msan_param_tls, align 8
1368; CHECK-NEXT:    [[TMP3:%.*]] = load <2 x i64>, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 32) to ptr), align 8
1369; CHECK-NEXT:    call void @llvm.donothing()
1370; CHECK-NEXT:    [[TMP4:%.*]] = ptrtoint ptr [[P]] to i64
1371; CHECK-NEXT:    [[TMP5:%.*]] = xor i64 [[TMP4]], 193514046488576
1372; CHECK-NEXT:    [[TMP6:%.*]] = inttoptr i64 [[TMP5]] to ptr
1373; CHECK-NEXT:    call void @llvm.aarch64.neon.st3.v2i64.p0(<2 x i64> [[TMP2]], <2 x i64> splat (i64 -1), <2 x i64> [[TMP3]], ptr [[TMP6]])
1374; CHECK-NEXT:    [[_MSCMP:%.*]] = icmp ne i64 [[TMP1]], 0
1375; CHECK-NEXT:    br i1 [[_MSCMP]], label [[TMP7:%.*]], label [[TMP8:%.*]], !prof [[PROF1]]
1376; CHECK:       7:
1377; CHECK-NEXT:    call void @__msan_warning_noreturn() #[[ATTR4]]
1378; CHECK-NEXT:    unreachable
1379; CHECK:       8:
1380; CHECK-NEXT:    call void @llvm.aarch64.neon.st3.v2i64.p0(<2 x i64> [[A]], <2 x i64> undef, <2 x i64> [[C]], ptr [[P]])
1381; CHECK-NEXT:    ret void
1382;
1383  call void @llvm.aarch64.neon.st3.v2i64.p0(<2 x i64> %A, <2 x i64> undef, <2 x i64> %C, ptr %P)
1384  ret void
1385}
1386
1387define void @st3_2d_undefC(<2 x i64> %A, <2 x i64> %B, <2 x i64> %C, ptr %P) nounwind sanitize_memory {
1388;
1389;
1390; CHECK-LABEL: define void @st3_2d_undefC(
1391; CHECK-SAME: <2 x i64> [[A:%.*]], <2 x i64> [[B:%.*]], <2 x i64> [[C:%.*]], ptr [[P:%.*]]) #[[ATTR0]] {
1392; CHECK-NEXT:    [[TMP1:%.*]] = load i64, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 48) to ptr), align 8
1393; CHECK-NEXT:    [[TMP2:%.*]] = load <2 x i64>, ptr @__msan_param_tls, align 8
1394; CHECK-NEXT:    [[TMP3:%.*]] = load <2 x i64>, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 16) to ptr), align 8
1395; CHECK-NEXT:    call void @llvm.donothing()
1396; CHECK-NEXT:    [[TMP4:%.*]] = ptrtoint ptr [[P]] to i64
1397; CHECK-NEXT:    [[TMP5:%.*]] = xor i64 [[TMP4]], 193514046488576
1398; CHECK-NEXT:    [[TMP6:%.*]] = inttoptr i64 [[TMP5]] to ptr
1399; CHECK-NEXT:    call void @llvm.aarch64.neon.st3.v2i64.p0(<2 x i64> [[TMP2]], <2 x i64> [[TMP3]], <2 x i64> splat (i64 -1), ptr [[TMP6]])
1400; CHECK-NEXT:    [[_MSCMP:%.*]] = icmp ne i64 [[TMP1]], 0
1401; CHECK-NEXT:    br i1 [[_MSCMP]], label [[TMP7:%.*]], label [[TMP8:%.*]], !prof [[PROF1]]
1402; CHECK:       7:
1403; CHECK-NEXT:    call void @__msan_warning_noreturn() #[[ATTR4]]
1404; CHECK-NEXT:    unreachable
1405; CHECK:       8:
1406; CHECK-NEXT:    call void @llvm.aarch64.neon.st3.v2i64.p0(<2 x i64> [[A]], <2 x i64> [[B]], <2 x i64> undef, ptr [[P]])
1407; CHECK-NEXT:    ret void
1408;
1409  call void @llvm.aarch64.neon.st3.v2i64.p0(<2 x i64> %A, <2 x i64> %B, <2 x i64> undef, ptr %P)
1410  ret void
1411}
1412
1413define void @st3_2d_undefAB(<2 x i64> %A, <2 x i64> %B, <2 x i64> %C, ptr %P) nounwind sanitize_memory {
1414;
1415;
1416; CHECK-LABEL: define void @st3_2d_undefAB(
1417; CHECK-SAME: <2 x i64> [[A:%.*]], <2 x i64> [[B:%.*]], <2 x i64> [[C:%.*]], ptr [[P:%.*]]) #[[ATTR0]] {
1418; CHECK-NEXT:    [[TMP1:%.*]] = load i64, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 48) to ptr), align 8
1419; CHECK-NEXT:    [[TMP2:%.*]] = load <2 x i64>, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 32) to ptr), align 8
1420; CHECK-NEXT:    call void @llvm.donothing()
1421; CHECK-NEXT:    [[TMP3:%.*]] = ptrtoint ptr [[P]] to i64
1422; CHECK-NEXT:    [[TMP4:%.*]] = xor i64 [[TMP3]], 193514046488576
1423; CHECK-NEXT:    [[TMP5:%.*]] = inttoptr i64 [[TMP4]] to ptr
1424; CHECK-NEXT:    call void @llvm.aarch64.neon.st3.v2i64.p0(<2 x i64> splat (i64 -1), <2 x i64> splat (i64 -1), <2 x i64> [[TMP2]], ptr [[TMP5]])
1425; CHECK-NEXT:    [[_MSCMP:%.*]] = icmp ne i64 [[TMP1]], 0
1426; CHECK-NEXT:    br i1 [[_MSCMP]], label [[TMP6:%.*]], label [[TMP7:%.*]], !prof [[PROF1]]
1427; CHECK:       6:
1428; CHECK-NEXT:    call void @__msan_warning_noreturn() #[[ATTR4]]
1429; CHECK-NEXT:    unreachable
1430; CHECK:       7:
1431; CHECK-NEXT:    call void @llvm.aarch64.neon.st3.v2i64.p0(<2 x i64> undef, <2 x i64> undef, <2 x i64> [[C]], ptr [[P]])
1432; CHECK-NEXT:    ret void
1433;
1434  call void @llvm.aarch64.neon.st3.v2i64.p0(<2 x i64> undef, <2 x i64> undef, <2 x i64> %C, ptr %P)
1435  ret void
1436}
1437
1438define void @st3_2d_undefAC(<2 x i64> %A, <2 x i64> %B, <2 x i64> %C, ptr %P) nounwind sanitize_memory {
1439;
1440;
1441; CHECK-LABEL: define void @st3_2d_undefAC(
1442; CHECK-SAME: <2 x i64> [[A:%.*]], <2 x i64> [[B:%.*]], <2 x i64> [[C:%.*]], ptr [[P:%.*]]) #[[ATTR0]] {
1443; CHECK-NEXT:    [[TMP1:%.*]] = load i64, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 48) to ptr), align 8
1444; CHECK-NEXT:    [[TMP2:%.*]] = load <2 x i64>, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 16) to ptr), align 8
1445; CHECK-NEXT:    call void @llvm.donothing()
1446; CHECK-NEXT:    [[TMP3:%.*]] = ptrtoint ptr [[P]] to i64
1447; CHECK-NEXT:    [[TMP4:%.*]] = xor i64 [[TMP3]], 193514046488576
1448; CHECK-NEXT:    [[TMP5:%.*]] = inttoptr i64 [[TMP4]] to ptr
1449; CHECK-NEXT:    call void @llvm.aarch64.neon.st3.v2i64.p0(<2 x i64> splat (i64 -1), <2 x i64> [[TMP2]], <2 x i64> splat (i64 -1), ptr [[TMP5]])
1450; CHECK-NEXT:    [[_MSCMP:%.*]] = icmp ne i64 [[TMP1]], 0
1451; CHECK-NEXT:    br i1 [[_MSCMP]], label [[TMP6:%.*]], label [[TMP7:%.*]], !prof [[PROF1]]
1452; CHECK:       6:
1453; CHECK-NEXT:    call void @__msan_warning_noreturn() #[[ATTR4]]
1454; CHECK-NEXT:    unreachable
1455; CHECK:       7:
1456; CHECK-NEXT:    call void @llvm.aarch64.neon.st3.v2i64.p0(<2 x i64> undef, <2 x i64> [[B]], <2 x i64> undef, ptr [[P]])
1457; CHECK-NEXT:    ret void
1458;
1459  call void @llvm.aarch64.neon.st3.v2i64.p0(<2 x i64> undef, <2 x i64> %B, <2 x i64> undef, ptr %P)
1460  ret void
1461}
1462
1463define void @st3_2d_undefBC(<2 x i64> %A, <2 x i64> %B, <2 x i64> %C, ptr %P) nounwind sanitize_memory {
1464;
1465;
1466; CHECK-LABEL: define void @st3_2d_undefBC(
1467; CHECK-SAME: <2 x i64> [[A:%.*]], <2 x i64> [[B:%.*]], <2 x i64> [[C:%.*]], ptr [[P:%.*]]) #[[ATTR0]] {
1468; CHECK-NEXT:    [[TMP1:%.*]] = load i64, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 48) to ptr), align 8
1469; CHECK-NEXT:    [[TMP2:%.*]] = load <2 x i64>, ptr @__msan_param_tls, align 8
1470; CHECK-NEXT:    call void @llvm.donothing()
1471; CHECK-NEXT:    [[TMP3:%.*]] = ptrtoint ptr [[P]] to i64
1472; CHECK-NEXT:    [[TMP4:%.*]] = xor i64 [[TMP3]], 193514046488576
1473; CHECK-NEXT:    [[TMP5:%.*]] = inttoptr i64 [[TMP4]] to ptr
1474; CHECK-NEXT:    call void @llvm.aarch64.neon.st3.v2i64.p0(<2 x i64> [[TMP2]], <2 x i64> splat (i64 -1), <2 x i64> splat (i64 -1), ptr [[TMP5]])
1475; CHECK-NEXT:    [[_MSCMP:%.*]] = icmp ne i64 [[TMP1]], 0
1476; CHECK-NEXT:    br i1 [[_MSCMP]], label [[TMP6:%.*]], label [[TMP7:%.*]], !prof [[PROF1]]
1477; CHECK:       6:
1478; CHECK-NEXT:    call void @__msan_warning_noreturn() #[[ATTR4]]
1479; CHECK-NEXT:    unreachable
1480; CHECK:       7:
1481; CHECK-NEXT:    call void @llvm.aarch64.neon.st3.v2i64.p0(<2 x i64> [[A]], <2 x i64> undef, <2 x i64> undef, ptr [[P]])
1482; CHECK-NEXT:    ret void
1483;
1484  call void @llvm.aarch64.neon.st3.v2i64.p0(<2 x i64> %A, <2 x i64> undef, <2 x i64> undef, ptr %P)
1485  ret void
1486}
1487
1488define void @st3_2d_undefABC(<2 x i64> %A, <2 x i64> %B, <2 x i64> %C, ptr %P) nounwind sanitize_memory {
1489;
1490;
1491; CHECK-LABEL: define void @st3_2d_undefABC(
1492; CHECK-SAME: <2 x i64> [[A:%.*]], <2 x i64> [[B:%.*]], <2 x i64> [[C:%.*]], ptr [[P:%.*]]) #[[ATTR0]] {
1493; CHECK-NEXT:    [[TMP1:%.*]] = load i64, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 48) to ptr), align 8
1494; CHECK-NEXT:    call void @llvm.donothing()
1495; CHECK-NEXT:    [[TMP2:%.*]] = ptrtoint ptr [[P]] to i64
1496; CHECK-NEXT:    [[TMP3:%.*]] = xor i64 [[TMP2]], 193514046488576
1497; CHECK-NEXT:    [[TMP4:%.*]] = inttoptr i64 [[TMP3]] to ptr
1498; CHECK-NEXT:    call void @llvm.aarch64.neon.st3.v2i64.p0(<2 x i64> splat (i64 -1), <2 x i64> splat (i64 -1), <2 x i64> splat (i64 -1), ptr [[TMP4]])
1499; CHECK-NEXT:    [[_MSCMP:%.*]] = icmp ne i64 [[TMP1]], 0
1500; CHECK-NEXT:    br i1 [[_MSCMP]], label [[TMP5:%.*]], label [[TMP6:%.*]], !prof [[PROF1]]
1501; CHECK:       5:
1502; CHECK-NEXT:    call void @__msan_warning_noreturn() #[[ATTR4]]
1503; CHECK-NEXT:    unreachable
1504; CHECK:       6:
1505; CHECK-NEXT:    call void @llvm.aarch64.neon.st3.v2i64.p0(<2 x i64> undef, <2 x i64> undef, <2 x i64> undef, ptr [[P]])
1506; CHECK-NEXT:    ret void
1507;
1508  call void @llvm.aarch64.neon.st3.v2i64.p0(<2 x i64> undef, <2 x i64> undef, <2 x i64> undef, ptr %P)
1509  ret void
1510}
1511
1512define void @st4_2d(<2 x i64> %A, <2 x i64> %B, <2 x i64> %C, <2 x i64> %D, ptr %P) nounwind sanitize_memory {
1513;
1514;
1515; CHECK-LABEL: define void @st4_2d(
1516; CHECK-SAME: <2 x i64> [[A:%.*]], <2 x i64> [[B:%.*]], <2 x i64> [[C:%.*]], <2 x i64> [[D:%.*]], ptr [[P:%.*]]) #[[ATTR0]] {
1517; CHECK-NEXT:    [[TMP1:%.*]] = load i64, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 64) to ptr), align 8
1518; CHECK-NEXT:    [[TMP2:%.*]] = load <2 x i64>, ptr @__msan_param_tls, align 8
1519; CHECK-NEXT:    [[TMP3:%.*]] = load <2 x i64>, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 16) to ptr), align 8
1520; CHECK-NEXT:    [[TMP4:%.*]] = load <2 x i64>, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 32) to ptr), align 8
1521; CHECK-NEXT:    [[TMP5:%.*]] = load <2 x i64>, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 48) to ptr), align 8
1522; CHECK-NEXT:    call void @llvm.donothing()
1523; CHECK-NEXT:    [[TMP6:%.*]] = ptrtoint ptr [[P]] to i64
1524; CHECK-NEXT:    [[TMP7:%.*]] = xor i64 [[TMP6]], 193514046488576
1525; CHECK-NEXT:    [[TMP8:%.*]] = inttoptr i64 [[TMP7]] to ptr
1526; CHECK-NEXT:    call void @llvm.aarch64.neon.st4.v2i64.p0(<2 x i64> [[TMP2]], <2 x i64> [[TMP3]], <2 x i64> [[TMP4]], <2 x i64> [[TMP5]], ptr [[TMP8]])
1527; CHECK-NEXT:    [[_MSCMP:%.*]] = icmp ne i64 [[TMP1]], 0
1528; CHECK-NEXT:    br i1 [[_MSCMP]], label [[TMP9:%.*]], label [[TMP10:%.*]], !prof [[PROF1]]
1529; CHECK:       9:
1530; CHECK-NEXT:    call void @__msan_warning_noreturn() #[[ATTR4]]
1531; CHECK-NEXT:    unreachable
1532; CHECK:       10:
1533; CHECK-NEXT:    call void @llvm.aarch64.neon.st4.v2i64.p0(<2 x i64> [[A]], <2 x i64> [[B]], <2 x i64> [[C]], <2 x i64> [[D]], ptr [[P]])
1534; CHECK-NEXT:    ret void
1535;
1536  call void @llvm.aarch64.neon.st4.v2i64.p0(<2 x i64> %A, <2 x i64> %B, <2 x i64> %C, <2 x i64> %D, ptr %P)
1537  ret void
1538}
1539
1540declare void @llvm.aarch64.neon.st2.v2i64.p0(<2 x i64>, <2 x i64>, ptr) nounwind sanitize_memory readonly
1541declare void @llvm.aarch64.neon.st3.v2i64.p0(<2 x i64>, <2 x i64>, <2 x i64>, ptr) nounwind sanitize_memory readonly
1542declare void @llvm.aarch64.neon.st4.v2i64.p0(<2 x i64>, <2 x i64>, <2 x i64>, <2 x i64>, ptr) nounwind sanitize_memory readonly
1543
1544define void @st4_2d_undefA(<2 x i64> %A, <2 x i64> %B, <2 x i64> %C, <2 x i64> %D, ptr %P) nounwind sanitize_memory {
1545;
1546;
1547; CHECK-LABEL: define void @st4_2d_undefA(
1548; CHECK-SAME: <2 x i64> [[A:%.*]], <2 x i64> [[B:%.*]], <2 x i64> [[C:%.*]], <2 x i64> [[D:%.*]], ptr [[P:%.*]]) #[[ATTR0]] {
1549; CHECK-NEXT:    [[TMP1:%.*]] = load i64, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 64) to ptr), align 8
1550; CHECK-NEXT:    [[TMP2:%.*]] = load <2 x i64>, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 16) to ptr), align 8
1551; CHECK-NEXT:    [[TMP3:%.*]] = load <2 x i64>, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 32) to ptr), align 8
1552; CHECK-NEXT:    [[TMP4:%.*]] = load <2 x i64>, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 48) to ptr), align 8
1553; CHECK-NEXT:    call void @llvm.donothing()
1554; CHECK-NEXT:    [[TMP5:%.*]] = ptrtoint ptr [[P]] to i64
1555; CHECK-NEXT:    [[TMP6:%.*]] = xor i64 [[TMP5]], 193514046488576
1556; CHECK-NEXT:    [[TMP7:%.*]] = inttoptr i64 [[TMP6]] to ptr
1557; CHECK-NEXT:    call void @llvm.aarch64.neon.st4.v2i64.p0(<2 x i64> splat (i64 -1), <2 x i64> [[TMP2]], <2 x i64> [[TMP3]], <2 x i64> [[TMP4]], ptr [[TMP7]])
1558; CHECK-NEXT:    [[_MSCMP:%.*]] = icmp ne i64 [[TMP1]], 0
1559; CHECK-NEXT:    br i1 [[_MSCMP]], label [[TMP8:%.*]], label [[TMP9:%.*]], !prof [[PROF1]]
1560; CHECK:       8:
1561; CHECK-NEXT:    call void @__msan_warning_noreturn() #[[ATTR4]]
1562; CHECK-NEXT:    unreachable
1563; CHECK:       9:
1564; CHECK-NEXT:    call void @llvm.aarch64.neon.st4.v2i64.p0(<2 x i64> undef, <2 x i64> [[B]], <2 x i64> [[C]], <2 x i64> [[D]], ptr [[P]])
1565; CHECK-NEXT:    ret void
1566;
1567  call void @llvm.aarch64.neon.st4.v2i64.p0(<2 x i64> undef, <2 x i64> %B, <2 x i64> %C, <2 x i64> %D, ptr %P)
1568  ret void
1569}
1570
1571define void @st4_2d_undefB(<2 x i64> %A, <2 x i64> %B, <2 x i64> %C, <2 x i64> %D, ptr %P) nounwind sanitize_memory {
1572;
1573;
1574; CHECK-LABEL: define void @st4_2d_undefB(
1575; CHECK-SAME: <2 x i64> [[A:%.*]], <2 x i64> [[B:%.*]], <2 x i64> [[C:%.*]], <2 x i64> [[D:%.*]], ptr [[P:%.*]]) #[[ATTR0]] {
1576; CHECK-NEXT:    [[TMP1:%.*]] = load i64, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 64) to ptr), align 8
1577; CHECK-NEXT:    [[TMP2:%.*]] = load <2 x i64>, ptr @__msan_param_tls, align 8
1578; CHECK-NEXT:    [[TMP3:%.*]] = load <2 x i64>, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 32) to ptr), align 8
1579; CHECK-NEXT:    [[TMP4:%.*]] = load <2 x i64>, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 48) to ptr), align 8
1580; CHECK-NEXT:    call void @llvm.donothing()
1581; CHECK-NEXT:    [[TMP5:%.*]] = ptrtoint ptr [[P]] to i64
1582; CHECK-NEXT:    [[TMP6:%.*]] = xor i64 [[TMP5]], 193514046488576
1583; CHECK-NEXT:    [[TMP7:%.*]] = inttoptr i64 [[TMP6]] to ptr
1584; CHECK-NEXT:    call void @llvm.aarch64.neon.st4.v2i64.p0(<2 x i64> [[TMP2]], <2 x i64> splat (i64 -1), <2 x i64> [[TMP3]], <2 x i64> [[TMP4]], ptr [[TMP7]])
1585; CHECK-NEXT:    [[_MSCMP:%.*]] = icmp ne i64 [[TMP1]], 0
1586; CHECK-NEXT:    br i1 [[_MSCMP]], label [[TMP8:%.*]], label [[TMP9:%.*]], !prof [[PROF1]]
1587; CHECK:       8:
1588; CHECK-NEXT:    call void @__msan_warning_noreturn() #[[ATTR4]]
1589; CHECK-NEXT:    unreachable
1590; CHECK:       9:
1591; CHECK-NEXT:    call void @llvm.aarch64.neon.st4.v2i64.p0(<2 x i64> [[A]], <2 x i64> undef, <2 x i64> [[C]], <2 x i64> [[D]], ptr [[P]])
1592; CHECK-NEXT:    ret void
1593;
1594  call void @llvm.aarch64.neon.st4.v2i64.p0(<2 x i64> %A, <2 x i64> undef, <2 x i64> %C, <2 x i64> %D, ptr %P)
1595  ret void
1596}
1597
1598define void @st4_2d_undefC(<2 x i64> %A, <2 x i64> %B, <2 x i64> %C, <2 x i64> %D, ptr %P) nounwind sanitize_memory {
1599;
1600;
1601; CHECK-LABEL: define void @st4_2d_undefC(
1602; CHECK-SAME: <2 x i64> [[A:%.*]], <2 x i64> [[B:%.*]], <2 x i64> [[C:%.*]], <2 x i64> [[D:%.*]], ptr [[P:%.*]]) #[[ATTR0]] {
1603; CHECK-NEXT:    [[TMP1:%.*]] = load i64, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 64) to ptr), align 8
1604; CHECK-NEXT:    [[TMP2:%.*]] = load <2 x i64>, ptr @__msan_param_tls, align 8
1605; CHECK-NEXT:    [[TMP3:%.*]] = load <2 x i64>, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 16) to ptr), align 8
1606; CHECK-NEXT:    [[TMP4:%.*]] = load <2 x i64>, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 48) to ptr), align 8
1607; CHECK-NEXT:    call void @llvm.donothing()
1608; CHECK-NEXT:    [[TMP5:%.*]] = ptrtoint ptr [[P]] to i64
1609; CHECK-NEXT:    [[TMP6:%.*]] = xor i64 [[TMP5]], 193514046488576
1610; CHECK-NEXT:    [[TMP7:%.*]] = inttoptr i64 [[TMP6]] to ptr
1611; CHECK-NEXT:    call void @llvm.aarch64.neon.st4.v2i64.p0(<2 x i64> [[TMP2]], <2 x i64> [[TMP3]], <2 x i64> splat (i64 -1), <2 x i64> [[TMP4]], ptr [[TMP7]])
1612; CHECK-NEXT:    [[_MSCMP:%.*]] = icmp ne i64 [[TMP1]], 0
1613; CHECK-NEXT:    br i1 [[_MSCMP]], label [[TMP8:%.*]], label [[TMP9:%.*]], !prof [[PROF1]]
1614; CHECK:       8:
1615; CHECK-NEXT:    call void @__msan_warning_noreturn() #[[ATTR4]]
1616; CHECK-NEXT:    unreachable
1617; CHECK:       9:
1618; CHECK-NEXT:    call void @llvm.aarch64.neon.st4.v2i64.p0(<2 x i64> [[A]], <2 x i64> [[B]], <2 x i64> undef, <2 x i64> [[D]], ptr [[P]])
1619; CHECK-NEXT:    ret void
1620;
1621  call void @llvm.aarch64.neon.st4.v2i64.p0(<2 x i64> %A, <2 x i64> %B, <2 x i64> undef, <2 x i64> %D, ptr %P)
1622  ret void
1623}
1624
1625define void @st4_2d_undefD(<2 x i64> %A, <2 x i64> %B, <2 x i64> %C, <2 x i64> %D, ptr %P) nounwind sanitize_memory {
1626;
1627;
1628; CHECK-LABEL: define void @st4_2d_undefD(
1629; CHECK-SAME: <2 x i64> [[A:%.*]], <2 x i64> [[B:%.*]], <2 x i64> [[C:%.*]], <2 x i64> [[D:%.*]], ptr [[P:%.*]]) #[[ATTR0]] {
1630; CHECK-NEXT:    [[TMP1:%.*]] = load i64, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 64) to ptr), align 8
1631; CHECK-NEXT:    [[TMP2:%.*]] = load <2 x i64>, ptr @__msan_param_tls, align 8
1632; CHECK-NEXT:    [[TMP3:%.*]] = load <2 x i64>, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 16) to ptr), align 8
1633; CHECK-NEXT:    [[TMP4:%.*]] = load <2 x i64>, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 32) to ptr), align 8
1634; CHECK-NEXT:    call void @llvm.donothing()
1635; CHECK-NEXT:    [[TMP5:%.*]] = ptrtoint ptr [[P]] to i64
1636; CHECK-NEXT:    [[TMP6:%.*]] = xor i64 [[TMP5]], 193514046488576
1637; CHECK-NEXT:    [[TMP7:%.*]] = inttoptr i64 [[TMP6]] to ptr
1638; CHECK-NEXT:    call void @llvm.aarch64.neon.st4.v2i64.p0(<2 x i64> [[TMP2]], <2 x i64> [[TMP3]], <2 x i64> [[TMP4]], <2 x i64> splat (i64 -1), ptr [[TMP7]])
1639; CHECK-NEXT:    [[_MSCMP:%.*]] = icmp ne i64 [[TMP1]], 0
1640; CHECK-NEXT:    br i1 [[_MSCMP]], label [[TMP8:%.*]], label [[TMP9:%.*]], !prof [[PROF1]]
1641; CHECK:       8:
1642; CHECK-NEXT:    call void @__msan_warning_noreturn() #[[ATTR4]]
1643; CHECK-NEXT:    unreachable
1644; CHECK:       9:
1645; CHECK-NEXT:    call void @llvm.aarch64.neon.st4.v2i64.p0(<2 x i64> [[A]], <2 x i64> [[B]], <2 x i64> [[C]], <2 x i64> undef, ptr [[P]])
1646; CHECK-NEXT:    ret void
1647;
1648  call void @llvm.aarch64.neon.st4.v2i64.p0(<2 x i64> %A, <2 x i64> %B, <2 x i64> %C, <2 x i64> undef, ptr %P)
1649  ret void
1650}
1651
1652define void @st4_2d_undefAB(<2 x i64> %A, <2 x i64> %B, <2 x i64> %C, <2 x i64> %D, ptr %P) nounwind sanitize_memory {
1653;
1654;
1655; CHECK-LABEL: define void @st4_2d_undefAB(
1656; CHECK-SAME: <2 x i64> [[A:%.*]], <2 x i64> [[B:%.*]], <2 x i64> [[C:%.*]], <2 x i64> [[D:%.*]], ptr [[P:%.*]]) #[[ATTR0]] {
1657; CHECK-NEXT:    [[TMP1:%.*]] = load i64, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 64) to ptr), align 8
1658; CHECK-NEXT:    [[TMP2:%.*]] = load <2 x i64>, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 32) to ptr), align 8
1659; CHECK-NEXT:    [[TMP3:%.*]] = load <2 x i64>, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 48) to ptr), align 8
1660; CHECK-NEXT:    call void @llvm.donothing()
1661; CHECK-NEXT:    [[TMP4:%.*]] = ptrtoint ptr [[P]] to i64
1662; CHECK-NEXT:    [[TMP5:%.*]] = xor i64 [[TMP4]], 193514046488576
1663; CHECK-NEXT:    [[TMP6:%.*]] = inttoptr i64 [[TMP5]] to ptr
1664; CHECK-NEXT:    call void @llvm.aarch64.neon.st4.v2i64.p0(<2 x i64> splat (i64 -1), <2 x i64> splat (i64 -1), <2 x i64> [[TMP2]], <2 x i64> [[TMP3]], ptr [[TMP6]])
1665; CHECK-NEXT:    [[_MSCMP:%.*]] = icmp ne i64 [[TMP1]], 0
1666; CHECK-NEXT:    br i1 [[_MSCMP]], label [[TMP7:%.*]], label [[TMP8:%.*]], !prof [[PROF1]]
1667; CHECK:       7:
1668; CHECK-NEXT:    call void @__msan_warning_noreturn() #[[ATTR4]]
1669; CHECK-NEXT:    unreachable
1670; CHECK:       8:
1671; CHECK-NEXT:    call void @llvm.aarch64.neon.st4.v2i64.p0(<2 x i64> undef, <2 x i64> undef, <2 x i64> [[C]], <2 x i64> [[D]], ptr [[P]])
1672; CHECK-NEXT:    ret void
1673;
1674  call void @llvm.aarch64.neon.st4.v2i64.p0(<2 x i64> undef, <2 x i64> undef, <2 x i64> %C, <2 x i64> %D, ptr %P)
1675  ret void
1676}
1677
1678define void @st4_2d_undefAC(<2 x i64> %A, <2 x i64> %B, <2 x i64> %C, <2 x i64> %D, ptr %P) nounwind sanitize_memory {
1679;
1680;
1681; CHECK-LABEL: define void @st4_2d_undefAC(
1682; CHECK-SAME: <2 x i64> [[A:%.*]], <2 x i64> [[B:%.*]], <2 x i64> [[C:%.*]], <2 x i64> [[D:%.*]], ptr [[P:%.*]]) #[[ATTR0]] {
1683; CHECK-NEXT:    [[TMP1:%.*]] = load i64, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 64) to ptr), align 8
1684; CHECK-NEXT:    [[TMP2:%.*]] = load <2 x i64>, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 16) to ptr), align 8
1685; CHECK-NEXT:    [[TMP3:%.*]] = load <2 x i64>, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 48) to ptr), align 8
1686; CHECK-NEXT:    call void @llvm.donothing()
1687; CHECK-NEXT:    [[TMP4:%.*]] = ptrtoint ptr [[P]] to i64
1688; CHECK-NEXT:    [[TMP5:%.*]] = xor i64 [[TMP4]], 193514046488576
1689; CHECK-NEXT:    [[TMP6:%.*]] = inttoptr i64 [[TMP5]] to ptr
1690; CHECK-NEXT:    call void @llvm.aarch64.neon.st4.v2i64.p0(<2 x i64> splat (i64 -1), <2 x i64> [[TMP2]], <2 x i64> splat (i64 -1), <2 x i64> [[TMP3]], ptr [[TMP6]])
1691; CHECK-NEXT:    [[_MSCMP:%.*]] = icmp ne i64 [[TMP1]], 0
1692; CHECK-NEXT:    br i1 [[_MSCMP]], label [[TMP7:%.*]], label [[TMP8:%.*]], !prof [[PROF1]]
1693; CHECK:       7:
1694; CHECK-NEXT:    call void @__msan_warning_noreturn() #[[ATTR4]]
1695; CHECK-NEXT:    unreachable
1696; CHECK:       8:
1697; CHECK-NEXT:    call void @llvm.aarch64.neon.st4.v2i64.p0(<2 x i64> undef, <2 x i64> [[B]], <2 x i64> undef, <2 x i64> [[D]], ptr [[P]])
1698; CHECK-NEXT:    ret void
1699;
1700  call void @llvm.aarch64.neon.st4.v2i64.p0(<2 x i64> undef, <2 x i64> %B, <2 x i64> undef, <2 x i64> %D, ptr %P)
1701  ret void
1702}
1703
1704define void @st4_2d_undefAD(<2 x i64> %A, <2 x i64> %B, <2 x i64> %C, <2 x i64> %D, ptr %P) nounwind sanitize_memory {
1705;
1706;
1707; CHECK-LABEL: define void @st4_2d_undefAD(
1708; CHECK-SAME: <2 x i64> [[A:%.*]], <2 x i64> [[B:%.*]], <2 x i64> [[C:%.*]], <2 x i64> [[D:%.*]], ptr [[P:%.*]]) #[[ATTR0]] {
1709; CHECK-NEXT:    [[TMP1:%.*]] = load i64, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 64) to ptr), align 8
1710; CHECK-NEXT:    [[TMP2:%.*]] = load <2 x i64>, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 16) to ptr), align 8
1711; CHECK-NEXT:    [[TMP3:%.*]] = load <2 x i64>, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 32) to ptr), align 8
1712; CHECK-NEXT:    call void @llvm.donothing()
1713; CHECK-NEXT:    [[TMP4:%.*]] = ptrtoint ptr [[P]] to i64
1714; CHECK-NEXT:    [[TMP5:%.*]] = xor i64 [[TMP4]], 193514046488576
1715; CHECK-NEXT:    [[TMP6:%.*]] = inttoptr i64 [[TMP5]] to ptr
1716; CHECK-NEXT:    call void @llvm.aarch64.neon.st4.v2i64.p0(<2 x i64> splat (i64 -1), <2 x i64> [[TMP2]], <2 x i64> [[TMP3]], <2 x i64> splat (i64 -1), ptr [[TMP6]])
1717; CHECK-NEXT:    [[_MSCMP:%.*]] = icmp ne i64 [[TMP1]], 0
1718; CHECK-NEXT:    br i1 [[_MSCMP]], label [[TMP7:%.*]], label [[TMP8:%.*]], !prof [[PROF1]]
1719; CHECK:       7:
1720; CHECK-NEXT:    call void @__msan_warning_noreturn() #[[ATTR4]]
1721; CHECK-NEXT:    unreachable
1722; CHECK:       8:
1723; CHECK-NEXT:    call void @llvm.aarch64.neon.st4.v2i64.p0(<2 x i64> undef, <2 x i64> [[B]], <2 x i64> [[C]], <2 x i64> undef, ptr [[P]])
1724; CHECK-NEXT:    ret void
1725;
1726  call void @llvm.aarch64.neon.st4.v2i64.p0(<2 x i64> undef, <2 x i64> %B, <2 x i64> %C, <2 x i64> undef, ptr %P)
1727  ret void
1728}
1729
1730define void @st4_2d_undefBC(<2 x i64> %A, <2 x i64> %B, <2 x i64> %C, <2 x i64> %D, ptr %P) nounwind sanitize_memory {
1731;
1732;
1733; CHECK-LABEL: define void @st4_2d_undefBC(
1734; CHECK-SAME: <2 x i64> [[A:%.*]], <2 x i64> [[B:%.*]], <2 x i64> [[C:%.*]], <2 x i64> [[D:%.*]], ptr [[P:%.*]]) #[[ATTR0]] {
1735; CHECK-NEXT:    [[TMP1:%.*]] = load i64, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 64) to ptr), align 8
1736; CHECK-NEXT:    [[TMP2:%.*]] = load <2 x i64>, ptr @__msan_param_tls, align 8
1737; CHECK-NEXT:    [[TMP3:%.*]] = load <2 x i64>, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 48) to ptr), align 8
1738; CHECK-NEXT:    call void @llvm.donothing()
1739; CHECK-NEXT:    [[TMP4:%.*]] = ptrtoint ptr [[P]] to i64
1740; CHECK-NEXT:    [[TMP5:%.*]] = xor i64 [[TMP4]], 193514046488576
1741; CHECK-NEXT:    [[TMP6:%.*]] = inttoptr i64 [[TMP5]] to ptr
1742; CHECK-NEXT:    call void @llvm.aarch64.neon.st4.v2i64.p0(<2 x i64> [[TMP2]], <2 x i64> splat (i64 -1), <2 x i64> splat (i64 -1), <2 x i64> [[TMP3]], ptr [[TMP6]])
1743; CHECK-NEXT:    [[_MSCMP:%.*]] = icmp ne i64 [[TMP1]], 0
1744; CHECK-NEXT:    br i1 [[_MSCMP]], label [[TMP7:%.*]], label [[TMP8:%.*]], !prof [[PROF1]]
1745; CHECK:       7:
1746; CHECK-NEXT:    call void @__msan_warning_noreturn() #[[ATTR4]]
1747; CHECK-NEXT:    unreachable
1748; CHECK:       8:
1749; CHECK-NEXT:    call void @llvm.aarch64.neon.st4.v2i64.p0(<2 x i64> [[A]], <2 x i64> undef, <2 x i64> undef, <2 x i64> [[D]], ptr [[P]])
1750; CHECK-NEXT:    ret void
1751;
1752  call void @llvm.aarch64.neon.st4.v2i64.p0(<2 x i64> %A, <2 x i64> undef, <2 x i64> undef, <2 x i64> %D, ptr %P)
1753  ret void
1754}
1755
1756define void @st4_2d_undefBD(<2 x i64> %A, <2 x i64> %B, <2 x i64> %C, <2 x i64> %D, ptr %P) nounwind sanitize_memory {
1757;
1758;
1759; CHECK-LABEL: define void @st4_2d_undefBD(
1760; CHECK-SAME: <2 x i64> [[A:%.*]], <2 x i64> [[B:%.*]], <2 x i64> [[C:%.*]], <2 x i64> [[D:%.*]], ptr [[P:%.*]]) #[[ATTR0]] {
1761; CHECK-NEXT:    [[TMP1:%.*]] = load i64, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 64) to ptr), align 8
1762; CHECK-NEXT:    [[TMP2:%.*]] = load <2 x i64>, ptr @__msan_param_tls, align 8
1763; CHECK-NEXT:    [[TMP3:%.*]] = load <2 x i64>, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 32) to ptr), align 8
1764; CHECK-NEXT:    call void @llvm.donothing()
1765; CHECK-NEXT:    [[TMP4:%.*]] = ptrtoint ptr [[P]] to i64
1766; CHECK-NEXT:    [[TMP5:%.*]] = xor i64 [[TMP4]], 193514046488576
1767; CHECK-NEXT:    [[TMP6:%.*]] = inttoptr i64 [[TMP5]] to ptr
1768; CHECK-NEXT:    call void @llvm.aarch64.neon.st4.v2i64.p0(<2 x i64> [[TMP2]], <2 x i64> splat (i64 -1), <2 x i64> [[TMP3]], <2 x i64> splat (i64 -1), ptr [[TMP6]])
1769; CHECK-NEXT:    [[_MSCMP:%.*]] = icmp ne i64 [[TMP1]], 0
1770; CHECK-NEXT:    br i1 [[_MSCMP]], label [[TMP7:%.*]], label [[TMP8:%.*]], !prof [[PROF1]]
1771; CHECK:       7:
1772; CHECK-NEXT:    call void @__msan_warning_noreturn() #[[ATTR4]]
1773; CHECK-NEXT:    unreachable
1774; CHECK:       8:
1775; CHECK-NEXT:    call void @llvm.aarch64.neon.st4.v2i64.p0(<2 x i64> [[A]], <2 x i64> undef, <2 x i64> [[C]], <2 x i64> undef, ptr [[P]])
1776; CHECK-NEXT:    ret void
1777;
1778  call void @llvm.aarch64.neon.st4.v2i64.p0(<2 x i64> %A, <2 x i64> undef, <2 x i64> %C, <2 x i64> undef, ptr %P)
1779  ret void
1780}
1781
1782define void @st4_2d_undefCD(<2 x i64> %A, <2 x i64> %B, <2 x i64> %C, <2 x i64> %D, ptr %P) nounwind sanitize_memory {
1783;
1784;
1785; CHECK-LABEL: define void @st4_2d_undefCD(
1786; CHECK-SAME: <2 x i64> [[A:%.*]], <2 x i64> [[B:%.*]], <2 x i64> [[C:%.*]], <2 x i64> [[D:%.*]], ptr [[P:%.*]]) #[[ATTR0]] {
1787; CHECK-NEXT:    [[TMP1:%.*]] = load i64, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 64) to ptr), align 8
1788; CHECK-NEXT:    [[TMP2:%.*]] = load <2 x i64>, ptr @__msan_param_tls, align 8
1789; CHECK-NEXT:    [[TMP3:%.*]] = load <2 x i64>, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 16) to ptr), align 8
1790; CHECK-NEXT:    call void @llvm.donothing()
1791; CHECK-NEXT:    [[TMP4:%.*]] = ptrtoint ptr [[P]] to i64
1792; CHECK-NEXT:    [[TMP5:%.*]] = xor i64 [[TMP4]], 193514046488576
1793; CHECK-NEXT:    [[TMP6:%.*]] = inttoptr i64 [[TMP5]] to ptr
1794; CHECK-NEXT:    call void @llvm.aarch64.neon.st4.v2i64.p0(<2 x i64> [[TMP2]], <2 x i64> [[TMP3]], <2 x i64> splat (i64 -1), <2 x i64> splat (i64 -1), ptr [[TMP6]])
1795; CHECK-NEXT:    [[_MSCMP:%.*]] = icmp ne i64 [[TMP1]], 0
1796; CHECK-NEXT:    br i1 [[_MSCMP]], label [[TMP7:%.*]], label [[TMP8:%.*]], !prof [[PROF1]]
1797; CHECK:       7:
1798; CHECK-NEXT:    call void @__msan_warning_noreturn() #[[ATTR4]]
1799; CHECK-NEXT:    unreachable
1800; CHECK:       8:
1801; CHECK-NEXT:    call void @llvm.aarch64.neon.st4.v2i64.p0(<2 x i64> [[A]], <2 x i64> [[B]], <2 x i64> undef, <2 x i64> undef, ptr [[P]])
1802; CHECK-NEXT:    ret void
1803;
1804  call void @llvm.aarch64.neon.st4.v2i64.p0(<2 x i64> %A, <2 x i64> %B, <2 x i64> undef, <2 x i64> undef, ptr %P)
1805  ret void
1806}
1807
1808define void @st4_2d_undefABC(<2 x i64> %A, <2 x i64> %B, <2 x i64> %C, <2 x i64> %D, ptr %P) nounwind sanitize_memory {
1809;
1810;
1811; CHECK-LABEL: define void @st4_2d_undefABC(
1812; CHECK-SAME: <2 x i64> [[A:%.*]], <2 x i64> [[B:%.*]], <2 x i64> [[C:%.*]], <2 x i64> [[D:%.*]], ptr [[P:%.*]]) #[[ATTR0]] {
1813; CHECK-NEXT:    [[TMP1:%.*]] = load i64, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 64) to ptr), align 8
1814; CHECK-NEXT:    [[TMP2:%.*]] = load <2 x i64>, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 48) to ptr), align 8
1815; CHECK-NEXT:    call void @llvm.donothing()
1816; CHECK-NEXT:    [[TMP3:%.*]] = ptrtoint ptr [[P]] to i64
1817; CHECK-NEXT:    [[TMP4:%.*]] = xor i64 [[TMP3]], 193514046488576
1818; CHECK-NEXT:    [[TMP5:%.*]] = inttoptr i64 [[TMP4]] to ptr
1819; CHECK-NEXT:    call void @llvm.aarch64.neon.st4.v2i64.p0(<2 x i64> splat (i64 -1), <2 x i64> splat (i64 -1), <2 x i64> splat (i64 -1), <2 x i64> [[TMP2]], ptr [[TMP5]])
1820; CHECK-NEXT:    [[_MSCMP:%.*]] = icmp ne i64 [[TMP1]], 0
1821; CHECK-NEXT:    br i1 [[_MSCMP]], label [[TMP6:%.*]], label [[TMP7:%.*]], !prof [[PROF1]]
1822; CHECK:       6:
1823; CHECK-NEXT:    call void @__msan_warning_noreturn() #[[ATTR4]]
1824; CHECK-NEXT:    unreachable
1825; CHECK:       7:
1826; CHECK-NEXT:    call void @llvm.aarch64.neon.st4.v2i64.p0(<2 x i64> undef, <2 x i64> undef, <2 x i64> undef, <2 x i64> [[D]], ptr [[P]])
1827; CHECK-NEXT:    ret void
1828;
1829  call void @llvm.aarch64.neon.st4.v2i64.p0(<2 x i64> undef, <2 x i64> undef, <2 x i64> undef, <2 x i64> %D, ptr %P)
1830  ret void
1831}
1832
1833define void @st4_2d_undefABD(<2 x i64> %A, <2 x i64> %B, <2 x i64> %C, <2 x i64> %D, ptr %P) nounwind sanitize_memory {
1834;
1835;
1836; CHECK-LABEL: define void @st4_2d_undefABD(
1837; CHECK-SAME: <2 x i64> [[A:%.*]], <2 x i64> [[B:%.*]], <2 x i64> [[C:%.*]], <2 x i64> [[D:%.*]], ptr [[P:%.*]]) #[[ATTR0]] {
1838; CHECK-NEXT:    [[TMP1:%.*]] = load i64, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 64) to ptr), align 8
1839; CHECK-NEXT:    [[TMP2:%.*]] = load <2 x i64>, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 32) to ptr), align 8
1840; CHECK-NEXT:    call void @llvm.donothing()
1841; CHECK-NEXT:    [[TMP3:%.*]] = ptrtoint ptr [[P]] to i64
1842; CHECK-NEXT:    [[TMP4:%.*]] = xor i64 [[TMP3]], 193514046488576
1843; CHECK-NEXT:    [[TMP5:%.*]] = inttoptr i64 [[TMP4]] to ptr
1844; CHECK-NEXT:    call void @llvm.aarch64.neon.st4.v2i64.p0(<2 x i64> splat (i64 -1), <2 x i64> splat (i64 -1), <2 x i64> [[TMP2]], <2 x i64> splat (i64 -1), ptr [[TMP5]])
1845; CHECK-NEXT:    [[_MSCMP:%.*]] = icmp ne i64 [[TMP1]], 0
1846; CHECK-NEXT:    br i1 [[_MSCMP]], label [[TMP6:%.*]], label [[TMP7:%.*]], !prof [[PROF1]]
1847; CHECK:       6:
1848; CHECK-NEXT:    call void @__msan_warning_noreturn() #[[ATTR4]]
1849; CHECK-NEXT:    unreachable
1850; CHECK:       7:
1851; CHECK-NEXT:    call void @llvm.aarch64.neon.st4.v2i64.p0(<2 x i64> undef, <2 x i64> undef, <2 x i64> [[C]], <2 x i64> undef, ptr [[P]])
1852; CHECK-NEXT:    ret void
1853;
1854  call void @llvm.aarch64.neon.st4.v2i64.p0(<2 x i64> undef, <2 x i64> undef, <2 x i64> %C, <2 x i64> undef, ptr %P)
1855  ret void
1856}
1857
1858define void @st4_2d_undefACD(<2 x i64> %A, <2 x i64> %B, <2 x i64> %C, <2 x i64> %D, ptr %P) nounwind sanitize_memory {
1859;
1860;
1861; CHECK-LABEL: define void @st4_2d_undefACD(
1862; CHECK-SAME: <2 x i64> [[A:%.*]], <2 x i64> [[B:%.*]], <2 x i64> [[C:%.*]], <2 x i64> [[D:%.*]], ptr [[P:%.*]]) #[[ATTR0]] {
1863; CHECK-NEXT:    [[TMP1:%.*]] = load i64, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 64) to ptr), align 8
1864; CHECK-NEXT:    [[TMP2:%.*]] = load <2 x i64>, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 16) to ptr), align 8
1865; CHECK-NEXT:    call void @llvm.donothing()
1866; CHECK-NEXT:    [[TMP3:%.*]] = ptrtoint ptr [[P]] to i64
1867; CHECK-NEXT:    [[TMP4:%.*]] = xor i64 [[TMP3]], 193514046488576
1868; CHECK-NEXT:    [[TMP5:%.*]] = inttoptr i64 [[TMP4]] to ptr
1869; CHECK-NEXT:    call void @llvm.aarch64.neon.st4.v2i64.p0(<2 x i64> splat (i64 -1), <2 x i64> [[TMP2]], <2 x i64> splat (i64 -1), <2 x i64> splat (i64 -1), ptr [[TMP5]])
1870; CHECK-NEXT:    [[_MSCMP:%.*]] = icmp ne i64 [[TMP1]], 0
1871; CHECK-NEXT:    br i1 [[_MSCMP]], label [[TMP6:%.*]], label [[TMP7:%.*]], !prof [[PROF1]]
1872; CHECK:       6:
1873; CHECK-NEXT:    call void @__msan_warning_noreturn() #[[ATTR4]]
1874; CHECK-NEXT:    unreachable
1875; CHECK:       7:
1876; CHECK-NEXT:    call void @llvm.aarch64.neon.st4.v2i64.p0(<2 x i64> undef, <2 x i64> [[B]], <2 x i64> undef, <2 x i64> undef, ptr [[P]])
1877; CHECK-NEXT:    ret void
1878;
1879  call void @llvm.aarch64.neon.st4.v2i64.p0(<2 x i64> undef, <2 x i64> %B, <2 x i64> undef, <2 x i64> undef, ptr %P)
1880  ret void
1881}
1882
1883define void @st4_2d_undefBCD(<2 x i64> %A, <2 x i64> %B, <2 x i64> %C, <2 x i64> %D, ptr %P) nounwind sanitize_memory {
1884;
1885;
1886; CHECK-LABEL: define void @st4_2d_undefBCD(
1887; CHECK-SAME: <2 x i64> [[A:%.*]], <2 x i64> [[B:%.*]], <2 x i64> [[C:%.*]], <2 x i64> [[D:%.*]], ptr [[P:%.*]]) #[[ATTR0]] {
1888; CHECK-NEXT:    [[TMP1:%.*]] = load i64, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 64) to ptr), align 8
1889; CHECK-NEXT:    [[TMP2:%.*]] = load <2 x i64>, ptr @__msan_param_tls, align 8
1890; CHECK-NEXT:    call void @llvm.donothing()
1891; CHECK-NEXT:    [[TMP3:%.*]] = ptrtoint ptr [[P]] to i64
1892; CHECK-NEXT:    [[TMP4:%.*]] = xor i64 [[TMP3]], 193514046488576
1893; CHECK-NEXT:    [[TMP5:%.*]] = inttoptr i64 [[TMP4]] to ptr
1894; CHECK-NEXT:    call void @llvm.aarch64.neon.st4.v2i64.p0(<2 x i64> [[TMP2]], <2 x i64> splat (i64 -1), <2 x i64> splat (i64 -1), <2 x i64> splat (i64 -1), ptr [[TMP5]])
1895; CHECK-NEXT:    [[_MSCMP:%.*]] = icmp ne i64 [[TMP1]], 0
1896; CHECK-NEXT:    br i1 [[_MSCMP]], label [[TMP6:%.*]], label [[TMP7:%.*]], !prof [[PROF1]]
1897; CHECK:       6:
1898; CHECK-NEXT:    call void @__msan_warning_noreturn() #[[ATTR4]]
1899; CHECK-NEXT:    unreachable
1900; CHECK:       7:
1901; CHECK-NEXT:    call void @llvm.aarch64.neon.st4.v2i64.p0(<2 x i64> [[A]], <2 x i64> undef, <2 x i64> undef, <2 x i64> undef, ptr [[P]])
1902; CHECK-NEXT:    ret void
1903;
1904  call void @llvm.aarch64.neon.st4.v2i64.p0(<2 x i64> %A, <2 x i64> undef, <2 x i64> undef, <2 x i64> undef, ptr %P)
1905  ret void
1906}
1907
1908define void @st4_2d_undefABCD(<2 x i64> %A, <2 x i64> %B, <2 x i64> %C, <2 x i64> %D, ptr %P) nounwind sanitize_memory {
1909;
1910;
1911; CHECK-LABEL: define void @st4_2d_undefABCD(
1912; CHECK-SAME: <2 x i64> [[A:%.*]], <2 x i64> [[B:%.*]], <2 x i64> [[C:%.*]], <2 x i64> [[D:%.*]], ptr [[P:%.*]]) #[[ATTR0]] {
1913; CHECK-NEXT:    [[TMP1:%.*]] = load i64, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 64) to ptr), align 8
1914; CHECK-NEXT:    call void @llvm.donothing()
1915; CHECK-NEXT:    [[TMP2:%.*]] = ptrtoint ptr [[P]] to i64
1916; CHECK-NEXT:    [[TMP3:%.*]] = xor i64 [[TMP2]], 193514046488576
1917; CHECK-NEXT:    [[TMP4:%.*]] = inttoptr i64 [[TMP3]] to ptr
1918; CHECK-NEXT:    call void @llvm.aarch64.neon.st4.v2i64.p0(<2 x i64> splat (i64 -1), <2 x i64> splat (i64 -1), <2 x i64> splat (i64 -1), <2 x i64> splat (i64 -1), ptr [[TMP4]])
1919; CHECK-NEXT:    [[_MSCMP:%.*]] = icmp ne i64 [[TMP1]], 0
1920; CHECK-NEXT:    br i1 [[_MSCMP]], label [[TMP5:%.*]], label [[TMP6:%.*]], !prof [[PROF1]]
1921; CHECK:       5:
1922; CHECK-NEXT:    call void @__msan_warning_noreturn() #[[ATTR4]]
1923; CHECK-NEXT:    unreachable
1924; CHECK:       6:
1925; CHECK-NEXT:    call void @llvm.aarch64.neon.st4.v2i64.p0(<2 x i64> undef, <2 x i64> undef, <2 x i64> undef, <2 x i64> undef, ptr [[P]])
1926; CHECK-NEXT:    ret void
1927;
1928  call void @llvm.aarch64.neon.st4.v2i64.p0(<2 x i64> undef, <2 x i64> undef, <2 x i64> undef, <2 x i64> undef, ptr %P)
1929  ret void
1930}
1931;.
1932; CHECK: [[PROF1]] = !{!"branch_weights", i32 1, i32 1048575}
1933;.
1934