• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1; RUN: llc < %s -mtriple=arm64-eabi -aarch64-neon-syntax=apple -mcpu=cyclone -mattr=+slow-misaligned-128store | FileCheck %s
2%struct.X = type <{ i32, i64, i64 }>
3
4define void @foo1(i32* %p, i64 %val) nounwind {
5; CHECK-LABEL: foo1:
6; CHECK: 	stur	w1, [x0, #-4]
7; CHECK-NEXT: 	ret
8  %tmp1 = trunc i64 %val to i32
9  %ptr = getelementptr inbounds i32, i32* %p, i64 -1
10  store i32 %tmp1, i32* %ptr, align 4
11  ret void
12}
13define void @foo2(i16* %p, i64 %val) nounwind {
14; CHECK-LABEL: foo2:
15; CHECK: 	sturh	w1, [x0, #-2]
16; CHECK-NEXT: 	ret
17  %tmp1 = trunc i64 %val to i16
18  %ptr = getelementptr inbounds i16, i16* %p, i64 -1
19  store i16 %tmp1, i16* %ptr, align 2
20  ret void
21}
22define void @foo3(i8* %p, i64 %val) nounwind {
23; CHECK-LABEL: foo3:
24; CHECK: 	sturb	w1, [x0, #-1]
25; CHECK-NEXT: 	ret
26  %tmp1 = trunc i64 %val to i8
27  %ptr = getelementptr inbounds i8, i8* %p, i64 -1
28  store i8 %tmp1, i8* %ptr, align 1
29  ret void
30}
31define void @foo4(i16* %p, i32 %val) nounwind {
32; CHECK-LABEL: foo4:
33; CHECK: 	sturh	w1, [x0, #-2]
34; CHECK-NEXT: 	ret
35  %tmp1 = trunc i32 %val to i16
36  %ptr = getelementptr inbounds i16, i16* %p, i32 -1
37  store i16 %tmp1, i16* %ptr, align 2
38  ret void
39}
40define void @foo5(i8* %p, i32 %val) nounwind {
41; CHECK-LABEL: foo5:
42; CHECK: 	sturb	w1, [x0, #-1]
43; CHECK-NEXT: 	ret
44  %tmp1 = trunc i32 %val to i8
45  %ptr = getelementptr inbounds i8, i8* %p, i32 -1
46  store i8 %tmp1, i8* %ptr, align 1
47  ret void
48}
49
50define void @foo(%struct.X* nocapture %p) nounwind optsize ssp {
51; CHECK-LABEL: foo:
52; CHECK-NOT: str
53; CHECK: stur    xzr, [x0, #12]
54; CHECK-NEXT: stur    xzr, [x0, #4]
55; CHECK-NEXT: ret
56  %B = getelementptr inbounds %struct.X, %struct.X* %p, i64 0, i32 1
57  %val = bitcast i64* %B to i8*
58  call void @llvm.memset.p0i8.i64(i8* %val, i8 0, i64 16, i1 false)
59  ret void
60}
61
62declare void @llvm.memset.p0i8.i64(i8* nocapture, i8, i64, i1) nounwind
63
64; Unaligned 16b stores are split into 8b stores for performance.
65; radar://15424193
66
67; CHECK-LABEL: unaligned:
68; CHECK-NOT: str q0
69; CHECK: str     d[[REG:[0-9]+]], [x0]
70; CHECK: ext.16b v[[REG2:[0-9]+]], v[[REG]], v[[REG]], #8
71; CHECK: str     d[[REG2]], [x0, #8]
72define void @unaligned(<4 x i32>* %p, <4 x i32> %v) nounwind {
73  store <4 x i32> %v, <4 x i32>* %p, align 4
74  ret void
75}
76
77; CHECK-LABEL: aligned:
78; CHECK: str q0
79define void @aligned(<4 x i32>* %p, <4 x i32> %v) nounwind {
80  store <4 x i32> %v, <4 x i32>* %p
81  ret void
82}
83
84; Don't split one and two byte aligned stores.
85; radar://16349308
86
87; CHECK-LABEL: twobytealign:
88; CHECK: str q0
89define void @twobytealign(<4 x i32>* %p, <4 x i32> %v) nounwind {
90  store <4 x i32> %v, <4 x i32>* %p, align 2
91  ret void
92}
93; CHECK-LABEL: onebytealign:
94; CHECK: str q0
95define void @onebytealign(<4 x i32>* %p, <4 x i32> %v) nounwind {
96  store <4 x i32> %v, <4 x i32>* %p, align 1
97  ret void
98}
99