1; RUN: opt < %s -sroa -S | FileCheck %s
2target datalayout = "e-p:64:64:64-p1:16:16:16-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:32:64-f32:32:32-f64:64:64-f80:128-v64:64:64-v128:128:128-a0:0:64-n8:16:32:64"
3
4declare void @llvm.memcpy.p0i8.p0i8.i32(i8* nocapture, i8* nocapture, i32, i32, i1) nounwind
5declare void @llvm.memset.p0i8.i32(i8* nocapture, i8, i32, i32, i1) nounwind
6
7define void @no_split_on_non_byte_width(i32) {
8; This tests that allocas are not split into slices that are not byte width multiple
9  %arg = alloca i32 , align 8
10  store i32 %0, i32* %arg
11  br label %load_i32
12
13load_i32:
14; CHECK-LABEL: load_i32:
15; CHECK-NOT: bitcast {{.*}} to i1
16; CHECK-NOT: zext i1
17  %r0 = load i32, i32* %arg
18  br label %load_i1
19
20load_i1:
21; CHECK-LABEL: load_i1:
22; CHECK: bitcast {{.*}} to i1
23  %p1 = bitcast i32* %arg to i1*
24  %t1 = load i1, i1* %p1
25  ret void
26}
27
28; PR18726: Check that we use memcpy and memset to fill out padding when we have
29; a slice with a simple single type whose store size is smaller than the slice
30; size.
31
32%union.Foo = type { x86_fp80, i64, i64 }
33
34@foo_copy_source = external constant %union.Foo
35@i64_sink = global i64 0
36
37define void @memcpy_fp80_padding() {
38  %x = alloca %union.Foo
39
40  ; Copy from a global.
41  %x_i8 = bitcast %union.Foo* %x to i8*
42  call void @llvm.memcpy.p0i8.p0i8.i32(i8* %x_i8, i8* bitcast (%union.Foo* @foo_copy_source to i8*), i32 32, i32 16, i1 false)
43
44  ; Access a slice of the alloca to trigger SROA.
45  %mid_p = getelementptr %union.Foo, %union.Foo* %x, i32 0, i32 1
46  %elt = load i64, i64* %mid_p
47  store i64 %elt, i64* @i64_sink
48  ret void
49}
50; CHECK-LABEL: define void @memcpy_fp80_padding
51; CHECK: alloca x86_fp80
52; CHECK: call void @llvm.memcpy.p0i8.p0i8.i32
53; CHECK: load i64, i64* getelementptr inbounds (%union.Foo, %union.Foo* @foo_copy_source, i64 0, i32 1)
54; CHECK: load i64, i64* getelementptr inbounds (%union.Foo, %union.Foo* @foo_copy_source, i64 0, i32 2)
55
56define void @memset_fp80_padding() {
57  %x = alloca %union.Foo
58
59  ; Set to all ones.
60  %x_i8 = bitcast %union.Foo* %x to i8*
61  call void @llvm.memset.p0i8.i32(i8* %x_i8, i8 -1, i32 32, i32 16, i1 false)
62
63  ; Access a slice of the alloca to trigger SROA.
64  %mid_p = getelementptr %union.Foo, %union.Foo* %x, i32 0, i32 1
65  %elt = load i64, i64* %mid_p
66  store i64 %elt, i64* @i64_sink
67  ret void
68}
69; CHECK-LABEL: define void @memset_fp80_padding
70; CHECK: alloca x86_fp80
71; CHECK: call void @llvm.memset.p0i8.i32(i8* %{{.*}}, i8 -1, i32 16, i32 16, i1 false)
72; CHECK: store i64 -1, i64* @i64_sink
73
74%S.vec3float = type { float, float, float }
75%U.vec3float = type { <4 x float> }
76
77declare i32 @memcpy_vec3float_helper(%S.vec3float*)
78
79define i32 @memcpy_vec3float_widening(%S.vec3float* %x) {
80; CHECK-LABEL: @memcpy_vec3float_widening(
81; PR18726: Check that SROA does not rewrite a 12-byte memcpy into a 16-byte
82; vector store, hence accidentally putting gibberish onto the stack.
83entry:
84  ; Create a temporary variable %tmp1 and copy %x[0] into it
85  %tmp1 = alloca %S.vec3float, align 4
86  %0 = bitcast %S.vec3float* %tmp1 to i8*
87  %1 = bitcast %S.vec3float* %x to i8*
88  call void @llvm.memcpy.p0i8.p0i8.i32(i8* %0, i8* %1, i32 12, i32 4, i1 false)
89
90  ; The following block does nothing; but appears to confuse SROA
91  %unused1 = bitcast %S.vec3float* %tmp1 to %U.vec3float*
92  %unused2 = getelementptr inbounds %U.vec3float, %U.vec3float* %unused1, i32 0, i32 0
93  %unused3 = load <4 x float>, <4 x float>* %unused2, align 1
94
95  ; Create a second temporary and copy %tmp1 into it
96  %tmp2 = alloca %S.vec3float, align 4
97  %2 = bitcast %S.vec3float* %tmp2 to i8*
98  %3 = bitcast %S.vec3float* %tmp1 to i8*
99; CHECK: alloca
100; CHECK-NOT: store <4 x float>
101  call void @llvm.memcpy.p0i8.p0i8.i32(i8* %2, i8* %3, i32 12, i32 4, i1 false)
102
103  %result = call i32 @memcpy_vec3float_helper(%S.vec3float* %tmp2)
104  ret i32 %result
105; CHECK: ret i32 %result
106}
107