xref: /llvm-project/llvm/test/Transforms/InstCombine/load-store-forward.ll (revision f104cc38f45ebbca065e78959c23d09f16b65bd5)
1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
2; RUN: opt -S -instcombine < %s | FileCheck %s --check-prefixes=CHECK,LITTLE
3; RUN: opt -S -instcombine -data-layout="E" < %s | FileCheck %s --check-prefixes=CHECK,BIG
4
5define i8 @load_smaller_int(i16* %p) {
6; LITTLE-LABEL: @load_smaller_int(
7; LITTLE-NEXT:    store i16 258, i16* [[P:%.*]], align 2
8; LITTLE-NEXT:    ret i8 2
9;
10; BIG-LABEL: @load_smaller_int(
11; BIG-NEXT:    store i16 258, i16* [[P:%.*]], align 2
12; BIG-NEXT:    ret i8 1
13;
14  store i16 258, i16* %p
15  %p2 = bitcast i16* %p to i8*
16  %load = load i8, i8* %p2
17  ret i8 %load
18}
19
20; This case can *not* be forwarded, as we only see part of the stored value.
21define i32 @load_larger_int(i16* %p) {
22; CHECK-LABEL: @load_larger_int(
23; CHECK-NEXT:    store i16 258, i16* [[P:%.*]], align 2
24; CHECK-NEXT:    [[P2:%.*]] = bitcast i16* [[P]] to i32*
25; CHECK-NEXT:    [[LOAD:%.*]] = load i32, i32* [[P2]], align 4
26; CHECK-NEXT:    ret i32 [[LOAD]]
27;
28  store i16 258, i16* %p
29  %p2 = bitcast i16* %p to i32*
30  %load = load i32, i32* %p2
31  ret i32 %load
32}
33
34define i32 @vec_store_load_first(i32* %p) {
35; CHECK-LABEL: @vec_store_load_first(
36; CHECK-NEXT:    [[P2:%.*]] = bitcast i32* [[P:%.*]] to <2 x i32>*
37; CHECK-NEXT:    store <2 x i32> <i32 1, i32 2>, <2 x i32>* [[P2]], align 8
38; CHECK-NEXT:    ret i32 1
39;
40  %p2 = bitcast i32* %p to <2 x i32>*
41  store <2 x i32> <i32 1, i32 2>, <2 x i32>* %p2
42  %load = load i32, i32* %p
43  ret i32 %load
44}
45
46define i17 @vec_store_load_first_odd_size(i17* %p) {
47; CHECK-LABEL: @vec_store_load_first_odd_size(
48; CHECK-NEXT:    [[P2:%.*]] = bitcast i17* [[P:%.*]] to <2 x i17>*
49; CHECK-NEXT:    store <2 x i17> <i17 1, i17 2>, <2 x i17>* [[P2]], align 8
50; CHECK-NEXT:    [[LOAD:%.*]] = load i17, i17* [[P]], align 4
51; CHECK-NEXT:    ret i17 [[LOAD]]
52;
53  %p2 = bitcast i17* %p to <2 x i17>*
54  store <2 x i17> <i17 1, i17 2>, <2 x i17>* %p2
55  %load = load i17, i17* %p
56  ret i17 %load
57}
58
59define i32 @vec_store_load_first_constexpr(i32* %p) {
60; CHECK-LABEL: @vec_store_load_first_constexpr(
61; CHECK-NEXT:    [[P2:%.*]] = bitcast i32* [[P:%.*]] to <2 x i32>*
62; CHECK-NEXT:    store <2 x i32> bitcast (i64 ptrtoint (i32 (i32*)* @vec_store_load_first to i64) to <2 x i32>), <2 x i32>* [[P2]], align 8
63; CHECK-NEXT:    [[LOAD:%.*]] = load i32, i32* [[P]], align 4
64; CHECK-NEXT:    ret i32 [[LOAD]]
65;
66  %p2 = bitcast i32* %p to <2 x i32>*
67  store <2 x i32> bitcast (i64 ptrtoint (i32 (i32*)* @vec_store_load_first to i64) to <2 x i32>), <2 x i32>* %p2, align 8
68  %load = load i32, i32* %p, align 4
69  ret i32 %load
70}
71
72define i32 @vec_store_load_second(i32* %p) {
73; CHECK-LABEL: @vec_store_load_second(
74; CHECK-NEXT:    [[P2:%.*]] = bitcast i32* [[P:%.*]] to <2 x i32>*
75; CHECK-NEXT:    store <2 x i32> <i32 1, i32 2>, <2 x i32>* [[P2]], align 8
76; CHECK-NEXT:    [[P3:%.*]] = getelementptr i32, i32* [[P]], i64 1
77; CHECK-NEXT:    [[LOAD:%.*]] = load i32, i32* [[P3]], align 4
78; CHECK-NEXT:    ret i32 [[LOAD]]
79;
80  %p2 = bitcast i32* %p to <2 x i32>*
81  store <2 x i32> <i32 1, i32 2>, <2 x i32>* %p2
82  %p3 = getelementptr i32, i32* %p, i64 1
83  %load = load i32, i32* %p3
84  ret i32 %load
85}
86
87define i64 @vec_store_load_whole(i32* %p) {
88; LITTLE-LABEL: @vec_store_load_whole(
89; LITTLE-NEXT:    [[P2:%.*]] = bitcast i32* [[P:%.*]] to <2 x i32>*
90; LITTLE-NEXT:    store <2 x i32> <i32 1, i32 2>, <2 x i32>* [[P2]], align 8
91; LITTLE-NEXT:    ret i64 8589934593
92;
93; BIG-LABEL: @vec_store_load_whole(
94; BIG-NEXT:    [[P2:%.*]] = bitcast i32* [[P:%.*]] to <2 x i32>*
95; BIG-NEXT:    store <2 x i32> <i32 1, i32 2>, <2 x i32>* [[P2]], align 8
96; BIG-NEXT:    ret i64 4294967298
97;
98  %p2 = bitcast i32* %p to <2 x i32>*
99  store <2 x i32> <i32 1, i32 2>, <2 x i32>* %p2
100  %p3 = bitcast i32* %p to i64*
101  %load = load i64, i64* %p3
102  ret i64 %load
103}
104
105define i32 @vec_store_load_overlap(i32* %p) {
106; CHECK-LABEL: @vec_store_load_overlap(
107; CHECK-NEXT:    [[P2:%.*]] = bitcast i32* [[P:%.*]] to <2 x i32>*
108; CHECK-NEXT:    store <2 x i32> <i32 1, i32 2>, <2 x i32>* [[P2]], align 8
109; CHECK-NEXT:    [[P3:%.*]] = bitcast i32* [[P]] to i8*
110; CHECK-NEXT:    [[P4:%.*]] = getelementptr i8, i8* [[P3]], i64 2
111; CHECK-NEXT:    [[P5:%.*]] = bitcast i8* [[P4]] to i32*
112; CHECK-NEXT:    [[LOAD:%.*]] = load i32, i32* [[P5]], align 2
113; CHECK-NEXT:    ret i32 [[LOAD]]
114;
115  %p2 = bitcast i32* %p to <2 x i32>*
116  store <2 x i32> <i32 1, i32 2>, <2 x i32>* %p2
117  %p3 = bitcast i32* %p to i8*
118  %p4 = getelementptr i8, i8* %p3, i64 2
119  %p5 = bitcast i8* %p4 to i32*
120  %load = load i32, i32* %p5, align 2
121  ret i32 %load
122}
123