1; RUN: llc -march=amdgcn -mattr=-promote-alloca -amdgpu-sroa=0 -verify-machineinstrs < %s | FileCheck -check-prefix=GCN %s
2
3; GCN-LABEL: {{^}}stored_fi_to_lds:
4; GCN: s_load_dword [[LDSPTR:s[0-9]+]]
5; GCN: v_mov_b32_e32 [[ZERO1:v[0-9]+]], 0{{$}}
6; GCN: buffer_store_dword v{{[0-9]+}}, [[ZERO1]]
7; GCN: v_mov_b32_e32 [[ZERO0:v[0-9]+]], 0{{$}}
8; GCN: v_mov_b32_e32 [[VLDSPTR:v[0-9]+]], [[LDSPTR]]
9; GCN: ds_write_b32  [[VLDSPTR]], [[ZERO0]]
10define void @stored_fi_to_lds(float* addrspace(3)* %ptr) #0 {
11  %tmp = alloca float
12  store float 4.0, float *%tmp
13  store float* %tmp, float* addrspace(3)* %ptr
14  ret void
15}
16
17; Offset is applied
18; GCN-LABEL: {{^}}stored_fi_to_lds_2_small_objects:
19; GCN-DAG: v_mov_b32_e32 [[ZERO:v[0-9]+]], 0{{$}}
20; GCN-DAG: buffer_store_dword v{{[0-9]+}}, [[ZERO]], s{{\[[0-9]+:[0-9]+\]}}, s{{[0-9]+}} offen{{$}}
21; GCN-DAG: buffer_store_dword v{{[0-9]+}}, [[ZERO]], s{{\[[0-9]+:[0-9]+\]}}, s{{[0-9]+}} offen offset:4{{$}}
22
23; GCN-DAG: s_load_dword [[LDSPTR:s[0-9]+]]
24
25; GCN-DAG: v_mov_b32_e32 [[VLDSPTR:v[0-9]+]], [[LDSPTR]]
26; GCN: ds_write_b32  [[VLDSPTR]], [[ZERO]]
27
28; GCN-DAG: v_mov_b32_e32 [[FI1:v[0-9]+]], 4{{$}}
29; GCN: ds_write_b32  [[VLDSPTR]], [[FI1]]
30define void @stored_fi_to_lds_2_small_objects(float* addrspace(3)* %ptr) #0 {
31  %tmp0 = alloca float
32  %tmp1 = alloca float
33  store float 4.0, float* %tmp0
34  store float 4.0, float* %tmp1
35  store volatile float* %tmp0, float* addrspace(3)* %ptr
36  store volatile float* %tmp1, float* addrspace(3)* %ptr
37  ret void
38}
39
40; Same frame index is used multiple times in the store
41; GCN-LABEL: {{^}}stored_fi_to_self:
42; GCN-DAG: v_mov_b32_e32 [[K:v[0-9]+]], 0x4d2{{$}}
43; GCN-DAG: v_mov_b32_e32 [[ZERO:v[0-9]+]], 0{{$}}
44; GCN: buffer_store_dword [[K]], [[ZERO]], s{{\[[0-9]+:[0-9]+\]}}, s{{[0-9]+}} offen{{$}}
45; GCN: buffer_store_dword [[ZERO]], [[ZERO]], s{{\[[0-9]+:[0-9]+\]}}, s{{[0-9]+}} offen{{$}}
46define void @stored_fi_to_self() #0 {
47  %tmp = alloca i32*
48
49  ; Avoid optimizing everything out
50  store volatile i32* inttoptr (i32 1234 to i32*), i32** %tmp
51  %bitcast = bitcast i32** %tmp to i32*
52  store volatile i32* %bitcast, i32** %tmp
53  ret void
54}
55
56; GCN-LABEL: {{^}}stored_fi_to_self_offset:
57; GCN-DAG: v_mov_b32_e32 [[ZERO:v[0-9]+]], 0{{$}}
58; GCN-DAG: v_mov_b32_e32 [[K0:v[0-9]+]], 32{{$}}
59; GCN: buffer_store_dword [[K0]], [[ZERO]], s{{\[[0-9]+:[0-9]+\]}}, s{{[0-9]+}} offen{{$}}
60
61; GCN-DAG: v_mov_b32_e32 [[K1:v[0-9]+]], 0x4d2{{$}}
62; GCN: buffer_store_dword [[K1]], [[ZERO]], s{{\[[0-9]+:[0-9]+\]}}, s{{[0-9]+}} offen offset:2048{{$}}
63
64; GCN: v_mov_b32_e32 [[OFFSETK:v[0-9]+]], 0x800{{$}}
65; GCN: buffer_store_dword [[OFFSETK]], [[ZERO]], s{{\[[0-9]+:[0-9]+\]}}, s{{[0-9]+}} offen offset:2048{{$}}
66define void @stored_fi_to_self_offset() #0 {
67  %tmp0 = alloca [512 x i32]
68  %tmp1 = alloca i32*
69
70  ; Avoid optimizing everything out
71  %tmp0.cast = bitcast [512 x i32]* %tmp0 to i32*
72  store volatile i32 32, i32* %tmp0.cast
73
74  store volatile i32* inttoptr (i32 1234 to i32*), i32** %tmp1
75
76  %bitcast = bitcast i32** %tmp1 to i32*
77  store volatile i32* %bitcast, i32** %tmp1
78  ret void
79}
80
81; GCN-LABEL: {{^}}stored_fi_to_fi:
82; GCN: v_mov_b32_e32 [[ZERO:v[0-9]+]], 0{{$}}
83; GCN: buffer_store_dword v{{[0-9]+}}, [[ZERO]], s{{\[[0-9]+:[0-9]+\]}}, s{{[0-9]+}} offen{{$}}
84; GCN: buffer_store_dword v{{[0-9]+}}, [[ZERO]], s{{\[[0-9]+:[0-9]+\]}}, s{{[0-9]+}} offen offset:4{{$}}
85; GCN: buffer_store_dword v{{[0-9]+}}, [[ZERO]], s{{\[[0-9]+:[0-9]+\]}}, s{{[0-9]+}} offen offset:8{{$}}
86
87; GCN: v_mov_b32_e32 [[FI1:v[0-9]+]], 4{{$}}
88; GCN: buffer_store_dword [[FI1]], [[ZERO]], s{{\[[0-9]+:[0-9]+\]}}, s{{[0-9]+}} offen offset:8{{$}}
89
90; GCN: v_mov_b32_e32 [[FI2:v[0-9]+]], 8{{$}}
91; GCN: buffer_store_dword [[FI2]], [[ZERO]], s{{\[[0-9]+:[0-9]+\]}}, s{{[0-9]+}} offen offset:4{{$}}
92define void @stored_fi_to_fi() #0 {
93  %tmp0 = alloca i32*
94  %tmp1 = alloca i32*
95  %tmp2 = alloca i32*
96  store volatile i32* inttoptr (i32 1234 to i32*), i32** %tmp0
97  store volatile i32* inttoptr (i32 5678 to i32*), i32** %tmp1
98  store volatile i32* inttoptr (i32 9999 to i32*), i32** %tmp2
99
100  %bitcast1 = bitcast i32** %tmp1 to i32*
101  %bitcast2 = bitcast i32** %tmp2 to i32* ;  at offset 8
102
103  store volatile i32* %bitcast1, i32** %tmp2 ; store offset 4 at offset 8
104  store volatile i32* %bitcast2, i32** %tmp1 ; store offset 8 at offset 4
105  ret void
106}
107
108; GCN-LABEL: {{^}}stored_fi_to_global:
109; GCN: buffer_store_dword v{{[0-9]+}}, v{{[0-9]+}}, s{{\[[0-9]+:[0-9]+\]}}, s{{[0-9]+}} offen
110; GCN: v_mov_b32_e32 [[FI:v[0-9]+]], 0{{$}}
111; GCN: buffer_store_dword [[FI]]
112define void @stored_fi_to_global(float* addrspace(1)* %ptr) #0 {
113  %tmp = alloca float
114  store float 0.0, float *%tmp
115  store float* %tmp, float* addrspace(1)* %ptr
116  ret void
117}
118
119; Offset is applied
120; GCN-LABEL: {{^}}stored_fi_to_global_2_small_objects:
121; GCN: buffer_store_dword v{{[0-9]+}}, v{{[0-9]+}}, s{{\[[0-9]+:[0-9]+\]}}, s{{[0-9]+}} offen
122; GCN: buffer_store_dword v{{[0-9]+}}, v{{[0-9]+}}, s{{\[[0-9]+:[0-9]+\]}}, s{{[0-9]+}} offen
123; GCN: buffer_store_dword v{{[0-9]+}}, v{{[0-9]+}}, s{{\[[0-9]+:[0-9]+\]}}, s{{[0-9]+}} offen
124
125; GCN: v_mov_b32_e32 [[FI1:v[0-9]+]], 4{{$}}
126; GCN: buffer_store_dword [[FI1]], off, s{{\[[0-9]+:[0-9]+\]}}, 0{{$}}
127
128; GCN-DAG: v_mov_b32_e32 [[FI2:v[0-9]+]], 8{{$}}
129; GCN: buffer_store_dword [[FI2]], off, s{{\[[0-9]+:[0-9]+\]}}, 0{{$}}
130define void @stored_fi_to_global_2_small_objects(float* addrspace(1)* %ptr) #0 {
131  %tmp0 = alloca float
132  %tmp1 = alloca float
133  %tmp2 = alloca float
134  store volatile float 0.0, float *%tmp0
135  store volatile float 0.0, float *%tmp1
136  store volatile float 0.0, float *%tmp2
137  store volatile float* %tmp1, float* addrspace(1)* %ptr
138  store volatile float* %tmp2, float* addrspace(1)* %ptr
139  ret void
140}
141
142; GCN-LABEL: {{^}}stored_fi_to_global_huge_frame_offset:
143; GCN: s_add_i32 [[BASE_1_OFF_0:s[0-9]+]], 0, 0x3ffc
144; GCN: v_mov_b32_e32 [[BASE_0:v[0-9]+]], 0{{$}}
145; GCN: buffer_store_dword [[BASE_0]], v{{[0-9]+}}, s{{\[[0-9]+:[0-9]+\]}}, s{{[0-9]+}} offen
146
147; GCN: v_mov_b32_e32 [[V_BASE_1_OFF_0:v[0-9]+]], [[BASE_1_OFF_0]]
148; GCN: v_mov_b32_e32 [[K:v[0-9]+]], 0x3e7{{$}}
149; GCN: s_add_i32 [[BASE_1_OFF_1:s[0-9]+]], 0, 56
150; GCN: buffer_store_dword [[K]], [[V_BASE_1_OFF_0]], s{{\[[0-9]+:[0-9]+\]}}, s{{[0-9]+}} offen{{$}}
151
152; GCN: v_mov_b32_e32 [[V_BASE_1_OFF_1:v[0-9]+]], [[BASE_1_OFF_1]]
153; GCN: buffer_store_dword [[V_BASE_1_OFF_1]], off, s{{\[[0-9]+:[0-9]+\]}}, 0{{$}}
154define void @stored_fi_to_global_huge_frame_offset(i32* addrspace(1)* %ptr) #0 {
155  %tmp0 = alloca [4096 x i32]
156  %tmp1 = alloca [4096 x i32]
157  %gep0.tmp0 = getelementptr [4096 x i32], [4096 x i32]* %tmp0, i32 0, i32 0
158  store volatile i32 0, i32* %gep0.tmp0
159  %gep1.tmp0 = getelementptr [4096 x i32], [4096 x i32]* %tmp0, i32 0, i32 4095
160  store volatile i32 999, i32* %gep1.tmp0
161  %gep0.tmp1 = getelementptr [4096 x i32], [4096 x i32]* %tmp0, i32 0, i32 14
162  store i32* %gep0.tmp1, i32* addrspace(1)* %ptr
163  ret void
164}
165
166attributes #0 = { nounwind }
167