1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
| ; RUN: llc -amdgpu-scalarize-global-loads=false -march=amdgcn -verify-machineinstrs < %s | FileCheck -check-prefix=SI -check-prefix=FUNC %s
declare i32 @llvm.amdgcn.workitem.id.x() nounwind readnone
; Make sure we don't turn the 32-bit argument load into a 16-bit
; load. There aren't extending scalar lods, so that would require
; using a buffer_load instruction.
; FUNC-LABEL: {{^}}truncate_kernarg_i32_to_i16:
; SI: s_load_dword s
; SI: buffer_store_short v
define amdgpu_kernel void @truncate_kernarg_i32_to_i16(i16 addrspace(1)* %out, i32 %arg) nounwind {
%trunc = trunc i32 %arg to i16
store i16 %trunc, i16 addrspace(1)* %out
ret void
}
; It should be OK (and probably performance neutral) to reduce this,
; but we don't know if the load is uniform yet.
; FUNC-LABEL: {{^}}truncate_buffer_load_i32_to_i16:
; SI: buffer_load_dword v
; SI: buffer_store_short v
define amdgpu_kernel void @truncate_buffer_load_i32_to_i16(i16 addrspace(1)* %out, i32 addrspace(1)* %in) nounwind {
%tid = call i32 @llvm.amdgcn.workitem.id.x() nounwind readnone
%gep.in = getelementptr i32, i32 addrspace(1)* %in, i32 %tid
%gep.out = getelementptr i16, i16 addrspace(1)* %out, i32 %tid
%load = load i32, i32 addrspace(1)* %gep.in
%trunc = trunc i32 %load to i16
store i16 %trunc, i16 addrspace(1)* %gep.out
ret void
}
; FUNC-LABEL: {{^}}truncate_kernarg_i32_to_i8:
; SI: s_load_dword s
; SI: buffer_store_byte v
define amdgpu_kernel void @truncate_kernarg_i32_to_i8(i8 addrspace(1)* %out, i32 %arg) nounwind {
%trunc = trunc i32 %arg to i8
store i8 %trunc, i8 addrspace(1)* %out
ret void
}
; FUNC-LABEL: {{^}}truncate_buffer_load_i32_to_i8:
; SI: buffer_load_dword v
; SI: buffer_store_byte v
define amdgpu_kernel void @truncate_buffer_load_i32_to_i8(i8 addrspace(1)* %out, i32 addrspace(1)* %in) nounwind {
%tid = call i32 @llvm.amdgcn.workitem.id.x() nounwind readnone
%gep.in = getelementptr i32, i32 addrspace(1)* %in, i32 %tid
%gep.out = getelementptr i8, i8 addrspace(1)* %out, i32 %tid
%load = load i32, i32 addrspace(1)* %gep.in
%trunc = trunc i32 %load to i8
store i8 %trunc, i8 addrspace(1)* %gep.out
ret void
}
; FUNC-LABEL: {{^}}truncate_kernarg_i32_to_i1:
; SI: s_load_dword s
; SI: buffer_store_byte v
define amdgpu_kernel void @truncate_kernarg_i32_to_i1(i1 addrspace(1)* %out, i32 %arg) nounwind {
%trunc = trunc i32 %arg to i1
store i1 %trunc, i1 addrspace(1)* %out
ret void
}
; FUNC-LABEL: {{^}}truncate_buffer_load_i32_to_i1:
; SI: buffer_load_dword v
; SI: buffer_store_byte v
define amdgpu_kernel void @truncate_buffer_load_i32_to_i1(i1 addrspace(1)* %out, i32 addrspace(1)* %in) nounwind {
%tid = call i32 @llvm.amdgcn.workitem.id.x() nounwind readnone
%gep.in = getelementptr i32, i32 addrspace(1)* %in, i32 %tid
%gep.out = getelementptr i1, i1 addrspace(1)* %out, i32 %tid
%load = load i32, i32 addrspace(1)* %gep.in
%trunc = trunc i32 %load to i1
store i1 %trunc, i1 addrspace(1)* %gep.out
ret void
}
; FUNC-LABEL: {{^}}truncate_kernarg_i64_to_i32:
; SI: s_load_dword s
; SI: buffer_store_dword v
define amdgpu_kernel void @truncate_kernarg_i64_to_i32(i32 addrspace(1)* %out, [8 x i32], i64 %arg) nounwind {
%trunc = trunc i64 %arg to i32
store i32 %trunc, i32 addrspace(1)* %out
ret void
}
; FUNC-LABEL: {{^}}truncate_buffer_load_i64_to_i32:
; SI: buffer_load_dword v
; SI: buffer_store_dword v
define amdgpu_kernel void @truncate_buffer_load_i64_to_i32(i32 addrspace(1)* %out, i64 addrspace(1)* %in) nounwind {
%tid = call i32 @llvm.amdgcn.workitem.id.x() nounwind readnone
%gep.in = getelementptr i64, i64 addrspace(1)* %in, i32 %tid
%gep.out = getelementptr i32, i32 addrspace(1)* %out, i32 %tid
%load = load i64, i64 addrspace(1)* %gep.in
%trunc = trunc i64 %load to i32
store i32 %trunc, i32 addrspace(1)* %gep.out
ret void
}
; FUNC-LABEL: {{^}}srl_kernarg_i64_to_i32:
; SI: s_load_dword s
; SI: buffer_store_dword v
define amdgpu_kernel void @srl_kernarg_i64_to_i32(i32 addrspace(1)* %out, [8 x i32], i64 %arg) nounwind {
%srl = lshr i64 %arg, 32
%trunc = trunc i64 %srl to i32
store i32 %trunc, i32 addrspace(1)* %out
ret void
}
; FUNC-LABEL: {{^}}srl_buffer_load_i64_to_i32:
; SI: buffer_load_dword v
; SI: buffer_store_dword v
define amdgpu_kernel void @srl_buffer_load_i64_to_i32(i32 addrspace(1)* %out, i64 addrspace(1)* %in) nounwind {
%tid = call i32 @llvm.amdgcn.workitem.id.x() nounwind readnone
%gep.in = getelementptr i64, i64 addrspace(1)* %in, i32 %tid
%gep.out = getelementptr i32, i32 addrspace(1)* %out, i32 %tid
%load = load i64, i64 addrspace(1)* %gep.in
%srl = lshr i64 %load, 32
%trunc = trunc i64 %srl to i32
store i32 %trunc, i32 addrspace(1)* %gep.out
ret void
}
; Might as well reduce to 8-bit loads.
; FUNC-LABEL: {{^}}truncate_kernarg_i16_to_i8:
; SI: s_load_dword s
; SI: buffer_store_byte v
define amdgpu_kernel void @truncate_kernarg_i16_to_i8(i8 addrspace(1)* %out, i16 %arg) nounwind {
%trunc = trunc i16 %arg to i8
store i8 %trunc, i8 addrspace(1)* %out
ret void
}
; FUNC-LABEL: {{^}}truncate_buffer_load_i16_to_i8:
; SI: buffer_load_ubyte v
; SI: buffer_store_byte v
define amdgpu_kernel void @truncate_buffer_load_i16_to_i8(i8 addrspace(1)* %out, i16 addrspace(1)* %in) nounwind {
%tid = call i32 @llvm.amdgcn.workitem.id.x() nounwind readnone
%gep.in = getelementptr i16, i16 addrspace(1)* %in, i32 %tid
%gep.out = getelementptr i8, i8 addrspace(1)* %out, i32 %tid
%load = load i16, i16 addrspace(1)* %gep.in
%trunc = trunc i16 %load to i8
store i8 %trunc, i8 addrspace(1)* %gep.out
ret void
}
; FUNC-LABEL: {{^}}srl_kernarg_i64_to_i8:
; SI: s_load_dword s
; SI: buffer_store_byte v
define amdgpu_kernel void @srl_kernarg_i64_to_i8(i8 addrspace(1)* %out, [8 x i32], i64 %arg) nounwind {
%srl = lshr i64 %arg, 32
%trunc = trunc i64 %srl to i8
store i8 %trunc, i8 addrspace(1)* %out
ret void
}
; FUNC-LABEL: {{^}}srl_buffer_load_i64_to_i8:
; SI: buffer_load_dword v
; SI: buffer_store_byte v
define amdgpu_kernel void @srl_buffer_load_i64_to_i8(i8 addrspace(1)* %out, i64 addrspace(1)* %in) nounwind {
%tid = call i32 @llvm.amdgcn.workitem.id.x() nounwind readnone
%gep.in = getelementptr i64, i64 addrspace(1)* %in, i32 %tid
%gep.out = getelementptr i8, i8 addrspace(1)* %out, i32 %tid
%load = load i64, i64 addrspace(1)* %gep.in
%srl = lshr i64 %load, 32
%trunc = trunc i64 %srl to i8
store i8 %trunc, i8 addrspace(1)* %gep.out
ret void
}
; FUNC-LABEL: {{^}}truncate_kernarg_i64_to_i8:
; SI: s_load_dword s
; SI: buffer_store_byte v
define amdgpu_kernel void @truncate_kernarg_i64_to_i8(i8 addrspace(1)* %out, [8 x i32], i64 %arg) nounwind {
%trunc = trunc i64 %arg to i8
store i8 %trunc, i8 addrspace(1)* %out
ret void
}
; FUNC-LABEL: {{^}}truncate_buffer_load_i64_to_i8:
; SI: buffer_load_dword v
; SI: buffer_store_byte v
define amdgpu_kernel void @truncate_buffer_load_i64_to_i8(i8 addrspace(1)* %out, i64 addrspace(1)* %in) nounwind {
%tid = call i32 @llvm.amdgcn.workitem.id.x() nounwind readnone
%gep.in = getelementptr i64, i64 addrspace(1)* %in, i32 %tid
%gep.out = getelementptr i8, i8 addrspace(1)* %out, i32 %tid
%load = load i64, i64 addrspace(1)* %gep.in
%trunc = trunc i64 %load to i8
store i8 %trunc, i8 addrspace(1)* %gep.out
ret void
}
; FUNC-LABEL: {{^}}smrd_mask_i32_to_i16
; SI: s_load_dword [[LOAD:s[0-9]+]], s[{{[0-9]+}}:{{[0-9]+}}], 0x0
; SI: s_waitcnt lgkmcnt(0)
; SI: s_and_b32 s{{[0-9]+}}, [[LOAD]], 0xffff
define amdgpu_kernel void @smrd_mask_i32_to_i16(i32 addrspace(1)* %out, i32 addrspace(4)* %in) {
entry:
%val = load i32, i32 addrspace(4)* %in
%mask = and i32 %val, 65535
store i32 %mask, i32 addrspace(1)* %out
ret void
}
; FUNC-LABEL: {{^}}extract_hi_i64_bitcast_v2i32:
; SI: buffer_load_dword v
; SI: buffer_store_dword v
define amdgpu_kernel void @extract_hi_i64_bitcast_v2i32(i32 addrspace(1)* %out, <2 x i32> addrspace(1)* %in) nounwind {
%ld = load <2 x i32>, <2 x i32> addrspace(1)* %in
%bc = bitcast <2 x i32> %ld to i64
%hi = lshr i64 %bc, 32
%trunc = trunc i64 %hi to i32
store i32 %trunc, i32 addrspace(1)* %out
ret void
}
|