| blob | 1105caf980ef4c5526506e5dd40423ab8b434a7c |
1 ; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx700 --amdhsa-code-object-version=3 -filetype=obj -o - < %s | llvm-readelf --notes - | FileCheck --check-prefix=CHECK %s
2 ; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx803 --amdhsa-code-object-version=3 -filetype=obj -o - < %s | llvm-readelf --notes - | FileCheck --check-prefix=CHECK %s
3 ; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx900 --amdhsa-code-object-version=3 -filetype=obj -o - < %s | llvm-readelf --notes - | FileCheck --check-prefix=CHECK %s
5 ; CHECK: ---
6 ; CHECK: amdhsa.kernels:
8 ; CHECK: - .args:
9 ; CHECK-NEXT: - .address_space: global
10 ; CHECK-NEXT: .name: r
11 ; CHECK-NEXT: .offset: 0
12 ; CHECK-NEXT: .size: 8
13 ; CHECK-NEXT: .value_kind: global_buffer
14 ; CHECK-NEXT: - .address_space: global
15 ; CHECK-NEXT: .name: a
16 ; CHECK-NEXT: .offset: 8
17 ; CHECK-NEXT: .size: 8
18 ; CHECK-NEXT: .value_kind: global_buffer
19 ; CHECK-NEXT: - .address_space: global
20 ; CHECK-NEXT: .name: b
21 ; CHECK-NEXT: .offset: 16
22 ; CHECK-NEXT: .size: 8
23 ; CHECK-NEXT: .value_kind: global_buffer
24 ; CHECK: .name: test0
25 ; CHECK: .symbol: test0.kd
26 define amdgpu_kernel void @test0(
27 half addrspace(1)* %r,
28 half addrspace(1)* %a,
29 half addrspace(1)* %b) {
30 entry:
31 %a.val = load half, half addrspace(1)* %a
32 %b.val = load half, half addrspace(1)* %b
33 %r.val = fadd half %a.val, %b.val
34 store half %r.val, half addrspace(1)* %r
35 ret void
36 }
38 ; CHECK: - .args:
39 ; CHECK-NEXT: - .address_space: global
40 ; CHECK-NEXT: .name: r
41 ; CHECK-NEXT: .offset: 0
42 ; CHECK-NEXT: .size: 8
43 ; CHECK-NEXT: .value_kind: global_buffer
44 ; CHECK-NEXT: - .address_space: global
45 ; CHECK-NEXT: .name: a
46 ; CHECK-NEXT: .offset: 8
47 ; CHECK-NEXT: .size: 8
48 ; CHECK-NEXT: .value_kind: global_buffer
49 ; CHECK-NEXT: - .address_space: global
50 ; CHECK-NEXT: .name: b
51 ; CHECK-NEXT: .offset: 16
52 ; CHECK-NEXT: .size: 8
53 ; CHECK-NEXT: .value_kind: global_buffer
54 ; CHECK-NEXT: - .offset: 24
55 ; CHECK-NEXT: .size: 8
56 ; CHECK-NEXT: .value_kind: hidden_global_offset_x
57 ; CHECK: .name: test8
58 ; CHECK: .symbol: test8.kd
59 define amdgpu_kernel void @test8(
60 half addrspace(1)* %r,
61 half addrspace(1)* %a,
62 half addrspace(1)* %b) #0 {
63 entry:
64 %a.val = load half, half addrspace(1)* %a
65 %b.val = load half, half addrspace(1)* %b
66 %r.val = fadd half %a.val, %b.val
67 store half %r.val, half addrspace(1)* %r
68 ret void
69 }
71 ; CHECK: - .args:
72 ; CHECK-NEXT: - .address_space: global
73 ; CHECK-NEXT: .name: r
74 ; CHECK-NEXT: .offset: 0
75 ; CHECK-NEXT: .size: 8
76 ; CHECK-NEXT: .value_kind: global_buffer
77 ; CHECK-NEXT: - .address_space: global
78 ; CHECK-NEXT: .name: a
79 ; CHECK-NEXT: .offset: 8
80 ; CHECK-NEXT: .size: 8
81 ; CHECK-NEXT: .value_kind: global_buffer
82 ; CHECK-NEXT: - .address_space: global
83 ; CHECK-NEXT: .name: b
84 ; CHECK-NEXT: .offset: 16
85 ; CHECK-NEXT: .size: 8
86 ; CHECK-NEXT: .value_kind: global_buffer
87 ; CHECK-NEXT: - .offset: 24
88 ; CHECK-NEXT: .size: 8
89 ; CHECK-NEXT: .value_kind: hidden_global_offset_x
90 ; CHECK-NEXT: - .offset: 32
91 ; CHECK-NEXT: .size: 8
92 ; CHECK-NEXT: .value_kind: hidden_global_offset_y
93 ; CHECK: .name: test16
94 ; CHECK: .symbol: test16.kd
95 define amdgpu_kernel void @test16(
96 half addrspace(1)* %r,
97 half addrspace(1)* %a,
98 half addrspace(1)* %b) #1 {
99 entry:
100 %a.val = load half, half addrspace(1)* %a
101 %b.val = load half, half addrspace(1)* %b
102 %r.val = fadd half %a.val, %b.val
103 store half %r.val, half addrspace(1)* %r
104 ret void
105 }
107 ; CHECK: - .args:
108 ; CHECK-NEXT: - .address_space: global
109 ; CHECK-NEXT: .name: r
110 ; CHECK-NEXT: .offset: 0
111 ; CHECK-NEXT: .size: 8
112 ; CHECK-NEXT: .value_kind: global_buffer
113 ; CHECK-NEXT: - .address_space: global
114 ; CHECK-NEXT: .name: a
115 ; CHECK-NEXT: .offset: 8
116 ; CHECK-NEXT: .size: 8
117 ; CHECK-NEXT: .value_kind: global_buffer
118 ; CHECK-NEXT: - .address_space: global
119 ; CHECK-NEXT: .name: b
120 ; CHECK-NEXT: .offset: 16
121 ; CHECK-NEXT: .size: 8
122 ; CHECK-NEXT: .value_kind: global_buffer
123 ; CHECK-NEXT: - .offset: 24
124 ; CHECK-NEXT: .size: 8
125 ; CHECK-NEXT: .value_kind: hidden_global_offset_x
126 ; CHECK-NEXT: - .offset: 32
127 ; CHECK-NEXT: .size: 8
128 ; CHECK-NEXT: .value_kind: hidden_global_offset_y
129 ; CHECK-NEXT: - .offset: 40
130 ; CHECK-NEXT: .size: 8
131 ; CHECK-NEXT: .value_kind: hidden_global_offset_z
132 ; CHECK: .name: test24
133 ; CHECK: .symbol: test24.kd
134 define amdgpu_kernel void @test24(
135 half addrspace(1)* %r,
136 half addrspace(1)* %a,
137 half addrspace(1)* %b) #2 {
138 entry:
139 %a.val = load half, half addrspace(1)* %a
140 %b.val = load half, half addrspace(1)* %b
141 %r.val = fadd half %a.val, %b.val
142 store half %r.val, half addrspace(1)* %r
143 ret void
144 }
146 ; CHECK: - .args:
147 ; CHECK-NEXT: - .address_space: global
148 ; CHECK-NEXT: .name: r
149 ; CHECK-NEXT: .offset: 0
150 ; CHECK-NEXT: .size: 8
151 ; CHECK-NEXT: .value_kind: global_buffer
152 ; CHECK-NEXT: - .address_space: global
153 ; CHECK-NEXT: .name: a
154 ; CHECK-NEXT: .offset: 8
155 ; CHECK-NEXT: .size: 8
156 ; CHECK-NEXT: .value_kind: global_buffer
157 ; CHECK-NEXT: - .address_space: global
158 ; CHECK-NEXT: .name: b
159 ; CHECK-NEXT: .offset: 16
160 ; CHECK-NEXT: .size: 8
161 ; CHECK-NEXT: .value_kind: global_buffer
162 ; CHECK-NEXT: - .offset: 24
163 ; CHECK-NEXT: .size: 8
164 ; CHECK-NEXT: .value_kind: hidden_global_offset_x
165 ; CHECK-NEXT: - .offset: 32
166 ; CHECK-NEXT: .size: 8
167 ; CHECK-NEXT: .value_kind: hidden_global_offset_y
168 ; CHECK-NEXT: - .offset: 40
169 ; CHECK-NEXT: .size: 8
170 ; CHECK-NEXT: .value_kind: hidden_global_offset_z
171 ; CHECK-NEXT: - .address_space: global
172 ; CHECK-NEXT: .offset: 48
173 ; CHECK-NEXT: .size: 8
174 ; CHECK-NEXT: .value_kind: hidden_none
175 ; CHECK: .name: test32
176 ; CHECK: .symbol: test32.kd
177 define amdgpu_kernel void @test32(
178 half addrspace(1)* %r,
179 half addrspace(1)* %a,
180 half addrspace(1)* %b) #3 {
181 entry:
182 %a.val = load half, half addrspace(1)* %a
183 %b.val = load half, half addrspace(1)* %b
184 %r.val = fadd half %a.val, %b.val
185 store half %r.val, half addrspace(1)* %r
186 ret void
187 }
189 ; CHECK: - .args:
190 ; CHECK-NEXT: - .address_space: global
191 ; CHECK-NEXT: .name: r
192 ; CHECK-NEXT: .offset: 0
193 ; CHECK-NEXT: .size: 8
194 ; CHECK-NEXT: .value_kind: global_buffer
195 ; CHECK-NEXT: - .address_space: global
196 ; CHECK-NEXT: .name: a
197 ; CHECK-NEXT: .offset: 8
198 ; CHECK-NEXT: .size: 8
199 ; CHECK-NEXT: .value_kind: global_buffer
200 ; CHECK-NEXT: - .address_space: global
201 ; CHECK-NEXT: .name: b
202 ; CHECK-NEXT: .offset: 16
203 ; CHECK-NEXT: .size: 8
204 ; CHECK-NEXT: .value_kind: global_buffer
205 ; CHECK-NEXT: - .offset: 24
206 ; CHECK-NEXT: .size: 8
207 ; CHECK-NEXT: .value_kind: hidden_global_offset_x
208 ; CHECK-NEXT: - .offset: 32
209 ; CHECK-NEXT: .size: 8
210 ; CHECK-NEXT: .value_kind: hidden_global_offset_y
211 ; CHECK-NEXT: - .offset: 40
212 ; CHECK-NEXT: .size: 8
213 ; CHECK-NEXT: .value_kind: hidden_global_offset_z
214 ; CHECK-NEXT: - .address_space: global
215 ; CHECK-NEXT: .offset: 48
216 ; CHECK-NEXT: .size: 8
217 ; CHECK-NEXT: .value_kind: hidden_none
218 ; CHECK-NEXT: - .address_space: global
219 ; CHECK-NEXT: .offset: 56
220 ; CHECK-NEXT: .size: 8
221 ; CHECK-NEXT: .value_kind: hidden_none
222 ; CHECK-NEXT: - .address_space: global
223 ; CHECK-NEXT: .offset: 64
224 ; CHECK-NEXT: .size: 8
225 ; CHECK-NEXT: .value_kind: hidden_none
226 ; CHECK: .name: test48
227 ; CHECK: .symbol: test48.kd
228 define amdgpu_kernel void @test48(
229 half addrspace(1)* %r,
230 half addrspace(1)* %a,
231 half addrspace(1)* %b) #4 {
232 entry:
233 %a.val = load half, half addrspace(1)* %a
234 %b.val = load half, half addrspace(1)* %b
235 %r.val = fadd half %a.val, %b.val
236 store half %r.val, half addrspace(1)* %r
237 ret void
238 }
240 ; CHECK: - .args:
241 ; CHECK-NEXT: - .address_space: global
242 ; CHECK-NEXT: .name: r
243 ; CHECK-NEXT: .offset: 0
244 ; CHECK-NEXT: .size: 8
245 ; CHECK-NEXT: .value_kind: global_buffer
246 ; CHECK-NEXT: - .address_space: global
247 ; CHECK-NEXT: .name: a
248 ; CHECK-NEXT: .offset: 8
249 ; CHECK-NEXT: .size: 8
250 ; CHECK-NEXT: .value_kind: global_buffer
251 ; CHECK-NEXT: - .address_space: global
252 ; CHECK-NEXT: .name: b
253 ; CHECK-NEXT: .offset: 16
254 ; CHECK-NEXT: .size: 8
255 ; CHECK-NEXT: .value_kind: global_buffer
256 ; CHECK-NEXT: - .offset: 24
257 ; CHECK-NEXT: .size: 8
258 ; CHECK-NEXT: .value_kind: hidden_global_offset_x
259 ; CHECK-NEXT: - .offset: 32
260 ; CHECK-NEXT: .size: 8
261 ; CHECK-NEXT: .value_kind: hidden_global_offset_y
262 ; CHECK-NEXT: - .offset: 40
263 ; CHECK-NEXT: .size: 8
264 ; CHECK-NEXT: .value_kind: hidden_global_offset_z
265 ; CHECK-NEXT: - .address_space: global
266 ; CHECK-NEXT: .offset: 48
267 ; CHECK-NEXT: .size: 8
268 ; CHECK-NEXT: .value_kind: hidden_none
269 ; CHECK-NEXT: - .address_space: global
270 ; CHECK-NEXT: .offset: 56
271 ; CHECK-NEXT: .size: 8
272 ; CHECK-NEXT: .value_kind: hidden_none
273 ; CHECK-NEXT: - .address_space: global
274 ; CHECK-NEXT: .offset: 64
275 ; CHECK-NEXT: .size: 8
276 ; CHECK-NEXT: .value_kind: hidden_none
277 ; CHECK-NEXT: - .address_space: global
278 ; CHECK-NEXT: .offset: 72
279 ; CHECK-NEXT: .size: 8
280 ; CHECK-NEXT: .value_kind: hidden_multigrid_sync_arg
281 ; CHECK: .name: test56
282 ; CHECK: .symbol: test56.kd
283 define amdgpu_kernel void @test56(
284 half addrspace(1)* %r,
285 half addrspace(1)* %a,
286 half addrspace(1)* %b) #5 {
287 entry:
288 %a.val = load half, half addrspace(1)* %a
289 %b.val = load half, half addrspace(1)* %b
290 %r.val = fadd half %a.val, %b.val
291 store half %r.val, half addrspace(1)* %r
292 ret void
293 }
295 ; CHECK: amdhsa.version:
296 ; CHECK-NEXT: - 1
297 ; CHECK-NEXT: - 0
299 attributes #0 = { "amdgpu-implicitarg-num-bytes"="8" }
300 attributes #1 = { "amdgpu-implicitarg-num-bytes"="16" }
301 attributes #2 = { "amdgpu-implicitarg-num-bytes"="24" }
302 attributes #3 = { "amdgpu-implicitarg-num-bytes"="32" }
303 attributes #4 = { "amdgpu-implicitarg-num-bytes"="48" }
304 attributes #5 = { "amdgpu-implicitarg-num-bytes"="56" }