1// RUN: llvm-mc -arch=amdgcn -mcpu=gfx1010 -mattr=+wavefrontsize32,-wavefrontsize64 -show-encoding %s | FileCheck --check-prefix=GFX10 %s
2// RUN: llvm-mc -arch=amdgcn -mcpu=gfx1010 -mattr=-wavefrontsize32,+wavefrontsize64 -show-encoding %s | FileCheck --check-prefix=GFX10 %s
3
4//===----------------------------------------------------------------------===//
5// ENC_FLAT.
6//===----------------------------------------------------------------------===//
7
8flat_load_ubyte v5, v[1:2]
9// GFX10: encoding: [0x00,0x00,0x20,0xdc,0x01,0x00,0x7d,0x05]
10
11flat_load_sbyte v5, v[1:2]
12// GFX10: encoding: [0x00,0x00,0x24,0xdc,0x01,0x00,0x7d,0x05]
13
14flat_load_ushort v5, v[1:2]
15// GFX10: encoding: [0x00,0x00,0x28,0xdc,0x01,0x00,0x7d,0x05]
16
17flat_load_sshort v5, v[1:2]
18// GFX10: encoding: [0x00,0x00,0x2c,0xdc,0x01,0x00,0x7d,0x05]
19
20flat_load_dword v5, v[1:2]
21// GFX10: encoding: [0x00,0x00,0x30,0xdc,0x01,0x00,0x7d,0x05]
22
23flat_load_dwordx2 v[5:6], v[1:2]
24// GFX10: encoding: [0x00,0x00,0x34,0xdc,0x01,0x00,0x7d,0x05]
25
26flat_load_dwordx3 v[5:7], v[1:2]
27// GFX10: encoding: [0x00,0x00,0x3c,0xdc,0x01,0x00,0x7d,0x05]
28
29flat_load_dwordx4 v[5:8], v[1:2]
30// GFX10: encoding: [0x00,0x00,0x38,0xdc,0x01,0x00,0x7d,0x05]
31
32flat_store_byte v[1:2], v2
33// GFX10: encoding: [0x00,0x00,0x60,0xdc,0x01,0x02,0x7d,0x00]
34
35flat_store_short v[1:2], v2
36// GFX10: encoding: [0x00,0x00,0x68,0xdc,0x01,0x02,0x7d,0x00]
37
38flat_store_dword v[1:2], v2
39// GFX10: encoding: [0x00,0x00,0x70,0xdc,0x01,0x02,0x7d,0x00]
40
41flat_store_dwordx2 v[1:2], v[2:3]
42// GFX10: encoding: [0x00,0x00,0x74,0xdc,0x01,0x02,0x7d,0x00]
43
44flat_store_dwordx3 v[1:2], v[2:4]
45// GFX10: encoding: [0x00,0x00,0x7c,0xdc,0x01,0x02,0x7d,0x00]
46
47flat_store_dwordx4 v[1:2], v[2:5]
48// GFX10: encoding: [0x00,0x00,0x78,0xdc,0x01,0x02,0x7d,0x00]
49
50flat_atomic_swap v[1:2], v2
51// GFX10: encoding: [0x00,0x00,0xc0,0xdc,0x01,0x02,0x7d,0x00]
52
53flat_atomic_cmpswap v[1:2], v[2:3]
54// GFX10: encoding: [0x00,0x00,0xc4,0xdc,0x01,0x02,0x7d,0x00]
55
56flat_atomic_add v[1:2], v2
57// GFX10: encoding: [0x00,0x00,0xc8,0xdc,0x01,0x02,0x7d,0x00]
58
59flat_atomic_sub v[1:2], v2
60// GFX10: encoding: [0x00,0x00,0xcc,0xdc,0x01,0x02,0x7d,0x00]
61
62flat_atomic_smin v[1:2], v2
63// GFX10: encoding: [0x00,0x00,0xd4,0xdc,0x01,0x02,0x7d,0x00]
64
65flat_atomic_umin v[1:2], v2
66// GFX10: encoding: [0x00,0x00,0xd8,0xdc,0x01,0x02,0x7d,0x00]
67
68flat_atomic_smax v[1:2], v2
69// GFX10: encoding: [0x00,0x00,0xdc,0xdc,0x01,0x02,0x7d,0x00]
70
71flat_atomic_umax v[1:2], v2
72// GFX10: encoding: [0x00,0x00,0xe0,0xdc,0x01,0x02,0x7d,0x00]
73
74flat_atomic_and v[1:2], v2
75// GFX10: encoding: [0x00,0x00,0xe4,0xdc,0x01,0x02,0x7d,0x00]
76
77flat_atomic_or v[1:2], v2
78// GFX10: encoding: [0x00,0x00,0xe8,0xdc,0x01,0x02,0x7d,0x00]
79
80flat_atomic_xor v[1:2], v2
81// GFX10: encoding: [0x00,0x00,0xec,0xdc,0x01,0x02,0x7d,0x00]
82
83flat_atomic_inc v[1:2], v2
84// GFX10: encoding: [0x00,0x00,0xf0,0xdc,0x01,0x02,0x7d,0x00]
85
86flat_atomic_dec v[1:2], v2
87// GFX10: encoding: [0x00,0x00,0xf4,0xdc,0x01,0x02,0x7d,0x00]
88
89flat_atomic_swap_x2 v[1:2], v[2:3]
90// GFX10: encoding: [0x00,0x00,0x40,0xdd,0x01,0x02,0x7d,0x00]
91
92flat_atomic_cmpswap_x2 v[1:2], v[2:5]
93// GFX10: encoding: [0x00,0x00,0x44,0xdd,0x01,0x02,0x7d,0x00]
94
95flat_atomic_add_x2 v[1:2], v[2:3]
96// GFX10: encoding: [0x00,0x00,0x48,0xdd,0x01,0x02,0x7d,0x00]
97
98flat_atomic_sub_x2 v[1:2], v[2:3]
99// GFX10: encoding: [0x00,0x00,0x4c,0xdd,0x01,0x02,0x7d,0x00]
100
101flat_atomic_smin_x2 v[1:2], v[2:3]
102// GFX10: encoding: [0x00,0x00,0x54,0xdd,0x01,0x02,0x7d,0x00]
103
104flat_atomic_umin_x2 v[1:2], v[2:3]
105// GFX10: encoding: [0x00,0x00,0x58,0xdd,0x01,0x02,0x7d,0x00]
106
107flat_atomic_smax_x2 v[1:2], v[2:3]
108// GFX10: encoding: [0x00,0x00,0x5c,0xdd,0x01,0x02,0x7d,0x00]
109
110flat_atomic_umax_x2 v[1:2], v[2:3]
111// GFX10: encoding: [0x00,0x00,0x60,0xdd,0x01,0x02,0x7d,0x00]
112
113flat_atomic_and_x2 v[1:2], v[2:3]
114// GFX10: encoding: [0x00,0x00,0x64,0xdd,0x01,0x02,0x7d,0x00]
115
116flat_atomic_or_x2 v[1:2], v[2:3]
117// GFX10: encoding: [0x00,0x00,0x68,0xdd,0x01,0x02,0x7d,0x00]
118
119flat_atomic_xor_x2 v[1:2], v[2:3]
120// GFX10: encoding: [0x00,0x00,0x6c,0xdd,0x01,0x02,0x7d,0x00]
121
122flat_atomic_inc_x2 v[1:2], v[2:3]
123// GFX10: encoding: [0x00,0x00,0x70,0xdd,0x01,0x02,0x7d,0x00]
124
125flat_atomic_dec_x2 v[1:2], v[2:3]
126// GFX10: encoding: [0x00,0x00,0x74,0xdd,0x01,0x02,0x7d,0x00]
127
128//===----------------------------------------------------------------------===//
129// ENC_FLAT: dlc support for atomics.
130//===----------------------------------------------------------------------===//
131
132flat_atomic_add v[1:2], v2 dlc
133// GFX10: [0x00,0x10,0xc8,0xdc,0x01,0x02,0x7d,0x00]
134
135flat_atomic_add_x2 v[1:2], v[2:3] dlc
136// GFX10: [0x00,0x10,0x48,0xdd,0x01,0x02,0x7d,0x00]
137
138flat_atomic_and v[1:2], v2 dlc
139// GFX10: [0x00,0x10,0xe4,0xdc,0x01,0x02,0x7d,0x00]
140
141flat_atomic_and_x2 v[1:2], v[2:3] dlc
142// GFX10: [0x00,0x10,0x64,0xdd,0x01,0x02,0x7d,0x00]
143
144flat_atomic_cmpswap v[1:2], v[2:3] dlc
145// GFX10: [0x00,0x10,0xc4,0xdc,0x01,0x02,0x7d,0x00]
146
147flat_atomic_cmpswap_x2 v[1:2], v[2:5] dlc
148// GFX10: [0x00,0x10,0x44,0xdd,0x01,0x02,0x7d,0x00]
149
150flat_atomic_dec v[1:2], v2 dlc
151// GFX10: [0x00,0x10,0xf4,0xdc,0x01,0x02,0x7d,0x00]
152
153flat_atomic_dec_x2 v[1:2], v[2:3] dlc
154// GFX10: [0x00,0x10,0x74,0xdd,0x01,0x02,0x7d,0x00]
155
156flat_atomic_fcmpswap v[1:2], v[2:3] dlc
157// GFX10: [0x00,0x10,0xf8,0xdc,0x01,0x02,0x7d,0x00]
158
159flat_atomic_fcmpswap_x2 v[1:2], v[2:5] dlc
160// GFX10: [0x00,0x10,0x78,0xdd,0x01,0x02,0x7d,0x00]
161
162flat_atomic_fmax v[1:2], v2 dlc
163// GFX10: [0x00,0x10,0x00,0xdd,0x01,0x02,0x7d,0x00]
164
165flat_atomic_fmax_x2 v[1:2], v[2:3] dlc
166// GFX10: [0x00,0x10,0x80,0xdd,0x01,0x02,0x7d,0x00]
167
168flat_atomic_fmin v[1:2], v2 dlc
169// GFX10: [0x00,0x10,0xfc,0xdc,0x01,0x02,0x7d,0x00]
170
171flat_atomic_fmin_x2 v[1:2], v[2:3] dlc
172// GFX10: [0x00,0x10,0x7c,0xdd,0x01,0x02,0x7d,0x00]
173
174flat_atomic_inc v[1:2], v2 dlc
175// GFX10: [0x00,0x10,0xf0,0xdc,0x01,0x02,0x7d,0x00]
176
177flat_atomic_inc_x2 v[1:2], v[2:3] dlc
178// GFX10: [0x00,0x10,0x70,0xdd,0x01,0x02,0x7d,0x00]
179
180flat_atomic_or v[1:2], v2 dlc
181// GFX10: [0x00,0x10,0xe8,0xdc,0x01,0x02,0x7d,0x00]
182
183flat_atomic_or_x2 v[1:2], v[2:3] dlc
184// GFX10: [0x00,0x10,0x68,0xdd,0x01,0x02,0x7d,0x00]
185
186flat_atomic_smax v[1:2], v2 dlc
187// GFX10: [0x00,0x10,0xdc,0xdc,0x01,0x02,0x7d,0x00]
188
189flat_atomic_smax_x2 v[1:2], v[2:3] dlc
190// GFX10: [0x00,0x10,0x5c,0xdd,0x01,0x02,0x7d,0x00]
191
192flat_atomic_smin v[1:2], v2 dlc
193// GFX10: [0x00,0x10,0xd4,0xdc,0x01,0x02,0x7d,0x00]
194
195flat_atomic_smin_x2 v[1:2], v[2:3] dlc
196// GFX10: [0x00,0x10,0x54,0xdd,0x01,0x02,0x7d,0x00]
197
198flat_atomic_sub v[1:2], v2 dlc
199// GFX10: [0x00,0x10,0xcc,0xdc,0x01,0x02,0x7d,0x00]
200
201flat_atomic_sub_x2 v[1:2], v[2:3] dlc
202// GFX10: [0x00,0x10,0x4c,0xdd,0x01,0x02,0x7d,0x00]
203
204flat_atomic_swap v[1:2], v2 dlc
205// GFX10: [0x00,0x10,0xc0,0xdc,0x01,0x02,0x7d,0x00]
206
207flat_atomic_swap_x2 v[1:2], v[2:3] dlc
208// GFX10: [0x00,0x10,0x40,0xdd,0x01,0x02,0x7d,0x00]
209
210flat_atomic_umax v[1:2], v2 dlc
211// GFX10: [0x00,0x10,0xe0,0xdc,0x01,0x02,0x7d,0x00]
212
213flat_atomic_umax_x2 v[1:2], v[2:3] dlc
214// GFX10: [0x00,0x10,0x60,0xdd,0x01,0x02,0x7d,0x00]
215
216flat_atomic_umin v[1:2], v2 dlc
217// GFX10: [0x00,0x10,0xd8,0xdc,0x01,0x02,0x7d,0x00]
218
219flat_atomic_umin_x2 v[1:2], v[2:3] dlc
220// GFX10: [0x00,0x10,0x58,0xdd,0x01,0x02,0x7d,0x00]
221
222flat_atomic_xor v[1:2], v2 dlc
223// GFX10: [0x00,0x10,0xec,0xdc,0x01,0x02,0x7d,0x00]
224
225flat_atomic_xor_x2 v[1:2], v[2:3] dlc
226// GFX10: [0x00,0x10,0x6c,0xdd,0x01,0x02,0x7d,0x00]
227
228//===----------------------------------------------------------------------===//
229// ENC_FLAT_GLOBAL: dlc support for atomics.
230//===----------------------------------------------------------------------===//
231
232global_atomic_add v[1:2], v2, off dlc
233// GFX10: [0x00,0x90,0xc8,0xdc,0x01,0x02,0x7d,0x00]
234
235global_atomic_add_x2 v[1:2], v[2:3], off dlc
236// GFX10: [0x00,0x90,0x48,0xdd,0x01,0x02,0x7d,0x00]
237
238global_atomic_and v[1:2], v2, off dlc
239// GFX10: [0x00,0x90,0xe4,0xdc,0x01,0x02,0x7d,0x00]
240
241global_atomic_and_x2 v[1:2], v[2:3], off dlc
242// GFX10: [0x00,0x90,0x64,0xdd,0x01,0x02,0x7d,0x00]
243
244global_atomic_cmpswap v[1:2], v[2:3], off dlc
245// GFX10: [0x00,0x90,0xc4,0xdc,0x01,0x02,0x7d,0x00]
246
247global_atomic_cmpswap_x2 v[1:2], v[2:5], off dlc
248// GFX10: [0x00,0x90,0x44,0xdd,0x01,0x02,0x7d,0x00]
249
250global_atomic_dec v[1:2], v2, off dlc
251// GFX10: [0x00,0x90,0xf4,0xdc,0x01,0x02,0x7d,0x00]
252
253global_atomic_dec_x2 v[1:2], v[2:3], off dlc
254// GFX10: [0x00,0x90,0x74,0xdd,0x01,0x02,0x7d,0x00]
255
256global_atomic_fmax v[1:2], v2, off dlc
257// GFX10: [0x00,0x90,0x00,0xdd,0x01,0x02,0x7d,0x00]
258
259global_atomic_fmax_x2 v[1:2], v[2:3], off dlc
260// GFX10: [0x00,0x90,0x80,0xdd,0x01,0x02,0x7d,0x00]
261
262global_atomic_fmin v[1:2], v2, off dlc
263// GFX10: [0x00,0x90,0xfc,0xdc,0x01,0x02,0x7d,0x00]
264
265global_atomic_fmin_x2 v[1:2], v[2:3], off dlc
266// GFX10: [0x00,0x90,0x7c,0xdd,0x01,0x02,0x7d,0x00]
267
268global_atomic_inc v[1:2], v2, off dlc
269// GFX10: [0x00,0x90,0xf0,0xdc,0x01,0x02,0x7d,0x00]
270
271global_atomic_inc_x2 v[1:2], v[2:3], off dlc
272// GFX10: [0x00,0x90,0x70,0xdd,0x01,0x02,0x7d,0x00]
273
274global_atomic_or v[1:2], v2, off dlc
275// GFX10: [0x00,0x90,0xe8,0xdc,0x01,0x02,0x7d,0x00]
276
277global_atomic_or_x2 v[1:2], v[2:3], off dlc
278// GFX10: [0x00,0x90,0x68,0xdd,0x01,0x02,0x7d,0x00]
279
280global_atomic_smax v[1:2], v2, off dlc
281// GFX10: [0x00,0x90,0xdc,0xdc,0x01,0x02,0x7d,0x00]
282
283global_atomic_smax_x2 v[1:2], v[2:3], off dlc
284// GFX10: [0x00,0x90,0x5c,0xdd,0x01,0x02,0x7d,0x00]
285
286global_atomic_smin v[1:2], v2, off dlc
287// GFX10: [0x00,0x90,0xd4,0xdc,0x01,0x02,0x7d,0x00]
288
289global_atomic_smin_x2 v[1:2], v[2:3], off dlc
290// GFX10: [0x00,0x90,0x54,0xdd,0x01,0x02,0x7d,0x00]
291
292global_atomic_sub v[1:2], v2, off dlc
293// GFX10: [0x00,0x90,0xcc,0xdc,0x01,0x02,0x7d,0x00]
294
295global_atomic_sub_x2 v[1:2], v[2:3], off dlc
296// GFX10: [0x00,0x90,0x4c,0xdd,0x01,0x02,0x7d,0x00]
297
298global_atomic_swap v[1:2], v2, off dlc
299// GFX10: [0x00,0x90,0xc0,0xdc,0x01,0x02,0x7d,0x00]
300
301global_atomic_swap_x2 v[1:2], v[2:3], off dlc
302// GFX10: [0x00,0x90,0x40,0xdd,0x01,0x02,0x7d,0x00]
303
304global_atomic_umax v[1:2], v2, off dlc
305// GFX10: [0x00,0x90,0xe0,0xdc,0x01,0x02,0x7d,0x00]
306
307global_atomic_umax_x2 v[1:2], v[2:3], off dlc
308// GFX10: [0x00,0x90,0x60,0xdd,0x01,0x02,0x7d,0x00]
309
310global_atomic_umin v[1:2], v2, off dlc
311// GFX10: [0x00,0x90,0xd8,0xdc,0x01,0x02,0x7d,0x00]
312
313global_atomic_umin_x2 v[1:2], v[2:3], off dlc
314// GFX10: [0x00,0x90,0x58,0xdd,0x01,0x02,0x7d,0x00]
315
316global_atomic_xor v[1:2], v2, off dlc
317// GFX10: [0x00,0x90,0xec,0xdc,0x01,0x02,0x7d,0x00]
318
319global_atomic_xor_x2 v[1:2], v[2:3], off dlc
320// GFX10: [0x00,0x90,0x6c,0xdd,0x01,0x02,0x7d,0x00]
321
322//===----------------------------------------------------------------------===//
323// Also see flat-gfx10.s, flat-global.s, flat-scratch-instructions.s.
324//===----------------------------------------------------------------------===//
325