ae50673492bea1a90f0c969206e78953b251432a
[oota-llvm.git] / test / MC / AMDGPU / hsa.s
1 // RUN: llvm-mc -triple amdgcn--amdhsa -mcpu=kaveri -show-encoding %s | FileCheck %s --check-prefix=ASM
2 // RUN: llvm-mc -filetype=obj -triple amdgcn--amdhsa -mcpu=kaveri -show-encoding %s | llvm-readobj -s -sd | FileCheck %s --check-prefix=ELF
3
4 // ELF: Section {
5 // ELF: Name: .hsatext
6 // ELF: Type: SHT_PROGBITS (0x1)
7 // ELF: Flags [ (0xC00007)
8 // ELF: SHF_ALLOC (0x2)
9 // ELF: SHF_AMDGPU_HSA_AGENT (0x800000)
10 // ELF: SHF_AMDGPU_HSA_CODE (0x400000)
11 // ELF: SHF_EXECINSTR (0x4)
12 // ELF: SHF_WRITE (0x1)
13 // ELF: }
14
15 // ELF: SHT_NOTE
16 // ELF: 0000: 04000000 08000000 01000000 414D4400
17 // ELF: 0010: 01000000 00000000 04000000 1B000000
18 // ELF: 0020: 03000000 414D4400 04000700 07000000
19 // ELF: 0030: 00000000 00000000 414D4400 414D4447
20 // ELF: 0040: 50550000
21
22 .hsa_code_object_version 1,0
23 // ASM: .hsa_code_object_version 1,0
24
25 .hsa_code_object_isa 7,0,0,"AMD","AMDGPU"
26 // ASM: .hsa_code_object_isa 7,0,0,"AMD","AMDGPU"
27
28 .hsatext
29 // ASM: .hsatext
30
31 amd_kernel_code_t_test_all:
32 ; Test all amd_kernel_code_t members with non-default values.
33 .amd_kernel_code_t
34         kernel_code_version_major = 100
35         kernel_code_version_minor = 100
36         machine_kind = 0
37         machine_version_major = 5
38         machine_version_minor = 5
39         machine_version_stepping = 5
40         kernel_code_entry_byte_offset = 512
41         kernel_code_prefetch_byte_size = 1
42         max_scratch_backing_memory_byte_size = 1
43         compute_pgm_rsrc1_vgprs = 1
44         compute_pgm_rsrc1_sgprs = 1
45         compute_pgm_rsrc1_priority = 1
46         compute_pgm_rsrc1_float_mode = 1
47         compute_pgm_rsrc1_priv = 1
48         compute_pgm_rsrc1_dx10_clamp = 1
49         compute_pgm_rsrc1_debug_mode = 1
50         compute_pgm_rsrc1_ieee_mode = 1
51         compute_pgm_rsrc2_scratch_en = 1
52         compute_pgm_rsrc2_user_sgpr = 1
53         compute_pgm_rsrc2_tgid_x_en = 1
54         compute_pgm_rsrc2_tgid_y_en = 1
55         compute_pgm_rsrc2_tgid_z_en = 1
56         compute_pgm_rsrc2_tg_size_en = 1
57         compute_pgm_rsrc2_tidig_comp_cnt = 1
58         compute_pgm_rsrc2_excp_en_msb = 1
59         compute_pgm_rsrc2_lds_size = 1
60         compute_pgm_rsrc2_excp_en = 1
61         enable_sgpr_private_segment_buffer = 1
62         enable_sgpr_dispatch_ptr = 1
63         enable_sgpr_queue_ptr = 1
64         enable_sgpr_kernarg_segment_ptr = 1
65         enable_sgpr_dispatch_id = 1
66         enable_sgpr_flat_scratch_init = 1
67         enable_sgpr_private_segment_size = 1
68         enable_sgpr_grid_workgroup_count_x = 1
69         enable_sgpr_grid_workgroup_count_y = 1
70         enable_sgpr_grid_workgroup_count_z = 1
71         enable_ordered_append_gds = 1
72         private_element_size = 1
73         is_ptr64 = 1
74         is_dynamic_callstack = 1
75         is_debug_enabled = 1
76         is_xnack_enabled = 1
77         workitem_private_segment_byte_size = 1
78         workgroup_group_segment_byte_size = 1
79         gds_segment_byte_size = 1
80         kernarg_segment_byte_size = 1
81         workgroup_fbarrier_count = 1
82         wavefront_sgpr_count = 1
83         workitem_vgpr_count = 1
84         reserved_vgpr_first = 1
85         reserved_vgpr_count = 1
86         reserved_sgpr_first = 1
87         reserved_sgpr_count = 1
88         debug_wavefront_private_segment_offset_sgpr = 1
89         debug_private_segment_buffer_sgpr = 1
90         kernarg_segment_alignment = 5
91         group_segment_alignment = 5
92         private_segment_alignment = 5
93         wavefront_size = 5
94         call_convention = 1
95         runtime_loader_kernel_symbol = 1
96 .end_amd_kernel_code_t
97
98 // ASM-LABEL: {{^}}amd_kernel_code_t_test_all:
99 // ASM: .amd_kernel_code_t
100 // ASM: kernel_code_version_major = 100
101 // ASM: kernel_code_version_minor = 100
102 // ASM: machine_kind = 0
103 // ASM: machine_version_major = 5
104 // ASM: machine_version_minor = 5
105 // ASM: machine_version_stepping = 5
106 // ASM: kernel_code_entry_byte_offset = 512
107 // ASM: kernel_code_prefetch_byte_size = 1
108 // ASM: max_scratch_backing_memory_byte_size = 1
109 // ASM: compute_pgm_rsrc1_vgprs = 1
110 // ASM: compute_pgm_rsrc1_sgprs = 1
111 // ASM: compute_pgm_rsrc1_priority = 1
112 // ASM: compute_pgm_rsrc1_float_mode = 1 
113 // ASM: compute_pgm_rsrc1_priv = 1
114 // ASM: compute_pgm_rsrc1_dx10_clamp = 1
115 // ASM: compute_pgm_rsrc1_debug_mode = 1 
116 // ASM: compute_pgm_rsrc1_ieee_mode = 1
117 // ASM: compute_pgm_rsrc2_scratch_en = 1
118 // ASM: compute_pgm_rsrc2_user_sgpr = 1
119 // ASM: compute_pgm_rsrc2_tgid_x_en = 1
120 // ASM: compute_pgm_rsrc2_tgid_y_en = 1
121 // ASM: compute_pgm_rsrc2_tgid_z_en = 1
122 // ASM: compute_pgm_rsrc2_tg_size_en = 1
123 // ASM: compute_pgm_rsrc2_tidig_comp_cnt = 1
124 // ASM: compute_pgm_rsrc2_excp_en_msb = 1
125 // ASM: compute_pgm_rsrc2_lds_size = 1
126 // ASM: compute_pgm_rsrc2_excp_en = 1
127 // ASM: enable_sgpr_private_segment_buffer = 1
128 // ASM: enable_sgpr_dispatch_ptr = 1
129 // ASM: enable_sgpr_queue_ptr = 1
130 // ASM: enable_sgpr_kernarg_segment_ptr = 1
131 // ASM: enable_sgpr_dispatch_id = 1
132 // ASM: enable_sgpr_flat_scratch_init = 1
133 // ASM: enable_sgpr_private_segment_size = 1
134 // ASM: enable_sgpr_grid_workgroup_count_x = 1
135 // ASM: enable_sgpr_grid_workgroup_count_y = 1
136 // ASM: enable_sgpr_grid_workgroup_count_z = 1
137 // ASM: enable_ordered_append_gds = 1
138 // ASM: private_element_size = 1
139 // ASM: is_ptr64 = 1
140 // ASM: is_dynamic_callstack = 1
141 // ASM: is_debug_enabled = 1
142 // ASM: is_xnack_enabled = 1
143 // ASM: workitem_private_segment_byte_size = 1
144 // ASM: workgroup_group_segment_byte_size = 1
145 // ASM: gds_segment_byte_size = 1
146 // ASM: kernarg_segment_byte_size = 1
147 // ASM: workgroup_fbarrier_count = 1
148 // ASM: wavefront_sgpr_count = 1
149 // ASM: workitem_vgpr_count = 1
150 // ASM: reserved_vgpr_first = 1
151 // ASM: reserved_vgpr_count = 1
152 // ASM: reserved_sgpr_first = 1
153 // ASM: reserved_sgpr_count = 1
154 // ASM: debug_wavefront_private_segment_offset_sgpr = 1
155 // ASM: debug_private_segment_buffer_sgpr = 1
156 // ASM: kernarg_segment_alignment = 5
157 // ASM: group_segment_alignment = 5
158 // ASM: private_segment_alignment = 5
159 // ASM: wavefront_size = 5
160 // ASM: call_convention = 1
161 // ASM: runtime_loader_kernel_symbol = 1
162 // ASM: .end_amd_kernel_code_t
163
164 amd_kernel_code_t_minimal:
165 .amd_kernel_code_t
166         enable_sgpr_kernarg_segment_ptr = 1
167         is_ptr64 = 1
168         compute_pgm_rsrc1_vgprs = 1
169         compute_pgm_rsrc1_sgprs = 1
170         compute_pgm_rsrc2_user_sgpr = 2
171         kernarg_segment_byte_size = 16
172         wavefront_sgpr_count = 8
173 //      wavefront_sgpr_count = 7
174 ;       wavefront_sgpr_count = 7
175 // Make sure a blank line won't break anything:
176
177 // Make sure a line with whitespace won't break anything:
178    
179         workitem_vgpr_count = 16
180 .end_amd_kernel_code_t
181
182 // ASM-LABEL: {{^}}amd_kernel_code_t_minimal:
183 // ASM: .amd_kernel_code_t
184 // ASM: kernel_code_version_major = 1
185 // ASM: kernel_code_version_minor = 0
186 // ASM: machine_kind = 1
187 // ASM: machine_version_major = 7
188 // ASM: machine_version_minor = 0
189 // ASM: machine_version_stepping = 0
190 // ASM: kernel_code_entry_byte_offset = 256
191 // ASM: kernel_code_prefetch_byte_size = 0
192 // ASM: max_scratch_backing_memory_byte_size = 0
193 // ASM: compute_pgm_rsrc1_vgprs = 1
194 // ASM: compute_pgm_rsrc1_sgprs = 1
195 // ASM: compute_pgm_rsrc1_priority = 0
196 // ASM: compute_pgm_rsrc1_float_mode = 0
197 // ASM: compute_pgm_rsrc1_priv = 0
198 // ASM: compute_pgm_rsrc1_dx10_clamp = 0
199 // ASM: compute_pgm_rsrc1_debug_mode = 0
200 // ASM: compute_pgm_rsrc1_ieee_mode = 0
201 // ASM: compute_pgm_rsrc2_scratch_en = 0
202 // ASM: compute_pgm_rsrc2_user_sgpr = 2
203 // ASM: compute_pgm_rsrc2_tgid_x_en = 0
204 // ASM: compute_pgm_rsrc2_tgid_y_en = 0
205 // ASM: compute_pgm_rsrc2_tgid_z_en = 0
206 // ASM: compute_pgm_rsrc2_tg_size_en = 0
207 // ASM: compute_pgm_rsrc2_tidig_comp_cnt = 0
208 // ASM: compute_pgm_rsrc2_excp_en_msb = 0
209 // ASM: compute_pgm_rsrc2_lds_size = 0
210 // ASM: compute_pgm_rsrc2_excp_en = 0
211 // ASM: enable_sgpr_private_segment_buffer = 0
212 // ASM: enable_sgpr_dispatch_ptr = 0
213 // ASM: enable_sgpr_queue_ptr = 0
214 // ASM: enable_sgpr_kernarg_segment_ptr = 1
215 // ASM: enable_sgpr_dispatch_id = 0
216 // ASM: enable_sgpr_flat_scratch_init = 0
217 // ASM: enable_sgpr_private_segment_size = 0
218 // ASM: enable_sgpr_grid_workgroup_count_x = 0
219 // ASM: enable_sgpr_grid_workgroup_count_y = 0
220 // ASM: enable_sgpr_grid_workgroup_count_z = 0
221 // ASM: enable_ordered_append_gds = 0
222 // ASM: private_element_size = 0
223 // ASM: is_ptr64 = 1
224 // ASM: is_dynamic_callstack = 0
225 // ASM: is_debug_enabled = 0
226 // ASM: is_xnack_enabled = 0
227 // ASM: workitem_private_segment_byte_size = 0
228 // ASM: workgroup_group_segment_byte_size = 0
229 // ASM: gds_segment_byte_size = 0
230 // ASM: kernarg_segment_byte_size = 16
231 // ASM: workgroup_fbarrier_count = 0
232 // ASM: wavefront_sgpr_count = 8
233 // ASM: workitem_vgpr_count = 16
234 // ASM: reserved_vgpr_first = 0
235 // ASM: reserved_vgpr_count = 0
236 // ASM: reserved_sgpr_first = 0
237 // ASM: reserved_sgpr_count = 0
238 // ASM: debug_wavefront_private_segment_offset_sgpr = 0
239 // ASM: debug_private_segment_buffer_sgpr = 0
240 // ASM: kernarg_segment_alignment = 4
241 // ASM: group_segment_alignment = 4
242 // ASM: private_segment_alignment = 4
243 // ASM: wavefront_size = 6
244 // ASM: call_convention = 0
245 // ASM: runtime_loader_kernel_symbol = 0
246 // ASM: .end_amd_kernel_code_t