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