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