1 /*
2 * Copyright © 2018 Valve Corporation
3 *
4 * SPDX-License-Identifier: MIT
5 */
6
7 #include "aco_ir.h"
8
9 #include "util/u_debug.h"
10
11 #if AMD_LLVM_AVAILABLE
12 #if defined(_MSC_VER) && defined(restrict)
13 #undef restrict
14 #endif
15 #include "llvm/ac_llvm_util.h"
16
17 #include "llvm-c/Disassembler.h"
18 #include <llvm/ADT/StringRef.h>
19 #include <llvm/MC/MCDisassembler/MCDisassembler.h>
20 #endif
21
22 #include <array>
23 #include <iomanip>
24 #include <vector>
25
26 namespace aco {
27 namespace {
28
29 std::vector<bool>
get_referenced_blocks(Program * program)30 get_referenced_blocks(Program* program)
31 {
32 std::vector<bool> referenced_blocks(program->blocks.size());
33 referenced_blocks[0] = true;
34 for (Block& block : program->blocks) {
35 for (unsigned succ : block.linear_succs)
36 referenced_blocks[succ] = true;
37 }
38 return referenced_blocks;
39 }
40
41 void
print_block_markers(FILE * output,Program * program,const std::vector<bool> & referenced_blocks,unsigned * next_block,unsigned pos)42 print_block_markers(FILE* output, Program* program, const std::vector<bool>& referenced_blocks,
43 unsigned* next_block, unsigned pos)
44 {
45 while (*next_block < program->blocks.size() && pos == program->blocks[*next_block].offset) {
46 if (referenced_blocks[*next_block])
47 fprintf(output, "BB%u:\n", *next_block);
48 (*next_block)++;
49 }
50 }
51
52 void
print_instr(FILE * output,const std::vector<uint32_t> & binary,char * instr,unsigned size,unsigned pos)53 print_instr(FILE* output, const std::vector<uint32_t>& binary, char* instr, unsigned size,
54 unsigned pos)
55 {
56 fprintf(output, "%-60s ;", instr);
57
58 for (unsigned i = 0; i < size; i++)
59 fprintf(output, " %.8x", binary[pos + i]);
60 fputc('\n', output);
61 }
62
63 void
print_constant_data(FILE * output,Program * program)64 print_constant_data(FILE* output, Program* program)
65 {
66 if (program->constant_data.empty())
67 return;
68
69 fputs("\n/* constant data */\n", output);
70 for (unsigned i = 0; i < program->constant_data.size(); i += 32) {
71 fprintf(output, "[%.6u]", i);
72 unsigned line_size = std::min<size_t>(program->constant_data.size() - i, 32);
73 for (unsigned j = 0; j < line_size; j += 4) {
74 unsigned size = std::min<size_t>(program->constant_data.size() - (i + j), 4);
75 uint32_t v = 0;
76 memcpy(&v, &program->constant_data[i + j], size);
77 fprintf(output, " %.8x", v);
78 }
79 fputc('\n', output);
80 }
81 }
82
83 /**
84 * Determines the GPU type to use for CLRXdisasm
85 */
86 const char*
to_clrx_device_name(amd_gfx_level gfx_level,radeon_family family)87 to_clrx_device_name(amd_gfx_level gfx_level, radeon_family family)
88 {
89 switch (gfx_level) {
90 case GFX6:
91 switch (family) {
92 case CHIP_TAHITI: return "tahiti";
93 case CHIP_PITCAIRN: return "pitcairn";
94 case CHIP_VERDE: return "capeverde";
95 case CHIP_OLAND: return "oland";
96 case CHIP_HAINAN: return "hainan";
97 default: return nullptr;
98 }
99 case GFX7:
100 switch (family) {
101 case CHIP_BONAIRE: return "bonaire";
102 case CHIP_KAVERI: return "gfx700";
103 case CHIP_HAWAII: return "hawaii";
104 default: return nullptr;
105 }
106 case GFX8:
107 switch (family) {
108 case CHIP_TONGA: return "tonga";
109 case CHIP_ICELAND: return "iceland";
110 case CHIP_CARRIZO: return "carrizo";
111 case CHIP_FIJI: return "fiji";
112 case CHIP_STONEY: return "stoney";
113 case CHIP_POLARIS10: return "polaris10";
114 case CHIP_POLARIS11: return "polaris11";
115 case CHIP_POLARIS12: return "polaris12";
116 case CHIP_VEGAM: return "polaris11";
117 default: return nullptr;
118 }
119 case GFX9:
120 switch (family) {
121 case CHIP_VEGA10: return "vega10";
122 case CHIP_VEGA12: return "vega12";
123 case CHIP_VEGA20: return "vega20";
124 case CHIP_RAVEN: return "raven";
125 default: return nullptr;
126 }
127 case GFX10:
128 switch (family) {
129 case CHIP_NAVI10: return "gfx1010";
130 case CHIP_NAVI12: return "gfx1011";
131 default: return nullptr;
132 }
133 default: return nullptr;
134 }
135 }
136
137 bool
get_branch_target(char ** output,Program * program,const std::vector<bool> & referenced_blocks,char ** line_start)138 get_branch_target(char** output, Program* program, const std::vector<bool>& referenced_blocks,
139 char** line_start)
140 {
141 unsigned pos;
142 if (sscanf(*line_start, ".L%d_0", &pos) != 1)
143 return false;
144 pos /= 4;
145 *line_start = strchr(*line_start, '_') + 2;
146
147 for (Block& block : program->blocks) {
148 if (referenced_blocks[block.index] && block.offset == pos) {
149 *output += sprintf(*output, "BB%u", block.index);
150 return true;
151 }
152 }
153 return false;
154 }
155
156 bool
print_asm_clrx(Program * program,std::vector<uint32_t> & binary,unsigned exec_size,FILE * output)157 print_asm_clrx(Program* program, std::vector<uint32_t>& binary, unsigned exec_size, FILE* output)
158 {
159 #ifdef _WIN32
160 return true;
161 #else
162 char path[] = "/tmp/fileXXXXXX";
163 char line[2048], command[128];
164 FILE* p;
165 int fd;
166
167 const char* gpu_type = to_clrx_device_name(program->gfx_level, program->family);
168
169 /* Dump the binary into a temporary file. */
170 fd = mkstemp(path);
171 if (fd < 0)
172 return true;
173
174 for (unsigned i = 0; i < exec_size; i++) {
175 if (write(fd, &binary[i], 4) == -1)
176 goto fail;
177 }
178
179 sprintf(command, "clrxdisasm --gpuType=%s -r %s", gpu_type, path);
180
181 p = popen(command, "r");
182 if (p) {
183 if (!fgets(line, sizeof(line), p)) {
184 fprintf(output, "clrxdisasm not found\n");
185 pclose(p);
186 goto fail;
187 }
188
189 std::vector<bool> referenced_blocks = get_referenced_blocks(program);
190 unsigned next_block = 0;
191
192 char prev_instr[2048];
193 unsigned prev_pos = 0;
194 do {
195 char* line_start = line;
196 if (strncmp(line_start, "/*", 2))
197 continue;
198
199 unsigned pos;
200 if (sscanf(line_start, "/*%x*/", &pos) != 1)
201 continue;
202 pos /= 4u; /* get the dword position */
203
204 while (strncmp(line_start, "*/", 2))
205 line_start++;
206 line_start += 2;
207
208 while (line_start[0] == ' ')
209 line_start++;
210 *strchr(line_start, '\n') = 0;
211
212 if (*line_start == 0)
213 continue; /* not an instruction, only a comment */
214
215 if (pos != prev_pos) {
216 /* Print the previous instruction, now that we know the encoding size. */
217 print_instr(output, binary, prev_instr, pos - prev_pos, prev_pos);
218 prev_pos = pos;
219 }
220
221 print_block_markers(output, program, referenced_blocks, &next_block, pos);
222
223 char* dest = prev_instr;
224 *(dest++) = '\t';
225 while (*line_start) {
226 if (!strncmp(line_start, ".L", 2) &&
227 get_branch_target(&dest, program, referenced_blocks, &line_start))
228 continue;
229 *(dest++) = *(line_start++);
230 }
231 *(dest++) = 0;
232 } while (fgets(line, sizeof(line), p));
233
234 if (prev_pos != exec_size)
235 print_instr(output, binary, prev_instr, exec_size - prev_pos, prev_pos);
236
237 pclose(p);
238
239 print_constant_data(output, program);
240 }
241
242 return false;
243
244 fail:
245 close(fd);
246 unlink(path);
247 return true;
248 #endif
249 }
250
251 #if AMD_LLVM_AVAILABLE
252 std::pair<bool, size_t>
disasm_instr(amd_gfx_level gfx_level,LLVMDisasmContextRef disasm,uint32_t * binary,unsigned exec_size,size_t pos,char * outline,unsigned outline_size)253 disasm_instr(amd_gfx_level gfx_level, LLVMDisasmContextRef disasm, uint32_t* binary,
254 unsigned exec_size, size_t pos, char* outline, unsigned outline_size)
255 {
256 size_t l =
257 LLVMDisasmInstruction(disasm, (uint8_t*)&binary[pos], (exec_size - pos) * sizeof(uint32_t),
258 pos * 4, outline, outline_size);
259
260 if (gfx_level >= GFX10 && l == 8 && ((binary[pos] & 0xffff0000) == 0xd7610000) &&
261 ((binary[pos + 1] & 0x1ff) == 0xff)) {
262 /* v_writelane with literal uses 3 dwords but llvm consumes only 2 */
263 l += 4;
264 }
265
266 bool invalid = false;
267 size_t size;
268 if (!l &&
269 ((gfx_level >= GFX9 &&
270 (binary[pos] & 0xffff8000) == 0xd1348000) || /* v_add_u32_e64 + clamp */
271 (gfx_level >= GFX10 &&
272 (binary[pos] & 0xffff8000) == 0xd7038000) || /* v_add_u16_e64 + clamp */
273 (gfx_level <= GFX9 &&
274 (binary[pos] & 0xffff8000) == 0xd1268000) || /* v_add_u16_e64 + clamp */
275 (gfx_level >= GFX10 && (binary[pos] & 0xffff8000) == 0xd76d8000) || /* v_add3_u32 + clamp */
276 (gfx_level == GFX9 && (binary[pos] & 0xffff8000) == 0xd1ff8000)) /* v_add3_u32 + clamp */) {
277 strcpy(outline, "\tinteger addition + clamp");
278 bool has_literal = gfx_level >= GFX10 && (((binary[pos + 1] & 0x1ff) == 0xff) ||
279 (((binary[pos + 1] >> 9) & 0x1ff) == 0xff));
280 size = 2 + has_literal;
281 } else if (gfx_level >= GFX10 && l == 4 && ((binary[pos] & 0xfe0001ff) == 0x020000f9)) {
282 strcpy(outline, "\tv_cndmask_b32 + sdwa");
283 size = 2;
284 } else if (!l) {
285 strcpy(outline, "(invalid instruction)");
286 size = 1;
287 invalid = true;
288 } else {
289 assert(l % 4 == 0);
290 size = l / 4;
291 }
292
293 return std::make_pair(invalid, size);
294 }
295
296 bool
print_asm_llvm(Program * program,std::vector<uint32_t> & binary,unsigned exec_size,FILE * output)297 print_asm_llvm(Program* program, std::vector<uint32_t>& binary, unsigned exec_size, FILE* output)
298 {
299 std::vector<bool> referenced_blocks = get_referenced_blocks(program);
300
301 std::vector<llvm::SymbolInfoTy> symbols;
302 std::vector<std::array<char, 16>> block_names;
303 block_names.reserve(program->blocks.size());
304 for (Block& block : program->blocks) {
305 if (!referenced_blocks[block.index])
306 continue;
307 std::array<char, 16> name;
308 sprintf(name.data(), "BB%u", block.index);
309 block_names.push_back(name);
310 symbols.emplace_back(block.offset * 4,
311 llvm::StringRef(block_names[block_names.size() - 1].data()), 0);
312 }
313
314 const char* features = "";
315 if (program->gfx_level >= GFX10 && program->wave_size == 64) {
316 features = "+wavefrontsize64";
317 }
318
319 LLVMDisasmContextRef disasm =
320 LLVMCreateDisasmCPUFeatures("amdgcn-mesa-mesa3d", ac_get_llvm_processor_name(program->family),
321 features, &symbols, 0, NULL, NULL);
322
323 size_t pos = 0;
324 bool invalid = false;
325 unsigned next_block = 0;
326
327 unsigned prev_size = 0;
328 unsigned prev_pos = 0;
329 unsigned repeat_count = 0;
330 while (pos <= exec_size) {
331 bool new_block =
332 next_block < program->blocks.size() && pos == program->blocks[next_block].offset;
333 if (pos + prev_size <= exec_size && prev_pos != pos && !new_block &&
334 memcmp(&binary[prev_pos], &binary[pos], prev_size * 4) == 0) {
335 repeat_count++;
336 pos += prev_size;
337 continue;
338 } else {
339 if (repeat_count)
340 fprintf(output, "\t(then repeated %u times)\n", repeat_count);
341 repeat_count = 0;
342 }
343
344 print_block_markers(output, program, referenced_blocks, &next_block, pos);
345
346 /* For empty last block, only print block marker. */
347 if (pos == exec_size)
348 break;
349
350 char outline[1024];
351 std::pair<bool, size_t> res = disasm_instr(program->gfx_level, disasm, binary.data(),
352 exec_size, pos, outline, sizeof(outline));
353 invalid |= res.first;
354
355 print_instr(output, binary, outline, res.second, pos);
356
357 prev_size = res.second;
358 prev_pos = pos;
359 pos += res.second;
360 }
361 assert(next_block == program->blocks.size());
362
363 LLVMDisasmDispose(disasm);
364
365 print_constant_data(output, program);
366
367 return invalid;
368 }
369 #endif /* AMD_LLVM_AVAILABLE */
370
371 } /* end namespace */
372
373 bool
check_print_asm_support(Program * program)374 check_print_asm_support(Program* program)
375 {
376 #if AMD_LLVM_AVAILABLE
377 if (program->gfx_level >= GFX8) {
378 /* LLVM disassembler only supports GFX8+ */
379 const char* name = ac_get_llvm_processor_name(program->family);
380 const char* triple = "amdgcn--";
381 LLVMTargetRef target = ac_get_llvm_target(triple);
382
383 LLVMTargetMachineRef tm = LLVMCreateTargetMachine(
384 target, triple, name, "", LLVMCodeGenLevelDefault, LLVMRelocDefault, LLVMCodeModelDefault);
385
386 bool supported = ac_is_llvm_processor_supported(tm, name);
387 LLVMDisposeTargetMachine(tm);
388
389 if (supported)
390 return true;
391 }
392 #endif
393
394 #ifndef _WIN32
395 /* Check if CLRX disassembler binary is available and can disassemble the program */
396 return to_clrx_device_name(program->gfx_level, program->family) &&
397 system("clrxdisasm --version > /dev/null 2>&1") == 0;
398 #else
399 return false;
400 #endif
401 }
402
403 /* Returns true on failure */
404 bool
print_asm(Program * program,std::vector<uint32_t> & binary,unsigned exec_size,FILE * output)405 print_asm(Program* program, std::vector<uint32_t>& binary, unsigned exec_size, FILE* output)
406 {
407 #if AMD_LLVM_AVAILABLE
408 if (program->gfx_level >= GFX8) {
409 return print_asm_llvm(program, binary, exec_size, output);
410 }
411 #endif
412
413 return print_asm_clrx(program, binary, exec_size, output);
414 }
415
416 } // namespace aco
417