1 #include "aco_ir.h"
2 #include "aco_builder.h"
3
4 #include "sid.h"
5 #include "ac_shader_util.h"
6
7 namespace aco {
8
9 static const char *reduce_ops[] = {
10 [iadd8] = "iadd8",
11 [iadd16] = "iadd16",
12 [iadd32] = "iadd32",
13 [iadd64] = "iadd64",
14 [imul8] = "imul8",
15 [imul16] = "imul16",
16 [imul32] = "imul32",
17 [imul64] = "imul64",
18 [fadd16] = "fadd16",
19 [fadd32] = "fadd32",
20 [fadd64] = "fadd64",
21 [fmul16] = "fmul16",
22 [fmul32] = "fmul32",
23 [fmul64] = "fmul64",
24 [imin8] = "imin8",
25 [imin16] = "imin16",
26 [imin32] = "imin32",
27 [imin64] = "imin64",
28 [imax8] = "imax8",
29 [imax16] = "imax16",
30 [imax32] = "imax32",
31 [imax64] = "imax64",
32 [umin8] = "umin8",
33 [umin16] = "umin16",
34 [umin32] = "umin32",
35 [umin64] = "umin64",
36 [umax8] = "umax8",
37 [umax16] = "umax16",
38 [umax32] = "umax32",
39 [umax64] = "umax64",
40 [fmin16] = "fmin16",
41 [fmin32] = "fmin32",
42 [fmin64] = "fmin64",
43 [fmax16] = "fmax16",
44 [fmax32] = "fmax32",
45 [fmax64] = "fmax64",
46 [iand8] = "iand8",
47 [iand16] = "iand16",
48 [iand32] = "iand32",
49 [iand64] = "iand64",
50 [ior8] = "ior8",
51 [ior16] = "ior16",
52 [ior32] = "ior32",
53 [ior64] = "ior64",
54 [ixor8] = "ixor8",
55 [ixor16] = "ixor16",
56 [ixor32] = "ixor32",
57 [ixor64] = "ixor64",
58 };
59
print_reg_class(const RegClass rc,FILE * output)60 static void print_reg_class(const RegClass rc, FILE *output)
61 {
62 switch (rc) {
63 case RegClass::s1: fprintf(output, " s1: "); return;
64 case RegClass::s2: fprintf(output, " s2: "); return;
65 case RegClass::s3: fprintf(output, " s3: "); return;
66 case RegClass::s4: fprintf(output, " s4: "); return;
67 case RegClass::s6: fprintf(output, " s6: "); return;
68 case RegClass::s8: fprintf(output, " s8: "); return;
69 case RegClass::s16: fprintf(output, "s16: "); return;
70 case RegClass::v1: fprintf(output, " v1: "); return;
71 case RegClass::v2: fprintf(output, " v2: "); return;
72 case RegClass::v3: fprintf(output, " v3: "); return;
73 case RegClass::v4: fprintf(output, " v4: "); return;
74 case RegClass::v5: fprintf(output, " v5: "); return;
75 case RegClass::v6: fprintf(output, " v6: "); return;
76 case RegClass::v7: fprintf(output, " v7: "); return;
77 case RegClass::v8: fprintf(output, " v8: "); return;
78 case RegClass::v1b: fprintf(output, " v1b: "); return;
79 case RegClass::v2b: fprintf(output, " v2b: "); return;
80 case RegClass::v3b: fprintf(output, " v3b: "); return;
81 case RegClass::v4b: fprintf(output, " v4b: "); return;
82 case RegClass::v6b: fprintf(output, " v6b: "); return;
83 case RegClass::v8b: fprintf(output, " v8b: "); return;
84 case RegClass::v1_linear: fprintf(output, " v1: "); return;
85 case RegClass::v2_linear: fprintf(output, " v2: "); return;
86 }
87 }
88
print_physReg(PhysReg reg,unsigned bytes,FILE * output)89 void print_physReg(PhysReg reg, unsigned bytes, FILE *output)
90 {
91 if (reg == 124) {
92 fprintf(output, ":m0");
93 } else if (reg == 106) {
94 fprintf(output, ":vcc");
95 } else if (reg == 253) {
96 fprintf(output, ":scc");
97 } else if (reg == 126) {
98 fprintf(output, ":exec");
99 } else {
100 bool is_vgpr = reg / 256;
101 unsigned r = reg % 256;
102 unsigned size = DIV_ROUND_UP(bytes, 4);
103 fprintf(output, ":%c[%d", is_vgpr ? 'v' : 's', r);
104 if (size > 1)
105 fprintf(output, "-%d]", r + size -1);
106 else
107 fprintf(output, "]");
108 if (reg.byte() || bytes % 4)
109 fprintf(output, "[%d:%d]", reg.byte()*8, (reg.byte()+bytes) * 8);
110 }
111 }
112
print_constant(uint8_t reg,FILE * output)113 static void print_constant(uint8_t reg, FILE *output)
114 {
115 if (reg >= 128 && reg <= 192) {
116 fprintf(output, "%d", reg - 128);
117 return;
118 } else if (reg >= 192 && reg <= 208) {
119 fprintf(output, "%d", 192 - reg);
120 return;
121 }
122
123 switch (reg) {
124 case 240:
125 fprintf(output, "0.5");
126 break;
127 case 241:
128 fprintf(output, "-0.5");
129 break;
130 case 242:
131 fprintf(output, "1.0");
132 break;
133 case 243:
134 fprintf(output, "-1.0");
135 break;
136 case 244:
137 fprintf(output, "2.0");
138 break;
139 case 245:
140 fprintf(output, "-2.0");
141 break;
142 case 246:
143 fprintf(output, "4.0");
144 break;
145 case 247:
146 fprintf(output, "-4.0");
147 break;
148 case 248:
149 fprintf(output, "1/(2*PI)");
150 break;
151 }
152 }
153
print_operand(const Operand * operand,FILE * output)154 static void print_operand(const Operand *operand, FILE *output)
155 {
156 if (operand->isLiteral() || (operand->isConstant() && operand->bytes() == 1)) {
157 if (operand->bytes() == 1)
158 fprintf(output, "0x%.2x", operand->constantValue());
159 else if (operand->bytes() == 2)
160 fprintf(output, "0x%.4x", operand->constantValue());
161 else
162 fprintf(output, "0x%x", operand->constantValue());
163 } else if (operand->isConstant()) {
164 print_constant(operand->physReg().reg(), output);
165 } else if (operand->isUndefined()) {
166 print_reg_class(operand->regClass(), output);
167 fprintf(output, "undef");
168 } else {
169 if (operand->isLateKill())
170 fprintf(output, "(latekill)");
171
172 fprintf(output, "%%%d", operand->tempId());
173
174 if (operand->isFixed())
175 print_physReg(operand->physReg(), operand->bytes(), output);
176 }
177 }
178
print_definition(const Definition * definition,FILE * output)179 static void print_definition(const Definition *definition, FILE *output)
180 {
181 print_reg_class(definition->regClass(), output);
182 if (definition->isPrecise())
183 fprintf(output, "(precise)");
184 if (definition->isNUW())
185 fprintf(output, "(nuw)");
186 fprintf(output, "%%%d", definition->tempId());
187
188 if (definition->isFixed())
189 print_physReg(definition->physReg(), definition->bytes(), output);
190 }
191
print_storage(storage_class storage,FILE * output)192 static void print_storage(storage_class storage, FILE *output)
193 {
194 fprintf(output, " storage:");
195 int printed = 0;
196 if (storage & storage_buffer)
197 printed += fprintf(output, "%sbuffer", printed ? "," : "");
198 if (storage & storage_atomic_counter)
199 printed += fprintf(output, "%satomic_counter", printed ? "," : "");
200 if (storage & storage_image)
201 printed += fprintf(output, "%simage", printed ? "," : "");
202 if (storage & storage_shared)
203 printed += fprintf(output, "%sshared", printed ? "," : "");
204 if (storage & storage_vmem_output)
205 printed += fprintf(output, "%svmem_output", printed ? "," : "");
206 if (storage & storage_scratch)
207 printed += fprintf(output, "%sscratch", printed ? "," : "");
208 if (storage & storage_vgpr_spill)
209 printed += fprintf(output, "%svgpr_spill", printed ? "," : "");
210 }
211
print_semantics(memory_semantics sem,FILE * output)212 static void print_semantics(memory_semantics sem, FILE *output)
213 {
214 fprintf(output, " semantics:");
215 int printed = 0;
216 if (sem & semantic_acquire)
217 printed += fprintf(output, "%sacquire", printed ? "," : "");
218 if (sem & semantic_release)
219 printed += fprintf(output, "%srelease", printed ? "," : "");
220 if (sem & semantic_volatile)
221 printed += fprintf(output, "%svolatile", printed ? "," : "");
222 if (sem & semantic_private)
223 printed += fprintf(output, "%sprivate", printed ? "," : "");
224 if (sem & semantic_can_reorder)
225 printed += fprintf(output, "%sreorder", printed ? "," : "");
226 if (sem & semantic_atomic)
227 printed += fprintf(output, "%satomic", printed ? "," : "");
228 if (sem & semantic_rmw)
229 printed += fprintf(output, "%srmw", printed ? "," : "");
230 }
231
print_scope(sync_scope scope,FILE * output,const char * prefix="scope")232 static void print_scope(sync_scope scope, FILE *output, const char *prefix="scope")
233 {
234 fprintf(output, " %s:", prefix);
235 switch (scope) {
236 case scope_invocation:
237 fprintf(output, "invocation");
238 break;
239 case scope_subgroup:
240 fprintf(output, "subgroup");
241 break;
242 case scope_workgroup:
243 fprintf(output, "workgroup");
244 break;
245 case scope_queuefamily:
246 fprintf(output, "queuefamily");
247 break;
248 case scope_device:
249 fprintf(output, "device");
250 break;
251 }
252 }
253
print_sync(memory_sync_info sync,FILE * output)254 static void print_sync(memory_sync_info sync, FILE *output)
255 {
256 print_storage(sync.storage, output);
257 print_semantics(sync.semantics, output);
258 print_scope(sync.scope, output);
259 }
260
print_instr_format_specific(const Instruction * instr,FILE * output)261 static void print_instr_format_specific(const Instruction *instr, FILE *output)
262 {
263 switch (instr->format) {
264 case Format::SOPK: {
265 const SOPK_instruction* sopk = static_cast<const SOPK_instruction*>(instr);
266 fprintf(output, " imm:%d", sopk->imm & 0x8000 ? (sopk->imm - 65536) : sopk->imm);
267 break;
268 }
269 case Format::SOPP: {
270 const SOPP_instruction* sopp = static_cast<const SOPP_instruction*>(instr);
271 uint16_t imm = sopp->imm;
272 switch (instr->opcode) {
273 case aco_opcode::s_waitcnt: {
274 /* we usually should check the chip class for vmcnt/lgkm, but
275 * insert_waitcnt() should fill it in regardless. */
276 unsigned vmcnt = (imm & 0xF) | ((imm & (0x3 << 14)) >> 10);
277 if (vmcnt != 63) fprintf(output, " vmcnt(%d)", vmcnt);
278 if (((imm >> 4) & 0x7) < 0x7) fprintf(output, " expcnt(%d)", (imm >> 4) & 0x7);
279 if (((imm >> 8) & 0x3F) < 0x3F) fprintf(output, " lgkmcnt(%d)", (imm >> 8) & 0x3F);
280 break;
281 }
282 case aco_opcode::s_endpgm:
283 case aco_opcode::s_endpgm_saved:
284 case aco_opcode::s_endpgm_ordered_ps_done:
285 case aco_opcode::s_wakeup:
286 case aco_opcode::s_barrier:
287 case aco_opcode::s_icache_inv:
288 case aco_opcode::s_ttracedata:
289 case aco_opcode::s_set_gpr_idx_off: {
290 break;
291 }
292 case aco_opcode::s_sendmsg: {
293 unsigned id = imm & sendmsg_id_mask;
294 switch (id) {
295 case sendmsg_none:
296 fprintf(output, " sendmsg(MSG_NONE)");
297 break;
298 case _sendmsg_gs:
299 fprintf(output, " sendmsg(gs%s%s, %u)",
300 imm & 0x10 ? ", cut" : "", imm & 0x20 ? ", emit" : "", imm >> 8);
301 break;
302 case _sendmsg_gs_done:
303 fprintf(output, " sendmsg(gs_done%s%s, %u)",
304 imm & 0x10 ? ", cut" : "", imm & 0x20 ? ", emit" : "", imm >> 8);
305 break;
306 case sendmsg_save_wave:
307 fprintf(output, " sendmsg(save_wave)");
308 break;
309 case sendmsg_stall_wave_gen:
310 fprintf(output, " sendmsg(stall_wave_gen)");
311 break;
312 case sendmsg_halt_waves:
313 fprintf(output, " sendmsg(halt_waves)");
314 break;
315 case sendmsg_ordered_ps_done:
316 fprintf(output, " sendmsg(ordered_ps_done)");
317 break;
318 case sendmsg_early_prim_dealloc:
319 fprintf(output, " sendmsg(early_prim_dealloc)");
320 break;
321 case sendmsg_gs_alloc_req:
322 fprintf(output, " sendmsg(gs_alloc_req)");
323 break;
324 }
325 break;
326 }
327 default: {
328 if (imm)
329 fprintf(output, " imm:%u", imm);
330 break;
331 }
332 }
333 if (sopp->block != -1)
334 fprintf(output, " block:BB%d", sopp->block);
335 break;
336 }
337 case Format::SMEM: {
338 const SMEM_instruction* smem = static_cast<const SMEM_instruction*>(instr);
339 if (smem->glc)
340 fprintf(output, " glc");
341 if (smem->dlc)
342 fprintf(output, " dlc");
343 if (smem->nv)
344 fprintf(output, " nv");
345 print_sync(smem->sync, output);
346 break;
347 }
348 case Format::VINTRP: {
349 const Interp_instruction* vintrp = static_cast<const Interp_instruction*>(instr);
350 fprintf(output, " attr%d.%c", vintrp->attribute, "xyzw"[vintrp->component]);
351 break;
352 }
353 case Format::DS: {
354 const DS_instruction* ds = static_cast<const DS_instruction*>(instr);
355 if (ds->offset0)
356 fprintf(output, " offset0:%u", ds->offset0);
357 if (ds->offset1)
358 fprintf(output, " offset1:%u", ds->offset1);
359 if (ds->gds)
360 fprintf(output, " gds");
361 print_sync(ds->sync, output);
362 break;
363 }
364 case Format::MUBUF: {
365 const MUBUF_instruction* mubuf = static_cast<const MUBUF_instruction*>(instr);
366 if (mubuf->offset)
367 fprintf(output, " offset:%u", mubuf->offset);
368 if (mubuf->offen)
369 fprintf(output, " offen");
370 if (mubuf->idxen)
371 fprintf(output, " idxen");
372 if (mubuf->addr64)
373 fprintf(output, " addr64");
374 if (mubuf->glc)
375 fprintf(output, " glc");
376 if (mubuf->dlc)
377 fprintf(output, " dlc");
378 if (mubuf->slc)
379 fprintf(output, " slc");
380 if (mubuf->tfe)
381 fprintf(output, " tfe");
382 if (mubuf->lds)
383 fprintf(output, " lds");
384 if (mubuf->disable_wqm)
385 fprintf(output, " disable_wqm");
386 print_sync(mubuf->sync, output);
387 break;
388 }
389 case Format::MIMG: {
390 const MIMG_instruction* mimg = static_cast<const MIMG_instruction*>(instr);
391 unsigned identity_dmask = !instr->definitions.empty() ?
392 (1 << instr->definitions[0].size()) - 1 :
393 0xf;
394 if ((mimg->dmask & identity_dmask) != identity_dmask)
395 fprintf(output, " dmask:%s%s%s%s",
396 mimg->dmask & 0x1 ? "x" : "",
397 mimg->dmask & 0x2 ? "y" : "",
398 mimg->dmask & 0x4 ? "z" : "",
399 mimg->dmask & 0x8 ? "w" : "");
400 switch (mimg->dim) {
401 case ac_image_1d:
402 fprintf(output, " 1d");
403 break;
404 case ac_image_2d:
405 fprintf(output, " 2d");
406 break;
407 case ac_image_3d:
408 fprintf(output, " 3d");
409 break;
410 case ac_image_cube:
411 fprintf(output, " cube");
412 break;
413 case ac_image_1darray:
414 fprintf(output, " 1darray");
415 break;
416 case ac_image_2darray:
417 fprintf(output, " 2darray");
418 break;
419 case ac_image_2dmsaa:
420 fprintf(output, " 2dmsaa");
421 break;
422 case ac_image_2darraymsaa:
423 fprintf(output, " 2darraymsaa");
424 break;
425 }
426 if (mimg->unrm)
427 fprintf(output, " unrm");
428 if (mimg->glc)
429 fprintf(output, " glc");
430 if (mimg->dlc)
431 fprintf(output, " dlc");
432 if (mimg->slc)
433 fprintf(output, " slc");
434 if (mimg->tfe)
435 fprintf(output, " tfe");
436 if (mimg->da)
437 fprintf(output, " da");
438 if (mimg->lwe)
439 fprintf(output, " lwe");
440 if (mimg->r128 || mimg->a16)
441 fprintf(output, " r128/a16");
442 if (mimg->d16)
443 fprintf(output, " d16");
444 if (mimg->disable_wqm)
445 fprintf(output, " disable_wqm");
446 print_sync(mimg->sync, output);
447 break;
448 }
449 case Format::EXP: {
450 const Export_instruction* exp = static_cast<const Export_instruction*>(instr);
451 unsigned identity_mask = exp->compressed ? 0x5 : 0xf;
452 if ((exp->enabled_mask & identity_mask) != identity_mask)
453 fprintf(output, " en:%c%c%c%c",
454 exp->enabled_mask & 0x1 ? 'r' : '*',
455 exp->enabled_mask & 0x2 ? 'g' : '*',
456 exp->enabled_mask & 0x4 ? 'b' : '*',
457 exp->enabled_mask & 0x8 ? 'a' : '*');
458 if (exp->compressed)
459 fprintf(output, " compr");
460 if (exp->done)
461 fprintf(output, " done");
462 if (exp->valid_mask)
463 fprintf(output, " vm");
464
465 if (exp->dest <= V_008DFC_SQ_EXP_MRT + 7)
466 fprintf(output, " mrt%d", exp->dest - V_008DFC_SQ_EXP_MRT);
467 else if (exp->dest == V_008DFC_SQ_EXP_MRTZ)
468 fprintf(output, " mrtz");
469 else if (exp->dest == V_008DFC_SQ_EXP_NULL)
470 fprintf(output, " null");
471 else if (exp->dest >= V_008DFC_SQ_EXP_POS && exp->dest <= V_008DFC_SQ_EXP_POS + 3)
472 fprintf(output, " pos%d", exp->dest - V_008DFC_SQ_EXP_POS);
473 else if (exp->dest >= V_008DFC_SQ_EXP_PARAM && exp->dest <= V_008DFC_SQ_EXP_PARAM + 31)
474 fprintf(output, " param%d", exp->dest - V_008DFC_SQ_EXP_PARAM);
475 break;
476 }
477 case Format::PSEUDO_BRANCH: {
478 const Pseudo_branch_instruction* branch = static_cast<const Pseudo_branch_instruction*>(instr);
479 /* Note: BB0 cannot be a branch target */
480 if (branch->target[0] != 0)
481 fprintf(output, " BB%d", branch->target[0]);
482 if (branch->target[1] != 0)
483 fprintf(output, ", BB%d", branch->target[1]);
484 break;
485 }
486 case Format::PSEUDO_REDUCTION: {
487 const Pseudo_reduction_instruction* reduce = static_cast<const Pseudo_reduction_instruction*>(instr);
488 fprintf(output, " op:%s", reduce_ops[reduce->reduce_op]);
489 if (reduce->cluster_size)
490 fprintf(output, " cluster_size:%u", reduce->cluster_size);
491 break;
492 }
493 case Format::PSEUDO_BARRIER: {
494 const Pseudo_barrier_instruction* barrier = static_cast<const Pseudo_barrier_instruction*>(instr);
495 print_sync(barrier->sync, output);
496 print_scope(barrier->exec_scope, output, "exec_scope");
497 break;
498 }
499 case Format::FLAT:
500 case Format::GLOBAL:
501 case Format::SCRATCH: {
502 const FLAT_instruction* flat = static_cast<const FLAT_instruction*>(instr);
503 if (flat->offset)
504 fprintf(output, " offset:%u", flat->offset);
505 if (flat->glc)
506 fprintf(output, " glc");
507 if (flat->dlc)
508 fprintf(output, " dlc");
509 if (flat->slc)
510 fprintf(output, " slc");
511 if (flat->lds)
512 fprintf(output, " lds");
513 if (flat->nv)
514 fprintf(output, " nv");
515 if (flat->disable_wqm)
516 fprintf(output, " disable_wqm");
517 print_sync(flat->sync, output);
518 break;
519 }
520 case Format::MTBUF: {
521 const MTBUF_instruction* mtbuf = static_cast<const MTBUF_instruction*>(instr);
522 fprintf(output, " dfmt:");
523 switch (mtbuf->dfmt) {
524 case V_008F0C_BUF_DATA_FORMAT_8: fprintf(output, "8"); break;
525 case V_008F0C_BUF_DATA_FORMAT_16: fprintf(output, "16"); break;
526 case V_008F0C_BUF_DATA_FORMAT_8_8: fprintf(output, "8_8"); break;
527 case V_008F0C_BUF_DATA_FORMAT_32: fprintf(output, "32"); break;
528 case V_008F0C_BUF_DATA_FORMAT_16_16: fprintf(output, "16_16"); break;
529 case V_008F0C_BUF_DATA_FORMAT_10_11_11: fprintf(output, "10_11_11"); break;
530 case V_008F0C_BUF_DATA_FORMAT_11_11_10: fprintf(output, "11_11_10"); break;
531 case V_008F0C_BUF_DATA_FORMAT_10_10_10_2: fprintf(output, "10_10_10_2"); break;
532 case V_008F0C_BUF_DATA_FORMAT_2_10_10_10: fprintf(output, "2_10_10_10"); break;
533 case V_008F0C_BUF_DATA_FORMAT_8_8_8_8: fprintf(output, "8_8_8_8"); break;
534 case V_008F0C_BUF_DATA_FORMAT_32_32: fprintf(output, "32_32"); break;
535 case V_008F0C_BUF_DATA_FORMAT_16_16_16_16: fprintf(output, "16_16_16_16"); break;
536 case V_008F0C_BUF_DATA_FORMAT_32_32_32: fprintf(output, "32_32_32"); break;
537 case V_008F0C_BUF_DATA_FORMAT_32_32_32_32: fprintf(output, "32_32_32_32"); break;
538 case V_008F0C_BUF_DATA_FORMAT_RESERVED_15: fprintf(output, "reserved15"); break;
539 }
540 fprintf(output, " nfmt:");
541 switch (mtbuf->nfmt) {
542 case V_008F0C_BUF_NUM_FORMAT_UNORM: fprintf(output, "unorm"); break;
543 case V_008F0C_BUF_NUM_FORMAT_SNORM: fprintf(output, "snorm"); break;
544 case V_008F0C_BUF_NUM_FORMAT_USCALED: fprintf(output, "uscaled"); break;
545 case V_008F0C_BUF_NUM_FORMAT_SSCALED: fprintf(output, "sscaled"); break;
546 case V_008F0C_BUF_NUM_FORMAT_UINT: fprintf(output, "uint"); break;
547 case V_008F0C_BUF_NUM_FORMAT_SINT: fprintf(output, "sint"); break;
548 case V_008F0C_BUF_NUM_FORMAT_SNORM_OGL: fprintf(output, "snorm"); break;
549 case V_008F0C_BUF_NUM_FORMAT_FLOAT: fprintf(output, "float"); break;
550 }
551 if (mtbuf->offset)
552 fprintf(output, " offset:%u", mtbuf->offset);
553 if (mtbuf->offen)
554 fprintf(output, " offen");
555 if (mtbuf->idxen)
556 fprintf(output, " idxen");
557 if (mtbuf->glc)
558 fprintf(output, " glc");
559 if (mtbuf->dlc)
560 fprintf(output, " dlc");
561 if (mtbuf->slc)
562 fprintf(output, " slc");
563 if (mtbuf->tfe)
564 fprintf(output, " tfe");
565 if (mtbuf->disable_wqm)
566 fprintf(output, " disable_wqm");
567 print_sync(mtbuf->sync, output);
568 break;
569 }
570 case Format::VOP3P: {
571 if (static_cast<const VOP3P_instruction*>(instr)->clamp)
572 fprintf(output, " clamp");
573 break;
574 }
575 default: {
576 break;
577 }
578 }
579 if (instr->isVOP3()) {
580 const VOP3A_instruction* vop3 = static_cast<const VOP3A_instruction*>(instr);
581 switch (vop3->omod) {
582 case 1:
583 fprintf(output, " *2");
584 break;
585 case 2:
586 fprintf(output, " *4");
587 break;
588 case 3:
589 fprintf(output, " *0.5");
590 break;
591 }
592 if (vop3->clamp)
593 fprintf(output, " clamp");
594 if (vop3->opsel & (1 << 3))
595 fprintf(output, " opsel_hi");
596 } else if (instr->isDPP()) {
597 const DPP_instruction* dpp = static_cast<const DPP_instruction*>(instr);
598 if (dpp->dpp_ctrl <= 0xff) {
599 fprintf(output, " quad_perm:[%d,%d,%d,%d]",
600 dpp->dpp_ctrl & 0x3, (dpp->dpp_ctrl >> 2) & 0x3,
601 (dpp->dpp_ctrl >> 4) & 0x3, (dpp->dpp_ctrl >> 6) & 0x3);
602 } else if (dpp->dpp_ctrl >= 0x101 && dpp->dpp_ctrl <= 0x10f) {
603 fprintf(output, " row_shl:%d", dpp->dpp_ctrl & 0xf);
604 } else if (dpp->dpp_ctrl >= 0x111 && dpp->dpp_ctrl <= 0x11f) {
605 fprintf(output, " row_shr:%d", dpp->dpp_ctrl & 0xf);
606 } else if (dpp->dpp_ctrl >= 0x121 && dpp->dpp_ctrl <= 0x12f) {
607 fprintf(output, " row_ror:%d", dpp->dpp_ctrl & 0xf);
608 } else if (dpp->dpp_ctrl == dpp_wf_sl1) {
609 fprintf(output, " wave_shl:1");
610 } else if (dpp->dpp_ctrl == dpp_wf_rl1) {
611 fprintf(output, " wave_rol:1");
612 } else if (dpp->dpp_ctrl == dpp_wf_sr1) {
613 fprintf(output, " wave_shr:1");
614 } else if (dpp->dpp_ctrl == dpp_wf_rr1) {
615 fprintf(output, " wave_ror:1");
616 } else if (dpp->dpp_ctrl == dpp_row_mirror) {
617 fprintf(output, " row_mirror");
618 } else if (dpp->dpp_ctrl == dpp_row_half_mirror) {
619 fprintf(output, " row_half_mirror");
620 } else if (dpp->dpp_ctrl == dpp_row_bcast15) {
621 fprintf(output, " row_bcast:15");
622 } else if (dpp->dpp_ctrl == dpp_row_bcast31) {
623 fprintf(output, " row_bcast:31");
624 } else {
625 fprintf(output, " dpp_ctrl:0x%.3x", dpp->dpp_ctrl);
626 }
627 if (dpp->row_mask != 0xf)
628 fprintf(output, " row_mask:0x%.1x", dpp->row_mask);
629 if (dpp->bank_mask != 0xf)
630 fprintf(output, " bank_mask:0x%.1x", dpp->bank_mask);
631 if (dpp->bound_ctrl)
632 fprintf(output, " bound_ctrl:1");
633 } else if ((int)instr->format & (int)Format::SDWA) {
634 const SDWA_instruction* sdwa = static_cast<const SDWA_instruction*>(instr);
635 switch (sdwa->omod) {
636 case 1:
637 fprintf(output, " *2");
638 break;
639 case 2:
640 fprintf(output, " *4");
641 break;
642 case 3:
643 fprintf(output, " *0.5");
644 break;
645 }
646 if (sdwa->clamp)
647 fprintf(output, " clamp");
648 switch (sdwa->dst_sel & sdwa_asuint) {
649 case sdwa_udword:
650 break;
651 case sdwa_ubyte0:
652 case sdwa_ubyte1:
653 case sdwa_ubyte2:
654 case sdwa_ubyte3:
655 fprintf(output, " dst_sel:%sbyte%u", sdwa->dst_sel & sdwa_sext ? "s" : "u",
656 sdwa->dst_sel & sdwa_bytenum);
657 break;
658 case sdwa_uword0:
659 case sdwa_uword1:
660 fprintf(output, " dst_sel:%sword%u", sdwa->dst_sel & sdwa_sext ? "s" : "u",
661 sdwa->dst_sel & sdwa_wordnum);
662 break;
663 }
664 if (sdwa->dst_preserve)
665 fprintf(output, " dst_preserve");
666 }
667 }
668
aco_print_instr(const Instruction * instr,FILE * output)669 void aco_print_instr(const Instruction *instr, FILE *output)
670 {
671 if (!instr->definitions.empty()) {
672 for (unsigned i = 0; i < instr->definitions.size(); ++i) {
673 print_definition(&instr->definitions[i], output);
674 if (i + 1 != instr->definitions.size())
675 fprintf(output, ", ");
676 }
677 fprintf(output, " = ");
678 }
679 fprintf(output, "%s", instr_info.name[(int)instr->opcode]);
680 if (instr->operands.size()) {
681 bool *const abs = (bool *)alloca(instr->operands.size() * sizeof(bool));
682 bool *const neg = (bool *)alloca(instr->operands.size() * sizeof(bool));
683 bool *const opsel = (bool *)alloca(instr->operands.size() * sizeof(bool));
684 uint8_t *const sel = (uint8_t *)alloca(instr->operands.size() * sizeof(uint8_t));
685 if ((int)instr->format & (int)Format::VOP3A) {
686 const VOP3A_instruction* vop3 = static_cast<const VOP3A_instruction*>(instr);
687 for (unsigned i = 0; i < instr->operands.size(); ++i) {
688 abs[i] = vop3->abs[i];
689 neg[i] = vop3->neg[i];
690 opsel[i] = vop3->opsel & (1 << i);
691 sel[i] = sdwa_udword;
692 }
693 } else if (instr->isDPP()) {
694 const DPP_instruction* dpp = static_cast<const DPP_instruction*>(instr);
695 for (unsigned i = 0; i < instr->operands.size(); ++i) {
696 abs[i] = i < 2 ? dpp->abs[i] : false;
697 neg[i] = i < 2 ? dpp->neg[i] : false;
698 opsel[i] = false;
699 sel[i] = sdwa_udword;
700 }
701 } else if (instr->isSDWA()) {
702 const SDWA_instruction* sdwa = static_cast<const SDWA_instruction*>(instr);
703 for (unsigned i = 0; i < instr->operands.size(); ++i) {
704 abs[i] = i < 2 ? sdwa->abs[i] : false;
705 neg[i] = i < 2 ? sdwa->neg[i] : false;
706 opsel[i] = false;
707 sel[i] = i < 2 ? sdwa->sel[i] : sdwa_udword;
708 }
709 } else {
710 for (unsigned i = 0; i < instr->operands.size(); ++i) {
711 abs[i] = false;
712 neg[i] = false;
713 opsel[i] = false;
714 sel[i] = sdwa_udword;
715 }
716 }
717 for (unsigned i = 0; i < instr->operands.size(); ++i) {
718 if (i)
719 fprintf(output, ", ");
720 else
721 fprintf(output, " ");
722
723 if (neg[i])
724 fprintf(output, "-");
725 if (abs[i])
726 fprintf(output, "|");
727 if (opsel[i])
728 fprintf(output, "hi(");
729 else if (sel[i] & sdwa_sext)
730 fprintf(output, "sext(");
731 print_operand(&instr->operands[i], output);
732 if (opsel[i] || (sel[i] & sdwa_sext))
733 fprintf(output, ")");
734 if (!(sel[i] & sdwa_isra)) {
735 if (sel[i] == sdwa_udword || sel[i] == sdwa_sdword) {
736 /* print nothing */
737 } else if (sel[i] & sdwa_isword) {
738 unsigned index = sel[i] & sdwa_wordnum;
739 fprintf(output, "[%u:%u]", index * 16, index * 16 + 15);
740 } else {
741 unsigned index = sel[i] & sdwa_bytenum;
742 fprintf(output, "[%u:%u]", index * 8, index * 8 + 7);
743 }
744 }
745 if (abs[i])
746 fprintf(output, "|");
747
748 if (instr->format == Format::VOP3P) {
749 const VOP3P_instruction* vop3 = static_cast<const VOP3P_instruction*>(instr);
750 if ((vop3->opsel_lo & (1 << i)) || !(vop3->opsel_hi & (1 << i))) {
751 fprintf(output, ".%c%c",
752 vop3->opsel_lo & (1 << i) ? 'y' : 'x',
753 vop3->opsel_hi & (1 << i) ? 'y' : 'x');
754 }
755 if (vop3->neg_lo[i] && vop3->neg_hi[i])
756 fprintf(output, "*[-1,-1]");
757 else if (vop3->neg_lo[i])
758 fprintf(output, "*[-1,1]");
759 else if (vop3->neg_hi[i])
760 fprintf(output, "*[1,-1]");
761 }
762 }
763 }
764 print_instr_format_specific(instr, output);
765 }
766
print_block_kind(uint16_t kind,FILE * output)767 static void print_block_kind(uint16_t kind, FILE *output)
768 {
769 if (kind & block_kind_uniform)
770 fprintf(output, "uniform, ");
771 if (kind & block_kind_top_level)
772 fprintf(output, "top-level, ");
773 if (kind & block_kind_loop_preheader)
774 fprintf(output, "loop-preheader, ");
775 if (kind & block_kind_loop_header)
776 fprintf(output, "loop-header, ");
777 if (kind & block_kind_loop_exit)
778 fprintf(output, "loop-exit, ");
779 if (kind & block_kind_continue)
780 fprintf(output, "continue, ");
781 if (kind & block_kind_break)
782 fprintf(output, "break, ");
783 if (kind & block_kind_continue_or_break)
784 fprintf(output, "continue_or_break, ");
785 if (kind & block_kind_discard)
786 fprintf(output, "discard, ");
787 if (kind & block_kind_branch)
788 fprintf(output, "branch, ");
789 if (kind & block_kind_merge)
790 fprintf(output, "merge, ");
791 if (kind & block_kind_invert)
792 fprintf(output, "invert, ");
793 if (kind & block_kind_uses_discard_if)
794 fprintf(output, "discard_if, ");
795 if (kind & block_kind_needs_lowering)
796 fprintf(output, "needs_lowering, ");
797 if (kind & block_kind_uses_demote)
798 fprintf(output, "uses_demote, ");
799 if (kind & block_kind_export_end)
800 fprintf(output, "export_end, ");
801 }
802
print_stage(Stage stage,FILE * output)803 static void print_stage(Stage stage, FILE *output)
804 {
805 fprintf(output, "ACO shader stage: ");
806
807 if (stage == compute_cs)
808 fprintf(output, "compute_cs");
809 else if (stage == fragment_fs)
810 fprintf(output, "fragment_fs");
811 else if (stage == gs_copy_vs)
812 fprintf(output, "gs_copy_vs");
813 else if (stage == vertex_ls)
814 fprintf(output, "vertex_ls");
815 else if (stage == vertex_es)
816 fprintf(output, "vertex_es");
817 else if (stage == vertex_vs)
818 fprintf(output, "vertex_vs");
819 else if (stage == tess_control_hs)
820 fprintf(output, "tess_control_hs");
821 else if (stage == vertex_tess_control_hs)
822 fprintf(output, "vertex_tess_control_hs");
823 else if (stage == tess_eval_es)
824 fprintf(output, "tess_eval_es");
825 else if (stage == tess_eval_vs)
826 fprintf(output, "tess_eval_vs");
827 else if (stage == geometry_gs)
828 fprintf(output, "geometry_gs");
829 else if (stage == vertex_geometry_gs)
830 fprintf(output, "vertex_geometry_gs");
831 else if (stage == tess_eval_geometry_gs)
832 fprintf(output, "tess_eval_geometry_gs");
833 else if (stage == vertex_ngg)
834 fprintf(output, "vertex_ngg");
835 else if (stage == tess_eval_ngg)
836 fprintf(output, "tess_eval_ngg");
837 else if (stage == vertex_geometry_ngg)
838 fprintf(output, "vertex_geometry_ngg");
839 else if (stage == tess_eval_geometry_ngg)
840 fprintf(output, "tess_eval_geometry_ngg");
841 else
842 fprintf(output, "unknown");
843
844 fprintf(output, "\n");
845 }
846
aco_print_block(const Block * block,FILE * output)847 void aco_print_block(const Block* block, FILE *output)
848 {
849 fprintf(output, "BB%d\n", block->index);
850 fprintf(output, "/* logical preds: ");
851 for (unsigned pred : block->logical_preds)
852 fprintf(output, "BB%d, ", pred);
853 fprintf(output, "/ linear preds: ");
854 for (unsigned pred : block->linear_preds)
855 fprintf(output, "BB%d, ", pred);
856 fprintf(output, "/ kind: ");
857 print_block_kind(block->kind, output);
858 fprintf(output, "*/\n");
859 for (auto const& instr : block->instructions) {
860 fprintf(output, "\t");
861 aco_print_instr(instr.get(), output);
862 fprintf(output, "\n");
863 }
864 }
865
aco_print_program(const Program * program,FILE * output)866 void aco_print_program(const Program *program, FILE *output)
867 {
868 print_stage(program->stage, output);
869
870 for (Block const& block : program->blocks)
871 aco_print_block(&block, output);
872
873 if (program->constant_data.size()) {
874 fprintf(output, "\n/* constant data */\n");
875 for (unsigned i = 0; i < program->constant_data.size(); i += 32) {
876 fprintf(output, "[%06d] ", i);
877 unsigned line_size = std::min<size_t>(program->constant_data.size() - i, 32);
878 for (unsigned j = 0; j < line_size; j += 4) {
879 unsigned size = std::min<size_t>(program->constant_data.size() - (i + j), 4);
880 uint32_t v = 0;
881 memcpy(&v, &program->constant_data[i + j], size);
882 fprintf(output, " %08x", v);
883 }
884 fprintf(output, "\n");
885 }
886 }
887
888 fprintf(output, "\n");
889 }
890
891 }
892