aco: validate the CFG
[mesa.git] / src / amd / compiler / aco_validate.cpp
1 /*
2 * Copyright © 2018 Valve Corporation
3 *
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
10 *
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
13 * Software.
14 *
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
21 * IN THE SOFTWARE.
22 *
23 */
24
25 #include "aco_ir.h"
26
27 #include <array>
28 #include <map>
29
30 namespace aco {
31
32 #ifndef NDEBUG
33 void perfwarn(bool cond, const char *msg, Instruction *instr)
34 {
35 if (cond) {
36 fprintf(stderr, "ACO performance warning: %s\n", msg);
37 if (instr) {
38 fprintf(stderr, "instruction: ");
39 aco_print_instr(instr, stderr);
40 fprintf(stderr, "\n");
41 }
42
43 if (debug_flags & DEBUG_PERFWARN)
44 exit(1);
45 }
46 }
47 #endif
48
49 void validate(Program* program, FILE * output)
50 {
51 if (!(debug_flags & DEBUG_VALIDATE))
52 return;
53
54 bool is_valid = true;
55 auto check = [&output, &is_valid](bool check, const char * msg, aco::Instruction * instr) -> void {
56 if (!check) {
57 fprintf(output, "%s: ", msg);
58 aco_print_instr(instr, output);
59 fprintf(output, "\n");
60 is_valid = false;
61 }
62 };
63 auto check_block = [&output, &is_valid](bool check, const char * msg, aco::Block * block) -> void {
64 if (!check) {
65 fprintf(output, "%s: BB%u\n", msg, block->index);
66 is_valid = false;
67 }
68 };
69
70 for (Block& block : program->blocks) {
71 for (aco_ptr<Instruction>& instr : block.instructions) {
72
73 /* check base format */
74 Format base_format = instr->format;
75 base_format = (Format)((uint32_t)base_format & ~(uint32_t)Format::SDWA);
76 base_format = (Format)((uint32_t)base_format & ~(uint32_t)Format::DPP);
77 if ((uint32_t)base_format & (uint32_t)Format::VOP1)
78 base_format = Format::VOP1;
79 else if ((uint32_t)base_format & (uint32_t)Format::VOP2)
80 base_format = Format::VOP2;
81 else if ((uint32_t)base_format & (uint32_t)Format::VOPC)
82 base_format = Format::VOPC;
83 else if ((uint32_t)base_format & (uint32_t)Format::VINTRP)
84 base_format = Format::VINTRP;
85 check(base_format == instr_info.format[(int)instr->opcode], "Wrong base format for instruction", instr.get());
86
87 /* check VOP3 modifiers */
88 if (((uint32_t)instr->format & (uint32_t)Format::VOP3) && instr->format != Format::VOP3) {
89 check(base_format == Format::VOP2 ||
90 base_format == Format::VOP1 ||
91 base_format == Format::VOPC ||
92 base_format == Format::VINTRP,
93 "Format cannot have VOP3A/VOP3B applied", instr.get());
94 }
95
96 /* check for undefs */
97 for (unsigned i = 0; i < instr->operands.size(); i++) {
98 if (instr->operands[i].isUndefined()) {
99 bool flat = instr->format == Format::FLAT || instr->format == Format::SCRATCH || instr->format == Format::GLOBAL;
100 bool can_be_undef = is_phi(instr) || instr->format == Format::EXP ||
101 instr->format == Format::PSEUDO_REDUCTION ||
102 (flat && i == 1) || (instr->format == Format::MIMG && i == 2) ||
103 ((instr->format == Format::MUBUF || instr->format == Format::MTBUF) && i == 0);
104 check(can_be_undef, "Undefs can only be used in certain operands", instr.get());
105 }
106 }
107
108 /* check num literals */
109 if (instr->isSALU() || instr->isVALU()) {
110 unsigned num_literals = 0;
111 for (unsigned i = 0; i < instr->operands.size(); i++)
112 {
113 if (instr->operands[i].isLiteral() && instr->isVOP3() && program->chip_class >= GFX10) {
114 num_literals++;
115 } else if (instr->operands[i].isLiteral()) {
116 check(instr->format == Format::SOP1 ||
117 instr->format == Format::SOP2 ||
118 instr->format == Format::SOPC ||
119 instr->format == Format::VOP1 ||
120 instr->format == Format::VOP2 ||
121 instr->format == Format::VOPC,
122 "Literal applied on wrong instruction format", instr.get());
123
124 num_literals++;
125 check(!instr->isVALU() || i == 0 || i == 2, "Wrong source position for Literal argument", instr.get());
126 }
127 }
128 check(num_literals <= 1, "Only 1 Literal allowed", instr.get());
129
130 /* check num sgprs for VALU */
131 if (instr->isVALU()) {
132 check(instr->definitions[0].getTemp().type() == RegType::vgpr ||
133 (int) instr->format & (int) Format::VOPC ||
134 instr->opcode == aco_opcode::v_readfirstlane_b32 ||
135 instr->opcode == aco_opcode::v_readlane_b32,
136 "Wrong Definition type for VALU instruction", instr.get());
137 unsigned num_sgpr = 0;
138 unsigned sgpr_idx = instr->operands.size();
139 for (unsigned i = 0; i < instr->operands.size(); i++)
140 {
141 if (instr->operands[i].isTemp() && instr->operands[i].regClass().type() == RegType::sgpr) {
142 check(i != 1 || (int) instr->format & (int) Format::VOP3A, "Wrong source position for SGPR argument", instr.get());
143
144 if (sgpr_idx == instr->operands.size() || instr->operands[sgpr_idx].tempId() != instr->operands[i].tempId())
145 num_sgpr++;
146 sgpr_idx = i;
147 }
148
149 if (instr->operands[i].isConstant() && !instr->operands[i].isLiteral())
150 check(i == 0 || (int) instr->format & (int) Format::VOP3A, "Wrong source position for constant argument", instr.get());
151 }
152 check(num_sgpr + num_literals <= 1, "Only 1 Literal OR 1 SGPR allowed", instr.get());
153 }
154
155 if (instr->format == Format::SOP1 || instr->format == Format::SOP2) {
156 check(instr->definitions[0].getTemp().type() == RegType::sgpr, "Wrong Definition type for SALU instruction", instr.get());
157 for (const Operand& op : instr->operands) {
158 check(op.isConstant() || op.regClass().type() <= RegType::sgpr,
159 "Wrong Operand type for SALU instruction", instr.get());
160 }
161 }
162 }
163
164 switch (instr->format) {
165 case Format::PSEUDO: {
166 if (instr->opcode == aco_opcode::p_create_vector) {
167 unsigned size = 0;
168 for (const Operand& op : instr->operands) {
169 size += op.size();
170 }
171 check(size == instr->definitions[0].size(), "Definition size does not match operand sizes", instr.get());
172 if (instr->definitions[0].getTemp().type() == RegType::sgpr) {
173 for (const Operand& op : instr->operands) {
174 check(op.isConstant() || op.regClass().type() == RegType::sgpr,
175 "Wrong Operand type for scalar vector", instr.get());
176 }
177 }
178 } else if (instr->opcode == aco_opcode::p_extract_vector) {
179 check((instr->operands[0].isTemp()) && instr->operands[1].isConstant(), "Wrong Operand types", instr.get());
180 check(instr->operands[1].constantValue() < instr->operands[0].size(), "Index out of range", instr.get());
181 check(instr->definitions[0].getTemp().type() == RegType::vgpr || instr->operands[0].regClass().type() == RegType::sgpr,
182 "Cannot extract SGPR value from VGPR vector", instr.get());
183 } else if (instr->opcode == aco_opcode::p_parallelcopy) {
184 check(instr->definitions.size() == instr->operands.size(), "Number of Operands does not match number of Definitions", instr.get());
185 for (unsigned i = 0; i < instr->operands.size(); i++) {
186 if (instr->operands[i].isTemp())
187 check((instr->definitions[i].getTemp().type() == instr->operands[i].regClass().type()) ||
188 (instr->definitions[i].getTemp().type() == RegType::vgpr && instr->operands[i].regClass().type() == RegType::sgpr),
189 "Operand and Definition types do not match", instr.get());
190 }
191 } else if (instr->opcode == aco_opcode::p_phi) {
192 check(instr->operands.size() == block.logical_preds.size(), "Number of Operands does not match number of predecessors", instr.get());
193 check(instr->definitions[0].getTemp().type() == RegType::vgpr || instr->definitions[0].getTemp().regClass() == s2, "Logical Phi Definition must be vgpr or divergent boolean", instr.get());
194 } else if (instr->opcode == aco_opcode::p_linear_phi) {
195 for (const Operand& op : instr->operands)
196 check(!op.isTemp() || op.getTemp().is_linear(), "Wrong Operand type", instr.get());
197 check(instr->operands.size() == block.linear_preds.size(), "Number of Operands does not match number of predecessors", instr.get());
198 }
199 break;
200 }
201 case Format::SMEM: {
202 if (instr->operands.size() >= 1)
203 check(instr->operands[0].isTemp() && instr->operands[0].regClass().type() == RegType::sgpr, "SMEM operands must be sgpr", instr.get());
204 if (instr->operands.size() >= 2)
205 check(instr->operands[1].isConstant() || (instr->operands[1].isTemp() && instr->operands[1].regClass().type() == RegType::sgpr),
206 "SMEM offset must be constant or sgpr", instr.get());
207 if (!instr->definitions.empty())
208 check(instr->definitions[0].getTemp().type() == RegType::sgpr, "SMEM result must be sgpr", instr.get());
209 break;
210 }
211 case Format::MTBUF:
212 case Format::MUBUF:
213 case Format::MIMG: {
214 check(instr->operands.size() > 1, "VMEM instructions must have at least one operand", instr.get());
215 check(instr->operands[0].hasRegClass() && instr->operands[0].regClass().type() == RegType::vgpr,
216 "VADDR must be in vgpr for VMEM instructions", instr.get());
217 check(instr->operands[1].isTemp() && instr->operands[1].regClass().type() == RegType::sgpr, "VMEM resource constant must be sgpr", instr.get());
218 check(instr->operands.size() < 4 || (instr->operands[3].isTemp() && instr->operands[3].regClass().type() == RegType::vgpr), "VMEM write data must be vgpr", instr.get());
219 break;
220 }
221 case Format::DS: {
222 for (const Operand& op : instr->operands) {
223 check((op.isTemp() && op.regClass().type() == RegType::vgpr) || op.physReg() == m0,
224 "Only VGPRs are valid DS instruction operands", instr.get());
225 }
226 if (!instr->definitions.empty())
227 check(instr->definitions[0].getTemp().type() == RegType::vgpr, "DS instruction must return VGPR", instr.get());
228 break;
229 }
230 case Format::EXP: {
231 for (unsigned i = 0; i < 4; i++)
232 check(instr->operands[i].hasRegClass() && instr->operands[i].regClass().type() == RegType::vgpr,
233 "Only VGPRs are valid Export arguments", instr.get());
234 break;
235 }
236 case Format::FLAT:
237 check(instr->operands[1].isUndefined(), "Flat instructions don't support SADDR", instr.get());
238 /* fallthrough */
239 case Format::GLOBAL:
240 case Format::SCRATCH: {
241 check(instr->operands[0].isTemp() && instr->operands[0].regClass().type() == RegType::vgpr, "FLAT/GLOBAL/SCRATCH address must be vgpr", instr.get());
242 check(instr->operands[1].hasRegClass() && instr->operands[1].regClass().type() == RegType::sgpr,
243 "FLAT/GLOBAL/SCRATCH sgpr address must be undefined or sgpr", instr.get());
244 if (!instr->definitions.empty())
245 check(instr->definitions[0].getTemp().type() == RegType::vgpr, "FLAT/GLOBAL/SCRATCH result must be vgpr", instr.get());
246 else
247 check(instr->operands[2].regClass().type() == RegType::vgpr, "FLAT/GLOBAL/SCRATCH data must be vgpr", instr.get());
248 break;
249 }
250 default:
251 break;
252 }
253 }
254 }
255
256 /* validate CFG */
257 for (unsigned i = 0; i < program->blocks.size(); i++) {
258 Block& block = program->blocks[i];
259 check_block(block.index == i, "block.index must match actual index", &block);
260
261 /* predecessors/successors should be sorted */
262 for (unsigned j = 0; j + 1 < block.linear_preds.size(); j++)
263 check_block(block.linear_preds[j] < block.linear_preds[j + 1], "linear predecessors must be sorted", &block);
264 for (unsigned j = 0; j + 1 < block.logical_preds.size(); j++)
265 check_block(block.logical_preds[j] < block.logical_preds[j + 1], "logical predecessors must be sorted", &block);
266 for (unsigned j = 0; j + 1 < block.linear_succs.size(); j++)
267 check_block(block.linear_succs[j] < block.linear_succs[j + 1], "linear successors must be sorted", &block);
268 for (unsigned j = 0; j + 1 < block.logical_succs.size(); j++)
269 check_block(block.logical_succs[j] < block.logical_succs[j + 1], "logical successors must be sorted", &block);
270
271 /* critical edges are not allowed */
272 if (block.linear_preds.size() > 1) {
273 for (unsigned pred : block.linear_preds)
274 check_block(program->blocks[pred].linear_succs.size() == 1, "linear critical edges are not allowed", &program->blocks[pred]);
275 for (unsigned pred : block.logical_preds)
276 check_block(program->blocks[pred].logical_succs.size() == 1, "logical critical edges are not allowed", &program->blocks[pred]);
277 }
278 }
279
280 assert(is_valid);
281 }
282
283 /* RA validation */
284 namespace {
285
286 struct Location {
287 Location() : block(NULL), instr(NULL) {}
288
289 Block *block;
290 Instruction *instr; //NULL if it's the block's live-in
291 };
292
293 struct Assignment {
294 Location defloc;
295 Location firstloc;
296 PhysReg reg;
297 };
298
299 bool ra_fail(FILE *output, Location loc, Location loc2, const char *fmt, ...) {
300 va_list args;
301 va_start(args, fmt);
302 char msg[1024];
303 vsprintf(msg, fmt, args);
304 va_end(args);
305
306 fprintf(stderr, "RA error found at instruction in BB%d:\n", loc.block->index);
307 if (loc.instr) {
308 aco_print_instr(loc.instr, stderr);
309 fprintf(stderr, "\n%s", msg);
310 } else {
311 fprintf(stderr, "%s", msg);
312 }
313 if (loc2.block) {
314 fprintf(stderr, " in BB%d:\n", loc2.block->index);
315 aco_print_instr(loc2.instr, stderr);
316 }
317 fprintf(stderr, "\n\n");
318
319 return true;
320 }
321
322 } /* end namespace */
323
324 bool validate_ra(Program *program, const struct radv_nir_compiler_options *options, FILE *output) {
325 if (!(debug_flags & DEBUG_VALIDATE_RA))
326 return false;
327
328 bool err = false;
329 aco::live live_vars = aco::live_var_analysis(program, options);
330 std::vector<std::vector<Temp>> phi_sgpr_ops(program->blocks.size());
331
332 std::map<unsigned, Assignment> assignments;
333 for (Block& block : program->blocks) {
334 Location loc;
335 loc.block = &block;
336 for (aco_ptr<Instruction>& instr : block.instructions) {
337 if (instr->opcode == aco_opcode::p_phi) {
338 for (unsigned i = 0; i < instr->operands.size(); i++) {
339 if (instr->operands[i].isTemp() &&
340 instr->operands[i].getTemp().type() == RegType::sgpr &&
341 instr->operands[i].isFirstKill())
342 phi_sgpr_ops[block.logical_preds[i]].emplace_back(instr->operands[i].getTemp());
343 }
344 }
345
346 loc.instr = instr.get();
347 for (unsigned i = 0; i < instr->operands.size(); i++) {
348 Operand& op = instr->operands[i];
349 if (!op.isTemp())
350 continue;
351 if (!op.isFixed())
352 err |= ra_fail(output, loc, Location(), "Operand %d is not assigned a register", i);
353 if (assignments.count(op.tempId()) && assignments[op.tempId()].reg != op.physReg())
354 err |= ra_fail(output, loc, assignments.at(op.tempId()).firstloc, "Operand %d has an inconsistent register assignment with instruction", i);
355 if ((op.getTemp().type() == RegType::vgpr && op.physReg() + op.size() > 256 + program->config->num_vgprs) ||
356 (op.getTemp().type() == RegType::sgpr && op.physReg() + op.size() > program->config->num_sgprs && op.physReg() < program->sgpr_limit))
357 err |= ra_fail(output, loc, assignments.at(op.tempId()).firstloc, "Operand %d has an out-of-bounds register assignment", i);
358 if (!assignments[op.tempId()].firstloc.block)
359 assignments[op.tempId()].firstloc = loc;
360 if (!assignments[op.tempId()].defloc.block)
361 assignments[op.tempId()].reg = op.physReg();
362 }
363
364 for (unsigned i = 0; i < instr->definitions.size(); i++) {
365 Definition& def = instr->definitions[i];
366 if (!def.isTemp())
367 continue;
368 if (!def.isFixed())
369 err |= ra_fail(output, loc, Location(), "Definition %d is not assigned a register", i);
370 if (assignments[def.tempId()].defloc.block)
371 err |= ra_fail(output, loc, assignments.at(def.tempId()).defloc, "Temporary %%%d also defined by instruction", def.tempId());
372 if ((def.getTemp().type() == RegType::vgpr && def.physReg() + def.size() > 256 + program->config->num_vgprs) ||
373 (def.getTemp().type() == RegType::sgpr && def.physReg() + def.size() > program->config->num_sgprs && def.physReg() < program->sgpr_limit))
374 err |= ra_fail(output, loc, assignments.at(def.tempId()).firstloc, "Definition %d has an out-of-bounds register assignment", i);
375 if (!assignments[def.tempId()].firstloc.block)
376 assignments[def.tempId()].firstloc = loc;
377 assignments[def.tempId()].defloc = loc;
378 assignments[def.tempId()].reg = def.physReg();
379 }
380 }
381 }
382
383 for (Block& block : program->blocks) {
384 Location loc;
385 loc.block = &block;
386
387 std::array<unsigned, 512> regs;
388 regs.fill(0);
389
390 std::set<Temp> live;
391 live.insert(live_vars.live_out[block.index].begin(), live_vars.live_out[block.index].end());
392 /* remove killed p_phi sgpr operands */
393 for (Temp tmp : phi_sgpr_ops[block.index])
394 live.erase(tmp);
395
396 /* check live out */
397 for (Temp tmp : live) {
398 PhysReg reg = assignments.at(tmp.id()).reg;
399 for (unsigned i = 0; i < tmp.size(); i++) {
400 if (regs[reg + i]) {
401 err |= ra_fail(output, loc, Location(), "Assignment of element %d of %%%d already taken by %%%d in live-out", i, tmp.id(), regs[reg + i]);
402 }
403 regs[reg + i] = tmp.id();
404 }
405 }
406 regs.fill(0);
407
408 for (auto it = block.instructions.rbegin(); it != block.instructions.rend(); ++it) {
409 aco_ptr<Instruction>& instr = *it;
410
411 /* check killed p_phi sgpr operands */
412 if (instr->opcode == aco_opcode::p_logical_end) {
413 for (Temp tmp : phi_sgpr_ops[block.index]) {
414 PhysReg reg = assignments.at(tmp.id()).reg;
415 for (unsigned i = 0; i < tmp.size(); i++) {
416 if (regs[reg + i])
417 err |= ra_fail(output, loc, Location(), "Assignment of element %d of %%%d already taken by %%%d in live-out", i, tmp.id(), regs[reg + i]);
418 }
419 live.emplace(tmp);
420 }
421 }
422
423 for (const Definition& def : instr->definitions) {
424 if (!def.isTemp())
425 continue;
426 live.erase(def.getTemp());
427 }
428
429 /* don't count phi operands as live-in, since they are actually
430 * killed when they are copied at the predecessor */
431 if (instr->opcode != aco_opcode::p_phi && instr->opcode != aco_opcode::p_linear_phi) {
432 for (const Operand& op : instr->operands) {
433 if (!op.isTemp())
434 continue;
435 live.insert(op.getTemp());
436 }
437 }
438 }
439
440 for (Temp tmp : live) {
441 PhysReg reg = assignments.at(tmp.id()).reg;
442 for (unsigned i = 0; i < tmp.size(); i++)
443 regs[reg + i] = tmp.id();
444 }
445
446 for (aco_ptr<Instruction>& instr : block.instructions) {
447 loc.instr = instr.get();
448
449 /* remove killed p_phi operands from regs */
450 if (instr->opcode == aco_opcode::p_logical_end) {
451 for (Temp tmp : phi_sgpr_ops[block.index]) {
452 PhysReg reg = assignments.at(tmp.id()).reg;
453 regs[reg] = 0;
454 }
455 }
456
457 if (instr->opcode != aco_opcode::p_phi && instr->opcode != aco_opcode::p_linear_phi) {
458 for (const Operand& op : instr->operands) {
459 if (!op.isTemp())
460 continue;
461 if (op.isFirstKill()) {
462 for (unsigned j = 0; j < op.getTemp().size(); j++)
463 regs[op.physReg() + j] = 0;
464 }
465 }
466 }
467
468 for (unsigned i = 0; i < instr->definitions.size(); i++) {
469 Definition& def = instr->definitions[i];
470 if (!def.isTemp())
471 continue;
472 Temp tmp = def.getTemp();
473 PhysReg reg = assignments.at(tmp.id()).reg;
474 for (unsigned j = 0; j < tmp.size(); j++) {
475 if (regs[reg + j])
476 err |= ra_fail(output, loc, assignments.at(regs[reg + i]).defloc, "Assignment of element %d of %%%d already taken by %%%d from instruction", i, tmp.id(), regs[reg + j]);
477 regs[reg + j] = tmp.id();
478 }
479 }
480
481 for (const Definition& def : instr->definitions) {
482 if (!def.isTemp())
483 continue;
484 if (def.isKill()) {
485 for (unsigned j = 0; j < def.getTemp().size(); j++)
486 regs[def.physReg() + j] = 0;
487 }
488 }
489 }
490 }
491
492 return err;
493 }
494 }