1 /*
2 * Copyright © 2015 Intel 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 #include "vtn_private.h"
25 #include "nir/nir_vla.h"
26
27 static bool
vtn_cfg_handle_prepass_instruction(struct vtn_builder * b,SpvOp opcode,const uint32_t * w,unsigned count)28 vtn_cfg_handle_prepass_instruction(struct vtn_builder *b, SpvOp opcode,
29 const uint32_t *w, unsigned count)
30 {
31 switch (opcode) {
32 case SpvOpFunction: {
33 assert(b->func == NULL);
34 b->func = rzalloc(b, struct vtn_function);
35
36 list_inithead(&b->func->body);
37 b->func->control = w[3];
38
39 MAYBE_UNUSED const struct glsl_type *result_type =
40 vtn_value(b, w[1], vtn_value_type_type)->type->type;
41 struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_function);
42 val->func = b->func;
43
44 const struct glsl_type *func_type =
45 vtn_value(b, w[4], vtn_value_type_type)->type->type;
46
47 assert(glsl_get_function_return_type(func_type) == result_type);
48
49 nir_function *func =
50 nir_function_create(b->shader, ralloc_strdup(b->shader, val->name));
51
52 func->num_params = glsl_get_length(func_type);
53 func->params = ralloc_array(b->shader, nir_parameter, func->num_params);
54 for (unsigned i = 0; i < func->num_params; i++) {
55 const struct glsl_function_param *param =
56 glsl_get_function_param(func_type, i);
57 func->params[i].type = param->type;
58 if (param->in) {
59 if (param->out) {
60 func->params[i].param_type = nir_parameter_inout;
61 } else {
62 func->params[i].param_type = nir_parameter_in;
63 }
64 } else {
65 if (param->out) {
66 func->params[i].param_type = nir_parameter_out;
67 } else {
68 assert(!"Parameter is neither in nor out");
69 }
70 }
71 }
72
73 func->return_type = glsl_get_function_return_type(func_type);
74
75 b->func->impl = nir_function_impl_create(func);
76
77 b->func_param_idx = 0;
78 break;
79 }
80
81 case SpvOpFunctionEnd:
82 b->func->end = w;
83 b->func = NULL;
84 break;
85
86 case SpvOpFunctionParameter: {
87 struct vtn_value *val =
88 vtn_push_value(b, w[2], vtn_value_type_access_chain);
89
90 struct vtn_type *type = vtn_value(b, w[1], vtn_value_type_type)->type;
91
92 assert(b->func_param_idx < b->func->impl->num_params);
93 nir_variable *param = b->func->impl->params[b->func_param_idx++];
94
95 assert(param->type == type->type);
96
97 /* Name the parameter so it shows up nicely in NIR */
98 param->name = ralloc_strdup(param, val->name);
99
100 struct vtn_variable *vtn_var = rzalloc(b, struct vtn_variable);
101 vtn_var->type = type;
102 vtn_var->var = param;
103 vtn_var->chain.var = vtn_var;
104 vtn_var->chain.length = 0;
105
106 struct vtn_type *without_array = type;
107 while(glsl_type_is_array(without_array->type))
108 without_array = without_array->array_element;
109
110 if (glsl_type_is_image(without_array->type)) {
111 vtn_var->mode = vtn_variable_mode_image;
112 param->interface_type = without_array->type;
113 } else if (glsl_type_is_sampler(without_array->type)) {
114 vtn_var->mode = vtn_variable_mode_sampler;
115 param->interface_type = without_array->type;
116 } else {
117 vtn_var->mode = vtn_variable_mode_param;
118 }
119
120 val->access_chain = &vtn_var->chain;
121 break;
122 }
123
124 case SpvOpLabel: {
125 assert(b->block == NULL);
126 b->block = rzalloc(b, struct vtn_block);
127 b->block->node.type = vtn_cf_node_type_block;
128 b->block->label = w;
129 vtn_push_value(b, w[1], vtn_value_type_block)->block = b->block;
130
131 if (b->func->start_block == NULL) {
132 /* This is the first block encountered for this function. In this
133 * case, we set the start block and add it to the list of
134 * implemented functions that we'll walk later.
135 */
136 b->func->start_block = b->block;
137 exec_list_push_tail(&b->functions, &b->func->node);
138 }
139 break;
140 }
141
142 case SpvOpSelectionMerge:
143 case SpvOpLoopMerge:
144 assert(b->block && b->block->merge == NULL);
145 b->block->merge = w;
146 break;
147
148 case SpvOpBranch:
149 case SpvOpBranchConditional:
150 case SpvOpSwitch:
151 case SpvOpKill:
152 case SpvOpReturn:
153 case SpvOpReturnValue:
154 case SpvOpUnreachable:
155 assert(b->block && b->block->branch == NULL);
156 b->block->branch = w;
157 b->block = NULL;
158 break;
159
160 default:
161 /* Continue on as per normal */
162 return true;
163 }
164
165 return true;
166 }
167
168 static void
vtn_add_case(struct vtn_builder * b,struct vtn_switch * swtch,struct vtn_block * break_block,uint32_t block_id,uint32_t val,bool is_default)169 vtn_add_case(struct vtn_builder *b, struct vtn_switch *swtch,
170 struct vtn_block *break_block,
171 uint32_t block_id, uint32_t val, bool is_default)
172 {
173 struct vtn_block *case_block =
174 vtn_value(b, block_id, vtn_value_type_block)->block;
175
176 /* Don't create dummy cases that just break */
177 if (case_block == break_block)
178 return;
179
180 if (case_block->switch_case == NULL) {
181 struct vtn_case *c = ralloc(b, struct vtn_case);
182
183 list_inithead(&c->body);
184 c->start_block = case_block;
185 c->fallthrough = NULL;
186 nir_array_init(&c->values, b);
187 c->is_default = false;
188 c->visited = false;
189
190 list_addtail(&c->link, &swtch->cases);
191
192 case_block->switch_case = c;
193 }
194
195 if (is_default) {
196 case_block->switch_case->is_default = true;
197 } else {
198 nir_array_add(&case_block->switch_case->values, uint32_t, val);
199 }
200 }
201
202 /* This function performs a depth-first search of the cases and puts them
203 * in fall-through order.
204 */
205 static void
vtn_order_case(struct vtn_switch * swtch,struct vtn_case * cse)206 vtn_order_case(struct vtn_switch *swtch, struct vtn_case *cse)
207 {
208 if (cse->visited)
209 return;
210
211 cse->visited = true;
212
213 list_del(&cse->link);
214
215 if (cse->fallthrough) {
216 vtn_order_case(swtch, cse->fallthrough);
217
218 /* If we have a fall-through, place this case right before the case it
219 * falls through to. This ensures that fallthroughs come one after
220 * the other. These two can never get separated because that would
221 * imply something else falling through to the same case. Also, this
222 * can't break ordering because the DFS ensures that this case is
223 * visited before anything that falls through to it.
224 */
225 list_addtail(&cse->link, &cse->fallthrough->link);
226 } else {
227 list_add(&cse->link, &swtch->cases);
228 }
229 }
230
231 static enum vtn_branch_type
vtn_get_branch_type(struct vtn_block * block,struct vtn_case * swcase,struct vtn_block * switch_break,struct vtn_block * loop_break,struct vtn_block * loop_cont)232 vtn_get_branch_type(struct vtn_block *block,
233 struct vtn_case *swcase, struct vtn_block *switch_break,
234 struct vtn_block *loop_break, struct vtn_block *loop_cont)
235 {
236 if (block->switch_case) {
237 /* This branch is actually a fallthrough */
238 assert(swcase->fallthrough == NULL ||
239 swcase->fallthrough == block->switch_case);
240 swcase->fallthrough = block->switch_case;
241 return vtn_branch_type_switch_fallthrough;
242 } else if (block == loop_break) {
243 return vtn_branch_type_loop_break;
244 } else if (block == loop_cont) {
245 return vtn_branch_type_loop_continue;
246 } else if (block == switch_break) {
247 return vtn_branch_type_switch_break;
248 } else {
249 return vtn_branch_type_none;
250 }
251 }
252
253 static void
vtn_cfg_walk_blocks(struct vtn_builder * b,struct list_head * cf_list,struct vtn_block * start,struct vtn_case * switch_case,struct vtn_block * switch_break,struct vtn_block * loop_break,struct vtn_block * loop_cont,struct vtn_block * end)254 vtn_cfg_walk_blocks(struct vtn_builder *b, struct list_head *cf_list,
255 struct vtn_block *start, struct vtn_case *switch_case,
256 struct vtn_block *switch_break,
257 struct vtn_block *loop_break, struct vtn_block *loop_cont,
258 struct vtn_block *end)
259 {
260 struct vtn_block *block = start;
261 while (block != end) {
262 if (block->merge && (*block->merge & SpvOpCodeMask) == SpvOpLoopMerge &&
263 !block->loop) {
264 struct vtn_loop *loop = ralloc(b, struct vtn_loop);
265
266 loop->node.type = vtn_cf_node_type_loop;
267 list_inithead(&loop->body);
268 list_inithead(&loop->cont_body);
269 loop->control = block->merge[3];
270
271 list_addtail(&loop->node.link, cf_list);
272 block->loop = loop;
273
274 struct vtn_block *new_loop_break =
275 vtn_value(b, block->merge[1], vtn_value_type_block)->block;
276 struct vtn_block *new_loop_cont =
277 vtn_value(b, block->merge[2], vtn_value_type_block)->block;
278
279 /* Note: This recursive call will start with the current block as
280 * its start block. If we weren't careful, we would get here
281 * again and end up in infinite recursion. This is why we set
282 * block->loop above and check for it before creating one. This
283 * way, we only create the loop once and the second call that
284 * tries to handle this loop goes to the cases below and gets
285 * handled as a regular block.
286 *
287 * Note: When we make the recursive walk calls, we pass NULL for
288 * the switch break since you have to break out of the loop first.
289 * We do, however, still pass the current switch case because it's
290 * possible that the merge block for the loop is the start of
291 * another case.
292 */
293 vtn_cfg_walk_blocks(b, &loop->body, block, switch_case, NULL,
294 new_loop_break, new_loop_cont, NULL );
295 vtn_cfg_walk_blocks(b, &loop->cont_body, new_loop_cont, NULL, NULL,
296 new_loop_break, NULL, block);
297
298 block = new_loop_break;
299 continue;
300 }
301
302 assert(block->node.link.next == NULL);
303 list_addtail(&block->node.link, cf_list);
304
305 switch (*block->branch & SpvOpCodeMask) {
306 case SpvOpBranch: {
307 struct vtn_block *branch_block =
308 vtn_value(b, block->branch[1], vtn_value_type_block)->block;
309
310 block->branch_type = vtn_get_branch_type(branch_block,
311 switch_case, switch_break,
312 loop_break, loop_cont);
313
314 if (block->branch_type != vtn_branch_type_none)
315 return;
316
317 block = branch_block;
318 continue;
319 }
320
321 case SpvOpReturn:
322 case SpvOpReturnValue:
323 block->branch_type = vtn_branch_type_return;
324 return;
325
326 case SpvOpKill:
327 block->branch_type = vtn_branch_type_discard;
328 return;
329
330 case SpvOpBranchConditional: {
331 struct vtn_block *then_block =
332 vtn_value(b, block->branch[2], vtn_value_type_block)->block;
333 struct vtn_block *else_block =
334 vtn_value(b, block->branch[3], vtn_value_type_block)->block;
335
336 struct vtn_if *if_stmt = ralloc(b, struct vtn_if);
337
338 if_stmt->node.type = vtn_cf_node_type_if;
339 if_stmt->condition = block->branch[1];
340 list_inithead(&if_stmt->then_body);
341 list_inithead(&if_stmt->else_body);
342
343 list_addtail(&if_stmt->node.link, cf_list);
344
345 if (block->merge &&
346 (*block->merge & SpvOpCodeMask) == SpvOpSelectionMerge) {
347 if_stmt->control = block->merge[2];
348 }
349
350 if_stmt->then_type = vtn_get_branch_type(then_block,
351 switch_case, switch_break,
352 loop_break, loop_cont);
353 if_stmt->else_type = vtn_get_branch_type(else_block,
354 switch_case, switch_break,
355 loop_break, loop_cont);
356
357 if (if_stmt->then_type == vtn_branch_type_none &&
358 if_stmt->else_type == vtn_branch_type_none) {
359 /* Neither side of the if is something we can short-circuit. */
360 assert((*block->merge & SpvOpCodeMask) == SpvOpSelectionMerge);
361 struct vtn_block *merge_block =
362 vtn_value(b, block->merge[1], vtn_value_type_block)->block;
363
364 vtn_cfg_walk_blocks(b, &if_stmt->then_body, then_block,
365 switch_case, switch_break,
366 loop_break, loop_cont, merge_block);
367 vtn_cfg_walk_blocks(b, &if_stmt->else_body, else_block,
368 switch_case, switch_break,
369 loop_break, loop_cont, merge_block);
370
371 enum vtn_branch_type merge_type =
372 vtn_get_branch_type(merge_block, switch_case, switch_break,
373 loop_break, loop_cont);
374 if (merge_type == vtn_branch_type_none) {
375 block = merge_block;
376 continue;
377 } else {
378 return;
379 }
380 } else if (if_stmt->then_type != vtn_branch_type_none &&
381 if_stmt->else_type != vtn_branch_type_none) {
382 /* Both sides were short-circuited. We're done here. */
383 return;
384 } else {
385 /* Exeactly one side of the branch could be short-circuited.
386 * We set the branch up as a predicated break/continue and we
387 * continue on with the other side as if it were what comes
388 * after the if.
389 */
390 if (if_stmt->then_type == vtn_branch_type_none) {
391 block = then_block;
392 } else {
393 block = else_block;
394 }
395 continue;
396 }
397 unreachable("Should have returned or continued");
398 }
399
400 case SpvOpSwitch: {
401 assert((*block->merge & SpvOpCodeMask) == SpvOpSelectionMerge);
402 struct vtn_block *break_block =
403 vtn_value(b, block->merge[1], vtn_value_type_block)->block;
404
405 struct vtn_switch *swtch = ralloc(b, struct vtn_switch);
406
407 swtch->node.type = vtn_cf_node_type_switch;
408 swtch->selector = block->branch[1];
409 list_inithead(&swtch->cases);
410
411 list_addtail(&swtch->node.link, cf_list);
412
413 /* First, we go through and record all of the cases. */
414 const uint32_t *branch_end =
415 block->branch + (block->branch[0] >> SpvWordCountShift);
416
417 vtn_add_case(b, swtch, break_block, block->branch[2], 0, true);
418 for (const uint32_t *w = block->branch + 3; w < branch_end; w += 2)
419 vtn_add_case(b, swtch, break_block, w[1], w[0], false);
420
421 /* Now, we go through and walk the blocks. While we walk through
422 * the blocks, we also gather the much-needed fall-through
423 * information.
424 */
425 list_for_each_entry(struct vtn_case, cse, &swtch->cases, link) {
426 assert(cse->start_block != break_block);
427 vtn_cfg_walk_blocks(b, &cse->body, cse->start_block, cse,
428 break_block, NULL, loop_cont, NULL);
429 }
430
431 /* Finally, we walk over all of the cases one more time and put
432 * them in fall-through order.
433 */
434 for (const uint32_t *w = block->branch + 2; w < branch_end; w += 2) {
435 struct vtn_block *case_block =
436 vtn_value(b, *w, vtn_value_type_block)->block;
437
438 if (case_block == break_block)
439 continue;
440
441 assert(case_block->switch_case);
442
443 vtn_order_case(swtch, case_block->switch_case);
444 }
445
446 enum vtn_branch_type branch_type =
447 vtn_get_branch_type(break_block, switch_case, NULL,
448 loop_break, loop_cont);
449
450 if (branch_type != vtn_branch_type_none) {
451 /* It is possible that the break is actually the continue block
452 * for the containing loop. In this case, we need to bail and let
453 * the loop parsing code handle the continue properly.
454 */
455 assert(branch_type == vtn_branch_type_loop_continue);
456 return;
457 }
458
459 block = break_block;
460 continue;
461 }
462
463 case SpvOpUnreachable:
464 return;
465
466 default:
467 unreachable("Unhandled opcode");
468 }
469 }
470 }
471
472 void
vtn_build_cfg(struct vtn_builder * b,const uint32_t * words,const uint32_t * end)473 vtn_build_cfg(struct vtn_builder *b, const uint32_t *words, const uint32_t *end)
474 {
475 vtn_foreach_instruction(b, words, end,
476 vtn_cfg_handle_prepass_instruction);
477
478 foreach_list_typed(struct vtn_function, func, node, &b->functions) {
479 vtn_cfg_walk_blocks(b, &func->body, func->start_block,
480 NULL, NULL, NULL, NULL, NULL);
481 }
482 }
483
484 static bool
vtn_handle_phis_first_pass(struct vtn_builder * b,SpvOp opcode,const uint32_t * w,unsigned count)485 vtn_handle_phis_first_pass(struct vtn_builder *b, SpvOp opcode,
486 const uint32_t *w, unsigned count)
487 {
488 if (opcode == SpvOpLabel)
489 return true; /* Nothing to do */
490
491 /* If this isn't a phi node, stop. */
492 if (opcode != SpvOpPhi)
493 return false;
494
495 /* For handling phi nodes, we do a poor-man's out-of-ssa on the spot.
496 * For each phi, we create a variable with the appropreate type and
497 * do a load from that variable. Then, in a second pass, we add
498 * stores to that variable to each of the predecessor blocks.
499 *
500 * We could do something more intelligent here. However, in order to
501 * handle loops and things properly, we really need dominance
502 * information. It would end up basically being the into-SSA
503 * algorithm all over again. It's easier if we just let
504 * lower_vars_to_ssa do that for us instead of repeating it here.
505 */
506 struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_ssa);
507
508 struct vtn_type *type = vtn_value(b, w[1], vtn_value_type_type)->type;
509 nir_variable *phi_var =
510 nir_local_variable_create(b->nb.impl, type->type, "phi");
511 _mesa_hash_table_insert(b->phi_table, w, phi_var);
512
513 val->ssa = vtn_local_load(b, nir_deref_var_create(b, phi_var));
514
515 return true;
516 }
517
518 static bool
vtn_handle_phi_second_pass(struct vtn_builder * b,SpvOp opcode,const uint32_t * w,unsigned count)519 vtn_handle_phi_second_pass(struct vtn_builder *b, SpvOp opcode,
520 const uint32_t *w, unsigned count)
521 {
522 if (opcode != SpvOpPhi)
523 return true;
524
525 struct hash_entry *phi_entry = _mesa_hash_table_search(b->phi_table, w);
526 assert(phi_entry);
527 nir_variable *phi_var = phi_entry->data;
528
529 for (unsigned i = 3; i < count; i += 2) {
530 struct vtn_block *pred =
531 vtn_value(b, w[i + 1], vtn_value_type_block)->block;
532
533 b->nb.cursor = nir_after_instr(&pred->end_nop->instr);
534
535 struct vtn_ssa_value *src = vtn_ssa_value(b, w[i]);
536
537 vtn_local_store(b, src, nir_deref_var_create(b, phi_var));
538 }
539
540 return true;
541 }
542
543 static void
vtn_emit_branch(struct vtn_builder * b,enum vtn_branch_type branch_type,nir_variable * switch_fall_var,bool * has_switch_break)544 vtn_emit_branch(struct vtn_builder *b, enum vtn_branch_type branch_type,
545 nir_variable *switch_fall_var, bool *has_switch_break)
546 {
547 switch (branch_type) {
548 case vtn_branch_type_switch_break:
549 nir_store_var(&b->nb, switch_fall_var, nir_imm_int(&b->nb, NIR_FALSE), 1);
550 *has_switch_break = true;
551 break;
552 case vtn_branch_type_switch_fallthrough:
553 break; /* Nothing to do */
554 case vtn_branch_type_loop_break:
555 nir_jump(&b->nb, nir_jump_break);
556 break;
557 case vtn_branch_type_loop_continue:
558 nir_jump(&b->nb, nir_jump_continue);
559 break;
560 case vtn_branch_type_return:
561 nir_jump(&b->nb, nir_jump_return);
562 break;
563 case vtn_branch_type_discard: {
564 nir_intrinsic_instr *discard =
565 nir_intrinsic_instr_create(b->nb.shader, nir_intrinsic_discard);
566 nir_builder_instr_insert(&b->nb, &discard->instr);
567 break;
568 }
569 default:
570 unreachable("Invalid branch type");
571 }
572 }
573
574 static void
vtn_emit_cf_list(struct vtn_builder * b,struct list_head * cf_list,nir_variable * switch_fall_var,bool * has_switch_break,vtn_instruction_handler handler)575 vtn_emit_cf_list(struct vtn_builder *b, struct list_head *cf_list,
576 nir_variable *switch_fall_var, bool *has_switch_break,
577 vtn_instruction_handler handler)
578 {
579 list_for_each_entry(struct vtn_cf_node, node, cf_list, link) {
580 switch (node->type) {
581 case vtn_cf_node_type_block: {
582 struct vtn_block *block = (struct vtn_block *)node;
583
584 const uint32_t *block_start = block->label;
585 const uint32_t *block_end = block->merge ? block->merge :
586 block->branch;
587
588 block_start = vtn_foreach_instruction(b, block_start, block_end,
589 vtn_handle_phis_first_pass);
590
591 vtn_foreach_instruction(b, block_start, block_end, handler);
592
593 block->end_nop = nir_intrinsic_instr_create(b->nb.shader,
594 nir_intrinsic_nop);
595 nir_builder_instr_insert(&b->nb, &block->end_nop->instr);
596
597 if ((*block->branch & SpvOpCodeMask) == SpvOpReturnValue) {
598 struct vtn_ssa_value *src = vtn_ssa_value(b, block->branch[1]);
599 vtn_local_store(b, src,
600 nir_deref_var_create(b, b->impl->return_var));
601 }
602
603 if (block->branch_type != vtn_branch_type_none) {
604 vtn_emit_branch(b, block->branch_type,
605 switch_fall_var, has_switch_break);
606 }
607
608 break;
609 }
610
611 case vtn_cf_node_type_if: {
612 struct vtn_if *vtn_if = (struct vtn_if *)node;
613
614 nir_if *if_stmt = nir_if_create(b->shader);
615 if_stmt->condition =
616 nir_src_for_ssa(vtn_ssa_value(b, vtn_if->condition)->def);
617 nir_cf_node_insert(b->nb.cursor, &if_stmt->cf_node);
618
619 bool sw_break = false;
620
621 b->nb.cursor = nir_after_cf_list(&if_stmt->then_list);
622 if (vtn_if->then_type == vtn_branch_type_none) {
623 vtn_emit_cf_list(b, &vtn_if->then_body,
624 switch_fall_var, &sw_break, handler);
625 } else {
626 vtn_emit_branch(b, vtn_if->then_type, switch_fall_var, &sw_break);
627 }
628
629 b->nb.cursor = nir_after_cf_list(&if_stmt->else_list);
630 if (vtn_if->else_type == vtn_branch_type_none) {
631 vtn_emit_cf_list(b, &vtn_if->else_body,
632 switch_fall_var, &sw_break, handler);
633 } else {
634 vtn_emit_branch(b, vtn_if->else_type, switch_fall_var, &sw_break);
635 }
636
637 b->nb.cursor = nir_after_cf_node(&if_stmt->cf_node);
638
639 /* If we encountered a switch break somewhere inside of the if,
640 * then it would have been handled correctly by calling
641 * emit_cf_list or emit_branch for the interrior. However, we
642 * need to predicate everything following on wether or not we're
643 * still going.
644 */
645 if (sw_break) {
646 *has_switch_break = true;
647
648 nir_if *switch_if = nir_if_create(b->shader);
649 switch_if->condition =
650 nir_src_for_ssa(nir_load_var(&b->nb, switch_fall_var));
651 nir_cf_node_insert(b->nb.cursor, &switch_if->cf_node);
652
653 b->nb.cursor = nir_after_cf_list(&if_stmt->then_list);
654 }
655 break;
656 }
657
658 case vtn_cf_node_type_loop: {
659 struct vtn_loop *vtn_loop = (struct vtn_loop *)node;
660
661 nir_loop *loop = nir_loop_create(b->shader);
662 nir_cf_node_insert(b->nb.cursor, &loop->cf_node);
663
664 b->nb.cursor = nir_after_cf_list(&loop->body);
665 vtn_emit_cf_list(b, &vtn_loop->body, NULL, NULL, handler);
666
667 if (!list_empty(&vtn_loop->cont_body)) {
668 /* If we have a non-trivial continue body then we need to put
669 * it at the beginning of the loop with a flag to ensure that
670 * it doesn't get executed in the first iteration.
671 */
672 nir_variable *do_cont =
673 nir_local_variable_create(b->nb.impl, glsl_bool_type(), "cont");
674
675 b->nb.cursor = nir_before_cf_node(&loop->cf_node);
676 nir_store_var(&b->nb, do_cont, nir_imm_int(&b->nb, NIR_FALSE), 1);
677
678 b->nb.cursor = nir_before_cf_list(&loop->body);
679 nir_if *cont_if = nir_if_create(b->shader);
680 cont_if->condition = nir_src_for_ssa(nir_load_var(&b->nb, do_cont));
681 nir_cf_node_insert(b->nb.cursor, &cont_if->cf_node);
682
683 b->nb.cursor = nir_after_cf_list(&cont_if->then_list);
684 vtn_emit_cf_list(b, &vtn_loop->cont_body, NULL, NULL, handler);
685
686 b->nb.cursor = nir_after_cf_node(&cont_if->cf_node);
687 nir_store_var(&b->nb, do_cont, nir_imm_int(&b->nb, NIR_TRUE), 1);
688
689 b->has_loop_continue = true;
690 }
691
692 b->nb.cursor = nir_after_cf_node(&loop->cf_node);
693 break;
694 }
695
696 case vtn_cf_node_type_switch: {
697 struct vtn_switch *vtn_switch = (struct vtn_switch *)node;
698
699 /* First, we create a variable to keep track of whether or not the
700 * switch is still going at any given point. Any switch breaks
701 * will set this variable to false.
702 */
703 nir_variable *fall_var =
704 nir_local_variable_create(b->nb.impl, glsl_bool_type(), "fall");
705 nir_store_var(&b->nb, fall_var, nir_imm_int(&b->nb, NIR_FALSE), 1);
706
707 /* Next, we gather up all of the conditions. We have to do this
708 * up-front because we also need to build an "any" condition so
709 * that we can use !any for default.
710 */
711 const int num_cases = list_length(&vtn_switch->cases);
712 NIR_VLA(nir_ssa_def *, conditions, num_cases);
713
714 nir_ssa_def *sel = vtn_ssa_value(b, vtn_switch->selector)->def;
715 /* An accumulation of all conditions. Used for the default */
716 nir_ssa_def *any = NULL;
717
718 int i = 0;
719 list_for_each_entry(struct vtn_case, cse, &vtn_switch->cases, link) {
720 if (cse->is_default) {
721 conditions[i++] = NULL;
722 continue;
723 }
724
725 nir_ssa_def *cond = NULL;
726 nir_array_foreach(&cse->values, uint32_t, val) {
727 nir_ssa_def *is_val =
728 nir_ieq(&b->nb, sel, nir_imm_int(&b->nb, *val));
729
730 cond = cond ? nir_ior(&b->nb, cond, is_val) : is_val;
731 }
732
733 any = any ? nir_ior(&b->nb, any, cond) : cond;
734 conditions[i++] = cond;
735 }
736 assert(i == num_cases);
737
738 /* Now we can walk the list of cases and actually emit code */
739 i = 0;
740 list_for_each_entry(struct vtn_case, cse, &vtn_switch->cases, link) {
741 /* Figure out the condition */
742 nir_ssa_def *cond = conditions[i++];
743 if (cse->is_default) {
744 assert(cond == NULL);
745 cond = nir_inot(&b->nb, any);
746 }
747 /* Take fallthrough into account */
748 cond = nir_ior(&b->nb, cond, nir_load_var(&b->nb, fall_var));
749
750 nir_if *case_if = nir_if_create(b->nb.shader);
751 case_if->condition = nir_src_for_ssa(cond);
752 nir_cf_node_insert(b->nb.cursor, &case_if->cf_node);
753
754 bool has_break = false;
755 b->nb.cursor = nir_after_cf_list(&case_if->then_list);
756 nir_store_var(&b->nb, fall_var, nir_imm_int(&b->nb, NIR_TRUE), 1);
757 vtn_emit_cf_list(b, &cse->body, fall_var, &has_break, handler);
758 (void)has_break; /* We don't care */
759
760 b->nb.cursor = nir_after_cf_node(&case_if->cf_node);
761 }
762 assert(i == num_cases);
763
764 break;
765 }
766
767 default:
768 unreachable("Invalid CF node type");
769 }
770 }
771 }
772
773 void
vtn_function_emit(struct vtn_builder * b,struct vtn_function * func,vtn_instruction_handler instruction_handler)774 vtn_function_emit(struct vtn_builder *b, struct vtn_function *func,
775 vtn_instruction_handler instruction_handler)
776 {
777 nir_builder_init(&b->nb, func->impl);
778 b->nb.cursor = nir_after_cf_list(&func->impl->body);
779 b->has_loop_continue = false;
780 b->phi_table = _mesa_hash_table_create(b, _mesa_hash_pointer,
781 _mesa_key_pointer_equal);
782
783 vtn_emit_cf_list(b, &func->body, NULL, NULL, instruction_handler);
784
785 vtn_foreach_instruction(b, func->start_block->label, func->end,
786 vtn_handle_phi_second_pass);
787
788 /* Continue blocks for loops get inserted before the body of the loop
789 * but instructions in the continue may use SSA defs in the loop body.
790 * Therefore, we need to repair SSA to insert the needed phi nodes.
791 */
792 if (b->has_loop_continue)
793 nir_repair_ssa_impl(func->impl);
794 }
795