ralloc: use rzalloc where it's necessary
[mesa.git] / src / compiler / nir / nir.c
1 /*
2 * Copyright © 2014 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 * Authors:
24 * Connor Abbott (cwabbott0@gmail.com)
25 *
26 */
27
28 #include "nir.h"
29 #include "nir_control_flow_private.h"
30 #include <assert.h>
31
32 nir_shader *
33 nir_shader_create(void *mem_ctx,
34 gl_shader_stage stage,
35 const nir_shader_compiler_options *options,
36 shader_info *si)
37 {
38 nir_shader *shader = rzalloc(mem_ctx, nir_shader);
39
40 exec_list_make_empty(&shader->uniforms);
41 exec_list_make_empty(&shader->inputs);
42 exec_list_make_empty(&shader->outputs);
43 exec_list_make_empty(&shader->shared);
44
45 shader->options = options;
46
47 shader->info = si ? si : rzalloc(shader, shader_info);
48
49 exec_list_make_empty(&shader->functions);
50 exec_list_make_empty(&shader->registers);
51 exec_list_make_empty(&shader->globals);
52 exec_list_make_empty(&shader->system_values);
53 shader->reg_alloc = 0;
54
55 shader->num_inputs = 0;
56 shader->num_outputs = 0;
57 shader->num_uniforms = 0;
58 shader->num_shared = 0;
59
60 shader->stage = stage;
61
62 return shader;
63 }
64
65 static nir_register *
66 reg_create(void *mem_ctx, struct exec_list *list)
67 {
68 nir_register *reg = ralloc(mem_ctx, nir_register);
69
70 list_inithead(&reg->uses);
71 list_inithead(&reg->defs);
72 list_inithead(&reg->if_uses);
73
74 reg->num_components = 0;
75 reg->bit_size = 32;
76 reg->num_array_elems = 0;
77 reg->is_packed = false;
78 reg->name = NULL;
79
80 exec_list_push_tail(list, &reg->node);
81
82 return reg;
83 }
84
85 nir_register *
86 nir_global_reg_create(nir_shader *shader)
87 {
88 nir_register *reg = reg_create(shader, &shader->registers);
89 reg->index = shader->reg_alloc++;
90 reg->is_global = true;
91
92 return reg;
93 }
94
95 nir_register *
96 nir_local_reg_create(nir_function_impl *impl)
97 {
98 nir_register *reg = reg_create(ralloc_parent(impl), &impl->registers);
99 reg->index = impl->reg_alloc++;
100 reg->is_global = false;
101
102 return reg;
103 }
104
105 void
106 nir_reg_remove(nir_register *reg)
107 {
108 exec_node_remove(&reg->node);
109 }
110
111 void
112 nir_shader_add_variable(nir_shader *shader, nir_variable *var)
113 {
114 switch (var->data.mode) {
115 case nir_var_all:
116 assert(!"invalid mode");
117 break;
118
119 case nir_var_local:
120 assert(!"nir_shader_add_variable cannot be used for local variables");
121 break;
122
123 case nir_var_param:
124 assert(!"nir_shader_add_variable cannot be used for function parameters");
125 break;
126
127 case nir_var_global:
128 exec_list_push_tail(&shader->globals, &var->node);
129 break;
130
131 case nir_var_shader_in:
132 exec_list_push_tail(&shader->inputs, &var->node);
133 break;
134
135 case nir_var_shader_out:
136 exec_list_push_tail(&shader->outputs, &var->node);
137 break;
138
139 case nir_var_uniform:
140 case nir_var_shader_storage:
141 exec_list_push_tail(&shader->uniforms, &var->node);
142 break;
143
144 case nir_var_shared:
145 assert(shader->stage == MESA_SHADER_COMPUTE);
146 exec_list_push_tail(&shader->shared, &var->node);
147 break;
148
149 case nir_var_system_value:
150 exec_list_push_tail(&shader->system_values, &var->node);
151 break;
152 }
153 }
154
155 nir_variable *
156 nir_variable_create(nir_shader *shader, nir_variable_mode mode,
157 const struct glsl_type *type, const char *name)
158 {
159 nir_variable *var = rzalloc(shader, nir_variable);
160 var->name = ralloc_strdup(var, name);
161 var->type = type;
162 var->data.mode = mode;
163
164 if ((mode == nir_var_shader_in && shader->stage != MESA_SHADER_VERTEX) ||
165 (mode == nir_var_shader_out && shader->stage != MESA_SHADER_FRAGMENT))
166 var->data.interpolation = INTERP_MODE_SMOOTH;
167
168 if (mode == nir_var_shader_in || mode == nir_var_uniform)
169 var->data.read_only = true;
170
171 nir_shader_add_variable(shader, var);
172
173 return var;
174 }
175
176 nir_variable *
177 nir_local_variable_create(nir_function_impl *impl,
178 const struct glsl_type *type, const char *name)
179 {
180 nir_variable *var = rzalloc(impl->function->shader, nir_variable);
181 var->name = ralloc_strdup(var, name);
182 var->type = type;
183 var->data.mode = nir_var_local;
184
185 nir_function_impl_add_variable(impl, var);
186
187 return var;
188 }
189
190 nir_function *
191 nir_function_create(nir_shader *shader, const char *name)
192 {
193 nir_function *func = ralloc(shader, nir_function);
194
195 exec_list_push_tail(&shader->functions, &func->node);
196
197 func->name = ralloc_strdup(func, name);
198 func->shader = shader;
199 func->num_params = 0;
200 func->params = NULL;
201 func->return_type = glsl_void_type();
202 func->impl = NULL;
203
204 return func;
205 }
206
207 void nir_src_copy(nir_src *dest, const nir_src *src, void *mem_ctx)
208 {
209 dest->is_ssa = src->is_ssa;
210 if (src->is_ssa) {
211 dest->ssa = src->ssa;
212 } else {
213 dest->reg.base_offset = src->reg.base_offset;
214 dest->reg.reg = src->reg.reg;
215 if (src->reg.indirect) {
216 dest->reg.indirect = ralloc(mem_ctx, nir_src);
217 nir_src_copy(dest->reg.indirect, src->reg.indirect, mem_ctx);
218 } else {
219 dest->reg.indirect = NULL;
220 }
221 }
222 }
223
224 void nir_dest_copy(nir_dest *dest, const nir_dest *src, nir_instr *instr)
225 {
226 /* Copying an SSA definition makes no sense whatsoever. */
227 assert(!src->is_ssa);
228
229 dest->is_ssa = false;
230
231 dest->reg.base_offset = src->reg.base_offset;
232 dest->reg.reg = src->reg.reg;
233 if (src->reg.indirect) {
234 dest->reg.indirect = ralloc(instr, nir_src);
235 nir_src_copy(dest->reg.indirect, src->reg.indirect, instr);
236 } else {
237 dest->reg.indirect = NULL;
238 }
239 }
240
241 void
242 nir_alu_src_copy(nir_alu_src *dest, const nir_alu_src *src,
243 nir_alu_instr *instr)
244 {
245 nir_src_copy(&dest->src, &src->src, &instr->instr);
246 dest->abs = src->abs;
247 dest->negate = src->negate;
248 for (unsigned i = 0; i < 4; i++)
249 dest->swizzle[i] = src->swizzle[i];
250 }
251
252 void
253 nir_alu_dest_copy(nir_alu_dest *dest, const nir_alu_dest *src,
254 nir_alu_instr *instr)
255 {
256 nir_dest_copy(&dest->dest, &src->dest, &instr->instr);
257 dest->write_mask = src->write_mask;
258 dest->saturate = src->saturate;
259 }
260
261
262 static void
263 cf_init(nir_cf_node *node, nir_cf_node_type type)
264 {
265 exec_node_init(&node->node);
266 node->parent = NULL;
267 node->type = type;
268 }
269
270 nir_function_impl *
271 nir_function_impl_create_bare(nir_shader *shader)
272 {
273 nir_function_impl *impl = ralloc(shader, nir_function_impl);
274
275 impl->function = NULL;
276
277 cf_init(&impl->cf_node, nir_cf_node_function);
278
279 exec_list_make_empty(&impl->body);
280 exec_list_make_empty(&impl->registers);
281 exec_list_make_empty(&impl->locals);
282 impl->num_params = 0;
283 impl->params = NULL;
284 impl->return_var = NULL;
285 impl->reg_alloc = 0;
286 impl->ssa_alloc = 0;
287 impl->valid_metadata = nir_metadata_none;
288
289 /* create start & end blocks */
290 nir_block *start_block = nir_block_create(shader);
291 nir_block *end_block = nir_block_create(shader);
292 start_block->cf_node.parent = &impl->cf_node;
293 end_block->cf_node.parent = &impl->cf_node;
294 impl->end_block = end_block;
295
296 exec_list_push_tail(&impl->body, &start_block->cf_node.node);
297
298 start_block->successors[0] = end_block;
299 _mesa_set_add(end_block->predecessors, start_block);
300 return impl;
301 }
302
303 nir_function_impl *
304 nir_function_impl_create(nir_function *function)
305 {
306 assert(function->impl == NULL);
307
308 nir_function_impl *impl = nir_function_impl_create_bare(function->shader);
309
310 function->impl = impl;
311 impl->function = function;
312
313 impl->num_params = function->num_params;
314 impl->params = ralloc_array(function->shader,
315 nir_variable *, impl->num_params);
316
317 for (unsigned i = 0; i < impl->num_params; i++) {
318 impl->params[i] = rzalloc(function->shader, nir_variable);
319 impl->params[i]->type = function->params[i].type;
320 impl->params[i]->data.mode = nir_var_param;
321 impl->params[i]->data.location = i;
322 }
323
324 if (!glsl_type_is_void(function->return_type)) {
325 impl->return_var = rzalloc(function->shader, nir_variable);
326 impl->return_var->type = function->return_type;
327 impl->return_var->data.mode = nir_var_param;
328 impl->return_var->data.location = -1;
329 } else {
330 impl->return_var = NULL;
331 }
332
333 return impl;
334 }
335
336 nir_block *
337 nir_block_create(nir_shader *shader)
338 {
339 nir_block *block = rzalloc(shader, nir_block);
340
341 cf_init(&block->cf_node, nir_cf_node_block);
342
343 block->successors[0] = block->successors[1] = NULL;
344 block->predecessors = _mesa_set_create(block, _mesa_hash_pointer,
345 _mesa_key_pointer_equal);
346 block->imm_dom = NULL;
347 /* XXX maybe it would be worth it to defer allocation? This
348 * way it doesn't get allocated for shader ref's that never run
349 * nir_calc_dominance? For example, state-tracker creates an
350 * initial IR, clones that, runs appropriate lowering pass, passes
351 * to driver which does common lowering/opt, and then stores ref
352 * which is later used to do state specific lowering and futher
353 * opt. Do any of the references not need dominance metadata?
354 */
355 block->dom_frontier = _mesa_set_create(block, _mesa_hash_pointer,
356 _mesa_key_pointer_equal);
357
358 exec_list_make_empty(&block->instr_list);
359
360 return block;
361 }
362
363 static inline void
364 src_init(nir_src *src)
365 {
366 src->is_ssa = false;
367 src->reg.reg = NULL;
368 src->reg.indirect = NULL;
369 src->reg.base_offset = 0;
370 }
371
372 nir_if *
373 nir_if_create(nir_shader *shader)
374 {
375 nir_if *if_stmt = ralloc(shader, nir_if);
376
377 cf_init(&if_stmt->cf_node, nir_cf_node_if);
378 src_init(&if_stmt->condition);
379
380 nir_block *then = nir_block_create(shader);
381 exec_list_make_empty(&if_stmt->then_list);
382 exec_list_push_tail(&if_stmt->then_list, &then->cf_node.node);
383 then->cf_node.parent = &if_stmt->cf_node;
384
385 nir_block *else_stmt = nir_block_create(shader);
386 exec_list_make_empty(&if_stmt->else_list);
387 exec_list_push_tail(&if_stmt->else_list, &else_stmt->cf_node.node);
388 else_stmt->cf_node.parent = &if_stmt->cf_node;
389
390 return if_stmt;
391 }
392
393 nir_loop *
394 nir_loop_create(nir_shader *shader)
395 {
396 nir_loop *loop = ralloc(shader, nir_loop);
397
398 cf_init(&loop->cf_node, nir_cf_node_loop);
399
400 nir_block *body = nir_block_create(shader);
401 exec_list_make_empty(&loop->body);
402 exec_list_push_tail(&loop->body, &body->cf_node.node);
403 body->cf_node.parent = &loop->cf_node;
404
405 body->successors[0] = body;
406 _mesa_set_add(body->predecessors, body);
407
408 return loop;
409 }
410
411 static void
412 instr_init(nir_instr *instr, nir_instr_type type)
413 {
414 instr->type = type;
415 instr->block = NULL;
416 exec_node_init(&instr->node);
417 }
418
419 static void
420 dest_init(nir_dest *dest)
421 {
422 dest->is_ssa = false;
423 dest->reg.reg = NULL;
424 dest->reg.indirect = NULL;
425 dest->reg.base_offset = 0;
426 }
427
428 static void
429 alu_dest_init(nir_alu_dest *dest)
430 {
431 dest_init(&dest->dest);
432 dest->saturate = false;
433 dest->write_mask = 0xf;
434 }
435
436 static void
437 alu_src_init(nir_alu_src *src)
438 {
439 src_init(&src->src);
440 src->abs = src->negate = false;
441 src->swizzle[0] = 0;
442 src->swizzle[1] = 1;
443 src->swizzle[2] = 2;
444 src->swizzle[3] = 3;
445 }
446
447 nir_alu_instr *
448 nir_alu_instr_create(nir_shader *shader, nir_op op)
449 {
450 unsigned num_srcs = nir_op_infos[op].num_inputs;
451 /* TODO: don't use rzalloc */
452 nir_alu_instr *instr =
453 rzalloc_size(shader,
454 sizeof(nir_alu_instr) + num_srcs * sizeof(nir_alu_src));
455
456 instr_init(&instr->instr, nir_instr_type_alu);
457 instr->op = op;
458 alu_dest_init(&instr->dest);
459 for (unsigned i = 0; i < num_srcs; i++)
460 alu_src_init(&instr->src[i]);
461
462 return instr;
463 }
464
465 nir_jump_instr *
466 nir_jump_instr_create(nir_shader *shader, nir_jump_type type)
467 {
468 nir_jump_instr *instr = ralloc(shader, nir_jump_instr);
469 instr_init(&instr->instr, nir_instr_type_jump);
470 instr->type = type;
471 return instr;
472 }
473
474 nir_load_const_instr *
475 nir_load_const_instr_create(nir_shader *shader, unsigned num_components,
476 unsigned bit_size)
477 {
478 nir_load_const_instr *instr = ralloc(shader, nir_load_const_instr);
479 instr_init(&instr->instr, nir_instr_type_load_const);
480
481 nir_ssa_def_init(&instr->instr, &instr->def, num_components, bit_size, NULL);
482
483 return instr;
484 }
485
486 nir_intrinsic_instr *
487 nir_intrinsic_instr_create(nir_shader *shader, nir_intrinsic_op op)
488 {
489 unsigned num_srcs = nir_intrinsic_infos[op].num_srcs;
490 /* TODO: don't use rzalloc */
491 nir_intrinsic_instr *instr =
492 rzalloc_size(shader,
493 sizeof(nir_intrinsic_instr) + num_srcs * sizeof(nir_src));
494
495 instr_init(&instr->instr, nir_instr_type_intrinsic);
496 instr->intrinsic = op;
497
498 if (nir_intrinsic_infos[op].has_dest)
499 dest_init(&instr->dest);
500
501 for (unsigned i = 0; i < num_srcs; i++)
502 src_init(&instr->src[i]);
503
504 return instr;
505 }
506
507 nir_call_instr *
508 nir_call_instr_create(nir_shader *shader, nir_function *callee)
509 {
510 nir_call_instr *instr = ralloc(shader, nir_call_instr);
511 instr_init(&instr->instr, nir_instr_type_call);
512
513 instr->callee = callee;
514 instr->num_params = callee->num_params;
515 instr->params = ralloc_array(instr, nir_deref_var *, instr->num_params);
516 instr->return_deref = NULL;
517
518 return instr;
519 }
520
521 nir_tex_instr *
522 nir_tex_instr_create(nir_shader *shader, unsigned num_srcs)
523 {
524 nir_tex_instr *instr = rzalloc(shader, nir_tex_instr);
525 instr_init(&instr->instr, nir_instr_type_tex);
526
527 dest_init(&instr->dest);
528
529 instr->num_srcs = num_srcs;
530 instr->src = ralloc_array(instr, nir_tex_src, num_srcs);
531 for (unsigned i = 0; i < num_srcs; i++)
532 src_init(&instr->src[i].src);
533
534 instr->texture_index = 0;
535 instr->texture_array_size = 0;
536 instr->texture = NULL;
537 instr->sampler_index = 0;
538 instr->sampler = NULL;
539
540 return instr;
541 }
542
543 void
544 nir_tex_instr_remove_src(nir_tex_instr *tex, unsigned src_idx)
545 {
546 assert(src_idx < tex->num_srcs);
547
548 /* First rewrite the source to NIR_SRC_INIT */
549 nir_instr_rewrite_src(&tex->instr, &tex->src[src_idx].src, NIR_SRC_INIT);
550
551 /* Now, move all of the other sources down */
552 for (unsigned i = src_idx + 1; i < tex->num_srcs; i++) {
553 tex->src[i-1].src_type = tex->src[i].src_type;
554 nir_instr_move_src(&tex->instr, &tex->src[i-1].src, &tex->src[i].src);
555 }
556 tex->num_srcs--;
557 }
558
559 nir_phi_instr *
560 nir_phi_instr_create(nir_shader *shader)
561 {
562 nir_phi_instr *instr = ralloc(shader, nir_phi_instr);
563 instr_init(&instr->instr, nir_instr_type_phi);
564
565 dest_init(&instr->dest);
566 exec_list_make_empty(&instr->srcs);
567 return instr;
568 }
569
570 nir_parallel_copy_instr *
571 nir_parallel_copy_instr_create(nir_shader *shader)
572 {
573 nir_parallel_copy_instr *instr = ralloc(shader, nir_parallel_copy_instr);
574 instr_init(&instr->instr, nir_instr_type_parallel_copy);
575
576 exec_list_make_empty(&instr->entries);
577
578 return instr;
579 }
580
581 nir_ssa_undef_instr *
582 nir_ssa_undef_instr_create(nir_shader *shader,
583 unsigned num_components,
584 unsigned bit_size)
585 {
586 nir_ssa_undef_instr *instr = ralloc(shader, nir_ssa_undef_instr);
587 instr_init(&instr->instr, nir_instr_type_ssa_undef);
588
589 nir_ssa_def_init(&instr->instr, &instr->def, num_components, bit_size, NULL);
590
591 return instr;
592 }
593
594 nir_deref_var *
595 nir_deref_var_create(void *mem_ctx, nir_variable *var)
596 {
597 nir_deref_var *deref = ralloc(mem_ctx, nir_deref_var);
598 deref->deref.deref_type = nir_deref_type_var;
599 deref->deref.child = NULL;
600 deref->deref.type = var->type;
601 deref->var = var;
602 return deref;
603 }
604
605 nir_deref_array *
606 nir_deref_array_create(void *mem_ctx)
607 {
608 nir_deref_array *deref = ralloc(mem_ctx, nir_deref_array);
609 deref->deref.deref_type = nir_deref_type_array;
610 deref->deref.child = NULL;
611 deref->deref_array_type = nir_deref_array_type_direct;
612 src_init(&deref->indirect);
613 deref->base_offset = 0;
614 return deref;
615 }
616
617 nir_deref_struct *
618 nir_deref_struct_create(void *mem_ctx, unsigned field_index)
619 {
620 nir_deref_struct *deref = ralloc(mem_ctx, nir_deref_struct);
621 deref->deref.deref_type = nir_deref_type_struct;
622 deref->deref.child = NULL;
623 deref->index = field_index;
624 return deref;
625 }
626
627 static nir_deref_var *
628 copy_deref_var(void *mem_ctx, nir_deref_var *deref)
629 {
630 nir_deref_var *ret = nir_deref_var_create(mem_ctx, deref->var);
631 ret->deref.type = deref->deref.type;
632 if (deref->deref.child)
633 ret->deref.child = nir_copy_deref(ret, deref->deref.child);
634 return ret;
635 }
636
637 static nir_deref_array *
638 copy_deref_array(void *mem_ctx, nir_deref_array *deref)
639 {
640 nir_deref_array *ret = nir_deref_array_create(mem_ctx);
641 ret->base_offset = deref->base_offset;
642 ret->deref_array_type = deref->deref_array_type;
643 if (deref->deref_array_type == nir_deref_array_type_indirect) {
644 nir_src_copy(&ret->indirect, &deref->indirect, mem_ctx);
645 }
646 ret->deref.type = deref->deref.type;
647 if (deref->deref.child)
648 ret->deref.child = nir_copy_deref(ret, deref->deref.child);
649 return ret;
650 }
651
652 static nir_deref_struct *
653 copy_deref_struct(void *mem_ctx, nir_deref_struct *deref)
654 {
655 nir_deref_struct *ret = nir_deref_struct_create(mem_ctx, deref->index);
656 ret->deref.type = deref->deref.type;
657 if (deref->deref.child)
658 ret->deref.child = nir_copy_deref(ret, deref->deref.child);
659 return ret;
660 }
661
662 nir_deref *
663 nir_copy_deref(void *mem_ctx, nir_deref *deref)
664 {
665 if (deref == NULL)
666 return NULL;
667
668 switch (deref->deref_type) {
669 case nir_deref_type_var:
670 return &copy_deref_var(mem_ctx, nir_deref_as_var(deref))->deref;
671 case nir_deref_type_array:
672 return &copy_deref_array(mem_ctx, nir_deref_as_array(deref))->deref;
673 case nir_deref_type_struct:
674 return &copy_deref_struct(mem_ctx, nir_deref_as_struct(deref))->deref;
675 default:
676 unreachable("Invalid dereference type");
677 }
678
679 return NULL;
680 }
681
682 /* This is the second step in the recursion. We've found the tail and made a
683 * copy. Now we need to iterate over all possible leaves and call the
684 * callback on each one.
685 */
686 static bool
687 deref_foreach_leaf_build_recur(nir_deref_var *deref, nir_deref *tail,
688 nir_deref_foreach_leaf_cb cb, void *state)
689 {
690 unsigned length;
691 union {
692 nir_deref_array arr;
693 nir_deref_struct str;
694 } tmp;
695
696 assert(tail->child == NULL);
697 switch (glsl_get_base_type(tail->type)) {
698 case GLSL_TYPE_UINT:
699 case GLSL_TYPE_INT:
700 case GLSL_TYPE_FLOAT:
701 case GLSL_TYPE_DOUBLE:
702 case GLSL_TYPE_BOOL:
703 if (glsl_type_is_vector_or_scalar(tail->type))
704 return cb(deref, state);
705 /* Fall Through */
706
707 case GLSL_TYPE_ARRAY:
708 tmp.arr.deref.deref_type = nir_deref_type_array;
709 tmp.arr.deref.type = glsl_get_array_element(tail->type);
710 tmp.arr.deref_array_type = nir_deref_array_type_direct;
711 tmp.arr.indirect = NIR_SRC_INIT;
712 tail->child = &tmp.arr.deref;
713
714 length = glsl_get_length(tail->type);
715 for (unsigned i = 0; i < length; i++) {
716 tmp.arr.deref.child = NULL;
717 tmp.arr.base_offset = i;
718 if (!deref_foreach_leaf_build_recur(deref, &tmp.arr.deref, cb, state))
719 return false;
720 }
721 return true;
722
723 case GLSL_TYPE_STRUCT:
724 tmp.str.deref.deref_type = nir_deref_type_struct;
725 tail->child = &tmp.str.deref;
726
727 length = glsl_get_length(tail->type);
728 for (unsigned i = 0; i < length; i++) {
729 tmp.arr.deref.child = NULL;
730 tmp.str.deref.type = glsl_get_struct_field(tail->type, i);
731 tmp.str.index = i;
732 if (!deref_foreach_leaf_build_recur(deref, &tmp.arr.deref, cb, state))
733 return false;
734 }
735 return true;
736
737 default:
738 unreachable("Invalid type for dereference");
739 }
740 }
741
742 /* This is the first step of the foreach_leaf recursion. In this step we are
743 * walking to the end of the deref chain and making a copy in the stack as we
744 * go. This is because we don't want to mutate the deref chain that was
745 * passed in by the caller. The downside is that this deref chain is on the
746 * stack and , if the caller wants to do anything with it, they will have to
747 * make their own copy because this one will go away.
748 */
749 static bool
750 deref_foreach_leaf_copy_recur(nir_deref_var *deref, nir_deref *tail,
751 nir_deref_foreach_leaf_cb cb, void *state)
752 {
753 union {
754 nir_deref_array arr;
755 nir_deref_struct str;
756 } c;
757
758 if (tail->child) {
759 switch (tail->child->deref_type) {
760 case nir_deref_type_array:
761 c.arr = *nir_deref_as_array(tail->child);
762 tail->child = &c.arr.deref;
763 return deref_foreach_leaf_copy_recur(deref, &c.arr.deref, cb, state);
764
765 case nir_deref_type_struct:
766 c.str = *nir_deref_as_struct(tail->child);
767 tail->child = &c.str.deref;
768 return deref_foreach_leaf_copy_recur(deref, &c.str.deref, cb, state);
769
770 case nir_deref_type_var:
771 default:
772 unreachable("Invalid deref type for a child");
773 }
774 } else {
775 /* We've gotten to the end of the original deref. Time to start
776 * building our own derefs.
777 */
778 return deref_foreach_leaf_build_recur(deref, tail, cb, state);
779 }
780 }
781
782 /**
783 * This function iterates over all of the possible derefs that can be created
784 * with the given deref as the head. It then calls the provided callback with
785 * a full deref for each one.
786 *
787 * The deref passed to the callback will be allocated on the stack. You will
788 * need to make a copy if you want it to hang around.
789 */
790 bool
791 nir_deref_foreach_leaf(nir_deref_var *deref,
792 nir_deref_foreach_leaf_cb cb, void *state)
793 {
794 nir_deref_var copy = *deref;
795 return deref_foreach_leaf_copy_recur(&copy, &copy.deref, cb, state);
796 }
797
798 /* Returns a load_const instruction that represents the constant
799 * initializer for the given deref chain. The caller is responsible for
800 * ensuring that there actually is a constant initializer.
801 */
802 nir_load_const_instr *
803 nir_deref_get_const_initializer_load(nir_shader *shader, nir_deref_var *deref)
804 {
805 nir_constant *constant = deref->var->constant_initializer;
806 assert(constant);
807
808 const nir_deref *tail = &deref->deref;
809 unsigned matrix_offset = 0;
810 while (tail->child) {
811 switch (tail->child->deref_type) {
812 case nir_deref_type_array: {
813 nir_deref_array *arr = nir_deref_as_array(tail->child);
814 assert(arr->deref_array_type == nir_deref_array_type_direct);
815 if (glsl_type_is_matrix(tail->type)) {
816 assert(arr->deref.child == NULL);
817 matrix_offset = arr->base_offset;
818 } else {
819 constant = constant->elements[arr->base_offset];
820 }
821 break;
822 }
823
824 case nir_deref_type_struct: {
825 constant = constant->elements[nir_deref_as_struct(tail->child)->index];
826 break;
827 }
828
829 default:
830 unreachable("Invalid deref child type");
831 }
832
833 tail = tail->child;
834 }
835
836 unsigned bit_size = glsl_get_bit_size(tail->type);
837 nir_load_const_instr *load =
838 nir_load_const_instr_create(shader, glsl_get_vector_elements(tail->type),
839 bit_size);
840
841 matrix_offset *= load->def.num_components;
842 for (unsigned i = 0; i < load->def.num_components; i++) {
843 switch (glsl_get_base_type(tail->type)) {
844 case GLSL_TYPE_FLOAT:
845 case GLSL_TYPE_INT:
846 case GLSL_TYPE_UINT:
847 load->value.u32[i] = constant->value.u[matrix_offset + i];
848 break;
849 case GLSL_TYPE_DOUBLE:
850 load->value.f64[i] = constant->value.d[matrix_offset + i];
851 break;
852 case GLSL_TYPE_BOOL:
853 load->value.u32[i] = constant->value.b[matrix_offset + i] ?
854 NIR_TRUE : NIR_FALSE;
855 break;
856 default:
857 unreachable("Invalid immediate type");
858 }
859 }
860
861 return load;
862 }
863
864 nir_function_impl *
865 nir_cf_node_get_function(nir_cf_node *node)
866 {
867 while (node->type != nir_cf_node_function) {
868 node = node->parent;
869 }
870
871 return nir_cf_node_as_function(node);
872 }
873
874 /* Reduces a cursor by trying to convert everything to after and trying to
875 * go up to block granularity when possible.
876 */
877 static nir_cursor
878 reduce_cursor(nir_cursor cursor)
879 {
880 switch (cursor.option) {
881 case nir_cursor_before_block:
882 assert(nir_cf_node_prev(&cursor.block->cf_node) == NULL ||
883 nir_cf_node_prev(&cursor.block->cf_node)->type != nir_cf_node_block);
884 if (exec_list_is_empty(&cursor.block->instr_list)) {
885 /* Empty block. After is as good as before. */
886 cursor.option = nir_cursor_after_block;
887 }
888 return cursor;
889
890 case nir_cursor_after_block:
891 return cursor;
892
893 case nir_cursor_before_instr: {
894 nir_instr *prev_instr = nir_instr_prev(cursor.instr);
895 if (prev_instr) {
896 /* Before this instruction is after the previous */
897 cursor.instr = prev_instr;
898 cursor.option = nir_cursor_after_instr;
899 } else {
900 /* No previous instruction. Switch to before block */
901 cursor.block = cursor.instr->block;
902 cursor.option = nir_cursor_before_block;
903 }
904 return reduce_cursor(cursor);
905 }
906
907 case nir_cursor_after_instr:
908 if (nir_instr_next(cursor.instr) == NULL) {
909 /* This is the last instruction, switch to after block */
910 cursor.option = nir_cursor_after_block;
911 cursor.block = cursor.instr->block;
912 }
913 return cursor;
914
915 default:
916 unreachable("Inavlid cursor option");
917 }
918 }
919
920 bool
921 nir_cursors_equal(nir_cursor a, nir_cursor b)
922 {
923 /* Reduced cursors should be unique */
924 a = reduce_cursor(a);
925 b = reduce_cursor(b);
926
927 return a.block == b.block && a.option == b.option;
928 }
929
930 static bool
931 add_use_cb(nir_src *src, void *state)
932 {
933 nir_instr *instr = state;
934
935 src->parent_instr = instr;
936 list_addtail(&src->use_link,
937 src->is_ssa ? &src->ssa->uses : &src->reg.reg->uses);
938
939 return true;
940 }
941
942 static bool
943 add_ssa_def_cb(nir_ssa_def *def, void *state)
944 {
945 nir_instr *instr = state;
946
947 if (instr->block && def->index == UINT_MAX) {
948 nir_function_impl *impl =
949 nir_cf_node_get_function(&instr->block->cf_node);
950
951 def->index = impl->ssa_alloc++;
952 }
953
954 return true;
955 }
956
957 static bool
958 add_reg_def_cb(nir_dest *dest, void *state)
959 {
960 nir_instr *instr = state;
961
962 if (!dest->is_ssa) {
963 dest->reg.parent_instr = instr;
964 list_addtail(&dest->reg.def_link, &dest->reg.reg->defs);
965 }
966
967 return true;
968 }
969
970 static void
971 add_defs_uses(nir_instr *instr)
972 {
973 nir_foreach_src(instr, add_use_cb, instr);
974 nir_foreach_dest(instr, add_reg_def_cb, instr);
975 nir_foreach_ssa_def(instr, add_ssa_def_cb, instr);
976 }
977
978 void
979 nir_instr_insert(nir_cursor cursor, nir_instr *instr)
980 {
981 switch (cursor.option) {
982 case nir_cursor_before_block:
983 /* Only allow inserting jumps into empty blocks. */
984 if (instr->type == nir_instr_type_jump)
985 assert(exec_list_is_empty(&cursor.block->instr_list));
986
987 instr->block = cursor.block;
988 add_defs_uses(instr);
989 exec_list_push_head(&cursor.block->instr_list, &instr->node);
990 break;
991 case nir_cursor_after_block: {
992 /* Inserting instructions after a jump is illegal. */
993 nir_instr *last = nir_block_last_instr(cursor.block);
994 assert(last == NULL || last->type != nir_instr_type_jump);
995 (void) last;
996
997 instr->block = cursor.block;
998 add_defs_uses(instr);
999 exec_list_push_tail(&cursor.block->instr_list, &instr->node);
1000 break;
1001 }
1002 case nir_cursor_before_instr:
1003 assert(instr->type != nir_instr_type_jump);
1004 instr->block = cursor.instr->block;
1005 add_defs_uses(instr);
1006 exec_node_insert_node_before(&cursor.instr->node, &instr->node);
1007 break;
1008 case nir_cursor_after_instr:
1009 /* Inserting instructions after a jump is illegal. */
1010 assert(cursor.instr->type != nir_instr_type_jump);
1011
1012 /* Only allow inserting jumps at the end of the block. */
1013 if (instr->type == nir_instr_type_jump)
1014 assert(cursor.instr == nir_block_last_instr(cursor.instr->block));
1015
1016 instr->block = cursor.instr->block;
1017 add_defs_uses(instr);
1018 exec_node_insert_after(&cursor.instr->node, &instr->node);
1019 break;
1020 }
1021
1022 if (instr->type == nir_instr_type_jump)
1023 nir_handle_add_jump(instr->block);
1024 }
1025
1026 static bool
1027 src_is_valid(const nir_src *src)
1028 {
1029 return src->is_ssa ? (src->ssa != NULL) : (src->reg.reg != NULL);
1030 }
1031
1032 static bool
1033 remove_use_cb(nir_src *src, void *state)
1034 {
1035 (void) state;
1036
1037 if (src_is_valid(src))
1038 list_del(&src->use_link);
1039
1040 return true;
1041 }
1042
1043 static bool
1044 remove_def_cb(nir_dest *dest, void *state)
1045 {
1046 (void) state;
1047
1048 if (!dest->is_ssa)
1049 list_del(&dest->reg.def_link);
1050
1051 return true;
1052 }
1053
1054 static void
1055 remove_defs_uses(nir_instr *instr)
1056 {
1057 nir_foreach_dest(instr, remove_def_cb, instr);
1058 nir_foreach_src(instr, remove_use_cb, instr);
1059 }
1060
1061 void nir_instr_remove(nir_instr *instr)
1062 {
1063 remove_defs_uses(instr);
1064 exec_node_remove(&instr->node);
1065
1066 if (instr->type == nir_instr_type_jump) {
1067 nir_jump_instr *jump_instr = nir_instr_as_jump(instr);
1068 nir_handle_remove_jump(instr->block, jump_instr->type);
1069 }
1070 }
1071
1072 /*@}*/
1073
1074 void
1075 nir_index_local_regs(nir_function_impl *impl)
1076 {
1077 unsigned index = 0;
1078 foreach_list_typed(nir_register, reg, node, &impl->registers) {
1079 reg->index = index++;
1080 }
1081 impl->reg_alloc = index;
1082 }
1083
1084 void
1085 nir_index_global_regs(nir_shader *shader)
1086 {
1087 unsigned index = 0;
1088 foreach_list_typed(nir_register, reg, node, &shader->registers) {
1089 reg->index = index++;
1090 }
1091 shader->reg_alloc = index;
1092 }
1093
1094 static bool
1095 visit_alu_dest(nir_alu_instr *instr, nir_foreach_dest_cb cb, void *state)
1096 {
1097 return cb(&instr->dest.dest, state);
1098 }
1099
1100 static bool
1101 visit_intrinsic_dest(nir_intrinsic_instr *instr, nir_foreach_dest_cb cb,
1102 void *state)
1103 {
1104 if (nir_intrinsic_infos[instr->intrinsic].has_dest)
1105 return cb(&instr->dest, state);
1106
1107 return true;
1108 }
1109
1110 static bool
1111 visit_texture_dest(nir_tex_instr *instr, nir_foreach_dest_cb cb,
1112 void *state)
1113 {
1114 return cb(&instr->dest, state);
1115 }
1116
1117 static bool
1118 visit_phi_dest(nir_phi_instr *instr, nir_foreach_dest_cb cb, void *state)
1119 {
1120 return cb(&instr->dest, state);
1121 }
1122
1123 static bool
1124 visit_parallel_copy_dest(nir_parallel_copy_instr *instr,
1125 nir_foreach_dest_cb cb, void *state)
1126 {
1127 nir_foreach_parallel_copy_entry(entry, instr) {
1128 if (!cb(&entry->dest, state))
1129 return false;
1130 }
1131
1132 return true;
1133 }
1134
1135 bool
1136 nir_foreach_dest(nir_instr *instr, nir_foreach_dest_cb cb, void *state)
1137 {
1138 switch (instr->type) {
1139 case nir_instr_type_alu:
1140 return visit_alu_dest(nir_instr_as_alu(instr), cb, state);
1141 case nir_instr_type_intrinsic:
1142 return visit_intrinsic_dest(nir_instr_as_intrinsic(instr), cb, state);
1143 case nir_instr_type_tex:
1144 return visit_texture_dest(nir_instr_as_tex(instr), cb, state);
1145 case nir_instr_type_phi:
1146 return visit_phi_dest(nir_instr_as_phi(instr), cb, state);
1147 case nir_instr_type_parallel_copy:
1148 return visit_parallel_copy_dest(nir_instr_as_parallel_copy(instr),
1149 cb, state);
1150
1151 case nir_instr_type_load_const:
1152 case nir_instr_type_ssa_undef:
1153 case nir_instr_type_call:
1154 case nir_instr_type_jump:
1155 break;
1156
1157 default:
1158 unreachable("Invalid instruction type");
1159 break;
1160 }
1161
1162 return true;
1163 }
1164
1165 struct foreach_ssa_def_state {
1166 nir_foreach_ssa_def_cb cb;
1167 void *client_state;
1168 };
1169
1170 static inline bool
1171 nir_ssa_def_visitor(nir_dest *dest, void *void_state)
1172 {
1173 struct foreach_ssa_def_state *state = void_state;
1174
1175 if (dest->is_ssa)
1176 return state->cb(&dest->ssa, state->client_state);
1177 else
1178 return true;
1179 }
1180
1181 bool
1182 nir_foreach_ssa_def(nir_instr *instr, nir_foreach_ssa_def_cb cb, void *state)
1183 {
1184 switch (instr->type) {
1185 case nir_instr_type_alu:
1186 case nir_instr_type_tex:
1187 case nir_instr_type_intrinsic:
1188 case nir_instr_type_phi:
1189 case nir_instr_type_parallel_copy: {
1190 struct foreach_ssa_def_state foreach_state = {cb, state};
1191 return nir_foreach_dest(instr, nir_ssa_def_visitor, &foreach_state);
1192 }
1193
1194 case nir_instr_type_load_const:
1195 return cb(&nir_instr_as_load_const(instr)->def, state);
1196 case nir_instr_type_ssa_undef:
1197 return cb(&nir_instr_as_ssa_undef(instr)->def, state);
1198 case nir_instr_type_call:
1199 case nir_instr_type_jump:
1200 return true;
1201 default:
1202 unreachable("Invalid instruction type");
1203 }
1204 }
1205
1206 static bool
1207 visit_src(nir_src *src, nir_foreach_src_cb cb, void *state)
1208 {
1209 if (!cb(src, state))
1210 return false;
1211 if (!src->is_ssa && src->reg.indirect)
1212 return cb(src->reg.indirect, state);
1213 return true;
1214 }
1215
1216 static bool
1217 visit_deref_array_src(nir_deref_array *deref, nir_foreach_src_cb cb,
1218 void *state)
1219 {
1220 if (deref->deref_array_type == nir_deref_array_type_indirect)
1221 return visit_src(&deref->indirect, cb, state);
1222 return true;
1223 }
1224
1225 static bool
1226 visit_deref_src(nir_deref_var *deref, nir_foreach_src_cb cb, void *state)
1227 {
1228 nir_deref *cur = &deref->deref;
1229 while (cur != NULL) {
1230 if (cur->deref_type == nir_deref_type_array) {
1231 if (!visit_deref_array_src(nir_deref_as_array(cur), cb, state))
1232 return false;
1233 }
1234
1235 cur = cur->child;
1236 }
1237
1238 return true;
1239 }
1240
1241 static bool
1242 visit_alu_src(nir_alu_instr *instr, nir_foreach_src_cb cb, void *state)
1243 {
1244 for (unsigned i = 0; i < nir_op_infos[instr->op].num_inputs; i++)
1245 if (!visit_src(&instr->src[i].src, cb, state))
1246 return false;
1247
1248 return true;
1249 }
1250
1251 static bool
1252 visit_tex_src(nir_tex_instr *instr, nir_foreach_src_cb cb, void *state)
1253 {
1254 for (unsigned i = 0; i < instr->num_srcs; i++) {
1255 if (!visit_src(&instr->src[i].src, cb, state))
1256 return false;
1257 }
1258
1259 if (instr->texture != NULL) {
1260 if (!visit_deref_src(instr->texture, cb, state))
1261 return false;
1262 }
1263
1264 if (instr->sampler != NULL) {
1265 if (!visit_deref_src(instr->sampler, cb, state))
1266 return false;
1267 }
1268
1269 return true;
1270 }
1271
1272 static bool
1273 visit_intrinsic_src(nir_intrinsic_instr *instr, nir_foreach_src_cb cb,
1274 void *state)
1275 {
1276 unsigned num_srcs = nir_intrinsic_infos[instr->intrinsic].num_srcs;
1277 for (unsigned i = 0; i < num_srcs; i++) {
1278 if (!visit_src(&instr->src[i], cb, state))
1279 return false;
1280 }
1281
1282 unsigned num_vars =
1283 nir_intrinsic_infos[instr->intrinsic].num_variables;
1284 for (unsigned i = 0; i < num_vars; i++) {
1285 if (!visit_deref_src(instr->variables[i], cb, state))
1286 return false;
1287 }
1288
1289 return true;
1290 }
1291
1292 static bool
1293 visit_phi_src(nir_phi_instr *instr, nir_foreach_src_cb cb, void *state)
1294 {
1295 nir_foreach_phi_src(src, instr) {
1296 if (!visit_src(&src->src, cb, state))
1297 return false;
1298 }
1299
1300 return true;
1301 }
1302
1303 static bool
1304 visit_parallel_copy_src(nir_parallel_copy_instr *instr,
1305 nir_foreach_src_cb cb, void *state)
1306 {
1307 nir_foreach_parallel_copy_entry(entry, instr) {
1308 if (!visit_src(&entry->src, cb, state))
1309 return false;
1310 }
1311
1312 return true;
1313 }
1314
1315 typedef struct {
1316 void *state;
1317 nir_foreach_src_cb cb;
1318 } visit_dest_indirect_state;
1319
1320 static bool
1321 visit_dest_indirect(nir_dest *dest, void *_state)
1322 {
1323 visit_dest_indirect_state *state = (visit_dest_indirect_state *) _state;
1324
1325 if (!dest->is_ssa && dest->reg.indirect)
1326 return state->cb(dest->reg.indirect, state->state);
1327
1328 return true;
1329 }
1330
1331 bool
1332 nir_foreach_src(nir_instr *instr, nir_foreach_src_cb cb, void *state)
1333 {
1334 switch (instr->type) {
1335 case nir_instr_type_alu:
1336 if (!visit_alu_src(nir_instr_as_alu(instr), cb, state))
1337 return false;
1338 break;
1339 case nir_instr_type_intrinsic:
1340 if (!visit_intrinsic_src(nir_instr_as_intrinsic(instr), cb, state))
1341 return false;
1342 break;
1343 case nir_instr_type_tex:
1344 if (!visit_tex_src(nir_instr_as_tex(instr), cb, state))
1345 return false;
1346 break;
1347 case nir_instr_type_call:
1348 /* Call instructions have no regular sources */
1349 break;
1350 case nir_instr_type_load_const:
1351 /* Constant load instructions have no regular sources */
1352 break;
1353 case nir_instr_type_phi:
1354 if (!visit_phi_src(nir_instr_as_phi(instr), cb, state))
1355 return false;
1356 break;
1357 case nir_instr_type_parallel_copy:
1358 if (!visit_parallel_copy_src(nir_instr_as_parallel_copy(instr),
1359 cb, state))
1360 return false;
1361 break;
1362 case nir_instr_type_jump:
1363 case nir_instr_type_ssa_undef:
1364 return true;
1365
1366 default:
1367 unreachable("Invalid instruction type");
1368 break;
1369 }
1370
1371 visit_dest_indirect_state dest_state;
1372 dest_state.state = state;
1373 dest_state.cb = cb;
1374 return nir_foreach_dest(instr, visit_dest_indirect, &dest_state);
1375 }
1376
1377 nir_const_value *
1378 nir_src_as_const_value(nir_src src)
1379 {
1380 if (!src.is_ssa)
1381 return NULL;
1382
1383 if (src.ssa->parent_instr->type != nir_instr_type_load_const)
1384 return NULL;
1385
1386 nir_load_const_instr *load = nir_instr_as_load_const(src.ssa->parent_instr);
1387
1388 return &load->value;
1389 }
1390
1391 /**
1392 * Returns true if the source is known to be dynamically uniform. Otherwise it
1393 * returns false which means it may or may not be dynamically uniform but it
1394 * can't be determined.
1395 */
1396 bool
1397 nir_src_is_dynamically_uniform(nir_src src)
1398 {
1399 if (!src.is_ssa)
1400 return false;
1401
1402 /* Constants are trivially dynamically uniform */
1403 if (src.ssa->parent_instr->type == nir_instr_type_load_const)
1404 return true;
1405
1406 /* As are uniform variables */
1407 if (src.ssa->parent_instr->type == nir_instr_type_intrinsic) {
1408 nir_intrinsic_instr *intr = nir_instr_as_intrinsic(src.ssa->parent_instr);
1409
1410 if (intr->intrinsic == nir_intrinsic_load_uniform)
1411 return true;
1412 }
1413
1414 /* XXX: this could have many more tests, such as when a sampler function is
1415 * called with dynamically uniform arguments.
1416 */
1417 return false;
1418 }
1419
1420 static void
1421 src_remove_all_uses(nir_src *src)
1422 {
1423 for (; src; src = src->is_ssa ? NULL : src->reg.indirect) {
1424 if (!src_is_valid(src))
1425 continue;
1426
1427 list_del(&src->use_link);
1428 }
1429 }
1430
1431 static void
1432 src_add_all_uses(nir_src *src, nir_instr *parent_instr, nir_if *parent_if)
1433 {
1434 for (; src; src = src->is_ssa ? NULL : src->reg.indirect) {
1435 if (!src_is_valid(src))
1436 continue;
1437
1438 if (parent_instr) {
1439 src->parent_instr = parent_instr;
1440 if (src->is_ssa)
1441 list_addtail(&src->use_link, &src->ssa->uses);
1442 else
1443 list_addtail(&src->use_link, &src->reg.reg->uses);
1444 } else {
1445 assert(parent_if);
1446 src->parent_if = parent_if;
1447 if (src->is_ssa)
1448 list_addtail(&src->use_link, &src->ssa->if_uses);
1449 else
1450 list_addtail(&src->use_link, &src->reg.reg->if_uses);
1451 }
1452 }
1453 }
1454
1455 void
1456 nir_instr_rewrite_src(nir_instr *instr, nir_src *src, nir_src new_src)
1457 {
1458 assert(!src_is_valid(src) || src->parent_instr == instr);
1459
1460 src_remove_all_uses(src);
1461 *src = new_src;
1462 src_add_all_uses(src, instr, NULL);
1463 }
1464
1465 void
1466 nir_instr_move_src(nir_instr *dest_instr, nir_src *dest, nir_src *src)
1467 {
1468 assert(!src_is_valid(dest) || dest->parent_instr == dest_instr);
1469
1470 src_remove_all_uses(dest);
1471 src_remove_all_uses(src);
1472 *dest = *src;
1473 *src = NIR_SRC_INIT;
1474 src_add_all_uses(dest, dest_instr, NULL);
1475 }
1476
1477 void
1478 nir_if_rewrite_condition(nir_if *if_stmt, nir_src new_src)
1479 {
1480 nir_src *src = &if_stmt->condition;
1481 assert(!src_is_valid(src) || src->parent_if == if_stmt);
1482
1483 src_remove_all_uses(src);
1484 *src = new_src;
1485 src_add_all_uses(src, NULL, if_stmt);
1486 }
1487
1488 void
1489 nir_instr_rewrite_dest(nir_instr *instr, nir_dest *dest, nir_dest new_dest)
1490 {
1491 if (dest->is_ssa) {
1492 /* We can only overwrite an SSA destination if it has no uses. */
1493 assert(list_empty(&dest->ssa.uses) && list_empty(&dest->ssa.if_uses));
1494 } else {
1495 list_del(&dest->reg.def_link);
1496 if (dest->reg.indirect)
1497 src_remove_all_uses(dest->reg.indirect);
1498 }
1499
1500 /* We can't re-write with an SSA def */
1501 assert(!new_dest.is_ssa);
1502
1503 nir_dest_copy(dest, &new_dest, instr);
1504
1505 dest->reg.parent_instr = instr;
1506 list_addtail(&dest->reg.def_link, &new_dest.reg.reg->defs);
1507
1508 if (dest->reg.indirect)
1509 src_add_all_uses(dest->reg.indirect, instr, NULL);
1510 }
1511
1512 /* note: does *not* take ownership of 'name' */
1513 void
1514 nir_ssa_def_init(nir_instr *instr, nir_ssa_def *def,
1515 unsigned num_components,
1516 unsigned bit_size, const char *name)
1517 {
1518 def->name = ralloc_strdup(instr, name);
1519 def->parent_instr = instr;
1520 list_inithead(&def->uses);
1521 list_inithead(&def->if_uses);
1522 def->num_components = num_components;
1523 def->bit_size = bit_size;
1524
1525 if (instr->block) {
1526 nir_function_impl *impl =
1527 nir_cf_node_get_function(&instr->block->cf_node);
1528
1529 def->index = impl->ssa_alloc++;
1530 } else {
1531 def->index = UINT_MAX;
1532 }
1533 }
1534
1535 /* note: does *not* take ownership of 'name' */
1536 void
1537 nir_ssa_dest_init(nir_instr *instr, nir_dest *dest,
1538 unsigned num_components, unsigned bit_size,
1539 const char *name)
1540 {
1541 dest->is_ssa = true;
1542 nir_ssa_def_init(instr, &dest->ssa, num_components, bit_size, name);
1543 }
1544
1545 void
1546 nir_ssa_def_rewrite_uses(nir_ssa_def *def, nir_src new_src)
1547 {
1548 assert(!new_src.is_ssa || def != new_src.ssa);
1549
1550 nir_foreach_use_safe(use_src, def)
1551 nir_instr_rewrite_src(use_src->parent_instr, use_src, new_src);
1552
1553 nir_foreach_if_use_safe(use_src, def)
1554 nir_if_rewrite_condition(use_src->parent_if, new_src);
1555 }
1556
1557 static bool
1558 is_instr_between(nir_instr *start, nir_instr *end, nir_instr *between)
1559 {
1560 assert(start->block == end->block);
1561
1562 if (between->block != start->block)
1563 return false;
1564
1565 /* Search backwards looking for "between" */
1566 while (start != end) {
1567 if (between == end)
1568 return true;
1569
1570 end = nir_instr_prev(end);
1571 assert(end);
1572 }
1573
1574 return false;
1575 }
1576
1577 /* Replaces all uses of the given SSA def with the given source but only if
1578 * the use comes after the after_me instruction. This can be useful if you
1579 * are emitting code to fix up the result of some instruction: you can freely
1580 * use the result in that code and then call rewrite_uses_after and pass the
1581 * last fixup instruction as after_me and it will replace all of the uses you
1582 * want without touching the fixup code.
1583 *
1584 * This function assumes that after_me is in the same block as
1585 * def->parent_instr and that after_me comes after def->parent_instr.
1586 */
1587 void
1588 nir_ssa_def_rewrite_uses_after(nir_ssa_def *def, nir_src new_src,
1589 nir_instr *after_me)
1590 {
1591 assert(!new_src.is_ssa || def != new_src.ssa);
1592
1593 nir_foreach_use_safe(use_src, def) {
1594 assert(use_src->parent_instr != def->parent_instr);
1595 /* Since def already dominates all of its uses, the only way a use can
1596 * not be dominated by after_me is if it is between def and after_me in
1597 * the instruction list.
1598 */
1599 if (!is_instr_between(def->parent_instr, after_me, use_src->parent_instr))
1600 nir_instr_rewrite_src(use_src->parent_instr, use_src, new_src);
1601 }
1602
1603 nir_foreach_if_use_safe(use_src, def)
1604 nir_if_rewrite_condition(use_src->parent_if, new_src);
1605 }
1606
1607 uint8_t
1608 nir_ssa_def_components_read(nir_ssa_def *def)
1609 {
1610 uint8_t read_mask = 0;
1611 nir_foreach_use(use, def) {
1612 if (use->parent_instr->type == nir_instr_type_alu) {
1613 nir_alu_instr *alu = nir_instr_as_alu(use->parent_instr);
1614 nir_alu_src *alu_src = exec_node_data(nir_alu_src, use, src);
1615 int src_idx = alu_src - &alu->src[0];
1616 assert(src_idx >= 0 && src_idx < nir_op_infos[alu->op].num_inputs);
1617
1618 for (unsigned c = 0; c < 4; c++) {
1619 if (!nir_alu_instr_channel_used(alu, src_idx, c))
1620 continue;
1621
1622 read_mask |= (1 << alu_src->swizzle[c]);
1623 }
1624 } else {
1625 return (1 << def->num_components) - 1;
1626 }
1627 }
1628
1629 return read_mask;
1630 }
1631
1632 nir_block *
1633 nir_block_cf_tree_next(nir_block *block)
1634 {
1635 if (block == NULL) {
1636 /* nir_foreach_block_safe() will call this function on a NULL block
1637 * after the last iteration, but it won't use the result so just return
1638 * NULL here.
1639 */
1640 return NULL;
1641 }
1642
1643 nir_cf_node *cf_next = nir_cf_node_next(&block->cf_node);
1644 if (cf_next)
1645 return nir_cf_node_cf_tree_first(cf_next);
1646
1647 nir_cf_node *parent = block->cf_node.parent;
1648
1649 switch (parent->type) {
1650 case nir_cf_node_if: {
1651 /* Are we at the end of the if? Go to the beginning of the else */
1652 nir_if *if_stmt = nir_cf_node_as_if(parent);
1653 if (block == nir_if_last_then_block(if_stmt))
1654 return nir_if_first_else_block(if_stmt);
1655
1656 assert(block == nir_if_last_else_block(if_stmt));
1657 /* fall through */
1658 }
1659
1660 case nir_cf_node_loop:
1661 return nir_cf_node_as_block(nir_cf_node_next(parent));
1662
1663 case nir_cf_node_function:
1664 return NULL;
1665
1666 default:
1667 unreachable("unknown cf node type");
1668 }
1669 }
1670
1671 nir_block *
1672 nir_block_cf_tree_prev(nir_block *block)
1673 {
1674 if (block == NULL) {
1675 /* do this for consistency with nir_block_cf_tree_next() */
1676 return NULL;
1677 }
1678
1679 nir_cf_node *cf_prev = nir_cf_node_prev(&block->cf_node);
1680 if (cf_prev)
1681 return nir_cf_node_cf_tree_last(cf_prev);
1682
1683 nir_cf_node *parent = block->cf_node.parent;
1684
1685 switch (parent->type) {
1686 case nir_cf_node_if: {
1687 /* Are we at the beginning of the else? Go to the end of the if */
1688 nir_if *if_stmt = nir_cf_node_as_if(parent);
1689 if (block == nir_if_first_else_block(if_stmt))
1690 return nir_if_last_then_block(if_stmt);
1691
1692 assert(block == nir_if_first_then_block(if_stmt));
1693 /* fall through */
1694 }
1695
1696 case nir_cf_node_loop:
1697 return nir_cf_node_as_block(nir_cf_node_prev(parent));
1698
1699 case nir_cf_node_function:
1700 return NULL;
1701
1702 default:
1703 unreachable("unknown cf node type");
1704 }
1705 }
1706
1707 nir_block *nir_cf_node_cf_tree_first(nir_cf_node *node)
1708 {
1709 switch (node->type) {
1710 case nir_cf_node_function: {
1711 nir_function_impl *impl = nir_cf_node_as_function(node);
1712 return nir_start_block(impl);
1713 }
1714
1715 case nir_cf_node_if: {
1716 nir_if *if_stmt = nir_cf_node_as_if(node);
1717 return nir_if_first_then_block(if_stmt);
1718 }
1719
1720 case nir_cf_node_loop: {
1721 nir_loop *loop = nir_cf_node_as_loop(node);
1722 return nir_loop_first_block(loop);
1723 }
1724
1725 case nir_cf_node_block: {
1726 return nir_cf_node_as_block(node);
1727 }
1728
1729 default:
1730 unreachable("unknown node type");
1731 }
1732 }
1733
1734 nir_block *nir_cf_node_cf_tree_last(nir_cf_node *node)
1735 {
1736 switch (node->type) {
1737 case nir_cf_node_function: {
1738 nir_function_impl *impl = nir_cf_node_as_function(node);
1739 return nir_impl_last_block(impl);
1740 }
1741
1742 case nir_cf_node_if: {
1743 nir_if *if_stmt = nir_cf_node_as_if(node);
1744 return nir_if_last_else_block(if_stmt);
1745 }
1746
1747 case nir_cf_node_loop: {
1748 nir_loop *loop = nir_cf_node_as_loop(node);
1749 return nir_loop_last_block(loop);
1750 }
1751
1752 case nir_cf_node_block: {
1753 return nir_cf_node_as_block(node);
1754 }
1755
1756 default:
1757 unreachable("unknown node type");
1758 }
1759 }
1760
1761 nir_block *nir_cf_node_cf_tree_next(nir_cf_node *node)
1762 {
1763 if (node->type == nir_cf_node_block)
1764 return nir_cf_node_cf_tree_first(nir_cf_node_next(node));
1765 else if (node->type == nir_cf_node_function)
1766 return NULL;
1767 else
1768 return nir_cf_node_as_block(nir_cf_node_next(node));
1769 }
1770
1771 nir_if *
1772 nir_block_get_following_if(nir_block *block)
1773 {
1774 if (exec_node_is_tail_sentinel(&block->cf_node.node))
1775 return NULL;
1776
1777 if (nir_cf_node_is_last(&block->cf_node))
1778 return NULL;
1779
1780 nir_cf_node *next_node = nir_cf_node_next(&block->cf_node);
1781
1782 if (next_node->type != nir_cf_node_if)
1783 return NULL;
1784
1785 return nir_cf_node_as_if(next_node);
1786 }
1787
1788 nir_loop *
1789 nir_block_get_following_loop(nir_block *block)
1790 {
1791 if (exec_node_is_tail_sentinel(&block->cf_node.node))
1792 return NULL;
1793
1794 if (nir_cf_node_is_last(&block->cf_node))
1795 return NULL;
1796
1797 nir_cf_node *next_node = nir_cf_node_next(&block->cf_node);
1798
1799 if (next_node->type != nir_cf_node_loop)
1800 return NULL;
1801
1802 return nir_cf_node_as_loop(next_node);
1803 }
1804
1805 void
1806 nir_index_blocks(nir_function_impl *impl)
1807 {
1808 unsigned index = 0;
1809
1810 if (impl->valid_metadata & nir_metadata_block_index)
1811 return;
1812
1813 nir_foreach_block(block, impl) {
1814 block->index = index++;
1815 }
1816
1817 impl->num_blocks = index;
1818 }
1819
1820 static bool
1821 index_ssa_def_cb(nir_ssa_def *def, void *state)
1822 {
1823 unsigned *index = (unsigned *) state;
1824 def->index = (*index)++;
1825
1826 return true;
1827 }
1828
1829 /**
1830 * The indices are applied top-to-bottom which has the very nice property
1831 * that, if A dominates B, then A->index <= B->index.
1832 */
1833 void
1834 nir_index_ssa_defs(nir_function_impl *impl)
1835 {
1836 unsigned index = 0;
1837
1838 nir_foreach_block(block, impl) {
1839 nir_foreach_instr(instr, block)
1840 nir_foreach_ssa_def(instr, index_ssa_def_cb, &index);
1841 }
1842
1843 impl->ssa_alloc = index;
1844 }
1845
1846 /**
1847 * The indices are applied top-to-bottom which has the very nice property
1848 * that, if A dominates B, then A->index <= B->index.
1849 */
1850 unsigned
1851 nir_index_instrs(nir_function_impl *impl)
1852 {
1853 unsigned index = 0;
1854
1855 nir_foreach_block(block, impl) {
1856 nir_foreach_instr(instr, block)
1857 instr->index = index++;
1858 }
1859
1860 return index;
1861 }
1862
1863 nir_intrinsic_op
1864 nir_intrinsic_from_system_value(gl_system_value val)
1865 {
1866 switch (val) {
1867 case SYSTEM_VALUE_VERTEX_ID:
1868 return nir_intrinsic_load_vertex_id;
1869 case SYSTEM_VALUE_INSTANCE_ID:
1870 return nir_intrinsic_load_instance_id;
1871 case SYSTEM_VALUE_DRAW_ID:
1872 return nir_intrinsic_load_draw_id;
1873 case SYSTEM_VALUE_BASE_INSTANCE:
1874 return nir_intrinsic_load_base_instance;
1875 case SYSTEM_VALUE_VERTEX_ID_ZERO_BASE:
1876 return nir_intrinsic_load_vertex_id_zero_base;
1877 case SYSTEM_VALUE_BASE_VERTEX:
1878 return nir_intrinsic_load_base_vertex;
1879 case SYSTEM_VALUE_INVOCATION_ID:
1880 return nir_intrinsic_load_invocation_id;
1881 case SYSTEM_VALUE_FRONT_FACE:
1882 return nir_intrinsic_load_front_face;
1883 case SYSTEM_VALUE_SAMPLE_ID:
1884 return nir_intrinsic_load_sample_id;
1885 case SYSTEM_VALUE_SAMPLE_POS:
1886 return nir_intrinsic_load_sample_pos;
1887 case SYSTEM_VALUE_SAMPLE_MASK_IN:
1888 return nir_intrinsic_load_sample_mask_in;
1889 case SYSTEM_VALUE_LOCAL_INVOCATION_ID:
1890 return nir_intrinsic_load_local_invocation_id;
1891 case SYSTEM_VALUE_LOCAL_INVOCATION_INDEX:
1892 return nir_intrinsic_load_local_invocation_index;
1893 case SYSTEM_VALUE_WORK_GROUP_ID:
1894 return nir_intrinsic_load_work_group_id;
1895 case SYSTEM_VALUE_NUM_WORK_GROUPS:
1896 return nir_intrinsic_load_num_work_groups;
1897 case SYSTEM_VALUE_PRIMITIVE_ID:
1898 return nir_intrinsic_load_primitive_id;
1899 case SYSTEM_VALUE_TESS_COORD:
1900 return nir_intrinsic_load_tess_coord;
1901 case SYSTEM_VALUE_TESS_LEVEL_OUTER:
1902 return nir_intrinsic_load_tess_level_outer;
1903 case SYSTEM_VALUE_TESS_LEVEL_INNER:
1904 return nir_intrinsic_load_tess_level_inner;
1905 case SYSTEM_VALUE_VERTICES_IN:
1906 return nir_intrinsic_load_patch_vertices_in;
1907 case SYSTEM_VALUE_HELPER_INVOCATION:
1908 return nir_intrinsic_load_helper_invocation;
1909 default:
1910 unreachable("system value does not directly correspond to intrinsic");
1911 }
1912 }
1913
1914 gl_system_value
1915 nir_system_value_from_intrinsic(nir_intrinsic_op intrin)
1916 {
1917 switch (intrin) {
1918 case nir_intrinsic_load_vertex_id:
1919 return SYSTEM_VALUE_VERTEX_ID;
1920 case nir_intrinsic_load_instance_id:
1921 return SYSTEM_VALUE_INSTANCE_ID;
1922 case nir_intrinsic_load_draw_id:
1923 return SYSTEM_VALUE_DRAW_ID;
1924 case nir_intrinsic_load_base_instance:
1925 return SYSTEM_VALUE_BASE_INSTANCE;
1926 case nir_intrinsic_load_vertex_id_zero_base:
1927 return SYSTEM_VALUE_VERTEX_ID_ZERO_BASE;
1928 case nir_intrinsic_load_base_vertex:
1929 return SYSTEM_VALUE_BASE_VERTEX;
1930 case nir_intrinsic_load_invocation_id:
1931 return SYSTEM_VALUE_INVOCATION_ID;
1932 case nir_intrinsic_load_front_face:
1933 return SYSTEM_VALUE_FRONT_FACE;
1934 case nir_intrinsic_load_sample_id:
1935 return SYSTEM_VALUE_SAMPLE_ID;
1936 case nir_intrinsic_load_sample_pos:
1937 return SYSTEM_VALUE_SAMPLE_POS;
1938 case nir_intrinsic_load_sample_mask_in:
1939 return SYSTEM_VALUE_SAMPLE_MASK_IN;
1940 case nir_intrinsic_load_local_invocation_id:
1941 return SYSTEM_VALUE_LOCAL_INVOCATION_ID;
1942 case nir_intrinsic_load_local_invocation_index:
1943 return SYSTEM_VALUE_LOCAL_INVOCATION_INDEX;
1944 case nir_intrinsic_load_num_work_groups:
1945 return SYSTEM_VALUE_NUM_WORK_GROUPS;
1946 case nir_intrinsic_load_work_group_id:
1947 return SYSTEM_VALUE_WORK_GROUP_ID;
1948 case nir_intrinsic_load_primitive_id:
1949 return SYSTEM_VALUE_PRIMITIVE_ID;
1950 case nir_intrinsic_load_tess_coord:
1951 return SYSTEM_VALUE_TESS_COORD;
1952 case nir_intrinsic_load_tess_level_outer:
1953 return SYSTEM_VALUE_TESS_LEVEL_OUTER;
1954 case nir_intrinsic_load_tess_level_inner:
1955 return SYSTEM_VALUE_TESS_LEVEL_INNER;
1956 case nir_intrinsic_load_patch_vertices_in:
1957 return SYSTEM_VALUE_VERTICES_IN;
1958 case nir_intrinsic_load_helper_invocation:
1959 return SYSTEM_VALUE_HELPER_INVOCATION;
1960 default:
1961 unreachable("intrinsic doesn't produce a system value");
1962 }
1963 }