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_builder.h"
30 #include "nir_control_flow_private.h"
31 #include "nir_worklist.h"
32 #include "util/half_float.h"
33 #include <limits.h>
34 #include <assert.h>
35 #include <math.h>
36 #include "util/u_math.h"
37 #include "util/u_qsort.h"
38
39 #include "main/menums.h" /* BITFIELD64_MASK */
40
41 #ifndef NDEBUG
42 uint32_t nir_debug = 0;
43 bool nir_debug_print_shader[MESA_SHADER_KERNEL + 1] = { 0 };
44
45 static const struct debug_named_value nir_debug_control[] = {
46 { "clone", NIR_DEBUG_CLONE,
47 "Test cloning a shader at each successful lowering/optimization call" },
48 { "serialize", NIR_DEBUG_SERIALIZE,
49 "Test serialize and deserialize shader at each successful lowering/optimization call" },
50 { "novalidate", NIR_DEBUG_NOVALIDATE,
51 "Disable shader validation at each successful lowering/optimization call" },
52 { "validate_ssa_dominance", NIR_DEBUG_VALIDATE_SSA_DOMINANCE,
53 "Validate SSA dominance in shader at each successful lowering/optimization call" },
54 { "validate_gc_list", NIR_DEBUG_VALIDATE_GC_LIST,
55 "Validate the instruction GC list at each successful lowering/optimization call" },
56 { "tgsi", NIR_DEBUG_TGSI,
57 "Dump NIR/TGSI shaders when doing a NIR<->TGSI translation" },
58 { "print", NIR_DEBUG_PRINT,
59 "Dump resulting shader after each successful lowering/optimization call" },
60 { "print_vs", NIR_DEBUG_PRINT_VS,
61 "Dump resulting vertex shader after each successful lowering/optimization call" },
62 { "print_tcs", NIR_DEBUG_PRINT_TCS,
63 "Dump resulting tessellation control shader after each successful lowering/optimization call" },
64 { "print_tes", NIR_DEBUG_PRINT_TES,
65 "Dump resulting tessellation evaluation shader after each successful lowering/optimization call" },
66 { "print_gs", NIR_DEBUG_PRINT_GS,
67 "Dump resulting geometry shader after each successful lowering/optimization call" },
68 { "print_fs", NIR_DEBUG_PRINT_FS,
69 "Dump resulting fragment shader after each successful lowering/optimization call" },
70 { "print_cs", NIR_DEBUG_PRINT_CS,
71 "Dump resulting compute shader after each successful lowering/optimization call" },
72 { "print_ts", NIR_DEBUG_PRINT_TS,
73 "Dump resulting task shader after each successful lowering/optimization call" },
74 { "print_ms", NIR_DEBUG_PRINT_MS,
75 "Dump resulting mesh shader after each successful lowering/optimization call" },
76 { "print_rgs", NIR_DEBUG_PRINT_RGS,
77 "Dump resulting raygen shader after each successful lowering/optimization call" },
78 { "print_ahs", NIR_DEBUG_PRINT_AHS,
79 "Dump resulting any-hit shader after each successful lowering/optimization call" },
80 { "print_chs", NIR_DEBUG_PRINT_CHS,
81 "Dump resulting closest-hit shader after each successful lowering/optimization call" },
82 { "print_mhs", NIR_DEBUG_PRINT_MHS,
83 "Dump resulting miss-hit shader after each successful lowering/optimization call" },
84 { "print_is", NIR_DEBUG_PRINT_IS,
85 "Dump resulting intersection shader after each successful lowering/optimization call" },
86 { "print_cbs", NIR_DEBUG_PRINT_CBS,
87 "Dump resulting callable shader after each successful lowering/optimization call" },
88 { "print_ks", NIR_DEBUG_PRINT_KS,
89 "Dump resulting kernel shader after each successful lowering/optimization call" },
90 { "print_consts", NIR_DEBUG_PRINT_CONSTS,
91 "Print const value near each use of const SSA variable" },
92 { NULL }
93 };
94
95 DEBUG_GET_ONCE_FLAGS_OPTION(nir_debug, "NIR_DEBUG", nir_debug_control, 0)
96
97 static void
nir_process_debug_variable_once(void)98 nir_process_debug_variable_once(void)
99 {
100 nir_debug = debug_get_option_nir_debug();
101 nir_debug_print_shader[MESA_SHADER_VERTEX] = NIR_DEBUG(PRINT_VS);
102 nir_debug_print_shader[MESA_SHADER_TESS_CTRL] = NIR_DEBUG(PRINT_TCS);
103 nir_debug_print_shader[MESA_SHADER_TESS_EVAL] = NIR_DEBUG(PRINT_TES);
104 nir_debug_print_shader[MESA_SHADER_GEOMETRY] = NIR_DEBUG(PRINT_GS);
105 nir_debug_print_shader[MESA_SHADER_FRAGMENT] = NIR_DEBUG(PRINT_FS);
106 nir_debug_print_shader[MESA_SHADER_COMPUTE] = NIR_DEBUG(PRINT_CS);
107 nir_debug_print_shader[MESA_SHADER_TASK] = NIR_DEBUG(PRINT_TS);
108 nir_debug_print_shader[MESA_SHADER_MESH] = NIR_DEBUG(PRINT_MS);
109 nir_debug_print_shader[MESA_SHADER_RAYGEN] = NIR_DEBUG(PRINT_RGS);
110 nir_debug_print_shader[MESA_SHADER_ANY_HIT] = NIR_DEBUG(PRINT_AHS);
111 nir_debug_print_shader[MESA_SHADER_CLOSEST_HIT] = NIR_DEBUG(PRINT_CHS);
112 nir_debug_print_shader[MESA_SHADER_MISS] = NIR_DEBUG(PRINT_MHS);
113 nir_debug_print_shader[MESA_SHADER_INTERSECTION] = NIR_DEBUG(PRINT_IS);
114 nir_debug_print_shader[MESA_SHADER_CALLABLE] = NIR_DEBUG(PRINT_CBS);
115 nir_debug_print_shader[MESA_SHADER_KERNEL] = NIR_DEBUG(PRINT_KS);
116 }
117
118 void
nir_process_debug_variable(void)119 nir_process_debug_variable(void)
120 {
121 static once_flag flag = ONCE_FLAG_INIT;
122 call_once(&flag, nir_process_debug_variable_once);
123 }
124 #endif
125
126 /** Return true if the component mask "mask" with bit size "old_bit_size" can
127 * be re-interpreted to be used with "new_bit_size".
128 */
129 bool
nir_component_mask_can_reinterpret(nir_component_mask_t mask,unsigned old_bit_size,unsigned new_bit_size)130 nir_component_mask_can_reinterpret(nir_component_mask_t mask,
131 unsigned old_bit_size,
132 unsigned new_bit_size)
133 {
134 assert(util_is_power_of_two_nonzero(old_bit_size));
135 assert(util_is_power_of_two_nonzero(new_bit_size));
136
137 if (old_bit_size == new_bit_size)
138 return true;
139
140 if (old_bit_size == 1 || new_bit_size == 1)
141 return false;
142
143 if (old_bit_size > new_bit_size) {
144 unsigned ratio = old_bit_size / new_bit_size;
145 return util_last_bit(mask) * ratio <= NIR_MAX_VEC_COMPONENTS;
146 }
147
148 unsigned iter = mask;
149 while (iter) {
150 int start, count;
151 u_bit_scan_consecutive_range(&iter, &start, &count);
152 start *= old_bit_size;
153 count *= old_bit_size;
154 if (start % new_bit_size != 0)
155 return false;
156 if (count % new_bit_size != 0)
157 return false;
158 }
159 return true;
160 }
161
162 /** Re-interprets a component mask "mask" with bit size "old_bit_size" so that
163 * it can be used can be used with "new_bit_size".
164 */
165 nir_component_mask_t
nir_component_mask_reinterpret(nir_component_mask_t mask,unsigned old_bit_size,unsigned new_bit_size)166 nir_component_mask_reinterpret(nir_component_mask_t mask,
167 unsigned old_bit_size,
168 unsigned new_bit_size)
169 {
170 assert(nir_component_mask_can_reinterpret(mask, old_bit_size, new_bit_size));
171
172 if (old_bit_size == new_bit_size)
173 return mask;
174
175 nir_component_mask_t new_mask = 0;
176 unsigned iter = mask;
177 while (iter) {
178 int start, count;
179 u_bit_scan_consecutive_range(&iter, &start, &count);
180 start = start * old_bit_size / new_bit_size;
181 count = count * old_bit_size / new_bit_size;
182 new_mask |= BITFIELD_RANGE(start, count);
183 }
184 return new_mask;
185 }
186
187 static void
nir_shader_destructor(void * ptr)188 nir_shader_destructor(void *ptr)
189 {
190 nir_shader *shader = ptr;
191
192 /* Free all instrs from the shader, since they're not ralloced. */
193 list_for_each_entry_safe(nir_instr, instr, &shader->gc_list, gc_node) {
194 nir_instr_free(instr);
195 }
196 }
197
198 nir_shader *
nir_shader_create(void * mem_ctx,gl_shader_stage stage,const nir_shader_compiler_options * options,shader_info * si)199 nir_shader_create(void *mem_ctx,
200 gl_shader_stage stage,
201 const nir_shader_compiler_options *options,
202 shader_info *si)
203 {
204 nir_shader *shader = rzalloc(mem_ctx, nir_shader);
205 ralloc_set_destructor(shader, nir_shader_destructor);
206
207 #ifndef NDEBUG
208 nir_process_debug_variable();
209 #endif
210
211 exec_list_make_empty(&shader->variables);
212
213 shader->options = options;
214
215 if (si) {
216 assert(si->stage == stage);
217 shader->info = *si;
218 } else {
219 shader->info.stage = stage;
220 }
221
222 exec_list_make_empty(&shader->functions);
223
224 list_inithead(&shader->gc_list);
225
226 shader->num_inputs = 0;
227 shader->num_outputs = 0;
228 shader->num_uniforms = 0;
229
230 return shader;
231 }
232
233 static nir_register *
reg_create(void * mem_ctx,struct exec_list * list)234 reg_create(void *mem_ctx, struct exec_list *list)
235 {
236 nir_register *reg = ralloc(mem_ctx, nir_register);
237
238 list_inithead(®->uses);
239 list_inithead(®->defs);
240 list_inithead(®->if_uses);
241
242 reg->num_components = 0;
243 reg->bit_size = 32;
244 reg->num_array_elems = 0;
245 reg->divergent = false;
246
247 exec_list_push_tail(list, ®->node);
248
249 return reg;
250 }
251
252 nir_register *
nir_local_reg_create(nir_function_impl * impl)253 nir_local_reg_create(nir_function_impl *impl)
254 {
255 nir_register *reg = reg_create(ralloc_parent(impl), &impl->registers);
256 reg->index = impl->reg_alloc++;
257
258 return reg;
259 }
260
261 void
nir_reg_remove(nir_register * reg)262 nir_reg_remove(nir_register *reg)
263 {
264 exec_node_remove(®->node);
265 }
266
267 void
nir_shader_add_variable(nir_shader * shader,nir_variable * var)268 nir_shader_add_variable(nir_shader *shader, nir_variable *var)
269 {
270 switch (var->data.mode) {
271 case nir_var_function_temp:
272 assert(!"nir_shader_add_variable cannot be used for local variables");
273 return;
274
275 case nir_var_shader_temp:
276 case nir_var_shader_in:
277 case nir_var_shader_out:
278 case nir_var_uniform:
279 case nir_var_mem_ubo:
280 case nir_var_mem_ssbo:
281 case nir_var_image:
282 case nir_var_mem_shared:
283 case nir_var_system_value:
284 case nir_var_mem_push_const:
285 case nir_var_mem_constant:
286 case nir_var_shader_call_data:
287 case nir_var_ray_hit_attrib:
288 case nir_var_mem_task_payload:
289 case nir_var_mem_global:
290 break;
291
292 default:
293 assert(!"invalid mode");
294 return;
295 }
296
297 exec_list_push_tail(&shader->variables, &var->node);
298 }
299
300 nir_variable *
nir_variable_create(nir_shader * shader,nir_variable_mode mode,const struct glsl_type * type,const char * name)301 nir_variable_create(nir_shader *shader, nir_variable_mode mode,
302 const struct glsl_type *type, const char *name)
303 {
304 nir_variable *var = rzalloc(shader, nir_variable);
305 var->name = ralloc_strdup(var, name);
306 var->type = type;
307 var->data.mode = mode;
308 var->data.how_declared = nir_var_declared_normally;
309
310 if ((mode == nir_var_shader_in &&
311 shader->info.stage != MESA_SHADER_VERTEX &&
312 shader->info.stage != MESA_SHADER_KERNEL) ||
313 (mode == nir_var_shader_out &&
314 shader->info.stage != MESA_SHADER_FRAGMENT))
315 var->data.interpolation = INTERP_MODE_SMOOTH;
316
317 if (mode == nir_var_shader_in || mode == nir_var_uniform)
318 var->data.read_only = true;
319
320 nir_shader_add_variable(shader, var);
321
322 return var;
323 }
324
325 nir_variable *
nir_local_variable_create(nir_function_impl * impl,const struct glsl_type * type,const char * name)326 nir_local_variable_create(nir_function_impl *impl,
327 const struct glsl_type *type, const char *name)
328 {
329 nir_variable *var = rzalloc(impl->function->shader, nir_variable);
330 var->name = ralloc_strdup(var, name);
331 var->type = type;
332 var->data.mode = nir_var_function_temp;
333
334 nir_function_impl_add_variable(impl, var);
335
336 return var;
337 }
338
339 nir_variable *
nir_find_variable_with_location(nir_shader * shader,nir_variable_mode mode,unsigned location)340 nir_find_variable_with_location(nir_shader *shader,
341 nir_variable_mode mode,
342 unsigned location)
343 {
344 assert(util_bitcount(mode) == 1 && mode != nir_var_function_temp);
345 nir_foreach_variable_with_modes(var, shader, mode) {
346 if (var->data.location == location)
347 return var;
348 }
349 return NULL;
350 }
351
352 nir_variable *
nir_find_variable_with_driver_location(nir_shader * shader,nir_variable_mode mode,unsigned location)353 nir_find_variable_with_driver_location(nir_shader *shader,
354 nir_variable_mode mode,
355 unsigned location)
356 {
357 assert(util_bitcount(mode) == 1 && mode != nir_var_function_temp);
358 nir_foreach_variable_with_modes(var, shader, mode) {
359 if (var->data.driver_location == location)
360 return var;
361 }
362 return NULL;
363 }
364
365 /* Annoyingly, qsort_r is not in the C standard library and, in particular, we
366 * can't count on it on MSV and Android. So we stuff the CMP function into
367 * each array element. It's a bit messy and burns more memory but the list of
368 * variables should hever be all that long.
369 */
370 struct var_cmp {
371 nir_variable *var;
372 int (*cmp)(const nir_variable *, const nir_variable *);
373 };
374
375 static int
var_sort_cmp(const void * _a,const void * _b,void * _cmp)376 var_sort_cmp(const void *_a, const void *_b, void *_cmp)
377 {
378 const struct var_cmp *a = _a;
379 const struct var_cmp *b = _b;
380 assert(a->cmp == b->cmp);
381 return a->cmp(a->var, b->var);
382 }
383
384 void
nir_sort_variables_with_modes(nir_shader * shader,int (* cmp)(const nir_variable *,const nir_variable *),nir_variable_mode modes)385 nir_sort_variables_with_modes(nir_shader *shader,
386 int (*cmp)(const nir_variable *,
387 const nir_variable *),
388 nir_variable_mode modes)
389 {
390 unsigned num_vars = 0;
391 nir_foreach_variable_with_modes(var, shader, modes) {
392 ++num_vars;
393 }
394 struct var_cmp *vars = ralloc_array(shader, struct var_cmp, num_vars);
395 unsigned i = 0;
396 nir_foreach_variable_with_modes_safe(var, shader, modes) {
397 exec_node_remove(&var->node);
398 vars[i++] = (struct var_cmp){
399 .var = var,
400 .cmp = cmp,
401 };
402 }
403 assert(i == num_vars);
404
405 util_qsort_r(vars, num_vars, sizeof(*vars), var_sort_cmp, cmp);
406
407 for (i = 0; i < num_vars; i++)
408 exec_list_push_tail(&shader->variables, &vars[i].var->node);
409
410 ralloc_free(vars);
411 }
412
413 nir_function *
nir_function_create(nir_shader * shader,const char * name)414 nir_function_create(nir_shader *shader, const char *name)
415 {
416 nir_function *func = ralloc(shader, nir_function);
417
418 exec_list_push_tail(&shader->functions, &func->node);
419
420 func->name = ralloc_strdup(func, name);
421 func->shader = shader;
422 func->num_params = 0;
423 func->params = NULL;
424 func->impl = NULL;
425 func->is_entrypoint = false;
426 func->is_preamble = false;
427
428 return func;
429 }
430
src_has_indirect(nir_src * src)431 static bool src_has_indirect(nir_src *src)
432 {
433 return !src->is_ssa && src->reg.indirect;
434 }
435
src_free_indirects(nir_src * src)436 static void src_free_indirects(nir_src *src)
437 {
438 if (src_has_indirect(src)) {
439 assert(src->reg.indirect->is_ssa || !src->reg.indirect->reg.indirect);
440 free(src->reg.indirect);
441 src->reg.indirect = NULL;
442 }
443 }
444
dest_free_indirects(nir_dest * dest)445 static void dest_free_indirects(nir_dest *dest)
446 {
447 if (!dest->is_ssa && dest->reg.indirect) {
448 assert(dest->reg.indirect->is_ssa || !dest->reg.indirect->reg.indirect);
449 free(dest->reg.indirect);
450 dest->reg.indirect = NULL;
451 }
452 }
453
454 /* NOTE: if the instruction you are copying a src to is already added
455 * to the IR, use nir_instr_rewrite_src() instead.
456 */
nir_src_copy(nir_src * dest,const nir_src * src)457 void nir_src_copy(nir_src *dest, const nir_src *src)
458 {
459 src_free_indirects(dest);
460
461 dest->is_ssa = src->is_ssa;
462 if (src->is_ssa) {
463 dest->ssa = src->ssa;
464 } else {
465 dest->reg.base_offset = src->reg.base_offset;
466 dest->reg.reg = src->reg.reg;
467 if (src->reg.indirect) {
468 dest->reg.indirect = calloc(1, sizeof(nir_src));
469 nir_src_copy(dest->reg.indirect, src->reg.indirect);
470 } else {
471 dest->reg.indirect = NULL;
472 }
473 }
474 }
475
nir_dest_copy(nir_dest * dest,const nir_dest * src)476 void nir_dest_copy(nir_dest *dest, const nir_dest *src)
477 {
478 /* Copying an SSA definition makes no sense whatsoever. */
479 assert(!src->is_ssa);
480
481 dest_free_indirects(dest);
482
483 dest->is_ssa = false;
484
485 dest->reg.base_offset = src->reg.base_offset;
486 dest->reg.reg = src->reg.reg;
487 if (src->reg.indirect) {
488 dest->reg.indirect = calloc(1, sizeof(nir_src));
489 nir_src_copy(dest->reg.indirect, src->reg.indirect);
490 } else {
491 dest->reg.indirect = NULL;
492 }
493 }
494
495 void
nir_alu_src_copy(nir_alu_src * dest,const nir_alu_src * src)496 nir_alu_src_copy(nir_alu_src *dest, const nir_alu_src *src)
497 {
498 nir_src_copy(&dest->src, &src->src);
499 dest->abs = src->abs;
500 dest->negate = src->negate;
501 for (unsigned i = 0; i < NIR_MAX_VEC_COMPONENTS; i++)
502 dest->swizzle[i] = src->swizzle[i];
503 }
504
505 void
nir_alu_dest_copy(nir_alu_dest * dest,const nir_alu_dest * src)506 nir_alu_dest_copy(nir_alu_dest *dest, const nir_alu_dest *src)
507 {
508 nir_dest_copy(&dest->dest, &src->dest);
509 dest->write_mask = src->write_mask;
510 dest->saturate = src->saturate;
511 }
512
513 bool
nir_alu_src_is_trivial_ssa(const nir_alu_instr * alu,unsigned srcn)514 nir_alu_src_is_trivial_ssa(const nir_alu_instr *alu, unsigned srcn)
515 {
516 static uint8_t trivial_swizzle[] = { 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15 };
517 STATIC_ASSERT(ARRAY_SIZE(trivial_swizzle) == NIR_MAX_VEC_COMPONENTS);
518
519 const nir_alu_src *src = &alu->src[srcn];
520 unsigned num_components = nir_ssa_alu_instr_src_components(alu, srcn);
521
522 return src->src.is_ssa && (src->src.ssa->num_components == num_components) &&
523 !src->abs && !src->negate &&
524 (memcmp(src->swizzle, trivial_swizzle, num_components) == 0);
525 }
526
527
528 static void
cf_init(nir_cf_node * node,nir_cf_node_type type)529 cf_init(nir_cf_node *node, nir_cf_node_type type)
530 {
531 exec_node_init(&node->node);
532 node->parent = NULL;
533 node->type = type;
534 }
535
536 nir_function_impl *
nir_function_impl_create_bare(nir_shader * shader)537 nir_function_impl_create_bare(nir_shader *shader)
538 {
539 nir_function_impl *impl = ralloc(shader, nir_function_impl);
540
541 impl->function = NULL;
542 impl->preamble = NULL;
543
544 cf_init(&impl->cf_node, nir_cf_node_function);
545
546 exec_list_make_empty(&impl->body);
547 exec_list_make_empty(&impl->registers);
548 exec_list_make_empty(&impl->locals);
549 impl->reg_alloc = 0;
550 impl->ssa_alloc = 0;
551 impl->num_blocks = 0;
552 impl->valid_metadata = nir_metadata_none;
553 impl->structured = true;
554
555 /* create start & end blocks */
556 nir_block *start_block = nir_block_create(shader);
557 nir_block *end_block = nir_block_create(shader);
558 start_block->cf_node.parent = &impl->cf_node;
559 end_block->cf_node.parent = &impl->cf_node;
560 impl->end_block = end_block;
561
562 exec_list_push_tail(&impl->body, &start_block->cf_node.node);
563
564 start_block->successors[0] = end_block;
565 _mesa_set_add(end_block->predecessors, start_block);
566 return impl;
567 }
568
569 nir_function_impl *
nir_function_impl_create(nir_function * function)570 nir_function_impl_create(nir_function *function)
571 {
572 assert(function->impl == NULL);
573
574 nir_function_impl *impl = nir_function_impl_create_bare(function->shader);
575
576 function->impl = impl;
577 impl->function = function;
578
579 return impl;
580 }
581
582 nir_block *
nir_block_create(nir_shader * shader)583 nir_block_create(nir_shader *shader)
584 {
585 nir_block *block = rzalloc(shader, nir_block);
586
587 cf_init(&block->cf_node, nir_cf_node_block);
588
589 block->successors[0] = block->successors[1] = NULL;
590 block->predecessors = _mesa_pointer_set_create(block);
591 block->imm_dom = NULL;
592 /* XXX maybe it would be worth it to defer allocation? This
593 * way it doesn't get allocated for shader refs that never run
594 * nir_calc_dominance? For example, state-tracker creates an
595 * initial IR, clones that, runs appropriate lowering pass, passes
596 * to driver which does common lowering/opt, and then stores ref
597 * which is later used to do state specific lowering and futher
598 * opt. Do any of the references not need dominance metadata?
599 */
600 block->dom_frontier = _mesa_pointer_set_create(block);
601
602 exec_list_make_empty(&block->instr_list);
603
604 return block;
605 }
606
607 static inline void
src_init(nir_src * src)608 src_init(nir_src *src)
609 {
610 src->is_ssa = false;
611 src->reg.reg = NULL;
612 src->reg.indirect = NULL;
613 src->reg.base_offset = 0;
614 }
615
616 nir_if *
nir_if_create(nir_shader * shader)617 nir_if_create(nir_shader *shader)
618 {
619 nir_if *if_stmt = ralloc(shader, nir_if);
620
621 if_stmt->control = nir_selection_control_none;
622
623 cf_init(&if_stmt->cf_node, nir_cf_node_if);
624 src_init(&if_stmt->condition);
625
626 nir_block *then = nir_block_create(shader);
627 exec_list_make_empty(&if_stmt->then_list);
628 exec_list_push_tail(&if_stmt->then_list, &then->cf_node.node);
629 then->cf_node.parent = &if_stmt->cf_node;
630
631 nir_block *else_stmt = nir_block_create(shader);
632 exec_list_make_empty(&if_stmt->else_list);
633 exec_list_push_tail(&if_stmt->else_list, &else_stmt->cf_node.node);
634 else_stmt->cf_node.parent = &if_stmt->cf_node;
635
636 return if_stmt;
637 }
638
639 nir_loop *
nir_loop_create(nir_shader * shader)640 nir_loop_create(nir_shader *shader)
641 {
642 nir_loop *loop = rzalloc(shader, nir_loop);
643
644 cf_init(&loop->cf_node, nir_cf_node_loop);
645 /* Assume that loops are divergent until proven otherwise */
646 loop->divergent = true;
647
648 nir_block *body = nir_block_create(shader);
649 exec_list_make_empty(&loop->body);
650 exec_list_push_tail(&loop->body, &body->cf_node.node);
651 body->cf_node.parent = &loop->cf_node;
652
653 body->successors[0] = body;
654 _mesa_set_add(body->predecessors, body);
655
656 return loop;
657 }
658
659 static void
instr_init(nir_instr * instr,nir_instr_type type)660 instr_init(nir_instr *instr, nir_instr_type type)
661 {
662 instr->type = type;
663 instr->block = NULL;
664 exec_node_init(&instr->node);
665 }
666
667 static void
dest_init(nir_dest * dest)668 dest_init(nir_dest *dest)
669 {
670 dest->is_ssa = false;
671 dest->reg.reg = NULL;
672 dest->reg.indirect = NULL;
673 dest->reg.base_offset = 0;
674 }
675
676 static void
alu_dest_init(nir_alu_dest * dest)677 alu_dest_init(nir_alu_dest *dest)
678 {
679 dest_init(&dest->dest);
680 dest->saturate = false;
681 dest->write_mask = 0xf;
682 }
683
684 static void
alu_src_init(nir_alu_src * src)685 alu_src_init(nir_alu_src *src)
686 {
687 src_init(&src->src);
688 src->abs = src->negate = false;
689 for (int i = 0; i < NIR_MAX_VEC_COMPONENTS; ++i)
690 src->swizzle[i] = i;
691 }
692
693 nir_alu_instr *
nir_alu_instr_create(nir_shader * shader,nir_op op)694 nir_alu_instr_create(nir_shader *shader, nir_op op)
695 {
696 unsigned num_srcs = nir_op_infos[op].num_inputs;
697 /* TODO: don't use calloc */
698 nir_alu_instr *instr = calloc(1, sizeof(nir_alu_instr) + num_srcs * sizeof(nir_alu_src));
699
700 instr_init(&instr->instr, nir_instr_type_alu);
701 instr->op = op;
702 alu_dest_init(&instr->dest);
703 for (unsigned i = 0; i < num_srcs; i++)
704 alu_src_init(&instr->src[i]);
705
706 list_add(&instr->instr.gc_node, &shader->gc_list);
707
708 return instr;
709 }
710
711 nir_deref_instr *
nir_deref_instr_create(nir_shader * shader,nir_deref_type deref_type)712 nir_deref_instr_create(nir_shader *shader, nir_deref_type deref_type)
713 {
714 nir_deref_instr *instr = calloc(1, sizeof(*instr));
715
716 instr_init(&instr->instr, nir_instr_type_deref);
717
718 instr->deref_type = deref_type;
719 if (deref_type != nir_deref_type_var)
720 src_init(&instr->parent);
721
722 if (deref_type == nir_deref_type_array ||
723 deref_type == nir_deref_type_ptr_as_array)
724 src_init(&instr->arr.index);
725
726 dest_init(&instr->dest);
727
728 list_add(&instr->instr.gc_node, &shader->gc_list);
729
730 return instr;
731 }
732
733 nir_jump_instr *
nir_jump_instr_create(nir_shader * shader,nir_jump_type type)734 nir_jump_instr_create(nir_shader *shader, nir_jump_type type)
735 {
736 nir_jump_instr *instr = malloc(sizeof(*instr));
737 instr_init(&instr->instr, nir_instr_type_jump);
738 src_init(&instr->condition);
739 instr->type = type;
740 instr->target = NULL;
741 instr->else_target = NULL;
742
743 list_add(&instr->instr.gc_node, &shader->gc_list);
744
745 return instr;
746 }
747
748 nir_load_const_instr *
nir_load_const_instr_create(nir_shader * shader,unsigned num_components,unsigned bit_size)749 nir_load_const_instr_create(nir_shader *shader, unsigned num_components,
750 unsigned bit_size)
751 {
752 nir_load_const_instr *instr =
753 calloc(1, sizeof(*instr) + num_components * sizeof(*instr->value));
754 instr_init(&instr->instr, nir_instr_type_load_const);
755
756 nir_ssa_def_init(&instr->instr, &instr->def, num_components, bit_size);
757
758 list_add(&instr->instr.gc_node, &shader->gc_list);
759
760 return instr;
761 }
762
763 nir_intrinsic_instr *
nir_intrinsic_instr_create(nir_shader * shader,nir_intrinsic_op op)764 nir_intrinsic_instr_create(nir_shader *shader, nir_intrinsic_op op)
765 {
766 unsigned num_srcs = nir_intrinsic_infos[op].num_srcs;
767 /* TODO: don't use calloc */
768 nir_intrinsic_instr *instr =
769 calloc(1, sizeof(nir_intrinsic_instr) + num_srcs * sizeof(nir_src));
770
771 instr_init(&instr->instr, nir_instr_type_intrinsic);
772 instr->intrinsic = op;
773
774 if (nir_intrinsic_infos[op].has_dest)
775 dest_init(&instr->dest);
776
777 for (unsigned i = 0; i < num_srcs; i++)
778 src_init(&instr->src[i]);
779
780 list_add(&instr->instr.gc_node, &shader->gc_list);
781
782 return instr;
783 }
784
785 nir_call_instr *
nir_call_instr_create(nir_shader * shader,nir_function * callee)786 nir_call_instr_create(nir_shader *shader, nir_function *callee)
787 {
788 const unsigned num_params = callee->num_params;
789 nir_call_instr *instr =
790 calloc(1, sizeof(*instr) + num_params * sizeof(instr->params[0]));
791
792 instr_init(&instr->instr, nir_instr_type_call);
793 instr->callee = callee;
794 instr->num_params = num_params;
795 for (unsigned i = 0; i < num_params; i++)
796 src_init(&instr->params[i]);
797
798 list_add(&instr->instr.gc_node, &shader->gc_list);
799
800 return instr;
801 }
802
803 static int8_t default_tg4_offsets[4][2] =
804 {
805 { 0, 1 },
806 { 1, 1 },
807 { 1, 0 },
808 { 0, 0 },
809 };
810
811 nir_tex_instr *
nir_tex_instr_create(nir_shader * shader,unsigned num_srcs)812 nir_tex_instr_create(nir_shader *shader, unsigned num_srcs)
813 {
814 nir_tex_instr *instr = calloc(1, sizeof(*instr));
815 instr_init(&instr->instr, nir_instr_type_tex);
816
817 dest_init(&instr->dest);
818
819 instr->num_srcs = num_srcs;
820 instr->src = malloc(sizeof(nir_tex_src) * num_srcs);
821 for (unsigned i = 0; i < num_srcs; i++)
822 src_init(&instr->src[i].src);
823
824 instr->texture_index = 0;
825 instr->sampler_index = 0;
826 memcpy(instr->tg4_offsets, default_tg4_offsets, sizeof(instr->tg4_offsets));
827
828 list_add(&instr->instr.gc_node, &shader->gc_list);
829
830 return instr;
831 }
832
833 void
nir_tex_instr_add_src(nir_tex_instr * tex,nir_tex_src_type src_type,nir_src src)834 nir_tex_instr_add_src(nir_tex_instr *tex,
835 nir_tex_src_type src_type,
836 nir_src src)
837 {
838 nir_tex_src *new_srcs = calloc(sizeof(*new_srcs),
839 tex->num_srcs + 1);
840
841 for (unsigned i = 0; i < tex->num_srcs; i++) {
842 new_srcs[i].src_type = tex->src[i].src_type;
843 nir_instr_move_src(&tex->instr, &new_srcs[i].src,
844 &tex->src[i].src);
845 }
846
847 free(tex->src);
848 tex->src = new_srcs;
849
850 tex->src[tex->num_srcs].src_type = src_type;
851 nir_instr_rewrite_src(&tex->instr, &tex->src[tex->num_srcs].src, src);
852 tex->num_srcs++;
853 }
854
855 void
nir_tex_instr_remove_src(nir_tex_instr * tex,unsigned src_idx)856 nir_tex_instr_remove_src(nir_tex_instr *tex, unsigned src_idx)
857 {
858 assert(src_idx < tex->num_srcs);
859
860 /* First rewrite the source to NIR_SRC_INIT */
861 nir_instr_rewrite_src(&tex->instr, &tex->src[src_idx].src, NIR_SRC_INIT);
862
863 /* Now, move all of the other sources down */
864 for (unsigned i = src_idx + 1; i < tex->num_srcs; i++) {
865 tex->src[i-1].src_type = tex->src[i].src_type;
866 nir_instr_move_src(&tex->instr, &tex->src[i-1].src, &tex->src[i].src);
867 }
868 tex->num_srcs--;
869 }
870
871 bool
nir_tex_instr_has_explicit_tg4_offsets(nir_tex_instr * tex)872 nir_tex_instr_has_explicit_tg4_offsets(nir_tex_instr *tex)
873 {
874 if (tex->op != nir_texop_tg4)
875 return false;
876 return memcmp(tex->tg4_offsets, default_tg4_offsets,
877 sizeof(tex->tg4_offsets)) != 0;
878 }
879
880 nir_phi_instr *
nir_phi_instr_create(nir_shader * shader)881 nir_phi_instr_create(nir_shader *shader)
882 {
883 nir_phi_instr *instr = malloc(sizeof(*instr));
884 instr_init(&instr->instr, nir_instr_type_phi);
885
886 dest_init(&instr->dest);
887 exec_list_make_empty(&instr->srcs);
888
889 list_add(&instr->instr.gc_node, &shader->gc_list);
890
891 return instr;
892 }
893
894 /**
895 * Adds a new source to a NIR instruction.
896 *
897 * Note that this does not update the def/use relationship for src, assuming
898 * that the instr is not in the shader. If it is, you have to do:
899 *
900 * list_addtail(&phi_src->src.use_link, &src.ssa->uses);
901 */
902 nir_phi_src *
nir_phi_instr_add_src(nir_phi_instr * instr,nir_block * pred,nir_src src)903 nir_phi_instr_add_src(nir_phi_instr *instr, nir_block *pred, nir_src src)
904 {
905 nir_phi_src *phi_src;
906
907 phi_src = calloc(1, sizeof(nir_phi_src));
908 phi_src->pred = pred;
909 phi_src->src = src;
910 phi_src->src.parent_instr = &instr->instr;
911 exec_list_push_tail(&instr->srcs, &phi_src->node);
912
913 return phi_src;
914 }
915
916 nir_parallel_copy_instr *
nir_parallel_copy_instr_create(nir_shader * shader)917 nir_parallel_copy_instr_create(nir_shader *shader)
918 {
919 nir_parallel_copy_instr *instr = malloc(sizeof(*instr));
920 instr_init(&instr->instr, nir_instr_type_parallel_copy);
921
922 exec_list_make_empty(&instr->entries);
923
924 list_add(&instr->instr.gc_node, &shader->gc_list);
925
926 return instr;
927 }
928
929 nir_ssa_undef_instr *
nir_ssa_undef_instr_create(nir_shader * shader,unsigned num_components,unsigned bit_size)930 nir_ssa_undef_instr_create(nir_shader *shader,
931 unsigned num_components,
932 unsigned bit_size)
933 {
934 nir_ssa_undef_instr *instr = malloc(sizeof(*instr));
935 instr_init(&instr->instr, nir_instr_type_ssa_undef);
936
937 nir_ssa_def_init(&instr->instr, &instr->def, num_components, bit_size);
938
939 list_add(&instr->instr.gc_node, &shader->gc_list);
940
941 return instr;
942 }
943
944 static nir_const_value
const_value_float(double d,unsigned bit_size)945 const_value_float(double d, unsigned bit_size)
946 {
947 nir_const_value v;
948 memset(&v, 0, sizeof(v));
949 switch (bit_size) {
950 case 16: v.u16 = _mesa_float_to_half(d); break;
951 case 32: v.f32 = d; break;
952 case 64: v.f64 = d; break;
953 default:
954 unreachable("Invalid bit size");
955 }
956 return v;
957 }
958
959 static nir_const_value
const_value_int(int64_t i,unsigned bit_size)960 const_value_int(int64_t i, unsigned bit_size)
961 {
962 nir_const_value v;
963 memset(&v, 0, sizeof(v));
964 switch (bit_size) {
965 case 1: v.b = i & 1; break;
966 case 8: v.i8 = i; break;
967 case 16: v.i16 = i; break;
968 case 32: v.i32 = i; break;
969 case 64: v.i64 = i; break;
970 default:
971 unreachable("Invalid bit size");
972 }
973 return v;
974 }
975
976 nir_const_value
nir_alu_binop_identity(nir_op binop,unsigned bit_size)977 nir_alu_binop_identity(nir_op binop, unsigned bit_size)
978 {
979 const int64_t max_int = (1ull << (bit_size - 1)) - 1;
980 const int64_t min_int = -max_int - 1;
981 switch (binop) {
982 case nir_op_iadd:
983 return const_value_int(0, bit_size);
984 case nir_op_fadd:
985 return const_value_float(0, bit_size);
986 case nir_op_imul:
987 return const_value_int(1, bit_size);
988 case nir_op_fmul:
989 return const_value_float(1, bit_size);
990 case nir_op_imin:
991 return const_value_int(max_int, bit_size);
992 case nir_op_umin:
993 return const_value_int(~0ull, bit_size);
994 case nir_op_fmin:
995 return const_value_float(INFINITY, bit_size);
996 case nir_op_imax:
997 return const_value_int(min_int, bit_size);
998 case nir_op_umax:
999 return const_value_int(0, bit_size);
1000 case nir_op_fmax:
1001 return const_value_float(-INFINITY, bit_size);
1002 case nir_op_iand:
1003 return const_value_int(~0ull, bit_size);
1004 case nir_op_ior:
1005 return const_value_int(0, bit_size);
1006 case nir_op_ixor:
1007 return const_value_int(0, bit_size);
1008 default:
1009 unreachable("Invalid reduction operation");
1010 }
1011 }
1012
1013 nir_function_impl *
nir_cf_node_get_function(nir_cf_node * node)1014 nir_cf_node_get_function(nir_cf_node *node)
1015 {
1016 while (node->type != nir_cf_node_function) {
1017 node = node->parent;
1018 }
1019
1020 return nir_cf_node_as_function(node);
1021 }
1022
1023 /* Reduces a cursor by trying to convert everything to after and trying to
1024 * go up to block granularity when possible.
1025 */
1026 static nir_cursor
reduce_cursor(nir_cursor cursor)1027 reduce_cursor(nir_cursor cursor)
1028 {
1029 switch (cursor.option) {
1030 case nir_cursor_before_block:
1031 if (exec_list_is_empty(&cursor.block->instr_list)) {
1032 /* Empty block. After is as good as before. */
1033 cursor.option = nir_cursor_after_block;
1034 }
1035 return cursor;
1036
1037 case nir_cursor_after_block:
1038 return cursor;
1039
1040 case nir_cursor_before_instr: {
1041 nir_instr *prev_instr = nir_instr_prev(cursor.instr);
1042 if (prev_instr) {
1043 /* Before this instruction is after the previous */
1044 cursor.instr = prev_instr;
1045 cursor.option = nir_cursor_after_instr;
1046 } else {
1047 /* No previous instruction. Switch to before block */
1048 cursor.block = cursor.instr->block;
1049 cursor.option = nir_cursor_before_block;
1050 }
1051 return reduce_cursor(cursor);
1052 }
1053
1054 case nir_cursor_after_instr:
1055 if (nir_instr_next(cursor.instr) == NULL) {
1056 /* This is the last instruction, switch to after block */
1057 cursor.option = nir_cursor_after_block;
1058 cursor.block = cursor.instr->block;
1059 }
1060 return cursor;
1061
1062 default:
1063 unreachable("Inavlid cursor option");
1064 }
1065 }
1066
1067 bool
nir_cursors_equal(nir_cursor a,nir_cursor b)1068 nir_cursors_equal(nir_cursor a, nir_cursor b)
1069 {
1070 /* Reduced cursors should be unique */
1071 a = reduce_cursor(a);
1072 b = reduce_cursor(b);
1073
1074 return a.block == b.block && a.option == b.option;
1075 }
1076
1077 static bool
add_use_cb(nir_src * src,void * state)1078 add_use_cb(nir_src *src, void *state)
1079 {
1080 nir_instr *instr = state;
1081
1082 src->parent_instr = instr;
1083 list_addtail(&src->use_link,
1084 src->is_ssa ? &src->ssa->uses : &src->reg.reg->uses);
1085
1086 return true;
1087 }
1088
1089 static bool
add_ssa_def_cb(nir_ssa_def * def,void * state)1090 add_ssa_def_cb(nir_ssa_def *def, void *state)
1091 {
1092 nir_instr *instr = state;
1093
1094 if (instr->block && def->index == UINT_MAX) {
1095 nir_function_impl *impl =
1096 nir_cf_node_get_function(&instr->block->cf_node);
1097
1098 def->index = impl->ssa_alloc++;
1099
1100 impl->valid_metadata &= ~nir_metadata_live_ssa_defs;
1101 }
1102
1103 return true;
1104 }
1105
1106 static bool
add_reg_def_cb(nir_dest * dest,void * state)1107 add_reg_def_cb(nir_dest *dest, void *state)
1108 {
1109 nir_instr *instr = state;
1110
1111 if (!dest->is_ssa) {
1112 dest->reg.parent_instr = instr;
1113 list_addtail(&dest->reg.def_link, &dest->reg.reg->defs);
1114 }
1115
1116 return true;
1117 }
1118
1119 static void
add_defs_uses(nir_instr * instr)1120 add_defs_uses(nir_instr *instr)
1121 {
1122 nir_foreach_src(instr, add_use_cb, instr);
1123 nir_foreach_dest(instr, add_reg_def_cb, instr);
1124 nir_foreach_ssa_def(instr, add_ssa_def_cb, instr);
1125 }
1126
1127 void
nir_instr_insert(nir_cursor cursor,nir_instr * instr)1128 nir_instr_insert(nir_cursor cursor, nir_instr *instr)
1129 {
1130 switch (cursor.option) {
1131 case nir_cursor_before_block:
1132 /* Only allow inserting jumps into empty blocks. */
1133 if (instr->type == nir_instr_type_jump)
1134 assert(exec_list_is_empty(&cursor.block->instr_list));
1135
1136 instr->block = cursor.block;
1137 add_defs_uses(instr);
1138 exec_list_push_head(&cursor.block->instr_list, &instr->node);
1139 break;
1140 case nir_cursor_after_block: {
1141 /* Inserting instructions after a jump is illegal. */
1142 nir_instr *last = nir_block_last_instr(cursor.block);
1143 assert(last == NULL || last->type != nir_instr_type_jump);
1144 (void) last;
1145
1146 instr->block = cursor.block;
1147 add_defs_uses(instr);
1148 exec_list_push_tail(&cursor.block->instr_list, &instr->node);
1149 break;
1150 }
1151 case nir_cursor_before_instr:
1152 assert(instr->type != nir_instr_type_jump);
1153 instr->block = cursor.instr->block;
1154 add_defs_uses(instr);
1155 exec_node_insert_node_before(&cursor.instr->node, &instr->node);
1156 break;
1157 case nir_cursor_after_instr:
1158 /* Inserting instructions after a jump is illegal. */
1159 assert(cursor.instr->type != nir_instr_type_jump);
1160
1161 /* Only allow inserting jumps at the end of the block. */
1162 if (instr->type == nir_instr_type_jump)
1163 assert(cursor.instr == nir_block_last_instr(cursor.instr->block));
1164
1165 instr->block = cursor.instr->block;
1166 add_defs_uses(instr);
1167 exec_node_insert_after(&cursor.instr->node, &instr->node);
1168 break;
1169 }
1170
1171 if (instr->type == nir_instr_type_jump)
1172 nir_handle_add_jump(instr->block);
1173
1174 nir_function_impl *impl = nir_cf_node_get_function(&instr->block->cf_node);
1175 impl->valid_metadata &= ~nir_metadata_instr_index;
1176 }
1177
1178 bool
nir_instr_move(nir_cursor cursor,nir_instr * instr)1179 nir_instr_move(nir_cursor cursor, nir_instr *instr)
1180 {
1181 /* If the cursor happens to refer to this instruction (either before or
1182 * after), don't do anything.
1183 */
1184 if ((cursor.option == nir_cursor_before_instr ||
1185 cursor.option == nir_cursor_after_instr) &&
1186 cursor.instr == instr)
1187 return false;
1188
1189 nir_instr_remove(instr);
1190 nir_instr_insert(cursor, instr);
1191 return true;
1192 }
1193
1194 static bool
src_is_valid(const nir_src * src)1195 src_is_valid(const nir_src *src)
1196 {
1197 return src->is_ssa ? (src->ssa != NULL) : (src->reg.reg != NULL);
1198 }
1199
1200 static bool
remove_use_cb(nir_src * src,void * state)1201 remove_use_cb(nir_src *src, void *state)
1202 {
1203 (void) state;
1204
1205 if (src_is_valid(src))
1206 list_del(&src->use_link);
1207
1208 return true;
1209 }
1210
1211 static bool
remove_def_cb(nir_dest * dest,void * state)1212 remove_def_cb(nir_dest *dest, void *state)
1213 {
1214 (void) state;
1215
1216 if (!dest->is_ssa)
1217 list_del(&dest->reg.def_link);
1218
1219 return true;
1220 }
1221
1222 static void
remove_defs_uses(nir_instr * instr)1223 remove_defs_uses(nir_instr *instr)
1224 {
1225 nir_foreach_dest(instr, remove_def_cb, instr);
1226 nir_foreach_src(instr, remove_use_cb, instr);
1227 }
1228
nir_instr_remove_v(nir_instr * instr)1229 void nir_instr_remove_v(nir_instr *instr)
1230 {
1231 remove_defs_uses(instr);
1232 exec_node_remove(&instr->node);
1233
1234 if (instr->type == nir_instr_type_jump) {
1235 nir_jump_instr *jump_instr = nir_instr_as_jump(instr);
1236 nir_handle_remove_jump(instr->block, jump_instr->type);
1237 }
1238 }
1239
free_src_indirects_cb(nir_src * src,void * state)1240 static bool free_src_indirects_cb(nir_src *src, void *state)
1241 {
1242 src_free_indirects(src);
1243 return true;
1244 }
1245
free_dest_indirects_cb(nir_dest * dest,void * state)1246 static bool free_dest_indirects_cb(nir_dest *dest, void *state)
1247 {
1248 dest_free_indirects(dest);
1249 return true;
1250 }
1251
nir_instr_free(nir_instr * instr)1252 void nir_instr_free(nir_instr *instr)
1253 {
1254 nir_foreach_src(instr, free_src_indirects_cb, NULL);
1255 nir_foreach_dest(instr, free_dest_indirects_cb, NULL);
1256
1257 switch (instr->type) {
1258 case nir_instr_type_tex:
1259 free(nir_instr_as_tex(instr)->src);
1260 break;
1261
1262 case nir_instr_type_phi: {
1263 nir_phi_instr *phi = nir_instr_as_phi(instr);
1264 nir_foreach_phi_src_safe(phi_src, phi) {
1265 free(phi_src);
1266 }
1267 break;
1268 }
1269
1270 default:
1271 break;
1272 }
1273
1274 list_del(&instr->gc_node);
1275 free(instr);
1276 }
1277
1278 void
nir_instr_free_list(struct exec_list * list)1279 nir_instr_free_list(struct exec_list *list)
1280 {
1281 struct exec_node *node;
1282 while ((node = exec_list_pop_head(list))) {
1283 nir_instr *removed_instr = exec_node_data(nir_instr, node, node);
1284 nir_instr_free(removed_instr);
1285 }
1286 }
1287
nir_instr_free_and_dce_live_cb(nir_ssa_def * def,void * state)1288 static bool nir_instr_free_and_dce_live_cb(nir_ssa_def *def, void *state)
1289 {
1290 bool *live = state;
1291
1292 if (!nir_ssa_def_is_unused(def)) {
1293 *live = true;
1294 return false;
1295 } else {
1296 return true;
1297 }
1298 }
1299
nir_instr_free_and_dce_is_live(nir_instr * instr)1300 static bool nir_instr_free_and_dce_is_live(nir_instr *instr)
1301 {
1302 /* Note: don't have to worry about jumps because they don't have dests to
1303 * become unused.
1304 */
1305 if (instr->type == nir_instr_type_intrinsic) {
1306 nir_intrinsic_instr *intr = nir_instr_as_intrinsic(instr);
1307 const nir_intrinsic_info *info = &nir_intrinsic_infos[intr->intrinsic];
1308 if (!(info->flags & NIR_INTRINSIC_CAN_ELIMINATE))
1309 return true;
1310 }
1311
1312 bool live = false;
1313 nir_foreach_ssa_def(instr, nir_instr_free_and_dce_live_cb, &live);
1314 return live;
1315 }
1316
1317 static bool
nir_instr_dce_add_dead_srcs_cb(nir_src * src,void * state)1318 nir_instr_dce_add_dead_srcs_cb(nir_src *src, void *state)
1319 {
1320 nir_instr_worklist *wl = state;
1321
1322 if (src->is_ssa) {
1323 list_del(&src->use_link);
1324 if (!nir_instr_free_and_dce_is_live(src->ssa->parent_instr))
1325 nir_instr_worklist_push_tail(wl, src->ssa->parent_instr);
1326
1327 /* Stop nir_instr_remove from trying to delete the link again. */
1328 src->ssa = NULL;
1329 }
1330
1331 return true;
1332 }
1333
1334 static void
nir_instr_dce_add_dead_ssa_srcs(nir_instr_worklist * wl,nir_instr * instr)1335 nir_instr_dce_add_dead_ssa_srcs(nir_instr_worklist *wl, nir_instr *instr)
1336 {
1337 nir_foreach_src(instr, nir_instr_dce_add_dead_srcs_cb, wl);
1338 }
1339
1340 /**
1341 * Frees an instruction and any SSA defs that it used that are now dead,
1342 * returning a nir_cursor where the instruction previously was.
1343 */
1344 nir_cursor
nir_instr_free_and_dce(nir_instr * instr)1345 nir_instr_free_and_dce(nir_instr *instr)
1346 {
1347 nir_instr_worklist *worklist = nir_instr_worklist_create();
1348
1349 nir_instr_dce_add_dead_ssa_srcs(worklist, instr);
1350 nir_cursor c = nir_instr_remove(instr);
1351
1352 struct exec_list to_free;
1353 exec_list_make_empty(&to_free);
1354
1355 nir_instr *dce_instr;
1356 while ((dce_instr = nir_instr_worklist_pop_head(worklist))) {
1357 nir_instr_dce_add_dead_ssa_srcs(worklist, dce_instr);
1358
1359 /* If we're removing the instr where our cursor is, then we have to
1360 * point the cursor elsewhere.
1361 */
1362 if ((c.option == nir_cursor_before_instr ||
1363 c.option == nir_cursor_after_instr) &&
1364 c.instr == dce_instr)
1365 c = nir_instr_remove(dce_instr);
1366 else
1367 nir_instr_remove(dce_instr);
1368 exec_list_push_tail(&to_free, &dce_instr->node);
1369 }
1370
1371 nir_instr_free_list(&to_free);
1372
1373 nir_instr_worklist_destroy(worklist);
1374
1375 return c;
1376 }
1377
1378 /*@}*/
1379
1380 void
nir_index_local_regs(nir_function_impl * impl)1381 nir_index_local_regs(nir_function_impl *impl)
1382 {
1383 unsigned index = 0;
1384 foreach_list_typed(nir_register, reg, node, &impl->registers) {
1385 reg->index = index++;
1386 }
1387 impl->reg_alloc = index;
1388 }
1389
1390 struct foreach_ssa_def_state {
1391 nir_foreach_ssa_def_cb cb;
1392 void *client_state;
1393 };
1394
1395 static inline bool
nir_ssa_def_visitor(nir_dest * dest,void * void_state)1396 nir_ssa_def_visitor(nir_dest *dest, void *void_state)
1397 {
1398 struct foreach_ssa_def_state *state = void_state;
1399
1400 if (dest->is_ssa)
1401 return state->cb(&dest->ssa, state->client_state);
1402 else
1403 return true;
1404 }
1405
1406 bool
nir_foreach_ssa_def(nir_instr * instr,nir_foreach_ssa_def_cb cb,void * state)1407 nir_foreach_ssa_def(nir_instr *instr, nir_foreach_ssa_def_cb cb, void *state)
1408 {
1409 switch (instr->type) {
1410 case nir_instr_type_alu:
1411 case nir_instr_type_deref:
1412 case nir_instr_type_tex:
1413 case nir_instr_type_intrinsic:
1414 case nir_instr_type_phi:
1415 case nir_instr_type_parallel_copy: {
1416 struct foreach_ssa_def_state foreach_state = {cb, state};
1417 return nir_foreach_dest(instr, nir_ssa_def_visitor, &foreach_state);
1418 }
1419
1420 case nir_instr_type_load_const:
1421 return cb(&nir_instr_as_load_const(instr)->def, state);
1422 case nir_instr_type_ssa_undef:
1423 return cb(&nir_instr_as_ssa_undef(instr)->def, state);
1424 case nir_instr_type_call:
1425 case nir_instr_type_jump:
1426 return true;
1427 default:
1428 unreachable("Invalid instruction type");
1429 }
1430 }
1431
1432 nir_ssa_def *
nir_instr_ssa_def(nir_instr * instr)1433 nir_instr_ssa_def(nir_instr *instr)
1434 {
1435 switch (instr->type) {
1436 case nir_instr_type_alu:
1437 assert(nir_instr_as_alu(instr)->dest.dest.is_ssa);
1438 return &nir_instr_as_alu(instr)->dest.dest.ssa;
1439
1440 case nir_instr_type_deref:
1441 assert(nir_instr_as_deref(instr)->dest.is_ssa);
1442 return &nir_instr_as_deref(instr)->dest.ssa;
1443
1444 case nir_instr_type_tex:
1445 assert(nir_instr_as_tex(instr)->dest.is_ssa);
1446 return &nir_instr_as_tex(instr)->dest.ssa;
1447
1448 case nir_instr_type_intrinsic: {
1449 nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
1450 if (nir_intrinsic_infos[intrin->intrinsic].has_dest) {
1451 assert(intrin->dest.is_ssa);
1452 return &intrin->dest.ssa;
1453 } else {
1454 return NULL;
1455 }
1456 }
1457
1458 case nir_instr_type_phi:
1459 assert(nir_instr_as_phi(instr)->dest.is_ssa);
1460 return &nir_instr_as_phi(instr)->dest.ssa;
1461
1462 case nir_instr_type_parallel_copy:
1463 unreachable("Parallel copies are unsupported by this function");
1464
1465 case nir_instr_type_load_const:
1466 return &nir_instr_as_load_const(instr)->def;
1467
1468 case nir_instr_type_ssa_undef:
1469 return &nir_instr_as_ssa_undef(instr)->def;
1470
1471 case nir_instr_type_call:
1472 case nir_instr_type_jump:
1473 return NULL;
1474 }
1475
1476 unreachable("Invalid instruction type");
1477 }
1478
1479 bool
nir_instr_def_is_register(nir_instr * instr)1480 nir_instr_def_is_register(nir_instr *instr)
1481 {
1482 switch (instr->type) {
1483 case nir_instr_type_alu:
1484 return !nir_instr_as_alu(instr)->dest.dest.is_ssa;
1485
1486 case nir_instr_type_deref:
1487 return !nir_instr_as_deref(instr)->dest.is_ssa;
1488
1489 case nir_instr_type_tex:
1490 return !nir_instr_as_tex(instr)->dest.is_ssa;
1491
1492 case nir_instr_type_intrinsic: {
1493 nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
1494 return nir_intrinsic_infos[intrin->intrinsic].has_dest &&
1495 !intrin->dest.is_ssa;
1496 }
1497
1498 case nir_instr_type_phi:
1499 return !nir_instr_as_phi(instr)->dest.is_ssa;
1500
1501 case nir_instr_type_parallel_copy:
1502 unreachable("Parallel copies are unsupported by this function");
1503
1504 case nir_instr_type_load_const:
1505 case nir_instr_type_ssa_undef:
1506 return false;
1507
1508 case nir_instr_type_call:
1509 case nir_instr_type_jump:
1510 return false;
1511 }
1512
1513 unreachable("Invalid instruction type");
1514 }
1515
1516 bool
nir_foreach_phi_src_leaving_block(nir_block * block,nir_foreach_src_cb cb,void * state)1517 nir_foreach_phi_src_leaving_block(nir_block *block,
1518 nir_foreach_src_cb cb,
1519 void *state)
1520 {
1521 for (unsigned i = 0; i < ARRAY_SIZE(block->successors); i++) {
1522 if (block->successors[i] == NULL)
1523 continue;
1524
1525 nir_foreach_instr(instr, block->successors[i]) {
1526 if (instr->type != nir_instr_type_phi)
1527 break;
1528
1529 nir_phi_instr *phi = nir_instr_as_phi(instr);
1530 nir_foreach_phi_src(phi_src, phi) {
1531 if (phi_src->pred == block) {
1532 if (!cb(&phi_src->src, state))
1533 return false;
1534 }
1535 }
1536 }
1537 }
1538
1539 return true;
1540 }
1541
1542 nir_const_value
nir_const_value_for_float(double f,unsigned bit_size)1543 nir_const_value_for_float(double f, unsigned bit_size)
1544 {
1545 nir_const_value v;
1546 memset(&v, 0, sizeof(v));
1547
1548 switch (bit_size) {
1549 case 16:
1550 v.u16 = _mesa_float_to_half(f);
1551 break;
1552 case 32:
1553 v.f32 = f;
1554 break;
1555 case 64:
1556 v.f64 = f;
1557 break;
1558 default:
1559 unreachable("Invalid bit size");
1560 }
1561
1562 return v;
1563 }
1564
1565 double
nir_const_value_as_float(nir_const_value value,unsigned bit_size)1566 nir_const_value_as_float(nir_const_value value, unsigned bit_size)
1567 {
1568 switch (bit_size) {
1569 case 16: return _mesa_half_to_float(value.u16);
1570 case 32: return value.f32;
1571 case 64: return value.f64;
1572 default:
1573 unreachable("Invalid bit size");
1574 }
1575 }
1576
1577 nir_const_value *
nir_src_as_const_value(nir_src src)1578 nir_src_as_const_value(nir_src src)
1579 {
1580 if (!src.is_ssa)
1581 return NULL;
1582
1583 if (src.ssa->parent_instr->type != nir_instr_type_load_const)
1584 return NULL;
1585
1586 nir_load_const_instr *load = nir_instr_as_load_const(src.ssa->parent_instr);
1587
1588 return load->value;
1589 }
1590
1591 /**
1592 * Returns true if the source is known to be always uniform. Otherwise it
1593 * returns false which means it may or may not be uniform but it can't be
1594 * determined.
1595 *
1596 * For a more precise analysis of uniform values, use nir_divergence_analysis.
1597 */
1598 bool
nir_src_is_always_uniform(nir_src src)1599 nir_src_is_always_uniform(nir_src src)
1600 {
1601 if (!src.is_ssa)
1602 return false;
1603
1604 /* Constants are trivially uniform */
1605 if (src.ssa->parent_instr->type == nir_instr_type_load_const)
1606 return true;
1607
1608 if (src.ssa->parent_instr->type == nir_instr_type_intrinsic) {
1609 nir_intrinsic_instr *intr = nir_instr_as_intrinsic(src.ssa->parent_instr);
1610 /* As are uniform variables */
1611 if (intr->intrinsic == nir_intrinsic_load_uniform &&
1612 nir_src_is_always_uniform(intr->src[0]))
1613 return true;
1614 /* From the Vulkan specification 15.6.1. Push Constant Interface:
1615 * "Any member of a push constant block that is declared as an array must
1616 * only be accessed with dynamically uniform indices."
1617 */
1618 if (intr->intrinsic == nir_intrinsic_load_push_constant)
1619 return true;
1620 if (intr->intrinsic == nir_intrinsic_load_deref &&
1621 nir_deref_mode_is(nir_src_as_deref(intr->src[0]), nir_var_mem_push_const))
1622 return true;
1623 }
1624
1625 /* Operating together uniform expressions produces a uniform result */
1626 if (src.ssa->parent_instr->type == nir_instr_type_alu) {
1627 nir_alu_instr *alu = nir_instr_as_alu(src.ssa->parent_instr);
1628 for (int i = 0; i < nir_op_infos[alu->op].num_inputs; i++) {
1629 if (!nir_src_is_always_uniform(alu->src[i].src))
1630 return false;
1631 }
1632
1633 return true;
1634 }
1635
1636 /* XXX: this could have many more tests, such as when a sampler function is
1637 * called with uniform arguments.
1638 */
1639 return false;
1640 }
1641
1642 static void
src_remove_all_uses(nir_src * src)1643 src_remove_all_uses(nir_src *src)
1644 {
1645 for (; src; src = src->is_ssa ? NULL : src->reg.indirect) {
1646 if (!src_is_valid(src))
1647 continue;
1648
1649 list_del(&src->use_link);
1650 }
1651 }
1652
1653 static void
src_add_all_uses(nir_src * src,nir_instr * parent_instr,nir_if * parent_if)1654 src_add_all_uses(nir_src *src, nir_instr *parent_instr, nir_if *parent_if)
1655 {
1656 for (; src; src = src->is_ssa ? NULL : src->reg.indirect) {
1657 if (!src_is_valid(src))
1658 continue;
1659
1660 if (parent_instr) {
1661 src->parent_instr = parent_instr;
1662 if (src->is_ssa)
1663 list_addtail(&src->use_link, &src->ssa->uses);
1664 else
1665 list_addtail(&src->use_link, &src->reg.reg->uses);
1666 } else {
1667 assert(parent_if);
1668 src->parent_if = parent_if;
1669 if (src->is_ssa)
1670 list_addtail(&src->use_link, &src->ssa->if_uses);
1671 else
1672 list_addtail(&src->use_link, &src->reg.reg->if_uses);
1673 }
1674 }
1675 }
1676
1677 void
nir_instr_rewrite_src(nir_instr * instr,nir_src * src,nir_src new_src)1678 nir_instr_rewrite_src(nir_instr *instr, nir_src *src, nir_src new_src)
1679 {
1680 assert(!src_is_valid(src) || src->parent_instr == instr);
1681
1682 src_remove_all_uses(src);
1683 nir_src_copy(src, &new_src);
1684 src_add_all_uses(src, instr, NULL);
1685 }
1686
1687 void
nir_instr_move_src(nir_instr * dest_instr,nir_src * dest,nir_src * src)1688 nir_instr_move_src(nir_instr *dest_instr, nir_src *dest, nir_src *src)
1689 {
1690 assert(!src_is_valid(dest) || dest->parent_instr == dest_instr);
1691
1692 src_remove_all_uses(dest);
1693 src_free_indirects(dest);
1694 src_remove_all_uses(src);
1695 *dest = *src;
1696 *src = NIR_SRC_INIT;
1697 src_add_all_uses(dest, dest_instr, NULL);
1698 }
1699
1700 void
nir_if_rewrite_condition(nir_if * if_stmt,nir_src new_src)1701 nir_if_rewrite_condition(nir_if *if_stmt, nir_src new_src)
1702 {
1703 nir_src *src = &if_stmt->condition;
1704 assert(!src_is_valid(src) || src->parent_if == if_stmt);
1705
1706 src_remove_all_uses(src);
1707 nir_src_copy(src, &new_src);
1708 src_add_all_uses(src, NULL, if_stmt);
1709 }
1710
1711 void
nir_instr_rewrite_dest(nir_instr * instr,nir_dest * dest,nir_dest new_dest)1712 nir_instr_rewrite_dest(nir_instr *instr, nir_dest *dest, nir_dest new_dest)
1713 {
1714 if (dest->is_ssa) {
1715 /* We can only overwrite an SSA destination if it has no uses. */
1716 assert(nir_ssa_def_is_unused(&dest->ssa));
1717 } else {
1718 list_del(&dest->reg.def_link);
1719 if (dest->reg.indirect)
1720 src_remove_all_uses(dest->reg.indirect);
1721 }
1722
1723 /* We can't re-write with an SSA def */
1724 assert(!new_dest.is_ssa);
1725
1726 nir_dest_copy(dest, &new_dest);
1727
1728 dest->reg.parent_instr = instr;
1729 list_addtail(&dest->reg.def_link, &new_dest.reg.reg->defs);
1730
1731 if (dest->reg.indirect)
1732 src_add_all_uses(dest->reg.indirect, instr, NULL);
1733 }
1734
1735 /* note: does *not* take ownership of 'name' */
1736 void
nir_ssa_def_init(nir_instr * instr,nir_ssa_def * def,unsigned num_components,unsigned bit_size)1737 nir_ssa_def_init(nir_instr *instr, nir_ssa_def *def,
1738 unsigned num_components,
1739 unsigned bit_size)
1740 {
1741 def->parent_instr = instr;
1742 list_inithead(&def->uses);
1743 list_inithead(&def->if_uses);
1744 def->num_components = num_components;
1745 def->bit_size = bit_size;
1746 def->divergent = true; /* This is the safer default */
1747
1748 if (instr->block) {
1749 nir_function_impl *impl =
1750 nir_cf_node_get_function(&instr->block->cf_node);
1751
1752 def->index = impl->ssa_alloc++;
1753
1754 impl->valid_metadata &= ~nir_metadata_live_ssa_defs;
1755 } else {
1756 def->index = UINT_MAX;
1757 }
1758 }
1759
1760 /* note: does *not* take ownership of 'name' */
1761 void
nir_ssa_dest_init(nir_instr * instr,nir_dest * dest,unsigned num_components,unsigned bit_size,const char * name)1762 nir_ssa_dest_init(nir_instr *instr, nir_dest *dest,
1763 unsigned num_components, unsigned bit_size,
1764 const char *name)
1765 {
1766 dest->is_ssa = true;
1767 nir_ssa_def_init(instr, &dest->ssa, num_components, bit_size);
1768 }
1769
1770 void
nir_ssa_def_rewrite_uses(nir_ssa_def * def,nir_ssa_def * new_ssa)1771 nir_ssa_def_rewrite_uses(nir_ssa_def *def, nir_ssa_def *new_ssa)
1772 {
1773 assert(def != new_ssa);
1774 nir_foreach_use_safe(use_src, def)
1775 nir_instr_rewrite_src_ssa(use_src->parent_instr, use_src, new_ssa);
1776
1777 nir_foreach_if_use_safe(use_src, def)
1778 nir_if_rewrite_condition_ssa(use_src->parent_if, use_src, new_ssa);
1779 }
1780
1781 void
nir_ssa_def_rewrite_uses_src(nir_ssa_def * def,nir_src new_src)1782 nir_ssa_def_rewrite_uses_src(nir_ssa_def *def, nir_src new_src)
1783 {
1784 if (new_src.is_ssa) {
1785 nir_ssa_def_rewrite_uses(def, new_src.ssa);
1786 } else {
1787 nir_foreach_use_safe(use_src, def)
1788 nir_instr_rewrite_src(use_src->parent_instr, use_src, new_src);
1789
1790 nir_foreach_if_use_safe(use_src, def)
1791 nir_if_rewrite_condition(use_src->parent_if, new_src);
1792 }
1793 }
1794
1795 static bool
is_instr_between(nir_instr * start,nir_instr * end,nir_instr * between)1796 is_instr_between(nir_instr *start, nir_instr *end, nir_instr *between)
1797 {
1798 assert(start->block == end->block);
1799
1800 if (between->block != start->block)
1801 return false;
1802
1803 /* Search backwards looking for "between" */
1804 while (start != end) {
1805 if (between == end)
1806 return true;
1807
1808 end = nir_instr_prev(end);
1809 assert(end);
1810 }
1811
1812 return false;
1813 }
1814
1815 /* Replaces all uses of the given SSA def with the given source but only if
1816 * the use comes after the after_me instruction. This can be useful if you
1817 * are emitting code to fix up the result of some instruction: you can freely
1818 * use the result in that code and then call rewrite_uses_after and pass the
1819 * last fixup instruction as after_me and it will replace all of the uses you
1820 * want without touching the fixup code.
1821 *
1822 * This function assumes that after_me is in the same block as
1823 * def->parent_instr and that after_me comes after def->parent_instr.
1824 */
1825 void
nir_ssa_def_rewrite_uses_after(nir_ssa_def * def,nir_ssa_def * new_ssa,nir_instr * after_me)1826 nir_ssa_def_rewrite_uses_after(nir_ssa_def *def, nir_ssa_def *new_ssa,
1827 nir_instr *after_me)
1828 {
1829 if (def == new_ssa)
1830 return;
1831
1832 nir_foreach_use_safe(use_src, def) {
1833 assert(use_src->parent_instr != def->parent_instr);
1834 /* Since def already dominates all of its uses, the only way a use can
1835 * not be dominated by after_me is if it is between def and after_me in
1836 * the instruction list.
1837 */
1838 if (!is_instr_between(def->parent_instr, after_me, use_src->parent_instr))
1839 nir_instr_rewrite_src_ssa(use_src->parent_instr, use_src, new_ssa);
1840 }
1841
1842 nir_foreach_if_use_safe(use_src, def) {
1843 nir_if_rewrite_condition_ssa(use_src->parent_if,
1844 &use_src->parent_if->condition,
1845 new_ssa);
1846 }
1847 }
1848
1849 static nir_ssa_def *
get_store_value(nir_intrinsic_instr * intrin)1850 get_store_value(nir_intrinsic_instr *intrin)
1851 {
1852 assert(nir_intrinsic_has_write_mask(intrin));
1853 /* deref stores have the deref in src[0] and the store value in src[1] */
1854 if (intrin->intrinsic == nir_intrinsic_store_deref ||
1855 intrin->intrinsic == nir_intrinsic_store_deref_block_intel)
1856 return intrin->src[1].ssa;
1857
1858 /* all other stores have the store value in src[0] */
1859 return intrin->src[0].ssa;
1860 }
1861
1862 nir_component_mask_t
nir_src_components_read(const nir_src * src)1863 nir_src_components_read(const nir_src *src)
1864 {
1865 assert(src->is_ssa && src->parent_instr);
1866
1867 if (src->parent_instr->type == nir_instr_type_alu) {
1868 nir_alu_instr *alu = nir_instr_as_alu(src->parent_instr);
1869 nir_alu_src *alu_src = exec_node_data(nir_alu_src, src, src);
1870 int src_idx = alu_src - &alu->src[0];
1871 assert(src_idx >= 0 && src_idx < nir_op_infos[alu->op].num_inputs);
1872 return nir_alu_instr_src_read_mask(alu, src_idx);
1873 } else if (src->parent_instr->type == nir_instr_type_intrinsic) {
1874 nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(src->parent_instr);
1875 if (nir_intrinsic_has_write_mask(intrin) && src->ssa == get_store_value(intrin))
1876 return nir_intrinsic_write_mask(intrin);
1877 else
1878 return (1 << src->ssa->num_components) - 1;
1879 } else {
1880 return (1 << src->ssa->num_components) - 1;
1881 }
1882 }
1883
1884 nir_component_mask_t
nir_ssa_def_components_read(const nir_ssa_def * def)1885 nir_ssa_def_components_read(const nir_ssa_def *def)
1886 {
1887 nir_component_mask_t read_mask = 0;
1888
1889 if (!list_is_empty(&def->if_uses))
1890 read_mask |= 1;
1891
1892 nir_foreach_use(use, def) {
1893 read_mask |= nir_src_components_read(use);
1894 if (read_mask == (1 << def->num_components) - 1)
1895 return read_mask;
1896 }
1897
1898 return read_mask;
1899 }
1900
1901 nir_block *
nir_block_unstructured_next(nir_block * block)1902 nir_block_unstructured_next(nir_block *block)
1903 {
1904 if (block == NULL) {
1905 /* nir_foreach_block_unstructured_safe() will call this function on a
1906 * NULL block after the last iteration, but it won't use the result so
1907 * just return NULL here.
1908 */
1909 return NULL;
1910 }
1911
1912 nir_cf_node *cf_next = nir_cf_node_next(&block->cf_node);
1913 if (cf_next == NULL && block->cf_node.parent->type == nir_cf_node_function)
1914 return NULL;
1915
1916 if (cf_next && cf_next->type == nir_cf_node_block)
1917 return nir_cf_node_as_block(cf_next);
1918
1919 return nir_block_cf_tree_next(block);
1920 }
1921
1922 nir_block *
nir_unstructured_start_block(nir_function_impl * impl)1923 nir_unstructured_start_block(nir_function_impl *impl)
1924 {
1925 return nir_start_block(impl);
1926 }
1927
1928 nir_block *
nir_block_cf_tree_next(nir_block * block)1929 nir_block_cf_tree_next(nir_block *block)
1930 {
1931 if (block == NULL) {
1932 /* nir_foreach_block_safe() will call this function on a NULL block
1933 * after the last iteration, but it won't use the result so just return
1934 * NULL here.
1935 */
1936 return NULL;
1937 }
1938
1939 assert(nir_cf_node_get_function(&block->cf_node)->structured);
1940
1941 nir_cf_node *cf_next = nir_cf_node_next(&block->cf_node);
1942 if (cf_next)
1943 return nir_cf_node_cf_tree_first(cf_next);
1944
1945 nir_cf_node *parent = block->cf_node.parent;
1946
1947 switch (parent->type) {
1948 case nir_cf_node_if: {
1949 /* Are we at the end of the if? Go to the beginning of the else */
1950 nir_if *if_stmt = nir_cf_node_as_if(parent);
1951 if (block == nir_if_last_then_block(if_stmt))
1952 return nir_if_first_else_block(if_stmt);
1953
1954 assert(block == nir_if_last_else_block(if_stmt));
1955 }
1956 FALLTHROUGH;
1957
1958 case nir_cf_node_loop:
1959 return nir_cf_node_as_block(nir_cf_node_next(parent));
1960
1961 case nir_cf_node_function:
1962 return NULL;
1963
1964 default:
1965 unreachable("unknown cf node type");
1966 }
1967 }
1968
1969 nir_block *
nir_block_cf_tree_prev(nir_block * block)1970 nir_block_cf_tree_prev(nir_block *block)
1971 {
1972 if (block == NULL) {
1973 /* do this for consistency with nir_block_cf_tree_next() */
1974 return NULL;
1975 }
1976
1977 assert(nir_cf_node_get_function(&block->cf_node)->structured);
1978
1979 nir_cf_node *cf_prev = nir_cf_node_prev(&block->cf_node);
1980 if (cf_prev)
1981 return nir_cf_node_cf_tree_last(cf_prev);
1982
1983 nir_cf_node *parent = block->cf_node.parent;
1984
1985 switch (parent->type) {
1986 case nir_cf_node_if: {
1987 /* Are we at the beginning of the else? Go to the end of the if */
1988 nir_if *if_stmt = nir_cf_node_as_if(parent);
1989 if (block == nir_if_first_else_block(if_stmt))
1990 return nir_if_last_then_block(if_stmt);
1991
1992 assert(block == nir_if_first_then_block(if_stmt));
1993 }
1994 FALLTHROUGH;
1995
1996 case nir_cf_node_loop:
1997 return nir_cf_node_as_block(nir_cf_node_prev(parent));
1998
1999 case nir_cf_node_function:
2000 return NULL;
2001
2002 default:
2003 unreachable("unknown cf node type");
2004 }
2005 }
2006
nir_cf_node_cf_tree_first(nir_cf_node * node)2007 nir_block *nir_cf_node_cf_tree_first(nir_cf_node *node)
2008 {
2009 switch (node->type) {
2010 case nir_cf_node_function: {
2011 nir_function_impl *impl = nir_cf_node_as_function(node);
2012 return nir_start_block(impl);
2013 }
2014
2015 case nir_cf_node_if: {
2016 nir_if *if_stmt = nir_cf_node_as_if(node);
2017 return nir_if_first_then_block(if_stmt);
2018 }
2019
2020 case nir_cf_node_loop: {
2021 nir_loop *loop = nir_cf_node_as_loop(node);
2022 return nir_loop_first_block(loop);
2023 }
2024
2025 case nir_cf_node_block: {
2026 return nir_cf_node_as_block(node);
2027 }
2028
2029 default:
2030 unreachable("unknown node type");
2031 }
2032 }
2033
nir_cf_node_cf_tree_last(nir_cf_node * node)2034 nir_block *nir_cf_node_cf_tree_last(nir_cf_node *node)
2035 {
2036 switch (node->type) {
2037 case nir_cf_node_function: {
2038 nir_function_impl *impl = nir_cf_node_as_function(node);
2039 return nir_impl_last_block(impl);
2040 }
2041
2042 case nir_cf_node_if: {
2043 nir_if *if_stmt = nir_cf_node_as_if(node);
2044 return nir_if_last_else_block(if_stmt);
2045 }
2046
2047 case nir_cf_node_loop: {
2048 nir_loop *loop = nir_cf_node_as_loop(node);
2049 return nir_loop_last_block(loop);
2050 }
2051
2052 case nir_cf_node_block: {
2053 return nir_cf_node_as_block(node);
2054 }
2055
2056 default:
2057 unreachable("unknown node type");
2058 }
2059 }
2060
nir_cf_node_cf_tree_next(nir_cf_node * node)2061 nir_block *nir_cf_node_cf_tree_next(nir_cf_node *node)
2062 {
2063 if (node->type == nir_cf_node_block)
2064 return nir_block_cf_tree_next(nir_cf_node_as_block(node));
2065 else if (node->type == nir_cf_node_function)
2066 return NULL;
2067 else
2068 return nir_cf_node_as_block(nir_cf_node_next(node));
2069 }
2070
2071 nir_if *
nir_block_get_following_if(nir_block * block)2072 nir_block_get_following_if(nir_block *block)
2073 {
2074 if (exec_node_is_tail_sentinel(&block->cf_node.node))
2075 return NULL;
2076
2077 if (nir_cf_node_is_last(&block->cf_node))
2078 return NULL;
2079
2080 nir_cf_node *next_node = nir_cf_node_next(&block->cf_node);
2081
2082 if (next_node->type != nir_cf_node_if)
2083 return NULL;
2084
2085 return nir_cf_node_as_if(next_node);
2086 }
2087
2088 nir_loop *
nir_block_get_following_loop(nir_block * block)2089 nir_block_get_following_loop(nir_block *block)
2090 {
2091 if (exec_node_is_tail_sentinel(&block->cf_node.node))
2092 return NULL;
2093
2094 if (nir_cf_node_is_last(&block->cf_node))
2095 return NULL;
2096
2097 nir_cf_node *next_node = nir_cf_node_next(&block->cf_node);
2098
2099 if (next_node->type != nir_cf_node_loop)
2100 return NULL;
2101
2102 return nir_cf_node_as_loop(next_node);
2103 }
2104
2105 static int
compare_block_index(const void * p1,const void * p2)2106 compare_block_index(const void *p1, const void *p2)
2107 {
2108 const nir_block *block1 = *((const nir_block **) p1);
2109 const nir_block *block2 = *((const nir_block **) p2);
2110
2111 return (int) block1->index - (int) block2->index;
2112 }
2113
2114 nir_block **
nir_block_get_predecessors_sorted(const nir_block * block,void * mem_ctx)2115 nir_block_get_predecessors_sorted(const nir_block *block, void *mem_ctx)
2116 {
2117 nir_block **preds =
2118 ralloc_array(mem_ctx, nir_block *, block->predecessors->entries);
2119
2120 unsigned i = 0;
2121 set_foreach(block->predecessors, entry)
2122 preds[i++] = (nir_block *) entry->key;
2123 assert(i == block->predecessors->entries);
2124
2125 qsort(preds, block->predecessors->entries, sizeof(nir_block *),
2126 compare_block_index);
2127
2128 return preds;
2129 }
2130
2131 void
nir_index_blocks(nir_function_impl * impl)2132 nir_index_blocks(nir_function_impl *impl)
2133 {
2134 unsigned index = 0;
2135
2136 if (impl->valid_metadata & nir_metadata_block_index)
2137 return;
2138
2139 nir_foreach_block_unstructured(block, impl) {
2140 block->index = index++;
2141 }
2142
2143 /* The end_block isn't really part of the program, which is why its index
2144 * is >= num_blocks.
2145 */
2146 impl->num_blocks = impl->end_block->index = index;
2147 }
2148
2149 static bool
index_ssa_def_cb(nir_ssa_def * def,void * state)2150 index_ssa_def_cb(nir_ssa_def *def, void *state)
2151 {
2152 unsigned *index = (unsigned *) state;
2153 def->index = (*index)++;
2154
2155 return true;
2156 }
2157
2158 /**
2159 * The indices are applied top-to-bottom which has the very nice property
2160 * that, if A dominates B, then A->index <= B->index.
2161 */
2162 void
nir_index_ssa_defs(nir_function_impl * impl)2163 nir_index_ssa_defs(nir_function_impl *impl)
2164 {
2165 unsigned index = 0;
2166
2167 impl->valid_metadata &= ~nir_metadata_live_ssa_defs;
2168
2169 nir_foreach_block_unstructured(block, impl) {
2170 nir_foreach_instr(instr, block)
2171 nir_foreach_ssa_def(instr, index_ssa_def_cb, &index);
2172 }
2173
2174 impl->ssa_alloc = index;
2175 }
2176
2177 /**
2178 * The indices are applied top-to-bottom which has the very nice property
2179 * that, if A dominates B, then A->index <= B->index.
2180 */
2181 unsigned
nir_index_instrs(nir_function_impl * impl)2182 nir_index_instrs(nir_function_impl *impl)
2183 {
2184 unsigned index = 0;
2185
2186 nir_foreach_block(block, impl) {
2187 block->start_ip = index++;
2188
2189 nir_foreach_instr(instr, block)
2190 instr->index = index++;
2191
2192 block->end_ip = index++;
2193 }
2194
2195 return index;
2196 }
2197
2198 unsigned
nir_shader_index_vars(nir_shader * shader,nir_variable_mode modes)2199 nir_shader_index_vars(nir_shader *shader, nir_variable_mode modes)
2200 {
2201 unsigned count = 0;
2202 nir_foreach_variable_with_modes(var, shader, modes)
2203 var->index = count++;
2204 return count;
2205 }
2206
2207 unsigned
nir_function_impl_index_vars(nir_function_impl * impl)2208 nir_function_impl_index_vars(nir_function_impl *impl)
2209 {
2210 unsigned count = 0;
2211 nir_foreach_function_temp_variable(var, impl)
2212 var->index = count++;
2213 return count;
2214 }
2215
2216 static nir_instr *
cursor_next_instr(nir_cursor cursor)2217 cursor_next_instr(nir_cursor cursor)
2218 {
2219 switch (cursor.option) {
2220 case nir_cursor_before_block:
2221 for (nir_block *block = cursor.block; block;
2222 block = nir_block_cf_tree_next(block)) {
2223 nir_instr *instr = nir_block_first_instr(block);
2224 if (instr)
2225 return instr;
2226 }
2227 return NULL;
2228
2229 case nir_cursor_after_block:
2230 cursor.block = nir_block_cf_tree_next(cursor.block);
2231 if (cursor.block == NULL)
2232 return NULL;
2233
2234 cursor.option = nir_cursor_before_block;
2235 return cursor_next_instr(cursor);
2236
2237 case nir_cursor_before_instr:
2238 return cursor.instr;
2239
2240 case nir_cursor_after_instr:
2241 if (nir_instr_next(cursor.instr))
2242 return nir_instr_next(cursor.instr);
2243
2244 cursor.option = nir_cursor_after_block;
2245 cursor.block = cursor.instr->block;
2246 return cursor_next_instr(cursor);
2247 }
2248
2249 unreachable("Inavlid cursor option");
2250 }
2251
2252 ASSERTED static bool
dest_is_ssa(nir_dest * dest,void * _state)2253 dest_is_ssa(nir_dest *dest, void *_state)
2254 {
2255 (void) _state;
2256 return dest->is_ssa;
2257 }
2258
2259 bool
nir_function_impl_lower_instructions(nir_function_impl * impl,nir_instr_filter_cb filter,nir_lower_instr_cb lower,void * cb_data)2260 nir_function_impl_lower_instructions(nir_function_impl *impl,
2261 nir_instr_filter_cb filter,
2262 nir_lower_instr_cb lower,
2263 void *cb_data)
2264 {
2265 nir_builder b;
2266 nir_builder_init(&b, impl);
2267
2268 nir_metadata preserved = nir_metadata_block_index |
2269 nir_metadata_dominance;
2270
2271 bool progress = false;
2272 nir_cursor iter = nir_before_cf_list(&impl->body);
2273 nir_instr *instr;
2274 while ((instr = cursor_next_instr(iter)) != NULL) {
2275 if (filter && !filter(instr, cb_data)) {
2276 iter = nir_after_instr(instr);
2277 continue;
2278 }
2279
2280 assert(nir_foreach_dest(instr, dest_is_ssa, NULL));
2281 nir_ssa_def *old_def = nir_instr_ssa_def(instr);
2282 struct list_head old_uses, old_if_uses;
2283 if (old_def != NULL) {
2284 /* We're about to ask the callback to generate a replacement for instr.
2285 * Save off the uses from instr's SSA def so we know what uses to
2286 * rewrite later. If we use nir_ssa_def_rewrite_uses, it fails in the
2287 * case where the generated replacement code uses the result of instr
2288 * itself. If we use nir_ssa_def_rewrite_uses_after (which is the
2289 * normal solution to this problem), it doesn't work well if control-
2290 * flow is inserted as part of the replacement, doesn't handle cases
2291 * where the replacement is something consumed by instr, and suffers
2292 * from performance issues. This is the only way to 100% guarantee
2293 * that we rewrite the correct set efficiently.
2294 */
2295
2296 list_replace(&old_def->uses, &old_uses);
2297 list_inithead(&old_def->uses);
2298 list_replace(&old_def->if_uses, &old_if_uses);
2299 list_inithead(&old_def->if_uses);
2300 }
2301
2302 b.cursor = nir_after_instr(instr);
2303 nir_ssa_def *new_def = lower(&b, instr, cb_data);
2304 if (new_def && new_def != NIR_LOWER_INSTR_PROGRESS &&
2305 new_def != NIR_LOWER_INSTR_PROGRESS_REPLACE) {
2306 assert(old_def != NULL);
2307 if (new_def->parent_instr->block != instr->block)
2308 preserved = nir_metadata_none;
2309
2310 nir_src new_src = nir_src_for_ssa(new_def);
2311 list_for_each_entry_safe(nir_src, use_src, &old_uses, use_link)
2312 nir_instr_rewrite_src(use_src->parent_instr, use_src, new_src);
2313
2314 list_for_each_entry_safe(nir_src, use_src, &old_if_uses, use_link)
2315 nir_if_rewrite_condition(use_src->parent_if, new_src);
2316
2317 if (nir_ssa_def_is_unused(old_def)) {
2318 iter = nir_instr_free_and_dce(instr);
2319 } else {
2320 iter = nir_after_instr(instr);
2321 }
2322 progress = true;
2323 } else {
2324 /* We didn't end up lowering after all. Put the uses back */
2325 if (old_def) {
2326 list_replace(&old_uses, &old_def->uses);
2327 list_replace(&old_if_uses, &old_def->if_uses);
2328 }
2329 if (new_def == NIR_LOWER_INSTR_PROGRESS_REPLACE) {
2330 /* Only instructions without a return value can be removed like this */
2331 assert(!old_def);
2332 iter = nir_instr_free_and_dce(instr);
2333 progress = true;
2334 } else
2335 iter = nir_after_instr(instr);
2336
2337 if (new_def == NIR_LOWER_INSTR_PROGRESS)
2338 progress = true;
2339 }
2340 }
2341
2342 if (progress) {
2343 nir_metadata_preserve(impl, preserved);
2344 } else {
2345 nir_metadata_preserve(impl, nir_metadata_all);
2346 }
2347
2348 return progress;
2349 }
2350
2351 bool
nir_shader_lower_instructions(nir_shader * shader,nir_instr_filter_cb filter,nir_lower_instr_cb lower,void * cb_data)2352 nir_shader_lower_instructions(nir_shader *shader,
2353 nir_instr_filter_cb filter,
2354 nir_lower_instr_cb lower,
2355 void *cb_data)
2356 {
2357 bool progress = false;
2358
2359 nir_foreach_function(function, shader) {
2360 if (function->impl &&
2361 nir_function_impl_lower_instructions(function->impl,
2362 filter, lower, cb_data))
2363 progress = true;
2364 }
2365
2366 return progress;
2367 }
2368
2369 /**
2370 * Returns true if the shader supports quad-based implicit derivatives on
2371 * texture sampling.
2372 */
nir_shader_supports_implicit_lod(nir_shader * shader)2373 bool nir_shader_supports_implicit_lod(nir_shader *shader)
2374 {
2375 return (shader->info.stage == MESA_SHADER_FRAGMENT ||
2376 (shader->info.stage == MESA_SHADER_COMPUTE &&
2377 shader->info.cs.derivative_group != DERIVATIVE_GROUP_NONE));
2378 }
2379
2380 nir_intrinsic_op
nir_intrinsic_from_system_value(gl_system_value val)2381 nir_intrinsic_from_system_value(gl_system_value val)
2382 {
2383 switch (val) {
2384 case SYSTEM_VALUE_VERTEX_ID:
2385 return nir_intrinsic_load_vertex_id;
2386 case SYSTEM_VALUE_INSTANCE_ID:
2387 return nir_intrinsic_load_instance_id;
2388 case SYSTEM_VALUE_DRAW_ID:
2389 return nir_intrinsic_load_draw_id;
2390 case SYSTEM_VALUE_BASE_INSTANCE:
2391 return nir_intrinsic_load_base_instance;
2392 case SYSTEM_VALUE_VERTEX_ID_ZERO_BASE:
2393 return nir_intrinsic_load_vertex_id_zero_base;
2394 case SYSTEM_VALUE_IS_INDEXED_DRAW:
2395 return nir_intrinsic_load_is_indexed_draw;
2396 case SYSTEM_VALUE_FIRST_VERTEX:
2397 return nir_intrinsic_load_first_vertex;
2398 case SYSTEM_VALUE_BASE_VERTEX:
2399 return nir_intrinsic_load_base_vertex;
2400 case SYSTEM_VALUE_INVOCATION_ID:
2401 return nir_intrinsic_load_invocation_id;
2402 case SYSTEM_VALUE_FRAG_COORD:
2403 return nir_intrinsic_load_frag_coord;
2404 case SYSTEM_VALUE_POINT_COORD:
2405 return nir_intrinsic_load_point_coord;
2406 case SYSTEM_VALUE_LINE_COORD:
2407 return nir_intrinsic_load_line_coord;
2408 case SYSTEM_VALUE_FRONT_FACE:
2409 return nir_intrinsic_load_front_face;
2410 case SYSTEM_VALUE_SAMPLE_ID:
2411 return nir_intrinsic_load_sample_id;
2412 case SYSTEM_VALUE_SAMPLE_POS:
2413 return nir_intrinsic_load_sample_pos;
2414 case SYSTEM_VALUE_SAMPLE_POS_OR_CENTER:
2415 return nir_intrinsic_load_sample_pos_or_center;
2416 case SYSTEM_VALUE_SAMPLE_MASK_IN:
2417 return nir_intrinsic_load_sample_mask_in;
2418 case SYSTEM_VALUE_LOCAL_INVOCATION_ID:
2419 return nir_intrinsic_load_local_invocation_id;
2420 case SYSTEM_VALUE_LOCAL_INVOCATION_INDEX:
2421 return nir_intrinsic_load_local_invocation_index;
2422 case SYSTEM_VALUE_WORKGROUP_ID:
2423 return nir_intrinsic_load_workgroup_id;
2424 case SYSTEM_VALUE_WORKGROUP_INDEX:
2425 return nir_intrinsic_load_workgroup_index;
2426 case SYSTEM_VALUE_NUM_WORKGROUPS:
2427 return nir_intrinsic_load_num_workgroups;
2428 case SYSTEM_VALUE_PRIMITIVE_ID:
2429 return nir_intrinsic_load_primitive_id;
2430 case SYSTEM_VALUE_TESS_COORD:
2431 return nir_intrinsic_load_tess_coord;
2432 case SYSTEM_VALUE_TESS_LEVEL_OUTER:
2433 return nir_intrinsic_load_tess_level_outer;
2434 case SYSTEM_VALUE_TESS_LEVEL_INNER:
2435 return nir_intrinsic_load_tess_level_inner;
2436 case SYSTEM_VALUE_TESS_LEVEL_OUTER_DEFAULT:
2437 return nir_intrinsic_load_tess_level_outer_default;
2438 case SYSTEM_VALUE_TESS_LEVEL_INNER_DEFAULT:
2439 return nir_intrinsic_load_tess_level_inner_default;
2440 case SYSTEM_VALUE_VERTICES_IN:
2441 return nir_intrinsic_load_patch_vertices_in;
2442 case SYSTEM_VALUE_HELPER_INVOCATION:
2443 return nir_intrinsic_load_helper_invocation;
2444 case SYSTEM_VALUE_COLOR0:
2445 return nir_intrinsic_load_color0;
2446 case SYSTEM_VALUE_COLOR1:
2447 return nir_intrinsic_load_color1;
2448 case SYSTEM_VALUE_VIEW_INDEX:
2449 return nir_intrinsic_load_view_index;
2450 case SYSTEM_VALUE_SUBGROUP_SIZE:
2451 return nir_intrinsic_load_subgroup_size;
2452 case SYSTEM_VALUE_SUBGROUP_INVOCATION:
2453 return nir_intrinsic_load_subgroup_invocation;
2454 case SYSTEM_VALUE_SUBGROUP_EQ_MASK:
2455 return nir_intrinsic_load_subgroup_eq_mask;
2456 case SYSTEM_VALUE_SUBGROUP_GE_MASK:
2457 return nir_intrinsic_load_subgroup_ge_mask;
2458 case SYSTEM_VALUE_SUBGROUP_GT_MASK:
2459 return nir_intrinsic_load_subgroup_gt_mask;
2460 case SYSTEM_VALUE_SUBGROUP_LE_MASK:
2461 return nir_intrinsic_load_subgroup_le_mask;
2462 case SYSTEM_VALUE_SUBGROUP_LT_MASK:
2463 return nir_intrinsic_load_subgroup_lt_mask;
2464 case SYSTEM_VALUE_NUM_SUBGROUPS:
2465 return nir_intrinsic_load_num_subgroups;
2466 case SYSTEM_VALUE_SUBGROUP_ID:
2467 return nir_intrinsic_load_subgroup_id;
2468 case SYSTEM_VALUE_WORKGROUP_SIZE:
2469 return nir_intrinsic_load_workgroup_size;
2470 case SYSTEM_VALUE_GLOBAL_INVOCATION_ID:
2471 return nir_intrinsic_load_global_invocation_id;
2472 case SYSTEM_VALUE_BASE_GLOBAL_INVOCATION_ID:
2473 return nir_intrinsic_load_base_global_invocation_id;
2474 case SYSTEM_VALUE_GLOBAL_INVOCATION_INDEX:
2475 return nir_intrinsic_load_global_invocation_index;
2476 case SYSTEM_VALUE_WORK_DIM:
2477 return nir_intrinsic_load_work_dim;
2478 case SYSTEM_VALUE_USER_DATA_AMD:
2479 return nir_intrinsic_load_user_data_amd;
2480 case SYSTEM_VALUE_RAY_LAUNCH_ID:
2481 return nir_intrinsic_load_ray_launch_id;
2482 case SYSTEM_VALUE_RAY_LAUNCH_SIZE:
2483 return nir_intrinsic_load_ray_launch_size;
2484 case SYSTEM_VALUE_RAY_LAUNCH_SIZE_ADDR_AMD:
2485 return nir_intrinsic_load_ray_launch_size_addr_amd;
2486 case SYSTEM_VALUE_RAY_WORLD_ORIGIN:
2487 return nir_intrinsic_load_ray_world_origin;
2488 case SYSTEM_VALUE_RAY_WORLD_DIRECTION:
2489 return nir_intrinsic_load_ray_world_direction;
2490 case SYSTEM_VALUE_RAY_OBJECT_ORIGIN:
2491 return nir_intrinsic_load_ray_object_origin;
2492 case SYSTEM_VALUE_RAY_OBJECT_DIRECTION:
2493 return nir_intrinsic_load_ray_object_direction;
2494 case SYSTEM_VALUE_RAY_T_MIN:
2495 return nir_intrinsic_load_ray_t_min;
2496 case SYSTEM_VALUE_RAY_T_MAX:
2497 return nir_intrinsic_load_ray_t_max;
2498 case SYSTEM_VALUE_RAY_OBJECT_TO_WORLD:
2499 return nir_intrinsic_load_ray_object_to_world;
2500 case SYSTEM_VALUE_RAY_WORLD_TO_OBJECT:
2501 return nir_intrinsic_load_ray_world_to_object;
2502 case SYSTEM_VALUE_RAY_HIT_KIND:
2503 return nir_intrinsic_load_ray_hit_kind;
2504 case SYSTEM_VALUE_RAY_FLAGS:
2505 return nir_intrinsic_load_ray_flags;
2506 case SYSTEM_VALUE_RAY_GEOMETRY_INDEX:
2507 return nir_intrinsic_load_ray_geometry_index;
2508 case SYSTEM_VALUE_RAY_INSTANCE_CUSTOM_INDEX:
2509 return nir_intrinsic_load_ray_instance_custom_index;
2510 case SYSTEM_VALUE_CULL_MASK:
2511 return nir_intrinsic_load_cull_mask;
2512 case SYSTEM_VALUE_MESH_VIEW_COUNT:
2513 return nir_intrinsic_load_mesh_view_count;
2514 case SYSTEM_VALUE_FRAG_SHADING_RATE:
2515 return nir_intrinsic_load_frag_shading_rate;
2516 default:
2517 unreachable("system value does not directly correspond to intrinsic");
2518 }
2519 }
2520
2521 gl_system_value
nir_system_value_from_intrinsic(nir_intrinsic_op intrin)2522 nir_system_value_from_intrinsic(nir_intrinsic_op intrin)
2523 {
2524 switch (intrin) {
2525 case nir_intrinsic_load_vertex_id:
2526 return SYSTEM_VALUE_VERTEX_ID;
2527 case nir_intrinsic_load_instance_id:
2528 return SYSTEM_VALUE_INSTANCE_ID;
2529 case nir_intrinsic_load_draw_id:
2530 return SYSTEM_VALUE_DRAW_ID;
2531 case nir_intrinsic_load_base_instance:
2532 return SYSTEM_VALUE_BASE_INSTANCE;
2533 case nir_intrinsic_load_vertex_id_zero_base:
2534 return SYSTEM_VALUE_VERTEX_ID_ZERO_BASE;
2535 case nir_intrinsic_load_first_vertex:
2536 return SYSTEM_VALUE_FIRST_VERTEX;
2537 case nir_intrinsic_load_is_indexed_draw:
2538 return SYSTEM_VALUE_IS_INDEXED_DRAW;
2539 case nir_intrinsic_load_base_vertex:
2540 return SYSTEM_VALUE_BASE_VERTEX;
2541 case nir_intrinsic_load_invocation_id:
2542 return SYSTEM_VALUE_INVOCATION_ID;
2543 case nir_intrinsic_load_frag_coord:
2544 return SYSTEM_VALUE_FRAG_COORD;
2545 case nir_intrinsic_load_point_coord:
2546 return SYSTEM_VALUE_POINT_COORD;
2547 case nir_intrinsic_load_line_coord:
2548 return SYSTEM_VALUE_LINE_COORD;
2549 case nir_intrinsic_load_front_face:
2550 return SYSTEM_VALUE_FRONT_FACE;
2551 case nir_intrinsic_load_sample_id:
2552 return SYSTEM_VALUE_SAMPLE_ID;
2553 case nir_intrinsic_load_sample_pos:
2554 return SYSTEM_VALUE_SAMPLE_POS;
2555 case nir_intrinsic_load_sample_pos_or_center:
2556 return SYSTEM_VALUE_SAMPLE_POS_OR_CENTER;
2557 case nir_intrinsic_load_sample_mask_in:
2558 return SYSTEM_VALUE_SAMPLE_MASK_IN;
2559 case nir_intrinsic_load_local_invocation_id:
2560 return SYSTEM_VALUE_LOCAL_INVOCATION_ID;
2561 case nir_intrinsic_load_local_invocation_index:
2562 return SYSTEM_VALUE_LOCAL_INVOCATION_INDEX;
2563 case nir_intrinsic_load_num_workgroups:
2564 return SYSTEM_VALUE_NUM_WORKGROUPS;
2565 case nir_intrinsic_load_workgroup_id:
2566 return SYSTEM_VALUE_WORKGROUP_ID;
2567 case nir_intrinsic_load_workgroup_index:
2568 return SYSTEM_VALUE_WORKGROUP_INDEX;
2569 case nir_intrinsic_load_primitive_id:
2570 return SYSTEM_VALUE_PRIMITIVE_ID;
2571 case nir_intrinsic_load_tess_coord:
2572 return SYSTEM_VALUE_TESS_COORD;
2573 case nir_intrinsic_load_tess_level_outer:
2574 return SYSTEM_VALUE_TESS_LEVEL_OUTER;
2575 case nir_intrinsic_load_tess_level_inner:
2576 return SYSTEM_VALUE_TESS_LEVEL_INNER;
2577 case nir_intrinsic_load_tess_level_outer_default:
2578 return SYSTEM_VALUE_TESS_LEVEL_OUTER_DEFAULT;
2579 case nir_intrinsic_load_tess_level_inner_default:
2580 return SYSTEM_VALUE_TESS_LEVEL_INNER_DEFAULT;
2581 case nir_intrinsic_load_patch_vertices_in:
2582 return SYSTEM_VALUE_VERTICES_IN;
2583 case nir_intrinsic_load_helper_invocation:
2584 return SYSTEM_VALUE_HELPER_INVOCATION;
2585 case nir_intrinsic_load_color0:
2586 return SYSTEM_VALUE_COLOR0;
2587 case nir_intrinsic_load_color1:
2588 return SYSTEM_VALUE_COLOR1;
2589 case nir_intrinsic_load_view_index:
2590 return SYSTEM_VALUE_VIEW_INDEX;
2591 case nir_intrinsic_load_subgroup_size:
2592 return SYSTEM_VALUE_SUBGROUP_SIZE;
2593 case nir_intrinsic_load_subgroup_invocation:
2594 return SYSTEM_VALUE_SUBGROUP_INVOCATION;
2595 case nir_intrinsic_load_subgroup_eq_mask:
2596 return SYSTEM_VALUE_SUBGROUP_EQ_MASK;
2597 case nir_intrinsic_load_subgroup_ge_mask:
2598 return SYSTEM_VALUE_SUBGROUP_GE_MASK;
2599 case nir_intrinsic_load_subgroup_gt_mask:
2600 return SYSTEM_VALUE_SUBGROUP_GT_MASK;
2601 case nir_intrinsic_load_subgroup_le_mask:
2602 return SYSTEM_VALUE_SUBGROUP_LE_MASK;
2603 case nir_intrinsic_load_subgroup_lt_mask:
2604 return SYSTEM_VALUE_SUBGROUP_LT_MASK;
2605 case nir_intrinsic_load_num_subgroups:
2606 return SYSTEM_VALUE_NUM_SUBGROUPS;
2607 case nir_intrinsic_load_subgroup_id:
2608 return SYSTEM_VALUE_SUBGROUP_ID;
2609 case nir_intrinsic_load_workgroup_size:
2610 return SYSTEM_VALUE_WORKGROUP_SIZE;
2611 case nir_intrinsic_load_global_invocation_id:
2612 return SYSTEM_VALUE_GLOBAL_INVOCATION_ID;
2613 case nir_intrinsic_load_base_global_invocation_id:
2614 return SYSTEM_VALUE_BASE_GLOBAL_INVOCATION_ID;
2615 case nir_intrinsic_load_global_invocation_index:
2616 return SYSTEM_VALUE_GLOBAL_INVOCATION_INDEX;
2617 case nir_intrinsic_load_work_dim:
2618 return SYSTEM_VALUE_WORK_DIM;
2619 case nir_intrinsic_load_user_data_amd:
2620 return SYSTEM_VALUE_USER_DATA_AMD;
2621 case nir_intrinsic_load_barycentric_model:
2622 return SYSTEM_VALUE_BARYCENTRIC_PULL_MODEL;
2623 case nir_intrinsic_load_gs_header_ir3:
2624 return SYSTEM_VALUE_GS_HEADER_IR3;
2625 case nir_intrinsic_load_tcs_header_ir3:
2626 return SYSTEM_VALUE_TCS_HEADER_IR3;
2627 case nir_intrinsic_load_ray_launch_id:
2628 return SYSTEM_VALUE_RAY_LAUNCH_ID;
2629 case nir_intrinsic_load_ray_launch_size:
2630 return SYSTEM_VALUE_RAY_LAUNCH_SIZE;
2631 case nir_intrinsic_load_ray_launch_size_addr_amd:
2632 return SYSTEM_VALUE_RAY_LAUNCH_SIZE_ADDR_AMD;
2633 case nir_intrinsic_load_ray_world_origin:
2634 return SYSTEM_VALUE_RAY_WORLD_ORIGIN;
2635 case nir_intrinsic_load_ray_world_direction:
2636 return SYSTEM_VALUE_RAY_WORLD_DIRECTION;
2637 case nir_intrinsic_load_ray_object_origin:
2638 return SYSTEM_VALUE_RAY_OBJECT_ORIGIN;
2639 case nir_intrinsic_load_ray_object_direction:
2640 return SYSTEM_VALUE_RAY_OBJECT_DIRECTION;
2641 case nir_intrinsic_load_ray_t_min:
2642 return SYSTEM_VALUE_RAY_T_MIN;
2643 case nir_intrinsic_load_ray_t_max:
2644 return SYSTEM_VALUE_RAY_T_MAX;
2645 case nir_intrinsic_load_ray_object_to_world:
2646 return SYSTEM_VALUE_RAY_OBJECT_TO_WORLD;
2647 case nir_intrinsic_load_ray_world_to_object:
2648 return SYSTEM_VALUE_RAY_WORLD_TO_OBJECT;
2649 case nir_intrinsic_load_ray_hit_kind:
2650 return SYSTEM_VALUE_RAY_HIT_KIND;
2651 case nir_intrinsic_load_ray_flags:
2652 return SYSTEM_VALUE_RAY_FLAGS;
2653 case nir_intrinsic_load_ray_geometry_index:
2654 return SYSTEM_VALUE_RAY_GEOMETRY_INDEX;
2655 case nir_intrinsic_load_ray_instance_custom_index:
2656 return SYSTEM_VALUE_RAY_INSTANCE_CUSTOM_INDEX;
2657 case nir_intrinsic_load_cull_mask:
2658 return SYSTEM_VALUE_CULL_MASK;
2659 case nir_intrinsic_load_frag_shading_rate:
2660 return SYSTEM_VALUE_FRAG_SHADING_RATE;
2661 case nir_intrinsic_load_mesh_view_count:
2662 return SYSTEM_VALUE_MESH_VIEW_COUNT;
2663 default:
2664 unreachable("intrinsic doesn't produce a system value");
2665 }
2666 }
2667
2668 /* OpenGL utility method that remaps the location attributes if they are
2669 * doubles. Not needed for vulkan due the differences on the input location
2670 * count for doubles on vulkan vs OpenGL
2671 *
2672 * The bitfield returned in dual_slot is one bit for each double input slot in
2673 * the original OpenGL single-slot input numbering. The mapping from old
2674 * locations to new locations is as follows:
2675 *
2676 * new_loc = loc + util_bitcount(dual_slot & BITFIELD64_MASK(loc))
2677 */
2678 void
nir_remap_dual_slot_attributes(nir_shader * shader,uint64_t * dual_slot)2679 nir_remap_dual_slot_attributes(nir_shader *shader, uint64_t *dual_slot)
2680 {
2681 assert(shader->info.stage == MESA_SHADER_VERTEX);
2682
2683 *dual_slot = 0;
2684 nir_foreach_shader_in_variable(var, shader) {
2685 if (glsl_type_is_dual_slot(glsl_without_array(var->type))) {
2686 unsigned slots = glsl_count_attribute_slots(var->type, true);
2687 *dual_slot |= BITFIELD64_MASK(slots) << var->data.location;
2688 }
2689 }
2690
2691 nir_foreach_shader_in_variable(var, shader) {
2692 var->data.location +=
2693 util_bitcount64(*dual_slot & BITFIELD64_MASK(var->data.location));
2694 }
2695 }
2696
2697 /* Returns an attribute mask that has been re-compacted using the given
2698 * dual_slot mask.
2699 */
2700 uint64_t
nir_get_single_slot_attribs_mask(uint64_t attribs,uint64_t dual_slot)2701 nir_get_single_slot_attribs_mask(uint64_t attribs, uint64_t dual_slot)
2702 {
2703 while (dual_slot) {
2704 unsigned loc = u_bit_scan64(&dual_slot);
2705 /* mask of all bits up to and including loc */
2706 uint64_t mask = BITFIELD64_MASK(loc + 1);
2707 attribs = (attribs & mask) | ((attribs & ~mask) >> 1);
2708 }
2709 return attribs;
2710 }
2711
2712 void
nir_rewrite_image_intrinsic(nir_intrinsic_instr * intrin,nir_ssa_def * src,bool bindless)2713 nir_rewrite_image_intrinsic(nir_intrinsic_instr *intrin, nir_ssa_def *src,
2714 bool bindless)
2715 {
2716 enum gl_access_qualifier access = nir_intrinsic_access(intrin);
2717
2718 /* Image intrinsics only have one of these */
2719 assert(!nir_intrinsic_has_src_type(intrin) ||
2720 !nir_intrinsic_has_dest_type(intrin));
2721
2722 nir_alu_type data_type = nir_type_invalid;
2723 if (nir_intrinsic_has_src_type(intrin))
2724 data_type = nir_intrinsic_src_type(intrin);
2725 if (nir_intrinsic_has_dest_type(intrin))
2726 data_type = nir_intrinsic_dest_type(intrin);
2727
2728 switch (intrin->intrinsic) {
2729 #define CASE(op) \
2730 case nir_intrinsic_image_deref_##op: \
2731 intrin->intrinsic = bindless ? nir_intrinsic_bindless_image_##op \
2732 : nir_intrinsic_image_##op; \
2733 break;
2734 CASE(load)
2735 CASE(sparse_load)
2736 CASE(store)
2737 CASE(atomic_add)
2738 CASE(atomic_imin)
2739 CASE(atomic_umin)
2740 CASE(atomic_imax)
2741 CASE(atomic_umax)
2742 CASE(atomic_and)
2743 CASE(atomic_or)
2744 CASE(atomic_xor)
2745 CASE(atomic_exchange)
2746 CASE(atomic_comp_swap)
2747 CASE(atomic_fadd)
2748 CASE(atomic_fmin)
2749 CASE(atomic_fmax)
2750 CASE(atomic_inc_wrap)
2751 CASE(atomic_dec_wrap)
2752 CASE(size)
2753 CASE(samples)
2754 CASE(load_raw_intel)
2755 CASE(store_raw_intel)
2756 #undef CASE
2757 default:
2758 unreachable("Unhanded image intrinsic");
2759 }
2760
2761 nir_deref_instr *deref = nir_src_as_deref(intrin->src[0]);
2762 nir_variable *var = nir_deref_instr_get_variable(deref);
2763
2764 /* Only update the format if the intrinsic doesn't have one set */
2765 if (nir_intrinsic_format(intrin) == PIPE_FORMAT_NONE)
2766 nir_intrinsic_set_format(intrin, var->data.image.format);
2767
2768 nir_intrinsic_set_access(intrin, access | var->data.access);
2769 if (nir_intrinsic_has_src_type(intrin))
2770 nir_intrinsic_set_src_type(intrin, data_type);
2771 if (nir_intrinsic_has_dest_type(intrin))
2772 nir_intrinsic_set_dest_type(intrin, data_type);
2773
2774 nir_instr_rewrite_src(&intrin->instr, &intrin->src[0],
2775 nir_src_for_ssa(src));
2776 }
2777
2778 unsigned
nir_image_intrinsic_coord_components(const nir_intrinsic_instr * instr)2779 nir_image_intrinsic_coord_components(const nir_intrinsic_instr *instr)
2780 {
2781 enum glsl_sampler_dim dim = nir_intrinsic_image_dim(instr);
2782 int coords = glsl_get_sampler_dim_coordinate_components(dim);
2783 if (dim == GLSL_SAMPLER_DIM_CUBE)
2784 return coords;
2785 else
2786 return coords + nir_intrinsic_image_array(instr);
2787 }
2788
2789 nir_src *
nir_get_shader_call_payload_src(nir_intrinsic_instr * call)2790 nir_get_shader_call_payload_src(nir_intrinsic_instr *call)
2791 {
2792 switch (call->intrinsic) {
2793 case nir_intrinsic_trace_ray:
2794 case nir_intrinsic_rt_trace_ray:
2795 return &call->src[10];
2796 case nir_intrinsic_execute_callable:
2797 case nir_intrinsic_rt_execute_callable:
2798 return &call->src[1];
2799 default:
2800 unreachable("Not a call intrinsic");
2801 return NULL;
2802 }
2803 }
2804
nir_chase_binding(nir_src rsrc)2805 nir_binding nir_chase_binding(nir_src rsrc)
2806 {
2807 nir_binding res = {0};
2808 if (rsrc.ssa->parent_instr->type == nir_instr_type_deref) {
2809 const struct glsl_type *type = glsl_without_array(nir_src_as_deref(rsrc)->type);
2810 bool is_image = glsl_type_is_image(type) || glsl_type_is_sampler(type);
2811 while (rsrc.ssa->parent_instr->type == nir_instr_type_deref) {
2812 nir_deref_instr *deref = nir_src_as_deref(rsrc);
2813
2814 if (deref->deref_type == nir_deref_type_var) {
2815 res.success = true;
2816 res.var = deref->var;
2817 res.desc_set = deref->var->data.descriptor_set;
2818 res.binding = deref->var->data.binding;
2819 return res;
2820 } else if (deref->deref_type == nir_deref_type_array && is_image) {
2821 if (res.num_indices == ARRAY_SIZE(res.indices))
2822 return (nir_binding){0};
2823 res.indices[res.num_indices++] = deref->arr.index;
2824 }
2825
2826 rsrc = deref->parent;
2827 }
2828 }
2829
2830 /* Skip copies and trimming. Trimming can appear as nir_op_mov instructions
2831 * when removing the offset from addresses. We also consider nir_op_is_vec()
2832 * instructions to skip trimming of vec2_index_32bit_offset addresses after
2833 * lowering ALU to scalar.
2834 */
2835 while (true) {
2836 nir_alu_instr *alu = nir_src_as_alu_instr(rsrc);
2837 nir_intrinsic_instr *intrin = nir_src_as_intrinsic(rsrc);
2838 if (alu && alu->op == nir_op_mov) {
2839 for (unsigned i = 0; i < alu->dest.dest.ssa.num_components; i++) {
2840 if (alu->src[0].swizzle[i] != i)
2841 return (nir_binding){0};
2842 }
2843 rsrc = alu->src[0].src;
2844 } else if (alu && nir_op_is_vec(alu->op)) {
2845 for (unsigned i = 0; i < nir_op_infos[alu->op].num_inputs; i++) {
2846 if (alu->src[i].swizzle[0] != i || alu->src[i].src.ssa != alu->src[0].src.ssa)
2847 return (nir_binding){0};
2848 }
2849 rsrc = alu->src[0].src;
2850 } else if (intrin && intrin->intrinsic == nir_intrinsic_read_first_invocation) {
2851 /* The caller might want to be aware if only the first invocation of
2852 * the indices are used.
2853 */
2854 res.read_first_invocation = true;
2855 rsrc = intrin->src[0];
2856 } else {
2857 break;
2858 }
2859 }
2860
2861 if (nir_src_is_const(rsrc)) {
2862 /* GL binding model after deref lowering */
2863 res.success = true;
2864 /* Can't use just nir_src_as_uint. Vulkan resource index produces a
2865 * vec2. Some drivers lower it to vec1 (to handle get_ssbo_size for
2866 * example) but others just keep it around as a vec2 (v3dv).
2867 */
2868 res.binding = nir_src_comp_as_uint(rsrc, 0);
2869 return res;
2870 }
2871
2872 /* otherwise, must be Vulkan binding model after deref lowering or GL bindless */
2873
2874 nir_intrinsic_instr *intrin = nir_src_as_intrinsic(rsrc);
2875 if (!intrin)
2876 return (nir_binding){0};
2877
2878 /* skip load_vulkan_descriptor */
2879 if (intrin->intrinsic == nir_intrinsic_load_vulkan_descriptor) {
2880 intrin = nir_src_as_intrinsic(intrin->src[0]);
2881 if (!intrin)
2882 return (nir_binding){0};
2883 }
2884
2885 if (intrin->intrinsic != nir_intrinsic_vulkan_resource_index)
2886 return (nir_binding){0};
2887
2888 assert(res.num_indices == 0);
2889 res.success = true;
2890 res.desc_set = nir_intrinsic_desc_set(intrin);
2891 res.binding = nir_intrinsic_binding(intrin);
2892 res.num_indices = 1;
2893 res.indices[0] = intrin->src[0];
2894 return res;
2895 }
2896
nir_get_binding_variable(nir_shader * shader,nir_binding binding)2897 nir_variable *nir_get_binding_variable(nir_shader *shader, nir_binding binding)
2898 {
2899 nir_variable *binding_var = NULL;
2900 unsigned count = 0;
2901
2902 if (!binding.success)
2903 return NULL;
2904
2905 if (binding.var)
2906 return binding.var;
2907
2908 nir_foreach_variable_with_modes(var, shader, nir_var_mem_ubo | nir_var_mem_ssbo) {
2909 if (var->data.descriptor_set == binding.desc_set && var->data.binding == binding.binding) {
2910 binding_var = var;
2911 count++;
2912 }
2913 }
2914
2915 /* Be conservative if another variable is using the same binding/desc_set
2916 * because the access mask might be different and we can't get it reliably.
2917 */
2918 if (count > 1)
2919 return NULL;
2920
2921 return binding_var;
2922 }
2923
2924 bool
nir_alu_instr_is_copy(nir_alu_instr * instr)2925 nir_alu_instr_is_copy(nir_alu_instr *instr)
2926 {
2927 assert(instr->src[0].src.is_ssa);
2928
2929 if (instr->op == nir_op_mov) {
2930 return !instr->dest.saturate &&
2931 !instr->src[0].abs &&
2932 !instr->src[0].negate;
2933 } else if (nir_op_is_vec(instr->op)) {
2934 for (unsigned i = 0; i < instr->dest.dest.ssa.num_components; i++) {
2935 if (instr->src[i].abs || instr->src[i].negate)
2936 return false;
2937 }
2938 return !instr->dest.saturate;
2939 } else {
2940 return false;
2941 }
2942 }
2943
2944 nir_ssa_scalar
nir_ssa_scalar_chase_movs(nir_ssa_scalar s)2945 nir_ssa_scalar_chase_movs(nir_ssa_scalar s)
2946 {
2947 while (nir_ssa_scalar_is_alu(s)) {
2948 nir_alu_instr *alu = nir_instr_as_alu(s.def->parent_instr);
2949 if (!nir_alu_instr_is_copy(alu))
2950 break;
2951
2952 if (alu->op == nir_op_mov) {
2953 s.def = alu->src[0].src.ssa;
2954 s.comp = alu->src[0].swizzle[s.comp];
2955 } else {
2956 assert(nir_op_is_vec(alu->op));
2957 s.def = alu->src[s.comp].src.ssa;
2958 s.comp = alu->src[s.comp].swizzle[0];
2959 }
2960 }
2961
2962 return s;
2963 }
2964
2965 nir_alu_type
nir_get_nir_type_for_glsl_base_type(enum glsl_base_type base_type)2966 nir_get_nir_type_for_glsl_base_type(enum glsl_base_type base_type)
2967 {
2968 switch (base_type) {
2969 case GLSL_TYPE_BOOL:
2970 return nir_type_bool1;
2971 break;
2972 case GLSL_TYPE_UINT:
2973 return nir_type_uint32;
2974 break;
2975 case GLSL_TYPE_INT:
2976 return nir_type_int32;
2977 break;
2978 case GLSL_TYPE_UINT16:
2979 return nir_type_uint16;
2980 break;
2981 case GLSL_TYPE_INT16:
2982 return nir_type_int16;
2983 break;
2984 case GLSL_TYPE_UINT8:
2985 return nir_type_uint8;
2986 case GLSL_TYPE_INT8:
2987 return nir_type_int8;
2988 case GLSL_TYPE_UINT64:
2989 return nir_type_uint64;
2990 break;
2991 case GLSL_TYPE_INT64:
2992 return nir_type_int64;
2993 break;
2994 case GLSL_TYPE_FLOAT:
2995 return nir_type_float32;
2996 break;
2997 case GLSL_TYPE_FLOAT16:
2998 return nir_type_float16;
2999 break;
3000 case GLSL_TYPE_DOUBLE:
3001 return nir_type_float64;
3002 break;
3003
3004 case GLSL_TYPE_SAMPLER:
3005 case GLSL_TYPE_TEXTURE:
3006 case GLSL_TYPE_IMAGE:
3007 case GLSL_TYPE_ATOMIC_UINT:
3008 case GLSL_TYPE_STRUCT:
3009 case GLSL_TYPE_INTERFACE:
3010 case GLSL_TYPE_ARRAY:
3011 case GLSL_TYPE_VOID:
3012 case GLSL_TYPE_SUBROUTINE:
3013 case GLSL_TYPE_FUNCTION:
3014 case GLSL_TYPE_ERROR:
3015 return nir_type_invalid;
3016 }
3017
3018 unreachable("unknown type");
3019 }
3020
3021 enum glsl_base_type
nir_get_glsl_base_type_for_nir_type(nir_alu_type base_type)3022 nir_get_glsl_base_type_for_nir_type(nir_alu_type base_type)
3023 {
3024 switch (base_type) {
3025 case nir_type_bool1:
3026 return GLSL_TYPE_BOOL;
3027 case nir_type_uint32:
3028 return GLSL_TYPE_UINT;
3029 case nir_type_int32:
3030 return GLSL_TYPE_INT;
3031 case nir_type_uint16:
3032 return GLSL_TYPE_UINT16;
3033 case nir_type_int16:
3034 return GLSL_TYPE_INT16;
3035 case nir_type_uint8:
3036 return GLSL_TYPE_UINT8;
3037 case nir_type_int8:
3038 return GLSL_TYPE_INT8;
3039 case nir_type_uint64:
3040 return GLSL_TYPE_UINT64;
3041 case nir_type_int64:
3042 return GLSL_TYPE_INT64;
3043 case nir_type_float32:
3044 return GLSL_TYPE_FLOAT;
3045 case nir_type_float16:
3046 return GLSL_TYPE_FLOAT16;
3047 case nir_type_float64:
3048 return GLSL_TYPE_DOUBLE;
3049
3050 default: unreachable("Not a sized nir_alu_type");
3051 }
3052 }
3053
3054 nir_op
nir_op_vec(unsigned components)3055 nir_op_vec(unsigned components)
3056 {
3057 switch (components) {
3058 case 1: return nir_op_mov;
3059 case 2: return nir_op_vec2;
3060 case 3: return nir_op_vec3;
3061 case 4: return nir_op_vec4;
3062 case 5: return nir_op_vec5;
3063 case 8: return nir_op_vec8;
3064 case 16: return nir_op_vec16;
3065 default: unreachable("bad component count");
3066 }
3067 }
3068
3069 bool
nir_op_is_vec(nir_op op)3070 nir_op_is_vec(nir_op op)
3071 {
3072 switch (op) {
3073 case nir_op_mov:
3074 case nir_op_vec2:
3075 case nir_op_vec3:
3076 case nir_op_vec4:
3077 case nir_op_vec5:
3078 case nir_op_vec8:
3079 case nir_op_vec16:
3080 return true;
3081 default:
3082 return false;
3083 }
3084 }
3085
3086 bool
nir_alu_instr_channel_used(const nir_alu_instr * instr,unsigned src,unsigned channel)3087 nir_alu_instr_channel_used(const nir_alu_instr *instr, unsigned src,
3088 unsigned channel)
3089 {
3090 if (nir_op_infos[instr->op].input_sizes[src] > 0)
3091 return channel < nir_op_infos[instr->op].input_sizes[src];
3092
3093 return (instr->dest.write_mask >> channel) & 1;
3094 }
3095
3096 nir_component_mask_t
nir_alu_instr_src_read_mask(const nir_alu_instr * instr,unsigned src)3097 nir_alu_instr_src_read_mask(const nir_alu_instr *instr, unsigned src)
3098 {
3099 nir_component_mask_t read_mask = 0;
3100 for (unsigned c = 0; c < NIR_MAX_VEC_COMPONENTS; c++) {
3101 if (!nir_alu_instr_channel_used(instr, src, c))
3102 continue;
3103
3104 read_mask |= (1 << instr->src[src].swizzle[c]);
3105 }
3106 return read_mask;
3107 }
3108
3109 unsigned
nir_ssa_alu_instr_src_components(const nir_alu_instr * instr,unsigned src)3110 nir_ssa_alu_instr_src_components(const nir_alu_instr *instr, unsigned src)
3111 {
3112 if (nir_op_infos[instr->op].input_sizes[src] > 0)
3113 return nir_op_infos[instr->op].input_sizes[src];
3114
3115 return nir_dest_num_components(instr->dest.dest);
3116 }
3117
3118 #define CASE_ALL_SIZES(op) \
3119 case op: \
3120 case op ## 8: \
3121 case op ## 16: \
3122 case op ## 32: \
3123
3124 bool
nir_alu_instr_is_comparison(const nir_alu_instr * instr)3125 nir_alu_instr_is_comparison(const nir_alu_instr *instr)
3126 {
3127 switch (instr->op) {
3128 CASE_ALL_SIZES(nir_op_flt)
3129 CASE_ALL_SIZES(nir_op_fge)
3130 CASE_ALL_SIZES(nir_op_feq)
3131 CASE_ALL_SIZES(nir_op_fneu)
3132 CASE_ALL_SIZES(nir_op_ilt)
3133 CASE_ALL_SIZES(nir_op_ult)
3134 CASE_ALL_SIZES(nir_op_ige)
3135 CASE_ALL_SIZES(nir_op_uge)
3136 CASE_ALL_SIZES(nir_op_ieq)
3137 CASE_ALL_SIZES(nir_op_ine)
3138 case nir_op_i2b1:
3139 case nir_op_i2b8:
3140 case nir_op_i2b16:
3141 case nir_op_i2b32:
3142 case nir_op_f2b1:
3143 case nir_op_f2b8:
3144 case nir_op_f2b16:
3145 case nir_op_f2b32:
3146 case nir_op_inot:
3147 return true;
3148 default:
3149 return false;
3150 }
3151 }
3152
3153 #undef CASE_ALL_SIZES
3154
3155 unsigned
nir_intrinsic_src_components(const nir_intrinsic_instr * intr,unsigned srcn)3156 nir_intrinsic_src_components(const nir_intrinsic_instr *intr, unsigned srcn)
3157 {
3158 const nir_intrinsic_info *info = &nir_intrinsic_infos[intr->intrinsic];
3159 assert(srcn < info->num_srcs);
3160 if (info->src_components[srcn] > 0)
3161 return info->src_components[srcn];
3162 else if (info->src_components[srcn] == 0)
3163 return intr->num_components;
3164 else
3165 return nir_src_num_components(intr->src[srcn]);
3166 }
3167
3168 unsigned
nir_intrinsic_dest_components(nir_intrinsic_instr * intr)3169 nir_intrinsic_dest_components(nir_intrinsic_instr *intr)
3170 {
3171 const nir_intrinsic_info *info = &nir_intrinsic_infos[intr->intrinsic];
3172 if (!info->has_dest)
3173 return 0;
3174 else if (info->dest_components)
3175 return info->dest_components;
3176 else
3177 return intr->num_components;
3178 }
3179
3180 /**
3181 * Helper to copy const_index[] from src to dst, without assuming they
3182 * match in order.
3183 */
3184 void
nir_intrinsic_copy_const_indices(nir_intrinsic_instr * dst,nir_intrinsic_instr * src)3185 nir_intrinsic_copy_const_indices(nir_intrinsic_instr *dst, nir_intrinsic_instr *src)
3186 {
3187 if (src->intrinsic == dst->intrinsic) {
3188 memcpy(dst->const_index, src->const_index, sizeof(dst->const_index));
3189 return;
3190 }
3191
3192 const nir_intrinsic_info *src_info = &nir_intrinsic_infos[src->intrinsic];
3193 const nir_intrinsic_info *dst_info = &nir_intrinsic_infos[dst->intrinsic];
3194
3195 for (unsigned i = 0; i < NIR_INTRINSIC_NUM_INDEX_FLAGS; i++) {
3196 if (src_info->index_map[i] == 0)
3197 continue;
3198
3199 /* require that dst instruction also uses the same const_index[]: */
3200 assert(dst_info->index_map[i] > 0);
3201
3202 dst->const_index[dst_info->index_map[i] - 1] =
3203 src->const_index[src_info->index_map[i] - 1];
3204 }
3205 }
3206
3207
3208 bool
nir_tex_instr_need_sampler(const nir_tex_instr * instr)3209 nir_tex_instr_need_sampler(const nir_tex_instr *instr)
3210 {
3211 switch (instr->op) {
3212 case nir_texop_txf:
3213 case nir_texop_txf_ms:
3214 case nir_texop_txs:
3215 case nir_texop_query_levels:
3216 case nir_texop_texture_samples:
3217 case nir_texop_samples_identical:
3218 return false;
3219 default:
3220 return true;
3221 }
3222 }
3223
3224 unsigned
nir_tex_instr_result_size(const nir_tex_instr * instr)3225 nir_tex_instr_result_size(const nir_tex_instr *instr)
3226 {
3227 switch (instr->op) {
3228 case nir_texop_txs: {
3229 unsigned ret;
3230 switch (instr->sampler_dim) {
3231 case GLSL_SAMPLER_DIM_1D:
3232 case GLSL_SAMPLER_DIM_BUF:
3233 ret = 1;
3234 break;
3235 case GLSL_SAMPLER_DIM_2D:
3236 case GLSL_SAMPLER_DIM_CUBE:
3237 case GLSL_SAMPLER_DIM_MS:
3238 case GLSL_SAMPLER_DIM_RECT:
3239 case GLSL_SAMPLER_DIM_EXTERNAL:
3240 case GLSL_SAMPLER_DIM_SUBPASS:
3241 ret = 2;
3242 break;
3243 case GLSL_SAMPLER_DIM_3D:
3244 ret = 3;
3245 break;
3246 default:
3247 unreachable("not reached");
3248 }
3249 if (instr->is_array)
3250 ret++;
3251 return ret;
3252 }
3253
3254 case nir_texop_lod:
3255 return 2;
3256
3257 case nir_texop_texture_samples:
3258 case nir_texop_query_levels:
3259 case nir_texop_samples_identical:
3260 case nir_texop_fragment_mask_fetch_amd:
3261 return 1;
3262
3263 default:
3264 if (instr->is_shadow && instr->is_new_style_shadow)
3265 return 1;
3266
3267 return 4;
3268 }
3269 }
3270
3271 bool
nir_tex_instr_is_query(const nir_tex_instr * instr)3272 nir_tex_instr_is_query(const nir_tex_instr *instr)
3273 {
3274 switch (instr->op) {
3275 case nir_texop_txs:
3276 case nir_texop_lod:
3277 case nir_texop_texture_samples:
3278 case nir_texop_query_levels:
3279 return true;
3280 case nir_texop_tex:
3281 case nir_texop_txb:
3282 case nir_texop_txl:
3283 case nir_texop_txd:
3284 case nir_texop_txf:
3285 case nir_texop_txf_ms:
3286 case nir_texop_txf_ms_fb:
3287 case nir_texop_txf_ms_mcs_intel:
3288 case nir_texop_tg4:
3289 return false;
3290 default:
3291 unreachable("Invalid texture opcode");
3292 }
3293 }
3294
3295 bool
nir_tex_instr_has_implicit_derivative(const nir_tex_instr * instr)3296 nir_tex_instr_has_implicit_derivative(const nir_tex_instr *instr)
3297 {
3298 switch (instr->op) {
3299 case nir_texop_tex:
3300 case nir_texop_txb:
3301 case nir_texop_lod:
3302 return true;
3303 default:
3304 return false;
3305 }
3306 }
3307
3308 nir_alu_type
nir_tex_instr_src_type(const nir_tex_instr * instr,unsigned src)3309 nir_tex_instr_src_type(const nir_tex_instr *instr, unsigned src)
3310 {
3311 switch (instr->src[src].src_type) {
3312 case nir_tex_src_coord:
3313 switch (instr->op) {
3314 case nir_texop_txf:
3315 case nir_texop_txf_ms:
3316 case nir_texop_txf_ms_fb:
3317 case nir_texop_txf_ms_mcs_intel:
3318 case nir_texop_samples_identical:
3319 case nir_texop_fragment_fetch_amd:
3320 case nir_texop_fragment_mask_fetch_amd:
3321 return nir_type_int;
3322
3323 default:
3324 return nir_type_float;
3325 }
3326
3327 case nir_tex_src_lod:
3328 switch (instr->op) {
3329 case nir_texop_txs:
3330 case nir_texop_txf:
3331 case nir_texop_txf_ms:
3332 case nir_texop_fragment_fetch_amd:
3333 case nir_texop_fragment_mask_fetch_amd:
3334 return nir_type_int;
3335
3336 default:
3337 return nir_type_float;
3338 }
3339
3340 case nir_tex_src_projector:
3341 case nir_tex_src_comparator:
3342 case nir_tex_src_bias:
3343 case nir_tex_src_min_lod:
3344 case nir_tex_src_ddx:
3345 case nir_tex_src_ddy:
3346 case nir_tex_src_backend1:
3347 case nir_tex_src_backend2:
3348 return nir_type_float;
3349
3350 case nir_tex_src_offset:
3351 case nir_tex_src_ms_index:
3352 case nir_tex_src_plane:
3353 return nir_type_int;
3354
3355 case nir_tex_src_ms_mcs_intel:
3356 case nir_tex_src_texture_deref:
3357 case nir_tex_src_sampler_deref:
3358 case nir_tex_src_texture_offset:
3359 case nir_tex_src_sampler_offset:
3360 case nir_tex_src_texture_handle:
3361 case nir_tex_src_sampler_handle:
3362 return nir_type_uint;
3363
3364 case nir_num_tex_src_types:
3365 unreachable("nir_num_tex_src_types is not a valid source type");
3366 }
3367
3368 unreachable("Invalid texture source type");
3369 }
3370
3371 unsigned
nir_tex_instr_src_size(const nir_tex_instr * instr,unsigned src)3372 nir_tex_instr_src_size(const nir_tex_instr *instr, unsigned src)
3373 {
3374 if (instr->src[src].src_type == nir_tex_src_coord)
3375 return instr->coord_components;
3376
3377 /* The MCS value is expected to be a vec4 returned by a txf_ms_mcs_intel */
3378 if (instr->src[src].src_type == nir_tex_src_ms_mcs_intel)
3379 return 4;
3380
3381 if (instr->src[src].src_type == nir_tex_src_ddx ||
3382 instr->src[src].src_type == nir_tex_src_ddy) {
3383
3384 if (instr->is_array && !instr->array_is_lowered_cube)
3385 return instr->coord_components - 1;
3386 else
3387 return instr->coord_components;
3388 }
3389
3390 if (instr->src[src].src_type == nir_tex_src_offset) {
3391 if (instr->is_array)
3392 return instr->coord_components - 1;
3393 else
3394 return instr->coord_components;
3395 }
3396
3397 if (instr->src[src].src_type == nir_tex_src_backend1 ||
3398 instr->src[src].src_type == nir_tex_src_backend2)
3399 return nir_src_num_components(instr->src[src].src);
3400
3401 /* For AMD, this can be a vec8/vec4 image/sampler descriptor. */
3402 if (instr->src[src].src_type == nir_tex_src_texture_handle ||
3403 instr->src[src].src_type == nir_tex_src_sampler_handle)
3404 return 0;
3405
3406 return 1;
3407 }
3408
3409 /**
3410 * Return which components are written into transform feedback buffers.
3411 * The result is relative to 0, not "component".
3412 */
3413 unsigned
nir_instr_xfb_write_mask(nir_intrinsic_instr * instr)3414 nir_instr_xfb_write_mask(nir_intrinsic_instr *instr)
3415 {
3416 unsigned mask = 0;
3417
3418 if (nir_intrinsic_has_io_xfb(instr)) {
3419 unsigned wr_mask = nir_intrinsic_write_mask(instr) <<
3420 nir_intrinsic_component(instr);
3421 assert((wr_mask & ~0xf) == 0); /* only 4 components allowed */
3422
3423 unsigned iter_mask = wr_mask;
3424 while (iter_mask) {
3425 unsigned i = u_bit_scan(&iter_mask);
3426 nir_io_xfb xfb = i < 2 ? nir_intrinsic_io_xfb(instr) :
3427 nir_intrinsic_io_xfb2(instr);
3428 if (xfb.out[i % 2].num_components)
3429 mask |= BITFIELD_RANGE(i, xfb.out[i % 2].num_components) & wr_mask;
3430 }
3431 }
3432
3433 return mask;
3434 }
3435
3436 /**
3437 * Whether an output slot is consumed by fixed-function logic.
3438 */
3439 bool
nir_slot_is_sysval_output(gl_varying_slot slot)3440 nir_slot_is_sysval_output(gl_varying_slot slot)
3441 {
3442 return slot == VARYING_SLOT_POS ||
3443 slot == VARYING_SLOT_PSIZ ||
3444 slot == VARYING_SLOT_EDGE ||
3445 slot == VARYING_SLOT_CLIP_VERTEX ||
3446 slot == VARYING_SLOT_CLIP_DIST0 ||
3447 slot == VARYING_SLOT_CLIP_DIST1 ||
3448 slot == VARYING_SLOT_CULL_DIST0 ||
3449 slot == VARYING_SLOT_CULL_DIST1 ||
3450 slot == VARYING_SLOT_LAYER ||
3451 slot == VARYING_SLOT_VIEWPORT ||
3452 slot == VARYING_SLOT_TESS_LEVEL_OUTER ||
3453 slot == VARYING_SLOT_TESS_LEVEL_INNER ||
3454 slot == VARYING_SLOT_BOUNDING_BOX0 ||
3455 slot == VARYING_SLOT_BOUNDING_BOX1 ||
3456 slot == VARYING_SLOT_VIEW_INDEX ||
3457 slot == VARYING_SLOT_VIEWPORT_MASK ||
3458 slot == VARYING_SLOT_PRIMITIVE_SHADING_RATE ||
3459 slot == VARYING_SLOT_PRIMITIVE_COUNT ||
3460 slot == VARYING_SLOT_PRIMITIVE_INDICES ||
3461 slot == VARYING_SLOT_TASK_COUNT;
3462 }
3463
3464 /**
3465 * Whether an input/output slot is consumed by the next shader stage,
3466 * or written by the previous shader stage.
3467 */
3468 bool
nir_slot_is_varying(gl_varying_slot slot)3469 nir_slot_is_varying(gl_varying_slot slot)
3470 {
3471 return slot >= VARYING_SLOT_VAR0 ||
3472 slot == VARYING_SLOT_COL0 ||
3473 slot == VARYING_SLOT_COL1 ||
3474 slot == VARYING_SLOT_BFC0 ||
3475 slot == VARYING_SLOT_BFC1 ||
3476 slot == VARYING_SLOT_FOGC ||
3477 (slot >= VARYING_SLOT_TEX0 && slot <= VARYING_SLOT_TEX7) ||
3478 slot == VARYING_SLOT_PNTC ||
3479 slot == VARYING_SLOT_CLIP_DIST0 ||
3480 slot == VARYING_SLOT_CLIP_DIST1 ||
3481 slot == VARYING_SLOT_CULL_DIST0 ||
3482 slot == VARYING_SLOT_CULL_DIST1 ||
3483 slot == VARYING_SLOT_PRIMITIVE_ID ||
3484 slot == VARYING_SLOT_LAYER ||
3485 slot == VARYING_SLOT_VIEWPORT ||
3486 slot == VARYING_SLOT_TESS_LEVEL_OUTER ||
3487 slot == VARYING_SLOT_TESS_LEVEL_INNER;
3488 }
3489
3490 bool
nir_slot_is_sysval_output_and_varying(gl_varying_slot slot)3491 nir_slot_is_sysval_output_and_varying(gl_varying_slot slot)
3492 {
3493 return nir_slot_is_sysval_output(slot) &&
3494 nir_slot_is_varying(slot);
3495 }
3496
3497 /**
3498 * This marks the output store instruction as not feeding the next shader
3499 * stage. If the instruction has no other use, it's removed.
3500 */
nir_remove_varying(nir_intrinsic_instr * intr)3501 void nir_remove_varying(nir_intrinsic_instr *intr)
3502 {
3503 nir_io_semantics sem = nir_intrinsic_io_semantics(intr);
3504
3505 if ((!sem.no_sysval_output && nir_slot_is_sysval_output(sem.location)) ||
3506 nir_instr_xfb_write_mask(intr)) {
3507 /* Demote the store instruction. */
3508 sem.no_varying = true;
3509 nir_intrinsic_set_io_semantics(intr, sem);
3510 } else {
3511 nir_instr_remove(&intr->instr);
3512 }
3513 }
3514
3515 /**
3516 * This marks the output store instruction as not feeding fixed-function
3517 * logic. If the instruction has no other use, it's removed.
3518 */
nir_remove_sysval_output(nir_intrinsic_instr * intr)3519 void nir_remove_sysval_output(nir_intrinsic_instr *intr)
3520 {
3521 nir_io_semantics sem = nir_intrinsic_io_semantics(intr);
3522
3523 if ((!sem.no_varying && nir_slot_is_varying(sem.location)) ||
3524 nir_instr_xfb_write_mask(intr)) {
3525 /* Demote the store instruction. */
3526 sem.no_sysval_output = true;
3527 nir_intrinsic_set_io_semantics(intr, sem);
3528 } else {
3529 nir_instr_remove(&intr->instr);
3530 }
3531 }
3532
nir_remove_non_entrypoints(nir_shader * nir)3533 void nir_remove_non_entrypoints(nir_shader *nir)
3534 {
3535 foreach_list_typed_safe(nir_function, func, node, &nir->functions) {
3536 if (!func->is_entrypoint)
3537 exec_node_remove(&func->node);
3538 }
3539 assert(exec_list_length(&nir->functions) == 1);
3540 }
3541