1 /*
2 * Copyright © 2020 Intel Corporation
3 *
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
10 *
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
13 * Software.
14 *
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
21 * IN THE SOFTWARE.
22 */
23
24 #include "nir.h"
25 #include "nir_builder.h"
26 #include "nir_phi_builder.h"
27 #include "util/u_math.h"
28
29 static bool
move_system_values_to_top(nir_shader * shader)30 move_system_values_to_top(nir_shader *shader)
31 {
32 nir_function_impl *impl = nir_shader_get_entrypoint(shader);
33
34 bool progress = false;
35 nir_foreach_block(block, impl) {
36 nir_foreach_instr_safe(instr, block) {
37 if (instr->type != nir_instr_type_intrinsic)
38 continue;
39
40 /* These intrinsics not only can't be re-materialized but aren't
41 * preserved when moving to the continuation shader. We have to move
42 * them to the top to ensure they get spilled as needed.
43 */
44 nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
45 switch (intrin->intrinsic) {
46 case nir_intrinsic_load_shader_record_ptr:
47 case nir_intrinsic_load_btd_local_arg_addr_intel:
48 nir_instr_remove(instr);
49 nir_instr_insert(nir_before_cf_list(&impl->body), instr);
50 progress = true;
51 break;
52
53 default:
54 break;
55 }
56 }
57 }
58
59 if (progress) {
60 nir_metadata_preserve(impl, nir_metadata_block_index |
61 nir_metadata_dominance);
62 } else {
63 nir_metadata_preserve(impl, nir_metadata_all);
64 }
65
66 return progress;
67 }
68
69 static bool
instr_is_shader_call(nir_instr * instr)70 instr_is_shader_call(nir_instr *instr)
71 {
72 if (instr->type != nir_instr_type_intrinsic)
73 return false;
74
75 nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
76 return intrin->intrinsic == nir_intrinsic_trace_ray ||
77 intrin->intrinsic == nir_intrinsic_report_ray_intersection ||
78 intrin->intrinsic == nir_intrinsic_execute_callable;
79 }
80
81 /* Previously named bitset, it had to be renamed as FreeBSD defines a struct
82 * named bitset in sys/_bitset.h required by pthread_np.h which is included
83 * from src/util/u_thread.h that is indirectly included by this file.
84 */
85 struct brw_bitset {
86 BITSET_WORD *set;
87 unsigned size;
88 };
89
90 static struct brw_bitset
bitset_create(void * mem_ctx,unsigned size)91 bitset_create(void *mem_ctx, unsigned size)
92 {
93 return (struct brw_bitset) {
94 .set = rzalloc_array(mem_ctx, BITSET_WORD, BITSET_WORDS(size)),
95 .size = size,
96 };
97 }
98
99 static bool
src_is_in_bitset(nir_src * src,void * _set)100 src_is_in_bitset(nir_src *src, void *_set)
101 {
102 struct brw_bitset *set = _set;
103 assert(src->is_ssa);
104
105 /* Any SSA values which were added after we generated liveness information
106 * are things generated by this pass and, while most of it is arithmetic
107 * which we could re-materialize, we don't need to because it's only used
108 * for a single load/store and so shouldn't cross any shader calls.
109 */
110 if (src->ssa->index >= set->size)
111 return false;
112
113 return BITSET_TEST(set->set, src->ssa->index);
114 }
115
116 static void
add_ssa_def_to_bitset(nir_ssa_def * def,struct brw_bitset * set)117 add_ssa_def_to_bitset(nir_ssa_def *def, struct brw_bitset *set)
118 {
119 if (def->index >= set->size)
120 return;
121
122 BITSET_SET(set->set, def->index);
123 }
124
125 static bool
can_remat_instr(nir_instr * instr,struct brw_bitset * remat)126 can_remat_instr(nir_instr *instr, struct brw_bitset *remat)
127 {
128 /* Set of all values which are trivially re-materializable and we shouldn't
129 * ever spill them. This includes:
130 *
131 * - Undef values
132 * - Constants
133 * - Uniforms (UBO or push constant)
134 * - ALU combinations of any of the above
135 * - Derefs which are either complete or casts of any of the above
136 *
137 * Because this pass rewrites things in-order and phis are always turned
138 * into register writes, We can use "is it SSA?" to answer the question
139 * "can my source be re-materialized?".
140 */
141 switch (instr->type) {
142 case nir_instr_type_alu:
143 if (!nir_instr_as_alu(instr)->dest.dest.is_ssa)
144 return false;
145
146 return nir_foreach_src(instr, src_is_in_bitset, remat);
147
148 case nir_instr_type_deref:
149 return nir_foreach_src(instr, src_is_in_bitset, remat);
150
151 case nir_instr_type_intrinsic: {
152 nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
153 switch (intrin->intrinsic) {
154 case nir_intrinsic_load_ubo:
155 case nir_intrinsic_vulkan_resource_index:
156 case nir_intrinsic_vulkan_resource_reindex:
157 case nir_intrinsic_load_vulkan_descriptor:
158 case nir_intrinsic_load_push_constant:
159 /* These intrinsics don't need to be spilled as long as they don't
160 * depend on any spilled values.
161 */
162 return nir_foreach_src(instr, src_is_in_bitset, remat);
163
164 case nir_intrinsic_load_scratch_base_ptr:
165 case nir_intrinsic_load_ray_launch_id:
166 case nir_intrinsic_load_topology_id_intel:
167 case nir_intrinsic_load_btd_global_arg_addr_intel:
168 case nir_intrinsic_load_btd_resume_sbt_addr_intel:
169 case nir_intrinsic_load_ray_base_mem_addr_intel:
170 case nir_intrinsic_load_ray_hw_stack_size_intel:
171 case nir_intrinsic_load_ray_sw_stack_size_intel:
172 case nir_intrinsic_load_ray_num_dss_rt_stacks_intel:
173 case nir_intrinsic_load_ray_hit_sbt_addr_intel:
174 case nir_intrinsic_load_ray_hit_sbt_stride_intel:
175 case nir_intrinsic_load_ray_miss_sbt_addr_intel:
176 case nir_intrinsic_load_ray_miss_sbt_stride_intel:
177 case nir_intrinsic_load_callable_sbt_addr_intel:
178 case nir_intrinsic_load_callable_sbt_stride_intel:
179 case nir_intrinsic_load_reloc_const_intel:
180 case nir_intrinsic_load_ray_query_global_intel:
181 /* Notably missing from the above list is btd_local_arg_addr_intel.
182 * This is because the resume shader will have a different local
183 * argument pointer because it has a different BSR. Any access of
184 * the original shader's local arguments needs to be preserved so
185 * that pointer has to be saved on the stack.
186 *
187 * TODO: There may be some system values we want to avoid
188 * re-materializing as well but we have to be very careful
189 * to ensure that it's a system value which cannot change
190 * across a shader call.
191 */
192 return true;
193
194 default:
195 return false;
196 }
197 }
198
199 case nir_instr_type_ssa_undef:
200 case nir_instr_type_load_const:
201 return true;
202
203 default:
204 return false;
205 }
206 }
207
208 static bool
can_remat_ssa_def(nir_ssa_def * def,struct brw_bitset * remat)209 can_remat_ssa_def(nir_ssa_def *def, struct brw_bitset *remat)
210 {
211 return can_remat_instr(def->parent_instr, remat);
212 }
213
214 static nir_ssa_def *
remat_ssa_def(nir_builder * b,nir_ssa_def * def)215 remat_ssa_def(nir_builder *b, nir_ssa_def *def)
216 {
217 nir_instr *clone = nir_instr_clone(b->shader, def->parent_instr);
218 nir_builder_instr_insert(b, clone);
219 return nir_instr_ssa_def(clone);
220 }
221
222 struct pbv_array {
223 struct nir_phi_builder_value **arr;
224 unsigned len;
225 };
226
227 static struct nir_phi_builder_value *
get_phi_builder_value_for_def(nir_ssa_def * def,struct pbv_array * pbv_arr)228 get_phi_builder_value_for_def(nir_ssa_def *def,
229 struct pbv_array *pbv_arr)
230 {
231 if (def->index >= pbv_arr->len)
232 return NULL;
233
234 return pbv_arr->arr[def->index];
235 }
236
237 static nir_ssa_def *
get_phi_builder_def_for_src(nir_src * src,struct pbv_array * pbv_arr,nir_block * block)238 get_phi_builder_def_for_src(nir_src *src, struct pbv_array *pbv_arr,
239 nir_block *block)
240 {
241 assert(src->is_ssa);
242
243 struct nir_phi_builder_value *pbv =
244 get_phi_builder_value_for_def(src->ssa, pbv_arr);
245 if (pbv == NULL)
246 return NULL;
247
248 return nir_phi_builder_value_get_block_def(pbv, block);
249 }
250
251 static bool
rewrite_instr_src_from_phi_builder(nir_src * src,void * _pbv_arr)252 rewrite_instr_src_from_phi_builder(nir_src *src, void *_pbv_arr)
253 {
254 nir_block *block;
255 if (src->parent_instr->type == nir_instr_type_phi) {
256 nir_phi_src *phi_src = exec_node_data(nir_phi_src, src, src);
257 block = phi_src->pred;
258 } else {
259 block = src->parent_instr->block;
260 }
261
262 nir_ssa_def *new_def = get_phi_builder_def_for_src(src, _pbv_arr, block);
263 if (new_def != NULL)
264 nir_instr_rewrite_src(src->parent_instr, src, nir_src_for_ssa(new_def));
265 return true;
266 }
267
268 static nir_ssa_def *
spill_fill(nir_builder * before,nir_builder * after,nir_ssa_def * def,unsigned offset,nir_address_format address_format,unsigned stack_alignment)269 spill_fill(nir_builder *before, nir_builder *after, nir_ssa_def *def, unsigned offset,
270 nir_address_format address_format, unsigned stack_alignment)
271 {
272 const unsigned comp_size = def->bit_size / 8;
273
274 switch(address_format) {
275 case nir_address_format_32bit_offset:
276 nir_store_scratch(before, def, nir_imm_int(before, offset),
277 .align_mul = MIN2(comp_size, stack_alignment),
278 .write_mask = BITFIELD_MASK(def->num_components));
279 def = nir_load_scratch(after, def->num_components, def->bit_size,
280 nir_imm_int(after, offset), .align_mul = MIN2(comp_size, stack_alignment));
281 break;
282 case nir_address_format_64bit_global: {
283 nir_ssa_def *addr = nir_iadd_imm(before, nir_load_scratch_base_ptr(before, 1, 64, 1), offset);
284 nir_store_global(before, addr, MIN2(comp_size, stack_alignment), def, ~0);
285 addr = nir_iadd_imm(after, nir_load_scratch_base_ptr(after, 1, 64, 1), offset);
286 def = nir_load_global(after, addr, MIN2(comp_size, stack_alignment),
287 def->num_components, def->bit_size);
288 break;
289 }
290 default:
291 unreachable("Unimplemented address format");
292 }
293 return def;
294 }
295
296 static void
spill_ssa_defs_and_lower_shader_calls(nir_shader * shader,uint32_t num_calls,nir_address_format address_format,unsigned stack_alignment)297 spill_ssa_defs_and_lower_shader_calls(nir_shader *shader, uint32_t num_calls,
298 nir_address_format address_format,
299 unsigned stack_alignment)
300 {
301 /* TODO: If a SSA def is filled more than once, we probably want to just
302 * spill it at the LCM of the fill sites so we avoid unnecessary
303 * extra spills
304 *
305 * TODO: If a SSA def is defined outside a loop but live through some call
306 * inside the loop, we probably want to spill outside the loop. We
307 * may also want to fill outside the loop if it's not used in the
308 * loop.
309 *
310 * TODO: Right now, we only re-materialize things if their immediate
311 * sources are things which we filled. We probably want to expand
312 * that to re-materialize things whose sources are things we can
313 * re-materialize from things we filled. We may want some DAG depth
314 * heuristic on this.
315 */
316
317 /* This happens per-shader rather than per-impl because we mess with
318 * nir_shader::scratch_size.
319 */
320 nir_function_impl *impl = nir_shader_get_entrypoint(shader);
321
322 nir_metadata_require(impl, nir_metadata_live_ssa_defs |
323 nir_metadata_dominance |
324 nir_metadata_block_index);
325
326 void *mem_ctx = ralloc_context(shader);
327
328 const unsigned num_ssa_defs = impl->ssa_alloc;
329 const unsigned live_words = BITSET_WORDS(num_ssa_defs);
330 struct brw_bitset trivial_remat = bitset_create(mem_ctx, num_ssa_defs);
331
332 /* Array of all live SSA defs which are spill candidates */
333 nir_ssa_def **spill_defs =
334 rzalloc_array(mem_ctx, nir_ssa_def *, num_ssa_defs);
335
336 /* For each spill candidate, an array of every time it's defined by a fill,
337 * indexed by call instruction index.
338 */
339 nir_ssa_def ***fill_defs =
340 rzalloc_array(mem_ctx, nir_ssa_def **, num_ssa_defs);
341
342 /* For each call instruction, the liveness set at the call */
343 const BITSET_WORD **call_live =
344 rzalloc_array(mem_ctx, const BITSET_WORD *, num_calls);
345
346 /* For each call instruction, the block index of the block it lives in */
347 uint32_t *call_block_indices = rzalloc_array(mem_ctx, uint32_t, num_calls);
348
349 /* Walk the call instructions and fetch the liveness set and block index
350 * for each one. We need to do this before we start modifying the shader
351 * so that liveness doesn't complain that it's been invalidated. Don't
352 * worry, we'll be very careful with our live sets. :-)
353 */
354 unsigned call_idx = 0;
355 nir_foreach_block(block, impl) {
356 nir_foreach_instr(instr, block) {
357 if (!instr_is_shader_call(instr))
358 continue;
359
360 call_block_indices[call_idx] = block->index;
361
362 /* The objective here is to preserve values around shader call
363 * instructions. Therefore, we use the live set after the
364 * instruction as the set of things we want to preserve. Because
365 * none of our shader call intrinsics return anything, we don't have
366 * to worry about spilling over a return value.
367 *
368 * TODO: This isn't quite true for report_intersection.
369 */
370 call_live[call_idx] =
371 nir_get_live_ssa_defs(nir_after_instr(instr), mem_ctx);
372
373 call_idx++;
374 }
375 }
376
377 nir_builder before, after;
378 nir_builder_init(&before, impl);
379 nir_builder_init(&after, impl);
380
381 call_idx = 0;
382 unsigned max_scratch_size = shader->scratch_size;
383 nir_foreach_block(block, impl) {
384 nir_foreach_instr_safe(instr, block) {
385 nir_ssa_def *def = nir_instr_ssa_def(instr);
386 if (def != NULL) {
387 if (can_remat_ssa_def(def, &trivial_remat)) {
388 add_ssa_def_to_bitset(def, &trivial_remat);
389 } else {
390 spill_defs[def->index] = def;
391 }
392 }
393
394 if (!instr_is_shader_call(instr))
395 continue;
396
397 const BITSET_WORD *live = call_live[call_idx];
398
399 /* Make a copy of trivial_remat that we'll update as we crawl through
400 * the live SSA defs and unspill them.
401 */
402 struct brw_bitset remat = bitset_create(mem_ctx, num_ssa_defs);
403 memcpy(remat.set, trivial_remat.set, live_words * sizeof(BITSET_WORD));
404
405 /* Before the two builders are always separated by the call
406 * instruction, it won't break anything to have two of them.
407 */
408 before.cursor = nir_before_instr(instr);
409 after.cursor = nir_after_instr(instr);
410
411 unsigned offset = shader->scratch_size;
412 for (unsigned w = 0; w < live_words; w++) {
413 BITSET_WORD spill_mask = live[w] & ~trivial_remat.set[w];
414 while (spill_mask) {
415 int i = u_bit_scan(&spill_mask);
416 assert(i >= 0);
417 unsigned index = w * BITSET_WORDBITS + i;
418 assert(index < num_ssa_defs);
419
420 nir_ssa_def *def = spill_defs[index];
421 if (can_remat_ssa_def(def, &remat)) {
422 /* If this SSA def is re-materializable or based on other
423 * things we've already spilled, re-materialize it rather
424 * than spilling and filling. Anything which is trivially
425 * re-materializable won't even get here because we take
426 * those into account in spill_mask above.
427 */
428 def = remat_ssa_def(&after, def);
429 } else {
430 bool is_bool = def->bit_size == 1;
431 if (is_bool)
432 def = nir_b2b32(&before, def);
433
434 const unsigned comp_size = def->bit_size / 8;
435 offset = ALIGN(offset, comp_size);
436
437 def = spill_fill(&before, &after, def, offset,
438 address_format,stack_alignment);
439
440 if (is_bool)
441 def = nir_b2b1(&after, def);
442
443 offset += def->num_components * comp_size;
444 }
445
446 /* Mark this SSA def as available in the remat set so that, if
447 * some other SSA def we need is computed based on it, we can
448 * just re-compute instead of fetching from memory.
449 */
450 BITSET_SET(remat.set, index);
451
452 /* For now, we just make a note of this new SSA def. We'll
453 * fix things up with the phi builder as a second pass.
454 */
455 if (fill_defs[index] == NULL) {
456 fill_defs[index] =
457 rzalloc_array(mem_ctx, nir_ssa_def *, num_calls);
458 }
459 fill_defs[index][call_idx] = def;
460 }
461 }
462
463 nir_builder *b = &before;
464
465 offset = ALIGN(offset, stack_alignment);
466 max_scratch_size = MAX2(max_scratch_size, offset);
467
468 /* First thing on the called shader's stack is the resume address
469 * followed by a pointer to the payload.
470 */
471 nir_intrinsic_instr *call = nir_instr_as_intrinsic(instr);
472
473 /* Lower to generic intrinsics with information about the stack & resume shader. */
474 switch (call->intrinsic) {
475 case nir_intrinsic_trace_ray: {
476 nir_rt_trace_ray(b, call->src[0].ssa, call->src[1].ssa,
477 call->src[2].ssa, call->src[3].ssa,
478 call->src[4].ssa, call->src[5].ssa,
479 call->src[6].ssa, call->src[7].ssa,
480 call->src[8].ssa, call->src[9].ssa,
481 call->src[10].ssa,
482 .call_idx = call_idx, .stack_size = offset);
483 break;
484 }
485
486 case nir_intrinsic_report_ray_intersection:
487 unreachable("Any-hit shaders must be inlined");
488
489 case nir_intrinsic_execute_callable: {
490 nir_rt_execute_callable(b, call->src[0].ssa, call->src[1].ssa, .call_idx = call_idx, .stack_size = offset);
491 break;
492 }
493
494 default:
495 unreachable("Invalid shader call instruction");
496 }
497
498 nir_rt_resume(b, .call_idx = call_idx, .stack_size = offset);
499
500 nir_instr_remove(&call->instr);
501
502 call_idx++;
503 }
504 }
505 assert(call_idx == num_calls);
506 shader->scratch_size = max_scratch_size;
507
508 struct nir_phi_builder *pb = nir_phi_builder_create(impl);
509 struct pbv_array pbv_arr = {
510 .arr = rzalloc_array(mem_ctx, struct nir_phi_builder_value *,
511 num_ssa_defs),
512 .len = num_ssa_defs,
513 };
514
515 const unsigned block_words = BITSET_WORDS(impl->num_blocks);
516 BITSET_WORD *def_blocks = ralloc_array(mem_ctx, BITSET_WORD, block_words);
517
518 /* Go through and set up phi builder values for each spillable value which
519 * we ever needed to spill at any point.
520 */
521 for (unsigned index = 0; index < num_ssa_defs; index++) {
522 if (fill_defs[index] == NULL)
523 continue;
524
525 nir_ssa_def *def = spill_defs[index];
526
527 memset(def_blocks, 0, block_words * sizeof(BITSET_WORD));
528 BITSET_SET(def_blocks, def->parent_instr->block->index);
529 for (unsigned call_idx = 0; call_idx < num_calls; call_idx++) {
530 if (fill_defs[index][call_idx] != NULL)
531 BITSET_SET(def_blocks, call_block_indices[call_idx]);
532 }
533
534 pbv_arr.arr[index] = nir_phi_builder_add_value(pb, def->num_components,
535 def->bit_size, def_blocks);
536 }
537
538 /* Walk the shader one more time and rewrite SSA defs as needed using the
539 * phi builder.
540 */
541 nir_foreach_block(block, impl) {
542 nir_foreach_instr_safe(instr, block) {
543 nir_ssa_def *def = nir_instr_ssa_def(instr);
544 if (def != NULL) {
545 struct nir_phi_builder_value *pbv =
546 get_phi_builder_value_for_def(def, &pbv_arr);
547 if (pbv != NULL)
548 nir_phi_builder_value_set_block_def(pbv, block, def);
549 }
550
551 if (instr->type == nir_instr_type_phi)
552 continue;
553
554 nir_foreach_src(instr, rewrite_instr_src_from_phi_builder, &pbv_arr);
555
556 if (instr->type != nir_instr_type_intrinsic)
557 continue;
558
559 nir_intrinsic_instr *resume = nir_instr_as_intrinsic(instr);
560 if (resume->intrinsic != nir_intrinsic_rt_resume)
561 continue;
562
563 call_idx = nir_intrinsic_call_idx(resume);
564
565 /* Technically, this is the wrong place to add the fill defs to the
566 * phi builder values because we haven't seen any of the load_scratch
567 * instructions for this call yet. However, we know based on how we
568 * emitted them that no value ever gets used until after the load
569 * instruction has been emitted so this should be safe. If we ever
570 * fail validation due this it likely means a bug in our spilling
571 * code and not the phi re-construction code here.
572 */
573 for (unsigned index = 0; index < num_ssa_defs; index++) {
574 if (fill_defs[index] && fill_defs[index][call_idx]) {
575 nir_phi_builder_value_set_block_def(pbv_arr.arr[index], block,
576 fill_defs[index][call_idx]);
577 }
578 }
579 }
580
581 nir_if *following_if = nir_block_get_following_if(block);
582 if (following_if) {
583 nir_ssa_def *new_def =
584 get_phi_builder_def_for_src(&following_if->condition,
585 &pbv_arr, block);
586 if (new_def != NULL)
587 nir_if_rewrite_condition(following_if, nir_src_for_ssa(new_def));
588 }
589
590 /* Handle phi sources that source from this block. We have to do this
591 * as a separate pass because the phi builder assumes that uses and
592 * defs are processed in an order that respects dominance. When we have
593 * loops, a phi source may be a back-edge so we have to handle it as if
594 * it were one of the last instructions in the predecessor block.
595 */
596 nir_foreach_phi_src_leaving_block(block,
597 rewrite_instr_src_from_phi_builder,
598 &pbv_arr);
599 }
600
601 nir_phi_builder_finish(pb);
602
603 ralloc_free(mem_ctx);
604
605 nir_metadata_preserve(impl, nir_metadata_block_index |
606 nir_metadata_dominance);
607 }
608
609 static nir_instr *
find_resume_instr(nir_function_impl * impl,unsigned call_idx)610 find_resume_instr(nir_function_impl *impl, unsigned call_idx)
611 {
612 nir_foreach_block(block, impl) {
613 nir_foreach_instr(instr, block) {
614 if (instr->type != nir_instr_type_intrinsic)
615 continue;
616
617 nir_intrinsic_instr *resume = nir_instr_as_intrinsic(instr);
618 if (resume->intrinsic != nir_intrinsic_rt_resume)
619 continue;
620
621 if (nir_intrinsic_call_idx(resume) == call_idx)
622 return &resume->instr;
623 }
624 }
625 unreachable("Couldn't find resume instruction");
626 }
627
628 /* Walk the CF tree and duplicate the contents of every loop, one half runs on
629 * resume and the other half is for any post-resume loop iterations. We are
630 * careful in our duplication to ensure that resume_instr is in the resume
631 * half of the loop though a copy of resume_instr will remain in the other
632 * half as well in case the same shader call happens twice.
633 */
634 static bool
duplicate_loop_bodies(nir_function_impl * impl,nir_instr * resume_instr)635 duplicate_loop_bodies(nir_function_impl *impl, nir_instr *resume_instr)
636 {
637 nir_register *resume_reg = NULL;
638 for (nir_cf_node *node = resume_instr->block->cf_node.parent;
639 node->type != nir_cf_node_function; node = node->parent) {
640 if (node->type != nir_cf_node_loop)
641 continue;
642
643 nir_loop *loop = nir_cf_node_as_loop(node);
644
645 if (resume_reg == NULL) {
646 /* We only create resume_reg if we encounter a loop. This way we can
647 * avoid re-validating the shader and calling ssa_to_regs in the case
648 * where it's just if-ladders.
649 */
650 resume_reg = nir_local_reg_create(impl);
651 resume_reg->num_components = 1;
652 resume_reg->bit_size = 1;
653
654 nir_builder b;
655 nir_builder_init(&b, impl);
656
657 /* Initialize resume to true */
658 b.cursor = nir_before_cf_list(&impl->body);
659 nir_store_reg(&b, resume_reg, nir_imm_true(&b), 1);
660
661 /* Set resume to false right after the resume instruction */
662 b.cursor = nir_after_instr(resume_instr);
663 nir_store_reg(&b, resume_reg, nir_imm_false(&b), 1);
664 }
665
666 /* Before we go any further, make sure that everything which exits the
667 * loop or continues around to the top of the loop does so through
668 * registers. We're about to duplicate the loop body and we'll have
669 * serious trouble if we don't do this.
670 */
671 nir_convert_loop_to_lcssa(loop);
672 nir_lower_phis_to_regs_block(nir_loop_first_block(loop));
673 nir_lower_phis_to_regs_block(
674 nir_cf_node_as_block(nir_cf_node_next(&loop->cf_node)));
675
676 nir_cf_list cf_list;
677 nir_cf_list_extract(&cf_list, &loop->body);
678
679 nir_if *_if = nir_if_create(impl->function->shader);
680 _if->condition = nir_src_for_reg(resume_reg);
681 nir_cf_node_insert(nir_after_cf_list(&loop->body), &_if->cf_node);
682
683 nir_cf_list clone;
684 nir_cf_list_clone(&clone, &cf_list, &loop->cf_node, NULL);
685
686 /* Insert the clone in the else and the original in the then so that
687 * the resume_instr remains valid even after the duplication.
688 */
689 nir_cf_reinsert(&cf_list, nir_before_cf_list(&_if->then_list));
690 nir_cf_reinsert(&clone, nir_before_cf_list(&_if->else_list));
691 }
692
693 if (resume_reg != NULL)
694 nir_metadata_preserve(impl, nir_metadata_none);
695
696 return resume_reg != NULL;
697 }
698
699 static bool
cf_node_contains_block(nir_cf_node * node,nir_block * block)700 cf_node_contains_block(nir_cf_node *node, nir_block *block)
701 {
702 for (nir_cf_node *n = &block->cf_node; n != NULL; n = n->parent) {
703 if (n == node)
704 return true;
705 }
706
707 return false;
708 }
709
710 static void
rewrite_phis_to_pred(nir_block * block,nir_block * pred)711 rewrite_phis_to_pred(nir_block *block, nir_block *pred)
712 {
713 nir_foreach_instr(instr, block) {
714 if (instr->type != nir_instr_type_phi)
715 break;
716
717 nir_phi_instr *phi = nir_instr_as_phi(instr);
718
719 ASSERTED bool found = false;
720 nir_foreach_phi_src(phi_src, phi) {
721 if (phi_src->pred == pred) {
722 found = true;
723 assert(phi_src->src.is_ssa);
724 nir_ssa_def_rewrite_uses(&phi->dest.ssa, phi_src->src.ssa);
725 break;
726 }
727 }
728 assert(found);
729 }
730 }
731
732 static bool
cursor_is_after_jump(nir_cursor cursor)733 cursor_is_after_jump(nir_cursor cursor)
734 {
735 switch (cursor.option) {
736 case nir_cursor_before_instr:
737 case nir_cursor_before_block:
738 return false;
739 case nir_cursor_after_instr:
740 return cursor.instr->type == nir_instr_type_jump;
741 case nir_cursor_after_block:
742 return nir_block_ends_in_jump(cursor.block);;
743 }
744 unreachable("Invalid cursor option");
745 }
746
747 /** Flattens if ladders leading up to a resume
748 *
749 * Given a resume_instr, this function flattens any if ladders leading to the
750 * resume instruction and deletes any code that cannot be encountered on a
751 * direct path to the resume instruction. This way we get, for the most part,
752 * straight-line control-flow up to the resume instruction.
753 *
754 * While we do this flattening, we also move any code which is in the remat
755 * set up to the top of the function or to the top of the resume portion of
756 * the current loop. We don't worry about control-flow as we do this because
757 * phis will never be in the remat set (see can_remat_instr) and so nothing
758 * control-dependent will ever need to be re-materialized. It is possible
759 * that this algorithm will preserve too many instructions by moving them to
760 * the top but we leave that for DCE to clean up. Any code not in the remat
761 * set is deleted because it's either unused in the continuation or else
762 * unspilled from a previous continuation and the unspill code is after the
763 * resume instruction.
764 *
765 * If, for instance, we have something like this:
766 *
767 * // block 0
768 * if (cond1) {
769 * // block 1
770 * } else {
771 * // block 2
772 * if (cond2) {
773 * // block 3
774 * resume;
775 * if (cond3) {
776 * // block 4
777 * }
778 * } else {
779 * // block 5
780 * }
781 * }
782 *
783 * then we know, because we know the resume instruction had to be encoutered,
784 * that cond1 = false and cond2 = true and we lower as follows:
785 *
786 * // block 0
787 * // block 2
788 * // block 3
789 * resume;
790 * if (cond3) {
791 * // block 4
792 * }
793 *
794 * As you can see, the code in blocks 1 and 5 was removed because there is no
795 * path from the start of the shader to the resume instruction which execute
796 * blocks 1 or 5. Any remat code from blocks 0, 2, and 3 is preserved and
797 * moved to the top. If the resume instruction is inside a loop then we know
798 * a priori that it is of the form
799 *
800 * loop {
801 * if (resume) {
802 * // Contents containing resume_instr
803 * } else {
804 * // Second copy of contents
805 * }
806 * }
807 *
808 * In this case, we only descend into the first half of the loop. The second
809 * half is left alone as that portion is only ever executed after the resume
810 * instruction.
811 */
812 static bool
flatten_resume_if_ladder(nir_builder * b,nir_cf_node * parent_node,struct exec_list * child_list,bool child_list_contains_cursor,nir_instr * resume_instr,struct brw_bitset * remat)813 flatten_resume_if_ladder(nir_builder *b,
814 nir_cf_node *parent_node,
815 struct exec_list *child_list,
816 bool child_list_contains_cursor,
817 nir_instr *resume_instr,
818 struct brw_bitset *remat)
819 {
820 nir_cf_list cf_list;
821
822 /* If our child list contains the cursor instruction then we start out
823 * before the cursor instruction. We need to know this so that we can skip
824 * moving instructions which are already before the cursor.
825 */
826 bool before_cursor = child_list_contains_cursor;
827
828 nir_cf_node *resume_node = NULL;
829 foreach_list_typed_safe(nir_cf_node, child, node, child_list) {
830 switch (child->type) {
831 case nir_cf_node_block: {
832 nir_block *block = nir_cf_node_as_block(child);
833 if (b->cursor.option == nir_cursor_before_block &&
834 b->cursor.block == block) {
835 assert(before_cursor);
836 before_cursor = false;
837 }
838 nir_foreach_instr_safe(instr, block) {
839 if ((b->cursor.option == nir_cursor_before_instr ||
840 b->cursor.option == nir_cursor_after_instr) &&
841 b->cursor.instr == instr) {
842 assert(nir_cf_node_is_first(&block->cf_node));
843 assert(before_cursor);
844 before_cursor = false;
845 continue;
846 }
847
848 if (instr == resume_instr)
849 goto found_resume;
850
851 if (!before_cursor && can_remat_instr(instr, remat)) {
852 nir_instr_remove(instr);
853 nir_instr_insert(b->cursor, instr);
854 b->cursor = nir_after_instr(instr);
855
856 nir_ssa_def *def = nir_instr_ssa_def(instr);
857 BITSET_SET(remat->set, def->index);
858 }
859 }
860 if (b->cursor.option == nir_cursor_after_block &&
861 b->cursor.block == block) {
862 assert(before_cursor);
863 before_cursor = false;
864 }
865 break;
866 }
867
868 case nir_cf_node_if: {
869 nir_if *_if = nir_cf_node_as_if(child);
870
871 /* Because of the dummy blocks inserted in the first if block of the
872 * loops, it's possible we find an empty if block that contains our
873 * cursor. At this point, the block should still be empty and we can
874 * just skip it and consider we're after the cursor.
875 */
876 if (cf_node_contains_block(&_if->cf_node,
877 nir_cursor_current_block(b->cursor))) {
878 /* Some sanity checks to verify this is actually a dummy block */
879 assert(nir_src_as_bool(_if->condition) == true);
880 assert(nir_cf_list_is_empty_block(&_if->then_list));
881 assert(nir_cf_list_is_empty_block(&_if->else_list));
882 before_cursor = false;
883 break;
884 }
885 assert(!before_cursor);
886
887 if (flatten_resume_if_ladder(b, &_if->cf_node, &_if->then_list,
888 false, resume_instr, remat)) {
889 resume_node = child;
890 rewrite_phis_to_pred(nir_cf_node_as_block(nir_cf_node_next(child)),
891 nir_if_last_then_block(_if));
892 goto found_resume;
893 }
894
895 if (flatten_resume_if_ladder(b, &_if->cf_node, &_if->else_list,
896 false, resume_instr, remat)) {
897 resume_node = child;
898 rewrite_phis_to_pred(nir_cf_node_as_block(nir_cf_node_next(child)),
899 nir_if_last_else_block(_if));
900 goto found_resume;
901 }
902 break;
903 }
904
905 case nir_cf_node_loop: {
906 assert(!before_cursor);
907 nir_loop *loop = nir_cf_node_as_loop(child);
908
909 if (cf_node_contains_block(&loop->cf_node, resume_instr->block)) {
910 /* Thanks to our loop body duplication pass, every level of loop
911 * containing the resume instruction contains exactly three nodes:
912 * two blocks and an if. We don't want to lower away this if
913 * because it's the resume selection if. The resume half is
914 * always the then_list so that's what we want to flatten.
915 */
916 nir_block *header = nir_loop_first_block(loop);
917 nir_if *_if = nir_cf_node_as_if(nir_cf_node_next(&header->cf_node));
918
919 nir_builder bl;
920 nir_builder_init(&bl, b->impl);
921 bl.cursor = nir_before_cf_list(&_if->then_list);
922 /* We want to place anything re-materialized from inside the loop
923 * at the top of the resume half of the loop.
924 *
925 * Because we're inside a loop, we might run into a break/continue
926 * instructions. We can't place those within a block of
927 * instructions, they need to be at the end of a block. So we
928 * build our own dummy block to place them.
929 */
930 nir_push_if(&bl, nir_imm_true(&bl));
931 {
932 ASSERTED bool found =
933 flatten_resume_if_ladder(&bl, &_if->cf_node, &_if->then_list,
934 true, resume_instr, remat);
935 assert(found);
936 }
937 nir_pop_if(&bl, NULL);
938
939 resume_node = child;
940 goto found_resume;
941 } else {
942 ASSERTED bool found =
943 flatten_resume_if_ladder(b, &loop->cf_node, &loop->body,
944 false, resume_instr, remat);
945 assert(!found);
946 }
947 break;
948 }
949
950 case nir_cf_node_function:
951 unreachable("Unsupported CF node type");
952 }
953 }
954 assert(!before_cursor);
955
956 /* If we got here, we didn't find the resume node or instruction. */
957 return false;
958
959 found_resume:
960 /* If we got here then we found either the resume node or the resume
961 * instruction in this CF list.
962 */
963 if (resume_node) {
964 /* If the resume instruction is buried in side one of our children CF
965 * nodes, resume_node now points to that child.
966 */
967 if (resume_node->type == nir_cf_node_if) {
968 /* Thanks to the recursive call, all of the interesting contents of
969 * resume_node have been copied before the cursor. We just need to
970 * copy the stuff after resume_node.
971 */
972 nir_cf_extract(&cf_list, nir_after_cf_node(resume_node),
973 nir_after_cf_list(child_list));
974 } else {
975 /* The loop contains its own cursor and still has useful stuff in it.
976 * We want to move everything after and including the loop to before
977 * the cursor.
978 */
979 assert(resume_node->type == nir_cf_node_loop);
980 nir_cf_extract(&cf_list, nir_before_cf_node(resume_node),
981 nir_after_cf_list(child_list));
982 }
983 } else {
984 /* If we found the resume instruction in one of our blocks, grab
985 * everything after it in the entire list (not just the one block), and
986 * place it before the cursor instr.
987 */
988 nir_cf_extract(&cf_list, nir_after_instr(resume_instr),
989 nir_after_cf_list(child_list));
990 }
991
992 if (cursor_is_after_jump(b->cursor)) {
993 /* If the resume instruction is in a loop, it's possible cf_list ends
994 * in a break or continue instruction, in which case we don't want to
995 * insert anything. It's also possible we have an early return if
996 * someone hasn't lowered those yet. In either case, nothing after that
997 * point executes in this context so we can delete it.
998 */
999 nir_cf_delete(&cf_list);
1000 } else {
1001 b->cursor = nir_cf_reinsert(&cf_list, b->cursor);
1002 }
1003
1004 if (!resume_node) {
1005 /* We want the resume to be the first "interesting" instruction */
1006 nir_instr_remove(resume_instr);
1007 nir_instr_insert(nir_before_cf_list(&b->impl->body), resume_instr);
1008 }
1009
1010 /* We've copied everything interesting out of this CF list to before the
1011 * cursor. Delete everything else.
1012 */
1013 if (child_list_contains_cursor) {
1014 /* If the cursor is in child_list, then we're either a loop or function
1015 * that contains the cursor. Cursors are always placed in a wrapper if
1016 * (true) to deal with break/continue and early returns. We've already
1017 * moved everything interesting inside the wrapper if and we want to
1018 * remove whatever is left after it.
1019 */
1020 nir_block *cursor_block = nir_cursor_current_block(b->cursor);
1021 nir_if *wrapper_if = nir_cf_node_as_if(cursor_block->cf_node.parent);
1022 assert(wrapper_if->cf_node.parent == parent_node);
1023 /* The wrapper if blocks are either put into the body of the main
1024 * function, or within the resume if block of the loops.
1025 */
1026 assert(parent_node->type == nir_cf_node_function ||
1027 (parent_node->type == nir_cf_node_if &&
1028 parent_node->parent->type == nir_cf_node_loop));
1029 nir_cf_extract(&cf_list, nir_after_cf_node(&wrapper_if->cf_node),
1030 nir_after_cf_list(child_list));
1031 } else {
1032 nir_cf_list_extract(&cf_list, child_list);
1033 }
1034 nir_cf_delete(&cf_list);
1035
1036 return true;
1037 }
1038
1039 static nir_instr *
lower_resume(nir_shader * shader,int call_idx)1040 lower_resume(nir_shader *shader, int call_idx)
1041 {
1042 nir_function_impl *impl = nir_shader_get_entrypoint(shader);
1043
1044 nir_instr *resume_instr = find_resume_instr(impl, call_idx);
1045
1046 if (duplicate_loop_bodies(impl, resume_instr)) {
1047 nir_validate_shader(shader, "after duplicate_loop_bodies in "
1048 "brw_nir_lower_shader_calls");
1049 /* If we duplicated the bodies of any loops, run regs_to_ssa to get rid
1050 * of all those pesky registers we just added.
1051 */
1052 NIR_PASS_V(shader, nir_lower_regs_to_ssa);
1053 }
1054
1055 /* Re-index nir_ssa_def::index. We don't care about actual liveness in
1056 * this pass but, so we can use the same helpers as the spilling pass, we
1057 * need to make sure that live_index is something sane. It's used
1058 * constantly for determining if an SSA value has been added since the
1059 * start of the pass.
1060 */
1061 nir_index_ssa_defs(impl);
1062
1063 void *mem_ctx = ralloc_context(shader);
1064
1065 /* Used to track which things may have been assumed to be re-materialized
1066 * by the spilling pass and which we shouldn't delete.
1067 */
1068 struct brw_bitset remat = bitset_create(mem_ctx, impl->ssa_alloc);
1069
1070 /* Create a nop instruction to use as a cursor as we extract and re-insert
1071 * stuff into the CFG.
1072 */
1073 nir_builder b;
1074 nir_builder_init(&b, impl);
1075 b.cursor = nir_before_cf_list(&impl->body);
1076
1077 nir_push_if(&b, nir_imm_true(&b));
1078 {
1079 ASSERTED bool found =
1080 flatten_resume_if_ladder(&b, &impl->cf_node, &impl->body,
1081 true, resume_instr, &remat);
1082 assert(found);
1083 }
1084 nir_pop_if(&b, NULL);
1085
1086 ralloc_free(mem_ctx);
1087
1088 nir_validate_shader(shader, "after flatten_resume_if_ladder in "
1089 "brw_nir_lower_shader_calls");
1090
1091 nir_metadata_preserve(impl, nir_metadata_none);
1092
1093 return resume_instr;
1094 }
1095
1096 static void
replace_resume_with_halt(nir_shader * shader,nir_instr * keep)1097 replace_resume_with_halt(nir_shader *shader, nir_instr *keep)
1098 {
1099 nir_function_impl *impl = nir_shader_get_entrypoint(shader);
1100
1101 nir_builder b;
1102 nir_builder_init(&b, impl);
1103
1104 nir_foreach_block_safe(block, impl) {
1105 nir_foreach_instr_safe(instr, block) {
1106 if (instr == keep)
1107 continue;
1108
1109 if (instr->type != nir_instr_type_intrinsic)
1110 continue;
1111
1112 nir_intrinsic_instr *resume = nir_instr_as_intrinsic(instr);
1113 if (resume->intrinsic != nir_intrinsic_rt_resume)
1114 continue;
1115
1116 /* If this is some other resume, then we've kicked off a ray or
1117 * bindless thread and we don't want to go any further in this
1118 * shader. Insert a halt so that NIR will delete any instructions
1119 * dominated by this call instruction including the scratch_load
1120 * instructions we inserted.
1121 */
1122 nir_cf_list cf_list;
1123 nir_cf_extract(&cf_list, nir_after_instr(&resume->instr),
1124 nir_after_block(block));
1125 nir_cf_delete(&cf_list);
1126 b.cursor = nir_instr_remove(&resume->instr);
1127 nir_jump(&b, nir_jump_halt);
1128 break;
1129 }
1130 }
1131 }
1132
1133 /** Lower shader call instructions to split shaders.
1134 *
1135 * Shader calls can be split into an initial shader and a series of "resume"
1136 * shaders. When the shader is first invoked, it is the initial shader which
1137 * is executed. At any point in the initial shader or any one of the resume
1138 * shaders, a shader call operation may be performed. The possible shader call
1139 * operations are:
1140 *
1141 * - trace_ray
1142 * - report_ray_intersection
1143 * - execute_callable
1144 *
1145 * When a shader call operation is performed, we push all live values to the
1146 * stack,call rt_trace_ray/rt_execute_callable and then kill the shader. Once
1147 * the operation we invoked is complete, a callee shader will return execution
1148 * to the respective resume shader. The resume shader pops the contents off
1149 * the stack and picks up where the calling shader left off.
1150 *
1151 * Stack management is assumed to be done after this pass. Call
1152 * instructions and their resumes get annotated with stack information that
1153 * should be enough for the backend to implement proper stack management.
1154 */
1155 bool
nir_lower_shader_calls(nir_shader * shader,nir_address_format address_format,unsigned stack_alignment,nir_shader *** resume_shaders_out,uint32_t * num_resume_shaders_out,void * mem_ctx)1156 nir_lower_shader_calls(nir_shader *shader,
1157 nir_address_format address_format,
1158 unsigned stack_alignment,
1159 nir_shader ***resume_shaders_out,
1160 uint32_t *num_resume_shaders_out,
1161 void *mem_ctx)
1162 {
1163 nir_function_impl *impl = nir_shader_get_entrypoint(shader);
1164
1165 nir_builder b;
1166 nir_builder_init(&b, impl);
1167
1168 int num_calls = 0;
1169 nir_foreach_block(block, impl) {
1170 nir_foreach_instr_safe(instr, block) {
1171 if (instr_is_shader_call(instr))
1172 num_calls++;
1173 }
1174 }
1175
1176 if (num_calls == 0) {
1177 nir_shader_preserve_all_metadata(shader);
1178 *num_resume_shaders_out = 0;
1179 return false;
1180 }
1181
1182 /* Some intrinsics not only can't be re-materialized but aren't preserved
1183 * when moving to the continuation shader. We have to move them to the top
1184 * to ensure they get spilled as needed.
1185 */
1186 {
1187 bool progress = false;
1188 NIR_PASS(progress, shader, move_system_values_to_top);
1189 if (progress)
1190 NIR_PASS(progress, shader, nir_opt_cse);
1191 }
1192
1193 NIR_PASS_V(shader, spill_ssa_defs_and_lower_shader_calls,
1194 num_calls, address_format, stack_alignment);
1195
1196 nir_opt_remove_phis(shader);
1197
1198 /* Make N copies of our shader */
1199 nir_shader **resume_shaders = ralloc_array(mem_ctx, nir_shader *, num_calls);
1200 for (unsigned i = 0; i < num_calls; i++) {
1201 resume_shaders[i] = nir_shader_clone(mem_ctx, shader);
1202
1203 /* Give them a recognizable name */
1204 resume_shaders[i]->info.name =
1205 ralloc_asprintf(mem_ctx, "%s%sresume_%u",
1206 shader->info.name ? shader->info.name : "",
1207 shader->info.name ? "-" : "",
1208 i);
1209 }
1210
1211 replace_resume_with_halt(shader, NULL);
1212 for (unsigned i = 0; i < num_calls; i++) {
1213 nir_instr *resume_instr = lower_resume(resume_shaders[i], i);
1214 replace_resume_with_halt(resume_shaders[i], resume_instr);
1215 nir_opt_remove_phis(resume_shaders[i]);
1216 /* Remove the dummy blocks added by flatten_resume_if_ladder() */
1217 nir_opt_if(resume_shaders[i], nir_opt_if_optimize_phi_true_false);
1218 }
1219
1220 *resume_shaders_out = resume_shaders;
1221 *num_resume_shaders_out = num_calls;
1222
1223 return true;
1224 }
1225