2 * Copyright © 2016 Intel Corporation
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:
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
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
21 * DEALINGS IN THE SOFTWARE.
25 #include "nir_builder.h"
26 #include "nir_control_flow.h"
27 #include "nir_loop_analyze.h"
30 /* This limit is chosen fairly arbitrarily. GLSL IR max iteration is 32
31 * instructions. (Multiply counting nodes and magic number 5.) But there is
32 * no 1:1 mapping between GLSL IR and NIR so 25 was picked because it seemed
33 * to give about the same results. Around 5 instructions per node. But some
34 * loops that would unroll with GLSL IR fail to unroll if we set this to 25 so
37 #define LOOP_UNROLL_LIMIT 26
39 /* Prepare this loop for unrolling by first converting to lcssa and then
40 * converting the phis from the top level of the loop body to regs.
41 * Partially converting out of SSA allows us to unroll the loop without having
42 * to keep track of and update phis along the way which gets tricky and
43 * doesn't add much value over converting to regs.
45 * The loop may have a continue instruction at the end of the loop which does
46 * nothing. Once we're out of SSA, we can safely delete it so we don't have
47 * to deal with it later.
50 loop_prepare_for_unroll(nir_loop
*loop
)
52 nir_rematerialize_derefs_in_use_blocks_impl(
53 nir_cf_node_get_function(&loop
->cf_node
));
55 nir_convert_loop_to_lcssa(loop
);
57 /* Lower phis at the top level of the loop body */
58 foreach_list_typed_safe(nir_cf_node
, node
, node
, &loop
->body
) {
59 if (nir_cf_node_block
== node
->type
) {
60 nir_lower_phis_to_regs_block(nir_cf_node_as_block(node
));
64 /* Lower phis after the loop */
65 nir_block
*block_after_loop
=
66 nir_cf_node_as_block(nir_cf_node_next(&loop
->cf_node
));
68 nir_lower_phis_to_regs_block(block_after_loop
);
70 /* Remove continue if its the last instruction in the loop */
71 nir_instr
*last_instr
= nir_block_last_instr(nir_loop_last_block(loop
));
72 if (last_instr
&& last_instr
->type
== nir_instr_type_jump
) {
73 nir_instr_remove(last_instr
);
78 get_first_blocks_in_terminator(nir_loop_terminator
*term
,
79 nir_block
**first_break_block
,
80 nir_block
**first_continue_block
)
82 if (term
->continue_from_then
) {
83 *first_continue_block
= nir_if_first_then_block(term
->nif
);
84 *first_break_block
= nir_if_first_else_block(term
->nif
);
86 *first_continue_block
= nir_if_first_else_block(term
->nif
);
87 *first_break_block
= nir_if_first_then_block(term
->nif
);
92 * Unroll a loop where we know exactly how many iterations there are and there
93 * is only a single exit point. Note here we can unroll loops with multiple
94 * theoretical exits that only have a single terminating exit that we always
95 * know is the "real" exit.
101 * And the iteration count is 3, the output will be:
103 * ...instrs... ...instrs... ...instrs...
106 simple_unroll(nir_loop
*loop
)
108 nir_loop_terminator
*limiting_term
= loop
->info
->limiting_terminator
;
109 assert(nir_is_trivial_loop_if(limiting_term
->nif
,
110 limiting_term
->break_block
));
112 loop_prepare_for_unroll(loop
);
114 /* Skip over loop terminator and get the loop body. */
115 list_for_each_entry(nir_loop_terminator
, terminator
,
116 &loop
->info
->loop_terminator_list
,
117 loop_terminator_link
) {
119 /* Remove all but the limiting terminator as we know the other exit
120 * conditions can never be met. Note we need to extract any instructions
121 * in the continue from branch and insert then into the loop body before
124 if (terminator
->nif
!= limiting_term
->nif
) {
125 nir_block
*first_break_block
;
126 nir_block
*first_continue_block
;
127 get_first_blocks_in_terminator(terminator
, &first_break_block
,
128 &first_continue_block
);
130 assert(nir_is_trivial_loop_if(terminator
->nif
,
131 terminator
->break_block
));
133 nir_cf_list continue_from_lst
;
134 nir_cf_extract(&continue_from_lst
,
135 nir_before_block(first_continue_block
),
136 nir_after_block(terminator
->continue_from_block
));
137 nir_cf_reinsert(&continue_from_lst
,
138 nir_after_cf_node(&terminator
->nif
->cf_node
));
140 nir_cf_node_remove(&terminator
->nif
->cf_node
);
144 nir_block
*first_break_block
;
145 nir_block
*first_continue_block
;
146 get_first_blocks_in_terminator(limiting_term
, &first_break_block
,
147 &first_continue_block
);
149 /* Pluck out the loop header */
150 nir_block
*header_blk
= nir_loop_first_block(loop
);
151 nir_cf_list lp_header
;
152 nir_cf_extract(&lp_header
, nir_before_block(header_blk
),
153 nir_before_cf_node(&limiting_term
->nif
->cf_node
));
155 /* Add the continue from block of the limiting terminator to the loop body
157 nir_cf_list continue_from_lst
;
158 nir_cf_extract(&continue_from_lst
, nir_before_block(first_continue_block
),
159 nir_after_block(limiting_term
->continue_from_block
));
160 nir_cf_reinsert(&continue_from_lst
,
161 nir_after_cf_node(&limiting_term
->nif
->cf_node
));
163 /* Pluck out the loop body */
164 nir_cf_list loop_body
;
165 nir_cf_extract(&loop_body
, nir_after_cf_node(&limiting_term
->nif
->cf_node
),
166 nir_after_block(nir_loop_last_block(loop
)));
168 struct hash_table
*remap_table
=
169 _mesa_hash_table_create(NULL
, _mesa_hash_pointer
,
170 _mesa_key_pointer_equal
);
172 /* Clone the loop header and insert before the loop */
173 nir_cf_list_clone_and_reinsert(&lp_header
, loop
->cf_node
.parent
,
174 nir_before_cf_node(&loop
->cf_node
),
177 for (unsigned i
= 0; i
< loop
->info
->max_trip_count
; i
++) {
178 /* Clone loop body and insert before the loop */
179 nir_cf_list_clone_and_reinsert(&loop_body
, loop
->cf_node
.parent
,
180 nir_before_cf_node(&loop
->cf_node
),
183 /* Clone loop header and insert after loop body */
184 nir_cf_list_clone_and_reinsert(&lp_header
, loop
->cf_node
.parent
,
185 nir_before_cf_node(&loop
->cf_node
),
189 /* Remove the break from the loop terminator and add instructions from
190 * the break block after the unrolled loop.
192 nir_instr
*break_instr
= nir_block_last_instr(limiting_term
->break_block
);
193 nir_instr_remove(break_instr
);
194 nir_cf_list break_list
;
195 nir_cf_extract(&break_list
, nir_before_block(first_break_block
),
196 nir_after_block(limiting_term
->break_block
));
198 /* Clone so things get properly remapped */
199 nir_cf_list_clone_and_reinsert(&break_list
, loop
->cf_node
.parent
,
200 nir_before_cf_node(&loop
->cf_node
),
203 /* Remove the loop */
204 nir_cf_node_remove(&loop
->cf_node
);
206 /* Delete the original loop body, break block & header */
207 nir_cf_delete(&lp_header
);
208 nir_cf_delete(&loop_body
);
209 nir_cf_delete(&break_list
);
211 _mesa_hash_table_destroy(remap_table
, NULL
);
215 move_cf_list_into_loop_term(nir_cf_list
*lst
, nir_loop_terminator
*term
)
217 /* Move the rest of the loop inside the continue-from-block */
218 nir_cf_reinsert(lst
, nir_after_block(term
->continue_from_block
));
220 /* Remove the break */
221 nir_instr_remove(nir_block_last_instr(term
->break_block
));
225 get_complex_unroll_insert_location(nir_cf_node
*node
, bool continue_from_then
)
227 if (node
->type
== nir_cf_node_loop
) {
228 return nir_before_cf_node(node
);
230 nir_if
*if_stmt
= nir_cf_node_as_if(node
);
231 if (continue_from_then
) {
232 return nir_after_block(nir_if_last_then_block(if_stmt
));
234 return nir_after_block(nir_if_last_else_block(if_stmt
));
240 complex_unroll_loop_body(nir_loop
*loop
, nir_loop_terminator
*unlimit_term
,
241 nir_cf_list
*lp_header
, nir_cf_list
*lp_body
,
242 struct hash_table
*remap_table
,
243 unsigned num_times_to_clone
)
245 /* In the terminator that we have no trip count for move everything after
246 * the terminator into the continue from branch.
248 nir_cf_list loop_end
;
249 nir_cf_extract(&loop_end
, nir_after_cf_node(&unlimit_term
->nif
->cf_node
),
250 nir_after_block(nir_loop_last_block(loop
)));
251 move_cf_list_into_loop_term(&loop_end
, unlimit_term
);
253 /* Pluck out the loop body. */
254 nir_cf_extract(lp_body
, nir_before_block(nir_loop_first_block(loop
)),
255 nir_after_block(nir_loop_last_block(loop
)));
257 /* Set unroll_loc to the loop as we will insert the unrolled loop before it
259 nir_cf_node
*unroll_loc
= &loop
->cf_node
;
261 /* Temp list to store the cloned loop as we unroll */
262 nir_cf_list unrolled_lp_body
;
264 for (unsigned i
= 0; i
< num_times_to_clone
; i
++) {
267 get_complex_unroll_insert_location(unroll_loc
,
268 unlimit_term
->continue_from_then
);
270 /* Clone loop header and insert in if branch */
271 nir_cf_list_clone_and_reinsert(lp_header
, loop
->cf_node
.parent
,
272 cursor
, remap_table
);
275 get_complex_unroll_insert_location(unroll_loc
,
276 unlimit_term
->continue_from_then
);
278 /* Clone loop body */
279 nir_cf_list_clone(&unrolled_lp_body
, lp_body
, loop
->cf_node
.parent
,
282 unroll_loc
= exec_node_data(nir_cf_node
,
283 exec_list_get_tail(&unrolled_lp_body
.list
),
285 assert(unroll_loc
->type
== nir_cf_node_block
&&
286 exec_list_is_empty(&nir_cf_node_as_block(unroll_loc
)->instr_list
));
288 /* Get the unrolled if node */
289 unroll_loc
= nir_cf_node_prev(unroll_loc
);
291 /* Insert unrolled loop body */
292 nir_cf_reinsert(&unrolled_lp_body
, cursor
);
299 * Unroll a loop with two exists when the trip count of one of the exits is
300 * unknown. If continue_from_then is true, the loop is repeated only when the
301 * "then" branch of the if is taken; otherwise it is repeated only
302 * when the "else" branch of the if is taken.
304 * For example, if the input is:
307 * ...phis/condition...
309 * ...then instructions...
311 * ...continue instructions...
317 * And the iteration count is 3, and unlimit_term->continue_from_then is true,
318 * then the output will be:
322 * ...then instructions...
325 * ...then instructions...
328 * ...then instructions...
331 * ...continue instructions...
334 * ...continue instructions...
337 * ...continue instructions...
341 complex_unroll(nir_loop
*loop
, nir_loop_terminator
*unlimit_term
,
342 bool limiting_term_second
)
344 assert(nir_is_trivial_loop_if(unlimit_term
->nif
,
345 unlimit_term
->break_block
));
347 nir_loop_terminator
*limiting_term
= loop
->info
->limiting_terminator
;
348 assert(nir_is_trivial_loop_if(limiting_term
->nif
,
349 limiting_term
->break_block
));
351 loop_prepare_for_unroll(loop
);
353 nir_block
*header_blk
= nir_loop_first_block(loop
);
355 nir_cf_list lp_header
;
356 nir_cf_list limit_break_list
;
357 unsigned num_times_to_clone
;
358 if (limiting_term_second
) {
359 /* Pluck out the loop header */
360 nir_cf_extract(&lp_header
, nir_before_block(header_blk
),
361 nir_before_cf_node(&unlimit_term
->nif
->cf_node
));
363 /* We need some special handling when its the second terminator causing
364 * us to exit the loop for example:
366 * for (int i = 0; i < uniform_lp_count; i++) {
367 * colour = vec4(0.0, 1.0, 0.0, 1.0);
372 * ... any further code is unreachable after i == 1 ...
375 nir_cf_list after_lt
;
376 nir_if
*limit_if
= limiting_term
->nif
;
377 nir_cf_extract(&after_lt
, nir_after_cf_node(&limit_if
->cf_node
),
378 nir_after_block(nir_loop_last_block(loop
)));
379 move_cf_list_into_loop_term(&after_lt
, limiting_term
);
381 /* Because the trip count is the number of times we pass over the entire
382 * loop before hitting a break when the second terminator is the
383 * limiting terminator we can actually execute code inside the loop when
384 * trip count == 0 e.g. the code above the break. So we need to bump
385 * the trip_count in order for the code below to clone anything. When
386 * trip count == 1 we execute the code above the break twice and the
387 * code below it once so we need clone things twice and so on.
389 num_times_to_clone
= loop
->info
->max_trip_count
+ 1;
391 /* Pluck out the loop header */
392 nir_cf_extract(&lp_header
, nir_before_block(header_blk
),
393 nir_before_cf_node(&limiting_term
->nif
->cf_node
));
395 nir_block
*first_break_block
;
396 nir_block
*first_continue_block
;
397 get_first_blocks_in_terminator(limiting_term
, &first_break_block
,
398 &first_continue_block
);
400 /* Remove the break then extract instructions from the break block so we
401 * can insert them in the innermost else of the unrolled loop.
403 nir_instr
*break_instr
= nir_block_last_instr(limiting_term
->break_block
);
404 nir_instr_remove(break_instr
);
405 nir_cf_extract(&limit_break_list
, nir_before_block(first_break_block
),
406 nir_after_block(limiting_term
->break_block
));
408 nir_cf_list continue_list
;
409 nir_cf_extract(&continue_list
, nir_before_block(first_continue_block
),
410 nir_after_block(limiting_term
->continue_from_block
));
412 nir_cf_reinsert(&continue_list
,
413 nir_after_cf_node(&limiting_term
->nif
->cf_node
));
415 nir_cf_node_remove(&limiting_term
->nif
->cf_node
);
417 num_times_to_clone
= loop
->info
->max_trip_count
;
420 struct hash_table
*remap_table
=
421 _mesa_hash_table_create(NULL
, _mesa_hash_pointer
,
422 _mesa_key_pointer_equal
);
425 nir_cf_node
*unroll_loc
=
426 complex_unroll_loop_body(loop
, unlimit_term
, &lp_header
, &lp_body
,
427 remap_table
, num_times_to_clone
);
429 if (!limiting_term_second
) {
430 assert(unroll_loc
->type
== nir_cf_node_if
);
433 get_complex_unroll_insert_location(unroll_loc
,
434 unlimit_term
->continue_from_then
);
436 /* Clone loop header and insert in if branch */
437 nir_cf_list_clone_and_reinsert(&lp_header
, loop
->cf_node
.parent
,
438 cursor
, remap_table
);
441 get_complex_unroll_insert_location(unroll_loc
,
442 unlimit_term
->continue_from_then
);
444 /* Clone so things get properly remapped, and insert break block from
445 * the limiting terminator.
447 nir_cf_list_clone_and_reinsert(&limit_break_list
, loop
->cf_node
.parent
,
448 cursor
, remap_table
);
450 nir_cf_delete(&limit_break_list
);
453 /* The loop has been unrolled so remove it. */
454 nir_cf_node_remove(&loop
->cf_node
);
456 /* Delete the original loop header and body */
457 nir_cf_delete(&lp_header
);
458 nir_cf_delete(&lp_body
);
460 _mesa_hash_table_destroy(remap_table
, NULL
);
463 /* Unrolls the classic wrapper loops e.g
470 wrapper_unroll(nir_loop
*loop
)
472 if (!list_empty(&loop
->info
->loop_terminator_list
)) {
474 /* Unrolling a loop with a large number of exits can result in a
475 * large inrease in register pressure. For now we just skip
476 * unrolling if we have more than 3 exits (not including the break
477 * at the end of the loop).
479 * TODO: Most loops that fit this pattern are simply switch
480 * statements that are converted to a loop to take advantage of
481 * exiting jump instruction handling. In this case we could make
482 * use of a binary seach pattern like we do in
483 * nir_lower_indirect_derefs(), this should allow us to unroll the
484 * loops in an optimal way and should also avoid some of the
485 * register pressure that comes from simply nesting the
486 * terminators one after the other.
488 if (list_length(&loop
->info
->loop_terminator_list
) > 3)
491 loop_prepare_for_unroll(loop
);
493 nir_cursor loop_end
= nir_after_block(nir_loop_last_block(loop
));
494 list_for_each_entry(nir_loop_terminator
, terminator
,
495 &loop
->info
->loop_terminator_list
,
496 loop_terminator_link
) {
498 /* Remove break from the terminator */
499 nir_instr
*break_instr
=
500 nir_block_last_instr(terminator
->break_block
);
501 nir_instr_remove(break_instr
);
503 /* Pluck out the loop body. */
504 nir_cf_list loop_body
;
505 nir_cf_extract(&loop_body
,
506 nir_after_cf_node(&terminator
->nif
->cf_node
),
509 /* Reinsert loop body into continue from block */
510 nir_cf_reinsert(&loop_body
,
511 nir_after_block(terminator
->continue_from_block
));
513 loop_end
= terminator
->continue_from_then
?
514 nir_after_block(nir_if_last_then_block(terminator
->nif
)) :
515 nir_after_block(nir_if_last_else_block(terminator
->nif
));
518 nir_block
*blk_after_loop
=
519 nir_cursor_current_block(nir_after_cf_node(&loop
->cf_node
));
521 /* There may still be some single src phis following the loop that
522 * have not yet been cleaned up by another pass. Tidy those up
523 * before unrolling the loop.
525 nir_foreach_instr_safe(instr
, blk_after_loop
) {
526 if (instr
->type
!= nir_instr_type_phi
)
529 nir_phi_instr
*phi
= nir_instr_as_phi(instr
);
530 assert(exec_list_length(&phi
->srcs
) == 1);
532 nir_phi_src
*phi_src
=
533 exec_node_data(nir_phi_src
, exec_list_get_head(&phi
->srcs
), node
);
535 nir_ssa_def_rewrite_uses(&phi
->dest
.ssa
, phi_src
->src
);
536 nir_instr_remove(instr
);
539 /* Remove break at end of the loop */
540 nir_block
*last_loop_blk
= nir_loop_last_block(loop
);
541 nir_instr
*break_instr
= nir_block_last_instr(last_loop_blk
);
542 nir_instr_remove(break_instr
);
545 /* Pluck out the loop body. */
546 nir_cf_list loop_body
;
547 nir_cf_extract(&loop_body
, nir_before_block(nir_loop_first_block(loop
)),
548 nir_after_block(nir_loop_last_block(loop
)));
550 /* Reinsert loop body after the loop */
551 nir_cf_reinsert(&loop_body
, nir_after_cf_node(&loop
->cf_node
));
553 /* The loop has been unrolled so remove it. */
554 nir_cf_node_remove(&loop
->cf_node
);
560 is_loop_small_enough_to_unroll(nir_shader
*shader
, nir_loop_info
*li
)
562 unsigned max_iter
= shader
->options
->max_unroll_iterations
;
564 if (li
->max_trip_count
> max_iter
)
567 if (li
->force_unroll
)
570 bool loop_not_too_large
=
571 li
->num_instructions
* li
->max_trip_count
<= max_iter
* LOOP_UNROLL_LIMIT
;
573 return loop_not_too_large
;
577 process_loops(nir_shader
*sh
, nir_cf_node
*cf_node
, bool *has_nested_loop_out
)
579 bool progress
= false;
580 bool has_nested_loop
= false;
583 switch (cf_node
->type
) {
584 case nir_cf_node_block
:
586 case nir_cf_node_if
: {
587 nir_if
*if_stmt
= nir_cf_node_as_if(cf_node
);
588 foreach_list_typed_safe(nir_cf_node
, nested_node
, node
, &if_stmt
->then_list
)
589 progress
|= process_loops(sh
, nested_node
, has_nested_loop_out
);
590 foreach_list_typed_safe(nir_cf_node
, nested_node
, node
, &if_stmt
->else_list
)
591 progress
|= process_loops(sh
, nested_node
, has_nested_loop_out
);
594 case nir_cf_node_loop
: {
595 loop
= nir_cf_node_as_loop(cf_node
);
596 foreach_list_typed_safe(nir_cf_node
, nested_node
, node
, &loop
->body
)
597 progress
|= process_loops(sh
, nested_node
, &has_nested_loop
);
602 unreachable("unknown cf node type");
605 /* Don't attempt to unroll a second inner loop in this pass, wait until the
606 * next pass as we have altered the cf.
610 /* Check for the classic
616 * that is used to wrap multi-line macros. GLSL IR also wraps switch
617 * statements in a loop like this.
619 if (loop
->info
->limiting_terminator
== NULL
&&
620 !loop
->info
->complex_loop
) {
622 nir_block
*last_loop_blk
= nir_loop_last_block(loop
);
623 if (!nir_block_ends_in_break(last_loop_blk
))
626 progress
= wrapper_unroll(loop
);
631 if (has_nested_loop
|| loop
->info
->limiting_terminator
== NULL
)
634 if (!is_loop_small_enough_to_unroll(sh
, loop
->info
))
637 if (loop
->info
->exact_trip_count_known
) {
641 /* Attempt to unroll loops with two terminators. */
642 unsigned num_lt
= list_length(&loop
->info
->loop_terminator_list
);
644 bool limiting_term_second
= true;
645 nir_loop_terminator
*terminator
=
646 list_first_entry(&loop
->info
->loop_terminator_list
,
647 nir_loop_terminator
, loop_terminator_link
);
650 if (terminator
->nif
== loop
->info
->limiting_terminator
->nif
) {
651 limiting_term_second
= false;
653 list_last_entry(&loop
->info
->loop_terminator_list
,
654 nir_loop_terminator
, loop_terminator_link
);
657 /* If the first terminator has a trip count of zero and is the
658 * limiting terminator just do a simple unroll as the second
659 * terminator can never be reached.
661 if (loop
->info
->max_trip_count
== 0 && !limiting_term_second
) {
664 complex_unroll(loop
, terminator
, limiting_term_second
);
672 *has_nested_loop_out
= true;
677 nir_opt_loop_unroll_impl(nir_function_impl
*impl
,
678 nir_variable_mode indirect_mask
)
680 bool progress
= false;
681 nir_metadata_require(impl
, nir_metadata_loop_analysis
, indirect_mask
);
682 nir_metadata_require(impl
, nir_metadata_block_index
);
684 foreach_list_typed_safe(nir_cf_node
, node
, node
, &impl
->body
) {
685 bool has_nested_loop
= false;
686 progress
|= process_loops(impl
->function
->shader
, node
,
691 nir_lower_regs_to_ssa_impl(impl
);
697 * indirect_mask specifies which type of indirectly accessed variables
698 * should force loop unrolling.
701 nir_opt_loop_unroll(nir_shader
*shader
, nir_variable_mode indirect_mask
)
703 bool progress
= false;
705 nir_foreach_function(function
, shader
) {
706 if (function
->impl
) {
707 progress
|= nir_opt_loop_unroll_impl(function
->impl
, indirect_mask
);