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 DEALINGS
25 #include "nir/nir_builder.h"
26 #include "nir_constant_expressions.h"
27 #include "nir_control_flow.h"
28 #include "nir_loop_analyze.h"
31 * Gets the single block that jumps back to the loop header. Already assumes
32 * there is exactly one such block.
35 find_continue_block(nir_loop
*loop
)
37 nir_block
*header_block
= nir_loop_first_block(loop
);
38 nir_block
*prev_block
=
39 nir_cf_node_as_block(nir_cf_node_prev(&loop
->cf_node
));
41 assert(header_block
->predecessors
->entries
== 2);
43 set_foreach(header_block
->predecessors
, pred_entry
) {
44 if (pred_entry
->key
!= prev_block
)
45 return (nir_block
*)pred_entry
->key
;
48 unreachable("Continue block not found!");
52 * This optimization detects if statements at the tops of loops where the
53 * condition is a phi node of two constants and moves half of the if to above
54 * the loop and the other half of the if to the end of the loop. A simple for
55 * loop "for (int i = 0; i < 4; i++)", when run through the SPIR-V front-end,
56 * ends up looking something like this:
58 * vec1 32 ssa_0 = load_const (0x00000000)
59 * vec1 32 ssa_1 = load_const (0xffffffff)
62 * vec1 32 ssa_2 = phi block_0: ssa_0, block_7: ssa_5
63 * vec1 32 ssa_3 = phi block_0: ssa_0, block_7: ssa_1
66 * vec1 32 ssa_4 = load_const (0x00000001)
67 * vec1 32 ssa_5 = iadd ssa_2, ssa_4
72 * vec1 32 ssa_6 = load_const (0x00000004)
73 * vec1 32 ssa_7 = ilt ssa_5, ssa_6
83 * This turns it into something like this:
85 * // Stuff from block 1
86 * // Stuff from block 3
89 * vec1 32 ssa_3 = phi block_0: ssa_0, block_7: ssa_1
90 * vec1 32 ssa_6 = load_const (0x00000004)
91 * vec1 32 ssa_7 = ilt ssa_5, ssa_6
99 * // Stuff from block 1
100 * // Stuff from block 2
101 * vec1 32 ssa_4 = load_const (0x00000001)
102 * vec1 32 ssa_5 = iadd ssa_2, ssa_4
106 opt_peel_loop_initial_if(nir_loop
*loop
)
108 nir_block
*header_block
= nir_loop_first_block(loop
);
109 MAYBE_UNUSED nir_block
*prev_block
=
110 nir_cf_node_as_block(nir_cf_node_prev(&loop
->cf_node
));
112 /* It would be insane if this were not true */
113 assert(_mesa_set_search(header_block
->predecessors
, prev_block
));
115 /* The loop must have exactly one continue block which could be a block
116 * ending in a continue instruction or the "natural" continue from the
117 * last block in the loop back to the top.
119 if (header_block
->predecessors
->entries
!= 2)
122 nir_block
*continue_block
= find_continue_block(loop
);
124 nir_cf_node
*if_node
= nir_cf_node_next(&header_block
->cf_node
);
125 if (!if_node
|| if_node
->type
!= nir_cf_node_if
)
128 nir_if
*nif
= nir_cf_node_as_if(if_node
);
129 assert(nif
->condition
.is_ssa
);
131 nir_ssa_def
*cond
= nif
->condition
.ssa
;
132 if (cond
->parent_instr
->type
!= nir_instr_type_phi
)
135 nir_phi_instr
*cond_phi
= nir_instr_as_phi(cond
->parent_instr
);
136 if (cond
->parent_instr
->block
!= header_block
)
139 /* We already know we have exactly one continue */
140 assert(exec_list_length(&cond_phi
->srcs
) == 2);
142 uint32_t entry_val
= 0, continue_val
= 0;
143 nir_foreach_phi_src(src
, cond_phi
) {
144 assert(src
->src
.is_ssa
);
145 nir_const_value
*const_src
= nir_src_as_const_value(src
->src
);
149 if (src
->pred
== continue_block
) {
150 continue_val
= const_src
->u32
[0];
152 assert(src
->pred
== prev_block
);
153 entry_val
= const_src
->u32
[0];
157 /* If they both execute or both don't execute, this is a job for
158 * nir_dead_cf, not this pass.
160 if ((entry_val
&& continue_val
) || (!entry_val
&& !continue_val
))
163 struct exec_list
*continue_list
, *entry_list
;
165 continue_list
= &nif
->then_list
;
166 entry_list
= &nif
->else_list
;
168 continue_list
= &nif
->else_list
;
169 entry_list
= &nif
->then_list
;
172 /* We want to be moving the contents of entry_list to above the loop so it
173 * can't contain any break or continue instructions.
175 foreach_list_typed(nir_cf_node
, cf_node
, node
, entry_list
) {
176 nir_foreach_block_in_cf_node(block
, cf_node
) {
177 nir_instr
*last_instr
= nir_block_last_instr(block
);
178 if (last_instr
&& last_instr
->type
== nir_instr_type_jump
)
183 /* We're about to re-arrange a bunch of blocks so make sure that we don't
184 * have deref uses which cross block boundaries. We don't want a deref
185 * accidentally ending up in a phi.
187 nir_rematerialize_derefs_in_use_blocks_impl(
188 nir_cf_node_get_function(&loop
->cf_node
));
190 /* Before we do anything, convert the loop to LCSSA. We're about to
191 * replace a bunch of SSA defs with registers and this will prevent any of
192 * it from leaking outside the loop.
194 nir_convert_loop_to_lcssa(loop
);
196 nir_block
*after_if_block
=
197 nir_cf_node_as_block(nir_cf_node_next(&nif
->cf_node
));
199 /* Get rid of phis in the header block since we will be duplicating it */
200 nir_lower_phis_to_regs_block(header_block
);
201 /* Get rid of phis after the if since dominance will change */
202 nir_lower_phis_to_regs_block(after_if_block
);
204 /* Get rid of SSA defs in the pieces we're about to move around */
205 nir_lower_ssa_defs_to_regs_block(header_block
);
206 nir_foreach_block_in_cf_node(block
, &nif
->cf_node
)
207 nir_lower_ssa_defs_to_regs_block(block
);
209 nir_cf_list header
, tmp
;
210 nir_cf_extract(&header
, nir_before_block(header_block
),
211 nir_after_block(header_block
));
213 nir_cf_list_clone(&tmp
, &header
, &loop
->cf_node
, NULL
);
214 nir_cf_reinsert(&tmp
, nir_before_cf_node(&loop
->cf_node
));
215 nir_cf_extract(&tmp
, nir_before_cf_list(entry_list
),
216 nir_after_cf_list(entry_list
));
217 nir_cf_reinsert(&tmp
, nir_before_cf_node(&loop
->cf_node
));
219 nir_cf_reinsert(&header
, nir_after_block_before_jump(continue_block
));
221 /* Get continue block again as the previous reinsert might have removed the block. */
222 continue_block
= find_continue_block(loop
);
224 nir_cf_extract(&tmp
, nir_before_cf_list(continue_list
),
225 nir_after_cf_list(continue_list
));
226 nir_cf_reinsert(&tmp
, nir_after_block_before_jump(continue_block
));
228 nir_cf_node_remove(&nif
->cf_node
);
234 is_block_empty(nir_block
*block
)
236 return nir_cf_node_is_last(&block
->cf_node
) &&
237 exec_list_is_empty(&block
->instr_list
);
241 nir_block_ends_in_continue(nir_block
*block
)
243 if (exec_list_is_empty(&block
->instr_list
))
246 nir_instr
*instr
= nir_block_last_instr(block
);
247 return instr
->type
== nir_instr_type_jump
&&
248 nir_instr_as_jump(instr
)->type
== nir_jump_continue
;
252 * This optimization turns:
276 * The continue should then be removed by nir_opt_trivial_continues() and the
277 * loop can potentially be unrolled.
279 * Note: do_work_2() is only ever blocks and nested loops. We could also nest
280 * other if-statments in the branch which would allow further continues to
281 * be removed. However in practice this can result in increased register
285 opt_if_loop_last_continue(nir_loop
*loop
)
287 /* Get the last if-stament in the loop */
288 nir_block
*last_block
= nir_loop_last_block(loop
);
289 nir_cf_node
*if_node
= nir_cf_node_prev(&last_block
->cf_node
);
291 if (if_node
->type
== nir_cf_node_if
)
294 if_node
= nir_cf_node_prev(if_node
);
297 if (!if_node
|| if_node
->type
!= nir_cf_node_if
)
300 nir_if
*nif
= nir_cf_node_as_if(if_node
);
301 nir_block
*then_block
= nir_if_last_then_block(nif
);
302 nir_block
*else_block
= nir_if_last_else_block(nif
);
304 bool then_ends_in_continue
= nir_block_ends_in_continue(then_block
);
305 bool else_ends_in_continue
= nir_block_ends_in_continue(else_block
);
307 /* If both branches end in a continue do nothing, this should be handled
308 * by nir_opt_dead_cf().
310 if (then_ends_in_continue
&& else_ends_in_continue
)
313 if (!then_ends_in_continue
&& !else_ends_in_continue
)
316 /* Move the last block of the loop inside the last if-statement */
318 nir_cf_extract(&tmp
, nir_after_cf_node(if_node
),
319 nir_after_block(last_block
));
320 if (then_ends_in_continue
) {
321 nir_cf_reinsert(&tmp
, nir_after_cf_list(&nif
->else_list
));
323 nir_cf_reinsert(&tmp
, nir_after_cf_list(&nif
->then_list
));
326 /* In order to avoid running nir_lower_regs_to_ssa_impl() every time an if
327 * opt makes progress we leave nir_opt_trivial_continues() to remove the
328 * continue now that the end of the loop has been simplified.
335 * This optimization turns:
350 opt_if_simplification(nir_builder
*b
, nir_if
*nif
)
352 /* Only simplify if the then block is empty and the else block is not. */
353 if (!is_block_empty(nir_if_first_then_block(nif
)) ||
354 is_block_empty(nir_if_first_else_block(nif
)))
357 /* Make sure the condition is a comparison operation. */
358 nir_instr
*src_instr
= nif
->condition
.ssa
->parent_instr
;
359 if (src_instr
->type
!= nir_instr_type_alu
)
362 nir_alu_instr
*alu_instr
= nir_instr_as_alu(src_instr
);
363 if (!nir_alu_instr_is_comparison(alu_instr
))
366 /* Insert the inverted instruction and rewrite the condition. */
367 b
->cursor
= nir_after_instr(&alu_instr
->instr
);
369 nir_ssa_def
*new_condition
=
370 nir_inot(b
, &alu_instr
->dest
.dest
.ssa
);
372 nir_if_rewrite_condition(nif
, nir_src_for_ssa(new_condition
));
374 /* Grab pointers to the last then/else blocks for fixing up the phis. */
375 nir_block
*then_block
= nir_if_last_then_block(nif
);
376 nir_block
*else_block
= nir_if_last_else_block(nif
);
378 /* Walk all the phis in the block immediately following the if statement and
381 nir_block
*after_if_block
=
382 nir_cf_node_as_block(nir_cf_node_next(&nif
->cf_node
));
384 nir_foreach_instr(instr
, after_if_block
) {
385 if (instr
->type
!= nir_instr_type_phi
)
388 nir_phi_instr
*phi
= nir_instr_as_phi(instr
);
390 foreach_list_typed(nir_phi_src
, src
, node
, &phi
->srcs
) {
391 if (src
->pred
== else_block
) {
392 src
->pred
= then_block
;
393 } else if (src
->pred
== then_block
) {
394 src
->pred
= else_block
;
399 /* Finally, move the else block to the then block. */
401 nir_cf_extract(&tmp
, nir_before_cf_list(&nif
->else_list
),
402 nir_after_cf_list(&nif
->else_list
));
403 nir_cf_reinsert(&tmp
, nir_before_cf_list(&nif
->then_list
));
409 * This optimization simplifies potential loop terminators which then allows
410 * other passes such as opt_if_simplification() and loop unrolling to progress
414 * ... then block instructions ...
427 * ... then block instructions ...
430 opt_if_loop_terminator(nir_if
*nif
)
432 nir_block
*break_blk
= NULL
;
433 nir_block
*continue_from_blk
= NULL
;
434 bool continue_from_then
= true;
436 nir_block
*last_then
= nir_if_last_then_block(nif
);
437 nir_block
*last_else
= nir_if_last_else_block(nif
);
439 if (nir_block_ends_in_break(last_then
)) {
440 break_blk
= last_then
;
441 continue_from_blk
= last_else
;
442 continue_from_then
= false;
443 } else if (nir_block_ends_in_break(last_else
)) {
444 break_blk
= last_else
;
445 continue_from_blk
= last_then
;
448 /* Continue if the if-statement contained no jumps at all */
452 /* If the continue from block is empty then return as there is nothing to
455 nir_block
*first_continue_from_blk
= continue_from_then
?
456 nir_if_first_then_block(nif
) :
457 nir_if_first_else_block(nif
);
458 if (is_block_empty(first_continue_from_blk
))
461 if (!nir_is_trivial_loop_if(nif
, break_blk
))
464 /* Finally, move the continue from branch after the if-statement. */
466 nir_cf_extract(&tmp
, nir_before_block(first_continue_from_blk
),
467 nir_after_block(continue_from_blk
));
468 nir_cf_reinsert(&tmp
, nir_after_cf_node(&nif
->cf_node
));
474 evaluate_if_condition(nir_if
*nif
, nir_cursor cursor
, bool *value
)
476 nir_block
*use_block
= nir_cursor_current_block(cursor
);
477 if (nir_block_dominates(nir_if_first_then_block(nif
), use_block
)) {
480 } else if (nir_block_dominates(nir_if_first_else_block(nif
), use_block
)) {
489 clone_alu_and_replace_src_defs(nir_builder
*b
, const nir_alu_instr
*alu
,
490 nir_ssa_def
**src_defs
)
492 nir_alu_instr
*nalu
= nir_alu_instr_create(b
->shader
, alu
->op
);
493 nalu
->exact
= alu
->exact
;
495 nir_ssa_dest_init(&nalu
->instr
, &nalu
->dest
.dest
,
496 alu
->dest
.dest
.ssa
.num_components
,
497 alu
->dest
.dest
.ssa
.bit_size
, alu
->dest
.dest
.ssa
.name
);
499 nalu
->dest
.saturate
= alu
->dest
.saturate
;
500 nalu
->dest
.write_mask
= alu
->dest
.write_mask
;
502 for (unsigned i
= 0; i
< nir_op_infos
[alu
->op
].num_inputs
; i
++) {
503 assert(alu
->src
[i
].src
.is_ssa
);
504 nalu
->src
[i
].src
= nir_src_for_ssa(src_defs
[i
]);
505 nalu
->src
[i
].negate
= alu
->src
[i
].negate
;
506 nalu
->src
[i
].abs
= alu
->src
[i
].abs
;
507 memcpy(nalu
->src
[i
].swizzle
, alu
->src
[i
].swizzle
,
508 sizeof(nalu
->src
[i
].swizzle
));
511 nir_builder_instr_insert(b
, &nalu
->instr
);
513 return &nalu
->dest
.dest
.ssa
;;
517 * This propagates if condition evaluation down the chain of some alu
518 * instructions. For example by checking the use of some of the following alu
519 * instruction we can eventually replace ssa_107 with NIR_TRUE.
523 * vec1 32 ssa_85 = load_const (0x00000002)
524 * vec1 32 ssa_86 = ieq ssa_48, ssa_85
525 * vec1 32 ssa_87 = load_const (0x00000001)
526 * vec1 32 ssa_88 = ieq ssa_48, ssa_87
527 * vec1 32 ssa_89 = ior ssa_86, ssa_88
528 * vec1 32 ssa_90 = ieq ssa_48, ssa_0
529 * vec1 32 ssa_91 = ior ssa_89, ssa_90
554 * vec1 32 ssa_107 = inot ssa_91
564 propagate_condition_eval(nir_builder
*b
, nir_if
*nif
, nir_src
*use_src
,
565 nir_src
*alu_use
, nir_alu_instr
*alu
,
566 bool is_if_condition
)
569 b
->cursor
= nir_before_src(alu_use
, is_if_condition
);
570 if (!evaluate_if_condition(nif
, b
->cursor
, &bool_value
))
573 nir_ssa_def
*def
[4] = {0};
574 for (unsigned i
= 0; i
< nir_op_infos
[alu
->op
].num_inputs
; i
++) {
575 if (alu
->src
[i
].src
.ssa
== use_src
->ssa
) {
576 def
[i
] = nir_imm_bool(b
, bool_value
);
578 def
[i
] = alu
->src
[i
].src
.ssa
;
582 nir_ssa_def
*nalu
= clone_alu_and_replace_src_defs(b
, alu
, def
);
584 /* Rewrite use to use new alu instruction */
585 nir_src new_src
= nir_src_for_ssa(nalu
);
588 nir_if_rewrite_condition(alu_use
->parent_if
, new_src
);
590 nir_instr_rewrite_src(alu_use
->parent_instr
, alu_use
, new_src
);
596 can_propagate_through_alu(nir_src
*src
)
598 if (src
->parent_instr
->type
!= nir_instr_type_alu
)
601 nir_alu_instr
*alu
= nir_instr_as_alu(src
->parent_instr
);
609 return src
== &alu
->src
[0].src
;
616 evaluate_condition_use(nir_builder
*b
, nir_if
*nif
, nir_src
*use_src
,
617 bool is_if_condition
)
619 bool progress
= false;
621 b
->cursor
= nir_before_src(use_src
, is_if_condition
);
624 if (evaluate_if_condition(nif
, b
->cursor
, &bool_value
)) {
625 /* Rewrite use to use const */
626 nir_src imm_src
= nir_src_for_ssa(nir_imm_bool(b
, bool_value
));
628 nir_if_rewrite_condition(use_src
->parent_if
, imm_src
);
630 nir_instr_rewrite_src(use_src
->parent_instr
, use_src
, imm_src
);
635 if (!is_if_condition
&& can_propagate_through_alu(use_src
)) {
636 nir_alu_instr
*alu
= nir_instr_as_alu(use_src
->parent_instr
);
638 nir_foreach_use_safe(alu_use
, &alu
->dest
.dest
.ssa
) {
639 progress
|= propagate_condition_eval(b
, nif
, use_src
, alu_use
, alu
,
643 nir_foreach_if_use_safe(alu_use
, &alu
->dest
.dest
.ssa
) {
644 progress
|= propagate_condition_eval(b
, nif
, use_src
, alu_use
, alu
,
653 opt_if_evaluate_condition_use(nir_builder
*b
, nir_if
*nif
)
655 bool progress
= false;
657 /* Evaluate any uses of the if condition inside the if branches */
658 assert(nif
->condition
.is_ssa
);
659 nir_foreach_use_safe(use_src
, nif
->condition
.ssa
) {
660 progress
|= evaluate_condition_use(b
, nif
, use_src
, false);
663 nir_foreach_if_use_safe(use_src
, nif
->condition
.ssa
) {
664 if (use_src
->parent_if
!= nif
)
665 progress
|= evaluate_condition_use(b
, nif
, use_src
, true);
672 opt_if_cf_list(nir_builder
*b
, struct exec_list
*cf_list
)
674 bool progress
= false;
675 foreach_list_typed(nir_cf_node
, cf_node
, node
, cf_list
) {
676 switch (cf_node
->type
) {
677 case nir_cf_node_block
:
680 case nir_cf_node_if
: {
681 nir_if
*nif
= nir_cf_node_as_if(cf_node
);
682 progress
|= opt_if_cf_list(b
, &nif
->then_list
);
683 progress
|= opt_if_cf_list(b
, &nif
->else_list
);
684 progress
|= opt_if_loop_terminator(nif
);
685 progress
|= opt_if_simplification(b
, nif
);
689 case nir_cf_node_loop
: {
690 nir_loop
*loop
= nir_cf_node_as_loop(cf_node
);
691 progress
|= opt_if_cf_list(b
, &loop
->body
);
692 progress
|= opt_peel_loop_initial_if(loop
);
693 progress
|= opt_if_loop_last_continue(loop
);
697 case nir_cf_node_function
:
698 unreachable("Invalid cf type");
706 * These optimisations depend on nir_metadata_block_index and therefore must
707 * not do anything to cause the metadata to become invalid.
710 opt_if_safe_cf_list(nir_builder
*b
, struct exec_list
*cf_list
)
712 bool progress
= false;
713 foreach_list_typed(nir_cf_node
, cf_node
, node
, cf_list
) {
714 switch (cf_node
->type
) {
715 case nir_cf_node_block
:
718 case nir_cf_node_if
: {
719 nir_if
*nif
= nir_cf_node_as_if(cf_node
);
720 progress
|= opt_if_safe_cf_list(b
, &nif
->then_list
);
721 progress
|= opt_if_safe_cf_list(b
, &nif
->else_list
);
722 progress
|= opt_if_evaluate_condition_use(b
, nif
);
726 case nir_cf_node_loop
: {
727 nir_loop
*loop
= nir_cf_node_as_loop(cf_node
);
728 progress
|= opt_if_safe_cf_list(b
, &loop
->body
);
732 case nir_cf_node_function
:
733 unreachable("Invalid cf type");
741 nir_opt_if(nir_shader
*shader
)
743 bool progress
= false;
745 nir_foreach_function(function
, shader
) {
746 if (function
->impl
== NULL
)
750 nir_builder_init(&b
, function
->impl
);
752 nir_metadata_require(function
->impl
, nir_metadata_block_index
|
753 nir_metadata_dominance
);
754 progress
= opt_if_safe_cf_list(&b
, &function
->impl
->body
);
755 nir_metadata_preserve(function
->impl
, nir_metadata_block_index
|
756 nir_metadata_dominance
);
758 if (opt_if_cf_list(&b
, &function
->impl
->body
)) {
759 nir_metadata_preserve(function
->impl
, nir_metadata_none
);
761 /* If that made progress, we're no longer really in SSA form. We
762 * need to convert registers back into SSA defs and clean up SSA defs
763 * that don't dominate their uses.
765 nir_lower_regs_to_ssa_impl(function
->impl
);