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_control_flow.h"
27 #include "nir_loop_analyze.h"
30 * Gets the single block that jumps back to the loop header. Already assumes
31 * there is exactly one such block.
34 find_continue_block(nir_loop
*loop
)
36 nir_block
*header_block
= nir_loop_first_block(loop
);
37 nir_block
*prev_block
=
38 nir_cf_node_as_block(nir_cf_node_prev(&loop
->cf_node
));
40 assert(header_block
->predecessors
->entries
== 2);
42 struct set_entry
*pred_entry
;
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 /* Before we do anything, convert the loop to LCSSA. We're about to
184 * replace a bunch of SSA defs with registers and this will prevent any of
185 * it from leaking outside the loop.
187 nir_convert_loop_to_lcssa(loop
);
189 nir_block
*after_if_block
=
190 nir_cf_node_as_block(nir_cf_node_next(&nif
->cf_node
));
192 /* Get rid of phis in the header block since we will be duplicating it */
193 nir_lower_phis_to_regs_block(header_block
);
194 /* Get rid of phis after the if since dominance will change */
195 nir_lower_phis_to_regs_block(after_if_block
);
197 /* Get rid of SSA defs in the pieces we're about to move around */
198 nir_lower_ssa_defs_to_regs_block(header_block
);
199 nir_foreach_block_in_cf_node(block
, &nif
->cf_node
)
200 nir_lower_ssa_defs_to_regs_block(block
);
202 nir_cf_list header
, tmp
;
203 nir_cf_extract(&header
, nir_before_block(header_block
),
204 nir_after_block(header_block
));
206 nir_cf_list_clone(&tmp
, &header
, &loop
->cf_node
, NULL
);
207 nir_cf_reinsert(&tmp
, nir_before_cf_node(&loop
->cf_node
));
208 nir_cf_extract(&tmp
, nir_before_cf_list(entry_list
),
209 nir_after_cf_list(entry_list
));
210 nir_cf_reinsert(&tmp
, nir_before_cf_node(&loop
->cf_node
));
212 nir_cf_reinsert(&header
, nir_after_block_before_jump(continue_block
));
214 /* Get continue block again as the previous reinsert might have removed the block. */
215 continue_block
= find_continue_block(loop
);
217 nir_cf_extract(&tmp
, nir_before_cf_list(continue_list
),
218 nir_after_cf_list(continue_list
));
219 nir_cf_reinsert(&tmp
, nir_after_block_before_jump(continue_block
));
221 nir_cf_node_remove(&nif
->cf_node
);
227 is_block_empty(nir_block
*block
)
229 return nir_cf_node_is_last(&block
->cf_node
) &&
230 exec_list_is_empty(&block
->instr_list
);
234 * This optimization turns:
249 opt_if_simplification(nir_builder
*b
, nir_if
*nif
)
251 /* Only simplify if the then block is empty and the else block is not. */
252 if (!is_block_empty(nir_if_first_then_block(nif
)) ||
253 is_block_empty(nir_if_first_else_block(nif
)))
256 /* Make sure the condition is a comparison operation. */
257 nir_instr
*src_instr
= nif
->condition
.ssa
->parent_instr
;
258 if (src_instr
->type
!= nir_instr_type_alu
)
261 nir_alu_instr
*alu_instr
= nir_instr_as_alu(src_instr
);
262 if (!nir_alu_instr_is_comparison(alu_instr
))
265 /* Insert the inverted instruction and rewrite the condition. */
266 b
->cursor
= nir_after_instr(&alu_instr
->instr
);
268 nir_ssa_def
*new_condition
=
269 nir_inot(b
, &alu_instr
->dest
.dest
.ssa
);
271 nir_if_rewrite_condition(nif
, nir_src_for_ssa(new_condition
));
273 /* Grab pointers to the last then/else blocks for fixing up the phis. */
274 nir_block
*then_block
= nir_if_last_then_block(nif
);
275 nir_block
*else_block
= nir_if_last_else_block(nif
);
277 /* Walk all the phis in the block immediately following the if statement and
280 nir_block
*after_if_block
=
281 nir_cf_node_as_block(nir_cf_node_next(&nif
->cf_node
));
283 nir_foreach_instr(instr
, after_if_block
) {
284 if (instr
->type
!= nir_instr_type_phi
)
287 nir_phi_instr
*phi
= nir_instr_as_phi(instr
);
289 foreach_list_typed(nir_phi_src
, src
, node
, &phi
->srcs
) {
290 if (src
->pred
== else_block
) {
291 src
->pred
= then_block
;
292 } else if (src
->pred
== then_block
) {
293 src
->pred
= else_block
;
298 /* Finally, move the else block to the then block. */
300 nir_cf_extract(&tmp
, nir_before_cf_list(&nif
->else_list
),
301 nir_after_cf_list(&nif
->else_list
));
302 nir_cf_reinsert(&tmp
, nir_before_cf_list(&nif
->then_list
));
308 * This optimization simplifies potential loop terminators which then allows
309 * other passes such as opt_if_simplification() and loop unrolling to progress
313 * ... then block instructions ...
326 * ... then block instructions ...
329 opt_if_loop_terminator(nir_if
*nif
)
331 nir_block
*break_blk
= NULL
;
332 nir_block
*continue_from_blk
= NULL
;
333 bool continue_from_then
= true;
335 nir_block
*last_then
= nir_if_last_then_block(nif
);
336 nir_block
*last_else
= nir_if_last_else_block(nif
);
338 if (nir_block_ends_in_break(last_then
)) {
339 break_blk
= last_then
;
340 continue_from_blk
= last_else
;
341 continue_from_then
= false;
342 } else if (nir_block_ends_in_break(last_else
)) {
343 break_blk
= last_else
;
344 continue_from_blk
= last_then
;
347 /* Continue if the if-statement contained no jumps at all */
351 /* If the continue from block is empty then return as there is nothing to
354 nir_block
*first_continue_from_blk
= continue_from_then
?
355 nir_if_first_then_block(nif
) :
356 nir_if_first_else_block(nif
);
357 if (is_block_empty(first_continue_from_blk
))
360 if (!nir_is_trivial_loop_if(nif
, break_blk
))
363 /* Finally, move the continue from branch after the if-statement. */
365 nir_cf_extract(&tmp
, nir_before_block(first_continue_from_blk
),
366 nir_after_block(continue_from_blk
));
367 nir_cf_reinsert(&tmp
, nir_after_cf_node(&nif
->cf_node
));
373 replace_if_condition_use_with_const(nir_builder
*b
, nir_src
*use
,
374 nir_const_value nir_boolean
,
378 nir_ssa_def
*const_def
= nir_build_imm(b
, 1, 32, nir_boolean
);
380 /* Rewrite use to use const */
381 nir_src new_src
= nir_src_for_ssa(const_def
);
383 nir_if_rewrite_condition(use
->parent_if
, new_src
);
385 nir_instr_rewrite_src(use
->parent_instr
, use
, new_src
);
389 evaluate_if_condition(nir_if
*nif
, nir_cursor cursor
, uint32_t *value
)
391 nir_block
*use_block
= nir_cursor_current_block(cursor
);
392 if (nir_block_dominates(nir_if_first_then_block(nif
), use_block
)) {
395 } else if (nir_block_dominates(nir_if_first_else_block(nif
), use_block
)) {
404 evaluate_condition_use(nir_builder
*b
, nir_if
*nif
, nir_src
*use_src
,
405 bool is_if_condition
)
407 bool progress
= false;
409 nir_const_value value
;
410 b
->cursor
= nir_before_src(use_src
, is_if_condition
);
412 if (evaluate_if_condition(nif
, b
->cursor
, &value
.u32
[0])) {
413 replace_if_condition_use_with_const(b
, use_src
, value
, is_if_condition
);
421 opt_if_evaluate_condition_use(nir_builder
*b
, nir_if
*nif
)
423 bool progress
= false;
425 /* Evaluate any uses of the if condition inside the if branches */
426 assert(nif
->condition
.is_ssa
);
427 nir_foreach_use_safe(use_src
, nif
->condition
.ssa
) {
428 progress
|= evaluate_condition_use(b
, nif
, use_src
, false);
431 nir_foreach_if_use_safe(use_src
, nif
->condition
.ssa
) {
432 if (use_src
->parent_if
!= nif
)
433 progress
|= evaluate_condition_use(b
, nif
, use_src
, true);
440 opt_if_cf_list(nir_builder
*b
, struct exec_list
*cf_list
)
442 bool progress
= false;
443 foreach_list_typed(nir_cf_node
, cf_node
, node
, cf_list
) {
444 switch (cf_node
->type
) {
445 case nir_cf_node_block
:
448 case nir_cf_node_if
: {
449 nir_if
*nif
= nir_cf_node_as_if(cf_node
);
450 progress
|= opt_if_cf_list(b
, &nif
->then_list
);
451 progress
|= opt_if_cf_list(b
, &nif
->else_list
);
452 progress
|= opt_if_loop_terminator(nif
);
453 progress
|= opt_if_simplification(b
, nif
);
457 case nir_cf_node_loop
: {
458 nir_loop
*loop
= nir_cf_node_as_loop(cf_node
);
459 progress
|= opt_if_cf_list(b
, &loop
->body
);
460 progress
|= opt_peel_loop_initial_if(loop
);
464 case nir_cf_node_function
:
465 unreachable("Invalid cf type");
473 * These optimisations depend on nir_metadata_block_index and therefore must
474 * not do anything to cause the metadata to become invalid.
477 opt_if_safe_cf_list(nir_builder
*b
, struct exec_list
*cf_list
)
479 bool progress
= false;
480 foreach_list_typed(nir_cf_node
, cf_node
, node
, cf_list
) {
481 switch (cf_node
->type
) {
482 case nir_cf_node_block
:
485 case nir_cf_node_if
: {
486 nir_if
*nif
= nir_cf_node_as_if(cf_node
);
487 progress
|= opt_if_safe_cf_list(b
, &nif
->then_list
);
488 progress
|= opt_if_safe_cf_list(b
, &nif
->else_list
);
489 progress
|= opt_if_evaluate_condition_use(b
, nif
);
493 case nir_cf_node_loop
: {
494 nir_loop
*loop
= nir_cf_node_as_loop(cf_node
);
495 progress
|= opt_if_safe_cf_list(b
, &loop
->body
);
499 case nir_cf_node_function
:
500 unreachable("Invalid cf type");
508 nir_opt_if(nir_shader
*shader
)
510 bool progress
= false;
512 nir_foreach_function(function
, shader
) {
513 if (function
->impl
== NULL
)
517 nir_builder_init(&b
, function
->impl
);
519 nir_metadata_require(function
->impl
, nir_metadata_block_index
|
520 nir_metadata_dominance
);
521 progress
= opt_if_safe_cf_list(&b
, &function
->impl
->body
);
522 nir_metadata_preserve(function
->impl
, nir_metadata_block_index
|
523 nir_metadata_dominance
);
525 if (opt_if_cf_list(&b
, &function
->impl
->body
)) {
526 nir_metadata_preserve(function
->impl
, nir_metadata_none
);
528 /* If that made progress, we're no longer really in SSA form. We
529 * need to convert registers back into SSA defs and clean up SSA defs
530 * that don't dominate their uses.
532 nir_lower_regs_to_ssa_impl(function
->impl
);
534 /* Calling nir_convert_loop_to_lcssa() in opt_peel_loop_initial_if()
535 * adds extra phi nodes which may not be valid if they're used for
536 * something such as a deref. Remove any unneeded phis.
538 nir_opt_remove_phis_impl(function
->impl
);