1 /*
2 * Copyright © 2012 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 /** @file brw_fs_copy_propagation.cpp
25 *
26 * Support for global copy propagation in two passes: A local pass that does
27 * intra-block copy (and constant) propagation, and a global pass that uses
28 * dataflow analysis on the copies available at the end of each block to re-do
29 * local copy propagation with more copies available.
30 *
31 * See Muchnick's Advanced Compiler Design and Implementation, section
32 * 12.5 (p356).
33 */
34
35 #define ACP_HASH_SIZE 16
36
37 #include "util/bitset.h"
38 #include "brw_fs.h"
39 #include "brw_cfg.h"
40 #include "brw_eu.h"
41
42 namespace { /* avoid conflict with opt_copy_propagation_elements */
43 struct acp_entry : public exec_node {
44 fs_reg dst;
45 fs_reg src;
46 uint8_t size_written;
47 uint8_t size_read;
48 enum opcode opcode;
49 bool saturate;
50 };
51
52 struct block_data {
53 /**
54 * Which entries in the fs_copy_prop_dataflow acp table are live at the
55 * start of this block. This is the useful output of the analysis, since
56 * it lets us plug those into the local copy propagation on the second
57 * pass.
58 */
59 BITSET_WORD *livein;
60
61 /**
62 * Which entries in the fs_copy_prop_dataflow acp table are live at the end
63 * of this block. This is done in initial setup from the per-block acps
64 * returned by the first local copy prop pass.
65 */
66 BITSET_WORD *liveout;
67
68 /**
69 * Which entries in the fs_copy_prop_dataflow acp table are generated by
70 * instructions in this block which reach the end of the block without
71 * being killed.
72 */
73 BITSET_WORD *copy;
74
75 /**
76 * Which entries in the fs_copy_prop_dataflow acp table are killed over the
77 * course of this block.
78 */
79 BITSET_WORD *kill;
80 };
81
82 class fs_copy_prop_dataflow
83 {
84 public:
85 fs_copy_prop_dataflow(void *mem_ctx, cfg_t *cfg,
86 exec_list *out_acp[ACP_HASH_SIZE]);
87
88 void setup_initial_values();
89 void run();
90
91 void dump_block_data() const UNUSED;
92
93 void *mem_ctx;
94 cfg_t *cfg;
95
96 acp_entry **acp;
97 int num_acp;
98 int bitset_words;
99
100 struct block_data *bd;
101 };
102 } /* anonymous namespace */
103
fs_copy_prop_dataflow(void * mem_ctx,cfg_t * cfg,exec_list * out_acp[ACP_HASH_SIZE])104 fs_copy_prop_dataflow::fs_copy_prop_dataflow(void *mem_ctx, cfg_t *cfg,
105 exec_list *out_acp[ACP_HASH_SIZE])
106 : mem_ctx(mem_ctx), cfg(cfg)
107 {
108 bd = rzalloc_array(mem_ctx, struct block_data, cfg->num_blocks);
109
110 num_acp = 0;
111 foreach_block (block, cfg) {
112 for (int i = 0; i < ACP_HASH_SIZE; i++) {
113 num_acp += out_acp[block->num][i].length();
114 }
115 }
116
117 acp = rzalloc_array(mem_ctx, struct acp_entry *, num_acp);
118
119 bitset_words = BITSET_WORDS(num_acp);
120
121 int next_acp = 0;
122 foreach_block (block, cfg) {
123 bd[block->num].livein = rzalloc_array(bd, BITSET_WORD, bitset_words);
124 bd[block->num].liveout = rzalloc_array(bd, BITSET_WORD, bitset_words);
125 bd[block->num].copy = rzalloc_array(bd, BITSET_WORD, bitset_words);
126 bd[block->num].kill = rzalloc_array(bd, BITSET_WORD, bitset_words);
127
128 for (int i = 0; i < ACP_HASH_SIZE; i++) {
129 foreach_in_list(acp_entry, entry, &out_acp[block->num][i]) {
130 acp[next_acp] = entry;
131
132 /* opt_copy_propagation_local populates out_acp with copies created
133 * in a block which are still live at the end of the block. This
134 * is exactly what we want in the COPY set.
135 */
136 BITSET_SET(bd[block->num].copy, next_acp);
137
138 next_acp++;
139 }
140 }
141 }
142
143 assert(next_acp == num_acp);
144
145 setup_initial_values();
146 run();
147 }
148
149 /**
150 * Set up initial values for each of the data flow sets, prior to running
151 * the fixed-point algorithm.
152 */
153 void
setup_initial_values()154 fs_copy_prop_dataflow::setup_initial_values()
155 {
156 /* Initialize the COPY and KILL sets. */
157 foreach_block (block, cfg) {
158 foreach_inst_in_block(fs_inst, inst, block) {
159 if (inst->dst.file != VGRF)
160 continue;
161
162 /* Mark ACP entries which are killed by this instruction. */
163 for (int i = 0; i < num_acp; i++) {
164 if (regions_overlap(inst->dst, inst->size_written,
165 acp[i]->dst, acp[i]->size_written) ||
166 regions_overlap(inst->dst, inst->size_written,
167 acp[i]->src, acp[i]->size_read)) {
168 BITSET_SET(bd[block->num].kill, i);
169 }
170 }
171 }
172 }
173
174 /* Populate the initial values for the livein and liveout sets. For the
175 * block at the start of the program, livein = 0 and liveout = copy.
176 * For the others, set liveout to 0 (the empty set) and livein to ~0
177 * (the universal set).
178 */
179 foreach_block (block, cfg) {
180 if (block->parents.is_empty()) {
181 for (int i = 0; i < bitset_words; i++) {
182 bd[block->num].livein[i] = 0u;
183 bd[block->num].liveout[i] = bd[block->num].copy[i];
184 }
185 } else {
186 for (int i = 0; i < bitset_words; i++) {
187 bd[block->num].liveout[i] = 0u;
188 bd[block->num].livein[i] = ~0u;
189 }
190 }
191 }
192 }
193
194 /**
195 * Walk the set of instructions in the block, marking which entries in the acp
196 * are killed by the block.
197 */
198 void
run()199 fs_copy_prop_dataflow::run()
200 {
201 bool progress;
202
203 do {
204 progress = false;
205
206 /* Update liveout for all blocks. */
207 foreach_block (block, cfg) {
208 if (block->parents.is_empty())
209 continue;
210
211 for (int i = 0; i < bitset_words; i++) {
212 const BITSET_WORD old_liveout = bd[block->num].liveout[i];
213
214 bd[block->num].liveout[i] =
215 bd[block->num].copy[i] | (bd[block->num].livein[i] &
216 ~bd[block->num].kill[i]);
217
218 if (old_liveout != bd[block->num].liveout[i])
219 progress = true;
220 }
221 }
222
223 /* Update livein for all blocks. If a copy is live out of all parent
224 * blocks, it's live coming in to this block.
225 */
226 foreach_block (block, cfg) {
227 if (block->parents.is_empty())
228 continue;
229
230 for (int i = 0; i < bitset_words; i++) {
231 const BITSET_WORD old_livein = bd[block->num].livein[i];
232
233 bd[block->num].livein[i] = ~0u;
234 foreach_list_typed(bblock_link, parent_link, link, &block->parents) {
235 bblock_t *parent = parent_link->block;
236 bd[block->num].livein[i] &= bd[parent->num].liveout[i];
237 }
238
239 if (old_livein != bd[block->num].livein[i])
240 progress = true;
241 }
242 }
243 } while (progress);
244 }
245
246 void
dump_block_data() const247 fs_copy_prop_dataflow::dump_block_data() const
248 {
249 foreach_block (block, cfg) {
250 fprintf(stderr, "Block %d [%d, %d] (parents ", block->num,
251 block->start_ip, block->end_ip);
252 foreach_list_typed(bblock_link, link, link, &block->parents) {
253 bblock_t *parent = link->block;
254 fprintf(stderr, "%d ", parent->num);
255 }
256 fprintf(stderr, "):\n");
257 fprintf(stderr, " livein = 0x");
258 for (int i = 0; i < bitset_words; i++)
259 fprintf(stderr, "%08x", bd[block->num].livein[i]);
260 fprintf(stderr, ", liveout = 0x");
261 for (int i = 0; i < bitset_words; i++)
262 fprintf(stderr, "%08x", bd[block->num].liveout[i]);
263 fprintf(stderr, ",\n copy = 0x");
264 for (int i = 0; i < bitset_words; i++)
265 fprintf(stderr, "%08x", bd[block->num].copy[i]);
266 fprintf(stderr, ", kill = 0x");
267 for (int i = 0; i < bitset_words; i++)
268 fprintf(stderr, "%08x", bd[block->num].kill[i]);
269 fprintf(stderr, "\n");
270 }
271 }
272
273 static bool
is_logic_op(enum opcode opcode)274 is_logic_op(enum opcode opcode)
275 {
276 return (opcode == BRW_OPCODE_AND ||
277 opcode == BRW_OPCODE_OR ||
278 opcode == BRW_OPCODE_XOR ||
279 opcode == BRW_OPCODE_NOT);
280 }
281
282 static bool
can_take_stride(fs_inst * inst,unsigned arg,unsigned stride,const gen_device_info * devinfo)283 can_take_stride(fs_inst *inst, unsigned arg, unsigned stride,
284 const gen_device_info *devinfo)
285 {
286 if (stride > 4)
287 return false;
288
289 /* 3-source instructions can only be Align16, which restricts what strides
290 * they can take. They can only take a stride of 1 (the usual case), or 0
291 * with a special "repctrl" bit. But the repctrl bit doesn't work for
292 * 64-bit datatypes, so if the source type is 64-bit then only a stride of
293 * 1 is allowed. From the Broadwell PRM, Volume 7 "3D Media GPGPU", page
294 * 944:
295 *
296 * This is applicable to 32b datatypes and 16b datatype. 64b datatypes
297 * cannot use the replicate control.
298 */
299 if (inst->is_3src(devinfo)) {
300 if (type_sz(inst->src[arg].type) > 4)
301 return stride == 1;
302 else
303 return stride == 1 || stride == 0;
304 }
305
306 /* From the Broadwell PRM, Volume 2a "Command Reference - Instructions",
307 * page 391 ("Extended Math Function"):
308 *
309 * The following restrictions apply for align1 mode: Scalar source is
310 * supported. Source and destination horizontal stride must be the
311 * same.
312 *
313 * From the Haswell PRM Volume 2b "Command Reference - Instructions", page
314 * 134 ("Extended Math Function"):
315 *
316 * Scalar source is supported. Source and destination horizontal stride
317 * must be 1.
318 *
319 * and similar language exists for IVB and SNB. Pre-SNB, math instructions
320 * are sends, so the sources are moved to MRF's and there are no
321 * restrictions.
322 */
323 if (inst->is_math()) {
324 if (devinfo->gen == 6 || devinfo->gen == 7) {
325 assert(inst->dst.stride == 1);
326 return stride == 1 || stride == 0;
327 } else if (devinfo->gen >= 8) {
328 return stride == inst->dst.stride || stride == 0;
329 }
330 }
331
332 return true;
333 }
334
335 bool
try_copy_propagate(fs_inst * inst,int arg,acp_entry * entry)336 fs_visitor::try_copy_propagate(fs_inst *inst, int arg, acp_entry *entry)
337 {
338 if (inst->src[arg].file != VGRF)
339 return false;
340
341 if (entry->src.file == IMM)
342 return false;
343 assert(entry->src.file == VGRF || entry->src.file == UNIFORM ||
344 entry->src.file == ATTR);
345
346 if (entry->opcode == SHADER_OPCODE_LOAD_PAYLOAD &&
347 inst->opcode == SHADER_OPCODE_LOAD_PAYLOAD)
348 return false;
349
350 assert(entry->dst.file == VGRF);
351 if (inst->src[arg].nr != entry->dst.nr)
352 return false;
353
354 /* Bail if inst is reading a range that isn't contained in the range
355 * that entry is writing.
356 */
357 if (!region_contained_in(inst->src[arg], inst->size_read(arg),
358 entry->dst, entry->size_written))
359 return false;
360
361 /* we can't generally copy-propagate UD negations because we
362 * can end up accessing the resulting values as signed integers
363 * instead. See also resolve_ud_negate() and comment in
364 * fs_generator::generate_code.
365 */
366 if (entry->src.type == BRW_REGISTER_TYPE_UD &&
367 entry->src.negate)
368 return false;
369
370 bool has_source_modifiers = entry->src.abs || entry->src.negate;
371
372 if ((has_source_modifiers || entry->src.file == UNIFORM ||
373 !entry->src.is_contiguous()) &&
374 !inst->can_do_source_mods(devinfo))
375 return false;
376
377 if (has_source_modifiers &&
378 inst->opcode == SHADER_OPCODE_GEN4_SCRATCH_WRITE)
379 return false;
380
381 /* Bail if the result of composing both strides would exceed the
382 * hardware limit.
383 */
384 if (!can_take_stride(inst, arg, entry->src.stride * inst->src[arg].stride,
385 devinfo))
386 return false;
387
388 /* Bail if the instruction type is larger than the execution type of the
389 * copy, what implies that each channel is reading multiple channels of the
390 * destination of the copy, and simply replacing the sources would give a
391 * program with different semantics.
392 */
393 if (type_sz(entry->dst.type) < type_sz(inst->src[arg].type))
394 return false;
395
396 /* Bail if the result of composing both strides cannot be expressed
397 * as another stride. This avoids, for example, trying to transform
398 * this:
399 *
400 * MOV (8) rX<1>UD rY<0;1,0>UD
401 * FOO (8) ... rX<8;8,1>UW
402 *
403 * into this:
404 *
405 * FOO (8) ... rY<0;1,0>UW
406 *
407 * Which would have different semantics.
408 */
409 if (entry->src.stride != 1 &&
410 (inst->src[arg].stride *
411 type_sz(inst->src[arg].type)) % type_sz(entry->src.type) != 0)
412 return false;
413
414 /* Since semantics of source modifiers are type-dependent we need to
415 * ensure that the meaning of the instruction remains the same if we
416 * change the type. If the sizes of the types are different the new
417 * instruction will read a different amount of data than the original
418 * and the semantics will always be different.
419 */
420 if (has_source_modifiers &&
421 entry->dst.type != inst->src[arg].type &&
422 (!inst->can_change_types() ||
423 type_sz(entry->dst.type) != type_sz(inst->src[arg].type)))
424 return false;
425
426 if (devinfo->gen >= 8 && (entry->src.negate || entry->src.abs) &&
427 is_logic_op(inst->opcode)) {
428 return false;
429 }
430
431 if (entry->saturate) {
432 switch(inst->opcode) {
433 case BRW_OPCODE_SEL:
434 if ((inst->conditional_mod != BRW_CONDITIONAL_GE &&
435 inst->conditional_mod != BRW_CONDITIONAL_L) ||
436 inst->src[1].file != IMM ||
437 inst->src[1].f < 0.0 ||
438 inst->src[1].f > 1.0) {
439 return false;
440 }
441 break;
442 default:
443 return false;
444 }
445 }
446
447 inst->src[arg].file = entry->src.file;
448 inst->src[arg].nr = entry->src.nr;
449 inst->src[arg].stride *= entry->src.stride;
450 inst->saturate = inst->saturate || entry->saturate;
451
452 /* Compute the offset of inst->src[arg] relative to entry->dst */
453 const unsigned rel_offset = inst->src[arg].offset - entry->dst.offset;
454
455 /* Compute the first component of the copy that the instruction is
456 * reading, and the base byte offset within that component.
457 */
458 assert(entry->dst.offset % REG_SIZE == 0 && entry->dst.stride == 1);
459 const unsigned component = rel_offset / type_sz(entry->dst.type);
460 const unsigned suboffset = rel_offset % type_sz(entry->dst.type);
461
462 /* Calculate the byte offset at the origin of the copy of the given
463 * component and suboffset.
464 */
465 inst->src[arg].offset = suboffset +
466 component * entry->src.stride * type_sz(entry->src.type) +
467 entry->src.offset;
468
469 if (has_source_modifiers) {
470 if (entry->dst.type != inst->src[arg].type) {
471 /* We are propagating source modifiers from a MOV with a different
472 * type. If we got here, then we can just change the source and
473 * destination types of the instruction and keep going.
474 */
475 assert(inst->can_change_types());
476 for (int i = 0; i < inst->sources; i++) {
477 inst->src[i].type = entry->dst.type;
478 }
479 inst->dst.type = entry->dst.type;
480 }
481
482 if (!inst->src[arg].abs) {
483 inst->src[arg].abs = entry->src.abs;
484 inst->src[arg].negate ^= entry->src.negate;
485 }
486 }
487
488 return true;
489 }
490
491
492 bool
try_constant_propagate(fs_inst * inst,acp_entry * entry)493 fs_visitor::try_constant_propagate(fs_inst *inst, acp_entry *entry)
494 {
495 bool progress = false;
496
497 if (entry->src.file != IMM)
498 return false;
499 if (type_sz(entry->src.type) > 4)
500 return false;
501 if (entry->saturate)
502 return false;
503
504 for (int i = inst->sources - 1; i >= 0; i--) {
505 if (inst->src[i].file != VGRF)
506 continue;
507
508 assert(entry->dst.file == VGRF);
509 if (inst->src[i].nr != entry->dst.nr)
510 continue;
511
512 /* Bail if inst is reading a range that isn't contained in the range
513 * that entry is writing.
514 */
515 if (!region_contained_in(inst->src[i], inst->size_read(i),
516 entry->dst, entry->size_written))
517 continue;
518
519 /* If the type sizes don't match each channel of the instruction is
520 * either extracting a portion of the constant (which could be handled
521 * with some effort but the code below doesn't) or reading multiple
522 * channels of the source at once.
523 */
524 if (type_sz(inst->src[i].type) != type_sz(entry->dst.type))
525 continue;
526
527 fs_reg val = entry->src;
528 val.type = inst->src[i].type;
529
530 if (inst->src[i].abs) {
531 if ((devinfo->gen >= 8 && is_logic_op(inst->opcode)) ||
532 !brw_abs_immediate(val.type, &val.as_brw_reg())) {
533 continue;
534 }
535 }
536
537 if (inst->src[i].negate) {
538 if ((devinfo->gen >= 8 && is_logic_op(inst->opcode)) ||
539 !brw_negate_immediate(val.type, &val.as_brw_reg())) {
540 continue;
541 }
542 }
543
544 switch (inst->opcode) {
545 case BRW_OPCODE_MOV:
546 case SHADER_OPCODE_LOAD_PAYLOAD:
547 case FS_OPCODE_PACK:
548 inst->src[i] = val;
549 progress = true;
550 break;
551
552 case SHADER_OPCODE_INT_QUOTIENT:
553 case SHADER_OPCODE_INT_REMAINDER:
554 /* FINISHME: Promote non-float constants and remove this. */
555 if (devinfo->gen < 8)
556 break;
557 /* fallthrough */
558 case SHADER_OPCODE_POW:
559 /* Allow constant propagation into src1 (except on Gen 6 which
560 * doesn't support scalar source math), and let constant combining
561 * promote the constant on Gen < 8.
562 */
563 if (devinfo->gen == 6)
564 break;
565 /* fallthrough */
566 case BRW_OPCODE_BFI1:
567 case BRW_OPCODE_ASR:
568 case BRW_OPCODE_SHL:
569 case BRW_OPCODE_SHR:
570 case BRW_OPCODE_SUBB:
571 if (i == 1) {
572 inst->src[i] = val;
573 progress = true;
574 }
575 break;
576
577 case BRW_OPCODE_MACH:
578 case BRW_OPCODE_MUL:
579 case SHADER_OPCODE_MULH:
580 case BRW_OPCODE_ADD:
581 case BRW_OPCODE_OR:
582 case BRW_OPCODE_AND:
583 case BRW_OPCODE_XOR:
584 case BRW_OPCODE_ADDC:
585 if (i == 1) {
586 inst->src[i] = val;
587 progress = true;
588 } else if (i == 0 && inst->src[1].file != IMM) {
589 /* Fit this constant in by commuting the operands.
590 * Exception: we can't do this for 32-bit integer MUL/MACH
591 * because it's asymmetric.
592 *
593 * The BSpec says for Broadwell that
594 *
595 * "When multiplying DW x DW, the dst cannot be accumulator."
596 *
597 * Integer MUL with a non-accumulator destination will be lowered
598 * by lower_integer_multiplication(), so don't restrict it.
599 */
600 if (((inst->opcode == BRW_OPCODE_MUL &&
601 inst->dst.is_accumulator()) ||
602 inst->opcode == BRW_OPCODE_MACH) &&
603 (inst->src[1].type == BRW_REGISTER_TYPE_D ||
604 inst->src[1].type == BRW_REGISTER_TYPE_UD))
605 break;
606 inst->src[0] = inst->src[1];
607 inst->src[1] = val;
608 progress = true;
609 }
610 break;
611
612 case BRW_OPCODE_CMP:
613 case BRW_OPCODE_IF:
614 if (i == 1) {
615 inst->src[i] = val;
616 progress = true;
617 } else if (i == 0 && inst->src[1].file != IMM) {
618 enum brw_conditional_mod new_cmod;
619
620 new_cmod = brw_swap_cmod(inst->conditional_mod);
621 if (new_cmod != BRW_CONDITIONAL_NONE) {
622 /* Fit this constant in by swapping the operands and
623 * flipping the test
624 */
625 inst->src[0] = inst->src[1];
626 inst->src[1] = val;
627 inst->conditional_mod = new_cmod;
628 progress = true;
629 }
630 }
631 break;
632
633 case BRW_OPCODE_SEL:
634 if (i == 1) {
635 inst->src[i] = val;
636 progress = true;
637 } else if (i == 0 && inst->src[1].file != IMM) {
638 inst->src[0] = inst->src[1];
639 inst->src[1] = val;
640
641 /* If this was predicated, flipping operands means
642 * we also need to flip the predicate.
643 */
644 if (inst->conditional_mod == BRW_CONDITIONAL_NONE) {
645 inst->predicate_inverse =
646 !inst->predicate_inverse;
647 }
648 progress = true;
649 }
650 break;
651
652 case SHADER_OPCODE_UNTYPED_ATOMIC:
653 case SHADER_OPCODE_UNTYPED_SURFACE_READ:
654 case SHADER_OPCODE_UNTYPED_SURFACE_WRITE:
655 case SHADER_OPCODE_TYPED_ATOMIC:
656 case SHADER_OPCODE_TYPED_SURFACE_READ:
657 case SHADER_OPCODE_TYPED_SURFACE_WRITE:
658 /* We only propagate into the surface argument of the
659 * instruction. Everything else goes through LOAD_PAYLOAD.
660 */
661 if (i == 1) {
662 inst->src[i] = val;
663 progress = true;
664 }
665 break;
666
667 case FS_OPCODE_FB_WRITE_LOGICAL:
668 /* The stencil and omask sources of FS_OPCODE_FB_WRITE_LOGICAL are
669 * bit-cast using a strided region so they cannot be immediates.
670 */
671 if (i != FB_WRITE_LOGICAL_SRC_SRC_STENCIL &&
672 i != FB_WRITE_LOGICAL_SRC_OMASK) {
673 inst->src[i] = val;
674 progress = true;
675 }
676 break;
677
678 case SHADER_OPCODE_TEX_LOGICAL:
679 case SHADER_OPCODE_TXD_LOGICAL:
680 case SHADER_OPCODE_TXF_LOGICAL:
681 case SHADER_OPCODE_TXL_LOGICAL:
682 case SHADER_OPCODE_TXS_LOGICAL:
683 case FS_OPCODE_TXB_LOGICAL:
684 case SHADER_OPCODE_TXF_CMS_LOGICAL:
685 case SHADER_OPCODE_TXF_CMS_W_LOGICAL:
686 case SHADER_OPCODE_TXF_UMS_LOGICAL:
687 case SHADER_OPCODE_TXF_MCS_LOGICAL:
688 case SHADER_OPCODE_LOD_LOGICAL:
689 case SHADER_OPCODE_TG4_LOGICAL:
690 case SHADER_OPCODE_TG4_OFFSET_LOGICAL:
691 case SHADER_OPCODE_UNTYPED_ATOMIC_LOGICAL:
692 case SHADER_OPCODE_UNTYPED_SURFACE_READ_LOGICAL:
693 case SHADER_OPCODE_UNTYPED_SURFACE_WRITE_LOGICAL:
694 case SHADER_OPCODE_TYPED_ATOMIC_LOGICAL:
695 case SHADER_OPCODE_TYPED_SURFACE_READ_LOGICAL:
696 case SHADER_OPCODE_TYPED_SURFACE_WRITE_LOGICAL:
697 inst->src[i] = val;
698 progress = true;
699 break;
700
701 case FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD:
702 case SHADER_OPCODE_BROADCAST:
703 inst->src[i] = val;
704 progress = true;
705 break;
706
707 case BRW_OPCODE_MAD:
708 case BRW_OPCODE_LRP:
709 inst->src[i] = val;
710 progress = true;
711 break;
712
713 default:
714 break;
715 }
716 }
717
718 return progress;
719 }
720
721 static bool
can_propagate_from(fs_inst * inst)722 can_propagate_from(fs_inst *inst)
723 {
724 return (inst->opcode == BRW_OPCODE_MOV &&
725 inst->dst.file == VGRF &&
726 ((inst->src[0].file == VGRF &&
727 !regions_overlap(inst->dst, inst->size_written,
728 inst->src[0], inst->size_read(0))) ||
729 inst->src[0].file == ATTR ||
730 inst->src[0].file == UNIFORM ||
731 inst->src[0].file == IMM) &&
732 inst->src[0].type == inst->dst.type &&
733 !inst->is_partial_write());
734 }
735
736 /* Walks a basic block and does copy propagation on it using the acp
737 * list.
738 */
739 bool
opt_copy_propagation_local(void * copy_prop_ctx,bblock_t * block,exec_list * acp)740 fs_visitor::opt_copy_propagation_local(void *copy_prop_ctx, bblock_t *block,
741 exec_list *acp)
742 {
743 bool progress = false;
744
745 foreach_inst_in_block(fs_inst, inst, block) {
746 /* Try propagating into this instruction. */
747 for (int i = 0; i < inst->sources; i++) {
748 if (inst->src[i].file != VGRF)
749 continue;
750
751 foreach_in_list(acp_entry, entry, &acp[inst->src[i].nr % ACP_HASH_SIZE]) {
752 if (try_constant_propagate(inst, entry))
753 progress = true;
754 else if (try_copy_propagate(inst, i, entry))
755 progress = true;
756 }
757 }
758
759 /* kill the destination from the ACP */
760 if (inst->dst.file == VGRF) {
761 foreach_in_list_safe(acp_entry, entry, &acp[inst->dst.nr % ACP_HASH_SIZE]) {
762 if (regions_overlap(entry->dst, entry->size_written,
763 inst->dst, inst->size_written))
764 entry->remove();
765 }
766
767 /* Oops, we only have the chaining hash based on the destination, not
768 * the source, so walk across the entire table.
769 */
770 for (int i = 0; i < ACP_HASH_SIZE; i++) {
771 foreach_in_list_safe(acp_entry, entry, &acp[i]) {
772 /* Make sure we kill the entry if this instruction overwrites
773 * _any_ of the registers that it reads
774 */
775 if (regions_overlap(entry->src, entry->size_read,
776 inst->dst, inst->size_written))
777 entry->remove();
778 }
779 }
780 }
781
782 /* If this instruction's source could potentially be folded into the
783 * operand of another instruction, add it to the ACP.
784 */
785 if (can_propagate_from(inst)) {
786 acp_entry *entry = ralloc(copy_prop_ctx, acp_entry);
787 entry->dst = inst->dst;
788 entry->src = inst->src[0];
789 entry->size_written = inst->size_written;
790 entry->size_read = inst->size_read(0);
791 entry->opcode = inst->opcode;
792 entry->saturate = inst->saturate;
793 acp[entry->dst.nr % ACP_HASH_SIZE].push_tail(entry);
794 } else if (inst->opcode == SHADER_OPCODE_LOAD_PAYLOAD &&
795 inst->dst.file == VGRF) {
796 int offset = 0;
797 for (int i = 0; i < inst->sources; i++) {
798 int effective_width = i < inst->header_size ? 8 : inst->exec_size;
799 assert(effective_width * type_sz(inst->src[i].type) % REG_SIZE == 0);
800 const unsigned size_written = effective_width *
801 type_sz(inst->src[i].type);
802 if (inst->src[i].file == VGRF) {
803 acp_entry *entry = rzalloc(copy_prop_ctx, acp_entry);
804 entry->dst = byte_offset(inst->dst, offset);
805 entry->src = inst->src[i];
806 entry->size_written = size_written;
807 entry->size_read = inst->size_read(i);
808 entry->opcode = inst->opcode;
809 if (!entry->dst.equals(inst->src[i])) {
810 acp[entry->dst.nr % ACP_HASH_SIZE].push_tail(entry);
811 } else {
812 ralloc_free(entry);
813 }
814 }
815 offset += size_written;
816 }
817 }
818 }
819
820 return progress;
821 }
822
823 bool
opt_copy_propagation()824 fs_visitor::opt_copy_propagation()
825 {
826 bool progress = false;
827 void *copy_prop_ctx = ralloc_context(NULL);
828 exec_list *out_acp[cfg->num_blocks];
829
830 for (int i = 0; i < cfg->num_blocks; i++)
831 out_acp[i] = new exec_list [ACP_HASH_SIZE];
832
833 /* First, walk through each block doing local copy propagation and getting
834 * the set of copies available at the end of the block.
835 */
836 foreach_block (block, cfg) {
837 progress = opt_copy_propagation_local(copy_prop_ctx, block,
838 out_acp[block->num]) || progress;
839 }
840
841 /* Do dataflow analysis for those available copies. */
842 fs_copy_prop_dataflow dataflow(copy_prop_ctx, cfg, out_acp);
843
844 /* Next, re-run local copy propagation, this time with the set of copies
845 * provided by the dataflow analysis available at the start of a block.
846 */
847 foreach_block (block, cfg) {
848 exec_list in_acp[ACP_HASH_SIZE];
849
850 for (int i = 0; i < dataflow.num_acp; i++) {
851 if (BITSET_TEST(dataflow.bd[block->num].livein, i)) {
852 struct acp_entry *entry = dataflow.acp[i];
853 in_acp[entry->dst.nr % ACP_HASH_SIZE].push_tail(entry);
854 }
855 }
856
857 progress = opt_copy_propagation_local(copy_prop_ctx, block, in_acp) ||
858 progress;
859 }
860
861 for (int i = 0; i < cfg->num_blocks; i++)
862 delete [] out_acp[i];
863 ralloc_free(copy_prop_ctx);
864
865 if (progress)
866 invalidate_live_intervals();
867
868 return progress;
869 }
870