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_fs_live_variables.h"
40#include "brw_cfg.h"
41#include "brw_eu.h"
42
43using namespace brw;
44
45namespace { /* avoid conflict with opt_copy_propagation_elements */
46struct acp_entry : public exec_node {
47   fs_reg dst;
48   fs_reg src;
49   uint8_t size_written;
50   uint8_t size_read;
51   enum opcode opcode;
52   bool saturate;
53};
54
55struct block_data {
56   /**
57    * Which entries in the fs_copy_prop_dataflow acp table are live at the
58    * start of this block.  This is the useful output of the analysis, since
59    * it lets us plug those into the local copy propagation on the second
60    * pass.
61    */
62   BITSET_WORD *livein;
63
64   /**
65    * Which entries in the fs_copy_prop_dataflow acp table are live at the end
66    * of this block.  This is done in initial setup from the per-block acps
67    * returned by the first local copy prop pass.
68    */
69   BITSET_WORD *liveout;
70
71   /**
72    * Which entries in the fs_copy_prop_dataflow acp table are generated by
73    * instructions in this block which reach the end of the block without
74    * being killed.
75    */
76   BITSET_WORD *copy;
77
78   /**
79    * Which entries in the fs_copy_prop_dataflow acp table are killed over the
80    * course of this block.
81    */
82   BITSET_WORD *kill;
83
84   /**
85    * Which entries in the fs_copy_prop_dataflow acp table are guaranteed to
86    * have a fully uninitialized destination at the end of this block.
87    */
88   BITSET_WORD *undef;
89};
90
91class fs_copy_prop_dataflow
92{
93public:
94   fs_copy_prop_dataflow(void *mem_ctx, cfg_t *cfg,
95                         const fs_live_variables *live,
96                         exec_list *out_acp[ACP_HASH_SIZE]);
97
98   void setup_initial_values();
99   void run();
100
101   void dump_block_data() const UNUSED;
102
103   void *mem_ctx;
104   cfg_t *cfg;
105   const fs_live_variables *live;
106
107   acp_entry **acp;
108   int num_acp;
109   int bitset_words;
110
111  struct block_data *bd;
112};
113} /* anonymous namespace */
114
115fs_copy_prop_dataflow::fs_copy_prop_dataflow(void *mem_ctx, cfg_t *cfg,
116                                             const fs_live_variables *live,
117                                             exec_list *out_acp[ACP_HASH_SIZE])
118   : mem_ctx(mem_ctx), cfg(cfg), live(live)
119{
120   bd = rzalloc_array(mem_ctx, struct block_data, cfg->num_blocks);
121
122   num_acp = 0;
123   foreach_block (block, cfg) {
124      for (int i = 0; i < ACP_HASH_SIZE; i++) {
125         num_acp += out_acp[block->num][i].length();
126      }
127   }
128
129   acp = rzalloc_array(mem_ctx, struct acp_entry *, num_acp);
130
131   bitset_words = BITSET_WORDS(num_acp);
132
133   int next_acp = 0;
134   foreach_block (block, cfg) {
135      bd[block->num].livein = rzalloc_array(bd, BITSET_WORD, bitset_words);
136      bd[block->num].liveout = rzalloc_array(bd, BITSET_WORD, bitset_words);
137      bd[block->num].copy = rzalloc_array(bd, BITSET_WORD, bitset_words);
138      bd[block->num].kill = rzalloc_array(bd, BITSET_WORD, bitset_words);
139      bd[block->num].undef = rzalloc_array(bd, BITSET_WORD, bitset_words);
140
141      for (int i = 0; i < ACP_HASH_SIZE; i++) {
142         foreach_in_list(acp_entry, entry, &out_acp[block->num][i]) {
143            acp[next_acp] = entry;
144
145            /* opt_copy_propagation_local populates out_acp with copies created
146             * in a block which are still live at the end of the block.  This
147             * is exactly what we want in the COPY set.
148             */
149            BITSET_SET(bd[block->num].copy, next_acp);
150
151            next_acp++;
152         }
153      }
154   }
155
156   assert(next_acp == num_acp);
157
158   setup_initial_values();
159   run();
160}
161
162/**
163 * Set up initial values for each of the data flow sets, prior to running
164 * the fixed-point algorithm.
165 */
166void
167fs_copy_prop_dataflow::setup_initial_values()
168{
169   /* Initialize the COPY and KILL sets. */
170   foreach_block (block, cfg) {
171      foreach_inst_in_block(fs_inst, inst, block) {
172         if (inst->dst.file != VGRF)
173            continue;
174
175         /* Mark ACP entries which are killed by this instruction. */
176         for (int i = 0; i < num_acp; i++) {
177            if (regions_overlap(inst->dst, inst->size_written,
178                                acp[i]->dst, acp[i]->size_written) ||
179                regions_overlap(inst->dst, inst->size_written,
180                                acp[i]->src, acp[i]->size_read)) {
181               BITSET_SET(bd[block->num].kill, i);
182            }
183         }
184      }
185   }
186
187   /* Populate the initial values for the livein and liveout sets.  For the
188    * block at the start of the program, livein = 0 and liveout = copy.
189    * For the others, set liveout and livein to ~0 (the universal set).
190    */
191   foreach_block (block, cfg) {
192      if (block->parents.is_empty()) {
193         for (int i = 0; i < bitset_words; i++) {
194            bd[block->num].livein[i] = 0u;
195            bd[block->num].liveout[i] = bd[block->num].copy[i];
196         }
197      } else {
198         for (int i = 0; i < bitset_words; i++) {
199            bd[block->num].liveout[i] = ~0u;
200            bd[block->num].livein[i] = ~0u;
201         }
202      }
203   }
204
205   /* Initialize the undef set. */
206   foreach_block (block, cfg) {
207      for (int i = 0; i < num_acp; i++) {
208         BITSET_SET(bd[block->num].undef, i);
209         for (unsigned off = 0; off < acp[i]->size_written; off += REG_SIZE) {
210            if (BITSET_TEST(live->block_data[block->num].defout,
211                            live->var_from_reg(byte_offset(acp[i]->dst, off))))
212               BITSET_CLEAR(bd[block->num].undef, i);
213         }
214      }
215   }
216}
217
218/**
219 * Walk the set of instructions in the block, marking which entries in the acp
220 * are killed by the block.
221 */
222void
223fs_copy_prop_dataflow::run()
224{
225   bool progress;
226
227   do {
228      progress = false;
229
230      foreach_block (block, cfg) {
231         if (block->parents.is_empty())
232            continue;
233
234         for (int i = 0; i < bitset_words; i++) {
235            const BITSET_WORD old_liveout = bd[block->num].liveout[i];
236            BITSET_WORD livein_from_any_block = 0;
237
238            /* Update livein for this block.  If a copy is live out of all
239             * parent blocks, it's live coming in to this block.
240             */
241            bd[block->num].livein[i] = ~0u;
242            foreach_list_typed(bblock_link, parent_link, link, &block->parents) {
243               bblock_t *parent = parent_link->block;
244               /* Consider ACP entries with a known-undefined destination to
245                * be available from the parent.  This is valid because we're
246                * free to set the undefined variable equal to the source of
247                * the ACP entry without breaking the application's
248                * expectations, since the variable is undefined.
249                */
250               bd[block->num].livein[i] &= (bd[parent->num].liveout[i] |
251                                            bd[parent->num].undef[i]);
252               livein_from_any_block |= bd[parent->num].liveout[i];
253            }
254
255            /* Limit to the set of ACP entries that can possibly be available
256             * at the start of the block, since propagating from a variable
257             * which is guaranteed to be undefined (rather than potentially
258             * undefined for some dynamic control-flow paths) doesn't seem
259             * particularly useful.
260             */
261            bd[block->num].livein[i] &= livein_from_any_block;
262
263            /* Update liveout for this block. */
264            bd[block->num].liveout[i] =
265               bd[block->num].copy[i] | (bd[block->num].livein[i] &
266                                         ~bd[block->num].kill[i]);
267
268            if (old_liveout != bd[block->num].liveout[i])
269               progress = true;
270         }
271      }
272   } while (progress);
273}
274
275void
276fs_copy_prop_dataflow::dump_block_data() const
277{
278   foreach_block (block, cfg) {
279      fprintf(stderr, "Block %d [%d, %d] (parents ", block->num,
280             block->start_ip, block->end_ip);
281      foreach_list_typed(bblock_link, link, link, &block->parents) {
282         bblock_t *parent = link->block;
283         fprintf(stderr, "%d ", parent->num);
284      }
285      fprintf(stderr, "):\n");
286      fprintf(stderr, "       livein = 0x");
287      for (int i = 0; i < bitset_words; i++)
288         fprintf(stderr, "%08x", bd[block->num].livein[i]);
289      fprintf(stderr, ", liveout = 0x");
290      for (int i = 0; i < bitset_words; i++)
291         fprintf(stderr, "%08x", bd[block->num].liveout[i]);
292      fprintf(stderr, ",\n       copy   = 0x");
293      for (int i = 0; i < bitset_words; i++)
294         fprintf(stderr, "%08x", bd[block->num].copy[i]);
295      fprintf(stderr, ", kill    = 0x");
296      for (int i = 0; i < bitset_words; i++)
297         fprintf(stderr, "%08x", bd[block->num].kill[i]);
298      fprintf(stderr, "\n");
299   }
300}
301
302static bool
303is_logic_op(enum opcode opcode)
304{
305   return (opcode == BRW_OPCODE_AND ||
306           opcode == BRW_OPCODE_OR  ||
307           opcode == BRW_OPCODE_XOR ||
308           opcode == BRW_OPCODE_NOT);
309}
310
311static bool
312can_take_stride(fs_inst *inst, unsigned arg, unsigned stride,
313                const gen_device_info *devinfo)
314{
315   if (stride > 4)
316      return false;
317
318   /* Bail if the channels of the source need to be aligned to the byte offset
319    * of the corresponding channel of the destination, and the provided stride
320    * would break this restriction.
321    */
322   if (has_dst_aligned_region_restriction(devinfo, inst) &&
323       !(type_sz(inst->src[arg].type) * stride ==
324           type_sz(inst->dst.type) * inst->dst.stride ||
325         stride == 0))
326      return false;
327
328   /* 3-source instructions can only be Align16, which restricts what strides
329    * they can take. They can only take a stride of 1 (the usual case), or 0
330    * with a special "repctrl" bit. But the repctrl bit doesn't work for
331    * 64-bit datatypes, so if the source type is 64-bit then only a stride of
332    * 1 is allowed. From the Broadwell PRM, Volume 7 "3D Media GPGPU", page
333    * 944:
334    *
335    *    This is applicable to 32b datatypes and 16b datatype. 64b datatypes
336    *    cannot use the replicate control.
337    */
338   if (inst->is_3src(devinfo)) {
339      if (type_sz(inst->src[arg].type) > 4)
340         return stride == 1;
341      else
342         return stride == 1 || stride == 0;
343   }
344
345   /* From the Broadwell PRM, Volume 2a "Command Reference - Instructions",
346    * page 391 ("Extended Math Function"):
347    *
348    *     The following restrictions apply for align1 mode: Scalar source is
349    *     supported. Source and destination horizontal stride must be the
350    *     same.
351    *
352    * From the Haswell PRM Volume 2b "Command Reference - Instructions", page
353    * 134 ("Extended Math Function"):
354    *
355    *    Scalar source is supported. Source and destination horizontal stride
356    *    must be 1.
357    *
358    * and similar language exists for IVB and SNB. Pre-SNB, math instructions
359    * are sends, so the sources are moved to MRF's and there are no
360    * restrictions.
361    */
362   if (inst->is_math()) {
363      if (devinfo->gen == 6 || devinfo->gen == 7) {
364         assert(inst->dst.stride == 1);
365         return stride == 1 || stride == 0;
366      } else if (devinfo->gen >= 8) {
367         return stride == inst->dst.stride || stride == 0;
368      }
369   }
370
371   return true;
372}
373
374static bool
375instruction_requires_packed_data(fs_inst *inst)
376{
377   switch (inst->opcode) {
378   case FS_OPCODE_DDX_FINE:
379   case FS_OPCODE_DDX_COARSE:
380   case FS_OPCODE_DDY_FINE:
381   case FS_OPCODE_DDY_COARSE:
382      return true;
383   default:
384      return false;
385   }
386}
387
388bool
389fs_visitor::try_copy_propagate(fs_inst *inst, int arg, acp_entry *entry)
390{
391   if (inst->src[arg].file != VGRF)
392      return false;
393
394   if (entry->src.file == IMM)
395      return false;
396   assert(entry->src.file == VGRF || entry->src.file == UNIFORM ||
397          entry->src.file == ATTR);
398
399   if (entry->opcode == SHADER_OPCODE_LOAD_PAYLOAD &&
400       inst->opcode == SHADER_OPCODE_LOAD_PAYLOAD)
401      return false;
402
403   assert(entry->dst.file == VGRF);
404   if (inst->src[arg].nr != entry->dst.nr)
405      return false;
406
407   /* Bail if inst is reading a range that isn't contained in the range
408    * that entry is writing.
409    */
410   if (!region_contained_in(inst->src[arg], inst->size_read(arg),
411                            entry->dst, entry->size_written))
412      return false;
413
414   /* we can't generally copy-propagate UD negations because we
415    * can end up accessing the resulting values as signed integers
416    * instead. See also resolve_ud_negate() and comment in
417    * fs_generator::generate_code.
418    */
419   if (entry->src.type == BRW_REGISTER_TYPE_UD &&
420       entry->src.negate)
421      return false;
422
423   bool has_source_modifiers = entry->src.abs || entry->src.negate;
424
425   if ((has_source_modifiers || entry->src.file == UNIFORM ||
426        !entry->src.is_contiguous()) &&
427       !inst->can_do_source_mods(devinfo))
428      return false;
429
430   if (has_source_modifiers &&
431       inst->opcode == SHADER_OPCODE_GEN4_SCRATCH_WRITE)
432      return false;
433
434   /* Some instructions implemented in the generator backend, such as
435    * derivatives, assume that their operands are packed so we can't
436    * generally propagate strided regions to them.
437    */
438   if (instruction_requires_packed_data(inst) && entry->src.stride > 1)
439      return false;
440
441   /* Bail if the result of composing both strides would exceed the
442    * hardware limit.
443    */
444   if (!can_take_stride(inst, arg, entry->src.stride * inst->src[arg].stride,
445                        devinfo))
446      return false;
447
448   /* Bail if the instruction type is larger than the execution type of the
449    * copy, what implies that each channel is reading multiple channels of the
450    * destination of the copy, and simply replacing the sources would give a
451    * program with different semantics.
452    */
453   if (type_sz(entry->dst.type) < type_sz(inst->src[arg].type))
454      return false;
455
456   /* Bail if the result of composing both strides cannot be expressed
457    * as another stride. This avoids, for example, trying to transform
458    * this:
459    *
460    *     MOV (8) rX<1>UD rY<0;1,0>UD
461    *     FOO (8) ...     rX<8;8,1>UW
462    *
463    * into this:
464    *
465    *     FOO (8) ...     rY<0;1,0>UW
466    *
467    * Which would have different semantics.
468    */
469   if (entry->src.stride != 1 &&
470       (inst->src[arg].stride *
471        type_sz(inst->src[arg].type)) % type_sz(entry->src.type) != 0)
472      return false;
473
474   /* Since semantics of source modifiers are type-dependent we need to
475    * ensure that the meaning of the instruction remains the same if we
476    * change the type. If the sizes of the types are different the new
477    * instruction will read a different amount of data than the original
478    * and the semantics will always be different.
479    */
480   if (has_source_modifiers &&
481       entry->dst.type != inst->src[arg].type &&
482       (!inst->can_change_types() ||
483        type_sz(entry->dst.type) != type_sz(inst->src[arg].type)))
484      return false;
485
486   if (devinfo->gen >= 8 && (entry->src.negate || entry->src.abs) &&
487       is_logic_op(inst->opcode)) {
488      return false;
489   }
490
491   if (entry->saturate) {
492      switch(inst->opcode) {
493      case BRW_OPCODE_SEL:
494         if ((inst->conditional_mod != BRW_CONDITIONAL_GE &&
495              inst->conditional_mod != BRW_CONDITIONAL_L) ||
496             inst->src[1].file != IMM ||
497             inst->src[1].f < 0.0 ||
498             inst->src[1].f > 1.0) {
499            return false;
500         }
501         break;
502      default:
503         return false;
504      }
505   }
506
507   inst->src[arg].file = entry->src.file;
508   inst->src[arg].nr = entry->src.nr;
509   inst->src[arg].stride *= entry->src.stride;
510   inst->saturate = inst->saturate || entry->saturate;
511
512   /* Compute the offset of inst->src[arg] relative to entry->dst */
513   const unsigned rel_offset = inst->src[arg].offset - entry->dst.offset;
514
515   /* Compute the first component of the copy that the instruction is
516    * reading, and the base byte offset within that component.
517    */
518   assert(entry->dst.offset % REG_SIZE == 0 && entry->dst.stride == 1);
519   const unsigned component = rel_offset / type_sz(entry->dst.type);
520   const unsigned suboffset = rel_offset % type_sz(entry->dst.type);
521
522   /* Calculate the byte offset at the origin of the copy of the given
523    * component and suboffset.
524    */
525   inst->src[arg].offset = suboffset +
526      component * entry->src.stride * type_sz(entry->src.type) +
527      entry->src.offset;
528
529   if (has_source_modifiers) {
530      if (entry->dst.type != inst->src[arg].type) {
531         /* We are propagating source modifiers from a MOV with a different
532          * type.  If we got here, then we can just change the source and
533          * destination types of the instruction and keep going.
534          */
535         assert(inst->can_change_types());
536         for (int i = 0; i < inst->sources; i++) {
537            inst->src[i].type = entry->dst.type;
538         }
539         inst->dst.type = entry->dst.type;
540      }
541
542      if (!inst->src[arg].abs) {
543         inst->src[arg].abs = entry->src.abs;
544         inst->src[arg].negate ^= entry->src.negate;
545      }
546   }
547
548   return true;
549}
550
551
552bool
553fs_visitor::try_constant_propagate(fs_inst *inst, acp_entry *entry)
554{
555   bool progress = false;
556
557   if (entry->src.file != IMM)
558      return false;
559   if (type_sz(entry->src.type) > 4)
560      return false;
561   if (entry->saturate)
562      return false;
563
564   for (int i = inst->sources - 1; i >= 0; i--) {
565      if (inst->src[i].file != VGRF)
566         continue;
567
568      assert(entry->dst.file == VGRF);
569      if (inst->src[i].nr != entry->dst.nr)
570         continue;
571
572      /* Bail if inst is reading a range that isn't contained in the range
573       * that entry is writing.
574       */
575      if (!region_contained_in(inst->src[i], inst->size_read(i),
576                               entry->dst, entry->size_written))
577         continue;
578
579      /* If the type sizes don't match each channel of the instruction is
580       * either extracting a portion of the constant (which could be handled
581       * with some effort but the code below doesn't) or reading multiple
582       * channels of the source at once.
583       */
584      if (type_sz(inst->src[i].type) != type_sz(entry->dst.type))
585         continue;
586
587      fs_reg val = entry->src;
588      val.type = inst->src[i].type;
589
590      if (inst->src[i].abs) {
591         if ((devinfo->gen >= 8 && is_logic_op(inst->opcode)) ||
592             !brw_abs_immediate(val.type, &val.as_brw_reg())) {
593            continue;
594         }
595      }
596
597      if (inst->src[i].negate) {
598         if ((devinfo->gen >= 8 && is_logic_op(inst->opcode)) ||
599             !brw_negate_immediate(val.type, &val.as_brw_reg())) {
600            continue;
601         }
602      }
603
604      switch (inst->opcode) {
605      case BRW_OPCODE_MOV:
606      case SHADER_OPCODE_LOAD_PAYLOAD:
607      case FS_OPCODE_PACK:
608         inst->src[i] = val;
609         progress = true;
610         break;
611
612      case SHADER_OPCODE_INT_QUOTIENT:
613      case SHADER_OPCODE_INT_REMAINDER:
614         /* FINISHME: Promote non-float constants and remove this. */
615         if (devinfo->gen < 8)
616            break;
617         /* fallthrough */
618      case SHADER_OPCODE_POW:
619         /* Allow constant propagation into src1 (except on Gen 6 which
620          * doesn't support scalar source math), and let constant combining
621          * promote the constant on Gen < 8.
622          */
623         if (devinfo->gen == 6)
624            break;
625         /* fallthrough */
626      case BRW_OPCODE_BFI1:
627      case BRW_OPCODE_ASR:
628      case BRW_OPCODE_SHL:
629      case BRW_OPCODE_SHR:
630      case BRW_OPCODE_SUBB:
631         if (i == 1) {
632            inst->src[i] = val;
633            progress = true;
634         }
635         break;
636
637      case BRW_OPCODE_MACH:
638      case BRW_OPCODE_MUL:
639      case SHADER_OPCODE_MULH:
640      case BRW_OPCODE_ADD:
641      case BRW_OPCODE_OR:
642      case BRW_OPCODE_AND:
643      case BRW_OPCODE_XOR:
644      case BRW_OPCODE_ADDC:
645         if (i == 1) {
646            inst->src[i] = val;
647            progress = true;
648         } else if (i == 0 && inst->src[1].file != IMM) {
649            /* Fit this constant in by commuting the operands.
650             * Exception: we can't do this for 32-bit integer MUL/MACH
651             * because it's asymmetric.
652             *
653             * The BSpec says for Broadwell that
654             *
655             *    "When multiplying DW x DW, the dst cannot be accumulator."
656             *
657             * Integer MUL with a non-accumulator destination will be lowered
658             * by lower_integer_multiplication(), so don't restrict it.
659             */
660            if (((inst->opcode == BRW_OPCODE_MUL &&
661                  inst->dst.is_accumulator()) ||
662                 inst->opcode == BRW_OPCODE_MACH) &&
663                (inst->src[1].type == BRW_REGISTER_TYPE_D ||
664                 inst->src[1].type == BRW_REGISTER_TYPE_UD))
665               break;
666            inst->src[0] = inst->src[1];
667            inst->src[1] = val;
668            progress = true;
669         }
670         break;
671
672      case BRW_OPCODE_CMP:
673      case BRW_OPCODE_IF:
674         if (i == 1) {
675            inst->src[i] = val;
676            progress = true;
677         } else if (i == 0 && inst->src[1].file != IMM) {
678            enum brw_conditional_mod new_cmod;
679
680            new_cmod = brw_swap_cmod(inst->conditional_mod);
681            if (new_cmod != BRW_CONDITIONAL_NONE) {
682               /* Fit this constant in by swapping the operands and
683                * flipping the test
684                */
685               inst->src[0] = inst->src[1];
686               inst->src[1] = val;
687               inst->conditional_mod = new_cmod;
688               progress = true;
689            }
690         }
691         break;
692
693      case BRW_OPCODE_SEL:
694         if (i == 1) {
695            inst->src[i] = val;
696            progress = true;
697         } else if (i == 0 && inst->src[1].file != IMM) {
698            inst->src[0] = inst->src[1];
699            inst->src[1] = val;
700
701            /* If this was predicated, flipping operands means
702             * we also need to flip the predicate.
703             */
704            if (inst->conditional_mod == BRW_CONDITIONAL_NONE) {
705               inst->predicate_inverse =
706                  !inst->predicate_inverse;
707            }
708            progress = true;
709         }
710         break;
711
712      case FS_OPCODE_FB_WRITE_LOGICAL:
713         /* The stencil and omask sources of FS_OPCODE_FB_WRITE_LOGICAL are
714          * bit-cast using a strided region so they cannot be immediates.
715          */
716         if (i != FB_WRITE_LOGICAL_SRC_SRC_STENCIL &&
717             i != FB_WRITE_LOGICAL_SRC_OMASK) {
718            inst->src[i] = val;
719            progress = true;
720         }
721         break;
722
723      case SHADER_OPCODE_TEX_LOGICAL:
724      case SHADER_OPCODE_TXD_LOGICAL:
725      case SHADER_OPCODE_TXF_LOGICAL:
726      case SHADER_OPCODE_TXL_LOGICAL:
727      case SHADER_OPCODE_TXS_LOGICAL:
728      case FS_OPCODE_TXB_LOGICAL:
729      case SHADER_OPCODE_TXF_CMS_LOGICAL:
730      case SHADER_OPCODE_TXF_CMS_W_LOGICAL:
731      case SHADER_OPCODE_TXF_UMS_LOGICAL:
732      case SHADER_OPCODE_TXF_MCS_LOGICAL:
733      case SHADER_OPCODE_LOD_LOGICAL:
734      case SHADER_OPCODE_TG4_LOGICAL:
735      case SHADER_OPCODE_TG4_OFFSET_LOGICAL:
736      case SHADER_OPCODE_UNTYPED_ATOMIC_LOGICAL:
737      case SHADER_OPCODE_UNTYPED_ATOMIC_FLOAT_LOGICAL:
738      case SHADER_OPCODE_UNTYPED_SURFACE_READ_LOGICAL:
739      case SHADER_OPCODE_UNTYPED_SURFACE_WRITE_LOGICAL:
740      case SHADER_OPCODE_TYPED_ATOMIC_LOGICAL:
741      case SHADER_OPCODE_TYPED_SURFACE_READ_LOGICAL:
742      case SHADER_OPCODE_TYPED_SURFACE_WRITE_LOGICAL:
743      case SHADER_OPCODE_BYTE_SCATTERED_WRITE_LOGICAL:
744      case SHADER_OPCODE_BYTE_SCATTERED_READ_LOGICAL:
745         inst->src[i] = val;
746         progress = true;
747         break;
748
749      case FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD:
750      case SHADER_OPCODE_BROADCAST:
751         inst->src[i] = val;
752         progress = true;
753         break;
754
755      case BRW_OPCODE_MAD:
756      case BRW_OPCODE_LRP:
757         inst->src[i] = val;
758         progress = true;
759         break;
760
761      default:
762         break;
763      }
764   }
765
766   return progress;
767}
768
769static bool
770can_propagate_from(fs_inst *inst)
771{
772   return (inst->opcode == BRW_OPCODE_MOV &&
773           inst->dst.file == VGRF &&
774           ((inst->src[0].file == VGRF &&
775             !regions_overlap(inst->dst, inst->size_written,
776                              inst->src[0], inst->size_read(0))) ||
777            inst->src[0].file == ATTR ||
778            inst->src[0].file == UNIFORM ||
779            inst->src[0].file == IMM) &&
780           inst->src[0].type == inst->dst.type &&
781           !inst->is_partial_write());
782}
783
784/* Walks a basic block and does copy propagation on it using the acp
785 * list.
786 */
787bool
788fs_visitor::opt_copy_propagation_local(void *copy_prop_ctx, bblock_t *block,
789                                       exec_list *acp)
790{
791   bool progress = false;
792
793   foreach_inst_in_block(fs_inst, inst, block) {
794      /* Try propagating into this instruction. */
795      for (int i = 0; i < inst->sources; i++) {
796         if (inst->src[i].file != VGRF)
797            continue;
798
799         foreach_in_list(acp_entry, entry, &acp[inst->src[i].nr % ACP_HASH_SIZE]) {
800            if (try_constant_propagate(inst, entry))
801               progress = true;
802            else if (try_copy_propagate(inst, i, entry))
803               progress = true;
804         }
805      }
806
807      /* kill the destination from the ACP */
808      if (inst->dst.file == VGRF) {
809         foreach_in_list_safe(acp_entry, entry, &acp[inst->dst.nr % ACP_HASH_SIZE]) {
810            if (regions_overlap(entry->dst, entry->size_written,
811                                inst->dst, inst->size_written))
812               entry->remove();
813         }
814
815         /* Oops, we only have the chaining hash based on the destination, not
816          * the source, so walk across the entire table.
817          */
818         for (int i = 0; i < ACP_HASH_SIZE; i++) {
819            foreach_in_list_safe(acp_entry, entry, &acp[i]) {
820               /* Make sure we kill the entry if this instruction overwrites
821                * _any_ of the registers that it reads
822                */
823               if (regions_overlap(entry->src, entry->size_read,
824                                   inst->dst, inst->size_written))
825                  entry->remove();
826            }
827	 }
828      }
829
830      /* If this instruction's source could potentially be folded into the
831       * operand of another instruction, add it to the ACP.
832       */
833      if (can_propagate_from(inst)) {
834         acp_entry *entry = ralloc(copy_prop_ctx, acp_entry);
835         entry->dst = inst->dst;
836         entry->src = inst->src[0];
837         entry->size_written = inst->size_written;
838         entry->size_read = inst->size_read(0);
839         entry->opcode = inst->opcode;
840         entry->saturate = inst->saturate;
841         acp[entry->dst.nr % ACP_HASH_SIZE].push_tail(entry);
842      } else if (inst->opcode == SHADER_OPCODE_LOAD_PAYLOAD &&
843                 inst->dst.file == VGRF) {
844         int offset = 0;
845         for (int i = 0; i < inst->sources; i++) {
846            int effective_width = i < inst->header_size ? 8 : inst->exec_size;
847            assert(effective_width * type_sz(inst->src[i].type) % REG_SIZE == 0);
848            const unsigned size_written = effective_width *
849                                          type_sz(inst->src[i].type);
850            if (inst->src[i].file == VGRF) {
851               acp_entry *entry = rzalloc(copy_prop_ctx, acp_entry);
852               entry->dst = byte_offset(inst->dst, offset);
853               entry->src = inst->src[i];
854               entry->size_written = size_written;
855               entry->size_read = inst->size_read(i);
856               entry->opcode = inst->opcode;
857               if (!entry->dst.equals(inst->src[i])) {
858                  acp[entry->dst.nr % ACP_HASH_SIZE].push_tail(entry);
859               } else {
860                  ralloc_free(entry);
861               }
862            }
863            offset += size_written;
864         }
865      }
866   }
867
868   return progress;
869}
870
871bool
872fs_visitor::opt_copy_propagation()
873{
874   bool progress = false;
875   void *copy_prop_ctx = ralloc_context(NULL);
876   exec_list *out_acp[cfg->num_blocks];
877
878   for (int i = 0; i < cfg->num_blocks; i++)
879      out_acp[i] = new exec_list [ACP_HASH_SIZE];
880
881   calculate_live_intervals();
882
883   /* First, walk through each block doing local copy propagation and getting
884    * the set of copies available at the end of the block.
885    */
886   foreach_block (block, cfg) {
887      progress = opt_copy_propagation_local(copy_prop_ctx, block,
888                                            out_acp[block->num]) || progress;
889   }
890
891   /* Do dataflow analysis for those available copies. */
892   fs_copy_prop_dataflow dataflow(copy_prop_ctx, cfg, live_intervals, out_acp);
893
894   /* Next, re-run local copy propagation, this time with the set of copies
895    * provided by the dataflow analysis available at the start of a block.
896    */
897   foreach_block (block, cfg) {
898      exec_list in_acp[ACP_HASH_SIZE];
899
900      for (int i = 0; i < dataflow.num_acp; i++) {
901         if (BITSET_TEST(dataflow.bd[block->num].livein, i)) {
902            struct acp_entry *entry = dataflow.acp[i];
903            in_acp[entry->dst.nr % ACP_HASH_SIZE].push_tail(entry);
904         }
905      }
906
907      progress = opt_copy_propagation_local(copy_prop_ctx, block, in_acp) ||
908                 progress;
909   }
910
911   for (int i = 0; i < cfg->num_blocks; i++)
912      delete [] out_acp[i];
913   ralloc_free(copy_prop_ctx);
914
915   if (progress)
916      invalidate_live_intervals();
917
918   return progress;
919}
920