xref: /aosp_15_r20/external/mesa3d/src/intel/compiler/brw_fs_copy_propagation.cpp (revision 6104692788411f58d303aa86923a9ff6ecaded22)
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
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 #include "util/bitset.h"
36 #include "util/u_math.h"
37 #include "util/rb_tree.h"
38 #include "brw_fs.h"
39 #include "brw_fs_live_variables.h"
40 #include "brw_cfg.h"
41 #include "brw_eu.h"
42 
43 using namespace brw;
44 
45 namespace { /* avoid conflict with opt_copy_propagation_elements */
46 struct acp_entry {
47    struct rb_node by_dst;
48    struct rb_node by_src;
49    brw_reg dst;
50    brw_reg src;
51    unsigned global_idx;
52    unsigned size_written;
53    unsigned size_read;
54    enum opcode opcode;
55    bool is_partial_write;
56    bool force_writemask_all;
57 };
58 
59 /**
60  * Compare two acp_entry::src.nr
61  *
62  * This is intended to be used as the comparison function for rb_tree.
63  */
64 static int
cmp_entry_dst_entry_dst(const struct rb_node * a_node,const struct rb_node * b_node)65 cmp_entry_dst_entry_dst(const struct rb_node *a_node, const struct rb_node *b_node)
66 {
67    const struct acp_entry *a_entry =
68       rb_node_data(struct acp_entry, a_node, by_dst);
69 
70    const struct acp_entry *b_entry =
71       rb_node_data(struct acp_entry, b_node, by_dst);
72 
73    return a_entry->dst.nr - b_entry->dst.nr;
74 }
75 
76 static int
cmp_entry_dst_nr(const struct rb_node * a_node,const void * b_key)77 cmp_entry_dst_nr(const struct rb_node *a_node, const void *b_key)
78 {
79    const struct acp_entry *a_entry =
80       rb_node_data(struct acp_entry, a_node, by_dst);
81 
82    return a_entry->dst.nr - (uintptr_t) b_key;
83 }
84 
85 static int
cmp_entry_src_entry_src(const struct rb_node * a_node,const struct rb_node * b_node)86 cmp_entry_src_entry_src(const struct rb_node *a_node, const struct rb_node *b_node)
87 {
88    const struct acp_entry *a_entry =
89       rb_node_data(struct acp_entry, a_node, by_src);
90 
91    const struct acp_entry *b_entry =
92       rb_node_data(struct acp_entry, b_node, by_src);
93 
94    return a_entry->src.nr - b_entry->src.nr;
95 }
96 
97 /**
98  * Compare an acp_entry::src.nr with a raw nr.
99  *
100  * This is intended to be used as the comparison function for rb_tree.
101  */
102 static int
cmp_entry_src_nr(const struct rb_node * a_node,const void * b_key)103 cmp_entry_src_nr(const struct rb_node *a_node, const void *b_key)
104 {
105    const struct acp_entry *a_entry =
106       rb_node_data(struct acp_entry, a_node, by_src);
107 
108    return a_entry->src.nr - (uintptr_t) b_key;
109 }
110 
111 class acp_forward_iterator {
112 public:
acp_forward_iterator(struct rb_node * n,unsigned offset)113    acp_forward_iterator(struct rb_node *n, unsigned offset)
114       : curr(n), next(nullptr), offset(offset)
115    {
116       next = rb_node_next_or_null(curr);
117    }
118 
operator ++()119    acp_forward_iterator &operator++()
120    {
121       curr = next;
122       next = rb_node_next_or_null(curr);
123 
124       return *this;
125    }
126 
operator !=(const acp_forward_iterator & other) const127    bool operator!=(const acp_forward_iterator &other) const
128    {
129       return curr != other.curr;
130    }
131 
operator *() const132    struct acp_entry *operator*() const
133    {
134       /* This open-codes part of rb_node_data. */
135       return curr != NULL ? (struct acp_entry *)(((char *)curr) - offset)
136                           : NULL;
137    }
138 
139 private:
140    struct rb_node *curr;
141    struct rb_node *next;
142    unsigned offset;
143 };
144 
145 struct acp {
146    struct rb_tree by_dst;
147    struct rb_tree by_src;
148 
acp__anon439490700111::acp149    acp()
150    {
151       rb_tree_init(&by_dst);
152       rb_tree_init(&by_src);
153    }
154 
begin__anon439490700111::acp155    acp_forward_iterator begin()
156    {
157       return acp_forward_iterator(rb_tree_first(&by_src),
158                                   rb_tree_offsetof(struct acp_entry, by_src, 0));
159    }
160 
end__anon439490700111::acp161    const acp_forward_iterator end() const
162    {
163       return acp_forward_iterator(nullptr, 0);
164    }
165 
length__anon439490700111::acp166    unsigned length()
167    {
168       unsigned l = 0;
169 
170       for (rb_node *iter = rb_tree_first(&by_src);
171            iter != NULL; iter = rb_node_next(iter))
172          l++;
173 
174       return l;
175    }
176 
add__anon439490700111::acp177    void add(acp_entry *entry)
178    {
179       rb_tree_insert(&by_dst, &entry->by_dst, cmp_entry_dst_entry_dst);
180       rb_tree_insert(&by_src, &entry->by_src, cmp_entry_src_entry_src);
181    }
182 
remove__anon439490700111::acp183    void remove(acp_entry *entry)
184    {
185       rb_tree_remove(&by_dst, &entry->by_dst);
186       rb_tree_remove(&by_src, &entry->by_src);
187    }
188 
find_by_src__anon439490700111::acp189    acp_forward_iterator find_by_src(unsigned nr)
190    {
191       struct rb_node *rbn = rb_tree_search(&by_src,
192                                            (void *)(uintptr_t) nr,
193                                            cmp_entry_src_nr);
194 
195       return acp_forward_iterator(rbn, rb_tree_offsetof(struct acp_entry,
196                                                         by_src, rbn));
197    }
198 
find_by_dst__anon439490700111::acp199    acp_forward_iterator find_by_dst(unsigned nr)
200    {
201       struct rb_node *rbn = rb_tree_search(&by_dst,
202                                            (void *)(uintptr_t) nr,
203                                            cmp_entry_dst_nr);
204 
205       return acp_forward_iterator(rbn, rb_tree_offsetof(struct acp_entry,
206                                                         by_dst, rbn));
207    }
208 };
209 
210 struct block_data {
211    /**
212     * Which entries in the fs_copy_prop_dataflow acp table are live at the
213     * start of this block.  This is the useful output of the analysis, since
214     * it lets us plug those into the local copy propagation on the second
215     * pass.
216     */
217    BITSET_WORD *livein;
218 
219    /**
220     * Which entries in the fs_copy_prop_dataflow acp table are live at the end
221     * of this block.  This is done in initial setup from the per-block acps
222     * returned by the first local copy prop pass.
223     */
224    BITSET_WORD *liveout;
225 
226    /**
227     * Which entries in the fs_copy_prop_dataflow acp table are generated by
228     * instructions in this block which reach the end of the block without
229     * being killed.
230     */
231    BITSET_WORD *copy;
232 
233    /**
234     * Which entries in the fs_copy_prop_dataflow acp table are killed over the
235     * course of this block.
236     */
237    BITSET_WORD *kill;
238 
239    /**
240     * Which entries in the fs_copy_prop_dataflow acp table are guaranteed to
241     * have a fully uninitialized destination at the end of this block.
242     */
243    BITSET_WORD *undef;
244 
245    /**
246     * Which entries in the fs_copy_prop_dataflow acp table can the
247     * start of this block be reached from.  Note that this is a weaker
248     * condition than livein.
249     */
250    BITSET_WORD *reachin;
251 
252    /**
253     * Which entries in the fs_copy_prop_dataflow acp table are
254     * overwritten by an instruction with channel masks inconsistent
255     * with the copy instruction (e.g. due to force_writemask_all).
256     * Such an overwrite can cause the copy entry to become invalid
257     * even if the copy instruction is subsequently re-executed for any
258     * given channel i, since the execution of the overwrite for
259     * channel i may corrupt other channels j!=i inactive for the
260     * subsequent copy.
261     */
262    BITSET_WORD *exec_mismatch;
263 };
264 
265 class fs_copy_prop_dataflow
266 {
267 public:
268    fs_copy_prop_dataflow(linear_ctx *lin_ctx, cfg_t *cfg,
269                          const fs_live_variables &live,
270                          struct acp *out_acp);
271 
272    void setup_initial_values();
273    void run();
274 
275    void dump_block_data() const UNUSED;
276 
277    cfg_t *cfg;
278    const fs_live_variables &live;
279 
280    acp_entry **acp;
281    int num_acp;
282    int bitset_words;
283 
284   struct block_data *bd;
285 };
286 } /* anonymous namespace */
287 
fs_copy_prop_dataflow(linear_ctx * lin_ctx,cfg_t * cfg,const fs_live_variables & live,struct acp * out_acp)288 fs_copy_prop_dataflow::fs_copy_prop_dataflow(linear_ctx *lin_ctx, cfg_t *cfg,
289                                              const fs_live_variables &live,
290                                              struct acp *out_acp)
291    : cfg(cfg), live(live)
292 {
293    bd = linear_zalloc_array(lin_ctx, struct block_data, cfg->num_blocks);
294 
295    num_acp = 0;
296    foreach_block (block, cfg)
297       num_acp += out_acp[block->num].length();
298 
299    bitset_words = BITSET_WORDS(num_acp);
300 
301    foreach_block (block, cfg) {
302       bd[block->num].livein = linear_zalloc_array(lin_ctx, BITSET_WORD, bitset_words);
303       bd[block->num].liveout = linear_zalloc_array(lin_ctx, BITSET_WORD, bitset_words);
304       bd[block->num].copy = linear_zalloc_array(lin_ctx, BITSET_WORD, bitset_words);
305       bd[block->num].kill = linear_zalloc_array(lin_ctx, BITSET_WORD, bitset_words);
306       bd[block->num].undef = linear_zalloc_array(lin_ctx, BITSET_WORD, bitset_words);
307       bd[block->num].reachin = linear_zalloc_array(lin_ctx, BITSET_WORD, bitset_words);
308       bd[block->num].exec_mismatch = linear_zalloc_array(lin_ctx, BITSET_WORD, bitset_words);
309    }
310 
311    acp = linear_zalloc_array(lin_ctx, struct acp_entry *, num_acp);
312 
313    int next_acp = 0;
314    foreach_block (block, cfg) {
315       for (auto iter = out_acp[block->num].begin();
316            iter != out_acp[block->num].end(); ++iter) {
317          acp[next_acp] = *iter;
318 
319          (*iter)->global_idx = next_acp;
320 
321          /* opt_copy_propagation_local populates out_acp with copies created
322           * in a block which are still live at the end of the block.  This
323           * is exactly what we want in the COPY set.
324           */
325          BITSET_SET(bd[block->num].copy, next_acp);
326 
327          next_acp++;
328       }
329    }
330 
331    assert(next_acp == num_acp);
332 
333    setup_initial_values();
334    run();
335 }
336 
337 /**
338  * Like reg_offset, but register must be VGRF or FIXED_GRF.
339  */
340 static inline unsigned
grf_reg_offset(const brw_reg & r)341 grf_reg_offset(const brw_reg &r)
342 {
343    return (r.file == VGRF ? 0 : r.nr) * REG_SIZE +
344           r.offset +
345           (r.file == FIXED_GRF ? r.subnr : 0);
346 }
347 
348 /**
349  * Like regions_overlap, but register must be VGRF or FIXED_GRF.
350  */
351 static inline bool
grf_regions_overlap(const brw_reg & r,unsigned dr,const brw_reg & s,unsigned ds)352 grf_regions_overlap(const brw_reg &r, unsigned dr, const brw_reg &s, unsigned ds)
353 {
354    return reg_space(r) == reg_space(s) &&
355           !(grf_reg_offset(r) + dr <= grf_reg_offset(s) ||
356             grf_reg_offset(s) + ds <= grf_reg_offset(r));
357 }
358 
359 /**
360  * Set up initial values for each of the data flow sets, prior to running
361  * the fixed-point algorithm.
362  */
363 void
setup_initial_values()364 fs_copy_prop_dataflow::setup_initial_values()
365 {
366    /* Initialize the COPY and KILL sets. */
367    {
368       struct acp acp_table;
369 
370       /* First, get all the KILLs for instructions which overwrite ACP
371        * destinations.
372        */
373       for (int i = 0; i < num_acp; i++)
374          acp_table.add(acp[i]);
375 
376       foreach_block (block, cfg) {
377          foreach_inst_in_block(fs_inst, inst, block) {
378             if (inst->dst.file != VGRF &&
379                 inst->dst.file != FIXED_GRF)
380                continue;
381 
382             for (auto iter = acp_table.find_by_src(inst->dst.nr);
383               iter != acp_table.end() && (*iter)->src.nr == inst->dst.nr;
384               ++iter) {
385                if (grf_regions_overlap(inst->dst, inst->size_written,
386                                        (*iter)->src, (*iter)->size_read)) {
387                   BITSET_SET(bd[block->num].kill, (*iter)->global_idx);
388                   if (inst->force_writemask_all && !(*iter)->force_writemask_all)
389                      BITSET_SET(bd[block->num].exec_mismatch, (*iter)->global_idx);
390                }
391             }
392 
393             if (inst->dst.file != VGRF)
394                continue;
395 
396             for (auto iter = acp_table.find_by_dst(inst->dst.nr);
397               iter != acp_table.end() && (*iter)->dst.nr == inst->dst.nr;
398               ++iter) {
399                if (grf_regions_overlap(inst->dst, inst->size_written,
400                                        (*iter)->dst, (*iter)->size_written)) {
401                   BITSET_SET(bd[block->num].kill, (*iter)->global_idx);
402                   if (inst->force_writemask_all && !(*iter)->force_writemask_all)
403                      BITSET_SET(bd[block->num].exec_mismatch, (*iter)->global_idx);
404                }
405             }
406          }
407       }
408    }
409 
410    /* Populate the initial values for the livein and liveout sets.  For the
411     * block at the start of the program, livein = 0 and liveout = copy.
412     * For the others, set liveout and livein to ~0 (the universal set).
413     */
414    foreach_block (block, cfg) {
415       if (block->parents.is_empty()) {
416          for (int i = 0; i < bitset_words; i++) {
417             bd[block->num].livein[i] = 0u;
418             bd[block->num].liveout[i] = bd[block->num].copy[i];
419          }
420       } else {
421          for (int i = 0; i < bitset_words; i++) {
422             bd[block->num].liveout[i] = ~0u;
423             bd[block->num].livein[i] = ~0u;
424          }
425       }
426    }
427 
428    /* Initialize the undef set. */
429    foreach_block (block, cfg) {
430       for (int i = 0; i < num_acp; i++) {
431          BITSET_SET(bd[block->num].undef, i);
432          for (unsigned off = 0; off < acp[i]->size_written; off += REG_SIZE) {
433             if (BITSET_TEST(live.block_data[block->num].defout,
434                             live.var_from_reg(byte_offset(acp[i]->dst, off))))
435                BITSET_CLEAR(bd[block->num].undef, i);
436          }
437       }
438    }
439 }
440 
441 /**
442  * Walk the set of instructions in the block, marking which entries in the acp
443  * are killed by the block.
444  */
445 void
run()446 fs_copy_prop_dataflow::run()
447 {
448    bool progress;
449 
450    do {
451       progress = false;
452 
453       foreach_block (block, cfg) {
454          if (block->parents.is_empty())
455             continue;
456 
457          for (int i = 0; i < bitset_words; i++) {
458             const BITSET_WORD old_liveout = bd[block->num].liveout[i];
459             const BITSET_WORD old_reachin = bd[block->num].reachin[i];
460             BITSET_WORD livein_from_any_block = 0;
461 
462             /* Update livein for this block.  If a copy is live out of all
463              * parent blocks, it's live coming in to this block.
464              */
465             bd[block->num].livein[i] = ~0u;
466             foreach_list_typed(bblock_link, parent_link, link, &block->parents) {
467                bblock_t *parent = parent_link->block;
468                /* Consider ACP entries with a known-undefined destination to
469                 * be available from the parent.  This is valid because we're
470                 * free to set the undefined variable equal to the source of
471                 * the ACP entry without breaking the application's
472                 * expectations, since the variable is undefined.
473                 */
474                bd[block->num].livein[i] &= (bd[parent->num].liveout[i] |
475                                             bd[parent->num].undef[i]);
476                livein_from_any_block |= bd[parent->num].liveout[i];
477 
478                /* Update reachin for this block.  If the end of any
479                 * parent block is reachable from the copy, the start
480                 * of this block is reachable from it as well.
481                 */
482                bd[block->num].reachin[i] |= (bd[parent->num].reachin[i] |
483                                              bd[parent->num].copy[i]);
484             }
485 
486             /* Limit to the set of ACP entries that can possibly be available
487              * at the start of the block, since propagating from a variable
488              * which is guaranteed to be undefined (rather than potentially
489              * undefined for some dynamic control-flow paths) doesn't seem
490              * particularly useful.
491              */
492             bd[block->num].livein[i] &= livein_from_any_block;
493 
494             /* Update liveout for this block. */
495             bd[block->num].liveout[i] =
496                bd[block->num].copy[i] | (bd[block->num].livein[i] &
497                                          ~bd[block->num].kill[i]);
498 
499             if (old_liveout != bd[block->num].liveout[i] ||
500                 old_reachin != bd[block->num].reachin[i])
501                progress = true;
502          }
503       }
504    } while (progress);
505 
506    /* Perform a second fixed-point pass in order to propagate the
507     * exec_mismatch bitsets.  Note that this requires an accurate
508     * value of the reachin bitsets as input, which isn't available
509     * until the end of the first propagation pass, so this loop cannot
510     * be folded into the previous one.
511     */
512    do {
513       progress = false;
514 
515       foreach_block (block, cfg) {
516          for (int i = 0; i < bitset_words; i++) {
517             const BITSET_WORD old_exec_mismatch = bd[block->num].exec_mismatch[i];
518 
519             /* Update exec_mismatch for this block.  If the end of a
520              * parent block is reachable by an overwrite with
521              * inconsistent execution masking, the start of this block
522              * is reachable by such an overwrite as well.
523              */
524             foreach_list_typed(bblock_link, parent_link, link, &block->parents) {
525                bblock_t *parent = parent_link->block;
526                bd[block->num].exec_mismatch[i] |= (bd[parent->num].exec_mismatch[i] &
527                                                    bd[parent->num].reachin[i]);
528             }
529 
530             /* Only consider overwrites with inconsistent execution
531              * masking if they are reachable from the copy, since
532              * overwrites unreachable from a copy are harmless to that
533              * copy.
534              */
535             bd[block->num].exec_mismatch[i] &= bd[block->num].reachin[i];
536             if (old_exec_mismatch != bd[block->num].exec_mismatch[i])
537                progress = true;
538          }
539       }
540    } while (progress);
541 }
542 
543 void
dump_block_data() const544 fs_copy_prop_dataflow::dump_block_data() const
545 {
546    foreach_block (block, cfg) {
547       fprintf(stderr, "Block %d [%d, %d] (parents ", block->num,
548              block->start_ip, block->end_ip);
549       foreach_list_typed(bblock_link, link, link, &block->parents) {
550          bblock_t *parent = link->block;
551          fprintf(stderr, "%d ", parent->num);
552       }
553       fprintf(stderr, "):\n");
554       fprintf(stderr, "       livein = 0x");
555       for (int i = 0; i < bitset_words; i++)
556          fprintf(stderr, "%08x", bd[block->num].livein[i]);
557       fprintf(stderr, ", liveout = 0x");
558       for (int i = 0; i < bitset_words; i++)
559          fprintf(stderr, "%08x", bd[block->num].liveout[i]);
560       fprintf(stderr, ",\n       copy   = 0x");
561       for (int i = 0; i < bitset_words; i++)
562          fprintf(stderr, "%08x", bd[block->num].copy[i]);
563       fprintf(stderr, ", kill    = 0x");
564       for (int i = 0; i < bitset_words; i++)
565          fprintf(stderr, "%08x", bd[block->num].kill[i]);
566       fprintf(stderr, "\n");
567    }
568 }
569 
570 static bool
is_logic_op(enum opcode opcode)571 is_logic_op(enum opcode opcode)
572 {
573    return (opcode == BRW_OPCODE_AND ||
574            opcode == BRW_OPCODE_OR  ||
575            opcode == BRW_OPCODE_XOR ||
576            opcode == BRW_OPCODE_NOT);
577 }
578 
579 static bool
can_take_stride(fs_inst * inst,brw_reg_type dst_type,unsigned arg,unsigned stride,const struct brw_compiler * compiler)580 can_take_stride(fs_inst *inst, brw_reg_type dst_type,
581                 unsigned arg, unsigned stride,
582                 const struct brw_compiler *compiler)
583 {
584    const struct intel_device_info *devinfo = compiler->devinfo;
585 
586    if (stride > 4)
587       return false;
588 
589    /* Bail if the channels of the source need to be aligned to the byte offset
590     * of the corresponding channel of the destination, and the provided stride
591     * would break this restriction.
592     */
593    if (has_dst_aligned_region_restriction(devinfo, inst, dst_type) &&
594        !(brw_type_size_bytes(inst->src[arg].type) * stride ==
595            brw_type_size_bytes(dst_type) * inst->dst.stride ||
596          stride == 0))
597       return false;
598 
599    /* 3-source instructions can only be Align16, which restricts what strides
600     * they can take. They can only take a stride of 1 (the usual case), or 0
601     * with a special "repctrl" bit. But the repctrl bit doesn't work for
602     * 64-bit datatypes, so if the source type is 64-bit then only a stride of
603     * 1 is allowed. From the Broadwell PRM, Volume 7 "3D Media GPGPU", page
604     * 944:
605     *
606     *    This is applicable to 32b datatypes and 16b datatype. 64b datatypes
607     *    cannot use the replicate control.
608     */
609    if (inst->is_3src(compiler)) {
610       if (brw_type_size_bytes(inst->src[arg].type) > 4)
611          return stride == 1;
612       else
613          return stride == 1 || stride == 0;
614    }
615 
616    if (inst->is_math()) {
617       /* Wa_22016140776:
618        *
619        *    Scalar broadcast on HF math (packed or unpacked) must not be used.
620        *    Compiler must use a mov instruction to expand the scalar value to
621        *    a vector before using in a HF (packed or unpacked) math operation.
622        *
623        * Prevent copy propagating a scalar value into a math instruction.
624        */
625       if (intel_needs_workaround(devinfo, 22016140776) &&
626           stride == 0 && inst->src[arg].type == BRW_TYPE_HF) {
627          return false;
628       }
629 
630       /* From the Broadwell PRM, Volume 2a "Command Reference - Instructions",
631        * page 391 ("Extended Math Function"):
632        *
633        *     The following restrictions apply for align1 mode: Scalar source
634        *     is supported. Source and destination horizontal stride must be
635        *     the same.
636        */
637       return stride == inst->dst.stride || stride == 0;
638    }
639 
640    return true;
641 }
642 
643 static bool
instruction_requires_packed_data(fs_inst * inst)644 instruction_requires_packed_data(fs_inst *inst)
645 {
646    switch (inst->opcode) {
647    case FS_OPCODE_DDX_FINE:
648    case FS_OPCODE_DDX_COARSE:
649    case FS_OPCODE_DDY_FINE:
650    case FS_OPCODE_DDY_COARSE:
651    case SHADER_OPCODE_QUAD_SWIZZLE:
652       return true;
653    default:
654       return false;
655    }
656 }
657 
658 static bool
try_copy_propagate(const brw_compiler * compiler,fs_inst * inst,acp_entry * entry,int arg,const brw::simple_allocator & alloc,uint8_t max_polygons)659 try_copy_propagate(const brw_compiler *compiler, fs_inst *inst,
660                    acp_entry *entry, int arg,
661                    const brw::simple_allocator &alloc,
662                    uint8_t max_polygons)
663 {
664    if (inst->src[arg].file != VGRF)
665       return false;
666 
667    const struct intel_device_info *devinfo = compiler->devinfo;
668 
669    assert(entry->src.file == VGRF || entry->src.file == UNIFORM ||
670           entry->src.file == ATTR || entry->src.file == FIXED_GRF);
671 
672    /* Avoid propagating a LOAD_PAYLOAD instruction into another if there is a
673     * good chance that we'll be able to eliminate the latter through register
674     * coalescing.  If only part of the sources of the second LOAD_PAYLOAD can
675     * be simplified through copy propagation we would be making register
676     * coalescing impossible, ending up with unnecessary copies in the program.
677     * This is also the case for is_multi_copy_payload() copies that can only
678     * be coalesced when the instruction is lowered into a sequence of MOVs.
679     *
680     * Worse -- In cases where the ACP entry was the result of CSE combining
681     * multiple LOAD_PAYLOAD subexpressions, propagating the first LOAD_PAYLOAD
682     * into the second would undo the work of CSE, leading to an infinite
683     * optimization loop.  Avoid this by detecting LOAD_PAYLOAD copies from CSE
684     * temporaries which should match is_coalescing_payload().
685     */
686    if (entry->opcode == SHADER_OPCODE_LOAD_PAYLOAD &&
687        (is_coalescing_payload(alloc, inst) || is_multi_copy_payload(inst)))
688       return false;
689 
690    assert(entry->dst.file == VGRF);
691    if (inst->src[arg].nr != entry->dst.nr)
692       return false;
693 
694    /* Bail if inst is reading a range that isn't contained in the range
695     * that entry is writing.
696     */
697    if (!region_contained_in(inst->src[arg], inst->size_read(arg),
698                             entry->dst, entry->size_written))
699       return false;
700 
701    /* Send messages with EOT set are restricted to use g112-g127 (and we
702     * sometimes need g127 for other purposes), so avoid copy propagating
703     * anything that would make it impossible to satisfy that restriction.
704     */
705    if (inst->eot) {
706       /* Don't propagate things that are already pinned. */
707       if (entry->src.file != VGRF)
708          return false;
709 
710       /* We might be propagating from a large register, while the SEND only
711        * is reading a portion of it (say the .A channel in an RGBA value).
712        * We need to pin both split SEND sources in g112-g126/127, so only
713        * allow this if the registers aren't too large.
714        */
715       if (inst->opcode == SHADER_OPCODE_SEND && inst->sources >= 4 &&
716           entry->src.file == VGRF) {
717          int other_src = arg == 2 ? 3 : 2;
718          unsigned other_size = inst->src[other_src].file == VGRF ?
719                                alloc.sizes[inst->src[other_src].nr] :
720                                inst->size_read(other_src);
721          unsigned prop_src_size = alloc.sizes[entry->src.nr];
722          if (other_size + prop_src_size > 15)
723             return false;
724       }
725    }
726 
727    /* we can't generally copy-propagate UD negations because we
728     * can end up accessing the resulting values as signed integers
729     * instead. See also resolve_ud_negate() and comment in
730     * fs_generator::generate_code.
731     */
732    if (entry->src.type == BRW_TYPE_UD &&
733        entry->src.negate)
734       return false;
735 
736    bool has_source_modifiers = entry->src.abs || entry->src.negate;
737 
738    if (has_source_modifiers && !inst->can_do_source_mods(devinfo))
739       return false;
740 
741    /* Reject cases that would violate register regioning restrictions. */
742    if ((entry->src.file == UNIFORM || !entry->src.is_contiguous()) &&
743        (inst->is_send_from_grf() ||
744         inst->uses_indirect_addressing())) {
745       return false;
746    }
747 
748    /* Some instructions implemented in the generator backend, such as
749     * derivatives, assume that their operands are packed so we can't
750     * generally propagate strided regions to them.
751     */
752    const unsigned entry_stride = (entry->src.file == FIXED_GRF ? 1 :
753                                   entry->src.stride);
754    if (instruction_requires_packed_data(inst) && entry_stride != 1)
755       return false;
756 
757    const brw_reg_type dst_type = (has_source_modifiers &&
758                                   entry->dst.type != inst->src[arg].type) ?
759       entry->dst.type : inst->dst.type;
760 
761    /* Bail if the result of composing both strides would exceed the
762     * hardware limit.
763     */
764    if (!can_take_stride(inst, dst_type, arg,
765                         entry_stride * inst->src[arg].stride,
766                         compiler))
767       return false;
768 
769    /* From the Cherry Trail/Braswell PRMs, Volume 7: 3D Media GPGPU:
770     *    EU Overview
771     *       Register Region Restrictions
772     *          Special Requirements for Handling Double Precision Data Types :
773     *
774     *   "When source or destination datatype is 64b or operation is integer
775     *    DWord multiply, regioning in Align1 must follow these rules:
776     *
777     *      1. Source and Destination horizontal stride must be aligned to the
778     *         same qword.
779     *      2. Regioning must ensure Src.Vstride = Src.Width * Src.Hstride.
780     *      3. Source and Destination offset must be the same, except the case
781     *         of scalar source."
782     *
783     * Most of this is already checked in can_take_stride(), we're only left
784     * with checking 3.
785     */
786    if (has_dst_aligned_region_restriction(devinfo, inst, dst_type) &&
787        entry_stride != 0 &&
788        (reg_offset(inst->dst) % (REG_SIZE * reg_unit(devinfo))) != (reg_offset(entry->src) % (REG_SIZE * reg_unit(devinfo))))
789       return false;
790 
791    /*
792     * Bail if the composition of both regions would be affected by the Xe2+
793     * regioning restrictions that apply to integer types smaller than a dword.
794     * See BSpec #56640 for details.
795     */
796    const brw_reg tmp = horiz_stride(entry->src, inst->src[arg].stride);
797    if (has_subdword_integer_region_restriction(devinfo, inst, &tmp, 1))
798       return false;
799 
800    /* The <8;8,0> regions used for FS attributes in multipolygon
801     * dispatch mode could violate regioning restrictions, don't copy
802     * propagate them in such cases.
803     */
804    if (entry->src.file == ATTR && max_polygons > 1 &&
805        (has_dst_aligned_region_restriction(devinfo, inst, dst_type) ||
806 	instruction_requires_packed_data(inst) ||
807 	(inst->is_3src(compiler) && arg == 2) ||
808 	entry->dst.type != inst->src[arg].type))
809       return false;
810 
811    /* Bail if the source FIXED_GRF region of the copy cannot be trivially
812     * composed with the source region of the instruction -- E.g. because the
813     * copy uses some extended stride greater than 4 not supported natively by
814     * the hardware as a horizontal stride, or because instruction compression
815     * could require us to use a vertical stride shorter than a GRF.
816     */
817    if (entry->src.file == FIXED_GRF &&
818        (inst->src[arg].stride > 4 ||
819         inst->dst.component_size(inst->exec_size) >
820         inst->src[arg].component_size(inst->exec_size)))
821       return false;
822 
823    /* Bail if the instruction type is larger than the execution type of the
824     * copy, what implies that each channel is reading multiple channels of the
825     * destination of the copy, and simply replacing the sources would give a
826     * program with different semantics.
827     */
828    if ((brw_type_size_bits(entry->dst.type) < brw_type_size_bits(inst->src[arg].type) ||
829         entry->is_partial_write) &&
830        inst->opcode != BRW_OPCODE_MOV) {
831       return false;
832    }
833 
834    /* Bail if the result of composing both strides cannot be expressed
835     * as another stride. This avoids, for example, trying to transform
836     * this:
837     *
838     *     MOV (8) rX<1>UD rY<0;1,0>UD
839     *     FOO (8) ...     rX<8;8,1>UW
840     *
841     * into this:
842     *
843     *     FOO (8) ...     rY<0;1,0>UW
844     *
845     * Which would have different semantics.
846     */
847    if (entry_stride != 1 &&
848        (inst->src[arg].stride *
849         brw_type_size_bytes(inst->src[arg].type)) % brw_type_size_bytes(entry->src.type) != 0)
850       return false;
851 
852    /* Since semantics of source modifiers are type-dependent we need to
853     * ensure that the meaning of the instruction remains the same if we
854     * change the type. If the sizes of the types are different the new
855     * instruction will read a different amount of data than the original
856     * and the semantics will always be different.
857     */
858    if (has_source_modifiers &&
859        entry->dst.type != inst->src[arg].type &&
860        (!inst->can_change_types() ||
861         brw_type_size_bits(entry->dst.type) != brw_type_size_bits(inst->src[arg].type)))
862       return false;
863 
864    if ((entry->src.negate || entry->src.abs) &&
865        is_logic_op(inst->opcode)) {
866       return false;
867    }
868 
869    /* Save the offset of inst->src[arg] relative to entry->dst for it to be
870     * applied later.
871     */
872    const unsigned rel_offset = inst->src[arg].offset - entry->dst.offset;
873 
874    /* Fold the copy into the instruction consuming it. */
875    inst->src[arg].file = entry->src.file;
876    inst->src[arg].nr = entry->src.nr;
877    inst->src[arg].subnr = entry->src.subnr;
878    inst->src[arg].offset = entry->src.offset;
879 
880    /* Compose the strides of both regions. */
881    if (entry->src.file == FIXED_GRF) {
882       if (inst->src[arg].stride) {
883          const unsigned orig_width = 1 << entry->src.width;
884          const unsigned reg_width =
885             REG_SIZE / (brw_type_size_bytes(inst->src[arg].type) *
886                         inst->src[arg].stride);
887          inst->src[arg].width = cvt(MIN2(orig_width, reg_width)) - 1;
888          inst->src[arg].hstride = cvt(inst->src[arg].stride);
889          inst->src[arg].vstride = inst->src[arg].hstride + inst->src[arg].width;
890       } else {
891          inst->src[arg].vstride = inst->src[arg].hstride =
892             inst->src[arg].width = 0;
893       }
894 
895       inst->src[arg].stride = 1;
896 
897       /* Hopefully no Align16 around here... */
898       assert(entry->src.swizzle == BRW_SWIZZLE_XYZW);
899       inst->src[arg].swizzle = entry->src.swizzle;
900    } else {
901       inst->src[arg].stride *= entry->src.stride;
902    }
903 
904    /* Compute the first component of the copy that the instruction is
905     * reading, and the base byte offset within that component.
906     */
907    assert(entry->dst.stride == 1);
908    const unsigned component = rel_offset / brw_type_size_bytes(entry->dst.type);
909    const unsigned suboffset = rel_offset % brw_type_size_bytes(entry->dst.type);
910 
911    /* Calculate the byte offset at the origin of the copy of the given
912     * component and suboffset.
913     */
914    inst->src[arg] = byte_offset(inst->src[arg],
915       component * entry_stride * brw_type_size_bytes(entry->src.type) + suboffset);
916 
917    if (has_source_modifiers) {
918       if (entry->dst.type != inst->src[arg].type) {
919          /* We are propagating source modifiers from a MOV with a different
920           * type.  If we got here, then we can just change the source and
921           * destination types of the instruction and keep going.
922           */
923          for (int i = 0; i < inst->sources; i++) {
924             inst->src[i].type = entry->dst.type;
925          }
926          inst->dst.type = entry->dst.type;
927       }
928 
929       if (!inst->src[arg].abs) {
930          inst->src[arg].abs = entry->src.abs;
931          inst->src[arg].negate ^= entry->src.negate;
932       }
933    }
934 
935    return true;
936 }
937 
938 static bool
try_constant_propagate_value(brw_reg val,brw_reg_type dst_type,fs_inst * inst,int arg)939 try_constant_propagate_value(brw_reg val, brw_reg_type dst_type,
940                              fs_inst *inst, int arg)
941 {
942    bool progress = false;
943 
944    if (brw_type_size_bytes(val.type) > 4)
945       return false;
946 
947    /* If the size of the use type is smaller than the size of the entry,
948     * clamp the value to the range of the use type.  This enables constant
949     * copy propagation in cases like
950     *
951     *
952     *    mov(8)          g12<1>UD        0x0000000cUD
953     *    ...
954     *    mul(8)          g47<1>D         g86<8,8,1>D     g12<16,8,2>W
955     */
956    if (brw_type_size_bits(inst->src[arg].type) <
957        brw_type_size_bits(dst_type)) {
958       if (brw_type_size_bytes(inst->src[arg].type) != 2 ||
959           brw_type_size_bytes(dst_type) != 4)
960          return false;
961 
962       assert(inst->src[arg].subnr == 0 || inst->src[arg].subnr == 2);
963 
964       /* When subnr is 0, we want the lower 16-bits, and when it's 2, we
965        * want the upper 16-bits. No other values of subnr are valid for a
966        * UD source.
967        */
968       const uint16_t v = inst->src[arg].subnr == 2 ? val.ud >> 16 : val.ud;
969 
970       val.ud = v | (uint32_t(v) << 16);
971    }
972 
973    val.type = inst->src[arg].type;
974 
975    if (inst->src[arg].abs) {
976       if (is_logic_op(inst->opcode) ||
977           !brw_reg_abs_immediate(&val)) {
978          return false;
979       }
980    }
981 
982    if (inst->src[arg].negate) {
983       if (is_logic_op(inst->opcode) ||
984           !brw_reg_negate_immediate(&val)) {
985          return false;
986       }
987    }
988 
989    switch (inst->opcode) {
990    case BRW_OPCODE_MOV:
991    case SHADER_OPCODE_LOAD_PAYLOAD:
992    case SHADER_OPCODE_POW:
993    case FS_OPCODE_PACK:
994       inst->src[arg] = val;
995       progress = true;
996       break;
997 
998    case BRW_OPCODE_SUBB:
999       if (arg == 1) {
1000          inst->src[arg] = val;
1001          progress = true;
1002       }
1003       break;
1004 
1005    case BRW_OPCODE_MACH:
1006    case BRW_OPCODE_MUL:
1007    case SHADER_OPCODE_MULH:
1008    case BRW_OPCODE_ADD:
1009    case BRW_OPCODE_XOR:
1010    case BRW_OPCODE_ADDC:
1011       if (arg == 1) {
1012          inst->src[arg] = val;
1013          progress = true;
1014       } else if (arg == 0 && inst->src[1].file != IMM) {
1015          /* We used to not copy propagate the constant in situations like
1016           *
1017           *    mov(8)          g8<1>D          0x7fffffffD
1018           *    mul(8)          g16<1>D         g8<8,8,1>D      g15<16,8,2>W
1019           *
1020           * On platforms that only have a 32x16 multiplier, this would
1021           * result in lowering the multiply to
1022           *
1023           *    mul(8)          g15<1>D         g14<8,8,1>D     0xffffUW
1024           *    mul(8)          g16<1>D         g14<8,8,1>D     0x7fffUW
1025           *    add(8)          g15.1<2>UW      g15.1<16,8,2>UW g16<16,8,2>UW
1026           *
1027           * On Gfx8 and Gfx9, which have the full 32x32 multiplier, it
1028           * would results in
1029           *
1030           *    mul(8)          g16<1>D         g15<16,8,2>W    0x7fffffffD
1031           *
1032           * Volume 2a of the Skylake PRM says:
1033           *
1034           *    When multiplying a DW and any lower precision integer, the
1035           *    DW operand must on src0.
1036           *
1037           * So it would have been invalid. However, brw_fs_combine_constants
1038           * will now "fix" the constant.
1039           */
1040          if (inst->opcode == BRW_OPCODE_MUL &&
1041              brw_type_size_bytes(inst->src[1].type) < 4 &&
1042              (inst->src[0].type == BRW_TYPE_D ||
1043               inst->src[0].type == BRW_TYPE_UD)) {
1044             inst->src[0] = val;
1045             inst->src[0].type = BRW_TYPE_D;
1046             progress = true;
1047             break;
1048          }
1049 
1050          /* Fit this constant in by commuting the operands.
1051           * Exception: we can't do this for 32-bit integer MUL/MACH
1052           * because it's asymmetric.
1053           *
1054           * The BSpec says for Broadwell that
1055           *
1056           *    "When multiplying DW x DW, the dst cannot be accumulator."
1057           *
1058           * Integer MUL with a non-accumulator destination will be lowered
1059           * by lower_integer_multiplication(), so don't restrict it.
1060           */
1061          if (((inst->opcode == BRW_OPCODE_MUL &&
1062                inst->dst.is_accumulator()) ||
1063               inst->opcode == BRW_OPCODE_MACH) &&
1064              (inst->src[1].type == BRW_TYPE_D ||
1065               inst->src[1].type == BRW_TYPE_UD))
1066             break;
1067          inst->src[0] = inst->src[1];
1068          inst->src[1] = val;
1069          progress = true;
1070       }
1071       break;
1072 
1073    case BRW_OPCODE_ADD3:
1074       /* add3 can have a single imm16 source. Proceed if the source type is
1075        * already W or UW or the value can be coerced to one of those types.
1076        */
1077       if (val.type == BRW_TYPE_W || val.type == BRW_TYPE_UW)
1078          ; /* Nothing to do. */
1079       else if (val.ud <= 0xffff)
1080          val = brw_imm_uw(val.ud);
1081       else if (val.d >= -0x8000 && val.d <= 0x7fff)
1082          val = brw_imm_w(val.d);
1083       else
1084          break;
1085 
1086       if (arg == 2) {
1087          inst->src[arg] = val;
1088          progress = true;
1089       } else if (inst->src[2].file != IMM) {
1090          inst->src[arg] = inst->src[2];
1091          inst->src[2] = val;
1092          progress = true;
1093       }
1094 
1095       break;
1096 
1097    case BRW_OPCODE_CMP:
1098       if (arg == 1) {
1099          inst->src[arg] = val;
1100          progress = true;
1101       } else if (arg == 0 && inst->src[1].file != IMM) {
1102          enum brw_conditional_mod new_cmod;
1103 
1104          new_cmod = brw_swap_cmod(inst->conditional_mod);
1105          if (new_cmod != BRW_CONDITIONAL_NONE) {
1106             /* Fit this constant in by swapping the operands and
1107              * flipping the test
1108              */
1109             inst->src[0] = inst->src[1];
1110             inst->src[1] = val;
1111             inst->conditional_mod = new_cmod;
1112             progress = true;
1113          }
1114       }
1115       break;
1116 
1117    case BRW_OPCODE_SEL:
1118       if (arg == 1) {
1119          inst->src[arg] = val;
1120          progress = true;
1121       } else if (arg == 0) {
1122          if (inst->src[1].file != IMM &&
1123              (inst->conditional_mod == BRW_CONDITIONAL_NONE ||
1124               /* Only GE and L are commutative. */
1125               inst->conditional_mod == BRW_CONDITIONAL_GE ||
1126               inst->conditional_mod == BRW_CONDITIONAL_L)) {
1127             inst->src[0] = inst->src[1];
1128             inst->src[1] = val;
1129 
1130             /* If this was predicated, flipping operands means
1131              * we also need to flip the predicate.
1132              */
1133             if (inst->conditional_mod == BRW_CONDITIONAL_NONE) {
1134                inst->predicate_inverse =
1135                   !inst->predicate_inverse;
1136             }
1137          } else {
1138             inst->src[0] = val;
1139          }
1140 
1141          progress = true;
1142       }
1143       break;
1144 
1145    case BRW_OPCODE_CSEL:
1146       assert(inst->conditional_mod != BRW_CONDITIONAL_NONE);
1147 
1148       if (arg == 0 &&
1149           inst->src[1].file != IMM &&
1150           (!brw_type_is_float(inst->src[1].type) ||
1151            inst->conditional_mod == BRW_CONDITIONAL_NZ ||
1152            inst->conditional_mod == BRW_CONDITIONAL_Z)) {
1153          /* Only EQ and NE are commutative due to NaN issues. */
1154          inst->src[0] = inst->src[1];
1155          inst->src[1] = val;
1156          inst->conditional_mod = brw_negate_cmod(inst->conditional_mod);
1157       } else {
1158          /* While CSEL is a 3-source instruction, the last source should never
1159           * be a constant.  We'll support that, but should it ever happen, we
1160           * should add support to the constant folding pass.
1161           */
1162          inst->src[arg] = val;
1163       }
1164 
1165       progress = true;
1166       break;
1167 
1168    case FS_OPCODE_FB_WRITE_LOGICAL:
1169       /* The stencil and omask sources of FS_OPCODE_FB_WRITE_LOGICAL are
1170        * bit-cast using a strided region so they cannot be immediates.
1171        */
1172       if (arg != FB_WRITE_LOGICAL_SRC_SRC_STENCIL &&
1173           arg != FB_WRITE_LOGICAL_SRC_OMASK) {
1174          inst->src[arg] = val;
1175          progress = true;
1176       }
1177       break;
1178 
1179    case SHADER_OPCODE_INT_QUOTIENT:
1180    case SHADER_OPCODE_INT_REMAINDER:
1181    case BRW_OPCODE_AND:
1182    case BRW_OPCODE_ASR:
1183    case BRW_OPCODE_BFE:
1184    case BRW_OPCODE_BFI1:
1185    case BRW_OPCODE_BFI2:
1186    case BRW_OPCODE_ROL:
1187    case BRW_OPCODE_ROR:
1188    case BRW_OPCODE_SHL:
1189    case BRW_OPCODE_SHR:
1190    case BRW_OPCODE_OR:
1191    case SHADER_OPCODE_TEX_LOGICAL:
1192    case SHADER_OPCODE_TXD_LOGICAL:
1193    case SHADER_OPCODE_TXF_LOGICAL:
1194    case SHADER_OPCODE_TXL_LOGICAL:
1195    case SHADER_OPCODE_TXS_LOGICAL:
1196    case FS_OPCODE_TXB_LOGICAL:
1197    case SHADER_OPCODE_TXF_CMS_W_LOGICAL:
1198    case SHADER_OPCODE_TXF_CMS_W_GFX12_LOGICAL:
1199    case SHADER_OPCODE_TXF_MCS_LOGICAL:
1200    case SHADER_OPCODE_LOD_LOGICAL:
1201    case SHADER_OPCODE_TG4_BIAS_LOGICAL:
1202    case SHADER_OPCODE_TG4_EXPLICIT_LOD_LOGICAL:
1203    case SHADER_OPCODE_TG4_IMPLICIT_LOD_LOGICAL:
1204    case SHADER_OPCODE_TG4_LOGICAL:
1205    case SHADER_OPCODE_TG4_OFFSET_LOGICAL:
1206    case SHADER_OPCODE_TG4_OFFSET_LOD_LOGICAL:
1207    case SHADER_OPCODE_TG4_OFFSET_BIAS_LOGICAL:
1208    case SHADER_OPCODE_SAMPLEINFO_LOGICAL:
1209    case SHADER_OPCODE_IMAGE_SIZE_LOGICAL:
1210    case SHADER_OPCODE_MEMORY_LOAD_LOGICAL:
1211    case SHADER_OPCODE_MEMORY_STORE_LOGICAL:
1212    case SHADER_OPCODE_MEMORY_ATOMIC_LOGICAL:
1213    case FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD:
1214    case FS_OPCODE_VARYING_PULL_CONSTANT_LOAD_LOGICAL:
1215    case SHADER_OPCODE_BROADCAST:
1216    case BRW_OPCODE_MAD:
1217    case BRW_OPCODE_LRP:
1218    case FS_OPCODE_PACK_HALF_2x16_SPLIT:
1219    case SHADER_OPCODE_SHUFFLE:
1220       inst->src[arg] = val;
1221       progress = true;
1222       break;
1223 
1224    default:
1225       break;
1226    }
1227 
1228    return progress;
1229 }
1230 
1231 
1232 static bool
try_constant_propagate(fs_inst * inst,acp_entry * entry,int arg)1233 try_constant_propagate(fs_inst *inst, acp_entry *entry, int arg)
1234 {
1235    if (inst->src[arg].file != VGRF)
1236       return false;
1237 
1238    assert(entry->dst.file == VGRF);
1239    if (inst->src[arg].nr != entry->dst.nr)
1240       return false;
1241 
1242    /* Bail if inst is reading a range that isn't contained in the range
1243     * that entry is writing.
1244     */
1245    if (!region_contained_in(inst->src[arg], inst->size_read(arg),
1246                             entry->dst, entry->size_written))
1247       return false;
1248 
1249    /* If the size of the use type is larger than the size of the entry
1250     * type, the entry doesn't contain all of the data that the user is
1251     * trying to use.
1252     */
1253    if (brw_type_size_bits(inst->src[arg].type) >
1254        brw_type_size_bits(entry->dst.type))
1255       return false;
1256 
1257    return try_constant_propagate_value(entry->src, entry->dst.type, inst, arg);
1258 }
1259 
1260 static bool
can_propagate_from(fs_inst * inst)1261 can_propagate_from(fs_inst *inst)
1262 {
1263    return (inst->opcode == BRW_OPCODE_MOV &&
1264            inst->dst.file == VGRF &&
1265            ((inst->src[0].file == VGRF &&
1266              !grf_regions_overlap(inst->dst, inst->size_written,
1267                                   inst->src[0], inst->size_read(0))) ||
1268             inst->src[0].file == ATTR ||
1269             inst->src[0].file == UNIFORM ||
1270             inst->src[0].file == IMM ||
1271             (inst->src[0].file == FIXED_GRF &&
1272              inst->src[0].is_contiguous())) &&
1273            /* is_raw_move also rejects source modifiers, but copy propagation
1274             * can handle that if the types are the same.
1275             */
1276            ((inst->src[0].type == inst->dst.type &&
1277              !inst->saturate) ||
1278             inst->is_raw_move()) &&
1279            /* Subset of !is_partial_write() conditions. */
1280            !inst->predicate && inst->dst.is_contiguous()) ||
1281           is_identity_payload(FIXED_GRF, inst);
1282 }
1283 
1284 static void
commute_immediates(fs_inst * inst)1285 commute_immediates(fs_inst *inst)
1286 {
1287    /* ADD3 can only have the immediate as src0. */
1288    if (inst->opcode == BRW_OPCODE_ADD3) {
1289       if (inst->src[2].file == IMM) {
1290          const auto src0 = inst->src[0];
1291          inst->src[0] = inst->src[2];
1292          inst->src[2] = src0;
1293       }
1294    }
1295 
1296    /* If only one of the sources of a 2-source, commutative instruction (e.g.,
1297     * AND) is immediate, it must be src1. If both are immediate, opt_algebraic
1298     * should fold it away.
1299     */
1300    if (inst->sources == 2 && inst->is_commutative() &&
1301        inst->src[0].file == IMM && inst->src[1].file != IMM) {
1302       const auto src1 = inst->src[1];
1303       inst->src[1] = inst->src[0];
1304       inst->src[0] = src1;
1305    }
1306 }
1307 
1308 /* Walks a basic block and does copy propagation on it using the acp
1309  * list.
1310  */
1311 static bool
opt_copy_propagation_local(const brw_compiler * compiler,linear_ctx * lin_ctx,bblock_t * block,struct acp & acp,const brw::simple_allocator & alloc,uint8_t max_polygons)1312 opt_copy_propagation_local(const brw_compiler *compiler, linear_ctx *lin_ctx,
1313                            bblock_t *block, struct acp &acp,
1314                            const brw::simple_allocator &alloc,
1315                            uint8_t max_polygons)
1316 {
1317    bool progress = false;
1318 
1319    foreach_inst_in_block(fs_inst, inst, block) {
1320       /* Try propagating into this instruction. */
1321       bool instruction_progress = false;
1322       for (int i = inst->sources - 1; i >= 0; i--) {
1323          if (inst->src[i].file != VGRF)
1324             continue;
1325 
1326          for (auto iter = acp.find_by_dst(inst->src[i].nr);
1327               iter != acp.end() && (*iter)->dst.nr == inst->src[i].nr;
1328               ++iter) {
1329             if ((*iter)->src.file == IMM) {
1330                if (try_constant_propagate(inst, *iter, i)) {
1331                   instruction_progress = true;
1332                   break;
1333                }
1334             } else {
1335                if (try_copy_propagate(compiler, inst, *iter, i, alloc,
1336                                       max_polygons)) {
1337                   instruction_progress = true;
1338                   break;
1339                }
1340             }
1341          }
1342       }
1343 
1344       if (instruction_progress) {
1345          progress = true;
1346          commute_immediates(inst);
1347       }
1348 
1349       /* kill the destination from the ACP */
1350       if (inst->dst.file == VGRF || inst->dst.file == FIXED_GRF) {
1351          for (auto iter = acp.find_by_dst(inst->dst.nr);
1352               iter != acp.end() && (*iter)->dst.nr == inst->dst.nr;
1353               ++iter) {
1354             if (grf_regions_overlap((*iter)->dst, (*iter)->size_written,
1355                                     inst->dst, inst->size_written))
1356                acp.remove(*iter);
1357          }
1358 
1359          for (auto iter = acp.find_by_src(inst->dst.nr);
1360               iter != acp.end() && (*iter)->src.nr == inst->dst.nr;
1361               ++iter) {
1362             /* Make sure we kill the entry if this instruction overwrites
1363              * _any_ of the registers that it reads
1364              */
1365             if (grf_regions_overlap((*iter)->src, (*iter)->size_read,
1366                                     inst->dst, inst->size_written))
1367                acp.remove(*iter);
1368          }
1369       }
1370 
1371       /* If this instruction's source could potentially be folded into the
1372        * operand of another instruction, add it to the ACP.
1373        */
1374       if (can_propagate_from(inst)) {
1375          acp_entry *entry = linear_zalloc(lin_ctx, acp_entry);
1376          entry->dst = inst->dst;
1377          entry->src = inst->src[0];
1378          entry->size_written = inst->size_written;
1379          for (unsigned i = 0; i < inst->sources; i++)
1380             entry->size_read += inst->size_read(i);
1381          entry->opcode = inst->opcode;
1382          entry->is_partial_write = inst->is_partial_write();
1383          entry->force_writemask_all = inst->force_writemask_all;
1384          acp.add(entry);
1385       } else if (inst->opcode == SHADER_OPCODE_LOAD_PAYLOAD &&
1386                  inst->dst.file == VGRF) {
1387          int offset = 0;
1388          for (int i = 0; i < inst->sources; i++) {
1389             int effective_width = i < inst->header_size ? 8 : inst->exec_size;
1390             const unsigned size_written =
1391                effective_width * brw_type_size_bytes(inst->src[i].type);
1392             if (inst->src[i].file == VGRF ||
1393                 (inst->src[i].file == FIXED_GRF &&
1394                  inst->src[i].is_contiguous())) {
1395                const brw_reg_type t = i < inst->header_size ?
1396                   BRW_TYPE_UD : inst->src[i].type;
1397                brw_reg dst = byte_offset(retype(inst->dst, t), offset);
1398                if (!dst.equals(inst->src[i])) {
1399                   acp_entry *entry = linear_zalloc(lin_ctx, acp_entry);
1400                   entry->dst = dst;
1401                   entry->src = retype(inst->src[i], t);
1402                   entry->size_written = size_written;
1403                   entry->size_read = inst->size_read(i);
1404                   entry->opcode = inst->opcode;
1405                   entry->force_writemask_all = inst->force_writemask_all;
1406                   acp.add(entry);
1407                }
1408             }
1409             offset += size_written;
1410          }
1411       }
1412    }
1413 
1414    return progress;
1415 }
1416 
1417 bool
brw_fs_opt_copy_propagation(fs_visitor & s)1418 brw_fs_opt_copy_propagation(fs_visitor &s)
1419 {
1420    bool progress = false;
1421    void *copy_prop_ctx = ralloc_context(NULL);
1422    linear_ctx *lin_ctx = linear_context(copy_prop_ctx);
1423    struct acp out_acp[s.cfg->num_blocks];
1424 
1425    const fs_live_variables &live = s.live_analysis.require();
1426 
1427    /* First, walk through each block doing local copy propagation and getting
1428     * the set of copies available at the end of the block.
1429     */
1430    foreach_block (block, s.cfg) {
1431       progress = opt_copy_propagation_local(s.compiler, lin_ctx, block,
1432                                             out_acp[block->num], s.alloc,
1433                                             s.max_polygons) || progress;
1434 
1435       /* If the destination of an ACP entry exists only within this block,
1436        * then there's no need to keep it for dataflow analysis.  We can delete
1437        * it from the out_acp table and avoid growing the bitsets any bigger
1438        * than we absolutely have to.
1439        *
1440        * Because nothing in opt_copy_propagation_local touches the block
1441        * start/end IPs and opt_copy_propagation_local is incapable of
1442        * extending the live range of an ACP destination beyond the block,
1443        * it's safe to use the liveness information in this way.
1444        */
1445       for (auto iter = out_acp[block->num].begin();
1446            iter != out_acp[block->num].end(); ++iter) {
1447          assert((*iter)->dst.file == VGRF);
1448          if (block->start_ip <= live.vgrf_start[(*iter)->dst.nr] &&
1449              live.vgrf_end[(*iter)->dst.nr] <= block->end_ip) {
1450             out_acp[block->num].remove(*iter);
1451          }
1452       }
1453    }
1454 
1455    /* Do dataflow analysis for those available copies. */
1456    fs_copy_prop_dataflow dataflow(lin_ctx, s.cfg, live, out_acp);
1457 
1458    /* Next, re-run local copy propagation, this time with the set of copies
1459     * provided by the dataflow analysis available at the start of a block.
1460     */
1461    foreach_block (block, s.cfg) {
1462       struct acp in_acp;
1463 
1464       for (int i = 0; i < dataflow.num_acp; i++) {
1465          if (BITSET_TEST(dataflow.bd[block->num].livein, i) &&
1466              !BITSET_TEST(dataflow.bd[block->num].exec_mismatch, i)) {
1467             struct acp_entry *entry = dataflow.acp[i];
1468             in_acp.add(entry);
1469          }
1470       }
1471 
1472       progress = opt_copy_propagation_local(s.compiler, lin_ctx, block,
1473                                             in_acp, s.alloc, s.max_polygons) ||
1474                  progress;
1475    }
1476 
1477    ralloc_free(copy_prop_ctx);
1478 
1479    if (progress)
1480       s.invalidate_analysis(DEPENDENCY_INSTRUCTION_DATA_FLOW |
1481                             DEPENDENCY_INSTRUCTION_DETAIL);
1482 
1483    return progress;
1484 }
1485 
1486 static bool
try_copy_propagate_def(const brw_compiler * compiler,const brw::simple_allocator & alloc,fs_inst * def,const brw_reg & val,fs_inst * inst,int arg,uint8_t max_polygons)1487 try_copy_propagate_def(const brw_compiler *compiler,
1488                        const brw::simple_allocator &alloc,
1489                        fs_inst *def, const brw_reg &val,
1490                        fs_inst *inst, int arg,
1491                        uint8_t max_polygons)
1492 {
1493    const struct intel_device_info *devinfo = compiler->devinfo;
1494 
1495    assert(val.file != BAD_FILE);
1496 
1497    /* We can't generally copy-propagate UD negations because we can end up
1498     * accessing the resulting values as signed integers instead.
1499     */
1500    if (val.negate && val.type == BRW_TYPE_UD)
1501       return false;
1502 
1503    /* Bail if the instruction type is larger than the execution type of the
1504     * copy, what implies that each channel is reading multiple channels of the
1505     * destination of the copy, and simply replacing the sources would give a
1506     * program with different semantics.
1507     */
1508    if (inst->opcode != BRW_OPCODE_MOV &&
1509        brw_type_size_bits(def->dst.type) <
1510        brw_type_size_bits(inst->src[arg].type))
1511       return false;
1512 
1513    const bool has_source_modifiers = val.abs || val.negate;
1514 
1515    if (has_source_modifiers) {
1516       if (is_logic_op(inst->opcode) || !inst->can_do_source_mods(devinfo))
1517          return false;
1518 
1519       /* Since semantics of source modifiers are type-dependent we need to
1520        * ensure that the meaning of the instruction remains the same if we
1521        * change the type. If the sizes of the types are different the new
1522        * instruction will read a different amount of data than the original
1523        * and the semantics will always be different.
1524        */
1525       if (def->dst.type != inst->src[arg].type &&
1526           (!inst->can_change_types() ||
1527            brw_type_size_bits(def->dst.type) !=
1528            brw_type_size_bits(inst->src[arg].type)))
1529          return false;
1530    }
1531 
1532    /* Send messages with EOT set are restricted to use g112-g127 (and we
1533     * sometimes need g127 for other purposes), so avoid copy propagating
1534     * anything that would make it impossible to satisfy that restriction.
1535     */
1536    if (inst->eot) {
1537       /* Don't propagate things that are already pinned. */
1538       if (val.file != VGRF)
1539          return false;
1540 
1541       /* We might be propagating from a large register, while the SEND only
1542        * is reading a portion of it (say the .A channel in an RGBA value).
1543        * We need to pin both split SEND sources in g112-g126/127, so only
1544        * allow this if the registers aren't too large.
1545        */
1546       if (inst->opcode == SHADER_OPCODE_SEND && inst->sources >= 4 &&
1547           val.file == VGRF) {
1548          int other_src = arg == 2 ? 3 : 2;
1549          unsigned other_size = inst->src[other_src].file == VGRF ?
1550                                alloc.sizes[inst->src[other_src].nr] :
1551                                inst->size_read(other_src);
1552          unsigned prop_src_size = alloc.sizes[val.nr];
1553          if (other_size + prop_src_size > 15)
1554             return false;
1555       }
1556    }
1557 
1558    /* Reject cases that would violate register regioning restrictions. */
1559    if ((val.file == UNIFORM || !val.is_contiguous()) &&
1560        (inst->is_send_from_grf() || inst->uses_indirect_addressing())) {
1561       return false;
1562    }
1563 
1564    /* Some instructions implemented in the generator backend, such as
1565     * derivatives, assume that their operands are packed so we can't
1566     * generally propagate strided regions to them.
1567     */
1568    const unsigned entry_stride = val.file == FIXED_GRF ? 1 : val.stride;
1569    if (instruction_requires_packed_data(inst) && entry_stride != 1)
1570       return false;
1571 
1572    const brw_reg_type dst_type = (has_source_modifiers &&
1573                                   def->dst.type != inst->src[arg].type) ?
1574       def->dst.type : inst->dst.type;
1575 
1576    /* Bail if the result of composing both strides would exceed the
1577     * hardware limit.
1578     */
1579    if (!can_take_stride(inst, dst_type, arg,
1580                         entry_stride * inst->src[arg].stride,
1581                         compiler))
1582       return false;
1583 
1584    /* Bail if the source FIXED_GRF region of the copy cannot be trivially
1585     * composed with the source region of the instruction -- E.g. because the
1586     * copy uses some extended stride greater than 4 not supported natively by
1587     * the hardware as a horizontal stride, or because instruction compression
1588     * could require us to use a vertical stride shorter than a GRF.
1589     */
1590    if (val.file == FIXED_GRF &&
1591        (inst->src[arg].stride > 4 ||
1592         inst->dst.component_size(inst->exec_size) >
1593         inst->src[arg].component_size(inst->exec_size)))
1594       return false;
1595 
1596    /* Bail if the result of composing both strides cannot be expressed
1597     * as another stride. This avoids, for example, trying to transform
1598     * this:
1599     *
1600     *     MOV (8) rX<1>UD rY<0;1,0>UD
1601     *     FOO (8) ...     rX<8;8,1>UW
1602     *
1603     * into this:
1604     *
1605     *     FOO (8) ...     rY<0;1,0>UW
1606     *
1607     * Which would have different semantics.
1608     */
1609    if (entry_stride != 1 &&
1610        (inst->src[arg].stride *
1611         brw_type_size_bytes(inst->src[arg].type)) % brw_type_size_bytes(val.type) != 0)
1612       return false;
1613 
1614    /* From the Cherry Trail/Braswell PRMs, Volume 7: 3D Media GPGPU:
1615     *    EU Overview
1616     *       Register Region Restrictions
1617     *          Special Requirements for Handling Double Precision Data Types :
1618     *
1619     *   "When source or destination datatype is 64b or operation is integer
1620     *    DWord multiply, regioning in Align1 must follow these rules:
1621     *
1622     *      1. Source and Destination horizontal stride must be aligned to the
1623     *         same qword.
1624     *      2. Regioning must ensure Src.Vstride = Src.Width * Src.Hstride.
1625     *      3. Source and Destination offset must be the same, except the case
1626     *         of scalar source."
1627     *
1628     * Most of this is already checked in can_take_stride(), we're only left
1629     * with checking 3.
1630     */
1631    if (has_dst_aligned_region_restriction(devinfo, inst, dst_type) &&
1632        entry_stride != 0 &&
1633        (reg_offset(inst->dst) % (REG_SIZE * reg_unit(devinfo))) != (reg_offset(val) % (REG_SIZE * reg_unit(devinfo))))
1634       return false;
1635 
1636    /* The <8;8,0> regions used for FS attributes in multipolygon
1637     * dispatch mode could violate regioning restrictions, don't copy
1638     * propagate them in such cases.
1639     */
1640    if (max_polygons > 1 && val.file == ATTR &&
1641        (has_dst_aligned_region_restriction(devinfo, inst, dst_type) ||
1642         instruction_requires_packed_data(inst) ||
1643         (inst->is_3src(compiler) && arg == 2) ||
1644         def->dst.type != inst->src[arg].type))
1645       return false;
1646 
1647    /* Fold the copy into the instruction consuming it. */
1648    inst->src[arg].file = val.file;
1649    inst->src[arg].nr = val.nr;
1650    inst->src[arg].subnr = val.subnr;
1651    inst->src[arg].offset = val.offset;
1652 
1653    /* Compose the strides of both regions. */
1654    if (val.file == FIXED_GRF) {
1655       if (inst->src[arg].stride) {
1656          const unsigned orig_width = 1 << val.width;
1657          const unsigned reg_width =
1658             REG_SIZE / (brw_type_size_bytes(inst->src[arg].type) *
1659                         inst->src[arg].stride);
1660          inst->src[arg].width = cvt(MIN2(orig_width, reg_width)) - 1;
1661          inst->src[arg].hstride = cvt(inst->src[arg].stride);
1662          inst->src[arg].vstride = inst->src[arg].hstride + inst->src[arg].width;
1663       } else {
1664          inst->src[arg].vstride = inst->src[arg].hstride =
1665             inst->src[arg].width = 0;
1666       }
1667 
1668       inst->src[arg].stride = 1;
1669 
1670       /* Hopefully no Align16 around here... */
1671       assert(val.swizzle == BRW_SWIZZLE_XYZW);
1672       inst->src[arg].swizzle = val.swizzle;
1673    } else {
1674       inst->src[arg].stride *= val.stride;
1675    }
1676 
1677    /* Handle NoMask cases where the def replicates a small scalar to a number
1678     * of channels, but the use is a lower SIMD width but larger type, so each
1679     * invocation reads multiple channels worth of data, e.g.
1680     *
1681     *    mov(16) vgrf1:UW, u0<0>:UW NoMask
1682     *    mov(8)  vgrf2:UD, vgrf1:UD NoMask group0
1683     *
1684     * In this case, we should just use the scalar's type.
1685     */
1686    if (val.stride == 0 &&
1687        inst->opcode == BRW_OPCODE_MOV &&
1688        inst->force_writemask_all && def->force_writemask_all &&
1689        inst->exec_size < def->exec_size &&
1690        (inst->exec_size * brw_type_size_bytes(inst->src[arg].type) ==
1691         def->exec_size * brw_type_size_bytes(val.type))) {
1692       inst->src[arg].type = val.type;
1693       inst->dst.type = val.type;
1694       inst->exec_size = def->exec_size;
1695    }
1696 
1697    if (has_source_modifiers) {
1698       if (def->dst.type != inst->src[arg].type) {
1699          /* We are propagating source modifiers from a MOV with a different
1700           * type.  If we got here, then we can just change the source and
1701           * destination types of the instruction and keep going.
1702           */
1703          for (int i = 0; i < inst->sources; i++) {
1704             inst->src[i].type = def->dst.type;
1705          }
1706          inst->dst.type = def->dst.type;
1707       }
1708 
1709       if (!inst->src[arg].abs) {
1710          inst->src[arg].abs = val.abs;
1711          inst->src[arg].negate ^= val.negate;
1712       }
1713    }
1714 
1715    return true;
1716 }
1717 
1718 static bool
try_constant_propagate_def(fs_inst * def,brw_reg val,fs_inst * inst,int arg)1719 try_constant_propagate_def(fs_inst *def, brw_reg val, fs_inst *inst, int arg)
1720 {
1721    /* Bail if inst is reading more than a single vector component of entry */
1722    if (inst->size_read(arg) > def->dst.component_size(inst->exec_size))
1723       return false;
1724 
1725    return try_constant_propagate_value(val, def->dst.type, inst, arg);
1726 }
1727 
1728 /**
1729  * Handle cases like UW subreads of a UD immediate, with an offset.
1730  */
1731 static brw_reg
extract_imm(brw_reg val,brw_reg_type type,unsigned offset)1732 extract_imm(brw_reg val, brw_reg_type type, unsigned offset)
1733 {
1734    assert(val.file == IMM);
1735 
1736    const unsigned bitsize = brw_type_size_bits(type);
1737 
1738    if (offset == 0 || bitsize == brw_type_size_bits(val.type))
1739       return val;
1740 
1741    assert(bitsize < brw_type_size_bits(val.type));
1742 
1743    val.u64 = (val.u64 >> (bitsize * offset)) & ((1ull << bitsize) - 1);
1744 
1745    return val;
1746 }
1747 
1748 static brw_reg
find_value_for_offset(fs_inst * def,const brw_reg & src,unsigned src_size)1749 find_value_for_offset(fs_inst *def, const brw_reg &src, unsigned src_size)
1750 {
1751    brw_reg val;
1752 
1753    switch (def->opcode) {
1754    case BRW_OPCODE_MOV:
1755       /* is_raw_move also rejects source modifiers, but copy propagation
1756        * can handle that if the tyeps are the same.
1757        */
1758       if ((def->dst.type == def->src[0].type || def->is_raw_move()) &&
1759           def->src[0].stride <= 1) {
1760          val = def->src[0];
1761 
1762          unsigned rel_offset = src.offset - def->dst.offset;
1763 
1764          if (val.stride == 0)
1765             rel_offset %= brw_type_size_bytes(def->dst.type);
1766 
1767          if (val.file == IMM)
1768             val = extract_imm(val, src.type, rel_offset);
1769          else
1770             val = byte_offset(def->src[0], rel_offset);
1771       }
1772       break;
1773    case SHADER_OPCODE_LOAD_PAYLOAD: {
1774       unsigned offset = 0;
1775       for (int i = def->header_size; i < def->sources; i++) {
1776          const unsigned splat = def->src[i].stride == 0 ? def->exec_size : 1;
1777          if (offset == src.offset) {
1778             if (def->dst.type == def->src[i].type &&
1779                 def->src[i].stride <= 1 &&
1780                 def->src[i].component_size(def->exec_size) * splat == src_size)
1781                val = def->src[i];
1782 
1783             break;
1784          }
1785 
1786          offset += def->exec_size * brw_type_size_bytes(def->src[i].type);
1787       }
1788       break;
1789    }
1790    default:
1791       break;
1792    }
1793 
1794    return val;
1795 }
1796 
1797 bool
brw_fs_opt_copy_propagation_defs(fs_visitor & s)1798 brw_fs_opt_copy_propagation_defs(fs_visitor &s)
1799 {
1800    const brw::def_analysis &defs = s.def_analysis.require();
1801    unsigned *uses_deleted = new unsigned[defs.count()]();
1802    bool progress = false;
1803 
1804    foreach_block_and_inst_safe(block, fs_inst, inst, s.cfg) {
1805       /* Try propagating into this instruction. */
1806       bool instruction_progress = false;
1807 
1808       for (int i = inst->sources - 1; i >= 0; i--) {
1809          fs_inst *def = defs.get(inst->src[i]);
1810 
1811          if (!def || def->saturate)
1812             continue;
1813 
1814          bool source_progress = false;
1815 
1816          if (def->opcode == SHADER_OPCODE_LOAD_PAYLOAD) {
1817             if (inst->size_read(i) == def->size_written &&
1818                 def->src[0].file != BAD_FILE && def->src[0].file != IMM &&
1819                 is_identity_payload(def->src[0].file, def)) {
1820                source_progress =
1821                   try_copy_propagate_def(s.compiler, s.alloc, def, def->src[0],
1822                                          inst, i, s.max_polygons);
1823 
1824                if (source_progress) {
1825                   instruction_progress = true;
1826                   ++uses_deleted[def->dst.nr];
1827                   if (defs.get_use_count(def->dst) == uses_deleted[def->dst.nr])
1828                      def->remove(defs.get_block(def->dst), true);
1829                }
1830 
1831                continue;
1832             }
1833          }
1834 
1835          brw_reg val =
1836             find_value_for_offset(def, inst->src[i], inst->size_read(i));
1837 
1838          if (val.file == IMM) {
1839             source_progress =
1840                try_constant_propagate_def(def, val, inst, i);
1841          } else if (val.file == VGRF ||
1842                     val.file == ATTR || val.file == UNIFORM ||
1843                     (val.file == FIXED_GRF && val.is_contiguous())) {
1844             source_progress =
1845                try_copy_propagate_def(s.compiler, s.alloc, def, val, inst, i,
1846                                       s.max_polygons);
1847          }
1848 
1849          if (source_progress) {
1850             instruction_progress = true;
1851             ++uses_deleted[def->dst.nr];
1852             if (defs.get_use_count(def->dst) == uses_deleted[def->dst.nr])
1853                def->remove(defs.get_block(def->dst), true);
1854          }
1855       }
1856 
1857       if (instruction_progress) {
1858          progress = true;
1859          commute_immediates(inst);
1860       }
1861    }
1862 
1863    if (progress) {
1864       s.cfg->adjust_block_ips();
1865       s.invalidate_analysis(DEPENDENCY_INSTRUCTION_DATA_FLOW |
1866                             DEPENDENCY_INSTRUCTION_DETAIL);
1867    }
1868 
1869    delete [] uses_deleted;
1870 
1871    return progress;
1872 }
1873