1/*
2 * Copyright (C) 2014 Rob Clark <robclark@freedesktop.org>
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 FROM,
20 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
21 * SOFTWARE.
22 *
23 * Authors:
24 *    Rob Clark <robclark@freedesktop.org>
25 */
26
27#include <math.h>
28#include "util/half_float.h"
29#include "util/u_math.h"
30
31#include "ir3.h"
32#include "ir3_compiler.h"
33#include "ir3_shader.h"
34
35#define swap(a, b)                                                             \
36   do {                                                                        \
37      __typeof(a) __tmp = (a);                                                 \
38      (a) = (b);                                                               \
39      (b) = __tmp;                                                             \
40   } while (0)
41
42/*
43 * Copy Propagate:
44 */
45
46struct ir3_cp_ctx {
47   struct ir3 *shader;
48   struct ir3_shader_variant *so;
49   bool progress;
50};
51
52/* is it a type preserving mov, with ok flags?
53 *
54 * @instr: the mov to consider removing
55 * @dst_instr: the instruction consuming the mov (instr)
56 *
57 * TODO maybe drop allow_flags since this is only false when dst is
58 * NULL (ie. outputs)
59 */
60static bool
61is_eligible_mov(struct ir3_instruction *instr,
62                struct ir3_instruction *dst_instr, bool allow_flags)
63{
64   if (is_same_type_mov(instr)) {
65      struct ir3_register *dst = instr->dsts[0];
66      struct ir3_register *src = instr->srcs[0];
67      struct ir3_instruction *src_instr = ssa(src);
68
69      /* only if mov src is SSA (not const/immed): */
70      if (!src_instr)
71         return false;
72
73      /* no indirect: */
74      if (dst->flags & IR3_REG_RELATIV)
75         return false;
76      if (src->flags & IR3_REG_RELATIV)
77         return false;
78
79      if (src->flags & IR3_REG_ARRAY)
80         return false;
81
82      if (!allow_flags)
83         if (src->flags & (IR3_REG_FABS | IR3_REG_FNEG | IR3_REG_SABS |
84                           IR3_REG_SNEG | IR3_REG_BNOT))
85            return false;
86
87      return true;
88   }
89   return false;
90}
91
92/* we can end up with extra cmps.s from frontend, which uses a
93 *
94 *    cmps.s p0.x, cond, 0
95 *
96 * as a way to mov into the predicate register.  But frequently 'cond'
97 * is itself a cmps.s/cmps.f/cmps.u. So detect this special case.
98 */
99static bool
100is_foldable_double_cmp(struct ir3_instruction *cmp)
101{
102   struct ir3_instruction *cond = ssa(cmp->srcs[0]);
103   return (cmp->dsts[0]->num == regid(REG_P0, 0)) && cond &&
104          (cmp->srcs[1]->flags & IR3_REG_IMMED) &&
105          (cmp->srcs[1]->iim_val == 0) &&
106          (cmp->cat2.condition == IR3_COND_NE) &&
107          (!cond->address || cond->address->def->instr->block == cmp->block);
108}
109
110/* propagate register flags from src to dst.. negates need special
111 * handling to cancel each other out.
112 */
113static void
114combine_flags(unsigned *dstflags, struct ir3_instruction *src)
115{
116   unsigned srcflags = src->srcs[0]->flags;
117
118   /* if what we are combining into already has (abs) flags,
119    * we can drop (neg) from src:
120    */
121   if (*dstflags & IR3_REG_FABS)
122      srcflags &= ~IR3_REG_FNEG;
123   if (*dstflags & IR3_REG_SABS)
124      srcflags &= ~IR3_REG_SNEG;
125
126   if (srcflags & IR3_REG_FABS)
127      *dstflags |= IR3_REG_FABS;
128   if (srcflags & IR3_REG_SABS)
129      *dstflags |= IR3_REG_SABS;
130   if (srcflags & IR3_REG_FNEG)
131      *dstflags ^= IR3_REG_FNEG;
132   if (srcflags & IR3_REG_SNEG)
133      *dstflags ^= IR3_REG_SNEG;
134   if (srcflags & IR3_REG_BNOT)
135      *dstflags ^= IR3_REG_BNOT;
136
137   *dstflags &= ~IR3_REG_SSA;
138   *dstflags |= srcflags & IR3_REG_SSA;
139   *dstflags |= srcflags & IR3_REG_CONST;
140   *dstflags |= srcflags & IR3_REG_IMMED;
141   *dstflags |= srcflags & IR3_REG_RELATIV;
142   *dstflags |= srcflags & IR3_REG_ARRAY;
143   *dstflags |= srcflags & IR3_REG_SHARED;
144
145   /* if src of the src is boolean we can drop the (abs) since we know
146    * the source value is already a postitive integer.  This cleans
147    * up the absnegs that get inserted when converting between nir and
148    * native boolean (see ir3_b2n/n2b)
149    */
150   struct ir3_instruction *srcsrc = ssa(src->srcs[0]);
151   if (srcsrc && is_bool(srcsrc))
152      *dstflags &= ~IR3_REG_SABS;
153}
154
155/* Tries lowering an immediate register argument to a const buffer access by
156 * adding to the list of immediates to be pushed to the const buffer when
157 * switching to this shader.
158 */
159static bool
160lower_immed(struct ir3_cp_ctx *ctx, struct ir3_instruction *instr, unsigned n,
161            struct ir3_register *reg, unsigned new_flags)
162{
163   if (!(new_flags & IR3_REG_IMMED))
164      return false;
165
166   new_flags &= ~IR3_REG_IMMED;
167   new_flags |= IR3_REG_CONST;
168
169   if (!ir3_valid_flags(instr, n, new_flags))
170      return false;
171
172   reg = ir3_reg_clone(ctx->shader, reg);
173
174   /* Half constant registers seems to handle only 32-bit values
175    * within floating-point opcodes. So convert back to 32-bit values.
176    */
177   bool f_opcode =
178      (is_cat2_float(instr->opc) || is_cat3_float(instr->opc)) ? true : false;
179   if (f_opcode && (new_flags & IR3_REG_HALF))
180      reg->uim_val = fui(_mesa_half_to_float(reg->uim_val));
181
182   /* in some cases, there are restrictions on (abs)/(neg) plus const..
183    * so just evaluate those and clear the flags:
184    */
185   if (new_flags & IR3_REG_SABS) {
186      reg->iim_val = abs(reg->iim_val);
187      new_flags &= ~IR3_REG_SABS;
188   }
189
190   if (new_flags & IR3_REG_FABS) {
191      reg->fim_val = fabs(reg->fim_val);
192      new_flags &= ~IR3_REG_FABS;
193   }
194
195   if (new_flags & IR3_REG_SNEG) {
196      reg->iim_val = -reg->iim_val;
197      new_flags &= ~IR3_REG_SNEG;
198   }
199
200   if (new_flags & IR3_REG_FNEG) {
201      reg->fim_val = -reg->fim_val;
202      new_flags &= ~IR3_REG_FNEG;
203   }
204
205   /* Reallocate for 4 more elements whenever it's necessary.  Note that ir3
206    * printing relies on having groups of 4 dwords, so we fill the unused
207    * slots with a dummy value.
208    */
209   struct ir3_const_state *const_state = ir3_const_state(ctx->so);
210   if (const_state->immediates_count == const_state->immediates_size) {
211      const_state->immediates = rerzalloc(
212         const_state, const_state->immediates,
213         __typeof__(const_state->immediates[0]), const_state->immediates_size,
214         const_state->immediates_size + 4);
215      const_state->immediates_size += 4;
216
217      for (int i = const_state->immediates_count;
218           i < const_state->immediates_size; i++)
219         const_state->immediates[i] = 0xd0d0d0d0;
220   }
221
222   int i;
223   for (i = 0; i < const_state->immediates_count; i++) {
224      if (const_state->immediates[i] == reg->uim_val)
225         break;
226   }
227
228   if (i == const_state->immediates_count) {
229      /* Add on a new immediate to be pushed, if we have space left in the
230       * constbuf.
231       */
232      if (const_state->offsets.immediate + const_state->immediates_count / 4 >=
233          ir3_max_const(ctx->so))
234         return false;
235
236      const_state->immediates[i] = reg->uim_val;
237      const_state->immediates_count++;
238   }
239
240   reg->flags = new_flags;
241   reg->num = i + (4 * const_state->offsets.immediate);
242
243   instr->srcs[n] = reg;
244
245   return true;
246}
247
248static void
249unuse(struct ir3_instruction *instr)
250{
251   debug_assert(instr->use_count > 0);
252
253   if (--instr->use_count == 0) {
254      struct ir3_block *block = instr->block;
255
256      instr->barrier_class = 0;
257      instr->barrier_conflict = 0;
258
259      /* we don't want to remove anything in keeps (which could
260       * be things like array store's)
261       */
262      for (unsigned i = 0; i < block->keeps_count; i++) {
263         debug_assert(block->keeps[i] != instr);
264      }
265   }
266}
267
268/**
269 * Handles the special case of the 2nd src (n == 1) to "normal" mad
270 * instructions, which cannot reference a constant.  See if it is
271 * possible to swap the 1st and 2nd sources.
272 */
273static bool
274try_swap_mad_two_srcs(struct ir3_instruction *instr, unsigned new_flags)
275{
276   if (!is_mad(instr->opc))
277      return false;
278
279   /* NOTE: pre-swap first two src's before valid_flags(),
280    * which might try to dereference the n'th src:
281    */
282   swap(instr->srcs[0], instr->srcs[1]);
283
284   /* cat3 doesn't encode immediate, but we can lower immediate
285    * to const if that helps:
286    */
287   if (new_flags & IR3_REG_IMMED) {
288      new_flags &= ~IR3_REG_IMMED;
289      new_flags |= IR3_REG_CONST;
290   }
291
292   bool valid_swap =
293      /* can we propagate mov if we move 2nd src to first? */
294      ir3_valid_flags(instr, 0, new_flags) &&
295      /* and does first src fit in second slot? */
296      ir3_valid_flags(instr, 1, instr->srcs[1]->flags);
297
298   if (!valid_swap) {
299      /* put things back the way they were: */
300      swap(instr->srcs[0], instr->srcs[1]);
301   } /* otherwise leave things swapped */
302
303   return valid_swap;
304}
305
306/* Values that are uniform inside a loop can become divergent outside
307 * it if the loop has a divergent trip count. This means that we can't
308 * propagate a copy of a shared to non-shared register if it would
309 * make the shared reg's live range extend outside of its loop. Users
310 * outside the loop would see the value for the thread(s) that last
311 * exited the loop, rather than for their own thread.
312 */
313static bool
314is_valid_shared_copy(struct ir3_instruction *dst_instr,
315                     struct ir3_instruction *src_instr,
316                     struct ir3_register *src_reg)
317{
318   return !(src_reg->flags & IR3_REG_SHARED) ||
319      dst_instr->block->loop_id == src_instr->block->loop_id;
320}
321
322/**
323 * Handle cp for a given src register.  This additionally handles
324 * the cases of collapsing immedate/const (which replace the src
325 * register with a non-ssa src) or collapsing mov's from relative
326 * src (which needs to also fixup the address src reference by the
327 * instruction).
328 */
329static bool
330reg_cp(struct ir3_cp_ctx *ctx, struct ir3_instruction *instr,
331       struct ir3_register *reg, unsigned n)
332{
333   struct ir3_instruction *src = ssa(reg);
334
335   if (is_eligible_mov(src, instr, true)) {
336      /* simple case, no immed/const/relativ, only mov's w/ ssa src: */
337      struct ir3_register *src_reg = src->srcs[0];
338      unsigned new_flags = reg->flags;
339
340      if (!is_valid_shared_copy(instr, src, src_reg))
341         return false;
342
343      combine_flags(&new_flags, src);
344
345      if (ir3_valid_flags(instr, n, new_flags)) {
346         if (new_flags & IR3_REG_ARRAY) {
347            debug_assert(!(reg->flags & IR3_REG_ARRAY));
348            reg->array = src_reg->array;
349         }
350         reg->flags = new_flags;
351         reg->def = src_reg->def;
352
353         instr->barrier_class |= src->barrier_class;
354         instr->barrier_conflict |= src->barrier_conflict;
355
356         unuse(src);
357         reg->def->instr->use_count++;
358
359         return true;
360      }
361   } else if ((is_same_type_mov(src) || is_const_mov(src)) &&
362              /* cannot collapse const/immed/etc into control flow: */
363              opc_cat(instr->opc) != 0) {
364      /* immed/const/etc cases, which require some special handling: */
365      struct ir3_register *src_reg = src->srcs[0];
366      unsigned new_flags = reg->flags;
367
368      if (!is_valid_shared_copy(instr, src, src_reg))
369         return false;
370
371      if (src_reg->flags & IR3_REG_ARRAY)
372         return false;
373
374      combine_flags(&new_flags, src);
375
376      if (!ir3_valid_flags(instr, n, new_flags)) {
377         /* See if lowering an immediate to const would help. */
378         if (lower_immed(ctx, instr, n, src_reg, new_flags))
379            return true;
380
381         /* special case for "normal" mad instructions, we can
382          * try swapping the first two args if that fits better.
383          *
384          * the "plain" MAD's (ie. the ones that don't shift first
385          * src prior to multiply) can swap their first two srcs if
386          * src[0] is !CONST and src[1] is CONST:
387          */
388         if ((n == 1) && try_swap_mad_two_srcs(instr, new_flags)) {
389            return true;
390         } else {
391            return false;
392         }
393      }
394
395      /* Here we handle the special case of mov from
396       * CONST and/or RELATIV.  These need to be handled
397       * specially, because in the case of move from CONST
398       * there is no src ir3_instruction so we need to
399       * replace the ir3_register.  And in the case of
400       * RELATIV we need to handle the address register
401       * dependency.
402       */
403      if (src_reg->flags & IR3_REG_CONST) {
404         /* an instruction cannot reference two different
405          * address registers:
406          */
407         if ((src_reg->flags & IR3_REG_RELATIV) &&
408             conflicts(instr->address, reg->def->instr->address))
409            return false;
410
411         /* This seems to be a hw bug, or something where the timings
412          * just somehow don't work out.  This restriction may only
413          * apply if the first src is also CONST.
414          */
415         if ((opc_cat(instr->opc) == 3) && (n == 2) &&
416             (src_reg->flags & IR3_REG_RELATIV) && (src_reg->array.offset == 0))
417            return false;
418
419         /* When narrowing constant from 32b to 16b, it seems
420          * to work only for float. So we should do this only with
421          * float opcodes.
422          */
423         if (src->cat1.dst_type == TYPE_F16) {
424            /* TODO: should we have a way to tell phi/collect to use a
425             * float move so that this is legal?
426             */
427            if (is_meta(instr))
428               return false;
429            if (instr->opc == OPC_MOV && !type_float(instr->cat1.src_type))
430               return false;
431            if (!is_cat2_float(instr->opc) && !is_cat3_float(instr->opc))
432               return false;
433         } else if (src->cat1.dst_type == TYPE_U16) {
434            /* Since we set CONSTANT_DEMOTION_ENABLE, a float reference of
435             * what was a U16 value read from the constbuf would incorrectly
436             * do 32f->16f conversion, when we want to read a 16f value.
437             */
438            if (is_cat2_float(instr->opc) || is_cat3_float(instr->opc))
439               return false;
440            if (instr->opc == OPC_MOV && type_float(instr->cat1.src_type))
441               return false;
442         }
443
444         src_reg = ir3_reg_clone(instr->block->shader, src_reg);
445         src_reg->flags = new_flags;
446         instr->srcs[n] = src_reg;
447
448         if (src_reg->flags & IR3_REG_RELATIV)
449            ir3_instr_set_address(instr, reg->def->instr->address->def->instr);
450
451         return true;
452      }
453
454      if (src_reg->flags & IR3_REG_IMMED) {
455         int32_t iim_val = src_reg->iim_val;
456
457         debug_assert((opc_cat(instr->opc) == 1) ||
458                      (opc_cat(instr->opc) == 2) ||
459                      (opc_cat(instr->opc) == 6) ||
460                      is_meta(instr) ||
461                      (is_mad(instr->opc) && (n == 0)));
462
463         if ((opc_cat(instr->opc) == 2) &&
464               !ir3_cat2_int(instr->opc)) {
465            iim_val = ir3_flut(src_reg);
466            if (iim_val < 0) {
467               /* Fall back to trying to load the immediate as a const: */
468               return lower_immed(ctx, instr, n, src_reg, new_flags);
469            }
470         }
471
472         if (new_flags & IR3_REG_SABS)
473            iim_val = abs(iim_val);
474
475         if (new_flags & IR3_REG_SNEG)
476            iim_val = -iim_val;
477
478         if (new_flags & IR3_REG_BNOT)
479            iim_val = ~iim_val;
480
481         if (ir3_valid_flags(instr, n, new_flags) &&
482             ir3_valid_immediate(instr, iim_val)) {
483            new_flags &= ~(IR3_REG_SABS | IR3_REG_SNEG | IR3_REG_BNOT);
484            src_reg = ir3_reg_clone(instr->block->shader, src_reg);
485            src_reg->flags = new_flags;
486            src_reg->iim_val = iim_val;
487            instr->srcs[n] = src_reg;
488
489            return true;
490         } else {
491            /* Fall back to trying to load the immediate as a const: */
492            return lower_immed(ctx, instr, n, src_reg, new_flags);
493         }
494      }
495   }
496
497   return false;
498}
499
500/* Handle special case of eliminating output mov, and similar cases where
501 * there isn't a normal "consuming" instruction.  In this case we cannot
502 * collapse flags (ie. output mov from const, or w/ abs/neg flags, cannot
503 * be eliminated)
504 */
505static struct ir3_instruction *
506eliminate_output_mov(struct ir3_cp_ctx *ctx, struct ir3_instruction *instr)
507{
508   if (is_eligible_mov(instr, NULL, false)) {
509      struct ir3_register *reg = instr->srcs[0];
510      if (!(reg->flags & IR3_REG_ARRAY)) {
511         struct ir3_instruction *src_instr = ssa(reg);
512         debug_assert(src_instr);
513         ctx->progress = true;
514         return src_instr;
515      }
516   }
517   return instr;
518}
519
520/**
521 * Find instruction src's which are mov's that can be collapsed, replacing
522 * the mov dst with the mov src
523 */
524static void
525instr_cp(struct ir3_cp_ctx *ctx, struct ir3_instruction *instr)
526{
527   if (instr->srcs_count == 0)
528      return;
529
530   if (ir3_instr_check_mark(instr))
531      return;
532
533   /* walk down the graph from each src: */
534   bool progress;
535   do {
536      progress = false;
537      foreach_src_n (reg, n, instr) {
538         struct ir3_instruction *src = ssa(reg);
539
540         if (!src)
541            continue;
542
543         instr_cp(ctx, src);
544
545         /* TODO non-indirect access we could figure out which register
546          * we actually want and allow cp..
547          */
548         if ((reg->flags & IR3_REG_ARRAY) && src->opc != OPC_META_PHI)
549            continue;
550
551         /* Don't CP absneg into meta instructions, that won't end well: */
552         if (is_meta(instr) &&
553             (src->opc == OPC_ABSNEG_F || src->opc == OPC_ABSNEG_S))
554            continue;
555
556         /* Don't CP mova and mova1 into their users */
557         if (writes_addr0(src) || writes_addr1(src))
558            continue;
559
560         progress |= reg_cp(ctx, instr, reg, n);
561         ctx->progress |= progress;
562      }
563   } while (progress);
564
565   /* After folding a mov's source we may wind up with a type-converting mov
566    * of an immediate. This happens e.g. with texture descriptors, since we
567    * narrow the descriptor (which may be a constant) to a half-reg in ir3.
568    * By converting the immediate in-place to the destination type, we can
569    * turn the mov into a same-type mov so that it can be further propagated.
570    */
571   if (instr->opc == OPC_MOV && (instr->srcs[0]->flags & IR3_REG_IMMED) &&
572       instr->cat1.src_type != instr->cat1.dst_type &&
573       /* Only do uint types for now, until we generate other types of
574        * mov's during instruction selection.
575        */
576       full_type(instr->cat1.src_type) == TYPE_U32 &&
577       full_type(instr->cat1.dst_type) == TYPE_U32) {
578      uint32_t uimm = instr->srcs[0]->uim_val;
579      if (instr->cat1.dst_type == TYPE_U16)
580         uimm &= 0xffff;
581      instr->srcs[0]->uim_val = uimm;
582      if (instr->dsts[0]->flags & IR3_REG_HALF)
583         instr->srcs[0]->flags |= IR3_REG_HALF;
584      else
585         instr->srcs[0]->flags &= ~IR3_REG_HALF;
586      instr->cat1.src_type = instr->cat1.dst_type;
587      ctx->progress = true;
588   }
589
590   /* Re-write the instruction writing predicate register to get rid
591    * of the double cmps.
592    */
593   if ((instr->opc == OPC_CMPS_S) && is_foldable_double_cmp(instr)) {
594      struct ir3_instruction *cond = ssa(instr->srcs[0]);
595      switch (cond->opc) {
596      case OPC_CMPS_S:
597      case OPC_CMPS_F:
598      case OPC_CMPS_U:
599         instr->opc = cond->opc;
600         instr->flags = cond->flags;
601         instr->cat2 = cond->cat2;
602         if (cond->address)
603            ir3_instr_set_address(instr, cond->address->def->instr);
604         instr->srcs[0] = ir3_reg_clone(ctx->shader, cond->srcs[0]);
605         instr->srcs[1] = ir3_reg_clone(ctx->shader, cond->srcs[1]);
606         instr->barrier_class |= cond->barrier_class;
607         instr->barrier_conflict |= cond->barrier_conflict;
608         unuse(cond);
609         ctx->progress = true;
610         break;
611      default:
612         break;
613      }
614   }
615
616   /* Handle converting a sam.s2en (taking samp/tex idx params via register)
617    * into a normal sam (encoding immediate samp/tex idx) if they are
618    * immediate. This saves some instructions and regs in the common case
619    * where we know samp/tex at compile time. This needs to be done in the
620    * frontend for bindless tex, though, so don't replicate it here.
621    */
622   if (is_tex(instr) && (instr->flags & IR3_INSTR_S2EN) &&
623       !(instr->flags & IR3_INSTR_B) &&
624       !(ir3_shader_debug & IR3_DBG_FORCES2EN)) {
625      /* The first src will be a collect, if both of it's
626       * two sources are mov from imm, then we can
627       */
628      struct ir3_instruction *samp_tex = ssa(instr->srcs[0]);
629
630      debug_assert(samp_tex->opc == OPC_META_COLLECT);
631
632      struct ir3_register *samp = samp_tex->srcs[0];
633      struct ir3_register *tex = samp_tex->srcs[1];
634
635      if ((samp->flags & IR3_REG_IMMED) && (tex->flags & IR3_REG_IMMED)) {
636         instr->flags &= ~IR3_INSTR_S2EN;
637         instr->cat5.samp = samp->iim_val;
638         instr->cat5.tex = tex->iim_val;
639
640         /* shuffle around the regs to remove the first src: */
641         instr->srcs_count--;
642         for (unsigned i = 0; i < instr->srcs_count; i++) {
643            instr->srcs[i] = instr->srcs[i + 1];
644         }
645
646         ctx->progress = true;
647      }
648   }
649}
650
651bool
652ir3_cp(struct ir3 *ir, struct ir3_shader_variant *so)
653{
654   struct ir3_cp_ctx ctx = {
655      .shader = ir,
656      .so = so,
657   };
658
659   /* This is a bit annoying, and probably wouldn't be necessary if we
660    * tracked a reverse link from producing instruction to consumer.
661    * But we need to know when we've eliminated the last consumer of
662    * a mov, so we need to do a pass to first count consumers of a
663    * mov.
664    */
665   foreach_block (block, &ir->block_list) {
666      foreach_instr (instr, &block->instr_list) {
667
668         /* by the way, we don't account for false-dep's, so the CP
669          * pass should always happen before false-dep's are inserted
670          */
671         debug_assert(instr->deps_count == 0);
672
673         foreach_ssa_src (src, instr) {
674            src->use_count++;
675         }
676      }
677   }
678
679   ir3_clear_mark(ir);
680
681   foreach_block (block, &ir->block_list) {
682      if (block->condition) {
683         instr_cp(&ctx, block->condition);
684         block->condition = eliminate_output_mov(&ctx, block->condition);
685      }
686
687      for (unsigned i = 0; i < block->keeps_count; i++) {
688         instr_cp(&ctx, block->keeps[i]);
689         block->keeps[i] = eliminate_output_mov(&ctx, block->keeps[i]);
690      }
691   }
692
693   return ctx.progress;
694}
695