1/*
2 * Copyright © 2014 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 * Authors:
24 *    Jason Ekstrand (jason@jlekstrand.net)
25 *
26 */
27
28#include <inttypes.h>
29#include "nir_search.h"
30#include "nir_builder.h"
31#include "nir_worklist.h"
32#include "util/half_float.h"
33
34/* This should be the same as nir_search_max_comm_ops in nir_algebraic.py. */
35#define NIR_SEARCH_MAX_COMM_OPS 8
36
37struct match_state {
38   bool inexact_match;
39   bool has_exact_alu;
40   uint8_t comm_op_direction;
41   unsigned variables_seen;
42
43   /* Used for running the automaton on newly-constructed instructions. */
44   struct util_dynarray *states;
45   const struct per_op_table *pass_op_table;
46
47   nir_alu_src variables[NIR_SEARCH_MAX_VARIABLES];
48   struct hash_table *range_ht;
49};
50
51static bool
52match_expression(const nir_search_expression *expr, nir_alu_instr *instr,
53                 unsigned num_components, const uint8_t *swizzle,
54                 struct match_state *state);
55static bool
56nir_algebraic_automaton(nir_instr *instr, struct util_dynarray *states,
57                        const struct per_op_table *pass_op_table);
58
59static const uint8_t identity_swizzle[NIR_MAX_VEC_COMPONENTS] =
60{
61    0,  1,  2,  3,
62    4,  5,  6,  7,
63    8,  9, 10, 11,
64   12, 13, 14, 15,
65};
66
67/**
68 * Check if a source produces a value of the given type.
69 *
70 * Used for satisfying 'a@type' constraints.
71 */
72static bool
73src_is_type(nir_src src, nir_alu_type type)
74{
75   assert(type != nir_type_invalid);
76
77   if (!src.is_ssa)
78      return false;
79
80   if (src.ssa->parent_instr->type == nir_instr_type_alu) {
81      nir_alu_instr *src_alu = nir_instr_as_alu(src.ssa->parent_instr);
82      nir_alu_type output_type = nir_op_infos[src_alu->op].output_type;
83
84      if (type == nir_type_bool) {
85         switch (src_alu->op) {
86         case nir_op_iand:
87         case nir_op_ior:
88         case nir_op_ixor:
89            return src_is_type(src_alu->src[0].src, nir_type_bool) &&
90                   src_is_type(src_alu->src[1].src, nir_type_bool);
91         case nir_op_inot:
92            return src_is_type(src_alu->src[0].src, nir_type_bool);
93         default:
94            break;
95         }
96      }
97
98      return nir_alu_type_get_base_type(output_type) == type;
99   } else if (src.ssa->parent_instr->type == nir_instr_type_intrinsic) {
100      nir_intrinsic_instr *intr = nir_instr_as_intrinsic(src.ssa->parent_instr);
101
102      if (type == nir_type_bool) {
103         return intr->intrinsic == nir_intrinsic_load_front_face ||
104                intr->intrinsic == nir_intrinsic_load_helper_invocation;
105      }
106   }
107
108   /* don't know */
109   return false;
110}
111
112static bool
113nir_op_matches_search_op(nir_op nop, uint16_t sop)
114{
115   if (sop <= nir_last_opcode)
116      return nop == sop;
117
118#define MATCH_FCONV_CASE(op) \
119   case nir_search_op_##op: \
120      return nop == nir_op_##op##16 || \
121             nop == nir_op_##op##32 || \
122             nop == nir_op_##op##64;
123
124#define MATCH_ICONV_CASE(op) \
125   case nir_search_op_##op: \
126      return nop == nir_op_##op##8 || \
127             nop == nir_op_##op##16 || \
128             nop == nir_op_##op##32 || \
129             nop == nir_op_##op##64;
130
131#define MATCH_BCONV_CASE(op) \
132   case nir_search_op_##op: \
133      return nop == nir_op_##op##1 || \
134             nop == nir_op_##op##32;
135
136   switch (sop) {
137   MATCH_FCONV_CASE(i2f)
138   MATCH_FCONV_CASE(u2f)
139   MATCH_FCONV_CASE(f2f)
140   MATCH_ICONV_CASE(f2u)
141   MATCH_ICONV_CASE(f2i)
142   MATCH_ICONV_CASE(u2u)
143   MATCH_ICONV_CASE(i2i)
144   MATCH_FCONV_CASE(b2f)
145   MATCH_ICONV_CASE(b2i)
146   MATCH_BCONV_CASE(i2b)
147   MATCH_BCONV_CASE(f2b)
148   default:
149      unreachable("Invalid nir_search_op");
150   }
151
152#undef MATCH_FCONV_CASE
153#undef MATCH_ICONV_CASE
154#undef MATCH_BCONV_CASE
155}
156
157uint16_t
158nir_search_op_for_nir_op(nir_op nop)
159{
160#define MATCH_FCONV_CASE(op) \
161   case nir_op_##op##16: \
162   case nir_op_##op##32: \
163   case nir_op_##op##64: \
164      return nir_search_op_##op;
165
166#define MATCH_ICONV_CASE(op) \
167   case nir_op_##op##8: \
168   case nir_op_##op##16: \
169   case nir_op_##op##32: \
170   case nir_op_##op##64: \
171      return nir_search_op_##op;
172
173#define MATCH_BCONV_CASE(op) \
174   case nir_op_##op##1: \
175   case nir_op_##op##32: \
176      return nir_search_op_##op;
177
178
179   switch (nop) {
180   MATCH_FCONV_CASE(i2f)
181   MATCH_FCONV_CASE(u2f)
182   MATCH_FCONV_CASE(f2f)
183   MATCH_ICONV_CASE(f2u)
184   MATCH_ICONV_CASE(f2i)
185   MATCH_ICONV_CASE(u2u)
186   MATCH_ICONV_CASE(i2i)
187   MATCH_FCONV_CASE(b2f)
188   MATCH_ICONV_CASE(b2i)
189   MATCH_BCONV_CASE(i2b)
190   MATCH_BCONV_CASE(f2b)
191   default:
192      return nop;
193   }
194
195#undef MATCH_FCONV_CASE
196#undef MATCH_ICONV_CASE
197#undef MATCH_BCONV_CASE
198}
199
200static nir_op
201nir_op_for_search_op(uint16_t sop, unsigned bit_size)
202{
203   if (sop <= nir_last_opcode)
204      return sop;
205
206#define RET_FCONV_CASE(op) \
207   case nir_search_op_##op: \
208      switch (bit_size) { \
209      case 16: return nir_op_##op##16; \
210      case 32: return nir_op_##op##32; \
211      case 64: return nir_op_##op##64; \
212      default: unreachable("Invalid bit size"); \
213      }
214
215#define RET_ICONV_CASE(op) \
216   case nir_search_op_##op: \
217      switch (bit_size) { \
218      case 8:  return nir_op_##op##8; \
219      case 16: return nir_op_##op##16; \
220      case 32: return nir_op_##op##32; \
221      case 64: return nir_op_##op##64; \
222      default: unreachable("Invalid bit size"); \
223      }
224
225#define RET_BCONV_CASE(op) \
226   case nir_search_op_##op: \
227      switch (bit_size) { \
228      case 1: return nir_op_##op##1; \
229      case 32: return nir_op_##op##32; \
230      default: unreachable("Invalid bit size"); \
231      }
232
233   switch (sop) {
234   RET_FCONV_CASE(i2f)
235   RET_FCONV_CASE(u2f)
236   RET_FCONV_CASE(f2f)
237   RET_ICONV_CASE(f2u)
238   RET_ICONV_CASE(f2i)
239   RET_ICONV_CASE(u2u)
240   RET_ICONV_CASE(i2i)
241   RET_FCONV_CASE(b2f)
242   RET_ICONV_CASE(b2i)
243   RET_BCONV_CASE(i2b)
244   RET_BCONV_CASE(f2b)
245   default:
246      unreachable("Invalid nir_search_op");
247   }
248
249#undef RET_FCONV_CASE
250#undef RET_ICONV_CASE
251#undef RET_BCONV_CASE
252}
253
254static bool
255match_value(const nir_search_value *value, nir_alu_instr *instr, unsigned src,
256            unsigned num_components, const uint8_t *swizzle,
257            struct match_state *state)
258{
259   uint8_t new_swizzle[NIR_MAX_VEC_COMPONENTS];
260
261   /* Searching only works on SSA values because, if it's not SSA, we can't
262    * know if the value changed between one instance of that value in the
263    * expression and another.  Also, the replace operation will place reads of
264    * that value right before the last instruction in the expression we're
265    * replacing so those reads will happen after the original reads and may
266    * not be valid if they're register reads.
267    */
268   assert(instr->src[src].src.is_ssa);
269
270   /* If the source is an explicitly sized source, then we need to reset
271    * both the number of components and the swizzle.
272    */
273   if (nir_op_infos[instr->op].input_sizes[src] != 0) {
274      num_components = nir_op_infos[instr->op].input_sizes[src];
275      swizzle = identity_swizzle;
276   }
277
278   for (unsigned i = 0; i < num_components; ++i)
279      new_swizzle[i] = instr->src[src].swizzle[swizzle[i]];
280
281   /* If the value has a specific bit size and it doesn't match, bail */
282   if (value->bit_size > 0 &&
283       nir_src_bit_size(instr->src[src].src) != value->bit_size)
284      return false;
285
286   switch (value->type) {
287   case nir_search_value_expression:
288      if (instr->src[src].src.ssa->parent_instr->type != nir_instr_type_alu)
289         return false;
290
291      return match_expression(nir_search_value_as_expression(value),
292                              nir_instr_as_alu(instr->src[src].src.ssa->parent_instr),
293                              num_components, new_swizzle, state);
294
295   case nir_search_value_variable: {
296      nir_search_variable *var = nir_search_value_as_variable(value);
297      assert(var->variable < NIR_SEARCH_MAX_VARIABLES);
298
299      if (state->variables_seen & (1 << var->variable)) {
300         if (state->variables[var->variable].src.ssa != instr->src[src].src.ssa)
301            return false;
302
303         assert(!instr->src[src].abs && !instr->src[src].negate);
304
305         for (unsigned i = 0; i < num_components; ++i) {
306            if (state->variables[var->variable].swizzle[i] != new_swizzle[i])
307               return false;
308         }
309
310         return true;
311      } else {
312         if (var->is_constant &&
313             instr->src[src].src.ssa->parent_instr->type != nir_instr_type_load_const)
314            return false;
315
316         if (var->cond && !var->cond(state->range_ht, instr,
317                                     src, num_components, new_swizzle))
318            return false;
319
320         if (var->type != nir_type_invalid &&
321             !src_is_type(instr->src[src].src, var->type))
322            return false;
323
324         state->variables_seen |= (1 << var->variable);
325         state->variables[var->variable].src = instr->src[src].src;
326         state->variables[var->variable].abs = false;
327         state->variables[var->variable].negate = false;
328
329         for (unsigned i = 0; i < NIR_MAX_VEC_COMPONENTS; ++i) {
330            if (i < num_components)
331               state->variables[var->variable].swizzle[i] = new_swizzle[i];
332            else
333               state->variables[var->variable].swizzle[i] = 0;
334         }
335
336         return true;
337      }
338   }
339
340   case nir_search_value_constant: {
341      nir_search_constant *const_val = nir_search_value_as_constant(value);
342
343      if (!nir_src_is_const(instr->src[src].src))
344         return false;
345
346      switch (const_val->type) {
347      case nir_type_float: {
348         nir_load_const_instr *const load =
349            nir_instr_as_load_const(instr->src[src].src.ssa->parent_instr);
350
351         /* There are 8-bit and 1-bit integer types, but there are no 8-bit or
352          * 1-bit float types.  This prevents potential assertion failures in
353          * nir_src_comp_as_float.
354          */
355         if (load->def.bit_size < 16)
356            return false;
357
358         for (unsigned i = 0; i < num_components; ++i) {
359            double val = nir_src_comp_as_float(instr->src[src].src,
360                                               new_swizzle[i]);
361            if (val != const_val->data.d)
362               return false;
363         }
364         return true;
365      }
366
367      case nir_type_int:
368      case nir_type_uint:
369      case nir_type_bool: {
370         unsigned bit_size = nir_src_bit_size(instr->src[src].src);
371         uint64_t mask = u_uintN_max(bit_size);
372         for (unsigned i = 0; i < num_components; ++i) {
373            uint64_t val = nir_src_comp_as_uint(instr->src[src].src,
374                                                new_swizzle[i]);
375            if ((val & mask) != (const_val->data.u & mask))
376               return false;
377         }
378         return true;
379      }
380
381      default:
382         unreachable("Invalid alu source type");
383      }
384   }
385
386   default:
387      unreachable("Invalid search value type");
388   }
389}
390
391static bool
392match_expression(const nir_search_expression *expr, nir_alu_instr *instr,
393                 unsigned num_components, const uint8_t *swizzle,
394                 struct match_state *state)
395{
396   if (expr->cond && !expr->cond(instr))
397      return false;
398
399   if (!nir_op_matches_search_op(instr->op, expr->opcode))
400      return false;
401
402   assert(instr->dest.dest.is_ssa);
403
404   if (expr->value.bit_size > 0 &&
405       instr->dest.dest.ssa.bit_size != expr->value.bit_size)
406      return false;
407
408   state->inexact_match = expr->inexact || state->inexact_match;
409   state->has_exact_alu = instr->exact || state->has_exact_alu;
410   if (state->inexact_match && state->has_exact_alu)
411      return false;
412
413   assert(!instr->dest.saturate);
414   assert(nir_op_infos[instr->op].num_inputs > 0);
415
416   /* If we have an explicitly sized destination, we can only handle the
417    * identity swizzle.  While dot(vec3(a, b, c).zxy) is a valid
418    * expression, we don't have the information right now to propagate that
419    * swizzle through.  We can only properly propagate swizzles if the
420    * instruction is vectorized.
421    */
422   if (nir_op_infos[instr->op].output_size != 0) {
423      for (unsigned i = 0; i < num_components; i++) {
424         if (swizzle[i] != i)
425            return false;
426      }
427   }
428
429   /* If this is a commutative expression and it's one of the first few, look
430    * up its direction for the current search operation.  We'll use that value
431    * to possibly flip the sources for the match.
432    */
433   unsigned comm_op_flip =
434      (expr->comm_expr_idx >= 0 &&
435       expr->comm_expr_idx < NIR_SEARCH_MAX_COMM_OPS) ?
436      ((state->comm_op_direction >> expr->comm_expr_idx) & 1) : 0;
437
438   bool matched = true;
439   for (unsigned i = 0; i < nir_op_infos[instr->op].num_inputs; i++) {
440      /* 2src_commutative instructions that have 3 sources are only commutative
441       * in the first two sources.  Source 2 is always source 2.
442       */
443      if (!match_value(expr->srcs[i], instr,
444                       i < 2 ? i ^ comm_op_flip : i,
445                       num_components, swizzle, state)) {
446         matched = false;
447         break;
448      }
449   }
450
451   return matched;
452}
453
454static unsigned
455replace_bitsize(const nir_search_value *value, unsigned search_bitsize,
456                struct match_state *state)
457{
458   if (value->bit_size > 0)
459      return value->bit_size;
460   if (value->bit_size < 0)
461      return nir_src_bit_size(state->variables[-value->bit_size - 1].src);
462   return search_bitsize;
463}
464
465static nir_alu_src
466construct_value(nir_builder *build,
467                const nir_search_value *value,
468                unsigned num_components, unsigned search_bitsize,
469                struct match_state *state,
470                nir_instr *instr)
471{
472   switch (value->type) {
473   case nir_search_value_expression: {
474      const nir_search_expression *expr = nir_search_value_as_expression(value);
475      unsigned dst_bit_size = replace_bitsize(value, search_bitsize, state);
476      nir_op op = nir_op_for_search_op(expr->opcode, dst_bit_size);
477
478      if (nir_op_infos[op].output_size != 0)
479         num_components = nir_op_infos[op].output_size;
480
481      nir_alu_instr *alu = nir_alu_instr_create(build->shader, op);
482      nir_ssa_dest_init(&alu->instr, &alu->dest.dest, num_components,
483                        dst_bit_size, NULL);
484      alu->dest.write_mask = (1 << num_components) - 1;
485      alu->dest.saturate = false;
486
487      /* We have no way of knowing what values in a given search expression
488       * map to a particular replacement value.  Therefore, if the
489       * expression we are replacing has any exact values, the entire
490       * replacement should be exact.
491       */
492      alu->exact = state->has_exact_alu || expr->exact;
493
494      for (unsigned i = 0; i < nir_op_infos[op].num_inputs; i++) {
495         /* If the source is an explicitly sized source, then we need to reset
496          * the number of components to match.
497          */
498         if (nir_op_infos[alu->op].input_sizes[i] != 0)
499            num_components = nir_op_infos[alu->op].input_sizes[i];
500
501         alu->src[i] = construct_value(build, expr->srcs[i],
502                                       num_components, search_bitsize,
503                                       state, instr);
504      }
505
506      nir_builder_instr_insert(build, &alu->instr);
507
508      assert(alu->dest.dest.ssa.index ==
509             util_dynarray_num_elements(state->states, uint16_t));
510      util_dynarray_append(state->states, uint16_t, 0);
511      nir_algebraic_automaton(&alu->instr, state->states, state->pass_op_table);
512
513      nir_alu_src val;
514      val.src = nir_src_for_ssa(&alu->dest.dest.ssa);
515      val.negate = false;
516      val.abs = false,
517      memcpy(val.swizzle, identity_swizzle, sizeof val.swizzle);
518
519      return val;
520   }
521
522   case nir_search_value_variable: {
523      const nir_search_variable *var = nir_search_value_as_variable(value);
524      assert(state->variables_seen & (1 << var->variable));
525
526      nir_alu_src val = { NIR_SRC_INIT };
527      nir_alu_src_copy(&val, &state->variables[var->variable]);
528      assert(!var->is_constant);
529
530      for (unsigned i = 0; i < NIR_MAX_VEC_COMPONENTS; i++)
531         val.swizzle[i] = state->variables[var->variable].swizzle[var->swizzle[i]];
532
533      return val;
534   }
535
536   case nir_search_value_constant: {
537      const nir_search_constant *c = nir_search_value_as_constant(value);
538      unsigned bit_size = replace_bitsize(value, search_bitsize, state);
539
540      nir_ssa_def *cval;
541      switch (c->type) {
542      case nir_type_float:
543         cval = nir_imm_floatN_t(build, c->data.d, bit_size);
544         break;
545
546      case nir_type_int:
547      case nir_type_uint:
548         cval = nir_imm_intN_t(build, c->data.i, bit_size);
549         break;
550
551      case nir_type_bool:
552         cval = nir_imm_boolN_t(build, c->data.u, bit_size);
553         break;
554
555      default:
556         unreachable("Invalid alu source type");
557      }
558
559      assert(cval->index ==
560             util_dynarray_num_elements(state->states, uint16_t));
561      util_dynarray_append(state->states, uint16_t, 0);
562      nir_algebraic_automaton(cval->parent_instr, state->states,
563                              state->pass_op_table);
564
565      nir_alu_src val;
566      val.src = nir_src_for_ssa(cval);
567      val.negate = false;
568      val.abs = false,
569      memset(val.swizzle, 0, sizeof val.swizzle);
570
571      return val;
572   }
573
574   default:
575      unreachable("Invalid search value type");
576   }
577}
578
579UNUSED static void dump_value(const nir_search_value *val)
580{
581   switch (val->type) {
582   case nir_search_value_constant: {
583      const nir_search_constant *sconst = nir_search_value_as_constant(val);
584      switch (sconst->type) {
585      case nir_type_float:
586         fprintf(stderr, "%f", sconst->data.d);
587         break;
588      case nir_type_int:
589         fprintf(stderr, "%"PRId64, sconst->data.i);
590         break;
591      case nir_type_uint:
592         fprintf(stderr, "0x%"PRIx64, sconst->data.u);
593         break;
594      case nir_type_bool:
595         fprintf(stderr, "%s", sconst->data.u != 0 ? "True" : "False");
596         break;
597      default:
598         unreachable("bad const type");
599      }
600      break;
601   }
602
603   case nir_search_value_variable: {
604      const nir_search_variable *var = nir_search_value_as_variable(val);
605      if (var->is_constant)
606         fprintf(stderr, "#");
607      fprintf(stderr, "%c", var->variable + 'a');
608      break;
609   }
610
611   case nir_search_value_expression: {
612      const nir_search_expression *expr = nir_search_value_as_expression(val);
613      fprintf(stderr, "(");
614      if (expr->inexact)
615         fprintf(stderr, "~");
616      switch (expr->opcode) {
617#define CASE(n) \
618      case nir_search_op_##n: fprintf(stderr, #n); break;
619      CASE(f2b)
620      CASE(b2f)
621      CASE(b2i)
622      CASE(i2b)
623      CASE(i2i)
624      CASE(f2i)
625      CASE(i2f)
626#undef CASE
627      default:
628         fprintf(stderr, "%s", nir_op_infos[expr->opcode].name);
629      }
630
631      unsigned num_srcs = 1;
632      if (expr->opcode <= nir_last_opcode)
633         num_srcs = nir_op_infos[expr->opcode].num_inputs;
634
635      for (unsigned i = 0; i < num_srcs; i++) {
636         fprintf(stderr, " ");
637         dump_value(expr->srcs[i]);
638      }
639
640      fprintf(stderr, ")");
641      break;
642   }
643   }
644
645   if (val->bit_size > 0)
646      fprintf(stderr, "@%d", val->bit_size);
647}
648
649static void
650add_uses_to_worklist(nir_instr *instr,
651                     nir_instr_worklist *worklist,
652                     struct util_dynarray *states,
653                     const struct per_op_table *pass_op_table)
654{
655   nir_ssa_def *def = nir_instr_ssa_def(instr);
656
657   nir_foreach_use_safe(use_src, def) {
658      if (nir_algebraic_automaton(use_src->parent_instr, states, pass_op_table))
659         nir_instr_worklist_push_tail(worklist, use_src->parent_instr);
660   }
661}
662
663static void
664nir_algebraic_update_automaton(nir_instr *new_instr,
665                               nir_instr_worklist *algebraic_worklist,
666                               struct util_dynarray *states,
667                               const struct per_op_table *pass_op_table)
668{
669
670   nir_instr_worklist *automaton_worklist = nir_instr_worklist_create();
671
672   /* Walk through the tree of uses of our new instruction's SSA value,
673    * recursively updating the automaton state until it stabilizes.
674    */
675   add_uses_to_worklist(new_instr, automaton_worklist, states, pass_op_table);
676
677   nir_instr *instr;
678   while ((instr = nir_instr_worklist_pop_head(automaton_worklist))) {
679      nir_instr_worklist_push_tail(algebraic_worklist, instr);
680      add_uses_to_worklist(instr, automaton_worklist, states, pass_op_table);
681   }
682
683   nir_instr_worklist_destroy(automaton_worklist);
684}
685
686nir_ssa_def *
687nir_replace_instr(nir_builder *build, nir_alu_instr *instr,
688                  struct hash_table *range_ht,
689                  struct util_dynarray *states,
690                  const struct per_op_table *pass_op_table,
691                  const nir_search_expression *search,
692                  const nir_search_value *replace,
693                  nir_instr_worklist *algebraic_worklist)
694{
695   uint8_t swizzle[NIR_MAX_VEC_COMPONENTS] = { 0 };
696
697   for (unsigned i = 0; i < instr->dest.dest.ssa.num_components; ++i)
698      swizzle[i] = i;
699
700   assert(instr->dest.dest.is_ssa);
701
702   struct match_state state;
703   state.inexact_match = false;
704   state.has_exact_alu = false;
705   state.range_ht = range_ht;
706   state.pass_op_table = pass_op_table;
707
708   STATIC_ASSERT(sizeof(state.comm_op_direction) * 8 >= NIR_SEARCH_MAX_COMM_OPS);
709
710   unsigned comm_expr_combinations =
711      1 << MIN2(search->comm_exprs, NIR_SEARCH_MAX_COMM_OPS);
712
713   bool found = false;
714   for (unsigned comb = 0; comb < comm_expr_combinations; comb++) {
715      /* The bitfield of directions is just the current iteration.  Hooray for
716       * binary.
717       */
718      state.comm_op_direction = comb;
719      state.variables_seen = 0;
720
721      if (match_expression(search, instr,
722                           instr->dest.dest.ssa.num_components,
723                           swizzle, &state)) {
724         found = true;
725         break;
726      }
727   }
728   if (!found)
729      return NULL;
730
731#if 0
732   fprintf(stderr, "matched: ");
733   dump_value(&search->value);
734   fprintf(stderr, " -> ");
735   dump_value(replace);
736   fprintf(stderr, " ssa_%d\n", instr->dest.dest.ssa.index);
737#endif
738
739   /* If the instruction at the root of the expression tree being replaced is
740    * a unary operation, insert the replacement instructions at the location
741    * of the source of the unary operation.  Otherwise, insert the replacement
742    * instructions at the location of the expression tree root.
743    *
744    * For the unary operation case, this is done to prevent some spurious code
745    * motion that can dramatically extend live ranges.  Imagine an expression
746    * like -(A+B) where the addtion and the negation are separated by flow
747    * control and thousands of instructions.  If this expression is replaced
748    * with -A+-B, inserting the new instructions at the site of the negation
749    * could extend the live range of A and B dramtically.  This could increase
750    * register pressure and cause spilling.
751    *
752    * It may well be that moving instructions around is a good thing, but
753    * keeping algebraic optimizations and code motion optimizations separate
754    * seems safest.
755    */
756   nir_alu_instr *const src_instr = nir_src_as_alu_instr(instr->src[0].src);
757   if (src_instr != NULL &&
758       (instr->op == nir_op_fneg || instr->op == nir_op_fabs ||
759        instr->op == nir_op_ineg || instr->op == nir_op_iabs ||
760        instr->op == nir_op_inot)) {
761      /* Insert new instructions *after*.  Otherwise a hypothetical
762       * replacement fneg(X) -> fabs(X) would insert the fabs() instruction
763       * before X!  This can also occur for things like fneg(X.wzyx) -> X.wzyx
764       * in vector mode.  A move instruction to handle the swizzle will get
765       * inserted before X.
766       *
767       * This manifested in a single OpenGL ES 2.0 CTS vertex shader test on
768       * older Intel GPU that use vector-mode vertex processing.
769       */
770      build->cursor = nir_after_instr(&src_instr->instr);
771   } else {
772      build->cursor = nir_before_instr(&instr->instr);
773   }
774
775   state.states = states;
776
777   nir_alu_src val = construct_value(build, replace,
778                                     instr->dest.dest.ssa.num_components,
779                                     instr->dest.dest.ssa.bit_size,
780                                     &state, &instr->instr);
781
782   /* Note that NIR builder will elide the MOV if it's a no-op, which may
783    * allow more work to be done in a single pass through algebraic.
784    */
785   nir_ssa_def *ssa_val =
786      nir_mov_alu(build, val, instr->dest.dest.ssa.num_components);
787   if (ssa_val->index == util_dynarray_num_elements(states, uint16_t)) {
788      util_dynarray_append(states, uint16_t, 0);
789      nir_algebraic_automaton(ssa_val->parent_instr, states, pass_op_table);
790   }
791
792   /* Rewrite the uses of the old SSA value to the new one, and recurse
793    * through the uses updating the automaton's state.
794    */
795   nir_ssa_def_rewrite_uses(&instr->dest.dest.ssa, ssa_val);
796   nir_algebraic_update_automaton(ssa_val->parent_instr, algebraic_worklist,
797                                  states, pass_op_table);
798
799   /* Nothing uses the instr any more, so drop it out of the program.  Note
800    * that the instr may be in the worklist still, so we can't free it
801    * directly.
802    */
803   nir_instr_remove(&instr->instr);
804
805   return ssa_val;
806}
807
808static bool
809nir_algebraic_automaton(nir_instr *instr, struct util_dynarray *states,
810                        const struct per_op_table *pass_op_table)
811{
812   switch (instr->type) {
813   case nir_instr_type_alu: {
814      nir_alu_instr *alu = nir_instr_as_alu(instr);
815      nir_op op = alu->op;
816      uint16_t search_op = nir_search_op_for_nir_op(op);
817      const struct per_op_table *tbl = &pass_op_table[search_op];
818      if (tbl->num_filtered_states == 0)
819         return false;
820
821      /* Calculate the index into the transition table. Note the index
822       * calculated must match the iteration order of Python's
823       * itertools.product(), which was used to emit the transition
824       * table.
825       */
826      unsigned index = 0;
827      for (unsigned i = 0; i < nir_op_infos[op].num_inputs; i++) {
828         index *= tbl->num_filtered_states;
829         index += tbl->filter[*util_dynarray_element(states, uint16_t,
830                                                     alu->src[i].src.ssa->index)];
831      }
832
833      uint16_t *state = util_dynarray_element(states, uint16_t,
834                                              alu->dest.dest.ssa.index);
835      if (*state != tbl->table[index]) {
836         *state = tbl->table[index];
837         return true;
838      }
839      return false;
840   }
841
842   case nir_instr_type_load_const: {
843      nir_load_const_instr *load_const = nir_instr_as_load_const(instr);
844      uint16_t *state = util_dynarray_element(states, uint16_t,
845                                              load_const->def.index);
846      if (*state != CONST_STATE) {
847         *state = CONST_STATE;
848         return true;
849      }
850      return false;
851   }
852
853   default:
854      return false;
855   }
856}
857
858static bool
859nir_algebraic_instr(nir_builder *build, nir_instr *instr,
860                    struct hash_table *range_ht,
861                    const bool *condition_flags,
862                    const struct transform **transforms,
863                    const uint16_t *transform_counts,
864                    struct util_dynarray *states,
865                    const struct per_op_table *pass_op_table,
866                    nir_instr_worklist *worklist)
867{
868
869   if (instr->type != nir_instr_type_alu)
870      return false;
871
872   nir_alu_instr *alu = nir_instr_as_alu(instr);
873   if (!alu->dest.dest.is_ssa)
874      return false;
875
876   unsigned bit_size = alu->dest.dest.ssa.bit_size;
877   const unsigned execution_mode =
878      build->shader->info.float_controls_execution_mode;
879   const bool ignore_inexact =
880      nir_is_float_control_signed_zero_inf_nan_preserve(execution_mode, bit_size) ||
881      nir_is_denorm_flush_to_zero(execution_mode, bit_size);
882
883   int xform_idx = *util_dynarray_element(states, uint16_t,
884                                          alu->dest.dest.ssa.index);
885   for (uint16_t i = 0; i < transform_counts[xform_idx]; i++) {
886      const struct transform *xform = &transforms[xform_idx][i];
887      if (condition_flags[xform->condition_offset] &&
888          !(xform->search->inexact && ignore_inexact) &&
889          nir_replace_instr(build, alu, range_ht, states, pass_op_table,
890                            xform->search, xform->replace, worklist)) {
891         _mesa_hash_table_clear(range_ht, NULL);
892         return true;
893      }
894   }
895
896   return false;
897}
898
899bool
900nir_algebraic_impl(nir_function_impl *impl,
901                   const bool *condition_flags,
902                   const struct transform **transforms,
903                   const uint16_t *transform_counts,
904                   const struct per_op_table *pass_op_table)
905{
906   bool progress = false;
907
908   nir_builder build;
909   nir_builder_init(&build, impl);
910
911   /* Note: it's important here that we're allocating a zeroed array, since
912    * state 0 is the default state, which means we don't have to visit
913    * anything other than constants and ALU instructions.
914    */
915   struct util_dynarray states = {0};
916   if (!util_dynarray_resize(&states, uint16_t, impl->ssa_alloc)) {
917      nir_metadata_preserve(impl, nir_metadata_all);
918      return false;
919   }
920   memset(states.data, 0, states.size);
921
922   struct hash_table *range_ht = _mesa_pointer_hash_table_create(NULL);
923
924   nir_instr_worklist *worklist = nir_instr_worklist_create();
925
926   /* Walk top-to-bottom setting up the automaton state. */
927   nir_foreach_block(block, impl) {
928      nir_foreach_instr(instr, block) {
929         nir_algebraic_automaton(instr, &states, pass_op_table);
930      }
931   }
932
933   /* Put our instrs in the worklist such that we're popping the last instr
934    * first.  This will encourage us to match the biggest source patterns when
935    * possible.
936    */
937   nir_foreach_block_reverse(block, impl) {
938      nir_foreach_instr_reverse(instr, block) {
939         if (instr->type == nir_instr_type_alu)
940            nir_instr_worklist_push_tail(worklist, instr);
941      }
942   }
943
944   nir_instr *instr;
945   while ((instr = nir_instr_worklist_pop_head(worklist))) {
946      /* The worklist can have an instr pushed to it multiple times if it was
947       * the src of multiple instrs that also got optimized, so make sure that
948       * we don't try to re-optimize an instr we already handled.
949       */
950      if (exec_node_is_tail_sentinel(&instr->node))
951         continue;
952
953      progress |= nir_algebraic_instr(&build, instr,
954                                      range_ht, condition_flags,
955                                      transforms, transform_counts, &states,
956                                      pass_op_table, worklist);
957   }
958
959   nir_instr_worklist_destroy(worklist);
960   ralloc_free(range_ht);
961   util_dynarray_fini(&states);
962
963   if (progress) {
964      nir_metadata_preserve(impl, nir_metadata_block_index |
965                                  nir_metadata_dominance);
966   } else {
967      nir_metadata_preserve(impl, nir_metadata_all);
968   }
969
970   return progress;
971}
972