1 /*
2 * Copyright © 2015 Thomas Helland
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 #include "nir_loop_analyze.h"
25 #include "util/bitset.h"
26 #include "nir.h"
27 #include "nir_constant_expressions.h"
28
29 typedef enum {
30 undefined,
31 basic_induction
32 } nir_loop_variable_type;
33
34 typedef struct {
35 /* A link for the work list */
36 struct list_head process_link;
37
38 bool in_loop;
39
40 /* The ssa_def associated with this info */
41 nir_def *def;
42
43 /* The type of this ssa_def */
44 nir_loop_variable_type type;
45
46 /* True if variable is in an if branch */
47 bool in_if_branch;
48
49 /* True if variable is in a nested loop */
50 bool in_nested_loop;
51
52 /* Could be a basic_induction if following uniforms are inlined */
53 nir_src *init_src;
54 nir_alu_src *update_src;
55
56 /**
57 * SSA def of the phi-node associated with this induction variable.
58 *
59 * Every loop induction variable has an associated phi node in the loop
60 * header. This may point to the same SSA def as \c def. If, however, \c def
61 * is the increment of the induction variable, this will point to the SSA
62 * def being incremented.
63 */
64 nir_def *basis;
65 } nir_loop_variable;
66
67 typedef struct {
68 /* The loop we store information for */
69 nir_loop *loop;
70
71 /* Loop_variable for all ssa_defs in function */
72 nir_loop_variable *loop_vars;
73 BITSET_WORD *loop_vars_init;
74
75 /* A list of the loop_vars to analyze */
76 struct list_head process_list;
77
78 nir_variable_mode indirect_mask;
79
80 bool force_unroll_sampler_indirect;
81 } loop_info_state;
82
83 static nir_loop_variable *
get_loop_var(nir_def * value,loop_info_state * state)84 get_loop_var(nir_def *value, loop_info_state *state)
85 {
86 nir_loop_variable *var = &(state->loop_vars[value->index]);
87
88 if (!BITSET_TEST(state->loop_vars_init, value->index)) {
89 var->in_loop = false;
90 var->def = value;
91 var->in_if_branch = false;
92 var->in_nested_loop = false;
93 var->init_src = NULL;
94 var->update_src = NULL;
95 var->type = undefined;
96
97 BITSET_SET(state->loop_vars_init, value->index);
98 }
99
100 return var;
101 }
102
103 typedef struct {
104 loop_info_state *state;
105 bool in_if_branch;
106 bool in_nested_loop;
107 } init_loop_state;
108
109 static bool
init_loop_def(nir_def * def,void * void_init_loop_state)110 init_loop_def(nir_def *def, void *void_init_loop_state)
111 {
112 init_loop_state *loop_init_state = void_init_loop_state;
113 nir_loop_variable *var = get_loop_var(def, loop_init_state->state);
114
115 if (loop_init_state->in_nested_loop) {
116 var->in_nested_loop = true;
117 } else if (loop_init_state->in_if_branch) {
118 var->in_if_branch = true;
119 } else {
120 /* Add to the tail of the list. That way we start at the beginning of
121 * the defs in the loop instead of the end when walking the list. This
122 * means less recursive calls. Only add defs that are not in nested
123 * loops or conditional blocks.
124 */
125 list_addtail(&var->process_link, &loop_init_state->state->process_list);
126 }
127
128 var->in_loop = true;
129
130 return true;
131 }
132
133 /** Calculate an estimated cost in number of instructions
134 *
135 * We do this so that we don't unroll loops which will later get massively
136 * inflated due to int64 or fp64 lowering. The estimates provided here don't
137 * have to be massively accurate; they just have to be good enough that loop
138 * unrolling doesn't cause things to blow up too much.
139 */
140 static unsigned
instr_cost(loop_info_state * state,nir_instr * instr,const nir_shader_compiler_options * options)141 instr_cost(loop_info_state *state, nir_instr *instr,
142 const nir_shader_compiler_options *options)
143 {
144 if (instr->type == nir_instr_type_intrinsic ||
145 instr->type == nir_instr_type_tex)
146 return 1;
147
148 if (instr->type != nir_instr_type_alu)
149 return 0;
150
151 nir_alu_instr *alu = nir_instr_as_alu(instr);
152 const nir_op_info *info = &nir_op_infos[alu->op];
153 unsigned cost = 1;
154
155 if (nir_op_is_selection(alu->op)) {
156 nir_scalar cond_scalar = { alu->src[0].src.ssa, 0 };
157 if (nir_is_terminator_condition_with_two_inputs(cond_scalar)) {
158 nir_instr *sel_cond = alu->src[0].src.ssa->parent_instr;
159 nir_alu_instr *sel_alu = nir_instr_as_alu(sel_cond);
160
161 nir_scalar rhs, lhs;
162 lhs = nir_scalar_chase_alu_src(cond_scalar, 0);
163 rhs = nir_scalar_chase_alu_src(cond_scalar, 1);
164
165 /* If the selects condition is a comparision between a constant and
166 * a basic induction variable we know that it will be eliminated once
167 * the loop is unrolled so here we assign it a cost of 0.
168 */
169 if ((nir_src_is_const(sel_alu->src[0].src) &&
170 get_loop_var(rhs.def, state)->type == basic_induction) ||
171 (nir_src_is_const(sel_alu->src[1].src) &&
172 get_loop_var(lhs.def, state)->type == basic_induction)) {
173 /* Also if the selects condition is only used by the select then
174 * remove that alu instructons cost from the cost total also.
175 */
176 if (!list_is_singular(&sel_alu->def.uses) ||
177 nir_def_used_by_if(&sel_alu->def))
178 return 0;
179 else
180 return -1;
181 }
182 }
183 }
184
185 if (alu->op == nir_op_flrp) {
186 if ((options->lower_flrp16 && alu->def.bit_size == 16) ||
187 (options->lower_flrp32 && alu->def.bit_size == 32) ||
188 (options->lower_flrp64 && alu->def.bit_size == 64))
189 cost *= 3;
190 }
191
192 /* Assume everything 16 or 32-bit is cheap.
193 *
194 * There are no 64-bit ops that don't have a 64-bit thing as their
195 * destination or first source.
196 */
197 if (alu->def.bit_size < 64 &&
198 nir_src_bit_size(alu->src[0].src) < 64)
199 return cost;
200
201 bool is_fp64 = alu->def.bit_size == 64 &&
202 nir_alu_type_get_base_type(info->output_type) == nir_type_float;
203 for (unsigned i = 0; i < info->num_inputs; i++) {
204 if (nir_src_bit_size(alu->src[i].src) == 64 &&
205 nir_alu_type_get_base_type(info->input_types[i]) == nir_type_float)
206 is_fp64 = true;
207 }
208
209 if (is_fp64) {
210 /* If it's something lowered normally, it's expensive. */
211 if (options->lower_doubles_options &
212 nir_lower_doubles_op_to_options_mask(alu->op))
213 cost *= 20;
214
215 /* If it's full software, it's even more expensive */
216 if (options->lower_doubles_options & nir_lower_fp64_full_software) {
217 cost *= 100;
218 state->loop->info->has_soft_fp64 = true;
219 }
220
221 return cost;
222 } else {
223 if (options->lower_int64_options &
224 nir_lower_int64_op_to_options_mask(alu->op)) {
225 /* These require a doing the division algorithm. */
226 if (alu->op == nir_op_idiv || alu->op == nir_op_udiv ||
227 alu->op == nir_op_imod || alu->op == nir_op_umod ||
228 alu->op == nir_op_irem)
229 return cost * 100;
230
231 /* Other int64 lowering isn't usually all that expensive */
232 return cost * 5;
233 }
234
235 return cost;
236 }
237 }
238
239 static bool
init_loop_block(nir_block * block,loop_info_state * state,bool in_if_branch,bool in_nested_loop)240 init_loop_block(nir_block *block, loop_info_state *state,
241 bool in_if_branch, bool in_nested_loop)
242 {
243 init_loop_state init_state = { .in_if_branch = in_if_branch,
244 .in_nested_loop = in_nested_loop,
245 .state = state };
246
247 nir_foreach_instr(instr, block) {
248 nir_foreach_def(instr, init_loop_def, &init_state);
249 }
250
251 return true;
252 }
253
254 static inline bool
is_var_alu(nir_loop_variable * var)255 is_var_alu(nir_loop_variable *var)
256 {
257 return var->def->parent_instr->type == nir_instr_type_alu;
258 }
259
260 static inline bool
is_var_phi(nir_loop_variable * var)261 is_var_phi(nir_loop_variable *var)
262 {
263 return var->def->parent_instr->type == nir_instr_type_phi;
264 }
265
266 /* If all of the instruction sources point to identical ALU instructions (as
267 * per nir_instrs_equal), return one of the ALU instructions. Otherwise,
268 * return NULL.
269 */
270 static nir_alu_instr *
phi_instr_as_alu(nir_phi_instr * phi)271 phi_instr_as_alu(nir_phi_instr *phi)
272 {
273 nir_alu_instr *first = NULL;
274 nir_foreach_phi_src(src, phi) {
275 if (src->src.ssa->parent_instr->type != nir_instr_type_alu)
276 return NULL;
277
278 nir_alu_instr *alu = nir_instr_as_alu(src->src.ssa->parent_instr);
279 if (first == NULL) {
280 first = alu;
281 } else {
282 if (!nir_instrs_equal(&first->instr, &alu->instr))
283 return NULL;
284 }
285 }
286
287 return first;
288 }
289
290 static bool
alu_src_has_identity_swizzle(nir_alu_instr * alu,unsigned src_idx)291 alu_src_has_identity_swizzle(nir_alu_instr *alu, unsigned src_idx)
292 {
293 assert(nir_op_infos[alu->op].input_sizes[src_idx] == 0);
294 for (unsigned i = 0; i < alu->def.num_components; i++) {
295 if (alu->src[src_idx].swizzle[i] != i)
296 return false;
297 }
298
299 return true;
300 }
301
302 static bool
is_only_uniform_src(nir_src * src)303 is_only_uniform_src(nir_src *src)
304 {
305 nir_instr *instr = src->ssa->parent_instr;
306
307 switch (instr->type) {
308 case nir_instr_type_alu: {
309 /* Return true if all sources return true. */
310 nir_alu_instr *alu = nir_instr_as_alu(instr);
311 for (unsigned i = 0; i < nir_op_infos[alu->op].num_inputs; i++) {
312 if (!is_only_uniform_src(&alu->src[i].src))
313 return false;
314 }
315 return true;
316 }
317
318 case nir_instr_type_intrinsic: {
319 nir_intrinsic_instr *inst = nir_instr_as_intrinsic(instr);
320 /* current uniform inline only support load ubo */
321 return inst->intrinsic == nir_intrinsic_load_ubo;
322 }
323
324 case nir_instr_type_load_const:
325 /* Always return true for constants. */
326 return true;
327
328 default:
329 return false;
330 }
331 }
332
333 static bool
compute_induction_information(loop_info_state * state)334 compute_induction_information(loop_info_state *state)
335 {
336 unsigned num_induction_vars = 0;
337
338 list_for_each_entry_safe(nir_loop_variable, var, &state->process_list,
339 process_link) {
340
341 /* Things in nested loops or conditionals should not have been added into
342 * the procss_list.
343 */
344 assert(!var->in_if_branch && !var->in_nested_loop);
345
346 /* We are only interested in checking phis for the basic induction
347 * variable case as its simple to detect. All basic induction variables
348 * have a phi node
349 */
350 if (!is_var_phi(var))
351 continue;
352
353 nir_phi_instr *phi = nir_instr_as_phi(var->def->parent_instr);
354
355 nir_loop_variable *alu_src_var = NULL;
356 nir_foreach_phi_src(src, phi) {
357 nir_loop_variable *src_var = get_loop_var(src->src.ssa, state);
358
359 /* If one of the sources is in an if branch or nested loop then don't
360 * attempt to go any further.
361 */
362 if (src_var->in_if_branch || src_var->in_nested_loop)
363 break;
364
365 /* Detect inductions variables that are incremented in both branches
366 * of an unnested if rather than in a loop block.
367 */
368 if (is_var_phi(src_var)) {
369 nir_phi_instr *src_phi =
370 nir_instr_as_phi(src_var->def->parent_instr);
371 nir_alu_instr *src_phi_alu = phi_instr_as_alu(src_phi);
372 if (src_phi_alu) {
373 src_var = get_loop_var(&src_phi_alu->def, state);
374 if (!src_var->in_if_branch)
375 break;
376 }
377 }
378
379 if (!src_var->in_loop && !var->init_src) {
380 var->init_src = &src->src;
381 } else if (is_var_alu(src_var) && !var->update_src) {
382 alu_src_var = src_var;
383 nir_alu_instr *alu = nir_instr_as_alu(src_var->def->parent_instr);
384
385 /* Check for unsupported alu operations */
386 if (alu->op != nir_op_iadd && alu->op != nir_op_fadd &&
387 alu->op != nir_op_imul && alu->op != nir_op_fmul &&
388 alu->op != nir_op_ishl && alu->op != nir_op_ishr &&
389 alu->op != nir_op_ushr)
390 break;
391
392 if (nir_op_infos[alu->op].num_inputs == 2) {
393 for (unsigned i = 0; i < 2; i++) {
394 /* Is one of the operands const or uniform, and the other the phi.
395 * The phi source can't be swizzled in any way.
396 */
397 if (alu->src[1 - i].src.ssa == &phi->def &&
398 alu_src_has_identity_swizzle(alu, 1 - i)) {
399 if (is_only_uniform_src(&alu->src[i].src))
400 var->update_src = alu->src + i;
401 }
402 }
403 }
404
405 if (!var->update_src)
406 break;
407 } else {
408 var->update_src = NULL;
409 break;
410 }
411 }
412
413 if (var->update_src && var->init_src &&
414 is_only_uniform_src(var->init_src)) {
415 alu_src_var->init_src = var->init_src;
416 alu_src_var->update_src = var->update_src;
417 alu_src_var->basis = var->def;
418 alu_src_var->type = basic_induction;
419
420 var->basis = var->def;
421 var->type = basic_induction;
422
423 num_induction_vars += 2;
424 } else {
425 var->init_src = NULL;
426 var->update_src = NULL;
427 var->basis = NULL;
428 }
429 }
430
431 nir_loop_info *info = state->loop->info;
432 ralloc_free(info->induction_vars);
433 info->num_induction_vars = 0;
434
435 /* record induction variables into nir_loop_info */
436 if (num_induction_vars) {
437 info->induction_vars = ralloc_array(info, nir_loop_induction_variable,
438 num_induction_vars);
439
440 list_for_each_entry(nir_loop_variable, var, &state->process_list,
441 process_link) {
442 if (var->type == basic_induction) {
443 nir_loop_induction_variable *ivar =
444 &info->induction_vars[info->num_induction_vars++];
445 ivar->def = var->def;
446 ivar->init_src = var->init_src;
447 ivar->update_src = var->update_src;
448 }
449 }
450 /* don't overflow */
451 assert(info->num_induction_vars <= num_induction_vars);
452 }
453
454 return num_induction_vars != 0;
455 }
456
457 static bool
find_loop_terminators(loop_info_state * state)458 find_loop_terminators(loop_info_state *state)
459 {
460 bool success = false;
461 foreach_list_typed_safe(nir_cf_node, node, node, &state->loop->body) {
462 if (node->type == nir_cf_node_if) {
463 nir_if *nif = nir_cf_node_as_if(node);
464
465 nir_block *break_blk = NULL;
466 nir_block *continue_from_blk = NULL;
467 bool continue_from_then = true;
468
469 nir_block *last_then = nir_if_last_then_block(nif);
470 nir_block *last_else = nir_if_last_else_block(nif);
471 if (nir_block_ends_in_break(last_then)) {
472 break_blk = last_then;
473 continue_from_blk = last_else;
474 continue_from_then = false;
475 } else if (nir_block_ends_in_break(last_else)) {
476 break_blk = last_else;
477 continue_from_blk = last_then;
478 }
479
480 /* If there is a break then we should find a terminator. If we can
481 * not find a loop terminator, but there is a break-statement then
482 * we should return false so that we do not try to find trip-count
483 */
484 if (!nir_is_trivial_loop_if(nif, break_blk)) {
485 state->loop->info->complex_loop = true;
486 return false;
487 }
488
489 /* Continue if the if contained no jumps at all */
490 if (!break_blk)
491 continue;
492
493 if (nif->condition.ssa->parent_instr->type == nir_instr_type_phi) {
494 state->loop->info->complex_loop = true;
495 return false;
496 }
497
498 nir_loop_terminator *terminator =
499 rzalloc(state->loop->info, nir_loop_terminator);
500
501 list_addtail(&terminator->loop_terminator_link,
502 &state->loop->info->loop_terminator_list);
503
504 terminator->nif = nif;
505 terminator->break_block = break_blk;
506 terminator->continue_from_block = continue_from_blk;
507 terminator->continue_from_then = continue_from_then;
508 terminator->conditional_instr = nif->condition.ssa->parent_instr;
509
510 success = true;
511 }
512 }
513
514 return success;
515 }
516
517 /* This function looks for an array access within a loop that uses an
518 * induction variable for the array index. If found it returns the size of the
519 * array, otherwise 0 is returned. If we find an induction var we pass it back
520 * to the caller via array_index_out.
521 */
522 static unsigned
find_array_access_via_induction(loop_info_state * state,nir_deref_instr * deref,nir_loop_variable ** array_index_out)523 find_array_access_via_induction(loop_info_state *state,
524 nir_deref_instr *deref,
525 nir_loop_variable **array_index_out)
526 {
527 for (nir_deref_instr *d = deref; d; d = nir_deref_instr_parent(d)) {
528 if (d->deref_type != nir_deref_type_array)
529 continue;
530
531 nir_loop_variable *array_index = get_loop_var(d->arr.index.ssa, state);
532
533 if (array_index->type != basic_induction)
534 continue;
535
536 if (array_index_out)
537 *array_index_out = array_index;
538
539 nir_deref_instr *parent = nir_deref_instr_parent(d);
540
541 if (glsl_type_is_array_or_matrix(parent->type)) {
542 return glsl_get_length(parent->type);
543 } else {
544 assert(glsl_type_is_vector(parent->type));
545 return glsl_get_vector_elements(parent->type);
546 }
547 }
548
549 return 0;
550 }
551
552 static bool
guess_loop_limit(loop_info_state * state,nir_const_value * limit_val,nir_scalar basic_ind)553 guess_loop_limit(loop_info_state *state, nir_const_value *limit_val,
554 nir_scalar basic_ind)
555 {
556 unsigned min_array_size = 0;
557
558 nir_foreach_block_in_cf_node(block, &state->loop->cf_node) {
559 nir_foreach_instr(instr, block) {
560 if (instr->type != nir_instr_type_intrinsic)
561 continue;
562
563 nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
564
565 /* Check for arrays variably-indexed by a loop induction variable. */
566 if (intrin->intrinsic == nir_intrinsic_load_deref ||
567 intrin->intrinsic == nir_intrinsic_store_deref ||
568 intrin->intrinsic == nir_intrinsic_copy_deref) {
569
570 nir_loop_variable *array_idx = NULL;
571 unsigned array_size =
572 find_array_access_via_induction(state,
573 nir_src_as_deref(intrin->src[0]),
574 &array_idx);
575 if (array_idx && basic_ind.def == array_idx->def &&
576 (min_array_size == 0 || min_array_size > array_size)) {
577 /* Array indices are scalars */
578 assert(basic_ind.def->num_components == 1);
579 min_array_size = array_size;
580 }
581
582 if (intrin->intrinsic != nir_intrinsic_copy_deref)
583 continue;
584
585 array_size =
586 find_array_access_via_induction(state,
587 nir_src_as_deref(intrin->src[1]),
588 &array_idx);
589 if (array_idx && basic_ind.def == array_idx->def &&
590 (min_array_size == 0 || min_array_size > array_size)) {
591 /* Array indices are scalars */
592 assert(basic_ind.def->num_components == 1);
593 min_array_size = array_size;
594 }
595 }
596 }
597 }
598
599 if (min_array_size) {
600 *limit_val = nir_const_value_for_uint(min_array_size,
601 basic_ind.def->bit_size);
602 return true;
603 }
604
605 return false;
606 }
607
608 static nir_op invert_comparison_if_needed(nir_op alu_op, bool invert);
609
610 /* Returns whether "limit_op(a, b) alu_op c" is equivalent to "(a alu_op c) || (b alu_op c)". */
611 static bool
is_minmax_compatible(nir_op limit_op,nir_op alu_op,bool limit_rhs,bool invert_cond)612 is_minmax_compatible(nir_op limit_op, nir_op alu_op, bool limit_rhs, bool invert_cond)
613 {
614 bool is_max;
615 switch (limit_op) {
616 case nir_op_imin:
617 case nir_op_fmin:
618 case nir_op_umin:
619 is_max = false;
620 break;
621 case nir_op_imax:
622 case nir_op_fmax:
623 case nir_op_umax:
624 is_max = true;
625 break;
626 default:
627 return false;
628 }
629
630 if (nir_op_infos[limit_op].input_types[0] != nir_op_infos[alu_op].input_types[0])
631 return false;
632
633 /* Comparisons we can split are:
634 * - min(a, b) < c
635 * - c < max(a, b)
636 * - max(a, b) >= c
637 * - c >= min(a, b)
638 */
639 switch (invert_comparison_if_needed(alu_op, invert_cond)) {
640 case nir_op_ilt:
641 case nir_op_flt:
642 case nir_op_ult:
643 return (!limit_rhs && !is_max) || (limit_rhs && is_max);
644 case nir_op_ige:
645 case nir_op_fge:
646 case nir_op_uge:
647 return (!limit_rhs && is_max) || (limit_rhs && !is_max);
648 default:
649 return false;
650 }
651 }
652
653 static bool
try_find_limit_of_alu(nir_scalar limit,nir_const_value * limit_val,nir_op alu_op,bool invert_cond,nir_loop_terminator * terminator,loop_info_state * state)654 try_find_limit_of_alu(nir_scalar limit, nir_const_value *limit_val, nir_op alu_op,
655 bool invert_cond, nir_loop_terminator *terminator,
656 loop_info_state *state)
657 {
658 if (!nir_scalar_is_alu(limit))
659 return false;
660
661 nir_op limit_op = nir_scalar_alu_op(limit);
662 if (is_minmax_compatible(limit_op, alu_op, !terminator->induction_rhs, invert_cond)) {
663 for (unsigned i = 0; i < 2; i++) {
664 nir_scalar src = nir_scalar_chase_alu_src(limit, i);
665 if (nir_scalar_is_const(src)) {
666 *limit_val = nir_scalar_as_const_value(src);
667 terminator->exact_trip_count_unknown = true;
668 return true;
669 }
670 }
671 }
672
673 return false;
674 }
675
676 static nir_const_value
eval_const_unop(nir_op op,unsigned bit_size,nir_const_value src0,unsigned execution_mode)677 eval_const_unop(nir_op op, unsigned bit_size, nir_const_value src0,
678 unsigned execution_mode)
679 {
680 assert(nir_op_infos[op].num_inputs == 1);
681 nir_const_value dest;
682 nir_const_value *src[1] = { &src0 };
683 nir_eval_const_opcode(op, &dest, 1, bit_size, src, execution_mode);
684 return dest;
685 }
686
687 static nir_const_value
eval_const_binop(nir_op op,unsigned bit_size,nir_const_value src0,nir_const_value src1,unsigned execution_mode)688 eval_const_binop(nir_op op, unsigned bit_size,
689 nir_const_value src0, nir_const_value src1,
690 unsigned execution_mode)
691 {
692 assert(nir_op_infos[op].num_inputs == 2);
693 nir_const_value dest;
694 nir_const_value *src[2] = { &src0, &src1 };
695 nir_eval_const_opcode(op, &dest, 1, bit_size, src, execution_mode);
696 return dest;
697 }
698
699 static int
find_replacement(const nir_scalar * originals,nir_scalar key,unsigned num_replacements)700 find_replacement(const nir_scalar *originals, nir_scalar key,
701 unsigned num_replacements)
702 {
703 for (int i = 0; i < num_replacements; i++) {
704 if (nir_scalar_equal(originals[i], key))
705 return i;
706 }
707
708 return -1;
709 }
710
711 /**
712 * Try to evaluate an ALU instruction as a constant with a replacement
713 *
714 * Much like \c nir_opt_constant_folding.c:try_fold_alu, this method attempts
715 * to evaluate an ALU instruction as a constant. There are two significant
716 * differences.
717 *
718 * First, this method performs the evaluation recursively. If any source of
719 * the ALU instruction is not itself a constant, it is first evaluated.
720 *
721 * Second, if the SSA value \c original is encountered as a source of the ALU
722 * instruction, the value \c replacement is substituted.
723 *
724 * The intended purpose of this function is to evaluate an arbitrary
725 * expression involving a loop induction variable. In this case, \c original
726 * would be the phi node associated with the induction variable, and
727 * \c replacement is the initial value of the induction variable.
728 *
729 * \returns true if the ALU instruction can be evaluated as constant (after
730 * applying the previously described substitution) or false otherwise.
731 */
732 static bool
try_eval_const_alu(nir_const_value * dest,nir_scalar alu_s,const nir_scalar * originals,const nir_const_value * replacements,unsigned num_replacements,unsigned execution_mode)733 try_eval_const_alu(nir_const_value *dest, nir_scalar alu_s, const nir_scalar *originals,
734 const nir_const_value *replacements,
735 unsigned num_replacements, unsigned execution_mode)
736 {
737 nir_alu_instr *alu = nir_instr_as_alu(alu_s.def->parent_instr);
738
739 if (nir_op_infos[alu->op].output_size)
740 return false;
741
742 /* In the case that any outputs/inputs have unsized types, then we need to
743 * guess the bit-size. In this case, the validator ensures that all
744 * bit-sizes match so we can just take the bit-size from first
745 * output/input with an unsized type. If all the outputs/inputs are sized
746 * then we don't need to guess the bit-size at all because the code we
747 * generate for constant opcodes in this case already knows the sizes of
748 * the types involved and does not need the provided bit-size for anything
749 * (although it still requires to receive a valid bit-size).
750 */
751 unsigned bit_size = 0;
752 if (!nir_alu_type_get_type_size(nir_op_infos[alu->op].output_type)) {
753 bit_size = alu->def.bit_size;
754 } else {
755 for (unsigned i = 0; i < nir_op_infos[alu->op].num_inputs; i++) {
756 if (!nir_alu_type_get_type_size(nir_op_infos[alu->op].input_types[i]))
757 bit_size = alu->src[i].src.ssa->bit_size;
758 }
759
760 if (bit_size == 0)
761 bit_size = 32;
762 }
763
764 nir_const_value src[NIR_MAX_VEC_COMPONENTS];
765 nir_const_value *src_ptrs[NIR_MAX_VEC_COMPONENTS];
766
767 for (unsigned i = 0; i < nir_op_infos[alu->op].num_inputs; i++) {
768 nir_scalar src_s = nir_scalar_chase_alu_src(alu_s, i);
769
770 src_ptrs[i] = &src[i];
771 if (nir_scalar_is_const(src_s)) {
772 src[i] = nir_scalar_as_const_value(src_s);
773 continue;
774 }
775
776 int r = find_replacement(originals, src_s, num_replacements);
777 if (r >= 0) {
778 src[i] = replacements[r];
779 } else if (!nir_scalar_is_alu(src_s) ||
780 !try_eval_const_alu(&src[i], src_s,
781 originals, replacements,
782 num_replacements, execution_mode)) {
783 return false;
784 }
785 }
786
787 nir_eval_const_opcode(alu->op, dest, 1, bit_size, src_ptrs, execution_mode);
788
789 return true;
790 }
791
792 static nir_op
invert_comparison_if_needed(nir_op alu_op,bool invert)793 invert_comparison_if_needed(nir_op alu_op, bool invert)
794 {
795 if (!invert)
796 return alu_op;
797
798 switch (alu_op) {
799 case nir_op_fge:
800 return nir_op_flt;
801 case nir_op_ige:
802 return nir_op_ilt;
803 case nir_op_uge:
804 return nir_op_ult;
805 case nir_op_flt:
806 return nir_op_fge;
807 case nir_op_ilt:
808 return nir_op_ige;
809 case nir_op_ult:
810 return nir_op_uge;
811 case nir_op_feq:
812 return nir_op_fneu;
813 case nir_op_ieq:
814 return nir_op_ine;
815 case nir_op_fneu:
816 return nir_op_feq;
817 case nir_op_ine:
818 return nir_op_ieq;
819 default:
820 unreachable("Unsuported comparison!");
821 }
822 }
823
824 static int32_t
get_iteration(nir_op cond_op,nir_const_value initial,nir_const_value step,nir_const_value limit,bool invert_cond,unsigned bit_size,unsigned execution_mode)825 get_iteration(nir_op cond_op, nir_const_value initial, nir_const_value step,
826 nir_const_value limit, bool invert_cond, unsigned bit_size,
827 unsigned execution_mode)
828 {
829 nir_const_value span, iter;
830 unsigned iter_bit_size = bit_size;
831
832 switch (invert_comparison_if_needed(cond_op, invert_cond)) {
833 case nir_op_ine:
834 /* In order for execution to be here, limit must be the same as initial.
835 * Otherwise will_break_on_first_iteration would have returned false.
836 * If step is zero, the loop is infinite. Otherwise the loop will
837 * execute once.
838 */
839 return step.u64 == 0 ? -1 : 1;
840
841 case nir_op_ige:
842 case nir_op_ilt:
843 case nir_op_ieq:
844 span = eval_const_binop(nir_op_isub, bit_size, limit, initial,
845 execution_mode);
846 iter = eval_const_binop(nir_op_idiv, bit_size, span, step,
847 execution_mode);
848 break;
849
850 case nir_op_uge:
851 case nir_op_ult:
852 span = eval_const_binop(nir_op_isub, bit_size, limit, initial,
853 execution_mode);
854 iter = eval_const_binop(nir_op_udiv, bit_size, span, step,
855 execution_mode);
856 break;
857
858 case nir_op_fneu:
859 /* In order for execution to be here, limit must be the same as initial.
860 * Otherwise will_break_on_first_iteration would have returned false.
861 * If step is zero, the loop is infinite. Otherwise the loop will
862 * execute once.
863 *
864 * This is a little more tricky for floating point since X-Y might still
865 * be X even if Y is not zero. Instead check that (initial + step) !=
866 * initial.
867 */
868 span = eval_const_binop(nir_op_fadd, bit_size, initial, step,
869 execution_mode);
870 iter = eval_const_binop(nir_op_feq, bit_size, initial,
871 span, execution_mode);
872
873 /* return (initial + step) == initial ? -1 : 1 */
874 return iter.b ? -1 : 1;
875
876 case nir_op_fge:
877 case nir_op_flt:
878 case nir_op_feq:
879 span = eval_const_binop(nir_op_fsub, bit_size, limit, initial,
880 execution_mode);
881 iter = eval_const_binop(nir_op_fdiv, bit_size, span,
882 step, execution_mode);
883 iter = eval_const_unop(nir_op_f2i64, bit_size, iter, execution_mode);
884 iter_bit_size = 64;
885 break;
886
887 default:
888 return -1;
889 }
890
891 uint64_t iter_u64 = nir_const_value_as_uint(iter, iter_bit_size);
892 return iter_u64 > u_intN_max(iter_bit_size) ? -1 : (int)iter_u64;
893 }
894
895 static int32_t
get_iteration_empirical(nir_scalar cond,nir_alu_instr * incr_alu,nir_scalar basis,nir_const_value initial,nir_scalar limit_basis,nir_const_value limit,bool invert_cond,unsigned execution_mode,unsigned max_unroll_iterations)896 get_iteration_empirical(nir_scalar cond, nir_alu_instr *incr_alu,
897 nir_scalar basis, nir_const_value initial,
898 nir_scalar limit_basis, nir_const_value limit,
899 bool invert_cond, unsigned execution_mode,
900 unsigned max_unroll_iterations)
901 {
902 int iter_count = 0;
903 nir_const_value result;
904
905 const nir_scalar incr = nir_get_scalar(&incr_alu->def, basis.comp);
906
907 const nir_scalar original[] = {basis, limit_basis};
908 nir_const_value replacement[] = {initial, limit};
909
910 while (iter_count <= max_unroll_iterations) {
911 bool success;
912
913 success = try_eval_const_alu(&result, cond, original, replacement,
914 2, execution_mode);
915 if (!success)
916 return -1;
917
918 const bool cond_succ = invert_cond ? !result.b : result.b;
919 if (cond_succ)
920 return iter_count;
921
922 iter_count++;
923
924 success = try_eval_const_alu(&result, incr, original, replacement,
925 2, execution_mode);
926 assert(success);
927
928 replacement[0] = result;
929 }
930
931 return -1;
932 }
933
934 static bool
will_break_on_first_iteration(nir_scalar cond,nir_scalar basis,nir_scalar limit_basis,nir_const_value initial,nir_const_value limit,bool invert_cond,unsigned execution_mode)935 will_break_on_first_iteration(nir_scalar cond, nir_scalar basis,
936 nir_scalar limit_basis,
937 nir_const_value initial, nir_const_value limit,
938 bool invert_cond, unsigned execution_mode)
939 {
940 nir_const_value result;
941
942 const nir_scalar originals[2] = { basis, limit_basis };
943 const nir_const_value replacements[2] = { initial, limit };
944
945 ASSERTED bool success = try_eval_const_alu(&result, cond, originals,
946 replacements, 2, execution_mode);
947
948 assert(success);
949
950 return invert_cond ? !result.b : result.b;
951 }
952
953 static bool
test_iterations(int32_t iter_int,nir_const_value step,nir_const_value limit,nir_op cond_op,unsigned bit_size,nir_alu_type induction_base_type,nir_const_value initial,bool limit_rhs,bool invert_cond,unsigned execution_mode)954 test_iterations(int32_t iter_int, nir_const_value step,
955 nir_const_value limit, nir_op cond_op, unsigned bit_size,
956 nir_alu_type induction_base_type,
957 nir_const_value initial, bool limit_rhs, bool invert_cond,
958 unsigned execution_mode)
959 {
960 assert(nir_op_infos[cond_op].num_inputs == 2);
961
962 nir_const_value iter_src;
963 nir_op mul_op;
964 nir_op add_op;
965 switch (induction_base_type) {
966 case nir_type_float:
967 iter_src = nir_const_value_for_float(iter_int, bit_size);
968 mul_op = nir_op_fmul;
969 add_op = nir_op_fadd;
970 break;
971 case nir_type_int:
972 case nir_type_uint:
973 iter_src = nir_const_value_for_int(iter_int, bit_size);
974 mul_op = nir_op_imul;
975 add_op = nir_op_iadd;
976 break;
977 default:
978 unreachable("Unhandled induction variable base type!");
979 }
980
981 /* Multiple the iteration count we are testing by the number of times we
982 * step the induction variable each iteration.
983 */
984 nir_const_value mul_result =
985 eval_const_binop(mul_op, bit_size, iter_src, step, execution_mode);
986
987 /* Add the initial value to the accumulated induction variable total */
988 nir_const_value add_result =
989 eval_const_binop(add_op, bit_size, mul_result, initial, execution_mode);
990
991 nir_const_value *src[2];
992 src[limit_rhs ? 0 : 1] = &add_result;
993 src[limit_rhs ? 1 : 0] = &limit;
994
995 /* Evaluate the loop exit condition */
996 nir_const_value result;
997 nir_eval_const_opcode(cond_op, &result, 1, bit_size, src, execution_mode);
998
999 return invert_cond ? !result.b : result.b;
1000 }
1001
1002 static int
calculate_iterations(nir_scalar basis,nir_scalar limit_basis,nir_const_value initial,nir_const_value step,nir_const_value limit,nir_alu_instr * alu,nir_scalar cond,nir_op alu_op,bool limit_rhs,bool invert_cond,unsigned execution_mode,unsigned max_unroll_iterations)1003 calculate_iterations(nir_scalar basis, nir_scalar limit_basis,
1004 nir_const_value initial, nir_const_value step,
1005 nir_const_value limit, nir_alu_instr *alu,
1006 nir_scalar cond, nir_op alu_op, bool limit_rhs,
1007 bool invert_cond, unsigned execution_mode,
1008 unsigned max_unroll_iterations)
1009 {
1010 /* nir_op_isub should have been lowered away by this point */
1011 assert(alu->op != nir_op_isub);
1012
1013 /* Make sure the alu type for our induction variable is compatible with the
1014 * conditional alus input type. If its not something has gone really wrong.
1015 */
1016 nir_alu_type induction_base_type =
1017 nir_alu_type_get_base_type(nir_op_infos[alu->op].output_type);
1018 if (induction_base_type == nir_type_int || induction_base_type == nir_type_uint) {
1019 assert(nir_alu_type_get_base_type(nir_op_infos[alu_op].input_types[1]) == nir_type_int ||
1020 nir_alu_type_get_base_type(nir_op_infos[alu_op].input_types[1]) == nir_type_uint);
1021 } else {
1022 assert(nir_alu_type_get_base_type(nir_op_infos[alu_op].input_types[0]) ==
1023 induction_base_type);
1024 }
1025
1026 /* do-while loops can increment the starting value before the condition is
1027 * checked. e.g.
1028 *
1029 * do {
1030 * ndx++;
1031 * } while (ndx < 3);
1032 *
1033 * Here we check if the induction variable is used directly by the loop
1034 * condition and if so we assume we need to step the initial value.
1035 */
1036 unsigned trip_offset = 0;
1037 nir_alu_instr *cond_alu = nir_instr_as_alu(cond.def->parent_instr);
1038 if (cond_alu->src[0].src.ssa == &alu->def ||
1039 cond_alu->src[1].src.ssa == &alu->def) {
1040 trip_offset = 1;
1041 }
1042
1043 unsigned bit_size = nir_src_bit_size(alu->src[0].src);
1044
1045 /* get_iteration works under assumption that iterator will be
1046 * incremented or decremented until it hits the limit,
1047 * however if the loop condition is false on the first iteration
1048 * get_iteration's assumption is broken. Handle such loops first.
1049 */
1050 if (will_break_on_first_iteration(cond, basis, limit_basis, initial,
1051 limit, invert_cond, execution_mode)) {
1052 return 0;
1053 }
1054
1055 /* For loops incremented with addition operation, it's easy to
1056 * calculate the number of iterations theoretically. Even though it
1057 * is possible for other operations as well, it is much more error
1058 * prone, and doesn't cover all possible cases. So, we try to
1059 * emulate the loop.
1060 */
1061 int iter_int;
1062 switch (alu->op) {
1063 case nir_op_iadd:
1064 case nir_op_fadd:
1065 assert(nir_src_bit_size(alu->src[0].src) ==
1066 nir_src_bit_size(alu->src[1].src));
1067
1068 iter_int = get_iteration(alu_op, initial, step, limit, invert_cond,
1069 bit_size, execution_mode);
1070 break;
1071 case nir_op_fmul:
1072 /* Detecting non-zero loop counts when the loop increment is floating
1073 * point multiplication triggers a preexisting problem in
1074 * glsl-fs-loop-unroll-mul-fp64.shader_test. See
1075 * https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/3445#note_1779438.
1076 */
1077 return -1;
1078 case nir_op_imul:
1079 case nir_op_ishl:
1080 case nir_op_ishr:
1081 case nir_op_ushr:
1082 return get_iteration_empirical(cond, alu, basis, initial,
1083 limit_basis, limit, invert_cond,
1084 execution_mode, max_unroll_iterations);
1085 default:
1086 unreachable("Invalid induction variable increment operation.");
1087 }
1088
1089 /* If iter_int is negative the loop is ill-formed or is the conditional is
1090 * unsigned with a huge iteration count so don't bother going any further.
1091 */
1092 if (iter_int < 0)
1093 return -1;
1094
1095 nir_op actual_alu_op = invert_comparison_if_needed(alu_op, invert_cond);
1096 if (actual_alu_op == nir_op_ine || actual_alu_op == nir_op_fneu)
1097 return iter_int;
1098
1099 /* An explanation from the GLSL unrolling pass:
1100 *
1101 * Make sure that the calculated number of iterations satisfies the exit
1102 * condition. This is needed to catch off-by-one errors and some types of
1103 * ill-formed loops. For example, we need to detect that the following
1104 * loop does not have a maximum iteration count.
1105 *
1106 * for (float x = 0.0; x != 0.9; x += 0.2);
1107 */
1108 for (int bias = -1; bias <= 1; bias++) {
1109 const int iter_bias = iter_int + bias;
1110 if (iter_bias < 1)
1111 continue;
1112
1113 if (test_iterations(iter_bias, step, limit, alu_op, bit_size,
1114 induction_base_type, initial,
1115 limit_rhs, invert_cond, execution_mode)) {
1116 return iter_bias - trip_offset;
1117 }
1118 }
1119
1120 return -1;
1121 }
1122
1123 static bool
get_induction_and_limit_vars(nir_scalar cond,nir_scalar * ind,nir_scalar * limit,bool * limit_rhs,loop_info_state * state)1124 get_induction_and_limit_vars(nir_scalar cond,
1125 nir_scalar *ind,
1126 nir_scalar *limit,
1127 bool *limit_rhs,
1128 loop_info_state *state)
1129 {
1130 nir_scalar rhs, lhs;
1131 lhs = nir_scalar_chase_alu_src(cond, 0);
1132 rhs = nir_scalar_chase_alu_src(cond, 1);
1133
1134 nir_loop_variable *src0_lv = get_loop_var(lhs.def, state);
1135 nir_loop_variable *src1_lv = get_loop_var(rhs.def, state);
1136
1137 if (src0_lv->type == basic_induction) {
1138 if (!nir_src_is_const(*src0_lv->init_src))
1139 return false;
1140
1141 *ind = lhs;
1142 *limit = rhs;
1143 *limit_rhs = true;
1144 return true;
1145 } else if (src1_lv->type == basic_induction) {
1146 if (!nir_src_is_const(*src1_lv->init_src))
1147 return false;
1148
1149 *ind = rhs;
1150 *limit = lhs;
1151 *limit_rhs = false;
1152 return true;
1153 } else {
1154 return false;
1155 }
1156 }
1157
1158 static bool
try_find_trip_count_vars_in_logical_op(nir_scalar * cond,nir_scalar * ind,nir_scalar * limit,bool * limit_rhs,loop_info_state * state)1159 try_find_trip_count_vars_in_logical_op(nir_scalar *cond,
1160 nir_scalar *ind,
1161 nir_scalar *limit,
1162 bool *limit_rhs,
1163 loop_info_state *state)
1164 {
1165 const nir_op alu_op = nir_scalar_alu_op(*cond);
1166 bool exit_loop_on_false = alu_op == nir_op_ieq || alu_op == nir_op_inot;
1167 nir_scalar logical_op = exit_loop_on_false ?
1168 nir_scalar_chase_alu_src(*cond, 0) : *cond;
1169
1170 if (alu_op == nir_op_ieq) {
1171 nir_scalar zero = nir_scalar_chase_alu_src(*cond, 1);
1172
1173 if (!nir_scalar_is_alu(logical_op) || !nir_scalar_is_const(zero)) {
1174 /* Maybe we had it the wrong way, flip things around */
1175 nir_scalar tmp = zero;
1176 zero = logical_op;
1177 logical_op = tmp;
1178
1179 /* If we still didn't find what we need then return */
1180 if (!nir_scalar_is_const(zero))
1181 return false;
1182 }
1183
1184 /* If the loop is not breaking on (x && y) == 0 then return */
1185 if (nir_scalar_as_uint(zero) != 0)
1186 return false;
1187 }
1188
1189 if (!nir_scalar_is_alu(logical_op))
1190 return false;
1191
1192 if ((exit_loop_on_false && (nir_scalar_alu_op(logical_op) != nir_op_iand)) ||
1193 (!exit_loop_on_false && (nir_scalar_alu_op(logical_op) != nir_op_ior)))
1194 return false;
1195
1196 /* Check if iand src is a terminator condition and try get induction var
1197 * and trip limit var.
1198 */
1199 bool found_induction_var = false;
1200 for (unsigned i = 0; i < 2; i++) {
1201 nir_scalar src = nir_scalar_chase_alu_src(logical_op, i);
1202 if (nir_is_terminator_condition_with_two_inputs(src) &&
1203 get_induction_and_limit_vars(src, ind, limit, limit_rhs, state)) {
1204 *cond = src;
1205 found_induction_var = true;
1206
1207 /* If we've found one with a constant limit, stop. */
1208 if (nir_scalar_is_const(*limit))
1209 return true;
1210 }
1211 }
1212
1213 return found_induction_var;
1214 }
1215
1216 /* Run through each of the terminators of the loop and try to infer a possible
1217 * trip-count. We need to check them all, and set the lowest trip-count as the
1218 * trip-count of our loop. If one of the terminators has an undecidable
1219 * trip-count we can not safely assume anything about the duration of the
1220 * loop.
1221 */
1222 static void
find_trip_count(loop_info_state * state,unsigned execution_mode,unsigned max_unroll_iterations)1223 find_trip_count(loop_info_state *state, unsigned execution_mode,
1224 unsigned max_unroll_iterations)
1225 {
1226 bool trip_count_known = true;
1227 bool guessed_trip_count = false;
1228 nir_loop_terminator *limiting_terminator = NULL;
1229 int max_trip_count = -1;
1230
1231 list_for_each_entry(nir_loop_terminator, terminator,
1232 &state->loop->info->loop_terminator_list,
1233 loop_terminator_link) {
1234 nir_scalar cond = { terminator->nif->condition.ssa, 0 };
1235
1236 if (!nir_scalar_is_alu(cond)) {
1237 /* If we get here the loop is dead and will get cleaned up by the
1238 * nir_opt_dead_cf pass.
1239 */
1240 trip_count_known = false;
1241 terminator->exact_trip_count_unknown = true;
1242 continue;
1243 }
1244
1245 nir_op alu_op = nir_scalar_alu_op(cond);
1246
1247 bool invert_cond = terminator->continue_from_then;
1248
1249 bool limit_rhs;
1250 nir_scalar basic_ind = { NULL, 0 };
1251 nir_scalar limit;
1252
1253 if ((alu_op == nir_op_inot || alu_op == nir_op_ieq || alu_op == nir_op_ior) &&
1254 try_find_trip_count_vars_in_logical_op(&cond, &basic_ind, &limit,
1255 &limit_rhs, state)) {
1256
1257 /* The loop is exiting on (x && y) == 0 so we need to get the
1258 * inverse of x or y (i.e. which ever contained the induction var) in
1259 * order to compute the trip count.
1260 */
1261 if (alu_op == nir_op_inot || alu_op == nir_op_ieq)
1262 invert_cond = !invert_cond;
1263
1264 alu_op = nir_scalar_alu_op(cond);
1265 trip_count_known = false;
1266 terminator->conditional_instr = cond.def->parent_instr;
1267 terminator->exact_trip_count_unknown = true;
1268 }
1269
1270 if (!basic_ind.def) {
1271 if (nir_is_supported_terminator_condition(cond)) {
1272 /* Extract and inverse the comparision if it is wrapped in an inot
1273 */
1274 if (alu_op == nir_op_inot) {
1275 cond = nir_scalar_chase_alu_src(cond, 0);
1276 alu_op = nir_scalar_alu_op(cond);
1277 invert_cond = !invert_cond;
1278 }
1279
1280 get_induction_and_limit_vars(cond, &basic_ind,
1281 &limit, &limit_rhs, state);
1282 }
1283 }
1284
1285 /* The comparison has to have a basic induction variable for us to be
1286 * able to find trip counts.
1287 */
1288 if (!basic_ind.def) {
1289 trip_count_known = false;
1290 terminator->exact_trip_count_unknown = true;
1291 continue;
1292 }
1293
1294 terminator->induction_rhs = !limit_rhs;
1295
1296 /* Attempt to find a constant limit for the loop */
1297 nir_const_value limit_val;
1298 if (nir_scalar_is_const(limit)) {
1299 limit_val = nir_scalar_as_const_value(limit);
1300 } else {
1301 trip_count_known = false;
1302
1303 if (!try_find_limit_of_alu(limit, &limit_val, alu_op, invert_cond, terminator, state)) {
1304 /* Guess loop limit based on array access */
1305 if (!guess_loop_limit(state, &limit_val, basic_ind)) {
1306 terminator->exact_trip_count_unknown = true;
1307 continue;
1308 }
1309
1310 guessed_trip_count = true;
1311 }
1312 }
1313
1314 /* We have determined that we have the following constants:
1315 * (With the typical int i = 0; i < x; i++; as an example)
1316 * - Upper limit.
1317 * - Starting value
1318 * - Step / iteration size
1319 * Thats all thats needed to calculate the trip-count
1320 */
1321
1322 nir_loop_variable *lv = get_loop_var(basic_ind.def, state);
1323
1324 /* The basic induction var might be a vector but, because we guarantee
1325 * earlier that the phi source has a scalar swizzle, we can take the
1326 * component from basic_ind.
1327 */
1328 nir_scalar initial_s = { lv->init_src->ssa, basic_ind.comp };
1329 nir_scalar alu_s = {
1330 lv->update_src->src.ssa,
1331 lv->update_src->swizzle[basic_ind.comp]
1332 };
1333
1334 /* We are not guaranteed by that at one of these sources is a constant.
1335 * Try to find one.
1336 */
1337 if (!nir_scalar_is_const(initial_s) ||
1338 !nir_scalar_is_const(alu_s))
1339 continue;
1340
1341 nir_const_value initial_val = nir_scalar_as_const_value(initial_s);
1342 nir_const_value step_val = nir_scalar_as_const_value(alu_s);
1343
1344 int iterations = calculate_iterations(nir_get_scalar(lv->basis, basic_ind.comp), limit,
1345 initial_val, step_val, limit_val,
1346 nir_instr_as_alu(nir_src_parent_instr(&lv->update_src->src)),
1347 cond,
1348 alu_op, limit_rhs,
1349 invert_cond,
1350 execution_mode,
1351 max_unroll_iterations);
1352
1353 /* Where we not able to calculate the iteration count */
1354 if (iterations == -1) {
1355 trip_count_known = false;
1356 guessed_trip_count = false;
1357 terminator->exact_trip_count_unknown = true;
1358 continue;
1359 }
1360
1361 if (guessed_trip_count) {
1362 guessed_trip_count = false;
1363 terminator->exact_trip_count_unknown = true;
1364 if (state->loop->info->guessed_trip_count == 0 ||
1365 state->loop->info->guessed_trip_count > iterations)
1366 state->loop->info->guessed_trip_count = iterations;
1367
1368 continue;
1369 }
1370
1371 /* If this is the first run or we have found a smaller amount of
1372 * iterations than previously (we have identified a more limiting
1373 * terminator) set the trip count and limiting terminator.
1374 */
1375 if (max_trip_count == -1 || iterations < max_trip_count) {
1376 max_trip_count = iterations;
1377 limiting_terminator = terminator;
1378 }
1379 }
1380
1381 state->loop->info->exact_trip_count_known = trip_count_known;
1382 if (max_trip_count > -1)
1383 state->loop->info->max_trip_count = max_trip_count;
1384 state->loop->info->limiting_terminator = limiting_terminator;
1385 }
1386
1387 static bool
force_unroll_array_access(loop_info_state * state,nir_deref_instr * deref,bool contains_sampler)1388 force_unroll_array_access(loop_info_state *state, nir_deref_instr *deref,
1389 bool contains_sampler)
1390 {
1391 unsigned array_size = find_array_access_via_induction(state, deref, NULL);
1392 if (array_size) {
1393 if ((array_size == state->loop->info->max_trip_count) &&
1394 nir_deref_mode_must_be(deref, nir_var_shader_in |
1395 nir_var_shader_out |
1396 nir_var_shader_temp |
1397 nir_var_function_temp))
1398 return true;
1399
1400 if (nir_deref_mode_must_be(deref, state->indirect_mask))
1401 return true;
1402
1403 if (contains_sampler && state->force_unroll_sampler_indirect)
1404 return true;
1405 }
1406
1407 return false;
1408 }
1409
1410 static bool
force_unroll_heuristics(loop_info_state * state,nir_block * block)1411 force_unroll_heuristics(loop_info_state *state, nir_block *block)
1412 {
1413 nir_foreach_instr(instr, block) {
1414 if (instr->type == nir_instr_type_tex) {
1415 nir_tex_instr *tex_instr = nir_instr_as_tex(instr);
1416 int sampler_idx =
1417 nir_tex_instr_src_index(tex_instr,
1418 nir_tex_src_sampler_deref);
1419
1420 if (sampler_idx >= 0) {
1421 nir_deref_instr *deref =
1422 nir_instr_as_deref(tex_instr->src[sampler_idx].src.ssa->parent_instr);
1423 if (force_unroll_array_access(state, deref, true))
1424 return true;
1425 }
1426 }
1427
1428 if (instr->type != nir_instr_type_intrinsic)
1429 continue;
1430
1431 nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
1432
1433 /* Check for arrays variably-indexed by a loop induction variable.
1434 * Unrolling the loop may convert that access into constant-indexing.
1435 */
1436 if (intrin->intrinsic == nir_intrinsic_load_deref ||
1437 intrin->intrinsic == nir_intrinsic_store_deref ||
1438 intrin->intrinsic == nir_intrinsic_copy_deref) {
1439 if (force_unroll_array_access(state,
1440 nir_src_as_deref(intrin->src[0]),
1441 false))
1442 return true;
1443
1444 if (intrin->intrinsic == nir_intrinsic_copy_deref &&
1445 force_unroll_array_access(state,
1446 nir_src_as_deref(intrin->src[1]),
1447 false))
1448 return true;
1449 }
1450 }
1451
1452 return false;
1453 }
1454
1455 static void
get_loop_info(loop_info_state * state,nir_function_impl * impl)1456 get_loop_info(loop_info_state *state, nir_function_impl *impl)
1457 {
1458 nir_shader *shader = impl->function->shader;
1459 const nir_shader_compiler_options *options = shader->options;
1460
1461 /* Add all entries in the outermost part of the loop to the processing list
1462 * Mark the entries in conditionals or in nested loops accordingly
1463 */
1464 foreach_list_typed_safe(nir_cf_node, node, node, &state->loop->body) {
1465 switch (node->type) {
1466
1467 case nir_cf_node_block:
1468 init_loop_block(nir_cf_node_as_block(node), state, false, false);
1469 break;
1470
1471 case nir_cf_node_if:
1472 nir_foreach_block_in_cf_node(block, node)
1473 init_loop_block(block, state, true, false);
1474 break;
1475
1476 case nir_cf_node_loop:
1477 nir_foreach_block_in_cf_node(block, node) {
1478 init_loop_block(block, state, false, true);
1479 }
1480 break;
1481
1482 case nir_cf_node_function:
1483 break;
1484 }
1485 }
1486
1487 /* Try to find all simple terminators of the loop. If we can't find any,
1488 * or we find possible terminators that have side effects then bail.
1489 */
1490 if (!find_loop_terminators(state)) {
1491 list_for_each_entry_safe(nir_loop_terminator, terminator,
1492 &state->loop->info->loop_terminator_list,
1493 loop_terminator_link) {
1494 list_del(&terminator->loop_terminator_link);
1495 ralloc_free(terminator);
1496 }
1497 return;
1498 }
1499
1500 if (!compute_induction_information(state))
1501 return;
1502
1503 /* Run through each of the terminators and try to compute a trip-count */
1504 find_trip_count(state,
1505 impl->function->shader->info.float_controls_execution_mode,
1506 impl->function->shader->options->max_unroll_iterations);
1507
1508 nir_foreach_block_in_cf_node(block, &state->loop->cf_node) {
1509 nir_foreach_instr(instr, block) {
1510 state->loop->info->instr_cost += instr_cost(state, instr, options);
1511 }
1512
1513 if (state->loop->info->force_unroll)
1514 continue;
1515
1516 if (force_unroll_heuristics(state, block)) {
1517 state->loop->info->force_unroll = true;
1518 }
1519 }
1520 }
1521
1522 static loop_info_state *
initialize_loop_info_state(nir_loop * loop,void * mem_ctx,nir_function_impl * impl)1523 initialize_loop_info_state(nir_loop *loop, void *mem_ctx,
1524 nir_function_impl *impl)
1525 {
1526 loop_info_state *state = rzalloc(mem_ctx, loop_info_state);
1527 state->loop_vars = ralloc_array(mem_ctx, nir_loop_variable,
1528 impl->ssa_alloc);
1529 state->loop_vars_init = rzalloc_array(mem_ctx, BITSET_WORD,
1530 BITSET_WORDS(impl->ssa_alloc));
1531 state->loop = loop;
1532
1533 list_inithead(&state->process_list);
1534
1535 if (loop->info)
1536 ralloc_free(loop->info);
1537
1538 loop->info = rzalloc(loop, nir_loop_info);
1539
1540 list_inithead(&loop->info->loop_terminator_list);
1541
1542 return state;
1543 }
1544
1545 static void
process_loops(nir_cf_node * cf_node,nir_variable_mode indirect_mask,bool force_unroll_sampler_indirect)1546 process_loops(nir_cf_node *cf_node, nir_variable_mode indirect_mask,
1547 bool force_unroll_sampler_indirect)
1548 {
1549 switch (cf_node->type) {
1550 case nir_cf_node_block:
1551 return;
1552 case nir_cf_node_if: {
1553 nir_if *if_stmt = nir_cf_node_as_if(cf_node);
1554 foreach_list_typed(nir_cf_node, nested_node, node, &if_stmt->then_list)
1555 process_loops(nested_node, indirect_mask, force_unroll_sampler_indirect);
1556 foreach_list_typed(nir_cf_node, nested_node, node, &if_stmt->else_list)
1557 process_loops(nested_node, indirect_mask, force_unroll_sampler_indirect);
1558 return;
1559 }
1560 case nir_cf_node_loop: {
1561 nir_loop *loop = nir_cf_node_as_loop(cf_node);
1562 assert(!nir_loop_has_continue_construct(loop));
1563
1564 foreach_list_typed(nir_cf_node, nested_node, node, &loop->body)
1565 process_loops(nested_node, indirect_mask, force_unroll_sampler_indirect);
1566 break;
1567 }
1568 default:
1569 unreachable("unknown cf node type");
1570 }
1571
1572 nir_loop *loop = nir_cf_node_as_loop(cf_node);
1573 nir_function_impl *impl = nir_cf_node_get_function(cf_node);
1574 void *mem_ctx = ralloc_context(NULL);
1575
1576 loop_info_state *state = initialize_loop_info_state(loop, mem_ctx, impl);
1577 state->indirect_mask = indirect_mask;
1578 state->force_unroll_sampler_indirect = force_unroll_sampler_indirect;
1579
1580 get_loop_info(state, impl);
1581
1582 ralloc_free(mem_ctx);
1583 }
1584
1585 void
nir_loop_analyze_impl(nir_function_impl * impl,nir_variable_mode indirect_mask,bool force_unroll_sampler_indirect)1586 nir_loop_analyze_impl(nir_function_impl *impl,
1587 nir_variable_mode indirect_mask,
1588 bool force_unroll_sampler_indirect)
1589 {
1590 nir_index_ssa_defs(impl);
1591 foreach_list_typed(nir_cf_node, node, node, &impl->body)
1592 process_loops(node, indirect_mask, force_unroll_sampler_indirect);
1593 }
1594