1 /*
2 * Copyright © 2018 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 #include "nir_range_analysis.h"
24 #include <float.h>
25 #include <math.h>
26 #include "util/hash_table.h"
27 #include "util/u_dynarray.h"
28 #include "util/u_math.h"
29 #include "c99_alloca.h"
30 #include "nir.h"
31
32 /**
33 * Analyzes a sequence of operations to determine some aspects of the range of
34 * the result.
35 */
36
37 struct analysis_query {
38 uint32_t pushed_queries;
39 uint32_t result_index;
40 };
41
42 struct analysis_state {
43 nir_shader *shader;
44 const nir_unsigned_upper_bound_config *config;
45 struct hash_table *range_ht;
46
47 struct util_dynarray query_stack;
48 struct util_dynarray result_stack;
49
50 size_t query_size;
51 uintptr_t (*get_key)(struct analysis_query *q);
52 void (*process_query)(struct analysis_state *state, struct analysis_query *q,
53 uint32_t *result, const uint32_t *src);
54 };
55
56 static void *
push_analysis_query(struct analysis_state * state,size_t size)57 push_analysis_query(struct analysis_state *state, size_t size)
58 {
59 struct analysis_query *q = util_dynarray_grow_bytes(&state->query_stack, 1, size);
60 q->pushed_queries = 0;
61 q->result_index = util_dynarray_num_elements(&state->result_stack, uint32_t);
62
63 util_dynarray_append(&state->result_stack, uint32_t, 0);
64
65 return q;
66 }
67
68 /* Helper for performing range analysis without recursion. */
69 static uint32_t
perform_analysis(struct analysis_state * state)70 perform_analysis(struct analysis_state *state)
71 {
72 while (state->query_stack.size) {
73 struct analysis_query *cur =
74 (struct analysis_query *)((char *)util_dynarray_end(&state->query_stack) - state->query_size);
75 uint32_t *result = util_dynarray_element(&state->result_stack, uint32_t, cur->result_index);
76
77 uintptr_t key = state->get_key(cur);
78 struct hash_entry *he = NULL;
79 /* There might be a cycle-resolving entry for loop header phis. Ignore this when finishing
80 * them by testing pushed_queries.
81 */
82 if (cur->pushed_queries == 0 && key &&
83 (he = _mesa_hash_table_search(state->range_ht, (void *)key))) {
84 *result = (uintptr_t)he->data;
85 state->query_stack.size -= state->query_size;
86 continue;
87 }
88
89 uint32_t *src = (uint32_t *)util_dynarray_end(&state->result_stack) - cur->pushed_queries;
90 state->result_stack.size -= sizeof(uint32_t) * cur->pushed_queries;
91
92 uint32_t prev_num_queries = state->query_stack.size;
93 state->process_query(state, cur, result, src);
94
95 uint32_t num_queries = state->query_stack.size;
96 if (num_queries > prev_num_queries) {
97 cur = (struct analysis_query *)util_dynarray_element(&state->query_stack, char,
98 prev_num_queries - state->query_size);
99 cur->pushed_queries = (num_queries - prev_num_queries) / state->query_size;
100 continue;
101 }
102
103 if (key)
104 _mesa_hash_table_insert(state->range_ht, (void *)key, (void *)(uintptr_t)*result);
105
106 state->query_stack.size -= state->query_size;
107 }
108
109 assert(state->result_stack.size == sizeof(uint32_t));
110
111 uint32_t res = util_dynarray_top(&state->result_stack, uint32_t);
112 util_dynarray_fini(&state->query_stack);
113 util_dynarray_fini(&state->result_stack);
114
115 return res;
116 }
117
118 static bool
is_not_negative(enum ssa_ranges r)119 is_not_negative(enum ssa_ranges r)
120 {
121 return r == gt_zero || r == ge_zero || r == eq_zero;
122 }
123
124 static bool
is_not_zero(enum ssa_ranges r)125 is_not_zero(enum ssa_ranges r)
126 {
127 return r == gt_zero || r == lt_zero || r == ne_zero;
128 }
129
130 static uint32_t
pack_data(const struct ssa_result_range r)131 pack_data(const struct ssa_result_range r)
132 {
133 return r.range | r.is_integral << 8 | r.is_finite << 9 | r.is_a_number << 10;
134 }
135
136 static struct ssa_result_range
unpack_data(uint32_t v)137 unpack_data(uint32_t v)
138 {
139 return (struct ssa_result_range){
140 .range = v & 0xff,
141 .is_integral = (v & 0x00100) != 0,
142 .is_finite = (v & 0x00200) != 0,
143 .is_a_number = (v & 0x00400) != 0
144 };
145 }
146
147 static nir_alu_type
nir_alu_src_type(const nir_alu_instr * instr,unsigned src)148 nir_alu_src_type(const nir_alu_instr *instr, unsigned src)
149 {
150 return nir_alu_type_get_base_type(nir_op_infos[instr->op].input_types[src]) |
151 nir_src_bit_size(instr->src[src].src);
152 }
153
154 static struct ssa_result_range
analyze_constant(const struct nir_alu_instr * instr,unsigned src,nir_alu_type use_type)155 analyze_constant(const struct nir_alu_instr *instr, unsigned src,
156 nir_alu_type use_type)
157 {
158 uint8_t swizzle[NIR_MAX_VEC_COMPONENTS] = { 0, 1, 2, 3,
159 4, 5, 6, 7,
160 8, 9, 10, 11,
161 12, 13, 14, 15 };
162
163 /* If the source is an explicitly sized source, then we need to reset
164 * both the number of components and the swizzle.
165 */
166 const unsigned num_components = nir_ssa_alu_instr_src_components(instr, src);
167
168 for (unsigned i = 0; i < num_components; ++i)
169 swizzle[i] = instr->src[src].swizzle[i];
170
171 const nir_load_const_instr *const load =
172 nir_instr_as_load_const(instr->src[src].src.ssa->parent_instr);
173
174 struct ssa_result_range r = { unknown, false, false, false };
175
176 switch (nir_alu_type_get_base_type(use_type)) {
177 case nir_type_float: {
178 double min_value = NAN;
179 double max_value = NAN;
180 bool any_zero = false;
181 bool all_zero = true;
182
183 r.is_integral = true;
184 r.is_a_number = true;
185 r.is_finite = true;
186
187 for (unsigned i = 0; i < num_components; ++i) {
188 const double v = nir_const_value_as_float(load->value[swizzle[i]],
189 load->def.bit_size);
190
191 if (floor(v) != v)
192 r.is_integral = false;
193
194 if (isnan(v))
195 r.is_a_number = false;
196
197 if (!isfinite(v))
198 r.is_finite = false;
199
200 any_zero = any_zero || (v == 0.0);
201 all_zero = all_zero && (v == 0.0);
202 min_value = fmin(min_value, v);
203 max_value = fmax(max_value, v);
204 }
205
206 assert(any_zero >= all_zero);
207 assert(isnan(max_value) || max_value >= min_value);
208
209 if (all_zero)
210 r.range = eq_zero;
211 else if (min_value > 0.0)
212 r.range = gt_zero;
213 else if (min_value == 0.0)
214 r.range = ge_zero;
215 else if (max_value < 0.0)
216 r.range = lt_zero;
217 else if (max_value == 0.0)
218 r.range = le_zero;
219 else if (!any_zero)
220 r.range = ne_zero;
221 else
222 r.range = unknown;
223
224 return r;
225 }
226
227 case nir_type_int:
228 case nir_type_bool: {
229 int64_t min_value = INT_MAX;
230 int64_t max_value = INT_MIN;
231 bool any_zero = false;
232 bool all_zero = true;
233
234 for (unsigned i = 0; i < num_components; ++i) {
235 const int64_t v = nir_const_value_as_int(load->value[swizzle[i]],
236 load->def.bit_size);
237
238 any_zero = any_zero || (v == 0);
239 all_zero = all_zero && (v == 0);
240 min_value = MIN2(min_value, v);
241 max_value = MAX2(max_value, v);
242 }
243
244 assert(any_zero >= all_zero);
245 assert(max_value >= min_value);
246
247 if (all_zero)
248 r.range = eq_zero;
249 else if (min_value > 0)
250 r.range = gt_zero;
251 else if (min_value == 0)
252 r.range = ge_zero;
253 else if (max_value < 0)
254 r.range = lt_zero;
255 else if (max_value == 0)
256 r.range = le_zero;
257 else if (!any_zero)
258 r.range = ne_zero;
259 else
260 r.range = unknown;
261
262 return r;
263 }
264
265 case nir_type_uint: {
266 bool any_zero = false;
267 bool all_zero = true;
268
269 for (unsigned i = 0; i < num_components; ++i) {
270 const uint64_t v = nir_const_value_as_uint(load->value[swizzle[i]],
271 load->def.bit_size);
272
273 any_zero = any_zero || (v == 0);
274 all_zero = all_zero && (v == 0);
275 }
276
277 assert(any_zero >= all_zero);
278
279 if (all_zero)
280 r.range = eq_zero;
281 else if (any_zero)
282 r.range = ge_zero;
283 else
284 r.range = gt_zero;
285
286 return r;
287 }
288
289 default:
290 unreachable("Invalid alu source type");
291 }
292 }
293
294 /**
295 * Short-hand name for use in the tables in process_fp_query. If this name
296 * becomes a problem on some compiler, we can change it to _.
297 */
298 #define _______ unknown
299
300 #if defined(__clang__)
301 /* clang wants _Pragma("unroll X") */
302 #define pragma_unroll_5 _Pragma("unroll 5")
303 #define pragma_unroll_7 _Pragma("unroll 7")
304 /* gcc wants _Pragma("GCC unroll X") */
305 #elif defined(__GNUC__)
306 #if __GNUC__ >= 8
307 #define pragma_unroll_5 _Pragma("GCC unroll 5")
308 #define pragma_unroll_7 _Pragma("GCC unroll 7")
309 #else
310 #pragma GCC optimize("unroll-loops")
311 #define pragma_unroll_5
312 #define pragma_unroll_7
313 #endif
314 #else
315 /* MSVC doesn't have C99's _Pragma() */
316 #define pragma_unroll_5
317 #define pragma_unroll_7
318 #endif
319
320 #ifndef NDEBUG
321 #define ASSERT_TABLE_IS_COMMUTATIVE(t) \
322 do { \
323 static bool first = true; \
324 if (first) { \
325 first = false; \
326 pragma_unroll_7 for (unsigned r = 0; r < ARRAY_SIZE(t); r++) \
327 { \
328 pragma_unroll_7 for (unsigned c = 0; c < ARRAY_SIZE(t[0]); c++) \
329 assert(t[r][c] == t[c][r]); \
330 } \
331 } \
332 } while (false)
333
334 #define ASSERT_TABLE_IS_DIAGONAL(t) \
335 do { \
336 static bool first = true; \
337 if (first) { \
338 first = false; \
339 pragma_unroll_7 for (unsigned r = 0; r < ARRAY_SIZE(t); r++) \
340 assert(t[r][r] == r); \
341 } \
342 } while (false)
343
344 #else
345 #define ASSERT_TABLE_IS_COMMUTATIVE(t)
346 #define ASSERT_TABLE_IS_DIAGONAL(t)
347 #endif /* !defined(NDEBUG) */
348
349 static enum ssa_ranges
union_ranges(enum ssa_ranges a,enum ssa_ranges b)350 union_ranges(enum ssa_ranges a, enum ssa_ranges b)
351 {
352 static const enum ssa_ranges union_table[last_range + 1][last_range + 1] = {
353 /* left\right unknown lt_zero le_zero gt_zero ge_zero ne_zero eq_zero */
354 /* unknown */ { _______, _______, _______, _______, _______, _______, _______ },
355 /* lt_zero */ { _______, lt_zero, le_zero, ne_zero, _______, ne_zero, le_zero },
356 /* le_zero */ { _______, le_zero, le_zero, _______, _______, _______, le_zero },
357 /* gt_zero */ { _______, ne_zero, _______, gt_zero, ge_zero, ne_zero, ge_zero },
358 /* ge_zero */ { _______, _______, _______, ge_zero, ge_zero, _______, ge_zero },
359 /* ne_zero */ { _______, ne_zero, _______, ne_zero, _______, ne_zero, _______ },
360 /* eq_zero */ { _______, le_zero, le_zero, ge_zero, ge_zero, _______, eq_zero },
361 };
362
363 ASSERT_TABLE_IS_COMMUTATIVE(union_table);
364 ASSERT_TABLE_IS_DIAGONAL(union_table);
365
366 return union_table[a][b];
367 }
368
369 #ifndef NDEBUG
370 /* Verify that the 'unknown' entry in each row (or column) of the table is the
371 * union of all the other values in the row (or column).
372 */
373 #define ASSERT_UNION_OF_OTHERS_MATCHES_UNKNOWN_2_SOURCE(t) \
374 do { \
375 static bool first = true; \
376 if (first) { \
377 first = false; \
378 pragma_unroll_7 for (unsigned i = 0; i < last_range; i++) \
379 { \
380 enum ssa_ranges col_range = t[i][unknown + 1]; \
381 enum ssa_ranges row_range = t[unknown + 1][i]; \
382 \
383 pragma_unroll_5 for (unsigned j = unknown + 2; j < last_range; j++) \
384 { \
385 col_range = union_ranges(col_range, t[i][j]); \
386 row_range = union_ranges(row_range, t[j][i]); \
387 } \
388 \
389 assert(col_range == t[i][unknown]); \
390 assert(row_range == t[unknown][i]); \
391 } \
392 } \
393 } while (false)
394
395 /* For most operations, the union of ranges for a strict inequality and
396 * equality should be the range of the non-strict inequality (e.g.,
397 * union_ranges(range(op(lt_zero), range(op(eq_zero))) == range(op(le_zero)).
398 *
399 * Does not apply to selection-like opcodes (bcsel, fmin, fmax, etc.).
400 */
401 #define ASSERT_UNION_OF_EQ_AND_STRICT_INEQ_MATCHES_NONSTRICT_1_SOURCE(t) \
402 do { \
403 assert(union_ranges(t[lt_zero], t[eq_zero]) == t[le_zero]); \
404 assert(union_ranges(t[gt_zero], t[eq_zero]) == t[ge_zero]); \
405 } while (false)
406
407 #define ASSERT_UNION_OF_EQ_AND_STRICT_INEQ_MATCHES_NONSTRICT_2_SOURCE(t) \
408 do { \
409 static bool first = true; \
410 if (first) { \
411 first = false; \
412 pragma_unroll_7 for (unsigned i = 0; i < last_range; i++) \
413 { \
414 assert(union_ranges(t[i][lt_zero], t[i][eq_zero]) == t[i][le_zero]); \
415 assert(union_ranges(t[i][gt_zero], t[i][eq_zero]) == t[i][ge_zero]); \
416 assert(union_ranges(t[lt_zero][i], t[eq_zero][i]) == t[le_zero][i]); \
417 assert(union_ranges(t[gt_zero][i], t[eq_zero][i]) == t[ge_zero][i]); \
418 } \
419 } \
420 } while (false)
421
422 /* Several other unordered tuples span the range of "everything." Each should
423 * have the same value as unknown: (lt_zero, ge_zero), (le_zero, gt_zero), and
424 * (eq_zero, ne_zero). union_ranges is already commutative, so only one
425 * ordering needs to be checked.
426 *
427 * Does not apply to selection-like opcodes (bcsel, fmin, fmax, etc.).
428 *
429 * In cases where this can be used, it is unnecessary to also use
430 * ASSERT_UNION_OF_OTHERS_MATCHES_UNKNOWN_*_SOURCE. For any range X,
431 * union_ranges(X, X) == X. The disjoint ranges cover all of the non-unknown
432 * possibilities, so the union of all the unions of disjoint ranges is
433 * equivalent to the union of "others."
434 */
435 #define ASSERT_UNION_OF_DISJOINT_MATCHES_UNKNOWN_1_SOURCE(t) \
436 do { \
437 assert(union_ranges(t[lt_zero], t[ge_zero]) == t[unknown]); \
438 assert(union_ranges(t[le_zero], t[gt_zero]) == t[unknown]); \
439 assert(union_ranges(t[eq_zero], t[ne_zero]) == t[unknown]); \
440 } while (false)
441
442 #define ASSERT_UNION_OF_DISJOINT_MATCHES_UNKNOWN_2_SOURCE(t) \
443 do { \
444 static bool first = true; \
445 if (first) { \
446 first = false; \
447 pragma_unroll_7 for (unsigned i = 0; i < last_range; i++) \
448 { \
449 assert(union_ranges(t[i][lt_zero], t[i][ge_zero]) == \
450 t[i][unknown]); \
451 assert(union_ranges(t[i][le_zero], t[i][gt_zero]) == \
452 t[i][unknown]); \
453 assert(union_ranges(t[i][eq_zero], t[i][ne_zero]) == \
454 t[i][unknown]); \
455 \
456 assert(union_ranges(t[lt_zero][i], t[ge_zero][i]) == \
457 t[unknown][i]); \
458 assert(union_ranges(t[le_zero][i], t[gt_zero][i]) == \
459 t[unknown][i]); \
460 assert(union_ranges(t[eq_zero][i], t[ne_zero][i]) == \
461 t[unknown][i]); \
462 } \
463 } \
464 } while (false)
465
466 #else
467 #define ASSERT_UNION_OF_OTHERS_MATCHES_UNKNOWN_2_SOURCE(t)
468 #define ASSERT_UNION_OF_EQ_AND_STRICT_INEQ_MATCHES_NONSTRICT_1_SOURCE(t)
469 #define ASSERT_UNION_OF_EQ_AND_STRICT_INEQ_MATCHES_NONSTRICT_2_SOURCE(t)
470 #define ASSERT_UNION_OF_DISJOINT_MATCHES_UNKNOWN_1_SOURCE(t)
471 #define ASSERT_UNION_OF_DISJOINT_MATCHES_UNKNOWN_2_SOURCE(t)
472 #endif /* !defined(NDEBUG) */
473
474 struct fp_query {
475 struct analysis_query head;
476 const nir_alu_instr *instr;
477 unsigned src;
478 nir_alu_type use_type;
479 };
480
481 static void
push_fp_query(struct analysis_state * state,const nir_alu_instr * alu,unsigned src,nir_alu_type type)482 push_fp_query(struct analysis_state *state, const nir_alu_instr *alu, unsigned src, nir_alu_type type)
483 {
484 struct fp_query *pushed_q = push_analysis_query(state, sizeof(struct fp_query));
485 pushed_q->instr = alu;
486 pushed_q->src = src;
487 pushed_q->use_type = type == nir_type_invalid ? nir_alu_src_type(alu, src) : type;
488 }
489
490 static uintptr_t
get_fp_key(struct analysis_query * q)491 get_fp_key(struct analysis_query *q)
492 {
493 struct fp_query *fp_q = (struct fp_query *)q;
494 const nir_src *src = &fp_q->instr->src[fp_q->src].src;
495
496 if (src->ssa->parent_instr->type != nir_instr_type_alu)
497 return 0;
498
499 uintptr_t type_encoding;
500 uintptr_t ptr = (uintptr_t)nir_instr_as_alu(src->ssa->parent_instr);
501
502 /* The low 2 bits have to be zero or this whole scheme falls apart. */
503 assert((ptr & 0x3) == 0);
504
505 /* NIR is typeless in the sense that sequences of bits have whatever
506 * meaning is attached to them by the instruction that consumes them.
507 * However, the number of bits must match between producer and consumer.
508 * As a result, the number of bits does not need to be encoded here.
509 */
510 switch (nir_alu_type_get_base_type(fp_q->use_type)) {
511 case nir_type_int:
512 type_encoding = 0;
513 break;
514 case nir_type_uint:
515 type_encoding = 1;
516 break;
517 case nir_type_bool:
518 type_encoding = 2;
519 break;
520 case nir_type_float:
521 type_encoding = 3;
522 break;
523 default:
524 unreachable("Invalid base type.");
525 }
526
527 return ptr | type_encoding;
528 }
529
530 /**
531 * Analyze an expression to determine the range of its result
532 *
533 * The end result of this analysis is a token that communicates something
534 * about the range of values. There's an implicit grammar that produces
535 * tokens from sequences of literal values, other tokens, and operations.
536 * This function implements this grammar as a recursive-descent parser. Some
537 * (but not all) of the grammar is listed in-line in the function.
538 */
539 static void
process_fp_query(struct analysis_state * state,struct analysis_query * aq,uint32_t * result,const uint32_t * src_res)540 process_fp_query(struct analysis_state *state, struct analysis_query *aq, uint32_t *result,
541 const uint32_t *src_res)
542 {
543 /* Ensure that the _Pragma("GCC unroll 7") above are correct. */
544 STATIC_ASSERT(last_range + 1 == 7);
545
546 struct fp_query q = *(struct fp_query *)aq;
547 const nir_alu_instr *instr = q.instr;
548 unsigned src = q.src;
549 nir_alu_type use_type = q.use_type;
550
551 if (nir_src_is_const(instr->src[src].src)) {
552 *result = pack_data(analyze_constant(instr, src, use_type));
553 return;
554 }
555
556 if (instr->src[src].src.ssa->parent_instr->type != nir_instr_type_alu) {
557 *result = pack_data((struct ssa_result_range){ unknown, false, false, false });
558 return;
559 }
560
561 const struct nir_alu_instr *const alu =
562 nir_instr_as_alu(instr->src[src].src.ssa->parent_instr);
563
564 /* Bail if the type of the instruction generating the value does not match
565 * the type the value will be interpreted as. int/uint/bool can be
566 * reinterpreted trivially. The most important cases are between float and
567 * non-float.
568 */
569 if (alu->op != nir_op_mov && alu->op != nir_op_bcsel) {
570 const nir_alu_type use_base_type =
571 nir_alu_type_get_base_type(use_type);
572 const nir_alu_type src_base_type =
573 nir_alu_type_get_base_type(nir_op_infos[alu->op].output_type);
574
575 if (use_base_type != src_base_type &&
576 (use_base_type == nir_type_float ||
577 src_base_type == nir_type_float)) {
578 *result = pack_data((struct ssa_result_range){ unknown, false, false, false });
579 return;
580 }
581 }
582
583 if (!aq->pushed_queries) {
584 switch (alu->op) {
585 case nir_op_bcsel:
586 push_fp_query(state, alu, 1, use_type);
587 push_fp_query(state, alu, 2, use_type);
588 return;
589 case nir_op_mov:
590 push_fp_query(state, alu, 0, use_type);
591 return;
592 case nir_op_i2f32:
593 case nir_op_u2f32:
594 case nir_op_fabs:
595 case nir_op_fexp2:
596 case nir_op_frcp:
597 case nir_op_fneg:
598 case nir_op_fsat:
599 case nir_op_fsign:
600 case nir_op_ffloor:
601 case nir_op_fceil:
602 case nir_op_ftrunc:
603 case nir_op_fdot2:
604 case nir_op_fdot3:
605 case nir_op_fdot4:
606 case nir_op_fdot8:
607 case nir_op_fdot16:
608 case nir_op_fdot2_replicated:
609 case nir_op_fdot3_replicated:
610 case nir_op_fdot4_replicated:
611 case nir_op_fdot8_replicated:
612 case nir_op_fdot16_replicated:
613 push_fp_query(state, alu, 0, nir_type_invalid);
614 return;
615 case nir_op_fadd:
616 case nir_op_fmax:
617 case nir_op_fmin:
618 case nir_op_fmul:
619 case nir_op_fmulz:
620 case nir_op_fpow:
621 push_fp_query(state, alu, 0, nir_type_invalid);
622 push_fp_query(state, alu, 1, nir_type_invalid);
623 return;
624 case nir_op_ffma:
625 case nir_op_flrp:
626 push_fp_query(state, alu, 0, nir_type_invalid);
627 push_fp_query(state, alu, 1, nir_type_invalid);
628 push_fp_query(state, alu, 2, nir_type_invalid);
629 return;
630 default:
631 break;
632 }
633 }
634
635 struct ssa_result_range r = { unknown, false, false, false };
636
637 /* ge_zero: ge_zero + ge_zero
638 *
639 * gt_zero: gt_zero + eq_zero
640 * | gt_zero + ge_zero
641 * | eq_zero + gt_zero # Addition is commutative
642 * | ge_zero + gt_zero # Addition is commutative
643 * | gt_zero + gt_zero
644 * ;
645 *
646 * le_zero: le_zero + le_zero
647 *
648 * lt_zero: lt_zero + eq_zero
649 * | lt_zero + le_zero
650 * | eq_zero + lt_zero # Addition is commutative
651 * | le_zero + lt_zero # Addition is commutative
652 * | lt_zero + lt_zero
653 * ;
654 *
655 * ne_zero: eq_zero + ne_zero
656 * | ne_zero + eq_zero # Addition is commutative
657 * ;
658 *
659 * eq_zero: eq_zero + eq_zero
660 * ;
661 *
662 * All other cases are 'unknown'. The seeming odd entry is (ne_zero,
663 * ne_zero), but that could be (-5, +5) which is not ne_zero.
664 */
665 static const enum ssa_ranges fadd_table[last_range + 1][last_range + 1] = {
666 /* left\right unknown lt_zero le_zero gt_zero ge_zero ne_zero eq_zero */
667 /* unknown */ { _______, _______, _______, _______, _______, _______, _______ },
668 /* lt_zero */ { _______, lt_zero, lt_zero, _______, _______, _______, lt_zero },
669 /* le_zero */ { _______, lt_zero, le_zero, _______, _______, _______, le_zero },
670 /* gt_zero */ { _______, _______, _______, gt_zero, gt_zero, _______, gt_zero },
671 /* ge_zero */ { _______, _______, _______, gt_zero, ge_zero, _______, ge_zero },
672 /* ne_zero */ { _______, _______, _______, _______, _______, _______, ne_zero },
673 /* eq_zero */ { _______, lt_zero, le_zero, gt_zero, ge_zero, ne_zero, eq_zero },
674 };
675
676 ASSERT_TABLE_IS_COMMUTATIVE(fadd_table);
677 ASSERT_UNION_OF_DISJOINT_MATCHES_UNKNOWN_2_SOURCE(fadd_table);
678 ASSERT_UNION_OF_EQ_AND_STRICT_INEQ_MATCHES_NONSTRICT_2_SOURCE(fadd_table);
679
680 /* Due to flush-to-zero semanatics of floating-point numbers with very
681 * small mangnitudes, we can never really be sure a result will be
682 * non-zero.
683 *
684 * ge_zero: ge_zero * ge_zero
685 * | ge_zero * gt_zero
686 * | ge_zero * eq_zero
687 * | le_zero * lt_zero
688 * | lt_zero * le_zero # Multiplication is commutative
689 * | le_zero * le_zero
690 * | gt_zero * ge_zero # Multiplication is commutative
691 * | eq_zero * ge_zero # Multiplication is commutative
692 * | a * a # Left source == right source
693 * | gt_zero * gt_zero
694 * | lt_zero * lt_zero
695 * ;
696 *
697 * le_zero: ge_zero * le_zero
698 * | ge_zero * lt_zero
699 * | lt_zero * ge_zero # Multiplication is commutative
700 * | le_zero * ge_zero # Multiplication is commutative
701 * | le_zero * gt_zero
702 * | lt_zero * gt_zero
703 * | gt_zero * lt_zero # Multiplication is commutative
704 * ;
705 *
706 * eq_zero: eq_zero * <any>
707 * <any> * eq_zero # Multiplication is commutative
708 *
709 * All other cases are 'unknown'.
710 */
711 static const enum ssa_ranges fmul_table[last_range + 1][last_range + 1] = {
712 /* left\right unknown lt_zero le_zero gt_zero ge_zero ne_zero eq_zero */
713 /* unknown */ { _______, _______, _______, _______, _______, _______, eq_zero },
714 /* lt_zero */ { _______, ge_zero, ge_zero, le_zero, le_zero, _______, eq_zero },
715 /* le_zero */ { _______, ge_zero, ge_zero, le_zero, le_zero, _______, eq_zero },
716 /* gt_zero */ { _______, le_zero, le_zero, ge_zero, ge_zero, _______, eq_zero },
717 /* ge_zero */ { _______, le_zero, le_zero, ge_zero, ge_zero, _______, eq_zero },
718 /* ne_zero */ { _______, _______, _______, _______, _______, _______, eq_zero },
719 /* eq_zero */ { eq_zero, eq_zero, eq_zero, eq_zero, eq_zero, eq_zero, eq_zero }
720 };
721
722 ASSERT_TABLE_IS_COMMUTATIVE(fmul_table);
723 ASSERT_UNION_OF_DISJOINT_MATCHES_UNKNOWN_2_SOURCE(fmul_table);
724 ASSERT_UNION_OF_EQ_AND_STRICT_INEQ_MATCHES_NONSTRICT_2_SOURCE(fmul_table);
725
726 static const enum ssa_ranges fneg_table[last_range + 1] = {
727 /* unknown lt_zero le_zero gt_zero ge_zero ne_zero eq_zero */
728 _______, gt_zero, ge_zero, lt_zero, le_zero, ne_zero, eq_zero
729 };
730
731 ASSERT_UNION_OF_DISJOINT_MATCHES_UNKNOWN_1_SOURCE(fneg_table);
732 ASSERT_UNION_OF_EQ_AND_STRICT_INEQ_MATCHES_NONSTRICT_1_SOURCE(fneg_table);
733
734 switch (alu->op) {
735 case nir_op_b2f32:
736 case nir_op_b2i32:
737 /* b2f32 will generate either 0.0 or 1.0. This case is trivial.
738 *
739 * b2i32 will generate either 0x00000000 or 0x00000001. When those bit
740 * patterns are interpreted as floating point, they are 0.0 and
741 * 1.401298464324817e-45. The latter is subnormal, but it is finite and
742 * a number.
743 */
744 r = (struct ssa_result_range){ ge_zero, alu->op == nir_op_b2f32, true, true };
745 break;
746
747 case nir_op_bcsel: {
748 const struct ssa_result_range left = unpack_data(src_res[0]);
749 const struct ssa_result_range right = unpack_data(src_res[1]);
750
751 r.is_integral = left.is_integral && right.is_integral;
752
753 /* This could be better, but it would require a lot of work. For
754 * example, the result of the following is a number:
755 *
756 * bcsel(a > 0.0, a, 38.6)
757 *
758 * If the result of 'a > 0.0' is true, then the use of 'a' in the true
759 * part of the bcsel must be a number.
760 *
761 * Other cases are even more challenging.
762 *
763 * bcsel(a > 0.5, a - 0.5, 0.0)
764 */
765 r.is_a_number = left.is_a_number && right.is_a_number;
766 r.is_finite = left.is_finite && right.is_finite;
767
768 r.range = union_ranges(left.range, right.range);
769 break;
770 }
771
772 case nir_op_i2f32:
773 case nir_op_u2f32:
774 r = unpack_data(src_res[0]);
775
776 r.is_integral = true;
777 r.is_a_number = true;
778 r.is_finite = true;
779
780 if (r.range == unknown && alu->op == nir_op_u2f32)
781 r.range = ge_zero;
782
783 break;
784
785 case nir_op_fabs:
786 r = unpack_data(src_res[0]);
787
788 switch (r.range) {
789 case unknown:
790 case le_zero:
791 case ge_zero:
792 r.range = ge_zero;
793 break;
794
795 case lt_zero:
796 case gt_zero:
797 case ne_zero:
798 r.range = gt_zero;
799 break;
800
801 case eq_zero:
802 break;
803 }
804
805 break;
806
807 case nir_op_fadd: {
808 const struct ssa_result_range left = unpack_data(src_res[0]);
809 const struct ssa_result_range right = unpack_data(src_res[1]);
810
811 r.is_integral = left.is_integral && right.is_integral;
812 r.range = fadd_table[left.range][right.range];
813
814 /* X + Y is NaN if either operand is NaN or if one operand is +Inf and
815 * the other is -Inf. If neither operand is NaN and at least one of the
816 * operands is finite, then the result cannot be NaN.
817 */
818 r.is_a_number = left.is_a_number && right.is_a_number &&
819 (left.is_finite || right.is_finite);
820 break;
821 }
822
823 case nir_op_fexp2: {
824 /* If the parameter might be less than zero, the mathematically result
825 * will be on (0, 1). For sufficiently large magnitude negative
826 * parameters, the result will flush to zero.
827 */
828 static const enum ssa_ranges table[last_range + 1] = {
829 /* unknown lt_zero le_zero gt_zero ge_zero ne_zero eq_zero */
830 ge_zero, ge_zero, ge_zero, gt_zero, gt_zero, ge_zero, gt_zero
831 };
832
833 r = unpack_data(src_res[0]);
834
835 ASSERT_UNION_OF_DISJOINT_MATCHES_UNKNOWN_1_SOURCE(table);
836 ASSERT_UNION_OF_EQ_AND_STRICT_INEQ_MATCHES_NONSTRICT_1_SOURCE(table);
837
838 r.is_integral = r.is_integral && is_not_negative(r.range);
839 r.range = table[r.range];
840
841 /* Various cases can result in NaN, so assume the worst. */
842 r.is_finite = false;
843 r.is_a_number = false;
844 break;
845 }
846
847 case nir_op_fmax: {
848 const struct ssa_result_range left = unpack_data(src_res[0]);
849 const struct ssa_result_range right = unpack_data(src_res[1]);
850
851 r.is_integral = left.is_integral && right.is_integral;
852
853 /* This is conservative. It may be possible to determine that the
854 * result must be finite in more cases, but it would take some effort to
855 * work out all the corners. For example, fmax({lt_zero, finite},
856 * {lt_zero}) should result in {lt_zero, finite}.
857 */
858 r.is_finite = left.is_finite && right.is_finite;
859
860 /* If one source is NaN, fmax always picks the other source. */
861 r.is_a_number = left.is_a_number || right.is_a_number;
862
863 /* gt_zero: fmax(gt_zero, *)
864 * | fmax(*, gt_zero) # Treat fmax as commutative
865 * ;
866 *
867 * ge_zero: fmax(ge_zero, ne_zero)
868 * | fmax(ge_zero, lt_zero)
869 * | fmax(ge_zero, le_zero)
870 * | fmax(ge_zero, eq_zero)
871 * | fmax(ne_zero, ge_zero) # Treat fmax as commutative
872 * | fmax(lt_zero, ge_zero) # Treat fmax as commutative
873 * | fmax(le_zero, ge_zero) # Treat fmax as commutative
874 * | fmax(eq_zero, ge_zero) # Treat fmax as commutative
875 * | fmax(ge_zero, ge_zero)
876 * ;
877 *
878 * le_zero: fmax(le_zero, lt_zero)
879 * | fmax(lt_zero, le_zero) # Treat fmax as commutative
880 * | fmax(le_zero, le_zero)
881 * ;
882 *
883 * lt_zero: fmax(lt_zero, lt_zero)
884 * ;
885 *
886 * ne_zero: fmax(ne_zero, lt_zero)
887 * | fmax(lt_zero, ne_zero) # Treat fmax as commutative
888 * | fmax(ne_zero, ne_zero)
889 * ;
890 *
891 * eq_zero: fmax(eq_zero, le_zero)
892 * | fmax(eq_zero, lt_zero)
893 * | fmax(le_zero, eq_zero) # Treat fmax as commutative
894 * | fmax(lt_zero, eq_zero) # Treat fmax as commutative
895 * | fmax(eq_zero, eq_zero)
896 * ;
897 *
898 * All other cases are 'unknown'.
899 */
900 static const enum ssa_ranges table[last_range + 1][last_range + 1] = {
901 /* left\right unknown lt_zero le_zero gt_zero ge_zero ne_zero eq_zero */
902 /* unknown */ { _______, _______, _______, gt_zero, ge_zero, _______, ge_zero },
903 /* lt_zero */ { _______, lt_zero, le_zero, gt_zero, ge_zero, ne_zero, eq_zero },
904 /* le_zero */ { _______, le_zero, le_zero, gt_zero, ge_zero, _______, eq_zero },
905 /* gt_zero */ { gt_zero, gt_zero, gt_zero, gt_zero, gt_zero, gt_zero, gt_zero },
906 /* ge_zero */ { ge_zero, ge_zero, ge_zero, gt_zero, ge_zero, ge_zero, ge_zero },
907 /* ne_zero */ { _______, ne_zero, _______, gt_zero, ge_zero, ne_zero, ge_zero },
908 /* eq_zero */ { ge_zero, eq_zero, eq_zero, gt_zero, ge_zero, ge_zero, eq_zero }
909 };
910
911 /* Treat fmax as commutative. */
912 ASSERT_TABLE_IS_COMMUTATIVE(table);
913 ASSERT_TABLE_IS_DIAGONAL(table);
914 ASSERT_UNION_OF_OTHERS_MATCHES_UNKNOWN_2_SOURCE(table);
915
916 r.range = table[left.range][right.range];
917
918 /* Recall that when either value is NaN, fmax will pick the other value.
919 * This means the result range of the fmax will either be the "ideal"
920 * result range (calculated above) or the range of the non-NaN value.
921 */
922 if (!left.is_a_number)
923 r.range = union_ranges(r.range, right.range);
924
925 if (!right.is_a_number)
926 r.range = union_ranges(r.range, left.range);
927
928 break;
929 }
930
931 case nir_op_fmin: {
932 const struct ssa_result_range left = unpack_data(src_res[0]);
933 const struct ssa_result_range right = unpack_data(src_res[1]);
934
935 r.is_integral = left.is_integral && right.is_integral;
936
937 /* This is conservative. It may be possible to determine that the
938 * result must be finite in more cases, but it would take some effort to
939 * work out all the corners. For example, fmin({gt_zero, finite},
940 * {gt_zero}) should result in {gt_zero, finite}.
941 */
942 r.is_finite = left.is_finite && right.is_finite;
943
944 /* If one source is NaN, fmin always picks the other source. */
945 r.is_a_number = left.is_a_number || right.is_a_number;
946
947 /* lt_zero: fmin(lt_zero, *)
948 * | fmin(*, lt_zero) # Treat fmin as commutative
949 * ;
950 *
951 * le_zero: fmin(le_zero, ne_zero)
952 * | fmin(le_zero, gt_zero)
953 * | fmin(le_zero, ge_zero)
954 * | fmin(le_zero, eq_zero)
955 * | fmin(ne_zero, le_zero) # Treat fmin as commutative
956 * | fmin(gt_zero, le_zero) # Treat fmin as commutative
957 * | fmin(ge_zero, le_zero) # Treat fmin as commutative
958 * | fmin(eq_zero, le_zero) # Treat fmin as commutative
959 * | fmin(le_zero, le_zero)
960 * ;
961 *
962 * ge_zero: fmin(ge_zero, gt_zero)
963 * | fmin(gt_zero, ge_zero) # Treat fmin as commutative
964 * | fmin(ge_zero, ge_zero)
965 * ;
966 *
967 * gt_zero: fmin(gt_zero, gt_zero)
968 * ;
969 *
970 * ne_zero: fmin(ne_zero, gt_zero)
971 * | fmin(gt_zero, ne_zero) # Treat fmin as commutative
972 * | fmin(ne_zero, ne_zero)
973 * ;
974 *
975 * eq_zero: fmin(eq_zero, ge_zero)
976 * | fmin(eq_zero, gt_zero)
977 * | fmin(ge_zero, eq_zero) # Treat fmin as commutative
978 * | fmin(gt_zero, eq_zero) # Treat fmin as commutative
979 * | fmin(eq_zero, eq_zero)
980 * ;
981 *
982 * All other cases are 'unknown'.
983 */
984 static const enum ssa_ranges table[last_range + 1][last_range + 1] = {
985 /* left\right unknown lt_zero le_zero gt_zero ge_zero ne_zero eq_zero */
986 /* unknown */ { _______, lt_zero, le_zero, _______, _______, _______, le_zero },
987 /* lt_zero */ { lt_zero, lt_zero, lt_zero, lt_zero, lt_zero, lt_zero, lt_zero },
988 /* le_zero */ { le_zero, lt_zero, le_zero, le_zero, le_zero, le_zero, le_zero },
989 /* gt_zero */ { _______, lt_zero, le_zero, gt_zero, ge_zero, ne_zero, eq_zero },
990 /* ge_zero */ { _______, lt_zero, le_zero, ge_zero, ge_zero, _______, eq_zero },
991 /* ne_zero */ { _______, lt_zero, le_zero, ne_zero, _______, ne_zero, le_zero },
992 /* eq_zero */ { le_zero, lt_zero, le_zero, eq_zero, eq_zero, le_zero, eq_zero }
993 };
994
995 /* Treat fmin as commutative. */
996 ASSERT_TABLE_IS_COMMUTATIVE(table);
997 ASSERT_TABLE_IS_DIAGONAL(table);
998 ASSERT_UNION_OF_OTHERS_MATCHES_UNKNOWN_2_SOURCE(table);
999
1000 r.range = table[left.range][right.range];
1001
1002 /* Recall that when either value is NaN, fmin will pick the other value.
1003 * This means the result range of the fmin will either be the "ideal"
1004 * result range (calculated above) or the range of the non-NaN value.
1005 */
1006 if (!left.is_a_number)
1007 r.range = union_ranges(r.range, right.range);
1008
1009 if (!right.is_a_number)
1010 r.range = union_ranges(r.range, left.range);
1011
1012 break;
1013 }
1014
1015 case nir_op_fmul:
1016 case nir_op_fmulz: {
1017 const struct ssa_result_range left = unpack_data(src_res[0]);
1018 const struct ssa_result_range right = unpack_data(src_res[1]);
1019
1020 r.is_integral = left.is_integral && right.is_integral;
1021
1022 /* x * x => ge_zero */
1023 if (left.range != eq_zero && nir_alu_srcs_equal(alu, alu, 0, 1)) {
1024 /* Even if x > 0, the result of x*x can be zero when x is, for
1025 * example, a subnormal number.
1026 */
1027 r.range = ge_zero;
1028 } else if (left.range != eq_zero && nir_alu_srcs_negative_equal(alu, alu, 0, 1)) {
1029 /* -x * x => le_zero. */
1030 r.range = le_zero;
1031 } else
1032 r.range = fmul_table[left.range][right.range];
1033
1034 if (alu->op == nir_op_fmul) {
1035 /* Mulitpliation produces NaN for X * NaN and for 0 * ±Inf. If both
1036 * operands are numbers and either both are finite or one is finite and
1037 * the other cannot be zero, then the result must be a number.
1038 */
1039 r.is_a_number = (left.is_a_number && right.is_a_number) &&
1040 ((left.is_finite && right.is_finite) ||
1041 (!is_not_zero(left.range) && right.is_finite) ||
1042 (left.is_finite && !is_not_zero(right.range)));
1043 } else {
1044 /* nir_op_fmulz: unlike nir_op_fmul, 0 * ±Inf is a number. */
1045 r.is_a_number = left.is_a_number && right.is_a_number;
1046 }
1047
1048 break;
1049 }
1050
1051 case nir_op_frcp:
1052 r = (struct ssa_result_range){
1053 unpack_data(src_res[0]).range,
1054 false,
1055 false, /* Various cases can result in NaN, so assume the worst. */
1056 false /* " " " " " " " " " " */
1057 };
1058 break;
1059
1060 case nir_op_mov:
1061 r = unpack_data(src_res[0]);
1062 break;
1063
1064 case nir_op_fneg:
1065 r = unpack_data(src_res[0]);
1066 r.range = fneg_table[r.range];
1067 break;
1068
1069 case nir_op_fsat: {
1070 const struct ssa_result_range left = unpack_data(src_res[0]);
1071
1072 /* fsat(NaN) = 0. */
1073 r.is_a_number = true;
1074 r.is_finite = true;
1075
1076 switch (left.range) {
1077 case le_zero:
1078 case lt_zero:
1079 case eq_zero:
1080 r.range = eq_zero;
1081 r.is_integral = true;
1082 break;
1083
1084 case gt_zero:
1085 /* fsat is equivalent to fmin(fmax(X, 0.0), 1.0), so if X is not a
1086 * number, the result will be 0.
1087 */
1088 r.range = left.is_a_number ? gt_zero : ge_zero;
1089 r.is_integral = left.is_integral;
1090 break;
1091
1092 case ge_zero:
1093 case ne_zero:
1094 case unknown:
1095 /* Since the result must be in [0, 1], the value must be >= 0. */
1096 r.range = ge_zero;
1097 r.is_integral = left.is_integral;
1098 break;
1099 }
1100 break;
1101 }
1102
1103 case nir_op_fsign:
1104 r = (struct ssa_result_range){
1105 unpack_data(src_res[0]).range,
1106 true,
1107 true, /* fsign is -1, 0, or 1, even for NaN, so it must be a number. */
1108 true /* fsign is -1, 0, or 1, even for NaN, so it must be finite. */
1109 };
1110 break;
1111
1112 case nir_op_fsqrt:
1113 case nir_op_frsq:
1114 r = (struct ssa_result_range){ ge_zero, false, false, false };
1115 break;
1116
1117 case nir_op_ffloor: {
1118 const struct ssa_result_range left = unpack_data(src_res[0]);
1119
1120 r.is_integral = true;
1121
1122 /* In IEEE 754, floor(NaN) is NaN, and floor(±Inf) is ±Inf. See
1123 * https://pubs.opengroup.org/onlinepubs/9699919799.2016edition/functions/floor.html
1124 */
1125 r.is_a_number = left.is_a_number;
1126 r.is_finite = left.is_finite;
1127
1128 if (left.is_integral || left.range == le_zero || left.range == lt_zero)
1129 r.range = left.range;
1130 else if (left.range == ge_zero || left.range == gt_zero)
1131 r.range = ge_zero;
1132 else if (left.range == ne_zero)
1133 r.range = unknown;
1134
1135 break;
1136 }
1137
1138 case nir_op_fceil: {
1139 const struct ssa_result_range left = unpack_data(src_res[0]);
1140
1141 r.is_integral = true;
1142
1143 /* In IEEE 754, ceil(NaN) is NaN, and ceil(±Inf) is ±Inf. See
1144 * https://pubs.opengroup.org/onlinepubs/9699919799.2016edition/functions/ceil.html
1145 */
1146 r.is_a_number = left.is_a_number;
1147 r.is_finite = left.is_finite;
1148
1149 if (left.is_integral || left.range == ge_zero || left.range == gt_zero)
1150 r.range = left.range;
1151 else if (left.range == le_zero || left.range == lt_zero)
1152 r.range = le_zero;
1153 else if (left.range == ne_zero)
1154 r.range = unknown;
1155
1156 break;
1157 }
1158
1159 case nir_op_ftrunc: {
1160 const struct ssa_result_range left = unpack_data(src_res[0]);
1161
1162 r.is_integral = true;
1163
1164 /* In IEEE 754, trunc(NaN) is NaN, and trunc(±Inf) is ±Inf. See
1165 * https://pubs.opengroup.org/onlinepubs/9699919799.2016edition/functions/trunc.html
1166 */
1167 r.is_a_number = left.is_a_number;
1168 r.is_finite = left.is_finite;
1169
1170 if (left.is_integral)
1171 r.range = left.range;
1172 else if (left.range == ge_zero || left.range == gt_zero)
1173 r.range = ge_zero;
1174 else if (left.range == le_zero || left.range == lt_zero)
1175 r.range = le_zero;
1176 else if (left.range == ne_zero)
1177 r.range = unknown;
1178
1179 break;
1180 }
1181
1182 case nir_op_flt:
1183 case nir_op_fge:
1184 case nir_op_feq:
1185 case nir_op_fneu:
1186 case nir_op_ilt:
1187 case nir_op_ige:
1188 case nir_op_ieq:
1189 case nir_op_ine:
1190 case nir_op_ult:
1191 case nir_op_uge:
1192 /* Boolean results are 0 or -1. */
1193 r = (struct ssa_result_range){ le_zero, false, true, false };
1194 break;
1195
1196 case nir_op_fdot2:
1197 case nir_op_fdot3:
1198 case nir_op_fdot4:
1199 case nir_op_fdot8:
1200 case nir_op_fdot16:
1201 case nir_op_fdot2_replicated:
1202 case nir_op_fdot3_replicated:
1203 case nir_op_fdot4_replicated:
1204 case nir_op_fdot8_replicated:
1205 case nir_op_fdot16_replicated: {
1206 const struct ssa_result_range left = unpack_data(src_res[0]);
1207
1208 /* If the two sources are the same SSA value, then the result is either
1209 * NaN or some number >= 0. If one source is the negation of the other,
1210 * the result is either NaN or some number <= 0.
1211 *
1212 * In either of these two cases, if one source is a number, then the
1213 * other must also be a number. Since it should not be possible to get
1214 * Inf-Inf in the dot-product, the result must also be a number.
1215 */
1216 if (nir_alu_srcs_equal(alu, alu, 0, 1)) {
1217 r = (struct ssa_result_range){ ge_zero, false, left.is_a_number, false };
1218 } else if (nir_alu_srcs_negative_equal(alu, alu, 0, 1)) {
1219 r = (struct ssa_result_range){ le_zero, false, left.is_a_number, false };
1220 } else {
1221 r = (struct ssa_result_range){ unknown, false, false, false };
1222 }
1223 break;
1224 }
1225
1226 case nir_op_fpow: {
1227 /* Due to flush-to-zero semanatics of floating-point numbers with very
1228 * small mangnitudes, we can never really be sure a result will be
1229 * non-zero.
1230 *
1231 * NIR uses pow() and powf() to constant evaluate nir_op_fpow. The man
1232 * page for that function says:
1233 *
1234 * If y is 0, the result is 1.0 (even if x is a NaN).
1235 *
1236 * gt_zero: pow(*, eq_zero)
1237 * | pow(eq_zero, lt_zero) # 0^-y = +inf
1238 * | pow(eq_zero, le_zero) # 0^-y = +inf or 0^0 = 1.0
1239 * ;
1240 *
1241 * eq_zero: pow(eq_zero, gt_zero)
1242 * ;
1243 *
1244 * ge_zero: pow(gt_zero, gt_zero)
1245 * | pow(gt_zero, ge_zero)
1246 * | pow(gt_zero, lt_zero)
1247 * | pow(gt_zero, le_zero)
1248 * | pow(gt_zero, ne_zero)
1249 * | pow(gt_zero, unknown)
1250 * | pow(ge_zero, gt_zero)
1251 * | pow(ge_zero, ge_zero)
1252 * | pow(ge_zero, lt_zero)
1253 * | pow(ge_zero, le_zero)
1254 * | pow(ge_zero, ne_zero)
1255 * | pow(ge_zero, unknown)
1256 * | pow(eq_zero, ge_zero) # 0^0 = 1.0 or 0^+y = 0.0
1257 * | pow(eq_zero, ne_zero) # 0^-y = +inf or 0^+y = 0.0
1258 * | pow(eq_zero, unknown) # union of all other y cases
1259 * ;
1260 *
1261 * All other cases are unknown.
1262 *
1263 * We could do better if the right operand is a constant, integral
1264 * value.
1265 */
1266 static const enum ssa_ranges table[last_range + 1][last_range + 1] = {
1267 /* left\right unknown lt_zero le_zero gt_zero ge_zero ne_zero eq_zero */
1268 /* unknown */ { _______, _______, _______, _______, _______, _______, gt_zero },
1269 /* lt_zero */ { _______, _______, _______, _______, _______, _______, gt_zero },
1270 /* le_zero */ { _______, _______, _______, _______, _______, _______, gt_zero },
1271 /* gt_zero */ { ge_zero, ge_zero, ge_zero, ge_zero, ge_zero, ge_zero, gt_zero },
1272 /* ge_zero */ { ge_zero, ge_zero, ge_zero, ge_zero, ge_zero, ge_zero, gt_zero },
1273 /* ne_zero */ { _______, _______, _______, _______, _______, _______, gt_zero },
1274 /* eq_zero */ { ge_zero, gt_zero, gt_zero, eq_zero, ge_zero, ge_zero, gt_zero },
1275 };
1276
1277 const struct ssa_result_range left = unpack_data(src_res[0]);
1278 const struct ssa_result_range right = unpack_data(src_res[1]);
1279
1280 ASSERT_UNION_OF_DISJOINT_MATCHES_UNKNOWN_2_SOURCE(table);
1281 ASSERT_UNION_OF_EQ_AND_STRICT_INEQ_MATCHES_NONSTRICT_2_SOURCE(table);
1282
1283 r.is_integral = left.is_integral && right.is_integral &&
1284 is_not_negative(right.range);
1285 r.range = table[left.range][right.range];
1286
1287 /* Various cases can result in NaN, so assume the worst. */
1288 r.is_a_number = false;
1289
1290 break;
1291 }
1292
1293 case nir_op_ffma: {
1294 const struct ssa_result_range first = unpack_data(src_res[0]);
1295 const struct ssa_result_range second = unpack_data(src_res[1]);
1296 const struct ssa_result_range third = unpack_data(src_res[2]);
1297
1298 r.is_integral = first.is_integral && second.is_integral &&
1299 third.is_integral;
1300
1301 /* Various cases can result in NaN, so assume the worst. */
1302 r.is_a_number = false;
1303
1304 enum ssa_ranges fmul_range;
1305
1306 if (first.range != eq_zero && nir_alu_srcs_equal(alu, alu, 0, 1)) {
1307 /* See handling of nir_op_fmul for explanation of why ge_zero is the
1308 * range.
1309 */
1310 fmul_range = ge_zero;
1311 } else if (first.range != eq_zero && nir_alu_srcs_negative_equal(alu, alu, 0, 1)) {
1312 /* -x * x => le_zero */
1313 fmul_range = le_zero;
1314 } else
1315 fmul_range = fmul_table[first.range][second.range];
1316
1317 r.range = fadd_table[fmul_range][third.range];
1318 break;
1319 }
1320
1321 case nir_op_flrp: {
1322 const struct ssa_result_range first = unpack_data(src_res[0]);
1323 const struct ssa_result_range second = unpack_data(src_res[1]);
1324 const struct ssa_result_range third = unpack_data(src_res[2]);
1325
1326 r.is_integral = first.is_integral && second.is_integral &&
1327 third.is_integral;
1328
1329 /* Various cases can result in NaN, so assume the worst. */
1330 r.is_a_number = false;
1331
1332 /* Decompose the flrp to first + third * (second + -first) */
1333 const enum ssa_ranges inner_fadd_range =
1334 fadd_table[second.range][fneg_table[first.range]];
1335
1336 const enum ssa_ranges fmul_range =
1337 fmul_table[third.range][inner_fadd_range];
1338
1339 r.range = fadd_table[first.range][fmul_range];
1340 break;
1341 }
1342
1343 default:
1344 r = (struct ssa_result_range){ unknown, false, false, false };
1345 break;
1346 }
1347
1348 if (r.range == eq_zero)
1349 r.is_integral = true;
1350
1351 /* Just like isfinite(), the is_finite flag implies the value is a number. */
1352 assert((int)r.is_finite <= (int)r.is_a_number);
1353
1354 *result = pack_data(r);
1355 }
1356
1357 #undef _______
1358
1359 struct ssa_result_range
nir_analyze_range(struct hash_table * range_ht,const nir_alu_instr * alu,unsigned src)1360 nir_analyze_range(struct hash_table *range_ht,
1361 const nir_alu_instr *alu, unsigned src)
1362 {
1363 struct fp_query query_alloc[64];
1364 uint32_t result_alloc[64];
1365
1366 struct analysis_state state;
1367 state.range_ht = range_ht;
1368 util_dynarray_init_from_stack(&state.query_stack, query_alloc, sizeof(query_alloc));
1369 util_dynarray_init_from_stack(&state.result_stack, result_alloc, sizeof(result_alloc));
1370 state.query_size = sizeof(struct fp_query);
1371 state.get_key = &get_fp_key;
1372 state.process_query = &process_fp_query;
1373
1374 push_fp_query(&state, alu, src, nir_type_invalid);
1375
1376 return unpack_data(perform_analysis(&state));
1377 }
1378
1379 static uint32_t
bitmask(uint32_t size)1380 bitmask(uint32_t size)
1381 {
1382 return size >= 32 ? 0xffffffffu : ((uint32_t)1 << size) - 1u;
1383 }
1384
1385 static uint64_t
mul_clamp(uint32_t a,uint32_t b)1386 mul_clamp(uint32_t a, uint32_t b)
1387 {
1388 if (a != 0 && (a * b) / a != b)
1389 return (uint64_t)UINT32_MAX + 1;
1390 else
1391 return a * b;
1392 }
1393
1394 /* recursively gather at most "buf_size" phi/bcsel sources */
1395 static unsigned
search_phi_bcsel(nir_scalar scalar,nir_scalar * buf,unsigned buf_size,struct set * visited)1396 search_phi_bcsel(nir_scalar scalar, nir_scalar *buf, unsigned buf_size, struct set *visited)
1397 {
1398 if (_mesa_set_search(visited, scalar.def))
1399 return 0;
1400 _mesa_set_add(visited, scalar.def);
1401
1402 if (scalar.def->parent_instr->type == nir_instr_type_phi) {
1403 nir_phi_instr *phi = nir_instr_as_phi(scalar.def->parent_instr);
1404 unsigned num_sources_left = exec_list_length(&phi->srcs);
1405 if (buf_size >= num_sources_left) {
1406 unsigned total_added = 0;
1407 nir_foreach_phi_src(src, phi) {
1408 num_sources_left--;
1409 unsigned added = search_phi_bcsel(nir_get_scalar(src->src.ssa, scalar.comp),
1410 buf + total_added, buf_size - num_sources_left, visited);
1411 assert(added <= buf_size);
1412 buf_size -= added;
1413 total_added += added;
1414 }
1415 return total_added;
1416 }
1417 }
1418
1419 if (nir_scalar_is_alu(scalar)) {
1420 nir_op op = nir_scalar_alu_op(scalar);
1421
1422 if ((op == nir_op_bcsel || op == nir_op_b32csel) && buf_size >= 2) {
1423 nir_scalar src1 = nir_scalar_chase_alu_src(scalar, 1);
1424 nir_scalar src2 = nir_scalar_chase_alu_src(scalar, 2);
1425
1426 unsigned added = search_phi_bcsel(src1, buf, buf_size - 1, visited);
1427 buf_size -= added;
1428 added += search_phi_bcsel(src2, buf + added, buf_size, visited);
1429 return added;
1430 }
1431 }
1432
1433 buf[0] = scalar;
1434 return 1;
1435 }
1436
1437 static nir_variable *
lookup_input(nir_shader * shader,unsigned driver_location)1438 lookup_input(nir_shader *shader, unsigned driver_location)
1439 {
1440 return nir_find_variable_with_driver_location(shader, nir_var_shader_in,
1441 driver_location);
1442 }
1443
1444 /* The config here should be generic enough to be correct on any HW. */
1445 static const nir_unsigned_upper_bound_config default_ub_config = {
1446 .min_subgroup_size = 1u,
1447 .max_subgroup_size = UINT16_MAX,
1448 .max_workgroup_invocations = UINT16_MAX,
1449
1450 /* max_workgroup_count represents the maximum compute shader / kernel
1451 * dispatchable work size. On most hardware, this is essentially
1452 * unbounded. On some hardware max_workgroup_count[1] and
1453 * max_workgroup_count[2] may be smaller.
1454 */
1455 .max_workgroup_count = { UINT32_MAX, UINT32_MAX, UINT32_MAX },
1456
1457 /* max_workgroup_size is the local invocation maximum. This is generally
1458 * small the OpenGL 4.2 minimum maximum is 1024.
1459 */
1460 .max_workgroup_size = { UINT16_MAX, UINT16_MAX, UINT16_MAX },
1461
1462 .vertex_attrib_max = {
1463 UINT32_MAX,
1464 UINT32_MAX,
1465 UINT32_MAX,
1466 UINT32_MAX,
1467 UINT32_MAX,
1468 UINT32_MAX,
1469 UINT32_MAX,
1470 UINT32_MAX,
1471 UINT32_MAX,
1472 UINT32_MAX,
1473 UINT32_MAX,
1474 UINT32_MAX,
1475 UINT32_MAX,
1476 UINT32_MAX,
1477 UINT32_MAX,
1478 UINT32_MAX,
1479 UINT32_MAX,
1480 UINT32_MAX,
1481 UINT32_MAX,
1482 UINT32_MAX,
1483 UINT32_MAX,
1484 UINT32_MAX,
1485 UINT32_MAX,
1486 UINT32_MAX,
1487 UINT32_MAX,
1488 UINT32_MAX,
1489 UINT32_MAX,
1490 UINT32_MAX,
1491 UINT32_MAX,
1492 UINT32_MAX,
1493 UINT32_MAX,
1494 UINT32_MAX,
1495 },
1496 };
1497
1498 struct uub_query {
1499 struct analysis_query head;
1500 nir_scalar scalar;
1501 };
1502
1503 static void
push_uub_query(struct analysis_state * state,nir_scalar scalar)1504 push_uub_query(struct analysis_state *state, nir_scalar scalar)
1505 {
1506 struct uub_query *pushed_q = push_analysis_query(state, sizeof(struct uub_query));
1507 pushed_q->scalar = scalar;
1508 }
1509
1510 static uintptr_t
get_uub_key(struct analysis_query * q)1511 get_uub_key(struct analysis_query *q)
1512 {
1513 nir_scalar scalar = ((struct uub_query *)q)->scalar;
1514 /* keys can't be 0, so we have to add 1 to the index */
1515 unsigned shift_amount = ffs(NIR_MAX_VEC_COMPONENTS) - 1;
1516 return nir_scalar_is_const(scalar)
1517 ? 0
1518 : ((uintptr_t)(scalar.def->index + 1) << shift_amount) | scalar.comp;
1519 }
1520
1521 static void
get_intrinsic_uub(struct analysis_state * state,struct uub_query q,uint32_t * result,const uint32_t * src)1522 get_intrinsic_uub(struct analysis_state *state, struct uub_query q, uint32_t *result,
1523 const uint32_t *src)
1524 {
1525 nir_shader *shader = state->shader;
1526 const nir_unsigned_upper_bound_config *config = state->config;
1527
1528 nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(q.scalar.def->parent_instr);
1529 switch (intrin->intrinsic) {
1530 case nir_intrinsic_load_local_invocation_index:
1531 /* The local invocation index is used under the hood by RADV for
1532 * some non-compute-like shaders (eg. LS and NGG). These technically
1533 * run in workgroups on the HW, even though this fact is not exposed
1534 * by the API.
1535 * They can safely use the same code path here as variable sized
1536 * compute-like shader stages.
1537 */
1538 if (!gl_shader_stage_uses_workgroup(shader->info.stage) ||
1539 shader->info.workgroup_size_variable) {
1540 *result = config->max_workgroup_invocations - 1;
1541 } else {
1542 *result = (shader->info.workgroup_size[0] *
1543 shader->info.workgroup_size[1] *
1544 shader->info.workgroup_size[2]) -
1545 1u;
1546 }
1547 break;
1548 case nir_intrinsic_load_local_invocation_id:
1549 if (shader->info.workgroup_size_variable)
1550 *result = config->max_workgroup_size[q.scalar.comp] - 1u;
1551 else
1552 *result = shader->info.workgroup_size[q.scalar.comp] - 1u;
1553 break;
1554 case nir_intrinsic_load_workgroup_id:
1555 *result = config->max_workgroup_count[q.scalar.comp] - 1u;
1556 break;
1557 case nir_intrinsic_load_num_workgroups:
1558 *result = config->max_workgroup_count[q.scalar.comp];
1559 break;
1560 case nir_intrinsic_load_global_invocation_id:
1561 if (shader->info.workgroup_size_variable) {
1562 *result = mul_clamp(config->max_workgroup_size[q.scalar.comp],
1563 config->max_workgroup_count[q.scalar.comp]) -
1564 1u;
1565 } else {
1566 *result = (shader->info.workgroup_size[q.scalar.comp] *
1567 config->max_workgroup_count[q.scalar.comp]) -
1568 1u;
1569 }
1570 break;
1571 case nir_intrinsic_load_invocation_id:
1572 if (shader->info.stage == MESA_SHADER_TESS_CTRL)
1573 *result = shader->info.tess.tcs_vertices_out
1574 ? (shader->info.tess.tcs_vertices_out - 1)
1575 : 511; /* Generous maximum output patch size of 512 */
1576 break;
1577 case nir_intrinsic_load_subgroup_invocation:
1578 case nir_intrinsic_first_invocation:
1579 *result = config->max_subgroup_size - 1;
1580 break;
1581 case nir_intrinsic_mbcnt_amd: {
1582 if (!q.head.pushed_queries) {
1583 push_uub_query(state, nir_get_scalar(intrin->src[1].ssa, 0));
1584 return;
1585 } else {
1586 uint32_t src0 = config->max_subgroup_size - 1;
1587 uint32_t src1 = src[0];
1588 if (src0 + src1 >= src0) /* check overflow */
1589 *result = src0 + src1;
1590 }
1591 break;
1592 }
1593 case nir_intrinsic_load_subgroup_size:
1594 *result = config->max_subgroup_size;
1595 break;
1596 case nir_intrinsic_load_subgroup_id:
1597 case nir_intrinsic_load_num_subgroups: {
1598 uint32_t workgroup_size = config->max_workgroup_invocations;
1599 if (gl_shader_stage_uses_workgroup(shader->info.stage) &&
1600 !shader->info.workgroup_size_variable) {
1601 workgroup_size = shader->info.workgroup_size[0] *
1602 shader->info.workgroup_size[1] *
1603 shader->info.workgroup_size[2];
1604 }
1605 *result = DIV_ROUND_UP(workgroup_size, config->min_subgroup_size);
1606 if (intrin->intrinsic == nir_intrinsic_load_subgroup_id)
1607 (*result)--;
1608 break;
1609 }
1610 case nir_intrinsic_load_input: {
1611 if (shader->info.stage == MESA_SHADER_VERTEX && nir_src_is_const(intrin->src[0])) {
1612 nir_variable *var = lookup_input(shader, nir_intrinsic_base(intrin));
1613 if (var) {
1614 int loc = var->data.location - VERT_ATTRIB_GENERIC0;
1615 if (loc >= 0)
1616 *result = config->vertex_attrib_max[loc];
1617 }
1618 }
1619 break;
1620 }
1621 case nir_intrinsic_reduce:
1622 case nir_intrinsic_inclusive_scan:
1623 case nir_intrinsic_exclusive_scan: {
1624 nir_op op = nir_intrinsic_reduction_op(intrin);
1625 if (op == nir_op_umin || op == nir_op_umax || op == nir_op_imin || op == nir_op_imax) {
1626 if (!q.head.pushed_queries) {
1627 push_uub_query(state, nir_get_scalar(intrin->src[0].ssa, q.scalar.comp));
1628 return;
1629 } else {
1630 *result = src[0];
1631 }
1632 }
1633 break;
1634 }
1635 case nir_intrinsic_read_first_invocation:
1636 case nir_intrinsic_read_invocation:
1637 case nir_intrinsic_shuffle:
1638 case nir_intrinsic_shuffle_xor:
1639 case nir_intrinsic_shuffle_up:
1640 case nir_intrinsic_shuffle_down:
1641 case nir_intrinsic_quad_broadcast:
1642 case nir_intrinsic_quad_swap_horizontal:
1643 case nir_intrinsic_quad_swap_vertical:
1644 case nir_intrinsic_quad_swap_diagonal:
1645 case nir_intrinsic_quad_swizzle_amd:
1646 case nir_intrinsic_masked_swizzle_amd:
1647 if (!q.head.pushed_queries) {
1648 push_uub_query(state, nir_get_scalar(intrin->src[0].ssa, q.scalar.comp));
1649 return;
1650 } else {
1651 *result = src[0];
1652 }
1653 break;
1654 case nir_intrinsic_write_invocation_amd:
1655 if (!q.head.pushed_queries) {
1656 push_uub_query(state, nir_get_scalar(intrin->src[0].ssa, q.scalar.comp));
1657 push_uub_query(state, nir_get_scalar(intrin->src[1].ssa, q.scalar.comp));
1658 return;
1659 } else {
1660 *result = MAX2(src[0], src[1]);
1661 }
1662 break;
1663 case nir_intrinsic_load_tess_rel_patch_id_amd:
1664 case nir_intrinsic_load_tcs_num_patches_amd:
1665 /* Very generous maximum: TCS/TES executed by largest possible workgroup */
1666 *result = config->max_workgroup_invocations / MAX2(shader->info.tess.tcs_vertices_out, 1u);
1667 break;
1668 case nir_intrinsic_load_typed_buffer_amd: {
1669 const enum pipe_format format = nir_intrinsic_format(intrin);
1670 if (format == PIPE_FORMAT_NONE)
1671 break;
1672
1673 const struct util_format_description *desc = util_format_description(format);
1674 if (desc->channel[q.scalar.comp].type != UTIL_FORMAT_TYPE_UNSIGNED)
1675 break;
1676
1677 if (desc->channel[q.scalar.comp].normalized) {
1678 *result = fui(1.0);
1679 break;
1680 }
1681
1682 const uint32_t chan_max = u_uintN_max(desc->channel[q.scalar.comp].size);
1683 *result = desc->channel[q.scalar.comp].pure_integer ? chan_max : fui(chan_max);
1684 break;
1685 }
1686 case nir_intrinsic_load_scalar_arg_amd:
1687 case nir_intrinsic_load_vector_arg_amd: {
1688 uint32_t upper_bound = nir_intrinsic_arg_upper_bound_u32_amd(intrin);
1689 if (upper_bound)
1690 *result = upper_bound;
1691 break;
1692 }
1693 default:
1694 break;
1695 }
1696 }
1697
1698 static void
get_alu_uub(struct analysis_state * state,struct uub_query q,uint32_t * result,const uint32_t * src)1699 get_alu_uub(struct analysis_state *state, struct uub_query q, uint32_t *result, const uint32_t *src)
1700 {
1701 nir_op op = nir_scalar_alu_op(q.scalar);
1702
1703 /* Early exit for unsupported ALU opcodes. */
1704 switch (op) {
1705 case nir_op_umin:
1706 case nir_op_imin:
1707 case nir_op_imax:
1708 case nir_op_umax:
1709 case nir_op_iand:
1710 case nir_op_ior:
1711 case nir_op_ixor:
1712 case nir_op_ishl:
1713 case nir_op_imul:
1714 case nir_op_ushr:
1715 case nir_op_ishr:
1716 case nir_op_iadd:
1717 case nir_op_umod:
1718 case nir_op_udiv:
1719 case nir_op_bcsel:
1720 case nir_op_b32csel:
1721 case nir_op_ubfe:
1722 case nir_op_bfm:
1723 case nir_op_fmul:
1724 case nir_op_fmulz:
1725 case nir_op_extract_u8:
1726 case nir_op_extract_i8:
1727 case nir_op_extract_u16:
1728 case nir_op_extract_i16:
1729 case nir_op_b2i8:
1730 case nir_op_b2i16:
1731 case nir_op_b2i32:
1732 break;
1733 case nir_op_u2u1:
1734 case nir_op_u2u8:
1735 case nir_op_u2u16:
1736 case nir_op_u2u32:
1737 case nir_op_f2u32:
1738 if (nir_scalar_chase_alu_src(q.scalar, 0).def->bit_size > 32) {
1739 /* If src is >32 bits, return max */
1740 return;
1741 }
1742 break;
1743 default:
1744 return;
1745 }
1746
1747 if (!q.head.pushed_queries) {
1748 for (unsigned i = 0; i < nir_op_infos[op].num_inputs; i++)
1749 push_uub_query(state, nir_scalar_chase_alu_src(q.scalar, i));
1750 return;
1751 }
1752
1753 uint32_t max = bitmask(q.scalar.def->bit_size);
1754 switch (op) {
1755 case nir_op_umin:
1756 *result = src[0] < src[1] ? src[0] : src[1];
1757 break;
1758 case nir_op_imin:
1759 case nir_op_imax:
1760 case nir_op_umax:
1761 *result = src[0] > src[1] ? src[0] : src[1];
1762 break;
1763 case nir_op_iand:
1764 *result = bitmask(util_last_bit64(src[0])) & bitmask(util_last_bit64(src[1]));
1765 break;
1766 case nir_op_ior:
1767 case nir_op_ixor:
1768 *result = bitmask(util_last_bit64(src[0])) | bitmask(util_last_bit64(src[1]));
1769 break;
1770 case nir_op_ishl: {
1771 uint32_t src1 = MIN2(src[1], q.scalar.def->bit_size - 1u);
1772 if (util_last_bit64(src[0]) + src1 <= q.scalar.def->bit_size) /* check overflow */
1773 *result = src[0] << src1;
1774 break;
1775 }
1776 case nir_op_imul:
1777 if (src[0] == 0 || (src[0] * src[1]) / src[0] == src[1]) /* check overflow */
1778 *result = src[0] * src[1];
1779 break;
1780 case nir_op_ushr: {
1781 nir_scalar src1_scalar = nir_scalar_chase_alu_src(q.scalar, 1);
1782 uint32_t mask = q.scalar.def->bit_size - 1u;
1783 if (nir_scalar_is_const(src1_scalar))
1784 *result = src[0] >> (nir_scalar_as_uint(src1_scalar) & mask);
1785 else
1786 *result = src[0];
1787 break;
1788 }
1789 case nir_op_ishr: {
1790 nir_scalar src1_scalar = nir_scalar_chase_alu_src(q.scalar, 1);
1791 uint32_t mask = q.scalar.def->bit_size - 1u;
1792 if (src[0] <= 2147483647 && nir_scalar_is_const(src1_scalar))
1793 *result = src[0] >> (nir_scalar_as_uint(src1_scalar) & mask);
1794 else
1795 *result = src[0];
1796 break;
1797 }
1798 case nir_op_iadd:
1799 if (src[0] + src[1] >= src[0]) /* check overflow */
1800 *result = src[0] + src[1];
1801 break;
1802 case nir_op_umod:
1803 *result = src[1] ? src[1] - 1 : 0;
1804 break;
1805 case nir_op_udiv: {
1806 nir_scalar src1_scalar = nir_scalar_chase_alu_src(q.scalar, 1);
1807 if (nir_scalar_is_const(src1_scalar))
1808 *result = nir_scalar_as_uint(src1_scalar)
1809 ? src[0] / nir_scalar_as_uint(src1_scalar)
1810 : 0;
1811 else
1812 *result = src[0];
1813 break;
1814 }
1815 case nir_op_bcsel:
1816 case nir_op_b32csel:
1817 *result = src[1] > src[2] ? src[1] : src[2];
1818 break;
1819 case nir_op_ubfe:
1820 *result = bitmask(MIN2(src[2], q.scalar.def->bit_size));
1821 break;
1822 case nir_op_bfm: {
1823 nir_scalar src1_scalar = nir_scalar_chase_alu_src(q.scalar, 1);
1824 if (nir_scalar_is_const(src1_scalar)) {
1825 uint32_t src0 = MIN2(src[0], 31);
1826 uint32_t src1 = nir_scalar_as_uint(src1_scalar) & 0x1fu;
1827 *result = bitmask(src0) << src1;
1828 } else {
1829 uint32_t src0 = MIN2(src[0], 31);
1830 uint32_t src1 = MIN2(src[1], 31);
1831 *result = bitmask(MIN2(src0 + src1, 32));
1832 }
1833 break;
1834 }
1835 /* limited floating-point support for f2u32(fmul(load_input(), <constant>)) */
1836 case nir_op_f2u32:
1837 /* infinity/NaN starts at 0x7f800000u, negative numbers at 0x80000000 */
1838 if (src[0] < 0x7f800000u) {
1839 float val;
1840 memcpy(&val, &src[0], 4);
1841 *result = (uint32_t)val;
1842 }
1843 break;
1844 case nir_op_fmul:
1845 case nir_op_fmulz:
1846 /* infinity/NaN starts at 0x7f800000u, negative numbers at 0x80000000 */
1847 if (src[0] < 0x7f800000u && src[1] < 0x7f800000u) {
1848 float src0_f, src1_f;
1849 memcpy(&src0_f, &src[0], 4);
1850 memcpy(&src1_f, &src[1], 4);
1851 /* not a proper rounding-up multiplication, but should be good enough */
1852 float max_f = ceilf(src0_f) * ceilf(src1_f);
1853 memcpy(result, &max_f, 4);
1854 }
1855 break;
1856 case nir_op_u2u1:
1857 case nir_op_u2u8:
1858 case nir_op_u2u16:
1859 case nir_op_u2u32:
1860 *result = MIN2(src[0], max);
1861 break;
1862 case nir_op_b2i8:
1863 case nir_op_b2i16:
1864 case nir_op_b2i32:
1865 *result = 1;
1866 break;
1867 case nir_op_msad_4x8:
1868 *result = MIN2((uint64_t)src[2] + 4 * 255, UINT32_MAX);
1869 break;
1870 case nir_op_extract_u8:
1871 *result = MIN2(src[0], UINT8_MAX);
1872 break;
1873 case nir_op_extract_i8:
1874 *result = (src[0] >= 0x80) ? max : MIN2(src[0], INT8_MAX);
1875 break;
1876 case nir_op_extract_u16:
1877 *result = MIN2(src[0], UINT16_MAX);
1878 break;
1879 case nir_op_extract_i16:
1880 *result = (src[0] >= 0x8000) ? max : MIN2(src[0], INT16_MAX);
1881 break;
1882 default:
1883 break;
1884 }
1885 }
1886
1887 static void
get_phi_uub(struct analysis_state * state,struct uub_query q,uint32_t * result,const uint32_t * src)1888 get_phi_uub(struct analysis_state *state, struct uub_query q, uint32_t *result, const uint32_t *src)
1889 {
1890 nir_phi_instr *phi = nir_instr_as_phi(q.scalar.def->parent_instr);
1891
1892 if (exec_list_is_empty(&phi->srcs))
1893 return;
1894
1895 if (q.head.pushed_queries) {
1896 *result = src[0];
1897 for (unsigned i = 1; i < q.head.pushed_queries; i++)
1898 *result = MAX2(*result, src[i]);
1899 return;
1900 }
1901
1902 nir_cf_node *prev = nir_cf_node_prev(&phi->instr.block->cf_node);
1903 if (!prev || prev->type == nir_cf_node_block) {
1904 /* Resolve cycles by inserting max into range_ht. */
1905 uint32_t max = bitmask(q.scalar.def->bit_size);
1906 _mesa_hash_table_insert(state->range_ht, (void *)get_uub_key(&q.head), (void *)(uintptr_t)max);
1907
1908 struct set *visited = _mesa_pointer_set_create(NULL);
1909 nir_scalar *defs = alloca(sizeof(nir_scalar) * 64);
1910 unsigned def_count = search_phi_bcsel(q.scalar, defs, 64, visited);
1911 _mesa_set_destroy(visited, NULL);
1912
1913 for (unsigned i = 0; i < def_count; i++)
1914 push_uub_query(state, defs[i]);
1915 } else {
1916 nir_foreach_phi_src(src, phi)
1917 push_uub_query(state, nir_get_scalar(src->src.ssa, q.scalar.comp));
1918 }
1919 }
1920
1921 static void
process_uub_query(struct analysis_state * state,struct analysis_query * aq,uint32_t * result,const uint32_t * src)1922 process_uub_query(struct analysis_state *state, struct analysis_query *aq, uint32_t *result,
1923 const uint32_t *src)
1924 {
1925 struct uub_query q = *(struct uub_query *)aq;
1926
1927 *result = bitmask(q.scalar.def->bit_size);
1928 if (nir_scalar_is_const(q.scalar))
1929 *result = nir_scalar_as_uint(q.scalar);
1930 else if (nir_scalar_is_intrinsic(q.scalar))
1931 get_intrinsic_uub(state, q, result, src);
1932 else if (nir_scalar_is_alu(q.scalar))
1933 get_alu_uub(state, q, result, src);
1934 else if (q.scalar.def->parent_instr->type == nir_instr_type_phi)
1935 get_phi_uub(state, q, result, src);
1936 }
1937
1938 uint32_t
nir_unsigned_upper_bound(nir_shader * shader,struct hash_table * range_ht,nir_scalar scalar,const nir_unsigned_upper_bound_config * config)1939 nir_unsigned_upper_bound(nir_shader *shader, struct hash_table *range_ht,
1940 nir_scalar scalar,
1941 const nir_unsigned_upper_bound_config *config)
1942 {
1943 if (!config)
1944 config = &default_ub_config;
1945
1946 struct uub_query query_alloc[16];
1947 uint32_t result_alloc[16];
1948
1949 struct analysis_state state;
1950 state.shader = shader;
1951 state.config = config;
1952 state.range_ht = range_ht;
1953 util_dynarray_init_from_stack(&state.query_stack, query_alloc, sizeof(query_alloc));
1954 util_dynarray_init_from_stack(&state.result_stack, result_alloc, sizeof(result_alloc));
1955 state.query_size = sizeof(struct uub_query);
1956 state.get_key = &get_uub_key;
1957 state.process_query = &process_uub_query;
1958
1959 push_uub_query(&state, scalar);
1960
1961 return perform_analysis(&state);
1962 }
1963
1964 bool
nir_addition_might_overflow(nir_shader * shader,struct hash_table * range_ht,nir_scalar ssa,unsigned const_val,const nir_unsigned_upper_bound_config * config)1965 nir_addition_might_overflow(nir_shader *shader, struct hash_table *range_ht,
1966 nir_scalar ssa, unsigned const_val,
1967 const nir_unsigned_upper_bound_config *config)
1968 {
1969 if (nir_scalar_is_alu(ssa)) {
1970 nir_op alu_op = nir_scalar_alu_op(ssa);
1971
1972 /* iadd(imul(a, #b), #c) */
1973 if (alu_op == nir_op_imul || alu_op == nir_op_ishl) {
1974 nir_scalar mul_src0 = nir_scalar_chase_alu_src(ssa, 0);
1975 nir_scalar mul_src1 = nir_scalar_chase_alu_src(ssa, 1);
1976 uint32_t stride = 1;
1977 if (nir_scalar_is_const(mul_src0))
1978 stride = nir_scalar_as_uint(mul_src0);
1979 else if (nir_scalar_is_const(mul_src1))
1980 stride = nir_scalar_as_uint(mul_src1);
1981
1982 if (alu_op == nir_op_ishl)
1983 stride = 1u << (stride % 32u);
1984
1985 if (!stride || const_val <= UINT32_MAX - (UINT32_MAX / stride * stride))
1986 return false;
1987 }
1988
1989 /* iadd(iand(a, #b), #c) */
1990 if (alu_op == nir_op_iand) {
1991 nir_scalar and_src0 = nir_scalar_chase_alu_src(ssa, 0);
1992 nir_scalar and_src1 = nir_scalar_chase_alu_src(ssa, 1);
1993 uint32_t mask = 0xffffffff;
1994 if (nir_scalar_is_const(and_src0))
1995 mask = nir_scalar_as_uint(and_src0);
1996 else if (nir_scalar_is_const(and_src1))
1997 mask = nir_scalar_as_uint(and_src1);
1998 if (mask == 0 || const_val < (1u << (ffs(mask) - 1)))
1999 return false;
2000 }
2001 }
2002
2003 uint32_t ub = nir_unsigned_upper_bound(shader, range_ht, ssa, config);
2004 return const_val + ub < const_val;
2005 }
2006
2007 static uint64_t
ssa_def_bits_used(const nir_def * def,int recur)2008 ssa_def_bits_used(const nir_def *def, int recur)
2009 {
2010 uint64_t bits_used = 0;
2011 uint64_t all_bits = BITFIELD64_MASK(def->bit_size);
2012
2013 /* Querying the bits used from a vector is too hard of a question to
2014 * answer. Return the conservative answer that all bits are used. To
2015 * handle this, the function would need to be extended to be a query of a
2016 * single component of the vector. That would also necessary to fully
2017 * handle the 'num_components > 1' inside the loop below.
2018 *
2019 * FINISHME: This restriction will eventually need to be restricted to be
2020 * useful for hardware that uses u16vec2 as the native 16-bit integer type.
2021 */
2022 if (def->num_components > 1)
2023 return all_bits;
2024
2025 /* Limit recursion */
2026 if (recur-- <= 0)
2027 return all_bits;
2028
2029 nir_foreach_use(src, def) {
2030 switch (nir_src_parent_instr(src)->type) {
2031 case nir_instr_type_alu: {
2032 nir_alu_instr *use_alu = nir_instr_as_alu(nir_src_parent_instr(src));
2033 unsigned src_idx = container_of(src, nir_alu_src, src) - use_alu->src;
2034
2035 /* If a user of the value produces a vector result, return the
2036 * conservative answer that all bits are used. It is possible to
2037 * answer this query by looping over the components used. For example,
2038 *
2039 * vec4 32 ssa_5 = load_const(0x0000f000, 0x00000f00, 0x000000f0, 0x0000000f)
2040 * ...
2041 * vec4 32 ssa_8 = iand ssa_7.xxxx, ssa_5
2042 *
2043 * could conceivably return 0x0000ffff when queyring the bits used of
2044 * ssa_7. This is unlikely to be worth the effort because the
2045 * question can eventually answered after the shader has been
2046 * scalarized.
2047 */
2048 if (use_alu->def.num_components > 1)
2049 return all_bits;
2050
2051 switch (use_alu->op) {
2052 case nir_op_u2u8:
2053 case nir_op_i2i8:
2054 bits_used |= 0xff;
2055 break;
2056
2057 case nir_op_u2u16:
2058 case nir_op_i2i16:
2059 bits_used |= all_bits & 0xffff;
2060 break;
2061
2062 case nir_op_u2u32:
2063 case nir_op_i2i32:
2064 bits_used |= all_bits & 0xffffffff;
2065 break;
2066
2067 case nir_op_extract_u8:
2068 case nir_op_extract_i8:
2069 if (src_idx == 0 && nir_src_is_const(use_alu->src[1].src)) {
2070 unsigned chunk = nir_src_comp_as_uint(use_alu->src[1].src,
2071 use_alu->src[1].swizzle[0]);
2072 bits_used |= 0xffull << (chunk * 8);
2073 break;
2074 } else {
2075 return all_bits;
2076 }
2077
2078 case nir_op_extract_u16:
2079 case nir_op_extract_i16:
2080 if (src_idx == 0 && nir_src_is_const(use_alu->src[1].src)) {
2081 unsigned chunk = nir_src_comp_as_uint(use_alu->src[1].src,
2082 use_alu->src[1].swizzle[0]);
2083 bits_used |= 0xffffull << (chunk * 16);
2084 break;
2085 } else {
2086 return all_bits;
2087 }
2088
2089 case nir_op_ishl:
2090 case nir_op_ishr:
2091 case nir_op_ushr:
2092 if (src_idx == 1) {
2093 bits_used |= (nir_src_bit_size(use_alu->src[0].src) - 1);
2094 break;
2095 } else {
2096 return all_bits;
2097 }
2098
2099 case nir_op_iand:
2100 assert(src_idx < 2);
2101 if (nir_src_is_const(use_alu->src[1 - src_idx].src)) {
2102 uint64_t u64 = nir_src_comp_as_uint(use_alu->src[1 - src_idx].src,
2103 use_alu->src[1 - src_idx].swizzle[0]);
2104 bits_used |= u64;
2105 break;
2106 } else {
2107 return all_bits;
2108 }
2109
2110 case nir_op_ior:
2111 assert(src_idx < 2);
2112 if (nir_src_is_const(use_alu->src[1 - src_idx].src)) {
2113 uint64_t u64 = nir_src_comp_as_uint(use_alu->src[1 - src_idx].src,
2114 use_alu->src[1 - src_idx].swizzle[0]);
2115 bits_used |= all_bits & ~u64;
2116 break;
2117 } else {
2118 return all_bits;
2119 }
2120
2121 default:
2122 /* We don't know what this op does */
2123 return all_bits;
2124 }
2125 break;
2126 }
2127
2128 case nir_instr_type_intrinsic: {
2129 nir_intrinsic_instr *use_intrin =
2130 nir_instr_as_intrinsic(nir_src_parent_instr(src));
2131 unsigned src_idx = src - use_intrin->src;
2132
2133 switch (use_intrin->intrinsic) {
2134 case nir_intrinsic_read_invocation:
2135 case nir_intrinsic_shuffle:
2136 case nir_intrinsic_shuffle_up:
2137 case nir_intrinsic_shuffle_down:
2138 case nir_intrinsic_shuffle_xor:
2139 case nir_intrinsic_quad_broadcast:
2140 case nir_intrinsic_quad_swap_horizontal:
2141 case nir_intrinsic_quad_swap_vertical:
2142 case nir_intrinsic_quad_swap_diagonal:
2143 if (src_idx == 0) {
2144 bits_used |= ssa_def_bits_used(&use_intrin->def, recur);
2145 } else {
2146 if (use_intrin->intrinsic == nir_intrinsic_quad_broadcast) {
2147 bits_used |= 3;
2148 } else {
2149 /* Subgroups larger than 128 are not a thing */
2150 bits_used |= 127;
2151 }
2152 }
2153 break;
2154
2155 case nir_intrinsic_reduce:
2156 case nir_intrinsic_inclusive_scan:
2157 case nir_intrinsic_exclusive_scan:
2158 assert(src_idx == 0);
2159 switch (nir_intrinsic_reduction_op(use_intrin)) {
2160 case nir_op_iadd:
2161 case nir_op_imul:
2162 case nir_op_ior:
2163 case nir_op_iand:
2164 case nir_op_ixor:
2165 bits_used |= ssa_def_bits_used(&use_intrin->def, recur);
2166 break;
2167
2168 default:
2169 return all_bits;
2170 }
2171 break;
2172
2173 default:
2174 /* We don't know what this op does */
2175 return all_bits;
2176 }
2177 break;
2178 }
2179
2180 case nir_instr_type_phi: {
2181 nir_phi_instr *use_phi = nir_instr_as_phi(nir_src_parent_instr(src));
2182 bits_used |= ssa_def_bits_used(&use_phi->def, recur);
2183 break;
2184 }
2185
2186 default:
2187 return all_bits;
2188 }
2189
2190 /* If we've somehow shown that all our bits are used, we're done */
2191 assert((bits_used & ~all_bits) == 0);
2192 if (bits_used == all_bits)
2193 return all_bits;
2194 }
2195
2196 return bits_used;
2197 }
2198
2199 uint64_t
nir_def_bits_used(const nir_def * def)2200 nir_def_bits_used(const nir_def *def)
2201 {
2202 return ssa_def_bits_used(def, 2);
2203 }
2204