spirv: Replace unreachable with vtn_fail
[mesa.git] / src / compiler / spirv / spirv_to_nir.c
1 /*
2 * Copyright © 2015 Intel Corporation
3 *
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
10 *
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
13 * Software.
14 *
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
21 * IN THE SOFTWARE.
22 *
23 * Authors:
24 * Jason Ekstrand (jason@jlekstrand.net)
25 *
26 */
27
28 #include "vtn_private.h"
29 #include "nir/nir_vla.h"
30 #include "nir/nir_control_flow.h"
31 #include "nir/nir_constant_expressions.h"
32 #include "spirv_info.h"
33
34 void
35 vtn_log(struct vtn_builder *b, enum nir_spirv_debug_level level,
36 size_t spirv_offset, const char *message)
37 {
38 if (b->options->debug.func) {
39 b->options->debug.func(b->options->debug.private_data,
40 level, spirv_offset, message);
41 }
42
43 #ifndef NDEBUG
44 if (level >= NIR_SPIRV_DEBUG_LEVEL_WARNING)
45 fprintf(stderr, "%s\n", message);
46 #endif
47 }
48
49 void
50 vtn_logf(struct vtn_builder *b, enum nir_spirv_debug_level level,
51 size_t spirv_offset, const char *fmt, ...)
52 {
53 va_list args;
54 char *msg;
55
56 va_start(args, fmt);
57 msg = ralloc_vasprintf(NULL, fmt, args);
58 va_end(args);
59
60 vtn_log(b, level, spirv_offset, msg);
61
62 ralloc_free(msg);
63 }
64
65 static void
66 vtn_log_err(struct vtn_builder *b,
67 enum nir_spirv_debug_level level, const char *prefix,
68 const char *file, unsigned line,
69 const char *fmt, va_list args)
70 {
71 char *msg;
72
73 msg = ralloc_strdup(NULL, prefix);
74
75 #ifndef NDEBUG
76 ralloc_asprintf_append(&msg, " In file %s:%u\n", file, line);
77 #endif
78
79 ralloc_asprintf_append(&msg, " ");
80
81 ralloc_vasprintf_append(&msg, fmt, args);
82
83 ralloc_asprintf_append(&msg, "\n %zu bytes into the SPIR-V binary",
84 b->spirv_offset);
85
86 if (b->file) {
87 ralloc_asprintf_append(&msg,
88 "\n in SPIR-V source file %s, line %d, col %d",
89 b->file, b->line, b->col);
90 }
91
92 vtn_log(b, level, b->spirv_offset, msg);
93
94 ralloc_free(msg);
95 }
96
97 void
98 _vtn_warn(struct vtn_builder *b, const char *file, unsigned line,
99 const char *fmt, ...)
100 {
101 va_list args;
102
103 va_start(args, fmt);
104 vtn_log_err(b, NIR_SPIRV_DEBUG_LEVEL_WARNING, "SPIR-V WARNING:\n",
105 file, line, fmt, args);
106 va_end(args);
107 }
108
109 void
110 _vtn_fail(struct vtn_builder *b, const char *file, unsigned line,
111 const char *fmt, ...)
112 {
113 va_list args;
114
115 va_start(args, fmt);
116 vtn_log_err(b, NIR_SPIRV_DEBUG_LEVEL_ERROR, "SPIR-V parsing FAILED:\n",
117 file, line, fmt, args);
118 va_end(args);
119
120 longjmp(b->fail_jump, 1);
121 }
122
123 struct spec_constant_value {
124 bool is_double;
125 union {
126 uint32_t data32;
127 uint64_t data64;
128 };
129 };
130
131 static struct vtn_ssa_value *
132 vtn_undef_ssa_value(struct vtn_builder *b, const struct glsl_type *type)
133 {
134 struct vtn_ssa_value *val = rzalloc(b, struct vtn_ssa_value);
135 val->type = type;
136
137 if (glsl_type_is_vector_or_scalar(type)) {
138 unsigned num_components = glsl_get_vector_elements(val->type);
139 unsigned bit_size = glsl_get_bit_size(val->type);
140 val->def = nir_ssa_undef(&b->nb, num_components, bit_size);
141 } else {
142 unsigned elems = glsl_get_length(val->type);
143 val->elems = ralloc_array(b, struct vtn_ssa_value *, elems);
144 if (glsl_type_is_matrix(type)) {
145 const struct glsl_type *elem_type =
146 glsl_vector_type(glsl_get_base_type(type),
147 glsl_get_vector_elements(type));
148
149 for (unsigned i = 0; i < elems; i++)
150 val->elems[i] = vtn_undef_ssa_value(b, elem_type);
151 } else if (glsl_type_is_array(type)) {
152 const struct glsl_type *elem_type = glsl_get_array_element(type);
153 for (unsigned i = 0; i < elems; i++)
154 val->elems[i] = vtn_undef_ssa_value(b, elem_type);
155 } else {
156 for (unsigned i = 0; i < elems; i++) {
157 const struct glsl_type *elem_type = glsl_get_struct_field(type, i);
158 val->elems[i] = vtn_undef_ssa_value(b, elem_type);
159 }
160 }
161 }
162
163 return val;
164 }
165
166 static struct vtn_ssa_value *
167 vtn_const_ssa_value(struct vtn_builder *b, nir_constant *constant,
168 const struct glsl_type *type)
169 {
170 struct hash_entry *entry = _mesa_hash_table_search(b->const_table, constant);
171
172 if (entry)
173 return entry->data;
174
175 struct vtn_ssa_value *val = rzalloc(b, struct vtn_ssa_value);
176 val->type = type;
177
178 switch (glsl_get_base_type(type)) {
179 case GLSL_TYPE_INT:
180 case GLSL_TYPE_UINT:
181 case GLSL_TYPE_INT64:
182 case GLSL_TYPE_UINT64:
183 case GLSL_TYPE_BOOL:
184 case GLSL_TYPE_FLOAT:
185 case GLSL_TYPE_DOUBLE: {
186 int bit_size = glsl_get_bit_size(type);
187 if (glsl_type_is_vector_or_scalar(type)) {
188 unsigned num_components = glsl_get_vector_elements(val->type);
189 nir_load_const_instr *load =
190 nir_load_const_instr_create(b->shader, num_components, bit_size);
191
192 load->value = constant->values[0];
193
194 nir_instr_insert_before_cf_list(&b->nb.impl->body, &load->instr);
195 val->def = &load->def;
196 } else {
197 assert(glsl_type_is_matrix(type));
198 unsigned rows = glsl_get_vector_elements(val->type);
199 unsigned columns = glsl_get_matrix_columns(val->type);
200 val->elems = ralloc_array(b, struct vtn_ssa_value *, columns);
201
202 for (unsigned i = 0; i < columns; i++) {
203 struct vtn_ssa_value *col_val = rzalloc(b, struct vtn_ssa_value);
204 col_val->type = glsl_get_column_type(val->type);
205 nir_load_const_instr *load =
206 nir_load_const_instr_create(b->shader, rows, bit_size);
207
208 load->value = constant->values[i];
209
210 nir_instr_insert_before_cf_list(&b->nb.impl->body, &load->instr);
211 col_val->def = &load->def;
212
213 val->elems[i] = col_val;
214 }
215 }
216 break;
217 }
218
219 case GLSL_TYPE_ARRAY: {
220 unsigned elems = glsl_get_length(val->type);
221 val->elems = ralloc_array(b, struct vtn_ssa_value *, elems);
222 const struct glsl_type *elem_type = glsl_get_array_element(val->type);
223 for (unsigned i = 0; i < elems; i++)
224 val->elems[i] = vtn_const_ssa_value(b, constant->elements[i],
225 elem_type);
226 break;
227 }
228
229 case GLSL_TYPE_STRUCT: {
230 unsigned elems = glsl_get_length(val->type);
231 val->elems = ralloc_array(b, struct vtn_ssa_value *, elems);
232 for (unsigned i = 0; i < elems; i++) {
233 const struct glsl_type *elem_type =
234 glsl_get_struct_field(val->type, i);
235 val->elems[i] = vtn_const_ssa_value(b, constant->elements[i],
236 elem_type);
237 }
238 break;
239 }
240
241 default:
242 vtn_fail("bad constant type");
243 }
244
245 return val;
246 }
247
248 struct vtn_ssa_value *
249 vtn_ssa_value(struct vtn_builder *b, uint32_t value_id)
250 {
251 struct vtn_value *val = vtn_untyped_value(b, value_id);
252 switch (val->value_type) {
253 case vtn_value_type_undef:
254 return vtn_undef_ssa_value(b, val->type->type);
255
256 case vtn_value_type_constant:
257 return vtn_const_ssa_value(b, val->constant, val->const_type);
258
259 case vtn_value_type_ssa:
260 return val->ssa;
261
262 case vtn_value_type_pointer:
263 vtn_assert(val->pointer->ptr_type && val->pointer->ptr_type->type);
264 struct vtn_ssa_value *ssa =
265 vtn_create_ssa_value(b, val->pointer->ptr_type->type);
266 ssa->def = vtn_pointer_to_ssa(b, val->pointer);
267 return ssa;
268
269 default:
270 vtn_fail("Invalid type for an SSA value");
271 }
272 }
273
274 static char *
275 vtn_string_literal(struct vtn_builder *b, const uint32_t *words,
276 unsigned word_count, unsigned *words_used)
277 {
278 char *dup = ralloc_strndup(b, (char *)words, word_count * sizeof(*words));
279 if (words_used) {
280 /* Ammount of space taken by the string (including the null) */
281 unsigned len = strlen(dup) + 1;
282 *words_used = DIV_ROUND_UP(len, sizeof(*words));
283 }
284 return dup;
285 }
286
287 const uint32_t *
288 vtn_foreach_instruction(struct vtn_builder *b, const uint32_t *start,
289 const uint32_t *end, vtn_instruction_handler handler)
290 {
291 b->file = NULL;
292 b->line = -1;
293 b->col = -1;
294
295 const uint32_t *w = start;
296 while (w < end) {
297 SpvOp opcode = w[0] & SpvOpCodeMask;
298 unsigned count = w[0] >> SpvWordCountShift;
299 vtn_assert(count >= 1 && w + count <= end);
300
301 b->spirv_offset = (uint8_t *)w - (uint8_t *)b->spirv;
302
303 switch (opcode) {
304 case SpvOpNop:
305 break; /* Do nothing */
306
307 case SpvOpLine:
308 b->file = vtn_value(b, w[1], vtn_value_type_string)->str;
309 b->line = w[2];
310 b->col = w[3];
311 break;
312
313 case SpvOpNoLine:
314 b->file = NULL;
315 b->line = -1;
316 b->col = -1;
317 break;
318
319 default:
320 if (!handler(b, opcode, w, count))
321 return w;
322 break;
323 }
324
325 w += count;
326 }
327
328 b->spirv_offset = 0;
329 b->file = NULL;
330 b->line = -1;
331 b->col = -1;
332
333 assert(w == end);
334 return w;
335 }
336
337 static void
338 vtn_handle_extension(struct vtn_builder *b, SpvOp opcode,
339 const uint32_t *w, unsigned count)
340 {
341 switch (opcode) {
342 case SpvOpExtInstImport: {
343 struct vtn_value *val = vtn_push_value(b, w[1], vtn_value_type_extension);
344 if (strcmp((const char *)&w[2], "GLSL.std.450") == 0) {
345 val->ext_handler = vtn_handle_glsl450_instruction;
346 } else {
347 vtn_fail("Unsupported extension");
348 }
349 break;
350 }
351
352 case SpvOpExtInst: {
353 struct vtn_value *val = vtn_value(b, w[3], vtn_value_type_extension);
354 bool handled = val->ext_handler(b, w[4], w, count);
355 vtn_assert(handled);
356 break;
357 }
358
359 default:
360 vtn_fail("Unhandled opcode");
361 }
362 }
363
364 static void
365 _foreach_decoration_helper(struct vtn_builder *b,
366 struct vtn_value *base_value,
367 int parent_member,
368 struct vtn_value *value,
369 vtn_decoration_foreach_cb cb, void *data)
370 {
371 for (struct vtn_decoration *dec = value->decoration; dec; dec = dec->next) {
372 int member;
373 if (dec->scope == VTN_DEC_DECORATION) {
374 member = parent_member;
375 } else if (dec->scope >= VTN_DEC_STRUCT_MEMBER0) {
376 vtn_assert(parent_member == -1);
377 member = dec->scope - VTN_DEC_STRUCT_MEMBER0;
378 } else {
379 /* Not a decoration */
380 continue;
381 }
382
383 if (dec->group) {
384 vtn_assert(dec->group->value_type == vtn_value_type_decoration_group);
385 _foreach_decoration_helper(b, base_value, member, dec->group,
386 cb, data);
387 } else {
388 cb(b, base_value, member, dec, data);
389 }
390 }
391 }
392
393 /** Iterates (recursively if needed) over all of the decorations on a value
394 *
395 * This function iterates over all of the decorations applied to a given
396 * value. If it encounters a decoration group, it recurses into the group
397 * and iterates over all of those decorations as well.
398 */
399 void
400 vtn_foreach_decoration(struct vtn_builder *b, struct vtn_value *value,
401 vtn_decoration_foreach_cb cb, void *data)
402 {
403 _foreach_decoration_helper(b, value, -1, value, cb, data);
404 }
405
406 void
407 vtn_foreach_execution_mode(struct vtn_builder *b, struct vtn_value *value,
408 vtn_execution_mode_foreach_cb cb, void *data)
409 {
410 for (struct vtn_decoration *dec = value->decoration; dec; dec = dec->next) {
411 if (dec->scope != VTN_DEC_EXECUTION_MODE)
412 continue;
413
414 vtn_assert(dec->group == NULL);
415 cb(b, value, dec, data);
416 }
417 }
418
419 static void
420 vtn_handle_decoration(struct vtn_builder *b, SpvOp opcode,
421 const uint32_t *w, unsigned count)
422 {
423 const uint32_t *w_end = w + count;
424 const uint32_t target = w[1];
425 w += 2;
426
427 switch (opcode) {
428 case SpvOpDecorationGroup:
429 vtn_push_value(b, target, vtn_value_type_decoration_group);
430 break;
431
432 case SpvOpDecorate:
433 case SpvOpMemberDecorate:
434 case SpvOpExecutionMode: {
435 struct vtn_value *val = &b->values[target];
436
437 struct vtn_decoration *dec = rzalloc(b, struct vtn_decoration);
438 switch (opcode) {
439 case SpvOpDecorate:
440 dec->scope = VTN_DEC_DECORATION;
441 break;
442 case SpvOpMemberDecorate:
443 dec->scope = VTN_DEC_STRUCT_MEMBER0 + *(w++);
444 break;
445 case SpvOpExecutionMode:
446 dec->scope = VTN_DEC_EXECUTION_MODE;
447 break;
448 default:
449 vtn_fail("Invalid decoration opcode");
450 }
451 dec->decoration = *(w++);
452 dec->literals = w;
453
454 /* Link into the list */
455 dec->next = val->decoration;
456 val->decoration = dec;
457 break;
458 }
459
460 case SpvOpGroupMemberDecorate:
461 case SpvOpGroupDecorate: {
462 struct vtn_value *group =
463 vtn_value(b, target, vtn_value_type_decoration_group);
464
465 for (; w < w_end; w++) {
466 struct vtn_value *val = vtn_untyped_value(b, *w);
467 struct vtn_decoration *dec = rzalloc(b, struct vtn_decoration);
468
469 dec->group = group;
470 if (opcode == SpvOpGroupDecorate) {
471 dec->scope = VTN_DEC_DECORATION;
472 } else {
473 dec->scope = VTN_DEC_STRUCT_MEMBER0 + *(++w);
474 }
475
476 /* Link into the list */
477 dec->next = val->decoration;
478 val->decoration = dec;
479 }
480 break;
481 }
482
483 default:
484 vtn_fail("Unhandled opcode");
485 }
486 }
487
488 struct member_decoration_ctx {
489 unsigned num_fields;
490 struct glsl_struct_field *fields;
491 struct vtn_type *type;
492 };
493
494 /* does a shallow copy of a vtn_type */
495
496 static struct vtn_type *
497 vtn_type_copy(struct vtn_builder *b, struct vtn_type *src)
498 {
499 struct vtn_type *dest = ralloc(b, struct vtn_type);
500 *dest = *src;
501
502 switch (src->base_type) {
503 case vtn_base_type_void:
504 case vtn_base_type_scalar:
505 case vtn_base_type_vector:
506 case vtn_base_type_matrix:
507 case vtn_base_type_array:
508 case vtn_base_type_pointer:
509 case vtn_base_type_image:
510 case vtn_base_type_sampler:
511 /* Nothing more to do */
512 break;
513
514 case vtn_base_type_struct:
515 dest->members = ralloc_array(b, struct vtn_type *, src->length);
516 memcpy(dest->members, src->members,
517 src->length * sizeof(src->members[0]));
518
519 dest->offsets = ralloc_array(b, unsigned, src->length);
520 memcpy(dest->offsets, src->offsets,
521 src->length * sizeof(src->offsets[0]));
522 break;
523
524 case vtn_base_type_function:
525 dest->params = ralloc_array(b, struct vtn_type *, src->length);
526 memcpy(dest->params, src->params, src->length * sizeof(src->params[0]));
527 break;
528 }
529
530 return dest;
531 }
532
533 static struct vtn_type *
534 mutable_matrix_member(struct vtn_builder *b, struct vtn_type *type, int member)
535 {
536 type->members[member] = vtn_type_copy(b, type->members[member]);
537 type = type->members[member];
538
539 /* We may have an array of matrices.... Oh, joy! */
540 while (glsl_type_is_array(type->type)) {
541 type->array_element = vtn_type_copy(b, type->array_element);
542 type = type->array_element;
543 }
544
545 vtn_assert(glsl_type_is_matrix(type->type));
546
547 return type;
548 }
549
550 static void
551 struct_member_decoration_cb(struct vtn_builder *b,
552 struct vtn_value *val, int member,
553 const struct vtn_decoration *dec, void *void_ctx)
554 {
555 struct member_decoration_ctx *ctx = void_ctx;
556
557 if (member < 0)
558 return;
559
560 vtn_assert(member < ctx->num_fields);
561
562 switch (dec->decoration) {
563 case SpvDecorationNonWritable:
564 case SpvDecorationNonReadable:
565 case SpvDecorationRelaxedPrecision:
566 case SpvDecorationVolatile:
567 case SpvDecorationCoherent:
568 case SpvDecorationUniform:
569 break; /* FIXME: Do nothing with this for now. */
570 case SpvDecorationNoPerspective:
571 ctx->fields[member].interpolation = INTERP_MODE_NOPERSPECTIVE;
572 break;
573 case SpvDecorationFlat:
574 ctx->fields[member].interpolation = INTERP_MODE_FLAT;
575 break;
576 case SpvDecorationCentroid:
577 ctx->fields[member].centroid = true;
578 break;
579 case SpvDecorationSample:
580 ctx->fields[member].sample = true;
581 break;
582 case SpvDecorationStream:
583 /* Vulkan only allows one GS stream */
584 vtn_assert(dec->literals[0] == 0);
585 break;
586 case SpvDecorationLocation:
587 ctx->fields[member].location = dec->literals[0];
588 break;
589 case SpvDecorationComponent:
590 break; /* FIXME: What should we do with these? */
591 case SpvDecorationBuiltIn:
592 ctx->type->members[member] = vtn_type_copy(b, ctx->type->members[member]);
593 ctx->type->members[member]->is_builtin = true;
594 ctx->type->members[member]->builtin = dec->literals[0];
595 ctx->type->builtin_block = true;
596 break;
597 case SpvDecorationOffset:
598 ctx->type->offsets[member] = dec->literals[0];
599 break;
600 case SpvDecorationMatrixStride:
601 /* Handled as a second pass */
602 break;
603 case SpvDecorationColMajor:
604 break; /* Nothing to do here. Column-major is the default. */
605 case SpvDecorationRowMajor:
606 mutable_matrix_member(b, ctx->type, member)->row_major = true;
607 break;
608
609 case SpvDecorationPatch:
610 break;
611
612 case SpvDecorationSpecId:
613 case SpvDecorationBlock:
614 case SpvDecorationBufferBlock:
615 case SpvDecorationArrayStride:
616 case SpvDecorationGLSLShared:
617 case SpvDecorationGLSLPacked:
618 case SpvDecorationInvariant:
619 case SpvDecorationRestrict:
620 case SpvDecorationAliased:
621 case SpvDecorationConstant:
622 case SpvDecorationIndex:
623 case SpvDecorationBinding:
624 case SpvDecorationDescriptorSet:
625 case SpvDecorationLinkageAttributes:
626 case SpvDecorationNoContraction:
627 case SpvDecorationInputAttachmentIndex:
628 vtn_warn("Decoration not allowed on struct members: %s",
629 spirv_decoration_to_string(dec->decoration));
630 break;
631
632 case SpvDecorationXfbBuffer:
633 case SpvDecorationXfbStride:
634 vtn_warn("Vulkan does not have transform feedback");
635 break;
636
637 case SpvDecorationCPacked:
638 case SpvDecorationSaturatedConversion:
639 case SpvDecorationFuncParamAttr:
640 case SpvDecorationFPRoundingMode:
641 case SpvDecorationFPFastMathMode:
642 case SpvDecorationAlignment:
643 vtn_warn("Decoration only allowed for CL-style kernels: %s",
644 spirv_decoration_to_string(dec->decoration));
645 break;
646
647 default:
648 vtn_fail("Unhandled decoration");
649 }
650 }
651
652 /* Matrix strides are handled as a separate pass because we need to know
653 * whether the matrix is row-major or not first.
654 */
655 static void
656 struct_member_matrix_stride_cb(struct vtn_builder *b,
657 struct vtn_value *val, int member,
658 const struct vtn_decoration *dec,
659 void *void_ctx)
660 {
661 if (dec->decoration != SpvDecorationMatrixStride)
662 return;
663 vtn_assert(member >= 0);
664
665 struct member_decoration_ctx *ctx = void_ctx;
666
667 struct vtn_type *mat_type = mutable_matrix_member(b, ctx->type, member);
668 if (mat_type->row_major) {
669 mat_type->array_element = vtn_type_copy(b, mat_type->array_element);
670 mat_type->stride = mat_type->array_element->stride;
671 mat_type->array_element->stride = dec->literals[0];
672 } else {
673 vtn_assert(mat_type->array_element->stride > 0);
674 mat_type->stride = dec->literals[0];
675 }
676 }
677
678 static void
679 type_decoration_cb(struct vtn_builder *b,
680 struct vtn_value *val, int member,
681 const struct vtn_decoration *dec, void *ctx)
682 {
683 struct vtn_type *type = val->type;
684
685 if (member != -1)
686 return;
687
688 switch (dec->decoration) {
689 case SpvDecorationArrayStride:
690 vtn_assert(type->base_type == vtn_base_type_matrix ||
691 type->base_type == vtn_base_type_array ||
692 type->base_type == vtn_base_type_pointer);
693 type->stride = dec->literals[0];
694 break;
695 case SpvDecorationBlock:
696 vtn_assert(type->base_type == vtn_base_type_struct);
697 type->block = true;
698 break;
699 case SpvDecorationBufferBlock:
700 vtn_assert(type->base_type == vtn_base_type_struct);
701 type->buffer_block = true;
702 break;
703 case SpvDecorationGLSLShared:
704 case SpvDecorationGLSLPacked:
705 /* Ignore these, since we get explicit offsets anyways */
706 break;
707
708 case SpvDecorationRowMajor:
709 case SpvDecorationColMajor:
710 case SpvDecorationMatrixStride:
711 case SpvDecorationBuiltIn:
712 case SpvDecorationNoPerspective:
713 case SpvDecorationFlat:
714 case SpvDecorationPatch:
715 case SpvDecorationCentroid:
716 case SpvDecorationSample:
717 case SpvDecorationVolatile:
718 case SpvDecorationCoherent:
719 case SpvDecorationNonWritable:
720 case SpvDecorationNonReadable:
721 case SpvDecorationUniform:
722 case SpvDecorationStream:
723 case SpvDecorationLocation:
724 case SpvDecorationComponent:
725 case SpvDecorationOffset:
726 case SpvDecorationXfbBuffer:
727 case SpvDecorationXfbStride:
728 vtn_warn("Decoration only allowed for struct members: %s",
729 spirv_decoration_to_string(dec->decoration));
730 break;
731
732 case SpvDecorationRelaxedPrecision:
733 case SpvDecorationSpecId:
734 case SpvDecorationInvariant:
735 case SpvDecorationRestrict:
736 case SpvDecorationAliased:
737 case SpvDecorationConstant:
738 case SpvDecorationIndex:
739 case SpvDecorationBinding:
740 case SpvDecorationDescriptorSet:
741 case SpvDecorationLinkageAttributes:
742 case SpvDecorationNoContraction:
743 case SpvDecorationInputAttachmentIndex:
744 vtn_warn("Decoration not allowed on types: %s",
745 spirv_decoration_to_string(dec->decoration));
746 break;
747
748 case SpvDecorationCPacked:
749 case SpvDecorationSaturatedConversion:
750 case SpvDecorationFuncParamAttr:
751 case SpvDecorationFPRoundingMode:
752 case SpvDecorationFPFastMathMode:
753 case SpvDecorationAlignment:
754 vtn_warn("Decoration only allowed for CL-style kernels: %s",
755 spirv_decoration_to_string(dec->decoration));
756 break;
757
758 default:
759 vtn_fail("Unhandled decoration");
760 }
761 }
762
763 static unsigned
764 translate_image_format(struct vtn_builder *b, SpvImageFormat format)
765 {
766 switch (format) {
767 case SpvImageFormatUnknown: return 0; /* GL_NONE */
768 case SpvImageFormatRgba32f: return 0x8814; /* GL_RGBA32F */
769 case SpvImageFormatRgba16f: return 0x881A; /* GL_RGBA16F */
770 case SpvImageFormatR32f: return 0x822E; /* GL_R32F */
771 case SpvImageFormatRgba8: return 0x8058; /* GL_RGBA8 */
772 case SpvImageFormatRgba8Snorm: return 0x8F97; /* GL_RGBA8_SNORM */
773 case SpvImageFormatRg32f: return 0x8230; /* GL_RG32F */
774 case SpvImageFormatRg16f: return 0x822F; /* GL_RG16F */
775 case SpvImageFormatR11fG11fB10f: return 0x8C3A; /* GL_R11F_G11F_B10F */
776 case SpvImageFormatR16f: return 0x822D; /* GL_R16F */
777 case SpvImageFormatRgba16: return 0x805B; /* GL_RGBA16 */
778 case SpvImageFormatRgb10A2: return 0x8059; /* GL_RGB10_A2 */
779 case SpvImageFormatRg16: return 0x822C; /* GL_RG16 */
780 case SpvImageFormatRg8: return 0x822B; /* GL_RG8 */
781 case SpvImageFormatR16: return 0x822A; /* GL_R16 */
782 case SpvImageFormatR8: return 0x8229; /* GL_R8 */
783 case SpvImageFormatRgba16Snorm: return 0x8F9B; /* GL_RGBA16_SNORM */
784 case SpvImageFormatRg16Snorm: return 0x8F99; /* GL_RG16_SNORM */
785 case SpvImageFormatRg8Snorm: return 0x8F95; /* GL_RG8_SNORM */
786 case SpvImageFormatR16Snorm: return 0x8F98; /* GL_R16_SNORM */
787 case SpvImageFormatR8Snorm: return 0x8F94; /* GL_R8_SNORM */
788 case SpvImageFormatRgba32i: return 0x8D82; /* GL_RGBA32I */
789 case SpvImageFormatRgba16i: return 0x8D88; /* GL_RGBA16I */
790 case SpvImageFormatRgba8i: return 0x8D8E; /* GL_RGBA8I */
791 case SpvImageFormatR32i: return 0x8235; /* GL_R32I */
792 case SpvImageFormatRg32i: return 0x823B; /* GL_RG32I */
793 case SpvImageFormatRg16i: return 0x8239; /* GL_RG16I */
794 case SpvImageFormatRg8i: return 0x8237; /* GL_RG8I */
795 case SpvImageFormatR16i: return 0x8233; /* GL_R16I */
796 case SpvImageFormatR8i: return 0x8231; /* GL_R8I */
797 case SpvImageFormatRgba32ui: return 0x8D70; /* GL_RGBA32UI */
798 case SpvImageFormatRgba16ui: return 0x8D76; /* GL_RGBA16UI */
799 case SpvImageFormatRgba8ui: return 0x8D7C; /* GL_RGBA8UI */
800 case SpvImageFormatR32ui: return 0x8236; /* GL_R32UI */
801 case SpvImageFormatRgb10a2ui: return 0x906F; /* GL_RGB10_A2UI */
802 case SpvImageFormatRg32ui: return 0x823C; /* GL_RG32UI */
803 case SpvImageFormatRg16ui: return 0x823A; /* GL_RG16UI */
804 case SpvImageFormatRg8ui: return 0x8238; /* GL_RG8UI */
805 case SpvImageFormatR16ui: return 0x8234; /* GL_R16UI */
806 case SpvImageFormatR8ui: return 0x8232; /* GL_R8UI */
807 default:
808 vtn_fail("Invalid image format");
809 }
810 }
811
812 static void
813 vtn_handle_type(struct vtn_builder *b, SpvOp opcode,
814 const uint32_t *w, unsigned count)
815 {
816 struct vtn_value *val = vtn_push_value(b, w[1], vtn_value_type_type);
817
818 val->type = rzalloc(b, struct vtn_type);
819 val->type->val = val;
820
821 switch (opcode) {
822 case SpvOpTypeVoid:
823 val->type->base_type = vtn_base_type_void;
824 val->type->type = glsl_void_type();
825 break;
826 case SpvOpTypeBool:
827 val->type->base_type = vtn_base_type_scalar;
828 val->type->type = glsl_bool_type();
829 break;
830 case SpvOpTypeInt: {
831 int bit_size = w[2];
832 const bool signedness = w[3];
833 val->type->base_type = vtn_base_type_scalar;
834 if (bit_size == 64)
835 val->type->type = (signedness ? glsl_int64_t_type() : glsl_uint64_t_type());
836 else
837 val->type->type = (signedness ? glsl_int_type() : glsl_uint_type());
838 break;
839 }
840 case SpvOpTypeFloat: {
841 int bit_size = w[2];
842 val->type->base_type = vtn_base_type_scalar;
843 val->type->type = bit_size == 64 ? glsl_double_type() : glsl_float_type();
844 break;
845 }
846
847 case SpvOpTypeVector: {
848 struct vtn_type *base = vtn_value(b, w[2], vtn_value_type_type)->type;
849 unsigned elems = w[3];
850
851 vtn_assert(glsl_type_is_scalar(base->type));
852 val->type->base_type = vtn_base_type_vector;
853 val->type->type = glsl_vector_type(glsl_get_base_type(base->type), elems);
854 val->type->stride = glsl_get_bit_size(base->type) / 8;
855 val->type->array_element = base;
856 break;
857 }
858
859 case SpvOpTypeMatrix: {
860 struct vtn_type *base = vtn_value(b, w[2], vtn_value_type_type)->type;
861 unsigned columns = w[3];
862
863 vtn_assert(glsl_type_is_vector(base->type));
864 val->type->base_type = vtn_base_type_matrix;
865 val->type->type = glsl_matrix_type(glsl_get_base_type(base->type),
866 glsl_get_vector_elements(base->type),
867 columns);
868 vtn_assert(!glsl_type_is_error(val->type->type));
869 val->type->length = columns;
870 val->type->array_element = base;
871 val->type->row_major = false;
872 val->type->stride = 0;
873 break;
874 }
875
876 case SpvOpTypeRuntimeArray:
877 case SpvOpTypeArray: {
878 struct vtn_type *array_element =
879 vtn_value(b, w[2], vtn_value_type_type)->type;
880
881 if (opcode == SpvOpTypeRuntimeArray) {
882 /* A length of 0 is used to denote unsized arrays */
883 val->type->length = 0;
884 } else {
885 val->type->length =
886 vtn_value(b, w[3], vtn_value_type_constant)->constant->values[0].u32[0];
887 }
888
889 val->type->base_type = vtn_base_type_array;
890 val->type->type = glsl_array_type(array_element->type, val->type->length);
891 val->type->array_element = array_element;
892 val->type->stride = 0;
893 break;
894 }
895
896 case SpvOpTypeStruct: {
897 unsigned num_fields = count - 2;
898 val->type->base_type = vtn_base_type_struct;
899 val->type->length = num_fields;
900 val->type->members = ralloc_array(b, struct vtn_type *, num_fields);
901 val->type->offsets = ralloc_array(b, unsigned, num_fields);
902
903 NIR_VLA(struct glsl_struct_field, fields, count);
904 for (unsigned i = 0; i < num_fields; i++) {
905 val->type->members[i] =
906 vtn_value(b, w[i + 2], vtn_value_type_type)->type;
907 fields[i] = (struct glsl_struct_field) {
908 .type = val->type->members[i]->type,
909 .name = ralloc_asprintf(b, "field%d", i),
910 .location = -1,
911 };
912 }
913
914 struct member_decoration_ctx ctx = {
915 .num_fields = num_fields,
916 .fields = fields,
917 .type = val->type
918 };
919
920 vtn_foreach_decoration(b, val, struct_member_decoration_cb, &ctx);
921 vtn_foreach_decoration(b, val, struct_member_matrix_stride_cb, &ctx);
922
923 const char *name = val->name ? val->name : "struct";
924
925 val->type->type = glsl_struct_type(fields, num_fields, name);
926 break;
927 }
928
929 case SpvOpTypeFunction: {
930 val->type->base_type = vtn_base_type_function;
931 val->type->type = NULL;
932
933 val->type->return_type = vtn_value(b, w[2], vtn_value_type_type)->type;
934
935 const unsigned num_params = count - 3;
936 val->type->length = num_params;
937 val->type->params = ralloc_array(b, struct vtn_type *, num_params);
938 for (unsigned i = 0; i < count - 3; i++) {
939 val->type->params[i] =
940 vtn_value(b, w[i + 3], vtn_value_type_type)->type;
941 }
942 break;
943 }
944
945 case SpvOpTypePointer: {
946 SpvStorageClass storage_class = w[2];
947 struct vtn_type *deref_type =
948 vtn_value(b, w[3], vtn_value_type_type)->type;
949
950 val->type->base_type = vtn_base_type_pointer;
951 val->type->storage_class = storage_class;
952 val->type->deref = deref_type;
953
954 if (storage_class == SpvStorageClassUniform ||
955 storage_class == SpvStorageClassStorageBuffer) {
956 /* These can actually be stored to nir_variables and used as SSA
957 * values so they need a real glsl_type.
958 */
959 val->type->type = glsl_vector_type(GLSL_TYPE_UINT, 2);
960 }
961 break;
962 }
963
964 case SpvOpTypeImage: {
965 val->type->base_type = vtn_base_type_image;
966
967 const struct glsl_type *sampled_type =
968 vtn_value(b, w[2], vtn_value_type_type)->type->type;
969
970 vtn_assert(glsl_type_is_vector_or_scalar(sampled_type));
971
972 enum glsl_sampler_dim dim;
973 switch ((SpvDim)w[3]) {
974 case SpvDim1D: dim = GLSL_SAMPLER_DIM_1D; break;
975 case SpvDim2D: dim = GLSL_SAMPLER_DIM_2D; break;
976 case SpvDim3D: dim = GLSL_SAMPLER_DIM_3D; break;
977 case SpvDimCube: dim = GLSL_SAMPLER_DIM_CUBE; break;
978 case SpvDimRect: dim = GLSL_SAMPLER_DIM_RECT; break;
979 case SpvDimBuffer: dim = GLSL_SAMPLER_DIM_BUF; break;
980 case SpvDimSubpassData: dim = GLSL_SAMPLER_DIM_SUBPASS; break;
981 default:
982 vtn_fail("Invalid SPIR-V Sampler dimension");
983 }
984
985 bool is_shadow = w[4];
986 bool is_array = w[5];
987 bool multisampled = w[6];
988 unsigned sampled = w[7];
989 SpvImageFormat format = w[8];
990
991 if (count > 9)
992 val->type->access_qualifier = w[9];
993 else
994 val->type->access_qualifier = SpvAccessQualifierReadWrite;
995
996 if (multisampled) {
997 if (dim == GLSL_SAMPLER_DIM_2D)
998 dim = GLSL_SAMPLER_DIM_MS;
999 else if (dim == GLSL_SAMPLER_DIM_SUBPASS)
1000 dim = GLSL_SAMPLER_DIM_SUBPASS_MS;
1001 else
1002 vtn_fail("Unsupported multisampled image type");
1003 }
1004
1005 val->type->image_format = translate_image_format(b, format);
1006
1007 if (sampled == 1) {
1008 val->type->sampled = true;
1009 val->type->type = glsl_sampler_type(dim, is_shadow, is_array,
1010 glsl_get_base_type(sampled_type));
1011 } else if (sampled == 2) {
1012 vtn_assert(!is_shadow);
1013 val->type->sampled = false;
1014 val->type->type = glsl_image_type(dim, is_array,
1015 glsl_get_base_type(sampled_type));
1016 } else {
1017 vtn_fail("We need to know if the image will be sampled");
1018 }
1019 break;
1020 }
1021
1022 case SpvOpTypeSampledImage:
1023 val->type = vtn_value(b, w[2], vtn_value_type_type)->type;
1024 break;
1025
1026 case SpvOpTypeSampler:
1027 /* The actual sampler type here doesn't really matter. It gets
1028 * thrown away the moment you combine it with an image. What really
1029 * matters is that it's a sampler type as opposed to an integer type
1030 * so the backend knows what to do.
1031 */
1032 val->type->base_type = vtn_base_type_sampler;
1033 val->type->type = glsl_bare_sampler_type();
1034 break;
1035
1036 case SpvOpTypeOpaque:
1037 case SpvOpTypeEvent:
1038 case SpvOpTypeDeviceEvent:
1039 case SpvOpTypeReserveId:
1040 case SpvOpTypeQueue:
1041 case SpvOpTypePipe:
1042 default:
1043 vtn_fail("Unhandled opcode");
1044 }
1045
1046 vtn_foreach_decoration(b, val, type_decoration_cb, NULL);
1047 }
1048
1049 static nir_constant *
1050 vtn_null_constant(struct vtn_builder *b, const struct glsl_type *type)
1051 {
1052 nir_constant *c = rzalloc(b, nir_constant);
1053
1054 /* For pointers and other typeless things, we have to return something but
1055 * it doesn't matter what.
1056 */
1057 if (!type)
1058 return c;
1059
1060 switch (glsl_get_base_type(type)) {
1061 case GLSL_TYPE_INT:
1062 case GLSL_TYPE_UINT:
1063 case GLSL_TYPE_INT64:
1064 case GLSL_TYPE_UINT64:
1065 case GLSL_TYPE_BOOL:
1066 case GLSL_TYPE_FLOAT:
1067 case GLSL_TYPE_DOUBLE:
1068 /* Nothing to do here. It's already initialized to zero */
1069 break;
1070
1071 case GLSL_TYPE_ARRAY:
1072 vtn_assert(glsl_get_length(type) > 0);
1073 c->num_elements = glsl_get_length(type);
1074 c->elements = ralloc_array(b, nir_constant *, c->num_elements);
1075
1076 c->elements[0] = vtn_null_constant(b, glsl_get_array_element(type));
1077 for (unsigned i = 1; i < c->num_elements; i++)
1078 c->elements[i] = c->elements[0];
1079 break;
1080
1081 case GLSL_TYPE_STRUCT:
1082 c->num_elements = glsl_get_length(type);
1083 c->elements = ralloc_array(b, nir_constant *, c->num_elements);
1084
1085 for (unsigned i = 0; i < c->num_elements; i++) {
1086 c->elements[i] = vtn_null_constant(b, glsl_get_struct_field(type, i));
1087 }
1088 break;
1089
1090 default:
1091 vtn_fail("Invalid type for null constant");
1092 }
1093
1094 return c;
1095 }
1096
1097 static void
1098 spec_constant_decoration_cb(struct vtn_builder *b, struct vtn_value *v,
1099 int member, const struct vtn_decoration *dec,
1100 void *data)
1101 {
1102 vtn_assert(member == -1);
1103 if (dec->decoration != SpvDecorationSpecId)
1104 return;
1105
1106 struct spec_constant_value *const_value = data;
1107
1108 for (unsigned i = 0; i < b->num_specializations; i++) {
1109 if (b->specializations[i].id == dec->literals[0]) {
1110 if (const_value->is_double)
1111 const_value->data64 = b->specializations[i].data64;
1112 else
1113 const_value->data32 = b->specializations[i].data32;
1114 return;
1115 }
1116 }
1117 }
1118
1119 static uint32_t
1120 get_specialization(struct vtn_builder *b, struct vtn_value *val,
1121 uint32_t const_value)
1122 {
1123 struct spec_constant_value data;
1124 data.is_double = false;
1125 data.data32 = const_value;
1126 vtn_foreach_decoration(b, val, spec_constant_decoration_cb, &data);
1127 return data.data32;
1128 }
1129
1130 static uint64_t
1131 get_specialization64(struct vtn_builder *b, struct vtn_value *val,
1132 uint64_t const_value)
1133 {
1134 struct spec_constant_value data;
1135 data.is_double = true;
1136 data.data64 = const_value;
1137 vtn_foreach_decoration(b, val, spec_constant_decoration_cb, &data);
1138 return data.data64;
1139 }
1140
1141 static void
1142 handle_workgroup_size_decoration_cb(struct vtn_builder *b,
1143 struct vtn_value *val,
1144 int member,
1145 const struct vtn_decoration *dec,
1146 void *data)
1147 {
1148 vtn_assert(member == -1);
1149 if (dec->decoration != SpvDecorationBuiltIn ||
1150 dec->literals[0] != SpvBuiltInWorkgroupSize)
1151 return;
1152
1153 vtn_assert(val->const_type == glsl_vector_type(GLSL_TYPE_UINT, 3));
1154
1155 b->shader->info.cs.local_size[0] = val->constant->values[0].u32[0];
1156 b->shader->info.cs.local_size[1] = val->constant->values[0].u32[1];
1157 b->shader->info.cs.local_size[2] = val->constant->values[0].u32[2];
1158 }
1159
1160 static void
1161 vtn_handle_constant(struct vtn_builder *b, SpvOp opcode,
1162 const uint32_t *w, unsigned count)
1163 {
1164 struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_constant);
1165 val->const_type = vtn_value(b, w[1], vtn_value_type_type)->type->type;
1166 val->constant = rzalloc(b, nir_constant);
1167 switch (opcode) {
1168 case SpvOpConstantTrue:
1169 vtn_assert(val->const_type == glsl_bool_type());
1170 val->constant->values[0].u32[0] = NIR_TRUE;
1171 break;
1172 case SpvOpConstantFalse:
1173 vtn_assert(val->const_type == glsl_bool_type());
1174 val->constant->values[0].u32[0] = NIR_FALSE;
1175 break;
1176
1177 case SpvOpSpecConstantTrue:
1178 case SpvOpSpecConstantFalse: {
1179 vtn_assert(val->const_type == glsl_bool_type());
1180 uint32_t int_val =
1181 get_specialization(b, val, (opcode == SpvOpSpecConstantTrue));
1182 val->constant->values[0].u32[0] = int_val ? NIR_TRUE : NIR_FALSE;
1183 break;
1184 }
1185
1186 case SpvOpConstant: {
1187 vtn_assert(glsl_type_is_scalar(val->const_type));
1188 int bit_size = glsl_get_bit_size(val->const_type);
1189 if (bit_size == 64) {
1190 val->constant->values->u32[0] = w[3];
1191 val->constant->values->u32[1] = w[4];
1192 } else {
1193 vtn_assert(bit_size == 32);
1194 val->constant->values->u32[0] = w[3];
1195 }
1196 break;
1197 }
1198 case SpvOpSpecConstant: {
1199 vtn_assert(glsl_type_is_scalar(val->const_type));
1200 val->constant->values[0].u32[0] = get_specialization(b, val, w[3]);
1201 int bit_size = glsl_get_bit_size(val->const_type);
1202 if (bit_size == 64)
1203 val->constant->values[0].u64[0] =
1204 get_specialization64(b, val, vtn_u64_literal(&w[3]));
1205 else
1206 val->constant->values[0].u32[0] = get_specialization(b, val, w[3]);
1207 break;
1208 }
1209 case SpvOpSpecConstantComposite:
1210 case SpvOpConstantComposite: {
1211 unsigned elem_count = count - 3;
1212 nir_constant **elems = ralloc_array(b, nir_constant *, elem_count);
1213 for (unsigned i = 0; i < elem_count; i++)
1214 elems[i] = vtn_value(b, w[i + 3], vtn_value_type_constant)->constant;
1215
1216 switch (glsl_get_base_type(val->const_type)) {
1217 case GLSL_TYPE_UINT:
1218 case GLSL_TYPE_INT:
1219 case GLSL_TYPE_UINT64:
1220 case GLSL_TYPE_INT64:
1221 case GLSL_TYPE_FLOAT:
1222 case GLSL_TYPE_BOOL:
1223 case GLSL_TYPE_DOUBLE: {
1224 int bit_size = glsl_get_bit_size(val->const_type);
1225 if (glsl_type_is_matrix(val->const_type)) {
1226 vtn_assert(glsl_get_matrix_columns(val->const_type) == elem_count);
1227 for (unsigned i = 0; i < elem_count; i++)
1228 val->constant->values[i] = elems[i]->values[0];
1229 } else {
1230 vtn_assert(glsl_type_is_vector(val->const_type));
1231 vtn_assert(glsl_get_vector_elements(val->const_type) == elem_count);
1232 for (unsigned i = 0; i < elem_count; i++) {
1233 if (bit_size == 64) {
1234 val->constant->values[0].u64[i] = elems[i]->values[0].u64[0];
1235 } else {
1236 vtn_assert(bit_size == 32);
1237 val->constant->values[0].u32[i] = elems[i]->values[0].u32[0];
1238 }
1239 }
1240 }
1241 ralloc_free(elems);
1242 break;
1243 }
1244 case GLSL_TYPE_STRUCT:
1245 case GLSL_TYPE_ARRAY:
1246 ralloc_steal(val->constant, elems);
1247 val->constant->num_elements = elem_count;
1248 val->constant->elements = elems;
1249 break;
1250
1251 default:
1252 vtn_fail("Unsupported type for constants");
1253 }
1254 break;
1255 }
1256
1257 case SpvOpSpecConstantOp: {
1258 SpvOp opcode = get_specialization(b, val, w[3]);
1259 switch (opcode) {
1260 case SpvOpVectorShuffle: {
1261 struct vtn_value *v0 = &b->values[w[4]];
1262 struct vtn_value *v1 = &b->values[w[5]];
1263
1264 vtn_assert(v0->value_type == vtn_value_type_constant ||
1265 v0->value_type == vtn_value_type_undef);
1266 vtn_assert(v1->value_type == vtn_value_type_constant ||
1267 v1->value_type == vtn_value_type_undef);
1268
1269 unsigned len0 = v0->value_type == vtn_value_type_constant ?
1270 glsl_get_vector_elements(v0->const_type) :
1271 glsl_get_vector_elements(v0->type->type);
1272 unsigned len1 = v1->value_type == vtn_value_type_constant ?
1273 glsl_get_vector_elements(v1->const_type) :
1274 glsl_get_vector_elements(v1->type->type);
1275
1276 vtn_assert(len0 + len1 < 16);
1277
1278 unsigned bit_size = glsl_get_bit_size(val->const_type);
1279 unsigned bit_size0 = v0->value_type == vtn_value_type_constant ?
1280 glsl_get_bit_size(v0->const_type) :
1281 glsl_get_bit_size(v0->type->type);
1282 unsigned bit_size1 = v1->value_type == vtn_value_type_constant ?
1283 glsl_get_bit_size(v1->const_type) :
1284 glsl_get_bit_size(v1->type->type);
1285
1286 vtn_assert(bit_size == bit_size0 && bit_size == bit_size1);
1287 (void)bit_size0; (void)bit_size1;
1288
1289 if (bit_size == 64) {
1290 uint64_t u64[8];
1291 if (v0->value_type == vtn_value_type_constant) {
1292 for (unsigned i = 0; i < len0; i++)
1293 u64[i] = v0->constant->values[0].u64[i];
1294 }
1295 if (v1->value_type == vtn_value_type_constant) {
1296 for (unsigned i = 0; i < len1; i++)
1297 u64[len0 + i] = v1->constant->values[0].u64[i];
1298 }
1299
1300 for (unsigned i = 0, j = 0; i < count - 6; i++, j++) {
1301 uint32_t comp = w[i + 6];
1302 /* If component is not used, set the value to a known constant
1303 * to detect if it is wrongly used.
1304 */
1305 if (comp == (uint32_t)-1)
1306 val->constant->values[0].u64[j] = 0xdeadbeefdeadbeef;
1307 else
1308 val->constant->values[0].u64[j] = u64[comp];
1309 }
1310 } else {
1311 uint32_t u32[8];
1312 if (v0->value_type == vtn_value_type_constant) {
1313 for (unsigned i = 0; i < len0; i++)
1314 u32[i] = v0->constant->values[0].u32[i];
1315 }
1316 if (v1->value_type == vtn_value_type_constant) {
1317 for (unsigned i = 0; i < len1; i++)
1318 u32[len0 + i] = v1->constant->values[0].u32[i];
1319 }
1320
1321 for (unsigned i = 0, j = 0; i < count - 6; i++, j++) {
1322 uint32_t comp = w[i + 6];
1323 /* If component is not used, set the value to a known constant
1324 * to detect if it is wrongly used.
1325 */
1326 if (comp == (uint32_t)-1)
1327 val->constant->values[0].u32[j] = 0xdeadbeef;
1328 else
1329 val->constant->values[0].u32[j] = u32[comp];
1330 }
1331 }
1332 break;
1333 }
1334
1335 case SpvOpCompositeExtract:
1336 case SpvOpCompositeInsert: {
1337 struct vtn_value *comp;
1338 unsigned deref_start;
1339 struct nir_constant **c;
1340 if (opcode == SpvOpCompositeExtract) {
1341 comp = vtn_value(b, w[4], vtn_value_type_constant);
1342 deref_start = 5;
1343 c = &comp->constant;
1344 } else {
1345 comp = vtn_value(b, w[5], vtn_value_type_constant);
1346 deref_start = 6;
1347 val->constant = nir_constant_clone(comp->constant,
1348 (nir_variable *)b);
1349 c = &val->constant;
1350 }
1351
1352 int elem = -1;
1353 int col = 0;
1354 const struct glsl_type *type = comp->const_type;
1355 for (unsigned i = deref_start; i < count; i++) {
1356 switch (glsl_get_base_type(type)) {
1357 case GLSL_TYPE_UINT:
1358 case GLSL_TYPE_INT:
1359 case GLSL_TYPE_UINT64:
1360 case GLSL_TYPE_INT64:
1361 case GLSL_TYPE_FLOAT:
1362 case GLSL_TYPE_DOUBLE:
1363 case GLSL_TYPE_BOOL:
1364 /* If we hit this granularity, we're picking off an element */
1365 if (glsl_type_is_matrix(type)) {
1366 vtn_assert(col == 0 && elem == -1);
1367 col = w[i];
1368 elem = 0;
1369 type = glsl_get_column_type(type);
1370 } else {
1371 vtn_assert(elem <= 0 && glsl_type_is_vector(type));
1372 elem = w[i];
1373 type = glsl_scalar_type(glsl_get_base_type(type));
1374 }
1375 continue;
1376
1377 case GLSL_TYPE_ARRAY:
1378 c = &(*c)->elements[w[i]];
1379 type = glsl_get_array_element(type);
1380 continue;
1381
1382 case GLSL_TYPE_STRUCT:
1383 c = &(*c)->elements[w[i]];
1384 type = glsl_get_struct_field(type, w[i]);
1385 continue;
1386
1387 default:
1388 vtn_fail("Invalid constant type");
1389 }
1390 }
1391
1392 if (opcode == SpvOpCompositeExtract) {
1393 if (elem == -1) {
1394 val->constant = *c;
1395 } else {
1396 unsigned num_components = glsl_get_vector_elements(type);
1397 unsigned bit_size = glsl_get_bit_size(type);
1398 for (unsigned i = 0; i < num_components; i++)
1399 if (bit_size == 64) {
1400 val->constant->values[0].u64[i] = (*c)->values[col].u64[elem + i];
1401 } else {
1402 vtn_assert(bit_size == 32);
1403 val->constant->values[0].u32[i] = (*c)->values[col].u32[elem + i];
1404 }
1405 }
1406 } else {
1407 struct vtn_value *insert =
1408 vtn_value(b, w[4], vtn_value_type_constant);
1409 vtn_assert(insert->const_type == type);
1410 if (elem == -1) {
1411 *c = insert->constant;
1412 } else {
1413 unsigned num_components = glsl_get_vector_elements(type);
1414 unsigned bit_size = glsl_get_bit_size(type);
1415 for (unsigned i = 0; i < num_components; i++)
1416 if (bit_size == 64) {
1417 (*c)->values[col].u64[elem + i] = insert->constant->values[0].u64[i];
1418 } else {
1419 vtn_assert(bit_size == 32);
1420 (*c)->values[col].u32[elem + i] = insert->constant->values[0].u32[i];
1421 }
1422 }
1423 }
1424 break;
1425 }
1426
1427 default: {
1428 bool swap;
1429 nir_alu_type dst_alu_type = nir_get_nir_type_for_glsl_type(val->const_type);
1430 nir_alu_type src_alu_type = dst_alu_type;
1431 nir_op op = vtn_nir_alu_op_for_spirv_opcode(b, opcode, &swap,
1432 src_alu_type,
1433 dst_alu_type);
1434
1435 unsigned num_components = glsl_get_vector_elements(val->const_type);
1436 unsigned bit_size =
1437 glsl_get_bit_size(val->const_type);
1438
1439 nir_const_value src[4];
1440 vtn_assert(count <= 7);
1441 for (unsigned i = 0; i < count - 4; i++) {
1442 nir_constant *c =
1443 vtn_value(b, w[4 + i], vtn_value_type_constant)->constant;
1444
1445 unsigned j = swap ? 1 - i : i;
1446 vtn_assert(bit_size == 32);
1447 src[j] = c->values[0];
1448 }
1449
1450 val->constant->values[0] =
1451 nir_eval_const_opcode(op, num_components, bit_size, src);
1452 break;
1453 } /* default */
1454 }
1455 break;
1456 }
1457
1458 case SpvOpConstantNull:
1459 val->constant = vtn_null_constant(b, val->const_type);
1460 break;
1461
1462 case SpvOpConstantSampler:
1463 vtn_fail("OpConstantSampler requires Kernel Capability");
1464 break;
1465
1466 default:
1467 vtn_fail("Unhandled opcode");
1468 }
1469
1470 /* Now that we have the value, update the workgroup size if needed */
1471 vtn_foreach_decoration(b, val, handle_workgroup_size_decoration_cb, NULL);
1472 }
1473
1474 static void
1475 vtn_handle_function_call(struct vtn_builder *b, SpvOp opcode,
1476 const uint32_t *w, unsigned count)
1477 {
1478 struct vtn_type *res_type = vtn_value(b, w[1], vtn_value_type_type)->type;
1479 struct vtn_function *vtn_callee =
1480 vtn_value(b, w[3], vtn_value_type_function)->func;
1481 struct nir_function *callee = vtn_callee->impl->function;
1482
1483 vtn_callee->referenced = true;
1484
1485 nir_call_instr *call = nir_call_instr_create(b->nb.shader, callee);
1486 for (unsigned i = 0; i < call->num_params; i++) {
1487 unsigned arg_id = w[4 + i];
1488 struct vtn_value *arg = vtn_untyped_value(b, arg_id);
1489 if (arg->value_type == vtn_value_type_pointer &&
1490 arg->pointer->ptr_type->type == NULL) {
1491 nir_deref_var *d = vtn_pointer_to_deref(b, arg->pointer);
1492 call->params[i] = nir_deref_var_clone(d, call);
1493 } else {
1494 struct vtn_ssa_value *arg_ssa = vtn_ssa_value(b, arg_id);
1495
1496 /* Make a temporary to store the argument in */
1497 nir_variable *tmp =
1498 nir_local_variable_create(b->nb.impl, arg_ssa->type, "arg_tmp");
1499 call->params[i] = nir_deref_var_create(call, tmp);
1500
1501 vtn_local_store(b, arg_ssa, call->params[i]);
1502 }
1503 }
1504
1505 nir_variable *out_tmp = NULL;
1506 vtn_assert(res_type->type == callee->return_type);
1507 if (!glsl_type_is_void(callee->return_type)) {
1508 out_tmp = nir_local_variable_create(b->nb.impl, callee->return_type,
1509 "out_tmp");
1510 call->return_deref = nir_deref_var_create(call, out_tmp);
1511 }
1512
1513 nir_builder_instr_insert(&b->nb, &call->instr);
1514
1515 if (glsl_type_is_void(callee->return_type)) {
1516 vtn_push_value(b, w[2], vtn_value_type_undef);
1517 } else {
1518 vtn_push_ssa(b, w[2], res_type, vtn_local_load(b, call->return_deref));
1519 }
1520 }
1521
1522 struct vtn_ssa_value *
1523 vtn_create_ssa_value(struct vtn_builder *b, const struct glsl_type *type)
1524 {
1525 struct vtn_ssa_value *val = rzalloc(b, struct vtn_ssa_value);
1526 val->type = type;
1527
1528 if (!glsl_type_is_vector_or_scalar(type)) {
1529 unsigned elems = glsl_get_length(type);
1530 val->elems = ralloc_array(b, struct vtn_ssa_value *, elems);
1531 for (unsigned i = 0; i < elems; i++) {
1532 const struct glsl_type *child_type;
1533
1534 switch (glsl_get_base_type(type)) {
1535 case GLSL_TYPE_INT:
1536 case GLSL_TYPE_UINT:
1537 case GLSL_TYPE_INT64:
1538 case GLSL_TYPE_UINT64:
1539 case GLSL_TYPE_BOOL:
1540 case GLSL_TYPE_FLOAT:
1541 case GLSL_TYPE_DOUBLE:
1542 child_type = glsl_get_column_type(type);
1543 break;
1544 case GLSL_TYPE_ARRAY:
1545 child_type = glsl_get_array_element(type);
1546 break;
1547 case GLSL_TYPE_STRUCT:
1548 child_type = glsl_get_struct_field(type, i);
1549 break;
1550 default:
1551 vtn_fail("unkown base type");
1552 }
1553
1554 val->elems[i] = vtn_create_ssa_value(b, child_type);
1555 }
1556 }
1557
1558 return val;
1559 }
1560
1561 static nir_tex_src
1562 vtn_tex_src(struct vtn_builder *b, unsigned index, nir_tex_src_type type)
1563 {
1564 nir_tex_src src;
1565 src.src = nir_src_for_ssa(vtn_ssa_value(b, index)->def);
1566 src.src_type = type;
1567 return src;
1568 }
1569
1570 static void
1571 vtn_handle_texture(struct vtn_builder *b, SpvOp opcode,
1572 const uint32_t *w, unsigned count)
1573 {
1574 if (opcode == SpvOpSampledImage) {
1575 struct vtn_value *val =
1576 vtn_push_value(b, w[2], vtn_value_type_sampled_image);
1577 val->sampled_image = ralloc(b, struct vtn_sampled_image);
1578 val->sampled_image->type =
1579 vtn_value(b, w[1], vtn_value_type_type)->type;
1580 val->sampled_image->image =
1581 vtn_value(b, w[3], vtn_value_type_pointer)->pointer;
1582 val->sampled_image->sampler =
1583 vtn_value(b, w[4], vtn_value_type_pointer)->pointer;
1584 return;
1585 } else if (opcode == SpvOpImage) {
1586 struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_pointer);
1587 struct vtn_value *src_val = vtn_untyped_value(b, w[3]);
1588 if (src_val->value_type == vtn_value_type_sampled_image) {
1589 val->pointer = src_val->sampled_image->image;
1590 } else {
1591 vtn_assert(src_val->value_type == vtn_value_type_pointer);
1592 val->pointer = src_val->pointer;
1593 }
1594 return;
1595 }
1596
1597 struct vtn_type *ret_type = vtn_value(b, w[1], vtn_value_type_type)->type;
1598 struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_ssa);
1599
1600 struct vtn_sampled_image sampled;
1601 struct vtn_value *sampled_val = vtn_untyped_value(b, w[3]);
1602 if (sampled_val->value_type == vtn_value_type_sampled_image) {
1603 sampled = *sampled_val->sampled_image;
1604 } else {
1605 vtn_assert(sampled_val->value_type == vtn_value_type_pointer);
1606 sampled.type = sampled_val->pointer->type;
1607 sampled.image = NULL;
1608 sampled.sampler = sampled_val->pointer;
1609 }
1610
1611 const struct glsl_type *image_type = sampled.type->type;
1612 const enum glsl_sampler_dim sampler_dim = glsl_get_sampler_dim(image_type);
1613 const bool is_array = glsl_sampler_type_is_array(image_type);
1614 const bool is_shadow = glsl_sampler_type_is_shadow(image_type);
1615
1616 /* Figure out the base texture operation */
1617 nir_texop texop;
1618 switch (opcode) {
1619 case SpvOpImageSampleImplicitLod:
1620 case SpvOpImageSampleDrefImplicitLod:
1621 case SpvOpImageSampleProjImplicitLod:
1622 case SpvOpImageSampleProjDrefImplicitLod:
1623 texop = nir_texop_tex;
1624 break;
1625
1626 case SpvOpImageSampleExplicitLod:
1627 case SpvOpImageSampleDrefExplicitLod:
1628 case SpvOpImageSampleProjExplicitLod:
1629 case SpvOpImageSampleProjDrefExplicitLod:
1630 texop = nir_texop_txl;
1631 break;
1632
1633 case SpvOpImageFetch:
1634 if (glsl_get_sampler_dim(image_type) == GLSL_SAMPLER_DIM_MS) {
1635 texop = nir_texop_txf_ms;
1636 } else {
1637 texop = nir_texop_txf;
1638 }
1639 break;
1640
1641 case SpvOpImageGather:
1642 case SpvOpImageDrefGather:
1643 texop = nir_texop_tg4;
1644 break;
1645
1646 case SpvOpImageQuerySizeLod:
1647 case SpvOpImageQuerySize:
1648 texop = nir_texop_txs;
1649 break;
1650
1651 case SpvOpImageQueryLod:
1652 texop = nir_texop_lod;
1653 break;
1654
1655 case SpvOpImageQueryLevels:
1656 texop = nir_texop_query_levels;
1657 break;
1658
1659 case SpvOpImageQuerySamples:
1660 texop = nir_texop_texture_samples;
1661 break;
1662
1663 default:
1664 vtn_fail("Unhandled opcode");
1665 }
1666
1667 nir_tex_src srcs[8]; /* 8 should be enough */
1668 nir_tex_src *p = srcs;
1669
1670 unsigned idx = 4;
1671
1672 struct nir_ssa_def *coord;
1673 unsigned coord_components;
1674 switch (opcode) {
1675 case SpvOpImageSampleImplicitLod:
1676 case SpvOpImageSampleExplicitLod:
1677 case SpvOpImageSampleDrefImplicitLod:
1678 case SpvOpImageSampleDrefExplicitLod:
1679 case SpvOpImageSampleProjImplicitLod:
1680 case SpvOpImageSampleProjExplicitLod:
1681 case SpvOpImageSampleProjDrefImplicitLod:
1682 case SpvOpImageSampleProjDrefExplicitLod:
1683 case SpvOpImageFetch:
1684 case SpvOpImageGather:
1685 case SpvOpImageDrefGather:
1686 case SpvOpImageQueryLod: {
1687 /* All these types have the coordinate as their first real argument */
1688 switch (sampler_dim) {
1689 case GLSL_SAMPLER_DIM_1D:
1690 case GLSL_SAMPLER_DIM_BUF:
1691 coord_components = 1;
1692 break;
1693 case GLSL_SAMPLER_DIM_2D:
1694 case GLSL_SAMPLER_DIM_RECT:
1695 case GLSL_SAMPLER_DIM_MS:
1696 coord_components = 2;
1697 break;
1698 case GLSL_SAMPLER_DIM_3D:
1699 case GLSL_SAMPLER_DIM_CUBE:
1700 coord_components = 3;
1701 break;
1702 default:
1703 vtn_fail("Invalid sampler type");
1704 }
1705
1706 if (is_array && texop != nir_texop_lod)
1707 coord_components++;
1708
1709 coord = vtn_ssa_value(b, w[idx++])->def;
1710 p->src = nir_src_for_ssa(nir_channels(&b->nb, coord,
1711 (1 << coord_components) - 1));
1712 p->src_type = nir_tex_src_coord;
1713 p++;
1714 break;
1715 }
1716
1717 default:
1718 coord = NULL;
1719 coord_components = 0;
1720 break;
1721 }
1722
1723 switch (opcode) {
1724 case SpvOpImageSampleProjImplicitLod:
1725 case SpvOpImageSampleProjExplicitLod:
1726 case SpvOpImageSampleProjDrefImplicitLod:
1727 case SpvOpImageSampleProjDrefExplicitLod:
1728 /* These have the projector as the last coordinate component */
1729 p->src = nir_src_for_ssa(nir_channel(&b->nb, coord, coord_components));
1730 p->src_type = nir_tex_src_projector;
1731 p++;
1732 break;
1733
1734 default:
1735 break;
1736 }
1737
1738 unsigned gather_component = 0;
1739 switch (opcode) {
1740 case SpvOpImageSampleDrefImplicitLod:
1741 case SpvOpImageSampleDrefExplicitLod:
1742 case SpvOpImageSampleProjDrefImplicitLod:
1743 case SpvOpImageSampleProjDrefExplicitLod:
1744 case SpvOpImageDrefGather:
1745 /* These all have an explicit depth value as their next source */
1746 (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_comparator);
1747 break;
1748
1749 case SpvOpImageGather:
1750 /* This has a component as its next source */
1751 gather_component =
1752 vtn_value(b, w[idx++], vtn_value_type_constant)->constant->values[0].u32[0];
1753 break;
1754
1755 default:
1756 break;
1757 }
1758
1759 /* For OpImageQuerySizeLod, we always have an LOD */
1760 if (opcode == SpvOpImageQuerySizeLod)
1761 (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_lod);
1762
1763 /* Now we need to handle some number of optional arguments */
1764 const struct vtn_ssa_value *gather_offsets = NULL;
1765 if (idx < count) {
1766 uint32_t operands = w[idx++];
1767
1768 if (operands & SpvImageOperandsBiasMask) {
1769 vtn_assert(texop == nir_texop_tex);
1770 texop = nir_texop_txb;
1771 (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_bias);
1772 }
1773
1774 if (operands & SpvImageOperandsLodMask) {
1775 vtn_assert(texop == nir_texop_txl || texop == nir_texop_txf ||
1776 texop == nir_texop_txs);
1777 (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_lod);
1778 }
1779
1780 if (operands & SpvImageOperandsGradMask) {
1781 vtn_assert(texop == nir_texop_txl);
1782 texop = nir_texop_txd;
1783 (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_ddx);
1784 (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_ddy);
1785 }
1786
1787 if (operands & SpvImageOperandsOffsetMask ||
1788 operands & SpvImageOperandsConstOffsetMask)
1789 (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_offset);
1790
1791 if (operands & SpvImageOperandsConstOffsetsMask) {
1792 gather_offsets = vtn_ssa_value(b, w[idx++]);
1793 (*p++) = (nir_tex_src){};
1794 }
1795
1796 if (operands & SpvImageOperandsSampleMask) {
1797 vtn_assert(texop == nir_texop_txf_ms);
1798 texop = nir_texop_txf_ms;
1799 (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_ms_index);
1800 }
1801 }
1802 /* We should have now consumed exactly all of the arguments */
1803 vtn_assert(idx == count);
1804
1805 nir_tex_instr *instr = nir_tex_instr_create(b->shader, p - srcs);
1806 instr->op = texop;
1807
1808 memcpy(instr->src, srcs, instr->num_srcs * sizeof(*instr->src));
1809
1810 instr->coord_components = coord_components;
1811 instr->sampler_dim = sampler_dim;
1812 instr->is_array = is_array;
1813 instr->is_shadow = is_shadow;
1814 instr->is_new_style_shadow =
1815 is_shadow && glsl_get_components(ret_type->type) == 1;
1816 instr->component = gather_component;
1817
1818 switch (glsl_get_sampler_result_type(image_type)) {
1819 case GLSL_TYPE_FLOAT: instr->dest_type = nir_type_float; break;
1820 case GLSL_TYPE_INT: instr->dest_type = nir_type_int; break;
1821 case GLSL_TYPE_UINT: instr->dest_type = nir_type_uint; break;
1822 case GLSL_TYPE_BOOL: instr->dest_type = nir_type_bool; break;
1823 default:
1824 vtn_fail("Invalid base type for sampler result");
1825 }
1826
1827 nir_deref_var *sampler = vtn_pointer_to_deref(b, sampled.sampler);
1828 nir_deref_var *texture;
1829 if (sampled.image) {
1830 nir_deref_var *image = vtn_pointer_to_deref(b, sampled.image);
1831 texture = image;
1832 } else {
1833 texture = sampler;
1834 }
1835
1836 instr->texture = nir_deref_var_clone(texture, instr);
1837
1838 switch (instr->op) {
1839 case nir_texop_tex:
1840 case nir_texop_txb:
1841 case nir_texop_txl:
1842 case nir_texop_txd:
1843 case nir_texop_tg4:
1844 /* These operations require a sampler */
1845 instr->sampler = nir_deref_var_clone(sampler, instr);
1846 break;
1847 case nir_texop_txf:
1848 case nir_texop_txf_ms:
1849 case nir_texop_txs:
1850 case nir_texop_lod:
1851 case nir_texop_query_levels:
1852 case nir_texop_texture_samples:
1853 case nir_texop_samples_identical:
1854 /* These don't */
1855 instr->sampler = NULL;
1856 break;
1857 case nir_texop_txf_ms_mcs:
1858 vtn_fail("unexpected nir_texop_txf_ms_mcs");
1859 }
1860
1861 nir_ssa_dest_init(&instr->instr, &instr->dest,
1862 nir_tex_instr_dest_size(instr), 32, NULL);
1863
1864 vtn_assert(glsl_get_vector_elements(ret_type->type) ==
1865 nir_tex_instr_dest_size(instr));
1866
1867 nir_ssa_def *def;
1868 nir_instr *instruction;
1869 if (gather_offsets) {
1870 vtn_assert(glsl_get_base_type(gather_offsets->type) == GLSL_TYPE_ARRAY);
1871 vtn_assert(glsl_get_length(gather_offsets->type) == 4);
1872 nir_tex_instr *instrs[4] = {instr, NULL, NULL, NULL};
1873
1874 /* Copy the current instruction 4x */
1875 for (uint32_t i = 1; i < 4; i++) {
1876 instrs[i] = nir_tex_instr_create(b->shader, instr->num_srcs);
1877 instrs[i]->op = instr->op;
1878 instrs[i]->coord_components = instr->coord_components;
1879 instrs[i]->sampler_dim = instr->sampler_dim;
1880 instrs[i]->is_array = instr->is_array;
1881 instrs[i]->is_shadow = instr->is_shadow;
1882 instrs[i]->is_new_style_shadow = instr->is_new_style_shadow;
1883 instrs[i]->component = instr->component;
1884 instrs[i]->dest_type = instr->dest_type;
1885 instrs[i]->texture = nir_deref_var_clone(texture, instrs[i]);
1886 instrs[i]->sampler = NULL;
1887
1888 memcpy(instrs[i]->src, srcs, instr->num_srcs * sizeof(*instr->src));
1889
1890 nir_ssa_dest_init(&instrs[i]->instr, &instrs[i]->dest,
1891 nir_tex_instr_dest_size(instr), 32, NULL);
1892 }
1893
1894 /* Fill in the last argument with the offset from the passed in offsets
1895 * and insert the instruction into the stream.
1896 */
1897 for (uint32_t i = 0; i < 4; i++) {
1898 nir_tex_src src;
1899 src.src = nir_src_for_ssa(gather_offsets->elems[i]->def);
1900 src.src_type = nir_tex_src_offset;
1901 instrs[i]->src[instrs[i]->num_srcs - 1] = src;
1902 nir_builder_instr_insert(&b->nb, &instrs[i]->instr);
1903 }
1904
1905 /* Combine the results of the 4 instructions by taking their .w
1906 * components
1907 */
1908 nir_alu_instr *vec4 = nir_alu_instr_create(b->shader, nir_op_vec4);
1909 nir_ssa_dest_init(&vec4->instr, &vec4->dest.dest, 4, 32, NULL);
1910 vec4->dest.write_mask = 0xf;
1911 for (uint32_t i = 0; i < 4; i++) {
1912 vec4->src[i].src = nir_src_for_ssa(&instrs[i]->dest.ssa);
1913 vec4->src[i].swizzle[0] = 3;
1914 }
1915 def = &vec4->dest.dest.ssa;
1916 instruction = &vec4->instr;
1917 } else {
1918 def = &instr->dest.ssa;
1919 instruction = &instr->instr;
1920 }
1921
1922 val->ssa = vtn_create_ssa_value(b, ret_type->type);
1923 val->ssa->def = def;
1924
1925 nir_builder_instr_insert(&b->nb, instruction);
1926 }
1927
1928 static void
1929 fill_common_atomic_sources(struct vtn_builder *b, SpvOp opcode,
1930 const uint32_t *w, nir_src *src)
1931 {
1932 switch (opcode) {
1933 case SpvOpAtomicIIncrement:
1934 src[0] = nir_src_for_ssa(nir_imm_int(&b->nb, 1));
1935 break;
1936
1937 case SpvOpAtomicIDecrement:
1938 src[0] = nir_src_for_ssa(nir_imm_int(&b->nb, -1));
1939 break;
1940
1941 case SpvOpAtomicISub:
1942 src[0] =
1943 nir_src_for_ssa(nir_ineg(&b->nb, vtn_ssa_value(b, w[6])->def));
1944 break;
1945
1946 case SpvOpAtomicCompareExchange:
1947 src[0] = nir_src_for_ssa(vtn_ssa_value(b, w[8])->def);
1948 src[1] = nir_src_for_ssa(vtn_ssa_value(b, w[7])->def);
1949 break;
1950
1951 case SpvOpAtomicExchange:
1952 case SpvOpAtomicIAdd:
1953 case SpvOpAtomicSMin:
1954 case SpvOpAtomicUMin:
1955 case SpvOpAtomicSMax:
1956 case SpvOpAtomicUMax:
1957 case SpvOpAtomicAnd:
1958 case SpvOpAtomicOr:
1959 case SpvOpAtomicXor:
1960 src[0] = nir_src_for_ssa(vtn_ssa_value(b, w[6])->def);
1961 break;
1962
1963 default:
1964 vtn_fail("Invalid SPIR-V atomic");
1965 }
1966 }
1967
1968 static nir_ssa_def *
1969 get_image_coord(struct vtn_builder *b, uint32_t value)
1970 {
1971 struct vtn_ssa_value *coord = vtn_ssa_value(b, value);
1972
1973 /* The image_load_store intrinsics assume a 4-dim coordinate */
1974 unsigned dim = glsl_get_vector_elements(coord->type);
1975 unsigned swizzle[4];
1976 for (unsigned i = 0; i < 4; i++)
1977 swizzle[i] = MIN2(i, dim - 1);
1978
1979 return nir_swizzle(&b->nb, coord->def, swizzle, 4, false);
1980 }
1981
1982 static void
1983 vtn_handle_image(struct vtn_builder *b, SpvOp opcode,
1984 const uint32_t *w, unsigned count)
1985 {
1986 /* Just get this one out of the way */
1987 if (opcode == SpvOpImageTexelPointer) {
1988 struct vtn_value *val =
1989 vtn_push_value(b, w[2], vtn_value_type_image_pointer);
1990 val->image = ralloc(b, struct vtn_image_pointer);
1991
1992 val->image->image = vtn_value(b, w[3], vtn_value_type_pointer)->pointer;
1993 val->image->coord = get_image_coord(b, w[4]);
1994 val->image->sample = vtn_ssa_value(b, w[5])->def;
1995 return;
1996 }
1997
1998 struct vtn_image_pointer image;
1999
2000 switch (opcode) {
2001 case SpvOpAtomicExchange:
2002 case SpvOpAtomicCompareExchange:
2003 case SpvOpAtomicCompareExchangeWeak:
2004 case SpvOpAtomicIIncrement:
2005 case SpvOpAtomicIDecrement:
2006 case SpvOpAtomicIAdd:
2007 case SpvOpAtomicISub:
2008 case SpvOpAtomicLoad:
2009 case SpvOpAtomicSMin:
2010 case SpvOpAtomicUMin:
2011 case SpvOpAtomicSMax:
2012 case SpvOpAtomicUMax:
2013 case SpvOpAtomicAnd:
2014 case SpvOpAtomicOr:
2015 case SpvOpAtomicXor:
2016 image = *vtn_value(b, w[3], vtn_value_type_image_pointer)->image;
2017 break;
2018
2019 case SpvOpAtomicStore:
2020 image = *vtn_value(b, w[1], vtn_value_type_image_pointer)->image;
2021 break;
2022
2023 case SpvOpImageQuerySize:
2024 image.image = vtn_value(b, w[3], vtn_value_type_pointer)->pointer;
2025 image.coord = NULL;
2026 image.sample = NULL;
2027 break;
2028
2029 case SpvOpImageRead:
2030 image.image = vtn_value(b, w[3], vtn_value_type_pointer)->pointer;
2031 image.coord = get_image_coord(b, w[4]);
2032
2033 if (count > 5 && (w[5] & SpvImageOperandsSampleMask)) {
2034 vtn_assert(w[5] == SpvImageOperandsSampleMask);
2035 image.sample = vtn_ssa_value(b, w[6])->def;
2036 } else {
2037 image.sample = nir_ssa_undef(&b->nb, 1, 32);
2038 }
2039 break;
2040
2041 case SpvOpImageWrite:
2042 image.image = vtn_value(b, w[1], vtn_value_type_pointer)->pointer;
2043 image.coord = get_image_coord(b, w[2]);
2044
2045 /* texel = w[3] */
2046
2047 if (count > 4 && (w[4] & SpvImageOperandsSampleMask)) {
2048 vtn_assert(w[4] == SpvImageOperandsSampleMask);
2049 image.sample = vtn_ssa_value(b, w[5])->def;
2050 } else {
2051 image.sample = nir_ssa_undef(&b->nb, 1, 32);
2052 }
2053 break;
2054
2055 default:
2056 vtn_fail("Invalid image opcode");
2057 }
2058
2059 nir_intrinsic_op op;
2060 switch (opcode) {
2061 #define OP(S, N) case SpvOp##S: op = nir_intrinsic_image_##N; break;
2062 OP(ImageQuerySize, size)
2063 OP(ImageRead, load)
2064 OP(ImageWrite, store)
2065 OP(AtomicLoad, load)
2066 OP(AtomicStore, store)
2067 OP(AtomicExchange, atomic_exchange)
2068 OP(AtomicCompareExchange, atomic_comp_swap)
2069 OP(AtomicIIncrement, atomic_add)
2070 OP(AtomicIDecrement, atomic_add)
2071 OP(AtomicIAdd, atomic_add)
2072 OP(AtomicISub, atomic_add)
2073 OP(AtomicSMin, atomic_min)
2074 OP(AtomicUMin, atomic_min)
2075 OP(AtomicSMax, atomic_max)
2076 OP(AtomicUMax, atomic_max)
2077 OP(AtomicAnd, atomic_and)
2078 OP(AtomicOr, atomic_or)
2079 OP(AtomicXor, atomic_xor)
2080 #undef OP
2081 default:
2082 vtn_fail("Invalid image opcode");
2083 }
2084
2085 nir_intrinsic_instr *intrin = nir_intrinsic_instr_create(b->shader, op);
2086
2087 nir_deref_var *image_deref = vtn_pointer_to_deref(b, image.image);
2088 intrin->variables[0] = nir_deref_var_clone(image_deref, intrin);
2089
2090 /* ImageQuerySize doesn't take any extra parameters */
2091 if (opcode != SpvOpImageQuerySize) {
2092 /* The image coordinate is always 4 components but we may not have that
2093 * many. Swizzle to compensate.
2094 */
2095 unsigned swiz[4];
2096 for (unsigned i = 0; i < 4; i++)
2097 swiz[i] = i < image.coord->num_components ? i : 0;
2098 intrin->src[0] = nir_src_for_ssa(nir_swizzle(&b->nb, image.coord,
2099 swiz, 4, false));
2100 intrin->src[1] = nir_src_for_ssa(image.sample);
2101 }
2102
2103 switch (opcode) {
2104 case SpvOpAtomicLoad:
2105 case SpvOpImageQuerySize:
2106 case SpvOpImageRead:
2107 break;
2108 case SpvOpAtomicStore:
2109 intrin->src[2] = nir_src_for_ssa(vtn_ssa_value(b, w[4])->def);
2110 break;
2111 case SpvOpImageWrite:
2112 intrin->src[2] = nir_src_for_ssa(vtn_ssa_value(b, w[3])->def);
2113 break;
2114
2115 case SpvOpAtomicCompareExchange:
2116 case SpvOpAtomicIIncrement:
2117 case SpvOpAtomicIDecrement:
2118 case SpvOpAtomicExchange:
2119 case SpvOpAtomicIAdd:
2120 case SpvOpAtomicISub:
2121 case SpvOpAtomicSMin:
2122 case SpvOpAtomicUMin:
2123 case SpvOpAtomicSMax:
2124 case SpvOpAtomicUMax:
2125 case SpvOpAtomicAnd:
2126 case SpvOpAtomicOr:
2127 case SpvOpAtomicXor:
2128 fill_common_atomic_sources(b, opcode, w, &intrin->src[2]);
2129 break;
2130
2131 default:
2132 vtn_fail("Invalid image opcode");
2133 }
2134
2135 if (opcode != SpvOpImageWrite) {
2136 struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_ssa);
2137 struct vtn_type *type = vtn_value(b, w[1], vtn_value_type_type)->type;
2138
2139 unsigned dest_components =
2140 nir_intrinsic_infos[intrin->intrinsic].dest_components;
2141 if (intrin->intrinsic == nir_intrinsic_image_size) {
2142 dest_components = intrin->num_components =
2143 glsl_get_vector_elements(type->type);
2144 }
2145
2146 nir_ssa_dest_init(&intrin->instr, &intrin->dest,
2147 dest_components, 32, NULL);
2148
2149 nir_builder_instr_insert(&b->nb, &intrin->instr);
2150
2151 val->ssa = vtn_create_ssa_value(b, type->type);
2152 val->ssa->def = &intrin->dest.ssa;
2153 } else {
2154 nir_builder_instr_insert(&b->nb, &intrin->instr);
2155 }
2156 }
2157
2158 static nir_intrinsic_op
2159 get_ssbo_nir_atomic_op(struct vtn_builder *b, SpvOp opcode)
2160 {
2161 switch (opcode) {
2162 case SpvOpAtomicLoad: return nir_intrinsic_load_ssbo;
2163 case SpvOpAtomicStore: return nir_intrinsic_store_ssbo;
2164 #define OP(S, N) case SpvOp##S: return nir_intrinsic_ssbo_##N;
2165 OP(AtomicExchange, atomic_exchange)
2166 OP(AtomicCompareExchange, atomic_comp_swap)
2167 OP(AtomicIIncrement, atomic_add)
2168 OP(AtomicIDecrement, atomic_add)
2169 OP(AtomicIAdd, atomic_add)
2170 OP(AtomicISub, atomic_add)
2171 OP(AtomicSMin, atomic_imin)
2172 OP(AtomicUMin, atomic_umin)
2173 OP(AtomicSMax, atomic_imax)
2174 OP(AtomicUMax, atomic_umax)
2175 OP(AtomicAnd, atomic_and)
2176 OP(AtomicOr, atomic_or)
2177 OP(AtomicXor, atomic_xor)
2178 #undef OP
2179 default:
2180 vtn_fail("Invalid SSBO atomic");
2181 }
2182 }
2183
2184 static nir_intrinsic_op
2185 get_shared_nir_atomic_op(struct vtn_builder *b, SpvOp opcode)
2186 {
2187 switch (opcode) {
2188 case SpvOpAtomicLoad: return nir_intrinsic_load_var;
2189 case SpvOpAtomicStore: return nir_intrinsic_store_var;
2190 #define OP(S, N) case SpvOp##S: return nir_intrinsic_var_##N;
2191 OP(AtomicExchange, atomic_exchange)
2192 OP(AtomicCompareExchange, atomic_comp_swap)
2193 OP(AtomicIIncrement, atomic_add)
2194 OP(AtomicIDecrement, atomic_add)
2195 OP(AtomicIAdd, atomic_add)
2196 OP(AtomicISub, atomic_add)
2197 OP(AtomicSMin, atomic_imin)
2198 OP(AtomicUMin, atomic_umin)
2199 OP(AtomicSMax, atomic_imax)
2200 OP(AtomicUMax, atomic_umax)
2201 OP(AtomicAnd, atomic_and)
2202 OP(AtomicOr, atomic_or)
2203 OP(AtomicXor, atomic_xor)
2204 #undef OP
2205 default:
2206 vtn_fail("Invalid shared atomic");
2207 }
2208 }
2209
2210 static void
2211 vtn_handle_ssbo_or_shared_atomic(struct vtn_builder *b, SpvOp opcode,
2212 const uint32_t *w, unsigned count)
2213 {
2214 struct vtn_pointer *ptr;
2215 nir_intrinsic_instr *atomic;
2216
2217 switch (opcode) {
2218 case SpvOpAtomicLoad:
2219 case SpvOpAtomicExchange:
2220 case SpvOpAtomicCompareExchange:
2221 case SpvOpAtomicCompareExchangeWeak:
2222 case SpvOpAtomicIIncrement:
2223 case SpvOpAtomicIDecrement:
2224 case SpvOpAtomicIAdd:
2225 case SpvOpAtomicISub:
2226 case SpvOpAtomicSMin:
2227 case SpvOpAtomicUMin:
2228 case SpvOpAtomicSMax:
2229 case SpvOpAtomicUMax:
2230 case SpvOpAtomicAnd:
2231 case SpvOpAtomicOr:
2232 case SpvOpAtomicXor:
2233 ptr = vtn_value(b, w[3], vtn_value_type_pointer)->pointer;
2234 break;
2235
2236 case SpvOpAtomicStore:
2237 ptr = vtn_value(b, w[1], vtn_value_type_pointer)->pointer;
2238 break;
2239
2240 default:
2241 vtn_fail("Invalid SPIR-V atomic");
2242 }
2243
2244 /*
2245 SpvScope scope = w[4];
2246 SpvMemorySemanticsMask semantics = w[5];
2247 */
2248
2249 if (ptr->mode == vtn_variable_mode_workgroup) {
2250 nir_deref_var *deref = vtn_pointer_to_deref(b, ptr);
2251 const struct glsl_type *deref_type = nir_deref_tail(&deref->deref)->type;
2252 nir_intrinsic_op op = get_shared_nir_atomic_op(b, opcode);
2253 atomic = nir_intrinsic_instr_create(b->nb.shader, op);
2254 atomic->variables[0] = nir_deref_var_clone(deref, atomic);
2255
2256 switch (opcode) {
2257 case SpvOpAtomicLoad:
2258 atomic->num_components = glsl_get_vector_elements(deref_type);
2259 break;
2260
2261 case SpvOpAtomicStore:
2262 atomic->num_components = glsl_get_vector_elements(deref_type);
2263 nir_intrinsic_set_write_mask(atomic, (1 << atomic->num_components) - 1);
2264 atomic->src[0] = nir_src_for_ssa(vtn_ssa_value(b, w[4])->def);
2265 break;
2266
2267 case SpvOpAtomicExchange:
2268 case SpvOpAtomicCompareExchange:
2269 case SpvOpAtomicCompareExchangeWeak:
2270 case SpvOpAtomicIIncrement:
2271 case SpvOpAtomicIDecrement:
2272 case SpvOpAtomicIAdd:
2273 case SpvOpAtomicISub:
2274 case SpvOpAtomicSMin:
2275 case SpvOpAtomicUMin:
2276 case SpvOpAtomicSMax:
2277 case SpvOpAtomicUMax:
2278 case SpvOpAtomicAnd:
2279 case SpvOpAtomicOr:
2280 case SpvOpAtomicXor:
2281 fill_common_atomic_sources(b, opcode, w, &atomic->src[0]);
2282 break;
2283
2284 default:
2285 vtn_fail("Invalid SPIR-V atomic");
2286
2287 }
2288 } else {
2289 vtn_assert(ptr->mode == vtn_variable_mode_ssbo);
2290 nir_ssa_def *offset, *index;
2291 offset = vtn_pointer_to_offset(b, ptr, &index, NULL);
2292
2293 nir_intrinsic_op op = get_ssbo_nir_atomic_op(b, opcode);
2294
2295 atomic = nir_intrinsic_instr_create(b->nb.shader, op);
2296
2297 switch (opcode) {
2298 case SpvOpAtomicLoad:
2299 atomic->num_components = glsl_get_vector_elements(ptr->type->type);
2300 atomic->src[0] = nir_src_for_ssa(index);
2301 atomic->src[1] = nir_src_for_ssa(offset);
2302 break;
2303
2304 case SpvOpAtomicStore:
2305 atomic->num_components = glsl_get_vector_elements(ptr->type->type);
2306 nir_intrinsic_set_write_mask(atomic, (1 << atomic->num_components) - 1);
2307 atomic->src[0] = nir_src_for_ssa(vtn_ssa_value(b, w[4])->def);
2308 atomic->src[1] = nir_src_for_ssa(index);
2309 atomic->src[2] = nir_src_for_ssa(offset);
2310 break;
2311
2312 case SpvOpAtomicExchange:
2313 case SpvOpAtomicCompareExchange:
2314 case SpvOpAtomicCompareExchangeWeak:
2315 case SpvOpAtomicIIncrement:
2316 case SpvOpAtomicIDecrement:
2317 case SpvOpAtomicIAdd:
2318 case SpvOpAtomicISub:
2319 case SpvOpAtomicSMin:
2320 case SpvOpAtomicUMin:
2321 case SpvOpAtomicSMax:
2322 case SpvOpAtomicUMax:
2323 case SpvOpAtomicAnd:
2324 case SpvOpAtomicOr:
2325 case SpvOpAtomicXor:
2326 atomic->src[0] = nir_src_for_ssa(index);
2327 atomic->src[1] = nir_src_for_ssa(offset);
2328 fill_common_atomic_sources(b, opcode, w, &atomic->src[2]);
2329 break;
2330
2331 default:
2332 vtn_fail("Invalid SPIR-V atomic");
2333 }
2334 }
2335
2336 if (opcode != SpvOpAtomicStore) {
2337 struct vtn_type *type = vtn_value(b, w[1], vtn_value_type_type)->type;
2338
2339 nir_ssa_dest_init(&atomic->instr, &atomic->dest,
2340 glsl_get_vector_elements(type->type),
2341 glsl_get_bit_size(type->type), NULL);
2342
2343 struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_ssa);
2344 val->ssa = rzalloc(b, struct vtn_ssa_value);
2345 val->ssa->def = &atomic->dest.ssa;
2346 val->ssa->type = type->type;
2347 }
2348
2349 nir_builder_instr_insert(&b->nb, &atomic->instr);
2350 }
2351
2352 static nir_alu_instr *
2353 create_vec(struct vtn_builder *b, unsigned num_components, unsigned bit_size)
2354 {
2355 nir_op op;
2356 switch (num_components) {
2357 case 1: op = nir_op_fmov; break;
2358 case 2: op = nir_op_vec2; break;
2359 case 3: op = nir_op_vec3; break;
2360 case 4: op = nir_op_vec4; break;
2361 default: vtn_fail("bad vector size");
2362 }
2363
2364 nir_alu_instr *vec = nir_alu_instr_create(b->shader, op);
2365 nir_ssa_dest_init(&vec->instr, &vec->dest.dest, num_components,
2366 bit_size, NULL);
2367 vec->dest.write_mask = (1 << num_components) - 1;
2368
2369 return vec;
2370 }
2371
2372 struct vtn_ssa_value *
2373 vtn_ssa_transpose(struct vtn_builder *b, struct vtn_ssa_value *src)
2374 {
2375 if (src->transposed)
2376 return src->transposed;
2377
2378 struct vtn_ssa_value *dest =
2379 vtn_create_ssa_value(b, glsl_transposed_type(src->type));
2380
2381 for (unsigned i = 0; i < glsl_get_matrix_columns(dest->type); i++) {
2382 nir_alu_instr *vec = create_vec(b, glsl_get_matrix_columns(src->type),
2383 glsl_get_bit_size(src->type));
2384 if (glsl_type_is_vector_or_scalar(src->type)) {
2385 vec->src[0].src = nir_src_for_ssa(src->def);
2386 vec->src[0].swizzle[0] = i;
2387 } else {
2388 for (unsigned j = 0; j < glsl_get_matrix_columns(src->type); j++) {
2389 vec->src[j].src = nir_src_for_ssa(src->elems[j]->def);
2390 vec->src[j].swizzle[0] = i;
2391 }
2392 }
2393 nir_builder_instr_insert(&b->nb, &vec->instr);
2394 dest->elems[i]->def = &vec->dest.dest.ssa;
2395 }
2396
2397 dest->transposed = src;
2398
2399 return dest;
2400 }
2401
2402 nir_ssa_def *
2403 vtn_vector_extract(struct vtn_builder *b, nir_ssa_def *src, unsigned index)
2404 {
2405 unsigned swiz[4] = { index };
2406 return nir_swizzle(&b->nb, src, swiz, 1, true);
2407 }
2408
2409 nir_ssa_def *
2410 vtn_vector_insert(struct vtn_builder *b, nir_ssa_def *src, nir_ssa_def *insert,
2411 unsigned index)
2412 {
2413 nir_alu_instr *vec = create_vec(b, src->num_components,
2414 src->bit_size);
2415
2416 for (unsigned i = 0; i < src->num_components; i++) {
2417 if (i == index) {
2418 vec->src[i].src = nir_src_for_ssa(insert);
2419 } else {
2420 vec->src[i].src = nir_src_for_ssa(src);
2421 vec->src[i].swizzle[0] = i;
2422 }
2423 }
2424
2425 nir_builder_instr_insert(&b->nb, &vec->instr);
2426
2427 return &vec->dest.dest.ssa;
2428 }
2429
2430 nir_ssa_def *
2431 vtn_vector_extract_dynamic(struct vtn_builder *b, nir_ssa_def *src,
2432 nir_ssa_def *index)
2433 {
2434 nir_ssa_def *dest = vtn_vector_extract(b, src, 0);
2435 for (unsigned i = 1; i < src->num_components; i++)
2436 dest = nir_bcsel(&b->nb, nir_ieq(&b->nb, index, nir_imm_int(&b->nb, i)),
2437 vtn_vector_extract(b, src, i), dest);
2438
2439 return dest;
2440 }
2441
2442 nir_ssa_def *
2443 vtn_vector_insert_dynamic(struct vtn_builder *b, nir_ssa_def *src,
2444 nir_ssa_def *insert, nir_ssa_def *index)
2445 {
2446 nir_ssa_def *dest = vtn_vector_insert(b, src, insert, 0);
2447 for (unsigned i = 1; i < src->num_components; i++)
2448 dest = nir_bcsel(&b->nb, nir_ieq(&b->nb, index, nir_imm_int(&b->nb, i)),
2449 vtn_vector_insert(b, src, insert, i), dest);
2450
2451 return dest;
2452 }
2453
2454 static nir_ssa_def *
2455 vtn_vector_shuffle(struct vtn_builder *b, unsigned num_components,
2456 nir_ssa_def *src0, nir_ssa_def *src1,
2457 const uint32_t *indices)
2458 {
2459 nir_alu_instr *vec = create_vec(b, num_components, src0->bit_size);
2460
2461 for (unsigned i = 0; i < num_components; i++) {
2462 uint32_t index = indices[i];
2463 if (index == 0xffffffff) {
2464 vec->src[i].src =
2465 nir_src_for_ssa(nir_ssa_undef(&b->nb, 1, src0->bit_size));
2466 } else if (index < src0->num_components) {
2467 vec->src[i].src = nir_src_for_ssa(src0);
2468 vec->src[i].swizzle[0] = index;
2469 } else {
2470 vec->src[i].src = nir_src_for_ssa(src1);
2471 vec->src[i].swizzle[0] = index - src0->num_components;
2472 }
2473 }
2474
2475 nir_builder_instr_insert(&b->nb, &vec->instr);
2476
2477 return &vec->dest.dest.ssa;
2478 }
2479
2480 /*
2481 * Concatentates a number of vectors/scalars together to produce a vector
2482 */
2483 static nir_ssa_def *
2484 vtn_vector_construct(struct vtn_builder *b, unsigned num_components,
2485 unsigned num_srcs, nir_ssa_def **srcs)
2486 {
2487 nir_alu_instr *vec = create_vec(b, num_components, srcs[0]->bit_size);
2488
2489 /* From the SPIR-V 1.1 spec for OpCompositeConstruct:
2490 *
2491 * "When constructing a vector, there must be at least two Constituent
2492 * operands."
2493 */
2494 vtn_assert(num_srcs >= 2);
2495
2496 unsigned dest_idx = 0;
2497 for (unsigned i = 0; i < num_srcs; i++) {
2498 nir_ssa_def *src = srcs[i];
2499 vtn_assert(dest_idx + src->num_components <= num_components);
2500 for (unsigned j = 0; j < src->num_components; j++) {
2501 vec->src[dest_idx].src = nir_src_for_ssa(src);
2502 vec->src[dest_idx].swizzle[0] = j;
2503 dest_idx++;
2504 }
2505 }
2506
2507 /* From the SPIR-V 1.1 spec for OpCompositeConstruct:
2508 *
2509 * "When constructing a vector, the total number of components in all
2510 * the operands must equal the number of components in Result Type."
2511 */
2512 vtn_assert(dest_idx == num_components);
2513
2514 nir_builder_instr_insert(&b->nb, &vec->instr);
2515
2516 return &vec->dest.dest.ssa;
2517 }
2518
2519 static struct vtn_ssa_value *
2520 vtn_composite_copy(void *mem_ctx, struct vtn_ssa_value *src)
2521 {
2522 struct vtn_ssa_value *dest = rzalloc(mem_ctx, struct vtn_ssa_value);
2523 dest->type = src->type;
2524
2525 if (glsl_type_is_vector_or_scalar(src->type)) {
2526 dest->def = src->def;
2527 } else {
2528 unsigned elems = glsl_get_length(src->type);
2529
2530 dest->elems = ralloc_array(mem_ctx, struct vtn_ssa_value *, elems);
2531 for (unsigned i = 0; i < elems; i++)
2532 dest->elems[i] = vtn_composite_copy(mem_ctx, src->elems[i]);
2533 }
2534
2535 return dest;
2536 }
2537
2538 static struct vtn_ssa_value *
2539 vtn_composite_insert(struct vtn_builder *b, struct vtn_ssa_value *src,
2540 struct vtn_ssa_value *insert, const uint32_t *indices,
2541 unsigned num_indices)
2542 {
2543 struct vtn_ssa_value *dest = vtn_composite_copy(b, src);
2544
2545 struct vtn_ssa_value *cur = dest;
2546 unsigned i;
2547 for (i = 0; i < num_indices - 1; i++) {
2548 cur = cur->elems[indices[i]];
2549 }
2550
2551 if (glsl_type_is_vector_or_scalar(cur->type)) {
2552 /* According to the SPIR-V spec, OpCompositeInsert may work down to
2553 * the component granularity. In that case, the last index will be
2554 * the index to insert the scalar into the vector.
2555 */
2556
2557 cur->def = vtn_vector_insert(b, cur->def, insert->def, indices[i]);
2558 } else {
2559 cur->elems[indices[i]] = insert;
2560 }
2561
2562 return dest;
2563 }
2564
2565 static struct vtn_ssa_value *
2566 vtn_composite_extract(struct vtn_builder *b, struct vtn_ssa_value *src,
2567 const uint32_t *indices, unsigned num_indices)
2568 {
2569 struct vtn_ssa_value *cur = src;
2570 for (unsigned i = 0; i < num_indices; i++) {
2571 if (glsl_type_is_vector_or_scalar(cur->type)) {
2572 vtn_assert(i == num_indices - 1);
2573 /* According to the SPIR-V spec, OpCompositeExtract may work down to
2574 * the component granularity. The last index will be the index of the
2575 * vector to extract.
2576 */
2577
2578 struct vtn_ssa_value *ret = rzalloc(b, struct vtn_ssa_value);
2579 ret->type = glsl_scalar_type(glsl_get_base_type(cur->type));
2580 ret->def = vtn_vector_extract(b, cur->def, indices[i]);
2581 return ret;
2582 } else {
2583 cur = cur->elems[indices[i]];
2584 }
2585 }
2586
2587 return cur;
2588 }
2589
2590 static void
2591 vtn_handle_composite(struct vtn_builder *b, SpvOp opcode,
2592 const uint32_t *w, unsigned count)
2593 {
2594 struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_ssa);
2595 const struct glsl_type *type =
2596 vtn_value(b, w[1], vtn_value_type_type)->type->type;
2597 val->ssa = vtn_create_ssa_value(b, type);
2598
2599 switch (opcode) {
2600 case SpvOpVectorExtractDynamic:
2601 val->ssa->def = vtn_vector_extract_dynamic(b, vtn_ssa_value(b, w[3])->def,
2602 vtn_ssa_value(b, w[4])->def);
2603 break;
2604
2605 case SpvOpVectorInsertDynamic:
2606 val->ssa->def = vtn_vector_insert_dynamic(b, vtn_ssa_value(b, w[3])->def,
2607 vtn_ssa_value(b, w[4])->def,
2608 vtn_ssa_value(b, w[5])->def);
2609 break;
2610
2611 case SpvOpVectorShuffle:
2612 val->ssa->def = vtn_vector_shuffle(b, glsl_get_vector_elements(type),
2613 vtn_ssa_value(b, w[3])->def,
2614 vtn_ssa_value(b, w[4])->def,
2615 w + 5);
2616 break;
2617
2618 case SpvOpCompositeConstruct: {
2619 unsigned elems = count - 3;
2620 if (glsl_type_is_vector_or_scalar(type)) {
2621 nir_ssa_def *srcs[4];
2622 for (unsigned i = 0; i < elems; i++)
2623 srcs[i] = vtn_ssa_value(b, w[3 + i])->def;
2624 val->ssa->def =
2625 vtn_vector_construct(b, glsl_get_vector_elements(type),
2626 elems, srcs);
2627 } else {
2628 val->ssa->elems = ralloc_array(b, struct vtn_ssa_value *, elems);
2629 for (unsigned i = 0; i < elems; i++)
2630 val->ssa->elems[i] = vtn_ssa_value(b, w[3 + i]);
2631 }
2632 break;
2633 }
2634 case SpvOpCompositeExtract:
2635 val->ssa = vtn_composite_extract(b, vtn_ssa_value(b, w[3]),
2636 w + 4, count - 4);
2637 break;
2638
2639 case SpvOpCompositeInsert:
2640 val->ssa = vtn_composite_insert(b, vtn_ssa_value(b, w[4]),
2641 vtn_ssa_value(b, w[3]),
2642 w + 5, count - 5);
2643 break;
2644
2645 case SpvOpCopyObject:
2646 val->ssa = vtn_composite_copy(b, vtn_ssa_value(b, w[3]));
2647 break;
2648
2649 default:
2650 vtn_fail("unknown composite operation");
2651 }
2652 }
2653
2654 static void
2655 vtn_handle_barrier(struct vtn_builder *b, SpvOp opcode,
2656 const uint32_t *w, unsigned count)
2657 {
2658 nir_intrinsic_op intrinsic_op;
2659 switch (opcode) {
2660 case SpvOpEmitVertex:
2661 case SpvOpEmitStreamVertex:
2662 intrinsic_op = nir_intrinsic_emit_vertex;
2663 break;
2664 case SpvOpEndPrimitive:
2665 case SpvOpEndStreamPrimitive:
2666 intrinsic_op = nir_intrinsic_end_primitive;
2667 break;
2668 case SpvOpMemoryBarrier:
2669 intrinsic_op = nir_intrinsic_memory_barrier;
2670 break;
2671 case SpvOpControlBarrier:
2672 intrinsic_op = nir_intrinsic_barrier;
2673 break;
2674 default:
2675 vtn_fail("unknown barrier instruction");
2676 }
2677
2678 nir_intrinsic_instr *intrin =
2679 nir_intrinsic_instr_create(b->shader, intrinsic_op);
2680
2681 if (opcode == SpvOpEmitStreamVertex || opcode == SpvOpEndStreamPrimitive)
2682 nir_intrinsic_set_stream_id(intrin, w[1]);
2683
2684 nir_builder_instr_insert(&b->nb, &intrin->instr);
2685 }
2686
2687 static unsigned
2688 gl_primitive_from_spv_execution_mode(struct vtn_builder *b,
2689 SpvExecutionMode mode)
2690 {
2691 switch (mode) {
2692 case SpvExecutionModeInputPoints:
2693 case SpvExecutionModeOutputPoints:
2694 return 0; /* GL_POINTS */
2695 case SpvExecutionModeInputLines:
2696 return 1; /* GL_LINES */
2697 case SpvExecutionModeInputLinesAdjacency:
2698 return 0x000A; /* GL_LINE_STRIP_ADJACENCY_ARB */
2699 case SpvExecutionModeTriangles:
2700 return 4; /* GL_TRIANGLES */
2701 case SpvExecutionModeInputTrianglesAdjacency:
2702 return 0x000C; /* GL_TRIANGLES_ADJACENCY_ARB */
2703 case SpvExecutionModeQuads:
2704 return 7; /* GL_QUADS */
2705 case SpvExecutionModeIsolines:
2706 return 0x8E7A; /* GL_ISOLINES */
2707 case SpvExecutionModeOutputLineStrip:
2708 return 3; /* GL_LINE_STRIP */
2709 case SpvExecutionModeOutputTriangleStrip:
2710 return 5; /* GL_TRIANGLE_STRIP */
2711 default:
2712 vtn_fail("Invalid primitive type");
2713 }
2714 }
2715
2716 static unsigned
2717 vertices_in_from_spv_execution_mode(struct vtn_builder *b,
2718 SpvExecutionMode mode)
2719 {
2720 switch (mode) {
2721 case SpvExecutionModeInputPoints:
2722 return 1;
2723 case SpvExecutionModeInputLines:
2724 return 2;
2725 case SpvExecutionModeInputLinesAdjacency:
2726 return 4;
2727 case SpvExecutionModeTriangles:
2728 return 3;
2729 case SpvExecutionModeInputTrianglesAdjacency:
2730 return 6;
2731 default:
2732 vtn_fail("Invalid GS input mode");
2733 }
2734 }
2735
2736 static gl_shader_stage
2737 stage_for_execution_model(struct vtn_builder *b, SpvExecutionModel model)
2738 {
2739 switch (model) {
2740 case SpvExecutionModelVertex:
2741 return MESA_SHADER_VERTEX;
2742 case SpvExecutionModelTessellationControl:
2743 return MESA_SHADER_TESS_CTRL;
2744 case SpvExecutionModelTessellationEvaluation:
2745 return MESA_SHADER_TESS_EVAL;
2746 case SpvExecutionModelGeometry:
2747 return MESA_SHADER_GEOMETRY;
2748 case SpvExecutionModelFragment:
2749 return MESA_SHADER_FRAGMENT;
2750 case SpvExecutionModelGLCompute:
2751 return MESA_SHADER_COMPUTE;
2752 default:
2753 vtn_fail("Unsupported execution model");
2754 }
2755 }
2756
2757 #define spv_check_supported(name, cap) do { \
2758 if (!(b->options && b->options->caps.name)) \
2759 vtn_warn("Unsupported SPIR-V capability: %s", \
2760 spirv_capability_to_string(cap)); \
2761 } while(0)
2762
2763 static bool
2764 vtn_handle_preamble_instruction(struct vtn_builder *b, SpvOp opcode,
2765 const uint32_t *w, unsigned count)
2766 {
2767 switch (opcode) {
2768 case SpvOpSource: {
2769 const char *lang;
2770 switch (w[1]) {
2771 default:
2772 case SpvSourceLanguageUnknown: lang = "unknown"; break;
2773 case SpvSourceLanguageESSL: lang = "ESSL"; break;
2774 case SpvSourceLanguageGLSL: lang = "GLSL"; break;
2775 case SpvSourceLanguageOpenCL_C: lang = "OpenCL C"; break;
2776 case SpvSourceLanguageOpenCL_CPP: lang = "OpenCL C++"; break;
2777 case SpvSourceLanguageHLSL: lang = "HLSL"; break;
2778 }
2779
2780 uint32_t version = w[2];
2781
2782 const char *file =
2783 (count > 3) ? vtn_value(b, w[3], vtn_value_type_string)->str : "";
2784
2785 vtn_info("Parsing SPIR-V from %s %u source file %s", lang, version, file);
2786 break;
2787 }
2788
2789 case SpvOpSourceExtension:
2790 case SpvOpSourceContinued:
2791 case SpvOpExtension:
2792 /* Unhandled, but these are for debug so that's ok. */
2793 break;
2794
2795 case SpvOpCapability: {
2796 SpvCapability cap = w[1];
2797 switch (cap) {
2798 case SpvCapabilityMatrix:
2799 case SpvCapabilityShader:
2800 case SpvCapabilityGeometry:
2801 case SpvCapabilityGeometryPointSize:
2802 case SpvCapabilityUniformBufferArrayDynamicIndexing:
2803 case SpvCapabilitySampledImageArrayDynamicIndexing:
2804 case SpvCapabilityStorageBufferArrayDynamicIndexing:
2805 case SpvCapabilityStorageImageArrayDynamicIndexing:
2806 case SpvCapabilityImageRect:
2807 case SpvCapabilitySampledRect:
2808 case SpvCapabilitySampled1D:
2809 case SpvCapabilityImage1D:
2810 case SpvCapabilitySampledCubeArray:
2811 case SpvCapabilityImageCubeArray:
2812 case SpvCapabilitySampledBuffer:
2813 case SpvCapabilityImageBuffer:
2814 case SpvCapabilityImageQuery:
2815 case SpvCapabilityDerivativeControl:
2816 case SpvCapabilityInterpolationFunction:
2817 case SpvCapabilityMultiViewport:
2818 case SpvCapabilitySampleRateShading:
2819 case SpvCapabilityClipDistance:
2820 case SpvCapabilityCullDistance:
2821 case SpvCapabilityInputAttachment:
2822 case SpvCapabilityImageGatherExtended:
2823 case SpvCapabilityStorageImageExtendedFormats:
2824 break;
2825
2826 case SpvCapabilityGeometryStreams:
2827 case SpvCapabilityLinkage:
2828 case SpvCapabilityVector16:
2829 case SpvCapabilityFloat16Buffer:
2830 case SpvCapabilityFloat16:
2831 case SpvCapabilityInt64Atomics:
2832 case SpvCapabilityAtomicStorage:
2833 case SpvCapabilityInt16:
2834 case SpvCapabilityStorageImageMultisample:
2835 case SpvCapabilityInt8:
2836 case SpvCapabilitySparseResidency:
2837 case SpvCapabilityMinLod:
2838 case SpvCapabilityTransformFeedback:
2839 vtn_warn("Unsupported SPIR-V capability: %s",
2840 spirv_capability_to_string(cap));
2841 break;
2842
2843 case SpvCapabilityFloat64:
2844 spv_check_supported(float64, cap);
2845 break;
2846 case SpvCapabilityInt64:
2847 spv_check_supported(int64, cap);
2848 break;
2849
2850 case SpvCapabilityAddresses:
2851 case SpvCapabilityKernel:
2852 case SpvCapabilityImageBasic:
2853 case SpvCapabilityImageReadWrite:
2854 case SpvCapabilityImageMipmap:
2855 case SpvCapabilityPipes:
2856 case SpvCapabilityGroups:
2857 case SpvCapabilityDeviceEnqueue:
2858 case SpvCapabilityLiteralSampler:
2859 case SpvCapabilityGenericPointer:
2860 vtn_warn("Unsupported OpenCL-style SPIR-V capability: %s",
2861 spirv_capability_to_string(cap));
2862 break;
2863
2864 case SpvCapabilityImageMSArray:
2865 spv_check_supported(image_ms_array, cap);
2866 break;
2867
2868 case SpvCapabilityTessellation:
2869 case SpvCapabilityTessellationPointSize:
2870 spv_check_supported(tessellation, cap);
2871 break;
2872
2873 case SpvCapabilityDrawParameters:
2874 spv_check_supported(draw_parameters, cap);
2875 break;
2876
2877 case SpvCapabilityStorageImageReadWithoutFormat:
2878 spv_check_supported(image_read_without_format, cap);
2879 break;
2880
2881 case SpvCapabilityStorageImageWriteWithoutFormat:
2882 spv_check_supported(image_write_without_format, cap);
2883 break;
2884
2885 case SpvCapabilityMultiView:
2886 spv_check_supported(multiview, cap);
2887 break;
2888
2889 case SpvCapabilityVariablePointersStorageBuffer:
2890 case SpvCapabilityVariablePointers:
2891 spv_check_supported(variable_pointers, cap);
2892 break;
2893
2894 default:
2895 vtn_fail("Unhandled capability");
2896 }
2897 break;
2898 }
2899
2900 case SpvOpExtInstImport:
2901 vtn_handle_extension(b, opcode, w, count);
2902 break;
2903
2904 case SpvOpMemoryModel:
2905 vtn_assert(w[1] == SpvAddressingModelLogical);
2906 vtn_assert(w[2] == SpvMemoryModelSimple ||
2907 w[2] == SpvMemoryModelGLSL450);
2908 break;
2909
2910 case SpvOpEntryPoint: {
2911 struct vtn_value *entry_point = &b->values[w[2]];
2912 /* Let this be a name label regardless */
2913 unsigned name_words;
2914 entry_point->name = vtn_string_literal(b, &w[3], count - 3, &name_words);
2915
2916 if (strcmp(entry_point->name, b->entry_point_name) != 0 ||
2917 stage_for_execution_model(b, w[1]) != b->entry_point_stage)
2918 break;
2919
2920 vtn_assert(b->entry_point == NULL);
2921 b->entry_point = entry_point;
2922 break;
2923 }
2924
2925 case SpvOpString:
2926 vtn_push_value(b, w[1], vtn_value_type_string)->str =
2927 vtn_string_literal(b, &w[2], count - 2, NULL);
2928 break;
2929
2930 case SpvOpName:
2931 b->values[w[1]].name = vtn_string_literal(b, &w[2], count - 2, NULL);
2932 break;
2933
2934 case SpvOpMemberName:
2935 /* TODO */
2936 break;
2937
2938 case SpvOpExecutionMode:
2939 case SpvOpDecorationGroup:
2940 case SpvOpDecorate:
2941 case SpvOpMemberDecorate:
2942 case SpvOpGroupDecorate:
2943 case SpvOpGroupMemberDecorate:
2944 vtn_handle_decoration(b, opcode, w, count);
2945 break;
2946
2947 default:
2948 return false; /* End of preamble */
2949 }
2950
2951 return true;
2952 }
2953
2954 static void
2955 vtn_handle_execution_mode(struct vtn_builder *b, struct vtn_value *entry_point,
2956 const struct vtn_decoration *mode, void *data)
2957 {
2958 vtn_assert(b->entry_point == entry_point);
2959
2960 switch(mode->exec_mode) {
2961 case SpvExecutionModeOriginUpperLeft:
2962 case SpvExecutionModeOriginLowerLeft:
2963 b->origin_upper_left =
2964 (mode->exec_mode == SpvExecutionModeOriginUpperLeft);
2965 break;
2966
2967 case SpvExecutionModeEarlyFragmentTests:
2968 vtn_assert(b->shader->info.stage == MESA_SHADER_FRAGMENT);
2969 b->shader->info.fs.early_fragment_tests = true;
2970 break;
2971
2972 case SpvExecutionModeInvocations:
2973 vtn_assert(b->shader->info.stage == MESA_SHADER_GEOMETRY);
2974 b->shader->info.gs.invocations = MAX2(1, mode->literals[0]);
2975 break;
2976
2977 case SpvExecutionModeDepthReplacing:
2978 vtn_assert(b->shader->info.stage == MESA_SHADER_FRAGMENT);
2979 b->shader->info.fs.depth_layout = FRAG_DEPTH_LAYOUT_ANY;
2980 break;
2981 case SpvExecutionModeDepthGreater:
2982 vtn_assert(b->shader->info.stage == MESA_SHADER_FRAGMENT);
2983 b->shader->info.fs.depth_layout = FRAG_DEPTH_LAYOUT_GREATER;
2984 break;
2985 case SpvExecutionModeDepthLess:
2986 vtn_assert(b->shader->info.stage == MESA_SHADER_FRAGMENT);
2987 b->shader->info.fs.depth_layout = FRAG_DEPTH_LAYOUT_LESS;
2988 break;
2989 case SpvExecutionModeDepthUnchanged:
2990 vtn_assert(b->shader->info.stage == MESA_SHADER_FRAGMENT);
2991 b->shader->info.fs.depth_layout = FRAG_DEPTH_LAYOUT_UNCHANGED;
2992 break;
2993
2994 case SpvExecutionModeLocalSize:
2995 vtn_assert(b->shader->info.stage == MESA_SHADER_COMPUTE);
2996 b->shader->info.cs.local_size[0] = mode->literals[0];
2997 b->shader->info.cs.local_size[1] = mode->literals[1];
2998 b->shader->info.cs.local_size[2] = mode->literals[2];
2999 break;
3000 case SpvExecutionModeLocalSizeHint:
3001 break; /* Nothing to do with this */
3002
3003 case SpvExecutionModeOutputVertices:
3004 if (b->shader->info.stage == MESA_SHADER_TESS_CTRL ||
3005 b->shader->info.stage == MESA_SHADER_TESS_EVAL) {
3006 b->shader->info.tess.tcs_vertices_out = mode->literals[0];
3007 } else {
3008 vtn_assert(b->shader->info.stage == MESA_SHADER_GEOMETRY);
3009 b->shader->info.gs.vertices_out = mode->literals[0];
3010 }
3011 break;
3012
3013 case SpvExecutionModeInputPoints:
3014 case SpvExecutionModeInputLines:
3015 case SpvExecutionModeInputLinesAdjacency:
3016 case SpvExecutionModeTriangles:
3017 case SpvExecutionModeInputTrianglesAdjacency:
3018 case SpvExecutionModeQuads:
3019 case SpvExecutionModeIsolines:
3020 if (b->shader->info.stage == MESA_SHADER_TESS_CTRL ||
3021 b->shader->info.stage == MESA_SHADER_TESS_EVAL) {
3022 b->shader->info.tess.primitive_mode =
3023 gl_primitive_from_spv_execution_mode(b, mode->exec_mode);
3024 } else {
3025 vtn_assert(b->shader->info.stage == MESA_SHADER_GEOMETRY);
3026 b->shader->info.gs.vertices_in =
3027 vertices_in_from_spv_execution_mode(b, mode->exec_mode);
3028 }
3029 break;
3030
3031 case SpvExecutionModeOutputPoints:
3032 case SpvExecutionModeOutputLineStrip:
3033 case SpvExecutionModeOutputTriangleStrip:
3034 vtn_assert(b->shader->info.stage == MESA_SHADER_GEOMETRY);
3035 b->shader->info.gs.output_primitive =
3036 gl_primitive_from_spv_execution_mode(b, mode->exec_mode);
3037 break;
3038
3039 case SpvExecutionModeSpacingEqual:
3040 vtn_assert(b->shader->info.stage == MESA_SHADER_TESS_CTRL ||
3041 b->shader->info.stage == MESA_SHADER_TESS_EVAL);
3042 b->shader->info.tess.spacing = TESS_SPACING_EQUAL;
3043 break;
3044 case SpvExecutionModeSpacingFractionalEven:
3045 vtn_assert(b->shader->info.stage == MESA_SHADER_TESS_CTRL ||
3046 b->shader->info.stage == MESA_SHADER_TESS_EVAL);
3047 b->shader->info.tess.spacing = TESS_SPACING_FRACTIONAL_EVEN;
3048 break;
3049 case SpvExecutionModeSpacingFractionalOdd:
3050 vtn_assert(b->shader->info.stage == MESA_SHADER_TESS_CTRL ||
3051 b->shader->info.stage == MESA_SHADER_TESS_EVAL);
3052 b->shader->info.tess.spacing = TESS_SPACING_FRACTIONAL_ODD;
3053 break;
3054 case SpvExecutionModeVertexOrderCw:
3055 vtn_assert(b->shader->info.stage == MESA_SHADER_TESS_CTRL ||
3056 b->shader->info.stage == MESA_SHADER_TESS_EVAL);
3057 b->shader->info.tess.ccw = false;
3058 break;
3059 case SpvExecutionModeVertexOrderCcw:
3060 vtn_assert(b->shader->info.stage == MESA_SHADER_TESS_CTRL ||
3061 b->shader->info.stage == MESA_SHADER_TESS_EVAL);
3062 b->shader->info.tess.ccw = true;
3063 break;
3064 case SpvExecutionModePointMode:
3065 vtn_assert(b->shader->info.stage == MESA_SHADER_TESS_CTRL ||
3066 b->shader->info.stage == MESA_SHADER_TESS_EVAL);
3067 b->shader->info.tess.point_mode = true;
3068 break;
3069
3070 case SpvExecutionModePixelCenterInteger:
3071 b->pixel_center_integer = true;
3072 break;
3073
3074 case SpvExecutionModeXfb:
3075 vtn_fail("Unhandled execution mode");
3076 break;
3077
3078 case SpvExecutionModeVecTypeHint:
3079 case SpvExecutionModeContractionOff:
3080 break; /* OpenCL */
3081
3082 default:
3083 vtn_fail("Unhandled execution mode");
3084 }
3085 }
3086
3087 static bool
3088 vtn_handle_variable_or_type_instruction(struct vtn_builder *b, SpvOp opcode,
3089 const uint32_t *w, unsigned count)
3090 {
3091 switch (opcode) {
3092 case SpvOpSource:
3093 case SpvOpSourceContinued:
3094 case SpvOpSourceExtension:
3095 case SpvOpExtension:
3096 case SpvOpCapability:
3097 case SpvOpExtInstImport:
3098 case SpvOpMemoryModel:
3099 case SpvOpEntryPoint:
3100 case SpvOpExecutionMode:
3101 case SpvOpString:
3102 case SpvOpName:
3103 case SpvOpMemberName:
3104 case SpvOpDecorationGroup:
3105 case SpvOpDecorate:
3106 case SpvOpMemberDecorate:
3107 case SpvOpGroupDecorate:
3108 case SpvOpGroupMemberDecorate:
3109 vtn_fail("Invalid opcode types and variables section");
3110 break;
3111
3112 case SpvOpTypeVoid:
3113 case SpvOpTypeBool:
3114 case SpvOpTypeInt:
3115 case SpvOpTypeFloat:
3116 case SpvOpTypeVector:
3117 case SpvOpTypeMatrix:
3118 case SpvOpTypeImage:
3119 case SpvOpTypeSampler:
3120 case SpvOpTypeSampledImage:
3121 case SpvOpTypeArray:
3122 case SpvOpTypeRuntimeArray:
3123 case SpvOpTypeStruct:
3124 case SpvOpTypeOpaque:
3125 case SpvOpTypePointer:
3126 case SpvOpTypeFunction:
3127 case SpvOpTypeEvent:
3128 case SpvOpTypeDeviceEvent:
3129 case SpvOpTypeReserveId:
3130 case SpvOpTypeQueue:
3131 case SpvOpTypePipe:
3132 vtn_handle_type(b, opcode, w, count);
3133 break;
3134
3135 case SpvOpConstantTrue:
3136 case SpvOpConstantFalse:
3137 case SpvOpConstant:
3138 case SpvOpConstantComposite:
3139 case SpvOpConstantSampler:
3140 case SpvOpConstantNull:
3141 case SpvOpSpecConstantTrue:
3142 case SpvOpSpecConstantFalse:
3143 case SpvOpSpecConstant:
3144 case SpvOpSpecConstantComposite:
3145 case SpvOpSpecConstantOp:
3146 vtn_handle_constant(b, opcode, w, count);
3147 break;
3148
3149 case SpvOpUndef:
3150 case SpvOpVariable:
3151 vtn_handle_variables(b, opcode, w, count);
3152 break;
3153
3154 default:
3155 return false; /* End of preamble */
3156 }
3157
3158 return true;
3159 }
3160
3161 static bool
3162 vtn_handle_body_instruction(struct vtn_builder *b, SpvOp opcode,
3163 const uint32_t *w, unsigned count)
3164 {
3165 switch (opcode) {
3166 case SpvOpLabel:
3167 break;
3168
3169 case SpvOpLoopMerge:
3170 case SpvOpSelectionMerge:
3171 /* This is handled by cfg pre-pass and walk_blocks */
3172 break;
3173
3174 case SpvOpUndef: {
3175 struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_undef);
3176 val->type = vtn_value(b, w[1], vtn_value_type_type)->type;
3177 break;
3178 }
3179
3180 case SpvOpExtInst:
3181 vtn_handle_extension(b, opcode, w, count);
3182 break;
3183
3184 case SpvOpVariable:
3185 case SpvOpLoad:
3186 case SpvOpStore:
3187 case SpvOpCopyMemory:
3188 case SpvOpCopyMemorySized:
3189 case SpvOpAccessChain:
3190 case SpvOpPtrAccessChain:
3191 case SpvOpInBoundsAccessChain:
3192 case SpvOpArrayLength:
3193 vtn_handle_variables(b, opcode, w, count);
3194 break;
3195
3196 case SpvOpFunctionCall:
3197 vtn_handle_function_call(b, opcode, w, count);
3198 break;
3199
3200 case SpvOpSampledImage:
3201 case SpvOpImage:
3202 case SpvOpImageSampleImplicitLod:
3203 case SpvOpImageSampleExplicitLod:
3204 case SpvOpImageSampleDrefImplicitLod:
3205 case SpvOpImageSampleDrefExplicitLod:
3206 case SpvOpImageSampleProjImplicitLod:
3207 case SpvOpImageSampleProjExplicitLod:
3208 case SpvOpImageSampleProjDrefImplicitLod:
3209 case SpvOpImageSampleProjDrefExplicitLod:
3210 case SpvOpImageFetch:
3211 case SpvOpImageGather:
3212 case SpvOpImageDrefGather:
3213 case SpvOpImageQuerySizeLod:
3214 case SpvOpImageQueryLod:
3215 case SpvOpImageQueryLevels:
3216 case SpvOpImageQuerySamples:
3217 vtn_handle_texture(b, opcode, w, count);
3218 break;
3219
3220 case SpvOpImageRead:
3221 case SpvOpImageWrite:
3222 case SpvOpImageTexelPointer:
3223 vtn_handle_image(b, opcode, w, count);
3224 break;
3225
3226 case SpvOpImageQuerySize: {
3227 struct vtn_pointer *image =
3228 vtn_value(b, w[3], vtn_value_type_pointer)->pointer;
3229 if (image->mode == vtn_variable_mode_image) {
3230 vtn_handle_image(b, opcode, w, count);
3231 } else {
3232 vtn_assert(image->mode == vtn_variable_mode_sampler);
3233 vtn_handle_texture(b, opcode, w, count);
3234 }
3235 break;
3236 }
3237
3238 case SpvOpAtomicLoad:
3239 case SpvOpAtomicExchange:
3240 case SpvOpAtomicCompareExchange:
3241 case SpvOpAtomicCompareExchangeWeak:
3242 case SpvOpAtomicIIncrement:
3243 case SpvOpAtomicIDecrement:
3244 case SpvOpAtomicIAdd:
3245 case SpvOpAtomicISub:
3246 case SpvOpAtomicSMin:
3247 case SpvOpAtomicUMin:
3248 case SpvOpAtomicSMax:
3249 case SpvOpAtomicUMax:
3250 case SpvOpAtomicAnd:
3251 case SpvOpAtomicOr:
3252 case SpvOpAtomicXor: {
3253 struct vtn_value *pointer = vtn_untyped_value(b, w[3]);
3254 if (pointer->value_type == vtn_value_type_image_pointer) {
3255 vtn_handle_image(b, opcode, w, count);
3256 } else {
3257 vtn_assert(pointer->value_type == vtn_value_type_pointer);
3258 vtn_handle_ssbo_or_shared_atomic(b, opcode, w, count);
3259 }
3260 break;
3261 }
3262
3263 case SpvOpAtomicStore: {
3264 struct vtn_value *pointer = vtn_untyped_value(b, w[1]);
3265 if (pointer->value_type == vtn_value_type_image_pointer) {
3266 vtn_handle_image(b, opcode, w, count);
3267 } else {
3268 vtn_assert(pointer->value_type == vtn_value_type_pointer);
3269 vtn_handle_ssbo_or_shared_atomic(b, opcode, w, count);
3270 }
3271 break;
3272 }
3273
3274 case SpvOpSelect: {
3275 /* Handle OpSelect up-front here because it needs to be able to handle
3276 * pointers and not just regular vectors and scalars.
3277 */
3278 struct vtn_type *res_type = vtn_value(b, w[1], vtn_value_type_type)->type;
3279 struct vtn_ssa_value *ssa = vtn_create_ssa_value(b, res_type->type);
3280 ssa->def = nir_bcsel(&b->nb, vtn_ssa_value(b, w[3])->def,
3281 vtn_ssa_value(b, w[4])->def,
3282 vtn_ssa_value(b, w[5])->def);
3283 vtn_push_ssa(b, w[2], res_type, ssa);
3284 break;
3285 }
3286
3287 case SpvOpSNegate:
3288 case SpvOpFNegate:
3289 case SpvOpNot:
3290 case SpvOpAny:
3291 case SpvOpAll:
3292 case SpvOpConvertFToU:
3293 case SpvOpConvertFToS:
3294 case SpvOpConvertSToF:
3295 case SpvOpConvertUToF:
3296 case SpvOpUConvert:
3297 case SpvOpSConvert:
3298 case SpvOpFConvert:
3299 case SpvOpQuantizeToF16:
3300 case SpvOpConvertPtrToU:
3301 case SpvOpConvertUToPtr:
3302 case SpvOpPtrCastToGeneric:
3303 case SpvOpGenericCastToPtr:
3304 case SpvOpBitcast:
3305 case SpvOpIsNan:
3306 case SpvOpIsInf:
3307 case SpvOpIsFinite:
3308 case SpvOpIsNormal:
3309 case SpvOpSignBitSet:
3310 case SpvOpLessOrGreater:
3311 case SpvOpOrdered:
3312 case SpvOpUnordered:
3313 case SpvOpIAdd:
3314 case SpvOpFAdd:
3315 case SpvOpISub:
3316 case SpvOpFSub:
3317 case SpvOpIMul:
3318 case SpvOpFMul:
3319 case SpvOpUDiv:
3320 case SpvOpSDiv:
3321 case SpvOpFDiv:
3322 case SpvOpUMod:
3323 case SpvOpSRem:
3324 case SpvOpSMod:
3325 case SpvOpFRem:
3326 case SpvOpFMod:
3327 case SpvOpVectorTimesScalar:
3328 case SpvOpDot:
3329 case SpvOpIAddCarry:
3330 case SpvOpISubBorrow:
3331 case SpvOpUMulExtended:
3332 case SpvOpSMulExtended:
3333 case SpvOpShiftRightLogical:
3334 case SpvOpShiftRightArithmetic:
3335 case SpvOpShiftLeftLogical:
3336 case SpvOpLogicalEqual:
3337 case SpvOpLogicalNotEqual:
3338 case SpvOpLogicalOr:
3339 case SpvOpLogicalAnd:
3340 case SpvOpLogicalNot:
3341 case SpvOpBitwiseOr:
3342 case SpvOpBitwiseXor:
3343 case SpvOpBitwiseAnd:
3344 case SpvOpIEqual:
3345 case SpvOpFOrdEqual:
3346 case SpvOpFUnordEqual:
3347 case SpvOpINotEqual:
3348 case SpvOpFOrdNotEqual:
3349 case SpvOpFUnordNotEqual:
3350 case SpvOpULessThan:
3351 case SpvOpSLessThan:
3352 case SpvOpFOrdLessThan:
3353 case SpvOpFUnordLessThan:
3354 case SpvOpUGreaterThan:
3355 case SpvOpSGreaterThan:
3356 case SpvOpFOrdGreaterThan:
3357 case SpvOpFUnordGreaterThan:
3358 case SpvOpULessThanEqual:
3359 case SpvOpSLessThanEqual:
3360 case SpvOpFOrdLessThanEqual:
3361 case SpvOpFUnordLessThanEqual:
3362 case SpvOpUGreaterThanEqual:
3363 case SpvOpSGreaterThanEqual:
3364 case SpvOpFOrdGreaterThanEqual:
3365 case SpvOpFUnordGreaterThanEqual:
3366 case SpvOpDPdx:
3367 case SpvOpDPdy:
3368 case SpvOpFwidth:
3369 case SpvOpDPdxFine:
3370 case SpvOpDPdyFine:
3371 case SpvOpFwidthFine:
3372 case SpvOpDPdxCoarse:
3373 case SpvOpDPdyCoarse:
3374 case SpvOpFwidthCoarse:
3375 case SpvOpBitFieldInsert:
3376 case SpvOpBitFieldSExtract:
3377 case SpvOpBitFieldUExtract:
3378 case SpvOpBitReverse:
3379 case SpvOpBitCount:
3380 case SpvOpTranspose:
3381 case SpvOpOuterProduct:
3382 case SpvOpMatrixTimesScalar:
3383 case SpvOpVectorTimesMatrix:
3384 case SpvOpMatrixTimesVector:
3385 case SpvOpMatrixTimesMatrix:
3386 vtn_handle_alu(b, opcode, w, count);
3387 break;
3388
3389 case SpvOpVectorExtractDynamic:
3390 case SpvOpVectorInsertDynamic:
3391 case SpvOpVectorShuffle:
3392 case SpvOpCompositeConstruct:
3393 case SpvOpCompositeExtract:
3394 case SpvOpCompositeInsert:
3395 case SpvOpCopyObject:
3396 vtn_handle_composite(b, opcode, w, count);
3397 break;
3398
3399 case SpvOpEmitVertex:
3400 case SpvOpEndPrimitive:
3401 case SpvOpEmitStreamVertex:
3402 case SpvOpEndStreamPrimitive:
3403 case SpvOpControlBarrier:
3404 case SpvOpMemoryBarrier:
3405 vtn_handle_barrier(b, opcode, w, count);
3406 break;
3407
3408 default:
3409 vtn_fail("Unhandled opcode");
3410 }
3411
3412 return true;
3413 }
3414
3415 nir_function *
3416 spirv_to_nir(const uint32_t *words, size_t word_count,
3417 struct nir_spirv_specialization *spec, unsigned num_spec,
3418 gl_shader_stage stage, const char *entry_point_name,
3419 const struct spirv_to_nir_options *options,
3420 const nir_shader_compiler_options *nir_options)
3421 {
3422 /* Initialize the stn_builder object */
3423 struct vtn_builder *b = rzalloc(NULL, struct vtn_builder);
3424 b->spirv = words;
3425 b->file = NULL;
3426 b->line = -1;
3427 b->col = -1;
3428 exec_list_make_empty(&b->functions);
3429 b->entry_point_stage = stage;
3430 b->entry_point_name = entry_point_name;
3431 b->options = options;
3432
3433 /* See also _vtn_fail() */
3434 if (setjmp(b->fail_jump)) {
3435 ralloc_free(b);
3436 return NULL;
3437 }
3438
3439 const uint32_t *word_end = words + word_count;
3440
3441 /* Handle the SPIR-V header (first 4 dwords) */
3442 vtn_assert(word_count > 5);
3443
3444 vtn_assert(words[0] == SpvMagicNumber);
3445 vtn_assert(words[1] >= 0x10000);
3446 /* words[2] == generator magic */
3447 unsigned value_id_bound = words[3];
3448 vtn_assert(words[4] == 0);
3449
3450 words+= 5;
3451
3452 b->value_id_bound = value_id_bound;
3453 b->values = rzalloc_array(b, struct vtn_value, value_id_bound);
3454
3455 /* Handle all the preamble instructions */
3456 words = vtn_foreach_instruction(b, words, word_end,
3457 vtn_handle_preamble_instruction);
3458
3459 if (b->entry_point == NULL) {
3460 vtn_fail("Entry point not found");
3461 ralloc_free(b);
3462 return NULL;
3463 }
3464
3465 b->shader = nir_shader_create(b, stage, nir_options, NULL);
3466
3467 /* Set shader info defaults */
3468 b->shader->info.gs.invocations = 1;
3469
3470 /* Parse execution modes */
3471 vtn_foreach_execution_mode(b, b->entry_point,
3472 vtn_handle_execution_mode, NULL);
3473
3474 b->specializations = spec;
3475 b->num_specializations = num_spec;
3476
3477 /* Handle all variable, type, and constant instructions */
3478 words = vtn_foreach_instruction(b, words, word_end,
3479 vtn_handle_variable_or_type_instruction);
3480
3481 vtn_build_cfg(b, words, word_end);
3482
3483 assert(b->entry_point->value_type == vtn_value_type_function);
3484 b->entry_point->func->referenced = true;
3485
3486 bool progress;
3487 do {
3488 progress = false;
3489 foreach_list_typed(struct vtn_function, func, node, &b->functions) {
3490 if (func->referenced && !func->emitted) {
3491 b->const_table = _mesa_hash_table_create(b, _mesa_hash_pointer,
3492 _mesa_key_pointer_equal);
3493
3494 vtn_function_emit(b, func, vtn_handle_body_instruction);
3495 progress = true;
3496 }
3497 }
3498 } while (progress);
3499
3500 vtn_assert(b->entry_point->value_type == vtn_value_type_function);
3501 nir_function *entry_point = b->entry_point->func->impl->function;
3502 vtn_assert(entry_point);
3503
3504 /* Unparent the shader from the vtn_builder before we delete the builder */
3505 ralloc_steal(NULL, b->shader);
3506
3507 ralloc_free(b);
3508
3509 return entry_point;
3510 }