nir: Make nir_copy_deref follow the "clone" pattern
[mesa.git] / src / compiler / spirv / vtn_variables.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 "spirv_info.h"
30
31 static struct vtn_access_chain *
32 vtn_access_chain_extend(struct vtn_builder *b, struct vtn_access_chain *old,
33 unsigned new_ids)
34 {
35 struct vtn_access_chain *chain;
36
37 unsigned new_len = old->length + new_ids;
38 /* TODO: don't use rzalloc */
39 chain = rzalloc_size(b, sizeof(*chain) + new_len * sizeof(chain->link[0]));
40
41 chain->var = old->var;
42 chain->length = new_len;
43
44 for (unsigned i = 0; i < old->length; i++)
45 chain->link[i] = old->link[i];
46
47 return chain;
48 }
49
50 static nir_ssa_def *
51 vtn_access_link_as_ssa(struct vtn_builder *b, struct vtn_access_link link,
52 unsigned stride)
53 {
54 assert(stride > 0);
55 if (link.mode == vtn_access_mode_literal) {
56 return nir_imm_int(&b->nb, link.id * stride);
57 } else if (stride == 1) {
58 return vtn_ssa_value(b, link.id)->def;
59 } else {
60 return nir_imul(&b->nb, vtn_ssa_value(b, link.id)->def,
61 nir_imm_int(&b->nb, stride));
62 }
63 }
64
65 static struct vtn_type *
66 vtn_access_chain_tail_type(struct vtn_builder *b,
67 struct vtn_access_chain *chain)
68 {
69 struct vtn_type *type = chain->var->type;
70 for (unsigned i = 0; i < chain->length; i++) {
71 if (glsl_type_is_struct(type->type)) {
72 assert(chain->link[i].mode == vtn_access_mode_literal);
73 type = type->members[chain->link[i].id];
74 } else {
75 type = type->array_element;
76 }
77 }
78 return type;
79 }
80
81 /* Crawls a chain of array derefs and rewrites the types so that the
82 * lengths stay the same but the terminal type is the one given by
83 * tail_type. This is useful for split structures.
84 */
85 static void
86 rewrite_deref_types(nir_deref *deref, const struct glsl_type *type)
87 {
88 deref->type = type;
89 if (deref->child) {
90 assert(deref->child->deref_type == nir_deref_type_array);
91 assert(glsl_type_is_array(deref->type));
92 rewrite_deref_types(deref->child, glsl_get_array_element(type));
93 }
94 }
95
96 nir_deref_var *
97 vtn_access_chain_to_deref(struct vtn_builder *b, struct vtn_access_chain *chain)
98 {
99 nir_deref_var *deref_var;
100 if (chain->var->var) {
101 deref_var = nir_deref_var_create(b, chain->var->var);
102 } else {
103 assert(chain->var->members);
104 /* Create the deref_var manually. It will get filled out later. */
105 deref_var = rzalloc(b, nir_deref_var);
106 deref_var->deref.deref_type = nir_deref_type_var;
107 }
108
109 struct vtn_type *deref_type = chain->var->type;
110 nir_deref *tail = &deref_var->deref;
111 nir_variable **members = chain->var->members;
112
113 for (unsigned i = 0; i < chain->length; i++) {
114 enum glsl_base_type base_type = glsl_get_base_type(deref_type->type);
115 switch (base_type) {
116 case GLSL_TYPE_UINT:
117 case GLSL_TYPE_INT:
118 case GLSL_TYPE_FLOAT:
119 case GLSL_TYPE_DOUBLE:
120 case GLSL_TYPE_BOOL:
121 case GLSL_TYPE_ARRAY: {
122 deref_type = deref_type->array_element;
123
124 nir_deref_array *deref_arr = nir_deref_array_create(b);
125 deref_arr->deref.type = deref_type->type;
126
127 if (chain->link[i].mode == vtn_access_mode_literal) {
128 deref_arr->deref_array_type = nir_deref_array_type_direct;
129 deref_arr->base_offset = chain->link[i].id;
130 } else {
131 assert(chain->link[i].mode == vtn_access_mode_id);
132 deref_arr->deref_array_type = nir_deref_array_type_indirect;
133 deref_arr->base_offset = 0;
134 deref_arr->indirect =
135 nir_src_for_ssa(vtn_ssa_value(b, chain->link[i].id)->def);
136 }
137 tail->child = &deref_arr->deref;
138 tail = tail->child;
139 break;
140 }
141
142 case GLSL_TYPE_STRUCT: {
143 assert(chain->link[i].mode == vtn_access_mode_literal);
144 unsigned idx = chain->link[i].id;
145 deref_type = deref_type->members[idx];
146 if (members) {
147 /* This is a pre-split structure. */
148 deref_var->var = members[idx];
149 rewrite_deref_types(&deref_var->deref, members[idx]->type);
150 assert(tail->type == deref_type->type);
151 members = NULL;
152 } else {
153 nir_deref_struct *deref_struct = nir_deref_struct_create(b, idx);
154 deref_struct->deref.type = deref_type->type;
155 tail->child = &deref_struct->deref;
156 tail = tail->child;
157 }
158 break;
159 }
160 default:
161 unreachable("Invalid type for deref");
162 }
163 }
164
165 assert(members == NULL);
166 return deref_var;
167 }
168
169 static void
170 _vtn_local_load_store(struct vtn_builder *b, bool load, nir_deref_var *deref,
171 nir_deref *tail, struct vtn_ssa_value *inout)
172 {
173 /* The deref tail may contain a deref to select a component of a vector (in
174 * other words, it might not be an actual tail) so we have to save it away
175 * here since we overwrite it later.
176 */
177 nir_deref *old_child = tail->child;
178
179 if (glsl_type_is_vector_or_scalar(tail->type)) {
180 /* Terminate the deref chain in case there is one more link to pick
181 * off a component of the vector.
182 */
183 tail->child = NULL;
184
185 nir_intrinsic_op op = load ? nir_intrinsic_load_var :
186 nir_intrinsic_store_var;
187
188 nir_intrinsic_instr *intrin = nir_intrinsic_instr_create(b->shader, op);
189 intrin->variables[0] = nir_deref_var_clone(deref, intrin);
190 intrin->num_components = glsl_get_vector_elements(tail->type);
191
192 if (load) {
193 nir_ssa_dest_init(&intrin->instr, &intrin->dest,
194 intrin->num_components,
195 glsl_get_bit_size(tail->type),
196 NULL);
197 inout->def = &intrin->dest.ssa;
198 } else {
199 nir_intrinsic_set_write_mask(intrin, (1 << intrin->num_components) - 1);
200 intrin->src[0] = nir_src_for_ssa(inout->def);
201 }
202
203 nir_builder_instr_insert(&b->nb, &intrin->instr);
204 } else if (glsl_get_base_type(tail->type) == GLSL_TYPE_ARRAY ||
205 glsl_type_is_matrix(tail->type)) {
206 unsigned elems = glsl_get_length(tail->type);
207 nir_deref_array *deref_arr = nir_deref_array_create(b);
208 deref_arr->deref_array_type = nir_deref_array_type_direct;
209 deref_arr->deref.type = glsl_get_array_element(tail->type);
210 tail->child = &deref_arr->deref;
211 for (unsigned i = 0; i < elems; i++) {
212 deref_arr->base_offset = i;
213 _vtn_local_load_store(b, load, deref, tail->child, inout->elems[i]);
214 }
215 } else {
216 assert(glsl_get_base_type(tail->type) == GLSL_TYPE_STRUCT);
217 unsigned elems = glsl_get_length(tail->type);
218 nir_deref_struct *deref_struct = nir_deref_struct_create(b, 0);
219 tail->child = &deref_struct->deref;
220 for (unsigned i = 0; i < elems; i++) {
221 deref_struct->index = i;
222 deref_struct->deref.type = glsl_get_struct_field(tail->type, i);
223 _vtn_local_load_store(b, load, deref, tail->child, inout->elems[i]);
224 }
225 }
226
227 tail->child = old_child;
228 }
229
230 nir_deref_var *
231 vtn_nir_deref(struct vtn_builder *b, uint32_t id)
232 {
233 struct vtn_access_chain *chain =
234 vtn_value(b, id, vtn_value_type_access_chain)->access_chain;
235
236 return vtn_access_chain_to_deref(b, chain);
237 }
238
239 /*
240 * Gets the NIR-level deref tail, which may have as a child an array deref
241 * selecting which component due to OpAccessChain supporting per-component
242 * indexing in SPIR-V.
243 */
244 static nir_deref *
245 get_deref_tail(nir_deref_var *deref)
246 {
247 nir_deref *cur = &deref->deref;
248 while (!glsl_type_is_vector_or_scalar(cur->type) && cur->child)
249 cur = cur->child;
250
251 return cur;
252 }
253
254 struct vtn_ssa_value *
255 vtn_local_load(struct vtn_builder *b, nir_deref_var *src)
256 {
257 nir_deref *src_tail = get_deref_tail(src);
258 struct vtn_ssa_value *val = vtn_create_ssa_value(b, src_tail->type);
259 _vtn_local_load_store(b, true, src, src_tail, val);
260
261 if (src_tail->child) {
262 nir_deref_array *vec_deref = nir_deref_as_array(src_tail->child);
263 assert(vec_deref->deref.child == NULL);
264 val->type = vec_deref->deref.type;
265 if (vec_deref->deref_array_type == nir_deref_array_type_direct)
266 val->def = vtn_vector_extract(b, val->def, vec_deref->base_offset);
267 else
268 val->def = vtn_vector_extract_dynamic(b, val->def,
269 vec_deref->indirect.ssa);
270 }
271
272 return val;
273 }
274
275 void
276 vtn_local_store(struct vtn_builder *b, struct vtn_ssa_value *src,
277 nir_deref_var *dest)
278 {
279 nir_deref *dest_tail = get_deref_tail(dest);
280
281 if (dest_tail->child) {
282 struct vtn_ssa_value *val = vtn_create_ssa_value(b, dest_tail->type);
283 _vtn_local_load_store(b, true, dest, dest_tail, val);
284 nir_deref_array *deref = nir_deref_as_array(dest_tail->child);
285 assert(deref->deref.child == NULL);
286 if (deref->deref_array_type == nir_deref_array_type_direct)
287 val->def = vtn_vector_insert(b, val->def, src->def,
288 deref->base_offset);
289 else
290 val->def = vtn_vector_insert_dynamic(b, val->def, src->def,
291 deref->indirect.ssa);
292 _vtn_local_load_store(b, false, dest, dest_tail, val);
293 } else {
294 _vtn_local_load_store(b, false, dest, dest_tail, src);
295 }
296 }
297
298 static nir_ssa_def *
299 get_vulkan_resource_index(struct vtn_builder *b, struct vtn_access_chain *chain,
300 struct vtn_type **type, unsigned *chain_idx)
301 {
302 /* Push constants have no explicit binding */
303 if (chain->var->mode == vtn_variable_mode_push_constant) {
304 *chain_idx = 0;
305 *type = chain->var->type;
306 return NULL;
307 }
308
309 nir_ssa_def *array_index;
310 if (glsl_type_is_array(chain->var->type->type)) {
311 assert(chain->length > 0);
312 array_index = vtn_access_link_as_ssa(b, chain->link[0], 1);
313 *chain_idx = 1;
314 *type = chain->var->type->array_element;
315 } else {
316 array_index = nir_imm_int(&b->nb, 0);
317 *chain_idx = 0;
318 *type = chain->var->type;
319 }
320
321 nir_intrinsic_instr *instr =
322 nir_intrinsic_instr_create(b->nb.shader,
323 nir_intrinsic_vulkan_resource_index);
324 instr->src[0] = nir_src_for_ssa(array_index);
325 nir_intrinsic_set_desc_set(instr, chain->var->descriptor_set);
326 nir_intrinsic_set_binding(instr, chain->var->binding);
327
328 nir_ssa_dest_init(&instr->instr, &instr->dest, 1, 32, NULL);
329 nir_builder_instr_insert(&b->nb, &instr->instr);
330
331 return &instr->dest.ssa;
332 }
333
334 nir_ssa_def *
335 vtn_access_chain_to_offset(struct vtn_builder *b,
336 struct vtn_access_chain *chain,
337 nir_ssa_def **index_out, struct vtn_type **type_out,
338 unsigned *end_idx_out, bool stop_at_matrix)
339 {
340 unsigned idx = 0;
341 struct vtn_type *type;
342 *index_out = get_vulkan_resource_index(b, chain, &type, &idx);
343
344 nir_ssa_def *offset = nir_imm_int(&b->nb, 0);
345 for (; idx < chain->length; idx++) {
346 enum glsl_base_type base_type = glsl_get_base_type(type->type);
347 switch (base_type) {
348 case GLSL_TYPE_UINT:
349 case GLSL_TYPE_INT:
350 case GLSL_TYPE_FLOAT:
351 case GLSL_TYPE_DOUBLE:
352 case GLSL_TYPE_BOOL:
353 /* Some users may not want matrix or vector derefs */
354 if (stop_at_matrix)
355 goto end;
356 /* Fall through */
357
358 case GLSL_TYPE_ARRAY:
359 offset = nir_iadd(&b->nb, offset,
360 vtn_access_link_as_ssa(b, chain->link[idx],
361 type->stride));
362
363 type = type->array_element;
364 break;
365
366 case GLSL_TYPE_STRUCT: {
367 assert(chain->link[idx].mode == vtn_access_mode_literal);
368 unsigned member = chain->link[idx].id;
369 offset = nir_iadd(&b->nb, offset,
370 nir_imm_int(&b->nb, type->offsets[member]));
371 type = type->members[member];
372 break;
373 }
374
375 default:
376 unreachable("Invalid type for deref");
377 }
378 }
379
380 end:
381 *type_out = type;
382 if (end_idx_out)
383 *end_idx_out = idx;
384
385 return offset;
386 }
387
388 static void
389 _vtn_load_store_tail(struct vtn_builder *b, nir_intrinsic_op op, bool load,
390 nir_ssa_def *index, nir_ssa_def *offset,
391 struct vtn_ssa_value **inout, const struct glsl_type *type)
392 {
393 nir_intrinsic_instr *instr = nir_intrinsic_instr_create(b->nb.shader, op);
394 instr->num_components = glsl_get_vector_elements(type);
395
396 int src = 0;
397 if (!load) {
398 nir_intrinsic_set_write_mask(instr, (1 << instr->num_components) - 1);
399 instr->src[src++] = nir_src_for_ssa((*inout)->def);
400 }
401
402 /* We set the base and size for push constant load to the entire push
403 * constant block for now.
404 */
405 if (op == nir_intrinsic_load_push_constant) {
406 nir_intrinsic_set_base(instr, 0);
407 nir_intrinsic_set_range(instr, 128);
408 }
409
410 if (index)
411 instr->src[src++] = nir_src_for_ssa(index);
412
413 instr->src[src++] = nir_src_for_ssa(offset);
414
415 if (load) {
416 nir_ssa_dest_init(&instr->instr, &instr->dest,
417 instr->num_components,
418 glsl_get_bit_size(type), NULL);
419 (*inout)->def = &instr->dest.ssa;
420 }
421
422 nir_builder_instr_insert(&b->nb, &instr->instr);
423
424 if (load && glsl_get_base_type(type) == GLSL_TYPE_BOOL)
425 (*inout)->def = nir_ine(&b->nb, (*inout)->def, nir_imm_int(&b->nb, 0));
426 }
427
428 static void
429 _vtn_block_load_store(struct vtn_builder *b, nir_intrinsic_op op, bool load,
430 nir_ssa_def *index, nir_ssa_def *offset,
431 struct vtn_access_chain *chain, unsigned chain_idx,
432 struct vtn_type *type, struct vtn_ssa_value **inout)
433 {
434 if (chain && chain_idx >= chain->length)
435 chain = NULL;
436
437 if (load && chain == NULL && *inout == NULL)
438 *inout = vtn_create_ssa_value(b, type->type);
439
440 enum glsl_base_type base_type = glsl_get_base_type(type->type);
441 switch (base_type) {
442 case GLSL_TYPE_UINT:
443 case GLSL_TYPE_INT:
444 case GLSL_TYPE_FLOAT:
445 case GLSL_TYPE_BOOL:
446 /* This is where things get interesting. At this point, we've hit
447 * a vector, a scalar, or a matrix.
448 */
449 if (glsl_type_is_matrix(type->type)) {
450 if (chain == NULL) {
451 /* Loading the whole matrix */
452 struct vtn_ssa_value *transpose;
453 unsigned num_ops, vec_width;
454 if (type->row_major) {
455 num_ops = glsl_get_vector_elements(type->type);
456 vec_width = glsl_get_matrix_columns(type->type);
457 if (load) {
458 const struct glsl_type *transpose_type =
459 glsl_matrix_type(base_type, vec_width, num_ops);
460 *inout = vtn_create_ssa_value(b, transpose_type);
461 } else {
462 transpose = vtn_ssa_transpose(b, *inout);
463 inout = &transpose;
464 }
465 } else {
466 num_ops = glsl_get_matrix_columns(type->type);
467 vec_width = glsl_get_vector_elements(type->type);
468 }
469
470 for (unsigned i = 0; i < num_ops; i++) {
471 nir_ssa_def *elem_offset =
472 nir_iadd(&b->nb, offset,
473 nir_imm_int(&b->nb, i * type->stride));
474 _vtn_load_store_tail(b, op, load, index, elem_offset,
475 &(*inout)->elems[i],
476 glsl_vector_type(base_type, vec_width));
477 }
478
479 if (load && type->row_major)
480 *inout = vtn_ssa_transpose(b, *inout);
481 } else if (type->row_major) {
482 /* Row-major but with an access chiain. */
483 nir_ssa_def *col_offset =
484 vtn_access_link_as_ssa(b, chain->link[chain_idx],
485 type->array_element->stride);
486 offset = nir_iadd(&b->nb, offset, col_offset);
487
488 if (chain_idx + 1 < chain->length) {
489 /* Picking off a single element */
490 nir_ssa_def *row_offset =
491 vtn_access_link_as_ssa(b, chain->link[chain_idx + 1],
492 type->stride);
493 offset = nir_iadd(&b->nb, offset, row_offset);
494 if (load)
495 *inout = vtn_create_ssa_value(b, glsl_scalar_type(base_type));
496 _vtn_load_store_tail(b, op, load, index, offset, inout,
497 glsl_scalar_type(base_type));
498 } else {
499 /* Grabbing a column; picking one element off each row */
500 unsigned num_comps = glsl_get_vector_elements(type->type);
501 const struct glsl_type *column_type =
502 glsl_get_column_type(type->type);
503
504 nir_ssa_def *comps[4];
505 for (unsigned i = 0; i < num_comps; i++) {
506 nir_ssa_def *elem_offset =
507 nir_iadd(&b->nb, offset,
508 nir_imm_int(&b->nb, i * type->stride));
509
510 struct vtn_ssa_value *comp, temp_val;
511 if (!load) {
512 temp_val.def = nir_channel(&b->nb, (*inout)->def, i);
513 temp_val.type = glsl_scalar_type(base_type);
514 }
515 comp = &temp_val;
516 _vtn_load_store_tail(b, op, load, index, elem_offset,
517 &comp, glsl_scalar_type(base_type));
518 comps[i] = comp->def;
519 }
520
521 if (load) {
522 if (*inout == NULL)
523 *inout = vtn_create_ssa_value(b, column_type);
524
525 (*inout)->def = nir_vec(&b->nb, comps, num_comps);
526 }
527 }
528 } else {
529 /* Column-major with a deref. Fall through to array case. */
530 nir_ssa_def *col_offset =
531 vtn_access_link_as_ssa(b, chain->link[chain_idx], type->stride);
532 offset = nir_iadd(&b->nb, offset, col_offset);
533
534 _vtn_block_load_store(b, op, load, index, offset,
535 chain, chain_idx + 1,
536 type->array_element, inout);
537 }
538 } else if (chain == NULL) {
539 /* Single whole vector */
540 assert(glsl_type_is_vector_or_scalar(type->type));
541 _vtn_load_store_tail(b, op, load, index, offset, inout, type->type);
542 } else {
543 /* Single component of a vector. Fall through to array case. */
544 nir_ssa_def *elem_offset =
545 vtn_access_link_as_ssa(b, chain->link[chain_idx], type->stride);
546 offset = nir_iadd(&b->nb, offset, elem_offset);
547
548 _vtn_block_load_store(b, op, load, index, offset, NULL, 0,
549 type->array_element, inout);
550 }
551 return;
552
553 case GLSL_TYPE_ARRAY: {
554 unsigned elems = glsl_get_length(type->type);
555 for (unsigned i = 0; i < elems; i++) {
556 nir_ssa_def *elem_off =
557 nir_iadd(&b->nb, offset, nir_imm_int(&b->nb, i * type->stride));
558 _vtn_block_load_store(b, op, load, index, elem_off, NULL, 0,
559 type->array_element, &(*inout)->elems[i]);
560 }
561 return;
562 }
563
564 case GLSL_TYPE_STRUCT: {
565 unsigned elems = glsl_get_length(type->type);
566 for (unsigned i = 0; i < elems; i++) {
567 nir_ssa_def *elem_off =
568 nir_iadd(&b->nb, offset, nir_imm_int(&b->nb, type->offsets[i]));
569 _vtn_block_load_store(b, op, load, index, elem_off, NULL, 0,
570 type->members[i], &(*inout)->elems[i]);
571 }
572 return;
573 }
574
575 default:
576 unreachable("Invalid block member type");
577 }
578 }
579
580 static struct vtn_ssa_value *
581 vtn_block_load(struct vtn_builder *b, struct vtn_access_chain *src)
582 {
583 nir_intrinsic_op op;
584 switch (src->var->mode) {
585 case vtn_variable_mode_ubo:
586 op = nir_intrinsic_load_ubo;
587 break;
588 case vtn_variable_mode_ssbo:
589 op = nir_intrinsic_load_ssbo;
590 break;
591 case vtn_variable_mode_push_constant:
592 op = nir_intrinsic_load_push_constant;
593 break;
594 default:
595 assert(!"Invalid block variable mode");
596 }
597
598 nir_ssa_def *offset, *index = NULL;
599 struct vtn_type *type;
600 unsigned chain_idx;
601 offset = vtn_access_chain_to_offset(b, src, &index, &type, &chain_idx, true);
602
603 struct vtn_ssa_value *value = NULL;
604 _vtn_block_load_store(b, op, true, index, offset,
605 src, chain_idx, type, &value);
606 return value;
607 }
608
609 static void
610 vtn_block_store(struct vtn_builder *b, struct vtn_ssa_value *src,
611 struct vtn_access_chain *dst)
612 {
613 nir_ssa_def *offset, *index = NULL;
614 struct vtn_type *type;
615 unsigned chain_idx;
616 offset = vtn_access_chain_to_offset(b, dst, &index, &type, &chain_idx, true);
617
618 _vtn_block_load_store(b, nir_intrinsic_store_ssbo, false, index, offset,
619 dst, chain_idx, type, &src);
620 }
621
622 static bool
623 vtn_variable_is_external_block(struct vtn_variable *var)
624 {
625 return var->mode == vtn_variable_mode_ssbo ||
626 var->mode == vtn_variable_mode_ubo ||
627 var->mode == vtn_variable_mode_push_constant;
628 }
629
630 static void
631 _vtn_variable_load_store(struct vtn_builder *b, bool load,
632 struct vtn_access_chain *chain,
633 struct vtn_type *tail_type,
634 struct vtn_ssa_value **inout)
635 {
636 enum glsl_base_type base_type = glsl_get_base_type(tail_type->type);
637 switch (base_type) {
638 case GLSL_TYPE_UINT:
639 case GLSL_TYPE_INT:
640 case GLSL_TYPE_FLOAT:
641 case GLSL_TYPE_BOOL:
642 /* At this point, we have a scalar, vector, or matrix so we know that
643 * there cannot be any structure splitting still in the way. By
644 * stopping at the matrix level rather than the vector level, we
645 * ensure that matrices get loaded in the optimal way even if they
646 * are storred row-major in a UBO.
647 */
648 if (load) {
649 *inout = vtn_local_load(b, vtn_access_chain_to_deref(b, chain));
650 } else {
651 vtn_local_store(b, *inout, vtn_access_chain_to_deref(b, chain));
652 }
653 return;
654
655 case GLSL_TYPE_ARRAY:
656 case GLSL_TYPE_STRUCT: {
657 struct vtn_access_chain *new_chain =
658 vtn_access_chain_extend(b, chain, 1);
659 new_chain->link[chain->length].mode = vtn_access_mode_literal;
660 unsigned elems = glsl_get_length(tail_type->type);
661 if (load) {
662 assert(*inout == NULL);
663 *inout = rzalloc(b, struct vtn_ssa_value);
664 (*inout)->type = tail_type->type;
665 (*inout)->elems = rzalloc_array(b, struct vtn_ssa_value *, elems);
666 }
667 for (unsigned i = 0; i < elems; i++) {
668 new_chain->link[chain->length].id = i;
669 struct vtn_type *elem_type = base_type == GLSL_TYPE_ARRAY ?
670 tail_type->array_element : tail_type->members[i];
671 _vtn_variable_load_store(b, load, new_chain, elem_type,
672 &(*inout)->elems[i]);
673 }
674 return;
675 }
676
677 default:
678 unreachable("Invalid access chain type");
679 }
680 }
681
682 struct vtn_ssa_value *
683 vtn_variable_load(struct vtn_builder *b, struct vtn_access_chain *src)
684 {
685 if (vtn_variable_is_external_block(src->var)) {
686 return vtn_block_load(b, src);
687 } else {
688 struct vtn_type *tail_type = vtn_access_chain_tail_type(b, src);
689 struct vtn_ssa_value *val = NULL;
690 _vtn_variable_load_store(b, true, src, tail_type, &val);
691 return val;
692 }
693 }
694
695 void
696 vtn_variable_store(struct vtn_builder *b, struct vtn_ssa_value *src,
697 struct vtn_access_chain *dest)
698 {
699 if (vtn_variable_is_external_block(dest->var)) {
700 assert(dest->var->mode == vtn_variable_mode_ssbo);
701 vtn_block_store(b, src, dest);
702 } else {
703 struct vtn_type *tail_type = vtn_access_chain_tail_type(b, dest);
704 _vtn_variable_load_store(b, false, dest, tail_type, &src);
705 }
706 }
707
708 static void
709 _vtn_variable_copy(struct vtn_builder *b, struct vtn_access_chain *dest,
710 struct vtn_access_chain *src, struct vtn_type *tail_type)
711 {
712 enum glsl_base_type base_type = glsl_get_base_type(tail_type->type);
713 switch (base_type) {
714 case GLSL_TYPE_UINT:
715 case GLSL_TYPE_INT:
716 case GLSL_TYPE_FLOAT:
717 case GLSL_TYPE_BOOL:
718 /* At this point, we have a scalar, vector, or matrix so we know that
719 * there cannot be any structure splitting still in the way. By
720 * stopping at the matrix level rather than the vector level, we
721 * ensure that matrices get loaded in the optimal way even if they
722 * are storred row-major in a UBO.
723 */
724 vtn_variable_store(b, vtn_variable_load(b, src), dest);
725 return;
726
727 case GLSL_TYPE_ARRAY:
728 case GLSL_TYPE_STRUCT: {
729 struct vtn_access_chain *new_src, *new_dest;
730 new_src = vtn_access_chain_extend(b, src, 1);
731 new_dest = vtn_access_chain_extend(b, dest, 1);
732 new_src->link[src->length].mode = vtn_access_mode_literal;
733 new_dest->link[dest->length].mode = vtn_access_mode_literal;
734 unsigned elems = glsl_get_length(tail_type->type);
735 for (unsigned i = 0; i < elems; i++) {
736 new_src->link[src->length].id = i;
737 new_dest->link[dest->length].id = i;
738 struct vtn_type *elem_type = base_type == GLSL_TYPE_ARRAY ?
739 tail_type->array_element : tail_type->members[i];
740 _vtn_variable_copy(b, new_dest, new_src, elem_type);
741 }
742 return;
743 }
744
745 default:
746 unreachable("Invalid access chain type");
747 }
748 }
749
750 static void
751 vtn_variable_copy(struct vtn_builder *b, struct vtn_access_chain *dest,
752 struct vtn_access_chain *src)
753 {
754 struct vtn_type *tail_type = vtn_access_chain_tail_type(b, src);
755 assert(vtn_access_chain_tail_type(b, dest)->type == tail_type->type);
756
757 /* TODO: At some point, we should add a special-case for when we can
758 * just emit a copy_var intrinsic.
759 */
760 _vtn_variable_copy(b, dest, src, tail_type);
761 }
762
763 static void
764 set_mode_system_value(nir_variable_mode *mode)
765 {
766 assert(*mode == nir_var_system_value || *mode == nir_var_shader_in);
767 *mode = nir_var_system_value;
768 }
769
770 static void
771 vtn_get_builtin_location(struct vtn_builder *b,
772 SpvBuiltIn builtin, int *location,
773 nir_variable_mode *mode)
774 {
775 switch (builtin) {
776 case SpvBuiltInPosition:
777 *location = VARYING_SLOT_POS;
778 break;
779 case SpvBuiltInPointSize:
780 *location = VARYING_SLOT_PSIZ;
781 break;
782 case SpvBuiltInClipDistance:
783 *location = VARYING_SLOT_CLIP_DIST0; /* XXX CLIP_DIST1? */
784 break;
785 case SpvBuiltInCullDistance:
786 *location = VARYING_SLOT_CULL_DIST0;
787 break;
788 case SpvBuiltInVertexIndex:
789 *location = SYSTEM_VALUE_VERTEX_ID;
790 set_mode_system_value(mode);
791 break;
792 case SpvBuiltInVertexId:
793 /* Vulkan defines VertexID to be zero-based and reserves the new
794 * builtin keyword VertexIndex to indicate the non-zero-based value.
795 */
796 *location = SYSTEM_VALUE_VERTEX_ID_ZERO_BASE;
797 set_mode_system_value(mode);
798 break;
799 case SpvBuiltInInstanceIndex:
800 *location = SYSTEM_VALUE_INSTANCE_INDEX;
801 set_mode_system_value(mode);
802 break;
803 case SpvBuiltInInstanceId:
804 *location = SYSTEM_VALUE_INSTANCE_ID;
805 set_mode_system_value(mode);
806 break;
807 case SpvBuiltInPrimitiveId:
808 if (*mode == nir_var_shader_out) {
809 *location = VARYING_SLOT_PRIMITIVE_ID;
810 } else {
811 *location = SYSTEM_VALUE_PRIMITIVE_ID;
812 set_mode_system_value(mode);
813 }
814 break;
815 case SpvBuiltInInvocationId:
816 *location = SYSTEM_VALUE_INVOCATION_ID;
817 set_mode_system_value(mode);
818 break;
819 case SpvBuiltInLayer:
820 *location = VARYING_SLOT_LAYER;
821 if (b->shader->stage == MESA_SHADER_FRAGMENT)
822 *mode = nir_var_shader_in;
823 else if (b->shader->stage == MESA_SHADER_GEOMETRY)
824 *mode = nir_var_shader_out;
825 else
826 unreachable("invalid stage for SpvBuiltInLayer");
827 break;
828 case SpvBuiltInViewportIndex:
829 *location = VARYING_SLOT_VIEWPORT;
830 if (b->shader->stage == MESA_SHADER_GEOMETRY)
831 *mode = nir_var_shader_out;
832 else if (b->shader->stage == MESA_SHADER_FRAGMENT)
833 *mode = nir_var_shader_in;
834 else
835 unreachable("invalid stage for SpvBuiltInViewportIndex");
836 break;
837 case SpvBuiltInTessLevelOuter:
838 case SpvBuiltInTessLevelInner:
839 case SpvBuiltInTessCoord:
840 case SpvBuiltInPatchVertices:
841 unreachable("no tessellation support");
842 case SpvBuiltInFragCoord:
843 *location = VARYING_SLOT_POS;
844 assert(*mode == nir_var_shader_in);
845 break;
846 case SpvBuiltInPointCoord:
847 *location = VARYING_SLOT_PNTC;
848 assert(*mode == nir_var_shader_in);
849 break;
850 case SpvBuiltInFrontFacing:
851 *location = SYSTEM_VALUE_FRONT_FACE;
852 set_mode_system_value(mode);
853 break;
854 case SpvBuiltInSampleId:
855 *location = SYSTEM_VALUE_SAMPLE_ID;
856 set_mode_system_value(mode);
857 break;
858 case SpvBuiltInSamplePosition:
859 *location = SYSTEM_VALUE_SAMPLE_POS;
860 set_mode_system_value(mode);
861 break;
862 case SpvBuiltInSampleMask:
863 *location = SYSTEM_VALUE_SAMPLE_MASK_IN; /* XXX out? */
864 set_mode_system_value(mode);
865 break;
866 case SpvBuiltInFragDepth:
867 *location = FRAG_RESULT_DEPTH;
868 assert(*mode == nir_var_shader_out);
869 break;
870 case SpvBuiltInNumWorkgroups:
871 *location = SYSTEM_VALUE_NUM_WORK_GROUPS;
872 set_mode_system_value(mode);
873 break;
874 case SpvBuiltInWorkgroupSize:
875 /* This should already be handled */
876 unreachable("unsupported builtin");
877 break;
878 case SpvBuiltInWorkgroupId:
879 *location = SYSTEM_VALUE_WORK_GROUP_ID;
880 set_mode_system_value(mode);
881 break;
882 case SpvBuiltInLocalInvocationId:
883 *location = SYSTEM_VALUE_LOCAL_INVOCATION_ID;
884 set_mode_system_value(mode);
885 break;
886 case SpvBuiltInLocalInvocationIndex:
887 *location = SYSTEM_VALUE_LOCAL_INVOCATION_INDEX;
888 set_mode_system_value(mode);
889 break;
890 case SpvBuiltInGlobalInvocationId:
891 *location = SYSTEM_VALUE_GLOBAL_INVOCATION_ID;
892 set_mode_system_value(mode);
893 break;
894 case SpvBuiltInHelperInvocation:
895 default:
896 unreachable("unsupported builtin");
897 }
898 }
899
900 static void
901 apply_var_decoration(struct vtn_builder *b, nir_variable *nir_var,
902 const struct vtn_decoration *dec)
903 {
904 switch (dec->decoration) {
905 case SpvDecorationRelaxedPrecision:
906 break; /* FIXME: Do nothing with this for now. */
907 case SpvDecorationNoPerspective:
908 nir_var->data.interpolation = INTERP_MODE_NOPERSPECTIVE;
909 break;
910 case SpvDecorationFlat:
911 nir_var->data.interpolation = INTERP_MODE_FLAT;
912 break;
913 case SpvDecorationCentroid:
914 nir_var->data.centroid = true;
915 break;
916 case SpvDecorationSample:
917 nir_var->data.sample = true;
918 break;
919 case SpvDecorationInvariant:
920 nir_var->data.invariant = true;
921 break;
922 case SpvDecorationConstant:
923 assert(nir_var->constant_initializer != NULL);
924 nir_var->data.read_only = true;
925 break;
926 case SpvDecorationNonWritable:
927 nir_var->data.read_only = true;
928 break;
929 case SpvDecorationComponent:
930 nir_var->data.location_frac = dec->literals[0];
931 break;
932 case SpvDecorationIndex:
933 nir_var->data.index = dec->literals[0];
934 break;
935 case SpvDecorationBuiltIn: {
936 SpvBuiltIn builtin = dec->literals[0];
937
938 if (builtin == SpvBuiltInWorkgroupSize) {
939 /* This shouldn't be a builtin. It's actually a constant. */
940 nir_var->data.mode = nir_var_global;
941 nir_var->data.read_only = true;
942
943 nir_constant *c = rzalloc(nir_var, nir_constant);
944 c->values[0].u32[0] = b->shader->info->cs.local_size[0];
945 c->values[0].u32[1] = b->shader->info->cs.local_size[1];
946 c->values[0].u32[2] = b->shader->info->cs.local_size[2];
947 nir_var->constant_initializer = c;
948 break;
949 }
950
951 nir_variable_mode mode = nir_var->data.mode;
952 vtn_get_builtin_location(b, builtin, &nir_var->data.location, &mode);
953 nir_var->data.mode = mode;
954
955 if (builtin == SpvBuiltInFragCoord || builtin == SpvBuiltInSamplePosition)
956 nir_var->data.origin_upper_left = b->origin_upper_left;
957
958 if (builtin == SpvBuiltInFragCoord)
959 nir_var->data.pixel_center_integer = b->pixel_center_integer;
960 break;
961 }
962
963 case SpvDecorationSpecId:
964 case SpvDecorationRowMajor:
965 case SpvDecorationColMajor:
966 case SpvDecorationMatrixStride:
967 case SpvDecorationRestrict:
968 case SpvDecorationAliased:
969 case SpvDecorationVolatile:
970 case SpvDecorationCoherent:
971 case SpvDecorationNonReadable:
972 case SpvDecorationUniform:
973 case SpvDecorationStream:
974 case SpvDecorationOffset:
975 case SpvDecorationLinkageAttributes:
976 break; /* Do nothing with these here */
977
978 case SpvDecorationPatch:
979 vtn_warn("Tessellation not yet supported");
980 break;
981
982 case SpvDecorationLocation:
983 unreachable("Handled above");
984
985 case SpvDecorationBlock:
986 case SpvDecorationBufferBlock:
987 case SpvDecorationArrayStride:
988 case SpvDecorationGLSLShared:
989 case SpvDecorationGLSLPacked:
990 break; /* These can apply to a type but we don't care about them */
991
992 case SpvDecorationBinding:
993 case SpvDecorationDescriptorSet:
994 case SpvDecorationNoContraction:
995 case SpvDecorationInputAttachmentIndex:
996 vtn_warn("Decoration not allowed for variable or structure member: %s",
997 spirv_decoration_to_string(dec->decoration));
998 break;
999
1000 case SpvDecorationXfbBuffer:
1001 case SpvDecorationXfbStride:
1002 vtn_warn("Vulkan does not have transform feedback: %s",
1003 spirv_decoration_to_string(dec->decoration));
1004 break;
1005
1006 case SpvDecorationCPacked:
1007 case SpvDecorationSaturatedConversion:
1008 case SpvDecorationFuncParamAttr:
1009 case SpvDecorationFPRoundingMode:
1010 case SpvDecorationFPFastMathMode:
1011 case SpvDecorationAlignment:
1012 vtn_warn("Decoraiton only allowed for CL-style kernels: %s",
1013 spirv_decoration_to_string(dec->decoration));
1014 break;
1015 }
1016 }
1017
1018 static void
1019 var_decoration_cb(struct vtn_builder *b, struct vtn_value *val, int member,
1020 const struct vtn_decoration *dec, void *void_var)
1021 {
1022 struct vtn_variable *vtn_var = void_var;
1023
1024 /* Handle decorations that apply to a vtn_variable as a whole */
1025 switch (dec->decoration) {
1026 case SpvDecorationBinding:
1027 vtn_var->binding = dec->literals[0];
1028 return;
1029 case SpvDecorationDescriptorSet:
1030 vtn_var->descriptor_set = dec->literals[0];
1031 return;
1032 case SpvDecorationInputAttachmentIndex:
1033 vtn_var->input_attachment_index = dec->literals[0];
1034 return;
1035 default:
1036 break;
1037 }
1038
1039 if (val->value_type == vtn_value_type_access_chain) {
1040 assert(val->access_chain->length == 0);
1041 assert(val->access_chain->var == void_var);
1042 assert(member == -1);
1043 } else {
1044 assert(val->value_type == vtn_value_type_type);
1045 }
1046
1047 /* Location is odd. If applied to a split structure, we have to walk the
1048 * whole thing and accumulate the location. It's easier to handle as a
1049 * special case.
1050 */
1051 if (dec->decoration == SpvDecorationLocation) {
1052 unsigned location = dec->literals[0];
1053 bool is_vertex_input;
1054 if (b->shader->stage == MESA_SHADER_FRAGMENT &&
1055 vtn_var->mode == vtn_variable_mode_output) {
1056 is_vertex_input = false;
1057 location += FRAG_RESULT_DATA0;
1058 } else if (b->shader->stage == MESA_SHADER_VERTEX &&
1059 vtn_var->mode == vtn_variable_mode_input) {
1060 is_vertex_input = true;
1061 location += VERT_ATTRIB_GENERIC0;
1062 } else if (vtn_var->mode == vtn_variable_mode_input ||
1063 vtn_var->mode == vtn_variable_mode_output) {
1064 is_vertex_input = false;
1065 location += VARYING_SLOT_VAR0;
1066 } else {
1067 unreachable("Location must be on input or output variable");
1068 }
1069
1070 if (vtn_var->var) {
1071 /* This handles the member and lone variable cases */
1072 vtn_var->var->data.location = location;
1073 } else {
1074 /* This handles the structure member case */
1075 assert(vtn_var->members);
1076 unsigned length =
1077 glsl_get_length(glsl_without_array(vtn_var->type->type));
1078 for (unsigned i = 0; i < length; i++) {
1079 vtn_var->members[i]->data.location = location;
1080 location +=
1081 glsl_count_attribute_slots(vtn_var->members[i]->interface_type,
1082 is_vertex_input);
1083 }
1084 }
1085 return;
1086 } else {
1087 if (vtn_var->var) {
1088 assert(member <= 0);
1089 apply_var_decoration(b, vtn_var->var, dec);
1090 } else if (vtn_var->members) {
1091 if (member >= 0) {
1092 assert(vtn_var->members);
1093 apply_var_decoration(b, vtn_var->members[member], dec);
1094 } else {
1095 unsigned length =
1096 glsl_get_length(glsl_without_array(vtn_var->type->type));
1097 for (unsigned i = 0; i < length; i++)
1098 apply_var_decoration(b, vtn_var->members[i], dec);
1099 }
1100 } else {
1101 /* A few variables, those with external storage, have no actual
1102 * nir_variables associated with them. Fortunately, all decorations
1103 * we care about for those variables are on the type only.
1104 */
1105 assert(vtn_var->mode == vtn_variable_mode_ubo ||
1106 vtn_var->mode == vtn_variable_mode_ssbo ||
1107 vtn_var->mode == vtn_variable_mode_push_constant);
1108 }
1109 }
1110 }
1111
1112 /* Tries to compute the size of an interface block based on the strides and
1113 * offsets that are provided to us in the SPIR-V source.
1114 */
1115 static unsigned
1116 vtn_type_block_size(struct vtn_type *type)
1117 {
1118 enum glsl_base_type base_type = glsl_get_base_type(type->type);
1119 switch (base_type) {
1120 case GLSL_TYPE_UINT:
1121 case GLSL_TYPE_INT:
1122 case GLSL_TYPE_FLOAT:
1123 case GLSL_TYPE_BOOL:
1124 case GLSL_TYPE_DOUBLE: {
1125 unsigned cols = type->row_major ? glsl_get_vector_elements(type->type) :
1126 glsl_get_matrix_columns(type->type);
1127 if (cols > 1) {
1128 assert(type->stride > 0);
1129 return type->stride * cols;
1130 } else if (base_type == GLSL_TYPE_DOUBLE) {
1131 return glsl_get_vector_elements(type->type) * 8;
1132 } else {
1133 return glsl_get_vector_elements(type->type) * 4;
1134 }
1135 }
1136
1137 case GLSL_TYPE_STRUCT:
1138 case GLSL_TYPE_INTERFACE: {
1139 unsigned size = 0;
1140 unsigned num_fields = glsl_get_length(type->type);
1141 for (unsigned f = 0; f < num_fields; f++) {
1142 unsigned field_end = type->offsets[f] +
1143 vtn_type_block_size(type->members[f]);
1144 size = MAX2(size, field_end);
1145 }
1146 return size;
1147 }
1148
1149 case GLSL_TYPE_ARRAY:
1150 assert(type->stride > 0);
1151 assert(glsl_get_length(type->type) > 0);
1152 return type->stride * glsl_get_length(type->type);
1153
1154 default:
1155 assert(!"Invalid block type");
1156 return 0;
1157 }
1158 }
1159
1160 void
1161 vtn_handle_variables(struct vtn_builder *b, SpvOp opcode,
1162 const uint32_t *w, unsigned count)
1163 {
1164 switch (opcode) {
1165 case SpvOpVariable: {
1166 struct vtn_variable *var = rzalloc(b, struct vtn_variable);
1167 var->type = vtn_value(b, w[1], vtn_value_type_type)->type;
1168
1169 var->chain.var = var;
1170 var->chain.length = 0;
1171
1172 struct vtn_value *val =
1173 vtn_push_value(b, w[2], vtn_value_type_access_chain);
1174 val->access_chain = &var->chain;
1175
1176 struct vtn_type *without_array = var->type;
1177 while(glsl_type_is_array(without_array->type))
1178 without_array = without_array->array_element;
1179
1180 nir_variable_mode nir_mode;
1181 switch ((SpvStorageClass)w[3]) {
1182 case SpvStorageClassUniform:
1183 case SpvStorageClassUniformConstant:
1184 if (without_array->block) {
1185 var->mode = vtn_variable_mode_ubo;
1186 b->shader->info->num_ubos++;
1187 } else if (without_array->buffer_block) {
1188 var->mode = vtn_variable_mode_ssbo;
1189 b->shader->info->num_ssbos++;
1190 } else if (glsl_type_is_image(without_array->type)) {
1191 var->mode = vtn_variable_mode_image;
1192 nir_mode = nir_var_uniform;
1193 b->shader->info->num_images++;
1194 } else if (glsl_type_is_sampler(without_array->type)) {
1195 var->mode = vtn_variable_mode_sampler;
1196 nir_mode = nir_var_uniform;
1197 b->shader->info->num_textures++;
1198 } else {
1199 assert(!"Invalid uniform variable type");
1200 }
1201 break;
1202 case SpvStorageClassPushConstant:
1203 var->mode = vtn_variable_mode_push_constant;
1204 assert(b->shader->num_uniforms == 0);
1205 b->shader->num_uniforms = vtn_type_block_size(var->type);
1206 break;
1207 case SpvStorageClassInput:
1208 var->mode = vtn_variable_mode_input;
1209 nir_mode = nir_var_shader_in;
1210 break;
1211 case SpvStorageClassOutput:
1212 var->mode = vtn_variable_mode_output;
1213 nir_mode = nir_var_shader_out;
1214 break;
1215 case SpvStorageClassPrivate:
1216 var->mode = vtn_variable_mode_global;
1217 nir_mode = nir_var_global;
1218 break;
1219 case SpvStorageClassFunction:
1220 var->mode = vtn_variable_mode_local;
1221 nir_mode = nir_var_local;
1222 break;
1223 case SpvStorageClassWorkgroup:
1224 var->mode = vtn_variable_mode_workgroup;
1225 nir_mode = nir_var_shared;
1226 break;
1227 case SpvStorageClassCrossWorkgroup:
1228 case SpvStorageClassGeneric:
1229 case SpvStorageClassAtomicCounter:
1230 default:
1231 unreachable("Unhandled variable storage class");
1232 }
1233
1234 switch (var->mode) {
1235 case vtn_variable_mode_local:
1236 case vtn_variable_mode_global:
1237 case vtn_variable_mode_image:
1238 case vtn_variable_mode_sampler:
1239 case vtn_variable_mode_workgroup:
1240 /* For these, we create the variable normally */
1241 var->var = rzalloc(b->shader, nir_variable);
1242 var->var->name = ralloc_strdup(var->var, val->name);
1243 var->var->type = var->type->type;
1244 var->var->data.mode = nir_mode;
1245
1246 switch (var->mode) {
1247 case vtn_variable_mode_image:
1248 case vtn_variable_mode_sampler:
1249 var->var->interface_type = without_array->type;
1250 break;
1251 default:
1252 var->var->interface_type = NULL;
1253 break;
1254 }
1255 break;
1256
1257 case vtn_variable_mode_input:
1258 case vtn_variable_mode_output: {
1259 /* For inputs and outputs, we immediately split structures. This
1260 * is for a couple of reasons. For one, builtins may all come in
1261 * a struct and we really want those split out into separate
1262 * variables. For another, interpolation qualifiers can be
1263 * applied to members of the top-level struct ane we need to be
1264 * able to preserve that information.
1265 */
1266
1267 int array_length = -1;
1268 struct vtn_type *interface_type = var->type;
1269 if (b->shader->stage == MESA_SHADER_GEOMETRY &&
1270 glsl_type_is_array(var->type->type)) {
1271 /* In Geometry shaders (and some tessellation), inputs come
1272 * in per-vertex arrays. However, some builtins come in
1273 * non-per-vertex, hence the need for the is_array check. In
1274 * any case, there are no non-builtin arrays allowed so this
1275 * check should be sufficient.
1276 */
1277 interface_type = var->type->array_element;
1278 array_length = glsl_get_length(var->type->type);
1279 }
1280
1281 if (glsl_type_is_struct(interface_type->type)) {
1282 /* It's a struct. Split it. */
1283 unsigned num_members = glsl_get_length(interface_type->type);
1284 var->members = ralloc_array(b, nir_variable *, num_members);
1285
1286 for (unsigned i = 0; i < num_members; i++) {
1287 const struct glsl_type *mtype = interface_type->members[i]->type;
1288 if (array_length >= 0)
1289 mtype = glsl_array_type(mtype, array_length);
1290
1291 var->members[i] = rzalloc(b->shader, nir_variable);
1292 var->members[i]->name =
1293 ralloc_asprintf(var->members[i], "%s.%d", val->name, i);
1294 var->members[i]->type = mtype;
1295 var->members[i]->interface_type =
1296 interface_type->members[i]->type;
1297 var->members[i]->data.mode = nir_mode;
1298 }
1299 } else {
1300 var->var = rzalloc(b->shader, nir_variable);
1301 var->var->name = ralloc_strdup(var->var, val->name);
1302 var->var->type = var->type->type;
1303 var->var->interface_type = interface_type->type;
1304 var->var->data.mode = nir_mode;
1305 }
1306
1307 /* For inputs and outputs, we need to grab locations and builtin
1308 * information from the interface type.
1309 */
1310 vtn_foreach_decoration(b, interface_type->val, var_decoration_cb, var);
1311 break;
1312
1313 case vtn_variable_mode_param:
1314 unreachable("Not created through OpVariable");
1315 }
1316
1317 case vtn_variable_mode_ubo:
1318 case vtn_variable_mode_ssbo:
1319 case vtn_variable_mode_push_constant:
1320 /* These don't need actual variables. */
1321 break;
1322 }
1323
1324 if (count > 4) {
1325 assert(count == 5);
1326 nir_constant *constant =
1327 vtn_value(b, w[4], vtn_value_type_constant)->constant;
1328 var->var->constant_initializer =
1329 nir_constant_clone(constant, var->var);
1330 }
1331
1332 vtn_foreach_decoration(b, val, var_decoration_cb, var);
1333
1334 if (var->mode == vtn_variable_mode_image ||
1335 var->mode == vtn_variable_mode_sampler) {
1336 /* XXX: We still need the binding information in the nir_variable
1337 * for these. We should fix that.
1338 */
1339 var->var->data.binding = var->binding;
1340 var->var->data.descriptor_set = var->descriptor_set;
1341 var->var->data.index = var->input_attachment_index;
1342
1343 if (var->mode == vtn_variable_mode_image)
1344 var->var->data.image.format = without_array->image_format;
1345 }
1346
1347 if (var->mode == vtn_variable_mode_local) {
1348 assert(var->members == NULL && var->var != NULL);
1349 nir_function_impl_add_variable(b->impl, var->var);
1350 } else if (var->var) {
1351 nir_shader_add_variable(b->shader, var->var);
1352 } else if (var->members) {
1353 unsigned count = glsl_get_length(without_array->type);
1354 for (unsigned i = 0; i < count; i++) {
1355 assert(var->members[i]->data.mode != nir_var_local);
1356 nir_shader_add_variable(b->shader, var->members[i]);
1357 }
1358 } else {
1359 assert(var->mode == vtn_variable_mode_ubo ||
1360 var->mode == vtn_variable_mode_ssbo ||
1361 var->mode == vtn_variable_mode_push_constant);
1362 }
1363 break;
1364 }
1365
1366 case SpvOpAccessChain:
1367 case SpvOpInBoundsAccessChain: {
1368 struct vtn_access_chain *base, *chain;
1369 struct vtn_value *base_val = vtn_untyped_value(b, w[3]);
1370 if (base_val->value_type == vtn_value_type_sampled_image) {
1371 /* This is rather insane. SPIR-V allows you to use OpSampledImage
1372 * to combine an array of images with a single sampler to get an
1373 * array of sampled images that all share the same sampler.
1374 * Fortunately, this means that we can more-or-less ignore the
1375 * sampler when crawling the access chain, but it does leave us
1376 * with this rather awkward little special-case.
1377 */
1378 base = base_val->sampled_image->image;
1379 } else {
1380 assert(base_val->value_type == vtn_value_type_access_chain);
1381 base = base_val->access_chain;
1382 }
1383
1384 chain = vtn_access_chain_extend(b, base, count - 4);
1385
1386 unsigned idx = base->length;
1387 for (int i = 4; i < count; i++) {
1388 struct vtn_value *link_val = vtn_untyped_value(b, w[i]);
1389 if (link_val->value_type == vtn_value_type_constant) {
1390 chain->link[idx].mode = vtn_access_mode_literal;
1391 chain->link[idx].id = link_val->constant->values[0].u32[0];
1392 } else {
1393 chain->link[idx].mode = vtn_access_mode_id;
1394 chain->link[idx].id = w[i];
1395 }
1396 idx++;
1397 }
1398
1399 if (base_val->value_type == vtn_value_type_sampled_image) {
1400 struct vtn_value *val =
1401 vtn_push_value(b, w[2], vtn_value_type_sampled_image);
1402 val->sampled_image = ralloc(b, struct vtn_sampled_image);
1403 val->sampled_image->image = chain;
1404 val->sampled_image->sampler = base_val->sampled_image->sampler;
1405 } else {
1406 struct vtn_value *val =
1407 vtn_push_value(b, w[2], vtn_value_type_access_chain);
1408 val->access_chain = chain;
1409 }
1410 break;
1411 }
1412
1413 case SpvOpCopyMemory: {
1414 struct vtn_value *dest = vtn_value(b, w[1], vtn_value_type_access_chain);
1415 struct vtn_value *src = vtn_value(b, w[2], vtn_value_type_access_chain);
1416
1417 vtn_variable_copy(b, dest->access_chain, src->access_chain);
1418 break;
1419 }
1420
1421 case SpvOpLoad: {
1422 struct vtn_access_chain *src =
1423 vtn_value(b, w[3], vtn_value_type_access_chain)->access_chain;
1424
1425 if (src->var->mode == vtn_variable_mode_image ||
1426 src->var->mode == vtn_variable_mode_sampler) {
1427 vtn_push_value(b, w[2], vtn_value_type_access_chain)->access_chain = src;
1428 return;
1429 }
1430
1431 struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_ssa);
1432 val->ssa = vtn_variable_load(b, src);
1433 break;
1434 }
1435
1436 case SpvOpStore: {
1437 struct vtn_access_chain *dest =
1438 vtn_value(b, w[1], vtn_value_type_access_chain)->access_chain;
1439 struct vtn_ssa_value *src = vtn_ssa_value(b, w[2]);
1440 vtn_variable_store(b, src, dest);
1441 break;
1442 }
1443
1444 case SpvOpArrayLength: {
1445 struct vtn_access_chain *chain =
1446 vtn_value(b, w[3], vtn_value_type_access_chain)->access_chain;
1447
1448 const uint32_t offset = chain->var->type->offsets[w[4]];
1449 const uint32_t stride = chain->var->type->members[w[4]]->stride;
1450
1451 unsigned chain_idx;
1452 struct vtn_type *type;
1453 nir_ssa_def *index =
1454 get_vulkan_resource_index(b, chain, &type, &chain_idx);
1455
1456 nir_intrinsic_instr *instr =
1457 nir_intrinsic_instr_create(b->nb.shader,
1458 nir_intrinsic_get_buffer_size);
1459 instr->src[0] = nir_src_for_ssa(index);
1460 nir_ssa_dest_init(&instr->instr, &instr->dest, 1, 32, NULL);
1461 nir_builder_instr_insert(&b->nb, &instr->instr);
1462 nir_ssa_def *buf_size = &instr->dest.ssa;
1463
1464 /* array_length = max(buffer_size - offset, 0) / stride */
1465 nir_ssa_def *array_length =
1466 nir_idiv(&b->nb,
1467 nir_imax(&b->nb,
1468 nir_isub(&b->nb,
1469 buf_size,
1470 nir_imm_int(&b->nb, offset)),
1471 nir_imm_int(&b->nb, 0u)),
1472 nir_imm_int(&b->nb, stride));
1473
1474 struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_ssa);
1475 val->ssa = vtn_create_ssa_value(b, glsl_uint_type());
1476 val->ssa->def = array_length;
1477 break;
1478 }
1479
1480 case SpvOpCopyMemorySized:
1481 default:
1482 unreachable("Unhandled opcode");
1483 }
1484 }