spirv: Calculate properly 16-bit vector sizes
[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_create(struct vtn_builder *b, unsigned length)
33 {
34 struct vtn_access_chain *chain;
35
36 /* Subtract 1 from the length since there's already one built in */
37 size_t size = sizeof(*chain) +
38 (MAX2(length, 1) - 1) * sizeof(chain->link[0]);
39 chain = rzalloc_size(b, size);
40 chain->length = length;
41
42 return chain;
43 }
44
45 static struct vtn_access_chain *
46 vtn_access_chain_extend(struct vtn_builder *b, struct vtn_access_chain *old,
47 unsigned new_ids)
48 {
49 struct vtn_access_chain *chain;
50
51 unsigned old_len = old ? old->length : 0;
52 chain = vtn_access_chain_create(b, old_len + new_ids);
53
54 for (unsigned i = 0; i < old_len; i++)
55 chain->link[i] = old->link[i];
56
57 return chain;
58 }
59
60 static bool
61 vtn_pointer_uses_ssa_offset(struct vtn_builder *b,
62 struct vtn_pointer *ptr)
63 {
64 return ptr->mode == vtn_variable_mode_ubo ||
65 ptr->mode == vtn_variable_mode_ssbo ||
66 (ptr->mode == vtn_variable_mode_workgroup &&
67 b->options->lower_workgroup_access_to_offsets);
68 }
69
70 static bool
71 vtn_pointer_is_external_block(struct vtn_builder *b,
72 struct vtn_pointer *ptr)
73 {
74 return ptr->mode == vtn_variable_mode_ssbo ||
75 ptr->mode == vtn_variable_mode_ubo ||
76 ptr->mode == vtn_variable_mode_push_constant ||
77 (ptr->mode == vtn_variable_mode_workgroup &&
78 b->options->lower_workgroup_access_to_offsets);
79 }
80
81 /* Dereference the given base pointer by the access chain */
82 static struct vtn_pointer *
83 vtn_access_chain_pointer_dereference(struct vtn_builder *b,
84 struct vtn_pointer *base,
85 struct vtn_access_chain *deref_chain)
86 {
87 struct vtn_access_chain *chain =
88 vtn_access_chain_extend(b, base->chain, deref_chain->length);
89 struct vtn_type *type = base->type;
90
91 /* OpPtrAccessChain is only allowed on things which support variable
92 * pointers. For everything else, the client is expected to just pass us
93 * the right access chain.
94 */
95 vtn_assert(!deref_chain->ptr_as_array);
96
97 unsigned start = base->chain ? base->chain->length : 0;
98 for (unsigned i = 0; i < deref_chain->length; i++) {
99 chain->link[start + i] = deref_chain->link[i];
100
101 if (glsl_type_is_struct(type->type)) {
102 vtn_assert(deref_chain->link[i].mode == vtn_access_mode_literal);
103 type = type->members[deref_chain->link[i].id];
104 } else {
105 type = type->array_element;
106 }
107 }
108
109 struct vtn_pointer *ptr = rzalloc(b, struct vtn_pointer);
110 ptr->mode = base->mode;
111 ptr->type = type;
112 ptr->var = base->var;
113 ptr->chain = chain;
114
115 return ptr;
116 }
117
118 static nir_ssa_def *
119 vtn_access_link_as_ssa(struct vtn_builder *b, struct vtn_access_link link,
120 unsigned stride)
121 {
122 vtn_assert(stride > 0);
123 if (link.mode == vtn_access_mode_literal) {
124 return nir_imm_int(&b->nb, link.id * stride);
125 } else if (stride == 1) {
126 nir_ssa_def *ssa = vtn_ssa_value(b, link.id)->def;
127 if (ssa->bit_size != 32)
128 ssa = nir_u2u32(&b->nb, ssa);
129 return ssa;
130 } else {
131 nir_ssa_def *src0 = vtn_ssa_value(b, link.id)->def;
132 if (src0->bit_size != 32)
133 src0 = nir_u2u32(&b->nb, src0);
134 return nir_imul(&b->nb, src0, nir_imm_int(&b->nb, stride));
135 }
136 }
137
138 static nir_ssa_def *
139 vtn_variable_resource_index(struct vtn_builder *b, struct vtn_variable *var,
140 nir_ssa_def *desc_array_index)
141 {
142 if (!desc_array_index) {
143 vtn_assert(glsl_type_is_struct(var->type->type));
144 desc_array_index = nir_imm_int(&b->nb, 0);
145 }
146
147 nir_intrinsic_instr *instr =
148 nir_intrinsic_instr_create(b->nb.shader,
149 nir_intrinsic_vulkan_resource_index);
150 instr->src[0] = nir_src_for_ssa(desc_array_index);
151 nir_intrinsic_set_desc_set(instr, var->descriptor_set);
152 nir_intrinsic_set_binding(instr, var->binding);
153
154 nir_ssa_dest_init(&instr->instr, &instr->dest, 1, 32, NULL);
155 nir_builder_instr_insert(&b->nb, &instr->instr);
156
157 return &instr->dest.ssa;
158 }
159
160 static nir_ssa_def *
161 vtn_resource_reindex(struct vtn_builder *b, nir_ssa_def *base_index,
162 nir_ssa_def *offset_index)
163 {
164 nir_intrinsic_instr *instr =
165 nir_intrinsic_instr_create(b->nb.shader,
166 nir_intrinsic_vulkan_resource_reindex);
167 instr->src[0] = nir_src_for_ssa(base_index);
168 instr->src[1] = nir_src_for_ssa(offset_index);
169
170 nir_ssa_dest_init(&instr->instr, &instr->dest, 1, 32, NULL);
171 nir_builder_instr_insert(&b->nb, &instr->instr);
172
173 return &instr->dest.ssa;
174 }
175
176 static struct vtn_pointer *
177 vtn_ssa_offset_pointer_dereference(struct vtn_builder *b,
178 struct vtn_pointer *base,
179 struct vtn_access_chain *deref_chain)
180 {
181 nir_ssa_def *block_index = base->block_index;
182 nir_ssa_def *offset = base->offset;
183 struct vtn_type *type = base->type;
184
185 unsigned idx = 0;
186 if (base->mode == vtn_variable_mode_ubo ||
187 base->mode == vtn_variable_mode_ssbo) {
188 if (!block_index) {
189 vtn_assert(base->var && base->type);
190 nir_ssa_def *desc_arr_idx;
191 if (glsl_type_is_array(type->type)) {
192 if (deref_chain->length >= 1) {
193 desc_arr_idx =
194 vtn_access_link_as_ssa(b, deref_chain->link[0], 1);
195 idx++;
196 /* This consumes a level of type */
197 type = type->array_element;
198 } else {
199 /* This is annoying. We've been asked for a pointer to the
200 * array of UBOs/SSBOs and not a specifc buffer. Return a
201 * pointer with a descriptor index of 0 and we'll have to do
202 * a reindex later to adjust it to the right thing.
203 */
204 desc_arr_idx = nir_imm_int(&b->nb, 0);
205 }
206 } else if (deref_chain->ptr_as_array) {
207 /* You can't have a zero-length OpPtrAccessChain */
208 vtn_assert(deref_chain->length >= 1);
209 desc_arr_idx = vtn_access_link_as_ssa(b, deref_chain->link[0], 1);
210 } else {
211 /* We have a regular non-array SSBO. */
212 desc_arr_idx = NULL;
213 }
214 block_index = vtn_variable_resource_index(b, base->var, desc_arr_idx);
215 } else if (deref_chain->ptr_as_array &&
216 type->base_type == vtn_base_type_struct && type->block) {
217 /* We are doing an OpPtrAccessChain on a pointer to a struct that is
218 * decorated block. This is an interesting corner in the SPIR-V
219 * spec. One interpretation would be that they client is clearly
220 * trying to treat that block as if it's an implicit array of blocks
221 * repeated in the buffer. However, the SPIR-V spec for the
222 * OpPtrAccessChain says:
223 *
224 * "Base is treated as the address of the first element of an
225 * array, and the Element element’s address is computed to be the
226 * base for the Indexes, as per OpAccessChain."
227 *
228 * Taken literally, that would mean that your struct type is supposed
229 * to be treated as an array of such a struct and, since it's
230 * decorated block, that means an array of blocks which corresponds
231 * to an array descriptor. Therefore, we need to do a reindex
232 * operation to add the index from the first link in the access chain
233 * to the index we recieved.
234 *
235 * The downside to this interpretation (there always is one) is that
236 * this might be somewhat surprising behavior to apps if they expect
237 * the implicit array behavior described above.
238 */
239 vtn_assert(deref_chain->length >= 1);
240 nir_ssa_def *offset_index =
241 vtn_access_link_as_ssa(b, deref_chain->link[0], 1);
242 idx++;
243
244 block_index = vtn_resource_reindex(b, block_index, offset_index);
245 }
246 }
247
248 if (!offset) {
249 if (base->mode == vtn_variable_mode_workgroup) {
250 /* SLM doesn't need nor have a block index */
251 vtn_assert(!block_index);
252
253 /* We need the variable for the base offset */
254 vtn_assert(base->var);
255
256 /* We need ptr_type for size and alignment */
257 vtn_assert(base->ptr_type);
258
259 /* Assign location on first use so that we don't end up bloating SLM
260 * address space for variables which are never statically used.
261 */
262 if (base->var->shared_location < 0) {
263 vtn_assert(base->ptr_type->length > 0 && base->ptr_type->align > 0);
264 b->shader->num_shared = vtn_align_u32(b->shader->num_shared,
265 base->ptr_type->align);
266 base->var->shared_location = b->shader->num_shared;
267 b->shader->num_shared += base->ptr_type->length;
268 }
269
270 offset = nir_imm_int(&b->nb, base->var->shared_location);
271 } else {
272 /* The code above should have ensured a block_index when needed. */
273 vtn_assert(block_index);
274
275 /* Start off with at the start of the buffer. */
276 offset = nir_imm_int(&b->nb, 0);
277 }
278 }
279
280 if (deref_chain->ptr_as_array && idx == 0) {
281 /* We need ptr_type for the stride */
282 vtn_assert(base->ptr_type);
283
284 /* We need at least one element in the chain */
285 vtn_assert(deref_chain->length >= 1);
286
287 nir_ssa_def *elem_offset =
288 vtn_access_link_as_ssa(b, deref_chain->link[idx],
289 base->ptr_type->stride);
290 offset = nir_iadd(&b->nb, offset, elem_offset);
291 idx++;
292 }
293
294 for (; idx < deref_chain->length; idx++) {
295 switch (glsl_get_base_type(type->type)) {
296 case GLSL_TYPE_UINT:
297 case GLSL_TYPE_INT:
298 case GLSL_TYPE_UINT16:
299 case GLSL_TYPE_INT16:
300 case GLSL_TYPE_UINT64:
301 case GLSL_TYPE_INT64:
302 case GLSL_TYPE_FLOAT:
303 case GLSL_TYPE_FLOAT16:
304 case GLSL_TYPE_DOUBLE:
305 case GLSL_TYPE_BOOL:
306 case GLSL_TYPE_ARRAY: {
307 nir_ssa_def *elem_offset =
308 vtn_access_link_as_ssa(b, deref_chain->link[idx], type->stride);
309 offset = nir_iadd(&b->nb, offset, elem_offset);
310 type = type->array_element;
311 break;
312 }
313
314 case GLSL_TYPE_STRUCT: {
315 vtn_assert(deref_chain->link[idx].mode == vtn_access_mode_literal);
316 unsigned member = deref_chain->link[idx].id;
317 nir_ssa_def *mem_offset = nir_imm_int(&b->nb, type->offsets[member]);
318 offset = nir_iadd(&b->nb, offset, mem_offset);
319 type = type->members[member];
320 break;
321 }
322
323 default:
324 vtn_fail("Invalid type for deref");
325 }
326 }
327
328 struct vtn_pointer *ptr = rzalloc(b, struct vtn_pointer);
329 ptr->mode = base->mode;
330 ptr->type = type;
331 ptr->block_index = block_index;
332 ptr->offset = offset;
333
334 return ptr;
335 }
336
337 /* Dereference the given base pointer by the access chain */
338 static struct vtn_pointer *
339 vtn_pointer_dereference(struct vtn_builder *b,
340 struct vtn_pointer *base,
341 struct vtn_access_chain *deref_chain)
342 {
343 if (vtn_pointer_uses_ssa_offset(b, base)) {
344 return vtn_ssa_offset_pointer_dereference(b, base, deref_chain);
345 } else {
346 return vtn_access_chain_pointer_dereference(b, base, deref_chain);
347 }
348 }
349
350 /* Crawls a chain of array derefs and rewrites the types so that the
351 * lengths stay the same but the terminal type is the one given by
352 * tail_type. This is useful for split structures.
353 */
354 static void
355 rewrite_deref_types(struct vtn_builder *b, nir_deref *deref,
356 const struct glsl_type *type)
357 {
358 deref->type = type;
359 if (deref->child) {
360 vtn_assert(deref->child->deref_type == nir_deref_type_array);
361 vtn_assert(glsl_type_is_array(deref->type));
362 rewrite_deref_types(b, deref->child, glsl_get_array_element(type));
363 }
364 }
365
366 struct vtn_pointer *
367 vtn_pointer_for_variable(struct vtn_builder *b,
368 struct vtn_variable *var, struct vtn_type *ptr_type)
369 {
370 struct vtn_pointer *pointer = rzalloc(b, struct vtn_pointer);
371
372 pointer->mode = var->mode;
373 pointer->type = var->type;
374 vtn_assert(ptr_type->base_type == vtn_base_type_pointer);
375 vtn_assert(ptr_type->deref->type == var->type->type);
376 pointer->ptr_type = ptr_type;
377 pointer->var = var;
378
379 return pointer;
380 }
381
382 nir_deref_var *
383 vtn_pointer_to_deref(struct vtn_builder *b, struct vtn_pointer *ptr)
384 {
385 /* Do on-the-fly copy propagation for samplers. */
386 if (ptr->var->copy_prop_sampler)
387 return vtn_pointer_to_deref(b, ptr->var->copy_prop_sampler);
388
389 nir_deref_var *deref_var;
390 if (ptr->var->var) {
391 deref_var = nir_deref_var_create(b, ptr->var->var);
392 /* Raw variable access */
393 if (!ptr->chain)
394 return deref_var;
395 } else {
396 vtn_assert(ptr->var->members);
397 /* Create the deref_var manually. It will get filled out later. */
398 deref_var = rzalloc(b, nir_deref_var);
399 deref_var->deref.deref_type = nir_deref_type_var;
400 }
401
402 struct vtn_access_chain *chain = ptr->chain;
403 vtn_assert(chain);
404
405 struct vtn_type *deref_type = ptr->var->type;
406 nir_deref *tail = &deref_var->deref;
407 nir_variable **members = ptr->var->members;
408
409 for (unsigned i = 0; i < chain->length; i++) {
410 enum glsl_base_type base_type = glsl_get_base_type(deref_type->type);
411 switch (base_type) {
412 case GLSL_TYPE_UINT:
413 case GLSL_TYPE_INT:
414 case GLSL_TYPE_UINT16:
415 case GLSL_TYPE_INT16:
416 case GLSL_TYPE_UINT64:
417 case GLSL_TYPE_INT64:
418 case GLSL_TYPE_FLOAT:
419 case GLSL_TYPE_FLOAT16:
420 case GLSL_TYPE_DOUBLE:
421 case GLSL_TYPE_BOOL:
422 case GLSL_TYPE_ARRAY: {
423 deref_type = deref_type->array_element;
424
425 nir_deref_array *deref_arr = nir_deref_array_create(b);
426 deref_arr->deref.type = deref_type->type;
427
428 if (chain->link[i].mode == vtn_access_mode_literal) {
429 deref_arr->deref_array_type = nir_deref_array_type_direct;
430 deref_arr->base_offset = chain->link[i].id;
431 } else {
432 vtn_assert(chain->link[i].mode == vtn_access_mode_id);
433 deref_arr->deref_array_type = nir_deref_array_type_indirect;
434 deref_arr->base_offset = 0;
435 deref_arr->indirect =
436 nir_src_for_ssa(vtn_ssa_value(b, chain->link[i].id)->def);
437 }
438 tail->child = &deref_arr->deref;
439 tail = tail->child;
440 break;
441 }
442
443 case GLSL_TYPE_STRUCT: {
444 vtn_assert(chain->link[i].mode == vtn_access_mode_literal);
445 unsigned idx = chain->link[i].id;
446 deref_type = deref_type->members[idx];
447 if (members) {
448 /* This is a pre-split structure. */
449 deref_var->var = members[idx];
450 rewrite_deref_types(b, &deref_var->deref, members[idx]->type);
451 vtn_assert(tail->type == deref_type->type);
452 members = NULL;
453 } else {
454 nir_deref_struct *deref_struct = nir_deref_struct_create(b, idx);
455 deref_struct->deref.type = deref_type->type;
456 tail->child = &deref_struct->deref;
457 tail = tail->child;
458 }
459 break;
460 }
461 default:
462 vtn_fail("Invalid type for deref");
463 }
464 }
465
466 vtn_assert(members == NULL);
467 return deref_var;
468 }
469
470 static void
471 _vtn_local_load_store(struct vtn_builder *b, bool load, nir_deref_var *deref,
472 nir_deref *tail, struct vtn_ssa_value *inout)
473 {
474 /* The deref tail may contain a deref to select a component of a vector (in
475 * other words, it might not be an actual tail) so we have to save it away
476 * here since we overwrite it later.
477 */
478 nir_deref *old_child = tail->child;
479
480 if (glsl_type_is_vector_or_scalar(tail->type)) {
481 /* Terminate the deref chain in case there is one more link to pick
482 * off a component of the vector.
483 */
484 tail->child = NULL;
485
486 nir_intrinsic_op op = load ? nir_intrinsic_load_var :
487 nir_intrinsic_store_var;
488
489 nir_intrinsic_instr *intrin = nir_intrinsic_instr_create(b->shader, op);
490 intrin->variables[0] = nir_deref_var_clone(deref, intrin);
491 intrin->num_components = glsl_get_vector_elements(tail->type);
492
493 if (load) {
494 nir_ssa_dest_init(&intrin->instr, &intrin->dest,
495 intrin->num_components,
496 glsl_get_bit_size(tail->type),
497 NULL);
498 inout->def = &intrin->dest.ssa;
499 } else {
500 nir_intrinsic_set_write_mask(intrin, (1 << intrin->num_components) - 1);
501 intrin->src[0] = nir_src_for_ssa(inout->def);
502 }
503
504 nir_builder_instr_insert(&b->nb, &intrin->instr);
505 } else if (glsl_get_base_type(tail->type) == GLSL_TYPE_ARRAY ||
506 glsl_type_is_matrix(tail->type)) {
507 unsigned elems = glsl_get_length(tail->type);
508 nir_deref_array *deref_arr = nir_deref_array_create(b);
509 deref_arr->deref_array_type = nir_deref_array_type_direct;
510 deref_arr->deref.type = glsl_get_array_element(tail->type);
511 tail->child = &deref_arr->deref;
512 for (unsigned i = 0; i < elems; i++) {
513 deref_arr->base_offset = i;
514 _vtn_local_load_store(b, load, deref, tail->child, inout->elems[i]);
515 }
516 } else {
517 vtn_assert(glsl_get_base_type(tail->type) == GLSL_TYPE_STRUCT);
518 unsigned elems = glsl_get_length(tail->type);
519 nir_deref_struct *deref_struct = nir_deref_struct_create(b, 0);
520 tail->child = &deref_struct->deref;
521 for (unsigned i = 0; i < elems; i++) {
522 deref_struct->index = i;
523 deref_struct->deref.type = glsl_get_struct_field(tail->type, i);
524 _vtn_local_load_store(b, load, deref, tail->child, inout->elems[i]);
525 }
526 }
527
528 tail->child = old_child;
529 }
530
531 nir_deref_var *
532 vtn_nir_deref(struct vtn_builder *b, uint32_t id)
533 {
534 struct vtn_pointer *ptr = vtn_value(b, id, vtn_value_type_pointer)->pointer;
535 return vtn_pointer_to_deref(b, ptr);
536 }
537
538 /*
539 * Gets the NIR-level deref tail, which may have as a child an array deref
540 * selecting which component due to OpAccessChain supporting per-component
541 * indexing in SPIR-V.
542 */
543 static nir_deref *
544 get_deref_tail(nir_deref_var *deref)
545 {
546 nir_deref *cur = &deref->deref;
547 while (!glsl_type_is_vector_or_scalar(cur->type) && cur->child)
548 cur = cur->child;
549
550 return cur;
551 }
552
553 struct vtn_ssa_value *
554 vtn_local_load(struct vtn_builder *b, nir_deref_var *src)
555 {
556 nir_deref *src_tail = get_deref_tail(src);
557 struct vtn_ssa_value *val = vtn_create_ssa_value(b, src_tail->type);
558 _vtn_local_load_store(b, true, src, src_tail, val);
559
560 if (src_tail->child) {
561 nir_deref_array *vec_deref = nir_deref_as_array(src_tail->child);
562 vtn_assert(vec_deref->deref.child == NULL);
563 val->type = vec_deref->deref.type;
564 if (vec_deref->deref_array_type == nir_deref_array_type_direct)
565 val->def = vtn_vector_extract(b, val->def, vec_deref->base_offset);
566 else
567 val->def = vtn_vector_extract_dynamic(b, val->def,
568 vec_deref->indirect.ssa);
569 }
570
571 return val;
572 }
573
574 void
575 vtn_local_store(struct vtn_builder *b, struct vtn_ssa_value *src,
576 nir_deref_var *dest)
577 {
578 nir_deref *dest_tail = get_deref_tail(dest);
579
580 if (dest_tail->child) {
581 struct vtn_ssa_value *val = vtn_create_ssa_value(b, dest_tail->type);
582 _vtn_local_load_store(b, true, dest, dest_tail, val);
583 nir_deref_array *deref = nir_deref_as_array(dest_tail->child);
584 vtn_assert(deref->deref.child == NULL);
585 if (deref->deref_array_type == nir_deref_array_type_direct)
586 val->def = vtn_vector_insert(b, val->def, src->def,
587 deref->base_offset);
588 else
589 val->def = vtn_vector_insert_dynamic(b, val->def, src->def,
590 deref->indirect.ssa);
591 _vtn_local_load_store(b, false, dest, dest_tail, val);
592 } else {
593 _vtn_local_load_store(b, false, dest, dest_tail, src);
594 }
595 }
596
597 nir_ssa_def *
598 vtn_pointer_to_offset(struct vtn_builder *b, struct vtn_pointer *ptr,
599 nir_ssa_def **index_out, unsigned *end_idx_out)
600 {
601 if (vtn_pointer_uses_ssa_offset(b, ptr)) {
602 if (!ptr->offset) {
603 struct vtn_access_chain chain = {
604 .length = 0,
605 };
606 ptr = vtn_ssa_offset_pointer_dereference(b, ptr, &chain);
607 }
608 *index_out = ptr->block_index;
609 return ptr->offset;
610 }
611
612 vtn_assert(ptr->mode == vtn_variable_mode_push_constant);
613 *index_out = NULL;
614
615 unsigned idx = 0;
616 struct vtn_type *type = ptr->var->type;
617 nir_ssa_def *offset = nir_imm_int(&b->nb, 0);
618
619 if (ptr->chain) {
620 for (; idx < ptr->chain->length; idx++) {
621 enum glsl_base_type base_type = glsl_get_base_type(type->type);
622 switch (base_type) {
623 case GLSL_TYPE_UINT:
624 case GLSL_TYPE_INT:
625 case GLSL_TYPE_UINT16:
626 case GLSL_TYPE_INT16:
627 case GLSL_TYPE_UINT64:
628 case GLSL_TYPE_INT64:
629 case GLSL_TYPE_FLOAT:
630 case GLSL_TYPE_FLOAT16:
631 case GLSL_TYPE_DOUBLE:
632 case GLSL_TYPE_BOOL:
633 case GLSL_TYPE_ARRAY:
634 offset = nir_iadd(&b->nb, offset,
635 vtn_access_link_as_ssa(b, ptr->chain->link[idx],
636 type->stride));
637
638 type = type->array_element;
639 break;
640
641 case GLSL_TYPE_STRUCT: {
642 vtn_assert(ptr->chain->link[idx].mode == vtn_access_mode_literal);
643 unsigned member = ptr->chain->link[idx].id;
644 offset = nir_iadd(&b->nb, offset,
645 nir_imm_int(&b->nb, type->offsets[member]));
646 type = type->members[member];
647 break;
648 }
649
650 default:
651 vtn_fail("Invalid type for deref");
652 }
653 }
654 }
655
656 vtn_assert(type == ptr->type);
657 if (end_idx_out)
658 *end_idx_out = idx;
659
660 return offset;
661 }
662
663 /* Tries to compute the size of an interface block based on the strides and
664 * offsets that are provided to us in the SPIR-V source.
665 */
666 static unsigned
667 vtn_type_block_size(struct vtn_builder *b, struct vtn_type *type)
668 {
669 enum glsl_base_type base_type = glsl_get_base_type(type->type);
670 switch (base_type) {
671 case GLSL_TYPE_UINT:
672 case GLSL_TYPE_INT:
673 case GLSL_TYPE_UINT16:
674 case GLSL_TYPE_INT16:
675 case GLSL_TYPE_UINT64:
676 case GLSL_TYPE_INT64:
677 case GLSL_TYPE_FLOAT:
678 case GLSL_TYPE_FLOAT16:
679 case GLSL_TYPE_BOOL:
680 case GLSL_TYPE_DOUBLE: {
681 unsigned cols = type->row_major ? glsl_get_vector_elements(type->type) :
682 glsl_get_matrix_columns(type->type);
683 if (cols > 1) {
684 vtn_assert(type->stride > 0);
685 return type->stride * cols;
686 } else {
687 unsigned type_size = glsl_get_bit_size(type->type) / 8;
688 return glsl_get_vector_elements(type->type) * type_size;
689 }
690 }
691
692 case GLSL_TYPE_STRUCT:
693 case GLSL_TYPE_INTERFACE: {
694 unsigned size = 0;
695 unsigned num_fields = glsl_get_length(type->type);
696 for (unsigned f = 0; f < num_fields; f++) {
697 unsigned field_end = type->offsets[f] +
698 vtn_type_block_size(b, type->members[f]);
699 size = MAX2(size, field_end);
700 }
701 return size;
702 }
703
704 case GLSL_TYPE_ARRAY:
705 vtn_assert(type->stride > 0);
706 vtn_assert(glsl_get_length(type->type) > 0);
707 return type->stride * glsl_get_length(type->type);
708
709 default:
710 vtn_fail("Invalid block type");
711 return 0;
712 }
713 }
714
715 static void
716 vtn_access_chain_get_offset_size(struct vtn_builder *b,
717 struct vtn_access_chain *chain,
718 struct vtn_type *type,
719 unsigned *access_offset,
720 unsigned *access_size)
721 {
722 *access_offset = 0;
723
724 for (unsigned i = 0; i < chain->length; i++) {
725 if (chain->link[i].mode != vtn_access_mode_literal)
726 break;
727
728 if (glsl_type_is_struct(type->type)) {
729 *access_offset += type->offsets[chain->link[i].id];
730 type = type->members[chain->link[i].id];
731 } else {
732 *access_offset += type->stride * chain->link[i].id;
733 type = type->array_element;
734 }
735 }
736
737 *access_size = vtn_type_block_size(b, type);
738 }
739
740 static void
741 _vtn_load_store_tail(struct vtn_builder *b, nir_intrinsic_op op, bool load,
742 nir_ssa_def *index, nir_ssa_def *offset,
743 unsigned access_offset, unsigned access_size,
744 struct vtn_ssa_value **inout, const struct glsl_type *type)
745 {
746 nir_intrinsic_instr *instr = nir_intrinsic_instr_create(b->nb.shader, op);
747 instr->num_components = glsl_get_vector_elements(type);
748
749 int src = 0;
750 if (!load) {
751 nir_intrinsic_set_write_mask(instr, (1 << instr->num_components) - 1);
752 instr->src[src++] = nir_src_for_ssa((*inout)->def);
753 }
754
755 if (op == nir_intrinsic_load_push_constant) {
756 vtn_assert(access_offset % 4 == 0);
757
758 nir_intrinsic_set_base(instr, access_offset);
759 nir_intrinsic_set_range(instr, access_size);
760 }
761
762 if (index)
763 instr->src[src++] = nir_src_for_ssa(index);
764
765 if (op == nir_intrinsic_load_push_constant) {
766 /* We need to subtract the offset from where the intrinsic will load the
767 * data. */
768 instr->src[src++] =
769 nir_src_for_ssa(nir_isub(&b->nb, offset,
770 nir_imm_int(&b->nb, access_offset)));
771 } else {
772 instr->src[src++] = nir_src_for_ssa(offset);
773 }
774
775 if (load) {
776 nir_ssa_dest_init(&instr->instr, &instr->dest,
777 instr->num_components,
778 glsl_get_bit_size(type), NULL);
779 (*inout)->def = &instr->dest.ssa;
780 }
781
782 nir_builder_instr_insert(&b->nb, &instr->instr);
783
784 if (load && glsl_get_base_type(type) == GLSL_TYPE_BOOL)
785 (*inout)->def = nir_ine(&b->nb, (*inout)->def, nir_imm_int(&b->nb, 0));
786 }
787
788 static void
789 _vtn_block_load_store(struct vtn_builder *b, nir_intrinsic_op op, bool load,
790 nir_ssa_def *index, nir_ssa_def *offset,
791 unsigned access_offset, unsigned access_size,
792 struct vtn_access_chain *chain, unsigned chain_idx,
793 struct vtn_type *type, struct vtn_ssa_value **inout)
794 {
795 if (chain && chain_idx >= chain->length)
796 chain = NULL;
797
798 if (load && chain == NULL && *inout == NULL)
799 *inout = vtn_create_ssa_value(b, type->type);
800
801 enum glsl_base_type base_type = glsl_get_base_type(type->type);
802 switch (base_type) {
803 case GLSL_TYPE_UINT:
804 case GLSL_TYPE_INT:
805 case GLSL_TYPE_UINT16:
806 case GLSL_TYPE_INT16:
807 case GLSL_TYPE_UINT64:
808 case GLSL_TYPE_INT64:
809 case GLSL_TYPE_FLOAT:
810 case GLSL_TYPE_FLOAT16:
811 case GLSL_TYPE_DOUBLE:
812 case GLSL_TYPE_BOOL:
813 /* This is where things get interesting. At this point, we've hit
814 * a vector, a scalar, or a matrix.
815 */
816 if (glsl_type_is_matrix(type->type)) {
817 /* Loading the whole matrix */
818 struct vtn_ssa_value *transpose;
819 unsigned num_ops, vec_width, col_stride;
820 if (type->row_major) {
821 num_ops = glsl_get_vector_elements(type->type);
822 vec_width = glsl_get_matrix_columns(type->type);
823 col_stride = type->array_element->stride;
824 if (load) {
825 const struct glsl_type *transpose_type =
826 glsl_matrix_type(base_type, vec_width, num_ops);
827 *inout = vtn_create_ssa_value(b, transpose_type);
828 } else {
829 transpose = vtn_ssa_transpose(b, *inout);
830 inout = &transpose;
831 }
832 } else {
833 num_ops = glsl_get_matrix_columns(type->type);
834 vec_width = glsl_get_vector_elements(type->type);
835 col_stride = type->stride;
836 }
837
838 for (unsigned i = 0; i < num_ops; i++) {
839 nir_ssa_def *elem_offset =
840 nir_iadd(&b->nb, offset, nir_imm_int(&b->nb, i * col_stride));
841 _vtn_load_store_tail(b, op, load, index, elem_offset,
842 access_offset, access_size,
843 &(*inout)->elems[i],
844 glsl_vector_type(base_type, vec_width));
845 }
846
847 if (load && type->row_major)
848 *inout = vtn_ssa_transpose(b, *inout);
849 } else {
850 unsigned elems = glsl_get_vector_elements(type->type);
851 unsigned type_size = glsl_get_bit_size(type->type) / 8;
852 if (elems == 1 || type->stride == type_size) {
853 /* This is a tightly-packed normal scalar or vector load */
854 vtn_assert(glsl_type_is_vector_or_scalar(type->type));
855 _vtn_load_store_tail(b, op, load, index, offset,
856 access_offset, access_size,
857 inout, type->type);
858 } else {
859 /* This is a strided load. We have to load N things separately.
860 * This is the single column of a row-major matrix case.
861 */
862 vtn_assert(type->stride > type_size);
863 vtn_assert(type->stride % type_size == 0);
864
865 nir_ssa_def *per_comp[4];
866 for (unsigned i = 0; i < elems; i++) {
867 nir_ssa_def *elem_offset =
868 nir_iadd(&b->nb, offset,
869 nir_imm_int(&b->nb, i * type->stride));
870 struct vtn_ssa_value *comp, temp_val;
871 if (!load) {
872 temp_val.def = nir_channel(&b->nb, (*inout)->def, i);
873 temp_val.type = glsl_scalar_type(base_type);
874 }
875 comp = &temp_val;
876 _vtn_load_store_tail(b, op, load, index, elem_offset,
877 access_offset, access_size,
878 &comp, glsl_scalar_type(base_type));
879 per_comp[i] = comp->def;
880 }
881
882 if (load) {
883 if (*inout == NULL)
884 *inout = vtn_create_ssa_value(b, type->type);
885 (*inout)->def = nir_vec(&b->nb, per_comp, elems);
886 }
887 }
888 }
889 return;
890
891 case GLSL_TYPE_ARRAY: {
892 unsigned elems = glsl_get_length(type->type);
893 for (unsigned i = 0; i < elems; i++) {
894 nir_ssa_def *elem_off =
895 nir_iadd(&b->nb, offset, nir_imm_int(&b->nb, i * type->stride));
896 _vtn_block_load_store(b, op, load, index, elem_off,
897 access_offset, access_size,
898 NULL, 0,
899 type->array_element, &(*inout)->elems[i]);
900 }
901 return;
902 }
903
904 case GLSL_TYPE_STRUCT: {
905 unsigned elems = glsl_get_length(type->type);
906 for (unsigned i = 0; i < elems; i++) {
907 nir_ssa_def *elem_off =
908 nir_iadd(&b->nb, offset, nir_imm_int(&b->nb, type->offsets[i]));
909 _vtn_block_load_store(b, op, load, index, elem_off,
910 access_offset, access_size,
911 NULL, 0,
912 type->members[i], &(*inout)->elems[i]);
913 }
914 return;
915 }
916
917 default:
918 vtn_fail("Invalid block member type");
919 }
920 }
921
922 static struct vtn_ssa_value *
923 vtn_block_load(struct vtn_builder *b, struct vtn_pointer *src)
924 {
925 nir_intrinsic_op op;
926 unsigned access_offset = 0, access_size = 0;
927 switch (src->mode) {
928 case vtn_variable_mode_ubo:
929 op = nir_intrinsic_load_ubo;
930 break;
931 case vtn_variable_mode_ssbo:
932 op = nir_intrinsic_load_ssbo;
933 break;
934 case vtn_variable_mode_push_constant:
935 op = nir_intrinsic_load_push_constant;
936 vtn_access_chain_get_offset_size(b, src->chain, src->var->type,
937 &access_offset, &access_size);
938 break;
939 case vtn_variable_mode_workgroup:
940 op = nir_intrinsic_load_shared;
941 break;
942 default:
943 vtn_fail("Invalid block variable mode");
944 }
945
946 nir_ssa_def *offset, *index = NULL;
947 unsigned chain_idx;
948 offset = vtn_pointer_to_offset(b, src, &index, &chain_idx);
949
950 struct vtn_ssa_value *value = NULL;
951 _vtn_block_load_store(b, op, true, index, offset,
952 access_offset, access_size,
953 src->chain, chain_idx, src->type, &value);
954 return value;
955 }
956
957 static void
958 vtn_block_store(struct vtn_builder *b, struct vtn_ssa_value *src,
959 struct vtn_pointer *dst)
960 {
961 nir_intrinsic_op op;
962 switch (dst->mode) {
963 case vtn_variable_mode_ssbo:
964 op = nir_intrinsic_store_ssbo;
965 break;
966 case vtn_variable_mode_workgroup:
967 op = nir_intrinsic_store_shared;
968 break;
969 default:
970 vtn_fail("Invalid block variable mode");
971 }
972
973 nir_ssa_def *offset, *index = NULL;
974 unsigned chain_idx;
975 offset = vtn_pointer_to_offset(b, dst, &index, &chain_idx);
976
977 _vtn_block_load_store(b, op, false, index, offset,
978 0, 0, dst->chain, chain_idx, dst->type, &src);
979 }
980
981 static void
982 _vtn_variable_load_store(struct vtn_builder *b, bool load,
983 struct vtn_pointer *ptr,
984 struct vtn_ssa_value **inout)
985 {
986 enum glsl_base_type base_type = glsl_get_base_type(ptr->type->type);
987 switch (base_type) {
988 case GLSL_TYPE_UINT:
989 case GLSL_TYPE_INT:
990 case GLSL_TYPE_UINT16:
991 case GLSL_TYPE_INT16:
992 case GLSL_TYPE_UINT64:
993 case GLSL_TYPE_INT64:
994 case GLSL_TYPE_FLOAT:
995 case GLSL_TYPE_FLOAT16:
996 case GLSL_TYPE_BOOL:
997 case GLSL_TYPE_DOUBLE:
998 /* At this point, we have a scalar, vector, or matrix so we know that
999 * there cannot be any structure splitting still in the way. By
1000 * stopping at the matrix level rather than the vector level, we
1001 * ensure that matrices get loaded in the optimal way even if they
1002 * are storred row-major in a UBO.
1003 */
1004 if (load) {
1005 *inout = vtn_local_load(b, vtn_pointer_to_deref(b, ptr));
1006 } else {
1007 vtn_local_store(b, *inout, vtn_pointer_to_deref(b, ptr));
1008 }
1009 return;
1010
1011 case GLSL_TYPE_ARRAY:
1012 case GLSL_TYPE_STRUCT: {
1013 unsigned elems = glsl_get_length(ptr->type->type);
1014 if (load) {
1015 vtn_assert(*inout == NULL);
1016 *inout = rzalloc(b, struct vtn_ssa_value);
1017 (*inout)->type = ptr->type->type;
1018 (*inout)->elems = rzalloc_array(b, struct vtn_ssa_value *, elems);
1019 }
1020
1021 struct vtn_access_chain chain = {
1022 .length = 1,
1023 .link = {
1024 { .mode = vtn_access_mode_literal, },
1025 }
1026 };
1027 for (unsigned i = 0; i < elems; i++) {
1028 chain.link[0].id = i;
1029 struct vtn_pointer *elem = vtn_pointer_dereference(b, ptr, &chain);
1030 _vtn_variable_load_store(b, load, elem, &(*inout)->elems[i]);
1031 }
1032 return;
1033 }
1034
1035 default:
1036 vtn_fail("Invalid access chain type");
1037 }
1038 }
1039
1040 struct vtn_ssa_value *
1041 vtn_variable_load(struct vtn_builder *b, struct vtn_pointer *src)
1042 {
1043 if (vtn_pointer_is_external_block(b, src)) {
1044 return vtn_block_load(b, src);
1045 } else {
1046 struct vtn_ssa_value *val = NULL;
1047 _vtn_variable_load_store(b, true, src, &val);
1048 return val;
1049 }
1050 }
1051
1052 void
1053 vtn_variable_store(struct vtn_builder *b, struct vtn_ssa_value *src,
1054 struct vtn_pointer *dest)
1055 {
1056 if (vtn_pointer_is_external_block(b, dest)) {
1057 vtn_assert(dest->mode == vtn_variable_mode_ssbo ||
1058 dest->mode == vtn_variable_mode_workgroup);
1059 vtn_block_store(b, src, dest);
1060 } else {
1061 _vtn_variable_load_store(b, false, dest, &src);
1062 }
1063 }
1064
1065 static void
1066 _vtn_variable_copy(struct vtn_builder *b, struct vtn_pointer *dest,
1067 struct vtn_pointer *src)
1068 {
1069 vtn_assert(src->type->type == dest->type->type);
1070 enum glsl_base_type base_type = glsl_get_base_type(src->type->type);
1071 switch (base_type) {
1072 case GLSL_TYPE_UINT:
1073 case GLSL_TYPE_INT:
1074 case GLSL_TYPE_UINT16:
1075 case GLSL_TYPE_INT16:
1076 case GLSL_TYPE_UINT64:
1077 case GLSL_TYPE_INT64:
1078 case GLSL_TYPE_FLOAT:
1079 case GLSL_TYPE_FLOAT16:
1080 case GLSL_TYPE_DOUBLE:
1081 case GLSL_TYPE_BOOL:
1082 /* At this point, we have a scalar, vector, or matrix so we know that
1083 * there cannot be any structure splitting still in the way. By
1084 * stopping at the matrix level rather than the vector level, we
1085 * ensure that matrices get loaded in the optimal way even if they
1086 * are storred row-major in a UBO.
1087 */
1088 vtn_variable_store(b, vtn_variable_load(b, src), dest);
1089 return;
1090
1091 case GLSL_TYPE_ARRAY:
1092 case GLSL_TYPE_STRUCT: {
1093 struct vtn_access_chain chain = {
1094 .length = 1,
1095 .link = {
1096 { .mode = vtn_access_mode_literal, },
1097 }
1098 };
1099 unsigned elems = glsl_get_length(src->type->type);
1100 for (unsigned i = 0; i < elems; i++) {
1101 chain.link[0].id = i;
1102 struct vtn_pointer *src_elem =
1103 vtn_pointer_dereference(b, src, &chain);
1104 struct vtn_pointer *dest_elem =
1105 vtn_pointer_dereference(b, dest, &chain);
1106
1107 _vtn_variable_copy(b, dest_elem, src_elem);
1108 }
1109 return;
1110 }
1111
1112 default:
1113 vtn_fail("Invalid access chain type");
1114 }
1115 }
1116
1117 static void
1118 vtn_variable_copy(struct vtn_builder *b, struct vtn_pointer *dest,
1119 struct vtn_pointer *src)
1120 {
1121 /* TODO: At some point, we should add a special-case for when we can
1122 * just emit a copy_var intrinsic.
1123 */
1124 _vtn_variable_copy(b, dest, src);
1125 }
1126
1127 static void
1128 set_mode_system_value(struct vtn_builder *b, nir_variable_mode *mode)
1129 {
1130 vtn_assert(*mode == nir_var_system_value || *mode == nir_var_shader_in);
1131 *mode = nir_var_system_value;
1132 }
1133
1134 static void
1135 vtn_get_builtin_location(struct vtn_builder *b,
1136 SpvBuiltIn builtin, int *location,
1137 nir_variable_mode *mode)
1138 {
1139 switch (builtin) {
1140 case SpvBuiltInPosition:
1141 *location = VARYING_SLOT_POS;
1142 break;
1143 case SpvBuiltInPointSize:
1144 *location = VARYING_SLOT_PSIZ;
1145 break;
1146 case SpvBuiltInClipDistance:
1147 *location = VARYING_SLOT_CLIP_DIST0; /* XXX CLIP_DIST1? */
1148 break;
1149 case SpvBuiltInCullDistance:
1150 *location = VARYING_SLOT_CULL_DIST0;
1151 break;
1152 case SpvBuiltInVertexIndex:
1153 *location = SYSTEM_VALUE_VERTEX_ID;
1154 set_mode_system_value(b, mode);
1155 break;
1156 case SpvBuiltInVertexId:
1157 /* Vulkan defines VertexID to be zero-based and reserves the new
1158 * builtin keyword VertexIndex to indicate the non-zero-based value.
1159 */
1160 *location = SYSTEM_VALUE_VERTEX_ID_ZERO_BASE;
1161 set_mode_system_value(b, mode);
1162 break;
1163 case SpvBuiltInInstanceIndex:
1164 *location = SYSTEM_VALUE_INSTANCE_INDEX;
1165 set_mode_system_value(b, mode);
1166 break;
1167 case SpvBuiltInInstanceId:
1168 *location = SYSTEM_VALUE_INSTANCE_ID;
1169 set_mode_system_value(b, mode);
1170 break;
1171 case SpvBuiltInPrimitiveId:
1172 if (b->shader->info.stage == MESA_SHADER_FRAGMENT) {
1173 vtn_assert(*mode == nir_var_shader_in);
1174 *location = VARYING_SLOT_PRIMITIVE_ID;
1175 } else if (*mode == nir_var_shader_out) {
1176 *location = VARYING_SLOT_PRIMITIVE_ID;
1177 } else {
1178 *location = SYSTEM_VALUE_PRIMITIVE_ID;
1179 set_mode_system_value(b, mode);
1180 }
1181 break;
1182 case SpvBuiltInInvocationId:
1183 *location = SYSTEM_VALUE_INVOCATION_ID;
1184 set_mode_system_value(b, mode);
1185 break;
1186 case SpvBuiltInLayer:
1187 *location = VARYING_SLOT_LAYER;
1188 if (b->shader->info.stage == MESA_SHADER_FRAGMENT)
1189 *mode = nir_var_shader_in;
1190 else if (b->shader->info.stage == MESA_SHADER_GEOMETRY)
1191 *mode = nir_var_shader_out;
1192 else
1193 vtn_fail("invalid stage for SpvBuiltInLayer");
1194 break;
1195 case SpvBuiltInViewportIndex:
1196 *location = VARYING_SLOT_VIEWPORT;
1197 if (b->shader->info.stage == MESA_SHADER_GEOMETRY)
1198 *mode = nir_var_shader_out;
1199 else if (b->shader->info.stage == MESA_SHADER_FRAGMENT)
1200 *mode = nir_var_shader_in;
1201 else
1202 vtn_fail("invalid stage for SpvBuiltInViewportIndex");
1203 break;
1204 case SpvBuiltInTessLevelOuter:
1205 *location = VARYING_SLOT_TESS_LEVEL_OUTER;
1206 break;
1207 case SpvBuiltInTessLevelInner:
1208 *location = VARYING_SLOT_TESS_LEVEL_INNER;
1209 break;
1210 case SpvBuiltInTessCoord:
1211 *location = SYSTEM_VALUE_TESS_COORD;
1212 set_mode_system_value(b, mode);
1213 break;
1214 case SpvBuiltInPatchVertices:
1215 *location = SYSTEM_VALUE_VERTICES_IN;
1216 set_mode_system_value(b, mode);
1217 break;
1218 case SpvBuiltInFragCoord:
1219 *location = VARYING_SLOT_POS;
1220 vtn_assert(*mode == nir_var_shader_in);
1221 break;
1222 case SpvBuiltInPointCoord:
1223 *location = VARYING_SLOT_PNTC;
1224 vtn_assert(*mode == nir_var_shader_in);
1225 break;
1226 case SpvBuiltInFrontFacing:
1227 *location = SYSTEM_VALUE_FRONT_FACE;
1228 set_mode_system_value(b, mode);
1229 break;
1230 case SpvBuiltInSampleId:
1231 *location = SYSTEM_VALUE_SAMPLE_ID;
1232 set_mode_system_value(b, mode);
1233 break;
1234 case SpvBuiltInSamplePosition:
1235 *location = SYSTEM_VALUE_SAMPLE_POS;
1236 set_mode_system_value(b, mode);
1237 break;
1238 case SpvBuiltInSampleMask:
1239 if (*mode == nir_var_shader_out) {
1240 *location = FRAG_RESULT_SAMPLE_MASK;
1241 } else {
1242 *location = SYSTEM_VALUE_SAMPLE_MASK_IN;
1243 set_mode_system_value(b, mode);
1244 }
1245 break;
1246 case SpvBuiltInFragDepth:
1247 *location = FRAG_RESULT_DEPTH;
1248 vtn_assert(*mode == nir_var_shader_out);
1249 break;
1250 case SpvBuiltInHelperInvocation:
1251 *location = SYSTEM_VALUE_HELPER_INVOCATION;
1252 set_mode_system_value(b, mode);
1253 break;
1254 case SpvBuiltInNumWorkgroups:
1255 *location = SYSTEM_VALUE_NUM_WORK_GROUPS;
1256 set_mode_system_value(b, mode);
1257 break;
1258 case SpvBuiltInWorkgroupSize:
1259 /* This should already be handled */
1260 vtn_fail("unsupported builtin");
1261 break;
1262 case SpvBuiltInWorkgroupId:
1263 *location = SYSTEM_VALUE_WORK_GROUP_ID;
1264 set_mode_system_value(b, mode);
1265 break;
1266 case SpvBuiltInLocalInvocationId:
1267 *location = SYSTEM_VALUE_LOCAL_INVOCATION_ID;
1268 set_mode_system_value(b, mode);
1269 break;
1270 case SpvBuiltInLocalInvocationIndex:
1271 *location = SYSTEM_VALUE_LOCAL_INVOCATION_INDEX;
1272 set_mode_system_value(b, mode);
1273 break;
1274 case SpvBuiltInGlobalInvocationId:
1275 *location = SYSTEM_VALUE_GLOBAL_INVOCATION_ID;
1276 set_mode_system_value(b, mode);
1277 break;
1278 case SpvBuiltInBaseVertex:
1279 *location = SYSTEM_VALUE_BASE_VERTEX;
1280 set_mode_system_value(b, mode);
1281 break;
1282 case SpvBuiltInBaseInstance:
1283 *location = SYSTEM_VALUE_BASE_INSTANCE;
1284 set_mode_system_value(b, mode);
1285 break;
1286 case SpvBuiltInDrawIndex:
1287 *location = SYSTEM_VALUE_DRAW_ID;
1288 set_mode_system_value(b, mode);
1289 break;
1290 case SpvBuiltInViewIndex:
1291 *location = SYSTEM_VALUE_VIEW_INDEX;
1292 set_mode_system_value(b, mode);
1293 break;
1294 default:
1295 vtn_fail("unsupported builtin");
1296 }
1297 }
1298
1299 static void
1300 apply_var_decoration(struct vtn_builder *b, nir_variable *nir_var,
1301 const struct vtn_decoration *dec)
1302 {
1303 switch (dec->decoration) {
1304 case SpvDecorationRelaxedPrecision:
1305 break; /* FIXME: Do nothing with this for now. */
1306 case SpvDecorationNoPerspective:
1307 nir_var->data.interpolation = INTERP_MODE_NOPERSPECTIVE;
1308 break;
1309 case SpvDecorationFlat:
1310 nir_var->data.interpolation = INTERP_MODE_FLAT;
1311 break;
1312 case SpvDecorationCentroid:
1313 nir_var->data.centroid = true;
1314 break;
1315 case SpvDecorationSample:
1316 nir_var->data.sample = true;
1317 break;
1318 case SpvDecorationInvariant:
1319 nir_var->data.invariant = true;
1320 break;
1321 case SpvDecorationConstant:
1322 vtn_assert(nir_var->constant_initializer != NULL);
1323 nir_var->data.read_only = true;
1324 break;
1325 case SpvDecorationNonReadable:
1326 nir_var->data.image.write_only = true;
1327 break;
1328 case SpvDecorationNonWritable:
1329 nir_var->data.read_only = true;
1330 nir_var->data.image.read_only = true;
1331 break;
1332 case SpvDecorationRestrict:
1333 nir_var->data.image.restrict_flag = true;
1334 break;
1335 case SpvDecorationVolatile:
1336 nir_var->data.image._volatile = true;
1337 break;
1338 case SpvDecorationCoherent:
1339 nir_var->data.image.coherent = true;
1340 break;
1341 case SpvDecorationComponent:
1342 nir_var->data.location_frac = dec->literals[0];
1343 break;
1344 case SpvDecorationIndex:
1345 nir_var->data.index = dec->literals[0];
1346 break;
1347 case SpvDecorationBuiltIn: {
1348 SpvBuiltIn builtin = dec->literals[0];
1349
1350 if (builtin == SpvBuiltInWorkgroupSize) {
1351 /* This shouldn't be a builtin. It's actually a constant. */
1352 nir_var->data.mode = nir_var_global;
1353 nir_var->data.read_only = true;
1354
1355 nir_constant *c = rzalloc(nir_var, nir_constant);
1356 c->values[0].u32[0] = b->shader->info.cs.local_size[0];
1357 c->values[0].u32[1] = b->shader->info.cs.local_size[1];
1358 c->values[0].u32[2] = b->shader->info.cs.local_size[2];
1359 nir_var->constant_initializer = c;
1360 break;
1361 }
1362
1363 nir_variable_mode mode = nir_var->data.mode;
1364 vtn_get_builtin_location(b, builtin, &nir_var->data.location, &mode);
1365 nir_var->data.mode = mode;
1366
1367 switch (builtin) {
1368 case SpvBuiltInTessLevelOuter:
1369 case SpvBuiltInTessLevelInner:
1370 nir_var->data.compact = true;
1371 break;
1372 case SpvBuiltInSamplePosition:
1373 nir_var->data.origin_upper_left = b->origin_upper_left;
1374 /* fallthrough */
1375 case SpvBuiltInFragCoord:
1376 nir_var->data.pixel_center_integer = b->pixel_center_integer;
1377 break;
1378 default:
1379 break;
1380 }
1381 }
1382
1383 case SpvDecorationSpecId:
1384 case SpvDecorationRowMajor:
1385 case SpvDecorationColMajor:
1386 case SpvDecorationMatrixStride:
1387 case SpvDecorationAliased:
1388 case SpvDecorationUniform:
1389 case SpvDecorationStream:
1390 case SpvDecorationOffset:
1391 case SpvDecorationLinkageAttributes:
1392 break; /* Do nothing with these here */
1393
1394 case SpvDecorationPatch:
1395 nir_var->data.patch = true;
1396 break;
1397
1398 case SpvDecorationLocation:
1399 vtn_fail("Handled above");
1400
1401 case SpvDecorationBlock:
1402 case SpvDecorationBufferBlock:
1403 case SpvDecorationArrayStride:
1404 case SpvDecorationGLSLShared:
1405 case SpvDecorationGLSLPacked:
1406 break; /* These can apply to a type but we don't care about them */
1407
1408 case SpvDecorationBinding:
1409 case SpvDecorationDescriptorSet:
1410 case SpvDecorationNoContraction:
1411 case SpvDecorationInputAttachmentIndex:
1412 vtn_warn("Decoration not allowed for variable or structure member: %s",
1413 spirv_decoration_to_string(dec->decoration));
1414 break;
1415
1416 case SpvDecorationXfbBuffer:
1417 case SpvDecorationXfbStride:
1418 vtn_warn("Vulkan does not have transform feedback: %s",
1419 spirv_decoration_to_string(dec->decoration));
1420 break;
1421
1422 case SpvDecorationCPacked:
1423 case SpvDecorationSaturatedConversion:
1424 case SpvDecorationFuncParamAttr:
1425 case SpvDecorationFPRoundingMode:
1426 case SpvDecorationFPFastMathMode:
1427 case SpvDecorationAlignment:
1428 vtn_warn("Decoration only allowed for CL-style kernels: %s",
1429 spirv_decoration_to_string(dec->decoration));
1430 break;
1431
1432 default:
1433 vtn_fail("Unhandled decoration");
1434 }
1435 }
1436
1437 static void
1438 var_is_patch_cb(struct vtn_builder *b, struct vtn_value *val, int member,
1439 const struct vtn_decoration *dec, void *out_is_patch)
1440 {
1441 if (dec->decoration == SpvDecorationPatch) {
1442 *((bool *) out_is_patch) = true;
1443 }
1444 }
1445
1446 static void
1447 var_decoration_cb(struct vtn_builder *b, struct vtn_value *val, int member,
1448 const struct vtn_decoration *dec, void *void_var)
1449 {
1450 struct vtn_variable *vtn_var = void_var;
1451
1452 /* Handle decorations that apply to a vtn_variable as a whole */
1453 switch (dec->decoration) {
1454 case SpvDecorationBinding:
1455 vtn_var->binding = dec->literals[0];
1456 return;
1457 case SpvDecorationDescriptorSet:
1458 vtn_var->descriptor_set = dec->literals[0];
1459 return;
1460 case SpvDecorationInputAttachmentIndex:
1461 vtn_var->input_attachment_index = dec->literals[0];
1462 return;
1463 case SpvDecorationPatch:
1464 vtn_var->patch = true;
1465 break;
1466 default:
1467 break;
1468 }
1469
1470 if (val->value_type == vtn_value_type_pointer) {
1471 assert(val->pointer->var == void_var);
1472 assert(val->pointer->chain == NULL);
1473 assert(member == -1);
1474 } else {
1475 assert(val->value_type == vtn_value_type_type);
1476 }
1477
1478 /* Location is odd. If applied to a split structure, we have to walk the
1479 * whole thing and accumulate the location. It's easier to handle as a
1480 * special case.
1481 */
1482 if (dec->decoration == SpvDecorationLocation) {
1483 unsigned location = dec->literals[0];
1484 bool is_vertex_input;
1485 if (b->shader->info.stage == MESA_SHADER_FRAGMENT &&
1486 vtn_var->mode == vtn_variable_mode_output) {
1487 is_vertex_input = false;
1488 location += FRAG_RESULT_DATA0;
1489 } else if (b->shader->info.stage == MESA_SHADER_VERTEX &&
1490 vtn_var->mode == vtn_variable_mode_input) {
1491 is_vertex_input = true;
1492 location += VERT_ATTRIB_GENERIC0;
1493 } else if (vtn_var->mode == vtn_variable_mode_input ||
1494 vtn_var->mode == vtn_variable_mode_output) {
1495 is_vertex_input = false;
1496 location += vtn_var->patch ? VARYING_SLOT_PATCH0 : VARYING_SLOT_VAR0;
1497 } else {
1498 vtn_warn("Location must be on input or output variable");
1499 return;
1500 }
1501
1502 if (vtn_var->var) {
1503 /* This handles the member and lone variable cases */
1504 vtn_var->var->data.location = location;
1505 } else {
1506 /* This handles the structure member case */
1507 assert(vtn_var->members);
1508 unsigned length =
1509 glsl_get_length(glsl_without_array(vtn_var->type->type));
1510 for (unsigned i = 0; i < length; i++) {
1511 vtn_var->members[i]->data.location = location;
1512 location +=
1513 glsl_count_attribute_slots(vtn_var->members[i]->interface_type,
1514 is_vertex_input);
1515 }
1516 }
1517 return;
1518 } else {
1519 if (vtn_var->var) {
1520 assert(member == -1);
1521 apply_var_decoration(b, vtn_var->var, dec);
1522 } else if (vtn_var->members) {
1523 if (member >= 0) {
1524 /* Member decorations must come from a type */
1525 assert(val->value_type == vtn_value_type_type);
1526 apply_var_decoration(b, vtn_var->members[member], dec);
1527 } else {
1528 unsigned length =
1529 glsl_get_length(glsl_without_array(vtn_var->type->type));
1530 for (unsigned i = 0; i < length; i++)
1531 apply_var_decoration(b, vtn_var->members[i], dec);
1532 }
1533 } else {
1534 /* A few variables, those with external storage, have no actual
1535 * nir_variables associated with them. Fortunately, all decorations
1536 * we care about for those variables are on the type only.
1537 */
1538 vtn_assert(vtn_var->mode == vtn_variable_mode_ubo ||
1539 vtn_var->mode == vtn_variable_mode_ssbo ||
1540 vtn_var->mode == vtn_variable_mode_push_constant ||
1541 (vtn_var->mode == vtn_variable_mode_workgroup &&
1542 b->options->lower_workgroup_access_to_offsets));
1543 }
1544 }
1545 }
1546
1547 static enum vtn_variable_mode
1548 vtn_storage_class_to_mode(struct vtn_builder *b,
1549 SpvStorageClass class,
1550 struct vtn_type *interface_type,
1551 nir_variable_mode *nir_mode_out)
1552 {
1553 enum vtn_variable_mode mode;
1554 nir_variable_mode nir_mode;
1555 switch (class) {
1556 case SpvStorageClassUniform:
1557 if (interface_type->block) {
1558 mode = vtn_variable_mode_ubo;
1559 nir_mode = 0;
1560 } else if (interface_type->buffer_block) {
1561 mode = vtn_variable_mode_ssbo;
1562 nir_mode = 0;
1563 } else {
1564 vtn_fail("Invalid uniform variable type");
1565 }
1566 break;
1567 case SpvStorageClassStorageBuffer:
1568 mode = vtn_variable_mode_ssbo;
1569 nir_mode = 0;
1570 break;
1571 case SpvStorageClassUniformConstant:
1572 if (glsl_type_is_image(interface_type->type)) {
1573 mode = vtn_variable_mode_image;
1574 nir_mode = nir_var_uniform;
1575 } else if (glsl_type_is_sampler(interface_type->type)) {
1576 mode = vtn_variable_mode_sampler;
1577 nir_mode = nir_var_uniform;
1578 } else {
1579 vtn_fail("Invalid uniform constant variable type");
1580 }
1581 break;
1582 case SpvStorageClassPushConstant:
1583 mode = vtn_variable_mode_push_constant;
1584 nir_mode = nir_var_uniform;
1585 break;
1586 case SpvStorageClassInput:
1587 mode = vtn_variable_mode_input;
1588 nir_mode = nir_var_shader_in;
1589 break;
1590 case SpvStorageClassOutput:
1591 mode = vtn_variable_mode_output;
1592 nir_mode = nir_var_shader_out;
1593 break;
1594 case SpvStorageClassPrivate:
1595 mode = vtn_variable_mode_global;
1596 nir_mode = nir_var_global;
1597 break;
1598 case SpvStorageClassFunction:
1599 mode = vtn_variable_mode_local;
1600 nir_mode = nir_var_local;
1601 break;
1602 case SpvStorageClassWorkgroup:
1603 mode = vtn_variable_mode_workgroup;
1604 nir_mode = nir_var_shared;
1605 break;
1606 case SpvStorageClassCrossWorkgroup:
1607 case SpvStorageClassGeneric:
1608 case SpvStorageClassAtomicCounter:
1609 default:
1610 vtn_fail("Unhandled variable storage class");
1611 }
1612
1613 if (nir_mode_out)
1614 *nir_mode_out = nir_mode;
1615
1616 return mode;
1617 }
1618
1619 nir_ssa_def *
1620 vtn_pointer_to_ssa(struct vtn_builder *b, struct vtn_pointer *ptr)
1621 {
1622 /* This pointer needs to have a pointer type with actual storage */
1623 vtn_assert(ptr->ptr_type);
1624 vtn_assert(ptr->ptr_type->type);
1625
1626 if (!ptr->offset) {
1627 /* If we don't have an offset then we must be a pointer to the variable
1628 * itself.
1629 */
1630 vtn_assert(!ptr->offset && !ptr->block_index);
1631
1632 struct vtn_access_chain chain = {
1633 .length = 0,
1634 };
1635 ptr = vtn_ssa_offset_pointer_dereference(b, ptr, &chain);
1636 }
1637
1638 vtn_assert(ptr->offset);
1639 if (ptr->block_index) {
1640 vtn_assert(ptr->mode == vtn_variable_mode_ubo ||
1641 ptr->mode == vtn_variable_mode_ssbo);
1642 return nir_vec2(&b->nb, ptr->block_index, ptr->offset);
1643 } else {
1644 vtn_assert(ptr->mode == vtn_variable_mode_workgroup);
1645 return ptr->offset;
1646 }
1647 }
1648
1649 struct vtn_pointer *
1650 vtn_pointer_from_ssa(struct vtn_builder *b, nir_ssa_def *ssa,
1651 struct vtn_type *ptr_type)
1652 {
1653 vtn_assert(ssa->num_components <= 2 && ssa->bit_size == 32);
1654 vtn_assert(ptr_type->base_type == vtn_base_type_pointer);
1655 vtn_assert(ptr_type->deref->base_type != vtn_base_type_pointer);
1656 /* This pointer type needs to have actual storage */
1657 vtn_assert(ptr_type->type);
1658
1659 struct vtn_pointer *ptr = rzalloc(b, struct vtn_pointer);
1660 ptr->mode = vtn_storage_class_to_mode(b, ptr_type->storage_class,
1661 ptr_type, NULL);
1662 ptr->type = ptr_type->deref;
1663 ptr->ptr_type = ptr_type;
1664
1665 if (ssa->num_components > 1) {
1666 vtn_assert(ssa->num_components == 2);
1667 vtn_assert(ptr->mode == vtn_variable_mode_ubo ||
1668 ptr->mode == vtn_variable_mode_ssbo);
1669 ptr->block_index = nir_channel(&b->nb, ssa, 0);
1670 ptr->offset = nir_channel(&b->nb, ssa, 1);
1671 } else {
1672 vtn_assert(ssa->num_components == 1);
1673 vtn_assert(ptr->mode == vtn_variable_mode_workgroup);
1674 ptr->block_index = NULL;
1675 ptr->offset = ssa;
1676 }
1677
1678 return ptr;
1679 }
1680
1681 static bool
1682 is_per_vertex_inout(const struct vtn_variable *var, gl_shader_stage stage)
1683 {
1684 if (var->patch || !glsl_type_is_array(var->type->type))
1685 return false;
1686
1687 if (var->mode == vtn_variable_mode_input) {
1688 return stage == MESA_SHADER_TESS_CTRL ||
1689 stage == MESA_SHADER_TESS_EVAL ||
1690 stage == MESA_SHADER_GEOMETRY;
1691 }
1692
1693 if (var->mode == vtn_variable_mode_output)
1694 return stage == MESA_SHADER_TESS_CTRL;
1695
1696 return false;
1697 }
1698
1699 static void
1700 vtn_create_variable(struct vtn_builder *b, struct vtn_value *val,
1701 struct vtn_type *ptr_type, SpvStorageClass storage_class,
1702 nir_constant *initializer)
1703 {
1704 vtn_assert(ptr_type->base_type == vtn_base_type_pointer);
1705 struct vtn_type *type = ptr_type->deref;
1706
1707 struct vtn_type *without_array = type;
1708 while(glsl_type_is_array(without_array->type))
1709 without_array = without_array->array_element;
1710
1711 enum vtn_variable_mode mode;
1712 nir_variable_mode nir_mode;
1713 mode = vtn_storage_class_to_mode(b, storage_class, without_array, &nir_mode);
1714
1715 switch (mode) {
1716 case vtn_variable_mode_ubo:
1717 b->shader->info.num_ubos++;
1718 break;
1719 case vtn_variable_mode_ssbo:
1720 b->shader->info.num_ssbos++;
1721 break;
1722 case vtn_variable_mode_image:
1723 b->shader->info.num_images++;
1724 break;
1725 case vtn_variable_mode_sampler:
1726 b->shader->info.num_textures++;
1727 break;
1728 case vtn_variable_mode_push_constant:
1729 b->shader->num_uniforms = vtn_type_block_size(b, type);
1730 break;
1731 default:
1732 /* No tallying is needed */
1733 break;
1734 }
1735
1736 struct vtn_variable *var = rzalloc(b, struct vtn_variable);
1737 var->type = type;
1738 var->mode = mode;
1739
1740 vtn_assert(val->value_type == vtn_value_type_pointer);
1741 val->pointer = vtn_pointer_for_variable(b, var, ptr_type);
1742
1743 switch (var->mode) {
1744 case vtn_variable_mode_local:
1745 case vtn_variable_mode_global:
1746 case vtn_variable_mode_image:
1747 case vtn_variable_mode_sampler:
1748 /* For these, we create the variable normally */
1749 var->var = rzalloc(b->shader, nir_variable);
1750 var->var->name = ralloc_strdup(var->var, val->name);
1751 var->var->type = var->type->type;
1752 var->var->data.mode = nir_mode;
1753
1754 switch (var->mode) {
1755 case vtn_variable_mode_image:
1756 case vtn_variable_mode_sampler:
1757 var->var->interface_type = without_array->type;
1758 break;
1759 default:
1760 var->var->interface_type = NULL;
1761 break;
1762 }
1763 break;
1764
1765 case vtn_variable_mode_workgroup:
1766 if (b->options->lower_workgroup_access_to_offsets) {
1767 var->shared_location = -1;
1768 } else {
1769 /* Create the variable normally */
1770 var->var = rzalloc(b->shader, nir_variable);
1771 var->var->name = ralloc_strdup(var->var, val->name);
1772 var->var->type = var->type->type;
1773 var->var->data.mode = nir_var_shared;
1774 }
1775 break;
1776
1777 case vtn_variable_mode_input:
1778 case vtn_variable_mode_output: {
1779 /* In order to know whether or not we're a per-vertex inout, we need
1780 * the patch qualifier. This means walking the variable decorations
1781 * early before we actually create any variables. Not a big deal.
1782 *
1783 * GLSLang really likes to place decorations in the most interior
1784 * thing it possibly can. In particular, if you have a struct, it
1785 * will place the patch decorations on the struct members. This
1786 * should be handled by the variable splitting below just fine.
1787 *
1788 * If you have an array-of-struct, things get even more weird as it
1789 * will place the patch decorations on the struct even though it's
1790 * inside an array and some of the members being patch and others not
1791 * makes no sense whatsoever. Since the only sensible thing is for
1792 * it to be all or nothing, we'll call it patch if any of the members
1793 * are declared patch.
1794 */
1795 var->patch = false;
1796 vtn_foreach_decoration(b, val, var_is_patch_cb, &var->patch);
1797 if (glsl_type_is_array(var->type->type) &&
1798 glsl_type_is_struct(without_array->type)) {
1799 vtn_foreach_decoration(b, vtn_value(b, without_array->id,
1800 vtn_value_type_type),
1801 var_is_patch_cb, &var->patch);
1802 }
1803
1804 /* For inputs and outputs, we immediately split structures. This
1805 * is for a couple of reasons. For one, builtins may all come in
1806 * a struct and we really want those split out into separate
1807 * variables. For another, interpolation qualifiers can be
1808 * applied to members of the top-level struct ane we need to be
1809 * able to preserve that information.
1810 */
1811
1812 int array_length = -1;
1813 struct vtn_type *interface_type = var->type;
1814 if (is_per_vertex_inout(var, b->shader->info.stage)) {
1815 /* In Geometry shaders (and some tessellation), inputs come
1816 * in per-vertex arrays. However, some builtins come in
1817 * non-per-vertex, hence the need for the is_array check. In
1818 * any case, there are no non-builtin arrays allowed so this
1819 * check should be sufficient.
1820 */
1821 interface_type = var->type->array_element;
1822 array_length = glsl_get_length(var->type->type);
1823 }
1824
1825 if (glsl_type_is_struct(interface_type->type)) {
1826 /* It's a struct. Split it. */
1827 unsigned num_members = glsl_get_length(interface_type->type);
1828 var->members = ralloc_array(b, nir_variable *, num_members);
1829
1830 for (unsigned i = 0; i < num_members; i++) {
1831 const struct glsl_type *mtype = interface_type->members[i]->type;
1832 if (array_length >= 0)
1833 mtype = glsl_array_type(mtype, array_length);
1834
1835 var->members[i] = rzalloc(b->shader, nir_variable);
1836 var->members[i]->name =
1837 ralloc_asprintf(var->members[i], "%s.%d", val->name, i);
1838 var->members[i]->type = mtype;
1839 var->members[i]->interface_type =
1840 interface_type->members[i]->type;
1841 var->members[i]->data.mode = nir_mode;
1842 var->members[i]->data.patch = var->patch;
1843
1844 if (initializer) {
1845 assert(i < initializer->num_elements);
1846 var->members[i]->constant_initializer =
1847 nir_constant_clone(initializer->elements[i], var->members[i]);
1848 }
1849 }
1850
1851 initializer = NULL;
1852 } else {
1853 var->var = rzalloc(b->shader, nir_variable);
1854 var->var->name = ralloc_strdup(var->var, val->name);
1855 var->var->type = var->type->type;
1856 var->var->interface_type = interface_type->type;
1857 var->var->data.mode = nir_mode;
1858 var->var->data.patch = var->patch;
1859 }
1860
1861 /* For inputs and outputs, we need to grab locations and builtin
1862 * information from the interface type.
1863 */
1864 vtn_foreach_decoration(b, vtn_value(b, interface_type->id,
1865 vtn_value_type_type),
1866 var_decoration_cb, var);
1867 break;
1868 }
1869
1870 case vtn_variable_mode_param:
1871 vtn_fail("Not created through OpVariable");
1872
1873 case vtn_variable_mode_ubo:
1874 case vtn_variable_mode_ssbo:
1875 case vtn_variable_mode_push_constant:
1876 /* These don't need actual variables. */
1877 break;
1878 }
1879
1880 if (initializer) {
1881 var->var->constant_initializer =
1882 nir_constant_clone(initializer, var->var);
1883 }
1884
1885 vtn_foreach_decoration(b, val, var_decoration_cb, var);
1886
1887 if (var->mode == vtn_variable_mode_image ||
1888 var->mode == vtn_variable_mode_sampler) {
1889 /* XXX: We still need the binding information in the nir_variable
1890 * for these. We should fix that.
1891 */
1892 var->var->data.binding = var->binding;
1893 var->var->data.descriptor_set = var->descriptor_set;
1894 var->var->data.index = var->input_attachment_index;
1895
1896 if (var->mode == vtn_variable_mode_image)
1897 var->var->data.image.format = without_array->image_format;
1898 }
1899
1900 if (var->mode == vtn_variable_mode_local) {
1901 vtn_assert(var->members == NULL && var->var != NULL);
1902 nir_function_impl_add_variable(b->nb.impl, var->var);
1903 } else if (var->var) {
1904 nir_shader_add_variable(b->shader, var->var);
1905 } else if (var->members) {
1906 unsigned count = glsl_get_length(without_array->type);
1907 for (unsigned i = 0; i < count; i++) {
1908 vtn_assert(var->members[i]->data.mode != nir_var_local);
1909 nir_shader_add_variable(b->shader, var->members[i]);
1910 }
1911 } else {
1912 vtn_assert(vtn_pointer_is_external_block(b, val->pointer));
1913 }
1914 }
1915
1916 static void
1917 vtn_assert_types_equal(struct vtn_builder *b, SpvOp opcode,
1918 struct vtn_type *dst_type,
1919 struct vtn_type *src_type)
1920 {
1921 if (dst_type->id == src_type->id)
1922 return;
1923
1924 if (vtn_types_compatible(b, dst_type, src_type)) {
1925 /* Early versions of GLSLang would re-emit types unnecessarily and you
1926 * would end up with OpLoad, OpStore, or OpCopyMemory opcodes which have
1927 * mismatched source and destination types.
1928 *
1929 * https://github.com/KhronosGroup/glslang/issues/304
1930 * https://github.com/KhronosGroup/glslang/issues/307
1931 * https://bugs.freedesktop.org/show_bug.cgi?id=104338
1932 * https://bugs.freedesktop.org/show_bug.cgi?id=104424
1933 */
1934 vtn_warn("Source and destination types of %s do not have the same "
1935 "ID (but are compatible): %u vs %u",
1936 spirv_op_to_string(opcode), dst_type->id, src_type->id);
1937 return;
1938 }
1939
1940 vtn_fail("Source and destination types of %s do not match: %s vs. %s",
1941 spirv_op_to_string(opcode),
1942 glsl_get_type_name(dst_type->type),
1943 glsl_get_type_name(src_type->type));
1944 }
1945
1946 void
1947 vtn_handle_variables(struct vtn_builder *b, SpvOp opcode,
1948 const uint32_t *w, unsigned count)
1949 {
1950 switch (opcode) {
1951 case SpvOpUndef: {
1952 struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_undef);
1953 val->type = vtn_value(b, w[1], vtn_value_type_type)->type;
1954 break;
1955 }
1956
1957 case SpvOpVariable: {
1958 struct vtn_type *ptr_type = vtn_value(b, w[1], vtn_value_type_type)->type;
1959
1960 struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_pointer);
1961
1962 SpvStorageClass storage_class = w[3];
1963 nir_constant *initializer = NULL;
1964 if (count > 4)
1965 initializer = vtn_value(b, w[4], vtn_value_type_constant)->constant;
1966
1967 vtn_create_variable(b, val, ptr_type, storage_class, initializer);
1968 break;
1969 }
1970
1971 case SpvOpAccessChain:
1972 case SpvOpPtrAccessChain:
1973 case SpvOpInBoundsAccessChain: {
1974 struct vtn_access_chain *chain = vtn_access_chain_create(b, count - 4);
1975 chain->ptr_as_array = (opcode == SpvOpPtrAccessChain);
1976
1977 unsigned idx = 0;
1978 for (int i = 4; i < count; i++) {
1979 struct vtn_value *link_val = vtn_untyped_value(b, w[i]);
1980 if (link_val->value_type == vtn_value_type_constant) {
1981 chain->link[idx].mode = vtn_access_mode_literal;
1982 chain->link[idx].id = link_val->constant->values[0].u32[0];
1983 } else {
1984 chain->link[idx].mode = vtn_access_mode_id;
1985 chain->link[idx].id = w[i];
1986
1987 }
1988 idx++;
1989 }
1990
1991 struct vtn_type *ptr_type = vtn_value(b, w[1], vtn_value_type_type)->type;
1992 struct vtn_value *base_val = vtn_untyped_value(b, w[3]);
1993 if (base_val->value_type == vtn_value_type_sampled_image) {
1994 /* This is rather insane. SPIR-V allows you to use OpSampledImage
1995 * to combine an array of images with a single sampler to get an
1996 * array of sampled images that all share the same sampler.
1997 * Fortunately, this means that we can more-or-less ignore the
1998 * sampler when crawling the access chain, but it does leave us
1999 * with this rather awkward little special-case.
2000 */
2001 struct vtn_value *val =
2002 vtn_push_value(b, w[2], vtn_value_type_sampled_image);
2003 val->sampled_image = ralloc(b, struct vtn_sampled_image);
2004 val->sampled_image->type = base_val->sampled_image->type;
2005 val->sampled_image->image =
2006 vtn_pointer_dereference(b, base_val->sampled_image->image, chain);
2007 val->sampled_image->sampler = base_val->sampled_image->sampler;
2008 } else {
2009 vtn_assert(base_val->value_type == vtn_value_type_pointer);
2010 struct vtn_value *val =
2011 vtn_push_value(b, w[2], vtn_value_type_pointer);
2012 val->pointer = vtn_pointer_dereference(b, base_val->pointer, chain);
2013 val->pointer->ptr_type = ptr_type;
2014 }
2015 break;
2016 }
2017
2018 case SpvOpCopyMemory: {
2019 struct vtn_value *dest = vtn_value(b, w[1], vtn_value_type_pointer);
2020 struct vtn_value *src = vtn_value(b, w[2], vtn_value_type_pointer);
2021
2022 vtn_assert_types_equal(b, opcode, dest->type->deref, src->type->deref);
2023
2024 vtn_variable_copy(b, dest->pointer, src->pointer);
2025 break;
2026 }
2027
2028 case SpvOpLoad: {
2029 struct vtn_type *res_type =
2030 vtn_value(b, w[1], vtn_value_type_type)->type;
2031 struct vtn_value *src_val = vtn_value(b, w[3], vtn_value_type_pointer);
2032 struct vtn_pointer *src = src_val->pointer;
2033
2034 vtn_assert_types_equal(b, opcode, res_type, src_val->type->deref);
2035
2036 if (src->mode == vtn_variable_mode_image ||
2037 src->mode == vtn_variable_mode_sampler) {
2038 vtn_push_value(b, w[2], vtn_value_type_pointer)->pointer = src;
2039 return;
2040 }
2041
2042 vtn_push_ssa(b, w[2], res_type, vtn_variable_load(b, src));
2043 break;
2044 }
2045
2046 case SpvOpStore: {
2047 struct vtn_value *dest_val = vtn_value(b, w[1], vtn_value_type_pointer);
2048 struct vtn_pointer *dest = dest_val->pointer;
2049 struct vtn_value *src_val = vtn_untyped_value(b, w[2]);
2050
2051 /* OpStore requires us to actually have a storage type */
2052 vtn_fail_if(dest->type->type == NULL,
2053 "Invalid destination type for OpStore");
2054
2055 if (glsl_get_base_type(dest->type->type) == GLSL_TYPE_BOOL &&
2056 glsl_get_base_type(src_val->type->type) == GLSL_TYPE_UINT) {
2057 /* Early versions of GLSLang would use uint types for UBOs/SSBOs but
2058 * would then store them to a local variable as bool. Work around
2059 * the issue by doing an implicit conversion.
2060 *
2061 * https://github.com/KhronosGroup/glslang/issues/170
2062 * https://bugs.freedesktop.org/show_bug.cgi?id=104424
2063 */
2064 vtn_warn("OpStore of value of type OpTypeInt to a pointer to type "
2065 "OpTypeBool. Doing an implicit conversion to work around "
2066 "the problem.");
2067 struct vtn_ssa_value *bool_ssa =
2068 vtn_create_ssa_value(b, dest->type->type);
2069 bool_ssa->def = nir_i2b(&b->nb, vtn_ssa_value(b, w[2])->def);
2070 vtn_variable_store(b, bool_ssa, dest);
2071 break;
2072 }
2073
2074 vtn_assert_types_equal(b, opcode, dest_val->type->deref, src_val->type);
2075
2076 if (glsl_type_is_sampler(dest->type->type)) {
2077 vtn_warn("OpStore of a sampler detected. Doing on-the-fly copy "
2078 "propagation to workaround the problem.");
2079 vtn_assert(dest->var->copy_prop_sampler == NULL);
2080 dest->var->copy_prop_sampler =
2081 vtn_value(b, w[2], vtn_value_type_pointer)->pointer;
2082 break;
2083 }
2084
2085 struct vtn_ssa_value *src = vtn_ssa_value(b, w[2]);
2086 vtn_variable_store(b, src, dest);
2087 break;
2088 }
2089
2090 case SpvOpArrayLength: {
2091 struct vtn_pointer *ptr =
2092 vtn_value(b, w[3], vtn_value_type_pointer)->pointer;
2093
2094 const uint32_t offset = ptr->var->type->offsets[w[4]];
2095 const uint32_t stride = ptr->var->type->members[w[4]]->stride;
2096
2097 if (!ptr->block_index) {
2098 struct vtn_access_chain chain = {
2099 .length = 0,
2100 };
2101 ptr = vtn_ssa_offset_pointer_dereference(b, ptr, &chain);
2102 vtn_assert(ptr->block_index);
2103 }
2104
2105 nir_intrinsic_instr *instr =
2106 nir_intrinsic_instr_create(b->nb.shader,
2107 nir_intrinsic_get_buffer_size);
2108 instr->src[0] = nir_src_for_ssa(ptr->block_index);
2109 nir_ssa_dest_init(&instr->instr, &instr->dest, 1, 32, NULL);
2110 nir_builder_instr_insert(&b->nb, &instr->instr);
2111 nir_ssa_def *buf_size = &instr->dest.ssa;
2112
2113 /* array_length = max(buffer_size - offset, 0) / stride */
2114 nir_ssa_def *array_length =
2115 nir_idiv(&b->nb,
2116 nir_imax(&b->nb,
2117 nir_isub(&b->nb,
2118 buf_size,
2119 nir_imm_int(&b->nb, offset)),
2120 nir_imm_int(&b->nb, 0u)),
2121 nir_imm_int(&b->nb, stride));
2122
2123 struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_ssa);
2124 val->ssa = vtn_create_ssa_value(b, glsl_uint_type());
2125 val->ssa->def = array_length;
2126 break;
2127 }
2128
2129 case SpvOpCopyMemorySized:
2130 default:
2131 vtn_fail("Unhandled opcode");
2132 }
2133 }