spirv/i965/anv: Relax push constant offset assertions being 32-bit aligned
[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 nir_intrinsic_set_base(instr, access_offset);
757 nir_intrinsic_set_range(instr, access_size);
758 }
759
760 if (index)
761 instr->src[src++] = nir_src_for_ssa(index);
762
763 if (op == nir_intrinsic_load_push_constant) {
764 /* We need to subtract the offset from where the intrinsic will load the
765 * data. */
766 instr->src[src++] =
767 nir_src_for_ssa(nir_isub(&b->nb, offset,
768 nir_imm_int(&b->nb, access_offset)));
769 } else {
770 instr->src[src++] = nir_src_for_ssa(offset);
771 }
772
773 if (load) {
774 nir_ssa_dest_init(&instr->instr, &instr->dest,
775 instr->num_components,
776 glsl_get_bit_size(type), NULL);
777 (*inout)->def = &instr->dest.ssa;
778 }
779
780 nir_builder_instr_insert(&b->nb, &instr->instr);
781
782 if (load && glsl_get_base_type(type) == GLSL_TYPE_BOOL)
783 (*inout)->def = nir_ine(&b->nb, (*inout)->def, nir_imm_int(&b->nb, 0));
784 }
785
786 static void
787 _vtn_block_load_store(struct vtn_builder *b, nir_intrinsic_op op, bool load,
788 nir_ssa_def *index, nir_ssa_def *offset,
789 unsigned access_offset, unsigned access_size,
790 struct vtn_access_chain *chain, unsigned chain_idx,
791 struct vtn_type *type, struct vtn_ssa_value **inout)
792 {
793 if (chain && chain_idx >= chain->length)
794 chain = NULL;
795
796 if (load && chain == NULL && *inout == NULL)
797 *inout = vtn_create_ssa_value(b, type->type);
798
799 enum glsl_base_type base_type = glsl_get_base_type(type->type);
800 switch (base_type) {
801 case GLSL_TYPE_UINT:
802 case GLSL_TYPE_INT:
803 case GLSL_TYPE_UINT16:
804 case GLSL_TYPE_INT16:
805 case GLSL_TYPE_UINT64:
806 case GLSL_TYPE_INT64:
807 case GLSL_TYPE_FLOAT:
808 case GLSL_TYPE_FLOAT16:
809 case GLSL_TYPE_DOUBLE:
810 case GLSL_TYPE_BOOL:
811 /* This is where things get interesting. At this point, we've hit
812 * a vector, a scalar, or a matrix.
813 */
814 if (glsl_type_is_matrix(type->type)) {
815 /* Loading the whole matrix */
816 struct vtn_ssa_value *transpose;
817 unsigned num_ops, vec_width, col_stride;
818 if (type->row_major) {
819 num_ops = glsl_get_vector_elements(type->type);
820 vec_width = glsl_get_matrix_columns(type->type);
821 col_stride = type->array_element->stride;
822 if (load) {
823 const struct glsl_type *transpose_type =
824 glsl_matrix_type(base_type, vec_width, num_ops);
825 *inout = vtn_create_ssa_value(b, transpose_type);
826 } else {
827 transpose = vtn_ssa_transpose(b, *inout);
828 inout = &transpose;
829 }
830 } else {
831 num_ops = glsl_get_matrix_columns(type->type);
832 vec_width = glsl_get_vector_elements(type->type);
833 col_stride = type->stride;
834 }
835
836 for (unsigned i = 0; i < num_ops; i++) {
837 nir_ssa_def *elem_offset =
838 nir_iadd(&b->nb, offset, nir_imm_int(&b->nb, i * col_stride));
839 _vtn_load_store_tail(b, op, load, index, elem_offset,
840 access_offset, access_size,
841 &(*inout)->elems[i],
842 glsl_vector_type(base_type, vec_width));
843 }
844
845 if (load && type->row_major)
846 *inout = vtn_ssa_transpose(b, *inout);
847 } else {
848 unsigned elems = glsl_get_vector_elements(type->type);
849 unsigned type_size = glsl_get_bit_size(type->type) / 8;
850 if (elems == 1 || type->stride == type_size) {
851 /* This is a tightly-packed normal scalar or vector load */
852 vtn_assert(glsl_type_is_vector_or_scalar(type->type));
853 _vtn_load_store_tail(b, op, load, index, offset,
854 access_offset, access_size,
855 inout, type->type);
856 } else {
857 /* This is a strided load. We have to load N things separately.
858 * This is the single column of a row-major matrix case.
859 */
860 vtn_assert(type->stride > type_size);
861 vtn_assert(type->stride % type_size == 0);
862
863 nir_ssa_def *per_comp[4];
864 for (unsigned i = 0; i < elems; i++) {
865 nir_ssa_def *elem_offset =
866 nir_iadd(&b->nb, offset,
867 nir_imm_int(&b->nb, i * type->stride));
868 struct vtn_ssa_value *comp, temp_val;
869 if (!load) {
870 temp_val.def = nir_channel(&b->nb, (*inout)->def, i);
871 temp_val.type = glsl_scalar_type(base_type);
872 }
873 comp = &temp_val;
874 _vtn_load_store_tail(b, op, load, index, elem_offset,
875 access_offset, access_size,
876 &comp, glsl_scalar_type(base_type));
877 per_comp[i] = comp->def;
878 }
879
880 if (load) {
881 if (*inout == NULL)
882 *inout = vtn_create_ssa_value(b, type->type);
883 (*inout)->def = nir_vec(&b->nb, per_comp, elems);
884 }
885 }
886 }
887 return;
888
889 case GLSL_TYPE_ARRAY: {
890 unsigned elems = glsl_get_length(type->type);
891 for (unsigned i = 0; i < elems; i++) {
892 nir_ssa_def *elem_off =
893 nir_iadd(&b->nb, offset, nir_imm_int(&b->nb, i * type->stride));
894 _vtn_block_load_store(b, op, load, index, elem_off,
895 access_offset, access_size,
896 NULL, 0,
897 type->array_element, &(*inout)->elems[i]);
898 }
899 return;
900 }
901
902 case GLSL_TYPE_STRUCT: {
903 unsigned elems = glsl_get_length(type->type);
904 for (unsigned i = 0; i < elems; i++) {
905 nir_ssa_def *elem_off =
906 nir_iadd(&b->nb, offset, nir_imm_int(&b->nb, type->offsets[i]));
907 _vtn_block_load_store(b, op, load, index, elem_off,
908 access_offset, access_size,
909 NULL, 0,
910 type->members[i], &(*inout)->elems[i]);
911 }
912 return;
913 }
914
915 default:
916 vtn_fail("Invalid block member type");
917 }
918 }
919
920 static struct vtn_ssa_value *
921 vtn_block_load(struct vtn_builder *b, struct vtn_pointer *src)
922 {
923 nir_intrinsic_op op;
924 unsigned access_offset = 0, access_size = 0;
925 switch (src->mode) {
926 case vtn_variable_mode_ubo:
927 op = nir_intrinsic_load_ubo;
928 break;
929 case vtn_variable_mode_ssbo:
930 op = nir_intrinsic_load_ssbo;
931 break;
932 case vtn_variable_mode_push_constant:
933 op = nir_intrinsic_load_push_constant;
934 vtn_access_chain_get_offset_size(b, src->chain, src->var->type,
935 &access_offset, &access_size);
936 break;
937 case vtn_variable_mode_workgroup:
938 op = nir_intrinsic_load_shared;
939 break;
940 default:
941 vtn_fail("Invalid block variable mode");
942 }
943
944 nir_ssa_def *offset, *index = NULL;
945 unsigned chain_idx;
946 offset = vtn_pointer_to_offset(b, src, &index, &chain_idx);
947
948 struct vtn_ssa_value *value = NULL;
949 _vtn_block_load_store(b, op, true, index, offset,
950 access_offset, access_size,
951 src->chain, chain_idx, src->type, &value);
952 return value;
953 }
954
955 static void
956 vtn_block_store(struct vtn_builder *b, struct vtn_ssa_value *src,
957 struct vtn_pointer *dst)
958 {
959 nir_intrinsic_op op;
960 switch (dst->mode) {
961 case vtn_variable_mode_ssbo:
962 op = nir_intrinsic_store_ssbo;
963 break;
964 case vtn_variable_mode_workgroup:
965 op = nir_intrinsic_store_shared;
966 break;
967 default:
968 vtn_fail("Invalid block variable mode");
969 }
970
971 nir_ssa_def *offset, *index = NULL;
972 unsigned chain_idx;
973 offset = vtn_pointer_to_offset(b, dst, &index, &chain_idx);
974
975 _vtn_block_load_store(b, op, false, index, offset,
976 0, 0, dst->chain, chain_idx, dst->type, &src);
977 }
978
979 static void
980 _vtn_variable_load_store(struct vtn_builder *b, bool load,
981 struct vtn_pointer *ptr,
982 struct vtn_ssa_value **inout)
983 {
984 enum glsl_base_type base_type = glsl_get_base_type(ptr->type->type);
985 switch (base_type) {
986 case GLSL_TYPE_UINT:
987 case GLSL_TYPE_INT:
988 case GLSL_TYPE_UINT16:
989 case GLSL_TYPE_INT16:
990 case GLSL_TYPE_UINT64:
991 case GLSL_TYPE_INT64:
992 case GLSL_TYPE_FLOAT:
993 case GLSL_TYPE_FLOAT16:
994 case GLSL_TYPE_BOOL:
995 case GLSL_TYPE_DOUBLE:
996 /* At this point, we have a scalar, vector, or matrix so we know that
997 * there cannot be any structure splitting still in the way. By
998 * stopping at the matrix level rather than the vector level, we
999 * ensure that matrices get loaded in the optimal way even if they
1000 * are storred row-major in a UBO.
1001 */
1002 if (load) {
1003 *inout = vtn_local_load(b, vtn_pointer_to_deref(b, ptr));
1004 } else {
1005 vtn_local_store(b, *inout, vtn_pointer_to_deref(b, ptr));
1006 }
1007 return;
1008
1009 case GLSL_TYPE_ARRAY:
1010 case GLSL_TYPE_STRUCT: {
1011 unsigned elems = glsl_get_length(ptr->type->type);
1012 if (load) {
1013 vtn_assert(*inout == NULL);
1014 *inout = rzalloc(b, struct vtn_ssa_value);
1015 (*inout)->type = ptr->type->type;
1016 (*inout)->elems = rzalloc_array(b, struct vtn_ssa_value *, elems);
1017 }
1018
1019 struct vtn_access_chain chain = {
1020 .length = 1,
1021 .link = {
1022 { .mode = vtn_access_mode_literal, },
1023 }
1024 };
1025 for (unsigned i = 0; i < elems; i++) {
1026 chain.link[0].id = i;
1027 struct vtn_pointer *elem = vtn_pointer_dereference(b, ptr, &chain);
1028 _vtn_variable_load_store(b, load, elem, &(*inout)->elems[i]);
1029 }
1030 return;
1031 }
1032
1033 default:
1034 vtn_fail("Invalid access chain type");
1035 }
1036 }
1037
1038 struct vtn_ssa_value *
1039 vtn_variable_load(struct vtn_builder *b, struct vtn_pointer *src)
1040 {
1041 if (vtn_pointer_is_external_block(b, src)) {
1042 return vtn_block_load(b, src);
1043 } else {
1044 struct vtn_ssa_value *val = NULL;
1045 _vtn_variable_load_store(b, true, src, &val);
1046 return val;
1047 }
1048 }
1049
1050 void
1051 vtn_variable_store(struct vtn_builder *b, struct vtn_ssa_value *src,
1052 struct vtn_pointer *dest)
1053 {
1054 if (vtn_pointer_is_external_block(b, dest)) {
1055 vtn_assert(dest->mode == vtn_variable_mode_ssbo ||
1056 dest->mode == vtn_variable_mode_workgroup);
1057 vtn_block_store(b, src, dest);
1058 } else {
1059 _vtn_variable_load_store(b, false, dest, &src);
1060 }
1061 }
1062
1063 static void
1064 _vtn_variable_copy(struct vtn_builder *b, struct vtn_pointer *dest,
1065 struct vtn_pointer *src)
1066 {
1067 vtn_assert(src->type->type == dest->type->type);
1068 enum glsl_base_type base_type = glsl_get_base_type(src->type->type);
1069 switch (base_type) {
1070 case GLSL_TYPE_UINT:
1071 case GLSL_TYPE_INT:
1072 case GLSL_TYPE_UINT16:
1073 case GLSL_TYPE_INT16:
1074 case GLSL_TYPE_UINT64:
1075 case GLSL_TYPE_INT64:
1076 case GLSL_TYPE_FLOAT:
1077 case GLSL_TYPE_FLOAT16:
1078 case GLSL_TYPE_DOUBLE:
1079 case GLSL_TYPE_BOOL:
1080 /* At this point, we have a scalar, vector, or matrix so we know that
1081 * there cannot be any structure splitting still in the way. By
1082 * stopping at the matrix level rather than the vector level, we
1083 * ensure that matrices get loaded in the optimal way even if they
1084 * are storred row-major in a UBO.
1085 */
1086 vtn_variable_store(b, vtn_variable_load(b, src), dest);
1087 return;
1088
1089 case GLSL_TYPE_ARRAY:
1090 case GLSL_TYPE_STRUCT: {
1091 struct vtn_access_chain chain = {
1092 .length = 1,
1093 .link = {
1094 { .mode = vtn_access_mode_literal, },
1095 }
1096 };
1097 unsigned elems = glsl_get_length(src->type->type);
1098 for (unsigned i = 0; i < elems; i++) {
1099 chain.link[0].id = i;
1100 struct vtn_pointer *src_elem =
1101 vtn_pointer_dereference(b, src, &chain);
1102 struct vtn_pointer *dest_elem =
1103 vtn_pointer_dereference(b, dest, &chain);
1104
1105 _vtn_variable_copy(b, dest_elem, src_elem);
1106 }
1107 return;
1108 }
1109
1110 default:
1111 vtn_fail("Invalid access chain type");
1112 }
1113 }
1114
1115 static void
1116 vtn_variable_copy(struct vtn_builder *b, struct vtn_pointer *dest,
1117 struct vtn_pointer *src)
1118 {
1119 /* TODO: At some point, we should add a special-case for when we can
1120 * just emit a copy_var intrinsic.
1121 */
1122 _vtn_variable_copy(b, dest, src);
1123 }
1124
1125 static void
1126 set_mode_system_value(struct vtn_builder *b, nir_variable_mode *mode)
1127 {
1128 vtn_assert(*mode == nir_var_system_value || *mode == nir_var_shader_in);
1129 *mode = nir_var_system_value;
1130 }
1131
1132 static void
1133 vtn_get_builtin_location(struct vtn_builder *b,
1134 SpvBuiltIn builtin, int *location,
1135 nir_variable_mode *mode)
1136 {
1137 switch (builtin) {
1138 case SpvBuiltInPosition:
1139 *location = VARYING_SLOT_POS;
1140 break;
1141 case SpvBuiltInPointSize:
1142 *location = VARYING_SLOT_PSIZ;
1143 break;
1144 case SpvBuiltInClipDistance:
1145 *location = VARYING_SLOT_CLIP_DIST0; /* XXX CLIP_DIST1? */
1146 break;
1147 case SpvBuiltInCullDistance:
1148 *location = VARYING_SLOT_CULL_DIST0;
1149 break;
1150 case SpvBuiltInVertexIndex:
1151 *location = SYSTEM_VALUE_VERTEX_ID;
1152 set_mode_system_value(b, mode);
1153 break;
1154 case SpvBuiltInVertexId:
1155 /* Vulkan defines VertexID to be zero-based and reserves the new
1156 * builtin keyword VertexIndex to indicate the non-zero-based value.
1157 */
1158 *location = SYSTEM_VALUE_VERTEX_ID_ZERO_BASE;
1159 set_mode_system_value(b, mode);
1160 break;
1161 case SpvBuiltInInstanceIndex:
1162 *location = SYSTEM_VALUE_INSTANCE_INDEX;
1163 set_mode_system_value(b, mode);
1164 break;
1165 case SpvBuiltInInstanceId:
1166 *location = SYSTEM_VALUE_INSTANCE_ID;
1167 set_mode_system_value(b, mode);
1168 break;
1169 case SpvBuiltInPrimitiveId:
1170 if (b->shader->info.stage == MESA_SHADER_FRAGMENT) {
1171 vtn_assert(*mode == nir_var_shader_in);
1172 *location = VARYING_SLOT_PRIMITIVE_ID;
1173 } else if (*mode == nir_var_shader_out) {
1174 *location = VARYING_SLOT_PRIMITIVE_ID;
1175 } else {
1176 *location = SYSTEM_VALUE_PRIMITIVE_ID;
1177 set_mode_system_value(b, mode);
1178 }
1179 break;
1180 case SpvBuiltInInvocationId:
1181 *location = SYSTEM_VALUE_INVOCATION_ID;
1182 set_mode_system_value(b, mode);
1183 break;
1184 case SpvBuiltInLayer:
1185 *location = VARYING_SLOT_LAYER;
1186 if (b->shader->info.stage == MESA_SHADER_FRAGMENT)
1187 *mode = nir_var_shader_in;
1188 else if (b->shader->info.stage == MESA_SHADER_GEOMETRY)
1189 *mode = nir_var_shader_out;
1190 else
1191 vtn_fail("invalid stage for SpvBuiltInLayer");
1192 break;
1193 case SpvBuiltInViewportIndex:
1194 *location = VARYING_SLOT_VIEWPORT;
1195 if (b->shader->info.stage == MESA_SHADER_GEOMETRY)
1196 *mode = nir_var_shader_out;
1197 else if (b->shader->info.stage == MESA_SHADER_FRAGMENT)
1198 *mode = nir_var_shader_in;
1199 else
1200 vtn_fail("invalid stage for SpvBuiltInViewportIndex");
1201 break;
1202 case SpvBuiltInTessLevelOuter:
1203 *location = VARYING_SLOT_TESS_LEVEL_OUTER;
1204 break;
1205 case SpvBuiltInTessLevelInner:
1206 *location = VARYING_SLOT_TESS_LEVEL_INNER;
1207 break;
1208 case SpvBuiltInTessCoord:
1209 *location = SYSTEM_VALUE_TESS_COORD;
1210 set_mode_system_value(b, mode);
1211 break;
1212 case SpvBuiltInPatchVertices:
1213 *location = SYSTEM_VALUE_VERTICES_IN;
1214 set_mode_system_value(b, mode);
1215 break;
1216 case SpvBuiltInFragCoord:
1217 *location = VARYING_SLOT_POS;
1218 vtn_assert(*mode == nir_var_shader_in);
1219 break;
1220 case SpvBuiltInPointCoord:
1221 *location = VARYING_SLOT_PNTC;
1222 vtn_assert(*mode == nir_var_shader_in);
1223 break;
1224 case SpvBuiltInFrontFacing:
1225 *location = SYSTEM_VALUE_FRONT_FACE;
1226 set_mode_system_value(b, mode);
1227 break;
1228 case SpvBuiltInSampleId:
1229 *location = SYSTEM_VALUE_SAMPLE_ID;
1230 set_mode_system_value(b, mode);
1231 break;
1232 case SpvBuiltInSamplePosition:
1233 *location = SYSTEM_VALUE_SAMPLE_POS;
1234 set_mode_system_value(b, mode);
1235 break;
1236 case SpvBuiltInSampleMask:
1237 if (*mode == nir_var_shader_out) {
1238 *location = FRAG_RESULT_SAMPLE_MASK;
1239 } else {
1240 *location = SYSTEM_VALUE_SAMPLE_MASK_IN;
1241 set_mode_system_value(b, mode);
1242 }
1243 break;
1244 case SpvBuiltInFragDepth:
1245 *location = FRAG_RESULT_DEPTH;
1246 vtn_assert(*mode == nir_var_shader_out);
1247 break;
1248 case SpvBuiltInHelperInvocation:
1249 *location = SYSTEM_VALUE_HELPER_INVOCATION;
1250 set_mode_system_value(b, mode);
1251 break;
1252 case SpvBuiltInNumWorkgroups:
1253 *location = SYSTEM_VALUE_NUM_WORK_GROUPS;
1254 set_mode_system_value(b, mode);
1255 break;
1256 case SpvBuiltInWorkgroupSize:
1257 /* This should already be handled */
1258 vtn_fail("unsupported builtin");
1259 break;
1260 case SpvBuiltInWorkgroupId:
1261 *location = SYSTEM_VALUE_WORK_GROUP_ID;
1262 set_mode_system_value(b, mode);
1263 break;
1264 case SpvBuiltInLocalInvocationId:
1265 *location = SYSTEM_VALUE_LOCAL_INVOCATION_ID;
1266 set_mode_system_value(b, mode);
1267 break;
1268 case SpvBuiltInLocalInvocationIndex:
1269 *location = SYSTEM_VALUE_LOCAL_INVOCATION_INDEX;
1270 set_mode_system_value(b, mode);
1271 break;
1272 case SpvBuiltInGlobalInvocationId:
1273 *location = SYSTEM_VALUE_GLOBAL_INVOCATION_ID;
1274 set_mode_system_value(b, mode);
1275 break;
1276 case SpvBuiltInBaseVertex:
1277 *location = SYSTEM_VALUE_BASE_VERTEX;
1278 set_mode_system_value(b, mode);
1279 break;
1280 case SpvBuiltInBaseInstance:
1281 *location = SYSTEM_VALUE_BASE_INSTANCE;
1282 set_mode_system_value(b, mode);
1283 break;
1284 case SpvBuiltInDrawIndex:
1285 *location = SYSTEM_VALUE_DRAW_ID;
1286 set_mode_system_value(b, mode);
1287 break;
1288 case SpvBuiltInViewIndex:
1289 *location = SYSTEM_VALUE_VIEW_INDEX;
1290 set_mode_system_value(b, mode);
1291 break;
1292 default:
1293 vtn_fail("unsupported builtin");
1294 }
1295 }
1296
1297 static void
1298 apply_var_decoration(struct vtn_builder *b, nir_variable *nir_var,
1299 const struct vtn_decoration *dec)
1300 {
1301 switch (dec->decoration) {
1302 case SpvDecorationRelaxedPrecision:
1303 break; /* FIXME: Do nothing with this for now. */
1304 case SpvDecorationNoPerspective:
1305 nir_var->data.interpolation = INTERP_MODE_NOPERSPECTIVE;
1306 break;
1307 case SpvDecorationFlat:
1308 nir_var->data.interpolation = INTERP_MODE_FLAT;
1309 break;
1310 case SpvDecorationCentroid:
1311 nir_var->data.centroid = true;
1312 break;
1313 case SpvDecorationSample:
1314 nir_var->data.sample = true;
1315 break;
1316 case SpvDecorationInvariant:
1317 nir_var->data.invariant = true;
1318 break;
1319 case SpvDecorationConstant:
1320 vtn_assert(nir_var->constant_initializer != NULL);
1321 nir_var->data.read_only = true;
1322 break;
1323 case SpvDecorationNonReadable:
1324 nir_var->data.image.write_only = true;
1325 break;
1326 case SpvDecorationNonWritable:
1327 nir_var->data.read_only = true;
1328 nir_var->data.image.read_only = true;
1329 break;
1330 case SpvDecorationRestrict:
1331 nir_var->data.image.restrict_flag = true;
1332 break;
1333 case SpvDecorationVolatile:
1334 nir_var->data.image._volatile = true;
1335 break;
1336 case SpvDecorationCoherent:
1337 nir_var->data.image.coherent = true;
1338 break;
1339 case SpvDecorationComponent:
1340 nir_var->data.location_frac = dec->literals[0];
1341 break;
1342 case SpvDecorationIndex:
1343 nir_var->data.index = dec->literals[0];
1344 break;
1345 case SpvDecorationBuiltIn: {
1346 SpvBuiltIn builtin = dec->literals[0];
1347
1348 if (builtin == SpvBuiltInWorkgroupSize) {
1349 /* This shouldn't be a builtin. It's actually a constant. */
1350 nir_var->data.mode = nir_var_global;
1351 nir_var->data.read_only = true;
1352
1353 nir_constant *c = rzalloc(nir_var, nir_constant);
1354 c->values[0].u32[0] = b->shader->info.cs.local_size[0];
1355 c->values[0].u32[1] = b->shader->info.cs.local_size[1];
1356 c->values[0].u32[2] = b->shader->info.cs.local_size[2];
1357 nir_var->constant_initializer = c;
1358 break;
1359 }
1360
1361 nir_variable_mode mode = nir_var->data.mode;
1362 vtn_get_builtin_location(b, builtin, &nir_var->data.location, &mode);
1363 nir_var->data.mode = mode;
1364
1365 switch (builtin) {
1366 case SpvBuiltInTessLevelOuter:
1367 case SpvBuiltInTessLevelInner:
1368 nir_var->data.compact = true;
1369 break;
1370 case SpvBuiltInSamplePosition:
1371 nir_var->data.origin_upper_left = b->origin_upper_left;
1372 /* fallthrough */
1373 case SpvBuiltInFragCoord:
1374 nir_var->data.pixel_center_integer = b->pixel_center_integer;
1375 break;
1376 default:
1377 break;
1378 }
1379 }
1380
1381 case SpvDecorationSpecId:
1382 case SpvDecorationRowMajor:
1383 case SpvDecorationColMajor:
1384 case SpvDecorationMatrixStride:
1385 case SpvDecorationAliased:
1386 case SpvDecorationUniform:
1387 case SpvDecorationStream:
1388 case SpvDecorationOffset:
1389 case SpvDecorationLinkageAttributes:
1390 break; /* Do nothing with these here */
1391
1392 case SpvDecorationPatch:
1393 nir_var->data.patch = true;
1394 break;
1395
1396 case SpvDecorationLocation:
1397 vtn_fail("Handled above");
1398
1399 case SpvDecorationBlock:
1400 case SpvDecorationBufferBlock:
1401 case SpvDecorationArrayStride:
1402 case SpvDecorationGLSLShared:
1403 case SpvDecorationGLSLPacked:
1404 break; /* These can apply to a type but we don't care about them */
1405
1406 case SpvDecorationBinding:
1407 case SpvDecorationDescriptorSet:
1408 case SpvDecorationNoContraction:
1409 case SpvDecorationInputAttachmentIndex:
1410 vtn_warn("Decoration not allowed for variable or structure member: %s",
1411 spirv_decoration_to_string(dec->decoration));
1412 break;
1413
1414 case SpvDecorationXfbBuffer:
1415 case SpvDecorationXfbStride:
1416 vtn_warn("Vulkan does not have transform feedback: %s",
1417 spirv_decoration_to_string(dec->decoration));
1418 break;
1419
1420 case SpvDecorationCPacked:
1421 case SpvDecorationSaturatedConversion:
1422 case SpvDecorationFuncParamAttr:
1423 case SpvDecorationFPRoundingMode:
1424 case SpvDecorationFPFastMathMode:
1425 case SpvDecorationAlignment:
1426 vtn_warn("Decoration only allowed for CL-style kernels: %s",
1427 spirv_decoration_to_string(dec->decoration));
1428 break;
1429
1430 default:
1431 vtn_fail("Unhandled decoration");
1432 }
1433 }
1434
1435 static void
1436 var_is_patch_cb(struct vtn_builder *b, struct vtn_value *val, int member,
1437 const struct vtn_decoration *dec, void *out_is_patch)
1438 {
1439 if (dec->decoration == SpvDecorationPatch) {
1440 *((bool *) out_is_patch) = true;
1441 }
1442 }
1443
1444 static void
1445 var_decoration_cb(struct vtn_builder *b, struct vtn_value *val, int member,
1446 const struct vtn_decoration *dec, void *void_var)
1447 {
1448 struct vtn_variable *vtn_var = void_var;
1449
1450 /* Handle decorations that apply to a vtn_variable as a whole */
1451 switch (dec->decoration) {
1452 case SpvDecorationBinding:
1453 vtn_var->binding = dec->literals[0];
1454 return;
1455 case SpvDecorationDescriptorSet:
1456 vtn_var->descriptor_set = dec->literals[0];
1457 return;
1458 case SpvDecorationInputAttachmentIndex:
1459 vtn_var->input_attachment_index = dec->literals[0];
1460 return;
1461 case SpvDecorationPatch:
1462 vtn_var->patch = true;
1463 break;
1464 default:
1465 break;
1466 }
1467
1468 if (val->value_type == vtn_value_type_pointer) {
1469 assert(val->pointer->var == void_var);
1470 assert(val->pointer->chain == NULL);
1471 assert(member == -1);
1472 } else {
1473 assert(val->value_type == vtn_value_type_type);
1474 }
1475
1476 /* Location is odd. If applied to a split structure, we have to walk the
1477 * whole thing and accumulate the location. It's easier to handle as a
1478 * special case.
1479 */
1480 if (dec->decoration == SpvDecorationLocation) {
1481 unsigned location = dec->literals[0];
1482 bool is_vertex_input;
1483 if (b->shader->info.stage == MESA_SHADER_FRAGMENT &&
1484 vtn_var->mode == vtn_variable_mode_output) {
1485 is_vertex_input = false;
1486 location += FRAG_RESULT_DATA0;
1487 } else if (b->shader->info.stage == MESA_SHADER_VERTEX &&
1488 vtn_var->mode == vtn_variable_mode_input) {
1489 is_vertex_input = true;
1490 location += VERT_ATTRIB_GENERIC0;
1491 } else if (vtn_var->mode == vtn_variable_mode_input ||
1492 vtn_var->mode == vtn_variable_mode_output) {
1493 is_vertex_input = false;
1494 location += vtn_var->patch ? VARYING_SLOT_PATCH0 : VARYING_SLOT_VAR0;
1495 } else {
1496 vtn_warn("Location must be on input or output variable");
1497 return;
1498 }
1499
1500 if (vtn_var->var) {
1501 /* This handles the member and lone variable cases */
1502 vtn_var->var->data.location = location;
1503 } else {
1504 /* This handles the structure member case */
1505 assert(vtn_var->members);
1506 unsigned length =
1507 glsl_get_length(glsl_without_array(vtn_var->type->type));
1508 for (unsigned i = 0; i < length; i++) {
1509 vtn_var->members[i]->data.location = location;
1510 location +=
1511 glsl_count_attribute_slots(vtn_var->members[i]->interface_type,
1512 is_vertex_input);
1513 }
1514 }
1515 return;
1516 } else {
1517 if (vtn_var->var) {
1518 assert(member == -1);
1519 apply_var_decoration(b, vtn_var->var, dec);
1520 } else if (vtn_var->members) {
1521 if (member >= 0) {
1522 /* Member decorations must come from a type */
1523 assert(val->value_type == vtn_value_type_type);
1524 apply_var_decoration(b, vtn_var->members[member], dec);
1525 } else {
1526 unsigned length =
1527 glsl_get_length(glsl_without_array(vtn_var->type->type));
1528 for (unsigned i = 0; i < length; i++)
1529 apply_var_decoration(b, vtn_var->members[i], dec);
1530 }
1531 } else {
1532 /* A few variables, those with external storage, have no actual
1533 * nir_variables associated with them. Fortunately, all decorations
1534 * we care about for those variables are on the type only.
1535 */
1536 vtn_assert(vtn_var->mode == vtn_variable_mode_ubo ||
1537 vtn_var->mode == vtn_variable_mode_ssbo ||
1538 vtn_var->mode == vtn_variable_mode_push_constant ||
1539 (vtn_var->mode == vtn_variable_mode_workgroup &&
1540 b->options->lower_workgroup_access_to_offsets));
1541 }
1542 }
1543 }
1544
1545 static enum vtn_variable_mode
1546 vtn_storage_class_to_mode(struct vtn_builder *b,
1547 SpvStorageClass class,
1548 struct vtn_type *interface_type,
1549 nir_variable_mode *nir_mode_out)
1550 {
1551 enum vtn_variable_mode mode;
1552 nir_variable_mode nir_mode;
1553 switch (class) {
1554 case SpvStorageClassUniform:
1555 if (interface_type->block) {
1556 mode = vtn_variable_mode_ubo;
1557 nir_mode = 0;
1558 } else if (interface_type->buffer_block) {
1559 mode = vtn_variable_mode_ssbo;
1560 nir_mode = 0;
1561 } else {
1562 vtn_fail("Invalid uniform variable type");
1563 }
1564 break;
1565 case SpvStorageClassStorageBuffer:
1566 mode = vtn_variable_mode_ssbo;
1567 nir_mode = 0;
1568 break;
1569 case SpvStorageClassUniformConstant:
1570 if (glsl_type_is_image(interface_type->type)) {
1571 mode = vtn_variable_mode_image;
1572 nir_mode = nir_var_uniform;
1573 } else if (glsl_type_is_sampler(interface_type->type)) {
1574 mode = vtn_variable_mode_sampler;
1575 nir_mode = nir_var_uniform;
1576 } else {
1577 vtn_fail("Invalid uniform constant variable type");
1578 }
1579 break;
1580 case SpvStorageClassPushConstant:
1581 mode = vtn_variable_mode_push_constant;
1582 nir_mode = nir_var_uniform;
1583 break;
1584 case SpvStorageClassInput:
1585 mode = vtn_variable_mode_input;
1586 nir_mode = nir_var_shader_in;
1587 break;
1588 case SpvStorageClassOutput:
1589 mode = vtn_variable_mode_output;
1590 nir_mode = nir_var_shader_out;
1591 break;
1592 case SpvStorageClassPrivate:
1593 mode = vtn_variable_mode_global;
1594 nir_mode = nir_var_global;
1595 break;
1596 case SpvStorageClassFunction:
1597 mode = vtn_variable_mode_local;
1598 nir_mode = nir_var_local;
1599 break;
1600 case SpvStorageClassWorkgroup:
1601 mode = vtn_variable_mode_workgroup;
1602 nir_mode = nir_var_shared;
1603 break;
1604 case SpvStorageClassCrossWorkgroup:
1605 case SpvStorageClassGeneric:
1606 case SpvStorageClassAtomicCounter:
1607 default:
1608 vtn_fail("Unhandled variable storage class");
1609 }
1610
1611 if (nir_mode_out)
1612 *nir_mode_out = nir_mode;
1613
1614 return mode;
1615 }
1616
1617 nir_ssa_def *
1618 vtn_pointer_to_ssa(struct vtn_builder *b, struct vtn_pointer *ptr)
1619 {
1620 /* This pointer needs to have a pointer type with actual storage */
1621 vtn_assert(ptr->ptr_type);
1622 vtn_assert(ptr->ptr_type->type);
1623
1624 if (!ptr->offset) {
1625 /* If we don't have an offset then we must be a pointer to the variable
1626 * itself.
1627 */
1628 vtn_assert(!ptr->offset && !ptr->block_index);
1629
1630 struct vtn_access_chain chain = {
1631 .length = 0,
1632 };
1633 ptr = vtn_ssa_offset_pointer_dereference(b, ptr, &chain);
1634 }
1635
1636 vtn_assert(ptr->offset);
1637 if (ptr->block_index) {
1638 vtn_assert(ptr->mode == vtn_variable_mode_ubo ||
1639 ptr->mode == vtn_variable_mode_ssbo);
1640 return nir_vec2(&b->nb, ptr->block_index, ptr->offset);
1641 } else {
1642 vtn_assert(ptr->mode == vtn_variable_mode_workgroup);
1643 return ptr->offset;
1644 }
1645 }
1646
1647 struct vtn_pointer *
1648 vtn_pointer_from_ssa(struct vtn_builder *b, nir_ssa_def *ssa,
1649 struct vtn_type *ptr_type)
1650 {
1651 vtn_assert(ssa->num_components <= 2 && ssa->bit_size == 32);
1652 vtn_assert(ptr_type->base_type == vtn_base_type_pointer);
1653 vtn_assert(ptr_type->deref->base_type != vtn_base_type_pointer);
1654 /* This pointer type needs to have actual storage */
1655 vtn_assert(ptr_type->type);
1656
1657 struct vtn_pointer *ptr = rzalloc(b, struct vtn_pointer);
1658 ptr->mode = vtn_storage_class_to_mode(b, ptr_type->storage_class,
1659 ptr_type, NULL);
1660 ptr->type = ptr_type->deref;
1661 ptr->ptr_type = ptr_type;
1662
1663 if (ssa->num_components > 1) {
1664 vtn_assert(ssa->num_components == 2);
1665 vtn_assert(ptr->mode == vtn_variable_mode_ubo ||
1666 ptr->mode == vtn_variable_mode_ssbo);
1667 ptr->block_index = nir_channel(&b->nb, ssa, 0);
1668 ptr->offset = nir_channel(&b->nb, ssa, 1);
1669 } else {
1670 vtn_assert(ssa->num_components == 1);
1671 vtn_assert(ptr->mode == vtn_variable_mode_workgroup);
1672 ptr->block_index = NULL;
1673 ptr->offset = ssa;
1674 }
1675
1676 return ptr;
1677 }
1678
1679 static bool
1680 is_per_vertex_inout(const struct vtn_variable *var, gl_shader_stage stage)
1681 {
1682 if (var->patch || !glsl_type_is_array(var->type->type))
1683 return false;
1684
1685 if (var->mode == vtn_variable_mode_input) {
1686 return stage == MESA_SHADER_TESS_CTRL ||
1687 stage == MESA_SHADER_TESS_EVAL ||
1688 stage == MESA_SHADER_GEOMETRY;
1689 }
1690
1691 if (var->mode == vtn_variable_mode_output)
1692 return stage == MESA_SHADER_TESS_CTRL;
1693
1694 return false;
1695 }
1696
1697 static void
1698 vtn_create_variable(struct vtn_builder *b, struct vtn_value *val,
1699 struct vtn_type *ptr_type, SpvStorageClass storage_class,
1700 nir_constant *initializer)
1701 {
1702 vtn_assert(ptr_type->base_type == vtn_base_type_pointer);
1703 struct vtn_type *type = ptr_type->deref;
1704
1705 struct vtn_type *without_array = type;
1706 while(glsl_type_is_array(without_array->type))
1707 without_array = without_array->array_element;
1708
1709 enum vtn_variable_mode mode;
1710 nir_variable_mode nir_mode;
1711 mode = vtn_storage_class_to_mode(b, storage_class, without_array, &nir_mode);
1712
1713 switch (mode) {
1714 case vtn_variable_mode_ubo:
1715 b->shader->info.num_ubos++;
1716 break;
1717 case vtn_variable_mode_ssbo:
1718 b->shader->info.num_ssbos++;
1719 break;
1720 case vtn_variable_mode_image:
1721 b->shader->info.num_images++;
1722 break;
1723 case vtn_variable_mode_sampler:
1724 b->shader->info.num_textures++;
1725 break;
1726 case vtn_variable_mode_push_constant:
1727 b->shader->num_uniforms = vtn_type_block_size(b, type);
1728 break;
1729 default:
1730 /* No tallying is needed */
1731 break;
1732 }
1733
1734 struct vtn_variable *var = rzalloc(b, struct vtn_variable);
1735 var->type = type;
1736 var->mode = mode;
1737
1738 vtn_assert(val->value_type == vtn_value_type_pointer);
1739 val->pointer = vtn_pointer_for_variable(b, var, ptr_type);
1740
1741 switch (var->mode) {
1742 case vtn_variable_mode_local:
1743 case vtn_variable_mode_global:
1744 case vtn_variable_mode_image:
1745 case vtn_variable_mode_sampler:
1746 /* For these, we create the variable normally */
1747 var->var = rzalloc(b->shader, nir_variable);
1748 var->var->name = ralloc_strdup(var->var, val->name);
1749 var->var->type = var->type->type;
1750 var->var->data.mode = nir_mode;
1751
1752 switch (var->mode) {
1753 case vtn_variable_mode_image:
1754 case vtn_variable_mode_sampler:
1755 var->var->interface_type = without_array->type;
1756 break;
1757 default:
1758 var->var->interface_type = NULL;
1759 break;
1760 }
1761 break;
1762
1763 case vtn_variable_mode_workgroup:
1764 if (b->options->lower_workgroup_access_to_offsets) {
1765 var->shared_location = -1;
1766 } else {
1767 /* Create the variable normally */
1768 var->var = rzalloc(b->shader, nir_variable);
1769 var->var->name = ralloc_strdup(var->var, val->name);
1770 var->var->type = var->type->type;
1771 var->var->data.mode = nir_var_shared;
1772 }
1773 break;
1774
1775 case vtn_variable_mode_input:
1776 case vtn_variable_mode_output: {
1777 /* In order to know whether or not we're a per-vertex inout, we need
1778 * the patch qualifier. This means walking the variable decorations
1779 * early before we actually create any variables. Not a big deal.
1780 *
1781 * GLSLang really likes to place decorations in the most interior
1782 * thing it possibly can. In particular, if you have a struct, it
1783 * will place the patch decorations on the struct members. This
1784 * should be handled by the variable splitting below just fine.
1785 *
1786 * If you have an array-of-struct, things get even more weird as it
1787 * will place the patch decorations on the struct even though it's
1788 * inside an array and some of the members being patch and others not
1789 * makes no sense whatsoever. Since the only sensible thing is for
1790 * it to be all or nothing, we'll call it patch if any of the members
1791 * are declared patch.
1792 */
1793 var->patch = false;
1794 vtn_foreach_decoration(b, val, var_is_patch_cb, &var->patch);
1795 if (glsl_type_is_array(var->type->type) &&
1796 glsl_type_is_struct(without_array->type)) {
1797 vtn_foreach_decoration(b, vtn_value(b, without_array->id,
1798 vtn_value_type_type),
1799 var_is_patch_cb, &var->patch);
1800 }
1801
1802 /* For inputs and outputs, we immediately split structures. This
1803 * is for a couple of reasons. For one, builtins may all come in
1804 * a struct and we really want those split out into separate
1805 * variables. For another, interpolation qualifiers can be
1806 * applied to members of the top-level struct ane we need to be
1807 * able to preserve that information.
1808 */
1809
1810 int array_length = -1;
1811 struct vtn_type *interface_type = var->type;
1812 if (is_per_vertex_inout(var, b->shader->info.stage)) {
1813 /* In Geometry shaders (and some tessellation), inputs come
1814 * in per-vertex arrays. However, some builtins come in
1815 * non-per-vertex, hence the need for the is_array check. In
1816 * any case, there are no non-builtin arrays allowed so this
1817 * check should be sufficient.
1818 */
1819 interface_type = var->type->array_element;
1820 array_length = glsl_get_length(var->type->type);
1821 }
1822
1823 if (glsl_type_is_struct(interface_type->type)) {
1824 /* It's a struct. Split it. */
1825 unsigned num_members = glsl_get_length(interface_type->type);
1826 var->members = ralloc_array(b, nir_variable *, num_members);
1827
1828 for (unsigned i = 0; i < num_members; i++) {
1829 const struct glsl_type *mtype = interface_type->members[i]->type;
1830 if (array_length >= 0)
1831 mtype = glsl_array_type(mtype, array_length);
1832
1833 var->members[i] = rzalloc(b->shader, nir_variable);
1834 var->members[i]->name =
1835 ralloc_asprintf(var->members[i], "%s.%d", val->name, i);
1836 var->members[i]->type = mtype;
1837 var->members[i]->interface_type =
1838 interface_type->members[i]->type;
1839 var->members[i]->data.mode = nir_mode;
1840 var->members[i]->data.patch = var->patch;
1841
1842 if (initializer) {
1843 assert(i < initializer->num_elements);
1844 var->members[i]->constant_initializer =
1845 nir_constant_clone(initializer->elements[i], var->members[i]);
1846 }
1847 }
1848
1849 initializer = NULL;
1850 } else {
1851 var->var = rzalloc(b->shader, nir_variable);
1852 var->var->name = ralloc_strdup(var->var, val->name);
1853 var->var->type = var->type->type;
1854 var->var->interface_type = interface_type->type;
1855 var->var->data.mode = nir_mode;
1856 var->var->data.patch = var->patch;
1857 }
1858
1859 /* For inputs and outputs, we need to grab locations and builtin
1860 * information from the interface type.
1861 */
1862 vtn_foreach_decoration(b, vtn_value(b, interface_type->id,
1863 vtn_value_type_type),
1864 var_decoration_cb, var);
1865 break;
1866 }
1867
1868 case vtn_variable_mode_param:
1869 vtn_fail("Not created through OpVariable");
1870
1871 case vtn_variable_mode_ubo:
1872 case vtn_variable_mode_ssbo:
1873 case vtn_variable_mode_push_constant:
1874 /* These don't need actual variables. */
1875 break;
1876 }
1877
1878 if (initializer) {
1879 var->var->constant_initializer =
1880 nir_constant_clone(initializer, var->var);
1881 }
1882
1883 vtn_foreach_decoration(b, val, var_decoration_cb, var);
1884
1885 if (var->mode == vtn_variable_mode_image ||
1886 var->mode == vtn_variable_mode_sampler) {
1887 /* XXX: We still need the binding information in the nir_variable
1888 * for these. We should fix that.
1889 */
1890 var->var->data.binding = var->binding;
1891 var->var->data.descriptor_set = var->descriptor_set;
1892 var->var->data.index = var->input_attachment_index;
1893
1894 if (var->mode == vtn_variable_mode_image)
1895 var->var->data.image.format = without_array->image_format;
1896 }
1897
1898 if (var->mode == vtn_variable_mode_local) {
1899 vtn_assert(var->members == NULL && var->var != NULL);
1900 nir_function_impl_add_variable(b->nb.impl, var->var);
1901 } else if (var->var) {
1902 nir_shader_add_variable(b->shader, var->var);
1903 } else if (var->members) {
1904 unsigned count = glsl_get_length(without_array->type);
1905 for (unsigned i = 0; i < count; i++) {
1906 vtn_assert(var->members[i]->data.mode != nir_var_local);
1907 nir_shader_add_variable(b->shader, var->members[i]);
1908 }
1909 } else {
1910 vtn_assert(vtn_pointer_is_external_block(b, val->pointer));
1911 }
1912 }
1913
1914 static void
1915 vtn_assert_types_equal(struct vtn_builder *b, SpvOp opcode,
1916 struct vtn_type *dst_type,
1917 struct vtn_type *src_type)
1918 {
1919 if (dst_type->id == src_type->id)
1920 return;
1921
1922 if (vtn_types_compatible(b, dst_type, src_type)) {
1923 /* Early versions of GLSLang would re-emit types unnecessarily and you
1924 * would end up with OpLoad, OpStore, or OpCopyMemory opcodes which have
1925 * mismatched source and destination types.
1926 *
1927 * https://github.com/KhronosGroup/glslang/issues/304
1928 * https://github.com/KhronosGroup/glslang/issues/307
1929 * https://bugs.freedesktop.org/show_bug.cgi?id=104338
1930 * https://bugs.freedesktop.org/show_bug.cgi?id=104424
1931 */
1932 vtn_warn("Source and destination types of %s do not have the same "
1933 "ID (but are compatible): %u vs %u",
1934 spirv_op_to_string(opcode), dst_type->id, src_type->id);
1935 return;
1936 }
1937
1938 vtn_fail("Source and destination types of %s do not match: %s vs. %s",
1939 spirv_op_to_string(opcode),
1940 glsl_get_type_name(dst_type->type),
1941 glsl_get_type_name(src_type->type));
1942 }
1943
1944 void
1945 vtn_handle_variables(struct vtn_builder *b, SpvOp opcode,
1946 const uint32_t *w, unsigned count)
1947 {
1948 switch (opcode) {
1949 case SpvOpUndef: {
1950 struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_undef);
1951 val->type = vtn_value(b, w[1], vtn_value_type_type)->type;
1952 break;
1953 }
1954
1955 case SpvOpVariable: {
1956 struct vtn_type *ptr_type = vtn_value(b, w[1], vtn_value_type_type)->type;
1957
1958 struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_pointer);
1959
1960 SpvStorageClass storage_class = w[3];
1961 nir_constant *initializer = NULL;
1962 if (count > 4)
1963 initializer = vtn_value(b, w[4], vtn_value_type_constant)->constant;
1964
1965 vtn_create_variable(b, val, ptr_type, storage_class, initializer);
1966 break;
1967 }
1968
1969 case SpvOpAccessChain:
1970 case SpvOpPtrAccessChain:
1971 case SpvOpInBoundsAccessChain: {
1972 struct vtn_access_chain *chain = vtn_access_chain_create(b, count - 4);
1973 chain->ptr_as_array = (opcode == SpvOpPtrAccessChain);
1974
1975 unsigned idx = 0;
1976 for (int i = 4; i < count; i++) {
1977 struct vtn_value *link_val = vtn_untyped_value(b, w[i]);
1978 if (link_val->value_type == vtn_value_type_constant) {
1979 chain->link[idx].mode = vtn_access_mode_literal;
1980 chain->link[idx].id = link_val->constant->values[0].u32[0];
1981 } else {
1982 chain->link[idx].mode = vtn_access_mode_id;
1983 chain->link[idx].id = w[i];
1984
1985 }
1986 idx++;
1987 }
1988
1989 struct vtn_type *ptr_type = vtn_value(b, w[1], vtn_value_type_type)->type;
1990 struct vtn_value *base_val = vtn_untyped_value(b, w[3]);
1991 if (base_val->value_type == vtn_value_type_sampled_image) {
1992 /* This is rather insane. SPIR-V allows you to use OpSampledImage
1993 * to combine an array of images with a single sampler to get an
1994 * array of sampled images that all share the same sampler.
1995 * Fortunately, this means that we can more-or-less ignore the
1996 * sampler when crawling the access chain, but it does leave us
1997 * with this rather awkward little special-case.
1998 */
1999 struct vtn_value *val =
2000 vtn_push_value(b, w[2], vtn_value_type_sampled_image);
2001 val->sampled_image = ralloc(b, struct vtn_sampled_image);
2002 val->sampled_image->type = base_val->sampled_image->type;
2003 val->sampled_image->image =
2004 vtn_pointer_dereference(b, base_val->sampled_image->image, chain);
2005 val->sampled_image->sampler = base_val->sampled_image->sampler;
2006 } else {
2007 vtn_assert(base_val->value_type == vtn_value_type_pointer);
2008 struct vtn_value *val =
2009 vtn_push_value(b, w[2], vtn_value_type_pointer);
2010 val->pointer = vtn_pointer_dereference(b, base_val->pointer, chain);
2011 val->pointer->ptr_type = ptr_type;
2012 }
2013 break;
2014 }
2015
2016 case SpvOpCopyMemory: {
2017 struct vtn_value *dest = vtn_value(b, w[1], vtn_value_type_pointer);
2018 struct vtn_value *src = vtn_value(b, w[2], vtn_value_type_pointer);
2019
2020 vtn_assert_types_equal(b, opcode, dest->type->deref, src->type->deref);
2021
2022 vtn_variable_copy(b, dest->pointer, src->pointer);
2023 break;
2024 }
2025
2026 case SpvOpLoad: {
2027 struct vtn_type *res_type =
2028 vtn_value(b, w[1], vtn_value_type_type)->type;
2029 struct vtn_value *src_val = vtn_value(b, w[3], vtn_value_type_pointer);
2030 struct vtn_pointer *src = src_val->pointer;
2031
2032 vtn_assert_types_equal(b, opcode, res_type, src_val->type->deref);
2033
2034 if (src->mode == vtn_variable_mode_image ||
2035 src->mode == vtn_variable_mode_sampler) {
2036 vtn_push_value(b, w[2], vtn_value_type_pointer)->pointer = src;
2037 return;
2038 }
2039
2040 vtn_push_ssa(b, w[2], res_type, vtn_variable_load(b, src));
2041 break;
2042 }
2043
2044 case SpvOpStore: {
2045 struct vtn_value *dest_val = vtn_value(b, w[1], vtn_value_type_pointer);
2046 struct vtn_pointer *dest = dest_val->pointer;
2047 struct vtn_value *src_val = vtn_untyped_value(b, w[2]);
2048
2049 /* OpStore requires us to actually have a storage type */
2050 vtn_fail_if(dest->type->type == NULL,
2051 "Invalid destination type for OpStore");
2052
2053 if (glsl_get_base_type(dest->type->type) == GLSL_TYPE_BOOL &&
2054 glsl_get_base_type(src_val->type->type) == GLSL_TYPE_UINT) {
2055 /* Early versions of GLSLang would use uint types for UBOs/SSBOs but
2056 * would then store them to a local variable as bool. Work around
2057 * the issue by doing an implicit conversion.
2058 *
2059 * https://github.com/KhronosGroup/glslang/issues/170
2060 * https://bugs.freedesktop.org/show_bug.cgi?id=104424
2061 */
2062 vtn_warn("OpStore of value of type OpTypeInt to a pointer to type "
2063 "OpTypeBool. Doing an implicit conversion to work around "
2064 "the problem.");
2065 struct vtn_ssa_value *bool_ssa =
2066 vtn_create_ssa_value(b, dest->type->type);
2067 bool_ssa->def = nir_i2b(&b->nb, vtn_ssa_value(b, w[2])->def);
2068 vtn_variable_store(b, bool_ssa, dest);
2069 break;
2070 }
2071
2072 vtn_assert_types_equal(b, opcode, dest_val->type->deref, src_val->type);
2073
2074 if (glsl_type_is_sampler(dest->type->type)) {
2075 vtn_warn("OpStore of a sampler detected. Doing on-the-fly copy "
2076 "propagation to workaround the problem.");
2077 vtn_assert(dest->var->copy_prop_sampler == NULL);
2078 dest->var->copy_prop_sampler =
2079 vtn_value(b, w[2], vtn_value_type_pointer)->pointer;
2080 break;
2081 }
2082
2083 struct vtn_ssa_value *src = vtn_ssa_value(b, w[2]);
2084 vtn_variable_store(b, src, dest);
2085 break;
2086 }
2087
2088 case SpvOpArrayLength: {
2089 struct vtn_pointer *ptr =
2090 vtn_value(b, w[3], vtn_value_type_pointer)->pointer;
2091
2092 const uint32_t offset = ptr->var->type->offsets[w[4]];
2093 const uint32_t stride = ptr->var->type->members[w[4]]->stride;
2094
2095 if (!ptr->block_index) {
2096 struct vtn_access_chain chain = {
2097 .length = 0,
2098 };
2099 ptr = vtn_ssa_offset_pointer_dereference(b, ptr, &chain);
2100 vtn_assert(ptr->block_index);
2101 }
2102
2103 nir_intrinsic_instr *instr =
2104 nir_intrinsic_instr_create(b->nb.shader,
2105 nir_intrinsic_get_buffer_size);
2106 instr->src[0] = nir_src_for_ssa(ptr->block_index);
2107 nir_ssa_dest_init(&instr->instr, &instr->dest, 1, 32, NULL);
2108 nir_builder_instr_insert(&b->nb, &instr->instr);
2109 nir_ssa_def *buf_size = &instr->dest.ssa;
2110
2111 /* array_length = max(buffer_size - offset, 0) / stride */
2112 nir_ssa_def *array_length =
2113 nir_idiv(&b->nb,
2114 nir_imax(&b->nb,
2115 nir_isub(&b->nb,
2116 buf_size,
2117 nir_imm_int(&b->nb, offset)),
2118 nir_imm_int(&b->nb, 0u)),
2119 nir_imm_int(&b->nb, stride));
2120
2121 struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_ssa);
2122 val->ssa = vtn_create_ssa_value(b, glsl_uint_type());
2123 val->ssa->def = array_length;
2124 break;
2125 }
2126
2127 case SpvOpCopyMemorySized:
2128 default:
2129 vtn_fail("Unhandled opcode");
2130 }
2131 }