spirv: Call nir_builder directly for vector_extract
[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 #include "nir_deref.h"
31 #include <vulkan/vulkan_core.h>
32
33 static void
34 ptr_decoration_cb(struct vtn_builder *b, struct vtn_value *val, int member,
35 const struct vtn_decoration *dec, void *void_ptr)
36 {
37 struct vtn_pointer *ptr = void_ptr;
38
39 switch (dec->decoration) {
40 case SpvDecorationNonUniformEXT:
41 ptr->access |= ACCESS_NON_UNIFORM;
42 break;
43
44 default:
45 break;
46 }
47 }
48
49 static struct vtn_pointer*
50 vtn_decorate_pointer(struct vtn_builder *b, struct vtn_value *val,
51 struct vtn_pointer *ptr)
52 {
53 struct vtn_pointer dummy = { .access = 0 };
54 vtn_foreach_decoration(b, val, ptr_decoration_cb, &dummy);
55
56 /* If we're adding access flags, make a copy of the pointer. We could
57 * probably just OR them in without doing so but this prevents us from
58 * leaking them any further than actually specified in the SPIR-V.
59 */
60 if (dummy.access & ~ptr->access) {
61 struct vtn_pointer *copy = ralloc(b, struct vtn_pointer);
62 *copy = *ptr;
63 copy->access |= dummy.access;
64 return copy;
65 }
66
67 return ptr;
68 }
69
70 struct vtn_value *
71 vtn_push_value_pointer(struct vtn_builder *b, uint32_t value_id,
72 struct vtn_pointer *ptr)
73 {
74 struct vtn_value *val = vtn_push_value(b, value_id, vtn_value_type_pointer);
75 val->pointer = vtn_decorate_pointer(b, val, ptr);
76 return val;
77 }
78
79 static void
80 ssa_decoration_cb(struct vtn_builder *b, struct vtn_value *val, int member,
81 const struct vtn_decoration *dec, void *void_ssa)
82 {
83 struct vtn_ssa_value *ssa = void_ssa;
84
85 switch (dec->decoration) {
86 case SpvDecorationNonUniformEXT:
87 ssa->access |= ACCESS_NON_UNIFORM;
88 break;
89
90 default:
91 break;
92 }
93 }
94
95 struct vtn_value *
96 vtn_push_ssa(struct vtn_builder *b, uint32_t value_id,
97 struct vtn_type *type, struct vtn_ssa_value *ssa)
98 {
99 struct vtn_value *val;
100 if (type->base_type == vtn_base_type_pointer) {
101 val = vtn_push_value_pointer(b, value_id, vtn_pointer_from_ssa(b, ssa->def, type));
102 } else {
103 val = vtn_push_value(b, value_id, vtn_value_type_ssa);
104 val->ssa = ssa;
105 vtn_foreach_decoration(b, val, ssa_decoration_cb, val->ssa);
106 }
107 return val;
108 }
109
110 static struct vtn_access_chain *
111 vtn_access_chain_create(struct vtn_builder *b, unsigned length)
112 {
113 struct vtn_access_chain *chain;
114
115 /* Subtract 1 from the length since there's already one built in */
116 size_t size = sizeof(*chain) +
117 (MAX2(length, 1) - 1) * sizeof(chain->link[0]);
118 chain = rzalloc_size(b, size);
119 chain->length = length;
120
121 return chain;
122 }
123
124 bool
125 vtn_mode_uses_ssa_offset(struct vtn_builder *b,
126 enum vtn_variable_mode mode)
127 {
128 return ((mode == vtn_variable_mode_ubo ||
129 mode == vtn_variable_mode_ssbo) &&
130 b->options->lower_ubo_ssbo_access_to_offsets) ||
131 mode == vtn_variable_mode_push_constant;
132 }
133
134 static bool
135 vtn_pointer_is_external_block(struct vtn_builder *b,
136 struct vtn_pointer *ptr)
137 {
138 return ptr->mode == vtn_variable_mode_ssbo ||
139 ptr->mode == vtn_variable_mode_ubo ||
140 ptr->mode == vtn_variable_mode_phys_ssbo ||
141 ptr->mode == vtn_variable_mode_push_constant;
142 }
143
144 static nir_ssa_def *
145 vtn_access_link_as_ssa(struct vtn_builder *b, struct vtn_access_link link,
146 unsigned stride, unsigned bit_size)
147 {
148 vtn_assert(stride > 0);
149 if (link.mode == vtn_access_mode_literal) {
150 return nir_imm_intN_t(&b->nb, link.id * stride, bit_size);
151 } else {
152 nir_ssa_def *ssa = vtn_ssa_value(b, link.id)->def;
153 if (ssa->bit_size != bit_size)
154 ssa = nir_i2i(&b->nb, ssa, bit_size);
155 return nir_imul_imm(&b->nb, ssa, stride);
156 }
157 }
158
159 static VkDescriptorType
160 vk_desc_type_for_mode(struct vtn_builder *b, enum vtn_variable_mode mode)
161 {
162 switch (mode) {
163 case vtn_variable_mode_ubo:
164 return VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER;
165 case vtn_variable_mode_ssbo:
166 return VK_DESCRIPTOR_TYPE_STORAGE_BUFFER;
167 default:
168 vtn_fail("Invalid mode for vulkan_resource_index");
169 }
170 }
171
172 static nir_ssa_def *
173 vtn_variable_resource_index(struct vtn_builder *b, struct vtn_variable *var,
174 nir_ssa_def *desc_array_index)
175 {
176 vtn_assert(b->options->environment == NIR_SPIRV_VULKAN);
177
178 if (!desc_array_index) {
179 vtn_assert(glsl_type_is_struct_or_ifc(var->type->type));
180 desc_array_index = nir_imm_int(&b->nb, 0);
181 }
182
183 nir_intrinsic_instr *instr =
184 nir_intrinsic_instr_create(b->nb.shader,
185 nir_intrinsic_vulkan_resource_index);
186 instr->src[0] = nir_src_for_ssa(desc_array_index);
187 nir_intrinsic_set_desc_set(instr, var->descriptor_set);
188 nir_intrinsic_set_binding(instr, var->binding);
189 nir_intrinsic_set_desc_type(instr, vk_desc_type_for_mode(b, var->mode));
190
191 vtn_fail_if(var->mode != vtn_variable_mode_ubo &&
192 var->mode != vtn_variable_mode_ssbo,
193 "Invalid mode for vulkan_resource_index");
194
195 nir_address_format addr_format = vtn_mode_to_address_format(b, var->mode);
196 const struct glsl_type *index_type =
197 b->options->lower_ubo_ssbo_access_to_offsets ?
198 glsl_uint_type() : nir_address_format_to_glsl_type(addr_format);
199
200 instr->num_components = glsl_get_vector_elements(index_type);
201 nir_ssa_dest_init(&instr->instr, &instr->dest, instr->num_components,
202 glsl_get_bit_size(index_type), NULL);
203 nir_builder_instr_insert(&b->nb, &instr->instr);
204
205 return &instr->dest.ssa;
206 }
207
208 static nir_ssa_def *
209 vtn_resource_reindex(struct vtn_builder *b, enum vtn_variable_mode mode,
210 nir_ssa_def *base_index, nir_ssa_def *offset_index)
211 {
212 vtn_assert(b->options->environment == NIR_SPIRV_VULKAN);
213
214 nir_intrinsic_instr *instr =
215 nir_intrinsic_instr_create(b->nb.shader,
216 nir_intrinsic_vulkan_resource_reindex);
217 instr->src[0] = nir_src_for_ssa(base_index);
218 instr->src[1] = nir_src_for_ssa(offset_index);
219 nir_intrinsic_set_desc_type(instr, vk_desc_type_for_mode(b, mode));
220
221 vtn_fail_if(mode != vtn_variable_mode_ubo && mode != vtn_variable_mode_ssbo,
222 "Invalid mode for vulkan_resource_reindex");
223
224 nir_address_format addr_format = vtn_mode_to_address_format(b, mode);
225 const struct glsl_type *index_type =
226 b->options->lower_ubo_ssbo_access_to_offsets ?
227 glsl_uint_type() : nir_address_format_to_glsl_type(addr_format);
228
229 instr->num_components = glsl_get_vector_elements(index_type);
230 nir_ssa_dest_init(&instr->instr, &instr->dest, instr->num_components,
231 glsl_get_bit_size(index_type), NULL);
232 nir_builder_instr_insert(&b->nb, &instr->instr);
233
234 return &instr->dest.ssa;
235 }
236
237 static nir_ssa_def *
238 vtn_descriptor_load(struct vtn_builder *b, enum vtn_variable_mode mode,
239 nir_ssa_def *desc_index)
240 {
241 vtn_assert(b->options->environment == NIR_SPIRV_VULKAN);
242
243 nir_intrinsic_instr *desc_load =
244 nir_intrinsic_instr_create(b->nb.shader,
245 nir_intrinsic_load_vulkan_descriptor);
246 desc_load->src[0] = nir_src_for_ssa(desc_index);
247 nir_intrinsic_set_desc_type(desc_load, vk_desc_type_for_mode(b, mode));
248
249 vtn_fail_if(mode != vtn_variable_mode_ubo && mode != vtn_variable_mode_ssbo,
250 "Invalid mode for load_vulkan_descriptor");
251
252 nir_address_format addr_format = vtn_mode_to_address_format(b, mode);
253 const struct glsl_type *ptr_type =
254 nir_address_format_to_glsl_type(addr_format);
255
256 desc_load->num_components = glsl_get_vector_elements(ptr_type);
257 nir_ssa_dest_init(&desc_load->instr, &desc_load->dest,
258 desc_load->num_components,
259 glsl_get_bit_size(ptr_type), NULL);
260 nir_builder_instr_insert(&b->nb, &desc_load->instr);
261
262 return &desc_load->dest.ssa;
263 }
264
265 /* Dereference the given base pointer by the access chain */
266 static struct vtn_pointer *
267 vtn_nir_deref_pointer_dereference(struct vtn_builder *b,
268 struct vtn_pointer *base,
269 struct vtn_access_chain *deref_chain)
270 {
271 struct vtn_type *type = base->type;
272 enum gl_access_qualifier access = base->access | deref_chain->access;
273 unsigned idx = 0;
274
275 nir_deref_instr *tail;
276 if (base->deref) {
277 tail = base->deref;
278 } else if (b->options->environment == NIR_SPIRV_VULKAN &&
279 vtn_pointer_is_external_block(b, base)) {
280 nir_ssa_def *block_index = base->block_index;
281
282 /* We dereferencing an external block pointer. Correctness of this
283 * operation relies on one particular line in the SPIR-V spec, section
284 * entitled "Validation Rules for Shader Capabilities":
285 *
286 * "Block and BufferBlock decorations cannot decorate a structure
287 * type that is nested at any level inside another structure type
288 * decorated with Block or BufferBlock."
289 *
290 * This means that we can detect the point where we cross over from
291 * descriptor indexing to buffer indexing by looking for the block
292 * decorated struct type. Anything before the block decorated struct
293 * type is a descriptor indexing operation and anything after the block
294 * decorated struct is a buffer offset operation.
295 */
296
297 /* Figure out the descriptor array index if any
298 *
299 * Some of the Vulkan CTS tests with hand-rolled SPIR-V have been known
300 * to forget the Block or BufferBlock decoration from time to time.
301 * It's more robust if we check for both !block_index and for the type
302 * to contain a block. This way there's a decent chance that arrays of
303 * UBOs/SSBOs will work correctly even if variable pointers are
304 * completley toast.
305 */
306 nir_ssa_def *desc_arr_idx = NULL;
307 if (!block_index || vtn_type_contains_block(b, type)) {
308 /* If our type contains a block, then we're still outside the block
309 * and we need to process enough levels of dereferences to get inside
310 * of it.
311 */
312 if (deref_chain->ptr_as_array) {
313 unsigned aoa_size = glsl_get_aoa_size(type->type);
314 desc_arr_idx = vtn_access_link_as_ssa(b, deref_chain->link[idx],
315 MAX2(aoa_size, 1), 32);
316 idx++;
317 }
318
319 for (; idx < deref_chain->length; idx++) {
320 if (type->base_type != vtn_base_type_array) {
321 vtn_assert(type->base_type == vtn_base_type_struct);
322 break;
323 }
324
325 unsigned aoa_size = glsl_get_aoa_size(type->array_element->type);
326 nir_ssa_def *arr_offset =
327 vtn_access_link_as_ssa(b, deref_chain->link[idx],
328 MAX2(aoa_size, 1), 32);
329 if (desc_arr_idx)
330 desc_arr_idx = nir_iadd(&b->nb, desc_arr_idx, arr_offset);
331 else
332 desc_arr_idx = arr_offset;
333
334 type = type->array_element;
335 access |= type->access;
336 }
337 }
338
339 if (!block_index) {
340 vtn_assert(base->var && base->type);
341 block_index = vtn_variable_resource_index(b, base->var, desc_arr_idx);
342 } else if (desc_arr_idx) {
343 block_index = vtn_resource_reindex(b, base->mode,
344 block_index, desc_arr_idx);
345 }
346
347 if (idx == deref_chain->length) {
348 /* The entire deref was consumed in finding the block index. Return
349 * a pointer which just has a block index and a later access chain
350 * will dereference deeper.
351 */
352 struct vtn_pointer *ptr = rzalloc(b, struct vtn_pointer);
353 ptr->mode = base->mode;
354 ptr->type = type;
355 ptr->block_index = block_index;
356 ptr->access = access;
357 return ptr;
358 }
359
360 /* If we got here, there's more access chain to handle and we have the
361 * final block index. Insert a descriptor load and cast to a deref to
362 * start the deref chain.
363 */
364 nir_ssa_def *desc = vtn_descriptor_load(b, base->mode, block_index);
365
366 assert(base->mode == vtn_variable_mode_ssbo ||
367 base->mode == vtn_variable_mode_ubo);
368 nir_variable_mode nir_mode =
369 base->mode == vtn_variable_mode_ssbo ? nir_var_mem_ssbo : nir_var_mem_ubo;
370
371 tail = nir_build_deref_cast(&b->nb, desc, nir_mode, type->type,
372 base->ptr_type->stride);
373 } else {
374 assert(base->var && base->var->var);
375 tail = nir_build_deref_var(&b->nb, base->var->var);
376 if (base->ptr_type && base->ptr_type->type) {
377 tail->dest.ssa.num_components =
378 glsl_get_vector_elements(base->ptr_type->type);
379 tail->dest.ssa.bit_size = glsl_get_bit_size(base->ptr_type->type);
380 }
381 }
382
383 if (idx == 0 && deref_chain->ptr_as_array) {
384 /* We start with a deref cast to get the stride. Hopefully, we'll be
385 * able to delete that cast eventually.
386 */
387 tail = nir_build_deref_cast(&b->nb, &tail->dest.ssa, tail->mode,
388 tail->type, base->ptr_type->stride);
389
390 nir_ssa_def *index = vtn_access_link_as_ssa(b, deref_chain->link[0], 1,
391 tail->dest.ssa.bit_size);
392 tail = nir_build_deref_ptr_as_array(&b->nb, tail, index);
393 idx++;
394 }
395
396 for (; idx < deref_chain->length; idx++) {
397 if (glsl_type_is_struct_or_ifc(type->type)) {
398 vtn_assert(deref_chain->link[idx].mode == vtn_access_mode_literal);
399 unsigned field = deref_chain->link[idx].id;
400 tail = nir_build_deref_struct(&b->nb, tail, field);
401 type = type->members[field];
402 } else {
403 nir_ssa_def *arr_index =
404 vtn_access_link_as_ssa(b, deref_chain->link[idx], 1,
405 tail->dest.ssa.bit_size);
406 tail = nir_build_deref_array(&b->nb, tail, arr_index);
407 type = type->array_element;
408 }
409
410 access |= type->access;
411 }
412
413 struct vtn_pointer *ptr = rzalloc(b, struct vtn_pointer);
414 ptr->mode = base->mode;
415 ptr->type = type;
416 ptr->var = base->var;
417 ptr->deref = tail;
418 ptr->access = access;
419
420 return ptr;
421 }
422
423 static struct vtn_pointer *
424 vtn_ssa_offset_pointer_dereference(struct vtn_builder *b,
425 struct vtn_pointer *base,
426 struct vtn_access_chain *deref_chain)
427 {
428 nir_ssa_def *block_index = base->block_index;
429 nir_ssa_def *offset = base->offset;
430 struct vtn_type *type = base->type;
431 enum gl_access_qualifier access = base->access;
432
433 unsigned idx = 0;
434 if (base->mode == vtn_variable_mode_ubo ||
435 base->mode == vtn_variable_mode_ssbo) {
436 if (!block_index) {
437 vtn_assert(base->var && base->type);
438 nir_ssa_def *desc_arr_idx;
439 if (glsl_type_is_array(type->type)) {
440 if (deref_chain->length >= 1) {
441 desc_arr_idx =
442 vtn_access_link_as_ssa(b, deref_chain->link[0], 1, 32);
443 idx++;
444 /* This consumes a level of type */
445 type = type->array_element;
446 access |= type->access;
447 } else {
448 /* This is annoying. We've been asked for a pointer to the
449 * array of UBOs/SSBOs and not a specifc buffer. Return a
450 * pointer with a descriptor index of 0 and we'll have to do
451 * a reindex later to adjust it to the right thing.
452 */
453 desc_arr_idx = nir_imm_int(&b->nb, 0);
454 }
455 } else if (deref_chain->ptr_as_array) {
456 /* You can't have a zero-length OpPtrAccessChain */
457 vtn_assert(deref_chain->length >= 1);
458 desc_arr_idx = vtn_access_link_as_ssa(b, deref_chain->link[0], 1, 32);
459 } else {
460 /* We have a regular non-array SSBO. */
461 desc_arr_idx = NULL;
462 }
463 block_index = vtn_variable_resource_index(b, base->var, desc_arr_idx);
464 } else if (deref_chain->ptr_as_array &&
465 type->base_type == vtn_base_type_struct && type->block) {
466 /* We are doing an OpPtrAccessChain on a pointer to a struct that is
467 * decorated block. This is an interesting corner in the SPIR-V
468 * spec. One interpretation would be that they client is clearly
469 * trying to treat that block as if it's an implicit array of blocks
470 * repeated in the buffer. However, the SPIR-V spec for the
471 * OpPtrAccessChain says:
472 *
473 * "Base is treated as the address of the first element of an
474 * array, and the Element element’s address is computed to be the
475 * base for the Indexes, as per OpAccessChain."
476 *
477 * Taken literally, that would mean that your struct type is supposed
478 * to be treated as an array of such a struct and, since it's
479 * decorated block, that means an array of blocks which corresponds
480 * to an array descriptor. Therefore, we need to do a reindex
481 * operation to add the index from the first link in the access chain
482 * to the index we recieved.
483 *
484 * The downside to this interpretation (there always is one) is that
485 * this might be somewhat surprising behavior to apps if they expect
486 * the implicit array behavior described above.
487 */
488 vtn_assert(deref_chain->length >= 1);
489 nir_ssa_def *offset_index =
490 vtn_access_link_as_ssa(b, deref_chain->link[0], 1, 32);
491 idx++;
492
493 block_index = vtn_resource_reindex(b, base->mode,
494 block_index, offset_index);
495 }
496 }
497
498 if (!offset) {
499 if (base->mode == vtn_variable_mode_workgroup) {
500 /* SLM doesn't need nor have a block index */
501 vtn_assert(!block_index);
502
503 /* We need the variable for the base offset */
504 vtn_assert(base->var);
505
506 /* We need ptr_type for size and alignment */
507 vtn_assert(base->ptr_type);
508
509 /* Assign location on first use so that we don't end up bloating SLM
510 * address space for variables which are never statically used.
511 */
512 if (base->var->shared_location < 0) {
513 vtn_assert(base->ptr_type->length > 0 && base->ptr_type->align > 0);
514 b->shader->num_shared = vtn_align_u32(b->shader->num_shared,
515 base->ptr_type->align);
516 base->var->shared_location = b->shader->num_shared;
517 b->shader->num_shared += base->ptr_type->length;
518 }
519
520 offset = nir_imm_int(&b->nb, base->var->shared_location);
521 } else if (base->mode == vtn_variable_mode_push_constant) {
522 /* Push constants neither need nor have a block index */
523 vtn_assert(!block_index);
524
525 /* Start off with at the start of the push constant block. */
526 offset = nir_imm_int(&b->nb, 0);
527 } else {
528 /* The code above should have ensured a block_index when needed. */
529 vtn_assert(block_index);
530
531 /* Start off with at the start of the buffer. */
532 offset = nir_imm_int(&b->nb, 0);
533 }
534 }
535
536 if (deref_chain->ptr_as_array && idx == 0) {
537 /* We need ptr_type for the stride */
538 vtn_assert(base->ptr_type);
539
540 /* We need at least one element in the chain */
541 vtn_assert(deref_chain->length >= 1);
542
543 nir_ssa_def *elem_offset =
544 vtn_access_link_as_ssa(b, deref_chain->link[idx],
545 base->ptr_type->stride, offset->bit_size);
546 offset = nir_iadd(&b->nb, offset, elem_offset);
547 idx++;
548 }
549
550 for (; idx < deref_chain->length; idx++) {
551 switch (glsl_get_base_type(type->type)) {
552 case GLSL_TYPE_UINT:
553 case GLSL_TYPE_INT:
554 case GLSL_TYPE_UINT16:
555 case GLSL_TYPE_INT16:
556 case GLSL_TYPE_UINT8:
557 case GLSL_TYPE_INT8:
558 case GLSL_TYPE_UINT64:
559 case GLSL_TYPE_INT64:
560 case GLSL_TYPE_FLOAT:
561 case GLSL_TYPE_FLOAT16:
562 case GLSL_TYPE_DOUBLE:
563 case GLSL_TYPE_BOOL:
564 case GLSL_TYPE_ARRAY: {
565 nir_ssa_def *elem_offset =
566 vtn_access_link_as_ssa(b, deref_chain->link[idx],
567 type->stride, offset->bit_size);
568 offset = nir_iadd(&b->nb, offset, elem_offset);
569 type = type->array_element;
570 access |= type->access;
571 break;
572 }
573
574 case GLSL_TYPE_INTERFACE:
575 case GLSL_TYPE_STRUCT: {
576 vtn_assert(deref_chain->link[idx].mode == vtn_access_mode_literal);
577 unsigned member = deref_chain->link[idx].id;
578 offset = nir_iadd_imm(&b->nb, offset, type->offsets[member]);
579 type = type->members[member];
580 access |= type->access;
581 break;
582 }
583
584 default:
585 vtn_fail("Invalid type for deref");
586 }
587 }
588
589 struct vtn_pointer *ptr = rzalloc(b, struct vtn_pointer);
590 ptr->mode = base->mode;
591 ptr->type = type;
592 ptr->block_index = block_index;
593 ptr->offset = offset;
594 ptr->access = access;
595
596 return ptr;
597 }
598
599 /* Dereference the given base pointer by the access chain */
600 static struct vtn_pointer *
601 vtn_pointer_dereference(struct vtn_builder *b,
602 struct vtn_pointer *base,
603 struct vtn_access_chain *deref_chain)
604 {
605 if (vtn_pointer_uses_ssa_offset(b, base)) {
606 return vtn_ssa_offset_pointer_dereference(b, base, deref_chain);
607 } else {
608 return vtn_nir_deref_pointer_dereference(b, base, deref_chain);
609 }
610 }
611
612 struct vtn_pointer *
613 vtn_pointer_for_variable(struct vtn_builder *b,
614 struct vtn_variable *var, struct vtn_type *ptr_type)
615 {
616 struct vtn_pointer *pointer = rzalloc(b, struct vtn_pointer);
617
618 pointer->mode = var->mode;
619 pointer->type = var->type;
620 vtn_assert(ptr_type->base_type == vtn_base_type_pointer);
621 vtn_assert(ptr_type->deref->type == var->type->type);
622 pointer->ptr_type = ptr_type;
623 pointer->var = var;
624 pointer->access = var->access | var->type->access;
625
626 return pointer;
627 }
628
629 /* Returns an atomic_uint type based on the original uint type. The returned
630 * type will be equivalent to the original one but will have an atomic_uint
631 * type as leaf instead of an uint.
632 *
633 * Manages uint scalars, arrays, and arrays of arrays of any nested depth.
634 */
635 static const struct glsl_type *
636 repair_atomic_type(const struct glsl_type *type)
637 {
638 assert(glsl_get_base_type(glsl_without_array(type)) == GLSL_TYPE_UINT);
639 assert(glsl_type_is_scalar(glsl_without_array(type)));
640
641 if (glsl_type_is_array(type)) {
642 const struct glsl_type *atomic =
643 repair_atomic_type(glsl_get_array_element(type));
644
645 return glsl_array_type(atomic, glsl_get_length(type),
646 glsl_get_explicit_stride(type));
647 } else {
648 return glsl_atomic_uint_type();
649 }
650 }
651
652 nir_deref_instr *
653 vtn_pointer_to_deref(struct vtn_builder *b, struct vtn_pointer *ptr)
654 {
655 if (b->wa_glslang_179) {
656 /* Do on-the-fly copy propagation for samplers. */
657 if (ptr->var && ptr->var->copy_prop_sampler)
658 return vtn_pointer_to_deref(b, ptr->var->copy_prop_sampler);
659 }
660
661 vtn_assert(!vtn_pointer_uses_ssa_offset(b, ptr));
662 if (!ptr->deref) {
663 struct vtn_access_chain chain = {
664 .length = 0,
665 };
666 ptr = vtn_nir_deref_pointer_dereference(b, ptr, &chain);
667 }
668
669 return ptr->deref;
670 }
671
672 static void
673 _vtn_local_load_store(struct vtn_builder *b, bool load, nir_deref_instr *deref,
674 struct vtn_ssa_value *inout,
675 enum gl_access_qualifier access)
676 {
677 if (glsl_type_is_vector_or_scalar(deref->type)) {
678 if (load) {
679 inout->def = nir_load_deref_with_access(&b->nb, deref, access);
680 } else {
681 nir_store_deref_with_access(&b->nb, deref, inout->def, ~0, access);
682 }
683 } else if (glsl_type_is_array(deref->type) ||
684 glsl_type_is_matrix(deref->type)) {
685 unsigned elems = glsl_get_length(deref->type);
686 for (unsigned i = 0; i < elems; i++) {
687 nir_deref_instr *child =
688 nir_build_deref_array_imm(&b->nb, deref, i);
689 _vtn_local_load_store(b, load, child, inout->elems[i], access);
690 }
691 } else {
692 vtn_assert(glsl_type_is_struct_or_ifc(deref->type));
693 unsigned elems = glsl_get_length(deref->type);
694 for (unsigned i = 0; i < elems; i++) {
695 nir_deref_instr *child = nir_build_deref_struct(&b->nb, deref, i);
696 _vtn_local_load_store(b, load, child, inout->elems[i], access);
697 }
698 }
699 }
700
701 nir_deref_instr *
702 vtn_nir_deref(struct vtn_builder *b, uint32_t id)
703 {
704 struct vtn_pointer *ptr = vtn_value(b, id, vtn_value_type_pointer)->pointer;
705 return vtn_pointer_to_deref(b, ptr);
706 }
707
708 /*
709 * Gets the NIR-level deref tail, which may have as a child an array deref
710 * selecting which component due to OpAccessChain supporting per-component
711 * indexing in SPIR-V.
712 */
713 static nir_deref_instr *
714 get_deref_tail(nir_deref_instr *deref)
715 {
716 if (deref->deref_type != nir_deref_type_array)
717 return deref;
718
719 nir_deref_instr *parent =
720 nir_instr_as_deref(deref->parent.ssa->parent_instr);
721
722 if (glsl_type_is_vector(parent->type))
723 return parent;
724 else
725 return deref;
726 }
727
728 struct vtn_ssa_value *
729 vtn_local_load(struct vtn_builder *b, nir_deref_instr *src,
730 enum gl_access_qualifier access)
731 {
732 nir_deref_instr *src_tail = get_deref_tail(src);
733 struct vtn_ssa_value *val = vtn_create_ssa_value(b, src_tail->type);
734 _vtn_local_load_store(b, true, src_tail, val, access);
735
736 if (src_tail != src) {
737 val->type = src->type;
738 val->def = nir_vector_extract(&b->nb, val->def, src->arr.index.ssa);
739 }
740
741 return val;
742 }
743
744 void
745 vtn_local_store(struct vtn_builder *b, struct vtn_ssa_value *src,
746 nir_deref_instr *dest, enum gl_access_qualifier access)
747 {
748 nir_deref_instr *dest_tail = get_deref_tail(dest);
749
750 if (dest_tail != dest) {
751 struct vtn_ssa_value *val = vtn_create_ssa_value(b, dest_tail->type);
752 _vtn_local_load_store(b, true, dest_tail, val, access);
753
754 if (nir_src_is_const(dest->arr.index))
755 val->def = vtn_vector_insert(b, val->def, src->def,
756 nir_src_as_uint(dest->arr.index));
757 else
758 val->def = vtn_vector_insert_dynamic(b, val->def, src->def,
759 dest->arr.index.ssa);
760 _vtn_local_load_store(b, false, dest_tail, val, access);
761 } else {
762 _vtn_local_load_store(b, false, dest_tail, src, access);
763 }
764 }
765
766 nir_ssa_def *
767 vtn_pointer_to_offset(struct vtn_builder *b, struct vtn_pointer *ptr,
768 nir_ssa_def **index_out)
769 {
770 assert(vtn_pointer_uses_ssa_offset(b, ptr));
771 if (!ptr->offset) {
772 struct vtn_access_chain chain = {
773 .length = 0,
774 };
775 ptr = vtn_ssa_offset_pointer_dereference(b, ptr, &chain);
776 }
777 *index_out = ptr->block_index;
778 return ptr->offset;
779 }
780
781 /* Tries to compute the size of an interface block based on the strides and
782 * offsets that are provided to us in the SPIR-V source.
783 */
784 static unsigned
785 vtn_type_block_size(struct vtn_builder *b, struct vtn_type *type)
786 {
787 enum glsl_base_type base_type = glsl_get_base_type(type->type);
788 switch (base_type) {
789 case GLSL_TYPE_UINT:
790 case GLSL_TYPE_INT:
791 case GLSL_TYPE_UINT16:
792 case GLSL_TYPE_INT16:
793 case GLSL_TYPE_UINT8:
794 case GLSL_TYPE_INT8:
795 case GLSL_TYPE_UINT64:
796 case GLSL_TYPE_INT64:
797 case GLSL_TYPE_FLOAT:
798 case GLSL_TYPE_FLOAT16:
799 case GLSL_TYPE_BOOL:
800 case GLSL_TYPE_DOUBLE: {
801 unsigned cols = type->row_major ? glsl_get_vector_elements(type->type) :
802 glsl_get_matrix_columns(type->type);
803 if (cols > 1) {
804 vtn_assert(type->stride > 0);
805 return type->stride * cols;
806 } else {
807 unsigned type_size = glsl_get_bit_size(type->type) / 8;
808 return glsl_get_vector_elements(type->type) * type_size;
809 }
810 }
811
812 case GLSL_TYPE_STRUCT:
813 case GLSL_TYPE_INTERFACE: {
814 unsigned size = 0;
815 unsigned num_fields = glsl_get_length(type->type);
816 for (unsigned f = 0; f < num_fields; f++) {
817 unsigned field_end = type->offsets[f] +
818 vtn_type_block_size(b, type->members[f]);
819 size = MAX2(size, field_end);
820 }
821 return size;
822 }
823
824 case GLSL_TYPE_ARRAY:
825 vtn_assert(type->stride > 0);
826 vtn_assert(glsl_get_length(type->type) > 0);
827 return type->stride * glsl_get_length(type->type);
828
829 default:
830 vtn_fail("Invalid block type");
831 return 0;
832 }
833 }
834
835 static void
836 _vtn_load_store_tail(struct vtn_builder *b, nir_intrinsic_op op, bool load,
837 nir_ssa_def *index, nir_ssa_def *offset,
838 unsigned access_offset, unsigned access_size,
839 struct vtn_ssa_value **inout, const struct glsl_type *type,
840 enum gl_access_qualifier access)
841 {
842 nir_intrinsic_instr *instr = nir_intrinsic_instr_create(b->nb.shader, op);
843 instr->num_components = glsl_get_vector_elements(type);
844
845 /* Booleans usually shouldn't show up in external memory in SPIR-V.
846 * However, they do for certain older GLSLang versions and can for shared
847 * memory when we lower access chains internally.
848 */
849 const unsigned data_bit_size = glsl_type_is_boolean(type) ? 32 :
850 glsl_get_bit_size(type);
851
852 int src = 0;
853 if (!load) {
854 nir_intrinsic_set_write_mask(instr, (1 << instr->num_components) - 1);
855 instr->src[src++] = nir_src_for_ssa((*inout)->def);
856 }
857
858 if (op == nir_intrinsic_load_push_constant) {
859 nir_intrinsic_set_base(instr, access_offset);
860 nir_intrinsic_set_range(instr, access_size);
861 }
862
863 if (op == nir_intrinsic_load_ubo ||
864 op == nir_intrinsic_load_ssbo ||
865 op == nir_intrinsic_store_ssbo) {
866 nir_intrinsic_set_access(instr, access);
867 }
868
869 /* With extensions like relaxed_block_layout, we really can't guarantee
870 * much more than scalar alignment.
871 */
872 if (op != nir_intrinsic_load_push_constant)
873 nir_intrinsic_set_align(instr, data_bit_size / 8, 0);
874
875 if (index)
876 instr->src[src++] = nir_src_for_ssa(index);
877
878 if (op == nir_intrinsic_load_push_constant) {
879 /* We need to subtract the offset from where the intrinsic will load the
880 * data. */
881 instr->src[src++] =
882 nir_src_for_ssa(nir_isub(&b->nb, offset,
883 nir_imm_int(&b->nb, access_offset)));
884 } else {
885 instr->src[src++] = nir_src_for_ssa(offset);
886 }
887
888 if (load) {
889 nir_ssa_dest_init(&instr->instr, &instr->dest,
890 instr->num_components, data_bit_size, NULL);
891 (*inout)->def = &instr->dest.ssa;
892 }
893
894 nir_builder_instr_insert(&b->nb, &instr->instr);
895
896 if (load && glsl_get_base_type(type) == GLSL_TYPE_BOOL)
897 (*inout)->def = nir_ine(&b->nb, (*inout)->def, nir_imm_int(&b->nb, 0));
898 }
899
900 static void
901 _vtn_block_load_store(struct vtn_builder *b, nir_intrinsic_op op, bool load,
902 nir_ssa_def *index, nir_ssa_def *offset,
903 unsigned access_offset, unsigned access_size,
904 struct vtn_type *type, enum gl_access_qualifier access,
905 struct vtn_ssa_value **inout)
906 {
907 if (load && *inout == NULL)
908 *inout = vtn_create_ssa_value(b, type->type);
909
910 enum glsl_base_type base_type = glsl_get_base_type(type->type);
911 switch (base_type) {
912 case GLSL_TYPE_UINT:
913 case GLSL_TYPE_INT:
914 case GLSL_TYPE_UINT16:
915 case GLSL_TYPE_INT16:
916 case GLSL_TYPE_UINT8:
917 case GLSL_TYPE_INT8:
918 case GLSL_TYPE_UINT64:
919 case GLSL_TYPE_INT64:
920 case GLSL_TYPE_FLOAT:
921 case GLSL_TYPE_FLOAT16:
922 case GLSL_TYPE_DOUBLE:
923 case GLSL_TYPE_BOOL:
924 /* This is where things get interesting. At this point, we've hit
925 * a vector, a scalar, or a matrix.
926 */
927 if (glsl_type_is_matrix(type->type)) {
928 /* Loading the whole matrix */
929 struct vtn_ssa_value *transpose;
930 unsigned num_ops, vec_width, col_stride;
931 if (type->row_major) {
932 num_ops = glsl_get_vector_elements(type->type);
933 vec_width = glsl_get_matrix_columns(type->type);
934 col_stride = type->array_element->stride;
935 if (load) {
936 const struct glsl_type *transpose_type =
937 glsl_matrix_type(base_type, vec_width, num_ops);
938 *inout = vtn_create_ssa_value(b, transpose_type);
939 } else {
940 transpose = vtn_ssa_transpose(b, *inout);
941 inout = &transpose;
942 }
943 } else {
944 num_ops = glsl_get_matrix_columns(type->type);
945 vec_width = glsl_get_vector_elements(type->type);
946 col_stride = type->stride;
947 }
948
949 for (unsigned i = 0; i < num_ops; i++) {
950 nir_ssa_def *elem_offset =
951 nir_iadd_imm(&b->nb, offset, i * col_stride);
952 _vtn_load_store_tail(b, op, load, index, elem_offset,
953 access_offset, access_size,
954 &(*inout)->elems[i],
955 glsl_vector_type(base_type, vec_width),
956 type->access | access);
957 }
958
959 if (load && type->row_major)
960 *inout = vtn_ssa_transpose(b, *inout);
961 } else {
962 unsigned elems = glsl_get_vector_elements(type->type);
963 unsigned type_size = glsl_get_bit_size(type->type) / 8;
964 if (elems == 1 || type->stride == type_size) {
965 /* This is a tightly-packed normal scalar or vector load */
966 vtn_assert(glsl_type_is_vector_or_scalar(type->type));
967 _vtn_load_store_tail(b, op, load, index, offset,
968 access_offset, access_size,
969 inout, type->type,
970 type->access | access);
971 } else {
972 /* This is a strided load. We have to load N things separately.
973 * This is the single column of a row-major matrix case.
974 */
975 vtn_assert(type->stride > type_size);
976 vtn_assert(type->stride % type_size == 0);
977
978 nir_ssa_def *per_comp[4];
979 for (unsigned i = 0; i < elems; i++) {
980 nir_ssa_def *elem_offset =
981 nir_iadd_imm(&b->nb, offset, i * type->stride);
982 struct vtn_ssa_value *comp, temp_val;
983 if (!load) {
984 temp_val.def = nir_channel(&b->nb, (*inout)->def, i);
985 temp_val.type = glsl_scalar_type(base_type);
986 }
987 comp = &temp_val;
988 _vtn_load_store_tail(b, op, load, index, elem_offset,
989 access_offset, access_size,
990 &comp, glsl_scalar_type(base_type),
991 type->access | access);
992 per_comp[i] = comp->def;
993 }
994
995 if (load) {
996 if (*inout == NULL)
997 *inout = vtn_create_ssa_value(b, type->type);
998 (*inout)->def = nir_vec(&b->nb, per_comp, elems);
999 }
1000 }
1001 }
1002 return;
1003
1004 case GLSL_TYPE_ARRAY: {
1005 unsigned elems = glsl_get_length(type->type);
1006 for (unsigned i = 0; i < elems; i++) {
1007 nir_ssa_def *elem_off =
1008 nir_iadd_imm(&b->nb, offset, i * type->stride);
1009 _vtn_block_load_store(b, op, load, index, elem_off,
1010 access_offset, access_size,
1011 type->array_element,
1012 type->array_element->access | access,
1013 &(*inout)->elems[i]);
1014 }
1015 return;
1016 }
1017
1018 case GLSL_TYPE_INTERFACE:
1019 case GLSL_TYPE_STRUCT: {
1020 unsigned elems = glsl_get_length(type->type);
1021 for (unsigned i = 0; i < elems; i++) {
1022 nir_ssa_def *elem_off =
1023 nir_iadd_imm(&b->nb, offset, type->offsets[i]);
1024 _vtn_block_load_store(b, op, load, index, elem_off,
1025 access_offset, access_size,
1026 type->members[i],
1027 type->members[i]->access | access,
1028 &(*inout)->elems[i]);
1029 }
1030 return;
1031 }
1032
1033 default:
1034 vtn_fail("Invalid block member type");
1035 }
1036 }
1037
1038 static struct vtn_ssa_value *
1039 vtn_block_load(struct vtn_builder *b, struct vtn_pointer *src)
1040 {
1041 nir_intrinsic_op op;
1042 unsigned access_offset = 0, access_size = 0;
1043 switch (src->mode) {
1044 case vtn_variable_mode_ubo:
1045 op = nir_intrinsic_load_ubo;
1046 break;
1047 case vtn_variable_mode_ssbo:
1048 op = nir_intrinsic_load_ssbo;
1049 break;
1050 case vtn_variable_mode_push_constant:
1051 op = nir_intrinsic_load_push_constant;
1052 access_size = b->shader->num_uniforms;
1053 break;
1054 case vtn_variable_mode_workgroup:
1055 op = nir_intrinsic_load_shared;
1056 break;
1057 default:
1058 vtn_fail("Invalid block variable mode");
1059 }
1060
1061 nir_ssa_def *offset, *index = NULL;
1062 offset = vtn_pointer_to_offset(b, src, &index);
1063
1064 struct vtn_ssa_value *value = NULL;
1065 _vtn_block_load_store(b, op, true, index, offset,
1066 access_offset, access_size,
1067 src->type, src->access, &value);
1068 return value;
1069 }
1070
1071 static void
1072 vtn_block_store(struct vtn_builder *b, struct vtn_ssa_value *src,
1073 struct vtn_pointer *dst)
1074 {
1075 nir_intrinsic_op op;
1076 switch (dst->mode) {
1077 case vtn_variable_mode_ssbo:
1078 op = nir_intrinsic_store_ssbo;
1079 break;
1080 case vtn_variable_mode_workgroup:
1081 op = nir_intrinsic_store_shared;
1082 break;
1083 default:
1084 vtn_fail("Invalid block variable mode");
1085 }
1086
1087 nir_ssa_def *offset, *index = NULL;
1088 offset = vtn_pointer_to_offset(b, dst, &index);
1089
1090 _vtn_block_load_store(b, op, false, index, offset,
1091 0, 0, dst->type, dst->access, &src);
1092 }
1093
1094 static void
1095 _vtn_variable_load_store(struct vtn_builder *b, bool load,
1096 struct vtn_pointer *ptr,
1097 enum gl_access_qualifier access,
1098 struct vtn_ssa_value **inout)
1099 {
1100 enum glsl_base_type base_type = glsl_get_base_type(ptr->type->type);
1101 switch (base_type) {
1102 case GLSL_TYPE_UINT:
1103 case GLSL_TYPE_INT:
1104 case GLSL_TYPE_UINT16:
1105 case GLSL_TYPE_INT16:
1106 case GLSL_TYPE_UINT8:
1107 case GLSL_TYPE_INT8:
1108 case GLSL_TYPE_UINT64:
1109 case GLSL_TYPE_INT64:
1110 case GLSL_TYPE_FLOAT:
1111 case GLSL_TYPE_FLOAT16:
1112 case GLSL_TYPE_BOOL:
1113 case GLSL_TYPE_DOUBLE:
1114 if (glsl_type_is_vector_or_scalar(ptr->type->type)) {
1115 /* We hit a vector or scalar; go ahead and emit the load[s] */
1116 nir_deref_instr *deref = vtn_pointer_to_deref(b, ptr);
1117 if (vtn_pointer_is_external_block(b, ptr)) {
1118 /* If it's external, we call nir_load/store_deref directly. The
1119 * vtn_local_load/store helpers are too clever and do magic to
1120 * avoid array derefs of vectors. That magic is both less
1121 * efficient than the direct load/store and, in the case of
1122 * stores, is broken because it creates a race condition if two
1123 * threads are writing to different components of the same vector
1124 * due to the load+insert+store it uses to emulate the array
1125 * deref.
1126 */
1127 if (load) {
1128 *inout = vtn_create_ssa_value(b, ptr->type->type);
1129 (*inout)->def = nir_load_deref_with_access(&b->nb, deref,
1130 ptr->type->access | access);
1131 } else {
1132 nir_store_deref_with_access(&b->nb, deref, (*inout)->def, ~0,
1133 ptr->type->access | access);
1134 }
1135 } else {
1136 if (load) {
1137 *inout = vtn_local_load(b, deref, ptr->type->access | access);
1138 } else {
1139 vtn_local_store(b, *inout, deref, ptr->type->access | access);
1140 }
1141 }
1142 return;
1143 }
1144 /* Fall through */
1145
1146 case GLSL_TYPE_INTERFACE:
1147 case GLSL_TYPE_ARRAY:
1148 case GLSL_TYPE_STRUCT: {
1149 unsigned elems = glsl_get_length(ptr->type->type);
1150 if (load) {
1151 vtn_assert(*inout == NULL);
1152 *inout = rzalloc(b, struct vtn_ssa_value);
1153 (*inout)->type = ptr->type->type;
1154 (*inout)->elems = rzalloc_array(b, struct vtn_ssa_value *, elems);
1155 }
1156
1157 struct vtn_access_chain chain = {
1158 .length = 1,
1159 .link = {
1160 { .mode = vtn_access_mode_literal, },
1161 }
1162 };
1163 for (unsigned i = 0; i < elems; i++) {
1164 chain.link[0].id = i;
1165 struct vtn_pointer *elem = vtn_pointer_dereference(b, ptr, &chain);
1166 _vtn_variable_load_store(b, load, elem, ptr->type->access | access,
1167 &(*inout)->elems[i]);
1168 }
1169 return;
1170 }
1171
1172 default:
1173 vtn_fail("Invalid access chain type");
1174 }
1175 }
1176
1177 struct vtn_ssa_value *
1178 vtn_variable_load(struct vtn_builder *b, struct vtn_pointer *src)
1179 {
1180 if (vtn_pointer_uses_ssa_offset(b, src)) {
1181 return vtn_block_load(b, src);
1182 } else {
1183 struct vtn_ssa_value *val = NULL;
1184 _vtn_variable_load_store(b, true, src, src->access, &val);
1185 return val;
1186 }
1187 }
1188
1189 void
1190 vtn_variable_store(struct vtn_builder *b, struct vtn_ssa_value *src,
1191 struct vtn_pointer *dest)
1192 {
1193 if (vtn_pointer_uses_ssa_offset(b, dest)) {
1194 vtn_assert(dest->mode == vtn_variable_mode_ssbo ||
1195 dest->mode == vtn_variable_mode_workgroup);
1196 vtn_block_store(b, src, dest);
1197 } else {
1198 _vtn_variable_load_store(b, false, dest, dest->access, &src);
1199 }
1200 }
1201
1202 static void
1203 _vtn_variable_copy(struct vtn_builder *b, struct vtn_pointer *dest,
1204 struct vtn_pointer *src)
1205 {
1206 vtn_assert(src->type->type == dest->type->type);
1207 enum glsl_base_type base_type = glsl_get_base_type(src->type->type);
1208 switch (base_type) {
1209 case GLSL_TYPE_UINT:
1210 case GLSL_TYPE_INT:
1211 case GLSL_TYPE_UINT16:
1212 case GLSL_TYPE_INT16:
1213 case GLSL_TYPE_UINT8:
1214 case GLSL_TYPE_INT8:
1215 case GLSL_TYPE_UINT64:
1216 case GLSL_TYPE_INT64:
1217 case GLSL_TYPE_FLOAT:
1218 case GLSL_TYPE_FLOAT16:
1219 case GLSL_TYPE_DOUBLE:
1220 case GLSL_TYPE_BOOL:
1221 /* At this point, we have a scalar, vector, or matrix so we know that
1222 * there cannot be any structure splitting still in the way. By
1223 * stopping at the matrix level rather than the vector level, we
1224 * ensure that matrices get loaded in the optimal way even if they
1225 * are storred row-major in a UBO.
1226 */
1227 vtn_variable_store(b, vtn_variable_load(b, src), dest);
1228 return;
1229
1230 case GLSL_TYPE_INTERFACE:
1231 case GLSL_TYPE_ARRAY:
1232 case GLSL_TYPE_STRUCT: {
1233 struct vtn_access_chain chain = {
1234 .length = 1,
1235 .link = {
1236 { .mode = vtn_access_mode_literal, },
1237 }
1238 };
1239 unsigned elems = glsl_get_length(src->type->type);
1240 for (unsigned i = 0; i < elems; i++) {
1241 chain.link[0].id = i;
1242 struct vtn_pointer *src_elem =
1243 vtn_pointer_dereference(b, src, &chain);
1244 struct vtn_pointer *dest_elem =
1245 vtn_pointer_dereference(b, dest, &chain);
1246
1247 _vtn_variable_copy(b, dest_elem, src_elem);
1248 }
1249 return;
1250 }
1251
1252 default:
1253 vtn_fail("Invalid access chain type");
1254 }
1255 }
1256
1257 static void
1258 vtn_variable_copy(struct vtn_builder *b, struct vtn_pointer *dest,
1259 struct vtn_pointer *src)
1260 {
1261 /* TODO: At some point, we should add a special-case for when we can
1262 * just emit a copy_var intrinsic.
1263 */
1264 _vtn_variable_copy(b, dest, src);
1265 }
1266
1267 static void
1268 set_mode_system_value(struct vtn_builder *b, nir_variable_mode *mode)
1269 {
1270 vtn_assert(*mode == nir_var_system_value || *mode == nir_var_shader_in);
1271 *mode = nir_var_system_value;
1272 }
1273
1274 static void
1275 vtn_get_builtin_location(struct vtn_builder *b,
1276 SpvBuiltIn builtin, int *location,
1277 nir_variable_mode *mode)
1278 {
1279 switch (builtin) {
1280 case SpvBuiltInPosition:
1281 *location = VARYING_SLOT_POS;
1282 break;
1283 case SpvBuiltInPointSize:
1284 *location = VARYING_SLOT_PSIZ;
1285 break;
1286 case SpvBuiltInClipDistance:
1287 *location = VARYING_SLOT_CLIP_DIST0; /* XXX CLIP_DIST1? */
1288 break;
1289 case SpvBuiltInCullDistance:
1290 *location = VARYING_SLOT_CULL_DIST0;
1291 break;
1292 case SpvBuiltInVertexId:
1293 case SpvBuiltInVertexIndex:
1294 /* The Vulkan spec defines VertexIndex to be non-zero-based and doesn't
1295 * allow VertexId. The ARB_gl_spirv spec defines VertexId to be the
1296 * same as gl_VertexID, which is non-zero-based, and removes
1297 * VertexIndex. Since they're both defined to be non-zero-based, we use
1298 * SYSTEM_VALUE_VERTEX_ID for both.
1299 */
1300 *location = SYSTEM_VALUE_VERTEX_ID;
1301 set_mode_system_value(b, mode);
1302 break;
1303 case SpvBuiltInInstanceIndex:
1304 *location = SYSTEM_VALUE_INSTANCE_INDEX;
1305 set_mode_system_value(b, mode);
1306 break;
1307 case SpvBuiltInInstanceId:
1308 *location = SYSTEM_VALUE_INSTANCE_ID;
1309 set_mode_system_value(b, mode);
1310 break;
1311 case SpvBuiltInPrimitiveId:
1312 if (b->shader->info.stage == MESA_SHADER_FRAGMENT) {
1313 vtn_assert(*mode == nir_var_shader_in);
1314 *location = VARYING_SLOT_PRIMITIVE_ID;
1315 } else if (*mode == nir_var_shader_out) {
1316 *location = VARYING_SLOT_PRIMITIVE_ID;
1317 } else {
1318 *location = SYSTEM_VALUE_PRIMITIVE_ID;
1319 set_mode_system_value(b, mode);
1320 }
1321 break;
1322 case SpvBuiltInInvocationId:
1323 *location = SYSTEM_VALUE_INVOCATION_ID;
1324 set_mode_system_value(b, mode);
1325 break;
1326 case SpvBuiltInLayer:
1327 *location = VARYING_SLOT_LAYER;
1328 if (b->shader->info.stage == MESA_SHADER_FRAGMENT)
1329 *mode = nir_var_shader_in;
1330 else if (b->shader->info.stage == MESA_SHADER_GEOMETRY)
1331 *mode = nir_var_shader_out;
1332 else if (b->options && b->options->caps.shader_viewport_index_layer &&
1333 (b->shader->info.stage == MESA_SHADER_VERTEX ||
1334 b->shader->info.stage == MESA_SHADER_TESS_EVAL))
1335 *mode = nir_var_shader_out;
1336 else
1337 vtn_fail("invalid stage for SpvBuiltInLayer");
1338 break;
1339 case SpvBuiltInViewportIndex:
1340 *location = VARYING_SLOT_VIEWPORT;
1341 if (b->shader->info.stage == MESA_SHADER_GEOMETRY)
1342 *mode = nir_var_shader_out;
1343 else if (b->options && b->options->caps.shader_viewport_index_layer &&
1344 (b->shader->info.stage == MESA_SHADER_VERTEX ||
1345 b->shader->info.stage == MESA_SHADER_TESS_EVAL))
1346 *mode = nir_var_shader_out;
1347 else if (b->shader->info.stage == MESA_SHADER_FRAGMENT)
1348 *mode = nir_var_shader_in;
1349 else
1350 vtn_fail("invalid stage for SpvBuiltInViewportIndex");
1351 break;
1352 case SpvBuiltInTessLevelOuter:
1353 *location = VARYING_SLOT_TESS_LEVEL_OUTER;
1354 break;
1355 case SpvBuiltInTessLevelInner:
1356 *location = VARYING_SLOT_TESS_LEVEL_INNER;
1357 break;
1358 case SpvBuiltInTessCoord:
1359 *location = SYSTEM_VALUE_TESS_COORD;
1360 set_mode_system_value(b, mode);
1361 break;
1362 case SpvBuiltInPatchVertices:
1363 *location = SYSTEM_VALUE_VERTICES_IN;
1364 set_mode_system_value(b, mode);
1365 break;
1366 case SpvBuiltInFragCoord:
1367 vtn_assert(*mode == nir_var_shader_in);
1368 if (b->options && b->options->frag_coord_is_sysval) {
1369 *mode = nir_var_system_value;
1370 *location = SYSTEM_VALUE_FRAG_COORD;
1371 } else {
1372 *location = VARYING_SLOT_POS;
1373 }
1374 break;
1375 case SpvBuiltInPointCoord:
1376 *location = VARYING_SLOT_PNTC;
1377 vtn_assert(*mode == nir_var_shader_in);
1378 break;
1379 case SpvBuiltInFrontFacing:
1380 *location = SYSTEM_VALUE_FRONT_FACE;
1381 set_mode_system_value(b, mode);
1382 break;
1383 case SpvBuiltInSampleId:
1384 *location = SYSTEM_VALUE_SAMPLE_ID;
1385 set_mode_system_value(b, mode);
1386 break;
1387 case SpvBuiltInSamplePosition:
1388 *location = SYSTEM_VALUE_SAMPLE_POS;
1389 set_mode_system_value(b, mode);
1390 break;
1391 case SpvBuiltInSampleMask:
1392 if (*mode == nir_var_shader_out) {
1393 *location = FRAG_RESULT_SAMPLE_MASK;
1394 } else {
1395 *location = SYSTEM_VALUE_SAMPLE_MASK_IN;
1396 set_mode_system_value(b, mode);
1397 }
1398 break;
1399 case SpvBuiltInFragDepth:
1400 *location = FRAG_RESULT_DEPTH;
1401 vtn_assert(*mode == nir_var_shader_out);
1402 break;
1403 case SpvBuiltInHelperInvocation:
1404 *location = SYSTEM_VALUE_HELPER_INVOCATION;
1405 set_mode_system_value(b, mode);
1406 break;
1407 case SpvBuiltInNumWorkgroups:
1408 *location = SYSTEM_VALUE_NUM_WORK_GROUPS;
1409 set_mode_system_value(b, mode);
1410 break;
1411 case SpvBuiltInWorkgroupSize:
1412 *location = SYSTEM_VALUE_LOCAL_GROUP_SIZE;
1413 set_mode_system_value(b, mode);
1414 break;
1415 case SpvBuiltInWorkgroupId:
1416 *location = SYSTEM_VALUE_WORK_GROUP_ID;
1417 set_mode_system_value(b, mode);
1418 break;
1419 case SpvBuiltInLocalInvocationId:
1420 *location = SYSTEM_VALUE_LOCAL_INVOCATION_ID;
1421 set_mode_system_value(b, mode);
1422 break;
1423 case SpvBuiltInLocalInvocationIndex:
1424 *location = SYSTEM_VALUE_LOCAL_INVOCATION_INDEX;
1425 set_mode_system_value(b, mode);
1426 break;
1427 case SpvBuiltInGlobalInvocationId:
1428 *location = SYSTEM_VALUE_GLOBAL_INVOCATION_ID;
1429 set_mode_system_value(b, mode);
1430 break;
1431 case SpvBuiltInGlobalLinearId:
1432 *location = SYSTEM_VALUE_GLOBAL_INVOCATION_INDEX;
1433 set_mode_system_value(b, mode);
1434 break;
1435 case SpvBuiltInBaseVertex:
1436 /* OpenGL gl_BaseVertex (SYSTEM_VALUE_BASE_VERTEX) is not the same
1437 * semantic as Vulkan BaseVertex (SYSTEM_VALUE_FIRST_VERTEX).
1438 */
1439 if (b->options->environment == NIR_SPIRV_OPENGL)
1440 *location = SYSTEM_VALUE_BASE_VERTEX;
1441 else
1442 *location = SYSTEM_VALUE_FIRST_VERTEX;
1443 set_mode_system_value(b, mode);
1444 break;
1445 case SpvBuiltInBaseInstance:
1446 *location = SYSTEM_VALUE_BASE_INSTANCE;
1447 set_mode_system_value(b, mode);
1448 break;
1449 case SpvBuiltInDrawIndex:
1450 *location = SYSTEM_VALUE_DRAW_ID;
1451 set_mode_system_value(b, mode);
1452 break;
1453 case SpvBuiltInSubgroupSize:
1454 *location = SYSTEM_VALUE_SUBGROUP_SIZE;
1455 set_mode_system_value(b, mode);
1456 break;
1457 case SpvBuiltInSubgroupId:
1458 *location = SYSTEM_VALUE_SUBGROUP_ID;
1459 set_mode_system_value(b, mode);
1460 break;
1461 case SpvBuiltInSubgroupLocalInvocationId:
1462 *location = SYSTEM_VALUE_SUBGROUP_INVOCATION;
1463 set_mode_system_value(b, mode);
1464 break;
1465 case SpvBuiltInNumSubgroups:
1466 *location = SYSTEM_VALUE_NUM_SUBGROUPS;
1467 set_mode_system_value(b, mode);
1468 break;
1469 case SpvBuiltInDeviceIndex:
1470 *location = SYSTEM_VALUE_DEVICE_INDEX;
1471 set_mode_system_value(b, mode);
1472 break;
1473 case SpvBuiltInViewIndex:
1474 *location = SYSTEM_VALUE_VIEW_INDEX;
1475 set_mode_system_value(b, mode);
1476 break;
1477 case SpvBuiltInSubgroupEqMask:
1478 *location = SYSTEM_VALUE_SUBGROUP_EQ_MASK,
1479 set_mode_system_value(b, mode);
1480 break;
1481 case SpvBuiltInSubgroupGeMask:
1482 *location = SYSTEM_VALUE_SUBGROUP_GE_MASK,
1483 set_mode_system_value(b, mode);
1484 break;
1485 case SpvBuiltInSubgroupGtMask:
1486 *location = SYSTEM_VALUE_SUBGROUP_GT_MASK,
1487 set_mode_system_value(b, mode);
1488 break;
1489 case SpvBuiltInSubgroupLeMask:
1490 *location = SYSTEM_VALUE_SUBGROUP_LE_MASK,
1491 set_mode_system_value(b, mode);
1492 break;
1493 case SpvBuiltInSubgroupLtMask:
1494 *location = SYSTEM_VALUE_SUBGROUP_LT_MASK,
1495 set_mode_system_value(b, mode);
1496 break;
1497 case SpvBuiltInFragStencilRefEXT:
1498 *location = FRAG_RESULT_STENCIL;
1499 vtn_assert(*mode == nir_var_shader_out);
1500 break;
1501 case SpvBuiltInWorkDim:
1502 *location = SYSTEM_VALUE_WORK_DIM;
1503 set_mode_system_value(b, mode);
1504 break;
1505 case SpvBuiltInGlobalSize:
1506 *location = SYSTEM_VALUE_GLOBAL_GROUP_SIZE;
1507 set_mode_system_value(b, mode);
1508 break;
1509 case SpvBuiltInBaryCoordNoPerspAMD:
1510 *location = SYSTEM_VALUE_BARYCENTRIC_LINEAR_PIXEL;
1511 set_mode_system_value(b, mode);
1512 break;
1513 case SpvBuiltInBaryCoordNoPerspCentroidAMD:
1514 *location = SYSTEM_VALUE_BARYCENTRIC_LINEAR_CENTROID;
1515 set_mode_system_value(b, mode);
1516 break;
1517 case SpvBuiltInBaryCoordNoPerspSampleAMD:
1518 *location = SYSTEM_VALUE_BARYCENTRIC_LINEAR_SAMPLE;
1519 set_mode_system_value(b, mode);
1520 break;
1521 case SpvBuiltInBaryCoordSmoothAMD:
1522 *location = SYSTEM_VALUE_BARYCENTRIC_PERSP_PIXEL;
1523 set_mode_system_value(b, mode);
1524 break;
1525 case SpvBuiltInBaryCoordSmoothCentroidAMD:
1526 *location = SYSTEM_VALUE_BARYCENTRIC_PERSP_CENTROID;
1527 set_mode_system_value(b, mode);
1528 break;
1529 case SpvBuiltInBaryCoordSmoothSampleAMD:
1530 *location = SYSTEM_VALUE_BARYCENTRIC_PERSP_SAMPLE;
1531 set_mode_system_value(b, mode);
1532 break;
1533 case SpvBuiltInBaryCoordPullModelAMD:
1534 *location = SYSTEM_VALUE_BARYCENTRIC_PULL_MODEL;
1535 set_mode_system_value(b, mode);
1536 break;
1537 default:
1538 vtn_fail("Unsupported builtin: %s (%u)",
1539 spirv_builtin_to_string(builtin), builtin);
1540 }
1541 }
1542
1543 static void
1544 apply_var_decoration(struct vtn_builder *b,
1545 struct nir_variable_data *var_data,
1546 const struct vtn_decoration *dec)
1547 {
1548 switch (dec->decoration) {
1549 case SpvDecorationRelaxedPrecision:
1550 break; /* FIXME: Do nothing with this for now. */
1551 case SpvDecorationNoPerspective:
1552 var_data->interpolation = INTERP_MODE_NOPERSPECTIVE;
1553 break;
1554 case SpvDecorationFlat:
1555 var_data->interpolation = INTERP_MODE_FLAT;
1556 break;
1557 case SpvDecorationExplicitInterpAMD:
1558 var_data->interpolation = INTERP_MODE_EXPLICIT;
1559 break;
1560 case SpvDecorationCentroid:
1561 var_data->centroid = true;
1562 break;
1563 case SpvDecorationSample:
1564 var_data->sample = true;
1565 break;
1566 case SpvDecorationInvariant:
1567 var_data->invariant = true;
1568 break;
1569 case SpvDecorationConstant:
1570 var_data->read_only = true;
1571 break;
1572 case SpvDecorationNonReadable:
1573 var_data->access |= ACCESS_NON_READABLE;
1574 break;
1575 case SpvDecorationNonWritable:
1576 var_data->read_only = true;
1577 var_data->access |= ACCESS_NON_WRITEABLE;
1578 break;
1579 case SpvDecorationRestrict:
1580 var_data->access |= ACCESS_RESTRICT;
1581 break;
1582 case SpvDecorationVolatile:
1583 var_data->access |= ACCESS_VOLATILE;
1584 break;
1585 case SpvDecorationCoherent:
1586 var_data->access |= ACCESS_COHERENT;
1587 break;
1588 case SpvDecorationComponent:
1589 var_data->location_frac = dec->operands[0];
1590 break;
1591 case SpvDecorationIndex:
1592 var_data->index = dec->operands[0];
1593 break;
1594 case SpvDecorationBuiltIn: {
1595 SpvBuiltIn builtin = dec->operands[0];
1596
1597 nir_variable_mode mode = var_data->mode;
1598 vtn_get_builtin_location(b, builtin, &var_data->location, &mode);
1599 var_data->mode = mode;
1600
1601 switch (builtin) {
1602 case SpvBuiltInTessLevelOuter:
1603 case SpvBuiltInTessLevelInner:
1604 case SpvBuiltInClipDistance:
1605 case SpvBuiltInCullDistance:
1606 var_data->compact = true;
1607 break;
1608 default:
1609 break;
1610 }
1611 }
1612
1613 case SpvDecorationSpecId:
1614 case SpvDecorationRowMajor:
1615 case SpvDecorationColMajor:
1616 case SpvDecorationMatrixStride:
1617 case SpvDecorationAliased:
1618 case SpvDecorationUniform:
1619 case SpvDecorationUniformId:
1620 case SpvDecorationLinkageAttributes:
1621 break; /* Do nothing with these here */
1622
1623 case SpvDecorationPatch:
1624 var_data->patch = true;
1625 break;
1626
1627 case SpvDecorationLocation:
1628 vtn_fail("Handled above");
1629
1630 case SpvDecorationBlock:
1631 case SpvDecorationBufferBlock:
1632 case SpvDecorationArrayStride:
1633 case SpvDecorationGLSLShared:
1634 case SpvDecorationGLSLPacked:
1635 break; /* These can apply to a type but we don't care about them */
1636
1637 case SpvDecorationBinding:
1638 case SpvDecorationDescriptorSet:
1639 case SpvDecorationNoContraction:
1640 case SpvDecorationInputAttachmentIndex:
1641 vtn_warn("Decoration not allowed for variable or structure member: %s",
1642 spirv_decoration_to_string(dec->decoration));
1643 break;
1644
1645 case SpvDecorationXfbBuffer:
1646 var_data->explicit_xfb_buffer = true;
1647 var_data->xfb.buffer = dec->operands[0];
1648 var_data->always_active_io = true;
1649 break;
1650 case SpvDecorationXfbStride:
1651 var_data->explicit_xfb_stride = true;
1652 var_data->xfb.stride = dec->operands[0];
1653 break;
1654 case SpvDecorationOffset:
1655 var_data->explicit_offset = true;
1656 var_data->offset = dec->operands[0];
1657 break;
1658
1659 case SpvDecorationStream:
1660 var_data->stream = dec->operands[0];
1661 break;
1662
1663 case SpvDecorationCPacked:
1664 case SpvDecorationSaturatedConversion:
1665 case SpvDecorationFuncParamAttr:
1666 case SpvDecorationFPRoundingMode:
1667 case SpvDecorationFPFastMathMode:
1668 case SpvDecorationAlignment:
1669 if (b->shader->info.stage != MESA_SHADER_KERNEL) {
1670 vtn_warn("Decoration only allowed for CL-style kernels: %s",
1671 spirv_decoration_to_string(dec->decoration));
1672 }
1673 break;
1674
1675 case SpvDecorationUserSemantic:
1676 /* User semantic decorations can safely be ignored by the driver. */
1677 break;
1678
1679 case SpvDecorationRestrictPointerEXT:
1680 case SpvDecorationAliasedPointerEXT:
1681 /* TODO: We should actually plumb alias information through NIR. */
1682 break;
1683
1684 default:
1685 vtn_fail_with_decoration("Unhandled decoration", dec->decoration);
1686 }
1687 }
1688
1689 static void
1690 var_is_patch_cb(struct vtn_builder *b, struct vtn_value *val, int member,
1691 const struct vtn_decoration *dec, void *out_is_patch)
1692 {
1693 if (dec->decoration == SpvDecorationPatch) {
1694 *((bool *) out_is_patch) = true;
1695 }
1696 }
1697
1698 static void
1699 var_decoration_cb(struct vtn_builder *b, struct vtn_value *val, int member,
1700 const struct vtn_decoration *dec, void *void_var)
1701 {
1702 struct vtn_variable *vtn_var = void_var;
1703
1704 /* Handle decorations that apply to a vtn_variable as a whole */
1705 switch (dec->decoration) {
1706 case SpvDecorationBinding:
1707 vtn_var->binding = dec->operands[0];
1708 vtn_var->explicit_binding = true;
1709 return;
1710 case SpvDecorationDescriptorSet:
1711 vtn_var->descriptor_set = dec->operands[0];
1712 return;
1713 case SpvDecorationInputAttachmentIndex:
1714 vtn_var->input_attachment_index = dec->operands[0];
1715 return;
1716 case SpvDecorationPatch:
1717 vtn_var->patch = true;
1718 break;
1719 case SpvDecorationOffset:
1720 vtn_var->offset = dec->operands[0];
1721 break;
1722 case SpvDecorationNonWritable:
1723 vtn_var->access |= ACCESS_NON_WRITEABLE;
1724 break;
1725 case SpvDecorationNonReadable:
1726 vtn_var->access |= ACCESS_NON_READABLE;
1727 break;
1728 case SpvDecorationVolatile:
1729 vtn_var->access |= ACCESS_VOLATILE;
1730 break;
1731 case SpvDecorationCoherent:
1732 vtn_var->access |= ACCESS_COHERENT;
1733 break;
1734 case SpvDecorationCounterBuffer:
1735 /* Counter buffer decorations can safely be ignored by the driver. */
1736 return;
1737 default:
1738 break;
1739 }
1740
1741 if (val->value_type == vtn_value_type_pointer) {
1742 assert(val->pointer->var == void_var);
1743 assert(member == -1);
1744 } else {
1745 assert(val->value_type == vtn_value_type_type);
1746 }
1747
1748 /* Location is odd. If applied to a split structure, we have to walk the
1749 * whole thing and accumulate the location. It's easier to handle as a
1750 * special case.
1751 */
1752 if (dec->decoration == SpvDecorationLocation) {
1753 unsigned location = dec->operands[0];
1754 if (b->shader->info.stage == MESA_SHADER_FRAGMENT &&
1755 vtn_var->mode == vtn_variable_mode_output) {
1756 location += FRAG_RESULT_DATA0;
1757 } else if (b->shader->info.stage == MESA_SHADER_VERTEX &&
1758 vtn_var->mode == vtn_variable_mode_input) {
1759 location += VERT_ATTRIB_GENERIC0;
1760 } else if (vtn_var->mode == vtn_variable_mode_input ||
1761 vtn_var->mode == vtn_variable_mode_output) {
1762 location += vtn_var->patch ? VARYING_SLOT_PATCH0 : VARYING_SLOT_VAR0;
1763 } else if (vtn_var->mode != vtn_variable_mode_uniform) {
1764 vtn_warn("Location must be on input, output, uniform, sampler or "
1765 "image variable");
1766 return;
1767 }
1768
1769 if (vtn_var->var->num_members == 0) {
1770 /* This handles the member and lone variable cases */
1771 vtn_var->var->data.location = location;
1772 } else {
1773 /* This handles the structure member case */
1774 assert(vtn_var->var->members);
1775
1776 if (member == -1)
1777 vtn_var->base_location = location;
1778 else
1779 vtn_var->var->members[member].location = location;
1780 }
1781
1782 return;
1783 } else {
1784 if (vtn_var->var) {
1785 if (vtn_var->var->num_members == 0) {
1786 /* We call this function on types as well as variables and not all
1787 * struct types get split so we can end up having stray member
1788 * decorations; just ignore them.
1789 */
1790 if (member == -1)
1791 apply_var_decoration(b, &vtn_var->var->data, dec);
1792 } else if (member >= 0) {
1793 /* Member decorations must come from a type */
1794 assert(val->value_type == vtn_value_type_type);
1795 apply_var_decoration(b, &vtn_var->var->members[member], dec);
1796 } else {
1797 unsigned length =
1798 glsl_get_length(glsl_without_array(vtn_var->type->type));
1799 for (unsigned i = 0; i < length; i++)
1800 apply_var_decoration(b, &vtn_var->var->members[i], dec);
1801 }
1802 } else {
1803 /* A few variables, those with external storage, have no actual
1804 * nir_variables associated with them. Fortunately, all decorations
1805 * we care about for those variables are on the type only.
1806 */
1807 vtn_assert(vtn_var->mode == vtn_variable_mode_ubo ||
1808 vtn_var->mode == vtn_variable_mode_ssbo ||
1809 vtn_var->mode == vtn_variable_mode_push_constant);
1810 }
1811 }
1812 }
1813
1814 enum vtn_variable_mode
1815 vtn_storage_class_to_mode(struct vtn_builder *b,
1816 SpvStorageClass class,
1817 struct vtn_type *interface_type,
1818 nir_variable_mode *nir_mode_out)
1819 {
1820 enum vtn_variable_mode mode;
1821 nir_variable_mode nir_mode;
1822 switch (class) {
1823 case SpvStorageClassUniform:
1824 /* Assume it's an UBO if we lack the interface_type. */
1825 if (!interface_type || interface_type->block) {
1826 mode = vtn_variable_mode_ubo;
1827 nir_mode = nir_var_mem_ubo;
1828 } else if (interface_type->buffer_block) {
1829 mode = vtn_variable_mode_ssbo;
1830 nir_mode = nir_var_mem_ssbo;
1831 } else {
1832 /* Default-block uniforms, coming from gl_spirv */
1833 mode = vtn_variable_mode_uniform;
1834 nir_mode = nir_var_uniform;
1835 }
1836 break;
1837 case SpvStorageClassStorageBuffer:
1838 mode = vtn_variable_mode_ssbo;
1839 nir_mode = nir_var_mem_ssbo;
1840 break;
1841 case SpvStorageClassPhysicalStorageBuffer:
1842 mode = vtn_variable_mode_phys_ssbo;
1843 nir_mode = nir_var_mem_global;
1844 break;
1845 case SpvStorageClassUniformConstant:
1846 if (b->shader->info.stage == MESA_SHADER_KERNEL) {
1847 if (b->options->constant_as_global) {
1848 mode = vtn_variable_mode_cross_workgroup;
1849 nir_mode = nir_var_mem_global;
1850 } else {
1851 mode = vtn_variable_mode_ubo;
1852 nir_mode = nir_var_mem_ubo;
1853 }
1854 } else {
1855 mode = vtn_variable_mode_uniform;
1856 nir_mode = nir_var_uniform;
1857 }
1858 break;
1859 case SpvStorageClassPushConstant:
1860 mode = vtn_variable_mode_push_constant;
1861 nir_mode = nir_var_uniform;
1862 break;
1863 case SpvStorageClassInput:
1864 mode = vtn_variable_mode_input;
1865 nir_mode = nir_var_shader_in;
1866 break;
1867 case SpvStorageClassOutput:
1868 mode = vtn_variable_mode_output;
1869 nir_mode = nir_var_shader_out;
1870 break;
1871 case SpvStorageClassPrivate:
1872 mode = vtn_variable_mode_private;
1873 nir_mode = nir_var_shader_temp;
1874 break;
1875 case SpvStorageClassFunction:
1876 mode = vtn_variable_mode_function;
1877 nir_mode = nir_var_function_temp;
1878 break;
1879 case SpvStorageClassWorkgroup:
1880 mode = vtn_variable_mode_workgroup;
1881 nir_mode = nir_var_mem_shared;
1882 break;
1883 case SpvStorageClassAtomicCounter:
1884 mode = vtn_variable_mode_uniform;
1885 nir_mode = nir_var_uniform;
1886 break;
1887 case SpvStorageClassCrossWorkgroup:
1888 mode = vtn_variable_mode_cross_workgroup;
1889 nir_mode = nir_var_mem_global;
1890 break;
1891 case SpvStorageClassImage:
1892 mode = vtn_variable_mode_image;
1893 nir_mode = nir_var_mem_ubo;
1894 break;
1895 case SpvStorageClassGeneric:
1896 default:
1897 vtn_fail("Unhandled variable storage class: %s (%u)",
1898 spirv_storageclass_to_string(class), class);
1899 }
1900
1901 if (nir_mode_out)
1902 *nir_mode_out = nir_mode;
1903
1904 return mode;
1905 }
1906
1907 nir_address_format
1908 vtn_mode_to_address_format(struct vtn_builder *b, enum vtn_variable_mode mode)
1909 {
1910 switch (mode) {
1911 case vtn_variable_mode_ubo:
1912 return b->options->ubo_addr_format;
1913
1914 case vtn_variable_mode_ssbo:
1915 return b->options->ssbo_addr_format;
1916
1917 case vtn_variable_mode_phys_ssbo:
1918 return b->options->phys_ssbo_addr_format;
1919
1920 case vtn_variable_mode_push_constant:
1921 return b->options->push_const_addr_format;
1922
1923 case vtn_variable_mode_workgroup:
1924 return b->options->shared_addr_format;
1925
1926 case vtn_variable_mode_cross_workgroup:
1927 return b->options->global_addr_format;
1928
1929 case vtn_variable_mode_function:
1930 if (b->physical_ptrs)
1931 return b->options->temp_addr_format;
1932 /* Fall through. */
1933
1934 case vtn_variable_mode_private:
1935 case vtn_variable_mode_uniform:
1936 case vtn_variable_mode_input:
1937 case vtn_variable_mode_output:
1938 case vtn_variable_mode_image:
1939 return nir_address_format_logical;
1940 }
1941
1942 unreachable("Invalid variable mode");
1943 }
1944
1945 nir_ssa_def *
1946 vtn_pointer_to_ssa(struct vtn_builder *b, struct vtn_pointer *ptr)
1947 {
1948 if (vtn_pointer_uses_ssa_offset(b, ptr)) {
1949 /* This pointer needs to have a pointer type with actual storage */
1950 vtn_assert(ptr->ptr_type);
1951 vtn_assert(ptr->ptr_type->type);
1952
1953 if (!ptr->offset) {
1954 /* If we don't have an offset then we must be a pointer to the variable
1955 * itself.
1956 */
1957 vtn_assert(!ptr->offset && !ptr->block_index);
1958
1959 struct vtn_access_chain chain = {
1960 .length = 0,
1961 };
1962 ptr = vtn_ssa_offset_pointer_dereference(b, ptr, &chain);
1963 }
1964
1965 vtn_assert(ptr->offset);
1966 if (ptr->block_index) {
1967 vtn_assert(ptr->mode == vtn_variable_mode_ubo ||
1968 ptr->mode == vtn_variable_mode_ssbo);
1969 return nir_vec2(&b->nb, ptr->block_index, ptr->offset);
1970 } else {
1971 vtn_assert(ptr->mode == vtn_variable_mode_workgroup);
1972 return ptr->offset;
1973 }
1974 } else {
1975 if (vtn_pointer_is_external_block(b, ptr) &&
1976 vtn_type_contains_block(b, ptr->type) &&
1977 ptr->mode != vtn_variable_mode_phys_ssbo) {
1978 /* In this case, we're looking for a block index and not an actual
1979 * deref.
1980 *
1981 * For PhysicalStorageBuffer pointers, we don't have a block index
1982 * at all because we get the pointer directly from the client. This
1983 * assumes that there will never be a SSBO binding variable using the
1984 * PhysicalStorageBuffer storage class. This assumption appears
1985 * to be correct according to the Vulkan spec because the table,
1986 * "Shader Resource and Storage Class Correspondence," the only the
1987 * Uniform storage class with BufferBlock or the StorageBuffer
1988 * storage class with Block can be used.
1989 */
1990 if (!ptr->block_index) {
1991 /* If we don't have a block_index then we must be a pointer to the
1992 * variable itself.
1993 */
1994 vtn_assert(!ptr->deref);
1995
1996 struct vtn_access_chain chain = {
1997 .length = 0,
1998 };
1999 ptr = vtn_nir_deref_pointer_dereference(b, ptr, &chain);
2000 }
2001
2002 return ptr->block_index;
2003 } else {
2004 return &vtn_pointer_to_deref(b, ptr)->dest.ssa;
2005 }
2006 }
2007 }
2008
2009 struct vtn_pointer *
2010 vtn_pointer_from_ssa(struct vtn_builder *b, nir_ssa_def *ssa,
2011 struct vtn_type *ptr_type)
2012 {
2013 vtn_assert(ptr_type->base_type == vtn_base_type_pointer);
2014
2015 struct vtn_pointer *ptr = rzalloc(b, struct vtn_pointer);
2016 struct vtn_type *without_array =
2017 vtn_type_without_array(ptr_type->deref);
2018
2019 nir_variable_mode nir_mode;
2020 ptr->mode = vtn_storage_class_to_mode(b, ptr_type->storage_class,
2021 without_array, &nir_mode);
2022 ptr->type = ptr_type->deref;
2023 ptr->ptr_type = ptr_type;
2024
2025 if (b->wa_glslang_179) {
2026 /* To work around https://github.com/KhronosGroup/glslang/issues/179 we
2027 * need to whack the mode because it creates a function parameter with
2028 * the Function storage class even though it's a pointer to a sampler.
2029 * If we don't do this, then NIR won't get rid of the deref_cast for us.
2030 */
2031 if (ptr->mode == vtn_variable_mode_function &&
2032 (ptr->type->base_type == vtn_base_type_sampler ||
2033 ptr->type->base_type == vtn_base_type_sampled_image)) {
2034 ptr->mode = vtn_variable_mode_uniform;
2035 nir_mode = nir_var_uniform;
2036 }
2037 }
2038
2039 if (vtn_pointer_uses_ssa_offset(b, ptr)) {
2040 /* This pointer type needs to have actual storage */
2041 vtn_assert(ptr_type->type);
2042 if (ptr->mode == vtn_variable_mode_ubo ||
2043 ptr->mode == vtn_variable_mode_ssbo) {
2044 vtn_assert(ssa->num_components == 2);
2045 ptr->block_index = nir_channel(&b->nb, ssa, 0);
2046 ptr->offset = nir_channel(&b->nb, ssa, 1);
2047 } else {
2048 vtn_assert(ssa->num_components == 1);
2049 ptr->block_index = NULL;
2050 ptr->offset = ssa;
2051 }
2052 } else {
2053 const struct glsl_type *deref_type = ptr_type->deref->type;
2054 if (!vtn_pointer_is_external_block(b, ptr)) {
2055 ptr->deref = nir_build_deref_cast(&b->nb, ssa, nir_mode,
2056 deref_type, ptr_type->stride);
2057 } else if (vtn_type_contains_block(b, ptr->type) &&
2058 ptr->mode != vtn_variable_mode_phys_ssbo) {
2059 /* This is a pointer to somewhere in an array of blocks, not a
2060 * pointer to somewhere inside the block. Set the block index
2061 * instead of making a cast.
2062 */
2063 ptr->block_index = ssa;
2064 } else {
2065 /* This is a pointer to something internal or a pointer inside a
2066 * block. It's just a regular cast.
2067 *
2068 * For PhysicalStorageBuffer pointers, we don't have a block index
2069 * at all because we get the pointer directly from the client. This
2070 * assumes that there will never be a SSBO binding variable using the
2071 * PhysicalStorageBuffer storage class. This assumption appears
2072 * to be correct according to the Vulkan spec because the table,
2073 * "Shader Resource and Storage Class Correspondence," the only the
2074 * Uniform storage class with BufferBlock or the StorageBuffer
2075 * storage class with Block can be used.
2076 */
2077 ptr->deref = nir_build_deref_cast(&b->nb, ssa, nir_mode,
2078 ptr_type->deref->type,
2079 ptr_type->stride);
2080 ptr->deref->dest.ssa.num_components =
2081 glsl_get_vector_elements(ptr_type->type);
2082 ptr->deref->dest.ssa.bit_size = glsl_get_bit_size(ptr_type->type);
2083 }
2084 }
2085
2086 return ptr;
2087 }
2088
2089 static bool
2090 is_per_vertex_inout(const struct vtn_variable *var, gl_shader_stage stage)
2091 {
2092 if (var->patch || !glsl_type_is_array(var->type->type))
2093 return false;
2094
2095 if (var->mode == vtn_variable_mode_input) {
2096 return stage == MESA_SHADER_TESS_CTRL ||
2097 stage == MESA_SHADER_TESS_EVAL ||
2098 stage == MESA_SHADER_GEOMETRY;
2099 }
2100
2101 if (var->mode == vtn_variable_mode_output)
2102 return stage == MESA_SHADER_TESS_CTRL;
2103
2104 return false;
2105 }
2106
2107 static void
2108 assign_missing_member_locations(struct vtn_variable *var)
2109 {
2110 unsigned length =
2111 glsl_get_length(glsl_without_array(var->type->type));
2112 int location = var->base_location;
2113
2114 for (unsigned i = 0; i < length; i++) {
2115 /* From the Vulkan spec:
2116 *
2117 * “If the structure type is a Block but without a Location, then each
2118 * of its members must have a Location decoration.”
2119 *
2120 */
2121 if (var->type->block) {
2122 assert(var->base_location != -1 ||
2123 var->var->members[i].location != -1);
2124 }
2125
2126 /* From the Vulkan spec:
2127 *
2128 * “Any member with its own Location decoration is assigned that
2129 * location. Each remaining member is assigned the location after the
2130 * immediately preceding member in declaration order.”
2131 */
2132 if (var->var->members[i].location != -1)
2133 location = var->var->members[i].location;
2134 else
2135 var->var->members[i].location = location;
2136
2137 /* Below we use type instead of interface_type, because interface_type
2138 * is only available when it is a Block. This code also supports
2139 * input/outputs that are just structs
2140 */
2141 const struct glsl_type *member_type =
2142 glsl_get_struct_field(glsl_without_array(var->type->type), i);
2143
2144 location +=
2145 glsl_count_attribute_slots(member_type,
2146 false /* is_gl_vertex_input */);
2147 }
2148 }
2149
2150
2151 static void
2152 vtn_create_variable(struct vtn_builder *b, struct vtn_value *val,
2153 struct vtn_type *ptr_type, SpvStorageClass storage_class,
2154 nir_constant *const_initializer, nir_variable *var_initializer)
2155 {
2156 vtn_assert(ptr_type->base_type == vtn_base_type_pointer);
2157 struct vtn_type *type = ptr_type->deref;
2158
2159 struct vtn_type *without_array = vtn_type_without_array(ptr_type->deref);
2160
2161 enum vtn_variable_mode mode;
2162 nir_variable_mode nir_mode;
2163 mode = vtn_storage_class_to_mode(b, storage_class, without_array, &nir_mode);
2164
2165 switch (mode) {
2166 case vtn_variable_mode_ubo:
2167 /* There's no other way to get vtn_variable_mode_ubo */
2168 vtn_assert(without_array->block);
2169 b->shader->info.num_ubos++;
2170 break;
2171 case vtn_variable_mode_ssbo:
2172 if (storage_class == SpvStorageClassStorageBuffer &&
2173 !without_array->block) {
2174 if (b->variable_pointers) {
2175 vtn_fail("Variables in the StorageBuffer storage class must "
2176 "have a struct type with the Block decoration");
2177 } else {
2178 /* If variable pointers are not present, it's still malformed
2179 * SPIR-V but we can parse it and do the right thing anyway.
2180 * Since some of the 8-bit storage tests have bugs in this are,
2181 * just make it a warning for now.
2182 */
2183 vtn_warn("Variables in the StorageBuffer storage class must "
2184 "have a struct type with the Block decoration");
2185 }
2186 }
2187 b->shader->info.num_ssbos++;
2188 break;
2189 case vtn_variable_mode_uniform:
2190 if (glsl_type_is_image(without_array->type))
2191 b->shader->info.num_images++;
2192 else if (glsl_type_is_sampler(without_array->type))
2193 b->shader->info.num_textures++;
2194 break;
2195 case vtn_variable_mode_push_constant:
2196 b->shader->num_uniforms = vtn_type_block_size(b, type);
2197 break;
2198
2199 case vtn_variable_mode_image:
2200 vtn_fail("Cannot create a variable with the Image storage class");
2201 break;
2202
2203 case vtn_variable_mode_phys_ssbo:
2204 vtn_fail("Cannot create a variable with the "
2205 "PhysicalStorageBuffer storage class");
2206 break;
2207
2208 default:
2209 /* No tallying is needed */
2210 break;
2211 }
2212
2213 struct vtn_variable *var = rzalloc(b, struct vtn_variable);
2214 var->type = type;
2215 var->mode = mode;
2216 var->base_location = -1;
2217
2218 vtn_assert(val->value_type == vtn_value_type_pointer);
2219 val->pointer = vtn_pointer_for_variable(b, var, ptr_type);
2220
2221 switch (var->mode) {
2222 case vtn_variable_mode_function:
2223 case vtn_variable_mode_private:
2224 case vtn_variable_mode_uniform:
2225 /* For these, we create the variable normally */
2226 var->var = rzalloc(b->shader, nir_variable);
2227 var->var->name = ralloc_strdup(var->var, val->name);
2228
2229 if (storage_class == SpvStorageClassAtomicCounter) {
2230 /* Need to tweak the nir type here as at vtn_handle_type we don't
2231 * have the access to storage_class, that is the one that points us
2232 * that is an atomic uint.
2233 */
2234 var->var->type = repair_atomic_type(var->type->type);
2235 } else {
2236 /* Private variables don't have any explicit layout but some layouts
2237 * may have leaked through due to type deduplication in the SPIR-V.
2238 */
2239 var->var->type = var->type->type;
2240 }
2241 var->var->data.mode = nir_mode;
2242 var->var->data.location = -1;
2243 var->var->interface_type = NULL;
2244 break;
2245
2246 case vtn_variable_mode_ubo:
2247 case vtn_variable_mode_ssbo:
2248 var->var = rzalloc(b->shader, nir_variable);
2249 var->var->name = ralloc_strdup(var->var, val->name);
2250
2251 var->var->type = var->type->type;
2252 var->var->interface_type = var->type->type;
2253
2254 var->var->data.mode = nir_mode;
2255 var->var->data.location = -1;
2256
2257 break;
2258
2259 case vtn_variable_mode_workgroup:
2260 /* Create the variable normally */
2261 var->var = rzalloc(b->shader, nir_variable);
2262 var->var->name = ralloc_strdup(var->var, val->name);
2263 /* Workgroup variables don't have any explicit layout but some
2264 * layouts may have leaked through due to type deduplication in the
2265 * SPIR-V.
2266 */
2267 var->var->type = var->type->type;
2268 var->var->data.mode = nir_var_mem_shared;
2269 break;
2270
2271 case vtn_variable_mode_input:
2272 case vtn_variable_mode_output: {
2273 /* In order to know whether or not we're a per-vertex inout, we need
2274 * the patch qualifier. This means walking the variable decorations
2275 * early before we actually create any variables. Not a big deal.
2276 *
2277 * GLSLang really likes to place decorations in the most interior
2278 * thing it possibly can. In particular, if you have a struct, it
2279 * will place the patch decorations on the struct members. This
2280 * should be handled by the variable splitting below just fine.
2281 *
2282 * If you have an array-of-struct, things get even more weird as it
2283 * will place the patch decorations on the struct even though it's
2284 * inside an array and some of the members being patch and others not
2285 * makes no sense whatsoever. Since the only sensible thing is for
2286 * it to be all or nothing, we'll call it patch if any of the members
2287 * are declared patch.
2288 */
2289 var->patch = false;
2290 vtn_foreach_decoration(b, val, var_is_patch_cb, &var->patch);
2291 if (glsl_type_is_array(var->type->type) &&
2292 glsl_type_is_struct_or_ifc(without_array->type)) {
2293 vtn_foreach_decoration(b, vtn_value(b, without_array->id,
2294 vtn_value_type_type),
2295 var_is_patch_cb, &var->patch);
2296 }
2297
2298 /* For inputs and outputs, we immediately split structures. This
2299 * is for a couple of reasons. For one, builtins may all come in
2300 * a struct and we really want those split out into separate
2301 * variables. For another, interpolation qualifiers can be
2302 * applied to members of the top-level struct ane we need to be
2303 * able to preserve that information.
2304 */
2305
2306 struct vtn_type *per_vertex_type = var->type;
2307 if (is_per_vertex_inout(var, b->shader->info.stage)) {
2308 /* In Geometry shaders (and some tessellation), inputs come
2309 * in per-vertex arrays. However, some builtins come in
2310 * non-per-vertex, hence the need for the is_array check. In
2311 * any case, there are no non-builtin arrays allowed so this
2312 * check should be sufficient.
2313 */
2314 per_vertex_type = var->type->array_element;
2315 }
2316
2317 var->var = rzalloc(b->shader, nir_variable);
2318 var->var->name = ralloc_strdup(var->var, val->name);
2319 /* In Vulkan, shader I/O variables don't have any explicit layout but
2320 * some layouts may have leaked through due to type deduplication in
2321 * the SPIR-V. We do, however, keep the layouts in the variable's
2322 * interface_type because we need offsets for XFB arrays of blocks.
2323 */
2324 var->var->type = var->type->type;
2325 var->var->data.mode = nir_mode;
2326 var->var->data.patch = var->patch;
2327
2328 /* Figure out the interface block type. */
2329 struct vtn_type *iface_type = per_vertex_type;
2330 if (var->mode == vtn_variable_mode_output &&
2331 (b->shader->info.stage == MESA_SHADER_VERTEX ||
2332 b->shader->info.stage == MESA_SHADER_TESS_EVAL ||
2333 b->shader->info.stage == MESA_SHADER_GEOMETRY)) {
2334 /* For vertex data outputs, we can end up with arrays of blocks for
2335 * transform feedback where each array element corresponds to a
2336 * different XFB output buffer.
2337 */
2338 while (iface_type->base_type == vtn_base_type_array)
2339 iface_type = iface_type->array_element;
2340 }
2341 if (iface_type->base_type == vtn_base_type_struct && iface_type->block)
2342 var->var->interface_type = iface_type->type;
2343
2344 if (per_vertex_type->base_type == vtn_base_type_struct &&
2345 per_vertex_type->block) {
2346 /* It's a struct. Set it up as per-member. */
2347 var->var->num_members = glsl_get_length(per_vertex_type->type);
2348 var->var->members = rzalloc_array(var->var, struct nir_variable_data,
2349 var->var->num_members);
2350
2351 for (unsigned i = 0; i < var->var->num_members; i++) {
2352 var->var->members[i].mode = nir_mode;
2353 var->var->members[i].patch = var->patch;
2354 var->var->members[i].location = -1;
2355 }
2356 }
2357
2358 /* For inputs and outputs, we need to grab locations and builtin
2359 * information from the per-vertex type.
2360 */
2361 vtn_foreach_decoration(b, vtn_value(b, per_vertex_type->id,
2362 vtn_value_type_type),
2363 var_decoration_cb, var);
2364 break;
2365 }
2366
2367 case vtn_variable_mode_push_constant:
2368 case vtn_variable_mode_cross_workgroup:
2369 /* These don't need actual variables. */
2370 break;
2371
2372 case vtn_variable_mode_image:
2373 case vtn_variable_mode_phys_ssbo:
2374 unreachable("Should have been caught before");
2375 }
2376
2377 /* We can only have one type of initializer */
2378 assert(!(const_initializer && var_initializer));
2379 if (const_initializer) {
2380 var->var->constant_initializer =
2381 nir_constant_clone(const_initializer, var->var);
2382 }
2383 if (var_initializer)
2384 var->var->pointer_initializer = var_initializer;
2385
2386 vtn_foreach_decoration(b, val, var_decoration_cb, var);
2387 vtn_foreach_decoration(b, val, ptr_decoration_cb, val->pointer);
2388
2389 if ((var->mode == vtn_variable_mode_input ||
2390 var->mode == vtn_variable_mode_output) &&
2391 var->var->members) {
2392 assign_missing_member_locations(var);
2393 }
2394
2395 if (var->mode == vtn_variable_mode_uniform ||
2396 var->mode == vtn_variable_mode_ubo ||
2397 var->mode == vtn_variable_mode_ssbo) {
2398 /* XXX: We still need the binding information in the nir_variable
2399 * for these. We should fix that.
2400 */
2401 var->var->data.binding = var->binding;
2402 var->var->data.explicit_binding = var->explicit_binding;
2403 var->var->data.descriptor_set = var->descriptor_set;
2404 var->var->data.index = var->input_attachment_index;
2405 var->var->data.offset = var->offset;
2406
2407 if (glsl_type_is_image(without_array->type))
2408 var->var->data.image.format = without_array->image_format;
2409 }
2410
2411 if (var->mode == vtn_variable_mode_function) {
2412 vtn_assert(var->var != NULL && var->var->members == NULL);
2413 nir_function_impl_add_variable(b->nb.impl, var->var);
2414 } else if (var->var) {
2415 nir_shader_add_variable(b->shader, var->var);
2416 } else {
2417 vtn_assert(vtn_pointer_is_external_block(b, val->pointer));
2418 }
2419 }
2420
2421 static void
2422 vtn_assert_types_equal(struct vtn_builder *b, SpvOp opcode,
2423 struct vtn_type *dst_type,
2424 struct vtn_type *src_type)
2425 {
2426 if (dst_type->id == src_type->id)
2427 return;
2428
2429 if (vtn_types_compatible(b, dst_type, src_type)) {
2430 /* Early versions of GLSLang would re-emit types unnecessarily and you
2431 * would end up with OpLoad, OpStore, or OpCopyMemory opcodes which have
2432 * mismatched source and destination types.
2433 *
2434 * https://github.com/KhronosGroup/glslang/issues/304
2435 * https://github.com/KhronosGroup/glslang/issues/307
2436 * https://bugs.freedesktop.org/show_bug.cgi?id=104338
2437 * https://bugs.freedesktop.org/show_bug.cgi?id=104424
2438 */
2439 vtn_warn("Source and destination types of %s do not have the same "
2440 "ID (but are compatible): %u vs %u",
2441 spirv_op_to_string(opcode), dst_type->id, src_type->id);
2442 return;
2443 }
2444
2445 vtn_fail("Source and destination types of %s do not match: %s vs. %s",
2446 spirv_op_to_string(opcode),
2447 glsl_get_type_name(dst_type->type),
2448 glsl_get_type_name(src_type->type));
2449 }
2450
2451 static nir_ssa_def *
2452 nir_shrink_zero_pad_vec(nir_builder *b, nir_ssa_def *val,
2453 unsigned num_components)
2454 {
2455 if (val->num_components == num_components)
2456 return val;
2457
2458 nir_ssa_def *comps[NIR_MAX_VEC_COMPONENTS];
2459 for (unsigned i = 0; i < num_components; i++) {
2460 if (i < val->num_components)
2461 comps[i] = nir_channel(b, val, i);
2462 else
2463 comps[i] = nir_imm_intN_t(b, 0, val->bit_size);
2464 }
2465 return nir_vec(b, comps, num_components);
2466 }
2467
2468 static nir_ssa_def *
2469 nir_sloppy_bitcast(nir_builder *b, nir_ssa_def *val,
2470 const struct glsl_type *type)
2471 {
2472 const unsigned num_components = glsl_get_vector_elements(type);
2473 const unsigned bit_size = glsl_get_bit_size(type);
2474
2475 /* First, zero-pad to ensure that the value is big enough that when we
2476 * bit-cast it, we don't loose anything.
2477 */
2478 if (val->bit_size < bit_size) {
2479 const unsigned src_num_components_needed =
2480 vtn_align_u32(val->num_components, bit_size / val->bit_size);
2481 val = nir_shrink_zero_pad_vec(b, val, src_num_components_needed);
2482 }
2483
2484 val = nir_bitcast_vector(b, val, bit_size);
2485
2486 return nir_shrink_zero_pad_vec(b, val, num_components);
2487 }
2488
2489 void
2490 vtn_handle_variables(struct vtn_builder *b, SpvOp opcode,
2491 const uint32_t *w, unsigned count)
2492 {
2493 switch (opcode) {
2494 case SpvOpUndef: {
2495 struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_undef);
2496 val->type = vtn_value(b, w[1], vtn_value_type_type)->type;
2497 break;
2498 }
2499
2500 case SpvOpVariable: {
2501 struct vtn_type *ptr_type = vtn_value(b, w[1], vtn_value_type_type)->type;
2502
2503 struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_pointer);
2504
2505 SpvStorageClass storage_class = w[3];
2506 nir_constant *const_initializer = NULL;
2507 nir_variable *var_initializer = NULL;
2508 if (count > 4) {
2509 struct vtn_value *init = vtn_untyped_value(b, w[4]);
2510 switch (init->value_type) {
2511 case vtn_value_type_constant:
2512 const_initializer = init->constant;
2513 break;
2514 case vtn_value_type_pointer:
2515 var_initializer = init->pointer->var->var;
2516 break;
2517 default:
2518 vtn_fail("SPIR-V variable initializer %u must be constant or pointer",
2519 w[4]);
2520 }
2521 }
2522
2523 vtn_create_variable(b, val, ptr_type, storage_class, const_initializer, var_initializer);
2524
2525 break;
2526 }
2527
2528 case SpvOpAccessChain:
2529 case SpvOpPtrAccessChain:
2530 case SpvOpInBoundsAccessChain:
2531 case SpvOpInBoundsPtrAccessChain: {
2532 struct vtn_access_chain *chain = vtn_access_chain_create(b, count - 4);
2533 enum gl_access_qualifier access = 0;
2534 chain->ptr_as_array = (opcode == SpvOpPtrAccessChain || opcode == SpvOpInBoundsPtrAccessChain);
2535
2536 unsigned idx = 0;
2537 for (int i = 4; i < count; i++) {
2538 struct vtn_value *link_val = vtn_untyped_value(b, w[i]);
2539 if (link_val->value_type == vtn_value_type_constant) {
2540 chain->link[idx].mode = vtn_access_mode_literal;
2541 chain->link[idx].id = vtn_constant_int(b, w[i]);
2542 } else {
2543 chain->link[idx].mode = vtn_access_mode_id;
2544 chain->link[idx].id = w[i];
2545 }
2546 access |= vtn_value_access(link_val);
2547 idx++;
2548 }
2549
2550 struct vtn_type *ptr_type = vtn_value(b, w[1], vtn_value_type_type)->type;
2551 struct vtn_value *base_val = vtn_untyped_value(b, w[3]);
2552 if (base_val->value_type == vtn_value_type_sampled_image) {
2553 /* This is rather insane. SPIR-V allows you to use OpSampledImage
2554 * to combine an array of images with a single sampler to get an
2555 * array of sampled images that all share the same sampler.
2556 * Fortunately, this means that we can more-or-less ignore the
2557 * sampler when crawling the access chain, but it does leave us
2558 * with this rather awkward little special-case.
2559 */
2560 struct vtn_value *val =
2561 vtn_push_value(b, w[2], vtn_value_type_sampled_image);
2562 val->sampled_image = ralloc(b, struct vtn_sampled_image);
2563 val->sampled_image->image =
2564 vtn_pointer_dereference(b, base_val->sampled_image->image, chain);
2565 val->sampled_image->sampler = base_val->sampled_image->sampler;
2566 val->sampled_image->image =
2567 vtn_decorate_pointer(b, val, val->sampled_image->image);
2568 val->sampled_image->sampler =
2569 vtn_decorate_pointer(b, val, val->sampled_image->sampler);
2570 } else {
2571 vtn_assert(base_val->value_type == vtn_value_type_pointer);
2572 struct vtn_pointer *ptr =
2573 vtn_pointer_dereference(b, base_val->pointer, chain);
2574 ptr->ptr_type = ptr_type;
2575 ptr->access |= access;
2576 vtn_push_value_pointer(b, w[2], ptr);
2577 }
2578 break;
2579 }
2580
2581 case SpvOpCopyMemory: {
2582 struct vtn_value *dest = vtn_value(b, w[1], vtn_value_type_pointer);
2583 struct vtn_value *src = vtn_value(b, w[2], vtn_value_type_pointer);
2584
2585 vtn_assert_types_equal(b, opcode, dest->type->deref, src->type->deref);
2586
2587 vtn_variable_copy(b, dest->pointer, src->pointer);
2588 break;
2589 }
2590
2591 case SpvOpLoad: {
2592 struct vtn_type *res_type =
2593 vtn_value(b, w[1], vtn_value_type_type)->type;
2594 struct vtn_value *src_val = vtn_value(b, w[3], vtn_value_type_pointer);
2595 struct vtn_pointer *src = src_val->pointer;
2596
2597 vtn_assert_types_equal(b, opcode, res_type, src_val->type->deref);
2598
2599 if (res_type->base_type == vtn_base_type_image ||
2600 res_type->base_type == vtn_base_type_sampler) {
2601 vtn_push_value_pointer(b, w[2], src);
2602 return;
2603 } else if (res_type->base_type == vtn_base_type_sampled_image) {
2604 struct vtn_value *val =
2605 vtn_push_value(b, w[2], vtn_value_type_sampled_image);
2606 val->sampled_image = ralloc(b, struct vtn_sampled_image);
2607 val->sampled_image->image = val->sampled_image->sampler =
2608 vtn_decorate_pointer(b, val, src);
2609 return;
2610 }
2611
2612 if (count > 4) {
2613 unsigned idx = 5;
2614 SpvMemoryAccessMask access = w[4];
2615 if (access & SpvMemoryAccessAlignedMask)
2616 idx++;
2617
2618 if (access & SpvMemoryAccessMakePointerVisibleMask) {
2619 SpvMemorySemanticsMask semantics =
2620 SpvMemorySemanticsMakeVisibleMask |
2621 vtn_storage_class_to_memory_semantics(src->ptr_type->storage_class);
2622
2623 SpvScope scope = vtn_constant_uint(b, w[idx]);
2624 vtn_emit_memory_barrier(b, scope, semantics);
2625 }
2626 }
2627
2628 vtn_push_ssa(b, w[2], res_type, vtn_variable_load(b, src));
2629 break;
2630 }
2631
2632 case SpvOpStore: {
2633 struct vtn_value *dest_val = vtn_value(b, w[1], vtn_value_type_pointer);
2634 struct vtn_pointer *dest = dest_val->pointer;
2635 struct vtn_value *src_val = vtn_untyped_value(b, w[2]);
2636
2637 /* OpStore requires us to actually have a storage type */
2638 vtn_fail_if(dest->type->type == NULL,
2639 "Invalid destination type for OpStore");
2640
2641 if (glsl_get_base_type(dest->type->type) == GLSL_TYPE_BOOL &&
2642 glsl_get_base_type(src_val->type->type) == GLSL_TYPE_UINT) {
2643 /* Early versions of GLSLang would use uint types for UBOs/SSBOs but
2644 * would then store them to a local variable as bool. Work around
2645 * the issue by doing an implicit conversion.
2646 *
2647 * https://github.com/KhronosGroup/glslang/issues/170
2648 * https://bugs.freedesktop.org/show_bug.cgi?id=104424
2649 */
2650 vtn_warn("OpStore of value of type OpTypeInt to a pointer to type "
2651 "OpTypeBool. Doing an implicit conversion to work around "
2652 "the problem.");
2653 struct vtn_ssa_value *bool_ssa =
2654 vtn_create_ssa_value(b, dest->type->type);
2655 bool_ssa->def = nir_i2b(&b->nb, vtn_ssa_value(b, w[2])->def);
2656 vtn_variable_store(b, bool_ssa, dest);
2657 break;
2658 }
2659
2660 vtn_assert_types_equal(b, opcode, dest_val->type->deref, src_val->type);
2661
2662 if (glsl_type_is_sampler(dest->type->type)) {
2663 if (b->wa_glslang_179) {
2664 vtn_warn("OpStore of a sampler detected. Doing on-the-fly copy "
2665 "propagation to workaround the problem.");
2666 vtn_assert(dest->var->copy_prop_sampler == NULL);
2667 struct vtn_value *v = vtn_untyped_value(b, w[2]);
2668 if (v->value_type == vtn_value_type_sampled_image) {
2669 dest->var->copy_prop_sampler = v->sampled_image->sampler;
2670 } else {
2671 vtn_assert(v->value_type == vtn_value_type_pointer);
2672 dest->var->copy_prop_sampler = v->pointer;
2673 }
2674 } else {
2675 vtn_fail("Vulkan does not allow OpStore of a sampler or image.");
2676 }
2677 break;
2678 }
2679
2680 struct vtn_ssa_value *src = vtn_ssa_value(b, w[2]);
2681 vtn_variable_store(b, src, dest);
2682
2683 if (count > 3) {
2684 unsigned idx = 4;
2685 SpvMemoryAccessMask access = w[3];
2686
2687 if (access & SpvMemoryAccessAlignedMask)
2688 idx++;
2689
2690 if (access & SpvMemoryAccessMakePointerAvailableMask) {
2691 SpvMemorySemanticsMask semantics =
2692 SpvMemorySemanticsMakeAvailableMask |
2693 vtn_storage_class_to_memory_semantics(dest->ptr_type->storage_class);
2694 SpvScope scope = vtn_constant_uint(b, w[idx]);
2695 vtn_emit_memory_barrier(b, scope, semantics);
2696 }
2697 }
2698 break;
2699 }
2700
2701 case SpvOpArrayLength: {
2702 struct vtn_pointer *ptr =
2703 vtn_value(b, w[3], vtn_value_type_pointer)->pointer;
2704 const uint32_t field = w[4];
2705
2706 vtn_fail_if(ptr->type->base_type != vtn_base_type_struct,
2707 "OpArrayLength must take a pointer to a structure type");
2708 vtn_fail_if(field != ptr->type->length - 1 ||
2709 ptr->type->members[field]->base_type != vtn_base_type_array,
2710 "OpArrayLength must reference the last memeber of the "
2711 "structure and that must be an array");
2712
2713 const uint32_t offset = ptr->type->offsets[field];
2714 const uint32_t stride = ptr->type->members[field]->stride;
2715
2716 if (!ptr->block_index) {
2717 struct vtn_access_chain chain = {
2718 .length = 0,
2719 };
2720 ptr = vtn_pointer_dereference(b, ptr, &chain);
2721 vtn_assert(ptr->block_index);
2722 }
2723
2724 nir_intrinsic_instr *instr =
2725 nir_intrinsic_instr_create(b->nb.shader,
2726 nir_intrinsic_get_buffer_size);
2727 instr->src[0] = nir_src_for_ssa(ptr->block_index);
2728 nir_ssa_dest_init(&instr->instr, &instr->dest, 1, 32, NULL);
2729 nir_builder_instr_insert(&b->nb, &instr->instr);
2730 nir_ssa_def *buf_size = &instr->dest.ssa;
2731
2732 /* array_length = max(buffer_size - offset, 0) / stride */
2733 nir_ssa_def *array_length =
2734 nir_idiv(&b->nb,
2735 nir_imax(&b->nb,
2736 nir_isub(&b->nb,
2737 buf_size,
2738 nir_imm_int(&b->nb, offset)),
2739 nir_imm_int(&b->nb, 0u)),
2740 nir_imm_int(&b->nb, stride));
2741
2742 struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_ssa);
2743 val->ssa = vtn_create_ssa_value(b, glsl_uint_type());
2744 val->ssa->def = array_length;
2745 break;
2746 }
2747
2748 case SpvOpConvertPtrToU: {
2749 struct vtn_value *u_val = vtn_push_value(b, w[2], vtn_value_type_ssa);
2750
2751 vtn_fail_if(u_val->type->base_type != vtn_base_type_vector &&
2752 u_val->type->base_type != vtn_base_type_scalar,
2753 "OpConvertPtrToU can only be used to cast to a vector or "
2754 "scalar type");
2755
2756 /* The pointer will be converted to an SSA value automatically */
2757 struct vtn_ssa_value *ptr_ssa = vtn_ssa_value(b, w[3]);
2758
2759 u_val->ssa = vtn_create_ssa_value(b, u_val->type->type);
2760 u_val->ssa->def = nir_sloppy_bitcast(&b->nb, ptr_ssa->def, u_val->type->type);
2761 u_val->ssa->access |= ptr_ssa->access;
2762 break;
2763 }
2764
2765 case SpvOpConvertUToPtr: {
2766 struct vtn_value *ptr_val =
2767 vtn_push_value(b, w[2], vtn_value_type_pointer);
2768 struct vtn_value *u_val = vtn_value(b, w[3], vtn_value_type_ssa);
2769
2770 vtn_fail_if(ptr_val->type->type == NULL,
2771 "OpConvertUToPtr can only be used on physical pointers");
2772
2773 vtn_fail_if(u_val->type->base_type != vtn_base_type_vector &&
2774 u_val->type->base_type != vtn_base_type_scalar,
2775 "OpConvertUToPtr can only be used to cast from a vector or "
2776 "scalar type");
2777
2778 nir_ssa_def *ptr_ssa = nir_sloppy_bitcast(&b->nb, u_val->ssa->def,
2779 ptr_val->type->type);
2780 ptr_val->pointer = vtn_pointer_from_ssa(b, ptr_ssa, ptr_val->type);
2781 vtn_foreach_decoration(b, ptr_val, ptr_decoration_cb, ptr_val->pointer);
2782 ptr_val->pointer->access |= u_val->ssa->access;
2783 break;
2784 }
2785
2786 case SpvOpCopyMemorySized:
2787 default:
2788 vtn_fail_with_opcode("Unhandled opcode", opcode);
2789 }
2790 }