spirv: add support for SpvBuiltInBaryCoord*
[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 if (nir_src_is_const(src->arr.index))
739 val->def = vtn_vector_extract(b, val->def,
740 nir_src_as_uint(src->arr.index));
741 else
742 val->def = vtn_vector_extract_dynamic(b, val->def, src->arr.index.ssa);
743 }
744
745 return val;
746 }
747
748 void
749 vtn_local_store(struct vtn_builder *b, struct vtn_ssa_value *src,
750 nir_deref_instr *dest, enum gl_access_qualifier access)
751 {
752 nir_deref_instr *dest_tail = get_deref_tail(dest);
753
754 if (dest_tail != dest) {
755 struct vtn_ssa_value *val = vtn_create_ssa_value(b, dest_tail->type);
756 _vtn_local_load_store(b, true, dest_tail, val, access);
757
758 if (nir_src_is_const(dest->arr.index))
759 val->def = vtn_vector_insert(b, val->def, src->def,
760 nir_src_as_uint(dest->arr.index));
761 else
762 val->def = vtn_vector_insert_dynamic(b, val->def, src->def,
763 dest->arr.index.ssa);
764 _vtn_local_load_store(b, false, dest_tail, val, access);
765 } else {
766 _vtn_local_load_store(b, false, dest_tail, src, access);
767 }
768 }
769
770 nir_ssa_def *
771 vtn_pointer_to_offset(struct vtn_builder *b, struct vtn_pointer *ptr,
772 nir_ssa_def **index_out)
773 {
774 assert(vtn_pointer_uses_ssa_offset(b, ptr));
775 if (!ptr->offset) {
776 struct vtn_access_chain chain = {
777 .length = 0,
778 };
779 ptr = vtn_ssa_offset_pointer_dereference(b, ptr, &chain);
780 }
781 *index_out = ptr->block_index;
782 return ptr->offset;
783 }
784
785 /* Tries to compute the size of an interface block based on the strides and
786 * offsets that are provided to us in the SPIR-V source.
787 */
788 static unsigned
789 vtn_type_block_size(struct vtn_builder *b, struct vtn_type *type)
790 {
791 enum glsl_base_type base_type = glsl_get_base_type(type->type);
792 switch (base_type) {
793 case GLSL_TYPE_UINT:
794 case GLSL_TYPE_INT:
795 case GLSL_TYPE_UINT16:
796 case GLSL_TYPE_INT16:
797 case GLSL_TYPE_UINT8:
798 case GLSL_TYPE_INT8:
799 case GLSL_TYPE_UINT64:
800 case GLSL_TYPE_INT64:
801 case GLSL_TYPE_FLOAT:
802 case GLSL_TYPE_FLOAT16:
803 case GLSL_TYPE_BOOL:
804 case GLSL_TYPE_DOUBLE: {
805 unsigned cols = type->row_major ? glsl_get_vector_elements(type->type) :
806 glsl_get_matrix_columns(type->type);
807 if (cols > 1) {
808 vtn_assert(type->stride > 0);
809 return type->stride * cols;
810 } else {
811 unsigned type_size = glsl_get_bit_size(type->type) / 8;
812 return glsl_get_vector_elements(type->type) * type_size;
813 }
814 }
815
816 case GLSL_TYPE_STRUCT:
817 case GLSL_TYPE_INTERFACE: {
818 unsigned size = 0;
819 unsigned num_fields = glsl_get_length(type->type);
820 for (unsigned f = 0; f < num_fields; f++) {
821 unsigned field_end = type->offsets[f] +
822 vtn_type_block_size(b, type->members[f]);
823 size = MAX2(size, field_end);
824 }
825 return size;
826 }
827
828 case GLSL_TYPE_ARRAY:
829 vtn_assert(type->stride > 0);
830 vtn_assert(glsl_get_length(type->type) > 0);
831 return type->stride * glsl_get_length(type->type);
832
833 default:
834 vtn_fail("Invalid block type");
835 return 0;
836 }
837 }
838
839 static void
840 _vtn_load_store_tail(struct vtn_builder *b, nir_intrinsic_op op, bool load,
841 nir_ssa_def *index, nir_ssa_def *offset,
842 unsigned access_offset, unsigned access_size,
843 struct vtn_ssa_value **inout, const struct glsl_type *type,
844 enum gl_access_qualifier access)
845 {
846 nir_intrinsic_instr *instr = nir_intrinsic_instr_create(b->nb.shader, op);
847 instr->num_components = glsl_get_vector_elements(type);
848
849 /* Booleans usually shouldn't show up in external memory in SPIR-V.
850 * However, they do for certain older GLSLang versions and can for shared
851 * memory when we lower access chains internally.
852 */
853 const unsigned data_bit_size = glsl_type_is_boolean(type) ? 32 :
854 glsl_get_bit_size(type);
855
856 int src = 0;
857 if (!load) {
858 nir_intrinsic_set_write_mask(instr, (1 << instr->num_components) - 1);
859 instr->src[src++] = nir_src_for_ssa((*inout)->def);
860 }
861
862 if (op == nir_intrinsic_load_push_constant) {
863 nir_intrinsic_set_base(instr, access_offset);
864 nir_intrinsic_set_range(instr, access_size);
865 }
866
867 if (op == nir_intrinsic_load_ubo ||
868 op == nir_intrinsic_load_ssbo ||
869 op == nir_intrinsic_store_ssbo) {
870 nir_intrinsic_set_access(instr, access);
871 }
872
873 /* With extensions like relaxed_block_layout, we really can't guarantee
874 * much more than scalar alignment.
875 */
876 if (op != nir_intrinsic_load_push_constant)
877 nir_intrinsic_set_align(instr, data_bit_size / 8, 0);
878
879 if (index)
880 instr->src[src++] = nir_src_for_ssa(index);
881
882 if (op == nir_intrinsic_load_push_constant) {
883 /* We need to subtract the offset from where the intrinsic will load the
884 * data. */
885 instr->src[src++] =
886 nir_src_for_ssa(nir_isub(&b->nb, offset,
887 nir_imm_int(&b->nb, access_offset)));
888 } else {
889 instr->src[src++] = nir_src_for_ssa(offset);
890 }
891
892 if (load) {
893 nir_ssa_dest_init(&instr->instr, &instr->dest,
894 instr->num_components, data_bit_size, NULL);
895 (*inout)->def = &instr->dest.ssa;
896 }
897
898 nir_builder_instr_insert(&b->nb, &instr->instr);
899
900 if (load && glsl_get_base_type(type) == GLSL_TYPE_BOOL)
901 (*inout)->def = nir_ine(&b->nb, (*inout)->def, nir_imm_int(&b->nb, 0));
902 }
903
904 static void
905 _vtn_block_load_store(struct vtn_builder *b, nir_intrinsic_op op, bool load,
906 nir_ssa_def *index, nir_ssa_def *offset,
907 unsigned access_offset, unsigned access_size,
908 struct vtn_type *type, enum gl_access_qualifier access,
909 struct vtn_ssa_value **inout)
910 {
911 if (load && *inout == NULL)
912 *inout = vtn_create_ssa_value(b, type->type);
913
914 enum glsl_base_type base_type = glsl_get_base_type(type->type);
915 switch (base_type) {
916 case GLSL_TYPE_UINT:
917 case GLSL_TYPE_INT:
918 case GLSL_TYPE_UINT16:
919 case GLSL_TYPE_INT16:
920 case GLSL_TYPE_UINT8:
921 case GLSL_TYPE_INT8:
922 case GLSL_TYPE_UINT64:
923 case GLSL_TYPE_INT64:
924 case GLSL_TYPE_FLOAT:
925 case GLSL_TYPE_FLOAT16:
926 case GLSL_TYPE_DOUBLE:
927 case GLSL_TYPE_BOOL:
928 /* This is where things get interesting. At this point, we've hit
929 * a vector, a scalar, or a matrix.
930 */
931 if (glsl_type_is_matrix(type->type)) {
932 /* Loading the whole matrix */
933 struct vtn_ssa_value *transpose;
934 unsigned num_ops, vec_width, col_stride;
935 if (type->row_major) {
936 num_ops = glsl_get_vector_elements(type->type);
937 vec_width = glsl_get_matrix_columns(type->type);
938 col_stride = type->array_element->stride;
939 if (load) {
940 const struct glsl_type *transpose_type =
941 glsl_matrix_type(base_type, vec_width, num_ops);
942 *inout = vtn_create_ssa_value(b, transpose_type);
943 } else {
944 transpose = vtn_ssa_transpose(b, *inout);
945 inout = &transpose;
946 }
947 } else {
948 num_ops = glsl_get_matrix_columns(type->type);
949 vec_width = glsl_get_vector_elements(type->type);
950 col_stride = type->stride;
951 }
952
953 for (unsigned i = 0; i < num_ops; i++) {
954 nir_ssa_def *elem_offset =
955 nir_iadd_imm(&b->nb, offset, i * col_stride);
956 _vtn_load_store_tail(b, op, load, index, elem_offset,
957 access_offset, access_size,
958 &(*inout)->elems[i],
959 glsl_vector_type(base_type, vec_width),
960 type->access | access);
961 }
962
963 if (load && type->row_major)
964 *inout = vtn_ssa_transpose(b, *inout);
965 } else {
966 unsigned elems = glsl_get_vector_elements(type->type);
967 unsigned type_size = glsl_get_bit_size(type->type) / 8;
968 if (elems == 1 || type->stride == type_size) {
969 /* This is a tightly-packed normal scalar or vector load */
970 vtn_assert(glsl_type_is_vector_or_scalar(type->type));
971 _vtn_load_store_tail(b, op, load, index, offset,
972 access_offset, access_size,
973 inout, type->type,
974 type->access | access);
975 } else {
976 /* This is a strided load. We have to load N things separately.
977 * This is the single column of a row-major matrix case.
978 */
979 vtn_assert(type->stride > type_size);
980 vtn_assert(type->stride % type_size == 0);
981
982 nir_ssa_def *per_comp[4];
983 for (unsigned i = 0; i < elems; i++) {
984 nir_ssa_def *elem_offset =
985 nir_iadd_imm(&b->nb, offset, i * type->stride);
986 struct vtn_ssa_value *comp, temp_val;
987 if (!load) {
988 temp_val.def = nir_channel(&b->nb, (*inout)->def, i);
989 temp_val.type = glsl_scalar_type(base_type);
990 }
991 comp = &temp_val;
992 _vtn_load_store_tail(b, op, load, index, elem_offset,
993 access_offset, access_size,
994 &comp, glsl_scalar_type(base_type),
995 type->access | access);
996 per_comp[i] = comp->def;
997 }
998
999 if (load) {
1000 if (*inout == NULL)
1001 *inout = vtn_create_ssa_value(b, type->type);
1002 (*inout)->def = nir_vec(&b->nb, per_comp, elems);
1003 }
1004 }
1005 }
1006 return;
1007
1008 case GLSL_TYPE_ARRAY: {
1009 unsigned elems = glsl_get_length(type->type);
1010 for (unsigned i = 0; i < elems; i++) {
1011 nir_ssa_def *elem_off =
1012 nir_iadd_imm(&b->nb, offset, i * type->stride);
1013 _vtn_block_load_store(b, op, load, index, elem_off,
1014 access_offset, access_size,
1015 type->array_element,
1016 type->array_element->access | access,
1017 &(*inout)->elems[i]);
1018 }
1019 return;
1020 }
1021
1022 case GLSL_TYPE_INTERFACE:
1023 case GLSL_TYPE_STRUCT: {
1024 unsigned elems = glsl_get_length(type->type);
1025 for (unsigned i = 0; i < elems; i++) {
1026 nir_ssa_def *elem_off =
1027 nir_iadd_imm(&b->nb, offset, type->offsets[i]);
1028 _vtn_block_load_store(b, op, load, index, elem_off,
1029 access_offset, access_size,
1030 type->members[i],
1031 type->members[i]->access | access,
1032 &(*inout)->elems[i]);
1033 }
1034 return;
1035 }
1036
1037 default:
1038 vtn_fail("Invalid block member type");
1039 }
1040 }
1041
1042 static struct vtn_ssa_value *
1043 vtn_block_load(struct vtn_builder *b, struct vtn_pointer *src)
1044 {
1045 nir_intrinsic_op op;
1046 unsigned access_offset = 0, access_size = 0;
1047 switch (src->mode) {
1048 case vtn_variable_mode_ubo:
1049 op = nir_intrinsic_load_ubo;
1050 break;
1051 case vtn_variable_mode_ssbo:
1052 op = nir_intrinsic_load_ssbo;
1053 break;
1054 case vtn_variable_mode_push_constant:
1055 op = nir_intrinsic_load_push_constant;
1056 access_size = b->shader->num_uniforms;
1057 break;
1058 case vtn_variable_mode_workgroup:
1059 op = nir_intrinsic_load_shared;
1060 break;
1061 default:
1062 vtn_fail("Invalid block variable mode");
1063 }
1064
1065 nir_ssa_def *offset, *index = NULL;
1066 offset = vtn_pointer_to_offset(b, src, &index);
1067
1068 struct vtn_ssa_value *value = NULL;
1069 _vtn_block_load_store(b, op, true, index, offset,
1070 access_offset, access_size,
1071 src->type, src->access, &value);
1072 return value;
1073 }
1074
1075 static void
1076 vtn_block_store(struct vtn_builder *b, struct vtn_ssa_value *src,
1077 struct vtn_pointer *dst)
1078 {
1079 nir_intrinsic_op op;
1080 switch (dst->mode) {
1081 case vtn_variable_mode_ssbo:
1082 op = nir_intrinsic_store_ssbo;
1083 break;
1084 case vtn_variable_mode_workgroup:
1085 op = nir_intrinsic_store_shared;
1086 break;
1087 default:
1088 vtn_fail("Invalid block variable mode");
1089 }
1090
1091 nir_ssa_def *offset, *index = NULL;
1092 offset = vtn_pointer_to_offset(b, dst, &index);
1093
1094 _vtn_block_load_store(b, op, false, index, offset,
1095 0, 0, dst->type, dst->access, &src);
1096 }
1097
1098 static void
1099 _vtn_variable_load_store(struct vtn_builder *b, bool load,
1100 struct vtn_pointer *ptr,
1101 enum gl_access_qualifier access,
1102 struct vtn_ssa_value **inout)
1103 {
1104 enum glsl_base_type base_type = glsl_get_base_type(ptr->type->type);
1105 switch (base_type) {
1106 case GLSL_TYPE_UINT:
1107 case GLSL_TYPE_INT:
1108 case GLSL_TYPE_UINT16:
1109 case GLSL_TYPE_INT16:
1110 case GLSL_TYPE_UINT8:
1111 case GLSL_TYPE_INT8:
1112 case GLSL_TYPE_UINT64:
1113 case GLSL_TYPE_INT64:
1114 case GLSL_TYPE_FLOAT:
1115 case GLSL_TYPE_FLOAT16:
1116 case GLSL_TYPE_BOOL:
1117 case GLSL_TYPE_DOUBLE:
1118 if (glsl_type_is_vector_or_scalar(ptr->type->type)) {
1119 /* We hit a vector or scalar; go ahead and emit the load[s] */
1120 nir_deref_instr *deref = vtn_pointer_to_deref(b, ptr);
1121 if (vtn_pointer_is_external_block(b, ptr)) {
1122 /* If it's external, we call nir_load/store_deref directly. The
1123 * vtn_local_load/store helpers are too clever and do magic to
1124 * avoid array derefs of vectors. That magic is both less
1125 * efficient than the direct load/store and, in the case of
1126 * stores, is broken because it creates a race condition if two
1127 * threads are writing to different components of the same vector
1128 * due to the load+insert+store it uses to emulate the array
1129 * deref.
1130 */
1131 if (load) {
1132 *inout = vtn_create_ssa_value(b, ptr->type->type);
1133 (*inout)->def = nir_load_deref_with_access(&b->nb, deref,
1134 ptr->type->access | access);
1135 } else {
1136 nir_store_deref_with_access(&b->nb, deref, (*inout)->def, ~0,
1137 ptr->type->access | access);
1138 }
1139 } else {
1140 if (load) {
1141 *inout = vtn_local_load(b, deref, ptr->type->access | access);
1142 } else {
1143 vtn_local_store(b, *inout, deref, ptr->type->access | access);
1144 }
1145 }
1146 return;
1147 }
1148 /* Fall through */
1149
1150 case GLSL_TYPE_INTERFACE:
1151 case GLSL_TYPE_ARRAY:
1152 case GLSL_TYPE_STRUCT: {
1153 unsigned elems = glsl_get_length(ptr->type->type);
1154 if (load) {
1155 vtn_assert(*inout == NULL);
1156 *inout = rzalloc(b, struct vtn_ssa_value);
1157 (*inout)->type = ptr->type->type;
1158 (*inout)->elems = rzalloc_array(b, struct vtn_ssa_value *, elems);
1159 }
1160
1161 struct vtn_access_chain chain = {
1162 .length = 1,
1163 .link = {
1164 { .mode = vtn_access_mode_literal, },
1165 }
1166 };
1167 for (unsigned i = 0; i < elems; i++) {
1168 chain.link[0].id = i;
1169 struct vtn_pointer *elem = vtn_pointer_dereference(b, ptr, &chain);
1170 _vtn_variable_load_store(b, load, elem, ptr->type->access | access,
1171 &(*inout)->elems[i]);
1172 }
1173 return;
1174 }
1175
1176 default:
1177 vtn_fail("Invalid access chain type");
1178 }
1179 }
1180
1181 struct vtn_ssa_value *
1182 vtn_variable_load(struct vtn_builder *b, struct vtn_pointer *src)
1183 {
1184 if (vtn_pointer_uses_ssa_offset(b, src)) {
1185 return vtn_block_load(b, src);
1186 } else {
1187 struct vtn_ssa_value *val = NULL;
1188 _vtn_variable_load_store(b, true, src, src->access, &val);
1189 return val;
1190 }
1191 }
1192
1193 void
1194 vtn_variable_store(struct vtn_builder *b, struct vtn_ssa_value *src,
1195 struct vtn_pointer *dest)
1196 {
1197 if (vtn_pointer_uses_ssa_offset(b, dest)) {
1198 vtn_assert(dest->mode == vtn_variable_mode_ssbo ||
1199 dest->mode == vtn_variable_mode_workgroup);
1200 vtn_block_store(b, src, dest);
1201 } else {
1202 _vtn_variable_load_store(b, false, dest, dest->access, &src);
1203 }
1204 }
1205
1206 static void
1207 _vtn_variable_copy(struct vtn_builder *b, struct vtn_pointer *dest,
1208 struct vtn_pointer *src)
1209 {
1210 vtn_assert(src->type->type == dest->type->type);
1211 enum glsl_base_type base_type = glsl_get_base_type(src->type->type);
1212 switch (base_type) {
1213 case GLSL_TYPE_UINT:
1214 case GLSL_TYPE_INT:
1215 case GLSL_TYPE_UINT16:
1216 case GLSL_TYPE_INT16:
1217 case GLSL_TYPE_UINT8:
1218 case GLSL_TYPE_INT8:
1219 case GLSL_TYPE_UINT64:
1220 case GLSL_TYPE_INT64:
1221 case GLSL_TYPE_FLOAT:
1222 case GLSL_TYPE_FLOAT16:
1223 case GLSL_TYPE_DOUBLE:
1224 case GLSL_TYPE_BOOL:
1225 /* At this point, we have a scalar, vector, or matrix so we know that
1226 * there cannot be any structure splitting still in the way. By
1227 * stopping at the matrix level rather than the vector level, we
1228 * ensure that matrices get loaded in the optimal way even if they
1229 * are storred row-major in a UBO.
1230 */
1231 vtn_variable_store(b, vtn_variable_load(b, src), dest);
1232 return;
1233
1234 case GLSL_TYPE_INTERFACE:
1235 case GLSL_TYPE_ARRAY:
1236 case GLSL_TYPE_STRUCT: {
1237 struct vtn_access_chain chain = {
1238 .length = 1,
1239 .link = {
1240 { .mode = vtn_access_mode_literal, },
1241 }
1242 };
1243 unsigned elems = glsl_get_length(src->type->type);
1244 for (unsigned i = 0; i < elems; i++) {
1245 chain.link[0].id = i;
1246 struct vtn_pointer *src_elem =
1247 vtn_pointer_dereference(b, src, &chain);
1248 struct vtn_pointer *dest_elem =
1249 vtn_pointer_dereference(b, dest, &chain);
1250
1251 _vtn_variable_copy(b, dest_elem, src_elem);
1252 }
1253 return;
1254 }
1255
1256 default:
1257 vtn_fail("Invalid access chain type");
1258 }
1259 }
1260
1261 static void
1262 vtn_variable_copy(struct vtn_builder *b, struct vtn_pointer *dest,
1263 struct vtn_pointer *src)
1264 {
1265 /* TODO: At some point, we should add a special-case for when we can
1266 * just emit a copy_var intrinsic.
1267 */
1268 _vtn_variable_copy(b, dest, src);
1269 }
1270
1271 static void
1272 set_mode_system_value(struct vtn_builder *b, nir_variable_mode *mode)
1273 {
1274 vtn_assert(*mode == nir_var_system_value || *mode == nir_var_shader_in);
1275 *mode = nir_var_system_value;
1276 }
1277
1278 static void
1279 vtn_get_builtin_location(struct vtn_builder *b,
1280 SpvBuiltIn builtin, int *location,
1281 nir_variable_mode *mode)
1282 {
1283 switch (builtin) {
1284 case SpvBuiltInPosition:
1285 *location = VARYING_SLOT_POS;
1286 break;
1287 case SpvBuiltInPointSize:
1288 *location = VARYING_SLOT_PSIZ;
1289 break;
1290 case SpvBuiltInClipDistance:
1291 *location = VARYING_SLOT_CLIP_DIST0; /* XXX CLIP_DIST1? */
1292 break;
1293 case SpvBuiltInCullDistance:
1294 *location = VARYING_SLOT_CULL_DIST0;
1295 break;
1296 case SpvBuiltInVertexId:
1297 case SpvBuiltInVertexIndex:
1298 /* The Vulkan spec defines VertexIndex to be non-zero-based and doesn't
1299 * allow VertexId. The ARB_gl_spirv spec defines VertexId to be the
1300 * same as gl_VertexID, which is non-zero-based, and removes
1301 * VertexIndex. Since they're both defined to be non-zero-based, we use
1302 * SYSTEM_VALUE_VERTEX_ID for both.
1303 */
1304 *location = SYSTEM_VALUE_VERTEX_ID;
1305 set_mode_system_value(b, mode);
1306 break;
1307 case SpvBuiltInInstanceIndex:
1308 *location = SYSTEM_VALUE_INSTANCE_INDEX;
1309 set_mode_system_value(b, mode);
1310 break;
1311 case SpvBuiltInInstanceId:
1312 *location = SYSTEM_VALUE_INSTANCE_ID;
1313 set_mode_system_value(b, mode);
1314 break;
1315 case SpvBuiltInPrimitiveId:
1316 if (b->shader->info.stage == MESA_SHADER_FRAGMENT) {
1317 vtn_assert(*mode == nir_var_shader_in);
1318 *location = VARYING_SLOT_PRIMITIVE_ID;
1319 } else if (*mode == nir_var_shader_out) {
1320 *location = VARYING_SLOT_PRIMITIVE_ID;
1321 } else {
1322 *location = SYSTEM_VALUE_PRIMITIVE_ID;
1323 set_mode_system_value(b, mode);
1324 }
1325 break;
1326 case SpvBuiltInInvocationId:
1327 *location = SYSTEM_VALUE_INVOCATION_ID;
1328 set_mode_system_value(b, mode);
1329 break;
1330 case SpvBuiltInLayer:
1331 *location = VARYING_SLOT_LAYER;
1332 if (b->shader->info.stage == MESA_SHADER_FRAGMENT)
1333 *mode = nir_var_shader_in;
1334 else if (b->shader->info.stage == MESA_SHADER_GEOMETRY)
1335 *mode = nir_var_shader_out;
1336 else if (b->options && b->options->caps.shader_viewport_index_layer &&
1337 (b->shader->info.stage == MESA_SHADER_VERTEX ||
1338 b->shader->info.stage == MESA_SHADER_TESS_EVAL))
1339 *mode = nir_var_shader_out;
1340 else
1341 vtn_fail("invalid stage for SpvBuiltInLayer");
1342 break;
1343 case SpvBuiltInViewportIndex:
1344 *location = VARYING_SLOT_VIEWPORT;
1345 if (b->shader->info.stage == MESA_SHADER_GEOMETRY)
1346 *mode = nir_var_shader_out;
1347 else if (b->options && b->options->caps.shader_viewport_index_layer &&
1348 (b->shader->info.stage == MESA_SHADER_VERTEX ||
1349 b->shader->info.stage == MESA_SHADER_TESS_EVAL))
1350 *mode = nir_var_shader_out;
1351 else if (b->shader->info.stage == MESA_SHADER_FRAGMENT)
1352 *mode = nir_var_shader_in;
1353 else
1354 vtn_fail("invalid stage for SpvBuiltInViewportIndex");
1355 break;
1356 case SpvBuiltInTessLevelOuter:
1357 *location = VARYING_SLOT_TESS_LEVEL_OUTER;
1358 break;
1359 case SpvBuiltInTessLevelInner:
1360 *location = VARYING_SLOT_TESS_LEVEL_INNER;
1361 break;
1362 case SpvBuiltInTessCoord:
1363 *location = SYSTEM_VALUE_TESS_COORD;
1364 set_mode_system_value(b, mode);
1365 break;
1366 case SpvBuiltInPatchVertices:
1367 *location = SYSTEM_VALUE_VERTICES_IN;
1368 set_mode_system_value(b, mode);
1369 break;
1370 case SpvBuiltInFragCoord:
1371 vtn_assert(*mode == nir_var_shader_in);
1372 if (b->options && b->options->frag_coord_is_sysval) {
1373 *mode = nir_var_system_value;
1374 *location = SYSTEM_VALUE_FRAG_COORD;
1375 } else {
1376 *location = VARYING_SLOT_POS;
1377 }
1378 break;
1379 case SpvBuiltInPointCoord:
1380 *location = VARYING_SLOT_PNTC;
1381 vtn_assert(*mode == nir_var_shader_in);
1382 break;
1383 case SpvBuiltInFrontFacing:
1384 *location = SYSTEM_VALUE_FRONT_FACE;
1385 set_mode_system_value(b, mode);
1386 break;
1387 case SpvBuiltInSampleId:
1388 *location = SYSTEM_VALUE_SAMPLE_ID;
1389 set_mode_system_value(b, mode);
1390 break;
1391 case SpvBuiltInSamplePosition:
1392 *location = SYSTEM_VALUE_SAMPLE_POS;
1393 set_mode_system_value(b, mode);
1394 break;
1395 case SpvBuiltInSampleMask:
1396 if (*mode == nir_var_shader_out) {
1397 *location = FRAG_RESULT_SAMPLE_MASK;
1398 } else {
1399 *location = SYSTEM_VALUE_SAMPLE_MASK_IN;
1400 set_mode_system_value(b, mode);
1401 }
1402 break;
1403 case SpvBuiltInFragDepth:
1404 *location = FRAG_RESULT_DEPTH;
1405 vtn_assert(*mode == nir_var_shader_out);
1406 break;
1407 case SpvBuiltInHelperInvocation:
1408 *location = SYSTEM_VALUE_HELPER_INVOCATION;
1409 set_mode_system_value(b, mode);
1410 break;
1411 case SpvBuiltInNumWorkgroups:
1412 *location = SYSTEM_VALUE_NUM_WORK_GROUPS;
1413 set_mode_system_value(b, mode);
1414 break;
1415 case SpvBuiltInWorkgroupSize:
1416 *location = SYSTEM_VALUE_LOCAL_GROUP_SIZE;
1417 set_mode_system_value(b, mode);
1418 break;
1419 case SpvBuiltInWorkgroupId:
1420 *location = SYSTEM_VALUE_WORK_GROUP_ID;
1421 set_mode_system_value(b, mode);
1422 break;
1423 case SpvBuiltInLocalInvocationId:
1424 *location = SYSTEM_VALUE_LOCAL_INVOCATION_ID;
1425 set_mode_system_value(b, mode);
1426 break;
1427 case SpvBuiltInLocalInvocationIndex:
1428 *location = SYSTEM_VALUE_LOCAL_INVOCATION_INDEX;
1429 set_mode_system_value(b, mode);
1430 break;
1431 case SpvBuiltInGlobalInvocationId:
1432 *location = SYSTEM_VALUE_GLOBAL_INVOCATION_ID;
1433 set_mode_system_value(b, mode);
1434 break;
1435 case SpvBuiltInGlobalLinearId:
1436 *location = SYSTEM_VALUE_GLOBAL_INVOCATION_INDEX;
1437 set_mode_system_value(b, mode);
1438 break;
1439 case SpvBuiltInBaseVertex:
1440 /* OpenGL gl_BaseVertex (SYSTEM_VALUE_BASE_VERTEX) is not the same
1441 * semantic as Vulkan BaseVertex (SYSTEM_VALUE_FIRST_VERTEX).
1442 */
1443 if (b->options->environment == NIR_SPIRV_OPENGL)
1444 *location = SYSTEM_VALUE_BASE_VERTEX;
1445 else
1446 *location = SYSTEM_VALUE_FIRST_VERTEX;
1447 set_mode_system_value(b, mode);
1448 break;
1449 case SpvBuiltInBaseInstance:
1450 *location = SYSTEM_VALUE_BASE_INSTANCE;
1451 set_mode_system_value(b, mode);
1452 break;
1453 case SpvBuiltInDrawIndex:
1454 *location = SYSTEM_VALUE_DRAW_ID;
1455 set_mode_system_value(b, mode);
1456 break;
1457 case SpvBuiltInSubgroupSize:
1458 *location = SYSTEM_VALUE_SUBGROUP_SIZE;
1459 set_mode_system_value(b, mode);
1460 break;
1461 case SpvBuiltInSubgroupId:
1462 *location = SYSTEM_VALUE_SUBGROUP_ID;
1463 set_mode_system_value(b, mode);
1464 break;
1465 case SpvBuiltInSubgroupLocalInvocationId:
1466 *location = SYSTEM_VALUE_SUBGROUP_INVOCATION;
1467 set_mode_system_value(b, mode);
1468 break;
1469 case SpvBuiltInNumSubgroups:
1470 *location = SYSTEM_VALUE_NUM_SUBGROUPS;
1471 set_mode_system_value(b, mode);
1472 break;
1473 case SpvBuiltInDeviceIndex:
1474 *location = SYSTEM_VALUE_DEVICE_INDEX;
1475 set_mode_system_value(b, mode);
1476 break;
1477 case SpvBuiltInViewIndex:
1478 *location = SYSTEM_VALUE_VIEW_INDEX;
1479 set_mode_system_value(b, mode);
1480 break;
1481 case SpvBuiltInSubgroupEqMask:
1482 *location = SYSTEM_VALUE_SUBGROUP_EQ_MASK,
1483 set_mode_system_value(b, mode);
1484 break;
1485 case SpvBuiltInSubgroupGeMask:
1486 *location = SYSTEM_VALUE_SUBGROUP_GE_MASK,
1487 set_mode_system_value(b, mode);
1488 break;
1489 case SpvBuiltInSubgroupGtMask:
1490 *location = SYSTEM_VALUE_SUBGROUP_GT_MASK,
1491 set_mode_system_value(b, mode);
1492 break;
1493 case SpvBuiltInSubgroupLeMask:
1494 *location = SYSTEM_VALUE_SUBGROUP_LE_MASK,
1495 set_mode_system_value(b, mode);
1496 break;
1497 case SpvBuiltInSubgroupLtMask:
1498 *location = SYSTEM_VALUE_SUBGROUP_LT_MASK,
1499 set_mode_system_value(b, mode);
1500 break;
1501 case SpvBuiltInFragStencilRefEXT:
1502 *location = FRAG_RESULT_STENCIL;
1503 vtn_assert(*mode == nir_var_shader_out);
1504 break;
1505 case SpvBuiltInWorkDim:
1506 *location = SYSTEM_VALUE_WORK_DIM;
1507 set_mode_system_value(b, mode);
1508 break;
1509 case SpvBuiltInGlobalSize:
1510 *location = SYSTEM_VALUE_GLOBAL_GROUP_SIZE;
1511 set_mode_system_value(b, mode);
1512 break;
1513 case SpvBuiltInBaryCoordNoPerspAMD:
1514 *location = SYSTEM_VALUE_BARYCENTRIC_LINEAR_PIXEL;
1515 set_mode_system_value(b, mode);
1516 break;
1517 case SpvBuiltInBaryCoordNoPerspCentroidAMD:
1518 *location = SYSTEM_VALUE_BARYCENTRIC_LINEAR_CENTROID;
1519 set_mode_system_value(b, mode);
1520 break;
1521 case SpvBuiltInBaryCoordNoPerspSampleAMD:
1522 *location = SYSTEM_VALUE_BARYCENTRIC_LINEAR_SAMPLE;
1523 set_mode_system_value(b, mode);
1524 break;
1525 case SpvBuiltInBaryCoordSmoothAMD:
1526 *location = SYSTEM_VALUE_BARYCENTRIC_PERSP_PIXEL;
1527 set_mode_system_value(b, mode);
1528 break;
1529 case SpvBuiltInBaryCoordSmoothCentroidAMD:
1530 *location = SYSTEM_VALUE_BARYCENTRIC_PERSP_CENTROID;
1531 set_mode_system_value(b, mode);
1532 break;
1533 case SpvBuiltInBaryCoordSmoothSampleAMD:
1534 *location = SYSTEM_VALUE_BARYCENTRIC_PERSP_SAMPLE;
1535 set_mode_system_value(b, mode);
1536 break;
1537 case SpvBuiltInBaryCoordPullModelAMD:
1538 *location = SYSTEM_VALUE_BARYCENTRIC_PULL_MODEL;
1539 set_mode_system_value(b, mode);
1540 break;
1541 default:
1542 vtn_fail("Unsupported builtin: %s (%u)",
1543 spirv_builtin_to_string(builtin), builtin);
1544 }
1545 }
1546
1547 static void
1548 apply_var_decoration(struct vtn_builder *b,
1549 struct nir_variable_data *var_data,
1550 const struct vtn_decoration *dec)
1551 {
1552 switch (dec->decoration) {
1553 case SpvDecorationRelaxedPrecision:
1554 break; /* FIXME: Do nothing with this for now. */
1555 case SpvDecorationNoPerspective:
1556 var_data->interpolation = INTERP_MODE_NOPERSPECTIVE;
1557 break;
1558 case SpvDecorationFlat:
1559 var_data->interpolation = INTERP_MODE_FLAT;
1560 break;
1561 case SpvDecorationExplicitInterpAMD:
1562 var_data->interpolation = INTERP_MODE_EXPLICIT;
1563 break;
1564 case SpvDecorationCentroid:
1565 var_data->centroid = true;
1566 break;
1567 case SpvDecorationSample:
1568 var_data->sample = true;
1569 break;
1570 case SpvDecorationInvariant:
1571 var_data->invariant = true;
1572 break;
1573 case SpvDecorationConstant:
1574 var_data->read_only = true;
1575 break;
1576 case SpvDecorationNonReadable:
1577 var_data->access |= ACCESS_NON_READABLE;
1578 break;
1579 case SpvDecorationNonWritable:
1580 var_data->read_only = true;
1581 var_data->access |= ACCESS_NON_WRITEABLE;
1582 break;
1583 case SpvDecorationRestrict:
1584 var_data->access |= ACCESS_RESTRICT;
1585 break;
1586 case SpvDecorationVolatile:
1587 var_data->access |= ACCESS_VOLATILE;
1588 break;
1589 case SpvDecorationCoherent:
1590 var_data->access |= ACCESS_COHERENT;
1591 break;
1592 case SpvDecorationComponent:
1593 var_data->location_frac = dec->operands[0];
1594 break;
1595 case SpvDecorationIndex:
1596 var_data->index = dec->operands[0];
1597 break;
1598 case SpvDecorationBuiltIn: {
1599 SpvBuiltIn builtin = dec->operands[0];
1600
1601 nir_variable_mode mode = var_data->mode;
1602 vtn_get_builtin_location(b, builtin, &var_data->location, &mode);
1603 var_data->mode = mode;
1604
1605 switch (builtin) {
1606 case SpvBuiltInTessLevelOuter:
1607 case SpvBuiltInTessLevelInner:
1608 case SpvBuiltInClipDistance:
1609 case SpvBuiltInCullDistance:
1610 var_data->compact = true;
1611 break;
1612 default:
1613 break;
1614 }
1615 }
1616
1617 case SpvDecorationSpecId:
1618 case SpvDecorationRowMajor:
1619 case SpvDecorationColMajor:
1620 case SpvDecorationMatrixStride:
1621 case SpvDecorationAliased:
1622 case SpvDecorationUniform:
1623 case SpvDecorationUniformId:
1624 case SpvDecorationLinkageAttributes:
1625 break; /* Do nothing with these here */
1626
1627 case SpvDecorationPatch:
1628 var_data->patch = true;
1629 break;
1630
1631 case SpvDecorationLocation:
1632 vtn_fail("Handled above");
1633
1634 case SpvDecorationBlock:
1635 case SpvDecorationBufferBlock:
1636 case SpvDecorationArrayStride:
1637 case SpvDecorationGLSLShared:
1638 case SpvDecorationGLSLPacked:
1639 break; /* These can apply to a type but we don't care about them */
1640
1641 case SpvDecorationBinding:
1642 case SpvDecorationDescriptorSet:
1643 case SpvDecorationNoContraction:
1644 case SpvDecorationInputAttachmentIndex:
1645 vtn_warn("Decoration not allowed for variable or structure member: %s",
1646 spirv_decoration_to_string(dec->decoration));
1647 break;
1648
1649 case SpvDecorationXfbBuffer:
1650 var_data->explicit_xfb_buffer = true;
1651 var_data->xfb.buffer = dec->operands[0];
1652 var_data->always_active_io = true;
1653 break;
1654 case SpvDecorationXfbStride:
1655 var_data->explicit_xfb_stride = true;
1656 var_data->xfb.stride = dec->operands[0];
1657 break;
1658 case SpvDecorationOffset:
1659 var_data->explicit_offset = true;
1660 var_data->offset = dec->operands[0];
1661 break;
1662
1663 case SpvDecorationStream:
1664 var_data->stream = dec->operands[0];
1665 break;
1666
1667 case SpvDecorationCPacked:
1668 case SpvDecorationSaturatedConversion:
1669 case SpvDecorationFuncParamAttr:
1670 case SpvDecorationFPRoundingMode:
1671 case SpvDecorationFPFastMathMode:
1672 case SpvDecorationAlignment:
1673 if (b->shader->info.stage != MESA_SHADER_KERNEL) {
1674 vtn_warn("Decoration only allowed for CL-style kernels: %s",
1675 spirv_decoration_to_string(dec->decoration));
1676 }
1677 break;
1678
1679 case SpvDecorationUserSemantic:
1680 /* User semantic decorations can safely be ignored by the driver. */
1681 break;
1682
1683 case SpvDecorationRestrictPointerEXT:
1684 case SpvDecorationAliasedPointerEXT:
1685 /* TODO: We should actually plumb alias information through NIR. */
1686 break;
1687
1688 default:
1689 vtn_fail_with_decoration("Unhandled decoration", dec->decoration);
1690 }
1691 }
1692
1693 static void
1694 var_is_patch_cb(struct vtn_builder *b, struct vtn_value *val, int member,
1695 const struct vtn_decoration *dec, void *out_is_patch)
1696 {
1697 if (dec->decoration == SpvDecorationPatch) {
1698 *((bool *) out_is_patch) = true;
1699 }
1700 }
1701
1702 static void
1703 var_decoration_cb(struct vtn_builder *b, struct vtn_value *val, int member,
1704 const struct vtn_decoration *dec, void *void_var)
1705 {
1706 struct vtn_variable *vtn_var = void_var;
1707
1708 /* Handle decorations that apply to a vtn_variable as a whole */
1709 switch (dec->decoration) {
1710 case SpvDecorationBinding:
1711 vtn_var->binding = dec->operands[0];
1712 vtn_var->explicit_binding = true;
1713 return;
1714 case SpvDecorationDescriptorSet:
1715 vtn_var->descriptor_set = dec->operands[0];
1716 return;
1717 case SpvDecorationInputAttachmentIndex:
1718 vtn_var->input_attachment_index = dec->operands[0];
1719 return;
1720 case SpvDecorationPatch:
1721 vtn_var->patch = true;
1722 break;
1723 case SpvDecorationOffset:
1724 vtn_var->offset = dec->operands[0];
1725 break;
1726 case SpvDecorationNonWritable:
1727 vtn_var->access |= ACCESS_NON_WRITEABLE;
1728 break;
1729 case SpvDecorationNonReadable:
1730 vtn_var->access |= ACCESS_NON_READABLE;
1731 break;
1732 case SpvDecorationVolatile:
1733 vtn_var->access |= ACCESS_VOLATILE;
1734 break;
1735 case SpvDecorationCoherent:
1736 vtn_var->access |= ACCESS_COHERENT;
1737 break;
1738 case SpvDecorationCounterBuffer:
1739 /* Counter buffer decorations can safely be ignored by the driver. */
1740 return;
1741 default:
1742 break;
1743 }
1744
1745 if (val->value_type == vtn_value_type_pointer) {
1746 assert(val->pointer->var == void_var);
1747 assert(member == -1);
1748 } else {
1749 assert(val->value_type == vtn_value_type_type);
1750 }
1751
1752 /* Location is odd. If applied to a split structure, we have to walk the
1753 * whole thing and accumulate the location. It's easier to handle as a
1754 * special case.
1755 */
1756 if (dec->decoration == SpvDecorationLocation) {
1757 unsigned location = dec->operands[0];
1758 if (b->shader->info.stage == MESA_SHADER_FRAGMENT &&
1759 vtn_var->mode == vtn_variable_mode_output) {
1760 location += FRAG_RESULT_DATA0;
1761 } else if (b->shader->info.stage == MESA_SHADER_VERTEX &&
1762 vtn_var->mode == vtn_variable_mode_input) {
1763 location += VERT_ATTRIB_GENERIC0;
1764 } else if (vtn_var->mode == vtn_variable_mode_input ||
1765 vtn_var->mode == vtn_variable_mode_output) {
1766 location += vtn_var->patch ? VARYING_SLOT_PATCH0 : VARYING_SLOT_VAR0;
1767 } else if (vtn_var->mode != vtn_variable_mode_uniform) {
1768 vtn_warn("Location must be on input, output, uniform, sampler or "
1769 "image variable");
1770 return;
1771 }
1772
1773 if (vtn_var->var->num_members == 0) {
1774 /* This handles the member and lone variable cases */
1775 vtn_var->var->data.location = location;
1776 } else {
1777 /* This handles the structure member case */
1778 assert(vtn_var->var->members);
1779
1780 if (member == -1)
1781 vtn_var->base_location = location;
1782 else
1783 vtn_var->var->members[member].location = location;
1784 }
1785
1786 return;
1787 } else {
1788 if (vtn_var->var) {
1789 if (vtn_var->var->num_members == 0) {
1790 /* We call this function on types as well as variables and not all
1791 * struct types get split so we can end up having stray member
1792 * decorations; just ignore them.
1793 */
1794 if (member == -1)
1795 apply_var_decoration(b, &vtn_var->var->data, dec);
1796 } else if (member >= 0) {
1797 /* Member decorations must come from a type */
1798 assert(val->value_type == vtn_value_type_type);
1799 apply_var_decoration(b, &vtn_var->var->members[member], dec);
1800 } else {
1801 unsigned length =
1802 glsl_get_length(glsl_without_array(vtn_var->type->type));
1803 for (unsigned i = 0; i < length; i++)
1804 apply_var_decoration(b, &vtn_var->var->members[i], dec);
1805 }
1806 } else {
1807 /* A few variables, those with external storage, have no actual
1808 * nir_variables associated with them. Fortunately, all decorations
1809 * we care about for those variables are on the type only.
1810 */
1811 vtn_assert(vtn_var->mode == vtn_variable_mode_ubo ||
1812 vtn_var->mode == vtn_variable_mode_ssbo ||
1813 vtn_var->mode == vtn_variable_mode_push_constant);
1814 }
1815 }
1816 }
1817
1818 enum vtn_variable_mode
1819 vtn_storage_class_to_mode(struct vtn_builder *b,
1820 SpvStorageClass class,
1821 struct vtn_type *interface_type,
1822 nir_variable_mode *nir_mode_out)
1823 {
1824 enum vtn_variable_mode mode;
1825 nir_variable_mode nir_mode;
1826 switch (class) {
1827 case SpvStorageClassUniform:
1828 /* Assume it's an UBO if we lack the interface_type. */
1829 if (!interface_type || interface_type->block) {
1830 mode = vtn_variable_mode_ubo;
1831 nir_mode = nir_var_mem_ubo;
1832 } else if (interface_type->buffer_block) {
1833 mode = vtn_variable_mode_ssbo;
1834 nir_mode = nir_var_mem_ssbo;
1835 } else {
1836 /* Default-block uniforms, coming from gl_spirv */
1837 mode = vtn_variable_mode_uniform;
1838 nir_mode = nir_var_uniform;
1839 }
1840 break;
1841 case SpvStorageClassStorageBuffer:
1842 mode = vtn_variable_mode_ssbo;
1843 nir_mode = nir_var_mem_ssbo;
1844 break;
1845 case SpvStorageClassPhysicalStorageBuffer:
1846 mode = vtn_variable_mode_phys_ssbo;
1847 nir_mode = nir_var_mem_global;
1848 break;
1849 case SpvStorageClassUniformConstant:
1850 if (b->shader->info.stage == MESA_SHADER_KERNEL) {
1851 if (b->options->constant_as_global) {
1852 mode = vtn_variable_mode_cross_workgroup;
1853 nir_mode = nir_var_mem_global;
1854 } else {
1855 mode = vtn_variable_mode_ubo;
1856 nir_mode = nir_var_mem_ubo;
1857 }
1858 } else {
1859 mode = vtn_variable_mode_uniform;
1860 nir_mode = nir_var_uniform;
1861 }
1862 break;
1863 case SpvStorageClassPushConstant:
1864 mode = vtn_variable_mode_push_constant;
1865 nir_mode = nir_var_uniform;
1866 break;
1867 case SpvStorageClassInput:
1868 mode = vtn_variable_mode_input;
1869 nir_mode = nir_var_shader_in;
1870 break;
1871 case SpvStorageClassOutput:
1872 mode = vtn_variable_mode_output;
1873 nir_mode = nir_var_shader_out;
1874 break;
1875 case SpvStorageClassPrivate:
1876 mode = vtn_variable_mode_private;
1877 nir_mode = nir_var_shader_temp;
1878 break;
1879 case SpvStorageClassFunction:
1880 mode = vtn_variable_mode_function;
1881 nir_mode = nir_var_function_temp;
1882 break;
1883 case SpvStorageClassWorkgroup:
1884 mode = vtn_variable_mode_workgroup;
1885 nir_mode = nir_var_mem_shared;
1886 break;
1887 case SpvStorageClassAtomicCounter:
1888 mode = vtn_variable_mode_uniform;
1889 nir_mode = nir_var_uniform;
1890 break;
1891 case SpvStorageClassCrossWorkgroup:
1892 mode = vtn_variable_mode_cross_workgroup;
1893 nir_mode = nir_var_mem_global;
1894 break;
1895 case SpvStorageClassImage:
1896 mode = vtn_variable_mode_image;
1897 nir_mode = nir_var_mem_ubo;
1898 break;
1899 case SpvStorageClassGeneric:
1900 default:
1901 vtn_fail("Unhandled variable storage class: %s (%u)",
1902 spirv_storageclass_to_string(class), class);
1903 }
1904
1905 if (nir_mode_out)
1906 *nir_mode_out = nir_mode;
1907
1908 return mode;
1909 }
1910
1911 nir_address_format
1912 vtn_mode_to_address_format(struct vtn_builder *b, enum vtn_variable_mode mode)
1913 {
1914 switch (mode) {
1915 case vtn_variable_mode_ubo:
1916 return b->options->ubo_addr_format;
1917
1918 case vtn_variable_mode_ssbo:
1919 return b->options->ssbo_addr_format;
1920
1921 case vtn_variable_mode_phys_ssbo:
1922 return b->options->phys_ssbo_addr_format;
1923
1924 case vtn_variable_mode_push_constant:
1925 return b->options->push_const_addr_format;
1926
1927 case vtn_variable_mode_workgroup:
1928 return b->options->shared_addr_format;
1929
1930 case vtn_variable_mode_cross_workgroup:
1931 return b->options->global_addr_format;
1932
1933 case vtn_variable_mode_function:
1934 if (b->physical_ptrs)
1935 return b->options->temp_addr_format;
1936 /* Fall through. */
1937
1938 case vtn_variable_mode_private:
1939 case vtn_variable_mode_uniform:
1940 case vtn_variable_mode_input:
1941 case vtn_variable_mode_output:
1942 case vtn_variable_mode_image:
1943 return nir_address_format_logical;
1944 }
1945
1946 unreachable("Invalid variable mode");
1947 }
1948
1949 nir_ssa_def *
1950 vtn_pointer_to_ssa(struct vtn_builder *b, struct vtn_pointer *ptr)
1951 {
1952 if (vtn_pointer_uses_ssa_offset(b, ptr)) {
1953 /* This pointer needs to have a pointer type with actual storage */
1954 vtn_assert(ptr->ptr_type);
1955 vtn_assert(ptr->ptr_type->type);
1956
1957 if (!ptr->offset) {
1958 /* If we don't have an offset then we must be a pointer to the variable
1959 * itself.
1960 */
1961 vtn_assert(!ptr->offset && !ptr->block_index);
1962
1963 struct vtn_access_chain chain = {
1964 .length = 0,
1965 };
1966 ptr = vtn_ssa_offset_pointer_dereference(b, ptr, &chain);
1967 }
1968
1969 vtn_assert(ptr->offset);
1970 if (ptr->block_index) {
1971 vtn_assert(ptr->mode == vtn_variable_mode_ubo ||
1972 ptr->mode == vtn_variable_mode_ssbo);
1973 return nir_vec2(&b->nb, ptr->block_index, ptr->offset);
1974 } else {
1975 vtn_assert(ptr->mode == vtn_variable_mode_workgroup);
1976 return ptr->offset;
1977 }
1978 } else {
1979 if (vtn_pointer_is_external_block(b, ptr) &&
1980 vtn_type_contains_block(b, ptr->type) &&
1981 ptr->mode != vtn_variable_mode_phys_ssbo) {
1982 /* In this case, we're looking for a block index and not an actual
1983 * deref.
1984 *
1985 * For PhysicalStorageBuffer pointers, we don't have a block index
1986 * at all because we get the pointer directly from the client. This
1987 * assumes that there will never be a SSBO binding variable using the
1988 * PhysicalStorageBuffer storage class. This assumption appears
1989 * to be correct according to the Vulkan spec because the table,
1990 * "Shader Resource and Storage Class Correspondence," the only the
1991 * Uniform storage class with BufferBlock or the StorageBuffer
1992 * storage class with Block can be used.
1993 */
1994 if (!ptr->block_index) {
1995 /* If we don't have a block_index then we must be a pointer to the
1996 * variable itself.
1997 */
1998 vtn_assert(!ptr->deref);
1999
2000 struct vtn_access_chain chain = {
2001 .length = 0,
2002 };
2003 ptr = vtn_nir_deref_pointer_dereference(b, ptr, &chain);
2004 }
2005
2006 return ptr->block_index;
2007 } else {
2008 return &vtn_pointer_to_deref(b, ptr)->dest.ssa;
2009 }
2010 }
2011 }
2012
2013 struct vtn_pointer *
2014 vtn_pointer_from_ssa(struct vtn_builder *b, nir_ssa_def *ssa,
2015 struct vtn_type *ptr_type)
2016 {
2017 vtn_assert(ptr_type->base_type == vtn_base_type_pointer);
2018
2019 struct vtn_pointer *ptr = rzalloc(b, struct vtn_pointer);
2020 struct vtn_type *without_array =
2021 vtn_type_without_array(ptr_type->deref);
2022
2023 nir_variable_mode nir_mode;
2024 ptr->mode = vtn_storage_class_to_mode(b, ptr_type->storage_class,
2025 without_array, &nir_mode);
2026 ptr->type = ptr_type->deref;
2027 ptr->ptr_type = ptr_type;
2028
2029 if (b->wa_glslang_179) {
2030 /* To work around https://github.com/KhronosGroup/glslang/issues/179 we
2031 * need to whack the mode because it creates a function parameter with
2032 * the Function storage class even though it's a pointer to a sampler.
2033 * If we don't do this, then NIR won't get rid of the deref_cast for us.
2034 */
2035 if (ptr->mode == vtn_variable_mode_function &&
2036 (ptr->type->base_type == vtn_base_type_sampler ||
2037 ptr->type->base_type == vtn_base_type_sampled_image)) {
2038 ptr->mode = vtn_variable_mode_uniform;
2039 nir_mode = nir_var_uniform;
2040 }
2041 }
2042
2043 if (vtn_pointer_uses_ssa_offset(b, ptr)) {
2044 /* This pointer type needs to have actual storage */
2045 vtn_assert(ptr_type->type);
2046 if (ptr->mode == vtn_variable_mode_ubo ||
2047 ptr->mode == vtn_variable_mode_ssbo) {
2048 vtn_assert(ssa->num_components == 2);
2049 ptr->block_index = nir_channel(&b->nb, ssa, 0);
2050 ptr->offset = nir_channel(&b->nb, ssa, 1);
2051 } else {
2052 vtn_assert(ssa->num_components == 1);
2053 ptr->block_index = NULL;
2054 ptr->offset = ssa;
2055 }
2056 } else {
2057 const struct glsl_type *deref_type = ptr_type->deref->type;
2058 if (!vtn_pointer_is_external_block(b, ptr)) {
2059 ptr->deref = nir_build_deref_cast(&b->nb, ssa, nir_mode,
2060 deref_type, ptr_type->stride);
2061 } else if (vtn_type_contains_block(b, ptr->type) &&
2062 ptr->mode != vtn_variable_mode_phys_ssbo) {
2063 /* This is a pointer to somewhere in an array of blocks, not a
2064 * pointer to somewhere inside the block. Set the block index
2065 * instead of making a cast.
2066 */
2067 ptr->block_index = ssa;
2068 } else {
2069 /* This is a pointer to something internal or a pointer inside a
2070 * block. It's just a regular cast.
2071 *
2072 * For PhysicalStorageBuffer pointers, we don't have a block index
2073 * at all because we get the pointer directly from the client. This
2074 * assumes that there will never be a SSBO binding variable using the
2075 * PhysicalStorageBuffer storage class. This assumption appears
2076 * to be correct according to the Vulkan spec because the table,
2077 * "Shader Resource and Storage Class Correspondence," the only the
2078 * Uniform storage class with BufferBlock or the StorageBuffer
2079 * storage class with Block can be used.
2080 */
2081 ptr->deref = nir_build_deref_cast(&b->nb, ssa, nir_mode,
2082 ptr_type->deref->type,
2083 ptr_type->stride);
2084 ptr->deref->dest.ssa.num_components =
2085 glsl_get_vector_elements(ptr_type->type);
2086 ptr->deref->dest.ssa.bit_size = glsl_get_bit_size(ptr_type->type);
2087 }
2088 }
2089
2090 return ptr;
2091 }
2092
2093 static bool
2094 is_per_vertex_inout(const struct vtn_variable *var, gl_shader_stage stage)
2095 {
2096 if (var->patch || !glsl_type_is_array(var->type->type))
2097 return false;
2098
2099 if (var->mode == vtn_variable_mode_input) {
2100 return stage == MESA_SHADER_TESS_CTRL ||
2101 stage == MESA_SHADER_TESS_EVAL ||
2102 stage == MESA_SHADER_GEOMETRY;
2103 }
2104
2105 if (var->mode == vtn_variable_mode_output)
2106 return stage == MESA_SHADER_TESS_CTRL;
2107
2108 return false;
2109 }
2110
2111 static void
2112 assign_missing_member_locations(struct vtn_variable *var)
2113 {
2114 unsigned length =
2115 glsl_get_length(glsl_without_array(var->type->type));
2116 int location = var->base_location;
2117
2118 for (unsigned i = 0; i < length; i++) {
2119 /* From the Vulkan spec:
2120 *
2121 * “If the structure type is a Block but without a Location, then each
2122 * of its members must have a Location decoration.”
2123 *
2124 */
2125 if (var->type->block) {
2126 assert(var->base_location != -1 ||
2127 var->var->members[i].location != -1);
2128 }
2129
2130 /* From the Vulkan spec:
2131 *
2132 * “Any member with its own Location decoration is assigned that
2133 * location. Each remaining member is assigned the location after the
2134 * immediately preceding member in declaration order.”
2135 */
2136 if (var->var->members[i].location != -1)
2137 location = var->var->members[i].location;
2138 else
2139 var->var->members[i].location = location;
2140
2141 /* Below we use type instead of interface_type, because interface_type
2142 * is only available when it is a Block. This code also supports
2143 * input/outputs that are just structs
2144 */
2145 const struct glsl_type *member_type =
2146 glsl_get_struct_field(glsl_without_array(var->type->type), i);
2147
2148 location +=
2149 glsl_count_attribute_slots(member_type,
2150 false /* is_gl_vertex_input */);
2151 }
2152 }
2153
2154
2155 static void
2156 vtn_create_variable(struct vtn_builder *b, struct vtn_value *val,
2157 struct vtn_type *ptr_type, SpvStorageClass storage_class,
2158 nir_constant *initializer)
2159 {
2160 vtn_assert(ptr_type->base_type == vtn_base_type_pointer);
2161 struct vtn_type *type = ptr_type->deref;
2162
2163 struct vtn_type *without_array = vtn_type_without_array(ptr_type->deref);
2164
2165 enum vtn_variable_mode mode;
2166 nir_variable_mode nir_mode;
2167 mode = vtn_storage_class_to_mode(b, storage_class, without_array, &nir_mode);
2168
2169 switch (mode) {
2170 case vtn_variable_mode_ubo:
2171 /* There's no other way to get vtn_variable_mode_ubo */
2172 vtn_assert(without_array->block);
2173 b->shader->info.num_ubos++;
2174 break;
2175 case vtn_variable_mode_ssbo:
2176 if (storage_class == SpvStorageClassStorageBuffer &&
2177 !without_array->block) {
2178 if (b->variable_pointers) {
2179 vtn_fail("Variables in the StorageBuffer storage class must "
2180 "have a struct type with the Block decoration");
2181 } else {
2182 /* If variable pointers are not present, it's still malformed
2183 * SPIR-V but we can parse it and do the right thing anyway.
2184 * Since some of the 8-bit storage tests have bugs in this are,
2185 * just make it a warning for now.
2186 */
2187 vtn_warn("Variables in the StorageBuffer storage class must "
2188 "have a struct type with the Block decoration");
2189 }
2190 }
2191 b->shader->info.num_ssbos++;
2192 break;
2193 case vtn_variable_mode_uniform:
2194 if (glsl_type_is_image(without_array->type))
2195 b->shader->info.num_images++;
2196 else if (glsl_type_is_sampler(without_array->type))
2197 b->shader->info.num_textures++;
2198 break;
2199 case vtn_variable_mode_push_constant:
2200 b->shader->num_uniforms = vtn_type_block_size(b, type);
2201 break;
2202
2203 case vtn_variable_mode_image:
2204 vtn_fail("Cannot create a variable with the Image storage class");
2205 break;
2206
2207 case vtn_variable_mode_phys_ssbo:
2208 vtn_fail("Cannot create a variable with the "
2209 "PhysicalStorageBuffer storage class");
2210 break;
2211
2212 default:
2213 /* No tallying is needed */
2214 break;
2215 }
2216
2217 struct vtn_variable *var = rzalloc(b, struct vtn_variable);
2218 var->type = type;
2219 var->mode = mode;
2220 var->base_location = -1;
2221
2222 vtn_assert(val->value_type == vtn_value_type_pointer);
2223 val->pointer = vtn_pointer_for_variable(b, var, ptr_type);
2224
2225 switch (var->mode) {
2226 case vtn_variable_mode_function:
2227 case vtn_variable_mode_private:
2228 case vtn_variable_mode_uniform:
2229 /* For these, we create the variable normally */
2230 var->var = rzalloc(b->shader, nir_variable);
2231 var->var->name = ralloc_strdup(var->var, val->name);
2232
2233 if (storage_class == SpvStorageClassAtomicCounter) {
2234 /* Need to tweak the nir type here as at vtn_handle_type we don't
2235 * have the access to storage_class, that is the one that points us
2236 * that is an atomic uint.
2237 */
2238 var->var->type = repair_atomic_type(var->type->type);
2239 } else {
2240 /* Private variables don't have any explicit layout but some layouts
2241 * may have leaked through due to type deduplication in the SPIR-V.
2242 */
2243 var->var->type = var->type->type;
2244 }
2245 var->var->data.mode = nir_mode;
2246 var->var->data.location = -1;
2247 var->var->interface_type = NULL;
2248 break;
2249
2250 case vtn_variable_mode_ubo:
2251 case vtn_variable_mode_ssbo:
2252 var->var = rzalloc(b->shader, nir_variable);
2253 var->var->name = ralloc_strdup(var->var, val->name);
2254
2255 var->var->type = var->type->type;
2256 var->var->interface_type = var->type->type;
2257
2258 var->var->data.mode = nir_mode;
2259 var->var->data.location = -1;
2260
2261 break;
2262
2263 case vtn_variable_mode_workgroup:
2264 /* Create the variable normally */
2265 var->var = rzalloc(b->shader, nir_variable);
2266 var->var->name = ralloc_strdup(var->var, val->name);
2267 /* Workgroup variables don't have any explicit layout but some
2268 * layouts may have leaked through due to type deduplication in the
2269 * SPIR-V.
2270 */
2271 var->var->type = var->type->type;
2272 var->var->data.mode = nir_var_mem_shared;
2273 break;
2274
2275 case vtn_variable_mode_input:
2276 case vtn_variable_mode_output: {
2277 /* In order to know whether or not we're a per-vertex inout, we need
2278 * the patch qualifier. This means walking the variable decorations
2279 * early before we actually create any variables. Not a big deal.
2280 *
2281 * GLSLang really likes to place decorations in the most interior
2282 * thing it possibly can. In particular, if you have a struct, it
2283 * will place the patch decorations on the struct members. This
2284 * should be handled by the variable splitting below just fine.
2285 *
2286 * If you have an array-of-struct, things get even more weird as it
2287 * will place the patch decorations on the struct even though it's
2288 * inside an array and some of the members being patch and others not
2289 * makes no sense whatsoever. Since the only sensible thing is for
2290 * it to be all or nothing, we'll call it patch if any of the members
2291 * are declared patch.
2292 */
2293 var->patch = false;
2294 vtn_foreach_decoration(b, val, var_is_patch_cb, &var->patch);
2295 if (glsl_type_is_array(var->type->type) &&
2296 glsl_type_is_struct_or_ifc(without_array->type)) {
2297 vtn_foreach_decoration(b, vtn_value(b, without_array->id,
2298 vtn_value_type_type),
2299 var_is_patch_cb, &var->patch);
2300 }
2301
2302 /* For inputs and outputs, we immediately split structures. This
2303 * is for a couple of reasons. For one, builtins may all come in
2304 * a struct and we really want those split out into separate
2305 * variables. For another, interpolation qualifiers can be
2306 * applied to members of the top-level struct ane we need to be
2307 * able to preserve that information.
2308 */
2309
2310 struct vtn_type *per_vertex_type = var->type;
2311 if (is_per_vertex_inout(var, b->shader->info.stage)) {
2312 /* In Geometry shaders (and some tessellation), inputs come
2313 * in per-vertex arrays. However, some builtins come in
2314 * non-per-vertex, hence the need for the is_array check. In
2315 * any case, there are no non-builtin arrays allowed so this
2316 * check should be sufficient.
2317 */
2318 per_vertex_type = var->type->array_element;
2319 }
2320
2321 var->var = rzalloc(b->shader, nir_variable);
2322 var->var->name = ralloc_strdup(var->var, val->name);
2323 /* In Vulkan, shader I/O variables don't have any explicit layout but
2324 * some layouts may have leaked through due to type deduplication in
2325 * the SPIR-V. We do, however, keep the layouts in the variable's
2326 * interface_type because we need offsets for XFB arrays of blocks.
2327 */
2328 var->var->type = var->type->type;
2329 var->var->data.mode = nir_mode;
2330 var->var->data.patch = var->patch;
2331
2332 /* Figure out the interface block type. */
2333 struct vtn_type *iface_type = per_vertex_type;
2334 if (var->mode == vtn_variable_mode_output &&
2335 (b->shader->info.stage == MESA_SHADER_VERTEX ||
2336 b->shader->info.stage == MESA_SHADER_TESS_EVAL ||
2337 b->shader->info.stage == MESA_SHADER_GEOMETRY)) {
2338 /* For vertex data outputs, we can end up with arrays of blocks for
2339 * transform feedback where each array element corresponds to a
2340 * different XFB output buffer.
2341 */
2342 while (iface_type->base_type == vtn_base_type_array)
2343 iface_type = iface_type->array_element;
2344 }
2345 if (iface_type->base_type == vtn_base_type_struct && iface_type->block)
2346 var->var->interface_type = iface_type->type;
2347
2348 if (per_vertex_type->base_type == vtn_base_type_struct &&
2349 per_vertex_type->block) {
2350 /* It's a struct. Set it up as per-member. */
2351 var->var->num_members = glsl_get_length(per_vertex_type->type);
2352 var->var->members = rzalloc_array(var->var, struct nir_variable_data,
2353 var->var->num_members);
2354
2355 for (unsigned i = 0; i < var->var->num_members; i++) {
2356 var->var->members[i].mode = nir_mode;
2357 var->var->members[i].patch = var->patch;
2358 var->var->members[i].location = -1;
2359 }
2360 }
2361
2362 /* For inputs and outputs, we need to grab locations and builtin
2363 * information from the per-vertex type.
2364 */
2365 vtn_foreach_decoration(b, vtn_value(b, per_vertex_type->id,
2366 vtn_value_type_type),
2367 var_decoration_cb, var);
2368 break;
2369 }
2370
2371 case vtn_variable_mode_push_constant:
2372 case vtn_variable_mode_cross_workgroup:
2373 /* These don't need actual variables. */
2374 break;
2375
2376 case vtn_variable_mode_image:
2377 case vtn_variable_mode_phys_ssbo:
2378 unreachable("Should have been caught before");
2379 }
2380
2381 if (initializer) {
2382 var->var->constant_initializer =
2383 nir_constant_clone(initializer, var->var);
2384 }
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 *initializer = NULL;
2507 if (count > 4)
2508 initializer = vtn_value(b, w[4], vtn_value_type_constant)->constant;
2509
2510 vtn_create_variable(b, val, ptr_type, storage_class, initializer);
2511 break;
2512 }
2513
2514 case SpvOpAccessChain:
2515 case SpvOpPtrAccessChain:
2516 case SpvOpInBoundsAccessChain:
2517 case SpvOpInBoundsPtrAccessChain: {
2518 struct vtn_access_chain *chain = vtn_access_chain_create(b, count - 4);
2519 enum gl_access_qualifier access = 0;
2520 chain->ptr_as_array = (opcode == SpvOpPtrAccessChain || opcode == SpvOpInBoundsPtrAccessChain);
2521
2522 unsigned idx = 0;
2523 for (int i = 4; i < count; i++) {
2524 struct vtn_value *link_val = vtn_untyped_value(b, w[i]);
2525 if (link_val->value_type == vtn_value_type_constant) {
2526 chain->link[idx].mode = vtn_access_mode_literal;
2527 chain->link[idx].id = vtn_constant_int(b, w[i]);
2528 } else {
2529 chain->link[idx].mode = vtn_access_mode_id;
2530 chain->link[idx].id = w[i];
2531 }
2532 access |= vtn_value_access(link_val);
2533 idx++;
2534 }
2535
2536 struct vtn_type *ptr_type = vtn_value(b, w[1], vtn_value_type_type)->type;
2537 struct vtn_value *base_val = vtn_untyped_value(b, w[3]);
2538 if (base_val->value_type == vtn_value_type_sampled_image) {
2539 /* This is rather insane. SPIR-V allows you to use OpSampledImage
2540 * to combine an array of images with a single sampler to get an
2541 * array of sampled images that all share the same sampler.
2542 * Fortunately, this means that we can more-or-less ignore the
2543 * sampler when crawling the access chain, but it does leave us
2544 * with this rather awkward little special-case.
2545 */
2546 struct vtn_value *val =
2547 vtn_push_value(b, w[2], vtn_value_type_sampled_image);
2548 val->sampled_image = ralloc(b, struct vtn_sampled_image);
2549 val->sampled_image->image =
2550 vtn_pointer_dereference(b, base_val->sampled_image->image, chain);
2551 val->sampled_image->sampler = base_val->sampled_image->sampler;
2552 val->sampled_image->image =
2553 vtn_decorate_pointer(b, val, val->sampled_image->image);
2554 val->sampled_image->sampler =
2555 vtn_decorate_pointer(b, val, val->sampled_image->sampler);
2556 } else {
2557 vtn_assert(base_val->value_type == vtn_value_type_pointer);
2558 struct vtn_pointer *ptr =
2559 vtn_pointer_dereference(b, base_val->pointer, chain);
2560 ptr->ptr_type = ptr_type;
2561 ptr->access |= access;
2562 vtn_push_value_pointer(b, w[2], ptr);
2563 }
2564 break;
2565 }
2566
2567 case SpvOpCopyMemory: {
2568 struct vtn_value *dest = vtn_value(b, w[1], vtn_value_type_pointer);
2569 struct vtn_value *src = vtn_value(b, w[2], vtn_value_type_pointer);
2570
2571 vtn_assert_types_equal(b, opcode, dest->type->deref, src->type->deref);
2572
2573 vtn_variable_copy(b, dest->pointer, src->pointer);
2574 break;
2575 }
2576
2577 case SpvOpLoad: {
2578 struct vtn_type *res_type =
2579 vtn_value(b, w[1], vtn_value_type_type)->type;
2580 struct vtn_value *src_val = vtn_value(b, w[3], vtn_value_type_pointer);
2581 struct vtn_pointer *src = src_val->pointer;
2582
2583 vtn_assert_types_equal(b, opcode, res_type, src_val->type->deref);
2584
2585 if (res_type->base_type == vtn_base_type_image ||
2586 res_type->base_type == vtn_base_type_sampler) {
2587 vtn_push_value_pointer(b, w[2], src);
2588 return;
2589 } else if (res_type->base_type == vtn_base_type_sampled_image) {
2590 struct vtn_value *val =
2591 vtn_push_value(b, w[2], vtn_value_type_sampled_image);
2592 val->sampled_image = ralloc(b, struct vtn_sampled_image);
2593 val->sampled_image->image = val->sampled_image->sampler =
2594 vtn_decorate_pointer(b, val, src);
2595 return;
2596 }
2597
2598 if (count > 4) {
2599 unsigned idx = 5;
2600 SpvMemoryAccessMask access = w[4];
2601 if (access & SpvMemoryAccessAlignedMask)
2602 idx++;
2603
2604 if (access & SpvMemoryAccessMakePointerVisibleMask) {
2605 SpvMemorySemanticsMask semantics =
2606 SpvMemorySemanticsMakeVisibleMask |
2607 vtn_storage_class_to_memory_semantics(src->ptr_type->storage_class);
2608
2609 SpvScope scope = vtn_constant_uint(b, w[idx]);
2610 vtn_emit_memory_barrier(b, scope, semantics);
2611 }
2612 }
2613
2614 vtn_push_ssa(b, w[2], res_type, vtn_variable_load(b, src));
2615 break;
2616 }
2617
2618 case SpvOpStore: {
2619 struct vtn_value *dest_val = vtn_value(b, w[1], vtn_value_type_pointer);
2620 struct vtn_pointer *dest = dest_val->pointer;
2621 struct vtn_value *src_val = vtn_untyped_value(b, w[2]);
2622
2623 /* OpStore requires us to actually have a storage type */
2624 vtn_fail_if(dest->type->type == NULL,
2625 "Invalid destination type for OpStore");
2626
2627 if (glsl_get_base_type(dest->type->type) == GLSL_TYPE_BOOL &&
2628 glsl_get_base_type(src_val->type->type) == GLSL_TYPE_UINT) {
2629 /* Early versions of GLSLang would use uint types for UBOs/SSBOs but
2630 * would then store them to a local variable as bool. Work around
2631 * the issue by doing an implicit conversion.
2632 *
2633 * https://github.com/KhronosGroup/glslang/issues/170
2634 * https://bugs.freedesktop.org/show_bug.cgi?id=104424
2635 */
2636 vtn_warn("OpStore of value of type OpTypeInt to a pointer to type "
2637 "OpTypeBool. Doing an implicit conversion to work around "
2638 "the problem.");
2639 struct vtn_ssa_value *bool_ssa =
2640 vtn_create_ssa_value(b, dest->type->type);
2641 bool_ssa->def = nir_i2b(&b->nb, vtn_ssa_value(b, w[2])->def);
2642 vtn_variable_store(b, bool_ssa, dest);
2643 break;
2644 }
2645
2646 vtn_assert_types_equal(b, opcode, dest_val->type->deref, src_val->type);
2647
2648 if (glsl_type_is_sampler(dest->type->type)) {
2649 if (b->wa_glslang_179) {
2650 vtn_warn("OpStore of a sampler detected. Doing on-the-fly copy "
2651 "propagation to workaround the problem.");
2652 vtn_assert(dest->var->copy_prop_sampler == NULL);
2653 struct vtn_value *v = vtn_untyped_value(b, w[2]);
2654 if (v->value_type == vtn_value_type_sampled_image) {
2655 dest->var->copy_prop_sampler = v->sampled_image->sampler;
2656 } else {
2657 vtn_assert(v->value_type == vtn_value_type_pointer);
2658 dest->var->copy_prop_sampler = v->pointer;
2659 }
2660 } else {
2661 vtn_fail("Vulkan does not allow OpStore of a sampler or image.");
2662 }
2663 break;
2664 }
2665
2666 struct vtn_ssa_value *src = vtn_ssa_value(b, w[2]);
2667 vtn_variable_store(b, src, dest);
2668
2669 if (count > 3) {
2670 unsigned idx = 4;
2671 SpvMemoryAccessMask access = w[3];
2672
2673 if (access & SpvMemoryAccessAlignedMask)
2674 idx++;
2675
2676 if (access & SpvMemoryAccessMakePointerAvailableMask) {
2677 SpvMemorySemanticsMask semantics =
2678 SpvMemorySemanticsMakeAvailableMask |
2679 vtn_storage_class_to_memory_semantics(dest->ptr_type->storage_class);
2680 SpvScope scope = vtn_constant_uint(b, w[idx]);
2681 vtn_emit_memory_barrier(b, scope, semantics);
2682 }
2683 }
2684 break;
2685 }
2686
2687 case SpvOpArrayLength: {
2688 struct vtn_pointer *ptr =
2689 vtn_value(b, w[3], vtn_value_type_pointer)->pointer;
2690 const uint32_t field = w[4];
2691
2692 vtn_fail_if(ptr->type->base_type != vtn_base_type_struct,
2693 "OpArrayLength must take a pointer to a structure type");
2694 vtn_fail_if(field != ptr->type->length - 1 ||
2695 ptr->type->members[field]->base_type != vtn_base_type_array,
2696 "OpArrayLength must reference the last memeber of the "
2697 "structure and that must be an array");
2698
2699 const uint32_t offset = ptr->type->offsets[field];
2700 const uint32_t stride = ptr->type->members[field]->stride;
2701
2702 if (!ptr->block_index) {
2703 struct vtn_access_chain chain = {
2704 .length = 0,
2705 };
2706 ptr = vtn_pointer_dereference(b, ptr, &chain);
2707 vtn_assert(ptr->block_index);
2708 }
2709
2710 nir_intrinsic_instr *instr =
2711 nir_intrinsic_instr_create(b->nb.shader,
2712 nir_intrinsic_get_buffer_size);
2713 instr->src[0] = nir_src_for_ssa(ptr->block_index);
2714 nir_ssa_dest_init(&instr->instr, &instr->dest, 1, 32, NULL);
2715 nir_builder_instr_insert(&b->nb, &instr->instr);
2716 nir_ssa_def *buf_size = &instr->dest.ssa;
2717
2718 /* array_length = max(buffer_size - offset, 0) / stride */
2719 nir_ssa_def *array_length =
2720 nir_idiv(&b->nb,
2721 nir_imax(&b->nb,
2722 nir_isub(&b->nb,
2723 buf_size,
2724 nir_imm_int(&b->nb, offset)),
2725 nir_imm_int(&b->nb, 0u)),
2726 nir_imm_int(&b->nb, stride));
2727
2728 struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_ssa);
2729 val->ssa = vtn_create_ssa_value(b, glsl_uint_type());
2730 val->ssa->def = array_length;
2731 break;
2732 }
2733
2734 case SpvOpConvertPtrToU: {
2735 struct vtn_value *u_val = vtn_push_value(b, w[2], vtn_value_type_ssa);
2736
2737 vtn_fail_if(u_val->type->base_type != vtn_base_type_vector &&
2738 u_val->type->base_type != vtn_base_type_scalar,
2739 "OpConvertPtrToU can only be used to cast to a vector or "
2740 "scalar type");
2741
2742 /* The pointer will be converted to an SSA value automatically */
2743 struct vtn_ssa_value *ptr_ssa = vtn_ssa_value(b, w[3]);
2744
2745 u_val->ssa = vtn_create_ssa_value(b, u_val->type->type);
2746 u_val->ssa->def = nir_sloppy_bitcast(&b->nb, ptr_ssa->def, u_val->type->type);
2747 u_val->ssa->access |= ptr_ssa->access;
2748 break;
2749 }
2750
2751 case SpvOpConvertUToPtr: {
2752 struct vtn_value *ptr_val =
2753 vtn_push_value(b, w[2], vtn_value_type_pointer);
2754 struct vtn_value *u_val = vtn_value(b, w[3], vtn_value_type_ssa);
2755
2756 vtn_fail_if(ptr_val->type->type == NULL,
2757 "OpConvertUToPtr can only be used on physical pointers");
2758
2759 vtn_fail_if(u_val->type->base_type != vtn_base_type_vector &&
2760 u_val->type->base_type != vtn_base_type_scalar,
2761 "OpConvertUToPtr can only be used to cast from a vector or "
2762 "scalar type");
2763
2764 nir_ssa_def *ptr_ssa = nir_sloppy_bitcast(&b->nb, u_val->ssa->def,
2765 ptr_val->type->type);
2766 ptr_val->pointer = vtn_pointer_from_ssa(b, ptr_ssa, ptr_val->type);
2767 vtn_foreach_decoration(b, ptr_val, ptr_decoration_cb, ptr_val->pointer);
2768 ptr_val->pointer->access |= u_val->ssa->access;
2769 break;
2770 }
2771
2772 case SpvOpCopyMemorySized:
2773 default:
2774 vtn_fail_with_opcode("Unhandled opcode", opcode);
2775 }
2776 }