Revert "nir: Add an option for lowering TessLevelInner/Outer to vecs"
[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_ctx)
82 {
83 switch (dec->decoration) {
84 case SpvDecorationNonUniformEXT:
85 if (val->value_type == vtn_value_type_ssa) {
86 val->ssa->access |= ACCESS_NON_UNIFORM;
87 } else if (val->value_type == vtn_value_type_pointer) {
88 val->pointer->access |= ACCESS_NON_UNIFORM;
89 } else if (val->value_type == vtn_value_type_sampled_image) {
90 val->sampled_image->image->access |= ACCESS_NON_UNIFORM;
91 } else if (val->value_type == vtn_value_type_image_pointer) {
92 val->image->image->access |= ACCESS_NON_UNIFORM;
93 }
94 break;
95
96 default:
97 break;
98 }
99 }
100
101 struct vtn_value *
102 vtn_push_ssa(struct vtn_builder *b, uint32_t value_id,
103 struct vtn_type *type, struct vtn_ssa_value *ssa)
104 {
105 struct vtn_value *val;
106 if (type->base_type == vtn_base_type_pointer) {
107 val = vtn_push_value_pointer(b, value_id, vtn_pointer_from_ssa(b, ssa->def, type));
108 } else {
109 val = vtn_push_value(b, value_id, vtn_value_type_ssa);
110 val->ssa = ssa;
111 vtn_foreach_decoration(b, val, ssa_decoration_cb, NULL);
112 }
113 return val;
114 }
115
116 void
117 vtn_copy_value(struct vtn_builder *b, uint32_t src_value_id,
118 uint32_t dst_value_id)
119 {
120 struct vtn_value *src = vtn_untyped_value(b, src_value_id);
121 struct vtn_value *dst = vtn_push_value(b, dst_value_id, src->value_type);
122 struct vtn_value src_copy = *src;
123
124 vtn_fail_if(dst->type->id != src->type->id,
125 "Result Type must equal Operand type");
126
127 src_copy.name = dst->name;
128 src_copy.decoration = dst->decoration;
129 src_copy.type = dst->type;
130 *dst = src_copy;
131
132 vtn_foreach_decoration(b, dst, ssa_decoration_cb, NULL);
133 }
134
135 static struct vtn_access_chain *
136 vtn_access_chain_create(struct vtn_builder *b, unsigned length)
137 {
138 struct vtn_access_chain *chain;
139
140 /* Subtract 1 from the length since there's already one built in */
141 size_t size = sizeof(*chain) +
142 (MAX2(length, 1) - 1) * sizeof(chain->link[0]);
143 chain = rzalloc_size(b, size);
144 chain->length = length;
145
146 return chain;
147 }
148
149 bool
150 vtn_mode_uses_ssa_offset(struct vtn_builder *b,
151 enum vtn_variable_mode mode)
152 {
153 return ((mode == vtn_variable_mode_ubo ||
154 mode == vtn_variable_mode_ssbo) &&
155 b->options->lower_ubo_ssbo_access_to_offsets) ||
156 mode == vtn_variable_mode_push_constant;
157 }
158
159 static bool
160 vtn_pointer_is_external_block(struct vtn_builder *b,
161 struct vtn_pointer *ptr)
162 {
163 return ptr->mode == vtn_variable_mode_ssbo ||
164 ptr->mode == vtn_variable_mode_ubo ||
165 ptr->mode == vtn_variable_mode_phys_ssbo ||
166 ptr->mode == vtn_variable_mode_push_constant;
167 }
168
169 static nir_ssa_def *
170 vtn_access_link_as_ssa(struct vtn_builder *b, struct vtn_access_link link,
171 unsigned stride, unsigned bit_size)
172 {
173 vtn_assert(stride > 0);
174 if (link.mode == vtn_access_mode_literal) {
175 return nir_imm_intN_t(&b->nb, link.id * stride, bit_size);
176 } else {
177 nir_ssa_def *ssa = vtn_ssa_value(b, link.id)->def;
178 if (ssa->bit_size != bit_size)
179 ssa = nir_i2i(&b->nb, ssa, bit_size);
180 return nir_imul_imm(&b->nb, ssa, stride);
181 }
182 }
183
184 static VkDescriptorType
185 vk_desc_type_for_mode(struct vtn_builder *b, enum vtn_variable_mode mode)
186 {
187 switch (mode) {
188 case vtn_variable_mode_ubo:
189 return VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER;
190 case vtn_variable_mode_ssbo:
191 return VK_DESCRIPTOR_TYPE_STORAGE_BUFFER;
192 default:
193 vtn_fail("Invalid mode for vulkan_resource_index");
194 }
195 }
196
197 static nir_ssa_def *
198 vtn_variable_resource_index(struct vtn_builder *b, struct vtn_variable *var,
199 nir_ssa_def *desc_array_index)
200 {
201 vtn_assert(b->options->environment == NIR_SPIRV_VULKAN);
202
203 if (!desc_array_index) {
204 vtn_assert(glsl_type_is_struct_or_ifc(var->type->type));
205 desc_array_index = nir_imm_int(&b->nb, 0);
206 }
207
208 nir_intrinsic_instr *instr =
209 nir_intrinsic_instr_create(b->nb.shader,
210 nir_intrinsic_vulkan_resource_index);
211 instr->src[0] = nir_src_for_ssa(desc_array_index);
212 nir_intrinsic_set_desc_set(instr, var->descriptor_set);
213 nir_intrinsic_set_binding(instr, var->binding);
214 nir_intrinsic_set_desc_type(instr, vk_desc_type_for_mode(b, var->mode));
215
216 vtn_fail_if(var->mode != vtn_variable_mode_ubo &&
217 var->mode != vtn_variable_mode_ssbo,
218 "Invalid mode for vulkan_resource_index");
219
220 nir_address_format addr_format = vtn_mode_to_address_format(b, var->mode);
221 const struct glsl_type *index_type =
222 b->options->lower_ubo_ssbo_access_to_offsets ?
223 glsl_uint_type() : nir_address_format_to_glsl_type(addr_format);
224
225 instr->num_components = glsl_get_vector_elements(index_type);
226 nir_ssa_dest_init(&instr->instr, &instr->dest, instr->num_components,
227 glsl_get_bit_size(index_type), NULL);
228 nir_builder_instr_insert(&b->nb, &instr->instr);
229
230 return &instr->dest.ssa;
231 }
232
233 static nir_ssa_def *
234 vtn_resource_reindex(struct vtn_builder *b, enum vtn_variable_mode mode,
235 nir_ssa_def *base_index, nir_ssa_def *offset_index)
236 {
237 vtn_assert(b->options->environment == NIR_SPIRV_VULKAN);
238
239 nir_intrinsic_instr *instr =
240 nir_intrinsic_instr_create(b->nb.shader,
241 nir_intrinsic_vulkan_resource_reindex);
242 instr->src[0] = nir_src_for_ssa(base_index);
243 instr->src[1] = nir_src_for_ssa(offset_index);
244 nir_intrinsic_set_desc_type(instr, vk_desc_type_for_mode(b, mode));
245
246 vtn_fail_if(mode != vtn_variable_mode_ubo && mode != vtn_variable_mode_ssbo,
247 "Invalid mode for vulkan_resource_reindex");
248
249 nir_address_format addr_format = vtn_mode_to_address_format(b, mode);
250 const struct glsl_type *index_type =
251 b->options->lower_ubo_ssbo_access_to_offsets ?
252 glsl_uint_type() : nir_address_format_to_glsl_type(addr_format);
253
254 instr->num_components = glsl_get_vector_elements(index_type);
255 nir_ssa_dest_init(&instr->instr, &instr->dest, instr->num_components,
256 glsl_get_bit_size(index_type), NULL);
257 nir_builder_instr_insert(&b->nb, &instr->instr);
258
259 return &instr->dest.ssa;
260 }
261
262 static nir_ssa_def *
263 vtn_descriptor_load(struct vtn_builder *b, enum vtn_variable_mode mode,
264 nir_ssa_def *desc_index)
265 {
266 vtn_assert(b->options->environment == NIR_SPIRV_VULKAN);
267
268 nir_intrinsic_instr *desc_load =
269 nir_intrinsic_instr_create(b->nb.shader,
270 nir_intrinsic_load_vulkan_descriptor);
271 desc_load->src[0] = nir_src_for_ssa(desc_index);
272 nir_intrinsic_set_desc_type(desc_load, vk_desc_type_for_mode(b, mode));
273
274 vtn_fail_if(mode != vtn_variable_mode_ubo && mode != vtn_variable_mode_ssbo,
275 "Invalid mode for load_vulkan_descriptor");
276
277 nir_address_format addr_format = vtn_mode_to_address_format(b, mode);
278 const struct glsl_type *ptr_type =
279 nir_address_format_to_glsl_type(addr_format);
280
281 desc_load->num_components = glsl_get_vector_elements(ptr_type);
282 nir_ssa_dest_init(&desc_load->instr, &desc_load->dest,
283 desc_load->num_components,
284 glsl_get_bit_size(ptr_type), NULL);
285 nir_builder_instr_insert(&b->nb, &desc_load->instr);
286
287 return &desc_load->dest.ssa;
288 }
289
290 /* Dereference the given base pointer by the access chain */
291 static struct vtn_pointer *
292 vtn_nir_deref_pointer_dereference(struct vtn_builder *b,
293 struct vtn_pointer *base,
294 struct vtn_access_chain *deref_chain)
295 {
296 struct vtn_type *type = base->type;
297 enum gl_access_qualifier access = base->access | deref_chain->access;
298 unsigned idx = 0;
299
300 nir_deref_instr *tail;
301 if (base->deref) {
302 tail = base->deref;
303 } else if (b->options->environment == NIR_SPIRV_VULKAN &&
304 vtn_pointer_is_external_block(b, base)) {
305 nir_ssa_def *block_index = base->block_index;
306
307 /* We dereferencing an external block pointer. Correctness of this
308 * operation relies on one particular line in the SPIR-V spec, section
309 * entitled "Validation Rules for Shader Capabilities":
310 *
311 * "Block and BufferBlock decorations cannot decorate a structure
312 * type that is nested at any level inside another structure type
313 * decorated with Block or BufferBlock."
314 *
315 * This means that we can detect the point where we cross over from
316 * descriptor indexing to buffer indexing by looking for the block
317 * decorated struct type. Anything before the block decorated struct
318 * type is a descriptor indexing operation and anything after the block
319 * decorated struct is a buffer offset operation.
320 */
321
322 /* Figure out the descriptor array index if any
323 *
324 * Some of the Vulkan CTS tests with hand-rolled SPIR-V have been known
325 * to forget the Block or BufferBlock decoration from time to time.
326 * It's more robust if we check for both !block_index and for the type
327 * to contain a block. This way there's a decent chance that arrays of
328 * UBOs/SSBOs will work correctly even if variable pointers are
329 * completley toast.
330 */
331 nir_ssa_def *desc_arr_idx = NULL;
332 if (!block_index || vtn_type_contains_block(b, type)) {
333 /* If our type contains a block, then we're still outside the block
334 * and we need to process enough levels of dereferences to get inside
335 * of it.
336 */
337 if (deref_chain->ptr_as_array) {
338 unsigned aoa_size = glsl_get_aoa_size(type->type);
339 desc_arr_idx = vtn_access_link_as_ssa(b, deref_chain->link[idx],
340 MAX2(aoa_size, 1), 32);
341 idx++;
342 }
343
344 for (; idx < deref_chain->length; idx++) {
345 if (type->base_type != vtn_base_type_array) {
346 vtn_assert(type->base_type == vtn_base_type_struct);
347 break;
348 }
349
350 unsigned aoa_size = glsl_get_aoa_size(type->array_element->type);
351 nir_ssa_def *arr_offset =
352 vtn_access_link_as_ssa(b, deref_chain->link[idx],
353 MAX2(aoa_size, 1), 32);
354 if (desc_arr_idx)
355 desc_arr_idx = nir_iadd(&b->nb, desc_arr_idx, arr_offset);
356 else
357 desc_arr_idx = arr_offset;
358
359 type = type->array_element;
360 access |= type->access;
361 }
362 }
363
364 if (!block_index) {
365 vtn_assert(base->var && base->type);
366 block_index = vtn_variable_resource_index(b, base->var, desc_arr_idx);
367 } else if (desc_arr_idx) {
368 block_index = vtn_resource_reindex(b, base->mode,
369 block_index, desc_arr_idx);
370 }
371
372 if (idx == deref_chain->length) {
373 /* The entire deref was consumed in finding the block index. Return
374 * a pointer which just has a block index and a later access chain
375 * will dereference deeper.
376 */
377 struct vtn_pointer *ptr = rzalloc(b, struct vtn_pointer);
378 ptr->mode = base->mode;
379 ptr->type = type;
380 ptr->block_index = block_index;
381 ptr->access = access;
382 return ptr;
383 }
384
385 /* If we got here, there's more access chain to handle and we have the
386 * final block index. Insert a descriptor load and cast to a deref to
387 * start the deref chain.
388 */
389 nir_ssa_def *desc = vtn_descriptor_load(b, base->mode, block_index);
390
391 assert(base->mode == vtn_variable_mode_ssbo ||
392 base->mode == vtn_variable_mode_ubo);
393 nir_variable_mode nir_mode =
394 base->mode == vtn_variable_mode_ssbo ? nir_var_mem_ssbo : nir_var_mem_ubo;
395
396 tail = nir_build_deref_cast(&b->nb, desc, nir_mode, type->type,
397 base->ptr_type->stride);
398 } else {
399 assert(base->var && base->var->var);
400 tail = nir_build_deref_var(&b->nb, base->var->var);
401 if (base->ptr_type && base->ptr_type->type) {
402 tail->dest.ssa.num_components =
403 glsl_get_vector_elements(base->ptr_type->type);
404 tail->dest.ssa.bit_size = glsl_get_bit_size(base->ptr_type->type);
405 }
406 }
407
408 if (idx == 0 && deref_chain->ptr_as_array) {
409 /* We start with a deref cast to get the stride. Hopefully, we'll be
410 * able to delete that cast eventually.
411 */
412 tail = nir_build_deref_cast(&b->nb, &tail->dest.ssa, tail->mode,
413 tail->type, base->ptr_type->stride);
414
415 nir_ssa_def *index = vtn_access_link_as_ssa(b, deref_chain->link[0], 1,
416 tail->dest.ssa.bit_size);
417 tail = nir_build_deref_ptr_as_array(&b->nb, tail, index);
418 idx++;
419 }
420
421 for (; idx < deref_chain->length; idx++) {
422 if (glsl_type_is_struct_or_ifc(type->type)) {
423 vtn_assert(deref_chain->link[idx].mode == vtn_access_mode_literal);
424 unsigned field = deref_chain->link[idx].id;
425 tail = nir_build_deref_struct(&b->nb, tail, field);
426 type = type->members[field];
427 } else {
428 nir_ssa_def *arr_index =
429 vtn_access_link_as_ssa(b, deref_chain->link[idx], 1,
430 tail->dest.ssa.bit_size);
431 tail = nir_build_deref_array(&b->nb, tail, arr_index);
432 type = type->array_element;
433 }
434
435 access |= type->access;
436 }
437
438 struct vtn_pointer *ptr = rzalloc(b, struct vtn_pointer);
439 ptr->mode = base->mode;
440 ptr->type = type;
441 ptr->var = base->var;
442 ptr->deref = tail;
443 ptr->access = access;
444
445 return ptr;
446 }
447
448 static struct vtn_pointer *
449 vtn_ssa_offset_pointer_dereference(struct vtn_builder *b,
450 struct vtn_pointer *base,
451 struct vtn_access_chain *deref_chain)
452 {
453 nir_ssa_def *block_index = base->block_index;
454 nir_ssa_def *offset = base->offset;
455 struct vtn_type *type = base->type;
456 enum gl_access_qualifier access = base->access;
457
458 unsigned idx = 0;
459 if (base->mode == vtn_variable_mode_ubo ||
460 base->mode == vtn_variable_mode_ssbo) {
461 if (!block_index) {
462 vtn_assert(base->var && base->type);
463 nir_ssa_def *desc_arr_idx;
464 if (glsl_type_is_array(type->type)) {
465 if (deref_chain->length >= 1) {
466 desc_arr_idx =
467 vtn_access_link_as_ssa(b, deref_chain->link[0], 1, 32);
468 idx++;
469 /* This consumes a level of type */
470 type = type->array_element;
471 access |= type->access;
472 } else {
473 /* This is annoying. We've been asked for a pointer to the
474 * array of UBOs/SSBOs and not a specifc buffer. Return a
475 * pointer with a descriptor index of 0 and we'll have to do
476 * a reindex later to adjust it to the right thing.
477 */
478 desc_arr_idx = nir_imm_int(&b->nb, 0);
479 }
480 } else if (deref_chain->ptr_as_array) {
481 /* You can't have a zero-length OpPtrAccessChain */
482 vtn_assert(deref_chain->length >= 1);
483 desc_arr_idx = vtn_access_link_as_ssa(b, deref_chain->link[0], 1, 32);
484 } else {
485 /* We have a regular non-array SSBO. */
486 desc_arr_idx = NULL;
487 }
488 block_index = vtn_variable_resource_index(b, base->var, desc_arr_idx);
489 } else if (deref_chain->ptr_as_array &&
490 type->base_type == vtn_base_type_struct && type->block) {
491 /* We are doing an OpPtrAccessChain on a pointer to a struct that is
492 * decorated block. This is an interesting corner in the SPIR-V
493 * spec. One interpretation would be that they client is clearly
494 * trying to treat that block as if it's an implicit array of blocks
495 * repeated in the buffer. However, the SPIR-V spec for the
496 * OpPtrAccessChain says:
497 *
498 * "Base is treated as the address of the first element of an
499 * array, and the Element element’s address is computed to be the
500 * base for the Indexes, as per OpAccessChain."
501 *
502 * Taken literally, that would mean that your struct type is supposed
503 * to be treated as an array of such a struct and, since it's
504 * decorated block, that means an array of blocks which corresponds
505 * to an array descriptor. Therefore, we need to do a reindex
506 * operation to add the index from the first link in the access chain
507 * to the index we recieved.
508 *
509 * The downside to this interpretation (there always is one) is that
510 * this might be somewhat surprising behavior to apps if they expect
511 * the implicit array behavior described above.
512 */
513 vtn_assert(deref_chain->length >= 1);
514 nir_ssa_def *offset_index =
515 vtn_access_link_as_ssa(b, deref_chain->link[0], 1, 32);
516 idx++;
517
518 block_index = vtn_resource_reindex(b, base->mode,
519 block_index, offset_index);
520 }
521 }
522
523 if (!offset) {
524 if (base->mode == vtn_variable_mode_workgroup) {
525 /* SLM doesn't need nor have a block index */
526 vtn_assert(!block_index);
527
528 /* We need the variable for the base offset */
529 vtn_assert(base->var);
530
531 /* We need ptr_type for size and alignment */
532 vtn_assert(base->ptr_type);
533
534 /* Assign location on first use so that we don't end up bloating SLM
535 * address space for variables which are never statically used.
536 */
537 if (base->var->shared_location < 0) {
538 vtn_assert(base->ptr_type->length > 0 && base->ptr_type->align > 0);
539 b->shader->num_shared = vtn_align_u32(b->shader->num_shared,
540 base->ptr_type->align);
541 base->var->shared_location = b->shader->num_shared;
542 b->shader->num_shared += base->ptr_type->length;
543 }
544
545 offset = nir_imm_int(&b->nb, base->var->shared_location);
546 } else if (base->mode == vtn_variable_mode_push_constant) {
547 /* Push constants neither need nor have a block index */
548 vtn_assert(!block_index);
549
550 /* Start off with at the start of the push constant block. */
551 offset = nir_imm_int(&b->nb, 0);
552 } else {
553 /* The code above should have ensured a block_index when needed. */
554 vtn_assert(block_index);
555
556 /* Start off with at the start of the buffer. */
557 offset = nir_imm_int(&b->nb, 0);
558 }
559 }
560
561 if (deref_chain->ptr_as_array && idx == 0) {
562 /* We need ptr_type for the stride */
563 vtn_assert(base->ptr_type);
564
565 /* We need at least one element in the chain */
566 vtn_assert(deref_chain->length >= 1);
567
568 nir_ssa_def *elem_offset =
569 vtn_access_link_as_ssa(b, deref_chain->link[idx],
570 base->ptr_type->stride, offset->bit_size);
571 offset = nir_iadd(&b->nb, offset, elem_offset);
572 idx++;
573 }
574
575 for (; idx < deref_chain->length; idx++) {
576 switch (glsl_get_base_type(type->type)) {
577 case GLSL_TYPE_UINT:
578 case GLSL_TYPE_INT:
579 case GLSL_TYPE_UINT16:
580 case GLSL_TYPE_INT16:
581 case GLSL_TYPE_UINT8:
582 case GLSL_TYPE_INT8:
583 case GLSL_TYPE_UINT64:
584 case GLSL_TYPE_INT64:
585 case GLSL_TYPE_FLOAT:
586 case GLSL_TYPE_FLOAT16:
587 case GLSL_TYPE_DOUBLE:
588 case GLSL_TYPE_BOOL:
589 case GLSL_TYPE_ARRAY: {
590 nir_ssa_def *elem_offset =
591 vtn_access_link_as_ssa(b, deref_chain->link[idx],
592 type->stride, offset->bit_size);
593 offset = nir_iadd(&b->nb, offset, elem_offset);
594 type = type->array_element;
595 access |= type->access;
596 break;
597 }
598
599 case GLSL_TYPE_INTERFACE:
600 case GLSL_TYPE_STRUCT: {
601 vtn_assert(deref_chain->link[idx].mode == vtn_access_mode_literal);
602 unsigned member = deref_chain->link[idx].id;
603 offset = nir_iadd_imm(&b->nb, offset, type->offsets[member]);
604 type = type->members[member];
605 access |= type->access;
606 break;
607 }
608
609 default:
610 vtn_fail("Invalid type for deref");
611 }
612 }
613
614 struct vtn_pointer *ptr = rzalloc(b, struct vtn_pointer);
615 ptr->mode = base->mode;
616 ptr->type = type;
617 ptr->block_index = block_index;
618 ptr->offset = offset;
619 ptr->access = access;
620
621 return ptr;
622 }
623
624 /* Dereference the given base pointer by the access chain */
625 static struct vtn_pointer *
626 vtn_pointer_dereference(struct vtn_builder *b,
627 struct vtn_pointer *base,
628 struct vtn_access_chain *deref_chain)
629 {
630 if (vtn_pointer_uses_ssa_offset(b, base)) {
631 return vtn_ssa_offset_pointer_dereference(b, base, deref_chain);
632 } else {
633 return vtn_nir_deref_pointer_dereference(b, base, deref_chain);
634 }
635 }
636
637 /* Returns an atomic_uint type based on the original uint type. The returned
638 * type will be equivalent to the original one but will have an atomic_uint
639 * type as leaf instead of an uint.
640 *
641 * Manages uint scalars, arrays, and arrays of arrays of any nested depth.
642 */
643 static const struct glsl_type *
644 repair_atomic_type(const struct glsl_type *type)
645 {
646 assert(glsl_get_base_type(glsl_without_array(type)) == GLSL_TYPE_UINT);
647 assert(glsl_type_is_scalar(glsl_without_array(type)));
648
649 if (glsl_type_is_array(type)) {
650 const struct glsl_type *atomic =
651 repair_atomic_type(glsl_get_array_element(type));
652
653 return glsl_array_type(atomic, glsl_get_length(type),
654 glsl_get_explicit_stride(type));
655 } else {
656 return glsl_atomic_uint_type();
657 }
658 }
659
660 nir_deref_instr *
661 vtn_pointer_to_deref(struct vtn_builder *b, struct vtn_pointer *ptr)
662 {
663 if (b->wa_glslang_179) {
664 /* Do on-the-fly copy propagation for samplers. */
665 if (ptr->var && ptr->var->copy_prop_sampler)
666 return vtn_pointer_to_deref(b, ptr->var->copy_prop_sampler);
667 }
668
669 vtn_assert(!vtn_pointer_uses_ssa_offset(b, ptr));
670 if (!ptr->deref) {
671 struct vtn_access_chain chain = {
672 .length = 0,
673 };
674 ptr = vtn_nir_deref_pointer_dereference(b, ptr, &chain);
675 }
676
677 return ptr->deref;
678 }
679
680 static void
681 _vtn_local_load_store(struct vtn_builder *b, bool load, nir_deref_instr *deref,
682 struct vtn_ssa_value *inout,
683 enum gl_access_qualifier access)
684 {
685 if (glsl_type_is_vector_or_scalar(deref->type)) {
686 if (load) {
687 inout->def = nir_load_deref_with_access(&b->nb, deref, access);
688 } else {
689 nir_store_deref_with_access(&b->nb, deref, inout->def, ~0, access);
690 }
691 } else if (glsl_type_is_array(deref->type) ||
692 glsl_type_is_matrix(deref->type)) {
693 unsigned elems = glsl_get_length(deref->type);
694 for (unsigned i = 0; i < elems; i++) {
695 nir_deref_instr *child =
696 nir_build_deref_array_imm(&b->nb, deref, i);
697 _vtn_local_load_store(b, load, child, inout->elems[i], access);
698 }
699 } else {
700 vtn_assert(glsl_type_is_struct_or_ifc(deref->type));
701 unsigned elems = glsl_get_length(deref->type);
702 for (unsigned i = 0; i < elems; i++) {
703 nir_deref_instr *child = nir_build_deref_struct(&b->nb, deref, i);
704 _vtn_local_load_store(b, load, child, inout->elems[i], access);
705 }
706 }
707 }
708
709 nir_deref_instr *
710 vtn_nir_deref(struct vtn_builder *b, uint32_t id)
711 {
712 struct vtn_pointer *ptr = vtn_value(b, id, vtn_value_type_pointer)->pointer;
713 return vtn_pointer_to_deref(b, ptr);
714 }
715
716 /*
717 * Gets the NIR-level deref tail, which may have as a child an array deref
718 * selecting which component due to OpAccessChain supporting per-component
719 * indexing in SPIR-V.
720 */
721 static nir_deref_instr *
722 get_deref_tail(nir_deref_instr *deref)
723 {
724 if (deref->deref_type != nir_deref_type_array)
725 return deref;
726
727 nir_deref_instr *parent =
728 nir_instr_as_deref(deref->parent.ssa->parent_instr);
729
730 if (glsl_type_is_vector(parent->type))
731 return parent;
732 else
733 return deref;
734 }
735
736 struct vtn_ssa_value *
737 vtn_local_load(struct vtn_builder *b, nir_deref_instr *src,
738 enum gl_access_qualifier access)
739 {
740 nir_deref_instr *src_tail = get_deref_tail(src);
741 struct vtn_ssa_value *val = vtn_create_ssa_value(b, src_tail->type);
742 _vtn_local_load_store(b, true, src_tail, val, access);
743
744 if (src_tail != src) {
745 val->type = src->type;
746 val->def = nir_vector_extract(&b->nb, val->def, src->arr.index.ssa);
747 }
748
749 return val;
750 }
751
752 void
753 vtn_local_store(struct vtn_builder *b, struct vtn_ssa_value *src,
754 nir_deref_instr *dest, enum gl_access_qualifier access)
755 {
756 nir_deref_instr *dest_tail = get_deref_tail(dest);
757
758 if (dest_tail != dest) {
759 struct vtn_ssa_value *val = vtn_create_ssa_value(b, dest_tail->type);
760 _vtn_local_load_store(b, true, dest_tail, val, access);
761
762 val->def = nir_vector_insert(&b->nb, 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 if (b->options && b->options->tess_levels_are_sysvals &&
1358 *mode == nir_var_shader_in) {
1359 *location = SYSTEM_VALUE_TESS_LEVEL_OUTER;
1360 set_mode_system_value(b, mode);
1361 } else {
1362 *location = VARYING_SLOT_TESS_LEVEL_OUTER;
1363 }
1364 break;
1365 case SpvBuiltInTessLevelInner:
1366 if (b->options && b->options->tess_levels_are_sysvals &&
1367 *mode == nir_var_shader_in) {
1368 *location = SYSTEM_VALUE_TESS_LEVEL_INNER;
1369 set_mode_system_value(b, mode);
1370 } else {
1371 *location = VARYING_SLOT_TESS_LEVEL_INNER;
1372 }
1373 break;
1374 case SpvBuiltInTessCoord:
1375 *location = SYSTEM_VALUE_TESS_COORD;
1376 set_mode_system_value(b, mode);
1377 break;
1378 case SpvBuiltInPatchVertices:
1379 *location = SYSTEM_VALUE_VERTICES_IN;
1380 set_mode_system_value(b, mode);
1381 break;
1382 case SpvBuiltInFragCoord:
1383 vtn_assert(*mode == nir_var_shader_in);
1384 if (b->options && b->options->frag_coord_is_sysval) {
1385 *mode = nir_var_system_value;
1386 *location = SYSTEM_VALUE_FRAG_COORD;
1387 } else {
1388 *location = VARYING_SLOT_POS;
1389 }
1390 break;
1391 case SpvBuiltInPointCoord:
1392 *location = VARYING_SLOT_PNTC;
1393 vtn_assert(*mode == nir_var_shader_in);
1394 break;
1395 case SpvBuiltInFrontFacing:
1396 *location = SYSTEM_VALUE_FRONT_FACE;
1397 set_mode_system_value(b, mode);
1398 break;
1399 case SpvBuiltInSampleId:
1400 *location = SYSTEM_VALUE_SAMPLE_ID;
1401 set_mode_system_value(b, mode);
1402 break;
1403 case SpvBuiltInSamplePosition:
1404 *location = SYSTEM_VALUE_SAMPLE_POS;
1405 set_mode_system_value(b, mode);
1406 break;
1407 case SpvBuiltInSampleMask:
1408 if (*mode == nir_var_shader_out) {
1409 *location = FRAG_RESULT_SAMPLE_MASK;
1410 } else {
1411 *location = SYSTEM_VALUE_SAMPLE_MASK_IN;
1412 set_mode_system_value(b, mode);
1413 }
1414 break;
1415 case SpvBuiltInFragDepth:
1416 *location = FRAG_RESULT_DEPTH;
1417 vtn_assert(*mode == nir_var_shader_out);
1418 break;
1419 case SpvBuiltInHelperInvocation:
1420 *location = SYSTEM_VALUE_HELPER_INVOCATION;
1421 set_mode_system_value(b, mode);
1422 break;
1423 case SpvBuiltInNumWorkgroups:
1424 *location = SYSTEM_VALUE_NUM_WORK_GROUPS;
1425 set_mode_system_value(b, mode);
1426 break;
1427 case SpvBuiltInWorkgroupSize:
1428 *location = SYSTEM_VALUE_LOCAL_GROUP_SIZE;
1429 set_mode_system_value(b, mode);
1430 break;
1431 case SpvBuiltInWorkgroupId:
1432 *location = SYSTEM_VALUE_WORK_GROUP_ID;
1433 set_mode_system_value(b, mode);
1434 break;
1435 case SpvBuiltInLocalInvocationId:
1436 *location = SYSTEM_VALUE_LOCAL_INVOCATION_ID;
1437 set_mode_system_value(b, mode);
1438 break;
1439 case SpvBuiltInLocalInvocationIndex:
1440 *location = SYSTEM_VALUE_LOCAL_INVOCATION_INDEX;
1441 set_mode_system_value(b, mode);
1442 break;
1443 case SpvBuiltInGlobalInvocationId:
1444 *location = SYSTEM_VALUE_GLOBAL_INVOCATION_ID;
1445 set_mode_system_value(b, mode);
1446 break;
1447 case SpvBuiltInGlobalLinearId:
1448 *location = SYSTEM_VALUE_GLOBAL_INVOCATION_INDEX;
1449 set_mode_system_value(b, mode);
1450 break;
1451 case SpvBuiltInBaseVertex:
1452 /* OpenGL gl_BaseVertex (SYSTEM_VALUE_BASE_VERTEX) is not the same
1453 * semantic as Vulkan BaseVertex (SYSTEM_VALUE_FIRST_VERTEX).
1454 */
1455 if (b->options->environment == NIR_SPIRV_OPENGL)
1456 *location = SYSTEM_VALUE_BASE_VERTEX;
1457 else
1458 *location = SYSTEM_VALUE_FIRST_VERTEX;
1459 set_mode_system_value(b, mode);
1460 break;
1461 case SpvBuiltInBaseInstance:
1462 *location = SYSTEM_VALUE_BASE_INSTANCE;
1463 set_mode_system_value(b, mode);
1464 break;
1465 case SpvBuiltInDrawIndex:
1466 *location = SYSTEM_VALUE_DRAW_ID;
1467 set_mode_system_value(b, mode);
1468 break;
1469 case SpvBuiltInSubgroupSize:
1470 *location = SYSTEM_VALUE_SUBGROUP_SIZE;
1471 set_mode_system_value(b, mode);
1472 break;
1473 case SpvBuiltInSubgroupId:
1474 *location = SYSTEM_VALUE_SUBGROUP_ID;
1475 set_mode_system_value(b, mode);
1476 break;
1477 case SpvBuiltInSubgroupLocalInvocationId:
1478 *location = SYSTEM_VALUE_SUBGROUP_INVOCATION;
1479 set_mode_system_value(b, mode);
1480 break;
1481 case SpvBuiltInNumSubgroups:
1482 *location = SYSTEM_VALUE_NUM_SUBGROUPS;
1483 set_mode_system_value(b, mode);
1484 break;
1485 case SpvBuiltInDeviceIndex:
1486 *location = SYSTEM_VALUE_DEVICE_INDEX;
1487 set_mode_system_value(b, mode);
1488 break;
1489 case SpvBuiltInViewIndex:
1490 *location = SYSTEM_VALUE_VIEW_INDEX;
1491 set_mode_system_value(b, mode);
1492 break;
1493 case SpvBuiltInSubgroupEqMask:
1494 *location = SYSTEM_VALUE_SUBGROUP_EQ_MASK,
1495 set_mode_system_value(b, mode);
1496 break;
1497 case SpvBuiltInSubgroupGeMask:
1498 *location = SYSTEM_VALUE_SUBGROUP_GE_MASK,
1499 set_mode_system_value(b, mode);
1500 break;
1501 case SpvBuiltInSubgroupGtMask:
1502 *location = SYSTEM_VALUE_SUBGROUP_GT_MASK,
1503 set_mode_system_value(b, mode);
1504 break;
1505 case SpvBuiltInSubgroupLeMask:
1506 *location = SYSTEM_VALUE_SUBGROUP_LE_MASK,
1507 set_mode_system_value(b, mode);
1508 break;
1509 case SpvBuiltInSubgroupLtMask:
1510 *location = SYSTEM_VALUE_SUBGROUP_LT_MASK,
1511 set_mode_system_value(b, mode);
1512 break;
1513 case SpvBuiltInFragStencilRefEXT:
1514 *location = FRAG_RESULT_STENCIL;
1515 vtn_assert(*mode == nir_var_shader_out);
1516 break;
1517 case SpvBuiltInWorkDim:
1518 *location = SYSTEM_VALUE_WORK_DIM;
1519 set_mode_system_value(b, mode);
1520 break;
1521 case SpvBuiltInGlobalSize:
1522 *location = SYSTEM_VALUE_GLOBAL_GROUP_SIZE;
1523 set_mode_system_value(b, mode);
1524 break;
1525 case SpvBuiltInBaryCoordNoPerspAMD:
1526 *location = SYSTEM_VALUE_BARYCENTRIC_LINEAR_PIXEL;
1527 set_mode_system_value(b, mode);
1528 break;
1529 case SpvBuiltInBaryCoordNoPerspCentroidAMD:
1530 *location = SYSTEM_VALUE_BARYCENTRIC_LINEAR_CENTROID;
1531 set_mode_system_value(b, mode);
1532 break;
1533 case SpvBuiltInBaryCoordNoPerspSampleAMD:
1534 *location = SYSTEM_VALUE_BARYCENTRIC_LINEAR_SAMPLE;
1535 set_mode_system_value(b, mode);
1536 break;
1537 case SpvBuiltInBaryCoordSmoothAMD:
1538 *location = SYSTEM_VALUE_BARYCENTRIC_PERSP_PIXEL;
1539 set_mode_system_value(b, mode);
1540 break;
1541 case SpvBuiltInBaryCoordSmoothCentroidAMD:
1542 *location = SYSTEM_VALUE_BARYCENTRIC_PERSP_CENTROID;
1543 set_mode_system_value(b, mode);
1544 break;
1545 case SpvBuiltInBaryCoordSmoothSampleAMD:
1546 *location = SYSTEM_VALUE_BARYCENTRIC_PERSP_SAMPLE;
1547 set_mode_system_value(b, mode);
1548 break;
1549 case SpvBuiltInBaryCoordPullModelAMD:
1550 *location = SYSTEM_VALUE_BARYCENTRIC_PULL_MODEL;
1551 set_mode_system_value(b, mode);
1552 break;
1553 default:
1554 vtn_fail("Unsupported builtin: %s (%u)",
1555 spirv_builtin_to_string(builtin), builtin);
1556 }
1557 }
1558
1559 static void
1560 apply_var_decoration(struct vtn_builder *b,
1561 struct nir_variable_data *var_data,
1562 const struct vtn_decoration *dec)
1563 {
1564 switch (dec->decoration) {
1565 case SpvDecorationRelaxedPrecision:
1566 break; /* FIXME: Do nothing with this for now. */
1567 case SpvDecorationNoPerspective:
1568 var_data->interpolation = INTERP_MODE_NOPERSPECTIVE;
1569 break;
1570 case SpvDecorationFlat:
1571 var_data->interpolation = INTERP_MODE_FLAT;
1572 break;
1573 case SpvDecorationExplicitInterpAMD:
1574 var_data->interpolation = INTERP_MODE_EXPLICIT;
1575 break;
1576 case SpvDecorationCentroid:
1577 var_data->centroid = true;
1578 break;
1579 case SpvDecorationSample:
1580 var_data->sample = true;
1581 break;
1582 case SpvDecorationInvariant:
1583 var_data->invariant = true;
1584 break;
1585 case SpvDecorationConstant:
1586 var_data->read_only = true;
1587 break;
1588 case SpvDecorationNonReadable:
1589 var_data->access |= ACCESS_NON_READABLE;
1590 break;
1591 case SpvDecorationNonWritable:
1592 var_data->read_only = true;
1593 var_data->access |= ACCESS_NON_WRITEABLE;
1594 break;
1595 case SpvDecorationRestrict:
1596 var_data->access |= ACCESS_RESTRICT;
1597 break;
1598 case SpvDecorationAliased:
1599 var_data->access &= ~ACCESS_RESTRICT;
1600 break;
1601 case SpvDecorationVolatile:
1602 var_data->access |= ACCESS_VOLATILE;
1603 break;
1604 case SpvDecorationCoherent:
1605 var_data->access |= ACCESS_COHERENT;
1606 break;
1607 case SpvDecorationComponent:
1608 var_data->location_frac = dec->operands[0];
1609 break;
1610 case SpvDecorationIndex:
1611 var_data->index = dec->operands[0];
1612 break;
1613 case SpvDecorationBuiltIn: {
1614 SpvBuiltIn builtin = dec->operands[0];
1615
1616 nir_variable_mode mode = var_data->mode;
1617 vtn_get_builtin_location(b, builtin, &var_data->location, &mode);
1618 var_data->mode = mode;
1619
1620 switch (builtin) {
1621 case SpvBuiltInTessLevelOuter:
1622 case SpvBuiltInTessLevelInner:
1623 case SpvBuiltInClipDistance:
1624 case SpvBuiltInCullDistance:
1625 var_data->compact = true;
1626 break;
1627 default:
1628 break;
1629 }
1630 }
1631
1632 case SpvDecorationSpecId:
1633 case SpvDecorationRowMajor:
1634 case SpvDecorationColMajor:
1635 case SpvDecorationMatrixStride:
1636 case SpvDecorationUniform:
1637 case SpvDecorationUniformId:
1638 case SpvDecorationLinkageAttributes:
1639 break; /* Do nothing with these here */
1640
1641 case SpvDecorationPatch:
1642 var_data->patch = true;
1643 break;
1644
1645 case SpvDecorationLocation:
1646 vtn_fail("Handled above");
1647
1648 case SpvDecorationBlock:
1649 case SpvDecorationBufferBlock:
1650 case SpvDecorationArrayStride:
1651 case SpvDecorationGLSLShared:
1652 case SpvDecorationGLSLPacked:
1653 break; /* These can apply to a type but we don't care about them */
1654
1655 case SpvDecorationBinding:
1656 case SpvDecorationDescriptorSet:
1657 case SpvDecorationNoContraction:
1658 case SpvDecorationInputAttachmentIndex:
1659 vtn_warn("Decoration not allowed for variable or structure member: %s",
1660 spirv_decoration_to_string(dec->decoration));
1661 break;
1662
1663 case SpvDecorationXfbBuffer:
1664 var_data->explicit_xfb_buffer = true;
1665 var_data->xfb.buffer = dec->operands[0];
1666 var_data->always_active_io = true;
1667 break;
1668 case SpvDecorationXfbStride:
1669 var_data->explicit_xfb_stride = true;
1670 var_data->xfb.stride = dec->operands[0];
1671 break;
1672 case SpvDecorationOffset:
1673 var_data->explicit_offset = true;
1674 var_data->offset = dec->operands[0];
1675 break;
1676
1677 case SpvDecorationStream:
1678 var_data->stream = dec->operands[0];
1679 break;
1680
1681 case SpvDecorationCPacked:
1682 case SpvDecorationSaturatedConversion:
1683 case SpvDecorationFuncParamAttr:
1684 case SpvDecorationFPRoundingMode:
1685 case SpvDecorationFPFastMathMode:
1686 case SpvDecorationAlignment:
1687 if (b->shader->info.stage != MESA_SHADER_KERNEL) {
1688 vtn_warn("Decoration only allowed for CL-style kernels: %s",
1689 spirv_decoration_to_string(dec->decoration));
1690 }
1691 break;
1692
1693 case SpvDecorationUserSemantic:
1694 case SpvDecorationUserTypeGOOGLE:
1695 /* User semantic decorations can safely be ignored by the driver. */
1696 break;
1697
1698 case SpvDecorationRestrictPointerEXT:
1699 case SpvDecorationAliasedPointerEXT:
1700 /* TODO: We should actually plumb alias information through NIR. */
1701 break;
1702
1703 default:
1704 vtn_fail_with_decoration("Unhandled decoration", dec->decoration);
1705 }
1706 }
1707
1708 static void
1709 var_is_patch_cb(struct vtn_builder *b, struct vtn_value *val, int member,
1710 const struct vtn_decoration *dec, void *out_is_patch)
1711 {
1712 if (dec->decoration == SpvDecorationPatch) {
1713 *((bool *) out_is_patch) = true;
1714 }
1715 }
1716
1717 static void
1718 var_decoration_cb(struct vtn_builder *b, struct vtn_value *val, int member,
1719 const struct vtn_decoration *dec, void *void_var)
1720 {
1721 struct vtn_variable *vtn_var = void_var;
1722
1723 /* Handle decorations that apply to a vtn_variable as a whole */
1724 switch (dec->decoration) {
1725 case SpvDecorationBinding:
1726 vtn_var->binding = dec->operands[0];
1727 vtn_var->explicit_binding = true;
1728 return;
1729 case SpvDecorationDescriptorSet:
1730 vtn_var->descriptor_set = dec->operands[0];
1731 return;
1732 case SpvDecorationInputAttachmentIndex:
1733 vtn_var->input_attachment_index = dec->operands[0];
1734 return;
1735 case SpvDecorationPatch:
1736 vtn_var->patch = true;
1737 break;
1738 case SpvDecorationOffset:
1739 vtn_var->offset = dec->operands[0];
1740 break;
1741 case SpvDecorationNonWritable:
1742 vtn_var->access |= ACCESS_NON_WRITEABLE;
1743 break;
1744 case SpvDecorationNonReadable:
1745 vtn_var->access |= ACCESS_NON_READABLE;
1746 break;
1747 case SpvDecorationVolatile:
1748 vtn_var->access |= ACCESS_VOLATILE;
1749 break;
1750 case SpvDecorationCoherent:
1751 vtn_var->access |= ACCESS_COHERENT;
1752 break;
1753 case SpvDecorationCounterBuffer:
1754 /* Counter buffer decorations can safely be ignored by the driver. */
1755 return;
1756 default:
1757 break;
1758 }
1759
1760 if (val->value_type == vtn_value_type_pointer) {
1761 assert(val->pointer->var == void_var);
1762 assert(member == -1);
1763 } else {
1764 assert(val->value_type == vtn_value_type_type);
1765 }
1766
1767 /* Location is odd. If applied to a split structure, we have to walk the
1768 * whole thing and accumulate the location. It's easier to handle as a
1769 * special case.
1770 */
1771 if (dec->decoration == SpvDecorationLocation) {
1772 unsigned location = dec->operands[0];
1773 if (b->shader->info.stage == MESA_SHADER_FRAGMENT &&
1774 vtn_var->mode == vtn_variable_mode_output) {
1775 location += FRAG_RESULT_DATA0;
1776 } else if (b->shader->info.stage == MESA_SHADER_VERTEX &&
1777 vtn_var->mode == vtn_variable_mode_input) {
1778 location += VERT_ATTRIB_GENERIC0;
1779 } else if (vtn_var->mode == vtn_variable_mode_input ||
1780 vtn_var->mode == vtn_variable_mode_output) {
1781 location += vtn_var->patch ? VARYING_SLOT_PATCH0 : VARYING_SLOT_VAR0;
1782 } else if (vtn_var->mode != vtn_variable_mode_uniform) {
1783 vtn_warn("Location must be on input, output, uniform, sampler or "
1784 "image variable");
1785 return;
1786 }
1787
1788 if (vtn_var->var->num_members == 0) {
1789 /* This handles the member and lone variable cases */
1790 vtn_var->var->data.location = location;
1791 } else {
1792 /* This handles the structure member case */
1793 assert(vtn_var->var->members);
1794
1795 if (member == -1)
1796 vtn_var->base_location = location;
1797 else
1798 vtn_var->var->members[member].location = location;
1799 }
1800
1801 return;
1802 } else {
1803 if (vtn_var->var) {
1804 if (vtn_var->var->num_members == 0) {
1805 /* We call this function on types as well as variables and not all
1806 * struct types get split so we can end up having stray member
1807 * decorations; just ignore them.
1808 */
1809 if (member == -1)
1810 apply_var_decoration(b, &vtn_var->var->data, dec);
1811 } else if (member >= 0) {
1812 /* Member decorations must come from a type */
1813 assert(val->value_type == vtn_value_type_type);
1814 apply_var_decoration(b, &vtn_var->var->members[member], dec);
1815 } else {
1816 unsigned length =
1817 glsl_get_length(glsl_without_array(vtn_var->type->type));
1818 for (unsigned i = 0; i < length; i++)
1819 apply_var_decoration(b, &vtn_var->var->members[i], dec);
1820 }
1821 } else {
1822 /* A few variables, those with external storage, have no actual
1823 * nir_variables associated with them. Fortunately, all decorations
1824 * we care about for those variables are on the type only.
1825 */
1826 vtn_assert(vtn_var->mode == vtn_variable_mode_ubo ||
1827 vtn_var->mode == vtn_variable_mode_ssbo ||
1828 vtn_var->mode == vtn_variable_mode_push_constant);
1829 }
1830 }
1831 }
1832
1833 enum vtn_variable_mode
1834 vtn_storage_class_to_mode(struct vtn_builder *b,
1835 SpvStorageClass class,
1836 struct vtn_type *interface_type,
1837 nir_variable_mode *nir_mode_out)
1838 {
1839 enum vtn_variable_mode mode;
1840 nir_variable_mode nir_mode;
1841 switch (class) {
1842 case SpvStorageClassUniform:
1843 /* Assume it's an UBO if we lack the interface_type. */
1844 if (!interface_type || interface_type->block) {
1845 mode = vtn_variable_mode_ubo;
1846 nir_mode = nir_var_mem_ubo;
1847 } else if (interface_type->buffer_block) {
1848 mode = vtn_variable_mode_ssbo;
1849 nir_mode = nir_var_mem_ssbo;
1850 } else {
1851 /* Default-block uniforms, coming from gl_spirv */
1852 mode = vtn_variable_mode_uniform;
1853 nir_mode = nir_var_uniform;
1854 }
1855 break;
1856 case SpvStorageClassStorageBuffer:
1857 mode = vtn_variable_mode_ssbo;
1858 nir_mode = nir_var_mem_ssbo;
1859 break;
1860 case SpvStorageClassPhysicalStorageBuffer:
1861 mode = vtn_variable_mode_phys_ssbo;
1862 nir_mode = nir_var_mem_global;
1863 break;
1864 case SpvStorageClassUniformConstant:
1865 if (b->shader->info.stage == MESA_SHADER_KERNEL) {
1866 if (b->options->constant_as_global) {
1867 mode = vtn_variable_mode_cross_workgroup;
1868 nir_mode = nir_var_mem_global;
1869 } else {
1870 mode = vtn_variable_mode_ubo;
1871 nir_mode = nir_var_mem_ubo;
1872 }
1873 } else {
1874 mode = vtn_variable_mode_uniform;
1875 nir_mode = nir_var_uniform;
1876 }
1877 break;
1878 case SpvStorageClassPushConstant:
1879 mode = vtn_variable_mode_push_constant;
1880 nir_mode = nir_var_uniform;
1881 break;
1882 case SpvStorageClassInput:
1883 mode = vtn_variable_mode_input;
1884 nir_mode = nir_var_shader_in;
1885 break;
1886 case SpvStorageClassOutput:
1887 mode = vtn_variable_mode_output;
1888 nir_mode = nir_var_shader_out;
1889 break;
1890 case SpvStorageClassPrivate:
1891 mode = vtn_variable_mode_private;
1892 nir_mode = nir_var_shader_temp;
1893 break;
1894 case SpvStorageClassFunction:
1895 mode = vtn_variable_mode_function;
1896 nir_mode = nir_var_function_temp;
1897 break;
1898 case SpvStorageClassWorkgroup:
1899 mode = vtn_variable_mode_workgroup;
1900 nir_mode = nir_var_mem_shared;
1901 break;
1902 case SpvStorageClassAtomicCounter:
1903 mode = vtn_variable_mode_uniform;
1904 nir_mode = nir_var_uniform;
1905 break;
1906 case SpvStorageClassCrossWorkgroup:
1907 mode = vtn_variable_mode_cross_workgroup;
1908 nir_mode = nir_var_mem_global;
1909 break;
1910 case SpvStorageClassImage:
1911 mode = vtn_variable_mode_image;
1912 nir_mode = nir_var_mem_ubo;
1913 break;
1914 case SpvStorageClassGeneric:
1915 default:
1916 vtn_fail("Unhandled variable storage class: %s (%u)",
1917 spirv_storageclass_to_string(class), class);
1918 }
1919
1920 if (nir_mode_out)
1921 *nir_mode_out = nir_mode;
1922
1923 return mode;
1924 }
1925
1926 nir_address_format
1927 vtn_mode_to_address_format(struct vtn_builder *b, enum vtn_variable_mode mode)
1928 {
1929 switch (mode) {
1930 case vtn_variable_mode_ubo:
1931 return b->options->ubo_addr_format;
1932
1933 case vtn_variable_mode_ssbo:
1934 return b->options->ssbo_addr_format;
1935
1936 case vtn_variable_mode_phys_ssbo:
1937 return b->options->phys_ssbo_addr_format;
1938
1939 case vtn_variable_mode_push_constant:
1940 return b->options->push_const_addr_format;
1941
1942 case vtn_variable_mode_workgroup:
1943 return b->options->shared_addr_format;
1944
1945 case vtn_variable_mode_cross_workgroup:
1946 return b->options->global_addr_format;
1947
1948 case vtn_variable_mode_function:
1949 if (b->physical_ptrs)
1950 return b->options->temp_addr_format;
1951 /* Fall through. */
1952
1953 case vtn_variable_mode_private:
1954 case vtn_variable_mode_uniform:
1955 case vtn_variable_mode_input:
1956 case vtn_variable_mode_output:
1957 case vtn_variable_mode_image:
1958 return nir_address_format_logical;
1959 }
1960
1961 unreachable("Invalid variable mode");
1962 }
1963
1964 nir_ssa_def *
1965 vtn_pointer_to_ssa(struct vtn_builder *b, struct vtn_pointer *ptr)
1966 {
1967 if (vtn_pointer_uses_ssa_offset(b, ptr)) {
1968 /* This pointer needs to have a pointer type with actual storage */
1969 vtn_assert(ptr->ptr_type);
1970 vtn_assert(ptr->ptr_type->type);
1971
1972 if (!ptr->offset) {
1973 /* If we don't have an offset then we must be a pointer to the variable
1974 * itself.
1975 */
1976 vtn_assert(!ptr->offset && !ptr->block_index);
1977
1978 struct vtn_access_chain chain = {
1979 .length = 0,
1980 };
1981 ptr = vtn_ssa_offset_pointer_dereference(b, ptr, &chain);
1982 }
1983
1984 vtn_assert(ptr->offset);
1985 if (ptr->block_index) {
1986 vtn_assert(ptr->mode == vtn_variable_mode_ubo ||
1987 ptr->mode == vtn_variable_mode_ssbo);
1988 return nir_vec2(&b->nb, ptr->block_index, ptr->offset);
1989 } else {
1990 vtn_assert(ptr->mode == vtn_variable_mode_workgroup);
1991 return ptr->offset;
1992 }
1993 } else {
1994 if (vtn_pointer_is_external_block(b, ptr) &&
1995 vtn_type_contains_block(b, ptr->type) &&
1996 ptr->mode != vtn_variable_mode_phys_ssbo) {
1997 /* In this case, we're looking for a block index and not an actual
1998 * deref.
1999 *
2000 * For PhysicalStorageBuffer pointers, we don't have a block index
2001 * at all because we get the pointer directly from the client. This
2002 * assumes that there will never be a SSBO binding variable using the
2003 * PhysicalStorageBuffer storage class. This assumption appears
2004 * to be correct according to the Vulkan spec because the table,
2005 * "Shader Resource and Storage Class Correspondence," the only the
2006 * Uniform storage class with BufferBlock or the StorageBuffer
2007 * storage class with Block can be used.
2008 */
2009 if (!ptr->block_index) {
2010 /* If we don't have a block_index then we must be a pointer to the
2011 * variable itself.
2012 */
2013 vtn_assert(!ptr->deref);
2014
2015 struct vtn_access_chain chain = {
2016 .length = 0,
2017 };
2018 ptr = vtn_nir_deref_pointer_dereference(b, ptr, &chain);
2019 }
2020
2021 return ptr->block_index;
2022 } else {
2023 return &vtn_pointer_to_deref(b, ptr)->dest.ssa;
2024 }
2025 }
2026 }
2027
2028 struct vtn_pointer *
2029 vtn_pointer_from_ssa(struct vtn_builder *b, nir_ssa_def *ssa,
2030 struct vtn_type *ptr_type)
2031 {
2032 vtn_assert(ptr_type->base_type == vtn_base_type_pointer);
2033
2034 struct vtn_pointer *ptr = rzalloc(b, struct vtn_pointer);
2035 struct vtn_type *without_array =
2036 vtn_type_without_array(ptr_type->deref);
2037
2038 nir_variable_mode nir_mode;
2039 ptr->mode = vtn_storage_class_to_mode(b, ptr_type->storage_class,
2040 without_array, &nir_mode);
2041 ptr->type = ptr_type->deref;
2042 ptr->ptr_type = ptr_type;
2043
2044 if (b->wa_glslang_179) {
2045 /* To work around https://github.com/KhronosGroup/glslang/issues/179 we
2046 * need to whack the mode because it creates a function parameter with
2047 * the Function storage class even though it's a pointer to a sampler.
2048 * If we don't do this, then NIR won't get rid of the deref_cast for us.
2049 */
2050 if (ptr->mode == vtn_variable_mode_function &&
2051 (ptr->type->base_type == vtn_base_type_sampler ||
2052 ptr->type->base_type == vtn_base_type_sampled_image)) {
2053 ptr->mode = vtn_variable_mode_uniform;
2054 nir_mode = nir_var_uniform;
2055 }
2056 }
2057
2058 if (vtn_pointer_uses_ssa_offset(b, ptr)) {
2059 /* This pointer type needs to have actual storage */
2060 vtn_assert(ptr_type->type);
2061 if (ptr->mode == vtn_variable_mode_ubo ||
2062 ptr->mode == vtn_variable_mode_ssbo) {
2063 vtn_assert(ssa->num_components == 2);
2064 ptr->block_index = nir_channel(&b->nb, ssa, 0);
2065 ptr->offset = nir_channel(&b->nb, ssa, 1);
2066 } else {
2067 vtn_assert(ssa->num_components == 1);
2068 ptr->block_index = NULL;
2069 ptr->offset = ssa;
2070 }
2071 } else {
2072 const struct glsl_type *deref_type = ptr_type->deref->type;
2073 if (!vtn_pointer_is_external_block(b, ptr)) {
2074 ptr->deref = nir_build_deref_cast(&b->nb, ssa, nir_mode,
2075 deref_type, ptr_type->stride);
2076 } else if (vtn_type_contains_block(b, ptr->type) &&
2077 ptr->mode != vtn_variable_mode_phys_ssbo) {
2078 /* This is a pointer to somewhere in an array of blocks, not a
2079 * pointer to somewhere inside the block. Set the block index
2080 * instead of making a cast.
2081 */
2082 ptr->block_index = ssa;
2083 } else {
2084 /* This is a pointer to something internal or a pointer inside a
2085 * block. It's just a regular cast.
2086 *
2087 * For PhysicalStorageBuffer pointers, we don't have a block index
2088 * at all because we get the pointer directly from the client. This
2089 * assumes that there will never be a SSBO binding variable using the
2090 * PhysicalStorageBuffer storage class. This assumption appears
2091 * to be correct according to the Vulkan spec because the table,
2092 * "Shader Resource and Storage Class Correspondence," the only the
2093 * Uniform storage class with BufferBlock or the StorageBuffer
2094 * storage class with Block can be used.
2095 */
2096 ptr->deref = nir_build_deref_cast(&b->nb, ssa, nir_mode,
2097 ptr_type->deref->type,
2098 ptr_type->stride);
2099 ptr->deref->dest.ssa.num_components =
2100 glsl_get_vector_elements(ptr_type->type);
2101 ptr->deref->dest.ssa.bit_size = glsl_get_bit_size(ptr_type->type);
2102 }
2103 }
2104
2105 return ptr;
2106 }
2107
2108 static bool
2109 is_per_vertex_inout(const struct vtn_variable *var, gl_shader_stage stage)
2110 {
2111 if (var->patch || !glsl_type_is_array(var->type->type))
2112 return false;
2113
2114 if (var->mode == vtn_variable_mode_input) {
2115 return stage == MESA_SHADER_TESS_CTRL ||
2116 stage == MESA_SHADER_TESS_EVAL ||
2117 stage == MESA_SHADER_GEOMETRY;
2118 }
2119
2120 if (var->mode == vtn_variable_mode_output)
2121 return stage == MESA_SHADER_TESS_CTRL;
2122
2123 return false;
2124 }
2125
2126 static void
2127 assign_missing_member_locations(struct vtn_variable *var)
2128 {
2129 unsigned length =
2130 glsl_get_length(glsl_without_array(var->type->type));
2131 int location = var->base_location;
2132
2133 for (unsigned i = 0; i < length; i++) {
2134 /* From the Vulkan spec:
2135 *
2136 * “If the structure type is a Block but without a Location, then each
2137 * of its members must have a Location decoration.”
2138 *
2139 */
2140 if (var->type->block) {
2141 assert(var->base_location != -1 ||
2142 var->var->members[i].location != -1);
2143 }
2144
2145 /* From the Vulkan spec:
2146 *
2147 * “Any member with its own Location decoration is assigned that
2148 * location. Each remaining member is assigned the location after the
2149 * immediately preceding member in declaration order.”
2150 */
2151 if (var->var->members[i].location != -1)
2152 location = var->var->members[i].location;
2153 else
2154 var->var->members[i].location = location;
2155
2156 /* Below we use type instead of interface_type, because interface_type
2157 * is only available when it is a Block. This code also supports
2158 * input/outputs that are just structs
2159 */
2160 const struct glsl_type *member_type =
2161 glsl_get_struct_field(glsl_without_array(var->type->type), i);
2162
2163 location +=
2164 glsl_count_attribute_slots(member_type,
2165 false /* is_gl_vertex_input */);
2166 }
2167 }
2168
2169
2170 static void
2171 vtn_create_variable(struct vtn_builder *b, struct vtn_value *val,
2172 struct vtn_type *ptr_type, SpvStorageClass storage_class,
2173 nir_constant *const_initializer, nir_variable *var_initializer)
2174 {
2175 vtn_assert(ptr_type->base_type == vtn_base_type_pointer);
2176 struct vtn_type *type = ptr_type->deref;
2177
2178 struct vtn_type *without_array = vtn_type_without_array(ptr_type->deref);
2179
2180 enum vtn_variable_mode mode;
2181 nir_variable_mode nir_mode;
2182 mode = vtn_storage_class_to_mode(b, storage_class, without_array, &nir_mode);
2183
2184 switch (mode) {
2185 case vtn_variable_mode_ubo:
2186 /* There's no other way to get vtn_variable_mode_ubo */
2187 vtn_assert(without_array->block);
2188 b->shader->info.num_ubos++;
2189 break;
2190 case vtn_variable_mode_ssbo:
2191 if (storage_class == SpvStorageClassStorageBuffer &&
2192 !without_array->block) {
2193 if (b->variable_pointers) {
2194 vtn_fail("Variables in the StorageBuffer storage class must "
2195 "have a struct type with the Block decoration");
2196 } else {
2197 /* If variable pointers are not present, it's still malformed
2198 * SPIR-V but we can parse it and do the right thing anyway.
2199 * Since some of the 8-bit storage tests have bugs in this are,
2200 * just make it a warning for now.
2201 */
2202 vtn_warn("Variables in the StorageBuffer storage class must "
2203 "have a struct type with the Block decoration");
2204 }
2205 }
2206 b->shader->info.num_ssbos++;
2207 break;
2208 case vtn_variable_mode_uniform:
2209 if (glsl_type_is_image(without_array->type))
2210 b->shader->info.num_images++;
2211 else if (glsl_type_is_sampler(without_array->type))
2212 b->shader->info.num_textures++;
2213 break;
2214 case vtn_variable_mode_push_constant:
2215 b->shader->num_uniforms = vtn_type_block_size(b, type);
2216 break;
2217
2218 case vtn_variable_mode_image:
2219 vtn_fail("Cannot create a variable with the Image storage class");
2220 break;
2221
2222 case vtn_variable_mode_phys_ssbo:
2223 vtn_fail("Cannot create a variable with the "
2224 "PhysicalStorageBuffer storage class");
2225 break;
2226
2227 default:
2228 /* No tallying is needed */
2229 break;
2230 }
2231
2232 struct vtn_variable *var = rzalloc(b, struct vtn_variable);
2233 var->type = type;
2234 var->mode = mode;
2235 var->base_location = -1;
2236
2237 val->pointer = rzalloc(b, struct vtn_pointer);
2238 val->pointer->mode = var->mode;
2239 val->pointer->type = var->type;
2240 val->pointer->ptr_type = ptr_type;
2241 val->pointer->var = var;
2242 val->pointer->access = var->type->access;
2243
2244 switch (var->mode) {
2245 case vtn_variable_mode_function:
2246 case vtn_variable_mode_private:
2247 case vtn_variable_mode_uniform:
2248 /* For these, we create the variable normally */
2249 var->var = rzalloc(b->shader, nir_variable);
2250 var->var->name = ralloc_strdup(var->var, val->name);
2251
2252 if (storage_class == SpvStorageClassAtomicCounter) {
2253 /* Need to tweak the nir type here as at vtn_handle_type we don't
2254 * have the access to storage_class, that is the one that points us
2255 * that is an atomic uint.
2256 */
2257 var->var->type = repair_atomic_type(var->type->type);
2258 } else {
2259 /* Private variables don't have any explicit layout but some layouts
2260 * may have leaked through due to type deduplication in the SPIR-V.
2261 */
2262 var->var->type = var->type->type;
2263 }
2264 var->var->data.mode = nir_mode;
2265 var->var->data.location = -1;
2266 var->var->interface_type = NULL;
2267 break;
2268
2269 case vtn_variable_mode_ubo:
2270 case vtn_variable_mode_ssbo:
2271 var->var = rzalloc(b->shader, nir_variable);
2272 var->var->name = ralloc_strdup(var->var, val->name);
2273
2274 var->var->type = var->type->type;
2275 var->var->interface_type = var->type->type;
2276
2277 var->var->data.mode = nir_mode;
2278 var->var->data.location = -1;
2279
2280 break;
2281
2282 case vtn_variable_mode_workgroup:
2283 /* Create the variable normally */
2284 var->var = rzalloc(b->shader, nir_variable);
2285 var->var->name = ralloc_strdup(var->var, val->name);
2286 /* Workgroup variables don't have any explicit layout but some
2287 * layouts may have leaked through due to type deduplication in the
2288 * SPIR-V.
2289 */
2290 var->var->type = var->type->type;
2291 var->var->data.mode = nir_var_mem_shared;
2292 break;
2293
2294 case vtn_variable_mode_input:
2295 case vtn_variable_mode_output: {
2296 /* In order to know whether or not we're a per-vertex inout, we need
2297 * the patch qualifier. This means walking the variable decorations
2298 * early before we actually create any variables. Not a big deal.
2299 *
2300 * GLSLang really likes to place decorations in the most interior
2301 * thing it possibly can. In particular, if you have a struct, it
2302 * will place the patch decorations on the struct members. This
2303 * should be handled by the variable splitting below just fine.
2304 *
2305 * If you have an array-of-struct, things get even more weird as it
2306 * will place the patch decorations on the struct even though it's
2307 * inside an array and some of the members being patch and others not
2308 * makes no sense whatsoever. Since the only sensible thing is for
2309 * it to be all or nothing, we'll call it patch if any of the members
2310 * are declared patch.
2311 */
2312 var->patch = false;
2313 vtn_foreach_decoration(b, val, var_is_patch_cb, &var->patch);
2314 if (glsl_type_is_array(var->type->type) &&
2315 glsl_type_is_struct_or_ifc(without_array->type)) {
2316 vtn_foreach_decoration(b, vtn_value(b, without_array->id,
2317 vtn_value_type_type),
2318 var_is_patch_cb, &var->patch);
2319 }
2320
2321 /* For inputs and outputs, we immediately split structures. This
2322 * is for a couple of reasons. For one, builtins may all come in
2323 * a struct and we really want those split out into separate
2324 * variables. For another, interpolation qualifiers can be
2325 * applied to members of the top-level struct ane we need to be
2326 * able to preserve that information.
2327 */
2328
2329 struct vtn_type *per_vertex_type = var->type;
2330 if (is_per_vertex_inout(var, b->shader->info.stage)) {
2331 /* In Geometry shaders (and some tessellation), inputs come
2332 * in per-vertex arrays. However, some builtins come in
2333 * non-per-vertex, hence the need for the is_array check. In
2334 * any case, there are no non-builtin arrays allowed so this
2335 * check should be sufficient.
2336 */
2337 per_vertex_type = var->type->array_element;
2338 }
2339
2340 var->var = rzalloc(b->shader, nir_variable);
2341 var->var->name = ralloc_strdup(var->var, val->name);
2342 /* In Vulkan, shader I/O variables don't have any explicit layout but
2343 * some layouts may have leaked through due to type deduplication in
2344 * the SPIR-V. We do, however, keep the layouts in the variable's
2345 * interface_type because we need offsets for XFB arrays of blocks.
2346 */
2347 var->var->type = var->type->type;
2348 var->var->data.mode = nir_mode;
2349 var->var->data.patch = var->patch;
2350
2351 /* Figure out the interface block type. */
2352 struct vtn_type *iface_type = per_vertex_type;
2353 if (var->mode == vtn_variable_mode_output &&
2354 (b->shader->info.stage == MESA_SHADER_VERTEX ||
2355 b->shader->info.stage == MESA_SHADER_TESS_EVAL ||
2356 b->shader->info.stage == MESA_SHADER_GEOMETRY)) {
2357 /* For vertex data outputs, we can end up with arrays of blocks for
2358 * transform feedback where each array element corresponds to a
2359 * different XFB output buffer.
2360 */
2361 while (iface_type->base_type == vtn_base_type_array)
2362 iface_type = iface_type->array_element;
2363 }
2364 if (iface_type->base_type == vtn_base_type_struct && iface_type->block)
2365 var->var->interface_type = iface_type->type;
2366
2367 if (per_vertex_type->base_type == vtn_base_type_struct &&
2368 per_vertex_type->block) {
2369 /* It's a struct. Set it up as per-member. */
2370 var->var->num_members = glsl_get_length(per_vertex_type->type);
2371 var->var->members = rzalloc_array(var->var, struct nir_variable_data,
2372 var->var->num_members);
2373
2374 for (unsigned i = 0; i < var->var->num_members; i++) {
2375 var->var->members[i].mode = nir_mode;
2376 var->var->members[i].patch = var->patch;
2377 var->var->members[i].location = -1;
2378 }
2379 }
2380
2381 /* For inputs and outputs, we need to grab locations and builtin
2382 * information from the per-vertex type.
2383 */
2384 vtn_foreach_decoration(b, vtn_value(b, per_vertex_type->id,
2385 vtn_value_type_type),
2386 var_decoration_cb, var);
2387 break;
2388 }
2389
2390 case vtn_variable_mode_push_constant:
2391 case vtn_variable_mode_cross_workgroup:
2392 /* These don't need actual variables. */
2393 break;
2394
2395 case vtn_variable_mode_image:
2396 case vtn_variable_mode_phys_ssbo:
2397 unreachable("Should have been caught before");
2398 }
2399
2400 /* We can only have one type of initializer */
2401 assert(!(const_initializer && var_initializer));
2402 if (const_initializer) {
2403 var->var->constant_initializer =
2404 nir_constant_clone(const_initializer, var->var);
2405 }
2406 if (var_initializer)
2407 var->var->pointer_initializer = var_initializer;
2408
2409 if (var->mode == vtn_variable_mode_uniform ||
2410 var->mode == vtn_variable_mode_ssbo) {
2411 /* SSBOs and images are assumed to not alias in the Simple, GLSL and Vulkan memory models */
2412 var->var->data.access |= b->mem_model != SpvMemoryModelOpenCL ? ACCESS_RESTRICT : 0;
2413 }
2414
2415 vtn_foreach_decoration(b, val, var_decoration_cb, var);
2416 vtn_foreach_decoration(b, val, ptr_decoration_cb, val->pointer);
2417
2418 /* Propagate access flags from the OpVariable decorations. */
2419 val->pointer->access |= var->access;
2420
2421 if ((var->mode == vtn_variable_mode_input ||
2422 var->mode == vtn_variable_mode_output) &&
2423 var->var->members) {
2424 assign_missing_member_locations(var);
2425 }
2426
2427 if (var->mode == vtn_variable_mode_uniform ||
2428 var->mode == vtn_variable_mode_ubo ||
2429 var->mode == vtn_variable_mode_ssbo) {
2430 /* XXX: We still need the binding information in the nir_variable
2431 * for these. We should fix that.
2432 */
2433 var->var->data.binding = var->binding;
2434 var->var->data.explicit_binding = var->explicit_binding;
2435 var->var->data.descriptor_set = var->descriptor_set;
2436 var->var->data.index = var->input_attachment_index;
2437 var->var->data.offset = var->offset;
2438
2439 if (glsl_type_is_image(without_array->type))
2440 var->var->data.image.format = without_array->image_format;
2441 }
2442
2443 if (var->mode == vtn_variable_mode_function) {
2444 vtn_assert(var->var != NULL && var->var->members == NULL);
2445 nir_function_impl_add_variable(b->nb.impl, var->var);
2446 } else if (var->var) {
2447 nir_shader_add_variable(b->shader, var->var);
2448 } else {
2449 vtn_assert(vtn_pointer_is_external_block(b, val->pointer));
2450 }
2451 }
2452
2453 static void
2454 vtn_assert_types_equal(struct vtn_builder *b, SpvOp opcode,
2455 struct vtn_type *dst_type,
2456 struct vtn_type *src_type)
2457 {
2458 if (dst_type->id == src_type->id)
2459 return;
2460
2461 if (vtn_types_compatible(b, dst_type, src_type)) {
2462 /* Early versions of GLSLang would re-emit types unnecessarily and you
2463 * would end up with OpLoad, OpStore, or OpCopyMemory opcodes which have
2464 * mismatched source and destination types.
2465 *
2466 * https://github.com/KhronosGroup/glslang/issues/304
2467 * https://github.com/KhronosGroup/glslang/issues/307
2468 * https://bugs.freedesktop.org/show_bug.cgi?id=104338
2469 * https://bugs.freedesktop.org/show_bug.cgi?id=104424
2470 */
2471 vtn_warn("Source and destination types of %s do not have the same "
2472 "ID (but are compatible): %u vs %u",
2473 spirv_op_to_string(opcode), dst_type->id, src_type->id);
2474 return;
2475 }
2476
2477 vtn_fail("Source and destination types of %s do not match: %s vs. %s",
2478 spirv_op_to_string(opcode),
2479 glsl_get_type_name(dst_type->type),
2480 glsl_get_type_name(src_type->type));
2481 }
2482
2483 static nir_ssa_def *
2484 nir_shrink_zero_pad_vec(nir_builder *b, nir_ssa_def *val,
2485 unsigned num_components)
2486 {
2487 if (val->num_components == num_components)
2488 return val;
2489
2490 nir_ssa_def *comps[NIR_MAX_VEC_COMPONENTS];
2491 for (unsigned i = 0; i < num_components; i++) {
2492 if (i < val->num_components)
2493 comps[i] = nir_channel(b, val, i);
2494 else
2495 comps[i] = nir_imm_intN_t(b, 0, val->bit_size);
2496 }
2497 return nir_vec(b, comps, num_components);
2498 }
2499
2500 static nir_ssa_def *
2501 nir_sloppy_bitcast(nir_builder *b, nir_ssa_def *val,
2502 const struct glsl_type *type)
2503 {
2504 const unsigned num_components = glsl_get_vector_elements(type);
2505 const unsigned bit_size = glsl_get_bit_size(type);
2506
2507 /* First, zero-pad to ensure that the value is big enough that when we
2508 * bit-cast it, we don't loose anything.
2509 */
2510 if (val->bit_size < bit_size) {
2511 const unsigned src_num_components_needed =
2512 vtn_align_u32(val->num_components, bit_size / val->bit_size);
2513 val = nir_shrink_zero_pad_vec(b, val, src_num_components_needed);
2514 }
2515
2516 val = nir_bitcast_vector(b, val, bit_size);
2517
2518 return nir_shrink_zero_pad_vec(b, val, num_components);
2519 }
2520
2521 void
2522 vtn_handle_variables(struct vtn_builder *b, SpvOp opcode,
2523 const uint32_t *w, unsigned count)
2524 {
2525 switch (opcode) {
2526 case SpvOpUndef: {
2527 struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_undef);
2528 val->type = vtn_value(b, w[1], vtn_value_type_type)->type;
2529 break;
2530 }
2531
2532 case SpvOpVariable: {
2533 struct vtn_type *ptr_type = vtn_value(b, w[1], vtn_value_type_type)->type;
2534
2535 struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_pointer);
2536
2537 SpvStorageClass storage_class = w[3];
2538 nir_constant *const_initializer = NULL;
2539 nir_variable *var_initializer = NULL;
2540 if (count > 4) {
2541 struct vtn_value *init = vtn_untyped_value(b, w[4]);
2542 switch (init->value_type) {
2543 case vtn_value_type_constant:
2544 const_initializer = init->constant;
2545 break;
2546 case vtn_value_type_pointer:
2547 var_initializer = init->pointer->var->var;
2548 break;
2549 default:
2550 vtn_fail("SPIR-V variable initializer %u must be constant or pointer",
2551 w[4]);
2552 }
2553 }
2554
2555 vtn_create_variable(b, val, ptr_type, storage_class, const_initializer, var_initializer);
2556
2557 break;
2558 }
2559
2560 case SpvOpAccessChain:
2561 case SpvOpPtrAccessChain:
2562 case SpvOpInBoundsAccessChain:
2563 case SpvOpInBoundsPtrAccessChain: {
2564 struct vtn_access_chain *chain = vtn_access_chain_create(b, count - 4);
2565 enum gl_access_qualifier access = 0;
2566 chain->ptr_as_array = (opcode == SpvOpPtrAccessChain || opcode == SpvOpInBoundsPtrAccessChain);
2567
2568 unsigned idx = 0;
2569 for (int i = 4; i < count; i++) {
2570 struct vtn_value *link_val = vtn_untyped_value(b, w[i]);
2571 if (link_val->value_type == vtn_value_type_constant) {
2572 chain->link[idx].mode = vtn_access_mode_literal;
2573 chain->link[idx].id = vtn_constant_int(b, w[i]);
2574 } else {
2575 chain->link[idx].mode = vtn_access_mode_id;
2576 chain->link[idx].id = w[i];
2577 }
2578 access |= vtn_value_access(link_val);
2579 idx++;
2580 }
2581
2582 struct vtn_type *ptr_type = vtn_value(b, w[1], vtn_value_type_type)->type;
2583 struct vtn_value *base_val = vtn_untyped_value(b, w[3]);
2584 if (base_val->value_type == vtn_value_type_sampled_image) {
2585 /* This is rather insane. SPIR-V allows you to use OpSampledImage
2586 * to combine an array of images with a single sampler to get an
2587 * array of sampled images that all share the same sampler.
2588 * Fortunately, this means that we can more-or-less ignore the
2589 * sampler when crawling the access chain, but it does leave us
2590 * with this rather awkward little special-case.
2591 */
2592 struct vtn_value *val =
2593 vtn_push_value(b, w[2], vtn_value_type_sampled_image);
2594 val->sampled_image = ralloc(b, struct vtn_sampled_image);
2595 val->sampled_image->image =
2596 vtn_pointer_dereference(b, base_val->sampled_image->image, chain);
2597 val->sampled_image->sampler = base_val->sampled_image->sampler;
2598 val->sampled_image->image =
2599 vtn_decorate_pointer(b, val, val->sampled_image->image);
2600 val->sampled_image->sampler =
2601 vtn_decorate_pointer(b, val, val->sampled_image->sampler);
2602 } else {
2603 vtn_assert(base_val->value_type == vtn_value_type_pointer);
2604 struct vtn_pointer *ptr =
2605 vtn_pointer_dereference(b, base_val->pointer, chain);
2606 ptr->ptr_type = ptr_type;
2607 ptr->access |= access;
2608 vtn_push_value_pointer(b, w[2], ptr);
2609 }
2610 break;
2611 }
2612
2613 case SpvOpCopyMemory: {
2614 struct vtn_value *dest = vtn_value(b, w[1], vtn_value_type_pointer);
2615 struct vtn_value *src = vtn_value(b, w[2], vtn_value_type_pointer);
2616
2617 vtn_assert_types_equal(b, opcode, dest->type->deref, src->type->deref);
2618
2619 vtn_variable_copy(b, dest->pointer, src->pointer);
2620 break;
2621 }
2622
2623 case SpvOpLoad: {
2624 struct vtn_type *res_type =
2625 vtn_value(b, w[1], vtn_value_type_type)->type;
2626 struct vtn_value *src_val = vtn_value(b, w[3], vtn_value_type_pointer);
2627 struct vtn_pointer *src = src_val->pointer;
2628
2629 vtn_assert_types_equal(b, opcode, res_type, src_val->type->deref);
2630
2631 if (res_type->base_type == vtn_base_type_image ||
2632 res_type->base_type == vtn_base_type_sampler) {
2633 vtn_push_value_pointer(b, w[2], src);
2634 return;
2635 } else if (res_type->base_type == vtn_base_type_sampled_image) {
2636 struct vtn_value *val =
2637 vtn_push_value(b, w[2], vtn_value_type_sampled_image);
2638 val->sampled_image = ralloc(b, struct vtn_sampled_image);
2639 val->sampled_image->image = val->sampled_image->sampler =
2640 vtn_decorate_pointer(b, val, src);
2641 return;
2642 }
2643
2644 if (count > 4) {
2645 unsigned idx = 5;
2646 SpvMemoryAccessMask access = w[4];
2647 if (access & SpvMemoryAccessAlignedMask)
2648 idx++;
2649
2650 if (access & SpvMemoryAccessMakePointerVisibleMask) {
2651 SpvMemorySemanticsMask semantics =
2652 SpvMemorySemanticsMakeVisibleMask |
2653 vtn_storage_class_to_memory_semantics(src->ptr_type->storage_class);
2654
2655 SpvScope scope = vtn_constant_uint(b, w[idx]);
2656 vtn_emit_memory_barrier(b, scope, semantics);
2657 }
2658 }
2659
2660 vtn_push_ssa(b, w[2], res_type, vtn_variable_load(b, src));
2661 break;
2662 }
2663
2664 case SpvOpStore: {
2665 struct vtn_value *dest_val = vtn_value(b, w[1], vtn_value_type_pointer);
2666 struct vtn_pointer *dest = dest_val->pointer;
2667 struct vtn_value *src_val = vtn_untyped_value(b, w[2]);
2668
2669 /* OpStore requires us to actually have a storage type */
2670 vtn_fail_if(dest->type->type == NULL,
2671 "Invalid destination type for OpStore");
2672
2673 if (glsl_get_base_type(dest->type->type) == GLSL_TYPE_BOOL &&
2674 glsl_get_base_type(src_val->type->type) == GLSL_TYPE_UINT) {
2675 /* Early versions of GLSLang would use uint types for UBOs/SSBOs but
2676 * would then store them to a local variable as bool. Work around
2677 * the issue by doing an implicit conversion.
2678 *
2679 * https://github.com/KhronosGroup/glslang/issues/170
2680 * https://bugs.freedesktop.org/show_bug.cgi?id=104424
2681 */
2682 vtn_warn("OpStore of value of type OpTypeInt to a pointer to type "
2683 "OpTypeBool. Doing an implicit conversion to work around "
2684 "the problem.");
2685 struct vtn_ssa_value *bool_ssa =
2686 vtn_create_ssa_value(b, dest->type->type);
2687 bool_ssa->def = nir_i2b(&b->nb, vtn_ssa_value(b, w[2])->def);
2688 vtn_variable_store(b, bool_ssa, dest);
2689 break;
2690 }
2691
2692 vtn_assert_types_equal(b, opcode, dest_val->type->deref, src_val->type);
2693
2694 if (glsl_type_is_sampler(dest->type->type)) {
2695 if (b->wa_glslang_179) {
2696 vtn_warn("OpStore of a sampler detected. Doing on-the-fly copy "
2697 "propagation to workaround the problem.");
2698 vtn_assert(dest->var->copy_prop_sampler == NULL);
2699 struct vtn_value *v = vtn_untyped_value(b, w[2]);
2700 if (v->value_type == vtn_value_type_sampled_image) {
2701 dest->var->copy_prop_sampler = v->sampled_image->sampler;
2702 } else {
2703 vtn_assert(v->value_type == vtn_value_type_pointer);
2704 dest->var->copy_prop_sampler = v->pointer;
2705 }
2706 } else {
2707 vtn_fail("Vulkan does not allow OpStore of a sampler or image.");
2708 }
2709 break;
2710 }
2711
2712 struct vtn_ssa_value *src = vtn_ssa_value(b, w[2]);
2713 vtn_variable_store(b, src, dest);
2714
2715 if (count > 3) {
2716 unsigned idx = 4;
2717 SpvMemoryAccessMask access = w[3];
2718
2719 if (access & SpvMemoryAccessAlignedMask)
2720 idx++;
2721
2722 if (access & SpvMemoryAccessMakePointerAvailableMask) {
2723 SpvMemorySemanticsMask semantics =
2724 SpvMemorySemanticsMakeAvailableMask |
2725 vtn_storage_class_to_memory_semantics(dest->ptr_type->storage_class);
2726 SpvScope scope = vtn_constant_uint(b, w[idx]);
2727 vtn_emit_memory_barrier(b, scope, semantics);
2728 }
2729 }
2730 break;
2731 }
2732
2733 case SpvOpArrayLength: {
2734 struct vtn_pointer *ptr =
2735 vtn_value(b, w[3], vtn_value_type_pointer)->pointer;
2736 const uint32_t field = w[4];
2737
2738 vtn_fail_if(ptr->type->base_type != vtn_base_type_struct,
2739 "OpArrayLength must take a pointer to a structure type");
2740 vtn_fail_if(field != ptr->type->length - 1 ||
2741 ptr->type->members[field]->base_type != vtn_base_type_array,
2742 "OpArrayLength must reference the last memeber of the "
2743 "structure and that must be an array");
2744
2745 const uint32_t offset = ptr->type->offsets[field];
2746 const uint32_t stride = ptr->type->members[field]->stride;
2747
2748 if (!ptr->block_index) {
2749 struct vtn_access_chain chain = {
2750 .length = 0,
2751 };
2752 ptr = vtn_pointer_dereference(b, ptr, &chain);
2753 vtn_assert(ptr->block_index);
2754 }
2755
2756 nir_intrinsic_instr *instr =
2757 nir_intrinsic_instr_create(b->nb.shader,
2758 nir_intrinsic_get_buffer_size);
2759 instr->src[0] = nir_src_for_ssa(ptr->block_index);
2760 nir_ssa_dest_init(&instr->instr, &instr->dest, 1, 32, NULL);
2761 nir_builder_instr_insert(&b->nb, &instr->instr);
2762 nir_ssa_def *buf_size = &instr->dest.ssa;
2763
2764 /* array_length = max(buffer_size - offset, 0) / stride */
2765 nir_ssa_def *array_length =
2766 nir_idiv(&b->nb,
2767 nir_imax(&b->nb,
2768 nir_isub(&b->nb,
2769 buf_size,
2770 nir_imm_int(&b->nb, offset)),
2771 nir_imm_int(&b->nb, 0u)),
2772 nir_imm_int(&b->nb, stride));
2773
2774 struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_ssa);
2775 val->ssa = vtn_create_ssa_value(b, glsl_uint_type());
2776 val->ssa->def = array_length;
2777 break;
2778 }
2779
2780 case SpvOpConvertPtrToU: {
2781 struct vtn_value *u_val = vtn_push_value(b, w[2], vtn_value_type_ssa);
2782
2783 vtn_fail_if(u_val->type->base_type != vtn_base_type_vector &&
2784 u_val->type->base_type != vtn_base_type_scalar,
2785 "OpConvertPtrToU can only be used to cast to a vector or "
2786 "scalar type");
2787
2788 /* The pointer will be converted to an SSA value automatically */
2789 struct vtn_ssa_value *ptr_ssa = vtn_ssa_value(b, w[3]);
2790
2791 u_val->ssa = vtn_create_ssa_value(b, u_val->type->type);
2792 u_val->ssa->def = nir_sloppy_bitcast(&b->nb, ptr_ssa->def, u_val->type->type);
2793 u_val->ssa->access |= ptr_ssa->access;
2794 break;
2795 }
2796
2797 case SpvOpConvertUToPtr: {
2798 struct vtn_value *ptr_val =
2799 vtn_push_value(b, w[2], vtn_value_type_pointer);
2800 struct vtn_value *u_val = vtn_untyped_value(b, w[3]);
2801
2802 vtn_fail_if(ptr_val->type->type == NULL,
2803 "OpConvertUToPtr can only be used on physical pointers");
2804
2805 vtn_fail_if(u_val->type->base_type != vtn_base_type_vector &&
2806 u_val->type->base_type != vtn_base_type_scalar,
2807 "OpConvertUToPtr can only be used to cast from a vector or "
2808 "scalar type");
2809
2810 struct vtn_ssa_value *u_ssa = vtn_ssa_value(b, w[3]);
2811 nir_ssa_def *ptr_ssa = nir_sloppy_bitcast(&b->nb, u_ssa->def,
2812 ptr_val->type->type);
2813 ptr_val->pointer = vtn_pointer_from_ssa(b, ptr_ssa, ptr_val->type);
2814 vtn_foreach_decoration(b, ptr_val, ptr_decoration_cb, ptr_val->pointer);
2815 ptr_val->pointer->access |= u_val->ssa->access;
2816 break;
2817 }
2818
2819 case SpvOpCopyMemorySized:
2820 default:
2821 vtn_fail_with_opcode("Unhandled opcode", opcode);
2822 }
2823 }