1 /*
2 * Copyright © 2015 Intel Corporation
3 *
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
10 *
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
13 * Software.
14 *
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
21 * IN THE SOFTWARE.
22 *
23 * Authors:
24 * Jason Ekstrand (jason@jlekstrand.net)
25 *
26 */
27
28 #include "vtn_private.h"
29 #include "spirv_info.h"
30
31 static struct vtn_access_chain *
vtn_access_chain_create(struct vtn_builder * b,unsigned length)32 vtn_access_chain_create(struct vtn_builder *b, unsigned length)
33 {
34 struct vtn_access_chain *chain;
35
36 /* Subtract 1 from the length since there's already one built in */
37 size_t size = sizeof(*chain) +
38 (MAX2(length, 1) - 1) * sizeof(chain->link[0]);
39 chain = rzalloc_size(b, size);
40 chain->length = length;
41
42 return chain;
43 }
44
45 static struct vtn_access_chain *
vtn_access_chain_extend(struct vtn_builder * b,struct vtn_access_chain * old,unsigned new_ids)46 vtn_access_chain_extend(struct vtn_builder *b, struct vtn_access_chain *old,
47 unsigned new_ids)
48 {
49 struct vtn_access_chain *chain;
50
51 unsigned old_len = old ? old->length : 0;
52 chain = vtn_access_chain_create(b, old_len + new_ids);
53
54 for (unsigned i = 0; i < old_len; i++)
55 chain->link[i] = old->link[i];
56
57 return chain;
58 }
59
60 static bool
vtn_pointer_uses_ssa_offset(struct vtn_builder * b,struct vtn_pointer * ptr)61 vtn_pointer_uses_ssa_offset(struct vtn_builder *b,
62 struct vtn_pointer *ptr)
63 {
64 return ptr->mode == vtn_variable_mode_ubo ||
65 ptr->mode == vtn_variable_mode_ssbo ||
66 (ptr->mode == vtn_variable_mode_workgroup &&
67 b->options->lower_workgroup_access_to_offsets);
68 }
69
70 static bool
vtn_pointer_is_external_block(struct vtn_builder * b,struct vtn_pointer * ptr)71 vtn_pointer_is_external_block(struct vtn_builder *b,
72 struct vtn_pointer *ptr)
73 {
74 return ptr->mode == vtn_variable_mode_ssbo ||
75 ptr->mode == vtn_variable_mode_ubo ||
76 ptr->mode == vtn_variable_mode_push_constant ||
77 (ptr->mode == vtn_variable_mode_workgroup &&
78 b->options->lower_workgroup_access_to_offsets);
79 }
80
81 /* Dereference the given base pointer by the access chain */
82 static struct vtn_pointer *
vtn_access_chain_pointer_dereference(struct vtn_builder * b,struct vtn_pointer * base,struct vtn_access_chain * deref_chain)83 vtn_access_chain_pointer_dereference(struct vtn_builder *b,
84 struct vtn_pointer *base,
85 struct vtn_access_chain *deref_chain)
86 {
87 struct vtn_access_chain *chain =
88 vtn_access_chain_extend(b, base->chain, deref_chain->length);
89 struct vtn_type *type = base->type;
90
91 /* OpPtrAccessChain is only allowed on things which support variable
92 * pointers. For everything else, the client is expected to just pass us
93 * the right access chain.
94 */
95 vtn_assert(!deref_chain->ptr_as_array);
96
97 unsigned start = base->chain ? base->chain->length : 0;
98 for (unsigned i = 0; i < deref_chain->length; i++) {
99 chain->link[start + i] = deref_chain->link[i];
100
101 if (glsl_type_is_struct(type->type)) {
102 vtn_assert(deref_chain->link[i].mode == vtn_access_mode_literal);
103 type = type->members[deref_chain->link[i].id];
104 } else {
105 type = type->array_element;
106 }
107 }
108
109 struct vtn_pointer *ptr = rzalloc(b, struct vtn_pointer);
110 ptr->mode = base->mode;
111 ptr->type = type;
112 ptr->var = base->var;
113 ptr->chain = chain;
114
115 return ptr;
116 }
117
118 static nir_ssa_def *
vtn_access_link_as_ssa(struct vtn_builder * b,struct vtn_access_link link,unsigned stride)119 vtn_access_link_as_ssa(struct vtn_builder *b, struct vtn_access_link link,
120 unsigned stride)
121 {
122 vtn_assert(stride > 0);
123 if (link.mode == vtn_access_mode_literal) {
124 return nir_imm_int(&b->nb, link.id * stride);
125 } else if (stride == 1) {
126 nir_ssa_def *ssa = vtn_ssa_value(b, link.id)->def;
127 if (ssa->bit_size != 32)
128 ssa = nir_u2u32(&b->nb, ssa);
129 return ssa;
130 } else {
131 nir_ssa_def *src0 = vtn_ssa_value(b, link.id)->def;
132 if (src0->bit_size != 32)
133 src0 = nir_u2u32(&b->nb, src0);
134 return nir_imul(&b->nb, src0, nir_imm_int(&b->nb, stride));
135 }
136 }
137
138 static nir_ssa_def *
vtn_variable_resource_index(struct vtn_builder * b,struct vtn_variable * var,nir_ssa_def * desc_array_index)139 vtn_variable_resource_index(struct vtn_builder *b, struct vtn_variable *var,
140 nir_ssa_def *desc_array_index)
141 {
142 if (!desc_array_index) {
143 vtn_assert(glsl_type_is_struct(var->type->type));
144 desc_array_index = nir_imm_int(&b->nb, 0);
145 }
146
147 nir_intrinsic_instr *instr =
148 nir_intrinsic_instr_create(b->nb.shader,
149 nir_intrinsic_vulkan_resource_index);
150 instr->src[0] = nir_src_for_ssa(desc_array_index);
151 nir_intrinsic_set_desc_set(instr, var->descriptor_set);
152 nir_intrinsic_set_binding(instr, var->binding);
153
154 nir_ssa_dest_init(&instr->instr, &instr->dest, 1, 32, NULL);
155 nir_builder_instr_insert(&b->nb, &instr->instr);
156
157 return &instr->dest.ssa;
158 }
159
160 static nir_ssa_def *
vtn_resource_reindex(struct vtn_builder * b,nir_ssa_def * base_index,nir_ssa_def * offset_index)161 vtn_resource_reindex(struct vtn_builder *b, nir_ssa_def *base_index,
162 nir_ssa_def *offset_index)
163 {
164 nir_intrinsic_instr *instr =
165 nir_intrinsic_instr_create(b->nb.shader,
166 nir_intrinsic_vulkan_resource_reindex);
167 instr->src[0] = nir_src_for_ssa(base_index);
168 instr->src[1] = nir_src_for_ssa(offset_index);
169
170 nir_ssa_dest_init(&instr->instr, &instr->dest, 1, 32, NULL);
171 nir_builder_instr_insert(&b->nb, &instr->instr);
172
173 return &instr->dest.ssa;
174 }
175
176 static struct vtn_pointer *
vtn_ssa_offset_pointer_dereference(struct vtn_builder * b,struct vtn_pointer * base,struct vtn_access_chain * deref_chain)177 vtn_ssa_offset_pointer_dereference(struct vtn_builder *b,
178 struct vtn_pointer *base,
179 struct vtn_access_chain *deref_chain)
180 {
181 nir_ssa_def *block_index = base->block_index;
182 nir_ssa_def *offset = base->offset;
183 struct vtn_type *type = base->type;
184
185 unsigned idx = 0;
186 if (base->mode == vtn_variable_mode_ubo ||
187 base->mode == vtn_variable_mode_ssbo) {
188 if (!block_index) {
189 vtn_assert(base->var && base->type);
190 nir_ssa_def *desc_arr_idx;
191 if (glsl_type_is_array(type->type)) {
192 if (deref_chain->length >= 1) {
193 desc_arr_idx =
194 vtn_access_link_as_ssa(b, deref_chain->link[0], 1);
195 idx++;
196 /* This consumes a level of type */
197 type = type->array_element;
198 } else {
199 /* This is annoying. We've been asked for a pointer to the
200 * array of UBOs/SSBOs and not a specifc buffer. Return a
201 * pointer with a descriptor index of 0 and we'll have to do
202 * a reindex later to adjust it to the right thing.
203 */
204 desc_arr_idx = nir_imm_int(&b->nb, 0);
205 }
206 } else if (deref_chain->ptr_as_array) {
207 /* You can't have a zero-length OpPtrAccessChain */
208 vtn_assert(deref_chain->length >= 1);
209 desc_arr_idx = vtn_access_link_as_ssa(b, deref_chain->link[0], 1);
210 } else {
211 /* We have a regular non-array SSBO. */
212 desc_arr_idx = NULL;
213 }
214 block_index = vtn_variable_resource_index(b, base->var, desc_arr_idx);
215 } else if (deref_chain->ptr_as_array &&
216 type->base_type == vtn_base_type_struct && type->block) {
217 /* We are doing an OpPtrAccessChain on a pointer to a struct that is
218 * decorated block. This is an interesting corner in the SPIR-V
219 * spec. One interpretation would be that they client is clearly
220 * trying to treat that block as if it's an implicit array of blocks
221 * repeated in the buffer. However, the SPIR-V spec for the
222 * OpPtrAccessChain says:
223 *
224 * "Base is treated as the address of the first element of an
225 * array, and the Element element’s address is computed to be the
226 * base for the Indexes, as per OpAccessChain."
227 *
228 * Taken literally, that would mean that your struct type is supposed
229 * to be treated as an array of such a struct and, since it's
230 * decorated block, that means an array of blocks which corresponds
231 * to an array descriptor. Therefore, we need to do a reindex
232 * operation to add the index from the first link in the access chain
233 * to the index we recieved.
234 *
235 * The downside to this interpretation (there always is one) is that
236 * this might be somewhat surprising behavior to apps if they expect
237 * the implicit array behavior described above.
238 */
239 vtn_assert(deref_chain->length >= 1);
240 nir_ssa_def *offset_index =
241 vtn_access_link_as_ssa(b, deref_chain->link[0], 1);
242 idx++;
243
244 block_index = vtn_resource_reindex(b, block_index, offset_index);
245 }
246 }
247
248 if (!offset) {
249 if (base->mode == vtn_variable_mode_workgroup) {
250 /* SLM doesn't need nor have a block index */
251 vtn_assert(!block_index);
252
253 /* We need the variable for the base offset */
254 vtn_assert(base->var);
255
256 /* We need ptr_type for size and alignment */
257 vtn_assert(base->ptr_type);
258
259 /* Assign location on first use so that we don't end up bloating SLM
260 * address space for variables which are never statically used.
261 */
262 if (base->var->shared_location < 0) {
263 vtn_assert(base->ptr_type->length > 0 && base->ptr_type->align > 0);
264 b->shader->num_shared = vtn_align_u32(b->shader->num_shared,
265 base->ptr_type->align);
266 base->var->shared_location = b->shader->num_shared;
267 b->shader->num_shared += base->ptr_type->length;
268 }
269
270 offset = nir_imm_int(&b->nb, base->var->shared_location);
271 } else {
272 /* The code above should have ensured a block_index when needed. */
273 vtn_assert(block_index);
274
275 /* Start off with at the start of the buffer. */
276 offset = nir_imm_int(&b->nb, 0);
277 }
278 }
279
280 if (deref_chain->ptr_as_array && idx == 0) {
281 /* We need ptr_type for the stride */
282 vtn_assert(base->ptr_type);
283
284 /* We need at least one element in the chain */
285 vtn_assert(deref_chain->length >= 1);
286
287 nir_ssa_def *elem_offset =
288 vtn_access_link_as_ssa(b, deref_chain->link[idx],
289 base->ptr_type->stride);
290 offset = nir_iadd(&b->nb, offset, elem_offset);
291 idx++;
292 }
293
294 for (; idx < deref_chain->length; idx++) {
295 switch (glsl_get_base_type(type->type)) {
296 case GLSL_TYPE_UINT:
297 case GLSL_TYPE_INT:
298 case GLSL_TYPE_UINT16:
299 case GLSL_TYPE_INT16:
300 case GLSL_TYPE_UINT64:
301 case GLSL_TYPE_INT64:
302 case GLSL_TYPE_FLOAT:
303 case GLSL_TYPE_FLOAT16:
304 case GLSL_TYPE_DOUBLE:
305 case GLSL_TYPE_BOOL:
306 case GLSL_TYPE_ARRAY: {
307 nir_ssa_def *elem_offset =
308 vtn_access_link_as_ssa(b, deref_chain->link[idx], type->stride);
309 offset = nir_iadd(&b->nb, offset, elem_offset);
310 type = type->array_element;
311 break;
312 }
313
314 case GLSL_TYPE_STRUCT: {
315 vtn_assert(deref_chain->link[idx].mode == vtn_access_mode_literal);
316 unsigned member = deref_chain->link[idx].id;
317 nir_ssa_def *mem_offset = nir_imm_int(&b->nb, type->offsets[member]);
318 offset = nir_iadd(&b->nb, offset, mem_offset);
319 type = type->members[member];
320 break;
321 }
322
323 default:
324 vtn_fail("Invalid type for deref");
325 }
326 }
327
328 struct vtn_pointer *ptr = rzalloc(b, struct vtn_pointer);
329 ptr->mode = base->mode;
330 ptr->type = type;
331 ptr->block_index = block_index;
332 ptr->offset = offset;
333
334 return ptr;
335 }
336
337 /* Dereference the given base pointer by the access chain */
338 static struct vtn_pointer *
vtn_pointer_dereference(struct vtn_builder * b,struct vtn_pointer * base,struct vtn_access_chain * deref_chain)339 vtn_pointer_dereference(struct vtn_builder *b,
340 struct vtn_pointer *base,
341 struct vtn_access_chain *deref_chain)
342 {
343 if (vtn_pointer_uses_ssa_offset(b, base)) {
344 return vtn_ssa_offset_pointer_dereference(b, base, deref_chain);
345 } else {
346 return vtn_access_chain_pointer_dereference(b, base, deref_chain);
347 }
348 }
349
350 /* Crawls a chain of array derefs and rewrites the types so that the
351 * lengths stay the same but the terminal type is the one given by
352 * tail_type. This is useful for split structures.
353 */
354 static void
rewrite_deref_types(struct vtn_builder * b,nir_deref * deref,const struct glsl_type * type)355 rewrite_deref_types(struct vtn_builder *b, nir_deref *deref,
356 const struct glsl_type *type)
357 {
358 deref->type = type;
359 if (deref->child) {
360 vtn_assert(deref->child->deref_type == nir_deref_type_array);
361 vtn_assert(glsl_type_is_array(deref->type));
362 rewrite_deref_types(b, deref->child, glsl_get_array_element(type));
363 }
364 }
365
366 struct vtn_pointer *
vtn_pointer_for_variable(struct vtn_builder * b,struct vtn_variable * var,struct vtn_type * ptr_type)367 vtn_pointer_for_variable(struct vtn_builder *b,
368 struct vtn_variable *var, struct vtn_type *ptr_type)
369 {
370 struct vtn_pointer *pointer = rzalloc(b, struct vtn_pointer);
371
372 pointer->mode = var->mode;
373 pointer->type = var->type;
374 vtn_assert(ptr_type->base_type == vtn_base_type_pointer);
375 vtn_assert(ptr_type->deref->type == var->type->type);
376 pointer->ptr_type = ptr_type;
377 pointer->var = var;
378
379 return pointer;
380 }
381
382 nir_deref_var *
vtn_pointer_to_deref(struct vtn_builder * b,struct vtn_pointer * ptr)383 vtn_pointer_to_deref(struct vtn_builder *b, struct vtn_pointer *ptr)
384 {
385 /* Do on-the-fly copy propagation for samplers. */
386 if (ptr->var->copy_prop_sampler)
387 return vtn_pointer_to_deref(b, ptr->var->copy_prop_sampler);
388
389 nir_deref_var *deref_var;
390 if (ptr->var->var) {
391 deref_var = nir_deref_var_create(b, ptr->var->var);
392 /* Raw variable access */
393 if (!ptr->chain)
394 return deref_var;
395 } else {
396 vtn_assert(ptr->var->members);
397 /* Create the deref_var manually. It will get filled out later. */
398 deref_var = rzalloc(b, nir_deref_var);
399 deref_var->deref.deref_type = nir_deref_type_var;
400 }
401
402 struct vtn_access_chain *chain = ptr->chain;
403 vtn_assert(chain);
404
405 struct vtn_type *deref_type = ptr->var->type;
406 nir_deref *tail = &deref_var->deref;
407 nir_variable **members = ptr->var->members;
408
409 for (unsigned i = 0; i < chain->length; i++) {
410 enum glsl_base_type base_type = glsl_get_base_type(deref_type->type);
411 switch (base_type) {
412 case GLSL_TYPE_UINT:
413 case GLSL_TYPE_INT:
414 case GLSL_TYPE_UINT16:
415 case GLSL_TYPE_INT16:
416 case GLSL_TYPE_UINT64:
417 case GLSL_TYPE_INT64:
418 case GLSL_TYPE_FLOAT:
419 case GLSL_TYPE_FLOAT16:
420 case GLSL_TYPE_DOUBLE:
421 case GLSL_TYPE_BOOL:
422 case GLSL_TYPE_ARRAY: {
423 deref_type = deref_type->array_element;
424
425 nir_deref_array *deref_arr = nir_deref_array_create(b);
426 deref_arr->deref.type = deref_type->type;
427
428 if (chain->link[i].mode == vtn_access_mode_literal) {
429 deref_arr->deref_array_type = nir_deref_array_type_direct;
430 deref_arr->base_offset = chain->link[i].id;
431 } else {
432 vtn_assert(chain->link[i].mode == vtn_access_mode_id);
433 deref_arr->deref_array_type = nir_deref_array_type_indirect;
434 deref_arr->base_offset = 0;
435 deref_arr->indirect =
436 nir_src_for_ssa(vtn_ssa_value(b, chain->link[i].id)->def);
437 }
438 tail->child = &deref_arr->deref;
439 tail = tail->child;
440 break;
441 }
442
443 case GLSL_TYPE_STRUCT: {
444 vtn_assert(chain->link[i].mode == vtn_access_mode_literal);
445 unsigned idx = chain->link[i].id;
446 deref_type = deref_type->members[idx];
447 if (members) {
448 /* This is a pre-split structure. */
449 deref_var->var = members[idx];
450 rewrite_deref_types(b, &deref_var->deref, members[idx]->type);
451 vtn_assert(tail->type == deref_type->type);
452 members = NULL;
453 } else {
454 nir_deref_struct *deref_struct = nir_deref_struct_create(b, idx);
455 deref_struct->deref.type = deref_type->type;
456 tail->child = &deref_struct->deref;
457 tail = tail->child;
458 }
459 break;
460 }
461 default:
462 vtn_fail("Invalid type for deref");
463 }
464 }
465
466 vtn_assert(members == NULL);
467 return deref_var;
468 }
469
470 static void
_vtn_local_load_store(struct vtn_builder * b,bool load,nir_deref_var * deref,nir_deref * tail,struct vtn_ssa_value * inout)471 _vtn_local_load_store(struct vtn_builder *b, bool load, nir_deref_var *deref,
472 nir_deref *tail, struct vtn_ssa_value *inout)
473 {
474 /* The deref tail may contain a deref to select a component of a vector (in
475 * other words, it might not be an actual tail) so we have to save it away
476 * here since we overwrite it later.
477 */
478 nir_deref *old_child = tail->child;
479
480 if (glsl_type_is_vector_or_scalar(tail->type)) {
481 /* Terminate the deref chain in case there is one more link to pick
482 * off a component of the vector.
483 */
484 tail->child = NULL;
485
486 nir_intrinsic_op op = load ? nir_intrinsic_load_var :
487 nir_intrinsic_store_var;
488
489 nir_intrinsic_instr *intrin = nir_intrinsic_instr_create(b->shader, op);
490 intrin->variables[0] = nir_deref_var_clone(deref, intrin);
491 intrin->num_components = glsl_get_vector_elements(tail->type);
492
493 if (load) {
494 nir_ssa_dest_init(&intrin->instr, &intrin->dest,
495 intrin->num_components,
496 glsl_get_bit_size(tail->type),
497 NULL);
498 inout->def = &intrin->dest.ssa;
499 } else {
500 nir_intrinsic_set_write_mask(intrin, (1 << intrin->num_components) - 1);
501 intrin->src[0] = nir_src_for_ssa(inout->def);
502 }
503
504 nir_builder_instr_insert(&b->nb, &intrin->instr);
505 } else if (glsl_get_base_type(tail->type) == GLSL_TYPE_ARRAY ||
506 glsl_type_is_matrix(tail->type)) {
507 unsigned elems = glsl_get_length(tail->type);
508 nir_deref_array *deref_arr = nir_deref_array_create(b);
509 deref_arr->deref_array_type = nir_deref_array_type_direct;
510 deref_arr->deref.type = glsl_get_array_element(tail->type);
511 tail->child = &deref_arr->deref;
512 for (unsigned i = 0; i < elems; i++) {
513 deref_arr->base_offset = i;
514 _vtn_local_load_store(b, load, deref, tail->child, inout->elems[i]);
515 }
516 } else {
517 vtn_assert(glsl_get_base_type(tail->type) == GLSL_TYPE_STRUCT);
518 unsigned elems = glsl_get_length(tail->type);
519 nir_deref_struct *deref_struct = nir_deref_struct_create(b, 0);
520 tail->child = &deref_struct->deref;
521 for (unsigned i = 0; i < elems; i++) {
522 deref_struct->index = i;
523 deref_struct->deref.type = glsl_get_struct_field(tail->type, i);
524 _vtn_local_load_store(b, load, deref, tail->child, inout->elems[i]);
525 }
526 }
527
528 tail->child = old_child;
529 }
530
531 nir_deref_var *
vtn_nir_deref(struct vtn_builder * b,uint32_t id)532 vtn_nir_deref(struct vtn_builder *b, uint32_t id)
533 {
534 struct vtn_pointer *ptr = vtn_value(b, id, vtn_value_type_pointer)->pointer;
535 return vtn_pointer_to_deref(b, ptr);
536 }
537
538 /*
539 * Gets the NIR-level deref tail, which may have as a child an array deref
540 * selecting which component due to OpAccessChain supporting per-component
541 * indexing in SPIR-V.
542 */
543 static nir_deref *
get_deref_tail(nir_deref_var * deref)544 get_deref_tail(nir_deref_var *deref)
545 {
546 nir_deref *cur = &deref->deref;
547 while (!glsl_type_is_vector_or_scalar(cur->type) && cur->child)
548 cur = cur->child;
549
550 return cur;
551 }
552
553 struct vtn_ssa_value *
vtn_local_load(struct vtn_builder * b,nir_deref_var * src)554 vtn_local_load(struct vtn_builder *b, nir_deref_var *src)
555 {
556 nir_deref *src_tail = get_deref_tail(src);
557 struct vtn_ssa_value *val = vtn_create_ssa_value(b, src_tail->type);
558 _vtn_local_load_store(b, true, src, src_tail, val);
559
560 if (src_tail->child) {
561 nir_deref_array *vec_deref = nir_deref_as_array(src_tail->child);
562 vtn_assert(vec_deref->deref.child == NULL);
563 val->type = vec_deref->deref.type;
564 if (vec_deref->deref_array_type == nir_deref_array_type_direct)
565 val->def = vtn_vector_extract(b, val->def, vec_deref->base_offset);
566 else
567 val->def = vtn_vector_extract_dynamic(b, val->def,
568 vec_deref->indirect.ssa);
569 }
570
571 return val;
572 }
573
574 void
vtn_local_store(struct vtn_builder * b,struct vtn_ssa_value * src,nir_deref_var * dest)575 vtn_local_store(struct vtn_builder *b, struct vtn_ssa_value *src,
576 nir_deref_var *dest)
577 {
578 nir_deref *dest_tail = get_deref_tail(dest);
579
580 if (dest_tail->child) {
581 struct vtn_ssa_value *val = vtn_create_ssa_value(b, dest_tail->type);
582 _vtn_local_load_store(b, true, dest, dest_tail, val);
583 nir_deref_array *deref = nir_deref_as_array(dest_tail->child);
584 vtn_assert(deref->deref.child == NULL);
585 if (deref->deref_array_type == nir_deref_array_type_direct)
586 val->def = vtn_vector_insert(b, val->def, src->def,
587 deref->base_offset);
588 else
589 val->def = vtn_vector_insert_dynamic(b, val->def, src->def,
590 deref->indirect.ssa);
591 _vtn_local_load_store(b, false, dest, dest_tail, val);
592 } else {
593 _vtn_local_load_store(b, false, dest, dest_tail, src);
594 }
595 }
596
597 nir_ssa_def *
vtn_pointer_to_offset(struct vtn_builder * b,struct vtn_pointer * ptr,nir_ssa_def ** index_out,unsigned * end_idx_out)598 vtn_pointer_to_offset(struct vtn_builder *b, struct vtn_pointer *ptr,
599 nir_ssa_def **index_out, unsigned *end_idx_out)
600 {
601 if (vtn_pointer_uses_ssa_offset(b, ptr)) {
602 if (!ptr->offset) {
603 struct vtn_access_chain chain = {
604 .length = 0,
605 };
606 ptr = vtn_ssa_offset_pointer_dereference(b, ptr, &chain);
607 }
608 *index_out = ptr->block_index;
609 return ptr->offset;
610 }
611
612 vtn_assert(ptr->mode == vtn_variable_mode_push_constant);
613 *index_out = NULL;
614
615 unsigned idx = 0;
616 struct vtn_type *type = ptr->var->type;
617 nir_ssa_def *offset = nir_imm_int(&b->nb, 0);
618
619 if (ptr->chain) {
620 for (; idx < ptr->chain->length; idx++) {
621 enum glsl_base_type base_type = glsl_get_base_type(type->type);
622 switch (base_type) {
623 case GLSL_TYPE_UINT:
624 case GLSL_TYPE_INT:
625 case GLSL_TYPE_UINT16:
626 case GLSL_TYPE_INT16:
627 case GLSL_TYPE_UINT64:
628 case GLSL_TYPE_INT64:
629 case GLSL_TYPE_FLOAT:
630 case GLSL_TYPE_FLOAT16:
631 case GLSL_TYPE_DOUBLE:
632 case GLSL_TYPE_BOOL:
633 case GLSL_TYPE_ARRAY:
634 offset = nir_iadd(&b->nb, offset,
635 vtn_access_link_as_ssa(b, ptr->chain->link[idx],
636 type->stride));
637
638 type = type->array_element;
639 break;
640
641 case GLSL_TYPE_STRUCT: {
642 vtn_assert(ptr->chain->link[idx].mode == vtn_access_mode_literal);
643 unsigned member = ptr->chain->link[idx].id;
644 offset = nir_iadd(&b->nb, offset,
645 nir_imm_int(&b->nb, type->offsets[member]));
646 type = type->members[member];
647 break;
648 }
649
650 default:
651 vtn_fail("Invalid type for deref");
652 }
653 }
654 }
655
656 vtn_assert(type == ptr->type);
657 if (end_idx_out)
658 *end_idx_out = idx;
659
660 return offset;
661 }
662
663 /* Tries to compute the size of an interface block based on the strides and
664 * offsets that are provided to us in the SPIR-V source.
665 */
666 static unsigned
vtn_type_block_size(struct vtn_builder * b,struct vtn_type * type)667 vtn_type_block_size(struct vtn_builder *b, struct vtn_type *type)
668 {
669 enum glsl_base_type base_type = glsl_get_base_type(type->type);
670 switch (base_type) {
671 case GLSL_TYPE_UINT:
672 case GLSL_TYPE_INT:
673 case GLSL_TYPE_UINT16:
674 case GLSL_TYPE_INT16:
675 case GLSL_TYPE_UINT64:
676 case GLSL_TYPE_INT64:
677 case GLSL_TYPE_FLOAT:
678 case GLSL_TYPE_FLOAT16:
679 case GLSL_TYPE_BOOL:
680 case GLSL_TYPE_DOUBLE: {
681 unsigned cols = type->row_major ? glsl_get_vector_elements(type->type) :
682 glsl_get_matrix_columns(type->type);
683 if (cols > 1) {
684 vtn_assert(type->stride > 0);
685 return type->stride * cols;
686 } else if (base_type == GLSL_TYPE_DOUBLE ||
687 base_type == GLSL_TYPE_UINT64 ||
688 base_type == GLSL_TYPE_INT64) {
689 return glsl_get_vector_elements(type->type) * 8;
690 } else {
691 return glsl_get_vector_elements(type->type) * 4;
692 }
693 }
694
695 case GLSL_TYPE_STRUCT:
696 case GLSL_TYPE_INTERFACE: {
697 unsigned size = 0;
698 unsigned num_fields = glsl_get_length(type->type);
699 for (unsigned f = 0; f < num_fields; f++) {
700 unsigned field_end = type->offsets[f] +
701 vtn_type_block_size(b, type->members[f]);
702 size = MAX2(size, field_end);
703 }
704 return size;
705 }
706
707 case GLSL_TYPE_ARRAY:
708 vtn_assert(type->stride > 0);
709 vtn_assert(glsl_get_length(type->type) > 0);
710 return type->stride * glsl_get_length(type->type);
711
712 default:
713 vtn_fail("Invalid block type");
714 return 0;
715 }
716 }
717
718 static void
vtn_access_chain_get_offset_size(struct vtn_builder * b,struct vtn_access_chain * chain,struct vtn_type * type,unsigned * access_offset,unsigned * access_size)719 vtn_access_chain_get_offset_size(struct vtn_builder *b,
720 struct vtn_access_chain *chain,
721 struct vtn_type *type,
722 unsigned *access_offset,
723 unsigned *access_size)
724 {
725 *access_offset = 0;
726
727 for (unsigned i = 0; i < chain->length; i++) {
728 if (chain->link[i].mode != vtn_access_mode_literal)
729 break;
730
731 if (glsl_type_is_struct(type->type)) {
732 *access_offset += type->offsets[chain->link[i].id];
733 type = type->members[chain->link[i].id];
734 } else {
735 *access_offset += type->stride * chain->link[i].id;
736 type = type->array_element;
737 }
738 }
739
740 *access_size = vtn_type_block_size(b, type);
741 }
742
743 static void
_vtn_load_store_tail(struct vtn_builder * b,nir_intrinsic_op op,bool load,nir_ssa_def * index,nir_ssa_def * offset,unsigned access_offset,unsigned access_size,struct vtn_ssa_value ** inout,const struct glsl_type * type)744 _vtn_load_store_tail(struct vtn_builder *b, nir_intrinsic_op op, bool load,
745 nir_ssa_def *index, nir_ssa_def *offset,
746 unsigned access_offset, unsigned access_size,
747 struct vtn_ssa_value **inout, const struct glsl_type *type)
748 {
749 nir_intrinsic_instr *instr = nir_intrinsic_instr_create(b->nb.shader, op);
750 instr->num_components = glsl_get_vector_elements(type);
751
752 int src = 0;
753 if (!load) {
754 nir_intrinsic_set_write_mask(instr, (1 << instr->num_components) - 1);
755 instr->src[src++] = nir_src_for_ssa((*inout)->def);
756 }
757
758 if (op == nir_intrinsic_load_push_constant) {
759 vtn_assert(access_offset % 4 == 0);
760
761 nir_intrinsic_set_base(instr, access_offset);
762 nir_intrinsic_set_range(instr, access_size);
763 }
764
765 if (index)
766 instr->src[src++] = nir_src_for_ssa(index);
767
768 if (op == nir_intrinsic_load_push_constant) {
769 /* We need to subtract the offset from where the intrinsic will load the
770 * data. */
771 instr->src[src++] =
772 nir_src_for_ssa(nir_isub(&b->nb, offset,
773 nir_imm_int(&b->nb, access_offset)));
774 } else {
775 instr->src[src++] = nir_src_for_ssa(offset);
776 }
777
778 if (load) {
779 nir_ssa_dest_init(&instr->instr, &instr->dest,
780 instr->num_components,
781 glsl_get_bit_size(type), NULL);
782 (*inout)->def = &instr->dest.ssa;
783 }
784
785 nir_builder_instr_insert(&b->nb, &instr->instr);
786
787 if (load && glsl_get_base_type(type) == GLSL_TYPE_BOOL)
788 (*inout)->def = nir_ine(&b->nb, (*inout)->def, nir_imm_int(&b->nb, 0));
789 }
790
791 static void
_vtn_block_load_store(struct vtn_builder * b,nir_intrinsic_op op,bool load,nir_ssa_def * index,nir_ssa_def * offset,unsigned access_offset,unsigned access_size,struct vtn_access_chain * chain,unsigned chain_idx,struct vtn_type * type,struct vtn_ssa_value ** inout)792 _vtn_block_load_store(struct vtn_builder *b, nir_intrinsic_op op, bool load,
793 nir_ssa_def *index, nir_ssa_def *offset,
794 unsigned access_offset, unsigned access_size,
795 struct vtn_access_chain *chain, unsigned chain_idx,
796 struct vtn_type *type, struct vtn_ssa_value **inout)
797 {
798 if (chain && chain_idx >= chain->length)
799 chain = NULL;
800
801 if (load && chain == NULL && *inout == NULL)
802 *inout = vtn_create_ssa_value(b, type->type);
803
804 enum glsl_base_type base_type = glsl_get_base_type(type->type);
805 switch (base_type) {
806 case GLSL_TYPE_UINT:
807 case GLSL_TYPE_INT:
808 case GLSL_TYPE_UINT16:
809 case GLSL_TYPE_INT16:
810 case GLSL_TYPE_UINT64:
811 case GLSL_TYPE_INT64:
812 case GLSL_TYPE_FLOAT:
813 case GLSL_TYPE_FLOAT16:
814 case GLSL_TYPE_DOUBLE:
815 case GLSL_TYPE_BOOL:
816 /* This is where things get interesting. At this point, we've hit
817 * a vector, a scalar, or a matrix.
818 */
819 if (glsl_type_is_matrix(type->type)) {
820 /* Loading the whole matrix */
821 struct vtn_ssa_value *transpose;
822 unsigned num_ops, vec_width, col_stride;
823 if (type->row_major) {
824 num_ops = glsl_get_vector_elements(type->type);
825 vec_width = glsl_get_matrix_columns(type->type);
826 col_stride = type->array_element->stride;
827 if (load) {
828 const struct glsl_type *transpose_type =
829 glsl_matrix_type(base_type, vec_width, num_ops);
830 *inout = vtn_create_ssa_value(b, transpose_type);
831 } else {
832 transpose = vtn_ssa_transpose(b, *inout);
833 inout = &transpose;
834 }
835 } else {
836 num_ops = glsl_get_matrix_columns(type->type);
837 vec_width = glsl_get_vector_elements(type->type);
838 col_stride = type->stride;
839 }
840
841 for (unsigned i = 0; i < num_ops; i++) {
842 nir_ssa_def *elem_offset =
843 nir_iadd(&b->nb, offset, nir_imm_int(&b->nb, i * col_stride));
844 _vtn_load_store_tail(b, op, load, index, elem_offset,
845 access_offset, access_size,
846 &(*inout)->elems[i],
847 glsl_vector_type(base_type, vec_width));
848 }
849
850 if (load && type->row_major)
851 *inout = vtn_ssa_transpose(b, *inout);
852 } else {
853 unsigned elems = glsl_get_vector_elements(type->type);
854 unsigned type_size = glsl_get_bit_size(type->type) / 8;
855 if (elems == 1 || type->stride == type_size) {
856 /* This is a tightly-packed normal scalar or vector load */
857 vtn_assert(glsl_type_is_vector_or_scalar(type->type));
858 _vtn_load_store_tail(b, op, load, index, offset,
859 access_offset, access_size,
860 inout, type->type);
861 } else {
862 /* This is a strided load. We have to load N things separately.
863 * This is the single column of a row-major matrix case.
864 */
865 vtn_assert(type->stride > type_size);
866 vtn_assert(type->stride % type_size == 0);
867
868 nir_ssa_def *per_comp[4];
869 for (unsigned i = 0; i < elems; i++) {
870 nir_ssa_def *elem_offset =
871 nir_iadd(&b->nb, offset,
872 nir_imm_int(&b->nb, i * type->stride));
873 struct vtn_ssa_value *comp, temp_val;
874 if (!load) {
875 temp_val.def = nir_channel(&b->nb, (*inout)->def, i);
876 temp_val.type = glsl_scalar_type(base_type);
877 }
878 comp = &temp_val;
879 _vtn_load_store_tail(b, op, load, index, elem_offset,
880 access_offset, access_size,
881 &comp, glsl_scalar_type(base_type));
882 per_comp[i] = comp->def;
883 }
884
885 if (load) {
886 if (*inout == NULL)
887 *inout = vtn_create_ssa_value(b, type->type);
888 (*inout)->def = nir_vec(&b->nb, per_comp, elems);
889 }
890 }
891 }
892 return;
893
894 case GLSL_TYPE_ARRAY: {
895 unsigned elems = glsl_get_length(type->type);
896 for (unsigned i = 0; i < elems; i++) {
897 nir_ssa_def *elem_off =
898 nir_iadd(&b->nb, offset, nir_imm_int(&b->nb, i * type->stride));
899 _vtn_block_load_store(b, op, load, index, elem_off,
900 access_offset, access_size,
901 NULL, 0,
902 type->array_element, &(*inout)->elems[i]);
903 }
904 return;
905 }
906
907 case GLSL_TYPE_STRUCT: {
908 unsigned elems = glsl_get_length(type->type);
909 for (unsigned i = 0; i < elems; i++) {
910 nir_ssa_def *elem_off =
911 nir_iadd(&b->nb, offset, nir_imm_int(&b->nb, type->offsets[i]));
912 _vtn_block_load_store(b, op, load, index, elem_off,
913 access_offset, access_size,
914 NULL, 0,
915 type->members[i], &(*inout)->elems[i]);
916 }
917 return;
918 }
919
920 default:
921 vtn_fail("Invalid block member type");
922 }
923 }
924
925 static struct vtn_ssa_value *
vtn_block_load(struct vtn_builder * b,struct vtn_pointer * src)926 vtn_block_load(struct vtn_builder *b, struct vtn_pointer *src)
927 {
928 nir_intrinsic_op op;
929 unsigned access_offset = 0, access_size = 0;
930 switch (src->mode) {
931 case vtn_variable_mode_ubo:
932 op = nir_intrinsic_load_ubo;
933 break;
934 case vtn_variable_mode_ssbo:
935 op = nir_intrinsic_load_ssbo;
936 break;
937 case vtn_variable_mode_push_constant:
938 op = nir_intrinsic_load_push_constant;
939 vtn_access_chain_get_offset_size(b, src->chain, src->var->type,
940 &access_offset, &access_size);
941 break;
942 case vtn_variable_mode_workgroup:
943 op = nir_intrinsic_load_shared;
944 break;
945 default:
946 vtn_fail("Invalid block variable mode");
947 }
948
949 nir_ssa_def *offset, *index = NULL;
950 unsigned chain_idx;
951 offset = vtn_pointer_to_offset(b, src, &index, &chain_idx);
952
953 struct vtn_ssa_value *value = NULL;
954 _vtn_block_load_store(b, op, true, index, offset,
955 access_offset, access_size,
956 src->chain, chain_idx, src->type, &value);
957 return value;
958 }
959
960 static void
vtn_block_store(struct vtn_builder * b,struct vtn_ssa_value * src,struct vtn_pointer * dst)961 vtn_block_store(struct vtn_builder *b, struct vtn_ssa_value *src,
962 struct vtn_pointer *dst)
963 {
964 nir_intrinsic_op op;
965 switch (dst->mode) {
966 case vtn_variable_mode_ssbo:
967 op = nir_intrinsic_store_ssbo;
968 break;
969 case vtn_variable_mode_workgroup:
970 op = nir_intrinsic_store_shared;
971 break;
972 default:
973 vtn_fail("Invalid block variable mode");
974 }
975
976 nir_ssa_def *offset, *index = NULL;
977 unsigned chain_idx;
978 offset = vtn_pointer_to_offset(b, dst, &index, &chain_idx);
979
980 _vtn_block_load_store(b, op, false, index, offset,
981 0, 0, dst->chain, chain_idx, dst->type, &src);
982 }
983
984 static void
_vtn_variable_load_store(struct vtn_builder * b,bool load,struct vtn_pointer * ptr,struct vtn_ssa_value ** inout)985 _vtn_variable_load_store(struct vtn_builder *b, bool load,
986 struct vtn_pointer *ptr,
987 struct vtn_ssa_value **inout)
988 {
989 enum glsl_base_type base_type = glsl_get_base_type(ptr->type->type);
990 switch (base_type) {
991 case GLSL_TYPE_UINT:
992 case GLSL_TYPE_INT:
993 case GLSL_TYPE_UINT16:
994 case GLSL_TYPE_INT16:
995 case GLSL_TYPE_UINT64:
996 case GLSL_TYPE_INT64:
997 case GLSL_TYPE_FLOAT:
998 case GLSL_TYPE_FLOAT16:
999 case GLSL_TYPE_BOOL:
1000 case GLSL_TYPE_DOUBLE:
1001 /* At this point, we have a scalar, vector, or matrix so we know that
1002 * there cannot be any structure splitting still in the way. By
1003 * stopping at the matrix level rather than the vector level, we
1004 * ensure that matrices get loaded in the optimal way even if they
1005 * are storred row-major in a UBO.
1006 */
1007 if (load) {
1008 *inout = vtn_local_load(b, vtn_pointer_to_deref(b, ptr));
1009 } else {
1010 vtn_local_store(b, *inout, vtn_pointer_to_deref(b, ptr));
1011 }
1012 return;
1013
1014 case GLSL_TYPE_ARRAY:
1015 case GLSL_TYPE_STRUCT: {
1016 unsigned elems = glsl_get_length(ptr->type->type);
1017 if (load) {
1018 vtn_assert(*inout == NULL);
1019 *inout = rzalloc(b, struct vtn_ssa_value);
1020 (*inout)->type = ptr->type->type;
1021 (*inout)->elems = rzalloc_array(b, struct vtn_ssa_value *, elems);
1022 }
1023
1024 struct vtn_access_chain chain = {
1025 .length = 1,
1026 .link = {
1027 { .mode = vtn_access_mode_literal, },
1028 }
1029 };
1030 for (unsigned i = 0; i < elems; i++) {
1031 chain.link[0].id = i;
1032 struct vtn_pointer *elem = vtn_pointer_dereference(b, ptr, &chain);
1033 _vtn_variable_load_store(b, load, elem, &(*inout)->elems[i]);
1034 }
1035 return;
1036 }
1037
1038 default:
1039 vtn_fail("Invalid access chain type");
1040 }
1041 }
1042
1043 struct vtn_ssa_value *
vtn_variable_load(struct vtn_builder * b,struct vtn_pointer * src)1044 vtn_variable_load(struct vtn_builder *b, struct vtn_pointer *src)
1045 {
1046 if (vtn_pointer_is_external_block(b, src)) {
1047 return vtn_block_load(b, src);
1048 } else {
1049 struct vtn_ssa_value *val = NULL;
1050 _vtn_variable_load_store(b, true, src, &val);
1051 return val;
1052 }
1053 }
1054
1055 void
vtn_variable_store(struct vtn_builder * b,struct vtn_ssa_value * src,struct vtn_pointer * dest)1056 vtn_variable_store(struct vtn_builder *b, struct vtn_ssa_value *src,
1057 struct vtn_pointer *dest)
1058 {
1059 if (vtn_pointer_is_external_block(b, dest)) {
1060 vtn_assert(dest->mode == vtn_variable_mode_ssbo ||
1061 dest->mode == vtn_variable_mode_workgroup);
1062 vtn_block_store(b, src, dest);
1063 } else {
1064 _vtn_variable_load_store(b, false, dest, &src);
1065 }
1066 }
1067
1068 static void
_vtn_variable_copy(struct vtn_builder * b,struct vtn_pointer * dest,struct vtn_pointer * src)1069 _vtn_variable_copy(struct vtn_builder *b, struct vtn_pointer *dest,
1070 struct vtn_pointer *src)
1071 {
1072 vtn_assert(src->type->type == dest->type->type);
1073 enum glsl_base_type base_type = glsl_get_base_type(src->type->type);
1074 switch (base_type) {
1075 case GLSL_TYPE_UINT:
1076 case GLSL_TYPE_INT:
1077 case GLSL_TYPE_UINT16:
1078 case GLSL_TYPE_INT16:
1079 case GLSL_TYPE_UINT64:
1080 case GLSL_TYPE_INT64:
1081 case GLSL_TYPE_FLOAT:
1082 case GLSL_TYPE_FLOAT16:
1083 case GLSL_TYPE_DOUBLE:
1084 case GLSL_TYPE_BOOL:
1085 /* At this point, we have a scalar, vector, or matrix so we know that
1086 * there cannot be any structure splitting still in the way. By
1087 * stopping at the matrix level rather than the vector level, we
1088 * ensure that matrices get loaded in the optimal way even if they
1089 * are storred row-major in a UBO.
1090 */
1091 vtn_variable_store(b, vtn_variable_load(b, src), dest);
1092 return;
1093
1094 case GLSL_TYPE_ARRAY:
1095 case GLSL_TYPE_STRUCT: {
1096 struct vtn_access_chain chain = {
1097 .length = 1,
1098 .link = {
1099 { .mode = vtn_access_mode_literal, },
1100 }
1101 };
1102 unsigned elems = glsl_get_length(src->type->type);
1103 for (unsigned i = 0; i < elems; i++) {
1104 chain.link[0].id = i;
1105 struct vtn_pointer *src_elem =
1106 vtn_pointer_dereference(b, src, &chain);
1107 struct vtn_pointer *dest_elem =
1108 vtn_pointer_dereference(b, dest, &chain);
1109
1110 _vtn_variable_copy(b, dest_elem, src_elem);
1111 }
1112 return;
1113 }
1114
1115 default:
1116 vtn_fail("Invalid access chain type");
1117 }
1118 }
1119
1120 static void
vtn_variable_copy(struct vtn_builder * b,struct vtn_pointer * dest,struct vtn_pointer * src)1121 vtn_variable_copy(struct vtn_builder *b, struct vtn_pointer *dest,
1122 struct vtn_pointer *src)
1123 {
1124 /* TODO: At some point, we should add a special-case for when we can
1125 * just emit a copy_var intrinsic.
1126 */
1127 _vtn_variable_copy(b, dest, src);
1128 }
1129
1130 static void
set_mode_system_value(struct vtn_builder * b,nir_variable_mode * mode)1131 set_mode_system_value(struct vtn_builder *b, nir_variable_mode *mode)
1132 {
1133 vtn_assert(*mode == nir_var_system_value || *mode == nir_var_shader_in);
1134 *mode = nir_var_system_value;
1135 }
1136
1137 static void
vtn_get_builtin_location(struct vtn_builder * b,SpvBuiltIn builtin,int * location,nir_variable_mode * mode)1138 vtn_get_builtin_location(struct vtn_builder *b,
1139 SpvBuiltIn builtin, int *location,
1140 nir_variable_mode *mode)
1141 {
1142 switch (builtin) {
1143 case SpvBuiltInPosition:
1144 *location = VARYING_SLOT_POS;
1145 break;
1146 case SpvBuiltInPointSize:
1147 *location = VARYING_SLOT_PSIZ;
1148 break;
1149 case SpvBuiltInClipDistance:
1150 *location = VARYING_SLOT_CLIP_DIST0; /* XXX CLIP_DIST1? */
1151 break;
1152 case SpvBuiltInCullDistance:
1153 *location = VARYING_SLOT_CULL_DIST0;
1154 break;
1155 case SpvBuiltInVertexIndex:
1156 *location = SYSTEM_VALUE_VERTEX_ID;
1157 set_mode_system_value(b, mode);
1158 break;
1159 case SpvBuiltInVertexId:
1160 /* Vulkan defines VertexID to be zero-based and reserves the new
1161 * builtin keyword VertexIndex to indicate the non-zero-based value.
1162 */
1163 *location = SYSTEM_VALUE_VERTEX_ID_ZERO_BASE;
1164 set_mode_system_value(b, mode);
1165 break;
1166 case SpvBuiltInInstanceIndex:
1167 *location = SYSTEM_VALUE_INSTANCE_INDEX;
1168 set_mode_system_value(b, mode);
1169 break;
1170 case SpvBuiltInInstanceId:
1171 *location = SYSTEM_VALUE_INSTANCE_ID;
1172 set_mode_system_value(b, mode);
1173 break;
1174 case SpvBuiltInPrimitiveId:
1175 if (b->shader->info.stage == MESA_SHADER_FRAGMENT) {
1176 vtn_assert(*mode == nir_var_shader_in);
1177 *location = VARYING_SLOT_PRIMITIVE_ID;
1178 } else if (*mode == nir_var_shader_out) {
1179 *location = VARYING_SLOT_PRIMITIVE_ID;
1180 } else {
1181 *location = SYSTEM_VALUE_PRIMITIVE_ID;
1182 set_mode_system_value(b, mode);
1183 }
1184 break;
1185 case SpvBuiltInInvocationId:
1186 *location = SYSTEM_VALUE_INVOCATION_ID;
1187 set_mode_system_value(b, mode);
1188 break;
1189 case SpvBuiltInLayer:
1190 *location = VARYING_SLOT_LAYER;
1191 if (b->shader->info.stage == MESA_SHADER_FRAGMENT)
1192 *mode = nir_var_shader_in;
1193 else if (b->shader->info.stage == MESA_SHADER_GEOMETRY)
1194 *mode = nir_var_shader_out;
1195 else
1196 vtn_fail("invalid stage for SpvBuiltInLayer");
1197 break;
1198 case SpvBuiltInViewportIndex:
1199 *location = VARYING_SLOT_VIEWPORT;
1200 if (b->shader->info.stage == MESA_SHADER_GEOMETRY)
1201 *mode = nir_var_shader_out;
1202 else if (b->shader->info.stage == MESA_SHADER_FRAGMENT)
1203 *mode = nir_var_shader_in;
1204 else
1205 vtn_fail("invalid stage for SpvBuiltInViewportIndex");
1206 break;
1207 case SpvBuiltInTessLevelOuter:
1208 *location = VARYING_SLOT_TESS_LEVEL_OUTER;
1209 break;
1210 case SpvBuiltInTessLevelInner:
1211 *location = VARYING_SLOT_TESS_LEVEL_INNER;
1212 break;
1213 case SpvBuiltInTessCoord:
1214 *location = SYSTEM_VALUE_TESS_COORD;
1215 set_mode_system_value(b, mode);
1216 break;
1217 case SpvBuiltInPatchVertices:
1218 *location = SYSTEM_VALUE_VERTICES_IN;
1219 set_mode_system_value(b, mode);
1220 break;
1221 case SpvBuiltInFragCoord:
1222 *location = VARYING_SLOT_POS;
1223 vtn_assert(*mode == nir_var_shader_in);
1224 break;
1225 case SpvBuiltInPointCoord:
1226 *location = VARYING_SLOT_PNTC;
1227 vtn_assert(*mode == nir_var_shader_in);
1228 break;
1229 case SpvBuiltInFrontFacing:
1230 *location = SYSTEM_VALUE_FRONT_FACE;
1231 set_mode_system_value(b, mode);
1232 break;
1233 case SpvBuiltInSampleId:
1234 *location = SYSTEM_VALUE_SAMPLE_ID;
1235 set_mode_system_value(b, mode);
1236 break;
1237 case SpvBuiltInSamplePosition:
1238 *location = SYSTEM_VALUE_SAMPLE_POS;
1239 set_mode_system_value(b, mode);
1240 break;
1241 case SpvBuiltInSampleMask:
1242 if (*mode == nir_var_shader_out) {
1243 *location = FRAG_RESULT_SAMPLE_MASK;
1244 } else {
1245 *location = SYSTEM_VALUE_SAMPLE_MASK_IN;
1246 set_mode_system_value(b, mode);
1247 }
1248 break;
1249 case SpvBuiltInFragDepth:
1250 *location = FRAG_RESULT_DEPTH;
1251 vtn_assert(*mode == nir_var_shader_out);
1252 break;
1253 case SpvBuiltInHelperInvocation:
1254 *location = SYSTEM_VALUE_HELPER_INVOCATION;
1255 set_mode_system_value(b, mode);
1256 break;
1257 case SpvBuiltInNumWorkgroups:
1258 *location = SYSTEM_VALUE_NUM_WORK_GROUPS;
1259 set_mode_system_value(b, mode);
1260 break;
1261 case SpvBuiltInWorkgroupSize:
1262 /* This should already be handled */
1263 vtn_fail("unsupported builtin");
1264 break;
1265 case SpvBuiltInWorkgroupId:
1266 *location = SYSTEM_VALUE_WORK_GROUP_ID;
1267 set_mode_system_value(b, mode);
1268 break;
1269 case SpvBuiltInLocalInvocationId:
1270 *location = SYSTEM_VALUE_LOCAL_INVOCATION_ID;
1271 set_mode_system_value(b, mode);
1272 break;
1273 case SpvBuiltInLocalInvocationIndex:
1274 *location = SYSTEM_VALUE_LOCAL_INVOCATION_INDEX;
1275 set_mode_system_value(b, mode);
1276 break;
1277 case SpvBuiltInGlobalInvocationId:
1278 *location = SYSTEM_VALUE_GLOBAL_INVOCATION_ID;
1279 set_mode_system_value(b, mode);
1280 break;
1281 case SpvBuiltInBaseVertex:
1282 *location = SYSTEM_VALUE_BASE_VERTEX;
1283 set_mode_system_value(b, mode);
1284 break;
1285 case SpvBuiltInBaseInstance:
1286 *location = SYSTEM_VALUE_BASE_INSTANCE;
1287 set_mode_system_value(b, mode);
1288 break;
1289 case SpvBuiltInDrawIndex:
1290 *location = SYSTEM_VALUE_DRAW_ID;
1291 set_mode_system_value(b, mode);
1292 break;
1293 case SpvBuiltInViewIndex:
1294 *location = SYSTEM_VALUE_VIEW_INDEX;
1295 set_mode_system_value(b, mode);
1296 break;
1297 default:
1298 vtn_fail("unsupported builtin");
1299 }
1300 }
1301
1302 static void
apply_var_decoration(struct vtn_builder * b,nir_variable * nir_var,const struct vtn_decoration * dec)1303 apply_var_decoration(struct vtn_builder *b, nir_variable *nir_var,
1304 const struct vtn_decoration *dec)
1305 {
1306 switch (dec->decoration) {
1307 case SpvDecorationRelaxedPrecision:
1308 break; /* FIXME: Do nothing with this for now. */
1309 case SpvDecorationNoPerspective:
1310 nir_var->data.interpolation = INTERP_MODE_NOPERSPECTIVE;
1311 break;
1312 case SpvDecorationFlat:
1313 nir_var->data.interpolation = INTERP_MODE_FLAT;
1314 break;
1315 case SpvDecorationCentroid:
1316 nir_var->data.centroid = true;
1317 break;
1318 case SpvDecorationSample:
1319 nir_var->data.sample = true;
1320 break;
1321 case SpvDecorationInvariant:
1322 nir_var->data.invariant = true;
1323 break;
1324 case SpvDecorationConstant:
1325 vtn_assert(nir_var->constant_initializer != NULL);
1326 nir_var->data.read_only = true;
1327 break;
1328 case SpvDecorationNonReadable:
1329 nir_var->data.image.write_only = true;
1330 break;
1331 case SpvDecorationNonWritable:
1332 nir_var->data.read_only = true;
1333 nir_var->data.image.read_only = true;
1334 break;
1335 case SpvDecorationComponent:
1336 nir_var->data.location_frac = dec->literals[0];
1337 break;
1338 case SpvDecorationIndex:
1339 nir_var->data.index = dec->literals[0];
1340 break;
1341 case SpvDecorationBuiltIn: {
1342 SpvBuiltIn builtin = dec->literals[0];
1343
1344 if (builtin == SpvBuiltInWorkgroupSize) {
1345 /* This shouldn't be a builtin. It's actually a constant. */
1346 nir_var->data.mode = nir_var_global;
1347 nir_var->data.read_only = true;
1348
1349 nir_constant *c = rzalloc(nir_var, nir_constant);
1350 c->values[0].u32[0] = b->shader->info.cs.local_size[0];
1351 c->values[0].u32[1] = b->shader->info.cs.local_size[1];
1352 c->values[0].u32[2] = b->shader->info.cs.local_size[2];
1353 nir_var->constant_initializer = c;
1354 break;
1355 }
1356
1357 nir_variable_mode mode = nir_var->data.mode;
1358 vtn_get_builtin_location(b, builtin, &nir_var->data.location, &mode);
1359 nir_var->data.mode = mode;
1360
1361 switch (builtin) {
1362 case SpvBuiltInTessLevelOuter:
1363 case SpvBuiltInTessLevelInner:
1364 nir_var->data.compact = true;
1365 break;
1366 case SpvBuiltInSamplePosition:
1367 nir_var->data.origin_upper_left = b->origin_upper_left;
1368 /* fallthrough */
1369 case SpvBuiltInFragCoord:
1370 nir_var->data.pixel_center_integer = b->pixel_center_integer;
1371 break;
1372 default:
1373 break;
1374 }
1375 }
1376
1377 case SpvDecorationSpecId:
1378 case SpvDecorationRowMajor:
1379 case SpvDecorationColMajor:
1380 case SpvDecorationMatrixStride:
1381 case SpvDecorationRestrict:
1382 case SpvDecorationAliased:
1383 case SpvDecorationVolatile:
1384 case SpvDecorationCoherent:
1385 case SpvDecorationUniform:
1386 case SpvDecorationStream:
1387 case SpvDecorationOffset:
1388 case SpvDecorationLinkageAttributes:
1389 break; /* Do nothing with these here */
1390
1391 case SpvDecorationPatch:
1392 nir_var->data.patch = true;
1393 break;
1394
1395 case SpvDecorationLocation:
1396 vtn_fail("Handled above");
1397
1398 case SpvDecorationBlock:
1399 case SpvDecorationBufferBlock:
1400 case SpvDecorationArrayStride:
1401 case SpvDecorationGLSLShared:
1402 case SpvDecorationGLSLPacked:
1403 break; /* These can apply to a type but we don't care about them */
1404
1405 case SpvDecorationBinding:
1406 case SpvDecorationDescriptorSet:
1407 case SpvDecorationNoContraction:
1408 case SpvDecorationInputAttachmentIndex:
1409 vtn_warn("Decoration not allowed for variable or structure member: %s",
1410 spirv_decoration_to_string(dec->decoration));
1411 break;
1412
1413 case SpvDecorationXfbBuffer:
1414 case SpvDecorationXfbStride:
1415 vtn_warn("Vulkan does not have transform feedback: %s",
1416 spirv_decoration_to_string(dec->decoration));
1417 break;
1418
1419 case SpvDecorationCPacked:
1420 case SpvDecorationSaturatedConversion:
1421 case SpvDecorationFuncParamAttr:
1422 case SpvDecorationFPRoundingMode:
1423 case SpvDecorationFPFastMathMode:
1424 case SpvDecorationAlignment:
1425 vtn_warn("Decoration only allowed for CL-style kernels: %s",
1426 spirv_decoration_to_string(dec->decoration));
1427 break;
1428
1429 default:
1430 vtn_fail("Unhandled decoration");
1431 }
1432 }
1433
1434 static void
var_is_patch_cb(struct vtn_builder * b,struct vtn_value * val,int member,const struct vtn_decoration * dec,void * out_is_patch)1435 var_is_patch_cb(struct vtn_builder *b, struct vtn_value *val, int member,
1436 const struct vtn_decoration *dec, void *out_is_patch)
1437 {
1438 if (dec->decoration == SpvDecorationPatch) {
1439 *((bool *) out_is_patch) = true;
1440 }
1441 }
1442
1443 static void
var_decoration_cb(struct vtn_builder * b,struct vtn_value * val,int member,const struct vtn_decoration * dec,void * void_var)1444 var_decoration_cb(struct vtn_builder *b, struct vtn_value *val, int member,
1445 const struct vtn_decoration *dec, void *void_var)
1446 {
1447 struct vtn_variable *vtn_var = void_var;
1448
1449 /* Handle decorations that apply to a vtn_variable as a whole */
1450 switch (dec->decoration) {
1451 case SpvDecorationBinding:
1452 vtn_var->binding = dec->literals[0];
1453 return;
1454 case SpvDecorationDescriptorSet:
1455 vtn_var->descriptor_set = dec->literals[0];
1456 return;
1457 case SpvDecorationInputAttachmentIndex:
1458 vtn_var->input_attachment_index = dec->literals[0];
1459 return;
1460 case SpvDecorationPatch:
1461 vtn_var->patch = true;
1462 break;
1463 default:
1464 break;
1465 }
1466
1467 if (val->value_type == vtn_value_type_pointer) {
1468 assert(val->pointer->var == void_var);
1469 assert(val->pointer->chain == NULL);
1470 assert(member == -1);
1471 } else {
1472 assert(val->value_type == vtn_value_type_type);
1473 }
1474
1475 /* Location is odd. If applied to a split structure, we have to walk the
1476 * whole thing and accumulate the location. It's easier to handle as a
1477 * special case.
1478 */
1479 if (dec->decoration == SpvDecorationLocation) {
1480 unsigned location = dec->literals[0];
1481 bool is_vertex_input;
1482 if (b->shader->info.stage == MESA_SHADER_FRAGMENT &&
1483 vtn_var->mode == vtn_variable_mode_output) {
1484 is_vertex_input = false;
1485 location += FRAG_RESULT_DATA0;
1486 } else if (b->shader->info.stage == MESA_SHADER_VERTEX &&
1487 vtn_var->mode == vtn_variable_mode_input) {
1488 is_vertex_input = true;
1489 location += VERT_ATTRIB_GENERIC0;
1490 } else if (vtn_var->mode == vtn_variable_mode_input ||
1491 vtn_var->mode == vtn_variable_mode_output) {
1492 is_vertex_input = false;
1493 location += vtn_var->patch ? VARYING_SLOT_PATCH0 : VARYING_SLOT_VAR0;
1494 } else {
1495 vtn_warn("Location must be on input or output variable");
1496 return;
1497 }
1498
1499 if (vtn_var->var) {
1500 /* This handles the member and lone variable cases */
1501 vtn_var->var->data.location = location;
1502 } else {
1503 /* This handles the structure member case */
1504 assert(vtn_var->members);
1505 unsigned length =
1506 glsl_get_length(glsl_without_array(vtn_var->type->type));
1507 for (unsigned i = 0; i < length; i++) {
1508 vtn_var->members[i]->data.location = location;
1509 location +=
1510 glsl_count_attribute_slots(vtn_var->members[i]->interface_type,
1511 is_vertex_input);
1512 }
1513 }
1514 return;
1515 } else {
1516 if (vtn_var->var) {
1517 assert(member == -1);
1518 apply_var_decoration(b, vtn_var->var, dec);
1519 } else if (vtn_var->members) {
1520 if (member >= 0) {
1521 /* Member decorations must come from a type */
1522 assert(val->value_type == vtn_value_type_type);
1523 apply_var_decoration(b, vtn_var->members[member], dec);
1524 } else {
1525 unsigned length =
1526 glsl_get_length(glsl_without_array(vtn_var->type->type));
1527 for (unsigned i = 0; i < length; i++)
1528 apply_var_decoration(b, vtn_var->members[i], dec);
1529 }
1530 } else {
1531 /* A few variables, those with external storage, have no actual
1532 * nir_variables associated with them. Fortunately, all decorations
1533 * we care about for those variables are on the type only.
1534 */
1535 vtn_assert(vtn_var->mode == vtn_variable_mode_ubo ||
1536 vtn_var->mode == vtn_variable_mode_ssbo ||
1537 vtn_var->mode == vtn_variable_mode_push_constant ||
1538 (vtn_var->mode == vtn_variable_mode_workgroup &&
1539 b->options->lower_workgroup_access_to_offsets));
1540 }
1541 }
1542 }
1543
1544 static enum vtn_variable_mode
vtn_storage_class_to_mode(struct vtn_builder * b,SpvStorageClass class,struct vtn_type * interface_type,nir_variable_mode * nir_mode_out)1545 vtn_storage_class_to_mode(struct vtn_builder *b,
1546 SpvStorageClass class,
1547 struct vtn_type *interface_type,
1548 nir_variable_mode *nir_mode_out)
1549 {
1550 enum vtn_variable_mode mode;
1551 nir_variable_mode nir_mode;
1552 switch (class) {
1553 case SpvStorageClassUniform:
1554 if (interface_type->block) {
1555 mode = vtn_variable_mode_ubo;
1556 nir_mode = 0;
1557 } else if (interface_type->buffer_block) {
1558 mode = vtn_variable_mode_ssbo;
1559 nir_mode = 0;
1560 } else {
1561 vtn_fail("Invalid uniform variable type");
1562 }
1563 break;
1564 case SpvStorageClassStorageBuffer:
1565 mode = vtn_variable_mode_ssbo;
1566 nir_mode = 0;
1567 break;
1568 case SpvStorageClassUniformConstant:
1569 if (glsl_type_is_image(interface_type->type)) {
1570 mode = vtn_variable_mode_image;
1571 nir_mode = nir_var_uniform;
1572 } else if (glsl_type_is_sampler(interface_type->type)) {
1573 mode = vtn_variable_mode_sampler;
1574 nir_mode = nir_var_uniform;
1575 } else {
1576 vtn_fail("Invalid uniform constant variable type");
1577 }
1578 break;
1579 case SpvStorageClassPushConstant:
1580 mode = vtn_variable_mode_push_constant;
1581 nir_mode = nir_var_uniform;
1582 break;
1583 case SpvStorageClassInput:
1584 mode = vtn_variable_mode_input;
1585 nir_mode = nir_var_shader_in;
1586 break;
1587 case SpvStorageClassOutput:
1588 mode = vtn_variable_mode_output;
1589 nir_mode = nir_var_shader_out;
1590 break;
1591 case SpvStorageClassPrivate:
1592 mode = vtn_variable_mode_global;
1593 nir_mode = nir_var_global;
1594 break;
1595 case SpvStorageClassFunction:
1596 mode = vtn_variable_mode_local;
1597 nir_mode = nir_var_local;
1598 break;
1599 case SpvStorageClassWorkgroup:
1600 mode = vtn_variable_mode_workgroup;
1601 nir_mode = nir_var_shared;
1602 break;
1603 case SpvStorageClassCrossWorkgroup:
1604 case SpvStorageClassGeneric:
1605 case SpvStorageClassAtomicCounter:
1606 default:
1607 vtn_fail("Unhandled variable storage class");
1608 }
1609
1610 if (nir_mode_out)
1611 *nir_mode_out = nir_mode;
1612
1613 return mode;
1614 }
1615
1616 nir_ssa_def *
vtn_pointer_to_ssa(struct vtn_builder * b,struct vtn_pointer * ptr)1617 vtn_pointer_to_ssa(struct vtn_builder *b, struct vtn_pointer *ptr)
1618 {
1619 /* This pointer needs to have a pointer type with actual storage */
1620 vtn_assert(ptr->ptr_type);
1621 vtn_assert(ptr->ptr_type->type);
1622
1623 if (!ptr->offset) {
1624 /* If we don't have an offset then we must be a pointer to the variable
1625 * itself.
1626 */
1627 vtn_assert(!ptr->offset && !ptr->block_index);
1628
1629 struct vtn_access_chain chain = {
1630 .length = 0,
1631 };
1632 ptr = vtn_ssa_offset_pointer_dereference(b, ptr, &chain);
1633 }
1634
1635 vtn_assert(ptr->offset);
1636 if (ptr->block_index) {
1637 vtn_assert(ptr->mode == vtn_variable_mode_ubo ||
1638 ptr->mode == vtn_variable_mode_ssbo);
1639 return nir_vec2(&b->nb, ptr->block_index, ptr->offset);
1640 } else {
1641 vtn_assert(ptr->mode == vtn_variable_mode_workgroup);
1642 return ptr->offset;
1643 }
1644 }
1645
1646 struct vtn_pointer *
vtn_pointer_from_ssa(struct vtn_builder * b,nir_ssa_def * ssa,struct vtn_type * ptr_type)1647 vtn_pointer_from_ssa(struct vtn_builder *b, nir_ssa_def *ssa,
1648 struct vtn_type *ptr_type)
1649 {
1650 vtn_assert(ssa->num_components <= 2 && ssa->bit_size == 32);
1651 vtn_assert(ptr_type->base_type == vtn_base_type_pointer);
1652 vtn_assert(ptr_type->deref->base_type != vtn_base_type_pointer);
1653 /* This pointer type needs to have actual storage */
1654 vtn_assert(ptr_type->type);
1655
1656 struct vtn_pointer *ptr = rzalloc(b, struct vtn_pointer);
1657 ptr->mode = vtn_storage_class_to_mode(b, ptr_type->storage_class,
1658 ptr_type, NULL);
1659 ptr->type = ptr_type->deref;
1660 ptr->ptr_type = ptr_type;
1661
1662 if (ssa->num_components > 1) {
1663 vtn_assert(ssa->num_components == 2);
1664 vtn_assert(ptr->mode == vtn_variable_mode_ubo ||
1665 ptr->mode == vtn_variable_mode_ssbo);
1666 ptr->block_index = nir_channel(&b->nb, ssa, 0);
1667 ptr->offset = nir_channel(&b->nb, ssa, 1);
1668 } else {
1669 vtn_assert(ssa->num_components == 1);
1670 vtn_assert(ptr->mode == vtn_variable_mode_workgroup);
1671 ptr->block_index = NULL;
1672 ptr->offset = ssa;
1673 }
1674
1675 return ptr;
1676 }
1677
1678 static bool
is_per_vertex_inout(const struct vtn_variable * var,gl_shader_stage stage)1679 is_per_vertex_inout(const struct vtn_variable *var, gl_shader_stage stage)
1680 {
1681 if (var->patch || !glsl_type_is_array(var->type->type))
1682 return false;
1683
1684 if (var->mode == vtn_variable_mode_input) {
1685 return stage == MESA_SHADER_TESS_CTRL ||
1686 stage == MESA_SHADER_TESS_EVAL ||
1687 stage == MESA_SHADER_GEOMETRY;
1688 }
1689
1690 if (var->mode == vtn_variable_mode_output)
1691 return stage == MESA_SHADER_TESS_CTRL;
1692
1693 return false;
1694 }
1695
1696 static void
vtn_create_variable(struct vtn_builder * b,struct vtn_value * val,struct vtn_type * ptr_type,SpvStorageClass storage_class,nir_constant * initializer)1697 vtn_create_variable(struct vtn_builder *b, struct vtn_value *val,
1698 struct vtn_type *ptr_type, SpvStorageClass storage_class,
1699 nir_constant *initializer)
1700 {
1701 vtn_assert(ptr_type->base_type == vtn_base_type_pointer);
1702 struct vtn_type *type = ptr_type->deref;
1703
1704 struct vtn_type *without_array = type;
1705 while(glsl_type_is_array(without_array->type))
1706 without_array = without_array->array_element;
1707
1708 enum vtn_variable_mode mode;
1709 nir_variable_mode nir_mode;
1710 mode = vtn_storage_class_to_mode(b, storage_class, without_array, &nir_mode);
1711
1712 switch (mode) {
1713 case vtn_variable_mode_ubo:
1714 b->shader->info.num_ubos++;
1715 break;
1716 case vtn_variable_mode_ssbo:
1717 b->shader->info.num_ssbos++;
1718 break;
1719 case vtn_variable_mode_image:
1720 b->shader->info.num_images++;
1721 break;
1722 case vtn_variable_mode_sampler:
1723 b->shader->info.num_textures++;
1724 break;
1725 case vtn_variable_mode_push_constant:
1726 b->shader->num_uniforms = vtn_type_block_size(b, type);
1727 break;
1728 default:
1729 /* No tallying is needed */
1730 break;
1731 }
1732
1733 struct vtn_variable *var = rzalloc(b, struct vtn_variable);
1734 var->type = type;
1735 var->mode = mode;
1736
1737 vtn_assert(val->value_type == vtn_value_type_pointer);
1738 val->pointer = vtn_pointer_for_variable(b, var, ptr_type);
1739
1740 switch (var->mode) {
1741 case vtn_variable_mode_local:
1742 case vtn_variable_mode_global:
1743 case vtn_variable_mode_image:
1744 case vtn_variable_mode_sampler:
1745 /* For these, we create the variable normally */
1746 var->var = rzalloc(b->shader, nir_variable);
1747 var->var->name = ralloc_strdup(var->var, val->name);
1748 var->var->type = var->type->type;
1749 var->var->data.mode = nir_mode;
1750
1751 switch (var->mode) {
1752 case vtn_variable_mode_image:
1753 case vtn_variable_mode_sampler:
1754 var->var->interface_type = without_array->type;
1755 break;
1756 default:
1757 var->var->interface_type = NULL;
1758 break;
1759 }
1760 break;
1761
1762 case vtn_variable_mode_workgroup:
1763 if (b->options->lower_workgroup_access_to_offsets) {
1764 var->shared_location = -1;
1765 } else {
1766 /* Create the variable normally */
1767 var->var = rzalloc(b->shader, nir_variable);
1768 var->var->name = ralloc_strdup(var->var, val->name);
1769 var->var->type = var->type->type;
1770 var->var->data.mode = nir_var_shared;
1771 }
1772 break;
1773
1774 case vtn_variable_mode_input:
1775 case vtn_variable_mode_output: {
1776 /* In order to know whether or not we're a per-vertex inout, we need
1777 * the patch qualifier. This means walking the variable decorations
1778 * early before we actually create any variables. Not a big deal.
1779 *
1780 * GLSLang really likes to place decorations in the most interior
1781 * thing it possibly can. In particular, if you have a struct, it
1782 * will place the patch decorations on the struct members. This
1783 * should be handled by the variable splitting below just fine.
1784 *
1785 * If you have an array-of-struct, things get even more weird as it
1786 * will place the patch decorations on the struct even though it's
1787 * inside an array and some of the members being patch and others not
1788 * makes no sense whatsoever. Since the only sensible thing is for
1789 * it to be all or nothing, we'll call it patch if any of the members
1790 * are declared patch.
1791 */
1792 var->patch = false;
1793 vtn_foreach_decoration(b, val, var_is_patch_cb, &var->patch);
1794 if (glsl_type_is_array(var->type->type) &&
1795 glsl_type_is_struct(without_array->type)) {
1796 vtn_foreach_decoration(b, vtn_value(b, without_array->id,
1797 vtn_value_type_type),
1798 var_is_patch_cb, &var->patch);
1799 }
1800
1801 /* For inputs and outputs, we immediately split structures. This
1802 * is for a couple of reasons. For one, builtins may all come in
1803 * a struct and we really want those split out into separate
1804 * variables. For another, interpolation qualifiers can be
1805 * applied to members of the top-level struct ane we need to be
1806 * able to preserve that information.
1807 */
1808
1809 int array_length = -1;
1810 struct vtn_type *interface_type = var->type;
1811 if (is_per_vertex_inout(var, b->shader->info.stage)) {
1812 /* In Geometry shaders (and some tessellation), inputs come
1813 * in per-vertex arrays. However, some builtins come in
1814 * non-per-vertex, hence the need for the is_array check. In
1815 * any case, there are no non-builtin arrays allowed so this
1816 * check should be sufficient.
1817 */
1818 interface_type = var->type->array_element;
1819 array_length = glsl_get_length(var->type->type);
1820 }
1821
1822 if (glsl_type_is_struct(interface_type->type)) {
1823 /* It's a struct. Split it. */
1824 unsigned num_members = glsl_get_length(interface_type->type);
1825 var->members = ralloc_array(b, nir_variable *, num_members);
1826
1827 for (unsigned i = 0; i < num_members; i++) {
1828 const struct glsl_type *mtype = interface_type->members[i]->type;
1829 if (array_length >= 0)
1830 mtype = glsl_array_type(mtype, array_length);
1831
1832 var->members[i] = rzalloc(b->shader, nir_variable);
1833 var->members[i]->name =
1834 ralloc_asprintf(var->members[i], "%s.%d", val->name, i);
1835 var->members[i]->type = mtype;
1836 var->members[i]->interface_type =
1837 interface_type->members[i]->type;
1838 var->members[i]->data.mode = nir_mode;
1839 var->members[i]->data.patch = var->patch;
1840 }
1841 } else {
1842 var->var = rzalloc(b->shader, nir_variable);
1843 var->var->name = ralloc_strdup(var->var, val->name);
1844 var->var->type = var->type->type;
1845 var->var->interface_type = interface_type->type;
1846 var->var->data.mode = nir_mode;
1847 var->var->data.patch = var->patch;
1848 }
1849
1850 /* For inputs and outputs, we need to grab locations and builtin
1851 * information from the interface type.
1852 */
1853 vtn_foreach_decoration(b, vtn_value(b, interface_type->id,
1854 vtn_value_type_type),
1855 var_decoration_cb, var);
1856 break;
1857 }
1858
1859 case vtn_variable_mode_param:
1860 vtn_fail("Not created through OpVariable");
1861
1862 case vtn_variable_mode_ubo:
1863 case vtn_variable_mode_ssbo:
1864 case vtn_variable_mode_push_constant:
1865 /* These don't need actual variables. */
1866 break;
1867 }
1868
1869 if (initializer) {
1870 var->var->constant_initializer =
1871 nir_constant_clone(initializer, var->var);
1872 }
1873
1874 vtn_foreach_decoration(b, val, var_decoration_cb, var);
1875
1876 if (var->mode == vtn_variable_mode_image ||
1877 var->mode == vtn_variable_mode_sampler) {
1878 /* XXX: We still need the binding information in the nir_variable
1879 * for these. We should fix that.
1880 */
1881 var->var->data.binding = var->binding;
1882 var->var->data.descriptor_set = var->descriptor_set;
1883 var->var->data.index = var->input_attachment_index;
1884
1885 if (var->mode == vtn_variable_mode_image)
1886 var->var->data.image.format = without_array->image_format;
1887 }
1888
1889 if (var->mode == vtn_variable_mode_local) {
1890 vtn_assert(var->members == NULL && var->var != NULL);
1891 nir_function_impl_add_variable(b->nb.impl, var->var);
1892 } else if (var->var) {
1893 nir_shader_add_variable(b->shader, var->var);
1894 } else if (var->members) {
1895 unsigned count = glsl_get_length(without_array->type);
1896 for (unsigned i = 0; i < count; i++) {
1897 vtn_assert(var->members[i]->data.mode != nir_var_local);
1898 nir_shader_add_variable(b->shader, var->members[i]);
1899 }
1900 } else {
1901 vtn_assert(vtn_pointer_is_external_block(b, val->pointer));
1902 }
1903 }
1904
1905 static void
vtn_assert_types_equal(struct vtn_builder * b,SpvOp opcode,struct vtn_type * dst_type,struct vtn_type * src_type)1906 vtn_assert_types_equal(struct vtn_builder *b, SpvOp opcode,
1907 struct vtn_type *dst_type,
1908 struct vtn_type *src_type)
1909 {
1910 if (dst_type->id == src_type->id)
1911 return;
1912
1913 if (vtn_types_compatible(b, dst_type, src_type)) {
1914 /* Early versions of GLSLang would re-emit types unnecessarily and you
1915 * would end up with OpLoad, OpStore, or OpCopyMemory opcodes which have
1916 * mismatched source and destination types.
1917 *
1918 * https://github.com/KhronosGroup/glslang/issues/304
1919 * https://github.com/KhronosGroup/glslang/issues/307
1920 * https://bugs.freedesktop.org/show_bug.cgi?id=104338
1921 * https://bugs.freedesktop.org/show_bug.cgi?id=104424
1922 */
1923 vtn_warn("Source and destination types of %s do not have the same "
1924 "ID (but are compatible): %u vs %u",
1925 spirv_op_to_string(opcode), dst_type->id, src_type->id);
1926 return;
1927 }
1928
1929 vtn_fail("Source and destination types of %s do not match: %s vs. %s",
1930 spirv_op_to_string(opcode),
1931 glsl_get_type_name(dst_type->type),
1932 glsl_get_type_name(src_type->type));
1933 }
1934
1935 void
vtn_handle_variables(struct vtn_builder * b,SpvOp opcode,const uint32_t * w,unsigned count)1936 vtn_handle_variables(struct vtn_builder *b, SpvOp opcode,
1937 const uint32_t *w, unsigned count)
1938 {
1939 switch (opcode) {
1940 case SpvOpUndef: {
1941 struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_undef);
1942 val->type = vtn_value(b, w[1], vtn_value_type_type)->type;
1943 break;
1944 }
1945
1946 case SpvOpVariable: {
1947 struct vtn_type *ptr_type = vtn_value(b, w[1], vtn_value_type_type)->type;
1948
1949 struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_pointer);
1950
1951 SpvStorageClass storage_class = w[3];
1952 nir_constant *initializer = NULL;
1953 if (count > 4)
1954 initializer = vtn_value(b, w[4], vtn_value_type_constant)->constant;
1955
1956 vtn_create_variable(b, val, ptr_type, storage_class, initializer);
1957 break;
1958 }
1959
1960 case SpvOpAccessChain:
1961 case SpvOpPtrAccessChain:
1962 case SpvOpInBoundsAccessChain: {
1963 struct vtn_access_chain *chain = vtn_access_chain_create(b, count - 4);
1964 chain->ptr_as_array = (opcode == SpvOpPtrAccessChain);
1965
1966 unsigned idx = 0;
1967 for (int i = 4; i < count; i++) {
1968 struct vtn_value *link_val = vtn_untyped_value(b, w[i]);
1969 if (link_val->value_type == vtn_value_type_constant) {
1970 chain->link[idx].mode = vtn_access_mode_literal;
1971 chain->link[idx].id = link_val->constant->values[0].u32[0];
1972 } else {
1973 chain->link[idx].mode = vtn_access_mode_id;
1974 chain->link[idx].id = w[i];
1975
1976 }
1977 idx++;
1978 }
1979
1980 struct vtn_type *ptr_type = vtn_value(b, w[1], vtn_value_type_type)->type;
1981 struct vtn_value *base_val = vtn_untyped_value(b, w[3]);
1982 if (base_val->value_type == vtn_value_type_sampled_image) {
1983 /* This is rather insane. SPIR-V allows you to use OpSampledImage
1984 * to combine an array of images with a single sampler to get an
1985 * array of sampled images that all share the same sampler.
1986 * Fortunately, this means that we can more-or-less ignore the
1987 * sampler when crawling the access chain, but it does leave us
1988 * with this rather awkward little special-case.
1989 */
1990 struct vtn_value *val =
1991 vtn_push_value(b, w[2], vtn_value_type_sampled_image);
1992 val->sampled_image = ralloc(b, struct vtn_sampled_image);
1993 val->sampled_image->type = base_val->sampled_image->type;
1994 val->sampled_image->image =
1995 vtn_pointer_dereference(b, base_val->sampled_image->image, chain);
1996 val->sampled_image->sampler = base_val->sampled_image->sampler;
1997 } else {
1998 vtn_assert(base_val->value_type == vtn_value_type_pointer);
1999 struct vtn_value *val =
2000 vtn_push_value(b, w[2], vtn_value_type_pointer);
2001 val->pointer = vtn_pointer_dereference(b, base_val->pointer, chain);
2002 val->pointer->ptr_type = ptr_type;
2003 }
2004 break;
2005 }
2006
2007 case SpvOpCopyMemory: {
2008 struct vtn_value *dest = vtn_value(b, w[1], vtn_value_type_pointer);
2009 struct vtn_value *src = vtn_value(b, w[2], vtn_value_type_pointer);
2010
2011 vtn_assert_types_equal(b, opcode, dest->type->deref, src->type->deref);
2012
2013 vtn_variable_copy(b, dest->pointer, src->pointer);
2014 break;
2015 }
2016
2017 case SpvOpLoad: {
2018 struct vtn_type *res_type =
2019 vtn_value(b, w[1], vtn_value_type_type)->type;
2020 struct vtn_value *src_val = vtn_value(b, w[3], vtn_value_type_pointer);
2021 struct vtn_pointer *src = src_val->pointer;
2022
2023 vtn_assert_types_equal(b, opcode, res_type, src_val->type->deref);
2024
2025 if (src->mode == vtn_variable_mode_image ||
2026 src->mode == vtn_variable_mode_sampler) {
2027 vtn_push_value(b, w[2], vtn_value_type_pointer)->pointer = src;
2028 return;
2029 }
2030
2031 vtn_push_ssa(b, w[2], res_type, vtn_variable_load(b, src));
2032 break;
2033 }
2034
2035 case SpvOpStore: {
2036 struct vtn_value *dest_val = vtn_value(b, w[1], vtn_value_type_pointer);
2037 struct vtn_pointer *dest = dest_val->pointer;
2038 struct vtn_value *src_val = vtn_untyped_value(b, w[2]);
2039
2040 /* OpStore requires us to actually have a storage type */
2041 vtn_fail_if(dest->type->type == NULL,
2042 "Invalid destination type for OpStore");
2043
2044 if (glsl_get_base_type(dest->type->type) == GLSL_TYPE_BOOL &&
2045 glsl_get_base_type(src_val->type->type) == GLSL_TYPE_UINT) {
2046 /* Early versions of GLSLang would use uint types for UBOs/SSBOs but
2047 * would then store them to a local variable as bool. Work around
2048 * the issue by doing an implicit conversion.
2049 *
2050 * https://github.com/KhronosGroup/glslang/issues/170
2051 * https://bugs.freedesktop.org/show_bug.cgi?id=104424
2052 */
2053 vtn_warn("OpStore of value of type OpTypeInt to a pointer to type "
2054 "OpTypeBool. Doing an implicit conversion to work around "
2055 "the problem.");
2056 struct vtn_ssa_value *bool_ssa =
2057 vtn_create_ssa_value(b, dest->type->type);
2058 bool_ssa->def = nir_i2b(&b->nb, vtn_ssa_value(b, w[2])->def);
2059 vtn_variable_store(b, bool_ssa, dest);
2060 break;
2061 }
2062
2063 vtn_assert_types_equal(b, opcode, dest_val->type->deref, src_val->type);
2064
2065 if (glsl_type_is_sampler(dest->type->type)) {
2066 vtn_warn("OpStore of a sampler detected. Doing on-the-fly copy "
2067 "propagation to workaround the problem.");
2068 vtn_assert(dest->var->copy_prop_sampler == NULL);
2069 dest->var->copy_prop_sampler =
2070 vtn_value(b, w[2], vtn_value_type_pointer)->pointer;
2071 break;
2072 }
2073
2074 struct vtn_ssa_value *src = vtn_ssa_value(b, w[2]);
2075 vtn_variable_store(b, src, dest);
2076 break;
2077 }
2078
2079 case SpvOpArrayLength: {
2080 struct vtn_pointer *ptr =
2081 vtn_value(b, w[3], vtn_value_type_pointer)->pointer;
2082
2083 const uint32_t offset = ptr->var->type->offsets[w[4]];
2084 const uint32_t stride = ptr->var->type->members[w[4]]->stride;
2085
2086 if (!ptr->block_index) {
2087 struct vtn_access_chain chain = {
2088 .length = 0,
2089 };
2090 ptr = vtn_ssa_offset_pointer_dereference(b, ptr, &chain);
2091 vtn_assert(ptr->block_index);
2092 }
2093
2094 nir_intrinsic_instr *instr =
2095 nir_intrinsic_instr_create(b->nb.shader,
2096 nir_intrinsic_get_buffer_size);
2097 instr->src[0] = nir_src_for_ssa(ptr->block_index);
2098 nir_ssa_dest_init(&instr->instr, &instr->dest, 1, 32, NULL);
2099 nir_builder_instr_insert(&b->nb, &instr->instr);
2100 nir_ssa_def *buf_size = &instr->dest.ssa;
2101
2102 /* array_length = max(buffer_size - offset, 0) / stride */
2103 nir_ssa_def *array_length =
2104 nir_idiv(&b->nb,
2105 nir_imax(&b->nb,
2106 nir_isub(&b->nb,
2107 buf_size,
2108 nir_imm_int(&b->nb, offset)),
2109 nir_imm_int(&b->nb, 0u)),
2110 nir_imm_int(&b->nb, stride));
2111
2112 struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_ssa);
2113 val->ssa = vtn_create_ssa_value(b, glsl_uint_type());
2114 val->ssa->def = array_length;
2115 break;
2116 }
2117
2118 case SpvOpCopyMemorySized:
2119 default:
2120 vtn_fail("Unhandled opcode");
2121 }
2122 }
2123