Added few more stubs so that control reaches to DestroyDevice().
[mesa.git] / src / compiler / glsl / lower_ubo_reference.cpp
1 /*
2 * Copyright © 2012 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
21 * DEALINGS IN THE SOFTWARE.
22 */
23
24 /**
25 * \file lower_ubo_reference.cpp
26 *
27 * IR lower pass to replace dereferences of variables in a uniform
28 * buffer object with usage of ir_binop_ubo_load expressions, each of
29 * which can read data up to the size of a vec4.
30 *
31 * This relieves drivers of the responsibility to deal with tricky UBO
32 * layout issues like std140 structures and row_major matrices on
33 * their own.
34 */
35
36 #include "lower_buffer_access.h"
37 #include "ir_builder.h"
38 #include "main/macros.h"
39 #include "glsl_parser_extras.h"
40 #include "main/mtypes.h"
41
42 using namespace ir_builder;
43
44 namespace {
45 class lower_ubo_reference_visitor :
46 public lower_buffer_access::lower_buffer_access {
47 public:
48 lower_ubo_reference_visitor(struct gl_linked_shader *shader,
49 bool clamp_block_indices,
50 bool use_std430_as_default)
51 : shader(shader), clamp_block_indices(clamp_block_indices),
52 struct_field(NULL), variable(NULL)
53 {
54 this->use_std430_as_default = use_std430_as_default;
55 }
56
57 void handle_rvalue(ir_rvalue **rvalue);
58 ir_visitor_status visit_enter(ir_assignment *ir);
59
60 void setup_for_load_or_store(void *mem_ctx,
61 ir_variable *var,
62 ir_rvalue *deref,
63 ir_rvalue **offset,
64 unsigned *const_offset,
65 bool *row_major,
66 const glsl_type **matrix_type,
67 enum glsl_interface_packing packing);
68 uint32_t ssbo_access_params();
69 ir_expression *ubo_load(void *mem_ctx, const struct glsl_type *type,
70 ir_rvalue *offset);
71 ir_call *ssbo_load(void *mem_ctx, const struct glsl_type *type,
72 ir_rvalue *offset);
73
74 bool check_for_buffer_array_copy(ir_assignment *ir);
75 bool check_for_buffer_struct_copy(ir_assignment *ir);
76 void check_for_ssbo_store(ir_assignment *ir);
77 void write_to_memory(void *mem_ctx, ir_dereference *deref, ir_variable *var,
78 ir_variable *write_var, unsigned write_mask);
79 ir_call *ssbo_store(void *mem_ctx, ir_rvalue *deref, ir_rvalue *offset,
80 unsigned write_mask);
81
82 enum {
83 ubo_load_access,
84 ssbo_load_access,
85 ssbo_store_access,
86 ssbo_unsized_array_length_access,
87 ssbo_atomic_access,
88 } buffer_access_type;
89
90 void insert_buffer_access(void *mem_ctx, ir_dereference *deref,
91 const glsl_type *type, ir_rvalue *offset,
92 unsigned mask, int channel);
93
94 ir_visitor_status visit_enter(class ir_expression *);
95 ir_expression *calculate_ssbo_unsized_array_length(ir_expression *expr);
96 void check_ssbo_unsized_array_length_expression(class ir_expression *);
97 void check_ssbo_unsized_array_length_assignment(ir_assignment *ir);
98
99 ir_expression *process_ssbo_unsized_array_length(ir_rvalue **,
100 ir_dereference *,
101 ir_variable *);
102 ir_expression *emit_ssbo_get_buffer_size(void *mem_ctx);
103
104 unsigned calculate_unsized_array_stride(ir_dereference *deref,
105 enum glsl_interface_packing packing);
106
107 ir_call *lower_ssbo_atomic_intrinsic(ir_call *ir);
108 ir_call *check_for_ssbo_atomic_intrinsic(ir_call *ir);
109 ir_visitor_status visit_enter(ir_call *ir);
110 ir_visitor_status visit_enter(ir_texture *ir);
111
112 struct gl_linked_shader *shader;
113 bool clamp_block_indices;
114 const struct glsl_struct_field *struct_field;
115 ir_variable *variable;
116 ir_rvalue *uniform_block;
117 bool progress;
118 };
119
120 /**
121 * Determine the name of the interface block field
122 *
123 * This is the name of the specific member as it would appear in the
124 * \c gl_uniform_buffer_variable::Name field in the shader's
125 * \c UniformBlocks array.
126 */
127 static const char *
128 interface_field_name(void *mem_ctx, char *base_name, ir_rvalue *d,
129 ir_rvalue **nonconst_block_index)
130 {
131 *nonconst_block_index = NULL;
132 char *name_copy = NULL;
133 size_t base_length = 0;
134
135 /* Loop back through the IR until we find the uniform block */
136 ir_rvalue *ir = d;
137 while (ir != NULL) {
138 switch (ir->ir_type) {
139 case ir_type_dereference_variable: {
140 /* Exit loop */
141 ir = NULL;
142 break;
143 }
144
145 case ir_type_dereference_record: {
146 ir_dereference_record *r = (ir_dereference_record *) ir;
147 ir = r->record->as_dereference();
148
149 /* If we got here it means any previous array subscripts belong to
150 * block members and not the block itself so skip over them in the
151 * next pass.
152 */
153 d = ir;
154 break;
155 }
156
157 case ir_type_dereference_array: {
158 ir_dereference_array *a = (ir_dereference_array *) ir;
159 ir = a->array->as_dereference();
160 break;
161 }
162
163 case ir_type_swizzle: {
164 ir_swizzle *s = (ir_swizzle *) ir;
165 ir = s->val->as_dereference();
166 /* Skip swizzle in the next pass */
167 d = ir;
168 break;
169 }
170
171 default:
172 assert(!"Should not get here.");
173 break;
174 }
175 }
176
177 while (d != NULL) {
178 switch (d->ir_type) {
179 case ir_type_dereference_variable: {
180 ir_dereference_variable *v = (ir_dereference_variable *) d;
181 if (name_copy != NULL &&
182 v->var->is_interface_instance() &&
183 v->var->type->is_array()) {
184 return name_copy;
185 } else {
186 *nonconst_block_index = NULL;
187 return base_name;
188 }
189
190 break;
191 }
192
193 case ir_type_dereference_array: {
194 ir_dereference_array *a = (ir_dereference_array *) d;
195 size_t new_length;
196
197 if (name_copy == NULL) {
198 name_copy = ralloc_strdup(mem_ctx, base_name);
199 base_length = strlen(name_copy);
200 }
201
202 /* For arrays of arrays we start at the innermost array and work our
203 * way out so we need to insert the subscript at the base of the
204 * name string rather than just attaching it to the end.
205 */
206 new_length = base_length;
207 ir_constant *const_index = a->array_index->as_constant();
208 char *end = ralloc_strdup(NULL, &name_copy[new_length]);
209 if (!const_index) {
210 ir_rvalue *array_index = a->array_index;
211 if (array_index->type != glsl_type::uint_type)
212 array_index = i2u(array_index);
213
214 if (a->array->type->is_array() &&
215 a->array->type->fields.array->is_array()) {
216 ir_constant *base_size = new(mem_ctx)
217 ir_constant(a->array->type->fields.array->arrays_of_arrays_size());
218 array_index = mul(array_index, base_size);
219 }
220
221 if (*nonconst_block_index) {
222 *nonconst_block_index = add(*nonconst_block_index, array_index);
223 } else {
224 *nonconst_block_index = array_index;
225 }
226
227 ralloc_asprintf_rewrite_tail(&name_copy, &new_length, "[0]%s",
228 end);
229 } else {
230 ralloc_asprintf_rewrite_tail(&name_copy, &new_length, "[%d]%s",
231 const_index->get_uint_component(0),
232 end);
233 }
234 ralloc_free(end);
235
236 d = a->array->as_dereference();
237
238 break;
239 }
240
241 default:
242 assert(!"Should not get here.");
243 break;
244 }
245 }
246
247 assert(!"Should not get here.");
248 return NULL;
249 }
250
251 static ir_rvalue *
252 clamp_to_array_bounds(void *mem_ctx, ir_rvalue *index, const glsl_type *type)
253 {
254 assert(type->is_array());
255
256 const unsigned array_size = type->arrays_of_arrays_size();
257
258 ir_constant *max_index = new(mem_ctx) ir_constant(array_size - 1);
259 max_index->type = index->type;
260
261 ir_constant *zero = new(mem_ctx) ir_constant(0);
262 zero->type = index->type;
263
264 if (index->type->base_type == GLSL_TYPE_INT)
265 index = max2(index, zero);
266 index = min2(index, max_index);
267
268 return index;
269 }
270
271 void
272 lower_ubo_reference_visitor::setup_for_load_or_store(void *mem_ctx,
273 ir_variable *var,
274 ir_rvalue *deref,
275 ir_rvalue **offset,
276 unsigned *const_offset,
277 bool *row_major,
278 const glsl_type **matrix_type,
279 enum glsl_interface_packing packing)
280 {
281 /* Determine the name of the interface block */
282 ir_rvalue *nonconst_block_index;
283 const char *const field_name =
284 interface_field_name(mem_ctx, (char *) var->get_interface_type()->name,
285 deref, &nonconst_block_index);
286
287 if (nonconst_block_index && clamp_block_indices) {
288 nonconst_block_index =
289 clamp_to_array_bounds(mem_ctx, nonconst_block_index, var->type);
290 }
291
292 /* Locate the block by interface name */
293 unsigned num_blocks;
294 struct gl_uniform_block **blocks;
295 if (this->buffer_access_type != ubo_load_access) {
296 num_blocks = shader->Program->info.num_ssbos;
297 blocks = shader->Program->sh.ShaderStorageBlocks;
298 } else {
299 num_blocks = shader->Program->info.num_ubos;
300 blocks = shader->Program->sh.UniformBlocks;
301 }
302 this->uniform_block = NULL;
303 for (unsigned i = 0; i < num_blocks; i++) {
304 if (strcmp(field_name, blocks[i]->Name) == 0) {
305
306 ir_constant *index = new(mem_ctx) ir_constant(i);
307
308 if (nonconst_block_index) {
309 this->uniform_block = add(nonconst_block_index, index);
310 } else {
311 this->uniform_block = index;
312 }
313
314 if (var->is_interface_instance()) {
315 *const_offset = 0;
316 } else {
317 *const_offset = blocks[i]->Uniforms[var->data.location].Offset;
318 }
319
320 break;
321 }
322 }
323
324 assert(this->uniform_block);
325
326 this->struct_field = NULL;
327 setup_buffer_access(mem_ctx, deref, offset, const_offset, row_major,
328 matrix_type, &this->struct_field, packing);
329 }
330
331 void
332 lower_ubo_reference_visitor::handle_rvalue(ir_rvalue **rvalue)
333 {
334 if (!*rvalue)
335 return;
336
337 ir_dereference *deref = (*rvalue)->as_dereference();
338 if (!deref)
339 return;
340
341 ir_variable *var = deref->variable_referenced();
342 if (!var || !var->is_in_buffer_block())
343 return;
344
345 void *mem_ctx = ralloc_parent(shader->ir);
346
347 ir_rvalue *offset = NULL;
348 unsigned const_offset;
349 bool row_major;
350 const glsl_type *matrix_type;
351
352 enum glsl_interface_packing packing =
353 var->get_interface_type()->
354 get_internal_ifc_packing(use_std430_as_default);
355
356 this->buffer_access_type =
357 var->is_in_shader_storage_block() ?
358 ssbo_load_access : ubo_load_access;
359 this->variable = var;
360
361 /* Compute the offset to the start if the dereference as well as other
362 * information we need to configure the write
363 */
364 setup_for_load_or_store(mem_ctx, var, deref,
365 &offset, &const_offset,
366 &row_major, &matrix_type,
367 packing);
368 assert(offset);
369
370 /* Now that we've calculated the offset to the start of the
371 * dereference, walk over the type and emit loads into a temporary.
372 */
373 const glsl_type *type = (*rvalue)->type;
374 ir_variable *load_var = new(mem_ctx) ir_variable(type,
375 "ubo_load_temp",
376 ir_var_temporary);
377 base_ir->insert_before(load_var);
378
379 ir_variable *load_offset = new(mem_ctx) ir_variable(glsl_type::uint_type,
380 "ubo_load_temp_offset",
381 ir_var_temporary);
382 base_ir->insert_before(load_offset);
383 base_ir->insert_before(assign(load_offset, offset));
384
385 deref = new(mem_ctx) ir_dereference_variable(load_var);
386 emit_access(mem_ctx, false, deref, load_offset, const_offset,
387 row_major, matrix_type, packing, 0);
388 *rvalue = deref;
389
390 progress = true;
391 }
392
393 ir_expression *
394 lower_ubo_reference_visitor::ubo_load(void *mem_ctx,
395 const glsl_type *type,
396 ir_rvalue *offset)
397 {
398 ir_rvalue *block_ref = this->uniform_block->clone(mem_ctx, NULL);
399 return new(mem_ctx)
400 ir_expression(ir_binop_ubo_load,
401 type,
402 block_ref,
403 offset);
404
405 }
406
407 static bool
408 shader_storage_buffer_object(const _mesa_glsl_parse_state *state)
409 {
410 return state->has_shader_storage_buffer_objects();
411 }
412
413 uint32_t
414 lower_ubo_reference_visitor::ssbo_access_params()
415 {
416 assert(variable);
417
418 if (variable->is_interface_instance()) {
419 assert(struct_field);
420
421 return ((struct_field->memory_coherent ? ACCESS_COHERENT : 0) |
422 (struct_field->memory_restrict ? ACCESS_RESTRICT : 0) |
423 (struct_field->memory_volatile ? ACCESS_VOLATILE : 0));
424 } else {
425 return ((variable->data.memory_coherent ? ACCESS_COHERENT : 0) |
426 (variable->data.memory_restrict ? ACCESS_RESTRICT : 0) |
427 (variable->data.memory_volatile ? ACCESS_VOLATILE : 0));
428 }
429 }
430
431 ir_call *
432 lower_ubo_reference_visitor::ssbo_store(void *mem_ctx,
433 ir_rvalue *deref,
434 ir_rvalue *offset,
435 unsigned write_mask)
436 {
437 exec_list sig_params;
438
439 ir_variable *block_ref = new(mem_ctx)
440 ir_variable(glsl_type::uint_type, "block_ref" , ir_var_function_in);
441 sig_params.push_tail(block_ref);
442
443 ir_variable *offset_ref = new(mem_ctx)
444 ir_variable(glsl_type::uint_type, "offset" , ir_var_function_in);
445 sig_params.push_tail(offset_ref);
446
447 ir_variable *val_ref = new(mem_ctx)
448 ir_variable(deref->type, "value" , ir_var_function_in);
449 sig_params.push_tail(val_ref);
450
451 ir_variable *writemask_ref = new(mem_ctx)
452 ir_variable(glsl_type::uint_type, "write_mask" , ir_var_function_in);
453 sig_params.push_tail(writemask_ref);
454
455 ir_variable *access_ref = new(mem_ctx)
456 ir_variable(glsl_type::uint_type, "access" , ir_var_function_in);
457 sig_params.push_tail(access_ref);
458
459 ir_function_signature *sig = new(mem_ctx)
460 ir_function_signature(glsl_type::void_type, shader_storage_buffer_object);
461 assert(sig);
462 sig->replace_parameters(&sig_params);
463 sig->intrinsic_id = ir_intrinsic_ssbo_store;
464
465 ir_function *f = new(mem_ctx) ir_function("__intrinsic_store_ssbo");
466 f->add_signature(sig);
467
468 exec_list call_params;
469 call_params.push_tail(this->uniform_block->clone(mem_ctx, NULL));
470 call_params.push_tail(offset->clone(mem_ctx, NULL));
471 call_params.push_tail(deref->clone(mem_ctx, NULL));
472 call_params.push_tail(new(mem_ctx) ir_constant(write_mask));
473 call_params.push_tail(new(mem_ctx) ir_constant(ssbo_access_params()));
474 return new(mem_ctx) ir_call(sig, NULL, &call_params);
475 }
476
477 ir_call *
478 lower_ubo_reference_visitor::ssbo_load(void *mem_ctx,
479 const struct glsl_type *type,
480 ir_rvalue *offset)
481 {
482 exec_list sig_params;
483
484 ir_variable *block_ref = new(mem_ctx)
485 ir_variable(glsl_type::uint_type, "block_ref" , ir_var_function_in);
486 sig_params.push_tail(block_ref);
487
488 ir_variable *offset_ref = new(mem_ctx)
489 ir_variable(glsl_type::uint_type, "offset_ref" , ir_var_function_in);
490 sig_params.push_tail(offset_ref);
491
492 ir_variable *access_ref = new(mem_ctx)
493 ir_variable(glsl_type::uint_type, "access" , ir_var_function_in);
494 sig_params.push_tail(access_ref);
495
496 ir_function_signature *sig =
497 new(mem_ctx) ir_function_signature(type, shader_storage_buffer_object);
498 assert(sig);
499 sig->replace_parameters(&sig_params);
500 sig->intrinsic_id = ir_intrinsic_ssbo_load;
501
502 ir_function *f = new(mem_ctx) ir_function("__intrinsic_load_ssbo");
503 f->add_signature(sig);
504
505 ir_variable *result = new(mem_ctx)
506 ir_variable(type, "ssbo_load_result", ir_var_temporary);
507 base_ir->insert_before(result);
508 ir_dereference_variable *deref_result = new(mem_ctx)
509 ir_dereference_variable(result);
510
511 exec_list call_params;
512 call_params.push_tail(this->uniform_block->clone(mem_ctx, NULL));
513 call_params.push_tail(offset->clone(mem_ctx, NULL));
514 call_params.push_tail(new(mem_ctx) ir_constant(ssbo_access_params()));
515
516 return new(mem_ctx) ir_call(sig, deref_result, &call_params);
517 }
518
519 void
520 lower_ubo_reference_visitor::insert_buffer_access(void *mem_ctx,
521 ir_dereference *deref,
522 const glsl_type *type,
523 ir_rvalue *offset,
524 unsigned mask,
525 int channel)
526 {
527 switch (this->buffer_access_type) {
528 case ubo_load_access:
529 base_ir->insert_before(assign(deref->clone(mem_ctx, NULL),
530 ubo_load(mem_ctx, type, offset),
531 mask));
532 break;
533 case ssbo_load_access: {
534 ir_call *load_ssbo = ssbo_load(mem_ctx, type, offset);
535 base_ir->insert_before(load_ssbo);
536 ir_rvalue *value = load_ssbo->return_deref->as_rvalue()->clone(mem_ctx, NULL);
537 ir_assignment *assignment =
538 assign(deref->clone(mem_ctx, NULL), value, mask);
539 base_ir->insert_before(assignment);
540 break;
541 }
542 case ssbo_store_access:
543 if (channel >= 0) {
544 base_ir->insert_after(ssbo_store(mem_ctx,
545 swizzle(deref, channel, 1),
546 offset, 1));
547 } else {
548 base_ir->insert_after(ssbo_store(mem_ctx, deref, offset, mask));
549 }
550 break;
551 default:
552 unreachable("invalid buffer_access_type in insert_buffer_access");
553 }
554 }
555
556 void
557 lower_ubo_reference_visitor::write_to_memory(void *mem_ctx,
558 ir_dereference *deref,
559 ir_variable *var,
560 ir_variable *write_var,
561 unsigned write_mask)
562 {
563 ir_rvalue *offset = NULL;
564 unsigned const_offset;
565 bool row_major;
566 const glsl_type *matrix_type;
567
568 enum glsl_interface_packing packing =
569 var->get_interface_type()->
570 get_internal_ifc_packing(use_std430_as_default);
571
572 this->buffer_access_type = ssbo_store_access;
573 this->variable = var;
574
575 /* Compute the offset to the start if the dereference as well as other
576 * information we need to configure the write
577 */
578 setup_for_load_or_store(mem_ctx, var, deref,
579 &offset, &const_offset,
580 &row_major, &matrix_type,
581 packing);
582 assert(offset);
583
584 /* Now emit writes from the temporary to memory */
585 ir_variable *write_offset =
586 new(mem_ctx) ir_variable(glsl_type::uint_type,
587 "ssbo_store_temp_offset",
588 ir_var_temporary);
589
590 base_ir->insert_before(write_offset);
591 base_ir->insert_before(assign(write_offset, offset));
592
593 deref = new(mem_ctx) ir_dereference_variable(write_var);
594 emit_access(mem_ctx, true, deref, write_offset, const_offset,
595 row_major, matrix_type, packing, write_mask);
596 }
597
598 ir_visitor_status
599 lower_ubo_reference_visitor::visit_enter(ir_expression *ir)
600 {
601 check_ssbo_unsized_array_length_expression(ir);
602 return rvalue_visit(ir);
603 }
604
605 ir_expression *
606 lower_ubo_reference_visitor::calculate_ssbo_unsized_array_length(ir_expression *expr)
607 {
608 if (expr->operation !=
609 ir_expression_operation(ir_unop_ssbo_unsized_array_length))
610 return NULL;
611
612 ir_rvalue *rvalue = expr->operands[0]->as_rvalue();
613 if (!rvalue ||
614 !rvalue->type->is_array() || !rvalue->type->is_unsized_array())
615 return NULL;
616
617 ir_dereference *deref = expr->operands[0]->as_dereference();
618 if (!deref)
619 return NULL;
620
621 ir_variable *var = expr->operands[0]->variable_referenced();
622 if (!var || !var->is_in_shader_storage_block())
623 return NULL;
624 return process_ssbo_unsized_array_length(&rvalue, deref, var);
625 }
626
627 void
628 lower_ubo_reference_visitor::check_ssbo_unsized_array_length_expression(ir_expression *ir)
629 {
630 if (ir->operation ==
631 ir_expression_operation(ir_unop_ssbo_unsized_array_length)) {
632 /* Don't replace this unop if it is found alone. It is going to be
633 * removed by the optimization passes or replaced if it is part of
634 * an ir_assignment or another ir_expression.
635 */
636 return;
637 }
638
639 for (unsigned i = 0; i < ir->num_operands; i++) {
640 if (ir->operands[i]->ir_type != ir_type_expression)
641 continue;
642 ir_expression *expr = (ir_expression *) ir->operands[i];
643 ir_expression *temp = calculate_ssbo_unsized_array_length(expr);
644 if (!temp)
645 continue;
646
647 delete expr;
648 ir->operands[i] = temp;
649 }
650 }
651
652 void
653 lower_ubo_reference_visitor::check_ssbo_unsized_array_length_assignment(ir_assignment *ir)
654 {
655 if (!ir->rhs || ir->rhs->ir_type != ir_type_expression)
656 return;
657
658 ir_expression *expr = (ir_expression *) ir->rhs;
659 ir_expression *temp = calculate_ssbo_unsized_array_length(expr);
660 if (!temp)
661 return;
662
663 delete expr;
664 ir->rhs = temp;
665 return;
666 }
667
668 ir_expression *
669 lower_ubo_reference_visitor::emit_ssbo_get_buffer_size(void *mem_ctx)
670 {
671 ir_rvalue *block_ref = this->uniform_block->clone(mem_ctx, NULL);
672 return new(mem_ctx) ir_expression(ir_unop_get_buffer_size,
673 glsl_type::int_type,
674 block_ref);
675 }
676
677 unsigned
678 lower_ubo_reference_visitor::calculate_unsized_array_stride(ir_dereference *deref,
679 enum glsl_interface_packing packing)
680 {
681 unsigned array_stride = 0;
682
683 switch (deref->ir_type) {
684 case ir_type_dereference_variable:
685 {
686 ir_dereference_variable *deref_var = (ir_dereference_variable *)deref;
687 const struct glsl_type *unsized_array_type = NULL;
688 /* An unsized array can be sized by other lowering passes, so pick
689 * the first field of the array which has the data type of the unsized
690 * array.
691 */
692 unsized_array_type = deref_var->var->type->fields.array;
693
694 /* Whether or not the field is row-major (because it might be a
695 * bvec2 or something) does not affect the array itself. We need
696 * to know whether an array element in its entirety is row-major.
697 */
698 const bool array_row_major =
699 is_dereferenced_thing_row_major(deref_var);
700
701 if (packing == GLSL_INTERFACE_PACKING_STD430) {
702 array_stride = unsized_array_type->std430_array_stride(array_row_major);
703 } else {
704 array_stride = unsized_array_type->std140_size(array_row_major);
705 array_stride = glsl_align(array_stride, 16);
706 }
707 break;
708 }
709 case ir_type_dereference_record:
710 {
711 ir_dereference_record *deref_record = (ir_dereference_record *) deref;
712 ir_dereference *interface_deref =
713 deref_record->record->as_dereference();
714 assert(interface_deref != NULL);
715 const struct glsl_type *interface_type = interface_deref->type;
716 unsigned record_length = interface_type->length;
717 /* Unsized array is always the last element of the interface */
718 const struct glsl_type *unsized_array_type =
719 interface_type->fields.structure[record_length - 1].type->fields.array;
720
721 const bool array_row_major =
722 is_dereferenced_thing_row_major(deref_record);
723
724 if (packing == GLSL_INTERFACE_PACKING_STD430) {
725 array_stride = unsized_array_type->std430_array_stride(array_row_major);
726 } else {
727 array_stride = unsized_array_type->std140_size(array_row_major);
728 array_stride = glsl_align(array_stride, 16);
729 }
730 break;
731 }
732 default:
733 unreachable("Unsupported dereference type");
734 }
735 return array_stride;
736 }
737
738 ir_expression *
739 lower_ubo_reference_visitor::process_ssbo_unsized_array_length(ir_rvalue **rvalue,
740 ir_dereference *deref,
741 ir_variable *var)
742 {
743 void *mem_ctx = ralloc_parent(*rvalue);
744
745 ir_rvalue *base_offset = NULL;
746 unsigned const_offset;
747 bool row_major;
748 const glsl_type *matrix_type;
749
750 enum glsl_interface_packing packing =
751 var->get_interface_type()->
752 get_internal_ifc_packing(use_std430_as_default);
753 int unsized_array_stride =
754 calculate_unsized_array_stride(deref, packing);
755
756 this->buffer_access_type = ssbo_unsized_array_length_access;
757 this->variable = var;
758
759 /* Compute the offset to the start if the dereference as well as other
760 * information we need to calculate the length.
761 */
762 setup_for_load_or_store(mem_ctx, var, deref,
763 &base_offset, &const_offset,
764 &row_major, &matrix_type,
765 packing);
766 /* array.length() =
767 * max((buffer_object_size - offset_of_array) / stride_of_array, 0)
768 */
769 ir_expression *buffer_size = emit_ssbo_get_buffer_size(mem_ctx);
770
771 ir_expression *offset_of_array = new(mem_ctx)
772 ir_expression(ir_binop_add, base_offset,
773 new(mem_ctx) ir_constant(const_offset));
774 ir_expression *offset_of_array_int = new(mem_ctx)
775 ir_expression(ir_unop_u2i, offset_of_array);
776
777 ir_expression *sub = new(mem_ctx)
778 ir_expression(ir_binop_sub, buffer_size, offset_of_array_int);
779 ir_expression *div = new(mem_ctx)
780 ir_expression(ir_binop_div, sub,
781 new(mem_ctx) ir_constant(unsized_array_stride));
782 ir_expression *max = new(mem_ctx)
783 ir_expression(ir_binop_max, div, new(mem_ctx) ir_constant(0));
784
785 return max;
786 }
787
788 void
789 lower_ubo_reference_visitor::check_for_ssbo_store(ir_assignment *ir)
790 {
791 if (!ir || !ir->lhs)
792 return;
793
794 ir_rvalue *rvalue = ir->lhs->as_rvalue();
795 if (!rvalue)
796 return;
797
798 ir_dereference *deref = ir->lhs->as_dereference();
799 if (!deref)
800 return;
801
802 ir_variable *var = ir->lhs->variable_referenced();
803 if (!var || !var->is_in_shader_storage_block())
804 return;
805
806 /* We have a write to a buffer variable, so declare a temporary and rewrite
807 * the assignment so that the temporary is the LHS.
808 */
809 void *mem_ctx = ralloc_parent(shader->ir);
810
811 const glsl_type *type = rvalue->type;
812 ir_variable *write_var = new(mem_ctx) ir_variable(type,
813 "ssbo_store_temp",
814 ir_var_temporary);
815 base_ir->insert_before(write_var);
816 ir->lhs = new(mem_ctx) ir_dereference_variable(write_var);
817
818 /* Now we have to write the value assigned to the temporary back to memory */
819 write_to_memory(mem_ctx, deref, var, write_var, ir->write_mask);
820 progress = true;
821 }
822
823 static bool
824 is_buffer_backed_variable(ir_variable *var)
825 {
826 return var->is_in_buffer_block() ||
827 var->data.mode == ir_var_shader_shared;
828 }
829
830 bool
831 lower_ubo_reference_visitor::check_for_buffer_array_copy(ir_assignment *ir)
832 {
833 if (!ir || !ir->lhs || !ir->rhs)
834 return false;
835
836 /* LHS and RHS must be arrays
837 * FIXME: arrays of arrays?
838 */
839 if (!ir->lhs->type->is_array() || !ir->rhs->type->is_array())
840 return false;
841
842 /* RHS must be a buffer-backed variable. This is what can cause the problem
843 * since it would lead to a series of loads that need to live until we
844 * see the writes to the LHS.
845 */
846 ir_variable *rhs_var = ir->rhs->variable_referenced();
847 if (!rhs_var || !is_buffer_backed_variable(rhs_var))
848 return false;
849
850 /* Split the array copy into individual element copies to reduce
851 * register pressure
852 */
853 ir_dereference *rhs_deref = ir->rhs->as_dereference();
854 if (!rhs_deref)
855 return false;
856
857 ir_dereference *lhs_deref = ir->lhs->as_dereference();
858 if (!lhs_deref)
859 return false;
860
861 assert(lhs_deref->type->length == rhs_deref->type->length);
862 void *mem_ctx = ralloc_parent(shader->ir);
863
864 for (unsigned i = 0; i < lhs_deref->type->length; i++) {
865 ir_dereference *lhs_i =
866 new(mem_ctx) ir_dereference_array(lhs_deref->clone(mem_ctx, NULL),
867 new(mem_ctx) ir_constant(i));
868
869 ir_dereference *rhs_i =
870 new(mem_ctx) ir_dereference_array(rhs_deref->clone(mem_ctx, NULL),
871 new(mem_ctx) ir_constant(i));
872 ir->insert_after(assign(lhs_i, rhs_i));
873 }
874
875 ir->remove();
876 progress = true;
877 return true;
878 }
879
880 bool
881 lower_ubo_reference_visitor::check_for_buffer_struct_copy(ir_assignment *ir)
882 {
883 if (!ir || !ir->lhs || !ir->rhs)
884 return false;
885
886 /* LHS and RHS must be records */
887 if (!ir->lhs->type->is_struct() || !ir->rhs->type->is_struct())
888 return false;
889
890 /* RHS must be a buffer-backed variable. This is what can cause the problem
891 * since it would lead to a series of loads that need to live until we
892 * see the writes to the LHS.
893 */
894 ir_variable *rhs_var = ir->rhs->variable_referenced();
895 if (!rhs_var || !is_buffer_backed_variable(rhs_var))
896 return false;
897
898 /* Split the struct copy into individual element copies to reduce
899 * register pressure
900 */
901 ir_dereference *rhs_deref = ir->rhs->as_dereference();
902 if (!rhs_deref)
903 return false;
904
905 ir_dereference *lhs_deref = ir->lhs->as_dereference();
906 if (!lhs_deref)
907 return false;
908
909 assert(lhs_deref->type == rhs_deref->type);
910 void *mem_ctx = ralloc_parent(shader->ir);
911
912 for (unsigned i = 0; i < lhs_deref->type->length; i++) {
913 const char *field_name = lhs_deref->type->fields.structure[i].name;
914 ir_dereference *lhs_field =
915 new(mem_ctx) ir_dereference_record(lhs_deref->clone(mem_ctx, NULL),
916 field_name);
917 ir_dereference *rhs_field =
918 new(mem_ctx) ir_dereference_record(rhs_deref->clone(mem_ctx, NULL),
919 field_name);
920 ir->insert_after(assign(lhs_field, rhs_field));
921 }
922
923 ir->remove();
924 progress = true;
925 return true;
926 }
927
928 ir_visitor_status
929 lower_ubo_reference_visitor::visit_enter(ir_assignment *ir)
930 {
931 /* Array and struct copies could involve large amounts of load/store
932 * operations. To improve register pressure we want to special-case
933 * these and split them into individual element copies.
934 * This way we avoid emitting all the loads for the RHS first and
935 * all the writes for the LHS second and register usage is more
936 * efficient.
937 */
938 if (check_for_buffer_array_copy(ir))
939 return visit_continue_with_parent;
940
941 if (check_for_buffer_struct_copy(ir))
942 return visit_continue_with_parent;
943
944 check_ssbo_unsized_array_length_assignment(ir);
945 check_for_ssbo_store(ir);
946 return rvalue_visit(ir);
947 }
948
949 /* Lowers the intrinsic call to a new internal intrinsic that swaps the
950 * access to the buffer variable in the first parameter by an offset
951 * and block index. This involves creating the new internal intrinsic
952 * (i.e. the new function signature).
953 */
954 ir_call *
955 lower_ubo_reference_visitor::lower_ssbo_atomic_intrinsic(ir_call *ir)
956 {
957 /* SSBO atomics usually have 2 parameters, the buffer variable and an
958 * integer argument. The exception is CompSwap, that has an additional
959 * integer parameter.
960 */
961 int param_count = ir->actual_parameters.length();
962 assert(param_count == 2 || param_count == 3);
963
964 /* First argument must be a scalar integer buffer variable */
965 exec_node *param = ir->actual_parameters.get_head();
966 ir_instruction *inst = (ir_instruction *) param;
967 assert(inst->ir_type == ir_type_dereference_variable ||
968 inst->ir_type == ir_type_dereference_array ||
969 inst->ir_type == ir_type_dereference_record ||
970 inst->ir_type == ir_type_swizzle);
971
972 ir_rvalue *deref = (ir_rvalue *) inst;
973 assert(deref->type->is_scalar() &&
974 (deref->type->is_integer_32() || deref->type->is_float()));
975
976 ir_variable *var = deref->variable_referenced();
977 assert(var);
978
979 /* Compute the offset to the start if the dereference and the
980 * block index
981 */
982 void *mem_ctx = ralloc_parent(shader->ir);
983
984 ir_rvalue *offset = NULL;
985 unsigned const_offset;
986 bool row_major;
987 const glsl_type *matrix_type;
988
989 enum glsl_interface_packing packing =
990 var->get_interface_type()->
991 get_internal_ifc_packing(use_std430_as_default);
992
993 this->buffer_access_type = ssbo_atomic_access;
994 this->variable = var;
995
996 setup_for_load_or_store(mem_ctx, var, deref,
997 &offset, &const_offset,
998 &row_major, &matrix_type,
999 packing);
1000 assert(offset);
1001 assert(!row_major);
1002 assert(matrix_type == NULL);
1003
1004 ir_rvalue *deref_offset =
1005 add(offset, new(mem_ctx) ir_constant(const_offset));
1006 ir_rvalue *block_index = this->uniform_block->clone(mem_ctx, NULL);
1007
1008 /* Create the new internal function signature that will take a block
1009 * index and offset instead of a buffer variable
1010 */
1011 exec_list sig_params;
1012 ir_variable *sig_param = new(mem_ctx)
1013 ir_variable(glsl_type::uint_type, "block_ref" , ir_var_function_in);
1014 sig_params.push_tail(sig_param);
1015
1016 sig_param = new(mem_ctx)
1017 ir_variable(glsl_type::uint_type, "offset" , ir_var_function_in);
1018 sig_params.push_tail(sig_param);
1019
1020 const glsl_type *type = deref->type->get_scalar_type();
1021 sig_param = new(mem_ctx)
1022 ir_variable(type, "data1", ir_var_function_in);
1023 sig_params.push_tail(sig_param);
1024
1025 if (param_count == 3) {
1026 sig_param = new(mem_ctx)
1027 ir_variable(type, "data2", ir_var_function_in);
1028 sig_params.push_tail(sig_param);
1029 }
1030
1031 ir_function_signature *sig =
1032 new(mem_ctx) ir_function_signature(deref->type,
1033 shader_storage_buffer_object);
1034 assert(sig);
1035 sig->replace_parameters(&sig_params);
1036
1037 assert(ir->callee->intrinsic_id >= ir_intrinsic_generic_load);
1038 assert(ir->callee->intrinsic_id <= ir_intrinsic_generic_atomic_comp_swap);
1039 sig->intrinsic_id = MAP_INTRINSIC_TO_TYPE(ir->callee->intrinsic_id, ssbo);
1040
1041 char func_name[64];
1042 sprintf(func_name, "%s_ssbo", ir->callee_name());
1043 ir_function *f = new(mem_ctx) ir_function(func_name);
1044 f->add_signature(sig);
1045
1046 /* Now, create the call to the internal intrinsic */
1047 exec_list call_params;
1048 call_params.push_tail(block_index);
1049 call_params.push_tail(deref_offset);
1050 param = ir->actual_parameters.get_head()->get_next();
1051 ir_rvalue *param_as_rvalue = ((ir_instruction *) param)->as_rvalue();
1052 call_params.push_tail(param_as_rvalue->clone(mem_ctx, NULL));
1053 if (param_count == 3) {
1054 param = param->get_next();
1055 param_as_rvalue = ((ir_instruction *) param)->as_rvalue();
1056 call_params.push_tail(param_as_rvalue->clone(mem_ctx, NULL));
1057 }
1058 ir_dereference_variable *return_deref =
1059 ir->return_deref->clone(mem_ctx, NULL);
1060 return new(mem_ctx) ir_call(sig, return_deref, &call_params);
1061 }
1062
1063 ir_call *
1064 lower_ubo_reference_visitor::check_for_ssbo_atomic_intrinsic(ir_call *ir)
1065 {
1066 exec_list& params = ir->actual_parameters;
1067
1068 if (params.length() < 2 || params.length() > 3)
1069 return ir;
1070
1071 ir_rvalue *rvalue =
1072 ((ir_instruction *) params.get_head())->as_rvalue();
1073 if (!rvalue)
1074 return ir;
1075
1076 ir_variable *var = rvalue->variable_referenced();
1077 if (!var || !var->is_in_shader_storage_block())
1078 return ir;
1079
1080 const enum ir_intrinsic_id id = ir->callee->intrinsic_id;
1081 if (id == ir_intrinsic_generic_atomic_add ||
1082 id == ir_intrinsic_generic_atomic_min ||
1083 id == ir_intrinsic_generic_atomic_max ||
1084 id == ir_intrinsic_generic_atomic_and ||
1085 id == ir_intrinsic_generic_atomic_or ||
1086 id == ir_intrinsic_generic_atomic_xor ||
1087 id == ir_intrinsic_generic_atomic_exchange ||
1088 id == ir_intrinsic_generic_atomic_comp_swap) {
1089 return lower_ssbo_atomic_intrinsic(ir);
1090 }
1091
1092 return ir;
1093 }
1094
1095
1096 ir_visitor_status
1097 lower_ubo_reference_visitor::visit_enter(ir_call *ir)
1098 {
1099 ir_call *new_ir = check_for_ssbo_atomic_intrinsic(ir);
1100 if (new_ir != ir) {
1101 progress = true;
1102 base_ir->replace_with(new_ir);
1103 return visit_continue_with_parent;
1104 }
1105
1106 return rvalue_visit(ir);
1107 }
1108
1109
1110 ir_visitor_status
1111 lower_ubo_reference_visitor::visit_enter(ir_texture *ir)
1112 {
1113 ir_dereference *sampler = ir->sampler;
1114
1115 if (sampler->ir_type == ir_type_dereference_record) {
1116 handle_rvalue((ir_rvalue **)&ir->sampler);
1117 return visit_continue_with_parent;
1118 }
1119
1120 return rvalue_visit(ir);
1121 }
1122
1123
1124 } /* unnamed namespace */
1125
1126 void
1127 lower_ubo_reference(struct gl_linked_shader *shader,
1128 bool clamp_block_indices, bool use_std430_as_default)
1129 {
1130 lower_ubo_reference_visitor v(shader, clamp_block_indices,
1131 use_std430_as_default);
1132
1133 /* Loop over the instructions lowering references, because we take
1134 * a deref of a UBO array using a UBO dereference as the index will
1135 * produce a collection of instructions all of which have cloned
1136 * UBO dereferences for that array index.
1137 */
1138 do {
1139 v.progress = false;
1140 visit_list_elements(&v, shader->ir);
1141 } while (v.progress);
1142 }