2 * Copyright © 2012 Intel Corporation
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:
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
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.
25 * \file lower_ubo_reference.cpp
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.
31 * This relieves drivers of the responsibility to deal with tricky UBO
32 * layout issues like std140 structures and row_major matrices on
36 #include "lower_buffer_access.h"
37 #include "ir_builder.h"
38 #include "main/macros.h"
39 #include "glsl_parser_extras.h"
41 using namespace ir_builder
;
44 class lower_ubo_reference_visitor
:
45 public lower_buffer_access::lower_buffer_access
{
47 lower_ubo_reference_visitor(struct gl_linked_shader
*shader
,
48 bool clamp_block_indices
)
49 : shader(shader
), clamp_block_indices(clamp_block_indices
),
50 struct_field(NULL
), variable(NULL
)
54 void handle_rvalue(ir_rvalue
**rvalue
);
55 ir_visitor_status
visit_enter(ir_assignment
*ir
);
57 void setup_for_load_or_store(void *mem_ctx
,
61 unsigned *const_offset
,
64 enum glsl_interface_packing packing
);
65 uint32_t ssbo_access_params();
66 ir_expression
*ubo_load(void *mem_ctx
, const struct glsl_type
*type
,
68 ir_call
*ssbo_load(void *mem_ctx
, const struct glsl_type
*type
,
71 bool check_for_buffer_array_copy(ir_assignment
*ir
);
72 bool check_for_buffer_struct_copy(ir_assignment
*ir
);
73 void check_for_ssbo_store(ir_assignment
*ir
);
74 void write_to_memory(void *mem_ctx
, ir_dereference
*deref
, ir_variable
*var
,
75 ir_variable
*write_var
, unsigned write_mask
);
76 ir_call
*ssbo_store(void *mem_ctx
, ir_rvalue
*deref
, ir_rvalue
*offset
,
83 ssbo_unsized_array_length_access
,
87 void insert_buffer_access(void *mem_ctx
, ir_dereference
*deref
,
88 const glsl_type
*type
, ir_rvalue
*offset
,
89 unsigned mask
, int channel
);
91 ir_visitor_status
visit_enter(class ir_expression
*);
92 ir_expression
*calculate_ssbo_unsized_array_length(ir_expression
*expr
);
93 void check_ssbo_unsized_array_length_expression(class ir_expression
*);
94 void check_ssbo_unsized_array_length_assignment(ir_assignment
*ir
);
96 ir_expression
*process_ssbo_unsized_array_length(ir_rvalue
**,
99 ir_expression
*emit_ssbo_get_buffer_size(void *mem_ctx
);
101 unsigned calculate_unsized_array_stride(ir_dereference
*deref
,
102 enum glsl_interface_packing packing
);
104 ir_call
*lower_ssbo_atomic_intrinsic(ir_call
*ir
);
105 ir_call
*check_for_ssbo_atomic_intrinsic(ir_call
*ir
);
106 ir_visitor_status
visit_enter(ir_call
*ir
);
108 struct gl_linked_shader
*shader
;
109 bool clamp_block_indices
;
110 struct gl_uniform_buffer_variable
*ubo_var
;
111 const struct glsl_struct_field
*struct_field
;
112 ir_variable
*variable
;
113 ir_rvalue
*uniform_block
;
118 * Determine the name of the interface block field
120 * This is the name of the specific member as it would appear in the
121 * \c gl_uniform_buffer_variable::Name field in the shader's
122 * \c UniformBlocks array.
125 interface_field_name(void *mem_ctx
, char *base_name
, ir_rvalue
*d
,
126 ir_rvalue
**nonconst_block_index
)
128 *nonconst_block_index
= NULL
;
129 char *name_copy
= NULL
;
130 size_t base_length
= 0;
132 /* Loop back through the IR until we find the uniform block */
135 switch (ir
->ir_type
) {
136 case ir_type_dereference_variable
: {
142 case ir_type_dereference_record
: {
143 ir_dereference_record
*r
= (ir_dereference_record
*) ir
;
144 ir
= r
->record
->as_dereference();
146 /* If we got here it means any previous array subscripts belong to
147 * block members and not the block itself so skip over them in the
154 case ir_type_dereference_array
: {
155 ir_dereference_array
*a
= (ir_dereference_array
*) ir
;
156 ir
= a
->array
->as_dereference();
160 case ir_type_swizzle
: {
161 ir_swizzle
*s
= (ir_swizzle
*) ir
;
162 ir
= s
->val
->as_dereference();
163 /* Skip swizzle in the next pass */
169 assert(!"Should not get here.");
175 switch (d
->ir_type
) {
176 case ir_type_dereference_variable
: {
177 ir_dereference_variable
*v
= (ir_dereference_variable
*) d
;
178 if (name_copy
!= NULL
&&
179 v
->var
->is_interface_instance() &&
180 v
->var
->type
->is_array()) {
183 *nonconst_block_index
= NULL
;
190 case ir_type_dereference_array
: {
191 ir_dereference_array
*a
= (ir_dereference_array
*) d
;
194 if (name_copy
== NULL
) {
195 name_copy
= ralloc_strdup(mem_ctx
, base_name
);
196 base_length
= strlen(name_copy
);
199 /* For arrays of arrays we start at the innermost array and work our
200 * way out so we need to insert the subscript at the base of the
201 * name string rather than just attaching it to the end.
203 new_length
= base_length
;
204 ir_constant
*const_index
= a
->array_index
->as_constant();
205 char *end
= ralloc_strdup(NULL
, &name_copy
[new_length
]);
207 ir_rvalue
*array_index
= a
->array_index
;
208 if (array_index
->type
!= glsl_type::uint_type
)
209 array_index
= i2u(array_index
);
211 if (a
->array
->type
->is_array() &&
212 a
->array
->type
->fields
.array
->is_array()) {
213 ir_constant
*base_size
= new(mem_ctx
)
214 ir_constant(a
->array
->type
->fields
.array
->arrays_of_arrays_size());
215 array_index
= mul(array_index
, base_size
);
218 if (*nonconst_block_index
) {
219 *nonconst_block_index
= add(*nonconst_block_index
, array_index
);
221 *nonconst_block_index
= array_index
;
224 ralloc_asprintf_rewrite_tail(&name_copy
, &new_length
, "[0]%s",
227 ralloc_asprintf_rewrite_tail(&name_copy
, &new_length
, "[%d]%s",
228 const_index
->get_uint_component(0),
233 d
= a
->array
->as_dereference();
239 assert(!"Should not get here.");
244 assert(!"Should not get here.");
249 clamp_to_array_bounds(void *mem_ctx
, ir_rvalue
*index
, const glsl_type
*type
)
251 assert(type
->is_array());
253 const unsigned array_size
= type
->arrays_of_arrays_size();
255 ir_constant
*max_index
= new(mem_ctx
) ir_constant(array_size
- 1);
256 max_index
->type
= index
->type
;
258 ir_constant
*zero
= new(mem_ctx
) ir_constant(0);
259 zero
->type
= index
->type
;
261 if (index
->type
->base_type
== GLSL_TYPE_INT
)
262 index
= max2(index
, zero
);
263 index
= min2(index
, max_index
);
269 lower_ubo_reference_visitor::setup_for_load_or_store(void *mem_ctx
,
273 unsigned *const_offset
,
276 enum glsl_interface_packing packing
)
278 /* Determine the name of the interface block */
279 ir_rvalue
*nonconst_block_index
;
280 const char *const field_name
=
281 interface_field_name(mem_ctx
, (char *) var
->get_interface_type()->name
,
282 deref
, &nonconst_block_index
);
284 if (nonconst_block_index
&& clamp_block_indices
) {
285 nonconst_block_index
=
286 clamp_to_array_bounds(mem_ctx
, nonconst_block_index
, var
->type
);
289 /* Locate the block by interface name */
291 struct gl_uniform_block
**blocks
;
292 if (this->buffer_access_type
!= ubo_load_access
) {
293 num_blocks
= shader
->NumShaderStorageBlocks
;
294 blocks
= shader
->ShaderStorageBlocks
;
296 num_blocks
= shader
->NumUniformBlocks
;
297 blocks
= shader
->UniformBlocks
;
299 this->uniform_block
= NULL
;
300 for (unsigned i
= 0; i
< num_blocks
; i
++) {
301 if (strcmp(field_name
, blocks
[i
]->Name
) == 0) {
303 ir_constant
*index
= new(mem_ctx
) ir_constant(i
);
305 if (nonconst_block_index
) {
306 this->uniform_block
= add(nonconst_block_index
, index
);
308 this->uniform_block
= index
;
311 this->ubo_var
= var
->is_interface_instance()
312 ? &blocks
[i
]->Uniforms
[0] : &blocks
[i
]->Uniforms
[var
->data
.location
];
318 assert(this->uniform_block
);
320 *const_offset
= ubo_var
->Offset
;
322 this->struct_field
= NULL
;
323 setup_buffer_access(mem_ctx
, deref
, offset
, const_offset
, row_major
,
324 matrix_columns
, &this->struct_field
, packing
);
328 lower_ubo_reference_visitor::handle_rvalue(ir_rvalue
**rvalue
)
333 ir_dereference
*deref
= (*rvalue
)->as_dereference();
337 ir_variable
*var
= deref
->variable_referenced();
338 if (!var
|| !var
->is_in_buffer_block())
341 void *mem_ctx
= ralloc_parent(shader
->ir
);
343 ir_rvalue
*offset
= NULL
;
344 unsigned const_offset
;
347 enum glsl_interface_packing packing
= var
->get_interface_type_packing();
349 this->buffer_access_type
=
350 var
->is_in_shader_storage_block() ?
351 ssbo_load_access
: ubo_load_access
;
352 this->variable
= var
;
354 /* Compute the offset to the start if the dereference as well as other
355 * information we need to configure the write
357 setup_for_load_or_store(mem_ctx
, var
, deref
,
358 &offset
, &const_offset
,
359 &row_major
, &matrix_columns
,
363 /* Now that we've calculated the offset to the start of the
364 * dereference, walk over the type and emit loads into a temporary.
366 const glsl_type
*type
= (*rvalue
)->type
;
367 ir_variable
*load_var
= new(mem_ctx
) ir_variable(type
,
370 base_ir
->insert_before(load_var
);
372 ir_variable
*load_offset
= new(mem_ctx
) ir_variable(glsl_type::uint_type
,
373 "ubo_load_temp_offset",
375 base_ir
->insert_before(load_offset
);
376 base_ir
->insert_before(assign(load_offset
, offset
));
378 deref
= new(mem_ctx
) ir_dereference_variable(load_var
);
379 emit_access(mem_ctx
, false, deref
, load_offset
, const_offset
,
380 row_major
, matrix_columns
, packing
, 0);
387 lower_ubo_reference_visitor::ubo_load(void *mem_ctx
,
388 const glsl_type
*type
,
391 ir_rvalue
*block_ref
= this->uniform_block
->clone(mem_ctx
, NULL
);
393 ir_expression(ir_binop_ubo_load
,
401 shader_storage_buffer_object(const _mesa_glsl_parse_state
*state
)
403 return state
->has_shader_storage_buffer_objects();
407 lower_ubo_reference_visitor::ssbo_access_params()
411 if (variable
->is_interface_instance()) {
412 assert(struct_field
);
414 return ((struct_field
->image_coherent
? ACCESS_COHERENT
: 0) |
415 (struct_field
->image_restrict
? ACCESS_RESTRICT
: 0) |
416 (struct_field
->image_volatile
? ACCESS_VOLATILE
: 0));
418 return ((variable
->data
.image_coherent
? ACCESS_COHERENT
: 0) |
419 (variable
->data
.image_restrict
? ACCESS_RESTRICT
: 0) |
420 (variable
->data
.image_volatile
? ACCESS_VOLATILE
: 0));
425 lower_ubo_reference_visitor::ssbo_store(void *mem_ctx
,
430 exec_list sig_params
;
432 ir_variable
*block_ref
= new(mem_ctx
)
433 ir_variable(glsl_type::uint_type
, "block_ref" , ir_var_function_in
);
434 sig_params
.push_tail(block_ref
);
436 ir_variable
*offset_ref
= new(mem_ctx
)
437 ir_variable(glsl_type::uint_type
, "offset" , ir_var_function_in
);
438 sig_params
.push_tail(offset_ref
);
440 ir_variable
*val_ref
= new(mem_ctx
)
441 ir_variable(deref
->type
, "value" , ir_var_function_in
);
442 sig_params
.push_tail(val_ref
);
444 ir_variable
*writemask_ref
= new(mem_ctx
)
445 ir_variable(glsl_type::uint_type
, "write_mask" , ir_var_function_in
);
446 sig_params
.push_tail(writemask_ref
);
448 ir_variable
*access_ref
= new(mem_ctx
)
449 ir_variable(glsl_type::uint_type
, "access" , ir_var_function_in
);
450 sig_params
.push_tail(access_ref
);
452 ir_function_signature
*sig
= new(mem_ctx
)
453 ir_function_signature(glsl_type::void_type
, shader_storage_buffer_object
);
455 sig
->replace_parameters(&sig_params
);
456 sig
->is_intrinsic
= true;
458 ir_function
*f
= new(mem_ctx
) ir_function("__intrinsic_store_ssbo");
459 f
->add_signature(sig
);
461 exec_list call_params
;
462 call_params
.push_tail(this->uniform_block
->clone(mem_ctx
, NULL
));
463 call_params
.push_tail(offset
->clone(mem_ctx
, NULL
));
464 call_params
.push_tail(deref
->clone(mem_ctx
, NULL
));
465 call_params
.push_tail(new(mem_ctx
) ir_constant(write_mask
));
466 call_params
.push_tail(new(mem_ctx
) ir_constant(ssbo_access_params()));
467 return new(mem_ctx
) ir_call(sig
, NULL
, &call_params
);
471 lower_ubo_reference_visitor::ssbo_load(void *mem_ctx
,
472 const struct glsl_type
*type
,
475 exec_list sig_params
;
477 ir_variable
*block_ref
= new(mem_ctx
)
478 ir_variable(glsl_type::uint_type
, "block_ref" , ir_var_function_in
);
479 sig_params
.push_tail(block_ref
);
481 ir_variable
*offset_ref
= new(mem_ctx
)
482 ir_variable(glsl_type::uint_type
, "offset_ref" , ir_var_function_in
);
483 sig_params
.push_tail(offset_ref
);
485 ir_variable
*access_ref
= new(mem_ctx
)
486 ir_variable(glsl_type::uint_type
, "access" , ir_var_function_in
);
487 sig_params
.push_tail(access_ref
);
489 ir_function_signature
*sig
=
490 new(mem_ctx
) ir_function_signature(type
, shader_storage_buffer_object
);
492 sig
->replace_parameters(&sig_params
);
493 sig
->is_intrinsic
= true;
495 ir_function
*f
= new(mem_ctx
) ir_function("__intrinsic_load_ssbo");
496 f
->add_signature(sig
);
498 ir_variable
*result
= new(mem_ctx
)
499 ir_variable(type
, "ssbo_load_result", ir_var_temporary
);
500 base_ir
->insert_before(result
);
501 ir_dereference_variable
*deref_result
= new(mem_ctx
)
502 ir_dereference_variable(result
);
504 exec_list call_params
;
505 call_params
.push_tail(this->uniform_block
->clone(mem_ctx
, NULL
));
506 call_params
.push_tail(offset
->clone(mem_ctx
, NULL
));
507 call_params
.push_tail(new(mem_ctx
) ir_constant(ssbo_access_params()));
509 return new(mem_ctx
) ir_call(sig
, deref_result
, &call_params
);
513 lower_ubo_reference_visitor::insert_buffer_access(void *mem_ctx
,
514 ir_dereference
*deref
,
515 const glsl_type
*type
,
520 switch (this->buffer_access_type
) {
521 case ubo_load_access
:
522 base_ir
->insert_before(assign(deref
->clone(mem_ctx
, NULL
),
523 ubo_load(mem_ctx
, type
, offset
),
526 case ssbo_load_access
: {
527 ir_call
*load_ssbo
= ssbo_load(mem_ctx
, type
, offset
);
528 base_ir
->insert_before(load_ssbo
);
529 ir_rvalue
*value
= load_ssbo
->return_deref
->as_rvalue()->clone(mem_ctx
, NULL
);
530 ir_assignment
*assignment
=
531 assign(deref
->clone(mem_ctx
, NULL
), value
, mask
);
532 base_ir
->insert_before(assignment
);
535 case ssbo_store_access
:
537 base_ir
->insert_after(ssbo_store(mem_ctx
,
538 swizzle(deref
, channel
, 1),
541 base_ir
->insert_after(ssbo_store(mem_ctx
, deref
, offset
, mask
));
545 unreachable("invalid buffer_access_type in insert_buffer_access");
550 lower_ubo_reference_visitor::write_to_memory(void *mem_ctx
,
551 ir_dereference
*deref
,
553 ir_variable
*write_var
,
556 ir_rvalue
*offset
= NULL
;
557 unsigned const_offset
;
560 enum glsl_interface_packing packing
= var
->get_interface_type_packing();
562 this->buffer_access_type
= ssbo_store_access
;
563 this->variable
= var
;
565 /* Compute the offset to the start if the dereference as well as other
566 * information we need to configure the write
568 setup_for_load_or_store(mem_ctx
, var
, deref
,
569 &offset
, &const_offset
,
570 &row_major
, &matrix_columns
,
574 /* Now emit writes from the temporary to memory */
575 ir_variable
*write_offset
=
576 new(mem_ctx
) ir_variable(glsl_type::uint_type
,
577 "ssbo_store_temp_offset",
580 base_ir
->insert_before(write_offset
);
581 base_ir
->insert_before(assign(write_offset
, offset
));
583 deref
= new(mem_ctx
) ir_dereference_variable(write_var
);
584 emit_access(mem_ctx
, true, deref
, write_offset
, const_offset
,
585 row_major
, matrix_columns
, packing
, write_mask
);
589 lower_ubo_reference_visitor::visit_enter(ir_expression
*ir
)
591 check_ssbo_unsized_array_length_expression(ir
);
592 return rvalue_visit(ir
);
596 lower_ubo_reference_visitor::calculate_ssbo_unsized_array_length(ir_expression
*expr
)
598 if (expr
->operation
!=
599 ir_expression_operation(ir_unop_ssbo_unsized_array_length
))
602 ir_rvalue
*rvalue
= expr
->operands
[0]->as_rvalue();
604 !rvalue
->type
->is_array() || !rvalue
->type
->is_unsized_array())
607 ir_dereference
*deref
= expr
->operands
[0]->as_dereference();
611 ir_variable
*var
= expr
->operands
[0]->variable_referenced();
612 if (!var
|| !var
->is_in_shader_storage_block())
614 return process_ssbo_unsized_array_length(&rvalue
, deref
, var
);
618 lower_ubo_reference_visitor::check_ssbo_unsized_array_length_expression(ir_expression
*ir
)
621 ir_expression_operation(ir_unop_ssbo_unsized_array_length
)) {
622 /* Don't replace this unop if it is found alone. It is going to be
623 * removed by the optimization passes or replaced if it is part of
624 * an ir_assignment or another ir_expression.
629 for (unsigned i
= 0; i
< ir
->get_num_operands(); i
++) {
630 if (ir
->operands
[i
]->ir_type
!= ir_type_expression
)
632 ir_expression
*expr
= (ir_expression
*) ir
->operands
[i
];
633 ir_expression
*temp
= calculate_ssbo_unsized_array_length(expr
);
638 ir
->operands
[i
] = temp
;
643 lower_ubo_reference_visitor::check_ssbo_unsized_array_length_assignment(ir_assignment
*ir
)
645 if (!ir
->rhs
|| ir
->rhs
->ir_type
!= ir_type_expression
)
648 ir_expression
*expr
= (ir_expression
*) ir
->rhs
;
649 ir_expression
*temp
= calculate_ssbo_unsized_array_length(expr
);
659 lower_ubo_reference_visitor::emit_ssbo_get_buffer_size(void *mem_ctx
)
661 ir_rvalue
*block_ref
= this->uniform_block
->clone(mem_ctx
, NULL
);
662 return new(mem_ctx
) ir_expression(ir_unop_get_buffer_size
,
668 lower_ubo_reference_visitor::calculate_unsized_array_stride(ir_dereference
*deref
,
669 enum glsl_interface_packing packing
)
671 unsigned array_stride
= 0;
673 switch (deref
->ir_type
) {
674 case ir_type_dereference_variable
:
676 ir_dereference_variable
*deref_var
= (ir_dereference_variable
*)deref
;
677 const struct glsl_type
*unsized_array_type
= NULL
;
678 /* An unsized array can be sized by other lowering passes, so pick
679 * the first field of the array which has the data type of the unsized
682 unsized_array_type
= deref_var
->var
->type
->fields
.array
;
684 /* Whether or not the field is row-major (because it might be a
685 * bvec2 or something) does not affect the array itself. We need
686 * to know whether an array element in its entirety is row-major.
688 const bool array_row_major
=
689 is_dereferenced_thing_row_major(deref_var
);
691 if (packing
== GLSL_INTERFACE_PACKING_STD430
) {
692 array_stride
= unsized_array_type
->std430_array_stride(array_row_major
);
694 array_stride
= unsized_array_type
->std140_size(array_row_major
);
695 array_stride
= glsl_align(array_stride
, 16);
699 case ir_type_dereference_record
:
701 ir_dereference_record
*deref_record
= (ir_dereference_record
*) deref
;
702 ir_dereference
*interface_deref
=
703 deref_record
->record
->as_dereference();
704 assert(interface_deref
!= NULL
);
705 const struct glsl_type
*interface_type
= interface_deref
->type
;
706 unsigned record_length
= interface_type
->length
;
707 /* Unsized array is always the last element of the interface */
708 const struct glsl_type
*unsized_array_type
=
709 interface_type
->fields
.structure
[record_length
- 1].type
->fields
.array
;
711 const bool array_row_major
=
712 is_dereferenced_thing_row_major(deref_record
);
714 if (packing
== GLSL_INTERFACE_PACKING_STD430
) {
715 array_stride
= unsized_array_type
->std430_array_stride(array_row_major
);
717 array_stride
= unsized_array_type
->std140_size(array_row_major
);
718 array_stride
= glsl_align(array_stride
, 16);
723 unreachable("Unsupported dereference type");
729 lower_ubo_reference_visitor::process_ssbo_unsized_array_length(ir_rvalue
**rvalue
,
730 ir_dereference
*deref
,
733 void *mem_ctx
= ralloc_parent(*rvalue
);
735 ir_rvalue
*base_offset
= NULL
;
736 unsigned const_offset
;
739 enum glsl_interface_packing packing
= var
->get_interface_type_packing();
740 int unsized_array_stride
= calculate_unsized_array_stride(deref
, packing
);
742 this->buffer_access_type
= ssbo_unsized_array_length_access
;
743 this->variable
= var
;
745 /* Compute the offset to the start if the dereference as well as other
746 * information we need to calculate the length.
748 setup_for_load_or_store(mem_ctx
, var
, deref
,
749 &base_offset
, &const_offset
,
750 &row_major
, &matrix_columns
,
753 * max((buffer_object_size - offset_of_array) / stride_of_array, 0)
755 ir_expression
*buffer_size
= emit_ssbo_get_buffer_size(mem_ctx
);
757 ir_expression
*offset_of_array
= new(mem_ctx
)
758 ir_expression(ir_binop_add
, base_offset
,
759 new(mem_ctx
) ir_constant(const_offset
));
760 ir_expression
*offset_of_array_int
= new(mem_ctx
)
761 ir_expression(ir_unop_u2i
, offset_of_array
);
763 ir_expression
*sub
= new(mem_ctx
)
764 ir_expression(ir_binop_sub
, buffer_size
, offset_of_array_int
);
765 ir_expression
*div
= new(mem_ctx
)
766 ir_expression(ir_binop_div
, sub
,
767 new(mem_ctx
) ir_constant(unsized_array_stride
));
768 ir_expression
*max
= new(mem_ctx
)
769 ir_expression(ir_binop_max
, div
, new(mem_ctx
) ir_constant(0));
775 lower_ubo_reference_visitor::check_for_ssbo_store(ir_assignment
*ir
)
780 ir_rvalue
*rvalue
= ir
->lhs
->as_rvalue();
784 ir_dereference
*deref
= ir
->lhs
->as_dereference();
788 ir_variable
*var
= ir
->lhs
->variable_referenced();
789 if (!var
|| !var
->is_in_shader_storage_block())
792 /* We have a write to a buffer variable, so declare a temporary and rewrite
793 * the assignment so that the temporary is the LHS.
795 void *mem_ctx
= ralloc_parent(shader
->ir
);
797 const glsl_type
*type
= rvalue
->type
;
798 ir_variable
*write_var
= new(mem_ctx
) ir_variable(type
,
801 base_ir
->insert_before(write_var
);
802 ir
->lhs
= new(mem_ctx
) ir_dereference_variable(write_var
);
804 /* Now we have to write the value assigned to the temporary back to memory */
805 write_to_memory(mem_ctx
, deref
, var
, write_var
, ir
->write_mask
);
810 is_buffer_backed_variable(ir_variable
*var
)
812 return var
->is_in_buffer_block() ||
813 var
->data
.mode
== ir_var_shader_shared
;
817 lower_ubo_reference_visitor::check_for_buffer_array_copy(ir_assignment
*ir
)
819 if (!ir
|| !ir
->lhs
|| !ir
->rhs
)
822 /* LHS and RHS must be arrays
823 * FIXME: arrays of arrays?
825 if (!ir
->lhs
->type
->is_array() || !ir
->rhs
->type
->is_array())
828 /* RHS must be a buffer-backed variable. This is what can cause the problem
829 * since it would lead to a series of loads that need to live until we
830 * see the writes to the LHS.
832 ir_variable
*rhs_var
= ir
->rhs
->variable_referenced();
833 if (!rhs_var
|| !is_buffer_backed_variable(rhs_var
))
836 /* Split the array copy into individual element copies to reduce
839 ir_dereference
*rhs_deref
= ir
->rhs
->as_dereference();
843 ir_dereference
*lhs_deref
= ir
->lhs
->as_dereference();
847 assert(lhs_deref
->type
->length
== rhs_deref
->type
->length
);
848 void *mem_ctx
= ralloc_parent(shader
->ir
);
850 for (unsigned i
= 0; i
< lhs_deref
->type
->length
; i
++) {
851 ir_dereference
*lhs_i
=
852 new(mem_ctx
) ir_dereference_array(lhs_deref
->clone(mem_ctx
, NULL
),
853 new(mem_ctx
) ir_constant(i
));
855 ir_dereference
*rhs_i
=
856 new(mem_ctx
) ir_dereference_array(rhs_deref
->clone(mem_ctx
, NULL
),
857 new(mem_ctx
) ir_constant(i
));
858 ir
->insert_after(assign(lhs_i
, rhs_i
));
867 lower_ubo_reference_visitor::check_for_buffer_struct_copy(ir_assignment
*ir
)
869 if (!ir
|| !ir
->lhs
|| !ir
->rhs
)
872 /* LHS and RHS must be records */
873 if (!ir
->lhs
->type
->is_record() || !ir
->rhs
->type
->is_record())
876 /* RHS must be a buffer-backed variable. This is what can cause the problem
877 * since it would lead to a series of loads that need to live until we
878 * see the writes to the LHS.
880 ir_variable
*rhs_var
= ir
->rhs
->variable_referenced();
881 if (!rhs_var
|| !is_buffer_backed_variable(rhs_var
))
884 /* Split the struct copy into individual element copies to reduce
887 ir_dereference
*rhs_deref
= ir
->rhs
->as_dereference();
891 ir_dereference
*lhs_deref
= ir
->lhs
->as_dereference();
895 assert(lhs_deref
->type
->record_compare(rhs_deref
->type
));
896 void *mem_ctx
= ralloc_parent(shader
->ir
);
898 for (unsigned i
= 0; i
< lhs_deref
->type
->length
; i
++) {
899 const char *field_name
= lhs_deref
->type
->fields
.structure
[i
].name
;
900 ir_dereference
*lhs_field
=
901 new(mem_ctx
) ir_dereference_record(lhs_deref
->clone(mem_ctx
, NULL
),
903 ir_dereference
*rhs_field
=
904 new(mem_ctx
) ir_dereference_record(rhs_deref
->clone(mem_ctx
, NULL
),
906 ir
->insert_after(assign(lhs_field
, rhs_field
));
915 lower_ubo_reference_visitor::visit_enter(ir_assignment
*ir
)
917 /* Array and struct copies could involve large amounts of load/store
918 * operations. To improve register pressure we want to special-case
919 * these and split them into individual element copies.
920 * This way we avoid emitting all the loads for the RHS first and
921 * all the writes for the LHS second and register usage is more
924 if (check_for_buffer_array_copy(ir
))
925 return visit_continue_with_parent
;
927 if (check_for_buffer_struct_copy(ir
))
928 return visit_continue_with_parent
;
930 check_ssbo_unsized_array_length_assignment(ir
);
931 check_for_ssbo_store(ir
);
932 return rvalue_visit(ir
);
935 /* Lowers the intrinsic call to a new internal intrinsic that swaps the
936 * access to the buffer variable in the first parameter by an offset
937 * and block index. This involves creating the new internal intrinsic
938 * (i.e. the new function signature).
941 lower_ubo_reference_visitor::lower_ssbo_atomic_intrinsic(ir_call
*ir
)
943 /* SSBO atomics usually have 2 parameters, the buffer variable and an
944 * integer argument. The exception is CompSwap, that has an additional
947 int param_count
= ir
->actual_parameters
.length();
948 assert(param_count
== 2 || param_count
== 3);
950 /* First argument must be a scalar integer buffer variable */
951 exec_node
*param
= ir
->actual_parameters
.get_head();
952 ir_instruction
*inst
= (ir_instruction
*) param
;
953 assert(inst
->ir_type
== ir_type_dereference_variable
||
954 inst
->ir_type
== ir_type_dereference_array
||
955 inst
->ir_type
== ir_type_dereference_record
||
956 inst
->ir_type
== ir_type_swizzle
);
958 ir_rvalue
*deref
= (ir_rvalue
*) inst
;
959 assert(deref
->type
->is_scalar() && deref
->type
->is_integer());
961 ir_variable
*var
= deref
->variable_referenced();
964 /* Compute the offset to the start if the dereference and the
967 void *mem_ctx
= ralloc_parent(shader
->ir
);
969 ir_rvalue
*offset
= NULL
;
970 unsigned const_offset
;
973 enum glsl_interface_packing packing
= var
->get_interface_type_packing();
975 this->buffer_access_type
= ssbo_atomic_access
;
976 this->variable
= var
;
978 setup_for_load_or_store(mem_ctx
, var
, deref
,
979 &offset
, &const_offset
,
980 &row_major
, &matrix_columns
,
984 assert(matrix_columns
== 1);
986 ir_rvalue
*deref_offset
=
987 add(offset
, new(mem_ctx
) ir_constant(const_offset
));
988 ir_rvalue
*block_index
= this->uniform_block
->clone(mem_ctx
, NULL
);
990 /* Create the new internal function signature that will take a block
991 * index and offset instead of a buffer variable
993 exec_list sig_params
;
994 ir_variable
*sig_param
= new(mem_ctx
)
995 ir_variable(glsl_type::uint_type
, "block_ref" , ir_var_function_in
);
996 sig_params
.push_tail(sig_param
);
998 sig_param
= new(mem_ctx
)
999 ir_variable(glsl_type::uint_type
, "offset" , ir_var_function_in
);
1000 sig_params
.push_tail(sig_param
);
1002 const glsl_type
*type
= deref
->type
->base_type
== GLSL_TYPE_INT
?
1003 glsl_type::int_type
: glsl_type::uint_type
;
1004 sig_param
= new(mem_ctx
)
1005 ir_variable(type
, "data1", ir_var_function_in
);
1006 sig_params
.push_tail(sig_param
);
1008 if (param_count
== 3) {
1009 sig_param
= new(mem_ctx
)
1010 ir_variable(type
, "data2", ir_var_function_in
);
1011 sig_params
.push_tail(sig_param
);
1014 ir_function_signature
*sig
=
1015 new(mem_ctx
) ir_function_signature(deref
->type
,
1016 shader_storage_buffer_object
);
1018 sig
->replace_parameters(&sig_params
);
1019 sig
->is_intrinsic
= true;
1022 sprintf(func_name
, "%s_ssbo", ir
->callee_name());
1023 ir_function
*f
= new(mem_ctx
) ir_function(func_name
);
1024 f
->add_signature(sig
);
1026 /* Now, create the call to the internal intrinsic */
1027 exec_list call_params
;
1028 call_params
.push_tail(block_index
);
1029 call_params
.push_tail(deref_offset
);
1030 param
= ir
->actual_parameters
.get_head()->get_next();
1031 ir_rvalue
*param_as_rvalue
= ((ir_instruction
*) param
)->as_rvalue();
1032 call_params
.push_tail(param_as_rvalue
->clone(mem_ctx
, NULL
));
1033 if (param_count
== 3) {
1034 param
= param
->get_next();
1035 param_as_rvalue
= ((ir_instruction
*) param
)->as_rvalue();
1036 call_params
.push_tail(param_as_rvalue
->clone(mem_ctx
, NULL
));
1038 ir_dereference_variable
*return_deref
=
1039 ir
->return_deref
->clone(mem_ctx
, NULL
);
1040 return new(mem_ctx
) ir_call(sig
, return_deref
, &call_params
);
1044 lower_ubo_reference_visitor::check_for_ssbo_atomic_intrinsic(ir_call
*ir
)
1046 exec_list
& params
= ir
->actual_parameters
;
1048 if (params
.length() < 2 || params
.length() > 3)
1052 ((ir_instruction
*) params
.get_head())->as_rvalue();
1056 ir_variable
*var
= rvalue
->variable_referenced();
1057 if (!var
|| !var
->is_in_shader_storage_block())
1060 const char *callee
= ir
->callee_name();
1061 if (!strcmp("__intrinsic_atomic_add", callee
) ||
1062 !strcmp("__intrinsic_atomic_min", callee
) ||
1063 !strcmp("__intrinsic_atomic_max", callee
) ||
1064 !strcmp("__intrinsic_atomic_and", callee
) ||
1065 !strcmp("__intrinsic_atomic_or", callee
) ||
1066 !strcmp("__intrinsic_atomic_xor", callee
) ||
1067 !strcmp("__intrinsic_atomic_exchange", callee
) ||
1068 !strcmp("__intrinsic_atomic_comp_swap", callee
)) {
1069 return lower_ssbo_atomic_intrinsic(ir
);
1077 lower_ubo_reference_visitor::visit_enter(ir_call
*ir
)
1079 ir_call
*new_ir
= check_for_ssbo_atomic_intrinsic(ir
);
1082 base_ir
->replace_with(new_ir
);
1083 return visit_continue_with_parent
;
1086 return rvalue_visit(ir
);
1090 } /* unnamed namespace */
1093 lower_ubo_reference(struct gl_linked_shader
*shader
, bool clamp_block_indices
)
1095 lower_ubo_reference_visitor
v(shader
, clamp_block_indices
);
1097 /* Loop over the instructions lowering references, because we take
1098 * a deref of a UBO array using a UBO dereference as the index will
1099 * produce a collection of instructions all of which have cloned
1100 * UBO dereferences for that array index.
1104 visit_list_elements(&v
, shader
->ir
);
1105 } while (v
.progress
);