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
);
107 ir_visitor_status
visit_enter(ir_texture
*ir
);
109 struct gl_linked_shader
*shader
;
110 bool clamp_block_indices
;
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
->Program
->info
.num_ssbos
;
294 blocks
= shader
->Program
->sh
.ShaderStorageBlocks
;
296 num_blocks
= shader
->Program
->info
.num_ubos
;
297 blocks
= shader
->Program
->sh
.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 if (var
->is_interface_instance()) {
314 *const_offset
= blocks
[i
]->Uniforms
[var
->data
.location
].Offset
;
321 assert(this->uniform_block
);
323 this->struct_field
= NULL
;
324 setup_buffer_access(mem_ctx
, deref
, offset
, const_offset
, row_major
,
325 matrix_columns
, &this->struct_field
, packing
);
329 lower_ubo_reference_visitor::handle_rvalue(ir_rvalue
**rvalue
)
334 ir_dereference
*deref
= (*rvalue
)->as_dereference();
338 ir_variable
*var
= deref
->variable_referenced();
339 if (!var
|| !var
->is_in_buffer_block())
342 void *mem_ctx
= ralloc_parent(shader
->ir
);
344 ir_rvalue
*offset
= NULL
;
345 unsigned const_offset
;
348 enum glsl_interface_packing packing
= var
->get_interface_type_packing();
350 this->buffer_access_type
=
351 var
->is_in_shader_storage_block() ?
352 ssbo_load_access
: ubo_load_access
;
353 this->variable
= var
;
355 /* Compute the offset to the start if the dereference as well as other
356 * information we need to configure the write
358 setup_for_load_or_store(mem_ctx
, var
, deref
,
359 &offset
, &const_offset
,
360 &row_major
, &matrix_columns
,
364 /* Now that we've calculated the offset to the start of the
365 * dereference, walk over the type and emit loads into a temporary.
367 const glsl_type
*type
= (*rvalue
)->type
;
368 ir_variable
*load_var
= new(mem_ctx
) ir_variable(type
,
371 base_ir
->insert_before(load_var
);
373 ir_variable
*load_offset
= new(mem_ctx
) ir_variable(glsl_type::uint_type
,
374 "ubo_load_temp_offset",
376 base_ir
->insert_before(load_offset
);
377 base_ir
->insert_before(assign(load_offset
, offset
));
379 deref
= new(mem_ctx
) ir_dereference_variable(load_var
);
380 emit_access(mem_ctx
, false, deref
, load_offset
, const_offset
,
381 row_major
, matrix_columns
, packing
, 0);
388 lower_ubo_reference_visitor::ubo_load(void *mem_ctx
,
389 const glsl_type
*type
,
392 ir_rvalue
*block_ref
= this->uniform_block
->clone(mem_ctx
, NULL
);
394 ir_expression(ir_binop_ubo_load
,
402 shader_storage_buffer_object(const _mesa_glsl_parse_state
*state
)
404 return state
->has_shader_storage_buffer_objects();
408 lower_ubo_reference_visitor::ssbo_access_params()
412 if (variable
->is_interface_instance()) {
413 assert(struct_field
);
415 return ((struct_field
->memory_coherent
? ACCESS_COHERENT
: 0) |
416 (struct_field
->memory_restrict
? ACCESS_RESTRICT
: 0) |
417 (struct_field
->memory_volatile
? ACCESS_VOLATILE
: 0));
419 return ((variable
->data
.memory_coherent
? ACCESS_COHERENT
: 0) |
420 (variable
->data
.memory_restrict
? ACCESS_RESTRICT
: 0) |
421 (variable
->data
.memory_volatile
? ACCESS_VOLATILE
: 0));
426 lower_ubo_reference_visitor::ssbo_store(void *mem_ctx
,
431 exec_list sig_params
;
433 ir_variable
*block_ref
= new(mem_ctx
)
434 ir_variable(glsl_type::uint_type
, "block_ref" , ir_var_function_in
);
435 sig_params
.push_tail(block_ref
);
437 ir_variable
*offset_ref
= new(mem_ctx
)
438 ir_variable(glsl_type::uint_type
, "offset" , ir_var_function_in
);
439 sig_params
.push_tail(offset_ref
);
441 ir_variable
*val_ref
= new(mem_ctx
)
442 ir_variable(deref
->type
, "value" , ir_var_function_in
);
443 sig_params
.push_tail(val_ref
);
445 ir_variable
*writemask_ref
= new(mem_ctx
)
446 ir_variable(glsl_type::uint_type
, "write_mask" , ir_var_function_in
);
447 sig_params
.push_tail(writemask_ref
);
449 ir_variable
*access_ref
= new(mem_ctx
)
450 ir_variable(glsl_type::uint_type
, "access" , ir_var_function_in
);
451 sig_params
.push_tail(access_ref
);
453 ir_function_signature
*sig
= new(mem_ctx
)
454 ir_function_signature(glsl_type::void_type
, shader_storage_buffer_object
);
456 sig
->replace_parameters(&sig_params
);
457 sig
->intrinsic_id
= ir_intrinsic_ssbo_store
;
459 ir_function
*f
= new(mem_ctx
) ir_function("__intrinsic_store_ssbo");
460 f
->add_signature(sig
);
462 exec_list call_params
;
463 call_params
.push_tail(this->uniform_block
->clone(mem_ctx
, NULL
));
464 call_params
.push_tail(offset
->clone(mem_ctx
, NULL
));
465 call_params
.push_tail(deref
->clone(mem_ctx
, NULL
));
466 call_params
.push_tail(new(mem_ctx
) ir_constant(write_mask
));
467 call_params
.push_tail(new(mem_ctx
) ir_constant(ssbo_access_params()));
468 return new(mem_ctx
) ir_call(sig
, NULL
, &call_params
);
472 lower_ubo_reference_visitor::ssbo_load(void *mem_ctx
,
473 const struct glsl_type
*type
,
476 exec_list sig_params
;
478 ir_variable
*block_ref
= new(mem_ctx
)
479 ir_variable(glsl_type::uint_type
, "block_ref" , ir_var_function_in
);
480 sig_params
.push_tail(block_ref
);
482 ir_variable
*offset_ref
= new(mem_ctx
)
483 ir_variable(glsl_type::uint_type
, "offset_ref" , ir_var_function_in
);
484 sig_params
.push_tail(offset_ref
);
486 ir_variable
*access_ref
= new(mem_ctx
)
487 ir_variable(glsl_type::uint_type
, "access" , ir_var_function_in
);
488 sig_params
.push_tail(access_ref
);
490 ir_function_signature
*sig
=
491 new(mem_ctx
) ir_function_signature(type
, shader_storage_buffer_object
);
493 sig
->replace_parameters(&sig_params
);
494 sig
->intrinsic_id
= ir_intrinsic_ssbo_load
;
496 ir_function
*f
= new(mem_ctx
) ir_function("__intrinsic_load_ssbo");
497 f
->add_signature(sig
);
499 ir_variable
*result
= new(mem_ctx
)
500 ir_variable(type
, "ssbo_load_result", ir_var_temporary
);
501 base_ir
->insert_before(result
);
502 ir_dereference_variable
*deref_result
= new(mem_ctx
)
503 ir_dereference_variable(result
);
505 exec_list call_params
;
506 call_params
.push_tail(this->uniform_block
->clone(mem_ctx
, NULL
));
507 call_params
.push_tail(offset
->clone(mem_ctx
, NULL
));
508 call_params
.push_tail(new(mem_ctx
) ir_constant(ssbo_access_params()));
510 return new(mem_ctx
) ir_call(sig
, deref_result
, &call_params
);
514 lower_ubo_reference_visitor::insert_buffer_access(void *mem_ctx
,
515 ir_dereference
*deref
,
516 const glsl_type
*type
,
521 switch (this->buffer_access_type
) {
522 case ubo_load_access
:
523 base_ir
->insert_before(assign(deref
->clone(mem_ctx
, NULL
),
524 ubo_load(mem_ctx
, type
, offset
),
527 case ssbo_load_access
: {
528 ir_call
*load_ssbo
= ssbo_load(mem_ctx
, type
, offset
);
529 base_ir
->insert_before(load_ssbo
);
530 ir_rvalue
*value
= load_ssbo
->return_deref
->as_rvalue()->clone(mem_ctx
, NULL
);
531 ir_assignment
*assignment
=
532 assign(deref
->clone(mem_ctx
, NULL
), value
, mask
);
533 base_ir
->insert_before(assignment
);
536 case ssbo_store_access
:
538 base_ir
->insert_after(ssbo_store(mem_ctx
,
539 swizzle(deref
, channel
, 1),
542 base_ir
->insert_after(ssbo_store(mem_ctx
, deref
, offset
, mask
));
546 unreachable("invalid buffer_access_type in insert_buffer_access");
551 lower_ubo_reference_visitor::write_to_memory(void *mem_ctx
,
552 ir_dereference
*deref
,
554 ir_variable
*write_var
,
557 ir_rvalue
*offset
= NULL
;
558 unsigned const_offset
;
561 enum glsl_interface_packing packing
= var
->get_interface_type_packing();
563 this->buffer_access_type
= ssbo_store_access
;
564 this->variable
= var
;
566 /* Compute the offset to the start if the dereference as well as other
567 * information we need to configure the write
569 setup_for_load_or_store(mem_ctx
, var
, deref
,
570 &offset
, &const_offset
,
571 &row_major
, &matrix_columns
,
575 /* Now emit writes from the temporary to memory */
576 ir_variable
*write_offset
=
577 new(mem_ctx
) ir_variable(glsl_type::uint_type
,
578 "ssbo_store_temp_offset",
581 base_ir
->insert_before(write_offset
);
582 base_ir
->insert_before(assign(write_offset
, offset
));
584 deref
= new(mem_ctx
) ir_dereference_variable(write_var
);
585 emit_access(mem_ctx
, true, deref
, write_offset
, const_offset
,
586 row_major
, matrix_columns
, packing
, write_mask
);
590 lower_ubo_reference_visitor::visit_enter(ir_expression
*ir
)
592 check_ssbo_unsized_array_length_expression(ir
);
593 return rvalue_visit(ir
);
597 lower_ubo_reference_visitor::calculate_ssbo_unsized_array_length(ir_expression
*expr
)
599 if (expr
->operation
!=
600 ir_expression_operation(ir_unop_ssbo_unsized_array_length
))
603 ir_rvalue
*rvalue
= expr
->operands
[0]->as_rvalue();
605 !rvalue
->type
->is_array() || !rvalue
->type
->is_unsized_array())
608 ir_dereference
*deref
= expr
->operands
[0]->as_dereference();
612 ir_variable
*var
= expr
->operands
[0]->variable_referenced();
613 if (!var
|| !var
->is_in_shader_storage_block())
615 return process_ssbo_unsized_array_length(&rvalue
, deref
, var
);
619 lower_ubo_reference_visitor::check_ssbo_unsized_array_length_expression(ir_expression
*ir
)
622 ir_expression_operation(ir_unop_ssbo_unsized_array_length
)) {
623 /* Don't replace this unop if it is found alone. It is going to be
624 * removed by the optimization passes or replaced if it is part of
625 * an ir_assignment or another ir_expression.
630 for (unsigned i
= 0; i
< ir
->get_num_operands(); i
++) {
631 if (ir
->operands
[i
]->ir_type
!= ir_type_expression
)
633 ir_expression
*expr
= (ir_expression
*) ir
->operands
[i
];
634 ir_expression
*temp
= calculate_ssbo_unsized_array_length(expr
);
639 ir
->operands
[i
] = temp
;
644 lower_ubo_reference_visitor::check_ssbo_unsized_array_length_assignment(ir_assignment
*ir
)
646 if (!ir
->rhs
|| ir
->rhs
->ir_type
!= ir_type_expression
)
649 ir_expression
*expr
= (ir_expression
*) ir
->rhs
;
650 ir_expression
*temp
= calculate_ssbo_unsized_array_length(expr
);
660 lower_ubo_reference_visitor::emit_ssbo_get_buffer_size(void *mem_ctx
)
662 ir_rvalue
*block_ref
= this->uniform_block
->clone(mem_ctx
, NULL
);
663 return new(mem_ctx
) ir_expression(ir_unop_get_buffer_size
,
669 lower_ubo_reference_visitor::calculate_unsized_array_stride(ir_dereference
*deref
,
670 enum glsl_interface_packing packing
)
672 unsigned array_stride
= 0;
674 switch (deref
->ir_type
) {
675 case ir_type_dereference_variable
:
677 ir_dereference_variable
*deref_var
= (ir_dereference_variable
*)deref
;
678 const struct glsl_type
*unsized_array_type
= NULL
;
679 /* An unsized array can be sized by other lowering passes, so pick
680 * the first field of the array which has the data type of the unsized
683 unsized_array_type
= deref_var
->var
->type
->fields
.array
;
685 /* Whether or not the field is row-major (because it might be a
686 * bvec2 or something) does not affect the array itself. We need
687 * to know whether an array element in its entirety is row-major.
689 const bool array_row_major
=
690 is_dereferenced_thing_row_major(deref_var
);
692 if (packing
== GLSL_INTERFACE_PACKING_STD430
) {
693 array_stride
= unsized_array_type
->std430_array_stride(array_row_major
);
695 array_stride
= unsized_array_type
->std140_size(array_row_major
);
696 array_stride
= glsl_align(array_stride
, 16);
700 case ir_type_dereference_record
:
702 ir_dereference_record
*deref_record
= (ir_dereference_record
*) deref
;
703 ir_dereference
*interface_deref
=
704 deref_record
->record
->as_dereference();
705 assert(interface_deref
!= NULL
);
706 const struct glsl_type
*interface_type
= interface_deref
->type
;
707 unsigned record_length
= interface_type
->length
;
708 /* Unsized array is always the last element of the interface */
709 const struct glsl_type
*unsized_array_type
=
710 interface_type
->fields
.structure
[record_length
- 1].type
->fields
.array
;
712 const bool array_row_major
=
713 is_dereferenced_thing_row_major(deref_record
);
715 if (packing
== GLSL_INTERFACE_PACKING_STD430
) {
716 array_stride
= unsized_array_type
->std430_array_stride(array_row_major
);
718 array_stride
= unsized_array_type
->std140_size(array_row_major
);
719 array_stride
= glsl_align(array_stride
, 16);
724 unreachable("Unsupported dereference type");
730 lower_ubo_reference_visitor::process_ssbo_unsized_array_length(ir_rvalue
**rvalue
,
731 ir_dereference
*deref
,
734 void *mem_ctx
= ralloc_parent(*rvalue
);
736 ir_rvalue
*base_offset
= NULL
;
737 unsigned const_offset
;
740 enum glsl_interface_packing packing
= var
->get_interface_type_packing();
741 int unsized_array_stride
= calculate_unsized_array_stride(deref
, packing
);
743 this->buffer_access_type
= ssbo_unsized_array_length_access
;
744 this->variable
= var
;
746 /* Compute the offset to the start if the dereference as well as other
747 * information we need to calculate the length.
749 setup_for_load_or_store(mem_ctx
, var
, deref
,
750 &base_offset
, &const_offset
,
751 &row_major
, &matrix_columns
,
754 * max((buffer_object_size - offset_of_array) / stride_of_array, 0)
756 ir_expression
*buffer_size
= emit_ssbo_get_buffer_size(mem_ctx
);
758 ir_expression
*offset_of_array
= new(mem_ctx
)
759 ir_expression(ir_binop_add
, base_offset
,
760 new(mem_ctx
) ir_constant(const_offset
));
761 ir_expression
*offset_of_array_int
= new(mem_ctx
)
762 ir_expression(ir_unop_u2i
, offset_of_array
);
764 ir_expression
*sub
= new(mem_ctx
)
765 ir_expression(ir_binop_sub
, buffer_size
, offset_of_array_int
);
766 ir_expression
*div
= new(mem_ctx
)
767 ir_expression(ir_binop_div
, sub
,
768 new(mem_ctx
) ir_constant(unsized_array_stride
));
769 ir_expression
*max
= new(mem_ctx
)
770 ir_expression(ir_binop_max
, div
, new(mem_ctx
) ir_constant(0));
776 lower_ubo_reference_visitor::check_for_ssbo_store(ir_assignment
*ir
)
781 ir_rvalue
*rvalue
= ir
->lhs
->as_rvalue();
785 ir_dereference
*deref
= ir
->lhs
->as_dereference();
789 ir_variable
*var
= ir
->lhs
->variable_referenced();
790 if (!var
|| !var
->is_in_shader_storage_block())
793 /* We have a write to a buffer variable, so declare a temporary and rewrite
794 * the assignment so that the temporary is the LHS.
796 void *mem_ctx
= ralloc_parent(shader
->ir
);
798 const glsl_type
*type
= rvalue
->type
;
799 ir_variable
*write_var
= new(mem_ctx
) ir_variable(type
,
802 base_ir
->insert_before(write_var
);
803 ir
->lhs
= new(mem_ctx
) ir_dereference_variable(write_var
);
805 /* Now we have to write the value assigned to the temporary back to memory */
806 write_to_memory(mem_ctx
, deref
, var
, write_var
, ir
->write_mask
);
811 is_buffer_backed_variable(ir_variable
*var
)
813 return var
->is_in_buffer_block() ||
814 var
->data
.mode
== ir_var_shader_shared
;
818 lower_ubo_reference_visitor::check_for_buffer_array_copy(ir_assignment
*ir
)
820 if (!ir
|| !ir
->lhs
|| !ir
->rhs
)
823 /* LHS and RHS must be arrays
824 * FIXME: arrays of arrays?
826 if (!ir
->lhs
->type
->is_array() || !ir
->rhs
->type
->is_array())
829 /* RHS must be a buffer-backed variable. This is what can cause the problem
830 * since it would lead to a series of loads that need to live until we
831 * see the writes to the LHS.
833 ir_variable
*rhs_var
= ir
->rhs
->variable_referenced();
834 if (!rhs_var
|| !is_buffer_backed_variable(rhs_var
))
837 /* Split the array copy into individual element copies to reduce
840 ir_dereference
*rhs_deref
= ir
->rhs
->as_dereference();
844 ir_dereference
*lhs_deref
= ir
->lhs
->as_dereference();
848 assert(lhs_deref
->type
->length
== rhs_deref
->type
->length
);
849 void *mem_ctx
= ralloc_parent(shader
->ir
);
851 for (unsigned i
= 0; i
< lhs_deref
->type
->length
; i
++) {
852 ir_dereference
*lhs_i
=
853 new(mem_ctx
) ir_dereference_array(lhs_deref
->clone(mem_ctx
, NULL
),
854 new(mem_ctx
) ir_constant(i
));
856 ir_dereference
*rhs_i
=
857 new(mem_ctx
) ir_dereference_array(rhs_deref
->clone(mem_ctx
, NULL
),
858 new(mem_ctx
) ir_constant(i
));
859 ir
->insert_after(assign(lhs_i
, rhs_i
));
868 lower_ubo_reference_visitor::check_for_buffer_struct_copy(ir_assignment
*ir
)
870 if (!ir
|| !ir
->lhs
|| !ir
->rhs
)
873 /* LHS and RHS must be records */
874 if (!ir
->lhs
->type
->is_record() || !ir
->rhs
->type
->is_record())
877 /* RHS must be a buffer-backed variable. This is what can cause the problem
878 * since it would lead to a series of loads that need to live until we
879 * see the writes to the LHS.
881 ir_variable
*rhs_var
= ir
->rhs
->variable_referenced();
882 if (!rhs_var
|| !is_buffer_backed_variable(rhs_var
))
885 /* Split the struct copy into individual element copies to reduce
888 ir_dereference
*rhs_deref
= ir
->rhs
->as_dereference();
892 ir_dereference
*lhs_deref
= ir
->lhs
->as_dereference();
896 assert(lhs_deref
->type
== rhs_deref
->type
);
897 void *mem_ctx
= ralloc_parent(shader
->ir
);
899 for (unsigned i
= 0; i
< lhs_deref
->type
->length
; i
++) {
900 const char *field_name
= lhs_deref
->type
->fields
.structure
[i
].name
;
901 ir_dereference
*lhs_field
=
902 new(mem_ctx
) ir_dereference_record(lhs_deref
->clone(mem_ctx
, NULL
),
904 ir_dereference
*rhs_field
=
905 new(mem_ctx
) ir_dereference_record(rhs_deref
->clone(mem_ctx
, NULL
),
907 ir
->insert_after(assign(lhs_field
, rhs_field
));
916 lower_ubo_reference_visitor::visit_enter(ir_assignment
*ir
)
918 /* Array and struct copies could involve large amounts of load/store
919 * operations. To improve register pressure we want to special-case
920 * these and split them into individual element copies.
921 * This way we avoid emitting all the loads for the RHS first and
922 * all the writes for the LHS second and register usage is more
925 if (check_for_buffer_array_copy(ir
))
926 return visit_continue_with_parent
;
928 if (check_for_buffer_struct_copy(ir
))
929 return visit_continue_with_parent
;
931 check_ssbo_unsized_array_length_assignment(ir
);
932 check_for_ssbo_store(ir
);
933 return rvalue_visit(ir
);
936 /* Lowers the intrinsic call to a new internal intrinsic that swaps the
937 * access to the buffer variable in the first parameter by an offset
938 * and block index. This involves creating the new internal intrinsic
939 * (i.e. the new function signature).
942 lower_ubo_reference_visitor::lower_ssbo_atomic_intrinsic(ir_call
*ir
)
944 /* SSBO atomics usually have 2 parameters, the buffer variable and an
945 * integer argument. The exception is CompSwap, that has an additional
948 int param_count
= ir
->actual_parameters
.length();
949 assert(param_count
== 2 || param_count
== 3);
951 /* First argument must be a scalar integer buffer variable */
952 exec_node
*param
= ir
->actual_parameters
.get_head();
953 ir_instruction
*inst
= (ir_instruction
*) param
;
954 assert(inst
->ir_type
== ir_type_dereference_variable
||
955 inst
->ir_type
== ir_type_dereference_array
||
956 inst
->ir_type
== ir_type_dereference_record
||
957 inst
->ir_type
== ir_type_swizzle
);
959 ir_rvalue
*deref
= (ir_rvalue
*) inst
;
960 assert(deref
->type
->is_scalar() && deref
->type
->is_integer());
962 ir_variable
*var
= deref
->variable_referenced();
965 /* Compute the offset to the start if the dereference and the
968 void *mem_ctx
= ralloc_parent(shader
->ir
);
970 ir_rvalue
*offset
= NULL
;
971 unsigned const_offset
;
974 enum glsl_interface_packing packing
= var
->get_interface_type_packing();
976 this->buffer_access_type
= ssbo_atomic_access
;
977 this->variable
= var
;
979 setup_for_load_or_store(mem_ctx
, var
, deref
,
980 &offset
, &const_offset
,
981 &row_major
, &matrix_columns
,
985 assert(matrix_columns
== 1);
987 ir_rvalue
*deref_offset
=
988 add(offset
, new(mem_ctx
) ir_constant(const_offset
));
989 ir_rvalue
*block_index
= this->uniform_block
->clone(mem_ctx
, NULL
);
991 /* Create the new internal function signature that will take a block
992 * index and offset instead of a buffer variable
994 exec_list sig_params
;
995 ir_variable
*sig_param
= new(mem_ctx
)
996 ir_variable(glsl_type::uint_type
, "block_ref" , ir_var_function_in
);
997 sig_params
.push_tail(sig_param
);
999 sig_param
= new(mem_ctx
)
1000 ir_variable(glsl_type::uint_type
, "offset" , ir_var_function_in
);
1001 sig_params
.push_tail(sig_param
);
1003 const glsl_type
*type
= deref
->type
->base_type
== GLSL_TYPE_INT
?
1004 glsl_type::int_type
: glsl_type::uint_type
;
1005 sig_param
= new(mem_ctx
)
1006 ir_variable(type
, "data1", ir_var_function_in
);
1007 sig_params
.push_tail(sig_param
);
1009 if (param_count
== 3) {
1010 sig_param
= new(mem_ctx
)
1011 ir_variable(type
, "data2", ir_var_function_in
);
1012 sig_params
.push_tail(sig_param
);
1015 ir_function_signature
*sig
=
1016 new(mem_ctx
) ir_function_signature(deref
->type
,
1017 shader_storage_buffer_object
);
1019 sig
->replace_parameters(&sig_params
);
1021 assert(ir
->callee
->intrinsic_id
>= ir_intrinsic_generic_load
);
1022 assert(ir
->callee
->intrinsic_id
<= ir_intrinsic_generic_atomic_comp_swap
);
1023 sig
->intrinsic_id
= MAP_INTRINSIC_TO_TYPE(ir
->callee
->intrinsic_id
, ssbo
);
1026 sprintf(func_name
, "%s_ssbo", ir
->callee_name());
1027 ir_function
*f
= new(mem_ctx
) ir_function(func_name
);
1028 f
->add_signature(sig
);
1030 /* Now, create the call to the internal intrinsic */
1031 exec_list call_params
;
1032 call_params
.push_tail(block_index
);
1033 call_params
.push_tail(deref_offset
);
1034 param
= ir
->actual_parameters
.get_head()->get_next();
1035 ir_rvalue
*param_as_rvalue
= ((ir_instruction
*) param
)->as_rvalue();
1036 call_params
.push_tail(param_as_rvalue
->clone(mem_ctx
, NULL
));
1037 if (param_count
== 3) {
1038 param
= param
->get_next();
1039 param_as_rvalue
= ((ir_instruction
*) param
)->as_rvalue();
1040 call_params
.push_tail(param_as_rvalue
->clone(mem_ctx
, NULL
));
1042 ir_dereference_variable
*return_deref
=
1043 ir
->return_deref
->clone(mem_ctx
, NULL
);
1044 return new(mem_ctx
) ir_call(sig
, return_deref
, &call_params
);
1048 lower_ubo_reference_visitor::check_for_ssbo_atomic_intrinsic(ir_call
*ir
)
1050 exec_list
& params
= ir
->actual_parameters
;
1052 if (params
.length() < 2 || params
.length() > 3)
1056 ((ir_instruction
*) params
.get_head())->as_rvalue();
1060 ir_variable
*var
= rvalue
->variable_referenced();
1061 if (!var
|| !var
->is_in_shader_storage_block())
1064 const enum ir_intrinsic_id id
= ir
->callee
->intrinsic_id
;
1065 if (id
== ir_intrinsic_generic_atomic_add
||
1066 id
== ir_intrinsic_generic_atomic_min
||
1067 id
== ir_intrinsic_generic_atomic_max
||
1068 id
== ir_intrinsic_generic_atomic_and
||
1069 id
== ir_intrinsic_generic_atomic_or
||
1070 id
== ir_intrinsic_generic_atomic_xor
||
1071 id
== ir_intrinsic_generic_atomic_exchange
||
1072 id
== ir_intrinsic_generic_atomic_comp_swap
) {
1073 return lower_ssbo_atomic_intrinsic(ir
);
1081 lower_ubo_reference_visitor::visit_enter(ir_call
*ir
)
1083 ir_call
*new_ir
= check_for_ssbo_atomic_intrinsic(ir
);
1086 base_ir
->replace_with(new_ir
);
1087 return visit_continue_with_parent
;
1090 return rvalue_visit(ir
);
1095 lower_ubo_reference_visitor::visit_enter(ir_texture
*ir
)
1097 ir_dereference
*sampler
= ir
->sampler
;
1099 if (sampler
->ir_type
== ir_type_dereference_record
) {
1100 handle_rvalue((ir_rvalue
**)&ir
->sampler
);
1101 return visit_continue_with_parent
;
1104 return rvalue_visit(ir
);
1108 } /* unnamed namespace */
1111 lower_ubo_reference(struct gl_linked_shader
*shader
, bool clamp_block_indices
)
1113 lower_ubo_reference_visitor
v(shader
, clamp_block_indices
);
1115 /* Loop over the instructions lowering references, because we take
1116 * a deref of a UBO array using a UBO dereference as the index will
1117 * produce a collection of instructions all of which have cloned
1118 * UBO dereferences for that array index.
1122 visit_list_elements(&v
, shader
->ir
);
1123 } while (v
.progress
);