2 * Copyright © 2010 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 opt_array_splitting.cpp
27 * If an array is always dereferenced with a constant index, then
28 * split it apart into its elements, making it more amenable to other
29 * optimization passes.
31 * This skips uniform/varying arrays, which would need careful
32 * handling due to their ir->location fields tying them to the GL API
33 * and other shader stages.
37 #include "ir_visitor.h"
38 #include "ir_rvalue_visitor.h"
39 #include "ir_print_visitor.h"
40 #include "glsl_types.h"
42 static bool debug
= false;
44 namespace opt_array_splitting
{
46 class variable_entry
: public exec_node
49 variable_entry(ir_variable
*var
)
53 this->declaration
= false;
54 this->components
= NULL
;
56 if (var
->type
->is_array())
57 this->size
= var
->type
->length
;
59 this->size
= var
->type
->matrix_columns
;
62 ir_variable
*var
; /* The key: the variable's pointer. */
63 unsigned size
; /* array length or matrix columns */
65 /** Whether this array should be split or not. */
68 /* If the variable had a decl we can work with in the instruction
69 * stream. We can't do splitting on function arguments, which
70 * don't get this variable set.
74 ir_variable
**components
;
76 /** ralloc_parent(this->var) -- the shader's talloc context. */
81 using namespace opt_array_splitting
;
84 * This class does a walk over the tree, coming up with the set of
85 * variables that could be split by looking to see if they are arrays
86 * that are only ever constant-index dereferenced.
88 class ir_array_reference_visitor
: public ir_hierarchical_visitor
{
90 ir_array_reference_visitor(void)
92 this->mem_ctx
= ralloc_context(NULL
);
93 this->variable_list
.make_empty();
96 ~ir_array_reference_visitor(void)
101 bool get_split_list(exec_list
*instructions
, bool linked
);
103 virtual ir_visitor_status
visit(ir_variable
*);
104 virtual ir_visitor_status
visit(ir_dereference_variable
*);
105 virtual ir_visitor_status
visit_enter(ir_dereference_array
*);
106 virtual ir_visitor_status
visit_enter(ir_function_signature
*);
108 variable_entry
*get_variable_entry(ir_variable
*var
);
110 /* List of variable_entry */
111 exec_list variable_list
;
117 ir_array_reference_visitor::get_variable_entry(ir_variable
*var
)
121 if (var
->mode
!= ir_var_auto
&&
122 var
->mode
!= ir_var_temporary
)
125 if (!(var
->type
->is_array() || var
->type
->is_matrix()))
128 /* If the array hasn't been sized yet, we can't split it. After
129 * linking, this should be resolved.
131 if (var
->type
->is_array() && var
->type
->length
== 0)
134 foreach_iter(exec_list_iterator
, iter
, this->variable_list
) {
135 variable_entry
*entry
= (variable_entry
*)iter
.get();
136 if (entry
->var
== var
)
140 variable_entry
*entry
= new(mem_ctx
) variable_entry(var
);
141 this->variable_list
.push_tail(entry
);
147 ir_array_reference_visitor::visit(ir_variable
*ir
)
149 variable_entry
*entry
= this->get_variable_entry(ir
);
152 entry
->declaration
= true;
154 return visit_continue
;
158 ir_array_reference_visitor::visit(ir_dereference_variable
*ir
)
160 variable_entry
*entry
= this->get_variable_entry(ir
->var
);
162 /* If we made it to here without seeing an ir_dereference_array,
163 * then the dereference of this array didn't have a constant index
164 * (see the visit_continue_with_parent below), so we can't split
168 entry
->split
= false;
170 return visit_continue
;
174 ir_array_reference_visitor::visit_enter(ir_dereference_array
*ir
)
176 ir_dereference_variable
*deref
= ir
->array
->as_dereference_variable();
178 return visit_continue
;
180 variable_entry
*entry
= this->get_variable_entry(deref
->var
);
182 /* If the access to the array has a variable index, we wouldn't
183 * know which split variable this dereference should go to.
185 if (entry
&& !ir
->array_index
->as_constant())
186 entry
->split
= false;
188 return visit_continue_with_parent
;
192 ir_array_reference_visitor::visit_enter(ir_function_signature
*ir
)
194 /* We don't have logic for array-splitting function arguments,
195 * so just look at the body instructions and not the parameter
198 visit_list_elements(this, &ir
->body
);
199 return visit_continue_with_parent
;
203 ir_array_reference_visitor::get_split_list(exec_list
*instructions
,
206 visit_list_elements(this, instructions
);
208 /* If the shaders aren't linked yet, we can't mess with global
209 * declarations, which need to be matched by name across shaders.
212 foreach_list(node
, instructions
) {
213 ir_variable
*var
= ((ir_instruction
*)node
)->as_variable();
215 variable_entry
*entry
= get_variable_entry(var
);
222 /* Trim out variables we found that we can't split. */
223 foreach_iter(exec_list_iterator
, iter
, variable_list
) {
224 variable_entry
*entry
= (variable_entry
*)iter
.get();
227 printf("array %s@%p: decl %d, split %d\n",
228 entry
->var
->name
, (void *) entry
->var
, entry
->declaration
,
232 if (!(entry
->declaration
&& entry
->split
)) {
237 return !variable_list
.is_empty();
241 * This class rewrites the dereferences of arrays that have been split
242 * to use the newly created ir_variables for each component.
244 class ir_array_splitting_visitor
: public ir_rvalue_visitor
{
246 ir_array_splitting_visitor(exec_list
*vars
)
248 this->variable_list
= vars
;
251 virtual ~ir_array_splitting_visitor()
255 virtual ir_visitor_status
visit_leave(ir_assignment
*);
257 void split_deref(ir_dereference
**deref
);
258 void handle_rvalue(ir_rvalue
**rvalue
);
259 variable_entry
*get_splitting_entry(ir_variable
*var
);
261 exec_list
*variable_list
;
265 ir_array_splitting_visitor::get_splitting_entry(ir_variable
*var
)
269 foreach_iter(exec_list_iterator
, iter
, *this->variable_list
) {
270 variable_entry
*entry
= (variable_entry
*)iter
.get();
271 if (entry
->var
== var
) {
280 ir_array_splitting_visitor::split_deref(ir_dereference
**deref
)
282 ir_dereference_array
*deref_array
= (*deref
)->as_dereference_array();
286 ir_dereference_variable
*deref_var
= deref_array
->array
->as_dereference_variable();
289 ir_variable
*var
= deref_var
->var
;
291 variable_entry
*entry
= get_splitting_entry(var
);
295 ir_constant
*constant
= deref_array
->array_index
->as_constant();
298 if (constant
->value
.i
[0] < (int)entry
->size
) {
299 *deref
= new(entry
->mem_ctx
)
300 ir_dereference_variable(entry
->components
[constant
->value
.i
[0]]);
302 /* There was a constant array access beyond the end of the
303 * array. This might have happened due to constant folding
304 * after the initial parse. This produces an undefined value,
305 * but shouldn't crash. Just give them an uninitialized
308 ir_variable
*temp
= new(entry
->mem_ctx
) ir_variable(deref_array
->type
,
311 entry
->components
[0]->insert_before(temp
);
312 *deref
= new(entry
->mem_ctx
) ir_dereference_variable(temp
);
317 ir_array_splitting_visitor::handle_rvalue(ir_rvalue
**rvalue
)
322 ir_dereference
*deref
= (*rvalue
)->as_dereference();
332 ir_array_splitting_visitor::visit_leave(ir_assignment
*ir
)
334 /* The normal rvalue visitor skips the LHS of assignments, but we
335 * need to process those just the same.
337 ir_rvalue
*lhs
= ir
->lhs
;
340 ir
->lhs
= lhs
->as_dereference();
342 ir
->lhs
->accept(this);
344 handle_rvalue(&ir
->rhs
);
345 ir
->rhs
->accept(this);
348 handle_rvalue(&ir
->condition
);
349 ir
->condition
->accept(this);
352 return visit_continue
;
356 optimize_split_arrays(exec_list
*instructions
, bool linked
)
358 ir_array_reference_visitor refs
;
359 if (!refs
.get_split_list(instructions
, linked
))
362 void *mem_ctx
= ralloc_context(NULL
);
364 /* Replace the decls of the arrays to be split with their split
367 foreach_iter(exec_list_iterator
, iter
, refs
.variable_list
) {
368 variable_entry
*entry
= (variable_entry
*)iter
.get();
369 const struct glsl_type
*type
= entry
->var
->type
;
370 const struct glsl_type
*subtype
;
372 if (type
->is_matrix())
373 subtype
= type
->column_type();
375 subtype
= type
->fields
.array
;
377 entry
->mem_ctx
= ralloc_parent(entry
->var
);
379 entry
->components
= ralloc_array(mem_ctx
,
383 for (unsigned int i
= 0; i
< entry
->size
; i
++) {
384 const char *name
= ralloc_asprintf(mem_ctx
, "%s_%d",
385 entry
->var
->name
, i
);
387 entry
->components
[i
] =
388 new(entry
->mem_ctx
) ir_variable(subtype
, name
, ir_var_temporary
);
389 entry
->var
->insert_before(entry
->components
[i
]);
392 entry
->var
->remove();
395 ir_array_splitting_visitor
split(&refs
.variable_list
);
396 visit_list_elements(&split
, instructions
);
399 _mesa_print_ir(instructions
, NULL
);
401 ralloc_free(mem_ctx
);