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_function_inlining.cpp
27 * Replaces calls to functions with the body of the function.
31 #include "ir_visitor.h"
32 #include "ir_function_inlining.h"
33 #include "ir_expression_flattening.h"
34 #include "compiler/glsl_types.h"
35 #include "util/hash_table.h"
38 do_variable_replacement(exec_list
*instructions
,
40 ir_dereference
*repl
);
44 class ir_function_inlining_visitor
: public ir_hierarchical_visitor
{
46 ir_function_inlining_visitor()
51 virtual ~ir_function_inlining_visitor()
56 virtual ir_visitor_status
visit_enter(ir_expression
*);
57 virtual ir_visitor_status
visit_enter(ir_call
*);
58 virtual ir_visitor_status
visit_enter(ir_return
*);
59 virtual ir_visitor_status
visit_enter(ir_texture
*);
60 virtual ir_visitor_status
visit_enter(ir_swizzle
*);
65 class ir_save_lvalue_visitor
: public ir_hierarchical_visitor
{
67 virtual ir_visitor_status
visit_enter(ir_dereference_array
*);
70 } /* unnamed namespace */
73 do_function_inlining(exec_list
*instructions
)
75 ir_function_inlining_visitor v
;
83 replace_return_with_assignment(ir_instruction
*ir
, void *data
)
85 void *ctx
= ralloc_parent(ir
);
86 ir_dereference
*orig_deref
= (ir_dereference
*) data
;
87 ir_return
*ret
= ir
->as_return();
91 ir_rvalue
*lhs
= orig_deref
->clone(ctx
, NULL
);
92 ret
->replace_with(new(ctx
) ir_assignment(lhs
, ret
->value
, NULL
));
94 /* un-valued return has to be the last return, or we shouldn't
95 * have reached here. (see can_inline()).
97 assert(ret
->next
->is_tail_sentinel());
103 /* Save the given lvalue before the given instruction.
105 * This is done by adding temporary variables into which the current value
106 * of any array indices are saved, and then modifying the dereference chain
107 * in-place to point to those temporary variables.
109 * The hierarchical visitor is only used to traverse the left-hand-side chain
113 ir_save_lvalue_visitor::visit_enter(ir_dereference_array
*deref
)
115 if (deref
->array_index
->ir_type
!= ir_type_constant
) {
116 void *ctx
= ralloc_parent(deref
);
118 ir_assignment
*assignment
;
120 index
= new(ctx
) ir_variable(deref
->array_index
->type
, "saved_idx", ir_var_temporary
);
121 base_ir
->insert_before(index
);
123 assignment
= new(ctx
) ir_assignment(new(ctx
) ir_dereference_variable(index
),
124 deref
->array_index
, 0);
125 base_ir
->insert_before(assignment
);
127 deref
->array_index
= new(ctx
) ir_dereference_variable(index
);
130 deref
->array
->accept(this);
135 ir_call::generate_inline(ir_instruction
*next_ir
)
137 void *ctx
= ralloc_parent(this);
138 ir_variable
**parameters
;
139 unsigned num_parameters
;
141 struct hash_table
*ht
;
143 ht
= _mesa_hash_table_create(NULL
, _mesa_hash_pointer
, _mesa_key_pointer_equal
);
145 num_parameters
= this->callee
->parameters
.length();
146 parameters
= new ir_variable
*[num_parameters
];
148 /* Generate the declarations for the parameters to our inlined code,
149 * and set up the mapping of real function body variables to ours.
152 foreach_two_lists(formal_node
, &this->callee
->parameters
,
153 actual_node
, &this->actual_parameters
) {
154 ir_variable
*sig_param
= (ir_variable
*) formal_node
;
155 ir_rvalue
*param
= (ir_rvalue
*) actual_node
;
157 /* Generate a new variable for the parameter. */
158 if (sig_param
->type
->contains_opaque()) {
159 /* For opaque types, we want the inlined variable references
160 * referencing the passed in variable, since that will have
161 * the location information, which an assignment of an opaque
162 * variable wouldn't. Fix it up below.
164 parameters
[i
] = NULL
;
166 parameters
[i
] = sig_param
->clone(ctx
, ht
);
167 parameters
[i
]->data
.mode
= ir_var_auto
;
169 /* Remove the read-only decoration because we're going to write
170 * directly to this variable. If the cloned variable is left
171 * read-only and the inlined function is inside a loop, the loop
172 * analysis code will get confused.
174 parameters
[i
]->data
.read_only
= false;
175 next_ir
->insert_before(parameters
[i
]);
178 /* Section 6.1.1 (Function Calling Conventions) of the OpenGL Shading
179 * Language 4.5 spec says:
181 * "All arguments are evaluated at call time, exactly once, in order,
182 * from left to right. [...] Evaluation of an out parameter results
183 * in an l-value that is used to copy out a value when the function
186 * I.e., we have to take temporary copies of any relevant array indices
187 * before the function body is executed.
190 * (a) if an array index expressions refers to a variable that is
191 * modified by the execution of the function body, we use the
192 * original value as intended, and
193 * (b) if an array index expression has side effects, those side effects
194 * are only executed once and at the right time.
197 if (sig_param
->data
.mode
== ir_var_function_in
||
198 sig_param
->data
.mode
== ir_var_const_in
) {
199 ir_assignment
*assign
;
201 assign
= new(ctx
) ir_assignment(new(ctx
) ir_dereference_variable(parameters
[i
]),
203 next_ir
->insert_before(assign
);
205 assert(sig_param
->data
.mode
== ir_var_function_out
||
206 sig_param
->data
.mode
== ir_var_function_inout
);
207 assert(param
->is_lvalue());
209 ir_save_lvalue_visitor v
;
214 if (sig_param
->data
.mode
== ir_var_function_inout
) {
215 ir_assignment
*assign
;
217 assign
= new(ctx
) ir_assignment(new(ctx
) ir_dereference_variable(parameters
[i
]),
218 param
->clone(ctx
, NULL
)->as_rvalue(), NULL
);
219 next_ir
->insert_before(assign
);
227 exec_list new_instructions
;
229 /* Generate the inlined body of the function to a new list */
230 foreach_in_list(ir_instruction
, ir
, &callee
->body
) {
231 ir_instruction
*new_ir
= ir
->clone(ctx
, ht
);
233 new_instructions
.push_tail(new_ir
);
234 visit_tree(new_ir
, replace_return_with_assignment
, this->return_deref
);
237 /* If any opaque types were passed in, replace any deref of the
238 * opaque variable with a deref of the argument.
240 foreach_two_lists(formal_node
, &this->callee
->parameters
,
241 actual_node
, &this->actual_parameters
) {
242 ir_rvalue
*const param
= (ir_rvalue
*) actual_node
;
243 ir_variable
*sig_param
= (ir_variable
*) formal_node
;
245 if (sig_param
->type
->contains_opaque()) {
246 ir_dereference
*deref
= param
->as_dereference();
249 do_variable_replacement(&new_instructions
, sig_param
, deref
);
253 /* Now push those new instructions in. */
254 next_ir
->insert_before(&new_instructions
);
256 /* Copy back the value of any 'out' parameters from the function body
257 * variables to our own.
260 foreach_two_lists(formal_node
, &this->callee
->parameters
,
261 actual_node
, &this->actual_parameters
) {
262 ir_rvalue
*const param
= (ir_rvalue
*) actual_node
;
263 const ir_variable
*const sig_param
= (ir_variable
*) formal_node
;
265 /* Move our param variable into the actual param if it's an 'out' type. */
266 if (parameters
[i
] && (sig_param
->data
.mode
== ir_var_function_out
||
267 sig_param
->data
.mode
== ir_var_function_inout
)) {
268 ir_assignment
*assign
;
270 assign
= new(ctx
) ir_assignment(param
,
271 new(ctx
) ir_dereference_variable(parameters
[i
]),
273 next_ir
->insert_before(assign
);
279 delete [] parameters
;
281 _mesa_hash_table_destroy(ht
, NULL
);
286 ir_function_inlining_visitor::visit_enter(ir_expression
*ir
)
289 return visit_continue_with_parent
;
294 ir_function_inlining_visitor::visit_enter(ir_return
*ir
)
297 return visit_continue_with_parent
;
302 ir_function_inlining_visitor::visit_enter(ir_texture
*ir
)
305 return visit_continue_with_parent
;
310 ir_function_inlining_visitor::visit_enter(ir_swizzle
*ir
)
313 return visit_continue_with_parent
;
318 ir_function_inlining_visitor::visit_enter(ir_call
*ir
)
320 if (can_inline(ir
)) {
321 ir
->generate_inline(ir
);
323 this->progress
= true;
326 return visit_continue
;
331 * Replaces references to the "orig" variable with a clone of "repl."
333 * From the spec, opaque types can appear in the tree as function
334 * (non-out) parameters and as the result of array indexing and
335 * structure field selection. In our builtin implementation, they
336 * also appear in the sampler field of an ir_tex instruction.
339 class ir_variable_replacement_visitor
: public ir_hierarchical_visitor
{
341 ir_variable_replacement_visitor(ir_variable
*orig
, ir_dereference
*repl
)
347 virtual ~ir_variable_replacement_visitor()
351 virtual ir_visitor_status
visit_leave(ir_call
*);
352 virtual ir_visitor_status
visit_leave(ir_dereference_array
*);
353 virtual ir_visitor_status
visit_leave(ir_dereference_record
*);
354 virtual ir_visitor_status
visit_leave(ir_texture
*);
356 void replace_deref(ir_dereference
**deref
);
357 void replace_rvalue(ir_rvalue
**rvalue
);
360 ir_dereference
*repl
;
364 ir_variable_replacement_visitor::replace_deref(ir_dereference
**deref
)
366 ir_dereference_variable
*deref_var
= (*deref
)->as_dereference_variable();
367 if (deref_var
&& deref_var
->var
== this->orig
) {
368 *deref
= this->repl
->clone(ralloc_parent(*deref
), NULL
);
373 ir_variable_replacement_visitor::replace_rvalue(ir_rvalue
**rvalue
)
378 ir_dereference
*deref
= (*rvalue
)->as_dereference();
383 replace_deref(&deref
);
388 ir_variable_replacement_visitor::visit_leave(ir_texture
*ir
)
390 replace_deref(&ir
->sampler
);
392 return visit_continue
;
396 ir_variable_replacement_visitor::visit_leave(ir_dereference_array
*ir
)
398 replace_rvalue(&ir
->array
);
399 return visit_continue
;
403 ir_variable_replacement_visitor::visit_leave(ir_dereference_record
*ir
)
405 replace_rvalue(&ir
->record
);
406 return visit_continue
;
410 ir_variable_replacement_visitor::visit_leave(ir_call
*ir
)
412 foreach_in_list_safe(ir_rvalue
, param
, &ir
->actual_parameters
) {
413 ir_rvalue
*new_param
= param
;
414 replace_rvalue(&new_param
);
416 if (new_param
!= param
) {
417 param
->replace_with(new_param
);
420 return visit_continue
;
424 do_variable_replacement(exec_list
*instructions
,
426 ir_dereference
*repl
)
428 ir_variable_replacement_visitor
v(orig
, repl
);
430 visit_list_elements(&v
, instructions
);