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 brw_wm_vector_splitting.cpp
27 * If a vector is only ever referenced by its components, then
28 * split those components out to individual variables so they can be
29 * handled normally by other optimization passes.
31 * This skips vectors in uniforms and varyings, which need to be
32 * accessible as vectors for their access by the GL. Also, vector
33 * results of non-variable-derefs in assignments aren't handled
34 * because to do so we would have to store the vector result to a
35 * temporary in order to unload each channel, and to do so would just
36 * loop us back to where we started. For the 965, this is exactly the
37 * behavior we want for the results of texture lookups, but probably not for
41 #include "main/core.h"
42 #include "brw_context.h"
45 #include "glsl/ir_visitor.h"
46 #include "glsl/ir_rvalue_visitor.h"
47 #include "glsl/glsl_types.h"
49 static bool debug
= false;
51 class variable_entry
: public exec_node
54 variable_entry(ir_variable
*var
)
57 this->whole_vector_access
= 0;
61 ir_variable
*var
; /* The key: the variable's pointer. */
63 /** Number of times the variable is referenced, including assignments. */
64 unsigned whole_vector_access
;
66 ir_variable
*components
[4];
68 /** ralloc_parent(this->var) -- the shader's ralloc context. */
72 class ir_vector_reference_visitor
: public ir_hierarchical_visitor
{
74 ir_vector_reference_visitor(void)
76 this->mem_ctx
= ralloc_context(NULL
);
77 this->variable_list
.make_empty();
80 ~ir_vector_reference_visitor(void)
85 virtual ir_visitor_status
visit(ir_variable
*);
86 virtual ir_visitor_status
visit(ir_dereference_variable
*);
87 virtual ir_visitor_status
visit_enter(ir_swizzle
*);
88 virtual ir_visitor_status
visit_enter(ir_assignment
*);
89 virtual ir_visitor_status
visit_enter(ir_function_signature
*);
91 variable_entry
*get_variable_entry(ir_variable
*var
);
93 /* List of variable_entry */
94 exec_list variable_list
;
100 ir_vector_reference_visitor::get_variable_entry(ir_variable
*var
)
104 if (!var
->type
->is_vector())
107 switch (var
->data
.mode
) {
109 case ir_var_shader_in
:
110 case ir_var_shader_out
:
111 case ir_var_system_value
:
112 case ir_var_function_in
:
113 case ir_var_function_out
:
114 case ir_var_function_inout
:
115 /* Can't split varyings or uniforms. Function in/outs won't get split
120 case ir_var_temporary
:
124 foreach_list(node
, &this->variable_list
) {
125 variable_entry
*entry
= (variable_entry
*)node
;
126 if (entry
->var
== var
)
130 variable_entry
*entry
= new(mem_ctx
) variable_entry(var
);
131 this->variable_list
.push_tail(entry
);
137 ir_vector_reference_visitor::visit(ir_variable
*ir
)
139 /* Make sure splitting looks at splitting this variable */
140 (void)this->get_variable_entry(ir
);
142 return visit_continue
;
146 ir_vector_reference_visitor::visit(ir_dereference_variable
*ir
)
148 ir_variable
*const var
= ir
->var
;
149 variable_entry
*entry
= this->get_variable_entry(var
);
152 entry
->whole_vector_access
++;
154 return visit_continue
;
158 ir_vector_reference_visitor::visit_enter(ir_swizzle
*ir
)
160 /* Don't descend into a vector ir_dereference_variable below. */
161 if (ir
->val
->as_dereference_variable() && ir
->type
->is_scalar())
162 return visit_continue_with_parent
;
164 return visit_continue
;
168 ir_vector_reference_visitor::visit_enter(ir_assignment
*ir
)
170 if (ir
->lhs
->as_dereference_variable() &&
171 ir
->rhs
->as_dereference_variable() &&
173 /* We'll split copies of a vector to copies of channels, so don't
174 * descend to the ir_dereference_variables.
176 return visit_continue_with_parent
;
178 if (ir
->lhs
->as_dereference_variable() &&
179 is_power_of_two(ir
->write_mask
) &&
181 /* If we're writing just a channel, then channel-splitting the LHS is OK.
183 ir
->rhs
->accept(this);
184 return visit_continue_with_parent
;
186 return visit_continue
;
190 ir_vector_reference_visitor::visit_enter(ir_function_signature
*ir
)
192 /* We don't want to descend into the function parameters and
193 * split them, so just accept the body here.
195 visit_list_elements(this, &ir
->body
);
196 return visit_continue_with_parent
;
199 class ir_vector_splitting_visitor
: public ir_rvalue_visitor
{
201 ir_vector_splitting_visitor(exec_list
*vars
)
203 this->variable_list
= vars
;
206 virtual ir_visitor_status
visit_leave(ir_assignment
*);
208 void handle_rvalue(ir_rvalue
**rvalue
);
209 variable_entry
*get_splitting_entry(ir_variable
*var
);
211 exec_list
*variable_list
;
215 ir_vector_splitting_visitor::get_splitting_entry(ir_variable
*var
)
219 if (!var
->type
->is_vector())
222 foreach_list(node
, &*this->variable_list
) {
223 variable_entry
*entry
= (variable_entry
*)node
;
224 if (entry
->var
== var
) {
233 ir_vector_splitting_visitor::handle_rvalue(ir_rvalue
**rvalue
)
238 ir_swizzle
*swiz
= (*rvalue
)->as_swizzle();
239 if (!swiz
|| !swiz
->type
->is_scalar())
242 ir_dereference_variable
*deref_var
= swiz
->val
->as_dereference_variable();
246 variable_entry
*entry
= get_splitting_entry(deref_var
->var
);
250 ir_variable
*var
= entry
->components
[swiz
->mask
.x
];
251 *rvalue
= new(entry
->mem_ctx
) ir_dereference_variable(var
);
255 ir_vector_splitting_visitor::visit_leave(ir_assignment
*ir
)
257 ir_dereference_variable
*lhs_deref
= ir
->lhs
->as_dereference_variable();
258 ir_dereference_variable
*rhs_deref
= ir
->rhs
->as_dereference_variable();
259 variable_entry
*lhs
= lhs_deref
? get_splitting_entry(lhs_deref
->var
) : NULL
;
260 variable_entry
*rhs
= rhs_deref
? get_splitting_entry(rhs_deref
->var
) : NULL
;
262 if (lhs_deref
&& rhs_deref
&& (lhs
|| rhs
) && !ir
->condition
) {
263 unsigned int rhs_chan
= 0;
265 /* Straight assignment of vector variables. */
266 for (unsigned int i
= 0; i
< ir
->lhs
->type
->vector_elements
; i
++) {
267 ir_dereference
*new_lhs
;
269 void *mem_ctx
= lhs
? lhs
->mem_ctx
: rhs
->mem_ctx
;
270 unsigned int writemask
;
272 if (!(ir
->write_mask
& (1 << i
)))
276 new_lhs
= new(mem_ctx
) ir_dereference_variable(lhs
->components
[i
]);
279 new_lhs
= ir
->lhs
->clone(mem_ctx
, NULL
);
285 new(mem_ctx
) ir_dereference_variable(rhs
->components
[rhs_chan
]);
287 new_rhs
= new(mem_ctx
) ir_swizzle(ir
->rhs
->clone(mem_ctx
, NULL
),
288 rhs_chan
, 0, 0, 0, 1);
291 ir
->insert_before(new(mem_ctx
) ir_assignment(new_lhs
,
299 void *mem_ctx
= lhs
->mem_ctx
;
302 switch (ir
->write_mask
) {
317 assert(!"not reached: non-channelwise dereference of LHS.");
320 ir
->lhs
= new(mem_ctx
) ir_dereference_variable(lhs
->components
[elem
]);
321 ir
->write_mask
= (1 << 0);
323 handle_rvalue(&ir
->rhs
);
325 handle_rvalue(&ir
->rhs
);
328 handle_rvalue(&ir
->condition
);
330 return visit_continue
;
334 brw_do_vector_splitting(exec_list
*instructions
)
336 ir_vector_reference_visitor refs
;
338 visit_list_elements(&refs
, instructions
);
340 /* Trim out variables we can't split. */
341 foreach_list_safe(node
, &refs
.variable_list
) {
342 variable_entry
*entry
= (variable_entry
*)node
;
345 fprintf(stderr
, "vector %s@%p: whole_access %d\n",
346 entry
->var
->name
, (void *) entry
->var
,
347 entry
->whole_vector_access
);
350 if (entry
->whole_vector_access
) {
355 if (refs
.variable_list
.is_empty())
358 void *mem_ctx
= ralloc_context(NULL
);
360 /* Replace the decls of the vectors to be split with their split
363 foreach_list(node
, &refs
.variable_list
) {
364 variable_entry
*entry
= (variable_entry
*)node
;
365 const struct glsl_type
*type
;
366 type
= glsl_type::get_instance(entry
->var
->type
->base_type
, 1, 1);
368 entry
->mem_ctx
= ralloc_parent(entry
->var
);
370 for (unsigned int i
= 0; i
< entry
->var
->type
->vector_elements
; i
++) {
371 const char *name
= ralloc_asprintf(mem_ctx
, "%s_%c",
375 entry
->components
[i
] = new(entry
->mem_ctx
) ir_variable(type
, name
,
377 entry
->var
->insert_before(entry
->components
[i
]);
380 entry
->var
->remove();
383 ir_vector_splitting_visitor
split(&refs
.variable_list
);
384 visit_list_elements(&split
, instructions
);
386 ralloc_free(mem_ctx
);