Go to most recent revision | Details | Last modification | View Log | RSS feed
Rev | Author | Line No. | Line |
---|---|---|---|
4358 | Serge | 1 | /* |
2 | * Copyright © 2010 Intel Corporation |
||
3 | * |
||
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: |
||
10 | * |
||
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 |
||
13 | * Software. |
||
14 | * |
||
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. |
||
22 | */ |
||
23 | |||
24 | /** |
||
25 | * \file opt_array_splitting.cpp |
||
26 | * |
||
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. |
||
30 | * |
||
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. |
||
34 | */ |
||
35 | |||
36 | #include "ir.h" |
||
37 | #include "ir_visitor.h" |
||
38 | #include "ir_rvalue_visitor.h" |
||
39 | #include "glsl_types.h" |
||
40 | |||
41 | static bool debug = false; |
||
42 | |||
43 | namespace opt_array_splitting { |
||
44 | |||
45 | class variable_entry : public exec_node |
||
46 | { |
||
47 | public: |
||
48 | variable_entry(ir_variable *var) |
||
49 | { |
||
50 | this->var = var; |
||
51 | this->split = true; |
||
52 | this->declaration = false; |
||
53 | this->components = NULL; |
||
54 | this->mem_ctx = NULL; |
||
55 | if (var->type->is_array()) |
||
56 | this->size = var->type->length; |
||
57 | else |
||
58 | this->size = var->type->matrix_columns; |
||
59 | } |
||
60 | |||
61 | ir_variable *var; /* The key: the variable's pointer. */ |
||
62 | unsigned size; /* array length or matrix columns */ |
||
63 | |||
64 | /** Whether this array should be split or not. */ |
||
65 | bool split; |
||
66 | |||
67 | /* If the variable had a decl we can work with in the instruction |
||
68 | * stream. We can't do splitting on function arguments, which |
||
69 | * don't get this variable set. |
||
70 | */ |
||
71 | bool declaration; |
||
72 | |||
73 | ir_variable **components; |
||
74 | |||
75 | /** ralloc_parent(this->var) -- the shader's talloc context. */ |
||
76 | void *mem_ctx; |
||
77 | }; |
||
78 | |||
79 | } /* namespace */ |
||
80 | using namespace opt_array_splitting; |
||
81 | |||
82 | /** |
||
83 | * This class does a walk over the tree, coming up with the set of |
||
84 | * variables that could be split by looking to see if they are arrays |
||
85 | * that are only ever constant-index dereferenced. |
||
86 | */ |
||
87 | class ir_array_reference_visitor : public ir_hierarchical_visitor { |
||
88 | public: |
||
89 | ir_array_reference_visitor(void) |
||
90 | { |
||
91 | this->mem_ctx = ralloc_context(NULL); |
||
92 | this->variable_list.make_empty(); |
||
93 | } |
||
94 | |||
95 | ~ir_array_reference_visitor(void) |
||
96 | { |
||
97 | ralloc_free(mem_ctx); |
||
98 | } |
||
99 | |||
100 | bool get_split_list(exec_list *instructions, bool linked); |
||
101 | |||
102 | virtual ir_visitor_status visit(ir_variable *); |
||
103 | virtual ir_visitor_status visit(ir_dereference_variable *); |
||
104 | virtual ir_visitor_status visit_enter(ir_dereference_array *); |
||
105 | virtual ir_visitor_status visit_enter(ir_function_signature *); |
||
106 | |||
107 | variable_entry *get_variable_entry(ir_variable *var); |
||
108 | |||
109 | /* List of variable_entry */ |
||
110 | exec_list variable_list; |
||
111 | |||
112 | void *mem_ctx; |
||
113 | }; |
||
114 | |||
115 | variable_entry * |
||
116 | ir_array_reference_visitor::get_variable_entry(ir_variable *var) |
||
117 | { |
||
118 | assert(var); |
||
119 | |||
120 | if (var->mode != ir_var_auto && |
||
121 | var->mode != ir_var_temporary) |
||
122 | return NULL; |
||
123 | |||
124 | if (!(var->type->is_array() || var->type->is_matrix())) |
||
125 | return NULL; |
||
126 | |||
127 | /* If the array hasn't been sized yet, we can't split it. After |
||
128 | * linking, this should be resolved. |
||
129 | */ |
||
130 | if (var->type->is_array() && var->type->length == 0) |
||
131 | return NULL; |
||
132 | |||
133 | foreach_iter(exec_list_iterator, iter, this->variable_list) { |
||
134 | variable_entry *entry = (variable_entry *)iter.get(); |
||
135 | if (entry->var == var) |
||
136 | return entry; |
||
137 | } |
||
138 | |||
139 | variable_entry *entry = new(mem_ctx) variable_entry(var); |
||
140 | this->variable_list.push_tail(entry); |
||
141 | return entry; |
||
142 | } |
||
143 | |||
144 | |||
145 | ir_visitor_status |
||
146 | ir_array_reference_visitor::visit(ir_variable *ir) |
||
147 | { |
||
148 | variable_entry *entry = this->get_variable_entry(ir); |
||
149 | |||
150 | if (entry) |
||
151 | entry->declaration = true; |
||
152 | |||
153 | return visit_continue; |
||
154 | } |
||
155 | |||
156 | ir_visitor_status |
||
157 | ir_array_reference_visitor::visit(ir_dereference_variable *ir) |
||
158 | { |
||
159 | variable_entry *entry = this->get_variable_entry(ir->var); |
||
160 | |||
161 | /* If we made it to here without seeing an ir_dereference_array, |
||
162 | * then the dereference of this array didn't have a constant index |
||
163 | * (see the visit_continue_with_parent below), so we can't split |
||
164 | * the variable. |
||
165 | */ |
||
166 | if (entry) |
||
167 | entry->split = false; |
||
168 | |||
169 | return visit_continue; |
||
170 | } |
||
171 | |||
172 | ir_visitor_status |
||
173 | ir_array_reference_visitor::visit_enter(ir_dereference_array *ir) |
||
174 | { |
||
175 | ir_dereference_variable *deref = ir->array->as_dereference_variable(); |
||
176 | if (!deref) |
||
177 | return visit_continue; |
||
178 | |||
179 | variable_entry *entry = this->get_variable_entry(deref->var); |
||
180 | |||
181 | /* If the access to the array has a variable index, we wouldn't |
||
182 | * know which split variable this dereference should go to. |
||
183 | */ |
||
184 | if (entry && !ir->array_index->as_constant()) |
||
185 | entry->split = false; |
||
186 | |||
187 | return visit_continue_with_parent; |
||
188 | } |
||
189 | |||
190 | ir_visitor_status |
||
191 | ir_array_reference_visitor::visit_enter(ir_function_signature *ir) |
||
192 | { |
||
193 | /* We don't have logic for array-splitting function arguments, |
||
194 | * so just look at the body instructions and not the parameter |
||
195 | * declarations. |
||
196 | */ |
||
197 | visit_list_elements(this, &ir->body); |
||
198 | return visit_continue_with_parent; |
||
199 | } |
||
200 | |||
201 | bool |
||
202 | ir_array_reference_visitor::get_split_list(exec_list *instructions, |
||
203 | bool linked) |
||
204 | { |
||
205 | visit_list_elements(this, instructions); |
||
206 | |||
207 | /* If the shaders aren't linked yet, we can't mess with global |
||
208 | * declarations, which need to be matched by name across shaders. |
||
209 | */ |
||
210 | if (!linked) { |
||
211 | foreach_list(node, instructions) { |
||
212 | ir_variable *var = ((ir_instruction *)node)->as_variable(); |
||
213 | if (var) { |
||
214 | variable_entry *entry = get_variable_entry(var); |
||
215 | if (entry) |
||
216 | entry->remove(); |
||
217 | } |
||
218 | } |
||
219 | } |
||
220 | |||
221 | /* Trim out variables we found that we can't split. */ |
||
222 | foreach_iter(exec_list_iterator, iter, variable_list) { |
||
223 | variable_entry *entry = (variable_entry *)iter.get(); |
||
224 | |||
225 | if (debug) { |
||
226 | printf("array %s@%p: decl %d, split %d\n", |
||
227 | entry->var->name, (void *) entry->var, entry->declaration, |
||
228 | entry->split); |
||
229 | } |
||
230 | |||
231 | if (!(entry->declaration && entry->split)) { |
||
232 | entry->remove(); |
||
233 | } |
||
234 | } |
||
235 | |||
236 | return !variable_list.is_empty(); |
||
237 | } |
||
238 | |||
239 | /** |
||
240 | * This class rewrites the dereferences of arrays that have been split |
||
241 | * to use the newly created ir_variables for each component. |
||
242 | */ |
||
243 | class ir_array_splitting_visitor : public ir_rvalue_visitor { |
||
244 | public: |
||
245 | ir_array_splitting_visitor(exec_list *vars) |
||
246 | { |
||
247 | this->variable_list = vars; |
||
248 | } |
||
249 | |||
250 | virtual ~ir_array_splitting_visitor() |
||
251 | { |
||
252 | } |
||
253 | |||
254 | virtual ir_visitor_status visit_leave(ir_assignment *); |
||
255 | |||
256 | void split_deref(ir_dereference **deref); |
||
257 | void handle_rvalue(ir_rvalue **rvalue); |
||
258 | variable_entry *get_splitting_entry(ir_variable *var); |
||
259 | |||
260 | exec_list *variable_list; |
||
261 | }; |
||
262 | |||
263 | variable_entry * |
||
264 | ir_array_splitting_visitor::get_splitting_entry(ir_variable *var) |
||
265 | { |
||
266 | assert(var); |
||
267 | |||
268 | foreach_iter(exec_list_iterator, iter, *this->variable_list) { |
||
269 | variable_entry *entry = (variable_entry *)iter.get(); |
||
270 | if (entry->var == var) { |
||
271 | return entry; |
||
272 | } |
||
273 | } |
||
274 | |||
275 | return NULL; |
||
276 | } |
||
277 | |||
278 | void |
||
279 | ir_array_splitting_visitor::split_deref(ir_dereference **deref) |
||
280 | { |
||
281 | ir_dereference_array *deref_array = (*deref)->as_dereference_array(); |
||
282 | if (!deref_array) |
||
283 | return; |
||
284 | |||
285 | ir_dereference_variable *deref_var = deref_array->array->as_dereference_variable(); |
||
286 | if (!deref_var) |
||
287 | return; |
||
288 | ir_variable *var = deref_var->var; |
||
289 | |||
290 | variable_entry *entry = get_splitting_entry(var); |
||
291 | if (!entry) |
||
292 | return; |
||
293 | |||
294 | ir_constant *constant = deref_array->array_index->as_constant(); |
||
295 | assert(constant); |
||
296 | |||
297 | if (constant->value.i[0] < (int)entry->size) { |
||
298 | *deref = new(entry->mem_ctx) |
||
299 | ir_dereference_variable(entry->components[constant->value.i[0]]); |
||
300 | } else { |
||
301 | /* There was a constant array access beyond the end of the |
||
302 | * array. This might have happened due to constant folding |
||
303 | * after the initial parse. This produces an undefined value, |
||
304 | * but shouldn't crash. Just give them an uninitialized |
||
305 | * variable. |
||
306 | */ |
||
307 | ir_variable *temp = new(entry->mem_ctx) ir_variable(deref_array->type, |
||
308 | "undef", |
||
309 | ir_var_temporary); |
||
310 | entry->components[0]->insert_before(temp); |
||
311 | *deref = new(entry->mem_ctx) ir_dereference_variable(temp); |
||
312 | } |
||
313 | } |
||
314 | |||
315 | void |
||
316 | ir_array_splitting_visitor::handle_rvalue(ir_rvalue **rvalue) |
||
317 | { |
||
318 | if (!*rvalue) |
||
319 | return; |
||
320 | |||
321 | ir_dereference *deref = (*rvalue)->as_dereference(); |
||
322 | |||
323 | if (!deref) |
||
324 | return; |
||
325 | |||
326 | split_deref(&deref); |
||
327 | *rvalue = deref; |
||
328 | } |
||
329 | |||
330 | ir_visitor_status |
||
331 | ir_array_splitting_visitor::visit_leave(ir_assignment *ir) |
||
332 | { |
||
333 | /* The normal rvalue visitor skips the LHS of assignments, but we |
||
334 | * need to process those just the same. |
||
335 | */ |
||
336 | ir_rvalue *lhs = ir->lhs; |
||
337 | |||
338 | handle_rvalue(&lhs); |
||
339 | ir->lhs = lhs->as_dereference(); |
||
340 | |||
341 | ir->lhs->accept(this); |
||
342 | |||
343 | handle_rvalue(&ir->rhs); |
||
344 | ir->rhs->accept(this); |
||
345 | |||
346 | if (ir->condition) { |
||
347 | handle_rvalue(&ir->condition); |
||
348 | ir->condition->accept(this); |
||
349 | } |
||
350 | |||
351 | return visit_continue; |
||
352 | } |
||
353 | |||
354 | bool |
||
355 | optimize_split_arrays(exec_list *instructions, bool linked) |
||
356 | { |
||
357 | ir_array_reference_visitor refs; |
||
358 | if (!refs.get_split_list(instructions, linked)) |
||
359 | return false; |
||
360 | |||
361 | void *mem_ctx = ralloc_context(NULL); |
||
362 | |||
363 | /* Replace the decls of the arrays to be split with their split |
||
364 | * components. |
||
365 | */ |
||
366 | foreach_iter(exec_list_iterator, iter, refs.variable_list) { |
||
367 | variable_entry *entry = (variable_entry *)iter.get(); |
||
368 | const struct glsl_type *type = entry->var->type; |
||
369 | const struct glsl_type *subtype; |
||
370 | |||
371 | if (type->is_matrix()) |
||
372 | subtype = type->column_type(); |
||
373 | else |
||
374 | subtype = type->fields.array; |
||
375 | |||
376 | entry->mem_ctx = ralloc_parent(entry->var); |
||
377 | |||
378 | entry->components = ralloc_array(mem_ctx, |
||
379 | ir_variable *, |
||
380 | entry->size); |
||
381 | |||
382 | for (unsigned int i = 0; i < entry->size; i++) { |
||
383 | const char *name = ralloc_asprintf(mem_ctx, "%s_%d", |
||
384 | entry->var->name, i); |
||
385 | |||
386 | entry->components[i] = |
||
387 | new(entry->mem_ctx) ir_variable(subtype, name, ir_var_temporary); |
||
388 | entry->var->insert_before(entry->components[i]); |
||
389 | } |
||
390 | |||
391 | entry->var->remove(); |
||
392 | } |
||
393 | |||
394 | ir_array_splitting_visitor split(&refs.variable_list); |
||
395 | visit_list_elements(&split, instructions); |
||
396 | |||
397 | if (debug) |
||
398 | _mesa_print_ir(instructions, NULL); |
||
399 | |||
400 | ralloc_free(mem_ctx); |
||
401 | |||
402 | return true; |
||
403 | |||
404 | }>> |