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.
24 #include "glsl_types.h"
25 #include "loop_analysis.h"
26 #include "ir_hierarchical_visitor.h"
28 static bool is_loop_terminator(ir_if *ir);
30 static bool all_expression_operands_are_loop_constant(ir_rvalue *,
33 static ir_rvalue *get_basic_induction_increment(ir_assignment *, hash_table *);
36 loop_state::loop_state()
38 this->ht = hash_table_ctor(0, hash_table_pointer_hash,
39 hash_table_pointer_compare);
40 this->mem_ctx = ralloc_context(NULL);
41 this->loop_found = false;
45 loop_state::~loop_state()
47 hash_table_dtor(this->ht);
48 ralloc_free(this->mem_ctx);
53 loop_state::insert(ir_loop *ir)
55 loop_variable_state *ls = new(this->mem_ctx) loop_variable_state;
57 hash_table_insert(this->ht, ls, ir);
58 this->loop_found = true;
65 loop_state::get(const ir_loop *ir)
67 return (loop_variable_state *) hash_table_find(this->ht, ir);
72 loop_variable_state::get(const ir_variable *ir)
74 return (loop_variable *) hash_table_find(this->var_hash, ir);
79 loop_variable_state::insert(ir_variable *var)
81 void *mem_ctx = ralloc_parent(this);
82 loop_variable *lv = rzalloc(mem_ctx, loop_variable);
86 hash_table_insert(this->var_hash, lv, lv->var);
87 this->variables.push_tail(lv);
94 loop_variable_state::insert(ir_if *if_stmt)
96 void *mem_ctx = ralloc_parent(this);
97 loop_terminator *t = rzalloc(mem_ctx, loop_terminator);
100 this->terminators.push_tail(t);
106 class loop_analysis : public ir_hierarchical_visitor {
108 loop_analysis(loop_state *loops);
110 virtual ir_visitor_status visit(ir_loop_jump *);
111 virtual ir_visitor_status visit(ir_dereference_variable *);
113 virtual ir_visitor_status visit_enter(ir_call *);
115 virtual ir_visitor_status visit_enter(ir_loop *);
116 virtual ir_visitor_status visit_leave(ir_loop *);
117 virtual ir_visitor_status visit_enter(ir_assignment *);
118 virtual ir_visitor_status visit_leave(ir_assignment *);
119 virtual ir_visitor_status visit_enter(ir_if *);
120 virtual ir_visitor_status visit_leave(ir_if *);
124 int if_statement_depth;
126 ir_assignment *current_assignment;
132 loop_analysis::loop_analysis(loop_state *loops)
133 : loops(loops), if_statement_depth(0), current_assignment(NULL)
140 loop_analysis::visit(ir_loop_jump *ir)
144 assert(!this->state.is_empty());
146 loop_variable_state *const ls =
147 (loop_variable_state *) this->state.get_head();
149 ls->num_loop_jumps++;
151 return visit_continue;
156 loop_analysis::visit_enter(ir_call *ir)
158 /* If we're not somewhere inside a loop, there's nothing to do. */
159 if (this->state.is_empty())
160 return visit_continue;
162 loop_variable_state *const ls =
163 (loop_variable_state *) this->state.get_head();
165 ls->contains_calls = true;
166 return visit_continue_with_parent;
171 loop_analysis::visit(ir_dereference_variable *ir)
173 /* If we're not somewhere inside a loop, there's nothing to do.
175 if (this->state.is_empty())
176 return visit_continue;
178 loop_variable_state *const ls =
179 (loop_variable_state *) this->state.get_head();
181 ir_variable *var = ir->variable_referenced();
182 loop_variable *lv = ls->get(var);
185 lv = ls->insert(var);
186 lv->read_before_write = !this->in_assignee;
189 if (this->in_assignee) {
190 assert(this->current_assignment != NULL);
192 lv->conditional_assignment = (this->if_statement_depth > 0)
193 || (this->current_assignment->condition != NULL);
195 if (lv->first_assignment == NULL) {
196 assert(lv->num_assignments == 0);
198 lv->first_assignment = this->current_assignment;
201 lv->num_assignments++;
202 } else if (lv->first_assignment == this->current_assignment) {
203 /* This catches the case where the variable is used in the RHS of an
204 * assignment where it is also in the LHS.
206 lv->read_before_write = true;
209 return visit_continue;
213 loop_analysis::visit_enter(ir_loop *ir)
215 loop_variable_state *ls = this->loops->insert(ir);
216 this->state.push_head(ls);
218 return visit_continue;
222 loop_analysis::visit_leave(ir_loop *ir)
224 loop_variable_state *const ls =
225 (loop_variable_state *) this->state.pop_head();
227 /* Function calls may contain side effects. These could alter any of our
228 * variables in ways that cannot be known, and may even terminate shader
229 * execution (say, calling discard in the fragment shader). So we can't
230 * rely on any of our analysis about assignments to variables.
232 * We could perform some conservative analysis (prove there's no statically
233 * possible assignment, etc.) but it isn't worth it for now; function
234 * inlining will allow us to unroll loops anyway.
236 if (ls->contains_calls)
237 return visit_continue;
239 foreach_list(node, &ir->body_instructions) {
240 /* Skip over declarations at the start of a loop.
242 if (((ir_instruction *) node)->as_variable())
245 ir_if *if_stmt = ((ir_instruction *) node)->as_if();
247 if ((if_stmt != NULL) && is_loop_terminator(if_stmt))
254 foreach_list_safe(node, &ls->variables) {
255 loop_variable *lv = (loop_variable *) node;
257 /* Move variables that are already marked as being loop constant to
258 * a separate list. These trivially don't need to be tested.
260 if (lv->is_loop_constant()) {
262 ls->constants.push_tail(lv);
266 /* Each variable assigned in the loop that isn't already marked as being loop
267 * constant might still be loop constant. The requirements at this point
270 * - Variable is written before it is read.
272 * - Only one assignment to the variable.
274 * - All operands on the RHS of the assignment are also loop constants.
276 * The last requirement is the reason for the progress loop. A variable
277 * marked as a loop constant on one pass may allow other variables to be
278 * marked as loop constant on following passes.
284 foreach_list_safe(node, &ls->variables) {
285 loop_variable *lv = (loop_variable *) node;
287 if (lv->conditional_assignment || (lv->num_assignments > 1))
290 /* Process the RHS of the assignment. If all of the variables
291 * accessed there are loop constants, then add this
293 ir_rvalue *const rhs = lv->first_assignment->rhs;
294 if (all_expression_operands_are_loop_constant(rhs, ls->var_hash)) {
295 lv->rhs_clean = true;
297 if (lv->is_loop_constant()) {
301 ls->constants.push_tail(lv);
307 /* The remaining variables that are not loop invariant might be loop
308 * induction variables.
310 foreach_list_safe(node, &ls->variables) {
311 loop_variable *lv = (loop_variable *) node;
313 /* If there is more than one assignment to a variable, it cannot be a
314 * loop induction variable. This isn't strictly true, but this is a
315 * very simple induction variable detector, and it can't handle more
318 if (lv->num_assignments > 1)
321 /* All of the variables with zero assignments in the loop are loop
322 * invariant, and they should have already been filtered out.
324 assert(lv->num_assignments == 1);
325 assert(lv->first_assignment != NULL);
327 /* The assignmnet to the variable in the loop must be unconditional.
329 if (lv->conditional_assignment)
332 /* Basic loop induction variables have a single assignment in the loop
333 * that has the form 'VAR = VAR + i' or 'VAR = VAR - i' where i is a
336 ir_rvalue *const inc =
337 get_basic_induction_increment(lv->first_assignment, ls->var_hash);
344 ls->induction_variables.push_tail(lv);
348 return visit_continue;
352 loop_analysis::visit_enter(ir_if *ir)
356 if (!this->state.is_empty())
357 this->if_statement_depth++;
359 return visit_continue;
363 loop_analysis::visit_leave(ir_if *ir)
367 if (!this->state.is_empty())
368 this->if_statement_depth--;
370 return visit_continue;
374 loop_analysis::visit_enter(ir_assignment *ir)
376 /* If we're not somewhere inside a loop, there's nothing to do.
378 if (this->state.is_empty())
379 return visit_continue_with_parent;
381 this->current_assignment = ir;
383 return visit_continue;
387 loop_analysis::visit_leave(ir_assignment *ir)
389 /* Since the visit_enter exits with visit_continue_with_parent for this
390 * case, the loop state stack should never be empty here.
392 assert(!this->state.is_empty());
394 assert(this->current_assignment == ir);
395 this->current_assignment = NULL;
397 return visit_continue;
401 class examine_rhs : public ir_hierarchical_visitor {
403 examine_rhs(hash_table *loop_variables)
405 this->only_uses_loop_constants = true;
406 this->loop_variables = loop_variables;
409 virtual ir_visitor_status visit(ir_dereference_variable *ir)
412 (loop_variable *) hash_table_find(this->loop_variables, ir->var);
416 if (lv->is_loop_constant()) {
417 return visit_continue;
419 this->only_uses_loop_constants = false;
424 hash_table *loop_variables;
425 bool only_uses_loop_constants;
430 all_expression_operands_are_loop_constant(ir_rvalue *ir, hash_table *variables)
432 examine_rhs v(variables);
436 return v.only_uses_loop_constants;
441 get_basic_induction_increment(ir_assignment *ir, hash_table *var_hash)
443 /* The RHS must be a binary expression.
445 ir_expression *const rhs = ir->rhs->as_expression();
447 || ((rhs->operation != ir_binop_add)
448 && (rhs->operation != ir_binop_sub)))
451 /* One of the of operands of the expression must be the variable assigned.
452 * If the operation is subtraction, the variable in question must be the
455 ir_variable *const var = ir->lhs->variable_referenced();
457 ir_variable *const op0 = rhs->operands[0]->variable_referenced();
458 ir_variable *const op1 = rhs->operands[1]->variable_referenced();
460 if (((op0 != var) && (op1 != var))
461 || ((op1 == var) && (rhs->operation == ir_binop_sub)))
464 ir_rvalue *inc = (op0 == var) ? rhs->operands[1] : rhs->operands[0];
466 if (inc->as_constant() == NULL) {
467 ir_variable *const inc_var = inc->variable_referenced();
468 if (inc_var != NULL) {
470 (loop_variable *) hash_table_find(var_hash, inc_var);
472 if (!lv->is_loop_constant())
478 if ((inc != NULL) && (rhs->operation == ir_binop_sub)) {
479 void *mem_ctx = ralloc_parent(ir);
481 inc = new(mem_ctx) ir_expression(ir_unop_neg,
483 inc->clone(mem_ctx, NULL),
492 * Detect whether an if-statement is a loop terminating condition
494 * Detects if-statements of the form
496 * (if (expression bool ...) (break))
499 is_loop_terminator(ir_if *ir)
501 if (!ir->else_instructions.is_empty())
504 ir_instruction *const inst =
505 (ir_instruction *) ir->then_instructions.get_head();
506 assert(inst != NULL);
508 if (inst->ir_type != ir_type_loop_jump)
511 ir_loop_jump *const jump = (ir_loop_jump *) inst;
512 if (jump->mode != ir_loop_jump::jump_break)
520 analyze_loop_variables(exec_list *instructions)
522 loop_state *loops = new loop_state;
523 loop_analysis v(loops);