class loop_unroll_visitor : public ir_hierarchical_visitor {
public:
- loop_unroll_visitor(loop_state *state)
+ loop_unroll_visitor(loop_state *state, unsigned max_iterations)
{
this->state = state;
this->progress = false;
+ this->max_iterations = max_iterations;
}
virtual ir_visitor_status visit_leave(ir_loop *ir);
loop_state *state;
bool progress;
+ unsigned max_iterations;
+};
+
+
+static bool
+is_break(ir_instruction *ir)
+{
+ return ir != NULL && ir->ir_type == ir_type_loop_jump
+ && ((ir_loop_jump *) ir)->is_break();
+}
+
+class loop_unroll_count : public ir_hierarchical_visitor {
+public:
+ int nodes;
+ bool fail;
+
+ loop_unroll_count(exec_list *list)
+ {
+ nodes = 0;
+ fail = false;
+
+ run(list);
+ }
+
+ virtual ir_visitor_status visit_enter(ir_assignment *ir)
+ {
+ nodes++;
+ return visit_continue;
+ }
+
+ virtual ir_visitor_status visit_enter(ir_expression *ir)
+ {
+ nodes++;
+ return visit_continue;
+ }
+
+ virtual ir_visitor_status visit_enter(ir_loop *ir)
+ {
+ fail = true;
+ return visit_continue;
+ }
};
loop_unroll_visitor::visit_leave(ir_loop *ir)
{
loop_variable_state *const ls = this->state->get(ir);
+ int iterations;
/* If we've entered a loop that hasn't been analyzed, something really,
* really bad has happened.
return visit_continue;
}
+ iterations = ls->max_iterations;
+
/* Don't try to unroll loops where the number of iterations is not known
* at compile-time.
*/
- if (ls->max_iterations < 0)
+ if (iterations < 0)
return visit_continue;
/* Don't try to unroll loops that have zillions of iterations either.
*/
- if (ls->max_iterations > 32)
+ if (iterations > (int) max_iterations)
return visit_continue;
- if (ls->num_loop_jumps > 0)
+ /* Don't try to unroll nested loops and loops with a huge body.
+ */
+ loop_unroll_count count(&ir->body_instructions);
+
+ if (count.fail || count.nodes * iterations > (int)max_iterations * 5)
return visit_continue;
- void *const mem_ctx = talloc_parent(ir);
+ if (ls->num_loop_jumps > 1)
+ return visit_continue;
+ else if (ls->num_loop_jumps) {
+ ir_instruction *last_ir = (ir_instruction *) ir->body_instructions.get_tail();
+ assert(last_ir != NULL);
+
+ if (is_break(last_ir)) {
+ /* If the only loop-jump is a break at the end of the loop, the loop
+ * will execute exactly once. Remove the break, set the iteration
+ * count, and fall through to the normal unroller.
+ */
+ last_ir->remove();
+ iterations = 1;
+
+ this->progress = true;
+ } else {
+ ir_if *ir_if = NULL;
+ ir_instruction *break_ir = NULL;
+ bool continue_from_then_branch = false;
+
+ foreach_list(node, &ir->body_instructions) {
+ /* recognize loops in the form produced by ir_lower_jumps */
+ ir_instruction *cur_ir = (ir_instruction *) node;
+
+ ir_if = cur_ir->as_if();
+ if (ir_if != NULL) {
+ /* Determine which if-statement branch, if any, ends with a
+ * break. The branch that did *not* have the break will get a
+ * temporary continue inserted in each iteration of the loop
+ * unroll.
+ *
+ * Note that since ls->num_loop_jumps is <= 1, it is impossible
+ * for both branches to end with a break.
+ */
+ ir_instruction *ir_if_last =
+ (ir_instruction *) ir_if->then_instructions.get_tail();
+
+ if (is_break(ir_if_last)) {
+ continue_from_then_branch = false;
+ break_ir = ir_if_last;
+ break;
+ } else {
+ ir_if_last =
+ (ir_instruction *) ir_if->else_instructions.get_tail();
+
+ if (is_break(ir_if_last)) {
+ break_ir = ir_if_last;
+ continue_from_then_branch = true;
+ break;
+ }
+ }
+ }
+ }
+
+ if (break_ir == NULL)
+ return visit_continue;
+
+ /* move instructions after then if in the continue branch */
+ while (!ir_if->get_next()->is_tail_sentinel()) {
+ ir_instruction *move_ir = (ir_instruction *) ir_if->get_next();
+
+ move_ir->remove();
+ if (continue_from_then_branch)
+ ir_if->then_instructions.push_tail(move_ir);
+ else
+ ir_if->else_instructions.push_tail(move_ir);
+ }
+
+ /* Remove the break from the if-statement.
+ */
+ break_ir->remove();
+
+ void *const mem_ctx = ralloc_parent(ir);
+ ir_instruction *ir_to_replace = ir;
+
+ for (int i = 0; i < iterations; i++) {
+ exec_list copy_list;
+
+ copy_list.make_empty();
+ clone_ir_list(mem_ctx, ©_list, &ir->body_instructions);
+
+ ir_if = ((ir_instruction *) copy_list.get_tail())->as_if();
+ assert(ir_if != NULL);
+
+ ir_to_replace->insert_before(©_list);
+ ir_to_replace->remove();
+
+ /* placeholder that will be removed in the next iteration */
+ ir_to_replace =
+ new(mem_ctx) ir_loop_jump(ir_loop_jump::jump_continue);
+
+ exec_list *const list = (continue_from_then_branch)
+ ? &ir_if->then_instructions : &ir_if->else_instructions;
+
+ list->push_tail(ir_to_replace);
+ }
+
+ ir_to_replace->remove();
+
+ this->progress = true;
+ return visit_continue;
+ }
+ }
+
+ void *const mem_ctx = ralloc_parent(ir);
- for (int i = 0; i < ls->max_iterations; i++) {
+ for (int i = 0; i < iterations; i++) {
exec_list copy_list;
copy_list.make_empty();
bool
-unroll_loops(exec_list *instructions, loop_state *ls)
+unroll_loops(exec_list *instructions, loop_state *ls, unsigned max_iterations)
{
- loop_unroll_visitor v(ls);
+ loop_unroll_visitor v(ls, max_iterations);
v.run(instructions);