3 * @file loop_unrolling.c
6 * File name: ir/opt/loop_unrolling.c
7 * Purpose: Make loop unrolling.
8 * Author: Beyhan Veliev
12 * Copyright: (c) 2004 Universität Karlsruhe
13 * Licence: This file protected by GPL - GNU GENERAL PUBLIC LICENSE.
23 #include "loop_unrolling.h"
35 #include "strength_red.h"
36 #include "compute_loop_info.h"
41 /* We will unroll maximal 4-times. */
44 /* We will know if the head of the loop to be copy.
45 * Default "0" don't copy.*/
46 static int copy_loop_head;
49 ir_node *irn ; /* Node of the loop to be unrolling*/
50 ir_node *copy[MAX_UNROLL] ; /* The copy of the node */
54 * Compare two elements of the copies_t set
56 static int set_cmp(const void *elt, const void *key, size_t size)
58 const copies_t *c1 = elt;
59 const copies_t *c2 = key;
61 return c1->irn != c2->irn;
65 * Remember the new node in the old node by using a field all nodes have.
68 set_new_node (ir_node *old, ir_node *new)
73 /*Check if phi in the loop head is.
75 * @param phi Must be a phi node from the loop.
76 * @param loop_head The loop head .
78 static int is_Phi_in_loop_head(ir_node *phi, ir_node *loop_head) {
79 assert(is_Block(loop_head));
80 return is_Phi(phi) && (get_nodes_block(phi) == loop_head);
84 * Copies predecessors of node to copies of node.
85 * If the predecessor is a loop invariant, then the copy get it
86 * as predecessor, else the copy of the predecessor.
88 * @param l_n A set, where the node of the loop are saved.
89 * @param value A element of the set.
90 * @param info Contains information about the induction variable.
91 * @param unroll_factor A integer 2 <= unroll_factor <= 4.
92 * @param env Free environment pointer.
95 set_preds (set *l_n, copies_t *value, induct_var_info *info, int unroll_factor, void *env)
99 copies_t key, *value_pred;
100 if(value->copy[0] == NULL ||
101 get_irn_op(value->irn) != get_irn_op(value->copy[0]))
104 /* The head of the unrolling loop. */
105 loop_head = get_loop_node(info->l_itervar_phi, 0);
106 irn_arity = get_irn_arity(value->irn);
108 for (i = 0; i < irn_arity; i++) {
109 ir_node *pred = get_irn_n(value->irn, i);
112 value_pred = set_find(l_n, &key, sizeof(key), HASH_PTR(pred));
114 if (value->irn != loop_head && !is_Phi_in_loop_head(value->irn, loop_head)) {
115 if (value_pred == NULL) {
116 /* Is loop invariant. */
117 for (p = 0; p < unroll_factor - 1; p++)
118 set_irn_n(value->copy[p], i, pred);
119 /* pred is a loop invariant. The copies of the successors get it as predecessor. */
122 for (p = 0; p < unroll_factor - 1; p++)
123 set_irn_n(value->copy[p], i, value_pred->copy[p]);
124 /* value_pred->irn is a node from the unrolling loop.
125 * The copies of the successors get the
126 * copies of value_pred as predecessor.
133 /* Set the backedge of phi in the loop head.The backedge of phis in the loop head
134 * must now point to the value defined
135 * in the last copy of the loop body.
137 * @param l_n A set, where the node of the loop are saved.
138 * @param value A phi in the loop head.
139 * @param info Contains information about the induction variable.
140 * @param unroll_factor A integer 2 <= unroll_factor <= 4.
143 set_phi_backedge(set *l_n, copies_t *value, induct_var_info *info, int unroll_factor)
145 copies_t key, *value_pred;
147 /* info->op_pred_pos is the backedge position. */
148 key.irn = get_irn_n(value->irn, info->op_pred_pos);
149 value_pred = set_find(l_n, &key, sizeof(key), HASH_PTR(key.irn));
151 /* value->copy[unroll_factor - 2] is the last copy. */
152 set_Phi_pred(value->irn, info->op_pred_pos, value_pred->copy[unroll_factor - 2]);
156 * Test for a loop head.
158 * Returns true if the node has predecessors in the loop _and_ out of
159 * the loop. Then it is a loop head: The loop can be entered through
162 * @param n The node to be tested. Must be a block.
163 * @param info Contains the loop information.
166 is_loop_head(ir_node *n, induct_var_info *info)
169 int some_outof_loop = 0, some_in_loop = 0;
171 assert(get_irn_op(n) == op_Block);
172 arity = get_Block_n_cfgpreds(n);
174 for (i = 0; i < arity; i++) {
175 ir_node *pred = get_Block_cfgpred(n, i);
177 if (is_loop_invariant(pred, get_loop_node(info->l_itervar_phi, 0))) {
183 return some_outof_loop && some_in_loop;
188 * Test whether the passed loop is a natural loop.
190 * Returns true if the loop has only one loop header and only a single
193 * @param info Contains the loop information.
196 is_natural_loop(induct_var_info *info)
201 l_n_node = get_loop_n_nodes (info->l_itervar_phi);
203 for (i = 1; i < (l_n_node); i ++) {
204 l_node = get_loop_node (info->l_itervar_phi, i);
205 if (is_loop_head(l_node, info)) return 0;
207 if (has_backedges(l_node) && i != l_n_node-1) return 0;
214 * Search for all nodes of a loop.
216 * @param node The induction variable of the loop.
217 * @param loop_head The head of the loop.
218 * @param l_n A set, where the node of the loop are saved.
221 find_loop_nodes(ir_node *node, ir_node *loop_head, set *l_n)
224 copies_t key, *value;
226 /* Add this node to the list. */
229 /* Initialize all copies of the added node with NULL.*/
230 for (i = 0; i < 4; ++i)
232 value = set_insert(l_n, &key, sizeof(key), HASH_PTR(key.irn));
234 /* Add all outs of this node to the list, if they are within the loop. */
235 for (i = get_irn_n_outs(node) - 1; i >= 0; --i) {
236 ir_node *pred = get_irn_out(node, i);
239 if (!is_loop_invariant(pred, loop_head) &&
240 set_find( l_n, &key, sizeof(key), HASH_PTR(key.irn)) == NULL ) {
241 find_loop_nodes(pred, loop_head, l_n);
245 /* Add all ins if they are within the loop. */
246 for (i = get_irn_arity(node) - 1; i >=0; --i) {
247 ir_node *pred = get_irn_n(node, i);
250 if (!is_loop_invariant(pred, loop_head) &&
251 set_find( l_n, &key, sizeof(key), HASH_PTR(key.irn)) == NULL ) {
252 find_loop_nodes(pred, loop_head, l_n);
258 * Make a new loop head if copy_head = 1.
260 * @param l_n A set, where the node of the loop are saved.
261 * @param info Contains the loop information.
262 * @param value A element of the set.
263 * @param unroll_factor A integer 2 <= unroll_factor <= 4.
264 * @param copy_head if non-zero, the loop head will be copied
268 new_loop_head(set *l_n, induct_var_info *info, copies_t *value, int unroll_factor, int copy_head)
271 copies_t block, *value_backedge_jmp, *backedge_jmp_block;
272 /* The predecessor of the loop head in the backedge position */
273 ir_node *backedge_jmp = get_Block_cfgpred(value->irn, info->op_pred_pos);
275 block.irn = backedge_jmp;
276 value_backedge_jmp = set_find(l_n, &block, sizeof(block), HASH_PTR(block.irn));
279 /* The first copy of the loop head must point to the loop head.*/
280 value->copy[0] = new_Block(1, &backedge_jmp);
281 if(info->l_itervar_phi->flags & do_loop){
282 value_backedge_jmp->copy[0] = new_r_Jmp(current_ir_graph, value->copy[0]);
283 for (i = 1; i < unroll_factor - 1; ++i) {
284 /* all other copies must point to the copy before it in the array. */
285 value->copy[i] = new_Block(1, &backedge_jmp);
286 value_backedge_jmp->copy[i] = new_r_Jmp(current_ir_graph, value->copy[i]);
288 for (i = 1; i < unroll_factor - 1; ++i){
289 set_nodes_block(value_backedge_jmp->copy[i-1], get_nodes_block(backedge_jmp));
290 set_Block_cfgpred (value->copy[i], 0, value_backedge_jmp->copy[i-1]);
293 for (i = 1; i < unroll_factor - 1; ++i) {
294 /* all other copies must point to the copy before it in the array. */
295 set_nodes_block(value_backedge_jmp->copy[i-1], get_nodes_block(backedge_jmp));
296 value->copy[i] = new_Block(1, &value_backedge_jmp->copy[i-1]);
300 /* If the loop head must not be copy. block.irn is the successor of the loop head.*/
301 block.irn = get_nodes_block(backedge_jmp);
302 backedge_jmp_block = set_find(l_n, &block, sizeof(block), HASH_PTR(block.irn));
304 /* The first copy of block.irn point to it.
305 The another copies muss point to the copy before it in the array.*/
306 set_irn_n(backedge_jmp_block->copy[0], 0, value_backedge_jmp->irn) ;
308 for (i = 1; i < unroll_factor - 1; ++i)
309 set_irn_n(backedge_jmp_block->copy[i], 0, value_backedge_jmp->copy[i - 1]);
313 /* Set all copies of the induction variable.
315 * @param phi A phi node in the loop head block.
316 * @param phi_pred The predecessor of the phi along the backedge.
317 * @param unroll_factor A integer 2 <= unroll_factor <= 4.
320 set_Phi_copies(copies_t *phi, copies_t *phi_pred, int unroll_factor)
324 if(phi_pred != NULL){
325 /* The first copy of Phi node get the node along the backedge as predecessor. The next copies
326 the copies of this node.*/
327 phi->copy[0] = phi_pred->irn;
328 for (p = 1; p < unroll_factor - 1; ++p) {
329 /* If two phi nodes are in cycle. */
330 if (phi_pred->copy[p - 1] == NULL && get_irn_op(phi_pred->irn) == op_Phi) {
332 phi->copy[p] = phi->irn;
334 phi->copy[p] = phi_pred->irn;
336 phi->copy[p] = phi_pred->copy[p - 1];
339 /* The predecessors of phi are loop invariant and the copies von phi
340 must be phi it self.*/
341 for(p = 0; p < unroll_factor - 1; ++p)
342 phi->copy[p] = phi->irn;
346 * Decide if the loop head must be copied. A head with important nodes
349 * @param l_n A set, where the node of the loop are saved.
350 * @param info Contains information about the induction variable.
352 * @return non-zero if the loop head must be copied
355 loop_head_nodes(set *l_n, induct_var_info *info)
359 ir_node *loop_head = get_loop_node(info->l_itervar_phi, 0);
361 for (value = set_first(l_n); value != NULL; value = set_next(l_n))
362 if (is_no_Block(value->irn) && get_nodes_block(value->irn) == loop_head)
363 /* If the loop head contains just this nodes, than must not be copy. */
364 switch (get_irn_opcode(value->irn)) {
381 /** Copy all loop nodes.
383 * @param l_n Contains all nodes of the loop.
384 * @param info Contains the loop information.
385 * @param unroll_factor A integer 2 <= unroll_factor <= 4.
388 copy_loop_body(set *l_n, induct_var_info *info, int unroll_factor)
391 copies_t *value, *info_op, *phi, *loop_h = NULL, key, *value_block;
392 ir_node *proj_b, *cond, *proj_1, *proj_0, *loop_head;
393 proj_b = get_irn_out(info->cmp, 0);
394 cond = get_irn_out(proj_b, 0);
395 proj_0 = get_irn_out(cond, 0);
396 proj_1 = get_irn_out(cond, 1);
398 loop_head = get_loop_node(info->l_itervar_phi, 0);
399 /* I will create some nodes and need the optimization to
401 opt = get_optimize();
404 for (value = set_first(l_n); value != NULL; value = set_next(l_n)) {
405 if(value->irn == loop_head)
407 else if (is_Phi_in_loop_head(value->irn, loop_head))
409 else if (copy_loop_head){
410 if(value->irn == proj_0){
411 for (i = 0; i < unroll_factor - 1; i++)
412 /* In the copies of the loop head we replace the cmp branche with a jmp node.
413 This is just for "for" loops."do" loops are handled in function
415 value->copy[i] = new_r_Jmp(current_ir_graph, get_nodes_block(value->irn));
416 }else if(value->irn != proj_b && value->irn != cond &&
417 value->irn != proj_1)
418 /* The loop head must be copied. */
419 for (i = 0; i < unroll_factor - 1; i++){
420 copy_irn_to_irg(value->irn, current_ir_graph);
421 value->copy[i] = get_irn_link(value->irn);
424 /* The loop head and its nodes must not be copied. */
425 if((value->irn->op == op_Block &&
426 value->irn != loop_head) ||
427 (value->irn->op != op_Block &&
428 get_nodes_block(value->irn) != loop_head)) {
429 for (i = 0; i<unroll_factor -1; i++) {
430 copy_irn_to_irg(value->irn, current_ir_graph);
431 value->copy[i] = get_irn_link(value->irn);
437 /* Treat the loop head block */
438 new_loop_head(l_n, info, loop_h, unroll_factor, copy_loop_head);
439 /* I have created the nodes and I set the optimization
443 /* Similarly treat the Phis in the loop head block. info->op is the node
444 along the backedges. */
446 info_op = set_find( l_n, &key, sizeof(key), HASH_PTR(key.irn));
447 assert(info_op->irn == get_Phi_pred(info->itervar_phi, info->op_pred_pos));
448 for (i = 0; i < get_irn_n_outs(loop_head); ++i) {
449 ir_node *phi = get_irn_out(loop_head, i);
452 copies_t *phi_pred, *phi_op;
454 key.irn = get_Phi_pred(phi, info->op_pred_pos); // info->op;
455 phi_pred = set_find(l_n, &key, sizeof(key), HASH_PTR(key.irn));
457 phi_op = set_find(l_n, &key, sizeof(key), HASH_PTR(key.irn));
458 set_Phi_copies(phi_op, phi_pred, unroll_factor);
462 for (value = set_first(l_n); value != NULL; value = set_next(l_n)) {
464 /* Set the predecessors of the copies. */
466 set_preds(l_n, value, info, unroll_factor, NULL);
467 }else if ((value->irn->op != op_Block) && get_nodes_block(value->irn) != loop_head)
468 set_preds(l_n, value, info, unroll_factor, NULL);
470 if (is_Phi_in_loop_head(value->irn, loop_head) &&
471 has_backedges(value->irn))
472 /* Set the backedge of phis in the loop head. */
473 set_phi_backedge(l_n, value, info, unroll_factor);
475 if ((value->irn->op != op_Block) && !is_Phi_in_loop_head(value->irn, loop_head) &&
476 value->copy[0] != NULL ) {
477 ir_node *nodes_block = get_nodes_block(value->irn);
479 if (!copy_loop_head && nodes_block == loop_head)
482 key.irn = nodes_block;
483 value_block = set_find( l_n, &key, sizeof(key), HASH_PTR(key.irn));
484 /* Set the copy of the node in the accordant copy of its block. */
485 for(i = 0; i < unroll_factor - 1; i++){
486 set_nodes_block(value->copy[i], value_block->copy[i]);
490 /* I must repair some control flow edges, when the loop head
492 if (copy_loop_head && !(info->l_itervar_phi->flags & do_loop)){
494 value = set_find( l_n, &key, sizeof(key), HASH_PTR(key.irn));
495 key.irn = get_irn_out(proj_0, 0);
496 info_op = set_find( l_n, &key, sizeof(key), HASH_PTR(key.irn));
497 for (i = 0; i < unroll_factor - 1; i++){
498 set_Block_cfgpred(info_op->copy[i], 0, value->copy[i]);
504 * Returns non-zero if a Proj node from the loop has a predecessor that
505 * can throw an exception.
507 * @param node A Proj node from the loop.
509 * @return non-zero if the predecessor of the Proj node can throw an exception
511 static int is_exception_possible(ir_node *node)
513 ir_node *pred = get_Proj_pred(node);
515 /* only fragile ops can throw an exception */
516 if (! is_fragile_op(pred))
520 * fragile ops that are known NOT to throw
521 * an exception if their pin_state is op_pin_state_floats
523 return get_irn_pinned(pred) != op_pin_state_floats;
527 * If a node from the loop is predecessor of the end block, then the end block must get all copies
528 * of this node as predecessors. This is possible with function calls in the unrolling loop.
530 * @param end_block The end block.
531 * @param loop_head The head of the unrolling loop.
532 * @param l_n Contains all nodes of the loop.
533 * @param loop_endblock_outs The set loop_endblock_outs contains all predecessors
534 * of the end block from the unrolling loop.
535 * @param unroll_factor A integer 2 <= unroll_factor <= 4.
538 new_end_block(ir_node *end_block, ir_node *loop_head, set *l_n, set *loop_endblock_outs, int unroll_factor)
540 copies_t key, *value;
541 int set_el, new_preds, all_new_preds, i, q;
542 int old_preds = get_Block_n_cfgpreds(end_block); /* All old predecessors of the end block. */
545 for (i = 0; i < old_preds; i++) {
546 ir_node *pred = get_Block_cfgpred(end_block, i);
549 value = set_find( l_n, &key, sizeof(key), HASH_PTR(key.irn));
551 /* If a predecessor of the end block is a Proj from the unrolling loop (for function calls) .*/
552 if (get_irn_op(pred) == op_Proj && is_exception_possible(pred) && value != NULL &&
553 !(!copy_loop_head && get_nodes_block(pred) == loop_head))
554 value = set_insert(loop_endblock_outs, &key, sizeof(key), HASH_PTR(key.irn));
557 /* The set loop_endblock_outs contains all predecessors of the end block from the unrolling loop,
558 * set_el the amount of such nodes.
560 set_el = set_count (loop_endblock_outs);
562 /* If the end block haven't such predecessors, we are finished. */
565 new_preds = (unroll_factor - 1) * set_el; /* All new predecessors of the end block */
566 all_new_preds = old_preds + new_preds; /* All predecessors of this block. */
567 all_in = alloca(sizeof(*all_in) * all_new_preds); /* A array with size for all predecessors of this block. */
569 for (i = 0; i < old_preds; i++)
570 all_in[i] = get_Block_cfgpred(end_block, i); /* The old predecessors. */
572 value = set_first(loop_endblock_outs);
574 for (; value != NULL; value = set_next(loop_endblock_outs)) {
575 key.irn = value->irn;
576 value = set_find( l_n, &key, sizeof(key), HASH_PTR(key.irn));
577 for (q = 0; q < unroll_factor - 1 ; q++) {
578 all_in[i++] = value->copy[q]; /* The new predecessors. */
583 /* Replace the old predecessors of the end block wit´h the new ones. */
584 set_irn_in(end_block, all_new_preds, all_in);
587 /** new_after_loop_block
589 * The after loop block must examine the possible exceptions in the loop.
590 * If a (Proj) node from the loop is predecessor of this block, then the
591 * after loop block must have as well all copies of this node as predecessors.
593 * @param l_n Contains all nodes of the loop.
594 * @param block A block after the loop.
595 * @param loop_in A node from the loop, that is predecessor of the end block.
596 * @param unroll_factor An integer 2 <= unroll_factor <= 4.
599 new_after_loop_block (set *l_n, ir_node* block, copies_t *loop_in, int unroll_factor)
601 copies_t key, *value;
602 int i, p, q, s, old_preds, new_preds, all_new_preds ;
605 /* The node from the unrolling loop must be a Proj. */
606 if (loop_in->irn->op != op_Proj) return;
608 old_preds = get_Block_n_cfgpreds(block); /* All old predecessors of this block. */
609 new_preds = old_preds * (unroll_factor - 1); /* All new predecessors of this block. */
610 all_new_preds = old_preds + new_preds; /* All predecessors of this block. */
612 all_in = alloca(sizeof(*all_in) * all_new_preds); /* An array with size for all predecessors of this block. */
614 for (i = 0 ; i < old_preds; i++)
615 all_in[i] = get_Block_cfgpred(block, i); /* The old predecessors. */
618 for (i = 0; i < old_preds; i++){
620 value = set_find( l_n, &key, sizeof(key), HASH_PTR(key.irn));
622 for (s = 0; s < (unroll_factor - 1); s++) {
623 all_in[q++] = value->copy[p++]; /* The new predecessors. */
626 /* Replace the old predecessors of the end block whit the new one. */
627 set_irn_in(block, all_new_preds, all_in);
630 /** new_after_loop_node
632 * An after loop node (phi or call) must examine the possible exceptions in the loop.
633 * If a (Proj) node from the loop is predecessor of this node, then the after loop
634 * node must have as well all copies of this node as predecessors.
636 * @param l_n Contains all nodes of the loop.
637 * @param loop_outs Contains nodes after the loop,that have as predecessor a node from the loop.
638 * @param node A node after the loop.
639 * @param loop_in A node (Proj) from the loop, that is predecessor of *node.
640 * @param unroll_factor An integer 2 <= unroll_factor <= 4.
643 new_after_loop_node(set *l_n, set *loop_outs, ir_node *node, copies_t *loop_in, int unroll_factor)
645 ir_node *pred, *block_pred = NULL, *node_block, *new_phi;
646 int phi = 0, old_preds, new_preds, all_new_preds, p, q, i, s;
647 copies_t key, *value = NULL;
650 old_preds = get_irn_arity(node); /* All old predecessors of this node. */
651 new_preds =old_preds * (unroll_factor - 1); /* All new predecessors of this block. */
652 all_new_preds = old_preds + new_preds; /* All predecessors of this block. */
654 all_in = alloca(sizeof(*all_in) * all_new_preds); /* An array with size for all predecessors of this block. */
656 /* Verification Predecessors, successors and operation of node and loop_in.
657 * loop_in must be a Proj node.
659 if (loop_in->irn->op != op_Proj) return;
660 /* Node must be operation Phi with mode memory or a Call node. */
661 if (get_irn_op(node) == op_Phi &&
662 get_irn_mode(node) == mode_M){
663 /* If node is a Phi node,then must have a Call node as successor. */
664 for (i = 0; i < get_irn_n_outs(node); i++)
665 if (get_irn_op(get_irn_out(node, i)) == op_Call) {
673 /* The predecessor of loop_in must not be a loop invariant. */
674 pred = get_Proj_pred(loop_in->irn);
676 value = set_find( l_n, &key, sizeof(key), HASH_PTR(key.irn));
679 node_block = get_nodes_block(node);
681 /* The block of node must also have a (Proj) predecessor from the unrolling loop. */
682 for (i = 0; i < get_Block_n_cfgpreds(node_block); i++) {
683 block_pred = get_Block_cfgpred(node_block, i);
685 if (get_irn_op(block_pred) == op_Proj) {
686 if (get_Proj_pred(block_pred) == pred)
694 if (! block_pred) return;
696 key.irn = block_pred;
697 value = set_find(l_n, &key, sizeof(key), HASH_PTR(key.irn));
702 new_after_loop_block(l_n, node_block, value, unroll_factor);
704 for (i = 0; i < old_preds; ++i)
705 all_in[i] = get_irn_n(node, i); /* The old predecessors. */
708 for (i = 0; i < old_preds; ++i) {
710 value = set_find( l_n, &key, sizeof(key), HASH_PTR(key.irn));
712 for (s = 0; s < (unroll_factor - 1); ++s) {
713 all_in[q] = value->copy[p]; /* The new predecessors. */
718 /* A new phi node with the new predecessors. */
719 new_phi = new_r_Phi(current_ir_graph, get_nodes_block(node), all_new_preds,all_in,
723 exchange(node, new_phi);
725 for (i = 0; i < get_irn_arity(node); ++i)
726 if (get_irn_n(node, i) == pred)
727 set_irn_n(node, i, new_phi);
730 /* The set loop_outs contains the visited nodes and their blocks. */
732 value = set_insert(loop_outs, &key, sizeof(key), HASH_PTR(key.irn));
733 key.irn = get_nodes_block(node);
734 value = set_insert(loop_outs, &key, sizeof(key), HASH_PTR(key.irn));
738 /* Set the outs of the unrolling loop. All loop outs of a node muss now
739 * point to the last copy of it. Just phi nodes in the loop head and proj
740 * nodes save it outs. The all copies of some Projs have too outs.
742 * @param l_n Contains all nodes of the loop.
743 * @param loop_outs The set contains the visited and changed loop outs.
744 * @param loop_endblock_outs The set loop_endblock_outs contains all predecessors
745 * of the end block from the unrolling loop.
746 * @param info Contains the loop information.
747 * @param unroll_factor A integer 2 <= unroll_factor <= 4.
750 set_loop_outs(set *l_n, set *loop_outs, set *loop_endblock_outs, induct_var_info *info, int unroll_factor)
752 copies_t *value, key;
754 ir_node *loop_head = get_loop_node(info->l_itervar_phi, 0);
755 ir_node *end_block = get_irg_end_block(current_ir_graph);
757 for (value = set_first(l_n); value != NULL; value = set_next(l_n)){
758 if(get_irn_node_nr(value->irn) == 35047)
760 if (! is_Phi_in_loop_head(value->irn, loop_head) &&
761 (get_irn_opcode(value->irn) == iro_Proj && value->copy[0] != NULL))
762 for (i = 0; i < get_irn_n_outs(value->irn); ++i) {
763 key.irn = get_irn_out(value->irn, i);
765 /* Search for loop outs. */
766 if (set_find( l_n, &key, sizeof(key), HASH_PTR(key.irn)) == NULL)
767 if ((key.irn->op == op_Block && get_Block_dom_depth(key.irn) >
768 get_Block_dom_depth(loop_head)) ||
769 (key.irn->op != op_Block && get_Block_dom_depth(get_nodes_block(key.irn)) >
770 get_Block_dom_depth(loop_head))) {
772 for (p = 0; p < get_irn_arity(key.irn); p++)
773 if (value->irn == get_irn_n(key.irn, p)) {
774 if (get_irn_op(value->irn) == op_Proj && is_exception_possible(value->irn)) {
775 if (set_find(loop_outs, &key, sizeof(key), HASH_PTR(key.irn)) == NULL) {
776 /* If the loop out is for exceptions in the loop. */
777 if ((get_irn_op(key.irn) == op_Phi && get_irn_mode(key.irn) == mode_M) ||
778 get_irn_op(key.irn) == op_Call)
779 new_after_loop_node(l_n,loop_outs, key.irn, value, unroll_factor);
785 /* Loop outs in the loop head must be not changed.*/
786 if(get_nodes_block(value->irn) != loop_head)
787 set_irn_n(key.irn, p, value->copy[unroll_factor-2]);
792 /* This function searches for loop outs associated with function call in the unrolling loop. */
793 new_end_block (end_block, loop_head, l_n, loop_endblock_outs, unroll_factor);
797 * Unroll the loop body with a factor that must be power of two.
798 * Called from a post walker.
800 * @param n An IR node.
801 * @param env points to a result
803 static void do_loop_unroll(ir_node *n, void *env)
805 int *unroll_done = env;
806 induct_var_info info;
808 set *loop_nodes, *loop_outs, *loop_endblock_outs;
809 ir_node *phi_init, *loop_head;
810 ir_node *backedge_jmp;
811 int l_sons = 0, unroll_factor = 0;
812 tarval *init, *iter_end, *iter_increment,*tarval_null, *tarval_one, *tarval_three, *tarval_two, *diff, *iter_number;
816 info.itervar_phi = n;
818 /* The IR node must be a induction variable. */
819 if (get_irn_op(n) == op_Phi) {
820 if (is_induction_variable (&info) == NULL)
826 /* Brute force limiting of loop body size. */
827 if (get_loop_n_nodes(info.l_itervar_phi) > 2 ) return;
829 /* Only unroll loops that compare against a constant for exiting. */
830 if (info.cmp == NULL) return;
832 /* We only want to unroll innermost loops. */
833 l_sons = get_loop_n_sons(info.l_itervar_phi);
836 /* "do" loops are not implementit gut for this reason
837 at the time we don't unroll them.*/
838 if(info.l_itervar_phi->flags & do_loop)
841 phi_init = get_Phi_pred(info.itervar_phi, info.init_pred_pos);
843 if ((!(info.l_itervar_phi->flags & loop_is_count_loop)) ||
844 (info.l_itervar_phi->flags & loop_is_endless) ||
845 (info.l_itervar_phi->flags & loop_is_dead) ||
846 (info.l_itervar_phi->flags & once))
849 init = info.l_itervar_phi->loop_iter_start;
850 iter_end = info.l_itervar_phi->loop_iter_end;
851 iter_increment = info.l_itervar_phi->loop_iter_increment;
852 tarval_null = get_mode_null(get_tarval_mode(iter_increment));
853 tarval_one = get_mode_one(get_tarval_mode(init));
855 if(tarval_is_double(init) || tarval_is_double(iter_end) || tarval_is_double(iter_increment))
858 if((tarval_is_negative(init) && tarval_is_negative(iter_end)) ||
859 (!tarval_is_negative(init) && !tarval_is_negative(iter_end)) ||
860 (tarval_is_null(init) || tarval_is_null(iter_end)))
861 diff = tarval_abs(tarval_sub(iter_end, init));
863 diff = tarval_add(tarval_abs(iter_end),tarval_abs(init));
865 iter_increment = tarval_abs(iter_increment);
867 if(!(info.l_itervar_phi->flags & do_loop))
868 diff = tarval_add(diff, tarval_one);
869 /* Test for the value of unroll factor. */
870 if (tarval_cmp(tarval_mod(diff,iter_increment), tarval_null) == pn_Cmp_Eq)
871 iter_number = tarval_div(diff, iter_increment);
874 tarval_two = tarval_add(tarval_one, tarval_one);
875 tarval_three = tarval_add(tarval_two,tarval_one);
877 if (tarval_cmp(tarval_mod(iter_number, tarval_three), tarval_null) == pn_Cmp_Eq)
879 else if (tarval_cmp(tarval_mod(iter_number, tarval_two), tarval_null) == pn_Cmp_Eq)
883 if (get_firm_verbosity())
884 printf("\nloop unrolling with factor %d \n", unroll_factor);
886 /* ok, we will do unrolling */
889 /* The unroll factor must be less than 4. */
890 assert(unroll_factor <= MAX_UNROLL);
892 loop_head = (is_natural_loop(&info)) ? get_loop_node(info.l_itervar_phi, 0) : NULL;
894 assert(loop_head != NULL && is_Block(loop_head));
896 /* We assume, the loop head has exactly one backedge. The position of
897 the backedge is in the following variable: */
898 backedge_pos = (is_backedge(loop_head, 0)) ? 0:1;
900 /* A set with the nodes to copy. */
901 loop_nodes = new_set(set_cmp, 8);
902 /* A set with the loop outs for exceptions. */
903 loop_outs = new_set(set_cmp, 8);
905 /* A set containing all predecessors
906 of the end block from the unrolling loop. */
907 loop_endblock_outs = new_set(set_cmp, 8);
908 /* Find all nodes of the unrolling loop. */
909 find_loop_nodes(info.itervar_phi, get_loop_node(info.l_itervar_phi, 0), loop_nodes);
911 /* Decide if the loop head to be copy.*/
912 copy_loop_head = loop_head_nodes(loop_nodes, &info);
914 /* Copy all nodes of the unrolling loop, that muss be copy. */
915 copy_loop_body(loop_nodes, &info, unroll_factor);
917 backedge_jmp = get_Block_cfgpred(loop_head, backedge_pos);
919 /* Set the backedge of the loop head. */
920 for (value = set_first(loop_nodes); value != NULL; value = set_next(loop_nodes)) {
921 if (value->irn == backedge_jmp)
922 set_Block_cfgpred(loop_head, backedge_pos, value->copy[unroll_factor-2]);
924 set_loop_outs(loop_nodes, loop_outs, loop_endblock_outs, &info, unroll_factor);
927 /* Performs loop unrolling for the passed graph. */
928 void optimize_loop_unrolling(ir_graph *irg /* unroll factor, max body size */)
933 if ( !get_opt_loop_unrolling()) return;
935 rem = current_ir_graph;
936 current_ir_graph = irg;
938 /* -- Precompute some information -- */
940 /* Call algorithm that computes the loop information */
941 // construct_backedges(irg);
942 compute_loop_info(irg);
943 /* Call algorithm that computes the backedges */
944 // construct_cf_backedges(irg);
946 /* Call algorithm that computes the dominator trees. */
949 /* Call algorithm that computes the out edges */
950 assure_irg_outs(irg);
952 collect_phiprojs(irg);
954 /* -- Search expressions that can be optimized -- */
955 irg_walk_graph(irg, NULL, do_loop_unroll, &unroll_done);
958 /* unrolling was done, all info is invalid */
959 set_irg_doms_inconsistent(irg);
960 set_irg_outs_inconsistent(irg);
961 set_irg_loopinfo_state(current_ir_graph, loopinfo_cf_inconsistent);
962 set_trouts_inconsistent();
963 set_irg_callee_info_state(irg, irg_callee_info_inconsistent);
966 current_ir_graph = rem;