2 * Copyright (C) 1995-2008 University of Karlsruhe. All right reserved.
4 * This file is part of libFirm.
6 * This file may be distributed and/or modified under the terms of the
7 * GNU General Public License version 2 as published by the Free Software
8 * Foundation and appearing in the file LICENSE.GPL included in the
9 * packaging of this file.
11 * Licensees holding valid libFirm Professional Edition licenses may use
12 * this file in accordance with the libFirm Commercial License.
13 * Agreement provided with the Software.
15 * This file is provided AS IS with NO WARRANTY OF ANY KIND, INCLUDING THE
16 * WARRANTY OF DESIGN, MERCHANTABILITY AND FITNESS FOR A PARTICULAR
22 * @brief Operator Strength Reduction.
24 * @author Michael Beck
27 * Implementation of the Operator Strength Reduction algorithm
28 * by Keith D. Cooper, L. Taylor Simpson, Christopher A. Vick.
34 #include "iroptimize.h"
56 /** The debug handle. */
57 DEBUG_ONLY(static firm_dbg_module_t *dbg;)
61 ir_node *head; /**< the head of the list */
62 tarval *init; /**< the init value iff only one exists. */
63 tarval *incr; /**< the induction variable increment if only a single const exists. */
64 unsigned code; /**< == iro_Add if +incr, iro_Sub if -incr, 0 if not analysed, iro_Bad else */
68 typedef struct node_entry {
69 unsigned DFSnum; /**< the DFS number of this node */
70 unsigned low; /**< the low number of this node */
71 ir_node *header; /**< the header of this node */
72 int in_stack; /**< flag, set if the node is on the stack */
73 ir_node *next; /**< link to the next node the the same scc */
74 scc *pscc; /**< the scc of this node */
75 unsigned POnum; /**< the post order number for blocks */
78 /** The environment. */
79 typedef struct iv_env {
80 struct obstack obst; /**< an obstack for allocations */
81 ir_node **stack; /**< the node stack */
82 int tos; /**< tos index */
83 unsigned nextDFSnum; /**< the current DFS number */
84 unsigned POnum; /**< current post order number */
85 set *quad_map; /**< a map from (op, iv, rc) to node */
86 set *lftr_edges; /**< the set of lftr edges */
87 unsigned replaced; /**< number of replaced ops */
88 unsigned lftr_replaced; /**< number of applied linear function test replacements */
89 unsigned osr_flags; /**< additional flags steering the transformation */
90 unsigned need_postpass; /**< set, if a post pass is needed to fix Add and Sub nodes */
91 /** Function called to process a SCC. */
92 void (*process_scc)(scc *pscc, struct iv_env *env);
96 * An entry in the (op, node, node) -> node map.
98 typedef struct quadruple_t {
99 ir_opcode code; /**< the opcode of the reduced operation */
100 ir_node *op1; /**< the first operand the reduced operation */
101 ir_node *op2; /**< the second operand of the reduced operation */
103 ir_node *res; /**< the reduced operation */
109 typedef struct LFTR_edge {
110 ir_node *src; /**< the source node */
111 ir_node *dst; /**< the destination node */
112 ir_opcode code; /**< the opcode that must be applied */
113 ir_node *rc; /**< the region const that must be applied */
117 static ir_node *reduce(ir_node *orig, ir_node *iv, ir_node *rc, iv_env *env);
120 * Compare two LFTR edges.
122 static int LFTR_cmp(const void *e1, const void *e2, size_t size)
124 const LFTR_edge *l1 = e1;
125 const LFTR_edge *l2 = e2;
128 return l1->src != l2->src;
134 * @param src the source node of the transition
136 static LFTR_edge *LFTR_find(ir_node *src, iv_env *env)
142 return set_find(env->lftr_edges, &key, sizeof(key), HASH_PTR(src));
148 * @param src the source node of the edge
149 * @param dst the destination node of the edge
150 * @param code the opcode of the transformed transition
151 * @param rc the region const used in the transition
152 * @param env the environment
154 static void LFTR_add(ir_node *src, ir_node *dst, ir_opcode code, ir_node *rc, iv_env *env)
164 * There might be more than one edge here. This is rather bad
165 * because we currently store only one.
167 // assert(LFTR_find(src, env) == NULL);
168 set_insert(env->lftr_edges, &key, sizeof(key), HASH_PTR(src));
172 * Gets the node_entry of a node.
174 * @param irn the node
175 * @param env the environment
177 static node_entry *get_irn_ne(ir_node *irn, iv_env *env)
179 node_entry *e = get_irn_link(irn);
182 e = OALLOCZ(&env->obst, node_entry);
183 set_irn_link(irn, e);
189 * Gets the scc from an induction variable.
191 * @param iv any node of the induction variable
192 * @param env the environment
194 static scc *get_iv_scc(ir_node *iv, iv_env *env)
196 node_entry *e = get_irn_ne(iv, env);
201 * Check if irn is an IV.
203 * @param irn the node to check
204 * @param env the environment
206 * @returns the header if it is one, NULL else
208 static ir_node *is_iv(ir_node *irn, iv_env *env)
210 return get_irn_ne(irn, env)->header;
214 * Check if irn is a region constant.
215 * The block or irn must strictly dominate the header block.
217 * @param irn the node to check
218 * @param header_block the header block of the induction variable
220 static int is_rc(ir_node *irn, ir_node *header_block)
222 ir_node *block = get_nodes_block(irn);
224 return (block != header_block) && block_dominates(block, header_block);
228 * Set compare function for the quad set.
230 static int quad_cmp(const void *e1, const void *e2, size_t size)
232 const quadruple_t *c1 = e1;
233 const quadruple_t *c2 = e2;
236 return c1->code != c2->code || c1->op1 != c2->op1 || c1->op2 != c2->op2;
240 * Check if an reduced operation was already calculated.
242 * @param code the opcode of the operation
243 * @param op1 the first operand of the operation
244 * @param op2 the second operand of the operation
245 * @param env the environment
247 * @return the already reduced node or NULL if this operation is not yet reduced
249 static ir_node *search(ir_opcode code, ir_node *op1, ir_node *op2, iv_env *env)
251 quadruple_t key, *entry;
257 entry = set_find(env->quad_map, &key, sizeof(key),
258 (code * 9) ^ HASH_PTR(op1) ^HASH_PTR(op2));
265 * Add an reduced operation.
267 * @param code the opcode of the operation
268 * @param op1 the first operand of the operation
269 * @param op2 the second operand of the operation
270 * @param result the result of the reduced operation
271 * @param env the environment
273 static void add(ir_opcode code, ir_node *op1, ir_node *op2, ir_node *result, iv_env *env)
282 set_insert(env->quad_map, &key, sizeof(key),
283 (code * 9) ^ HASH_PTR(op1) ^HASH_PTR(op2));
287 * Find a location where to place a bin-op whose operands are in
290 * @param block1 the block of the first operand
291 * @param block2 the block of the second operand
293 * Note that we know here that such a place must exists. Moreover, this means
294 * that either block1 dominates block2 or vice versa. So, just return
297 static ir_node *find_location(ir_node *block1, ir_node *block2)
299 if (block_dominates(block1, block2))
301 assert(block_dominates(block2, block1));
303 } /* find_location */
306 * Create a node that executes an op1 code op1 operation.
308 * @param code the opcode to execute
309 * @param db debug info to add to the new node
310 * @param op1 the first operand
311 * @param op2 the second operand
312 * @param mode the mode of the new operation
314 * @return the newly created node
316 static ir_node *do_apply(ir_opcode code, dbg_info *db, ir_node *op1, ir_node *op2, ir_mode *mode)
319 ir_node *block = find_location(get_nodes_block(op1), get_nodes_block(op2));
323 result = new_rd_Mul(db, block, op1, op2, mode);
326 result = new_rd_Add(db, block, op1, op2, mode);
329 result = new_rd_Sub(db, block, op1, op2, mode);
332 panic("Unsupported opcode");
338 * The Apply operation.
340 * @param orig the node that represent the original operation and determines
341 * the opcode, debug-info and mode of a newly created one
342 * @param op1 the first operand
343 * @param op2 the second operand
344 * @param env the environment
346 * @return the newly created node
348 static ir_node *apply(ir_node *header, ir_node *orig, ir_node *op1, ir_node *op2, iv_env *env)
350 ir_opcode code = get_irn_opcode(orig);
351 ir_node *result = search(code, op1, op2, env);
353 if (result == NULL) {
354 dbg_info *db = get_irn_dbg_info(orig);
355 ir_node *op1_header = get_irn_ne(op1, env)->header;
356 ir_node *op2_header = get_irn_ne(op2, env)->header;
358 if (op1_header == header && is_rc(op2, op1_header)) {
359 result = reduce(orig, op1, op2, env);
361 else if (op2_header == header && is_rc(op1, op2_header)) {
362 result = reduce(orig, op2, op1, env);
365 result = do_apply(code, db, op1, op2, get_irn_mode(orig));
366 get_irn_ne(result, env)->header = NULL;
373 * The Reduce operation.
375 * @param orig the node that represent the original operation and determines
376 * the opcode, debug-info and mode of a newly created one
377 * @param iv the induction variable
378 * @param rc the region constant
379 * @param env the environment
381 * @return the reduced node
383 static ir_node *reduce(ir_node *orig, ir_node *iv, ir_node *rc, iv_env *env)
385 ir_opcode code = get_irn_opcode(orig);
386 ir_node *result = search(code, iv, rc, env);
388 /* check if we have already done this operation on the iv */
389 if (result == NULL) {
390 node_entry *e, *iv_e;
392 ir_mode *mode = get_irn_mode(orig);
394 result = exact_copy(iv);
396 if (get_irn_mode(result) != mode) {
398 * Beware: we must always create a new induction variable with the same mode
399 * as the node we are replacing. Especially this means the mode might be changed
400 * from P to I and back. This is always possible, because we have only Phi, Add
402 * However, this might lead to AddIs(Iu,Is) which we must fix. The best way to do this
403 * seems to be a post-pass, or we might end with useless Conv's.
405 set_irn_mode(result, mode);
406 env->need_postpass = 1;
408 add(code, iv, rc, result, env);
409 DB((dbg, LEVEL_3, " Created new %+F for %+F (%s %+F)\n", result, iv,
410 get_irn_opname(orig), rc));
412 iv_e = get_irn_ne(iv, env);
413 e = get_irn_ne(result, env);
414 e->header = iv_e->header;
416 /* create the LFTR edge */
417 LFTR_add(iv, result, code, rc, env);
419 for (i = get_irn_arity(result) - 1; i >= 0; --i) {
420 ir_node *o = get_irn_n(result, i);
422 e = get_irn_ne(o, env);
423 if (e->header == iv_e->header)
424 o = reduce(orig, o, rc, env);
425 else if (is_Phi(result) || code == iro_Mul)
426 o = apply(iv_e->header, orig, o, rc, env);
427 set_irn_n(result, i, o);
430 DB((dbg, LEVEL_3, " Already Created %+F for %+F (%s %+F)\n", result, iv,
431 get_irn_opname(orig), rc));
437 * Update the scc for a newly created IV.
439 static void update_scc(ir_node *iv, node_entry *e, iv_env *env)
442 ir_node *header = e->header;
443 waitq *wq = new_waitq();
445 DB((dbg, LEVEL_2, " Creating SCC for new an induction variable:\n "));
449 ir_node *irn = waitq_get(wq);
450 node_entry *ne = get_irn_ne(irn, env);
454 ne->next = pscc->head;
456 DB((dbg, LEVEL_2, " %+F,", irn));
458 for (i = get_irn_arity(irn) - 1; i >= 0; --i) {
459 ir_node *pred = get_irn_n(irn, i);
460 node_entry *pe = get_irn_ne(pred, env);
462 if (pe->header == header && pe->pscc == NULL) {
463 /* set the pscc here to ensure that the node is NOT enqueued another time */
468 } while (! waitq_empty(wq));
470 DB((dbg, LEVEL_2, "\n"));
474 * The Replace operation. We found a node representing iv (+,-,*) rc
475 * that can be removed by replacing the induction variable iv by a new
476 * one that 'applies' the operation 'irn'.
478 * @param irn the node that will be replaced
479 * @param iv the induction variable
480 * @param rc the region constant
481 * @param env the environment
483 static int replace(ir_node *irn, ir_node *iv, ir_node *rc, iv_env *env)
487 DB((dbg, LEVEL_2, " Replacing %+F\n", irn));
489 result = reduce(irn, iv, rc, env);
493 hook_strength_red(current_ir_graph, irn);
494 exchange(irn, result);
495 e = get_irn_ne(result, env);
496 if (e->pscc == NULL) {
497 e->pscc = OALLOCZ(&env->obst, scc);
498 update_scc(result, e, env);
508 * check if a given node is a mul with 2, 4, 8
510 static int is_x86_shift_const(ir_node *mul)
517 /* normalization put constants on the right side */
518 rc = get_Mul_right(mul);
520 tarval *tv = get_Const_tarval(rc);
522 if (tarval_is_long(tv)) {
523 long value = get_tarval_long(tv);
525 if (value == 2 || value == 4 || value == 8) {
526 /* do not reduce multiplications by 2, 4, 8 */
532 } /* is_x86_shift_const */
536 * Check if an IV represents a counter with constant limits.
538 * @param iv any node of the induction variable
539 * @param env the environment
541 static int is_counter_iv(ir_node *iv, iv_env *env)
543 node_entry *e = get_irn_ne(iv, env);
545 ir_node *have_init = NULL;
546 ir_node *have_incr = NULL;
547 ir_opcode code = iro_Bad;
550 if (pscc->code != 0) {
551 /* already analysed */
552 return pscc->code != iro_Bad;
555 pscc->code = iro_Bad;
556 for (irn = pscc->head; irn != NULL; irn = e->next) {
558 if (have_incr != NULL)
561 have_incr = get_Add_right(irn);
562 if (! is_Const(have_incr)) {
563 have_incr = get_Add_left(irn);
564 if (! is_Const(have_incr))
568 } else if (is_Sub(irn)) {
569 if (have_incr != NULL)
572 have_incr = get_Sub_right(irn);
573 if (! is_Const(have_incr))
576 } else if (is_Phi(irn)) {
579 for (i = get_Phi_n_preds(irn) - 1; i >= 0; --i) {
580 ir_node *pred = get_Phi_pred(irn, i);
581 node_entry *ne = get_irn_ne(pred, env);
583 if (ne->header == e->header)
585 if (have_init != NULL)
588 if (! is_Const(pred))
593 e = get_irn_ne(irn, env);
595 pscc->init = get_Const_tarval(have_init);
596 pscc->incr = get_Const_tarval(have_incr);
598 return code != iro_Bad;
599 } /* is_counter_iv */
602 * Check the users of an induction variable for register pressure.
604 * @param iv any node of the induction variable
605 * @param env the environment
607 * @return non-zero if the register pressure is estimated
608 * to not increase, zero else
610 static int check_users_for_reg_pressure(ir_node *iv, iv_env *env)
612 ir_node *irn, *header;
613 ir_node *have_user = NULL;
614 ir_node *have_cmp = NULL;
615 node_entry *e = get_irn_ne(iv, env);
619 for (irn = pscc->head; irn != NULL; irn = e->next) {
620 const ir_edge_t *edge;
622 foreach_out_edge(irn, edge) {
623 ir_node *user = get_edge_src_irn(edge);
624 node_entry *ne = get_irn_ne(user, env);
626 if (e->header == ne->header) {
627 /* found user from the same IV */
631 if (have_cmp != NULL) {
632 /* more than one cmp, for now end here */
637 /* user is a real user of the IV */
638 if (have_user != NULL) {
639 /* found the second user */
645 e = get_irn_ne(irn, env);
648 if (have_user == NULL) {
649 /* no user, ignore */
653 if (have_cmp == NULL) {
654 /* fine, only one user, try to reduce */
658 * We found one user AND at least one cmp.
659 * We should check here if we can transform the Cmp.
661 * For now our capabilities for doing linear function test
662 * are limited, so check if the iv has the right form: Only ONE
663 * Phi, only one Add/Sub with a Const.
665 if (! is_counter_iv(iv, env))
669 * Ok, we have only one increment AND it is a Const, we might be able
670 * to do a linear function test replacement, so go on.
673 } /* check_users_for_reg_pressure */
676 * Check if a node can be replaced (+, -, *).
678 * @param irn the node to check
679 * @param env the environment
681 * @return non-zero if irn should be Replace'd
683 static int check_replace(ir_node *irn, iv_env *env)
685 ir_node *left, *right, *iv, *rc;
686 ir_op *op = get_irn_op(irn);
687 ir_opcode code = get_op_code(op);
696 left = get_binop_left(irn);
697 right = get_binop_right(irn);
699 liv = is_iv(left, env);
700 riv = is_iv(right, env);
701 if (liv && is_rc(right, liv)) {
702 iv = left; rc = right;
704 else if (riv && is_op_commutative(op) &&
706 iv = right; rc = left;
710 if (env->osr_flags & osr_flag_keep_reg_pressure) {
711 if (! check_users_for_reg_pressure(iv, env))
714 return replace(irn, iv, rc, env);
721 } /* check_replace */
724 * Check which SCC's are induction variables.
727 * @param env the environment
729 static void classify_iv(scc *pscc, iv_env *env)
731 ir_node *irn, *next, *header = NULL;
732 node_entry *b, *h = NULL;
733 int j, only_phi, num_outside;
736 /* find the header block for this scc */
737 for (irn = pscc->head; irn; irn = next) {
738 node_entry *e = get_irn_link(irn);
739 ir_node *block = get_nodes_block(irn);
742 b = get_irn_ne(block, env);
745 if (h->POnum < b->POnum) {
756 /* check if this scc contains only Phi, Add or Sub nodes */
760 for (irn = pscc->head; irn; irn = next) {
761 node_entry *e = get_irn_ne(irn, env);
764 switch (get_irn_opcode(irn)) {
770 for (j = get_irn_arity(irn) - 1; j >= 0; --j) {
771 ir_node *pred = get_irn_n(irn, j);
772 node_entry *pe = get_irn_ne(pred, env);
774 if (pe->pscc != e->pscc) {
775 /* not in the same SCC, must be a region const */
776 if (! is_rc(pred, header)) {
777 /* not an induction variable */
783 } else if (out_rc != pred) {
790 /* not an induction variable */
794 /* found an induction variable */
795 DB((dbg, LEVEL_2, " Found an induction variable:\n "));
796 if (only_phi && num_outside == 1) {
797 /* a phi cycle with only one real predecessor can be collapsed */
798 DB((dbg, LEVEL_2, " Found an USELESS Phi cycle:\n "));
800 for (irn = pscc->head; irn; irn = next) {
801 node_entry *e = get_irn_ne(irn, env);
804 exchange(irn, out_rc);
810 /* set the header for every node in this scc */
811 for (irn = pscc->head; irn; irn = next) {
812 node_entry *e = get_irn_ne(irn, env);
815 DB((dbg, LEVEL_2, " %+F,", irn));
817 DB((dbg, LEVEL_2, "\n"));
821 for (irn = pscc->head; irn; irn = next) {
822 node_entry *e = get_irn_ne(irn, env);
830 * Process an SCC for the operator strength reduction.
832 * @param pscc the SCC
833 * @param env the environment
835 static void process_scc(scc *pscc, iv_env *env)
837 ir_node *head = pscc->head;
838 node_entry *e = get_irn_link(head);
844 DB((dbg, LEVEL_4, " SCC at %p:\n ", pscc));
845 for (irn = pscc->head; irn != NULL; irn = next) {
846 node_entry *e = get_irn_link(irn);
850 DB((dbg, LEVEL_4, " %+F,", irn));
852 DB((dbg, LEVEL_4, "\n"));
856 if (e->next == NULL) {
857 /* this SCC has only a single member */
858 check_replace(head, env);
860 classify_iv(pscc, env);
865 * If an SCC is a Phi only cycle, remove it.
867 * @param pscc an SCC that consists of Phi nodes only
868 * @param env the environment
870 static void remove_phi_cycle(scc *pscc, iv_env *env)
876 /* check if this scc contains only Phi, Add or Sub nodes */
878 for (irn = pscc->head; irn; irn = next) {
879 node_entry *e = get_irn_ne(irn, env);
885 for (j = get_irn_arity(irn) - 1; j >= 0; --j) {
886 ir_node *pred = get_irn_n(irn, j);
887 node_entry *pe = get_irn_ne(pred, env);
889 if (pe->pscc != e->pscc) {
890 /* not in the same SCC, must be the only input */
893 } else if (out_rc != pred) {
899 /* found a Phi cycle */
900 DB((dbg, LEVEL_2, " Found an USELESS Phi cycle:\n "));
902 for (irn = pscc->head; irn; irn = next) {
903 node_entry *e = get_irn_ne(irn, env);
906 exchange(irn, out_rc);
909 } /* remove_phi_cycle */
912 * Process a SCC for the Phi cycle remove.
914 * @param pscc the SCC
915 * @param env the environment
917 static void process_phi_only_scc(scc *pscc, iv_env *env)
919 ir_node *head = pscc->head;
920 node_entry *e = get_irn_link(head);
926 DB((dbg, LEVEL_4, " SCC at %p:\n ", pscc));
927 for (irn = pscc->head; irn; irn = next) {
928 node_entry *e = get_irn_link(irn);
932 DB((dbg, LEVEL_4, " %+F,", irn));
934 DB((dbg, LEVEL_4, "\n"));
939 remove_phi_cycle(pscc, env);
940 } /* process_phi_only_scc */
944 * Push a node onto the stack.
946 * @param env the environment
947 * @param n the node to push
949 static void push(iv_env *env, ir_node *n)
953 if (env->tos == ARR_LEN(env->stack)) {
954 int nlen = ARR_LEN(env->stack) * 2;
955 ARR_RESIZE(ir_node *, env->stack, nlen);
957 env->stack[env->tos++] = n;
958 e = get_irn_ne(n, env);
963 * Pop a node from the stack.
965 * @param env the environment
967 * @return The topmost node
969 static ir_node *pop(iv_env *env)
971 ir_node *n = env->stack[--env->tos];
972 node_entry *e = get_irn_ne(n, env);
979 * Do Tarjan's SCC algorithm and drive OSR.
981 * @param irn start at this node
982 * @param env the environment
984 static void dfs(ir_node *irn, iv_env *env)
987 node_entry *node = get_irn_ne(irn, env);
989 mark_irn_visited(irn);
991 /* do not put blocks into the scc */
993 n = get_irn_arity(irn);
994 for (i = 0; i < n; ++i) {
995 ir_node *pred = get_irn_n(irn, i);
997 if (!irn_visited(pred))
1001 ir_node *block = get_nodes_block(irn);
1003 node->DFSnum = env->nextDFSnum++;
1004 node->low = node->DFSnum;
1007 /* handle the block */
1008 if (!irn_visited(block))
1011 n = get_irn_arity(irn);
1012 for (i = 0; i < n; ++i) {
1013 ir_node *pred = get_irn_n(irn, i);
1014 node_entry *o = get_irn_ne(pred, env);
1016 if (!irn_visited(pred)) {
1018 node->low = MIN(node->low, o->low);
1020 if (o->DFSnum < node->DFSnum && o->in_stack)
1021 node->low = MIN(o->DFSnum, node->low);
1023 if (node->low == node->DFSnum) {
1024 scc *pscc = OALLOCZ(&env->obst, scc);
1031 e = get_irn_ne(x, env);
1033 e->next = pscc->head;
1037 env->process_scc(pscc, env);
1043 * Do the DFS by starting at the End node of a graph.
1045 * @param irg the graph to process
1046 * @param env the environment
1048 static void do_dfs(ir_graph *irg, iv_env *env)
1050 ir_graph *rem = current_ir_graph;
1051 ir_node *end = get_irg_end(irg);
1054 ir_reserve_resources(irg, IR_RESOURCE_IRN_VISITED);
1056 current_ir_graph = irg;
1057 inc_irg_visited(irg);
1059 /* visit all visible nodes */
1062 /* visit the keep-alives */
1063 for (i = get_End_n_keepalives(end) - 1; i >= 0; --i) {
1064 ir_node *ka = get_End_keepalive(end, i);
1066 if (!irn_visited(ka))
1070 ir_free_resources(irg, IR_RESOURCE_IRN_VISITED);
1072 current_ir_graph = rem;
1076 * Post-block-walker: assign the post-order number.
1078 static void assign_po(ir_node *block, void *ctx)
1081 node_entry *e = get_irn_ne(block, env);
1083 e->POnum = env->POnum++;
1087 * Apply one LFTR edge operation.
1088 * Return NULL if the transformation cannot be done safely without
1091 * @param iv the induction variable
1092 * @param rc the constant that should be translated
1093 * @param e the LFTR edge
1094 * @param env the IV environment
1096 * @return the translated region constant or NULL
1097 * if the translation was not possible
1100 * In the current implementation only the last edge is stored, so
1101 * only one chain exists. That's why we might miss some opportunities.
1103 static ir_node *applyOneEdge(ir_node *iv, ir_node *rc, LFTR_edge *e, iv_env *env)
1105 if (env->osr_flags & osr_flag_lftr_with_ov_check) {
1106 tarval *tv_l, *tv_r, *tv, *tv_init, *tv_incr, *tv_end;
1107 tarval_int_overflow_mode_t ovmode;
1110 if (! is_counter_iv(iv, env)) {
1111 DB((dbg, LEVEL_4, " not counter IV"));
1115 /* overflow can only be decided for Consts */
1116 if (! is_Const(e->rc)) {
1117 if (e->code == iro_Add && mode_is_reference(get_irn_mode(e->rc))) {
1118 /* However we allow ONE Pointer Add, as pointer arithmetic with wrap
1119 around is undefined anyway */
1120 return do_apply(e->code, NULL, rc, e->rc, get_irn_mode(e->rc));
1122 DB((dbg, LEVEL_4, " = UNKNOWN (%+F)", e->rc));
1126 tv_l = get_Const_tarval(rc);
1127 tv_r = get_Const_tarval(e->rc);
1129 ovmode = tarval_get_integer_overflow_mode();
1130 tarval_set_integer_overflow_mode(TV_OVERFLOW_BAD);
1132 pscc = get_iv_scc(iv, env);
1133 tv_incr = pscc->incr;
1134 tv_init = pscc->init;
1137 * Check that no overflow occurs:
1138 * init must be transformed without overflow
1139 * the new rc must be transformed without overflow
1140 * rc +/- incr must be possible without overflow
1144 tv = tarval_mul(tv_l, tv_r);
1145 tv_init = tarval_mul(tv_init, tv_r);
1146 tv_incr = tarval_mul(tv_incr, tv_r);
1147 DB((dbg, LEVEL_4, " * %+F", tv_r));
1150 tv = tarval_add(tv_l, tv_r);
1151 tv_init = tarval_add(tv_init, tv_r);
1152 DB((dbg, LEVEL_4, " + %+F", tv_r));
1155 tv = tarval_sub(tv_l, tv_r, NULL);
1156 tv_init = tarval_sub(tv_init, tv_r, NULL);
1157 DB((dbg, LEVEL_4, " - %+F", tv_r));
1160 panic("Unsupported opcode");
1163 if (pscc->code == iro_Add) {
1164 tv_end = tarval_add(tv, tv_incr);
1166 assert(pscc->code == iro_Sub);
1167 tv_end = tarval_sub(tv, tv_incr, NULL);
1170 tarval_set_integer_overflow_mode(ovmode);
1172 if (tv == tarval_bad || tv_init == tarval_bad || tv_end == tarval_bad) {
1173 DB((dbg, LEVEL_4, " = OVERFLOW"));
1176 return new_Const(tv);
1178 return do_apply(e->code, NULL, rc, e->rc, get_irn_mode(e->dst));
1179 } /* applyOneEdge */
1182 * Applies the operations represented by the LFTR edges to a
1183 * region constant and returns the value.
1184 * Return NULL if the transformation cannot be done safely without
1187 * @param pIV points to the IV node that starts the LFTR edge chain
1188 * after translation points to the new IV
1189 * @param rc the region constant that should be translated
1190 * @param env the IV environment
1192 * @return the translated region constant or NULL
1193 * if the translation was not possible
1195 static ir_node *applyEdges(ir_node **pIV, ir_node *rc, iv_env *env)
1198 if (env->osr_flags & osr_flag_lftr_with_ov_check) {
1199 /* overflow can only be decided for Consts */
1200 if (! is_counter_iv(iv, env)) {
1201 DB((dbg, LEVEL_4, "not counter IV\n", rc));
1204 if (! is_Const(rc)) {
1205 DB((dbg, LEVEL_4, " = UNKNOWN (%+F)\n", rc));
1208 DB((dbg, LEVEL_4, "%+F", get_Const_tarval(rc)));
1212 LFTR_edge *e = LFTR_find(iv, env);
1214 rc = applyOneEdge(iv, rc, e, env);
1219 DB((dbg, LEVEL_3, "\n"));
1225 * Walker, finds Cmp(iv, rc) or Cmp(rc, iv)
1226 * and tries to optimize them.
1228 static void do_lftr(ir_node *cmp, void *ctx)
1231 ir_node *left, *right, *liv, *riv;
1233 ir_node *nleft = NULL, *nright = NULL;
1238 left = get_Cmp_left(cmp);
1239 right = get_Cmp_right(cmp);
1241 liv = is_iv(left, env);
1242 riv = is_iv(right, env);
1243 if (liv && is_rc(right, liv)) {
1244 iv = left; rc = right;
1246 nright = applyEdges(&iv, rc, env);
1249 else if (riv && is_rc(left, riv)) {
1250 iv = right; rc = left;
1252 nleft = applyEdges(&iv, rc, env);
1256 if (nleft && nright) {
1257 DB((dbg, LEVEL_2, " LFTR for %+F\n", cmp));
1258 set_Cmp_left(cmp, nleft);
1259 set_Cmp_right(cmp, nright);
1260 ++env->lftr_replaced;
1265 * do linear function test replacement.
1267 * @param irg the graph that should be optimized
1268 * @param env the IV environment
1270 static void lftr(ir_graph *irg, iv_env *env)
1272 irg_walk_graph(irg, NULL, do_lftr, env);
1276 * Pre-walker: set all node links to NULL and fix the
1277 * block of Proj nodes.
1279 static void clear_and_fix(ir_node *irn, void *env)
1282 set_irn_link(irn, NULL);
1285 ir_node *pred = get_Proj_pred(irn);
1286 ir_node *pred_block = get_nodes_block(pred);
1288 if (get_nodes_block(irn) != pred_block) {
1289 set_nodes_block(irn, pred_block);
1293 } /* clear_and_fix */
1296 /* Remove any Phi cycles with only one real input. */
1297 void remove_phi_cycles(ir_graph *irg)
1303 rem = current_ir_graph;
1304 current_ir_graph = irg;
1306 FIRM_DBG_REGISTER(dbg, "firm.opt.remove_phi");
1308 DB((dbg, LEVEL_1, "Doing Phi cycle removement for %+F\n", irg));
1310 obstack_init(&env.obst);
1311 env.stack = NEW_ARR_F(ir_node *, 128);
1315 env.quad_map = NULL;
1316 env.lftr_edges = NULL;
1318 env.lftr_replaced = 0;
1320 env.need_postpass = 0;
1321 env.process_scc = process_phi_only_scc;
1323 /* Clear all links and move Proj nodes into the
1324 * the same block as their predecessors.
1325 * This can improve the placement of new nodes.
1328 irg_walk_graph(irg, NULL, clear_and_fix, &projs_moved);
1330 set_irg_outs_inconsistent(irg);
1332 /* we need outs for calculating the post order */
1333 assure_irg_outs(irg);
1335 /* calculate the post order number for blocks. */
1336 irg_out_block_walk(get_irg_start_block(irg), NULL, assign_po, &env);
1338 /* calculate the SCC's and drive OSR. */
1339 ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK);
1341 ir_free_resources(irg, IR_RESOURCE_IRN_LINK);
1344 set_irg_outs_inconsistent(irg);
1345 DB((dbg, LEVEL_1, "remove_phi_cycles: %u Cycles removed\n\n", env.replaced));
1348 DEL_ARR_F(env.stack);
1349 obstack_free(&env.obst, NULL);
1351 current_ir_graph = rem;
1352 } /* remove_phi_cycles */
1354 ir_graph_pass_t *remove_phi_cycles_pass(const char *name)
1356 return def_graph_pass(name ? name : "remove_phi_cycles", remove_phi_cycles);
1357 } /* remove_phi_cycles_pass */
1360 * Post-walker: fix Add and Sub nodes that where results of I<->P conversions.
1362 static void fix_adds_and_subs(ir_node *irn, void *ctx)
1367 ir_mode *mode = get_irn_mode(irn);
1369 if (mode_is_int(mode)) {
1372 pred = get_Add_left(irn);
1373 if (get_irn_mode(pred) != mode) {
1374 ir_node *block = get_nodes_block(pred);
1376 pred = new_r_Conv(block, pred, mode);
1377 set_Add_left(irn, pred);
1379 pred = get_Add_right(irn);
1380 if (get_irn_mode(pred) != mode) {
1381 ir_node *block = get_nodes_block(pred);
1383 pred = new_r_Conv(block, pred, mode);
1384 set_Add_right(irn, pred);
1387 } else if (is_Sub(irn)) {
1388 ir_mode *mode = get_irn_mode(irn);
1390 if (mode_is_int(mode)) {
1391 ir_node *left = get_Sub_left(irn);
1392 ir_node *right = get_Sub_right(irn);
1393 ir_mode *l_mode = get_irn_mode(left);
1394 ir_mode *r_mode = get_irn_mode(right);
1396 if (mode_is_int(l_mode) && mode_is_int(r_mode)) {
1397 if (l_mode != mode) {
1398 ir_node *block = get_nodes_block(left);
1400 left = new_r_Conv(block, left, mode);
1401 set_Sub_left(irn, left);
1403 if (r_mode != mode) {
1404 ir_node *block = get_nodes_block(right);
1406 right = new_r_Conv(block, right, mode);
1407 set_Sub_right(irn, right);
1412 } /* fix_adds_and_subs */
1414 /* Performs Operator Strength Reduction for the passed graph. */
1415 void opt_osr(ir_graph *irg, unsigned flags)
1422 rem = current_ir_graph;
1423 current_ir_graph = irg;
1425 FIRM_DBG_REGISTER(dbg, "firm.opt.osr");
1427 DB((dbg, LEVEL_1, "Doing Operator Strength Reduction for %+F\n", irg));
1429 obstack_init(&env.obst);
1430 env.stack = NEW_ARR_F(ir_node *, 128);
1434 env.quad_map = new_set(quad_cmp, 64);
1435 env.lftr_edges = new_set(LFTR_cmp, 64);
1437 env.lftr_replaced = 0;
1438 env.osr_flags = flags;
1439 env.need_postpass = 0;
1440 env.process_scc = process_scc;
1442 /* Clear all links and move Proj nodes into the
1443 the same block as it's predecessors.
1444 This can improve the placement of new nodes.
1447 irg_walk_graph(irg, NULL, clear_and_fix, &projs_moved);
1449 set_irg_outs_inconsistent(irg);
1451 /* we need dominance */
1454 edges = edges_assure(irg);
1456 /* calculate the post order number for blocks by walking the out edges. */
1457 assure_irg_outs(irg);
1458 irg_block_edges_walk(get_irg_start_block(irg), NULL, assign_po, &env);
1460 /* calculate the SCC's and drive OSR. */
1461 ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK);
1465 if (env.need_postpass)
1466 irg_walk_graph(irg, NULL, fix_adds_and_subs, &env);
1468 /* try linear function test replacements */
1472 set_irg_outs_inconsistent(irg);
1473 DB((dbg, LEVEL_1, "Replacements: %u + %u (lftr)\n\n", env.replaced, env.lftr_replaced));
1475 ir_free_resources(irg, IR_RESOURCE_IRN_LINK);
1477 del_set(env.lftr_edges);
1478 del_set(env.quad_map);
1479 DEL_ARR_F(env.stack);
1480 obstack_free(&env.obst, NULL);
1483 edges_deactivate(irg);
1485 current_ir_graph = rem;
1489 ir_graph_pass_t pass;
1494 * Wrapper for running opt_osr() as an ir_graph pass.
1496 static int pass_wrapper(ir_graph *irg, void *context)
1498 struct pass_t *pass = context;
1499 opt_osr(irg, pass->flags);
1501 } /* pass_wrapper */
1503 ir_graph_pass_t *opt_osr_pass(const char *name, unsigned flags)
1505 struct pass_t *pass = XMALLOCZ(struct pass_t);
1507 pass->flags = flags;
1508 return def_graph_pass_constructor(
1509 &pass->pass, name ? name : "osr", pass_wrapper);
1510 } /* opt_osr_pass */