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");
339 * The Apply operation.
341 * @param orig the node that represent the original operation and determines
342 * the opcode, debug-info and mode of a newly created one
343 * @param op1 the first operand
344 * @param op2 the second operand
345 * @param env the environment
347 * @return the newly created node
349 static ir_node *apply(ir_node *header, ir_node *orig, ir_node *op1, ir_node *op2, iv_env *env)
351 ir_opcode code = get_irn_opcode(orig);
352 ir_node *result = search(code, op1, op2, env);
354 if (result == NULL) {
355 dbg_info *db = get_irn_dbg_info(orig);
356 ir_node *op1_header = get_irn_ne(op1, env)->header;
357 ir_node *op2_header = get_irn_ne(op2, env)->header;
359 if (op1_header == header && is_rc(op2, op1_header)) {
360 result = reduce(orig, op1, op2, env);
362 else if (op2_header == header && is_rc(op1, op2_header)) {
363 result = reduce(orig, op2, op1, env);
366 result = do_apply(code, db, op1, op2, get_irn_mode(orig));
367 get_irn_ne(result, env)->header = NULL;
374 * The Reduce operation.
376 * @param orig the node that represent the original operation and determines
377 * the opcode, debug-info and mode of a newly created one
378 * @param iv the induction variable
379 * @param rc the region constant
380 * @param env the environment
382 * @return the reduced node
384 static ir_node *reduce(ir_node *orig, ir_node *iv, ir_node *rc, iv_env *env)
386 ir_opcode code = get_irn_opcode(orig);
387 ir_node *result = search(code, iv, rc, env);
389 /* check if we have already done this operation on the iv */
390 if (result == NULL) {
391 node_entry *e, *iv_e;
393 ir_mode *mode = get_irn_mode(orig);
395 result = exact_copy(iv);
397 if (get_irn_mode(result) != mode) {
399 * Beware: we must always create a new induction variable with the same mode
400 * as the node we are replacing. Especially this means the mode might be changed
401 * from P to I and back. This is always possible, because we have only Phi, Add
403 * However, this might lead to AddIs(Iu,Is) which we must fix. The best way to do this
404 * seems to be a post-pass, or we might end with useless Conv's.
406 set_irn_mode(result, mode);
407 env->need_postpass = 1;
409 add(code, iv, rc, result, env);
410 DB((dbg, LEVEL_3, " Created new %+F for %+F (%s %+F)\n", result, iv,
411 get_irn_opname(orig), rc));
413 iv_e = get_irn_ne(iv, env);
414 e = get_irn_ne(result, env);
415 e->header = iv_e->header;
417 /* create the LFTR edge */
418 LFTR_add(iv, result, code, rc, env);
420 for (i = get_irn_arity(result) - 1; i >= 0; --i) {
421 ir_node *o = get_irn_n(result, i);
423 e = get_irn_ne(o, env);
424 if (e->header == iv_e->header)
425 o = reduce(orig, o, rc, env);
426 else if (is_Phi(result) || code == iro_Mul)
427 o = apply(iv_e->header, orig, o, rc, env);
428 set_irn_n(result, i, o);
431 DB((dbg, LEVEL_3, " Already Created %+F for %+F (%s %+F)\n", result, iv,
432 get_irn_opname(orig), rc));
438 * Update the scc for a newly created IV.
440 static void update_scc(ir_node *iv, node_entry *e, iv_env *env)
443 ir_node *header = e->header;
444 waitq *wq = new_waitq();
446 DB((dbg, LEVEL_2, " Creating SCC for new an induction variable:\n "));
450 ir_node *irn = waitq_get(wq);
451 node_entry *ne = get_irn_ne(irn, env);
455 ne->next = pscc->head;
457 DB((dbg, LEVEL_2, " %+F,", irn));
459 for (i = get_irn_arity(irn) - 1; i >= 0; --i) {
460 ir_node *pred = get_irn_n(irn, i);
461 node_entry *pe = get_irn_ne(pred, env);
463 if (pe->header == header && pe->pscc == NULL) {
464 /* set the pscc here to ensure that the node is NOT enqueued another time */
469 } while (! waitq_empty(wq));
471 DB((dbg, LEVEL_2, "\n"));
475 * The Replace operation. We found a node representing iv (+,-,*) rc
476 * that can be removed by replacing the induction variable iv by a new
477 * one that 'applies' the operation 'irn'.
479 * @param irn the node that will be replaced
480 * @param iv the induction variable
481 * @param rc the region constant
482 * @param env the environment
484 static int replace(ir_node *irn, ir_node *iv, ir_node *rc, iv_env *env)
488 DB((dbg, LEVEL_2, " Replacing %+F\n", irn));
490 result = reduce(irn, iv, rc, env);
494 hook_strength_red(current_ir_graph, irn);
495 exchange(irn, result);
496 e = get_irn_ne(result, env);
497 if (e->pscc == NULL) {
498 e->pscc = OALLOCZ(&env->obst, scc);
499 update_scc(result, e, env);
509 * check if a given node is a mul with 2, 4, 8
511 static int is_x86_shift_const(ir_node *mul)
518 /* normalization put constants on the right side */
519 rc = get_Mul_right(mul);
521 tarval *tv = get_Const_tarval(rc);
523 if (tarval_is_long(tv)) {
524 long value = get_tarval_long(tv);
526 if (value == 2 || value == 4 || value == 8) {
527 /* do not reduce multiplications by 2, 4, 8 */
533 } /* is_x86_shift_const */
537 * Check if an IV represents a counter with constant limits.
539 * @param iv any node of the induction variable
540 * @param env the environment
542 static int is_counter_iv(ir_node *iv, iv_env *env)
544 node_entry *e = get_irn_ne(iv, env);
546 ir_node *have_init = NULL;
547 ir_node *have_incr = NULL;
548 ir_opcode code = iro_Bad;
551 if (pscc->code != 0) {
552 /* already analysed */
553 return pscc->code != iro_Bad;
556 pscc->code = iro_Bad;
557 for (irn = pscc->head; irn != NULL; irn = e->next) {
559 if (have_incr != NULL)
562 have_incr = get_Add_right(irn);
563 if (! is_Const(have_incr)) {
564 have_incr = get_Add_left(irn);
565 if (! is_Const(have_incr))
569 } else if (is_Sub(irn)) {
570 if (have_incr != NULL)
573 have_incr = get_Sub_right(irn);
574 if (! is_Const(have_incr))
577 } else if (is_Phi(irn)) {
580 for (i = get_Phi_n_preds(irn) - 1; i >= 0; --i) {
581 ir_node *pred = get_Phi_pred(irn, i);
582 node_entry *ne = get_irn_ne(pred, env);
584 if (ne->header == e->header)
586 if (have_init != NULL)
589 if (! is_Const(pred))
594 e = get_irn_ne(irn, env);
596 pscc->init = get_Const_tarval(have_init);
597 pscc->incr = get_Const_tarval(have_incr);
599 return code != iro_Bad;
600 } /* is_counter_iv */
603 * Check the users of an induction variable for register pressure.
605 * @param iv any node of the induction variable
606 * @param env the environment
608 * @return non-zero if the register pressure is estimated
609 * to not increase, zero else
611 static int check_users_for_reg_pressure(ir_node *iv, iv_env *env)
613 ir_node *irn, *header;
614 ir_node *have_user = NULL;
615 ir_node *have_cmp = NULL;
616 node_entry *e = get_irn_ne(iv, env);
620 for (irn = pscc->head; irn != NULL; irn = e->next) {
621 const ir_edge_t *edge;
623 foreach_out_edge(irn, edge) {
624 ir_node *user = get_edge_src_irn(edge);
625 node_entry *ne = get_irn_ne(user, env);
627 if (e->header == ne->header) {
628 /* found user from the same IV */
632 if (have_cmp != NULL) {
633 /* more than one cmp, for now end here */
638 /* user is a real user of the IV */
639 if (have_user != NULL) {
640 /* found the second user */
646 e = get_irn_ne(irn, env);
649 if (have_user == NULL) {
650 /* no user, ignore */
654 if (have_cmp == NULL) {
655 /* fine, only one user, try to reduce */
659 * We found one user AND at least one cmp.
660 * We should check here if we can transform the Cmp.
662 * For now our capabilities for doing linear function test
663 * are limited, so check if the iv has the right form: Only ONE
664 * Phi, only one Add/Sub with a Const.
666 if (! is_counter_iv(iv, env))
670 * Ok, we have only one increment AND it is a Const, we might be able
671 * to do a linear function test replacement, so go on.
674 } /* check_users_for_reg_pressure */
677 * Check if a node can be replaced (+, -, *).
679 * @param irn the node to check
680 * @param env the environment
682 * @return non-zero if irn should be Replace'd
684 static int check_replace(ir_node *irn, iv_env *env)
686 ir_node *left, *right, *iv, *rc;
687 ir_op *op = get_irn_op(irn);
688 ir_opcode code = get_op_code(op);
697 left = get_binop_left(irn);
698 right = get_binop_right(irn);
700 liv = is_iv(left, env);
701 riv = is_iv(right, env);
702 if (liv && is_rc(right, liv)) {
703 iv = left; rc = right;
705 else if (riv && is_op_commutative(op) &&
707 iv = right; rc = left;
711 if (env->osr_flags & osr_flag_keep_reg_pressure) {
712 if (! check_users_for_reg_pressure(iv, env))
715 return replace(irn, iv, rc, env);
722 } /* check_replace */
725 * Check which SCC's are induction variables.
728 * @param env the environment
730 static void classify_iv(scc *pscc, iv_env *env)
732 ir_node *irn, *next, *header = NULL;
733 node_entry *b, *h = NULL;
734 int j, only_phi, num_outside;
737 /* find the header block for this scc */
738 for (irn = pscc->head; irn; irn = next) {
739 node_entry *e = get_irn_link(irn);
740 ir_node *block = get_nodes_block(irn);
743 b = get_irn_ne(block, env);
746 if (h->POnum < b->POnum) {
757 /* check if this scc contains only Phi, Add or Sub nodes */
761 for (irn = pscc->head; irn; irn = next) {
762 node_entry *e = get_irn_ne(irn, env);
765 switch (get_irn_opcode(irn)) {
771 for (j = get_irn_arity(irn) - 1; j >= 0; --j) {
772 ir_node *pred = get_irn_n(irn, j);
773 node_entry *pe = get_irn_ne(pred, env);
775 if (pe->pscc != e->pscc) {
776 /* not in the same SCC, must be a region const */
777 if (! is_rc(pred, header)) {
778 /* not an induction variable */
784 } else if (out_rc != pred) {
791 /* not an induction variable */
795 /* found an induction variable */
796 DB((dbg, LEVEL_2, " Found an induction variable:\n "));
797 if (only_phi && num_outside == 1) {
798 /* a phi cycle with only one real predecessor can be collapsed */
799 DB((dbg, LEVEL_2, " Found an USELESS Phi cycle:\n "));
801 for (irn = pscc->head; irn; irn = next) {
802 node_entry *e = get_irn_ne(irn, env);
805 exchange(irn, out_rc);
811 /* set the header for every node in this scc */
812 for (irn = pscc->head; irn; irn = next) {
813 node_entry *e = get_irn_ne(irn, env);
816 DB((dbg, LEVEL_2, " %+F,", irn));
818 DB((dbg, LEVEL_2, "\n"));
822 for (irn = pscc->head; irn; irn = next) {
823 node_entry *e = get_irn_ne(irn, env);
831 * Process an SCC for the operator strength reduction.
833 * @param pscc the SCC
834 * @param env the environment
836 static void process_scc(scc *pscc, iv_env *env)
838 ir_node *head = pscc->head;
839 node_entry *e = get_irn_link(head);
845 DB((dbg, LEVEL_4, " SCC at %p:\n ", pscc));
846 for (irn = pscc->head; irn != NULL; irn = next) {
847 node_entry *e = get_irn_link(irn);
851 DB((dbg, LEVEL_4, " %+F,", irn));
853 DB((dbg, LEVEL_4, "\n"));
857 if (e->next == NULL) {
858 /* this SCC has only a single member */
859 check_replace(head, env);
861 classify_iv(pscc, env);
866 * If an SCC is a Phi only cycle, remove it.
868 * @param pscc an SCC that consists of Phi nodes only
869 * @param env the environment
871 static void remove_phi_cycle(scc *pscc, iv_env *env)
877 /* check if this scc contains only Phi, Add or Sub nodes */
879 for (irn = pscc->head; irn; irn = next) {
880 node_entry *e = get_irn_ne(irn, env);
886 for (j = get_irn_arity(irn) - 1; j >= 0; --j) {
887 ir_node *pred = get_irn_n(irn, j);
888 node_entry *pe = get_irn_ne(pred, env);
890 if (pe->pscc != e->pscc) {
891 /* not in the same SCC, must be the only input */
894 } else if (out_rc != pred) {
900 /* found a Phi cycle */
901 DB((dbg, LEVEL_2, " Found an USELESS Phi cycle:\n "));
903 for (irn = pscc->head; irn; irn = next) {
904 node_entry *e = get_irn_ne(irn, env);
907 exchange(irn, out_rc);
910 } /* remove_phi_cycle */
913 * Process a SCC for the Phi cycle remove.
915 * @param pscc the SCC
916 * @param env the environment
918 static void process_phi_only_scc(scc *pscc, iv_env *env)
920 ir_node *head = pscc->head;
921 node_entry *e = get_irn_link(head);
927 DB((dbg, LEVEL_4, " SCC at %p:\n ", pscc));
928 for (irn = pscc->head; irn; irn = next) {
929 node_entry *e = get_irn_link(irn);
933 DB((dbg, LEVEL_4, " %+F,", irn));
935 DB((dbg, LEVEL_4, "\n"));
940 remove_phi_cycle(pscc, env);
941 } /* process_phi_only_scc */
945 * Push a node onto the stack.
947 * @param env the environment
948 * @param n the node to push
950 static void push(iv_env *env, ir_node *n)
954 if (env->tos == ARR_LEN(env->stack)) {
955 int nlen = ARR_LEN(env->stack) * 2;
956 ARR_RESIZE(ir_node *, env->stack, nlen);
958 env->stack[env->tos++] = n;
959 e = get_irn_ne(n, env);
964 * Pop a node from the stack.
966 * @param env the environment
968 * @return The topmost node
970 static ir_node *pop(iv_env *env)
972 ir_node *n = env->stack[--env->tos];
973 node_entry *e = get_irn_ne(n, env);
980 * Do Tarjan's SCC algorithm and drive OSR.
982 * @param irn start at this node
983 * @param env the environment
985 static void dfs(ir_node *irn, iv_env *env)
988 node_entry *node = get_irn_ne(irn, env);
990 mark_irn_visited(irn);
992 /* do not put blocks into the scc */
994 n = get_irn_arity(irn);
995 for (i = 0; i < n; ++i) {
996 ir_node *pred = get_irn_n(irn, i);
998 if (!irn_visited(pred))
1002 ir_node *block = get_nodes_block(irn);
1004 node->DFSnum = env->nextDFSnum++;
1005 node->low = node->DFSnum;
1008 /* handle the block */
1009 if (!irn_visited(block))
1012 n = get_irn_arity(irn);
1013 for (i = 0; i < n; ++i) {
1014 ir_node *pred = get_irn_n(irn, i);
1015 node_entry *o = get_irn_ne(pred, env);
1017 if (!irn_visited(pred)) {
1019 node->low = MIN(node->low, o->low);
1021 if (o->DFSnum < node->DFSnum && o->in_stack)
1022 node->low = MIN(o->DFSnum, node->low);
1024 if (node->low == node->DFSnum) {
1025 scc *pscc = OALLOCZ(&env->obst, scc);
1032 e = get_irn_ne(x, env);
1034 e->next = pscc->head;
1038 env->process_scc(pscc, env);
1044 * Do the DFS by starting at the End node of a graph.
1046 * @param irg the graph to process
1047 * @param env the environment
1049 static void do_dfs(ir_graph *irg, iv_env *env)
1051 ir_graph *rem = current_ir_graph;
1052 ir_node *end = get_irg_end(irg);
1055 ir_reserve_resources(irg, IR_RESOURCE_IRN_VISITED);
1057 current_ir_graph = irg;
1058 inc_irg_visited(irg);
1060 /* visit all visible nodes */
1063 /* visit the keep-alives */
1064 for (i = get_End_n_keepalives(end) - 1; i >= 0; --i) {
1065 ir_node *ka = get_End_keepalive(end, i);
1067 if (!irn_visited(ka))
1071 ir_free_resources(irg, IR_RESOURCE_IRN_VISITED);
1073 current_ir_graph = rem;
1077 * Post-block-walker: assign the post-order number.
1079 static void assign_po(ir_node *block, void *ctx)
1082 node_entry *e = get_irn_ne(block, env);
1084 e->POnum = env->POnum++;
1088 * Apply one LFTR edge operation.
1089 * Return NULL if the transformation cannot be done safely without
1092 * @param iv the induction variable
1093 * @param rc the constant that should be translated
1094 * @param e the LFTR edge
1095 * @param env the IV environment
1097 * @return the translated region constant or NULL
1098 * if the translation was not possible
1101 * In the current implementation only the last edge is stored, so
1102 * only one chain exists. That's why we might miss some opportunities.
1104 static ir_node *applyOneEdge(ir_node *iv, ir_node *rc, LFTR_edge *e, iv_env *env)
1106 if (env->osr_flags & osr_flag_lftr_with_ov_check) {
1107 tarval *tv_l, *tv_r, *tv, *tv_init, *tv_incr, *tv_end;
1108 tarval_int_overflow_mode_t ovmode;
1111 if (! is_counter_iv(iv, env)) {
1112 DB((dbg, LEVEL_4, " not counter IV"));
1116 /* overflow can only be decided for Consts */
1117 if (! is_Const(e->rc)) {
1118 if (e->code == iro_Add && mode_is_reference(get_irn_mode(e->rc))) {
1119 /* However we allow ONE Pointer Add, as pointer arithmetic with wrap
1120 around is undefined anyway */
1121 return do_apply(e->code, NULL, rc, e->rc, get_irn_mode(e->rc));
1123 DB((dbg, LEVEL_4, " = UNKNOWN (%+F)", e->rc));
1127 tv_l = get_Const_tarval(rc);
1128 tv_r = get_Const_tarval(e->rc);
1130 ovmode = tarval_get_integer_overflow_mode();
1131 tarval_set_integer_overflow_mode(TV_OVERFLOW_BAD);
1133 pscc = get_iv_scc(iv, env);
1134 tv_incr = pscc->incr;
1135 tv_init = pscc->init;
1138 * Check that no overflow occurs:
1139 * init must be transformed without overflow
1140 * the new rc must be transformed without overflow
1141 * rc +/- incr must be possible without overflow
1145 tv = tarval_mul(tv_l, tv_r);
1146 tv_init = tarval_mul(tv_init, tv_r);
1147 tv_incr = tarval_mul(tv_incr, tv_r);
1148 DB((dbg, LEVEL_4, " * %+F", tv_r));
1151 tv = tarval_add(tv_l, tv_r);
1152 tv_init = tarval_add(tv_init, tv_r);
1153 DB((dbg, LEVEL_4, " + %+F", tv_r));
1156 tv = tarval_sub(tv_l, tv_r, NULL);
1157 tv_init = tarval_sub(tv_init, tv_r, NULL);
1158 DB((dbg, LEVEL_4, " - %+F", tv_r));
1161 panic("Unsupported opcode");
1165 if (pscc->code == iro_Add) {
1166 tv_end = tarval_add(tv, tv_incr);
1168 assert(pscc->code == iro_Sub);
1169 tv_end = tarval_sub(tv, tv_incr, NULL);
1172 tarval_set_integer_overflow_mode(ovmode);
1174 if (tv == tarval_bad || tv_init == tarval_bad || tv_end == tarval_bad) {
1175 DB((dbg, LEVEL_4, " = OVERFLOW"));
1178 return new_Const(tv);
1180 return do_apply(e->code, NULL, rc, e->rc, get_irn_mode(e->dst));
1181 } /* applyOneEdge */
1184 * Applies the operations represented by the LFTR edges to a
1185 * region constant and returns the value.
1186 * Return NULL if the transformation cannot be done safely without
1189 * @param pIV points to the IV node that starts the LFTR edge chain
1190 * after translation points to the new IV
1191 * @param rc the region constant that should be translated
1192 * @param env the IV environment
1194 * @return the translated region constant or NULL
1195 * if the translation was not possible
1197 static ir_node *applyEdges(ir_node **pIV, ir_node *rc, iv_env *env)
1200 if (env->osr_flags & osr_flag_lftr_with_ov_check) {
1201 /* overflow can only be decided for Consts */
1202 if (! is_counter_iv(iv, env)) {
1203 DB((dbg, LEVEL_4, "not counter IV\n", rc));
1206 if (! is_Const(rc)) {
1207 DB((dbg, LEVEL_4, " = UNKNOWN (%+F)\n", rc));
1210 DB((dbg, LEVEL_4, "%+F", get_Const_tarval(rc)));
1214 LFTR_edge *e = LFTR_find(iv, env);
1216 rc = applyOneEdge(iv, rc, e, env);
1221 DB((dbg, LEVEL_3, "\n"));
1227 * Walker, finds Cmp(iv, rc) or Cmp(rc, iv)
1228 * and tries to optimize them.
1230 static void do_lftr(ir_node *cmp, void *ctx)
1233 ir_node *left, *right, *liv, *riv;
1235 ir_node *nleft = NULL, *nright = NULL;
1240 left = get_Cmp_left(cmp);
1241 right = get_Cmp_right(cmp);
1243 liv = is_iv(left, env);
1244 riv = is_iv(right, env);
1245 if (liv && is_rc(right, liv)) {
1246 iv = left; rc = right;
1248 nright = applyEdges(&iv, rc, env);
1251 else if (riv && is_rc(left, riv)) {
1252 iv = right; rc = left;
1254 nleft = applyEdges(&iv, rc, env);
1258 if (nleft && nright) {
1259 DB((dbg, LEVEL_2, " LFTR for %+F\n", cmp));
1260 set_Cmp_left(cmp, nleft);
1261 set_Cmp_right(cmp, nright);
1262 ++env->lftr_replaced;
1267 * do linear function test replacement.
1269 * @param irg the graph that should be optimized
1270 * @param env the IV environment
1272 static void lftr(ir_graph *irg, iv_env *env)
1274 irg_walk_graph(irg, NULL, do_lftr, env);
1278 * Pre-walker: set all node links to NULL and fix the
1279 * block of Proj nodes.
1281 static void clear_and_fix(ir_node *irn, void *env)
1284 set_irn_link(irn, NULL);
1287 ir_node *pred = get_Proj_pred(irn);
1288 ir_node *pred_block = get_nodes_block(pred);
1290 if (get_nodes_block(irn) != pred_block) {
1291 set_nodes_block(irn, pred_block);
1295 } /* clear_and_fix */
1298 /* Remove any Phi cycles with only one real input. */
1299 void remove_phi_cycles(ir_graph *irg)
1305 rem = current_ir_graph;
1306 current_ir_graph = irg;
1308 FIRM_DBG_REGISTER(dbg, "firm.opt.remove_phi");
1310 DB((dbg, LEVEL_1, "Doing Phi cycle removement for %+F\n", irg));
1312 obstack_init(&env.obst);
1313 env.stack = NEW_ARR_F(ir_node *, 128);
1317 env.quad_map = NULL;
1318 env.lftr_edges = NULL;
1320 env.lftr_replaced = 0;
1322 env.need_postpass = 0;
1323 env.process_scc = process_phi_only_scc;
1325 /* Clear all links and move Proj nodes into the
1326 * the same block as their predecessors.
1327 * This can improve the placement of new nodes.
1330 irg_walk_graph(irg, NULL, clear_and_fix, &projs_moved);
1332 set_irg_outs_inconsistent(irg);
1334 /* we need outs for calculating the post order */
1335 assure_irg_outs(irg);
1337 /* calculate the post order number for blocks. */
1338 irg_out_block_walk(get_irg_start_block(irg), NULL, assign_po, &env);
1340 /* calculate the SCC's and drive OSR. */
1341 ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK);
1343 ir_free_resources(irg, IR_RESOURCE_IRN_LINK);
1346 set_irg_outs_inconsistent(irg);
1347 DB((dbg, LEVEL_1, "remove_phi_cycles: %u Cycles removed\n\n", env.replaced));
1350 DEL_ARR_F(env.stack);
1351 obstack_free(&env.obst, NULL);
1353 current_ir_graph = rem;
1354 } /* remove_phi_cycles */
1356 ir_graph_pass_t *remove_phi_cycles_pass(const char *name)
1358 return def_graph_pass(name ? name : "remove_phi_cycles", remove_phi_cycles);
1359 } /* remove_phi_cycles_pass */
1362 * Post-walker: fix Add and Sub nodes that where results of I<->P conversions.
1364 static void fix_adds_and_subs(ir_node *irn, void *ctx)
1369 ir_mode *mode = get_irn_mode(irn);
1371 if (mode_is_int(mode)) {
1374 pred = get_Add_left(irn);
1375 if (get_irn_mode(pred) != mode) {
1376 ir_node *block = get_nodes_block(pred);
1378 pred = new_r_Conv(block, pred, mode);
1379 set_Add_left(irn, pred);
1381 pred = get_Add_right(irn);
1382 if (get_irn_mode(pred) != mode) {
1383 ir_node *block = get_nodes_block(pred);
1385 pred = new_r_Conv(block, pred, mode);
1386 set_Add_right(irn, pred);
1389 } else if (is_Sub(irn)) {
1390 ir_mode *mode = get_irn_mode(irn);
1392 if (mode_is_int(mode)) {
1393 ir_node *left = get_Sub_left(irn);
1394 ir_node *right = get_Sub_right(irn);
1395 ir_mode *l_mode = get_irn_mode(left);
1396 ir_mode *r_mode = get_irn_mode(right);
1398 if (mode_is_int(l_mode) && mode_is_int(r_mode)) {
1399 if (l_mode != mode) {
1400 ir_node *block = get_nodes_block(left);
1402 left = new_r_Conv(block, left, mode);
1403 set_Sub_left(irn, left);
1405 if (r_mode != mode) {
1406 ir_node *block = get_nodes_block(right);
1408 right = new_r_Conv(block, right, mode);
1409 set_Sub_right(irn, right);
1414 } /* fix_adds_and_subs */
1416 /* Performs Operator Strength Reduction for the passed graph. */
1417 void opt_osr(ir_graph *irg, unsigned flags)
1424 rem = current_ir_graph;
1425 current_ir_graph = irg;
1427 FIRM_DBG_REGISTER(dbg, "firm.opt.osr");
1429 DB((dbg, LEVEL_1, "Doing Operator Strength Reduction for %+F\n", irg));
1431 obstack_init(&env.obst);
1432 env.stack = NEW_ARR_F(ir_node *, 128);
1436 env.quad_map = new_set(quad_cmp, 64);
1437 env.lftr_edges = new_set(LFTR_cmp, 64);
1439 env.lftr_replaced = 0;
1440 env.osr_flags = flags;
1441 env.need_postpass = 0;
1442 env.process_scc = process_scc;
1444 /* Clear all links and move Proj nodes into the
1445 the same block as it's predecessors.
1446 This can improve the placement of new nodes.
1449 irg_walk_graph(irg, NULL, clear_and_fix, &projs_moved);
1451 set_irg_outs_inconsistent(irg);
1453 /* we need dominance */
1456 edges = edges_assure(irg);
1458 /* calculate the post order number for blocks by walking the out edges. */
1459 assure_irg_outs(irg);
1460 irg_block_edges_walk(get_irg_start_block(irg), NULL, assign_po, &env);
1462 /* calculate the SCC's and drive OSR. */
1463 ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK);
1467 if (env.need_postpass)
1468 irg_walk_graph(irg, NULL, fix_adds_and_subs, &env);
1470 /* try linear function test replacements */
1474 set_irg_outs_inconsistent(irg);
1475 DB((dbg, LEVEL_1, "Replacements: %u + %u (lftr)\n\n", env.replaced, env.lftr_replaced));
1477 ir_free_resources(irg, IR_RESOURCE_IRN_LINK);
1479 del_set(env.lftr_edges);
1480 del_set(env.quad_map);
1481 DEL_ARR_F(env.stack);
1482 obstack_free(&env.obst, NULL);
1485 edges_deactivate(irg);
1487 current_ir_graph = rem;
1491 ir_graph_pass_t pass;
1496 * Wrapper for running opt_osr() as an ir_graph pass.
1498 static int pass_wrapper(ir_graph *irg, void *context)
1500 struct pass_t *pass = context;
1501 opt_osr(irg, pass->flags);
1503 } /* pass_wrapper */
1505 ir_graph_pass_t *opt_osr_pass(const char *name, unsigned flags)
1507 struct pass_t *pass = XMALLOCZ(struct pass_t);
1509 pass->flags = flags;
1510 return def_graph_pass_constructor(
1511 &pass->pass, name ? name : "osr", pass_wrapper);
1512 } /* opt_osr_pass */