2 * Copyright (C) 1995-2011 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 ir_tarval *init; /**< the init value iff only one exists. */
63 ir_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 size_t 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 unsigned 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 unsigned 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 = (const LFTR_edge*)e1;
125 const LFTR_edge *l2 = (const LFTR_edge*)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 (LFTR_edge*)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, unsigned 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 set_insert(env->lftr_edges, &key, sizeof(key), HASH_PTR(src));
171 * Gets the node_entry of a node.
173 * @param irn the node
174 * @param env the environment
176 static node_entry *get_irn_ne(ir_node *irn, iv_env *env)
178 node_entry *e = (node_entry*)get_irn_link(irn);
181 e = OALLOCZ(&env->obst, node_entry);
182 set_irn_link(irn, e);
188 * Gets the scc from an induction variable.
190 * @param iv any node of the induction variable
191 * @param env the environment
193 static scc *get_iv_scc(ir_node *iv, iv_env *env)
195 node_entry *e = get_irn_ne(iv, env);
200 * Check if irn is an IV.
202 * @param irn the node to check
203 * @param env the environment
205 * @returns the header if it is one, NULL else
207 static ir_node *is_iv(ir_node *irn, iv_env *env)
209 return get_irn_ne(irn, env)->header;
213 * Check if irn is a region constant.
214 * The block or irn must strictly dominate the header block.
216 * @param irn the node to check
217 * @param header_block the header block of the induction variable
219 static int is_rc(ir_node *irn, ir_node *header_block)
221 ir_node *block = get_nodes_block(irn);
223 return (block != header_block) && block_dominates(block, header_block);
227 * Set compare function for the quad set.
229 static int quad_cmp(const void *e1, const void *e2, size_t size)
231 const quadruple_t *c1 = (const quadruple_t*)e1;
232 const quadruple_t *c2 = (const quadruple_t*)e2;
235 return c1->code != c2->code || c1->op1 != c2->op1 || c1->op2 != c2->op2;
239 * Check if an reduced operation was already calculated.
241 * @param code the opcode of the operation
242 * @param op1 the first operand of the operation
243 * @param op2 the second operand of the operation
244 * @param env the environment
246 * @return the already reduced node or NULL if this operation is not yet reduced
248 static ir_node *search(unsigned code, ir_node *op1, ir_node *op2, iv_env *env)
250 quadruple_t key, *entry;
256 entry = (quadruple_t*)set_find(env->quad_map, &key, sizeof(key),
257 (code * 9) ^ HASH_PTR(op1) ^HASH_PTR(op2));
264 * Add an reduced operation.
266 * @param code the opcode of the operation
267 * @param op1 the first operand of the operation
268 * @param op2 the second operand of the operation
269 * @param result the result of the reduced operation
270 * @param env the environment
272 static void add(unsigned code, ir_node *op1, ir_node *op2, ir_node *result, iv_env *env)
281 set_insert(env->quad_map, &key, sizeof(key),
282 (code * 9) ^ HASH_PTR(op1) ^HASH_PTR(op2));
286 * Find a location where to place a bin-op whose operands are in
289 * @param block1 the block of the first operand
290 * @param block2 the block of the second operand
292 * Note that we know here that such a place must exists. Moreover, this means
293 * that either block1 dominates block2 or vice versa. So, just return
296 static ir_node *find_location(ir_node *block1, ir_node *block2)
298 if (block_dominates(block1, block2))
300 assert(block_dominates(block2, block1));
302 } /* find_location */
305 * Create a node that executes an op1 code op1 operation.
307 * @param code the opcode to execute
308 * @param db debug info to add to the new node
309 * @param op1 the first operand
310 * @param op2 the second operand
311 * @param mode the mode of the new operation
313 * @return the newly created node
315 static ir_node *do_apply(unsigned code, dbg_info *db, ir_node *op1, ir_node *op2, ir_mode *mode)
318 ir_node *block = find_location(get_nodes_block(op1), get_nodes_block(op2));
322 result = new_rd_Mul(db, block, op1, op2, mode);
325 result = new_rd_Add(db, block, op1, op2, mode);
328 result = new_rd_Sub(db, block, op1, op2, mode);
331 panic("Unsupported opcode");
337 * The Apply operation.
339 * @param orig the node that represent the original operation and determines
340 * the opcode, debug-info and mode of a newly created one
341 * @param op1 the first operand
342 * @param op2 the second operand
343 * @param env the environment
345 * @return the newly created node
347 static ir_node *apply(ir_node *header, ir_node *orig, ir_node *op1, ir_node *op2, iv_env *env)
349 unsigned code = get_irn_opcode(orig);
350 ir_node *result = search(code, op1, op2, env);
352 if (result == NULL) {
353 dbg_info *db = get_irn_dbg_info(orig);
354 ir_node *op1_header = get_irn_ne(op1, env)->header;
355 ir_node *op2_header = get_irn_ne(op2, env)->header;
357 if (op1_header == header && is_rc(op2, op1_header)) {
358 result = reduce(orig, op1, op2, env);
360 else if (op2_header == header && is_rc(op1, op2_header)) {
361 result = reduce(orig, op2, op1, env);
364 result = do_apply(code, db, op1, op2, get_irn_mode(orig));
365 get_irn_ne(result, env)->header = NULL;
372 * The Reduce operation.
374 * @param orig the node that represent the original operation and determines
375 * the opcode, debug-info and mode of a newly created one
376 * @param iv the induction variable
377 * @param rc the region constant
378 * @param env the environment
380 * @return the reduced node
382 static ir_node *reduce(ir_node *orig, ir_node *iv, ir_node *rc, iv_env *env)
384 unsigned code = get_irn_opcode(orig);
385 ir_node *result = search(code, iv, rc, env);
387 /* check if we have already done this operation on the iv */
388 if (result == NULL) {
389 node_entry *e, *iv_e;
391 ir_mode *mode = get_irn_mode(orig);
393 result = exact_copy(iv);
395 if (get_irn_mode(result) != mode) {
397 * Beware: we must always create a new induction variable with the same mode
398 * as the node we are replacing. Especially this means the mode might be changed
399 * from P to I and back. This is always possible, because we have only Phi, Add
401 * However, this might lead to AddIs(Iu,Is) which we must fix. The best way to do this
402 * seems to be a post-pass, or we might end with useless Conv's.
404 set_irn_mode(result, mode);
405 env->need_postpass = 1;
407 add(code, iv, rc, result, env);
408 DB((dbg, LEVEL_3, " Created new %+F for %+F (%s %+F)\n", result, iv,
409 get_irn_opname(orig), rc));
411 iv_e = get_irn_ne(iv, env);
412 e = get_irn_ne(result, env);
413 e->header = iv_e->header;
415 /* create the LFTR edge */
416 LFTR_add(iv, result, code, rc, env);
418 for (i = get_irn_arity(result) - 1; i >= 0; --i) {
419 ir_node *o = get_irn_n(result, i);
421 e = get_irn_ne(o, env);
422 if (e->header == iv_e->header)
423 o = reduce(orig, o, rc, env);
424 else if (is_Phi(result) || code == iro_Mul)
425 o = apply(iv_e->header, orig, o, rc, env);
426 set_irn_n(result, i, o);
429 DB((dbg, LEVEL_3, " Already Created %+F for %+F (%s %+F)\n", result, iv,
430 get_irn_opname(orig), rc));
436 * Update the scc for a newly created IV.
438 static void update_scc(ir_node *iv, node_entry *e, iv_env *env)
441 ir_node *header = e->header;
442 waitq *wq = new_waitq();
444 DB((dbg, LEVEL_2, " Creating SCC for new an induction variable:\n "));
448 ir_node *irn = (ir_node*)waitq_get(wq);
449 node_entry *ne = get_irn_ne(irn, env);
453 ne->next = pscc->head;
455 DB((dbg, LEVEL_2, " %+F,", irn));
457 for (i = get_irn_arity(irn) - 1; i >= 0; --i) {
458 ir_node *pred = get_irn_n(irn, i);
459 node_entry *pe = get_irn_ne(pred, env);
461 if (pe->header == header && pe->pscc == NULL) {
462 /* set the pscc here to ensure that the node is NOT enqueued another time */
467 } while (! waitq_empty(wq));
469 DB((dbg, LEVEL_2, "\n"));
473 * The Replace operation. We found a node representing iv (+,-,*) rc
474 * that can be removed by replacing the induction variable iv by a new
475 * one that 'applies' the operation 'irn'.
477 * @param irn the node that will be replaced
478 * @param iv the induction variable
479 * @param rc the region constant
480 * @param env the environment
482 static int replace(ir_node *irn, ir_node *iv, ir_node *rc, iv_env *env)
486 DB((dbg, LEVEL_2, " Replacing %+F\n", irn));
488 result = reduce(irn, iv, rc, env);
492 hook_strength_red(get_irn_irg(irn), irn);
493 exchange(irn, result);
494 e = get_irn_ne(result, env);
495 if (e->pscc == NULL) {
496 e->pscc = OALLOCZ(&env->obst, scc);
497 update_scc(result, e, env);
507 * check if a given node is a mul with 2, 4, 8
509 static int is_x86_shift_const(ir_node *mul)
516 /* normalization put constants on the right side */
517 rc = get_Mul_right(mul);
519 ir_tarval *tv = get_Const_tarval(rc);
521 if (tarval_is_long(tv)) {
522 long value = get_tarval_long(tv);
524 if (value == 2 || value == 4 || value == 8) {
525 /* do not reduce multiplications by 2, 4, 8 */
531 } /* is_x86_shift_const */
535 * Check if an IV represents a counter with constant limits.
537 * @param iv any node of the induction variable
538 * @param env the environment
540 static int is_counter_iv(ir_node *iv, iv_env *env)
542 node_entry *e = get_irn_ne(iv, env);
544 ir_node *have_init = NULL;
545 ir_node *have_incr = NULL;
546 ir_opcode code = iro_Bad;
549 if (pscc->code != 0) {
550 /* already analysed */
551 return pscc->code != iro_Bad;
554 pscc->code = iro_Bad;
555 for (irn = pscc->head; irn != NULL; irn = e->next) {
557 if (have_incr != NULL)
560 have_incr = get_Add_right(irn);
561 if (! is_Const(have_incr)) {
562 have_incr = get_Add_left(irn);
563 if (! is_Const(have_incr))
567 } else if (is_Sub(irn)) {
568 if (have_incr != NULL)
571 have_incr = get_Sub_right(irn);
572 if (! is_Const(have_incr))
575 } else if (is_Phi(irn)) {
578 for (i = get_Phi_n_preds(irn) - 1; i >= 0; --i) {
579 ir_node *pred = get_Phi_pred(irn, i);
580 node_entry *ne = get_irn_ne(pred, env);
582 if (ne->header == e->header)
584 if (have_init != NULL)
587 if (! is_Const(pred))
592 e = get_irn_ne(irn, env);
594 pscc->init = get_Const_tarval(have_init);
595 pscc->incr = get_Const_tarval(have_incr);
597 return code != iro_Bad;
598 } /* is_counter_iv */
601 * Check the users of an induction variable for register pressure.
603 * @param iv any node of the induction variable
604 * @param env the environment
606 * @return non-zero if the register pressure is estimated
607 * to not increase, zero else
609 static int check_users_for_reg_pressure(ir_node *iv, iv_env *env)
611 ir_node *irn, *header;
612 ir_node *have_user = NULL;
613 ir_node *have_cmp = NULL;
614 node_entry *e = get_irn_ne(iv, env);
618 for (irn = pscc->head; irn != NULL; irn = e->next) {
619 const ir_edge_t *edge;
621 foreach_out_edge(irn, edge) {
622 ir_node *user = get_edge_src_irn(edge);
623 node_entry *ne = get_irn_ne(user, env);
625 if (e->header == ne->header) {
626 /* found user from the same IV */
630 if (have_cmp != NULL) {
631 /* more than one cmp, for now end here */
636 /* user is a real user of the IV */
637 if (have_user != NULL) {
638 /* found the second user */
644 e = get_irn_ne(irn, env);
647 if (have_user == NULL) {
648 /* no user, ignore */
652 if (have_cmp == NULL) {
653 /* fine, only one user, try to reduce */
657 * We found one user AND at least one cmp.
658 * We should check here if we can transform the Cmp.
660 * For now our capabilities for doing linear function test
661 * are limited, so check if the iv has the right form: Only ONE
662 * Phi, only one Add/Sub with a Const.
664 if (! is_counter_iv(iv, env))
668 * Ok, we have only one increment AND it is a Const, we might be able
669 * to do a linear function test replacement, so go on.
672 } /* check_users_for_reg_pressure */
675 * Check if a node can be replaced (+, -, *).
677 * @param irn the node to check
678 * @param env the environment
680 * @return non-zero if irn should be Replace'd
682 static int check_replace(ir_node *irn, iv_env *env)
684 ir_node *left, *right, *iv, *rc;
685 ir_op *op = get_irn_op(irn);
686 unsigned code = get_op_code(op);
695 left = get_binop_left(irn);
696 right = get_binop_right(irn);
698 liv = is_iv(left, env);
699 riv = is_iv(right, env);
700 if (liv && is_rc(right, liv)) {
701 iv = left; rc = right;
703 else if (riv && is_op_commutative(op) &&
705 iv = right; rc = left;
709 if (env->osr_flags & osr_flag_keep_reg_pressure) {
710 if (! check_users_for_reg_pressure(iv, env))
713 return replace(irn, iv, rc, env);
720 } /* check_replace */
723 * Check which SCC's are induction variables.
726 * @param env the environment
728 static void classify_iv(scc *pscc, iv_env *env)
730 ir_node *irn, *next, *header = NULL;
731 node_entry *b, *h = NULL;
732 int j, only_phi, num_outside;
735 /* find the header block for this scc */
736 for (irn = pscc->head; irn; irn = next) {
737 node_entry *e = (node_entry*)get_irn_link(irn);
738 ir_node *block = get_nodes_block(irn);
741 b = get_irn_ne(block, env);
744 if (h->POnum < b->POnum) {
755 /* check if this scc contains only Phi, Add or Sub nodes */
759 for (irn = pscc->head; irn; irn = next) {
760 node_entry *e = get_irn_ne(irn, env);
763 switch (get_irn_opcode(irn)) {
769 for (j = get_irn_arity(irn) - 1; j >= 0; --j) {
770 ir_node *pred = get_irn_n(irn, j);
771 node_entry *pe = get_irn_ne(pred, env);
773 if (pe->pscc != e->pscc) {
774 /* not in the same SCC, must be a region const */
775 if (! is_rc(pred, header)) {
776 /* not an induction variable */
782 } else if (out_rc != pred) {
789 /* not an induction variable */
793 /* found an induction variable */
794 DB((dbg, LEVEL_2, " Found an induction variable:\n "));
795 if (only_phi && num_outside == 1) {
796 /* a phi cycle with only one real predecessor can be collapsed */
797 DB((dbg, LEVEL_2, " Found an USELESS Phi cycle:\n "));
799 for (irn = pscc->head; irn; irn = next) {
800 node_entry *e = get_irn_ne(irn, env);
803 exchange(irn, out_rc);
809 /* set the header for every node in this scc */
810 for (irn = pscc->head; irn; irn = next) {
811 node_entry *e = get_irn_ne(irn, env);
814 DB((dbg, LEVEL_2, " %+F,", irn));
816 DB((dbg, LEVEL_2, "\n"));
820 for (irn = pscc->head; irn; irn = next) {
821 node_entry *e = get_irn_ne(irn, env);
829 * Process an SCC for the operator strength reduction.
831 * @param pscc the SCC
832 * @param env the environment
834 static void process_scc(scc *pscc, iv_env *env)
836 ir_node *head = pscc->head;
837 node_entry *e = (node_entry*)get_irn_link(head);
843 DB((dbg, LEVEL_4, " SCC at %p:\n ", pscc));
844 for (irn = pscc->head; irn != NULL; irn = next) {
845 node_entry *e = (node_entry*)get_irn_link(irn);
849 DB((dbg, LEVEL_4, " %+F,", irn));
851 DB((dbg, LEVEL_4, "\n"));
855 if (e->next == NULL) {
856 /* this SCC has only a single member */
857 check_replace(head, env);
859 classify_iv(pscc, env);
864 * If an SCC is a Phi only cycle, remove it.
866 * @param pscc an SCC that consists of Phi nodes only
867 * @param env the environment
869 static void remove_phi_cycle(scc *pscc, iv_env *env)
875 /* check if this scc contains only Phi, Add or Sub nodes */
877 for (irn = pscc->head; irn; irn = next) {
878 node_entry *e = get_irn_ne(irn, env);
884 for (j = get_irn_arity(irn) - 1; j >= 0; --j) {
885 ir_node *pred = get_irn_n(irn, j);
886 node_entry *pe = get_irn_ne(pred, env);
888 if (pe->pscc != e->pscc) {
889 /* not in the same SCC, must be the only input */
892 } else if (out_rc != pred) {
898 /* found a Phi cycle */
899 DB((dbg, LEVEL_2, " Found an USELESS Phi cycle:\n "));
901 for (irn = pscc->head; irn; irn = next) {
902 node_entry *e = get_irn_ne(irn, env);
905 exchange(irn, out_rc);
908 } /* remove_phi_cycle */
911 * Process a SCC for the Phi cycle remove.
913 * @param pscc the SCC
914 * @param env the environment
916 static void process_phi_only_scc(scc *pscc, iv_env *env)
918 ir_node *head = pscc->head;
919 node_entry *e = (node_entry*)get_irn_link(head);
925 DB((dbg, LEVEL_4, " SCC at %p:\n ", pscc));
926 for (irn = pscc->head; irn; irn = next) {
927 node_entry *e = (node_entry*)get_irn_link(irn);
931 DB((dbg, LEVEL_4, " %+F,", irn));
933 DB((dbg, LEVEL_4, "\n"));
938 remove_phi_cycle(pscc, env);
939 } /* process_phi_only_scc */
943 * Push a node onto the stack.
945 * @param env the environment
946 * @param n the node to push
948 static void push(iv_env *env, ir_node *n)
952 if (env->tos == ARR_LEN(env->stack)) {
953 size_t nlen = ARR_LEN(env->stack) * 2;
954 ARR_RESIZE(ir_node *, env->stack, nlen);
956 env->stack[env->tos++] = n;
957 e = get_irn_ne(n, env);
962 * Pop a node from the stack.
964 * @param env the environment
966 * @return The topmost node
968 static ir_node *pop(iv_env *env)
970 ir_node *n = env->stack[--env->tos];
971 node_entry *e = get_irn_ne(n, env);
978 * Do Tarjan's SCC algorithm and drive OSR.
980 * @param irn start at this node
981 * @param env the environment
983 static void dfs(ir_node *irn, iv_env *env)
986 node_entry *node = get_irn_ne(irn, env);
988 mark_irn_visited(irn);
990 /* do not put blocks into the scc */
992 n = get_irn_arity(irn);
993 for (i = 0; i < n; ++i) {
994 ir_node *pred = get_irn_n(irn, i);
996 if (!irn_visited(pred))
1000 ir_node *block = get_nodes_block(irn);
1002 node->DFSnum = env->nextDFSnum++;
1003 node->low = node->DFSnum;
1006 /* handle the block */
1007 if (!irn_visited(block))
1010 n = get_irn_arity(irn);
1011 for (i = 0; i < n; ++i) {
1012 ir_node *pred = get_irn_n(irn, i);
1013 node_entry *o = get_irn_ne(pred, env);
1015 if (!irn_visited(pred)) {
1017 node->low = MIN(node->low, o->low);
1019 if (o->DFSnum < node->DFSnum && o->in_stack)
1020 node->low = MIN(o->DFSnum, node->low);
1022 if (node->low == node->DFSnum) {
1023 scc *pscc = OALLOCZ(&env->obst, scc);
1030 e = get_irn_ne(x, env);
1032 e->next = pscc->head;
1036 env->process_scc(pscc, env);
1042 * Do the DFS by starting at the End node of a graph.
1044 * @param irg the graph to process
1045 * @param env the environment
1047 static void do_dfs(ir_graph *irg, iv_env *env)
1049 ir_node *end = get_irg_end(irg);
1052 ir_reserve_resources(irg, IR_RESOURCE_IRN_VISITED);
1054 inc_irg_visited(irg);
1056 /* visit all visible nodes */
1059 /* visit the keep-alives */
1060 for (i = get_End_n_keepalives(end) - 1; i >= 0; --i) {
1061 ir_node *ka = get_End_keepalive(end, i);
1063 if (!irn_visited(ka))
1067 ir_free_resources(irg, IR_RESOURCE_IRN_VISITED);
1071 * Post-block-walker: assign the post-order number.
1073 static void assign_po(ir_node *block, void *ctx)
1075 iv_env *env = (iv_env*)ctx;
1076 node_entry *e = get_irn_ne(block, env);
1078 e->POnum = env->POnum++;
1082 * Apply one LFTR edge operation.
1083 * Return NULL if the transformation cannot be done safely without
1086 * @param iv the induction variable
1087 * @param rc the constant that should be translated
1088 * @param e the LFTR edge
1089 * @param env the IV environment
1091 * @return the translated region constant or NULL
1092 * if the translation was not possible
1095 * In the current implementation only the last edge is stored, so
1096 * only one chain exists. That's why we might miss some opportunities.
1098 static ir_node *applyOneEdge(ir_node *iv, ir_node *rc, LFTR_edge *e, iv_env *env)
1100 if (env->osr_flags & osr_flag_lftr_with_ov_check) {
1101 ir_tarval *tv_l, *tv_r, *tv, *tv_init, *tv_incr, *tv_end;
1102 tarval_int_overflow_mode_t ovmode;
1106 if (! is_counter_iv(iv, env)) {
1107 DB((dbg, LEVEL_4, " not counter IV"));
1111 /* overflow can only be decided for Consts */
1112 if (! is_Const(e->rc)) {
1113 if (e->code == iro_Add && mode_is_reference(get_irn_mode(e->rc))) {
1114 /* However we allow ONE Pointer Add, as pointer arithmetic with wrap
1115 around is undefined anyway */
1116 return do_apply(e->code, NULL, rc, e->rc, get_irn_mode(e->rc));
1118 DB((dbg, LEVEL_4, " = UNKNOWN (%+F)", e->rc));
1122 tv_l = get_Const_tarval(rc);
1123 tv_r = get_Const_tarval(e->rc);
1125 ovmode = tarval_get_integer_overflow_mode();
1126 tarval_set_integer_overflow_mode(TV_OVERFLOW_BAD);
1128 pscc = get_iv_scc(iv, env);
1129 tv_incr = pscc->incr;
1130 tv_init = pscc->init;
1133 * Check that no overflow occurs:
1134 * init must be transformed without overflow
1135 * the new rc must be transformed without overflow
1136 * rc +/- incr must be possible without overflow
1140 tv = tarval_mul(tv_l, tv_r);
1141 tv_init = tarval_mul(tv_init, tv_r);
1142 tv_incr = tarval_mul(tv_incr, tv_r);
1143 DB((dbg, LEVEL_4, " * %+F", tv_r));
1146 tv = tarval_add(tv_l, tv_r);
1147 tv_init = tarval_add(tv_init, tv_r);
1148 DB((dbg, LEVEL_4, " + %+F", tv_r));
1151 tv = tarval_sub(tv_l, tv_r, NULL);
1152 tv_init = tarval_sub(tv_init, tv_r, NULL);
1153 DB((dbg, LEVEL_4, " - %+F", tv_r));
1156 panic("Unsupported opcode");
1159 if (pscc->code == iro_Add) {
1160 tv_end = tarval_add(tv, tv_incr);
1162 assert(pscc->code == iro_Sub);
1163 tv_end = tarval_sub(tv, tv_incr, NULL);
1166 tarval_set_integer_overflow_mode(ovmode);
1168 if (tv == tarval_bad || tv_init == tarval_bad || tv_end == tarval_bad) {
1169 DB((dbg, LEVEL_4, " = OVERFLOW"));
1172 irg = get_irn_irg(iv);
1173 return new_r_Const(irg, tv);
1175 return do_apply(e->code, NULL, rc, e->rc, get_irn_mode(e->dst));
1176 } /* applyOneEdge */
1179 * Applies the operations represented by the LFTR edges to a
1180 * region constant and returns the value.
1181 * Return NULL if the transformation cannot be done safely without
1184 * @param pIV points to the IV node that starts the LFTR edge chain
1185 * after translation points to the new IV
1186 * @param rc the region constant that should be translated
1187 * @param env the IV environment
1189 * @return the translated region constant or NULL
1190 * if the translation was not possible
1192 static ir_node *applyEdges(ir_node **pIV, ir_node *rc, iv_env *env)
1195 if (env->osr_flags & osr_flag_lftr_with_ov_check) {
1196 /* overflow can only be decided for Consts */
1197 if (! is_counter_iv(iv, env)) {
1198 DB((dbg, LEVEL_4, "not counter IV\n", rc));
1201 if (! is_Const(rc)) {
1202 DB((dbg, LEVEL_4, " = UNKNOWN (%+F)\n", rc));
1205 DB((dbg, LEVEL_4, "%+F", get_Const_tarval(rc)));
1209 LFTR_edge *e = LFTR_find(iv, env);
1211 rc = applyOneEdge(iv, rc, e, env);
1216 DB((dbg, LEVEL_3, "\n"));
1222 * Walker, finds Cmp(iv, rc) or Cmp(rc, iv)
1223 * and tries to optimize them.
1225 static void do_lftr(ir_node *cmp, void *ctx)
1227 iv_env *env = (iv_env*)ctx;
1228 ir_node *left, *right, *liv, *riv;
1230 ir_node *nleft = NULL, *nright = NULL;
1235 left = get_Cmp_left(cmp);
1236 right = get_Cmp_right(cmp);
1238 liv = is_iv(left, env);
1239 riv = is_iv(right, env);
1240 if (liv && is_rc(right, liv)) {
1241 iv = left; rc = right;
1243 nright = applyEdges(&iv, rc, env);
1246 else if (riv && is_rc(left, riv)) {
1247 iv = right; rc = left;
1249 nleft = applyEdges(&iv, rc, env);
1253 if (nleft && nright) {
1254 DB((dbg, LEVEL_2, " LFTR for %+F\n", cmp));
1255 set_Cmp_left(cmp, nleft);
1256 set_Cmp_right(cmp, nright);
1257 ++env->lftr_replaced;
1262 * do linear function test replacement.
1264 * @param irg the graph that should be optimized
1265 * @param env the IV environment
1267 static void lftr(ir_graph *irg, iv_env *env)
1269 irg_walk_graph(irg, NULL, do_lftr, env);
1273 * Pre-walker: set all node links to NULL and fix the
1274 * block of Proj nodes.
1276 static void clear_and_fix(ir_node *irn, void *env)
1278 set_irn_link(irn, NULL);
1281 ir_node *pred = get_Proj_pred(irn);
1282 ir_node *pred_block = get_nodes_block(pred);
1284 if (get_nodes_block(irn) != pred_block) {
1285 set_nodes_block(irn, pred_block);
1288 } /* clear_and_fix */
1291 /* Remove any Phi cycles with only one real input. */
1292 void remove_phi_cycles(ir_graph *irg)
1296 FIRM_DBG_REGISTER(dbg, "firm.opt.remove_phi");
1298 DB((dbg, LEVEL_1, "Doing Phi cycle removement for %+F\n", irg));
1300 obstack_init(&env.obst);
1301 env.stack = NEW_ARR_F(ir_node *, 128);
1305 env.quad_map = NULL;
1306 env.lftr_edges = NULL;
1308 env.lftr_replaced = 0;
1310 env.need_postpass = 0;
1311 env.process_scc = process_phi_only_scc;
1313 /* Clear all links and move Proj nodes into the
1314 * the same block as their predecessors.
1315 * This can improve the placement of new nodes.
1317 irg_walk_graph(irg, NULL, clear_and_fix, NULL);
1319 /* we need outs for calculating the post order */
1320 assure_irg_outs(irg);
1322 /* calculate the post order number for blocks. */
1323 irg_out_block_walk(get_irg_start_block(irg), NULL, assign_po, &env);
1325 /* calculate the SCC's and drive OSR. */
1326 ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK);
1328 ir_free_resources(irg, IR_RESOURCE_IRN_LINK);
1331 DB((dbg, LEVEL_1, "remove_phi_cycles: %u Cycles removed\n\n", env.replaced));
1334 DEL_ARR_F(env.stack);
1335 obstack_free(&env.obst, NULL);
1336 } /* remove_phi_cycles */
1338 ir_graph_pass_t *remove_phi_cycles_pass(const char *name)
1340 return def_graph_pass(name ? name : "remove_phi_cycles", remove_phi_cycles);
1341 } /* remove_phi_cycles_pass */
1344 * Post-walker: fix Add and Sub nodes that where results of I<->P conversions.
1346 static void fix_adds_and_subs(ir_node *irn, void *ctx)
1351 ir_mode *mode = get_irn_mode(irn);
1353 if (mode_is_int(mode)) {
1356 pred = get_Add_left(irn);
1357 if (get_irn_mode(pred) != mode) {
1358 ir_node *block = get_nodes_block(pred);
1360 pred = new_r_Conv(block, pred, mode);
1361 set_Add_left(irn, pred);
1363 pred = get_Add_right(irn);
1364 if (get_irn_mode(pred) != mode) {
1365 ir_node *block = get_nodes_block(pred);
1367 pred = new_r_Conv(block, pred, mode);
1368 set_Add_right(irn, pred);
1371 } else if (is_Sub(irn)) {
1372 ir_mode *mode = get_irn_mode(irn);
1374 if (mode_is_int(mode)) {
1375 ir_node *left = get_Sub_left(irn);
1376 ir_node *right = get_Sub_right(irn);
1377 ir_mode *l_mode = get_irn_mode(left);
1378 ir_mode *r_mode = get_irn_mode(right);
1380 if (mode_is_int(l_mode) && mode_is_int(r_mode)) {
1381 if (l_mode != mode) {
1382 ir_node *block = get_nodes_block(left);
1384 left = new_r_Conv(block, left, mode);
1385 set_Sub_left(irn, left);
1387 if (r_mode != mode) {
1388 ir_node *block = get_nodes_block(right);
1390 right = new_r_Conv(block, right, mode);
1391 set_Sub_right(irn, right);
1394 } else if (mode_is_reference(mode)) {
1395 ir_node *left = get_Sub_left(irn);
1396 ir_node *right = get_Sub_right(irn);
1397 ir_mode *l_mode = get_irn_mode(left);
1398 ir_mode *r_mode = get_irn_mode(right);
1399 if (mode_is_int(l_mode)) {
1400 /* Usually, Sub(I*,P) is an error, hence the verifier rejects it.
1401 * However, it is correct in this case, so add Conv to make verifier happy. */
1402 ir_node *block = get_nodes_block(right);
1403 ir_node *lconv = new_r_Conv(block, left, r_mode);
1404 assert(mode_is_reference(r_mode));
1405 set_Sub_left(irn, lconv);
1409 } /* fix_adds_and_subs */
1411 /* Performs Operator Strength Reduction for the passed graph. */
1412 void opt_osr(ir_graph *irg, unsigned flags)
1417 FIRM_DBG_REGISTER(dbg, "firm.opt.osr");
1419 DB((dbg, LEVEL_1, "Doing Operator Strength Reduction for %+F\n", irg));
1421 obstack_init(&env.obst);
1422 env.stack = NEW_ARR_F(ir_node *, 128);
1426 env.quad_map = new_set(quad_cmp, 64);
1427 env.lftr_edges = new_set(LFTR_cmp, 64);
1429 env.lftr_replaced = 0;
1430 env.osr_flags = flags;
1431 env.need_postpass = 0;
1432 env.process_scc = process_scc;
1434 /* Clear all links and move Proj nodes into the
1435 * the same block as its predecessors.
1436 * This can improve the placement of new nodes.
1438 irg_walk_graph(irg, NULL, clear_and_fix, NULL);
1440 /* we need dominance */
1443 edges = edges_assure(irg);
1445 /* calculate the post order number for blocks by walking the out edges. */
1446 assure_irg_outs(irg);
1447 irg_block_edges_walk(get_irg_start_block(irg), NULL, assign_po, &env);
1449 /* calculate the SCC's and drive OSR. */
1450 ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK);
1454 if (env.need_postpass)
1455 irg_walk_graph(irg, NULL, fix_adds_and_subs, &env);
1457 /* try linear function test replacements */
1461 DB((dbg, LEVEL_1, "Replacements: %u + %u (lftr)\n\n", env.replaced, env.lftr_replaced));
1463 ir_free_resources(irg, IR_RESOURCE_IRN_LINK);
1465 del_set(env.lftr_edges);
1466 del_set(env.quad_map);
1467 DEL_ARR_F(env.stack);
1468 obstack_free(&env.obst, NULL);
1471 edges_deactivate(irg);
1474 typedef struct pass_t {
1475 ir_graph_pass_t pass;
1480 * Wrapper for running opt_osr() as an ir_graph pass.
1482 static int pass_wrapper(ir_graph *irg, void *context)
1484 pass_t *pass = (pass_t*)context;
1485 opt_osr(irg, pass->flags);
1487 } /* pass_wrapper */
1489 ir_graph_pass_t *opt_osr_pass(const char *name, unsigned flags)
1491 pass_t *pass = XMALLOCZ(pass_t);
1493 pass->flags = flags;
1494 return def_graph_pass_constructor(
1495 &pass->pass, name ? name : "osr", pass_wrapper);
1496 } /* opt_osr_pass */