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 Cliff Click's Combined Analysis/Optimization
23 * @author Michael Beck
26 * Note further that we use the terminology from Click's work here, which is different
27 * in some cases from Firm terminology. Especially, Click's type is a
28 * Firm tarval/entity, nevertheless we call it type here for "maximum compatibility".
36 #include "iroptimize.h"
44 #include "irgraph_t.h"
59 /* define this to check that all type translations are monotone */
60 #define VERIFY_MONOTONE
64 typedef struct node_t node_t;
65 typedef struct partition_t partition_t;
66 typedef struct opcode_key_t opcode_key_t;
67 typedef struct listmap_entry_t listmap_entry_t;
69 /** The type of the compute function. */
70 typedef void (*compute_func)(node_t *node);
76 ir_opcode code; /**< The Firm opcode. */
77 ir_mode *mode; /**< The mode of all nodes in the partition. */
79 long proj; /**< For Proj nodes, its proj number */
80 ir_entity *ent; /**< For Sel Nodes, its entity */
85 * An entry in the list_map.
87 struct listmap_entry_t {
88 void *id; /**< The id. */
89 node_t *list; /**< The associated list for this id. */
90 listmap_entry_t *next; /**< Link to the next entry in the map. */
93 /** We must map id's to lists. */
94 typedef struct listmap_t {
95 set *map; /**< Map id's to listmap_entry_t's */
96 listmap_entry_t *values; /**< List of all values in the map. */
100 * A lattice element. Because we handle constants and symbolic constants different, we
101 * have to use this union.
112 ir_node *node; /**< The IR-node itself. */
113 list_head node_list; /**< Double-linked list of leader/follower entries. */
114 list_head cprop_list; /**< Double-linked partition.cprop list. */
115 partition_t *part; /**< points to the partition this node belongs to */
116 node_t *next; /**< Next node on local list (partition.touched, fallen). */
117 node_t *race_next; /**< Next node on race list. */
118 lattice_elem_t type; /**< The associated lattice element "type". */
119 int max_user_input; /**< Maximum input number of Def-Use edges. */
120 int next_edge; /**< Index of the next Def-Use edge to use. */
121 int n_followers; /**< Number of Follower in the outs set. */
122 unsigned on_touched:1; /**< Set, if this node is on the partition.touched set. */
123 unsigned on_cprop:1; /**< Set, if this node is on the partition.cprop list. */
124 unsigned on_fallen:1; /**< Set, if this node is on the fallen list. */
125 unsigned is_follower:1; /**< Set, if this node is a follower. */
126 unsigned is_flagged:1; /**< Set, if this node is flagged by step(). */
127 unsigned by_all_const:1; /**< Set, if this node was once evaluated by all constants. */
131 * A partition containing congruent nodes.
134 list_head Leader; /**< The head of partition Leader node list. */
135 list_head Follower; /**< The head of partition Follower node list. */
136 list_head cprop; /**< The head of partition.cprop list. */
137 partition_t *wl_next; /**< Next entry in the work list if any. */
138 partition_t *touched_next; /**< Points to the next partition in the touched set. */
139 partition_t *cprop_next; /**< Points to the next partition in the cprop list. */
140 partition_t *split_next; /**< Points to the next partition in the list that must be split by split_by(). */
141 node_t *touched; /**< The partition.touched set of this partition. */
142 unsigned n_leader; /**< Number of entries in this partition.Leader. */
143 unsigned n_touched; /**< Number of entries in the partition.touched. */
144 int max_user_inputs; /**< Maximum number of user inputs of all entries. */
145 unsigned on_worklist:1; /**< Set, if this partition is in the work list. */
146 unsigned on_touched:1; /**< Set, if this partition is on the touched set. */
147 unsigned on_cprop:1; /**< Set, if this partition is on the cprop list. */
148 unsigned type_is_T_or_C:1;/**< Set, if all nodes in this partition have type Top or Constant. */
150 partition_t *dbg_next; /**< Link all partitions for debugging */
151 unsigned nr; /**< A unique number for (what-)mapping, >0. */
155 typedef struct environment_t {
156 struct obstack obst; /**< obstack to allocate data structures. */
157 partition_t *worklist; /**< The work list. */
158 partition_t *cprop; /**< The constant propagation list. */
159 partition_t *touched; /**< the touched set. */
160 partition_t *initial; /**< The initial partition. */
161 set *opcode2id_map; /**< The opcodeMode->id map. */
162 pmap *type2id_map; /**< The type->id map. */
163 int end_idx; /**< -1 for local and 0 for global congruences. */
164 int lambda_input; /**< Captured argument for lambda_partition(). */
165 int modified; /**< Set, if the graph was modified. */
167 partition_t *dbg_list; /**< List of all partitions. */
171 /** Type of the what function. */
172 typedef void *(*what_func)(const node_t *node, environment_t *env);
174 #define get_irn_node(follower) ((node_t *)get_irn_link(follower))
175 #define set_irn_node(follower, node) set_irn_link(follower, node)
177 /* we do NOT use tarval_unreachable here, instead we use Top for this purpose */
178 #undef tarval_unreachable
179 #define tarval_unreachable tarval_top
182 /** The debug module handle. */
183 DEBUG_ONLY(static firm_dbg_module_t *dbg;)
185 /** Next partition number. */
186 DEBUG_ONLY(static unsigned part_nr = 0);
189 static INLINE lattice_elem_t get_partition_type(const partition_t *X);
192 * Dump partition to output.
194 static void dump_partition(const char *msg, const partition_t *part) {
197 lattice_elem_t type = get_partition_type(part);
199 DB((dbg, LEVEL_2, "%s part%u%s (%u, %+F) {\n ",
200 msg, part->nr, part->type_is_T_or_C ? "*" : "",
201 part->n_leader, type));
202 list_for_each_entry(node_t, node, &part->Leader, node_list) {
203 DB((dbg, LEVEL_2, "%s%+F", first ? "" : ", ", node->node));
206 if (! list_empty(&part->Follower)) {
207 DB((dbg, LEVEL_2, "\n---\n "));
209 list_for_each_entry(node_t, node, &part->Follower, node_list) {
210 DB((dbg, LEVEL_2, "%s%+F", first ? "" : ", ", node->node));
214 DB((dbg, LEVEL_2, "\n}\n"));
215 } /* dump_partition */
220 static void do_dump_list(const char *msg, const node_t *node, int ofs) {
224 #define GET_LINK(p, ofs) *((const node_t **)((char *)(p) + (ofs)))
226 DB((dbg, LEVEL_3, "%s = {\n ", msg));
227 for (p = node; p != NULL; p = GET_LINK(p, ofs)) {
228 DB((dbg, LEVEL_3, "%s%+F", first ? "" : ", ", p->node));
231 DB((dbg, LEVEL_3, "\n}\n"));
239 static void dump_race_list(const char *msg, const node_t *list) {
240 do_dump_list(msg, list, offsetof(node_t, race_next));
244 * Dumps a local list.
246 static void dump_list(const char *msg, const node_t *list) {
247 do_dump_list(msg, list, offsetof(node_t, next));
251 * Dump all partitions.
253 static void dump_all_partitions(const environment_t *env) {
254 const partition_t *P;
256 DB((dbg, LEVEL_2, "All partitions\n===============\n"));
257 for (P = env->dbg_list; P != NULL; P = P->dbg_next)
258 dump_partition("", P);
262 #define dump_partition(msg, part)
263 #define dump_race_list(msg, list)
264 #define dump_list(msg, list)
265 #define dump_all_partitions(env)
268 #if defined(VERIFY_MONOTONE) && defined (DEBUG_libfirm)
270 * Verify that a type transition is monotone
272 static void verify_type(const lattice_elem_t old_type, const lattice_elem_t new_type) {
273 if (old_type.tv == new_type.tv) {
277 if (old_type.tv == tarval_top) {
278 /* from Top down-to is always allowed */
281 if (old_type.tv == tarval_reachable) {
282 panic("verify_type(): wrong translation from %+F to %+F", old_type, new_type);
284 if (new_type.tv == tarval_bottom || new_type.tv == tarval_reachable) {
288 panic("verify_type(): wrong translation from %+F to %+F", old_type, new_type);
291 #define verify_type(old_type, new_type)
295 * Compare two pointer values of a listmap.
297 static int listmap_cmp_ptr(const void *elt, const void *key, size_t size) {
298 const listmap_entry_t *e1 = elt;
299 const listmap_entry_t *e2 = key;
302 return e1->id != e2->id;
303 } /* listmap_cmp_ptr */
306 * Initializes a listmap.
308 * @param map the listmap
310 static void listmap_init(listmap_t *map) {
311 map->map = new_set(listmap_cmp_ptr, 16);
316 * Terminates a listmap.
318 * @param map the listmap
320 static void listmap_term(listmap_t *map) {
325 * Return the associated listmap entry for a given id.
327 * @param map the listmap
328 * @param id the id to search for
330 * @return the asociated listmap entry for the given id
332 static listmap_entry_t *listmap_find(listmap_t *map, void *id) {
333 listmap_entry_t key, *entry;
338 entry = set_insert(map->map, &key, sizeof(key), HASH_PTR(id));
340 if (entry->list == NULL) {
341 /* a new entry, put into the list */
342 entry->next = map->values;
349 * Calculate the hash value for an opcode map entry.
351 * @param entry an opcode map entry
353 * @return a hash value for the given opcode map entry
355 static unsigned opcode_hash(const opcode_key_t *entry) {
356 return (entry->mode - (ir_mode *)0) * 9 + entry->code + entry->u.proj * 3 + HASH_PTR(entry->u.ent);
360 * Compare two entries in the opcode map.
362 static int cmp_opcode(const void *elt, const void *key, size_t size) {
363 const opcode_key_t *o1 = elt;
364 const opcode_key_t *o2 = key;
367 return o1->code != o2->code || o1->mode != o2->mode ||
368 o1->u.proj != o2->u.proj || o1->u.ent != o2->u.ent;
372 * Compare two Def-Use edges for input position.
374 static int cmp_def_use_edge(const void *a, const void *b) {
375 const ir_def_use_edge *ea = a;
376 const ir_def_use_edge *eb = b;
378 /* no overrun, because range is [-1, MAXINT] */
379 return ea->pos - eb->pos;
380 } /* cmp_def_use_edge */
383 * We need the Def-Use edges sorted.
385 static void sort_irn_outs(node_t *node) {
386 ir_node *irn = node->node;
387 int n_outs = get_irn_n_outs(irn);
390 qsort(&irn->out[1], n_outs, sizeof(irn->out[0]), cmp_def_use_edge);
392 node->max_user_input = irn->out[n_outs].pos;
393 } /* sort_irn_outs */
396 * Return the type of a node.
398 * @param irn an IR-node
400 * @return the associated type of this node
402 static INLINE lattice_elem_t get_node_type(const ir_node *irn) {
403 return get_irn_node(irn)->type;
404 } /* get_node_type */
407 * Return the tarval of a node.
409 * @param irn an IR-node
411 * @return the associated type of this node
413 static INLINE tarval *get_node_tarval(const ir_node *irn) {
414 lattice_elem_t type = get_node_type(irn);
416 if (is_tarval(type.tv))
418 return tarval_bottom;
419 } /* get_node_type */
422 * Add a partition to the worklist.
424 static INLINE void add_to_worklist(partition_t *X, environment_t *env) {
425 assert(X->on_worklist == 0);
426 X->wl_next = env->worklist;
429 } /* add_to_worklist */
432 * Create a new empty partition.
434 * @param env the environment
436 * @return a newly allocated partition
438 static INLINE partition_t *new_partition(environment_t *env) {
439 partition_t *part = obstack_alloc(&env->obst, sizeof(*part));
441 INIT_LIST_HEAD(&part->Leader);
442 INIT_LIST_HEAD(&part->Follower);
443 INIT_LIST_HEAD(&part->cprop);
444 part->wl_next = NULL;
445 part->touched_next = NULL;
446 part->cprop_next = NULL;
447 part->split_next = NULL;
448 part->touched = NULL;
451 part->max_user_inputs = 0;
452 part->on_worklist = 0;
453 part->on_touched = 0;
455 part->type_is_T_or_C = 0;
457 part->dbg_next = env->dbg_list;
458 env->dbg_list = part;
459 part->nr = part_nr++;
463 } /* new_partition */
466 * Get the first node from a partition.
468 static INLINE node_t *get_first_node(const partition_t *X) {
469 return list_entry(X->Leader.next, node_t, node_list);
473 * Return the type of a partition (assuming partition is non-empty and
474 * all elements have the same type).
476 * @param X a partition
478 * @return the type of the first element of the partition
480 static INLINE lattice_elem_t get_partition_type(const partition_t *X) {
481 const node_t *first = get_first_node(X);
483 } /* get_partition_type */
486 * Creates a partition node for the given IR-node and place it
487 * into the given partition.
489 * @param irn an IR-node
490 * @param part a partition to place the node in
491 * @param env the environment
493 * @return the created node
495 static node_t *create_partition_node(ir_node *irn, partition_t *part, environment_t *env) {
496 /* create a partition node and place it in the partition */
497 node_t *node = obstack_alloc(&env->obst, sizeof(*node));
499 INIT_LIST_HEAD(&node->node_list);
500 INIT_LIST_HEAD(&node->cprop_list);
504 node->race_next = NULL;
505 node->type.tv = tarval_top;
506 node->max_user_input = 0;
508 node->n_followers = 0;
509 node->on_touched = 0;
512 node->is_follower = 0;
513 node->is_flagged = 0;
514 node->by_all_const = 0;
515 set_irn_node(irn, node);
517 list_add_tail(&node->node_list, &part->Leader);
521 } /* create_partition_node */
524 * Pre-Walker, init all Block-Phi lists.
526 static void init_block_phis(ir_node *irn, void *env) {
530 set_Block_phis(irn, NULL);
535 * Post-Walker, initialize all Nodes' type to U or top and place
536 * all nodes into the TOP partition.
538 static void create_initial_partitions(ir_node *irn, void *ctx) {
539 environment_t *env = ctx;
540 partition_t *part = env->initial;
543 node = create_partition_node(irn, part, env);
545 if (node->max_user_input > part->max_user_inputs)
546 part->max_user_inputs = node->max_user_input;
549 add_Block_phi(get_nodes_block(irn), irn);
551 } /* create_initial_partitions */
554 * Add a node to the entry.partition.touched set and
555 * node->partition to the touched set if not already there.
558 * @param env the environment
560 static INLINE void add_to_touched(node_t *y, environment_t *env) {
561 if (y->on_touched == 0) {
562 partition_t *part = y->part;
564 y->next = part->touched;
569 if (part->on_touched == 0) {
570 part->touched_next = env->touched;
572 part->on_touched = 1;
575 } /* add_to_touched */
578 * Place a node on the cprop list.
581 * @param env the environment
583 static void add_to_cprop(node_t *y, environment_t *env) {
584 /* Add y to y.partition.cprop. */
585 if (y->on_cprop == 0) {
586 partition_t *Y = y->part;
588 list_add_tail(&y->cprop_list, &Y->cprop);
591 DB((dbg, LEVEL_3, "Add %+F to part%u.cprop\n", y->node, Y->nr));
593 /* place its partition on the cprop list */
594 if (Y->on_cprop == 0) {
595 Y->cprop_next = env->cprop;
600 if (get_irn_mode(y->node) == mode_T) {
601 /* mode_T nodes always produce tarval_bottom, so we must explicitly
602 add it's Proj's to get constant evaluation to work */
605 for (i = get_irn_n_outs(y->node) - 1; i >= 0; --i) {
606 node_t *proj = get_irn_node(get_irn_out(y->node, i));
608 add_to_cprop(proj, env);
610 } else if (is_Block(y->node)) {
611 /* Due to the way we handle Phi's, we must place all Phis of a block on the list
612 * if someone placed the block. The Block is only placed if the reachability
613 * changes, and this must be re-evaluated in compute_Phi(). */
615 for (phi = get_Block_phis(y->node); phi != NULL; phi = get_Phi_next(phi)) {
616 node_t *p = get_irn_node(phi);
617 add_to_cprop(p, env);
623 * Update the worklist: If Z is on worklist then add Z' to worklist.
624 * Else add the smaller of Z and Z' to worklist.
626 * @param Z the Z partition
627 * @param Z_prime the Z' partition, a previous part of Z
628 * @param env the environment
630 static void update_worklist(partition_t *Z, partition_t *Z_prime, environment_t *env) {
631 if (Z->on_worklist || Z_prime->n_leader < Z->n_leader) {
632 add_to_worklist(Z_prime, env);
634 add_to_worklist(Z, env);
636 } /* update_worklist */
639 * Split a partition that has NO followers by a local list.
641 * @param Z partition to split
642 * @param g a (non-empty) node list
643 * @param env the environment
645 * @return a new partition containing the nodes of g
647 static partition_t *split_no_followers(partition_t *Z, node_t *g, environment_t *env) {
648 partition_t *Z_prime;
653 dump_partition("Splitting ", Z);
654 dump_list("by list ", g);
658 /* Remove g from Z. */
659 for (node = g; node != NULL; node = node->next) {
660 assert(node->part == Z);
661 list_del(&node->node_list);
664 assert(n < Z->n_leader);
667 /* Move g to a new partition, Z
\92. */
668 Z_prime = new_partition(env);
670 for (node = g; node != NULL; node = node->next) {
671 list_add(&node->node_list, &Z_prime->Leader);
672 node->part = Z_prime;
673 if (node->max_user_input > max_input)
674 max_input = node->max_user_input;
676 Z_prime->max_user_inputs = max_input;
677 Z_prime->n_leader = n;
679 /* for now, copy the type info tag, it will be adjusted in split_by(). */
680 Z_prime->type_is_T_or_C = Z->type_is_T_or_C;
682 update_worklist(Z, Z_prime, env);
684 dump_partition("Now ", Z);
685 dump_partition("Created new ", Z_prime);
687 } /* split_no_followers */
691 #define split(Z, g, env) split_no_followers(*(Z), g, env)
696 * The environment for one race step.
698 typedef struct step_env {
699 node_t *initial; /**< The initial node list. */
700 node_t *unwalked; /**< The unwalked node list. */
701 node_t *walked; /**< The walked node list. */
702 int index; /**< Next index of Follower use_def edge. */
703 int n_leader; /**< number of Leader in initial. */
707 * Do one step in the race.
709 static int step(step_env *env) {
712 if (env->initial != NULL) {
713 /* Move node from initial to unwalked */
715 env->initial = n->race_next;
717 n->race_next = env->unwalked;
723 while (env->unwalked != NULL) {
724 /* let n be the first node in unwalked */
726 while (env->index < n->n_followers) {
727 /* let m be n.F.def_use[index] */
728 node_t *m = get_irn_node(n->node->out[1 + env->index].use);
730 assert(m->is_follower);
733 /* only followers from our partition */
734 if (m->part != n->part)
737 if (! m->is_flagged) {
740 /* add m to unwalked not as first node (we might still need to
741 check for more follower node */
742 m->race_next = n->race_next;
747 /* move n to walked */
748 env->unwalked = n->race_next;
749 n->race_next = env->walked;
757 * Clear the flags from a list.
759 * @param list the list
761 static void clear_flags(node_t *list) {
764 for (n = list; n != NULL; n = n->race_next)
769 * Split a partition by a local list using the race.
771 * @param pX pointer to the partition to split, might be changed!
772 * @param gg a (non-empty) node list
773 * @param env the environment
775 * @return a new partition containing the nodes of gg
777 static partition_t *split(partition_t **pX, node_t *gg, environment_t *env) {
778 partition_t *X = *pX;
779 partition_t *X_prime;
781 step_env env1, env2, *winner;
782 node_t *g, *h, *node;
785 if (list_empty(&X->Follower)) {
786 /* if the partition has NO follower, we can use the fast
787 splitting algorithm. */
788 return split_no_followers(X, gg, env);
790 /* else do the race */
792 /* Note: there might be n-Input followers in this partition. When we split it
793 there inputs might end up in different partitions and these nodes must
794 do the Follower->Leader transition. Put them on the cprop list to let
796 list_for_each_entry(node_t, node, &X->Follower, node_list) {
797 assert(node->is_follower);
798 add_to_cprop(node, env);
801 dump_partition("Splitting ", X);
802 dump_list("by list ", gg);
804 INIT_LIST_HEAD(&tmp);
806 /* Remove gg from X.Leader and put into g */
809 for (node = gg; node != NULL; node = node->next) {
810 assert(node->part == X);
811 assert(node->is_follower == 0);
813 list_del(&node->node_list);
814 list_add_tail(&node->node_list, &tmp);
822 list_for_each_entry(node_t, node, &X->Leader, node_list) {
827 /* restore X.Leader */
828 list_splice(&tmp, &X->Leader);
831 env1.unwalked = NULL;
837 env2.unwalked = NULL;
852 assert(winner->initial == NULL);
853 assert(winner->unwalked == NULL);
855 /* clear flags from walked/unwalked */
856 clear_flags(env1.unwalked);
857 clear_flags(env1.walked);
858 clear_flags(env2.unwalked);
859 clear_flags(env2.walked);
861 dump_race_list("winner ", winner->walked);
863 /* Move walked_{winner} to a new partition, X'. */
864 X_prime = new_partition(env);
866 for (node = winner->walked; node != NULL; node = node->race_next) {
867 list_del(&node->node_list);
868 if (node->is_follower) {
869 list_add(&node->node_list, &X_prime->Follower);
871 list_add(&node->node_list, &X_prime->Leader);
874 node->part = X_prime;
875 if (node->max_user_input > max_input)
876 max_input = node->max_user_input;
878 X_prime->max_user_inputs = max_input;
879 X->n_leader -= winner->n_leader;
881 /* for now, copy the type info tag, it will be adjusted in split_by(). */
882 X_prime->type_is_T_or_C = X->type_is_T_or_C;
884 /* X' is the smaller part */
885 add_to_worklist(X_prime, env);
887 dump_partition("Now ", X);
888 dump_partition("Created new ", X_prime);
890 /* we have to ensure that the partition containing g is returned */
891 if (winner == &env2) {
898 #endif /* NO_FOLLOWER */
901 * Returns non-zero if the i'th input of a Phi node is live.
903 * @param phi a Phi-node
904 * @param i an input number
906 * @return non-zero if the i'th input of the given Phi node is live
908 static int is_live_input(ir_node *phi, int i) {
910 ir_node *block = get_nodes_block(phi);
911 ir_node *pred = get_Block_cfgpred(block, i);
912 lattice_elem_t type = get_node_type(pred);
914 return type.tv != tarval_unreachable;
916 /* else it's the control input, always live */
918 } /* is_live_input */
921 * Return non-zero if a type is a constant.
923 static int is_constant_type(lattice_elem_t type) {
924 if (type.tv != tarval_bottom && type.tv != tarval_top)
927 } /* is_constant_type */
930 * Check whether a type is neither Top or a constant.
931 * Note: U is handled like Top here, R is a constant.
933 * @param type the type to check
935 static int type_is_neither_top_nor_const(const lattice_elem_t type) {
936 if (is_tarval(type.tv)) {
937 if (type.tv == tarval_top)
939 if (tarval_is_constant(type.tv))
949 * Collect nodes to the touched list.
951 * @param list the list which contains the nodes that must be evaluated
952 * @param idx the index of the def_use edge to evaluate
953 * @param env the environment
955 static void collect_touched(list_head *list, int idx, environment_t *env) {
957 int end_idx = env->end_idx;
959 list_for_each_entry(node_t, x, list, node_list) {
963 /* leader edges start AFTER follower edges */
964 x->next_edge = 1 + x->n_followers;
966 num_edges = get_irn_n_outs(x->node);
968 /* for all edges in x.L.def_use_{idx} */
969 while (x->next_edge <= num_edges) {
970 ir_def_use_edge *edge = &x->node->out[x->next_edge];
973 /* check if we have necessary edges */
981 /* ignore the "control input" for non-pinned nodes
982 if we are running in GCSE mode */
983 if (idx < end_idx && get_irn_pinned(succ) != op_pin_state_pinned)
986 y = get_irn_node(succ);
987 if (is_constant_type(y->type)) {
988 ir_opcode code = get_irn_opcode(succ);
989 if (code == iro_Sub || code == iro_Cmp)
990 add_to_cprop(y, env);
993 /* Partitions of constants should not be split simply because their Nodes have unequal
994 functions or incongruent inputs. */
995 if (type_is_neither_top_nor_const(y->type) &&
996 (! is_Phi(y->node) || is_live_input(y->node, idx))) {
997 add_to_touched(y, env);
1001 } /* collect_touched */
1004 * Split the partitions if caused by the first entry on the worklist.
1006 * @param env the environment
1008 static void cause_splits(environment_t *env) {
1009 partition_t *X, *Z, *N;
1012 /* remove the first partition from the worklist */
1014 env->worklist = X->wl_next;
1017 dump_partition("Cause_split: ", X);
1019 /* combine temporary leader and follower list */
1020 for (idx = -1; idx <= X->max_user_inputs; ++idx) {
1021 /* empty the touched set: already done, just clear the list */
1022 env->touched = NULL;
1024 collect_touched(&X->Leader, idx, env);
1025 collect_touched(&X->Follower, idx, env);
1027 for (Z = env->touched; Z != NULL; Z = N) {
1030 unsigned n_touched = Z->n_touched;
1032 assert(Z->touched != NULL);
1034 /* beware, split might change Z */
1035 N = Z->touched_next;
1037 /* remove it from the touched set */
1040 /* Empty local Z.touched AND filter out followers. */
1041 for (pe = &Z->touched; (*pe) != NULL; pe = &e->next) {
1044 if (e->is_follower) {
1045 DB((dbg, LEVEL_2, "Removed follower %+F from touched\n", e->node));
1050 touched = Z->touched;
1054 if (n_touched > 0 && Z->n_leader != n_touched) {
1055 DB((dbg, LEVEL_2, "Split part%d by touched\n", Z->nr));
1056 split(&Z, touched, env);
1060 } /* cause_splits */
1063 * Implements split_by_what(): Split a partition by characteristics given
1064 * by the what function.
1066 * @param X the partition to split
1067 * @param What a function returning an Id for every node of the partition X
1068 * @param P a list to store the result partitions
1069 * @param env the environment
1073 static partition_t *split_by_what(partition_t *X, what_func What,
1074 partition_t **P, environment_t *env) {
1077 listmap_entry_t *iter;
1080 /* Let map be an empty mapping from the range of What to (local) list of Nodes. */
1082 list_for_each_entry(node_t, x, &X->Leader, node_list) {
1083 void *id = What(x, env);
1084 listmap_entry_t *entry;
1087 /* input not allowed, ignore */
1090 /* Add x to map[What(x)]. */
1091 entry = listmap_find(&map, id);
1092 x->next = entry->list;
1095 /* Let P be a set of Partitions. */
1097 /* for all sets S except one in the range of map do */
1098 for (iter = map.values; iter != NULL; iter = iter->next) {
1099 if (iter->next == NULL) {
1100 /* this is the last entry, ignore */
1105 /* Add SPLIT( X, S ) to P. */
1106 DB((dbg, LEVEL_2, "Split part%d by what\n", X->nr));
1107 R = split(&X, S, env);
1117 } /* split_by_what */
1119 /** lambda n.(n.type) */
1120 static void *lambda_type(const node_t *node, environment_t *env) {
1122 return node->type.tv;
1125 /** lambda n.(n.opcode) */
1126 static void *lambda_opcode(const node_t *node, environment_t *env) {
1127 opcode_key_t key, *entry;
1128 ir_node *irn = node->node;
1130 key.code = get_irn_opcode(irn);
1131 key.mode = get_irn_mode(irn);
1135 switch (get_irn_opcode(irn)) {
1137 key.u.proj = get_Proj_proj(irn);
1140 key.u.ent = get_Sel_entity(irn);
1146 entry = set_insert(env->opcode2id_map, &key, sizeof(key), opcode_hash(&key));
1148 } /* lambda_opcode */
1150 /** lambda n.(n[i].partition) */
1151 static void *lambda_partition(const node_t *node, environment_t *env) {
1152 ir_node *skipped = skip_Proj(node->node);
1155 int i = env->lambda_input;
1157 if (i >= get_irn_arity(node->node)) {
1158 /* we are outside the allowed range */
1162 /* ignore the "control input" for non-pinned nodes
1163 if we are running in GCSE mode */
1164 if (i < env->end_idx && get_irn_pinned(skipped) != op_pin_state_pinned)
1167 pred = i == -1 ? get_irn_n(skipped, i) : get_irn_n(node->node, i);
1168 p = get_irn_node(pred);
1171 } /* lambda_partition */
1174 * Returns true if a type is a constant.
1176 static int is_con(const lattice_elem_t type) {
1177 /* be conservative */
1178 if (is_tarval(type.tv))
1179 return tarval_is_constant(type.tv);
1180 return is_entity(type.sym.entity_p);
1184 * Implements split_by().
1186 * @param X the partition to split
1187 * @param env the environment
1189 static void split_by(partition_t *X, environment_t *env) {
1190 partition_t *I, *P = NULL;
1193 dump_partition("split_by", X);
1195 if (X->n_leader == 1) {
1196 /* we have only one leader, no need to split, just check it's type */
1197 node_t *x = get_first_node(X);
1198 X->type_is_T_or_C = x->type.tv == tarval_top || is_con(x->type);
1202 DB((dbg, LEVEL_2, "WHAT = lambda n.(n.type) on part%d\n", X->nr));
1203 P = split_by_what(X, lambda_type, &P, env);
1205 /* adjust the type tags, we have split partitions by type */
1206 for (I = P; I != NULL; I = I->split_next) {
1207 node_t *x = get_first_node(I);
1208 I->type_is_T_or_C = x->type.tv == tarval_top || is_con(x->type);
1215 if (Y->n_leader > 1) {
1216 /* we do not want split the TOP or constant partitions */
1217 if (! Y->type_is_T_or_C) {
1218 partition_t *Q = NULL;
1220 DB((dbg, LEVEL_2, "WHAT = lambda n.(n.opcode) on part%d\n", Y->nr));
1221 Q = split_by_what(Y, lambda_opcode, &Q, env);
1227 if (Z->n_leader > 1) {
1228 const node_t *first = get_first_node(Z);
1229 int arity = get_irn_arity(first->node);
1233 * BEWARE: during splitting by input 2 for instance we might
1234 * create new partitions which are different by input 1, so collect
1235 * them and split further.
1237 Z->split_next = NULL;
1240 for (input = arity - 1; input >= -1; --input) {
1242 partition_t *Z_prime = R;
1245 if (Z_prime->n_leader > 1) {
1246 env->lambda_input = input;
1247 DB((dbg, LEVEL_2, "WHAT = lambda n.(n[%d].partition) on part%d\n", input, Z_prime->nr));
1248 S = split_by_what(Z_prime, lambda_partition, &S, env);
1250 Z_prime->split_next = S;
1253 } while (R != NULL);
1258 } while (Q != NULL);
1261 } while (P != NULL);
1265 * (Re-)compute the type for a given node.
1267 * @param node the node
1269 static void default_compute(node_t *node) {
1271 ir_node *irn = node->node;
1272 node_t *block = get_irn_node(get_nodes_block(irn));
1274 if (block->type.tv == tarval_unreachable) {
1275 node->type.tv = tarval_top;
1279 /* if any of the data inputs have type top, the result is type top */
1280 for (i = get_irn_arity(irn) - 1; i >= 0; --i) {
1281 ir_node *pred = get_irn_n(irn, i);
1282 node_t *p = get_irn_node(pred);
1284 if (p->type.tv == tarval_top) {
1285 node->type.tv = tarval_top;
1290 if (get_irn_mode(node->node) == mode_X)
1291 node->type.tv = tarval_reachable;
1293 node->type.tv = computed_value(irn);
1294 } /* default_compute */
1297 * (Re-)compute the type for a Block node.
1299 * @param node the node
1301 static void compute_Block(node_t *node) {
1303 ir_node *block = node->node;
1305 if (block == get_irg_start_block(current_ir_graph)) {
1306 /* start block is always reachable */
1307 node->type.tv = tarval_reachable;
1311 for (i = get_Block_n_cfgpreds(block) - 1; i >= 0; --i) {
1312 node_t *pred = get_irn_node(get_Block_cfgpred(block, i));
1314 if (pred->type.tv == tarval_reachable) {
1315 /* A block is reachable, if at least of predecessor is reachable. */
1316 node->type.tv = tarval_reachable;
1320 node->type.tv = tarval_top;
1321 } /* compute_Block */
1324 * (Re-)compute the type for a Bad node.
1326 * @param node the node
1328 static void compute_Bad(node_t *node) {
1329 /* Bad nodes ALWAYS compute Top */
1330 node->type.tv = tarval_top;
1334 * (Re-)compute the type for an Unknown node.
1336 * @param node the node
1338 static void compute_Unknown(node_t *node) {
1339 /* While Unknown nodes should compute Top this is dangerous:
1340 * a Top input to a Cond would lead to BOTH control flows unreachable.
1341 * While this is correct in the given semantics, it would destroy the Firm
1344 * It would be safe to compute Top IF it can be assured, that only Cmp
1345 * nodes are inputs to Conds. We check that first.
1346 * This is the way Frontends typically build Firm, but some optimizations
1347 * (cond_eval for instance) might replace them by Phib's...
1349 * For now, we compute bottom here.
1351 node->type.tv = tarval_bottom;
1352 } /* compute_Unknown */
1355 * (Re-)compute the type for a Jmp node.
1357 * @param node the node
1359 static void compute_Jmp(node_t *node) {
1360 node_t *block = get_irn_node(get_nodes_block(node->node));
1362 node->type = block->type;
1366 * (Re-)compute the type for the End node.
1368 * @param node the node
1370 static void compute_End(node_t *node) {
1371 /* the End node is NOT dead of course */
1372 node->type.tv = tarval_reachable;
1376 * (Re-)compute the type for a SymConst node.
1378 * @param node the node
1380 static void compute_SymConst(node_t *node) {
1381 ir_node *irn = node->node;
1382 node_t *block = get_irn_node(get_nodes_block(irn));
1384 if (block->type.tv == tarval_unreachable) {
1385 node->type.tv = tarval_top;
1388 switch (get_SymConst_kind(irn)) {
1389 case symconst_addr_ent:
1390 /* case symconst_addr_name: cannot handle this yet */
1391 node->type.sym = get_SymConst_symbol(irn);
1394 node->type.tv = computed_value(irn);
1396 } /* compute_SymConst */
1399 * (Re-)compute the type for a Phi node.
1401 * @param node the node
1403 static void compute_Phi(node_t *node) {
1405 ir_node *phi = node->node;
1406 lattice_elem_t type;
1408 /* if a Phi is in a unreachable block, its type is TOP */
1409 node_t *block = get_irn_node(get_nodes_block(phi));
1411 if (block->type.tv == tarval_unreachable) {
1412 node->type.tv = tarval_top;
1416 /* Phi implements the Meet operation */
1417 type.tv = tarval_top;
1418 for (i = get_Phi_n_preds(phi) - 1; i >= 0; --i) {
1419 node_t *pred = get_irn_node(get_Phi_pred(phi, i));
1420 node_t *pred_X = get_irn_node(get_Block_cfgpred(block->node, i));
1422 if (pred_X->type.tv == tarval_unreachable || pred->type.tv == tarval_top) {
1423 /* ignore TOP inputs: We must check here for unreachable blocks,
1424 because Firm constants live in the Start Block are NEVER Top.
1425 Else, a Phi (1,2) will produce Bottom, even if the 2 for instance
1426 comes from a unreachable input. */
1429 if (pred->type.tv == tarval_bottom) {
1430 node->type.tv = tarval_bottom;
1432 } else if (type.tv == tarval_top) {
1433 /* first constant found */
1435 } else if (type.tv != pred->type.tv) {
1436 /* different constants or tarval_bottom */
1437 node->type.tv = tarval_bottom;
1440 /* else nothing, constants are the same */
1446 * (Re-)compute the type for an Add. Special case: one nodes is a Zero Const.
1448 * @param node the node
1450 static void compute_Add(node_t *node) {
1451 ir_node *sub = node->node;
1452 node_t *l = get_irn_node(get_Add_left(sub));
1453 node_t *r = get_irn_node(get_Add_right(sub));
1454 lattice_elem_t a = l->type;
1455 lattice_elem_t b = r->type;
1458 if (a.tv == tarval_top || b.tv == tarval_top) {
1459 node->type.tv = tarval_top;
1460 } else if (a.tv == tarval_bottom || b.tv == tarval_bottom) {
1461 node->type.tv = tarval_bottom;
1463 /* x + 0 = 0 + x = x, but beware of floating point +0 + -0, so we
1464 must call tarval_add() first to handle this case! */
1465 if (is_tarval(a.tv)) {
1466 if (is_tarval(b.tv)) {
1467 node->type.tv = tarval_add(a.tv, b.tv);
1470 mode = get_tarval_mode(a.tv);
1471 if (a.tv == get_mode_null(mode)) {
1475 } else if (is_tarval(b.tv)) {
1476 mode = get_tarval_mode(b.tv);
1477 if (b.tv == get_mode_null(mode)) {
1482 node->type.tv = tarval_bottom;
1487 * (Re-)compute the type for a Sub. Special case: both nodes are congruent.
1489 * @param node the node
1491 static void compute_Sub(node_t *node) {
1492 ir_node *sub = node->node;
1493 node_t *l = get_irn_node(get_Sub_left(sub));
1494 node_t *r = get_irn_node(get_Sub_right(sub));
1495 lattice_elem_t a = l->type;
1496 lattice_elem_t b = r->type;
1499 if (a.tv == tarval_top || b.tv == tarval_top) {
1500 node->type.tv = tarval_top;
1501 } else if (is_con(a) && is_con(b)) {
1502 if (is_tarval(a.tv) && is_tarval(b.tv)) {
1503 node->type.tv = tarval_sub(a.tv, b.tv, get_irn_mode(sub));
1504 } else if (is_tarval(a.tv) && tarval_is_null(a.tv)) {
1506 } else if (is_tarval(b.tv) && tarval_is_null(b.tv)) {
1509 node->type.tv = tarval_bottom;
1511 node->by_all_const = 1;
1512 } else if (r->part == l->part &&
1513 (!mode_is_float(get_irn_mode(l->node)))) {
1515 * BEWARE: a - a is NOT always 0 for floating Point values, as
1516 * NaN op NaN = NaN, so we must check this here.
1518 ir_mode *mode = get_irn_mode(sub);
1519 tv = get_mode_null(mode);
1521 /* if the node was ONCE evaluated by all constants, but now
1522 this breakes AND we cat by partition a different result, switch to bottom.
1523 This happens because initially all nodes are in the same partition ... */
1524 if (node->by_all_const && node->type.tv != tv)
1528 node->type.tv = tarval_bottom;
1533 * (Re-)compute the type for Cmp.
1535 * @param node the node
1537 static void compute_Cmp(node_t *node) {
1538 ir_node *cmp = node->node;
1539 node_t *l = get_irn_node(get_Cmp_left(cmp));
1540 node_t *r = get_irn_node(get_Cmp_right(cmp));
1541 lattice_elem_t a = l->type;
1542 lattice_elem_t b = r->type;
1544 if (a.tv == tarval_top || b.tv == tarval_top) {
1545 node->type.tv = tarval_top;
1546 } else if (is_con(a) && is_con(b)) {
1547 /* both nodes are constants, we can probably do something */
1548 node->type.tv = tarval_b_true;
1549 } else if (r->part == l->part) {
1550 /* both nodes congruent, we can probably do something */
1551 node->type.tv = tarval_b_true;
1553 node->type.tv = tarval_bottom;
1555 } /* compute_Proj_Cmp */
1558 * (Re-)compute the type for a Proj(Cmp).
1560 * @param node the node
1561 * @param cond the predecessor Cmp node
1563 static void compute_Proj_Cmp(node_t *node, ir_node *cmp) {
1564 ir_node *proj = node->node;
1565 node_t *l = get_irn_node(get_Cmp_left(cmp));
1566 node_t *r = get_irn_node(get_Cmp_right(cmp));
1567 lattice_elem_t a = l->type;
1568 lattice_elem_t b = r->type;
1569 pn_Cmp pnc = get_Proj_proj(proj);
1572 if (a.tv == tarval_top || b.tv == tarval_top) {
1573 node->type.tv = tarval_top;
1574 } else if (is_con(a) && is_con(b)) {
1575 default_compute(node);
1576 node->by_all_const = 1;
1577 } else if (r->part == l->part &&
1578 (!mode_is_float(get_irn_mode(l->node)) || pnc == pn_Cmp_Lt || pnc == pn_Cmp_Gt)) {
1580 * BEWARE: a == a is NOT always True for floating Point values, as
1581 * NaN != NaN is defined, so we must check this here.
1583 tv = new_tarval_from_long(pnc & pn_Cmp_Eq, mode_b);
1585 /* if the node was ONCE evaluated by all constants, but now
1586 this breakes AND we cat by partition a different result, switch to bottom.
1587 This happens because initially all nodes are in the same partition ... */
1588 if (node->by_all_const && node->type.tv != tv)
1592 node->type.tv = tarval_bottom;
1594 } /* compute_Proj_Cmp */
1597 * (Re-)compute the type for a Proj(Cond).
1599 * @param node the node
1600 * @param cond the predecessor Cond node
1602 static void compute_Proj_Cond(node_t *node, ir_node *cond) {
1603 ir_node *proj = node->node;
1604 long pnc = get_Proj_proj(proj);
1605 ir_node *sel = get_Cond_selector(cond);
1606 node_t *selector = get_irn_node(sel);
1608 if (get_irn_mode(sel) == mode_b) {
1610 if (pnc == pn_Cond_true) {
1611 if (selector->type.tv == tarval_b_false) {
1612 node->type.tv = tarval_unreachable;
1613 } else if (selector->type.tv == tarval_b_true) {
1614 node->type.tv = tarval_reachable;
1615 } else if (selector->type.tv == tarval_bottom) {
1616 node->type.tv = tarval_reachable;
1618 assert(selector->type.tv == tarval_top);
1619 node->type.tv = tarval_unreachable;
1622 assert(pnc == pn_Cond_false);
1624 if (selector->type.tv == tarval_b_false) {
1625 node->type.tv = tarval_reachable;
1626 } else if (selector->type.tv == tarval_b_true) {
1627 node->type.tv = tarval_unreachable;
1628 } else if (selector->type.tv == tarval_bottom) {
1629 node->type.tv = tarval_reachable;
1631 assert(selector->type.tv == tarval_top);
1632 node->type.tv = tarval_unreachable;
1637 if (selector->type.tv == tarval_bottom) {
1638 node->type.tv = tarval_reachable;
1639 } else if (selector->type.tv == tarval_top) {
1640 node->type.tv = tarval_unreachable;
1642 long value = get_tarval_long(selector->type.tv);
1643 if (pnc == get_Cond_defaultProj(cond)) {
1644 /* default switch, have to check ALL other cases */
1647 for (i = get_irn_n_outs(cond) - 1; i >= 0; --i) {
1648 ir_node *succ = get_irn_out(cond, i);
1652 if (value == get_Proj_proj(succ)) {
1653 /* we found a match, will NOT take the default case */
1654 node->type.tv = tarval_unreachable;
1658 /* all cases checked, no match, will take default case */
1659 node->type.tv = tarval_reachable;
1662 node->type.tv = value == pnc ? tarval_reachable : tarval_unreachable;
1666 } /* compute_Proj_Cond */
1669 * (Re-)compute the type for a Proj-Node.
1671 * @param node the node
1673 static void compute_Proj(node_t *node) {
1674 ir_node *proj = node->node;
1675 ir_mode *mode = get_irn_mode(proj);
1676 node_t *block = get_irn_node(get_nodes_block(skip_Proj(proj)));
1677 ir_node *pred = get_Proj_pred(proj);
1679 if (block->type.tv == tarval_unreachable) {
1680 /* a Proj in a unreachable Block stay Top */
1681 node->type.tv = tarval_top;
1684 if (get_irn_node(pred)->type.tv == tarval_top) {
1685 /* if the predecessor is Top, its Proj follow */
1686 node->type.tv = tarval_top;
1690 if (mode == mode_M) {
1691 /* mode M is always bottom */
1692 node->type.tv = tarval_bottom;
1695 if (mode != mode_X) {
1697 compute_Proj_Cmp(node, pred);
1699 default_compute(node);
1702 /* handle mode_X nodes */
1704 switch (get_irn_opcode(pred)) {
1706 /* the Proj_X from the Start is always reachable.
1707 However this is already handled at the top. */
1708 node->type.tv = tarval_reachable;
1711 compute_Proj_Cond(node, pred);
1714 default_compute(node);
1716 } /* compute_Proj */
1719 * (Re-)compute the type for a Confirm.
1721 * @param node the node
1723 static void compute_Confirm(node_t *node) {
1724 ir_node *confirm = node->node;
1725 node_t *pred = get_irn_node(get_Confirm_value(confirm));
1727 if (get_Confirm_cmp(confirm) == pn_Cmp_Eq) {
1728 node_t *bound = get_irn_node(get_Confirm_bound(confirm));
1730 if (is_con(bound->type)) {
1731 /* is equal to a constant */
1732 node->type = bound->type;
1736 /* a Confirm is a copy OR a Const */
1737 node->type = pred->type;
1738 } /* compute_Confirm */
1741 * (Re-)compute the type for a Max.
1743 * @param node the node
1745 static void compute_Max(node_t *node) {
1746 ir_node *op = node->node;
1747 node_t *l = get_irn_node(get_binop_left(op));
1748 node_t *r = get_irn_node(get_binop_right(op));
1749 lattice_elem_t a = l->type;
1750 lattice_elem_t b = r->type;
1752 if (a.tv == tarval_top || b.tv == tarval_top) {
1753 node->type.tv = tarval_top;
1754 } else if (is_con(a) && is_con(b)) {
1755 /* both nodes are constants, we can probably do something */
1757 /* this case handles symconsts as well */
1760 ir_mode *mode = get_irn_mode(op);
1761 tarval *tv_min = get_mode_min(mode);
1765 else if (b.tv == tv_min)
1767 else if (is_tarval(a.tv) && is_tarval(b.tv)) {
1768 if (tarval_cmp(a.tv, b.tv) & pn_Cmp_Gt)
1769 node->type.tv = a.tv;
1771 node->type.tv = b.tv;
1773 node->type.tv = tarval_bad;
1776 } else if (r->part == l->part) {
1777 /* both nodes congruent, we can probably do something */
1780 node->type.tv = tarval_bottom;
1785 * (Re-)compute the type for a Min.
1787 * @param node the node
1789 static void compute_Min(node_t *node) {
1790 ir_node *op = node->node;
1791 node_t *l = get_irn_node(get_binop_left(op));
1792 node_t *r = get_irn_node(get_binop_right(op));
1793 lattice_elem_t a = l->type;
1794 lattice_elem_t b = r->type;
1796 if (a.tv == tarval_top || b.tv == tarval_top) {
1797 node->type.tv = tarval_top;
1798 } else if (is_con(a) && is_con(b)) {
1799 /* both nodes are constants, we can probably do something */
1801 /* this case handles symconsts as well */
1804 ir_mode *mode = get_irn_mode(op);
1805 tarval *tv_max = get_mode_max(mode);
1809 else if (b.tv == tv_max)
1811 else if (is_tarval(a.tv) && is_tarval(b.tv)) {
1812 if (tarval_cmp(a.tv, b.tv) & pn_Cmp_Gt)
1813 node->type.tv = a.tv;
1815 node->type.tv = b.tv;
1817 node->type.tv = tarval_bad;
1820 } else if (r->part == l->part) {
1821 /* both nodes congruent, we can probably do something */
1824 node->type.tv = tarval_bottom;
1829 * (Re-)compute the type for a given node.
1831 * @param node the node
1833 static void compute(node_t *node) {
1836 if (is_no_Block(node->node)) {
1837 node_t *block = get_irn_node(get_nodes_block(node->node));
1839 if (block->type.tv == tarval_unreachable) {
1840 node->type.tv = tarval_top;
1845 func = (compute_func)node->node->op->ops.generic;
1851 * Identity functions: Note that one might thing that identity() is just a
1852 * synonym for equivalent_node(). While this is true, we cannot use it for the algorithm
1853 * here, because it expects that the identity node is one of the inputs, which is NOT
1854 * always true for equivalent_node() which can handle (and does sometimes) DAGs.
1855 * So, we have our own implementation, which copies some parts of equivalent_node()
1859 * Calculates the Identity for Phi nodes
1861 static node_t *identity_Phi(node_t *node) {
1862 ir_node *phi = node->node;
1863 ir_node *block = get_nodes_block(phi);
1864 node_t *n_part = NULL;
1867 for (i = get_Phi_n_preds(phi) - 1; i >= 0; --i) {
1868 node_t *pred_X = get_irn_node(get_Block_cfgpred(block, i));
1870 if (pred_X->type.tv == tarval_reachable) {
1871 node_t *pred = get_irn_node(get_Phi_pred(phi, i));
1875 else if (n_part->part != pred->part) {
1876 /* incongruent inputs, not a follower */
1881 /* if n_part is NULL here, all inputs path are dead, the Phi computes
1882 * tarval_top, is in the TOP partition and should NOT being split! */
1883 assert(n_part != NULL);
1885 } /* identity_Phi */
1888 * Calculates the Identity for commutative 0 neutral nodes.
1890 static node_t *identity_comm_zero_binop(node_t *node) {
1891 ir_node *op = node->node;
1892 node_t *a = get_irn_node(get_binop_left(op));
1893 node_t *b = get_irn_node(get_binop_right(op));
1894 ir_mode *mode = get_irn_mode(op);
1897 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
1898 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
1901 /* node: no input should be tarval_top, else the binop would be also
1902 * Top and not being split. */
1903 zero = get_mode_null(mode);
1904 if (a->type.tv == zero)
1906 if (b->type.tv == zero)
1909 } /* identity_comm_zero_binop */
1911 #define identity_Add identity_comm_zero_binop
1912 #define identity_Or identity_comm_zero_binop
1915 * Calculates the Identity for Mul nodes.
1917 static node_t *identity_Mul(node_t *node) {
1918 ir_node *op = node->node;
1919 node_t *a = get_irn_node(get_Mul_left(op));
1920 node_t *b = get_irn_node(get_Mul_right(op));
1921 ir_mode *mode = get_irn_mode(op);
1924 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
1925 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
1928 /* node: no input should be tarval_top, else the binop would be also
1929 * Top and not being split. */
1930 one = get_mode_one(mode);
1931 if (a->type.tv == one)
1933 if (b->type.tv == one)
1936 } /* identity_Mul */
1939 * Calculates the Identity for Sub nodes.
1941 static node_t *identity_Sub(node_t *node) {
1942 ir_node *sub = node->node;
1943 node_t *b = get_irn_node(get_Sub_right(sub));
1944 ir_mode *mode = get_irn_mode(sub);
1946 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
1947 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
1950 /* node: no input should be tarval_top, else the binop would be also
1951 * Top and not being split. */
1952 if (b->type.tv == get_mode_null(mode))
1953 return get_irn_node(get_Sub_left(sub));
1955 } /* identity_Mul */
1958 * Calculates the Identity for And nodes.
1960 static node_t *identity_And(node_t *node) {
1961 ir_node *and = node->node;
1962 node_t *a = get_irn_node(get_And_left(and));
1963 node_t *b = get_irn_node(get_And_right(and));
1964 tarval *neutral = get_mode_all_one(get_irn_mode(and));
1966 /* node: no input should be tarval_top, else the And would be also
1967 * Top and not being split. */
1968 if (a->type.tv == neutral)
1970 if (b->type.tv == neutral)
1973 } /* identity_And */
1976 * Calculates the Identity for Confirm nodes.
1978 static node_t *identity_Confirm(node_t *node) {
1979 ir_node *confirm = node->node;
1981 /* a Confirm is always a Copy */
1982 return get_irn_node(get_Confirm_value(confirm));
1983 } /* identity_Confirm */
1986 * Calculates the Identity for Mux nodes.
1988 static node_t *identity_Mux(node_t *node) {
1989 ir_node *mux = node->node;
1990 node_t *sel = get_irn_node(get_Mux_sel(mux));
1991 node_t *t = get_irn_node(get_Mux_true(mux));
1992 node_t *f = get_irn_node(get_Mux_false(mux));
1994 if (t->part == f->part)
1997 /* Mux sel input is mode_b, so it is always a tarval */
1998 if (sel->type.tv == tarval_b_true)
2000 if (sel->type.tv == tarval_b_false)
2003 } /* identity_Mux */
2006 * Calculates the Identity for Min nodes.
2008 static node_t *identity_Min(node_t *node) {
2009 ir_node *op = node->node;
2010 node_t *a = get_irn_node(get_binop_left(op));
2011 node_t *b = get_irn_node(get_binop_right(op));
2012 ir_mode *mode = get_irn_mode(op);
2015 if (a->part == b->part) {
2016 /* leader of multiple predecessors */
2020 /* works even with NaN */
2021 tv_max = get_mode_max(mode);
2022 if (a->type.tv == tv_max)
2024 if (b->type.tv == tv_max)
2027 } /* identity_Min */
2030 * Calculates the Identity for Max nodes.
2032 static node_t *identity_Max(node_t *node) {
2033 ir_node *op = node->node;
2034 node_t *a = get_irn_node(get_binop_left(op));
2035 node_t *b = get_irn_node(get_binop_right(op));
2036 ir_mode *mode = get_irn_mode(op);
2039 if (a->part == b->part) {
2040 /* leader of multiple predecessors */
2044 /* works even with NaN */
2045 tv_min = get_mode_min(mode);
2046 if (a->type.tv == tv_min)
2048 if (b->type.tv == tv_min)
2051 } /* identity_Max */
2054 * Calculates the Identity for nodes.
2056 static node_t *identity(node_t *node) {
2057 ir_node *irn = node->node;
2059 switch (get_irn_opcode(irn)) {
2061 return identity_Phi(node);
2063 return identity_Add(node);
2065 return identity_Mul(node);
2067 return identity_Or(node);
2069 return identity_And(node);
2071 return identity_Sub(node);
2073 return identity_Confirm(node);
2075 return identity_Mux(node);
2077 return identity_Min(node);
2079 return identity_Max(node);
2086 * Node follower is a (new) follower of leader, segregate Leader
2089 static void segregate_def_use_chain_1(const ir_node *follower, node_t *leader) {
2090 ir_node *l = leader->node;
2091 int j, i, n = get_irn_n_outs(l);
2093 DB((dbg, LEVEL_2, "%+F is a follower of %+F\n", follower, leader->node));
2094 /* The leader edges must remain sorted, but follower edges can
2096 for (i = leader->n_followers + 1; i <= n; ++i) {
2097 if (l->out[i].use == follower) {
2098 ir_def_use_edge t = l->out[i];
2100 for (j = i - 1; j >= leader->n_followers + 1; --j)
2101 l->out[j + 1] = l->out[j];
2102 ++leader->n_followers;
2103 l->out[leader->n_followers] = t;
2107 } /* segregate_def_use_chain_1 */
2110 * Node follower is a (new) follower of leader, segregate Leader
2111 * out edges. If follower is a n-congruent Input identity, all follower
2112 * inputs congruent to follower are also leader.
2114 static void segregate_def_use_chain(const ir_node *follower, node_t *leader) {
2115 ir_op *op = get_irn_op(follower);
2118 /* n-Congruent Input Identity for Phi's */
2120 ir_node *block = get_nodes_block(follower);
2122 DB((dbg, LEVEL_2, "n-Congruent follower %+F\n", follower));
2123 for (i = get_irn_arity(follower) - 1; i >= 0; --i) {
2124 node_t *pred_X = get_irn_node(get_Block_cfgpred(block, i));
2126 /* beware: we are NOT followers of dead inputs */
2127 if (pred_X->type.tv == tarval_reachable) {
2128 node_t *pred = get_irn_node(get_irn_n(follower, i));
2130 if (pred->part == leader->part)
2131 segregate_def_use_chain_1(follower, pred);
2134 } else if (op == op_Mux || op == op_Max || op == op_Min) {
2135 /* n-Congruent Input Identity */
2138 DB((dbg, LEVEL_2, "n-Congruent follower %+F\n", follower));
2139 for (i = get_irn_arity(follower) - 1; i >= 0; --i) {
2140 node_t *pred = get_irn_node(get_irn_n(follower, i));
2142 if (pred->part == leader->part)
2143 segregate_def_use_chain_1(follower, pred);
2146 /* 1-Congruent Input Identity */
2147 segregate_def_use_chain_1(follower, leader);
2149 } /* segregate_def_use_chain */
2152 * Make all inputs to x from inside X no longer be F.def_use edges.
2154 static void move_edges_to_leader(node_t *x) {
2155 ir_node *irn = x->node;
2158 for (i = get_irn_arity(irn) - 1; i >= 0; --i) {
2159 node_t *pred = get_irn_node(get_irn_n(irn, i));
2164 n = get_irn_n_outs(p);
2165 for (j = 1; j <= pred->n_followers; ++j) {
2166 if (p->out[j].pos == i && p->out[j].use == irn) {
2167 /* found a follower edge to x, move it to the Leader */
2168 ir_def_use_edge edge = p->out[j];
2170 /* remove this edge from the Follower set */
2171 p->out[j] = p->out[pred->n_followers];
2172 --pred->n_followers;
2174 /* sort it into the leader set */
2175 for (k = pred->n_followers + 2; k <= n; ++k) {
2176 if (p->out[k].pos >= edge.pos)
2178 p->out[k - 1] = p->out[k];
2180 /* place the new edge here */
2181 p->out[k - 1] = edge;
2183 /* edge found and moved */
2191 * Propagate constant evaluation.
2193 * @param env the environment
2195 static void propagate(environment_t *env) {
2198 lattice_elem_t old_type;
2200 unsigned n_fallen, old_type_was_T_or_C;
2203 while (env->cprop != NULL) {
2204 void *oldopcode = NULL;
2206 /* remove the first partition X from cprop */
2209 env->cprop = X->cprop_next;
2211 old_type_was_T_or_C = X->type_is_T_or_C;
2213 DB((dbg, LEVEL_2, "Propagate type on part%d\n", X->nr));
2216 while (! list_empty(&X->cprop)) {
2217 /* remove the first Node x from X.cprop */
2218 x = list_entry(X->cprop.next, node_t, cprop_list);
2219 list_del(&x->cprop_list);
2222 if (x->is_follower && identity(x) == x) {
2223 /* x will make the follower -> leader transition */
2224 DB((dbg, LEVEL_2, "%+F make the follower -> leader transition\n", x->node));
2226 if (oldopcode == NULL) {
2227 oldopcode = lambda_opcode(get_first_node(X), env);
2229 if (oldopcode != lambda_opcode(x, env)) {
2230 if (x->on_fallen == 0) {
2231 /* different opcode -> x falls out of this partition */
2236 DB((dbg, LEVEL_2, "Add node %+F to fallen\n", x->node));
2240 /* move x from X.Follower to X.Leader */
2241 list_del(&x->node_list);
2242 list_add_tail(&x->node_list, &X->Leader);
2246 /* Make all inputs to x from inside X no longer be F.def_use edges */
2247 move_edges_to_leader(x);
2250 /* compute a new type for x */
2252 DB((dbg, LEVEL_3, "computing type of %+F\n", x->node));
2254 if (x->type.tv != old_type.tv) {
2255 verify_type(old_type, x->type);
2256 DB((dbg, LEVEL_2, "node %+F has changed type from %+F to %+F\n", x->node, old_type, x->type));
2258 if (x->on_fallen == 0) {
2259 /* Add x to fallen. Nodes might fall from T -> const -> _|_, so check that they are
2260 not already on the list. */
2265 DB((dbg, LEVEL_2, "Add node %+F to fallen\n", x->node));
2267 for (i = get_irn_n_outs(x->node) - 1; i >= 0; --i) {
2268 ir_node *succ = get_irn_out(x->node, i);
2269 node_t *y = get_irn_node(succ);
2271 /* Add y to y.partition.cprop. */
2272 add_to_cprop(y, env);
2277 if (n_fallen > 0 && n_fallen != X->n_leader) {
2278 DB((dbg, LEVEL_2, "Splitting part%d by fallen\n", X->nr));
2279 Y = split(&X, fallen, env);
2283 /* remove the flags from the fallen list */
2284 for (x = fallen; x != NULL; x = x->next)
2288 if (old_type_was_T_or_C) {
2291 if (Y->on_worklist == 0)
2292 add_to_worklist(Y, env);
2294 /* check if some nodes will make the leader -> follower transition */
2295 list_for_each_entry_safe(node_t, y, tmp, &Y->Leader, node_list) {
2296 if (y->type.tv != tarval_top && ! is_con(y->type)) {
2297 node_t *eq_node = identity(y);
2299 if (eq_node != y && eq_node->part == y->part) {
2300 DB((dbg, LEVEL_2, "Node %+F is a follower of %+F\n", y->node, eq_node->node));
2301 /* move to Follower */
2303 list_del(&y->node_list);
2304 list_add_tail(&y->node_list, &Y->Follower);
2307 segregate_def_use_chain(y->node, eq_node);
2318 * Get the leader for a given node from its congruence class.
2320 * @param irn the node
2322 static ir_node *get_leader(node_t *node) {
2323 partition_t *part = node->part;
2325 if (part->n_leader > 1 || node->is_follower) {
2326 DB((dbg, LEVEL_2, "Found congruence class for %+F\n", node->node));
2328 return get_first_node(part)->node;
2334 * Return non-zero if the control flow predecessor node pred
2335 * is the only reachable control flow exit of its block.
2337 * @param pred the control flow exit
2339 static int can_exchange(ir_node *pred) {
2342 else if (is_Jmp(pred))
2344 else if (get_irn_mode(pred) == mode_T) {
2347 /* if the predecessor block has more than one
2348 reachable outputs we cannot remove the block */
2350 for (i = get_irn_n_outs(pred) - 1; i >= 0; --i) {
2351 ir_node *proj = get_irn_out(pred, i);
2354 /* skip non-control flow Proj's */
2355 if (get_irn_mode(proj) != mode_X)
2358 node = get_irn_node(proj);
2359 if (node->type.tv == tarval_reachable) {
2370 * Block Post-Walker, apply the analysis results on control flow by
2371 * shortening Phi's and Block inputs.
2373 static void apply_cf(ir_node *block, void *ctx) {
2374 environment_t *env = ctx;
2375 node_t *node = get_irn_node(block);
2377 ir_node **ins, **in_X;
2378 ir_node *phi, *next;
2380 if (block == get_irg_end_block(current_ir_graph) ||
2381 block == get_irg_start_block(current_ir_graph)) {
2382 /* the EndBlock is always reachable even if the analysis
2383 finds out the opposite :-) */
2386 if (node->type.tv == tarval_unreachable) {
2387 /* mark dead blocks */
2388 set_Block_dead(block);
2392 n = get_Block_n_cfgpreds(block);
2395 /* only one predecessor combine */
2396 ir_node *pred = skip_Proj(get_Block_cfgpred(block, 0));
2398 if (can_exchange(pred)) {
2399 exchange(block, get_nodes_block(pred));
2405 NEW_ARR_A(ir_node *, in_X, n);
2407 for (i = 0; i < n; ++i) {
2408 ir_node *pred = get_Block_cfgpred(block, i);
2409 node_t *node = get_irn_node(pred);
2411 if (node->type.tv == tarval_reachable) {
2418 NEW_ARR_A(ir_node *, ins, n);
2419 for (phi = get_Block_phis(block); phi != NULL; phi = next) {
2420 node_t *node = get_irn_node(phi);
2422 next = get_Phi_next(phi);
2423 if (is_tarval(node->type.tv) && tarval_is_constant(node->type.tv)) {
2424 /* this Phi is replaced by a constant */
2425 tarval *tv = node->type.tv;
2426 ir_node *c = new_r_Const(current_ir_graph, block, get_tarval_mode(tv), tv);
2428 set_irn_node(c, node);
2430 DB((dbg, LEVEL_1, "%+F is replaced by %+F\n", phi, c));
2435 for (i = 0; i < n; ++i) {
2436 node_t *pred = get_irn_node(get_Block_cfgpred(block, i));
2438 if (pred->type.tv == tarval_reachable) {
2439 ins[j++] = get_Phi_pred(phi, i);
2443 /* this Phi is replaced by a single predecessor */
2444 ir_node *s = ins[0];
2447 DB((dbg, LEVEL_1, "%+F is replaced by %+F because of cf change\n", phi, s));
2451 set_irn_in(phi, j, ins);
2458 /* this Block has only one live predecessor */
2459 ir_node *pred = skip_Proj(in_X[0]);
2461 if (can_exchange(pred)) {
2462 exchange(block, get_nodes_block(pred));
2466 set_irn_in(block, k, in_X);
2472 * Post-Walker, apply the analysis results;
2474 static void apply_result(ir_node *irn, void *ctx) {
2475 environment_t *env = ctx;
2476 node_t *node = get_irn_node(irn);
2478 if (is_Block(irn) || is_End(irn) || is_Bad(irn)) {
2479 /* blocks already handled, do not touch the End node */
2481 node_t *block = get_irn_node(get_nodes_block(irn));
2483 if (block->type.tv == tarval_unreachable) {
2484 ir_node *bad = get_irg_bad(current_ir_graph);
2486 /* here, bad might already have a node, but this can be safely ignored
2487 as long as bad has at least ONE valid node */
2488 set_irn_node(bad, node);
2490 DB((dbg, LEVEL_1, "%+F is unreachable\n", irn));
2494 else if (node->type.tv == tarval_unreachable) {
2495 ir_node *bad = get_irg_bad(current_ir_graph);
2497 /* see comment above */
2498 set_irn_node(bad, node);
2500 DB((dbg, LEVEL_1, "%+F is unreachable\n", irn));
2504 else if (get_irn_mode(irn) == mode_X) {
2507 ir_node *cond = get_Proj_pred(irn);
2509 if (is_Cond(cond)) {
2510 node_t *sel = get_irn_node(get_Cond_selector(cond));
2512 if (is_tarval(sel->type.tv) && tarval_is_constant(sel->type.tv)) {
2513 /* Cond selector is a constant, make a Jmp */
2514 ir_node *jmp = new_r_Jmp(current_ir_graph, block->node);
2515 set_irn_node(jmp, node);
2517 DB((dbg, LEVEL_1, "%+F is replaced by %+F\n", irn, jmp));
2524 /* normal data node */
2525 if (is_tarval(node->type.tv) && tarval_is_constant(node->type.tv)) {
2526 tarval *tv = node->type.tv;
2529 * Beware: never replace mode_T nodes by constants. Currently we must mark
2530 * mode_T nodes with constants, but do NOT replace them.
2532 if (! is_Const(irn) && get_irn_mode(irn) != mode_T) {
2533 /* can be replaced by a constant */
2534 ir_node *c = new_r_Const(current_ir_graph, block->node, get_tarval_mode(tv), tv);
2535 set_irn_node(c, node);
2537 DB((dbg, LEVEL_1, "%+F is replaced by %+F\n", irn, c));
2541 } else if (is_entity(node->type.sym.entity_p)) {
2542 if (! is_SymConst(irn)) {
2543 /* can be replaced by a Symconst */
2544 ir_node *symc = new_r_SymConst(current_ir_graph, block->node, get_irn_mode(irn), node->type.sym, symconst_addr_ent);
2545 set_irn_node(symc, node);
2548 DB((dbg, LEVEL_1, "%+F is replaced by %+F\n", irn, symc));
2549 exchange(irn, symc);
2553 ir_node *leader = get_leader(node);
2555 if (leader != irn) {
2556 DB((dbg, LEVEL_1, "%+F from part%d is replaced by %+F\n", irn, node->part->nr, leader));
2557 exchange(irn, leader);
2563 } /* apply_result */
2566 * Fix the keep-alives by deleting unreachable ones.
2568 static void apply_end(ir_node *end, environment_t *env) {
2569 int i, j, n = get_End_n_keepalives(end);
2573 NEW_ARR_A(ir_node *, in, n);
2575 /* fix the keep alive */
2576 for (i = j = 0; i < n; i++) {
2577 ir_node *ka = get_End_keepalive(end, i);
2578 node_t *node = get_irn_node(ka);
2580 if (! node->is_flagged) {
2581 node->is_flagged = 1;
2584 node = get_irn_node(get_nodes_block(ka));
2586 if (node->type.tv != tarval_unreachable)
2591 set_End_keepalives(end, j, in);
2596 #define SET(code) op_##code->ops.generic = (op_func)compute_##code
2599 * sets the generic functions to compute.
2601 static void set_compute_functions(void) {
2604 /* set the default compute function */
2605 for (i = get_irp_n_opcodes() - 1; i >= 0; --i) {
2606 ir_op *op = get_irp_opcode(i);
2607 op->ops.generic = (op_func)default_compute;
2610 /* set specific functions */
2629 } /* set_compute_functions */
2631 static int dump_partition_hook(FILE *F, ir_node *n, ir_node *local) {
2632 ir_node *irn = local != NULL ? local : n;
2633 node_t *node = get_irn_node(irn);
2635 ir_fprintf(F, "info2 : \"partition %u type %+F\"\n", node->part->nr, node->type);
2639 void combo(ir_graph *irg) {
2641 ir_node *initial_bl;
2643 ir_graph *rem = current_ir_graph;
2645 current_ir_graph = irg;
2647 /* register a debug mask */
2648 FIRM_DBG_REGISTER(dbg, "firm.opt.combo");
2649 firm_dbg_set_mask(dbg, SET_LEVEL_3);
2651 DB((dbg, LEVEL_1, "Doing COMBO for %+F\n", irg));
2653 obstack_init(&env.obst);
2654 env.worklist = NULL;
2658 #ifdef DEBUG_libfirm
2659 env.dbg_list = NULL;
2661 env.opcode2id_map = new_set(cmp_opcode, iro_Last * 4);
2662 env.type2id_map = pmap_create();
2663 env.end_idx = get_opt_global_cse() ? 0 : -1;
2664 env.lambda_input = 0;
2667 assure_irg_outs(irg);
2669 /* we have our own value_of function */
2670 set_value_of_func(get_node_tarval);
2672 set_compute_functions();
2673 DEBUG_ONLY(part_nr = 0);
2675 /* create the initial partition and place it on the work list */
2676 env.initial = new_partition(&env);
2677 add_to_worklist(env.initial, &env);
2678 irg_walk_graph(irg, init_block_phis, create_initial_partitions, &env);
2680 /* all nodes on the initial partition have type Top */
2681 env.initial->type_is_T_or_C = 1;
2683 /* Place the START Node's partition on cprop.
2684 Place the START Node on its local worklist. */
2685 initial_bl = get_irg_start_block(irg);
2686 start = get_irn_node(initial_bl);
2687 add_to_cprop(start, &env);
2691 if (env.worklist != NULL)
2693 } while (env.cprop != NULL || env.worklist != NULL);
2695 dump_all_partitions(&env);
2698 set_dump_node_vcgattr_hook(dump_partition_hook);
2699 dump_ir_block_graph(irg, "-partition");
2700 set_dump_node_vcgattr_hook(NULL);
2702 (void)dump_partition_hook;
2705 /* apply the result */
2706 irg_block_walk_graph(irg, NULL, apply_cf, &env);
2707 irg_walk_graph(irg, NULL, apply_result, &env);
2708 apply_end(get_irg_end(irg), &env);
2711 /* control flow might changed */
2712 set_irg_outs_inconsistent(irg);
2713 set_irg_extblk_inconsistent(irg);
2714 set_irg_doms_inconsistent(irg);
2715 set_irg_loopinfo_inconsistent(irg);
2718 pmap_destroy(env.type2id_map);
2719 del_set(env.opcode2id_map);
2720 obstack_free(&env.obst, NULL);
2722 /* restore value_of() default behavior */
2723 set_value_of_func(NULL);
2724 current_ir_graph = rem;