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 partition to the touched set if not already there.
556 * @param part the partition
557 * @param env the environment
559 static INLINE void add_to_touched(partition_t *part, environment_t *env) {
560 if (part->on_touched == 0) {
561 part->touched_next = env->touched;
563 part->on_touched = 1;
565 } /* add_to_touched */
568 * Add a node to the entry.partition.touched set if not already there.
572 static INLINE void add_to_partition_touched(node_t *y) {
573 if (y->on_touched == 0) {
574 partition_t *part = y->part;
576 y->next = part->touched;
581 } /* add_to_partition_touched */
584 * Update the worklist: If Z is on worklist then add Z' to worklist.
585 * Else add the smaller of Z and Z' to worklist.
587 * @param Z the Z partition
588 * @param Z_prime the Z' partition, a previous part of Z
589 * @param env the environment
591 static void update_worklist(partition_t *Z, partition_t *Z_prime, environment_t *env) {
592 if (Z->on_worklist || Z_prime->n_leader < Z->n_leader) {
593 add_to_worklist(Z_prime, env);
595 add_to_worklist(Z, env);
597 } /* update_worklist */
600 * Split a partition that has NO followers by a local list.
602 * @param Z the Z partition to split
603 * @param g a (non-empty) node list
604 * @param env the environment
606 * @return a new partition containing the nodes of g
608 static partition_t *split_no_followers(partition_t *Z, node_t *g, environment_t *env) {
609 partition_t *Z_prime;
614 dump_partition("Splitting ", Z);
618 /* Remove g from Z. */
619 for (node = g; node != NULL; node = node->next) {
620 list_del(&node->node_list);
623 assert(n < Z->n_leader);
626 /* Move g to a new partition, Z
\92. */
627 Z_prime = new_partition(env);
629 for (node = g; node != NULL; node = node->next) {
630 list_add(&node->node_list, &Z_prime->Leader);
631 node->part = Z_prime;
632 if (node->max_user_input > max_input)
633 max_input = node->max_user_input;
635 Z_prime->max_user_inputs = max_input;
636 Z_prime->n_leader = n;
638 /* for now, copy the type info tag. it will be adjusted
640 Z_prime->type_is_T_or_C = Z->type_is_T_or_C;
642 update_worklist(Z, Z_prime, env);
644 dump_partition("Now ", Z);
645 dump_partition("Created new ", Z_prime);
647 } /* split_no_followers */
651 #define split(Z, g, env) split_no_followers(Z, g, env)
656 * The environment for one race step.
658 typedef struct step_env {
659 node_t *initial; /**< The initial node list. */
660 node_t *unwalked; /**< The unwalked node list. */
661 node_t *unwalked_last; /**< Points to the last element of the unwalked node list. */
662 node_t *walked; /**< The walked node list. */
663 int index; /**< Next index of Follower use_def edge. */
664 int n_leader; /**< number of Leader in initial. */
668 * Do one step in the race.
670 static int step(step_env *env) {
673 if (env->initial != NULL) {
674 /* Move node from initial to unwalked */
676 env->initial = n->race_next;
678 if (env->unwalked_last == NULL)
679 env->unwalked_last = n;
681 n->race_next = env->unwalked;
687 while (env->unwalked != NULL) {
688 /* let n be the first node in unwalked */
690 while (env->index < n->n_followers) {
691 /* let m be n.F.def_use[index] */
692 node_t *m = get_irn_node(n->node->out[1 + env->index].use);
694 assert(m->is_follower);
697 /* only followers from our partition */
698 if (m->part != n->part)
701 if (! m->is_flagged) {
704 /* add m to unwalked not as first node */
706 if (env->unwalked == NULL) {
709 env->unwalked_last->race_next = m;
711 env->unwalked_last = m;
715 /* move n to walked */
716 env->unwalked = n->race_next;
717 n->race_next = env->walked;
725 * Clear the flags from a list.
727 * @param list the list
729 static void clear_flags(node_t *list) {
732 for (n = list; n != NULL; n = n->race_next)
737 * Split a partition by a local list using the race.
739 * @param X the partition to split
740 * @param gg a (non-empty) node list
741 * @param env the environment
743 * @return a new partition containing the nodes of gg
745 static partition_t *split(partition_t *X, node_t *gg, environment_t *env) {
746 partition_t *X_prime;
748 step_env env1, env2, *winner;
749 node_t *g, *h, *node;
752 if (list_empty(&X->Follower)) {
753 /* if the partition has NO follower, we can use the fast
754 splitting algorithm. */
755 return split_no_followers(X, gg, env);
757 /* else do the race */
759 dump_partition("Splitting ", X);
760 dump_list("by list ", gg);
762 INIT_LIST_HEAD(&tmp);
764 /* Remove gg from X.Leader and put into g */
767 for (node = gg; node != NULL; node = node->next) {
768 list_del(&node->node_list);
769 list_add_tail(&node->node_list, &tmp);
777 list_for_each_entry(node_t, node, &X->Leader, node_list) {
782 /* restore X.Leader */
783 list_splice(&tmp, &X->Leader);
786 env1.unwalked = NULL;
787 env1.unwalked_last = NULL;
793 env2.unwalked = NULL;
794 env2.unwalked_last = NULL;
809 assert(winner->initial == NULL);
810 assert(winner->unwalked == NULL);
812 /* clear flags from walked/unwalked */
813 clear_flags(env1.unwalked);
814 clear_flags(env1.walked);
815 clear_flags(env2.unwalked);
816 clear_flags(env2.walked);
818 dump_race_list("winner ", winner->walked);
820 /* Move walked_{winner} to a new partition, X
\92. */
821 X_prime = new_partition(env);
823 for (node = winner->walked; node != NULL; node = node->race_next) {
824 list_del(&node->node_list);
825 if (node->is_follower) {
826 list_add(&node->node_list, &X_prime->Follower);
828 list_add(&node->node_list, &X_prime->Leader);
831 node->part = X_prime;
832 if (node->max_user_input > max_input)
833 max_input = node->max_user_input;
835 X_prime->max_user_inputs = max_input;
836 X->n_leader -= winner->n_leader;
838 /* for now, copy the type info tag. it will be adjusted
840 X_prime->type_is_T_or_C = X->type_is_T_or_C;
842 update_worklist(X, X_prime, env);
844 dump_partition("Now ", X);
845 dump_partition("Created new ", X_prime);
848 #endif /* NO_FOLLOWER */
851 * Returns non-zero if the i'th input of a Phi node is live.
853 * @param phi a Phi-node
854 * @param i an input number
856 * @return non-zero if the i'th input of the given Phi node is live
858 static int is_live_input(ir_node *phi, int i) {
860 ir_node *block = get_nodes_block(phi);
861 ir_node *pred = get_Block_cfgpred(block, i);
862 lattice_elem_t type = get_node_type(pred);
864 return type.tv != tarval_unreachable;
866 /* else it's the control input, always live */
868 } /* is_live_input */
871 * Return non-zero if a type is a constant.
873 static int is_constant_type(lattice_elem_t type) {
874 if (type.tv != tarval_bottom && type.tv != tarval_top)
877 } /* is_constant_type */
880 * Place a node on the cprop list.
883 * @param env the environment
885 static void add_node_to_cprop(node_t *y, environment_t *env) {
886 /* Add y to y.partition.cprop. */
887 if (y->on_cprop == 0) {
888 partition_t *Y = y->part;
890 list_add_tail(&y->cprop_list, &Y->cprop);
893 DB((dbg, LEVEL_3, "Add %+F to part%u.cprop\n", y->node, Y->nr));
895 /* place its partition on the cprop list */
896 if (Y->on_cprop == 0) {
897 Y->cprop_next = env->cprop;
902 if (get_irn_mode(y->node) == mode_T) {
903 /* mode_T nodes always produce tarval_bottom, so we must explicitly
904 add it's Proj's to get constant evaluation to work */
907 for (i = get_irn_n_outs(y->node) - 1; i >= 0; --i) {
908 node_t *proj = get_irn_node(get_irn_out(y->node, i));
910 add_node_to_cprop(proj, env);
914 if (is_Block(y->node)) {
915 /* Due to the way we handle Phi's, we must place all Phis of a block on the list
916 * if someone placed the block. The Block is only placed if the reachability
917 * changes, and this must be re-evaluated in compute_Phi(). */
919 for (phi = get_Block_phis(y->node); phi != NULL; phi = get_Phi_next(phi)) {
920 node_t *p = get_irn_node(phi);
921 add_node_to_cprop(p, env);
924 } /* add_node_to_cprop */
927 * Check whether a type is neither Top or a constant.
928 * Note: U is handled like Top here, R is a constant.
930 * @param type the type to check
932 static int type_is_neither_top_nor_const(const lattice_elem_t type) {
933 if (is_tarval(type.tv)) {
934 if (type.tv == tarval_top)
936 if (tarval_is_constant(type.tv))
946 * Collect nodes to the touched list.
948 * @param list the list which contains the nodes that must be evaluated
949 * @param idx the index of the def_use edge to evaluate
950 * @param env the environment
952 static void collect_touched(list_head *list, int idx, environment_t *env) {
954 int end_idx = env->end_idx;
956 list_for_each_entry(node_t, x, list, node_list) {
960 /* leader edges start AFTER follower edges */
961 x->next_edge = 1 + x->n_followers;
963 num_edges = get_irn_n_outs(x->node);
965 /* for all edges in x.L.def_use_{idx} */
966 while (x->next_edge <= num_edges) {
967 ir_def_use_edge *edge = &x->node->out[x->next_edge];
970 /* check if we have necessary edges */
978 /* ignore the "control input" for non-pinned nodes
979 if we are running in GCSE mode */
980 if (idx < end_idx && get_irn_pinned(succ) != op_pin_state_pinned)
983 y = get_irn_node(succ);
984 if (is_constant_type(y->type)) {
985 ir_opcode code = get_irn_opcode(succ);
986 if (code == iro_Sub || code == iro_Cmp)
987 add_node_to_cprop(y, env);
990 /* Partitions of constants should not be split simply because their Nodes have unequal
991 functions or incongruent inputs. */
992 if (type_is_neither_top_nor_const(y->type) &&
993 (! is_Phi(y->node) || is_live_input(y->node, idx))) {
994 partition_t *Y = y->part;
995 add_to_touched(Y, env);
996 add_to_partition_touched(y);
1002 * Split the partitions if caused by the first entry on the worklist.
1004 * @param env the environment
1006 static void cause_splits(environment_t *env) {
1011 /* remove the first partition from the worklist */
1013 env->worklist = X->wl_next;
1016 dump_partition("Cause_split: ", X);
1018 /* combine temporary leader and follower list */
1019 for (idx = -1; idx <= X->max_user_inputs; ++idx) {
1020 /* empty the touched set: already done, just clear the list */
1021 env->touched = NULL;
1023 collect_touched(&X->Leader, idx, env);
1024 collect_touched(&X->Follower, idx, env);
1026 for (Z = env->touched; Z != NULL; Z = Z->touched_next) {
1027 /* remove it from the touched set */
1030 if (Z->n_leader != Z->n_touched) {
1031 DB((dbg, LEVEL_2, "Split part%d by touched\n", Z->nr));
1032 split(Z, Z->touched, env);
1034 /* Empty local Z.touched. */
1035 for (e = Z->touched; e != NULL; e = e->next) {
1042 } /* cause_splits */
1045 * Implements split_by_what(): Split a partition by characteristics given
1046 * by the what function.
1048 * @param X the partition to split
1049 * @param What a function returning an Id for every node of the partition X
1050 * @param P a list to store the result partitions
1051 * @param env the environment
1055 static partition_t *split_by_what(partition_t *X, what_func What,
1056 partition_t **P, environment_t *env) {
1059 listmap_entry_t *iter;
1062 /* Let map be an empty mapping from the range of What to (local) list of Nodes. */
1064 list_for_each_entry(node_t, x, &X->Leader, node_list) {
1065 void *id = What(x, env);
1066 listmap_entry_t *entry;
1069 /* input not allowed, ignore */
1072 /* Add x to map[What(x)]. */
1073 entry = listmap_find(&map, id);
1074 x->next = entry->list;
1077 /* Let P be a set of Partitions. */
1079 /* for all sets S except one in the range of map do */
1080 for (iter = map.values; iter != NULL; iter = iter->next) {
1081 if (iter->next == NULL) {
1082 /* this is the last entry, ignore */
1087 /* Add SPLIT( X, S ) to P. */
1088 DB((dbg, LEVEL_2, "Split part%d by what\n", X->nr));
1089 R = split(X, S, env);
1099 } /* split_by_what */
1101 /** lambda n.(n.type) */
1102 static void *lambda_type(const node_t *node, environment_t *env) {
1104 return node->type.tv;
1107 /** lambda n.(n.opcode) */
1108 static void *lambda_opcode(const node_t *node, environment_t *env) {
1109 opcode_key_t key, *entry;
1110 ir_node *irn = node->node;
1112 key.code = get_irn_opcode(irn);
1113 key.mode = get_irn_mode(irn);
1117 switch (get_irn_opcode(irn)) {
1119 key.u.proj = get_Proj_proj(irn);
1122 key.u.ent = get_Sel_entity(irn);
1128 entry = set_insert(env->opcode2id_map, &key, sizeof(key), opcode_hash(&key));
1130 } /* lambda_opcode */
1132 /** lambda n.(n[i].partition) */
1133 static void *lambda_partition(const node_t *node, environment_t *env) {
1134 ir_node *skipped = skip_Proj(node->node);
1137 int i = env->lambda_input;
1139 if (i >= get_irn_arity(node->node)) {
1140 /* we are outside the allowed range */
1144 /* ignore the "control input" for non-pinned nodes
1145 if we are running in GCSE mode */
1146 if (i < env->end_idx && get_irn_pinned(skipped) != op_pin_state_pinned)
1149 pred = i == -1 ? get_irn_n(skipped, i) : get_irn_n(node->node, i);
1150 p = get_irn_node(pred);
1153 } /* lambda_partition */
1156 * Returns true if a type is a constant.
1158 static int is_con(const lattice_elem_t type) {
1159 /* be conservative */
1160 if (is_tarval(type.tv))
1161 return tarval_is_constant(type.tv);
1162 return is_entity(type.sym.entity_p);
1166 * Implements split_by().
1168 * @param X the partition to split
1169 * @param env the environment
1171 static void split_by(partition_t *X, environment_t *env) {
1172 partition_t *I, *P = NULL;
1175 dump_partition("split_by", X);
1177 if (X->n_leader == 1) {
1178 /* we have only one leader, no need to split, just check it's type */
1179 node_t *x = get_first_node(X);
1180 X->type_is_T_or_C = x->type.tv == tarval_top || is_con(x->type);
1184 DB((dbg, LEVEL_2, "WHAT = lambda n.(n.type) on part%d\n", X->nr));
1185 P = split_by_what(X, lambda_type, &P, env);
1187 /* adjust the type tags, we have split partitions by type */
1188 for (I = P; I != NULL; I = I->split_next) {
1189 node_t *x = get_first_node(I);
1190 I->type_is_T_or_C = x->type.tv == tarval_top || is_con(x->type);
1197 if (Y->n_leader > 1) {
1198 /* we do not want split the TOP or constant partitions */
1199 if (! Y->type_is_T_or_C) {
1200 partition_t *Q = NULL;
1202 DB((dbg, LEVEL_2, "WHAT = lambda n.(n.opcode) on part%d\n", Y->nr));
1203 Q = split_by_what(Y, lambda_opcode, &Q, env);
1209 if (Z->n_leader > 1) {
1210 const node_t *first = get_first_node(Z);
1211 int arity = get_irn_arity(first->node);
1215 * BEWARE: during splitting by input 2 for instance we might
1216 * create new partitions which are different by input 1, so collect
1217 * them and split further.
1219 Z->split_next = NULL;
1222 for (input = arity - 1; input >= -1; --input) {
1224 partition_t *Z_prime = R;
1227 if (Z_prime->n_leader > 1) {
1228 env->lambda_input = input;
1229 DB((dbg, LEVEL_2, "WHAT = lambda n.(n[%d].partition) on part%d\n", input, Z_prime->nr));
1230 S = split_by_what(Z_prime, lambda_partition, &S, env);
1232 Z_prime->split_next = S;
1235 } while (R != NULL);
1240 } while (Q != NULL);
1243 } while (P != NULL);
1247 * (Re-)compute the type for a given node.
1249 * @param node the node
1251 static void default_compute(node_t *node) {
1253 ir_node *irn = node->node;
1254 node_t *block = get_irn_node(get_nodes_block(irn));
1256 if (block->type.tv == tarval_unreachable) {
1257 node->type.tv = tarval_top;
1261 /* if any of the data inputs have type top, the result is type top */
1262 for (i = get_irn_arity(irn) - 1; i >= 0; --i) {
1263 ir_node *pred = get_irn_n(irn, i);
1264 node_t *p = get_irn_node(pred);
1266 if (p->type.tv == tarval_top) {
1267 node->type.tv = tarval_top;
1272 if (get_irn_mode(node->node) == mode_X)
1273 node->type.tv = tarval_reachable;
1275 node->type.tv = computed_value(irn);
1276 } /* default_compute */
1279 * (Re-)compute the type for a Block node.
1281 * @param node the node
1283 static void compute_Block(node_t *node) {
1285 ir_node *block = node->node;
1287 if (block == get_irg_start_block(current_ir_graph)) {
1288 /* start block is always reachable */
1289 node->type.tv = tarval_reachable;
1293 for (i = get_Block_n_cfgpreds(block) - 1; i >= 0; --i) {
1294 node_t *pred = get_irn_node(get_Block_cfgpred(block, i));
1296 if (pred->type.tv == tarval_reachable) {
1297 /* A block is reachable, if at least of predecessor is reachable. */
1298 node->type.tv = tarval_reachable;
1302 node->type.tv = tarval_top;
1303 } /* compute_Block */
1306 * (Re-)compute the type for a Bad node.
1308 * @param node the node
1310 static void compute_Bad(node_t *node) {
1311 /* Bad nodes ALWAYS compute Top */
1312 node->type.tv = tarval_top;
1316 * (Re-)compute the type for an Unknown node.
1318 * @param node the node
1320 static void compute_Unknown(node_t *node) {
1321 /* While Unknown nodes should compute Top this is dangerous:
1322 * a Top input to a Cond would lead to BOTH control flows unreachable.
1323 * While this is correct in the given semantics, it would destroy the Firm
1326 * It would be safe to compute Top IF it can be assured, that only Cmp
1327 * nodes are inputs to Conds. We check that first.
1328 * This is the way Frontends typically build Firm, but some optimizations
1329 * (cond_eval for instance) might replace them by Phib's...
1331 * For now, we compute bottom here.
1333 node->type.tv = tarval_bottom;
1334 } /* compute_Unknown */
1337 * (Re-)compute the type for a Jmp node.
1339 * @param node the node
1341 static void compute_Jmp(node_t *node) {
1342 node_t *block = get_irn_node(get_nodes_block(node->node));
1344 node->type = block->type;
1348 * (Re-)compute the type for the End node.
1350 * @param node the node
1352 static void compute_End(node_t *node) {
1353 /* the End node is NOT dead of course */
1354 node->type.tv = tarval_reachable;
1358 * (Re-)compute the type for a SymConst node.
1360 * @param node the node
1362 static void compute_SymConst(node_t *node) {
1363 ir_node *irn = node->node;
1364 node_t *block = get_irn_node(get_nodes_block(irn));
1366 if (block->type.tv == tarval_unreachable) {
1367 node->type.tv = tarval_top;
1370 switch (get_SymConst_kind(irn)) {
1371 case symconst_addr_ent:
1372 /* case symconst_addr_name: cannot handle this yet */
1373 node->type.sym = get_SymConst_symbol(irn);
1376 node->type.tv = computed_value(irn);
1378 } /* compute_SymConst */
1381 * (Re-)compute the type for a Phi node.
1383 * @param node the node
1385 static void compute_Phi(node_t *node) {
1387 ir_node *phi = node->node;
1388 lattice_elem_t type;
1390 /* if a Phi is in a unreachable block, its type is TOP */
1391 node_t *block = get_irn_node(get_nodes_block(phi));
1393 if (block->type.tv == tarval_unreachable) {
1394 node->type.tv = tarval_top;
1398 /* Phi implements the Meet operation */
1399 type.tv = tarval_top;
1400 for (i = get_Phi_n_preds(phi) - 1; i >= 0; --i) {
1401 node_t *pred = get_irn_node(get_Phi_pred(phi, i));
1402 node_t *pred_X = get_irn_node(get_Block_cfgpred(block->node, i));
1404 if (pred_X->type.tv == tarval_unreachable || pred->type.tv == tarval_top) {
1405 /* ignore TOP inputs: We must check here for unreachable blocks,
1406 because Firm constants live in the Start Block are NEVER Top.
1407 Else, a Phi (1,2) will produce Bottom, even if the 2 for instance
1408 comes from a unreachable input. */
1411 if (pred->type.tv == tarval_bottom) {
1412 node->type.tv = tarval_bottom;
1414 } else if (type.tv == tarval_top) {
1415 /* first constant found */
1417 } else if (type.tv != pred->type.tv) {
1418 /* different constants or tarval_bottom */
1419 node->type.tv = tarval_bottom;
1422 /* else nothing, constants are the same */
1428 * (Re-)compute the type for an Add. Special case: one nodes is a Zero Const.
1430 * @param node the node
1432 static void compute_Add(node_t *node) {
1433 ir_node *sub = node->node;
1434 node_t *l = get_irn_node(get_Add_left(sub));
1435 node_t *r = get_irn_node(get_Add_right(sub));
1436 lattice_elem_t a = l->type;
1437 lattice_elem_t b = r->type;
1440 if (a.tv == tarval_top || b.tv == tarval_top) {
1441 node->type.tv = tarval_top;
1442 } else if (a.tv == tarval_bottom || b.tv == tarval_bottom) {
1443 node->type.tv = tarval_bottom;
1445 /* x + 0 = 0 + x = x, but beware of floating point +0 + -0, so we
1446 must call tarval_add() first to handle this case! */
1447 if (is_tarval(a.tv)) {
1448 if (is_tarval(b.tv)) {
1449 node->type.tv = tarval_add(a.tv, b.tv);
1452 mode = get_tarval_mode(a.tv);
1453 if (a.tv == get_mode_null(mode)) {
1457 } else if (is_tarval(b.tv)) {
1458 mode = get_tarval_mode(b.tv);
1459 if (b.tv == get_mode_null(mode)) {
1464 node->type.tv = tarval_bottom;
1469 * (Re-)compute the type for a Sub. Special case: both nodes are congruent.
1471 * @param node the node
1473 static void compute_Sub(node_t *node) {
1474 ir_node *sub = node->node;
1475 node_t *l = get_irn_node(get_Sub_left(sub));
1476 node_t *r = get_irn_node(get_Sub_right(sub));
1477 lattice_elem_t a = l->type;
1478 lattice_elem_t b = r->type;
1481 if (a.tv == tarval_top || b.tv == tarval_top) {
1482 node->type.tv = tarval_top;
1483 } else if (is_con(a) && is_con(b)) {
1484 if (is_tarval(a.tv) && is_tarval(b.tv)) {
1485 node->type.tv = tarval_sub(a.tv, b.tv, get_irn_mode(sub));
1486 } else if (is_tarval(a.tv) && tarval_is_null(a.tv)) {
1488 } else if (is_tarval(b.tv) && tarval_is_null(b.tv)) {
1491 node->type.tv = tarval_bottom;
1493 node->by_all_const = 1;
1494 } else if (r->part == l->part &&
1495 (!mode_is_float(get_irn_mode(l->node)))) {
1497 * BEWARE: a - a is NOT always 0 for floating Point values, as
1498 * NaN op NaN = NaN, so we must check this here.
1500 ir_mode *mode = get_irn_mode(sub);
1501 tv = get_mode_null(mode);
1503 /* if the node was ONCE evaluated by all constants, but now
1504 this breakes AND we cat by partition a different result, switch to bottom.
1505 This happens because initially all nodes are in the same partition ... */
1506 if (node->by_all_const && node->type.tv != tv)
1510 node->type.tv = tarval_bottom;
1515 * (Re-)compute the type for Cmp.
1517 * @param node the node
1519 static void compute_Cmp(node_t *node) {
1520 ir_node *cmp = node->node;
1521 node_t *l = get_irn_node(get_Cmp_left(cmp));
1522 node_t *r = get_irn_node(get_Cmp_right(cmp));
1523 lattice_elem_t a = l->type;
1524 lattice_elem_t b = r->type;
1526 if (a.tv == tarval_top || b.tv == tarval_top) {
1527 node->type.tv = tarval_top;
1528 } else if (is_con(a) && is_con(b)) {
1529 /* both nodes are constants, we can probably do something */
1530 node->type.tv = tarval_b_true;
1531 } else if (r->part == l->part) {
1532 /* both nodes congruent, we can probably do something */
1533 node->type.tv = tarval_b_true;
1535 node->type.tv = tarval_bottom;
1537 } /* compute_Proj_Cmp */
1540 * (Re-)compute the type for a Proj(Cmp).
1542 * @param node the node
1543 * @param cond the predecessor Cmp node
1545 static void compute_Proj_Cmp(node_t *node, ir_node *cmp) {
1546 ir_node *proj = node->node;
1547 node_t *l = get_irn_node(get_Cmp_left(cmp));
1548 node_t *r = get_irn_node(get_Cmp_right(cmp));
1549 lattice_elem_t a = l->type;
1550 lattice_elem_t b = r->type;
1551 pn_Cmp pnc = get_Proj_proj(proj);
1554 if (a.tv == tarval_top || b.tv == tarval_top) {
1555 node->type.tv = tarval_top;
1556 } else if (is_con(a) && is_con(b)) {
1557 default_compute(node);
1558 node->by_all_const = 1;
1559 } else if (r->part == l->part &&
1560 (!mode_is_float(get_irn_mode(l->node)) || pnc == pn_Cmp_Lt || pnc == pn_Cmp_Gt)) {
1562 * BEWARE: a == a is NOT always True for floating Point values, as
1563 * NaN != NaN is defined, so we must check this here.
1565 tv = new_tarval_from_long(pnc & pn_Cmp_Eq, mode_b);
1567 /* if the node was ONCE evaluated by all constants, but now
1568 this breakes AND we cat by partition a different result, switch to bottom.
1569 This happens because initially all nodes are in the same partition ... */
1570 if (node->by_all_const && node->type.tv != tv)
1574 node->type.tv = tarval_bottom;
1576 } /* compute_Proj_Cmp */
1579 * (Re-)compute the type for a Proj(Cond).
1581 * @param node the node
1582 * @param cond the predecessor Cond node
1584 static void compute_Proj_Cond(node_t *node, ir_node *cond) {
1585 ir_node *proj = node->node;
1586 long pnc = get_Proj_proj(proj);
1587 ir_node *sel = get_Cond_selector(cond);
1588 node_t *selector = get_irn_node(sel);
1590 if (get_irn_mode(sel) == mode_b) {
1592 if (pnc == pn_Cond_true) {
1593 if (selector->type.tv == tarval_b_false) {
1594 node->type.tv = tarval_unreachable;
1595 } else if (selector->type.tv == tarval_b_true) {
1596 node->type.tv = tarval_reachable;
1597 } else if (selector->type.tv == tarval_bottom) {
1598 node->type.tv = tarval_reachable;
1600 assert(selector->type.tv == tarval_top);
1601 node->type.tv = tarval_unreachable;
1604 assert(pnc == pn_Cond_false);
1606 if (selector->type.tv == tarval_b_false) {
1607 node->type.tv = tarval_reachable;
1608 } else if (selector->type.tv == tarval_b_true) {
1609 node->type.tv = tarval_unreachable;
1610 } else if (selector->type.tv == tarval_bottom) {
1611 node->type.tv = tarval_reachable;
1613 assert(selector->type.tv == tarval_top);
1614 node->type.tv = tarval_unreachable;
1619 if (selector->type.tv == tarval_bottom) {
1620 node->type.tv = tarval_reachable;
1621 } else if (selector->type.tv == tarval_top) {
1622 node->type.tv = tarval_unreachable;
1624 long value = get_tarval_long(selector->type.tv);
1625 if (pnc == get_Cond_defaultProj(cond)) {
1626 /* default switch, have to check ALL other cases */
1629 for (i = get_irn_n_outs(cond) - 1; i >= 0; --i) {
1630 ir_node *succ = get_irn_out(cond, i);
1634 if (value == get_Proj_proj(succ)) {
1635 /* we found a match, will NOT take the default case */
1636 node->type.tv = tarval_unreachable;
1640 /* all cases checked, no match, will take default case */
1641 node->type.tv = tarval_reachable;
1644 node->type.tv = value == pnc ? tarval_reachable : tarval_unreachable;
1648 } /* compute_Proj_Cond */
1651 * (Re-)compute the type for a Proj-Node.
1653 * @param node the node
1655 static void compute_Proj(node_t *node) {
1656 ir_node *proj = node->node;
1657 ir_mode *mode = get_irn_mode(proj);
1658 node_t *block = get_irn_node(get_nodes_block(skip_Proj(proj)));
1659 ir_node *pred = get_Proj_pred(proj);
1661 if (block->type.tv == tarval_unreachable) {
1662 /* a Proj in a unreachable Block stay Top */
1663 node->type.tv = tarval_top;
1666 if (get_irn_node(pred)->type.tv == tarval_top) {
1667 /* if the predecessor is Top, its Proj follow */
1668 node->type.tv = tarval_top;
1672 if (mode == mode_M) {
1673 /* mode M is always bottom */
1674 node->type.tv = tarval_bottom;
1677 if (mode != mode_X) {
1679 compute_Proj_Cmp(node, pred);
1681 default_compute(node);
1684 /* handle mode_X nodes */
1686 switch (get_irn_opcode(pred)) {
1688 /* the Proj_X from the Start is always reachable.
1689 However this is already handled at the top. */
1690 node->type.tv = tarval_reachable;
1693 compute_Proj_Cond(node, pred);
1696 default_compute(node);
1698 } /* compute_Proj */
1701 * (Re-)compute the type for a Confirm.
1703 * @param node the node
1705 static void compute_Confirm(node_t *node) {
1706 ir_node *confirm = node->node;
1707 node_t *pred = get_irn_node(get_Confirm_value(confirm));
1709 if (get_Confirm_cmp(confirm) == pn_Cmp_Eq) {
1710 node_t *bound = get_irn_node(get_Confirm_bound(confirm));
1712 if (is_con(bound->type)) {
1713 /* is equal to a constant */
1714 node->type = bound->type;
1718 /* a Confirm is a copy OR a Const */
1719 node->type = pred->type;
1720 } /* compute_Confirm */
1723 * (Re-)compute the type for a Max.
1725 * @param node the node
1727 static void compute_Max(node_t *node) {
1728 ir_node *op = node->node;
1729 node_t *l = get_irn_node(get_binop_left(op));
1730 node_t *r = get_irn_node(get_binop_right(op));
1731 lattice_elem_t a = l->type;
1732 lattice_elem_t b = r->type;
1734 if (a.tv == tarval_top || b.tv == tarval_top) {
1735 node->type.tv = tarval_top;
1736 } else if (is_con(a) && is_con(b)) {
1737 /* both nodes are constants, we can probably do something */
1739 /* this case handles symconsts as well */
1742 ir_mode *mode = get_irn_mode(op);
1743 tarval *tv_min = get_mode_min(mode);
1747 else if (b.tv == tv_min)
1749 else if (is_tarval(a.tv) && is_tarval(b.tv)) {
1750 if (tarval_cmp(a.tv, b.tv) & pn_Cmp_Gt)
1751 node->type.tv = a.tv;
1753 node->type.tv = b.tv;
1755 node->type.tv = tarval_bad;
1758 } else if (r->part == l->part) {
1759 /* both nodes congruent, we can probably do something */
1762 node->type.tv = tarval_bottom;
1767 * (Re-)compute the type for a Min.
1769 * @param node the node
1771 static void compute_Min(node_t *node) {
1772 ir_node *op = node->node;
1773 node_t *l = get_irn_node(get_binop_left(op));
1774 node_t *r = get_irn_node(get_binop_right(op));
1775 lattice_elem_t a = l->type;
1776 lattice_elem_t b = r->type;
1778 if (a.tv == tarval_top || b.tv == tarval_top) {
1779 node->type.tv = tarval_top;
1780 } else if (is_con(a) && is_con(b)) {
1781 /* both nodes are constants, we can probably do something */
1783 /* this case handles symconsts as well */
1786 ir_mode *mode = get_irn_mode(op);
1787 tarval *tv_max = get_mode_max(mode);
1791 else if (b.tv == tv_max)
1793 else if (is_tarval(a.tv) && is_tarval(b.tv)) {
1794 if (tarval_cmp(a.tv, b.tv) & pn_Cmp_Gt)
1795 node->type.tv = a.tv;
1797 node->type.tv = b.tv;
1799 node->type.tv = tarval_bad;
1802 } else if (r->part == l->part) {
1803 /* both nodes congruent, we can probably do something */
1806 node->type.tv = tarval_bottom;
1811 * (Re-)compute the type for a given node.
1813 * @param node the node
1815 static void compute(node_t *node) {
1818 if (is_no_Block(node->node)) {
1819 node_t *block = get_irn_node(get_nodes_block(node->node));
1821 if (block->type.tv == tarval_unreachable) {
1822 node->type.tv = tarval_top;
1827 func = (compute_func)node->node->op->ops.generic;
1833 * Identity functions: Note that one might thing that identity() is just a
1834 * synonym for equivalent_node(). While this is true, we cannot use it for the algorithm
1835 * here, because it expects that the identity node is one of the inputs, which is NOT
1836 * always true for equivalent_node() which can handle (and does sometimes) DAGs.
1837 * So, we have our own implementation, which copies some parts of equivalent_node()
1841 * Calculates the Identity for Phi nodes
1843 static node_t *identity_Phi(node_t *node) {
1844 ir_node *phi = node->node;
1845 ir_node *block = get_nodes_block(phi);
1846 node_t *n_part = NULL;
1849 for (i = get_Phi_n_preds(phi) - 1; i >= 0; --i) {
1850 node_t *pred_X = get_irn_node(get_Block_cfgpred(block, i));
1852 if (pred_X->type.tv == tarval_reachable) {
1853 node_t *pred = get_irn_node(get_Phi_pred(phi, i));
1857 else if (n_part->part != pred->part) {
1858 /* incongruent inputs, not a follower */
1863 /* if n_part is NULL here, all inputs path are dead, the Phi computes
1864 * tarval_top, is in the TOP partition and should NOT being split! */
1865 assert(n_part != NULL);
1867 } /* identity_Phi */
1870 * Calculates the Identity for commutative 0 neutral nodes.
1872 static node_t *identity_comm_zero_binop(node_t *node) {
1873 ir_node *op = node->node;
1874 node_t *a = get_irn_node(get_binop_left(op));
1875 node_t *b = get_irn_node(get_binop_right(op));
1876 ir_mode *mode = get_irn_mode(op);
1879 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
1880 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
1883 /* node: no input should be tarval_top, else the binop would be also
1884 * Top and not being split. */
1885 zero = get_mode_null(mode);
1886 if (a->type.tv == zero)
1888 if (b->type.tv == zero)
1891 } /* identity_comm_zero_binop */
1893 #define identity_Add identity_comm_zero_binop
1894 #define identity_Or identity_comm_zero_binop
1897 * Calculates the Identity for Mul nodes.
1899 static node_t *identity_Mul(node_t *node) {
1900 ir_node *op = node->node;
1901 node_t *a = get_irn_node(get_Mul_left(op));
1902 node_t *b = get_irn_node(get_Mul_right(op));
1903 ir_mode *mode = get_irn_mode(op);
1906 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
1907 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
1910 /* node: no input should be tarval_top, else the binop would be also
1911 * Top and not being split. */
1912 one = get_mode_one(mode);
1913 if (a->type.tv == one)
1915 if (b->type.tv == one)
1918 } /* identity_Mul */
1921 * Calculates the Identity for Sub nodes.
1923 static node_t *identity_Sub(node_t *node) {
1924 ir_node *sub = node->node;
1925 node_t *b = get_irn_node(get_Sub_right(sub));
1926 ir_mode *mode = get_irn_mode(sub);
1928 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
1929 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
1932 /* node: no input should be tarval_top, else the binop would be also
1933 * Top and not being split. */
1934 if (b->type.tv == get_mode_null(mode))
1935 return get_irn_node(get_Sub_left(sub));
1937 } /* identity_Mul */
1940 * Calculates the Identity for And nodes.
1942 static node_t *identity_And(node_t *node) {
1943 ir_node *and = node->node;
1944 node_t *a = get_irn_node(get_And_left(and));
1945 node_t *b = get_irn_node(get_And_right(and));
1946 tarval *neutral = get_mode_all_one(get_irn_mode(and));
1948 /* node: no input should be tarval_top, else the And would be also
1949 * Top and not being split. */
1950 if (a->type.tv == neutral)
1952 if (b->type.tv == neutral)
1955 } /* identity_And */
1958 * Calculates the Identity for Confirm nodes.
1960 static node_t *identity_Confirm(node_t *node) {
1961 ir_node *confirm = node->node;
1963 /* a Confirm is always a Copy */
1964 return get_irn_node(get_Confirm_value(confirm));
1965 } /* identity_Confirm */
1968 * Calculates the Identity for Mux nodes.
1970 static node_t *identity_Mux(node_t *node) {
1971 ir_node *mux = node->node;
1972 node_t *sel = get_irn_node(get_Mux_sel(mux));
1973 node_t *t = get_irn_node(get_Mux_true(mux));
1974 node_t *f = get_irn_node(get_Mux_false(mux));
1976 if (t->part == f->part)
1979 /* Mux sel input is mode_b, so it is always a tarval */
1980 if (sel->type.tv == tarval_b_true)
1982 if (sel->type.tv == tarval_b_false)
1985 } /* identity_Mux */
1988 * Calculates the Identity for Min nodes.
1990 static node_t *identity_Min(node_t *node) {
1991 ir_node *op = node->node;
1992 node_t *a = get_irn_node(get_binop_left(op));
1993 node_t *b = get_irn_node(get_binop_right(op));
1994 ir_mode *mode = get_irn_mode(op);
1997 if (a->part == b->part) {
1998 /* leader of multiple predecessors */
2002 /* works even with NaN */
2003 tv_max = get_mode_max(mode);
2004 if (a->type.tv == tv_max)
2006 if (b->type.tv == tv_max)
2009 } /* identity_Min */
2012 * Calculates the Identity for Max nodes.
2014 static node_t *identity_Max(node_t *node) {
2015 ir_node *op = node->node;
2016 node_t *a = get_irn_node(get_binop_left(op));
2017 node_t *b = get_irn_node(get_binop_right(op));
2018 ir_mode *mode = get_irn_mode(op);
2021 if (a->part == b->part) {
2022 /* leader of multiple predecessors */
2026 /* works even with NaN */
2027 tv_min = get_mode_min(mode);
2028 if (a->type.tv == tv_min)
2030 if (b->type.tv == tv_min)
2033 } /* identity_Max */
2036 * Calculates the Identity for nodes.
2038 static node_t *identity(node_t *node) {
2039 ir_node *irn = node->node;
2041 switch (get_irn_opcode(irn)) {
2043 return identity_Phi(node);
2045 return identity_Add(node);
2047 return identity_Or(node);
2049 return identity_Sub(node);
2051 return identity_Add(node);
2053 return identity_Confirm(node);
2055 return identity_Mux(node);
2057 return identity_Min(node);
2059 return identity_Max(node);
2066 * Node follower is a (new) follower of leader, segregate Leader
2069 static void segregate_def_use_chain_1(const ir_node *follower, node_t *leader) {
2070 ir_node *l = leader->node;
2071 int j, i, n = get_irn_n_outs(l);
2073 DB((dbg, LEVEL_2, "%+F is a follower of %+F\n", follower, leader->node));
2074 /* The leader edges must remain sorted, but follower edges can
2076 for (i = leader->n_followers + 1; i <= n; ++i) {
2077 if (l->out[i].use == follower) {
2078 ir_def_use_edge t = l->out[i];
2080 for (j = i - 1; j >= leader->n_followers + 1; --j)
2081 l->out[j + 1] = l->out[j];
2082 ++leader->n_followers;
2083 l->out[leader->n_followers] = t;
2085 /* note: a node might be a n-fold follower, for instance
2086 * if x = max(a,a), so no break here. */
2089 } /* segregate_def_use_chain_1 */
2092 * Node follower is a (new) follower of leader, segregate Leader
2093 * out edges. If follower is a n-congruent Input identity, all follower
2094 * inputs congruent to follower are also leader.
2096 static void segregate_def_use_chain(const ir_node *follower, node_t *leader) {
2097 ir_op *op = get_irn_op(follower);
2100 /* n-Congruent Input Identity for Phi's */
2102 ir_node *block = get_nodes_block(follower);
2104 DB((dbg, LEVEL_2, "n-Congruent follower %+F\n", follower));
2105 for (i = get_irn_arity(follower) - 1; i >= 0; --i) {
2106 node_t *pred_X = get_irn_node(get_Block_cfgpred(block, i));
2108 /* beware: we are NOT followers of dead inputs */
2109 if (pred_X->type.tv == tarval_reachable) {
2110 node_t *pred = get_irn_node(get_irn_n(follower, i));
2112 if (pred->part == leader->part)
2113 segregate_def_use_chain_1(follower, pred);
2116 } else if (op == op_Mux || op == op_Max || op == op_Min) {
2117 /* n-Congruent Input Identity */
2120 DB((dbg, LEVEL_2, "n-Congruent follower %+F\n", follower));
2121 for (i = get_irn_arity(follower) - 1; i >= 0; --i) {
2122 node_t *pred = get_irn_node(get_irn_n(follower, i));
2124 if (pred->part == leader->part)
2125 segregate_def_use_chain_1(follower, pred);
2128 /* 1-Congruent Input Identity */
2129 segregate_def_use_chain_1(follower, leader);
2131 } /* segregate_def_use_chain */
2134 * Make all inputs to x from inside X no longer be F.def_use edges.
2136 static void move_edges_to_leader(node_t *x) {
2137 partition_t *X = x->part;
2138 ir_node *irn = x->node;
2141 for (i = get_irn_arity(irn) - 1; i >= 0; --i) {
2142 node_t *pred = get_irn_node(get_irn_n(irn, i));
2147 n = get_irn_n_outs(p);
2148 for (j = 1; j <= pred->n_followers; ++j) {
2149 if (p->out[j].pos == i && p->out[j].use == irn) {
2150 /* found a follower edge to x, move it to the Leader */
2151 ir_def_use_edge edge = p->out[j];
2153 /* remove this edge from the Follower set */
2154 p->out[j] = p->out[pred->n_followers];
2155 --pred->n_followers;
2157 /* sort it into the leader set */
2158 for (k = pred->n_followers + 2; k <= n; ++k) {
2159 if (p->out[k].pos >= edge.pos)
2161 p->out[k - 1] = p->out[k];
2163 /* place the new edge here */
2164 p->out[k - 1] = edge;
2166 /* edge found and moved */
2174 * Propagate constant evaluation.
2176 * @param env the environment
2178 static void propagate(environment_t *env) {
2181 lattice_elem_t old_type;
2183 unsigned n_fallen, old_type_was_T_or_C;
2186 while (env->cprop != NULL) {
2187 void *oldopcode = NULL;
2189 /* remove the first partition X from cprop */
2192 env->cprop = X->cprop_next;
2194 old_type_was_T_or_C = X->type_is_T_or_C;
2196 DB((dbg, LEVEL_2, "Propagate type on part%d\n", X->nr));
2199 while (! list_empty(&X->cprop)) {
2200 /* remove the first Node x from X.cprop */
2201 x = list_entry(X->cprop.next, node_t, cprop_list);
2202 list_del(&x->cprop_list);
2205 if (x->is_follower && identity(x) == x) {
2206 /* x will make the follower -> leader transition */
2207 DB((dbg, LEVEL_2, "%+F make the follower -> leader transition\n", x->node));
2208 if (oldopcode == NULL) {
2209 oldopcode = lambda_opcode(get_first_node(X), env);
2211 if (oldopcode != lambda_opcode(x, env)) {
2212 /* different opcode -> x falls out of this partition */
2217 DB((dbg, LEVEL_2, "Add node %+F to fallen\n", x->node));
2220 /* move x from X.Follower to X.Leader */
2221 list_del(&x->node_list);
2222 list_add_tail(&x->node_list, &X->Leader);
2226 /* Make all inputs to x from inside X no longer be F.def_use edges */
2227 move_edges_to_leader(x);
2230 /* compute a new type for x */
2232 DB((dbg, LEVEL_3, "computing type of %+F\n", x->node));
2234 if (x->type.tv != old_type.tv) {
2235 verify_type(old_type, x->type);
2236 DB((dbg, LEVEL_2, "node %+F has changed type from %+F to %+F\n", x->node, old_type, x->type));
2238 if (x->on_fallen == 0) {
2239 /* Add x to fallen. Nodes might fall from T -> const -> _|_, so check that they are
2240 not already on the list. */
2245 DB((dbg, LEVEL_2, "Add node %+F to fallen\n", x->node));
2247 for (i = get_irn_n_outs(x->node) - 1; i >= 0; --i) {
2248 ir_node *succ = get_irn_out(x->node, i);
2249 node_t *y = get_irn_node(succ);
2251 /* Add y to y.partition.cprop. */
2252 add_node_to_cprop(y, env);
2257 if (n_fallen > 0 && n_fallen != X->n_leader) {
2258 DB((dbg, LEVEL_2, "Splitting part%d by fallen\n", X->nr));
2259 Y = split(X, fallen, env);
2263 /* remove the flags from the fallen list */
2264 for (x = fallen; x != NULL; x = x->next)
2268 if (old_type_was_T_or_C) {
2271 /* check if some nodes will make the leader -> follower transition */
2272 list_for_each_entry_safe(node_t, y, tmp, &Y->Leader, node_list) {
2273 if (! is_con(y->type)) {
2274 node_t *eq_node = identity(y);
2277 DB((dbg, LEVEL_2, "Node %+F is a follower of %+F\n", y->node, eq_node->node));
2278 /* move to Follower */
2280 list_del(&y->node_list);
2283 list_add_tail(&y->node_list, &Y->Follower);
2284 segregate_def_use_chain(y->node, eq_node);
2295 * Get the leader for a given node from its congruence class.
2297 * @param irn the node
2299 static ir_node *get_leader(node_t *node) {
2300 partition_t *part = node->part;
2302 if (part->n_leader > 1 || node->is_follower) {
2303 DB((dbg, LEVEL_2, "Found congruence class for %+F\n", node->node));
2305 return get_first_node(part)->node;
2311 * Return non-zero if the control flow predecessor node pred
2312 * is the only reachable control flow exit of its block.
2314 * @param pred the control flow exit
2316 static int can_exchange(ir_node *pred) {
2319 else if (is_Jmp(pred))
2321 else if (get_irn_mode(pred) == mode_T) {
2324 /* if the predecessor block has more than one
2325 reachable outputs we cannot remove the block */
2327 for (i = get_irn_n_outs(pred) - 1; i >= 0; --i) {
2328 ir_node *proj = get_irn_out(pred, i);
2331 /* skip non-control flow Proj's */
2332 if (get_irn_mode(proj) != mode_X)
2335 node = get_irn_node(proj);
2336 if (node->type.tv == tarval_reachable) {
2347 * Block Post-Walker, apply the analysis results on control flow by
2348 * shortening Phi's and Block inputs.
2350 static void apply_cf(ir_node *block, void *ctx) {
2351 environment_t *env = ctx;
2352 node_t *node = get_irn_node(block);
2354 ir_node **ins, **in_X;
2355 ir_node *phi, *next;
2357 if (block == get_irg_end_block(current_ir_graph) ||
2358 block == get_irg_start_block(current_ir_graph)) {
2359 /* the EndBlock is always reachable even if the analysis
2360 finds out the opposite :-) */
2363 if (node->type.tv == tarval_unreachable) {
2364 /* mark dead blocks */
2365 set_Block_dead(block);
2369 n = get_Block_n_cfgpreds(block);
2372 /* only one predecessor combine */
2373 ir_node *pred = skip_Proj(get_Block_cfgpred(block, 0));
2375 if (can_exchange(pred)) {
2376 exchange(block, get_nodes_block(pred));
2382 NEW_ARR_A(ir_node *, in_X, n);
2384 for (i = 0; i < n; ++i) {
2385 ir_node *pred = get_Block_cfgpred(block, i);
2386 node_t *node = get_irn_node(pred);
2388 if (node->type.tv == tarval_reachable) {
2395 NEW_ARR_A(ir_node *, ins, n);
2396 for (phi = get_Block_phis(block); phi != NULL; phi = next) {
2397 node_t *node = get_irn_node(phi);
2399 next = get_Phi_next(phi);
2400 if (is_tarval(node->type.tv) && tarval_is_constant(node->type.tv)) {
2401 /* this Phi is replaced by a constant */
2402 tarval *tv = node->type.tv;
2403 ir_node *c = new_r_Const(current_ir_graph, block, get_tarval_mode(tv), tv);
2405 set_irn_node(c, node);
2407 DB((dbg, LEVEL_1, "%+F is replaced by %+F\n", phi, c));
2412 for (i = 0; i < n; ++i) {
2413 node_t *pred = get_irn_node(get_Block_cfgpred(block, i));
2415 if (pred->type.tv == tarval_reachable) {
2416 ins[j++] = get_Phi_pred(phi, i);
2420 /* this Phi is replaced by a single predecessor */
2421 ir_node *s = ins[0];
2424 DB((dbg, LEVEL_1, "%+F is replaced by %+F because of cf change\n", phi, s));
2428 set_irn_in(phi, j, ins);
2435 /* this Block has only one live predecessor */
2436 ir_node *pred = skip_Proj(in_X[0]);
2438 if (can_exchange(pred)) {
2439 exchange(block, get_nodes_block(pred));
2443 set_irn_in(block, k, in_X);
2449 * Post-Walker, apply the analysis results;
2451 static void apply_result(ir_node *irn, void *ctx) {
2452 environment_t *env = ctx;
2453 node_t *node = get_irn_node(irn);
2455 if (is_Block(irn) || is_End(irn) || is_Bad(irn)) {
2456 /* blocks already handled, do not touch the End node */
2458 node_t *block = get_irn_node(get_nodes_block(irn));
2460 if (block->type.tv == tarval_unreachable) {
2461 ir_node *bad = get_irg_bad(current_ir_graph);
2463 /* here, bad might already have a node, but this can be safely ignored
2464 as long as bad has at least ONE valid node */
2465 set_irn_node(bad, node);
2467 DB((dbg, LEVEL_1, "%+F is unreachable\n", irn));
2471 else if (node->type.tv == tarval_unreachable) {
2472 ir_node *bad = get_irg_bad(current_ir_graph);
2474 /* see comment above */
2475 set_irn_node(bad, node);
2477 DB((dbg, LEVEL_1, "%+F is unreachable\n", irn));
2481 else if (get_irn_mode(irn) == mode_X) {
2484 ir_node *cond = get_Proj_pred(irn);
2486 if (is_Cond(cond)) {
2487 node_t *sel = get_irn_node(get_Cond_selector(cond));
2489 if (is_tarval(sel->type.tv) && tarval_is_constant(sel->type.tv)) {
2490 /* Cond selector is a constant, make a Jmp */
2491 ir_node *jmp = new_r_Jmp(current_ir_graph, block->node);
2492 set_irn_node(jmp, node);
2494 DB((dbg, LEVEL_1, "%+F is replaced by %+F\n", irn, jmp));
2501 /* normal data node */
2502 if (is_tarval(node->type.tv) && tarval_is_constant(node->type.tv)) {
2503 tarval *tv = node->type.tv;
2506 * Beware: never replace mode_T nodes by constants. Currently we must mark
2507 * mode_T nodes with constants, but do NOT replace them.
2509 if (! is_Const(irn) && get_irn_mode(irn) != mode_T) {
2510 /* can be replaced by a constant */
2511 ir_node *c = new_r_Const(current_ir_graph, block->node, get_tarval_mode(tv), tv);
2512 set_irn_node(c, node);
2514 DB((dbg, LEVEL_1, "%+F is replaced by %+F\n", irn, c));
2518 } else if (is_entity(node->type.sym.entity_p)) {
2519 if (! is_SymConst(irn)) {
2520 /* can be replaced by a Symconst */
2521 ir_node *symc = new_r_SymConst(current_ir_graph, block->node, get_irn_mode(irn), node->type.sym, symconst_addr_ent);
2522 set_irn_node(symc, node);
2525 DB((dbg, LEVEL_1, "%+F is replaced by %+F\n", irn, symc));
2526 exchange(irn, symc);
2530 ir_node *leader = get_leader(node);
2532 if (leader != irn) {
2533 DB((dbg, LEVEL_1, "%+F from part%d is replaced by %+F\n", irn, node->part->nr, leader));
2534 exchange(irn, leader);
2540 } /* apply_result */
2542 #define SET(code) op_##code->ops.generic = (op_func)compute_##code
2545 * sets the generic functions to compute.
2547 static void set_compute_functions(void) {
2550 /* set the default compute function */
2551 for (i = get_irp_n_opcodes() - 1; i >= 0; --i) {
2552 ir_op *op = get_irp_opcode(i);
2553 op->ops.generic = (op_func)default_compute;
2556 /* set specific functions */
2575 } /* set_compute_functions */
2577 static int dump_partition_hook(FILE *F, ir_node *n, ir_node *local) {
2578 ir_node *irn = local != NULL ? local : n;
2579 node_t *node = get_irn_node(irn);
2581 ir_fprintf(F, "info2 : \"partition %u type %+F\"\n", node->part->nr, node->type);
2585 void combo(ir_graph *irg) {
2587 ir_node *initial_bl;
2589 ir_graph *rem = current_ir_graph;
2591 current_ir_graph = irg;
2593 /* register a debug mask */
2594 FIRM_DBG_REGISTER(dbg, "firm.opt.combo");
2595 //firm_dbg_set_mask(dbg, SET_LEVEL_3);
2597 DB((dbg, LEVEL_1, "Doing COMBO for %+F\n", irg));
2599 obstack_init(&env.obst);
2600 env.worklist = NULL;
2604 #ifdef DEBUG_libfirm
2605 env.dbg_list = NULL;
2607 env.opcode2id_map = new_set(cmp_opcode, iro_Last * 4);
2608 env.type2id_map = pmap_create();
2609 env.end_idx = get_opt_global_cse() ? 0 : -1;
2610 env.lambda_input = 0;
2613 assure_irg_outs(irg);
2615 /* we have our own value_of function */
2616 set_value_of_func(get_node_tarval);
2618 set_compute_functions();
2619 DEBUG_ONLY(part_nr = 0);
2621 /* create the initial partition and place it on the work list */
2622 env.initial = new_partition(&env);
2623 add_to_worklist(env.initial, &env);
2624 irg_walk_graph(irg, init_block_phis, create_initial_partitions, &env);
2626 /* all nodes on the initial partition have type Top */
2627 env.initial->type_is_T_or_C = 1;
2629 /* Place the START Node's partition on cprop.
2630 Place the START Node on its local worklist. */
2631 initial_bl = get_irg_start_block(irg);
2632 start = get_irn_node(initial_bl);
2633 add_node_to_cprop(start, &env);
2637 if (env.worklist != NULL)
2639 } while (env.cprop != NULL || env.worklist != NULL);
2641 dump_all_partitions(&env);
2644 set_dump_node_vcgattr_hook(dump_partition_hook);
2645 dump_ir_block_graph(irg, "-partition");
2646 set_dump_node_vcgattr_hook(NULL);
2648 (void)dump_partition_hook;
2651 /* apply the result */
2652 irg_block_walk_graph(irg, NULL, apply_cf, &env);
2653 irg_walk_graph(irg, NULL, apply_result, &env);
2656 /* control flow might changed */
2657 set_irg_outs_inconsistent(irg);
2658 set_irg_extblk_inconsistent(irg);
2659 set_irg_doms_inconsistent(irg);
2660 set_irg_loopinfo_inconsistent(irg);
2663 pmap_destroy(env.type2id_map);
2664 del_set(env.opcode2id_map);
2665 obstack_free(&env.obst, NULL);
2667 /* restore value_of() default behavior */
2668 set_value_of_func(NULL);
2669 current_ir_graph = rem;