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 * This is a slightly enhanced version of Cliff Clicks combo algorithm
27 * - support for commutative nodes is added, Add(a,b) and Add(b,a) ARE congruent
28 * - supports all Firm direct (by a data edge) identities except Mux
29 * (Mux can be a 2-input or 1-input identity, only 2-input is implemented yet)
30 * - supports Confirm nodes (handle them like Copies but do NOT remove them)
31 * - let Cmp nodes calculate Top like all othe data nodes: this would let
32 * Mux nodes to calculate Unknown instead of taking the true result
33 * - let Cond(Top) always select FALSE/default: This is tricky. Nodes are only reavaluated
34 * IFF the predecessor changed its type. Because nodes are initialized with Top
35 * this never happens, let all Proj(Cond) be unreachable.
36 * We avoid this condition by the same way we work around Phi: whenever a Block
37 * node is placed on the list, place its Cond nodes (and because they are Tuple
38 * all its Proj-nodes either on the cprop list)
39 * Especially, this changes the meaning of Click's example:
54 * using Click's version while is silent with our.
55 * - support for global congruences is implemented but not tested yet
57 * Note further that we use the terminology from Click's work here, which is different
58 * in some cases from Firm terminology. Especially, Click's type is a
59 * Firm tarval/entity, nevertheless we call it type here for "maximum compatibility".
65 #include "iroptimize.h"
73 #include "irgraph_t.h"
80 #include "iropt_dbg.h"
90 /* define this to check that all type translations are monotone */
91 #define VERIFY_MONOTONE
93 /* define this to check the consistency of partitions */
94 #define CHECK_PARTITIONS
97 /* allow optimization of non-strict programs */
100 typedef struct node_t node_t;
101 typedef struct partition_t partition_t;
102 typedef struct opcode_key_t opcode_key_t;
103 typedef struct listmap_entry_t listmap_entry_t;
105 /** The type of the compute function. */
106 typedef void (*compute_func)(node_t *node);
111 struct opcode_key_t {
112 ir_opcode code; /**< The Firm opcode. */
113 ir_mode *mode; /**< The mode of all nodes in the partition. */
114 int arity; /**< The arity of this opcode (needed for Phi etc. */
116 long proj; /**< For Proj nodes, its proj number */
117 ir_entity *ent; /**< For Sel Nodes, its entity */
122 * An entry in the list_map.
124 struct listmap_entry_t {
125 void *id; /**< The id. */
126 node_t *list; /**< The associated list for this id. */
127 listmap_entry_t *next; /**< Link to the next entry in the map. */
130 /** We must map id's to lists. */
131 typedef struct listmap_t {
132 set *map; /**< Map id's to listmap_entry_t's */
133 listmap_entry_t *values; /**< List of all values in the map. */
137 * A lattice element. Because we handle constants and symbolic constants different, we
138 * have to use this union.
149 ir_node *node; /**< The IR-node itself. */
150 list_head node_list; /**< Double-linked list of leader/follower entries. */
151 list_head cprop_list; /**< Double-linked partition.cprop list. */
152 partition_t *part; /**< points to the partition this node belongs to */
153 node_t *next; /**< Next node on local list (partition.touched, fallen). */
154 node_t *race_next; /**< Next node on race list. */
155 lattice_elem_t type; /**< The associated lattice element "type". */
156 int max_user_input; /**< Maximum input number of Def-Use edges. */
157 int next_edge; /**< Index of the next Def-Use edge to use. */
158 int n_followers; /**< Number of Follower in the outs set. */
159 unsigned on_touched:1; /**< Set, if this node is on the partition.touched set. */
160 unsigned on_cprop:1; /**< Set, if this node is on the partition.cprop list. */
161 unsigned on_fallen:1; /**< Set, if this node is on the fallen list. */
162 unsigned is_follower:1; /**< Set, if this node is a follower. */
163 unsigned by_all_const:1; /**< Set, if this node was once evaluated by all constants. */
164 unsigned flagged:2; /**< 2 Bits, set if this node was visited by race 1 or 2. */
168 * A partition containing congruent nodes.
171 list_head Leader; /**< The head of partition Leader node list. */
172 list_head Follower; /**< The head of partition Follower node list. */
173 list_head cprop; /**< The head of partition.cprop list. */
174 list_head cprop_X; /**< The head of partition.cprop (Cond nodes and its Projs) list. */
175 partition_t *wl_next; /**< Next entry in the work list if any. */
176 partition_t *touched_next; /**< Points to the next partition in the touched set. */
177 partition_t *cprop_next; /**< Points to the next partition in the cprop list. */
178 partition_t *split_next; /**< Points to the next partition in the list that must be split by split_by(). */
179 node_t *touched; /**< The partition.touched set of this partition. */
180 unsigned n_leader; /**< Number of entries in this partition.Leader. */
181 unsigned n_touched; /**< Number of entries in the partition.touched. */
182 int max_user_inputs; /**< Maximum number of user inputs of all entries. */
183 unsigned on_worklist:1; /**< Set, if this partition is in the work list. */
184 unsigned on_touched:1; /**< Set, if this partition is on the touched set. */
185 unsigned on_cprop:1; /**< Set, if this partition is on the cprop list. */
186 unsigned type_is_T_or_C:1;/**< Set, if all nodes in this partition have type Top or Constant. */
188 partition_t *dbg_next; /**< Link all partitions for debugging */
189 unsigned nr; /**< A unique number for (what-)mapping, >0. */
193 typedef struct environment_t {
194 struct obstack obst; /**< obstack to allocate data structures. */
195 partition_t *worklist; /**< The work list. */
196 partition_t *cprop; /**< The constant propagation list. */
197 partition_t *touched; /**< the touched set. */
198 partition_t *initial; /**< The initial partition. */
199 set *opcode2id_map; /**< The opcodeMode->id map. */
200 pmap *type2id_map; /**< The type->id map. */
201 int end_idx; /**< -1 for local and 0 for global congruences. */
202 int lambda_input; /**< Captured argument for lambda_partition(). */
203 char modified; /**< Set, if the graph was modified. */
204 char commutative; /**< Set, if commutation nodes should be handled specially. */
206 partition_t *dbg_list; /**< List of all partitions. */
210 /** Type of the what function. */
211 typedef void *(*what_func)(const node_t *node, environment_t *env);
213 #define get_irn_node(follower) ((node_t *)get_irn_link(follower))
214 #define set_irn_node(follower, node) set_irn_link(follower, node)
216 /* we do NOT use tarval_unreachable here, instead we use Top for this purpose */
217 #undef tarval_unreachable
218 #define tarval_unreachable tarval_top
221 /** The debug module handle. */
222 DEBUG_ONLY(static firm_dbg_module_t *dbg;)
224 /** The what reason. */
225 DEBUG_ONLY(static const char *what_reason;)
227 /** Next partition number. */
228 DEBUG_ONLY(static unsigned part_nr = 0);
230 /** The tarval returned by Unknown nodes. */
232 #define tarval_UNKNOWN tarval_top
234 #define tarval_UNKNOWN tarval_bad
238 static node_t *identity(node_t *node);
240 #ifdef CHECK_PARTITIONS
244 static void check_partition(const partition_t *T) {
248 list_for_each_entry(node_t, node, &T->Leader, node_list) {
249 assert(node->is_follower == 0);
250 assert(node->flagged == 0);
251 assert(node->part == T);
254 assert(n == T->n_leader);
256 list_for_each_entry(node_t, node, &T->Follower, node_list) {
257 assert(node->is_follower == 1);
258 assert(node->flagged == 0);
259 assert(node->part == T);
261 } /* check_partition */
264 * check that all leader nodes in the partition have the same opcode.
266 static void check_opcode(const partition_t *Z) {
271 list_for_each_entry(node_t, node, &Z->Leader, node_list) {
272 ir_node *irn = node->node;
275 key.code = get_irn_opcode(irn);
276 key.mode = get_irn_mode(irn);
277 key.arity = get_irn_arity(irn);
281 switch (get_irn_opcode(irn)) {
283 key.u.proj = get_Proj_proj(irn);
286 key.u.ent = get_Sel_entity(irn);
293 assert(key.code == get_irn_opcode(irn));
294 assert(key.mode == get_irn_mode(irn));
295 assert(key.arity == get_irn_arity(irn));
297 switch (get_irn_opcode(irn)) {
299 assert(key.u.proj == get_Proj_proj(irn));
302 assert(key.u.ent == get_Sel_entity(irn));
311 static void check_all_partitions(environment_t *env) {
316 for (P = env->dbg_list; P != NULL; P = P->dbg_next) {
318 if (! P->type_is_T_or_C)
320 list_for_each_entry(node_t, node, &P->Follower, node_list) {
321 node_t *leader = identity(node);
323 assert(leader != node && leader->part == node->part);
332 static void do_check_list(const node_t *list, int ofs, const partition_t *Z) {
335 #define NEXT(e) *((const node_t **)((char *)(e) + (ofs)))
336 for (e = list; e != NULL; e = NEXT(e)) {
337 assert(e->part == Z);
340 } /* ido_check_list */
343 * Check a local list.
345 static void check_list(const node_t *list, const partition_t *Z) {
346 do_check_list(list, offsetof(node_t, next), Z);
350 #define check_partition(T)
351 #define check_list(list, Z)
352 #define check_all_partitions(env)
353 #endif /* CHECK_PARTITIONS */
356 static inline lattice_elem_t get_partition_type(const partition_t *X);
359 * Dump partition to output.
361 static void dump_partition(const char *msg, const partition_t *part) {
364 lattice_elem_t type = get_partition_type(part);
366 DB((dbg, LEVEL_2, "%s part%u%s (%u, %+F) {\n ",
367 msg, part->nr, part->type_is_T_or_C ? "*" : "",
368 part->n_leader, type));
369 list_for_each_entry(node_t, node, &part->Leader, node_list) {
370 DB((dbg, LEVEL_2, "%s%+F", first ? "" : ", ", node->node));
373 if (! list_empty(&part->Follower)) {
374 DB((dbg, LEVEL_2, "\n---\n "));
376 list_for_each_entry(node_t, node, &part->Follower, node_list) {
377 DB((dbg, LEVEL_2, "%s%+F", first ? "" : ", ", node->node));
381 DB((dbg, LEVEL_2, "\n}\n"));
382 } /* dump_partition */
387 static void do_dump_list(const char *msg, const node_t *node, int ofs) {
391 #define GET_LINK(p, ofs) *((const node_t **)((char *)(p) + (ofs)))
393 DB((dbg, LEVEL_3, "%s = {\n ", msg));
394 for (p = node; p != NULL; p = GET_LINK(p, ofs)) {
395 DB((dbg, LEVEL_3, "%s%+F", first ? "" : ", ", p->node));
398 DB((dbg, LEVEL_3, "\n}\n"));
406 static void dump_race_list(const char *msg, const node_t *list) {
407 do_dump_list(msg, list, offsetof(node_t, race_next));
408 } /* dump_race_list */
411 * Dumps a local list.
413 static void dump_list(const char *msg, const node_t *list) {
414 do_dump_list(msg, list, offsetof(node_t, next));
418 * Dump all partitions.
420 static void dump_all_partitions(const environment_t *env) {
421 const partition_t *P;
423 DB((dbg, LEVEL_2, "All partitions\n===============\n"));
424 for (P = env->dbg_list; P != NULL; P = P->dbg_next)
425 dump_partition("", P);
426 } /* dump_all_partitions */
431 static void dump_split_list(const partition_t *list) {
432 const partition_t *p;
434 DB((dbg, LEVEL_2, "Split by %s produced = {\n", what_reason));
435 for (p = list; p != NULL; p = p->split_next)
436 DB((dbg, LEVEL_2, "part%u, ", p->nr));
437 DB((dbg, LEVEL_2, "\n}\n"));
438 } /* dump_split_list */
441 * Dump partition and type for a node.
443 static int dump_partition_hook(FILE *F, ir_node *n, ir_node *local) {
444 ir_node *irn = local != NULL ? local : n;
445 node_t *node = get_irn_node(irn);
447 ir_fprintf(F, "info2 : \"partition %u type %+F\"\n", node->part->nr, node->type);
449 } /* dump_partition_hook */
452 #define dump_partition(msg, part)
453 #define dump_race_list(msg, list)
454 #define dump_list(msg, list)
455 #define dump_all_partitions(env)
456 #define dump_split_list(list)
459 #if defined(VERIFY_MONOTONE) && defined (DEBUG_libfirm)
461 * Verify that a type transition is monotone
463 static void verify_type(const lattice_elem_t old_type, node_t *node) {
464 if (old_type.tv == node->type.tv) {
468 if (old_type.tv == tarval_top) {
469 /* from Top down-to is always allowed */
472 if (node->type.tv == tarval_bottom || node->type.tv == tarval_reachable) {
476 panic("combo: wrong translation from %+F to %+F on node %+F", old_type, node->type, node->node);
480 #define verify_type(old_type, node)
484 * Compare two pointer values of a listmap.
486 static int listmap_cmp_ptr(const void *elt, const void *key, size_t size) {
487 const listmap_entry_t *e1 = elt;
488 const listmap_entry_t *e2 = key;
491 return e1->id != e2->id;
492 } /* listmap_cmp_ptr */
495 * Initializes a listmap.
497 * @param map the listmap
499 static void listmap_init(listmap_t *map) {
500 map->map = new_set(listmap_cmp_ptr, 16);
505 * Terminates a listmap.
507 * @param map the listmap
509 static void listmap_term(listmap_t *map) {
514 * Return the associated listmap entry for a given id.
516 * @param map the listmap
517 * @param id the id to search for
519 * @return the associated listmap entry for the given id
521 static listmap_entry_t *listmap_find(listmap_t *map, void *id) {
522 listmap_entry_t key, *entry;
527 entry = set_insert(map->map, &key, sizeof(key), HASH_PTR(id));
529 if (entry->list == NULL) {
530 /* a new entry, put into the list */
531 entry->next = map->values;
538 * Calculate the hash value for an opcode map entry.
540 * @param entry an opcode map entry
542 * @return a hash value for the given opcode map entry
544 static unsigned opcode_hash(const opcode_key_t *entry) {
545 return (entry->mode - (ir_mode *)0) * 9 + entry->code + entry->u.proj * 3 + HASH_PTR(entry->u.ent) + entry->arity;
549 * Compare two entries in the opcode map.
551 static int cmp_opcode(const void *elt, const void *key, size_t size) {
552 const opcode_key_t *o1 = elt;
553 const opcode_key_t *o2 = key;
556 return o1->code != o2->code || o1->mode != o2->mode ||
557 o1->arity != o2->arity ||
558 o1->u.proj != o2->u.proj || o1->u.ent != o2->u.ent;
562 * Compare two Def-Use edges for input position.
564 static int cmp_def_use_edge(const void *a, const void *b) {
565 const ir_def_use_edge *ea = a;
566 const ir_def_use_edge *eb = b;
568 /* no overrun, because range is [-1, MAXINT] */
569 return ea->pos - eb->pos;
570 } /* cmp_def_use_edge */
573 * We need the Def-Use edges sorted.
575 static void sort_irn_outs(node_t *node) {
576 ir_node *irn = node->node;
577 int n_outs = get_irn_n_outs(irn);
580 qsort(&irn->out[1], n_outs, sizeof(irn->out[0]), cmp_def_use_edge);
582 node->max_user_input = irn->out[n_outs].pos;
583 } /* sort_irn_outs */
586 * Return the type of a node.
588 * @param irn an IR-node
590 * @return the associated type of this node
592 static inline lattice_elem_t get_node_type(const ir_node *irn) {
593 return get_irn_node(irn)->type;
594 } /* get_node_type */
597 * Return the tarval of a node.
599 * @param irn an IR-node
601 * @return the associated type of this node
603 static inline tarval *get_node_tarval(const ir_node *irn) {
604 lattice_elem_t type = get_node_type(irn);
606 if (is_tarval(type.tv))
608 return tarval_bottom;
609 } /* get_node_type */
612 * Add a partition to the worklist.
614 static inline void add_to_worklist(partition_t *X, environment_t *env) {
615 assert(X->on_worklist == 0);
616 DB((dbg, LEVEL_2, "Adding part%d to worklist\n", X->nr));
617 X->wl_next = env->worklist;
620 } /* add_to_worklist */
623 * Create a new empty partition.
625 * @param env the environment
627 * @return a newly allocated partition
629 static inline partition_t *new_partition(environment_t *env) {
630 partition_t *part = obstack_alloc(&env->obst, sizeof(*part));
632 INIT_LIST_HEAD(&part->Leader);
633 INIT_LIST_HEAD(&part->Follower);
634 INIT_LIST_HEAD(&part->cprop);
635 INIT_LIST_HEAD(&part->cprop_X);
636 part->wl_next = NULL;
637 part->touched_next = NULL;
638 part->cprop_next = NULL;
639 part->split_next = NULL;
640 part->touched = NULL;
643 part->max_user_inputs = 0;
644 part->on_worklist = 0;
645 part->on_touched = 0;
647 part->type_is_T_or_C = 0;
649 part->dbg_next = env->dbg_list;
650 env->dbg_list = part;
651 part->nr = part_nr++;
655 } /* new_partition */
658 * Get the first node from a partition.
660 static inline node_t *get_first_node(const partition_t *X) {
661 return list_entry(X->Leader.next, node_t, node_list);
662 } /* get_first_node */
665 * Return the type of a partition (assuming partition is non-empty and
666 * all elements have the same type).
668 * @param X a partition
670 * @return the type of the first element of the partition
672 static inline lattice_elem_t get_partition_type(const partition_t *X) {
673 const node_t *first = get_first_node(X);
675 } /* get_partition_type */
678 * Creates a partition node for the given IR-node and place it
679 * into the given partition.
681 * @param irn an IR-node
682 * @param part a partition to place the node in
683 * @param env the environment
685 * @return the created node
687 static node_t *create_partition_node(ir_node *irn, partition_t *part, environment_t *env) {
688 /* create a partition node and place it in the partition */
689 node_t *node = obstack_alloc(&env->obst, sizeof(*node));
691 INIT_LIST_HEAD(&node->node_list);
692 INIT_LIST_HEAD(&node->cprop_list);
696 node->race_next = NULL;
697 node->type.tv = tarval_top;
698 node->max_user_input = 0;
700 node->n_followers = 0;
701 node->on_touched = 0;
704 node->is_follower = 0;
705 node->by_all_const = 0;
707 set_irn_node(irn, node);
709 list_add_tail(&node->node_list, &part->Leader);
713 } /* create_partition_node */
716 * Pre-Walker, initialize all Nodes' type to U or top and place
717 * all nodes into the TOP partition.
719 static void create_initial_partitions(ir_node *irn, void *ctx) {
720 environment_t *env = ctx;
721 partition_t *part = env->initial;
724 node = create_partition_node(irn, part, env);
726 if (node->max_user_input > part->max_user_inputs)
727 part->max_user_inputs = node->max_user_input;
730 set_Block_phis(irn, NULL);
732 } /* create_initial_partitions */
735 * Post-Walker, collect all Block-Phi lists, set Cond.
737 static void init_block_phis(ir_node *irn, void *ctx) {
741 add_Block_phi(get_nodes_block(irn), irn);
743 } /* init_block_phis */
746 * Add a node to the entry.partition.touched set and
747 * node->partition to the touched set if not already there.
750 * @param env the environment
752 static inline void add_to_touched(node_t *y, environment_t *env) {
753 if (y->on_touched == 0) {
754 partition_t *part = y->part;
756 y->next = part->touched;
761 if (part->on_touched == 0) {
762 part->touched_next = env->touched;
764 part->on_touched = 1;
767 check_list(part->touched, part);
769 } /* add_to_touched */
772 * Place a node on the cprop list.
775 * @param env the environment
777 static void add_to_cprop(node_t *y, environment_t *env) {
780 /* Add y to y.partition.cprop. */
781 if (y->on_cprop == 0) {
782 partition_t *Y = y->part;
784 /* place Conds and its Proj nodes on the cprop_X list */
785 if (is_Cond(skip_Proj(y->node)))
786 list_add_tail(&y->cprop_list, &Y->cprop_X);
788 list_add_tail(&y->cprop_list, &Y->cprop);
791 DB((dbg, LEVEL_3, "Add %+F to part%u.cprop\n", y->node, Y->nr));
793 /* place its partition on the cprop list */
794 if (Y->on_cprop == 0) {
795 Y->cprop_next = env->cprop;
801 if (get_irn_mode(irn) == mode_T) {
802 /* mode_T nodes always produce tarval_bottom, so we must explicitly
803 add it's Proj's to get constant evaluation to work */
806 for (i = get_irn_n_outs(irn) - 1; i >= 0; --i) {
807 node_t *proj = get_irn_node(get_irn_out(irn, i));
809 add_to_cprop(proj, env);
811 } else if (is_Block(irn)) {
812 /* Due to the way we handle Phi's, we must place all Phis of a block on the list
813 * if someone placed the block. The Block is only placed if the reachability
814 * changes, and this must be re-evaluated in compute_Phi(). */
816 for (phi = get_Block_phis(irn); phi != NULL; phi = get_Phi_next(phi)) {
817 node_t *p = get_irn_node(phi);
818 add_to_cprop(p, env);
824 * Update the worklist: If Z is on worklist then add Z' to worklist.
825 * Else add the smaller of Z and Z' to worklist.
827 * @param Z the Z partition
828 * @param Z_prime the Z' partition, a previous part of Z
829 * @param env the environment
831 static void update_worklist(partition_t *Z, partition_t *Z_prime, environment_t *env) {
832 if (Z->on_worklist || Z_prime->n_leader < Z->n_leader) {
833 add_to_worklist(Z_prime, env);
835 add_to_worklist(Z, env);
837 } /* update_worklist */
840 * Make all inputs to x no longer be F.def_use edges.
844 static void move_edges_to_leader(node_t *x) {
845 ir_node *irn = x->node;
848 for (i = get_irn_arity(irn) - 1; i >= 0; --i) {
849 node_t *pred = get_irn_node(get_irn_n(irn, i));
854 n = get_irn_n_outs(p);
855 for (j = 1; j <= pred->n_followers; ++j) {
856 if (p->out[j].pos == i && p->out[j].use == irn) {
857 /* found a follower edge to x, move it to the Leader */
858 ir_def_use_edge edge = p->out[j];
860 /* remove this edge from the Follower set */
861 p->out[j] = p->out[pred->n_followers];
864 /* sort it into the leader set */
865 for (k = pred->n_followers + 2; k <= n; ++k) {
866 if (p->out[k].pos >= edge.pos)
868 p->out[k - 1] = p->out[k];
870 /* place the new edge here */
871 p->out[k - 1] = edge;
873 /* edge found and moved */
878 } /* move_edges_to_leader */
881 * Split a partition that has NO followers by a local list.
883 * @param Z partition to split
884 * @param g a (non-empty) node list
885 * @param env the environment
887 * @return a new partition containing the nodes of g
889 static partition_t *split_no_followers(partition_t *Z, node_t *g, environment_t *env) {
890 partition_t *Z_prime;
895 dump_partition("Splitting ", Z);
896 dump_list("by list ", g);
900 /* Remove g from Z. */
901 for (node = g; node != NULL; node = node->next) {
902 assert(node->part == Z);
903 list_del(&node->node_list);
906 assert(n < Z->n_leader);
909 /* Move g to a new partition, Z'. */
910 Z_prime = new_partition(env);
912 for (node = g; node != NULL; node = node->next) {
913 list_add_tail(&node->node_list, &Z_prime->Leader);
914 node->part = Z_prime;
915 if (node->max_user_input > max_input)
916 max_input = node->max_user_input;
918 Z_prime->max_user_inputs = max_input;
919 Z_prime->n_leader = n;
922 check_partition(Z_prime);
924 /* for now, copy the type info tag, it will be adjusted in split_by(). */
925 Z_prime->type_is_T_or_C = Z->type_is_T_or_C;
927 update_worklist(Z, Z_prime, env);
929 dump_partition("Now ", Z);
930 dump_partition("Created new ", Z_prime);
932 } /* split_no_followers */
935 * Make the Follower -> Leader transition for a node.
939 static void follower_to_leader(node_t *n) {
940 assert(n->is_follower == 1);
942 DB((dbg, LEVEL_2, "%+F make the follower -> leader transition\n", n->node));
944 move_edges_to_leader(n);
945 list_del(&n->node_list);
946 list_add_tail(&n->node_list, &n->part->Leader);
948 } /* follower_to_leader */
951 * The environment for one race step.
953 typedef struct step_env {
954 node_t *initial; /**< The initial node list. */
955 node_t *unwalked; /**< The unwalked node list. */
956 node_t *walked; /**< The walked node list. */
957 int index; /**< Next index of Follower use_def edge. */
958 unsigned side; /**< side number. */
962 * Return non-zero, if a input is a real follower
964 * @param irn the node to check
965 * @param input number of the input
967 static int is_real_follower(const ir_node *irn, int input) {
970 switch (get_irn_opcode(irn)) {
973 /* ignore the Confirm bound input */
979 /* ignore the Mux sel input */
984 /* dead inputs are not follower edges */
985 ir_node *block = get_nodes_block(irn);
986 node_t *pred = get_irn_node(get_Block_cfgpred(block, input));
988 if (pred->type.tv == tarval_unreachable)
998 /* only a Sub x,0 / Shift x,0 might be a follower */
1005 pred = get_irn_node(get_irn_n(irn, input));
1006 if (is_tarval(pred->type.tv) && tarval_is_null(pred->type.tv))
1010 pred = get_irn_node(get_irn_n(irn, input));
1011 if (is_tarval(pred->type.tv) && tarval_is_one(pred->type.tv))
1015 pred = get_irn_node(get_irn_n(irn, input));
1016 if (is_tarval(pred->type.tv) && tarval_is_all_one(pred->type.tv))
1021 /* all inputs are followers */
1024 assert(!"opcode not implemented yet");
1028 } /* is_real_follower */
1031 * Do one step in the race.
1033 static int step(step_env *env) {
1036 if (env->initial != NULL) {
1037 /* Move node from initial to unwalked */
1039 env->initial = n->race_next;
1041 n->race_next = env->unwalked;
1047 while (env->unwalked != NULL) {
1048 /* let n be the first node in unwalked */
1050 while (env->index < n->n_followers) {
1051 const ir_def_use_edge *edge = &n->node->out[1 + env->index];
1053 /* let m be n.F.def_use[index] */
1054 node_t *m = get_irn_node(edge->use);
1056 assert(m->is_follower);
1058 * Some inputs, like the get_Confirm_bound are NOT
1059 * real followers, sort them out.
1061 if (! is_real_follower(m->node, edge->pos)) {
1067 /* only followers from our partition */
1068 if (m->part != n->part)
1071 if ((m->flagged & env->side) == 0) {
1072 m->flagged |= env->side;
1074 if (m->flagged != 3) {
1075 /* visited the first time */
1076 /* add m to unwalked not as first node (we might still need to
1077 check for more follower node */
1078 m->race_next = n->race_next;
1082 /* else already visited by the other side and on the other list */
1085 /* move n to walked */
1086 env->unwalked = n->race_next;
1087 n->race_next = env->walked;
1095 * Clear the flags from a list and check for
1096 * nodes that where touched from both sides.
1098 * @param list the list
1100 static int clear_flags(node_t *list) {
1104 for (n = list; n != NULL; n = n->race_next) {
1105 if (n->flagged == 3) {
1106 /* we reach a follower from both sides, this will split congruent
1107 * inputs and make it a leader. */
1108 follower_to_leader(n);
1117 * Split a partition by a local list using the race.
1119 * @param pX pointer to the partition to split, might be changed!
1120 * @param gg a (non-empty) node list
1121 * @param env the environment
1123 * @return a new partition containing the nodes of gg
1125 static partition_t *split(partition_t **pX, node_t *gg, environment_t *env) {
1126 partition_t *X = *pX;
1127 partition_t *X_prime;
1130 node_t *g, *h, *node, *t;
1131 int max_input, transitions, winner, shf;
1133 DEBUG_ONLY(static int run = 0;)
1135 DB((dbg, LEVEL_2, "Run %d ", run++));
1136 if (list_empty(&X->Follower)) {
1137 /* if the partition has NO follower, we can use the fast
1138 splitting algorithm. */
1139 return split_no_followers(X, gg, env);
1141 /* else do the race */
1143 dump_partition("Splitting ", X);
1144 dump_list("by list ", gg);
1146 INIT_LIST_HEAD(&tmp);
1148 /* Remove gg from X.Leader and put into g */
1150 for (node = gg; node != NULL; node = node->next) {
1151 assert(node->part == X);
1152 assert(node->is_follower == 0);
1154 list_del(&node->node_list);
1155 list_add_tail(&node->node_list, &tmp);
1156 node->race_next = g;
1161 list_for_each_entry(node_t, node, &X->Leader, node_list) {
1162 node->race_next = h;
1165 /* restore X.Leader */
1166 list_splice(&tmp, &X->Leader);
1168 senv[0].initial = g;
1169 senv[0].unwalked = NULL;
1170 senv[0].walked = NULL;
1174 senv[1].initial = h;
1175 senv[1].unwalked = NULL;
1176 senv[1].walked = NULL;
1181 * Some informations on the race that are not stated clearly in Click's
1183 * 1) A follower stays on the side that reach him first.
1184 * 2) If the other side reches a follower, if will be converted to
1185 * a leader. /This must be done after the race is over, else the
1186 * edges we are iterating on are renumbered./
1187 * 3) /New leader might end up on both sides./
1188 * 4) /If one side ends up with new Leaders, we must ensure that
1189 * they can split out by opcode, hence we have to put _every_
1190 * partition with new Leader nodes on the cprop list, as
1191 * opcode splitting is done by split_by() at the end of
1192 * constant propagation./
1195 if (step(&senv[0])) {
1199 if (step(&senv[1])) {
1204 assert(senv[winner].initial == NULL);
1205 assert(senv[winner].unwalked == NULL);
1207 /* clear flags from walked/unwalked */
1209 transitions = clear_flags(senv[0].unwalked) << shf;
1210 transitions |= clear_flags(senv[0].walked) << shf;
1212 transitions |= clear_flags(senv[1].unwalked) << shf;
1213 transitions |= clear_flags(senv[1].walked) << shf;
1215 dump_race_list("winner ", senv[winner].walked);
1217 /* Move walked_{winner} to a new partition, X'. */
1218 X_prime = new_partition(env);
1221 for (node = senv[winner].walked; node != NULL; node = node->race_next) {
1222 list_del(&node->node_list);
1223 node->part = X_prime;
1224 if (node->is_follower) {
1225 list_add_tail(&node->node_list, &X_prime->Follower);
1227 list_add_tail(&node->node_list, &X_prime->Leader);
1230 if (node->max_user_input > max_input)
1231 max_input = node->max_user_input;
1233 X_prime->n_leader = n;
1234 X_prime->max_user_inputs = max_input;
1235 X->n_leader -= X_prime->n_leader;
1237 /* for now, copy the type info tag, it will be adjusted in split_by(). */
1238 X_prime->type_is_T_or_C = X->type_is_T_or_C;
1241 * Even if a follower was not checked by both sides, it might have
1242 * loose its congruence, so we need to check this case for all follower.
1244 list_for_each_entry_safe(node_t, node, t, &X_prime->Follower, node_list) {
1245 if (identity(node) == node) {
1246 follower_to_leader(node);
1252 check_partition(X_prime);
1254 /* X' is the smaller part */
1255 add_to_worklist(X_prime, env);
1258 * If there where follower to leader transitions, ensure that the nodes
1259 * can be split out if necessary.
1261 if (transitions & 1) {
1262 /* place winner partition on the cprop list */
1263 if (X_prime->on_cprop == 0) {
1264 X_prime->cprop_next = env->cprop;
1265 env->cprop = X_prime;
1266 X_prime->on_cprop = 1;
1269 if (transitions & 2) {
1270 /* place other partition on the cprop list */
1271 if (X->on_cprop == 0) {
1272 X->cprop_next = env->cprop;
1278 dump_partition("Now ", X);
1279 dump_partition("Created new ", X_prime);
1281 /* we have to ensure that the partition containing g is returned */
1291 * Returns non-zero if the i'th input of a Phi node is live.
1293 * @param phi a Phi-node
1294 * @param i an input number
1296 * @return non-zero if the i'th input of the given Phi node is live
1298 static int is_live_input(ir_node *phi, int i) {
1300 ir_node *block = get_nodes_block(phi);
1301 ir_node *pred = get_Block_cfgpred(block, i);
1302 lattice_elem_t type = get_node_type(pred);
1304 return type.tv != tarval_unreachable;
1306 /* else it's the control input, always live */
1308 } /* is_live_input */
1311 * Return non-zero if a type is a constant.
1313 static int is_constant_type(lattice_elem_t type) {
1314 if (type.tv != tarval_bottom && type.tv != tarval_top)
1317 } /* is_constant_type */
1320 * Check whether a type is neither Top or a constant.
1321 * Note: U is handled like Top here, R is a constant.
1323 * @param type the type to check
1325 static int type_is_neither_top_nor_const(const lattice_elem_t type) {
1326 if (is_tarval(type.tv)) {
1327 if (type.tv == tarval_top)
1329 if (tarval_is_constant(type.tv))
1336 } /* type_is_neither_top_nor_const */
1339 * Collect nodes to the touched list.
1341 * @param list the list which contains the nodes that must be evaluated
1342 * @param idx the index of the def_use edge to evaluate
1343 * @param env the environment
1345 static void collect_touched(list_head *list, int idx, environment_t *env) {
1347 int end_idx = env->end_idx;
1349 list_for_each_entry(node_t, x, list, node_list) {
1353 /* leader edges start AFTER follower edges */
1354 x->next_edge = x->n_followers + 1;
1356 num_edges = get_irn_n_outs(x->node);
1358 /* for all edges in x.L.def_use_{idx} */
1359 while (x->next_edge <= num_edges) {
1360 const ir_def_use_edge *edge = &x->node->out[x->next_edge];
1363 /* check if we have necessary edges */
1364 if (edge->pos > idx)
1371 /* only non-commutative nodes */
1372 if (env->commutative &&
1373 (idx == 0 || idx == 1) && is_op_commutative(get_irn_op(succ)))
1376 /* ignore the "control input" for non-pinned nodes
1377 if we are running in GCSE mode */
1378 if (idx < end_idx && get_irn_pinned(succ) != op_pin_state_pinned)
1381 y = get_irn_node(succ);
1382 assert(get_irn_n(succ, idx) == x->node);
1384 /* ignore block edges touching followers */
1385 if (idx == -1 && y->is_follower)
1388 if (is_constant_type(y->type)) {
1389 ir_opcode code = get_irn_opcode(succ);
1390 if (code == iro_Sub || code == iro_Cmp)
1391 add_to_cprop(y, env);
1394 /* Partitions of constants should not be split simply because their Nodes have unequal
1395 functions or incongruent inputs. */
1396 if (type_is_neither_top_nor_const(y->type) &&
1397 (! is_Phi(y->node) || is_live_input(y->node, idx))) {
1398 add_to_touched(y, env);
1402 } /* collect_touched */
1405 * Collect commutative nodes to the touched list.
1407 * @param list the list which contains the nodes that must be evaluated
1408 * @param env the environment
1410 static void collect_commutative_touched(list_head *list, environment_t *env) {
1413 list_for_each_entry(node_t, x, list, node_list) {
1416 num_edges = get_irn_n_outs(x->node);
1418 x->next_edge = x->n_followers + 1;
1420 /* for all edges in x.L.def_use_{idx} */
1421 while (x->next_edge <= num_edges) {
1422 const ir_def_use_edge *edge = &x->node->out[x->next_edge];
1425 /* check if we have necessary edges */
1435 /* only commutative nodes */
1436 if (!is_op_commutative(get_irn_op(succ)))
1439 y = get_irn_node(succ);
1440 if (is_constant_type(y->type)) {
1441 ir_opcode code = get_irn_opcode(succ);
1442 if (code == iro_Eor)
1443 add_to_cprop(y, env);
1446 /* Partitions of constants should not be split simply because their Nodes have unequal
1447 functions or incongruent inputs. */
1448 if (type_is_neither_top_nor_const(y->type)) {
1449 add_to_touched(y, env);
1453 } /* collect_commutative_touched */
1456 * Split the partitions if caused by the first entry on the worklist.
1458 * @param env the environment
1460 static void cause_splits(environment_t *env) {
1461 partition_t *X, *Z, *N;
1464 /* remove the first partition from the worklist */
1466 env->worklist = X->wl_next;
1469 dump_partition("Cause_split: ", X);
1471 if (env->commutative) {
1472 /* handle commutative nodes first */
1474 /* empty the touched set: already done, just clear the list */
1475 env->touched = NULL;
1477 collect_commutative_touched(&X->Leader, env);
1478 collect_commutative_touched(&X->Follower, env);
1480 for (Z = env->touched; Z != NULL; Z = N) {
1482 node_t *touched = Z->touched;
1483 unsigned n_touched = Z->n_touched;
1485 assert(Z->touched != NULL);
1487 /* beware, split might change Z */
1488 N = Z->touched_next;
1490 /* remove it from the touched set */
1493 /* Empty local Z.touched. */
1494 for (e = touched; e != NULL; e = e->next) {
1495 assert(e->is_follower == 0);
1501 if (0 < n_touched && n_touched < Z->n_leader) {
1502 DB((dbg, LEVEL_2, "Split part%d by touched\n", Z->nr));
1503 split(&Z, touched, env);
1505 assert(n_touched <= Z->n_leader);
1509 /* combine temporary leader and follower list */
1510 for (idx = -1; idx <= X->max_user_inputs; ++idx) {
1511 /* empty the touched set: already done, just clear the list */
1512 env->touched = NULL;
1514 collect_touched(&X->Leader, idx, env);
1515 collect_touched(&X->Follower, idx, env);
1517 for (Z = env->touched; Z != NULL; Z = N) {
1519 node_t *touched = Z->touched;
1520 unsigned n_touched = Z->n_touched;
1522 assert(Z->touched != NULL);
1524 /* beware, split might change Z */
1525 N = Z->touched_next;
1527 /* remove it from the touched set */
1530 /* Empty local Z.touched. */
1531 for (e = touched; e != NULL; e = e->next) {
1532 assert(e->is_follower == 0);
1538 if (0 < n_touched && n_touched < Z->n_leader) {
1539 DB((dbg, LEVEL_2, "Split part%d by touched\n", Z->nr));
1540 split(&Z, touched, env);
1542 assert(n_touched <= Z->n_leader);
1545 } /* cause_splits */
1548 * Implements split_by_what(): Split a partition by characteristics given
1549 * by the what function.
1551 * @param X the partition to split
1552 * @param What a function returning an Id for every node of the partition X
1553 * @param P a list to store the result partitions
1554 * @param env the environment
1558 static partition_t *split_by_what(partition_t *X, what_func What,
1559 partition_t **P, environment_t *env) {
1562 listmap_entry_t *iter;
1565 /* Let map be an empty mapping from the range of What to (local) list of Nodes. */
1567 list_for_each_entry(node_t, x, &X->Leader, node_list) {
1568 void *id = What(x, env);
1569 listmap_entry_t *entry;
1572 /* input not allowed, ignore */
1575 /* Add x to map[What(x)]. */
1576 entry = listmap_find(&map, id);
1577 x->next = entry->list;
1580 /* Let P be a set of Partitions. */
1582 /* for all sets S except one in the range of map do */
1583 for (iter = map.values; iter != NULL; iter = iter->next) {
1584 if (iter->next == NULL) {
1585 /* this is the last entry, ignore */
1590 /* Add SPLIT( X, S ) to P. */
1591 DB((dbg, LEVEL_2, "Split part%d by WHAT = %s\n", X->nr, what_reason));
1592 R = split(&X, S, env);
1602 } /* split_by_what */
1604 /** lambda n.(n.type) */
1605 static void *lambda_type(const node_t *node, environment_t *env) {
1607 return node->type.tv;
1610 /** lambda n.(n.opcode) */
1611 static void *lambda_opcode(const node_t *node, environment_t *env) {
1612 opcode_key_t key, *entry;
1613 ir_node *irn = node->node;
1615 key.code = get_irn_opcode(irn);
1616 key.mode = get_irn_mode(irn);
1617 key.arity = get_irn_arity(irn);
1621 switch (get_irn_opcode(irn)) {
1623 key.u.proj = get_Proj_proj(irn);
1626 key.u.ent = get_Sel_entity(irn);
1632 entry = set_insert(env->opcode2id_map, &key, sizeof(key), opcode_hash(&key));
1634 } /* lambda_opcode */
1636 /** lambda n.(n[i].partition) */
1637 static void *lambda_partition(const node_t *node, environment_t *env) {
1638 ir_node *skipped = skip_Proj(node->node);
1641 int i = env->lambda_input;
1643 if (i >= get_irn_arity(node->node)) {
1645 * We are outside the allowed range: This can happen even
1646 * if we have split by opcode first: doing so might move Followers
1647 * to Leaders and those will have a different opcode!
1648 * Note that in this case the partition is on the cprop list and will be
1654 /* ignore the "control input" for non-pinned nodes
1655 if we are running in GCSE mode */
1656 if (i < env->end_idx && get_irn_pinned(skipped) != op_pin_state_pinned)
1659 pred = i == -1 ? get_irn_n(skipped, i) : get_irn_n(node->node, i);
1660 p = get_irn_node(pred);
1663 } /* lambda_partition */
1665 /** lambda n.(n[i].partition) for commutative nodes */
1666 static void *lambda_commutative_partition(const node_t *node, environment_t *env) {
1667 ir_node *irn = node->node;
1668 ir_node *skipped = skip_Proj(irn);
1669 ir_node *pred, *left, *right;
1671 partition_t *pl, *pr;
1672 int i = env->lambda_input;
1674 if (i >= get_irn_arity(node->node)) {
1676 * We are outside the allowed range: This can happen even
1677 * if we have split by opcode first: doing so might move Followers
1678 * to Leaders and those will have a different opcode!
1679 * Note that in this case the partition is on the cprop list and will be
1685 /* ignore the "control input" for non-pinned nodes
1686 if we are running in GCSE mode */
1687 if (i < env->end_idx && get_irn_pinned(skipped) != op_pin_state_pinned)
1691 pred = get_irn_n(skipped, i);
1692 p = get_irn_node(pred);
1696 if (is_op_commutative(get_irn_op(irn))) {
1697 /* normalize partition order by returning the "smaller" on input 0,
1698 the "bigger" on input 1. */
1699 left = get_binop_left(irn);
1700 pl = get_irn_node(left)->part;
1701 right = get_binop_right(irn);
1702 pr = get_irn_node(right)->part;
1705 return pl < pr ? pl : pr;
1707 return pl > pr ? pl : pr;
1709 /* a not split out Follower */
1710 pred = get_irn_n(irn, i);
1711 p = get_irn_node(pred);
1715 } /* lambda_commutative_partition */
1718 * Returns true if a type is a constant (and NOT Top
1721 static int is_con(const lattice_elem_t type) {
1722 /* be conservative */
1723 if (is_tarval(type.tv))
1724 return tarval_is_constant(type.tv);
1725 return is_entity(type.sym.entity_p);
1729 * Implements split_by().
1731 * @param X the partition to split
1732 * @param env the environment
1734 static void split_by(partition_t *X, environment_t *env) {
1735 partition_t *I, *P = NULL;
1738 dump_partition("split_by", X);
1740 if (X->n_leader == 1) {
1741 /* we have only one leader, no need to split, just check it's type */
1742 node_t *x = get_first_node(X);
1743 X->type_is_T_or_C = x->type.tv == tarval_top || is_con(x->type);
1747 DEBUG_ONLY(what_reason = "lambda n.(n.type)";)
1748 P = split_by_what(X, lambda_type, &P, env);
1751 /* adjust the type tags, we have split partitions by type */
1752 for (I = P; I != NULL; I = I->split_next) {
1753 node_t *x = get_first_node(I);
1754 I->type_is_T_or_C = x->type.tv == tarval_top || is_con(x->type);
1761 if (Y->n_leader > 1) {
1762 /* we do not want split the TOP or constant partitions */
1763 if (! Y->type_is_T_or_C) {
1764 partition_t *Q = NULL;
1766 DEBUG_ONLY(what_reason = "lambda n.(n.opcode)";)
1767 Q = split_by_what(Y, lambda_opcode, &Q, env);
1774 if (Z->n_leader > 1) {
1775 const node_t *first = get_first_node(Z);
1776 int arity = get_irn_arity(first->node);
1778 what_func what = lambda_partition;
1779 DEBUG_ONLY(char buf[64];)
1781 if (env->commutative && is_op_commutative(get_irn_op(first->node)))
1782 what = lambda_commutative_partition;
1785 * BEWARE: during splitting by input 2 for instance we might
1786 * create new partitions which are different by input 1, so collect
1787 * them and split further.
1789 Z->split_next = NULL;
1792 for (input = arity - 1; input >= -1; --input) {
1794 partition_t *Z_prime = R;
1797 if (Z_prime->n_leader > 1) {
1798 env->lambda_input = input;
1799 DEBUG_ONLY(snprintf(buf, sizeof(buf), "lambda n.(n[%d].partition)", input);)
1800 DEBUG_ONLY(what_reason = buf;)
1801 S = split_by_what(Z_prime, what, &S, env);
1804 Z_prime->split_next = S;
1807 } while (R != NULL);
1812 } while (Q != NULL);
1815 } while (P != NULL);
1819 * (Re-)compute the type for a given node.
1821 * @param node the node
1823 static void default_compute(node_t *node) {
1825 ir_node *irn = node->node;
1826 node_t *block = get_irn_node(get_nodes_block(irn));
1828 if (block->type.tv == tarval_unreachable) {
1829 node->type.tv = tarval_top;
1833 /* if any of the data inputs have type top, the result is type top */
1834 for (i = get_irn_arity(irn) - 1; i >= 0; --i) {
1835 ir_node *pred = get_irn_n(irn, i);
1836 node_t *p = get_irn_node(pred);
1838 if (p->type.tv == tarval_top) {
1839 node->type.tv = tarval_top;
1844 if (get_irn_mode(node->node) == mode_X)
1845 node->type.tv = tarval_reachable;
1847 node->type.tv = computed_value(irn);
1848 } /* default_compute */
1851 * (Re-)compute the type for a Block node.
1853 * @param node the node
1855 static void compute_Block(node_t *node) {
1857 ir_node *block = node->node;
1859 if (block == get_irg_start_block(current_ir_graph)) {
1860 /* start block is always reachable */
1861 node->type.tv = tarval_reachable;
1865 for (i = get_Block_n_cfgpreds(block) - 1; i >= 0; --i) {
1866 node_t *pred = get_irn_node(get_Block_cfgpred(block, i));
1868 if (pred->type.tv == tarval_reachable) {
1869 /* A block is reachable, if at least of predecessor is reachable. */
1870 node->type.tv = tarval_reachable;
1874 node->type.tv = tarval_top;
1875 } /* compute_Block */
1878 * (Re-)compute the type for a Bad node.
1880 * @param node the node
1882 static void compute_Bad(node_t *node) {
1883 /* Bad nodes ALWAYS compute Top */
1884 node->type.tv = tarval_top;
1888 * (Re-)compute the type for an Unknown node.
1890 * @param node the node
1892 static void compute_Unknown(node_t *node) {
1893 /* While Unknown nodes should compute Top this is dangerous:
1894 * a Top input to a Cond would lead to BOTH control flows unreachable.
1895 * While this is correct in the given semantics, it would destroy the Firm
1898 * It would be safe to compute Top IF it can be assured, that only Cmp
1899 * nodes are inputs to Conds. We check that first.
1900 * This is the way Frontends typically build Firm, but some optimizations
1901 * (cond_eval for instance) might replace them by Phib's...
1903 node->type.tv = tarval_UNKNOWN;
1904 } /* compute_Unknown */
1907 * (Re-)compute the type for a Jmp node.
1909 * @param node the node
1911 static void compute_Jmp(node_t *node) {
1912 node_t *block = get_irn_node(get_nodes_block(node->node));
1914 node->type = block->type;
1918 * (Re-)compute the type for the Return node.
1920 * @param node the node
1922 static void compute_Return(node_t *node) {
1923 /* The Return node is NOT dead if it is in a reachable block.
1924 * This is already checked in compute(). so we can return
1925 * Reachable here. */
1926 node->type.tv = tarval_reachable;
1927 } /* compute_Return */
1930 * (Re-)compute the type for the End node.
1932 * @param node the node
1934 static void compute_End(node_t *node) {
1935 /* the End node is NOT dead of course */
1936 node->type.tv = tarval_reachable;
1940 * (Re-)compute the type for a SymConst node.
1942 * @param node the node
1944 static void compute_SymConst(node_t *node) {
1945 ir_node *irn = node->node;
1946 node_t *block = get_irn_node(get_nodes_block(irn));
1948 if (block->type.tv == tarval_unreachable) {
1949 node->type.tv = tarval_top;
1952 switch (get_SymConst_kind(irn)) {
1953 case symconst_addr_ent:
1954 /* case symconst_addr_name: cannot handle this yet */
1955 node->type.sym = get_SymConst_symbol(irn);
1958 node->type.tv = computed_value(irn);
1960 } /* compute_SymConst */
1963 * (Re-)compute the type for a Phi node.
1965 * @param node the node
1967 static void compute_Phi(node_t *node) {
1969 ir_node *phi = node->node;
1970 lattice_elem_t type;
1972 /* if a Phi is in a unreachable block, its type is TOP */
1973 node_t *block = get_irn_node(get_nodes_block(phi));
1975 if (block->type.tv == tarval_unreachable) {
1976 node->type.tv = tarval_top;
1980 /* Phi implements the Meet operation */
1981 type.tv = tarval_top;
1982 for (i = get_Phi_n_preds(phi) - 1; i >= 0; --i) {
1983 node_t *pred = get_irn_node(get_Phi_pred(phi, i));
1984 node_t *pred_X = get_irn_node(get_Block_cfgpred(block->node, i));
1986 if (pred_X->type.tv == tarval_unreachable || pred->type.tv == tarval_top) {
1987 /* ignore TOP inputs: We must check here for unreachable blocks,
1988 because Firm constants live in the Start Block are NEVER Top.
1989 Else, a Phi (1,2) will produce Bottom, even if the 2 for instance
1990 comes from a unreachable input. */
1993 if (pred->type.tv == tarval_bottom) {
1994 node->type.tv = tarval_bottom;
1996 } else if (type.tv == tarval_top) {
1997 /* first constant found */
1999 } else if (type.tv != pred->type.tv) {
2000 /* different constants or tarval_bottom */
2001 node->type.tv = tarval_bottom;
2004 /* else nothing, constants are the same */
2010 * (Re-)compute the type for an Add. Special case: one nodes is a Zero Const.
2012 * @param node the node
2014 static void compute_Add(node_t *node) {
2015 ir_node *sub = node->node;
2016 node_t *l = get_irn_node(get_Add_left(sub));
2017 node_t *r = get_irn_node(get_Add_right(sub));
2018 lattice_elem_t a = l->type;
2019 lattice_elem_t b = r->type;
2022 if (a.tv == tarval_top || b.tv == tarval_top) {
2023 node->type.tv = tarval_top;
2024 } else if (a.tv == tarval_bottom || b.tv == tarval_bottom) {
2025 node->type.tv = tarval_bottom;
2027 /* x + 0 = 0 + x = x, but beware of floating point +0 + -0, so we
2028 must call tarval_add() first to handle this case! */
2029 if (is_tarval(a.tv)) {
2030 if (is_tarval(b.tv)) {
2031 node->type.tv = tarval_add(a.tv, b.tv);
2034 mode = get_tarval_mode(a.tv);
2035 if (a.tv == get_mode_null(mode)) {
2039 } else if (is_tarval(b.tv)) {
2040 mode = get_tarval_mode(b.tv);
2041 if (b.tv == get_mode_null(mode)) {
2046 node->type.tv = tarval_bottom;
2051 * (Re-)compute the type for a Sub. Special case: both nodes are congruent.
2053 * @param node the node
2055 static void compute_Sub(node_t *node) {
2056 ir_node *sub = node->node;
2057 node_t *l = get_irn_node(get_Sub_left(sub));
2058 node_t *r = get_irn_node(get_Sub_right(sub));
2059 lattice_elem_t a = l->type;
2060 lattice_elem_t b = r->type;
2063 if (a.tv == tarval_top || b.tv == tarval_top) {
2064 node->type.tv = tarval_top;
2065 } else if (is_con(a) && is_con(b)) {
2066 if (is_tarval(a.tv) && is_tarval(b.tv)) {
2067 node->type.tv = tarval_sub(a.tv, b.tv, get_irn_mode(sub));
2068 } else if (is_tarval(a.tv) && tarval_is_null(a.tv)) {
2070 } else if (is_tarval(b.tv) && tarval_is_null(b.tv)) {
2073 node->type.tv = tarval_bottom;
2075 node->by_all_const = 1;
2076 } else if (r->part == l->part &&
2077 (!mode_is_float(get_irn_mode(l->node)))) {
2079 * BEWARE: a - a is NOT always 0 for floating Point values, as
2080 * NaN op NaN = NaN, so we must check this here.
2082 ir_mode *mode = get_irn_mode(sub);
2083 tv = get_mode_null(mode);
2085 /* if the node was ONCE evaluated by all constants, but now
2086 this breaks AND we get from the argument partitions a different
2087 result, switch to bottom.
2088 This happens because initially all nodes are in the same partition ... */
2089 if (node->by_all_const && node->type.tv != tv)
2093 node->type.tv = tarval_bottom;
2098 * (Re-)compute the type for an Eor. Special case: both nodes are congruent.
2100 * @param node the node
2102 static void compute_Eor(node_t *node) {
2103 ir_node *eor = node->node;
2104 node_t *l = get_irn_node(get_Eor_left(eor));
2105 node_t *r = get_irn_node(get_Eor_right(eor));
2106 lattice_elem_t a = l->type;
2107 lattice_elem_t b = r->type;
2110 if (a.tv == tarval_top || b.tv == tarval_top) {
2111 node->type.tv = tarval_top;
2112 } else if (is_con(a) && is_con(b)) {
2113 if (is_tarval(a.tv) && is_tarval(b.tv)) {
2114 node->type.tv = tarval_eor(a.tv, b.tv);
2115 } else if (is_tarval(a.tv) && tarval_is_null(a.tv)) {
2117 } else if (is_tarval(b.tv) && tarval_is_null(b.tv)) {
2120 node->type.tv = tarval_bottom;
2122 node->by_all_const = 1;
2123 } else if (r->part == l->part) {
2124 ir_mode *mode = get_irn_mode(eor);
2125 tv = get_mode_null(mode);
2127 /* if the node was ONCE evaluated by all constants, but now
2128 this breaks AND we get from the argument partitions a different
2129 result, switch to bottom.
2130 This happens because initially all nodes are in the same partition ... */
2131 if (node->by_all_const && node->type.tv != tv)
2135 node->type.tv = tarval_bottom;
2140 * (Re-)compute the type for Cmp.
2142 * @param node the node
2144 static void compute_Cmp(node_t *node) {
2145 ir_node *cmp = node->node;
2146 node_t *l = get_irn_node(get_Cmp_left(cmp));
2147 node_t *r = get_irn_node(get_Cmp_right(cmp));
2148 lattice_elem_t a = l->type;
2149 lattice_elem_t b = r->type;
2151 if (a.tv == tarval_top || b.tv == tarval_top) {
2152 node->type.tv = tarval_top;
2153 } else if (r->part == l->part) {
2154 /* both nodes congruent, we can probably do something */
2155 node->type.tv = tarval_b_true;
2156 } else if (is_con(a) && is_con(b)) {
2157 /* both nodes are constants, we can probably do something */
2158 node->type.tv = tarval_b_true;
2160 node->type.tv = tarval_bottom;
2165 * (Re-)compute the type for a Proj(Cmp).
2167 * @param node the node
2168 * @param cond the predecessor Cmp node
2170 static void compute_Proj_Cmp(node_t *node, ir_node *cmp) {
2171 ir_node *proj = node->node;
2172 node_t *l = get_irn_node(get_Cmp_left(cmp));
2173 node_t *r = get_irn_node(get_Cmp_right(cmp));
2174 lattice_elem_t a = l->type;
2175 lattice_elem_t b = r->type;
2176 pn_Cmp pnc = get_Proj_proj(proj);
2179 if (a.tv == tarval_top || b.tv == tarval_top) {
2180 node->type.tv = tarval_undefined;
2181 } else if (is_con(a) && is_con(b)) {
2182 default_compute(node);
2183 node->by_all_const = 1;
2184 } else if (r->part == l->part &&
2185 (!mode_is_float(get_irn_mode(l->node)) || pnc == pn_Cmp_Lt || pnc == pn_Cmp_Gt)) {
2187 * BEWARE: a == a is NOT always True for floating Point values, as
2188 * NaN != NaN is defined, so we must check this here.
2190 tv = pnc & pn_Cmp_Eq ? tarval_b_true: tarval_b_false;
2192 /* if the node was ONCE evaluated by all constants, but now
2193 this breaks AND we get from the argument partitions a different
2194 result, switch to bottom.
2195 This happens because initially all nodes are in the same partition ... */
2196 if (node->by_all_const && node->type.tv != tv)
2200 node->type.tv = tarval_bottom;
2202 } /* compute_Proj_Cmp */
2205 * (Re-)compute the type for a Proj(Cond).
2207 * @param node the node
2208 * @param cond the predecessor Cond node
2210 static void compute_Proj_Cond(node_t *node, ir_node *cond) {
2211 ir_node *proj = node->node;
2212 long pnc = get_Proj_proj(proj);
2213 ir_node *sel = get_Cond_selector(cond);
2214 node_t *selector = get_irn_node(sel);
2217 * Note: it is crucial for the monotony that the Proj(Cond)
2218 * are evaluates after all predecessors of the Cond selector are
2224 * Due to the fact that 0 is a const, the Cmp gets immediately
2225 * on the cprop list. It will be evaluated before x is evaluated,
2226 * might leaving x as Top. When later x is evaluated, the Cmp
2227 * might change its value.
2228 * BUT if the Cond is evaluated before this happens, Proj(Cond, FALSE)
2229 * gets R, and later changed to F if Cmp is evaluated to True!
2231 * We prevent this by putting Conds in an extra cprop_X queue, which
2232 * gets evaluated after the cprop queue is empty.
2234 * Note that this even happens with Click's original algorithm, if
2235 * Cmp(x, 0) is evaluated to True first and later changed to False
2236 * if x was Top first tiome and later Const ...
2237 * It is unclear how Click solved that problem ...
2239 if (get_irn_mode(sel) == mode_b) {
2241 if (pnc == pn_Cond_true) {
2242 if (selector->type.tv == tarval_b_false) {
2243 node->type.tv = tarval_unreachable;
2244 } else if (selector->type.tv == tarval_b_true) {
2245 node->type.tv = tarval_reachable;
2246 } else if (selector->type.tv == tarval_bottom) {
2247 node->type.tv = tarval_reachable;
2249 assert(selector->type.tv == tarval_top);
2250 /* any condition based on Top is "!=" */
2251 node->type.tv = tarval_unreachable;
2254 assert(pnc == pn_Cond_false);
2256 if (selector->type.tv == tarval_b_false) {
2257 node->type.tv = tarval_reachable;
2258 } else if (selector->type.tv == tarval_b_true) {
2259 node->type.tv = tarval_unreachable;
2260 } else if (selector->type.tv == tarval_bottom) {
2261 node->type.tv = tarval_reachable;
2263 assert(selector->type.tv == tarval_top);
2264 /* any condition based on Top is "!=" */
2265 node->type.tv = tarval_reachable;
2270 if (selector->type.tv == tarval_bottom) {
2271 node->type.tv = tarval_reachable;
2272 } else if (selector->type.tv == tarval_top) {
2273 if (pnc == get_Cond_defaultProj(cond)) {
2274 /* a switch based of Top is always "default" */
2275 node->type.tv = tarval_reachable;
2277 node->type.tv = tarval_unreachable;
2279 long value = get_tarval_long(selector->type.tv);
2280 if (pnc == get_Cond_defaultProj(cond)) {
2281 /* default switch, have to check ALL other cases */
2284 for (i = get_irn_n_outs(cond) - 1; i >= 0; --i) {
2285 ir_node *succ = get_irn_out(cond, i);
2289 if (value == get_Proj_proj(succ)) {
2290 /* we found a match, will NOT take the default case */
2291 node->type.tv = tarval_unreachable;
2295 /* all cases checked, no match, will take default case */
2296 node->type.tv = tarval_reachable;
2299 node->type.tv = value == pnc ? tarval_reachable : tarval_unreachable;
2303 } /* compute_Proj_Cond */
2306 * (Re-)compute the type for a Proj-Node.
2308 * @param node the node
2310 static void compute_Proj(node_t *node) {
2311 ir_node *proj = node->node;
2312 ir_mode *mode = get_irn_mode(proj);
2313 node_t *block = get_irn_node(get_nodes_block(skip_Proj(proj)));
2314 ir_node *pred = get_Proj_pred(proj);
2316 if (block->type.tv == tarval_unreachable) {
2317 /* a Proj in a unreachable Block stay Top */
2318 node->type.tv = tarval_top;
2321 if (get_irn_node(pred)->type.tv == tarval_top && !is_Cond(pred)) {
2322 /* if the predecessor is Top, its Proj follow */
2323 node->type.tv = tarval_top;
2327 if (mode == mode_M) {
2328 /* mode M is always bottom */
2329 node->type.tv = tarval_bottom;
2332 if (mode != mode_X) {
2334 compute_Proj_Cmp(node, pred);
2336 default_compute(node);
2339 /* handle mode_X nodes */
2341 switch (get_irn_opcode(pred)) {
2343 /* the Proj_X from the Start is always reachable.
2344 However this is already handled at the top. */
2345 node->type.tv = tarval_reachable;
2348 compute_Proj_Cond(node, pred);
2351 default_compute(node);
2353 } /* compute_Proj */
2356 * (Re-)compute the type for a Confirm.
2358 * @param node the node
2360 static void compute_Confirm(node_t *node) {
2361 ir_node *confirm = node->node;
2362 node_t *pred = get_irn_node(get_Confirm_value(confirm));
2364 if (get_Confirm_cmp(confirm) == pn_Cmp_Eq) {
2365 node_t *bound = get_irn_node(get_Confirm_bound(confirm));
2367 if (is_con(bound->type)) {
2368 /* is equal to a constant */
2369 node->type = bound->type;
2373 /* a Confirm is a copy OR a Const */
2374 node->type = pred->type;
2375 } /* compute_Confirm */
2378 * (Re-)compute the type for a Max.
2380 * @param node the node
2382 static void compute_Max(node_t *node) {
2383 ir_node *op = node->node;
2384 node_t *l = get_irn_node(get_binop_left(op));
2385 node_t *r = get_irn_node(get_binop_right(op));
2386 lattice_elem_t a = l->type;
2387 lattice_elem_t b = r->type;
2389 if (a.tv == tarval_top || b.tv == tarval_top) {
2390 node->type.tv = tarval_top;
2391 } else if (is_con(a) && is_con(b)) {
2392 /* both nodes are constants, we can probably do something */
2394 /* this case handles SymConsts as well */
2397 ir_mode *mode = get_irn_mode(op);
2398 tarval *tv_min = get_mode_min(mode);
2402 else if (b.tv == tv_min)
2404 else if (is_tarval(a.tv) && is_tarval(b.tv)) {
2405 if (tarval_cmp(a.tv, b.tv) & pn_Cmp_Gt)
2406 node->type.tv = a.tv;
2408 node->type.tv = b.tv;
2410 node->type.tv = tarval_bad;
2413 } else if (r->part == l->part) {
2414 /* both nodes congruent, we can probably do something */
2417 node->type.tv = tarval_bottom;
2422 * (Re-)compute the type for a Min.
2424 * @param node the node
2426 static void compute_Min(node_t *node) {
2427 ir_node *op = node->node;
2428 node_t *l = get_irn_node(get_binop_left(op));
2429 node_t *r = get_irn_node(get_binop_right(op));
2430 lattice_elem_t a = l->type;
2431 lattice_elem_t b = r->type;
2433 if (a.tv == tarval_top || b.tv == tarval_top) {
2434 node->type.tv = tarval_top;
2435 } else if (is_con(a) && is_con(b)) {
2436 /* both nodes are constants, we can probably do something */
2438 /* this case handles SymConsts as well */
2441 ir_mode *mode = get_irn_mode(op);
2442 tarval *tv_max = get_mode_max(mode);
2446 else if (b.tv == tv_max)
2448 else if (is_tarval(a.tv) && is_tarval(b.tv)) {
2449 if (tarval_cmp(a.tv, b.tv) & pn_Cmp_Gt)
2450 node->type.tv = a.tv;
2452 node->type.tv = b.tv;
2454 node->type.tv = tarval_bad;
2457 } else if (r->part == l->part) {
2458 /* both nodes congruent, we can probably do something */
2461 node->type.tv = tarval_bottom;
2466 * (Re-)compute the type for a given node.
2468 * @param node the node
2470 static void compute(node_t *node) {
2471 ir_node *irn = node->node;
2475 if (is_no_Block(irn)) {
2476 /* for pinned nodes, check its control input */
2477 if (get_irn_pinned(skip_Proj(irn)) == op_pin_state_pinned) {
2478 node_t *block = get_irn_node(get_nodes_block(irn));
2480 if (block->type.tv == tarval_unreachable) {
2481 node->type.tv = tarval_top;
2487 func = (compute_func)node->node->op->ops.generic;
2493 * Identity functions: Note that one might thing that identity() is just a
2494 * synonym for equivalent_node(). While this is true, we cannot use it for the algorithm
2495 * here, because it expects that the identity node is one of the inputs, which is NOT
2496 * always true for equivalent_node() which can handle (and does sometimes) DAGs.
2497 * So, we have our own implementation, which copies some parts of equivalent_node()
2501 * Calculates the Identity for Phi nodes
2503 static node_t *identity_Phi(node_t *node) {
2504 ir_node *phi = node->node;
2505 ir_node *block = get_nodes_block(phi);
2506 node_t *n_part = NULL;
2509 for (i = get_Phi_n_preds(phi) - 1; i >= 0; --i) {
2510 node_t *pred_X = get_irn_node(get_Block_cfgpred(block, i));
2512 if (pred_X->type.tv == tarval_reachable) {
2513 node_t *pred = get_irn_node(get_Phi_pred(phi, i));
2517 else if (n_part->part != pred->part) {
2518 /* incongruent inputs, not a follower */
2523 /* if n_part is NULL here, all inputs path are dead, the Phi computes
2524 * tarval_top, is in the TOP partition and should NOT being split! */
2525 assert(n_part != NULL);
2527 } /* identity_Phi */
2530 * Calculates the Identity for commutative 0 neutral nodes.
2532 static node_t *identity_comm_zero_binop(node_t *node) {
2533 ir_node *op = node->node;
2534 node_t *a = get_irn_node(get_binop_left(op));
2535 node_t *b = get_irn_node(get_binop_right(op));
2536 ir_mode *mode = get_irn_mode(op);
2539 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
2540 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
2543 /* node: no input should be tarval_top, else the binop would be also
2544 * Top and not being split. */
2545 zero = get_mode_null(mode);
2546 if (a->type.tv == zero)
2548 if (b->type.tv == zero)
2551 } /* identity_comm_zero_binop */
2554 * Calculates the Identity for Shift nodes.
2556 static node_t *identity_shift(node_t *node) {
2557 ir_node *op = node->node;
2558 node_t *b = get_irn_node(get_binop_right(op));
2559 ir_mode *mode = get_irn_mode(b->node);
2562 /* node: no input should be tarval_top, else the binop would be also
2563 * Top and not being split. */
2564 zero = get_mode_null(mode);
2565 if (b->type.tv == zero)
2566 return get_irn_node(get_binop_left(op));
2568 } /* identity_shift */
2571 * Calculates the Identity for Mul nodes.
2573 static node_t *identity_Mul(node_t *node) {
2574 ir_node *op = node->node;
2575 node_t *a = get_irn_node(get_Mul_left(op));
2576 node_t *b = get_irn_node(get_Mul_right(op));
2577 ir_mode *mode = get_irn_mode(op);
2580 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
2581 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
2584 /* node: no input should be tarval_top, else the binop would be also
2585 * Top and not being split. */
2586 one = get_mode_one(mode);
2587 if (a->type.tv == one)
2589 if (b->type.tv == one)
2592 } /* identity_Mul */
2595 * Calculates the Identity for Sub nodes.
2597 static node_t *identity_Sub(node_t *node) {
2598 ir_node *sub = node->node;
2599 node_t *b = get_irn_node(get_Sub_right(sub));
2600 ir_mode *mode = get_irn_mode(sub);
2602 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
2603 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
2606 /* node: no input should be tarval_top, else the binop would be also
2607 * Top and not being split. */
2608 if (b->type.tv == get_mode_null(mode))
2609 return get_irn_node(get_Sub_left(sub));
2611 } /* identity_Sub */
2614 * Calculates the Identity for And nodes.
2616 static node_t *identity_And(node_t *node) {
2617 ir_node *and = node->node;
2618 node_t *a = get_irn_node(get_And_left(and));
2619 node_t *b = get_irn_node(get_And_right(and));
2620 tarval *neutral = get_mode_all_one(get_irn_mode(and));
2622 /* node: no input should be tarval_top, else the And would be also
2623 * Top and not being split. */
2624 if (a->type.tv == neutral)
2626 if (b->type.tv == neutral)
2629 } /* identity_And */
2632 * Calculates the Identity for Confirm nodes.
2634 static node_t *identity_Confirm(node_t *node) {
2635 ir_node *confirm = node->node;
2637 /* a Confirm is always a Copy */
2638 return get_irn_node(get_Confirm_value(confirm));
2639 } /* identity_Confirm */
2642 * Calculates the Identity for Mux nodes.
2644 static node_t *identity_Mux(node_t *node) {
2645 ir_node *mux = node->node;
2646 node_t *t = get_irn_node(get_Mux_true(mux));
2647 node_t *f = get_irn_node(get_Mux_false(mux));
2650 if (t->part == f->part)
2653 /* for now, the 1-input identity is not supported */
2655 sel = get_irn_node(get_Mux_sel(mux));
2657 /* Mux sel input is mode_b, so it is always a tarval */
2658 if (sel->type.tv == tarval_b_true)
2660 if (sel->type.tv == tarval_b_false)
2664 } /* identity_Mux */
2667 * Calculates the Identity for Min nodes.
2669 static node_t *identity_Min(node_t *node) {
2670 ir_node *op = node->node;
2671 node_t *a = get_irn_node(get_binop_left(op));
2672 node_t *b = get_irn_node(get_binop_right(op));
2673 ir_mode *mode = get_irn_mode(op);
2676 if (a->part == b->part) {
2677 /* leader of multiple predecessors */
2681 /* works even with NaN */
2682 tv_max = get_mode_max(mode);
2683 if (a->type.tv == tv_max)
2685 if (b->type.tv == tv_max)
2688 } /* identity_Min */
2691 * Calculates the Identity for Max nodes.
2693 static node_t *identity_Max(node_t *node) {
2694 ir_node *op = node->node;
2695 node_t *a = get_irn_node(get_binop_left(op));
2696 node_t *b = get_irn_node(get_binop_right(op));
2697 ir_mode *mode = get_irn_mode(op);
2700 if (a->part == b->part) {
2701 /* leader of multiple predecessors */
2705 /* works even with NaN */
2706 tv_min = get_mode_min(mode);
2707 if (a->type.tv == tv_min)
2709 if (b->type.tv == tv_min)
2712 } /* identity_Max */
2715 * Calculates the Identity for nodes.
2717 static node_t *identity(node_t *node) {
2718 ir_node *irn = node->node;
2720 switch (get_irn_opcode(irn)) {
2722 return identity_Phi(node);
2724 return identity_Mul(node);
2728 return identity_comm_zero_binop(node);
2733 return identity_shift(node);
2735 return identity_And(node);
2737 return identity_Sub(node);
2739 return identity_Confirm(node);
2741 return identity_Mux(node);
2743 return identity_Min(node);
2745 return identity_Max(node);
2752 * Node follower is a (new) follower of leader, segregate Leader
2755 static void segregate_def_use_chain_1(const ir_node *follower, node_t *leader) {
2756 ir_node *l = leader->node;
2757 int j, i, n = get_irn_n_outs(l);
2759 DB((dbg, LEVEL_2, "%+F is a follower of %+F\n", follower, leader->node));
2760 /* The leader edges must remain sorted, but follower edges can
2762 for (i = leader->n_followers + 1; i <= n; ++i) {
2763 if (l->out[i].use == follower) {
2764 ir_def_use_edge t = l->out[i];
2766 for (j = i - 1; j >= leader->n_followers + 1; --j)
2767 l->out[j + 1] = l->out[j];
2768 ++leader->n_followers;
2769 l->out[leader->n_followers] = t;
2773 } /* segregate_def_use_chain_1 */
2776 * Node follower is a (new) follower segregate its Leader
2779 * @param follower the follower IR node
2781 static void segregate_def_use_chain(const ir_node *follower) {
2784 for (i = get_irn_arity(follower) - 1; i >= 0; --i) {
2785 node_t *pred = get_irn_node(get_irn_n(follower, i));
2787 segregate_def_use_chain_1(follower, pred);
2789 } /* segregate_def_use_chain */
2792 * Propagate constant evaluation.
2794 * @param env the environment
2796 static void propagate(environment_t *env) {
2799 lattice_elem_t old_type;
2801 unsigned n_fallen, old_type_was_T_or_C;
2804 while (env->cprop != NULL) {
2805 void *oldopcode = NULL;
2807 /* remove the first partition X from cprop */
2810 env->cprop = X->cprop_next;
2812 old_type_was_T_or_C = X->type_is_T_or_C;
2814 DB((dbg, LEVEL_2, "Propagate type on part%d\n", X->nr));
2818 int cprop_empty = list_empty(&X->cprop);
2819 int cprop_X_empty = list_empty(&X->cprop_X);
2821 if (cprop_empty && cprop_X_empty) {
2822 /* both cprop lists are empty */
2826 /* remove the first Node x from X.cprop */
2828 /* Get a node from the cprop_X list only if
2829 * all data nodes are processed.
2830 * This ensures, that all inputs of the Cond
2831 * predecessor are processed if its type is still Top.
2833 x = list_entry(X->cprop_X.next, node_t, cprop_list);
2835 x = list_entry(X->cprop.next, node_t, cprop_list);
2838 //assert(x->part == X);
2839 list_del(&x->cprop_list);
2842 if (x->is_follower && identity(x) == x) {
2843 /* check the opcode first */
2844 if (oldopcode == NULL) {
2845 oldopcode = lambda_opcode(get_first_node(X), env);
2847 if (oldopcode != lambda_opcode(x, env)) {
2848 if (x->on_fallen == 0) {
2849 /* different opcode -> x falls out of this partition */
2854 DB((dbg, LEVEL_2, "Add node %+F to fallen\n", x->node));
2858 /* x will make the follower -> leader transition */
2859 follower_to_leader(x);
2862 /* compute a new type for x */
2864 DB((dbg, LEVEL_3, "computing type of %+F\n", x->node));
2866 if (x->type.tv != old_type.tv) {
2867 DB((dbg, LEVEL_2, "node %+F has changed type from %+F to %+F\n", x->node, old_type, x->type));
2868 verify_type(old_type, x);
2870 if (x->on_fallen == 0) {
2871 /* Add x to fallen. Nodes might fall from T -> const -> _|_, so check that they are
2872 not already on the list. */
2877 DB((dbg, LEVEL_2, "Add node %+F to fallen\n", x->node));
2879 for (i = get_irn_n_outs(x->node) - 1; i >= 0; --i) {
2880 ir_node *succ = get_irn_out(x->node, i);
2881 node_t *y = get_irn_node(succ);
2883 /* Add y to y.partition.cprop. */
2884 add_to_cprop(y, env);
2889 if (n_fallen > 0 && n_fallen != X->n_leader) {
2890 DB((dbg, LEVEL_2, "Splitting part%d by fallen\n", X->nr));
2891 Y = split(&X, fallen, env);
2893 * We have split out fallen node. The type of the result
2894 * partition is NOT set yet.
2896 Y->type_is_T_or_C = 0;
2900 /* remove the flags from the fallen list */
2901 for (x = fallen; x != NULL; x = x->next)
2904 if (old_type_was_T_or_C) {
2907 /* check if some nodes will make the leader -> follower transition */
2908 list_for_each_entry_safe(node_t, y, tmp, &Y->Leader, node_list) {
2909 if (y->type.tv != tarval_top && ! is_con(y->type)) {
2910 node_t *eq_node = identity(y);
2912 if (eq_node != y && eq_node->part == y->part) {
2913 DB((dbg, LEVEL_2, "Node %+F is a follower of %+F\n", y->node, eq_node->node));
2914 /* move to Follower */
2916 list_del(&y->node_list);
2917 list_add_tail(&y->node_list, &Y->Follower);
2920 segregate_def_use_chain(y->node);
2930 * Get the leader for a given node from its congruence class.
2932 * @param irn the node
2934 static ir_node *get_leader(node_t *node) {
2935 partition_t *part = node->part;
2937 if (part->n_leader > 1 || node->is_follower) {
2938 if (node->is_follower) {
2939 DB((dbg, LEVEL_2, "Replacing follower %+F\n", node->node));
2942 DB((dbg, LEVEL_2, "Found congruence class for %+F\n", node->node));
2944 return get_first_node(part)->node;
2950 * Return non-zero if the control flow predecessor node pred
2951 * is the only reachable control flow exit of its block.
2953 * @param pred the control flow exit
2955 static int can_exchange(ir_node *pred) {
2958 else if (is_Jmp(pred))
2960 else if (get_irn_mode(pred) == mode_T) {
2963 /* if the predecessor block has more than one
2964 reachable outputs we cannot remove the block */
2966 for (i = get_irn_n_outs(pred) - 1; i >= 0; --i) {
2967 ir_node *proj = get_irn_out(pred, i);
2970 /* skip non-control flow Proj's */
2971 if (get_irn_mode(proj) != mode_X)
2974 node = get_irn_node(proj);
2975 if (node->type.tv == tarval_reachable) {
2983 } /* can_exchange */
2986 * Block Post-Walker, apply the analysis results on control flow by
2987 * shortening Phi's and Block inputs.
2989 static void apply_cf(ir_node *block, void *ctx) {
2990 environment_t *env = ctx;
2991 node_t *node = get_irn_node(block);
2993 ir_node **ins, **in_X;
2994 ir_node *phi, *next;
2996 n = get_Block_n_cfgpreds(block);
2998 if (node->type.tv == tarval_unreachable) {
3001 for (i = n - 1; i >= 0; --i) {
3002 ir_node *pred = get_Block_cfgpred(block, i);
3004 if (! is_Bad(pred)) {
3005 node_t *pred_bl = get_irn_node(get_nodes_block(skip_Proj(pred)));
3007 if (pred_bl->flagged == 0) {
3008 pred_bl->flagged = 3;
3010 if (pred_bl->type.tv == tarval_reachable) {
3012 * We will remove an edge from block to its pred.
3013 * This might leave the pred block as an endless loop
3015 if (! is_backedge(block, i))
3016 keep_alive(pred_bl->node);
3022 /* the EndBlock is always reachable even if the analysis
3023 finds out the opposite :-) */
3024 if (block != get_irg_end_block(current_ir_graph)) {
3025 /* mark dead blocks */
3026 set_Block_dead(block);
3027 DB((dbg, LEVEL_1, "Removing dead %+F\n", block));
3029 /* the endblock is unreachable */
3030 set_irn_in(block, 0, NULL);
3036 /* only one predecessor combine */
3037 ir_node *pred = skip_Proj(get_Block_cfgpred(block, 0));
3039 if (can_exchange(pred)) {
3040 ir_node *new_block = get_nodes_block(pred);
3041 DB((dbg, LEVEL_1, "Fuse %+F with %+F\n", block, new_block));
3042 DBG_OPT_COMBO(block, new_block, FS_OPT_COMBO_CF);
3043 exchange(block, new_block);
3044 node->node = new_block;
3050 NEW_ARR_A(ir_node *, in_X, n);
3052 for (i = 0; i < n; ++i) {
3053 ir_node *pred = get_Block_cfgpred(block, i);
3054 node_t *node = get_irn_node(pred);
3056 if (node->type.tv == tarval_reachable) {
3059 DB((dbg, LEVEL_1, "Removing dead input %d from %+F (%+F)\n", i, block, pred));
3060 if (! is_Bad(pred)) {
3061 node_t *pred_bl = get_irn_node(get_nodes_block(skip_Proj(pred)));
3063 if (pred_bl->flagged == 0) {
3064 pred_bl->flagged = 3;
3066 if (pred_bl->type.tv == tarval_reachable) {
3068 * We will remove an edge from block to its pred.
3069 * This might leave the pred block as an endless loop
3071 if (! is_backedge(block, i))
3072 keep_alive(pred_bl->node);
3082 NEW_ARR_A(ir_node *, ins, n);
3083 for (phi = get_Block_phis(block); phi != NULL; phi = next) {
3084 node_t *node = get_irn_node(phi);
3086 next = get_Phi_next(phi);
3087 if (is_tarval(node->type.tv) && tarval_is_constant(node->type.tv)) {
3088 /* this Phi is replaced by a constant */
3089 tarval *tv = node->type.tv;
3090 ir_node *c = new_r_Const(current_ir_graph, block, get_tarval_mode(tv), tv);
3092 set_irn_node(c, node);
3094 DB((dbg, LEVEL_1, "%+F is replaced by %+F\n", phi, c));
3095 DBG_OPT_COMBO(phi, c, FS_OPT_COMBO_CONST);
3100 for (i = 0; i < n; ++i) {
3101 node_t *pred = get_irn_node(get_Block_cfgpred(block, i));
3103 if (pred->type.tv == tarval_reachable) {
3104 ins[j++] = get_Phi_pred(phi, i);
3108 /* this Phi is replaced by a single predecessor */
3109 ir_node *s = ins[0];
3110 node_t *phi_node = get_irn_node(phi);
3113 DB((dbg, LEVEL_1, "%+F is replaced by %+F because of cf change\n", phi, s));
3114 DBG_OPT_COMBO(phi, s, FS_OPT_COMBO_FOLLOWER);
3119 set_irn_in(phi, j, ins);
3127 /* this Block has only one live predecessor */
3128 ir_node *pred = skip_Proj(in_X[0]);
3130 if (can_exchange(pred)) {
3131 ir_node *new_block = get_nodes_block(pred);
3132 DBG_OPT_COMBO(block, new_block, FS_OPT_COMBO_CF);
3133 exchange(block, new_block);
3134 node->node = new_block;
3139 set_irn_in(block, k, in_X);
3144 * Exchange a node by its leader.
3145 * Beware: in rare cases the mode might be wrong here, for instance
3146 * AddP(x, NULL) is a follower of x, but with different mode.
3149 static void exchange_leader(ir_node *irn, ir_node *leader) {
3150 ir_mode *mode = get_irn_mode(irn);
3151 if (mode != get_irn_mode(leader)) {
3152 /* The conv is a no-op, so we are fre to place in
3153 * either in the block of the leader OR in irn's block.
3154 * Probably placing it into leaders block might reduce
3155 * the number of Conv due to CSE. */
3156 ir_node *block = get_nodes_block(leader);
3157 dbg_info *dbg = get_irn_dbg_info(irn);
3159 leader = new_rd_Conv(dbg, current_ir_graph, block, leader, mode);
3161 exchange(irn, leader);
3165 * Post-Walker, apply the analysis results;
3167 static void apply_result(ir_node *irn, void *ctx) {
3168 environment_t *env = ctx;
3169 node_t *node = get_irn_node(irn);
3171 if (is_Block(irn) || is_End(irn) || is_Bad(irn)) {
3172 /* blocks already handled, do not touch the End node */
3174 node_t *block = get_irn_node(get_nodes_block(irn));
3176 if (block->type.tv == tarval_unreachable) {
3177 ir_node *bad = get_irg_bad(current_ir_graph);
3179 /* here, bad might already have a node, but this can be safely ignored
3180 as long as bad has at least ONE valid node */
3181 set_irn_node(bad, node);
3183 DB((dbg, LEVEL_1, "%+F is unreachable\n", irn));
3187 else if (node->type.tv == tarval_top) {
3188 /* don't kick away Unknown's, they might be still needed */
3189 if (! is_Unknown(irn)) {
3190 ir_mode *mode = get_irn_mode(irn);
3191 ir_node *unk = new_r_Unknown(current_ir_graph, mode);
3193 /* control flow should already be handled at apply_cf() */
3194 assert(mode != mode_X);
3196 /* see comment above */
3197 set_irn_node(unk, node);
3199 DB((dbg, LEVEL_1, "%+F computes Top\n", irn));
3204 else if (get_irn_mode(irn) == mode_X) {
3207 ir_node *cond = get_Proj_pred(irn);
3209 if (is_Cond(cond)) {
3210 node_t *sel = get_irn_node(get_Cond_selector(cond));
3212 if (is_tarval(sel->type.tv) && tarval_is_constant(sel->type.tv)) {
3213 /* Cond selector is a constant and the Proj is reachable, make a Jmp */
3214 ir_node *jmp = new_r_Jmp(current_ir_graph, block->node);
3215 set_irn_node(jmp, node);
3217 DB((dbg, LEVEL_1, "%+F is replaced by %+F\n", irn, jmp));
3218 DBG_OPT_COMBO(irn, jmp, FS_OPT_COMBO_CF);
3225 /* normal data node */
3226 if (is_tarval(node->type.tv) && tarval_is_constant(node->type.tv)) {
3227 tarval *tv = node->type.tv;
3230 * Beware: never replace mode_T nodes by constants. Currently we must mark
3231 * mode_T nodes with constants, but do NOT replace them.
3233 if (! is_Const(irn) && get_irn_mode(irn) != mode_T) {
3234 /* can be replaced by a constant */
3235 ir_node *c = new_r_Const(current_ir_graph, block->node, get_tarval_mode(tv), tv);
3236 set_irn_node(c, node);
3238 DB((dbg, LEVEL_1, "%+F is replaced by %+F\n", irn, c));
3239 DBG_OPT_COMBO(irn, c, FS_OPT_COMBO_CONST);
3240 exchange_leader(irn, c);
3243 } else if (is_entity(node->type.sym.entity_p)) {
3244 if (! is_SymConst(irn)) {
3245 /* can be replaced by a SymConst */
3246 ir_node *symc = new_r_SymConst(current_ir_graph, block->node, get_irn_mode(irn), node->type.sym, symconst_addr_ent);
3247 set_irn_node(symc, node);
3250 DB((dbg, LEVEL_1, "%+F is replaced by %+F\n", irn, symc));
3251 DBG_OPT_COMBO(irn, symc, FS_OPT_COMBO_CONST);
3252 exchange_leader(irn, symc);
3255 } else if (is_Confirm(irn)) {
3256 /* Confirms are always follower, but do not kill them here */
3258 ir_node *leader = get_leader(node);
3260 if (leader != irn) {
3261 int non_strict_phi = 0;
3264 * Beware: Do not remove Phi(Unknown, ..., x, ..., Unknown)
3265 * as this might create non-strict programs.
3267 if (node->is_follower && is_Phi(irn) && !is_Unknown(leader)) {
3270 for (i = get_Phi_n_preds(irn) - 1; i >= 0; --i) {
3271 ir_node *pred = get_Phi_pred(irn, i);
3273 if (is_Unknown(pred)) {
3279 if (! non_strict_phi) {
3280 DB((dbg, LEVEL_1, "%+F from part%d is replaced by %+F\n", irn, node->part->nr, leader));
3281 if (node->is_follower)
3282 DBG_OPT_COMBO(irn, leader, FS_OPT_COMBO_FOLLOWER);
3284 DBG_OPT_COMBO(irn, leader, FS_OPT_COMBO_CONGRUENT);
3285 exchange_leader(irn, leader);
3292 } /* apply_result */
3295 * Fix the keep-alives by deleting unreachable ones.
3297 static void apply_end(ir_node *end, environment_t *env) {
3298 int i, j, n = get_End_n_keepalives(end);
3302 NEW_ARR_A(ir_node *, in, n);
3304 /* fix the keep alive */
3305 for (i = j = 0; i < n; i++) {
3306 ir_node *ka = get_End_keepalive(end, i);
3307 node_t *node = get_irn_node(ka);
3310 node = get_irn_node(get_nodes_block(ka));
3312 if (node->type.tv != tarval_unreachable && !is_Bad(ka))
3316 set_End_keepalives(end, j, in);
3321 #define SET(code) op_##code->ops.generic = (op_func)compute_##code
3324 * sets the generic functions to compute.
3326 static void set_compute_functions(void) {
3329 /* set the default compute function */
3330 for (i = get_irp_n_opcodes() - 1; i >= 0; --i) {
3331 ir_op *op = get_irp_opcode(i);
3332 op->ops.generic = (op_func)default_compute;
3335 /* set specific functions */
3356 } /* set_compute_functions */
3358 void combo(ir_graph *irg) {
3360 ir_node *initial_bl;
3362 ir_graph *rem = current_ir_graph;
3364 current_ir_graph = irg;
3366 /* register a debug mask */
3367 FIRM_DBG_REGISTER(dbg, "firm.opt.combo");
3369 DB((dbg, LEVEL_1, "Doing COMBO for %+F\n", irg));
3371 obstack_init(&env.obst);
3372 env.worklist = NULL;
3376 #ifdef DEBUG_libfirm
3377 env.dbg_list = NULL;
3379 env.opcode2id_map = new_set(cmp_opcode, iro_Last * 4);
3380 env.type2id_map = pmap_create();
3381 env.end_idx = get_opt_global_cse() ? 0 : -1;
3382 env.lambda_input = 0;
3383 env.commutative = 1;
3386 assure_irg_outs(irg);
3387 assure_cf_loop(irg);
3389 /* we have our own value_of function */
3390 set_value_of_func(get_node_tarval);
3392 set_compute_functions();
3393 DEBUG_ONLY(part_nr = 0);
3395 ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK);
3397 /* create the initial partition and place it on the work list */
3398 env.initial = new_partition(&env);
3399 add_to_worklist(env.initial, &env);
3400 irg_walk_graph(irg, create_initial_partitions, init_block_phis, &env);
3402 /* set the hook: from now, every node has a partition and a type */
3403 DEBUG_ONLY(set_dump_node_vcgattr_hook(dump_partition_hook));
3405 /* all nodes on the initial partition have type Top */
3406 env.initial->type_is_T_or_C = 1;
3408 /* Place the START Node's partition on cprop.
3409 Place the START Node on its local worklist. */
3410 initial_bl = get_irg_start_block(irg);
3411 start = get_irn_node(initial_bl);
3412 add_to_cprop(start, &env);
3416 if (env.worklist != NULL)
3418 } while (env.cprop != NULL || env.worklist != NULL);
3420 dump_all_partitions(&env);
3421 check_all_partitions(&env);
3424 dump_ir_block_graph(irg, "-partition");
3427 /* apply the result */
3428 irg_block_walk_graph(irg, NULL, apply_cf, &env);
3429 /* Kill keep-alives of dead blocks: this speeds up apply_result()
3430 * and fixes assertion because dead cf to dead blocks is NOT removed by
3432 apply_end(get_irg_end(irg), &env);
3433 irg_walk_graph(irg, NULL, apply_result, &env);
3436 /* control flow might changed */
3437 set_irg_outs_inconsistent(irg);
3438 set_irg_extblk_inconsistent(irg);
3439 set_irg_doms_inconsistent(irg);
3440 set_irg_loopinfo_inconsistent(irg);
3443 ir_free_resources(irg, IR_RESOURCE_IRN_LINK);
3445 /* remove the partition hook */
3446 DEBUG_ONLY(set_dump_node_vcgattr_hook(NULL));
3448 pmap_destroy(env.type2id_map);
3449 del_set(env.opcode2id_map);
3450 obstack_free(&env.obst, NULL);
3452 /* restore value_of() default behavior */
3453 set_value_of_func(NULL);
3454 current_ir_graph = rem;