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
96 /* allow optimization of non-strict programs */
99 typedef struct node_t node_t;
100 typedef struct partition_t partition_t;
101 typedef struct opcode_key_t opcode_key_t;
102 typedef struct listmap_entry_t listmap_entry_t;
104 /** The type of the compute function. */
105 typedef void (*compute_func)(node_t *node);
110 struct opcode_key_t {
111 ir_opcode code; /**< The Firm opcode. */
112 ir_mode *mode; /**< The mode of all nodes in the partition. */
113 int arity; /**< The arity of this opcode (needed for Phi etc. */
115 long proj; /**< For Proj nodes, its proj number */
116 ir_entity *ent; /**< For Sel Nodes, its entity */
121 * An entry in the list_map.
123 struct listmap_entry_t {
124 void *id; /**< The id. */
125 node_t *list; /**< The associated list for this id. */
126 listmap_entry_t *next; /**< Link to the next entry in the map. */
129 /** We must map id's to lists. */
130 typedef struct listmap_t {
131 set *map; /**< Map id's to listmap_entry_t's */
132 listmap_entry_t *values; /**< List of all values in the map. */
136 * A lattice element. Because we handle constants and symbolic constants different, we
137 * have to use this union.
148 ir_node *node; /**< The IR-node itself. */
149 list_head node_list; /**< Double-linked list of leader/follower entries. */
150 list_head cprop_list; /**< Double-linked partition.cprop list. */
151 partition_t *part; /**< points to the partition this node belongs to */
152 node_t *next; /**< Next node on local list (partition.touched, fallen). */
153 node_t *race_next; /**< Next node on race list. */
154 lattice_elem_t type; /**< The associated lattice element "type". */
155 int max_user_input; /**< Maximum input number of Def-Use edges. */
156 int next_edge; /**< Index of the next Def-Use edge to use. */
157 int n_followers; /**< Number of Follower in the outs set. */
158 unsigned on_touched:1; /**< Set, if this node is on the partition.touched set. */
159 unsigned on_cprop:1; /**< Set, if this node is on the partition.cprop list. */
160 unsigned on_fallen:1; /**< Set, if this node is on the fallen list. */
161 unsigned is_follower:1; /**< Set, if this node is a follower. */
162 unsigned by_all_const:1; /**< Set, if this node was once evaluated by all constants. */
163 unsigned flagged:2; /**< 2 Bits, set if this node was visited by race 1 or 2. */
167 * A partition containing congruent nodes.
170 list_head Leader; /**< The head of partition Leader node list. */
171 list_head Follower; /**< The head of partition Follower node list. */
172 list_head cprop; /**< The head of partition.cprop list. */
173 list_head cprop_X; /**< The head of partition.cprop (Cond nodes and its Projs) list. */
174 partition_t *wl_next; /**< Next entry in the work list if any. */
175 partition_t *touched_next; /**< Points to the next partition in the touched set. */
176 partition_t *cprop_next; /**< Points to the next partition in the cprop list. */
177 partition_t *split_next; /**< Points to the next partition in the list that must be split by split_by(). */
178 node_t *touched; /**< The partition.touched set of this partition. */
179 unsigned n_leader; /**< Number of entries in this partition.Leader. */
180 unsigned n_touched; /**< Number of entries in the partition.touched. */
181 int max_user_inputs; /**< Maximum number of user inputs of all entries. */
182 unsigned on_worklist:1; /**< Set, if this partition is in the work list. */
183 unsigned on_touched:1; /**< Set, if this partition is on the touched set. */
184 unsigned on_cprop:1; /**< Set, if this partition is on the cprop list. */
185 unsigned type_is_T_or_C:1;/**< Set, if all nodes in this partition have type Top or Constant. */
187 partition_t *dbg_next; /**< Link all partitions for debugging */
188 unsigned nr; /**< A unique number for (what-)mapping, >0. */
192 typedef struct environment_t {
193 struct obstack obst; /**< obstack to allocate data structures. */
194 partition_t *worklist; /**< The work list. */
195 partition_t *cprop; /**< The constant propagation list. */
196 partition_t *touched; /**< the touched set. */
197 partition_t *initial; /**< The initial partition. */
198 set *opcode2id_map; /**< The opcodeMode->id map. */
199 pmap *type2id_map; /**< The type->id map. */
200 int end_idx; /**< -1 for local and 0 for global congruences. */
201 int lambda_input; /**< Captured argument for lambda_partition(). */
202 unsigned modified:1; /**< Set, if the graph was modified. */
203 unsigned commutative:1; /**< Set, if commutation nodes should be handled specially. */
204 unsigned unopt_cf:1; /**< If set, control flow is not optimized due to Unknown. */
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;
783 ir_node *irn = y->node;
785 /* place Conds and all its Projs on the cprop_X list */
786 if (is_Cond(skip_Proj(irn)))
787 list_add_tail(&y->cprop_list, &Y->cprop_X);
789 list_add_tail(&y->cprop_list, &Y->cprop);
792 DB((dbg, LEVEL_3, "Add %+F to part%u.cprop\n", y->node, Y->nr));
794 /* place its partition on the cprop list */
795 if (Y->on_cprop == 0) {
796 Y->cprop_next = env->cprop;
802 if (get_irn_mode(irn) == mode_T) {
803 /* mode_T nodes always produce tarval_bottom, so we must explicitly
804 add it's Proj's to get constant evaluation to work */
807 for (i = get_irn_n_outs(irn) - 1; i >= 0; --i) {
808 node_t *proj = get_irn_node(get_irn_out(irn, i));
810 add_to_cprop(proj, env);
812 } else if (is_Block(irn)) {
813 /* Due to the way we handle Phi's, we must place all Phis of a block on the list
814 * if someone placed the block. The Block is only placed if the reachability
815 * changes, and this must be re-evaluated in compute_Phi(). */
817 for (phi = get_Block_phis(irn); phi != NULL; phi = get_Phi_next(phi)) {
818 node_t *p = get_irn_node(phi);
819 add_to_cprop(p, env);
825 * Update the worklist: If Z is on worklist then add Z' to worklist.
826 * Else add the smaller of Z and Z' to worklist.
828 * @param Z the Z partition
829 * @param Z_prime the Z' partition, a previous part of Z
830 * @param env the environment
832 static void update_worklist(partition_t *Z, partition_t *Z_prime, environment_t *env) {
833 if (Z->on_worklist || Z_prime->n_leader < Z->n_leader) {
834 add_to_worklist(Z_prime, env);
836 add_to_worklist(Z, env);
838 } /* update_worklist */
841 * Make all inputs to x no longer be F.def_use edges.
845 static void move_edges_to_leader(node_t *x) {
846 ir_node *irn = x->node;
849 for (i = get_irn_arity(irn) - 1; i >= 0; --i) {
850 node_t *pred = get_irn_node(get_irn_n(irn, i));
855 n = get_irn_n_outs(p);
856 for (j = 1; j <= pred->n_followers; ++j) {
857 if (p->out[j].pos == i && p->out[j].use == irn) {
858 /* found a follower edge to x, move it to the Leader */
859 ir_def_use_edge edge = p->out[j];
861 /* remove this edge from the Follower set */
862 p->out[j] = p->out[pred->n_followers];
865 /* sort it into the leader set */
866 for (k = pred->n_followers + 2; k <= n; ++k) {
867 if (p->out[k].pos >= edge.pos)
869 p->out[k - 1] = p->out[k];
871 /* place the new edge here */
872 p->out[k - 1] = edge;
874 /* edge found and moved */
879 } /* move_edges_to_leader */
882 * Split a partition that has NO followers by a local list.
884 * @param Z partition to split
885 * @param g a (non-empty) node list
886 * @param env the environment
888 * @return a new partition containing the nodes of g
890 static partition_t *split_no_followers(partition_t *Z, node_t *g, environment_t *env) {
891 partition_t *Z_prime;
896 dump_partition("Splitting ", Z);
897 dump_list("by list ", g);
901 /* Remove g from Z. */
902 for (node = g; node != NULL; node = node->next) {
903 assert(node->part == Z);
904 list_del(&node->node_list);
907 assert(n < Z->n_leader);
910 /* Move g to a new partition, Z'. */
911 Z_prime = new_partition(env);
913 for (node = g; node != NULL; node = node->next) {
914 list_add_tail(&node->node_list, &Z_prime->Leader);
915 node->part = Z_prime;
916 if (node->max_user_input > max_input)
917 max_input = node->max_user_input;
919 Z_prime->max_user_inputs = max_input;
920 Z_prime->n_leader = n;
923 check_partition(Z_prime);
925 /* for now, copy the type info tag, it will be adjusted in split_by(). */
926 Z_prime->type_is_T_or_C = Z->type_is_T_or_C;
928 update_worklist(Z, Z_prime, env);
930 dump_partition("Now ", Z);
931 dump_partition("Created new ", Z_prime);
933 } /* split_no_followers */
936 * Make the Follower -> Leader transition for a node.
940 static void follower_to_leader(node_t *n) {
941 assert(n->is_follower == 1);
943 DB((dbg, LEVEL_2, "%+F make the follower -> leader transition\n", n->node));
945 move_edges_to_leader(n);
946 list_del(&n->node_list);
947 list_add_tail(&n->node_list, &n->part->Leader);
949 } /* follower_to_leader */
952 * The environment for one race step.
954 typedef struct step_env {
955 node_t *initial; /**< The initial node list. */
956 node_t *unwalked; /**< The unwalked node list. */
957 node_t *walked; /**< The walked node list. */
958 int index; /**< Next index of Follower use_def edge. */
959 unsigned side; /**< side number. */
963 * Return non-zero, if a input is a real follower
965 * @param irn the node to check
966 * @param input number of the input
968 static int is_real_follower(const ir_node *irn, int input) {
971 switch (get_irn_opcode(irn)) {
974 /* ignore the Confirm bound input */
980 /* ignore the Mux sel input */
985 /* dead inputs are not follower edges */
986 ir_node *block = get_nodes_block(irn);
987 node_t *pred = get_irn_node(get_Block_cfgpred(block, input));
989 if (pred->type.tv == tarval_unreachable)
999 /* only a Sub x,0 / Shift x,0 might be a follower */
1006 pred = get_irn_node(get_irn_n(irn, input));
1007 if (is_tarval(pred->type.tv) && tarval_is_null(pred->type.tv))
1011 pred = get_irn_node(get_irn_n(irn, input));
1012 if (is_tarval(pred->type.tv) && tarval_is_one(pred->type.tv))
1016 pred = get_irn_node(get_irn_n(irn, input));
1017 if (is_tarval(pred->type.tv) && tarval_is_all_one(pred->type.tv))
1022 /* all inputs are followers */
1025 assert(!"opcode not implemented yet");
1029 } /* is_real_follower */
1032 * Do one step in the race.
1034 static int step(step_env *env) {
1037 if (env->initial != NULL) {
1038 /* Move node from initial to unwalked */
1040 env->initial = n->race_next;
1042 n->race_next = env->unwalked;
1048 while (env->unwalked != NULL) {
1049 /* let n be the first node in unwalked */
1051 while (env->index < n->n_followers) {
1052 const ir_def_use_edge *edge = &n->node->out[1 + env->index];
1054 /* let m be n.F.def_use[index] */
1055 node_t *m = get_irn_node(edge->use);
1057 assert(m->is_follower);
1059 * Some inputs, like the get_Confirm_bound are NOT
1060 * real followers, sort them out.
1062 if (! is_real_follower(m->node, edge->pos)) {
1068 /* only followers from our partition */
1069 if (m->part != n->part)
1072 if ((m->flagged & env->side) == 0) {
1073 m->flagged |= env->side;
1075 if (m->flagged != 3) {
1076 /* visited the first time */
1077 /* add m to unwalked not as first node (we might still need to
1078 check for more follower node */
1079 m->race_next = n->race_next;
1083 /* else already visited by the other side and on the other list */
1086 /* move n to walked */
1087 env->unwalked = n->race_next;
1088 n->race_next = env->walked;
1096 * Clear the flags from a list and check for
1097 * nodes that where touched from both sides.
1099 * @param list the list
1101 static int clear_flags(node_t *list) {
1105 for (n = list; n != NULL; n = n->race_next) {
1106 if (n->flagged == 3) {
1107 /* we reach a follower from both sides, this will split congruent
1108 * inputs and make it a leader. */
1109 follower_to_leader(n);
1118 * Split a partition by a local list using the race.
1120 * @param pX pointer to the partition to split, might be changed!
1121 * @param gg a (non-empty) node list
1122 * @param env the environment
1124 * @return a new partition containing the nodes of gg
1126 static partition_t *split(partition_t **pX, node_t *gg, environment_t *env) {
1127 partition_t *X = *pX;
1128 partition_t *X_prime;
1131 node_t *g, *h, *node, *t;
1132 int max_input, transitions, winner, shf;
1134 DEBUG_ONLY(static int run = 0;)
1136 DB((dbg, LEVEL_2, "Run %d ", run++));
1137 if (list_empty(&X->Follower)) {
1138 /* if the partition has NO follower, we can use the fast
1139 splitting algorithm. */
1140 return split_no_followers(X, gg, env);
1142 /* else do the race */
1144 dump_partition("Splitting ", X);
1145 dump_list("by list ", gg);
1147 INIT_LIST_HEAD(&tmp);
1149 /* Remove gg from X.Leader and put into g */
1151 for (node = gg; node != NULL; node = node->next) {
1152 assert(node->part == X);
1153 assert(node->is_follower == 0);
1155 list_del(&node->node_list);
1156 list_add_tail(&node->node_list, &tmp);
1157 node->race_next = g;
1162 list_for_each_entry(node_t, node, &X->Leader, node_list) {
1163 node->race_next = h;
1166 /* restore X.Leader */
1167 list_splice(&tmp, &X->Leader);
1169 senv[0].initial = g;
1170 senv[0].unwalked = NULL;
1171 senv[0].walked = NULL;
1175 senv[1].initial = h;
1176 senv[1].unwalked = NULL;
1177 senv[1].walked = NULL;
1182 * Some informations on the race that are not stated clearly in Click's
1184 * 1) A follower stays on the side that reach him first.
1185 * 2) If the other side reches a follower, if will be converted to
1186 * a leader. /This must be done after the race is over, else the
1187 * edges we are iterating on are renumbered./
1188 * 3) /New leader might end up on both sides./
1189 * 4) /If one side ends up with new Leaders, we must ensure that
1190 * they can split out by opcode, hence we have to put _every_
1191 * partition with new Leader nodes on the cprop list, as
1192 * opcode splitting is done by split_by() at the end of
1193 * constant propagation./
1196 if (step(&senv[0])) {
1200 if (step(&senv[1])) {
1205 assert(senv[winner].initial == NULL);
1206 assert(senv[winner].unwalked == NULL);
1208 /* clear flags from walked/unwalked */
1210 transitions = clear_flags(senv[0].unwalked) << shf;
1211 transitions |= clear_flags(senv[0].walked) << shf;
1213 transitions |= clear_flags(senv[1].unwalked) << shf;
1214 transitions |= clear_flags(senv[1].walked) << shf;
1216 dump_race_list("winner ", senv[winner].walked);
1218 /* Move walked_{winner} to a new partition, X'. */
1219 X_prime = new_partition(env);
1222 for (node = senv[winner].walked; node != NULL; node = node->race_next) {
1223 list_del(&node->node_list);
1224 node->part = X_prime;
1225 if (node->is_follower) {
1226 list_add_tail(&node->node_list, &X_prime->Follower);
1228 list_add_tail(&node->node_list, &X_prime->Leader);
1231 if (node->max_user_input > max_input)
1232 max_input = node->max_user_input;
1234 X_prime->n_leader = n;
1235 X_prime->max_user_inputs = max_input;
1236 X->n_leader -= X_prime->n_leader;
1238 /* for now, copy the type info tag, it will be adjusted in split_by(). */
1239 X_prime->type_is_T_or_C = X->type_is_T_or_C;
1242 * Even if a follower was not checked by both sides, it might have
1243 * loose its congruence, so we need to check this case for all follower.
1245 list_for_each_entry_safe(node_t, node, t, &X_prime->Follower, node_list) {
1246 if (identity(node) == node) {
1247 follower_to_leader(node);
1253 check_partition(X_prime);
1255 /* X' is the smaller part */
1256 add_to_worklist(X_prime, env);
1259 * If there where follower to leader transitions, ensure that the nodes
1260 * can be split out if necessary.
1262 if (transitions & 1) {
1263 /* place winner partition on the cprop list */
1264 if (X_prime->on_cprop == 0) {
1265 X_prime->cprop_next = env->cprop;
1266 env->cprop = X_prime;
1267 X_prime->on_cprop = 1;
1270 if (transitions & 2) {
1271 /* place other partition on the cprop list */
1272 if (X->on_cprop == 0) {
1273 X->cprop_next = env->cprop;
1279 dump_partition("Now ", X);
1280 dump_partition("Created new ", X_prime);
1282 /* we have to ensure that the partition containing g is returned */
1292 * Returns non-zero if the i'th input of a Phi node is live.
1294 * @param phi a Phi-node
1295 * @param i an input number
1297 * @return non-zero if the i'th input of the given Phi node is live
1299 static int is_live_input(ir_node *phi, int i) {
1301 ir_node *block = get_nodes_block(phi);
1302 ir_node *pred = get_Block_cfgpred(block, i);
1303 lattice_elem_t type = get_node_type(pred);
1305 return type.tv != tarval_unreachable;
1307 /* else it's the control input, always live */
1309 } /* is_live_input */
1312 * Return non-zero if a type is a constant.
1314 static int is_constant_type(lattice_elem_t type) {
1315 if (type.tv != tarval_bottom && type.tv != tarval_top)
1318 } /* is_constant_type */
1321 * Check whether a type is neither Top or a constant.
1322 * Note: U is handled like Top here, R is a constant.
1324 * @param type the type to check
1326 static int type_is_neither_top_nor_const(const lattice_elem_t type) {
1327 if (is_tarval(type.tv)) {
1328 if (type.tv == tarval_top)
1330 if (tarval_is_constant(type.tv))
1337 } /* type_is_neither_top_nor_const */
1340 * Collect nodes to the touched list.
1342 * @param list the list which contains the nodes that must be evaluated
1343 * @param idx the index of the def_use edge to evaluate
1344 * @param env the environment
1346 static void collect_touched(list_head *list, int idx, environment_t *env) {
1348 int end_idx = env->end_idx;
1350 list_for_each_entry(node_t, x, list, node_list) {
1354 /* leader edges start AFTER follower edges */
1355 x->next_edge = x->n_followers + 1;
1357 num_edges = get_irn_n_outs(x->node);
1359 /* for all edges in x.L.def_use_{idx} */
1360 while (x->next_edge <= num_edges) {
1361 const ir_def_use_edge *edge = &x->node->out[x->next_edge];
1364 /* check if we have necessary edges */
1365 if (edge->pos > idx)
1372 /* only non-commutative nodes */
1373 if (env->commutative &&
1374 (idx == 0 || idx == 1) && is_op_commutative(get_irn_op(succ)))
1377 /* ignore the "control input" for non-pinned nodes
1378 if we are running in GCSE mode */
1379 if (idx < end_idx && get_irn_pinned(succ) != op_pin_state_pinned)
1382 y = get_irn_node(succ);
1383 assert(get_irn_n(succ, idx) == x->node);
1385 /* ignore block edges touching followers */
1386 if (idx == -1 && y->is_follower)
1389 if (is_constant_type(y->type)) {
1390 ir_opcode code = get_irn_opcode(succ);
1391 if (code == iro_Sub || code == iro_Cmp)
1392 add_to_cprop(y, env);
1395 /* Partitions of constants should not be split simply because their Nodes have unequal
1396 functions or incongruent inputs. */
1397 if (type_is_neither_top_nor_const(y->type) &&
1398 (! is_Phi(y->node) || is_live_input(y->node, idx))) {
1399 add_to_touched(y, env);
1403 } /* collect_touched */
1406 * Collect commutative nodes to the touched list.
1408 * @param list the list which contains the nodes that must be evaluated
1409 * @param env the environment
1411 static void collect_commutative_touched(list_head *list, environment_t *env) {
1414 list_for_each_entry(node_t, x, list, node_list) {
1417 num_edges = get_irn_n_outs(x->node);
1419 x->next_edge = x->n_followers + 1;
1421 /* for all edges in x.L.def_use_{idx} */
1422 while (x->next_edge <= num_edges) {
1423 const ir_def_use_edge *edge = &x->node->out[x->next_edge];
1426 /* check if we have necessary edges */
1436 /* only commutative nodes */
1437 if (!is_op_commutative(get_irn_op(succ)))
1440 y = get_irn_node(succ);
1441 if (is_constant_type(y->type)) {
1442 ir_opcode code = get_irn_opcode(succ);
1443 if (code == iro_Eor)
1444 add_to_cprop(y, env);
1447 /* Partitions of constants should not be split simply because their Nodes have unequal
1448 functions or incongruent inputs. */
1449 if (type_is_neither_top_nor_const(y->type)) {
1450 add_to_touched(y, env);
1454 } /* collect_commutative_touched */
1457 * Split the partitions if caused by the first entry on the worklist.
1459 * @param env the environment
1461 static void cause_splits(environment_t *env) {
1462 partition_t *X, *Z, *N;
1465 /* remove the first partition from the worklist */
1467 env->worklist = X->wl_next;
1470 dump_partition("Cause_split: ", X);
1472 if (env->commutative) {
1473 /* handle commutative nodes first */
1475 /* empty the touched set: already done, just clear the list */
1476 env->touched = NULL;
1478 collect_commutative_touched(&X->Leader, env);
1479 collect_commutative_touched(&X->Follower, env);
1481 for (Z = env->touched; Z != NULL; Z = N) {
1483 node_t *touched = Z->touched;
1484 unsigned n_touched = Z->n_touched;
1486 assert(Z->touched != NULL);
1488 /* beware, split might change Z */
1489 N = Z->touched_next;
1491 /* remove it from the touched set */
1494 /* Empty local Z.touched. */
1495 for (e = touched; e != NULL; e = e->next) {
1496 assert(e->is_follower == 0);
1502 if (0 < n_touched && n_touched < Z->n_leader) {
1503 DB((dbg, LEVEL_2, "Split part%d by touched\n", Z->nr));
1504 split(&Z, touched, env);
1506 assert(n_touched <= Z->n_leader);
1510 /* combine temporary leader and follower list */
1511 for (idx = -1; idx <= X->max_user_inputs; ++idx) {
1512 /* empty the touched set: already done, just clear the list */
1513 env->touched = NULL;
1515 collect_touched(&X->Leader, idx, env);
1516 collect_touched(&X->Follower, idx, env);
1518 for (Z = env->touched; Z != NULL; Z = N) {
1520 node_t *touched = Z->touched;
1521 unsigned n_touched = Z->n_touched;
1523 assert(Z->touched != NULL);
1525 /* beware, split might change Z */
1526 N = Z->touched_next;
1528 /* remove it from the touched set */
1531 /* Empty local Z.touched. */
1532 for (e = touched; e != NULL; e = e->next) {
1533 assert(e->is_follower == 0);
1539 if (0 < n_touched && n_touched < Z->n_leader) {
1540 DB((dbg, LEVEL_2, "Split part%d by touched\n", Z->nr));
1541 split(&Z, touched, env);
1543 assert(n_touched <= Z->n_leader);
1546 } /* cause_splits */
1549 * Implements split_by_what(): Split a partition by characteristics given
1550 * by the what function.
1552 * @param X the partition to split
1553 * @param What a function returning an Id for every node of the partition X
1554 * @param P a list to store the result partitions
1555 * @param env the environment
1559 static partition_t *split_by_what(partition_t *X, what_func What,
1560 partition_t **P, environment_t *env) {
1563 listmap_entry_t *iter;
1566 /* Let map be an empty mapping from the range of What to (local) list of Nodes. */
1568 list_for_each_entry(node_t, x, &X->Leader, node_list) {
1569 void *id = What(x, env);
1570 listmap_entry_t *entry;
1573 /* input not allowed, ignore */
1576 /* Add x to map[What(x)]. */
1577 entry = listmap_find(&map, id);
1578 x->next = entry->list;
1581 /* Let P be a set of Partitions. */
1583 /* for all sets S except one in the range of map do */
1584 for (iter = map.values; iter != NULL; iter = iter->next) {
1585 if (iter->next == NULL) {
1586 /* this is the last entry, ignore */
1591 /* Add SPLIT( X, S ) to P. */
1592 DB((dbg, LEVEL_2, "Split part%d by WHAT = %s\n", X->nr, what_reason));
1593 R = split(&X, S, env);
1603 } /* split_by_what */
1605 /** lambda n.(n.type) */
1606 static void *lambda_type(const node_t *node, environment_t *env) {
1608 return node->type.tv;
1611 /** lambda n.(n.opcode) */
1612 static void *lambda_opcode(const node_t *node, environment_t *env) {
1613 opcode_key_t key, *entry;
1614 ir_node *irn = node->node;
1616 key.code = get_irn_opcode(irn);
1617 key.mode = get_irn_mode(irn);
1618 key.arity = get_irn_arity(irn);
1622 switch (get_irn_opcode(irn)) {
1624 key.u.proj = get_Proj_proj(irn);
1627 key.u.ent = get_Sel_entity(irn);
1633 entry = set_insert(env->opcode2id_map, &key, sizeof(key), opcode_hash(&key));
1635 } /* lambda_opcode */
1637 /** lambda n.(n[i].partition) */
1638 static void *lambda_partition(const node_t *node, environment_t *env) {
1639 ir_node *skipped = skip_Proj(node->node);
1642 int i = env->lambda_input;
1644 if (i >= get_irn_arity(node->node)) {
1646 * We are outside the allowed range: This can happen even
1647 * if we have split by opcode first: doing so might move Followers
1648 * to Leaders and those will have a different opcode!
1649 * Note that in this case the partition is on the cprop list and will be
1655 /* ignore the "control input" for non-pinned nodes
1656 if we are running in GCSE mode */
1657 if (i < env->end_idx && get_irn_pinned(skipped) != op_pin_state_pinned)
1660 pred = i == -1 ? get_irn_n(skipped, i) : get_irn_n(node->node, i);
1661 p = get_irn_node(pred);
1664 } /* lambda_partition */
1666 /** lambda n.(n[i].partition) for commutative nodes */
1667 static void *lambda_commutative_partition(const node_t *node, environment_t *env) {
1668 ir_node *irn = node->node;
1669 ir_node *skipped = skip_Proj(irn);
1670 ir_node *pred, *left, *right;
1672 partition_t *pl, *pr;
1673 int i = env->lambda_input;
1675 if (i >= get_irn_arity(node->node)) {
1677 * We are outside the allowed range: This can happen even
1678 * if we have split by opcode first: doing so might move Followers
1679 * to Leaders and those will have a different opcode!
1680 * Note that in this case the partition is on the cprop list and will be
1686 /* ignore the "control input" for non-pinned nodes
1687 if we are running in GCSE mode */
1688 if (i < env->end_idx && get_irn_pinned(skipped) != op_pin_state_pinned)
1692 pred = get_irn_n(skipped, i);
1693 p = get_irn_node(pred);
1697 if (is_op_commutative(get_irn_op(irn))) {
1698 /* normalize partition order by returning the "smaller" on input 0,
1699 the "bigger" on input 1. */
1700 left = get_binop_left(irn);
1701 pl = get_irn_node(left)->part;
1702 right = get_binop_right(irn);
1703 pr = get_irn_node(right)->part;
1706 return pl < pr ? pl : pr;
1708 return pl > pr ? pl : pr;
1710 /* a not split out Follower */
1711 pred = get_irn_n(irn, i);
1712 p = get_irn_node(pred);
1716 } /* lambda_commutative_partition */
1719 * Returns true if a type is a constant (and NOT Top
1722 static int is_con(const lattice_elem_t type) {
1723 /* be conservative */
1724 if (is_tarval(type.tv))
1725 return tarval_is_constant(type.tv);
1726 return is_entity(type.sym.entity_p);
1730 * Implements split_by().
1732 * @param X the partition to split
1733 * @param env the environment
1735 static void split_by(partition_t *X, environment_t *env) {
1736 partition_t *I, *P = NULL;
1739 dump_partition("split_by", X);
1741 if (X->n_leader == 1) {
1742 /* we have only one leader, no need to split, just check it's type */
1743 node_t *x = get_first_node(X);
1744 X->type_is_T_or_C = x->type.tv == tarval_top || is_con(x->type);
1748 DEBUG_ONLY(what_reason = "lambda n.(n.type)";)
1749 P = split_by_what(X, lambda_type, &P, env);
1752 /* adjust the type tags, we have split partitions by type */
1753 for (I = P; I != NULL; I = I->split_next) {
1754 node_t *x = get_first_node(I);
1755 I->type_is_T_or_C = x->type.tv == tarval_top || is_con(x->type);
1762 if (Y->n_leader > 1) {
1763 /* we do not want split the TOP or constant partitions */
1764 if (! Y->type_is_T_or_C) {
1765 partition_t *Q = NULL;
1767 DEBUG_ONLY(what_reason = "lambda n.(n.opcode)";)
1768 Q = split_by_what(Y, lambda_opcode, &Q, env);
1775 if (Z->n_leader > 1) {
1776 const node_t *first = get_first_node(Z);
1777 int arity = get_irn_arity(first->node);
1779 what_func what = lambda_partition;
1780 DEBUG_ONLY(char buf[64];)
1782 if (env->commutative && is_op_commutative(get_irn_op(first->node)))
1783 what = lambda_commutative_partition;
1786 * BEWARE: during splitting by input 2 for instance we might
1787 * create new partitions which are different by input 1, so collect
1788 * them and split further.
1790 Z->split_next = NULL;
1793 for (input = arity - 1; input >= -1; --input) {
1795 partition_t *Z_prime = R;
1798 if (Z_prime->n_leader > 1) {
1799 env->lambda_input = input;
1800 DEBUG_ONLY(snprintf(buf, sizeof(buf), "lambda n.(n[%d].partition)", input);)
1801 DEBUG_ONLY(what_reason = buf;)
1802 S = split_by_what(Z_prime, what, &S, env);
1805 Z_prime->split_next = S;
1808 } while (R != NULL);
1813 } while (Q != NULL);
1816 } while (P != NULL);
1820 * (Re-)compute the type for a given node.
1822 * @param node the node
1824 static void default_compute(node_t *node) {
1826 ir_node *irn = node->node;
1827 node_t *block = get_irn_node(get_nodes_block(irn));
1829 if (block->type.tv == tarval_unreachable) {
1830 node->type.tv = tarval_top;
1834 /* if any of the data inputs have type top, the result is type top */
1835 for (i = get_irn_arity(irn) - 1; i >= 0; --i) {
1836 ir_node *pred = get_irn_n(irn, i);
1837 node_t *p = get_irn_node(pred);
1839 if (p->type.tv == tarval_top) {
1840 node->type.tv = tarval_top;
1845 if (get_irn_mode(node->node) == mode_X)
1846 node->type.tv = tarval_reachable;
1848 node->type.tv = computed_value(irn);
1849 } /* default_compute */
1852 * (Re-)compute the type for a Block node.
1854 * @param node the node
1856 static void compute_Block(node_t *node) {
1858 ir_node *block = node->node;
1860 if (block == get_irg_start_block(current_ir_graph)) {
1861 /* start block is always reachable */
1862 node->type.tv = tarval_reachable;
1866 for (i = get_Block_n_cfgpreds(block) - 1; i >= 0; --i) {
1867 node_t *pred = get_irn_node(get_Block_cfgpred(block, i));
1869 if (pred->type.tv == tarval_reachable) {
1870 /* A block is reachable, if at least of predecessor is reachable. */
1871 node->type.tv = tarval_reachable;
1875 node->type.tv = tarval_top;
1876 } /* compute_Block */
1879 * (Re-)compute the type for a Bad node.
1881 * @param node the node
1883 static void compute_Bad(node_t *node) {
1884 /* Bad nodes ALWAYS compute Top */
1885 node->type.tv = tarval_top;
1889 * (Re-)compute the type for an Unknown node.
1891 * @param node the node
1893 static void compute_Unknown(node_t *node) {
1894 /* While Unknown nodes should compute Top this is dangerous:
1895 * a Top input to a Cond would lead to BOTH control flows unreachable.
1896 * While this is correct in the given semantics, it would destroy the Firm
1899 * It would be safe to compute Top IF it can be assured, that only Cmp
1900 * nodes are inputs to Conds. We check that first.
1901 * This is the way Frontends typically build Firm, but some optimizations
1902 * (cond_eval for instance) might replace them by Phib's...
1904 node->type.tv = tarval_UNKNOWN;
1905 } /* compute_Unknown */
1908 * (Re-)compute the type for a Jmp node.
1910 * @param node the node
1912 static void compute_Jmp(node_t *node) {
1913 node_t *block = get_irn_node(get_nodes_block(node->node));
1915 node->type = block->type;
1919 * (Re-)compute the type for the Return node.
1921 * @param node the node
1923 static void compute_Return(node_t *node) {
1924 /* The Return node is NOT dead if it is in a reachable block.
1925 * This is already checked in compute(). so we can return
1926 * Reachable here. */
1927 node->type.tv = tarval_reachable;
1928 } /* compute_Return */
1931 * (Re-)compute the type for the End node.
1933 * @param node the node
1935 static void compute_End(node_t *node) {
1936 /* the End node is NOT dead of course */
1937 node->type.tv = tarval_reachable;
1941 * (Re-)compute the type for a SymConst node.
1943 * @param node the node
1945 static void compute_SymConst(node_t *node) {
1946 ir_node *irn = node->node;
1947 node_t *block = get_irn_node(get_nodes_block(irn));
1949 if (block->type.tv == tarval_unreachable) {
1950 node->type.tv = tarval_top;
1953 switch (get_SymConst_kind(irn)) {
1954 case symconst_addr_ent:
1955 /* case symconst_addr_name: cannot handle this yet */
1956 node->type.sym = get_SymConst_symbol(irn);
1959 node->type.tv = computed_value(irn);
1961 } /* compute_SymConst */
1964 * (Re-)compute the type for a Phi node.
1966 * @param node the node
1968 static void compute_Phi(node_t *node) {
1970 ir_node *phi = node->node;
1971 lattice_elem_t type;
1973 /* if a Phi is in a unreachable block, its type is TOP */
1974 node_t *block = get_irn_node(get_nodes_block(phi));
1976 if (block->type.tv == tarval_unreachable) {
1977 node->type.tv = tarval_top;
1981 /* Phi implements the Meet operation */
1982 type.tv = tarval_top;
1983 for (i = get_Phi_n_preds(phi) - 1; i >= 0; --i) {
1984 node_t *pred = get_irn_node(get_Phi_pred(phi, i));
1985 node_t *pred_X = get_irn_node(get_Block_cfgpred(block->node, i));
1987 if (pred_X->type.tv == tarval_unreachable || pred->type.tv == tarval_top) {
1988 /* ignore TOP inputs: We must check here for unreachable blocks,
1989 because Firm constants live in the Start Block are NEVER Top.
1990 Else, a Phi (1,2) will produce Bottom, even if the 2 for instance
1991 comes from a unreachable input. */
1994 if (pred->type.tv == tarval_bottom) {
1995 node->type.tv = tarval_bottom;
1997 } else if (type.tv == tarval_top) {
1998 /* first constant found */
2000 } else if (type.tv != pred->type.tv) {
2001 /* different constants or tarval_bottom */
2002 node->type.tv = tarval_bottom;
2005 /* else nothing, constants are the same */
2011 * (Re-)compute the type for an Add. Special case: one nodes is a Zero Const.
2013 * @param node the node
2015 static void compute_Add(node_t *node) {
2016 ir_node *sub = node->node;
2017 node_t *l = get_irn_node(get_Add_left(sub));
2018 node_t *r = get_irn_node(get_Add_right(sub));
2019 lattice_elem_t a = l->type;
2020 lattice_elem_t b = r->type;
2023 if (a.tv == tarval_top || b.tv == tarval_top) {
2024 node->type.tv = tarval_top;
2025 } else if (a.tv == tarval_bottom || b.tv == tarval_bottom) {
2026 node->type.tv = tarval_bottom;
2028 /* x + 0 = 0 + x = x, but beware of floating point +0 + -0, so we
2029 must call tarval_add() first to handle this case! */
2030 if (is_tarval(a.tv)) {
2031 if (is_tarval(b.tv)) {
2032 node->type.tv = tarval_add(a.tv, b.tv);
2035 mode = get_tarval_mode(a.tv);
2036 if (a.tv == get_mode_null(mode)) {
2040 } else if (is_tarval(b.tv)) {
2041 mode = get_tarval_mode(b.tv);
2042 if (b.tv == get_mode_null(mode)) {
2047 node->type.tv = tarval_bottom;
2052 * (Re-)compute the type for a Sub. Special case: both nodes are congruent.
2054 * @param node the node
2056 static void compute_Sub(node_t *node) {
2057 ir_node *sub = node->node;
2058 node_t *l = get_irn_node(get_Sub_left(sub));
2059 node_t *r = get_irn_node(get_Sub_right(sub));
2060 lattice_elem_t a = l->type;
2061 lattice_elem_t b = r->type;
2064 if (a.tv == tarval_top || b.tv == tarval_top) {
2065 node->type.tv = tarval_top;
2066 } else if (is_con(a) && is_con(b)) {
2067 if (is_tarval(a.tv) && is_tarval(b.tv)) {
2068 node->type.tv = tarval_sub(a.tv, b.tv, get_irn_mode(sub));
2069 } else if (is_tarval(a.tv) && tarval_is_null(a.tv)) {
2071 } else if (is_tarval(b.tv) && tarval_is_null(b.tv)) {
2074 node->type.tv = tarval_bottom;
2076 node->by_all_const = 1;
2077 } else if (r->part == l->part &&
2078 (!mode_is_float(get_irn_mode(l->node)))) {
2080 * BEWARE: a - a is NOT always 0 for floating Point values, as
2081 * NaN op NaN = NaN, so we must check this here.
2083 ir_mode *mode = get_irn_mode(sub);
2084 tv = get_mode_null(mode);
2086 /* if the node was ONCE evaluated by all constants, but now
2087 this breaks AND we get from the argument partitions a different
2088 result, switch to bottom.
2089 This happens because initially all nodes are in the same partition ... */
2090 if (node->by_all_const && node->type.tv != tv)
2094 node->type.tv = tarval_bottom;
2099 * (Re-)compute the type for an Eor. Special case: both nodes are congruent.
2101 * @param node the node
2103 static void compute_Eor(node_t *node) {
2104 ir_node *eor = node->node;
2105 node_t *l = get_irn_node(get_Eor_left(eor));
2106 node_t *r = get_irn_node(get_Eor_right(eor));
2107 lattice_elem_t a = l->type;
2108 lattice_elem_t b = r->type;
2111 if (a.tv == tarval_top || b.tv == tarval_top) {
2112 node->type.tv = tarval_top;
2113 } else if (is_con(a) && is_con(b)) {
2114 if (is_tarval(a.tv) && is_tarval(b.tv)) {
2115 node->type.tv = tarval_eor(a.tv, b.tv);
2116 } else if (is_tarval(a.tv) && tarval_is_null(a.tv)) {
2118 } else if (is_tarval(b.tv) && tarval_is_null(b.tv)) {
2121 node->type.tv = tarval_bottom;
2123 node->by_all_const = 1;
2124 } else if (r->part == l->part) {
2125 ir_mode *mode = get_irn_mode(eor);
2126 tv = get_mode_null(mode);
2128 /* if the node was ONCE evaluated by all constants, but now
2129 this breaks AND we get from the argument partitions a different
2130 result, switch to bottom.
2131 This happens because initially all nodes are in the same partition ... */
2132 if (node->by_all_const && node->type.tv != tv)
2136 node->type.tv = tarval_bottom;
2141 * (Re-)compute the type for Cmp.
2143 * @param node the node
2145 static void compute_Cmp(node_t *node) {
2146 ir_node *cmp = node->node;
2147 node_t *l = get_irn_node(get_Cmp_left(cmp));
2148 node_t *r = get_irn_node(get_Cmp_right(cmp));
2149 lattice_elem_t a = l->type;
2150 lattice_elem_t b = r->type;
2152 if (a.tv == tarval_top || b.tv == tarval_top) {
2153 node->type.tv = tarval_top;
2154 } else if (r->part == l->part) {
2155 /* both nodes congruent, we can probably do something */
2156 node->type.tv = tarval_b_true;
2157 } else if (is_con(a) && is_con(b)) {
2158 /* both nodes are constants, we can probably do something */
2159 node->type.tv = tarval_b_true;
2161 node->type.tv = tarval_bottom;
2166 * (Re-)compute the type for a Proj(Cmp).
2168 * @param node the node
2169 * @param cond the predecessor Cmp node
2171 static void compute_Proj_Cmp(node_t *node, ir_node *cmp) {
2172 ir_node *proj = node->node;
2173 node_t *l = get_irn_node(get_Cmp_left(cmp));
2174 node_t *r = get_irn_node(get_Cmp_right(cmp));
2175 lattice_elem_t a = l->type;
2176 lattice_elem_t b = r->type;
2177 pn_Cmp pnc = get_Proj_proj(proj);
2180 if (a.tv == tarval_top || b.tv == tarval_top) {
2181 node->type.tv = tarval_undefined;
2182 } else if (is_con(a) && is_con(b)) {
2183 default_compute(node);
2184 node->by_all_const = 1;
2185 } else if (r->part == l->part &&
2186 (!mode_is_float(get_irn_mode(l->node)) || pnc == pn_Cmp_Lt || pnc == pn_Cmp_Gt)) {
2188 * BEWARE: a == a is NOT always True for floating Point values, as
2189 * NaN != NaN is defined, so we must check this here.
2191 tv = pnc & pn_Cmp_Eq ? tarval_b_true: tarval_b_false;
2193 /* if the node was ONCE evaluated by all constants, but now
2194 this breaks AND we get from the argument partitions a different
2195 result, switch to bottom.
2196 This happens because initially all nodes are in the same partition ... */
2197 if (node->by_all_const && node->type.tv != tv)
2201 node->type.tv = tarval_bottom;
2203 } /* compute_Proj_Cmp */
2206 * (Re-)compute the type for a Proj(Cond).
2208 * @param node the node
2209 * @param cond the predecessor Cond node
2211 static void compute_Proj_Cond(node_t *node, ir_node *cond) {
2212 ir_node *proj = node->node;
2213 long pnc = get_Proj_proj(proj);
2214 ir_node *sel = get_Cond_selector(cond);
2215 node_t *selector = get_irn_node(sel);
2218 * Note: it is crucial for the monotony that the Proj(Cond)
2219 * are evaluates after all predecessors of the Cond selector are
2225 * Due to the fact that 0 is a const, the Cmp gets immediately
2226 * on the cprop list. It will be evaluated before x is evaluated,
2227 * might leaving x as Top. When later x is evaluated, the Cmp
2228 * might change its value.
2229 * BUT if the Cond is evaluated before this happens, Proj(Cond, FALSE)
2230 * gets R, and later changed to F if Cmp is evaluated to True!
2232 * We prevent this by putting Conds in an extra cprop_X queue, which
2233 * gets evaluated after the cprop queue is empty.
2235 * Note that this even happens with Click's original algorithm, if
2236 * Cmp(x, 0) is evaluated to True first and later changed to False
2237 * if x was Top first and later changed to a Const ...
2238 * It is unclear how Click solved that problem ...
2240 * However, in rare cases even this does not help, if a Top reaches
2241 * a compare through a Phi, than Proj(Cond) is evaluated changing
2242 * the type of the Phi to something other.
2243 * So, we take the last resort and bind the type to R once
2246 * (This might be even the way Click works around the whole problem).
2248 * Finally, we may miss some optimization possibilities due to this:
2253 * If Top reaches the if first, than we decide for != here.
2254 * If y later is evaluated to 0, we cannot revert this decision
2255 * and must live with both outputs enabled. If this happens,
2256 * we get an unresolved if (true) in the code ...
2258 * In Click's version where this decision is done at the Cmp,
2259 * the Cmp is NOT optimized away than (if y evaluated to 1
2260 * for instance) and we get a if (1 == 0) here ...
2262 * Both solutions are suboptimal.
2263 * At least, we could easily detect this problem and run
2264 * cf_opt() (or even combo) again :-(
2266 if (node->type.tv == tarval_reachable)
2269 if (get_irn_mode(sel) == mode_b) {
2271 if (pnc == pn_Cond_true) {
2272 if (selector->type.tv == tarval_b_false) {
2273 node->type.tv = tarval_unreachable;
2274 } else if (selector->type.tv == tarval_b_true) {
2275 node->type.tv = tarval_reachable;
2276 } else if (selector->type.tv == tarval_bottom) {
2277 node->type.tv = tarval_reachable;
2279 assert(selector->type.tv == tarval_top);
2281 /* any condition based on Top is "!=" */
2282 node->type.tv = tarval_unreachable;
2284 node->type.tv = tarval_unreachable;
2288 assert(pnc == pn_Cond_false);
2290 if (selector->type.tv == tarval_b_false) {
2291 node->type.tv = tarval_reachable;
2292 } else if (selector->type.tv == tarval_b_true) {
2293 node->type.tv = tarval_unreachable;
2294 } else if (selector->type.tv == tarval_bottom) {
2295 node->type.tv = tarval_reachable;
2297 assert(selector->type.tv == tarval_top);
2299 /* any condition based on Top is "!=" */
2300 node->type.tv = tarval_reachable;
2302 node->type.tv = tarval_unreachable;
2308 if (selector->type.tv == tarval_bottom) {
2309 node->type.tv = tarval_reachable;
2310 } else if (selector->type.tv == tarval_top) {
2312 if (pnc == get_Cond_defaultProj(cond)) {
2313 /* a switch based of Top is always "default" */
2314 node->type.tv = tarval_reachable;
2317 node->type.tv = tarval_unreachable;
2319 long value = get_tarval_long(selector->type.tv);
2320 if (pnc == get_Cond_defaultProj(cond)) {
2321 /* default switch, have to check ALL other cases */
2324 for (i = get_irn_n_outs(cond) - 1; i >= 0; --i) {
2325 ir_node *succ = get_irn_out(cond, i);
2329 if (value == get_Proj_proj(succ)) {
2330 /* we found a match, will NOT take the default case */
2331 node->type.tv = tarval_unreachable;
2335 /* all cases checked, no match, will take default case */
2336 node->type.tv = tarval_reachable;
2339 node->type.tv = value == pnc ? tarval_reachable : tarval_unreachable;
2343 } /* compute_Proj_Cond */
2346 * (Re-)compute the type for a Proj-Node.
2348 * @param node the node
2350 static void compute_Proj(node_t *node) {
2351 ir_node *proj = node->node;
2352 ir_mode *mode = get_irn_mode(proj);
2353 node_t *block = get_irn_node(get_nodes_block(skip_Proj(proj)));
2354 ir_node *pred = get_Proj_pred(proj);
2356 if (block->type.tv == tarval_unreachable) {
2357 /* a Proj in a unreachable Block stay Top */
2358 node->type.tv = tarval_top;
2361 if (get_irn_node(pred)->type.tv == tarval_top && !is_Cond(pred)) {
2362 /* if the predecessor is Top, its Proj follow */
2363 node->type.tv = tarval_top;
2367 if (mode == mode_M) {
2368 /* mode M is always bottom */
2369 node->type.tv = tarval_bottom;
2372 if (mode != mode_X) {
2374 compute_Proj_Cmp(node, pred);
2376 default_compute(node);
2379 /* handle mode_X nodes */
2381 switch (get_irn_opcode(pred)) {
2383 /* the Proj_X from the Start is always reachable.
2384 However this is already handled at the top. */
2385 node->type.tv = tarval_reachable;
2388 compute_Proj_Cond(node, pred);
2391 default_compute(node);
2393 } /* compute_Proj */
2396 * (Re-)compute the type for a Confirm.
2398 * @param node the node
2400 static void compute_Confirm(node_t *node) {
2401 ir_node *confirm = node->node;
2402 node_t *pred = get_irn_node(get_Confirm_value(confirm));
2404 if (get_Confirm_cmp(confirm) == pn_Cmp_Eq) {
2405 node_t *bound = get_irn_node(get_Confirm_bound(confirm));
2407 if (is_con(bound->type)) {
2408 /* is equal to a constant */
2409 node->type = bound->type;
2413 /* a Confirm is a copy OR a Const */
2414 node->type = pred->type;
2415 } /* compute_Confirm */
2418 * (Re-)compute the type for a Max.
2420 * @param node the node
2422 static void compute_Max(node_t *node) {
2423 ir_node *op = node->node;
2424 node_t *l = get_irn_node(get_binop_left(op));
2425 node_t *r = get_irn_node(get_binop_right(op));
2426 lattice_elem_t a = l->type;
2427 lattice_elem_t b = r->type;
2429 if (a.tv == tarval_top || b.tv == tarval_top) {
2430 node->type.tv = tarval_top;
2431 } else if (is_con(a) && is_con(b)) {
2432 /* both nodes are constants, we can probably do something */
2434 /* this case handles SymConsts as well */
2437 ir_mode *mode = get_irn_mode(op);
2438 tarval *tv_min = get_mode_min(mode);
2442 else if (b.tv == tv_min)
2444 else if (is_tarval(a.tv) && is_tarval(b.tv)) {
2445 if (tarval_cmp(a.tv, b.tv) & pn_Cmp_Gt)
2446 node->type.tv = a.tv;
2448 node->type.tv = b.tv;
2450 node->type.tv = tarval_bad;
2453 } else if (r->part == l->part) {
2454 /* both nodes congruent, we can probably do something */
2457 node->type.tv = tarval_bottom;
2462 * (Re-)compute the type for a Min.
2464 * @param node the node
2466 static void compute_Min(node_t *node) {
2467 ir_node *op = node->node;
2468 node_t *l = get_irn_node(get_binop_left(op));
2469 node_t *r = get_irn_node(get_binop_right(op));
2470 lattice_elem_t a = l->type;
2471 lattice_elem_t b = r->type;
2473 if (a.tv == tarval_top || b.tv == tarval_top) {
2474 node->type.tv = tarval_top;
2475 } else if (is_con(a) && is_con(b)) {
2476 /* both nodes are constants, we can probably do something */
2478 /* this case handles SymConsts as well */
2481 ir_mode *mode = get_irn_mode(op);
2482 tarval *tv_max = get_mode_max(mode);
2486 else if (b.tv == tv_max)
2488 else if (is_tarval(a.tv) && is_tarval(b.tv)) {
2489 if (tarval_cmp(a.tv, b.tv) & pn_Cmp_Gt)
2490 node->type.tv = a.tv;
2492 node->type.tv = b.tv;
2494 node->type.tv = tarval_bad;
2497 } else if (r->part == l->part) {
2498 /* both nodes congruent, we can probably do something */
2501 node->type.tv = tarval_bottom;
2506 * (Re-)compute the type for a given node.
2508 * @param node the node
2510 static void compute(node_t *node) {
2511 ir_node *irn = node->node;
2514 #ifndef VERIFY_MONOTONE
2516 * Once a node reaches bottom, the type cannot fall further
2517 * in the lattice and we can stop computation.
2518 * Do not take this exit if the monotony verifier is
2519 * enabled to catch errors.
2521 if (node->type.tv == tarval_bottom)
2524 if (is_no_Block(irn)) {
2525 /* for pinned nodes, check its control input */
2526 if (get_irn_pinned(skip_Proj(irn)) == op_pin_state_pinned) {
2527 node_t *block = get_irn_node(get_nodes_block(irn));
2529 if (block->type.tv == tarval_unreachable) {
2530 node->type.tv = tarval_top;
2536 func = (compute_func)node->node->op->ops.generic;
2542 * Identity functions: Note that one might thing that identity() is just a
2543 * synonym for equivalent_node(). While this is true, we cannot use it for the algorithm
2544 * here, because it expects that the identity node is one of the inputs, which is NOT
2545 * always true for equivalent_node() which can handle (and does sometimes) DAGs.
2546 * So, we have our own implementation, which copies some parts of equivalent_node()
2550 * Calculates the Identity for Phi nodes
2552 static node_t *identity_Phi(node_t *node) {
2553 ir_node *phi = node->node;
2554 ir_node *block = get_nodes_block(phi);
2555 node_t *n_part = NULL;
2558 for (i = get_Phi_n_preds(phi) - 1; i >= 0; --i) {
2559 node_t *pred_X = get_irn_node(get_Block_cfgpred(block, i));
2561 if (pred_X->type.tv == tarval_reachable) {
2562 node_t *pred = get_irn_node(get_Phi_pred(phi, i));
2566 else if (n_part->part != pred->part) {
2567 /* incongruent inputs, not a follower */
2572 /* if n_part is NULL here, all inputs path are dead, the Phi computes
2573 * tarval_top, is in the TOP partition and should NOT being split! */
2574 assert(n_part != NULL);
2576 } /* identity_Phi */
2579 * Calculates the Identity for commutative 0 neutral nodes.
2581 static node_t *identity_comm_zero_binop(node_t *node) {
2582 ir_node *op = node->node;
2583 node_t *a = get_irn_node(get_binop_left(op));
2584 node_t *b = get_irn_node(get_binop_right(op));
2585 ir_mode *mode = get_irn_mode(op);
2588 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
2589 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
2592 /* node: no input should be tarval_top, else the binop would be also
2593 * Top and not being split. */
2594 zero = get_mode_null(mode);
2595 if (a->type.tv == zero)
2597 if (b->type.tv == zero)
2600 } /* identity_comm_zero_binop */
2603 * Calculates the Identity for Shift nodes.
2605 static node_t *identity_shift(node_t *node) {
2606 ir_node *op = node->node;
2607 node_t *b = get_irn_node(get_binop_right(op));
2608 ir_mode *mode = get_irn_mode(b->node);
2611 /* node: no input should be tarval_top, else the binop would be also
2612 * Top and not being split. */
2613 zero = get_mode_null(mode);
2614 if (b->type.tv == zero)
2615 return get_irn_node(get_binop_left(op));
2617 } /* identity_shift */
2620 * Calculates the Identity for Mul nodes.
2622 static node_t *identity_Mul(node_t *node) {
2623 ir_node *op = node->node;
2624 node_t *a = get_irn_node(get_Mul_left(op));
2625 node_t *b = get_irn_node(get_Mul_right(op));
2626 ir_mode *mode = get_irn_mode(op);
2629 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
2630 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
2633 /* node: no input should be tarval_top, else the binop would be also
2634 * Top and not being split. */
2635 one = get_mode_one(mode);
2636 if (a->type.tv == one)
2638 if (b->type.tv == one)
2641 } /* identity_Mul */
2644 * Calculates the Identity for Sub nodes.
2646 static node_t *identity_Sub(node_t *node) {
2647 ir_node *sub = node->node;
2648 node_t *b = get_irn_node(get_Sub_right(sub));
2649 ir_mode *mode = get_irn_mode(sub);
2651 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
2652 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
2655 /* node: no input should be tarval_top, else the binop would be also
2656 * Top and not being split. */
2657 if (b->type.tv == get_mode_null(mode))
2658 return get_irn_node(get_Sub_left(sub));
2660 } /* identity_Sub */
2663 * Calculates the Identity for And nodes.
2665 static node_t *identity_And(node_t *node) {
2666 ir_node *and = node->node;
2667 node_t *a = get_irn_node(get_And_left(and));
2668 node_t *b = get_irn_node(get_And_right(and));
2669 tarval *neutral = get_mode_all_one(get_irn_mode(and));
2671 /* node: no input should be tarval_top, else the And would be also
2672 * Top and not being split. */
2673 if (a->type.tv == neutral)
2675 if (b->type.tv == neutral)
2678 } /* identity_And */
2681 * Calculates the Identity for Confirm nodes.
2683 static node_t *identity_Confirm(node_t *node) {
2684 ir_node *confirm = node->node;
2686 /* a Confirm is always a Copy */
2687 return get_irn_node(get_Confirm_value(confirm));
2688 } /* identity_Confirm */
2691 * Calculates the Identity for Mux nodes.
2693 static node_t *identity_Mux(node_t *node) {
2694 ir_node *mux = node->node;
2695 node_t *t = get_irn_node(get_Mux_true(mux));
2696 node_t *f = get_irn_node(get_Mux_false(mux));
2699 if (t->part == f->part)
2702 /* for now, the 1-input identity is not supported */
2704 sel = get_irn_node(get_Mux_sel(mux));
2706 /* Mux sel input is mode_b, so it is always a tarval */
2707 if (sel->type.tv == tarval_b_true)
2709 if (sel->type.tv == tarval_b_false)
2713 } /* identity_Mux */
2716 * Calculates the Identity for Min nodes.
2718 static node_t *identity_Min(node_t *node) {
2719 ir_node *op = node->node;
2720 node_t *a = get_irn_node(get_binop_left(op));
2721 node_t *b = get_irn_node(get_binop_right(op));
2722 ir_mode *mode = get_irn_mode(op);
2725 if (a->part == b->part) {
2726 /* leader of multiple predecessors */
2730 /* works even with NaN */
2731 tv_max = get_mode_max(mode);
2732 if (a->type.tv == tv_max)
2734 if (b->type.tv == tv_max)
2737 } /* identity_Min */
2740 * Calculates the Identity for Max nodes.
2742 static node_t *identity_Max(node_t *node) {
2743 ir_node *op = node->node;
2744 node_t *a = get_irn_node(get_binop_left(op));
2745 node_t *b = get_irn_node(get_binop_right(op));
2746 ir_mode *mode = get_irn_mode(op);
2749 if (a->part == b->part) {
2750 /* leader of multiple predecessors */
2754 /* works even with NaN */
2755 tv_min = get_mode_min(mode);
2756 if (a->type.tv == tv_min)
2758 if (b->type.tv == tv_min)
2761 } /* identity_Max */
2764 * Calculates the Identity for nodes.
2766 static node_t *identity(node_t *node) {
2767 ir_node *irn = node->node;
2769 switch (get_irn_opcode(irn)) {
2771 return identity_Phi(node);
2773 return identity_Mul(node);
2777 return identity_comm_zero_binop(node);
2782 return identity_shift(node);
2784 return identity_And(node);
2786 return identity_Sub(node);
2788 return identity_Confirm(node);
2790 return identity_Mux(node);
2792 return identity_Min(node);
2794 return identity_Max(node);
2801 * Node follower is a (new) follower of leader, segregate Leader
2804 static void segregate_def_use_chain_1(const ir_node *follower, node_t *leader) {
2805 ir_node *l = leader->node;
2806 int j, i, n = get_irn_n_outs(l);
2808 DB((dbg, LEVEL_2, "%+F is a follower of %+F\n", follower, leader->node));
2809 /* The leader edges must remain sorted, but follower edges can
2811 for (i = leader->n_followers + 1; i <= n; ++i) {
2812 if (l->out[i].use == follower) {
2813 ir_def_use_edge t = l->out[i];
2815 for (j = i - 1; j >= leader->n_followers + 1; --j)
2816 l->out[j + 1] = l->out[j];
2817 ++leader->n_followers;
2818 l->out[leader->n_followers] = t;
2822 } /* segregate_def_use_chain_1 */
2825 * Node follower is a (new) follower segregate its Leader
2828 * @param follower the follower IR node
2830 static void segregate_def_use_chain(const ir_node *follower) {
2833 for (i = get_irn_arity(follower) - 1; i >= 0; --i) {
2834 node_t *pred = get_irn_node(get_irn_n(follower, i));
2836 segregate_def_use_chain_1(follower, pred);
2838 } /* segregate_def_use_chain */
2841 * Propagate constant evaluation.
2843 * @param env the environment
2845 static void propagate(environment_t *env) {
2848 lattice_elem_t old_type;
2850 unsigned n_fallen, old_type_was_T_or_C;
2853 while (env->cprop != NULL) {
2854 void *oldopcode = NULL;
2856 /* remove the first partition X from cprop */
2859 env->cprop = X->cprop_next;
2861 old_type_was_T_or_C = X->type_is_T_or_C;
2863 DB((dbg, LEVEL_2, "Propagate type on part%d\n", X->nr));
2867 int cprop_empty = list_empty(&X->cprop);
2868 int cprop_X_empty = list_empty(&X->cprop_X);
2870 if (cprop_empty && cprop_X_empty) {
2871 /* both cprop lists are empty */
2875 /* remove the first Node x from X.cprop */
2877 /* Get a node from the cprop_X list only if
2878 * all data nodes are processed.
2879 * This ensures, that all inputs of the Cond
2880 * predecessor are processed if its type is still Top.
2882 x = list_entry(X->cprop_X.next, node_t, cprop_list);
2884 x = list_entry(X->cprop.next, node_t, cprop_list);
2887 //assert(x->part == X);
2888 list_del(&x->cprop_list);
2891 if (x->is_follower && identity(x) == x) {
2892 /* check the opcode first */
2893 if (oldopcode == NULL) {
2894 oldopcode = lambda_opcode(get_first_node(X), env);
2896 if (oldopcode != lambda_opcode(x, env)) {
2897 if (x->on_fallen == 0) {
2898 /* different opcode -> x falls out of this partition */
2903 DB((dbg, LEVEL_2, "Add node %+F to fallen\n", x->node));
2907 /* x will make the follower -> leader transition */
2908 follower_to_leader(x);
2911 /* compute a new type for x */
2913 DB((dbg, LEVEL_3, "computing type of %+F\n", x->node));
2915 if (x->type.tv != old_type.tv) {
2916 DB((dbg, LEVEL_2, "node %+F has changed type from %+F to %+F\n", x->node, old_type, x->type));
2917 verify_type(old_type, x);
2919 if (x->on_fallen == 0) {
2920 /* Add x to fallen. Nodes might fall from T -> const -> _|_, so check that they are
2921 not already on the list. */
2926 DB((dbg, LEVEL_2, "Add node %+F to fallen\n", x->node));
2928 for (i = get_irn_n_outs(x->node) - 1; i >= 0; --i) {
2929 ir_node *succ = get_irn_out(x->node, i);
2930 node_t *y = get_irn_node(succ);
2932 /* Add y to y.partition.cprop. */
2933 add_to_cprop(y, env);
2938 if (n_fallen > 0 && n_fallen != X->n_leader) {
2939 DB((dbg, LEVEL_2, "Splitting part%d by fallen\n", X->nr));
2940 Y = split(&X, fallen, env);
2942 * We have split out fallen node. The type of the result
2943 * partition is NOT set yet.
2945 Y->type_is_T_or_C = 0;
2949 /* remove the flags from the fallen list */
2950 for (x = fallen; x != NULL; x = x->next)
2953 if (old_type_was_T_or_C) {
2956 /* check if some nodes will make the leader -> follower transition */
2957 list_for_each_entry_safe(node_t, y, tmp, &Y->Leader, node_list) {
2958 if (y->type.tv != tarval_top && ! is_con(y->type)) {
2959 node_t *eq_node = identity(y);
2961 if (eq_node != y && eq_node->part == y->part) {
2962 DB((dbg, LEVEL_2, "Node %+F is a follower of %+F\n", y->node, eq_node->node));
2963 /* move to Follower */
2965 list_del(&y->node_list);
2966 list_add_tail(&y->node_list, &Y->Follower);
2969 segregate_def_use_chain(y->node);
2979 * Get the leader for a given node from its congruence class.
2981 * @param irn the node
2983 static ir_node *get_leader(node_t *node) {
2984 partition_t *part = node->part;
2986 if (part->n_leader > 1 || node->is_follower) {
2987 if (node->is_follower) {
2988 DB((dbg, LEVEL_2, "Replacing follower %+F\n", node->node));
2991 DB((dbg, LEVEL_2, "Found congruence class for %+F\n", node->node));
2993 return get_first_node(part)->node;
2999 * Returns non-zero if a mode_T node has only one reachable output.
3001 static int only_one_reachable_proj(ir_node *n) {
3004 for (i = get_irn_n_outs(n) - 1; i >= 0; --i) {
3005 ir_node *proj = get_irn_out(n, i);
3008 /* skip non-control flow Proj's */
3009 if (get_irn_mode(proj) != mode_X)
3012 node = get_irn_node(proj);
3013 if (node->type.tv == tarval_reachable) {
3019 } /* only_one_reachable_proj */
3022 * Return non-zero if the control flow predecessor node pred
3023 * is the only reachable control flow exit of its block.
3025 * @param pred the control flow exit
3027 static int can_exchange(ir_node *pred) {
3030 else if (is_Jmp(pred))
3032 else if (get_irn_mode(pred) == mode_T) {
3033 /* if the predecessor block has more than one
3034 reachable outputs we cannot remove the block */
3035 return only_one_reachable_proj(pred);
3038 } /* can_exchange */
3041 * Block Post-Walker, apply the analysis results on control flow by
3042 * shortening Phi's and Block inputs.
3044 static void apply_cf(ir_node *block, void *ctx) {
3045 environment_t *env = ctx;
3046 node_t *node = get_irn_node(block);
3048 ir_node **ins, **in_X;
3049 ir_node *phi, *next;
3051 n = get_Block_n_cfgpreds(block);
3053 if (node->type.tv == tarval_unreachable) {
3056 for (i = n - 1; i >= 0; --i) {
3057 ir_node *pred = get_Block_cfgpred(block, i);
3059 if (! is_Bad(pred)) {
3060 node_t *pred_bl = get_irn_node(get_nodes_block(skip_Proj(pred)));
3062 if (pred_bl->flagged == 0) {
3063 pred_bl->flagged = 3;
3065 if (pred_bl->type.tv == tarval_reachable) {
3067 * We will remove an edge from block to its pred.
3068 * This might leave the pred block as an endless loop
3070 if (! is_backedge(block, i))
3071 keep_alive(pred_bl->node);
3077 /* the EndBlock is always reachable even if the analysis
3078 finds out the opposite :-) */
3079 if (block != get_irg_end_block(current_ir_graph)) {
3080 /* mark dead blocks */
3081 set_Block_dead(block);
3082 DB((dbg, LEVEL_1, "Removing dead %+F\n", block));
3084 /* the endblock is unreachable */
3085 set_irn_in(block, 0, NULL);
3091 /* only one predecessor combine */
3092 ir_node *pred = skip_Proj(get_Block_cfgpred(block, 0));
3094 if (can_exchange(pred)) {
3095 ir_node *new_block = get_nodes_block(pred);
3096 DB((dbg, LEVEL_1, "Fuse %+F with %+F\n", block, new_block));
3097 DBG_OPT_COMBO(block, new_block, FS_OPT_COMBO_CF);
3098 exchange(block, new_block);
3099 node->node = new_block;
3105 NEW_ARR_A(ir_node *, in_X, n);
3107 for (i = 0; i < n; ++i) {
3108 ir_node *pred = get_Block_cfgpred(block, i);
3109 node_t *node = get_irn_node(pred);
3111 if (node->type.tv == tarval_reachable) {
3114 DB((dbg, LEVEL_1, "Removing dead input %d from %+F (%+F)\n", i, block, pred));
3115 if (! is_Bad(pred)) {
3116 node_t *pred_bl = get_irn_node(get_nodes_block(skip_Proj(pred)));
3118 if (pred_bl->flagged == 0) {
3119 pred_bl->flagged = 3;
3121 if (pred_bl->type.tv == tarval_reachable) {
3123 * We will remove an edge from block to its pred.
3124 * This might leave the pred block as an endless loop
3126 if (! is_backedge(block, i))
3127 keep_alive(pred_bl->node);
3137 NEW_ARR_A(ir_node *, ins, n);
3138 for (phi = get_Block_phis(block); phi != NULL; phi = next) {
3139 node_t *node = get_irn_node(phi);
3141 next = get_Phi_next(phi);
3142 if (is_tarval(node->type.tv) && tarval_is_constant(node->type.tv)) {
3143 /* this Phi is replaced by a constant */
3144 tarval *tv = node->type.tv;
3145 ir_node *c = new_r_Const(current_ir_graph, block, get_tarval_mode(tv), tv);
3147 set_irn_node(c, node);
3149 DB((dbg, LEVEL_1, "%+F is replaced by %+F\n", phi, c));
3150 DBG_OPT_COMBO(phi, c, FS_OPT_COMBO_CONST);
3155 for (i = 0; i < n; ++i) {
3156 node_t *pred = get_irn_node(get_Block_cfgpred(block, i));
3158 if (pred->type.tv == tarval_reachable) {
3159 ins[j++] = get_Phi_pred(phi, i);
3163 /* this Phi is replaced by a single predecessor */
3164 ir_node *s = ins[0];
3165 node_t *phi_node = get_irn_node(phi);
3168 DB((dbg, LEVEL_1, "%+F is replaced by %+F because of cf change\n", phi, s));
3169 DBG_OPT_COMBO(phi, s, FS_OPT_COMBO_FOLLOWER);
3174 set_irn_in(phi, j, ins);
3182 /* this Block has only one live predecessor */
3183 ir_node *pred = skip_Proj(in_X[0]);
3185 if (can_exchange(pred)) {
3186 ir_node *new_block = get_nodes_block(pred);
3187 DBG_OPT_COMBO(block, new_block, FS_OPT_COMBO_CF);
3188 exchange(block, new_block);
3189 node->node = new_block;
3194 set_irn_in(block, k, in_X);
3199 * Exchange a node by its leader.
3200 * Beware: in rare cases the mode might be wrong here, for instance
3201 * AddP(x, NULL) is a follower of x, but with different mode.
3204 static void exchange_leader(ir_node *irn, ir_node *leader) {
3205 ir_mode *mode = get_irn_mode(irn);
3206 if (mode != get_irn_mode(leader)) {
3207 /* The conv is a no-op, so we are fre to place in
3208 * either in the block of the leader OR in irn's block.
3209 * Probably placing it into leaders block might reduce
3210 * the number of Conv due to CSE. */
3211 ir_node *block = get_nodes_block(leader);
3212 dbg_info *dbg = get_irn_dbg_info(irn);
3214 leader = new_rd_Conv(dbg, current_ir_graph, block, leader, mode);
3216 exchange(irn, leader);
3220 * Post-Walker, apply the analysis results;
3222 static void apply_result(ir_node *irn, void *ctx) {
3223 environment_t *env = ctx;
3224 node_t *node = get_irn_node(irn);
3226 if (is_Block(irn) || is_End(irn) || is_Bad(irn)) {
3227 /* blocks already handled, do not touch the End node */
3229 node_t *block = get_irn_node(get_nodes_block(irn));
3231 if (block->type.tv == tarval_unreachable) {
3232 ir_node *bad = get_irg_bad(current_ir_graph);
3234 /* here, bad might already have a node, but this can be safely ignored
3235 as long as bad has at least ONE valid node */
3236 set_irn_node(bad, node);
3238 DB((dbg, LEVEL_1, "%+F is unreachable\n", irn));
3242 else if (node->type.tv == tarval_top) {
3243 /* don't kick away Unknown's, they might be still needed */
3244 if (! is_Unknown(irn)) {
3245 ir_mode *mode = get_irn_mode(irn);
3246 ir_node *unk = new_r_Unknown(current_ir_graph, mode);
3248 /* control flow should already be handled at apply_cf() */
3249 assert(mode != mode_X);
3251 /* see comment above */
3252 set_irn_node(unk, node);
3254 DB((dbg, LEVEL_1, "%+F computes Top\n", irn));
3259 else if (get_irn_mode(irn) == mode_X) {
3262 ir_node *cond = get_Proj_pred(irn);
3264 if (is_Cond(cond)) {
3265 if (only_one_reachable_proj(cond)) {
3266 ir_node *jmp = new_r_Jmp(current_ir_graph, block->node);
3267 set_irn_node(jmp, node);
3269 DB((dbg, LEVEL_1, "%+F is replaced by %+F\n", irn, jmp));
3270 DBG_OPT_COMBO(irn, jmp, FS_OPT_COMBO_CF);
3274 node_t *sel = get_irn_node(get_Cond_selector(cond));
3275 tarval *tv = sel->type.tv;
3277 if (is_tarval(tv) && tarval_is_constant(tv)) {
3278 /* The selector is a constant, but more
3279 * than one output is active: An unoptimized
3287 /* normal data node */
3288 if (is_tarval(node->type.tv) && tarval_is_constant(node->type.tv)) {
3289 tarval *tv = node->type.tv;
3292 * Beware: never replace mode_T nodes by constants. Currently we must mark
3293 * mode_T nodes with constants, but do NOT replace them.
3295 if (! is_Const(irn) && get_irn_mode(irn) != mode_T) {
3296 /* can be replaced by a constant */
3297 ir_node *c = new_r_Const(current_ir_graph, block->node, get_tarval_mode(tv), tv);
3298 set_irn_node(c, node);
3300 DB((dbg, LEVEL_1, "%+F is replaced by %+F\n", irn, c));
3301 DBG_OPT_COMBO(irn, c, FS_OPT_COMBO_CONST);
3302 exchange_leader(irn, c);
3305 } else if (is_entity(node->type.sym.entity_p)) {
3306 if (! is_SymConst(irn)) {
3307 /* can be replaced by a SymConst */
3308 ir_node *symc = new_r_SymConst(current_ir_graph, block->node, get_irn_mode(irn), node->type.sym, symconst_addr_ent);
3309 set_irn_node(symc, node);
3312 DB((dbg, LEVEL_1, "%+F is replaced by %+F\n", irn, symc));
3313 DBG_OPT_COMBO(irn, symc, FS_OPT_COMBO_CONST);
3314 exchange_leader(irn, symc);
3317 } else if (is_Confirm(irn)) {
3318 /* Confirms are always follower, but do not kill them here */
3320 ir_node *leader = get_leader(node);
3322 if (leader != irn) {
3323 int non_strict_phi = 0;
3326 * Beware: Do not remove Phi(Unknown, ..., x, ..., Unknown)
3327 * as this might create non-strict programs.
3329 if (node->is_follower && is_Phi(irn) && !is_Unknown(leader)) {
3332 for (i = get_Phi_n_preds(irn) - 1; i >= 0; --i) {
3333 ir_node *pred = get_Phi_pred(irn, i);
3335 if (is_Unknown(pred)) {
3341 if (! non_strict_phi) {
3342 DB((dbg, LEVEL_1, "%+F from part%d is replaced by %+F\n", irn, node->part->nr, leader));
3343 if (node->is_follower)
3344 DBG_OPT_COMBO(irn, leader, FS_OPT_COMBO_FOLLOWER);
3346 DBG_OPT_COMBO(irn, leader, FS_OPT_COMBO_CONGRUENT);
3347 exchange_leader(irn, leader);
3354 } /* apply_result */
3357 * Fix the keep-alives by deleting unreachable ones.
3359 static void apply_end(ir_node *end, environment_t *env) {
3360 int i, j, n = get_End_n_keepalives(end);
3364 NEW_ARR_A(ir_node *, in, n);
3366 /* fix the keep alive */
3367 for (i = j = 0; i < n; i++) {
3368 ir_node *ka = get_End_keepalive(end, i);
3369 node_t *node = get_irn_node(ka);
3372 node = get_irn_node(get_nodes_block(ka));
3374 if (node->type.tv != tarval_unreachable && !is_Bad(ka))
3378 set_End_keepalives(end, j, in);
3383 #define SET(code) op_##code->ops.generic = (op_func)compute_##code
3386 * sets the generic functions to compute.
3388 static void set_compute_functions(void) {
3391 /* set the default compute function */
3392 for (i = get_irp_n_opcodes() - 1; i >= 0; --i) {
3393 ir_op *op = get_irp_opcode(i);
3394 op->ops.generic = (op_func)default_compute;
3397 /* set specific functions */
3418 } /* set_compute_functions */
3420 void combo(ir_graph *irg) {
3422 ir_node *initial_bl;
3424 ir_graph *rem = current_ir_graph;
3426 current_ir_graph = irg;
3428 /* register a debug mask */
3429 FIRM_DBG_REGISTER(dbg, "firm.opt.combo");
3431 DB((dbg, LEVEL_1, "Doing COMBO for %+F\n", irg));
3433 obstack_init(&env.obst);
3434 env.worklist = NULL;
3438 #ifdef DEBUG_libfirm
3439 env.dbg_list = NULL;
3441 env.opcode2id_map = new_set(cmp_opcode, iro_Last * 4);
3442 env.type2id_map = pmap_create();
3443 env.end_idx = get_opt_global_cse() ? 0 : -1;
3444 env.lambda_input = 0;
3445 env.commutative = 1;
3449 assure_irg_outs(irg);
3450 assure_cf_loop(irg);
3452 /* we have our own value_of function */
3453 set_value_of_func(get_node_tarval);
3455 set_compute_functions();
3456 DEBUG_ONLY(part_nr = 0);
3458 ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK);
3460 /* create the initial partition and place it on the work list */
3461 env.initial = new_partition(&env);
3462 add_to_worklist(env.initial, &env);
3463 irg_walk_graph(irg, create_initial_partitions, init_block_phis, &env);
3465 /* set the hook: from now, every node has a partition and a type */
3466 DEBUG_ONLY(set_dump_node_vcgattr_hook(dump_partition_hook));
3468 /* all nodes on the initial partition have type Top */
3469 env.initial->type_is_T_or_C = 1;
3471 /* Place the START Node's partition on cprop.
3472 Place the START Node on its local worklist. */
3473 initial_bl = get_irg_start_block(irg);
3474 start = get_irn_node(initial_bl);
3475 add_to_cprop(start, &env);
3479 if (env.worklist != NULL)
3481 } while (env.cprop != NULL || env.worklist != NULL);
3483 dump_all_partitions(&env);
3484 check_all_partitions(&env);
3487 dump_ir_block_graph(irg, "-partition");
3490 /* apply the result */
3491 irg_block_walk_graph(irg, NULL, apply_cf, &env);
3492 /* Kill keep-alives of dead blocks: this speeds up apply_result()
3493 * and fixes assertion because dead cf to dead blocks is NOT removed by
3495 apply_end(get_irg_end(irg), &env);
3496 irg_walk_graph(irg, NULL, apply_result, &env);
3499 DB((dbg, LEVEL_1, "Unoptimized Control Flow left"));
3503 /* control flow might changed */
3504 set_irg_outs_inconsistent(irg);
3505 set_irg_extblk_inconsistent(irg);
3506 set_irg_doms_inconsistent(irg);
3507 set_irg_loopinfo_inconsistent(irg);
3510 ir_free_resources(irg, IR_RESOURCE_IRN_LINK);
3512 /* remove the partition hook */
3513 DEBUG_ONLY(set_dump_node_vcgattr_hook(NULL));
3515 pmap_destroy(env.type2id_map);
3516 del_set(env.opcode2id_map);
3517 obstack_free(&env.obst, NULL);
3519 /* restore value_of() default behavior */
3520 set_value_of_func(NULL);
3521 current_ir_graph = rem;