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 flagged:2; /**< 2 Bits, set if this node was visited by race 1 or 2. */
166 * A partition containing congruent nodes.
169 list_head Leader; /**< The head of partition Leader node list. */
170 list_head Follower; /**< The head of partition Follower node list. */
171 list_head cprop; /**< The head of partition.cprop list. */
172 list_head cprop_X; /**< The head of partition.cprop (Cond nodes and its Projs) list. */
173 partition_t *wl_next; /**< Next entry in the work list if any. */
174 partition_t *touched_next; /**< Points to the next partition in the touched set. */
175 partition_t *cprop_next; /**< Points to the next partition in the cprop list. */
176 partition_t *split_next; /**< Points to the next partition in the list that must be split by split_by(). */
177 node_t *touched; /**< The partition.touched set of this partition. */
178 unsigned n_leader; /**< Number of entries in this partition.Leader. */
179 unsigned n_touched; /**< Number of entries in the partition.touched. */
180 int max_user_inputs; /**< Maximum number of user inputs of all entries. */
181 unsigned on_worklist:1; /**< Set, if this partition is in the work list. */
182 unsigned on_touched:1; /**< Set, if this partition is on the touched set. */
183 unsigned on_cprop:1; /**< Set, if this partition is on the cprop list. */
184 unsigned type_is_T_or_C:1;/**< Set, if all nodes in this partition have type Top or Constant. */
186 partition_t *dbg_next; /**< Link all partitions for debugging */
187 unsigned nr; /**< A unique number for (what-)mapping, >0. */
191 typedef struct environment_t {
192 struct obstack obst; /**< obstack to allocate data structures. */
193 partition_t *worklist; /**< The work list. */
194 partition_t *cprop; /**< The constant propagation list. */
195 partition_t *touched; /**< the touched set. */
196 partition_t *initial; /**< The initial partition. */
197 set *opcode2id_map; /**< The opcodeMode->id map. */
198 pmap *type2id_map; /**< The type->id map. */
199 int end_idx; /**< -1 for local and 0 for global congruences. */
200 int lambda_input; /**< Captured argument for lambda_partition(). */
201 unsigned modified:1; /**< Set, if the graph was modified. */
202 unsigned commutative:1; /**< Set, if commutation nodes should be handled specially. */
203 unsigned unopt_cf:1; /**< If set, control flow is not optimized due to Unknown. */
205 partition_t *dbg_list; /**< List of all partitions. */
209 /** Type of the what function. */
210 typedef void *(*what_func)(const node_t *node, environment_t *env);
212 #define get_irn_node(follower) ((node_t *)get_irn_link(follower))
213 #define set_irn_node(follower, node) set_irn_link(follower, node)
215 /* we do NOT use tarval_unreachable here, instead we use Top for this purpose */
216 #undef tarval_unreachable
217 #define tarval_unreachable tarval_top
220 /** The debug module handle. */
221 DEBUG_ONLY(static firm_dbg_module_t *dbg;)
223 /** The what reason. */
224 DEBUG_ONLY(static const char *what_reason;)
226 /** Next partition number. */
227 DEBUG_ONLY(static unsigned part_nr = 0);
229 /** The tarval returned by Unknown nodes. */
231 #define tarval_UNKNOWN tarval_top
233 #define tarval_UNKNOWN tarval_bad
237 static node_t *identity(node_t *node);
239 #ifdef CHECK_PARTITIONS
243 static void check_partition(const partition_t *T) {
247 list_for_each_entry(node_t, node, &T->Leader, node_list) {
248 assert(node->is_follower == 0);
249 assert(node->flagged == 0);
250 assert(node->part == T);
253 assert(n == T->n_leader);
255 list_for_each_entry(node_t, node, &T->Follower, node_list) {
256 assert(node->is_follower == 1);
257 assert(node->flagged == 0);
258 assert(node->part == T);
260 } /* check_partition */
263 * check that all leader nodes in the partition have the same opcode.
265 static void check_opcode(const partition_t *Z) {
270 list_for_each_entry(node_t, node, &Z->Leader, node_list) {
271 ir_node *irn = node->node;
274 key.code = get_irn_opcode(irn);
275 key.mode = get_irn_mode(irn);
276 key.arity = get_irn_arity(irn);
280 switch (get_irn_opcode(irn)) {
282 key.u.proj = get_Proj_proj(irn);
285 key.u.ent = get_Sel_entity(irn);
292 assert(key.code == get_irn_opcode(irn));
293 assert(key.mode == get_irn_mode(irn));
294 assert(key.arity == get_irn_arity(irn));
296 switch (get_irn_opcode(irn)) {
298 assert(key.u.proj == get_Proj_proj(irn));
301 assert(key.u.ent == get_Sel_entity(irn));
310 static void check_all_partitions(environment_t *env) {
315 for (P = env->dbg_list; P != NULL; P = P->dbg_next) {
317 if (! P->type_is_T_or_C)
319 list_for_each_entry(node_t, node, &P->Follower, node_list) {
320 node_t *leader = identity(node);
322 assert(leader != node && leader->part == node->part);
331 static void do_check_list(const node_t *list, int ofs, const partition_t *Z) {
334 #define NEXT(e) *((const node_t **)((char *)(e) + (ofs)))
335 for (e = list; e != NULL; e = NEXT(e)) {
336 assert(e->part == Z);
339 } /* ido_check_list */
342 * Check a local list.
344 static void check_list(const node_t *list, const partition_t *Z) {
345 do_check_list(list, offsetof(node_t, next), Z);
349 #define check_partition(T)
350 #define check_list(list, Z)
351 #define check_all_partitions(env)
352 #endif /* CHECK_PARTITIONS */
355 static inline lattice_elem_t get_partition_type(const partition_t *X);
358 * Dump partition to output.
360 static void dump_partition(const char *msg, const partition_t *part) {
363 lattice_elem_t type = get_partition_type(part);
365 DB((dbg, LEVEL_2, "%s part%u%s (%u, %+F) {\n ",
366 msg, part->nr, part->type_is_T_or_C ? "*" : "",
367 part->n_leader, type));
368 list_for_each_entry(node_t, node, &part->Leader, node_list) {
369 DB((dbg, LEVEL_2, "%s%+F", first ? "" : ", ", node->node));
372 if (! list_empty(&part->Follower)) {
373 DB((dbg, LEVEL_2, "\n---\n "));
375 list_for_each_entry(node_t, node, &part->Follower, node_list) {
376 DB((dbg, LEVEL_2, "%s%+F", first ? "" : ", ", node->node));
380 DB((dbg, LEVEL_2, "\n}\n"));
381 } /* dump_partition */
386 static void do_dump_list(const char *msg, const node_t *node, int ofs) {
390 #define GET_LINK(p, ofs) *((const node_t **)((char *)(p) + (ofs)))
392 DB((dbg, LEVEL_3, "%s = {\n ", msg));
393 for (p = node; p != NULL; p = GET_LINK(p, ofs)) {
394 DB((dbg, LEVEL_3, "%s%+F", first ? "" : ", ", p->node));
397 DB((dbg, LEVEL_3, "\n}\n"));
405 static void dump_race_list(const char *msg, const node_t *list) {
406 do_dump_list(msg, list, offsetof(node_t, race_next));
407 } /* dump_race_list */
410 * Dumps a local list.
412 static void dump_list(const char *msg, const node_t *list) {
413 do_dump_list(msg, list, offsetof(node_t, next));
417 * Dump all partitions.
419 static void dump_all_partitions(const environment_t *env) {
420 const partition_t *P;
422 DB((dbg, LEVEL_2, "All partitions\n===============\n"));
423 for (P = env->dbg_list; P != NULL; P = P->dbg_next)
424 dump_partition("", P);
425 } /* dump_all_partitions */
430 static void dump_split_list(const partition_t *list) {
431 const partition_t *p;
433 DB((dbg, LEVEL_2, "Split by %s produced = {\n", what_reason));
434 for (p = list; p != NULL; p = p->split_next)
435 DB((dbg, LEVEL_2, "part%u, ", p->nr));
436 DB((dbg, LEVEL_2, "\n}\n"));
437 } /* dump_split_list */
440 * Dump partition and type for a node.
442 static int dump_partition_hook(FILE *F, ir_node *n, ir_node *local) {
443 ir_node *irn = local != NULL ? local : n;
444 node_t *node = get_irn_node(irn);
446 ir_fprintf(F, "info2 : \"partition %u type %+F\"\n", node->part->nr, node->type);
448 } /* dump_partition_hook */
451 #define dump_partition(msg, part)
452 #define dump_race_list(msg, list)
453 #define dump_list(msg, list)
454 #define dump_all_partitions(env)
455 #define dump_split_list(list)
458 #if defined(VERIFY_MONOTONE) && defined (DEBUG_libfirm)
460 * Verify that a type transition is monotone
462 static void verify_type(const lattice_elem_t old_type, node_t *node) {
463 if (old_type.tv == node->type.tv) {
467 if (old_type.tv == tarval_top) {
468 /* from Top down-to is always allowed */
471 if (node->type.tv == tarval_bottom || node->type.tv == tarval_reachable) {
475 panic("combo: wrong translation from %+F to %+F on node %+F", old_type, node->type, node->node);
479 #define verify_type(old_type, node)
483 * Compare two pointer values of a listmap.
485 static int listmap_cmp_ptr(const void *elt, const void *key, size_t size) {
486 const listmap_entry_t *e1 = elt;
487 const listmap_entry_t *e2 = key;
490 return e1->id != e2->id;
491 } /* listmap_cmp_ptr */
494 * Initializes a listmap.
496 * @param map the listmap
498 static void listmap_init(listmap_t *map) {
499 map->map = new_set(listmap_cmp_ptr, 16);
504 * Terminates a listmap.
506 * @param map the listmap
508 static void listmap_term(listmap_t *map) {
513 * Return the associated listmap entry for a given id.
515 * @param map the listmap
516 * @param id the id to search for
518 * @return the associated listmap entry for the given id
520 static listmap_entry_t *listmap_find(listmap_t *map, void *id) {
521 listmap_entry_t key, *entry;
526 entry = set_insert(map->map, &key, sizeof(key), HASH_PTR(id));
528 if (entry->list == NULL) {
529 /* a new entry, put into the list */
530 entry->next = map->values;
537 * Calculate the hash value for an opcode map entry.
539 * @param entry an opcode map entry
541 * @return a hash value for the given opcode map entry
543 static unsigned opcode_hash(const opcode_key_t *entry) {
544 return (entry->mode - (ir_mode *)0) * 9 + entry->code + entry->u.proj * 3 + HASH_PTR(entry->u.ent) + entry->arity;
548 * Compare two entries in the opcode map.
550 static int cmp_opcode(const void *elt, const void *key, size_t size) {
551 const opcode_key_t *o1 = elt;
552 const opcode_key_t *o2 = key;
555 return o1->code != o2->code || o1->mode != o2->mode ||
556 o1->arity != o2->arity ||
557 o1->u.proj != o2->u.proj || o1->u.ent != o2->u.ent;
561 * Compare two Def-Use edges for input position.
563 static int cmp_def_use_edge(const void *a, const void *b) {
564 const ir_def_use_edge *ea = a;
565 const ir_def_use_edge *eb = b;
567 /* no overrun, because range is [-1, MAXINT] */
568 return ea->pos - eb->pos;
569 } /* cmp_def_use_edge */
572 * We need the Def-Use edges sorted.
574 static void sort_irn_outs(node_t *node) {
575 ir_node *irn = node->node;
576 int n_outs = get_irn_n_outs(irn);
579 qsort(&irn->out[1], n_outs, sizeof(irn->out[0]), cmp_def_use_edge);
581 node->max_user_input = irn->out[n_outs].pos;
582 } /* sort_irn_outs */
585 * Return the type of a node.
587 * @param irn an IR-node
589 * @return the associated type of this node
591 static inline lattice_elem_t get_node_type(const ir_node *irn) {
592 return get_irn_node(irn)->type;
593 } /* get_node_type */
596 * Return the tarval of a node.
598 * @param irn an IR-node
600 * @return the associated type of this node
602 static inline tarval *get_node_tarval(const ir_node *irn) {
603 lattice_elem_t type = get_node_type(irn);
605 if (is_tarval(type.tv))
607 return tarval_bottom;
608 } /* get_node_type */
611 * Add a partition to the worklist.
613 static inline void add_to_worklist(partition_t *X, environment_t *env) {
614 assert(X->on_worklist == 0);
615 DB((dbg, LEVEL_2, "Adding part%d to worklist\n", X->nr));
616 X->wl_next = env->worklist;
619 } /* add_to_worklist */
622 * Create a new empty partition.
624 * @param env the environment
626 * @return a newly allocated partition
628 static inline partition_t *new_partition(environment_t *env) {
629 partition_t *part = obstack_alloc(&env->obst, sizeof(*part));
631 INIT_LIST_HEAD(&part->Leader);
632 INIT_LIST_HEAD(&part->Follower);
633 INIT_LIST_HEAD(&part->cprop);
634 INIT_LIST_HEAD(&part->cprop_X);
635 part->wl_next = NULL;
636 part->touched_next = NULL;
637 part->cprop_next = NULL;
638 part->split_next = NULL;
639 part->touched = NULL;
642 part->max_user_inputs = 0;
643 part->on_worklist = 0;
644 part->on_touched = 0;
646 part->type_is_T_or_C = 0;
648 part->dbg_next = env->dbg_list;
649 env->dbg_list = part;
650 part->nr = part_nr++;
654 } /* new_partition */
657 * Get the first node from a partition.
659 static inline node_t *get_first_node(const partition_t *X) {
660 return list_entry(X->Leader.next, node_t, node_list);
661 } /* get_first_node */
664 * Return the type of a partition (assuming partition is non-empty and
665 * all elements have the same type).
667 * @param X a partition
669 * @return the type of the first element of the partition
671 static inline lattice_elem_t get_partition_type(const partition_t *X) {
672 const node_t *first = get_first_node(X);
674 } /* get_partition_type */
677 * Creates a partition node for the given IR-node and place it
678 * into the given partition.
680 * @param irn an IR-node
681 * @param part a partition to place the node in
682 * @param env the environment
684 * @return the created node
686 static node_t *create_partition_node(ir_node *irn, partition_t *part, environment_t *env) {
687 /* create a partition node and place it in the partition */
688 node_t *node = obstack_alloc(&env->obst, sizeof(*node));
689 tarval *init = tarval_top;
691 if (is_Proj(irn) && get_irn_mode(irn) == mode_M) {
692 /* mode_M Proj's must never be removed */
693 init = tarval_bottom;
696 INIT_LIST_HEAD(&node->node_list);
697 INIT_LIST_HEAD(&node->cprop_list);
701 node->race_next = NULL;
702 node->type.tv = init;
703 node->max_user_input = 0;
705 node->n_followers = 0;
706 node->on_touched = 0;
709 node->is_follower = 0;
711 set_irn_node(irn, node);
713 list_add_tail(&node->node_list, &part->Leader);
717 } /* create_partition_node */
720 * Pre-Walker, initialize all Nodes' type to U or top and place
721 * all nodes into the TOP partition.
723 static void create_initial_partitions(ir_node *irn, void *ctx) {
724 environment_t *env = ctx;
725 partition_t *part = env->initial;
728 node = create_partition_node(irn, part, env);
730 if (node->max_user_input > part->max_user_inputs)
731 part->max_user_inputs = node->max_user_input;
734 set_Block_phis(irn, NULL);
736 } /* create_initial_partitions */
739 * Post-Walker, collect all Block-Phi lists, set Cond.
741 static void init_block_phis(ir_node *irn, void *ctx) {
745 add_Block_phi(get_nodes_block(irn), irn);
747 } /* init_block_phis */
750 * Add a node to the entry.partition.touched set and
751 * node->partition to the touched set if not already there.
754 * @param env the environment
756 static inline void add_to_touched(node_t *y, environment_t *env) {
757 if (y->on_touched == 0) {
758 partition_t *part = y->part;
760 y->next = part->touched;
765 if (part->on_touched == 0) {
766 part->touched_next = env->touched;
768 part->on_touched = 1;
771 check_list(part->touched, part);
773 } /* add_to_touched */
776 * Place a node on the cprop list.
779 * @param env the environment
781 static void add_to_cprop(node_t *y, environment_t *env) {
784 /* Add y to y.partition.cprop. */
785 if (y->on_cprop == 0) {
786 partition_t *Y = y->part;
787 ir_node *irn = y->node;
789 /* place Conds and all its Projs on the cprop_X list */
790 if (is_Cond(skip_Proj(irn)))
791 list_add_tail(&y->cprop_list, &Y->cprop_X);
793 list_add_tail(&y->cprop_list, &Y->cprop);
796 DB((dbg, LEVEL_3, "Add %+F to part%u.cprop\n", y->node, Y->nr));
798 /* place its partition on the cprop list */
799 if (Y->on_cprop == 0) {
800 Y->cprop_next = env->cprop;
806 if (get_irn_mode(irn) == mode_T) {
807 /* mode_T nodes always produce tarval_bottom, so we must explicitly
808 add it's Proj's to get constant evaluation to work */
811 for (i = get_irn_n_outs(irn) - 1; i >= 0; --i) {
812 node_t *proj = get_irn_node(get_irn_out(irn, i));
814 add_to_cprop(proj, env);
816 } else if (is_Block(irn)) {
817 /* Due to the way we handle Phi's, we must place all Phis of a block on the list
818 * if someone placed the block. The Block is only placed if the reachability
819 * changes, and this must be re-evaluated in compute_Phi(). */
821 for (phi = get_Block_phis(irn); phi != NULL; phi = get_Phi_next(phi)) {
822 node_t *p = get_irn_node(phi);
823 add_to_cprop(p, env);
829 * Update the worklist: If Z is on worklist then add Z' to worklist.
830 * Else add the smaller of Z and Z' to worklist.
832 * @param Z the Z partition
833 * @param Z_prime the Z' partition, a previous part of Z
834 * @param env the environment
836 static void update_worklist(partition_t *Z, partition_t *Z_prime, environment_t *env) {
837 if (Z->on_worklist || Z_prime->n_leader < Z->n_leader) {
838 add_to_worklist(Z_prime, env);
840 add_to_worklist(Z, env);
842 } /* update_worklist */
845 * Make all inputs to x no longer be F.def_use edges.
849 static void move_edges_to_leader(node_t *x) {
850 ir_node *irn = x->node;
853 for (i = get_irn_arity(irn) - 1; i >= 0; --i) {
854 node_t *pred = get_irn_node(get_irn_n(irn, i));
859 n = get_irn_n_outs(p);
860 for (j = 1; j <= pred->n_followers; ++j) {
861 if (p->out[j].pos == i && p->out[j].use == irn) {
862 /* found a follower edge to x, move it to the Leader */
863 ir_def_use_edge edge = p->out[j];
865 /* remove this edge from the Follower set */
866 p->out[j] = p->out[pred->n_followers];
869 /* sort it into the leader set */
870 for (k = pred->n_followers + 2; k <= n; ++k) {
871 if (p->out[k].pos >= edge.pos)
873 p->out[k - 1] = p->out[k];
875 /* place the new edge here */
876 p->out[k - 1] = edge;
878 /* edge found and moved */
883 } /* move_edges_to_leader */
886 * Split a partition that has NO followers by a local list.
888 * @param Z partition to split
889 * @param g a (non-empty) node list
890 * @param env the environment
892 * @return a new partition containing the nodes of g
894 static partition_t *split_no_followers(partition_t *Z, node_t *g, environment_t *env) {
895 partition_t *Z_prime;
900 dump_partition("Splitting ", Z);
901 dump_list("by list ", g);
905 /* Remove g from Z. */
906 for (node = g; node != NULL; node = node->next) {
907 assert(node->part == Z);
908 list_del(&node->node_list);
911 assert(n < Z->n_leader);
914 /* Move g to a new partition, Z'. */
915 Z_prime = new_partition(env);
917 for (node = g; node != NULL; node = node->next) {
918 list_add_tail(&node->node_list, &Z_prime->Leader);
919 node->part = Z_prime;
920 if (node->max_user_input > max_input)
921 max_input = node->max_user_input;
923 Z_prime->max_user_inputs = max_input;
924 Z_prime->n_leader = n;
927 check_partition(Z_prime);
929 /* for now, copy the type info tag, it will be adjusted in split_by(). */
930 Z_prime->type_is_T_or_C = Z->type_is_T_or_C;
932 update_worklist(Z, Z_prime, env);
934 dump_partition("Now ", Z);
935 dump_partition("Created new ", Z_prime);
937 } /* split_no_followers */
940 * Make the Follower -> Leader transition for a node.
944 static void follower_to_leader(node_t *n) {
945 assert(n->is_follower == 1);
947 DB((dbg, LEVEL_2, "%+F make the follower -> leader transition\n", n->node));
949 move_edges_to_leader(n);
950 list_del(&n->node_list);
951 list_add_tail(&n->node_list, &n->part->Leader);
953 } /* follower_to_leader */
956 * The environment for one race step.
958 typedef struct step_env {
959 node_t *initial; /**< The initial node list. */
960 node_t *unwalked; /**< The unwalked node list. */
961 node_t *walked; /**< The walked node list. */
962 int index; /**< Next index of Follower use_def edge. */
963 unsigned side; /**< side number. */
967 * Return non-zero, if a input is a real follower
969 * @param irn the node to check
970 * @param input number of the input
972 static int is_real_follower(const ir_node *irn, int input) {
975 switch (get_irn_opcode(irn)) {
978 /* ignore the Confirm bound input */
984 /* ignore the Mux sel input */
989 /* dead inputs are not follower edges */
990 ir_node *block = get_nodes_block(irn);
991 node_t *pred = get_irn_node(get_Block_cfgpred(block, input));
993 if (pred->type.tv == tarval_unreachable)
1003 /* only a Sub x,0 / Shift x,0 might be a follower */
1010 pred = get_irn_node(get_irn_n(irn, input));
1011 if (is_tarval(pred->type.tv) && tarval_is_null(pred->type.tv))
1015 pred = get_irn_node(get_irn_n(irn, input));
1016 if (is_tarval(pred->type.tv) && tarval_is_one(pred->type.tv))
1020 pred = get_irn_node(get_irn_n(irn, input));
1021 if (is_tarval(pred->type.tv) && tarval_is_all_one(pred->type.tv))
1026 /* all inputs are followers */
1029 assert(!"opcode not implemented yet");
1033 } /* is_real_follower */
1036 * Do one step in the race.
1038 static int step(step_env *env) {
1041 if (env->initial != NULL) {
1042 /* Move node from initial to unwalked */
1044 env->initial = n->race_next;
1046 n->race_next = env->unwalked;
1052 while (env->unwalked != NULL) {
1053 /* let n be the first node in unwalked */
1055 while (env->index < n->n_followers) {
1056 const ir_def_use_edge *edge = &n->node->out[1 + env->index];
1058 /* let m be n.F.def_use[index] */
1059 node_t *m = get_irn_node(edge->use);
1061 assert(m->is_follower);
1063 * Some inputs, like the get_Confirm_bound are NOT
1064 * real followers, sort them out.
1066 if (! is_real_follower(m->node, edge->pos)) {
1072 /* only followers from our partition */
1073 if (m->part != n->part)
1076 if ((m->flagged & env->side) == 0) {
1077 m->flagged |= env->side;
1079 if (m->flagged != 3) {
1080 /* visited the first time */
1081 /* add m to unwalked not as first node (we might still need to
1082 check for more follower node */
1083 m->race_next = n->race_next;
1087 /* else already visited by the other side and on the other list */
1090 /* move n to walked */
1091 env->unwalked = n->race_next;
1092 n->race_next = env->walked;
1100 * Clear the flags from a list and check for
1101 * nodes that where touched from both sides.
1103 * @param list the list
1105 static int clear_flags(node_t *list) {
1109 for (n = list; n != NULL; n = n->race_next) {
1110 if (n->flagged == 3) {
1111 /* we reach a follower from both sides, this will split congruent
1112 * inputs and make it a leader. */
1113 follower_to_leader(n);
1122 * Split a partition by a local list using the race.
1124 * @param pX pointer to the partition to split, might be changed!
1125 * @param gg a (non-empty) node list
1126 * @param env the environment
1128 * @return a new partition containing the nodes of gg
1130 static partition_t *split(partition_t **pX, node_t *gg, environment_t *env) {
1131 partition_t *X = *pX;
1132 partition_t *X_prime;
1135 node_t *g, *h, *node, *t;
1136 int max_input, transitions, winner, shf;
1138 DEBUG_ONLY(static int run = 0;)
1140 DB((dbg, LEVEL_2, "Run %d ", run++));
1141 if (list_empty(&X->Follower)) {
1142 /* if the partition has NO follower, we can use the fast
1143 splitting algorithm. */
1144 return split_no_followers(X, gg, env);
1146 /* else do the race */
1148 dump_partition("Splitting ", X);
1149 dump_list("by list ", gg);
1151 INIT_LIST_HEAD(&tmp);
1153 /* Remove gg from X.Leader and put into g */
1155 for (node = gg; node != NULL; node = node->next) {
1156 assert(node->part == X);
1157 assert(node->is_follower == 0);
1159 list_del(&node->node_list);
1160 list_add_tail(&node->node_list, &tmp);
1161 node->race_next = g;
1166 list_for_each_entry(node_t, node, &X->Leader, node_list) {
1167 node->race_next = h;
1170 /* restore X.Leader */
1171 list_splice(&tmp, &X->Leader);
1173 senv[0].initial = g;
1174 senv[0].unwalked = NULL;
1175 senv[0].walked = NULL;
1179 senv[1].initial = h;
1180 senv[1].unwalked = NULL;
1181 senv[1].walked = NULL;
1186 * Some informations on the race that are not stated clearly in Click's
1188 * 1) A follower stays on the side that reach him first.
1189 * 2) If the other side reches a follower, if will be converted to
1190 * a leader. /This must be done after the race is over, else the
1191 * edges we are iterating on are renumbered./
1192 * 3) /New leader might end up on both sides./
1193 * 4) /If one side ends up with new Leaders, we must ensure that
1194 * they can split out by opcode, hence we have to put _every_
1195 * partition with new Leader nodes on the cprop list, as
1196 * opcode splitting is done by split_by() at the end of
1197 * constant propagation./
1200 if (step(&senv[0])) {
1204 if (step(&senv[1])) {
1209 assert(senv[winner].initial == NULL);
1210 assert(senv[winner].unwalked == NULL);
1212 /* clear flags from walked/unwalked */
1214 transitions = clear_flags(senv[0].unwalked) << shf;
1215 transitions |= clear_flags(senv[0].walked) << shf;
1217 transitions |= clear_flags(senv[1].unwalked) << shf;
1218 transitions |= clear_flags(senv[1].walked) << shf;
1220 dump_race_list("winner ", senv[winner].walked);
1222 /* Move walked_{winner} to a new partition, X'. */
1223 X_prime = new_partition(env);
1226 for (node = senv[winner].walked; node != NULL; node = node->race_next) {
1227 list_del(&node->node_list);
1228 node->part = X_prime;
1229 if (node->is_follower) {
1230 list_add_tail(&node->node_list, &X_prime->Follower);
1232 list_add_tail(&node->node_list, &X_prime->Leader);
1235 if (node->max_user_input > max_input)
1236 max_input = node->max_user_input;
1238 X_prime->n_leader = n;
1239 X_prime->max_user_inputs = max_input;
1240 X->n_leader -= X_prime->n_leader;
1242 /* for now, copy the type info tag, it will be adjusted in split_by(). */
1243 X_prime->type_is_T_or_C = X->type_is_T_or_C;
1246 * Even if a follower was not checked by both sides, it might have
1247 * loose its congruence, so we need to check this case for all follower.
1249 list_for_each_entry_safe(node_t, node, t, &X_prime->Follower, node_list) {
1250 if (identity(node) == node) {
1251 follower_to_leader(node);
1257 check_partition(X_prime);
1259 /* X' is the smaller part */
1260 add_to_worklist(X_prime, env);
1263 * If there where follower to leader transitions, ensure that the nodes
1264 * can be split out if necessary.
1266 if (transitions & 1) {
1267 /* place winner partition on the cprop list */
1268 if (X_prime->on_cprop == 0) {
1269 X_prime->cprop_next = env->cprop;
1270 env->cprop = X_prime;
1271 X_prime->on_cprop = 1;
1274 if (transitions & 2) {
1275 /* place other partition on the cprop list */
1276 if (X->on_cprop == 0) {
1277 X->cprop_next = env->cprop;
1283 dump_partition("Now ", X);
1284 dump_partition("Created new ", X_prime);
1286 /* we have to ensure that the partition containing g is returned */
1296 * Returns non-zero if the i'th input of a Phi node is live.
1298 * @param phi a Phi-node
1299 * @param i an input number
1301 * @return non-zero if the i'th input of the given Phi node is live
1303 static int is_live_input(ir_node *phi, int i) {
1305 ir_node *block = get_nodes_block(phi);
1306 ir_node *pred = get_Block_cfgpred(block, i);
1307 lattice_elem_t type = get_node_type(pred);
1309 return type.tv != tarval_unreachable;
1311 /* else it's the control input, always live */
1313 } /* is_live_input */
1316 * Return non-zero if a type is a constant.
1318 static int is_constant_type(lattice_elem_t type) {
1319 if (type.tv != tarval_bottom && type.tv != tarval_top)
1322 } /* is_constant_type */
1325 * Check whether a type is neither Top or a constant.
1326 * Note: U is handled like Top here, R is a constant.
1328 * @param type the type to check
1330 static int type_is_neither_top_nor_const(const lattice_elem_t type) {
1331 if (is_tarval(type.tv)) {
1332 if (type.tv == tarval_top)
1334 if (tarval_is_constant(type.tv))
1341 } /* type_is_neither_top_nor_const */
1344 * Collect nodes to the touched list.
1346 * @param list the list which contains the nodes that must be evaluated
1347 * @param idx the index of the def_use edge to evaluate
1348 * @param env the environment
1350 static void collect_touched(list_head *list, int idx, environment_t *env) {
1352 int end_idx = env->end_idx;
1354 list_for_each_entry(node_t, x, list, node_list) {
1358 /* leader edges start AFTER follower edges */
1359 x->next_edge = x->n_followers + 1;
1361 num_edges = get_irn_n_outs(x->node);
1363 /* for all edges in x.L.def_use_{idx} */
1364 while (x->next_edge <= num_edges) {
1365 const ir_def_use_edge *edge = &x->node->out[x->next_edge];
1368 /* check if we have necessary edges */
1369 if (edge->pos > idx)
1376 /* only non-commutative nodes */
1377 if (env->commutative &&
1378 (idx == 0 || idx == 1) && is_op_commutative(get_irn_op(succ)))
1381 /* ignore the "control input" for non-pinned nodes
1382 if we are running in GCSE mode */
1383 if (idx < end_idx && get_irn_pinned(succ) != op_pin_state_pinned)
1386 y = get_irn_node(succ);
1387 assert(get_irn_n(succ, idx) == x->node);
1389 /* ignore block edges touching followers */
1390 if (idx == -1 && y->is_follower)
1393 if (is_constant_type(y->type)) {
1394 ir_opcode code = get_irn_opcode(succ);
1395 if (code == iro_Sub || code == iro_Cmp)
1396 add_to_cprop(y, env);
1399 /* Partitions of constants should not be split simply because their Nodes have unequal
1400 functions or incongruent inputs. */
1401 if (type_is_neither_top_nor_const(y->type) &&
1402 (! is_Phi(y->node) || is_live_input(y->node, idx))) {
1403 add_to_touched(y, env);
1407 } /* collect_touched */
1410 * Collect commutative nodes to the touched list.
1412 * @param list the list which contains the nodes that must be evaluated
1413 * @param env the environment
1415 static void collect_commutative_touched(list_head *list, environment_t *env) {
1418 list_for_each_entry(node_t, x, list, node_list) {
1421 num_edges = get_irn_n_outs(x->node);
1423 x->next_edge = x->n_followers + 1;
1425 /* for all edges in x.L.def_use_{idx} */
1426 while (x->next_edge <= num_edges) {
1427 const ir_def_use_edge *edge = &x->node->out[x->next_edge];
1430 /* check if we have necessary edges */
1440 /* only commutative nodes */
1441 if (!is_op_commutative(get_irn_op(succ)))
1444 y = get_irn_node(succ);
1445 if (is_constant_type(y->type)) {
1446 ir_opcode code = get_irn_opcode(succ);
1447 if (code == iro_Eor)
1448 add_to_cprop(y, env);
1451 /* Partitions of constants should not be split simply because their Nodes have unequal
1452 functions or incongruent inputs. */
1453 if (type_is_neither_top_nor_const(y->type)) {
1454 add_to_touched(y, env);
1458 } /* collect_commutative_touched */
1461 * Split the partitions if caused by the first entry on the worklist.
1463 * @param env the environment
1465 static void cause_splits(environment_t *env) {
1466 partition_t *X, *Z, *N;
1469 /* remove the first partition from the worklist */
1471 env->worklist = X->wl_next;
1474 dump_partition("Cause_split: ", X);
1476 if (env->commutative) {
1477 /* handle commutative nodes first */
1479 /* empty the touched set: already done, just clear the list */
1480 env->touched = NULL;
1482 collect_commutative_touched(&X->Leader, env);
1483 collect_commutative_touched(&X->Follower, env);
1485 for (Z = env->touched; Z != NULL; Z = N) {
1487 node_t *touched = Z->touched;
1488 unsigned n_touched = Z->n_touched;
1490 assert(Z->touched != NULL);
1492 /* beware, split might change Z */
1493 N = Z->touched_next;
1495 /* remove it from the touched set */
1498 /* Empty local Z.touched. */
1499 for (e = touched; e != NULL; e = e->next) {
1500 assert(e->is_follower == 0);
1506 if (0 < n_touched && n_touched < Z->n_leader) {
1507 DB((dbg, LEVEL_2, "Split part%d by touched\n", Z->nr));
1508 split(&Z, touched, env);
1510 assert(n_touched <= Z->n_leader);
1514 /* combine temporary leader and follower list */
1515 for (idx = -1; idx <= X->max_user_inputs; ++idx) {
1516 /* empty the touched set: already done, just clear the list */
1517 env->touched = NULL;
1519 collect_touched(&X->Leader, idx, env);
1520 collect_touched(&X->Follower, idx, env);
1522 for (Z = env->touched; Z != NULL; Z = N) {
1524 node_t *touched = Z->touched;
1525 unsigned n_touched = Z->n_touched;
1527 assert(Z->touched != NULL);
1529 /* beware, split might change Z */
1530 N = Z->touched_next;
1532 /* remove it from the touched set */
1535 /* Empty local Z.touched. */
1536 for (e = touched; e != NULL; e = e->next) {
1537 assert(e->is_follower == 0);
1543 if (0 < n_touched && n_touched < Z->n_leader) {
1544 DB((dbg, LEVEL_2, "Split part%d by touched\n", Z->nr));
1545 split(&Z, touched, env);
1547 assert(n_touched <= Z->n_leader);
1550 } /* cause_splits */
1553 * Implements split_by_what(): Split a partition by characteristics given
1554 * by the what function.
1556 * @param X the partition to split
1557 * @param What a function returning an Id for every node of the partition X
1558 * @param P a list to store the result partitions
1559 * @param env the environment
1563 static partition_t *split_by_what(partition_t *X, what_func What,
1564 partition_t **P, environment_t *env) {
1567 listmap_entry_t *iter;
1570 /* Let map be an empty mapping from the range of What to (local) list of Nodes. */
1572 list_for_each_entry(node_t, x, &X->Leader, node_list) {
1573 void *id = What(x, env);
1574 listmap_entry_t *entry;
1577 /* input not allowed, ignore */
1580 /* Add x to map[What(x)]. */
1581 entry = listmap_find(&map, id);
1582 x->next = entry->list;
1585 /* Let P be a set of Partitions. */
1587 /* for all sets S except one in the range of map do */
1588 for (iter = map.values; iter != NULL; iter = iter->next) {
1589 if (iter->next == NULL) {
1590 /* this is the last entry, ignore */
1595 /* Add SPLIT( X, S ) to P. */
1596 DB((dbg, LEVEL_2, "Split part%d by WHAT = %s\n", X->nr, what_reason));
1597 R = split(&X, S, env);
1607 } /* split_by_what */
1609 /** lambda n.(n.type) */
1610 static void *lambda_type(const node_t *node, environment_t *env) {
1612 return node->type.tv;
1615 /** lambda n.(n.opcode) */
1616 static void *lambda_opcode(const node_t *node, environment_t *env) {
1617 opcode_key_t key, *entry;
1618 ir_node *irn = node->node;
1620 key.code = get_irn_opcode(irn);
1621 key.mode = get_irn_mode(irn);
1622 key.arity = get_irn_arity(irn);
1626 switch (get_irn_opcode(irn)) {
1628 key.u.proj = get_Proj_proj(irn);
1631 key.u.ent = get_Sel_entity(irn);
1637 entry = set_insert(env->opcode2id_map, &key, sizeof(key), opcode_hash(&key));
1639 } /* lambda_opcode */
1641 /** lambda n.(n[i].partition) */
1642 static void *lambda_partition(const node_t *node, environment_t *env) {
1643 ir_node *skipped = skip_Proj(node->node);
1646 int i = env->lambda_input;
1648 if (i >= get_irn_arity(node->node)) {
1650 * We are outside the allowed range: This can happen even
1651 * if we have split by opcode first: doing so might move Followers
1652 * to Leaders and those will have a different opcode!
1653 * Note that in this case the partition is on the cprop list and will be
1659 /* ignore the "control input" for non-pinned nodes
1660 if we are running in GCSE mode */
1661 if (i < env->end_idx && get_irn_pinned(skipped) != op_pin_state_pinned)
1664 pred = i == -1 ? get_irn_n(skipped, i) : get_irn_n(node->node, i);
1665 p = get_irn_node(pred);
1668 } /* lambda_partition */
1670 /** lambda n.(n[i].partition) for commutative nodes */
1671 static void *lambda_commutative_partition(const node_t *node, environment_t *env) {
1672 ir_node *irn = node->node;
1673 ir_node *skipped = skip_Proj(irn);
1674 ir_node *pred, *left, *right;
1676 partition_t *pl, *pr;
1677 int i = env->lambda_input;
1679 if (i >= get_irn_arity(node->node)) {
1681 * We are outside the allowed range: This can happen even
1682 * if we have split by opcode first: doing so might move Followers
1683 * to Leaders and those will have a different opcode!
1684 * Note that in this case the partition is on the cprop list and will be
1690 /* ignore the "control input" for non-pinned nodes
1691 if we are running in GCSE mode */
1692 if (i < env->end_idx && get_irn_pinned(skipped) != op_pin_state_pinned)
1696 pred = get_irn_n(skipped, i);
1697 p = get_irn_node(pred);
1701 if (is_op_commutative(get_irn_op(irn))) {
1702 /* normalize partition order by returning the "smaller" on input 0,
1703 the "bigger" on input 1. */
1704 left = get_binop_left(irn);
1705 pl = get_irn_node(left)->part;
1706 right = get_binop_right(irn);
1707 pr = get_irn_node(right)->part;
1710 return pl < pr ? pl : pr;
1712 return pl > pr ? pl : pr;
1714 /* a not split out Follower */
1715 pred = get_irn_n(irn, i);
1716 p = get_irn_node(pred);
1720 } /* lambda_commutative_partition */
1723 * Returns true if a type is a constant (and NOT Top
1726 static int is_con(const lattice_elem_t type) {
1727 /* be conservative */
1728 if (is_tarval(type.tv))
1729 return tarval_is_constant(type.tv);
1730 return is_entity(type.sym.entity_p);
1734 * Implements split_by().
1736 * @param X the partition to split
1737 * @param env the environment
1739 static void split_by(partition_t *X, environment_t *env) {
1740 partition_t *I, *P = NULL;
1743 dump_partition("split_by", X);
1745 if (X->n_leader == 1) {
1746 /* we have only one leader, no need to split, just check it's type */
1747 node_t *x = get_first_node(X);
1748 X->type_is_T_or_C = x->type.tv == tarval_top || is_con(x->type);
1752 DEBUG_ONLY(what_reason = "lambda n.(n.type)";)
1753 P = split_by_what(X, lambda_type, &P, env);
1756 /* adjust the type tags, we have split partitions by type */
1757 for (I = P; I != NULL; I = I->split_next) {
1758 node_t *x = get_first_node(I);
1759 I->type_is_T_or_C = x->type.tv == tarval_top || is_con(x->type);
1766 if (Y->n_leader > 1) {
1767 /* we do not want split the TOP or constant partitions */
1768 if (! Y->type_is_T_or_C) {
1769 partition_t *Q = NULL;
1771 DEBUG_ONLY(what_reason = "lambda n.(n.opcode)";)
1772 Q = split_by_what(Y, lambda_opcode, &Q, env);
1779 if (Z->n_leader > 1) {
1780 const node_t *first = get_first_node(Z);
1781 int arity = get_irn_arity(first->node);
1783 what_func what = lambda_partition;
1784 DEBUG_ONLY(char buf[64];)
1786 if (env->commutative && is_op_commutative(get_irn_op(first->node)))
1787 what = lambda_commutative_partition;
1790 * BEWARE: during splitting by input 2 for instance we might
1791 * create new partitions which are different by input 1, so collect
1792 * them and split further.
1794 Z->split_next = NULL;
1797 for (input = arity - 1; input >= -1; --input) {
1799 partition_t *Z_prime = R;
1802 if (Z_prime->n_leader > 1) {
1803 env->lambda_input = input;
1804 DEBUG_ONLY(snprintf(buf, sizeof(buf), "lambda n.(n[%d].partition)", input);)
1805 DEBUG_ONLY(what_reason = buf;)
1806 S = split_by_what(Z_prime, what, &S, env);
1809 Z_prime->split_next = S;
1812 } while (R != NULL);
1817 } while (Q != NULL);
1820 } while (P != NULL);
1824 * (Re-)compute the type for a given node.
1826 * @param node the node
1828 static void default_compute(node_t *node) {
1830 ir_node *irn = node->node;
1831 node_t *block = get_irn_node(get_nodes_block(irn));
1833 if (block->type.tv == tarval_unreachable) {
1834 node->type.tv = tarval_top;
1838 /* if any of the data inputs have type top, the result is type top */
1839 for (i = get_irn_arity(irn) - 1; i >= 0; --i) {
1840 ir_node *pred = get_irn_n(irn, i);
1841 node_t *p = get_irn_node(pred);
1843 if (p->type.tv == tarval_top) {
1844 node->type.tv = tarval_top;
1849 if (get_irn_mode(node->node) == mode_X)
1850 node->type.tv = tarval_reachable;
1852 node->type.tv = computed_value(irn);
1853 } /* default_compute */
1856 * (Re-)compute the type for a Block node.
1858 * @param node the node
1860 static void compute_Block(node_t *node) {
1862 ir_node *block = node->node;
1864 if (block == get_irg_start_block(current_ir_graph)) {
1865 /* start block is always reachable */
1866 node->type.tv = tarval_reachable;
1870 for (i = get_Block_n_cfgpreds(block) - 1; i >= 0; --i) {
1871 node_t *pred = get_irn_node(get_Block_cfgpred(block, i));
1873 if (pred->type.tv == tarval_reachable) {
1874 /* A block is reachable, if at least of predecessor is reachable. */
1875 node->type.tv = tarval_reachable;
1879 node->type.tv = tarval_top;
1880 } /* compute_Block */
1883 * (Re-)compute the type for a Bad node.
1885 * @param node the node
1887 static void compute_Bad(node_t *node) {
1888 /* Bad nodes ALWAYS compute Top */
1889 node->type.tv = tarval_top;
1893 * (Re-)compute the type for an Unknown node.
1895 * @param node the node
1897 static void compute_Unknown(node_t *node) {
1898 /* While Unknown nodes should compute Top this is dangerous:
1899 * a Top input to a Cond would lead to BOTH control flows unreachable.
1900 * While this is correct in the given semantics, it would destroy the Firm
1903 * It would be safe to compute Top IF it can be assured, that only Cmp
1904 * nodes are inputs to Conds. We check that first.
1905 * This is the way Frontends typically build Firm, but some optimizations
1906 * (cond_eval for instance) might replace them by Phib's...
1908 node->type.tv = tarval_UNKNOWN;
1909 } /* compute_Unknown */
1912 * (Re-)compute the type for a Jmp node.
1914 * @param node the node
1916 static void compute_Jmp(node_t *node) {
1917 node_t *block = get_irn_node(get_nodes_block(node->node));
1919 node->type = block->type;
1923 * (Re-)compute the type for the Return node.
1925 * @param node the node
1927 static void compute_Return(node_t *node) {
1928 /* The Return node is NOT dead if it is in a reachable block.
1929 * This is already checked in compute(). so we can return
1930 * Reachable here. */
1931 node->type.tv = tarval_reachable;
1932 } /* compute_Return */
1935 * (Re-)compute the type for the End node.
1937 * @param node the node
1939 static void compute_End(node_t *node) {
1940 /* the End node is NOT dead of course */
1941 node->type.tv = tarval_reachable;
1945 * (Re-)compute the type for a Call.
1947 * @param node the node
1949 static void compute_Call(node_t *node) {
1951 * A Call computes always bottom, even if it has Unknown
1954 node->type.tv = tarval_bottom;
1955 } /* compute_Call */
1958 * (Re-)compute the type for a SymConst node.
1960 * @param node the node
1962 static void compute_SymConst(node_t *node) {
1963 ir_node *irn = node->node;
1964 node_t *block = get_irn_node(get_nodes_block(irn));
1966 if (block->type.tv == tarval_unreachable) {
1967 node->type.tv = tarval_top;
1970 switch (get_SymConst_kind(irn)) {
1971 case symconst_addr_ent:
1972 /* case symconst_addr_name: cannot handle this yet */
1973 node->type.sym = get_SymConst_symbol(irn);
1976 node->type.tv = computed_value(irn);
1978 } /* compute_SymConst */
1981 * (Re-)compute the type for a Phi node.
1983 * @param node the node
1985 static void compute_Phi(node_t *node) {
1987 ir_node *phi = node->node;
1988 lattice_elem_t type;
1990 /* if a Phi is in a unreachable block, its type is TOP */
1991 node_t *block = get_irn_node(get_nodes_block(phi));
1993 if (block->type.tv == tarval_unreachable) {
1994 node->type.tv = tarval_top;
1998 /* Phi implements the Meet operation */
1999 type.tv = tarval_top;
2000 for (i = get_Phi_n_preds(phi) - 1; i >= 0; --i) {
2001 node_t *pred = get_irn_node(get_Phi_pred(phi, i));
2002 node_t *pred_X = get_irn_node(get_Block_cfgpred(block->node, i));
2004 if (pred_X->type.tv == tarval_unreachable || pred->type.tv == tarval_top) {
2005 /* ignore TOP inputs: We must check here for unreachable blocks,
2006 because Firm constants live in the Start Block are NEVER Top.
2007 Else, a Phi (1,2) will produce Bottom, even if the 2 for instance
2008 comes from a unreachable input. */
2011 if (pred->type.tv == tarval_bottom) {
2012 node->type.tv = tarval_bottom;
2014 } else if (type.tv == tarval_top) {
2015 /* first constant found */
2017 } else if (type.tv != pred->type.tv) {
2018 /* different constants or tarval_bottom */
2019 node->type.tv = tarval_bottom;
2022 /* else nothing, constants are the same */
2028 * (Re-)compute the type for an Add. Special case: one nodes is a Zero Const.
2030 * @param node the node
2032 static void compute_Add(node_t *node) {
2033 ir_node *sub = node->node;
2034 node_t *l = get_irn_node(get_Add_left(sub));
2035 node_t *r = get_irn_node(get_Add_right(sub));
2036 lattice_elem_t a = l->type;
2037 lattice_elem_t b = r->type;
2040 if (a.tv == tarval_top || b.tv == tarval_top) {
2041 node->type.tv = tarval_top;
2042 } else if (a.tv == tarval_bottom || b.tv == tarval_bottom) {
2043 node->type.tv = tarval_bottom;
2045 /* x + 0 = 0 + x = x, but beware of floating point +0 + -0, so we
2046 must call tarval_add() first to handle this case! */
2047 if (is_tarval(a.tv)) {
2048 if (is_tarval(b.tv)) {
2049 node->type.tv = tarval_add(a.tv, b.tv);
2052 mode = get_tarval_mode(a.tv);
2053 if (a.tv == get_mode_null(mode)) {
2057 } else if (is_tarval(b.tv)) {
2058 mode = get_tarval_mode(b.tv);
2059 if (b.tv == get_mode_null(mode)) {
2064 node->type.tv = tarval_bottom;
2069 * (Re-)compute the type for a Sub. Special case: both nodes are congruent.
2071 * @param node the node
2073 static void compute_Sub(node_t *node) {
2074 ir_node *sub = node->node;
2075 node_t *l = get_irn_node(get_Sub_left(sub));
2076 node_t *r = get_irn_node(get_Sub_right(sub));
2077 lattice_elem_t a = l->type;
2078 lattice_elem_t b = r->type;
2081 if (a.tv == tarval_top || b.tv == tarval_top) {
2082 node->type.tv = tarval_top;
2083 } else if (is_con(a) && is_con(b)) {
2084 if (is_tarval(a.tv) && is_tarval(b.tv)) {
2085 node->type.tv = tarval_sub(a.tv, b.tv, get_irn_mode(sub));
2086 } else if (is_tarval(a.tv) && tarval_is_null(a.tv)) {
2088 } else if (is_tarval(b.tv) && tarval_is_null(b.tv)) {
2091 node->type.tv = tarval_bottom;
2093 } else if (r->part == l->part &&
2094 (!mode_is_float(get_irn_mode(l->node)))) {
2096 * BEWARE: a - a is NOT always 0 for floating Point values, as
2097 * NaN op NaN = NaN, so we must check this here.
2099 ir_mode *mode = get_irn_mode(sub);
2100 tv = get_mode_null(mode);
2102 /* if the node was ONCE evaluated by all constants, but now
2103 this breaks AND we get from the argument partitions a different
2104 result, switch to bottom.
2105 This happens because initially all nodes are in the same partition ... */
2106 if (node->type.tv != tv)
2110 node->type.tv = tarval_bottom;
2115 * (Re-)compute the type for an Eor. Special case: both nodes are congruent.
2117 * @param node the node
2119 static void compute_Eor(node_t *node) {
2120 ir_node *eor = node->node;
2121 node_t *l = get_irn_node(get_Eor_left(eor));
2122 node_t *r = get_irn_node(get_Eor_right(eor));
2123 lattice_elem_t a = l->type;
2124 lattice_elem_t b = r->type;
2127 if (a.tv == tarval_top || b.tv == tarval_top) {
2128 node->type.tv = tarval_top;
2129 } else if (is_con(a) && is_con(b)) {
2130 if (is_tarval(a.tv) && is_tarval(b.tv)) {
2131 node->type.tv = tarval_eor(a.tv, b.tv);
2132 } else if (is_tarval(a.tv) && tarval_is_null(a.tv)) {
2134 } else if (is_tarval(b.tv) && tarval_is_null(b.tv)) {
2137 node->type.tv = tarval_bottom;
2139 } else if (r->part == l->part) {
2140 ir_mode *mode = get_irn_mode(eor);
2141 tv = get_mode_null(mode);
2143 /* if the node was ONCE evaluated by all constants, but now
2144 this breaks AND we get from the argument partitions a different
2145 result, switch to bottom.
2146 This happens because initially all nodes are in the same partition ... */
2147 if (node->type.tv != tv)
2151 node->type.tv = tarval_bottom;
2156 * (Re-)compute the type for Cmp.
2158 * @param node the node
2160 static void compute_Cmp(node_t *node) {
2161 ir_node *cmp = node->node;
2162 node_t *l = get_irn_node(get_Cmp_left(cmp));
2163 node_t *r = get_irn_node(get_Cmp_right(cmp));
2164 lattice_elem_t a = l->type;
2165 lattice_elem_t b = r->type;
2167 if (a.tv == tarval_top || b.tv == tarval_top) {
2168 node->type.tv = tarval_top;
2169 } else if (r->part == l->part) {
2170 /* both nodes congruent, we can probably do something */
2171 node->type.tv = tarval_b_true;
2172 } else if (is_con(a) && is_con(b)) {
2173 /* both nodes are constants, we can probably do something */
2174 node->type.tv = tarval_b_true;
2176 node->type.tv = tarval_bottom;
2181 * (Re-)compute the type for a Proj(Cmp).
2183 * @param node the node
2184 * @param cond the predecessor Cmp node
2186 static void compute_Proj_Cmp(node_t *node, ir_node *cmp) {
2187 ir_node *proj = node->node;
2188 node_t *l = get_irn_node(get_Cmp_left(cmp));
2189 node_t *r = get_irn_node(get_Cmp_right(cmp));
2190 lattice_elem_t a = l->type;
2191 lattice_elem_t b = r->type;
2192 pn_Cmp pnc = get_Proj_proj(proj);
2195 if (a.tv == tarval_top || b.tv == tarval_top) {
2196 node->type.tv = tarval_undefined;
2197 } else if (is_con(a) && is_con(b)) {
2198 default_compute(node);
2199 } else if (r->part == l->part &&
2200 (!mode_is_float(get_irn_mode(l->node)) || pnc == pn_Cmp_Lt || pnc == pn_Cmp_Gt)) {
2202 * BEWARE: a == a is NOT always True for floating Point values, as
2203 * NaN != NaN is defined, so we must check this here.
2205 tv = pnc & pn_Cmp_Eq ? tarval_b_true: tarval_b_false;
2207 /* if the node was ONCE evaluated by all constants, but now
2208 this breaks AND we get from the argument partitions a different
2209 result, switch to bottom.
2210 This happens because initially all nodes are in the same partition ... */
2211 if (node->type.tv != tv)
2215 node->type.tv = tarval_bottom;
2217 } /* compute_Proj_Cmp */
2220 * (Re-)compute the type for a Proj(Cond).
2222 * @param node the node
2223 * @param cond the predecessor Cond node
2225 static void compute_Proj_Cond(node_t *node, ir_node *cond) {
2226 ir_node *proj = node->node;
2227 long pnc = get_Proj_proj(proj);
2228 ir_node *sel = get_Cond_selector(cond);
2229 node_t *selector = get_irn_node(sel);
2232 * Note: it is crucial for the monotony that the Proj(Cond)
2233 * are evaluates after all predecessors of the Cond selector are
2239 * Due to the fact that 0 is a const, the Cmp gets immediately
2240 * on the cprop list. It will be evaluated before x is evaluated,
2241 * might leaving x as Top. When later x is evaluated, the Cmp
2242 * might change its value.
2243 * BUT if the Cond is evaluated before this happens, Proj(Cond, FALSE)
2244 * gets R, and later changed to F if Cmp is evaluated to True!
2246 * We prevent this by putting Conds in an extra cprop_X queue, which
2247 * gets evaluated after the cprop queue is empty.
2249 * Note that this even happens with Click's original algorithm, if
2250 * Cmp(x, 0) is evaluated to True first and later changed to False
2251 * if x was Top first and later changed to a Const ...
2252 * It is unclear how Click solved that problem ...
2254 * However, in rare cases even this does not help, if a Top reaches
2255 * a compare through a Phi, than Proj(Cond) is evaluated changing
2256 * the type of the Phi to something other.
2257 * So, we take the last resort and bind the type to R once
2260 * (This might be even the way Click works around the whole problem).
2262 * Finally, we may miss some optimization possibilities due to this:
2267 * If Top reaches the if first, than we decide for != here.
2268 * If y later is evaluated to 0, we cannot revert this decision
2269 * and must live with both outputs enabled. If this happens,
2270 * we get an unresolved if (true) in the code ...
2272 * In Click's version where this decision is done at the Cmp,
2273 * the Cmp is NOT optimized away than (if y evaluated to 1
2274 * for instance) and we get a if (1 == 0) here ...
2276 * Both solutions are suboptimal.
2277 * At least, we could easily detect this problem and run
2278 * cf_opt() (or even combo) again :-(
2280 if (node->type.tv == tarval_reachable)
2283 if (get_irn_mode(sel) == mode_b) {
2285 if (pnc == pn_Cond_true) {
2286 if (selector->type.tv == tarval_b_false) {
2287 node->type.tv = tarval_unreachable;
2288 } else if (selector->type.tv == tarval_b_true) {
2289 node->type.tv = tarval_reachable;
2290 } else if (selector->type.tv == tarval_bottom) {
2291 node->type.tv = tarval_reachable;
2293 assert(selector->type.tv == tarval_top);
2295 /* any condition based on Top is "!=" */
2296 node->type.tv = tarval_unreachable;
2298 node->type.tv = tarval_unreachable;
2302 assert(pnc == pn_Cond_false);
2304 if (selector->type.tv == tarval_b_false) {
2305 node->type.tv = tarval_reachable;
2306 } else if (selector->type.tv == tarval_b_true) {
2307 node->type.tv = tarval_unreachable;
2308 } else if (selector->type.tv == tarval_bottom) {
2309 node->type.tv = tarval_reachable;
2311 assert(selector->type.tv == tarval_top);
2313 /* any condition based on Top is "!=" */
2314 node->type.tv = tarval_reachable;
2316 node->type.tv = tarval_unreachable;
2322 if (selector->type.tv == tarval_bottom) {
2323 node->type.tv = tarval_reachable;
2324 } else if (selector->type.tv == tarval_top) {
2326 if (pnc == get_Cond_defaultProj(cond)) {
2327 /* a switch based of Top is always "default" */
2328 node->type.tv = tarval_reachable;
2331 node->type.tv = tarval_unreachable;
2333 long value = get_tarval_long(selector->type.tv);
2334 if (pnc == get_Cond_defaultProj(cond)) {
2335 /* default switch, have to check ALL other cases */
2338 for (i = get_irn_n_outs(cond) - 1; i >= 0; --i) {
2339 ir_node *succ = get_irn_out(cond, i);
2343 if (value == get_Proj_proj(succ)) {
2344 /* we found a match, will NOT take the default case */
2345 node->type.tv = tarval_unreachable;
2349 /* all cases checked, no match, will take default case */
2350 node->type.tv = tarval_reachable;
2353 node->type.tv = value == pnc ? tarval_reachable : tarval_unreachable;
2357 } /* compute_Proj_Cond */
2360 * (Re-)compute the type for a Proj-Node.
2362 * @param node the node
2364 static void compute_Proj(node_t *node) {
2365 ir_node *proj = node->node;
2366 ir_mode *mode = get_irn_mode(proj);
2367 node_t *block = get_irn_node(get_nodes_block(skip_Proj(proj)));
2368 ir_node *pred = get_Proj_pred(proj);
2370 if (block->type.tv == tarval_unreachable) {
2371 /* a Proj in a unreachable Block stay Top */
2372 node->type.tv = tarval_top;
2375 if (get_irn_node(pred)->type.tv == tarval_top && !is_Cond(pred)) {
2376 /* if the predecessor is Top, its Proj follow */
2377 node->type.tv = tarval_top;
2381 if (mode == mode_M) {
2382 /* mode M is always bottom */
2383 node->type.tv = tarval_bottom;
2386 if (mode != mode_X) {
2388 compute_Proj_Cmp(node, pred);
2390 default_compute(node);
2393 /* handle mode_X nodes */
2395 switch (get_irn_opcode(pred)) {
2397 /* the Proj_X from the Start is always reachable.
2398 However this is already handled at the top. */
2399 node->type.tv = tarval_reachable;
2402 compute_Proj_Cond(node, pred);
2405 default_compute(node);
2407 } /* compute_Proj */
2410 * (Re-)compute the type for a Confirm.
2412 * @param node the node
2414 static void compute_Confirm(node_t *node) {
2415 ir_node *confirm = node->node;
2416 node_t *pred = get_irn_node(get_Confirm_value(confirm));
2418 if (get_Confirm_cmp(confirm) == pn_Cmp_Eq) {
2419 node_t *bound = get_irn_node(get_Confirm_bound(confirm));
2421 if (is_con(bound->type)) {
2422 /* is equal to a constant */
2423 node->type = bound->type;
2427 /* a Confirm is a copy OR a Const */
2428 node->type = pred->type;
2429 } /* compute_Confirm */
2432 * (Re-)compute the type for a Max.
2434 * @param node the node
2436 static void compute_Max(node_t *node) {
2437 ir_node *op = node->node;
2438 node_t *l = get_irn_node(get_binop_left(op));
2439 node_t *r = get_irn_node(get_binop_right(op));
2440 lattice_elem_t a = l->type;
2441 lattice_elem_t b = r->type;
2443 if (a.tv == tarval_top || b.tv == tarval_top) {
2444 node->type.tv = tarval_top;
2445 } else if (is_con(a) && is_con(b)) {
2446 /* both nodes are constants, we can probably do something */
2448 /* this case handles SymConsts as well */
2451 ir_mode *mode = get_irn_mode(op);
2452 tarval *tv_min = get_mode_min(mode);
2456 else if (b.tv == tv_min)
2458 else if (is_tarval(a.tv) && is_tarval(b.tv)) {
2459 if (tarval_cmp(a.tv, b.tv) & pn_Cmp_Gt)
2460 node->type.tv = a.tv;
2462 node->type.tv = b.tv;
2464 node->type.tv = tarval_bad;
2467 } else if (r->part == l->part) {
2468 /* both nodes congruent, we can probably do something */
2471 node->type.tv = tarval_bottom;
2476 * (Re-)compute the type for a Min.
2478 * @param node the node
2480 static void compute_Min(node_t *node) {
2481 ir_node *op = node->node;
2482 node_t *l = get_irn_node(get_binop_left(op));
2483 node_t *r = get_irn_node(get_binop_right(op));
2484 lattice_elem_t a = l->type;
2485 lattice_elem_t b = r->type;
2487 if (a.tv == tarval_top || b.tv == tarval_top) {
2488 node->type.tv = tarval_top;
2489 } else if (is_con(a) && is_con(b)) {
2490 /* both nodes are constants, we can probably do something */
2492 /* this case handles SymConsts as well */
2495 ir_mode *mode = get_irn_mode(op);
2496 tarval *tv_max = get_mode_max(mode);
2500 else if (b.tv == tv_max)
2502 else if (is_tarval(a.tv) && is_tarval(b.tv)) {
2503 if (tarval_cmp(a.tv, b.tv) & pn_Cmp_Gt)
2504 node->type.tv = a.tv;
2506 node->type.tv = b.tv;
2508 node->type.tv = tarval_bad;
2511 } else if (r->part == l->part) {
2512 /* both nodes congruent, we can probably do something */
2515 node->type.tv = tarval_bottom;
2520 * (Re-)compute the type for a given node.
2522 * @param node the node
2524 static void compute(node_t *node) {
2525 ir_node *irn = node->node;
2529 * Once a node reaches bottom, the type cannot fall further
2530 * in the lattice and we can stop computation.
2531 * This reduces further checking for ProjM not allowed to raise
2534 if (node->type.tv == tarval_bottom)
2537 if (is_no_Block(irn)) {
2538 /* for pinned nodes, check its control input */
2539 if (get_irn_pinned(skip_Proj(irn)) == op_pin_state_pinned) {
2540 node_t *block = get_irn_node(get_nodes_block(irn));
2542 if (block->type.tv == tarval_unreachable) {
2543 node->type.tv = tarval_top;
2549 func = (compute_func)node->node->op->ops.generic;
2555 * Identity functions: Note that one might thing that identity() is just a
2556 * synonym for equivalent_node(). While this is true, we cannot use it for the algorithm
2557 * here, because it expects that the identity node is one of the inputs, which is NOT
2558 * always true for equivalent_node() which can handle (and does sometimes) DAGs.
2559 * So, we have our own implementation, which copies some parts of equivalent_node()
2563 * Calculates the Identity for Phi nodes
2565 static node_t *identity_Phi(node_t *node) {
2566 ir_node *phi = node->node;
2567 ir_node *block = get_nodes_block(phi);
2568 node_t *n_part = NULL;
2571 for (i = get_Phi_n_preds(phi) - 1; i >= 0; --i) {
2572 node_t *pred_X = get_irn_node(get_Block_cfgpred(block, i));
2574 if (pred_X->type.tv == tarval_reachable) {
2575 node_t *pred = get_irn_node(get_Phi_pred(phi, i));
2579 else if (n_part->part != pred->part) {
2580 /* incongruent inputs, not a follower */
2585 /* if n_part is NULL here, all inputs path are dead, the Phi computes
2586 * tarval_top, is in the TOP partition and should NOT being split! */
2587 assert(n_part != NULL);
2589 } /* identity_Phi */
2592 * Calculates the Identity for commutative 0 neutral nodes.
2594 static node_t *identity_comm_zero_binop(node_t *node) {
2595 ir_node *op = node->node;
2596 node_t *a = get_irn_node(get_binop_left(op));
2597 node_t *b = get_irn_node(get_binop_right(op));
2598 ir_mode *mode = get_irn_mode(op);
2601 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
2602 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
2605 /* node: no input should be tarval_top, else the binop would be also
2606 * Top and not being split. */
2607 zero = get_mode_null(mode);
2608 if (a->type.tv == zero)
2610 if (b->type.tv == zero)
2613 } /* identity_comm_zero_binop */
2616 * Calculates the Identity for Shift nodes.
2618 static node_t *identity_shift(node_t *node) {
2619 ir_node *op = node->node;
2620 node_t *b = get_irn_node(get_binop_right(op));
2621 ir_mode *mode = get_irn_mode(b->node);
2624 /* node: no input should be tarval_top, else the binop would be also
2625 * Top and not being split. */
2626 zero = get_mode_null(mode);
2627 if (b->type.tv == zero)
2628 return get_irn_node(get_binop_left(op));
2630 } /* identity_shift */
2633 * Calculates the Identity for Mul nodes.
2635 static node_t *identity_Mul(node_t *node) {
2636 ir_node *op = node->node;
2637 node_t *a = get_irn_node(get_Mul_left(op));
2638 node_t *b = get_irn_node(get_Mul_right(op));
2639 ir_mode *mode = get_irn_mode(op);
2642 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
2643 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
2646 /* node: no input should be tarval_top, else the binop would be also
2647 * Top and not being split. */
2648 one = get_mode_one(mode);
2649 if (a->type.tv == one)
2651 if (b->type.tv == one)
2654 } /* identity_Mul */
2657 * Calculates the Identity for Sub nodes.
2659 static node_t *identity_Sub(node_t *node) {
2660 ir_node *sub = node->node;
2661 node_t *b = get_irn_node(get_Sub_right(sub));
2662 ir_mode *mode = get_irn_mode(sub);
2664 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
2665 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
2668 /* node: no input should be tarval_top, else the binop would be also
2669 * Top and not being split. */
2670 if (b->type.tv == get_mode_null(mode))
2671 return get_irn_node(get_Sub_left(sub));
2673 } /* identity_Sub */
2676 * Calculates the Identity for And nodes.
2678 static node_t *identity_And(node_t *node) {
2679 ir_node *and = node->node;
2680 node_t *a = get_irn_node(get_And_left(and));
2681 node_t *b = get_irn_node(get_And_right(and));
2682 tarval *neutral = get_mode_all_one(get_irn_mode(and));
2684 /* node: no input should be tarval_top, else the And would be also
2685 * Top and not being split. */
2686 if (a->type.tv == neutral)
2688 if (b->type.tv == neutral)
2691 } /* identity_And */
2694 * Calculates the Identity for Confirm nodes.
2696 static node_t *identity_Confirm(node_t *node) {
2697 ir_node *confirm = node->node;
2699 /* a Confirm is always a Copy */
2700 return get_irn_node(get_Confirm_value(confirm));
2701 } /* identity_Confirm */
2704 * Calculates the Identity for Mux nodes.
2706 static node_t *identity_Mux(node_t *node) {
2707 ir_node *mux = node->node;
2708 node_t *t = get_irn_node(get_Mux_true(mux));
2709 node_t *f = get_irn_node(get_Mux_false(mux));
2712 if (t->part == f->part)
2715 /* for now, the 1-input identity is not supported */
2717 sel = get_irn_node(get_Mux_sel(mux));
2719 /* Mux sel input is mode_b, so it is always a tarval */
2720 if (sel->type.tv == tarval_b_true)
2722 if (sel->type.tv == tarval_b_false)
2726 } /* identity_Mux */
2729 * Calculates the Identity for Min nodes.
2731 static node_t *identity_Min(node_t *node) {
2732 ir_node *op = node->node;
2733 node_t *a = get_irn_node(get_binop_left(op));
2734 node_t *b = get_irn_node(get_binop_right(op));
2735 ir_mode *mode = get_irn_mode(op);
2738 if (a->part == b->part) {
2739 /* leader of multiple predecessors */
2743 /* works even with NaN */
2744 tv_max = get_mode_max(mode);
2745 if (a->type.tv == tv_max)
2747 if (b->type.tv == tv_max)
2750 } /* identity_Min */
2753 * Calculates the Identity for Max nodes.
2755 static node_t *identity_Max(node_t *node) {
2756 ir_node *op = node->node;
2757 node_t *a = get_irn_node(get_binop_left(op));
2758 node_t *b = get_irn_node(get_binop_right(op));
2759 ir_mode *mode = get_irn_mode(op);
2762 if (a->part == b->part) {
2763 /* leader of multiple predecessors */
2767 /* works even with NaN */
2768 tv_min = get_mode_min(mode);
2769 if (a->type.tv == tv_min)
2771 if (b->type.tv == tv_min)
2774 } /* identity_Max */
2777 * Calculates the Identity for nodes.
2779 static node_t *identity(node_t *node) {
2780 ir_node *irn = node->node;
2782 switch (get_irn_opcode(irn)) {
2784 return identity_Phi(node);
2786 return identity_Mul(node);
2790 return identity_comm_zero_binop(node);
2795 return identity_shift(node);
2797 return identity_And(node);
2799 return identity_Sub(node);
2801 return identity_Confirm(node);
2803 return identity_Mux(node);
2805 return identity_Min(node);
2807 return identity_Max(node);
2814 * Node follower is a (new) follower of leader, segregate Leader
2817 static void segregate_def_use_chain_1(const ir_node *follower, node_t *leader) {
2818 ir_node *l = leader->node;
2819 int j, i, n = get_irn_n_outs(l);
2821 DB((dbg, LEVEL_2, "%+F is a follower of %+F\n", follower, leader->node));
2822 /* The leader edges must remain sorted, but follower edges can
2824 for (i = leader->n_followers + 1; i <= n; ++i) {
2825 if (l->out[i].use == follower) {
2826 ir_def_use_edge t = l->out[i];
2828 for (j = i - 1; j >= leader->n_followers + 1; --j)
2829 l->out[j + 1] = l->out[j];
2830 ++leader->n_followers;
2831 l->out[leader->n_followers] = t;
2835 } /* segregate_def_use_chain_1 */
2838 * Node follower is a (new) follower segregate its Leader
2841 * @param follower the follower IR node
2843 static void segregate_def_use_chain(const ir_node *follower) {
2846 for (i = get_irn_arity(follower) - 1; i >= 0; --i) {
2847 node_t *pred = get_irn_node(get_irn_n(follower, i));
2849 segregate_def_use_chain_1(follower, pred);
2851 } /* segregate_def_use_chain */
2854 * Propagate constant evaluation.
2856 * @param env the environment
2858 static void propagate(environment_t *env) {
2861 lattice_elem_t old_type;
2863 unsigned n_fallen, old_type_was_T_or_C;
2866 while (env->cprop != NULL) {
2867 void *oldopcode = NULL;
2869 /* remove the first partition X from cprop */
2872 env->cprop = X->cprop_next;
2874 old_type_was_T_or_C = X->type_is_T_or_C;
2876 DB((dbg, LEVEL_2, "Propagate type on part%d\n", X->nr));
2880 int cprop_empty = list_empty(&X->cprop);
2881 int cprop_X_empty = list_empty(&X->cprop_X);
2883 if (cprop_empty && cprop_X_empty) {
2884 /* both cprop lists are empty */
2888 /* remove the first Node x from X.cprop */
2890 /* Get a node from the cprop_X list only if
2891 * all data nodes are processed.
2892 * This ensures, that all inputs of the Cond
2893 * predecessor are processed if its type is still Top.
2895 x = list_entry(X->cprop_X.next, node_t, cprop_list);
2897 x = list_entry(X->cprop.next, node_t, cprop_list);
2900 //assert(x->part == X);
2901 list_del(&x->cprop_list);
2904 if (x->is_follower && identity(x) == x) {
2905 /* check the opcode first */
2906 if (oldopcode == NULL) {
2907 oldopcode = lambda_opcode(get_first_node(X), env);
2909 if (oldopcode != lambda_opcode(x, env)) {
2910 if (x->on_fallen == 0) {
2911 /* different opcode -> x falls out of this partition */
2916 DB((dbg, LEVEL_2, "Add node %+F to fallen\n", x->node));
2920 /* x will make the follower -> leader transition */
2921 follower_to_leader(x);
2924 /* compute a new type for x */
2926 DB((dbg, LEVEL_3, "computing type of %+F\n", x->node));
2928 if (x->type.tv != old_type.tv) {
2929 DB((dbg, LEVEL_2, "node %+F has changed type from %+F to %+F\n", x->node, old_type, x->type));
2930 verify_type(old_type, x);
2932 if (x->on_fallen == 0) {
2933 /* Add x to fallen. Nodes might fall from T -> const -> _|_, so check that they are
2934 not already on the list. */
2939 DB((dbg, LEVEL_2, "Add node %+F to fallen\n", x->node));
2941 for (i = get_irn_n_outs(x->node) - 1; i >= 0; --i) {
2942 ir_node *succ = get_irn_out(x->node, i);
2943 node_t *y = get_irn_node(succ);
2945 /* Add y to y.partition.cprop. */
2946 add_to_cprop(y, env);
2951 if (n_fallen > 0 && n_fallen != X->n_leader) {
2952 DB((dbg, LEVEL_2, "Splitting part%d by fallen\n", X->nr));
2953 Y = split(&X, fallen, env);
2955 * We have split out fallen node. The type of the result
2956 * partition is NOT set yet.
2958 Y->type_is_T_or_C = 0;
2962 /* remove the flags from the fallen list */
2963 for (x = fallen; x != NULL; x = x->next)
2966 if (old_type_was_T_or_C) {
2969 /* check if some nodes will make the leader -> follower transition */
2970 list_for_each_entry_safe(node_t, y, tmp, &Y->Leader, node_list) {
2971 if (y->type.tv != tarval_top && ! is_con(y->type)) {
2972 node_t *eq_node = identity(y);
2974 if (eq_node != y && eq_node->part == y->part) {
2975 DB((dbg, LEVEL_2, "Node %+F is a follower of %+F\n", y->node, eq_node->node));
2976 /* move to Follower */
2978 list_del(&y->node_list);
2979 list_add_tail(&y->node_list, &Y->Follower);
2982 segregate_def_use_chain(y->node);
2992 * Get the leader for a given node from its congruence class.
2994 * @param irn the node
2996 static ir_node *get_leader(node_t *node) {
2997 partition_t *part = node->part;
2999 if (part->n_leader > 1 || node->is_follower) {
3000 if (node->is_follower) {
3001 DB((dbg, LEVEL_2, "Replacing follower %+F\n", node->node));
3004 DB((dbg, LEVEL_2, "Found congruence class for %+F\n", node->node));
3006 return get_first_node(part)->node;
3012 * Returns non-zero if a mode_T node has only one reachable output.
3014 static int only_one_reachable_proj(ir_node *n) {
3017 for (i = get_irn_n_outs(n) - 1; i >= 0; --i) {
3018 ir_node *proj = get_irn_out(n, i);
3021 /* skip non-control flow Proj's */
3022 if (get_irn_mode(proj) != mode_X)
3025 node = get_irn_node(proj);
3026 if (node->type.tv == tarval_reachable) {
3032 } /* only_one_reachable_proj */
3035 * Return non-zero if the control flow predecessor node pred
3036 * is the only reachable control flow exit of its block.
3038 * @param pred the control flow exit
3040 static int can_exchange(ir_node *pred) {
3043 else if (is_Jmp(pred))
3045 else if (get_irn_mode(pred) == mode_T) {
3046 /* if the predecessor block has more than one
3047 reachable outputs we cannot remove the block */
3048 return only_one_reachable_proj(pred);
3051 } /* can_exchange */
3054 * Block Post-Walker, apply the analysis results on control flow by
3055 * shortening Phi's and Block inputs.
3057 static void apply_cf(ir_node *block, void *ctx) {
3058 environment_t *env = ctx;
3059 node_t *node = get_irn_node(block);
3061 ir_node **ins, **in_X;
3062 ir_node *phi, *next;
3064 n = get_Block_n_cfgpreds(block);
3066 if (node->type.tv == tarval_unreachable) {
3069 for (i = n - 1; i >= 0; --i) {
3070 ir_node *pred = get_Block_cfgpred(block, i);
3072 if (! is_Bad(pred)) {
3073 node_t *pred_bl = get_irn_node(get_nodes_block(skip_Proj(pred)));
3075 if (pred_bl->flagged == 0) {
3076 pred_bl->flagged = 3;
3078 if (pred_bl->type.tv == tarval_reachable) {
3080 * We will remove an edge from block to its pred.
3081 * This might leave the pred block as an endless loop
3083 if (! is_backedge(block, i))
3084 keep_alive(pred_bl->node);
3090 /* the EndBlock is always reachable even if the analysis
3091 finds out the opposite :-) */
3092 if (block != get_irg_end_block(current_ir_graph)) {
3093 /* mark dead blocks */
3094 set_Block_dead(block);
3095 DB((dbg, LEVEL_1, "Removing dead %+F\n", block));
3097 /* the endblock is unreachable */
3098 set_irn_in(block, 0, NULL);
3104 /* only one predecessor combine */
3105 ir_node *pred = skip_Proj(get_Block_cfgpred(block, 0));
3107 if (can_exchange(pred)) {
3108 ir_node *new_block = get_nodes_block(pred);
3109 DB((dbg, LEVEL_1, "Fuse %+F with %+F\n", block, new_block));
3110 DBG_OPT_COMBO(block, new_block, FS_OPT_COMBO_CF);
3111 exchange(block, new_block);
3112 node->node = new_block;
3118 NEW_ARR_A(ir_node *, in_X, n);
3120 for (i = 0; i < n; ++i) {
3121 ir_node *pred = get_Block_cfgpred(block, i);
3122 node_t *node = get_irn_node(pred);
3124 if (node->type.tv == tarval_reachable) {
3127 DB((dbg, LEVEL_1, "Removing dead input %d from %+F (%+F)\n", i, block, pred));
3128 if (! is_Bad(pred)) {
3129 node_t *pred_bl = get_irn_node(get_nodes_block(skip_Proj(pred)));
3131 if (pred_bl->flagged == 0) {
3132 pred_bl->flagged = 3;
3134 if (pred_bl->type.tv == tarval_reachable) {
3136 * We will remove an edge from block to its pred.
3137 * This might leave the pred block as an endless loop
3139 if (! is_backedge(block, i))
3140 keep_alive(pred_bl->node);
3150 NEW_ARR_A(ir_node *, ins, n);
3151 for (phi = get_Block_phis(block); phi != NULL; phi = next) {
3152 node_t *node = get_irn_node(phi);
3154 next = get_Phi_next(phi);
3155 if (is_tarval(node->type.tv) && tarval_is_constant(node->type.tv)) {
3156 /* this Phi is replaced by a constant */
3157 tarval *tv = node->type.tv;
3158 ir_node *c = new_r_Const(current_ir_graph, block, get_tarval_mode(tv), tv);
3160 set_irn_node(c, node);
3162 DB((dbg, LEVEL_1, "%+F is replaced by %+F\n", phi, c));
3163 DBG_OPT_COMBO(phi, c, FS_OPT_COMBO_CONST);
3168 for (i = 0; i < n; ++i) {
3169 node_t *pred = get_irn_node(get_Block_cfgpred(block, i));
3171 if (pred->type.tv == tarval_reachable) {
3172 ins[j++] = get_Phi_pred(phi, i);
3176 /* this Phi is replaced by a single predecessor */
3177 ir_node *s = ins[0];
3178 node_t *phi_node = get_irn_node(phi);
3181 DB((dbg, LEVEL_1, "%+F is replaced by %+F because of cf change\n", phi, s));
3182 DBG_OPT_COMBO(phi, s, FS_OPT_COMBO_FOLLOWER);
3187 set_irn_in(phi, j, ins);
3195 /* this Block has only one live predecessor */
3196 ir_node *pred = skip_Proj(in_X[0]);
3198 if (can_exchange(pred)) {
3199 ir_node *new_block = get_nodes_block(pred);
3200 DBG_OPT_COMBO(block, new_block, FS_OPT_COMBO_CF);
3201 exchange(block, new_block);
3202 node->node = new_block;
3207 set_irn_in(block, k, in_X);
3212 * Exchange a node by its leader.
3213 * Beware: in rare cases the mode might be wrong here, for instance
3214 * AddP(x, NULL) is a follower of x, but with different mode.
3217 static void exchange_leader(ir_node *irn, ir_node *leader) {
3218 ir_mode *mode = get_irn_mode(irn);
3219 if (mode != get_irn_mode(leader)) {
3220 /* The conv is a no-op, so we are fre to place in
3221 * either in the block of the leader OR in irn's block.
3222 * Probably placing it into leaders block might reduce
3223 * the number of Conv due to CSE. */
3224 ir_node *block = get_nodes_block(leader);
3225 dbg_info *dbg = get_irn_dbg_info(irn);
3227 leader = new_rd_Conv(dbg, current_ir_graph, block, leader, mode);
3229 exchange(irn, leader);
3233 * Post-Walker, apply the analysis results;
3235 static void apply_result(ir_node *irn, void *ctx) {
3236 environment_t *env = ctx;
3237 node_t *node = get_irn_node(irn);
3239 if (is_Block(irn) || is_End(irn) || is_Bad(irn)) {
3240 /* blocks already handled, do not touch the End node */
3242 node_t *block = get_irn_node(get_nodes_block(irn));
3243 ir_mode *mode = get_irn_mode(irn);
3245 if (block->type.tv == tarval_unreachable) {
3246 ir_node *bad = get_irg_bad(current_ir_graph);
3248 /* here, bad might already have a node, but this can be safely ignored
3249 as long as bad has at least ONE valid node */
3250 set_irn_node(bad, node);
3252 DB((dbg, LEVEL_1, "%+F is unreachable\n", irn));
3255 } else if (mode == mode_M && is_Proj(irn)) {
3256 ir_node *pred = get_Proj_pred(irn);
3257 node_t *pnode = get_irn_node(pred);
3259 if (pnode->type.tv == tarval_top) {
3260 /* skip the predecessor */
3261 ir_node *mem = get_memop_mem(pred);
3263 DB((dbg, LEVEL_1, "%+F computes Top, replaced by %+F\n", irn, mem));
3267 } else if (node->type.tv == tarval_top) {
3268 if (mode == mode_T) {
3269 /* Do not kill mode_T nodes, kill their Projs */
3270 } else if (! is_Unknown(irn)) {
3271 /* don't kick away Unknown's, they might be still needed */
3272 ir_node *unk = new_r_Unknown(current_ir_graph, mode);
3274 /* control flow should already be handled at apply_cf() */
3275 assert(mode != mode_X);
3277 /* see comment above */
3278 set_irn_node(unk, node);
3280 DB((dbg, LEVEL_1, "%+F computes Top\n", irn));
3285 else if (get_irn_mode(irn) == mode_X) {
3288 ir_node *cond = get_Proj_pred(irn);
3290 if (is_Cond(cond)) {
3291 if (only_one_reachable_proj(cond)) {
3292 ir_node *jmp = new_r_Jmp(current_ir_graph, block->node);
3293 set_irn_node(jmp, node);
3295 DB((dbg, LEVEL_1, "%+F is replaced by %+F\n", irn, jmp));
3296 DBG_OPT_COMBO(irn, jmp, FS_OPT_COMBO_CF);
3300 node_t *sel = get_irn_node(get_Cond_selector(cond));
3301 tarval *tv = sel->type.tv;
3303 if (is_tarval(tv) && tarval_is_constant(tv)) {
3304 /* The selector is a constant, but more
3305 * than one output is active: An unoptimized
3313 /* normal data node */
3314 if (is_tarval(node->type.tv) && tarval_is_constant(node->type.tv)) {
3315 tarval *tv = node->type.tv;
3318 * Beware: never replace mode_T nodes by constants. Currently we must mark
3319 * mode_T nodes with constants, but do NOT replace them.
3321 if (! is_Const(irn) && get_irn_mode(irn) != mode_T) {
3322 /* can be replaced by a constant */
3323 ir_node *c = new_r_Const(current_ir_graph, block->node, get_tarval_mode(tv), tv);
3324 set_irn_node(c, node);
3326 DB((dbg, LEVEL_1, "%+F is replaced by %+F\n", irn, c));
3327 DBG_OPT_COMBO(irn, c, FS_OPT_COMBO_CONST);
3328 exchange_leader(irn, c);
3331 } else if (is_entity(node->type.sym.entity_p)) {
3332 if (! is_SymConst(irn)) {
3333 /* can be replaced by a SymConst */
3334 ir_node *symc = new_r_SymConst(current_ir_graph, block->node, get_irn_mode(irn), node->type.sym, symconst_addr_ent);
3335 set_irn_node(symc, node);
3338 DB((dbg, LEVEL_1, "%+F is replaced by %+F\n", irn, symc));
3339 DBG_OPT_COMBO(irn, symc, FS_OPT_COMBO_CONST);
3340 exchange_leader(irn, symc);
3343 } else if (is_Confirm(irn)) {
3344 /* Confirms are always follower, but do not kill them here */
3346 ir_node *leader = get_leader(node);
3348 if (leader != irn) {
3349 int non_strict_phi = 0;
3352 * Beware: Do not remove Phi(Unknown, ..., x, ..., Unknown)
3353 * as this might create non-strict programs.
3355 if (node->is_follower && is_Phi(irn) && !is_Unknown(leader)) {
3358 for (i = get_Phi_n_preds(irn) - 1; i >= 0; --i) {
3359 ir_node *pred = get_Phi_pred(irn, i);
3361 if (is_Unknown(pred)) {
3367 if (! non_strict_phi) {
3368 DB((dbg, LEVEL_1, "%+F from part%d is replaced by %+F\n", irn, node->part->nr, leader));
3369 if (node->is_follower)
3370 DBG_OPT_COMBO(irn, leader, FS_OPT_COMBO_FOLLOWER);
3372 DBG_OPT_COMBO(irn, leader, FS_OPT_COMBO_CONGRUENT);
3373 exchange_leader(irn, leader);
3380 } /* apply_result */
3383 * Fix the keep-alives by deleting unreachable ones.
3385 static void apply_end(ir_node *end, environment_t *env) {
3386 int i, j, n = get_End_n_keepalives(end);
3390 NEW_ARR_A(ir_node *, in, n);
3392 /* fix the keep alive */
3393 for (i = j = 0; i < n; i++) {
3394 ir_node *ka = get_End_keepalive(end, i);
3395 node_t *node = get_irn_node(ka);
3398 node = get_irn_node(get_nodes_block(ka));
3400 if (node->type.tv != tarval_unreachable && !is_Bad(ka))
3404 set_End_keepalives(end, j, in);
3409 #define SET(code) op_##code->ops.generic = (op_func)compute_##code
3412 * sets the generic functions to compute.
3414 static void set_compute_functions(void) {
3417 /* set the default compute function */
3418 for (i = get_irp_n_opcodes() - 1; i >= 0; --i) {
3419 ir_op *op = get_irp_opcode(i);
3420 op->ops.generic = (op_func)default_compute;
3423 /* set specific functions */
3445 } /* set_compute_functions */
3447 void combo(ir_graph *irg) {
3449 ir_node *initial_bl;
3451 ir_graph *rem = current_ir_graph;
3453 current_ir_graph = irg;
3455 /* register a debug mask */
3456 FIRM_DBG_REGISTER(dbg, "firm.opt.combo");
3458 DB((dbg, LEVEL_1, "Doing COMBO for %+F\n", irg));
3460 obstack_init(&env.obst);
3461 env.worklist = NULL;
3465 #ifdef DEBUG_libfirm
3466 env.dbg_list = NULL;
3468 env.opcode2id_map = new_set(cmp_opcode, iro_Last * 4);
3469 env.type2id_map = pmap_create();
3470 env.end_idx = get_opt_global_cse() ? 0 : -1;
3471 env.lambda_input = 0;
3472 env.commutative = 1;
3476 assure_irg_outs(irg);
3477 assure_cf_loop(irg);
3479 /* we have our own value_of function */
3480 set_value_of_func(get_node_tarval);
3482 set_compute_functions();
3483 DEBUG_ONLY(part_nr = 0);
3485 ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK);
3487 /* create the initial partition and place it on the work list */
3488 env.initial = new_partition(&env);
3489 add_to_worklist(env.initial, &env);
3490 irg_walk_graph(irg, create_initial_partitions, init_block_phis, &env);
3492 /* set the hook: from now, every node has a partition and a type */
3493 DEBUG_ONLY(set_dump_node_vcgattr_hook(dump_partition_hook));
3495 /* all nodes on the initial partition have type Top */
3496 env.initial->type_is_T_or_C = 1;
3498 /* Place the START Node's partition on cprop.
3499 Place the START Node on its local worklist. */
3500 initial_bl = get_irg_start_block(irg);
3501 start = get_irn_node(initial_bl);
3502 add_to_cprop(start, &env);
3506 if (env.worklist != NULL)
3508 } while (env.cprop != NULL || env.worklist != NULL);
3510 dump_all_partitions(&env);
3511 check_all_partitions(&env);
3514 dump_ir_block_graph(irg, "-partition");
3517 /* apply the result */
3518 irg_block_walk_graph(irg, NULL, apply_cf, &env);
3519 /* Kill keep-alives of dead blocks: this speeds up apply_result()
3520 * and fixes assertion because dead cf to dead blocks is NOT removed by
3522 apply_end(get_irg_end(irg), &env);
3523 irg_walk_graph(irg, NULL, apply_result, &env);
3526 DB((dbg, LEVEL_1, "Unoptimized Control Flow left"));
3530 /* control flow might changed */
3531 set_irg_outs_inconsistent(irg);
3532 set_irg_extblk_inconsistent(irg);
3533 set_irg_doms_inconsistent(irg);
3534 set_irg_loopinfo_inconsistent(irg);
3537 ir_free_resources(irg, IR_RESOURCE_IRN_LINK);
3539 /* remove the partition hook */
3540 DEBUG_ONLY(set_dump_node_vcgattr_hook(NULL));
3542 pmap_destroy(env.type2id_map);
3543 del_set(env.opcode2id_map);
3544 obstack_free(&env.obst, NULL);
3546 /* restore value_of() default behavior */
3547 set_value_of_func(NULL);
3548 current_ir_graph = rem;