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"
72 #include "irgraph_t.h"
79 #include "iropt_dbg.h"
83 #include "irnodeset.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 typedef struct node_t node_t;
97 typedef struct partition_t partition_t;
98 typedef struct opcode_key_t opcode_key_t;
99 typedef struct listmap_entry_t listmap_entry_t;
101 /** The type of the compute function. */
102 typedef void (*compute_func)(node_t *node);
107 struct opcode_key_t {
108 ir_opcode code; /**< The Firm opcode. */
109 ir_mode *mode; /**< The mode of all nodes in the partition. */
110 int arity; /**< The arity of this opcode (needed for Phi etc. */
112 long proj; /**< For Proj nodes, its proj number */
113 ir_entity *ent; /**< For Sel Nodes, its entity */
114 int intVal; /**< For Conv/Div Nodes: strict/remainderless */
119 * An entry in the list_map.
121 struct listmap_entry_t {
122 void *id; /**< The id. */
123 node_t *list; /**< The associated list for this id. */
124 listmap_entry_t *next; /**< Link to the next entry in the map. */
127 /** We must map id's to lists. */
128 typedef struct listmap_t {
129 set *map; /**< Map id's to listmap_entry_t's */
130 listmap_entry_t *values; /**< List of all values in the map. */
134 * A lattice element. Because we handle constants and symbolic constants different, we
135 * have to use this union.
146 ir_node *node; /**< The IR-node itself. */
147 list_head node_list; /**< Double-linked list of leader/follower entries. */
148 list_head cprop_list; /**< Double-linked partition.cprop list. */
149 partition_t *part; /**< points to the partition this node belongs to */
150 node_t *next; /**< Next node on local list (partition.touched, fallen). */
151 node_t *race_next; /**< Next node on race list. */
152 lattice_elem_t type; /**< The associated lattice element "type". */
153 int max_user_input; /**< Maximum input number of Def-Use edges. */
154 int next_edge; /**< Index of the next Def-Use edge to use. */
155 int n_followers; /**< Number of Follower in the outs set. */
156 unsigned on_touched:1; /**< Set, if this node is on the partition.touched set. */
157 unsigned on_cprop:1; /**< Set, if this node is on the partition.cprop list. */
158 unsigned on_fallen:1; /**< Set, if this node is on the fallen list. */
159 unsigned is_follower:1; /**< Set, if this node is a follower. */
160 unsigned flagged:2; /**< 2 Bits, set if this node was visited by race 1 or 2. */
164 * A partition containing congruent nodes.
167 list_head Leader; /**< The head of partition Leader node list. */
168 list_head Follower; /**< The head of partition Follower node list. */
169 list_head cprop; /**< The head of partition.cprop list. */
170 list_head cprop_X; /**< The head of partition.cprop (Cond nodes and its Projs) list. */
171 partition_t *wl_next; /**< Next entry in the work list if any. */
172 partition_t *touched_next; /**< Points to the next partition in the touched set. */
173 partition_t *cprop_next; /**< Points to the next partition in the cprop list. */
174 partition_t *split_next; /**< Points to the next partition in the list that must be split by split_by(). */
175 node_t *touched; /**< The partition.touched set of this partition. */
176 unsigned n_leader; /**< Number of entries in this partition.Leader. */
177 unsigned n_touched; /**< Number of entries in the partition.touched. */
178 int max_user_inputs; /**< Maximum number of user inputs of all entries. */
179 unsigned on_worklist:1; /**< Set, if this partition is in the work list. */
180 unsigned on_touched:1; /**< Set, if this partition is on the touched set. */
181 unsigned on_cprop:1; /**< Set, if this partition is on the cprop list. */
182 unsigned type_is_T_or_C:1;/**< Set, if all nodes in this partition have type Top or Constant. */
184 partition_t *dbg_next; /**< Link all partitions for debugging */
185 unsigned nr; /**< A unique number for (what-)mapping, >0. */
189 typedef struct environment_t {
190 struct obstack obst; /**< obstack to allocate data structures. */
191 partition_t *worklist; /**< The work list. */
192 partition_t *cprop; /**< The constant propagation list. */
193 partition_t *touched; /**< the touched set. */
194 partition_t *initial; /**< The initial partition. */
195 set *opcode2id_map; /**< The opcodeMode->id map. */
196 pmap *type2id_map; /**< The type->id map. */
197 ir_node **kept_memory; /**< Array of memory nodes that must be kept. */
198 int end_idx; /**< -1 for local and 0 for global congruences. */
199 int lambda_input; /**< Captured argument for lambda_partition(). */
200 unsigned modified:1; /**< Set, if the graph was modified. */
201 unsigned unopt_cf:1; /**< If set, control flow is not optimized due to Unknown. */
202 /* options driving the optimization */
203 unsigned commutative:1; /**< Set, if commutation nodes should be handled specially. */
204 unsigned opt_unknown:1; /**< Set, if non-strict programs should be optimized. */
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(irn) ((node_t *)get_irn_link(irn))
214 #define set_irn_node(irn, node) set_irn_link(irn, 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: set to either tarval_bad OR tarval_top. */
231 static tarval *tarval_UNKNOWN;
234 static node_t *identity(node_t *node);
236 #ifdef CHECK_PARTITIONS
240 static void check_partition(const partition_t *T) {
244 list_for_each_entry(node_t, node, &T->Leader, node_list) {
245 assert(node->is_follower == 0);
246 assert(node->flagged == 0);
247 assert(node->part == T);
250 assert(n == T->n_leader);
252 list_for_each_entry(node_t, node, &T->Follower, node_list) {
253 assert(node->is_follower == 1);
254 assert(node->flagged == 0);
255 assert(node->part == T);
257 } /* check_partition */
260 * check that all leader nodes in the partition have the same opcode.
262 static void check_opcode(const partition_t *Z) {
267 list_for_each_entry(node_t, node, &Z->Leader, node_list) {
268 ir_node *irn = node->node;
271 key.code = get_irn_opcode(irn);
272 key.mode = get_irn_mode(irn);
273 key.arity = get_irn_arity(irn);
277 switch (get_irn_opcode(irn)) {
279 key.u.proj = get_Proj_proj(irn);
282 key.u.ent = get_Sel_entity(irn);
285 key.u.intVal = get_Conv_strict(irn);
288 key.u.intVal = is_Div_remainderless(irn);
295 assert(key.code == get_irn_opcode(irn));
296 assert(key.mode == get_irn_mode(irn));
297 assert(key.arity == get_irn_arity(irn));
299 switch (get_irn_opcode(irn)) {
301 assert(key.u.proj == get_Proj_proj(irn));
304 assert(key.u.ent == get_Sel_entity(irn));
307 assert(key.u.intVal == get_Conv_strict(irn));
310 assert(key.u.intVal == is_Div_remainderless(irn));
319 static void check_all_partitions(environment_t *env) {
324 for (P = env->dbg_list; P != NULL; P = P->dbg_next) {
326 if (! P->type_is_T_or_C)
328 list_for_each_entry(node_t, node, &P->Follower, node_list) {
329 node_t *leader = identity(node);
331 assert(leader != node && leader->part == node->part);
340 static void do_check_list(const node_t *list, int ofs, const partition_t *Z) {
343 #define NEXT(e) *((const node_t **)((char *)(e) + (ofs)))
344 for (e = list; e != NULL; e = NEXT(e)) {
345 assert(e->part == Z);
348 } /* ido_check_list */
351 * Check a local list.
353 static void check_list(const node_t *list, const partition_t *Z) {
354 do_check_list(list, offsetof(node_t, next), Z);
358 #define check_partition(T)
359 #define check_list(list, Z)
360 #define check_all_partitions(env)
361 #endif /* CHECK_PARTITIONS */
364 static inline lattice_elem_t get_partition_type(const partition_t *X);
367 * Dump partition to output.
369 static void dump_partition(const char *msg, const partition_t *part) {
372 lattice_elem_t type = get_partition_type(part);
374 DB((dbg, LEVEL_2, "%s part%u%s (%u, %+F) {\n ",
375 msg, part->nr, part->type_is_T_or_C ? "*" : "",
376 part->n_leader, type));
377 list_for_each_entry(node_t, node, &part->Leader, node_list) {
378 DB((dbg, LEVEL_2, "%s%+F", first ? "" : ", ", node->node));
381 if (! list_empty(&part->Follower)) {
382 DB((dbg, LEVEL_2, "\n---\n "));
384 list_for_each_entry(node_t, node, &part->Follower, node_list) {
385 DB((dbg, LEVEL_2, "%s%+F", first ? "" : ", ", node->node));
389 DB((dbg, LEVEL_2, "\n}\n"));
390 } /* dump_partition */
395 static void do_dump_list(const char *msg, const node_t *node, int ofs) {
399 #define GET_LINK(p, ofs) *((const node_t **)((char *)(p) + (ofs)))
401 DB((dbg, LEVEL_3, "%s = {\n ", msg));
402 for (p = node; p != NULL; p = GET_LINK(p, ofs)) {
403 DB((dbg, LEVEL_3, "%s%+F", first ? "" : ", ", p->node));
406 DB((dbg, LEVEL_3, "\n}\n"));
414 static void dump_race_list(const char *msg, const node_t *list) {
415 do_dump_list(msg, list, offsetof(node_t, race_next));
416 } /* dump_race_list */
419 * Dumps a local list.
421 static void dump_list(const char *msg, const node_t *list) {
422 do_dump_list(msg, list, offsetof(node_t, next));
426 * Dump all partitions.
428 static void dump_all_partitions(const environment_t *env) {
429 const partition_t *P;
431 DB((dbg, LEVEL_2, "All partitions\n===============\n"));
432 for (P = env->dbg_list; P != NULL; P = P->dbg_next)
433 dump_partition("", P);
434 } /* dump_all_partitions */
439 static void dump_split_list(const partition_t *list) {
440 const partition_t *p;
442 DB((dbg, LEVEL_2, "Split by %s produced = {\n", what_reason));
443 for (p = list; p != NULL; p = p->split_next)
444 DB((dbg, LEVEL_2, "part%u, ", p->nr));
445 DB((dbg, LEVEL_2, "\n}\n"));
446 } /* dump_split_list */
449 * Dump partition and type for a node.
451 static int dump_partition_hook(FILE *F, ir_node *n, ir_node *local) {
452 ir_node *irn = local != NULL ? local : n;
453 node_t *node = get_irn_node(irn);
455 ir_fprintf(F, "info2 : \"partition %u type %+F\"\n", node->part->nr, node->type);
457 } /* dump_partition_hook */
460 #define dump_partition(msg, part)
461 #define dump_race_list(msg, list)
462 #define dump_list(msg, list)
463 #define dump_all_partitions(env)
464 #define dump_split_list(list)
467 #if defined(VERIFY_MONOTONE) && defined (DEBUG_libfirm)
469 * Verify that a type transition is monotone
471 static void verify_type(const lattice_elem_t old_type, node_t *node) {
472 if (old_type.tv == node->type.tv) {
476 if (old_type.tv == tarval_top) {
477 /* from Top down-to is always allowed */
480 if (node->type.tv == tarval_bottom || node->type.tv == tarval_reachable) {
484 panic("combo: wrong translation from %+F to %+F on node %+F", old_type, node->type, node->node);
488 #define verify_type(old_type, node)
492 * Compare two pointer values of a listmap.
494 static int listmap_cmp_ptr(const void *elt, const void *key, size_t size) {
495 const listmap_entry_t *e1 = elt;
496 const listmap_entry_t *e2 = key;
499 return e1->id != e2->id;
500 } /* listmap_cmp_ptr */
503 * Initializes a listmap.
505 * @param map the listmap
507 static void listmap_init(listmap_t *map) {
508 map->map = new_set(listmap_cmp_ptr, 16);
513 * Terminates a listmap.
515 * @param map the listmap
517 static void listmap_term(listmap_t *map) {
522 * Return the associated listmap entry for a given id.
524 * @param map the listmap
525 * @param id the id to search for
527 * @return the associated listmap entry for the given id
529 static listmap_entry_t *listmap_find(listmap_t *map, void *id) {
530 listmap_entry_t key, *entry;
535 entry = set_insert(map->map, &key, sizeof(key), HASH_PTR(id));
537 if (entry->list == NULL) {
538 /* a new entry, put into the list */
539 entry->next = map->values;
546 * Calculate the hash value for an opcode map entry.
548 * @param entry an opcode map entry
550 * @return a hash value for the given opcode map entry
552 static unsigned opcode_hash(const opcode_key_t *entry) {
553 return (entry->mode - (ir_mode *)0) * 9 + entry->code + entry->u.proj * 3 + HASH_PTR(entry->u.ent) + entry->arity;
557 * Compare two entries in the opcode map.
559 static int cmp_opcode(const void *elt, const void *key, size_t size) {
560 const opcode_key_t *o1 = elt;
561 const opcode_key_t *o2 = key;
564 return o1->code != o2->code || o1->mode != o2->mode ||
565 o1->arity != o2->arity ||
566 o1->u.proj != o2->u.proj || o1->u.ent != o2->u.ent ||
567 o1->u.intVal != o2->u.intVal;
571 * Compare two Def-Use edges for input position.
573 static int cmp_def_use_edge(const void *a, const void *b) {
574 const ir_def_use_edge *ea = a;
575 const ir_def_use_edge *eb = b;
577 /* no overrun, because range is [-1, MAXINT] */
578 return ea->pos - eb->pos;
579 } /* cmp_def_use_edge */
582 * We need the Def-Use edges sorted.
584 static void sort_irn_outs(node_t *node) {
585 ir_node *irn = node->node;
586 int n_outs = get_irn_n_outs(irn);
589 qsort(&irn->out[1], n_outs, sizeof(irn->out[0]), cmp_def_use_edge);
591 node->max_user_input = irn->out[n_outs].pos;
592 } /* sort_irn_outs */
595 * Return the type of a node.
597 * @param irn an IR-node
599 * @return the associated type of this node
601 static inline lattice_elem_t get_node_type(const ir_node *irn) {
602 return get_irn_node(irn)->type;
603 } /* get_node_type */
606 * Return the tarval of a node.
608 * @param irn an IR-node
610 * @return the associated type of this node
612 static inline tarval *get_node_tarval(const ir_node *irn) {
613 lattice_elem_t type = get_node_type(irn);
615 if (is_tarval(type.tv))
617 return tarval_bottom;
618 } /* get_node_type */
621 * Add a partition to the worklist.
623 static inline void add_to_worklist(partition_t *X, environment_t *env) {
624 assert(X->on_worklist == 0);
625 DB((dbg, LEVEL_2, "Adding part%d to worklist\n", X->nr));
626 X->wl_next = env->worklist;
629 } /* add_to_worklist */
632 * Create a new empty partition.
634 * @param env the environment
636 * @return a newly allocated partition
638 static inline partition_t *new_partition(environment_t *env) {
639 partition_t *part = obstack_alloc(&env->obst, sizeof(*part));
641 INIT_LIST_HEAD(&part->Leader);
642 INIT_LIST_HEAD(&part->Follower);
643 INIT_LIST_HEAD(&part->cprop);
644 INIT_LIST_HEAD(&part->cprop_X);
645 part->wl_next = NULL;
646 part->touched_next = NULL;
647 part->cprop_next = NULL;
648 part->split_next = NULL;
649 part->touched = NULL;
652 part->max_user_inputs = 0;
653 part->on_worklist = 0;
654 part->on_touched = 0;
656 part->type_is_T_or_C = 0;
658 part->dbg_next = env->dbg_list;
659 env->dbg_list = part;
660 part->nr = part_nr++;
664 } /* new_partition */
667 * Get the first node from a partition.
669 static inline node_t *get_first_node(const partition_t *X) {
670 return list_entry(X->Leader.next, node_t, node_list);
671 } /* get_first_node */
674 * Return the type of a partition (assuming partition is non-empty and
675 * all elements have the same type).
677 * @param X a partition
679 * @return the type of the first element of the partition
681 static inline lattice_elem_t get_partition_type(const partition_t *X) {
682 const node_t *first = get_first_node(X);
684 } /* get_partition_type */
687 * Creates a partition node for the given IR-node and place it
688 * into the given partition.
690 * @param irn an IR-node
691 * @param part a partition to place the node in
692 * @param env the environment
694 * @return the created node
696 static node_t *create_partition_node(ir_node *irn, partition_t *part, environment_t *env) {
697 /* create a partition node and place it in the partition */
698 node_t *node = obstack_alloc(&env->obst, sizeof(*node));
700 INIT_LIST_HEAD(&node->node_list);
701 INIT_LIST_HEAD(&node->cprop_list);
705 node->race_next = NULL;
706 node->type.tv = tarval_top;
707 node->max_user_input = 0;
709 node->n_followers = 0;
710 node->on_touched = 0;
713 node->is_follower = 0;
715 set_irn_node(irn, node);
717 list_add_tail(&node->node_list, &part->Leader);
721 } /* create_partition_node */
724 * Pre-Walker, initialize all Nodes' type to U or top and place
725 * all nodes into the TOP partition.
727 static void create_initial_partitions(ir_node *irn, void *ctx) {
728 environment_t *env = ctx;
729 partition_t *part = env->initial;
732 node = create_partition_node(irn, part, env);
734 if (node->max_user_input > part->max_user_inputs)
735 part->max_user_inputs = node->max_user_input;
738 set_Block_phis(irn, NULL);
740 } /* create_initial_partitions */
743 * Post-Walker, collect all Block-Phi lists, set Cond.
745 static void init_block_phis(ir_node *irn, void *ctx) {
749 add_Block_phi(get_nodes_block(irn), irn);
751 } /* init_block_phis */
754 * Add a node to the entry.partition.touched set and
755 * node->partition to the touched set if not already there.
758 * @param env the environment
760 static inline void add_to_touched(node_t *y, environment_t *env) {
761 if (y->on_touched == 0) {
762 partition_t *part = y->part;
764 y->next = part->touched;
769 if (part->on_touched == 0) {
770 part->touched_next = env->touched;
772 part->on_touched = 1;
775 check_list(part->touched, part);
777 } /* add_to_touched */
780 * Place a node on the cprop list.
783 * @param env the environment
785 static void add_to_cprop(node_t *y, environment_t *env) {
788 /* Add y to y.partition.cprop. */
789 if (y->on_cprop == 0) {
790 partition_t *Y = y->part;
791 ir_node *irn = y->node;
793 /* place Conds and all its Projs on the cprop_X list */
794 if (is_Cond(skip_Proj(irn)))
795 list_add_tail(&y->cprop_list, &Y->cprop_X);
797 list_add_tail(&y->cprop_list, &Y->cprop);
800 DB((dbg, LEVEL_3, "Add %+F to part%u.cprop\n", y->node, Y->nr));
802 /* place its partition on the cprop list */
803 if (Y->on_cprop == 0) {
804 Y->cprop_next = env->cprop;
810 if (get_irn_mode(irn) == mode_T) {
811 /* mode_T nodes always produce tarval_bottom, so we must explicitly
812 add it's Proj's to get constant evaluation to work */
815 for (i = get_irn_n_outs(irn) - 1; i >= 0; --i) {
816 node_t *proj = get_irn_node(get_irn_out(irn, i));
818 add_to_cprop(proj, env);
820 } else if (is_Block(irn)) {
821 /* Due to the way we handle Phi's, we must place all Phis of a block on the list
822 * if someone placed the block. The Block is only placed if the reachability
823 * changes, and this must be re-evaluated in compute_Phi(). */
825 for (phi = get_Block_phis(irn); phi != NULL; phi = get_Phi_next(phi)) {
826 node_t *p = get_irn_node(phi);
827 add_to_cprop(p, env);
833 * Update the worklist: If Z is on worklist then add Z' to worklist.
834 * Else add the smaller of Z and Z' to worklist.
836 * @param Z the Z partition
837 * @param Z_prime the Z' partition, a previous part of Z
838 * @param env the environment
840 static void update_worklist(partition_t *Z, partition_t *Z_prime, environment_t *env) {
841 if (Z->on_worklist || Z_prime->n_leader < Z->n_leader) {
842 add_to_worklist(Z_prime, env);
844 add_to_worklist(Z, env);
846 } /* update_worklist */
849 * Make all inputs to x no longer be F.def_use edges.
853 static void move_edges_to_leader(node_t *x) {
854 ir_node *irn = x->node;
857 for (i = get_irn_arity(irn) - 1; i >= 0; --i) {
858 node_t *pred = get_irn_node(get_irn_n(irn, i));
863 n = get_irn_n_outs(p);
864 for (j = 1; j <= pred->n_followers; ++j) {
865 if (p->out[j].pos == i && p->out[j].use == irn) {
866 /* found a follower edge to x, move it to the Leader */
867 ir_def_use_edge edge = p->out[j];
869 /* remove this edge from the Follower set */
870 p->out[j] = p->out[pred->n_followers];
873 /* sort it into the leader set */
874 for (k = pred->n_followers + 2; k <= n; ++k) {
875 if (p->out[k].pos >= edge.pos)
877 p->out[k - 1] = p->out[k];
879 /* place the new edge here */
880 p->out[k - 1] = edge;
882 /* edge found and moved */
887 } /* move_edges_to_leader */
890 * Split a partition that has NO followers by a local list.
892 * @param Z partition to split
893 * @param g a (non-empty) node list
894 * @param env the environment
896 * @return a new partition containing the nodes of g
898 static partition_t *split_no_followers(partition_t *Z, node_t *g, environment_t *env) {
899 partition_t *Z_prime;
904 dump_partition("Splitting ", Z);
905 dump_list("by list ", g);
909 /* Remove g from Z. */
910 for (node = g; node != NULL; node = node->next) {
911 assert(node->part == Z);
912 list_del(&node->node_list);
915 assert(n < Z->n_leader);
918 /* Move g to a new partition, Z'. */
919 Z_prime = new_partition(env);
921 for (node = g; node != NULL; node = node->next) {
922 list_add_tail(&node->node_list, &Z_prime->Leader);
923 node->part = Z_prime;
924 if (node->max_user_input > max_input)
925 max_input = node->max_user_input;
927 Z_prime->max_user_inputs = max_input;
928 Z_prime->n_leader = n;
931 check_partition(Z_prime);
933 /* for now, copy the type info tag, it will be adjusted in split_by(). */
934 Z_prime->type_is_T_or_C = Z->type_is_T_or_C;
936 update_worklist(Z, Z_prime, env);
938 dump_partition("Now ", Z);
939 dump_partition("Created new ", Z_prime);
941 } /* split_no_followers */
944 * Make the Follower -> Leader transition for a node.
948 static void follower_to_leader(node_t *n) {
949 assert(n->is_follower == 1);
951 DB((dbg, LEVEL_2, "%+F make the follower -> leader transition\n", n->node));
953 move_edges_to_leader(n);
954 list_del(&n->node_list);
955 list_add_tail(&n->node_list, &n->part->Leader);
957 } /* follower_to_leader */
960 * The environment for one race step.
962 typedef struct step_env {
963 node_t *initial; /**< The initial node list. */
964 node_t *unwalked; /**< The unwalked node list. */
965 node_t *walked; /**< The walked node list. */
966 int index; /**< Next index of Follower use_def edge. */
967 unsigned side; /**< side number. */
971 * Return non-zero, if a input is a real follower
973 * @param irn the node to check
974 * @param input number of the input
976 static int is_real_follower(const ir_node *irn, int input) {
979 switch (get_irn_opcode(irn)) {
982 /* ignore the Confirm bound input */
988 /* ignore the Mux sel input */
993 /* dead inputs are not follower edges */
994 ir_node *block = get_nodes_block(irn);
995 node_t *pred = get_irn_node(get_Block_cfgpred(block, input));
997 if (pred->type.tv == tarval_unreachable)
1007 /* only a Sub x,0 / Shift x,0 might be a follower */
1014 pred = get_irn_node(get_irn_n(irn, input));
1015 if (is_tarval(pred->type.tv) && tarval_is_null(pred->type.tv))
1019 pred = get_irn_node(get_irn_n(irn, input));
1020 if (is_tarval(pred->type.tv) && tarval_is_one(pred->type.tv))
1024 pred = get_irn_node(get_irn_n(irn, input));
1025 if (is_tarval(pred->type.tv) && tarval_is_all_one(pred->type.tv))
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 X the partition of the list
1413 * @param list the list which contains the nodes that must be evaluated
1414 * @param env the environment
1416 static void collect_commutative_touched(partition_t *X, list_head *list, environment_t *env) {
1421 list_for_each_entry(node_t, x, list, node_list) {
1424 num_edges = get_irn_n_outs(x->node);
1426 x->next_edge = x->n_followers + 1;
1428 /* for all edges in x.L.def_use_{idx} */
1429 while (x->next_edge <= num_edges) {
1430 const ir_def_use_edge *edge = &x->node->out[x->next_edge];
1433 /* check if we have necessary edges */
1443 /* only commutative nodes */
1444 if (!is_op_commutative(get_irn_op(succ)))
1447 y = get_irn_node(succ);
1448 if (is_constant_type(y->type)) {
1449 ir_opcode code = get_irn_opcode(succ);
1450 if (code == iro_Eor)
1451 add_to_cprop(y, env);
1454 /* Partitions of constants should not be split simply because their Nodes have unequal
1455 functions or incongruent inputs. */
1456 if (type_is_neither_top_nor_const(y->type)) {
1457 int other_idx = edge->pos ^ 1;
1458 node_t *other = get_irn_node(get_irn_n(succ, other_idx));
1459 int equal = X == other->part;
1462 * Note: op(a, a) is NOT congruent to op(a, b).
1463 * So, either all touch nodes must have both inputs congruent,
1464 * or not. We decide this by the first occurred node.
1470 if (both_input == equal)
1471 add_to_touched(y, env);
1475 } /* collect_commutative_touched */
1478 * Split the partitions if caused by the first entry on the worklist.
1480 * @param env the environment
1482 static void cause_splits(environment_t *env) {
1483 partition_t *X, *Z, *N;
1486 /* remove the first partition from the worklist */
1488 env->worklist = X->wl_next;
1491 dump_partition("Cause_split: ", X);
1493 if (env->commutative) {
1494 /* handle commutative nodes first */
1496 /* empty the touched set: already done, just clear the list */
1497 env->touched = NULL;
1499 collect_commutative_touched(X, &X->Leader, env);
1500 collect_commutative_touched(X, &X->Follower, env);
1502 for (Z = env->touched; Z != NULL; Z = N) {
1504 node_t *touched = Z->touched;
1505 unsigned n_touched = Z->n_touched;
1507 assert(Z->touched != NULL);
1509 /* beware, split might change Z */
1510 N = Z->touched_next;
1512 /* remove it from the touched set */
1515 /* Empty local Z.touched. */
1516 for (e = touched; e != NULL; e = e->next) {
1517 assert(e->is_follower == 0);
1523 if (0 < n_touched && n_touched < Z->n_leader) {
1524 DB((dbg, LEVEL_2, "Split part%d by touched\n", Z->nr));
1525 split(&Z, touched, env);
1527 assert(n_touched <= Z->n_leader);
1531 /* combine temporary leader and follower list */
1532 for (idx = -1; idx <= X->max_user_inputs; ++idx) {
1533 /* empty the touched set: already done, just clear the list */
1534 env->touched = NULL;
1536 collect_touched(&X->Leader, idx, env);
1537 collect_touched(&X->Follower, idx, env);
1539 for (Z = env->touched; Z != NULL; Z = N) {
1541 node_t *touched = Z->touched;
1542 unsigned n_touched = Z->n_touched;
1544 assert(Z->touched != NULL);
1546 /* beware, split might change Z */
1547 N = Z->touched_next;
1549 /* remove it from the touched set */
1552 /* Empty local Z.touched. */
1553 for (e = touched; e != NULL; e = e->next) {
1554 assert(e->is_follower == 0);
1560 if (0 < n_touched && n_touched < Z->n_leader) {
1561 DB((dbg, LEVEL_2, "Split part%d by touched\n", Z->nr));
1562 split(&Z, touched, env);
1564 assert(n_touched <= Z->n_leader);
1567 } /* cause_splits */
1570 * Implements split_by_what(): Split a partition by characteristics given
1571 * by the what function.
1573 * @param X the partition to split
1574 * @param What a function returning an Id for every node of the partition X
1575 * @param P a list to store the result partitions
1576 * @param env the environment
1580 static partition_t *split_by_what(partition_t *X, what_func What,
1581 partition_t **P, environment_t *env) {
1584 listmap_entry_t *iter;
1587 /* Let map be an empty mapping from the range of What to (local) list of Nodes. */
1589 list_for_each_entry(node_t, x, &X->Leader, node_list) {
1590 void *id = What(x, env);
1591 listmap_entry_t *entry;
1594 /* input not allowed, ignore */
1597 /* Add x to map[What(x)]. */
1598 entry = listmap_find(&map, id);
1599 x->next = entry->list;
1602 /* Let P be a set of Partitions. */
1604 /* for all sets S except one in the range of map do */
1605 for (iter = map.values; iter != NULL; iter = iter->next) {
1606 if (iter->next == NULL) {
1607 /* this is the last entry, ignore */
1612 /* Add SPLIT( X, S ) to P. */
1613 DB((dbg, LEVEL_2, "Split part%d by WHAT = %s\n", X->nr, what_reason));
1614 R = split(&X, S, env);
1624 } /* split_by_what */
1626 /** lambda n.(n.type) */
1627 static void *lambda_type(const node_t *node, environment_t *env) {
1629 return node->type.tv;
1632 /** lambda n.(n.opcode) */
1633 static void *lambda_opcode(const node_t *node, environment_t *env) {
1634 opcode_key_t key, *entry;
1635 ir_node *irn = node->node;
1637 key.code = get_irn_opcode(irn);
1638 key.mode = get_irn_mode(irn);
1639 key.arity = get_irn_arity(irn);
1643 switch (get_irn_opcode(irn)) {
1645 key.u.proj = get_Proj_proj(irn);
1648 key.u.ent = get_Sel_entity(irn);
1651 key.u.intVal = get_Conv_strict(irn);
1654 key.u.intVal = is_Div_remainderless(irn);
1660 entry = set_insert(env->opcode2id_map, &key, sizeof(key), opcode_hash(&key));
1662 } /* lambda_opcode */
1664 /** lambda n.(n[i].partition) */
1665 static void *lambda_partition(const node_t *node, environment_t *env) {
1666 ir_node *skipped = skip_Proj(node->node);
1669 int i = env->lambda_input;
1671 if (i >= get_irn_arity(node->node)) {
1673 * We are outside the allowed range: This can happen even
1674 * if we have split by opcode first: doing so might move Followers
1675 * to Leaders and those will have a different opcode!
1676 * Note that in this case the partition is on the cprop list and will be
1682 /* ignore the "control input" for non-pinned nodes
1683 if we are running in GCSE mode */
1684 if (i < env->end_idx && get_irn_pinned(skipped) != op_pin_state_pinned)
1687 pred = i == -1 ? get_irn_n(skipped, i) : get_irn_n(node->node, i);
1688 p = get_irn_node(pred);
1691 } /* lambda_partition */
1693 /** lambda n.(n[i].partition) for commutative nodes */
1694 static void *lambda_commutative_partition(const node_t *node, environment_t *env) {
1695 ir_node *irn = node->node;
1696 ir_node *skipped = skip_Proj(irn);
1697 ir_node *pred, *left, *right;
1699 partition_t *pl, *pr;
1700 int i = env->lambda_input;
1702 if (i >= get_irn_arity(node->node)) {
1704 * We are outside the allowed range: This can happen even
1705 * if we have split by opcode first: doing so might move Followers
1706 * to Leaders and those will have a different opcode!
1707 * Note that in this case the partition is on the cprop list and will be
1713 /* ignore the "control input" for non-pinned nodes
1714 if we are running in GCSE mode */
1715 if (i < env->end_idx && get_irn_pinned(skipped) != op_pin_state_pinned)
1719 pred = get_irn_n(skipped, i);
1720 p = get_irn_node(pred);
1724 if (is_op_commutative(get_irn_op(irn))) {
1725 /* normalize partition order by returning the "smaller" on input 0,
1726 the "bigger" on input 1. */
1727 left = get_binop_left(irn);
1728 pl = get_irn_node(left)->part;
1729 right = get_binop_right(irn);
1730 pr = get_irn_node(right)->part;
1733 return pl < pr ? pl : pr;
1735 return pl > pr ? pl : pr;
1737 /* a not split out Follower */
1738 pred = get_irn_n(irn, i);
1739 p = get_irn_node(pred);
1743 } /* lambda_commutative_partition */
1746 * Returns true if a type is a constant (and NOT Top
1749 static int is_con(const lattice_elem_t type) {
1750 /* be conservative */
1751 if (is_tarval(type.tv))
1752 return tarval_is_constant(type.tv);
1753 return is_entity(type.sym.entity_p);
1757 * Implements split_by().
1759 * @param X the partition to split
1760 * @param env the environment
1762 static void split_by(partition_t *X, environment_t *env) {
1763 partition_t *I, *P = NULL;
1766 dump_partition("split_by", X);
1768 if (X->n_leader == 1) {
1769 /* we have only one leader, no need to split, just check it's type */
1770 node_t *x = get_first_node(X);
1771 X->type_is_T_or_C = x->type.tv == tarval_top || is_con(x->type);
1775 DEBUG_ONLY(what_reason = "lambda n.(n.type)";)
1776 P = split_by_what(X, lambda_type, &P, env);
1779 /* adjust the type tags, we have split partitions by type */
1780 for (I = P; I != NULL; I = I->split_next) {
1781 node_t *x = get_first_node(I);
1782 I->type_is_T_or_C = x->type.tv == tarval_top || is_con(x->type);
1789 if (Y->n_leader > 1) {
1790 /* we do not want split the TOP or constant partitions */
1791 if (! Y->type_is_T_or_C) {
1792 partition_t *Q = NULL;
1794 DEBUG_ONLY(what_reason = "lambda n.(n.opcode)";)
1795 Q = split_by_what(Y, lambda_opcode, &Q, env);
1802 if (Z->n_leader > 1) {
1803 const node_t *first = get_first_node(Z);
1804 int arity = get_irn_arity(first->node);
1806 what_func what = lambda_partition;
1807 DEBUG_ONLY(char buf[64];)
1809 if (env->commutative && is_op_commutative(get_irn_op(first->node)))
1810 what = lambda_commutative_partition;
1813 * BEWARE: during splitting by input 2 for instance we might
1814 * create new partitions which are different by input 1, so collect
1815 * them and split further.
1817 Z->split_next = NULL;
1820 for (input = arity - 1; input >= -1; --input) {
1822 partition_t *Z_prime = R;
1825 if (Z_prime->n_leader > 1) {
1826 env->lambda_input = input;
1827 DEBUG_ONLY(snprintf(buf, sizeof(buf), "lambda n.(n[%d].partition)", input);)
1828 DEBUG_ONLY(what_reason = buf;)
1829 S = split_by_what(Z_prime, what, &S, env);
1832 Z_prime->split_next = S;
1835 } while (R != NULL);
1840 } while (Q != NULL);
1843 } while (P != NULL);
1847 * (Re-)compute the type for a given node.
1849 * @param node the node
1851 static void default_compute(node_t *node) {
1853 ir_node *irn = node->node;
1855 /* if any of the data inputs have type top, the result is type top */
1856 for (i = get_irn_arity(irn) - 1; i >= 0; --i) {
1857 ir_node *pred = get_irn_n(irn, i);
1858 node_t *p = get_irn_node(pred);
1860 if (p->type.tv == tarval_top) {
1861 node->type.tv = tarval_top;
1866 if (get_irn_mode(node->node) == mode_X)
1867 node->type.tv = tarval_reachable;
1869 node->type.tv = computed_value(irn);
1870 } /* default_compute */
1873 * (Re-)compute the type for a Block node.
1875 * @param node the node
1877 static void compute_Block(node_t *node) {
1879 ir_node *block = node->node;
1881 if (block == get_irg_start_block(current_ir_graph) || has_Block_label(block)) {
1882 /* start block and labelled blocks are always reachable */
1883 node->type.tv = tarval_reachable;
1887 for (i = get_Block_n_cfgpreds(block) - 1; i >= 0; --i) {
1888 node_t *pred = get_irn_node(get_Block_cfgpred(block, i));
1890 if (pred->type.tv == tarval_reachable) {
1891 /* A block is reachable, if at least of predecessor is reachable. */
1892 node->type.tv = tarval_reachable;
1896 node->type.tv = tarval_top;
1897 } /* compute_Block */
1900 * (Re-)compute the type for a Bad node.
1902 * @param node the node
1904 static void compute_Bad(node_t *node) {
1905 /* Bad nodes ALWAYS compute Top */
1906 node->type.tv = tarval_top;
1910 * (Re-)compute the type for an Unknown node.
1912 * @param node the node
1914 static void compute_Unknown(node_t *node) {
1915 /* While Unknown nodes should compute Top this is dangerous:
1916 * a Top input to a Cond would lead to BOTH control flows unreachable.
1917 * While this is correct in the given semantics, it would destroy the Firm
1920 * It would be safe to compute Top IF it can be assured, that only Cmp
1921 * nodes are inputs to Conds. We check that first.
1922 * This is the way Frontends typically build Firm, but some optimizations
1923 * (cond_eval for instance) might replace them by Phib's...
1925 node->type.tv = tarval_UNKNOWN;
1926 } /* compute_Unknown */
1929 * (Re-)compute the type for a Jmp node.
1931 * @param node the node
1933 static void compute_Jmp(node_t *node) {
1934 node_t *block = get_irn_node(get_nodes_block(node->node));
1936 node->type = block->type;
1940 * (Re-)compute the type for the Return node.
1942 * @param node the node
1944 static void compute_Return(node_t *node) {
1945 /* The Return node is NOT dead if it is in a reachable block.
1946 * This is already checked in compute(). so we can return
1947 * Reachable here. */
1948 node->type.tv = tarval_reachable;
1949 } /* compute_Return */
1952 * (Re-)compute the type for the End node.
1954 * @param node the node
1956 static void compute_End(node_t *node) {
1957 /* the End node is NOT dead of course */
1958 node->type.tv = tarval_reachable;
1962 * (Re-)compute the type for a Call.
1964 * @param node the node
1966 static void compute_Call(node_t *node) {
1968 * A Call computes always bottom, even if it has Unknown
1971 node->type.tv = tarval_bottom;
1972 } /* compute_Call */
1975 * (Re-)compute the type for a SymConst node.
1977 * @param node the node
1979 static void compute_SymConst(node_t *node) {
1980 ir_node *irn = node->node;
1981 node_t *block = get_irn_node(get_nodes_block(irn));
1983 if (block->type.tv == tarval_unreachable) {
1984 node->type.tv = tarval_top;
1987 switch (get_SymConst_kind(irn)) {
1988 case symconst_addr_ent:
1989 /* case symconst_addr_name: cannot handle this yet */
1990 node->type.sym = get_SymConst_symbol(irn);
1993 node->type.tv = computed_value(irn);
1995 } /* compute_SymConst */
1998 * (Re-)compute the type for a Phi node.
2000 * @param node the node
2002 static void compute_Phi(node_t *node) {
2004 ir_node *phi = node->node;
2005 lattice_elem_t type;
2007 /* if a Phi is in a unreachable block, its type is TOP */
2008 node_t *block = get_irn_node(get_nodes_block(phi));
2010 if (block->type.tv == tarval_unreachable) {
2011 node->type.tv = tarval_top;
2015 /* Phi implements the Meet operation */
2016 type.tv = tarval_top;
2017 for (i = get_Phi_n_preds(phi) - 1; i >= 0; --i) {
2018 node_t *pred = get_irn_node(get_Phi_pred(phi, i));
2019 node_t *pred_X = get_irn_node(get_Block_cfgpred(block->node, i));
2021 if (pred_X->type.tv == tarval_unreachable || pred->type.tv == tarval_top) {
2022 /* ignore TOP inputs: We must check here for unreachable blocks,
2023 because Firm constants live in the Start Block are NEVER Top.
2024 Else, a Phi (1,2) will produce Bottom, even if the 2 for instance
2025 comes from a unreachable input. */
2028 if (pred->type.tv == tarval_bottom) {
2029 node->type.tv = tarval_bottom;
2031 } else if (type.tv == tarval_top) {
2032 /* first constant found */
2034 } else if (type.tv != pred->type.tv) {
2035 /* different constants or tarval_bottom */
2036 node->type.tv = tarval_bottom;
2039 /* else nothing, constants are the same */
2045 * (Re-)compute the type for an Add. Special case: one nodes is a Zero Const.
2047 * @param node the node
2049 static void compute_Add(node_t *node) {
2050 ir_node *sub = node->node;
2051 node_t *l = get_irn_node(get_Add_left(sub));
2052 node_t *r = get_irn_node(get_Add_right(sub));
2053 lattice_elem_t a = l->type;
2054 lattice_elem_t b = r->type;
2057 if (a.tv == tarval_top || b.tv == tarval_top) {
2058 node->type.tv = tarval_top;
2059 } else if (a.tv == tarval_bottom || b.tv == tarval_bottom) {
2060 node->type.tv = tarval_bottom;
2062 /* x + 0 = 0 + x = x, but beware of floating point +0 + -0, so we
2063 must call tarval_add() first to handle this case! */
2064 if (is_tarval(a.tv)) {
2065 if (is_tarval(b.tv)) {
2066 node->type.tv = tarval_add(a.tv, b.tv);
2069 mode = get_tarval_mode(a.tv);
2070 if (a.tv == get_mode_null(mode)) {
2074 } else if (is_tarval(b.tv)) {
2075 mode = get_tarval_mode(b.tv);
2076 if (b.tv == get_mode_null(mode)) {
2081 node->type.tv = tarval_bottom;
2086 * (Re-)compute the type for a Sub. Special case: both nodes are congruent.
2088 * @param node the node
2090 static void compute_Sub(node_t *node) {
2091 ir_node *sub = node->node;
2092 node_t *l = get_irn_node(get_Sub_left(sub));
2093 node_t *r = get_irn_node(get_Sub_right(sub));
2094 lattice_elem_t a = l->type;
2095 lattice_elem_t b = r->type;
2098 if (a.tv == tarval_top || b.tv == tarval_top) {
2099 node->type.tv = tarval_top;
2100 } else if (is_con(a) && is_con(b)) {
2101 if (is_tarval(a.tv) && is_tarval(b.tv)) {
2102 node->type.tv = tarval_sub(a.tv, b.tv, get_irn_mode(sub));
2103 } else if (is_tarval(a.tv) && tarval_is_null(a.tv)) {
2105 } else if (is_tarval(b.tv) && tarval_is_null(b.tv)) {
2108 node->type.tv = tarval_bottom;
2110 } else if (r->part == l->part &&
2111 (!mode_is_float(get_irn_mode(l->node)))) {
2113 * BEWARE: a - a is NOT always 0 for floating Point values, as
2114 * NaN op NaN = NaN, so we must check this here.
2116 ir_mode *mode = get_irn_mode(sub);
2117 tv = get_mode_null(mode);
2119 /* if the node was ONCE evaluated by all constants, but now
2120 this breaks AND we get from the argument partitions a different
2121 result, switch to bottom.
2122 This happens because initially all nodes are in the same partition ... */
2123 if (node->type.tv != tv)
2127 node->type.tv = tarval_bottom;
2132 * (Re-)compute the type for an Eor. Special case: both nodes are congruent.
2134 * @param node the node
2136 static void compute_Eor(node_t *node) {
2137 ir_node *eor = node->node;
2138 node_t *l = get_irn_node(get_Eor_left(eor));
2139 node_t *r = get_irn_node(get_Eor_right(eor));
2140 lattice_elem_t a = l->type;
2141 lattice_elem_t b = r->type;
2144 if (a.tv == tarval_top || b.tv == tarval_top) {
2145 node->type.tv = tarval_top;
2146 } else if (is_con(a) && is_con(b)) {
2147 if (is_tarval(a.tv) && is_tarval(b.tv)) {
2148 node->type.tv = tarval_eor(a.tv, b.tv);
2149 } else if (is_tarval(a.tv) && tarval_is_null(a.tv)) {
2151 } else if (is_tarval(b.tv) && tarval_is_null(b.tv)) {
2154 node->type.tv = tarval_bottom;
2156 } else if (r->part == l->part) {
2157 ir_mode *mode = get_irn_mode(eor);
2158 tv = get_mode_null(mode);
2160 /* if the node was ONCE evaluated by all constants, but now
2161 this breaks AND we get from the argument partitions a different
2162 result, switch to bottom.
2163 This happens because initially all nodes are in the same partition ... */
2164 if (node->type.tv != tv)
2168 node->type.tv = tarval_bottom;
2173 * (Re-)compute the type for Cmp.
2175 * @param node the node
2177 static void compute_Cmp(node_t *node) {
2178 ir_node *cmp = node->node;
2179 node_t *l = get_irn_node(get_Cmp_left(cmp));
2180 node_t *r = get_irn_node(get_Cmp_right(cmp));
2181 lattice_elem_t a = l->type;
2182 lattice_elem_t b = r->type;
2184 if (a.tv == tarval_top || b.tv == tarval_top) {
2185 node->type.tv = tarval_top;
2186 } else if (r->part == l->part) {
2187 /* both nodes congruent, we can probably do something */
2188 node->type.tv = tarval_b_true;
2189 } else if (is_con(a) && is_con(b)) {
2190 /* both nodes are constants, we can probably do something */
2191 node->type.tv = tarval_b_true;
2193 node->type.tv = tarval_bottom;
2198 * (Re-)compute the type for a Proj(Cmp).
2200 * @param node the node
2201 * @param cond the predecessor Cmp node
2203 static void compute_Proj_Cmp(node_t *node, ir_node *cmp) {
2204 ir_node *proj = node->node;
2205 node_t *l = get_irn_node(get_Cmp_left(cmp));
2206 node_t *r = get_irn_node(get_Cmp_right(cmp));
2207 lattice_elem_t a = l->type;
2208 lattice_elem_t b = r->type;
2209 pn_Cmp pnc = get_Proj_proj(proj);
2212 if (a.tv == tarval_top || b.tv == tarval_top) {
2213 node->type.tv = tarval_undefined;
2214 } else if (is_con(a) && is_con(b)) {
2215 default_compute(node);
2216 } else if (r->part == l->part &&
2217 (!mode_is_float(get_irn_mode(l->node)) || pnc == pn_Cmp_Lt || pnc == pn_Cmp_Gt)) {
2219 * BEWARE: a == a is NOT always True for floating Point values, as
2220 * NaN != NaN is defined, so we must check this here.
2222 tv = pnc & pn_Cmp_Eq ? tarval_b_true: tarval_b_false;
2224 /* if the node was ONCE evaluated by all constants, but now
2225 this breaks AND we get from the argument partitions a different
2226 result, switch to bottom.
2227 This happens because initially all nodes are in the same partition ... */
2228 if (node->type.tv != tv)
2232 node->type.tv = tarval_bottom;
2234 } /* compute_Proj_Cmp */
2237 * (Re-)compute the type for a Proj(Cond).
2239 * @param node the node
2240 * @param cond the predecessor Cond node
2242 static void compute_Proj_Cond(node_t *node, ir_node *cond) {
2243 ir_node *proj = node->node;
2244 long pnc = get_Proj_proj(proj);
2245 ir_node *sel = get_Cond_selector(cond);
2246 node_t *selector = get_irn_node(sel);
2249 * Note: it is crucial for the monotony that the Proj(Cond)
2250 * are evaluates after all predecessors of the Cond selector are
2256 * Due to the fact that 0 is a const, the Cmp gets immediately
2257 * on the cprop list. It will be evaluated before x is evaluated,
2258 * might leaving x as Top. When later x is evaluated, the Cmp
2259 * might change its value.
2260 * BUT if the Cond is evaluated before this happens, Proj(Cond, FALSE)
2261 * gets R, and later changed to F if Cmp is evaluated to True!
2263 * We prevent this by putting Conds in an extra cprop_X queue, which
2264 * gets evaluated after the cprop queue is empty.
2266 * Note that this even happens with Click's original algorithm, if
2267 * Cmp(x, 0) is evaluated to True first and later changed to False
2268 * if x was Top first and later changed to a Const ...
2269 * It is unclear how Click solved that problem ...
2271 * However, in rare cases even this does not help, if a Top reaches
2272 * a compare through a Phi, than Proj(Cond) is evaluated changing
2273 * the type of the Phi to something other.
2274 * So, we take the last resort and bind the type to R once
2277 * (This might be even the way Click works around the whole problem).
2279 * Finally, we may miss some optimization possibilities due to this:
2284 * If Top reaches the if first, than we decide for != here.
2285 * If y later is evaluated to 0, we cannot revert this decision
2286 * and must live with both outputs enabled. If this happens,
2287 * we get an unresolved if (true) in the code ...
2289 * In Click's version where this decision is done at the Cmp,
2290 * the Cmp is NOT optimized away than (if y evaluated to 1
2291 * for instance) and we get a if (1 == 0) here ...
2293 * Both solutions are suboptimal.
2294 * At least, we could easily detect this problem and run
2295 * cf_opt() (or even combo) again :-(
2297 if (node->type.tv == tarval_reachable)
2300 if (get_irn_mode(sel) == mode_b) {
2302 if (pnc == pn_Cond_true) {
2303 if (selector->type.tv == tarval_b_false) {
2304 node->type.tv = tarval_unreachable;
2305 } else if (selector->type.tv == tarval_b_true) {
2306 node->type.tv = tarval_reachable;
2307 } else if (selector->type.tv == tarval_bottom) {
2308 node->type.tv = tarval_reachable;
2310 assert(selector->type.tv == tarval_top);
2311 if (tarval_UNKNOWN == tarval_top) {
2312 /* any condition based on Top is "!=" */
2313 node->type.tv = tarval_unreachable;
2315 node->type.tv = tarval_unreachable;
2319 assert(pnc == pn_Cond_false);
2321 if (selector->type.tv == tarval_b_false) {
2322 node->type.tv = tarval_reachable;
2323 } else if (selector->type.tv == tarval_b_true) {
2324 node->type.tv = tarval_unreachable;
2325 } else if (selector->type.tv == tarval_bottom) {
2326 node->type.tv = tarval_reachable;
2328 assert(selector->type.tv == tarval_top);
2329 if (tarval_UNKNOWN == tarval_top) {
2330 /* any condition based on Top is "!=" */
2331 node->type.tv = tarval_reachable;
2333 node->type.tv = tarval_unreachable;
2339 if (selector->type.tv == tarval_bottom) {
2340 node->type.tv = tarval_reachable;
2341 } else if (selector->type.tv == tarval_top) {
2342 if (tarval_UNKNOWN == tarval_top &&
2343 pnc == get_Cond_defaultProj(cond)) {
2344 /* a switch based of Top is always "default" */
2345 node->type.tv = tarval_reachable;
2347 node->type.tv = tarval_unreachable;
2350 long value = get_tarval_long(selector->type.tv);
2351 if (pnc == get_Cond_defaultProj(cond)) {
2352 /* default switch, have to check ALL other cases */
2355 for (i = get_irn_n_outs(cond) - 1; i >= 0; --i) {
2356 ir_node *succ = get_irn_out(cond, i);
2360 if (value == get_Proj_proj(succ)) {
2361 /* we found a match, will NOT take the default case */
2362 node->type.tv = tarval_unreachable;
2366 /* all cases checked, no match, will take default case */
2367 node->type.tv = tarval_reachable;
2370 node->type.tv = value == pnc ? tarval_reachable : tarval_unreachable;
2374 } /* compute_Proj_Cond */
2377 * (Re-)compute the type for a Proj-Node.
2379 * @param node the node
2381 static void compute_Proj(node_t *node) {
2382 ir_node *proj = node->node;
2383 ir_mode *mode = get_irn_mode(proj);
2384 node_t *block = get_irn_node(get_nodes_block(skip_Proj(proj)));
2385 ir_node *pred = get_Proj_pred(proj);
2387 if (block->type.tv == tarval_unreachable) {
2388 /* a Proj in a unreachable Block stay Top */
2389 node->type.tv = tarval_top;
2392 if (get_irn_node(pred)->type.tv == tarval_top && !is_Cond(pred)) {
2393 /* if the predecessor is Top, its Proj follow */
2394 node->type.tv = tarval_top;
2398 if (mode == mode_M) {
2399 /* mode M is always bottom */
2400 node->type.tv = tarval_bottom;
2403 if (mode != mode_X) {
2405 compute_Proj_Cmp(node, pred);
2407 default_compute(node);
2410 /* handle mode_X nodes */
2412 switch (get_irn_opcode(pred)) {
2414 /* the Proj_X from the Start is always reachable.
2415 However this is already handled at the top. */
2416 node->type.tv = tarval_reachable;
2419 compute_Proj_Cond(node, pred);
2422 default_compute(node);
2424 } /* compute_Proj */
2427 * (Re-)compute the type for a Confirm.
2429 * @param node the node
2431 static void compute_Confirm(node_t *node) {
2432 ir_node *confirm = node->node;
2433 node_t *pred = get_irn_node(get_Confirm_value(confirm));
2435 if (get_Confirm_cmp(confirm) == pn_Cmp_Eq) {
2436 node_t *bound = get_irn_node(get_Confirm_bound(confirm));
2438 if (is_con(bound->type)) {
2439 /* is equal to a constant */
2440 node->type = bound->type;
2444 /* a Confirm is a copy OR a Const */
2445 node->type = pred->type;
2446 } /* compute_Confirm */
2449 * (Re-)compute the type for a given node.
2451 * @param node the node
2453 static void compute(node_t *node) {
2454 ir_node *irn = node->node;
2457 #ifndef VERIFY_MONOTONE
2459 * Once a node reaches bottom, the type cannot fall further
2460 * in the lattice and we can stop computation.
2461 * Do not take this exit if the monotony verifier is
2462 * enabled to catch errors.
2464 if (node->type.tv == tarval_bottom)
2468 if (is_no_Block(irn)) {
2469 /* for pinned nodes, check its control input */
2470 if (get_irn_pinned(skip_Proj(irn)) == op_pin_state_pinned) {
2471 node_t *block = get_irn_node(get_nodes_block(irn));
2473 if (block->type.tv == tarval_unreachable) {
2474 node->type.tv = tarval_top;
2480 func = (compute_func)node->node->op->ops.generic;
2486 * Identity functions: Note that one might thing that identity() is just a
2487 * synonym for equivalent_node(). While this is true, we cannot use it for the algorithm
2488 * here, because it expects that the identity node is one of the inputs, which is NOT
2489 * always true for equivalent_node() which can handle (and does sometimes) DAGs.
2490 * So, we have our own implementation, which copies some parts of equivalent_node()
2494 * Calculates the Identity for Phi nodes
2496 static node_t *identity_Phi(node_t *node) {
2497 ir_node *phi = node->node;
2498 ir_node *block = get_nodes_block(phi);
2499 node_t *n_part = NULL;
2502 for (i = get_Phi_n_preds(phi) - 1; i >= 0; --i) {
2503 node_t *pred_X = get_irn_node(get_Block_cfgpred(block, i));
2505 if (pred_X->type.tv == tarval_reachable) {
2506 node_t *pred = get_irn_node(get_Phi_pred(phi, i));
2510 else if (n_part->part != pred->part) {
2511 /* incongruent inputs, not a follower */
2516 /* if n_part is NULL here, all inputs path are dead, the Phi computes
2517 * tarval_top, is in the TOP partition and should NOT being split! */
2518 assert(n_part != NULL);
2520 } /* identity_Phi */
2523 * Calculates the Identity for commutative 0 neutral nodes.
2525 static node_t *identity_comm_zero_binop(node_t *node) {
2526 ir_node *op = node->node;
2527 node_t *a = get_irn_node(get_binop_left(op));
2528 node_t *b = get_irn_node(get_binop_right(op));
2529 ir_mode *mode = get_irn_mode(op);
2532 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
2533 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
2536 /* node: no input should be tarval_top, else the binop would be also
2537 * Top and not being split. */
2538 zero = get_mode_null(mode);
2539 if (a->type.tv == zero)
2541 if (b->type.tv == zero)
2544 } /* identity_comm_zero_binop */
2547 * Calculates the Identity for Shift nodes.
2549 static node_t *identity_shift(node_t *node) {
2550 ir_node *op = node->node;
2551 node_t *b = get_irn_node(get_binop_right(op));
2552 ir_mode *mode = get_irn_mode(b->node);
2555 /* node: no input should be tarval_top, else the binop would be also
2556 * Top and not being split. */
2557 zero = get_mode_null(mode);
2558 if (b->type.tv == zero)
2559 return get_irn_node(get_binop_left(op));
2561 } /* identity_shift */
2564 * Calculates the Identity for Mul nodes.
2566 static node_t *identity_Mul(node_t *node) {
2567 ir_node *op = node->node;
2568 node_t *a = get_irn_node(get_Mul_left(op));
2569 node_t *b = get_irn_node(get_Mul_right(op));
2570 ir_mode *mode = get_irn_mode(op);
2573 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
2574 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
2577 /* node: no input should be tarval_top, else the binop would be also
2578 * Top and not being split. */
2579 one = get_mode_one(mode);
2580 if (a->type.tv == one)
2582 if (b->type.tv == one)
2585 } /* identity_Mul */
2588 * Calculates the Identity for Sub nodes.
2590 static node_t *identity_Sub(node_t *node) {
2591 ir_node *sub = node->node;
2592 node_t *b = get_irn_node(get_Sub_right(sub));
2593 ir_mode *mode = get_irn_mode(sub);
2595 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
2596 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
2599 /* node: no input should be tarval_top, else the binop would be also
2600 * Top and not being split. */
2601 if (b->type.tv == get_mode_null(mode))
2602 return get_irn_node(get_Sub_left(sub));
2604 } /* identity_Sub */
2607 * Calculates the Identity for And nodes.
2609 static node_t *identity_And(node_t *node) {
2610 ir_node *and = node->node;
2611 node_t *a = get_irn_node(get_And_left(and));
2612 node_t *b = get_irn_node(get_And_right(and));
2613 tarval *neutral = get_mode_all_one(get_irn_mode(and));
2615 /* node: no input should be tarval_top, else the And would be also
2616 * Top and not being split. */
2617 if (a->type.tv == neutral)
2619 if (b->type.tv == neutral)
2622 } /* identity_And */
2625 * Calculates the Identity for Confirm nodes.
2627 static node_t *identity_Confirm(node_t *node) {
2628 ir_node *confirm = node->node;
2630 /* a Confirm is always a Copy */
2631 return get_irn_node(get_Confirm_value(confirm));
2632 } /* identity_Confirm */
2635 * Calculates the Identity for Mux nodes.
2637 static node_t *identity_Mux(node_t *node) {
2638 ir_node *mux = node->node;
2639 node_t *t = get_irn_node(get_Mux_true(mux));
2640 node_t *f = get_irn_node(get_Mux_false(mux));
2643 if (t->part == f->part)
2646 /* for now, the 1-input identity is not supported */
2648 sel = get_irn_node(get_Mux_sel(mux));
2650 /* Mux sel input is mode_b, so it is always a tarval */
2651 if (sel->type.tv == tarval_b_true)
2653 if (sel->type.tv == tarval_b_false)
2657 } /* identity_Mux */
2660 * Calculates the Identity for nodes.
2662 static node_t *identity(node_t *node) {
2663 ir_node *irn = node->node;
2665 switch (get_irn_opcode(irn)) {
2667 return identity_Phi(node);
2669 return identity_Mul(node);
2673 return identity_comm_zero_binop(node);
2678 return identity_shift(node);
2680 return identity_And(node);
2682 return identity_Sub(node);
2684 return identity_Confirm(node);
2686 return identity_Mux(node);
2693 * Node follower is a (new) follower of leader, segregate Leader
2696 static void segregate_def_use_chain_1(const ir_node *follower, node_t *leader) {
2697 ir_node *l = leader->node;
2698 int j, i, n = get_irn_n_outs(l);
2700 DB((dbg, LEVEL_2, "%+F is a follower of %+F\n", follower, leader->node));
2701 /* The leader edges must remain sorted, but follower edges can
2703 for (i = leader->n_followers + 1; i <= n; ++i) {
2704 if (l->out[i].use == follower) {
2705 ir_def_use_edge t = l->out[i];
2707 for (j = i - 1; j >= leader->n_followers + 1; --j)
2708 l->out[j + 1] = l->out[j];
2709 ++leader->n_followers;
2710 l->out[leader->n_followers] = t;
2714 } /* segregate_def_use_chain_1 */
2717 * Node follower is a (new) follower segregate its Leader
2720 * @param follower the follower IR node
2722 static void segregate_def_use_chain(const ir_node *follower) {
2725 for (i = get_irn_arity(follower) - 1; i >= 0; --i) {
2726 node_t *pred = get_irn_node(get_irn_n(follower, i));
2728 segregate_def_use_chain_1(follower, pred);
2730 } /* segregate_def_use_chain */
2733 * Propagate constant evaluation.
2735 * @param env the environment
2737 static void propagate(environment_t *env) {
2740 lattice_elem_t old_type;
2742 unsigned n_fallen, old_type_was_T_or_C;
2745 while (env->cprop != NULL) {
2746 void *oldopcode = NULL;
2748 /* remove the first partition X from cprop */
2751 env->cprop = X->cprop_next;
2753 old_type_was_T_or_C = X->type_is_T_or_C;
2755 DB((dbg, LEVEL_2, "Propagate type on part%d\n", X->nr));
2759 int cprop_empty = list_empty(&X->cprop);
2760 int cprop_X_empty = list_empty(&X->cprop_X);
2762 if (cprop_empty && cprop_X_empty) {
2763 /* both cprop lists are empty */
2767 /* remove the first Node x from X.cprop */
2769 /* Get a node from the cprop_X list only if
2770 * all data nodes are processed.
2771 * This ensures, that all inputs of the Cond
2772 * predecessor are processed if its type is still Top.
2774 x = list_entry(X->cprop_X.next, node_t, cprop_list);
2776 x = list_entry(X->cprop.next, node_t, cprop_list);
2779 //assert(x->part == X);
2780 list_del(&x->cprop_list);
2783 if (x->is_follower && identity(x) == x) {
2784 /* check the opcode first */
2785 if (oldopcode == NULL) {
2786 oldopcode = lambda_opcode(get_first_node(X), env);
2788 if (oldopcode != lambda_opcode(x, env)) {
2789 if (x->on_fallen == 0) {
2790 /* different opcode -> x falls out of this partition */
2795 DB((dbg, LEVEL_2, "Add node %+F to fallen\n", x->node));
2799 /* x will make the follower -> leader transition */
2800 follower_to_leader(x);
2803 /* compute a new type for x */
2805 DB((dbg, LEVEL_3, "computing type of %+F\n", x->node));
2807 if (x->type.tv != old_type.tv) {
2808 DB((dbg, LEVEL_2, "node %+F has changed type from %+F to %+F\n", x->node, old_type, x->type));
2809 verify_type(old_type, x);
2811 if (x->on_fallen == 0) {
2812 /* Add x to fallen. Nodes might fall from T -> const -> _|_, so check that they are
2813 not already on the list. */
2818 DB((dbg, LEVEL_2, "Add node %+F to fallen\n", x->node));
2820 for (i = get_irn_n_outs(x->node) - 1; i >= 0; --i) {
2821 ir_node *succ = get_irn_out(x->node, i);
2822 node_t *y = get_irn_node(succ);
2824 /* Add y to y.partition.cprop. */
2825 add_to_cprop(y, env);
2830 if (n_fallen > 0 && n_fallen != X->n_leader) {
2831 DB((dbg, LEVEL_2, "Splitting part%d by fallen\n", X->nr));
2832 Y = split(&X, fallen, env);
2834 * We have split out fallen node. The type of the result
2835 * partition is NOT set yet.
2837 Y->type_is_T_or_C = 0;
2841 /* remove the flags from the fallen list */
2842 for (x = fallen; x != NULL; x = x->next)
2845 if (old_type_was_T_or_C) {
2848 /* check if some nodes will make the leader -> follower transition */
2849 list_for_each_entry_safe(node_t, y, tmp, &Y->Leader, node_list) {
2850 if (y->type.tv != tarval_top && ! is_con(y->type)) {
2851 node_t *eq_node = identity(y);
2853 if (eq_node != y && eq_node->part == y->part) {
2854 DB((dbg, LEVEL_2, "Node %+F is a follower of %+F\n", y->node, eq_node->node));
2855 /* move to Follower */
2857 list_del(&y->node_list);
2858 list_add_tail(&y->node_list, &Y->Follower);
2861 segregate_def_use_chain(y->node);
2871 * Get the leader for a given node from its congruence class.
2873 * @param irn the node
2875 static ir_node *get_leader(node_t *node) {
2876 partition_t *part = node->part;
2878 if (part->n_leader > 1 || node->is_follower) {
2879 if (node->is_follower) {
2880 DB((dbg, LEVEL_2, "Replacing follower %+F\n", node->node));
2883 DB((dbg, LEVEL_2, "Found congruence class for %+F\n", node->node));
2885 return get_first_node(part)->node;
2891 * Returns non-zero if a mode_T node has only one reachable output.
2893 static int only_one_reachable_proj(ir_node *n) {
2896 for (i = get_irn_n_outs(n) - 1; i >= 0; --i) {
2897 ir_node *proj = get_irn_out(n, i);
2900 /* skip non-control flow Proj's */
2901 if (get_irn_mode(proj) != mode_X)
2904 node = get_irn_node(proj);
2905 if (node->type.tv == tarval_reachable) {
2911 } /* only_one_reachable_proj */
2914 * Return non-zero if the control flow predecessor node pred
2915 * is the only reachable control flow exit of its block.
2917 * @param pred the control flow exit
2918 * @param block the destination block
2920 static int can_exchange(ir_node *pred, ir_node *block) {
2921 if (is_Start(pred) || has_Block_label(block))
2923 else if (is_Jmp(pred))
2925 else if (get_irn_mode(pred) == mode_T) {
2926 /* if the predecessor block has more than one
2927 reachable outputs we cannot remove the block */
2928 return only_one_reachable_proj(pred);
2931 } /* can_exchange */
2934 * Block Post-Walker, apply the analysis results on control flow by
2935 * shortening Phi's and Block inputs.
2937 static void apply_cf(ir_node *block, void *ctx) {
2938 environment_t *env = ctx;
2939 node_t *node = get_irn_node(block);
2941 ir_node **ins, **in_X;
2942 ir_node *phi, *next;
2944 n = get_Block_n_cfgpreds(block);
2946 if (node->type.tv == tarval_unreachable) {
2949 for (i = n - 1; i >= 0; --i) {
2950 ir_node *pred = get_Block_cfgpred(block, i);
2952 if (! is_Bad(pred)) {
2953 node_t *pred_bl = get_irn_node(get_nodes_block(skip_Proj(pred)));
2955 if (pred_bl->flagged == 0) {
2956 pred_bl->flagged = 3;
2958 if (pred_bl->type.tv == tarval_reachable) {
2960 * We will remove an edge from block to its pred.
2961 * This might leave the pred block as an endless loop
2963 if (! is_backedge(block, i))
2964 keep_alive(pred_bl->node);
2970 /* the EndBlock is always reachable even if the analysis
2971 finds out the opposite :-) */
2972 if (block != get_irg_end_block(current_ir_graph)) {
2973 /* mark dead blocks */
2974 set_Block_dead(block);
2975 DB((dbg, LEVEL_1, "Removing dead %+F\n", block));
2977 /* the endblock is unreachable */
2978 set_irn_in(block, 0, NULL);
2984 /* only one predecessor combine */
2985 ir_node *pred = skip_Proj(get_Block_cfgpred(block, 0));
2987 if (can_exchange(pred, block)) {
2988 ir_node *new_block = get_nodes_block(pred);
2989 DB((dbg, LEVEL_1, "Fuse %+F with %+F\n", block, new_block));
2990 DBG_OPT_COMBO(block, new_block, FS_OPT_COMBO_CF);
2991 exchange(block, new_block);
2992 node->node = new_block;
2998 NEW_ARR_A(ir_node *, in_X, n);
3000 for (i = 0; i < n; ++i) {
3001 ir_node *pred = get_Block_cfgpred(block, i);
3002 node_t *node = get_irn_node(pred);
3004 if (node->type.tv == tarval_reachable) {
3007 DB((dbg, LEVEL_1, "Removing dead input %d from %+F (%+F)\n", i, block, pred));
3008 if (! is_Bad(pred)) {
3009 node_t *pred_bl = get_irn_node(get_nodes_block(skip_Proj(pred)));
3011 if (pred_bl->flagged == 0) {
3012 pred_bl->flagged = 3;
3014 if (pred_bl->type.tv == tarval_reachable) {
3016 * We will remove an edge from block to its pred.
3017 * This might leave the pred block as an endless loop
3019 if (! is_backedge(block, i))
3020 keep_alive(pred_bl->node);
3030 NEW_ARR_A(ir_node *, ins, n);
3031 for (phi = get_Block_phis(block); phi != NULL; phi = next) {
3032 node_t *node = get_irn_node(phi);
3034 next = get_Phi_next(phi);
3035 if (is_tarval(node->type.tv) && tarval_is_constant(node->type.tv)) {
3036 /* this Phi is replaced by a constant */
3037 tarval *tv = node->type.tv;
3038 ir_node *c = new_Const(tv);
3040 set_irn_node(c, node);
3042 DB((dbg, LEVEL_1, "%+F is replaced by %+F\n", phi, c));
3043 DBG_OPT_COMBO(phi, c, FS_OPT_COMBO_CONST);
3048 for (i = 0; i < n; ++i) {
3049 node_t *pred = get_irn_node(get_Block_cfgpred(block, i));
3051 if (pred->type.tv == tarval_reachable) {
3052 ins[j++] = get_Phi_pred(phi, i);
3056 /* this Phi is replaced by a single predecessor */
3057 ir_node *s = ins[0];
3058 node_t *phi_node = get_irn_node(phi);
3061 DB((dbg, LEVEL_1, "%+F is replaced by %+F because of cf change\n", phi, s));
3062 DBG_OPT_COMBO(phi, s, FS_OPT_COMBO_FOLLOWER);
3067 set_irn_in(phi, j, ins);
3075 /* this Block has only one live predecessor */
3076 ir_node *pred = skip_Proj(in_X[0]);
3078 if (can_exchange(pred, block)) {
3079 ir_node *new_block = get_nodes_block(pred);
3080 DBG_OPT_COMBO(block, new_block, FS_OPT_COMBO_CF);
3081 exchange(block, new_block);
3082 node->node = new_block;
3087 set_irn_in(block, k, in_X);
3092 * Exchange a node by its leader.
3093 * Beware: in rare cases the mode might be wrong here, for instance
3094 * AddP(x, NULL) is a follower of x, but with different mode.
3097 static void exchange_leader(ir_node *irn, ir_node *leader) {
3098 ir_mode *mode = get_irn_mode(irn);
3099 if (mode != get_irn_mode(leader)) {
3100 /* The conv is a no-op, so we are free to place it
3101 * either in the block of the leader OR in irn's block.
3102 * Probably placing it into leaders block might reduce
3103 * the number of Conv due to CSE. */
3104 ir_node *block = get_nodes_block(leader);
3105 dbg_info *dbg = get_irn_dbg_info(irn);
3107 leader = new_rd_Conv(dbg, current_ir_graph, block, leader, mode);
3109 exchange(irn, leader);
3110 } /* exchange_leader */
3113 * Check, if all users of a mode_M node are dead. Use
3114 * the Def-Use edges for this purpose, as they still
3115 * reflect the situation.
3117 static int all_users_are_dead(const ir_node *irn) {
3118 int i, n = get_irn_n_outs(irn);
3120 for (i = 1; i <= n; ++i) {
3121 const ir_node *succ = irn->out[i].use;
3122 const node_t *block = get_irn_node(get_nodes_block(succ));
3125 if (block->type.tv == tarval_unreachable) {
3126 /* block is unreachable */
3129 node = get_irn_node(succ);
3130 if (node->type.tv != tarval_top) {
3131 /* found a reachable user */
3135 /* all users are unreachable */
3137 } /* all_user_are_dead */
3140 * Walker: Find reachable mode_M nodes that have only
3141 * unreachable users. These nodes must be kept later.
3143 static void find_kept_memory(ir_node *irn, void *ctx) {
3144 environment_t *env = ctx;
3145 node_t *node, *block;
3147 if (get_irn_mode(irn) != mode_M)
3150 block = get_irn_node(get_nodes_block(irn));
3151 if (block->type.tv == tarval_unreachable)
3154 node = get_irn_node(irn);
3155 if (node->type.tv == tarval_top)
3158 /* ok, we found a live memory node. */
3159 if (all_users_are_dead(irn)) {
3160 DB((dbg, LEVEL_1, "%+F must be kept\n", irn));
3161 ARR_APP1(ir_node *, env->kept_memory, irn);
3163 } /* find_kept_memory */
3166 * Post-Walker, apply the analysis results;
3168 static void apply_result(ir_node *irn, void *ctx) {
3169 environment_t *env = ctx;
3170 node_t *node = get_irn_node(irn);
3172 if (is_Block(irn) || is_End(irn) || is_Bad(irn)) {
3173 /* blocks already handled, do not touch the End node */
3175 node_t *block = get_irn_node(get_nodes_block(irn));
3177 if (block->type.tv == tarval_unreachable) {
3178 ir_node *bad = get_irg_bad(current_ir_graph);
3180 /* here, bad might already have a node, but this can be safely ignored
3181 as long as bad has at least ONE valid node */
3182 set_irn_node(bad, node);
3184 DB((dbg, LEVEL_1, "%+F is unreachable\n", irn));
3187 } else if (node->type.tv == tarval_top) {
3188 ir_mode *mode = get_irn_mode(irn);
3190 if (mode == mode_M) {
3191 /* never kill a mode_M node */
3193 ir_node *pred = get_Proj_pred(irn);
3194 node_t *pnode = get_irn_node(pred);
3196 if (pnode->type.tv == tarval_top) {
3197 /* skip the predecessor */
3198 ir_node *mem = get_memop_mem(pred);
3200 DB((dbg, LEVEL_1, "%+F computes Top, replaced by %+F\n", irn, mem));
3205 /* leave other nodes, especially PhiM */
3206 } else if (mode == mode_T) {
3207 /* Do not kill mode_T nodes, kill their Projs */
3208 } else if (! is_Unknown(irn)) {
3209 /* don't kick away Unknown's, they might be still needed */
3210 ir_node *unk = new_r_Unknown(current_ir_graph, mode);
3212 /* control flow should already be handled at apply_cf() */
3213 assert(mode != mode_X);
3215 /* see comment above */
3216 set_irn_node(unk, node);
3218 DB((dbg, LEVEL_1, "%+F computes Top\n", irn));
3223 else if (get_irn_mode(irn) == mode_X) {
3226 ir_node *cond = get_Proj_pred(irn);
3228 if (is_Cond(cond)) {
3229 if (only_one_reachable_proj(cond)) {
3230 ir_node *jmp = new_r_Jmp(current_ir_graph, block->node);
3231 set_irn_node(jmp, node);
3233 DB((dbg, LEVEL_1, "%+F is replaced by %+F\n", irn, jmp));
3234 DBG_OPT_COMBO(irn, jmp, FS_OPT_COMBO_CF);
3238 node_t *sel = get_irn_node(get_Cond_selector(cond));
3239 tarval *tv = sel->type.tv;
3241 if (is_tarval(tv) && tarval_is_constant(tv)) {
3242 /* The selector is a constant, but more
3243 * than one output is active: An unoptimized
3251 /* normal data node */
3252 if (is_tarval(node->type.tv) && tarval_is_constant(node->type.tv)) {
3253 tarval *tv = node->type.tv;
3256 * Beware: never replace mode_T nodes by constants. Currently we must mark
3257 * mode_T nodes with constants, but do NOT replace them.
3259 if (! is_Const(irn) && get_irn_mode(irn) != mode_T) {
3260 /* can be replaced by a constant */
3261 ir_node *c = new_Const(tv);
3262 set_irn_node(c, node);
3264 DB((dbg, LEVEL_1, "%+F is replaced by %+F\n", irn, c));
3265 DBG_OPT_COMBO(irn, c, FS_OPT_COMBO_CONST);
3266 exchange_leader(irn, c);
3269 } else if (is_entity(node->type.sym.entity_p)) {
3270 if (! is_SymConst(irn)) {
3271 /* can be replaced by a SymConst */
3272 ir_node *symc = new_r_SymConst(current_ir_graph, block->node, get_irn_mode(irn), node->type.sym, symconst_addr_ent);
3273 set_irn_node(symc, node);
3276 DB((dbg, LEVEL_1, "%+F is replaced by %+F\n", irn, symc));
3277 DBG_OPT_COMBO(irn, symc, FS_OPT_COMBO_CONST);
3278 exchange_leader(irn, symc);
3281 } else if (is_Confirm(irn)) {
3282 /* Confirms are always follower, but do not kill them here */
3284 ir_node *leader = get_leader(node);
3286 if (leader != irn) {
3287 int non_strict_phi = 0;
3290 * Beware: Do not remove Phi(Unknown, ..., x, ..., Unknown)
3291 * as this might create non-strict programs.
3293 if (node->is_follower && is_Phi(irn) && !is_Unknown(leader)) {
3296 for (i = get_Phi_n_preds(irn) - 1; i >= 0; --i) {
3297 ir_node *pred = get_Phi_pred(irn, i);
3299 if (is_Unknown(pred)) {
3305 if (! non_strict_phi) {
3306 DB((dbg, LEVEL_1, "%+F from part%d is replaced by %+F\n", irn, node->part->nr, leader));
3307 if (node->is_follower)
3308 DBG_OPT_COMBO(irn, leader, FS_OPT_COMBO_FOLLOWER);
3310 DBG_OPT_COMBO(irn, leader, FS_OPT_COMBO_CONGRUENT);
3311 exchange_leader(irn, leader);
3318 } /* apply_result */
3321 * Fix the keep-alives by deleting unreachable ones.
3323 static void apply_end(ir_node *end, environment_t *env) {
3324 int i, j, n = get_End_n_keepalives(end);
3328 NEW_ARR_A(ir_node *, in, n);
3330 /* fix the keep alive */
3331 for (i = j = 0; i < n; i++) {
3332 ir_node *ka = get_End_keepalive(end, i);
3333 node_t *node = get_irn_node(ka);
3336 node = get_irn_node(get_nodes_block(ka));
3338 if (node->type.tv != tarval_unreachable && !is_Bad(ka))
3342 set_End_keepalives(end, j, in);
3347 #define SET(code) op_##code->ops.generic = (op_func)compute_##code
3350 * sets the generic functions to compute.
3352 static void set_compute_functions(void) {
3355 /* set the default compute function */
3356 for (i = get_irp_n_opcodes() - 1; i >= 0; --i) {
3357 ir_op *op = get_irp_opcode(i);
3358 op->ops.generic = (op_func)default_compute;
3361 /* set specific functions */
3377 } /* set_compute_functions */
3382 static void add_memory_keeps(ir_node **kept_memory, int len) {
3383 ir_node *end = get_irg_end(current_ir_graph);
3387 ir_nodeset_init(&set);
3389 /* check, if those nodes are already kept */
3390 for (i = get_End_n_keepalives(end) - 1; i >= 0; --i)
3391 ir_nodeset_insert(&set, get_End_keepalive(end, i));
3393 for (i = len - 1; i >= 0; --i) {
3394 ir_node *ka = kept_memory[i];
3396 if (! ir_nodeset_contains(&set, ka)) {
3397 add_End_keepalive(end, ka);
3400 ir_nodeset_destroy(&set);
3401 } /* add_memory_keeps */
3403 void combo(ir_graph *irg) {
3405 ir_node *initial_bl;
3407 ir_graph *rem = current_ir_graph;
3410 current_ir_graph = irg;
3412 /* register a debug mask */
3413 FIRM_DBG_REGISTER(dbg, "firm.opt.combo");
3415 DB((dbg, LEVEL_1, "Doing COMBO for %+F\n", irg));
3417 obstack_init(&env.obst);
3418 env.worklist = NULL;
3422 #ifdef DEBUG_libfirm
3423 env.dbg_list = NULL;
3425 env.opcode2id_map = new_set(cmp_opcode, iro_Last * 4);
3426 env.type2id_map = pmap_create();
3427 env.kept_memory = NEW_ARR_F(ir_node *, 0);
3428 env.end_idx = get_opt_global_cse() ? 0 : -1;
3429 env.lambda_input = 0;
3432 /* options driving the optimization */
3433 env.commutative = 1;
3434 env.opt_unknown = 1;
3436 assure_irg_outs(irg);
3437 assure_cf_loop(irg);
3439 /* we have our own value_of function */
3440 set_value_of_func(get_node_tarval);
3442 set_compute_functions();
3443 DEBUG_ONLY(part_nr = 0);
3445 ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK | IR_RESOURCE_PHI_LIST);
3447 if (env.opt_unknown)
3448 tarval_UNKNOWN = tarval_top;
3450 tarval_UNKNOWN = tarval_bad;
3452 /* create the initial partition and place it on the work list */
3453 env.initial = new_partition(&env);
3454 add_to_worklist(env.initial, &env);
3455 irg_walk_graph(irg, create_initial_partitions, init_block_phis, &env);
3457 /* set the hook: from now, every node has a partition and a type */
3458 DEBUG_ONLY(set_dump_node_vcgattr_hook(dump_partition_hook));
3460 /* all nodes on the initial partition have type Top */
3461 env.initial->type_is_T_or_C = 1;
3463 /* Place the START Node's partition on cprop.
3464 Place the START Node on its local worklist. */
3465 initial_bl = get_irg_start_block(irg);
3466 start = get_irn_node(initial_bl);
3467 add_to_cprop(start, &env);
3471 if (env.worklist != NULL)
3473 } while (env.cprop != NULL || env.worklist != NULL);
3475 dump_all_partitions(&env);
3476 check_all_partitions(&env);
3479 dump_ir_block_graph(irg, "-partition");
3482 /* apply the result */
3484 /* check, which nodes must be kept */
3485 irg_walk_graph(irg, NULL, find_kept_memory, &env);
3487 /* kill unreachable control flow */
3488 irg_block_walk_graph(irg, NULL, apply_cf, &env);
3489 /* Kill keep-alives of dead blocks: this speeds up apply_result()
3490 * and fixes assertion because dead cf to dead blocks is NOT removed by
3492 apply_end(get_irg_end(irg), &env);
3493 irg_walk_graph(irg, NULL, apply_result, &env);
3495 len = ARR_LEN(env.kept_memory);
3497 add_memory_keeps(env.kept_memory, len);
3500 DB((dbg, LEVEL_1, "Unoptimized Control Flow left"));
3504 /* control flow might changed */
3505 set_irg_outs_inconsistent(irg);
3506 set_irg_extblk_inconsistent(irg);
3507 set_irg_doms_inconsistent(irg);
3508 set_irg_loopinfo_inconsistent(irg);
3511 ir_free_resources(irg, IR_RESOURCE_IRN_LINK | IR_RESOURCE_PHI_LIST);
3513 /* remove the partition hook */
3514 DEBUG_ONLY(set_dump_node_vcgattr_hook(NULL));
3516 DEL_ARR_F(env.kept_memory);
3517 pmap_destroy(env.type2id_map);
3518 del_set(env.opcode2id_map);
3519 obstack_free(&env.obst, NULL);
3521 /* restore value_of() default behavior */
3522 set_value_of_func(NULL);
3523 current_ir_graph = rem;