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 */
115 ir_node *block; /**< for Block: itself */
116 void *ptr; /**< generic pointer for hash/cmp */
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 ir_node **kept_memory; /**< Array of memory nodes that must be kept. */
200 int end_idx; /**< -1 for local and 0 for global congruences. */
201 int lambda_input; /**< Captured argument for lambda_partition(). */
202 unsigned modified:1; /**< Set, if the graph was modified. */
203 unsigned unopt_cf:1; /**< If set, control flow is not optimized due to Unknown. */
204 /* options driving the optimization */
205 unsigned commutative:1; /**< Set, if commutation nodes should be handled specially. */
206 unsigned opt_unknown:1; /**< Set, if non-strict programs should be optimized. */
208 partition_t *dbg_list; /**< List of all partitions. */
212 /** Type of the what function. */
213 typedef void *(*what_func)(const node_t *node, environment_t *env);
215 #define get_irn_node(irn) ((node_t *)get_irn_link(irn))
216 #define set_irn_node(irn, node) set_irn_link(irn, node)
218 /* we do NOT use tarval_unreachable here, instead we use Top for this purpose */
219 #undef tarval_unreachable
220 #define tarval_unreachable tarval_top
223 /** The debug module handle. */
224 DEBUG_ONLY(static firm_dbg_module_t *dbg;)
226 /** The what reason. */
227 DEBUG_ONLY(static const char *what_reason;)
229 /** Next partition number. */
230 DEBUG_ONLY(static unsigned part_nr = 0);
232 /** The tarval returned by Unknown nodes: set to either tarval_bad OR tarval_top. */
233 static tarval *tarval_UNKNOWN;
236 static node_t *identity(node_t *node);
238 #ifdef CHECK_PARTITIONS
242 static void check_partition(const partition_t *T) {
246 list_for_each_entry(node_t, node, &T->Leader, node_list) {
247 assert(node->is_follower == 0);
248 assert(node->flagged == 0);
249 assert(node->part == T);
252 assert(n == T->n_leader);
254 list_for_each_entry(node_t, node, &T->Follower, node_list) {
255 assert(node->is_follower == 1);
256 assert(node->flagged == 0);
257 assert(node->part == T);
259 } /* check_partition */
262 * check that all leader nodes in the partition have the same opcode.
264 static void check_opcode(const partition_t *Z) {
269 list_for_each_entry(node_t, node, &Z->Leader, node_list) {
270 ir_node *irn = node->node;
273 key.code = get_irn_opcode(irn);
274 key.mode = get_irn_mode(irn);
275 key.arity = get_irn_arity(irn);
279 switch (get_irn_opcode(irn)) {
281 key.u.proj = get_Proj_proj(irn);
284 key.u.ent = get_Sel_entity(irn);
287 key.u.intVal = get_Conv_strict(irn);
290 key.u.intVal = is_Div_remainderless(irn);
296 key.mode = get_Load_mode(irn);
303 assert(key.code == get_irn_opcode(irn));
304 assert(key.mode == get_irn_mode(irn));
305 assert(key.arity == get_irn_arity(irn));
307 switch (get_irn_opcode(irn)) {
309 assert(key.u.proj == get_Proj_proj(irn));
312 assert(key.u.ent == get_Sel_entity(irn));
315 assert(key.u.intVal == get_Conv_strict(irn));
318 assert(key.u.intVal == is_Div_remainderless(irn));
321 assert(key.u.block == irn);
324 assert(key.mode == get_Load_mode(irn));
333 static void check_all_partitions(environment_t *env) {
338 for (P = env->dbg_list; P != NULL; P = P->dbg_next) {
340 if (! P->type_is_T_or_C)
342 list_for_each_entry(node_t, node, &P->Follower, node_list) {
343 node_t *leader = identity(node);
345 assert(leader != node && leader->part == node->part);
354 static void do_check_list(const node_t *list, int ofs, const partition_t *Z) {
357 #define NEXT(e) *((const node_t **)((char *)(e) + (ofs)))
358 for (e = list; e != NULL; e = NEXT(e)) {
359 assert(e->part == Z);
362 } /* ido_check_list */
365 * Check a local list.
367 static void check_list(const node_t *list, const partition_t *Z) {
368 do_check_list(list, offsetof(node_t, next), Z);
372 #define check_partition(T)
373 #define check_list(list, Z)
374 #define check_all_partitions(env)
375 #endif /* CHECK_PARTITIONS */
378 static inline lattice_elem_t get_partition_type(const partition_t *X);
381 * Dump partition to output.
383 static void dump_partition(const char *msg, const partition_t *part) {
386 lattice_elem_t type = get_partition_type(part);
388 DB((dbg, LEVEL_2, "%s part%u%s (%u, %+F) {\n ",
389 msg, part->nr, part->type_is_T_or_C ? "*" : "",
390 part->n_leader, type));
391 list_for_each_entry(node_t, node, &part->Leader, node_list) {
392 DB((dbg, LEVEL_2, "%s%+F", first ? "" : ", ", node->node));
395 if (! list_empty(&part->Follower)) {
396 DB((dbg, LEVEL_2, "\n---\n "));
398 list_for_each_entry(node_t, node, &part->Follower, node_list) {
399 DB((dbg, LEVEL_2, "%s%+F", first ? "" : ", ", node->node));
403 DB((dbg, LEVEL_2, "\n}\n"));
404 } /* dump_partition */
409 static void do_dump_list(const char *msg, const node_t *node, int ofs) {
413 #define GET_LINK(p, ofs) *((const node_t **)((char *)(p) + (ofs)))
415 DB((dbg, LEVEL_3, "%s = {\n ", msg));
416 for (p = node; p != NULL; p = GET_LINK(p, ofs)) {
417 DB((dbg, LEVEL_3, "%s%+F", first ? "" : ", ", p->node));
420 DB((dbg, LEVEL_3, "\n}\n"));
428 static void dump_race_list(const char *msg, const node_t *list) {
429 do_dump_list(msg, list, offsetof(node_t, race_next));
430 } /* dump_race_list */
433 * Dumps a local list.
435 static void dump_list(const char *msg, const node_t *list) {
436 do_dump_list(msg, list, offsetof(node_t, next));
440 * Dump all partitions.
442 static void dump_all_partitions(const environment_t *env) {
443 const partition_t *P;
445 DB((dbg, LEVEL_2, "All partitions\n===============\n"));
446 for (P = env->dbg_list; P != NULL; P = P->dbg_next)
447 dump_partition("", P);
448 } /* dump_all_partitions */
453 static void dump_split_list(const partition_t *list) {
454 const partition_t *p;
456 DB((dbg, LEVEL_2, "Split by %s produced = {\n", what_reason));
457 for (p = list; p != NULL; p = p->split_next)
458 DB((dbg, LEVEL_2, "part%u, ", p->nr));
459 DB((dbg, LEVEL_2, "\n}\n"));
460 } /* dump_split_list */
463 * Dump partition and type for a node.
465 static int dump_partition_hook(FILE *F, ir_node *n, ir_node *local) {
466 ir_node *irn = local != NULL ? local : n;
467 node_t *node = get_irn_node(irn);
469 ir_fprintf(F, "info2 : \"partition %u type %+F\"\n", node->part->nr, node->type);
471 } /* dump_partition_hook */
474 #define dump_partition(msg, part)
475 #define dump_race_list(msg, list)
476 #define dump_list(msg, list)
477 #define dump_all_partitions(env)
478 #define dump_split_list(list)
481 #if defined(VERIFY_MONOTONE) && defined (DEBUG_libfirm)
483 * Verify that a type transition is monotone
485 static void verify_type(const lattice_elem_t old_type, node_t *node) {
486 if (old_type.tv == node->type.tv) {
490 if (old_type.tv == tarval_top) {
491 /* from Top down-to is always allowed */
494 if (node->type.tv == tarval_bottom || node->type.tv == tarval_reachable) {
498 panic("combo: wrong translation from %+F to %+F on node %+F", old_type, node->type, node->node);
502 #define verify_type(old_type, node)
506 * Compare two pointer values of a listmap.
508 static int listmap_cmp_ptr(const void *elt, const void *key, size_t size) {
509 const listmap_entry_t *e1 = elt;
510 const listmap_entry_t *e2 = key;
513 return e1->id != e2->id;
514 } /* listmap_cmp_ptr */
517 * Initializes a listmap.
519 * @param map the listmap
521 static void listmap_init(listmap_t *map) {
522 map->map = new_set(listmap_cmp_ptr, 16);
527 * Terminates a listmap.
529 * @param map the listmap
531 static void listmap_term(listmap_t *map) {
536 * Return the associated listmap entry for a given id.
538 * @param map the listmap
539 * @param id the id to search for
541 * @return the associated listmap entry for the given id
543 static listmap_entry_t *listmap_find(listmap_t *map, void *id) {
544 listmap_entry_t key, *entry;
549 entry = set_insert(map->map, &key, sizeof(key), HASH_PTR(id));
551 if (entry->list == NULL) {
552 /* a new entry, put into the list */
553 entry->next = map->values;
560 * Calculate the hash value for an opcode map entry.
562 * @param entry an opcode map entry
564 * @return a hash value for the given opcode map entry
566 static unsigned opcode_hash(const opcode_key_t *entry) {
567 return (entry->mode - (ir_mode *)0) * 9 + entry->code + entry->u.proj * 3 + HASH_PTR(entry->u.ptr) + entry->arity;
571 * Compare two entries in the opcode map.
573 static int cmp_opcode(const void *elt, const void *key, size_t size) {
574 const opcode_key_t *o1 = elt;
575 const opcode_key_t *o2 = key;
578 return o1->code != o2->code || o1->mode != o2->mode ||
579 o1->arity != o2->arity ||
580 o1->u.proj != o2->u.proj ||
581 o1->u.intVal != o2->u.intVal ||
582 o1->u.ptr != o2->u.ptr;
586 * Compare two Def-Use edges for input position.
588 static int cmp_def_use_edge(const void *a, const void *b) {
589 const ir_def_use_edge *ea = a;
590 const ir_def_use_edge *eb = b;
592 /* no overrun, because range is [-1, MAXINT] */
593 return ea->pos - eb->pos;
594 } /* cmp_def_use_edge */
597 * We need the Def-Use edges sorted.
599 static void sort_irn_outs(node_t *node) {
600 ir_node *irn = node->node;
601 int n_outs = get_irn_n_outs(irn);
604 qsort(&irn->out[1], n_outs, sizeof(irn->out[0]), cmp_def_use_edge);
606 node->max_user_input = irn->out[n_outs].pos;
607 } /* sort_irn_outs */
610 * Return the type of a node.
612 * @param irn an IR-node
614 * @return the associated type of this node
616 static inline lattice_elem_t get_node_type(const ir_node *irn) {
617 return get_irn_node(irn)->type;
618 } /* get_node_type */
621 * Return the tarval of a node.
623 * @param irn an IR-node
625 * @return the associated type of this node
627 static inline tarval *get_node_tarval(const ir_node *irn) {
628 lattice_elem_t type = get_node_type(irn);
630 if (is_tarval(type.tv))
632 return tarval_bottom;
633 } /* get_node_type */
636 * Add a partition to the worklist.
638 static inline void add_to_worklist(partition_t *X, environment_t *env) {
639 assert(X->on_worklist == 0);
640 DB((dbg, LEVEL_2, "Adding part%d to worklist\n", X->nr));
641 X->wl_next = env->worklist;
644 } /* add_to_worklist */
647 * Create a new empty partition.
649 * @param env the environment
651 * @return a newly allocated partition
653 static inline partition_t *new_partition(environment_t *env) {
654 partition_t *part = obstack_alloc(&env->obst, sizeof(*part));
656 INIT_LIST_HEAD(&part->Leader);
657 INIT_LIST_HEAD(&part->Follower);
658 INIT_LIST_HEAD(&part->cprop);
659 INIT_LIST_HEAD(&part->cprop_X);
660 part->wl_next = NULL;
661 part->touched_next = NULL;
662 part->cprop_next = NULL;
663 part->split_next = NULL;
664 part->touched = NULL;
667 part->max_user_inputs = 0;
668 part->on_worklist = 0;
669 part->on_touched = 0;
671 part->type_is_T_or_C = 0;
673 part->dbg_next = env->dbg_list;
674 env->dbg_list = part;
675 part->nr = part_nr++;
679 } /* new_partition */
682 * Get the first node from a partition.
684 static inline node_t *get_first_node(const partition_t *X) {
685 return list_entry(X->Leader.next, node_t, node_list);
686 } /* get_first_node */
689 * Return the type of a partition (assuming partition is non-empty and
690 * all elements have the same type).
692 * @param X a partition
694 * @return the type of the first element of the partition
696 static inline lattice_elem_t get_partition_type(const partition_t *X) {
697 const node_t *first = get_first_node(X);
699 } /* get_partition_type */
702 * Creates a partition node for the given IR-node and place it
703 * into the given partition.
705 * @param irn an IR-node
706 * @param part a partition to place the node in
707 * @param env the environment
709 * @return the created node
711 static node_t *create_partition_node(ir_node *irn, partition_t *part, environment_t *env) {
712 /* create a partition node and place it in the partition */
713 node_t *node = obstack_alloc(&env->obst, sizeof(*node));
715 INIT_LIST_HEAD(&node->node_list);
716 INIT_LIST_HEAD(&node->cprop_list);
720 node->race_next = NULL;
721 node->type.tv = tarval_top;
722 node->max_user_input = 0;
724 node->n_followers = 0;
725 node->on_touched = 0;
728 node->is_follower = 0;
730 set_irn_node(irn, node);
732 list_add_tail(&node->node_list, &part->Leader);
736 } /* create_partition_node */
739 * Pre-Walker, initialize all Nodes' type to U or top and place
740 * all nodes into the TOP partition.
742 static void create_initial_partitions(ir_node *irn, void *ctx) {
743 environment_t *env = ctx;
744 partition_t *part = env->initial;
747 node = create_partition_node(irn, part, env);
749 if (node->max_user_input > part->max_user_inputs)
750 part->max_user_inputs = node->max_user_input;
753 set_Block_phis(irn, NULL);
755 } /* create_initial_partitions */
758 * Post-Walker, collect all Block-Phi lists, set Cond.
760 static void init_block_phis(ir_node *irn, void *ctx) {
764 add_Block_phi(get_nodes_block(irn), irn);
766 } /* init_block_phis */
769 * Add a node to the entry.partition.touched set and
770 * node->partition to the touched set if not already there.
773 * @param env the environment
775 static inline void add_to_touched(node_t *y, environment_t *env) {
776 if (y->on_touched == 0) {
777 partition_t *part = y->part;
779 y->next = part->touched;
784 if (part->on_touched == 0) {
785 part->touched_next = env->touched;
787 part->on_touched = 1;
790 check_list(part->touched, part);
792 } /* add_to_touched */
795 * Place a node on the cprop list.
798 * @param env the environment
800 static void add_to_cprop(node_t *y, environment_t *env) {
803 /* Add y to y.partition.cprop. */
804 if (y->on_cprop == 0) {
805 partition_t *Y = y->part;
806 ir_node *irn = y->node;
808 /* place Conds and all its Projs on the cprop_X list */
809 if (is_Cond(skip_Proj(irn)))
810 list_add_tail(&y->cprop_list, &Y->cprop_X);
812 list_add_tail(&y->cprop_list, &Y->cprop);
815 DB((dbg, LEVEL_3, "Add %+F to part%u.cprop\n", y->node, Y->nr));
817 /* place its partition on the cprop list */
818 if (Y->on_cprop == 0) {
819 Y->cprop_next = env->cprop;
825 if (get_irn_mode(irn) == mode_T) {
826 /* mode_T nodes always produce tarval_bottom, so we must explicitly
827 add it's Proj's to get constant evaluation to work */
830 for (i = get_irn_n_outs(irn) - 1; i >= 0; --i) {
831 node_t *proj = get_irn_node(get_irn_out(irn, i));
833 add_to_cprop(proj, env);
835 } else if (is_Block(irn)) {
836 /* Due to the way we handle Phi's, we must place all Phis of a block on the list
837 * if someone placed the block. The Block is only placed if the reachability
838 * changes, and this must be re-evaluated in compute_Phi(). */
840 for (phi = get_Block_phis(irn); phi != NULL; phi = get_Phi_next(phi)) {
841 node_t *p = get_irn_node(phi);
842 add_to_cprop(p, env);
848 * Update the worklist: If Z is on worklist then add Z' to worklist.
849 * Else add the smaller of Z and Z' to worklist.
851 * @param Z the Z partition
852 * @param Z_prime the Z' partition, a previous part of Z
853 * @param env the environment
855 static void update_worklist(partition_t *Z, partition_t *Z_prime, environment_t *env) {
856 if (Z->on_worklist || Z_prime->n_leader < Z->n_leader) {
857 add_to_worklist(Z_prime, env);
859 add_to_worklist(Z, env);
861 } /* update_worklist */
864 * Make all inputs to x no longer be F.def_use edges.
868 static void move_edges_to_leader(node_t *x) {
869 ir_node *irn = x->node;
872 for (i = get_irn_arity(irn) - 1; i >= 0; --i) {
873 node_t *pred = get_irn_node(get_irn_n(irn, i));
878 n = get_irn_n_outs(p);
879 for (j = 1; j <= pred->n_followers; ++j) {
880 if (p->out[j].pos == i && p->out[j].use == irn) {
881 /* found a follower edge to x, move it to the Leader */
882 ir_def_use_edge edge = p->out[j];
884 /* remove this edge from the Follower set */
885 p->out[j] = p->out[pred->n_followers];
888 /* sort it into the leader set */
889 for (k = pred->n_followers + 2; k <= n; ++k) {
890 if (p->out[k].pos >= edge.pos)
892 p->out[k - 1] = p->out[k];
894 /* place the new edge here */
895 p->out[k - 1] = edge;
897 /* edge found and moved */
902 } /* move_edges_to_leader */
905 * Split a partition that has NO followers by a local list.
907 * @param Z partition to split
908 * @param g a (non-empty) node list
909 * @param env the environment
911 * @return a new partition containing the nodes of g
913 static partition_t *split_no_followers(partition_t *Z, node_t *g, environment_t *env) {
914 partition_t *Z_prime;
919 dump_partition("Splitting ", Z);
920 dump_list("by list ", g);
924 /* Remove g from Z. */
925 for (node = g; node != NULL; node = node->next) {
926 assert(node->part == Z);
927 list_del(&node->node_list);
930 assert(n < Z->n_leader);
933 /* Move g to a new partition, Z'. */
934 Z_prime = new_partition(env);
936 for (node = g; node != NULL; node = node->next) {
937 list_add_tail(&node->node_list, &Z_prime->Leader);
938 node->part = Z_prime;
939 if (node->max_user_input > max_input)
940 max_input = node->max_user_input;
942 Z_prime->max_user_inputs = max_input;
943 Z_prime->n_leader = n;
946 check_partition(Z_prime);
948 /* for now, copy the type info tag, it will be adjusted in split_by(). */
949 Z_prime->type_is_T_or_C = Z->type_is_T_or_C;
951 update_worklist(Z, Z_prime, env);
953 dump_partition("Now ", Z);
954 dump_partition("Created new ", Z_prime);
956 } /* split_no_followers */
959 * Make the Follower -> Leader transition for a node.
963 static void follower_to_leader(node_t *n) {
964 assert(n->is_follower == 1);
966 DB((dbg, LEVEL_2, "%+F make the follower -> leader transition\n", n->node));
968 move_edges_to_leader(n);
969 list_del(&n->node_list);
970 list_add_tail(&n->node_list, &n->part->Leader);
972 } /* follower_to_leader */
975 * The environment for one race step.
977 typedef struct step_env {
978 node_t *initial; /**< The initial node list. */
979 node_t *unwalked; /**< The unwalked node list. */
980 node_t *walked; /**< The walked node list. */
981 int index; /**< Next index of Follower use_def edge. */
982 unsigned side; /**< side number. */
986 * Return non-zero, if a input is a real follower
988 * @param irn the node to check
989 * @param input number of the input
991 static int is_real_follower(const ir_node *irn, int input) {
994 switch (get_irn_opcode(irn)) {
997 /* ignore the Confirm bound input */
1003 /* ignore the Mux sel input */
1008 /* dead inputs are not follower edges */
1009 ir_node *block = get_nodes_block(irn);
1010 node_t *pred = get_irn_node(get_Block_cfgpred(block, input));
1012 if (pred->type.tv == tarval_unreachable)
1022 /* only a Sub x,0 / Shift x,0 might be a follower */
1029 pred = get_irn_node(get_irn_n(irn, input));
1030 if (is_tarval(pred->type.tv) && tarval_is_null(pred->type.tv))
1034 pred = get_irn_node(get_irn_n(irn, input));
1035 if (is_tarval(pred->type.tv) && tarval_is_one(pred->type.tv))
1039 pred = get_irn_node(get_irn_n(irn, input));
1040 if (is_tarval(pred->type.tv) && tarval_is_all_one(pred->type.tv))
1044 assert(!"opcode not implemented yet");
1048 } /* is_real_follower */
1051 * Do one step in the race.
1053 static int step(step_env *env) {
1056 if (env->initial != NULL) {
1057 /* Move node from initial to unwalked */
1059 env->initial = n->race_next;
1061 n->race_next = env->unwalked;
1067 while (env->unwalked != NULL) {
1068 /* let n be the first node in unwalked */
1070 while (env->index < n->n_followers) {
1071 const ir_def_use_edge *edge = &n->node->out[1 + env->index];
1073 /* let m be n.F.def_use[index] */
1074 node_t *m = get_irn_node(edge->use);
1076 assert(m->is_follower);
1078 * Some inputs, like the get_Confirm_bound are NOT
1079 * real followers, sort them out.
1081 if (! is_real_follower(m->node, edge->pos)) {
1087 /* only followers from our partition */
1088 if (m->part != n->part)
1091 if ((m->flagged & env->side) == 0) {
1092 m->flagged |= env->side;
1094 if (m->flagged != 3) {
1095 /* visited the first time */
1096 /* add m to unwalked not as first node (we might still need to
1097 check for more follower node */
1098 m->race_next = n->race_next;
1102 /* else already visited by the other side and on the other list */
1105 /* move n to walked */
1106 env->unwalked = n->race_next;
1107 n->race_next = env->walked;
1115 * Clear the flags from a list and check for
1116 * nodes that where touched from both sides.
1118 * @param list the list
1120 static int clear_flags(node_t *list) {
1124 for (n = list; n != NULL; n = n->race_next) {
1125 if (n->flagged == 3) {
1126 /* we reach a follower from both sides, this will split congruent
1127 * inputs and make it a leader. */
1128 follower_to_leader(n);
1137 * Split a partition by a local list using the race.
1139 * @param pX pointer to the partition to split, might be changed!
1140 * @param gg a (non-empty) node list
1141 * @param env the environment
1143 * @return a new partition containing the nodes of gg
1145 static partition_t *split(partition_t **pX, node_t *gg, environment_t *env) {
1146 partition_t *X = *pX;
1147 partition_t *X_prime;
1150 node_t *g, *h, *node, *t;
1151 int max_input, transitions, winner, shf;
1153 DEBUG_ONLY(static int run = 0;)
1155 DB((dbg, LEVEL_2, "Run %d ", run++));
1156 if (list_empty(&X->Follower)) {
1157 /* if the partition has NO follower, we can use the fast
1158 splitting algorithm. */
1159 return split_no_followers(X, gg, env);
1161 /* else do the race */
1163 dump_partition("Splitting ", X);
1164 dump_list("by list ", gg);
1166 INIT_LIST_HEAD(&tmp);
1168 /* Remove gg from X.Leader and put into g */
1170 for (node = gg; node != NULL; node = node->next) {
1171 assert(node->part == X);
1172 assert(node->is_follower == 0);
1174 list_del(&node->node_list);
1175 list_add_tail(&node->node_list, &tmp);
1176 node->race_next = g;
1181 list_for_each_entry(node_t, node, &X->Leader, node_list) {
1182 node->race_next = h;
1185 /* restore X.Leader */
1186 list_splice(&tmp, &X->Leader);
1188 senv[0].initial = g;
1189 senv[0].unwalked = NULL;
1190 senv[0].walked = NULL;
1194 senv[1].initial = h;
1195 senv[1].unwalked = NULL;
1196 senv[1].walked = NULL;
1201 * Some informations on the race that are not stated clearly in Click's
1203 * 1) A follower stays on the side that reach him first.
1204 * 2) If the other side reches a follower, if will be converted to
1205 * a leader. /This must be done after the race is over, else the
1206 * edges we are iterating on are renumbered./
1207 * 3) /New leader might end up on both sides./
1208 * 4) /If one side ends up with new Leaders, we must ensure that
1209 * they can split out by opcode, hence we have to put _every_
1210 * partition with new Leader nodes on the cprop list, as
1211 * opcode splitting is done by split_by() at the end of
1212 * constant propagation./
1215 if (step(&senv[0])) {
1219 if (step(&senv[1])) {
1224 assert(senv[winner].initial == NULL);
1225 assert(senv[winner].unwalked == NULL);
1227 /* clear flags from walked/unwalked */
1229 transitions = clear_flags(senv[0].unwalked) << shf;
1230 transitions |= clear_flags(senv[0].walked) << shf;
1232 transitions |= clear_flags(senv[1].unwalked) << shf;
1233 transitions |= clear_flags(senv[1].walked) << shf;
1235 dump_race_list("winner ", senv[winner].walked);
1237 /* Move walked_{winner} to a new partition, X'. */
1238 X_prime = new_partition(env);
1241 for (node = senv[winner].walked; node != NULL; node = node->race_next) {
1242 list_del(&node->node_list);
1243 node->part = X_prime;
1244 if (node->is_follower) {
1245 list_add_tail(&node->node_list, &X_prime->Follower);
1247 list_add_tail(&node->node_list, &X_prime->Leader);
1250 if (node->max_user_input > max_input)
1251 max_input = node->max_user_input;
1253 X_prime->n_leader = n;
1254 X_prime->max_user_inputs = max_input;
1255 X->n_leader -= X_prime->n_leader;
1257 /* for now, copy the type info tag, it will be adjusted in split_by(). */
1258 X_prime->type_is_T_or_C = X->type_is_T_or_C;
1261 * Even if a follower was not checked by both sides, it might have
1262 * loose its congruence, so we need to check this case for all follower.
1264 list_for_each_entry_safe(node_t, node, t, &X_prime->Follower, node_list) {
1265 if (identity(node) == node) {
1266 follower_to_leader(node);
1272 check_partition(X_prime);
1274 /* X' is the smaller part */
1275 add_to_worklist(X_prime, env);
1278 * If there where follower to leader transitions, ensure that the nodes
1279 * can be split out if necessary.
1281 if (transitions & 1) {
1282 /* place winner partition on the cprop list */
1283 if (X_prime->on_cprop == 0) {
1284 X_prime->cprop_next = env->cprop;
1285 env->cprop = X_prime;
1286 X_prime->on_cprop = 1;
1289 if (transitions & 2) {
1290 /* place other partition on the cprop list */
1291 if (X->on_cprop == 0) {
1292 X->cprop_next = env->cprop;
1298 dump_partition("Now ", X);
1299 dump_partition("Created new ", X_prime);
1301 /* we have to ensure that the partition containing g is returned */
1311 * Returns non-zero if the i'th input of a Phi node is live.
1313 * @param phi a Phi-node
1314 * @param i an input number
1316 * @return non-zero if the i'th input of the given Phi node is live
1318 static int is_live_input(ir_node *phi, int i) {
1320 ir_node *block = get_nodes_block(phi);
1321 ir_node *pred = get_Block_cfgpred(block, i);
1322 lattice_elem_t type = get_node_type(pred);
1324 return type.tv != tarval_unreachable;
1326 /* else it's the control input, always live */
1328 } /* is_live_input */
1331 * Return non-zero if a type is a constant.
1333 static int is_constant_type(lattice_elem_t type) {
1334 if (type.tv != tarval_bottom && type.tv != tarval_top)
1337 } /* is_constant_type */
1340 * Check whether a type is neither Top or a constant.
1341 * Note: U is handled like Top here, R is a constant.
1343 * @param type the type to check
1345 static int type_is_neither_top_nor_const(const lattice_elem_t type) {
1346 if (is_tarval(type.tv)) {
1347 if (type.tv == tarval_top)
1349 if (tarval_is_constant(type.tv))
1356 } /* type_is_neither_top_nor_const */
1359 * Collect nodes to the touched list.
1361 * @param list the list which contains the nodes that must be evaluated
1362 * @param idx the index of the def_use edge to evaluate
1363 * @param env the environment
1365 static void collect_touched(list_head *list, int idx, environment_t *env) {
1367 int end_idx = env->end_idx;
1369 list_for_each_entry(node_t, x, list, node_list) {
1373 /* leader edges start AFTER follower edges */
1374 x->next_edge = x->n_followers + 1;
1376 num_edges = get_irn_n_outs(x->node);
1378 /* for all edges in x.L.def_use_{idx} */
1379 while (x->next_edge <= num_edges) {
1380 const ir_def_use_edge *edge = &x->node->out[x->next_edge];
1383 /* check if we have necessary edges */
1384 if (edge->pos > idx)
1391 /* only non-commutative nodes */
1392 if (env->commutative &&
1393 (idx == 0 || idx == 1) && is_op_commutative(get_irn_op(succ)))
1396 /* ignore the "control input" for non-pinned nodes
1397 if we are running in GCSE mode */
1398 if (idx < end_idx && get_irn_pinned(succ) != op_pin_state_pinned)
1401 y = get_irn_node(succ);
1402 assert(get_irn_n(succ, idx) == x->node);
1404 /* ignore block edges touching followers */
1405 if (idx == -1 && y->is_follower)
1408 if (is_constant_type(y->type)) {
1409 ir_opcode code = get_irn_opcode(succ);
1410 if (code == iro_Sub || code == iro_Cmp)
1411 add_to_cprop(y, env);
1414 /* Partitions of constants should not be split simply because their Nodes have unequal
1415 functions or incongruent inputs. */
1416 if (type_is_neither_top_nor_const(y->type) &&
1417 (! is_Phi(y->node) || is_live_input(y->node, idx))) {
1418 add_to_touched(y, env);
1422 } /* collect_touched */
1425 * Collect commutative nodes to the touched list.
1427 * @param X the partition of the list
1428 * @param list the list which contains the nodes that must be evaluated
1429 * @param env the environment
1431 static void collect_commutative_touched(partition_t *X, list_head *list, environment_t *env) {
1436 list_for_each_entry(node_t, x, list, node_list) {
1439 num_edges = get_irn_n_outs(x->node);
1441 x->next_edge = x->n_followers + 1;
1443 /* for all edges in x.L.def_use_{idx} */
1444 while (x->next_edge <= num_edges) {
1445 const ir_def_use_edge *edge = &x->node->out[x->next_edge];
1448 /* check if we have necessary edges */
1458 /* only commutative nodes */
1459 if (!is_op_commutative(get_irn_op(succ)))
1462 y = get_irn_node(succ);
1463 if (is_constant_type(y->type)) {
1464 ir_opcode code = get_irn_opcode(succ);
1465 if (code == iro_Eor)
1466 add_to_cprop(y, env);
1469 /* Partitions of constants should not be split simply because their Nodes have unequal
1470 functions or incongruent inputs. */
1471 if (type_is_neither_top_nor_const(y->type)) {
1472 int other_idx = edge->pos ^ 1;
1473 node_t *other = get_irn_node(get_irn_n(succ, other_idx));
1474 int equal = X == other->part;
1477 * Note: op(a, a) is NOT congruent to op(a, b).
1478 * So, either all touch nodes must have both inputs congruent,
1479 * or not. We decide this by the first occurred node.
1485 if (both_input == equal)
1486 add_to_touched(y, env);
1490 } /* collect_commutative_touched */
1493 * Split the partitions if caused by the first entry on the worklist.
1495 * @param env the environment
1497 static void cause_splits(environment_t *env) {
1498 partition_t *X, *Z, *N;
1501 /* remove the first partition from the worklist */
1503 env->worklist = X->wl_next;
1506 dump_partition("Cause_split: ", X);
1508 if (env->commutative) {
1509 /* handle commutative nodes first */
1511 /* empty the touched set: already done, just clear the list */
1512 env->touched = NULL;
1514 collect_commutative_touched(X, &X->Leader, env);
1515 collect_commutative_touched(X, &X->Follower, env);
1517 for (Z = env->touched; Z != NULL; Z = N) {
1519 node_t *touched = Z->touched;
1520 unsigned n_touched = Z->n_touched;
1522 assert(Z->touched != NULL);
1524 /* beware, split might change Z */
1525 N = Z->touched_next;
1527 /* remove it from the touched set */
1530 /* Empty local Z.touched. */
1531 for (e = touched; e != NULL; e = e->next) {
1532 assert(e->is_follower == 0);
1538 if (0 < n_touched && n_touched < Z->n_leader) {
1539 DB((dbg, LEVEL_2, "Split part%d by touched\n", Z->nr));
1540 split(&Z, touched, env);
1542 assert(n_touched <= Z->n_leader);
1546 /* combine temporary leader and follower list */
1547 for (idx = -1; idx <= X->max_user_inputs; ++idx) {
1548 /* empty the touched set: already done, just clear the list */
1549 env->touched = NULL;
1551 collect_touched(&X->Leader, idx, env);
1552 collect_touched(&X->Follower, idx, env);
1554 for (Z = env->touched; Z != NULL; Z = N) {
1556 node_t *touched = Z->touched;
1557 unsigned n_touched = Z->n_touched;
1559 assert(Z->touched != NULL);
1561 /* beware, split might change Z */
1562 N = Z->touched_next;
1564 /* remove it from the touched set */
1567 /* Empty local Z.touched. */
1568 for (e = touched; e != NULL; e = e->next) {
1569 assert(e->is_follower == 0);
1575 if (0 < n_touched && n_touched < Z->n_leader) {
1576 DB((dbg, LEVEL_2, "Split part%d by touched\n", Z->nr));
1577 split(&Z, touched, env);
1579 assert(n_touched <= Z->n_leader);
1582 } /* cause_splits */
1585 * Implements split_by_what(): Split a partition by characteristics given
1586 * by the what function.
1588 * @param X the partition to split
1589 * @param What a function returning an Id for every node of the partition X
1590 * @param P a list to store the result partitions
1591 * @param env the environment
1595 static partition_t *split_by_what(partition_t *X, what_func What,
1596 partition_t **P, environment_t *env) {
1599 listmap_entry_t *iter;
1602 /* Let map be an empty mapping from the range of What to (local) list of Nodes. */
1604 list_for_each_entry(node_t, x, &X->Leader, node_list) {
1605 void *id = What(x, env);
1606 listmap_entry_t *entry;
1609 /* input not allowed, ignore */
1612 /* Add x to map[What(x)]. */
1613 entry = listmap_find(&map, id);
1614 x->next = entry->list;
1617 /* Let P be a set of Partitions. */
1619 /* for all sets S except one in the range of map do */
1620 for (iter = map.values; iter != NULL; iter = iter->next) {
1621 if (iter->next == NULL) {
1622 /* this is the last entry, ignore */
1627 /* Add SPLIT( X, S ) to P. */
1628 DB((dbg, LEVEL_2, "Split part%d by WHAT = %s\n", X->nr, what_reason));
1629 R = split(&X, S, env);
1639 } /* split_by_what */
1641 /** lambda n.(n.type) */
1642 static void *lambda_type(const node_t *node, environment_t *env) {
1644 return node->type.tv;
1647 /** lambda n.(n.opcode) */
1648 static void *lambda_opcode(const node_t *node, environment_t *env) {
1649 opcode_key_t key, *entry;
1650 ir_node *irn = node->node;
1652 key.code = get_irn_opcode(irn);
1653 key.mode = get_irn_mode(irn);
1654 key.arity = get_irn_arity(irn);
1658 switch (get_irn_opcode(irn)) {
1660 key.u.proj = get_Proj_proj(irn);
1663 key.u.ent = get_Sel_entity(irn);
1666 key.u.intVal = get_Conv_strict(irn);
1669 key.u.intVal = is_Div_remainderless(irn);
1673 * Some ugliness here: Two Blocks having the same
1674 * IJmp predecessor would be congruent, which of course is wrong.
1675 * We fix it by never letting blocks be congruent
1676 * which cannot be detected by combo either.
1681 key.mode = get_Load_mode(irn);
1687 entry = set_insert(env->opcode2id_map, &key, sizeof(key), opcode_hash(&key));
1689 } /* lambda_opcode */
1691 /** lambda n.(n[i].partition) */
1692 static void *lambda_partition(const node_t *node, environment_t *env) {
1693 ir_node *skipped = skip_Proj(node->node);
1696 int i = env->lambda_input;
1698 if (i >= get_irn_arity(node->node)) {
1700 * We are outside the allowed range: This can happen even
1701 * if we have split by opcode first: doing so might move Followers
1702 * to Leaders and those will have a different opcode!
1703 * Note that in this case the partition is on the cprop list and will be
1709 /* ignore the "control input" for non-pinned nodes
1710 if we are running in GCSE mode */
1711 if (i < env->end_idx && get_irn_pinned(skipped) != op_pin_state_pinned)
1714 pred = i == -1 ? get_irn_n(skipped, i) : get_irn_n(node->node, i);
1715 p = get_irn_node(pred);
1717 } /* lambda_partition */
1719 /** lambda n.(n[i].partition) for commutative nodes */
1720 static void *lambda_commutative_partition(const node_t *node, environment_t *env) {
1721 ir_node *irn = node->node;
1722 ir_node *skipped = skip_Proj(irn);
1723 ir_node *pred, *left, *right;
1725 partition_t *pl, *pr;
1726 int i = env->lambda_input;
1728 if (i >= get_irn_arity(node->node)) {
1730 * We are outside the allowed range: This can happen even
1731 * if we have split by opcode first: doing so might move Followers
1732 * to Leaders and those will have a different opcode!
1733 * Note that in this case the partition is on the cprop list and will be
1739 /* ignore the "control input" for non-pinned nodes
1740 if we are running in GCSE mode */
1741 if (i < env->end_idx && get_irn_pinned(skipped) != op_pin_state_pinned)
1745 pred = get_irn_n(skipped, i);
1746 p = get_irn_node(pred);
1750 if (is_op_commutative(get_irn_op(irn))) {
1751 /* normalize partition order by returning the "smaller" on input 0,
1752 the "bigger" on input 1. */
1753 left = get_binop_left(irn);
1754 pl = get_irn_node(left)->part;
1755 right = get_binop_right(irn);
1756 pr = get_irn_node(right)->part;
1759 return pl < pr ? pl : pr;
1761 return pl > pr ? pl : pr;
1763 /* a not split out Follower */
1764 pred = get_irn_n(irn, i);
1765 p = get_irn_node(pred);
1769 } /* lambda_commutative_partition */
1772 * Returns true if a type is a constant (and NOT Top
1775 static int is_con(const lattice_elem_t type) {
1776 /* be conservative */
1777 if (is_tarval(type.tv))
1778 return tarval_is_constant(type.tv);
1779 return is_entity(type.sym.entity_p);
1783 * Implements split_by().
1785 * @param X the partition to split
1786 * @param env the environment
1788 static void split_by(partition_t *X, environment_t *env) {
1789 partition_t *I, *P = NULL;
1792 dump_partition("split_by", X);
1794 if (X->n_leader == 1) {
1795 /* we have only one leader, no need to split, just check it's type */
1796 node_t *x = get_first_node(X);
1797 X->type_is_T_or_C = x->type.tv == tarval_top || is_con(x->type);
1801 DEBUG_ONLY(what_reason = "lambda n.(n.type)";)
1802 P = split_by_what(X, lambda_type, &P, env);
1805 /* adjust the type tags, we have split partitions by type */
1806 for (I = P; I != NULL; I = I->split_next) {
1807 node_t *x = get_first_node(I);
1808 I->type_is_T_or_C = x->type.tv == tarval_top || is_con(x->type);
1815 if (Y->n_leader > 1) {
1816 /* we do not want split the TOP or constant partitions */
1817 if (! Y->type_is_T_or_C) {
1818 partition_t *Q = NULL;
1820 DEBUG_ONLY(what_reason = "lambda n.(n.opcode)";)
1821 Q = split_by_what(Y, lambda_opcode, &Q, env);
1828 if (Z->n_leader > 1) {
1829 const node_t *first = get_first_node(Z);
1830 int arity = get_irn_arity(first->node);
1832 what_func what = lambda_partition;
1833 DEBUG_ONLY(char buf[64];)
1835 if (env->commutative && is_op_commutative(get_irn_op(first->node)))
1836 what = lambda_commutative_partition;
1839 * BEWARE: during splitting by input 2 for instance we might
1840 * create new partitions which are different by input 1, so collect
1841 * them and split further.
1843 Z->split_next = NULL;
1846 for (input = arity - 1; input >= -1; --input) {
1848 partition_t *Z_prime = R;
1851 if (Z_prime->n_leader > 1) {
1852 env->lambda_input = input;
1853 DEBUG_ONLY(snprintf(buf, sizeof(buf), "lambda n.(n[%d].partition)", input);)
1854 DEBUG_ONLY(what_reason = buf;)
1855 S = split_by_what(Z_prime, what, &S, env);
1858 Z_prime->split_next = S;
1861 } while (R != NULL);
1866 } while (Q != NULL);
1869 } while (P != NULL);
1873 * (Re-)compute the type for a given node.
1875 * @param node the node
1877 static void default_compute(node_t *node) {
1879 ir_node *irn = node->node;
1881 /* if any of the data inputs have type top, the result is type top */
1882 for (i = get_irn_arity(irn) - 1; i >= 0; --i) {
1883 ir_node *pred = get_irn_n(irn, i);
1884 node_t *p = get_irn_node(pred);
1886 if (p->type.tv == tarval_top) {
1887 node->type.tv = tarval_top;
1892 if (get_irn_mode(node->node) == mode_X)
1893 node->type.tv = tarval_reachable;
1895 node->type.tv = computed_value(irn);
1896 } /* default_compute */
1899 * (Re-)compute the type for a Block node.
1901 * @param node the node
1903 static void compute_Block(node_t *node) {
1905 ir_node *block = node->node;
1907 if (block == get_irg_start_block(current_ir_graph) || has_Block_label(block)) {
1908 /* start block and labelled blocks are always reachable */
1909 node->type.tv = tarval_reachable;
1913 for (i = get_Block_n_cfgpreds(block) - 1; i >= 0; --i) {
1914 node_t *pred = get_irn_node(get_Block_cfgpred(block, i));
1916 if (pred->type.tv == tarval_reachable) {
1917 /* A block is reachable, if at least of predecessor is reachable. */
1918 node->type.tv = tarval_reachable;
1922 node->type.tv = tarval_top;
1923 } /* compute_Block */
1926 * (Re-)compute the type for a Bad node.
1928 * @param node the node
1930 static void compute_Bad(node_t *node) {
1931 /* Bad nodes ALWAYS compute Top */
1932 node->type.tv = tarval_top;
1936 * (Re-)compute the type for an Unknown node.
1938 * @param node the node
1940 static void compute_Unknown(node_t *node) {
1941 /* While Unknown nodes should compute Top this is dangerous:
1942 * a Top input to a Cond would lead to BOTH control flows unreachable.
1943 * While this is correct in the given semantics, it would destroy the Firm
1946 * It would be safe to compute Top IF it can be assured, that only Cmp
1947 * nodes are inputs to Conds. We check that first.
1948 * This is the way Frontends typically build Firm, but some optimizations
1949 * (cond_eval for instance) might replace them by Phib's...
1951 node->type.tv = tarval_UNKNOWN;
1952 } /* compute_Unknown */
1955 * (Re-)compute the type for a Jmp node.
1957 * @param node the node
1959 static void compute_Jmp(node_t *node) {
1960 node_t *block = get_irn_node(get_nodes_block(node->node));
1962 node->type = block->type;
1966 * (Re-)compute the type for the Return node.
1968 * @param node the node
1970 static void compute_Return(node_t *node) {
1971 /* The Return node is NOT dead if it is in a reachable block.
1972 * This is already checked in compute(). so we can return
1973 * Reachable here. */
1974 node->type.tv = tarval_reachable;
1975 } /* compute_Return */
1978 * (Re-)compute the type for the End node.
1980 * @param node the node
1982 static void compute_End(node_t *node) {
1983 /* the End node is NOT dead of course */
1984 node->type.tv = tarval_reachable;
1988 * (Re-)compute the type for a Call.
1990 * @param node the node
1992 static void compute_Call(node_t *node) {
1994 * A Call computes always bottom, even if it has Unknown
1997 node->type.tv = tarval_bottom;
1998 } /* compute_Call */
2001 * (Re-)compute the type for a SymConst node.
2003 * @param node the node
2005 static void compute_SymConst(node_t *node) {
2006 ir_node *irn = node->node;
2007 node_t *block = get_irn_node(get_nodes_block(irn));
2009 if (block->type.tv == tarval_unreachable) {
2010 node->type.tv = tarval_top;
2013 switch (get_SymConst_kind(irn)) {
2014 case symconst_addr_ent:
2015 /* case symconst_addr_name: cannot handle this yet */
2016 node->type.sym = get_SymConst_symbol(irn);
2019 node->type.tv = computed_value(irn);
2021 } /* compute_SymConst */
2024 * (Re-)compute the type for a Phi node.
2026 * @param node the node
2028 static void compute_Phi(node_t *node) {
2030 ir_node *phi = node->node;
2031 lattice_elem_t type;
2033 /* if a Phi is in a unreachable block, its type is TOP */
2034 node_t *block = get_irn_node(get_nodes_block(phi));
2036 if (block->type.tv == tarval_unreachable) {
2037 node->type.tv = tarval_top;
2041 /* Phi implements the Meet operation */
2042 type.tv = tarval_top;
2043 for (i = get_Phi_n_preds(phi) - 1; i >= 0; --i) {
2044 node_t *pred = get_irn_node(get_Phi_pred(phi, i));
2045 node_t *pred_X = get_irn_node(get_Block_cfgpred(block->node, i));
2047 if (pred_X->type.tv == tarval_unreachable || pred->type.tv == tarval_top) {
2048 /* ignore TOP inputs: We must check here for unreachable blocks,
2049 because Firm constants live in the Start Block are NEVER Top.
2050 Else, a Phi (1,2) will produce Bottom, even if the 2 for instance
2051 comes from a unreachable input. */
2054 if (pred->type.tv == tarval_bottom) {
2055 node->type.tv = tarval_bottom;
2057 } else if (type.tv == tarval_top) {
2058 /* first constant found */
2060 } else if (type.tv != pred->type.tv) {
2061 /* different constants or tarval_bottom */
2062 node->type.tv = tarval_bottom;
2065 /* else nothing, constants are the same */
2071 * (Re-)compute the type for an Add. Special case: one nodes is a Zero Const.
2073 * @param node the node
2075 static void compute_Add(node_t *node) {
2076 ir_node *sub = node->node;
2077 node_t *l = get_irn_node(get_Add_left(sub));
2078 node_t *r = get_irn_node(get_Add_right(sub));
2079 lattice_elem_t a = l->type;
2080 lattice_elem_t b = r->type;
2083 if (a.tv == tarval_top || b.tv == tarval_top) {
2084 node->type.tv = tarval_top;
2085 } else if (a.tv == tarval_bottom || b.tv == tarval_bottom) {
2086 node->type.tv = tarval_bottom;
2088 /* x + 0 = 0 + x = x, but beware of floating point +0 + -0, so we
2089 must call tarval_add() first to handle this case! */
2090 if (is_tarval(a.tv)) {
2091 if (is_tarval(b.tv)) {
2092 node->type.tv = tarval_add(a.tv, b.tv);
2095 mode = get_tarval_mode(a.tv);
2096 if (a.tv == get_mode_null(mode)) {
2100 } else if (is_tarval(b.tv)) {
2101 mode = get_tarval_mode(b.tv);
2102 if (b.tv == get_mode_null(mode)) {
2107 node->type.tv = tarval_bottom;
2112 * (Re-)compute the type for a Sub. Special case: both nodes are congruent.
2114 * @param node the node
2116 static void compute_Sub(node_t *node) {
2117 ir_node *sub = node->node;
2118 node_t *l = get_irn_node(get_Sub_left(sub));
2119 node_t *r = get_irn_node(get_Sub_right(sub));
2120 lattice_elem_t a = l->type;
2121 lattice_elem_t b = r->type;
2124 if (a.tv == tarval_top || b.tv == tarval_top) {
2125 node->type.tv = tarval_top;
2126 } else if (is_con(a) && is_con(b)) {
2127 if (is_tarval(a.tv) && is_tarval(b.tv)) {
2128 node->type.tv = tarval_sub(a.tv, b.tv, get_irn_mode(sub));
2129 } else if (is_tarval(a.tv) && tarval_is_null(a.tv)) {
2131 } else if (is_tarval(b.tv) && tarval_is_null(b.tv)) {
2134 node->type.tv = tarval_bottom;
2136 } else if (r->part == l->part &&
2137 (!mode_is_float(get_irn_mode(l->node)))) {
2139 * BEWARE: a - a is NOT always 0 for floating Point values, as
2140 * NaN op NaN = NaN, so we must check this here.
2142 ir_mode *mode = get_irn_mode(sub);
2143 tv = get_mode_null(mode);
2145 /* if the node was ONCE evaluated by all constants, but now
2146 this breaks AND we get from the argument partitions a different
2147 result, switch to bottom.
2148 This happens because initially all nodes are in the same partition ... */
2149 if (node->type.tv != tv)
2153 node->type.tv = tarval_bottom;
2158 * (Re-)compute the type for an Eor. Special case: both nodes are congruent.
2160 * @param node the node
2162 static void compute_Eor(node_t *node) {
2163 ir_node *eor = node->node;
2164 node_t *l = get_irn_node(get_Eor_left(eor));
2165 node_t *r = get_irn_node(get_Eor_right(eor));
2166 lattice_elem_t a = l->type;
2167 lattice_elem_t b = r->type;
2170 if (a.tv == tarval_top || b.tv == tarval_top) {
2171 node->type.tv = tarval_top;
2172 } else if (is_con(a) && is_con(b)) {
2173 if (is_tarval(a.tv) && is_tarval(b.tv)) {
2174 node->type.tv = tarval_eor(a.tv, b.tv);
2175 } else if (is_tarval(a.tv) && tarval_is_null(a.tv)) {
2177 } else if (is_tarval(b.tv) && tarval_is_null(b.tv)) {
2180 node->type.tv = tarval_bottom;
2182 } else if (r->part == l->part) {
2183 ir_mode *mode = get_irn_mode(eor);
2184 tv = get_mode_null(mode);
2186 /* if the node was ONCE evaluated by all constants, but now
2187 this breaks AND we get from the argument partitions a different
2188 result, switch to bottom.
2189 This happens because initially all nodes are in the same partition ... */
2190 if (node->type.tv != tv)
2194 node->type.tv = tarval_bottom;
2199 * (Re-)compute the type for Cmp.
2201 * @param node the node
2203 static void compute_Cmp(node_t *node) {
2204 ir_node *cmp = 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;
2210 if (a.tv == tarval_top || b.tv == tarval_top) {
2211 node->type.tv = tarval_top;
2212 } else if (r->part == l->part) {
2213 /* both nodes congruent, we can probably do something */
2214 node->type.tv = tarval_b_true;
2215 } else if (is_con(a) && is_con(b)) {
2216 /* both nodes are constants, we can probably do something */
2217 node->type.tv = tarval_b_true;
2219 node->type.tv = tarval_bottom;
2224 * (Re-)compute the type for a Proj(Cmp).
2226 * @param node the node
2227 * @param cond the predecessor Cmp node
2229 static void compute_Proj_Cmp(node_t *node, ir_node *cmp) {
2230 ir_node *proj = node->node;
2231 node_t *l = get_irn_node(get_Cmp_left(cmp));
2232 node_t *r = get_irn_node(get_Cmp_right(cmp));
2233 lattice_elem_t a = l->type;
2234 lattice_elem_t b = r->type;
2235 pn_Cmp pnc = get_Proj_proj(proj);
2238 if (a.tv == tarval_top || b.tv == tarval_top) {
2239 node->type.tv = tarval_undefined;
2240 } else if (is_con(a) && is_con(b)) {
2241 default_compute(node);
2242 } else if (r->part == l->part &&
2243 (!mode_is_float(get_irn_mode(l->node)) || pnc == pn_Cmp_Lt || pnc == pn_Cmp_Gt)) {
2245 * BEWARE: a == a is NOT always True for floating Point values, as
2246 * NaN != NaN is defined, so we must check this here.
2248 tv = pnc & pn_Cmp_Eq ? tarval_b_true: tarval_b_false;
2250 /* if the node was ONCE evaluated by all constants, but now
2251 this breaks AND we get from the argument partitions a different
2252 result, switch to bottom.
2253 This happens because initially all nodes are in the same partition ... */
2254 if (node->type.tv != tv)
2258 node->type.tv = tarval_bottom;
2260 } /* compute_Proj_Cmp */
2263 * (Re-)compute the type for a Proj(Cond).
2265 * @param node the node
2266 * @param cond the predecessor Cond node
2268 static void compute_Proj_Cond(node_t *node, ir_node *cond) {
2269 ir_node *proj = node->node;
2270 long pnc = get_Proj_proj(proj);
2271 ir_node *sel = get_Cond_selector(cond);
2272 node_t *selector = get_irn_node(sel);
2275 * Note: it is crucial for the monotony that the Proj(Cond)
2276 * are evaluates after all predecessors of the Cond selector are
2282 * Due to the fact that 0 is a const, the Cmp gets immediately
2283 * on the cprop list. It will be evaluated before x is evaluated,
2284 * might leaving x as Top. When later x is evaluated, the Cmp
2285 * might change its value.
2286 * BUT if the Cond is evaluated before this happens, Proj(Cond, FALSE)
2287 * gets R, and later changed to F if Cmp is evaluated to True!
2289 * We prevent this by putting Conds in an extra cprop_X queue, which
2290 * gets evaluated after the cprop queue is empty.
2292 * Note that this even happens with Click's original algorithm, if
2293 * Cmp(x, 0) is evaluated to True first and later changed to False
2294 * if x was Top first and later changed to a Const ...
2295 * It is unclear how Click solved that problem ...
2297 * However, in rare cases even this does not help, if a Top reaches
2298 * a compare through a Phi, than Proj(Cond) is evaluated changing
2299 * the type of the Phi to something other.
2300 * So, we take the last resort and bind the type to R once
2303 * (This might be even the way Click works around the whole problem).
2305 * Finally, we may miss some optimization possibilities due to this:
2310 * If Top reaches the if first, than we decide for != here.
2311 * If y later is evaluated to 0, we cannot revert this decision
2312 * and must live with both outputs enabled. If this happens,
2313 * we get an unresolved if (true) in the code ...
2315 * In Click's version where this decision is done at the Cmp,
2316 * the Cmp is NOT optimized away than (if y evaluated to 1
2317 * for instance) and we get a if (1 == 0) here ...
2319 * Both solutions are suboptimal.
2320 * At least, we could easily detect this problem and run
2321 * cf_opt() (or even combo) again :-(
2323 if (node->type.tv == tarval_reachable)
2326 if (get_irn_mode(sel) == mode_b) {
2328 if (pnc == pn_Cond_true) {
2329 if (selector->type.tv == tarval_b_false) {
2330 node->type.tv = tarval_unreachable;
2331 } else if (selector->type.tv == tarval_b_true) {
2332 node->type.tv = tarval_reachable;
2333 } else if (selector->type.tv == tarval_bottom) {
2334 node->type.tv = tarval_reachable;
2336 assert(selector->type.tv == tarval_top);
2337 if (tarval_UNKNOWN == tarval_top) {
2338 /* any condition based on Top is "!=" */
2339 node->type.tv = tarval_unreachable;
2341 node->type.tv = tarval_unreachable;
2345 assert(pnc == pn_Cond_false);
2347 if (selector->type.tv == tarval_b_false) {
2348 node->type.tv = tarval_reachable;
2349 } else if (selector->type.tv == tarval_b_true) {
2350 node->type.tv = tarval_unreachable;
2351 } else if (selector->type.tv == tarval_bottom) {
2352 node->type.tv = tarval_reachable;
2354 assert(selector->type.tv == tarval_top);
2355 if (tarval_UNKNOWN == tarval_top) {
2356 /* any condition based on Top is "!=" */
2357 node->type.tv = tarval_reachable;
2359 node->type.tv = tarval_unreachable;
2365 if (selector->type.tv == tarval_bottom) {
2366 node->type.tv = tarval_reachable;
2367 } else if (selector->type.tv == tarval_top) {
2368 if (tarval_UNKNOWN == tarval_top &&
2369 pnc == get_Cond_defaultProj(cond)) {
2370 /* a switch based of Top is always "default" */
2371 node->type.tv = tarval_reachable;
2373 node->type.tv = tarval_unreachable;
2376 long value = get_tarval_long(selector->type.tv);
2377 if (pnc == get_Cond_defaultProj(cond)) {
2378 /* default switch, have to check ALL other cases */
2381 for (i = get_irn_n_outs(cond) - 1; i >= 0; --i) {
2382 ir_node *succ = get_irn_out(cond, i);
2386 if (value == get_Proj_proj(succ)) {
2387 /* we found a match, will NOT take the default case */
2388 node->type.tv = tarval_unreachable;
2392 /* all cases checked, no match, will take default case */
2393 node->type.tv = tarval_reachable;
2396 node->type.tv = value == pnc ? tarval_reachable : tarval_unreachable;
2400 } /* compute_Proj_Cond */
2403 * (Re-)compute the type for a Proj-Node.
2405 * @param node the node
2407 static void compute_Proj(node_t *node) {
2408 ir_node *proj = node->node;
2409 ir_mode *mode = get_irn_mode(proj);
2410 node_t *block = get_irn_node(get_nodes_block(skip_Proj(proj)));
2411 ir_node *pred = get_Proj_pred(proj);
2413 if (block->type.tv == tarval_unreachable) {
2414 /* a Proj in a unreachable Block stay Top */
2415 node->type.tv = tarval_top;
2418 if (get_irn_node(pred)->type.tv == tarval_top && !is_Cond(pred)) {
2419 /* if the predecessor is Top, its Proj follow */
2420 node->type.tv = tarval_top;
2424 if (mode == mode_M) {
2425 /* mode M is always bottom */
2426 node->type.tv = tarval_bottom;
2429 if (mode != mode_X) {
2431 compute_Proj_Cmp(node, pred);
2433 default_compute(node);
2436 /* handle mode_X nodes */
2438 switch (get_irn_opcode(pred)) {
2440 /* the Proj_X from the Start is always reachable.
2441 However this is already handled at the top. */
2442 node->type.tv = tarval_reachable;
2445 compute_Proj_Cond(node, pred);
2448 default_compute(node);
2450 } /* compute_Proj */
2453 * (Re-)compute the type for a Confirm.
2455 * @param node the node
2457 static void compute_Confirm(node_t *node) {
2458 ir_node *confirm = node->node;
2459 node_t *pred = get_irn_node(get_Confirm_value(confirm));
2461 if (get_Confirm_cmp(confirm) == pn_Cmp_Eq) {
2462 node_t *bound = get_irn_node(get_Confirm_bound(confirm));
2464 if (is_con(bound->type)) {
2465 /* is equal to a constant */
2466 node->type = bound->type;
2470 /* a Confirm is a copy OR a Const */
2471 node->type = pred->type;
2472 } /* compute_Confirm */
2475 * (Re-)compute the type for a given node.
2477 * @param node the node
2479 static void compute(node_t *node) {
2480 ir_node *irn = node->node;
2483 #ifndef VERIFY_MONOTONE
2485 * Once a node reaches bottom, the type cannot fall further
2486 * in the lattice and we can stop computation.
2487 * Do not take this exit if the monotony verifier is
2488 * enabled to catch errors.
2490 if (node->type.tv == tarval_bottom)
2494 if (is_no_Block(irn)) {
2495 /* for pinned nodes, check its control input */
2496 if (get_irn_pinned(skip_Proj(irn)) == op_pin_state_pinned) {
2497 node_t *block = get_irn_node(get_nodes_block(irn));
2499 if (block->type.tv == tarval_unreachable) {
2500 node->type.tv = tarval_top;
2506 func = (compute_func)node->node->op->ops.generic;
2512 * Identity functions: Note that one might thing that identity() is just a
2513 * synonym for equivalent_node(). While this is true, we cannot use it for the algorithm
2514 * here, because it expects that the identity node is one of the inputs, which is NOT
2515 * always true for equivalent_node() which can handle (and does sometimes) DAGs.
2516 * So, we have our own implementation, which copies some parts of equivalent_node()
2520 * Calculates the Identity for Phi nodes
2522 static node_t *identity_Phi(node_t *node) {
2523 ir_node *phi = node->node;
2524 ir_node *block = get_nodes_block(phi);
2525 node_t *n_part = NULL;
2528 for (i = get_Phi_n_preds(phi) - 1; i >= 0; --i) {
2529 node_t *pred_X = get_irn_node(get_Block_cfgpred(block, i));
2531 if (pred_X->type.tv == tarval_reachable) {
2532 node_t *pred = get_irn_node(get_Phi_pred(phi, i));
2536 else if (n_part->part != pred->part) {
2537 /* incongruent inputs, not a follower */
2542 /* if n_part is NULL here, all inputs path are dead, the Phi computes
2543 * tarval_top, is in the TOP partition and should NOT being split! */
2544 assert(n_part != NULL);
2546 } /* identity_Phi */
2549 * Calculates the Identity for commutative 0 neutral nodes.
2551 static node_t *identity_comm_zero_binop(node_t *node) {
2552 ir_node *op = node->node;
2553 node_t *a = get_irn_node(get_binop_left(op));
2554 node_t *b = get_irn_node(get_binop_right(op));
2555 ir_mode *mode = get_irn_mode(op);
2558 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
2559 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
2562 /* node: no input should be tarval_top, else the binop would be also
2563 * Top and not being split. */
2564 zero = get_mode_null(mode);
2565 if (a->type.tv == zero)
2567 if (b->type.tv == zero)
2570 } /* identity_comm_zero_binop */
2573 * Calculates the Identity for Shift nodes.
2575 static node_t *identity_shift(node_t *node) {
2576 ir_node *op = node->node;
2577 node_t *b = get_irn_node(get_binop_right(op));
2578 ir_mode *mode = get_irn_mode(b->node);
2581 /* node: no input should be tarval_top, else the binop would be also
2582 * Top and not being split. */
2583 zero = get_mode_null(mode);
2584 if (b->type.tv == zero)
2585 return get_irn_node(get_binop_left(op));
2587 } /* identity_shift */
2590 * Calculates the Identity for Mul nodes.
2592 static node_t *identity_Mul(node_t *node) {
2593 ir_node *op = node->node;
2594 node_t *a = get_irn_node(get_Mul_left(op));
2595 node_t *b = get_irn_node(get_Mul_right(op));
2596 ir_mode *mode = get_irn_mode(op);
2599 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
2600 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
2603 /* node: no input should be tarval_top, else the binop would be also
2604 * Top and not being split. */
2605 one = get_mode_one(mode);
2606 if (a->type.tv == one)
2608 if (b->type.tv == one)
2611 } /* identity_Mul */
2614 * Calculates the Identity for Sub nodes.
2616 static node_t *identity_Sub(node_t *node) {
2617 ir_node *sub = node->node;
2618 node_t *b = get_irn_node(get_Sub_right(sub));
2619 ir_mode *mode = get_irn_mode(sub);
2621 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
2622 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
2625 /* node: no input should be tarval_top, else the binop would be also
2626 * Top and not being split. */
2627 if (b->type.tv == get_mode_null(mode))
2628 return get_irn_node(get_Sub_left(sub));
2630 } /* identity_Sub */
2633 * Calculates the Identity for And nodes.
2635 static node_t *identity_And(node_t *node) {
2636 ir_node *and = node->node;
2637 node_t *a = get_irn_node(get_And_left(and));
2638 node_t *b = get_irn_node(get_And_right(and));
2639 tarval *neutral = get_mode_all_one(get_irn_mode(and));
2641 /* node: no input should be tarval_top, else the And would be also
2642 * Top and not being split. */
2643 if (a->type.tv == neutral)
2645 if (b->type.tv == neutral)
2648 } /* identity_And */
2651 * Calculates the Identity for Confirm nodes.
2653 static node_t *identity_Confirm(node_t *node) {
2654 ir_node *confirm = node->node;
2656 /* a Confirm is always a Copy */
2657 return get_irn_node(get_Confirm_value(confirm));
2658 } /* identity_Confirm */
2661 * Calculates the Identity for Mux nodes.
2663 static node_t *identity_Mux(node_t *node) {
2664 ir_node *mux = node->node;
2665 node_t *t = get_irn_node(get_Mux_true(mux));
2666 node_t *f = get_irn_node(get_Mux_false(mux));
2669 if (t->part == f->part)
2672 /* for now, the 1-input identity is not supported */
2674 sel = get_irn_node(get_Mux_sel(mux));
2676 /* Mux sel input is mode_b, so it is always a tarval */
2677 if (sel->type.tv == tarval_b_true)
2679 if (sel->type.tv == tarval_b_false)
2683 } /* identity_Mux */
2686 * Calculates the Identity for nodes.
2688 static node_t *identity(node_t *node) {
2689 ir_node *irn = node->node;
2691 switch (get_irn_opcode(irn)) {
2693 return identity_Phi(node);
2695 return identity_Mul(node);
2699 return identity_comm_zero_binop(node);
2704 return identity_shift(node);
2706 return identity_And(node);
2708 return identity_Sub(node);
2710 return identity_Confirm(node);
2712 return identity_Mux(node);
2719 * Node follower is a (new) follower of leader, segregate Leader
2722 static void segregate_def_use_chain_1(const ir_node *follower, node_t *leader) {
2723 ir_node *l = leader->node;
2724 int j, i, n = get_irn_n_outs(l);
2726 DB((dbg, LEVEL_2, "%+F is a follower of %+F\n", follower, leader->node));
2727 /* The leader edges must remain sorted, but follower edges can
2729 for (i = leader->n_followers + 1; i <= n; ++i) {
2730 if (l->out[i].use == follower) {
2731 ir_def_use_edge t = l->out[i];
2733 for (j = i - 1; j >= leader->n_followers + 1; --j)
2734 l->out[j + 1] = l->out[j];
2735 ++leader->n_followers;
2736 l->out[leader->n_followers] = t;
2740 } /* segregate_def_use_chain_1 */
2743 * Node follower is a (new) follower segregate its Leader
2746 * @param follower the follower IR node
2748 static void segregate_def_use_chain(const ir_node *follower) {
2751 for (i = get_irn_arity(follower) - 1; i >= 0; --i) {
2752 node_t *pred = get_irn_node(get_irn_n(follower, i));
2754 segregate_def_use_chain_1(follower, pred);
2756 } /* segregate_def_use_chain */
2759 * Propagate constant evaluation.
2761 * @param env the environment
2763 static void propagate(environment_t *env) {
2766 lattice_elem_t old_type;
2768 unsigned n_fallen, old_type_was_T_or_C;
2771 while (env->cprop != NULL) {
2772 void *oldopcode = NULL;
2774 /* remove the first partition X from cprop */
2777 env->cprop = X->cprop_next;
2779 old_type_was_T_or_C = X->type_is_T_or_C;
2781 DB((dbg, LEVEL_2, "Propagate type on part%d\n", X->nr));
2785 int cprop_empty = list_empty(&X->cprop);
2786 int cprop_X_empty = list_empty(&X->cprop_X);
2788 if (cprop_empty && cprop_X_empty) {
2789 /* both cprop lists are empty */
2793 /* remove the first Node x from X.cprop */
2795 /* Get a node from the cprop_X list only if
2796 * all data nodes are processed.
2797 * This ensures, that all inputs of the Cond
2798 * predecessor are processed if its type is still Top.
2800 x = list_entry(X->cprop_X.next, node_t, cprop_list);
2802 x = list_entry(X->cprop.next, node_t, cprop_list);
2805 //assert(x->part == X);
2806 list_del(&x->cprop_list);
2809 if (x->is_follower && identity(x) == x) {
2810 /* check the opcode first */
2811 if (oldopcode == NULL) {
2812 oldopcode = lambda_opcode(get_first_node(X), env);
2814 if (oldopcode != lambda_opcode(x, env)) {
2815 if (x->on_fallen == 0) {
2816 /* different opcode -> x falls out of this partition */
2821 DB((dbg, LEVEL_2, "Add node %+F to fallen\n", x->node));
2825 /* x will make the follower -> leader transition */
2826 follower_to_leader(x);
2829 /* compute a new type for x */
2831 DB((dbg, LEVEL_3, "computing type of %+F\n", x->node));
2833 if (x->type.tv != old_type.tv) {
2834 DB((dbg, LEVEL_2, "node %+F has changed type from %+F to %+F\n", x->node, old_type, x->type));
2835 verify_type(old_type, x);
2837 if (x->on_fallen == 0) {
2838 /* Add x to fallen. Nodes might fall from T -> const -> _|_, so check that they are
2839 not already on the list. */
2844 DB((dbg, LEVEL_2, "Add node %+F to fallen\n", x->node));
2846 for (i = get_irn_n_outs(x->node) - 1; i >= 0; --i) {
2847 ir_node *succ = get_irn_out(x->node, i);
2848 node_t *y = get_irn_node(succ);
2850 /* Add y to y.partition.cprop. */
2851 add_to_cprop(y, env);
2856 if (n_fallen > 0 && n_fallen != X->n_leader) {
2857 DB((dbg, LEVEL_2, "Splitting part%d by fallen\n", X->nr));
2858 Y = split(&X, fallen, env);
2860 * We have split out fallen node. The type of the result
2861 * partition is NOT set yet.
2863 Y->type_is_T_or_C = 0;
2867 /* remove the flags from the fallen list */
2868 for (x = fallen; x != NULL; x = x->next)
2871 if (old_type_was_T_or_C) {
2874 /* check if some nodes will make the leader -> follower transition */
2875 list_for_each_entry_safe(node_t, y, tmp, &Y->Leader, node_list) {
2876 if (y->type.tv != tarval_top && ! is_con(y->type)) {
2877 node_t *eq_node = identity(y);
2879 if (eq_node != y && eq_node->part == y->part) {
2880 DB((dbg, LEVEL_2, "Node %+F is a follower of %+F\n", y->node, eq_node->node));
2881 /* move to Follower */
2883 list_del(&y->node_list);
2884 list_add_tail(&y->node_list, &Y->Follower);
2887 segregate_def_use_chain(y->node);
2897 * Get the leader for a given node from its congruence class.
2899 * @param irn the node
2901 static ir_node *get_leader(node_t *node) {
2902 partition_t *part = node->part;
2904 if (part->n_leader > 1 || node->is_follower) {
2905 if (node->is_follower) {
2906 DB((dbg, LEVEL_2, "Replacing follower %+F\n", node->node));
2909 DB((dbg, LEVEL_2, "Found congruence class for %+F\n", node->node));
2911 return get_first_node(part)->node;
2917 * Returns non-zero if a mode_T node has only one reachable output.
2919 static int only_one_reachable_proj(ir_node *n) {
2922 for (i = get_irn_n_outs(n) - 1; i >= 0; --i) {
2923 ir_node *proj = get_irn_out(n, i);
2926 /* skip non-control flow Proj's */
2927 if (get_irn_mode(proj) != mode_X)
2930 node = get_irn_node(proj);
2931 if (node->type.tv == tarval_reachable) {
2937 } /* only_one_reachable_proj */
2940 * Return non-zero if the control flow predecessor node pred
2941 * is the only reachable control flow exit of its block.
2943 * @param pred the control flow exit
2944 * @param block the destination block
2946 static int can_exchange(ir_node *pred, ir_node *block) {
2947 if (is_Start(pred) || has_Block_label(block))
2949 else if (is_Jmp(pred))
2951 else if (get_irn_mode(pred) == mode_T) {
2952 /* if the predecessor block has more than one
2953 reachable outputs we cannot remove the block */
2954 return only_one_reachable_proj(pred);
2957 } /* can_exchange */
2960 * Block Post-Walker, apply the analysis results on control flow by
2961 * shortening Phi's and Block inputs.
2963 static void apply_cf(ir_node *block, void *ctx) {
2964 environment_t *env = ctx;
2965 node_t *node = get_irn_node(block);
2967 ir_node **ins, **in_X;
2968 ir_node *phi, *next;
2970 n = get_Block_n_cfgpreds(block);
2972 if (node->type.tv == tarval_unreachable) {
2975 for (i = n - 1; i >= 0; --i) {
2976 ir_node *pred = get_Block_cfgpred(block, i);
2978 if (! is_Bad(pred)) {
2979 node_t *pred_bl = get_irn_node(get_nodes_block(skip_Proj(pred)));
2981 if (pred_bl->flagged == 0) {
2982 pred_bl->flagged = 3;
2984 if (pred_bl->type.tv == tarval_reachable) {
2986 * We will remove an edge from block to its pred.
2987 * This might leave the pred block as an endless loop
2989 if (! is_backedge(block, i))
2990 keep_alive(pred_bl->node);
2996 /* the EndBlock is always reachable even if the analysis
2997 finds out the opposite :-) */
2998 if (block != get_irg_end_block(current_ir_graph)) {
2999 /* mark dead blocks */
3000 set_Block_dead(block);
3001 DB((dbg, LEVEL_1, "Removing dead %+F\n", block));
3003 /* the endblock is unreachable */
3004 set_irn_in(block, 0, NULL);
3010 /* only one predecessor combine */
3011 ir_node *pred = skip_Proj(get_Block_cfgpred(block, 0));
3013 if (can_exchange(pred, block)) {
3014 ir_node *new_block = get_nodes_block(pred);
3015 DB((dbg, LEVEL_1, "Fuse %+F with %+F\n", block, new_block));
3016 DBG_OPT_COMBO(block, new_block, FS_OPT_COMBO_CF);
3017 exchange(block, new_block);
3018 node->node = new_block;
3024 NEW_ARR_A(ir_node *, in_X, n);
3026 for (i = 0; i < n; ++i) {
3027 ir_node *pred = get_Block_cfgpred(block, i);
3028 node_t *node = get_irn_node(pred);
3030 if (node->type.tv == tarval_reachable) {
3033 DB((dbg, LEVEL_1, "Removing dead input %d from %+F (%+F)\n", i, block, pred));
3034 if (! is_Bad(pred)) {
3035 node_t *pred_bl = get_irn_node(get_nodes_block(skip_Proj(pred)));
3037 if (pred_bl->flagged == 0) {
3038 pred_bl->flagged = 3;
3040 if (pred_bl->type.tv == tarval_reachable) {
3042 * We will remove an edge from block to its pred.
3043 * This might leave the pred block as an endless loop
3045 if (! is_backedge(block, i))
3046 keep_alive(pred_bl->node);
3056 NEW_ARR_A(ir_node *, ins, n);
3057 for (phi = get_Block_phis(block); phi != NULL; phi = next) {
3058 node_t *node = get_irn_node(phi);
3060 next = get_Phi_next(phi);
3061 if (is_tarval(node->type.tv) && tarval_is_constant(node->type.tv)) {
3062 /* this Phi is replaced by a constant */
3063 tarval *tv = node->type.tv;
3064 ir_node *c = new_Const(tv);
3066 set_irn_node(c, node);
3068 DB((dbg, LEVEL_1, "%+F is replaced by %+F\n", phi, c));
3069 DBG_OPT_COMBO(phi, c, FS_OPT_COMBO_CONST);
3074 for (i = 0; i < n; ++i) {
3075 node_t *pred = get_irn_node(get_Block_cfgpred(block, i));
3077 if (pred->type.tv == tarval_reachable) {
3078 ins[j++] = get_Phi_pred(phi, i);
3082 /* this Phi is replaced by a single predecessor */
3083 ir_node *s = ins[0];
3084 node_t *phi_node = get_irn_node(phi);
3087 DB((dbg, LEVEL_1, "%+F is replaced by %+F because of cf change\n", phi, s));
3088 DBG_OPT_COMBO(phi, s, FS_OPT_COMBO_FOLLOWER);
3093 set_irn_in(phi, j, ins);
3101 /* this Block has only one live predecessor */
3102 ir_node *pred = skip_Proj(in_X[0]);
3104 if (can_exchange(pred, block)) {
3105 ir_node *new_block = get_nodes_block(pred);
3106 DBG_OPT_COMBO(block, new_block, FS_OPT_COMBO_CF);
3107 exchange(block, new_block);
3108 node->node = new_block;
3113 set_irn_in(block, k, in_X);
3118 * Exchange a node by its leader.
3119 * Beware: in rare cases the mode might be wrong here, for instance
3120 * AddP(x, NULL) is a follower of x, but with different mode.
3123 static void exchange_leader(ir_node *irn, ir_node *leader) {
3124 ir_mode *mode = get_irn_mode(irn);
3125 if (mode != get_irn_mode(leader)) {
3126 /* The conv is a no-op, so we are free to place it
3127 * either in the block of the leader OR in irn's block.
3128 * Probably placing it into leaders block might reduce
3129 * the number of Conv due to CSE. */
3130 ir_node *block = get_nodes_block(leader);
3131 dbg_info *dbg = get_irn_dbg_info(irn);
3133 leader = new_rd_Conv(dbg, current_ir_graph, block, leader, mode);
3135 exchange(irn, leader);
3136 } /* exchange_leader */
3139 * Check, if all users of a mode_M node are dead. Use
3140 * the Def-Use edges for this purpose, as they still
3141 * reflect the situation.
3143 static int all_users_are_dead(const ir_node *irn) {
3144 int i, n = get_irn_n_outs(irn);
3146 for (i = 1; i <= n; ++i) {
3147 const ir_node *succ = irn->out[i].use;
3148 const node_t *block = get_irn_node(get_nodes_block(succ));
3151 if (block->type.tv == tarval_unreachable) {
3152 /* block is unreachable */
3155 node = get_irn_node(succ);
3156 if (node->type.tv != tarval_top) {
3157 /* found a reachable user */
3161 /* all users are unreachable */
3163 } /* all_user_are_dead */
3166 * Walker: Find reachable mode_M nodes that have only
3167 * unreachable users. These nodes must be kept later.
3169 static void find_kept_memory(ir_node *irn, void *ctx) {
3170 environment_t *env = ctx;
3171 node_t *node, *block;
3173 if (get_irn_mode(irn) != mode_M)
3176 block = get_irn_node(get_nodes_block(irn));
3177 if (block->type.tv == tarval_unreachable)
3180 node = get_irn_node(irn);
3181 if (node->type.tv == tarval_top)
3184 /* ok, we found a live memory node. */
3185 if (all_users_are_dead(irn)) {
3186 DB((dbg, LEVEL_1, "%+F must be kept\n", irn));
3187 ARR_APP1(ir_node *, env->kept_memory, irn);
3189 } /* find_kept_memory */
3192 * Post-Walker, apply the analysis results;
3194 static void apply_result(ir_node *irn, void *ctx) {
3195 environment_t *env = ctx;
3196 node_t *node = get_irn_node(irn);
3198 if (is_Block(irn) || is_End(irn) || is_Bad(irn)) {
3199 /* blocks already handled, do not touch the End node */
3201 node_t *block = get_irn_node(get_nodes_block(irn));
3203 if (block->type.tv == tarval_unreachable) {
3204 ir_node *bad = get_irg_bad(current_ir_graph);
3206 /* here, bad might already have a node, but this can be safely ignored
3207 as long as bad has at least ONE valid node */
3208 set_irn_node(bad, node);
3210 DB((dbg, LEVEL_1, "%+F is unreachable\n", irn));
3213 } else if (node->type.tv == tarval_top) {
3214 ir_mode *mode = get_irn_mode(irn);
3216 if (mode == mode_M) {
3217 /* never kill a mode_M node */
3219 ir_node *pred = get_Proj_pred(irn);
3220 node_t *pnode = get_irn_node(pred);
3222 if (pnode->type.tv == tarval_top) {
3223 /* skip the predecessor */
3224 ir_node *mem = get_memop_mem(pred);
3226 DB((dbg, LEVEL_1, "%+F computes Top, replaced by %+F\n", irn, mem));
3231 /* leave other nodes, especially PhiM */
3232 } else if (mode == mode_T) {
3233 /* Do not kill mode_T nodes, kill their Projs */
3234 } else if (! is_Unknown(irn)) {
3235 /* don't kick away Unknown's, they might be still needed */
3236 ir_node *unk = new_r_Unknown(current_ir_graph, mode);
3238 /* control flow should already be handled at apply_cf() */
3239 assert(mode != mode_X);
3241 /* see comment above */
3242 set_irn_node(unk, node);
3244 DB((dbg, LEVEL_1, "%+F computes Top\n", irn));
3249 else if (get_irn_mode(irn) == mode_X) {
3252 ir_node *cond = get_Proj_pred(irn);
3254 if (is_Cond(cond)) {
3255 if (only_one_reachable_proj(cond)) {
3256 ir_node *jmp = new_r_Jmp(current_ir_graph, block->node);
3257 set_irn_node(jmp, node);
3259 DB((dbg, LEVEL_1, "%+F is replaced by %+F\n", irn, jmp));
3260 DBG_OPT_COMBO(irn, jmp, FS_OPT_COMBO_CF);
3264 node_t *sel = get_irn_node(get_Cond_selector(cond));
3265 tarval *tv = sel->type.tv;
3267 if (is_tarval(tv) && tarval_is_constant(tv)) {
3268 /* The selector is a constant, but more
3269 * than one output is active: An unoptimized
3277 /* normal data node */
3278 if (is_tarval(node->type.tv) && tarval_is_constant(node->type.tv)) {
3279 tarval *tv = node->type.tv;
3282 * Beware: never replace mode_T nodes by constants. Currently we must mark
3283 * mode_T nodes with constants, but do NOT replace them.
3285 if (! is_Const(irn) && get_irn_mode(irn) != mode_T) {
3286 /* can be replaced by a constant */
3287 ir_node *c = new_Const(tv);
3288 set_irn_node(c, node);
3290 DB((dbg, LEVEL_1, "%+F is replaced by %+F\n", irn, c));
3291 DBG_OPT_COMBO(irn, c, FS_OPT_COMBO_CONST);
3292 exchange_leader(irn, c);
3295 } else if (is_entity(node->type.sym.entity_p)) {
3296 if (! is_SymConst(irn)) {
3297 /* can be replaced by a SymConst */
3298 ir_node *symc = new_r_SymConst(current_ir_graph, block->node, get_irn_mode(irn), node->type.sym, symconst_addr_ent);
3299 set_irn_node(symc, node);
3302 DB((dbg, LEVEL_1, "%+F is replaced by %+F\n", irn, symc));
3303 DBG_OPT_COMBO(irn, symc, FS_OPT_COMBO_CONST);
3304 exchange_leader(irn, symc);
3307 } else if (is_Confirm(irn)) {
3308 /* Confirms are always follower, but do not kill them here */
3310 ir_node *leader = get_leader(node);
3312 if (leader != irn) {
3313 int non_strict_phi = 0;
3316 * Beware: Do not remove Phi(Unknown, ..., x, ..., Unknown)
3317 * as this might create non-strict programs.
3319 if (node->is_follower && is_Phi(irn) && !is_Unknown(leader)) {
3322 for (i = get_Phi_n_preds(irn) - 1; i >= 0; --i) {
3323 ir_node *pred = get_Phi_pred(irn, i);
3325 if (is_Unknown(pred)) {
3331 if (! non_strict_phi) {
3332 DB((dbg, LEVEL_1, "%+F from part%d is replaced by %+F\n", irn, node->part->nr, leader));
3333 if (node->is_follower)
3334 DBG_OPT_COMBO(irn, leader, FS_OPT_COMBO_FOLLOWER);
3336 DBG_OPT_COMBO(irn, leader, FS_OPT_COMBO_CONGRUENT);
3337 exchange_leader(irn, leader);
3344 } /* apply_result */
3347 * Fix the keep-alives by deleting unreachable ones.
3349 static void apply_end(ir_node *end, environment_t *env) {
3350 int i, j, n = get_End_n_keepalives(end);
3354 NEW_ARR_A(ir_node *, in, n);
3356 /* fix the keep alive */
3357 for (i = j = 0; i < n; i++) {
3358 ir_node *ka = get_End_keepalive(end, i);
3359 node_t *node = get_irn_node(ka);
3362 node = get_irn_node(get_nodes_block(ka));
3364 if (node->type.tv != tarval_unreachable && !is_Bad(ka))
3368 set_End_keepalives(end, j, in);
3373 #define SET(code) op_##code->ops.generic = (op_func)compute_##code
3376 * sets the generic functions to compute.
3378 static void set_compute_functions(void) {
3381 /* set the default compute function */
3382 for (i = get_irp_n_opcodes() - 1; i >= 0; --i) {
3383 ir_op *op = get_irp_opcode(i);
3384 op->ops.generic = (op_func)default_compute;
3387 /* set specific functions */
3403 } /* set_compute_functions */
3408 static void add_memory_keeps(ir_node **kept_memory, int len) {
3409 ir_node *end = get_irg_end(current_ir_graph);
3413 ir_nodeset_init(&set);
3415 /* check, if those nodes are already kept */
3416 for (i = get_End_n_keepalives(end) - 1; i >= 0; --i)
3417 ir_nodeset_insert(&set, get_End_keepalive(end, i));
3419 for (i = len - 1; i >= 0; --i) {
3420 ir_node *ka = kept_memory[i];
3422 if (! ir_nodeset_contains(&set, ka)) {
3423 add_End_keepalive(end, ka);
3426 ir_nodeset_destroy(&set);
3427 } /* add_memory_keeps */
3429 void combo(ir_graph *irg) {
3431 ir_node *initial_bl;
3433 ir_graph *rem = current_ir_graph;
3436 current_ir_graph = irg;
3438 /* register a debug mask */
3439 FIRM_DBG_REGISTER(dbg, "firm.opt.combo");
3441 DB((dbg, LEVEL_1, "Doing COMBO for %+F\n", irg));
3443 obstack_init(&env.obst);
3444 env.worklist = NULL;
3448 #ifdef DEBUG_libfirm
3449 env.dbg_list = NULL;
3451 env.opcode2id_map = new_set(cmp_opcode, iro_Last * 4);
3452 env.type2id_map = pmap_create();
3453 env.kept_memory = NEW_ARR_F(ir_node *, 0);
3454 env.end_idx = get_opt_global_cse() ? 0 : -1;
3455 env.lambda_input = 0;
3458 /* options driving the optimization */
3459 env.commutative = 1;
3460 env.opt_unknown = 1;
3462 assure_irg_outs(irg);
3463 assure_cf_loop(irg);
3465 /* we have our own value_of function */
3466 set_value_of_func(get_node_tarval);
3468 set_compute_functions();
3469 DEBUG_ONLY(part_nr = 0);
3471 ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK | IR_RESOURCE_PHI_LIST);
3473 if (env.opt_unknown)
3474 tarval_UNKNOWN = tarval_top;
3476 tarval_UNKNOWN = tarval_bad;
3478 /* create the initial partition and place it on the work list */
3479 env.initial = new_partition(&env);
3480 add_to_worklist(env.initial, &env);
3481 irg_walk_graph(irg, create_initial_partitions, init_block_phis, &env);
3483 /* set the hook: from now, every node has a partition and a type */
3484 DEBUG_ONLY(set_dump_node_vcgattr_hook(dump_partition_hook));
3486 /* all nodes on the initial partition have type Top */
3487 env.initial->type_is_T_or_C = 1;
3489 /* Place the START Node's partition on cprop.
3490 Place the START Node on its local worklist. */
3491 initial_bl = get_irg_start_block(irg);
3492 start = get_irn_node(initial_bl);
3493 add_to_cprop(start, &env);
3497 if (env.worklist != NULL)
3499 } while (env.cprop != NULL || env.worklist != NULL);
3501 dump_all_partitions(&env);
3502 check_all_partitions(&env);
3505 dump_ir_block_graph(irg, "-partition");
3508 /* apply the result */
3510 /* check, which nodes must be kept */
3511 irg_walk_graph(irg, NULL, find_kept_memory, &env);
3513 /* kill unreachable control flow */
3514 irg_block_walk_graph(irg, NULL, apply_cf, &env);
3515 /* Kill keep-alives of dead blocks: this speeds up apply_result()
3516 * and fixes assertion because dead cf to dead blocks is NOT removed by
3518 apply_end(get_irg_end(irg), &env);
3519 irg_walk_graph(irg, NULL, apply_result, &env);
3521 len = ARR_LEN(env.kept_memory);
3523 add_memory_keeps(env.kept_memory, len);
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 | IR_RESOURCE_PHI_LIST);
3539 /* remove the partition hook */
3540 DEBUG_ONLY(set_dump_node_vcgattr_hook(NULL));
3542 DEL_ARR_F(env.kept_memory);
3543 pmap_destroy(env.type2id_map);
3544 del_set(env.opcode2id_map);
3545 obstack_free(&env.obst, NULL);
3547 /* restore value_of() default behavior */
3548 set_value_of_func(NULL);
3549 current_ir_graph = rem;