2 * Copyright (C) 1995-2008 University of Karlsruhe. All right reserved.
4 * This file is part of libFirm.
6 * This file may be distributed and/or modified under the terms of the
7 * GNU General Public License version 2 as published by the Free Software
8 * Foundation and appearing in the file LICENSE.GPL included in the
9 * packaging of this file.
11 * Licensees holding valid libFirm Professional Edition licenses may use
12 * this file in accordance with the libFirm Commercial License.
13 * Agreement provided with the Software.
15 * This file is provided AS IS with NO WARRANTY OF ANY KIND, INCLUDING THE
16 * WARRANTY OF DESIGN, MERCHANTABILITY AND FITNESS FOR A PARTICULAR
22 * @brief Cliff Click's Combined Analysis/Optimization
23 * @author Michael Beck
26 * This is a slightly enhanced version of Cliff Clicks combo algorithm
27 * - support for commutative nodes is added, Add(a,b) and Add(b,a) ARE congruent
28 * - supports all Firm direct (by a data edge) identities except Mux
29 * (Mux can be a 2-input or 1-input identity, only 2-input is implemented yet)
30 * - supports Confirm nodes (handle them like Copies but do NOT remove them)
31 * - let Cmp nodes calculate Top like all othe data nodes: this would let
32 * Mux nodes to calculate Unknown instead of taking the true result
33 * - let Cond(Top) always select FALSE/default: This is tricky. Nodes are only reavaluated
34 * IFF the predecessor changed its type. Because nodes are initialized with Top
35 * this never happens, let all Proj(Cond) be unreachable.
36 * We avoid this condition by the same way we work around Phi: whenever a Block
37 * node is placed on the list, place its Cond nodes (and because they are Tuple
38 * all its Proj-nodes either on the cprop list)
39 * Especially, this changes the meaning of Click's example:
54 * using Click's version while is silent with our.
55 * - support for global congruences is implemented but not tested yet
57 * Note further that we use the terminology from Click's work here, which is different
58 * in some cases from Firm terminology. Especially, Click's type is a
59 * Firm tarval/entity, nevertheless we call it type here for "maximum compatibility".
65 #include "iroptimize.h"
73 #include "irgraph_t.h"
80 #include "iropt_dbg.h"
90 /* define this to check that all type translations are monotone */
91 #define VERIFY_MONOTONE
93 /* define this to check the consistency of partitions */
94 #define CHECK_PARTITIONS
97 /* allow optimization of non-strict programs */
100 typedef struct node_t node_t;
101 typedef struct partition_t partition_t;
102 typedef struct opcode_key_t opcode_key_t;
103 typedef struct listmap_entry_t listmap_entry_t;
105 /** The type of the compute function. */
106 typedef void (*compute_func)(node_t *node);
111 struct opcode_key_t {
112 ir_opcode code; /**< The Firm opcode. */
113 ir_mode *mode; /**< The mode of all nodes in the partition. */
114 int arity; /**< The arity of this opcode (needed for Phi etc. */
116 long proj; /**< For Proj nodes, its proj number */
117 ir_entity *ent; /**< For Sel Nodes, its entity */
122 * An entry in the list_map.
124 struct listmap_entry_t {
125 void *id; /**< The id. */
126 node_t *list; /**< The associated list for this id. */
127 listmap_entry_t *next; /**< Link to the next entry in the map. */
130 /** We must map id's to lists. */
131 typedef struct listmap_t {
132 set *map; /**< Map id's to listmap_entry_t's */
133 listmap_entry_t *values; /**< List of all values in the map. */
137 * A lattice element. Because we handle constants and symbolic constants different, we
138 * have to use this union.
149 ir_node *node; /**< The IR-node itself. */
150 list_head node_list; /**< Double-linked list of leader/follower entries. */
151 list_head cprop_list; /**< Double-linked partition.cprop list. */
152 partition_t *part; /**< points to the partition this node belongs to */
153 node_t *next; /**< Next node on local list (partition.touched, fallen). */
154 node_t *race_next; /**< Next node on race list. */
155 lattice_elem_t type; /**< The associated lattice element "type". */
156 int max_user_input; /**< Maximum input number of Def-Use edges. */
157 int next_edge; /**< Index of the next Def-Use edge to use. */
158 int n_followers; /**< Number of Follower in the outs set. */
159 unsigned on_touched:1; /**< Set, if this node is on the partition.touched set. */
160 unsigned on_cprop:1; /**< Set, if this node is on the partition.cprop list. */
161 unsigned on_fallen:1; /**< Set, if this node is on the fallen list. */
162 unsigned is_follower:1; /**< Set, if this node is a follower. */
163 unsigned by_all_const:1; /**< Set, if this node was once evaluated by all constants. */
164 unsigned flagged:2; /**< 2 Bits, set if this node was visited by race 1 or 2. */
168 * A partition containing congruent nodes.
171 list_head Leader; /**< The head of partition Leader node list. */
172 list_head Follower; /**< The head of partition Follower node list. */
173 list_head cprop; /**< The head of partition.cprop list. */
174 list_head cprop_X; /**< The head of partition.cprop (Cond nodes and its Projs) list. */
175 partition_t *wl_next; /**< Next entry in the work list if any. */
176 partition_t *touched_next; /**< Points to the next partition in the touched set. */
177 partition_t *cprop_next; /**< Points to the next partition in the cprop list. */
178 partition_t *split_next; /**< Points to the next partition in the list that must be split by split_by(). */
179 node_t *touched; /**< The partition.touched set of this partition. */
180 unsigned n_leader; /**< Number of entries in this partition.Leader. */
181 unsigned n_touched; /**< Number of entries in the partition.touched. */
182 int max_user_inputs; /**< Maximum number of user inputs of all entries. */
183 unsigned on_worklist:1; /**< Set, if this partition is in the work list. */
184 unsigned on_touched:1; /**< Set, if this partition is on the touched set. */
185 unsigned on_cprop:1; /**< Set, if this partition is on the cprop list. */
186 unsigned type_is_T_or_C:1;/**< Set, if all nodes in this partition have type Top or Constant. */
188 partition_t *dbg_next; /**< Link all partitions for debugging */
189 unsigned nr; /**< A unique number for (what-)mapping, >0. */
193 typedef struct environment_t {
194 struct obstack obst; /**< obstack to allocate data structures. */
195 partition_t *worklist; /**< The work list. */
196 partition_t *cprop; /**< The constant propagation list. */
197 partition_t *touched; /**< the touched set. */
198 partition_t *initial; /**< The initial partition. */
199 set *opcode2id_map; /**< The opcodeMode->id map. */
200 pmap *type2id_map; /**< The type->id map. */
201 int end_idx; /**< -1 for local and 0 for global congruences. */
202 int lambda_input; /**< Captured argument for lambda_partition(). */
203 char modified; /**< Set, if the graph was modified. */
204 char commutative; /**< Set, if commutation nodes should be handled specially. */
206 partition_t *dbg_list; /**< List of all partitions. */
210 /** Type of the what function. */
211 typedef void *(*what_func)(const node_t *node, environment_t *env);
213 #define get_irn_node(follower) ((node_t *)get_irn_link(follower))
214 #define set_irn_node(follower, node) set_irn_link(follower, node)
216 /* we do NOT use tarval_unreachable here, instead we use Top for this purpose */
217 #undef tarval_unreachable
218 #define tarval_unreachable tarval_top
221 /** The debug module handle. */
222 DEBUG_ONLY(static firm_dbg_module_t *dbg;)
224 /** The what reason. */
225 DEBUG_ONLY(static const char *what_reason;)
227 /** Next partition number. */
228 DEBUG_ONLY(static unsigned part_nr = 0);
230 /** The tarval returned by Unknown nodes. */
232 #define tarval_UNKNOWN tarval_top
234 #define tarval_UNKNOWN tarval_bad
238 static node_t *identity(node_t *node);
240 #ifdef CHECK_PARTITIONS
244 static void check_partition(const partition_t *T) {
248 list_for_each_entry(node_t, node, &T->Leader, node_list) {
249 assert(node->is_follower == 0);
250 assert(node->flagged == 0);
251 assert(node->part == T);
254 assert(n == T->n_leader);
256 list_for_each_entry(node_t, node, &T->Follower, node_list) {
257 assert(node->is_follower == 1);
258 assert(node->flagged == 0);
259 assert(node->part == T);
261 } /* check_partition */
264 * check that all leader nodes in the partition have the same opcode.
266 static void check_opcode(const partition_t *Z) {
271 list_for_each_entry(node_t, node, &Z->Leader, node_list) {
272 ir_node *irn = node->node;
275 key.code = get_irn_opcode(irn);
276 key.mode = get_irn_mode(irn);
277 key.arity = get_irn_arity(irn);
281 switch (get_irn_opcode(irn)) {
283 key.u.proj = get_Proj_proj(irn);
286 key.u.ent = get_Sel_entity(irn);
293 assert(key.code == get_irn_opcode(irn));
294 assert(key.mode == get_irn_mode(irn));
295 assert(key.arity == get_irn_arity(irn));
297 switch (get_irn_opcode(irn)) {
299 assert(key.u.proj == get_Proj_proj(irn));
302 assert(key.u.ent == get_Sel_entity(irn));
311 static void check_all_partitions(environment_t *env) {
316 for (P = env->dbg_list; P != NULL; P = P->dbg_next) {
318 if (! P->type_is_T_or_C)
320 list_for_each_entry(node_t, node, &P->Follower, node_list) {
321 node_t *leader = identity(node);
323 assert(leader != node && leader->part == node->part);
332 static void do_check_list(const node_t *list, int ofs, const partition_t *Z) {
335 #define NEXT(e) *((const node_t **)((char *)(e) + (ofs)))
336 for (e = list; e != NULL; e = NEXT(e)) {
337 assert(e->part == Z);
340 } /* ido_check_list */
343 * Check a local list.
345 static void check_list(const node_t *list, const partition_t *Z) {
346 do_check_list(list, offsetof(node_t, next), Z);
350 #define check_partition(T)
351 #define check_list(list, Z)
352 #define check_all_partitions(env)
353 #endif /* CHECK_PARTITIONS */
356 static inline lattice_elem_t get_partition_type(const partition_t *X);
359 * Dump partition to output.
361 static void dump_partition(const char *msg, const partition_t *part) {
364 lattice_elem_t type = get_partition_type(part);
366 DB((dbg, LEVEL_2, "%s part%u%s (%u, %+F) {\n ",
367 msg, part->nr, part->type_is_T_or_C ? "*" : "",
368 part->n_leader, type));
369 list_for_each_entry(node_t, node, &part->Leader, node_list) {
370 DB((dbg, LEVEL_2, "%s%+F", first ? "" : ", ", node->node));
373 if (! list_empty(&part->Follower)) {
374 DB((dbg, LEVEL_2, "\n---\n "));
376 list_for_each_entry(node_t, node, &part->Follower, node_list) {
377 DB((dbg, LEVEL_2, "%s%+F", first ? "" : ", ", node->node));
381 DB((dbg, LEVEL_2, "\n}\n"));
382 } /* dump_partition */
387 static void do_dump_list(const char *msg, const node_t *node, int ofs) {
391 #define GET_LINK(p, ofs) *((const node_t **)((char *)(p) + (ofs)))
393 DB((dbg, LEVEL_3, "%s = {\n ", msg));
394 for (p = node; p != NULL; p = GET_LINK(p, ofs)) {
395 DB((dbg, LEVEL_3, "%s%+F", first ? "" : ", ", p->node));
398 DB((dbg, LEVEL_3, "\n}\n"));
406 static void dump_race_list(const char *msg, const node_t *list) {
407 do_dump_list(msg, list, offsetof(node_t, race_next));
408 } /* dump_race_list */
411 * Dumps a local list.
413 static void dump_list(const char *msg, const node_t *list) {
414 do_dump_list(msg, list, offsetof(node_t, next));
418 * Dump all partitions.
420 static void dump_all_partitions(const environment_t *env) {
421 const partition_t *P;
423 DB((dbg, LEVEL_2, "All partitions\n===============\n"));
424 for (P = env->dbg_list; P != NULL; P = P->dbg_next)
425 dump_partition("", P);
426 } /* dump_all_partitions */
431 static void dump_split_list(const partition_t *list) {
432 const partition_t *p;
434 DB((dbg, LEVEL_2, "Split by %s produced = {\n", what_reason));
435 for (p = list; p != NULL; p = p->split_next)
436 DB((dbg, LEVEL_2, "part%u, ", p->nr));
437 DB((dbg, LEVEL_2, "\n}\n"));
438 } /* dump_split_list */
441 * Dump partition and type for a node.
443 static int dump_partition_hook(FILE *F, ir_node *n, ir_node *local) {
444 ir_node *irn = local != NULL ? local : n;
445 node_t *node = get_irn_node(irn);
447 ir_fprintf(F, "info2 : \"partition %u type %+F\"\n", node->part->nr, node->type);
449 } /* dump_partition_hook */
452 #define dump_partition(msg, part)
453 #define dump_race_list(msg, list)
454 #define dump_list(msg, list)
455 #define dump_all_partitions(env)
456 #define dump_split_list(list)
459 #if defined(VERIFY_MONOTONE) && defined (DEBUG_libfirm)
461 * Verify that a type transition is monotone
463 static void do_verify_type(const lattice_elem_t old_type, node_t *node) {
464 if (old_type.tv == node->type.tv) {
468 if (old_type.tv == tarval_top) {
469 /* from Top down-to is always allowed */
472 if (node->type.tv == tarval_bottom || node->type.tv == tarval_reachable) {
476 panic("combo: wrong translation from %+F to %+F on node %+F", old_type, node->type, node->node);
479 #define decl_verify(node) lattice_elem_t old_type = (node)->type;
480 #define verify_type(node) do_verify_type(old_type, node)
482 #define decl_verify(node) (void)0;
483 #define verify_type(node)
487 * Compare two pointer values of a listmap.
489 static int listmap_cmp_ptr(const void *elt, const void *key, size_t size) {
490 const listmap_entry_t *e1 = elt;
491 const listmap_entry_t *e2 = key;
494 return e1->id != e2->id;
495 } /* listmap_cmp_ptr */
498 * Initializes a listmap.
500 * @param map the listmap
502 static void listmap_init(listmap_t *map) {
503 map->map = new_set(listmap_cmp_ptr, 16);
508 * Terminates a listmap.
510 * @param map the listmap
512 static void listmap_term(listmap_t *map) {
517 * Return the associated listmap entry for a given id.
519 * @param map the listmap
520 * @param id the id to search for
522 * @return the associated listmap entry for the given id
524 static listmap_entry_t *listmap_find(listmap_t *map, void *id) {
525 listmap_entry_t key, *entry;
530 entry = set_insert(map->map, &key, sizeof(key), HASH_PTR(id));
532 if (entry->list == NULL) {
533 /* a new entry, put into the list */
534 entry->next = map->values;
541 * Calculate the hash value for an opcode map entry.
543 * @param entry an opcode map entry
545 * @return a hash value for the given opcode map entry
547 static unsigned opcode_hash(const opcode_key_t *entry) {
548 return (entry->mode - (ir_mode *)0) * 9 + entry->code + entry->u.proj * 3 + HASH_PTR(entry->u.ent) + entry->arity;
552 * Compare two entries in the opcode map.
554 static int cmp_opcode(const void *elt, const void *key, size_t size) {
555 const opcode_key_t *o1 = elt;
556 const opcode_key_t *o2 = key;
559 return o1->code != o2->code || o1->mode != o2->mode ||
560 o1->arity != o2->arity ||
561 o1->u.proj != o2->u.proj || o1->u.ent != o2->u.ent;
565 * Compare two Def-Use edges for input position.
567 static int cmp_def_use_edge(const void *a, const void *b) {
568 const ir_def_use_edge *ea = a;
569 const ir_def_use_edge *eb = b;
571 /* no overrun, because range is [-1, MAXINT] */
572 return ea->pos - eb->pos;
573 } /* cmp_def_use_edge */
576 * We need the Def-Use edges sorted.
578 static void sort_irn_outs(node_t *node) {
579 ir_node *irn = node->node;
580 int n_outs = get_irn_n_outs(irn);
583 qsort(&irn->out[1], n_outs, sizeof(irn->out[0]), cmp_def_use_edge);
585 node->max_user_input = irn->out[n_outs].pos;
586 } /* sort_irn_outs */
589 * Return the type of a node.
591 * @param irn an IR-node
593 * @return the associated type of this node
595 static inline lattice_elem_t get_node_type(const ir_node *irn) {
596 return get_irn_node(irn)->type;
597 } /* get_node_type */
600 * Return the tarval of a node.
602 * @param irn an IR-node
604 * @return the associated type of this node
606 static inline tarval *get_node_tarval(const ir_node *irn) {
607 lattice_elem_t type = get_node_type(irn);
609 if (is_tarval(type.tv))
611 return tarval_bottom;
612 } /* get_node_type */
615 * Add a partition to the worklist.
617 static inline void add_to_worklist(partition_t *X, environment_t *env) {
618 assert(X->on_worklist == 0);
619 DB((dbg, LEVEL_2, "Adding part%d to worklist\n", X->nr));
620 X->wl_next = env->worklist;
623 } /* add_to_worklist */
626 * Create a new empty partition.
628 * @param env the environment
630 * @return a newly allocated partition
632 static inline partition_t *new_partition(environment_t *env) {
633 partition_t *part = obstack_alloc(&env->obst, sizeof(*part));
635 INIT_LIST_HEAD(&part->Leader);
636 INIT_LIST_HEAD(&part->Follower);
637 INIT_LIST_HEAD(&part->cprop);
638 INIT_LIST_HEAD(&part->cprop_X);
639 part->wl_next = NULL;
640 part->touched_next = NULL;
641 part->cprop_next = NULL;
642 part->split_next = NULL;
643 part->touched = NULL;
646 part->max_user_inputs = 0;
647 part->on_worklist = 0;
648 part->on_touched = 0;
650 part->type_is_T_or_C = 0;
652 part->dbg_next = env->dbg_list;
653 env->dbg_list = part;
654 part->nr = part_nr++;
658 } /* new_partition */
661 * Get the first node from a partition.
663 static inline node_t *get_first_node(const partition_t *X) {
664 return list_entry(X->Leader.next, node_t, node_list);
665 } /* get_first_node */
668 * Return the type of a partition (assuming partition is non-empty and
669 * all elements have the same type).
671 * @param X a partition
673 * @return the type of the first element of the partition
675 static inline lattice_elem_t get_partition_type(const partition_t *X) {
676 const node_t *first = get_first_node(X);
678 } /* get_partition_type */
681 * Creates a partition node for the given IR-node and place it
682 * into the given partition.
684 * @param irn an IR-node
685 * @param part a partition to place the node in
686 * @param env the environment
688 * @return the created node
690 static node_t *create_partition_node(ir_node *irn, partition_t *part, environment_t *env) {
691 /* create a partition node and place it in the partition */
692 node_t *node = obstack_alloc(&env->obst, sizeof(*node));
694 INIT_LIST_HEAD(&node->node_list);
695 INIT_LIST_HEAD(&node->cprop_list);
699 node->race_next = NULL;
700 node->type.tv = tarval_top;
701 node->max_user_input = 0;
703 node->n_followers = 0;
704 node->on_touched = 0;
707 node->is_follower = 0;
708 node->by_all_const = 0;
710 set_irn_node(irn, node);
712 list_add_tail(&node->node_list, &part->Leader);
716 } /* create_partition_node */
719 * Pre-Walker, initialize all Nodes' type to U or top and place
720 * all nodes into the TOP partition.
722 static void create_initial_partitions(ir_node *irn, void *ctx) {
723 environment_t *env = ctx;
724 partition_t *part = env->initial;
727 node = create_partition_node(irn, part, env);
729 if (node->max_user_input > part->max_user_inputs)
730 part->max_user_inputs = node->max_user_input;
733 set_Block_phis(irn, NULL);
735 } /* create_initial_partitions */
738 * Post-Walker, collect all Block-Phi lists, set Cond.
740 static void init_block_phis(ir_node *irn, void *ctx) {
744 add_Block_phi(get_nodes_block(irn), irn);
746 } /* init_block_phis */
749 * Add a node to the entry.partition.touched set and
750 * node->partition to the touched set if not already there.
753 * @param env the environment
755 static inline void add_to_touched(node_t *y, environment_t *env) {
756 if (y->on_touched == 0) {
757 partition_t *part = y->part;
759 y->next = part->touched;
764 if (part->on_touched == 0) {
765 part->touched_next = env->touched;
767 part->on_touched = 1;
770 check_list(part->touched, part);
772 } /* add_to_touched */
775 * Place a node on the cprop list.
778 * @param env the environment
780 static void add_to_cprop(node_t *y, environment_t *env) {
783 /* Add y to y.partition.cprop. */
784 if (y->on_cprop == 0) {
785 partition_t *Y = y->part;
787 /* place Conds and its Proj nodes on the cprop_X list */
788 if (is_Cond(skip_Proj(y->node)))
789 list_add_tail(&y->cprop_list, &Y->cprop_X);
791 list_add_tail(&y->cprop_list, &Y->cprop);
794 DB((dbg, LEVEL_3, "Add %+F to part%u.cprop\n", y->node, Y->nr));
796 /* place its partition on the cprop list */
797 if (Y->on_cprop == 0) {
798 Y->cprop_next = env->cprop;
804 if (get_irn_mode(irn) == mode_T) {
805 /* mode_T nodes always produce tarval_bottom, so we must explicitly
806 add it's Proj's to get constant evaluation to work */
809 for (i = get_irn_n_outs(irn) - 1; i >= 0; --i) {
810 node_t *proj = get_irn_node(get_irn_out(irn, i));
812 add_to_cprop(proj, env);
814 } else if (is_Block(irn)) {
815 /* Due to the way we handle Phi's, we must place all Phis of a block on the list
816 * if someone placed the block. The Block is only placed if the reachability
817 * changes, and this must be re-evaluated in compute_Phi(). */
819 for (phi = get_Block_phis(irn); phi != NULL; phi = get_Phi_next(phi)) {
820 node_t *p = get_irn_node(phi);
821 add_to_cprop(p, env);
827 * Update the worklist: If Z is on worklist then add Z' to worklist.
828 * Else add the smaller of Z and Z' to worklist.
830 * @param Z the Z partition
831 * @param Z_prime the Z' partition, a previous part of Z
832 * @param env the environment
834 static void update_worklist(partition_t *Z, partition_t *Z_prime, environment_t *env) {
835 if (Z->on_worklist || Z_prime->n_leader < Z->n_leader) {
836 add_to_worklist(Z_prime, env);
838 add_to_worklist(Z, env);
840 } /* update_worklist */
843 * Make all inputs to x no longer be F.def_use edges.
847 static void move_edges_to_leader(node_t *x) {
848 ir_node *irn = x->node;
851 for (i = get_irn_arity(irn) - 1; i >= 0; --i) {
852 node_t *pred = get_irn_node(get_irn_n(irn, i));
857 n = get_irn_n_outs(p);
858 for (j = 1; j <= pred->n_followers; ++j) {
859 if (p->out[j].pos == i && p->out[j].use == irn) {
860 /* found a follower edge to x, move it to the Leader */
861 ir_def_use_edge edge = p->out[j];
863 /* remove this edge from the Follower set */
864 p->out[j] = p->out[pred->n_followers];
867 /* sort it into the leader set */
868 for (k = pred->n_followers + 2; k <= n; ++k) {
869 if (p->out[k].pos >= edge.pos)
871 p->out[k - 1] = p->out[k];
873 /* place the new edge here */
874 p->out[k - 1] = edge;
876 /* edge found and moved */
881 } /* move_edges_to_leader */
884 * Split a partition that has NO followers by a local list.
886 * @param Z partition to split
887 * @param g a (non-empty) node list
888 * @param env the environment
890 * @return a new partition containing the nodes of g
892 static partition_t *split_no_followers(partition_t *Z, node_t *g, environment_t *env) {
893 partition_t *Z_prime;
898 dump_partition("Splitting ", Z);
899 dump_list("by list ", g);
903 /* Remove g from Z. */
904 for (node = g; node != NULL; node = node->next) {
905 assert(node->part == Z);
906 list_del(&node->node_list);
909 assert(n < Z->n_leader);
912 /* Move g to a new partition, Z'. */
913 Z_prime = new_partition(env);
915 for (node = g; node != NULL; node = node->next) {
916 list_add_tail(&node->node_list, &Z_prime->Leader);
917 node->part = Z_prime;
918 if (node->max_user_input > max_input)
919 max_input = node->max_user_input;
921 Z_prime->max_user_inputs = max_input;
922 Z_prime->n_leader = n;
925 check_partition(Z_prime);
927 /* for now, copy the type info tag, it will be adjusted in split_by(). */
928 Z_prime->type_is_T_or_C = Z->type_is_T_or_C;
930 update_worklist(Z, Z_prime, env);
932 dump_partition("Now ", Z);
933 dump_partition("Created new ", Z_prime);
935 } /* split_no_followers */
938 * Make the Follower -> Leader transition for a node.
942 static void follower_to_leader(node_t *n) {
943 assert(n->is_follower == 1);
945 DB((dbg, LEVEL_2, "%+F make the follower -> leader transition\n", n->node));
947 move_edges_to_leader(n);
948 list_del(&n->node_list);
949 list_add_tail(&n->node_list, &n->part->Leader);
951 } /* follower_to_leader */
954 * The environment for one race step.
956 typedef struct step_env {
957 node_t *initial; /**< The initial node list. */
958 node_t *unwalked; /**< The unwalked node list. */
959 node_t *walked; /**< The walked node list. */
960 int index; /**< Next index of Follower use_def edge. */
961 unsigned side; /**< side number. */
965 * Return non-zero, if a input is a real follower
967 * @param irn the node to check
968 * @param input number of the input
970 static int is_real_follower(const ir_node *irn, int input) {
973 switch (get_irn_opcode(irn)) {
976 /* ignore the Confirm bound input */
982 /* ignore the Mux sel input */
987 /* dead inputs are not follower edges */
988 ir_node *block = get_nodes_block(irn);
989 node_t *pred = get_irn_node(get_Block_cfgpred(block, input));
991 if (pred->type.tv == tarval_unreachable)
1001 /* only a Sub x,0 / Shift x,0 might be a follower */
1008 pred = get_irn_node(get_irn_n(irn, input));
1009 if (is_tarval(pred->type.tv) && tarval_is_null(pred->type.tv))
1013 pred = get_irn_node(get_irn_n(irn, input));
1014 if (is_tarval(pred->type.tv) && tarval_is_one(pred->type.tv))
1018 pred = get_irn_node(get_irn_n(irn, input));
1019 if (is_tarval(pred->type.tv) && tarval_is_all_one(pred->type.tv))
1024 /* all inputs are followers */
1027 assert(!"opcode not implemented yet");
1031 } /* is_real_follower */
1034 * Do one step in the race.
1036 static int step(step_env *env) {
1039 if (env->initial != NULL) {
1040 /* Move node from initial to unwalked */
1042 env->initial = n->race_next;
1044 n->race_next = env->unwalked;
1050 while (env->unwalked != NULL) {
1051 /* let n be the first node in unwalked */
1053 while (env->index < n->n_followers) {
1054 const ir_def_use_edge *edge = &n->node->out[1 + env->index];
1056 /* let m be n.F.def_use[index] */
1057 node_t *m = get_irn_node(edge->use);
1059 assert(m->is_follower);
1061 * Some inputs, like the get_Confirm_bound are NOT
1062 * real followers, sort them out.
1064 if (! is_real_follower(m->node, edge->pos)) {
1070 /* only followers from our partition */
1071 if (m->part != n->part)
1074 if ((m->flagged & env->side) == 0) {
1075 m->flagged |= env->side;
1077 if (m->flagged != 3) {
1078 /* visited the first time */
1079 /* add m to unwalked not as first node (we might still need to
1080 check for more follower node */
1081 m->race_next = n->race_next;
1085 /* else already visited by the other side and on the other list */
1088 /* move n to walked */
1089 env->unwalked = n->race_next;
1090 n->race_next = env->walked;
1098 * Clear the flags from a list and check for
1099 * nodes that where touched from both sides.
1101 * @param list the list
1103 static int clear_flags(node_t *list) {
1107 for (n = list; n != NULL; n = n->race_next) {
1108 if (n->flagged == 3) {
1109 /* we reach a follower from both sides, this will split congruent
1110 * inputs and make it a leader. */
1111 follower_to_leader(n);
1120 * Split a partition by a local list using the race.
1122 * @param pX pointer to the partition to split, might be changed!
1123 * @param gg a (non-empty) node list
1124 * @param env the environment
1126 * @return a new partition containing the nodes of gg
1128 static partition_t *split(partition_t **pX, node_t *gg, environment_t *env) {
1129 partition_t *X = *pX;
1130 partition_t *X_prime;
1133 node_t *g, *h, *node, *t;
1134 int max_input, transitions, winner, shf;
1136 DEBUG_ONLY(static int run = 0;)
1138 DB((dbg, LEVEL_2, "Run %d ", run++));
1139 if (list_empty(&X->Follower)) {
1140 /* if the partition has NO follower, we can use the fast
1141 splitting algorithm. */
1142 return split_no_followers(X, gg, env);
1144 /* else do the race */
1146 dump_partition("Splitting ", X);
1147 dump_list("by list ", gg);
1149 INIT_LIST_HEAD(&tmp);
1151 /* Remove gg from X.Leader and put into g */
1153 for (node = gg; node != NULL; node = node->next) {
1154 assert(node->part == X);
1155 assert(node->is_follower == 0);
1157 list_del(&node->node_list);
1158 list_add_tail(&node->node_list, &tmp);
1159 node->race_next = g;
1164 list_for_each_entry(node_t, node, &X->Leader, node_list) {
1165 node->race_next = h;
1168 /* restore X.Leader */
1169 list_splice(&tmp, &X->Leader);
1171 senv[0].initial = g;
1172 senv[0].unwalked = NULL;
1173 senv[0].walked = NULL;
1177 senv[1].initial = h;
1178 senv[1].unwalked = NULL;
1179 senv[1].walked = NULL;
1184 * Some informations on the race that are not stated clearly in Click's
1186 * 1) A follower stays on the side that reach him first.
1187 * 2) If the other side reches a follower, if will be converted to
1188 * a leader. /This must be done after the race is over, else the
1189 * edges we are iterating on are renumbered./
1190 * 3) /New leader might end up on both sides./
1191 * 4) /If one side ends up with new Leaders, we must ensure that
1192 * they can split out by opcode, hence we have to put _every_
1193 * partition with new Leader nodes on the cprop list, as
1194 * opcode splitting is done by split_by() at the end of
1195 * constant propagation./
1198 if (step(&senv[0])) {
1202 if (step(&senv[1])) {
1207 assert(senv[winner].initial == NULL);
1208 assert(senv[winner].unwalked == NULL);
1210 /* clear flags from walked/unwalked */
1212 transitions = clear_flags(senv[0].unwalked) << shf;
1213 transitions |= clear_flags(senv[0].walked) << shf;
1215 transitions |= clear_flags(senv[1].unwalked) << shf;
1216 transitions |= clear_flags(senv[1].walked) << shf;
1218 dump_race_list("winner ", senv[winner].walked);
1220 /* Move walked_{winner} to a new partition, X'. */
1221 X_prime = new_partition(env);
1224 for (node = senv[winner].walked; node != NULL; node = node->race_next) {
1225 list_del(&node->node_list);
1226 node->part = X_prime;
1227 if (node->is_follower) {
1228 list_add_tail(&node->node_list, &X_prime->Follower);
1230 list_add_tail(&node->node_list, &X_prime->Leader);
1233 if (node->max_user_input > max_input)
1234 max_input = node->max_user_input;
1236 X_prime->n_leader = n;
1237 X_prime->max_user_inputs = max_input;
1238 X->n_leader -= X_prime->n_leader;
1240 /* for now, copy the type info tag, it will be adjusted in split_by(). */
1241 X_prime->type_is_T_or_C = X->type_is_T_or_C;
1244 * Even if a follower was not checked by both sides, it might have
1245 * loose its congruence, so we need to check this case for all follower.
1247 list_for_each_entry_safe(node_t, node, t, &X_prime->Follower, node_list) {
1248 if (identity(node) == node) {
1249 follower_to_leader(node);
1255 check_partition(X_prime);
1257 /* X' is the smaller part */
1258 add_to_worklist(X_prime, env);
1261 * If there where follower to leader transitions, ensure that the nodes
1262 * can be split out if necessary.
1264 if (transitions & 1) {
1265 /* place winner partition on the cprop list */
1266 if (X_prime->on_cprop == 0) {
1267 X_prime->cprop_next = env->cprop;
1268 env->cprop = X_prime;
1269 X_prime->on_cprop = 1;
1272 if (transitions & 2) {
1273 /* place other partition on the cprop list */
1274 if (X->on_cprop == 0) {
1275 X->cprop_next = env->cprop;
1281 dump_partition("Now ", X);
1282 dump_partition("Created new ", X_prime);
1284 /* we have to ensure that the partition containing g is returned */
1294 * Returns non-zero if the i'th input of a Phi node is live.
1296 * @param phi a Phi-node
1297 * @param i an input number
1299 * @return non-zero if the i'th input of the given Phi node is live
1301 static int is_live_input(ir_node *phi, int i) {
1303 ir_node *block = get_nodes_block(phi);
1304 ir_node *pred = get_Block_cfgpred(block, i);
1305 lattice_elem_t type = get_node_type(pred);
1307 return type.tv != tarval_unreachable;
1309 /* else it's the control input, always live */
1311 } /* is_live_input */
1314 * Return non-zero if a type is a constant.
1316 static int is_constant_type(lattice_elem_t type) {
1317 if (type.tv != tarval_bottom && type.tv != tarval_top)
1320 } /* is_constant_type */
1323 * Check whether a type is neither Top or a constant.
1324 * Note: U is handled like Top here, R is a constant.
1326 * @param type the type to check
1328 static int type_is_neither_top_nor_const(const lattice_elem_t type) {
1329 if (is_tarval(type.tv)) {
1330 if (type.tv == tarval_top)
1332 if (tarval_is_constant(type.tv))
1339 } /* type_is_neither_top_nor_const */
1342 * Collect nodes to the touched list.
1344 * @param list the list which contains the nodes that must be evaluated
1345 * @param idx the index of the def_use edge to evaluate
1346 * @param env the environment
1348 static void collect_touched(list_head *list, int idx, environment_t *env) {
1350 int end_idx = env->end_idx;
1352 list_for_each_entry(node_t, x, list, node_list) {
1356 /* leader edges start AFTER follower edges */
1357 x->next_edge = x->n_followers + 1;
1359 num_edges = get_irn_n_outs(x->node);
1361 /* for all edges in x.L.def_use_{idx} */
1362 while (x->next_edge <= num_edges) {
1363 const ir_def_use_edge *edge = &x->node->out[x->next_edge];
1366 /* check if we have necessary edges */
1367 if (edge->pos > idx)
1374 /* only non-commutative nodes */
1375 if (env->commutative &&
1376 (idx == 0 || idx == 1) && is_op_commutative(get_irn_op(succ)))
1379 /* ignore the "control input" for non-pinned nodes
1380 if we are running in GCSE mode */
1381 if (idx < end_idx && get_irn_pinned(succ) != op_pin_state_pinned)
1384 y = get_irn_node(succ);
1385 assert(get_irn_n(succ, idx) == x->node);
1387 /* ignore block edges touching followers */
1388 if (idx == -1 && y->is_follower)
1391 if (is_constant_type(y->type)) {
1392 ir_opcode code = get_irn_opcode(succ);
1393 if (code == iro_Sub || code == iro_Cmp)
1394 add_to_cprop(y, env);
1397 /* Partitions of constants should not be split simply because their Nodes have unequal
1398 functions or incongruent inputs. */
1399 if (type_is_neither_top_nor_const(y->type) &&
1400 (! is_Phi(y->node) || is_live_input(y->node, idx))) {
1401 add_to_touched(y, env);
1405 } /* collect_touched */
1408 * Collect commutative nodes to the touched list.
1410 * @param list the list which contains the nodes that must be evaluated
1411 * @param env the environment
1413 static void collect_commutative_touched(list_head *list, environment_t *env) {
1416 list_for_each_entry(node_t, x, list, node_list) {
1419 num_edges = get_irn_n_outs(x->node);
1421 x->next_edge = x->n_followers + 1;
1423 /* for all edges in x.L.def_use_{idx} */
1424 while (x->next_edge <= num_edges) {
1425 const ir_def_use_edge *edge = &x->node->out[x->next_edge];
1428 /* check if we have necessary edges */
1438 /* only commutative nodes */
1439 if (!is_op_commutative(get_irn_op(succ)))
1442 y = get_irn_node(succ);
1443 if (is_constant_type(y->type)) {
1444 ir_opcode code = get_irn_opcode(succ);
1445 if (code == iro_Eor)
1446 add_to_cprop(y, env);
1449 /* Partitions of constants should not be split simply because their Nodes have unequal
1450 functions or incongruent inputs. */
1451 if (type_is_neither_top_nor_const(y->type)) {
1452 add_to_touched(y, env);
1456 } /* collect_commutative_touched */
1459 * Split the partitions if caused by the first entry on the worklist.
1461 * @param env the environment
1463 static void cause_splits(environment_t *env) {
1464 partition_t *X, *Z, *N;
1467 /* remove the first partition from the worklist */
1469 env->worklist = X->wl_next;
1472 dump_partition("Cause_split: ", X);
1474 if (env->commutative) {
1475 /* handle commutative nodes first */
1477 /* empty the touched set: already done, just clear the list */
1478 env->touched = NULL;
1480 collect_commutative_touched(&X->Leader, env);
1481 collect_commutative_touched(&X->Follower, env);
1483 for (Z = env->touched; Z != NULL; Z = N) {
1485 node_t *touched = Z->touched;
1486 unsigned n_touched = Z->n_touched;
1488 assert(Z->touched != NULL);
1490 /* beware, split might change Z */
1491 N = Z->touched_next;
1493 /* remove it from the touched set */
1496 /* Empty local Z.touched. */
1497 for (e = touched; e != NULL; e = e->next) {
1498 assert(e->is_follower == 0);
1504 if (0 < n_touched && n_touched < Z->n_leader) {
1505 DB((dbg, LEVEL_2, "Split part%d by touched\n", Z->nr));
1506 split(&Z, touched, env);
1508 assert(n_touched <= Z->n_leader);
1512 /* combine temporary leader and follower list */
1513 for (idx = -1; idx <= X->max_user_inputs; ++idx) {
1514 /* empty the touched set: already done, just clear the list */
1515 env->touched = NULL;
1517 collect_touched(&X->Leader, idx, env);
1518 collect_touched(&X->Follower, idx, env);
1520 for (Z = env->touched; Z != NULL; Z = N) {
1522 node_t *touched = Z->touched;
1523 unsigned n_touched = Z->n_touched;
1525 assert(Z->touched != NULL);
1527 /* beware, split might change Z */
1528 N = Z->touched_next;
1530 /* remove it from the touched set */
1533 /* Empty local Z.touched. */
1534 for (e = touched; e != NULL; e = e->next) {
1535 assert(e->is_follower == 0);
1541 if (0 < n_touched && n_touched < Z->n_leader) {
1542 DB((dbg, LEVEL_2, "Split part%d by touched\n", Z->nr));
1543 split(&Z, touched, env);
1545 assert(n_touched <= Z->n_leader);
1548 } /* cause_splits */
1551 * Implements split_by_what(): Split a partition by characteristics given
1552 * by the what function.
1554 * @param X the partition to split
1555 * @param What a function returning an Id for every node of the partition X
1556 * @param P a list to store the result partitions
1557 * @param env the environment
1561 static partition_t *split_by_what(partition_t *X, what_func What,
1562 partition_t **P, environment_t *env) {
1565 listmap_entry_t *iter;
1568 /* Let map be an empty mapping from the range of What to (local) list of Nodes. */
1570 list_for_each_entry(node_t, x, &X->Leader, node_list) {
1571 void *id = What(x, env);
1572 listmap_entry_t *entry;
1575 /* input not allowed, ignore */
1578 /* Add x to map[What(x)]. */
1579 entry = listmap_find(&map, id);
1580 x->next = entry->list;
1583 /* Let P be a set of Partitions. */
1585 /* for all sets S except one in the range of map do */
1586 for (iter = map.values; iter != NULL; iter = iter->next) {
1587 if (iter->next == NULL) {
1588 /* this is the last entry, ignore */
1593 /* Add SPLIT( X, S ) to P. */
1594 DB((dbg, LEVEL_2, "Split part%d by WHAT = %s\n", X->nr, what_reason));
1595 R = split(&X, S, env);
1605 } /* split_by_what */
1607 /** lambda n.(n.type) */
1608 static void *lambda_type(const node_t *node, environment_t *env) {
1610 return node->type.tv;
1613 /** lambda n.(n.opcode) */
1614 static void *lambda_opcode(const node_t *node, environment_t *env) {
1615 opcode_key_t key, *entry;
1616 ir_node *irn = node->node;
1618 key.code = get_irn_opcode(irn);
1619 key.mode = get_irn_mode(irn);
1620 key.arity = get_irn_arity(irn);
1624 switch (get_irn_opcode(irn)) {
1626 key.u.proj = get_Proj_proj(irn);
1629 key.u.ent = get_Sel_entity(irn);
1635 entry = set_insert(env->opcode2id_map, &key, sizeof(key), opcode_hash(&key));
1637 } /* lambda_opcode */
1639 /** lambda n.(n[i].partition) */
1640 static void *lambda_partition(const node_t *node, environment_t *env) {
1641 ir_node *skipped = skip_Proj(node->node);
1644 int i = env->lambda_input;
1646 if (i >= get_irn_arity(node->node)) {
1648 * We are outside the allowed range: This can happen even
1649 * if we have split by opcode first: doing so might move Followers
1650 * to Leaders and those will have a different opcode!
1651 * Note that in this case the partition is on the cprop list and will be
1657 /* ignore the "control input" for non-pinned nodes
1658 if we are running in GCSE mode */
1659 if (i < env->end_idx && get_irn_pinned(skipped) != op_pin_state_pinned)
1662 pred = i == -1 ? get_irn_n(skipped, i) : get_irn_n(node->node, i);
1663 p = get_irn_node(pred);
1666 } /* lambda_partition */
1668 /** lambda n.(n[i].partition) for commutative nodes */
1669 static void *lambda_commutative_partition(const node_t *node, environment_t *env) {
1670 ir_node *irn = node->node;
1671 ir_node *skipped = skip_Proj(irn);
1672 ir_node *pred, *left, *right;
1674 partition_t *pl, *pr;
1675 int i = env->lambda_input;
1677 if (i >= get_irn_arity(node->node)) {
1679 * We are outside the allowed range: This can happen even
1680 * if we have split by opcode first: doing so might move Followers
1681 * to Leaders and those will have a different opcode!
1682 * Note that in this case the partition is on the cprop list and will be
1688 /* ignore the "control input" for non-pinned nodes
1689 if we are running in GCSE mode */
1690 if (i < env->end_idx && get_irn_pinned(skipped) != op_pin_state_pinned)
1694 pred = get_irn_n(skipped, i);
1695 p = get_irn_node(pred);
1699 if (is_op_commutative(get_irn_op(irn))) {
1700 /* normalize partition order by returning the "smaller" on input 0,
1701 the "bigger" on input 1. */
1702 left = get_binop_left(irn);
1703 pl = get_irn_node(left)->part;
1704 right = get_binop_right(irn);
1705 pr = get_irn_node(right)->part;
1708 return pl < pr ? pl : pr;
1710 return pl > pr ? pl : pr;
1712 /* a not split out Follower */
1713 pred = get_irn_n(irn, i);
1714 p = get_irn_node(pred);
1718 } /* lambda_commutative_partition */
1721 * Returns true if a type is a constant (and NOT Top
1724 static int is_con(const lattice_elem_t type) {
1725 /* be conservative */
1726 if (is_tarval(type.tv))
1727 return tarval_is_constant(type.tv);
1728 return is_entity(type.sym.entity_p);
1732 * Implements split_by().
1734 * @param X the partition to split
1735 * @param env the environment
1737 static void split_by(partition_t *X, environment_t *env) {
1738 partition_t *I, *P = NULL;
1741 dump_partition("split_by", X);
1743 if (X->n_leader == 1) {
1744 /* we have only one leader, no need to split, just check it's type */
1745 node_t *x = get_first_node(X);
1746 X->type_is_T_or_C = x->type.tv == tarval_top || is_con(x->type);
1750 DEBUG_ONLY(what_reason = "lambda n.(n.type)";)
1751 P = split_by_what(X, lambda_type, &P, env);
1754 /* adjust the type tags, we have split partitions by type */
1755 for (I = P; I != NULL; I = I->split_next) {
1756 node_t *x = get_first_node(I);
1757 I->type_is_T_or_C = x->type.tv == tarval_top || is_con(x->type);
1764 if (Y->n_leader > 1) {
1765 /* we do not want split the TOP or constant partitions */
1766 if (! Y->type_is_T_or_C) {
1767 partition_t *Q = NULL;
1769 DEBUG_ONLY(what_reason = "lambda n.(n.opcode)";)
1770 Q = split_by_what(Y, lambda_opcode, &Q, env);
1777 if (Z->n_leader > 1) {
1778 const node_t *first = get_first_node(Z);
1779 int arity = get_irn_arity(first->node);
1781 what_func what = lambda_partition;
1782 DEBUG_ONLY(char buf[64];)
1784 if (env->commutative && is_op_commutative(get_irn_op(first->node)))
1785 what = lambda_commutative_partition;
1788 * BEWARE: during splitting by input 2 for instance we might
1789 * create new partitions which are different by input 1, so collect
1790 * them and split further.
1792 Z->split_next = NULL;
1795 for (input = arity - 1; input >= -1; --input) {
1797 partition_t *Z_prime = R;
1800 if (Z_prime->n_leader > 1) {
1801 env->lambda_input = input;
1802 DEBUG_ONLY(snprintf(buf, sizeof(buf), "lambda n.(n[%d].partition)", input);)
1803 DEBUG_ONLY(what_reason = buf;)
1804 S = split_by_what(Z_prime, what, &S, env);
1807 Z_prime->split_next = S;
1810 } while (R != NULL);
1815 } while (Q != NULL);
1818 } while (P != NULL);
1822 * (Re-)compute the type for a given node.
1824 * @param node the node
1826 static void default_compute(node_t *node) {
1828 ir_node *irn = node->node;
1829 node_t *block = get_irn_node(get_nodes_block(irn));
1831 if (block->type.tv == tarval_unreachable) {
1832 node->type.tv = tarval_top;
1836 /* if any of the data inputs have type top, the result is type top */
1837 for (i = get_irn_arity(irn) - 1; i >= 0; --i) {
1838 ir_node *pred = get_irn_n(irn, i);
1839 node_t *p = get_irn_node(pred);
1841 if (p->type.tv == tarval_top) {
1842 node->type.tv = tarval_top;
1847 if (get_irn_mode(node->node) == mode_X)
1848 node->type.tv = tarval_reachable;
1850 node->type.tv = computed_value(irn);
1851 } /* default_compute */
1854 * (Re-)compute the type for a Block node.
1856 * @param node the node
1858 static void compute_Block(node_t *node) {
1860 ir_node *block = node->node;
1862 if (block == get_irg_start_block(current_ir_graph)) {
1863 /* start block is always reachable */
1864 node->type.tv = tarval_reachable;
1868 for (i = get_Block_n_cfgpreds(block) - 1; i >= 0; --i) {
1869 node_t *pred = get_irn_node(get_Block_cfgpred(block, i));
1871 if (pred->type.tv == tarval_reachable) {
1872 /* A block is reachable, if at least of predecessor is reachable. */
1873 node->type.tv = tarval_reachable;
1877 node->type.tv = tarval_top;
1878 } /* compute_Block */
1881 * (Re-)compute the type for a Bad node.
1883 * @param node the node
1885 static void compute_Bad(node_t *node) {
1886 /* Bad nodes ALWAYS compute Top */
1887 node->type.tv = tarval_top;
1891 * (Re-)compute the type for an Unknown node.
1893 * @param node the node
1895 static void compute_Unknown(node_t *node) {
1896 /* While Unknown nodes should compute Top this is dangerous:
1897 * a Top input to a Cond would lead to BOTH control flows unreachable.
1898 * While this is correct in the given semantics, it would destroy the Firm
1901 * It would be safe to compute Top IF it can be assured, that only Cmp
1902 * nodes are inputs to Conds. We check that first.
1903 * This is the way Frontends typically build Firm, but some optimizations
1904 * (cond_eval for instance) might replace them by Phib's...
1906 node->type.tv = tarval_UNKNOWN;
1907 } /* compute_Unknown */
1910 * (Re-)compute the type for a Jmp node.
1912 * @param node the node
1914 static void compute_Jmp(node_t *node) {
1915 node_t *block = get_irn_node(get_nodes_block(node->node));
1917 node->type = block->type;
1921 * (Re-)compute the type for the Return node.
1923 * @param node the node
1925 static void compute_Return(node_t *node) {
1926 /* The Return node is NOT dead if it is in a reachable block.
1927 * This is already checked in compute(). so we can return
1928 * Reachable here. */
1929 node->type.tv = tarval_reachable;
1930 } /* compute_Return */
1933 * (Re-)compute the type for the End node.
1935 * @param node the node
1937 static void compute_End(node_t *node) {
1938 /* the End node is NOT dead of course */
1939 node->type.tv = tarval_reachable;
1943 * (Re-)compute the type for a SymConst node.
1945 * @param node the node
1947 static void compute_SymConst(node_t *node) {
1948 ir_node *irn = node->node;
1949 node_t *block = get_irn_node(get_nodes_block(irn));
1951 if (block->type.tv == tarval_unreachable) {
1952 node->type.tv = tarval_top;
1955 switch (get_SymConst_kind(irn)) {
1956 case symconst_addr_ent:
1957 /* case symconst_addr_name: cannot handle this yet */
1958 node->type.sym = get_SymConst_symbol(irn);
1961 node->type.tv = computed_value(irn);
1963 } /* compute_SymConst */
1966 * (Re-)compute the type for a Phi node.
1968 * @param node the node
1970 static void compute_Phi(node_t *node) {
1972 ir_node *phi = node->node;
1973 lattice_elem_t type;
1975 /* if a Phi is in a unreachable block, its type is TOP */
1976 node_t *block = get_irn_node(get_nodes_block(phi));
1978 if (block->type.tv == tarval_unreachable) {
1979 node->type.tv = tarval_top;
1983 /* Phi implements the Meet operation */
1984 type.tv = tarval_top;
1985 for (i = get_Phi_n_preds(phi) - 1; i >= 0; --i) {
1986 node_t *pred = get_irn_node(get_Phi_pred(phi, i));
1987 node_t *pred_X = get_irn_node(get_Block_cfgpred(block->node, i));
1989 if (pred_X->type.tv == tarval_unreachable || pred->type.tv == tarval_top) {
1990 /* ignore TOP inputs: We must check here for unreachable blocks,
1991 because Firm constants live in the Start Block are NEVER Top.
1992 Else, a Phi (1,2) will produce Bottom, even if the 2 for instance
1993 comes from a unreachable input. */
1996 if (pred->type.tv == tarval_bottom) {
1997 node->type.tv = tarval_bottom;
1999 } else if (type.tv == tarval_top) {
2000 /* first constant found */
2002 } else if (type.tv != pred->type.tv) {
2003 /* different constants or tarval_bottom */
2004 node->type.tv = tarval_bottom;
2007 /* else nothing, constants are the same */
2013 * (Re-)compute the type for an Add. Special case: one nodes is a Zero Const.
2015 * @param node the node
2017 static void compute_Add(node_t *node) {
2018 ir_node *sub = node->node;
2019 node_t *l = get_irn_node(get_Add_left(sub));
2020 node_t *r = get_irn_node(get_Add_right(sub));
2021 lattice_elem_t a = l->type;
2022 lattice_elem_t b = r->type;
2025 if (a.tv == tarval_top || b.tv == tarval_top) {
2026 node->type.tv = tarval_top;
2027 } else if (a.tv == tarval_bottom || b.tv == tarval_bottom) {
2028 node->type.tv = tarval_bottom;
2030 /* x + 0 = 0 + x = x, but beware of floating point +0 + -0, so we
2031 must call tarval_add() first to handle this case! */
2032 if (is_tarval(a.tv)) {
2033 if (is_tarval(b.tv)) {
2034 node->type.tv = tarval_add(a.tv, b.tv);
2037 mode = get_tarval_mode(a.tv);
2038 if (a.tv == get_mode_null(mode)) {
2042 } else if (is_tarval(b.tv)) {
2043 mode = get_tarval_mode(b.tv);
2044 if (b.tv == get_mode_null(mode)) {
2049 node->type.tv = tarval_bottom;
2054 * (Re-)compute the type for a Sub. Special case: both nodes are congruent.
2056 * @param node the node
2058 static void compute_Sub(node_t *node) {
2059 ir_node *sub = node->node;
2060 node_t *l = get_irn_node(get_Sub_left(sub));
2061 node_t *r = get_irn_node(get_Sub_right(sub));
2062 lattice_elem_t a = l->type;
2063 lattice_elem_t b = r->type;
2066 if (a.tv == tarval_top || b.tv == tarval_top) {
2067 node->type.tv = tarval_top;
2068 } else if (is_con(a) && is_con(b)) {
2069 if (is_tarval(a.tv) && is_tarval(b.tv)) {
2070 node->type.tv = tarval_sub(a.tv, b.tv, get_irn_mode(sub));
2071 } else if (is_tarval(a.tv) && tarval_is_null(a.tv)) {
2073 } else if (is_tarval(b.tv) && tarval_is_null(b.tv)) {
2076 node->type.tv = tarval_bottom;
2078 node->by_all_const = 1;
2079 } else if (r->part == l->part &&
2080 (!mode_is_float(get_irn_mode(l->node)))) {
2082 * BEWARE: a - a is NOT always 0 for floating Point values, as
2083 * NaN op NaN = NaN, so we must check this here.
2085 ir_mode *mode = get_irn_mode(sub);
2086 tv = get_mode_null(mode);
2088 /* if the node was ONCE evaluated by all constants, but now
2089 this breaks AND we get from the argument partitions a different
2090 result, switch to bottom.
2091 This happens because initially all nodes are in the same partition ... */
2092 if (node->by_all_const && node->type.tv != tv)
2096 node->type.tv = tarval_bottom;
2101 * (Re-)compute the type for an Eor. Special case: both nodes are congruent.
2103 * @param node the node
2105 static void compute_Eor(node_t *node) {
2106 ir_node *eor = node->node;
2107 node_t *l = get_irn_node(get_Eor_left(eor));
2108 node_t *r = get_irn_node(get_Eor_right(eor));
2109 lattice_elem_t a = l->type;
2110 lattice_elem_t b = r->type;
2113 if (a.tv == tarval_top || b.tv == tarval_top) {
2114 node->type.tv = tarval_top;
2115 } else if (is_con(a) && is_con(b)) {
2116 if (is_tarval(a.tv) && is_tarval(b.tv)) {
2117 node->type.tv = tarval_eor(a.tv, b.tv);
2118 } else if (is_tarval(a.tv) && tarval_is_null(a.tv)) {
2120 } else if (is_tarval(b.tv) && tarval_is_null(b.tv)) {
2123 node->type.tv = tarval_bottom;
2125 node->by_all_const = 1;
2126 } else if (r->part == l->part) {
2127 ir_mode *mode = get_irn_mode(eor);
2128 tv = get_mode_null(mode);
2130 /* if the node was ONCE evaluated by all constants, but now
2131 this breaks AND we get from the argument partitions a different
2132 result, switch to bottom.
2133 This happens because initially all nodes are in the same partition ... */
2134 if (node->by_all_const && node->type.tv != tv)
2138 node->type.tv = tarval_bottom;
2143 * (Re-)compute the type for Cmp.
2145 * @param node the node
2147 static void compute_Cmp(node_t *node) {
2148 ir_node *cmp = node->node;
2149 node_t *l = get_irn_node(get_Cmp_left(cmp));
2150 node_t *r = get_irn_node(get_Cmp_right(cmp));
2151 lattice_elem_t a = l->type;
2152 lattice_elem_t b = r->type;
2154 if (a.tv == tarval_top || b.tv == tarval_top) {
2155 node->type.tv = tarval_top;
2156 } else if (r->part == l->part) {
2157 /* both nodes congruent, we can probably do something */
2158 node->type.tv = tarval_b_true;
2159 } else if (is_con(a) && is_con(b)) {
2160 /* both nodes are constants, we can probably do something */
2161 node->type.tv = tarval_b_true;
2163 node->type.tv = tarval_bottom;
2168 * (Re-)compute the type for a Proj(Cmp).
2170 * @param node the node
2171 * @param cond the predecessor Cmp node
2173 static void compute_Proj_Cmp(node_t *node, ir_node *cmp) {
2174 ir_node *proj = node->node;
2175 node_t *l = get_irn_node(get_Cmp_left(cmp));
2176 node_t *r = get_irn_node(get_Cmp_right(cmp));
2177 lattice_elem_t a = l->type;
2178 lattice_elem_t b = r->type;
2179 pn_Cmp pnc = get_Proj_proj(proj);
2182 if (a.tv == tarval_top || b.tv == tarval_top) {
2183 node->type.tv = tarval_undefined;
2184 } else if (is_con(a) && is_con(b)) {
2185 default_compute(node);
2186 node->by_all_const = 1;
2187 } else if (r->part == l->part &&
2188 (!mode_is_float(get_irn_mode(l->node)) || pnc == pn_Cmp_Lt || pnc == pn_Cmp_Gt)) {
2190 * BEWARE: a == a is NOT always True for floating Point values, as
2191 * NaN != NaN is defined, so we must check this here.
2193 tv = pnc & pn_Cmp_Eq ? tarval_b_true: tarval_b_false;
2195 /* if the node was ONCE evaluated by all constants, but now
2196 this breaks AND we get from the argument partitions a different
2197 result, switch to bottom.
2198 This happens because initially all nodes are in the same partition ... */
2199 if (node->by_all_const && node->type.tv != tv)
2203 node->type.tv = tarval_bottom;
2205 } /* compute_Proj_Cmp */
2208 * (Re-)compute the type for a Proj(Cond).
2210 * @param node the node
2211 * @param cond the predecessor Cond node
2213 static void compute_Proj_Cond(node_t *node, ir_node *cond) {
2214 ir_node *proj = node->node;
2215 long pnc = get_Proj_proj(proj);
2216 ir_node *sel = get_Cond_selector(cond);
2217 node_t *selector = get_irn_node(sel);
2220 * Note: it is crucial for the monotony that the Proj(Cond)
2221 * are evaluates after all predecessors of the Cond selector are
2227 * Due to the fact that 0 is a const, the Cmp gets immediately
2228 * on the cprop list. It will be evaluated before x is evaluated,
2229 * might leaving x as Top. When later x is evaluated, the Cmp
2230 * might change its value.
2231 * BUT if the Cond is evaluated before this happens, Proj(Cond, FALSE)
2232 * gets R, and later changed to F if Cmp is evaluated to True!
2234 * We prevent this by putting Conds in an extra cprop_X queue, which
2235 * gets evaluated after the cprop queue is empty.
2237 * Note that this even happens with Click's original algorithm, if
2238 * Cmp(x, 0) is evaluated to True first and later changed to False
2239 * if x was Top first tiome and later Const ...
2240 * It is unclear how Click solved that problem ...
2242 if (get_irn_mode(sel) == mode_b) {
2244 if (pnc == pn_Cond_true) {
2245 if (selector->type.tv == tarval_b_false) {
2246 node->type.tv = tarval_unreachable;
2247 } else if (selector->type.tv == tarval_b_true) {
2248 node->type.tv = tarval_reachable;
2249 } else if (selector->type.tv == tarval_bottom) {
2250 node->type.tv = tarval_reachable;
2252 assert(selector->type.tv == tarval_top);
2253 /* any condition based on Top is "!=" */
2254 node->type.tv = tarval_unreachable;
2257 assert(pnc == pn_Cond_false);
2259 if (selector->type.tv == tarval_b_false) {
2260 node->type.tv = tarval_reachable;
2261 } else if (selector->type.tv == tarval_b_true) {
2262 node->type.tv = tarval_unreachable;
2263 } else if (selector->type.tv == tarval_bottom) {
2264 node->type.tv = tarval_reachable;
2266 assert(selector->type.tv == tarval_top);
2267 /* any condition based on Top is "!=" */
2268 node->type.tv = tarval_reachable;
2273 if (selector->type.tv == tarval_bottom) {
2274 node->type.tv = tarval_reachable;
2275 } else if (selector->type.tv == tarval_top) {
2276 if (pnc == get_Cond_defaultProj(cond)) {
2277 /* a switch based of Top is always "default" */
2278 node->type.tv = tarval_reachable;
2280 node->type.tv = tarval_unreachable;
2282 long value = get_tarval_long(selector->type.tv);
2283 if (pnc == get_Cond_defaultProj(cond)) {
2284 /* default switch, have to check ALL other cases */
2287 for (i = get_irn_n_outs(cond) - 1; i >= 0; --i) {
2288 ir_node *succ = get_irn_out(cond, i);
2292 if (value == get_Proj_proj(succ)) {
2293 /* we found a match, will NOT take the default case */
2294 node->type.tv = tarval_unreachable;
2298 /* all cases checked, no match, will take default case */
2299 node->type.tv = tarval_reachable;
2302 node->type.tv = value == pnc ? tarval_reachable : tarval_unreachable;
2306 } /* compute_Proj_Cond */
2309 * (Re-)compute the type for a Proj-Node.
2311 * @param node the node
2313 static void compute_Proj(node_t *node) {
2314 ir_node *proj = node->node;
2315 ir_mode *mode = get_irn_mode(proj);
2316 node_t *block = get_irn_node(get_nodes_block(skip_Proj(proj)));
2317 ir_node *pred = get_Proj_pred(proj);
2319 if (block->type.tv == tarval_unreachable) {
2320 /* a Proj in a unreachable Block stay Top */
2321 node->type.tv = tarval_top;
2324 if (get_irn_node(pred)->type.tv == tarval_top && !is_Cond(pred)) {
2325 /* if the predecessor is Top, its Proj follow */
2326 node->type.tv = tarval_top;
2330 if (mode == mode_M) {
2331 /* mode M is always bottom */
2332 node->type.tv = tarval_bottom;
2335 if (mode != mode_X) {
2337 compute_Proj_Cmp(node, pred);
2339 default_compute(node);
2342 /* handle mode_X nodes */
2344 switch (get_irn_opcode(pred)) {
2346 /* the Proj_X from the Start is always reachable.
2347 However this is already handled at the top. */
2348 node->type.tv = tarval_reachable;
2351 compute_Proj_Cond(node, pred);
2354 default_compute(node);
2356 } /* compute_Proj */
2359 * (Re-)compute the type for a Confirm.
2361 * @param node the node
2363 static void compute_Confirm(node_t *node) {
2364 ir_node *confirm = node->node;
2365 node_t *pred = get_irn_node(get_Confirm_value(confirm));
2367 if (get_Confirm_cmp(confirm) == pn_Cmp_Eq) {
2368 node_t *bound = get_irn_node(get_Confirm_bound(confirm));
2370 if (is_con(bound->type)) {
2371 /* is equal to a constant */
2372 node->type = bound->type;
2376 /* a Confirm is a copy OR a Const */
2377 node->type = pred->type;
2378 } /* compute_Confirm */
2381 * (Re-)compute the type for a Max.
2383 * @param node the node
2385 static void compute_Max(node_t *node) {
2386 ir_node *op = node->node;
2387 node_t *l = get_irn_node(get_binop_left(op));
2388 node_t *r = get_irn_node(get_binop_right(op));
2389 lattice_elem_t a = l->type;
2390 lattice_elem_t b = r->type;
2392 if (a.tv == tarval_top || b.tv == tarval_top) {
2393 node->type.tv = tarval_top;
2394 } else if (is_con(a) && is_con(b)) {
2395 /* both nodes are constants, we can probably do something */
2397 /* this case handles SymConsts as well */
2400 ir_mode *mode = get_irn_mode(op);
2401 tarval *tv_min = get_mode_min(mode);
2405 else if (b.tv == tv_min)
2407 else if (is_tarval(a.tv) && is_tarval(b.tv)) {
2408 if (tarval_cmp(a.tv, b.tv) & pn_Cmp_Gt)
2409 node->type.tv = a.tv;
2411 node->type.tv = b.tv;
2413 node->type.tv = tarval_bad;
2416 } else if (r->part == l->part) {
2417 /* both nodes congruent, we can probably do something */
2420 node->type.tv = tarval_bottom;
2425 * (Re-)compute the type for a Min.
2427 * @param node the node
2429 static void compute_Min(node_t *node) {
2430 ir_node *op = node->node;
2431 node_t *l = get_irn_node(get_binop_left(op));
2432 node_t *r = get_irn_node(get_binop_right(op));
2433 lattice_elem_t a = l->type;
2434 lattice_elem_t b = r->type;
2436 if (a.tv == tarval_top || b.tv == tarval_top) {
2437 node->type.tv = tarval_top;
2438 } else if (is_con(a) && is_con(b)) {
2439 /* both nodes are constants, we can probably do something */
2441 /* this case handles SymConsts as well */
2444 ir_mode *mode = get_irn_mode(op);
2445 tarval *tv_max = get_mode_max(mode);
2449 else if (b.tv == tv_max)
2451 else if (is_tarval(a.tv) && is_tarval(b.tv)) {
2452 if (tarval_cmp(a.tv, b.tv) & pn_Cmp_Gt)
2453 node->type.tv = a.tv;
2455 node->type.tv = b.tv;
2457 node->type.tv = tarval_bad;
2460 } else if (r->part == l->part) {
2461 /* both nodes congruent, we can probably do something */
2464 node->type.tv = tarval_bottom;
2469 * (Re-)compute the type for a given node.
2471 * @param node the node
2473 static void compute(node_t *node) {
2474 ir_node *irn = node->node;
2478 if (is_no_Block(irn)) {
2479 /* for pinned nodes, check its control input */
2480 if (get_irn_pinned(skip_Proj(irn)) == op_pin_state_pinned) {
2481 node_t *block = get_irn_node(get_nodes_block(irn));
2483 if (block->type.tv == tarval_unreachable) {
2484 node->type.tv = tarval_top;
2490 func = (compute_func)node->node->op->ops.generic;
2498 * Identity functions: Note that one might thing that identity() is just a
2499 * synonym for equivalent_node(). While this is true, we cannot use it for the algorithm
2500 * here, because it expects that the identity node is one of the inputs, which is NOT
2501 * always true for equivalent_node() which can handle (and does sometimes) DAGs.
2502 * So, we have our own implementation, which copies some parts of equivalent_node()
2506 * Calculates the Identity for Phi nodes
2508 static node_t *identity_Phi(node_t *node) {
2509 ir_node *phi = node->node;
2510 ir_node *block = get_nodes_block(phi);
2511 node_t *n_part = NULL;
2514 for (i = get_Phi_n_preds(phi) - 1; i >= 0; --i) {
2515 node_t *pred_X = get_irn_node(get_Block_cfgpred(block, i));
2517 if (pred_X->type.tv == tarval_reachable) {
2518 node_t *pred = get_irn_node(get_Phi_pred(phi, i));
2522 else if (n_part->part != pred->part) {
2523 /* incongruent inputs, not a follower */
2528 /* if n_part is NULL here, all inputs path are dead, the Phi computes
2529 * tarval_top, is in the TOP partition and should NOT being split! */
2530 assert(n_part != NULL);
2532 } /* identity_Phi */
2535 * Calculates the Identity for commutative 0 neutral nodes.
2537 static node_t *identity_comm_zero_binop(node_t *node) {
2538 ir_node *op = node->node;
2539 node_t *a = get_irn_node(get_binop_left(op));
2540 node_t *b = get_irn_node(get_binop_right(op));
2541 ir_mode *mode = get_irn_mode(op);
2544 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
2545 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
2548 /* node: no input should be tarval_top, else the binop would be also
2549 * Top and not being split. */
2550 zero = get_mode_null(mode);
2551 if (a->type.tv == zero)
2553 if (b->type.tv == zero)
2556 } /* identity_comm_zero_binop */
2559 * Calculates the Identity for Shift nodes.
2561 static node_t *identity_shift(node_t *node) {
2562 ir_node *op = node->node;
2563 node_t *b = get_irn_node(get_binop_right(op));
2564 ir_mode *mode = get_irn_mode(b->node);
2567 /* node: no input should be tarval_top, else the binop would be also
2568 * Top and not being split. */
2569 zero = get_mode_null(mode);
2570 if (b->type.tv == zero)
2571 return get_irn_node(get_binop_left(op));
2573 } /* identity_shift */
2576 * Calculates the Identity for Mul nodes.
2578 static node_t *identity_Mul(node_t *node) {
2579 ir_node *op = node->node;
2580 node_t *a = get_irn_node(get_Mul_left(op));
2581 node_t *b = get_irn_node(get_Mul_right(op));
2582 ir_mode *mode = get_irn_mode(op);
2585 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
2586 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
2589 /* node: no input should be tarval_top, else the binop would be also
2590 * Top and not being split. */
2591 one = get_mode_one(mode);
2592 if (a->type.tv == one)
2594 if (b->type.tv == one)
2597 } /* identity_Mul */
2600 * Calculates the Identity for Sub nodes.
2602 static node_t *identity_Sub(node_t *node) {
2603 ir_node *sub = node->node;
2604 node_t *b = get_irn_node(get_Sub_right(sub));
2605 ir_mode *mode = get_irn_mode(sub);
2607 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
2608 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
2611 /* node: no input should be tarval_top, else the binop would be also
2612 * Top and not being split. */
2613 if (b->type.tv == get_mode_null(mode))
2614 return get_irn_node(get_Sub_left(sub));
2616 } /* identity_Sub */
2619 * Calculates the Identity for And nodes.
2621 static node_t *identity_And(node_t *node) {
2622 ir_node *and = node->node;
2623 node_t *a = get_irn_node(get_And_left(and));
2624 node_t *b = get_irn_node(get_And_right(and));
2625 tarval *neutral = get_mode_all_one(get_irn_mode(and));
2627 /* node: no input should be tarval_top, else the And would be also
2628 * Top and not being split. */
2629 if (a->type.tv == neutral)
2631 if (b->type.tv == neutral)
2634 } /* identity_And */
2637 * Calculates the Identity for Confirm nodes.
2639 static node_t *identity_Confirm(node_t *node) {
2640 ir_node *confirm = node->node;
2642 /* a Confirm is always a Copy */
2643 return get_irn_node(get_Confirm_value(confirm));
2644 } /* identity_Confirm */
2647 * Calculates the Identity for Mux nodes.
2649 static node_t *identity_Mux(node_t *node) {
2650 ir_node *mux = node->node;
2651 node_t *t = get_irn_node(get_Mux_true(mux));
2652 node_t *f = get_irn_node(get_Mux_false(mux));
2655 if (t->part == f->part)
2658 /* for now, the 1-input identity is not supported */
2660 sel = get_irn_node(get_Mux_sel(mux));
2662 /* Mux sel input is mode_b, so it is always a tarval */
2663 if (sel->type.tv == tarval_b_true)
2665 if (sel->type.tv == tarval_b_false)
2669 } /* identity_Mux */
2672 * Calculates the Identity for Min nodes.
2674 static node_t *identity_Min(node_t *node) {
2675 ir_node *op = node->node;
2676 node_t *a = get_irn_node(get_binop_left(op));
2677 node_t *b = get_irn_node(get_binop_right(op));
2678 ir_mode *mode = get_irn_mode(op);
2681 if (a->part == b->part) {
2682 /* leader of multiple predecessors */
2686 /* works even with NaN */
2687 tv_max = get_mode_max(mode);
2688 if (a->type.tv == tv_max)
2690 if (b->type.tv == tv_max)
2693 } /* identity_Min */
2696 * Calculates the Identity for Max nodes.
2698 static node_t *identity_Max(node_t *node) {
2699 ir_node *op = node->node;
2700 node_t *a = get_irn_node(get_binop_left(op));
2701 node_t *b = get_irn_node(get_binop_right(op));
2702 ir_mode *mode = get_irn_mode(op);
2705 if (a->part == b->part) {
2706 /* leader of multiple predecessors */
2710 /* works even with NaN */
2711 tv_min = get_mode_min(mode);
2712 if (a->type.tv == tv_min)
2714 if (b->type.tv == tv_min)
2717 } /* identity_Max */
2720 * Calculates the Identity for nodes.
2722 static node_t *identity(node_t *node) {
2723 ir_node *irn = node->node;
2725 switch (get_irn_opcode(irn)) {
2727 return identity_Phi(node);
2729 return identity_Mul(node);
2733 return identity_comm_zero_binop(node);
2738 return identity_shift(node);
2740 return identity_And(node);
2742 return identity_Sub(node);
2744 return identity_Confirm(node);
2746 return identity_Mux(node);
2748 return identity_Min(node);
2750 return identity_Max(node);
2757 * Node follower is a (new) follower of leader, segregate Leader
2760 static void segregate_def_use_chain_1(const ir_node *follower, node_t *leader) {
2761 ir_node *l = leader->node;
2762 int j, i, n = get_irn_n_outs(l);
2764 DB((dbg, LEVEL_2, "%+F is a follower of %+F\n", follower, leader->node));
2765 /* The leader edges must remain sorted, but follower edges can
2767 for (i = leader->n_followers + 1; i <= n; ++i) {
2768 if (l->out[i].use == follower) {
2769 ir_def_use_edge t = l->out[i];
2771 for (j = i - 1; j >= leader->n_followers + 1; --j)
2772 l->out[j + 1] = l->out[j];
2773 ++leader->n_followers;
2774 l->out[leader->n_followers] = t;
2778 } /* segregate_def_use_chain_1 */
2781 * Node follower is a (new) follower segregate its Leader
2784 * @param follower the follower IR node
2786 static void segregate_def_use_chain(const ir_node *follower) {
2789 for (i = get_irn_arity(follower) - 1; i >= 0; --i) {
2790 node_t *pred = get_irn_node(get_irn_n(follower, i));
2792 segregate_def_use_chain_1(follower, pred);
2794 } /* segregate_def_use_chain */
2797 * Propagate constant evaluation.
2799 * @param env the environment
2801 static void propagate(environment_t *env) {
2804 lattice_elem_t old_type;
2806 unsigned n_fallen, old_type_was_T_or_C;
2809 while (env->cprop != NULL) {
2810 void *oldopcode = NULL;
2812 /* remove the first partition X from cprop */
2815 env->cprop = X->cprop_next;
2817 old_type_was_T_or_C = X->type_is_T_or_C;
2819 DB((dbg, LEVEL_2, "Propagate type on part%d\n", X->nr));
2823 int cprop_empty = list_empty(&X->cprop);
2824 int cprop_X_empty = list_empty(&X->cprop_X);
2826 if (cprop_empty && cprop_X_empty) {
2827 /* both cprop lists are empty */
2831 /* remove the first Node x from X.cprop */
2833 /* Get a node from the cprop_X list only if
2834 * all data nodes are processed.
2835 * This ensures, that all inputs of the Cond
2836 * predecessor are processed if its type is still Top.
2838 x = list_entry(X->cprop_X.next, node_t, cprop_list);
2840 x = list_entry(X->cprop.next, node_t, cprop_list);
2843 //assert(x->part == X);
2844 list_del(&x->cprop_list);
2847 if (x->is_follower && identity(x) == x) {
2848 /* check the opcode first */
2849 if (oldopcode == NULL) {
2850 oldopcode = lambda_opcode(get_first_node(X), env);
2852 if (oldopcode != lambda_opcode(x, env)) {
2853 if (x->on_fallen == 0) {
2854 /* different opcode -> x falls out of this partition */
2859 DB((dbg, LEVEL_2, "Add node %+F to fallen\n", x->node));
2863 /* x will make the follower -> leader transition */
2864 follower_to_leader(x);
2867 /* compute a new type for x */
2869 DB((dbg, LEVEL_3, "computing type of %+F\n", x->node));
2871 if (x->type.tv != old_type.tv) {
2872 DB((dbg, LEVEL_2, "node %+F has changed type from %+F to %+F\n", x->node, old_type, x->type));
2874 if (x->on_fallen == 0) {
2875 /* Add x to fallen. Nodes might fall from T -> const -> _|_, so check that they are
2876 not already on the list. */
2881 DB((dbg, LEVEL_2, "Add node %+F to fallen\n", x->node));
2883 for (i = get_irn_n_outs(x->node) - 1; i >= 0; --i) {
2884 ir_node *succ = get_irn_out(x->node, i);
2885 node_t *y = get_irn_node(succ);
2887 /* Add y to y.partition.cprop. */
2888 add_to_cprop(y, env);
2893 if (n_fallen > 0 && n_fallen != X->n_leader) {
2894 DB((dbg, LEVEL_2, "Splitting part%d by fallen\n", X->nr));
2895 Y = split(&X, fallen, env);
2897 * We have split out fallen node. The type of the result
2898 * partition is NOT set yet.
2900 Y->type_is_T_or_C = 0;
2904 /* remove the flags from the fallen list */
2905 for (x = fallen; x != NULL; x = x->next)
2908 if (old_type_was_T_or_C) {
2911 /* check if some nodes will make the leader -> follower transition */
2912 list_for_each_entry_safe(node_t, y, tmp, &Y->Leader, node_list) {
2913 if (y->type.tv != tarval_top && ! is_con(y->type)) {
2914 node_t *eq_node = identity(y);
2916 if (eq_node != y && eq_node->part == y->part) {
2917 DB((dbg, LEVEL_2, "Node %+F is a follower of %+F\n", y->node, eq_node->node));
2918 /* move to Follower */
2920 list_del(&y->node_list);
2921 list_add_tail(&y->node_list, &Y->Follower);
2924 segregate_def_use_chain(y->node);
2934 * Get the leader for a given node from its congruence class.
2936 * @param irn the node
2938 static ir_node *get_leader(node_t *node) {
2939 partition_t *part = node->part;
2941 if (part->n_leader > 1 || node->is_follower) {
2942 if (node->is_follower) {
2943 DB((dbg, LEVEL_2, "Replacing follower %+F\n", node->node));
2946 DB((dbg, LEVEL_2, "Found congruence class for %+F\n", node->node));
2948 return get_first_node(part)->node;
2954 * Return non-zero if the control flow predecessor node pred
2955 * is the only reachable control flow exit of its block.
2957 * @param pred the control flow exit
2959 static int can_exchange(ir_node *pred) {
2962 else if (is_Jmp(pred))
2964 else if (get_irn_mode(pred) == mode_T) {
2967 /* if the predecessor block has more than one
2968 reachable outputs we cannot remove the block */
2970 for (i = get_irn_n_outs(pred) - 1; i >= 0; --i) {
2971 ir_node *proj = get_irn_out(pred, i);
2974 /* skip non-control flow Proj's */
2975 if (get_irn_mode(proj) != mode_X)
2978 node = get_irn_node(proj);
2979 if (node->type.tv == tarval_reachable) {
2987 } /* can_exchange */
2990 * Block Post-Walker, apply the analysis results on control flow by
2991 * shortening Phi's and Block inputs.
2993 static void apply_cf(ir_node *block, void *ctx) {
2994 environment_t *env = ctx;
2995 node_t *node = get_irn_node(block);
2997 ir_node **ins, **in_X;
2998 ir_node *phi, *next;
3000 n = get_Block_n_cfgpreds(block);
3002 if (node->type.tv == tarval_unreachable) {
3005 for (i = n - 1; i >= 0; --i) {
3006 ir_node *pred = get_Block_cfgpred(block, i);
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);
3026 /* the EndBlock is always reachable even if the analysis
3027 finds out the opposite :-) */
3028 if (block != get_irg_end_block(current_ir_graph)) {
3029 /* mark dead blocks */
3030 set_Block_dead(block);
3031 DB((dbg, LEVEL_1, "Removing dead %+F\n", block));
3033 /* the endblock is unreachable */
3034 set_irn_in(block, 0, NULL);
3040 /* only one predecessor combine */
3041 ir_node *pred = skip_Proj(get_Block_cfgpred(block, 0));
3043 if (can_exchange(pred)) {
3044 ir_node *new_block = get_nodes_block(pred);
3045 DB((dbg, LEVEL_1, "Fuse %+F with %+F\n", block, new_block));
3046 DBG_OPT_COMBO(block, new_block, FS_OPT_COMBO_CF);
3047 exchange(block, new_block);
3048 node->node = new_block;
3054 NEW_ARR_A(ir_node *, in_X, n);
3056 for (i = 0; i < n; ++i) {
3057 ir_node *pred = get_Block_cfgpred(block, i);
3058 node_t *node = get_irn_node(pred);
3060 if (node->type.tv == tarval_reachable) {
3063 DB((dbg, LEVEL_1, "Removing dead input %d from %+F (%+F)\n", i, block, pred));
3064 if (! is_Bad(pred)) {
3065 node_t *pred_bl = get_irn_node(get_nodes_block(skip_Proj(pred)));
3067 if (pred_bl->flagged == 0) {
3068 pred_bl->flagged = 3;
3070 if (pred_bl->type.tv == tarval_reachable) {
3072 * We will remove an edge from block to its pred.
3073 * This might leave the pred block as an endless loop
3075 if (! is_backedge(block, i))
3076 keep_alive(pred_bl->node);
3086 NEW_ARR_A(ir_node *, ins, n);
3087 for (phi = get_Block_phis(block); phi != NULL; phi = next) {
3088 node_t *node = get_irn_node(phi);
3090 next = get_Phi_next(phi);
3091 if (is_tarval(node->type.tv) && tarval_is_constant(node->type.tv)) {
3092 /* this Phi is replaced by a constant */
3093 tarval *tv = node->type.tv;
3094 ir_node *c = new_r_Const(current_ir_graph, block, get_tarval_mode(tv), tv);
3096 set_irn_node(c, node);
3098 DB((dbg, LEVEL_1, "%+F is replaced by %+F\n", phi, c));
3099 DBG_OPT_COMBO(phi, c, FS_OPT_COMBO_CONST);
3104 for (i = 0; i < n; ++i) {
3105 node_t *pred = get_irn_node(get_Block_cfgpred(block, i));
3107 if (pred->type.tv == tarval_reachable) {
3108 ins[j++] = get_Phi_pred(phi, i);
3112 /* this Phi is replaced by a single predecessor */
3113 ir_node *s = ins[0];
3114 node_t *phi_node = get_irn_node(phi);
3117 DB((dbg, LEVEL_1, "%+F is replaced by %+F because of cf change\n", phi, s));
3118 DBG_OPT_COMBO(phi, s, FS_OPT_COMBO_FOLLOWER);
3123 set_irn_in(phi, j, ins);
3131 /* this Block has only one live predecessor */
3132 ir_node *pred = skip_Proj(in_X[0]);
3134 if (can_exchange(pred)) {
3135 ir_node *new_block = get_nodes_block(pred);
3136 DBG_OPT_COMBO(block, new_block, FS_OPT_COMBO_CF);
3137 exchange(block, new_block);
3138 node->node = new_block;
3143 set_irn_in(block, k, in_X);
3148 * Exchange a node by its leader.
3149 * Beware: in rare cases the mode might be wrong here, for instance
3150 * AddP(x, NULL) is a follower of x, but with different mode.
3153 static void exchange_leader(ir_node *irn, ir_node *leader) {
3154 ir_mode *mode = get_irn_mode(irn);
3155 if (mode != get_irn_mode(leader)) {
3156 /* The conv is a no-op, so we are fre to place in
3157 * either in the block of the leader OR in irn's block.
3158 * Probably placing it into leaders block might reduce
3159 * the number of Conv due to CSE. */
3160 ir_node *block = get_nodes_block(leader);
3161 dbg_info *dbg = get_irn_dbg_info(irn);
3163 leader = new_rd_Conv(dbg, current_ir_graph, block, leader, mode);
3165 exchange(irn, leader);
3169 * Post-Walker, apply the analysis results;
3171 static void apply_result(ir_node *irn, void *ctx) {
3172 environment_t *env = ctx;
3173 node_t *node = get_irn_node(irn);
3175 if (is_Block(irn) || is_End(irn) || is_Bad(irn)) {
3176 /* blocks already handled, do not touch the End node */
3178 node_t *block = get_irn_node(get_nodes_block(irn));
3180 if (block->type.tv == tarval_unreachable) {
3181 ir_node *bad = get_irg_bad(current_ir_graph);
3183 /* here, bad might already have a node, but this can be safely ignored
3184 as long as bad has at least ONE valid node */
3185 set_irn_node(bad, node);
3187 DB((dbg, LEVEL_1, "%+F is unreachable\n", irn));
3191 else if (node->type.tv == tarval_top) {
3192 /* don't kick away Unknown's, they might be still needed */
3193 if (! is_Unknown(irn)) {
3194 ir_mode *mode = get_irn_mode(irn);
3195 ir_node *unk = new_r_Unknown(current_ir_graph, mode);
3197 /* control flow should already be handled at apply_cf() */
3198 assert(mode != mode_X);
3200 /* see comment above */
3201 set_irn_node(unk, node);
3203 DB((dbg, LEVEL_1, "%+F computes Top\n", irn));
3208 else if (get_irn_mode(irn) == mode_X) {
3211 ir_node *cond = get_Proj_pred(irn);
3213 if (is_Cond(cond)) {
3214 node_t *sel = get_irn_node(get_Cond_selector(cond));
3216 if (is_tarval(sel->type.tv) && tarval_is_constant(sel->type.tv)) {
3217 /* Cond selector is a constant and the Proj is reachable, make a Jmp */
3218 ir_node *jmp = new_r_Jmp(current_ir_graph, block->node);
3219 set_irn_node(jmp, node);
3221 DB((dbg, LEVEL_1, "%+F is replaced by %+F\n", irn, jmp));
3222 DBG_OPT_COMBO(irn, jmp, FS_OPT_COMBO_CF);
3229 /* normal data node */
3230 if (is_tarval(node->type.tv) && tarval_is_constant(node->type.tv)) {
3231 tarval *tv = node->type.tv;
3234 * Beware: never replace mode_T nodes by constants. Currently we must mark
3235 * mode_T nodes with constants, but do NOT replace them.
3237 if (! is_Const(irn) && get_irn_mode(irn) != mode_T) {
3238 /* can be replaced by a constant */
3239 ir_node *c = new_r_Const(current_ir_graph, block->node, get_tarval_mode(tv), tv);
3240 set_irn_node(c, node);
3242 DB((dbg, LEVEL_1, "%+F is replaced by %+F\n", irn, c));
3243 DBG_OPT_COMBO(irn, c, FS_OPT_COMBO_CONST);
3244 exchange_leader(irn, c);
3247 } else if (is_entity(node->type.sym.entity_p)) {
3248 if (! is_SymConst(irn)) {
3249 /* can be replaced by a SymConst */
3250 ir_node *symc = new_r_SymConst(current_ir_graph, block->node, get_irn_mode(irn), node->type.sym, symconst_addr_ent);
3251 set_irn_node(symc, node);
3254 DB((dbg, LEVEL_1, "%+F is replaced by %+F\n", irn, symc));
3255 DBG_OPT_COMBO(irn, symc, FS_OPT_COMBO_CONST);
3256 exchange_leader(irn, symc);
3259 } else if (is_Confirm(irn)) {
3260 /* Confirms are always follower, but do not kill them here */
3262 ir_node *leader = get_leader(node);
3264 if (leader != irn) {
3265 int non_strict_phi = 0;
3268 * Beware: Do not remove Phi(Unknown, ..., x, ..., Unknown)
3269 * as this might create non-strict programs.
3271 if (node->is_follower && is_Phi(irn) && !is_Unknown(leader)) {
3274 for (i = get_Phi_n_preds(irn) - 1; i >= 0; --i) {
3275 ir_node *pred = get_Phi_pred(irn, i);
3277 if (is_Unknown(pred)) {
3283 if (! non_strict_phi) {
3284 DB((dbg, LEVEL_1, "%+F from part%d is replaced by %+F\n", irn, node->part->nr, leader));
3285 if (node->is_follower)
3286 DBG_OPT_COMBO(irn, leader, FS_OPT_COMBO_FOLLOWER);
3288 DBG_OPT_COMBO(irn, leader, FS_OPT_COMBO_CONGRUENT);
3289 exchange_leader(irn, leader);
3296 } /* apply_result */
3299 * Fix the keep-alives by deleting unreachable ones.
3301 static void apply_end(ir_node *end, environment_t *env) {
3302 int i, j, n = get_End_n_keepalives(end);
3306 NEW_ARR_A(ir_node *, in, n);
3308 /* fix the keep alive */
3309 for (i = j = 0; i < n; i++) {
3310 ir_node *ka = get_End_keepalive(end, i);
3311 node_t *node = get_irn_node(ka);
3314 node = get_irn_node(get_nodes_block(ka));
3316 if (node->type.tv != tarval_unreachable && !is_Bad(ka))
3320 set_End_keepalives(end, j, in);
3325 #define SET(code) op_##code->ops.generic = (op_func)compute_##code
3328 * sets the generic functions to compute.
3330 static void set_compute_functions(void) {
3333 /* set the default compute function */
3334 for (i = get_irp_n_opcodes() - 1; i >= 0; --i) {
3335 ir_op *op = get_irp_opcode(i);
3336 op->ops.generic = (op_func)default_compute;
3339 /* set specific functions */
3360 } /* set_compute_functions */
3362 void combo(ir_graph *irg) {
3364 ir_node *initial_bl;
3366 ir_graph *rem = current_ir_graph;
3368 current_ir_graph = irg;
3370 /* register a debug mask */
3371 FIRM_DBG_REGISTER(dbg, "firm.opt.combo");
3373 DB((dbg, LEVEL_1, "Doing COMBO for %+F\n", irg));
3375 obstack_init(&env.obst);
3376 env.worklist = NULL;
3380 #ifdef DEBUG_libfirm
3381 env.dbg_list = NULL;
3383 env.opcode2id_map = new_set(cmp_opcode, iro_Last * 4);
3384 env.type2id_map = pmap_create();
3385 env.end_idx = get_opt_global_cse() ? 0 : -1;
3386 env.lambda_input = 0;
3387 env.commutative = 1;
3390 assure_irg_outs(irg);
3391 assure_cf_loop(irg);
3393 /* we have our own value_of function */
3394 set_value_of_func(get_node_tarval);
3396 set_compute_functions();
3397 DEBUG_ONLY(part_nr = 0);
3399 ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK);
3401 /* create the initial partition and place it on the work list */
3402 env.initial = new_partition(&env);
3403 add_to_worklist(env.initial, &env);
3404 irg_walk_graph(irg, create_initial_partitions, init_block_phis, &env);
3406 /* set the hook: from now, every node has a partition and a type */
3407 DEBUG_ONLY(set_dump_node_vcgattr_hook(dump_partition_hook));
3409 /* all nodes on the initial partition have type Top */
3410 env.initial->type_is_T_or_C = 1;
3412 /* Place the START Node's partition on cprop.
3413 Place the START Node on its local worklist. */
3414 initial_bl = get_irg_start_block(irg);
3415 start = get_irn_node(initial_bl);
3416 add_to_cprop(start, &env);
3420 if (env.worklist != NULL)
3422 } while (env.cprop != NULL || env.worklist != NULL);
3424 dump_all_partitions(&env);
3425 check_all_partitions(&env);
3428 dump_ir_block_graph(irg, "-partition");
3431 /* apply the result */
3432 irg_block_walk_graph(irg, NULL, apply_cf, &env);
3433 irg_walk_graph(irg, NULL, apply_result, &env);
3434 apply_end(get_irg_end(irg), &env);
3437 /* control flow might changed */
3438 set_irg_outs_inconsistent(irg);
3439 set_irg_extblk_inconsistent(irg);
3440 set_irg_doms_inconsistent(irg);
3441 set_irg_loopinfo_inconsistent(irg);
3444 ir_free_resources(irg, IR_RESOURCE_IRN_LINK);
3446 /* remove the partition hook */
3447 DEBUG_ONLY(set_dump_node_vcgattr_hook(NULL));
3449 pmap_destroy(env.type2id_map);
3450 del_set(env.opcode2id_map);
3451 obstack_free(&env.obst, NULL);
3453 /* restore value_of() default behavior */
3454 set_value_of_func(NULL);
3455 current_ir_graph = rem;