2 * Copyright (C) 1995-2011 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
25 * This is a slightly enhanced version of Cliff Clicks combo algorithm
26 * - support for commutative nodes is added, Add(a,b) and Add(b,a) ARE congruent
27 * - supports all Firm direct (by a data edge) identities except Mux
28 * (Mux can be a 2-input or 1-input identity, only 2-input is implemented yet)
29 * - supports Confirm nodes (handle them like Copies but do NOT remove them)
30 * - let Cmp nodes calculate Top like all othe data nodes: this would let
31 * Mux nodes to calculate Unknown instead of taking the true result
32 * - let Cond(Top) always select FALSE/default: This is tricky. Nodes are only reavaluated
33 * IFF the predecessor changed its type. Because nodes are initialized with Top
34 * this never happens, let all Proj(Cond) be unreachable.
35 * We avoid this condition by the same way we work around Phi: whenever a Block
36 * node is placed on the list, place its Cond nodes (and because they are Tuple
37 * all its Proj-nodes either on the cprop list)
38 * Especially, this changes the meaning of Click's example:
53 * using Click's version while is silent with our.
54 * - support for global congruences is implemented but not tested yet
56 * Note further that we use the terminology from Click's work here, which is different
57 * in some cases from Firm terminology. Especially, Click's type is a
58 * Firm tarval/entity, nevertheless we call it type here for "maximum compatibility".
64 #include "iroptimize.h"
71 #include "irgraph_t.h"
78 #include "iropt_dbg.h"
82 #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_node *irn; /**< An IR node representing this opcode. */
112 * An entry in the list_map.
114 struct listmap_entry_t {
115 void *id; /**< The id. */
116 node_t *list; /**< The associated list for this id. */
117 listmap_entry_t *next; /**< Link to the next entry in the map. */
120 /** We must map id's to lists. */
121 typedef struct listmap_t {
122 set *map; /**< Map id's to listmap_entry_t's */
123 listmap_entry_t *values; /**< List of all values in the map. */
127 * A lattice element. Because we handle constants and symbolic constants different, we
128 * have to use this union.
139 ir_node *node; /**< The IR-node itself. */
140 list_head node_list; /**< Double-linked list of leader/follower entries. */
141 list_head cprop_list; /**< Double-linked partition.cprop list. */
142 partition_t *part; /**< points to the partition this node belongs to */
143 node_t *next; /**< Next node on local list (partition.touched, fallen). */
144 node_t *race_next; /**< Next node on race list. */
145 lattice_elem_t type; /**< The associated lattice element "type". */
146 int max_user_input; /**< Maximum input number of Def-Use edges. */
147 int next_edge; /**< Index of the next Def-Use edge to use. */
148 int n_followers; /**< Number of Follower in the outs set. */
149 unsigned on_touched:1; /**< Set, if this node is on the partition.touched set. */
150 unsigned on_cprop:1; /**< Set, if this node is on the partition.cprop list. */
151 unsigned on_fallen:1; /**< Set, if this node is on the fallen list. */
152 unsigned is_follower:1; /**< Set, if this node is a follower. */
153 unsigned flagged:2; /**< 2 Bits, set if this node was visited by race 1 or 2. */
157 * A partition containing congruent nodes.
160 list_head Leader; /**< The head of partition Leader node list. */
161 list_head Follower; /**< The head of partition Follower node list. */
162 list_head cprop; /**< The head of partition.cprop list. */
163 list_head cprop_X; /**< The head of partition.cprop (Cond nodes and its Projs) list. */
164 partition_t *wl_next; /**< Next entry in the work list if any. */
165 partition_t *touched_next; /**< Points to the next partition in the touched set. */
166 partition_t *cprop_next; /**< Points to the next partition in the cprop list. */
167 partition_t *split_next; /**< Points to the next partition in the list that must be split by split_by(). */
168 node_t *touched; /**< The partition.touched set of this partition. */
169 unsigned n_leader; /**< Number of entries in this partition.Leader. */
170 unsigned n_touched; /**< Number of entries in the partition.touched. */
171 int max_user_inputs; /**< Maximum number of user inputs of all entries. */
172 unsigned on_worklist:1; /**< Set, if this partition is in the work list. */
173 unsigned on_touched:1; /**< Set, if this partition is on the touched set. */
174 unsigned on_cprop:1; /**< Set, if this partition is on the cprop list. */
175 unsigned type_is_T_or_C:1;/**< Set, if all nodes in this partition have type Top or Constant. */
177 partition_t *dbg_next; /**< Link all partitions for debugging */
178 unsigned nr; /**< A unique number for (what-)mapping, >0. */
182 typedef struct environment_t {
183 struct obstack obst; /**< obstack to allocate data structures. */
184 partition_t *worklist; /**< The work list. */
185 partition_t *cprop; /**< The constant propagation list. */
186 partition_t *touched; /**< the touched set. */
187 partition_t *initial; /**< The initial partition. */
188 set *opcode2id_map; /**< The opcodeMode->id map. */
189 ir_node **kept_memory; /**< Array of memory nodes that must be kept. */
190 int end_idx; /**< -1 for local and 0 for global congruences. */
191 int lambda_input; /**< Captured argument for lambda_partition(). */
192 unsigned modified:1; /**< Set, if the graph was modified. */
193 unsigned unopt_cf:1; /**< If set, control flow is not optimized due to Unknown. */
194 /* options driving the optimization */
195 unsigned commutative:1; /**< Set, if commutation nodes should be handled specially. */
196 unsigned opt_unknown:1; /**< Set, if non-strict programs should be optimized. */
198 partition_t *dbg_list; /**< List of all partitions. */
202 /** Type of the what function. */
203 typedef void *(*what_func)(const node_t *node, environment_t *env);
205 #define get_irn_node(irn) ((node_t *)get_irn_link(irn))
206 #define set_irn_node(irn, node) set_irn_link(irn, node)
208 /* we do NOT use tarval_unreachable here, instead we use Top for this purpose */
209 #undef tarval_unreachable
210 #define tarval_unreachable tarval_top
213 /** The debug module handle. */
214 DEBUG_ONLY(static firm_dbg_module_t *dbg;)
216 /** The what reason. */
217 DEBUG_ONLY(static const char *what_reason;)
219 /** Next partition number. */
220 DEBUG_ONLY(static unsigned part_nr = 0;)
222 /** The tarval returned by Unknown nodes: set to either tarval_bad OR tarval_top. */
223 static ir_tarval *tarval_UNKNOWN;
226 static node_t *identity(node_t *node);
229 * Compare two opcode representatives.
231 static int cmp_irn_opcode(const ir_node *a, const ir_node *b)
235 if ((get_irn_op(a) != get_irn_op(b)) ||
236 (get_irn_mode(a) != get_irn_mode(b)))
239 /* compare if a's in and b's in are of equal length */
240 arity = get_irn_arity(a);
241 if (arity != get_irn_arity(b))
246 * Some ugliness here: Two Blocks having the same
247 * IJmp predecessor would be congruent, which of course is wrong.
248 * We fix it by never letting blocks be congruent
249 * which cannot be detected by combo either.
255 * here, we already know that the nodes are identical except their
258 if (a->op->ops.node_cmp_attr)
259 return a->op->ops.node_cmp_attr(a, b);
262 } /* cmp_irn_opcode */
264 #ifdef CHECK_PARTITIONS
268 static void check_partition(const partition_t *T)
273 list_for_each_entry(node_t, node, &T->Leader, node_list) {
274 assert(node->is_follower == 0);
275 assert(node->flagged == 0);
276 assert(node->part == T);
279 assert(n == T->n_leader);
281 list_for_each_entry(node_t, node, &T->Follower, node_list) {
282 assert(node->is_follower == 1);
283 assert(node->flagged == 0);
284 assert(node->part == T);
286 } /* check_partition */
289 * check that all leader nodes in the partition have the same opcode.
291 static void check_opcode(const partition_t *Z)
294 const ir_node *repr = NULL;
296 list_for_each_entry(node_t, node, &Z->Leader, node_list) {
297 ir_node *irn = node->node;
302 assert(cmp_irn_opcode(repr, irn) == 0);
307 static void check_all_partitions(environment_t *env)
313 for (P = env->dbg_list; P != NULL; P = P->dbg_next) {
315 if (! P->type_is_T_or_C)
317 list_for_each_entry(node_t, node, &P->Follower, node_list) {
318 node_t *leader = identity(node);
320 assert(leader != node && leader->part == node->part);
331 static void do_check_list(const node_t *list, int ofs, const partition_t *Z)
336 #define NEXT(e) *((const node_t **)((char *)(e) + (ofs)))
337 for (e = list; e != NULL; e = NEXT(e)) {
338 assert(e->part == Z);
346 } /* ido_check_list */
349 * Check a local list.
351 static void check_list(const node_t *list, const partition_t *Z)
353 do_check_list(list, offsetof(node_t, next), Z);
357 #define check_partition(T)
358 #define check_list(list, Z)
359 #define check_all_partitions(env)
360 #endif /* CHECK_PARTITIONS */
363 static inline lattice_elem_t get_partition_type(const partition_t *X);
366 * Dump partition to output.
368 static void dump_partition(const char *msg, const partition_t *part)
372 lattice_elem_t type = get_partition_type(part);
374 DB((dbg, LEVEL_2, "%s part%u%s (%u, %+F) {\n ",
375 msg, part->nr, part->type_is_T_or_C ? "*" : "",
376 part->n_leader, type));
377 list_for_each_entry(node_t, node, &part->Leader, node_list) {
378 DB((dbg, LEVEL_2, "%s%+F", first ? "" : ", ", node->node));
381 if (! list_empty(&part->Follower)) {
382 DB((dbg, LEVEL_2, "\n---\n "));
384 list_for_each_entry(node_t, node, &part->Follower, node_list) {
385 DB((dbg, LEVEL_2, "%s%+F", first ? "" : ", ", node->node));
389 DB((dbg, LEVEL_2, "\n}\n"));
390 } /* dump_partition */
395 static void do_dump_list(const char *msg, const node_t *node, int ofs)
400 #define GET_LINK(p, ofs) *((const node_t **)((char *)(p) + (ofs)))
402 DB((dbg, LEVEL_3, "%s = {\n ", msg));
403 for (p = node; p != NULL; p = GET_LINK(p, ofs)) {
404 DB((dbg, LEVEL_3, "%s%+F", first ? "" : ", ", p->node));
407 DB((dbg, LEVEL_3, "\n}\n"));
415 static void dump_race_list(const char *msg, const node_t *list)
417 do_dump_list(msg, list, offsetof(node_t, race_next));
418 } /* dump_race_list */
421 * Dumps a local list.
423 static void dump_list(const char *msg, const node_t *list)
425 do_dump_list(msg, list, offsetof(node_t, next));
429 * Dump all partitions.
431 static void dump_all_partitions(const environment_t *env)
433 const partition_t *P;
435 DB((dbg, LEVEL_2, "All partitions\n===============\n"));
436 for (P = env->dbg_list; P != NULL; P = P->dbg_next)
437 dump_partition("", P);
438 } /* dump_all_partitions */
443 static void dump_split_list(const partition_t *list)
445 const partition_t *p;
447 DB((dbg, LEVEL_2, "Split by %s produced = {\n", what_reason));
448 for (p = list; p != NULL; p = p->split_next)
449 DB((dbg, LEVEL_2, "part%u, ", p->nr));
450 DB((dbg, LEVEL_2, "\n}\n"));
451 } /* dump_split_list */
454 * Dump partition and type for a node.
456 static int dump_partition_hook(FILE *F, const ir_node *n, const ir_node *local)
458 const ir_node *irn = local != NULL ? local : n;
459 node_t *node = get_irn_node(irn);
461 ir_fprintf(F, "info2 : \"partition %u type %+F\"\n", node->part->nr, node->type);
463 } /* dump_partition_hook */
466 #define dump_partition(msg, part)
467 #define dump_race_list(msg, list)
468 #define dump_list(msg, list)
469 #define dump_all_partitions(env)
470 #define dump_split_list(list)
473 #if defined(VERIFY_MONOTONE) && defined (DEBUG_libfirm)
475 * Verify that a type transition is monotone
477 static void verify_type(const lattice_elem_t old_type, node_t *node)
479 if (old_type.tv == node->type.tv) {
483 if (old_type.tv == tarval_top) {
484 /* from Top down-to is always allowed */
487 if (node->type.tv == tarval_bottom || node->type.tv == tarval_reachable) {
491 panic("combo: wrong translation from %+F to %+F on node %+F", old_type, node->type, node->node);
495 #define verify_type(old_type, node)
499 * Compare two pointer values of a listmap.
501 static int listmap_cmp_ptr(const void *elt, const void *key, size_t size)
503 const listmap_entry_t *e1 = (listmap_entry_t*)elt;
504 const listmap_entry_t *e2 = (listmap_entry_t*)key;
507 return e1->id != e2->id;
508 } /* listmap_cmp_ptr */
511 * Initializes a listmap.
513 * @param map the listmap
515 static void listmap_init(listmap_t *map)
517 map->map = new_set(listmap_cmp_ptr, 16);
522 * Terminates a listmap.
524 * @param map the listmap
526 static void listmap_term(listmap_t *map)
532 * Return the associated listmap entry for a given id.
534 * @param map the listmap
535 * @param id the id to search for
537 * @return the associated listmap entry for the given id
539 static listmap_entry_t *listmap_find(listmap_t *map, void *id)
541 listmap_entry_t key, *entry;
546 entry = (listmap_entry_t*)set_insert(map->map, &key, sizeof(key), hash_ptr(id));
548 if (entry->list == NULL) {
549 /* a new entry, put into the list */
550 entry->next = map->values;
557 * Calculate the hash value for an opcode map entry.
559 * @param entry an opcode map entry
561 * @return a hash value for the given opcode map entry
563 static unsigned opcode_hash(const opcode_key_t *entry)
565 /* we cannot use the ir ops hash function here, because it hashes the
567 const ir_node *n = entry->irn;
568 ir_opcode code = get_irn_opcode(n);
569 ir_mode *mode = get_irn_mode(n);
570 unsigned hash = (unsigned)(PTR_TO_INT(mode) * 9 + code) + get_irn_arity(n);
572 if (code == iro_Const)
573 hash ^= (unsigned)hash_ptr(get_Const_tarval(n));
574 else if (code == iro_Proj)
575 hash += (unsigned)get_Proj_proj(n);
580 * Compare two entries in the opcode map.
582 static int cmp_opcode(const void *elt, const void *key, size_t size)
584 const opcode_key_t *o1 = (opcode_key_t*)elt;
585 const opcode_key_t *o2 = (opcode_key_t*)key;
589 return cmp_irn_opcode(o1->irn, o2->irn);
593 * Compare two Def-Use edges for input position.
595 static int cmp_def_use_edge(const void *a, const void *b)
597 const ir_def_use_edge *ea = (const ir_def_use_edge*)a;
598 const ir_def_use_edge *eb = (const ir_def_use_edge*)b;
600 /* no overrun, because range is [-1, MAXINT] */
601 return ea->pos - eb->pos;
602 } /* cmp_def_use_edge */
605 * We need the Def-Use edges sorted.
607 static void sort_irn_outs(node_t *node)
609 ir_node *irn = node->node;
610 int n_outs = get_irn_n_outs(irn);
613 qsort(&irn->out[1], n_outs, sizeof(irn->out[0]), cmp_def_use_edge);
615 node->max_user_input = irn->out[n_outs].pos;
616 } /* sort_irn_outs */
619 * Return the type of a node.
621 * @param irn an IR-node
623 * @return the associated type of this node
625 static inline lattice_elem_t get_node_type(const ir_node *irn)
627 return get_irn_node(irn)->type;
628 } /* get_node_type */
631 * Return the tarval of a node.
633 * @param irn an IR-node
635 * @return the associated type of this node
637 static inline ir_tarval *get_node_tarval(const ir_node *irn)
639 lattice_elem_t type = get_node_type(irn);
641 if (is_tarval(type.tv))
643 return tarval_bottom;
644 } /* get_node_type */
647 * Add a partition to the worklist.
649 static inline void add_to_worklist(partition_t *X, environment_t *env)
651 assert(X->on_worklist == 0);
652 DB((dbg, LEVEL_2, "Adding part%d to worklist\n", X->nr));
653 X->wl_next = env->worklist;
656 } /* add_to_worklist */
659 * Create a new empty partition.
661 * @param env the environment
663 * @return a newly allocated partition
665 static inline partition_t *new_partition(environment_t *env)
667 partition_t *part = OALLOC(&env->obst, partition_t);
669 INIT_LIST_HEAD(&part->Leader);
670 INIT_LIST_HEAD(&part->Follower);
671 INIT_LIST_HEAD(&part->cprop);
672 INIT_LIST_HEAD(&part->cprop_X);
673 part->wl_next = NULL;
674 part->touched_next = NULL;
675 part->cprop_next = NULL;
676 part->split_next = NULL;
677 part->touched = NULL;
680 part->max_user_inputs = 0;
681 part->on_worklist = 0;
682 part->on_touched = 0;
684 part->type_is_T_or_C = 0;
686 part->dbg_next = env->dbg_list;
687 env->dbg_list = part;
688 part->nr = part_nr++;
692 } /* new_partition */
695 * Get the first node from a partition.
697 static inline node_t *get_first_node(const partition_t *X)
699 return list_entry(X->Leader.next, node_t, node_list);
700 } /* get_first_node */
703 * Return the type of a partition (assuming partition is non-empty and
704 * all elements have the same type).
706 * @param X a partition
708 * @return the type of the first element of the partition
710 static inline lattice_elem_t get_partition_type(const partition_t *X)
712 const node_t *first = get_first_node(X);
714 } /* get_partition_type */
717 * Creates a partition node for the given IR-node and place it
718 * into the given partition.
720 * @param irn an IR-node
721 * @param part a partition to place the node in
722 * @param env the environment
724 * @return the created node
726 static node_t *create_partition_node(ir_node *irn, partition_t *part, environment_t *env)
728 /* create a partition node and place it in the partition */
729 node_t *node = OALLOC(&env->obst, node_t);
731 INIT_LIST_HEAD(&node->node_list);
732 INIT_LIST_HEAD(&node->cprop_list);
736 node->race_next = NULL;
737 node->type.tv = tarval_top;
738 node->max_user_input = 0;
740 node->n_followers = 0;
741 node->on_touched = 0;
744 node->is_follower = 0;
746 set_irn_node(irn, node);
748 list_add_tail(&node->node_list, &part->Leader);
752 } /* create_partition_node */
755 * Pre-Walker, initialize all Nodes' type to U or top and place
756 * all nodes into the TOP partition.
758 static void create_initial_partitions(ir_node *irn, void *ctx)
760 environment_t *env = (environment_t*)ctx;
761 partition_t *part = env->initial;
764 node = create_partition_node(irn, part, env);
766 if (node->max_user_input > part->max_user_inputs)
767 part->max_user_inputs = node->max_user_input;
770 set_Block_phis(irn, NULL);
772 } /* create_initial_partitions */
775 * Post-Walker, collect all Block-Phi lists, set Cond.
777 static void init_block_phis(ir_node *irn, void *ctx)
782 ir_node *block = get_nodes_block(irn);
783 add_Block_phi(block, irn);
785 } /* init_block_phis */
788 * Add a node to the entry.partition.touched set and
789 * node->partition to the touched set if not already there.
792 * @param env the environment
794 static inline void add_to_touched(node_t *y, environment_t *env)
796 if (y->on_touched == 0) {
797 partition_t *part = y->part;
799 y->next = part->touched;
804 if (part->on_touched == 0) {
805 part->touched_next = env->touched;
807 part->on_touched = 1;
810 check_list(part->touched, part);
812 } /* add_to_touched */
815 * Place a node on the cprop list.
818 * @param env the environment
820 static void add_to_cprop(node_t *y, environment_t *env)
824 /* Add y to y.partition.cprop. */
825 if (y->on_cprop == 0) {
826 partition_t *Y = y->part;
827 ir_node *irn = y->node;
828 ir_node *skipped = skip_Proj(irn);
830 /* place Conds and all its Projs on the cprop_X list */
831 if (is_Cond(skipped) || is_Switch(skipped))
832 list_add_tail(&y->cprop_list, &Y->cprop_X);
834 list_add_tail(&y->cprop_list, &Y->cprop);
837 DB((dbg, LEVEL_3, "Add %+F to part%u.cprop\n", y->node, Y->nr));
839 /* place its partition on the cprop list */
840 if (Y->on_cprop == 0) {
841 Y->cprop_next = env->cprop;
847 if (get_irn_mode(irn) == mode_T) {
848 /* mode_T nodes always produce tarval_bottom, so we must explicitly
849 * add its Projs to get constant evaluation to work */
852 for (i = get_irn_n_outs(irn) - 1; i >= 0; --i) {
853 node_t *proj = get_irn_node(get_irn_out(irn, i));
855 add_to_cprop(proj, env);
857 } else if (is_Block(irn)) {
858 /* Due to the way we handle Phi's, we must place all Phis of a block on the list
859 * if someone placed the block. The Block is only placed if the reachability
860 * changes, and this must be re-evaluated in compute_Phi(). */
862 for (phi = get_Block_phis(irn); phi != NULL; phi = get_Phi_next(phi)) {
863 node_t *p = get_irn_node(phi);
864 add_to_cprop(p, env);
870 * Update the worklist: If Z is on worklist then add Z' to worklist.
871 * Else add the smaller of Z and Z' to worklist.
873 * @param Z the Z partition
874 * @param Z_prime the Z' partition, a previous part of Z
875 * @param env the environment
877 static void update_worklist(partition_t *Z, partition_t *Z_prime, environment_t *env)
879 if (Z->on_worklist || Z_prime->n_leader < Z->n_leader) {
880 add_to_worklist(Z_prime, env);
882 add_to_worklist(Z, env);
884 } /* update_worklist */
887 * Make all inputs to x no longer be F.def_use edges.
891 static void move_edges_to_leader(node_t *x)
893 ir_node *irn = x->node;
896 for (i = get_irn_arity(irn) - 1; i >= 0; --i) {
897 node_t *pred = get_irn_node(get_irn_n(irn, i));
902 n = get_irn_n_outs(p);
903 for (j = 1; j <= pred->n_followers; ++j) {
904 if (p->out[j].pos == i && p->out[j].use == irn) {
905 /* found a follower edge to x, move it to the Leader */
906 ir_def_use_edge edge = p->out[j];
908 /* remove this edge from the Follower set */
909 p->out[j] = p->out[pred->n_followers];
912 /* sort it into the leader set */
913 for (k = pred->n_followers + 2; k <= n; ++k) {
914 if (p->out[k].pos >= edge.pos)
916 p->out[k - 1] = p->out[k];
918 /* place the new edge here */
919 p->out[k - 1] = edge;
921 /* edge found and moved */
926 } /* move_edges_to_leader */
929 * Split a partition that has NO followers by a local list.
931 * @param Z partition to split
932 * @param g a (non-empty) node list
933 * @param env the environment
935 * @return a new partition containing the nodes of g
937 static partition_t *split_no_followers(partition_t *Z, node_t *g, environment_t *env)
939 partition_t *Z_prime;
944 dump_partition("Splitting ", Z);
945 dump_list("by list ", g);
949 /* Remove g from Z. */
950 for (node = g; node != NULL; node = node->next) {
951 assert(node->part == Z);
952 list_del(&node->node_list);
955 assert(n < Z->n_leader);
958 /* Move g to a new partition, Z'. */
959 Z_prime = new_partition(env);
961 for (node = g; node != NULL; node = node->next) {
962 list_add_tail(&node->node_list, &Z_prime->Leader);
963 node->part = Z_prime;
964 if (node->max_user_input > max_input)
965 max_input = node->max_user_input;
967 Z_prime->max_user_inputs = max_input;
968 Z_prime->n_leader = n;
971 check_partition(Z_prime);
973 /* for now, copy the type info tag, it will be adjusted in split_by(). */
974 Z_prime->type_is_T_or_C = Z->type_is_T_or_C;
976 update_worklist(Z, Z_prime, env);
978 dump_partition("Now ", Z);
979 dump_partition("Created new ", Z_prime);
981 } /* split_no_followers */
984 * Make the Follower -> Leader transition for a node.
988 static void follower_to_leader(node_t *n)
990 assert(n->is_follower == 1);
992 DB((dbg, LEVEL_2, "%+F make the follower -> leader transition\n", n->node));
994 move_edges_to_leader(n);
995 list_del(&n->node_list);
996 list_add_tail(&n->node_list, &n->part->Leader);
998 } /* follower_to_leader */
1001 * The environment for one race step.
1003 typedef struct step_env {
1004 node_t *initial; /**< The initial node list. */
1005 node_t *unwalked; /**< The unwalked node list. */
1006 node_t *walked; /**< The walked node list. */
1007 int index; /**< Next index of Follower use_def edge. */
1008 unsigned side; /**< side number. */
1012 * Return non-zero, if a input is a real follower
1014 * @param irn the node to check
1015 * @param input number of the input
1017 static int is_real_follower(const ir_node *irn, int input)
1021 switch (get_irn_opcode(irn)) {
1024 /* ignore the Confirm bound input */
1030 /* ignore the Mux sel input */
1035 /* dead inputs are not follower edges */
1036 ir_node *block = get_nodes_block(irn);
1037 node_t *pred = get_irn_node(get_Block_cfgpred(block, input));
1039 if (pred->type.tv == tarval_unreachable)
1049 /* only a Sub x,0 / Shift x,0 might be a follower */
1056 pred = get_irn_node(get_irn_n(irn, input));
1057 if (is_tarval(pred->type.tv) && tarval_is_null(pred->type.tv))
1061 pred = get_irn_node(get_irn_n(irn, input));
1062 if (is_tarval(pred->type.tv) && tarval_is_one(pred->type.tv))
1066 pred = get_irn_node(get_irn_n(irn, input));
1067 if (is_tarval(pred->type.tv) && tarval_is_all_one(pred->type.tv))
1071 assert(!"opcode not implemented yet");
1075 } /* is_real_follower */
1078 * Do one step in the race.
1080 static int step(step_env *env)
1084 if (env->initial != NULL) {
1085 /* Move node from initial to unwalked */
1087 env->initial = n->race_next;
1089 n->race_next = env->unwalked;
1095 while (env->unwalked != NULL) {
1096 /* let n be the first node in unwalked */
1098 while (env->index < n->n_followers) {
1099 const ir_def_use_edge *edge = &n->node->out[1 + env->index];
1101 /* let m be n.F.def_use[index] */
1102 node_t *m = get_irn_node(edge->use);
1104 assert(m->is_follower);
1106 * Some inputs, like the get_Confirm_bound are NOT
1107 * real followers, sort them out.
1109 if (! is_real_follower(m->node, edge->pos)) {
1115 /* only followers from our partition */
1116 if (m->part != n->part)
1119 if ((m->flagged & env->side) == 0) {
1120 m->flagged |= env->side;
1122 if (m->flagged != 3) {
1123 /* visited the first time */
1124 /* add m to unwalked not as first node (we might still need to
1125 check for more follower node */
1126 m->race_next = n->race_next;
1130 /* else already visited by the other side and on the other list */
1133 /* move n to walked */
1134 env->unwalked = n->race_next;
1135 n->race_next = env->walked;
1143 * Clear the flags from a list and check for
1144 * nodes that where touched from both sides.
1146 * @param list the list
1148 static int clear_flags(node_t *list)
1153 for (n = list; n != NULL; n = n->race_next) {
1154 if (n->flagged == 3) {
1155 /* we reach a follower from both sides, this will split congruent
1156 * inputs and make it a leader. */
1157 follower_to_leader(n);
1166 * Split a partition by a local list using the race.
1168 * @param pX pointer to the partition to split, might be changed!
1169 * @param gg a (non-empty) node list
1170 * @param env the environment
1172 * @return a new partition containing the nodes of gg
1174 static partition_t *split(partition_t **pX, node_t *gg, environment_t *env)
1176 partition_t *X = *pX;
1177 partition_t *X_prime;
1180 node_t *g, *h, *node, *t;
1181 int max_input, transitions, winner, shf;
1183 DEBUG_ONLY(static int run = 0;)
1185 DB((dbg, LEVEL_2, "Run %d ", run++));
1186 if (list_empty(&X->Follower)) {
1187 /* if the partition has NO follower, we can use the fast
1188 splitting algorithm. */
1189 return split_no_followers(X, gg, env);
1191 /* else do the race */
1193 dump_partition("Splitting ", X);
1194 dump_list("by list ", gg);
1196 INIT_LIST_HEAD(&tmp);
1198 /* Remove gg from X.Leader and put into g */
1200 for (node = gg; node != NULL; node = node->next) {
1201 assert(node->part == X);
1202 assert(node->is_follower == 0);
1204 list_del(&node->node_list);
1205 list_add_tail(&node->node_list, &tmp);
1206 node->race_next = g;
1211 list_for_each_entry(node_t, node, &X->Leader, node_list) {
1212 node->race_next = h;
1215 /* restore X.Leader */
1216 list_splice(&tmp, &X->Leader);
1218 senv[0].initial = g;
1219 senv[0].unwalked = NULL;
1220 senv[0].walked = NULL;
1224 senv[1].initial = h;
1225 senv[1].unwalked = NULL;
1226 senv[1].walked = NULL;
1231 * Some informations on the race that are not stated clearly in Click's
1233 * 1) A follower stays on the side that reach him first.
1234 * 2) If the other side reches a follower, if will be converted to
1235 * a leader. /This must be done after the race is over, else the
1236 * edges we are iterating on are renumbered./
1237 * 3) /New leader might end up on both sides./
1238 * 4) /If one side ends up with new Leaders, we must ensure that
1239 * they can split out by opcode, hence we have to put _every_
1240 * partition with new Leader nodes on the cprop list, as
1241 * opcode splitting is done by split_by() at the end of
1242 * constant propagation./
1245 if (step(&senv[0])) {
1249 if (step(&senv[1])) {
1254 assert(senv[winner].initial == NULL);
1255 assert(senv[winner].unwalked == NULL);
1257 /* clear flags from walked/unwalked */
1259 transitions = clear_flags(senv[0].unwalked) << shf;
1260 transitions |= clear_flags(senv[0].walked) << shf;
1262 transitions |= clear_flags(senv[1].unwalked) << shf;
1263 transitions |= clear_flags(senv[1].walked) << shf;
1265 dump_race_list("winner ", senv[winner].walked);
1267 /* Move walked_{winner} to a new partition, X'. */
1268 X_prime = new_partition(env);
1271 for (node = senv[winner].walked; node != NULL; node = node->race_next) {
1272 list_del(&node->node_list);
1273 node->part = X_prime;
1274 if (node->is_follower) {
1275 list_add_tail(&node->node_list, &X_prime->Follower);
1277 list_add_tail(&node->node_list, &X_prime->Leader);
1280 if (node->max_user_input > max_input)
1281 max_input = node->max_user_input;
1283 X_prime->n_leader = n;
1284 X_prime->max_user_inputs = max_input;
1285 X->n_leader -= X_prime->n_leader;
1287 /* for now, copy the type info tag, it will be adjusted in split_by(). */
1288 X_prime->type_is_T_or_C = X->type_is_T_or_C;
1291 * Even if a follower was not checked by both sides, it might have
1292 * loose its congruence, so we need to check this case for all follower.
1294 list_for_each_entry_safe(node_t, node, t, &X_prime->Follower, node_list) {
1295 if (identity(node) == node) {
1296 follower_to_leader(node);
1302 check_partition(X_prime);
1304 /* X' is the smaller part */
1305 add_to_worklist(X_prime, env);
1308 * If there where follower to leader transitions, ensure that the nodes
1309 * can be split out if necessary.
1311 if (transitions & 1) {
1312 /* place winner partition on the cprop list */
1313 if (X_prime->on_cprop == 0) {
1314 X_prime->cprop_next = env->cprop;
1315 env->cprop = X_prime;
1316 X_prime->on_cprop = 1;
1319 if (transitions & 2) {
1320 /* place other partition on the cprop list */
1321 if (X->on_cprop == 0) {
1322 X->cprop_next = env->cprop;
1328 dump_partition("Now ", X);
1329 dump_partition("Created new ", X_prime);
1331 /* we have to ensure that the partition containing g is returned */
1341 * Returns non-zero if the i'th input of a Phi node is live.
1343 * @param phi a Phi-node
1344 * @param i an input number
1346 * @return non-zero if the i'th input of the given Phi node is live
1348 static int is_live_input(ir_node *phi, int i)
1351 ir_node *block = get_nodes_block(phi);
1352 ir_node *pred = get_Block_cfgpred(block, i);
1353 lattice_elem_t type = get_node_type(pred);
1355 return type.tv != tarval_unreachable;
1357 /* else it's the control input, always live */
1359 } /* is_live_input */
1362 * Return non-zero if a type is a constant.
1364 static int is_constant_type(lattice_elem_t type)
1366 if (type.tv != tarval_bottom && type.tv != tarval_top)
1369 } /* is_constant_type */
1372 * Check whether a type is neither Top or a constant.
1373 * Note: U is handled like Top here, R is a constant.
1375 * @param type the type to check
1377 static int type_is_neither_top_nor_const(const lattice_elem_t type)
1379 if (is_tarval(type.tv)) {
1380 if (type.tv == tarval_top)
1382 if (tarval_is_constant(type.tv))
1389 } /* type_is_neither_top_nor_const */
1392 * Collect nodes to the touched list.
1394 * @param list the list which contains the nodes that must be evaluated
1395 * @param idx the index of the def_use edge to evaluate
1396 * @param env the environment
1398 static void collect_touched(list_head *list, int idx, environment_t *env)
1401 int end_idx = env->end_idx;
1403 list_for_each_entry(node_t, x, list, node_list) {
1407 /* leader edges start AFTER follower edges */
1408 x->next_edge = x->n_followers + 1;
1410 num_edges = get_irn_n_outs(x->node);
1412 /* for all edges in x.L.def_use_{idx} */
1413 while (x->next_edge <= num_edges) {
1414 const ir_def_use_edge *edge = &x->node->out[x->next_edge];
1417 /* check if we have necessary edges */
1418 if (edge->pos > idx)
1425 /* only non-commutative nodes */
1426 if (env->commutative &&
1427 (idx == 0 || idx == 1) && is_op_commutative(get_irn_op(succ)))
1430 /* ignore the "control input" for non-pinned nodes
1431 if we are running in GCSE mode */
1432 if (idx < end_idx && get_irn_pinned(succ) != op_pin_state_pinned)
1435 y = get_irn_node(succ);
1436 assert(get_irn_n(succ, idx) == x->node);
1438 /* ignore block edges touching followers */
1439 if (idx == -1 && y->is_follower)
1442 if (is_constant_type(y->type)) {
1443 unsigned code = get_irn_opcode(succ);
1444 if (code == iro_Sub || code == iro_Cmp)
1445 add_to_cprop(y, env);
1448 /* Partitions of constants should not be split simply because their Nodes have unequal
1449 functions or incongruent inputs. */
1450 if (type_is_neither_top_nor_const(y->type) &&
1451 (! is_Phi(y->node) || is_live_input(y->node, idx))) {
1452 add_to_touched(y, env);
1456 } /* collect_touched */
1459 * Collect commutative nodes to the touched list.
1461 * @param list the list which contains the nodes that must be evaluated
1462 * @param env the environment
1464 static void collect_commutative_touched(list_head *list, environment_t *env)
1468 list_for_each_entry(node_t, x, list, node_list) {
1471 num_edges = get_irn_n_outs(x->node);
1473 x->next_edge = x->n_followers + 1;
1475 /* for all edges in x.L.def_use_{idx} */
1476 while (x->next_edge <= num_edges) {
1477 const ir_def_use_edge *edge = &x->node->out[x->next_edge];
1480 /* check if we have necessary edges */
1490 /* only commutative nodes */
1491 if (!is_op_commutative(get_irn_op(succ)))
1494 y = get_irn_node(succ);
1495 if (is_constant_type(y->type)) {
1496 unsigned code = get_irn_opcode(succ);
1497 if (code == iro_Eor)
1498 add_to_cprop(y, env);
1501 /* Partitions of constants should not be split simply because their Nodes have unequal
1502 functions or incongruent inputs. */
1503 if (type_is_neither_top_nor_const(y->type)) {
1504 add_to_touched(y, env);
1508 } /* collect_commutative_touched */
1511 * Split the partitions if caused by the first entry on the worklist.
1513 * @param env the environment
1515 static void cause_splits(environment_t *env)
1517 partition_t *X, *Z, *N;
1520 /* remove the first partition from the worklist */
1522 env->worklist = X->wl_next;
1525 dump_partition("Cause_split: ", X);
1527 if (env->commutative) {
1528 /* handle commutative nodes first */
1530 /* empty the touched set: already done, just clear the list */
1531 env->touched = NULL;
1533 collect_commutative_touched(&X->Leader, env);
1534 collect_commutative_touched(&X->Follower, env);
1536 for (Z = env->touched; Z != NULL; Z = N) {
1538 node_t *touched = Z->touched;
1539 node_t *touched_aa = NULL;
1540 node_t *touched_ab = NULL;
1541 unsigned n_touched_aa = 0;
1542 unsigned n_touched_ab = 0;
1544 assert(Z->touched != NULL);
1546 /* beware, split might change Z */
1547 N = Z->touched_next;
1549 /* remove it from the touched set */
1552 /* Empty local Z.touched. */
1553 for (e = touched; e != NULL; e = n) {
1554 node_t *left = get_irn_node(get_irn_n(e->node, 0));
1555 node_t *right = get_irn_node(get_irn_n(e->node, 1));
1557 assert(e->is_follower == 0);
1562 * Note: op(a, a) is NOT congruent to op(a, b).
1563 * So, we must split the touched list.
1565 if (left->part == right->part) {
1566 e->next = touched_aa;
1570 e->next = touched_ab;
1575 assert(n_touched_aa + n_touched_ab == Z->n_touched);
1579 if (0 < n_touched_aa && n_touched_aa < Z->n_leader) {
1580 partition_t *Z_prime = Z;
1581 DB((dbg, LEVEL_2, "Split part%d by touched_aa\n", Z_prime->nr));
1582 split(&Z_prime, touched_aa, env);
1584 assert(n_touched_aa <= Z->n_leader);
1586 if (0 < n_touched_ab && n_touched_ab < Z->n_leader) {
1587 partition_t *Z_prime = Z;
1588 DB((dbg, LEVEL_2, "Split part%d by touched_ab\n", Z_prime->nr));
1589 split(&Z_prime, touched_ab, env);
1591 assert(n_touched_ab <= Z->n_leader);
1595 /* combine temporary leader and follower list */
1596 for (idx = -1; idx <= X->max_user_inputs; ++idx) {
1597 /* empty the touched set: already done, just clear the list */
1598 env->touched = NULL;
1600 collect_touched(&X->Leader, idx, env);
1601 collect_touched(&X->Follower, idx, env);
1603 for (Z = env->touched; Z != NULL; Z = N) {
1605 node_t *touched = Z->touched;
1606 unsigned n_touched = Z->n_touched;
1608 assert(Z->touched != NULL);
1610 /* beware, split might change Z */
1611 N = Z->touched_next;
1613 /* remove it from the touched set */
1616 /* Empty local Z.touched. */
1617 for (e = touched; e != NULL; e = e->next) {
1618 assert(e->is_follower == 0);
1624 if (0 < n_touched && n_touched < Z->n_leader) {
1625 DB((dbg, LEVEL_2, "Split part%d by touched\n", Z->nr));
1626 split(&Z, touched, env);
1628 assert(n_touched <= Z->n_leader);
1631 } /* cause_splits */
1634 * Implements split_by_what(): Split a partition by characteristics given
1635 * by the what function.
1637 * @param X the partition to split
1638 * @param What a function returning an Id for every node of the partition X
1639 * @param P a list to store the result partitions
1640 * @param env the environment
1644 static partition_t *split_by_what(partition_t *X, what_func What,
1645 partition_t **P, environment_t *env)
1649 listmap_entry_t *iter;
1652 /* Let map be an empty mapping from the range of What to (local) list of Nodes. */
1654 list_for_each_entry(node_t, x, &X->Leader, node_list) {
1655 void *id = What(x, env);
1656 listmap_entry_t *entry;
1659 /* input not allowed, ignore */
1662 /* Add x to map[What(x)]. */
1663 entry = listmap_find(&map, id);
1664 x->next = entry->list;
1667 /* Let P be a set of Partitions. */
1669 /* for all sets S except one in the range of map do */
1670 for (iter = map.values; iter != NULL; iter = iter->next) {
1671 if (iter->next == NULL) {
1672 /* this is the last entry, ignore */
1677 /* Add SPLIT( X, S ) to P. */
1678 DB((dbg, LEVEL_2, "Split part%d by WHAT = %s\n", X->nr, what_reason));
1679 R = split(&X, S, env);
1689 } /* split_by_what */
1691 /** lambda n.(n.type) */
1692 static void *lambda_type(const node_t *node, environment_t *env)
1695 return node->type.tv;
1698 /** lambda n.(n.opcode) */
1699 static void *lambda_opcode(const node_t *node, environment_t *env)
1701 opcode_key_t key, *entry;
1703 key.irn = node->node;
1705 entry = (opcode_key_t*)set_insert(env->opcode2id_map, &key, sizeof(key), opcode_hash(&key));
1707 } /* lambda_opcode */
1709 /** lambda n.(n[i].partition) */
1710 static void *lambda_partition(const node_t *node, environment_t *env)
1712 ir_node *skipped = skip_Proj(node->node);
1715 int i = env->lambda_input;
1717 if (i >= get_irn_arity(node->node)) {
1719 * We are outside the allowed range: This can happen even
1720 * if we have split by opcode first: doing so might move Followers
1721 * to Leaders and those will have a different opcode!
1722 * Note that in this case the partition is on the cprop list and will be
1728 /* ignore the "control input" for non-pinned nodes
1729 if we are running in GCSE mode */
1730 if (i < env->end_idx && get_irn_pinned(skipped) != op_pin_state_pinned)
1733 pred = i == -1 ? get_irn_n(skipped, i) : get_irn_n(node->node, i);
1734 p = get_irn_node(pred);
1736 } /* lambda_partition */
1738 /** lambda n.(n[i].partition) for commutative nodes */
1739 static void *lambda_commutative_partition(const node_t *node, environment_t *env)
1741 ir_node *irn = node->node;
1742 ir_node *skipped = skip_Proj(irn);
1743 ir_node *pred, *left, *right;
1745 partition_t *pl, *pr;
1746 int i = env->lambda_input;
1748 if (i >= get_irn_arity(node->node)) {
1750 * We are outside the allowed range: This can happen even
1751 * if we have split by opcode first: doing so might move Followers
1752 * to Leaders and those will have a different opcode!
1753 * Note that in this case the partition is on the cprop list and will be
1759 /* ignore the "control input" for non-pinned nodes
1760 if we are running in GCSE mode */
1761 if (i < env->end_idx && get_irn_pinned(skipped) != op_pin_state_pinned)
1765 pred = get_irn_n(skipped, i);
1766 p = get_irn_node(pred);
1770 if (is_op_commutative(get_irn_op(irn))) {
1771 /* normalize partition order by returning the "smaller" on input 0,
1772 the "bigger" on input 1. */
1773 left = get_binop_left(irn);
1774 pl = get_irn_node(left)->part;
1775 right = get_binop_right(irn);
1776 pr = get_irn_node(right)->part;
1779 return pl < pr ? pl : pr;
1781 return pl > pr ? pl : pr;
1783 /* a not split out Follower */
1784 pred = get_irn_n(irn, i);
1785 p = get_irn_node(pred);
1789 } /* lambda_commutative_partition */
1792 * Returns true if a type is a constant (and NOT Top
1795 static int is_con(const lattice_elem_t type)
1797 /* be conservative */
1798 if (is_tarval(type.tv))
1799 return tarval_is_constant(type.tv);
1800 return is_entity(type.sym.entity_p);
1804 * Implements split_by().
1806 * @param X the partition to split
1807 * @param env the environment
1809 static void split_by(partition_t *X, environment_t *env)
1811 partition_t *I, *P = NULL;
1814 dump_partition("split_by", X);
1816 if (X->n_leader == 1) {
1817 /* we have only one leader, no need to split, just check its type */
1818 node_t *x = get_first_node(X);
1819 X->type_is_T_or_C = x->type.tv == tarval_top || is_con(x->type);
1823 DEBUG_ONLY(what_reason = "lambda n.(n.type)";)
1824 P = split_by_what(X, lambda_type, &P, env);
1827 /* adjust the type tags, we have split partitions by type */
1828 for (I = P; I != NULL; I = I->split_next) {
1829 node_t *x = get_first_node(I);
1830 I->type_is_T_or_C = x->type.tv == tarval_top || is_con(x->type);
1837 if (Y->n_leader > 1) {
1838 /* we do not want split the TOP or constant partitions */
1839 if (! Y->type_is_T_or_C) {
1840 partition_t *Q = NULL;
1842 DEBUG_ONLY(what_reason = "lambda n.(n.opcode)";)
1843 Q = split_by_what(Y, lambda_opcode, &Q, env);
1850 if (Z->n_leader > 1) {
1851 const node_t *first = get_first_node(Z);
1852 int arity = get_irn_arity(first->node);
1854 what_func what = lambda_partition;
1855 DEBUG_ONLY(char buf[64];)
1857 if (env->commutative && is_op_commutative(get_irn_op(first->node)))
1858 what = lambda_commutative_partition;
1861 * BEWARE: during splitting by input 2 for instance we might
1862 * create new partitions which are different by input 1, so collect
1863 * them and split further.
1865 Z->split_next = NULL;
1868 for (input = arity - 1; input >= -1; --input) {
1870 partition_t *Z_prime = R;
1873 if (Z_prime->n_leader > 1) {
1874 env->lambda_input = input;
1875 DEBUG_ONLY(snprintf(buf, sizeof(buf), "lambda n.(n[%d].partition)", input);)
1876 DEBUG_ONLY(what_reason = buf;)
1877 S = split_by_what(Z_prime, what, &S, env);
1880 Z_prime->split_next = S;
1883 } while (R != NULL);
1888 } while (Q != NULL);
1891 } while (P != NULL);
1895 * (Re-)compute the type for a given node.
1897 * @param node the node
1899 static void default_compute(node_t *node)
1902 ir_node *irn = node->node;
1904 /* if any of the data inputs have type top, the result is type top */
1905 for (i = get_irn_arity(irn) - 1; i >= 0; --i) {
1906 ir_node *pred = get_irn_n(irn, i);
1907 node_t *p = get_irn_node(pred);
1909 if (p->type.tv == tarval_top) {
1910 node->type.tv = tarval_top;
1915 if (get_irn_mode(node->node) == mode_X)
1916 node->type.tv = tarval_reachable;
1918 node->type.tv = computed_value(irn);
1919 } /* default_compute */
1922 * (Re-)compute the type for a Block node.
1924 * @param node the node
1926 static void compute_Block(node_t *node)
1929 ir_node *block = node->node;
1931 if (block == get_irg_start_block(current_ir_graph) || get_Block_entity(block) != NULL) {
1932 /* start block and labelled blocks are always reachable */
1933 node->type.tv = tarval_reachable;
1937 for (i = get_Block_n_cfgpreds(block) - 1; i >= 0; --i) {
1938 node_t *pred = get_irn_node(get_Block_cfgpred(block, i));
1940 if (pred->type.tv == tarval_reachable) {
1941 /* A block is reachable, if at least of predecessor is reachable. */
1942 node->type.tv = tarval_reachable;
1946 node->type.tv = tarval_top;
1947 } /* compute_Block */
1950 * (Re-)compute the type for a Bad node.
1952 * @param node the node
1954 static void compute_Bad(node_t *node)
1956 /* Bad nodes ALWAYS compute Top */
1957 node->type.tv = tarval_top;
1961 * (Re-)compute the type for an Unknown node.
1963 * @param node the node
1965 static void compute_Unknown(node_t *node)
1967 /* While Unknown nodes should compute Top this is dangerous:
1968 * a Top input to a Cond would lead to BOTH control flows unreachable.
1969 * While this is correct in the given semantics, it would destroy the Firm
1972 * It would be safe to compute Top IF it can be assured, that only Cmp
1973 * nodes are inputs to Conds. We check that first.
1974 * This is the way Frontends typically build Firm, but some optimizations
1975 * (jump threading for instance) might replace them by Phib's...
1977 node->type.tv = tarval_UNKNOWN;
1978 } /* compute_Unknown */
1981 * (Re-)compute the type for a Jmp node.
1983 * @param node the node
1985 static void compute_Jmp(node_t *node)
1987 node_t *block = get_irn_node(get_nodes_block(node->node));
1989 node->type = block->type;
1993 * (Re-)compute the type for the Return node.
1995 * @param node the node
1997 static void compute_Return(node_t *node)
1999 /* The Return node is NOT dead if it is in a reachable block.
2000 * This is already checked in compute(). so we can return
2001 * Reachable here. */
2002 node->type.tv = tarval_reachable;
2003 } /* compute_Return */
2006 * (Re-)compute the type for the End node.
2008 * @param node the node
2010 static void compute_End(node_t *node)
2012 /* the End node is NOT dead of course */
2013 node->type.tv = tarval_reachable;
2017 * (Re-)compute the type for a Call.
2019 * @param node the node
2021 static void compute_Call(node_t *node)
2024 * A Call computes always bottom, even if it has Unknown
2027 node->type.tv = tarval_bottom;
2028 } /* compute_Call */
2031 * (Re-)compute the type for a SymConst node.
2033 * @param node the node
2035 static void compute_SymConst(node_t *node)
2037 ir_node *irn = node->node;
2038 node_t *block = get_irn_node(get_nodes_block(irn));
2040 if (block->type.tv == tarval_unreachable) {
2041 node->type.tv = tarval_top;
2044 switch (get_SymConst_kind(irn)) {
2045 case symconst_addr_ent:
2046 node->type.sym = get_SymConst_symbol(irn);
2049 node->type.tv = computed_value(irn);
2051 } /* compute_SymConst */
2054 * (Re-)compute the type for a Phi node.
2056 * @param node the node
2058 static void compute_Phi(node_t *node)
2061 ir_node *phi = node->node;
2062 lattice_elem_t type;
2064 /* if a Phi is in a unreachable block, its type is TOP */
2065 node_t *block = get_irn_node(get_nodes_block(phi));
2067 if (block->type.tv == tarval_unreachable) {
2068 node->type.tv = tarval_top;
2072 /* Phi implements the Meet operation */
2073 type.tv = tarval_top;
2074 for (i = get_Phi_n_preds(phi) - 1; i >= 0; --i) {
2075 node_t *pred = get_irn_node(get_Phi_pred(phi, i));
2076 node_t *pred_X = get_irn_node(get_Block_cfgpred(block->node, i));
2078 if (pred_X->type.tv == tarval_unreachable || pred->type.tv == tarval_top) {
2079 /* ignore TOP inputs: We must check here for unreachable blocks,
2080 because Firm constants live in the Start Block are NEVER Top.
2081 Else, a Phi (1,2) will produce Bottom, even if the 2 for instance
2082 comes from a unreachable input. */
2085 if (pred->type.tv == tarval_bottom) {
2086 node->type.tv = tarval_bottom;
2088 } else if (type.tv == tarval_top) {
2089 /* first constant found */
2091 } else if (type.tv != pred->type.tv) {
2092 /* different constants or tarval_bottom */
2093 node->type.tv = tarval_bottom;
2096 /* else nothing, constants are the same */
2102 * (Re-)compute the type for an Add. Special case: one nodes is a Zero Const.
2104 * @param node the node
2106 static void compute_Add(node_t *node)
2108 ir_node *sub = node->node;
2109 node_t *l = get_irn_node(get_Add_left(sub));
2110 node_t *r = get_irn_node(get_Add_right(sub));
2111 lattice_elem_t a = l->type;
2112 lattice_elem_t b = r->type;
2115 if (a.tv == tarval_top || b.tv == tarval_top) {
2116 node->type.tv = tarval_top;
2117 } else if (a.tv == tarval_bottom || b.tv == tarval_bottom) {
2118 node->type.tv = tarval_bottom;
2120 /* x + 0 = 0 + x = x, but beware of floating point +0 + -0, so we
2121 must call tarval_add() first to handle this case! */
2122 if (is_tarval(a.tv)) {
2123 if (is_tarval(b.tv)) {
2124 node->type.tv = tarval_add(a.tv, b.tv);
2127 mode = get_tarval_mode(a.tv);
2128 if (a.tv == get_mode_null(mode)) {
2132 } else if (is_tarval(b.tv)) {
2133 mode = get_tarval_mode(b.tv);
2134 if (b.tv == get_mode_null(mode)) {
2139 node->type.tv = tarval_bottom;
2144 * (Re-)compute the type for a Sub. Special case: both nodes are congruent.
2146 * @param node the node
2148 static void compute_Sub(node_t *node)
2150 ir_node *sub = node->node;
2151 node_t *l = get_irn_node(get_Sub_left(sub));
2152 node_t *r = get_irn_node(get_Sub_right(sub));
2153 lattice_elem_t a = l->type;
2154 lattice_elem_t b = r->type;
2157 if (a.tv == tarval_top || b.tv == tarval_top) {
2158 node->type.tv = tarval_top;
2159 } else if (is_con(a) && is_con(b)) {
2160 if (is_tarval(a.tv) && is_tarval(b.tv)) {
2161 node->type.tv = tarval_sub(a.tv, b.tv, get_irn_mode(sub));
2162 } else if (is_tarval(a.tv) && tarval_is_null(a.tv)) {
2164 } else if (is_tarval(b.tv) && tarval_is_null(b.tv)) {
2167 node->type.tv = tarval_bottom;
2169 } else if (r->part == l->part &&
2170 (!mode_is_float(get_irn_mode(l->node)))) {
2172 * BEWARE: a - a is NOT always 0 for floating Point values, as
2173 * NaN op NaN = NaN, so we must check this here.
2175 ir_mode *mode = get_irn_mode(sub);
2176 tv = get_mode_null(mode);
2178 /* if the node was ONCE evaluated by all constants, but now
2179 this breaks AND we get from the argument partitions a different
2180 result, switch to bottom.
2181 This happens because initially all nodes are in the same partition ... */
2182 if (node->type.tv != tv)
2186 node->type.tv = tarval_bottom;
2191 * (Re-)compute the type for an Eor. Special case: both nodes are congruent.
2193 * @param node the node
2195 static void compute_Eor(node_t *node)
2197 ir_node *eor = node->node;
2198 node_t *l = get_irn_node(get_Eor_left(eor));
2199 node_t *r = get_irn_node(get_Eor_right(eor));
2200 lattice_elem_t a = l->type;
2201 lattice_elem_t b = r->type;
2204 if (a.tv == tarval_top || b.tv == tarval_top) {
2205 node->type.tv = tarval_top;
2206 } else if (is_con(a) && is_con(b)) {
2207 if (is_tarval(a.tv) && is_tarval(b.tv)) {
2208 node->type.tv = tarval_eor(a.tv, b.tv);
2209 } else if (is_tarval(a.tv) && tarval_is_null(a.tv)) {
2211 } else if (is_tarval(b.tv) && tarval_is_null(b.tv)) {
2214 node->type.tv = tarval_bottom;
2216 } else if (r->part == l->part) {
2217 ir_mode *mode = get_irn_mode(eor);
2218 tv = get_mode_null(mode);
2220 /* if the node was ONCE evaluated by all constants, but now
2221 this breaks AND we get from the argument partitions a different
2222 result, switch to bottom.
2223 This happens because initially all nodes are in the same partition ... */
2224 if (node->type.tv != tv)
2228 node->type.tv = tarval_bottom;
2233 * (Re-)compute the type for Cmp.
2235 * @param node the node
2237 static void compute_Cmp(node_t *node)
2239 ir_node *cmp = node->node;
2240 node_t *l = get_irn_node(get_Cmp_left(cmp));
2241 node_t *r = get_irn_node(get_Cmp_right(cmp));
2242 lattice_elem_t a = l->type;
2243 lattice_elem_t b = r->type;
2244 ir_relation relation = get_Cmp_relation(cmp);
2247 if (a.tv == tarval_top || b.tv == tarval_top) {
2248 node->type.tv = tarval_undefined;
2249 } else if (is_con(a) && is_con(b)) {
2250 default_compute(node);
2253 * BEWARE: a == a is NOT always True for floating Point values, as
2254 * NaN != NaN is defined, so we must check this here.
2255 * (while for some pnc we could still optimize we have to stay
2256 * consistent with compute_Cmp, so don't do anything for floats)
2258 } else if (r->part == l->part && !mode_is_float(get_irn_mode(l->node))) {
2259 tv = relation & ir_relation_equal ? tarval_b_true : tarval_b_false;
2261 /* if the node was ONCE evaluated to a constant, but now
2262 this breaks AND we get from the argument partitions a different
2263 result, ensure monotony by fall to bottom.
2264 This happens because initially all nodes are in the same partition ... */
2265 if (node->type.tv == tarval_bottom)
2267 else if (node->type.tv != tv && is_constant_type(node->type))
2271 node->type.tv = tarval_bottom;
2276 * (Re-)compute the type for a Proj(Cond).
2278 * @param node the node
2279 * @param cond the predecessor Cond node
2281 static void compute_Proj_Cond(node_t *node, ir_node *cond)
2283 ir_node *proj = node->node;
2284 long pnc = get_Proj_proj(proj);
2285 ir_node *sel = get_Cond_selector(cond);
2286 node_t *selector = get_irn_node(sel);
2289 * Note: it is crucial for the monotony that the Proj(Cond)
2290 * are evaluates after all predecessors of the Cond selector are
2296 * Due to the fact that 0 is a const, the Cmp gets immediately
2297 * on the cprop list. It will be evaluated before x is evaluated,
2298 * might leaving x as Top. When later x is evaluated, the Cmp
2299 * might change its value.
2300 * BUT if the Cond is evaluated before this happens, Proj(Cond, FALSE)
2301 * gets R, and later changed to F if Cmp is evaluated to True!
2303 * We prevent this by putting Conds in an extra cprop_X queue, which
2304 * gets evaluated after the cprop queue is empty.
2306 * Note that this even happens with Click's original algorithm, if
2307 * Cmp(x, 0) is evaluated to True first and later changed to False
2308 * if x was Top first and later changed to a Const ...
2309 * It is unclear how Click solved that problem ...
2311 * However, in rare cases even this does not help, if a Top reaches
2312 * a compare through a Phi, than Proj(Cond) is evaluated changing
2313 * the type of the Phi to something other.
2314 * So, we take the last resort and bind the type to R once
2317 * (This might be even the way Click works around the whole problem).
2319 * Finally, we may miss some optimization possibilities due to this:
2324 * If Top reaches the if first, than we decide for != here.
2325 * If y later is evaluated to 0, we cannot revert this decision
2326 * and must live with both outputs enabled. If this happens,
2327 * we get an unresolved if (true) in the code ...
2329 * In Click's version where this decision is done at the Cmp,
2330 * the Cmp is NOT optimized away than (if y evaluated to 1
2331 * for instance) and we get a if (1 == 0) here ...
2333 * Both solutions are suboptimal.
2334 * At least, we could easily detect this problem and run
2335 * cf_opt() (or even combo) again :-(
2337 if (node->type.tv == tarval_reachable)
2340 if (pnc == pn_Cond_true) {
2341 if (selector->type.tv == tarval_b_false) {
2342 node->type.tv = tarval_unreachable;
2343 } else if (selector->type.tv == tarval_b_true) {
2344 node->type.tv = tarval_reachable;
2345 } else if (selector->type.tv == tarval_bottom) {
2346 node->type.tv = tarval_reachable;
2348 assert(selector->type.tv == tarval_top);
2349 if (tarval_UNKNOWN == tarval_top) {
2350 /* any condition based on Top is "!=" */
2351 node->type.tv = tarval_unreachable;
2353 node->type.tv = tarval_unreachable;
2357 assert(pnc == pn_Cond_false);
2359 if (selector->type.tv == tarval_b_false) {
2360 node->type.tv = tarval_reachable;
2361 } else if (selector->type.tv == tarval_b_true) {
2362 node->type.tv = tarval_unreachable;
2363 } else if (selector->type.tv == tarval_bottom) {
2364 node->type.tv = tarval_reachable;
2366 assert(selector->type.tv == tarval_top);
2367 if (tarval_UNKNOWN == tarval_top) {
2368 /* any condition based on Top is "!=" */
2369 node->type.tv = tarval_reachable;
2371 node->type.tv = tarval_unreachable;
2375 } /* compute_Proj_Cond */
2377 static void compute_Proj_Switch(node_t *node, ir_node *switchn)
2379 ir_node *proj = node->node;
2380 long pnc = get_Proj_proj(proj);
2381 ir_node *sel = get_Switch_selector(switchn);
2382 node_t *selector = get_irn_node(sel);
2384 /* see long comment in compute_Proj_Cond */
2385 if (node->type.tv == tarval_reachable)
2388 if (selector->type.tv == tarval_bottom) {
2389 node->type.tv = tarval_reachable;
2390 } else if (selector->type.tv == tarval_top) {
2391 if (tarval_UNKNOWN == tarval_top && pnc == pn_Switch_default) {
2392 /* a switch based of Top is always "default" */
2393 node->type.tv = tarval_reachable;
2395 node->type.tv = tarval_unreachable;
2398 long value = get_tarval_long(selector->type.tv);
2399 const ir_switch_table *table = get_Switch_table(switchn);
2400 size_t n_entries = ir_switch_table_get_n_entries(table);
2403 for (e = 0; e < n_entries; ++e) {
2404 const ir_switch_table_entry *entry
2405 = ir_switch_table_get_entry_const(table, e);
2406 ir_tarval *min = entry->min;
2407 ir_tarval *max = entry->max;
2409 if (selector->type.tv == min) {
2410 node->type.tv = entry->pn == pnc
2411 ? tarval_reachable : tarval_unreachable;
2415 long minval = get_tarval_long(min);
2416 long maxval = get_tarval_long(max);
2417 if (minval <= value && value <= maxval) {
2418 node->type.tv = entry->pn == pnc
2419 ? tarval_reachable : tarval_unreachable;
2425 /* no entry matched: default */
2427 = pnc == pn_Switch_default ? tarval_reachable : tarval_unreachable;
2432 * (Re-)compute the type for a Proj-Node.
2434 * @param node the node
2436 static void compute_Proj(node_t *node)
2438 ir_node *proj = node->node;
2439 ir_mode *mode = get_irn_mode(proj);
2440 node_t *block = get_irn_node(get_nodes_block(skip_Proj(proj)));
2441 ir_node *pred = get_Proj_pred(proj);
2443 if (block->type.tv == tarval_unreachable) {
2444 /* a Proj in a unreachable Block stay Top */
2445 node->type.tv = tarval_top;
2448 if (get_irn_node(pred)->type.tv == tarval_top && !is_Cond(pred) && !is_Switch(pred)) {
2449 /* if the predecessor is Top, its Proj follow */
2450 node->type.tv = tarval_top;
2454 if (mode == mode_M) {
2455 /* mode M is always bottom */
2456 node->type.tv = tarval_bottom;
2458 } else if (mode == mode_X) {
2459 /* handle mode_X nodes */
2460 switch (get_irn_opcode(pred)) {
2462 /* the Proj_X from the Start is always reachable.
2463 However this is already handled at the top. */
2464 node->type.tv = tarval_reachable;
2467 compute_Proj_Cond(node, pred);
2470 compute_Proj_Switch(node, pred);
2477 default_compute(node);
2478 } /* compute_Proj */
2481 * (Re-)compute the type for a Confirm.
2483 * @param node the node
2485 static void compute_Confirm(node_t *node)
2487 ir_node *confirm = node->node;
2488 node_t *pred = get_irn_node(get_Confirm_value(confirm));
2490 if (get_Confirm_relation(confirm) == ir_relation_equal) {
2491 node_t *bound = get_irn_node(get_Confirm_bound(confirm));
2493 if (is_con(bound->type)) {
2494 /* is equal to a constant */
2495 node->type = bound->type;
2499 /* a Confirm is a copy OR a Const */
2500 node->type = pred->type;
2501 } /* compute_Confirm */
2504 * (Re-)compute the type for a given node.
2506 * @param node the node
2508 static void compute(node_t *node)
2510 ir_node *irn = node->node;
2513 #ifndef VERIFY_MONOTONE
2515 * Once a node reaches bottom, the type cannot fall further
2516 * in the lattice and we can stop computation.
2517 * Do not take this exit if the monotony verifier is
2518 * enabled to catch errors.
2520 if (node->type.tv == tarval_bottom)
2524 if (!is_Block(irn)) {
2525 /* for pinned nodes, check its control input */
2526 if (get_irn_pinned(skip_Proj(irn)) == op_pin_state_pinned) {
2527 node_t *block = get_irn_node(get_nodes_block(irn));
2529 if (block->type.tv == tarval_unreachable) {
2530 node->type.tv = tarval_top;
2536 func = (compute_func)node->node->op->ops.generic;
2542 * Identity functions: Note that one might think that identity() is just a
2543 * synonym for equivalent_node(). While this is true, we cannot use it for the algorithm
2544 * here, because it expects that the identity node is one of the inputs, which is NOT
2545 * always true for equivalent_node() which can handle (and does sometimes) DAGs.
2546 * So, we have our own implementation, which copies some parts of equivalent_node()
2550 * Calculates the Identity for Phi nodes
2552 static node_t *identity_Phi(node_t *node)
2554 ir_node *phi = node->node;
2555 ir_node *block = get_nodes_block(phi);
2556 node_t *n_part = NULL;
2559 for (i = get_Phi_n_preds(phi) - 1; i >= 0; --i) {
2560 node_t *pred_X = get_irn_node(get_Block_cfgpred(block, i));
2562 if (pred_X->type.tv == tarval_reachable) {
2563 node_t *pred = get_irn_node(get_Phi_pred(phi, i));
2567 else if (n_part->part != pred->part) {
2568 /* incongruent inputs, not a follower */
2573 /* if n_part is NULL here, all inputs path are dead, the Phi computes
2574 * tarval_top, is in the TOP partition and should NOT being split! */
2575 assert(n_part != NULL);
2577 } /* identity_Phi */
2580 * Calculates the Identity for commutative 0 neutral nodes.
2582 static node_t *identity_comm_zero_binop(node_t *node)
2584 ir_node *op = node->node;
2585 node_t *a = get_irn_node(get_binop_left(op));
2586 node_t *b = get_irn_node(get_binop_right(op));
2587 ir_mode *mode = get_irn_mode(op);
2590 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
2591 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
2594 /* node: no input should be tarval_top, else the binop would be also
2595 * Top and not being split. */
2596 zero = get_mode_null(mode);
2597 if (a->type.tv == zero)
2599 if (b->type.tv == zero)
2602 } /* identity_comm_zero_binop */
2605 * Calculates the Identity for Shift nodes.
2607 static node_t *identity_shift(node_t *node)
2609 ir_node *op = node->node;
2610 node_t *b = get_irn_node(get_binop_right(op));
2611 ir_mode *mode = get_irn_mode(b->node);
2614 /* node: no input should be tarval_top, else the binop would be also
2615 * Top and not being split. */
2616 zero = get_mode_null(mode);
2617 if (b->type.tv == zero)
2618 return get_irn_node(get_binop_left(op));
2620 } /* identity_shift */
2623 * Calculates the Identity for Mul nodes.
2625 static node_t *identity_Mul(node_t *node)
2627 ir_node *op = node->node;
2628 node_t *a = get_irn_node(get_Mul_left(op));
2629 node_t *b = get_irn_node(get_Mul_right(op));
2630 ir_mode *mode = get_irn_mode(op);
2633 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
2634 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
2637 /* node: no input should be tarval_top, else the binop would be also
2638 * Top and not being split. */
2639 one = get_mode_one(mode);
2640 if (a->type.tv == one)
2642 if (b->type.tv == one)
2645 } /* identity_Mul */
2648 * Calculates the Identity for Sub nodes.
2650 static node_t *identity_Sub(node_t *node)
2652 ir_node *sub = node->node;
2653 node_t *b = get_irn_node(get_Sub_right(sub));
2654 ir_mode *mode = get_irn_mode(sub);
2656 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
2657 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
2660 /* node: no input should be tarval_top, else the binop would be also
2661 * Top and not being split. */
2662 if (b->type.tv == get_mode_null(mode))
2663 return get_irn_node(get_Sub_left(sub));
2665 } /* identity_Sub */
2668 * Calculates the Identity for And nodes.
2670 static node_t *identity_And(node_t *node)
2672 ir_node *andnode = node->node;
2673 node_t *a = get_irn_node(get_And_left(andnode));
2674 node_t *b = get_irn_node(get_And_right(andnode));
2675 ir_tarval *neutral = get_mode_all_one(get_irn_mode(andnode));
2677 /* node: no input should be tarval_top, else the And would be also
2678 * Top and not being split. */
2679 if (a->type.tv == neutral)
2681 if (b->type.tv == neutral)
2684 } /* identity_And */
2687 * Calculates the Identity for Confirm nodes.
2689 static node_t *identity_Confirm(node_t *node)
2691 ir_node *confirm = node->node;
2693 /* a Confirm is always a Copy */
2694 return get_irn_node(get_Confirm_value(confirm));
2695 } /* identity_Confirm */
2698 * Calculates the Identity for Mux nodes.
2700 static node_t *identity_Mux(node_t *node)
2702 ir_node *mux = node->node;
2703 node_t *t = get_irn_node(get_Mux_true(mux));
2704 node_t *f = get_irn_node(get_Mux_false(mux));
2707 if (t->part == f->part)
2710 /* for now, the 1-input identity is not supported */
2712 sel = get_irn_node(get_Mux_sel(mux));
2714 /* Mux sel input is mode_b, so it is always a tarval */
2715 if (sel->type.tv == tarval_b_true)
2717 if (sel->type.tv == tarval_b_false)
2721 } /* identity_Mux */
2724 * Calculates the Identity for nodes.
2726 static node_t *identity(node_t *node)
2728 ir_node *irn = node->node;
2730 switch (get_irn_opcode(irn)) {
2732 return identity_Phi(node);
2734 return identity_Mul(node);
2738 return identity_comm_zero_binop(node);
2743 return identity_shift(node);
2745 return identity_And(node);
2747 return identity_Sub(node);
2749 return identity_Confirm(node);
2751 return identity_Mux(node);
2758 * Node follower is a (new) follower of leader, segregate Leader
2761 static void segregate_def_use_chain_1(const ir_node *follower, node_t *leader)
2763 ir_node *l = leader->node;
2764 int j, i, n = get_irn_n_outs(l);
2766 DB((dbg, LEVEL_2, "%+F is a follower of %+F\n", follower, leader->node));
2767 /* The leader edges must remain sorted, but follower edges can
2769 for (i = leader->n_followers + 1; i <= n; ++i) {
2770 if (l->out[i].use == follower) {
2771 ir_def_use_edge t = l->out[i];
2773 for (j = i - 1; j >= leader->n_followers + 1; --j)
2774 l->out[j + 1] = l->out[j];
2775 ++leader->n_followers;
2776 l->out[leader->n_followers] = t;
2780 } /* segregate_def_use_chain_1 */
2783 * Node follower is a (new) follower segregate its Leader
2786 * @param follower the follower IR node
2788 static void segregate_def_use_chain(const ir_node *follower)
2792 for (i = get_irn_arity(follower) - 1; i >= 0; --i) {
2793 node_t *pred = get_irn_node(get_irn_n(follower, i));
2795 segregate_def_use_chain_1(follower, pred);
2797 } /* segregate_def_use_chain */
2800 * Propagate constant evaluation.
2802 * @param env the environment
2804 static void propagate(environment_t *env)
2808 lattice_elem_t old_type;
2810 unsigned n_fallen, old_type_was_T_or_C;
2813 while (env->cprop != NULL) {
2814 void *oldopcode = NULL;
2816 /* remove the first partition X from cprop */
2819 env->cprop = X->cprop_next;
2821 old_type_was_T_or_C = X->type_is_T_or_C;
2823 DB((dbg, LEVEL_2, "Propagate type on part%d\n", X->nr));
2827 int cprop_empty = list_empty(&X->cprop);
2828 int cprop_X_empty = list_empty(&X->cprop_X);
2830 if (cprop_empty && cprop_X_empty) {
2831 /* both cprop lists are empty */
2835 /* remove the first Node x from X.cprop */
2837 /* Get a node from the cprop_X list only if
2838 * all data nodes are processed.
2839 * This ensures, that all inputs of the Cond
2840 * predecessor are processed if its type is still Top.
2842 x = list_entry(X->cprop_X.next, node_t, cprop_list);
2844 x = list_entry(X->cprop.next, node_t, cprop_list);
2847 //assert(x->part == X);
2848 list_del(&x->cprop_list);
2851 if (x->is_follower && identity(x) == x) {
2852 /* check the opcode first */
2853 if (oldopcode == NULL) {
2854 oldopcode = lambda_opcode(get_first_node(X), env);
2856 if (oldopcode != lambda_opcode(x, env)) {
2857 if (x->on_fallen == 0) {
2858 /* different opcode -> x falls out of this partition */
2863 DB((dbg, LEVEL_2, "Add node %+F to fallen\n", x->node));
2867 /* x will make the follower -> leader transition */
2868 follower_to_leader(x);
2871 /* compute a new type for x */
2873 DB((dbg, LEVEL_3, "computing type of %+F\n", x->node));
2875 if (x->type.tv != old_type.tv) {
2876 DB((dbg, LEVEL_2, "node %+F has changed type from %+F to %+F\n", x->node, old_type, x->type));
2877 verify_type(old_type, x);
2879 if (x->on_fallen == 0) {
2880 /* Add x to fallen. Nodes might fall from T -> const -> _|_, so check that they are
2881 not already on the list. */
2886 DB((dbg, LEVEL_2, "Add node %+F to fallen\n", x->node));
2888 for (i = get_irn_n_outs(x->node) - 1; i >= 0; --i) {
2889 ir_node *succ = get_irn_out(x->node, i);
2890 node_t *y = get_irn_node(succ);
2892 /* Add y to y.partition.cprop. */
2893 add_to_cprop(y, env);
2898 if (n_fallen > 0 && n_fallen != X->n_leader) {
2899 DB((dbg, LEVEL_2, "Splitting part%d by fallen\n", X->nr));
2900 Y = split(&X, fallen, env);
2902 * We have split out fallen node. The type of the result
2903 * partition is NOT set yet.
2905 Y->type_is_T_or_C = 0;
2909 /* remove the flags from the fallen list */
2910 for (x = fallen; x != NULL; x = x->next)
2913 if (old_type_was_T_or_C) {
2916 /* check if some nodes will make the leader -> follower transition */
2917 list_for_each_entry_safe(node_t, y, tmp, &Y->Leader, node_list) {
2918 if (y->type.tv != tarval_top && ! is_con(y->type)) {
2919 node_t *eq_node = identity(y);
2921 if (eq_node != y && eq_node->part == y->part) {
2922 DB((dbg, LEVEL_2, "Node %+F is a follower of %+F\n", y->node, eq_node->node));
2923 /* move to Follower */
2925 list_del(&y->node_list);
2926 list_add_tail(&y->node_list, &Y->Follower);
2929 segregate_def_use_chain(y->node);
2939 * Get the leader for a given node from its congruence class.
2941 * @param irn the node
2943 static ir_node *get_leader(node_t *node)
2945 partition_t *part = node->part;
2947 if (part->n_leader > 1 || node->is_follower) {
2948 if (node->is_follower) {
2949 DB((dbg, LEVEL_2, "Replacing follower %+F\n", node->node));
2952 DB((dbg, LEVEL_2, "Found congruence class for %+F\n", node->node));
2954 return get_first_node(part)->node;
2960 * Returns non-zero if a mode_T node has only one reachable output.
2962 static int only_one_reachable_proj(ir_node *n)
2966 for (i = get_irn_n_outs(n) - 1; i >= 0; --i) {
2967 ir_node *proj = get_irn_out(n, i);
2970 /* skip non-control flow Proj's */
2971 if (get_irn_mode(proj) != mode_X)
2974 node = get_irn_node(proj);
2975 if (node->type.tv == tarval_reachable) {
2981 } /* only_one_reachable_proj */
2984 * Return non-zero if the control flow predecessor node pred
2985 * is the only reachable control flow exit of its block.
2987 * @param pred the control flow exit
2988 * @param block the destination block
2990 static int can_exchange(ir_node *pred, ir_node *block)
2992 if (is_Start(pred) || get_Block_entity(block) != NULL)
2994 else if (is_Jmp(pred))
2996 else if (is_Raise(pred)) {
2997 /* Raise is a tuple and usually has only one reachable ProjX,
2998 * but it must not be eliminated like a Jmp */
3001 else if (get_irn_mode(pred) == mode_T) {
3002 /* if the predecessor block has more than one
3003 reachable outputs we cannot remove the block */
3004 return only_one_reachable_proj(pred);
3007 } /* can_exchange */
3010 * Block Post-Walker, apply the analysis results on control flow by
3011 * shortening Phi's and Block inputs.
3013 static void apply_cf(ir_node *block, void *ctx)
3015 environment_t *env = (environment_t*)ctx;
3016 node_t *node = get_irn_node(block);
3018 ir_node **ins, **in_X;
3019 ir_node *phi, *next;
3021 n = get_Block_n_cfgpreds(block);
3023 if (node->type.tv == tarval_unreachable) {
3026 for (i = n - 1; i >= 0; --i) {
3027 ir_node *pred = get_Block_cfgpred(block, i);
3029 if (! is_Bad(pred)) {
3030 ir_node *pred_block = get_nodes_block(skip_Proj(pred));
3031 if (!is_Bad(pred_block)) {
3032 node_t *pred_bl = get_irn_node(pred_block);
3034 if (pred_bl->flagged == 0) {
3035 pred_bl->flagged = 3;
3037 if (pred_bl->type.tv == tarval_reachable) {
3039 * We will remove an edge from block to its pred.
3040 * This might leave the pred block as an endless loop
3042 if (! is_backedge(block, i))
3043 keep_alive(pred_bl->node);
3050 if (block == get_irg_end_block(current_ir_graph)) {
3051 /* Analysis found out that the end block is unreachable,
3052 * hence we remove all its control flow predecessors. */
3053 set_irn_in(block, 0, NULL);
3059 /* only one predecessor combine */
3060 ir_node *pred = skip_Proj(get_Block_cfgpred(block, 0));
3062 if (can_exchange(pred, block)) {
3063 ir_node *new_block = get_nodes_block(pred);
3064 DB((dbg, LEVEL_1, "Fuse %+F with %+F\n", block, new_block));
3065 DBG_OPT_COMBO(block, new_block, FS_OPT_COMBO_CF);
3066 exchange(block, new_block);
3067 node->node = new_block;
3073 NEW_ARR_A(ir_node *, in_X, n);
3075 for (i = 0; i < n; ++i) {
3076 ir_node *pred = get_Block_cfgpred(block, i);
3077 node_t *node = get_irn_node(pred);
3079 if (node->type.tv == tarval_reachable) {
3082 DB((dbg, LEVEL_1, "Removing dead input %d from %+F (%+F)\n", i, block, pred));
3083 if (! is_Bad(pred)) {
3084 ir_node *pred_block = get_nodes_block(skip_Proj(pred));
3085 if (!is_Bad(pred_block)) {
3086 node_t *pred_bl = get_irn_node(pred_block);
3088 if (!is_Bad(pred_bl->node) && pred_bl->flagged == 0) {
3089 pred_bl->flagged = 3;
3091 if (pred_bl->type.tv == tarval_reachable) {
3093 * We will remove an edge from block to its pred.
3094 * This might leave the pred block as an endless loop
3096 if (! is_backedge(block, i))
3097 keep_alive(pred_bl->node);
3108 NEW_ARR_A(ir_node *, ins, n);
3109 for (phi = get_Block_phis(block); phi != NULL; phi = next) {
3110 node_t *node = get_irn_node(phi);
3112 next = get_Phi_next(phi);
3113 if (is_tarval(node->type.tv) && tarval_is_constant(node->type.tv)) {
3114 /* this Phi is replaced by a constant */
3115 ir_tarval *tv = node->type.tv;
3116 ir_node *c = new_r_Const(current_ir_graph, tv);
3118 set_irn_node(c, node);
3120 DB((dbg, LEVEL_1, "%+F is replaced by %+F\n", phi, c));
3121 DBG_OPT_COMBO(phi, c, FS_OPT_COMBO_CONST);
3126 for (i = 0; i < n; ++i) {
3127 node_t *pred = get_irn_node(get_Block_cfgpred(block, i));
3129 if (pred->type.tv == tarval_reachable) {
3130 ins[j++] = get_Phi_pred(phi, i);
3134 /* this Phi is replaced by a single predecessor */
3135 ir_node *s = ins[0];
3136 node_t *phi_node = get_irn_node(phi);
3139 DB((dbg, LEVEL_1, "%+F is replaced by %+F because of cf change\n", phi, s));
3140 DBG_OPT_COMBO(phi, s, FS_OPT_COMBO_FOLLOWER);
3145 set_irn_in(phi, j, ins);
3153 /* this Block has only one live predecessor */
3154 ir_node *pred = skip_Proj(in_X[0]);
3156 if (can_exchange(pred, block)) {
3157 ir_node *new_block = get_nodes_block(pred);
3158 DBG_OPT_COMBO(block, new_block, FS_OPT_COMBO_CF);
3159 exchange(block, new_block);
3160 node->node = new_block;
3165 set_irn_in(block, k, in_X);
3170 * Exchange a node by its leader.
3171 * Beware: in rare cases the mode might be wrong here, for instance
3172 * AddP(x, NULL) is a follower of x, but with different mode.
3175 static void exchange_leader(ir_node *irn, ir_node *leader)
3177 ir_mode *mode = get_irn_mode(irn);
3178 if (mode != get_irn_mode(leader)) {
3179 /* The conv is a no-op, so we are free to place it
3180 * either in the block of the leader OR in irn's block.
3181 * Probably placing it into leaders block might reduce
3182 * the number of Conv due to CSE. */
3183 ir_node *block = get_nodes_block(leader);
3184 dbg_info *dbg = get_irn_dbg_info(irn);
3185 ir_node *nlead = new_rd_Conv(dbg, block, leader, mode);
3187 if (nlead != leader) {
3188 /* Note: this newly create irn has no node info because
3189 * it is created after the analysis. However, this node
3190 * replaces the node irn and should not be visited again,
3191 * so set its visited count to the count of irn.
3192 * Otherwise we might visited this node more than once if
3193 * irn had more than one user.
3195 set_irn_node(nlead, NULL);
3196 set_irn_visited(nlead, get_irn_visited(irn));
3200 exchange(irn, leader);
3201 } /* exchange_leader */
3204 * Check, if all users of a mode_M node are dead. Use
3205 * the Def-Use edges for this purpose, as they still
3206 * reflect the situation.
3208 static int all_users_are_dead(const ir_node *irn)
3210 int i, n = get_irn_n_outs(irn);
3212 for (i = 1; i <= n; ++i) {
3213 const ir_node *succ = irn->out[i].use;
3214 const node_t *block = get_irn_node(get_nodes_block(succ));
3217 if (block->type.tv == tarval_unreachable) {
3218 /* block is unreachable */
3221 node = get_irn_node(succ);
3222 if (node->type.tv != tarval_top) {
3223 /* found a reachable user */
3227 /* all users are unreachable */
3229 } /* all_user_are_dead */
3232 * Walker: Find reachable mode_M nodes that have only
3233 * unreachable users. These nodes must be kept later.
3235 static void find_kept_memory(ir_node *irn, void *ctx)
3237 environment_t *env = (environment_t*)ctx;
3238 node_t *node, *block;
3240 if (get_irn_mode(irn) != mode_M)
3243 block = get_irn_node(get_nodes_block(irn));
3244 if (block->type.tv == tarval_unreachable)
3247 node = get_irn_node(irn);
3248 if (node->type.tv == tarval_top)
3251 /* ok, we found a live memory node. */
3252 if (all_users_are_dead(irn)) {
3253 DB((dbg, LEVEL_1, "%+F must be kept\n", irn));
3254 ARR_APP1(ir_node *, env->kept_memory, irn);
3256 } /* find_kept_memory */
3259 * Post-Walker, apply the analysis results;
3261 static void apply_result(ir_node *irn, void *ctx)
3263 environment_t *env = (environment_t*)ctx;
3264 node_t *node = get_irn_node(irn);
3266 if (is_Block(irn) || is_End(irn) || is_Bad(irn)) {
3267 /* blocks already handled, do not touch the End node */
3269 node_t *block = get_irn_node(get_nodes_block(irn));
3271 if (block->type.tv == tarval_unreachable) {
3272 ir_graph *irg = get_irn_irg(irn);
3273 ir_mode *mode = get_irn_mode(node->node);
3274 ir_node *bad = new_r_Bad(irg, mode);
3276 /* here, bad might already have a node, but this can be safely ignored
3277 as long as bad has at least ONE valid node */
3278 set_irn_node(bad, node);
3280 DB((dbg, LEVEL_1, "%+F is unreachable\n", irn));
3283 } else if (node->type.tv == tarval_top) {
3284 ir_mode *mode = get_irn_mode(irn);
3286 if (mode == mode_M) {
3287 /* never kill a mode_M node */
3289 ir_node *pred = get_Proj_pred(irn);
3290 node_t *pnode = get_irn_node(pred);
3292 if (pnode->type.tv == tarval_top) {
3293 /* skip the predecessor */
3294 ir_node *mem = get_memop_mem(pred);
3296 DB((dbg, LEVEL_1, "%+F computes Top, replaced by %+F\n", irn, mem));
3301 /* leave other nodes, especially PhiM */
3302 } else if (mode == mode_T) {
3303 /* Do not kill mode_T nodes, kill their Projs */
3304 } else if (! is_Unknown(irn)) {
3305 /* don't kick away Unknown's, they might be still needed */
3306 ir_node *unk = new_r_Unknown(current_ir_graph, mode);
3308 /* control flow should already be handled at apply_cf() */
3309 assert(mode != mode_X);
3311 /* see comment above */
3312 set_irn_node(unk, node);
3314 DB((dbg, LEVEL_1, "%+F computes Top\n", irn));
3319 else if (get_irn_mode(irn) == mode_X) {
3322 ir_node *cond = get_Proj_pred(irn);
3324 if (is_Cond(cond) || is_Switch(cond)) {
3325 if (only_one_reachable_proj(cond)) {
3326 ir_node *jmp = new_r_Jmp(block->node);
3327 set_irn_node(jmp, node);
3329 DB((dbg, LEVEL_1, "%+F is replaced by %+F\n", irn, jmp));
3330 DBG_OPT_COMBO(irn, jmp, FS_OPT_COMBO_CF);
3334 if (is_Switch(cond)) {
3335 node_t *sel = get_irn_node(get_Switch_selector(cond));
3336 ir_tarval *tv = sel->type.tv;
3338 if (is_tarval(tv) && tarval_is_constant(tv)) {
3339 /* The selector is a constant, but more
3340 * than one output is active: An unoptimized
3349 /* normal data node */
3350 if (is_tarval(node->type.tv) && tarval_is_constant(node->type.tv)) {
3351 ir_tarval *tv = node->type.tv;
3354 * Beware: never replace mode_T nodes by constants. Currently we must mark
3355 * mode_T nodes with constants, but do NOT replace them.
3357 if (! is_Const(irn) && get_irn_mode(irn) != mode_T) {
3358 /* can be replaced by a constant */
3359 ir_node *c = new_r_Const(current_ir_graph, tv);
3360 set_irn_node(c, node);
3362 DB((dbg, LEVEL_1, "%+F is replaced by %+F\n", irn, c));
3363 DBG_OPT_COMBO(irn, c, FS_OPT_COMBO_CONST);
3364 exchange_leader(irn, c);
3367 } else if (is_entity(node->type.sym.entity_p)) {
3368 if (! is_SymConst(irn)) {
3369 /* can be replaced by a SymConst */
3370 ir_node *symc = new_r_SymConst(current_ir_graph, get_irn_mode(irn), node->type.sym, symconst_addr_ent);
3371 set_irn_node(symc, node);
3374 DB((dbg, LEVEL_1, "%+F is replaced by %+F\n", irn, symc));
3375 DBG_OPT_COMBO(irn, symc, FS_OPT_COMBO_CONST);
3376 exchange_leader(irn, symc);
3379 } else if (is_Confirm(irn)) {
3380 /* Confirms are always follower, but do not kill them here */
3382 ir_node *leader = get_leader(node);
3384 if (leader != irn) {
3385 int non_strict_phi = 0;
3388 * Beware: Do not remove Phi(Unknown, ..., x, ..., Unknown)
3389 * as this might create non-strict programs.
3391 if (node->is_follower && is_Phi(irn) && !is_Unknown(leader)) {
3394 for (i = get_Phi_n_preds(irn) - 1; i >= 0; --i) {
3395 ir_node *pred = get_Phi_pred(irn, i);
3397 if (is_Unknown(pred)) {
3403 if (! non_strict_phi) {
3404 DB((dbg, LEVEL_1, "%+F from part%d is replaced by %+F\n", irn, node->part->nr, leader));
3405 if (node->is_follower)
3406 DBG_OPT_COMBO(irn, leader, FS_OPT_COMBO_FOLLOWER);
3408 DBG_OPT_COMBO(irn, leader, FS_OPT_COMBO_CONGRUENT);
3409 exchange_leader(irn, leader);
3416 } /* apply_result */
3419 * Fix the keep-alives by deleting unreachable ones.
3421 static void apply_end(ir_node *end, environment_t *env)
3423 int i, j, n = get_End_n_keepalives(end);
3424 ir_node **in = NULL;
3427 NEW_ARR_A(ir_node *, in, n);
3429 /* fix the keep alive */
3430 for (i = j = 0; i < n; i++) {
3431 ir_node *ka = get_End_keepalive(end, i);
3437 if (!is_Block(ka)) {
3438 block = get_nodes_block(ka);
3445 node = get_irn_node(block);
3446 if (node->type.tv != tarval_unreachable)
3450 set_End_keepalives(end, j, in);
3455 #define SET(code) op_##code->ops.generic = (op_func)compute_##code
3458 * sets the generic functions to compute.
3460 static void set_compute_functions(void)
3464 /* set the default compute function */
3465 for (i = 0, n = ir_get_n_opcodes(); i < n; ++i) {
3466 ir_op *op = ir_get_opcode(i);
3467 op->ops.generic = (op_func)default_compute;
3470 /* set specific functions */
3486 } /* set_compute_functions */
3491 static void add_memory_keeps(ir_node **kept_memory, size_t len)
3493 ir_node *end = get_irg_end(current_ir_graph);
3498 ir_nodeset_init(&set);
3500 /* check, if those nodes are already kept */
3501 for (i = get_End_n_keepalives(end) - 1; i >= 0; --i)
3502 ir_nodeset_insert(&set, get_End_keepalive(end, i));
3504 for (idx = 0; idx < len; ++idx) {
3505 ir_node *ka = kept_memory[idx];
3507 if (! ir_nodeset_contains(&set, ka)) {
3508 add_End_keepalive(end, ka);
3511 ir_nodeset_destroy(&set);
3512 } /* add_memory_keeps */
3514 void combo(ir_graph *irg)
3517 ir_node *initial_bl;
3519 ir_graph *rem = current_ir_graph;
3522 assure_irg_properties(irg,
3523 IR_GRAPH_PROPERTY_NO_BADS
3524 | IR_GRAPH_PROPERTY_CONSISTENT_OUTS
3525 | IR_GRAPH_PROPERTY_CONSISTENT_LOOPINFO);
3527 current_ir_graph = irg;
3529 /* register a debug mask */
3530 FIRM_DBG_REGISTER(dbg, "firm.opt.combo");
3532 DB((dbg, LEVEL_1, "Doing COMBO for %+F\n", irg));
3534 obstack_init(&env.obst);
3535 env.worklist = NULL;
3539 #ifdef DEBUG_libfirm
3540 env.dbg_list = NULL;
3542 env.opcode2id_map = new_set(cmp_opcode, iro_Last * 4);
3543 env.kept_memory = NEW_ARR_F(ir_node *, 0);
3544 env.end_idx = get_opt_global_cse() ? 0 : -1;
3545 env.lambda_input = 0;
3548 /* options driving the optimization */
3549 env.commutative = 1;
3550 env.opt_unknown = 1;
3552 /* we have our own value_of function */
3553 set_value_of_func(get_node_tarval);
3555 set_compute_functions();
3556 DEBUG_ONLY(part_nr = 0;)
3558 ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK | IR_RESOURCE_PHI_LIST);
3560 if (env.opt_unknown)
3561 tarval_UNKNOWN = tarval_top;
3563 tarval_UNKNOWN = tarval_bad;
3565 /* create the initial partition and place it on the work list */
3566 env.initial = new_partition(&env);
3567 add_to_worklist(env.initial, &env);
3568 irg_walk_graph(irg, create_initial_partitions, init_block_phis, &env);
3570 /* set the hook: from now, every node has a partition and a type */
3571 DEBUG_ONLY(set_dump_node_vcgattr_hook(dump_partition_hook);)
3573 /* all nodes on the initial partition have type Top */
3574 env.initial->type_is_T_or_C = 1;
3576 /* Place the START Node's partition on cprop.
3577 Place the START Node on its local worklist. */
3578 initial_bl = get_irg_start_block(irg);
3579 start = get_irn_node(initial_bl);
3580 add_to_cprop(start, &env);
3584 if (env.worklist != NULL)
3586 } while (env.cprop != NULL || env.worklist != NULL);
3588 dump_all_partitions(&env);
3589 check_all_partitions(&env);
3592 dump_ir_block_graph(irg, "-partition");
3595 /* apply the result */
3597 /* check, which nodes must be kept */
3598 irg_walk_graph(irg, NULL, find_kept_memory, &env);
3600 /* kill unreachable control flow */
3601 irg_block_walk_graph(irg, NULL, apply_cf, &env);
3602 /* Kill keep-alives of dead blocks: this speeds up apply_result()
3603 * and fixes assertion because dead cf to dead blocks is NOT removed by
3605 apply_end(get_irg_end(irg), &env);
3606 irg_walk_graph(irg, NULL, apply_result, &env);
3608 len = ARR_LEN(env.kept_memory);
3610 add_memory_keeps(env.kept_memory, len);
3613 DB((dbg, LEVEL_1, "Unoptimized Control Flow left"));
3616 ir_free_resources(irg, IR_RESOURCE_IRN_LINK | IR_RESOURCE_PHI_LIST);
3618 /* remove the partition hook */
3619 DEBUG_ONLY(set_dump_node_vcgattr_hook(NULL);)
3621 DEL_ARR_F(env.kept_memory);
3622 del_set(env.opcode2id_map);
3623 obstack_free(&env.obst, NULL);
3625 /* restore value_of() default behavior */
3626 set_value_of_func(NULL);
3627 current_ir_graph = rem;
3629 confirm_irg_properties(irg, IR_GRAPH_PROPERTIES_NONE);
3632 /* Creates an ir_graph pass for combo. */
3633 ir_graph_pass_t *combo_pass(const char *name)
3635 return def_graph_pass(name ? name : "combo", combo);