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 other 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 reevaluated
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 unsigned next_edge; /**< Index of the next Def-Use edge to use. */
148 unsigned 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)
272 list_for_each_entry(node_t, node, &T->Leader, node_list) {
273 assert(node->is_follower == 0);
274 assert(node->flagged == 0);
275 assert(node->part == T);
278 assert(n == T->n_leader);
280 list_for_each_entry(node_t, node, &T->Follower, node_list) {
281 assert(node->is_follower == 1);
282 assert(node->flagged == 0);
283 assert(node->part == T);
285 } /* check_partition */
288 * check that all leader nodes in the partition have the same opcode.
290 static void check_opcode(const partition_t *Z)
292 const ir_node *repr = NULL;
294 list_for_each_entry(node_t, node, &Z->Leader, node_list) {
295 ir_node *irn = node->node;
300 assert(cmp_irn_opcode(repr, irn) == 0);
305 static void check_all_partitions(environment_t *env)
310 for (P = env->dbg_list; P != NULL; P = P->dbg_next) {
312 if (! P->type_is_T_or_C)
314 list_for_each_entry(node_t, node, &P->Follower, node_list) {
315 node_t *leader = identity(node);
317 assert(leader != node && leader->part == node->part);
328 static void do_check_list(const node_t *list, int ofs, const partition_t *Z)
333 #define NEXT(e) *((const node_t **)((char *)(e) + (ofs)))
334 for (e = list; e != NULL; e = NEXT(e)) {
335 assert(e->part == Z);
343 } /* ido_check_list */
346 * Check a local list.
348 static void check_list(const node_t *list, const partition_t *Z)
350 do_check_list(list, offsetof(node_t, next), Z);
354 #define check_partition(T)
355 #define check_list(list, Z)
356 #define check_all_partitions(env)
357 #endif /* CHECK_PARTITIONS */
360 static inline lattice_elem_t get_partition_type(const partition_t *X);
363 * Dump partition to output.
365 static void dump_partition(const char *msg, const partition_t *part)
368 lattice_elem_t type = get_partition_type(part);
370 DB((dbg, LEVEL_2, "%s part%u%s (%u, %+F) {\n ",
371 msg, part->nr, part->type_is_T_or_C ? "*" : "",
372 part->n_leader, type));
373 list_for_each_entry(node_t, node, &part->Leader, node_list) {
374 DB((dbg, LEVEL_2, "%s%+F", first ? "" : ", ", node->node));
377 if (! list_empty(&part->Follower)) {
378 DB((dbg, LEVEL_2, "\n---\n "));
380 list_for_each_entry(node_t, node, &part->Follower, node_list) {
381 DB((dbg, LEVEL_2, "%s%+F", first ? "" : ", ", node->node));
385 DB((dbg, LEVEL_2, "\n}\n"));
386 } /* dump_partition */
391 static void do_dump_list(const char *msg, const node_t *node, int ofs)
396 #define GET_LINK(p, ofs) *((const node_t **)((char *)(p) + (ofs)))
398 DB((dbg, LEVEL_3, "%s = {\n ", msg));
399 for (p = node; p != NULL; p = GET_LINK(p, ofs)) {
400 DB((dbg, LEVEL_3, "%s%+F", first ? "" : ", ", p->node));
403 DB((dbg, LEVEL_3, "\n}\n"));
411 static void dump_race_list(const char *msg, const node_t *list)
413 do_dump_list(msg, list, offsetof(node_t, race_next));
414 } /* dump_race_list */
417 * Dumps a local list.
419 static void dump_list(const char *msg, const node_t *list)
421 do_dump_list(msg, list, offsetof(node_t, next));
425 * Dump all partitions.
427 static void dump_all_partitions(const environment_t *env)
429 const partition_t *P;
431 DB((dbg, LEVEL_2, "All partitions\n===============\n"));
432 for (P = env->dbg_list; P != NULL; P = P->dbg_next)
433 dump_partition("", P);
434 } /* dump_all_partitions */
439 static void dump_split_list(const partition_t *list)
441 const partition_t *p;
443 DB((dbg, LEVEL_2, "Split by %s produced = {\n", what_reason));
444 for (p = list; p != NULL; p = p->split_next)
445 DB((dbg, LEVEL_2, "part%u, ", p->nr));
446 DB((dbg, LEVEL_2, "\n}\n"));
447 } /* dump_split_list */
450 * Dump partition and type for a node.
452 static int dump_partition_hook(FILE *F, const ir_node *n, const ir_node *local)
454 const ir_node *irn = local != NULL ? local : n;
455 node_t *node = get_irn_node(irn);
457 ir_fprintf(F, "info2 : \"partition %u type %+F\"\n", node->part->nr, node->type);
459 } /* dump_partition_hook */
462 #define dump_partition(msg, part)
463 #define dump_race_list(msg, list)
464 #define dump_list(msg, list)
465 #define dump_all_partitions(env)
466 #define dump_split_list(list)
469 #if defined(VERIFY_MONOTONE) && defined (DEBUG_libfirm)
471 * Verify that a type transition is monotone
473 static void verify_type(const lattice_elem_t old_type, node_t *node)
475 if (old_type.tv == node->type.tv) {
479 if (old_type.tv == tarval_top) {
480 /* from Top down-to is always allowed */
483 if (node->type.tv == tarval_bottom || node->type.tv == tarval_reachable) {
487 panic("wrong translation from %+F to %+F on node %+F", old_type, node->type, node->node);
491 #define verify_type(old_type, node)
495 * Compare two pointer values of a listmap.
497 static int listmap_cmp_ptr(const void *elt, const void *key, size_t size)
499 const listmap_entry_t *e1 = (listmap_entry_t*)elt;
500 const listmap_entry_t *e2 = (listmap_entry_t*)key;
503 return e1->id != e2->id;
504 } /* listmap_cmp_ptr */
507 * Initializes a listmap.
509 * @param map the listmap
511 static void listmap_init(listmap_t *map)
513 map->map = new_set(listmap_cmp_ptr, 16);
518 * Terminates a listmap.
520 * @param map the listmap
522 static void listmap_term(listmap_t *map)
528 * Return the associated listmap entry for a given id.
530 * @param map the listmap
531 * @param id the id to search for
533 * @return the associated listmap entry for the given id
535 static listmap_entry_t *listmap_find(listmap_t *map, void *id)
537 listmap_entry_t key, *entry;
542 entry = set_insert(listmap_entry_t, map->map, &key, sizeof(key), hash_ptr(id));
544 if (entry->list == NULL) {
545 /* a new entry, put into the list */
546 entry->next = map->values;
553 * Calculate the hash value for an opcode map entry.
555 * @param entry an opcode map entry
557 * @return a hash value for the given opcode map entry
559 static unsigned opcode_hash(const opcode_key_t *entry)
561 /* we cannot use the ir ops hash function here, because it hashes the
563 const ir_node *n = entry->irn;
564 ir_opcode code = (ir_opcode)get_irn_opcode(n);
565 ir_mode *mode = get_irn_mode(n);
566 unsigned hash = (unsigned)(PTR_TO_INT(mode) * 9 + code) + get_irn_arity(n);
568 if (code == iro_Const)
569 hash ^= (unsigned)hash_ptr(get_Const_tarval(n));
570 else if (code == iro_Proj)
571 hash += (unsigned)get_Proj_proj(n);
576 * Compare two entries in the opcode map.
578 static int cmp_opcode(const void *elt, const void *key, size_t size)
580 const opcode_key_t *o1 = (opcode_key_t*)elt;
581 const opcode_key_t *o2 = (opcode_key_t*)key;
585 return cmp_irn_opcode(o1->irn, o2->irn);
589 * Compare two Def-Use edges for input position.
591 static int cmp_def_use_edge(const void *a, const void *b)
593 const ir_def_use_edge *ea = (const ir_def_use_edge*)a;
594 const ir_def_use_edge *eb = (const ir_def_use_edge*)b;
596 /* no overrun, because range is [-1, MAXINT] */
597 return ea->pos - eb->pos;
598 } /* cmp_def_use_edge */
601 * We need the Def-Use edges sorted.
603 static void sort_irn_outs(node_t *node)
605 ir_node *irn = node->node;
606 unsigned n_outs = get_irn_n_outs(irn);
607 qsort(irn->o.out->edges, n_outs, sizeof(irn->o.out->edges[0]),
609 node->max_user_input = n_outs > 0 ? irn->o.out->edges[n_outs-1].pos : -1;
610 } /* sort_irn_outs */
613 * Return the type of a node.
615 * @param irn an IR-node
617 * @return the associated type of this node
619 static inline lattice_elem_t get_node_type(const ir_node *irn)
621 return get_irn_node(irn)->type;
622 } /* get_node_type */
625 * Return the tarval of a node.
627 * @param irn an IR-node
629 * @return the associated type of this node
631 static inline ir_tarval *get_node_tarval(const ir_node *irn)
633 lattice_elem_t type = get_node_type(irn);
635 if (is_tarval(type.tv))
637 return tarval_bottom;
638 } /* get_node_type */
641 * Add a partition to the worklist.
643 static inline void add_to_worklist(partition_t *X, environment_t *env)
645 assert(X->on_worklist == 0);
646 DB((dbg, LEVEL_2, "Adding part%d to worklist\n", X->nr));
647 X->wl_next = env->worklist;
650 } /* add_to_worklist */
653 * Create a new empty partition.
655 * @param env the environment
657 * @return a newly allocated partition
659 static inline partition_t *new_partition(environment_t *env)
661 partition_t *part = OALLOC(&env->obst, partition_t);
663 INIT_LIST_HEAD(&part->Leader);
664 INIT_LIST_HEAD(&part->Follower);
665 INIT_LIST_HEAD(&part->cprop);
666 INIT_LIST_HEAD(&part->cprop_X);
667 part->wl_next = NULL;
668 part->touched_next = NULL;
669 part->cprop_next = NULL;
670 part->split_next = NULL;
671 part->touched = NULL;
674 part->max_user_inputs = 0;
675 part->on_worklist = 0;
676 part->on_touched = 0;
678 part->type_is_T_or_C = 0;
680 part->dbg_next = env->dbg_list;
681 env->dbg_list = part;
682 part->nr = part_nr++;
686 } /* new_partition */
689 * Get the first node from a partition.
691 static inline node_t *get_first_node(const partition_t *X)
693 return list_entry(X->Leader.next, node_t, node_list);
694 } /* get_first_node */
697 * Return the type of a partition (assuming partition is non-empty and
698 * all elements have the same type).
700 * @param X a partition
702 * @return the type of the first element of the partition
704 static inline lattice_elem_t get_partition_type(const partition_t *X)
706 const node_t *first = get_first_node(X);
708 } /* get_partition_type */
711 * Creates a partition node for the given IR-node and place it
712 * into the given partition.
714 * @param irn an IR-node
715 * @param part a partition to place the node in
716 * @param env the environment
718 * @return the created node
720 static node_t *create_partition_node(ir_node *irn, partition_t *part, environment_t *env)
722 /* create a partition node and place it in the partition */
723 node_t *node = OALLOC(&env->obst, node_t);
725 INIT_LIST_HEAD(&node->node_list);
726 INIT_LIST_HEAD(&node->cprop_list);
730 node->race_next = NULL;
731 node->type.tv = tarval_top;
732 node->max_user_input = 0;
734 node->n_followers = 0;
735 node->on_touched = 0;
738 node->is_follower = 0;
740 set_irn_node(irn, node);
742 list_add_tail(&node->node_list, &part->Leader);
746 } /* create_partition_node */
749 * Pre-Walker, initialize all Nodes' type to U or top and place
750 * all nodes into the TOP partition.
752 static void create_initial_partitions(ir_node *irn, void *ctx)
754 environment_t *env = (environment_t*)ctx;
755 partition_t *part = env->initial;
758 node = create_partition_node(irn, part, env);
760 if (node->max_user_input > part->max_user_inputs)
761 part->max_user_inputs = node->max_user_input;
764 set_Block_phis(irn, NULL);
766 } /* create_initial_partitions */
769 * Post-Walker, collect all Block-Phi lists, set Cond.
771 static void init_block_phis(ir_node *irn, void *ctx)
776 ir_node *block = get_nodes_block(irn);
777 add_Block_phi(block, irn);
779 } /* init_block_phis */
782 * Add a node to the entry.partition.touched set and
783 * node->partition to the touched set if not already there.
786 * @param env the environment
788 static inline void add_to_touched(node_t *y, environment_t *env)
790 if (y->on_touched == 0) {
791 partition_t *part = y->part;
793 y->next = part->touched;
798 if (part->on_touched == 0) {
799 part->touched_next = env->touched;
801 part->on_touched = 1;
804 check_list(part->touched, part);
806 } /* add_to_touched */
809 * Place a node on the cprop list.
812 * @param env the environment
814 static void add_to_cprop(node_t *y, environment_t *env)
818 /* Add y to y.partition.cprop. */
819 if (y->on_cprop == 0) {
820 partition_t *Y = y->part;
821 ir_node *irn = y->node;
822 ir_node *skipped = skip_Proj(irn);
824 /* place Conds and all its Projs on the cprop_X list */
825 if (is_Cond(skipped) || is_Switch(skipped))
826 list_add_tail(&y->cprop_list, &Y->cprop_X);
828 list_add_tail(&y->cprop_list, &Y->cprop);
831 DB((dbg, LEVEL_3, "Add %+F to part%u.cprop\n", y->node, Y->nr));
833 /* place its partition on the cprop list */
834 if (Y->on_cprop == 0) {
835 Y->cprop_next = env->cprop;
841 if (get_irn_mode(irn) == mode_T) {
842 /* mode_T nodes always produce tarval_bottom, so we must explicitly
843 * add its Projs to get constant evaluation to work */
844 for (unsigned i = get_irn_n_outs(irn); i-- > 0; ) {
845 node_t *proj = get_irn_node(get_irn_out(irn, i));
847 add_to_cprop(proj, env);
849 } else if (is_Block(irn)) {
850 /* Due to the way we handle Phi's, we must place all Phis of a block on the list
851 * if someone placed the block. The Block is only placed if the reachability
852 * changes, and this must be re-evaluated in compute_Phi(). */
854 for (phi = get_Block_phis(irn); phi != NULL; phi = get_Phi_next(phi)) {
855 node_t *p = get_irn_node(phi);
856 add_to_cprop(p, env);
862 * Update the worklist: If Z is on worklist then add Z' to worklist.
863 * Else add the smaller of Z and Z' to worklist.
865 * @param Z the Z partition
866 * @param Z_prime the Z' partition, a previous part of Z
867 * @param env the environment
869 static void update_worklist(partition_t *Z, partition_t *Z_prime, environment_t *env)
871 if (Z->on_worklist || Z_prime->n_leader < Z->n_leader) {
872 add_to_worklist(Z_prime, env);
874 add_to_worklist(Z, env);
876 } /* update_worklist */
879 * Make all inputs to x no longer be F.def_use edges.
883 static void move_edges_to_leader(node_t *x)
885 ir_node *irn = x->node;
886 for (int i = get_irn_arity(irn) - 1; i >= 0; --i) {
887 node_t *pred = get_irn_node(get_irn_n(irn, i));
888 ir_node *p = pred->node;
889 unsigned n = get_irn_n_outs(p);
890 for (unsigned j = 0; j < pred->n_followers; ++j) {
891 ir_def_use_edge edge = p->o.out->edges[j];
892 if (edge.pos == i && edge.use == irn) {
893 /* found a follower edge to x, move it to the Leader */
894 /* remove this edge from the Follower set */
896 p->o.out->edges[j] = p->o.out->edges[pred->n_followers];
898 /* sort it into the leader set */
900 for (k = pred->n_followers+1; k < n; ++k) {
901 if (p->o.out->edges[k].pos >= edge.pos)
903 p->o.out->edges[k-1] = p->o.out->edges[k];
905 /* place the new edge here */
906 p->o.out->edges[k-1] = edge;
908 /* edge found and moved */
913 } /* move_edges_to_leader */
916 * Split a partition that has NO followers by a local list.
918 * @param Z partition to split
919 * @param g a (non-empty) node list
920 * @param env the environment
922 * @return a new partition containing the nodes of g
924 static partition_t *split_no_followers(partition_t *Z, node_t *g, environment_t *env)
926 partition_t *Z_prime;
931 dump_partition("Splitting ", Z);
932 dump_list("by list ", g);
936 /* Remove g from Z. */
937 for (node = g; node != NULL; node = node->next) {
938 assert(node->part == Z);
939 list_del(&node->node_list);
942 assert(n < Z->n_leader);
945 /* Move g to a new partition, Z'. */
946 Z_prime = new_partition(env);
948 for (node = g; node != NULL; node = node->next) {
949 list_add_tail(&node->node_list, &Z_prime->Leader);
950 node->part = Z_prime;
951 if (node->max_user_input > max_input)
952 max_input = node->max_user_input;
954 Z_prime->max_user_inputs = max_input;
955 Z_prime->n_leader = n;
958 check_partition(Z_prime);
960 /* for now, copy the type info tag, it will be adjusted in split_by(). */
961 Z_prime->type_is_T_or_C = Z->type_is_T_or_C;
963 dump_partition("Now ", Z);
964 dump_partition("Created new ", Z_prime);
966 update_worklist(Z, Z_prime, env);
969 } /* split_no_followers */
972 * Make the Follower -> Leader transition for a node.
976 static void follower_to_leader(node_t *n)
978 assert(n->is_follower == 1);
980 DB((dbg, LEVEL_2, "%+F make the follower -> leader transition\n", n->node));
982 move_edges_to_leader(n);
983 list_del(&n->node_list);
984 list_add_tail(&n->node_list, &n->part->Leader);
986 } /* follower_to_leader */
989 * The environment for one race step.
991 typedef struct step_env {
992 node_t *initial; /**< The initial node list. */
993 node_t *unwalked; /**< The unwalked node list. */
994 node_t *walked; /**< The walked node list. */
995 unsigned index; /**< Next index of Follower use_def edge. */
996 unsigned side; /**< side number. */
1000 * Return non-zero, if a input is a real follower
1002 * @param irn the node to check
1003 * @param input number of the input
1005 static int is_real_follower(const ir_node *irn, int input)
1009 switch (get_irn_opcode(irn)) {
1012 /* ignore the Confirm bound input */
1018 /* ignore the Mux sel input */
1023 /* dead inputs are not follower edges */
1024 ir_node *block = get_nodes_block(irn);
1025 node_t *pred = get_irn_node(get_Block_cfgpred(block, input));
1027 if (pred->type.tv == tarval_unreachable)
1037 /* only a Sub x,0 / Shift x,0 might be a follower */
1044 pred = get_irn_node(get_irn_n(irn, input));
1045 if (is_tarval(pred->type.tv) && tarval_is_null(pred->type.tv))
1049 pred = get_irn_node(get_irn_n(irn, input));
1050 if (is_tarval(pred->type.tv) && tarval_is_one(pred->type.tv))
1054 pred = get_irn_node(get_irn_n(irn, input));
1055 if (is_tarval(pred->type.tv) && tarval_is_all_one(pred->type.tv))
1059 assert(!"opcode not implemented yet");
1063 } /* is_real_follower */
1066 * Do one step in the race.
1068 static int step(step_env *env)
1072 if (env->initial != NULL) {
1073 /* Move node from initial to unwalked */
1075 env->initial = n->race_next;
1077 n->race_next = env->unwalked;
1083 while (env->unwalked != NULL) {
1084 /* let n be the first node in unwalked */
1086 while (env->index < n->n_followers) {
1087 const ir_def_use_edge *edge = &n->node->o.out->edges[env->index];
1089 /* let m be n.F.def_use[index] */
1090 node_t *m = get_irn_node(edge->use);
1092 assert(m->is_follower);
1094 * Some inputs, like the get_Confirm_bound are NOT
1095 * real followers, sort them out.
1097 if (! is_real_follower(m->node, edge->pos)) {
1103 /* only followers from our partition */
1104 if (m->part != n->part)
1107 if ((m->flagged & env->side) == 0) {
1108 m->flagged |= env->side;
1110 if (m->flagged != 3) {
1111 /* visited the first time */
1112 /* add m to unwalked not as first node (we might still need to
1113 check for more follower node */
1114 m->race_next = n->race_next;
1118 /* else already visited by the other side and on the other list */
1121 /* move n to walked */
1122 env->unwalked = n->race_next;
1123 n->race_next = env->walked;
1131 * Clear the flags from a list and check for
1132 * nodes that where touched from both sides.
1134 * @param list the list
1136 static int clear_flags(node_t *list)
1141 for (n = list; n != NULL; n = n->race_next) {
1142 if (n->flagged == 3) {
1143 /* we reach a follower from both sides, this will split congruent
1144 * inputs and make it a leader. */
1145 follower_to_leader(n);
1154 * Split a partition by a local list using the race.
1156 * @param pX pointer to the partition to split, might be changed!
1157 * @param gg a (non-empty) node list
1158 * @param env the environment
1160 * @return a new partition containing the nodes of gg
1162 static partition_t *split(partition_t **pX, node_t *gg, environment_t *env)
1164 partition_t *X = *pX;
1165 partition_t *X_prime;
1169 int max_input, transitions, winner, shf;
1171 DEBUG_ONLY(static int run = 0;)
1173 DB((dbg, LEVEL_2, "Run %d ", run++));
1174 if (list_empty(&X->Follower)) {
1175 /* if the partition has NO follower, we can use the fast
1176 splitting algorithm. */
1177 return split_no_followers(X, gg, env);
1179 /* else do the race */
1181 dump_partition("Splitting ", X);
1182 dump_list("by list ", gg);
1184 INIT_LIST_HEAD(&tmp);
1186 /* Remove gg from X.Leader and put into g */
1188 for (node_t *node = gg; node != NULL; node = node->next) {
1189 assert(node->part == X);
1190 assert(node->is_follower == 0);
1192 list_del(&node->node_list);
1193 list_add_tail(&node->node_list, &tmp);
1194 node->race_next = g;
1199 list_for_each_entry(node_t, node, &X->Leader, node_list) {
1200 node->race_next = h;
1203 /* restore X.Leader */
1204 list_splice(&tmp, &X->Leader);
1206 senv[0].initial = g;
1207 senv[0].unwalked = NULL;
1208 senv[0].walked = NULL;
1212 senv[1].initial = h;
1213 senv[1].unwalked = NULL;
1214 senv[1].walked = NULL;
1219 * Some informations on the race that are not stated clearly in Click's
1221 * 1) A follower stays on the side that reach him first.
1222 * 2) If the other side reaches a follower, if will be converted to
1223 * a leader. /This must be done after the race is over, else the
1224 * edges we are iterating on are renumbered./
1225 * 3) /New leader might end up on both sides./
1226 * 4) /If one side ends up with new Leaders, we must ensure that
1227 * they can split out by opcode, hence we have to put _every_
1228 * partition with new Leader nodes on the cprop list, as
1229 * opcode splitting is done by split_by() at the end of
1230 * constant propagation./
1233 if (step(&senv[0])) {
1237 if (step(&senv[1])) {
1242 assert(senv[winner].initial == NULL);
1243 assert(senv[winner].unwalked == NULL);
1245 /* clear flags from walked/unwalked */
1247 transitions = clear_flags(senv[0].unwalked) << shf;
1248 transitions |= clear_flags(senv[0].walked) << shf;
1250 transitions |= clear_flags(senv[1].unwalked) << shf;
1251 transitions |= clear_flags(senv[1].walked) << shf;
1253 dump_race_list("winner ", senv[winner].walked);
1255 /* Move walked_{winner} to a new partition, X'. */
1256 X_prime = new_partition(env);
1259 for (node_t *node = senv[winner].walked; node != NULL; node = node->race_next) {
1260 list_del(&node->node_list);
1261 node->part = X_prime;
1262 if (node->is_follower) {
1263 list_add_tail(&node->node_list, &X_prime->Follower);
1265 list_add_tail(&node->node_list, &X_prime->Leader);
1268 if (node->max_user_input > max_input)
1269 max_input = node->max_user_input;
1271 X_prime->n_leader = n;
1272 X_prime->max_user_inputs = max_input;
1273 X->n_leader -= X_prime->n_leader;
1275 /* for now, copy the type info tag, it will be adjusted in split_by(). */
1276 X_prime->type_is_T_or_C = X->type_is_T_or_C;
1279 * Even if a follower was not checked by both sides, it might have
1280 * loose its congruence, so we need to check this case for all follower.
1282 list_for_each_entry_safe(node_t, node, t, &X_prime->Follower, node_list) {
1283 if (identity(node) == node) {
1284 follower_to_leader(node);
1290 check_partition(X_prime);
1292 dump_partition("Now ", X);
1293 dump_partition("Created new ", X_prime);
1295 /* X' is the smaller part */
1296 add_to_worklist(X_prime, env);
1299 * If there where follower to leader transitions, ensure that the nodes
1300 * can be split out if necessary.
1302 if (transitions & 1) {
1303 /* place winner partition on the cprop list */
1304 if (X_prime->on_cprop == 0) {
1305 X_prime->cprop_next = env->cprop;
1306 env->cprop = X_prime;
1307 X_prime->on_cprop = 1;
1310 if (transitions & 2) {
1311 /* place other partition on the cprop list */
1312 if (X->on_cprop == 0) {
1313 X->cprop_next = env->cprop;
1319 /* we have to ensure that the partition containing g is returned */
1329 * Returns non-zero if the i'th input of a Phi node is live.
1331 * @param phi a Phi-node
1332 * @param i an input number
1334 * @return non-zero if the i'th input of the given Phi node is live
1336 static int is_live_input(ir_node *phi, int i)
1339 ir_node *block = get_nodes_block(phi);
1340 ir_node *pred = get_Block_cfgpred(block, i);
1341 lattice_elem_t type = get_node_type(pred);
1343 return type.tv != tarval_unreachable;
1345 /* else it's the control input, always live */
1347 } /* is_live_input */
1350 * Return non-zero if a type is a constant.
1352 static int is_constant_type(lattice_elem_t type)
1354 if (type.tv != tarval_bottom && type.tv != tarval_top)
1357 } /* is_constant_type */
1360 * Check whether a type is neither Top or a constant.
1361 * Note: U is handled like Top here, R is a constant.
1363 * @param type the type to check
1365 static int type_is_neither_top_nor_const(const lattice_elem_t type)
1367 if (is_tarval(type.tv)) {
1368 if (type.tv == tarval_top)
1370 if (tarval_is_constant(type.tv))
1377 } /* type_is_neither_top_nor_const */
1380 * Collect nodes to the touched list.
1382 * @param list the list which contains the nodes that must be evaluated
1383 * @param idx the index of the def_use edge to evaluate
1384 * @param env the environment
1386 static void collect_touched(list_head *list, int idx, environment_t *env)
1389 int end_idx = env->end_idx;
1391 list_for_each_entry(node_t, x, list, node_list) {
1393 /* leader edges start AFTER follower edges */
1394 x->next_edge = x->n_followers;
1396 unsigned num_edges = get_irn_n_outs(x->node);
1398 /* for all edges in x.L.def_use_{idx} */
1399 while (x->next_edge < num_edges) {
1400 const ir_def_use_edge *edge = &x->node->o.out->edges[x->next_edge];
1403 /* check if we have necessary edges */
1404 if (edge->pos > idx)
1411 /* only non-commutative nodes */
1412 if (env->commutative &&
1413 (idx == 0 || idx == 1) && is_op_commutative(get_irn_op(succ)))
1416 /* ignore the "control input" for non-pinned nodes
1417 if we are running in GCSE mode */
1418 if (idx < end_idx && get_irn_pinned(succ) != op_pin_state_pinned)
1421 y = get_irn_node(succ);
1422 assert(get_irn_n(succ, idx) == x->node);
1424 /* ignore block edges touching followers */
1425 if (idx == -1 && y->is_follower)
1428 if (is_constant_type(y->type)) {
1429 unsigned code = get_irn_opcode(succ);
1430 if (code == iro_Sub || code == iro_Cmp)
1431 add_to_cprop(y, env);
1434 /* Partitions of constants should not be split simply because their Nodes have unequal
1435 functions or incongruent inputs. */
1436 if (type_is_neither_top_nor_const(y->type) &&
1437 (! is_Phi(y->node) || is_live_input(y->node, idx))) {
1438 add_to_touched(y, env);
1442 } /* collect_touched */
1445 * Collect commutative nodes to the touched list.
1447 * @param list the list which contains the nodes that must be evaluated
1448 * @param env the environment
1450 static void collect_commutative_touched(list_head *list, environment_t *env)
1454 list_for_each_entry(node_t, x, list, node_list) {
1455 unsigned num_edges = get_irn_n_outs(x->node);
1457 x->next_edge = x->n_followers;
1459 /* for all edges in x.L.def_use_{idx} */
1460 while (x->next_edge < num_edges) {
1461 const ir_def_use_edge *edge = &x->node->o.out->edges[x->next_edge];
1464 /* check if we have necessary edges */
1474 /* only commutative nodes */
1475 if (!is_op_commutative(get_irn_op(succ)))
1478 y = get_irn_node(succ);
1479 if (is_constant_type(y->type)) {
1480 unsigned code = get_irn_opcode(succ);
1481 if (code == iro_Eor)
1482 add_to_cprop(y, env);
1485 /* Partitions of constants should not be split simply because their Nodes have unequal
1486 functions or incongruent inputs. */
1487 if (type_is_neither_top_nor_const(y->type)) {
1488 add_to_touched(y, env);
1492 } /* collect_commutative_touched */
1495 * Split the partitions if caused by the first entry on the worklist.
1497 * @param env the environment
1499 static void cause_splits(environment_t *env)
1501 partition_t *X, *Z, *N;
1504 /* remove the first partition from the worklist */
1506 env->worklist = X->wl_next;
1509 dump_partition("Cause_split: ", X);
1511 if (env->commutative) {
1512 /* handle commutative nodes first */
1514 /* empty the touched set: already done, just clear the list */
1515 env->touched = NULL;
1517 collect_commutative_touched(&X->Leader, env);
1518 collect_commutative_touched(&X->Follower, env);
1520 for (Z = env->touched; Z != NULL; Z = N) {
1522 node_t *touched = Z->touched;
1523 node_t *touched_aa = NULL;
1524 node_t *touched_ab = NULL;
1525 unsigned n_touched_aa = 0;
1526 unsigned n_touched_ab = 0;
1528 assert(Z->touched != NULL);
1530 /* beware, split might change Z */
1531 N = Z->touched_next;
1533 /* remove it from the touched set */
1536 /* Empty local Z.touched. */
1537 for (e = touched; e != NULL; e = n) {
1538 node_t *left = get_irn_node(get_irn_n(e->node, 0));
1539 node_t *right = get_irn_node(get_irn_n(e->node, 1));
1541 assert(e->is_follower == 0);
1546 * Note: op(a, a) is NOT congruent to op(a, b).
1547 * So, we must split the touched list.
1549 if (left->part == right->part) {
1550 e->next = touched_aa;
1554 e->next = touched_ab;
1559 assert(n_touched_aa + n_touched_ab == Z->n_touched);
1563 if (0 < n_touched_aa && n_touched_aa < Z->n_leader) {
1564 partition_t *Z_prime = Z;
1565 DB((dbg, LEVEL_2, "Split part%d by touched_aa\n", Z_prime->nr));
1566 split(&Z_prime, touched_aa, env);
1568 assert(n_touched_aa <= Z->n_leader);
1570 if (0 < n_touched_ab && n_touched_ab < Z->n_leader) {
1571 partition_t *Z_prime = Z;
1572 DB((dbg, LEVEL_2, "Split part%d by touched_ab\n", Z_prime->nr));
1573 split(&Z_prime, touched_ab, env);
1575 assert(n_touched_ab <= Z->n_leader);
1579 /* combine temporary leader and follower list */
1580 for (idx = -1; idx <= X->max_user_inputs; ++idx) {
1581 /* empty the touched set: already done, just clear the list */
1582 env->touched = NULL;
1584 collect_touched(&X->Leader, idx, env);
1585 collect_touched(&X->Follower, idx, env);
1587 for (Z = env->touched; Z != NULL; Z = N) {
1589 node_t *touched = Z->touched;
1590 unsigned n_touched = Z->n_touched;
1592 assert(Z->touched != NULL);
1594 /* beware, split might change Z */
1595 N = Z->touched_next;
1597 /* remove it from the touched set */
1600 /* Empty local Z.touched. */
1601 for (e = touched; e != NULL; e = e->next) {
1602 assert(e->is_follower == 0);
1608 if (0 < n_touched && n_touched < Z->n_leader) {
1609 DB((dbg, LEVEL_2, "Split part%d by touched\n", Z->nr));
1610 split(&Z, touched, env);
1612 assert(n_touched <= Z->n_leader);
1615 } /* cause_splits */
1618 * Implements split_by_what(): Split a partition by characteristics given
1619 * by the what function.
1621 * @param X the partition to split
1622 * @param What a function returning an Id for every node of the partition X
1623 * @param P a list to store the result partitions
1624 * @param env the environment
1628 static partition_t *split_by_what(partition_t *X, what_func What,
1629 partition_t **P, environment_t *env)
1633 listmap_entry_t *iter;
1636 /* Let map be an empty mapping from the range of What to (local) list of Nodes. */
1638 list_for_each_entry(node_t, x, &X->Leader, node_list) {
1639 void *id = What(x, env);
1640 listmap_entry_t *entry;
1643 /* input not allowed, ignore */
1646 /* Add x to map[What(x)]. */
1647 entry = listmap_find(&map, id);
1648 x->next = entry->list;
1651 /* Let P be a set of Partitions. */
1653 /* for all sets S except one in the range of map do */
1654 for (iter = map.values; iter != NULL; iter = iter->next) {
1655 if (iter->next == NULL) {
1656 /* this is the last entry, ignore */
1661 /* Add SPLIT( X, S ) to P. */
1662 DB((dbg, LEVEL_2, "Split part%d by WHAT = %s\n", X->nr, what_reason));
1663 R = split(&X, S, env);
1673 } /* split_by_what */
1675 /** lambda n.(n.type) */
1676 static void *lambda_type(const node_t *node, environment_t *env)
1679 return node->type.tv;
1682 /** lambda n.(n.opcode) */
1683 static void *lambda_opcode(const node_t *node, environment_t *env)
1685 opcode_key_t key, *entry;
1687 key.irn = node->node;
1689 entry = set_insert(opcode_key_t, env->opcode2id_map, &key, sizeof(key), opcode_hash(&key));
1691 } /* lambda_opcode */
1693 /** lambda n.(n[i].partition) */
1694 static void *lambda_partition(const node_t *node, environment_t *env)
1696 ir_node *skipped = skip_Proj(node->node);
1699 int i = env->lambda_input;
1701 if (i >= get_irn_arity(node->node)) {
1703 * We are outside the allowed range: This can happen even
1704 * if we have split by opcode first: doing so might move Followers
1705 * to Leaders and those will have a different opcode!
1706 * Note that in this case the partition is on the cprop list and will be
1712 /* ignore the "control input" for non-pinned nodes
1713 if we are running in GCSE mode */
1714 if (i < env->end_idx && get_irn_pinned(skipped) != op_pin_state_pinned)
1717 pred = i == -1 ? get_irn_n(skipped, i) : get_irn_n(node->node, i);
1718 p = get_irn_node(pred);
1720 } /* lambda_partition */
1722 /** lambda n.(n[i].partition) for commutative nodes */
1723 static void *lambda_commutative_partition(const node_t *node, environment_t *env)
1725 ir_node *irn = node->node;
1726 ir_node *skipped = skip_Proj(irn);
1727 ir_node *pred, *left, *right;
1729 partition_t *pl, *pr;
1730 int i = env->lambda_input;
1732 if (i >= get_irn_arity(node->node)) {
1734 * We are outside the allowed range: This can happen even
1735 * if we have split by opcode first: doing so might move Followers
1736 * to Leaders and those will have a different opcode!
1737 * Note that in this case the partition is on the cprop list and will be
1743 /* ignore the "control input" for non-pinned nodes
1744 if we are running in GCSE mode */
1745 if (i < env->end_idx && get_irn_pinned(skipped) != op_pin_state_pinned)
1749 pred = get_irn_n(skipped, i);
1750 p = get_irn_node(pred);
1754 if (is_op_commutative(get_irn_op(irn))) {
1755 /* normalize partition order by returning the "smaller" on input 0,
1756 the "bigger" on input 1. */
1757 left = get_binop_left(irn);
1758 pl = get_irn_node(left)->part;
1759 right = get_binop_right(irn);
1760 pr = get_irn_node(right)->part;
1763 return pl < pr ? pl : pr;
1765 return pl > pr ? pl : pr;
1767 /* a not split out Follower */
1768 pred = get_irn_n(irn, i);
1769 p = get_irn_node(pred);
1773 } /* lambda_commutative_partition */
1776 * Returns true if a type is a constant (and NOT Top
1779 static int is_con(const lattice_elem_t type)
1781 /* be conservative */
1782 if (is_tarval(type.tv))
1783 return tarval_is_constant(type.tv);
1784 return is_entity(type.sym.entity_p);
1788 * Implements split_by().
1790 * @param X the partition to split
1791 * @param env the environment
1793 static void split_by(partition_t *X, environment_t *env)
1795 partition_t *I, *P = NULL;
1798 dump_partition("split_by", X);
1800 if (X->n_leader == 1) {
1801 /* we have only one leader, no need to split, just check its type */
1802 node_t *x = get_first_node(X);
1803 X->type_is_T_or_C = x->type.tv == tarval_top || is_con(x->type);
1807 DEBUG_ONLY(what_reason = "lambda n.(n.type)";)
1808 P = split_by_what(X, lambda_type, &P, env);
1811 /* adjust the type tags, we have split partitions by type */
1812 for (I = P; I != NULL; I = I->split_next) {
1813 node_t *x = get_first_node(I);
1814 I->type_is_T_or_C = x->type.tv == tarval_top || is_con(x->type);
1821 if (Y->n_leader > 1) {
1822 /* we do not want split the TOP or constant partitions */
1823 if (! Y->type_is_T_or_C) {
1824 partition_t *Q = NULL;
1826 DEBUG_ONLY(what_reason = "lambda n.(n.opcode)";)
1827 Q = split_by_what(Y, lambda_opcode, &Q, env);
1834 if (Z->n_leader > 1) {
1835 const node_t *first = get_first_node(Z);
1836 int arity = get_irn_arity(first->node);
1838 what_func what = lambda_partition;
1839 DEBUG_ONLY(char buf[64];)
1841 if (env->commutative && is_op_commutative(get_irn_op(first->node)))
1842 what = lambda_commutative_partition;
1845 * BEWARE: during splitting by input 2 for instance we might
1846 * create new partitions which are different by input 1, so collect
1847 * them and split further.
1849 Z->split_next = NULL;
1852 for (input = arity - 1; input >= -1; --input) {
1854 partition_t *Z_prime = R;
1857 if (Z_prime->n_leader > 1) {
1858 env->lambda_input = input;
1859 DEBUG_ONLY(snprintf(buf, sizeof(buf), "lambda n.(n[%d].partition)", input);)
1860 DEBUG_ONLY(what_reason = buf;)
1861 S = split_by_what(Z_prime, what, &S, env);
1864 Z_prime->split_next = S;
1867 } while (R != NULL);
1872 } while (Q != NULL);
1875 } while (P != NULL);
1879 * (Re-)compute the type for a given node.
1881 * @param node the node
1883 static void default_compute(node_t *node)
1886 ir_node *irn = node->node;
1888 /* if any of the data inputs have type top, the result is type top */
1889 for (i = get_irn_arity(irn) - 1; i >= 0; --i) {
1890 ir_node *pred = get_irn_n(irn, i);
1891 node_t *p = get_irn_node(pred);
1893 if (p->type.tv == tarval_top) {
1894 node->type.tv = tarval_top;
1899 if (get_irn_mode(node->node) == mode_X)
1900 node->type.tv = tarval_reachable;
1902 node->type.tv = computed_value(irn);
1903 } /* default_compute */
1906 * (Re-)compute the type for a Block node.
1908 * @param node the node
1910 static void compute_Block(node_t *node)
1913 ir_node *block = node->node;
1915 if (block == get_irg_start_block(current_ir_graph) || get_Block_entity(block) != NULL) {
1916 /* start block and labelled blocks are always reachable */
1917 node->type.tv = tarval_reachable;
1921 for (i = get_Block_n_cfgpreds(block) - 1; i >= 0; --i) {
1922 node_t *pred = get_irn_node(get_Block_cfgpred(block, i));
1924 if (pred->type.tv == tarval_reachable) {
1925 /* A block is reachable, if at least of predecessor is reachable. */
1926 node->type.tv = tarval_reachable;
1930 node->type.tv = tarval_top;
1931 } /* compute_Block */
1934 * (Re-)compute the type for a Bad node.
1936 * @param node the node
1938 static void compute_Bad(node_t *node)
1940 /* Bad nodes ALWAYS compute Top */
1941 node->type.tv = tarval_top;
1945 * (Re-)compute the type for an Unknown node.
1947 * @param node the node
1949 static void compute_Unknown(node_t *node)
1951 /* While Unknown nodes should compute Top this is dangerous:
1952 * a Top input to a Cond would lead to BOTH control flows unreachable.
1953 * While this is correct in the given semantics, it would destroy the Firm
1956 * It would be safe to compute Top IF it can be assured, that only Cmp
1957 * nodes are inputs to Conds. We check that first.
1958 * This is the way Frontends typically build Firm, but some optimizations
1959 * (jump threading for instance) might replace them by Phib's...
1961 node->type.tv = tarval_UNKNOWN;
1962 } /* compute_Unknown */
1965 * (Re-)compute the type for a Jmp node.
1967 * @param node the node
1969 static void compute_Jmp(node_t *node)
1971 node_t *block = get_irn_node(get_nodes_block(node->node));
1973 node->type = block->type;
1977 * (Re-)compute the type for the Return node.
1979 * @param node the node
1981 static void compute_Return(node_t *node)
1983 /* The Return node is NOT dead if it is in a reachable block.
1984 * This is already checked in compute(). so we can return
1985 * Reachable here. */
1986 node->type.tv = tarval_reachable;
1987 } /* compute_Return */
1990 * (Re-)compute the type for the End node.
1992 * @param node the node
1994 static void compute_End(node_t *node)
1996 /* the End node is NOT dead of course */
1997 node->type.tv = tarval_reachable;
2001 * (Re-)compute the type for a Call.
2003 * @param node the node
2005 static void compute_Call(node_t *node)
2008 * A Call computes always bottom, even if it has Unknown
2011 node->type.tv = tarval_bottom;
2012 } /* compute_Call */
2015 * (Re-)compute the type for a SymConst node.
2017 * @param node the node
2019 static void compute_SymConst(node_t *node)
2021 ir_node *irn = node->node;
2022 node_t *block = get_irn_node(get_nodes_block(irn));
2024 if (block->type.tv == tarval_unreachable) {
2025 node->type.tv = tarval_top;
2028 switch (get_SymConst_kind(irn)) {
2029 case symconst_addr_ent:
2030 node->type.sym = get_SymConst_symbol(irn);
2033 node->type.tv = computed_value(irn);
2035 } /* compute_SymConst */
2038 * (Re-)compute the type for a Phi node.
2040 * @param node the node
2042 static void compute_Phi(node_t *node)
2045 ir_node *phi = node->node;
2046 lattice_elem_t type;
2048 /* if a Phi is in a unreachable block, its type is TOP */
2049 node_t *block = get_irn_node(get_nodes_block(phi));
2051 if (block->type.tv == tarval_unreachable) {
2052 node->type.tv = tarval_top;
2056 /* Phi implements the Meet operation */
2057 type.tv = tarval_top;
2058 for (i = get_Phi_n_preds(phi) - 1; i >= 0; --i) {
2059 node_t *pred = get_irn_node(get_Phi_pred(phi, i));
2060 node_t *pred_X = get_irn_node(get_Block_cfgpred(block->node, i));
2062 if (pred_X->type.tv == tarval_unreachable || pred->type.tv == tarval_top) {
2063 /* ignore TOP inputs: We must check here for unreachable blocks,
2064 because Firm constants live in the Start Block are NEVER Top.
2065 Else, a Phi (1,2) will produce Bottom, even if the 2 for instance
2066 comes from a unreachable input. */
2069 if (pred->type.tv == tarval_bottom) {
2070 node->type.tv = tarval_bottom;
2072 } else if (type.tv == tarval_top) {
2073 /* first constant found */
2075 } else if (type.tv != pred->type.tv) {
2076 /* different constants or tarval_bottom */
2077 node->type.tv = tarval_bottom;
2080 /* else nothing, constants are the same */
2086 * (Re-)compute the type for an Add. Special case: one nodes is a Zero Const.
2088 * @param node the node
2090 static void compute_Add(node_t *node)
2092 ir_node *sub = node->node;
2093 node_t *l = get_irn_node(get_Add_left(sub));
2094 node_t *r = get_irn_node(get_Add_right(sub));
2095 lattice_elem_t a = l->type;
2096 lattice_elem_t b = r->type;
2099 if (a.tv == tarval_top || b.tv == tarval_top) {
2100 node->type.tv = tarval_top;
2101 } else if (a.tv == tarval_bottom || b.tv == tarval_bottom) {
2102 node->type.tv = tarval_bottom;
2104 /* x + 0 = 0 + x = x, but beware of floating point +0 + -0, so we
2105 must call tarval_add() first to handle this case! */
2106 if (is_tarval(a.tv)) {
2107 if (is_tarval(b.tv)) {
2108 node->type.tv = tarval_add(a.tv, b.tv);
2111 mode = get_tarval_mode(a.tv);
2112 if (a.tv == get_mode_null(mode)) {
2116 } else if (is_tarval(b.tv)) {
2117 mode = get_tarval_mode(b.tv);
2118 if (b.tv == get_mode_null(mode)) {
2123 node->type.tv = tarval_bottom;
2128 * (Re-)compute the type for a Sub. Special case: both nodes are congruent.
2130 * @param node the node
2132 static void compute_Sub(node_t *node)
2134 ir_node *sub = node->node;
2135 node_t *l = get_irn_node(get_Sub_left(sub));
2136 node_t *r = get_irn_node(get_Sub_right(sub));
2137 lattice_elem_t a = l->type;
2138 lattice_elem_t b = r->type;
2141 if (a.tv == tarval_top || b.tv == tarval_top) {
2142 node->type.tv = tarval_top;
2143 } else if (is_con(a) && is_con(b)) {
2144 if (is_tarval(a.tv) && is_tarval(b.tv)) {
2145 node->type.tv = tarval_sub(a.tv, b.tv, get_irn_mode(sub));
2146 } else if (is_tarval(a.tv) && tarval_is_null(a.tv)) {
2148 } else if (is_tarval(b.tv) && tarval_is_null(b.tv)) {
2151 node->type.tv = tarval_bottom;
2153 } else if (r->part == l->part &&
2154 (!mode_is_float(get_irn_mode(l->node)))) {
2156 * BEWARE: a - a is NOT always 0 for floating Point values, as
2157 * NaN op NaN = NaN, so we must check this here.
2159 ir_mode *mode = get_irn_mode(sub);
2160 tv = get_mode_null(mode);
2162 /* if the node was ONCE evaluated by all constants, but now
2163 this breaks AND we get from the argument partitions a different
2164 result, switch to bottom.
2165 This happens because initially all nodes are in the same partition ... */
2166 if (node->type.tv != tv)
2170 node->type.tv = tarval_bottom;
2175 * (Re-)compute the type for an Eor. Special case: both nodes are congruent.
2177 * @param node the node
2179 static void compute_Eor(node_t *node)
2181 ir_node *eor = node->node;
2182 node_t *l = get_irn_node(get_Eor_left(eor));
2183 node_t *r = get_irn_node(get_Eor_right(eor));
2184 lattice_elem_t a = l->type;
2185 lattice_elem_t b = r->type;
2188 if (a.tv == tarval_top || b.tv == tarval_top) {
2189 node->type.tv = tarval_top;
2190 } else if (is_con(a) && is_con(b)) {
2191 if (is_tarval(a.tv) && is_tarval(b.tv)) {
2192 node->type.tv = tarval_eor(a.tv, b.tv);
2193 } else if (is_tarval(a.tv) && tarval_is_null(a.tv)) {
2195 } else if (is_tarval(b.tv) && tarval_is_null(b.tv)) {
2198 node->type.tv = tarval_bottom;
2200 } else if (r->part == l->part) {
2201 ir_mode *mode = get_irn_mode(eor);
2202 tv = get_mode_null(mode);
2204 /* if the node was ONCE evaluated by all constants, but now
2205 this breaks AND we get from the argument partitions a different
2206 result, switch to bottom.
2207 This happens because initially all nodes are in the same partition ... */
2208 if (node->type.tv != tv)
2212 node->type.tv = tarval_bottom;
2217 * (Re-)compute the type for Cmp.
2219 * @param node the node
2221 static void compute_Cmp(node_t *node)
2223 ir_node *cmp = node->node;
2224 node_t *l = get_irn_node(get_Cmp_left(cmp));
2225 node_t *r = get_irn_node(get_Cmp_right(cmp));
2226 lattice_elem_t a = l->type;
2227 lattice_elem_t b = r->type;
2228 ir_relation relation = get_Cmp_relation(cmp);
2231 if (a.tv == tarval_top || b.tv == tarval_top) {
2232 node->type.tv = tarval_undefined;
2233 } else if (is_con(a) && is_con(b)) {
2234 default_compute(node);
2237 * BEWARE: a == a is NOT always True for floating Point values, as
2238 * NaN != NaN is defined, so we must check this here.
2239 * (while for some pnc we could still optimize we have to stay
2240 * consistent with compute_Cmp, so don't do anything for floats)
2242 } else if (r->part == l->part && !mode_is_float(get_irn_mode(l->node))) {
2243 tv = relation & ir_relation_equal ? tarval_b_true : tarval_b_false;
2245 /* if the node was ONCE evaluated to a constant, but now
2246 this breaks AND we get from the argument partitions a different
2247 result, ensure monotony by fall to bottom.
2248 This happens because initially all nodes are in the same partition ... */
2249 if (node->type.tv == tarval_bottom)
2251 else if (node->type.tv != tv && is_constant_type(node->type))
2255 node->type.tv = tarval_bottom;
2260 * (Re-)compute the type for a Proj(Cond).
2262 * @param node the node
2263 * @param cond the predecessor Cond node
2265 static void compute_Proj_Cond(node_t *node, ir_node *cond)
2267 ir_node *proj = node->node;
2268 long pnc = get_Proj_proj(proj);
2269 ir_node *sel = get_Cond_selector(cond);
2270 node_t *selector = get_irn_node(sel);
2273 * Note: it is crucial for the monotony that the Proj(Cond)
2274 * are evaluates after all predecessors of the Cond selector are
2280 * Due to the fact that 0 is a const, the Cmp gets immediately
2281 * on the cprop list. It will be evaluated before x is evaluated,
2282 * might leaving x as Top. When later x is evaluated, the Cmp
2283 * might change its value.
2284 * BUT if the Cond is evaluated before this happens, Proj(Cond, FALSE)
2285 * gets R, and later changed to F if Cmp is evaluated to True!
2287 * We prevent this by putting Conds in an extra cprop_X queue, which
2288 * gets evaluated after the cprop queue is empty.
2290 * Note that this even happens with Click's original algorithm, if
2291 * Cmp(x, 0) is evaluated to True first and later changed to False
2292 * if x was Top first and later changed to a Const ...
2293 * It is unclear how Click solved that problem ...
2295 * However, in rare cases even this does not help, if a Top reaches
2296 * a compare through a Phi, than Proj(Cond) is evaluated changing
2297 * the type of the Phi to something other.
2298 * So, we take the last resort and bind the type to R once
2301 * (This might be even the way Click works around the whole problem).
2303 * Finally, we may miss some optimization possibilities due to this:
2308 * If Top reaches the if first, than we decide for != here.
2309 * If y later is evaluated to 0, we cannot revert this decision
2310 * and must live with both outputs enabled. If this happens,
2311 * we get an unresolved if (true) in the code ...
2313 * In Click's version where this decision is done at the Cmp,
2314 * the Cmp is NOT optimized away than (if y evaluated to 1
2315 * for instance) and we get a if (1 == 0) here ...
2317 * Both solutions are suboptimal.
2318 * At least, we could easily detect this problem and run
2319 * cf_opt() (or even combo) again :-(
2321 if (node->type.tv == tarval_reachable)
2324 if (pnc == pn_Cond_true) {
2325 if (selector->type.tv == tarval_b_false) {
2326 node->type.tv = tarval_unreachable;
2327 } else if (selector->type.tv == tarval_b_true) {
2328 node->type.tv = tarval_reachable;
2329 } else if (selector->type.tv == tarval_bottom) {
2330 node->type.tv = tarval_reachable;
2332 assert(selector->type.tv == tarval_top);
2333 if (tarval_UNKNOWN == tarval_top) {
2334 /* any condition based on Top is "!=" */
2335 node->type.tv = tarval_unreachable;
2337 node->type.tv = tarval_unreachable;
2341 assert(pnc == pn_Cond_false);
2343 if (selector->type.tv == tarval_b_false) {
2344 node->type.tv = tarval_reachable;
2345 } else if (selector->type.tv == tarval_b_true) {
2346 node->type.tv = tarval_unreachable;
2347 } else if (selector->type.tv == tarval_bottom) {
2348 node->type.tv = tarval_reachable;
2350 assert(selector->type.tv == tarval_top);
2351 if (tarval_UNKNOWN == tarval_top) {
2352 /* any condition based on Top is "!=" */
2353 node->type.tv = tarval_reachable;
2355 node->type.tv = tarval_unreachable;
2359 } /* compute_Proj_Cond */
2361 static void compute_Proj_Switch(node_t *node, ir_node *switchn)
2363 ir_node *proj = node->node;
2364 long pnc = get_Proj_proj(proj);
2365 ir_node *sel = get_Switch_selector(switchn);
2366 node_t *selector = get_irn_node(sel);
2368 /* see long comment in compute_Proj_Cond */
2369 if (node->type.tv == tarval_reachable)
2372 if (selector->type.tv == tarval_bottom) {
2373 node->type.tv = tarval_reachable;
2374 } else if (selector->type.tv == tarval_top) {
2375 if (tarval_UNKNOWN == tarval_top && pnc == pn_Switch_default) {
2376 /* a switch based of Top is always "default" */
2377 node->type.tv = tarval_reachable;
2379 node->type.tv = tarval_unreachable;
2382 long value = get_tarval_long(selector->type.tv);
2383 const ir_switch_table *table = get_Switch_table(switchn);
2384 size_t n_entries = ir_switch_table_get_n_entries(table);
2387 for (e = 0; e < n_entries; ++e) {
2388 const ir_switch_table_entry *entry
2389 = ir_switch_table_get_entry_const(table, e);
2390 ir_tarval *min = entry->min;
2391 ir_tarval *max = entry->max;
2393 if (selector->type.tv == min) {
2394 node->type.tv = entry->pn == pnc
2395 ? tarval_reachable : tarval_unreachable;
2399 long minval = get_tarval_long(min);
2400 long maxval = get_tarval_long(max);
2401 if (minval <= value && value <= maxval) {
2402 node->type.tv = entry->pn == pnc
2403 ? tarval_reachable : tarval_unreachable;
2409 /* no entry matched: default */
2411 = pnc == pn_Switch_default ? tarval_reachable : tarval_unreachable;
2416 * (Re-)compute the type for a Proj-Node.
2418 * @param node the node
2420 static void compute_Proj(node_t *node)
2422 ir_node *proj = node->node;
2423 ir_mode *mode = get_irn_mode(proj);
2424 node_t *block = get_irn_node(get_nodes_block(skip_Proj(proj)));
2425 ir_node *pred = get_Proj_pred(proj);
2427 if (block->type.tv == tarval_unreachable) {
2428 /* a Proj in a unreachable Block stay Top */
2429 node->type.tv = tarval_top;
2432 if (get_irn_node(pred)->type.tv == tarval_top && !is_Cond(pred) && !is_Switch(pred)) {
2433 /* if the predecessor is Top, its Proj follow */
2434 node->type.tv = tarval_top;
2438 if (mode == mode_M) {
2439 /* mode M is always bottom */
2440 node->type.tv = tarval_bottom;
2442 } else if (mode == mode_X) {
2443 /* handle mode_X nodes */
2444 switch (get_irn_opcode(pred)) {
2446 /* the Proj_X from the Start is always reachable.
2447 However this is already handled at the top. */
2448 node->type.tv = tarval_reachable;
2451 compute_Proj_Cond(node, pred);
2454 compute_Proj_Switch(node, pred);
2461 default_compute(node);
2462 } /* compute_Proj */
2465 * (Re-)compute the type for a Confirm.
2467 * @param node the node
2469 static void compute_Confirm(node_t *node)
2471 ir_node *confirm = node->node;
2472 node_t *pred = get_irn_node(get_Confirm_value(confirm));
2474 if (get_Confirm_relation(confirm) == ir_relation_equal) {
2475 node_t *bound = get_irn_node(get_Confirm_bound(confirm));
2477 if (is_con(bound->type)) {
2478 /* is equal to a constant */
2479 node->type = bound->type;
2483 /* a Confirm is a copy OR a Const */
2484 node->type = pred->type;
2485 } /* compute_Confirm */
2488 * (Re-)compute the type for a given node.
2490 * @param node the node
2492 static void compute(node_t *node)
2494 ir_node *irn = node->node;
2497 #ifndef VERIFY_MONOTONE
2499 * Once a node reaches bottom, the type cannot fall further
2500 * in the lattice and we can stop computation.
2501 * Do not take this exit if the monotony verifier is
2502 * enabled to catch errors.
2504 if (node->type.tv == tarval_bottom)
2508 if (!is_Block(irn)) {
2509 /* for pinned nodes, check its control input */
2510 if (get_irn_pinned(skip_Proj(irn)) == op_pin_state_pinned) {
2511 node_t *block = get_irn_node(get_nodes_block(irn));
2513 if (block->type.tv == tarval_unreachable) {
2514 node->type.tv = tarval_top;
2520 func = (compute_func)node->node->op->ops.generic;
2526 * Identity functions: Note that one might think that identity() is just a
2527 * synonym for equivalent_node(). While this is true, we cannot use it for the algorithm
2528 * here, because it expects that the identity node is one of the inputs, which is NOT
2529 * always true for equivalent_node() which can handle (and does sometimes) DAGs.
2530 * So, we have our own implementation, which copies some parts of equivalent_node()
2534 * Calculates the Identity for Phi nodes
2536 static node_t *identity_Phi(node_t *node)
2538 ir_node *phi = node->node;
2539 ir_node *block = get_nodes_block(phi);
2540 node_t *n_part = NULL;
2543 for (i = get_Phi_n_preds(phi) - 1; i >= 0; --i) {
2544 node_t *pred_X = get_irn_node(get_Block_cfgpred(block, i));
2546 if (pred_X->type.tv == tarval_reachable) {
2547 node_t *pred = get_irn_node(get_Phi_pred(phi, i));
2551 else if (n_part->part != pred->part) {
2552 /* incongruent inputs, not a follower */
2557 /* if n_part is NULL here, all inputs path are dead, the Phi computes
2558 * tarval_top, is in the TOP partition and should NOT being split! */
2559 assert(n_part != NULL);
2561 } /* identity_Phi */
2564 * Calculates the Identity for commutative 0 neutral nodes.
2566 static node_t *identity_comm_zero_binop(node_t *node)
2568 ir_node *op = node->node;
2569 node_t *a = get_irn_node(get_binop_left(op));
2570 node_t *b = get_irn_node(get_binop_right(op));
2571 ir_mode *mode = get_irn_mode(op);
2574 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
2575 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
2578 /* node: no input should be tarval_top, else the binop would be also
2579 * Top and not being split. */
2580 zero = get_mode_null(mode);
2581 if (a->type.tv == zero)
2583 if (b->type.tv == zero)
2586 } /* identity_comm_zero_binop */
2589 * Calculates the Identity for Shift nodes.
2591 static node_t *identity_shift(node_t *node)
2593 ir_node *op = node->node;
2594 node_t *b = get_irn_node(get_binop_right(op));
2595 ir_mode *mode = get_irn_mode(b->node);
2598 /* node: no input should be tarval_top, else the binop would be also
2599 * Top and not being split. */
2600 zero = get_mode_null(mode);
2601 if (b->type.tv == zero)
2602 return get_irn_node(get_binop_left(op));
2604 } /* identity_shift */
2607 * Calculates the Identity for Mul nodes.
2609 static node_t *identity_Mul(node_t *node)
2611 ir_node *op = node->node;
2612 node_t *a = get_irn_node(get_Mul_left(op));
2613 node_t *b = get_irn_node(get_Mul_right(op));
2614 ir_mode *mode = get_irn_mode(op);
2617 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
2618 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
2621 /* node: no input should be tarval_top, else the binop would be also
2622 * Top and not being split. */
2623 one = get_mode_one(mode);
2624 if (a->type.tv == one)
2626 if (b->type.tv == one)
2629 } /* identity_Mul */
2632 * Calculates the Identity for Sub nodes.
2634 static node_t *identity_Sub(node_t *node)
2636 ir_node *sub = node->node;
2637 node_t *b = get_irn_node(get_Sub_right(sub));
2638 ir_mode *mode = get_irn_mode(sub);
2640 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
2641 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
2644 /* node: no input should be tarval_top, else the binop would be also
2645 * Top and not being split. */
2646 if (b->type.tv == get_mode_null(mode))
2647 return get_irn_node(get_Sub_left(sub));
2649 } /* identity_Sub */
2652 * Calculates the Identity for And nodes.
2654 static node_t *identity_And(node_t *node)
2656 ir_node *andnode = node->node;
2657 node_t *a = get_irn_node(get_And_left(andnode));
2658 node_t *b = get_irn_node(get_And_right(andnode));
2659 ir_tarval *neutral = get_mode_all_one(get_irn_mode(andnode));
2661 /* node: no input should be tarval_top, else the And would be also
2662 * Top and not being split. */
2663 if (a->type.tv == neutral)
2665 if (b->type.tv == neutral)
2668 } /* identity_And */
2671 * Calculates the Identity for Confirm nodes.
2673 static node_t *identity_Confirm(node_t *node)
2675 ir_node *confirm = node->node;
2677 /* a Confirm is always a Copy */
2678 return get_irn_node(get_Confirm_value(confirm));
2679 } /* identity_Confirm */
2682 * Calculates the Identity for Mux nodes.
2684 static node_t *identity_Mux(node_t *node)
2686 ir_node *mux = node->node;
2687 node_t *t = get_irn_node(get_Mux_true(mux));
2688 node_t *f = get_irn_node(get_Mux_false(mux));
2691 if (t->part == f->part)
2694 /* for now, the 1-input identity is not supported */
2696 sel = get_irn_node(get_Mux_sel(mux));
2698 /* Mux sel input is mode_b, so it is always a tarval */
2699 if (sel->type.tv == tarval_b_true)
2701 if (sel->type.tv == tarval_b_false)
2705 } /* identity_Mux */
2708 * Calculates the Identity for nodes.
2710 static node_t *identity(node_t *node)
2712 ir_node *irn = node->node;
2714 switch (get_irn_opcode(irn)) {
2716 return identity_Phi(node);
2718 return identity_Mul(node);
2722 return identity_comm_zero_binop(node);
2727 return identity_shift(node);
2729 return identity_And(node);
2731 return identity_Sub(node);
2733 return identity_Confirm(node);
2735 return identity_Mux(node);
2742 * Node follower is a (new) follower of leader, segregate Leader
2745 static void segregate_def_use_chain_1(const ir_node *follower, node_t *leader)
2747 DB((dbg, LEVEL_2, "%+F is a follower of %+F\n", follower, leader->node));
2748 /* The leader edges must remain sorted, but follower edges can
2750 ir_node *l = leader->node;
2751 unsigned n = get_irn_n_outs(l);
2752 for (unsigned i = leader->n_followers; i < n; ++i) {
2753 if (l->o.out->edges[i].use == follower) {
2754 ir_def_use_edge t = l->o.out->edges[i];
2756 for (unsigned j = i; j-- > leader->n_followers; )
2757 l->o.out->edges[j+1] = l->o.out->edges[j];
2758 l->o.out->edges[leader->n_followers] = t;
2759 ++leader->n_followers;
2763 } /* segregate_def_use_chain_1 */
2766 * Node follower is a (new) follower segregate its Leader
2769 * @param follower the follower IR node
2771 static void segregate_def_use_chain(const ir_node *follower)
2775 for (i = get_irn_arity(follower) - 1; i >= 0; --i) {
2776 node_t *pred = get_irn_node(get_irn_n(follower, i));
2778 segregate_def_use_chain_1(follower, pred);
2780 } /* segregate_def_use_chain */
2783 * Propagate constant evaluation.
2785 * @param env the environment
2787 static void propagate(environment_t *env)
2791 lattice_elem_t old_type;
2793 unsigned n_fallen, old_type_was_T_or_C;
2795 while (env->cprop != NULL) {
2796 void *oldopcode = NULL;
2798 /* remove the first partition X from cprop */
2801 env->cprop = X->cprop_next;
2803 old_type_was_T_or_C = X->type_is_T_or_C;
2805 DB((dbg, LEVEL_2, "Propagate type on part%d\n", X->nr));
2809 int cprop_empty = list_empty(&X->cprop);
2810 int cprop_X_empty = list_empty(&X->cprop_X);
2812 if (cprop_empty && cprop_X_empty) {
2813 /* both cprop lists are empty */
2817 /* remove the first Node x from X.cprop */
2819 /* Get a node from the cprop_X list only if
2820 * all data nodes are processed.
2821 * This ensures, that all inputs of the Cond
2822 * predecessor are processed if its type is still Top.
2824 x = list_entry(X->cprop_X.next, node_t, cprop_list);
2826 x = list_entry(X->cprop.next, node_t, cprop_list);
2829 //assert(x->part == X);
2830 list_del(&x->cprop_list);
2833 if (x->is_follower && identity(x) == x) {
2834 /* check the opcode first */
2835 if (oldopcode == NULL) {
2836 oldopcode = lambda_opcode(get_first_node(X), env);
2838 if (oldopcode != lambda_opcode(x, env)) {
2839 if (x->on_fallen == 0) {
2840 /* different opcode -> x falls out of this partition */
2845 DB((dbg, LEVEL_2, "Add node %+F to fallen\n", x->node));
2849 /* x will make the follower -> leader transition */
2850 follower_to_leader(x);
2853 /* compute a new type for x */
2855 DB((dbg, LEVEL_3, "computing type of %+F\n", x->node));
2857 if (x->type.tv != old_type.tv) {
2858 DB((dbg, LEVEL_2, "node %+F has changed type from %+F to %+F\n", x->node, old_type, x->type));
2859 verify_type(old_type, x);
2861 if (x->on_fallen == 0) {
2862 /* Add x to fallen. Nodes might fall from T -> const -> _|_, so check that they are
2863 not already on the list. */
2868 DB((dbg, LEVEL_2, "Add node %+F to fallen\n", x->node));
2870 for (unsigned i = get_irn_n_outs(x->node); i-- > 0; ) {
2871 ir_node *succ = get_irn_out(x->node, i);
2872 node_t *y = get_irn_node(succ);
2874 /* Add y to y.partition.cprop. */
2875 add_to_cprop(y, env);
2880 if (n_fallen > 0 && n_fallen != X->n_leader) {
2881 DB((dbg, LEVEL_2, "Splitting part%d by fallen\n", X->nr));
2882 Y = split(&X, fallen, env);
2884 * We have split out fallen node. The type of the result
2885 * partition is NOT set yet.
2887 Y->type_is_T_or_C = 0;
2891 /* remove the flags from the fallen list */
2892 for (x = fallen; x != NULL; x = x->next)
2895 if (old_type_was_T_or_C) {
2896 /* check if some nodes will make the leader -> follower transition */
2897 list_for_each_entry_safe(node_t, y, tmp, &Y->Leader, node_list) {
2898 if (y->type.tv != tarval_top && ! is_con(y->type)) {
2899 node_t *eq_node = identity(y);
2901 if (eq_node != y && eq_node->part == y->part) {
2902 DB((dbg, LEVEL_2, "Node %+F is a follower of %+F\n", y->node, eq_node->node));
2903 /* move to Follower */
2905 list_del(&y->node_list);
2906 list_add_tail(&y->node_list, &Y->Follower);
2909 segregate_def_use_chain(y->node);
2919 * Get the leader for a given node from its congruence class.
2921 * @param irn the node
2923 static ir_node *get_leader(node_t *node)
2925 partition_t *part = node->part;
2927 if (part->n_leader > 1 || node->is_follower) {
2928 if (node->is_follower) {
2929 DB((dbg, LEVEL_2, "Replacing follower %+F\n", node->node));
2932 DB((dbg, LEVEL_2, "Found congruence class for %+F\n", node->node));
2934 return get_first_node(part)->node;
2940 * Returns non-zero if a mode_T node has only one reachable output.
2942 static int only_one_reachable_proj(ir_node *n)
2946 for (unsigned i = get_irn_n_outs(n); i-- > 0; ) {
2947 ir_node *proj = get_irn_out(n, i);
2950 /* skip non-control flow Proj's */
2951 if (get_irn_mode(proj) != mode_X)
2954 node = get_irn_node(proj);
2955 if (node->type.tv == tarval_reachable) {
2961 } /* only_one_reachable_proj */
2964 * Return non-zero if the control flow predecessor node pred
2965 * is the only reachable control flow exit of its block.
2967 * @param pred the control flow exit
2968 * @param block the destination block
2970 static int can_exchange(ir_node *pred, ir_node *block)
2972 if (is_Start(pred) || get_Block_entity(block) != NULL)
2974 else if (is_Jmp(pred))
2976 else if (is_Raise(pred)) {
2977 /* Raise is a tuple and usually has only one reachable ProjX,
2978 * but it must not be eliminated like a Jmp */
2981 else if (get_irn_mode(pred) == mode_T) {
2982 /* if the predecessor block has more than one
2983 reachable outputs we cannot remove the block */
2984 return only_one_reachable_proj(pred);
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)
2995 environment_t *env = (environment_t*)ctx;
2996 node_t *node = get_irn_node(block);
2998 ir_node **ins, **in_X;
2999 ir_node *phi, *next;
3001 n = get_Block_n_cfgpreds(block);
3003 if (node->type.tv == tarval_unreachable) {
3006 for (i = n - 1; i >= 0; --i) {
3007 ir_node *pred = get_Block_cfgpred(block, i);
3009 if (! is_Bad(pred)) {
3010 ir_node *pred_block = get_nodes_block(skip_Proj(pred));
3011 if (!is_Bad(pred_block)) {
3012 node_t *pred_bl = get_irn_node(pred_block);
3014 if (pred_bl->flagged == 0) {
3015 pred_bl->flagged = 3;
3017 if (pred_bl->type.tv == tarval_reachable) {
3019 * We will remove an edge from block to its pred.
3020 * This might leave the pred block as an endless loop
3022 if (! is_backedge(block, i))
3023 keep_alive(pred_bl->node);
3030 if (block == get_irg_end_block(current_ir_graph)) {
3031 /* Analysis found out that the end block is unreachable,
3032 * hence we remove all its control flow predecessors. */
3033 set_irn_in(block, 0, NULL);
3039 /* only one predecessor combine */
3040 ir_node *pred = skip_Proj(get_Block_cfgpred(block, 0));
3042 if (can_exchange(pred, block)) {
3043 ir_node *new_block = get_nodes_block(pred);
3044 DB((dbg, LEVEL_1, "Fuse %+F with %+F\n", block, new_block));
3045 DBG_OPT_COMBO(block, new_block, FS_OPT_COMBO_CF);
3046 exchange(block, new_block);
3047 node->node = new_block;
3053 NEW_ARR_A(ir_node *, in_X, n);
3055 for (i = 0; i < n; ++i) {
3056 ir_node *pred = get_Block_cfgpred(block, i);
3057 node_t *node = get_irn_node(pred);
3059 if (node->type.tv == tarval_reachable) {
3062 DB((dbg, LEVEL_1, "Removing dead input %d from %+F (%+F)\n", i, block, pred));
3063 if (! is_Bad(pred)) {
3064 ir_node *pred_block = get_nodes_block(skip_Proj(pred));
3065 if (!is_Bad(pred_block)) {
3066 node_t *pred_bl = get_irn_node(pred_block);
3068 if (!is_Bad(pred_bl->node) && pred_bl->flagged == 0) {
3069 pred_bl->flagged = 3;
3071 if (pred_bl->type.tv == tarval_reachable) {
3073 * We will remove an edge from block to its pred.
3074 * This might leave the pred block as an endless loop
3076 if (! is_backedge(block, i))
3077 keep_alive(pred_bl->node);
3088 NEW_ARR_A(ir_node *, ins, n);
3089 for (phi = get_Block_phis(block); phi != NULL; phi = next) {
3090 node_t *node = get_irn_node(phi);
3092 next = get_Phi_next(phi);
3093 if (is_tarval(node->type.tv) && tarval_is_constant(node->type.tv)) {
3094 /* this Phi is replaced by a constant */
3095 ir_tarval *tv = node->type.tv;
3096 ir_node *c = new_r_Const(current_ir_graph, tv);
3098 set_irn_node(c, node);
3100 DB((dbg, LEVEL_1, "%+F is replaced by %+F\n", phi, c));
3101 DBG_OPT_COMBO(phi, c, FS_OPT_COMBO_CONST);
3106 for (i = 0; i < n; ++i) {
3107 node_t *pred = get_irn_node(get_Block_cfgpred(block, i));
3109 if (pred->type.tv == tarval_reachable) {
3110 ins[j++] = get_Phi_pred(phi, i);
3114 /* this Phi is replaced by a single predecessor */
3115 ir_node *s = ins[0];
3116 node_t *phi_node = get_irn_node(phi);
3119 DB((dbg, LEVEL_1, "%+F is replaced by %+F because of cf change\n", phi, s));
3120 DBG_OPT_COMBO(phi, s, FS_OPT_COMBO_FOLLOWER);
3125 set_irn_in(phi, j, ins);
3133 /* this Block has only one live predecessor */
3134 ir_node *pred = skip_Proj(in_X[0]);
3136 if (can_exchange(pred, block)) {
3137 ir_node *new_block = get_nodes_block(pred);
3138 DBG_OPT_COMBO(block, new_block, FS_OPT_COMBO_CF);
3139 exchange(block, new_block);
3140 node->node = new_block;
3145 set_irn_in(block, k, in_X);
3150 * Exchange a node by its leader.
3151 * Beware: in rare cases the mode might be wrong here, for instance
3152 * AddP(x, NULL) is a follower of x, but with different mode.
3155 static void exchange_leader(ir_node *irn, ir_node *leader)
3157 ir_mode *mode = get_irn_mode(irn);
3158 if (mode != get_irn_mode(leader)) {
3159 /* The conv is a no-op, so we are free to place it
3160 * either in the block of the leader OR in irn's block.
3161 * Probably placing it into leaders block might reduce
3162 * the number of Conv due to CSE. */
3163 ir_node *block = get_nodes_block(leader);
3164 dbg_info *dbg = get_irn_dbg_info(irn);
3165 ir_node *nlead = new_rd_Conv(dbg, block, leader, mode);
3167 if (nlead != leader) {
3168 /* Note: this newly create irn has no node info because
3169 * it is created after the analysis. However, this node
3170 * replaces the node irn and should not be visited again,
3171 * so set its visited count to the count of irn.
3172 * Otherwise we might visited this node more than once if
3173 * irn had more than one user.
3175 set_irn_node(nlead, NULL);
3176 set_irn_visited(nlead, get_irn_visited(irn));
3180 exchange(irn, leader);
3181 } /* exchange_leader */
3184 * Check, if all users of a mode_M node are dead. Use
3185 * the Def-Use edges for this purpose, as they still
3186 * reflect the situation.
3188 static int all_users_are_dead(const ir_node *irn)
3190 unsigned n = get_irn_n_outs(irn);
3191 for (unsigned i = 0; i < n; ++i) {
3192 const ir_node *succ = get_irn_out(irn, i);
3193 const node_t *block = get_irn_node(get_nodes_block(succ));
3196 if (block->type.tv == tarval_unreachable) {
3197 /* block is unreachable */
3200 node = get_irn_node(succ);
3201 if (node->type.tv != tarval_top) {
3202 /* found a reachable user */
3206 /* all users are unreachable */
3208 } /* all_user_are_dead */
3211 * Walker: Find reachable mode_M nodes that have only
3212 * unreachable users. These nodes must be kept later.
3214 static void find_kept_memory(ir_node *irn, void *ctx)
3216 environment_t *env = (environment_t*)ctx;
3217 node_t *node, *block;
3219 if (get_irn_mode(irn) != mode_M)
3222 block = get_irn_node(get_nodes_block(irn));
3223 if (block->type.tv == tarval_unreachable)
3226 node = get_irn_node(irn);
3227 if (node->type.tv == tarval_top)
3230 /* ok, we found a live memory node. */
3231 if (all_users_are_dead(irn)) {
3232 DB((dbg, LEVEL_1, "%+F must be kept\n", irn));
3233 ARR_APP1(ir_node *, env->kept_memory, irn);
3235 } /* find_kept_memory */
3238 * Post-Walker, apply the analysis results;
3240 static void apply_result(ir_node *irn, void *ctx)
3242 environment_t *env = (environment_t*)ctx;
3243 node_t *node = get_irn_node(irn);
3245 if (is_Block(irn) || is_End(irn) || is_Bad(irn)) {
3246 /* blocks already handled, do not touch the End node */
3248 node_t *block = get_irn_node(get_nodes_block(irn));
3250 if (block->type.tv == tarval_unreachable) {
3251 ir_graph *irg = get_irn_irg(irn);
3252 ir_mode *mode = get_irn_mode(node->node);
3253 ir_node *bad = new_r_Bad(irg, mode);
3255 /* here, bad might already have a node, but this can be safely ignored
3256 as long as bad has at least ONE valid node */
3257 set_irn_node(bad, node);
3259 DB((dbg, LEVEL_1, "%+F is unreachable\n", irn));
3262 } else if (node->type.tv == tarval_top) {
3263 ir_mode *mode = get_irn_mode(irn);
3265 if (mode == mode_M) {
3266 /* never kill a mode_M node */
3268 ir_node *pred = get_Proj_pred(irn);
3269 node_t *pnode = get_irn_node(pred);
3271 if (pnode->type.tv == tarval_top) {
3272 /* skip the predecessor */
3273 ir_node *mem = get_memop_mem(pred);
3275 DB((dbg, LEVEL_1, "%+F computes Top, replaced by %+F\n", irn, mem));
3280 /* leave other nodes, especially PhiM */
3281 } else if (mode == mode_T) {
3282 /* Do not kill mode_T nodes, kill their Projs */
3283 } else if (! is_Unknown(irn)) {
3284 /* don't kick away Unknown's, they might be still needed */
3285 ir_node *unk = new_r_Unknown(current_ir_graph, mode);
3287 /* control flow should already be handled at apply_cf() */
3288 assert(mode != mode_X);
3290 /* see comment above */
3291 set_irn_node(unk, node);
3293 DB((dbg, LEVEL_1, "%+F computes Top\n", irn));
3298 else if (get_irn_mode(irn) == mode_X) {
3301 ir_node *cond = get_Proj_pred(irn);
3303 if (is_Cond(cond) || is_Switch(cond)) {
3304 if (only_one_reachable_proj(cond)) {
3305 ir_node *jmp = new_r_Jmp(block->node);
3306 set_irn_node(jmp, node);
3308 DB((dbg, LEVEL_1, "%+F is replaced by %+F\n", irn, jmp));
3309 DBG_OPT_COMBO(irn, jmp, FS_OPT_COMBO_CF);
3313 if (is_Switch(cond)) {
3314 node_t *sel = get_irn_node(get_Switch_selector(cond));
3315 ir_tarval *tv = sel->type.tv;
3317 if (is_tarval(tv) && tarval_is_constant(tv)) {
3318 /* The selector is a constant, but more
3319 * than one output is active: An unoptimized
3328 /* normal data node */
3329 if (is_tarval(node->type.tv) && tarval_is_constant(node->type.tv)) {
3330 ir_tarval *tv = node->type.tv;
3333 * Beware: never replace mode_T nodes by constants. Currently we must mark
3334 * mode_T nodes with constants, but do NOT replace them.
3336 if (! is_Const(irn) && get_irn_mode(irn) != mode_T) {
3337 /* can be replaced by a constant */
3338 ir_node *c = new_r_Const(current_ir_graph, tv);
3339 set_irn_node(c, node);
3341 DB((dbg, LEVEL_1, "%+F is replaced by %+F\n", irn, c));
3342 DBG_OPT_COMBO(irn, c, FS_OPT_COMBO_CONST);
3343 exchange_leader(irn, c);
3346 } else if (is_entity(node->type.sym.entity_p)) {
3347 if (! is_SymConst(irn)) {
3348 /* can be replaced by a SymConst */
3349 ir_node *symc = new_r_SymConst(current_ir_graph, get_irn_mode(irn), node->type.sym, symconst_addr_ent);
3350 set_irn_node(symc, node);
3353 DB((dbg, LEVEL_1, "%+F is replaced by %+F\n", irn, symc));
3354 DBG_OPT_COMBO(irn, symc, FS_OPT_COMBO_CONST);
3355 exchange_leader(irn, symc);
3358 } else if (is_Confirm(irn)) {
3359 /* Confirms are always follower, but do not kill them here */
3361 ir_node *leader = get_leader(node);
3363 if (leader != irn) {
3364 int non_strict_phi = 0;
3367 * Beware: Do not remove Phi(Unknown, ..., x, ..., Unknown)
3368 * as this might create non-strict programs.
3370 if (node->is_follower && is_Phi(irn) && !is_Unknown(leader)) {
3373 for (i = get_Phi_n_preds(irn) - 1; i >= 0; --i) {
3374 ir_node *pred = get_Phi_pred(irn, i);
3376 if (is_Unknown(pred)) {
3382 if (! non_strict_phi) {
3383 DB((dbg, LEVEL_1, "%+F from part%d is replaced by %+F\n", irn, node->part->nr, leader));
3384 if (node->is_follower)
3385 DBG_OPT_COMBO(irn, leader, FS_OPT_COMBO_FOLLOWER);
3387 DBG_OPT_COMBO(irn, leader, FS_OPT_COMBO_CONGRUENT);
3388 exchange_leader(irn, leader);
3395 } /* apply_result */
3398 * Fix the keep-alives by deleting unreachable ones.
3400 static void apply_end(ir_node *end, environment_t *env)
3402 int i, j, n = get_End_n_keepalives(end);
3403 ir_node **in = NULL;
3406 NEW_ARR_A(ir_node *, in, n);
3408 /* fix the keep alive */
3409 for (i = j = 0; i < n; i++) {
3410 ir_node *ka = get_End_keepalive(end, i);
3416 if (!is_Block(ka)) {
3417 block = get_nodes_block(ka);
3424 node = get_irn_node(block);
3425 if (node->type.tv != tarval_unreachable)
3429 set_End_keepalives(end, j, in);
3434 #define SET(code) op_##code->ops.generic = (op_func)compute_##code
3437 * sets the generic functions to compute.
3439 static void set_compute_functions(void)
3443 /* set the default compute function */
3444 for (i = 0, n = ir_get_n_opcodes(); i < n; ++i) {
3445 ir_op *op = ir_get_opcode(i);
3446 op->ops.generic = (op_func)default_compute;
3449 /* set specific functions */
3465 } /* set_compute_functions */
3470 static void add_memory_keeps(ir_node **kept_memory, size_t len)
3472 ir_node *end = get_irg_end(current_ir_graph);
3477 ir_nodeset_init(&set);
3479 /* check, if those nodes are already kept */
3480 for (i = get_End_n_keepalives(end) - 1; i >= 0; --i)
3481 ir_nodeset_insert(&set, get_End_keepalive(end, i));
3483 for (idx = 0; idx < len; ++idx) {
3484 ir_node *ka = kept_memory[idx];
3486 if (! ir_nodeset_contains(&set, ka)) {
3487 add_End_keepalive(end, ka);
3490 ir_nodeset_destroy(&set);
3491 } /* add_memory_keeps */
3493 void combo(ir_graph *irg)
3496 ir_node *initial_bl;
3498 ir_graph *rem = current_ir_graph;
3501 assure_irg_properties(irg,
3502 IR_GRAPH_PROPERTY_NO_BADS
3503 | IR_GRAPH_PROPERTY_CONSISTENT_OUTS
3504 | IR_GRAPH_PROPERTY_CONSISTENT_LOOPINFO);
3506 current_ir_graph = irg;
3508 /* register a debug mask */
3509 FIRM_DBG_REGISTER(dbg, "firm.opt.combo");
3511 DB((dbg, LEVEL_1, "Doing COMBO for %+F\n", irg));
3513 obstack_init(&env.obst);
3514 env.worklist = NULL;
3518 #ifdef DEBUG_libfirm
3519 env.dbg_list = NULL;
3521 env.opcode2id_map = new_set(cmp_opcode, iro_Last * 4);
3522 env.kept_memory = NEW_ARR_F(ir_node *, 0);
3523 env.end_idx = get_opt_global_cse() ? 0 : -1;
3524 env.lambda_input = 0;
3527 /* options driving the optimization */
3528 env.commutative = 1;
3529 env.opt_unknown = 1;
3531 /* we have our own value_of function */
3532 set_value_of_func(get_node_tarval);
3534 set_compute_functions();
3535 DEBUG_ONLY(part_nr = 0;)
3537 ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK | IR_RESOURCE_PHI_LIST);
3539 if (env.opt_unknown)
3540 tarval_UNKNOWN = tarval_top;
3542 tarval_UNKNOWN = tarval_bad;
3544 /* create the initial partition and place it on the work list */
3545 env.initial = new_partition(&env);
3546 add_to_worklist(env.initial, &env);
3547 irg_walk_graph(irg, create_initial_partitions, init_block_phis, &env);
3549 /* set the hook: from now, every node has a partition and a type */
3550 DEBUG_ONLY(set_dump_node_vcgattr_hook(dump_partition_hook);)
3552 /* all nodes on the initial partition have type Top */
3553 env.initial->type_is_T_or_C = 1;
3555 /* Place the START Node's partition on cprop.
3556 Place the START Node on its local worklist. */
3557 initial_bl = get_irg_start_block(irg);
3558 start = get_irn_node(initial_bl);
3559 add_to_cprop(start, &env);
3563 if (env.worklist != NULL)
3565 } while (env.cprop != NULL || env.worklist != NULL);
3567 dump_all_partitions(&env);
3568 check_all_partitions(&env);
3571 dump_ir_block_graph(irg, "-partition");
3574 /* apply the result */
3576 /* check, which nodes must be kept */
3577 irg_walk_graph(irg, NULL, find_kept_memory, &env);
3579 /* kill unreachable control flow */
3580 irg_block_walk_graph(irg, NULL, apply_cf, &env);
3581 /* Kill keep-alives of dead blocks: this speeds up apply_result()
3582 * and fixes assertion because dead cf to dead blocks is NOT removed by
3584 apply_end(get_irg_end(irg), &env);
3585 irg_walk_graph(irg, NULL, apply_result, &env);
3587 len = ARR_LEN(env.kept_memory);
3589 add_memory_keeps(env.kept_memory, len);
3592 DB((dbg, LEVEL_1, "Unoptimized Control Flow left"));
3595 ir_free_resources(irg, IR_RESOURCE_IRN_LINK | IR_RESOURCE_PHI_LIST);
3597 /* remove the partition hook */
3598 DEBUG_ONLY(set_dump_node_vcgattr_hook(NULL);)
3600 DEL_ARR_F(env.kept_memory);
3601 del_set(env.opcode2id_map);
3602 obstack_free(&env.obst, NULL);
3604 /* restore value_of() default behavior */
3605 set_value_of_func(NULL);
3606 current_ir_graph = rem;
3608 confirm_irg_properties(irg, IR_GRAPH_PROPERTIES_NONE);
3611 /* Creates an ir_graph pass for combo. */
3612 ir_graph_pass_t *combo_pass(const char *name)
3614 return def_graph_pass(name ? name : "combo", combo);
projects / libfirm / blob