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;
444 DB((dbg, LEVEL_2, "Split by %s produced = {\n", what_reason));
445 for (p = list; p != NULL; p = p->split_next) {
446 DB((dbg, LEVEL_2, "%c part%u", split, p->nr));
449 DB((dbg, LEVEL_2, "\n}\n"));
450 } /* dump_split_list */
453 * Dump partition and type for a node.
455 static int dump_partition_hook(FILE *F, const ir_node *n, const ir_node *local)
457 const ir_node *irn = local != NULL ? local : n;
458 node_t *node = get_irn_node(irn);
460 ir_fprintf(F, "info2 : \"partition %u type %+F\"\n", node->part->nr, node->type);
462 } /* dump_partition_hook */
465 #define dump_partition(msg, part)
466 #define dump_race_list(msg, list)
467 #define dump_list(msg, list)
468 #define dump_all_partitions(env)
469 #define dump_split_list(list)
472 #if defined(VERIFY_MONOTONE) && defined (DEBUG_libfirm)
474 * Verify that a type transition is monotone
476 static void verify_type(const lattice_elem_t old_type, node_t *node)
478 if (old_type.tv == node->type.tv) {
482 if (old_type.tv == tarval_top) {
483 /* from Top down-to is always allowed */
486 if (node->type.tv == tarval_bottom || node->type.tv == tarval_reachable) {
490 panic("wrong translation from %+F to %+F on node %+F", old_type, node->type, node->node);
494 #define verify_type(old_type, node)
498 * Compare two pointer values of a listmap.
500 static int listmap_cmp_ptr(const void *elt, const void *key, size_t size)
502 const listmap_entry_t *e1 = (listmap_entry_t*)elt;
503 const listmap_entry_t *e2 = (listmap_entry_t*)key;
506 return e1->id != e2->id;
507 } /* listmap_cmp_ptr */
510 * Initializes a listmap.
512 * @param map the listmap
514 static void listmap_init(listmap_t *map)
516 map->map = new_set(listmap_cmp_ptr, 16);
521 * Terminates a listmap.
523 * @param map the listmap
525 static void listmap_term(listmap_t *map)
531 * Return the associated listmap entry for a given id.
533 * @param map the listmap
534 * @param id the id to search for
536 * @return the associated listmap entry for the given id
538 static listmap_entry_t *listmap_find(listmap_t *map, void *id)
540 listmap_entry_t key, *entry;
545 entry = set_insert(listmap_entry_t, map->map, &key, sizeof(key), hash_ptr(id));
547 if (entry->list == NULL) {
548 /* a new entry, put into the list */
549 entry->next = map->values;
556 * Calculate the hash value for an opcode map entry.
558 * @param entry an opcode map entry
560 * @return a hash value for the given opcode map entry
562 static unsigned opcode_hash(const opcode_key_t *entry)
564 /* we cannot use the ir ops hash function here, because it hashes the
566 const ir_node *n = entry->irn;
567 ir_opcode code = (ir_opcode)get_irn_opcode(n);
568 ir_mode *mode = get_irn_mode(n);
569 unsigned hash = (unsigned)(PTR_TO_INT(mode) * 9 + code) + get_irn_arity(n);
571 if (code == iro_Const)
572 hash ^= (unsigned)hash_ptr(get_Const_tarval(n));
573 else if (code == iro_Proj)
574 hash += (unsigned)get_Proj_proj(n);
579 * Compare two entries in the opcode map.
581 static int cmp_opcode(const void *elt, const void *key, size_t size)
583 const opcode_key_t *o1 = (opcode_key_t*)elt;
584 const opcode_key_t *o2 = (opcode_key_t*)key;
588 return cmp_irn_opcode(o1->irn, o2->irn);
592 * Compare two Def-Use edges for input position.
594 static int cmp_def_use_edge(const void *a, const void *b)
596 const ir_def_use_edge *ea = (const ir_def_use_edge*)a;
597 const ir_def_use_edge *eb = (const ir_def_use_edge*)b;
599 /* no overrun, because range is [-1, MAXINT] */
600 return ea->pos - eb->pos;
601 } /* cmp_def_use_edge */
604 * We need the Def-Use edges sorted.
606 static void sort_irn_outs(node_t *node)
608 ir_node *irn = node->node;
609 unsigned n_outs = get_irn_n_outs(irn);
610 qsort(irn->o.out->edges, n_outs, sizeof(irn->o.out->edges[0]),
612 node->max_user_input = n_outs > 0 ? irn->o.out->edges[n_outs-1].pos : -1;
613 } /* sort_irn_outs */
616 * Return the type of a node.
618 * @param irn an IR-node
620 * @return the associated type of this node
622 static inline lattice_elem_t get_node_type(const ir_node *irn)
624 return get_irn_node(irn)->type;
625 } /* get_node_type */
628 * Return the tarval of a node.
630 * @param irn an IR-node
632 * @return the associated type of this node
634 static inline ir_tarval *get_node_tarval(const ir_node *irn)
636 lattice_elem_t type = get_node_type(irn);
638 if (is_tarval(type.tv))
640 return tarval_bottom;
641 } /* get_node_type */
644 * Add a partition to the worklist.
646 static inline void add_to_worklist(partition_t *X, environment_t *env)
648 assert(X->on_worklist == 0);
649 DB((dbg, LEVEL_2, "Adding part%d to worklist\n", X->nr));
650 X->wl_next = env->worklist;
653 } /* add_to_worklist */
656 * Create a new empty partition.
658 * @param env the environment
660 * @return a newly allocated partition
662 static inline partition_t *new_partition(environment_t *env)
664 partition_t *part = OALLOC(&env->obst, partition_t);
666 INIT_LIST_HEAD(&part->Leader);
667 INIT_LIST_HEAD(&part->Follower);
668 INIT_LIST_HEAD(&part->cprop);
669 INIT_LIST_HEAD(&part->cprop_X);
670 part->wl_next = NULL;
671 part->touched_next = NULL;
672 part->cprop_next = NULL;
673 part->split_next = NULL;
674 part->touched = NULL;
677 part->max_user_inputs = 0;
678 part->on_worklist = 0;
679 part->on_touched = 0;
681 part->type_is_T_or_C = 0;
683 part->dbg_next = env->dbg_list;
684 env->dbg_list = part;
685 part->nr = part_nr++;
689 } /* new_partition */
692 * Get the first node from a partition.
694 static inline node_t *get_first_node(const partition_t *X)
696 return list_entry(X->Leader.next, node_t, node_list);
697 } /* get_first_node */
700 * Return the type of a partition (assuming partition is non-empty and
701 * all elements have the same type).
703 * @param X a partition
705 * @return the type of the first element of the partition
707 static inline lattice_elem_t get_partition_type(const partition_t *X)
709 const node_t *first = get_first_node(X);
711 } /* get_partition_type */
714 * Creates a partition node for the given IR-node and place it
715 * into the given partition.
717 * @param irn an IR-node
718 * @param part a partition to place the node in
719 * @param env the environment
721 * @return the created node
723 static node_t *create_partition_node(ir_node *irn, partition_t *part, environment_t *env)
725 /* create a partition node and place it in the partition */
726 node_t *node = OALLOC(&env->obst, node_t);
728 INIT_LIST_HEAD(&node->node_list);
729 INIT_LIST_HEAD(&node->cprop_list);
733 node->race_next = NULL;
734 node->type.tv = tarval_top;
735 node->max_user_input = 0;
737 node->n_followers = 0;
738 node->on_touched = 0;
741 node->is_follower = 0;
743 set_irn_node(irn, node);
745 list_add_tail(&node->node_list, &part->Leader);
749 } /* create_partition_node */
752 * Pre-Walker, initialize all Nodes' type to U or top and place
753 * all nodes into the TOP partition.
755 static void create_initial_partitions(ir_node *irn, void *ctx)
757 environment_t *env = (environment_t*)ctx;
758 partition_t *part = env->initial;
761 node = create_partition_node(irn, part, env);
763 if (node->max_user_input > part->max_user_inputs)
764 part->max_user_inputs = node->max_user_input;
767 set_Block_phis(irn, NULL);
769 } /* create_initial_partitions */
772 * Post-Walker, collect all Block-Phi lists, set Cond.
774 static void init_block_phis(ir_node *irn, void *ctx)
779 ir_node *block = get_nodes_block(irn);
780 add_Block_phi(block, irn);
782 } /* init_block_phis */
785 * Add a node to the entry.partition.touched set and
786 * node->partition to the touched set if not already there.
789 * @param env the environment
791 static inline void add_to_touched(node_t *y, environment_t *env)
793 if (y->on_touched == 0) {
794 partition_t *part = y->part;
796 y->next = part->touched;
801 if (part->on_touched == 0) {
802 part->touched_next = env->touched;
804 part->on_touched = 1;
807 check_list(part->touched, part);
809 } /* add_to_touched */
812 * Place a node on the cprop list.
815 * @param env the environment
817 static void add_to_cprop(node_t *y, environment_t *env)
821 /* Add y to y.partition.cprop. */
822 if (y->on_cprop == 0) {
823 partition_t *Y = y->part;
824 ir_node *irn = y->node;
825 ir_node *skipped = skip_Proj(irn);
827 /* place Conds and all its Projs on the cprop_X list */
828 if (is_Cond(skipped) || is_Switch(skipped))
829 list_add_tail(&y->cprop_list, &Y->cprop_X);
831 list_add_tail(&y->cprop_list, &Y->cprop);
834 DB((dbg, LEVEL_3, "Add %+F to part%u.cprop\n", y->node, Y->nr));
836 /* place its partition on the cprop list */
837 if (Y->on_cprop == 0) {
838 Y->cprop_next = env->cprop;
844 if (get_irn_mode(irn) == mode_T) {
845 /* mode_T nodes always produce tarval_bottom, so we must explicitly
846 * add its Projs to get constant evaluation to work */
847 for (unsigned i = get_irn_n_outs(irn); i-- > 0; ) {
848 node_t *proj = get_irn_node(get_irn_out(irn, i));
850 add_to_cprop(proj, env);
852 } else if (is_Block(irn)) {
853 /* Due to the way we handle Phi's, we must place all Phis of a block on the list
854 * if someone placed the block. The Block is only placed if the reachability
855 * changes, and this must be re-evaluated in compute_Phi(). */
857 for (phi = get_Block_phis(irn); phi != NULL; phi = get_Phi_next(phi)) {
858 node_t *p = get_irn_node(phi);
859 add_to_cprop(p, env);
865 * Update the worklist: If Z is on worklist then add Z' to worklist.
866 * Else add the smaller of Z and Z' to worklist.
868 * @param Z the Z partition
869 * @param Z_prime the Z' partition, a previous part of Z
870 * @param env the environment
872 static void update_worklist(partition_t *Z, partition_t *Z_prime, environment_t *env)
874 if (Z->on_worklist || Z_prime->n_leader < Z->n_leader) {
875 add_to_worklist(Z_prime, env);
877 add_to_worklist(Z, env);
879 } /* update_worklist */
882 * Make all inputs to x no longer be F.def_use edges.
886 static void move_edges_to_leader(node_t *x)
888 ir_node *irn = x->node;
889 for (int i = get_irn_arity(irn) - 1; i >= 0; --i) {
890 node_t *pred = get_irn_node(get_irn_n(irn, i));
891 ir_node *p = pred->node;
892 unsigned n = get_irn_n_outs(p);
893 for (unsigned j = 0; j < pred->n_followers; ++j) {
894 ir_def_use_edge edge = p->o.out->edges[j];
895 if (edge.pos == i && edge.use == irn) {
896 /* found a follower edge to x, move it to the Leader */
897 /* remove this edge from the Follower set */
899 p->o.out->edges[j] = p->o.out->edges[pred->n_followers];
901 /* sort it into the leader set */
903 for (k = pred->n_followers+1; k < n; ++k) {
904 if (p->o.out->edges[k].pos >= edge.pos)
906 p->o.out->edges[k-1] = p->o.out->edges[k];
908 /* place the new edge here */
909 p->o.out->edges[k-1] = edge;
911 /* edge found and moved */
916 } /* move_edges_to_leader */
919 * Split a partition that has NO followers by a local list.
921 * @param Z partition to split
922 * @param g a (non-empty) node list
923 * @param env the environment
925 * @return a new partition containing the nodes of g
927 static partition_t *split_no_followers(partition_t *Z, node_t *g, environment_t *env)
929 partition_t *Z_prime;
934 dump_partition("Splitting ", Z);
935 dump_list("by list ", g);
939 /* Remove g from Z. */
940 for (node = g; node != NULL; node = node->next) {
941 assert(node->part == Z);
942 list_del(&node->node_list);
945 assert(n < Z->n_leader);
948 /* Move g to a new partition, Z'. */
949 Z_prime = new_partition(env);
951 for (node = g; node != NULL; node = node->next) {
952 list_add_tail(&node->node_list, &Z_prime->Leader);
953 node->part = Z_prime;
954 if (node->max_user_input > max_input)
955 max_input = node->max_user_input;
957 Z_prime->max_user_inputs = max_input;
958 Z_prime->n_leader = n;
961 check_partition(Z_prime);
963 /* for now, copy the type info tag, it will be adjusted in split_by(). */
964 Z_prime->type_is_T_or_C = Z->type_is_T_or_C;
966 dump_partition("Now ", Z);
967 dump_partition("Created new ", Z_prime);
969 update_worklist(Z, Z_prime, env);
972 } /* split_no_followers */
975 * Make the Follower -> Leader transition for a node.
979 static void follower_to_leader(node_t *n)
981 assert(n->is_follower == 1);
983 DB((dbg, LEVEL_2, "%+F make the follower -> leader transition\n", n->node));
985 move_edges_to_leader(n);
986 list_del(&n->node_list);
987 list_add_tail(&n->node_list, &n->part->Leader);
989 } /* follower_to_leader */
992 * The environment for one race step.
994 typedef struct step_env {
995 node_t *initial; /**< The initial node list. */
996 node_t *unwalked; /**< The unwalked node list. */
997 node_t *walked; /**< The walked node list. */
998 unsigned index; /**< Next index of Follower use_def edge. */
999 unsigned side; /**< side number. */
1003 * Return non-zero, if a input is a real follower
1005 * @param irn the node to check
1006 * @param input number of the input
1008 static int is_real_follower(const ir_node *irn, int input)
1012 switch (get_irn_opcode(irn)) {
1015 /* ignore the Confirm bound input */
1021 /* ignore the Mux sel input */
1026 /* dead inputs are not follower edges */
1027 ir_node *block = get_nodes_block(irn);
1028 node_t *pred = get_irn_node(get_Block_cfgpred(block, input));
1030 if (pred->type.tv == tarval_unreachable)
1040 /* only a Sub x,0 / Shift x,0 might be a follower */
1047 pred = get_irn_node(get_irn_n(irn, input));
1048 if (is_tarval(pred->type.tv) && tarval_is_null(pred->type.tv))
1052 pred = get_irn_node(get_irn_n(irn, input));
1053 if (is_tarval(pred->type.tv) && tarval_is_one(pred->type.tv))
1057 pred = get_irn_node(get_irn_n(irn, input));
1058 if (is_tarval(pred->type.tv) && tarval_is_all_one(pred->type.tv))
1062 assert(!"opcode not implemented yet");
1066 } /* is_real_follower */
1069 * Do one step in the race.
1071 static int step(step_env *env)
1075 if (env->initial != NULL) {
1076 /* Move node from initial to unwalked */
1078 env->initial = n->race_next;
1080 n->race_next = env->unwalked;
1086 while (env->unwalked != NULL) {
1087 /* let n be the first node in unwalked */
1089 while (env->index < n->n_followers) {
1090 const ir_def_use_edge *edge = &n->node->o.out->edges[env->index];
1092 /* let m be n.F.def_use[index] */
1093 node_t *m = get_irn_node(edge->use);
1095 assert(m->is_follower);
1097 * Some inputs, like the get_Confirm_bound are NOT
1098 * real followers, sort them out.
1100 if (! is_real_follower(m->node, edge->pos)) {
1106 /* only followers from our partition */
1107 if (m->part != n->part)
1110 if ((m->flagged & env->side) == 0) {
1111 m->flagged |= env->side;
1113 if (m->flagged != 3) {
1114 /* visited the first time */
1115 /* add m to unwalked not as first node (we might still need to
1116 check for more follower node */
1117 m->race_next = n->race_next;
1121 /* else already visited by the other side and on the other list */
1124 /* move n to walked */
1125 env->unwalked = n->race_next;
1126 n->race_next = env->walked;
1134 * Clear the flags from a list and check for
1135 * nodes that where touched from both sides.
1137 * @param list the list
1139 static int clear_flags(node_t *list)
1144 for (n = list; n != NULL; n = n->race_next) {
1145 if (n->flagged == 3) {
1146 /* we reach a follower from both sides, this will split congruent
1147 * inputs and make it a leader. */
1148 follower_to_leader(n);
1157 * Split a partition by a local list using the race.
1159 * @param pX pointer to the partition to split, might be changed!
1160 * @param gg a (non-empty) node list
1161 * @param env the environment
1163 * @return a new partition containing the nodes of gg
1165 static partition_t *split(partition_t **pX, node_t *gg, environment_t *env)
1167 partition_t *X = *pX;
1168 partition_t *X_prime;
1172 int max_input, transitions, winner, shf;
1174 DEBUG_ONLY(static int run = 0;)
1176 DB((dbg, LEVEL_2, "Run %d ", run++));
1177 if (list_empty(&X->Follower)) {
1178 /* if the partition has NO follower, we can use the fast
1179 splitting algorithm. */
1180 return split_no_followers(X, gg, env);
1182 /* else do the race */
1184 dump_partition("Splitting ", X);
1185 dump_list("by list ", gg);
1187 INIT_LIST_HEAD(&tmp);
1189 /* Remove gg from X.Leader and put into g */
1191 for (node_t *node = gg; node != NULL; node = node->next) {
1192 assert(node->part == X);
1193 assert(node->is_follower == 0);
1195 list_del(&node->node_list);
1196 list_add_tail(&node->node_list, &tmp);
1197 node->race_next = g;
1202 list_for_each_entry(node_t, node, &X->Leader, node_list) {
1203 node->race_next = h;
1206 /* restore X.Leader */
1207 list_splice(&tmp, &X->Leader);
1209 senv[0].initial = g;
1210 senv[0].unwalked = NULL;
1211 senv[0].walked = NULL;
1215 senv[1].initial = h;
1216 senv[1].unwalked = NULL;
1217 senv[1].walked = NULL;
1222 * Some informations on the race that are not stated clearly in Click's
1224 * 1) A follower stays on the side that reach him first.
1225 * 2) If the other side reaches a follower, if will be converted to
1226 * a leader. /This must be done after the race is over, else the
1227 * edges we are iterating on are renumbered./
1228 * 3) /New leader might end up on both sides./
1229 * 4) /If one side ends up with new Leaders, we must ensure that
1230 * they can split out by opcode, hence we have to put _every_
1231 * partition with new Leader nodes on the cprop list, as
1232 * opcode splitting is done by split_by() at the end of
1233 * constant propagation./
1236 if (step(&senv[0])) {
1240 if (step(&senv[1])) {
1245 assert(senv[winner].initial == NULL);
1246 assert(senv[winner].unwalked == NULL);
1248 /* clear flags from walked/unwalked */
1250 transitions = clear_flags(senv[0].unwalked) << shf;
1251 transitions |= clear_flags(senv[0].walked) << shf;
1253 transitions |= clear_flags(senv[1].unwalked) << shf;
1254 transitions |= clear_flags(senv[1].walked) << shf;
1256 dump_race_list("winner ", senv[winner].walked);
1258 /* Move walked_{winner} to a new partition, X'. */
1259 X_prime = new_partition(env);
1262 for (node_t *node = senv[winner].walked; node != NULL; node = node->race_next) {
1263 list_del(&node->node_list);
1264 node->part = X_prime;
1265 if (node->is_follower) {
1266 list_add_tail(&node->node_list, &X_prime->Follower);
1268 list_add_tail(&node->node_list, &X_prime->Leader);
1271 if (node->max_user_input > max_input)
1272 max_input = node->max_user_input;
1274 X_prime->n_leader = n;
1275 X_prime->max_user_inputs = max_input;
1276 X->n_leader -= X_prime->n_leader;
1278 /* for now, copy the type info tag, it will be adjusted in split_by(). */
1279 X_prime->type_is_T_or_C = X->type_is_T_or_C;
1282 * Even if a follower was not checked by both sides, it might have
1283 * loose its congruence, so we need to check this case for all follower.
1285 list_for_each_entry_safe(node_t, node, t, &X_prime->Follower, node_list) {
1286 if (identity(node) == node) {
1287 follower_to_leader(node);
1293 check_partition(X_prime);
1295 dump_partition("Now ", X);
1296 dump_partition("Created new ", X_prime);
1298 /* X' is the smaller part */
1299 add_to_worklist(X_prime, env);
1302 * If there where follower to leader transitions, ensure that the nodes
1303 * can be split out if necessary.
1305 if (transitions & 1) {
1306 /* place winner partition on the cprop list */
1307 if (X_prime->on_cprop == 0) {
1308 X_prime->cprop_next = env->cprop;
1309 env->cprop = X_prime;
1310 X_prime->on_cprop = 1;
1313 if (transitions & 2) {
1314 /* place other partition on the cprop list */
1315 if (X->on_cprop == 0) {
1316 X->cprop_next = env->cprop;
1322 /* we have to ensure that the partition containing g is returned */
1332 * Returns non-zero if the i'th input of a Phi node is live.
1334 * @param phi a Phi-node
1335 * @param i an input number
1337 * @return non-zero if the i'th input of the given Phi node is live
1339 static int is_live_input(ir_node *phi, int i)
1342 ir_node *block = get_nodes_block(phi);
1343 ir_node *pred = get_Block_cfgpred(block, i);
1344 lattice_elem_t type = get_node_type(pred);
1346 return type.tv != tarval_unreachable;
1348 /* else it's the control input, always live */
1350 } /* is_live_input */
1353 * Return non-zero if a type is a constant.
1355 static int is_constant_type(lattice_elem_t type)
1357 if (type.tv != tarval_bottom && type.tv != tarval_top)
1360 } /* is_constant_type */
1363 * Check whether a type is neither Top or a constant.
1364 * Note: U is handled like Top here, R is a constant.
1366 * @param type the type to check
1368 static int type_is_neither_top_nor_const(const lattice_elem_t type)
1370 if (is_tarval(type.tv)) {
1371 if (type.tv == tarval_top)
1373 if (tarval_is_constant(type.tv))
1380 } /* type_is_neither_top_nor_const */
1383 * Collect nodes to the touched list.
1385 * @param list the list which contains the nodes that must be evaluated
1386 * @param idx the index of the def_use edge to evaluate
1387 * @param env the environment
1389 static void collect_touched(list_head *list, int idx, environment_t *env)
1392 int end_idx = env->end_idx;
1394 list_for_each_entry(node_t, x, list, node_list) {
1396 /* leader edges start AFTER follower edges */
1397 x->next_edge = x->n_followers;
1399 unsigned num_edges = get_irn_n_outs(x->node);
1401 /* for all edges in x.L.def_use_{idx} */
1402 while (x->next_edge < num_edges) {
1403 const ir_def_use_edge *edge = &x->node->o.out->edges[x->next_edge];
1406 /* check if we have necessary edges */
1407 if (edge->pos > idx)
1414 /* only non-commutative nodes */
1415 if (env->commutative &&
1416 (idx == 0 || idx == 1) && is_op_commutative(get_irn_op(succ)))
1419 /* ignore the "control input" for non-pinned nodes
1420 if we are running in GCSE mode */
1421 if (idx < end_idx && get_irn_pinned(succ) != op_pin_state_pinned)
1424 y = get_irn_node(succ);
1425 assert(get_irn_n(succ, idx) == x->node);
1427 /* ignore block edges touching followers */
1428 if (idx == -1 && y->is_follower)
1431 if (is_constant_type(y->type)) {
1432 unsigned code = get_irn_opcode(succ);
1433 if (code == iro_Sub || code == iro_Cmp)
1434 add_to_cprop(y, env);
1437 /* Partitions of constants should not be split simply because their Nodes have unequal
1438 functions or incongruent inputs. */
1439 if (type_is_neither_top_nor_const(y->type) &&
1440 (! is_Phi(y->node) || is_live_input(y->node, idx))) {
1441 add_to_touched(y, env);
1445 } /* collect_touched */
1448 * Collect commutative nodes to the touched list.
1450 * @param list the list which contains the nodes that must be evaluated
1451 * @param env the environment
1453 static void collect_commutative_touched(list_head *list, environment_t *env)
1457 list_for_each_entry(node_t, x, list, node_list) {
1458 unsigned num_edges = get_irn_n_outs(x->node);
1460 x->next_edge = x->n_followers;
1462 /* for all edges in x.L.def_use_{idx} */
1463 while (x->next_edge < num_edges) {
1464 const ir_def_use_edge *edge = &x->node->o.out->edges[x->next_edge];
1467 /* check if we have necessary edges */
1477 /* only commutative nodes */
1478 if (!is_op_commutative(get_irn_op(succ)))
1481 y = get_irn_node(succ);
1482 if (is_constant_type(y->type)) {
1483 unsigned code = get_irn_opcode(succ);
1484 if (code == iro_Eor)
1485 add_to_cprop(y, env);
1488 /* Partitions of constants should not be split simply because their Nodes have unequal
1489 functions or incongruent inputs. */
1490 if (type_is_neither_top_nor_const(y->type)) {
1491 add_to_touched(y, env);
1495 } /* collect_commutative_touched */
1498 * Split the partitions if caused by the first entry on the worklist.
1500 * @param env the environment
1502 static void cause_splits(environment_t *env)
1504 partition_t *X, *Z, *N;
1507 /* remove the first partition from the worklist */
1509 env->worklist = X->wl_next;
1512 dump_partition("Cause_split: ", X);
1514 if (env->commutative) {
1515 /* handle commutative nodes first */
1517 /* empty the touched set: already done, just clear the list */
1518 env->touched = NULL;
1520 collect_commutative_touched(&X->Leader, env);
1521 collect_commutative_touched(&X->Follower, env);
1523 for (Z = env->touched; Z != NULL; Z = N) {
1525 node_t *touched = Z->touched;
1526 node_t *touched_aa = NULL;
1527 node_t *touched_ab = NULL;
1528 unsigned n_touched_aa = 0;
1529 unsigned n_touched_ab = 0;
1531 assert(Z->touched != NULL);
1533 /* beware, split might change Z */
1534 N = Z->touched_next;
1536 /* remove it from the touched set */
1539 /* Empty local Z.touched. */
1540 for (e = touched; e != NULL; e = n) {
1541 node_t *left = get_irn_node(get_irn_n(e->node, 0));
1542 node_t *right = get_irn_node(get_irn_n(e->node, 1));
1544 assert(e->is_follower == 0);
1549 * Note: op(a, a) is NOT congruent to op(a, b).
1550 * So, we must split the touched list.
1552 if (left->part == right->part) {
1553 e->next = touched_aa;
1557 e->next = touched_ab;
1562 assert(n_touched_aa + n_touched_ab == Z->n_touched);
1566 if (0 < n_touched_aa && n_touched_aa < Z->n_leader) {
1567 partition_t *Z_prime = Z;
1568 DB((dbg, LEVEL_2, "Split part%d by touched_aa\n", Z_prime->nr));
1569 split(&Z_prime, touched_aa, env);
1571 assert(n_touched_aa <= Z->n_leader);
1573 if (0 < n_touched_ab && n_touched_ab < Z->n_leader) {
1574 partition_t *Z_prime = Z;
1575 DB((dbg, LEVEL_2, "Split part%d by touched_ab\n", Z_prime->nr));
1576 split(&Z_prime, touched_ab, env);
1578 assert(n_touched_ab <= Z->n_leader);
1582 /* combine temporary leader and follower list */
1583 for (idx = -1; idx <= X->max_user_inputs; ++idx) {
1584 /* empty the touched set: already done, just clear the list */
1585 env->touched = NULL;
1587 collect_touched(&X->Leader, idx, env);
1588 collect_touched(&X->Follower, idx, env);
1590 for (Z = env->touched; Z != NULL; Z = N) {
1592 node_t *touched = Z->touched;
1593 unsigned n_touched = Z->n_touched;
1595 assert(Z->touched != NULL);
1597 /* beware, split might change Z */
1598 N = Z->touched_next;
1600 /* remove it from the touched set */
1603 /* Empty local Z.touched. */
1604 for (e = touched; e != NULL; e = e->next) {
1605 assert(e->is_follower == 0);
1611 if (0 < n_touched && n_touched < Z->n_leader) {
1612 DB((dbg, LEVEL_2, "Split part%d by touched\n", Z->nr));
1613 split(&Z, touched, env);
1615 assert(n_touched <= Z->n_leader);
1618 } /* cause_splits */
1621 * Implements split_by_what(): Split a partition by characteristics given
1622 * by the what function.
1624 * @param X the partition to split
1625 * @param What a function returning an Id for every node of the partition X
1626 * @param P a list to store the result partitions
1627 * @param env the environment
1631 static partition_t *split_by_what(partition_t *X, what_func What,
1632 partition_t **P, environment_t *env)
1636 listmap_entry_t *iter;
1639 /* Let map be an empty mapping from the range of What to (local) list of Nodes. */
1641 list_for_each_entry(node_t, x, &X->Leader, node_list) {
1642 void *id = What(x, env);
1643 listmap_entry_t *entry;
1646 /* input not allowed, ignore */
1649 /* Add x to map[What(x)]. */
1650 entry = listmap_find(&map, id);
1651 x->next = entry->list;
1654 /* Let P be a set of Partitions. */
1656 /* for all sets S except one in the range of map do */
1657 for (iter = map.values; iter != NULL; iter = iter->next) {
1658 if (iter->next == NULL) {
1659 /* this is the last entry, ignore */
1664 /* Add SPLIT( X, S ) to P. */
1665 DB((dbg, LEVEL_2, "Split part%d by WHAT = %s\n", X->nr, what_reason));
1666 R = split(&X, S, env);
1676 } /* split_by_what */
1678 /** lambda n.(n.type) */
1679 static void *lambda_type(const node_t *node, environment_t *env)
1682 return node->type.tv;
1685 /** lambda n.(n.opcode) */
1686 static void *lambda_opcode(const node_t *node, environment_t *env)
1688 opcode_key_t key, *entry;
1690 key.irn = node->node;
1692 entry = set_insert(opcode_key_t, env->opcode2id_map, &key, sizeof(key), opcode_hash(&key));
1694 } /* lambda_opcode */
1696 /** lambda n.(n[i].partition) */
1697 static void *lambda_partition(const node_t *node, environment_t *env)
1699 ir_node *skipped = skip_Proj(node->node);
1702 int i = env->lambda_input;
1704 if (i >= get_irn_arity(node->node)) {
1706 * We are outside the allowed range: This can happen even
1707 * if we have split by opcode first: doing so might move Followers
1708 * to Leaders and those will have a different opcode!
1709 * Note that in this case the partition is on the cprop list and will be
1715 /* ignore the "control input" for non-pinned nodes
1716 if we are running in GCSE mode */
1717 if (i < env->end_idx && get_irn_pinned(skipped) != op_pin_state_pinned)
1720 pred = i == -1 ? get_irn_n(skipped, i) : get_irn_n(node->node, i);
1721 p = get_irn_node(pred);
1723 } /* lambda_partition */
1725 /** lambda n.(n[i].partition) for commutative nodes */
1726 static void *lambda_commutative_partition(const node_t *node, environment_t *env)
1728 ir_node *irn = node->node;
1729 ir_node *skipped = skip_Proj(irn);
1730 ir_node *pred, *left, *right;
1732 partition_t *pl, *pr;
1733 int i = env->lambda_input;
1735 if (i >= get_irn_arity(node->node)) {
1737 * We are outside the allowed range: This can happen even
1738 * if we have split by opcode first: doing so might move Followers
1739 * to Leaders and those will have a different opcode!
1740 * Note that in this case the partition is on the cprop list and will be
1746 /* ignore the "control input" for non-pinned nodes
1747 if we are running in GCSE mode */
1748 if (i < env->end_idx && get_irn_pinned(skipped) != op_pin_state_pinned)
1752 pred = get_irn_n(skipped, i);
1753 p = get_irn_node(pred);
1757 if (is_op_commutative(get_irn_op(irn))) {
1758 /* normalize partition order by returning the "smaller" on input 0,
1759 the "bigger" on input 1. */
1760 left = get_binop_left(irn);
1761 pl = get_irn_node(left)->part;
1762 right = get_binop_right(irn);
1763 pr = get_irn_node(right)->part;
1766 return pl < pr ? pl : pr;
1768 return pl > pr ? pl : pr;
1770 /* a not split out Follower */
1771 pred = get_irn_n(irn, i);
1772 p = get_irn_node(pred);
1776 } /* lambda_commutative_partition */
1779 * Returns true if a type is a constant (and NOT Top
1782 static int is_con(const lattice_elem_t type)
1784 /* be conservative */
1785 if (is_tarval(type.tv))
1786 return tarval_is_constant(type.tv);
1787 return is_entity(type.sym.entity_p);
1791 * Implements split_by().
1793 * @param X the partition to split
1794 * @param env the environment
1796 static void split_by(partition_t *X, environment_t *env)
1798 partition_t *I, *P = NULL;
1801 dump_partition("split_by", X);
1803 if (X->n_leader == 1) {
1804 /* we have only one leader, no need to split, just check its type */
1805 node_t *x = get_first_node(X);
1806 X->type_is_T_or_C = x->type.tv == tarval_top || is_con(x->type);
1810 DEBUG_ONLY(what_reason = "lambda n.(n.type)";)
1811 P = split_by_what(X, lambda_type, &P, env);
1814 /* adjust the type tags, we have split partitions by type */
1815 for (I = P; I != NULL; I = I->split_next) {
1816 node_t *x = get_first_node(I);
1817 I->type_is_T_or_C = x->type.tv == tarval_top || is_con(x->type);
1824 if (Y->n_leader > 1) {
1825 /* we do not want split the TOP or constant partitions */
1826 if (! Y->type_is_T_or_C) {
1827 partition_t *Q = NULL;
1829 DEBUG_ONLY(what_reason = "lambda n.(n.opcode)";)
1830 Q = split_by_what(Y, lambda_opcode, &Q, env);
1837 if (Z->n_leader > 1) {
1838 const node_t *first = get_first_node(Z);
1839 int arity = get_irn_arity(first->node);
1841 what_func what = lambda_partition;
1842 DEBUG_ONLY(char buf[64];)
1844 if (env->commutative && is_op_commutative(get_irn_op(first->node)))
1845 what = lambda_commutative_partition;
1848 * BEWARE: during splitting by input 2 for instance we might
1849 * create new partitions which are different by input 1, so collect
1850 * them and split further.
1852 Z->split_next = NULL;
1855 for (input = arity - 1; input >= -1; --input) {
1857 partition_t *Z_prime = R;
1860 if (Z_prime->n_leader > 1) {
1861 env->lambda_input = input;
1862 DEBUG_ONLY(snprintf(buf, sizeof(buf), "lambda n.(n[%d].partition)", input);)
1863 DEBUG_ONLY(what_reason = buf;)
1864 S = split_by_what(Z_prime, what, &S, env);
1867 Z_prime->split_next = S;
1870 } while (R != NULL);
1875 } while (Q != NULL);
1878 } while (P != NULL);
1882 * (Re-)compute the type for a given node.
1884 * @param node the node
1886 static void default_compute(node_t *node)
1889 ir_node *irn = node->node;
1891 /* if any of the data inputs have type top, the result is type top */
1892 for (i = get_irn_arity(irn) - 1; i >= 0; --i) {
1893 ir_node *pred = get_irn_n(irn, i);
1894 node_t *p = get_irn_node(pred);
1896 if (p->type.tv == tarval_top) {
1897 node->type.tv = tarval_top;
1902 if (get_irn_mode(node->node) == mode_X)
1903 node->type.tv = tarval_reachable;
1905 node->type.tv = computed_value(irn);
1906 } /* default_compute */
1909 * (Re-)compute the type for a Block node.
1911 * @param node the node
1913 static void compute_Block(node_t *node)
1916 ir_node *block = node->node;
1918 ir_graph *const irg = get_Block_irg(block);
1919 if (block == get_irg_start_block(irg) || get_Block_entity(block) != NULL) {
1920 /* start block and labelled blocks are always reachable */
1921 node->type.tv = tarval_reachable;
1925 for (i = get_Block_n_cfgpreds(block) - 1; i >= 0; --i) {
1926 node_t *pred = get_irn_node(get_Block_cfgpred(block, i));
1928 if (pred->type.tv == tarval_reachable) {
1929 /* A block is reachable, if at least of predecessor is reachable. */
1930 node->type.tv = tarval_reachable;
1934 node->type.tv = tarval_top;
1935 } /* compute_Block */
1938 * (Re-)compute the type for a Bad node.
1940 * @param node the node
1942 static void compute_Bad(node_t *node)
1944 /* Bad nodes ALWAYS compute Top */
1945 node->type.tv = tarval_top;
1949 * (Re-)compute the type for an Unknown node.
1951 * @param node the node
1953 static void compute_Unknown(node_t *node)
1955 /* While Unknown nodes should compute Top this is dangerous:
1956 * a Top input to a Cond would lead to BOTH control flows unreachable.
1957 * While this is correct in the given semantics, it would destroy the Firm
1960 * It would be safe to compute Top IF it can be assured, that only Cmp
1961 * nodes are inputs to Conds. We check that first.
1962 * This is the way Frontends typically build Firm, but some optimizations
1963 * (jump threading for instance) might replace them by Phib's...
1965 node->type.tv = tarval_UNKNOWN;
1966 } /* compute_Unknown */
1969 * (Re-)compute the type for a Jmp node.
1971 * @param node the node
1973 static void compute_Jmp(node_t *node)
1975 node_t *block = get_irn_node(get_nodes_block(node->node));
1977 node->type = block->type;
1981 * (Re-)compute the type for the Return node.
1983 * @param node the node
1985 static void compute_Return(node_t *node)
1987 /* The Return node is NOT dead if it is in a reachable block.
1988 * This is already checked in compute(). so we can return
1989 * Reachable here. */
1990 node->type.tv = tarval_reachable;
1991 } /* compute_Return */
1994 * (Re-)compute the type for the End node.
1996 * @param node the node
1998 static void compute_End(node_t *node)
2000 /* the End node is NOT dead of course */
2001 node->type.tv = tarval_reachable;
2005 * (Re-)compute the type for a Call.
2007 * @param node the node
2009 static void compute_Call(node_t *node)
2012 * A Call computes always bottom, even if it has Unknown
2015 node->type.tv = tarval_bottom;
2016 } /* compute_Call */
2019 * (Re-)compute the type for a SymConst node.
2021 * @param node the node
2023 static void compute_SymConst(node_t *node)
2025 ir_node *irn = node->node;
2026 node_t *block = get_irn_node(get_nodes_block(irn));
2028 if (block->type.tv == tarval_unreachable) {
2029 node->type.tv = tarval_top;
2032 switch (get_SymConst_kind(irn)) {
2033 case symconst_addr_ent:
2034 node->type.sym = get_SymConst_symbol(irn);
2037 node->type.tv = computed_value(irn);
2039 } /* compute_SymConst */
2042 * (Re-)compute the type for a Phi node.
2044 * @param node the node
2046 static void compute_Phi(node_t *node)
2049 ir_node *phi = node->node;
2050 lattice_elem_t type;
2052 /* if a Phi is in a unreachable block, its type is TOP */
2053 node_t *block = get_irn_node(get_nodes_block(phi));
2055 if (block->type.tv == tarval_unreachable) {
2056 node->type.tv = tarval_top;
2060 /* Phi implements the Meet operation */
2061 type.tv = tarval_top;
2062 for (i = get_Phi_n_preds(phi) - 1; i >= 0; --i) {
2063 node_t *pred = get_irn_node(get_Phi_pred(phi, i));
2064 node_t *pred_X = get_irn_node(get_Block_cfgpred(block->node, i));
2066 if (pred_X->type.tv == tarval_unreachable || pred->type.tv == tarval_top) {
2067 /* ignore TOP inputs: We must check here for unreachable blocks,
2068 because Firm constants live in the Start Block are NEVER Top.
2069 Else, a Phi (1,2) will produce Bottom, even if the 2 for instance
2070 comes from a unreachable input. */
2073 if (pred->type.tv == tarval_bottom) {
2074 node->type.tv = tarval_bottom;
2076 } else if (type.tv == tarval_top) {
2077 /* first constant found */
2079 } else if (type.tv != pred->type.tv) {
2080 /* different constants or tarval_bottom */
2081 node->type.tv = tarval_bottom;
2084 /* else nothing, constants are the same */
2090 * (Re-)compute the type for an Add. Special case: one nodes is a Zero Const.
2092 * @param node the node
2094 static void compute_Add(node_t *node)
2096 ir_node *sub = node->node;
2097 node_t *l = get_irn_node(get_Add_left(sub));
2098 node_t *r = get_irn_node(get_Add_right(sub));
2099 lattice_elem_t a = l->type;
2100 lattice_elem_t b = r->type;
2103 if (a.tv == tarval_top || b.tv == tarval_top) {
2104 node->type.tv = tarval_top;
2105 } else if (a.tv == tarval_bottom || b.tv == tarval_bottom) {
2106 node->type.tv = tarval_bottom;
2108 /* x + 0 = 0 + x = x, but beware of floating point +0 + -0, so we
2109 must call tarval_add() first to handle this case! */
2110 if (is_tarval(a.tv)) {
2111 if (is_tarval(b.tv)) {
2112 node->type.tv = tarval_add(a.tv, b.tv);
2115 mode = get_tarval_mode(a.tv);
2116 if (a.tv == get_mode_null(mode)) {
2120 } else if (is_tarval(b.tv)) {
2121 mode = get_tarval_mode(b.tv);
2122 if (b.tv == get_mode_null(mode)) {
2127 node->type.tv = tarval_bottom;
2132 * (Re-)compute the type for a Sub. Special case: both nodes are congruent.
2134 * @param node the node
2136 static void compute_Sub(node_t *node)
2138 ir_node *sub = node->node;
2139 node_t *l = get_irn_node(get_Sub_left(sub));
2140 node_t *r = get_irn_node(get_Sub_right(sub));
2141 lattice_elem_t a = l->type;
2142 lattice_elem_t b = r->type;
2145 if (a.tv == tarval_top || b.tv == tarval_top) {
2146 node->type.tv = tarval_top;
2147 } else if (is_con(a) && is_con(b)) {
2148 if (is_tarval(a.tv) && is_tarval(b.tv)) {
2149 node->type.tv = tarval_sub(a.tv, b.tv, get_irn_mode(sub));
2150 } else if (is_tarval(a.tv) && tarval_is_null(a.tv)) {
2152 } else if (is_tarval(b.tv) && tarval_is_null(b.tv)) {
2155 node->type.tv = tarval_bottom;
2157 } else if (r->part == l->part &&
2158 (!mode_is_float(get_irn_mode(l->node)))) {
2160 * BEWARE: a - a is NOT always 0 for floating Point values, as
2161 * NaN op NaN = NaN, so we must check this here.
2163 ir_mode *mode = get_irn_mode(sub);
2164 tv = get_mode_null(mode);
2166 /* if the node was ONCE evaluated by all constants, but now
2167 this breaks AND we get from the argument partitions a different
2168 result, switch to bottom.
2169 This happens because initially all nodes are in the same partition ... */
2170 if (node->type.tv != tv)
2174 node->type.tv = tarval_bottom;
2179 * (Re-)compute the type for an Eor. Special case: both nodes are congruent.
2181 * @param node the node
2183 static void compute_Eor(node_t *node)
2185 ir_node *eor = node->node;
2186 node_t *l = get_irn_node(get_Eor_left(eor));
2187 node_t *r = get_irn_node(get_Eor_right(eor));
2188 lattice_elem_t a = l->type;
2189 lattice_elem_t b = r->type;
2192 if (a.tv == tarval_top || b.tv == tarval_top) {
2193 node->type.tv = tarval_top;
2194 } else if (is_con(a) && is_con(b)) {
2195 if (is_tarval(a.tv) && is_tarval(b.tv)) {
2196 node->type.tv = tarval_eor(a.tv, b.tv);
2197 } else if (is_tarval(a.tv) && tarval_is_null(a.tv)) {
2199 } else if (is_tarval(b.tv) && tarval_is_null(b.tv)) {
2202 node->type.tv = tarval_bottom;
2204 } else if (r->part == l->part) {
2205 ir_mode *mode = get_irn_mode(eor);
2206 tv = get_mode_null(mode);
2208 /* if the node was ONCE evaluated by all constants, but now
2209 this breaks AND we get from the argument partitions a different
2210 result, switch to bottom.
2211 This happens because initially all nodes are in the same partition ... */
2212 if (node->type.tv != tv)
2216 node->type.tv = tarval_bottom;
2221 * (Re-)compute the type for Cmp.
2223 * @param node the node
2225 static void compute_Cmp(node_t *node)
2227 ir_node *cmp = node->node;
2228 node_t *l = get_irn_node(get_Cmp_left(cmp));
2229 node_t *r = get_irn_node(get_Cmp_right(cmp));
2230 lattice_elem_t a = l->type;
2231 lattice_elem_t b = r->type;
2232 ir_relation relation = get_Cmp_relation(cmp);
2235 if (a.tv == tarval_top || b.tv == tarval_top) {
2236 node->type.tv = tarval_undefined;
2237 } else if (is_con(a) && is_con(b)) {
2238 default_compute(node);
2241 * BEWARE: a == a is NOT always True for floating Point values, as
2242 * NaN != NaN is defined, so we must check this here.
2243 * (while for some pnc we could still optimize we have to stay
2244 * consistent with compute_Cmp, so don't do anything for floats)
2246 } else if (r->part == l->part && !mode_is_float(get_irn_mode(l->node))) {
2247 tv = relation & ir_relation_equal ? tarval_b_true : tarval_b_false;
2249 /* if the node was ONCE evaluated to a constant, but now
2250 this breaks AND we get from the argument partitions a different
2251 result, ensure monotony by fall to bottom.
2252 This happens because initially all nodes are in the same partition ... */
2253 if (node->type.tv == tarval_bottom)
2255 else if (node->type.tv != tv && is_constant_type(node->type))
2259 node->type.tv = tarval_bottom;
2264 * (Re-)compute the type for a Proj(Cond).
2266 * @param node the node
2267 * @param cond the predecessor Cond node
2269 static void compute_Proj_Cond(node_t *node, ir_node *cond)
2271 ir_node *proj = node->node;
2272 long pnc = get_Proj_proj(proj);
2273 ir_node *sel = get_Cond_selector(cond);
2274 node_t *selector = get_irn_node(sel);
2277 * Note: it is crucial for the monotony that the Proj(Cond)
2278 * are evaluates after all predecessors of the Cond selector are
2284 * Due to the fact that 0 is a const, the Cmp gets immediately
2285 * on the cprop list. It will be evaluated before x is evaluated,
2286 * might leaving x as Top. When later x is evaluated, the Cmp
2287 * might change its value.
2288 * BUT if the Cond is evaluated before this happens, Proj(Cond, FALSE)
2289 * gets R, and later changed to F if Cmp is evaluated to True!
2291 * We prevent this by putting Conds in an extra cprop_X queue, which
2292 * gets evaluated after the cprop queue is empty.
2294 * Note that this even happens with Click's original algorithm, if
2295 * Cmp(x, 0) is evaluated to True first and later changed to False
2296 * if x was Top first and later changed to a Const ...
2297 * It is unclear how Click solved that problem ...
2299 * However, in rare cases even this does not help, if a Top reaches
2300 * a compare through a Phi, than Proj(Cond) is evaluated changing
2301 * the type of the Phi to something other.
2302 * So, we take the last resort and bind the type to R once
2305 * (This might be even the way Click works around the whole problem).
2307 * Finally, we may miss some optimization possibilities due to this:
2312 * If Top reaches the if first, than we decide for != here.
2313 * If y later is evaluated to 0, we cannot revert this decision
2314 * and must live with both outputs enabled. If this happens,
2315 * we get an unresolved if (true) in the code ...
2317 * In Click's version where this decision is done at the Cmp,
2318 * the Cmp is NOT optimized away than (if y evaluated to 1
2319 * for instance) and we get a if (1 == 0) here ...
2321 * Both solutions are suboptimal.
2322 * At least, we could easily detect this problem and run
2323 * cf_opt() (or even combo) again :-(
2325 if (node->type.tv == tarval_reachable)
2328 if (pnc == pn_Cond_true) {
2329 if (selector->type.tv == tarval_b_false) {
2330 node->type.tv = tarval_unreachable;
2331 } else if (selector->type.tv == tarval_b_true) {
2332 node->type.tv = tarval_reachable;
2333 } else if (selector->type.tv == tarval_bottom) {
2334 node->type.tv = tarval_reachable;
2336 assert(selector->type.tv == tarval_top);
2337 if (tarval_UNKNOWN == tarval_top) {
2338 /* any condition based on Top is "!=" */
2339 node->type.tv = tarval_unreachable;
2341 node->type.tv = tarval_unreachable;
2345 assert(pnc == pn_Cond_false);
2347 if (selector->type.tv == tarval_b_false) {
2348 node->type.tv = tarval_reachable;
2349 } else if (selector->type.tv == tarval_b_true) {
2350 node->type.tv = tarval_unreachable;
2351 } else if (selector->type.tv == tarval_bottom) {
2352 node->type.tv = tarval_reachable;
2354 assert(selector->type.tv == tarval_top);
2355 if (tarval_UNKNOWN == tarval_top) {
2356 /* any condition based on Top is "!=" */
2357 node->type.tv = tarval_reachable;
2359 node->type.tv = tarval_unreachable;
2363 } /* compute_Proj_Cond */
2365 static void compute_Proj_Switch(node_t *node, ir_node *switchn)
2367 ir_node *proj = node->node;
2368 long pnc = get_Proj_proj(proj);
2369 ir_node *sel = get_Switch_selector(switchn);
2370 node_t *selector = get_irn_node(sel);
2372 /* see long comment in compute_Proj_Cond */
2373 if (node->type.tv == tarval_reachable)
2376 if (selector->type.tv == tarval_bottom) {
2377 node->type.tv = tarval_reachable;
2378 } else if (selector->type.tv == tarval_top) {
2379 if (tarval_UNKNOWN == tarval_top && pnc == pn_Switch_default) {
2380 /* a switch based of Top is always "default" */
2381 node->type.tv = tarval_reachable;
2383 node->type.tv = tarval_unreachable;
2386 long value = get_tarval_long(selector->type.tv);
2387 const ir_switch_table *table = get_Switch_table(switchn);
2388 size_t n_entries = ir_switch_table_get_n_entries(table);
2391 for (e = 0; e < n_entries; ++e) {
2392 const ir_switch_table_entry *entry
2393 = ir_switch_table_get_entry_const(table, e);
2394 ir_tarval *min = entry->min;
2395 ir_tarval *max = entry->max;
2397 if (selector->type.tv == min) {
2398 node->type.tv = entry->pn == pnc
2399 ? tarval_reachable : tarval_unreachable;
2403 long minval = get_tarval_long(min);
2404 long maxval = get_tarval_long(max);
2405 if (minval <= value && value <= maxval) {
2406 node->type.tv = entry->pn == pnc
2407 ? tarval_reachable : tarval_unreachable;
2413 /* no entry matched: default */
2415 = pnc == pn_Switch_default ? tarval_reachable : tarval_unreachable;
2420 * (Re-)compute the type for a Proj-Node.
2422 * @param node the node
2424 static void compute_Proj(node_t *node)
2426 ir_node *proj = node->node;
2427 ir_mode *mode = get_irn_mode(proj);
2428 node_t *block = get_irn_node(get_nodes_block(skip_Proj(proj)));
2429 ir_node *pred = get_Proj_pred(proj);
2431 if (block->type.tv == tarval_unreachable) {
2432 /* a Proj in a unreachable Block stay Top */
2433 node->type.tv = tarval_top;
2436 if (get_irn_node(pred)->type.tv == tarval_top && !is_Cond(pred) && !is_Switch(pred)) {
2437 /* if the predecessor is Top, its Proj follow */
2438 node->type.tv = tarval_top;
2442 if (mode == mode_M) {
2443 /* mode M is always bottom */
2444 node->type.tv = tarval_bottom;
2446 } else if (mode == mode_X) {
2447 /* handle mode_X nodes */
2448 switch (get_irn_opcode(pred)) {
2450 /* the Proj_X from the Start is always reachable.
2451 However this is already handled at the top. */
2452 node->type.tv = tarval_reachable;
2455 compute_Proj_Cond(node, pred);
2458 compute_Proj_Switch(node, pred);
2465 default_compute(node);
2466 } /* compute_Proj */
2469 * (Re-)compute the type for a Confirm.
2471 * @param node the node
2473 static void compute_Confirm(node_t *node)
2475 ir_node *confirm = node->node;
2476 node_t *pred = get_irn_node(get_Confirm_value(confirm));
2478 if (get_Confirm_relation(confirm) == ir_relation_equal) {
2479 node_t *bound = get_irn_node(get_Confirm_bound(confirm));
2481 if (is_con(bound->type)) {
2482 /* is equal to a constant */
2483 node->type = bound->type;
2487 /* a Confirm is a copy OR a Const */
2488 node->type = pred->type;
2489 } /* compute_Confirm */
2492 * (Re-)compute the type for a given node.
2494 * @param node the node
2496 static void compute(node_t *node)
2498 ir_node *irn = node->node;
2501 #ifndef VERIFY_MONOTONE
2503 * Once a node reaches bottom, the type cannot fall further
2504 * in the lattice and we can stop computation.
2505 * Do not take this exit if the monotony verifier is
2506 * enabled to catch errors.
2508 if (node->type.tv == tarval_bottom)
2512 if (!is_Block(irn)) {
2513 /* for pinned nodes, check its control input */
2514 if (get_irn_pinned(skip_Proj(irn)) == op_pin_state_pinned) {
2515 node_t *block = get_irn_node(get_nodes_block(irn));
2517 if (block->type.tv == tarval_unreachable) {
2518 node->type.tv = tarval_top;
2524 func = (compute_func)node->node->op->ops.generic;
2530 * Identity functions: Note that one might think that identity() is just a
2531 * synonym for equivalent_node(). While this is true, we cannot use it for the algorithm
2532 * here, because it expects that the identity node is one of the inputs, which is NOT
2533 * always true for equivalent_node() which can handle (and does sometimes) DAGs.
2534 * So, we have our own implementation, which copies some parts of equivalent_node()
2538 * Calculates the Identity for Phi nodes
2540 static node_t *identity_Phi(node_t *node)
2542 ir_node *phi = node->node;
2543 ir_node *block = get_nodes_block(phi);
2544 node_t *n_part = NULL;
2547 for (i = get_Phi_n_preds(phi) - 1; i >= 0; --i) {
2548 node_t *pred_X = get_irn_node(get_Block_cfgpred(block, i));
2550 if (pred_X->type.tv == tarval_reachable) {
2551 node_t *pred = get_irn_node(get_Phi_pred(phi, i));
2555 else if (n_part->part != pred->part) {
2556 /* incongruent inputs, not a follower */
2561 /* if n_part is NULL here, all inputs path are dead, the Phi computes
2562 * tarval_top, is in the TOP partition and should NOT being split! */
2563 assert(n_part != NULL);
2565 } /* identity_Phi */
2568 * Calculates the Identity for commutative 0 neutral nodes.
2570 static node_t *identity_comm_zero_binop(node_t *node)
2572 ir_node *op = node->node;
2573 node_t *a = get_irn_node(get_binop_left(op));
2574 node_t *b = get_irn_node(get_binop_right(op));
2575 ir_mode *mode = get_irn_mode(op);
2578 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
2579 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
2582 /* node: no input should be tarval_top, else the binop would be also
2583 * Top and not being split. */
2584 zero = get_mode_null(mode);
2585 if (a->type.tv == zero)
2587 if (b->type.tv == zero)
2590 } /* identity_comm_zero_binop */
2593 * Calculates the Identity for Shift nodes.
2595 static node_t *identity_shift(node_t *node)
2597 ir_node *op = node->node;
2598 node_t *b = get_irn_node(get_binop_right(op));
2599 ir_mode *mode = get_irn_mode(b->node);
2602 /* node: no input should be tarval_top, else the binop would be also
2603 * Top and not being split. */
2604 zero = get_mode_null(mode);
2605 if (b->type.tv == zero)
2606 return get_irn_node(get_binop_left(op));
2608 } /* identity_shift */
2611 * Calculates the Identity for Mul nodes.
2613 static node_t *identity_Mul(node_t *node)
2615 ir_node *op = node->node;
2616 node_t *a = get_irn_node(get_Mul_left(op));
2617 node_t *b = get_irn_node(get_Mul_right(op));
2618 ir_mode *mode = get_irn_mode(op);
2621 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
2622 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
2625 /* node: no input should be tarval_top, else the binop would be also
2626 * Top and not being split. */
2627 one = get_mode_one(mode);
2628 if (a->type.tv == one)
2630 if (b->type.tv == one)
2633 } /* identity_Mul */
2636 * Calculates the Identity for Sub nodes.
2638 static node_t *identity_Sub(node_t *node)
2640 ir_node *sub = node->node;
2641 node_t *b = get_irn_node(get_Sub_right(sub));
2642 ir_mode *mode = get_irn_mode(sub);
2644 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
2645 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
2648 /* node: no input should be tarval_top, else the binop would be also
2649 * Top and not being split. */
2650 if (b->type.tv == get_mode_null(mode))
2651 return get_irn_node(get_Sub_left(sub));
2653 } /* identity_Sub */
2656 * Calculates the Identity for And nodes.
2658 static node_t *identity_And(node_t *node)
2660 ir_node *andnode = node->node;
2661 node_t *a = get_irn_node(get_And_left(andnode));
2662 node_t *b = get_irn_node(get_And_right(andnode));
2663 ir_tarval *neutral = get_mode_all_one(get_irn_mode(andnode));
2665 /* node: no input should be tarval_top, else the And would be also
2666 * Top and not being split. */
2667 if (a->type.tv == neutral)
2669 if (b->type.tv == neutral)
2672 } /* identity_And */
2675 * Calculates the Identity for Confirm nodes.
2677 static node_t *identity_Confirm(node_t *node)
2679 ir_node *confirm = node->node;
2681 /* a Confirm is always a Copy */
2682 return get_irn_node(get_Confirm_value(confirm));
2683 } /* identity_Confirm */
2686 * Calculates the Identity for Mux nodes.
2688 static node_t *identity_Mux(node_t *node)
2690 ir_node *mux = node->node;
2691 node_t *t = get_irn_node(get_Mux_true(mux));
2692 node_t *f = get_irn_node(get_Mux_false(mux));
2695 if (t->part == f->part)
2698 /* for now, the 1-input identity is not supported */
2700 sel = get_irn_node(get_Mux_sel(mux));
2702 /* Mux sel input is mode_b, so it is always a tarval */
2703 if (sel->type.tv == tarval_b_true)
2705 if (sel->type.tv == tarval_b_false)
2709 } /* identity_Mux */
2712 * Calculates the Identity for nodes.
2714 static node_t *identity(node_t *node)
2716 ir_node *irn = node->node;
2718 switch (get_irn_opcode(irn)) {
2720 return identity_Phi(node);
2722 return identity_Mul(node);
2726 return identity_comm_zero_binop(node);
2731 return identity_shift(node);
2733 return identity_And(node);
2735 return identity_Sub(node);
2737 return identity_Confirm(node);
2739 return identity_Mux(node);
2746 * Node follower is a (new) follower of leader, segregate Leader
2749 static void segregate_def_use_chain_1(const ir_node *follower, node_t *leader)
2751 DB((dbg, LEVEL_2, "%+F is a follower of %+F\n", follower, leader->node));
2752 /* The leader edges must remain sorted, but follower edges can
2754 ir_node *l = leader->node;
2755 unsigned n = get_irn_n_outs(l);
2756 for (unsigned i = leader->n_followers; i < n; ++i) {
2757 if (l->o.out->edges[i].use == follower) {
2758 ir_def_use_edge t = l->o.out->edges[i];
2760 for (unsigned j = i; j-- > leader->n_followers; )
2761 l->o.out->edges[j+1] = l->o.out->edges[j];
2762 l->o.out->edges[leader->n_followers] = t;
2763 ++leader->n_followers;
2767 } /* segregate_def_use_chain_1 */
2770 * Node follower is a (new) follower segregate its Leader
2773 * @param follower the follower IR node
2775 static void segregate_def_use_chain(const ir_node *follower)
2779 for (i = get_irn_arity(follower) - 1; i >= 0; --i) {
2780 node_t *pred = get_irn_node(get_irn_n(follower, i));
2782 segregate_def_use_chain_1(follower, pred);
2784 } /* segregate_def_use_chain */
2787 * Propagate constant evaluation.
2789 * @param env the environment
2791 static void propagate(environment_t *env)
2795 lattice_elem_t old_type;
2797 unsigned n_fallen, old_type_was_T_or_C;
2799 while (env->cprop != NULL) {
2800 void *oldopcode = NULL;
2802 /* remove the first partition X from cprop */
2805 env->cprop = X->cprop_next;
2807 old_type_was_T_or_C = X->type_is_T_or_C;
2809 DB((dbg, LEVEL_2, "Propagate type on part%d\n", X->nr));
2813 int cprop_empty = list_empty(&X->cprop);
2814 int cprop_X_empty = list_empty(&X->cprop_X);
2816 if (cprop_empty && cprop_X_empty) {
2817 /* both cprop lists are empty */
2821 /* remove the first Node x from X.cprop */
2823 /* Get a node from the cprop_X list only if
2824 * all data nodes are processed.
2825 * This ensures, that all inputs of the Cond
2826 * predecessor are processed if its type is still Top.
2828 x = list_entry(X->cprop_X.next, node_t, cprop_list);
2830 x = list_entry(X->cprop.next, node_t, cprop_list);
2833 //assert(x->part == X);
2834 list_del(&x->cprop_list);
2837 if (x->is_follower && identity(x) == x) {
2838 /* check the opcode first */
2839 if (oldopcode == NULL) {
2840 oldopcode = lambda_opcode(get_first_node(X), env);
2842 if (oldopcode != lambda_opcode(x, env)) {
2843 if (x->on_fallen == 0) {
2844 /* different opcode -> x falls out of this partition */
2849 DB((dbg, LEVEL_2, "Add node %+F to fallen\n", x->node));
2853 /* x will make the follower -> leader transition */
2854 follower_to_leader(x);
2856 /* In case of a follower -> leader transition of a Phi node
2857 * we have to ensure that the current partition will be split
2858 * by lambda n.(n[i].partition).
2860 * This split may already happened before when some predecessors
2861 * of the Phi's Block are unreachable. Thus, we have to put the
2862 * current partition in the worklist to repeat the check.
2864 if (is_Phi(x->node) && ! x->part->on_worklist)
2865 add_to_worklist(x->part, env);
2868 /* compute a new type for x */
2870 DB((dbg, LEVEL_3, "computing type of %+F\n", x->node));
2872 if (x->type.tv != old_type.tv) {
2873 DB((dbg, LEVEL_2, "node %+F has changed type from %+F to %+F\n", x->node, old_type, x->type));
2874 verify_type(old_type, x);
2876 if (x->on_fallen == 0) {
2877 /* Add x to fallen. Nodes might fall from T -> const -> _|_, so check that they are
2878 not already on the list. */
2883 DB((dbg, LEVEL_2, "Add node %+F to fallen\n", x->node));
2885 for (unsigned i = get_irn_n_outs(x->node); i-- > 0; ) {
2886 ir_node *succ = get_irn_out(x->node, i);
2887 node_t *y = get_irn_node(succ);
2889 /* Add y to y.partition.cprop. */
2890 add_to_cprop(y, env);
2895 if (n_fallen > 0 && n_fallen != X->n_leader) {
2896 DB((dbg, LEVEL_2, "Splitting part%d by fallen\n", X->nr));
2897 Y = split(&X, fallen, env);
2899 * We have split out fallen node. The type of the result
2900 * partition is NOT set yet.
2902 Y->type_is_T_or_C = 0;
2906 /* remove the flags from the fallen list */
2907 for (x = fallen; x != NULL; x = x->next)
2910 if (old_type_was_T_or_C) {
2911 /* check if some nodes will make the leader -> follower transition */
2912 list_for_each_entry_safe(node_t, y, tmp, &Y->Leader, node_list) {
2913 if (y->type.tv != tarval_top && ! is_con(y->type)) {
2914 node_t *eq_node = identity(y);
2916 if (eq_node != y && eq_node->part == y->part) {
2917 DB((dbg, LEVEL_2, "Node %+F is a follower of %+F\n", y->node, eq_node->node));
2918 /* move to Follower */
2920 list_del(&y->node_list);
2921 list_add_tail(&y->node_list, &Y->Follower);
2924 segregate_def_use_chain(y->node);
2934 * Get the leader for a given node from its congruence class.
2936 * @param irn the node
2938 static ir_node *get_leader(node_t *node)
2940 partition_t *part = node->part;
2942 if (part->n_leader > 1 || node->is_follower) {
2943 if (node->is_follower) {
2944 DB((dbg, LEVEL_2, "Replacing follower %+F\n", node->node));
2947 DB((dbg, LEVEL_2, "Found congruence class for %+F\n", node->node));
2949 return get_first_node(part)->node;
2955 * Returns non-zero if a mode_T node has only one reachable output.
2957 static int only_one_reachable_proj(ir_node *n)
2961 for (unsigned i = get_irn_n_outs(n); i-- > 0; ) {
2962 ir_node *proj = get_irn_out(n, i);
2965 /* skip non-control flow Proj's */
2966 if (get_irn_mode(proj) != mode_X)
2969 node = get_irn_node(proj);
2970 if (node->type.tv == tarval_reachable) {
2976 } /* only_one_reachable_proj */
2979 * Return non-zero if the control flow predecessor node pred
2980 * is the only reachable control flow exit of its block.
2982 * @param pred the control flow exit
2983 * @param block the destination block
2985 static int can_exchange(ir_node *pred, ir_node *block)
2987 if (is_Start(pred) || get_Block_entity(block) != NULL)
2989 else if (is_Jmp(pred))
2991 else if (is_Raise(pred)) {
2992 /* Raise is a tuple and usually has only one reachable ProjX,
2993 * but it must not be eliminated like a Jmp */
2996 else if (get_irn_mode(pred) == mode_T) {
2997 /* if the predecessor block has more than one
2998 reachable outputs we cannot remove the block */
2999 return only_one_reachable_proj(pred);
3002 } /* can_exchange */
3005 * Block Post-Walker, apply the analysis results on control flow by
3006 * shortening Phi's and Block inputs.
3008 static void apply_cf(ir_node *block, void *ctx)
3010 environment_t *env = (environment_t*)ctx;
3011 node_t *node = get_irn_node(block);
3013 ir_node **ins, **in_X;
3014 ir_node *phi, *next;
3016 n = get_Block_n_cfgpreds(block);
3018 if (node->type.tv == tarval_unreachable) {
3021 for (i = n - 1; i >= 0; --i) {
3022 ir_node *pred = get_Block_cfgpred(block, i);
3024 if (! is_Bad(pred)) {
3025 ir_node *pred_block = get_nodes_block(skip_Proj(pred));
3026 if (!is_Bad(pred_block)) {
3027 node_t *pred_bl = get_irn_node(pred_block);
3029 if (pred_bl->flagged == 0) {
3030 pred_bl->flagged = 3;
3032 if (pred_bl->type.tv == tarval_reachable) {
3034 * We will remove an edge from block to its pred.
3035 * This might leave the pred block as an endless loop
3037 if (! is_backedge(block, i))
3038 keep_alive(pred_bl->node);
3045 ir_graph *const irg = get_Block_irg(block);
3046 if (block == get_irg_end_block(irg)) {
3047 /* Analysis found out that the end block is unreachable,
3048 * hence we remove all its control flow predecessors. */
3049 set_irn_in(block, 0, NULL);
3055 /* only one predecessor combine */
3056 ir_node *pred = skip_Proj(get_Block_cfgpred(block, 0));
3058 if (can_exchange(pred, block)) {
3059 ir_node *new_block = get_nodes_block(pred);
3060 DB((dbg, LEVEL_1, "Fuse %+F with %+F\n", block, new_block));
3061 DBG_OPT_COMBO(block, new_block, FS_OPT_COMBO_CF);
3062 exchange(block, new_block);
3063 node->node = new_block;
3069 NEW_ARR_A(ir_node *, in_X, n);
3071 for (i = 0; i < n; ++i) {
3072 ir_node *pred = get_Block_cfgpred(block, i);
3073 node_t *node = get_irn_node(pred);
3075 if (node->type.tv == tarval_reachable) {
3078 DB((dbg, LEVEL_1, "Removing dead input %d from %+F (%+F)\n", i, block, pred));
3079 if (! is_Bad(pred)) {
3080 ir_node *pred_block = get_nodes_block(skip_Proj(pred));
3081 if (!is_Bad(pred_block)) {
3082 node_t *pred_bl = get_irn_node(pred_block);
3084 if (!is_Bad(pred_bl->node) && pred_bl->flagged == 0) {
3085 pred_bl->flagged = 3;
3087 if (pred_bl->type.tv == tarval_reachable) {
3089 * We will remove an edge from block to its pred.
3090 * This might leave the pred block as an endless loop
3092 if (! is_backedge(block, i))
3093 keep_alive(pred_bl->node);
3104 NEW_ARR_A(ir_node *, ins, n);
3105 for (phi = get_Block_phis(block); phi != NULL; phi = next) {
3106 node_t *node = get_irn_node(phi);
3108 next = get_Phi_next(phi);
3109 if (is_tarval(node->type.tv) && tarval_is_constant(node->type.tv)) {
3110 /* this Phi is replaced by a constant */
3111 ir_tarval *tv = node->type.tv;
3112 ir_node *c = new_r_Const(current_ir_graph, tv);
3114 set_irn_node(c, node);
3116 DB((dbg, LEVEL_1, "%+F is replaced by %+F\n", phi, c));
3117 DBG_OPT_COMBO(phi, c, FS_OPT_COMBO_CONST);
3122 for (i = 0; i < n; ++i) {
3123 node_t *pred = get_irn_node(get_Block_cfgpred(block, i));
3125 if (pred->type.tv == tarval_reachable) {
3126 ins[j++] = get_Phi_pred(phi, i);
3130 /* this Phi is replaced by a single predecessor */
3131 ir_node *s = ins[0];
3132 node_t *phi_node = get_irn_node(phi);
3135 DB((dbg, LEVEL_1, "%+F is replaced by %+F because of cf change\n", phi, s));
3136 DBG_OPT_COMBO(phi, s, FS_OPT_COMBO_FOLLOWER);
3141 set_irn_in(phi, j, ins);
3149 /* this Block has only one live predecessor */
3150 ir_node *pred = skip_Proj(in_X[0]);
3152 if (can_exchange(pred, block)) {
3153 ir_node *new_block = get_nodes_block(pred);
3154 DBG_OPT_COMBO(block, new_block, FS_OPT_COMBO_CF);
3155 exchange(block, new_block);
3156 node->node = new_block;
3161 set_irn_in(block, k, in_X);
3166 * Exchange a node by its leader.
3167 * Beware: in rare cases the mode might be wrong here, for instance
3168 * AddP(x, NULL) is a follower of x, but with different mode.
3171 static void exchange_leader(ir_node *irn, ir_node *leader)
3173 ir_mode *mode = get_irn_mode(irn);
3174 if (mode != get_irn_mode(leader)) {
3175 /* The conv is a no-op, so we are free to place it
3176 * either in the block of the leader OR in irn's block.
3177 * Probably placing it into leaders block might reduce
3178 * the number of Conv due to CSE. */
3179 ir_node *block = get_nodes_block(leader);
3180 dbg_info *dbg = get_irn_dbg_info(irn);
3181 ir_node *nlead = new_rd_Conv(dbg, block, leader, mode);
3183 if (nlead != leader) {
3184 /* Note: this newly create irn has no node info because
3185 * it is created after the analysis. However, this node
3186 * replaces the node irn and should not be visited again,
3187 * so set its visited count to the count of irn.
3188 * Otherwise we might visited this node more than once if
3189 * irn had more than one user.
3191 set_irn_node(nlead, NULL);
3192 set_irn_visited(nlead, get_irn_visited(irn));
3196 exchange(irn, leader);
3197 } /* exchange_leader */
3200 * Check, if all users of a mode_M node are dead. Use
3201 * the Def-Use edges for this purpose, as they still
3202 * reflect the situation.
3204 static int all_users_are_dead(const ir_node *irn)
3206 unsigned n = get_irn_n_outs(irn);
3207 for (unsigned i = 0; i < n; ++i) {
3208 const ir_node *succ = get_irn_out(irn, i);
3209 const node_t *block = get_irn_node(get_nodes_block(succ));
3212 if (block->type.tv == tarval_unreachable) {
3213 /* block is unreachable */
3216 node = get_irn_node(succ);
3217 if (node->type.tv != tarval_top) {
3218 /* found a reachable user */
3222 /* all users are unreachable */
3224 } /* all_user_are_dead */
3227 * Walker: Find reachable mode_M nodes that have only
3228 * unreachable users. These nodes must be kept later.
3230 static void find_kept_memory(ir_node *irn, void *ctx)
3232 environment_t *env = (environment_t*)ctx;
3233 node_t *node, *block;
3235 if (get_irn_mode(irn) != mode_M)
3238 block = get_irn_node(get_nodes_block(irn));
3239 if (block->type.tv == tarval_unreachable)
3242 node = get_irn_node(irn);
3243 if (node->type.tv == tarval_top)
3246 /* ok, we found a live memory node. */
3247 if (all_users_are_dead(irn)) {
3248 DB((dbg, LEVEL_1, "%+F must be kept\n", irn));
3249 ARR_APP1(ir_node *, env->kept_memory, irn);
3251 } /* find_kept_memory */
3254 * Post-Walker, apply the analysis results;
3256 static void apply_result(ir_node *irn, void *ctx)
3258 environment_t *env = (environment_t*)ctx;
3259 node_t *node = get_irn_node(irn);
3261 if (is_Block(irn) || is_End(irn) || is_Bad(irn)) {
3262 /* blocks already handled, do not touch the End node */
3264 node_t *block = get_irn_node(get_nodes_block(irn));
3266 if (block->type.tv == tarval_unreachable) {
3267 ir_graph *irg = get_irn_irg(irn);
3268 ir_mode *mode = get_irn_mode(node->node);
3269 ir_node *bad = new_r_Bad(irg, mode);
3271 /* here, bad might already have a node, but this can be safely ignored
3272 as long as bad has at least ONE valid node */
3273 set_irn_node(bad, node);
3275 DB((dbg, LEVEL_1, "%+F is unreachable\n", irn));
3278 } else if (node->type.tv == tarval_top) {
3279 ir_mode *mode = get_irn_mode(irn);
3281 if (mode == mode_M) {
3282 /* never kill a mode_M node */
3284 ir_node *pred = get_Proj_pred(irn);
3285 node_t *pnode = get_irn_node(pred);
3287 if (pnode->type.tv == tarval_top) {
3288 /* skip the predecessor */
3289 ir_node *mem = get_memop_mem(pred);
3291 DB((dbg, LEVEL_1, "%+F computes Top, replaced by %+F\n", irn, mem));
3296 /* leave other nodes, especially PhiM */
3297 } else if (mode == mode_T) {
3298 /* Do not kill mode_T nodes, kill their Projs */
3299 } else if (! is_Unknown(irn)) {
3300 /* don't kick away Unknown's, they might be still needed */
3301 ir_node *unk = new_r_Unknown(current_ir_graph, mode);
3303 /* control flow should already be handled at apply_cf() */
3304 assert(mode != mode_X);
3306 /* see comment above */
3307 set_irn_node(unk, node);
3309 DB((dbg, LEVEL_1, "%+F computes Top\n", irn));
3314 else if (get_irn_mode(irn) == mode_X) {
3317 ir_node *cond = get_Proj_pred(irn);
3319 if (is_Cond(cond) || is_Switch(cond)) {
3320 if (only_one_reachable_proj(cond)) {
3321 ir_node *jmp = new_r_Jmp(block->node);
3322 set_irn_node(jmp, node);
3324 DB((dbg, LEVEL_1, "%+F is replaced by %+F\n", irn, jmp));
3325 DBG_OPT_COMBO(irn, jmp, FS_OPT_COMBO_CF);
3329 if (is_Switch(cond)) {
3330 node_t *sel = get_irn_node(get_Switch_selector(cond));
3331 ir_tarval *tv = sel->type.tv;
3333 if (is_tarval(tv) && tarval_is_constant(tv)) {
3334 /* The selector is a constant, but more
3335 * than one output is active: An unoptimized
3344 /* normal data node */
3345 if (is_tarval(node->type.tv) && tarval_is_constant(node->type.tv)) {
3346 ir_tarval *tv = node->type.tv;
3349 * Beware: never replace mode_T nodes by constants. Currently we must mark
3350 * mode_T nodes with constants, but do NOT replace them.
3352 if (! is_Const(irn) && get_irn_mode(irn) != mode_T) {
3353 /* can be replaced by a constant */
3354 ir_node *c = new_r_Const(current_ir_graph, tv);
3355 set_irn_node(c, node);
3357 DB((dbg, LEVEL_1, "%+F is replaced by %+F\n", irn, c));
3358 DBG_OPT_COMBO(irn, c, FS_OPT_COMBO_CONST);
3359 exchange_leader(irn, c);
3362 } else if (is_entity(node->type.sym.entity_p)) {
3363 if (! is_SymConst(irn)) {
3364 /* can be replaced by a SymConst */
3365 ir_node *symc = new_r_SymConst(current_ir_graph, get_irn_mode(irn), node->type.sym, symconst_addr_ent);
3366 set_irn_node(symc, node);
3369 DB((dbg, LEVEL_1, "%+F is replaced by %+F\n", irn, symc));
3370 DBG_OPT_COMBO(irn, symc, FS_OPT_COMBO_CONST);
3371 exchange_leader(irn, symc);
3374 } else if (is_Confirm(irn)) {
3375 /* Confirms are always follower, but do not kill them here */
3377 ir_node *leader = get_leader(node);
3379 if (leader != irn) {
3380 int non_strict_phi = 0;
3383 * Beware: Do not remove Phi(Unknown, ..., x, ..., Unknown)
3384 * as this might create non-strict programs.
3386 if (node->is_follower && is_Phi(irn) && !is_Unknown(leader)) {
3389 for (i = get_Phi_n_preds(irn) - 1; i >= 0; --i) {
3390 ir_node *pred = get_Phi_pred(irn, i);
3392 if (is_Unknown(pred)) {
3398 if (! non_strict_phi) {
3399 DB((dbg, LEVEL_1, "%+F from part%d is replaced by %+F\n", irn, node->part->nr, leader));
3400 if (node->is_follower)
3401 DBG_OPT_COMBO(irn, leader, FS_OPT_COMBO_FOLLOWER);
3403 DBG_OPT_COMBO(irn, leader, FS_OPT_COMBO_CONGRUENT);
3404 exchange_leader(irn, leader);
3411 } /* apply_result */
3414 * Fix the keep-alives by deleting unreachable ones.
3416 static void apply_end(ir_node *end, environment_t *env)
3418 int i, j, n = get_End_n_keepalives(end);
3419 ir_node **in = NULL;
3422 NEW_ARR_A(ir_node *, in, n);
3424 /* fix the keep alive */
3425 for (i = j = 0; i < n; i++) {
3426 ir_node *ka = get_End_keepalive(end, i);
3432 if (!is_Block(ka)) {
3433 block = get_nodes_block(ka);
3440 node = get_irn_node(block);
3441 if (node->type.tv != tarval_unreachable)
3445 set_End_keepalives(end, j, in);
3450 #define SET(code) op_##code->ops.generic = (op_func)compute_##code
3453 * sets the generic functions to compute.
3455 static void set_compute_functions(void)
3459 /* set the default compute function */
3460 for (i = 0, n = ir_get_n_opcodes(); i < n; ++i) {
3461 ir_op *op = ir_get_opcode(i);
3462 op->ops.generic = (op_func)default_compute;
3465 /* set specific functions */
3481 } /* set_compute_functions */
3486 static void add_memory_keeps(ir_node **kept_memory, size_t len)
3488 ir_node *end = get_irg_end(current_ir_graph);
3493 ir_nodeset_init(&set);
3495 /* check, if those nodes are already kept */
3496 for (i = get_End_n_keepalives(end) - 1; i >= 0; --i)
3497 ir_nodeset_insert(&set, get_End_keepalive(end, i));
3499 for (idx = 0; idx < len; ++idx) {
3500 ir_node *ka = kept_memory[idx];
3502 if (! ir_nodeset_contains(&set, ka)) {
3503 add_End_keepalive(end, ka);
3506 ir_nodeset_destroy(&set);
3507 } /* add_memory_keeps */
3509 void combo(ir_graph *irg)
3512 ir_node *initial_bl;
3514 ir_graph *rem = current_ir_graph;
3517 assure_irg_properties(irg,
3518 IR_GRAPH_PROPERTY_NO_BADS
3519 | IR_GRAPH_PROPERTY_CONSISTENT_OUTS
3520 | IR_GRAPH_PROPERTY_CONSISTENT_LOOPINFO);
3522 current_ir_graph = irg;
3524 /* register a debug mask */
3525 FIRM_DBG_REGISTER(dbg, "firm.opt.combo");
3527 DB((dbg, LEVEL_1, "Doing COMBO for %+F\n", irg));
3529 obstack_init(&env.obst);
3530 env.worklist = NULL;
3534 #ifdef DEBUG_libfirm
3535 env.dbg_list = NULL;
3537 env.opcode2id_map = new_set(cmp_opcode, iro_Last * 4);
3538 env.kept_memory = NEW_ARR_F(ir_node *, 0);
3539 env.end_idx = get_opt_global_cse() ? 0 : -1;
3540 env.lambda_input = 0;
3543 /* options driving the optimization */
3544 env.commutative = 1;
3545 env.opt_unknown = 1;
3547 /* we have our own value_of function */
3548 set_value_of_func(get_node_tarval);
3550 set_compute_functions();
3551 DEBUG_ONLY(part_nr = 0;)
3553 ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK | IR_RESOURCE_PHI_LIST);
3555 if (env.opt_unknown)
3556 tarval_UNKNOWN = tarval_top;
3558 tarval_UNKNOWN = tarval_bad;
3560 /* create the initial partition and place it on the work list */
3561 env.initial = new_partition(&env);
3562 add_to_worklist(env.initial, &env);
3563 irg_walk_graph(irg, create_initial_partitions, init_block_phis, &env);
3565 /* set the hook: from now, every node has a partition and a type */
3566 DEBUG_ONLY(set_dump_node_vcgattr_hook(dump_partition_hook);)
3568 /* all nodes on the initial partition have type Top */
3569 env.initial->type_is_T_or_C = 1;
3571 /* Place the START Node's partition on cprop.
3572 Place the START Node on its local worklist. */
3573 initial_bl = get_irg_start_block(irg);
3574 start = get_irn_node(initial_bl);
3575 add_to_cprop(start, &env);
3579 if (env.worklist != NULL)
3581 } while (env.cprop != NULL || env.worklist != NULL);
3583 dump_all_partitions(&env);
3584 check_all_partitions(&env);
3587 dump_ir_block_graph(irg, "-partition");
3590 /* apply the result */
3592 /* check, which nodes must be kept */
3593 irg_walk_graph(irg, NULL, find_kept_memory, &env);
3595 /* kill unreachable control flow */
3596 irg_block_walk_graph(irg, NULL, apply_cf, &env);
3597 /* Kill keep-alives of dead blocks: this speeds up apply_result()
3598 * and fixes assertion because dead cf to dead blocks is NOT removed by
3600 apply_end(get_irg_end(irg), &env);
3601 irg_walk_graph(irg, NULL, apply_result, &env);
3603 len = ARR_LEN(env.kept_memory);
3605 add_memory_keeps(env.kept_memory, len);
3608 DB((dbg, LEVEL_1, "Unoptimized Control Flow left"));
3611 ir_free_resources(irg, IR_RESOURCE_IRN_LINK | IR_RESOURCE_PHI_LIST);
3613 /* remove the partition hook */
3614 DEBUG_ONLY(set_dump_node_vcgattr_hook(NULL);)
3616 DEL_ARR_F(env.kept_memory);
3617 del_set(env.opcode2id_map);
3618 obstack_free(&env.obst, NULL);
3620 /* restore value_of() default behavior */
3621 set_value_of_func(NULL);
3622 current_ir_graph = rem;
3624 confirm_irg_properties(irg, IR_GRAPH_PROPERTIES_NONE);
3627 /* Creates an ir_graph pass for combo. */
3628 ir_graph_pass_t *combo_pass(const char *name)
3630 return def_graph_pass(name ? name : "combo", combo);
projects / libfirm / blob