2 * Copyright (C) 1995-2008 University of Karlsruhe. All right reserved.
4 * This file is part of libFirm.
6 * This file may be distributed and/or modified under the terms of the
7 * GNU General Public License version 2 as published by the Free Software
8 * Foundation and appearing in the file LICENSE.GPL included in the
9 * packaging of this file.
11 * Licensees holding valid libFirm Professional Edition licenses may use
12 * this file in accordance with the libFirm Commercial License.
13 * Agreement provided with the Software.
15 * This file is provided AS IS with NO WARRANTY OF ANY KIND, INCLUDING THE
16 * WARRANTY OF DESIGN, MERCHANTABILITY AND FITNESS FOR A PARTICULAR
22 * @brief Cliff Click's Combined Analysis/Optimization
23 * @author Michael Beck
26 * This is a slightly enhanced version of Cliff Clicks combo algorithm
27 * - support for commutative nodes is added, Add(a,b) and Add(b,a) ARE congruent
28 * - supports all Firm direct (by a data edge) identities except Mux
29 * (Mux can be a 2-input or 1-input identity, only 2-input is implemented yet)
30 * - supports Confirm nodes (handle them like Copies but do NOT remove them)
31 * - let Cmp nodes calculate Top like all othe data nodes: this would let
32 * Mux nodes to calculate Unknown instead of taking the true result
33 * - let Cond(Top) always select FALSE/default: This is tricky. Nodes are only reavaluated
34 * IFF the predecessor changed its type. Because nodes are initialized with Top
35 * this never happens, let all Proj(Cond) be unreachable.
36 * We avoid this condition by the same way we work around Phi: whenever a Block
37 * node is placed on the list, place its Cond nodes (and because they are Tuple
38 * all its Proj-nodes either on the cprop list)
39 * Especially, this changes the meaning of Click's example:
54 * using Click's version while is silent with our.
55 * - support for global congruences is implemented but not tested yet
57 * Note further that we use the terminology from Click's work here, which is different
58 * in some cases from Firm terminology. Especially, Click's type is a
59 * Firm tarval/entity, nevertheless we call it type here for "maximum compatibility".
65 #include "iroptimize.h"
72 #include "irgraph_t.h"
79 #include "iropt_dbg.h"
83 #include "irnodeset.h"
90 /* define this to check that all type translations are monotone */
91 #define VERIFY_MONOTONE
93 /* define this to check the consistency of partitions */
94 #define CHECK_PARTITIONS
96 typedef struct node_t node_t;
97 typedef struct partition_t partition_t;
98 typedef struct opcode_key_t opcode_key_t;
99 typedef struct listmap_entry_t listmap_entry_t;
101 /** The type of the compute function. */
102 typedef void (*compute_func)(node_t *node);
107 struct opcode_key_t {
108 ir_opcode code; /**< The Firm opcode. */
109 ir_mode *mode; /**< The mode of all nodes in the partition. */
110 int arity; /**< The arity of this opcode (needed for Phi etc. */
112 long proj; /**< For Proj nodes, its proj number */
113 ir_entity *ent; /**< For Sel Nodes, its entity */
114 int intVal; /**< For Conv/Div Nodes: strict/remainderless */
115 unsigned uintVal;/**< for Builtin: the kind */
116 ir_node *block; /**< for Block: itself */
117 void *ptr; /**< generic pointer for hash/cmp */
122 * An entry in the list_map.
124 struct listmap_entry_t {
125 void *id; /**< The id. */
126 node_t *list; /**< The associated list for this id. */
127 listmap_entry_t *next; /**< Link to the next entry in the map. */
130 /** We must map id's to lists. */
131 typedef struct listmap_t {
132 set *map; /**< Map id's to listmap_entry_t's */
133 listmap_entry_t *values; /**< List of all values in the map. */
137 * A lattice element. Because we handle constants and symbolic constants different, we
138 * have to use this union.
149 ir_node *node; /**< The IR-node itself. */
150 list_head node_list; /**< Double-linked list of leader/follower entries. */
151 list_head cprop_list; /**< Double-linked partition.cprop list. */
152 partition_t *part; /**< points to the partition this node belongs to */
153 node_t *next; /**< Next node on local list (partition.touched, fallen). */
154 node_t *race_next; /**< Next node on race list. */
155 lattice_elem_t type; /**< The associated lattice element "type". */
156 int max_user_input; /**< Maximum input number of Def-Use edges. */
157 int next_edge; /**< Index of the next Def-Use edge to use. */
158 int n_followers; /**< Number of Follower in the outs set. */
159 unsigned on_touched:1; /**< Set, if this node is on the partition.touched set. */
160 unsigned on_cprop:1; /**< Set, if this node is on the partition.cprop list. */
161 unsigned on_fallen:1; /**< Set, if this node is on the fallen list. */
162 unsigned is_follower:1; /**< Set, if this node is a follower. */
163 unsigned flagged:2; /**< 2 Bits, set if this node was visited by race 1 or 2. */
167 * A partition containing congruent nodes.
170 list_head Leader; /**< The head of partition Leader node list. */
171 list_head Follower; /**< The head of partition Follower node list. */
172 list_head cprop; /**< The head of partition.cprop list. */
173 list_head cprop_X; /**< The head of partition.cprop (Cond nodes and its Projs) list. */
174 partition_t *wl_next; /**< Next entry in the work list if any. */
175 partition_t *touched_next; /**< Points to the next partition in the touched set. */
176 partition_t *cprop_next; /**< Points to the next partition in the cprop list. */
177 partition_t *split_next; /**< Points to the next partition in the list that must be split by split_by(). */
178 node_t *touched; /**< The partition.touched set of this partition. */
179 unsigned n_leader; /**< Number of entries in this partition.Leader. */
180 unsigned n_touched; /**< Number of entries in the partition.touched. */
181 int max_user_inputs; /**< Maximum number of user inputs of all entries. */
182 unsigned on_worklist:1; /**< Set, if this partition is in the work list. */
183 unsigned on_touched:1; /**< Set, if this partition is on the touched set. */
184 unsigned on_cprop:1; /**< Set, if this partition is on the cprop list. */
185 unsigned type_is_T_or_C:1;/**< Set, if all nodes in this partition have type Top or Constant. */
187 partition_t *dbg_next; /**< Link all partitions for debugging */
188 unsigned nr; /**< A unique number for (what-)mapping, >0. */
192 typedef struct environment_t {
193 struct obstack obst; /**< obstack to allocate data structures. */
194 partition_t *worklist; /**< The work list. */
195 partition_t *cprop; /**< The constant propagation list. */
196 partition_t *touched; /**< the touched set. */
197 partition_t *initial; /**< The initial partition. */
198 set *opcode2id_map; /**< The opcodeMode->id map. */
199 pmap *type2id_map; /**< The type->id map. */
200 ir_node **kept_memory; /**< Array of memory nodes that must be kept. */
201 int end_idx; /**< -1 for local and 0 for global congruences. */
202 int lambda_input; /**< Captured argument for lambda_partition(). */
203 unsigned modified:1; /**< Set, if the graph was modified. */
204 unsigned unopt_cf:1; /**< If set, control flow is not optimized due to Unknown. */
205 /* options driving the optimization */
206 unsigned commutative:1; /**< Set, if commutation nodes should be handled specially. */
207 unsigned opt_unknown:1; /**< Set, if non-strict programs should be optimized. */
209 partition_t *dbg_list; /**< List of all partitions. */
213 /** Type of the what function. */
214 typedef void *(*what_func)(const node_t *node, environment_t *env);
216 #define get_irn_node(irn) ((node_t *)get_irn_link(irn))
217 #define set_irn_node(irn, node) set_irn_link(irn, node)
219 /* we do NOT use tarval_unreachable here, instead we use Top for this purpose */
220 #undef tarval_unreachable
221 #define tarval_unreachable tarval_top
224 /** The debug module handle. */
225 DEBUG_ONLY(static firm_dbg_module_t *dbg;)
227 /** The what reason. */
228 DEBUG_ONLY(static const char *what_reason;)
230 /** Next partition number. */
231 DEBUG_ONLY(static unsigned part_nr = 0);
233 /** The tarval returned by Unknown nodes: set to either tarval_bad OR tarval_top. */
234 static tarval *tarval_UNKNOWN;
237 static node_t *identity(node_t *node);
239 #ifdef CHECK_PARTITIONS
243 static void check_partition(const partition_t *T) {
247 list_for_each_entry(node_t, node, &T->Leader, node_list) {
248 assert(node->is_follower == 0);
249 assert(node->flagged == 0);
250 assert(node->part == T);
253 assert(n == T->n_leader);
255 list_for_each_entry(node_t, node, &T->Follower, node_list) {
256 assert(node->is_follower == 1);
257 assert(node->flagged == 0);
258 assert(node->part == T);
260 } /* check_partition */
263 * check that all leader nodes in the partition have the same opcode.
265 static void check_opcode(const partition_t *Z) {
270 list_for_each_entry(node_t, node, &Z->Leader, node_list) {
271 ir_node *irn = node->node;
274 key.code = get_irn_opcode(irn);
275 key.mode = get_irn_mode(irn);
276 key.arity = get_irn_arity(irn);
280 switch (get_irn_opcode(irn)) {
282 key.u.proj = get_Proj_proj(irn);
285 key.u.ent = get_Sel_entity(irn);
288 key.u.intVal = get_Conv_strict(irn);
291 key.u.intVal = get_Div_no_remainder(irn);
297 key.mode = get_Load_mode(irn);
300 key.u.intVal = get_Builtin_kind(irn);
307 assert((unsigned)key.code == get_irn_opcode(irn));
308 assert(key.mode == get_irn_mode(irn));
309 assert(key.arity == get_irn_arity(irn));
311 switch (get_irn_opcode(irn)) {
313 assert(key.u.proj == get_Proj_proj(irn));
316 assert(key.u.ent == get_Sel_entity(irn));
319 assert(key.u.intVal == get_Conv_strict(irn));
322 assert(key.u.intVal == get_Div_no_remainder(irn));
325 assert(key.u.block == irn);
328 assert(key.mode == get_Load_mode(irn));
331 assert(key.u.intVal == (int) get_Builtin_kind(irn));
340 static void check_all_partitions(environment_t *env) {
345 for (P = env->dbg_list; P != NULL; P = P->dbg_next) {
347 if (! P->type_is_T_or_C)
349 list_for_each_entry(node_t, node, &P->Follower, node_list) {
350 node_t *leader = identity(node);
352 assert(leader != node && leader->part == node->part);
363 static void do_check_list(const node_t *list, int ofs, const partition_t *Z) {
367 #define NEXT(e) *((const node_t **)((char *)(e) + (ofs)))
368 for (e = list; e != NULL; e = NEXT(e)) {
369 assert(e->part == Z);
377 } /* ido_check_list */
380 * Check a local list.
382 static void check_list(const node_t *list, const partition_t *Z) {
383 do_check_list(list, offsetof(node_t, next), Z);
387 #define check_partition(T)
388 #define check_list(list, Z)
389 #define check_all_partitions(env)
390 #endif /* CHECK_PARTITIONS */
393 static inline lattice_elem_t get_partition_type(const partition_t *X);
396 * Dump partition to output.
398 static void dump_partition(const char *msg, const partition_t *part) {
401 lattice_elem_t type = get_partition_type(part);
403 DB((dbg, LEVEL_2, "%s part%u%s (%u, %+F) {\n ",
404 msg, part->nr, part->type_is_T_or_C ? "*" : "",
405 part->n_leader, type));
406 list_for_each_entry(node_t, node, &part->Leader, node_list) {
407 DB((dbg, LEVEL_2, "%s%+F", first ? "" : ", ", node->node));
410 if (! list_empty(&part->Follower)) {
411 DB((dbg, LEVEL_2, "\n---\n "));
413 list_for_each_entry(node_t, node, &part->Follower, node_list) {
414 DB((dbg, LEVEL_2, "%s%+F", first ? "" : ", ", node->node));
418 DB((dbg, LEVEL_2, "\n}\n"));
419 } /* dump_partition */
424 static void do_dump_list(const char *msg, const node_t *node, int ofs) {
428 #define GET_LINK(p, ofs) *((const node_t **)((char *)(p) + (ofs)))
430 DB((dbg, LEVEL_3, "%s = {\n ", msg));
431 for (p = node; p != NULL; p = GET_LINK(p, ofs)) {
432 DB((dbg, LEVEL_3, "%s%+F", first ? "" : ", ", p->node));
435 DB((dbg, LEVEL_3, "\n}\n"));
443 static void dump_race_list(const char *msg, const node_t *list) {
444 do_dump_list(msg, list, offsetof(node_t, race_next));
445 } /* dump_race_list */
448 * Dumps a local list.
450 static void dump_list(const char *msg, const node_t *list) {
451 do_dump_list(msg, list, offsetof(node_t, next));
455 * Dump all partitions.
457 static void dump_all_partitions(const environment_t *env) {
458 const partition_t *P;
460 DB((dbg, LEVEL_2, "All partitions\n===============\n"));
461 for (P = env->dbg_list; P != NULL; P = P->dbg_next)
462 dump_partition("", P);
463 } /* dump_all_partitions */
468 static void dump_split_list(const partition_t *list) {
469 const partition_t *p;
471 DB((dbg, LEVEL_2, "Split by %s produced = {\n", what_reason));
472 for (p = list; p != NULL; p = p->split_next)
473 DB((dbg, LEVEL_2, "part%u, ", p->nr));
474 DB((dbg, LEVEL_2, "\n}\n"));
475 } /* dump_split_list */
478 * Dump partition and type for a node.
480 static int dump_partition_hook(FILE *F, ir_node *n, ir_node *local) {
481 ir_node *irn = local != NULL ? local : n;
482 node_t *node = get_irn_node(irn);
484 ir_fprintf(F, "info2 : \"partition %u type %+F\"\n", node->part->nr, node->type);
486 } /* dump_partition_hook */
489 #define dump_partition(msg, part)
490 #define dump_race_list(msg, list)
491 #define dump_list(msg, list)
492 #define dump_all_partitions(env)
493 #define dump_split_list(list)
496 #if defined(VERIFY_MONOTONE) && defined (DEBUG_libfirm)
498 * Verify that a type transition is monotone
500 static void verify_type(const lattice_elem_t old_type, node_t *node) {
501 if (old_type.tv == node->type.tv) {
505 if (old_type.tv == tarval_top) {
506 /* from Top down-to is always allowed */
509 if (node->type.tv == tarval_bottom || node->type.tv == tarval_reachable) {
513 panic("combo: wrong translation from %+F to %+F on node %+F", old_type, node->type, node->node);
517 #define verify_type(old_type, node)
521 * Compare two pointer values of a listmap.
523 static int listmap_cmp_ptr(const void *elt, const void *key, size_t size) {
524 const listmap_entry_t *e1 = elt;
525 const listmap_entry_t *e2 = key;
528 return e1->id != e2->id;
529 } /* listmap_cmp_ptr */
532 * Initializes a listmap.
534 * @param map the listmap
536 static void listmap_init(listmap_t *map) {
537 map->map = new_set(listmap_cmp_ptr, 16);
542 * Terminates a listmap.
544 * @param map the listmap
546 static void listmap_term(listmap_t *map) {
551 * Return the associated listmap entry for a given id.
553 * @param map the listmap
554 * @param id the id to search for
556 * @return the associated listmap entry for the given id
558 static listmap_entry_t *listmap_find(listmap_t *map, void *id) {
559 listmap_entry_t key, *entry;
564 entry = set_insert(map->map, &key, sizeof(key), HASH_PTR(id));
566 if (entry->list == NULL) {
567 /* a new entry, put into the list */
568 entry->next = map->values;
575 * Calculate the hash value for an opcode map entry.
577 * @param entry an opcode map entry
579 * @return a hash value for the given opcode map entry
581 static unsigned opcode_hash(const opcode_key_t *entry) {
582 return (entry->mode - (ir_mode *)0) * 9 + entry->code + entry->u.proj * 3 + HASH_PTR(entry->u.ptr) + entry->arity;
586 * Compare two entries in the opcode map.
588 static int cmp_opcode(const void *elt, const void *key, size_t size) {
589 const opcode_key_t *o1 = elt;
590 const opcode_key_t *o2 = key;
593 return o1->code != o2->code || o1->mode != o2->mode ||
594 o1->arity != o2->arity ||
595 o1->u.proj != o2->u.proj ||
596 o1->u.intVal != o2->u.intVal || /* this already checks uIntVal */
597 o1->u.ptr != o2->u.ptr;
601 * Compare two Def-Use edges for input position.
603 static int cmp_def_use_edge(const void *a, const void *b) {
604 const ir_def_use_edge *ea = a;
605 const ir_def_use_edge *eb = b;
607 /* no overrun, because range is [-1, MAXINT] */
608 return ea->pos - eb->pos;
609 } /* cmp_def_use_edge */
612 * We need the Def-Use edges sorted.
614 static void sort_irn_outs(node_t *node) {
615 ir_node *irn = node->node;
616 int n_outs = get_irn_n_outs(irn);
619 qsort(&irn->out[1], n_outs, sizeof(irn->out[0]), cmp_def_use_edge);
621 node->max_user_input = irn->out[n_outs].pos;
622 } /* sort_irn_outs */
625 * Return the type of a node.
627 * @param irn an IR-node
629 * @return the associated type of this node
631 static inline lattice_elem_t get_node_type(const ir_node *irn) {
632 return get_irn_node(irn)->type;
633 } /* get_node_type */
636 * Return the tarval of a node.
638 * @param irn an IR-node
640 * @return the associated type of this node
642 static inline tarval *get_node_tarval(const ir_node *irn) {
643 lattice_elem_t type = get_node_type(irn);
645 if (is_tarval(type.tv))
647 return tarval_bottom;
648 } /* get_node_type */
651 * Add a partition to the worklist.
653 static inline void add_to_worklist(partition_t *X, environment_t *env) {
654 assert(X->on_worklist == 0);
655 DB((dbg, LEVEL_2, "Adding part%d to worklist\n", X->nr));
656 X->wl_next = env->worklist;
659 } /* add_to_worklist */
662 * Create a new empty partition.
664 * @param env the environment
666 * @return a newly allocated partition
668 static inline partition_t *new_partition(environment_t *env) {
669 partition_t *part = obstack_alloc(&env->obst, sizeof(*part));
671 INIT_LIST_HEAD(&part->Leader);
672 INIT_LIST_HEAD(&part->Follower);
673 INIT_LIST_HEAD(&part->cprop);
674 INIT_LIST_HEAD(&part->cprop_X);
675 part->wl_next = NULL;
676 part->touched_next = NULL;
677 part->cprop_next = NULL;
678 part->split_next = NULL;
679 part->touched = NULL;
682 part->max_user_inputs = 0;
683 part->on_worklist = 0;
684 part->on_touched = 0;
686 part->type_is_T_or_C = 0;
688 part->dbg_next = env->dbg_list;
689 env->dbg_list = part;
690 part->nr = part_nr++;
694 } /* new_partition */
697 * Get the first node from a partition.
699 static inline node_t *get_first_node(const partition_t *X) {
700 return list_entry(X->Leader.next, node_t, node_list);
701 } /* get_first_node */
704 * Return the type of a partition (assuming partition is non-empty and
705 * all elements have the same type).
707 * @param X a partition
709 * @return the type of the first element of the partition
711 static inline lattice_elem_t get_partition_type(const partition_t *X) {
712 const node_t *first = get_first_node(X);
714 } /* get_partition_type */
717 * Creates a partition node for the given IR-node and place it
718 * into the given partition.
720 * @param irn an IR-node
721 * @param part a partition to place the node in
722 * @param env the environment
724 * @return the created node
726 static node_t *create_partition_node(ir_node *irn, partition_t *part, environment_t *env) {
727 /* create a partition node and place it in the partition */
728 node_t *node = obstack_alloc(&env->obst, sizeof(*node));
730 INIT_LIST_HEAD(&node->node_list);
731 INIT_LIST_HEAD(&node->cprop_list);
735 node->race_next = NULL;
736 node->type.tv = tarval_top;
737 node->max_user_input = 0;
739 node->n_followers = 0;
740 node->on_touched = 0;
743 node->is_follower = 0;
745 set_irn_node(irn, node);
747 list_add_tail(&node->node_list, &part->Leader);
751 } /* create_partition_node */
754 * Pre-Walker, initialize all Nodes' type to U or top and place
755 * all nodes into the TOP partition.
757 static void create_initial_partitions(ir_node *irn, void *ctx) {
758 environment_t *env = ctx;
759 partition_t *part = env->initial;
762 node = create_partition_node(irn, part, env);
764 if (node->max_user_input > part->max_user_inputs)
765 part->max_user_inputs = node->max_user_input;
768 set_Block_phis(irn, NULL);
770 } /* create_initial_partitions */
773 * Post-Walker, collect all Block-Phi lists, set Cond.
775 static void init_block_phis(ir_node *irn, void *ctx) {
779 add_Block_phi(get_nodes_block(irn), irn);
781 } /* init_block_phis */
784 * Add a node to the entry.partition.touched set and
785 * node->partition to the touched set if not already there.
788 * @param env the environment
790 static inline void add_to_touched(node_t *y, environment_t *env) {
791 if (y->on_touched == 0) {
792 partition_t *part = y->part;
794 y->next = part->touched;
799 if (part->on_touched == 0) {
800 part->touched_next = env->touched;
802 part->on_touched = 1;
805 check_list(part->touched, part);
807 } /* add_to_touched */
810 * Place a node on the cprop list.
813 * @param env the environment
815 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;
823 /* place Conds and all its Projs on the cprop_X list */
824 if (is_Cond(skip_Proj(irn)))
825 list_add_tail(&y->cprop_list, &Y->cprop_X);
827 list_add_tail(&y->cprop_list, &Y->cprop);
830 DB((dbg, LEVEL_3, "Add %+F to part%u.cprop\n", y->node, Y->nr));
832 /* place its partition on the cprop list */
833 if (Y->on_cprop == 0) {
834 Y->cprop_next = env->cprop;
840 if (get_irn_mode(irn) == mode_T) {
841 /* mode_T nodes always produce tarval_bottom, so we must explicitly
842 add it's Proj's to get constant evaluation to work */
845 for (i = get_irn_n_outs(irn) - 1; i >= 0; --i) {
846 node_t *proj = get_irn_node(get_irn_out(irn, i));
848 add_to_cprop(proj, env);
850 } else if (is_Block(irn)) {
851 /* Due to the way we handle Phi's, we must place all Phis of a block on the list
852 * if someone placed the block. The Block is only placed if the reachability
853 * changes, and this must be re-evaluated in compute_Phi(). */
855 for (phi = get_Block_phis(irn); phi != NULL; phi = get_Phi_next(phi)) {
856 node_t *p = get_irn_node(phi);
857 add_to_cprop(p, env);
863 * Update the worklist: If Z is on worklist then add Z' to worklist.
864 * Else add the smaller of Z and Z' to worklist.
866 * @param Z the Z partition
867 * @param Z_prime the Z' partition, a previous part of Z
868 * @param env the environment
870 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) {
884 ir_node *irn = x->node;
887 for (i = get_irn_arity(irn) - 1; i >= 0; --i) {
888 node_t *pred = get_irn_node(get_irn_n(irn, i));
893 n = get_irn_n_outs(p);
894 for (j = 1; j <= pred->n_followers; ++j) {
895 if (p->out[j].pos == i && p->out[j].use == irn) {
896 /* found a follower edge to x, move it to the Leader */
897 ir_def_use_edge edge = p->out[j];
899 /* remove this edge from the Follower set */
900 p->out[j] = p->out[pred->n_followers];
903 /* sort it into the leader set */
904 for (k = pred->n_followers + 2; k <= n; ++k) {
905 if (p->out[k].pos >= edge.pos)
907 p->out[k - 1] = p->out[k];
909 /* place the new edge here */
910 p->out[k - 1] = edge;
912 /* edge found and moved */
917 } /* move_edges_to_leader */
920 * Split a partition that has NO followers by a local list.
922 * @param Z partition to split
923 * @param g a (non-empty) node list
924 * @param env the environment
926 * @return a new partition containing the nodes of g
928 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 update_worklist(Z, Z_prime, env);
968 dump_partition("Now ", Z);
969 dump_partition("Created new ", Z_prime);
971 } /* split_no_followers */
974 * Make the Follower -> Leader transition for a node.
978 static void follower_to_leader(node_t *n) {
979 assert(n->is_follower == 1);
981 DB((dbg, LEVEL_2, "%+F make the follower -> leader transition\n", n->node));
983 move_edges_to_leader(n);
984 list_del(&n->node_list);
985 list_add_tail(&n->node_list, &n->part->Leader);
987 } /* follower_to_leader */
990 * The environment for one race step.
992 typedef struct step_env {
993 node_t *initial; /**< The initial node list. */
994 node_t *unwalked; /**< The unwalked node list. */
995 node_t *walked; /**< The walked node list. */
996 int index; /**< Next index of Follower use_def edge. */
997 unsigned side; /**< side number. */
1001 * Return non-zero, if a input is a real follower
1003 * @param irn the node to check
1004 * @param input number of the input
1006 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) {
1071 if (env->initial != NULL) {
1072 /* Move node from initial to unwalked */
1074 env->initial = n->race_next;
1076 n->race_next = env->unwalked;
1082 while (env->unwalked != NULL) {
1083 /* let n be the first node in unwalked */
1085 while (env->index < n->n_followers) {
1086 const ir_def_use_edge *edge = &n->node->out[1 + env->index];
1088 /* let m be n.F.def_use[index] */
1089 node_t *m = get_irn_node(edge->use);
1091 assert(m->is_follower);
1093 * Some inputs, like the get_Confirm_bound are NOT
1094 * real followers, sort them out.
1096 if (! is_real_follower(m->node, edge->pos)) {
1102 /* only followers from our partition */
1103 if (m->part != n->part)
1106 if ((m->flagged & env->side) == 0) {
1107 m->flagged |= env->side;
1109 if (m->flagged != 3) {
1110 /* visited the first time */
1111 /* add m to unwalked not as first node (we might still need to
1112 check for more follower node */
1113 m->race_next = n->race_next;
1117 /* else already visited by the other side and on the other list */
1120 /* move n to walked */
1121 env->unwalked = n->race_next;
1122 n->race_next = env->walked;
1130 * Clear the flags from a list and check for
1131 * nodes that where touched from both sides.
1133 * @param list the list
1135 static int clear_flags(node_t *list) {
1139 for (n = list; n != NULL; n = n->race_next) {
1140 if (n->flagged == 3) {
1141 /* we reach a follower from both sides, this will split congruent
1142 * inputs and make it a leader. */
1143 follower_to_leader(n);
1152 * Split a partition by a local list using the race.
1154 * @param pX pointer to the partition to split, might be changed!
1155 * @param gg a (non-empty) node list
1156 * @param env the environment
1158 * @return a new partition containing the nodes of gg
1160 static partition_t *split(partition_t **pX, node_t *gg, environment_t *env) {
1161 partition_t *X = *pX;
1162 partition_t *X_prime;
1165 node_t *g, *h, *node, *t;
1166 int max_input, transitions, winner, shf;
1168 DEBUG_ONLY(static int run = 0;)
1170 DB((dbg, LEVEL_2, "Run %d ", run++));
1171 if (list_empty(&X->Follower)) {
1172 /* if the partition has NO follower, we can use the fast
1173 splitting algorithm. */
1174 return split_no_followers(X, gg, env);
1176 /* else do the race */
1178 dump_partition("Splitting ", X);
1179 dump_list("by list ", gg);
1181 INIT_LIST_HEAD(&tmp);
1183 /* Remove gg from X.Leader and put into g */
1185 for (node = gg; node != NULL; node = node->next) {
1186 assert(node->part == X);
1187 assert(node->is_follower == 0);
1189 list_del(&node->node_list);
1190 list_add_tail(&node->node_list, &tmp);
1191 node->race_next = g;
1196 list_for_each_entry(node_t, node, &X->Leader, node_list) {
1197 node->race_next = h;
1200 /* restore X.Leader */
1201 list_splice(&tmp, &X->Leader);
1203 senv[0].initial = g;
1204 senv[0].unwalked = NULL;
1205 senv[0].walked = NULL;
1209 senv[1].initial = h;
1210 senv[1].unwalked = NULL;
1211 senv[1].walked = NULL;
1216 * Some informations on the race that are not stated clearly in Click's
1218 * 1) A follower stays on the side that reach him first.
1219 * 2) If the other side reches a follower, if will be converted to
1220 * a leader. /This must be done after the race is over, else the
1221 * edges we are iterating on are renumbered./
1222 * 3) /New leader might end up on both sides./
1223 * 4) /If one side ends up with new Leaders, we must ensure that
1224 * they can split out by opcode, hence we have to put _every_
1225 * partition with new Leader nodes on the cprop list, as
1226 * opcode splitting is done by split_by() at the end of
1227 * constant propagation./
1230 if (step(&senv[0])) {
1234 if (step(&senv[1])) {
1239 assert(senv[winner].initial == NULL);
1240 assert(senv[winner].unwalked == NULL);
1242 /* clear flags from walked/unwalked */
1244 transitions = clear_flags(senv[0].unwalked) << shf;
1245 transitions |= clear_flags(senv[0].walked) << shf;
1247 transitions |= clear_flags(senv[1].unwalked) << shf;
1248 transitions |= clear_flags(senv[1].walked) << shf;
1250 dump_race_list("winner ", senv[winner].walked);
1252 /* Move walked_{winner} to a new partition, X'. */
1253 X_prime = new_partition(env);
1256 for (node = senv[winner].walked; node != NULL; node = node->race_next) {
1257 list_del(&node->node_list);
1258 node->part = X_prime;
1259 if (node->is_follower) {
1260 list_add_tail(&node->node_list, &X_prime->Follower);
1262 list_add_tail(&node->node_list, &X_prime->Leader);
1265 if (node->max_user_input > max_input)
1266 max_input = node->max_user_input;
1268 X_prime->n_leader = n;
1269 X_prime->max_user_inputs = max_input;
1270 X->n_leader -= X_prime->n_leader;
1272 /* for now, copy the type info tag, it will be adjusted in split_by(). */
1273 X_prime->type_is_T_or_C = X->type_is_T_or_C;
1276 * Even if a follower was not checked by both sides, it might have
1277 * loose its congruence, so we need to check this case for all follower.
1279 list_for_each_entry_safe(node_t, node, t, &X_prime->Follower, node_list) {
1280 if (identity(node) == node) {
1281 follower_to_leader(node);
1287 check_partition(X_prime);
1289 /* X' is the smaller part */
1290 add_to_worklist(X_prime, env);
1293 * If there where follower to leader transitions, ensure that the nodes
1294 * can be split out if necessary.
1296 if (transitions & 1) {
1297 /* place winner partition on the cprop list */
1298 if (X_prime->on_cprop == 0) {
1299 X_prime->cprop_next = env->cprop;
1300 env->cprop = X_prime;
1301 X_prime->on_cprop = 1;
1304 if (transitions & 2) {
1305 /* place other partition on the cprop list */
1306 if (X->on_cprop == 0) {
1307 X->cprop_next = env->cprop;
1313 dump_partition("Now ", X);
1314 dump_partition("Created new ", X_prime);
1316 /* we have to ensure that the partition containing g is returned */
1326 * Returns non-zero if the i'th input of a Phi node is live.
1328 * @param phi a Phi-node
1329 * @param i an input number
1331 * @return non-zero if the i'th input of the given Phi node is live
1333 static int is_live_input(ir_node *phi, int i) {
1335 ir_node *block = get_nodes_block(phi);
1336 ir_node *pred = get_Block_cfgpred(block, i);
1337 lattice_elem_t type = get_node_type(pred);
1339 return type.tv != tarval_unreachable;
1341 /* else it's the control input, always live */
1343 } /* is_live_input */
1346 * Return non-zero if a type is a constant.
1348 static int is_constant_type(lattice_elem_t type) {
1349 if (type.tv != tarval_bottom && type.tv != tarval_top)
1352 } /* is_constant_type */
1355 * Check whether a type is neither Top or a constant.
1356 * Note: U is handled like Top here, R is a constant.
1358 * @param type the type to check
1360 static int type_is_neither_top_nor_const(const lattice_elem_t type) {
1361 if (is_tarval(type.tv)) {
1362 if (type.tv == tarval_top)
1364 if (tarval_is_constant(type.tv))
1371 } /* type_is_neither_top_nor_const */
1374 * Collect nodes to the touched list.
1376 * @param list the list which contains the nodes that must be evaluated
1377 * @param idx the index of the def_use edge to evaluate
1378 * @param env the environment
1380 static void collect_touched(list_head *list, int idx, environment_t *env) {
1382 int end_idx = env->end_idx;
1384 list_for_each_entry(node_t, x, list, node_list) {
1388 /* leader edges start AFTER follower edges */
1389 x->next_edge = x->n_followers + 1;
1391 num_edges = get_irn_n_outs(x->node);
1393 /* for all edges in x.L.def_use_{idx} */
1394 while (x->next_edge <= num_edges) {
1395 const ir_def_use_edge *edge = &x->node->out[x->next_edge];
1398 /* check if we have necessary edges */
1399 if (edge->pos > idx)
1406 /* only non-commutative nodes */
1407 if (env->commutative &&
1408 (idx == 0 || idx == 1) && is_op_commutative(get_irn_op(succ)))
1411 /* ignore the "control input" for non-pinned nodes
1412 if we are running in GCSE mode */
1413 if (idx < end_idx && get_irn_pinned(succ) != op_pin_state_pinned)
1416 y = get_irn_node(succ);
1417 assert(get_irn_n(succ, idx) == x->node);
1419 /* ignore block edges touching followers */
1420 if (idx == -1 && y->is_follower)
1423 if (is_constant_type(y->type)) {
1424 ir_opcode code = get_irn_opcode(succ);
1425 if (code == iro_Sub || code == iro_Cmp)
1426 add_to_cprop(y, env);
1429 /* Partitions of constants should not be split simply because their Nodes have unequal
1430 functions or incongruent inputs. */
1431 if (type_is_neither_top_nor_const(y->type) &&
1432 (! is_Phi(y->node) || is_live_input(y->node, idx))) {
1433 add_to_touched(y, env);
1437 } /* collect_touched */
1440 * Collect commutative nodes to the touched list.
1442 * @param X the partition of the list
1443 * @param list the list which contains the nodes that must be evaluated
1444 * @param env the environment
1446 static void collect_commutative_touched(partition_t *X, list_head *list, environment_t *env) {
1451 list_for_each_entry(node_t, x, list, node_list) {
1454 num_edges = get_irn_n_outs(x->node);
1456 x->next_edge = x->n_followers + 1;
1458 /* for all edges in x.L.def_use_{idx} */
1459 while (x->next_edge <= num_edges) {
1460 const ir_def_use_edge *edge = &x->node->out[x->next_edge];
1463 /* check if we have necessary edges */
1473 /* only commutative nodes */
1474 if (!is_op_commutative(get_irn_op(succ)))
1477 y = get_irn_node(succ);
1478 if (is_constant_type(y->type)) {
1479 ir_opcode code = get_irn_opcode(succ);
1480 if (code == iro_Eor)
1481 add_to_cprop(y, env);
1484 /* Partitions of constants should not be split simply because their Nodes have unequal
1485 functions or incongruent inputs. */
1486 if (type_is_neither_top_nor_const(y->type)) {
1487 int other_idx = edge->pos ^ 1;
1488 node_t *other = get_irn_node(get_irn_n(succ, other_idx));
1489 int equal = X == other->part;
1492 * Note: op(a, a) is NOT congruent to op(a, b).
1493 * So, either all touch nodes must have both inputs congruent,
1494 * or not. We decide this by the first occurred node.
1500 if (both_input == equal)
1501 add_to_touched(y, env);
1505 } /* collect_commutative_touched */
1508 * Split the partitions if caused by the first entry on the worklist.
1510 * @param env the environment
1512 static void cause_splits(environment_t *env) {
1513 partition_t *X, *Z, *N;
1516 /* remove the first partition from the worklist */
1518 env->worklist = X->wl_next;
1521 dump_partition("Cause_split: ", X);
1523 if (env->commutative) {
1524 /* handle commutative nodes first */
1526 /* empty the touched set: already done, just clear the list */
1527 env->touched = NULL;
1529 collect_commutative_touched(X, &X->Leader, env);
1530 collect_commutative_touched(X, &X->Follower, env);
1532 for (Z = env->touched; Z != NULL; Z = N) {
1534 node_t *touched = Z->touched;
1535 unsigned n_touched = Z->n_touched;
1537 assert(Z->touched != NULL);
1539 /* beware, split might change Z */
1540 N = Z->touched_next;
1542 /* remove it from the touched set */
1545 /* Empty local Z.touched. */
1546 for (e = touched; e != NULL; e = e->next) {
1547 assert(e->is_follower == 0);
1553 if (0 < n_touched && n_touched < Z->n_leader) {
1554 DB((dbg, LEVEL_2, "Split part%d by touched\n", Z->nr));
1555 split(&Z, touched, env);
1557 assert(n_touched <= Z->n_leader);
1561 /* combine temporary leader and follower list */
1562 for (idx = -1; idx <= X->max_user_inputs; ++idx) {
1563 /* empty the touched set: already done, just clear the list */
1564 env->touched = NULL;
1566 collect_touched(&X->Leader, idx, env);
1567 collect_touched(&X->Follower, idx, env);
1569 for (Z = env->touched; Z != NULL; Z = N) {
1571 node_t *touched = Z->touched;
1572 unsigned n_touched = Z->n_touched;
1574 assert(Z->touched != NULL);
1576 /* beware, split might change Z */
1577 N = Z->touched_next;
1579 /* remove it from the touched set */
1582 /* Empty local Z.touched. */
1583 for (e = touched; e != NULL; e = e->next) {
1584 assert(e->is_follower == 0);
1590 if (0 < n_touched && n_touched < Z->n_leader) {
1591 DB((dbg, LEVEL_2, "Split part%d by touched\n", Z->nr));
1592 split(&Z, touched, env);
1594 assert(n_touched <= Z->n_leader);
1597 } /* cause_splits */
1600 * Implements split_by_what(): Split a partition by characteristics given
1601 * by the what function.
1603 * @param X the partition to split
1604 * @param What a function returning an Id for every node of the partition X
1605 * @param P a list to store the result partitions
1606 * @param env the environment
1610 static partition_t *split_by_what(partition_t *X, what_func What,
1611 partition_t **P, environment_t *env) {
1614 listmap_entry_t *iter;
1617 /* Let map be an empty mapping from the range of What to (local) list of Nodes. */
1619 list_for_each_entry(node_t, x, &X->Leader, node_list) {
1620 void *id = What(x, env);
1621 listmap_entry_t *entry;
1624 /* input not allowed, ignore */
1627 /* Add x to map[What(x)]. */
1628 entry = listmap_find(&map, id);
1629 x->next = entry->list;
1632 /* Let P be a set of Partitions. */
1634 /* for all sets S except one in the range of map do */
1635 for (iter = map.values; iter != NULL; iter = iter->next) {
1636 if (iter->next == NULL) {
1637 /* this is the last entry, ignore */
1642 /* Add SPLIT( X, S ) to P. */
1643 DB((dbg, LEVEL_2, "Split part%d by WHAT = %s\n", X->nr, what_reason));
1644 R = split(&X, S, env);
1654 } /* split_by_what */
1656 /** lambda n.(n.type) */
1657 static void *lambda_type(const node_t *node, environment_t *env) {
1659 return node->type.tv;
1662 /** lambda n.(n.opcode) */
1663 static void *lambda_opcode(const node_t *node, environment_t *env) {
1664 opcode_key_t key, *entry;
1665 ir_node *irn = node->node;
1667 key.code = get_irn_opcode(irn);
1668 key.mode = get_irn_mode(irn);
1669 key.arity = get_irn_arity(irn);
1673 switch (get_irn_opcode(irn)) {
1675 key.u.proj = get_Proj_proj(irn);
1678 key.u.ent = get_Sel_entity(irn);
1681 key.u.intVal = get_Conv_strict(irn);
1684 key.u.intVal = get_Div_no_remainder(irn);
1688 * Some ugliness here: Two Blocks having the same
1689 * IJmp predecessor would be congruent, which of course is wrong.
1690 * We fix it by never letting blocks be congruent
1691 * which cannot be detected by combo either.
1696 key.mode = get_Load_mode(irn);
1699 key.u.intVal = get_Builtin_kind(irn);
1705 entry = set_insert(env->opcode2id_map, &key, sizeof(key), opcode_hash(&key));
1707 } /* lambda_opcode */
1709 /** lambda n.(n[i].partition) */
1710 static void *lambda_partition(const node_t *node, environment_t *env) {
1711 ir_node *skipped = skip_Proj(node->node);
1714 int i = env->lambda_input;
1716 if (i >= get_irn_arity(node->node)) {
1718 * We are outside the allowed range: This can happen even
1719 * if we have split by opcode first: doing so might move Followers
1720 * to Leaders and those will have a different opcode!
1721 * Note that in this case the partition is on the cprop list and will be
1727 /* ignore the "control input" for non-pinned nodes
1728 if we are running in GCSE mode */
1729 if (i < env->end_idx && get_irn_pinned(skipped) != op_pin_state_pinned)
1732 pred = i == -1 ? get_irn_n(skipped, i) : get_irn_n(node->node, i);
1733 p = get_irn_node(pred);
1735 } /* lambda_partition */
1737 /** lambda n.(n[i].partition) for commutative nodes */
1738 static void *lambda_commutative_partition(const node_t *node, environment_t *env) {
1739 ir_node *irn = node->node;
1740 ir_node *skipped = skip_Proj(irn);
1741 ir_node *pred, *left, *right;
1743 partition_t *pl, *pr;
1744 int i = env->lambda_input;
1746 if (i >= get_irn_arity(node->node)) {
1748 * We are outside the allowed range: This can happen even
1749 * if we have split by opcode first: doing so might move Followers
1750 * to Leaders and those will have a different opcode!
1751 * Note that in this case the partition is on the cprop list and will be
1757 /* ignore the "control input" for non-pinned nodes
1758 if we are running in GCSE mode */
1759 if (i < env->end_idx && get_irn_pinned(skipped) != op_pin_state_pinned)
1763 pred = get_irn_n(skipped, i);
1764 p = get_irn_node(pred);
1768 if (is_op_commutative(get_irn_op(irn))) {
1769 /* normalize partition order by returning the "smaller" on input 0,
1770 the "bigger" on input 1. */
1771 left = get_binop_left(irn);
1772 pl = get_irn_node(left)->part;
1773 right = get_binop_right(irn);
1774 pr = get_irn_node(right)->part;
1777 return pl < pr ? pl : pr;
1779 return pl > pr ? pl : pr;
1781 /* a not split out Follower */
1782 pred = get_irn_n(irn, i);
1783 p = get_irn_node(pred);
1787 } /* lambda_commutative_partition */
1790 * Returns true if a type is a constant (and NOT Top
1793 static int is_con(const lattice_elem_t type) {
1794 /* be conservative */
1795 if (is_tarval(type.tv))
1796 return tarval_is_constant(type.tv);
1797 return is_entity(type.sym.entity_p);
1801 * Implements split_by().
1803 * @param X the partition to split
1804 * @param env the environment
1806 static void split_by(partition_t *X, environment_t *env) {
1807 partition_t *I, *P = NULL;
1810 dump_partition("split_by", X);
1812 if (X->n_leader == 1) {
1813 /* we have only one leader, no need to split, just check it's type */
1814 node_t *x = get_first_node(X);
1815 X->type_is_T_or_C = x->type.tv == tarval_top || is_con(x->type);
1819 DEBUG_ONLY(what_reason = "lambda n.(n.type)";)
1820 P = split_by_what(X, lambda_type, &P, env);
1823 /* adjust the type tags, we have split partitions by type */
1824 for (I = P; I != NULL; I = I->split_next) {
1825 node_t *x = get_first_node(I);
1826 I->type_is_T_or_C = x->type.tv == tarval_top || is_con(x->type);
1833 if (Y->n_leader > 1) {
1834 /* we do not want split the TOP or constant partitions */
1835 if (! Y->type_is_T_or_C) {
1836 partition_t *Q = NULL;
1838 DEBUG_ONLY(what_reason = "lambda n.(n.opcode)";)
1839 Q = split_by_what(Y, lambda_opcode, &Q, env);
1846 if (Z->n_leader > 1) {
1847 const node_t *first = get_first_node(Z);
1848 int arity = get_irn_arity(first->node);
1850 what_func what = lambda_partition;
1851 DEBUG_ONLY(char buf[64];)
1853 if (env->commutative && is_op_commutative(get_irn_op(first->node)))
1854 what = lambda_commutative_partition;
1857 * BEWARE: during splitting by input 2 for instance we might
1858 * create new partitions which are different by input 1, so collect
1859 * them and split further.
1861 Z->split_next = NULL;
1864 for (input = arity - 1; input >= -1; --input) {
1866 partition_t *Z_prime = R;
1869 if (Z_prime->n_leader > 1) {
1870 env->lambda_input = input;
1871 DEBUG_ONLY(snprintf(buf, sizeof(buf), "lambda n.(n[%d].partition)", input);)
1872 DEBUG_ONLY(what_reason = buf;)
1873 S = split_by_what(Z_prime, what, &S, env);
1876 Z_prime->split_next = S;
1879 } while (R != NULL);
1884 } while (Q != NULL);
1887 } while (P != NULL);
1891 * (Re-)compute the type for a given node.
1893 * @param node the node
1895 static void default_compute(node_t *node) {
1897 ir_node *irn = node->node;
1899 /* if any of the data inputs have type top, the result is type top */
1900 for (i = get_irn_arity(irn) - 1; i >= 0; --i) {
1901 ir_node *pred = get_irn_n(irn, i);
1902 node_t *p = get_irn_node(pred);
1904 if (p->type.tv == tarval_top) {
1905 node->type.tv = tarval_top;
1910 if (get_irn_mode(node->node) == mode_X)
1911 node->type.tv = tarval_reachable;
1913 node->type.tv = computed_value(irn);
1914 } /* default_compute */
1917 * (Re-)compute the type for a Block node.
1919 * @param node the node
1921 static void compute_Block(node_t *node) {
1923 ir_node *block = node->node;
1925 if (block == get_irg_start_block(current_ir_graph) || has_Block_entity(block)) {
1926 /* start block and labelled blocks are always reachable */
1927 node->type.tv = tarval_reachable;
1931 for (i = get_Block_n_cfgpreds(block) - 1; i >= 0; --i) {
1932 node_t *pred = get_irn_node(get_Block_cfgpred(block, i));
1934 if (pred->type.tv == tarval_reachable) {
1935 /* A block is reachable, if at least of predecessor is reachable. */
1936 node->type.tv = tarval_reachable;
1940 node->type.tv = tarval_top;
1941 } /* compute_Block */
1944 * (Re-)compute the type for a Bad node.
1946 * @param node the node
1948 static void compute_Bad(node_t *node) {
1949 /* Bad nodes ALWAYS compute Top */
1950 node->type.tv = tarval_top;
1954 * (Re-)compute the type for an Unknown node.
1956 * @param node the node
1958 static void compute_Unknown(node_t *node) {
1959 /* While Unknown nodes should compute Top this is dangerous:
1960 * a Top input to a Cond would lead to BOTH control flows unreachable.
1961 * While this is correct in the given semantics, it would destroy the Firm
1964 * It would be safe to compute Top IF it can be assured, that only Cmp
1965 * nodes are inputs to Conds. We check that first.
1966 * This is the way Frontends typically build Firm, but some optimizations
1967 * (jump threading for instance) might replace them by Phib's...
1969 node->type.tv = tarval_UNKNOWN;
1970 } /* compute_Unknown */
1973 * (Re-)compute the type for a Jmp node.
1975 * @param node the node
1977 static void compute_Jmp(node_t *node) {
1978 node_t *block = get_irn_node(get_nodes_block(node->node));
1980 node->type = block->type;
1984 * (Re-)compute the type for the Return node.
1986 * @param node the node
1988 static void compute_Return(node_t *node) {
1989 /* The Return node is NOT dead if it is in a reachable block.
1990 * This is already checked in compute(). so we can return
1991 * Reachable here. */
1992 node->type.tv = tarval_reachable;
1993 } /* compute_Return */
1996 * (Re-)compute the type for the End node.
1998 * @param node the node
2000 static void compute_End(node_t *node) {
2001 /* the End node is NOT dead of course */
2002 node->type.tv = tarval_reachable;
2006 * (Re-)compute the type for a Call.
2008 * @param node the node
2010 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) {
2024 ir_node *irn = node->node;
2025 node_t *block = get_irn_node(get_nodes_block(irn));
2027 if (block->type.tv == tarval_unreachable) {
2028 node->type.tv = tarval_top;
2031 switch (get_SymConst_kind(irn)) {
2032 case symconst_addr_ent:
2033 /* case symconst_addr_name: cannot handle this yet */
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) {
2048 ir_node *phi = node->node;
2049 lattice_elem_t type;
2051 /* if a Phi is in a unreachable block, its type is TOP */
2052 node_t *block = get_irn_node(get_nodes_block(phi));
2054 if (block->type.tv == tarval_unreachable) {
2055 node->type.tv = tarval_top;
2059 /* Phi implements the Meet operation */
2060 type.tv = tarval_top;
2061 for (i = get_Phi_n_preds(phi) - 1; i >= 0; --i) {
2062 node_t *pred = get_irn_node(get_Phi_pred(phi, i));
2063 node_t *pred_X = get_irn_node(get_Block_cfgpred(block->node, i));
2065 if (pred_X->type.tv == tarval_unreachable || pred->type.tv == tarval_top) {
2066 /* ignore TOP inputs: We must check here for unreachable blocks,
2067 because Firm constants live in the Start Block are NEVER Top.
2068 Else, a Phi (1,2) will produce Bottom, even if the 2 for instance
2069 comes from a unreachable input. */
2072 if (pred->type.tv == tarval_bottom) {
2073 node->type.tv = tarval_bottom;
2075 } else if (type.tv == tarval_top) {
2076 /* first constant found */
2078 } else if (type.tv != pred->type.tv) {
2079 /* different constants or tarval_bottom */
2080 node->type.tv = tarval_bottom;
2083 /* else nothing, constants are the same */
2089 * (Re-)compute the type for an Add. Special case: one nodes is a Zero Const.
2091 * @param node the node
2093 static void compute_Add(node_t *node) {
2094 ir_node *sub = node->node;
2095 node_t *l = get_irn_node(get_Add_left(sub));
2096 node_t *r = get_irn_node(get_Add_right(sub));
2097 lattice_elem_t a = l->type;
2098 lattice_elem_t b = r->type;
2101 if (a.tv == tarval_top || b.tv == tarval_top) {
2102 node->type.tv = tarval_top;
2103 } else if (a.tv == tarval_bottom || b.tv == tarval_bottom) {
2104 node->type.tv = tarval_bottom;
2106 /* x + 0 = 0 + x = x, but beware of floating point +0 + -0, so we
2107 must call tarval_add() first to handle this case! */
2108 if (is_tarval(a.tv)) {
2109 if (is_tarval(b.tv)) {
2110 node->type.tv = tarval_add(a.tv, b.tv);
2113 mode = get_tarval_mode(a.tv);
2114 if (a.tv == get_mode_null(mode)) {
2118 } else if (is_tarval(b.tv)) {
2119 mode = get_tarval_mode(b.tv);
2120 if (b.tv == get_mode_null(mode)) {
2125 node->type.tv = tarval_bottom;
2130 * (Re-)compute the type for a Sub. Special case: both nodes are congruent.
2132 * @param node the node
2134 static void compute_Sub(node_t *node) {
2135 ir_node *sub = node->node;
2136 node_t *l = get_irn_node(get_Sub_left(sub));
2137 node_t *r = get_irn_node(get_Sub_right(sub));
2138 lattice_elem_t a = l->type;
2139 lattice_elem_t b = r->type;
2142 if (a.tv == tarval_top || b.tv == tarval_top) {
2143 node->type.tv = tarval_top;
2144 } else if (is_con(a) && is_con(b)) {
2145 if (is_tarval(a.tv) && is_tarval(b.tv)) {
2146 node->type.tv = tarval_sub(a.tv, b.tv, get_irn_mode(sub));
2147 } else if (is_tarval(a.tv) && tarval_is_null(a.tv)) {
2149 } else if (is_tarval(b.tv) && tarval_is_null(b.tv)) {
2152 node->type.tv = tarval_bottom;
2154 } else if (r->part == l->part &&
2155 (!mode_is_float(get_irn_mode(l->node)))) {
2157 * BEWARE: a - a is NOT always 0 for floating Point values, as
2158 * NaN op NaN = NaN, so we must check this here.
2160 ir_mode *mode = get_irn_mode(sub);
2161 tv = get_mode_null(mode);
2163 /* if the node was ONCE evaluated by all constants, but now
2164 this breaks AND we get from the argument partitions a different
2165 result, switch to bottom.
2166 This happens because initially all nodes are in the same partition ... */
2167 if (node->type.tv != tv)
2171 node->type.tv = tarval_bottom;
2176 * (Re-)compute the type for an Eor. Special case: both nodes are congruent.
2178 * @param node the node
2180 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) {
2222 ir_node *cmp = node->node;
2223 node_t *l = get_irn_node(get_Cmp_left(cmp));
2224 node_t *r = get_irn_node(get_Cmp_right(cmp));
2225 lattice_elem_t a = l->type;
2226 lattice_elem_t b = r->type;
2228 if (a.tv == tarval_top || b.tv == tarval_top) {
2229 node->type.tv = tarval_top;
2230 } else if (r->part == l->part) {
2231 /* both nodes congruent, we can probably do something */
2232 node->type.tv = tarval_b_true;
2233 } else if (is_con(a) && is_con(b)) {
2234 /* both nodes are constants, we can probably do something */
2235 node->type.tv = tarval_b_true;
2237 node->type.tv = tarval_bottom;
2242 * (Re-)compute the type for a Proj(Cmp).
2244 * @param node the node
2245 * @param cond the predecessor Cmp node
2247 static void compute_Proj_Cmp(node_t *node, ir_node *cmp) {
2248 ir_node *proj = node->node;
2249 node_t *l = get_irn_node(get_Cmp_left(cmp));
2250 node_t *r = get_irn_node(get_Cmp_right(cmp));
2251 lattice_elem_t a = l->type;
2252 lattice_elem_t b = r->type;
2253 pn_Cmp pnc = get_Proj_proj(proj);
2256 if (a.tv == tarval_top || b.tv == tarval_top) {
2257 node->type.tv = tarval_undefined;
2258 } else if (is_con(a) && is_con(b)) {
2259 default_compute(node);
2260 } else if (r->part == l->part &&
2261 (!mode_is_float(get_irn_mode(l->node)) || pnc == pn_Cmp_Lt || pnc == pn_Cmp_Gt)) {
2263 * BEWARE: a == a is NOT always True for floating Point values, as
2264 * NaN != NaN is defined, so we must check this here.
2266 tv = pnc & pn_Cmp_Eq ? tarval_b_true: tarval_b_false;
2268 /* if the node was ONCE evaluated by all constants, but now
2269 this breaks AND we get from the argument partitions a different
2270 result, switch to bottom.
2271 This happens because initially all nodes are in the same partition ... */
2272 if (node->type.tv != tv)
2276 node->type.tv = tarval_bottom;
2278 } /* compute_Proj_Cmp */
2281 * (Re-)compute the type for a Proj(Cond).
2283 * @param node the node
2284 * @param cond the predecessor Cond node
2286 static void compute_Proj_Cond(node_t *node, ir_node *cond) {
2287 ir_node *proj = node->node;
2288 long pnc = get_Proj_proj(proj);
2289 ir_node *sel = get_Cond_selector(cond);
2290 node_t *selector = get_irn_node(sel);
2293 * Note: it is crucial for the monotony that the Proj(Cond)
2294 * are evaluates after all predecessors of the Cond selector are
2300 * Due to the fact that 0 is a const, the Cmp gets immediately
2301 * on the cprop list. It will be evaluated before x is evaluated,
2302 * might leaving x as Top. When later x is evaluated, the Cmp
2303 * might change its value.
2304 * BUT if the Cond is evaluated before this happens, Proj(Cond, FALSE)
2305 * gets R, and later changed to F if Cmp is evaluated to True!
2307 * We prevent this by putting Conds in an extra cprop_X queue, which
2308 * gets evaluated after the cprop queue is empty.
2310 * Note that this even happens with Click's original algorithm, if
2311 * Cmp(x, 0) is evaluated to True first and later changed to False
2312 * if x was Top first and later changed to a Const ...
2313 * It is unclear how Click solved that problem ...
2315 * However, in rare cases even this does not help, if a Top reaches
2316 * a compare through a Phi, than Proj(Cond) is evaluated changing
2317 * the type of the Phi to something other.
2318 * So, we take the last resort and bind the type to R once
2321 * (This might be even the way Click works around the whole problem).
2323 * Finally, we may miss some optimization possibilities due to this:
2328 * If Top reaches the if first, than we decide for != here.
2329 * If y later is evaluated to 0, we cannot revert this decision
2330 * and must live with both outputs enabled. If this happens,
2331 * we get an unresolved if (true) in the code ...
2333 * In Click's version where this decision is done at the Cmp,
2334 * the Cmp is NOT optimized away than (if y evaluated to 1
2335 * for instance) and we get a if (1 == 0) here ...
2337 * Both solutions are suboptimal.
2338 * At least, we could easily detect this problem and run
2339 * cf_opt() (or even combo) again :-(
2341 if (node->type.tv == tarval_reachable)
2344 if (get_irn_mode(sel) == mode_b) {
2346 if (pnc == pn_Cond_true) {
2347 if (selector->type.tv == tarval_b_false) {
2348 node->type.tv = tarval_unreachable;
2349 } else if (selector->type.tv == tarval_b_true) {
2350 node->type.tv = tarval_reachable;
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_unreachable;
2359 node->type.tv = tarval_unreachable;
2363 assert(pnc == pn_Cond_false);
2365 if (selector->type.tv == tarval_b_false) {
2366 node->type.tv = tarval_reachable;
2367 } else if (selector->type.tv == tarval_b_true) {
2368 node->type.tv = tarval_unreachable;
2369 } else if (selector->type.tv == tarval_bottom) {
2370 node->type.tv = tarval_reachable;
2372 assert(selector->type.tv == tarval_top);
2373 if (tarval_UNKNOWN == tarval_top) {
2374 /* any condition based on Top is "!=" */
2375 node->type.tv = tarval_reachable;
2377 node->type.tv = tarval_unreachable;
2383 if (selector->type.tv == tarval_bottom) {
2384 node->type.tv = tarval_reachable;
2385 } else if (selector->type.tv == tarval_top) {
2386 if (tarval_UNKNOWN == tarval_top &&
2387 pnc == get_Cond_default_proj(cond)) {
2388 /* a switch based of Top is always "default" */
2389 node->type.tv = tarval_reachable;
2391 node->type.tv = tarval_unreachable;
2394 long value = get_tarval_long(selector->type.tv);
2395 if (pnc == get_Cond_default_proj(cond)) {
2396 /* default switch, have to check ALL other cases */
2399 for (i = get_irn_n_outs(cond) - 1; i >= 0; --i) {
2400 ir_node *succ = get_irn_out(cond, i);
2404 if (value == get_Proj_proj(succ)) {
2405 /* we found a match, will NOT take the default case */
2406 node->type.tv = tarval_unreachable;
2410 /* all cases checked, no match, will take default case */
2411 node->type.tv = tarval_reachable;
2414 node->type.tv = value == pnc ? tarval_reachable : tarval_unreachable;
2418 } /* compute_Proj_Cond */
2421 * (Re-)compute the type for a Proj-Node.
2423 * @param node the node
2425 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)) {
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;
2447 if (mode != mode_X) {
2449 compute_Proj_Cmp(node, pred);
2451 default_compute(node);
2454 /* handle mode_X nodes */
2456 switch (get_irn_opcode(pred)) {
2458 /* the Proj_X from the Start is always reachable.
2459 However this is already handled at the top. */
2460 node->type.tv = tarval_reachable;
2463 compute_Proj_Cond(node, pred);
2466 default_compute(node);
2468 } /* compute_Proj */
2471 * (Re-)compute the type for a Confirm.
2473 * @param node the node
2475 static void compute_Confirm(node_t *node) {
2476 ir_node *confirm = node->node;
2477 node_t *pred = get_irn_node(get_Confirm_value(confirm));
2479 if (get_Confirm_cmp(confirm) == pn_Cmp_Eq) {
2480 node_t *bound = get_irn_node(get_Confirm_bound(confirm));
2482 if (is_con(bound->type)) {
2483 /* is equal to a constant */
2484 node->type = bound->type;
2488 /* a Confirm is a copy OR a Const */
2489 node->type = pred->type;
2490 } /* compute_Confirm */
2493 * (Re-)compute the type for a given node.
2495 * @param node the node
2497 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_no_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 thing 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) {
2541 ir_node *phi = node->node;
2542 ir_node *block = get_nodes_block(phi);
2543 node_t *n_part = NULL;
2546 for (i = get_Phi_n_preds(phi) - 1; i >= 0; --i) {
2547 node_t *pred_X = get_irn_node(get_Block_cfgpred(block, i));
2549 if (pred_X->type.tv == tarval_reachable) {
2550 node_t *pred = get_irn_node(get_Phi_pred(phi, i));
2554 else if (n_part->part != pred->part) {
2555 /* incongruent inputs, not a follower */
2560 /* if n_part is NULL here, all inputs path are dead, the Phi computes
2561 * tarval_top, is in the TOP partition and should NOT being split! */
2562 assert(n_part != NULL);
2564 } /* identity_Phi */
2567 * Calculates the Identity for commutative 0 neutral nodes.
2569 static node_t *identity_comm_zero_binop(node_t *node) {
2570 ir_node *op = node->node;
2571 node_t *a = get_irn_node(get_binop_left(op));
2572 node_t *b = get_irn_node(get_binop_right(op));
2573 ir_mode *mode = get_irn_mode(op);
2576 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
2577 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
2580 /* node: no input should be tarval_top, else the binop would be also
2581 * Top and not being split. */
2582 zero = get_mode_null(mode);
2583 if (a->type.tv == zero)
2585 if (b->type.tv == zero)
2588 } /* identity_comm_zero_binop */
2591 * Calculates the Identity for Shift nodes.
2593 static node_t *identity_shift(node_t *node) {
2594 ir_node *op = node->node;
2595 node_t *b = get_irn_node(get_binop_right(op));
2596 ir_mode *mode = get_irn_mode(b->node);
2599 /* node: no input should be tarval_top, else the binop would be also
2600 * Top and not being split. */
2601 zero = get_mode_null(mode);
2602 if (b->type.tv == zero)
2603 return get_irn_node(get_binop_left(op));
2605 } /* identity_shift */
2608 * Calculates the Identity for Mul nodes.
2610 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) {
2635 ir_node *sub = node->node;
2636 node_t *b = get_irn_node(get_Sub_right(sub));
2637 ir_mode *mode = get_irn_mode(sub);
2639 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
2640 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
2643 /* node: no input should be tarval_top, else the binop would be also
2644 * Top and not being split. */
2645 if (b->type.tv == get_mode_null(mode))
2646 return get_irn_node(get_Sub_left(sub));
2648 } /* identity_Sub */
2651 * Calculates the Identity for And nodes.
2653 static node_t *identity_And(node_t *node) {
2654 ir_node *and = node->node;
2655 node_t *a = get_irn_node(get_And_left(and));
2656 node_t *b = get_irn_node(get_And_right(and));
2657 tarval *neutral = get_mode_all_one(get_irn_mode(and));
2659 /* node: no input should be tarval_top, else the And would be also
2660 * Top and not being split. */
2661 if (a->type.tv == neutral)
2663 if (b->type.tv == neutral)
2666 } /* identity_And */
2669 * Calculates the Identity for Confirm nodes.
2671 static node_t *identity_Confirm(node_t *node) {
2672 ir_node *confirm = node->node;
2674 /* a Confirm is always a Copy */
2675 return get_irn_node(get_Confirm_value(confirm));
2676 } /* identity_Confirm */
2679 * Calculates the Identity for Mux nodes.
2681 static node_t *identity_Mux(node_t *node) {
2682 ir_node *mux = node->node;
2683 node_t *t = get_irn_node(get_Mux_true(mux));
2684 node_t *f = get_irn_node(get_Mux_false(mux));
2687 if (t->part == f->part)
2690 /* for now, the 1-input identity is not supported */
2692 sel = get_irn_node(get_Mux_sel(mux));
2694 /* Mux sel input is mode_b, so it is always a tarval */
2695 if (sel->type.tv == tarval_b_true)
2697 if (sel->type.tv == tarval_b_false)
2701 } /* identity_Mux */
2704 * Calculates the Identity for nodes.
2706 static node_t *identity(node_t *node) {
2707 ir_node *irn = node->node;
2709 switch (get_irn_opcode(irn)) {
2711 return identity_Phi(node);
2713 return identity_Mul(node);
2717 return identity_comm_zero_binop(node);
2722 return identity_shift(node);
2724 return identity_And(node);
2726 return identity_Sub(node);
2728 return identity_Confirm(node);
2730 return identity_Mux(node);
2737 * Node follower is a (new) follower of leader, segregate Leader
2740 static void segregate_def_use_chain_1(const ir_node *follower, node_t *leader) {
2741 ir_node *l = leader->node;
2742 int j, i, n = get_irn_n_outs(l);
2744 DB((dbg, LEVEL_2, "%+F is a follower of %+F\n", follower, leader->node));
2745 /* The leader edges must remain sorted, but follower edges can
2747 for (i = leader->n_followers + 1; i <= n; ++i) {
2748 if (l->out[i].use == follower) {
2749 ir_def_use_edge t = l->out[i];
2751 for (j = i - 1; j >= leader->n_followers + 1; --j)
2752 l->out[j + 1] = l->out[j];
2753 ++leader->n_followers;
2754 l->out[leader->n_followers] = t;
2758 } /* segregate_def_use_chain_1 */
2761 * Node follower is a (new) follower segregate its Leader
2764 * @param follower the follower IR node
2766 static void segregate_def_use_chain(const ir_node *follower) {
2769 for (i = get_irn_arity(follower) - 1; i >= 0; --i) {
2770 node_t *pred = get_irn_node(get_irn_n(follower, i));
2772 segregate_def_use_chain_1(follower, pred);
2774 } /* segregate_def_use_chain */
2777 * Propagate constant evaluation.
2779 * @param env the environment
2781 static void propagate(environment_t *env) {
2784 lattice_elem_t old_type;
2786 unsigned n_fallen, old_type_was_T_or_C;
2789 while (env->cprop != NULL) {
2790 void *oldopcode = NULL;
2792 /* remove the first partition X from cprop */
2795 env->cprop = X->cprop_next;
2797 old_type_was_T_or_C = X->type_is_T_or_C;
2799 DB((dbg, LEVEL_2, "Propagate type on part%d\n", X->nr));
2803 int cprop_empty = list_empty(&X->cprop);
2804 int cprop_X_empty = list_empty(&X->cprop_X);
2806 if (cprop_empty && cprop_X_empty) {
2807 /* both cprop lists are empty */
2811 /* remove the first Node x from X.cprop */
2813 /* Get a node from the cprop_X list only if
2814 * all data nodes are processed.
2815 * This ensures, that all inputs of the Cond
2816 * predecessor are processed if its type is still Top.
2818 x = list_entry(X->cprop_X.next, node_t, cprop_list);
2820 x = list_entry(X->cprop.next, node_t, cprop_list);
2823 //assert(x->part == X);
2824 list_del(&x->cprop_list);
2827 if (x->is_follower && identity(x) == x) {
2828 /* check the opcode first */
2829 if (oldopcode == NULL) {
2830 oldopcode = lambda_opcode(get_first_node(X), env);
2832 if (oldopcode != lambda_opcode(x, env)) {
2833 if (x->on_fallen == 0) {
2834 /* different opcode -> x falls out of this partition */
2839 DB((dbg, LEVEL_2, "Add node %+F to fallen\n", x->node));
2843 /* x will make the follower -> leader transition */
2844 follower_to_leader(x);
2847 /* compute a new type for x */
2849 DB((dbg, LEVEL_3, "computing type of %+F\n", x->node));
2851 if (x->type.tv != old_type.tv) {
2852 DB((dbg, LEVEL_2, "node %+F has changed type from %+F to %+F\n", x->node, old_type, x->type));
2853 verify_type(old_type, x);
2855 if (x->on_fallen == 0) {
2856 /* Add x to fallen. Nodes might fall from T -> const -> _|_, so check that they are
2857 not already on the list. */
2862 DB((dbg, LEVEL_2, "Add node %+F to fallen\n", x->node));
2864 for (i = get_irn_n_outs(x->node) - 1; i >= 0; --i) {
2865 ir_node *succ = get_irn_out(x->node, i);
2866 node_t *y = get_irn_node(succ);
2868 /* Add y to y.partition.cprop. */
2869 add_to_cprop(y, env);
2874 if (n_fallen > 0 && n_fallen != X->n_leader) {
2875 DB((dbg, LEVEL_2, "Splitting part%d by fallen\n", X->nr));
2876 Y = split(&X, fallen, env);
2878 * We have split out fallen node. The type of the result
2879 * partition is NOT set yet.
2881 Y->type_is_T_or_C = 0;
2885 /* remove the flags from the fallen list */
2886 for (x = fallen; x != NULL; x = x->next)
2889 if (old_type_was_T_or_C) {
2892 /* check if some nodes will make the leader -> follower transition */
2893 list_for_each_entry_safe(node_t, y, tmp, &Y->Leader, node_list) {
2894 if (y->type.tv != tarval_top && ! is_con(y->type)) {
2895 node_t *eq_node = identity(y);
2897 if (eq_node != y && eq_node->part == y->part) {
2898 DB((dbg, LEVEL_2, "Node %+F is a follower of %+F\n", y->node, eq_node->node));
2899 /* move to Follower */
2901 list_del(&y->node_list);
2902 list_add_tail(&y->node_list, &Y->Follower);
2905 segregate_def_use_chain(y->node);
2915 * Get the leader for a given node from its congruence class.
2917 * @param irn the node
2919 static ir_node *get_leader(node_t *node) {
2920 partition_t *part = node->part;
2922 if (part->n_leader > 1 || node->is_follower) {
2923 if (node->is_follower) {
2924 DB((dbg, LEVEL_2, "Replacing follower %+F\n", node->node));
2927 DB((dbg, LEVEL_2, "Found congruence class for %+F\n", node->node));
2929 return get_first_node(part)->node;
2935 * Returns non-zero if a mode_T node has only one reachable output.
2937 static int only_one_reachable_proj(ir_node *n) {
2940 for (i = get_irn_n_outs(n) - 1; i >= 0; --i) {
2941 ir_node *proj = get_irn_out(n, i);
2944 /* skip non-control flow Proj's */
2945 if (get_irn_mode(proj) != mode_X)
2948 node = get_irn_node(proj);
2949 if (node->type.tv == tarval_reachable) {
2955 } /* only_one_reachable_proj */
2958 * Return non-zero if the control flow predecessor node pred
2959 * is the only reachable control flow exit of its block.
2961 * @param pred the control flow exit
2962 * @param block the destination block
2964 static int can_exchange(ir_node *pred, ir_node *block) {
2965 if (is_Start(pred) || has_Block_entity(block))
2967 else if (is_Jmp(pred))
2969 else if (get_irn_mode(pred) == mode_T) {
2970 /* if the predecessor block has more than one
2971 reachable outputs we cannot remove the block */
2972 return only_one_reachable_proj(pred);
2975 } /* can_exchange */
2978 * Block Post-Walker, apply the analysis results on control flow by
2979 * shortening Phi's and Block inputs.
2981 static void apply_cf(ir_node *block, void *ctx) {
2982 environment_t *env = ctx;
2983 node_t *node = get_irn_node(block);
2985 ir_node **ins, **in_X;
2986 ir_node *phi, *next;
2988 n = get_Block_n_cfgpreds(block);
2990 if (node->type.tv == tarval_unreachable) {
2993 for (i = n - 1; i >= 0; --i) {
2994 ir_node *pred = get_Block_cfgpred(block, i);
2996 if (! is_Bad(pred)) {
2997 node_t *pred_bl = get_irn_node(get_nodes_block(skip_Proj(pred)));
2999 if (pred_bl->flagged == 0) {
3000 pred_bl->flagged = 3;
3002 if (pred_bl->type.tv == tarval_reachable) {
3004 * We will remove an edge from block to its pred.
3005 * This might leave the pred block as an endless loop
3007 if (! is_backedge(block, i))
3008 keep_alive(pred_bl->node);
3014 /* the EndBlock is always reachable even if the analysis
3015 finds out the opposite :-) */
3016 if (block != get_irg_end_block(current_ir_graph)) {
3017 /* mark dead blocks */
3018 set_Block_dead(block);
3019 DB((dbg, LEVEL_1, "Removing dead %+F\n", block));
3021 /* the endblock is unreachable */
3022 set_irn_in(block, 0, NULL);
3028 /* only one predecessor combine */
3029 ir_node *pred = skip_Proj(get_Block_cfgpred(block, 0));
3031 if (can_exchange(pred, block)) {
3032 ir_node *new_block = get_nodes_block(pred);
3033 DB((dbg, LEVEL_1, "Fuse %+F with %+F\n", block, new_block));
3034 DBG_OPT_COMBO(block, new_block, FS_OPT_COMBO_CF);
3035 exchange(block, new_block);
3036 node->node = new_block;
3042 NEW_ARR_A(ir_node *, in_X, n);
3044 for (i = 0; i < n; ++i) {
3045 ir_node *pred = get_Block_cfgpred(block, i);
3046 node_t *node = get_irn_node(pred);
3048 if (node->type.tv == tarval_reachable) {
3051 DB((dbg, LEVEL_1, "Removing dead input %d from %+F (%+F)\n", i, block, pred));
3052 if (! is_Bad(pred)) {
3053 node_t *pred_bl = get_irn_node(get_nodes_block(skip_Proj(pred)));
3055 if (pred_bl->flagged == 0) {
3056 pred_bl->flagged = 3;
3058 if (pred_bl->type.tv == tarval_reachable) {
3060 * We will remove an edge from block to its pred.
3061 * This might leave the pred block as an endless loop
3063 if (! is_backedge(block, i))
3064 keep_alive(pred_bl->node);
3074 NEW_ARR_A(ir_node *, ins, n);
3075 for (phi = get_Block_phis(block); phi != NULL; phi = next) {
3076 node_t *node = get_irn_node(phi);
3078 next = get_Phi_next(phi);
3079 if (is_tarval(node->type.tv) && tarval_is_constant(node->type.tv)) {
3080 /* this Phi is replaced by a constant */
3081 tarval *tv = node->type.tv;
3082 ir_node *c = new_Const(tv);
3084 set_irn_node(c, node);
3086 DB((dbg, LEVEL_1, "%+F is replaced by %+F\n", phi, c));
3087 DBG_OPT_COMBO(phi, c, FS_OPT_COMBO_CONST);
3092 for (i = 0; i < n; ++i) {
3093 node_t *pred = get_irn_node(get_Block_cfgpred(block, i));
3095 if (pred->type.tv == tarval_reachable) {
3096 ins[j++] = get_Phi_pred(phi, i);
3100 /* this Phi is replaced by a single predecessor */
3101 ir_node *s = ins[0];
3102 node_t *phi_node = get_irn_node(phi);
3105 DB((dbg, LEVEL_1, "%+F is replaced by %+F because of cf change\n", phi, s));
3106 DBG_OPT_COMBO(phi, s, FS_OPT_COMBO_FOLLOWER);
3111 set_irn_in(phi, j, ins);
3119 /* this Block has only one live predecessor */
3120 ir_node *pred = skip_Proj(in_X[0]);
3122 if (can_exchange(pred, block)) {
3123 ir_node *new_block = get_nodes_block(pred);
3124 DBG_OPT_COMBO(block, new_block, FS_OPT_COMBO_CF);
3125 exchange(block, new_block);
3126 node->node = new_block;
3131 set_irn_in(block, k, in_X);
3136 * Exchange a node by its leader.
3137 * Beware: in rare cases the mode might be wrong here, for instance
3138 * AddP(x, NULL) is a follower of x, but with different mode.
3141 static void exchange_leader(ir_node *irn, ir_node *leader) {
3142 ir_mode *mode = get_irn_mode(irn);
3143 if (mode != get_irn_mode(leader)) {
3144 /* The conv is a no-op, so we are free to place it
3145 * either in the block of the leader OR in irn's block.
3146 * Probably placing it into leaders block might reduce
3147 * the number of Conv due to CSE. */
3148 ir_node *block = get_nodes_block(leader);
3149 dbg_info *dbg = get_irn_dbg_info(irn);
3151 leader = new_rd_Conv(dbg, block, leader, mode);
3153 exchange(irn, leader);
3154 } /* exchange_leader */
3157 * Check, if all users of a mode_M node are dead. Use
3158 * the Def-Use edges for this purpose, as they still
3159 * reflect the situation.
3161 static int all_users_are_dead(const ir_node *irn) {
3162 int i, n = get_irn_n_outs(irn);
3164 for (i = 1; i <= n; ++i) {
3165 const ir_node *succ = irn->out[i].use;
3166 const node_t *block = get_irn_node(get_nodes_block(succ));
3169 if (block->type.tv == tarval_unreachable) {
3170 /* block is unreachable */
3173 node = get_irn_node(succ);
3174 if (node->type.tv != tarval_top) {
3175 /* found a reachable user */
3179 /* all users are unreachable */
3181 } /* all_user_are_dead */
3184 * Walker: Find reachable mode_M nodes that have only
3185 * unreachable users. These nodes must be kept later.
3187 static void find_kept_memory(ir_node *irn, void *ctx) {
3188 environment_t *env = ctx;
3189 node_t *node, *block;
3191 if (get_irn_mode(irn) != mode_M)
3194 block = get_irn_node(get_nodes_block(irn));
3195 if (block->type.tv == tarval_unreachable)
3198 node = get_irn_node(irn);
3199 if (node->type.tv == tarval_top)
3202 /* ok, we found a live memory node. */
3203 if (all_users_are_dead(irn)) {
3204 DB((dbg, LEVEL_1, "%+F must be kept\n", irn));
3205 ARR_APP1(ir_node *, env->kept_memory, irn);
3207 } /* find_kept_memory */
3210 * Post-Walker, apply the analysis results;
3212 static void apply_result(ir_node *irn, void *ctx) {
3213 environment_t *env = ctx;
3214 node_t *node = get_irn_node(irn);
3216 if (is_Block(irn) || is_End(irn) || is_Bad(irn)) {
3217 /* blocks already handled, do not touch the End node */
3219 node_t *block = get_irn_node(get_nodes_block(irn));
3221 if (block->type.tv == tarval_unreachable) {
3222 ir_node *bad = get_irg_bad(current_ir_graph);
3224 /* here, bad might already have a node, but this can be safely ignored
3225 as long as bad has at least ONE valid node */
3226 set_irn_node(bad, node);
3228 DB((dbg, LEVEL_1, "%+F is unreachable\n", irn));
3231 } else if (node->type.tv == tarval_top) {
3232 ir_mode *mode = get_irn_mode(irn);
3234 if (mode == mode_M) {
3235 /* never kill a mode_M node */
3237 ir_node *pred = get_Proj_pred(irn);
3238 node_t *pnode = get_irn_node(pred);
3240 if (pnode->type.tv == tarval_top) {
3241 /* skip the predecessor */
3242 ir_node *mem = get_memop_mem(pred);
3244 DB((dbg, LEVEL_1, "%+F computes Top, replaced by %+F\n", irn, mem));
3249 /* leave other nodes, especially PhiM */
3250 } else if (mode == mode_T) {
3251 /* Do not kill mode_T nodes, kill their Projs */
3252 } else if (! is_Unknown(irn)) {
3253 /* don't kick away Unknown's, they might be still needed */
3254 ir_node *unk = new_r_Unknown(current_ir_graph, mode);
3256 /* control flow should already be handled at apply_cf() */
3257 assert(mode != mode_X);
3259 /* see comment above */
3260 set_irn_node(unk, node);
3262 DB((dbg, LEVEL_1, "%+F computes Top\n", irn));
3267 else if (get_irn_mode(irn) == mode_X) {
3270 ir_node *cond = get_Proj_pred(irn);
3272 if (is_Cond(cond)) {
3273 if (only_one_reachable_proj(cond)) {
3274 ir_node *jmp = new_r_Jmp(block->node);
3275 set_irn_node(jmp, node);
3277 DB((dbg, LEVEL_1, "%+F is replaced by %+F\n", irn, jmp));
3278 DBG_OPT_COMBO(irn, jmp, FS_OPT_COMBO_CF);
3282 node_t *sel = get_irn_node(get_Cond_selector(cond));
3283 tarval *tv = sel->type.tv;
3285 if (is_tarval(tv) && tarval_is_constant(tv)) {
3286 /* The selector is a constant, but more
3287 * than one output is active: An unoptimized
3295 /* normal data node */
3296 if (is_tarval(node->type.tv) && tarval_is_constant(node->type.tv)) {
3297 tarval *tv = node->type.tv;
3300 * Beware: never replace mode_T nodes by constants. Currently we must mark
3301 * mode_T nodes with constants, but do NOT replace them.
3303 if (! is_Const(irn) && get_irn_mode(irn) != mode_T) {
3304 /* can be replaced by a constant */
3305 ir_node *c = new_Const(tv);
3306 set_irn_node(c, node);
3308 DB((dbg, LEVEL_1, "%+F is replaced by %+F\n", irn, c));
3309 DBG_OPT_COMBO(irn, c, FS_OPT_COMBO_CONST);
3310 exchange_leader(irn, c);
3313 } else if (is_entity(node->type.sym.entity_p)) {
3314 if (! is_SymConst(irn)) {
3315 /* can be replaced by a SymConst */
3316 ir_node *symc = new_r_SymConst(current_ir_graph, get_irn_mode(irn), node->type.sym, symconst_addr_ent);
3317 set_irn_node(symc, node);
3320 DB((dbg, LEVEL_1, "%+F is replaced by %+F\n", irn, symc));
3321 DBG_OPT_COMBO(irn, symc, FS_OPT_COMBO_CONST);
3322 exchange_leader(irn, symc);
3325 } else if (is_Confirm(irn)) {
3326 /* Confirms are always follower, but do not kill them here */
3328 ir_node *leader = get_leader(node);
3330 if (leader != irn) {
3331 int non_strict_phi = 0;
3334 * Beware: Do not remove Phi(Unknown, ..., x, ..., Unknown)
3335 * as this might create non-strict programs.
3337 if (node->is_follower && is_Phi(irn) && !is_Unknown(leader)) {
3340 for (i = get_Phi_n_preds(irn) - 1; i >= 0; --i) {
3341 ir_node *pred = get_Phi_pred(irn, i);
3343 if (is_Unknown(pred)) {
3349 if (! non_strict_phi) {
3350 DB((dbg, LEVEL_1, "%+F from part%d is replaced by %+F\n", irn, node->part->nr, leader));
3351 if (node->is_follower)
3352 DBG_OPT_COMBO(irn, leader, FS_OPT_COMBO_FOLLOWER);
3354 DBG_OPT_COMBO(irn, leader, FS_OPT_COMBO_CONGRUENT);
3355 exchange_leader(irn, leader);
3362 } /* apply_result */
3365 * Fix the keep-alives by deleting unreachable ones.
3367 static void apply_end(ir_node *end, environment_t *env) {
3368 int i, j, n = get_End_n_keepalives(end);
3372 NEW_ARR_A(ir_node *, in, n);
3374 /* fix the keep alive */
3375 for (i = j = 0; i < n; i++) {
3376 ir_node *ka = get_End_keepalive(end, i);
3377 node_t *node = get_irn_node(ka);
3380 node = get_irn_node(get_nodes_block(ka));
3382 if (node->type.tv != tarval_unreachable && !is_Bad(ka))
3386 set_End_keepalives(end, j, in);
3391 #define SET(code) op_##code->ops.generic = (op_func)compute_##code
3394 * sets the generic functions to compute.
3396 static void set_compute_functions(void) {
3399 /* set the default compute function */
3400 for (i = get_irp_n_opcodes() - 1; i >= 0; --i) {
3401 ir_op *op = get_irp_opcode(i);
3402 op->ops.generic = (op_func)default_compute;
3405 /* set specific functions */
3421 } /* set_compute_functions */
3426 static void add_memory_keeps(ir_node **kept_memory, int len) {
3427 ir_node *end = get_irg_end(current_ir_graph);
3431 ir_nodeset_init(&set);
3433 /* check, if those nodes are already kept */
3434 for (i = get_End_n_keepalives(end) - 1; i >= 0; --i)
3435 ir_nodeset_insert(&set, get_End_keepalive(end, i));
3437 for (i = len - 1; i >= 0; --i) {
3438 ir_node *ka = kept_memory[i];
3440 if (! ir_nodeset_contains(&set, ka)) {
3441 add_End_keepalive(end, ka);
3444 ir_nodeset_destroy(&set);
3445 } /* add_memory_keeps */
3447 void combo(ir_graph *irg) {
3449 ir_node *initial_bl;
3451 ir_graph *rem = current_ir_graph;
3454 current_ir_graph = irg;
3456 /* register a debug mask */
3457 FIRM_DBG_REGISTER(dbg, "firm.opt.combo");
3459 DB((dbg, LEVEL_1, "Doing COMBO for %+F\n", irg));
3461 obstack_init(&env.obst);
3462 env.worklist = NULL;
3466 #ifdef DEBUG_libfirm
3467 env.dbg_list = NULL;
3469 env.opcode2id_map = new_set(cmp_opcode, iro_Last * 4);
3470 env.type2id_map = pmap_create();
3471 env.kept_memory = NEW_ARR_F(ir_node *, 0);
3472 env.end_idx = get_opt_global_cse() ? 0 : -1;
3473 env.lambda_input = 0;
3476 /* options driving the optimization */
3477 env.commutative = 1;
3478 env.opt_unknown = 1;
3480 assure_irg_outs(irg);
3481 assure_cf_loop(irg);
3483 /* we have our own value_of function */
3484 set_value_of_func(get_node_tarval);
3486 set_compute_functions();
3487 DEBUG_ONLY(part_nr = 0);
3489 ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK | IR_RESOURCE_PHI_LIST);
3491 if (env.opt_unknown)
3492 tarval_UNKNOWN = tarval_top;
3494 tarval_UNKNOWN = tarval_bad;
3496 /* create the initial partition and place it on the work list */
3497 env.initial = new_partition(&env);
3498 add_to_worklist(env.initial, &env);
3499 irg_walk_graph(irg, create_initial_partitions, init_block_phis, &env);
3501 /* set the hook: from now, every node has a partition and a type */
3502 DEBUG_ONLY(set_dump_node_vcgattr_hook(dump_partition_hook));
3504 /* all nodes on the initial partition have type Top */
3505 env.initial->type_is_T_or_C = 1;
3507 /* Place the START Node's partition on cprop.
3508 Place the START Node on its local worklist. */
3509 initial_bl = get_irg_start_block(irg);
3510 start = get_irn_node(initial_bl);
3511 add_to_cprop(start, &env);
3515 if (env.worklist != NULL)
3517 } while (env.cprop != NULL || env.worklist != NULL);
3519 dump_all_partitions(&env);
3520 check_all_partitions(&env);
3523 dump_ir_block_graph(irg, "-partition");
3526 /* apply the result */
3528 /* check, which nodes must be kept */
3529 irg_walk_graph(irg, NULL, find_kept_memory, &env);
3531 /* kill unreachable control flow */
3532 irg_block_walk_graph(irg, NULL, apply_cf, &env);
3533 /* Kill keep-alives of dead blocks: this speeds up apply_result()
3534 * and fixes assertion because dead cf to dead blocks is NOT removed by
3536 apply_end(get_irg_end(irg), &env);
3537 irg_walk_graph(irg, NULL, apply_result, &env);
3539 len = ARR_LEN(env.kept_memory);
3541 add_memory_keeps(env.kept_memory, len);
3544 DB((dbg, LEVEL_1, "Unoptimized Control Flow left"));
3548 /* control flow might changed */
3549 set_irg_outs_inconsistent(irg);
3550 set_irg_extblk_inconsistent(irg);
3551 set_irg_doms_inconsistent(irg);
3552 set_irg_loopinfo_inconsistent(irg);
3553 set_irg_entity_usage_state(irg, ir_entity_usage_not_computed);
3556 ir_free_resources(irg, IR_RESOURCE_IRN_LINK | IR_RESOURCE_PHI_LIST);
3558 /* remove the partition hook */
3559 DEBUG_ONLY(set_dump_node_vcgattr_hook(NULL));
3561 DEL_ARR_F(env.kept_memory);
3562 pmap_destroy(env.type2id_map);
3563 del_set(env.opcode2id_map);
3564 obstack_free(&env.obst, NULL);
3566 /* restore value_of() default behavior */
3567 set_value_of_func(NULL);
3568 current_ir_graph = rem;
3571 /* Creates an ir_graph pass for combo. */
3572 ir_graph_pass_t *combo_pass(const char *name, int verify, int dump)
3574 return def_graph_pass(name ? name : "combo", verify, dump, combo);