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
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"
91 /* define this to check that all type translations are monotone */
92 #define VERIFY_MONOTONE
94 /* define this to check the consistency of partitions */
95 #define CHECK_PARTITIONS
97 typedef struct node_t node_t;
98 typedef struct partition_t partition_t;
99 typedef struct opcode_key_t opcode_key_t;
100 typedef struct listmap_entry_t listmap_entry_t;
102 /** The type of the compute function. */
103 typedef void (*compute_func)(node_t *node);
108 struct opcode_key_t {
109 unsigned code; /**< The Firm opcode. */
110 ir_mode *mode; /**< The mode of all nodes in the partition. */
111 int arity; /**< The arity of this opcode (needed for Phi etc. */
113 long proj; /**< For Proj nodes, its proj number */
114 ir_entity *ent; /**< For Sel Nodes, its entity */
115 int intVal; /**< For Conv/Div Nodes: strict/remainderless */
116 unsigned uintVal;/**< for Builtin: the kind */
117 ir_node *block; /**< for Block: itself */
118 void *ptr; /**< generic pointer for hash/cmp */
123 * An entry in the list_map.
125 struct listmap_entry_t {
126 void *id; /**< The id. */
127 node_t *list; /**< The associated list for this id. */
128 listmap_entry_t *next; /**< Link to the next entry in the map. */
131 /** We must map id's to lists. */
132 typedef struct listmap_t {
133 set *map; /**< Map id's to listmap_entry_t's */
134 listmap_entry_t *values; /**< List of all values in the map. */
138 * A lattice element. Because we handle constants and symbolic constants different, we
139 * have to use this union.
150 ir_node *node; /**< The IR-node itself. */
151 list_head node_list; /**< Double-linked list of leader/follower entries. */
152 list_head cprop_list; /**< Double-linked partition.cprop list. */
153 partition_t *part; /**< points to the partition this node belongs to */
154 node_t *next; /**< Next node on local list (partition.touched, fallen). */
155 node_t *race_next; /**< Next node on race list. */
156 lattice_elem_t type; /**< The associated lattice element "type". */
157 int max_user_input; /**< Maximum input number of Def-Use edges. */
158 int next_edge; /**< Index of the next Def-Use edge to use. */
159 int n_followers; /**< Number of Follower in the outs set. */
160 unsigned on_touched:1; /**< Set, if this node is on the partition.touched set. */
161 unsigned on_cprop:1; /**< Set, if this node is on the partition.cprop list. */
162 unsigned on_fallen:1; /**< Set, if this node is on the fallen list. */
163 unsigned is_follower:1; /**< Set, if this node is a follower. */
164 unsigned flagged:2; /**< 2 Bits, set if this node was visited by race 1 or 2. */
168 * A partition containing congruent nodes.
171 list_head Leader; /**< The head of partition Leader node list. */
172 list_head Follower; /**< The head of partition Follower node list. */
173 list_head cprop; /**< The head of partition.cprop list. */
174 list_head cprop_X; /**< The head of partition.cprop (Cond nodes and its Projs) list. */
175 partition_t *wl_next; /**< Next entry in the work list if any. */
176 partition_t *touched_next; /**< Points to the next partition in the touched set. */
177 partition_t *cprop_next; /**< Points to the next partition in the cprop list. */
178 partition_t *split_next; /**< Points to the next partition in the list that must be split by split_by(). */
179 node_t *touched; /**< The partition.touched set of this partition. */
180 unsigned n_leader; /**< Number of entries in this partition.Leader. */
181 unsigned n_touched; /**< Number of entries in the partition.touched. */
182 int max_user_inputs; /**< Maximum number of user inputs of all entries. */
183 unsigned on_worklist:1; /**< Set, if this partition is in the work list. */
184 unsigned on_touched:1; /**< Set, if this partition is on the touched set. */
185 unsigned on_cprop:1; /**< Set, if this partition is on the cprop list. */
186 unsigned type_is_T_or_C:1;/**< Set, if all nodes in this partition have type Top or Constant. */
188 partition_t *dbg_next; /**< Link all partitions for debugging */
189 unsigned nr; /**< A unique number for (what-)mapping, >0. */
193 typedef struct environment_t {
194 struct obstack obst; /**< obstack to allocate data structures. */
195 partition_t *worklist; /**< The work list. */
196 partition_t *cprop; /**< The constant propagation list. */
197 partition_t *touched; /**< the touched set. */
198 partition_t *initial; /**< The initial partition. */
199 set *opcode2id_map; /**< The opcodeMode->id map. */
200 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 ir_tarval *tarval_UNKNOWN;
237 static node_t *identity(node_t *node);
239 #ifdef CHECK_PARTITIONS
243 static void check_partition(const partition_t *T)
248 list_for_each_entry(node_t, node, &T->Leader, node_list) {
249 assert(node->is_follower == 0);
250 assert(node->flagged == 0);
251 assert(node->part == T);
254 assert(n == T->n_leader);
256 list_for_each_entry(node_t, node, &T->Follower, node_list) {
257 assert(node->is_follower == 1);
258 assert(node->flagged == 0);
259 assert(node->part == T);
261 } /* check_partition */
264 * check that all leader nodes in the partition have the same opcode.
266 static void check_opcode(const partition_t *Z)
272 list_for_each_entry(node_t, node, &Z->Leader, node_list) {
273 ir_node *irn = node->node;
276 key.code = get_irn_opcode(irn);
277 key.mode = get_irn_mode(irn);
278 key.arity = get_irn_arity(irn);
282 switch (get_irn_opcode(irn)) {
284 key.u.proj = get_Proj_proj(irn);
287 key.u.ent = get_Sel_entity(irn);
290 key.u.intVal = get_Conv_strict(irn);
293 key.u.intVal = get_Div_no_remainder(irn);
299 key.mode = get_Load_mode(irn);
302 key.u.intVal = get_Builtin_kind(irn);
309 assert((unsigned)key.code == get_irn_opcode(irn));
310 assert(key.mode == get_irn_mode(irn));
311 assert(key.arity == get_irn_arity(irn));
313 switch (get_irn_opcode(irn)) {
315 assert(key.u.proj == get_Proj_proj(irn));
318 assert(key.u.ent == get_Sel_entity(irn));
321 assert(key.u.intVal == get_Conv_strict(irn));
324 assert(key.u.intVal == get_Div_no_remainder(irn));
327 assert(key.u.block == irn);
330 assert(key.mode == get_Load_mode(irn));
333 assert(key.u.intVal == (int) get_Builtin_kind(irn));
342 static void check_all_partitions(environment_t *env)
348 for (P = env->dbg_list; P != NULL; P = P->dbg_next) {
350 if (! P->type_is_T_or_C)
352 list_for_each_entry(node_t, node, &P->Follower, node_list) {
353 node_t *leader = identity(node);
355 assert(leader != node && leader->part == node->part);
366 static void do_check_list(const node_t *list, int ofs, const partition_t *Z)
371 #define NEXT(e) *((const node_t **)((char *)(e) + (ofs)))
372 for (e = list; e != NULL; e = NEXT(e)) {
373 assert(e->part == Z);
381 } /* ido_check_list */
384 * Check a local list.
386 static void check_list(const node_t *list, const partition_t *Z)
388 do_check_list(list, offsetof(node_t, next), Z);
392 #define check_partition(T)
393 #define check_list(list, Z)
394 #define check_all_partitions(env)
395 #endif /* CHECK_PARTITIONS */
398 static inline lattice_elem_t get_partition_type(const partition_t *X);
401 * Dump partition to output.
403 static void dump_partition(const char *msg, const partition_t *part)
407 lattice_elem_t type = get_partition_type(part);
409 DB((dbg, LEVEL_2, "%s part%u%s (%u, %+F) {\n ",
410 msg, part->nr, part->type_is_T_or_C ? "*" : "",
411 part->n_leader, type));
412 list_for_each_entry(node_t, node, &part->Leader, node_list) {
413 DB((dbg, LEVEL_2, "%s%+F", first ? "" : ", ", node->node));
416 if (! list_empty(&part->Follower)) {
417 DB((dbg, LEVEL_2, "\n---\n "));
419 list_for_each_entry(node_t, node, &part->Follower, node_list) {
420 DB((dbg, LEVEL_2, "%s%+F", first ? "" : ", ", node->node));
424 DB((dbg, LEVEL_2, "\n}\n"));
425 } /* dump_partition */
430 static void do_dump_list(const char *msg, const node_t *node, int ofs)
435 #define GET_LINK(p, ofs) *((const node_t **)((char *)(p) + (ofs)))
437 DB((dbg, LEVEL_3, "%s = {\n ", msg));
438 for (p = node; p != NULL; p = GET_LINK(p, ofs)) {
439 DB((dbg, LEVEL_3, "%s%+F", first ? "" : ", ", p->node));
442 DB((dbg, LEVEL_3, "\n}\n"));
450 static void dump_race_list(const char *msg, const node_t *list)
452 do_dump_list(msg, list, offsetof(node_t, race_next));
453 } /* dump_race_list */
456 * Dumps a local list.
458 static void dump_list(const char *msg, const node_t *list)
460 do_dump_list(msg, list, offsetof(node_t, next));
464 * Dump all partitions.
466 static void dump_all_partitions(const environment_t *env)
468 const partition_t *P;
470 DB((dbg, LEVEL_2, "All partitions\n===============\n"));
471 for (P = env->dbg_list; P != NULL; P = P->dbg_next)
472 dump_partition("", P);
473 } /* dump_all_partitions */
478 static void dump_split_list(const partition_t *list)
480 const partition_t *p;
482 DB((dbg, LEVEL_2, "Split by %s produced = {\n", what_reason));
483 for (p = list; p != NULL; p = p->split_next)
484 DB((dbg, LEVEL_2, "part%u, ", p->nr));
485 DB((dbg, LEVEL_2, "\n}\n"));
486 } /* dump_split_list */
489 * Dump partition and type for a node.
491 static int dump_partition_hook(FILE *F, ir_node *n, ir_node *local)
493 ir_node *irn = local != NULL ? local : n;
494 node_t *node = get_irn_node(irn);
496 ir_fprintf(F, "info2 : \"partition %u type %+F\"\n", node->part->nr, node->type);
498 } /* dump_partition_hook */
501 #define dump_partition(msg, part)
502 #define dump_race_list(msg, list)
503 #define dump_list(msg, list)
504 #define dump_all_partitions(env)
505 #define dump_split_list(list)
508 #if defined(VERIFY_MONOTONE) && defined (DEBUG_libfirm)
510 * Verify that a type transition is monotone
512 static void verify_type(const lattice_elem_t old_type, node_t *node)
514 if (old_type.tv == node->type.tv) {
518 if (old_type.tv == tarval_top) {
519 /* from Top down-to is always allowed */
522 if (node->type.tv == tarval_bottom || node->type.tv == tarval_reachable) {
526 panic("combo: wrong translation from %+F to %+F on node %+F", old_type, node->type, node->node);
530 #define verify_type(old_type, node)
534 * Compare two pointer values of a listmap.
536 static int listmap_cmp_ptr(const void *elt, const void *key, size_t size)
538 const listmap_entry_t *e1 = (listmap_entry_t*)elt;
539 const listmap_entry_t *e2 = (listmap_entry_t*)key;
542 return e1->id != e2->id;
543 } /* listmap_cmp_ptr */
546 * Initializes a listmap.
548 * @param map the listmap
550 static void listmap_init(listmap_t *map)
552 map->map = new_set(listmap_cmp_ptr, 16);
557 * Terminates a listmap.
559 * @param map the listmap
561 static void listmap_term(listmap_t *map)
567 * Return the associated listmap entry for a given id.
569 * @param map the listmap
570 * @param id the id to search for
572 * @return the associated listmap entry for the given id
574 static listmap_entry_t *listmap_find(listmap_t *map, void *id)
576 listmap_entry_t key, *entry;
581 entry = (listmap_entry_t*)set_insert(map->map, &key, sizeof(key), HASH_PTR(id));
583 if (entry->list == NULL) {
584 /* a new entry, put into the list */
585 entry->next = map->values;
592 * Calculate the hash value for an opcode map entry.
594 * @param entry an opcode map entry
596 * @return a hash value for the given opcode map entry
598 static unsigned opcode_hash(const opcode_key_t *entry)
600 return (unsigned)(PTR_TO_INT(entry->mode) * 9 + entry->code + entry->u.proj * 3 + HASH_PTR(entry->u.ptr) + entry->arity);
604 * Compare two entries in the opcode map.
606 static int cmp_opcode(const void *elt, const void *key, size_t size)
608 const opcode_key_t *o1 = (opcode_key_t*)elt;
609 const opcode_key_t *o2 = (opcode_key_t*)key;
612 return o1->code != o2->code || o1->mode != o2->mode ||
613 o1->arity != o2->arity ||
614 o1->u.proj != o2->u.proj ||
615 o1->u.intVal != o2->u.intVal || /* this already checks uIntVal */
616 o1->u.ptr != o2->u.ptr;
620 * Compare two Def-Use edges for input position.
622 static int cmp_def_use_edge(const void *a, const void *b)
624 const ir_def_use_edge *ea = (const ir_def_use_edge*)a;
625 const ir_def_use_edge *eb = (const ir_def_use_edge*)b;
627 /* no overrun, because range is [-1, MAXINT] */
628 return ea->pos - eb->pos;
629 } /* cmp_def_use_edge */
632 * We need the Def-Use edges sorted.
634 static void sort_irn_outs(node_t *node)
636 ir_node *irn = node->node;
637 int n_outs = get_irn_n_outs(irn);
640 qsort(&irn->out[1], n_outs, sizeof(irn->out[0]), cmp_def_use_edge);
642 node->max_user_input = irn->out[n_outs].pos;
643 } /* sort_irn_outs */
646 * Return the type of a node.
648 * @param irn an IR-node
650 * @return the associated type of this node
652 static inline lattice_elem_t get_node_type(const ir_node *irn)
654 return get_irn_node(irn)->type;
655 } /* get_node_type */
658 * Return the tarval of a node.
660 * @param irn an IR-node
662 * @return the associated type of this node
664 static inline ir_tarval *get_node_tarval(const ir_node *irn)
666 lattice_elem_t type = get_node_type(irn);
668 if (is_tarval(type.tv))
670 return tarval_bottom;
671 } /* get_node_type */
674 * Add a partition to the worklist.
676 static inline void add_to_worklist(partition_t *X, environment_t *env)
678 assert(X->on_worklist == 0);
679 DB((dbg, LEVEL_2, "Adding part%d to worklist\n", X->nr));
680 X->wl_next = env->worklist;
683 } /* add_to_worklist */
686 * Create a new empty partition.
688 * @param env the environment
690 * @return a newly allocated partition
692 static inline partition_t *new_partition(environment_t *env)
694 partition_t *part = OALLOC(&env->obst, partition_t);
696 INIT_LIST_HEAD(&part->Leader);
697 INIT_LIST_HEAD(&part->Follower);
698 INIT_LIST_HEAD(&part->cprop);
699 INIT_LIST_HEAD(&part->cprop_X);
700 part->wl_next = NULL;
701 part->touched_next = NULL;
702 part->cprop_next = NULL;
703 part->split_next = NULL;
704 part->touched = NULL;
707 part->max_user_inputs = 0;
708 part->on_worklist = 0;
709 part->on_touched = 0;
711 part->type_is_T_or_C = 0;
713 part->dbg_next = env->dbg_list;
714 env->dbg_list = part;
715 part->nr = part_nr++;
719 } /* new_partition */
722 * Get the first node from a partition.
724 static inline node_t *get_first_node(const partition_t *X)
726 return list_entry(X->Leader.next, node_t, node_list);
727 } /* get_first_node */
730 * Return the type of a partition (assuming partition is non-empty and
731 * all elements have the same type).
733 * @param X a partition
735 * @return the type of the first element of the partition
737 static inline lattice_elem_t get_partition_type(const partition_t *X)
739 const node_t *first = get_first_node(X);
741 } /* get_partition_type */
744 * Creates a partition node for the given IR-node and place it
745 * into the given partition.
747 * @param irn an IR-node
748 * @param part a partition to place the node in
749 * @param env the environment
751 * @return the created node
753 static node_t *create_partition_node(ir_node *irn, partition_t *part, environment_t *env)
755 /* create a partition node and place it in the partition */
756 node_t *node = OALLOC(&env->obst, node_t);
758 INIT_LIST_HEAD(&node->node_list);
759 INIT_LIST_HEAD(&node->cprop_list);
763 node->race_next = NULL;
764 node->type.tv = tarval_top;
765 node->max_user_input = 0;
767 node->n_followers = 0;
768 node->on_touched = 0;
771 node->is_follower = 0;
773 set_irn_node(irn, node);
775 list_add_tail(&node->node_list, &part->Leader);
779 } /* create_partition_node */
782 * Pre-Walker, initialize all Nodes' type to U or top and place
783 * all nodes into the TOP partition.
785 static void create_initial_partitions(ir_node *irn, void *ctx)
787 environment_t *env = (environment_t*)ctx;
788 partition_t *part = env->initial;
791 node = create_partition_node(irn, part, env);
793 if (node->max_user_input > part->max_user_inputs)
794 part->max_user_inputs = node->max_user_input;
797 set_Block_phis(irn, NULL);
799 } /* create_initial_partitions */
802 * Post-Walker, collect all Block-Phi lists, set Cond.
804 static void init_block_phis(ir_node *irn, void *ctx)
809 add_Block_phi(get_nodes_block(irn), irn);
811 } /* init_block_phis */
814 * Add a node to the entry.partition.touched set and
815 * node->partition to the touched set if not already there.
818 * @param env the environment
820 static inline void add_to_touched(node_t *y, environment_t *env)
822 if (y->on_touched == 0) {
823 partition_t *part = y->part;
825 y->next = part->touched;
830 if (part->on_touched == 0) {
831 part->touched_next = env->touched;
833 part->on_touched = 1;
836 check_list(part->touched, part);
838 } /* add_to_touched */
841 * Place a node on the cprop list.
844 * @param env the environment
846 static void add_to_cprop(node_t *y, environment_t *env)
850 /* Add y to y.partition.cprop. */
851 if (y->on_cprop == 0) {
852 partition_t *Y = y->part;
853 ir_node *irn = y->node;
855 /* place Conds and all its Projs on the cprop_X list */
856 if (is_Cond(skip_Proj(irn)))
857 list_add_tail(&y->cprop_list, &Y->cprop_X);
859 list_add_tail(&y->cprop_list, &Y->cprop);
862 DB((dbg, LEVEL_3, "Add %+F to part%u.cprop\n", y->node, Y->nr));
864 /* place its partition on the cprop list */
865 if (Y->on_cprop == 0) {
866 Y->cprop_next = env->cprop;
872 if (get_irn_mode(irn) == mode_T) {
873 /* mode_T nodes always produce tarval_bottom, so we must explicitly
874 add it's Proj's to get constant evaluation to work */
877 for (i = get_irn_n_outs(irn) - 1; i >= 0; --i) {
878 node_t *proj = get_irn_node(get_irn_out(irn, i));
880 add_to_cprop(proj, env);
882 } else if (is_Block(irn)) {
883 /* Due to the way we handle Phi's, we must place all Phis of a block on the list
884 * if someone placed the block. The Block is only placed if the reachability
885 * changes, and this must be re-evaluated in compute_Phi(). */
887 for (phi = get_Block_phis(irn); phi != NULL; phi = get_Phi_next(phi)) {
888 node_t *p = get_irn_node(phi);
889 add_to_cprop(p, env);
895 * Update the worklist: If Z is on worklist then add Z' to worklist.
896 * Else add the smaller of Z and Z' to worklist.
898 * @param Z the Z partition
899 * @param Z_prime the Z' partition, a previous part of Z
900 * @param env the environment
902 static void update_worklist(partition_t *Z, partition_t *Z_prime, environment_t *env)
904 if (Z->on_worklist || Z_prime->n_leader < Z->n_leader) {
905 add_to_worklist(Z_prime, env);
907 add_to_worklist(Z, env);
909 } /* update_worklist */
912 * Make all inputs to x no longer be F.def_use edges.
916 static void move_edges_to_leader(node_t *x)
918 ir_node *irn = x->node;
921 for (i = get_irn_arity(irn) - 1; i >= 0; --i) {
922 node_t *pred = get_irn_node(get_irn_n(irn, i));
927 n = get_irn_n_outs(p);
928 for (j = 1; j <= pred->n_followers; ++j) {
929 if (p->out[j].pos == i && p->out[j].use == irn) {
930 /* found a follower edge to x, move it to the Leader */
931 ir_def_use_edge edge = p->out[j];
933 /* remove this edge from the Follower set */
934 p->out[j] = p->out[pred->n_followers];
937 /* sort it into the leader set */
938 for (k = pred->n_followers + 2; k <= n; ++k) {
939 if (p->out[k].pos >= edge.pos)
941 p->out[k - 1] = p->out[k];
943 /* place the new edge here */
944 p->out[k - 1] = edge;
946 /* edge found and moved */
951 } /* move_edges_to_leader */
954 * Split a partition that has NO followers by a local list.
956 * @param Z partition to split
957 * @param g a (non-empty) node list
958 * @param env the environment
960 * @return a new partition containing the nodes of g
962 static partition_t *split_no_followers(partition_t *Z, node_t *g, environment_t *env)
964 partition_t *Z_prime;
969 dump_partition("Splitting ", Z);
970 dump_list("by list ", g);
974 /* Remove g from Z. */
975 for (node = g; node != NULL; node = node->next) {
976 assert(node->part == Z);
977 list_del(&node->node_list);
980 assert(n < Z->n_leader);
983 /* Move g to a new partition, Z'. */
984 Z_prime = new_partition(env);
986 for (node = g; node != NULL; node = node->next) {
987 list_add_tail(&node->node_list, &Z_prime->Leader);
988 node->part = Z_prime;
989 if (node->max_user_input > max_input)
990 max_input = node->max_user_input;
992 Z_prime->max_user_inputs = max_input;
993 Z_prime->n_leader = n;
996 check_partition(Z_prime);
998 /* for now, copy the type info tag, it will be adjusted in split_by(). */
999 Z_prime->type_is_T_or_C = Z->type_is_T_or_C;
1001 update_worklist(Z, Z_prime, env);
1003 dump_partition("Now ", Z);
1004 dump_partition("Created new ", Z_prime);
1006 } /* split_no_followers */
1009 * Make the Follower -> Leader transition for a node.
1013 static void follower_to_leader(node_t *n)
1015 assert(n->is_follower == 1);
1017 DB((dbg, LEVEL_2, "%+F make the follower -> leader transition\n", n->node));
1019 move_edges_to_leader(n);
1020 list_del(&n->node_list);
1021 list_add_tail(&n->node_list, &n->part->Leader);
1022 ++n->part->n_leader;
1023 } /* follower_to_leader */
1026 * The environment for one race step.
1028 typedef struct step_env {
1029 node_t *initial; /**< The initial node list. */
1030 node_t *unwalked; /**< The unwalked node list. */
1031 node_t *walked; /**< The walked node list. */
1032 int index; /**< Next index of Follower use_def edge. */
1033 unsigned side; /**< side number. */
1037 * Return non-zero, if a input is a real follower
1039 * @param irn the node to check
1040 * @param input number of the input
1042 static int is_real_follower(const ir_node *irn, int input)
1046 switch (get_irn_opcode(irn)) {
1049 /* ignore the Confirm bound input */
1055 /* ignore the Mux sel input */
1060 /* dead inputs are not follower edges */
1061 ir_node *block = get_nodes_block(irn);
1062 node_t *pred = get_irn_node(get_Block_cfgpred(block, input));
1064 if (pred->type.tv == tarval_unreachable)
1074 /* only a Sub x,0 / Shift x,0 might be a follower */
1081 pred = get_irn_node(get_irn_n(irn, input));
1082 if (is_tarval(pred->type.tv) && tarval_is_null(pred->type.tv))
1086 pred = get_irn_node(get_irn_n(irn, input));
1087 if (is_tarval(pred->type.tv) && tarval_is_one(pred->type.tv))
1091 pred = get_irn_node(get_irn_n(irn, input));
1092 if (is_tarval(pred->type.tv) && tarval_is_all_one(pred->type.tv))
1096 assert(!"opcode not implemented yet");
1100 } /* is_real_follower */
1103 * Do one step in the race.
1105 static int step(step_env *env)
1109 if (env->initial != NULL) {
1110 /* Move node from initial to unwalked */
1112 env->initial = n->race_next;
1114 n->race_next = env->unwalked;
1120 while (env->unwalked != NULL) {
1121 /* let n be the first node in unwalked */
1123 while (env->index < n->n_followers) {
1124 const ir_def_use_edge *edge = &n->node->out[1 + env->index];
1126 /* let m be n.F.def_use[index] */
1127 node_t *m = get_irn_node(edge->use);
1129 assert(m->is_follower);
1131 * Some inputs, like the get_Confirm_bound are NOT
1132 * real followers, sort them out.
1134 if (! is_real_follower(m->node, edge->pos)) {
1140 /* only followers from our partition */
1141 if (m->part != n->part)
1144 if ((m->flagged & env->side) == 0) {
1145 m->flagged |= env->side;
1147 if (m->flagged != 3) {
1148 /* visited the first time */
1149 /* add m to unwalked not as first node (we might still need to
1150 check for more follower node */
1151 m->race_next = n->race_next;
1155 /* else already visited by the other side and on the other list */
1158 /* move n to walked */
1159 env->unwalked = n->race_next;
1160 n->race_next = env->walked;
1168 * Clear the flags from a list and check for
1169 * nodes that where touched from both sides.
1171 * @param list the list
1173 static int clear_flags(node_t *list)
1178 for (n = list; n != NULL; n = n->race_next) {
1179 if (n->flagged == 3) {
1180 /* we reach a follower from both sides, this will split congruent
1181 * inputs and make it a leader. */
1182 follower_to_leader(n);
1191 * Split a partition by a local list using the race.
1193 * @param pX pointer to the partition to split, might be changed!
1194 * @param gg a (non-empty) node list
1195 * @param env the environment
1197 * @return a new partition containing the nodes of gg
1199 static partition_t *split(partition_t **pX, node_t *gg, environment_t *env)
1201 partition_t *X = *pX;
1202 partition_t *X_prime;
1205 node_t *g, *h, *node, *t;
1206 int max_input, transitions, winner, shf;
1208 DEBUG_ONLY(static int run = 0;)
1210 DB((dbg, LEVEL_2, "Run %d ", run++));
1211 if (list_empty(&X->Follower)) {
1212 /* if the partition has NO follower, we can use the fast
1213 splitting algorithm. */
1214 return split_no_followers(X, gg, env);
1216 /* else do the race */
1218 dump_partition("Splitting ", X);
1219 dump_list("by list ", gg);
1221 INIT_LIST_HEAD(&tmp);
1223 /* Remove gg from X.Leader and put into g */
1225 for (node = gg; node != NULL; node = node->next) {
1226 assert(node->part == X);
1227 assert(node->is_follower == 0);
1229 list_del(&node->node_list);
1230 list_add_tail(&node->node_list, &tmp);
1231 node->race_next = g;
1236 list_for_each_entry(node_t, node, &X->Leader, node_list) {
1237 node->race_next = h;
1240 /* restore X.Leader */
1241 list_splice(&tmp, &X->Leader);
1243 senv[0].initial = g;
1244 senv[0].unwalked = NULL;
1245 senv[0].walked = NULL;
1249 senv[1].initial = h;
1250 senv[1].unwalked = NULL;
1251 senv[1].walked = NULL;
1256 * Some informations on the race that are not stated clearly in Click's
1258 * 1) A follower stays on the side that reach him first.
1259 * 2) If the other side reches a follower, if will be converted to
1260 * a leader. /This must be done after the race is over, else the
1261 * edges we are iterating on are renumbered./
1262 * 3) /New leader might end up on both sides./
1263 * 4) /If one side ends up with new Leaders, we must ensure that
1264 * they can split out by opcode, hence we have to put _every_
1265 * partition with new Leader nodes on the cprop list, as
1266 * opcode splitting is done by split_by() at the end of
1267 * constant propagation./
1270 if (step(&senv[0])) {
1274 if (step(&senv[1])) {
1279 assert(senv[winner].initial == NULL);
1280 assert(senv[winner].unwalked == NULL);
1282 /* clear flags from walked/unwalked */
1284 transitions = clear_flags(senv[0].unwalked) << shf;
1285 transitions |= clear_flags(senv[0].walked) << shf;
1287 transitions |= clear_flags(senv[1].unwalked) << shf;
1288 transitions |= clear_flags(senv[1].walked) << shf;
1290 dump_race_list("winner ", senv[winner].walked);
1292 /* Move walked_{winner} to a new partition, X'. */
1293 X_prime = new_partition(env);
1296 for (node = senv[winner].walked; node != NULL; node = node->race_next) {
1297 list_del(&node->node_list);
1298 node->part = X_prime;
1299 if (node->is_follower) {
1300 list_add_tail(&node->node_list, &X_prime->Follower);
1302 list_add_tail(&node->node_list, &X_prime->Leader);
1305 if (node->max_user_input > max_input)
1306 max_input = node->max_user_input;
1308 X_prime->n_leader = n;
1309 X_prime->max_user_inputs = max_input;
1310 X->n_leader -= X_prime->n_leader;
1312 /* for now, copy the type info tag, it will be adjusted in split_by(). */
1313 X_prime->type_is_T_or_C = X->type_is_T_or_C;
1316 * Even if a follower was not checked by both sides, it might have
1317 * loose its congruence, so we need to check this case for all follower.
1319 list_for_each_entry_safe(node_t, node, t, &X_prime->Follower, node_list) {
1320 if (identity(node) == node) {
1321 follower_to_leader(node);
1327 check_partition(X_prime);
1329 /* X' is the smaller part */
1330 add_to_worklist(X_prime, env);
1333 * If there where follower to leader transitions, ensure that the nodes
1334 * can be split out if necessary.
1336 if (transitions & 1) {
1337 /* place winner partition on the cprop list */
1338 if (X_prime->on_cprop == 0) {
1339 X_prime->cprop_next = env->cprop;
1340 env->cprop = X_prime;
1341 X_prime->on_cprop = 1;
1344 if (transitions & 2) {
1345 /* place other partition on the cprop list */
1346 if (X->on_cprop == 0) {
1347 X->cprop_next = env->cprop;
1353 dump_partition("Now ", X);
1354 dump_partition("Created new ", X_prime);
1356 /* we have to ensure that the partition containing g is returned */
1366 * Returns non-zero if the i'th input of a Phi node is live.
1368 * @param phi a Phi-node
1369 * @param i an input number
1371 * @return non-zero if the i'th input of the given Phi node is live
1373 static int is_live_input(ir_node *phi, int i)
1376 ir_node *block = get_nodes_block(phi);
1377 ir_node *pred = get_Block_cfgpred(block, i);
1378 lattice_elem_t type = get_node_type(pred);
1380 return type.tv != tarval_unreachable;
1382 /* else it's the control input, always live */
1384 } /* is_live_input */
1387 * Return non-zero if a type is a constant.
1389 static int is_constant_type(lattice_elem_t type)
1391 if (type.tv != tarval_bottom && type.tv != tarval_top)
1394 } /* is_constant_type */
1397 * Check whether a type is neither Top or a constant.
1398 * Note: U is handled like Top here, R is a constant.
1400 * @param type the type to check
1402 static int type_is_neither_top_nor_const(const lattice_elem_t type)
1404 if (is_tarval(type.tv)) {
1405 if (type.tv == tarval_top)
1407 if (tarval_is_constant(type.tv))
1414 } /* type_is_neither_top_nor_const */
1417 * Collect nodes to the touched list.
1419 * @param list the list which contains the nodes that must be evaluated
1420 * @param idx the index of the def_use edge to evaluate
1421 * @param env the environment
1423 static void collect_touched(list_head *list, int idx, environment_t *env)
1426 int end_idx = env->end_idx;
1428 list_for_each_entry(node_t, x, list, node_list) {
1432 /* leader edges start AFTER follower edges */
1433 x->next_edge = x->n_followers + 1;
1435 num_edges = get_irn_n_outs(x->node);
1437 /* for all edges in x.L.def_use_{idx} */
1438 while (x->next_edge <= num_edges) {
1439 const ir_def_use_edge *edge = &x->node->out[x->next_edge];
1442 /* check if we have necessary edges */
1443 if (edge->pos > idx)
1450 /* only non-commutative nodes */
1451 if (env->commutative &&
1452 (idx == 0 || idx == 1) && is_op_commutative(get_irn_op(succ)))
1455 /* ignore the "control input" for non-pinned nodes
1456 if we are running in GCSE mode */
1457 if (idx < end_idx && get_irn_pinned(succ) != op_pin_state_pinned)
1460 y = get_irn_node(succ);
1461 assert(get_irn_n(succ, idx) == x->node);
1463 /* ignore block edges touching followers */
1464 if (idx == -1 && y->is_follower)
1467 if (is_constant_type(y->type)) {
1468 unsigned code = get_irn_opcode(succ);
1469 if (code == iro_Sub || code == iro_Cmp)
1470 add_to_cprop(y, env);
1473 /* Partitions of constants should not be split simply because their Nodes have unequal
1474 functions or incongruent inputs. */
1475 if (type_is_neither_top_nor_const(y->type) &&
1476 (! is_Phi(y->node) || is_live_input(y->node, idx))) {
1477 add_to_touched(y, env);
1481 } /* collect_touched */
1484 * Collect commutative nodes to the touched list.
1486 * @param list the list which contains the nodes that must be evaluated
1487 * @param env the environment
1489 static void collect_commutative_touched(list_head *list, environment_t *env)
1493 list_for_each_entry(node_t, x, list, node_list) {
1496 num_edges = get_irn_n_outs(x->node);
1498 x->next_edge = x->n_followers + 1;
1500 /* for all edges in x.L.def_use_{idx} */
1501 while (x->next_edge <= num_edges) {
1502 const ir_def_use_edge *edge = &x->node->out[x->next_edge];
1505 /* check if we have necessary edges */
1515 /* only commutative nodes */
1516 if (!is_op_commutative(get_irn_op(succ)))
1519 y = get_irn_node(succ);
1520 if (is_constant_type(y->type)) {
1521 unsigned code = get_irn_opcode(succ);
1522 if (code == iro_Eor)
1523 add_to_cprop(y, env);
1526 /* Partitions of constants should not be split simply because their Nodes have unequal
1527 functions or incongruent inputs. */
1528 if (type_is_neither_top_nor_const(y->type)) {
1529 add_to_touched(y, env);
1533 } /* collect_commutative_touched */
1536 * Split the partitions if caused by the first entry on the worklist.
1538 * @param env the environment
1540 static void cause_splits(environment_t *env)
1542 partition_t *X, *Z, *N;
1545 /* remove the first partition from the worklist */
1547 env->worklist = X->wl_next;
1550 dump_partition("Cause_split: ", X);
1552 if (env->commutative) {
1553 /* handle commutative nodes first */
1555 /* empty the touched set: already done, just clear the list */
1556 env->touched = NULL;
1558 collect_commutative_touched(&X->Leader, env);
1559 collect_commutative_touched(&X->Follower, env);
1561 for (Z = env->touched; Z != NULL; Z = N) {
1563 node_t *touched = Z->touched;
1564 node_t *touched_aa = NULL;
1565 node_t *touched_ab = NULL;
1566 unsigned n_touched_aa = 0;
1567 unsigned n_touched_ab = 0;
1569 assert(Z->touched != NULL);
1571 /* beware, split might change Z */
1572 N = Z->touched_next;
1574 /* remove it from the touched set */
1577 /* Empty local Z.touched. */
1578 for (e = touched; e != NULL; e = n) {
1579 node_t *left = get_irn_node(get_irn_n(e->node, 0));
1580 node_t *right = get_irn_node(get_irn_n(e->node, 1));
1582 assert(e->is_follower == 0);
1587 * Note: op(a, a) is NOT congruent to op(a, b).
1588 * So, we must split the touched list.
1590 if (left->part == right->part) {
1591 e->next = touched_aa;
1595 e->next = touched_ab;
1600 assert(n_touched_aa + n_touched_ab == Z->n_touched);
1604 if (0 < n_touched_aa && n_touched_aa < Z->n_leader) {
1605 partition_t *Z_prime = Z;
1606 DB((dbg, LEVEL_2, "Split part%d by touched_aa\n", Z_prime->nr));
1607 split(&Z_prime, touched_aa, env);
1609 assert(n_touched_aa <= Z->n_leader);
1611 if (0 < n_touched_ab && n_touched_ab < Z->n_leader) {
1612 partition_t *Z_prime = Z;
1613 DB((dbg, LEVEL_2, "Split part%d by touched_ab\n", Z_prime->nr));
1614 split(&Z_prime, touched_ab, env);
1616 assert(n_touched_ab <= Z->n_leader);
1620 /* combine temporary leader and follower list */
1621 for (idx = -1; idx <= X->max_user_inputs; ++idx) {
1622 /* empty the touched set: already done, just clear the list */
1623 env->touched = NULL;
1625 collect_touched(&X->Leader, idx, env);
1626 collect_touched(&X->Follower, idx, env);
1628 for (Z = env->touched; Z != NULL; Z = N) {
1630 node_t *touched = Z->touched;
1631 unsigned n_touched = Z->n_touched;
1633 assert(Z->touched != NULL);
1635 /* beware, split might change Z */
1636 N = Z->touched_next;
1638 /* remove it from the touched set */
1641 /* Empty local Z.touched. */
1642 for (e = touched; e != NULL; e = e->next) {
1643 assert(e->is_follower == 0);
1649 if (0 < n_touched && n_touched < Z->n_leader) {
1650 DB((dbg, LEVEL_2, "Split part%d by touched\n", Z->nr));
1651 split(&Z, touched, env);
1653 assert(n_touched <= Z->n_leader);
1656 } /* cause_splits */
1659 * Implements split_by_what(): Split a partition by characteristics given
1660 * by the what function.
1662 * @param X the partition to split
1663 * @param What a function returning an Id for every node of the partition X
1664 * @param P a list to store the result partitions
1665 * @param env the environment
1669 static partition_t *split_by_what(partition_t *X, what_func What,
1670 partition_t **P, environment_t *env)
1674 listmap_entry_t *iter;
1677 /* Let map be an empty mapping from the range of What to (local) list of Nodes. */
1679 list_for_each_entry(node_t, x, &X->Leader, node_list) {
1680 void *id = What(x, env);
1681 listmap_entry_t *entry;
1684 /* input not allowed, ignore */
1687 /* Add x to map[What(x)]. */
1688 entry = listmap_find(&map, id);
1689 x->next = entry->list;
1692 /* Let P be a set of Partitions. */
1694 /* for all sets S except one in the range of map do */
1695 for (iter = map.values; iter != NULL; iter = iter->next) {
1696 if (iter->next == NULL) {
1697 /* this is the last entry, ignore */
1702 /* Add SPLIT( X, S ) to P. */
1703 DB((dbg, LEVEL_2, "Split part%d by WHAT = %s\n", X->nr, what_reason));
1704 R = split(&X, S, env);
1714 } /* split_by_what */
1716 /** lambda n.(n.type) */
1717 static void *lambda_type(const node_t *node, environment_t *env)
1720 return node->type.tv;
1723 /** lambda n.(n.opcode) */
1724 static void *lambda_opcode(const node_t *node, environment_t *env)
1726 opcode_key_t key, *entry;
1727 ir_node *irn = node->node;
1729 key.code = get_irn_opcode(irn);
1730 key.mode = get_irn_mode(irn);
1731 key.arity = get_irn_arity(irn);
1735 switch (get_irn_opcode(irn)) {
1737 key.u.proj = get_Proj_proj(irn);
1740 key.u.ent = get_Sel_entity(irn);
1743 key.u.intVal = get_Conv_strict(irn);
1746 key.u.intVal = get_Div_no_remainder(irn);
1750 * Some ugliness here: Two Blocks having the same
1751 * IJmp predecessor would be congruent, which of course is wrong.
1752 * We fix it by never letting blocks be congruent
1753 * which cannot be detected by combo either.
1758 key.mode = get_Load_mode(irn);
1761 key.u.intVal = get_Builtin_kind(irn);
1767 entry = (opcode_key_t*)set_insert(env->opcode2id_map, &key, sizeof(key), opcode_hash(&key));
1769 } /* lambda_opcode */
1771 /** lambda n.(n[i].partition) */
1772 static void *lambda_partition(const node_t *node, environment_t *env)
1774 ir_node *skipped = skip_Proj(node->node);
1777 int i = env->lambda_input;
1779 if (i >= get_irn_arity(node->node)) {
1781 * We are outside the allowed range: This can happen even
1782 * if we have split by opcode first: doing so might move Followers
1783 * to Leaders and those will have a different opcode!
1784 * Note that in this case the partition is on the cprop list and will be
1790 /* ignore the "control input" for non-pinned nodes
1791 if we are running in GCSE mode */
1792 if (i < env->end_idx && get_irn_pinned(skipped) != op_pin_state_pinned)
1795 pred = i == -1 ? get_irn_n(skipped, i) : get_irn_n(node->node, i);
1796 p = get_irn_node(pred);
1798 } /* lambda_partition */
1800 /** lambda n.(n[i].partition) for commutative nodes */
1801 static void *lambda_commutative_partition(const node_t *node, environment_t *env)
1803 ir_node *irn = node->node;
1804 ir_node *skipped = skip_Proj(irn);
1805 ir_node *pred, *left, *right;
1807 partition_t *pl, *pr;
1808 int i = env->lambda_input;
1810 if (i >= get_irn_arity(node->node)) {
1812 * We are outside the allowed range: This can happen even
1813 * if we have split by opcode first: doing so might move Followers
1814 * to Leaders and those will have a different opcode!
1815 * Note that in this case the partition is on the cprop list and will be
1821 /* ignore the "control input" for non-pinned nodes
1822 if we are running in GCSE mode */
1823 if (i < env->end_idx && get_irn_pinned(skipped) != op_pin_state_pinned)
1827 pred = get_irn_n(skipped, i);
1828 p = get_irn_node(pred);
1832 if (is_op_commutative(get_irn_op(irn))) {
1833 /* normalize partition order by returning the "smaller" on input 0,
1834 the "bigger" on input 1. */
1835 left = get_binop_left(irn);
1836 pl = get_irn_node(left)->part;
1837 right = get_binop_right(irn);
1838 pr = get_irn_node(right)->part;
1841 return pl < pr ? pl : pr;
1843 return pl > pr ? pl : pr;
1845 /* a not split out Follower */
1846 pred = get_irn_n(irn, i);
1847 p = get_irn_node(pred);
1851 } /* lambda_commutative_partition */
1854 * Returns true if a type is a constant (and NOT Top
1857 static int is_con(const lattice_elem_t type)
1859 /* be conservative */
1860 if (is_tarval(type.tv))
1861 return tarval_is_constant(type.tv);
1862 return is_entity(type.sym.entity_p);
1866 * Implements split_by().
1868 * @param X the partition to split
1869 * @param env the environment
1871 static void split_by(partition_t *X, environment_t *env)
1873 partition_t *I, *P = NULL;
1876 dump_partition("split_by", X);
1878 if (X->n_leader == 1) {
1879 /* we have only one leader, no need to split, just check it's type */
1880 node_t *x = get_first_node(X);
1881 X->type_is_T_or_C = x->type.tv == tarval_top || is_con(x->type);
1885 DEBUG_ONLY(what_reason = "lambda n.(n.type)";)
1886 P = split_by_what(X, lambda_type, &P, env);
1889 /* adjust the type tags, we have split partitions by type */
1890 for (I = P; I != NULL; I = I->split_next) {
1891 node_t *x = get_first_node(I);
1892 I->type_is_T_or_C = x->type.tv == tarval_top || is_con(x->type);
1899 if (Y->n_leader > 1) {
1900 /* we do not want split the TOP or constant partitions */
1901 if (! Y->type_is_T_or_C) {
1902 partition_t *Q = NULL;
1904 DEBUG_ONLY(what_reason = "lambda n.(n.opcode)";)
1905 Q = split_by_what(Y, lambda_opcode, &Q, env);
1912 if (Z->n_leader > 1) {
1913 const node_t *first = get_first_node(Z);
1914 int arity = get_irn_arity(first->node);
1916 what_func what = lambda_partition;
1917 DEBUG_ONLY(char buf[64];)
1919 if (env->commutative && is_op_commutative(get_irn_op(first->node)))
1920 what = lambda_commutative_partition;
1923 * BEWARE: during splitting by input 2 for instance we might
1924 * create new partitions which are different by input 1, so collect
1925 * them and split further.
1927 Z->split_next = NULL;
1930 for (input = arity - 1; input >= -1; --input) {
1932 partition_t *Z_prime = R;
1935 if (Z_prime->n_leader > 1) {
1936 env->lambda_input = input;
1937 DEBUG_ONLY(snprintf(buf, sizeof(buf), "lambda n.(n[%d].partition)", input);)
1938 DEBUG_ONLY(what_reason = buf;)
1939 S = split_by_what(Z_prime, what, &S, env);
1942 Z_prime->split_next = S;
1945 } while (R != NULL);
1950 } while (Q != NULL);
1953 } while (P != NULL);
1957 * (Re-)compute the type for a given node.
1959 * @param node the node
1961 static void default_compute(node_t *node)
1964 ir_node *irn = node->node;
1966 /* if any of the data inputs have type top, the result is type top */
1967 for (i = get_irn_arity(irn) - 1; i >= 0; --i) {
1968 ir_node *pred = get_irn_n(irn, i);
1969 node_t *p = get_irn_node(pred);
1971 if (p->type.tv == tarval_top) {
1972 node->type.tv = tarval_top;
1977 if (get_irn_mode(node->node) == mode_X)
1978 node->type.tv = tarval_reachable;
1980 node->type.tv = computed_value(irn);
1981 } /* default_compute */
1984 * (Re-)compute the type for a Block node.
1986 * @param node the node
1988 static void compute_Block(node_t *node)
1991 ir_node *block = node->node;
1993 if (block == get_irg_start_block(current_ir_graph) || has_Block_entity(block)) {
1994 /* start block and labelled blocks are always reachable */
1995 node->type.tv = tarval_reachable;
1999 for (i = get_Block_n_cfgpreds(block) - 1; i >= 0; --i) {
2000 node_t *pred = get_irn_node(get_Block_cfgpred(block, i));
2002 if (pred->type.tv == tarval_reachable) {
2003 /* A block is reachable, if at least of predecessor is reachable. */
2004 node->type.tv = tarval_reachable;
2008 node->type.tv = tarval_top;
2009 } /* compute_Block */
2012 * (Re-)compute the type for a Bad node.
2014 * @param node the node
2016 static void compute_Bad(node_t *node)
2018 /* Bad nodes ALWAYS compute Top */
2019 node->type.tv = tarval_top;
2023 * (Re-)compute the type for an Unknown node.
2025 * @param node the node
2027 static void compute_Unknown(node_t *node)
2029 /* While Unknown nodes should compute Top this is dangerous:
2030 * a Top input to a Cond would lead to BOTH control flows unreachable.
2031 * While this is correct in the given semantics, it would destroy the Firm
2034 * It would be safe to compute Top IF it can be assured, that only Cmp
2035 * nodes are inputs to Conds. We check that first.
2036 * This is the way Frontends typically build Firm, but some optimizations
2037 * (jump threading for instance) might replace them by Phib's...
2039 node->type.tv = tarval_UNKNOWN;
2040 } /* compute_Unknown */
2043 * (Re-)compute the type for a Jmp node.
2045 * @param node the node
2047 static void compute_Jmp(node_t *node)
2049 node_t *block = get_irn_node(get_nodes_block(node->node));
2051 node->type = block->type;
2055 * (Re-)compute the type for the Return node.
2057 * @param node the node
2059 static void compute_Return(node_t *node)
2061 /* The Return node is NOT dead if it is in a reachable block.
2062 * This is already checked in compute(). so we can return
2063 * Reachable here. */
2064 node->type.tv = tarval_reachable;
2065 } /* compute_Return */
2068 * (Re-)compute the type for the End node.
2070 * @param node the node
2072 static void compute_End(node_t *node)
2074 /* the End node is NOT dead of course */
2075 node->type.tv = tarval_reachable;
2079 * (Re-)compute the type for a Call.
2081 * @param node the node
2083 static void compute_Call(node_t *node)
2086 * A Call computes always bottom, even if it has Unknown
2089 node->type.tv = tarval_bottom;
2090 } /* compute_Call */
2093 * (Re-)compute the type for a SymConst node.
2095 * @param node the node
2097 static void compute_SymConst(node_t *node)
2099 ir_node *irn = node->node;
2100 node_t *block = get_irn_node(get_nodes_block(irn));
2102 if (block->type.tv == tarval_unreachable) {
2103 node->type.tv = tarval_top;
2106 switch (get_SymConst_kind(irn)) {
2107 case symconst_addr_ent:
2108 node->type.sym = get_SymConst_symbol(irn);
2111 node->type.tv = computed_value(irn);
2113 } /* compute_SymConst */
2116 * (Re-)compute the type for a Phi node.
2118 * @param node the node
2120 static void compute_Phi(node_t *node)
2123 ir_node *phi = node->node;
2124 lattice_elem_t type;
2126 /* if a Phi is in a unreachable block, its type is TOP */
2127 node_t *block = get_irn_node(get_nodes_block(phi));
2129 if (block->type.tv == tarval_unreachable) {
2130 node->type.tv = tarval_top;
2134 /* Phi implements the Meet operation */
2135 type.tv = tarval_top;
2136 for (i = get_Phi_n_preds(phi) - 1; i >= 0; --i) {
2137 node_t *pred = get_irn_node(get_Phi_pred(phi, i));
2138 node_t *pred_X = get_irn_node(get_Block_cfgpred(block->node, i));
2140 if (pred_X->type.tv == tarval_unreachable || pred->type.tv == tarval_top) {
2141 /* ignore TOP inputs: We must check here for unreachable blocks,
2142 because Firm constants live in the Start Block are NEVER Top.
2143 Else, a Phi (1,2) will produce Bottom, even if the 2 for instance
2144 comes from a unreachable input. */
2147 if (pred->type.tv == tarval_bottom) {
2148 node->type.tv = tarval_bottom;
2150 } else if (type.tv == tarval_top) {
2151 /* first constant found */
2153 } else if (type.tv != pred->type.tv) {
2154 /* different constants or tarval_bottom */
2155 node->type.tv = tarval_bottom;
2158 /* else nothing, constants are the same */
2164 * (Re-)compute the type for an Add. Special case: one nodes is a Zero Const.
2166 * @param node the node
2168 static void compute_Add(node_t *node)
2170 ir_node *sub = node->node;
2171 node_t *l = get_irn_node(get_Add_left(sub));
2172 node_t *r = get_irn_node(get_Add_right(sub));
2173 lattice_elem_t a = l->type;
2174 lattice_elem_t b = r->type;
2177 if (a.tv == tarval_top || b.tv == tarval_top) {
2178 node->type.tv = tarval_top;
2179 } else if (a.tv == tarval_bottom || b.tv == tarval_bottom) {
2180 node->type.tv = tarval_bottom;
2182 /* x + 0 = 0 + x = x, but beware of floating point +0 + -0, so we
2183 must call tarval_add() first to handle this case! */
2184 if (is_tarval(a.tv)) {
2185 if (is_tarval(b.tv)) {
2186 node->type.tv = tarval_add(a.tv, b.tv);
2189 mode = get_tarval_mode(a.tv);
2190 if (a.tv == get_mode_null(mode)) {
2194 } else if (is_tarval(b.tv)) {
2195 mode = get_tarval_mode(b.tv);
2196 if (b.tv == get_mode_null(mode)) {
2201 node->type.tv = tarval_bottom;
2206 * (Re-)compute the type for a Sub. Special case: both nodes are congruent.
2208 * @param node the node
2210 static void compute_Sub(node_t *node)
2212 ir_node *sub = node->node;
2213 node_t *l = get_irn_node(get_Sub_left(sub));
2214 node_t *r = get_irn_node(get_Sub_right(sub));
2215 lattice_elem_t a = l->type;
2216 lattice_elem_t b = r->type;
2219 if (a.tv == tarval_top || b.tv == tarval_top) {
2220 node->type.tv = tarval_top;
2221 } else if (is_con(a) && is_con(b)) {
2222 if (is_tarval(a.tv) && is_tarval(b.tv)) {
2223 node->type.tv = tarval_sub(a.tv, b.tv, get_irn_mode(sub));
2224 } else if (is_tarval(a.tv) && tarval_is_null(a.tv)) {
2226 } else if (is_tarval(b.tv) && tarval_is_null(b.tv)) {
2229 node->type.tv = tarval_bottom;
2231 } else if (r->part == l->part &&
2232 (!mode_is_float(get_irn_mode(l->node)))) {
2234 * BEWARE: a - a is NOT always 0 for floating Point values, as
2235 * NaN op NaN = NaN, so we must check this here.
2237 ir_mode *mode = get_irn_mode(sub);
2238 tv = get_mode_null(mode);
2240 /* if the node was ONCE evaluated by all constants, but now
2241 this breaks AND we get from the argument partitions a different
2242 result, switch to bottom.
2243 This happens because initially all nodes are in the same partition ... */
2244 if (node->type.tv != tv)
2248 node->type.tv = tarval_bottom;
2253 * (Re-)compute the type for an Eor. Special case: both nodes are congruent.
2255 * @param node the node
2257 static void compute_Eor(node_t *node)
2259 ir_node *eor = node->node;
2260 node_t *l = get_irn_node(get_Eor_left(eor));
2261 node_t *r = get_irn_node(get_Eor_right(eor));
2262 lattice_elem_t a = l->type;
2263 lattice_elem_t b = r->type;
2266 if (a.tv == tarval_top || b.tv == tarval_top) {
2267 node->type.tv = tarval_top;
2268 } else if (is_con(a) && is_con(b)) {
2269 if (is_tarval(a.tv) && is_tarval(b.tv)) {
2270 node->type.tv = tarval_eor(a.tv, b.tv);
2271 } else if (is_tarval(a.tv) && tarval_is_null(a.tv)) {
2273 } else if (is_tarval(b.tv) && tarval_is_null(b.tv)) {
2276 node->type.tv = tarval_bottom;
2278 } else if (r->part == l->part) {
2279 ir_mode *mode = get_irn_mode(eor);
2280 tv = get_mode_null(mode);
2282 /* if the node was ONCE evaluated by all constants, but now
2283 this breaks AND we get from the argument partitions a different
2284 result, switch to bottom.
2285 This happens because initially all nodes are in the same partition ... */
2286 if (node->type.tv != tv)
2290 node->type.tv = tarval_bottom;
2295 * (Re-)compute the type for Cmp.
2297 * @param node the node
2299 static void compute_Cmp(node_t *node)
2301 ir_node *cmp = node->node;
2302 node_t *l = get_irn_node(get_Cmp_left(cmp));
2303 node_t *r = get_irn_node(get_Cmp_right(cmp));
2304 lattice_elem_t a = l->type;
2305 lattice_elem_t b = r->type;
2306 ir_mode *mode = get_irn_mode(get_Cmp_left(cmp));
2308 if (a.tv == tarval_top || b.tv == tarval_top) {
2309 node->type.tv = tarval_top;
2310 } else if (r->part == l->part) {
2311 /* both nodes congruent, we can probably do something */
2312 if (mode_is_float(mode)) {
2313 /* beware of NaN's */
2314 node->type.tv = tarval_bottom;
2316 node->type.tv = tarval_b_true;
2318 } else if (is_con(a) && is_con(b)) {
2319 node->type.tv = tarval_b_true;
2321 node->type.tv = tarval_bottom;
2326 * (Re-)compute the type for a Proj(Cmp).
2328 * @param node the node
2329 * @param cond the predecessor Cmp node
2331 static void compute_Proj_Cmp(node_t *node, ir_node *cmp)
2333 ir_node *proj = node->node;
2334 node_t *l = get_irn_node(get_Cmp_left(cmp));
2335 node_t *r = get_irn_node(get_Cmp_right(cmp));
2336 lattice_elem_t a = l->type;
2337 lattice_elem_t b = r->type;
2338 pn_Cmp pnc = get_Proj_pn_cmp(proj);
2341 if (a.tv == tarval_top || b.tv == tarval_top) {
2342 node->type.tv = tarval_undefined;
2343 } else if (is_con(a) && is_con(b)) {
2344 default_compute(node);
2347 * BEWARE: a == a is NOT always True for floating Point values, as
2348 * NaN != NaN is defined, so we must check this here.
2349 * (while for some pnc we could still optimize we have to stay
2350 * consistent with compute_Cmp, so don't do anything for floats)
2352 } else if (r->part == l->part && !mode_is_float(get_irn_mode(l->node))) {
2353 tv = pnc & pn_Cmp_Eq ? tarval_b_true : tarval_b_false;
2355 /* if the node was ONCE evaluated by all constants, but now
2356 this breaks AND we get from the argument partitions a different
2357 result, switch to bottom.
2358 This happens because initially all nodes are in the same partition ... */
2359 if (node->type.tv != tv)
2363 node->type.tv = tarval_bottom;
2365 } /* compute_Proj_Cmp */
2368 * (Re-)compute the type for a Proj(Cond).
2370 * @param node the node
2371 * @param cond the predecessor Cond node
2373 static void compute_Proj_Cond(node_t *node, ir_node *cond)
2375 ir_node *proj = node->node;
2376 long pnc = get_Proj_proj(proj);
2377 ir_node *sel = get_Cond_selector(cond);
2378 node_t *selector = get_irn_node(sel);
2381 * Note: it is crucial for the monotony that the Proj(Cond)
2382 * are evaluates after all predecessors of the Cond selector are
2388 * Due to the fact that 0 is a const, the Cmp gets immediately
2389 * on the cprop list. It will be evaluated before x is evaluated,
2390 * might leaving x as Top. When later x is evaluated, the Cmp
2391 * might change its value.
2392 * BUT if the Cond is evaluated before this happens, Proj(Cond, FALSE)
2393 * gets R, and later changed to F if Cmp is evaluated to True!
2395 * We prevent this by putting Conds in an extra cprop_X queue, which
2396 * gets evaluated after the cprop queue is empty.
2398 * Note that this even happens with Click's original algorithm, if
2399 * Cmp(x, 0) is evaluated to True first and later changed to False
2400 * if x was Top first and later changed to a Const ...
2401 * It is unclear how Click solved that problem ...
2403 * However, in rare cases even this does not help, if a Top reaches
2404 * a compare through a Phi, than Proj(Cond) is evaluated changing
2405 * the type of the Phi to something other.
2406 * So, we take the last resort and bind the type to R once
2409 * (This might be even the way Click works around the whole problem).
2411 * Finally, we may miss some optimization possibilities due to this:
2416 * If Top reaches the if first, than we decide for != here.
2417 * If y later is evaluated to 0, we cannot revert this decision
2418 * and must live with both outputs enabled. If this happens,
2419 * we get an unresolved if (true) in the code ...
2421 * In Click's version where this decision is done at the Cmp,
2422 * the Cmp is NOT optimized away than (if y evaluated to 1
2423 * for instance) and we get a if (1 == 0) here ...
2425 * Both solutions are suboptimal.
2426 * At least, we could easily detect this problem and run
2427 * cf_opt() (or even combo) again :-(
2429 if (node->type.tv == tarval_reachable)
2432 if (get_irn_mode(sel) == mode_b) {
2434 if (pnc == pn_Cond_true) {
2435 if (selector->type.tv == tarval_b_false) {
2436 node->type.tv = tarval_unreachable;
2437 } else if (selector->type.tv == tarval_b_true) {
2438 node->type.tv = tarval_reachable;
2439 } else if (selector->type.tv == tarval_bottom) {
2440 node->type.tv = tarval_reachable;
2442 assert(selector->type.tv == tarval_top);
2443 if (tarval_UNKNOWN == tarval_top) {
2444 /* any condition based on Top is "!=" */
2445 node->type.tv = tarval_unreachable;
2447 node->type.tv = tarval_unreachable;
2451 assert(pnc == pn_Cond_false);
2453 if (selector->type.tv == tarval_b_false) {
2454 node->type.tv = tarval_reachable;
2455 } else if (selector->type.tv == tarval_b_true) {
2456 node->type.tv = tarval_unreachable;
2457 } else if (selector->type.tv == tarval_bottom) {
2458 node->type.tv = tarval_reachable;
2460 assert(selector->type.tv == tarval_top);
2461 if (tarval_UNKNOWN == tarval_top) {
2462 /* any condition based on Top is "!=" */
2463 node->type.tv = tarval_reachable;
2465 node->type.tv = tarval_unreachable;
2471 if (selector->type.tv == tarval_bottom) {
2472 node->type.tv = tarval_reachable;
2473 } else if (selector->type.tv == tarval_top) {
2474 if (tarval_UNKNOWN == tarval_top &&
2475 pnc == get_Cond_default_proj(cond)) {
2476 /* a switch based of Top is always "default" */
2477 node->type.tv = tarval_reachable;
2479 node->type.tv = tarval_unreachable;
2482 long value = get_tarval_long(selector->type.tv);
2483 if (pnc == get_Cond_default_proj(cond)) {
2484 /* default switch, have to check ALL other cases */
2487 for (i = get_irn_n_outs(cond) - 1; i >= 0; --i) {
2488 ir_node *succ = get_irn_out(cond, i);
2492 if (value == get_Proj_proj(succ)) {
2493 /* we found a match, will NOT take the default case */
2494 node->type.tv = tarval_unreachable;
2498 /* all cases checked, no match, will take default case */
2499 node->type.tv = tarval_reachable;
2502 node->type.tv = value == pnc ? tarval_reachable : tarval_unreachable;
2506 } /* compute_Proj_Cond */
2509 * (Re-)compute the type for a Proj-Node.
2511 * @param node the node
2513 static void compute_Proj(node_t *node)
2515 ir_node *proj = node->node;
2516 ir_mode *mode = get_irn_mode(proj);
2517 node_t *block = get_irn_node(get_nodes_block(skip_Proj(proj)));
2518 ir_node *pred = get_Proj_pred(proj);
2520 if (block->type.tv == tarval_unreachable) {
2521 /* a Proj in a unreachable Block stay Top */
2522 node->type.tv = tarval_top;
2525 if (get_irn_node(pred)->type.tv == tarval_top && !is_Cond(pred)) {
2526 /* if the predecessor is Top, its Proj follow */
2527 node->type.tv = tarval_top;
2531 if (mode == mode_M) {
2532 /* mode M is always bottom */
2533 node->type.tv = tarval_bottom;
2536 if (mode != mode_X) {
2538 compute_Proj_Cmp(node, pred);
2540 default_compute(node);
2543 /* handle mode_X nodes */
2545 switch (get_irn_opcode(pred)) {
2547 /* the Proj_X from the Start is always reachable.
2548 However this is already handled at the top. */
2549 node->type.tv = tarval_reachable;
2552 compute_Proj_Cond(node, pred);
2555 default_compute(node);
2557 } /* compute_Proj */
2560 * (Re-)compute the type for a Confirm.
2562 * @param node the node
2564 static void compute_Confirm(node_t *node)
2566 ir_node *confirm = node->node;
2567 node_t *pred = get_irn_node(get_Confirm_value(confirm));
2569 if (get_Confirm_cmp(confirm) == pn_Cmp_Eq) {
2570 node_t *bound = get_irn_node(get_Confirm_bound(confirm));
2572 if (is_con(bound->type)) {
2573 /* is equal to a constant */
2574 node->type = bound->type;
2578 /* a Confirm is a copy OR a Const */
2579 node->type = pred->type;
2580 } /* compute_Confirm */
2583 * (Re-)compute the type for a given node.
2585 * @param node the node
2587 static void compute(node_t *node)
2589 ir_node *irn = node->node;
2592 #ifndef VERIFY_MONOTONE
2594 * Once a node reaches bottom, the type cannot fall further
2595 * in the lattice and we can stop computation.
2596 * Do not take this exit if the monotony verifier is
2597 * enabled to catch errors.
2599 if (node->type.tv == tarval_bottom)
2603 if (!is_Block(irn)) {
2604 /* for pinned nodes, check its control input */
2605 if (get_irn_pinned(skip_Proj(irn)) == op_pin_state_pinned) {
2606 node_t *block = get_irn_node(get_nodes_block(irn));
2608 if (block->type.tv == tarval_unreachable) {
2609 node->type.tv = tarval_top;
2615 func = (compute_func)node->node->op->ops.generic;
2621 * Identity functions: Note that one might think that identity() is just a
2622 * synonym for equivalent_node(). While this is true, we cannot use it for the algorithm
2623 * here, because it expects that the identity node is one of the inputs, which is NOT
2624 * always true for equivalent_node() which can handle (and does sometimes) DAGs.
2625 * So, we have our own implementation, which copies some parts of equivalent_node()
2629 * Calculates the Identity for Phi nodes
2631 static node_t *identity_Phi(node_t *node)
2633 ir_node *phi = node->node;
2634 ir_node *block = get_nodes_block(phi);
2635 node_t *n_part = NULL;
2638 for (i = get_Phi_n_preds(phi) - 1; i >= 0; --i) {
2639 node_t *pred_X = get_irn_node(get_Block_cfgpred(block, i));
2641 if (pred_X->type.tv == tarval_reachable) {
2642 node_t *pred = get_irn_node(get_Phi_pred(phi, i));
2646 else if (n_part->part != pred->part) {
2647 /* incongruent inputs, not a follower */
2652 /* if n_part is NULL here, all inputs path are dead, the Phi computes
2653 * tarval_top, is in the TOP partition and should NOT being split! */
2654 assert(n_part != NULL);
2656 } /* identity_Phi */
2659 * Calculates the Identity for commutative 0 neutral nodes.
2661 static node_t *identity_comm_zero_binop(node_t *node)
2663 ir_node *op = node->node;
2664 node_t *a = get_irn_node(get_binop_left(op));
2665 node_t *b = get_irn_node(get_binop_right(op));
2666 ir_mode *mode = get_irn_mode(op);
2669 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
2670 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
2673 /* node: no input should be tarval_top, else the binop would be also
2674 * Top and not being split. */
2675 zero = get_mode_null(mode);
2676 if (a->type.tv == zero)
2678 if (b->type.tv == zero)
2681 } /* identity_comm_zero_binop */
2684 * Calculates the Identity for Shift nodes.
2686 static node_t *identity_shift(node_t *node)
2688 ir_node *op = node->node;
2689 node_t *b = get_irn_node(get_binop_right(op));
2690 ir_mode *mode = get_irn_mode(b->node);
2693 /* node: no input should be tarval_top, else the binop would be also
2694 * Top and not being split. */
2695 zero = get_mode_null(mode);
2696 if (b->type.tv == zero)
2697 return get_irn_node(get_binop_left(op));
2699 } /* identity_shift */
2702 * Calculates the Identity for Mul nodes.
2704 static node_t *identity_Mul(node_t *node)
2706 ir_node *op = node->node;
2707 node_t *a = get_irn_node(get_Mul_left(op));
2708 node_t *b = get_irn_node(get_Mul_right(op));
2709 ir_mode *mode = get_irn_mode(op);
2712 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
2713 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
2716 /* node: no input should be tarval_top, else the binop would be also
2717 * Top and not being split. */
2718 one = get_mode_one(mode);
2719 if (a->type.tv == one)
2721 if (b->type.tv == one)
2724 } /* identity_Mul */
2727 * Calculates the Identity for Sub nodes.
2729 static node_t *identity_Sub(node_t *node)
2731 ir_node *sub = node->node;
2732 node_t *b = get_irn_node(get_Sub_right(sub));
2733 ir_mode *mode = get_irn_mode(sub);
2735 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
2736 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
2739 /* node: no input should be tarval_top, else the binop would be also
2740 * Top and not being split. */
2741 if (b->type.tv == get_mode_null(mode))
2742 return get_irn_node(get_Sub_left(sub));
2744 } /* identity_Sub */
2747 * Calculates the Identity for And nodes.
2749 static node_t *identity_And(node_t *node)
2751 ir_node *andnode = node->node;
2752 node_t *a = get_irn_node(get_And_left(andnode));
2753 node_t *b = get_irn_node(get_And_right(andnode));
2754 ir_tarval *neutral = get_mode_all_one(get_irn_mode(andnode));
2756 /* node: no input should be tarval_top, else the And would be also
2757 * Top and not being split. */
2758 if (a->type.tv == neutral)
2760 if (b->type.tv == neutral)
2763 } /* identity_And */
2766 * Calculates the Identity for Confirm nodes.
2768 static node_t *identity_Confirm(node_t *node)
2770 ir_node *confirm = node->node;
2772 /* a Confirm is always a Copy */
2773 return get_irn_node(get_Confirm_value(confirm));
2774 } /* identity_Confirm */
2777 * Calculates the Identity for Mux nodes.
2779 static node_t *identity_Mux(node_t *node)
2781 ir_node *mux = node->node;
2782 node_t *t = get_irn_node(get_Mux_true(mux));
2783 node_t *f = get_irn_node(get_Mux_false(mux));
2786 if (t->part == f->part)
2789 /* for now, the 1-input identity is not supported */
2791 sel = get_irn_node(get_Mux_sel(mux));
2793 /* Mux sel input is mode_b, so it is always a tarval */
2794 if (sel->type.tv == tarval_b_true)
2796 if (sel->type.tv == tarval_b_false)
2800 } /* identity_Mux */
2803 * Calculates the Identity for nodes.
2805 static node_t *identity(node_t *node)
2807 ir_node *irn = node->node;
2809 switch (get_irn_opcode(irn)) {
2811 return identity_Phi(node);
2813 return identity_Mul(node);
2817 return identity_comm_zero_binop(node);
2822 return identity_shift(node);
2824 return identity_And(node);
2826 return identity_Sub(node);
2828 return identity_Confirm(node);
2830 return identity_Mux(node);
2837 * Node follower is a (new) follower of leader, segregate Leader
2840 static void segregate_def_use_chain_1(const ir_node *follower, node_t *leader)
2842 ir_node *l = leader->node;
2843 int j, i, n = get_irn_n_outs(l);
2845 DB((dbg, LEVEL_2, "%+F is a follower of %+F\n", follower, leader->node));
2846 /* The leader edges must remain sorted, but follower edges can
2848 for (i = leader->n_followers + 1; i <= n; ++i) {
2849 if (l->out[i].use == follower) {
2850 ir_def_use_edge t = l->out[i];
2852 for (j = i - 1; j >= leader->n_followers + 1; --j)
2853 l->out[j + 1] = l->out[j];
2854 ++leader->n_followers;
2855 l->out[leader->n_followers] = t;
2859 } /* segregate_def_use_chain_1 */
2862 * Node follower is a (new) follower segregate its Leader
2865 * @param follower the follower IR node
2867 static void segregate_def_use_chain(const ir_node *follower)
2871 for (i = get_irn_arity(follower) - 1; i >= 0; --i) {
2872 node_t *pred = get_irn_node(get_irn_n(follower, i));
2874 segregate_def_use_chain_1(follower, pred);
2876 } /* segregate_def_use_chain */
2879 * Propagate constant evaluation.
2881 * @param env the environment
2883 static void propagate(environment_t *env)
2887 lattice_elem_t old_type;
2889 unsigned n_fallen, old_type_was_T_or_C;
2892 while (env->cprop != NULL) {
2893 void *oldopcode = NULL;
2895 /* remove the first partition X from cprop */
2898 env->cprop = X->cprop_next;
2900 old_type_was_T_or_C = X->type_is_T_or_C;
2902 DB((dbg, LEVEL_2, "Propagate type on part%d\n", X->nr));
2906 int cprop_empty = list_empty(&X->cprop);
2907 int cprop_X_empty = list_empty(&X->cprop_X);
2909 if (cprop_empty && cprop_X_empty) {
2910 /* both cprop lists are empty */
2914 /* remove the first Node x from X.cprop */
2916 /* Get a node from the cprop_X list only if
2917 * all data nodes are processed.
2918 * This ensures, that all inputs of the Cond
2919 * predecessor are processed if its type is still Top.
2921 x = list_entry(X->cprop_X.next, node_t, cprop_list);
2923 x = list_entry(X->cprop.next, node_t, cprop_list);
2926 //assert(x->part == X);
2927 list_del(&x->cprop_list);
2930 if (x->is_follower && identity(x) == x) {
2931 /* check the opcode first */
2932 if (oldopcode == NULL) {
2933 oldopcode = lambda_opcode(get_first_node(X), env);
2935 if (oldopcode != lambda_opcode(x, env)) {
2936 if (x->on_fallen == 0) {
2937 /* different opcode -> x falls out of this partition */
2942 DB((dbg, LEVEL_2, "Add node %+F to fallen\n", x->node));
2946 /* x will make the follower -> leader transition */
2947 follower_to_leader(x);
2950 /* compute a new type for x */
2952 DB((dbg, LEVEL_3, "computing type of %+F\n", x->node));
2954 if (x->type.tv != old_type.tv) {
2955 DB((dbg, LEVEL_2, "node %+F has changed type from %+F to %+F\n", x->node, old_type, x->type));
2956 verify_type(old_type, x);
2958 if (x->on_fallen == 0) {
2959 /* Add x to fallen. Nodes might fall from T -> const -> _|_, so check that they are
2960 not already on the list. */
2965 DB((dbg, LEVEL_2, "Add node %+F to fallen\n", x->node));
2967 for (i = get_irn_n_outs(x->node) - 1; i >= 0; --i) {
2968 ir_node *succ = get_irn_out(x->node, i);
2969 node_t *y = get_irn_node(succ);
2971 /* Add y to y.partition.cprop. */
2972 add_to_cprop(y, env);
2977 if (n_fallen > 0 && n_fallen != X->n_leader) {
2978 DB((dbg, LEVEL_2, "Splitting part%d by fallen\n", X->nr));
2979 Y = split(&X, fallen, env);
2981 * We have split out fallen node. The type of the result
2982 * partition is NOT set yet.
2984 Y->type_is_T_or_C = 0;
2988 /* remove the flags from the fallen list */
2989 for (x = fallen; x != NULL; x = x->next)
2992 if (old_type_was_T_or_C) {
2995 /* check if some nodes will make the leader -> follower transition */
2996 list_for_each_entry_safe(node_t, y, tmp, &Y->Leader, node_list) {
2997 if (y->type.tv != tarval_top && ! is_con(y->type)) {
2998 node_t *eq_node = identity(y);
3000 if (eq_node != y && eq_node->part == y->part) {
3001 DB((dbg, LEVEL_2, "Node %+F is a follower of %+F\n", y->node, eq_node->node));
3002 /* move to Follower */
3004 list_del(&y->node_list);
3005 list_add_tail(&y->node_list, &Y->Follower);
3008 segregate_def_use_chain(y->node);
3018 * Get the leader for a given node from its congruence class.
3020 * @param irn the node
3022 static ir_node *get_leader(node_t *node)
3024 partition_t *part = node->part;
3026 if (part->n_leader > 1 || node->is_follower) {
3027 if (node->is_follower) {
3028 DB((dbg, LEVEL_2, "Replacing follower %+F\n", node->node));
3031 DB((dbg, LEVEL_2, "Found congruence class for %+F\n", node->node));
3033 return get_first_node(part)->node;
3039 * Returns non-zero if a mode_T node has only one reachable output.
3041 static int only_one_reachable_proj(ir_node *n)
3045 for (i = get_irn_n_outs(n) - 1; i >= 0; --i) {
3046 ir_node *proj = get_irn_out(n, i);
3049 /* skip non-control flow Proj's */
3050 if (get_irn_mode(proj) != mode_X)
3053 node = get_irn_node(proj);
3054 if (node->type.tv == tarval_reachable) {
3060 } /* only_one_reachable_proj */
3063 * Return non-zero if the control flow predecessor node pred
3064 * is the only reachable control flow exit of its block.
3066 * @param pred the control flow exit
3067 * @param block the destination block
3069 static int can_exchange(ir_node *pred, ir_node *block)
3071 if (is_Start(pred) || has_Block_entity(block))
3073 else if (is_Jmp(pred))
3075 else if (get_irn_mode(pred) == mode_T) {
3076 /* if the predecessor block has more than one
3077 reachable outputs we cannot remove the block */
3078 return only_one_reachable_proj(pred);
3081 } /* can_exchange */
3084 * Block Post-Walker, apply the analysis results on control flow by
3085 * shortening Phi's and Block inputs.
3087 static void apply_cf(ir_node *block, void *ctx)
3089 environment_t *env = (environment_t*)ctx;
3090 node_t *node = get_irn_node(block);
3092 ir_node **ins, **in_X;
3093 ir_node *phi, *next;
3095 n = get_Block_n_cfgpreds(block);
3097 if (node->type.tv == tarval_unreachable) {
3100 for (i = n - 1; i >= 0; --i) {
3101 ir_node *pred = get_Block_cfgpred(block, i);
3103 if (! is_Bad(pred)) {
3104 node_t *pred_bl = get_irn_node(get_nodes_block(skip_Proj(pred)));
3106 if (pred_bl->flagged == 0) {
3107 pred_bl->flagged = 3;
3109 if (pred_bl->type.tv == tarval_reachable) {
3111 * We will remove an edge from block to its pred.
3112 * This might leave the pred block as an endless loop
3114 if (! is_backedge(block, i))
3115 keep_alive(pred_bl->node);
3121 /* the EndBlock is always reachable even if the analysis
3122 finds out the opposite :-) */
3123 if (block != get_irg_end_block(current_ir_graph)) {
3124 /* mark dead blocks */
3125 set_Block_dead(block);
3126 DB((dbg, LEVEL_1, "Removing dead %+F\n", block));
3128 /* the endblock is unreachable */
3129 set_irn_in(block, 0, NULL);
3135 /* only one predecessor combine */
3136 ir_node *pred = skip_Proj(get_Block_cfgpred(block, 0));
3138 if (can_exchange(pred, block)) {
3139 ir_node *new_block = get_nodes_block(pred);
3140 DB((dbg, LEVEL_1, "Fuse %+F with %+F\n", block, new_block));
3141 DBG_OPT_COMBO(block, new_block, FS_OPT_COMBO_CF);
3142 exchange(block, new_block);
3143 node->node = new_block;
3149 NEW_ARR_A(ir_node *, in_X, n);
3151 for (i = 0; i < n; ++i) {
3152 ir_node *pred = get_Block_cfgpred(block, i);
3153 node_t *node = get_irn_node(pred);
3155 if (node->type.tv == tarval_reachable) {
3158 DB((dbg, LEVEL_1, "Removing dead input %d from %+F (%+F)\n", i, block, pred));
3159 if (! is_Bad(pred)) {
3160 node_t *pred_bl = get_irn_node(get_nodes_block(skip_Proj(pred)));
3162 if (pred_bl->flagged == 0) {
3163 pred_bl->flagged = 3;
3165 if (pred_bl->type.tv == tarval_reachable) {
3167 * We will remove an edge from block to its pred.
3168 * This might leave the pred block as an endless loop
3170 if (! is_backedge(block, i))
3171 keep_alive(pred_bl->node);
3181 NEW_ARR_A(ir_node *, ins, n);
3182 for (phi = get_Block_phis(block); phi != NULL; phi = next) {
3183 node_t *node = get_irn_node(phi);
3185 next = get_Phi_next(phi);
3186 if (is_tarval(node->type.tv) && tarval_is_constant(node->type.tv)) {
3187 /* this Phi is replaced by a constant */
3188 ir_tarval *tv = node->type.tv;
3189 ir_node *c = new_r_Const(current_ir_graph, tv);
3191 set_irn_node(c, node);
3193 DB((dbg, LEVEL_1, "%+F is replaced by %+F\n", phi, c));
3194 DBG_OPT_COMBO(phi, c, FS_OPT_COMBO_CONST);
3199 for (i = 0; i < n; ++i) {
3200 node_t *pred = get_irn_node(get_Block_cfgpred(block, i));
3202 if (pred->type.tv == tarval_reachable) {
3203 ins[j++] = get_Phi_pred(phi, i);
3207 /* this Phi is replaced by a single predecessor */
3208 ir_node *s = ins[0];
3209 node_t *phi_node = get_irn_node(phi);
3212 DB((dbg, LEVEL_1, "%+F is replaced by %+F because of cf change\n", phi, s));
3213 DBG_OPT_COMBO(phi, s, FS_OPT_COMBO_FOLLOWER);
3218 set_irn_in(phi, j, ins);
3226 /* this Block has only one live predecessor */
3227 ir_node *pred = skip_Proj(in_X[0]);
3229 if (can_exchange(pred, block)) {
3230 ir_node *new_block = get_nodes_block(pred);
3231 DBG_OPT_COMBO(block, new_block, FS_OPT_COMBO_CF);
3232 exchange(block, new_block);
3233 node->node = new_block;
3238 set_irn_in(block, k, in_X);
3243 * Exchange a node by its leader.
3244 * Beware: in rare cases the mode might be wrong here, for instance
3245 * AddP(x, NULL) is a follower of x, but with different mode.
3248 static void exchange_leader(ir_node *irn, ir_node *leader)
3250 ir_mode *mode = get_irn_mode(irn);
3251 if (mode != get_irn_mode(leader)) {
3252 /* The conv is a no-op, so we are free to place it
3253 * either in the block of the leader OR in irn's block.
3254 * Probably placing it into leaders block might reduce
3255 * the number of Conv due to CSE. */
3256 ir_node *block = get_nodes_block(leader);
3257 dbg_info *dbg = get_irn_dbg_info(irn);
3258 ir_node *nlead = new_rd_Conv(dbg, block, leader, mode);
3260 if (nlead != leader) {
3261 /* Note: this newly create irn has no node info because
3262 * it is created after the analysis. However, this node
3263 * replaces the node irn and should not be visited again,
3264 * so set its visited count to the count of irn.
3265 * Otherwise we might visited this node more than once if
3266 * irn had more than one user.
3268 set_irn_node(nlead, NULL);
3269 set_irn_visited(nlead, get_irn_visited(irn));
3273 exchange(irn, leader);
3274 } /* exchange_leader */
3277 * Check, if all users of a mode_M node are dead. Use
3278 * the Def-Use edges for this purpose, as they still
3279 * reflect the situation.
3281 static int all_users_are_dead(const ir_node *irn)
3283 int i, n = get_irn_n_outs(irn);
3285 for (i = 1; i <= n; ++i) {
3286 const ir_node *succ = irn->out[i].use;
3287 const node_t *block = get_irn_node(get_nodes_block(succ));
3290 if (block->type.tv == tarval_unreachable) {
3291 /* block is unreachable */
3294 node = get_irn_node(succ);
3295 if (node->type.tv != tarval_top) {
3296 /* found a reachable user */
3300 /* all users are unreachable */
3302 } /* all_user_are_dead */
3305 * Walker: Find reachable mode_M nodes that have only
3306 * unreachable users. These nodes must be kept later.
3308 static void find_kept_memory(ir_node *irn, void *ctx)
3310 environment_t *env = (environment_t*)ctx;
3311 node_t *node, *block;
3313 if (get_irn_mode(irn) != mode_M)
3316 block = get_irn_node(get_nodes_block(irn));
3317 if (block->type.tv == tarval_unreachable)
3320 node = get_irn_node(irn);
3321 if (node->type.tv == tarval_top)
3324 /* ok, we found a live memory node. */
3325 if (all_users_are_dead(irn)) {
3326 DB((dbg, LEVEL_1, "%+F must be kept\n", irn));
3327 ARR_APP1(ir_node *, env->kept_memory, irn);
3329 } /* find_kept_memory */
3332 * Post-Walker, apply the analysis results;
3334 static void apply_result(ir_node *irn, void *ctx)
3336 environment_t *env = (environment_t*)ctx;
3337 node_t *node = get_irn_node(irn);
3339 if (is_Block(irn) || is_End(irn) || is_Bad(irn)) {
3340 /* blocks already handled, do not touch the End node */
3342 node_t *block = get_irn_node(get_nodes_block(irn));
3344 if (block->type.tv == tarval_unreachable) {
3345 ir_node *bad = get_irg_bad(current_ir_graph);
3347 /* here, bad might already have a node, but this can be safely ignored
3348 as long as bad has at least ONE valid node */
3349 set_irn_node(bad, node);
3351 DB((dbg, LEVEL_1, "%+F is unreachable\n", irn));
3354 } else if (node->type.tv == tarval_top) {
3355 ir_mode *mode = get_irn_mode(irn);
3357 if (mode == mode_M) {
3358 /* never kill a mode_M node */
3360 ir_node *pred = get_Proj_pred(irn);
3361 node_t *pnode = get_irn_node(pred);
3363 if (pnode->type.tv == tarval_top) {
3364 /* skip the predecessor */
3365 ir_node *mem = get_memop_mem(pred);
3367 DB((dbg, LEVEL_1, "%+F computes Top, replaced by %+F\n", irn, mem));
3372 /* leave other nodes, especially PhiM */
3373 } else if (mode == mode_T) {
3374 /* Do not kill mode_T nodes, kill their Projs */
3375 } else if (! is_Unknown(irn)) {
3376 /* don't kick away Unknown's, they might be still needed */
3377 ir_node *unk = new_r_Unknown(current_ir_graph, mode);
3379 /* control flow should already be handled at apply_cf() */
3380 assert(mode != mode_X);
3382 /* see comment above */
3383 set_irn_node(unk, node);
3385 DB((dbg, LEVEL_1, "%+F computes Top\n", irn));
3390 else if (get_irn_mode(irn) == mode_X) {
3393 ir_node *cond = get_Proj_pred(irn);
3395 if (is_Cond(cond)) {
3396 if (only_one_reachable_proj(cond)) {
3397 ir_node *jmp = new_r_Jmp(block->node);
3398 set_irn_node(jmp, node);
3400 DB((dbg, LEVEL_1, "%+F is replaced by %+F\n", irn, jmp));
3401 DBG_OPT_COMBO(irn, jmp, FS_OPT_COMBO_CF);
3405 node_t *sel = get_irn_node(get_Cond_selector(cond));
3406 ir_tarval *tv = sel->type.tv;
3408 if (is_tarval(tv) && tarval_is_constant(tv)) {
3409 /* The selector is a constant, but more
3410 * than one output is active: An unoptimized
3418 /* normal data node */
3419 if (is_tarval(node->type.tv) && tarval_is_constant(node->type.tv)) {
3420 ir_tarval *tv = node->type.tv;
3423 * Beware: never replace mode_T nodes by constants. Currently we must mark
3424 * mode_T nodes with constants, but do NOT replace them.
3426 if (! is_Const(irn) && get_irn_mode(irn) != mode_T) {
3427 /* can be replaced by a constant */
3428 ir_node *c = new_r_Const(current_ir_graph, tv);
3429 set_irn_node(c, node);
3431 DB((dbg, LEVEL_1, "%+F is replaced by %+F\n", irn, c));
3432 DBG_OPT_COMBO(irn, c, FS_OPT_COMBO_CONST);
3433 exchange_leader(irn, c);
3436 } else if (is_entity(node->type.sym.entity_p)) {
3437 if (! is_SymConst(irn)) {
3438 /* can be replaced by a SymConst */
3439 ir_node *symc = new_r_SymConst(current_ir_graph, get_irn_mode(irn), node->type.sym, symconst_addr_ent);
3440 set_irn_node(symc, node);
3443 DB((dbg, LEVEL_1, "%+F is replaced by %+F\n", irn, symc));
3444 DBG_OPT_COMBO(irn, symc, FS_OPT_COMBO_CONST);
3445 exchange_leader(irn, symc);
3448 } else if (is_Confirm(irn)) {
3449 /* Confirms are always follower, but do not kill them here */
3451 ir_node *leader = get_leader(node);
3453 if (leader != irn) {
3454 int non_strict_phi = 0;
3457 * Beware: Do not remove Phi(Unknown, ..., x, ..., Unknown)
3458 * as this might create non-strict programs.
3460 if (node->is_follower && is_Phi(irn) && !is_Unknown(leader)) {
3463 for (i = get_Phi_n_preds(irn) - 1; i >= 0; --i) {
3464 ir_node *pred = get_Phi_pred(irn, i);
3466 if (is_Unknown(pred)) {
3472 if (! non_strict_phi) {
3473 DB((dbg, LEVEL_1, "%+F from part%d is replaced by %+F\n", irn, node->part->nr, leader));
3474 if (node->is_follower)
3475 DBG_OPT_COMBO(irn, leader, FS_OPT_COMBO_FOLLOWER);
3477 DBG_OPT_COMBO(irn, leader, FS_OPT_COMBO_CONGRUENT);
3478 exchange_leader(irn, leader);
3485 } /* apply_result */
3488 * Fix the keep-alives by deleting unreachable ones.
3490 static void apply_end(ir_node *end, environment_t *env)
3492 int i, j, n = get_End_n_keepalives(end);
3493 ir_node **in = NULL;
3496 NEW_ARR_A(ir_node *, in, n);
3498 /* fix the keep alive */
3499 for (i = j = 0; i < n; i++) {
3500 ir_node *ka = get_End_keepalive(end, i);
3501 node_t *node = get_irn_node(ka);
3504 node = get_irn_node(get_nodes_block(ka));
3506 if (node->type.tv != tarval_unreachable && !is_Bad(ka))
3510 set_End_keepalives(end, j, in);
3515 #define SET(code) op_##code->ops.generic = (op_func)compute_##code
3518 * sets the generic functions to compute.
3520 static void set_compute_functions(void)
3524 /* set the default compute function */
3525 for (i = 0, n = get_irp_n_opcodes(); i < n; ++i) {
3526 ir_op *op = get_irp_opcode(i);
3527 op->ops.generic = (op_func)default_compute;
3530 /* set specific functions */
3546 } /* set_compute_functions */
3551 static void add_memory_keeps(ir_node **kept_memory, size_t len)
3553 ir_node *end = get_irg_end(current_ir_graph);
3558 ir_nodeset_init(&set);
3560 /* check, if those nodes are already kept */
3561 for (i = get_End_n_keepalives(end) - 1; i >= 0; --i)
3562 ir_nodeset_insert(&set, get_End_keepalive(end, i));
3564 for (idx = 0; idx < len; ++idx) {
3565 ir_node *ka = kept_memory[idx];
3567 if (! ir_nodeset_contains(&set, ka)) {
3568 add_End_keepalive(end, ka);
3571 ir_nodeset_destroy(&set);
3572 } /* add_memory_keeps */
3574 void combo(ir_graph *irg)
3577 ir_node *initial_bl;
3579 ir_graph *rem = current_ir_graph;
3582 current_ir_graph = irg;
3584 /* register a debug mask */
3585 FIRM_DBG_REGISTER(dbg, "firm.opt.combo");
3587 DB((dbg, LEVEL_1, "Doing COMBO for %+F\n", irg));
3589 obstack_init(&env.obst);
3590 env.worklist = NULL;
3594 #ifdef DEBUG_libfirm
3595 env.dbg_list = NULL;
3597 env.opcode2id_map = new_set(cmp_opcode, iro_Last * 4);
3598 env.kept_memory = NEW_ARR_F(ir_node *, 0);
3599 env.end_idx = get_opt_global_cse() ? 0 : -1;
3600 env.lambda_input = 0;
3603 /* options driving the optimization */
3604 env.commutative = 1;
3605 env.opt_unknown = 1;
3607 assure_irg_outs(irg);
3608 assure_cf_loop(irg);
3610 /* we have our own value_of function */
3611 set_value_of_func(get_node_tarval);
3613 set_compute_functions();
3614 DEBUG_ONLY(part_nr = 0);
3616 ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK | IR_RESOURCE_PHI_LIST);
3618 if (env.opt_unknown)
3619 tarval_UNKNOWN = tarval_top;
3621 tarval_UNKNOWN = tarval_bad;
3623 /* create the initial partition and place it on the work list */
3624 env.initial = new_partition(&env);
3625 add_to_worklist(env.initial, &env);
3626 irg_walk_graph(irg, create_initial_partitions, init_block_phis, &env);
3628 /* set the hook: from now, every node has a partition and a type */
3629 DEBUG_ONLY(set_dump_node_vcgattr_hook(dump_partition_hook));
3631 /* all nodes on the initial partition have type Top */
3632 env.initial->type_is_T_or_C = 1;
3634 /* Place the START Node's partition on cprop.
3635 Place the START Node on its local worklist. */
3636 initial_bl = get_irg_start_block(irg);
3637 start = get_irn_node(initial_bl);
3638 add_to_cprop(start, &env);
3642 if (env.worklist != NULL)
3644 } while (env.cprop != NULL || env.worklist != NULL);
3646 dump_all_partitions(&env);
3647 check_all_partitions(&env);
3650 dump_ir_block_graph(irg, "-partition");
3653 /* apply the result */
3655 /* check, which nodes must be kept */
3656 irg_walk_graph(irg, NULL, find_kept_memory, &env);
3658 /* kill unreachable control flow */
3659 irg_block_walk_graph(irg, NULL, apply_cf, &env);
3660 /* Kill keep-alives of dead blocks: this speeds up apply_result()
3661 * and fixes assertion because dead cf to dead blocks is NOT removed by
3663 apply_end(get_irg_end(irg), &env);
3664 irg_walk_graph(irg, NULL, apply_result, &env);
3666 len = ARR_LEN(env.kept_memory);
3668 add_memory_keeps(env.kept_memory, len);
3671 DB((dbg, LEVEL_1, "Unoptimized Control Flow left"));
3675 /* control flow might changed */
3676 set_irg_outs_inconsistent(irg);
3677 set_irg_extblk_inconsistent(irg);
3678 set_irg_doms_inconsistent(irg);
3679 set_irg_loopinfo_inconsistent(irg);
3680 set_irg_entity_usage_state(irg, ir_entity_usage_not_computed);
3683 ir_free_resources(irg, IR_RESOURCE_IRN_LINK | IR_RESOURCE_PHI_LIST);
3685 /* remove the partition hook */
3686 DEBUG_ONLY(set_dump_node_vcgattr_hook(NULL));
3688 DEL_ARR_F(env.kept_memory);
3689 del_set(env.opcode2id_map);
3690 obstack_free(&env.obst, NULL);
3692 /* restore value_of() default behavior */
3693 set_value_of_func(NULL);
3694 current_ir_graph = rem;
3697 /* Creates an ir_graph pass for combo. */
3698 ir_graph_pass_t *combo_pass(const char *name)
3700 return def_graph_pass(name ? name : "combo", combo);