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 * Note that the current implementation lack the leaders/followers
29 * Note further that we use the terminology from Click's work here, which is different
30 * in some cases from Firm terminology. Especially, Click's type is a
31 * Firm tarval/entity, nevertheless we call it type here for "maximum compatibility".
39 #include "iroptimize.h"
46 #include "irgraph_t.h"
61 /* define this to check that all type translations are monotone */
62 #define VERIFY_MONOTONE
64 typedef struct node_t node_t;
65 typedef struct partition_t partition_t;
66 typedef struct opcode_key_t opcode_key_t;
67 typedef struct listmap_entry_t listmap_entry_t;
69 /** The type of the compute function. */
70 typedef void (*compute_func)(node_t *node);
76 ir_opcode code; /**< The Firm opcode. */
77 ir_mode *mode; /**< The mode of all nodes in the partition. */
79 long proj; /**< For Proj nodes, its proj number */
80 ir_entity *ent; /**< For Sel Nodes, its entity */
85 * An entry in the list_map.
87 struct listmap_entry_t {
88 void *id; /**< The id. */
89 node_t *list; /**< The associated list for this id. */
90 listmap_entry_t *next; /**< Link to the next entry in the map. */
93 /** We must map id's to lists. */
94 typedef struct listmap_t {
95 set *map; /**< Map id's to listmap_entry_t's */
96 listmap_entry_t *values; /**< List of all values in the map. */
100 * A lattice element. Because we handle constants and symbolic constants different, we
101 * have to use this union.
112 ir_node *node; /**< The IR-node itself. */
113 list_head node_list; /**< Double-linked list of entries. */
114 list_head cprop_list; /**< Double-linked partition.cprop list. */
115 partition_t *part; /**< points to the partition this node belongs to */
116 node_t *next; /**< Next node on local list (partition.touched, fallen). */
117 lattice_elem_t type; /**< The associated lattice element "type". */
118 int max_user_input; /**< Maximum input number of Def-Use edges. */
119 int next_edge; /**< Index of the next Def-Use edge to use. */
120 unsigned on_touched:1; /**< Set, if this node is on the partition.touched set. */
121 unsigned on_cprop:1; /**< Set, if this node is on the partition.cprop list. */
122 unsigned on_fallen:1; /**< Set, if this node is on the fallen list. */
126 * A partition containing congruent nodes.
129 list_head entries; /**< The head of partition node list. */
130 list_head cprop; /**< The head of partition.cprop list. */
131 partition_t *wl_next; /**< Next entry in the work list if any. */
132 partition_t *touched_next; /**< Points to the next partition in the touched set. */
133 partition_t *cprop_next; /**< Points to the next partition in the cprop list. */
134 partition_t *split_next; /**< Points to the next partition in the list that must be split by split_by(). */
135 node_t *touched; /**< The partition.touched set of this partition. */
136 unsigned n_nodes; /**< Number of entries in this partition. */
137 unsigned n_touched; /**< Number of entries in the partition.touched. */
138 int max_user_inputs; /**< Maximum number of user inputs of all entries. */
139 unsigned on_worklist:1; /**< Set, if this partition is in the work list. */
140 unsigned on_touched:1; /**< Set, if this partition is on the touched set. */
141 unsigned on_cprop:1; /**< Set, if this partition is on the cprop list. */
143 partition_t *dbg_next; /**< Link all partitions for debugging */
144 unsigned nr; /**< A unique number for (what-)mapping, >0. */
148 typedef struct environment_t {
149 struct obstack obst; /**< obstack to allocate data structures. */
150 partition_t *worklist; /**< The work list. */
151 partition_t *cprop; /**< The constant propagation list. */
152 partition_t *touched; /**< the touched set. */
153 partition_t *initial; /**< The initial partition. */
154 set *opcode2id_map; /**< The opcodeMode->id map. */
155 pmap *type2id_map; /**< The type->id map. */
156 int end_idx; /**< -1 for local and 0 for global congruences. */
157 int lambda_input; /**< Captured argument for lambda_partition(). */
159 partition_t *dbg_list; /**< List of all partitions. */
163 /** Type of the what function. */
164 typedef void *(*what_func)(const node_t *node, environment_t *env);
166 #define get_irn_node(irn) ((node_t *)get_irn_link(irn))
167 #define set_irn_node(irn, node) set_irn_link(irn, node)
169 /* we do NOT use tarval_unreachable here, instead we use Top for this purpose */
170 #undef tarval_unreachable
171 #define tarval_unreachable tarval_top
174 /** The debug module handle. */
175 DEBUG_ONLY(static firm_dbg_module_t *dbg;)
177 /** Next partition number. */
178 DEBUG_ONLY(static unsigned part_nr = 0);
181 static INLINE lattice_elem_t get_partition_type(const partition_t *X);
184 * Dump partition to output.
186 static void dump_partition(const char *msg, const partition_t *part) {
189 lattice_elem_t type = get_partition_type(part);
191 DB((dbg, LEVEL_2, "%s part%u (%u, %+F) {\n ", msg, part->nr, part->n_nodes, type));
192 list_for_each_entry(node_t, node, &part->entries, node_list) {
193 DB((dbg, LEVEL_2, "%s%+F", first ? "" : ", ", node->node));
196 DB((dbg, LEVEL_2, "\n}\n"));
200 * Dump all partitions.
202 static void dump_all_partitions(const environment_t *env) {
203 const partition_t *P;
205 DB((dbg, LEVEL_2, "All partitions\n===============\n"));
206 for (P = env->dbg_list; P != NULL; P = P->dbg_next)
207 dump_partition("", P);
211 #define dump_partition(msg, part)
212 #define dump_all_partitions(env)
215 #if defined(VERIFY_MONOTONE) && defined (DEBUG_libfirm)
217 * Verify that a type transition is monotone
219 static void verify_type(const lattice_elem_t old_type, const lattice_elem_t new_type) {
220 if (old_type.tv == new_type.tv) {
224 if (old_type.tv == tarval_top) {
225 /* from Top down-to is always allowed */
228 if (old_type.tv == tarval_reachable) {
229 panic("verify_type(): wrong translation from %+F to %+F", old_type, new_type);
231 if (new_type.tv == tarval_bottom || new_type.tv == tarval_reachable) {
235 panic("verify_type(): wrong translation from %+F to %+F", old_type, new_type);
238 #define verify_type(old_type, new_type)
242 * Compare two pointer values of a listmap.
244 static int listmap_cmp_ptr(const void *elt, const void *key, size_t size) {
245 const listmap_entry_t *e1 = elt;
246 const listmap_entry_t *e2 = key;
249 return e1->id != e2->id;
250 } /* listmap_cmp_ptr */
253 * Initializes a listmap.
255 * @param map the listmap
257 static void listmap_init(listmap_t *map) {
258 map->map = new_set(listmap_cmp_ptr, 16);
263 * Terminates a listmap.
265 * @param map the listmap
267 static void listmap_term(listmap_t *map) {
272 * Return the associated listmap entry for a given id.
274 * @param map the listmap
275 * @param id the id to search for
277 * @return the asociated listmap entry for the given id
279 static listmap_entry_t *listmap_find(listmap_t *map, void *id) {
280 listmap_entry_t key, *entry;
285 entry = set_insert(map->map, &key, sizeof(key), HASH_PTR(id));
287 if (entry->list == NULL) {
288 /* a new entry, put into the list */
289 entry->next = map->values;
296 * Calculate the hash value for an opcode map entry.
298 * @param entry an opcode map entry
300 * @return a hash value for the given opcode map entry
302 static unsigned opcode_hash(const opcode_key_t *entry) {
303 return (entry->mode - (ir_mode *)0) * 9 + entry->code + entry->u.proj * 3 + HASH_PTR(entry->u.ent);
307 * Compare two entries in the opcode map.
309 static int cmp_opcode(const void *elt, const void *key, size_t size) {
310 const opcode_key_t *o1 = elt;
311 const opcode_key_t *o2 = key;
314 return o1->code != o2->code || o1->mode != o2->mode ||
315 o1->u.proj != o2->u.proj || o1->u.ent != o2->u.ent;
319 * Compare two Def-Use edges for input position.
321 static int cmp_def_use_edge(const void *a, const void *b) {
322 const ir_def_use_edge *ea = a;
323 const ir_def_use_edge *eb = b;
325 /* no overrun, because range is [-1, MAXINT] */
326 return ea->pos - eb->pos;
327 } /* cmp_def_use_edge */
330 * We need the Def-Use edges sorted.
332 static void sort_irn_outs(node_t *node) {
333 ir_node *irn = node->node;
334 int n_outs = get_irn_n_outs(irn);
337 qsort(&irn->out[1], n_outs, sizeof(irn->out[0]), cmp_def_use_edge);
339 node->max_user_input = irn->out[n_outs].pos;
340 } /* sort_irn_outs */
343 * Return the type of a node.
345 * @param irn an IR-node
347 * @return the associated type of this node
349 static INLINE lattice_elem_t get_node_type(const ir_node *irn) {
350 return get_irn_node(irn)->type;
351 } /* get_node_type */
354 * Return the tarval of a node.
356 * @param irn an IR-node
358 * @return the associated type of this node
360 static INLINE tarval *get_node_tarval(const ir_node *irn) {
361 lattice_elem_t type = get_node_type(irn);
363 if (is_tarval(type.tv))
365 return tarval_bottom;
366 } /* get_node_type */
369 * Add a partition to the worklist.
371 static INLINE void add_to_worklist(partition_t *X, environment_t *env) {
372 assert(X->on_worklist == 0);
373 X->wl_next = env->worklist;
379 * Create a new empty partition.
381 * @param env the environment
383 * @return a newly allocated partition
385 static INLINE partition_t *new_partition(environment_t *env) {
386 partition_t *part = obstack_alloc(&env->obst, sizeof(*part));
388 INIT_LIST_HEAD(&part->entries);
389 INIT_LIST_HEAD(&part->cprop);
390 part->wl_next = NULL;
391 part->touched_next = NULL;
392 part->cprop_next = NULL;
393 part->split_next = NULL;
394 part->touched = NULL;
397 part->max_user_inputs = 0;
398 part->on_worklist = 0;
399 part->on_touched = 0;
402 part->dbg_next = env->dbg_list;
403 env->dbg_list = part;
404 part->nr = part_nr++;
408 } /* new_partition */
411 * Get the first node from a partition.
413 static INLINE node_t *get_first_node(const partition_t *X) {
414 return list_entry(X->entries.next, node_t, node_list);
418 * Return the type of a partition (assuming partition is non-empty and
419 * all elements have the same type).
421 * @param X a partition
423 * @return the type of the first element of the partition
425 static INLINE lattice_elem_t get_partition_type(const partition_t *X) {
426 const node_t *first = get_first_node(X);
428 } /* get_partition_type */
431 * Creates a partition node for the given IR-node and place it
432 * into the given partition.
434 * @param irn an IR-node
435 * @param part a partition to place the node in
436 * @param env the environment
438 * @return the created node
440 static node_t *create_partition_node(ir_node *irn, partition_t *part, environment_t *env) {
441 /* create a partition node and place it in the partition */
442 node_t *node = obstack_alloc(&env->obst, sizeof(*node));
444 INIT_LIST_HEAD(&node->node_list);
445 INIT_LIST_HEAD(&node->cprop_list);
449 node->type.tv = tarval_top;
450 node->max_user_input = 0;
452 node->on_touched = 0;
455 set_irn_node(irn, node);
457 list_add_tail(&node->node_list, &part->entries);
461 } /* create_partition_node */
464 * Pre-Walker, init all Block-Phi lists.
466 static void init_block_phis(ir_node *irn, void *env) {
470 set_Block_phis(irn, NULL);
475 * Post-Walker, initialize all Nodes' type to U or top and place
476 * all nodes into the TOP partition.
478 static void create_initial_partitions(ir_node *irn, void *ctx) {
479 environment_t *env = ctx;
480 partition_t *part = env->initial;
483 node = create_partition_node(irn, part, env);
485 if (node->max_user_input > part->max_user_inputs)
486 part->max_user_inputs = node->max_user_input;
489 add_Block_phi(get_nodes_block(irn), irn);
491 } /* create_initial_partitions */
494 * Add a partition to the touched set if not already there.
496 * @param part the partition
497 * @param env the environment
499 static INLINE void add_to_touched(partition_t *part, environment_t *env) {
500 if (part->on_touched == 0) {
501 part->touched_next = env->touched;
503 part->on_touched = 1;
505 } /* add_to_touched */
508 * Add a node to the entry.partition.touched set if not already there.
512 static INLINE void add_to_partition_touched(node_t *y) {
513 if (y->on_touched == 0) {
514 partition_t *part = y->part;
516 y->next = part->touched;
521 } /* add_to_partition_touched */
524 * Update the worklist: If Z is on worklist then add Z' to worklist.
525 * Else add the smaller of Z and Z' to worklist.
527 * @param Z the Z partition
528 * @param Z_prime the Z' partition, a previous part of Z
529 * @param env the environment
531 static void update_worklist(partition_t *Z, partition_t *Z_prime, environment_t *env) {
532 if (Z->on_worklist || Z_prime->n_nodes < Z->n_nodes) {
533 add_to_worklist(Z_prime, env);
535 add_to_worklist(Z, env);
537 } /* update_worklist */
540 * Split a partition by a local list.
542 * @param Z the Z partition to split
543 * @param g a (non-empty) node list
544 * @param env the environment
546 * @return a new partition containing the nodes of g
548 static partition_t *split(partition_t *Z, node_t *g, environment_t *env) {
549 partition_t *Z_prime;
552 int max_input, max_arity, arity;
554 dump_partition("Splitting ", Z);
558 /* Remove g from Z. */
559 for (node = g; node != NULL; node = node->next) {
560 list_del(&node->node_list);
563 assert(n < Z->n_nodes);
566 /* Move g to a new partition, Z
\92. */
567 Z_prime = new_partition(env);
568 max_arity = max_input = 0;
569 for (node = g; node != NULL; node = node->next) {
570 list_add(&node->node_list, &Z_prime->entries);
571 node->part = Z_prime;
572 arity = get_irn_arity(node->node);
573 if (arity > max_arity)
575 if (node->max_user_input > max_input)
576 max_input = node->max_user_input;
578 Z_prime->max_user_inputs = max_input;
579 Z_prime->n_nodes = n;
581 update_worklist(Z, Z_prime, env);
583 dump_partition("Now ", Z);
584 dump_partition("Created new ", Z_prime);
589 * Returns non-zero if the i'th input of a Phi node is live.
591 * @param phi a Phi-node
592 * @param i an input number
594 * @return non-zero if the i'th input of the given Phi node is live
596 static int is_live_input(ir_node *phi, int i) {
598 ir_node *block = get_nodes_block(phi);
599 ir_node *pred = get_Block_cfgpred(block, i);
600 lattice_elem_t type = get_node_type(pred);
602 return type.tv != tarval_unreachable;
604 /* else it's the control input, always live */
606 } /* is_live_input */
609 * Return non-zero if a type is a constant.
611 static int is_constant_type(lattice_elem_t type) {
612 if (type.tv != tarval_bottom && type.tv != tarval_top)
615 } /* is_constant_type */
618 * Place a node on the cprop list.
621 * @param env the environment
623 static void add_node_to_cprop(node_t *y, environment_t *env) {
624 /* Add y to y.partition.cprop. */
625 if (y->on_cprop == 0) {
626 partition_t *Y = y->part;
628 list_add_tail(&y->cprop_list, &Y->cprop);
631 DB((dbg, LEVEL_3, "Add %+F to part%u.cprop\n", y->node, Y->nr));
633 /* place its partition on the cprop list */
634 if (Y->on_cprop == 0) {
635 Y->cprop_next = env->cprop;
640 if (get_irn_mode(y->node) == mode_T) {
641 /* mode_T nodes always produce tarval_bottom, so we must explicitly
642 add it's Proj's to get constant evaluation to work */
645 for (i = get_irn_n_outs(y->node) - 1; i >= 0; --i) {
646 node_t *proj = get_irn_node(get_irn_out(y->node, i));
648 add_node_to_cprop(proj, env);
652 if (is_Block(y->node)) {
653 /* Due to the way we handle Phi's, we must place all Phis of a block on the list
654 * if someone placed the block. The Block is only placed if the reachability
655 * changes, and this must be re-evaluated in compute_Phi(). */
657 for (phi = get_Block_phis(y->node); phi != NULL; phi = get_Phi_next(phi)) {
658 node_t *p = get_irn_node(phi);
659 add_node_to_cprop(p, env);
662 } /* add_node_to_cprop */
665 * Check whether a type is neither Top or a constant.
666 * Note: U is handled like Top here, R is a constant.
668 * @param type the type to check
670 static int type_is_neither_top_nor_const(const lattice_elem_t type) {
671 if (is_tarval(type.tv)) {
672 if (type.tv == tarval_top)
674 if (tarval_is_constant(type.tv))
684 * Split the partitions if caused by the first entry on the worklist.
686 * @param env the environment
688 static void cause_splits(environment_t *env) {
689 partition_t *X, *Y, *Z;
695 /* remove the first partition from the worklist */
697 env->worklist = X->wl_next;
700 dump_partition("Cause_split: ", X);
701 end_idx = env->end_idx;
702 for (i = -1; i <= X->max_user_inputs; ++i) {
703 /* empty the touched set: already done, just clear the list */
706 list_for_each_entry(node_t, x, &X->entries, node_list) {
712 num_edges = get_irn_n_outs(x->node);
714 while (x->next_edge <= num_edges) {
715 ir_def_use_edge *edge = &x->node->out[x->next_edge];
717 /* check if we have necessary edges */
725 /* ignore the "control input" for non-pinned nodes
726 if we are running in GCSE mode */
727 if (i < end_idx && get_irn_pinned(succ) != op_pin_state_pinned)
730 y = get_irn_node(succ);
731 if (is_constant_type(y->type)) {
732 code = get_irn_opcode(succ);
733 if (code == iro_Sub || code == iro_Cmp)
734 add_node_to_cprop(y, env);
737 /* Partitions of constants should not be split simply because their Nodes have unequal
738 functions or incongruent inputs. */
739 if (type_is_neither_top_nor_const(y->type) &&
740 (! is_Phi(y->node) || is_live_input(y->node, i))) {
742 add_to_touched(Y, env);
743 add_to_partition_touched(y);
748 for (Z = env->touched; Z != NULL; Z = Z->touched_next) {
749 /* remove it from the touched set */
752 if (Z->n_nodes != Z->n_touched) {
753 DB((dbg, LEVEL_2, "Split part%d by touched\n", Z->nr));
754 split(Z, Z->touched, env);
756 /* Empty local Z.touched. */
757 for (e = Z->touched; e != NULL; e = e->next) {
767 * Implements split_by_what(): Split a partition by characteristics given
768 * by the what function.
770 * @param X the partition to split
771 * @param What a function returning an Id for every node of the partition X
772 * @param P a list to store the result partitions
773 * @param env the environment
777 static partition_t *split_by_what(partition_t *X, what_func What,
778 partition_t **P, environment_t *env) {
781 listmap_entry_t *iter;
784 /* Let map be an empty mapping from the range of What to (local) list of Nodes. */
786 list_for_each_entry(node_t, x, &X->entries, node_list) {
787 void *id = What(x, env);
788 listmap_entry_t *entry;
791 /* input not allowed, ignore */
794 /* Add x to map[What(x)]. */
795 entry = listmap_find(&map, id);
796 x->next = entry->list;
799 /* Let P be a set of Partitions. */
801 /* for all sets S except one in the range of map do */
802 for (iter = map.values; iter != NULL; iter = iter->next) {
803 if (iter->next == NULL) {
804 /* this is the last entry, ignore */
809 /* Add SPLIT( X, S ) to P. */
810 DB((dbg, LEVEL_2, "Split part%d by what\n", X->nr));
811 R = split(X, S, env);
821 } /* split_by_what */
823 /** lambda n.(n.type) */
824 static void *lambda_type(const node_t *node, environment_t *env) {
826 return node->type.tv;
829 /** lambda n.(n.opcode) */
830 static void *lambda_opcode(const node_t *node, environment_t *env) {
831 opcode_key_t key, *entry;
832 ir_node *irn = node->node;
834 key.code = get_irn_opcode(irn);
835 key.mode = get_irn_mode(irn);
839 switch (get_irn_opcode(irn)) {
841 key.u.proj = get_Proj_proj(irn);
844 key.u.ent = get_Sel_entity(irn);
850 entry = set_insert(env->opcode2id_map, &key, sizeof(key), opcode_hash(&key));
852 } /* lambda_opcode */
854 /** lambda n.(n[i].partition) */
855 static void *lambda_partition(const node_t *node, environment_t *env) {
856 ir_node *skipped = skip_Proj(node->node);
859 int i = env->lambda_input;
861 if (i >= get_irn_arity(node->node)) {
862 /* we are outside the allowed range */
866 /* ignore the "control input" for non-pinned nodes
867 if we are running in GCSE mode */
868 if (i < env->end_idx && get_irn_pinned(skipped) != op_pin_state_pinned)
871 pred = i == -1 ? get_irn_n(skipped, i) : get_irn_n(node->node, i);
872 p = get_irn_node(pred);
875 } /* lambda_partition */
878 * Checks whether a type is a constant.
880 static int is_type_constant(lattice_elem_t type) {
881 if (is_tarval(type.tv))
882 return tarval_is_constant(type.tv);
883 /* else it is a symconst */
888 * Implements split_by().
890 * @param X the partition to split
891 * @param env the environment
893 static void split_by(partition_t *X, environment_t *env) {
894 partition_t *P = NULL;
897 DB((dbg, LEVEL_2, "WHAT = lambda n.(n.type) on part%d\n", X->nr));
898 P = split_by_what(X, lambda_type, &P, env);
903 if (Y->n_nodes > 1) {
904 lattice_elem_t type = get_partition_type(Y);
906 /* we do not want split the TOP or constant partitions */
907 if (type.tv != tarval_top && !is_type_constant(type)) {
908 partition_t *Q = NULL;
910 DB((dbg, LEVEL_2, "WHAT = lambda n.(n.opcode) on part%d\n", Y->nr));
911 Q = split_by_what(Y, lambda_opcode, &Q, env);
917 if (Z->n_nodes > 1) {
918 const node_t *first = get_first_node(Z);
919 int arity = get_irn_arity(first->node);
923 * BEWARE: during splitting by input 2 for instance we might
924 * create new partitions which are different by input 1, so collect
925 * them and split further.
927 Z->split_next = NULL;
930 for (input = arity - 1; input >= -1; --input) {
932 partition_t *Z_prime = R;
935 if (Z_prime->n_nodes > 1) {
936 env->lambda_input = input;
937 DB((dbg, LEVEL_2, "WHAT = lambda n.(n[%d].partition) on part%d\n", input, Z_prime->nr));
938 S = split_by_what(Z_prime, lambda_partition, &S, env);
940 Z_prime->split_next = S;
955 * (Re-)compute the type for a given node.
957 * @param node the node
959 static void default_compute(node_t *node) {
961 ir_node *irn = node->node;
962 node_t *block = get_irn_node(get_nodes_block(irn));
964 if (block->type.tv == tarval_unreachable) {
965 node->type.tv = tarval_top;
969 /* if any of the data inputs have type top, the result is type top */
970 for (i = get_irn_arity(irn) - 1; i >= 0; --i) {
971 ir_node *pred = get_irn_n(irn, i);
972 node_t *p = get_irn_node(pred);
974 if (p->type.tv == tarval_top) {
975 node->type.tv = tarval_top;
980 if (get_irn_mode(node->node) == mode_X)
981 node->type.tv = tarval_reachable;
983 node->type.tv = computed_value(irn);
984 } /* default_compute */
987 * (Re-)compute the type for a Block node.
989 * @param node the node
991 static void compute_Block(node_t *node) {
993 ir_node *block = node->node;
995 for (i = get_Block_n_cfgpreds(block) - 1; i >= 0; --i) {
996 node_t *pred = get_irn_node(get_Block_cfgpred(block, i));
998 if (pred->type.tv == tarval_reachable) {
999 /* A block is reachable, if at least of predecessor is reachable. */
1000 node->type.tv = tarval_reachable;
1004 node->type.tv = tarval_top;
1005 } /* compute_Block */
1008 * (Re-)compute the type for a Bad node.
1010 * @param node the node
1012 static void compute_Bad(node_t *node) {
1013 /* Bad nodes ALWAYS compute Top */
1014 node->type.tv = tarval_top;
1018 * (Re-)compute the type for an Unknown node.
1020 * @param node the node
1022 static void compute_Unknown(node_t *node) {
1023 /* While Unknown nodes compute Top, but this is dangerous:
1024 * a if (unknown) would lead to BOTH control flows unreachable.
1025 * While this is correct in the given semantics, it would destroy the Firm
1027 * For now, we compute bottom here.
1029 node->type.tv = tarval_bottom;
1030 } /* compute_Unknown */
1033 * (Re-)compute the type for a Jmp node.
1035 * @param node the node
1037 static void compute_Jmp(node_t *node) {
1038 node_t *block = get_irn_node(get_nodes_block(node->node));
1040 node->type = block->type;
1044 * (Re-)compute the type for the End node.
1046 * @param node the node
1048 static void compute_End(node_t *node) {
1049 /* the End node is NOT dead of course */
1050 node->type.tv = tarval_reachable;
1054 * (Re-)compute the type for a SymConst node.
1056 * @param node the node
1058 static void compute_SymConst(node_t *node) {
1059 ir_node *irn = node->node;
1060 node_t *block = get_irn_node(get_nodes_block(irn));
1062 if (block->type.tv == tarval_unreachable) {
1063 node->type.tv = tarval_top;
1066 switch (get_SymConst_kind(irn)) {
1067 case symconst_addr_ent:
1068 /* case symconst_addr_name: cannot handle this yet */
1069 node->type.sym = get_SymConst_symbol(irn);
1072 node->type.tv = computed_value(irn);
1074 } /* compute_SymConst */
1077 * (Re-)compute the type for a Phi node.
1079 * @param node the node
1081 static void compute_Phi(node_t *node) {
1083 ir_node *phi = node->node;
1084 lattice_elem_t type;
1086 /* if a Phi is in a unreachable block, its type is TOP */
1087 node_t *block = get_irn_node(get_nodes_block(phi));
1089 if (block->type.tv == tarval_unreachable) {
1090 node->type.tv = tarval_top;
1094 /* Phi implements the Meet operation */
1095 type.tv = tarval_top;
1096 for (i = get_Phi_n_preds(phi) - 1; i >= 0; --i) {
1097 node_t *pred = get_irn_node(get_Phi_pred(phi, i));
1098 node_t *pred_X = get_irn_node(get_Block_cfgpred(block->node, i));
1100 if (pred_X->type.tv == tarval_unreachable || pred->type.tv == tarval_top) {
1101 /* ignore TOP inputs: We must check here for unreachable blocks,
1102 because Firm constants live in the Start Block are NEVER Top.
1103 Else, a Phi (1,2) will produce Bottom, even if the 2 for instance
1104 comes from a unreachable input. */
1107 if (pred->type.tv == tarval_bottom) {
1108 node->type.tv = tarval_bottom;
1110 } else if (type.tv == tarval_top) {
1111 /* first constant found */
1113 } else if (type.tv != pred->type.tv) {
1114 /* different constants or tarval_bottom */
1115 node->type.tv = tarval_bottom;
1118 /* else nothing, constants are the same */
1124 * (Re-)compute the type for an Add. Special case: one nodes is a Zero Const.
1126 * @param node the node
1128 static void compute_Add(node_t *node) {
1129 ir_node *sub = node->node;
1130 node_t *l = get_irn_node(get_Add_left(sub));
1131 node_t *r = get_irn_node(get_Add_right(sub));
1132 lattice_elem_t a = l->type;
1133 lattice_elem_t b = r->type;
1134 node_t *block = get_irn_node(get_nodes_block(sub));
1137 if (block->type.tv == tarval_unreachable) {
1138 node->type.tv = tarval_top;
1142 if (a.tv == tarval_top || b.tv == tarval_top) {
1143 node->type.tv = tarval_top;
1144 } else if (a.tv == tarval_bottom || b.tv == tarval_bottom) {
1145 node->type.tv = tarval_bottom;
1147 /* x + 0 = 0 + x = x, but beware of floating point +0 + -0, so we
1148 must call tarval_add() first to handle this case! */
1149 if (is_tarval(a.tv)) {
1150 if (is_tarval(b.tv)) {
1151 node->type.tv = tarval_add(a.tv, b.tv);
1154 mode = get_tarval_mode(a.tv);
1155 if (a.tv == get_mode_null(mode)) {
1159 } else if (is_tarval(b.tv)) {
1160 mode = get_tarval_mode(b.tv);
1161 if (b.tv == get_mode_null(mode)) {
1166 node->type.tv = tarval_bottom;
1171 * Returns true if a type is a constant.
1173 static int is_con(const lattice_elem_t type) {
1174 return is_entity(type.sym.entity_p) || tarval_is_constant(type.tv);
1178 * (Re-)compute the type for a Sub. Special case: both nodes are congruent.
1180 * @param node the node
1182 static void compute_Sub(node_t *node) {
1183 ir_node *sub = node->node;
1184 node_t *l = get_irn_node(get_Sub_left(sub));
1185 node_t *r = get_irn_node(get_Sub_right(sub));
1186 lattice_elem_t a = l->type;
1187 lattice_elem_t b = r->type;
1188 node_t *block = get_irn_node(get_nodes_block(sub));
1190 if (block->type.tv == tarval_unreachable) {
1191 node->type.tv = tarval_top;
1194 if (a.tv == tarval_top || b.tv == tarval_top) {
1195 node->type.tv = tarval_top;
1196 } else if (is_con(a) && is_con(b)) {
1197 if (is_tarval(a.tv) && is_tarval(b.tv)) {
1198 node->type.tv = tarval_sub(a.tv, b.tv, get_irn_mode(sub));
1199 } else if (is_tarval(a.tv) && tarval_is_null(a.tv)) {
1201 } else if (is_tarval(b.tv) && tarval_is_null(b.tv)) {
1204 node->type.tv = tarval_bottom;
1206 } else if (r->part == l->part &&
1207 (!mode_is_float(get_irn_mode(l->node)))) {
1208 if (node->type.tv == tarval_top) {
1210 * BEWARE: a - a is NOT always 0 for floating Point values, as
1211 * NaN op NaN = NaN, so we must check this here.
1213 ir_mode *mode = get_irn_mode(sub);
1214 node->type.tv = get_mode_null(mode);
1216 node->type.tv = tarval_bottom;
1219 node->type.tv = tarval_bottom;
1224 * (Re-)compute the type for Cmp.
1226 * @param node the node
1228 static void compute_Cmp(node_t *node) {
1229 ir_node *cmp = node->node;
1230 node_t *l = get_irn_node(get_Cmp_left(cmp));
1231 node_t *r = get_irn_node(get_Cmp_right(cmp));
1232 lattice_elem_t a = l->type;
1233 lattice_elem_t b = r->type;
1235 if (a.tv == tarval_top || b.tv == tarval_top) {
1236 node->type.tv = tarval_top;
1237 } else if (is_con(a) && is_con(b)) {
1238 /* both nodes are constants, we can propbably do something */
1239 node->type.tv = tarval_b_true;
1240 } else if (r->part == l->part) {
1241 /* both nodes congruent, we can probably do something */
1242 node->type.tv = tarval_b_true;
1244 node->type.tv = tarval_bottom;
1246 } /* compute_Proj_Cmp */
1249 * (Re-)compute the type for a Proj(Cmp).
1251 * @param node the node
1252 * @param cond the predecessor Cmp node
1254 static void compute_Proj_Cmp(node_t *node, ir_node *cmp) {
1255 ir_node *proj = node->node;
1256 node_t *l = get_irn_node(get_Cmp_left(cmp));
1257 node_t *r = get_irn_node(get_Cmp_right(cmp));
1258 lattice_elem_t a = l->type;
1259 lattice_elem_t b = r->type;
1260 pn_Cmp pnc = get_Proj_proj(proj);
1262 if (a.tv == tarval_top || b.tv == tarval_top) {
1263 node->type.tv = tarval_top;
1264 } else if (is_con(a) && is_con(b)) {
1265 default_compute(node);
1266 } else if (r->part == l->part &&
1267 (!mode_is_float(get_irn_mode(l->node)) || pnc == pn_Cmp_Lt || pnc == pn_Cmp_Gt)) {
1268 if (node->type.tv == tarval_top) {
1270 * BEWARE: a == a is NOT always True for floating Point values, as
1271 * NaN != NaN is defined, so we must check this here.
1273 node->type.tv = new_tarval_from_long(pnc & pn_Cmp_Eq, mode_b);
1275 node->type.tv = tarval_bottom;
1278 node->type.tv = tarval_bottom;
1280 } /* compute_Proj_Cmp */
1283 * (Re-)compute the type for a Proj(Cond).
1285 * @param node the node
1286 * @param cond the predecessor Cond node
1288 static void compute_Proj_Cond(node_t *node, ir_node *cond) {
1289 ir_node *proj = node->node;
1290 long pnc = get_Proj_proj(proj);
1291 ir_node *sel = get_Cond_selector(cond);
1292 node_t *selector = get_irn_node(sel);
1294 if (get_irn_mode(sel) == mode_b) {
1296 if (pnc == pn_Cond_true) {
1297 if (selector->type.tv == tarval_b_false) {
1298 node->type.tv = tarval_unreachable;
1299 } else if (selector->type.tv == tarval_b_true) {
1300 node->type.tv = tarval_reachable;
1301 } else if (selector->type.tv == tarval_bottom) {
1302 node->type.tv = tarval_reachable;
1304 assert(selector->type.tv == tarval_top);
1305 node->type.tv = tarval_unreachable;
1308 assert(pnc == pn_Cond_false);
1310 if (selector->type.tv == tarval_b_false) {
1311 node->type.tv = tarval_reachable;
1312 } else if (selector->type.tv == tarval_b_true) {
1313 node->type.tv = tarval_unreachable;
1314 } else if (selector->type.tv == tarval_bottom) {
1315 node->type.tv = tarval_reachable;
1317 assert(selector->type.tv == tarval_top);
1318 node->type.tv = tarval_unreachable;
1323 if (selector->type.tv == tarval_bottom) {
1324 node->type.tv = tarval_reachable;
1325 } else if (selector->type.tv == tarval_top) {
1326 node->type.tv = tarval_unreachable;
1328 long value = get_tarval_long(selector->type.tv);
1329 if (pnc == get_Cond_defaultProj(cond)) {
1330 /* default switch, have to check ALL other cases */
1333 for (i = get_irn_n_outs(cond) - 1; i >= 0; --i) {
1334 ir_node *succ = get_irn_out(cond, i);
1338 if (value == get_Proj_proj(succ)) {
1339 /* we found a match, will NOT take the default case */
1340 node->type.tv = tarval_unreachable;
1344 /* all cases checked, no match, will take default case */
1345 node->type.tv = tarval_reachable;
1348 node->type.tv = value == pnc ? tarval_reachable : tarval_unreachable;
1352 } /* compute_Proj_Cond */
1355 * (Re-)compute the type for a Proj-Nodes.
1357 * @param node the node
1359 static void compute_Proj(node_t *node) {
1360 ir_node *proj = node->node;
1361 ir_mode *mode = get_irn_mode(proj);
1362 node_t *block = get_irn_node(get_nodes_block(skip_Proj(proj)));
1363 ir_node *pred = get_Proj_pred(proj);
1365 if (get_Proj_proj(proj) == pn_Start_X_initial_exec && is_Start(pred)) {
1366 /* The initial_exec node is ALWAYS reachable. */
1367 node->type.tv = tarval_reachable;
1371 if (block->type.tv == tarval_unreachable) {
1372 /* a Proj in a unreachable Block stay Top */
1373 node->type.tv = tarval_top;
1376 if (get_irn_node(pred)->type.tv == tarval_top) {
1377 /* if the predecessor is Top, its Proj follow */
1378 node->type.tv = tarval_top;
1382 if (mode == mode_M) {
1383 /* mode M is always bottom */
1384 node->type.tv = tarval_bottom;
1387 if (mode != mode_X) {
1389 compute_Proj_Cmp(node, pred);
1391 default_compute(node);
1394 /* handle mode_X nodes */
1396 switch (get_irn_opcode(pred)) {
1398 /* the Proj_X from the Start is always reachable.
1399 However this is already handled at the top. */
1400 node->type.tv = tarval_reachable;
1403 compute_Proj_Cond(node, pred);
1406 default_compute(node);
1408 } /* compute_Proj */
1411 * (Re-)compute the type for a Confirm-Nodes.
1413 * @param node the node
1415 static void compute_Confirm(node_t *node) {
1416 ir_node *confirm = node->node;
1417 node_t *pred = get_irn_node(get_Confirm_value(confirm));
1419 if (get_Confirm_cmp(confirm) == pn_Cmp_Eq) {
1420 node_t *bound = get_irn_node(get_Confirm_bound(confirm));
1422 if (is_con(bound->type)) {
1423 /* is equal to a constant */
1424 node->type = bound->type;
1428 /* a Confirm is a copy OR a Const */
1429 node->type = pred->type;
1430 } /* compute_Confirm */
1433 * (Re-)compute the type for a given node.
1435 * @param node the node
1437 static void compute(node_t *node) {
1438 compute_func func = (compute_func)node->node->op->ops.generic;
1445 * Propagate constant evaluation.
1447 * @param env the environment
1449 static void propagate(environment_t *env) {
1452 lattice_elem_t old_type;
1457 while (env->cprop != NULL) {
1458 /* remove the first partition X from cprop */
1461 env->cprop = X->cprop_next;
1463 DB((dbg, LEVEL_2, "Propagate type on part%d\n", X->nr));
1466 while (! list_empty(&X->cprop)) {
1467 /* remove the first Node x from X.cprop */
1468 x = list_entry(X->cprop.next, node_t, cprop_list);
1469 list_del(&x->cprop_list);
1472 /* compute a new type for x */
1474 DB((dbg, LEVEL_3, "computing type of %+F\n", x->node));
1476 if (x->type.tv != old_type.tv) {
1477 verify_type(old_type, x->type);
1478 DB((dbg, LEVEL_2, "node %+F has changed type from %+F to %+F\n", x->node, old_type, x->type));
1480 if (x->on_fallen == 0) {
1481 /* Add x to fallen. Nodes might fall from T -> const -> _|_, so check that they are
1482 not already on the list. */
1487 DB((dbg, LEVEL_2, "Add node %+F to fallen\n", x->node));
1489 for (i = get_irn_n_outs(x->node) - 1; i >= 0; --i) {
1490 ir_node *succ = get_irn_out(x->node, i);
1491 node_t *y = get_irn_node(succ);
1493 /* Add y to y.partition.cprop. */
1494 add_node_to_cprop(y, env);
1499 if (n_fallen > 0 && n_fallen != X->n_nodes) {
1500 DB((dbg, LEVEL_2, "Splitting part%d by fallen\n", X->nr));
1501 Y = split(X, fallen, env);
1505 /* remove the nodes from the fallen list */
1506 for (x = fallen; x != NULL; x = x->next)
1515 * Get the leader for a given node from its congruence class.
1517 * @param irn the node
1519 static ir_node *get_leader(node_t *node) {
1520 partition_t *part = node->part;
1522 if (part->n_nodes > 1) {
1523 DB((dbg, LEVEL_2, "Found congruence class for %+F\n", node->node));
1525 return get_first_node(part)->node;
1531 * Return non-zero if the control flow predecessor node pred
1532 * is the only reachable control flow exit of its block.
1534 * @param pred the control flow exit
1536 static int can_exchange(ir_node *pred) {
1539 else if (is_Jmp(pred))
1541 else if (get_irn_mode(pred) == mode_T) {
1544 /* if the predecessor block has more than one
1545 reachable outputs we cannot remove the block */
1547 for (i = get_irn_n_outs(pred) - 1; i >= 0; --i) {
1548 ir_node *proj = get_irn_out(pred, i);
1551 /* skip non-control flow Proj's */
1552 if (get_irn_mode(proj) != mode_X)
1555 node = get_irn_node(proj);
1556 if (node->type.tv == tarval_reachable) {
1567 * Block Post-Walker, apply the analysis results on control flow by
1568 * shortening Phi's and Block inputs.
1570 static void apply_cf(ir_node *block, void *ctx) {
1571 node_t *node = get_irn_node(block);
1573 ir_node **ins, **in_X;
1574 ir_node *phi, *next;
1577 if (block == get_irg_end_block(current_ir_graph) ||
1578 block == get_irg_start_block(current_ir_graph)) {
1579 /* the EndBlock is always reachable even if the analysis
1580 finds out the opposite :-) */
1583 if (node->type.tv == tarval_unreachable) {
1584 /* mark dead blocks */
1585 set_Block_dead(block);
1589 n = get_Block_n_cfgpreds(block);
1592 /* only one predecessor combine */
1593 ir_node *pred = skip_Proj(get_Block_cfgpred(block, 0));
1595 if (can_exchange(pred))
1596 exchange(block, get_nodes_block(pred));
1600 NEW_ARR_A(ir_node *, in_X, n);
1602 for (i = 0; i < n; ++i) {
1603 ir_node *pred = get_Block_cfgpred(block, i);
1604 node_t *node = get_irn_node(pred);
1606 if (node->type.tv == tarval_reachable) {
1613 NEW_ARR_A(ir_node *, ins, n);
1614 for (phi = get_Block_phis(block); phi != NULL; phi = next) {
1615 node_t *node = get_irn_node(phi);
1617 next = get_Phi_next(phi);
1618 if (is_tarval(node->type.tv) && tarval_is_constant(node->type.tv)) {
1619 /* this Phi is replaced by a constant */
1620 tarval *tv = node->type.tv;
1621 ir_node *c = new_r_Const(current_ir_graph, block, get_tarval_mode(tv), tv);
1623 set_irn_node(c, node);
1625 DB((dbg, LEVEL_1, "%+F is replaced by %+F\n", phi, c));
1629 for (i = 0; i < n; ++i) {
1630 node_t *pred = get_irn_node(get_Block_cfgpred(block, i));
1632 if (pred->type.tv == tarval_reachable) {
1633 ins[j++] = get_Phi_pred(phi, i);
1637 /* this Phi is replaced by a single predecessor */
1638 ir_node *s = ins[0];
1641 DB((dbg, LEVEL_1, "%+F is replaced by %+F\n", phi, s));
1644 set_irn_in(phi, j, ins);
1650 /* this Block has only one live predecessor */
1651 ir_node *pred = skip_Proj(in_X[0]);
1653 if (can_exchange(pred))
1654 exchange(block, get_nodes_block(pred));
1656 set_irn_in(block, k, in_X);
1661 * Post-Walker, apply the analysis results;
1663 static void apply_result(ir_node *irn, void *ctx) {
1664 node_t *node = get_irn_node(irn);
1667 if (is_Block(irn) || is_End(irn) || is_Bad(irn)) {
1668 /* blocks already handled, do not touch the End node */
1670 node_t *block = get_irn_node(get_nodes_block(irn));
1672 if (block->type.tv == tarval_unreachable) {
1673 ir_node *bad = get_irg_bad(current_ir_graph);
1675 /* here, bad might already have a node, but this can be safely ignored
1676 as long as bad has at least ONE valid node */
1677 set_irn_node(bad, node);
1679 DB((dbg, LEVEL_1, "%+F is unreachable\n", irn));
1682 else if (node->type.tv == tarval_unreachable) {
1683 ir_node *bad = get_irg_bad(current_ir_graph);
1685 /* see comment above */
1686 set_irn_node(bad, node);
1688 DB((dbg, LEVEL_1, "%+F is unreachable\n", irn));
1691 else if (get_irn_mode(irn) == mode_X) {
1694 ir_node *cond = get_Proj_pred(irn);
1696 if (is_Cond(cond)) {
1697 node_t *sel = get_irn_node(get_Cond_selector(cond));
1699 if (is_tarval(sel->type.tv) && tarval_is_constant(sel->type.tv)) {
1700 /* Cond selector is a constant, make a Jmp */
1701 ir_node *jmp = new_r_Jmp(current_ir_graph, block->node);
1702 set_irn_node(jmp, node);
1704 DB((dbg, LEVEL_1, "%+F is replaced by %+F\n", irn, jmp));
1710 /* normal data node */
1711 if (is_tarval(node->type.tv) && tarval_is_constant(node->type.tv)) {
1712 tarval *tv = node->type.tv;
1714 if (! is_Const(irn)) {
1715 /* can be replaced by a constant */
1716 ir_node *c = new_r_Const(current_ir_graph, block->node, get_tarval_mode(tv), tv);
1717 set_irn_node(c, node);
1719 DB((dbg, LEVEL_1, "%+F is replaced by %+F\n", irn, c));
1722 } else if (is_entity(node->type.sym.entity_p)) {
1723 if (! is_SymConst(irn)) {
1724 /* can be replaced by a Symconst */
1725 ir_node *symc = new_r_SymConst(current_ir_graph, block->node, get_irn_mode(irn), node->type.sym, symconst_addr_ent);
1726 set_irn_node(symc, node);
1729 DB((dbg, LEVEL_1, "%+F is replaced by %+F\n", irn, symc));
1730 exchange(irn, symc);
1733 ir_node *leader = get_leader(node);
1735 if (leader != irn) {
1736 DB((dbg, LEVEL_1, "%+F from part%d is replaced by %+F\n", irn, node->part->nr, leader));
1737 exchange(irn, leader);
1742 } /* apply_result */
1744 #define SET(code) op_##code->ops.generic = (op_func)compute_##code
1747 * sets the generic functions to compute.
1749 static void set_compute_functions(void) {
1752 /* set the default compute function */
1753 for (i = get_irp_n_opcodes() - 1; i >= 0; --i) {
1754 ir_op *op = get_irp_opcode(i);
1755 op->ops.generic = (op_func)default_compute;
1758 /* set specific functions */
1771 } /* set_compute_functions */
1773 static int dump_partition_hook(FILE *F, ir_node *n, ir_node *local) {
1774 ir_node *irn = local != NULL ? local : n;
1775 node_t *node = get_irn_node(irn);
1777 ir_fprintf(F, "info2 : \"partition %u type %+F\"\n", node->part->nr, node->type);
1781 void combo(ir_graph *irg) {
1785 ir_graph *rem = current_ir_graph;
1787 current_ir_graph = irg;
1789 /* register a debug mask */
1790 FIRM_DBG_REGISTER(dbg, "firm.opt.combo");
1791 //firm_dbg_set_mask(dbg, SET_LEVEL_3);
1793 DB((dbg, LEVEL_1, "Doing COMBO for %+F\n", irg));
1795 obstack_init(&env.obst);
1796 env.worklist = NULL;
1800 #ifdef DEBUG_libfirm
1801 env.dbg_list = NULL;
1803 env.opcode2id_map = new_set(cmp_opcode, iro_Last * 4);
1804 env.type2id_map = pmap_create();
1805 env.end_idx = get_opt_global_cse() ? 0 : -1;
1806 env.lambda_input = 0;
1808 assure_irg_outs(irg);
1810 /* we have our own value_of function */
1811 set_value_of_func(get_node_tarval);
1813 set_compute_functions();
1814 DEBUG_ONLY(part_nr = 0);
1816 /* create the initial partition and place it on the work list */
1817 env.initial = new_partition(&env);
1818 add_to_worklist(env.initial, &env);
1819 irg_walk_graph(irg, init_block_phis, create_initial_partitions, &env);
1821 /* Place the START Node's partition on cprop.
1822 Place the START Node on its local worklist. */
1823 initial_X = get_irg_initial_exec(irg);
1824 start = get_irn_node(initial_X);
1825 add_node_to_cprop(start, &env);
1829 if (env.worklist != NULL)
1831 } while (env.cprop != NULL || env.worklist != NULL);
1833 dump_all_partitions(&env);
1836 set_dump_node_vcgattr_hook(dump_partition_hook);
1837 dump_ir_block_graph(irg, "-partition");
1838 set_dump_node_vcgattr_hook(NULL);
1840 (void)dump_partition_hook;
1843 /* apply the result */
1844 irg_block_walk_graph(irg, NULL, apply_cf, &env);
1845 irg_walk_graph(irg, NULL, apply_result, &env);
1847 pmap_destroy(env.type2id_map);
1848 del_set(env.opcode2id_map);
1849 obstack_free(&env.obst, NULL);
1851 /* restore value_of() default behavior */
1852 set_value_of_func(NULL);
1853 current_ir_graph = rem;