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 we use the terminology from Click's work here, which is different
27 * in some cases from Firm terminology. Especially, Click's type is a
28 * Firm tarval, nevertheless we call it type here for "maximum compatibility".
36 #include "iroptimize.h"
44 #include "irgraph_t.h"
58 typedef struct node_t node_t;
59 typedef struct partition_t partition_t;
60 typedef struct opcode_key_t opcode_key_t;
61 typedef struct listmap_entry_t listmap_entry_t;
63 /** The type of the compute function. */
64 typedef void (*compute_func)(node_t *node);
70 ir_opcode code; /**< The Firm opcode. */
71 ir_mode *mode; /**< The mode of all nodes in the partition. */
73 long proj; /**< For Proj nodes, its proj number */
74 ir_entity *ent; /**< For Sel Nodes, its entity */
79 * An entry in the list_map.
81 struct listmap_entry_t {
82 void *id; /**< The id. */
83 node_t *list; /**< The associated list for this id. */
84 listmap_entry_t *next; /**< Link to the next entry in the map. */
87 /** We must map id's to lists. */
88 typedef struct listmap_t {
89 set *map; /**< Map id's to listmap_entry_t's */
90 listmap_entry_t *values; /**< List of all values in the map. */
94 * A lattice element. Because we handle constants and symbolic constants different, we
95 * have to use this union.
106 ir_node *node; /**< The IR-node itself. */
107 list_head node_list; /**< Double-linked list of entries. */
108 list_head cprop_list; /**< Double-linked partition.cprop list. */
109 partition_t *part; /**< points to the partition this node belongs to */
110 node_t *next; /**< Next node on local list (partition.touched, fallen). */
111 lattice_elem_t type; /**< The associated lattice element "type". */
112 int max_user_input; /**< Maximum input number of Def-Use edges. */
113 int next_edge; /**< Index of the next Def-Use edge to use. */
114 unsigned on_touched:1; /**< Set, if this node is on the partition.touched set. */
115 unsigned on_cprop:1; /**< Set, if this node is on the partition.cprop list. */
116 unsigned on_fallen:1; /**< Set, if this node is on the fallen list. */
120 * A partition containing congruent nodes.
123 list_head entries; /**< The head of partition node list. */
124 list_head cprop; /**< The head of partition.cprop list. */
125 partition_t *wl_next; /**< Next entry in the work list if any. */
126 partition_t *touched_next; /**< Points to the next partition in the touched set. */
127 partition_t *cprop_next; /**< Points to the next partition in the cprop list. */
128 node_t *touched; /**< The partition.touched set of this partition. */
129 unsigned n_nodes; /**< Number of entries in this partition. */
130 unsigned n_touched; /**< Number of entries in the partition.touched. */
131 int max_arity; /**< Maximum arity of all entries. */
132 int max_user_inputs; /**< Maximum number of user inputs of all entries. */
133 unsigned on_worklist:1; /**< Set, if this partition is in the work list. */
134 unsigned on_touched:1; /**< Set, if this partition is on the touched set. */
135 unsigned on_cprop:1; /**< Set, if this partition is on the cprop list. */
137 partition_t *dbg_next; /**< Link all partitions for debugging */
138 unsigned nr; /**< A unique number for (what-)mapping, >0. */
142 typedef struct environment_t {
143 struct obstack obst; /**< obstack to allocate data structures. */
144 partition_t *worklist; /**< The work list. */
145 partition_t *cprop; /**< The constant propagation list. */
146 partition_t *touched; /**< the touched set. */
147 partition_t *initial; /**< The initial partition. */
149 partition_t *dbg_list; /**< List of all partitions. */
151 set *opcode2id_map; /**< The opcodeMode->id map. */
152 pmap *type2id_map; /**< The type->id map. */
153 int end_idx; /**< -1 for local and 0 for global congruences. */
154 int lambda_input; /**< Captured argument for lambda_partition(). */
157 /** Type of the what function. */
158 typedef void *(*what_func)(const node_t *node, environment_t *env);
160 #define get_irn_node(irn) ((node_t *)get_irn_link(irn))
161 #define set_irn_node(irn, node) set_irn_link(irn, node)
163 /** The debug module handle. */
164 DEBUG_ONLY(static firm_dbg_module_t *dbg;)
166 /** Next partition number. */
167 DEBUG_ONLY(static unsigned part_nr = 0);
170 static INLINE lattice_elem_t get_partition_type(const partition_t *X);
173 * Dump partition to output.
175 static void dump_partition(const char *msg, const partition_t *part) {
178 lattice_elem_t type = get_partition_type(part);
180 DB((dbg, LEVEL_2, "%s part%u (%u, %+F) {\n ", msg, part->nr, part->n_nodes, type));
181 list_for_each_entry(node_t, node, &part->entries, node_list) {
182 DB((dbg, LEVEL_2, "%s%+F", first ? "" : ", ", node->node));
185 DB((dbg, LEVEL_2, "\n}\n"));
189 * Dump all partitions.
191 static void dump_all_partitions(const environment_t *env) {
192 const partition_t *P;
194 DB((dbg, LEVEL_2, "All partitions\n===============\n"));
195 for (P = env->dbg_list; P != NULL; P = P->dbg_next)
196 dump_partition("", P);
200 #define dump_partition(msg, part)
201 #define dump_all_partitions(env)
205 * Return the "top" value depending on the mode
207 static tarval *get_top_value(const ir_mode *mode) {
208 return (mode == mode_X || mode == mode_BB) ? tarval_unreachable : tarval_top;
212 * Compare two pointer values of a listmap.
214 static int listmap_cmp_ptr(const void *elt, const void *key, size_t size) {
215 const listmap_entry_t *e1 = elt;
216 const listmap_entry_t *e2 = key;
219 return e1->id != e2->id;
220 } /* listmap_cmp_ptr */
223 * Initializes a listmap.
225 * @param map the listmap
227 static void listmap_init(listmap_t *map) {
228 map->map = new_set(listmap_cmp_ptr, 16);
233 * Terminates a listmap.
235 * @param map the listmap
237 static void listmap_term(listmap_t *map) {
242 * Return the associated listmap entry for a given id.
244 * @param map the listmap
245 * @param id the id to search for
247 * @return the asociated listmap entry for the given id
249 static listmap_entry_t *listmap_find(listmap_t *map, void *id) {
250 listmap_entry_t key, *entry;
255 entry = set_insert(map->map, &key, sizeof(key), HASH_PTR(id));
257 if (entry->list == NULL) {
258 /* a new entry, put into the list */
259 entry->next = map->values;
266 * Calculate the hash value for an opcode map entry.
268 * @param entry an opcode map entry
270 * @return a hash value for the given opcode map entry
272 static unsigned opcode_hash(const opcode_key_t *entry) {
273 return (entry->mode - (ir_mode *)0) * 9 + entry->code + entry->u.proj * 3 + HASH_PTR(entry->u.ent);
277 * Compare two entries in the opcode map.
279 static int cmp_opcode(const void *elt, const void *key, size_t size) {
280 const opcode_key_t *o1 = elt;
281 const opcode_key_t *o2 = key;
284 return o1->code != o2->code || o1->mode != o2->mode ||
285 o1->u.proj != o2->u.proj || o1->u.ent != o2->u.ent;
289 * Compare two Def-Use edges for input position.
291 static int cmp_def_use_edge(const void *a, const void *b) {
292 const ir_def_use_edge *ea = a;
293 const ir_def_use_edge *eb = b;
295 /* no overrun, because range is [-1, MAXINT] */
296 return ea->pos - eb->pos;
297 } /* cmp_def_use_edge */
300 * We need the Def-Use edges sorted.
302 static void sort_irn_outs(node_t *node) {
303 ir_node *irn = node->node;
304 int n_outs = get_irn_n_outs(irn);
307 qsort(&irn->out[1], n_outs, sizeof(irn->out[0]), cmp_def_use_edge);
309 node->max_user_input = irn->out[n_outs + 1].pos;
310 } /* sort_irn_outs */
313 * Return the type of a node.
315 * @param irn an IR-node
317 * @return the associated type of this node
319 static INLINE lattice_elem_t get_node_type(const ir_node *irn) {
320 return get_irn_node(irn)->type;
321 } /* get_node_type */
324 * Return the tarval of a node.
326 * @param irn an IR-node
328 * @return the associated type of this node
330 static INLINE tarval *get_node_tarval(const ir_node *irn) {
331 lattice_elem_t type = get_node_type(irn);
333 if (is_tarval(type.tv))
335 return tarval_bottom;
336 } /* get_node_type */
339 * Add a partition to the worklist.
341 static INLINE void add_to_worklist(partition_t *X, environment_t *env) {
342 assert(X->on_worklist == 0);
343 X->wl_next = env->worklist;
349 * Create a new empty partition.
351 * @param env the environment
353 * @return a newly allocated partition
355 static INLINE partition_t *new_partition(environment_t *env) {
356 partition_t *part = obstack_alloc(&env->obst, sizeof(*part));
358 INIT_LIST_HEAD(&part->entries);
359 INIT_LIST_HEAD(&part->cprop);
360 part->wl_next = NULL;
361 part->touched_next = NULL;
362 part->cprop_next = NULL;
363 part->touched = NULL;
367 part->max_user_inputs = 0;
368 part->on_worklist = 0;
369 part->on_touched = 0;
372 part->dbg_next = env->dbg_list;
373 env->dbg_list = part;
374 part->nr = part_nr++;
378 } /* new_partition */
381 * Get the first node from a partition.
383 static INLINE node_t *get_first_node(const partition_t *X) {
384 return list_entry(X->entries.next, node_t, node_list);
388 * Return the type of a partition (assuming partition is non-empty and
389 * all elements have the same type).
391 * @param X a partition
393 * @return the type of the first element of the partition
395 static INLINE lattice_elem_t get_partition_type(const partition_t *X) {
396 const node_t *first = get_first_node(X);
398 } /* get_partition_type */
401 * Creates a partition node for the given IR-node and place it
402 * into the given partition.
404 * @param irn an IR-node
405 * @param part a partition to place the node in
406 * @param env the environment
408 * @return the created node
410 static node_t *create_partition_node(ir_node *irn, partition_t *part, environment_t *env) {
411 /* create a partition node and place it in the partition */
412 node_t *node = obstack_alloc(&env->obst, sizeof(*node));
413 ir_mode *mode = get_irn_mode(irn);
415 INIT_LIST_HEAD(&node->node_list);
416 INIT_LIST_HEAD(&node->cprop_list);
420 node->type.tv = get_top_value(mode);
421 node->max_user_input = 0;
423 node->on_touched = 0;
426 set_irn_node(irn, node);
428 list_add_tail(&node->node_list, &part->entries);
432 } /* create_partition_node */
435 * Pre-Walker, init all Block-Phi lists.
437 static void init_block_phis(ir_node *irn, void *env) {
441 set_Block_phis(irn, NULL);
446 * Post-Walker, initialize all Nodes' type to U or top and place
447 * all nodes into the TOP partition.
449 static void create_initial_partitions(ir_node *irn, void *ctx) {
450 environment_t *env = ctx;
451 partition_t *part = env->initial;
455 node = create_partition_node(irn, part, env);
457 arity = get_irn_arity(irn);
458 if (arity > part->max_arity)
459 part->max_arity = arity;
460 if (node->max_user_input > part->max_user_inputs)
461 part->max_user_inputs = node->max_user_input;
464 add_Block_phi(get_nodes_block(irn), irn);
466 } /* create_initial_partitions */
469 * Add a partition to the touched set if not already there.
471 * @param part the partition
472 * @param env the environment
474 static INLINE void add_to_touched(partition_t *part, environment_t *env) {
475 if (part->on_touched == 0) {
476 part->touched_next = env->touched;
478 part->on_touched = 1;
480 } /* add_to_touched */
483 * Add a node to the entry.partition.touched set if not already there.
487 static INLINE void add_to_partition_touched(node_t *y) {
488 if (y->on_touched == 0) {
489 partition_t *part = y->part;
491 y->next = part->touched;
496 } /* add_to_partition_touched */
499 * Update the worklist: If Z is on worklist then add Z' to worklist.
500 * Else add the smaller of Z and Z' to worklist.
502 * @param Z the Z partition
503 * @param Z_prime the Z' partition, a previous part of Z
504 * @param env the environment
506 static void update_worklist(partition_t *Z, partition_t *Z_prime, environment_t *env) {
507 if (Z->on_worklist || Z_prime->n_nodes < Z->n_nodes) {
508 add_to_worklist(Z_prime, env);
510 add_to_worklist(Z, env);
512 } /* update_worklist */
515 * Split a partition by a local list.
517 * @param Z the Z partition to split
518 * @param g a (non-empty) node list
519 * @param env the environment
521 * @return a new partition containing the nodes of g
523 static partition_t *split(partition_t *Z, node_t *g, environment_t *env) {
524 partition_t *Z_prime;
527 int max_input, max_arity, arity;
529 dump_partition("Splitting ", Z);
533 /* Remove g from Z. */
534 for (node = g; node != NULL; node = node->next) {
535 list_del(&node->node_list);
538 assert(n < Z->n_nodes);
541 /* Move g to a new partition, Z
\92. */
542 Z_prime = new_partition(env);
543 max_arity = max_input = 0;
544 for (node = g; node != NULL; node = node->next) {
545 list_add(&node->node_list, &Z_prime->entries);
546 node->part = Z_prime;
547 arity = get_irn_arity(node->node);
548 if (arity > max_arity)
550 if (node->max_user_input > max_input)
551 max_input = node->max_user_input;
553 Z_prime->max_arity = max_arity;
554 Z_prime->max_user_inputs = max_input;
555 Z_prime->n_nodes = n;
557 update_worklist(Z, Z_prime, env);
559 dump_partition("Now ", Z);
560 dump_partition("Created new ", Z_prime);
565 * Returns non-zero if the i'th input of a Phi node is live.
567 * @param phi a Phi-node
568 * @param i an input number
570 * @return non-zero if the i'th input of the given Phi node is live
572 static int is_live_input(ir_node *phi, int i) {
574 ir_node *block = get_nodes_block(phi);
575 ir_node *pred = get_Block_cfgpred(block, i);
576 lattice_elem_t type = get_node_type(pred);
578 return type.tv != tarval_unreachable;
580 /* else it's the control input, always live */
582 } /* is_live_input */
585 * Return non-zero if a type is a constant.
587 static int is_constant_type(lattice_elem_t type) {
588 if (type.tv != tarval_bottom && type.tv != tarval_top)
591 } /* is_constant_type */
594 * Place a node on the cprop list.
597 * @param env the environment
599 static void add_node_to_cprop(node_t *y, environment_t *env) {
600 /* Add y to y.partition.cprop. */
601 if (y->on_cprop == 0) {
602 partition_t *Y = y->part;
604 list_add_tail(&y->cprop_list, &Y->cprop);
607 DB((dbg, LEVEL_3, "Add %+F to part%u.cprop\n", y->node, Y->nr));
609 /* place its partition on the cprop list */
610 if (Y->on_cprop == 0) {
611 Y->cprop_next = env->cprop;
616 if (get_irn_mode(y->node) == mode_T) {
617 /* mode_T nodes always produce tarval_bottom, so we must explicitly
618 add it's Proj's to get constant evaluation to work */
621 for (i = get_irn_n_outs(y->node) - 1; i >= 0; --i) {
622 node_t *proj = get_irn_node(get_irn_out(y->node, i));
624 add_node_to_cprop(proj, env);
627 if (is_Block(y->node)) {
628 /* Due to the way we handle Phi's, we must place all Phis of a block on the list
629 * if someone placeis the block. The Block is only placed if the reachability
630 * changes, and this must be re-evaluated in compute_Phi(). */
632 for (phi = get_Block_phis(y->node); phi != NULL; phi = get_Phi_next(phi)) {
633 node_t *p = get_irn_node(phi);
634 add_node_to_cprop(p, env);
637 } /* add_node_to_cprop */
640 * Split the partitions if caused by the first entry on the worklist.
642 * @param env the environment
644 static void cause_splits(environment_t *env) {
645 partition_t *X, *Y, *Z;
651 /* remove the first partition from the worklist */
653 env->worklist = X->wl_next;
656 dump_partition("Cause_split: ", X);
657 end_idx = env->end_idx;
658 for (i = -1; i <= X->max_user_inputs; ++i) {
659 /* empty the touched set: already done, just clear the list */
662 list_for_each_entry(node_t, x, &X->entries, node_list) {
668 num_edges = get_irn_n_outs(x->node);
670 while (x->next_edge <= num_edges) {
671 ir_def_use_edge *edge = &x->node->out[x->next_edge];
673 /* check if we have necessary edges */
681 /* ignore the "control input" for non-pinned nodes
682 if we are running in GCSE mode */
683 if (i < end_idx && get_irn_pinned(succ) != op_pin_state_pinned)
686 y = get_irn_node(succ);
687 if (is_constant_type(y->type)) {
688 code = get_irn_opcode(succ);
689 if (code == iro_Sub || (code == iro_Proj && is_Cmp(get_Proj_pred(succ))))
690 add_node_to_cprop(y, env);
693 /* Partitions of constants should not be split simply because their Nodes have unequal
694 functions or incongruent inputs. */
695 if (y->type.tv == tarval_bottom &&
696 (! is_Phi(y->node) || is_live_input(y->node, i))) {
698 add_to_touched(Y, env);
699 add_to_partition_touched(y);
704 for (Z = env->touched; Z != NULL; Z = Z->touched_next) {
705 /* remove it from the touched set */
708 if (Z->n_nodes != Z->n_touched) {
709 DB((dbg, LEVEL_2, "Split part%d by touched\n", Z->nr));
710 split(Z, Z->touched, env);
712 /* Empty local Z.touched. */
713 for (e = Z->touched; e != NULL; e = e->next) {
723 * Implements split_by_what(): Split a partition by characteristics given
724 * by the what function.
726 * @param X the partition to split
727 * @param What a function returning an Id for every node of the partition X
728 * @param P an flexible array to store the result partitions or NULL
729 * @param env the environment
731 * @return if P != NULL P will be filled with the resulting partitions and returned
733 static partition_t **split_by_what(partition_t *X, what_func What,
734 partition_t **P, environment_t *env) {
737 listmap_entry_t *iter;
740 /* Let map be an empty mapping from the range of What to (local) list of Nodes. */
742 list_for_each_entry(node_t, x, &X->entries, node_list) {
743 void *id = What(x, env);
744 listmap_entry_t *entry;
747 /* input not allowed, ignore */
750 /* Add x to map[What(x)]. */
751 entry = listmap_find(&map, id);
752 x->next = entry->list;
755 /* Let P be a set of Partitions. */
757 /* for all sets S except one in the range of map do */
758 for (iter = map.values; iter != NULL; iter = iter->next) {
759 if (iter->next == NULL) {
760 /* this is the last entry, ignore */
765 /* Add SPLIT( X, S ) to P. */
766 DB((dbg, LEVEL_2, "Split part%d by what\n", X->nr));
767 R = split(X, S, env);
769 ARR_APP1(partition_t *, P, R);
774 ARR_APP1(partition_t *, P, X);
779 } /* split_by_what */
781 /** lambda n.(n.type) */
782 static void *lambda_type(const node_t *node, environment_t *env) {
784 return node->type.tv;
787 /** lambda n.(n.opcode) */
788 static void *lambda_opcode(const node_t *node, environment_t *env) {
789 opcode_key_t key, *entry;
790 ir_node *irn = node->node;
792 key.code = get_irn_opcode(irn);
793 key.mode = get_irn_mode(irn);
797 switch (get_irn_opcode(irn)) {
799 key.u.proj = get_Proj_proj(irn);
802 key.u.ent = get_Sel_entity(irn);
808 entry = set_insert(env->opcode2id_map, &key, sizeof(key), opcode_hash(&key));
810 } /* lambda_opcode */
812 /** lambda n.(n[i].partition) */
813 static void *lambda_partition(const node_t *node, environment_t *env) {
816 int i = env->lambda_input;
818 if (i >= get_irn_arity(node->node)) {
819 /* we are outside the allowed range */
823 /* ignore the "control input" for non-pinned nodes
824 if we are running in GCSE mode */
825 if (i < env->end_idx && get_irn_pinned(node->node) != op_pin_state_pinned)
828 pred = get_irn_n(node->node, i);
829 p = get_irn_node(pred);
832 } /* lambda_partition */
835 * Checks whether a type is a constant.
837 static int is_type_constant(lattice_elem_t type) {
838 if (is_tarval(type.tv))
839 return tarval_is_constant(type.tv);
840 /* else it is a symconst */
845 * Implements split_by().
847 * @param X the partition to split
848 * @param env the environment
850 static void split_by(partition_t *X, environment_t *env) {
851 partition_t **P = NEW_ARR_F(partition_t *, 0);
854 DB((dbg, LEVEL_2, "WHAT = lambda n.(n.type) on part%d\n", X->nr));
855 P = split_by_what(X, lambda_type, P, env);
856 for (i = ARR_LEN(P) - 1; i >= 0; --i) {
857 partition_t *Y = P[i];
859 if (Y->n_nodes > 1) {
860 lattice_elem_t type = get_partition_type(Y);
862 /* we do not want split the TOP, unreachable or constant partitions */
863 if (type.tv != tarval_top && type.tv != tarval_unreachable && !is_type_constant(type)) {
864 partition_t **Q = NEW_ARR_F(partition_t *, 0);
866 DB((dbg, LEVEL_2, "WHAT = lambda n.(n.opcode) on part%d\n", Y->nr));
867 Q = split_by_what(Y, lambda_opcode, Q, env);
869 for (j = ARR_LEN(Q) - 1; j >= 0; --j) {
870 partition_t *Z = Q[j];
872 for (k = Z->max_arity - 1; k >= -1; --k) {
873 if (Z->n_nodes > 1) {
874 env->lambda_input = k;
875 DB((dbg, LEVEL_2, "WHAT = lambda n.(n[%d].partition) on part%d\n", k, Z->nr));
876 split_by_what(Z, lambda_partition, NULL, env);
888 * (Re-)compute the type for a given node.
890 * @param node the node
892 static void default_compute(node_t *node) {
894 ir_node *irn = node->node;
895 tarval *top = tarval_top;
897 if (get_irn_mode(node->node) == mode_X)
898 top = tarval_unreachable;
900 if (get_irn_pinned(irn) == op_pin_state_pinned) {
901 node_t *block = get_irn_node(get_nodes_block(irn));
903 if (block->type.tv == tarval_unreachable) {
909 /* if any of the data inputs have type top, the result is type top */
910 for (i = get_irn_arity(irn) - 1; i >= 0; --i) {
911 ir_node *pred = get_irn_n(irn, i);
912 node_t *p = get_irn_node(pred);
914 if (p->type.tv == tarval_top) {
920 if (get_irn_mode(node->node) == mode_X)
921 node->type.tv = tarval_reachable;
923 node->type.tv = computed_value(irn);
924 } /* default_compute */
927 * (Re-)compute the type for a Block node.
929 * @param node the node
931 static void compute_Block(node_t *node) {
933 ir_node *block = node->node;
935 for (i = get_Block_n_cfgpreds(block) - 1; i >= 0; --i) {
936 node_t *pred = get_irn_node(get_Block_cfgpred(block, i));
938 if (pred->type.tv == tarval_reachable) {
939 /* A block is reachable, if at least of predecessor is reachable. */
940 node->type.tv = tarval_reachable;
944 node->type.tv = tarval_unreachable;
945 } /* compute_Block */
948 * (Re-)compute the type for a Jmp node.
950 * @param node the node
952 static void compute_Jmp(node_t *node) {
953 node_t *block = get_irn_node(get_nodes_block(node->node));
955 node->type = block->type;
959 * (Re-)compute the type for the End node.
961 * @param node the node
963 static void compute_End(node_t *node) {
964 /* the End node is NOT dead of course */
965 node->type.tv = tarval_reachable;
969 * (Re-)compute the type for a SymConst node.
971 * @param node the node
973 static void compute_SymConst(node_t *node) {
974 ir_node *irn = node->node;
975 node_t *block = get_irn_node(get_nodes_block(irn));
977 if (block->type.tv == tarval_unreachable) {
978 node->type.tv = tarval_top;
981 switch (get_SymConst_kind(irn)) {
982 case symconst_addr_ent:
983 /* case symconst_addr_name: cannot handle this yet */
984 node->type.sym = get_SymConst_symbol(irn);
987 node->type.tv = computed_value(irn);
989 } /* compute_SymConst */
992 * (Re-)compute the type for a Phi node.
994 * @param node the node
996 static void compute_Phi(node_t *node) {
998 ir_node *phi = node->node;
1001 /* if a Phi is in a unreachable block, its type is TOP */
1002 node_t *block = get_irn_node(get_nodes_block(phi));
1004 if (block->type.tv == tarval_unreachable) {
1005 node->type.tv = tarval_top;
1009 /* Phi implements the Meet operation */
1010 type.tv = tarval_top;
1011 for (i = get_Phi_n_preds(phi) - 1; i >= 0; --i) {
1012 node_t *pred = get_irn_node(get_Phi_pred(phi, i));
1013 node_t *pred_X = get_irn_node(get_Block_cfgpred(block->node, i));
1015 if (pred_X->type.tv == tarval_unreachable || pred->type.tv == tarval_top) {
1016 /* ignore TOP inputs: We must check here for unreachable blocks,
1017 because Firm constants live in the Start Block are NEVER Top.
1018 Else, a Phi (1,2) will produce Bottom, even if the 2 for instance
1019 comes from a unreachable input. */
1022 if (pred->type.tv == tarval_bottom) {
1023 node->type.tv = tarval_bottom;
1025 } else if (type.tv == tarval_top) {
1026 /* first constant found */
1028 } else if (type.tv != pred->type.tv) {
1029 /* different constants or tarval_bottom */
1030 node->type.tv = tarval_bottom;
1033 /* else nothing, constants are the same */
1039 * (Re-)compute the type for an Add. Special case: one nodes is a Zero Const.
1041 * @param node the node
1043 static void compute_Add(node_t *node) {
1044 ir_node *sub = node->node;
1045 node_t *l = get_irn_node(get_Add_left(sub));
1046 node_t *r = get_irn_node(get_Add_right(sub));
1047 lattice_elem_t a = l->type;
1048 lattice_elem_t b = r->type;
1049 node_t *block = get_irn_node(get_nodes_block(sub));
1052 if (block->type.tv == tarval_unreachable) {
1053 node->type.tv = tarval_top;
1057 if (a.tv == tarval_top || b.tv == tarval_top) {
1058 node->type.tv = tarval_top;
1059 } else if (a.tv == tarval_bottom || b.tv == tarval_bottom) {
1060 node->type.tv = tarval_bottom;
1062 /* x + 0 = 0 + x = x, but beware of floating point +0 + -0, so we
1063 must call tarval_add() first to handle this case! */
1064 if (is_tarval(a.tv)) {
1065 if (is_tarval(b.tv)) {
1066 node->type.tv = tarval_add(a.tv, b.tv);
1069 mode = get_tarval_mode(a.tv);
1070 if (a.tv == get_mode_null(mode)) {
1074 } else if (is_tarval(b.tv)) {
1075 mode = get_tarval_mode(b.tv);
1076 if (b.tv == get_mode_null(mode)) {
1081 node->type.tv = tarval_bottom;
1086 * (Re-)compute the type for a Sub. Special case: both nodes are congruent.
1088 * @param node the node
1090 static void compute_Sub(node_t *node) {
1091 ir_node *sub = node->node;
1092 node_t *l = get_irn_node(get_Sub_left(sub));
1093 node_t *r = get_irn_node(get_Sub_right(sub));
1094 lattice_elem_t a = l->type;
1095 lattice_elem_t b = r->type;
1096 node_t *block = get_irn_node(get_nodes_block(sub));
1098 if (block->type.tv == tarval_unreachable) {
1099 node->type.tv = tarval_top;
1103 if (a.tv == tarval_top || b.tv == tarval_top) {
1104 node->type.tv = tarval_top;
1105 } else if (r->part == l->part) {
1106 ir_mode *mode = get_irn_mode(sub);
1107 node->type.tv = get_mode_null(mode);
1108 } else if (a.tv == tarval_bottom || b.tv == tarval_bottom) {
1109 node->type.tv = tarval_bottom;
1111 if (is_tarval(a.tv) && is_tarval(b.tv))
1112 node->type.tv = tarval_sub(a.tv, b.tv);
1114 node->type.tv = tarval_bottom;
1119 * (Re-)compute the type for a Proj(Cmp).
1121 * @param node the node
1122 * @param cond the predecessor Cmp node
1124 static void compute_Proj_Cmp(node_t *node, ir_node *cmp) {
1125 ir_node *proj = node->node;
1126 node_t *l = get_irn_node(get_Cmp_left(cmp));
1127 node_t *r = get_irn_node(get_Cmp_right(cmp));
1128 lattice_elem_t a = l->type;
1129 lattice_elem_t b = r->type;
1130 pn_Cmp pnc = get_Proj_proj(proj);
1133 * BEWARE: a == a is NOT always True for floating Point values, as
1134 * NaN != NaN is defined, so we must check this here.
1136 if (!mode_is_float(get_irn_mode(l->node)) || pnc == pn_Cmp_Lt || pnc == pn_Cmp_Gt) {
1137 if (a.tv == tarval_top || b.tv == tarval_top) {
1138 node->type.tv = tarval_top;
1139 } else if (r->part == l->part) {
1140 node->type.tv = new_tarval_from_long(pnc & pn_Cmp_Eq, mode_b);
1141 } else if (a.tv == tarval_bottom || b.tv == tarval_bottom) {
1142 node->type.tv = tarval_bottom;
1144 default_compute(node);
1147 default_compute(node);
1149 } /* compute_Proj_Cmp */
1152 * (Re-)compute the type for a Proj(Cond).
1154 * @param node the node
1155 * @param cond the predecessor Cond node
1157 static void compute_Proj_Cond(node_t *node, ir_node *cond) {
1158 ir_node *proj = node->node;
1159 long pnc = get_Proj_proj(proj);
1160 ir_node *sel = get_Cond_selector(cond);
1161 node_t *selector = get_irn_node(sel);
1163 if (get_irn_mode(sel) == mode_b) {
1165 if (pnc == pn_Cond_true) {
1166 if (selector->type.tv == tarval_b_false) {
1167 node->type.tv = tarval_unreachable;
1168 } else if (selector->type.tv == tarval_b_true) {
1169 node->type.tv = tarval_reachable;
1170 } else if (selector->type.tv == tarval_bottom) {
1171 node->type.tv = tarval_reachable;
1173 assert(selector->type.tv == tarval_top);
1174 node->type.tv = tarval_unreachable;
1177 assert(pnc == pn_Cond_false);
1179 if (selector->type.tv == tarval_b_false) {
1180 node->type.tv = tarval_reachable;
1181 } else if (selector->type.tv == tarval_b_true) {
1182 node->type.tv = tarval_unreachable;
1183 } else if (selector->type.tv == tarval_bottom) {
1184 node->type.tv = tarval_reachable;
1186 assert(selector->type.tv == tarval_top);
1187 node->type.tv = tarval_unreachable;
1192 if (selector->type.tv == tarval_bottom) {
1193 node->type.tv = tarval_reachable;
1194 } else if (selector->type.tv == tarval_top) {
1195 node->type.tv = tarval_unreachable;
1197 long value = get_tarval_long(selector->type.tv);
1198 if (pnc == get_Cond_defaultProj(cond)) {
1199 /* default switch, have to check ALL other cases */
1202 for (i = get_irn_n_outs(cond) - 1; i >= 0; --i) {
1203 ir_node *succ = get_irn_out(cond, i);
1207 if (value == get_Proj_proj(succ)) {
1208 /* we found a match, will NOT take the default case */
1209 node->type.tv = tarval_unreachable;
1213 /* all cases checked, no match, will take default case */
1214 node->type.tv = tarval_reachable;
1217 node->type.tv = value == pnc ? tarval_reachable : tarval_unreachable;
1221 } /* compute_Proj_Cond */
1224 * (Re-)compute the type for a Proj-Nodes.
1226 * @param node the node
1228 static void compute_Proj(node_t *node) {
1229 ir_node *proj = node->node;
1230 ir_mode *mode = get_irn_mode(proj);
1231 node_t *block = get_irn_node(get_nodes_block(skip_Proj(proj)));
1232 ir_node *pred = get_Proj_pred(proj);
1234 if (block->type.tv == tarval_unreachable) {
1235 /* a Proj node in an unreachable block computes Top
1236 except if it's the initial_exec node. */
1237 if (get_Proj_proj(proj) != pn_Start_X_initial_exec ||
1239 node->type.tv = get_top_value(mode);
1244 if (mode == mode_M) {
1245 /* mode M is always bottom */
1246 node->type.tv = tarval_bottom;
1249 if (mode != mode_X) {
1251 compute_Proj_Cmp(node, pred);
1253 default_compute(node);
1256 /* handle mode_X nodes */
1258 switch (get_irn_opcode(pred)) {
1260 /* the Proj_X from the Start is always reachable */
1261 node->type.tv = tarval_reachable;
1264 compute_Proj_Cond(node, pred);
1267 default_compute(node);
1269 } /* compute_Proj */
1272 * (Re-)compute the type for a given node.
1274 * @param node the node
1276 static void compute(node_t *node) {
1277 compute_func func = (compute_func)node->node->op->ops.generic;
1284 * Propagate constant evaluation.
1286 * @param env the environment
1288 static void propagate(environment_t *env) {
1291 lattice_elem_t old_type;
1296 while (env->cprop != NULL) {
1297 /* remove the first partition X from cprop */
1300 env->cprop = X->cprop_next;
1304 while (! list_empty(&X->cprop)) {
1305 /* remove the first Node x from X.cprop */
1306 x = list_entry(X->cprop.next, node_t, cprop_list);
1307 list_del(&x->cprop_list);
1310 /* compute a new type for x */
1312 DB((dbg, LEVEL_3, "computing type of %+F\n", x->node));
1314 if (x->type.tv != old_type.tv) {
1315 DB((dbg, LEVEL_2, "node %+F has changed type from %+F to %+F\n", x->node, old_type, x->type));
1317 if (x->on_fallen == 0) {
1318 /* Add x to fallen. Nodes might fall from T -> const -> _|_, so check that they are
1319 not already on the list. */
1325 for (i = get_irn_n_outs(x->node) - 1; i >= 0; --i) {
1326 ir_node *succ = get_irn_out(x->node, i);
1327 node_t *y = get_irn_node(succ);
1329 /* Add y to y.partition.cprop. */
1330 add_node_to_cprop(y, env);
1335 if (n_fallen > 0 && n_fallen != X->n_nodes) {
1336 DB((dbg, LEVEL_2, "Splitting part%d by fallen\n", X->nr));
1337 Y = split(X, fallen, env);
1341 /* remove the nodes from the fallen list */
1342 for (x = fallen; x != NULL; x = x->next)
1351 * Get the leader for a given node from its congruence class.
1353 * @param irn the node
1355 static ir_node *get_leader(node_t *node) {
1356 partition_t *part = node->part;
1358 if (part->n_nodes > 1) {
1359 DB((dbg, LEVEL_2, "Found congruence class for %+F\n", node->node));
1361 return get_first_node(part)->node;
1367 * Post-Walker, apply the analysis results;
1369 static void apply_result(ir_node *irn, void *ctx) {
1370 node_t *node = get_irn_node(irn);
1373 if (is_Block(irn)) {
1374 if (irn == get_irg_end_block(current_ir_graph)) {
1375 /* the EndBlock is always reachable even if the analysis
1376 finds out the opposite :-) */
1380 if (node->type.tv == tarval_unreachable) {
1381 /* mark dead blocks */
1382 set_Block_dead(irn);
1384 } else if (is_End(irn)) {
1385 /* do not touch the End node */
1387 node_t *block = get_irn_node(get_nodes_block(irn));
1389 if (block->type.tv == tarval_unreachable) {
1390 if (! is_Bad(irn)) {
1391 ir_node *bad = get_irg_bad(current_ir_graph);
1393 /* here, bad might already have a node, but this can be safely ignored
1394 as long as bad has at least ONE valid node */
1395 set_irn_node(bad, node);
1397 DB((dbg, LEVEL_1, "%+F is unreachable\n", irn));
1401 else if (get_irn_mode(irn) == mode_X) {
1402 if (node->type.tv == tarval_unreachable) {
1403 ir_node *bad = get_irg_bad(current_ir_graph);
1405 /* see comment above */
1406 set_irn_node(bad, node);
1408 DB((dbg, LEVEL_1, "%+F is unreachable\n", irn));
1411 else if (is_Proj(irn)) {
1413 ir_node *cond = get_Proj_pred(irn);
1415 if (is_Cond(cond)) {
1416 node_t *sel = get_irn_node(get_Cond_selector(cond));
1418 if (is_tarval(sel->type.tv) && tarval_is_constant(sel->type.tv)) {
1419 /* Cond selector is a constant, make a Jmp */
1420 ir_node *jmp = new_r_Jmp(current_ir_graph, block->node);
1421 set_irn_node(jmp, node);
1423 DB((dbg, LEVEL_1, "%+F is replaced by %+F\n", irn, jmp));
1429 /* normal data node */
1430 if (is_tarval(node->type.tv) && tarval_is_constant(node->type.tv)) {
1431 tarval *tv = node->type.tv;
1433 if (! is_Const(irn)) {
1434 /* can be replaced by a constant */
1435 ir_node *c = new_r_Const(current_ir_graph, block->node, get_tarval_mode(tv), tv);
1436 set_irn_node(c, node);
1438 DB((dbg, LEVEL_1, "%+F is replaced by %+F\n", irn, c));
1441 } else if (is_entity(node->type.sym.entity_p)) {
1442 if (! is_SymConst(irn)) {
1443 /* can be replaced by a Symconst */
1444 ir_node *symc = new_r_SymConst(current_ir_graph, block->node, get_irn_mode(irn), node->type.sym, symconst_addr_ent);
1445 set_irn_node(symc, node);
1448 DB((dbg, LEVEL_1, "%+F is replaced by %+F\n", irn, symc));
1449 exchange(irn, symc);
1452 ir_node *leader = get_leader(node);
1454 if (leader != irn) {
1455 DB((dbg, LEVEL_1, "%+F from part%d is replaced by %+F\n", irn, node->part->nr, leader));
1456 exchange(irn, leader);
1461 } /* static void apply_result(ir_node *irn, void *ctx) {
1464 #define SET(code) op_##code->ops.generic = (op_func)compute_##code
1467 * sets the generic functions to compute.
1469 static void set_compute_functions(void) {
1472 /* set the default compute function */
1473 for (i = get_irp_n_opcodes() - 1; i >= 0; --i) {
1474 ir_op *op = get_irp_opcode(i);
1475 op->ops.generic = (op_func)default_compute;
1478 /* set specific functions */
1487 } /* set_compute_functions */
1489 static int dump_partition_hook(FILE *F, ir_node *n, ir_node *local) {
1490 ir_node *irn = local != NULL ? local : n;
1491 node_t *node = get_irn_node(irn);
1493 ir_fprintf(F, "info2 : \"partition %u type %+F\"\n", node->part->nr, node->type);
1497 void combo(ir_graph *irg) {
1501 ir_graph *rem = current_ir_graph;
1503 current_ir_graph = irg;
1505 /* register a debug mask */
1506 FIRM_DBG_REGISTER(dbg, "firm.opt.combo");
1507 firm_dbg_set_mask(dbg, SET_LEVEL_3);
1509 DB((dbg, LEVEL_1, "Doing COMBO for %+F\n", irg));
1511 obstack_init(&env.obst);
1512 env.worklist = NULL;
1516 #ifdef DEBUG_libfirm
1517 env.dbg_list = NULL;
1519 env.opcode2id_map = new_set(cmp_opcode, iro_Last * 4);
1520 env.type2id_map = pmap_create();
1521 env.end_idx = get_opt_global_cse() ? 0 : -1;
1522 env.lambda_input = 0;
1524 assure_irg_outs(irg);
1526 /* we have our own value_of function */
1527 set_value_of_func(get_node_tarval);
1529 set_compute_functions();
1530 DEBUG_ONLY(part_nr = 0);
1532 /* create the initial partition and place it on the work list */
1533 env.initial = new_partition(&env);
1534 add_to_worklist(env.initial, &env);
1535 irg_walk_graph(irg, init_block_phis, create_initial_partitions, &env);
1537 /* Place the START Node's partition on cprop.
1538 Place the START Node on its local worklist. */
1539 initial_X = get_irg_initial_exec(irg);
1540 start = get_irn_node(initial_X);
1541 add_node_to_cprop(start, &env);
1545 if (env.worklist != NULL)
1547 } while (env.cprop != NULL || env.worklist != NULL);
1549 dump_all_partitions(&env);
1551 set_dump_node_vcgattr_hook(dump_partition_hook);
1552 dump_ir_block_graph(irg, "-partition");
1553 set_dump_node_vcgattr_hook(NULL);
1556 /* apply the result */
1557 irg_walk_graph(irg, NULL, apply_result, &env);
1559 pmap_destroy(env.type2id_map);
1560 del_set(env.opcode2id_map);
1561 obstack_free(&env.obst, NULL);
1563 /* restore value_of() default behavior */
1564 set_value_of_func(NULL);
1565 current_ir_graph = rem;