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"
87 #include "opt_manage.h"
92 /* define this to check that all type translations are monotone */
93 #define VERIFY_MONOTONE
95 /* define this to check the consistency of partitions */
96 #define CHECK_PARTITIONS
98 typedef struct node_t node_t;
99 typedef struct partition_t partition_t;
100 typedef struct opcode_key_t opcode_key_t;
101 typedef struct listmap_entry_t listmap_entry_t;
103 /** The type of the compute function. */
104 typedef void (*compute_func)(node_t *node);
109 struct opcode_key_t {
110 ir_node *irn; /**< An IR node representing this opcode. */
114 * An entry in the list_map.
116 struct listmap_entry_t {
117 void *id; /**< The id. */
118 node_t *list; /**< The associated list for this id. */
119 listmap_entry_t *next; /**< Link to the next entry in the map. */
122 /** We must map id's to lists. */
123 typedef struct listmap_t {
124 set *map; /**< Map id's to listmap_entry_t's */
125 listmap_entry_t *values; /**< List of all values in the map. */
129 * A lattice element. Because we handle constants and symbolic constants different, we
130 * have to use this union.
141 ir_node *node; /**< The IR-node itself. */
142 list_head node_list; /**< Double-linked list of leader/follower entries. */
143 list_head cprop_list; /**< Double-linked partition.cprop list. */
144 partition_t *part; /**< points to the partition this node belongs to */
145 node_t *next; /**< Next node on local list (partition.touched, fallen). */
146 node_t *race_next; /**< Next node on race list. */
147 lattice_elem_t type; /**< The associated lattice element "type". */
148 int max_user_input; /**< Maximum input number of Def-Use edges. */
149 int next_edge; /**< Index of the next Def-Use edge to use. */
150 int n_followers; /**< Number of Follower in the outs set. */
151 unsigned on_touched:1; /**< Set, if this node is on the partition.touched set. */
152 unsigned on_cprop:1; /**< Set, if this node is on the partition.cprop list. */
153 unsigned on_fallen:1; /**< Set, if this node is on the fallen list. */
154 unsigned is_follower:1; /**< Set, if this node is a follower. */
155 unsigned flagged:2; /**< 2 Bits, set if this node was visited by race 1 or 2. */
159 * A partition containing congruent nodes.
162 list_head Leader; /**< The head of partition Leader node list. */
163 list_head Follower; /**< The head of partition Follower node list. */
164 list_head cprop; /**< The head of partition.cprop list. */
165 list_head cprop_X; /**< The head of partition.cprop (Cond nodes and its Projs) list. */
166 partition_t *wl_next; /**< Next entry in the work list if any. */
167 partition_t *touched_next; /**< Points to the next partition in the touched set. */
168 partition_t *cprop_next; /**< Points to the next partition in the cprop list. */
169 partition_t *split_next; /**< Points to the next partition in the list that must be split by split_by(). */
170 node_t *touched; /**< The partition.touched set of this partition. */
171 unsigned n_leader; /**< Number of entries in this partition.Leader. */
172 unsigned n_touched; /**< Number of entries in the partition.touched. */
173 int max_user_inputs; /**< Maximum number of user inputs of all entries. */
174 unsigned on_worklist:1; /**< Set, if this partition is in the work list. */
175 unsigned on_touched:1; /**< Set, if this partition is on the touched set. */
176 unsigned on_cprop:1; /**< Set, if this partition is on the cprop list. */
177 unsigned type_is_T_or_C:1;/**< Set, if all nodes in this partition have type Top or Constant. */
179 partition_t *dbg_next; /**< Link all partitions for debugging */
180 unsigned nr; /**< A unique number for (what-)mapping, >0. */
184 typedef struct environment_t {
185 struct obstack obst; /**< obstack to allocate data structures. */
186 partition_t *worklist; /**< The work list. */
187 partition_t *cprop; /**< The constant propagation list. */
188 partition_t *touched; /**< the touched set. */
189 partition_t *initial; /**< The initial partition. */
190 set *opcode2id_map; /**< The opcodeMode->id map. */
191 ir_node **kept_memory; /**< Array of memory nodes that must be kept. */
192 int end_idx; /**< -1 for local and 0 for global congruences. */
193 int lambda_input; /**< Captured argument for lambda_partition(). */
194 unsigned modified:1; /**< Set, if the graph was modified. */
195 unsigned unopt_cf:1; /**< If set, control flow is not optimized due to Unknown. */
196 /* options driving the optimization */
197 unsigned commutative:1; /**< Set, if commutation nodes should be handled specially. */
198 unsigned opt_unknown:1; /**< Set, if non-strict programs should be optimized. */
200 partition_t *dbg_list; /**< List of all partitions. */
204 /** Type of the what function. */
205 typedef void *(*what_func)(const node_t *node, environment_t *env);
207 #define get_irn_node(irn) ((node_t *)get_irn_link(irn))
208 #define set_irn_node(irn, node) set_irn_link(irn, node)
210 /* we do NOT use tarval_unreachable here, instead we use Top for this purpose */
211 #undef tarval_unreachable
212 #define tarval_unreachable tarval_top
215 /** The debug module handle. */
216 DEBUG_ONLY(static firm_dbg_module_t *dbg;)
218 /** The what reason. */
219 DEBUG_ONLY(static const char *what_reason;)
221 /** Next partition number. */
222 DEBUG_ONLY(static unsigned part_nr = 0;)
224 /** The tarval returned by Unknown nodes: set to either tarval_bad OR tarval_top. */
225 static ir_tarval *tarval_UNKNOWN;
228 static node_t *identity(node_t *node);
231 * Compare two opcode representatives.
233 static int cmp_irn_opcode(const ir_node *a, const ir_node *b)
237 if ((get_irn_op(a) != get_irn_op(b)) ||
238 (get_irn_mode(a) != get_irn_mode(b)))
241 /* compare if a's in and b's in are of equal length */
242 arity = get_irn_arity(a);
243 if (arity != get_irn_arity(b))
248 * Some ugliness here: Two Blocks having the same
249 * IJmp predecessor would be congruent, which of course is wrong.
250 * We fix it by never letting blocks be congruent
251 * which cannot be detected by combo either.
257 * here, we already know that the nodes are identical except their
260 if (a->op->ops.node_cmp_attr)
261 return a->op->ops.node_cmp_attr(a, b);
264 } /* cmp_irn_opcode */
266 #ifdef CHECK_PARTITIONS
270 static void check_partition(const partition_t *T)
275 list_for_each_entry(node_t, node, &T->Leader, node_list) {
276 assert(node->is_follower == 0);
277 assert(node->flagged == 0);
278 assert(node->part == T);
281 assert(n == T->n_leader);
283 list_for_each_entry(node_t, node, &T->Follower, node_list) {
284 assert(node->is_follower == 1);
285 assert(node->flagged == 0);
286 assert(node->part == T);
288 } /* check_partition */
291 * check that all leader nodes in the partition have the same opcode.
293 static void check_opcode(const partition_t *Z)
296 const ir_node *repr = NULL;
298 list_for_each_entry(node_t, node, &Z->Leader, node_list) {
299 ir_node *irn = node->node;
304 assert(cmp_irn_opcode(repr, irn) == 0);
309 static void check_all_partitions(environment_t *env)
315 for (P = env->dbg_list; P != NULL; P = P->dbg_next) {
317 if (! P->type_is_T_or_C)
319 list_for_each_entry(node_t, node, &P->Follower, node_list) {
320 node_t *leader = identity(node);
322 assert(leader != node && leader->part == node->part);
333 static void do_check_list(const node_t *list, int ofs, const partition_t *Z)
338 #define NEXT(e) *((const node_t **)((char *)(e) + (ofs)))
339 for (e = list; e != NULL; e = NEXT(e)) {
340 assert(e->part == Z);
348 } /* ido_check_list */
351 * Check a local list.
353 static void check_list(const node_t *list, const partition_t *Z)
355 do_check_list(list, offsetof(node_t, next), Z);
359 #define check_partition(T)
360 #define check_list(list, Z)
361 #define check_all_partitions(env)
362 #endif /* CHECK_PARTITIONS */
365 static inline lattice_elem_t get_partition_type(const partition_t *X);
368 * Dump partition to output.
370 static void dump_partition(const char *msg, const partition_t *part)
374 lattice_elem_t type = get_partition_type(part);
376 DB((dbg, LEVEL_2, "%s part%u%s (%u, %+F) {\n ",
377 msg, part->nr, part->type_is_T_or_C ? "*" : "",
378 part->n_leader, type));
379 list_for_each_entry(node_t, node, &part->Leader, node_list) {
380 DB((dbg, LEVEL_2, "%s%+F", first ? "" : ", ", node->node));
383 if (! list_empty(&part->Follower)) {
384 DB((dbg, LEVEL_2, "\n---\n "));
386 list_for_each_entry(node_t, node, &part->Follower, node_list) {
387 DB((dbg, LEVEL_2, "%s%+F", first ? "" : ", ", node->node));
391 DB((dbg, LEVEL_2, "\n}\n"));
392 } /* dump_partition */
397 static void do_dump_list(const char *msg, const node_t *node, int ofs)
402 #define GET_LINK(p, ofs) *((const node_t **)((char *)(p) + (ofs)))
404 DB((dbg, LEVEL_3, "%s = {\n ", msg));
405 for (p = node; p != NULL; p = GET_LINK(p, ofs)) {
406 DB((dbg, LEVEL_3, "%s%+F", first ? "" : ", ", p->node));
409 DB((dbg, LEVEL_3, "\n}\n"));
417 static void dump_race_list(const char *msg, const node_t *list)
419 do_dump_list(msg, list, offsetof(node_t, race_next));
420 } /* dump_race_list */
423 * Dumps a local list.
425 static void dump_list(const char *msg, const node_t *list)
427 do_dump_list(msg, list, offsetof(node_t, next));
431 * Dump all partitions.
433 static void dump_all_partitions(const environment_t *env)
435 const partition_t *P;
437 DB((dbg, LEVEL_2, "All partitions\n===============\n"));
438 for (P = env->dbg_list; P != NULL; P = P->dbg_next)
439 dump_partition("", P);
440 } /* dump_all_partitions */
445 static void dump_split_list(const partition_t *list)
447 const partition_t *p;
449 DB((dbg, LEVEL_2, "Split by %s produced = {\n", what_reason));
450 for (p = list; p != NULL; p = p->split_next)
451 DB((dbg, LEVEL_2, "part%u, ", p->nr));
452 DB((dbg, LEVEL_2, "\n}\n"));
453 } /* dump_split_list */
456 * Dump partition and type for a node.
458 static int dump_partition_hook(FILE *F, ir_node *n, ir_node *local)
460 ir_node *irn = local != NULL ? local : n;
461 node_t *node = get_irn_node(irn);
463 ir_fprintf(F, "info2 : \"partition %u type %+F\"\n", node->part->nr, node->type);
465 } /* dump_partition_hook */
468 #define dump_partition(msg, part)
469 #define dump_race_list(msg, list)
470 #define dump_list(msg, list)
471 #define dump_all_partitions(env)
472 #define dump_split_list(list)
475 #if defined(VERIFY_MONOTONE) && defined (DEBUG_libfirm)
477 * Verify that a type transition is monotone
479 static void verify_type(const lattice_elem_t old_type, node_t *node)
481 if (old_type.tv == node->type.tv) {
485 if (old_type.tv == tarval_top) {
486 /* from Top down-to is always allowed */
489 if (node->type.tv == tarval_bottom || node->type.tv == tarval_reachable) {
493 panic("combo: wrong translation from %+F to %+F on node %+F", old_type, node->type, node->node);
497 #define verify_type(old_type, node)
501 * Compare two pointer values of a listmap.
503 static int listmap_cmp_ptr(const void *elt, const void *key, size_t size)
505 const listmap_entry_t *e1 = (listmap_entry_t*)elt;
506 const listmap_entry_t *e2 = (listmap_entry_t*)key;
509 return e1->id != e2->id;
510 } /* listmap_cmp_ptr */
513 * Initializes a listmap.
515 * @param map the listmap
517 static void listmap_init(listmap_t *map)
519 map->map = new_set(listmap_cmp_ptr, 16);
524 * Terminates a listmap.
526 * @param map the listmap
528 static void listmap_term(listmap_t *map)
534 * Return the associated listmap entry for a given id.
536 * @param map the listmap
537 * @param id the id to search for
539 * @return the associated listmap entry for the given id
541 static listmap_entry_t *listmap_find(listmap_t *map, void *id)
543 listmap_entry_t key, *entry;
548 entry = (listmap_entry_t*)set_insert(map->map, &key, sizeof(key), HASH_PTR(id));
550 if (entry->list == NULL) {
551 /* a new entry, put into the list */
552 entry->next = map->values;
559 * Calculate the hash value for an opcode map entry.
561 * @param entry an opcode map entry
563 * @return a hash value for the given opcode map entry
565 static unsigned opcode_hash(const opcode_key_t *entry)
567 /* we cannot use the ir ops hash function here, because it hashes the
569 const ir_node *n = entry->irn;
570 ir_opcode code = get_irn_opcode(n);
571 ir_mode *mode = get_irn_mode(n);
572 unsigned hash = (unsigned)(PTR_TO_INT(mode) * 9 + code) + get_irn_arity(n);
574 if (code == iro_Const)
575 hash ^= (unsigned)HASH_PTR(get_Const_tarval(n));
576 else if (code == iro_Proj)
577 hash += (unsigned)get_Proj_proj(n);
582 * Compare two entries in the opcode map.
584 static int cmp_opcode(const void *elt, const void *key, size_t size)
586 const opcode_key_t *o1 = (opcode_key_t*)elt;
587 const opcode_key_t *o2 = (opcode_key_t*)key;
591 return cmp_irn_opcode(o1->irn, o2->irn);
595 * Compare two Def-Use edges for input position.
597 static int cmp_def_use_edge(const void *a, const void *b)
599 const ir_def_use_edge *ea = (const ir_def_use_edge*)a;
600 const ir_def_use_edge *eb = (const ir_def_use_edge*)b;
602 /* no overrun, because range is [-1, MAXINT] */
603 return ea->pos - eb->pos;
604 } /* cmp_def_use_edge */
607 * We need the Def-Use edges sorted.
609 static void sort_irn_outs(node_t *node)
611 ir_node *irn = node->node;
612 int n_outs = get_irn_n_outs(irn);
615 qsort(&irn->out[1], n_outs, sizeof(irn->out[0]), cmp_def_use_edge);
617 node->max_user_input = irn->out[n_outs].pos;
618 } /* sort_irn_outs */
621 * Return the type of a node.
623 * @param irn an IR-node
625 * @return the associated type of this node
627 static inline lattice_elem_t get_node_type(const ir_node *irn)
629 return get_irn_node(irn)->type;
630 } /* get_node_type */
633 * Return the tarval of a node.
635 * @param irn an IR-node
637 * @return the associated type of this node
639 static inline ir_tarval *get_node_tarval(const ir_node *irn)
641 lattice_elem_t type = get_node_type(irn);
643 if (is_tarval(type.tv))
645 return tarval_bottom;
646 } /* get_node_type */
649 * Add a partition to the worklist.
651 static inline void add_to_worklist(partition_t *X, environment_t *env)
653 assert(X->on_worklist == 0);
654 DB((dbg, LEVEL_2, "Adding part%d to worklist\n", X->nr));
655 X->wl_next = env->worklist;
658 } /* add_to_worklist */
661 * Create a new empty partition.
663 * @param env the environment
665 * @return a newly allocated partition
667 static inline partition_t *new_partition(environment_t *env)
669 partition_t *part = OALLOC(&env->obst, partition_t);
671 INIT_LIST_HEAD(&part->Leader);
672 INIT_LIST_HEAD(&part->Follower);
673 INIT_LIST_HEAD(&part->cprop);
674 INIT_LIST_HEAD(&part->cprop_X);
675 part->wl_next = NULL;
676 part->touched_next = NULL;
677 part->cprop_next = NULL;
678 part->split_next = NULL;
679 part->touched = NULL;
682 part->max_user_inputs = 0;
683 part->on_worklist = 0;
684 part->on_touched = 0;
686 part->type_is_T_or_C = 0;
688 part->dbg_next = env->dbg_list;
689 env->dbg_list = part;
690 part->nr = part_nr++;
694 } /* new_partition */
697 * Get the first node from a partition.
699 static inline node_t *get_first_node(const partition_t *X)
701 return list_entry(X->Leader.next, node_t, node_list);
702 } /* get_first_node */
705 * Return the type of a partition (assuming partition is non-empty and
706 * all elements have the same type).
708 * @param X a partition
710 * @return the type of the first element of the partition
712 static inline lattice_elem_t get_partition_type(const partition_t *X)
714 const node_t *first = get_first_node(X);
716 } /* get_partition_type */
719 * Creates a partition node for the given IR-node and place it
720 * into the given partition.
722 * @param irn an IR-node
723 * @param part a partition to place the node in
724 * @param env the environment
726 * @return the created node
728 static node_t *create_partition_node(ir_node *irn, partition_t *part, environment_t *env)
730 /* create a partition node and place it in the partition */
731 node_t *node = OALLOC(&env->obst, node_t);
733 INIT_LIST_HEAD(&node->node_list);
734 INIT_LIST_HEAD(&node->cprop_list);
738 node->race_next = NULL;
739 node->type.tv = tarval_top;
740 node->max_user_input = 0;
742 node->n_followers = 0;
743 node->on_touched = 0;
746 node->is_follower = 0;
748 set_irn_node(irn, node);
750 list_add_tail(&node->node_list, &part->Leader);
754 } /* create_partition_node */
757 * Pre-Walker, initialize all Nodes' type to U or top and place
758 * all nodes into the TOP partition.
760 static void create_initial_partitions(ir_node *irn, void *ctx)
762 environment_t *env = (environment_t*)ctx;
763 partition_t *part = env->initial;
766 node = create_partition_node(irn, part, env);
768 if (node->max_user_input > part->max_user_inputs)
769 part->max_user_inputs = node->max_user_input;
772 set_Block_phis(irn, NULL);
774 } /* create_initial_partitions */
777 * Post-Walker, collect all Block-Phi lists, set Cond.
779 static void init_block_phis(ir_node *irn, void *ctx)
784 ir_node *block = get_nodes_block(irn);
785 add_Block_phi(block, irn);
787 } /* init_block_phis */
790 * Add a node to the entry.partition.touched set and
791 * node->partition to the touched set if not already there.
794 * @param env the environment
796 static inline void add_to_touched(node_t *y, environment_t *env)
798 if (y->on_touched == 0) {
799 partition_t *part = y->part;
801 y->next = part->touched;
806 if (part->on_touched == 0) {
807 part->touched_next = env->touched;
809 part->on_touched = 1;
812 check_list(part->touched, part);
814 } /* add_to_touched */
817 * Place a node on the cprop list.
820 * @param env the environment
822 static void add_to_cprop(node_t *y, environment_t *env)
826 /* Add y to y.partition.cprop. */
827 if (y->on_cprop == 0) {
828 partition_t *Y = y->part;
829 ir_node *irn = y->node;
830 ir_node *skipped = skip_Proj(irn);
832 /* place Conds and all its Projs on the cprop_X list */
833 if (is_Cond(skipped) || is_Switch(skipped))
834 list_add_tail(&y->cprop_list, &Y->cprop_X);
836 list_add_tail(&y->cprop_list, &Y->cprop);
839 DB((dbg, LEVEL_3, "Add %+F to part%u.cprop\n", y->node, Y->nr));
841 /* place its partition on the cprop list */
842 if (Y->on_cprop == 0) {
843 Y->cprop_next = env->cprop;
849 if (get_irn_mode(irn) == mode_T) {
850 /* mode_T nodes always produce tarval_bottom, so we must explicitly
851 * add its Projs to get constant evaluation to work */
854 for (i = get_irn_n_outs(irn) - 1; i >= 0; --i) {
855 node_t *proj = get_irn_node(get_irn_out(irn, i));
857 add_to_cprop(proj, env);
859 } else if (is_Block(irn)) {
860 /* Due to the way we handle Phi's, we must place all Phis of a block on the list
861 * if someone placed the block. The Block is only placed if the reachability
862 * changes, and this must be re-evaluated in compute_Phi(). */
864 for (phi = get_Block_phis(irn); phi != NULL; phi = get_Phi_next(phi)) {
865 node_t *p = get_irn_node(phi);
866 add_to_cprop(p, env);
872 * Update the worklist: If Z is on worklist then add Z' to worklist.
873 * Else add the smaller of Z and Z' to worklist.
875 * @param Z the Z partition
876 * @param Z_prime the Z' partition, a previous part of Z
877 * @param env the environment
879 static void update_worklist(partition_t *Z, partition_t *Z_prime, environment_t *env)
881 if (Z->on_worklist || Z_prime->n_leader < Z->n_leader) {
882 add_to_worklist(Z_prime, env);
884 add_to_worklist(Z, env);
886 } /* update_worklist */
889 * Make all inputs to x no longer be F.def_use edges.
893 static void move_edges_to_leader(node_t *x)
895 ir_node *irn = x->node;
898 for (i = get_irn_arity(irn) - 1; i >= 0; --i) {
899 node_t *pred = get_irn_node(get_irn_n(irn, i));
904 n = get_irn_n_outs(p);
905 for (j = 1; j <= pred->n_followers; ++j) {
906 if (p->out[j].pos == i && p->out[j].use == irn) {
907 /* found a follower edge to x, move it to the Leader */
908 ir_def_use_edge edge = p->out[j];
910 /* remove this edge from the Follower set */
911 p->out[j] = p->out[pred->n_followers];
914 /* sort it into the leader set */
915 for (k = pred->n_followers + 2; k <= n; ++k) {
916 if (p->out[k].pos >= edge.pos)
918 p->out[k - 1] = p->out[k];
920 /* place the new edge here */
921 p->out[k - 1] = edge;
923 /* edge found and moved */
928 } /* move_edges_to_leader */
931 * Split a partition that has NO followers by a local list.
933 * @param Z partition to split
934 * @param g a (non-empty) node list
935 * @param env the environment
937 * @return a new partition containing the nodes of g
939 static partition_t *split_no_followers(partition_t *Z, node_t *g, environment_t *env)
941 partition_t *Z_prime;
946 dump_partition("Splitting ", Z);
947 dump_list("by list ", g);
951 /* Remove g from Z. */
952 for (node = g; node != NULL; node = node->next) {
953 assert(node->part == Z);
954 list_del(&node->node_list);
957 assert(n < Z->n_leader);
960 /* Move g to a new partition, Z'. */
961 Z_prime = new_partition(env);
963 for (node = g; node != NULL; node = node->next) {
964 list_add_tail(&node->node_list, &Z_prime->Leader);
965 node->part = Z_prime;
966 if (node->max_user_input > max_input)
967 max_input = node->max_user_input;
969 Z_prime->max_user_inputs = max_input;
970 Z_prime->n_leader = n;
973 check_partition(Z_prime);
975 /* for now, copy the type info tag, it will be adjusted in split_by(). */
976 Z_prime->type_is_T_or_C = Z->type_is_T_or_C;
978 update_worklist(Z, Z_prime, env);
980 dump_partition("Now ", Z);
981 dump_partition("Created new ", Z_prime);
983 } /* split_no_followers */
986 * Make the Follower -> Leader transition for a node.
990 static void follower_to_leader(node_t *n)
992 assert(n->is_follower == 1);
994 DB((dbg, LEVEL_2, "%+F make the follower -> leader transition\n", n->node));
996 move_edges_to_leader(n);
997 list_del(&n->node_list);
998 list_add_tail(&n->node_list, &n->part->Leader);
1000 } /* follower_to_leader */
1003 * The environment for one race step.
1005 typedef struct step_env {
1006 node_t *initial; /**< The initial node list. */
1007 node_t *unwalked; /**< The unwalked node list. */
1008 node_t *walked; /**< The walked node list. */
1009 int index; /**< Next index of Follower use_def edge. */
1010 unsigned side; /**< side number. */
1014 * Return non-zero, if a input is a real follower
1016 * @param irn the node to check
1017 * @param input number of the input
1019 static int is_real_follower(const ir_node *irn, int input)
1023 switch (get_irn_opcode(irn)) {
1026 /* ignore the Confirm bound input */
1032 /* ignore the Mux sel input */
1037 /* dead inputs are not follower edges */
1038 ir_node *block = get_nodes_block(irn);
1039 node_t *pred = get_irn_node(get_Block_cfgpred(block, input));
1041 if (pred->type.tv == tarval_unreachable)
1051 /* only a Sub x,0 / Shift x,0 might be a follower */
1058 pred = get_irn_node(get_irn_n(irn, input));
1059 if (is_tarval(pred->type.tv) && tarval_is_null(pred->type.tv))
1063 pred = get_irn_node(get_irn_n(irn, input));
1064 if (is_tarval(pred->type.tv) && tarval_is_one(pred->type.tv))
1068 pred = get_irn_node(get_irn_n(irn, input));
1069 if (is_tarval(pred->type.tv) && tarval_is_all_one(pred->type.tv))
1073 assert(!"opcode not implemented yet");
1077 } /* is_real_follower */
1080 * Do one step in the race.
1082 static int step(step_env *env)
1086 if (env->initial != NULL) {
1087 /* Move node from initial to unwalked */
1089 env->initial = n->race_next;
1091 n->race_next = env->unwalked;
1097 while (env->unwalked != NULL) {
1098 /* let n be the first node in unwalked */
1100 while (env->index < n->n_followers) {
1101 const ir_def_use_edge *edge = &n->node->out[1 + env->index];
1103 /* let m be n.F.def_use[index] */
1104 node_t *m = get_irn_node(edge->use);
1106 assert(m->is_follower);
1108 * Some inputs, like the get_Confirm_bound are NOT
1109 * real followers, sort them out.
1111 if (! is_real_follower(m->node, edge->pos)) {
1117 /* only followers from our partition */
1118 if (m->part != n->part)
1121 if ((m->flagged & env->side) == 0) {
1122 m->flagged |= env->side;
1124 if (m->flagged != 3) {
1125 /* visited the first time */
1126 /* add m to unwalked not as first node (we might still need to
1127 check for more follower node */
1128 m->race_next = n->race_next;
1132 /* else already visited by the other side and on the other list */
1135 /* move n to walked */
1136 env->unwalked = n->race_next;
1137 n->race_next = env->walked;
1145 * Clear the flags from a list and check for
1146 * nodes that where touched from both sides.
1148 * @param list the list
1150 static int clear_flags(node_t *list)
1155 for (n = list; n != NULL; n = n->race_next) {
1156 if (n->flagged == 3) {
1157 /* we reach a follower from both sides, this will split congruent
1158 * inputs and make it a leader. */
1159 follower_to_leader(n);
1168 * Split a partition by a local list using the race.
1170 * @param pX pointer to the partition to split, might be changed!
1171 * @param gg a (non-empty) node list
1172 * @param env the environment
1174 * @return a new partition containing the nodes of gg
1176 static partition_t *split(partition_t **pX, node_t *gg, environment_t *env)
1178 partition_t *X = *pX;
1179 partition_t *X_prime;
1182 node_t *g, *h, *node, *t;
1183 int max_input, transitions, winner, shf;
1185 DEBUG_ONLY(static int run = 0;)
1187 DB((dbg, LEVEL_2, "Run %d ", run++));
1188 if (list_empty(&X->Follower)) {
1189 /* if the partition has NO follower, we can use the fast
1190 splitting algorithm. */
1191 return split_no_followers(X, gg, env);
1193 /* else do the race */
1195 dump_partition("Splitting ", X);
1196 dump_list("by list ", gg);
1198 INIT_LIST_HEAD(&tmp);
1200 /* Remove gg from X.Leader and put into g */
1202 for (node = gg; node != NULL; node = node->next) {
1203 assert(node->part == X);
1204 assert(node->is_follower == 0);
1206 list_del(&node->node_list);
1207 list_add_tail(&node->node_list, &tmp);
1208 node->race_next = g;
1213 list_for_each_entry(node_t, node, &X->Leader, node_list) {
1214 node->race_next = h;
1217 /* restore X.Leader */
1218 list_splice(&tmp, &X->Leader);
1220 senv[0].initial = g;
1221 senv[0].unwalked = NULL;
1222 senv[0].walked = NULL;
1226 senv[1].initial = h;
1227 senv[1].unwalked = NULL;
1228 senv[1].walked = NULL;
1233 * Some informations on the race that are not stated clearly in Click's
1235 * 1) A follower stays on the side that reach him first.
1236 * 2) If the other side reches a follower, if will be converted to
1237 * a leader. /This must be done after the race is over, else the
1238 * edges we are iterating on are renumbered./
1239 * 3) /New leader might end up on both sides./
1240 * 4) /If one side ends up with new Leaders, we must ensure that
1241 * they can split out by opcode, hence we have to put _every_
1242 * partition with new Leader nodes on the cprop list, as
1243 * opcode splitting is done by split_by() at the end of
1244 * constant propagation./
1247 if (step(&senv[0])) {
1251 if (step(&senv[1])) {
1256 assert(senv[winner].initial == NULL);
1257 assert(senv[winner].unwalked == NULL);
1259 /* clear flags from walked/unwalked */
1261 transitions = clear_flags(senv[0].unwalked) << shf;
1262 transitions |= clear_flags(senv[0].walked) << shf;
1264 transitions |= clear_flags(senv[1].unwalked) << shf;
1265 transitions |= clear_flags(senv[1].walked) << shf;
1267 dump_race_list("winner ", senv[winner].walked);
1269 /* Move walked_{winner} to a new partition, X'. */
1270 X_prime = new_partition(env);
1273 for (node = senv[winner].walked; node != NULL; node = node->race_next) {
1274 list_del(&node->node_list);
1275 node->part = X_prime;
1276 if (node->is_follower) {
1277 list_add_tail(&node->node_list, &X_prime->Follower);
1279 list_add_tail(&node->node_list, &X_prime->Leader);
1282 if (node->max_user_input > max_input)
1283 max_input = node->max_user_input;
1285 X_prime->n_leader = n;
1286 X_prime->max_user_inputs = max_input;
1287 X->n_leader -= X_prime->n_leader;
1289 /* for now, copy the type info tag, it will be adjusted in split_by(). */
1290 X_prime->type_is_T_or_C = X->type_is_T_or_C;
1293 * Even if a follower was not checked by both sides, it might have
1294 * loose its congruence, so we need to check this case for all follower.
1296 list_for_each_entry_safe(node_t, node, t, &X_prime->Follower, node_list) {
1297 if (identity(node) == node) {
1298 follower_to_leader(node);
1304 check_partition(X_prime);
1306 /* X' is the smaller part */
1307 add_to_worklist(X_prime, env);
1310 * If there where follower to leader transitions, ensure that the nodes
1311 * can be split out if necessary.
1313 if (transitions & 1) {
1314 /* place winner partition on the cprop list */
1315 if (X_prime->on_cprop == 0) {
1316 X_prime->cprop_next = env->cprop;
1317 env->cprop = X_prime;
1318 X_prime->on_cprop = 1;
1321 if (transitions & 2) {
1322 /* place other partition on the cprop list */
1323 if (X->on_cprop == 0) {
1324 X->cprop_next = env->cprop;
1330 dump_partition("Now ", X);
1331 dump_partition("Created new ", X_prime);
1333 /* we have to ensure that the partition containing g is returned */
1343 * Returns non-zero if the i'th input of a Phi node is live.
1345 * @param phi a Phi-node
1346 * @param i an input number
1348 * @return non-zero if the i'th input of the given Phi node is live
1350 static int is_live_input(ir_node *phi, int i)
1353 ir_node *block = get_nodes_block(phi);
1354 ir_node *pred = get_Block_cfgpred(block, i);
1355 lattice_elem_t type = get_node_type(pred);
1357 return type.tv != tarval_unreachable;
1359 /* else it's the control input, always live */
1361 } /* is_live_input */
1364 * Return non-zero if a type is a constant.
1366 static int is_constant_type(lattice_elem_t type)
1368 if (type.tv != tarval_bottom && type.tv != tarval_top)
1371 } /* is_constant_type */
1374 * Check whether a type is neither Top or a constant.
1375 * Note: U is handled like Top here, R is a constant.
1377 * @param type the type to check
1379 static int type_is_neither_top_nor_const(const lattice_elem_t type)
1381 if (is_tarval(type.tv)) {
1382 if (type.tv == tarval_top)
1384 if (tarval_is_constant(type.tv))
1391 } /* type_is_neither_top_nor_const */
1394 * Collect nodes to the touched list.
1396 * @param list the list which contains the nodes that must be evaluated
1397 * @param idx the index of the def_use edge to evaluate
1398 * @param env the environment
1400 static void collect_touched(list_head *list, int idx, environment_t *env)
1403 int end_idx = env->end_idx;
1405 list_for_each_entry(node_t, x, list, node_list) {
1409 /* leader edges start AFTER follower edges */
1410 x->next_edge = x->n_followers + 1;
1412 num_edges = get_irn_n_outs(x->node);
1414 /* for all edges in x.L.def_use_{idx} */
1415 while (x->next_edge <= num_edges) {
1416 const ir_def_use_edge *edge = &x->node->out[x->next_edge];
1419 /* check if we have necessary edges */
1420 if (edge->pos > idx)
1427 /* only non-commutative nodes */
1428 if (env->commutative &&
1429 (idx == 0 || idx == 1) && is_op_commutative(get_irn_op(succ)))
1432 /* ignore the "control input" for non-pinned nodes
1433 if we are running in GCSE mode */
1434 if (idx < end_idx && get_irn_pinned(succ) != op_pin_state_pinned)
1437 y = get_irn_node(succ);
1438 assert(get_irn_n(succ, idx) == x->node);
1440 /* ignore block edges touching followers */
1441 if (idx == -1 && y->is_follower)
1444 if (is_constant_type(y->type)) {
1445 unsigned code = get_irn_opcode(succ);
1446 if (code == iro_Sub || code == iro_Cmp)
1447 add_to_cprop(y, env);
1450 /* Partitions of constants should not be split simply because their Nodes have unequal
1451 functions or incongruent inputs. */
1452 if (type_is_neither_top_nor_const(y->type) &&
1453 (! is_Phi(y->node) || is_live_input(y->node, idx))) {
1454 add_to_touched(y, env);
1458 } /* collect_touched */
1461 * Collect commutative nodes to the touched list.
1463 * @param list the list which contains the nodes that must be evaluated
1464 * @param env the environment
1466 static void collect_commutative_touched(list_head *list, environment_t *env)
1470 list_for_each_entry(node_t, x, list, node_list) {
1473 num_edges = get_irn_n_outs(x->node);
1475 x->next_edge = x->n_followers + 1;
1477 /* for all edges in x.L.def_use_{idx} */
1478 while (x->next_edge <= num_edges) {
1479 const ir_def_use_edge *edge = &x->node->out[x->next_edge];
1482 /* check if we have necessary edges */
1492 /* only commutative nodes */
1493 if (!is_op_commutative(get_irn_op(succ)))
1496 y = get_irn_node(succ);
1497 if (is_constant_type(y->type)) {
1498 unsigned code = get_irn_opcode(succ);
1499 if (code == iro_Eor)
1500 add_to_cprop(y, env);
1503 /* Partitions of constants should not be split simply because their Nodes have unequal
1504 functions or incongruent inputs. */
1505 if (type_is_neither_top_nor_const(y->type)) {
1506 add_to_touched(y, env);
1510 } /* collect_commutative_touched */
1513 * Split the partitions if caused by the first entry on the worklist.
1515 * @param env the environment
1517 static void cause_splits(environment_t *env)
1519 partition_t *X, *Z, *N;
1522 /* remove the first partition from the worklist */
1524 env->worklist = X->wl_next;
1527 dump_partition("Cause_split: ", X);
1529 if (env->commutative) {
1530 /* handle commutative nodes first */
1532 /* empty the touched set: already done, just clear the list */
1533 env->touched = NULL;
1535 collect_commutative_touched(&X->Leader, env);
1536 collect_commutative_touched(&X->Follower, env);
1538 for (Z = env->touched; Z != NULL; Z = N) {
1540 node_t *touched = Z->touched;
1541 node_t *touched_aa = NULL;
1542 node_t *touched_ab = NULL;
1543 unsigned n_touched_aa = 0;
1544 unsigned n_touched_ab = 0;
1546 assert(Z->touched != NULL);
1548 /* beware, split might change Z */
1549 N = Z->touched_next;
1551 /* remove it from the touched set */
1554 /* Empty local Z.touched. */
1555 for (e = touched; e != NULL; e = n) {
1556 node_t *left = get_irn_node(get_irn_n(e->node, 0));
1557 node_t *right = get_irn_node(get_irn_n(e->node, 1));
1559 assert(e->is_follower == 0);
1564 * Note: op(a, a) is NOT congruent to op(a, b).
1565 * So, we must split the touched list.
1567 if (left->part == right->part) {
1568 e->next = touched_aa;
1572 e->next = touched_ab;
1577 assert(n_touched_aa + n_touched_ab == Z->n_touched);
1581 if (0 < n_touched_aa && n_touched_aa < Z->n_leader) {
1582 partition_t *Z_prime = Z;
1583 DB((dbg, LEVEL_2, "Split part%d by touched_aa\n", Z_prime->nr));
1584 split(&Z_prime, touched_aa, env);
1586 assert(n_touched_aa <= Z->n_leader);
1588 if (0 < n_touched_ab && n_touched_ab < Z->n_leader) {
1589 partition_t *Z_prime = Z;
1590 DB((dbg, LEVEL_2, "Split part%d by touched_ab\n", Z_prime->nr));
1591 split(&Z_prime, touched_ab, env);
1593 assert(n_touched_ab <= Z->n_leader);
1597 /* combine temporary leader and follower list */
1598 for (idx = -1; idx <= X->max_user_inputs; ++idx) {
1599 /* empty the touched set: already done, just clear the list */
1600 env->touched = NULL;
1602 collect_touched(&X->Leader, idx, env);
1603 collect_touched(&X->Follower, idx, env);
1605 for (Z = env->touched; Z != NULL; Z = N) {
1607 node_t *touched = Z->touched;
1608 unsigned n_touched = Z->n_touched;
1610 assert(Z->touched != NULL);
1612 /* beware, split might change Z */
1613 N = Z->touched_next;
1615 /* remove it from the touched set */
1618 /* Empty local Z.touched. */
1619 for (e = touched; e != NULL; e = e->next) {
1620 assert(e->is_follower == 0);
1626 if (0 < n_touched && n_touched < Z->n_leader) {
1627 DB((dbg, LEVEL_2, "Split part%d by touched\n", Z->nr));
1628 split(&Z, touched, env);
1630 assert(n_touched <= Z->n_leader);
1633 } /* cause_splits */
1636 * Implements split_by_what(): Split a partition by characteristics given
1637 * by the what function.
1639 * @param X the partition to split
1640 * @param What a function returning an Id for every node of the partition X
1641 * @param P a list to store the result partitions
1642 * @param env the environment
1646 static partition_t *split_by_what(partition_t *X, what_func What,
1647 partition_t **P, environment_t *env)
1651 listmap_entry_t *iter;
1654 /* Let map be an empty mapping from the range of What to (local) list of Nodes. */
1656 list_for_each_entry(node_t, x, &X->Leader, node_list) {
1657 void *id = What(x, env);
1658 listmap_entry_t *entry;
1661 /* input not allowed, ignore */
1664 /* Add x to map[What(x)]. */
1665 entry = listmap_find(&map, id);
1666 x->next = entry->list;
1669 /* Let P be a set of Partitions. */
1671 /* for all sets S except one in the range of map do */
1672 for (iter = map.values; iter != NULL; iter = iter->next) {
1673 if (iter->next == NULL) {
1674 /* this is the last entry, ignore */
1679 /* Add SPLIT( X, S ) to P. */
1680 DB((dbg, LEVEL_2, "Split part%d by WHAT = %s\n", X->nr, what_reason));
1681 R = split(&X, S, env);
1691 } /* split_by_what */
1693 /** lambda n.(n.type) */
1694 static void *lambda_type(const node_t *node, environment_t *env)
1697 return node->type.tv;
1700 /** lambda n.(n.opcode) */
1701 static void *lambda_opcode(const node_t *node, environment_t *env)
1703 opcode_key_t key, *entry;
1705 key.irn = node->node;
1707 entry = (opcode_key_t*)set_insert(env->opcode2id_map, &key, sizeof(key), opcode_hash(&key));
1709 } /* lambda_opcode */
1711 /** lambda n.(n[i].partition) */
1712 static void *lambda_partition(const node_t *node, environment_t *env)
1714 ir_node *skipped = skip_Proj(node->node);
1717 int i = env->lambda_input;
1719 if (i >= get_irn_arity(node->node)) {
1721 * We are outside the allowed range: This can happen even
1722 * if we have split by opcode first: doing so might move Followers
1723 * to Leaders and those will have a different opcode!
1724 * Note that in this case the partition is on the cprop list and will be
1730 /* ignore the "control input" for non-pinned nodes
1731 if we are running in GCSE mode */
1732 if (i < env->end_idx && get_irn_pinned(skipped) != op_pin_state_pinned)
1735 pred = i == -1 ? get_irn_n(skipped, i) : get_irn_n(node->node, i);
1736 p = get_irn_node(pred);
1738 } /* lambda_partition */
1740 /** lambda n.(n[i].partition) for commutative nodes */
1741 static void *lambda_commutative_partition(const node_t *node, environment_t *env)
1743 ir_node *irn = node->node;
1744 ir_node *skipped = skip_Proj(irn);
1745 ir_node *pred, *left, *right;
1747 partition_t *pl, *pr;
1748 int i = env->lambda_input;
1750 if (i >= get_irn_arity(node->node)) {
1752 * We are outside the allowed range: This can happen even
1753 * if we have split by opcode first: doing so might move Followers
1754 * to Leaders and those will have a different opcode!
1755 * Note that in this case the partition is on the cprop list and will be
1761 /* ignore the "control input" for non-pinned nodes
1762 if we are running in GCSE mode */
1763 if (i < env->end_idx && get_irn_pinned(skipped) != op_pin_state_pinned)
1767 pred = get_irn_n(skipped, i);
1768 p = get_irn_node(pred);
1772 if (is_op_commutative(get_irn_op(irn))) {
1773 /* normalize partition order by returning the "smaller" on input 0,
1774 the "bigger" on input 1. */
1775 left = get_binop_left(irn);
1776 pl = get_irn_node(left)->part;
1777 right = get_binop_right(irn);
1778 pr = get_irn_node(right)->part;
1781 return pl < pr ? pl : pr;
1783 return pl > pr ? pl : pr;
1785 /* a not split out Follower */
1786 pred = get_irn_n(irn, i);
1787 p = get_irn_node(pred);
1791 } /* lambda_commutative_partition */
1794 * Returns true if a type is a constant (and NOT Top
1797 static int is_con(const lattice_elem_t type)
1799 /* be conservative */
1800 if (is_tarval(type.tv))
1801 return tarval_is_constant(type.tv);
1802 return is_entity(type.sym.entity_p);
1806 * Implements split_by().
1808 * @param X the partition to split
1809 * @param env the environment
1811 static void split_by(partition_t *X, environment_t *env)
1813 partition_t *I, *P = NULL;
1816 dump_partition("split_by", X);
1818 if (X->n_leader == 1) {
1819 /* we have only one leader, no need to split, just check its type */
1820 node_t *x = get_first_node(X);
1821 X->type_is_T_or_C = x->type.tv == tarval_top || is_con(x->type);
1825 DEBUG_ONLY(what_reason = "lambda n.(n.type)";)
1826 P = split_by_what(X, lambda_type, &P, env);
1829 /* adjust the type tags, we have split partitions by type */
1830 for (I = P; I != NULL; I = I->split_next) {
1831 node_t *x = get_first_node(I);
1832 I->type_is_T_or_C = x->type.tv == tarval_top || is_con(x->type);
1839 if (Y->n_leader > 1) {
1840 /* we do not want split the TOP or constant partitions */
1841 if (! Y->type_is_T_or_C) {
1842 partition_t *Q = NULL;
1844 DEBUG_ONLY(what_reason = "lambda n.(n.opcode)";)
1845 Q = split_by_what(Y, lambda_opcode, &Q, env);
1852 if (Z->n_leader > 1) {
1853 const node_t *first = get_first_node(Z);
1854 int arity = get_irn_arity(first->node);
1856 what_func what = lambda_partition;
1857 DEBUG_ONLY(char buf[64];)
1859 if (env->commutative && is_op_commutative(get_irn_op(first->node)))
1860 what = lambda_commutative_partition;
1863 * BEWARE: during splitting by input 2 for instance we might
1864 * create new partitions which are different by input 1, so collect
1865 * them and split further.
1867 Z->split_next = NULL;
1870 for (input = arity - 1; input >= -1; --input) {
1872 partition_t *Z_prime = R;
1875 if (Z_prime->n_leader > 1) {
1876 env->lambda_input = input;
1877 DEBUG_ONLY(snprintf(buf, sizeof(buf), "lambda n.(n[%d].partition)", input);)
1878 DEBUG_ONLY(what_reason = buf;)
1879 S = split_by_what(Z_prime, what, &S, env);
1882 Z_prime->split_next = S;
1885 } while (R != NULL);
1890 } while (Q != NULL);
1893 } while (P != NULL);
1897 * (Re-)compute the type for a given node.
1899 * @param node the node
1901 static void default_compute(node_t *node)
1904 ir_node *irn = node->node;
1906 /* if any of the data inputs have type top, the result is type top */
1907 for (i = get_irn_arity(irn) - 1; i >= 0; --i) {
1908 ir_node *pred = get_irn_n(irn, i);
1909 node_t *p = get_irn_node(pred);
1911 if (p->type.tv == tarval_top) {
1912 node->type.tv = tarval_top;
1917 if (get_irn_mode(node->node) == mode_X)
1918 node->type.tv = tarval_reachable;
1920 node->type.tv = computed_value(irn);
1921 } /* default_compute */
1924 * (Re-)compute the type for a Block node.
1926 * @param node the node
1928 static void compute_Block(node_t *node)
1931 ir_node *block = node->node;
1933 if (block == get_irg_start_block(current_ir_graph) || has_Block_entity(block)) {
1934 /* start block and labelled blocks are always reachable */
1935 node->type.tv = tarval_reachable;
1939 for (i = get_Block_n_cfgpreds(block) - 1; i >= 0; --i) {
1940 node_t *pred = get_irn_node(get_Block_cfgpred(block, i));
1942 if (pred->type.tv == tarval_reachable) {
1943 /* A block is reachable, if at least of predecessor is reachable. */
1944 node->type.tv = tarval_reachable;
1948 node->type.tv = tarval_top;
1949 } /* compute_Block */
1952 * (Re-)compute the type for a Bad node.
1954 * @param node the node
1956 static void compute_Bad(node_t *node)
1958 /* Bad nodes ALWAYS compute Top */
1959 node->type.tv = tarval_top;
1963 * (Re-)compute the type for an Unknown node.
1965 * @param node the node
1967 static void compute_Unknown(node_t *node)
1969 /* While Unknown nodes should compute Top this is dangerous:
1970 * a Top input to a Cond would lead to BOTH control flows unreachable.
1971 * While this is correct in the given semantics, it would destroy the Firm
1974 * It would be safe to compute Top IF it can be assured, that only Cmp
1975 * nodes are inputs to Conds. We check that first.
1976 * This is the way Frontends typically build Firm, but some optimizations
1977 * (jump threading for instance) might replace them by Phib's...
1979 node->type.tv = tarval_UNKNOWN;
1980 } /* compute_Unknown */
1983 * (Re-)compute the type for a Jmp node.
1985 * @param node the node
1987 static void compute_Jmp(node_t *node)
1989 node_t *block = get_irn_node(get_nodes_block(node->node));
1991 node->type = block->type;
1995 * (Re-)compute the type for the Return node.
1997 * @param node the node
1999 static void compute_Return(node_t *node)
2001 /* The Return node is NOT dead if it is in a reachable block.
2002 * This is already checked in compute(). so we can return
2003 * Reachable here. */
2004 node->type.tv = tarval_reachable;
2005 } /* compute_Return */
2008 * (Re-)compute the type for the End node.
2010 * @param node the node
2012 static void compute_End(node_t *node)
2014 /* the End node is NOT dead of course */
2015 node->type.tv = tarval_reachable;
2019 * (Re-)compute the type for a Call.
2021 * @param node the node
2023 static void compute_Call(node_t *node)
2026 * A Call computes always bottom, even if it has Unknown
2029 node->type.tv = tarval_bottom;
2030 } /* compute_Call */
2033 * (Re-)compute the type for a SymConst node.
2035 * @param node the node
2037 static void compute_SymConst(node_t *node)
2039 ir_node *irn = node->node;
2040 node_t *block = get_irn_node(get_nodes_block(irn));
2042 if (block->type.tv == tarval_unreachable) {
2043 node->type.tv = tarval_top;
2046 switch (get_SymConst_kind(irn)) {
2047 case symconst_addr_ent:
2048 node->type.sym = get_SymConst_symbol(irn);
2051 node->type.tv = computed_value(irn);
2053 } /* compute_SymConst */
2056 * (Re-)compute the type for a Phi node.
2058 * @param node the node
2060 static void compute_Phi(node_t *node)
2063 ir_node *phi = node->node;
2064 lattice_elem_t type;
2066 /* if a Phi is in a unreachable block, its type is TOP */
2067 node_t *block = get_irn_node(get_nodes_block(phi));
2069 if (block->type.tv == tarval_unreachable) {
2070 node->type.tv = tarval_top;
2074 /* Phi implements the Meet operation */
2075 type.tv = tarval_top;
2076 for (i = get_Phi_n_preds(phi) - 1; i >= 0; --i) {
2077 node_t *pred = get_irn_node(get_Phi_pred(phi, i));
2078 node_t *pred_X = get_irn_node(get_Block_cfgpred(block->node, i));
2080 if (pred_X->type.tv == tarval_unreachable || pred->type.tv == tarval_top) {
2081 /* ignore TOP inputs: We must check here for unreachable blocks,
2082 because Firm constants live in the Start Block are NEVER Top.
2083 Else, a Phi (1,2) will produce Bottom, even if the 2 for instance
2084 comes from a unreachable input. */
2087 if (pred->type.tv == tarval_bottom) {
2088 node->type.tv = tarval_bottom;
2090 } else if (type.tv == tarval_top) {
2091 /* first constant found */
2093 } else if (type.tv != pred->type.tv) {
2094 /* different constants or tarval_bottom */
2095 node->type.tv = tarval_bottom;
2098 /* else nothing, constants are the same */
2104 * (Re-)compute the type for an Add. Special case: one nodes is a Zero Const.
2106 * @param node the node
2108 static void compute_Add(node_t *node)
2110 ir_node *sub = node->node;
2111 node_t *l = get_irn_node(get_Add_left(sub));
2112 node_t *r = get_irn_node(get_Add_right(sub));
2113 lattice_elem_t a = l->type;
2114 lattice_elem_t b = r->type;
2117 if (a.tv == tarval_top || b.tv == tarval_top) {
2118 node->type.tv = tarval_top;
2119 } else if (a.tv == tarval_bottom || b.tv == tarval_bottom) {
2120 node->type.tv = tarval_bottom;
2122 /* x + 0 = 0 + x = x, but beware of floating point +0 + -0, so we
2123 must call tarval_add() first to handle this case! */
2124 if (is_tarval(a.tv)) {
2125 if (is_tarval(b.tv)) {
2126 node->type.tv = tarval_add(a.tv, b.tv);
2129 mode = get_tarval_mode(a.tv);
2130 if (a.tv == get_mode_null(mode)) {
2134 } else if (is_tarval(b.tv)) {
2135 mode = get_tarval_mode(b.tv);
2136 if (b.tv == get_mode_null(mode)) {
2141 node->type.tv = tarval_bottom;
2146 * (Re-)compute the type for a Sub. Special case: both nodes are congruent.
2148 * @param node the node
2150 static void compute_Sub(node_t *node)
2152 ir_node *sub = node->node;
2153 node_t *l = get_irn_node(get_Sub_left(sub));
2154 node_t *r = get_irn_node(get_Sub_right(sub));
2155 lattice_elem_t a = l->type;
2156 lattice_elem_t b = r->type;
2159 if (a.tv == tarval_top || b.tv == tarval_top) {
2160 node->type.tv = tarval_top;
2161 } else if (is_con(a) && is_con(b)) {
2162 if (is_tarval(a.tv) && is_tarval(b.tv)) {
2163 node->type.tv = tarval_sub(a.tv, b.tv, get_irn_mode(sub));
2164 } else if (is_tarval(a.tv) && tarval_is_null(a.tv)) {
2166 } else if (is_tarval(b.tv) && tarval_is_null(b.tv)) {
2169 node->type.tv = tarval_bottom;
2171 } else if (r->part == l->part &&
2172 (!mode_is_float(get_irn_mode(l->node)))) {
2174 * BEWARE: a - a is NOT always 0 for floating Point values, as
2175 * NaN op NaN = NaN, so we must check this here.
2177 ir_mode *mode = get_irn_mode(sub);
2178 tv = get_mode_null(mode);
2180 /* if the node was ONCE evaluated by all constants, but now
2181 this breaks AND we get from the argument partitions a different
2182 result, switch to bottom.
2183 This happens because initially all nodes are in the same partition ... */
2184 if (node->type.tv != tv)
2188 node->type.tv = tarval_bottom;
2193 * (Re-)compute the type for an Eor. Special case: both nodes are congruent.
2195 * @param node the node
2197 static void compute_Eor(node_t *node)
2199 ir_node *eor = node->node;
2200 node_t *l = get_irn_node(get_Eor_left(eor));
2201 node_t *r = get_irn_node(get_Eor_right(eor));
2202 lattice_elem_t a = l->type;
2203 lattice_elem_t b = r->type;
2206 if (a.tv == tarval_top || b.tv == tarval_top) {
2207 node->type.tv = tarval_top;
2208 } else if (is_con(a) && is_con(b)) {
2209 if (is_tarval(a.tv) && is_tarval(b.tv)) {
2210 node->type.tv = tarval_eor(a.tv, b.tv);
2211 } else if (is_tarval(a.tv) && tarval_is_null(a.tv)) {
2213 } else if (is_tarval(b.tv) && tarval_is_null(b.tv)) {
2216 node->type.tv = tarval_bottom;
2218 } else if (r->part == l->part) {
2219 ir_mode *mode = get_irn_mode(eor);
2220 tv = get_mode_null(mode);
2222 /* if the node was ONCE evaluated by all constants, but now
2223 this breaks AND we get from the argument partitions a different
2224 result, switch to bottom.
2225 This happens because initially all nodes are in the same partition ... */
2226 if (node->type.tv != tv)
2230 node->type.tv = tarval_bottom;
2235 * (Re-)compute the type for Cmp.
2237 * @param node the node
2239 static void compute_Cmp(node_t *node)
2241 ir_node *cmp = node->node;
2242 node_t *l = get_irn_node(get_Cmp_left(cmp));
2243 node_t *r = get_irn_node(get_Cmp_right(cmp));
2244 lattice_elem_t a = l->type;
2245 lattice_elem_t b = r->type;
2246 ir_relation relation = get_Cmp_relation(cmp);
2249 if (a.tv == tarval_top || b.tv == tarval_top) {
2250 node->type.tv = tarval_undefined;
2251 } else if (is_con(a) && is_con(b)) {
2252 default_compute(node);
2255 * BEWARE: a == a is NOT always True for floating Point values, as
2256 * NaN != NaN is defined, so we must check this here.
2257 * (while for some pnc we could still optimize we have to stay
2258 * consistent with compute_Cmp, so don't do anything for floats)
2260 } else if (r->part == l->part && !mode_is_float(get_irn_mode(l->node))) {
2261 tv = relation & ir_relation_equal ? tarval_b_true : tarval_b_false;
2263 /* if the node was ONCE evaluated to a constant, but now
2264 this breaks AND we get from the argument partitions a different
2265 result, ensure monotony by fall to bottom.
2266 This happens because initially all nodes are in the same partition ... */
2267 if (node->type.tv == tarval_bottom)
2269 else if (node->type.tv != tv && is_constant_type(node->type))
2273 node->type.tv = tarval_bottom;
2278 * (Re-)compute the type for a Proj(Cond).
2280 * @param node the node
2281 * @param cond the predecessor Cond node
2283 static void compute_Proj_Cond(node_t *node, ir_node *cond)
2285 ir_node *proj = node->node;
2286 long pnc = get_Proj_proj(proj);
2287 ir_node *sel = get_Cond_selector(cond);
2288 node_t *selector = get_irn_node(sel);
2291 * Note: it is crucial for the monotony that the Proj(Cond)
2292 * are evaluates after all predecessors of the Cond selector are
2298 * Due to the fact that 0 is a const, the Cmp gets immediately
2299 * on the cprop list. It will be evaluated before x is evaluated,
2300 * might leaving x as Top. When later x is evaluated, the Cmp
2301 * might change its value.
2302 * BUT if the Cond is evaluated before this happens, Proj(Cond, FALSE)
2303 * gets R, and later changed to F if Cmp is evaluated to True!
2305 * We prevent this by putting Conds in an extra cprop_X queue, which
2306 * gets evaluated after the cprop queue is empty.
2308 * Note that this even happens with Click's original algorithm, if
2309 * Cmp(x, 0) is evaluated to True first and later changed to False
2310 * if x was Top first and later changed to a Const ...
2311 * It is unclear how Click solved that problem ...
2313 * However, in rare cases even this does not help, if a Top reaches
2314 * a compare through a Phi, than Proj(Cond) is evaluated changing
2315 * the type of the Phi to something other.
2316 * So, we take the last resort and bind the type to R once
2319 * (This might be even the way Click works around the whole problem).
2321 * Finally, we may miss some optimization possibilities due to this:
2326 * If Top reaches the if first, than we decide for != here.
2327 * If y later is evaluated to 0, we cannot revert this decision
2328 * and must live with both outputs enabled. If this happens,
2329 * we get an unresolved if (true) in the code ...
2331 * In Click's version where this decision is done at the Cmp,
2332 * the Cmp is NOT optimized away than (if y evaluated to 1
2333 * for instance) and we get a if (1 == 0) here ...
2335 * Both solutions are suboptimal.
2336 * At least, we could easily detect this problem and run
2337 * cf_opt() (or even combo) again :-(
2339 if (node->type.tv == tarval_reachable)
2342 if (pnc == pn_Cond_true) {
2343 if (selector->type.tv == tarval_b_false) {
2344 node->type.tv = tarval_unreachable;
2345 } else if (selector->type.tv == tarval_b_true) {
2346 node->type.tv = tarval_reachable;
2347 } else if (selector->type.tv == tarval_bottom) {
2348 node->type.tv = tarval_reachable;
2350 assert(selector->type.tv == tarval_top);
2351 if (tarval_UNKNOWN == tarval_top) {
2352 /* any condition based on Top is "!=" */
2353 node->type.tv = tarval_unreachable;
2355 node->type.tv = tarval_unreachable;
2359 assert(pnc == pn_Cond_false);
2361 if (selector->type.tv == tarval_b_false) {
2362 node->type.tv = tarval_reachable;
2363 } else if (selector->type.tv == tarval_b_true) {
2364 node->type.tv = tarval_unreachable;
2365 } else if (selector->type.tv == tarval_bottom) {
2366 node->type.tv = tarval_reachable;
2368 assert(selector->type.tv == tarval_top);
2369 if (tarval_UNKNOWN == tarval_top) {
2370 /* any condition based on Top is "!=" */
2371 node->type.tv = tarval_reachable;
2373 node->type.tv = tarval_unreachable;
2377 } /* compute_Proj_Cond */
2379 static void compute_Proj_Switch(node_t *node, ir_node *switchn)
2381 ir_node *proj = node->node;
2382 long pnc = get_Proj_proj(proj);
2383 ir_node *sel = get_Switch_selector(switchn);
2384 node_t *selector = get_irn_node(sel);
2386 /* see long comment in compute_Proj_Cond */
2387 if (node->type.tv == tarval_reachable)
2390 if (selector->type.tv == tarval_bottom) {
2391 node->type.tv = tarval_reachable;
2392 } else if (selector->type.tv == tarval_top) {
2393 if (tarval_UNKNOWN == tarval_top && pnc == pn_Switch_default) {
2394 /* a switch based of Top is always "default" */
2395 node->type.tv = tarval_reachable;
2397 node->type.tv = tarval_unreachable;
2400 long value = get_tarval_long(selector->type.tv);
2401 const ir_switch_table *table = get_Switch_table(switchn);
2402 size_t n_entries = ir_switch_table_get_n_entries(table);
2405 for (e = 0; e < n_entries; ++e) {
2406 const ir_switch_table_entry *entry
2407 = ir_switch_table_get_entry_const(table, e);
2408 ir_tarval *min = entry->min;
2409 ir_tarval *max = entry->max;
2411 if (selector->type.tv == min) {
2412 node->type.tv = entry->pn == pnc
2413 ? tarval_reachable : tarval_unreachable;
2417 long minval = get_tarval_long(min);
2418 long maxval = get_tarval_long(max);
2419 if (minval <= value && value <= maxval) {
2420 node->type.tv = entry->pn == pnc
2421 ? tarval_reachable : tarval_unreachable;
2427 /* no entry matched: default */
2429 = pnc == pn_Switch_default ? tarval_reachable : tarval_unreachable;
2434 * (Re-)compute the type for a Proj-Node.
2436 * @param node the node
2438 static void compute_Proj(node_t *node)
2440 ir_node *proj = node->node;
2441 ir_mode *mode = get_irn_mode(proj);
2442 node_t *block = get_irn_node(get_nodes_block(skip_Proj(proj)));
2443 ir_node *pred = get_Proj_pred(proj);
2445 if (block->type.tv == tarval_unreachable) {
2446 /* a Proj in a unreachable Block stay Top */
2447 node->type.tv = tarval_top;
2450 if (get_irn_node(pred)->type.tv == tarval_top && !is_Cond(pred) && !is_Switch(pred)) {
2451 /* if the predecessor is Top, its Proj follow */
2452 node->type.tv = tarval_top;
2456 if (mode == mode_M) {
2457 /* mode M is always bottom */
2458 node->type.tv = tarval_bottom;
2460 } else if (mode == mode_X) {
2461 /* handle mode_X nodes */
2462 switch (get_irn_opcode(pred)) {
2464 /* the Proj_X from the Start is always reachable.
2465 However this is already handled at the top. */
2466 node->type.tv = tarval_reachable;
2469 compute_Proj_Cond(node, pred);
2472 compute_Proj_Switch(node, pred);
2479 default_compute(node);
2480 } /* compute_Proj */
2483 * (Re-)compute the type for a Confirm.
2485 * @param node the node
2487 static void compute_Confirm(node_t *node)
2489 ir_node *confirm = node->node;
2490 node_t *pred = get_irn_node(get_Confirm_value(confirm));
2492 if (get_Confirm_relation(confirm) == ir_relation_equal) {
2493 node_t *bound = get_irn_node(get_Confirm_bound(confirm));
2495 if (is_con(bound->type)) {
2496 /* is equal to a constant */
2497 node->type = bound->type;
2501 /* a Confirm is a copy OR a Const */
2502 node->type = pred->type;
2503 } /* compute_Confirm */
2506 * (Re-)compute the type for a given node.
2508 * @param node the node
2510 static void compute(node_t *node)
2512 ir_node *irn = node->node;
2515 #ifndef VERIFY_MONOTONE
2517 * Once a node reaches bottom, the type cannot fall further
2518 * in the lattice and we can stop computation.
2519 * Do not take this exit if the monotony verifier is
2520 * enabled to catch errors.
2522 if (node->type.tv == tarval_bottom)
2526 if (!is_Block(irn)) {
2527 /* for pinned nodes, check its control input */
2528 if (get_irn_pinned(skip_Proj(irn)) == op_pin_state_pinned) {
2529 node_t *block = get_irn_node(get_nodes_block(irn));
2531 if (block->type.tv == tarval_unreachable) {
2532 node->type.tv = tarval_top;
2538 func = (compute_func)node->node->op->ops.generic;
2544 * Identity functions: Note that one might think that identity() is just a
2545 * synonym for equivalent_node(). While this is true, we cannot use it for the algorithm
2546 * here, because it expects that the identity node is one of the inputs, which is NOT
2547 * always true for equivalent_node() which can handle (and does sometimes) DAGs.
2548 * So, we have our own implementation, which copies some parts of equivalent_node()
2552 * Calculates the Identity for Phi nodes
2554 static node_t *identity_Phi(node_t *node)
2556 ir_node *phi = node->node;
2557 ir_node *block = get_nodes_block(phi);
2558 node_t *n_part = NULL;
2561 for (i = get_Phi_n_preds(phi) - 1; i >= 0; --i) {
2562 node_t *pred_X = get_irn_node(get_Block_cfgpred(block, i));
2564 if (pred_X->type.tv == tarval_reachable) {
2565 node_t *pred = get_irn_node(get_Phi_pred(phi, i));
2569 else if (n_part->part != pred->part) {
2570 /* incongruent inputs, not a follower */
2575 /* if n_part is NULL here, all inputs path are dead, the Phi computes
2576 * tarval_top, is in the TOP partition and should NOT being split! */
2577 assert(n_part != NULL);
2579 } /* identity_Phi */
2582 * Calculates the Identity for commutative 0 neutral nodes.
2584 static node_t *identity_comm_zero_binop(node_t *node)
2586 ir_node *op = node->node;
2587 node_t *a = get_irn_node(get_binop_left(op));
2588 node_t *b = get_irn_node(get_binop_right(op));
2589 ir_mode *mode = get_irn_mode(op);
2592 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
2593 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
2596 /* node: no input should be tarval_top, else the binop would be also
2597 * Top and not being split. */
2598 zero = get_mode_null(mode);
2599 if (a->type.tv == zero)
2601 if (b->type.tv == zero)
2604 } /* identity_comm_zero_binop */
2607 * Calculates the Identity for Shift nodes.
2609 static node_t *identity_shift(node_t *node)
2611 ir_node *op = node->node;
2612 node_t *b = get_irn_node(get_binop_right(op));
2613 ir_mode *mode = get_irn_mode(b->node);
2616 /* node: no input should be tarval_top, else the binop would be also
2617 * Top and not being split. */
2618 zero = get_mode_null(mode);
2619 if (b->type.tv == zero)
2620 return get_irn_node(get_binop_left(op));
2622 } /* identity_shift */
2625 * Calculates the Identity for Mul nodes.
2627 static node_t *identity_Mul(node_t *node)
2629 ir_node *op = node->node;
2630 node_t *a = get_irn_node(get_Mul_left(op));
2631 node_t *b = get_irn_node(get_Mul_right(op));
2632 ir_mode *mode = get_irn_mode(op);
2635 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
2636 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
2639 /* node: no input should be tarval_top, else the binop would be also
2640 * Top and not being split. */
2641 one = get_mode_one(mode);
2642 if (a->type.tv == one)
2644 if (b->type.tv == one)
2647 } /* identity_Mul */
2650 * Calculates the Identity for Sub nodes.
2652 static node_t *identity_Sub(node_t *node)
2654 ir_node *sub = node->node;
2655 node_t *b = get_irn_node(get_Sub_right(sub));
2656 ir_mode *mode = get_irn_mode(sub);
2658 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
2659 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
2662 /* node: no input should be tarval_top, else the binop would be also
2663 * Top and not being split. */
2664 if (b->type.tv == get_mode_null(mode))
2665 return get_irn_node(get_Sub_left(sub));
2667 } /* identity_Sub */
2670 * Calculates the Identity for And nodes.
2672 static node_t *identity_And(node_t *node)
2674 ir_node *andnode = node->node;
2675 node_t *a = get_irn_node(get_And_left(andnode));
2676 node_t *b = get_irn_node(get_And_right(andnode));
2677 ir_tarval *neutral = get_mode_all_one(get_irn_mode(andnode));
2679 /* node: no input should be tarval_top, else the And would be also
2680 * Top and not being split. */
2681 if (a->type.tv == neutral)
2683 if (b->type.tv == neutral)
2686 } /* identity_And */
2689 * Calculates the Identity for Confirm nodes.
2691 static node_t *identity_Confirm(node_t *node)
2693 ir_node *confirm = node->node;
2695 /* a Confirm is always a Copy */
2696 return get_irn_node(get_Confirm_value(confirm));
2697 } /* identity_Confirm */
2700 * Calculates the Identity for Mux nodes.
2702 static node_t *identity_Mux(node_t *node)
2704 ir_node *mux = node->node;
2705 node_t *t = get_irn_node(get_Mux_true(mux));
2706 node_t *f = get_irn_node(get_Mux_false(mux));
2709 if (t->part == f->part)
2712 /* for now, the 1-input identity is not supported */
2714 sel = get_irn_node(get_Mux_sel(mux));
2716 /* Mux sel input is mode_b, so it is always a tarval */
2717 if (sel->type.tv == tarval_b_true)
2719 if (sel->type.tv == tarval_b_false)
2723 } /* identity_Mux */
2726 * Calculates the Identity for nodes.
2728 static node_t *identity(node_t *node)
2730 ir_node *irn = node->node;
2732 switch (get_irn_opcode(irn)) {
2734 return identity_Phi(node);
2736 return identity_Mul(node);
2740 return identity_comm_zero_binop(node);
2745 return identity_shift(node);
2747 return identity_And(node);
2749 return identity_Sub(node);
2751 return identity_Confirm(node);
2753 return identity_Mux(node);
2760 * Node follower is a (new) follower of leader, segregate Leader
2763 static void segregate_def_use_chain_1(const ir_node *follower, node_t *leader)
2765 ir_node *l = leader->node;
2766 int j, i, n = get_irn_n_outs(l);
2768 DB((dbg, LEVEL_2, "%+F is a follower of %+F\n", follower, leader->node));
2769 /* The leader edges must remain sorted, but follower edges can
2771 for (i = leader->n_followers + 1; i <= n; ++i) {
2772 if (l->out[i].use == follower) {
2773 ir_def_use_edge t = l->out[i];
2775 for (j = i - 1; j >= leader->n_followers + 1; --j)
2776 l->out[j + 1] = l->out[j];
2777 ++leader->n_followers;
2778 l->out[leader->n_followers] = t;
2782 } /* segregate_def_use_chain_1 */
2785 * Node follower is a (new) follower segregate its Leader
2788 * @param follower the follower IR node
2790 static void segregate_def_use_chain(const ir_node *follower)
2794 for (i = get_irn_arity(follower) - 1; i >= 0; --i) {
2795 node_t *pred = get_irn_node(get_irn_n(follower, i));
2797 segregate_def_use_chain_1(follower, pred);
2799 } /* segregate_def_use_chain */
2802 * Propagate constant evaluation.
2804 * @param env the environment
2806 static void propagate(environment_t *env)
2810 lattice_elem_t old_type;
2812 unsigned n_fallen, old_type_was_T_or_C;
2815 while (env->cprop != NULL) {
2816 void *oldopcode = NULL;
2818 /* remove the first partition X from cprop */
2821 env->cprop = X->cprop_next;
2823 old_type_was_T_or_C = X->type_is_T_or_C;
2825 DB((dbg, LEVEL_2, "Propagate type on part%d\n", X->nr));
2829 int cprop_empty = list_empty(&X->cprop);
2830 int cprop_X_empty = list_empty(&X->cprop_X);
2832 if (cprop_empty && cprop_X_empty) {
2833 /* both cprop lists are empty */
2837 /* remove the first Node x from X.cprop */
2839 /* Get a node from the cprop_X list only if
2840 * all data nodes are processed.
2841 * This ensures, that all inputs of the Cond
2842 * predecessor are processed if its type is still Top.
2844 x = list_entry(X->cprop_X.next, node_t, cprop_list);
2846 x = list_entry(X->cprop.next, node_t, cprop_list);
2849 //assert(x->part == X);
2850 list_del(&x->cprop_list);
2853 if (x->is_follower && identity(x) == x) {
2854 /* check the opcode first */
2855 if (oldopcode == NULL) {
2856 oldopcode = lambda_opcode(get_first_node(X), env);
2858 if (oldopcode != lambda_opcode(x, env)) {
2859 if (x->on_fallen == 0) {
2860 /* different opcode -> x falls out of this partition */
2865 DB((dbg, LEVEL_2, "Add node %+F to fallen\n", x->node));
2869 /* x will make the follower -> leader transition */
2870 follower_to_leader(x);
2873 /* compute a new type for x */
2875 DB((dbg, LEVEL_3, "computing type of %+F\n", x->node));
2877 if (x->type.tv != old_type.tv) {
2878 DB((dbg, LEVEL_2, "node %+F has changed type from %+F to %+F\n", x->node, old_type, x->type));
2879 verify_type(old_type, x);
2881 if (x->on_fallen == 0) {
2882 /* Add x to fallen. Nodes might fall from T -> const -> _|_, so check that they are
2883 not already on the list. */
2888 DB((dbg, LEVEL_2, "Add node %+F to fallen\n", x->node));
2890 for (i = get_irn_n_outs(x->node) - 1; i >= 0; --i) {
2891 ir_node *succ = get_irn_out(x->node, i);
2892 node_t *y = get_irn_node(succ);
2894 /* Add y to y.partition.cprop. */
2895 add_to_cprop(y, env);
2900 if (n_fallen > 0 && n_fallen != X->n_leader) {
2901 DB((dbg, LEVEL_2, "Splitting part%d by fallen\n", X->nr));
2902 Y = split(&X, fallen, env);
2904 * We have split out fallen node. The type of the result
2905 * partition is NOT set yet.
2907 Y->type_is_T_or_C = 0;
2911 /* remove the flags from the fallen list */
2912 for (x = fallen; x != NULL; x = x->next)
2915 if (old_type_was_T_or_C) {
2918 /* check if some nodes will make the leader -> follower transition */
2919 list_for_each_entry_safe(node_t, y, tmp, &Y->Leader, node_list) {
2920 if (y->type.tv != tarval_top && ! is_con(y->type)) {
2921 node_t *eq_node = identity(y);
2923 if (eq_node != y && eq_node->part == y->part) {
2924 DB((dbg, LEVEL_2, "Node %+F is a follower of %+F\n", y->node, eq_node->node));
2925 /* move to Follower */
2927 list_del(&y->node_list);
2928 list_add_tail(&y->node_list, &Y->Follower);
2931 segregate_def_use_chain(y->node);
2941 * Get the leader for a given node from its congruence class.
2943 * @param irn the node
2945 static ir_node *get_leader(node_t *node)
2947 partition_t *part = node->part;
2949 if (part->n_leader > 1 || node->is_follower) {
2950 if (node->is_follower) {
2951 DB((dbg, LEVEL_2, "Replacing follower %+F\n", node->node));
2954 DB((dbg, LEVEL_2, "Found congruence class for %+F\n", node->node));
2956 return get_first_node(part)->node;
2962 * Returns non-zero if a mode_T node has only one reachable output.
2964 static int only_one_reachable_proj(ir_node *n)
2968 for (i = get_irn_n_outs(n) - 1; i >= 0; --i) {
2969 ir_node *proj = get_irn_out(n, i);
2972 /* skip non-control flow Proj's */
2973 if (get_irn_mode(proj) != mode_X)
2976 node = get_irn_node(proj);
2977 if (node->type.tv == tarval_reachable) {
2983 } /* only_one_reachable_proj */
2986 * Return non-zero if the control flow predecessor node pred
2987 * is the only reachable control flow exit of its block.
2989 * @param pred the control flow exit
2990 * @param block the destination block
2992 static int can_exchange(ir_node *pred, ir_node *block)
2994 if (is_Start(pred) || has_Block_entity(block))
2996 else if (is_Jmp(pred))
2998 else if (is_Raise(pred)) {
2999 /* Raise is a tuple and usually has only one reachable ProjX,
3000 * but it must not be eliminated like a Jmp */
3003 else if (get_irn_mode(pred) == mode_T) {
3004 /* if the predecessor block has more than one
3005 reachable outputs we cannot remove the block */
3006 return only_one_reachable_proj(pred);
3009 } /* can_exchange */
3012 * Block Post-Walker, apply the analysis results on control flow by
3013 * shortening Phi's and Block inputs.
3015 static void apply_cf(ir_node *block, void *ctx)
3017 environment_t *env = (environment_t*)ctx;
3018 node_t *node = get_irn_node(block);
3020 ir_node **ins, **in_X;
3021 ir_node *phi, *next;
3023 n = get_Block_n_cfgpreds(block);
3025 if (node->type.tv == tarval_unreachable) {
3028 for (i = n - 1; i >= 0; --i) {
3029 ir_node *pred = get_Block_cfgpred(block, i);
3031 if (! is_Bad(pred)) {
3032 ir_node *pred_block = get_nodes_block(skip_Proj(pred));
3033 if (!is_Bad(pred_block)) {
3034 node_t *pred_bl = get_irn_node(pred_block);
3036 if (pred_bl->flagged == 0) {
3037 pred_bl->flagged = 3;
3039 if (pred_bl->type.tv == tarval_reachable) {
3041 * We will remove an edge from block to its pred.
3042 * This might leave the pred block as an endless loop
3044 if (! is_backedge(block, i))
3045 keep_alive(pred_bl->node);
3052 if (block == get_irg_end_block(current_ir_graph)) {
3053 /* Analysis found out that the end block is unreachable,
3054 * hence we remove all its control flow predecessors. */
3055 set_irn_in(block, 0, NULL);
3061 /* only one predecessor combine */
3062 ir_node *pred = skip_Proj(get_Block_cfgpred(block, 0));
3064 if (can_exchange(pred, block)) {
3065 ir_node *new_block = get_nodes_block(pred);
3066 DB((dbg, LEVEL_1, "Fuse %+F with %+F\n", block, new_block));
3067 DBG_OPT_COMBO(block, new_block, FS_OPT_COMBO_CF);
3068 exchange(block, new_block);
3069 node->node = new_block;
3075 NEW_ARR_A(ir_node *, in_X, n);
3077 for (i = 0; i < n; ++i) {
3078 ir_node *pred = get_Block_cfgpred(block, i);
3079 node_t *node = get_irn_node(pred);
3081 if (node->type.tv == tarval_reachable) {
3084 DB((dbg, LEVEL_1, "Removing dead input %d from %+F (%+F)\n", i, block, pred));
3085 if (! is_Bad(pred)) {
3086 ir_node *pred_block = get_nodes_block(skip_Proj(pred));
3087 if (!is_Bad(pred_block)) {
3088 node_t *pred_bl = get_irn_node(pred_block);
3090 if (!is_Bad(pred_bl->node) && pred_bl->flagged == 0) {
3091 pred_bl->flagged = 3;
3093 if (pred_bl->type.tv == tarval_reachable) {
3095 * We will remove an edge from block to its pred.
3096 * This might leave the pred block as an endless loop
3098 if (! is_backedge(block, i))
3099 keep_alive(pred_bl->node);
3110 NEW_ARR_A(ir_node *, ins, n);
3111 for (phi = get_Block_phis(block); phi != NULL; phi = next) {
3112 node_t *node = get_irn_node(phi);
3114 next = get_Phi_next(phi);
3115 if (is_tarval(node->type.tv) && tarval_is_constant(node->type.tv)) {
3116 /* this Phi is replaced by a constant */
3117 ir_tarval *tv = node->type.tv;
3118 ir_node *c = new_r_Const(current_ir_graph, tv);
3120 set_irn_node(c, node);
3122 DB((dbg, LEVEL_1, "%+F is replaced by %+F\n", phi, c));
3123 DBG_OPT_COMBO(phi, c, FS_OPT_COMBO_CONST);
3128 for (i = 0; i < n; ++i) {
3129 node_t *pred = get_irn_node(get_Block_cfgpred(block, i));
3131 if (pred->type.tv == tarval_reachable) {
3132 ins[j++] = get_Phi_pred(phi, i);
3136 /* this Phi is replaced by a single predecessor */
3137 ir_node *s = ins[0];
3138 node_t *phi_node = get_irn_node(phi);
3141 DB((dbg, LEVEL_1, "%+F is replaced by %+F because of cf change\n", phi, s));
3142 DBG_OPT_COMBO(phi, s, FS_OPT_COMBO_FOLLOWER);
3147 set_irn_in(phi, j, ins);
3155 /* this Block has only one live predecessor */
3156 ir_node *pred = skip_Proj(in_X[0]);
3158 if (can_exchange(pred, block)) {
3159 ir_node *new_block = get_nodes_block(pred);
3160 DBG_OPT_COMBO(block, new_block, FS_OPT_COMBO_CF);
3161 exchange(block, new_block);
3162 node->node = new_block;
3167 set_irn_in(block, k, in_X);
3172 * Exchange a node by its leader.
3173 * Beware: in rare cases the mode might be wrong here, for instance
3174 * AddP(x, NULL) is a follower of x, but with different mode.
3177 static void exchange_leader(ir_node *irn, ir_node *leader)
3179 ir_mode *mode = get_irn_mode(irn);
3180 if (mode != get_irn_mode(leader)) {
3181 /* The conv is a no-op, so we are free to place it
3182 * either in the block of the leader OR in irn's block.
3183 * Probably placing it into leaders block might reduce
3184 * the number of Conv due to CSE. */
3185 ir_node *block = get_nodes_block(leader);
3186 dbg_info *dbg = get_irn_dbg_info(irn);
3187 ir_node *nlead = new_rd_Conv(dbg, block, leader, mode);
3189 if (nlead != leader) {
3190 /* Note: this newly create irn has no node info because
3191 * it is created after the analysis. However, this node
3192 * replaces the node irn and should not be visited again,
3193 * so set its visited count to the count of irn.
3194 * Otherwise we might visited this node more than once if
3195 * irn had more than one user.
3197 set_irn_node(nlead, NULL);
3198 set_irn_visited(nlead, get_irn_visited(irn));
3202 exchange(irn, leader);
3203 } /* exchange_leader */
3206 * Check, if all users of a mode_M node are dead. Use
3207 * the Def-Use edges for this purpose, as they still
3208 * reflect the situation.
3210 static int all_users_are_dead(const ir_node *irn)
3212 int i, n = get_irn_n_outs(irn);
3214 for (i = 1; i <= n; ++i) {
3215 const ir_node *succ = irn->out[i].use;
3216 const node_t *block = get_irn_node(get_nodes_block(succ));
3219 if (block->type.tv == tarval_unreachable) {
3220 /* block is unreachable */
3223 node = get_irn_node(succ);
3224 if (node->type.tv != tarval_top) {
3225 /* found a reachable user */
3229 /* all users are unreachable */
3231 } /* all_user_are_dead */
3234 * Walker: Find reachable mode_M nodes that have only
3235 * unreachable users. These nodes must be kept later.
3237 static void find_kept_memory(ir_node *irn, void *ctx)
3239 environment_t *env = (environment_t*)ctx;
3240 node_t *node, *block;
3242 if (get_irn_mode(irn) != mode_M)
3245 block = get_irn_node(get_nodes_block(irn));
3246 if (block->type.tv == tarval_unreachable)
3249 node = get_irn_node(irn);
3250 if (node->type.tv == tarval_top)
3253 /* ok, we found a live memory node. */
3254 if (all_users_are_dead(irn)) {
3255 DB((dbg, LEVEL_1, "%+F must be kept\n", irn));
3256 ARR_APP1(ir_node *, env->kept_memory, irn);
3258 } /* find_kept_memory */
3261 * Post-Walker, apply the analysis results;
3263 static void apply_result(ir_node *irn, void *ctx)
3265 environment_t *env = (environment_t*)ctx;
3266 node_t *node = get_irn_node(irn);
3268 if (is_Block(irn) || is_End(irn) || is_Bad(irn)) {
3269 /* blocks already handled, do not touch the End node */
3271 node_t *block = get_irn_node(get_nodes_block(irn));
3273 if (block->type.tv == tarval_unreachable) {
3274 ir_graph *irg = get_irn_irg(irn);
3275 ir_mode *mode = get_irn_mode(node->node);
3276 ir_node *bad = new_r_Bad(irg, mode);
3278 /* here, bad might already have a node, but this can be safely ignored
3279 as long as bad has at least ONE valid node */
3280 set_irn_node(bad, node);
3282 DB((dbg, LEVEL_1, "%+F is unreachable\n", irn));
3285 } else if (node->type.tv == tarval_top) {
3286 ir_mode *mode = get_irn_mode(irn);
3288 if (mode == mode_M) {
3289 /* never kill a mode_M node */
3291 ir_node *pred = get_Proj_pred(irn);
3292 node_t *pnode = get_irn_node(pred);
3294 if (pnode->type.tv == tarval_top) {
3295 /* skip the predecessor */
3296 ir_node *mem = get_memop_mem(pred);
3298 DB((dbg, LEVEL_1, "%+F computes Top, replaced by %+F\n", irn, mem));
3303 /* leave other nodes, especially PhiM */
3304 } else if (mode == mode_T) {
3305 /* Do not kill mode_T nodes, kill their Projs */
3306 } else if (! is_Unknown(irn)) {
3307 /* don't kick away Unknown's, they might be still needed */
3308 ir_node *unk = new_r_Unknown(current_ir_graph, mode);
3310 /* control flow should already be handled at apply_cf() */
3311 assert(mode != mode_X);
3313 /* see comment above */
3314 set_irn_node(unk, node);
3316 DB((dbg, LEVEL_1, "%+F computes Top\n", irn));
3321 else if (get_irn_mode(irn) == mode_X) {
3324 ir_node *cond = get_Proj_pred(irn);
3326 if (is_Cond(cond) || is_Switch(cond)) {
3327 if (only_one_reachable_proj(cond)) {
3328 ir_node *jmp = new_r_Jmp(block->node);
3329 set_irn_node(jmp, node);
3331 DB((dbg, LEVEL_1, "%+F is replaced by %+F\n", irn, jmp));
3332 DBG_OPT_COMBO(irn, jmp, FS_OPT_COMBO_CF);
3336 if (is_Switch(cond)) {
3337 node_t *sel = get_irn_node(get_Switch_selector(cond));
3338 ir_tarval *tv = sel->type.tv;
3340 if (is_tarval(tv) && tarval_is_constant(tv)) {
3341 /* The selector is a constant, but more
3342 * than one output is active: An unoptimized
3351 /* normal data node */
3352 if (is_tarval(node->type.tv) && tarval_is_constant(node->type.tv)) {
3353 ir_tarval *tv = node->type.tv;
3356 * Beware: never replace mode_T nodes by constants. Currently we must mark
3357 * mode_T nodes with constants, but do NOT replace them.
3359 if (! is_Const(irn) && get_irn_mode(irn) != mode_T) {
3360 /* can be replaced by a constant */
3361 ir_node *c = new_r_Const(current_ir_graph, tv);
3362 set_irn_node(c, node);
3364 DB((dbg, LEVEL_1, "%+F is replaced by %+F\n", irn, c));
3365 DBG_OPT_COMBO(irn, c, FS_OPT_COMBO_CONST);
3366 exchange_leader(irn, c);
3369 } else if (is_entity(node->type.sym.entity_p)) {
3370 if (! is_SymConst(irn)) {
3371 /* can be replaced by a SymConst */
3372 ir_node *symc = new_r_SymConst(current_ir_graph, get_irn_mode(irn), node->type.sym, symconst_addr_ent);
3373 set_irn_node(symc, node);
3376 DB((dbg, LEVEL_1, "%+F is replaced by %+F\n", irn, symc));
3377 DBG_OPT_COMBO(irn, symc, FS_OPT_COMBO_CONST);
3378 exchange_leader(irn, symc);
3381 } else if (is_Confirm(irn)) {
3382 /* Confirms are always follower, but do not kill them here */
3384 ir_node *leader = get_leader(node);
3386 if (leader != irn) {
3387 int non_strict_phi = 0;
3390 * Beware: Do not remove Phi(Unknown, ..., x, ..., Unknown)
3391 * as this might create non-strict programs.
3393 if (node->is_follower && is_Phi(irn) && !is_Unknown(leader)) {
3396 for (i = get_Phi_n_preds(irn) - 1; i >= 0; --i) {
3397 ir_node *pred = get_Phi_pred(irn, i);
3399 if (is_Unknown(pred)) {
3405 if (! non_strict_phi) {
3406 DB((dbg, LEVEL_1, "%+F from part%d is replaced by %+F\n", irn, node->part->nr, leader));
3407 if (node->is_follower)
3408 DBG_OPT_COMBO(irn, leader, FS_OPT_COMBO_FOLLOWER);
3410 DBG_OPT_COMBO(irn, leader, FS_OPT_COMBO_CONGRUENT);
3411 exchange_leader(irn, leader);
3418 } /* apply_result */
3421 * Fix the keep-alives by deleting unreachable ones.
3423 static void apply_end(ir_node *end, environment_t *env)
3425 int i, j, n = get_End_n_keepalives(end);
3426 ir_node **in = NULL;
3429 NEW_ARR_A(ir_node *, in, n);
3431 /* fix the keep alive */
3432 for (i = j = 0; i < n; i++) {
3433 ir_node *ka = get_End_keepalive(end, i);
3439 if (!is_Block(ka)) {
3440 block = get_nodes_block(ka);
3447 node = get_irn_node(block);
3448 if (node->type.tv != tarval_unreachable)
3452 set_End_keepalives(end, j, in);
3457 #define SET(code) op_##code->ops.generic = (op_func)compute_##code
3460 * sets the generic functions to compute.
3462 static void set_compute_functions(void)
3466 /* set the default compute function */
3467 for (i = 0, n = get_irp_n_opcodes(); i < n; ++i) {
3468 ir_op *op = get_irp_opcode(i);
3469 op->ops.generic = (op_func)default_compute;
3472 /* set specific functions */
3488 } /* set_compute_functions */
3493 static void add_memory_keeps(ir_node **kept_memory, size_t len)
3495 ir_node *end = get_irg_end(current_ir_graph);
3500 ir_nodeset_init(&set);
3502 /* check, if those nodes are already kept */
3503 for (i = get_End_n_keepalives(end) - 1; i >= 0; --i)
3504 ir_nodeset_insert(&set, get_End_keepalive(end, i));
3506 for (idx = 0; idx < len; ++idx) {
3507 ir_node *ka = kept_memory[idx];
3509 if (! ir_nodeset_contains(&set, ka)) {
3510 add_End_keepalive(end, ka);
3513 ir_nodeset_destroy(&set);
3514 } /* add_memory_keeps */
3516 static ir_graph_state_t do_combo(ir_graph *irg)
3519 ir_node *initial_bl;
3521 ir_graph *rem = current_ir_graph;
3524 current_ir_graph = irg;
3526 /* register a debug mask */
3527 FIRM_DBG_REGISTER(dbg, "firm.opt.combo");
3529 DB((dbg, LEVEL_1, "Doing COMBO for %+F\n", irg));
3531 obstack_init(&env.obst);
3532 env.worklist = NULL;
3536 #ifdef DEBUG_libfirm
3537 env.dbg_list = NULL;
3539 env.opcode2id_map = new_set(cmp_opcode, iro_Last * 4);
3540 env.kept_memory = NEW_ARR_F(ir_node *, 0);
3541 env.end_idx = get_opt_global_cse() ? 0 : -1;
3542 env.lambda_input = 0;
3545 /* options driving the optimization */
3546 env.commutative = 1;
3547 env.opt_unknown = 1;
3549 /* we have our own value_of function */
3550 set_value_of_func(get_node_tarval);
3552 set_compute_functions();
3553 DEBUG_ONLY(part_nr = 0;)
3555 ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK | IR_RESOURCE_PHI_LIST);
3557 if (env.opt_unknown)
3558 tarval_UNKNOWN = tarval_top;
3560 tarval_UNKNOWN = tarval_bad;
3562 /* create the initial partition and place it on the work list */
3563 env.initial = new_partition(&env);
3564 add_to_worklist(env.initial, &env);
3565 irg_walk_graph(irg, create_initial_partitions, init_block_phis, &env);
3567 /* set the hook: from now, every node has a partition and a type */
3568 DEBUG_ONLY(set_dump_node_vcgattr_hook(dump_partition_hook);)
3570 /* all nodes on the initial partition have type Top */
3571 env.initial->type_is_T_or_C = 1;
3573 /* Place the START Node's partition on cprop.
3574 Place the START Node on its local worklist. */
3575 initial_bl = get_irg_start_block(irg);
3576 start = get_irn_node(initial_bl);
3577 add_to_cprop(start, &env);
3581 if (env.worklist != NULL)
3583 } while (env.cprop != NULL || env.worklist != NULL);
3585 dump_all_partitions(&env);
3586 check_all_partitions(&env);
3589 dump_ir_block_graph(irg, "-partition");
3592 /* apply the result */
3594 /* check, which nodes must be kept */
3595 irg_walk_graph(irg, NULL, find_kept_memory, &env);
3597 /* kill unreachable control flow */
3598 irg_block_walk_graph(irg, NULL, apply_cf, &env);
3599 /* Kill keep-alives of dead blocks: this speeds up apply_result()
3600 * and fixes assertion because dead cf to dead blocks is NOT removed by
3602 apply_end(get_irg_end(irg), &env);
3603 irg_walk_graph(irg, NULL, apply_result, &env);
3605 len = ARR_LEN(env.kept_memory);
3607 add_memory_keeps(env.kept_memory, len);
3610 DB((dbg, LEVEL_1, "Unoptimized Control Flow left"));
3613 ir_free_resources(irg, IR_RESOURCE_IRN_LINK | IR_RESOURCE_PHI_LIST);
3615 /* remove the partition hook */
3616 DEBUG_ONLY(set_dump_node_vcgattr_hook(NULL);)
3618 DEL_ARR_F(env.kept_memory);
3619 del_set(env.opcode2id_map);
3620 obstack_free(&env.obst, NULL);
3622 /* restore value_of() default behavior */
3623 set_value_of_func(NULL);
3624 current_ir_graph = rem;
3626 return 0; // cannot guarantee anything
3629 static optdesc_t opt_combo = {
3631 IR_GRAPH_STATE_NO_BADS | IR_GRAPH_STATE_CONSISTENT_OUTS | IR_GRAPH_STATE_CONSISTENT_LOOPINFO,
3635 void combo(ir_graph *irg)
3637 perform_irg_optimization(irg, &opt_combo);
3640 /* Creates an ir_graph pass for combo. */
3641 ir_graph_pass_t *combo_pass(const char *name)
3643 return def_graph_pass(name ? name : "combo", combo);