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 Dataflow driven Load/Store optimizations, uses some ideas from
24 * @author Michael Beck
40 #include "iroptimize.h"
41 #include "irnodemap.h"
42 #include "raw_bitset.h"
46 /* maximum number of output Proj's */
47 #define MAX_PROJ (pn_Load_max > pn_Store_max ? pn_Load_max : pn_Store_max)
50 * Mapping an address to an dense ID.
52 typedef struct address_entry_t {
53 unsigned id; /**< The ID */
60 FLAG_KILL_ALL = 1, /**< KILL all addresses */
61 FLAG_KILLED_NODE = 2, /**< this node was killed */
62 FLAG_EXCEPTION = 4, /**< this node has exception flow */
63 FLAG_IGNORE = 8, /**< ignore this node (volatile or other) */
67 * A value: This represents a value stored at a given address in
68 * memory. Do not confuse with values from value numbering.
70 typedef struct value_t value_t;
72 ir_node *address; /**< the address of this value */
73 ir_node *value; /**< the value itself */
74 ir_mode *mode; /**< the mode of the value */
75 unsigned id; /**< address id */
79 * A memop describes an memory-related operation.
80 * These are Loads/Store and all other ops that might modify
81 * memory (Calls, CopyB) or causing exceptions.
83 typedef struct memop_t memop_t;
85 value_t value; /**< the value of this memop: only defined for Load/Store */
86 ir_node *node; /**< the memory op itself */
87 ir_node *mem; /**< the memory FROM this node */
88 ir_node *replace; /**< the replacement node if this memop is replaced */
89 memop_t *next; /**< links to the next memory op in the block in forward order. */
90 memop_t *prev; /**< links to the previous memory op in the block in forward order. */
91 unsigned flags; /**< memop flags */
92 ir_node *projs[MAX_PROJ]; /**< Projs of this memory op */
96 * Additional data for every basic block.
98 typedef struct block_t block_t;
100 memop_t *memop_forward; /**< topologically sorted list of memory ops in this block */
101 memop_t *memop_backward; /**< last memop in the list */
102 unsigned *avail_out; /**< out-set of available addresses */
103 memop_t **id_2_memop_avail; /**< maps avail address ids to memops */
104 unsigned *anticL_in; /**< in-set of anticipated Load addresses */
105 memop_t **id_2_memop_antic; /**< maps anticipated address ids to memops */
106 ir_node *block; /**< the associated block */
107 block_t *forward_next; /**< next block entry for forward iteration */
108 block_t *backward_next; /**< next block entry for backward iteration */
109 memop_t *avail; /**< used locally for the avail map */
110 memop_t **trans_results; /**< used to cached translated nodes due antic calculation. */
114 * Metadata for this pass.
116 typedef struct ldst_env_t {
117 struct obstack obst; /**< obstack for temporary data */
118 ir_nodemap_t adr_map; /**< Map addresses to */
119 block_t *forward; /**< Inverse post-order list of all blocks Start->End */
120 block_t *backward; /**< Inverse post-order list of all blocks End->Start */
121 ir_node *start_bl; /**< start block of the current graph */
122 ir_node *end_bl; /**< end block of the current graph */
123 unsigned *curr_set; /**< current set of addresses */
124 memop_t **curr_id_2_memop; /**< current map of address ids to memops */
125 unsigned curr_adr_id; /**< number for address mapping */
126 unsigned n_mem_ops; /**< number of memory operations (Loads/Stores) */
127 unsigned rbs_size; /**< size of all bitsets in bytes */
128 int max_cfg_preds; /**< maximum number of block cfg predecessors */
129 int changed; /**< Flags for changed graph state */
131 ir_node **id_2_address; /**< maps an id to the used address */
137 static firm_dbg_module_t *dbg;
139 /* the one and only environment */
143 * Dumps the block list.
145 * @param ldst environment
147 static void dump_block_list(ldst_env *env) {
152 for (entry = env->forward; entry != NULL; entry = entry->forward_next) {
153 DB((dbg, LEVEL_2, "%+F {", entry->block));
156 for (op = entry->memop_forward; op != NULL; op = op->next) {
158 DB((dbg, LEVEL_2, "\n\t"));
159 } DB((dbg, LEVEL_2, "%+F", op->node));
160 if ((op->flags & FLAG_KILL_ALL) == FLAG_KILL_ALL)
161 DB((dbg, LEVEL_2, "X"));
162 else if (op->flags & FLAG_KILL_ALL)
163 DB((dbg, LEVEL_2, "K"));
164 DB((dbg, LEVEL_2, ", "));
168 DB((dbg, LEVEL_2, "\n}\n\n"));
170 } /* dump_block_list */
173 * Dumps the current set.
175 * @param bl current block
176 * @param s name of the set
178 static void dump_curr(block_t *bl, const char *s) {
179 unsigned end = env.rbs_size - 1;
183 DB((dbg, LEVEL_2, "%s[%+F] = {", s, bl->block));
185 for (pos = rbitset_next(env.curr_set, 0, 1); pos < end; pos = rbitset_next(env.curr_set, pos + 1, 1)) {
186 memop_t *op = env.curr_id_2_memop[pos];
189 DB((dbg, LEVEL_2, "\n\t"));
192 DB((dbg, LEVEL_2, "<%+F, %+F>, ", op->value.address, op->value.value));
195 DB((dbg, LEVEL_2, "\n}\n"));
199 #define dump_block_list()
200 #define dump_curr(bl, s)
201 #endif /* DEBUG_libfirm */
203 /** Get the block entry for a block node */
204 static block_t *get_block_entry(const ir_node *block) {
205 assert(is_Block(block));
207 return get_irn_link(block);
208 } /* get_block_entry */
210 /** Get the memop entry for a memory operation node */
211 static memop_t *get_irn_memop(const ir_node *irn) {
212 assert(! is_Block(irn));
213 return get_irn_link(irn);
214 } /* get_irn_memop */
217 * Walk over the memory edges from definition to users.
218 * This ensures, that even operation without memory output are found.
220 * @param irn start node
221 * @param pre pre walker function
222 * @param post post walker function
223 * @param ctx context parameter for the walker functions
225 static void walk_memory(ir_node *irn, irg_walk_func *pre, irg_walk_func *post, void *ctx) {
229 mark_irn_visited(irn);
234 mode = get_irn_mode(irn);
235 if (mode == mode_M) {
236 /* every successor uses memory */
237 for (i = get_irn_n_outs(irn) - 1; i >= 0; --i) {
238 ir_node *succ = get_irn_out(irn, i);
240 if (! irn_visited(succ))
241 walk_memory(succ, pre, post, ctx);
243 } else if (mode == mode_T) {
244 /* only some Proj's uses memory */
245 for (i = get_irn_n_outs(irn) - 1; i >= 0; --i) {
246 ir_node *proj = get_irn_out(irn, i);
248 if (get_irn_mode(proj) == mode_M && ! irn_visited(proj))
249 walk_memory(proj, pre, post, ctx);
257 * Walks over all memory nodes of a graph.
260 * @param pre pre walker function
261 * @param post post walker function
262 * @param ctx context parameter for the walker functions
264 static void walk_memory_irg(ir_graph *irg, irg_walk_func pre, irg_walk_func post, void *ctx) {
265 inc_irg_visited(irg);
267 ir_reserve_resources(irg, IR_RESOURCE_IRN_VISITED);
270 * there are two possible sources for memory: initial_mem and nomem
271 * we ignore nomem as this should NOT change the memory
273 walk_memory(get_irg_initial_mem(irg), pre, post, ctx);
275 ir_free_resources(irg, IR_RESOURCE_IRN_VISITED);
276 } /* walk_memory_irg */
279 * Register an address and allocate a (sparse, 0..n) ID for it.
281 * @param adr the IR-node representing the address
283 * @return the allocated id
285 static unsigned register_address(ir_node *adr) {
286 address_entry *entry;
288 /* skip Confirms and Casts */
290 if (is_Confirm(adr)) {
291 adr = get_Confirm_value(adr);
295 adr = get_Cast_op(adr);
299 entry = ir_nodemap_get(&env.adr_map, adr);
303 entry = obstack_alloc(&env.obst, sizeof(*entry));
305 entry->id = env.curr_adr_id++;
306 ir_nodemap_insert(&env.adr_map, adr, entry);
308 DB((dbg, LEVEL_3, "ADDRESS %+F has ID %u\n", adr, entry->id));
310 ARR_APP1(ir_node *, env.id_2_address, adr);
314 } /* register_address */
318 * translate an address through a Phi node into a given predecessor
321 * @param address the address
322 * @param block the block
323 * @param pos the position of the predecessor in block
325 static ir_node *phi_translate(ir_node *address, const ir_node *block, int pos) {
326 if (is_Phi(address) && get_nodes_block(address) == block)
327 address = get_Phi_pred(address, pos);
329 } /* phi_translate */
332 * Get the effective block of an address in the pos'th predecessor
333 * of the given block.
335 * @param address the address
336 * @param block the block
337 * @param pos the position of the predecessor in block
339 static ir_node *get_effective_block(ir_node *address, ir_node *block, int pos) {
340 address = phi_translate(address, block, pos);
341 return get_nodes_block(address);
342 } /* get_effective_block */
345 * Walker: allocate an block entry for every block
346 * and register all potential addresses.
348 static void prepare_blocks(ir_node *irn, void *ctx) {
352 block_t *entry = obstack_alloc(&env.obst, sizeof(*entry));
355 entry->memop_forward = NULL;
356 entry->memop_backward = NULL;
357 entry->avail_out = NULL;
358 entry->id_2_memop_avail = NULL;
359 entry->anticL_in = NULL;
360 entry->id_2_memop_antic = NULL;
362 entry->forward_next = NULL;
363 entry->backward_next = NULL;
365 entry->trans_results = NULL;
366 set_irn_link(irn, entry);
368 set_Block_phis(irn, NULL);
370 /* use block marks to track unreachable blocks */
371 set_Block_mark(irn, 0);
373 n = get_Block_n_cfgpreds(irn);
374 if (n > env.max_cfg_preds)
375 env.max_cfg_preds = n;
377 ir_mode *mode = get_irn_mode(irn);
379 if (mode_is_reference(mode)) {
381 * Register ALL possible addresses: this is overkill yet but
382 * simpler then doing it for all possible translated addresses
383 * (which would be sufficient in the moment.
385 (void)register_address(irn);
388 } /* prepare_blocks */
391 * Post-Walker, link in all Phi's
393 static void link_phis(ir_node *irn, void *ctx) {
397 ir_node *block = get_nodes_block(irn);
398 add_Block_phi(block, irn);
403 * Block walker: creates the inverse post-order list for the CFG.
405 static void inverse_post_order(ir_node *block, void *ctx) {
406 block_t *entry = get_block_entry(block);
410 /* mark this block IS reachable from start */
411 set_Block_mark(block, 1);
413 /* create the list in inverse order */
414 entry->forward_next = env.forward;
417 /* remember the first visited (last in list) entry, needed for later */
418 if (env.backward == NULL)
419 env.backward = entry;
420 } /* inverse_post_order */
423 * Block walker: create backward links for the memops of a block.
425 static void collect_backward(ir_node *block, void *ctx) {
426 block_t *entry = get_block_entry(block);
432 * Do NOT link in the end block yet. We want it to be
433 * the first in the list. This is NOT guaranteed by the walker
434 * if we have endless loops.
436 if (block != env.end_bl) {
437 entry->backward_next = env.backward;
439 /* create the list in inverse order */
440 env.backward = entry;
443 /* create backward links for all memory ops */
445 for (op = entry->memop_forward; op != NULL; op = op->next) {
449 entry->memop_backward = last;
450 } /* collect_backward */
455 * @param irn the IR-node representing the memop or NULL
456 * if this is a translated (virtual) memop
458 * @return the allocated memop
460 static memop_t *alloc_memop(ir_node *irn) {
461 memop_t *m = obstack_alloc(&env.obst, sizeof(*m));
463 m->value.address = NULL;
464 m->value.value = NULL;
465 m->value.mode = NULL;
473 memset(m->projs, 0, sizeof(m->projs));
476 set_irn_link(irn, m);
481 * Create a memop for a Phi-replacement.
483 * @param op the memop to clone
484 * @param phi the Phi-node representing the new value
486 static memop_t *clone_memop_phi(memop_t *op, ir_node *phi) {
487 memop_t *m = obstack_alloc(&env.obst, sizeof(*m));
489 m->value = op->value;
490 m->value.value = phi;
497 set_irn_link(phi, m);
499 } /* clone_memop_phi */
502 * Return the memory properties of a call node.
504 * @param call the call node
506 * return a bitset of mtp_property_const and mtp_property_pure
508 static unsigned get_Call_memory_properties(ir_node *call) {
509 ir_type *call_tp = get_Call_type(call);
510 unsigned prop = get_method_additional_properties(call_tp);
512 /* check first the call type */
513 if ((prop & (mtp_property_const|mtp_property_pure)) == 0) {
514 /* try the called entity */
515 ir_node *ptr = get_Call_ptr(call);
517 if (is_Global(ptr)) {
518 ir_entity *ent = get_Global_entity(ptr);
520 prop = get_entity_additional_properties(ent);
523 return prop & (mtp_property_const|mtp_property_pure);
524 } /* get_Call_memory_properties */
527 * Returns an entity if the address ptr points to a constant one.
529 * @param ptr the address
531 * @return an entity or NULL
533 static ir_entity *find_constant_entity(ir_node *ptr) {
535 if (is_SymConst(ptr) && get_SymConst_kind(ptr) == symconst_addr_ent) {
536 return get_SymConst_entity(ptr);
537 } else if (is_Sel(ptr)) {
538 ir_entity *ent = get_Sel_entity(ptr);
539 ir_type *tp = get_entity_owner(ent);
541 /* Do not fiddle with polymorphism. */
542 if (is_Class_type(get_entity_owner(ent)) &&
543 ((get_entity_n_overwrites(ent) != 0) ||
544 (get_entity_n_overwrittenby(ent) != 0) ) )
547 if (is_Array_type(tp)) {
551 for (i = 0, n = get_Sel_n_indexs(ptr); i < n; ++i) {
553 tarval *tlower, *tupper;
554 ir_node *index = get_Sel_index(ptr, i);
555 tarval *tv = computed_value(index);
557 /* check if the index is constant */
558 if (tv == tarval_bad)
561 bound = get_array_lower_bound(tp, i);
562 tlower = computed_value(bound);
563 bound = get_array_upper_bound(tp, i);
564 tupper = computed_value(bound);
566 if (tlower == tarval_bad || tupper == tarval_bad)
569 if (tarval_cmp(tv, tlower) & pn_Cmp_Lt)
571 if (tarval_cmp(tupper, tv) & pn_Cmp_Lt)
574 /* ok, bounds check finished */
578 if (variability_constant == get_entity_variability(ent))
582 ptr = get_Sel_ptr(ptr);
583 } else if (is_Add(ptr)) {
584 ir_node *l = get_Add_left(ptr);
585 ir_node *r = get_Add_right(ptr);
587 if (get_irn_mode(l) == get_irn_mode(ptr) && is_Const(r))
589 else if (get_irn_mode(r) == get_irn_mode(ptr) && is_Const(l))
594 /* for now, we support only one addition, reassoc should fold all others */
595 if (! is_SymConst(ptr) && !is_Sel(ptr))
597 } else if (is_Sub(ptr)) {
598 ir_node *l = get_Sub_left(ptr);
599 ir_node *r = get_Sub_right(ptr);
601 if (get_irn_mode(l) == get_irn_mode(ptr) && is_Const(r))
605 /* for now, we support only one subtraction, reassoc should fold all others */
606 if (! is_SymConst(ptr) && !is_Sel(ptr))
611 } /* find_constant_entity */
614 * Return the Selection index of a Sel node from dimension n
616 static long get_Sel_array_index_long(ir_node *n, int dim) {
617 ir_node *index = get_Sel_index(n, dim);
618 assert(is_Const(index));
619 return get_tarval_long(get_Const_tarval(index));
620 } /* get_Sel_array_index_long */
623 * Returns the accessed component graph path for an
624 * node computing an address.
626 * @param ptr the node computing the address
627 * @param depth current depth in steps upward from the root
630 static compound_graph_path *rec_get_accessed_path(ir_node *ptr, int depth) {
631 compound_graph_path *res = NULL;
632 ir_entity *root, *field, *ent;
633 int path_len, pos, idx;
637 if (is_SymConst(ptr)) {
638 /* a SymConst. If the depth is 0, this is an access to a global
639 * entity and we don't need a component path, else we know
640 * at least its length.
642 assert(get_SymConst_kind(ptr) == symconst_addr_ent);
643 root = get_SymConst_entity(ptr);
644 res = (depth == 0) ? NULL : new_compound_graph_path(get_entity_type(root), depth);
645 } else if (is_Sel(ptr)) {
646 /* it's a Sel, go up until we find the root */
647 res = rec_get_accessed_path(get_Sel_ptr(ptr), depth+1);
651 /* fill up the step in the path at the current position */
652 field = get_Sel_entity(ptr);
653 path_len = get_compound_graph_path_length(res);
654 pos = path_len - depth - 1;
655 set_compound_graph_path_node(res, pos, field);
657 if (is_Array_type(get_entity_owner(field))) {
658 assert(get_Sel_n_indexs(ptr) == 1 && "multi dim arrays not implemented");
659 set_compound_graph_path_array_index(res, pos, get_Sel_array_index_long(ptr, 0));
661 } else if (is_Add(ptr)) {
662 ir_node *l = get_Add_left(ptr);
663 ir_node *r = get_Add_right(ptr);
664 ir_mode *mode = get_irn_mode(ptr);
667 if (is_Const(r) && get_irn_mode(l) == mode) {
669 tv = get_Const_tarval(r);
672 tv = get_Const_tarval(l);
675 mode = get_tarval_mode(tv);
678 /* ptr must be a Sel or a SymConst, this was checked in find_constant_entity() */
680 field = get_Sel_entity(ptr);
682 field = get_SymConst_entity(ptr);
685 for (ent = field;;) {
687 tarval *sz, *tv_index, *tlower, *tupper;
690 tp = get_entity_type(ent);
691 if (! is_Array_type(tp))
693 ent = get_array_element_entity(tp);
694 size = get_type_size_bytes(get_entity_type(ent));
695 sz = new_tarval_from_long(size, mode);
697 tv_index = tarval_div(tmp, sz);
698 tmp = tarval_mod(tmp, sz);
700 if (tv_index == tarval_bad || tmp == tarval_bad)
703 assert(get_array_n_dimensions(tp) == 1 && "multiarrays not implemented");
704 bound = get_array_lower_bound(tp, 0);
705 tlower = computed_value(bound);
706 bound = get_array_upper_bound(tp, 0);
707 tupper = computed_value(bound);
709 if (tlower == tarval_bad || tupper == tarval_bad)
712 if (tarval_cmp(tv_index, tlower) & pn_Cmp_Lt)
714 if (tarval_cmp(tupper, tv_index) & pn_Cmp_Lt)
717 /* ok, bounds check finished */
720 if (! tarval_is_null(tmp)) {
721 /* access to some struct/union member */
725 /* should be at least ONE array */
729 res = rec_get_accessed_path(ptr, depth + idx);
733 path_len = get_compound_graph_path_length(res);
734 pos = path_len - depth - idx;
736 for (ent = field;;) {
738 tarval *sz, *tv_index;
741 tp = get_entity_type(ent);
742 if (! is_Array_type(tp))
744 ent = get_array_element_entity(tp);
745 set_compound_graph_path_node(res, pos, ent);
747 size = get_type_size_bytes(get_entity_type(ent));
748 sz = new_tarval_from_long(size, mode);
750 tv_index = tarval_div(tv, sz);
751 tv = tarval_mod(tv, sz);
753 /* worked above, should work again */
754 assert(tv_index != tarval_bad && tv != tarval_bad);
756 /* bounds already checked above */
757 index = get_tarval_long(tv_index);
758 set_compound_graph_path_array_index(res, pos, index);
761 } else if (is_Sub(ptr)) {
762 ir_node *l = get_Sub_left(ptr);
763 ir_node *r = get_Sub_right(ptr);
766 tv = get_Const_tarval(r);
771 } /* rec_get_accessed_path */
774 * Returns an access path or NULL. The access path is only
775 * valid, if the graph is in phase_high and _no_ address computation is used.
777 static compound_graph_path *get_accessed_path(ir_node *ptr) {
778 compound_graph_path *gr = rec_get_accessed_path(ptr, 0);
780 } /* get_accessed_path */
782 typedef struct path_entry {
784 struct path_entry *next;
788 static ir_node *rec_find_compound_ent_value(ir_node *ptr, path_entry *next) {
789 path_entry entry, *p;
790 ir_entity *ent, *field;
791 ir_initializer_t *initializer;
797 if (is_SymConst(ptr)) {
799 ent = get_SymConst_entity(ptr);
800 initializer = get_entity_initializer(ent);
801 for (p = next; p != NULL;) {
802 if (initializer->kind != IR_INITIALIZER_COMPOUND)
804 n = get_initializer_compound_n_entries(initializer);
805 tp = get_entity_type(ent);
807 if (is_Array_type(tp)) {
808 ent = get_array_element_entity(tp);
813 initializer = get_initializer_compound_value(initializer, 0);
817 if (p->index >= (int) n)
819 initializer = get_initializer_compound_value(initializer, p->index);
824 tp = get_entity_type(ent);
825 while (is_Array_type(tp)) {
826 ent = get_array_element_entity(tp);
827 tp = get_entity_type(ent);
829 n = get_initializer_compound_n_entries(initializer);
832 initializer = get_initializer_compound_value(initializer, 0);
835 switch (initializer->kind) {
836 case IR_INITIALIZER_CONST:
837 return get_initializer_const_value(initializer);
838 case IR_INITIALIZER_TARVAL:
839 case IR_INITIALIZER_NULL:
843 } else if (is_Sel(ptr)) {
844 entry.ent = field = get_Sel_entity(ptr);
845 tp = get_entity_owner(field);
846 if (is_Array_type(tp)) {
847 assert(get_Sel_n_indexs(ptr) == 1 && "multi dim arrays not implemented");
848 entry.index = get_Sel_array_index_long(ptr, 0) - get_array_lower_bound_int(tp, 0);
850 int i, n_members = get_compound_n_members(tp);
851 for (i = 0; i < n_members; ++i) {
852 if (get_compound_member(tp, i) == field)
855 if (i >= n_members) {
856 /* not found: should NOT happen */
861 return rec_find_compound_ent_value(get_Sel_ptr(ptr), &entry);
862 } else if (is_Add(ptr)) {
863 ir_node *l = get_Add_left(ptr);
864 ir_node *r = get_Add_right(ptr);
870 tv = get_Const_tarval(r);
873 tv = get_Const_tarval(l);
876 mode = get_tarval_mode(tv);
878 /* ptr must be a Sel or a SymConst, this was checked in find_constant_entity() */
880 field = get_Sel_entity(ptr);
882 field = get_SymConst_entity(ptr);
885 /* count needed entries */
887 for (ent = field;;) {
888 tp = get_entity_type(ent);
889 if (! is_Array_type(tp))
891 ent = get_array_element_entity(tp);
894 /* should be at least ONE entry */
898 /* allocate the right number of entries */
899 NEW_ARR_A(path_entry, p, pos);
903 for (ent = field;;) {
905 tarval *sz, *tv_index, *tlower, *tupper;
909 tp = get_entity_type(ent);
910 if (! is_Array_type(tp))
912 ent = get_array_element_entity(tp);
914 p[pos].next = &p[pos + 1];
916 size = get_type_size_bytes(get_entity_type(ent));
917 sz = new_tarval_from_long(size, mode);
919 tv_index = tarval_div(tv, sz);
920 tv = tarval_mod(tv, sz);
922 if (tv_index == tarval_bad || tv == tarval_bad)
925 assert(get_array_n_dimensions(tp) == 1 && "multiarrays not implemented");
926 bound = get_array_lower_bound(tp, 0);
927 tlower = computed_value(bound);
928 bound = get_array_upper_bound(tp, 0);
929 tupper = computed_value(bound);
931 if (tlower == tarval_bad || tupper == tarval_bad)
934 if (tarval_cmp(tv_index, tlower) & pn_Cmp_Lt)
936 if (tarval_cmp(tupper, tv_index) & pn_Cmp_Lt)
939 /* ok, bounds check finished */
940 index = get_tarval_long(tv_index);
941 p[pos].index = index;
944 if (! tarval_is_null(tv)) {
945 /* hmm, wrong access */
948 p[pos - 1].next = next;
949 return rec_find_compound_ent_value(ptr, p);
950 } else if (is_Sub(ptr)) {
951 ir_node *l = get_Sub_left(ptr);
952 ir_node *r = get_Sub_right(ptr);
955 tv = get_Const_tarval(r);
960 } /* rec_find_compound_ent_value */
962 static ir_node *find_compound_ent_value(ir_node *ptr) {
963 return rec_find_compound_ent_value(ptr, NULL);
964 } /* find_compound_ent_value */
967 * Mark a Load memop to be replace by a definition
969 * @param op the Load memop
971 static void mark_replace_load(memop_t *op, ir_node *def) {
973 op->flags |= FLAG_KILLED_NODE;
975 } /* mark_replace_load */
978 * Mark a Store memop to be removed.
980 * @param op the Store memop
982 static void mark_remove_store(memop_t *op) {
983 op->flags |= FLAG_KILLED_NODE;
985 } /* mark_remove_store */
988 * Update a memop for a Load.
992 static void update_Load_memop(memop_t *m) {
994 ir_node *load = m->node;
998 if (get_Load_volatility(load) == volatility_is_volatile)
999 m->flags |= FLAG_IGNORE;
1001 ptr = get_Load_ptr(load);
1003 m->value.address = ptr;
1005 for (i = get_irn_n_outs(load) - 1; i >= 0; --i) {
1006 ir_node *proj = get_irn_out(load, i);
1009 /* beware of keep edges */
1013 pn = get_Proj_proj(proj);
1014 m->projs[pn] = proj;
1017 m->value.value = proj;
1018 m->value.mode = get_irn_mode(proj);
1020 case pn_Load_X_except:
1021 m->flags |= FLAG_EXCEPTION;
1026 case pn_Load_X_regular:
1029 panic("Unsupported Proj from Load %+F", proj);
1033 /* check if we can determine the entity that will be loaded */
1034 ent = find_constant_entity(ptr);
1037 allocation_static == get_entity_allocation(ent) &&
1038 visibility_external_allocated != get_entity_visibility(ent)) {
1039 /* a static allocation that is not external: there should be NO exception
1040 * when loading even if we cannot replace the load itself. */
1041 ir_node *value = NULL;
1043 /* no exception, clear the m fields as it might be checked later again */
1044 if (m->projs[pn_Load_X_except]) {
1045 exchange(m->projs[pn_Load_X_except], new_Bad());
1046 m->projs[pn_Load_X_except] = NULL;
1047 m->flags &= ~FLAG_EXCEPTION;
1050 if (m->projs[pn_Load_X_regular]) {
1051 exchange(m->projs[pn_Load_X_regular], new_r_Jmp(current_ir_graph, get_nodes_block(load)));
1052 m->projs[pn_Load_X_regular] = NULL;
1056 if (variability_constant == get_entity_variability(ent)) {
1057 if (is_atomic_entity(ent)) {
1058 /* Might not be atomic after lowering of Sels. In this case we
1059 * could also load, but it's more complicated. */
1060 /* more simpler case: we load the content of a constant value:
1061 * replace it by the constant itself */
1062 value = get_atomic_ent_value(ent);
1063 } else if (ent->has_initializer) {
1064 /* new style initializer */
1065 value = find_compound_ent_value(ptr);
1067 /* old style initializer */
1068 compound_graph_path *path = get_accessed_path(ptr);
1071 assert(is_proper_compound_graph_path(path, get_compound_graph_path_length(path)-1));
1073 value = get_compound_ent_value_by_path(ent, path);
1074 DB((dbg, LEVEL_1, " Constant access at %F%F resulted in %+F\n", ent, path, value));
1075 free_compound_graph_path(path);
1079 value = can_replace_load_by_const(load, value);
1082 if (value != NULL) {
1083 /* we completely replace the load by this value */
1084 DB((dbg, LEVEL_1, "Replacing Load %+F by constant %+F\n", m->node, value));
1085 mark_replace_load(m, value);
1090 if (m->value.value != NULL && !(m->flags & FLAG_IGNORE)) {
1091 /* only create an address if this node is NOT killed immediately or ignored */
1092 m->value.id = register_address(ptr);
1095 /* no user, KILL it */
1096 mark_replace_load(m, NULL);
1098 } /* update_Load_memop */
1101 * Update a memop for a Store.
1103 * @param m the memop
1105 static void update_Store_memop(memop_t *m) {
1107 ir_node *store = m->node;
1108 ir_node *adr = get_Store_ptr(store);
1110 if (get_Store_volatility(store) == volatility_is_volatile) {
1111 m->flags |= FLAG_IGNORE;
1113 /* only create an address if this node is NOT ignored */
1114 m->value.id = register_address(adr);
1118 m->value.address = adr;
1120 for (i = get_irn_n_outs(store) - 1; i >= 0; --i) {
1121 ir_node *proj = get_irn_out(store, i);
1124 /* beware of keep edges */
1128 pn = get_Proj_proj(proj);
1129 m->projs[pn] = proj;
1131 case pn_Store_X_except:
1132 m->flags |= FLAG_EXCEPTION;
1137 case pn_Store_X_regular:
1140 panic("Unsupported Proj from Store %+F", proj);
1143 m->value.value = get_Store_value(store);
1144 m->value.mode = get_irn_mode(m->value.value);
1145 } /* update_Store_memop */
1148 * Update a memop for a Call.
1150 * @param m the memop
1152 static void update_Call_memop(memop_t *m) {
1153 ir_node *call = m->node;
1154 unsigned prop = get_Call_memory_properties(call);
1157 if (prop & mtp_property_const) {
1158 /* A constant call did NOT use memory at all, we
1159 can kick it from the list. */
1160 } else if (prop & mtp_property_pure) {
1161 /* pure calls READ memory */
1164 m->flags = FLAG_KILL_ALL;
1166 for (i = get_irn_n_outs(call) - 1; i >= 0; --i) {
1167 ir_node *proj = get_irn_out(call, i);
1169 /* beware of keep edges */
1173 switch (get_Proj_proj(proj)) {
1174 case pn_Call_X_except:
1175 m->flags |= FLAG_EXCEPTION;
1177 case pn_Call_M_regular:
1182 } /* update_Call_memop */
1185 * Update a memop for a Div/Mod/Quot/DivMod.
1187 * @param m the memop
1189 static void update_DivOp_memop(memop_t *m) {
1190 ir_node *div = m->node;
1193 for (i = get_irn_n_outs(div) - 1; i >= 0; --i) {
1194 ir_node *proj = get_irn_out(div, i);
1196 /* beware of keep edges */
1200 switch (get_Proj_proj(proj)) {
1201 case pn_Generic_X_except:
1202 m->flags |= FLAG_EXCEPTION;
1204 case pn_Generic_M_regular:
1209 } /* update_DivOp_memop */
1212 * Update a memop for a Phi.
1214 * @param m the memop
1216 static void update_Phi_memop(memop_t *m) {
1217 /* the Phi is it's own mem */
1219 } /* update_Phi_memop */
1222 * Memory walker: collect all memory ops and build topological lists.
1224 static void collect_memops(ir_node *irn, void *ctx) {
1231 /* we can safely ignore ProjM's except the initial memory */
1232 if (irn != get_irg_initial_mem(current_ir_graph))
1236 op = alloc_memop(irn);
1237 block = get_nodes_block(irn);
1238 entry = get_block_entry(block);
1241 update_Phi_memop(op);
1242 /* Phis must be always placed first */
1243 op->next = entry->memop_forward;
1244 entry->memop_forward = op;
1245 if (entry->memop_backward == NULL)
1246 entry->memop_backward = op;
1248 switch (get_irn_opcode(irn)) {
1250 update_Load_memop(op);
1253 update_Store_memop(op);
1256 update_Call_memop(op);
1263 /* initial memory */
1268 /* we can those to find the memory edge */
1274 update_DivOp_memop(op);
1278 /* TODO: handle some builtins */
1280 /* unsupported operation */
1281 op->flags = FLAG_KILL_ALL;
1285 /* all other should be placed last */
1286 if (entry->memop_backward == NULL) {
1287 entry->memop_forward = entry->memop_backward = op;
1289 entry->memop_backward->next = op;
1290 entry->memop_backward = op;
1293 } /* collect_memops */
1296 * Find an address in the current set.
1298 * @param value the value to be searched for
1300 * @return a memop for the value or NULL if the value does
1301 * not exists in the set or cannot be converted into
1302 * the requested mode
1304 static memop_t *find_address(const value_t *value) {
1305 if (rbitset_is_set(env.curr_set, value->id)) {
1306 memop_t *res = env.curr_id_2_memop[value->id];
1308 if (res->value.mode == value->mode)
1310 /* allow hidden casts */
1311 if (get_mode_arithmetic(res->value.mode) == irma_twos_complement &&
1312 get_mode_arithmetic(value->mode) == irma_twos_complement &&
1313 get_mode_size_bits(res->value.mode) == get_mode_size_bits(value->mode))
1317 } /* find_address */
1320 * Find an address in the avail_out set.
1322 * @param bl the block
1323 * @param value the value to be searched for
1325 static memop_t *find_address_avail(const block_t *bl, const value_t *value) {
1326 if (rbitset_is_set(bl->avail_out, value->id)) {
1327 memop_t *res = bl->id_2_memop_avail[value->id];
1329 if (res->value.mode == value->mode)
1331 /* allow hidden casts */
1332 if (get_mode_arithmetic(res->value.mode) == irma_twos_complement &&
1333 get_mode_arithmetic(value->mode) == irma_twos_complement &&
1334 get_mode_size_bits(res->value.mode) == get_mode_size_bits(value->mode))
1338 } /* find_address_avail */
1341 * Kill all addresses from the current set.
1343 static void kill_all(void) {
1344 rbitset_clear_all(env.curr_set, env.rbs_size);
1347 rbitset_set(env.curr_set, env.rbs_size - 1);
1351 * Kill memops that are not alias free due to a Store value from the current set.
1353 * @param value the Store value
1355 static void kill_memops(const value_t *value) {
1356 unsigned end = env.rbs_size - 1;
1359 for (pos = rbitset_next(env.curr_set, 0, 1); pos < end; pos = rbitset_next(env.curr_set, pos + 1, 1)) {
1360 memop_t *op = env.curr_id_2_memop[pos];
1362 if (ir_no_alias != get_alias_relation(current_ir_graph, value->address, value->mode,
1363 op->value.address, op->value.mode)) {
1364 rbitset_clear(env.curr_set, pos);
1365 env.curr_id_2_memop[pos] = NULL;
1366 DB((dbg, LEVEL_2, "KILLING %+F because of possible alias address %+F\n", op->node, value->address));
1372 * Add the value of a memop to the current set.
1374 * @param op the memory op
1376 static void add_memop(memop_t *op) {
1377 rbitset_set(env.curr_set, op->value.id);
1378 env.curr_id_2_memop[op->value.id] = op;
1382 * Add the value of a memop to the avail_out set.
1384 * @param bl the block
1385 * @param op the memory op
1387 static void add_memop_avail(block_t *bl, memop_t *op) {
1388 rbitset_set(bl->avail_out, op->value.id);
1389 bl->id_2_memop_avail[op->value.id] = op;
1390 } /* add_memop_avail */
1393 * Check, if we can convert a value of one mode to another mode
1394 * without changing the representation of bits.
1396 * @param from the original mode
1397 * @param to the destination mode
1399 static int can_convert_to(const ir_mode *from, const ir_mode *to) {
1400 if (get_mode_arithmetic(from) == irma_twos_complement &&
1401 get_mode_arithmetic(to) == irma_twos_complement &&
1402 get_mode_size_bits(from) == get_mode_size_bits(to))
1405 } /* can_convert_to */
1408 * Add a Conv to the requested mode if needed.
1410 * @param irn the IR-node to convert
1411 * @param mode the destination mode
1413 * @return the possible converted node or NULL
1414 * if the conversion is not possible
1416 static ir_node *conv_to(ir_node *irn, ir_mode *mode) {
1417 ir_mode *other = get_irn_mode(irn);
1418 if (other != mode) {
1419 /* different modes: check if conversion is possible without changing the bits */
1420 if (can_convert_to(other, mode)) {
1421 ir_node *block = get_nodes_block(irn);
1422 return new_r_Conv(current_ir_graph, block, irn, mode);
1424 /* otherwise not possible ... yet */
1431 * Update the address of an value if this address was a load result
1432 * and the load is killed now.
1434 * @param value the value whose address is updated
1436 static void update_address(value_t *value) {
1437 if (is_Proj(value->address)) {
1438 ir_node *load = get_Proj_pred(value->address);
1440 if (is_Load(load)) {
1441 const memop_t *op = get_irn_memop(load);
1443 if (op->flags & FLAG_KILLED_NODE)
1444 value->address = op->replace;
1447 } /* update_address */
1450 * Do forward dataflow analysis on the given block and calculate the
1451 * GEN and KILL in the current (avail) set.
1453 * @param bl the block
1455 static void calc_gen_kill_avail(block_t *bl) {
1459 for (op = bl->memop_forward; op != NULL; op = op->next) {
1460 switch (get_irn_opcode(op->node)) {
1468 if (! (op->flags & (FLAG_KILLED_NODE|FLAG_IGNORE))) {
1469 /* do we have this already? */
1472 update_address(&op->value);
1473 other = find_address(&op->value);
1474 if (other != NULL && other != op) {
1475 def = conv_to(other->value.value, op->value.mode);
1477 #ifdef DEBUG_libfirm
1478 if (is_Store(other->node)) {
1480 DB((dbg, LEVEL_1, "RAW %+F <- %+F(%+F)\n", op->node, def, other->node));
1483 DB((dbg, LEVEL_1, "RAR %+F <- %+F(%+F)\n", op->node, def, other->node));
1486 mark_replace_load(op, def);
1487 /* do NOT change the memop table */
1491 /* add this value */
1496 if (! (op->flags & FLAG_KILLED_NODE)) {
1497 /* do we have this store already */
1500 update_address(&op->value);
1501 other = find_address(&op->value);
1502 if (other != NULL) {
1503 if (is_Store(other->node)) {
1504 if (op != other && !(other->flags & FLAG_IGNORE) &&
1505 get_nodes_block(other->node) == get_nodes_block(op->node)) {
1507 * A WAW in the same block we can kick the first store.
1508 * This is a shortcut: we know that the second Store will be anticipated
1511 DB((dbg, LEVEL_1, "WAW %+F <- %+F\n", other->node, op->node));
1512 mark_remove_store(other);
1513 /* FIXME: a Load might be get freed due to this killed store */
1515 } else if (other->value.value == op->value.value && !(op->flags & FLAG_IGNORE)) {
1517 DB((dbg, LEVEL_1, "WAR %+F <- %+F\n", op->node, other->node));
1518 mark_remove_store(op);
1519 /* do NOT change the memop table */
1523 /* KILL all possible aliases */
1524 kill_memops(&op->value);
1525 /* add this value */
1530 if (op->flags & FLAG_KILL_ALL)
1534 } /* calc_gen_kill_avail */
1536 #define BYTE_SIZE(x) (((x) + 7) >> 3)
1539 * Do forward dataflow analysis on a given block to calculate the avail_out set
1540 * for this block only.
1542 * @param block the block
1544 static void forward_avail(block_t *bl) {
1545 /* fill the data from the current block */
1546 env.curr_id_2_memop = bl->id_2_memop_avail;
1547 env.curr_set = bl->avail_out;
1549 calc_gen_kill_avail(bl);
1550 dump_curr(bl, "Avail_out");
1551 } /* forward_avail */
1554 * Do backward dataflow analysis on a given block to calculate the antic set
1555 * of Loaded addresses.
1557 * @param bl the block
1559 * @return non-zero if the set has changed since last iteration
1561 static int backward_antic(block_t *bl) {
1563 ir_node *block = bl->block;
1564 int n = get_Block_n_cfg_outs(block);
1567 ir_node *succ = get_Block_cfg_out(block, 0);
1568 block_t *succ_bl = get_block_entry(succ);
1569 int pred_pos = get_Block_cfgpred_pos(succ, block);
1570 unsigned end = env.rbs_size;
1575 if (bl->trans_results == NULL) {
1576 /* allocate the translate cache */
1577 unsigned size = env.curr_adr_id * sizeof(bl->trans_results[0]);
1578 bl->trans_results = obstack_alloc(&env.obst, size);
1579 memset(bl->trans_results, 0, size);
1582 /* check for partly redundant values */
1583 for (pos = rbitset_next(succ_bl->anticL_in, 0, 1);
1585 pos = rbitset_next(succ_bl->anticL_in, pos + 1, 1)) {
1587 * do Phi-translation here: Note that at this point the nodes are
1588 * not changed, so we can safely cache the results.
1589 * However: Loads of Load results ARE bad, because we have no way
1590 to translate them yet ...
1592 memop_t *op = bl->trans_results[pos];
1594 /* not yet translated */
1595 ir_node *adr, *trans_adr;
1597 op = succ_bl->id_2_memop_antic[pos];
1598 adr = op->value.address;
1600 trans_adr = phi_translate(adr, succ, pred_pos);
1601 if (trans_adr != adr) {
1602 /* create a new entry for the translated one */
1605 new_op = alloc_memop(NULL);
1606 new_op->value.address = trans_adr;
1607 new_op->value.id = register_address(trans_adr);
1608 new_op->value.mode = op->value.mode;
1609 new_op->node = op->node; /* we need the node to decide if Load/Store */
1610 new_op->flags = op->flags;
1612 bl->trans_results[pos] = op;
1616 rbitset_set(env.curr_set, op->value.id);
1617 env.curr_id_2_memop[pos] = op;
1620 ir_node *succ = get_Block_cfg_out(block, 0);
1621 block_t *succ_bl = get_block_entry(succ);
1624 rbitset_cpy(env.curr_set, succ_bl->anticL_in, env.rbs_size);
1625 memcpy(env.curr_id_2_memop, succ_bl->id_2_memop_antic, env.rbs_size * sizeof(env.curr_id_2_memop[0]));
1627 /* Hmm: probably we want kill merges of Loads ans Stores here */
1628 for (i = n - 1; i > 0; --i) {
1629 ir_node *succ = get_Block_cfg_out(bl->block, i);
1630 block_t *succ_bl = get_block_entry(succ);
1632 rbitset_and(env.curr_set, succ_bl->anticL_in, env.rbs_size);
1635 /* block ends with a noreturn call */
1639 dump_curr(bl, "AnticL_out");
1641 for (op = bl->memop_backward; op != NULL; op = op->prev) {
1642 switch (get_irn_opcode(op->node)) {
1650 if (! (op->flags & (FLAG_KILLED_NODE|FLAG_IGNORE))) {
1656 if (! (op->flags & FLAG_KILLED_NODE)) {
1657 /* a Store: check which memops must be killed */
1658 kill_memops(&op->value);
1662 if (op->flags & FLAG_KILL_ALL)
1667 memcpy(bl->id_2_memop_antic, env.curr_id_2_memop, env.rbs_size * sizeof(env.curr_id_2_memop[0]));
1668 if (! rbitset_equal(bl->anticL_in, env.curr_set, env.rbs_size)) {
1670 rbitset_cpy(bl->anticL_in, env.curr_set, env.rbs_size);
1671 dump_curr(bl, "AnticL_in*");
1674 dump_curr(bl, "AnticL_in");
1676 } /* backward_antic */
1679 * Replace a Load memop by a already known value.
1681 * @param op the Load memop
1683 static void replace_load(memop_t *op) {
1684 ir_node *load = op->node;
1685 ir_node *def = skip_Id(op->replace);
1690 DB((dbg, LEVEL_1, "Replacing %+F by definition %+F\n", load, is_Proj(def) ? get_Proj_pred(def) : def));
1692 if (op->flags & FLAG_EXCEPTION) {
1693 /* bad: this node is unused and executed for exception only */
1694 DB((dbg, LEVEL_1, "Unused %+F executed for exception only ...\n", load));
1697 DB((dbg, LEVEL_1, "Killing unused %+F\n", load));
1700 if (op->mem != NULL) {
1701 /* in rare cases a Load might have NO memory */
1702 exchange(op->mem, get_Load_mem(load));
1704 proj = op->projs[pn_Load_res];
1706 mode = get_irn_mode(proj);
1707 if (get_irn_mode(def) != mode) {
1709 dbg_info *db = get_irn_dbg_info(load);
1710 ir_node *block = get_nodes_block(proj);
1711 def = new_rd_Conv(db, current_ir_graph, block, def, mode);
1713 exchange(proj, def);
1715 proj = op->projs[pn_Load_X_except];
1717 exchange(proj, new_Bad());
1719 proj = op->projs[pn_Load_X_regular];
1721 exchange(proj, new_r_Jmp(current_ir_graph, get_nodes_block(load)));
1723 } /* replace_load */
1726 * Remove a Store memop.
1728 * @param op the Store memop
1730 static void remove_store(memop_t *op) {
1731 ir_node *store = op->node;
1734 DB((dbg, LEVEL_1, "Removing %+F\n", store));
1736 if (op->mem != NULL) {
1737 /* in rare cases a Store might have no memory */
1738 exchange(op->mem, get_Store_mem(store));
1740 proj = op->projs[pn_Store_X_except];
1742 exchange(proj, new_Bad());
1744 proj = op->projs[pn_Store_X_regular];
1746 exchange(proj, new_r_Jmp(current_ir_graph, get_nodes_block(store)));
1748 } /* remove_store */
1752 * Do all necessary replacements for a given block.
1754 * @param bl the block
1756 static void do_replacements(block_t *bl) {
1759 for (op = bl->memop_forward; op != NULL; op = op->next) {
1760 if (op->flags & FLAG_KILLED_NODE) {
1761 switch (get_irn_opcode(op->node)) {
1771 } /* do_replacements */
1774 * Calculate the Avail_out sets for all basic blocks.
1776 static void calcAvail(void) {
1777 memop_t **tmp_memop = env.curr_id_2_memop;
1778 unsigned *tmp_set = env.curr_set;
1781 /* calculate avail_out */
1782 DB((dbg, LEVEL_2, "Calculate Avail_out\n"));
1784 /* iterate over all blocks in in any order, skip the start block */
1785 for (bl = env.forward->forward_next; bl != NULL; bl = bl->forward_next) {
1789 /* restore the current sets */
1790 env.curr_id_2_memop = tmp_memop;
1791 env.curr_set = tmp_set;
1795 * Calculate the Antic_in sets for all basic blocks.
1797 static void calcAntic(void) {
1800 /* calculate antic_out */
1801 DB((dbg, LEVEL_2, "Calculate Antic_in\n"));
1806 DB((dbg, LEVEL_2, "Iteration %d:\n=========\n", i));
1810 /* over all blocks in reverse post order */
1811 for (bl = env.backward->backward_next; bl != NULL; bl = bl->backward_next) {
1812 need_iter |= backward_antic(bl);
1815 } while (need_iter);
1816 DB((dbg, LEVEL_2, "Get anticipated Load set after %d iterations\n", i));
1820 * Return the node representing the last memory in a block.
1822 * @param bl the block
1824 static ir_node *find_last_memory(block_t *bl) {
1826 if (bl->memop_backward != NULL) {
1827 return bl->memop_backward->mem;
1829 /* if there is NO memory in this block, go to the dominator */
1830 bl = get_block_entry(get_Block_idom(bl->block));
1832 } /* find_last_memory */
1835 * Reroute all memory users of old memory
1836 * to a new memory IR-node.
1838 * @param omem the old memory IR-node
1839 * @param nmem the new memory IR-node
1841 static void reroute_all_mem_users(ir_node *omem, ir_node *nmem) {
1844 for (i = get_irn_n_outs(omem) - 1; i >= 0; --i) {
1846 ir_node *user = get_irn_out_ex(omem, i, &n_pos);
1848 set_irn_n(user, n_pos, nmem);
1851 /* all edges previously point to omem now point to nmem */
1852 nmem->out = omem->out;
1853 } /* reroute_all_mem_users */
1856 * Reroute memory users of old memory that are dominated by a given block
1857 * to a new memory IR-node.
1859 * @param omem the old memory IR-node
1860 * @param nmem the new memory IR-node
1861 * @param pass_bl the block the memory must pass
1863 static void reroute_mem_through(ir_node *omem, ir_node *nmem, ir_node *pass_bl) {
1864 int i, j, n = get_irn_n_outs(omem);
1865 ir_def_use_edge *edges = NEW_ARR_D(ir_def_use_edge, &env.obst, n + 1);
1867 for (i = j = 0; i < n; ++i) {
1869 ir_node *user = get_irn_out_ex(omem, i, &n_pos);
1870 ir_node *use_bl = get_nodes_block(user);
1874 use_bl = get_Block_cfgpred_block(use_bl, n_pos);
1876 if (block_dominates(pass_bl, use_bl)) {
1877 /* found an user that is dominated */
1879 edges[j].pos = n_pos;
1880 edges[j].use = user;
1882 set_irn_n(user, n_pos, nmem);
1886 /* Modify the out structure: we create a new out edge array on our
1887 temporary obstack here. This should be no problem, as we invalidate the edges
1888 at the end either. */
1889 /* first entry is used for the length */
1892 } /* reroute_mem_through */
1895 * insert Loads, making partly redundant Loads fully redundant
1897 static int insert_Load(block_t *bl) {
1898 ir_node *block = bl->block;
1899 int i, n = get_Block_n_cfgpreds(block);
1900 unsigned end = env.rbs_size - 1;
1903 DB((dbg, LEVEL_3, "processing %+F\n", block));
1906 /* might still happen for an unreachable block (end for instance) */
1913 NEW_ARR_A(ir_node *, ins, n);
1915 rbitset_set_all(env.curr_set, env.rbs_size);
1917 /* More than one predecessors, calculate the join for all avail_outs ignoring unevaluated
1918 Blocks. These put in Top anyway. */
1919 for (i = n - 1; i >= 0; --i) {
1920 ir_node *pred = skip_Proj(get_Block_cfgpred(block, i));
1921 ir_node *blk = get_nodes_block(pred);
1924 pred_bl = get_block_entry(blk);
1925 rbitset_and(env.curr_set, pred_bl->avail_out, env.rbs_size);
1927 if (is_Load(pred) || is_Store(pred)) {
1928 /* We reached this block by an exception from a Load or Store:
1929 * the memop creating the exception was NOT completed than, kill it
1931 memop_t *exc_op = get_irn_memop(pred);
1932 rbitset_clear(env.curr_set, exc_op->value.id);
1937 * Ensure that all values are in the map: build Phi's if necessary:
1938 * Note: the last bit is the sentinel and ALWAYS set, so start with -2.
1940 for (pos = env.rbs_size - 2; pos >= 0; --pos) {
1941 if (! rbitset_is_set(env.curr_set, pos))
1942 env.curr_id_2_memop[pos] = NULL;
1944 ir_node *pred = get_Block_cfgpred_block(bl->block, 0);
1945 block_t *pred_bl = get_block_entry(pred);
1947 memop_t *first = NULL;
1950 for (i = 0; i < n; ++i) {
1953 pred = get_Block_cfgpred_block(bl->block, i);
1954 pred_bl = get_block_entry(pred);
1956 mop = pred_bl->id_2_memop_avail[pos];
1957 if (first == NULL) {
1959 ins[0] = first->value.value;
1960 mode = get_irn_mode(ins[0]);
1962 /* no Phi needed so far */
1963 env.curr_id_2_memop[pos] = first;
1965 ins[i] = conv_to(mop->value.value, mode);
1966 if (ins[i] != ins[0]) {
1967 if (ins[i] == NULL) {
1968 /* conversion failed */
1969 env.curr_id_2_memop[pos] = NULL;
1970 rbitset_clear(env.curr_set, pos);
1979 ir_node *phi = new_r_Phi(current_ir_graph, bl->block, n, ins, mode);
1980 memop_t *phiop = alloc_memop(phi);
1982 phiop->value = first->value;
1983 phiop->value.value = phi;
1985 /* no need to link it in, as it is a DATA phi */
1987 env.curr_id_2_memop[pos] = phiop;
1989 DB((dbg, LEVEL_3, "Created new %+F on merging value for address %+F\n", phi, first->value.address));
1994 /* only one predecessor, simply copy the map */
1995 ir_node *pred = get_Block_cfgpred_block(bl->block, 0);
1996 block_t *pred_bl = get_block_entry(pred);
1998 rbitset_cpy(env.curr_set, pred_bl->avail_out, env.rbs_size);
2000 memcpy(env.curr_id_2_memop, pred_bl->id_2_memop_avail, env.rbs_size * sizeof(bl->id_2_memop_avail[0]));
2004 /* check for partly redundant values */
2005 for (pos = rbitset_next(bl->anticL_in, 0, 1);
2007 pos = rbitset_next(bl->anticL_in, pos + 1, 1)) {
2008 memop_t *op = bl->id_2_memop_antic[pos];
2009 int have_some, all_same;
2012 if (rbitset_is_set(env.curr_set, pos)) {
2017 assert(is_Load(op->node));
2019 DB((dbg, LEVEL_3, "anticipated %+F\n", op->node));
2024 for (i = n - 1; i >= 0; --i) {
2025 ir_node *pred = get_Block_cfgpred_block(block, i);
2026 block_t *pred_bl = get_block_entry(pred);
2027 memop_t *e = find_address_avail(pred_bl, &op->value);
2028 ir_mode *mode = op->value.mode;
2031 ir_node *ef_block = get_effective_block(op->value.address, block, i);
2032 if (! block_dominates(ef_block, pred)) {
2033 /* cannot place a copy here */
2035 DB((dbg, LEVEL_3, "%+F is cannot be moved into predecessor %+F\n", op->node, pred));
2038 DB((dbg, LEVEL_3, "%+F is not available in predecessor %+F\n", op->node, pred));
2039 pred_bl->avail = NULL;
2042 if (e->value.mode != mode && !can_convert_to(e->value.mode, mode)) {
2043 /* cannot create a Phi due to different modes */
2049 DB((dbg, LEVEL_3, "%+F is available for %+F in predecessor %+F\n", e->node, op->node, pred));
2052 else if (first != e->node)
2056 if (have_some && !all_same) {
2057 ir_mode *mode = op->value.mode;
2061 NEW_ARR_A(ir_node *, in, n);
2063 for (i = n - 1; i >= 0; --i) {
2064 ir_node *pred = get_Block_cfgpred_block(block, i);
2065 block_t *pred_bl = get_block_entry(pred);
2067 if (pred_bl->avail == NULL) {
2068 /* create a new Load here and make to make it fully redundant */
2069 dbg_info *db = get_irn_dbg_info(op->node);
2070 ir_node *last_mem = find_last_memory(pred_bl);
2071 ir_node *load, *def, *adr;
2074 assert(last_mem != NULL);
2075 adr = phi_translate(op->value.address, block, i);
2076 load = new_rd_Load(db, current_ir_graph, pred, last_mem, adr, mode, cons_none);
2077 def = new_r_Proj(current_ir_graph, pred, load, mode, pn_Load_res);
2078 DB((dbg, LEVEL_1, "Created new %+F in %+F for party redundant %+F\n", load, pred, op->node));
2080 new_op = alloc_memop(load);
2081 new_op->mem = new_r_Proj(current_ir_graph, pred, load, mode_M, pn_Load_M);
2082 new_op->value.address = adr;
2083 new_op->value.id = op->value.id;
2084 new_op->value.mode = mode;
2085 new_op->value.value = def;
2087 new_op->projs[pn_Load_M] = new_op->mem;
2088 new_op->projs[pn_Load_res] = def;
2090 new_op->prev = pred_bl->memop_backward;
2091 if (pred_bl->memop_backward != NULL)
2092 pred_bl->memop_backward->next = new_op;
2094 pred_bl->memop_backward = new_op;
2096 if (pred_bl->memop_forward == NULL)
2097 pred_bl->memop_forward = new_op;
2099 if (get_nodes_block(last_mem) == pred) {
2100 /* We have add a new last memory op in pred block.
2101 If pred had already a last mem, reroute all memory
2103 reroute_all_mem_users(last_mem, new_op->mem);
2105 /* reroute only those memory going through the pre block */
2106 reroute_mem_through(last_mem, new_op->mem, pred);
2109 /* we added this load at the end, so it will be avail anyway */
2110 add_memop_avail(pred_bl, new_op);
2111 pred_bl->avail = new_op;
2113 in[i] = conv_to(pred_bl->avail->value.value, mode);
2115 phi = new_r_Phi(current_ir_graph, block, n, in, mode);
2116 DB((dbg, LEVEL_1, "Created new %+F in %+F for now redundant %+F\n", phi, block, op->node));
2118 phi_op = clone_memop_phi(op, phi);
2124 /* recalculate avail by gen and kill */
2125 calc_gen_kill_avail(bl);
2127 /* always update the map after gen/kill, as values might have been changed due to RAR/WAR/WAW */
2128 memcpy(bl->id_2_memop_avail, env.curr_id_2_memop, env.rbs_size * sizeof(env.curr_id_2_memop[0]));
2130 if (!rbitset_equal(bl->avail_out, env.curr_set, env.rbs_size)) {
2131 /* the avail set has changed */
2132 rbitset_cpy(bl->avail_out, env.curr_set, env.rbs_size);
2133 dump_curr(bl, "Avail_out*");
2136 dump_curr(bl, "Avail_out");
2141 * Insert Loads upwards.
2143 static void insert_Loads_upwards(void) {
2147 /* recalculate antic_out and insert Loads */
2148 DB((dbg, LEVEL_2, "Inserting Loads\n"));
2152 DB((dbg, LEVEL_2, "Iteration %d:\n=========\n", i));
2156 /* over all blocks in reverse post order, skip the start block */
2157 for (bl = env.forward->forward_next; bl != NULL; bl = bl->forward_next) {
2158 need_iter |= insert_Load(bl);
2161 } while (need_iter);
2163 DB((dbg, LEVEL_2, "Finished Load inserting after %d iterations\n", i));
2164 } /* insert_Loads_upwards */
2167 * Kill unreachable control flow.
2169 * @param irg the graph to operate on
2171 static void kill_unreachable_blocks(ir_graph *irg) {
2176 NEW_ARR_A(ir_node *, ins, env.max_cfg_preds);
2178 for (bl = env.forward; bl != NULL; bl = bl->forward_next) {
2179 ir_node *block = bl->block;
2182 assert(get_Block_mark(block));
2184 n = get_Block_n_cfgpreds(block);
2186 for (i = j = 0; i < n; ++i) {
2187 ir_node *pred = get_Block_cfgpred(block, i);
2193 pred_bl = get_nodes_block(skip_Proj(pred));
2194 if (! get_Block_mark(pred_bl))
2200 ir_node *phi, *next;
2202 /* some unreachable blocks detected */
2205 DB((dbg, LEVEL_1, "Killing dead block predecessors on %+F\n", block));
2207 set_irn_in(block, j, ins);
2209 /* shorten all Phi nodes */
2210 for (phi = get_Block_phis(block); phi != NULL; phi = next) {
2211 next = get_Phi_next(phi);
2213 for (i = k = 0; i < n; ++i) {
2214 ir_node *pred = get_Block_cfgpred_block(block, i);
2219 if (! get_Block_mark(pred))
2222 ins[k++] = get_Phi_pred(phi, i);
2225 exchange(phi, ins[0]);
2227 set_irn_in(phi, k, ins);
2234 /* kick keep alives */
2235 ir_node *end = get_irg_end(irg);
2236 int i, j, n = get_End_n_keepalives(end);
2238 NEW_ARR_A(ir_node *, ins, n);
2240 for (i = j = 0; i < n; ++i) {
2241 ir_node *ka = get_End_keepalive(end, i);
2249 ka_bl = get_nodes_block(skip_Proj(ka));
2250 if (get_Block_mark(ka_bl))
2254 set_End_keepalives(end, j, ins);
2258 /* this transformation do NOT invalidate the dominance */
2260 } /* kill_unreachable_blocks */
2262 int opt_ldst(ir_graph *irg) {
2264 ir_graph *rem = current_ir_graph;
2266 current_ir_graph = irg;
2268 FIRM_DBG_REGISTER(dbg, "firm.opt.ldst");
2269 // firm_dbg_set_mask(dbg, -1);
2271 DB((dbg, LEVEL_1, "\nDoing Load/Store optimization on %+F\n", irg));
2273 /* we need landing pads */
2274 remove_critical_cf_edges(irg);
2276 // dump_ir_block_graph(irg, "-XXX");
2278 if (get_opt_alias_analysis()) {
2279 assure_irg_entity_usage_computed(irg);
2280 assure_irp_globals_entity_usage_computed();
2283 obstack_init(&env.obst);
2284 ir_nodemap_init(&env.adr_map);
2287 env.backward = NULL;
2288 env.curr_adr_id = 0;
2290 env.max_cfg_preds = 0;
2292 env.start_bl = get_irg_start_block(irg);
2293 env.end_bl = get_irg_end_block(irg);
2294 #ifdef DEBUG_libfirm
2295 env.id_2_address = NEW_ARR_F(ir_node *, 0);
2298 assure_irg_outs(irg);
2300 ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK | IR_RESOURCE_BLOCK_MARK);
2302 /* first step: allocate block entries. Note that some blocks might be
2303 unreachable here. Using the normal walk ensures that ALL blocks are initialized. */
2304 irg_walk_graph(irg, prepare_blocks, link_phis, NULL);
2306 /* produce an inverse post-order list for the CFG: this links only reachable
2308 irg_out_block_walk(get_irg_start_block(irg), NULL, inverse_post_order, NULL);
2310 if (! get_Block_mark(env.end_bl)) {
2312 * The end block is NOT reachable due to endless loops
2313 * or no_return calls.
2314 * Place the end block last.
2315 * env.backward points to the last block in the list for this purpose.
2317 env.backward->forward_next = get_block_entry(env.end_bl);
2319 set_Block_mark(env.end_bl, 1);
2322 /* KILL unreachable blocks: these disturb the data flow analysis */
2323 kill_unreachable_blocks(irg);
2327 /* second step: find and sort all memory ops */
2328 walk_memory_irg(irg, collect_memops, NULL, NULL);
2330 #ifdef DEBUG_libfirm
2331 /* check that the backward map is correct */
2332 assert((unsigned)ARR_LEN(env.id_2_address) == env.curr_adr_id);
2335 if (env.n_mem_ops == 0) {
2340 /* create the backward links. */
2341 env.backward = NULL;
2342 irg_block_walk_graph(irg, NULL, collect_backward, NULL);
2344 /* link the end block in */
2345 bl = get_block_entry(env.end_bl);
2346 bl->backward_next = env.backward;
2349 /* check that we really start with the start / end block */
2350 assert(env.forward->block == env.start_bl);
2351 assert(env.backward->block == env.end_bl);
2353 /* create address sets: for now, only the existing addresses are allowed plus one
2354 needed for the sentinel */
2355 env.rbs_size = env.n_mem_ops + 1;
2357 /* create the current set */
2358 env.curr_set = rbitset_obstack_alloc(&env.obst, env.rbs_size);
2359 rbitset_set(env.curr_set, env.rbs_size - 1);
2360 env.curr_id_2_memop = NEW_ARR_D(memop_t *, &env.obst, env.rbs_size);
2361 memset(env.curr_id_2_memop, 0, env.rbs_size * sizeof(env.curr_id_2_memop[0]));
2363 for (bl = env.forward; bl != NULL; bl = bl->forward_next) {
2364 /* set sentinel bits */
2365 bl->avail_out = rbitset_obstack_alloc(&env.obst, env.rbs_size);
2366 rbitset_set(bl->avail_out, env.rbs_size - 1);
2368 bl->id_2_memop_avail = NEW_ARR_D(memop_t *, &env.obst, env.rbs_size);
2369 memset(bl->id_2_memop_avail, 0, env.rbs_size * sizeof(bl->id_2_memop_avail[0]));
2371 bl->anticL_in = rbitset_obstack_alloc(&env.obst, env.rbs_size);
2372 rbitset_set(bl->anticL_in, env.rbs_size - 1);
2374 bl->id_2_memop_antic = NEW_ARR_D(memop_t *, &env.obst, env.rbs_size);
2375 memset(bl->id_2_memop_antic, 0, env.rbs_size * sizeof(bl->id_2_memop_antic[0]));
2378 // dump_block_list(&env);
2383 insert_Loads_upwards();
2386 /* over all blocks in reverse post order */
2387 for (bl = env.forward; bl != NULL; bl = bl->forward_next) {
2388 do_replacements(bl);
2391 /* not only invalidate but free them. We might allocate new out arrays
2392 on our obstack which will be deleted yet. */
2394 set_irg_entity_usage_state(irg, ir_entity_usage_not_computed);
2398 ir_free_resources(irg, IR_RESOURCE_IRN_LINK | IR_RESOURCE_BLOCK_MARK);
2399 ir_nodemap_destroy(&env.adr_map);
2400 obstack_free(&env.obst, NULL);
2402 // dump_ir_block_graph(irg, "-YYY");
2404 #ifdef DEBUG_libfirm
2405 DEL_ARR_F(env.id_2_address);
2408 current_ir_graph = rem;
2409 return env.changed != 0;