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"
47 /* maximum number of output Proj's */
48 #define MAX_PROJ (pn_Load_max > pn_Store_max ? pn_Load_max : pn_Store_max)
51 * Mapping an address to an dense ID.
53 typedef struct address_entry_t {
54 unsigned id; /**< The ID */
61 FLAG_KILL_ALL = 1, /**< KILL all addresses */
62 FLAG_KILLED_NODE = 2, /**< this node was killed */
63 FLAG_EXCEPTION = 4, /**< this node has exception flow */
64 FLAG_IGNORE = 8, /**< ignore this node (volatile or other) */
68 * A value: This represents a value stored at a given address in
69 * memory. Do not confuse with values from value numbering.
71 typedef struct value_t value_t;
73 ir_node *address; /**< the address of this value */
74 ir_node *value; /**< the value itself */
75 ir_mode *mode; /**< the mode of the value */
76 unsigned id; /**< address id */
80 * A memop describes an memory-related operation.
81 * These are Loads/Store and all other ops that might modify
82 * memory (Calls, CopyB) or causing exceptions.
84 typedef struct memop_t memop_t;
86 value_t value; /**< the value of this memop: only defined for Load/Store */
87 ir_node *node; /**< the memory op itself */
88 ir_node *mem; /**< the memory FROM this node */
89 ir_node *replace; /**< the replacement node if this memop is replaced */
90 memop_t *next; /**< links to the next memory op in the block in forward order. */
91 memop_t *prev; /**< links to the previous memory op in the block in forward order. */
92 unsigned flags; /**< memop flags */
93 ir_node *projs[MAX_PROJ]; /**< Projs of this memory op */
97 * Additional data for every basic block.
99 typedef struct block_t block_t;
101 memop_t *memop_forward; /**< topologically sorted list of memory ops in this block */
102 memop_t *memop_backward; /**< last memop in the list */
103 unsigned *avail_out; /**< out-set of available addresses */
104 memop_t **id_2_memop_avail; /**< maps avail address ids to memops */
105 unsigned *anticL_in; /**< in-set of anticipated Load addresses */
106 memop_t **id_2_memop_antic; /**< maps anticipated address ids to memops */
107 ir_node *block; /**< the associated block */
108 block_t *forward_next; /**< next block entry for forward iteration */
109 block_t *backward_next; /**< next block entry for backward iteration */
110 memop_t *avail; /**< used locally for the avail map */
111 memop_t **trans_results; /**< used to cached translated nodes due antic calculation. */
115 * Metadata for this pass.
117 typedef struct ldst_env_t {
118 struct obstack obst; /**< obstack for temporary data */
119 ir_nodemap_t adr_map; /**< Map addresses to */
120 block_t *forward; /**< Inverse post-order list of all blocks Start->End */
121 block_t *backward; /**< Inverse post-order list of all blocks End->Start */
122 ir_node *start_bl; /**< start block of the current graph */
123 ir_node *end_bl; /**< end block of the current graph */
124 unsigned *curr_set; /**< current set of addresses */
125 memop_t **curr_id_2_memop; /**< current map of address ids to memops */
126 unsigned curr_adr_id; /**< number for address mapping */
127 unsigned n_mem_ops; /**< number of memory operations (Loads/Stores) */
128 unsigned rbs_size; /**< size of all bitsets in bytes */
129 int max_cfg_preds; /**< maximum number of block cfg predecessors */
130 int changed; /**< Flags for changed graph state */
132 ir_node **id_2_address; /**< maps an id to the used address */
136 /* the one and only environment */
141 static firm_dbg_module_t *dbg;
144 * Dumps the block list.
146 * @param ldst environment
148 static void dump_block_list(ldst_env *env)
154 for (entry = env->forward; entry != NULL; entry = entry->forward_next) {
155 DB((dbg, LEVEL_2, "%+F {", entry->block));
158 for (op = entry->memop_forward; op != NULL; op = op->next) {
160 DB((dbg, LEVEL_2, "\n\t"));
161 } DB((dbg, LEVEL_2, "%+F", op->node));
162 if ((op->flags & FLAG_KILL_ALL) == FLAG_KILL_ALL)
163 DB((dbg, LEVEL_2, "X"));
164 else if (op->flags & FLAG_KILL_ALL)
165 DB((dbg, LEVEL_2, "K"));
166 DB((dbg, LEVEL_2, ", "));
170 DB((dbg, LEVEL_2, "\n}\n\n"));
172 } /* dump_block_list */
175 * Dumps the current set.
177 * @param bl current block
178 * @param s name of the set
180 static void dump_curr(block_t *bl, const char *s)
182 unsigned end = env.rbs_size - 1;
186 DB((dbg, LEVEL_2, "%s[%+F] = {", s, bl->block));
188 for (pos = rbitset_next(env.curr_set, 0, 1); pos < end; pos = rbitset_next(env.curr_set, pos + 1, 1)) {
189 memop_t *op = env.curr_id_2_memop[pos];
192 DB((dbg, LEVEL_2, "\n\t"));
195 DB((dbg, LEVEL_2, "<%+F, %+F>, ", op->value.address, op->value.value));
198 DB((dbg, LEVEL_2, "\n}\n"));
202 static void dump_block_list(ldst_env *env)
206 static void dump_curr(block_t *bl, const char *s)
211 #endif /* DEBUG_libfirm */
213 /** Get the block entry for a block node */
214 static block_t *get_block_entry(const ir_node *block)
216 assert(is_Block(block));
218 return get_irn_link(block);
219 } /* get_block_entry */
221 /** Get the memop entry for a memory operation node */
222 static memop_t *get_irn_memop(const ir_node *irn)
224 assert(! is_Block(irn));
225 return get_irn_link(irn);
226 } /* get_irn_memop */
229 * Walk over the memory edges from definition to users.
230 * This ensures, that even operation without memory output are found.
232 * @param irn start node
233 * @param pre pre walker function
234 * @param post post walker function
235 * @param ctx context parameter for the walker functions
237 static void walk_memory(ir_node *irn, irg_walk_func *pre, irg_walk_func *post, void *ctx)
242 mark_irn_visited(irn);
247 mode = get_irn_mode(irn);
248 if (mode == mode_M) {
249 /* every successor uses memory */
250 for (i = get_irn_n_outs(irn) - 1; i >= 0; --i) {
251 ir_node *succ = get_irn_out(irn, i);
253 if (! irn_visited(succ))
254 walk_memory(succ, pre, post, ctx);
256 } else if (mode == mode_T) {
257 /* only some Proj's uses memory */
258 for (i = get_irn_n_outs(irn) - 1; i >= 0; --i) {
259 ir_node *proj = get_irn_out(irn, i);
261 if (get_irn_mode(proj) == mode_M && ! irn_visited(proj))
262 walk_memory(proj, pre, post, ctx);
270 * Walks over all memory nodes of a graph.
273 * @param pre pre walker function
274 * @param post post walker function
275 * @param ctx context parameter for the walker functions
277 static void walk_memory_irg(ir_graph *irg, irg_walk_func pre, irg_walk_func post, void *ctx)
279 inc_irg_visited(irg);
281 ir_reserve_resources(irg, IR_RESOURCE_IRN_VISITED);
284 * there are two possible sources for memory: initial_mem and nomem
285 * we ignore nomem as this should NOT change the memory
287 walk_memory(get_irg_initial_mem(irg), pre, post, ctx);
289 ir_free_resources(irg, IR_RESOURCE_IRN_VISITED);
290 } /* walk_memory_irg */
293 * Register an address and allocate a (sparse, 0..n) ID for it.
295 * @param adr the IR-node representing the address
297 * @return the allocated id
299 static unsigned register_address(ir_node *adr)
301 address_entry *entry;
303 /* skip Confirms and Casts */
305 if (is_Confirm(adr)) {
306 adr = get_Confirm_value(adr);
310 adr = get_Cast_op(adr);
314 entry = ir_nodemap_get(&env.adr_map, adr);
318 entry = OALLOC(&env.obst, address_entry);
320 entry->id = env.curr_adr_id++;
321 ir_nodemap_insert(&env.adr_map, adr, entry);
323 DB((dbg, LEVEL_3, "ADDRESS %+F has ID %u\n", adr, entry->id));
325 ARR_APP1(ir_node *, env.id_2_address, adr);
329 } /* register_address */
333 * translate an address through a Phi node into a given predecessor
336 * @param address the address
337 * @param block the block
338 * @param pos the position of the predecessor in block
340 static ir_node *phi_translate(ir_node *address, const ir_node *block, int pos)
342 if (is_Phi(address) && get_nodes_block(address) == block)
343 address = get_Phi_pred(address, pos);
345 } /* phi_translate */
348 * Walker: allocate an block entry for every block
349 * and register all potential addresses.
351 static void prepare_blocks(ir_node *irn, void *ctx)
356 block_t *entry = OALLOC(&env.obst, block_t);
359 entry->memop_forward = NULL;
360 entry->memop_backward = NULL;
361 entry->avail_out = NULL;
362 entry->id_2_memop_avail = NULL;
363 entry->anticL_in = NULL;
364 entry->id_2_memop_antic = NULL;
366 entry->forward_next = NULL;
367 entry->backward_next = NULL;
369 entry->trans_results = NULL;
370 set_irn_link(irn, entry);
372 set_Block_phis(irn, NULL);
374 /* use block marks to track unreachable blocks */
375 set_Block_mark(irn, 0);
377 n = get_Block_n_cfgpreds(irn);
378 if (n > env.max_cfg_preds)
379 env.max_cfg_preds = n;
381 ir_mode *mode = get_irn_mode(irn);
383 if (mode_is_reference(mode)) {
385 * Register ALL possible addresses: this is overkill yet but
386 * simpler then doing it for all possible translated addresses
387 * (which would be sufficient in the moment.
389 (void)register_address(irn);
392 } /* prepare_blocks */
395 * Post-Walker, link in all Phi's
397 static void link_phis(ir_node *irn, void *ctx)
402 ir_node *block = get_nodes_block(irn);
403 add_Block_phi(block, irn);
408 * Block walker: creates the inverse post-order list for the CFG.
410 static void inverse_post_order(ir_node *block, void *ctx)
412 block_t *entry = get_block_entry(block);
416 /* mark this block IS reachable from start */
417 set_Block_mark(block, 1);
419 /* create the list in inverse order */
420 entry->forward_next = env.forward;
423 /* remember the first visited (last in list) entry, needed for later */
424 if (env.backward == NULL)
425 env.backward = entry;
426 } /* inverse_post_order */
429 * Block walker: create backward links for the memops of a block.
431 static void collect_backward(ir_node *block, void *ctx)
433 block_t *entry = get_block_entry(block);
439 * Do NOT link in the end block yet. We want it to be
440 * the first in the list. This is NOT guaranteed by the walker
441 * if we have endless loops.
443 if (block != env.end_bl) {
444 entry->backward_next = env.backward;
446 /* create the list in inverse order */
447 env.backward = entry;
450 /* create backward links for all memory ops */
452 for (op = entry->memop_forward; op != NULL; op = op->next) {
456 entry->memop_backward = last;
457 } /* collect_backward */
462 * @param irn the IR-node representing the memop or NULL
463 * if this is a translated (virtual) memop
465 * @return the allocated memop
467 static memop_t *alloc_memop(ir_node *irn)
469 memop_t *m = OALLOC(&env.obst, memop_t);
471 m->value.address = NULL;
472 m->value.value = NULL;
473 m->value.mode = NULL;
481 memset(m->projs, 0, sizeof(m->projs));
484 set_irn_link(irn, m);
489 * Create a memop for a Phi-replacement.
491 * @param op the memop to clone
492 * @param phi the Phi-node representing the new value
494 static memop_t *clone_memop_phi(memop_t *op, ir_node *phi)
496 memop_t *m = OALLOC(&env.obst, memop_t);
498 m->value = op->value;
499 m->value.value = phi;
506 set_irn_link(phi, m);
508 } /* clone_memop_phi */
511 * Return the memory properties of a call node.
513 * @param call the call node
515 * return a bitset of mtp_property_const and mtp_property_pure
517 static unsigned get_Call_memory_properties(ir_node *call)
519 ir_type *call_tp = get_Call_type(call);
520 unsigned prop = get_method_additional_properties(call_tp);
522 /* check first the call type */
523 if ((prop & (mtp_property_const|mtp_property_pure)) == 0) {
524 /* try the called entity */
525 ir_node *ptr = get_Call_ptr(call);
527 if (is_Global(ptr)) {
528 ir_entity *ent = get_Global_entity(ptr);
530 prop = get_entity_additional_properties(ent);
533 return prop & (mtp_property_const|mtp_property_pure);
534 } /* get_Call_memory_properties */
537 * Returns an entity if the address ptr points to a constant one.
539 * @param ptr the address
541 * @return an entity or NULL
543 static ir_entity *find_constant_entity(ir_node *ptr)
546 if (is_SymConst(ptr) && get_SymConst_kind(ptr) == symconst_addr_ent) {
547 return get_SymConst_entity(ptr);
548 } else if (is_Sel(ptr)) {
549 ir_entity *ent = get_Sel_entity(ptr);
550 ir_type *tp = get_entity_owner(ent);
552 /* Do not fiddle with polymorphism. */
553 if (is_Class_type(get_entity_owner(ent)) &&
554 ((get_entity_n_overwrites(ent) != 0) ||
555 (get_entity_n_overwrittenby(ent) != 0) ) )
558 if (is_Array_type(tp)) {
562 for (i = 0, n = get_Sel_n_indexs(ptr); i < n; ++i) {
564 tarval *tlower, *tupper;
565 ir_node *index = get_Sel_index(ptr, i);
566 tarval *tv = computed_value(index);
568 /* check if the index is constant */
569 if (tv == tarval_bad)
572 bound = get_array_lower_bound(tp, i);
573 tlower = computed_value(bound);
574 bound = get_array_upper_bound(tp, i);
575 tupper = computed_value(bound);
577 if (tlower == tarval_bad || tupper == tarval_bad)
580 if (tarval_cmp(tv, tlower) & pn_Cmp_Lt)
582 if (tarval_cmp(tupper, tv) & pn_Cmp_Lt)
585 /* ok, bounds check finished */
589 if (get_entity_linkage(ent) == IR_LINKAGE_CONSTANT)
593 ptr = get_Sel_ptr(ptr);
594 } else if (is_Add(ptr)) {
595 ir_node *l = get_Add_left(ptr);
596 ir_node *r = get_Add_right(ptr);
598 if (get_irn_mode(l) == get_irn_mode(ptr) && is_Const(r))
600 else if (get_irn_mode(r) == get_irn_mode(ptr) && is_Const(l))
605 /* for now, we support only one addition, reassoc should fold all others */
606 if (! is_SymConst(ptr) && !is_Sel(ptr))
608 } else if (is_Sub(ptr)) {
609 ir_node *l = get_Sub_left(ptr);
610 ir_node *r = get_Sub_right(ptr);
612 if (get_irn_mode(l) == get_irn_mode(ptr) && is_Const(r))
616 /* for now, we support only one subtraction, reassoc should fold all others */
617 if (! is_SymConst(ptr) && !is_Sel(ptr))
622 } /* find_constant_entity */
625 * Return the Selection index of a Sel node from dimension n
627 static long get_Sel_array_index_long(ir_node *n, int dim)
629 ir_node *index = get_Sel_index(n, dim);
630 assert(is_Const(index));
631 return get_tarval_long(get_Const_tarval(index));
632 } /* get_Sel_array_index_long */
635 * Returns the accessed component graph path for an
636 * node computing an address.
638 * @param ptr the node computing the address
639 * @param depth current depth in steps upward from the root
642 static compound_graph_path *rec_get_accessed_path(ir_node *ptr, int depth)
644 compound_graph_path *res = NULL;
645 ir_entity *root, *field, *ent;
646 int path_len, pos, idx;
650 if (is_SymConst(ptr)) {
651 /* a SymConst. If the depth is 0, this is an access to a global
652 * entity and we don't need a component path, else we know
653 * at least its length.
655 assert(get_SymConst_kind(ptr) == symconst_addr_ent);
656 root = get_SymConst_entity(ptr);
657 res = (depth == 0) ? NULL : new_compound_graph_path(get_entity_type(root), depth);
658 } else if (is_Sel(ptr)) {
659 /* it's a Sel, go up until we find the root */
660 res = rec_get_accessed_path(get_Sel_ptr(ptr), depth+1);
664 /* fill up the step in the path at the current position */
665 field = get_Sel_entity(ptr);
666 path_len = get_compound_graph_path_length(res);
667 pos = path_len - depth - 1;
668 set_compound_graph_path_node(res, pos, field);
670 if (is_Array_type(get_entity_owner(field))) {
671 assert(get_Sel_n_indexs(ptr) == 1 && "multi dim arrays not implemented");
672 set_compound_graph_path_array_index(res, pos, get_Sel_array_index_long(ptr, 0));
674 } else if (is_Add(ptr)) {
675 ir_node *l = get_Add_left(ptr);
676 ir_node *r = get_Add_right(ptr);
677 ir_mode *mode = get_irn_mode(ptr);
680 if (is_Const(r) && get_irn_mode(l) == mode) {
682 tv = get_Const_tarval(r);
685 tv = get_Const_tarval(l);
688 mode = get_tarval_mode(tv);
691 /* ptr must be a Sel or a SymConst, this was checked in find_constant_entity() */
693 field = get_Sel_entity(ptr);
695 field = get_SymConst_entity(ptr);
698 for (ent = field;;) {
700 tarval *sz, *tv_index, *tlower, *tupper;
703 tp = get_entity_type(ent);
704 if (! is_Array_type(tp))
706 ent = get_array_element_entity(tp);
707 size = get_type_size_bytes(get_entity_type(ent));
708 sz = new_tarval_from_long(size, mode);
710 tv_index = tarval_div(tmp, sz);
711 tmp = tarval_mod(tmp, sz);
713 if (tv_index == tarval_bad || tmp == tarval_bad)
716 assert(get_array_n_dimensions(tp) == 1 && "multiarrays not implemented");
717 bound = get_array_lower_bound(tp, 0);
718 tlower = computed_value(bound);
719 bound = get_array_upper_bound(tp, 0);
720 tupper = computed_value(bound);
722 if (tlower == tarval_bad || tupper == tarval_bad)
725 if (tarval_cmp(tv_index, tlower) & pn_Cmp_Lt)
727 if (tarval_cmp(tupper, tv_index) & pn_Cmp_Lt)
730 /* ok, bounds check finished */
733 if (! tarval_is_null(tmp)) {
734 /* access to some struct/union member */
738 /* should be at least ONE array */
742 res = rec_get_accessed_path(ptr, depth + idx);
746 path_len = get_compound_graph_path_length(res);
747 pos = path_len - depth - idx;
749 for (ent = field;;) {
751 tarval *sz, *tv_index;
754 tp = get_entity_type(ent);
755 if (! is_Array_type(tp))
757 ent = get_array_element_entity(tp);
758 set_compound_graph_path_node(res, pos, ent);
760 size = get_type_size_bytes(get_entity_type(ent));
761 sz = new_tarval_from_long(size, mode);
763 tv_index = tarval_div(tv, sz);
764 tv = tarval_mod(tv, sz);
766 /* worked above, should work again */
767 assert(tv_index != tarval_bad && tv != tarval_bad);
769 /* bounds already checked above */
770 index = get_tarval_long(tv_index);
771 set_compound_graph_path_array_index(res, pos, index);
774 } else if (is_Sub(ptr)) {
775 ir_node *l = get_Sub_left(ptr);
776 ir_node *r = get_Sub_right(ptr);
779 tv = get_Const_tarval(r);
784 } /* rec_get_accessed_path */
787 * Returns an access path or NULL. The access path is only
788 * valid, if the graph is in phase_high and _no_ address computation is used.
790 static compound_graph_path *get_accessed_path(ir_node *ptr)
792 compound_graph_path *gr = rec_get_accessed_path(ptr, 0);
794 } /* get_accessed_path */
796 typedef struct path_entry {
798 struct path_entry *next;
802 static ir_node *rec_find_compound_ent_value(ir_node *ptr, path_entry *next)
804 path_entry entry, *p;
805 ir_entity *ent, *field;
806 ir_initializer_t *initializer;
812 if (is_SymConst(ptr)) {
814 ent = get_SymConst_entity(ptr);
815 initializer = get_entity_initializer(ent);
816 for (p = next; p != NULL;) {
817 if (initializer->kind != IR_INITIALIZER_COMPOUND)
819 n = get_initializer_compound_n_entries(initializer);
820 tp = get_entity_type(ent);
822 if (is_Array_type(tp)) {
823 ent = get_array_element_entity(tp);
828 initializer = get_initializer_compound_value(initializer, 0);
832 if (p->index >= (int) n)
834 initializer = get_initializer_compound_value(initializer, p->index);
839 tp = get_entity_type(ent);
840 while (is_Array_type(tp)) {
841 ent = get_array_element_entity(tp);
842 tp = get_entity_type(ent);
844 n = get_initializer_compound_n_entries(initializer);
847 initializer = get_initializer_compound_value(initializer, 0);
850 switch (initializer->kind) {
851 case IR_INITIALIZER_CONST:
852 return get_initializer_const_value(initializer);
853 case IR_INITIALIZER_TARVAL:
854 case IR_INITIALIZER_NULL:
858 } else if (is_Sel(ptr)) {
859 entry.ent = field = get_Sel_entity(ptr);
860 tp = get_entity_owner(field);
861 if (is_Array_type(tp)) {
862 assert(get_Sel_n_indexs(ptr) == 1 && "multi dim arrays not implemented");
863 entry.index = get_Sel_array_index_long(ptr, 0) - get_array_lower_bound_int(tp, 0);
865 int i, n_members = get_compound_n_members(tp);
866 for (i = 0; i < n_members; ++i) {
867 if (get_compound_member(tp, i) == field)
870 if (i >= n_members) {
871 /* not found: should NOT happen */
876 return rec_find_compound_ent_value(get_Sel_ptr(ptr), &entry);
877 } else if (is_Add(ptr)) {
878 ir_node *l = get_Add_left(ptr);
879 ir_node *r = get_Add_right(ptr);
885 tv = get_Const_tarval(r);
888 tv = get_Const_tarval(l);
891 mode = get_tarval_mode(tv);
893 /* ptr must be a Sel or a SymConst, this was checked in find_constant_entity() */
895 field = get_Sel_entity(ptr);
897 field = get_SymConst_entity(ptr);
900 /* count needed entries */
902 for (ent = field;;) {
903 tp = get_entity_type(ent);
904 if (! is_Array_type(tp))
906 ent = get_array_element_entity(tp);
909 /* should be at least ONE entry */
913 /* allocate the right number of entries */
914 NEW_ARR_A(path_entry, p, pos);
918 for (ent = field;;) {
920 tarval *sz, *tv_index, *tlower, *tupper;
924 tp = get_entity_type(ent);
925 if (! is_Array_type(tp))
927 ent = get_array_element_entity(tp);
929 p[pos].next = &p[pos + 1];
931 size = get_type_size_bytes(get_entity_type(ent));
932 sz = new_tarval_from_long(size, mode);
934 tv_index = tarval_div(tv, sz);
935 tv = tarval_mod(tv, sz);
937 if (tv_index == tarval_bad || tv == tarval_bad)
940 assert(get_array_n_dimensions(tp) == 1 && "multiarrays not implemented");
941 bound = get_array_lower_bound(tp, 0);
942 tlower = computed_value(bound);
943 bound = get_array_upper_bound(tp, 0);
944 tupper = computed_value(bound);
946 if (tlower == tarval_bad || tupper == tarval_bad)
949 if (tarval_cmp(tv_index, tlower) & pn_Cmp_Lt)
951 if (tarval_cmp(tupper, tv_index) & pn_Cmp_Lt)
954 /* ok, bounds check finished */
955 index = get_tarval_long(tv_index);
956 p[pos].index = index;
959 if (! tarval_is_null(tv)) {
960 /* hmm, wrong access */
963 p[pos - 1].next = next;
964 return rec_find_compound_ent_value(ptr, p);
965 } else if (is_Sub(ptr)) {
966 ir_node *l = get_Sub_left(ptr);
967 ir_node *r = get_Sub_right(ptr);
970 tv = get_Const_tarval(r);
975 } /* rec_find_compound_ent_value */
977 static ir_node *find_compound_ent_value(ir_node *ptr)
979 return rec_find_compound_ent_value(ptr, NULL);
980 } /* find_compound_ent_value */
983 * Mark a Load memop to be replace by a definition
985 * @param op the Load memop
987 static void mark_replace_load(memop_t *op, ir_node *def)
990 op->flags |= FLAG_KILLED_NODE;
992 } /* mark_replace_load */
995 * Mark a Store memop to be removed.
997 * @param op the Store memop
999 static void mark_remove_store(memop_t *op)
1001 op->flags |= FLAG_KILLED_NODE;
1003 } /* mark_remove_store */
1006 * Update a memop for a Load.
1008 * @param m the memop
1010 static void update_Load_memop(memop_t *m)
1013 ir_node *load = m->node;
1017 if (get_Load_volatility(load) == volatility_is_volatile)
1018 m->flags |= FLAG_IGNORE;
1020 ptr = get_Load_ptr(load);
1022 m->value.address = ptr;
1024 for (i = get_irn_n_outs(load) - 1; i >= 0; --i) {
1025 ir_node *proj = get_irn_out(load, i);
1028 /* beware of keep edges */
1032 pn = get_Proj_proj(proj);
1033 m->projs[pn] = proj;
1036 m->value.value = proj;
1037 m->value.mode = get_irn_mode(proj);
1039 case pn_Load_X_except:
1040 m->flags |= FLAG_EXCEPTION;
1045 case pn_Load_X_regular:
1048 panic("Unsupported Proj from Load %+F", proj);
1052 /* check if we can determine the entity that will be loaded */
1053 ent = find_constant_entity(ptr);
1055 if (ent != NULL && get_entity_visibility(ent) != ir_visibility_external) {
1056 /* a static allocation that is not external: there should be NO exception
1057 * when loading even if we cannot replace the load itself. */
1058 ir_node *value = NULL;
1060 /* no exception, clear the m fields as it might be checked later again */
1061 if (m->projs[pn_Load_X_except]) {
1062 exchange(m->projs[pn_Load_X_except], new_Bad());
1063 m->projs[pn_Load_X_except] = NULL;
1064 m->flags &= ~FLAG_EXCEPTION;
1067 if (m->projs[pn_Load_X_regular]) {
1068 exchange(m->projs[pn_Load_X_regular], new_r_Jmp(get_nodes_block(load)));
1069 m->projs[pn_Load_X_regular] = NULL;
1073 if (get_entity_linkage(ent) & IR_LINKAGE_CONSTANT) {
1074 if (ent->initializer) {
1075 /* new style initializer */
1076 value = find_compound_ent_value(ptr);
1077 } else if (entity_has_compound_ent_values(ent)) {
1078 /* old style initializer */
1079 compound_graph_path *path = get_accessed_path(ptr);
1082 assert(is_proper_compound_graph_path(path, get_compound_graph_path_length(path)-1));
1084 value = get_compound_ent_value_by_path(ent, path);
1085 DB((dbg, LEVEL_1, " Constant access at %F%F resulted in %+F\n", ent, path, value));
1086 free_compound_graph_path(path);
1090 value = can_replace_load_by_const(load, value);
1093 if (value != NULL) {
1094 /* we completely replace the load by this value */
1095 DB((dbg, LEVEL_1, "Replacing Load %+F by constant %+F\n", m->node, value));
1096 mark_replace_load(m, value);
1101 if (m->value.value != NULL && !(m->flags & FLAG_IGNORE)) {
1102 /* only create an address if this node is NOT killed immediately or ignored */
1103 m->value.id = register_address(ptr);
1106 /* no user, KILL it */
1107 mark_replace_load(m, NULL);
1109 } /* update_Load_memop */
1112 * Update a memop for a Store.
1114 * @param m the memop
1116 static void update_Store_memop(memop_t *m)
1119 ir_node *store = m->node;
1120 ir_node *adr = get_Store_ptr(store);
1122 if (get_Store_volatility(store) == volatility_is_volatile) {
1123 m->flags |= FLAG_IGNORE;
1125 /* only create an address if this node is NOT ignored */
1126 m->value.id = register_address(adr);
1130 m->value.address = adr;
1132 for (i = get_irn_n_outs(store) - 1; i >= 0; --i) {
1133 ir_node *proj = get_irn_out(store, i);
1136 /* beware of keep edges */
1140 pn = get_Proj_proj(proj);
1141 m->projs[pn] = proj;
1143 case pn_Store_X_except:
1144 m->flags |= FLAG_EXCEPTION;
1149 case pn_Store_X_regular:
1152 panic("Unsupported Proj from Store %+F", proj);
1155 m->value.value = get_Store_value(store);
1156 m->value.mode = get_irn_mode(m->value.value);
1157 } /* update_Store_memop */
1160 * Update a memop for a Call.
1162 * @param m the memop
1164 static void update_Call_memop(memop_t *m)
1166 ir_node *call = m->node;
1167 unsigned prop = get_Call_memory_properties(call);
1170 if (prop & mtp_property_const) {
1171 /* A constant call did NOT use memory at all, we
1172 can kick it from the list. */
1173 } else if (prop & mtp_property_pure) {
1174 /* pure calls READ memory */
1177 m->flags = FLAG_KILL_ALL;
1179 for (i = get_irn_n_outs(call) - 1; i >= 0; --i) {
1180 ir_node *proj = get_irn_out(call, i);
1182 /* beware of keep edges */
1186 switch (get_Proj_proj(proj)) {
1187 case pn_Call_X_except:
1188 m->flags |= FLAG_EXCEPTION;
1195 } /* update_Call_memop */
1198 * Update a memop for a Div/Mod/Quot/DivMod.
1200 * @param m the memop
1202 static void update_DivOp_memop(memop_t *m)
1204 ir_node *div = m->node;
1207 for (i = get_irn_n_outs(div) - 1; i >= 0; --i) {
1208 ir_node *proj = get_irn_out(div, i);
1210 /* beware of keep edges */
1214 switch (get_Proj_proj(proj)) {
1215 case pn_Generic_X_except:
1216 m->flags |= FLAG_EXCEPTION;
1223 } /* update_DivOp_memop */
1226 * Update a memop for a Phi.
1228 * @param m the memop
1230 static void update_Phi_memop(memop_t *m)
1232 /* the Phi is it's own mem */
1234 } /* update_Phi_memop */
1237 * Memory walker: collect all memory ops and build topological lists.
1239 static void collect_memops(ir_node *irn, void *ctx)
1247 /* we can safely ignore ProjM's except the initial memory */
1248 if (irn != get_irg_initial_mem(current_ir_graph))
1252 op = alloc_memop(irn);
1253 block = get_nodes_block(irn);
1254 entry = get_block_entry(block);
1257 update_Phi_memop(op);
1258 /* Phis must be always placed first */
1259 op->next = entry->memop_forward;
1260 entry->memop_forward = op;
1261 if (entry->memop_backward == NULL)
1262 entry->memop_backward = op;
1264 switch (get_irn_opcode(irn)) {
1266 update_Load_memop(op);
1269 update_Store_memop(op);
1272 update_Call_memop(op);
1279 /* initial memory */
1284 /* we can those to find the memory edge */
1290 update_DivOp_memop(op);
1294 /* TODO: handle some builtins */
1296 /* unsupported operation */
1297 op->flags = FLAG_KILL_ALL;
1301 /* all other should be placed last */
1302 if (entry->memop_backward == NULL) {
1303 entry->memop_forward = entry->memop_backward = op;
1305 entry->memop_backward->next = op;
1306 entry->memop_backward = op;
1309 } /* collect_memops */
1312 * Find an address in the current set.
1314 * @param value the value to be searched for
1316 * @return a memop for the value or NULL if the value does
1317 * not exists in the set or cannot be converted into
1318 * the requested mode
1320 static memop_t *find_address(const value_t *value)
1322 if (rbitset_is_set(env.curr_set, value->id)) {
1323 memop_t *res = env.curr_id_2_memop[value->id];
1325 if (res->value.mode == value->mode)
1327 /* allow hidden casts */
1328 if (get_mode_arithmetic(res->value.mode) == irma_twos_complement &&
1329 get_mode_arithmetic(value->mode) == irma_twos_complement &&
1330 get_mode_size_bits(res->value.mode) == get_mode_size_bits(value->mode))
1334 } /* find_address */
1337 * Find an address in the avail_out set.
1339 * @param bl the block
1341 static memop_t *find_address_avail(const block_t *bl, unsigned id, const ir_mode *mode)
1343 if (rbitset_is_set(bl->avail_out, id)) {
1344 memop_t *res = bl->id_2_memop_avail[id];
1346 if (res->value.mode == mode)
1348 /* allow hidden casts */
1349 if (get_mode_arithmetic(res->value.mode) == irma_twos_complement &&
1350 get_mode_arithmetic(mode) == irma_twos_complement &&
1351 get_mode_size_bits(res->value.mode) == get_mode_size_bits(mode))
1355 } /* find_address_avail */
1358 * Kill all addresses from the current set.
1360 static void kill_all(void)
1362 rbitset_clear_all(env.curr_set, env.rbs_size);
1365 rbitset_set(env.curr_set, env.rbs_size - 1);
1369 * Kill memops that are not alias free due to a Store value from the current set.
1371 * @param value the Store value
1373 static void kill_memops(const value_t *value)
1375 unsigned end = env.rbs_size - 1;
1378 for (pos = rbitset_next(env.curr_set, 0, 1); pos < end; pos = rbitset_next(env.curr_set, pos + 1, 1)) {
1379 memop_t *op = env.curr_id_2_memop[pos];
1381 if (ir_no_alias != get_alias_relation(current_ir_graph, value->address, value->mode,
1382 op->value.address, op->value.mode)) {
1383 rbitset_clear(env.curr_set, pos);
1384 env.curr_id_2_memop[pos] = NULL;
1385 DB((dbg, LEVEL_2, "KILLING %+F because of possible alias address %+F\n", op->node, value->address));
1391 * Add the value of a memop to the current set.
1393 * @param op the memory op
1395 static void add_memop(memop_t *op)
1397 rbitset_set(env.curr_set, op->value.id);
1398 env.curr_id_2_memop[op->value.id] = op;
1402 * Add the value of a memop to the avail_out set.
1404 * @param bl the block
1405 * @param op the memory op
1407 static void add_memop_avail(block_t *bl, memop_t *op)
1409 rbitset_set(bl->avail_out, op->value.id);
1410 bl->id_2_memop_avail[op->value.id] = op;
1411 } /* add_memop_avail */
1414 * Check, if we can convert a value of one mode to another mode
1415 * without changing the representation of bits.
1417 * @param from the original mode
1418 * @param to the destination mode
1420 static int can_convert_to(const ir_mode *from, const ir_mode *to)
1422 if (get_mode_arithmetic(from) == irma_twos_complement &&
1423 get_mode_arithmetic(to) == irma_twos_complement &&
1424 get_mode_size_bits(from) == get_mode_size_bits(to))
1427 } /* can_convert_to */
1430 * Add a Conv to the requested mode if needed.
1432 * @param irn the IR-node to convert
1433 * @param mode the destination mode
1435 * @return the possible converted node or NULL
1436 * if the conversion is not possible
1438 static ir_node *conv_to(ir_node *irn, ir_mode *mode)
1440 ir_mode *other = get_irn_mode(irn);
1441 if (other != mode) {
1442 /* different modes: check if conversion is possible without changing the bits */
1443 if (can_convert_to(other, mode)) {
1444 ir_node *block = get_nodes_block(irn);
1445 return new_r_Conv(block, irn, mode);
1447 /* otherwise not possible ... yet */
1454 * Update the address of an value if this address was a load result
1455 * and the load is killed now.
1457 * @param value the value whose address is updated
1459 static void update_address(value_t *value)
1461 if (is_Proj(value->address)) {
1462 ir_node *load = get_Proj_pred(value->address);
1464 if (is_Load(load)) {
1465 const memop_t *op = get_irn_memop(load);
1467 if (op->flags & FLAG_KILLED_NODE)
1468 value->address = op->replace;
1471 } /* update_address */
1474 * Do forward dataflow analysis on the given block and calculate the
1475 * GEN and KILL in the current (avail) set.
1477 * @param bl the block
1479 static void calc_gen_kill_avail(block_t *bl)
1484 for (op = bl->memop_forward; op != NULL; op = op->next) {
1485 switch (get_irn_opcode(op->node)) {
1493 if (! (op->flags & (FLAG_KILLED_NODE|FLAG_IGNORE))) {
1494 /* do we have this already? */
1497 update_address(&op->value);
1498 other = find_address(&op->value);
1499 if (other != NULL && other != op) {
1500 def = conv_to(other->value.value, op->value.mode);
1502 #ifdef DEBUG_libfirm
1503 if (is_Store(other->node)) {
1505 DB((dbg, LEVEL_1, "RAW %+F <- %+F(%+F)\n", op->node, def, other->node));
1508 DB((dbg, LEVEL_1, "RAR %+F <- %+F(%+F)\n", op->node, def, other->node));
1511 mark_replace_load(op, def);
1512 /* do NOT change the memop table */
1516 /* add this value */
1521 if (! (op->flags & FLAG_KILLED_NODE)) {
1522 /* do we have this store already */
1525 update_address(&op->value);
1526 other = find_address(&op->value);
1527 if (other != NULL) {
1528 if (is_Store(other->node)) {
1529 if (op != other && !(other->flags & FLAG_IGNORE) &&
1530 get_nodes_block(other->node) == get_nodes_block(op->node)) {
1532 * A WAW in the same block we can kick the first store.
1533 * This is a shortcut: we know that the second Store will be anticipated
1536 DB((dbg, LEVEL_1, "WAW %+F <- %+F\n", other->node, op->node));
1537 mark_remove_store(other);
1538 /* FIXME: a Load might be get freed due to this killed store */
1540 } else if (other->value.value == op->value.value && !(op->flags & FLAG_IGNORE)) {
1542 DB((dbg, LEVEL_1, "WAR %+F <- %+F\n", op->node, other->node));
1543 mark_remove_store(op);
1544 /* do NOT change the memop table */
1548 /* KILL all possible aliases */
1549 kill_memops(&op->value);
1550 /* add this value */
1555 if (op->flags & FLAG_KILL_ALL)
1559 } /* calc_gen_kill_avail */
1561 #define BYTE_SIZE(x) (((x) + 7) >> 3)
1564 * Do forward dataflow analysis on a given block to calculate the avail_out set
1565 * for this block only.
1567 * @param block the block
1569 static void forward_avail(block_t *bl)
1571 /* fill the data from the current block */
1572 env.curr_id_2_memop = bl->id_2_memop_avail;
1573 env.curr_set = bl->avail_out;
1575 calc_gen_kill_avail(bl);
1576 dump_curr(bl, "Avail_out");
1577 } /* forward_avail */
1580 * Do backward dataflow analysis on a given block to calculate the antic set
1581 * of Loaded addresses.
1583 * @param bl the block
1585 * @return non-zero if the set has changed since last iteration
1587 static int backward_antic(block_t *bl)
1590 ir_node *block = bl->block;
1591 int n = get_Block_n_cfg_outs(block);
1594 ir_node *succ = get_Block_cfg_out(block, 0);
1595 block_t *succ_bl = get_block_entry(succ);
1596 int pred_pos = get_Block_cfgpred_pos(succ, block);
1597 unsigned end = env.rbs_size - 1;
1602 if (bl->trans_results == NULL) {
1603 /* allocate the translate cache */
1604 bl->trans_results = OALLOCNZ(&env.obst, memop_t*, env.curr_adr_id);
1607 /* check for partly redundant values */
1608 for (pos = rbitset_next(succ_bl->anticL_in, 0, 1);
1610 pos = rbitset_next(succ_bl->anticL_in, pos + 1, 1)) {
1612 * do Phi-translation here: Note that at this point the nodes are
1613 * not changed, so we can safely cache the results.
1614 * However: Loads of Load results ARE bad, because we have no way
1615 to translate them yet ...
1617 memop_t *op = bl->trans_results[pos];
1619 /* not yet translated */
1620 ir_node *adr, *trans_adr;
1622 op = succ_bl->id_2_memop_antic[pos];
1623 adr = op->value.address;
1625 trans_adr = phi_translate(adr, succ, pred_pos);
1626 if (trans_adr != adr) {
1627 /* create a new entry for the translated one */
1630 new_op = alloc_memop(NULL);
1631 new_op->value.address = trans_adr;
1632 new_op->value.id = register_address(trans_adr);
1633 new_op->value.mode = op->value.mode;
1634 new_op->node = op->node; /* we need the node to decide if Load/Store */
1635 new_op->flags = op->flags;
1637 bl->trans_results[pos] = new_op;
1641 env.curr_id_2_memop[op->value.id] = op;
1642 rbitset_set(env.curr_set, op->value.id);
1645 ir_node *succ = get_Block_cfg_out(block, 0);
1646 block_t *succ_bl = get_block_entry(succ);
1649 rbitset_copy(env.curr_set, succ_bl->anticL_in, env.rbs_size);
1650 memcpy(env.curr_id_2_memop, succ_bl->id_2_memop_antic, env.rbs_size * sizeof(env.curr_id_2_memop[0]));
1652 /* Hmm: probably we want kill merges of Loads ans Stores here */
1653 for (i = n - 1; i > 0; --i) {
1654 ir_node *succ = get_Block_cfg_out(bl->block, i);
1655 block_t *succ_bl = get_block_entry(succ);
1657 rbitset_and(env.curr_set, succ_bl->anticL_in, env.rbs_size);
1660 /* block ends with a noreturn call */
1664 dump_curr(bl, "AnticL_out");
1666 for (op = bl->memop_backward; op != NULL; op = op->prev) {
1667 switch (get_irn_opcode(op->node)) {
1675 if (! (op->flags & (FLAG_KILLED_NODE|FLAG_IGNORE))) {
1681 if (! (op->flags & FLAG_KILLED_NODE)) {
1682 /* a Store: check which memops must be killed */
1683 kill_memops(&op->value);
1687 if (op->flags & FLAG_KILL_ALL)
1692 memcpy(bl->id_2_memop_antic, env.curr_id_2_memop, env.rbs_size * sizeof(env.curr_id_2_memop[0]));
1693 if (! rbitset_equal(bl->anticL_in, env.curr_set, env.rbs_size)) {
1695 rbitset_copy(bl->anticL_in, env.curr_set, env.rbs_size);
1696 dump_curr(bl, "AnticL_in*");
1699 dump_curr(bl, "AnticL_in");
1701 } /* backward_antic */
1704 * Replace a Load memop by a already known value.
1706 * @param op the Load memop
1708 static void replace_load(memop_t *op)
1710 ir_node *load = op->node;
1711 ir_node *def = skip_Id(op->replace);
1716 DB((dbg, LEVEL_1, "Replacing %+F by definition %+F\n", load, is_Proj(def) ? get_Proj_pred(def) : def));
1718 if (op->flags & FLAG_EXCEPTION) {
1719 /* bad: this node is unused and executed for exception only */
1720 DB((dbg, LEVEL_1, "Unused %+F executed for exception only ...\n", load));
1723 DB((dbg, LEVEL_1, "Killing unused %+F\n", load));
1726 if (op->mem != NULL) {
1727 /* in rare cases a Load might have NO memory */
1728 exchange(op->mem, get_Load_mem(load));
1730 proj = op->projs[pn_Load_res];
1732 mode = get_irn_mode(proj);
1733 if (get_irn_mode(def) != mode) {
1735 dbg_info *db = get_irn_dbg_info(load);
1736 ir_node *block = get_nodes_block(proj);
1737 def = new_rd_Conv(db, block, def, mode);
1739 exchange(proj, def);
1741 proj = op->projs[pn_Load_X_except];
1743 exchange(proj, new_Bad());
1745 proj = op->projs[pn_Load_X_regular];
1747 exchange(proj, new_r_Jmp(get_nodes_block(load)));
1749 } /* replace_load */
1752 * Remove a Store memop.
1754 * @param op the Store memop
1756 static void remove_store(memop_t *op)
1758 ir_node *store = op->node;
1761 DB((dbg, LEVEL_1, "Removing %+F\n", store));
1763 if (op->mem != NULL) {
1764 /* in rare cases a Store might have no memory */
1765 exchange(op->mem, get_Store_mem(store));
1767 proj = op->projs[pn_Store_X_except];
1769 exchange(proj, new_Bad());
1771 proj = op->projs[pn_Store_X_regular];
1773 exchange(proj, new_r_Jmp(get_nodes_block(store)));
1775 } /* remove_store */
1779 * Do all necessary replacements for a given block.
1781 * @param bl the block
1783 static void do_replacements(block_t *bl)
1787 for (op = bl->memop_forward; op != NULL; op = op->next) {
1788 if (op->flags & FLAG_KILLED_NODE) {
1789 switch (get_irn_opcode(op->node)) {
1799 } /* do_replacements */
1802 * Calculate the Avail_out sets for all basic blocks.
1804 static void calcAvail(void)
1806 memop_t **tmp_memop = env.curr_id_2_memop;
1807 unsigned *tmp_set = env.curr_set;
1810 /* calculate avail_out */
1811 DB((dbg, LEVEL_2, "Calculate Avail_out\n"));
1813 /* iterate over all blocks in in any order, skip the start block */
1814 for (bl = env.forward->forward_next; bl != NULL; bl = bl->forward_next) {
1818 /* restore the current sets */
1819 env.curr_id_2_memop = tmp_memop;
1820 env.curr_set = tmp_set;
1824 * Calculate the Antic_in sets for all basic blocks.
1826 static void calcAntic(void)
1830 /* calculate antic_out */
1831 DB((dbg, LEVEL_2, "Calculate Antic_in\n"));
1836 DB((dbg, LEVEL_2, "Iteration %d:\n=========\n", i));
1840 /* over all blocks in reverse post order */
1841 for (bl = env.backward->backward_next; bl != NULL; bl = bl->backward_next) {
1842 need_iter |= backward_antic(bl);
1845 } while (need_iter);
1846 DB((dbg, LEVEL_2, "Get anticipated Load set after %d iterations\n", i));
1850 * Return the node representing the last memory in a block.
1852 * @param bl the block
1854 static ir_node *find_last_memory(block_t *bl)
1857 if (bl->memop_backward != NULL) {
1858 return bl->memop_backward->mem;
1860 /* if there is NO memory in this block, go to the dominator */
1861 bl = get_block_entry(get_Block_idom(bl->block));
1863 } /* find_last_memory */
1866 * Reroute all memory users of old memory
1867 * to a new memory IR-node.
1869 * @param omem the old memory IR-node
1870 * @param nmem the new memory IR-node
1872 static void reroute_all_mem_users(ir_node *omem, ir_node *nmem)
1876 for (i = get_irn_n_outs(omem) - 1; i >= 0; --i) {
1878 ir_node *user = get_irn_out_ex(omem, i, &n_pos);
1880 set_irn_n(user, n_pos, nmem);
1883 /* all edges previously point to omem now point to nmem */
1884 nmem->out = omem->out;
1885 } /* reroute_all_mem_users */
1888 * Reroute memory users of old memory that are dominated by a given block
1889 * to a new memory IR-node.
1891 * @param omem the old memory IR-node
1892 * @param nmem the new memory IR-node
1893 * @param pass_bl the block the memory must pass
1895 static void reroute_mem_through(ir_node *omem, ir_node *nmem, ir_node *pass_bl)
1897 int i, j, n = get_irn_n_outs(omem);
1898 ir_def_use_edge *edges = NEW_ARR_D(ir_def_use_edge, &env.obst, n + 1);
1900 for (i = j = 0; i < n; ++i) {
1902 ir_node *user = get_irn_out_ex(omem, i, &n_pos);
1903 ir_node *use_bl = get_nodes_block(user);
1907 use_bl = get_Block_cfgpred_block(use_bl, n_pos);
1909 if (block_dominates(pass_bl, use_bl)) {
1910 /* found an user that is dominated */
1912 edges[j].pos = n_pos;
1913 edges[j].use = user;
1915 set_irn_n(user, n_pos, nmem);
1919 /* Modify the out structure: we create a new out edge array on our
1920 temporary obstack here. This should be no problem, as we invalidate the edges
1921 at the end either. */
1922 /* first entry is used for the length */
1925 } /* reroute_mem_through */
1928 * insert Loads, making partly redundant Loads fully redundant
1930 static int insert_Load(block_t *bl)
1932 ir_node *block = bl->block;
1933 int i, n = get_Block_n_cfgpreds(block);
1934 unsigned end = env.rbs_size - 1;
1937 DB((dbg, LEVEL_3, "processing %+F\n", block));
1940 /* might still happen for an unreachable block (end for instance) */
1948 NEW_ARR_A(ir_node *, ins, n);
1950 rbitset_set_all(env.curr_set, env.rbs_size);
1952 /* More than one predecessors, calculate the join for all avail_outs ignoring unevaluated
1953 Blocks. These put in Top anyway. */
1954 for (i = n - 1; i >= 0; --i) {
1955 ir_node *pred = skip_Proj(get_Block_cfgpred(block, i));
1956 ir_node *blk = get_nodes_block(pred);
1959 pred_bl = get_block_entry(blk);
1960 rbitset_and(env.curr_set, pred_bl->avail_out, env.rbs_size);
1962 if (is_Load(pred) || is_Store(pred)) {
1963 /* We reached this block by an exception from a Load or Store:
1964 * the memop creating the exception was NOT completed than, kill it
1966 memop_t *exc_op = get_irn_memop(pred);
1967 rbitset_clear(env.curr_set, exc_op->value.id);
1972 * Ensure that all values are in the map: build Phi's if necessary:
1973 * Note: the last bit is the sentinel and ALWAYS set, so start with -2.
1975 for (pos = env.rbs_size - 2; pos >= 0; --pos) {
1976 if (! rbitset_is_set(env.curr_set, pos))
1977 env.curr_id_2_memop[pos] = NULL;
1979 ir_node *pred = get_Block_cfgpred_block(bl->block, 0);
1980 block_t *pred_bl = get_block_entry(pred);
1982 memop_t *first = NULL;
1983 ir_mode *mode = NULL;
1985 for (i = 0; i < n; ++i) {
1988 pred = get_Block_cfgpred_block(bl->block, i);
1989 pred_bl = get_block_entry(pred);
1991 mop = pred_bl->id_2_memop_avail[pos];
1992 if (first == NULL) {
1994 ins[0] = first->value.value;
1995 mode = get_irn_mode(ins[0]);
1997 /* no Phi needed so far */
1998 env.curr_id_2_memop[pos] = first;
2000 ins[i] = conv_to(mop->value.value, mode);
2001 if (ins[i] != ins[0]) {
2002 if (ins[i] == NULL) {
2003 /* conversion failed */
2004 env.curr_id_2_memop[pos] = NULL;
2005 rbitset_clear(env.curr_set, pos);
2014 ir_node *phi = new_r_Phi(bl->block, n, ins, mode);
2015 memop_t *phiop = alloc_memop(phi);
2017 phiop->value = first->value;
2018 phiop->value.value = phi;
2020 /* no need to link it in, as it is a DATA phi */
2022 env.curr_id_2_memop[pos] = phiop;
2024 DB((dbg, LEVEL_3, "Created new %+F on merging value for address %+F\n", phi, first->value.address));
2029 /* only one predecessor, simply copy the map */
2030 ir_node *pred = get_Block_cfgpred_block(bl->block, 0);
2031 block_t *pred_bl = get_block_entry(pred);
2033 rbitset_copy(env.curr_set, pred_bl->avail_out, env.rbs_size);
2035 memcpy(env.curr_id_2_memop, pred_bl->id_2_memop_avail, env.rbs_size * sizeof(bl->id_2_memop_avail[0]));
2039 /* check for partly redundant values */
2040 for (pos = rbitset_next(bl->anticL_in, 0, 1);
2042 pos = rbitset_next(bl->anticL_in, pos + 1, 1)) {
2043 memop_t *op = bl->id_2_memop_antic[pos];
2044 int have_some, all_same;
2047 if (rbitset_is_set(env.curr_set, pos)) {
2052 assert(is_Load(op->node));
2054 DB((dbg, LEVEL_3, "anticipated %+F\n", op->node));
2059 for (i = n - 1; i >= 0; --i) {
2060 ir_node *pred = get_Block_cfgpred_block(block, i);
2061 block_t *pred_bl = get_block_entry(pred);
2062 ir_mode *mode = op->value.mode;
2066 adr = phi_translate(op->value.address, block, i);
2067 DB((dbg, LEVEL_3, ".. using address %+F in pred %d\n", adr, i));
2068 e = find_address_avail(pred_bl, register_address(adr), mode);
2070 ir_node *ef_block = get_nodes_block(adr);
2071 if (! block_dominates(ef_block, pred)) {
2072 /* cannot place a copy here */
2074 DB((dbg, LEVEL_3, "%+F cannot be moved into predecessor %+F\n", op->node, pred));
2077 DB((dbg, LEVEL_3, "%+F is not available in predecessor %+F\n", op->node, pred));
2078 pred_bl->avail = NULL;
2081 if (e->value.mode != mode && !can_convert_to(e->value.mode, mode)) {
2082 /* cannot create a Phi due to different modes */
2088 DB((dbg, LEVEL_3, "%+F is available for %+F in predecessor %+F\n", e->node, op->node, pred));
2091 else if (first != e->node)
2095 if (have_some && !all_same) {
2096 ir_mode *mode = op->value.mode;
2100 NEW_ARR_A(ir_node *, in, n);
2102 for (i = n - 1; i >= 0; --i) {
2103 ir_node *pred = get_Block_cfgpred_block(block, i);
2104 block_t *pred_bl = get_block_entry(pred);
2106 if (pred_bl->avail == NULL) {
2107 /* create a new Load here and make to make it fully redundant */
2108 dbg_info *db = get_irn_dbg_info(op->node);
2109 ir_node *last_mem = find_last_memory(pred_bl);
2110 ir_node *load, *def, *adr;
2113 assert(last_mem != NULL);
2114 adr = phi_translate(op->value.address, block, i);
2115 load = new_rd_Load(db, pred, last_mem, adr, mode, cons_none);
2116 def = new_r_Proj(load, mode, pn_Load_res);
2117 DB((dbg, LEVEL_1, "Created new %+F in %+F for party redundant %+F\n", load, pred, op->node));
2119 new_op = alloc_memop(load);
2120 new_op->mem = new_r_Proj(load, mode_M, pn_Load_M);
2121 new_op->value.address = adr;
2122 new_op->value.id = op->value.id;
2123 new_op->value.mode = mode;
2124 new_op->value.value = def;
2126 new_op->projs[pn_Load_M] = new_op->mem;
2127 new_op->projs[pn_Load_res] = def;
2129 new_op->prev = pred_bl->memop_backward;
2130 if (pred_bl->memop_backward != NULL)
2131 pred_bl->memop_backward->next = new_op;
2133 pred_bl->memop_backward = new_op;
2135 if (pred_bl->memop_forward == NULL)
2136 pred_bl->memop_forward = new_op;
2138 if (get_nodes_block(last_mem) == pred) {
2139 /* We have add a new last memory op in pred block.
2140 If pred had already a last mem, reroute all memory
2142 reroute_all_mem_users(last_mem, new_op->mem);
2144 /* reroute only those memory going through the pre block */
2145 reroute_mem_through(last_mem, new_op->mem, pred);
2148 /* we added this load at the end, so it will be avail anyway */
2149 add_memop_avail(pred_bl, new_op);
2150 pred_bl->avail = new_op;
2152 in[i] = conv_to(pred_bl->avail->value.value, mode);
2154 phi = new_r_Phi(block, n, in, mode);
2155 DB((dbg, LEVEL_1, "Created new %+F in %+F for now redundant %+F\n", phi, block, op->node));
2157 phi_op = clone_memop_phi(op, phi);
2163 /* recalculate avail by gen and kill */
2164 calc_gen_kill_avail(bl);
2166 /* always update the map after gen/kill, as values might have been changed due to RAR/WAR/WAW */
2167 memcpy(bl->id_2_memop_avail, env.curr_id_2_memop, env.rbs_size * sizeof(env.curr_id_2_memop[0]));
2169 if (!rbitset_equal(bl->avail_out, env.curr_set, env.rbs_size)) {
2170 /* the avail set has changed */
2171 rbitset_copy(bl->avail_out, env.curr_set, env.rbs_size);
2172 dump_curr(bl, "Avail_out*");
2175 dump_curr(bl, "Avail_out");
2180 * Insert Loads upwards.
2182 static void insert_Loads_upwards(void)
2187 /* recalculate antic_out and insert Loads */
2188 DB((dbg, LEVEL_2, "Inserting Loads\n"));
2192 DB((dbg, LEVEL_2, "Iteration %d:\n=========\n", i));
2196 /* over all blocks in reverse post order, skip the start block */
2197 for (bl = env.forward->forward_next; bl != NULL; bl = bl->forward_next) {
2198 need_iter |= insert_Load(bl);
2201 } while (need_iter);
2203 DB((dbg, LEVEL_2, "Finished Load inserting after %d iterations\n", i));
2204 } /* insert_Loads_upwards */
2207 * Kill unreachable control flow.
2209 * @param irg the graph to operate on
2211 static void kill_unreachable_blocks(ir_graph *irg)
2217 NEW_ARR_A(ir_node *, ins, env.max_cfg_preds);
2219 for (bl = env.forward; bl != NULL; bl = bl->forward_next) {
2220 ir_node *block = bl->block;
2223 assert(get_Block_mark(block));
2225 n = get_Block_n_cfgpreds(block);
2227 for (i = j = 0; i < n; ++i) {
2228 ir_node *pred = get_Block_cfgpred(block, i);
2234 pred_bl = get_nodes_block(skip_Proj(pred));
2235 if (! get_Block_mark(pred_bl))
2241 ir_node *phi, *next;
2243 /* some unreachable blocks detected */
2246 DB((dbg, LEVEL_1, "Killing dead block predecessors on %+F\n", block));
2248 set_irn_in(block, j, ins);
2250 /* shorten all Phi nodes */
2251 for (phi = get_Block_phis(block); phi != NULL; phi = next) {
2252 next = get_Phi_next(phi);
2254 for (i = k = 0; i < n; ++i) {
2255 ir_node *pred = get_Block_cfgpred_block(block, i);
2260 if (! get_Block_mark(pred))
2263 ins[k++] = get_Phi_pred(phi, i);
2266 exchange(phi, ins[0]);
2268 set_irn_in(phi, k, ins);
2275 /* kick keep alives */
2276 ir_node *end = get_irg_end(irg);
2277 int i, j, n = get_End_n_keepalives(end);
2279 NEW_ARR_A(ir_node *, ins, n);
2281 for (i = j = 0; i < n; ++i) {
2282 ir_node *ka = get_End_keepalive(end, i);
2290 ka_bl = get_nodes_block(skip_Proj(ka));
2291 if (get_Block_mark(ka_bl))
2295 set_End_keepalives(end, j, ins);
2299 /* this transformation do NOT invalidate the dominance */
2301 } /* kill_unreachable_blocks */
2303 int opt_ldst(ir_graph *irg)
2306 ir_graph *rem = current_ir_graph;
2308 current_ir_graph = irg;
2310 FIRM_DBG_REGISTER(dbg, "firm.opt.ldst");
2311 // firm_dbg_set_mask(dbg, -1);
2313 DB((dbg, LEVEL_1, "\nDoing Load/Store optimization on %+F\n", irg));
2315 /* we need landing pads */
2316 remove_critical_cf_edges(irg);
2318 // dump_ir_block_graph(irg, "-XXX");
2320 if (get_opt_alias_analysis()) {
2321 assure_irg_entity_usage_computed(irg);
2322 assure_irp_globals_entity_usage_computed();
2325 obstack_init(&env.obst);
2326 ir_nodemap_init(&env.adr_map);
2329 env.backward = NULL;
2330 env.curr_adr_id = 0;
2332 env.max_cfg_preds = 0;
2334 env.start_bl = get_irg_start_block(irg);
2335 env.end_bl = get_irg_end_block(irg);
2336 #ifdef DEBUG_libfirm
2337 env.id_2_address = NEW_ARR_F(ir_node *, 0);
2340 assure_irg_outs(irg);
2342 ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK | IR_RESOURCE_BLOCK_MARK);
2344 /* first step: allocate block entries. Note that some blocks might be
2345 unreachable here. Using the normal walk ensures that ALL blocks are initialized. */
2346 irg_walk_graph(irg, prepare_blocks, link_phis, NULL);
2348 /* produce an inverse post-order list for the CFG: this links only reachable
2350 irg_out_block_walk(get_irg_start_block(irg), NULL, inverse_post_order, NULL);
2352 if (! get_Block_mark(env.end_bl)) {
2354 * The end block is NOT reachable due to endless loops
2355 * or no_return calls.
2356 * Place the end block last.
2357 * env.backward points to the last block in the list for this purpose.
2359 env.backward->forward_next = get_block_entry(env.end_bl);
2361 set_Block_mark(env.end_bl, 1);
2364 /* KILL unreachable blocks: these disturb the data flow analysis */
2365 kill_unreachable_blocks(irg);
2369 /* second step: find and sort all memory ops */
2370 walk_memory_irg(irg, collect_memops, NULL, NULL);
2372 #ifdef DEBUG_libfirm
2373 /* check that the backward map is correct */
2374 assert((unsigned)ARR_LEN(env.id_2_address) == env.curr_adr_id);
2377 if (env.n_mem_ops == 0) {
2382 /* create the backward links. */
2383 env.backward = NULL;
2384 irg_block_walk_graph(irg, NULL, collect_backward, NULL);
2386 /* link the end block in */
2387 bl = get_block_entry(env.end_bl);
2388 bl->backward_next = env.backward;
2391 /* check that we really start with the start / end block */
2392 assert(env.forward->block == env.start_bl);
2393 assert(env.backward->block == env.end_bl);
2395 /* create address sets: for now, only the existing addresses are allowed plus one
2396 needed for the sentinel */
2397 env.rbs_size = env.curr_adr_id + 1;
2399 /* create the current set */
2400 env.curr_set = rbitset_obstack_alloc(&env.obst, env.rbs_size);
2401 rbitset_set(env.curr_set, env.rbs_size - 1);
2402 env.curr_id_2_memop = NEW_ARR_D(memop_t *, &env.obst, env.rbs_size);
2403 memset(env.curr_id_2_memop, 0, env.rbs_size * sizeof(env.curr_id_2_memop[0]));
2405 for (bl = env.forward; bl != NULL; bl = bl->forward_next) {
2406 /* set sentinel bits */
2407 bl->avail_out = rbitset_obstack_alloc(&env.obst, env.rbs_size);
2408 rbitset_set(bl->avail_out, env.rbs_size - 1);
2410 bl->id_2_memop_avail = NEW_ARR_D(memop_t *, &env.obst, env.rbs_size);
2411 memset(bl->id_2_memop_avail, 0, env.rbs_size * sizeof(bl->id_2_memop_avail[0]));
2413 bl->anticL_in = rbitset_obstack_alloc(&env.obst, env.rbs_size);
2414 rbitset_set(bl->anticL_in, env.rbs_size - 1);
2416 bl->id_2_memop_antic = NEW_ARR_D(memop_t *, &env.obst, env.rbs_size);
2417 memset(bl->id_2_memop_antic, 0, env.rbs_size * sizeof(bl->id_2_memop_antic[0]));
2420 // dump_block_list(&env);
2421 (void) dump_block_list;
2426 insert_Loads_upwards();
2429 /* over all blocks in reverse post order */
2430 for (bl = env.forward; bl != NULL; bl = bl->forward_next) {
2431 do_replacements(bl);
2434 /* not only invalidate but free them. We might allocate new out arrays
2435 on our obstack which will be deleted yet. */
2437 set_irg_entity_usage_state(irg, ir_entity_usage_not_computed);
2441 ir_free_resources(irg, IR_RESOURCE_IRN_LINK | IR_RESOURCE_BLOCK_MARK);
2442 ir_nodemap_destroy(&env.adr_map);
2443 obstack_free(&env.obst, NULL);
2445 // dump_ir_block_graph(irg, "-YYY");
2447 #ifdef DEBUG_libfirm
2448 DEL_ARR_F(env.id_2_address);
2451 current_ir_graph = rem;
2452 return env.changed != 0;
2455 ir_graph_pass_t *opt_ldst_pass(const char *name)
2457 return def_graph_pass_ret(name ? name : "ldst_df", opt_ldst);
2458 } /* opt_ldst_pass */