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 */
135 /* the one and only environment */
140 static firm_dbg_module_t *dbg;
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 * Walker: allocate an block entry for every block
333 * and register all potential addresses.
335 static void prepare_blocks(ir_node *irn, void *ctx) {
339 block_t *entry = obstack_alloc(&env.obst, sizeof(*entry));
342 entry->memop_forward = NULL;
343 entry->memop_backward = NULL;
344 entry->avail_out = NULL;
345 entry->id_2_memop_avail = NULL;
346 entry->anticL_in = NULL;
347 entry->id_2_memop_antic = NULL;
349 entry->forward_next = NULL;
350 entry->backward_next = NULL;
352 entry->trans_results = NULL;
353 set_irn_link(irn, entry);
355 set_Block_phis(irn, NULL);
357 /* use block marks to track unreachable blocks */
358 set_Block_mark(irn, 0);
360 n = get_Block_n_cfgpreds(irn);
361 if (n > env.max_cfg_preds)
362 env.max_cfg_preds = n;
364 ir_mode *mode = get_irn_mode(irn);
366 if (mode_is_reference(mode)) {
368 * Register ALL possible addresses: this is overkill yet but
369 * simpler then doing it for all possible translated addresses
370 * (which would be sufficient in the moment.
372 (void)register_address(irn);
375 } /* prepare_blocks */
378 * Post-Walker, link in all Phi's
380 static void link_phis(ir_node *irn, void *ctx) {
384 ir_node *block = get_nodes_block(irn);
385 add_Block_phi(block, irn);
390 * Block walker: creates the inverse post-order list for the CFG.
392 static void inverse_post_order(ir_node *block, void *ctx) {
393 block_t *entry = get_block_entry(block);
397 /* mark this block IS reachable from start */
398 set_Block_mark(block, 1);
400 /* create the list in inverse order */
401 entry->forward_next = env.forward;
404 /* remember the first visited (last in list) entry, needed for later */
405 if (env.backward == NULL)
406 env.backward = entry;
407 } /* inverse_post_order */
410 * Block walker: create backward links for the memops of a block.
412 static void collect_backward(ir_node *block, void *ctx) {
413 block_t *entry = get_block_entry(block);
419 * Do NOT link in the end block yet. We want it to be
420 * the first in the list. This is NOT guaranteed by the walker
421 * if we have endless loops.
423 if (block != env.end_bl) {
424 entry->backward_next = env.backward;
426 /* create the list in inverse order */
427 env.backward = entry;
430 /* create backward links for all memory ops */
432 for (op = entry->memop_forward; op != NULL; op = op->next) {
436 entry->memop_backward = last;
437 } /* collect_backward */
442 * @param irn the IR-node representing the memop or NULL
443 * if this is a translated (virtual) memop
445 * @return the allocated memop
447 static memop_t *alloc_memop(ir_node *irn) {
448 memop_t *m = obstack_alloc(&env.obst, sizeof(*m));
450 m->value.address = NULL;
451 m->value.value = NULL;
452 m->value.mode = NULL;
460 memset(m->projs, 0, sizeof(m->projs));
463 set_irn_link(irn, m);
468 * Create a memop for a Phi-replacement.
470 * @param op the memop to clone
471 * @param phi the Phi-node representing the new value
473 static memop_t *clone_memop_phi(memop_t *op, ir_node *phi) {
474 memop_t *m = obstack_alloc(&env.obst, sizeof(*m));
476 m->value = op->value;
477 m->value.value = phi;
484 set_irn_link(phi, m);
486 } /* clone_memop_phi */
489 * Return the memory properties of a call node.
491 * @param call the call node
493 * return a bitset of mtp_property_const and mtp_property_pure
495 static unsigned get_Call_memory_properties(ir_node *call) {
496 ir_type *call_tp = get_Call_type(call);
497 unsigned prop = get_method_additional_properties(call_tp);
499 /* check first the call type */
500 if ((prop & (mtp_property_const|mtp_property_pure)) == 0) {
501 /* try the called entity */
502 ir_node *ptr = get_Call_ptr(call);
504 if (is_Global(ptr)) {
505 ir_entity *ent = get_Global_entity(ptr);
507 prop = get_entity_additional_properties(ent);
510 return prop & (mtp_property_const|mtp_property_pure);
511 } /* get_Call_memory_properties */
514 * Returns an entity if the address ptr points to a constant one.
516 * @param ptr the address
518 * @return an entity or NULL
520 static ir_entity *find_constant_entity(ir_node *ptr) {
522 if (is_SymConst(ptr) && get_SymConst_kind(ptr) == symconst_addr_ent) {
523 return get_SymConst_entity(ptr);
524 } else if (is_Sel(ptr)) {
525 ir_entity *ent = get_Sel_entity(ptr);
526 ir_type *tp = get_entity_owner(ent);
528 /* Do not fiddle with polymorphism. */
529 if (is_Class_type(get_entity_owner(ent)) &&
530 ((get_entity_n_overwrites(ent) != 0) ||
531 (get_entity_n_overwrittenby(ent) != 0) ) )
534 if (is_Array_type(tp)) {
538 for (i = 0, n = get_Sel_n_indexs(ptr); i < n; ++i) {
540 tarval *tlower, *tupper;
541 ir_node *index = get_Sel_index(ptr, i);
542 tarval *tv = computed_value(index);
544 /* check if the index is constant */
545 if (tv == tarval_bad)
548 bound = get_array_lower_bound(tp, i);
549 tlower = computed_value(bound);
550 bound = get_array_upper_bound(tp, i);
551 tupper = computed_value(bound);
553 if (tlower == tarval_bad || tupper == tarval_bad)
556 if (tarval_cmp(tv, tlower) & pn_Cmp_Lt)
558 if (tarval_cmp(tupper, tv) & pn_Cmp_Lt)
561 /* ok, bounds check finished */
565 if (variability_constant == get_entity_variability(ent))
569 ptr = get_Sel_ptr(ptr);
570 } else if (is_Add(ptr)) {
571 ir_node *l = get_Add_left(ptr);
572 ir_node *r = get_Add_right(ptr);
574 if (get_irn_mode(l) == get_irn_mode(ptr) && is_Const(r))
576 else if (get_irn_mode(r) == get_irn_mode(ptr) && is_Const(l))
581 /* for now, we support only one addition, reassoc should fold all others */
582 if (! is_SymConst(ptr) && !is_Sel(ptr))
584 } else if (is_Sub(ptr)) {
585 ir_node *l = get_Sub_left(ptr);
586 ir_node *r = get_Sub_right(ptr);
588 if (get_irn_mode(l) == get_irn_mode(ptr) && is_Const(r))
592 /* for now, we support only one subtraction, reassoc should fold all others */
593 if (! is_SymConst(ptr) && !is_Sel(ptr))
598 } /* find_constant_entity */
601 * Return the Selection index of a Sel node from dimension n
603 static long get_Sel_array_index_long(ir_node *n, int dim) {
604 ir_node *index = get_Sel_index(n, dim);
605 assert(is_Const(index));
606 return get_tarval_long(get_Const_tarval(index));
607 } /* get_Sel_array_index_long */
610 * Returns the accessed component graph path for an
611 * node computing an address.
613 * @param ptr the node computing the address
614 * @param depth current depth in steps upward from the root
617 static compound_graph_path *rec_get_accessed_path(ir_node *ptr, int depth) {
618 compound_graph_path *res = NULL;
619 ir_entity *root, *field, *ent;
620 int path_len, pos, idx;
624 if (is_SymConst(ptr)) {
625 /* a SymConst. If the depth is 0, this is an access to a global
626 * entity and we don't need a component path, else we know
627 * at least its length.
629 assert(get_SymConst_kind(ptr) == symconst_addr_ent);
630 root = get_SymConst_entity(ptr);
631 res = (depth == 0) ? NULL : new_compound_graph_path(get_entity_type(root), depth);
632 } else if (is_Sel(ptr)) {
633 /* it's a Sel, go up until we find the root */
634 res = rec_get_accessed_path(get_Sel_ptr(ptr), depth+1);
638 /* fill up the step in the path at the current position */
639 field = get_Sel_entity(ptr);
640 path_len = get_compound_graph_path_length(res);
641 pos = path_len - depth - 1;
642 set_compound_graph_path_node(res, pos, field);
644 if (is_Array_type(get_entity_owner(field))) {
645 assert(get_Sel_n_indexs(ptr) == 1 && "multi dim arrays not implemented");
646 set_compound_graph_path_array_index(res, pos, get_Sel_array_index_long(ptr, 0));
648 } else if (is_Add(ptr)) {
649 ir_node *l = get_Add_left(ptr);
650 ir_node *r = get_Add_right(ptr);
651 ir_mode *mode = get_irn_mode(ptr);
654 if (is_Const(r) && get_irn_mode(l) == mode) {
656 tv = get_Const_tarval(r);
659 tv = get_Const_tarval(l);
662 mode = get_tarval_mode(tv);
665 /* ptr must be a Sel or a SymConst, this was checked in find_constant_entity() */
667 field = get_Sel_entity(ptr);
669 field = get_SymConst_entity(ptr);
672 for (ent = field;;) {
674 tarval *sz, *tv_index, *tlower, *tupper;
677 tp = get_entity_type(ent);
678 if (! is_Array_type(tp))
680 ent = get_array_element_entity(tp);
681 size = get_type_size_bytes(get_entity_type(ent));
682 sz = new_tarval_from_long(size, mode);
684 tv_index = tarval_div(tmp, sz);
685 tmp = tarval_mod(tmp, sz);
687 if (tv_index == tarval_bad || tmp == tarval_bad)
690 assert(get_array_n_dimensions(tp) == 1 && "multiarrays not implemented");
691 bound = get_array_lower_bound(tp, 0);
692 tlower = computed_value(bound);
693 bound = get_array_upper_bound(tp, 0);
694 tupper = computed_value(bound);
696 if (tlower == tarval_bad || tupper == tarval_bad)
699 if (tarval_cmp(tv_index, tlower) & pn_Cmp_Lt)
701 if (tarval_cmp(tupper, tv_index) & pn_Cmp_Lt)
704 /* ok, bounds check finished */
707 if (! tarval_is_null(tmp)) {
708 /* access to some struct/union member */
712 /* should be at least ONE array */
716 res = rec_get_accessed_path(ptr, depth + idx);
720 path_len = get_compound_graph_path_length(res);
721 pos = path_len - depth - idx;
723 for (ent = field;;) {
725 tarval *sz, *tv_index;
728 tp = get_entity_type(ent);
729 if (! is_Array_type(tp))
731 ent = get_array_element_entity(tp);
732 set_compound_graph_path_node(res, pos, ent);
734 size = get_type_size_bytes(get_entity_type(ent));
735 sz = new_tarval_from_long(size, mode);
737 tv_index = tarval_div(tv, sz);
738 tv = tarval_mod(tv, sz);
740 /* worked above, should work again */
741 assert(tv_index != tarval_bad && tv != tarval_bad);
743 /* bounds already checked above */
744 index = get_tarval_long(tv_index);
745 set_compound_graph_path_array_index(res, pos, index);
748 } else if (is_Sub(ptr)) {
749 ir_node *l = get_Sub_left(ptr);
750 ir_node *r = get_Sub_right(ptr);
753 tv = get_Const_tarval(r);
758 } /* rec_get_accessed_path */
761 * Returns an access path or NULL. The access path is only
762 * valid, if the graph is in phase_high and _no_ address computation is used.
764 static compound_graph_path *get_accessed_path(ir_node *ptr) {
765 compound_graph_path *gr = rec_get_accessed_path(ptr, 0);
767 } /* get_accessed_path */
769 typedef struct path_entry {
771 struct path_entry *next;
775 static ir_node *rec_find_compound_ent_value(ir_node *ptr, path_entry *next) {
776 path_entry entry, *p;
777 ir_entity *ent, *field;
778 ir_initializer_t *initializer;
784 if (is_SymConst(ptr)) {
786 ent = get_SymConst_entity(ptr);
787 initializer = get_entity_initializer(ent);
788 for (p = next; p != NULL;) {
789 if (initializer->kind != IR_INITIALIZER_COMPOUND)
791 n = get_initializer_compound_n_entries(initializer);
792 tp = get_entity_type(ent);
794 if (is_Array_type(tp)) {
795 ent = get_array_element_entity(tp);
800 initializer = get_initializer_compound_value(initializer, 0);
804 if (p->index >= (int) n)
806 initializer = get_initializer_compound_value(initializer, p->index);
811 tp = get_entity_type(ent);
812 while (is_Array_type(tp)) {
813 ent = get_array_element_entity(tp);
814 tp = get_entity_type(ent);
816 n = get_initializer_compound_n_entries(initializer);
819 initializer = get_initializer_compound_value(initializer, 0);
822 switch (initializer->kind) {
823 case IR_INITIALIZER_CONST:
824 return get_initializer_const_value(initializer);
825 case IR_INITIALIZER_TARVAL:
826 case IR_INITIALIZER_NULL:
830 } else if (is_Sel(ptr)) {
831 entry.ent = field = get_Sel_entity(ptr);
832 tp = get_entity_owner(field);
833 if (is_Array_type(tp)) {
834 assert(get_Sel_n_indexs(ptr) == 1 && "multi dim arrays not implemented");
835 entry.index = get_Sel_array_index_long(ptr, 0) - get_array_lower_bound_int(tp, 0);
837 int i, n_members = get_compound_n_members(tp);
838 for (i = 0; i < n_members; ++i) {
839 if (get_compound_member(tp, i) == field)
842 if (i >= n_members) {
843 /* not found: should NOT happen */
848 return rec_find_compound_ent_value(get_Sel_ptr(ptr), &entry);
849 } else if (is_Add(ptr)) {
850 ir_node *l = get_Add_left(ptr);
851 ir_node *r = get_Add_right(ptr);
857 tv = get_Const_tarval(r);
860 tv = get_Const_tarval(l);
863 mode = get_tarval_mode(tv);
865 /* ptr must be a Sel or a SymConst, this was checked in find_constant_entity() */
867 field = get_Sel_entity(ptr);
869 field = get_SymConst_entity(ptr);
872 /* count needed entries */
874 for (ent = field;;) {
875 tp = get_entity_type(ent);
876 if (! is_Array_type(tp))
878 ent = get_array_element_entity(tp);
881 /* should be at least ONE entry */
885 /* allocate the right number of entries */
886 NEW_ARR_A(path_entry, p, pos);
890 for (ent = field;;) {
892 tarval *sz, *tv_index, *tlower, *tupper;
896 tp = get_entity_type(ent);
897 if (! is_Array_type(tp))
899 ent = get_array_element_entity(tp);
901 p[pos].next = &p[pos + 1];
903 size = get_type_size_bytes(get_entity_type(ent));
904 sz = new_tarval_from_long(size, mode);
906 tv_index = tarval_div(tv, sz);
907 tv = tarval_mod(tv, sz);
909 if (tv_index == tarval_bad || tv == tarval_bad)
912 assert(get_array_n_dimensions(tp) == 1 && "multiarrays not implemented");
913 bound = get_array_lower_bound(tp, 0);
914 tlower = computed_value(bound);
915 bound = get_array_upper_bound(tp, 0);
916 tupper = computed_value(bound);
918 if (tlower == tarval_bad || tupper == tarval_bad)
921 if (tarval_cmp(tv_index, tlower) & pn_Cmp_Lt)
923 if (tarval_cmp(tupper, tv_index) & pn_Cmp_Lt)
926 /* ok, bounds check finished */
927 index = get_tarval_long(tv_index);
928 p[pos].index = index;
931 if (! tarval_is_null(tv)) {
932 /* hmm, wrong access */
935 p[pos - 1].next = next;
936 return rec_find_compound_ent_value(ptr, p);
937 } else if (is_Sub(ptr)) {
938 ir_node *l = get_Sub_left(ptr);
939 ir_node *r = get_Sub_right(ptr);
942 tv = get_Const_tarval(r);
947 } /* rec_find_compound_ent_value */
949 static ir_node *find_compound_ent_value(ir_node *ptr) {
950 return rec_find_compound_ent_value(ptr, NULL);
951 } /* find_compound_ent_value */
954 * Mark a Load memop to be replace by a definition
956 * @param op the Load memop
958 static void mark_replace_load(memop_t *op, ir_node *def) {
960 op->flags |= FLAG_KILLED_NODE;
962 } /* mark_replace_load */
965 * Mark a Store memop to be removed.
967 * @param op the Store memop
969 static void mark_remove_store(memop_t *op) {
970 op->flags |= FLAG_KILLED_NODE;
972 } /* mark_remove_store */
975 * Update a memop for a Load.
979 static void update_Load_memop(memop_t *m) {
981 ir_node *load = m->node;
985 if (get_Load_volatility(load) == volatility_is_volatile)
986 m->flags |= FLAG_IGNORE;
988 ptr = get_Load_ptr(load);
990 m->value.address = ptr;
992 for (i = get_irn_n_outs(load) - 1; i >= 0; --i) {
993 ir_node *proj = get_irn_out(load, i);
996 /* beware of keep edges */
1000 pn = get_Proj_proj(proj);
1001 m->projs[pn] = proj;
1004 m->value.value = proj;
1005 m->value.mode = get_irn_mode(proj);
1007 case pn_Load_X_except:
1008 m->flags |= FLAG_EXCEPTION;
1013 case pn_Load_X_regular:
1016 panic("Unsupported Proj from Load %+F", proj);
1020 /* check if we can determine the entity that will be loaded */
1021 ent = find_constant_entity(ptr);
1024 allocation_static == get_entity_allocation(ent) &&
1025 visibility_external_allocated != get_entity_visibility(ent)) {
1026 /* a static allocation that is not external: there should be NO exception
1027 * when loading even if we cannot replace the load itself. */
1028 ir_node *value = NULL;
1030 /* no exception, clear the m fields as it might be checked later again */
1031 if (m->projs[pn_Load_X_except]) {
1032 exchange(m->projs[pn_Load_X_except], new_Bad());
1033 m->projs[pn_Load_X_except] = NULL;
1034 m->flags &= ~FLAG_EXCEPTION;
1037 if (m->projs[pn_Load_X_regular]) {
1038 exchange(m->projs[pn_Load_X_regular], new_r_Jmp(get_nodes_block(load)));
1039 m->projs[pn_Load_X_regular] = NULL;
1043 if (variability_constant == get_entity_variability(ent)) {
1044 if (is_atomic_entity(ent)) {
1045 /* Might not be atomic after lowering of Sels. In this case we
1046 * could also load, but it's more complicated. */
1047 /* more simpler case: we load the content of a constant value:
1048 * replace it by the constant itself */
1049 value = get_atomic_ent_value(ent);
1050 } else if (ent->has_initializer) {
1051 /* new style initializer */
1052 value = find_compound_ent_value(ptr);
1054 /* old style initializer */
1055 compound_graph_path *path = get_accessed_path(ptr);
1058 assert(is_proper_compound_graph_path(path, get_compound_graph_path_length(path)-1));
1060 value = get_compound_ent_value_by_path(ent, path);
1061 DB((dbg, LEVEL_1, " Constant access at %F%F resulted in %+F\n", ent, path, value));
1062 free_compound_graph_path(path);
1066 value = can_replace_load_by_const(load, value);
1069 if (value != NULL) {
1070 /* we completely replace the load by this value */
1071 DB((dbg, LEVEL_1, "Replacing Load %+F by constant %+F\n", m->node, value));
1072 mark_replace_load(m, value);
1077 if (m->value.value != NULL && !(m->flags & FLAG_IGNORE)) {
1078 /* only create an address if this node is NOT killed immediately or ignored */
1079 m->value.id = register_address(ptr);
1082 /* no user, KILL it */
1083 mark_replace_load(m, NULL);
1085 } /* update_Load_memop */
1088 * Update a memop for a Store.
1090 * @param m the memop
1092 static void update_Store_memop(memop_t *m) {
1094 ir_node *store = m->node;
1095 ir_node *adr = get_Store_ptr(store);
1097 if (get_Store_volatility(store) == volatility_is_volatile) {
1098 m->flags |= FLAG_IGNORE;
1100 /* only create an address if this node is NOT ignored */
1101 m->value.id = register_address(adr);
1105 m->value.address = adr;
1107 for (i = get_irn_n_outs(store) - 1; i >= 0; --i) {
1108 ir_node *proj = get_irn_out(store, i);
1111 /* beware of keep edges */
1115 pn = get_Proj_proj(proj);
1116 m->projs[pn] = proj;
1118 case pn_Store_X_except:
1119 m->flags |= FLAG_EXCEPTION;
1124 case pn_Store_X_regular:
1127 panic("Unsupported Proj from Store %+F", proj);
1130 m->value.value = get_Store_value(store);
1131 m->value.mode = get_irn_mode(m->value.value);
1132 } /* update_Store_memop */
1135 * Update a memop for a Call.
1137 * @param m the memop
1139 static void update_Call_memop(memop_t *m) {
1140 ir_node *call = m->node;
1141 unsigned prop = get_Call_memory_properties(call);
1144 if (prop & mtp_property_const) {
1145 /* A constant call did NOT use memory at all, we
1146 can kick it from the list. */
1147 } else if (prop & mtp_property_pure) {
1148 /* pure calls READ memory */
1151 m->flags = FLAG_KILL_ALL;
1153 for (i = get_irn_n_outs(call) - 1; i >= 0; --i) {
1154 ir_node *proj = get_irn_out(call, i);
1156 /* beware of keep edges */
1160 switch (get_Proj_proj(proj)) {
1161 case pn_Call_X_except:
1162 m->flags |= FLAG_EXCEPTION;
1164 case pn_Call_M_regular:
1169 } /* update_Call_memop */
1172 * Update a memop for a Div/Mod/Quot/DivMod.
1174 * @param m the memop
1176 static void update_DivOp_memop(memop_t *m) {
1177 ir_node *div = m->node;
1180 for (i = get_irn_n_outs(div) - 1; i >= 0; --i) {
1181 ir_node *proj = get_irn_out(div, i);
1183 /* beware of keep edges */
1187 switch (get_Proj_proj(proj)) {
1188 case pn_Generic_X_except:
1189 m->flags |= FLAG_EXCEPTION;
1191 case pn_Generic_M_regular:
1196 } /* update_DivOp_memop */
1199 * Update a memop for a Phi.
1201 * @param m the memop
1203 static void update_Phi_memop(memop_t *m) {
1204 /* the Phi is it's own mem */
1206 } /* update_Phi_memop */
1209 * Memory walker: collect all memory ops and build topological lists.
1211 static void collect_memops(ir_node *irn, void *ctx) {
1218 /* we can safely ignore ProjM's except the initial memory */
1219 if (irn != get_irg_initial_mem(current_ir_graph))
1223 op = alloc_memop(irn);
1224 block = get_nodes_block(irn);
1225 entry = get_block_entry(block);
1228 update_Phi_memop(op);
1229 /* Phis must be always placed first */
1230 op->next = entry->memop_forward;
1231 entry->memop_forward = op;
1232 if (entry->memop_backward == NULL)
1233 entry->memop_backward = op;
1235 switch (get_irn_opcode(irn)) {
1237 update_Load_memop(op);
1240 update_Store_memop(op);
1243 update_Call_memop(op);
1250 /* initial memory */
1255 /* we can those to find the memory edge */
1261 update_DivOp_memop(op);
1265 /* TODO: handle some builtins */
1267 /* unsupported operation */
1268 op->flags = FLAG_KILL_ALL;
1272 /* all other should be placed last */
1273 if (entry->memop_backward == NULL) {
1274 entry->memop_forward = entry->memop_backward = op;
1276 entry->memop_backward->next = op;
1277 entry->memop_backward = op;
1280 } /* collect_memops */
1283 * Find an address in the current set.
1285 * @param value the value to be searched for
1287 * @return a memop for the value or NULL if the value does
1288 * not exists in the set or cannot be converted into
1289 * the requested mode
1291 static memop_t *find_address(const value_t *value) {
1292 if (rbitset_is_set(env.curr_set, value->id)) {
1293 memop_t *res = env.curr_id_2_memop[value->id];
1295 if (res->value.mode == value->mode)
1297 /* allow hidden casts */
1298 if (get_mode_arithmetic(res->value.mode) == irma_twos_complement &&
1299 get_mode_arithmetic(value->mode) == irma_twos_complement &&
1300 get_mode_size_bits(res->value.mode) == get_mode_size_bits(value->mode))
1304 } /* find_address */
1307 * Find an address in the avail_out set.
1309 * @param bl the block
1311 static memop_t *find_address_avail(const block_t *bl, unsigned id, const ir_mode *mode) {
1312 if (rbitset_is_set(bl->avail_out, id)) {
1313 memop_t *res = bl->id_2_memop_avail[id];
1315 if (res->value.mode == mode)
1317 /* allow hidden casts */
1318 if (get_mode_arithmetic(res->value.mode) == irma_twos_complement &&
1319 get_mode_arithmetic(mode) == irma_twos_complement &&
1320 get_mode_size_bits(res->value.mode) == get_mode_size_bits(mode))
1324 } /* find_address_avail */
1327 * Kill all addresses from the current set.
1329 static void kill_all(void) {
1330 rbitset_clear_all(env.curr_set, env.rbs_size);
1333 rbitset_set(env.curr_set, env.rbs_size - 1);
1337 * Kill memops that are not alias free due to a Store value from the current set.
1339 * @param value the Store value
1341 static void kill_memops(const value_t *value) {
1342 unsigned end = env.rbs_size - 1;
1345 for (pos = rbitset_next(env.curr_set, 0, 1); pos < end; pos = rbitset_next(env.curr_set, pos + 1, 1)) {
1346 memop_t *op = env.curr_id_2_memop[pos];
1348 if (ir_no_alias != get_alias_relation(current_ir_graph, value->address, value->mode,
1349 op->value.address, op->value.mode)) {
1350 rbitset_clear(env.curr_set, pos);
1351 env.curr_id_2_memop[pos] = NULL;
1352 DB((dbg, LEVEL_2, "KILLING %+F because of possible alias address %+F\n", op->node, value->address));
1358 * Add the value of a memop to the current set.
1360 * @param op the memory op
1362 static void add_memop(memop_t *op) {
1363 rbitset_set(env.curr_set, op->value.id);
1364 env.curr_id_2_memop[op->value.id] = op;
1368 * Add the value of a memop to the avail_out set.
1370 * @param bl the block
1371 * @param op the memory op
1373 static void add_memop_avail(block_t *bl, memop_t *op) {
1374 rbitset_set(bl->avail_out, op->value.id);
1375 bl->id_2_memop_avail[op->value.id] = op;
1376 } /* add_memop_avail */
1379 * Check, if we can convert a value of one mode to another mode
1380 * without changing the representation of bits.
1382 * @param from the original mode
1383 * @param to the destination mode
1385 static int can_convert_to(const ir_mode *from, const ir_mode *to) {
1386 if (get_mode_arithmetic(from) == irma_twos_complement &&
1387 get_mode_arithmetic(to) == irma_twos_complement &&
1388 get_mode_size_bits(from) == get_mode_size_bits(to))
1391 } /* can_convert_to */
1394 * Add a Conv to the requested mode if needed.
1396 * @param irn the IR-node to convert
1397 * @param mode the destination mode
1399 * @return the possible converted node or NULL
1400 * if the conversion is not possible
1402 static ir_node *conv_to(ir_node *irn, ir_mode *mode) {
1403 ir_mode *other = get_irn_mode(irn);
1404 if (other != mode) {
1405 /* different modes: check if conversion is possible without changing the bits */
1406 if (can_convert_to(other, mode)) {
1407 ir_node *block = get_nodes_block(irn);
1408 return new_r_Conv(block, irn, mode);
1410 /* otherwise not possible ... yet */
1417 * Update the address of an value if this address was a load result
1418 * and the load is killed now.
1420 * @param value the value whose address is updated
1422 static void update_address(value_t *value) {
1423 if (is_Proj(value->address)) {
1424 ir_node *load = get_Proj_pred(value->address);
1426 if (is_Load(load)) {
1427 const memop_t *op = get_irn_memop(load);
1429 if (op->flags & FLAG_KILLED_NODE)
1430 value->address = op->replace;
1433 } /* update_address */
1436 * Do forward dataflow analysis on the given block and calculate the
1437 * GEN and KILL in the current (avail) set.
1439 * @param bl the block
1441 static void calc_gen_kill_avail(block_t *bl) {
1445 for (op = bl->memop_forward; op != NULL; op = op->next) {
1446 switch (get_irn_opcode(op->node)) {
1454 if (! (op->flags & (FLAG_KILLED_NODE|FLAG_IGNORE))) {
1455 /* do we have this already? */
1458 update_address(&op->value);
1459 other = find_address(&op->value);
1460 if (other != NULL && other != op) {
1461 def = conv_to(other->value.value, op->value.mode);
1463 #ifdef DEBUG_libfirm
1464 if (is_Store(other->node)) {
1466 DB((dbg, LEVEL_1, "RAW %+F <- %+F(%+F)\n", op->node, def, other->node));
1469 DB((dbg, LEVEL_1, "RAR %+F <- %+F(%+F)\n", op->node, def, other->node));
1472 mark_replace_load(op, def);
1473 /* do NOT change the memop table */
1477 /* add this value */
1482 if (! (op->flags & FLAG_KILLED_NODE)) {
1483 /* do we have this store already */
1486 update_address(&op->value);
1487 other = find_address(&op->value);
1488 if (other != NULL) {
1489 if (is_Store(other->node)) {
1490 if (op != other && !(other->flags & FLAG_IGNORE) &&
1491 get_nodes_block(other->node) == get_nodes_block(op->node)) {
1493 * A WAW in the same block we can kick the first store.
1494 * This is a shortcut: we know that the second Store will be anticipated
1497 DB((dbg, LEVEL_1, "WAW %+F <- %+F\n", other->node, op->node));
1498 mark_remove_store(other);
1499 /* FIXME: a Load might be get freed due to this killed store */
1501 } else if (other->value.value == op->value.value && !(op->flags & FLAG_IGNORE)) {
1503 DB((dbg, LEVEL_1, "WAR %+F <- %+F\n", op->node, other->node));
1504 mark_remove_store(op);
1505 /* do NOT change the memop table */
1509 /* KILL all possible aliases */
1510 kill_memops(&op->value);
1511 /* add this value */
1516 if (op->flags & FLAG_KILL_ALL)
1520 } /* calc_gen_kill_avail */
1522 #define BYTE_SIZE(x) (((x) + 7) >> 3)
1525 * Do forward dataflow analysis on a given block to calculate the avail_out set
1526 * for this block only.
1528 * @param block the block
1530 static void forward_avail(block_t *bl) {
1531 /* fill the data from the current block */
1532 env.curr_id_2_memop = bl->id_2_memop_avail;
1533 env.curr_set = bl->avail_out;
1535 calc_gen_kill_avail(bl);
1536 dump_curr(bl, "Avail_out");
1537 } /* forward_avail */
1540 * Do backward dataflow analysis on a given block to calculate the antic set
1541 * of Loaded addresses.
1543 * @param bl the block
1545 * @return non-zero if the set has changed since last iteration
1547 static int backward_antic(block_t *bl) {
1549 ir_node *block = bl->block;
1550 int n = get_Block_n_cfg_outs(block);
1553 ir_node *succ = get_Block_cfg_out(block, 0);
1554 block_t *succ_bl = get_block_entry(succ);
1555 int pred_pos = get_Block_cfgpred_pos(succ, block);
1556 unsigned end = env.rbs_size - 1;
1561 if (bl->trans_results == NULL) {
1562 /* allocate the translate cache */
1563 unsigned size = env.curr_adr_id * sizeof(bl->trans_results[0]);
1564 bl->trans_results = obstack_alloc(&env.obst, size);
1565 memset(bl->trans_results, 0, size);
1568 /* check for partly redundant values */
1569 for (pos = rbitset_next(succ_bl->anticL_in, 0, 1);
1571 pos = rbitset_next(succ_bl->anticL_in, pos + 1, 1)) {
1573 * do Phi-translation here: Note that at this point the nodes are
1574 * not changed, so we can safely cache the results.
1575 * However: Loads of Load results ARE bad, because we have no way
1576 to translate them yet ...
1578 memop_t *op = bl->trans_results[pos];
1580 /* not yet translated */
1581 ir_node *adr, *trans_adr;
1583 op = succ_bl->id_2_memop_antic[pos];
1584 adr = op->value.address;
1586 trans_adr = phi_translate(adr, succ, pred_pos);
1587 if (trans_adr != adr) {
1588 /* create a new entry for the translated one */
1591 new_op = alloc_memop(NULL);
1592 new_op->value.address = trans_adr;
1593 new_op->value.id = register_address(trans_adr);
1594 new_op->value.mode = op->value.mode;
1595 new_op->node = op->node; /* we need the node to decide if Load/Store */
1596 new_op->flags = op->flags;
1598 bl->trans_results[pos] = new_op;
1602 env.curr_id_2_memop[op->value.id] = op;
1603 rbitset_set(env.curr_set, op->value.id);
1606 ir_node *succ = get_Block_cfg_out(block, 0);
1607 block_t *succ_bl = get_block_entry(succ);
1610 rbitset_copy(env.curr_set, succ_bl->anticL_in, env.rbs_size);
1611 memcpy(env.curr_id_2_memop, succ_bl->id_2_memop_antic, env.rbs_size * sizeof(env.curr_id_2_memop[0]));
1613 /* Hmm: probably we want kill merges of Loads ans Stores here */
1614 for (i = n - 1; i > 0; --i) {
1615 ir_node *succ = get_Block_cfg_out(bl->block, i);
1616 block_t *succ_bl = get_block_entry(succ);
1618 rbitset_and(env.curr_set, succ_bl->anticL_in, env.rbs_size);
1621 /* block ends with a noreturn call */
1625 dump_curr(bl, "AnticL_out");
1627 for (op = bl->memop_backward; op != NULL; op = op->prev) {
1628 switch (get_irn_opcode(op->node)) {
1636 if (! (op->flags & (FLAG_KILLED_NODE|FLAG_IGNORE))) {
1642 if (! (op->flags & FLAG_KILLED_NODE)) {
1643 /* a Store: check which memops must be killed */
1644 kill_memops(&op->value);
1648 if (op->flags & FLAG_KILL_ALL)
1653 memcpy(bl->id_2_memop_antic, env.curr_id_2_memop, env.rbs_size * sizeof(env.curr_id_2_memop[0]));
1654 if (! rbitset_equal(bl->anticL_in, env.curr_set, env.rbs_size)) {
1656 rbitset_copy(bl->anticL_in, env.curr_set, env.rbs_size);
1657 dump_curr(bl, "AnticL_in*");
1660 dump_curr(bl, "AnticL_in");
1662 } /* backward_antic */
1665 * Replace a Load memop by a already known value.
1667 * @param op the Load memop
1669 static void replace_load(memop_t *op) {
1670 ir_node *load = op->node;
1671 ir_node *def = skip_Id(op->replace);
1676 DB((dbg, LEVEL_1, "Replacing %+F by definition %+F\n", load, is_Proj(def) ? get_Proj_pred(def) : def));
1678 if (op->flags & FLAG_EXCEPTION) {
1679 /* bad: this node is unused and executed for exception only */
1680 DB((dbg, LEVEL_1, "Unused %+F executed for exception only ...\n", load));
1683 DB((dbg, LEVEL_1, "Killing unused %+F\n", load));
1686 if (op->mem != NULL) {
1687 /* in rare cases a Load might have NO memory */
1688 exchange(op->mem, get_Load_mem(load));
1690 proj = op->projs[pn_Load_res];
1692 mode = get_irn_mode(proj);
1693 if (get_irn_mode(def) != mode) {
1695 dbg_info *db = get_irn_dbg_info(load);
1696 ir_node *block = get_nodes_block(proj);
1697 def = new_rd_Conv(db, block, def, mode);
1699 exchange(proj, def);
1701 proj = op->projs[pn_Load_X_except];
1703 exchange(proj, new_Bad());
1705 proj = op->projs[pn_Load_X_regular];
1707 exchange(proj, new_r_Jmp(get_nodes_block(load)));
1709 } /* replace_load */
1712 * Remove a Store memop.
1714 * @param op the Store memop
1716 static void remove_store(memop_t *op) {
1717 ir_node *store = op->node;
1720 DB((dbg, LEVEL_1, "Removing %+F\n", store));
1722 if (op->mem != NULL) {
1723 /* in rare cases a Store might have no memory */
1724 exchange(op->mem, get_Store_mem(store));
1726 proj = op->projs[pn_Store_X_except];
1728 exchange(proj, new_Bad());
1730 proj = op->projs[pn_Store_X_regular];
1732 exchange(proj, new_r_Jmp(get_nodes_block(store)));
1734 } /* remove_store */
1738 * Do all necessary replacements for a given block.
1740 * @param bl the block
1742 static void do_replacements(block_t *bl) {
1745 for (op = bl->memop_forward; op != NULL; op = op->next) {
1746 if (op->flags & FLAG_KILLED_NODE) {
1747 switch (get_irn_opcode(op->node)) {
1757 } /* do_replacements */
1760 * Calculate the Avail_out sets for all basic blocks.
1762 static void calcAvail(void) {
1763 memop_t **tmp_memop = env.curr_id_2_memop;
1764 unsigned *tmp_set = env.curr_set;
1767 /* calculate avail_out */
1768 DB((dbg, LEVEL_2, "Calculate Avail_out\n"));
1770 /* iterate over all blocks in in any order, skip the start block */
1771 for (bl = env.forward->forward_next; bl != NULL; bl = bl->forward_next) {
1775 /* restore the current sets */
1776 env.curr_id_2_memop = tmp_memop;
1777 env.curr_set = tmp_set;
1781 * Calculate the Antic_in sets for all basic blocks.
1783 static void calcAntic(void) {
1786 /* calculate antic_out */
1787 DB((dbg, LEVEL_2, "Calculate Antic_in\n"));
1792 DB((dbg, LEVEL_2, "Iteration %d:\n=========\n", i));
1796 /* over all blocks in reverse post order */
1797 for (bl = env.backward->backward_next; bl != NULL; bl = bl->backward_next) {
1798 need_iter |= backward_antic(bl);
1801 } while (need_iter);
1802 DB((dbg, LEVEL_2, "Get anticipated Load set after %d iterations\n", i));
1806 * Return the node representing the last memory in a block.
1808 * @param bl the block
1810 static ir_node *find_last_memory(block_t *bl) {
1812 if (bl->memop_backward != NULL) {
1813 return bl->memop_backward->mem;
1815 /* if there is NO memory in this block, go to the dominator */
1816 bl = get_block_entry(get_Block_idom(bl->block));
1818 } /* find_last_memory */
1821 * Reroute all memory users of old memory
1822 * to a new memory IR-node.
1824 * @param omem the old memory IR-node
1825 * @param nmem the new memory IR-node
1827 static void reroute_all_mem_users(ir_node *omem, ir_node *nmem) {
1830 for (i = get_irn_n_outs(omem) - 1; i >= 0; --i) {
1832 ir_node *user = get_irn_out_ex(omem, i, &n_pos);
1834 set_irn_n(user, n_pos, nmem);
1837 /* all edges previously point to omem now point to nmem */
1838 nmem->out = omem->out;
1839 } /* reroute_all_mem_users */
1842 * Reroute memory users of old memory that are dominated by a given block
1843 * to a new memory IR-node.
1845 * @param omem the old memory IR-node
1846 * @param nmem the new memory IR-node
1847 * @param pass_bl the block the memory must pass
1849 static void reroute_mem_through(ir_node *omem, ir_node *nmem, ir_node *pass_bl) {
1850 int i, j, n = get_irn_n_outs(omem);
1851 ir_def_use_edge *edges = NEW_ARR_D(ir_def_use_edge, &env.obst, n + 1);
1853 for (i = j = 0; i < n; ++i) {
1855 ir_node *user = get_irn_out_ex(omem, i, &n_pos);
1856 ir_node *use_bl = get_nodes_block(user);
1860 use_bl = get_Block_cfgpred_block(use_bl, n_pos);
1862 if (block_dominates(pass_bl, use_bl)) {
1863 /* found an user that is dominated */
1865 edges[j].pos = n_pos;
1866 edges[j].use = user;
1868 set_irn_n(user, n_pos, nmem);
1872 /* Modify the out structure: we create a new out edge array on our
1873 temporary obstack here. This should be no problem, as we invalidate the edges
1874 at the end either. */
1875 /* first entry is used for the length */
1878 } /* reroute_mem_through */
1881 * insert Loads, making partly redundant Loads fully redundant
1883 static int insert_Load(block_t *bl) {
1884 ir_node *block = bl->block;
1885 int i, n = get_Block_n_cfgpreds(block);
1886 unsigned end = env.rbs_size - 1;
1889 DB((dbg, LEVEL_3, "processing %+F\n", block));
1892 /* might still happen for an unreachable block (end for instance) */
1900 NEW_ARR_A(ir_node *, ins, n);
1902 rbitset_set_all(env.curr_set, env.rbs_size);
1904 /* More than one predecessors, calculate the join for all avail_outs ignoring unevaluated
1905 Blocks. These put in Top anyway. */
1906 for (i = n - 1; i >= 0; --i) {
1907 ir_node *pred = skip_Proj(get_Block_cfgpred(block, i));
1908 ir_node *blk = get_nodes_block(pred);
1911 pred_bl = get_block_entry(blk);
1912 rbitset_and(env.curr_set, pred_bl->avail_out, env.rbs_size);
1914 if (is_Load(pred) || is_Store(pred)) {
1915 /* We reached this block by an exception from a Load or Store:
1916 * the memop creating the exception was NOT completed than, kill it
1918 memop_t *exc_op = get_irn_memop(pred);
1919 rbitset_clear(env.curr_set, exc_op->value.id);
1924 * Ensure that all values are in the map: build Phi's if necessary:
1925 * Note: the last bit is the sentinel and ALWAYS set, so start with -2.
1927 for (pos = env.rbs_size - 2; pos >= 0; --pos) {
1928 if (! rbitset_is_set(env.curr_set, pos))
1929 env.curr_id_2_memop[pos] = NULL;
1931 ir_node *pred = get_Block_cfgpred_block(bl->block, 0);
1932 block_t *pred_bl = get_block_entry(pred);
1934 memop_t *first = NULL;
1937 for (i = 0; i < n; ++i) {
1940 pred = get_Block_cfgpred_block(bl->block, i);
1941 pred_bl = get_block_entry(pred);
1943 mop = pred_bl->id_2_memop_avail[pos];
1944 if (first == NULL) {
1946 ins[0] = first->value.value;
1947 mode = get_irn_mode(ins[0]);
1949 /* no Phi needed so far */
1950 env.curr_id_2_memop[pos] = first;
1952 ins[i] = conv_to(mop->value.value, mode);
1953 if (ins[i] != ins[0]) {
1954 if (ins[i] == NULL) {
1955 /* conversion failed */
1956 env.curr_id_2_memop[pos] = NULL;
1957 rbitset_clear(env.curr_set, pos);
1966 ir_node *phi = new_r_Phi(bl->block, n, ins, mode);
1967 memop_t *phiop = alloc_memop(phi);
1969 phiop->value = first->value;
1970 phiop->value.value = phi;
1972 /* no need to link it in, as it is a DATA phi */
1974 env.curr_id_2_memop[pos] = phiop;
1976 DB((dbg, LEVEL_3, "Created new %+F on merging value for address %+F\n", phi, first->value.address));
1981 /* only one predecessor, simply copy the map */
1982 ir_node *pred = get_Block_cfgpred_block(bl->block, 0);
1983 block_t *pred_bl = get_block_entry(pred);
1985 rbitset_copy(env.curr_set, pred_bl->avail_out, env.rbs_size);
1987 memcpy(env.curr_id_2_memop, pred_bl->id_2_memop_avail, env.rbs_size * sizeof(bl->id_2_memop_avail[0]));
1991 /* check for partly redundant values */
1992 for (pos = rbitset_next(bl->anticL_in, 0, 1);
1994 pos = rbitset_next(bl->anticL_in, pos + 1, 1)) {
1995 memop_t *op = bl->id_2_memop_antic[pos];
1996 int have_some, all_same;
1999 if (rbitset_is_set(env.curr_set, pos)) {
2004 assert(is_Load(op->node));
2006 DB((dbg, LEVEL_3, "anticipated %+F\n", op->node));
2011 for (i = n - 1; i >= 0; --i) {
2012 ir_node *pred = get_Block_cfgpred_block(block, i);
2013 block_t *pred_bl = get_block_entry(pred);
2014 ir_mode *mode = op->value.mode;
2018 adr = phi_translate(op->value.address, block, i);
2019 DB((dbg, LEVEL_3, ".. using address %+F in pred %d\n", adr, i));
2020 e = find_address_avail(pred_bl, register_address(adr), mode);
2022 ir_node *ef_block = get_nodes_block(adr);
2023 if (! block_dominates(ef_block, pred)) {
2024 /* cannot place a copy here */
2026 DB((dbg, LEVEL_3, "%+F cannot be moved into predecessor %+F\n", op->node, pred));
2029 DB((dbg, LEVEL_3, "%+F is not available in predecessor %+F\n", op->node, pred));
2030 pred_bl->avail = NULL;
2033 if (e->value.mode != mode && !can_convert_to(e->value.mode, mode)) {
2034 /* cannot create a Phi due to different modes */
2040 DB((dbg, LEVEL_3, "%+F is available for %+F in predecessor %+F\n", e->node, op->node, pred));
2043 else if (first != e->node)
2047 if (have_some && !all_same) {
2048 ir_mode *mode = op->value.mode;
2052 NEW_ARR_A(ir_node *, in, n);
2054 for (i = n - 1; i >= 0; --i) {
2055 ir_node *pred = get_Block_cfgpred_block(block, i);
2056 block_t *pred_bl = get_block_entry(pred);
2058 if (pred_bl->avail == NULL) {
2059 /* create a new Load here and make to make it fully redundant */
2060 dbg_info *db = get_irn_dbg_info(op->node);
2061 ir_node *last_mem = find_last_memory(pred_bl);
2062 ir_node *load, *def, *adr;
2065 assert(last_mem != NULL);
2066 adr = phi_translate(op->value.address, block, i);
2067 load = new_rd_Load(db, pred, last_mem, adr, mode, cons_none);
2068 def = new_r_Proj(pred, load, mode, pn_Load_res);
2069 DB((dbg, LEVEL_1, "Created new %+F in %+F for party redundant %+F\n", load, pred, op->node));
2071 new_op = alloc_memop(load);
2072 new_op->mem = new_r_Proj(pred, load, mode_M, pn_Load_M);
2073 new_op->value.address = adr;
2074 new_op->value.id = op->value.id;
2075 new_op->value.mode = mode;
2076 new_op->value.value = def;
2078 new_op->projs[pn_Load_M] = new_op->mem;
2079 new_op->projs[pn_Load_res] = def;
2081 new_op->prev = pred_bl->memop_backward;
2082 if (pred_bl->memop_backward != NULL)
2083 pred_bl->memop_backward->next = new_op;
2085 pred_bl->memop_backward = new_op;
2087 if (pred_bl->memop_forward == NULL)
2088 pred_bl->memop_forward = new_op;
2090 if (get_nodes_block(last_mem) == pred) {
2091 /* We have add a new last memory op in pred block.
2092 If pred had already a last mem, reroute all memory
2094 reroute_all_mem_users(last_mem, new_op->mem);
2096 /* reroute only those memory going through the pre block */
2097 reroute_mem_through(last_mem, new_op->mem, pred);
2100 /* we added this load at the end, so it will be avail anyway */
2101 add_memop_avail(pred_bl, new_op);
2102 pred_bl->avail = new_op;
2104 in[i] = conv_to(pred_bl->avail->value.value, mode);
2106 phi = new_r_Phi(block, n, in, mode);
2107 DB((dbg, LEVEL_1, "Created new %+F in %+F for now redundant %+F\n", phi, block, op->node));
2109 phi_op = clone_memop_phi(op, phi);
2115 /* recalculate avail by gen and kill */
2116 calc_gen_kill_avail(bl);
2118 /* always update the map after gen/kill, as values might have been changed due to RAR/WAR/WAW */
2119 memcpy(bl->id_2_memop_avail, env.curr_id_2_memop, env.rbs_size * sizeof(env.curr_id_2_memop[0]));
2121 if (!rbitset_equal(bl->avail_out, env.curr_set, env.rbs_size)) {
2122 /* the avail set has changed */
2123 rbitset_copy(bl->avail_out, env.curr_set, env.rbs_size);
2124 dump_curr(bl, "Avail_out*");
2127 dump_curr(bl, "Avail_out");
2132 * Insert Loads upwards.
2134 static void insert_Loads_upwards(void) {
2138 /* recalculate antic_out and insert Loads */
2139 DB((dbg, LEVEL_2, "Inserting Loads\n"));
2143 DB((dbg, LEVEL_2, "Iteration %d:\n=========\n", i));
2147 /* over all blocks in reverse post order, skip the start block */
2148 for (bl = env.forward->forward_next; bl != NULL; bl = bl->forward_next) {
2149 need_iter |= insert_Load(bl);
2152 } while (need_iter);
2154 DB((dbg, LEVEL_2, "Finished Load inserting after %d iterations\n", i));
2155 } /* insert_Loads_upwards */
2158 * Kill unreachable control flow.
2160 * @param irg the graph to operate on
2162 static void kill_unreachable_blocks(ir_graph *irg) {
2167 NEW_ARR_A(ir_node *, ins, env.max_cfg_preds);
2169 for (bl = env.forward; bl != NULL; bl = bl->forward_next) {
2170 ir_node *block = bl->block;
2173 assert(get_Block_mark(block));
2175 n = get_Block_n_cfgpreds(block);
2177 for (i = j = 0; i < n; ++i) {
2178 ir_node *pred = get_Block_cfgpred(block, i);
2184 pred_bl = get_nodes_block(skip_Proj(pred));
2185 if (! get_Block_mark(pred_bl))
2191 ir_node *phi, *next;
2193 /* some unreachable blocks detected */
2196 DB((dbg, LEVEL_1, "Killing dead block predecessors on %+F\n", block));
2198 set_irn_in(block, j, ins);
2200 /* shorten all Phi nodes */
2201 for (phi = get_Block_phis(block); phi != NULL; phi = next) {
2202 next = get_Phi_next(phi);
2204 for (i = k = 0; i < n; ++i) {
2205 ir_node *pred = get_Block_cfgpred_block(block, i);
2210 if (! get_Block_mark(pred))
2213 ins[k++] = get_Phi_pred(phi, i);
2216 exchange(phi, ins[0]);
2218 set_irn_in(phi, k, ins);
2225 /* kick keep alives */
2226 ir_node *end = get_irg_end(irg);
2227 int i, j, n = get_End_n_keepalives(end);
2229 NEW_ARR_A(ir_node *, ins, n);
2231 for (i = j = 0; i < n; ++i) {
2232 ir_node *ka = get_End_keepalive(end, i);
2240 ka_bl = get_nodes_block(skip_Proj(ka));
2241 if (get_Block_mark(ka_bl))
2245 set_End_keepalives(end, j, ins);
2249 /* this transformation do NOT invalidate the dominance */
2251 } /* kill_unreachable_blocks */
2253 int opt_ldst(ir_graph *irg) {
2255 ir_graph *rem = current_ir_graph;
2257 current_ir_graph = irg;
2259 FIRM_DBG_REGISTER(dbg, "firm.opt.ldst");
2260 // firm_dbg_set_mask(dbg, -1);
2262 DB((dbg, LEVEL_1, "\nDoing Load/Store optimization on %+F\n", irg));
2264 /* we need landing pads */
2265 remove_critical_cf_edges(irg);
2267 // dump_ir_block_graph(irg, "-XXX");
2269 if (get_opt_alias_analysis()) {
2270 assure_irg_entity_usage_computed(irg);
2271 assure_irp_globals_entity_usage_computed();
2274 obstack_init(&env.obst);
2275 ir_nodemap_init(&env.adr_map);
2278 env.backward = NULL;
2279 env.curr_adr_id = 0;
2281 env.max_cfg_preds = 0;
2283 env.start_bl = get_irg_start_block(irg);
2284 env.end_bl = get_irg_end_block(irg);
2285 #ifdef DEBUG_libfirm
2286 env.id_2_address = NEW_ARR_F(ir_node *, 0);
2289 assure_irg_outs(irg);
2291 ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK | IR_RESOURCE_BLOCK_MARK);
2293 /* first step: allocate block entries. Note that some blocks might be
2294 unreachable here. Using the normal walk ensures that ALL blocks are initialized. */
2295 irg_walk_graph(irg, prepare_blocks, link_phis, NULL);
2297 /* produce an inverse post-order list for the CFG: this links only reachable
2299 irg_out_block_walk(get_irg_start_block(irg), NULL, inverse_post_order, NULL);
2301 if (! get_Block_mark(env.end_bl)) {
2303 * The end block is NOT reachable due to endless loops
2304 * or no_return calls.
2305 * Place the end block last.
2306 * env.backward points to the last block in the list for this purpose.
2308 env.backward->forward_next = get_block_entry(env.end_bl);
2310 set_Block_mark(env.end_bl, 1);
2313 /* KILL unreachable blocks: these disturb the data flow analysis */
2314 kill_unreachable_blocks(irg);
2318 /* second step: find and sort all memory ops */
2319 walk_memory_irg(irg, collect_memops, NULL, NULL);
2321 #ifdef DEBUG_libfirm
2322 /* check that the backward map is correct */
2323 assert((unsigned)ARR_LEN(env.id_2_address) == env.curr_adr_id);
2326 if (env.n_mem_ops == 0) {
2331 /* create the backward links. */
2332 env.backward = NULL;
2333 irg_block_walk_graph(irg, NULL, collect_backward, NULL);
2335 /* link the end block in */
2336 bl = get_block_entry(env.end_bl);
2337 bl->backward_next = env.backward;
2340 /* check that we really start with the start / end block */
2341 assert(env.forward->block == env.start_bl);
2342 assert(env.backward->block == env.end_bl);
2344 /* create address sets: for now, only the existing addresses are allowed plus one
2345 needed for the sentinel */
2346 env.rbs_size = env.curr_adr_id + 1;
2348 /* create the current set */
2349 env.curr_set = rbitset_obstack_alloc(&env.obst, env.rbs_size);
2350 rbitset_set(env.curr_set, env.rbs_size - 1);
2351 env.curr_id_2_memop = NEW_ARR_D(memop_t *, &env.obst, env.rbs_size);
2352 memset(env.curr_id_2_memop, 0, env.rbs_size * sizeof(env.curr_id_2_memop[0]));
2354 for (bl = env.forward; bl != NULL; bl = bl->forward_next) {
2355 /* set sentinel bits */
2356 bl->avail_out = rbitset_obstack_alloc(&env.obst, env.rbs_size);
2357 rbitset_set(bl->avail_out, env.rbs_size - 1);
2359 bl->id_2_memop_avail = NEW_ARR_D(memop_t *, &env.obst, env.rbs_size);
2360 memset(bl->id_2_memop_avail, 0, env.rbs_size * sizeof(bl->id_2_memop_avail[0]));
2362 bl->anticL_in = rbitset_obstack_alloc(&env.obst, env.rbs_size);
2363 rbitset_set(bl->anticL_in, env.rbs_size - 1);
2365 bl->id_2_memop_antic = NEW_ARR_D(memop_t *, &env.obst, env.rbs_size);
2366 memset(bl->id_2_memop_antic, 0, env.rbs_size * sizeof(bl->id_2_memop_antic[0]));
2369 // dump_block_list(&env);
2370 (void) dump_block_list;
2375 insert_Loads_upwards();
2378 /* over all blocks in reverse post order */
2379 for (bl = env.forward; bl != NULL; bl = bl->forward_next) {
2380 do_replacements(bl);
2383 /* not only invalidate but free them. We might allocate new out arrays
2384 on our obstack which will be deleted yet. */
2386 set_irg_entity_usage_state(irg, ir_entity_usage_not_computed);
2390 ir_free_resources(irg, IR_RESOURCE_IRN_LINK | IR_RESOURCE_BLOCK_MARK);
2391 ir_nodemap_destroy(&env.adr_map);
2392 obstack_free(&env.obst, NULL);
2394 // dump_ir_block_graph(irg, "-YYY");
2396 #ifdef DEBUG_libfirm
2397 DEL_ARR_F(env.id_2_address);
2400 current_ir_graph = rem;
2401 return env.changed != 0;