2 * Copyright (C) 1995-2011 University of Karlsruhe. All right reserved.
4 * This file is part of libFirm.
6 * This file may be distributed and/or modified under the terms of the
7 * GNU General Public License version 2 as published by the Free Software
8 * Foundation and appearing in the file LICENSE.GPL included in the
9 * packaging of this file.
11 * Licensees holding valid libFirm Professional Edition licenses may use
12 * this file in accordance with the libFirm Commercial License.
13 * Agreement provided with the Software.
15 * This file is provided AS IS with NO WARRANTY OF ANY KIND, INCLUDING THE
16 * WARRANTY OF DESIGN, MERCHANTABILITY AND FITNESS FOR A PARTICULAR
22 * @brief Dataflow driven Load/Store optimizations, uses some ideas from
24 * @author Michael Beck
39 #include "iroptimize.h"
40 #include "irnodehashmap.h"
41 #include "raw_bitset.h"
46 /* maximum number of output Proj's */
47 #define MAX_PROJ ((long)pn_Load_max > (long)pn_Store_max ? (long)pn_Load_max : (long)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+1]; /**< 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_nodehashmap_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 size_t 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)
153 for (entry = env->forward; entry != NULL; entry = entry->forward_next) {
154 DB((dbg, LEVEL_2, "%+F {", entry->block));
157 for (op = entry->memop_forward; op != NULL; op = op->next) {
159 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"));
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 size_t 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 (block_t*)get_irn_link(block);
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 (memop_t*)get_irn_link(irn);
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)
241 mark_irn_visited(irn);
246 mode = get_irn_mode(irn);
247 if (mode == mode_M) {
248 /* every successor uses memory */
249 for (unsigned i = get_irn_n_outs(irn); i-- > 0; ) {
250 ir_node *succ = get_irn_out(irn, i);
252 if (! irn_visited(succ))
253 walk_memory(succ, pre, post, ctx);
255 } else if (mode == mode_T) {
256 /* only some Proj's uses memory */
257 for (unsigned i = get_irn_n_outs(irn); i-- > 0; ) {
258 ir_node *proj = get_irn_out(irn, i);
260 if (get_irn_mode(proj) == mode_M && ! irn_visited(proj))
261 walk_memory(proj, pre, post, ctx);
269 * Walks over all memory nodes of a graph.
272 * @param pre pre walker function
273 * @param post post walker function
274 * @param ctx context parameter for the walker functions
276 static void walk_memory_irg(ir_graph *irg, irg_walk_func pre, irg_walk_func post, void *ctx)
278 inc_irg_visited(irg);
280 ir_reserve_resources(irg, IR_RESOURCE_IRN_VISITED);
283 * there are two possible sources for memory: initial_mem and nomem
284 * we ignore nomem as this should NOT change the memory
286 walk_memory(get_irg_initial_mem(irg), pre, post, ctx);
288 ir_free_resources(irg, IR_RESOURCE_IRN_VISITED);
292 * Register an address and allocate a (sparse, 0..n) ID for it.
294 * @param adr the IR-node representing the address
296 * @return the allocated id
298 static unsigned register_address(ir_node *adr)
300 address_entry *entry;
304 if (is_Confirm(adr)) {
305 adr = get_Confirm_value(adr);
309 entry = ir_nodehashmap_get(address_entry, &env.adr_map, adr);
313 entry = OALLOC(&env.obst, address_entry);
315 entry->id = env.curr_adr_id++;
316 ir_nodehashmap_insert(&env.adr_map, adr, entry);
318 DB((dbg, LEVEL_3, "ADDRESS %+F has ID %u\n", adr, entry->id));
320 ARR_APP1(ir_node *, env.id_2_address, adr);
328 * translate an address through a Phi node into a given predecessor
331 * @param address the address
332 * @param block the block
333 * @param pos the position of the predecessor in block
335 static ir_node *phi_translate(ir_node *address, const ir_node *block, int pos)
337 if (is_Phi(address) && get_nodes_block(address) == block)
338 address = get_Phi_pred(address, pos);
343 * Walker: allocate an block entry for every block
344 * and register all potential addresses.
346 static void prepare_blocks(ir_node *irn, void *ctx)
351 block_t *entry = OALLOC(&env.obst, block_t);
354 entry->memop_forward = NULL;
355 entry->memop_backward = NULL;
356 entry->avail_out = NULL;
357 entry->id_2_memop_avail = NULL;
358 entry->anticL_in = NULL;
359 entry->id_2_memop_antic = NULL;
361 entry->forward_next = NULL;
362 entry->backward_next = NULL;
364 entry->trans_results = NULL;
365 set_irn_link(irn, entry);
367 set_Block_phis(irn, NULL);
369 /* use block marks to track unreachable blocks */
370 set_Block_mark(irn, 0);
372 n = get_Block_n_cfgpreds(irn);
373 if (n > env.max_cfg_preds)
374 env.max_cfg_preds = n;
376 ir_mode *mode = get_irn_mode(irn);
378 if (mode_is_reference(mode)) {
380 * Register ALL possible addresses: this is overkill yet but
381 * simpler then doing it for all possible translated addresses
382 * (which would be sufficient in the moment.
384 (void)register_address(irn);
390 * Post-Walker, link in all Phi's
392 static void link_phis(ir_node *irn, void *ctx)
397 ir_node *block = get_nodes_block(irn);
398 add_Block_phi(block, irn);
403 * Block walker: creates the inverse post-order list for the CFG.
405 static void inverse_post_order(ir_node *block, void *ctx)
407 block_t *entry = get_block_entry(block);
411 /* mark this block IS reachable from start */
412 set_Block_mark(block, 1);
414 /* create the list in inverse order */
415 entry->forward_next = env.forward;
418 /* remember the first visited (last in list) entry, needed for later */
419 if (env.backward == NULL)
420 env.backward = entry;
424 * Block walker: create backward links for the memops of a block.
426 static void collect_backward(ir_node *block, void *ctx)
428 block_t *entry = get_block_entry(block);
434 * Do NOT link in the end block yet. We want it to be
435 * the first in the list. This is NOT guaranteed by the walker
436 * if we have endless loops.
438 if (block != env.end_bl) {
439 entry->backward_next = env.backward;
441 /* create the list in inverse order */
442 env.backward = entry;
445 /* create backward links for all memory ops */
447 for (op = entry->memop_forward; op != NULL; op = op->next) {
451 entry->memop_backward = last;
457 * @param irn the IR-node representing the memop or NULL
458 * if this is a translated (virtual) memop
460 * @return the allocated memop
462 static memop_t *alloc_memop(ir_node *irn)
464 memop_t *m = OALLOC(&env.obst, memop_t);
466 m->value.address = NULL;
467 m->value.value = NULL;
468 m->value.mode = NULL;
476 memset(m->projs, 0, sizeof(m->projs));
479 set_irn_link(irn, m);
484 * Create a memop for a Phi-replacement.
486 * @param op the memop to clone
487 * @param phi the Phi-node representing the new value
489 static memop_t *clone_memop_phi(memop_t *op, ir_node *phi)
491 memop_t *m = OALLOC(&env.obst, memop_t);
493 m->value = op->value;
494 m->value.value = phi;
501 set_irn_link(phi, m);
506 * Return the memory properties of a call node.
508 * @param call the call node
510 * return a bitset of mtp_property_const and mtp_property_pure
512 static unsigned get_Call_memory_properties(ir_node *call)
514 ir_type *call_tp = get_Call_type(call);
515 unsigned prop = get_method_additional_properties(call_tp);
517 /* check first the call type */
518 if ((prop & (mtp_property_const|mtp_property_pure)) == 0) {
519 /* try the called entity */
520 ir_node *ptr = get_Call_ptr(call);
522 if (is_SymConst_addr_ent(ptr)) {
523 ir_entity *ent = get_SymConst_entity(ptr);
525 prop = get_entity_additional_properties(ent);
528 return prop & (mtp_property_const|mtp_property_pure);
532 * Returns an entity if the address ptr points to a constant one.
534 * @param ptr the address
536 * @return an entity or NULL
538 static ir_entity *find_constant_entity(ir_node *ptr)
541 if (is_SymConst(ptr) && get_SymConst_kind(ptr) == symconst_addr_ent) {
542 return get_SymConst_entity(ptr);
543 } else if (is_Sel(ptr)) {
544 ir_entity *ent = get_Sel_entity(ptr);
545 ir_type *tp = get_entity_owner(ent);
547 /* Do not fiddle with polymorphism. */
548 if (is_Class_type(get_entity_owner(ent)) &&
549 ((get_entity_n_overwrites(ent) != 0) ||
550 (get_entity_n_overwrittenby(ent) != 0) ) )
553 if (is_Array_type(tp)) {
557 for (i = 0, n = get_Sel_n_indexs(ptr); i < n; ++i) {
559 ir_tarval *tlower, *tupper;
560 ir_node *index = get_Sel_index(ptr, i);
561 ir_tarval *tv = computed_value(index);
563 /* check if the index is constant */
564 if (tv == tarval_bad)
567 bound = get_array_lower_bound(tp, i);
568 tlower = computed_value(bound);
569 bound = get_array_upper_bound(tp, i);
570 tupper = computed_value(bound);
572 if (tlower == tarval_bad || tupper == tarval_bad)
575 if (tarval_cmp(tv, tlower) == ir_relation_less)
577 if (tarval_cmp(tupper, tv) == ir_relation_less)
580 /* ok, bounds check finished */
584 if (get_entity_linkage(ent) == IR_LINKAGE_CONSTANT)
588 ptr = get_Sel_ptr(ptr);
589 } else if (is_Add(ptr)) {
590 ir_node *l = get_Add_left(ptr);
591 ir_node *r = get_Add_right(ptr);
593 if (get_irn_mode(l) == get_irn_mode(ptr) && is_Const(r))
595 else if (get_irn_mode(r) == get_irn_mode(ptr) && is_Const(l))
600 /* for now, we support only one addition, reassoc should fold all others */
601 if (! is_SymConst(ptr) && !is_Sel(ptr))
603 } else if (is_Sub(ptr)) {
604 ir_node *l = get_Sub_left(ptr);
605 ir_node *r = get_Sub_right(ptr);
607 if (get_irn_mode(l) == get_irn_mode(ptr) && is_Const(r))
611 /* for now, we support only one subtraction, reassoc should fold all others */
612 if (! is_SymConst(ptr) && !is_Sel(ptr))
620 * Return the Selection index of a Sel node from dimension n
622 static long get_Sel_array_index_long(ir_node *n, int dim)
624 ir_node *index = get_Sel_index(n, dim);
625 return get_tarval_long(get_Const_tarval(index));
628 typedef struct path_entry {
630 struct path_entry *next;
634 static ir_node *rec_find_compound_ent_value(ir_node *ptr, path_entry *next)
636 path_entry entry, *p;
637 ir_entity *ent, *field;
638 ir_initializer_t *initializer;
644 if (is_SymConst(ptr)) {
646 ent = get_SymConst_entity(ptr);
647 initializer = get_entity_initializer(ent);
648 for (p = next; p != NULL;) {
649 if (initializer->kind != IR_INITIALIZER_COMPOUND)
651 n = get_initializer_compound_n_entries(initializer);
652 tp = get_entity_type(ent);
654 if (is_Array_type(tp)) {
655 ent = get_array_element_entity(tp);
660 initializer = get_initializer_compound_value(initializer, 0);
666 initializer = get_initializer_compound_value(initializer, p->index);
671 tp = get_entity_type(ent);
672 while (is_Array_type(tp)) {
673 ent = get_array_element_entity(tp);
674 tp = get_entity_type(ent);
676 n = get_initializer_compound_n_entries(initializer);
679 initializer = get_initializer_compound_value(initializer, 0);
682 switch (initializer->kind) {
683 case IR_INITIALIZER_CONST:
684 return get_initializer_const_value(initializer);
685 case IR_INITIALIZER_TARVAL:
686 case IR_INITIALIZER_NULL:
690 } else if (is_Sel(ptr)) {
691 entry.ent = field = get_Sel_entity(ptr);
692 tp = get_entity_owner(field);
693 if (is_Array_type(tp)) {
694 assert(get_Sel_n_indexs(ptr) == 1 && "multi dim arrays not implemented");
695 entry.index = get_Sel_array_index_long(ptr, 0) - get_array_lower_bound_int(tp, 0);
697 size_t i, n_members = get_compound_n_members(tp);
698 for (i = 0; i < n_members; ++i) {
699 if (get_compound_member(tp, i) == field)
702 if (i >= n_members) {
703 /* not found: should NOT happen */
708 return rec_find_compound_ent_value(get_Sel_ptr(ptr), &entry);
709 } else if (is_Add(ptr)) {
714 ir_node *l = get_Add_left(ptr);
715 ir_node *r = get_Add_right(ptr);
718 tv = get_Const_tarval(r);
721 tv = get_Const_tarval(l);
725 mode = get_tarval_mode(tv);
727 /* ptr must be a Sel or a SymConst, this was checked in find_constant_entity() */
729 field = get_Sel_entity(ptr);
731 field = get_SymConst_entity(ptr);
734 /* count needed entries */
736 for (ent = field;;) {
737 tp = get_entity_type(ent);
738 if (! is_Array_type(tp))
740 ent = get_array_element_entity(tp);
743 /* should be at least ONE entry */
747 /* allocate the right number of entries */
748 NEW_ARR_A(path_entry, p, pos);
752 for (ent = field;;) {
754 ir_tarval *sz, *tv_index, *tlower, *tupper;
758 tp = get_entity_type(ent);
759 if (! is_Array_type(tp))
761 ent = get_array_element_entity(tp);
763 p[pos].next = &p[pos + 1];
765 size = get_type_size_bytes(get_entity_type(ent));
766 sz = new_tarval_from_long(size, mode);
768 tv_index = tarval_div(tv, sz);
769 tv = tarval_mod(tv, sz);
771 if (tv_index == tarval_bad || tv == tarval_bad)
774 assert(get_array_n_dimensions(tp) == 1 && "multiarrays not implemented");
775 bound = get_array_lower_bound(tp, 0);
776 tlower = computed_value(bound);
777 bound = get_array_upper_bound(tp, 0);
778 tupper = computed_value(bound);
780 if (tlower == tarval_bad || tupper == tarval_bad)
783 if (tarval_cmp(tv_index, tlower) == ir_relation_less)
785 if (tarval_cmp(tupper, tv_index) == ir_relation_less)
788 /* ok, bounds check finished */
789 index = get_tarval_long(tv_index);
790 p[pos].index = index;
793 if (! tarval_is_null(tv)) {
794 /* hmm, wrong access */
797 p[pos - 1].next = next;
798 return rec_find_compound_ent_value(ptr, p);
799 } else if (is_Sub(ptr)) {
800 ir_node *l = get_Sub_left(ptr);
801 ir_node *r = get_Sub_right(ptr);
804 tv = get_Const_tarval(r);
811 static ir_node *find_compound_ent_value(ir_node *ptr)
813 return rec_find_compound_ent_value(ptr, NULL);
817 * Mark a Load memop to be replace by a definition
819 * @param op the Load memop
821 static void mark_replace_load(memop_t *op, ir_node *def)
824 op->flags |= FLAG_KILLED_NODE;
829 * Mark a Store memop to be removed.
831 * @param op the Store memop
833 static void mark_remove_store(memop_t *op)
835 op->flags |= FLAG_KILLED_NODE;
840 * Update a memop for a Load.
844 static void update_Load_memop(memop_t *m)
846 ir_node *load = m->node;
850 if (get_Load_volatility(load) == volatility_is_volatile)
851 m->flags |= FLAG_IGNORE;
853 ptr = get_Load_ptr(load);
855 m->value.address = ptr;
857 for (unsigned i = get_irn_n_outs(load); i-- > 0; ) {
858 ir_node *proj = get_irn_out(load, i);
861 /* beware of keep edges */
865 pn = get_Proj_proj(proj);
869 m->value.value = proj;
870 m->value.mode = get_irn_mode(proj);
872 case pn_Load_X_except:
873 m->flags |= FLAG_EXCEPTION;
878 case pn_Load_X_regular:
881 panic("Unsupported Proj from Load %+F", proj);
885 /* check if we can determine the entity that will be loaded */
886 ent = find_constant_entity(ptr);
888 if (ent != NULL && get_entity_visibility(ent) != ir_visibility_external) {
889 /* a static allocation that is not external: there should be NO exception
890 * when loading even if we cannot replace the load itself. */
891 ir_node *value = NULL;
893 /* no exception, clear the m fields as it might be checked later again */
894 if (m->projs[pn_Load_X_except]) {
895 ir_graph *irg = get_irn_irg(ptr);
896 exchange(m->projs[pn_Load_X_except], new_r_Bad(irg, mode_X));
897 m->projs[pn_Load_X_except] = NULL;
898 m->flags &= ~FLAG_EXCEPTION;
901 if (m->projs[pn_Load_X_regular]) {
902 exchange(m->projs[pn_Load_X_regular], new_r_Jmp(get_nodes_block(load)));
903 m->projs[pn_Load_X_regular] = NULL;
907 if (get_entity_linkage(ent) & IR_LINKAGE_CONSTANT) {
908 if (ent->initializer) {
909 /* new style initializer */
910 value = find_compound_ent_value(ptr);
913 value = can_replace_load_by_const(load, value);
917 /* we completely replace the load by this value */
918 DB((dbg, LEVEL_1, "Replacing Load %+F by constant %+F\n", m->node, value));
919 mark_replace_load(m, value);
924 if (m->value.value != NULL && !(m->flags & FLAG_IGNORE)) {
925 /* only create an address if this node is NOT killed immediately or ignored */
926 m->value.id = register_address(ptr);
929 /* no user, KILL it */
930 mark_replace_load(m, NULL);
935 * Update a memop for a Store.
939 static void update_Store_memop(memop_t *m)
941 ir_node *store = m->node;
942 ir_node *adr = get_Store_ptr(store);
944 if (get_Store_volatility(store) == volatility_is_volatile) {
945 m->flags |= FLAG_IGNORE;
947 /* only create an address if this node is NOT ignored */
948 m->value.id = register_address(adr);
952 m->value.address = adr;
954 for (unsigned i = get_irn_n_outs(store); i-- > 0; ) {
955 ir_node *proj = get_irn_out(store, i);
958 /* beware of keep edges */
962 pn = get_Proj_proj(proj);
965 case pn_Store_X_except:
966 m->flags |= FLAG_EXCEPTION;
971 case pn_Store_X_regular:
974 panic("Unsupported Proj from Store %+F", proj);
977 m->value.value = get_Store_value(store);
978 m->value.mode = get_irn_mode(m->value.value);
982 * Update a memop for a Call.
986 static void update_Call_memop(memop_t *m)
988 ir_node *call = m->node;
989 unsigned prop = get_Call_memory_properties(call);
991 if (prop & mtp_property_const) {
992 /* A constant call did NOT use memory at all, we
993 can kick it from the list. */
994 } else if (prop & mtp_property_pure) {
995 /* pure calls READ memory */
998 m->flags = FLAG_KILL_ALL;
1000 for (unsigned i = get_irn_n_outs(call); i-- > 0; ) {
1001 ir_node *proj = get_irn_out(call, i);
1003 /* beware of keep edges */
1007 switch (get_Proj_proj(proj)) {
1008 case pn_Call_X_except:
1009 m->flags |= FLAG_EXCEPTION;
1019 * Update a memop for a Div/Mod.
1021 * @param m the memop
1023 static void update_Div_memop(memop_t *m)
1025 ir_node *div = m->node;
1027 for (unsigned i = get_irn_n_outs(div); i-- > 0; ) {
1028 ir_node *proj = get_irn_out(div, i);
1030 /* beware of keep edges */
1034 switch (get_Proj_proj(proj)) {
1035 case pn_Div_X_except:
1036 m->flags |= FLAG_EXCEPTION;
1045 static void update_Mod_memop(memop_t *m)
1047 ir_node *div = m->node;
1049 for (unsigned i = get_irn_n_outs(div); i-- > 0; ) {
1050 ir_node *proj = get_irn_out(div, i);
1052 /* beware of keep edges */
1056 switch (get_Proj_proj(proj)) {
1057 case pn_Mod_X_except:
1058 m->flags |= FLAG_EXCEPTION;
1068 * Update a memop for a Phi.
1070 * @param m the memop
1072 static void update_Phi_memop(memop_t *m)
1074 /* the Phi is its own mem */
1079 * Memory walker: collect all memory ops and build topological lists.
1081 static void collect_memops(ir_node *irn, void *ctx)
1089 /* we can safely ignore ProjM's except the initial memory */
1090 ir_graph *irg = get_irn_irg(irn);
1091 if (irn != get_irg_initial_mem(irg))
1095 op = alloc_memop(irn);
1096 block = get_nodes_block(irn);
1097 entry = get_block_entry(block);
1100 update_Phi_memop(op);
1101 /* Phis must be always placed first */
1102 op->next = entry->memop_forward;
1103 entry->memop_forward = op;
1104 if (entry->memop_backward == NULL)
1105 entry->memop_backward = op;
1107 switch (get_irn_opcode(irn)) {
1109 update_Load_memop(op);
1112 update_Store_memop(op);
1115 update_Call_memop(op);
1122 /* initial memory */
1127 /* we can those to find the memory edge */
1130 update_Div_memop(op);
1133 update_Mod_memop(op);
1137 /* TODO: handle some builtins */
1139 /* unsupported operation */
1140 op->flags = FLAG_KILL_ALL;
1144 /* all other should be placed last */
1145 if (entry->memop_backward == NULL) {
1146 entry->memop_forward = entry->memop_backward = op;
1148 entry->memop_backward->next = op;
1149 entry->memop_backward = op;
1155 * Find an address in the current set.
1157 * @param value the value to be searched for
1159 * @return a memop for the value or NULL if the value does
1160 * not exists in the set or cannot be converted into
1161 * the requested mode
1163 static memop_t *find_address(const value_t *value)
1165 if (rbitset_is_set(env.curr_set, value->id)) {
1166 memop_t *res = env.curr_id_2_memop[value->id];
1168 if (res->value.mode == value->mode)
1170 /* allow hidden casts */
1171 if (get_mode_arithmetic(res->value.mode) == irma_twos_complement &&
1172 get_mode_arithmetic(value->mode) == irma_twos_complement &&
1173 get_mode_size_bits(res->value.mode) == get_mode_size_bits(value->mode))
1180 * Find an address in the avail_out set.
1182 * @param bl the block
1184 static memop_t *find_address_avail(const block_t *bl, unsigned id, const ir_mode *mode)
1186 if (rbitset_is_set(bl->avail_out, id)) {
1187 memop_t *res = bl->id_2_memop_avail[id];
1189 if (res->value.mode == mode)
1191 /* allow hidden casts */
1192 if (get_mode_arithmetic(res->value.mode) == irma_twos_complement &&
1193 get_mode_arithmetic(mode) == irma_twos_complement &&
1194 get_mode_size_bits(res->value.mode) == get_mode_size_bits(mode))
1201 * Kill all addresses from the current set.
1203 static void kill_all(void)
1205 rbitset_clear_all(env.curr_set, env.rbs_size);
1208 rbitset_set(env.curr_set, env.rbs_size - 1);
1212 * Kill memops that are not alias free due to a Store value from the current set.
1214 * @param value the Store value
1216 static void kill_memops(const value_t *value)
1218 size_t end = env.rbs_size - 1;
1221 for (pos = rbitset_next(env.curr_set, 0, 1); pos < end; pos = rbitset_next(env.curr_set, pos + 1, 1)) {
1222 memop_t *op = env.curr_id_2_memop[pos];
1224 if (ir_no_alias != get_alias_relation(value->address, value->mode,
1225 op->value.address, op->value.mode)) {
1226 rbitset_clear(env.curr_set, pos);
1227 env.curr_id_2_memop[pos] = NULL;
1228 DB((dbg, LEVEL_2, "KILLING %+F because of possible alias address %+F\n", op->node, value->address));
1234 * Add the value of a memop to the current set.
1236 * @param op the memory op
1238 static void add_memop(memop_t *op)
1240 rbitset_set(env.curr_set, op->value.id);
1241 env.curr_id_2_memop[op->value.id] = op;
1245 * Add the value of a memop to the avail_out set.
1247 * @param bl the block
1248 * @param op the memory op
1250 static void add_memop_avail(block_t *bl, memop_t *op)
1252 rbitset_set(bl->avail_out, op->value.id);
1253 bl->id_2_memop_avail[op->value.id] = op;
1257 * Check, if we can convert a value of one mode to another mode
1258 * without changing the representation of bits.
1260 * @param from the original mode
1261 * @param to the destination mode
1263 static int can_convert_to(const ir_mode *from, const ir_mode *to)
1265 if (get_mode_arithmetic(from) == irma_twos_complement &&
1266 get_mode_arithmetic(to) == irma_twos_complement &&
1267 get_mode_size_bits(from) == get_mode_size_bits(to))
1273 * Add a Conv to the requested mode if needed.
1275 * @param irn the IR-node to convert
1276 * @param mode the destination mode
1278 * @return the possible converted node or NULL
1279 * if the conversion is not possible
1281 static ir_node *conv_to(ir_node *irn, ir_mode *mode)
1283 ir_mode *other = get_irn_mode(irn);
1284 if (other != mode) {
1285 /* different modes: check if conversion is possible without changing the bits */
1286 if (can_convert_to(other, mode)) {
1287 ir_node *block = get_nodes_block(irn);
1288 return new_r_Conv(block, irn, mode);
1290 /* otherwise not possible ... yet */
1297 * Update the address of an value if this address was a load result
1298 * and the load is killed now.
1300 * @param value the value whose address is updated
1302 static void update_address(value_t *value)
1304 if (is_Proj(value->address)) {
1305 ir_node *load = get_Proj_pred(value->address);
1307 if (is_Load(load)) {
1308 const memop_t *op = get_irn_memop(load);
1310 if (op->flags & FLAG_KILLED_NODE)
1311 value->address = op->replace;
1317 * Do forward dataflow analysis on the given block and calculate the
1318 * GEN and KILL in the current (avail) set.
1320 * @param bl the block
1322 static void calc_gen_kill_avail(block_t *bl)
1327 for (op = bl->memop_forward; op != NULL; op = op->next) {
1328 switch (get_irn_opcode(op->node)) {
1336 if (! (op->flags & (FLAG_KILLED_NODE|FLAG_IGNORE))) {
1337 /* do we have this already? */
1340 update_address(&op->value);
1341 other = find_address(&op->value);
1342 if (other != NULL && other != op) {
1343 def = conv_to(other->value.value, op->value.mode);
1345 #ifdef DEBUG_libfirm
1346 if (is_Store(other->node)) {
1348 DB((dbg, LEVEL_1, "RAW %+F <- %+F(%+F)\n", op->node, def, other->node));
1351 DB((dbg, LEVEL_1, "RAR %+F <- %+F(%+F)\n", op->node, def, other->node));
1354 mark_replace_load(op, def);
1355 /* do NOT change the memop table */
1359 /* add this value */
1364 if (! (op->flags & FLAG_KILLED_NODE)) {
1365 /* do we have this store already */
1368 update_address(&op->value);
1369 other = find_address(&op->value);
1370 if (other != NULL) {
1371 if (is_Store(other->node)) {
1372 if (op != other && !(other->flags & FLAG_IGNORE) &&
1373 get_nodes_block(other->node) == get_nodes_block(op->node)) {
1375 * A WAW in the same block we can kick the first store.
1376 * This is a shortcut: we know that the second Store will be anticipated
1379 DB((dbg, LEVEL_1, "WAW %+F <- %+F\n", other->node, op->node));
1380 mark_remove_store(other);
1381 /* FIXME: a Load might be get freed due to this killed store */
1383 } else if (other->value.value == op->value.value && !(op->flags & FLAG_IGNORE)) {
1385 DB((dbg, LEVEL_1, "WAR %+F <- %+F\n", op->node, other->node));
1386 mark_remove_store(op);
1387 /* do NOT change the memop table */
1391 /* KILL all possible aliases */
1392 kill_memops(&op->value);
1393 /* add this value */
1398 if (op->flags & FLAG_KILL_ALL)
1405 * Do forward dataflow analysis on a given block to calculate the avail_out set
1406 * for this block only.
1408 * @param block the block
1410 static void forward_avail(block_t *bl)
1412 /* fill the data from the current block */
1413 env.curr_id_2_memop = bl->id_2_memop_avail;
1414 env.curr_set = bl->avail_out;
1416 calc_gen_kill_avail(bl);
1417 dump_curr(bl, "Avail_out");
1421 * Do backward dataflow analysis on a given block to calculate the antic set
1422 * of Loaded addresses.
1424 * @param bl the block
1426 * @return non-zero if the set has changed since last iteration
1428 static int backward_antic(block_t *bl)
1431 ir_node *block = bl->block;
1432 int n = get_Block_n_cfg_outs(block);
1435 ir_node *succ = get_Block_cfg_out(block, 0);
1436 block_t *succ_bl = get_block_entry(succ);
1437 int pred_pos = get_Block_cfgpred_pos(succ, block);
1438 size_t end = env.rbs_size - 1;
1443 if (bl->trans_results == NULL) {
1444 /* allocate the translate cache */
1445 bl->trans_results = OALLOCNZ(&env.obst, memop_t*, env.curr_adr_id);
1448 /* check for partly redundant values */
1449 for (pos = rbitset_next(succ_bl->anticL_in, 0, 1);
1451 pos = rbitset_next(succ_bl->anticL_in, pos + 1, 1)) {
1453 * do Phi-translation here: Note that at this point the nodes are
1454 * not changed, so we can safely cache the results.
1455 * However: Loads of Load results ARE bad, because we have no way
1456 to translate them yet ...
1458 memop_t *op = bl->trans_results[pos];
1460 /* not yet translated */
1461 ir_node *adr, *trans_adr;
1463 op = succ_bl->id_2_memop_antic[pos];
1464 adr = op->value.address;
1466 trans_adr = phi_translate(adr, succ, pred_pos);
1467 if (trans_adr != adr) {
1468 /* create a new entry for the translated one */
1471 new_op = alloc_memop(NULL);
1472 new_op->value.address = trans_adr;
1473 new_op->value.id = register_address(trans_adr);
1474 new_op->value.mode = op->value.mode;
1475 new_op->node = op->node; /* we need the node to decide if Load/Store */
1476 new_op->flags = op->flags;
1478 bl->trans_results[pos] = new_op;
1482 env.curr_id_2_memop[op->value.id] = op;
1483 rbitset_set(env.curr_set, op->value.id);
1486 ir_node *succ = get_Block_cfg_out(block, 0);
1487 block_t *succ_bl = get_block_entry(succ);
1490 rbitset_copy(env.curr_set, succ_bl->anticL_in, env.rbs_size);
1491 memcpy(env.curr_id_2_memop, succ_bl->id_2_memop_antic, env.rbs_size * sizeof(env.curr_id_2_memop[0]));
1493 /* Hmm: probably we want kill merges of Loads ans Stores here */
1494 for (i = n - 1; i > 0; --i) {
1495 ir_node *succ = get_Block_cfg_out(bl->block, i);
1496 block_t *succ_bl = get_block_entry(succ);
1498 rbitset_and(env.curr_set, succ_bl->anticL_in, env.rbs_size);
1501 /* block ends with a noreturn call */
1505 dump_curr(bl, "AnticL_out");
1507 for (op = bl->memop_backward; op != NULL; op = op->prev) {
1508 switch (get_irn_opcode(op->node)) {
1516 if (! (op->flags & (FLAG_KILLED_NODE|FLAG_IGNORE))) {
1522 if (! (op->flags & FLAG_KILLED_NODE)) {
1523 /* a Store: check which memops must be killed */
1524 kill_memops(&op->value);
1528 if (op->flags & FLAG_KILL_ALL)
1533 memcpy(bl->id_2_memop_antic, env.curr_id_2_memop, env.rbs_size * sizeof(env.curr_id_2_memop[0]));
1534 if (! rbitsets_equal(bl->anticL_in, env.curr_set, env.rbs_size)) {
1536 rbitset_copy(bl->anticL_in, env.curr_set, env.rbs_size);
1537 dump_curr(bl, "AnticL_in*");
1540 dump_curr(bl, "AnticL_in");
1545 * Replace a Load memop by a already known value.
1547 * @param op the Load memop
1549 static void replace_load(memop_t *op)
1551 ir_node *load = op->node;
1552 ir_node *def = skip_Id(op->replace);
1557 DB((dbg, LEVEL_1, "Replacing %+F by definition %+F\n", load, is_Proj(def) ? get_Proj_pred(def) : def));
1559 if (op->flags & FLAG_EXCEPTION) {
1560 /* bad: this node is unused and executed for exception only */
1561 DB((dbg, LEVEL_1, "Unused %+F executed for exception only ...\n", load));
1564 DB((dbg, LEVEL_1, "Killing unused %+F\n", load));
1567 if (op->mem != NULL) {
1568 /* in rare cases a Load might have NO memory */
1569 exchange(op->mem, get_Load_mem(load));
1571 proj = op->projs[pn_Load_res];
1573 mode = get_irn_mode(proj);
1574 if (get_irn_mode(def) != mode) {
1576 dbg_info *db = get_irn_dbg_info(load);
1577 ir_node *block = get_nodes_block(proj);
1578 def = new_rd_Conv(db, block, def, mode);
1580 exchange(proj, def);
1582 proj = op->projs[pn_Load_X_except];
1584 ir_graph *irg = get_irn_irg(load);
1585 exchange(proj, new_r_Bad(irg, mode_X));
1587 proj = op->projs[pn_Load_X_regular];
1589 exchange(proj, new_r_Jmp(get_nodes_block(load)));
1594 * Remove a Store memop.
1596 * @param op the Store memop
1598 static void remove_store(memop_t *op)
1600 ir_node *store = op->node;
1603 DB((dbg, LEVEL_1, "Removing %+F\n", store));
1605 if (op->mem != NULL) {
1606 /* in rare cases a Store might have no memory */
1607 exchange(op->mem, get_Store_mem(store));
1609 proj = op->projs[pn_Store_X_except];
1611 ir_graph *irg = get_irn_irg(store);
1612 exchange(proj, new_r_Bad(irg, mode_X));
1614 proj = op->projs[pn_Store_X_regular];
1616 exchange(proj, new_r_Jmp(get_nodes_block(store)));
1622 * Do all necessary replacements for a given block.
1624 * @param bl the block
1626 static void do_replacements(block_t *bl)
1630 for (op = bl->memop_forward; op != NULL; op = op->next) {
1631 if (op->flags & FLAG_KILLED_NODE) {
1632 switch (get_irn_opcode(op->node)) {
1645 * Calculate the Avail_out sets for all basic blocks.
1647 static void calcAvail(void)
1649 memop_t **tmp_memop = env.curr_id_2_memop;
1650 unsigned *tmp_set = env.curr_set;
1653 /* calculate avail_out */
1654 DB((dbg, LEVEL_2, "Calculate Avail_out\n"));
1656 /* iterate over all blocks in in any order, skip the start block */
1657 for (bl = env.forward->forward_next; bl != NULL; bl = bl->forward_next) {
1661 /* restore the current sets */
1662 env.curr_id_2_memop = tmp_memop;
1663 env.curr_set = tmp_set;
1667 * Calculate the Antic_in sets for all basic blocks.
1669 static void calcAntic(void)
1673 /* calculate antic_out */
1674 DB((dbg, LEVEL_2, "Calculate Antic_in\n"));
1679 DB((dbg, LEVEL_2, "Iteration %d:\n=========\n", i));
1683 /* over all blocks in reverse post order */
1684 for (bl = env.backward->backward_next; bl != NULL; bl = bl->backward_next) {
1685 need_iter |= backward_antic(bl);
1688 } while (need_iter);
1689 DB((dbg, LEVEL_2, "Get anticipated Load set after %d iterations\n", i));
1693 * Return the node representing the last memory in a block.
1695 * @param bl the block
1697 static ir_node *find_last_memory(block_t *bl)
1700 if (bl->memop_backward != NULL) {
1701 return bl->memop_backward->mem;
1703 /* if there is NO memory in this block, go to the dominator */
1704 bl = get_block_entry(get_Block_idom(bl->block));
1709 * Reroute all memory users of old memory
1710 * to a new memory IR-node.
1712 * @param omem the old memory IR-node
1713 * @param nmem the new memory IR-node
1715 static void reroute_all_mem_users(ir_node *omem, ir_node *nmem)
1717 for (unsigned i = get_irn_n_outs(omem); i-- > 0; ) {
1719 ir_node *user = get_irn_out_ex(omem, i, &n_pos);
1721 set_irn_n(user, n_pos, nmem);
1724 /* all edges previously point to omem now point to nmem */
1725 nmem->o.out = omem->o.out;
1729 * Reroute memory users of old memory that are dominated by a given block
1730 * to a new memory IR-node.
1732 * @param omem the old memory IR-node
1733 * @param nmem the new memory IR-node
1734 * @param pass_bl the block the memory must pass
1736 static void reroute_mem_through(ir_node *omem, ir_node *nmem, ir_node *pass_bl)
1738 unsigned n = get_irn_n_outs(omem);
1739 ir_def_use_edges *new_out = OALLOCF(&env.obst, ir_def_use_edges, edges, n);
1742 for (unsigned i = 0; i < n; ++i) {
1744 ir_node *user = get_irn_out_ex(omem, i, &n_pos);
1745 ir_node *use_bl = get_nodes_block(user);
1749 use_bl = get_Block_cfgpred_block(use_bl, n_pos);
1751 if (block_dominates(pass_bl, use_bl)) {
1752 /* found an user that is dominated */
1753 new_out->edges[j].pos = n_pos;
1754 new_out->edges[j].use = user;
1757 set_irn_n(user, n_pos, nmem);
1760 new_out->n_edges = j;
1762 /* Modify the out structure: we create a new out edge array on our
1763 temporary obstack here. This should be no problem, as we invalidate the
1764 edges at the end either. */
1765 /* first entry is used for the length */
1766 nmem->o.out = new_out;
1770 * insert Loads, making partly redundant Loads fully redundant
1772 static int insert_Load(block_t *bl)
1774 ir_node *block = bl->block;
1775 int i, n = get_Block_n_cfgpreds(block);
1776 size_t end = env.rbs_size - 1;
1778 DB((dbg, LEVEL_3, "processing %+F\n", block));
1781 /* might still happen for an unreachable block (end for instance) */
1789 NEW_ARR_A(ir_node *, ins, n);
1791 rbitset_set_all(env.curr_set, env.rbs_size);
1793 /* More than one predecessors, calculate the join for all avail_outs ignoring unevaluated
1794 Blocks. These put in Top anyway. */
1795 for (i = n - 1; i >= 0; --i) {
1796 ir_node *pred = skip_Proj(get_Block_cfgpred(block, i));
1797 ir_node *blk = get_nodes_block(pred);
1800 pred_bl = get_block_entry(blk);
1801 rbitset_and(env.curr_set, pred_bl->avail_out, env.rbs_size);
1803 if (is_Load(pred) || is_Store(pred)) {
1804 /* We reached this block by an exception from a Load or Store:
1805 * the memop creating the exception was NOT completed than, kill it
1807 memop_t *exc_op = get_irn_memop(pred);
1808 rbitset_clear(env.curr_set, exc_op->value.id);
1813 * Ensure that all values are in the map: build Phi's if necessary:
1814 * Note: the last bit is the sentinel and ALWAYS set, so end with -2.
1816 for (pos = 0; pos < env.rbs_size - 1; ++pos) {
1817 if (! rbitset_is_set(env.curr_set, pos))
1818 env.curr_id_2_memop[pos] = NULL;
1821 memop_t *first = NULL;
1822 ir_mode *mode = NULL;
1824 for (i = 0; i < n; ++i) {
1825 ir_node *pred = get_Block_cfgpred_block(bl->block, i);
1826 block_t *pred_bl = get_block_entry(pred);
1828 memop_t *mop = pred_bl->id_2_memop_avail[pos];
1829 if (first == NULL) {
1831 ins[0] = first->value.value;
1832 mode = get_irn_mode(ins[0]);
1834 /* no Phi needed so far */
1835 env.curr_id_2_memop[pos] = first;
1837 ins[i] = conv_to(mop->value.value, mode);
1838 if (ins[i] != ins[0]) {
1839 if (ins[i] == NULL) {
1840 /* conversion failed */
1841 env.curr_id_2_memop[pos] = NULL;
1842 rbitset_clear(env.curr_set, pos);
1851 ir_node *phi = new_r_Phi(bl->block, n, ins, mode);
1852 memop_t *phiop = alloc_memop(phi);
1854 phiop->value = first->value;
1855 phiop->value.value = phi;
1857 /* no need to link it in, as it is a DATA phi */
1859 env.curr_id_2_memop[pos] = phiop;
1861 DB((dbg, LEVEL_3, "Created new %+F on merging value for address %+F\n", phi, first->value.address));
1866 /* only one predecessor, simply copy the map */
1867 ir_node *pred = get_Block_cfgpred_block(bl->block, 0);
1868 block_t *pred_bl = get_block_entry(pred);
1870 rbitset_copy(env.curr_set, pred_bl->avail_out, env.rbs_size);
1872 memcpy(env.curr_id_2_memop, pred_bl->id_2_memop_avail, env.rbs_size * sizeof(bl->id_2_memop_avail[0]));
1878 /* check for partly redundant values */
1879 for (pos = rbitset_next(bl->anticL_in, 0, 1);
1881 pos = rbitset_next(bl->anticL_in, pos + 1, 1)) {
1882 memop_t *op = bl->id_2_memop_antic[pos];
1883 int have_some, all_same;
1886 if (rbitset_is_set(env.curr_set, pos)) {
1891 assert(is_Load(op->node));
1893 DB((dbg, LEVEL_3, "anticipated %+F\n", op->node));
1898 for (i = n - 1; i >= 0; --i) {
1899 ir_node *pred = get_Block_cfgpred_block(block, i);
1900 block_t *pred_bl = get_block_entry(pred);
1901 ir_mode *mode = op->value.mode;
1905 adr = phi_translate(op->value.address, block, i);
1906 DB((dbg, LEVEL_3, ".. using address %+F in pred %d\n", adr, i));
1907 e = find_address_avail(pred_bl, register_address(adr), mode);
1909 ir_node *ef_block = get_nodes_block(adr);
1910 if (! block_dominates(ef_block, pred)) {
1911 /* cannot place a copy here */
1913 DB((dbg, LEVEL_3, "%+F cannot be moved into predecessor %+F\n", op->node, pred));
1916 DB((dbg, LEVEL_3, "%+F is not available in predecessor %+F\n", op->node, pred));
1917 pred_bl->avail = NULL;
1920 if (e->value.mode != mode && !can_convert_to(e->value.mode, mode)) {
1921 /* cannot create a Phi due to different modes */
1927 DB((dbg, LEVEL_3, "%+F is available for %+F in predecessor %+F\n", e->node, op->node, pred));
1930 else if (first != e->node)
1934 if (have_some && !all_same) {
1935 ir_mode *mode = op->value.mode;
1939 NEW_ARR_A(ir_node *, in, n);
1941 for (i = n - 1; i >= 0; --i) {
1942 ir_node *pred = get_Block_cfgpred_block(block, i);
1943 block_t *pred_bl = get_block_entry(pred);
1945 if (pred_bl->avail == NULL) {
1946 /* create a new Load here and make to make it fully redundant */
1947 dbg_info *db = get_irn_dbg_info(op->node);
1948 ir_node *last_mem = find_last_memory(pred_bl);
1949 ir_node *load, *def, *adr;
1952 assert(last_mem != NULL);
1953 adr = phi_translate(op->value.address, block, i);
1954 load = new_rd_Load(db, pred, last_mem, adr, mode, cons_none);
1955 def = new_r_Proj(load, mode, pn_Load_res);
1956 DB((dbg, LEVEL_1, "Created new %+F in %+F for party redundant %+F\n", load, pred, op->node));
1958 new_op = alloc_memop(load);
1959 new_op->mem = new_r_Proj(load, mode_M, pn_Load_M);
1960 new_op->value.address = adr;
1961 new_op->value.id = op->value.id;
1962 new_op->value.mode = mode;
1963 new_op->value.value = def;
1965 new_op->projs[pn_Load_M] = new_op->mem;
1966 new_op->projs[pn_Load_res] = def;
1968 new_op->prev = pred_bl->memop_backward;
1969 if (pred_bl->memop_backward != NULL)
1970 pred_bl->memop_backward->next = new_op;
1972 pred_bl->memop_backward = new_op;
1974 if (pred_bl->memop_forward == NULL)
1975 pred_bl->memop_forward = new_op;
1977 if (get_nodes_block(last_mem) == pred) {
1978 /* We have add a new last memory op in pred block.
1979 If pred had already a last mem, reroute all memory
1981 reroute_all_mem_users(last_mem, new_op->mem);
1983 /* reroute only those memory going through the pre block */
1984 reroute_mem_through(last_mem, new_op->mem, pred);
1987 /* we added this load at the end, so it will be avail anyway */
1988 add_memop_avail(pred_bl, new_op);
1989 pred_bl->avail = new_op;
1991 in[i] = conv_to(pred_bl->avail->value.value, mode);
1993 phi = new_r_Phi(block, n, in, mode);
1994 DB((dbg, LEVEL_1, "Created new %+F in %+F for now redundant %+F\n", phi, block, op->node));
1996 phi_op = clone_memop_phi(op, phi);
2002 /* recalculate avail by gen and kill */
2003 calc_gen_kill_avail(bl);
2005 /* always update the map after gen/kill, as values might have been changed due to RAR/WAR/WAW */
2006 memcpy(bl->id_2_memop_avail, env.curr_id_2_memop, env.rbs_size * sizeof(env.curr_id_2_memop[0]));
2008 if (!rbitsets_equal(bl->avail_out, env.curr_set, env.rbs_size)) {
2009 /* the avail set has changed */
2010 rbitset_copy(bl->avail_out, env.curr_set, env.rbs_size);
2011 dump_curr(bl, "Avail_out*");
2014 dump_curr(bl, "Avail_out");
2019 * Insert Loads upwards.
2021 static void insert_Loads_upwards(void)
2026 /* recalculate antic_out and insert Loads */
2027 DB((dbg, LEVEL_2, "Inserting Loads\n"));
2031 DB((dbg, LEVEL_2, "Iteration %d:\n=========\n", i));
2035 /* over all blocks in reverse post order, skip the start block */
2036 for (bl = env.forward->forward_next; bl != NULL; bl = bl->forward_next) {
2037 need_iter |= insert_Load(bl);
2040 } while (need_iter);
2042 DB((dbg, LEVEL_2, "Finished Load inserting after %d iterations\n", i));
2045 void opt_ldst(ir_graph *irg)
2049 FIRM_DBG_REGISTER(dbg, "firm.opt.ldst");
2051 DB((dbg, LEVEL_1, "\nDoing Load/Store optimization on %+F\n", irg));
2053 assure_irg_properties(irg,
2054 IR_GRAPH_PROPERTY_NO_CRITICAL_EDGES /* we need landing pads */
2055 | IR_GRAPH_PROPERTY_CONSISTENT_ENTITY_USAGE
2056 | IR_GRAPH_PROPERTY_CONSISTENT_OUTS
2057 | IR_GRAPH_PROPERTY_NO_UNREACHABLE_CODE
2058 | IR_GRAPH_PROPERTY_CONSISTENT_DOMINANCE);
2060 if (get_opt_alias_analysis()) {
2061 assure_irp_globals_entity_usage_computed();
2064 obstack_init(&env.obst);
2065 ir_nodehashmap_init(&env.adr_map);
2068 env.backward = NULL;
2069 env.curr_adr_id = 0;
2071 env.max_cfg_preds = 0;
2073 env.start_bl = get_irg_start_block(irg);
2074 env.end_bl = get_irg_end_block(irg);
2075 #ifdef DEBUG_libfirm
2076 env.id_2_address = NEW_ARR_F(ir_node *, 0);
2079 ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK | IR_RESOURCE_BLOCK_MARK);
2081 /* first step: allocate block entries. Note that some blocks might be
2082 unreachable here. Using the normal walk ensures that ALL blocks are initialized. */
2083 irg_walk_graph(irg, prepare_blocks, link_phis, NULL);
2085 /* produce an inverse post-order list for the CFG: this links only reachable
2087 irg_out_block_walk(get_irg_start_block(irg), NULL, inverse_post_order, NULL);
2089 if (! get_Block_mark(env.end_bl)) {
2091 * The end block is NOT reachable due to endless loops
2092 * or no_return calls.
2093 * Place the end block last.
2094 * env.backward points to the last block in the list for this purpose.
2096 env.backward->forward_next = get_block_entry(env.end_bl);
2098 set_Block_mark(env.end_bl, 1);
2101 /* second step: find and sort all memory ops */
2102 walk_memory_irg(irg, collect_memops, NULL, NULL);
2104 #ifdef DEBUG_libfirm
2105 /* check that the backward map is correct */
2106 assert((unsigned)ARR_LEN(env.id_2_address) == env.curr_adr_id);
2109 if (env.n_mem_ops == 0) {
2114 /* create the backward links. */
2115 env.backward = NULL;
2116 irg_block_walk_graph(irg, NULL, collect_backward, NULL);
2118 /* link the end block in */
2119 bl = get_block_entry(env.end_bl);
2120 bl->backward_next = env.backward;
2123 /* check that we really start with the start / end block */
2124 assert(env.forward->block == env.start_bl);
2125 assert(env.backward->block == env.end_bl);
2127 /* create address sets: for now, only the existing addresses are allowed plus one
2128 needed for the sentinel */
2129 env.rbs_size = env.curr_adr_id + 1;
2131 /* create the current set */
2132 env.curr_set = rbitset_obstack_alloc(&env.obst, env.rbs_size);
2133 rbitset_set(env.curr_set, env.rbs_size - 1);
2134 env.curr_id_2_memop = NEW_ARR_DZ(memop_t*, &env.obst, env.rbs_size);
2136 for (bl = env.forward; bl != NULL; bl = bl->forward_next) {
2137 /* set sentinel bits */
2138 bl->avail_out = rbitset_obstack_alloc(&env.obst, env.rbs_size);
2139 rbitset_set(bl->avail_out, env.rbs_size - 1);
2141 bl->id_2_memop_avail = NEW_ARR_DZ(memop_t*, &env.obst, env.rbs_size);
2143 bl->anticL_in = rbitset_obstack_alloc(&env.obst, env.rbs_size);
2144 rbitset_set(bl->anticL_in, env.rbs_size - 1);
2146 bl->id_2_memop_antic = NEW_ARR_DZ(memop_t*, &env.obst, env.rbs_size);
2149 (void) dump_block_list;
2154 insert_Loads_upwards();
2157 /* over all blocks in reverse post order */
2158 for (bl = env.forward; bl != NULL; bl = bl->forward_next) {
2159 do_replacements(bl);
2162 /* not only invalidate but free them. We might allocate new out arrays
2163 on our obstack which will be deleted yet. */
2164 confirm_irg_properties(irg, IR_GRAPH_PROPERTIES_CONTROL_FLOW);
2167 confirm_irg_properties(irg, IR_GRAPH_PROPERTIES_ALL);
2170 ir_free_resources(irg, IR_RESOURCE_IRN_LINK | IR_RESOURCE_BLOCK_MARK);
2171 ir_nodehashmap_destroy(&env.adr_map);
2172 obstack_free(&env.obst, NULL);
2174 #ifdef DEBUG_libfirm
2175 DEL_ARR_F(env.id_2_address);
2179 ir_graph_pass_t *opt_ldst_pass(const char *name)
2181 return def_graph_pass(name ? name : "ldst_df", opt_ldst);