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 static void dump_block_list(ldst_env *env) {
202 static void dump_curr(block_t *bl, const char *s) {
206 #endif /* DEBUG_libfirm */
208 /** Get the block entry for a block node */
209 static block_t *get_block_entry(const ir_node *block) {
210 assert(is_Block(block));
212 return get_irn_link(block);
213 } /* get_block_entry */
215 /** Get the memop entry for a memory operation node */
216 static memop_t *get_irn_memop(const ir_node *irn) {
217 assert(! is_Block(irn));
218 return get_irn_link(irn);
219 } /* get_irn_memop */
222 * Walk over the memory edges from definition to users.
223 * This ensures, that even operation without memory output are found.
225 * @param irn start node
226 * @param pre pre walker function
227 * @param post post walker function
228 * @param ctx context parameter for the walker functions
230 static void walk_memory(ir_node *irn, irg_walk_func *pre, irg_walk_func *post, void *ctx) {
234 mark_irn_visited(irn);
239 mode = get_irn_mode(irn);
240 if (mode == mode_M) {
241 /* every successor uses memory */
242 for (i = get_irn_n_outs(irn) - 1; i >= 0; --i) {
243 ir_node *succ = get_irn_out(irn, i);
245 if (! irn_visited(succ))
246 walk_memory(succ, pre, post, ctx);
248 } else if (mode == mode_T) {
249 /* only some Proj's uses memory */
250 for (i = get_irn_n_outs(irn) - 1; i >= 0; --i) {
251 ir_node *proj = get_irn_out(irn, i);
253 if (get_irn_mode(proj) == mode_M && ! irn_visited(proj))
254 walk_memory(proj, pre, post, ctx);
262 * Walks over all memory nodes of a graph.
265 * @param pre pre walker function
266 * @param post post walker function
267 * @param ctx context parameter for the walker functions
269 static void walk_memory_irg(ir_graph *irg, irg_walk_func pre, irg_walk_func post, void *ctx) {
270 inc_irg_visited(irg);
272 ir_reserve_resources(irg, IR_RESOURCE_IRN_VISITED);
275 * there are two possible sources for memory: initial_mem and nomem
276 * we ignore nomem as this should NOT change the memory
278 walk_memory(get_irg_initial_mem(irg), pre, post, ctx);
280 ir_free_resources(irg, IR_RESOURCE_IRN_VISITED);
281 } /* walk_memory_irg */
284 * Register an address and allocate a (sparse, 0..n) ID for it.
286 * @param adr the IR-node representing the address
288 * @return the allocated id
290 static unsigned register_address(ir_node *adr) {
291 address_entry *entry;
293 /* skip Confirms and Casts */
295 if (is_Confirm(adr)) {
296 adr = get_Confirm_value(adr);
300 adr = get_Cast_op(adr);
304 entry = ir_nodemap_get(&env.adr_map, adr);
308 entry = obstack_alloc(&env.obst, sizeof(*entry));
310 entry->id = env.curr_adr_id++;
311 ir_nodemap_insert(&env.adr_map, adr, entry);
313 DB((dbg, LEVEL_3, "ADDRESS %+F has ID %u\n", adr, entry->id));
315 ARR_APP1(ir_node *, env.id_2_address, adr);
319 } /* register_address */
323 * translate an address through a Phi node into a given predecessor
326 * @param address the address
327 * @param block the block
328 * @param pos the position of the predecessor in block
330 static ir_node *phi_translate(ir_node *address, const ir_node *block, int pos) {
331 if (is_Phi(address) && get_nodes_block(address) == block)
332 address = get_Phi_pred(address, pos);
334 } /* phi_translate */
337 * Walker: allocate an block entry for every block
338 * and register all potential addresses.
340 static void prepare_blocks(ir_node *irn, void *ctx) {
344 block_t *entry = obstack_alloc(&env.obst, sizeof(*entry));
347 entry->memop_forward = NULL;
348 entry->memop_backward = NULL;
349 entry->avail_out = NULL;
350 entry->id_2_memop_avail = NULL;
351 entry->anticL_in = NULL;
352 entry->id_2_memop_antic = NULL;
354 entry->forward_next = NULL;
355 entry->backward_next = NULL;
357 entry->trans_results = NULL;
358 set_irn_link(irn, entry);
360 set_Block_phis(irn, NULL);
362 /* use block marks to track unreachable blocks */
363 set_Block_mark(irn, 0);
365 n = get_Block_n_cfgpreds(irn);
366 if (n > env.max_cfg_preds)
367 env.max_cfg_preds = n;
369 ir_mode *mode = get_irn_mode(irn);
371 if (mode_is_reference(mode)) {
373 * Register ALL possible addresses: this is overkill yet but
374 * simpler then doing it for all possible translated addresses
375 * (which would be sufficient in the moment.
377 (void)register_address(irn);
380 } /* prepare_blocks */
383 * Post-Walker, link in all Phi's
385 static void link_phis(ir_node *irn, void *ctx) {
389 ir_node *block = get_nodes_block(irn);
390 add_Block_phi(block, irn);
395 * Block walker: creates the inverse post-order list for the CFG.
397 static void inverse_post_order(ir_node *block, void *ctx) {
398 block_t *entry = get_block_entry(block);
402 /* mark this block IS reachable from start */
403 set_Block_mark(block, 1);
405 /* create the list in inverse order */
406 entry->forward_next = env.forward;
409 /* remember the first visited (last in list) entry, needed for later */
410 if (env.backward == NULL)
411 env.backward = entry;
412 } /* inverse_post_order */
415 * Block walker: create backward links for the memops of a block.
417 static void collect_backward(ir_node *block, void *ctx) {
418 block_t *entry = get_block_entry(block);
424 * Do NOT link in the end block yet. We want it to be
425 * the first in the list. This is NOT guaranteed by the walker
426 * if we have endless loops.
428 if (block != env.end_bl) {
429 entry->backward_next = env.backward;
431 /* create the list in inverse order */
432 env.backward = entry;
435 /* create backward links for all memory ops */
437 for (op = entry->memop_forward; op != NULL; op = op->next) {
441 entry->memop_backward = last;
442 } /* collect_backward */
447 * @param irn the IR-node representing the memop or NULL
448 * if this is a translated (virtual) memop
450 * @return the allocated memop
452 static memop_t *alloc_memop(ir_node *irn) {
453 memop_t *m = obstack_alloc(&env.obst, sizeof(*m));
455 m->value.address = NULL;
456 m->value.value = NULL;
457 m->value.mode = NULL;
465 memset(m->projs, 0, sizeof(m->projs));
468 set_irn_link(irn, m);
473 * Create a memop for a Phi-replacement.
475 * @param op the memop to clone
476 * @param phi the Phi-node representing the new value
478 static memop_t *clone_memop_phi(memop_t *op, ir_node *phi) {
479 memop_t *m = obstack_alloc(&env.obst, sizeof(*m));
481 m->value = op->value;
482 m->value.value = phi;
489 set_irn_link(phi, m);
491 } /* clone_memop_phi */
494 * Return the memory properties of a call node.
496 * @param call the call node
498 * return a bitset of mtp_property_const and mtp_property_pure
500 static unsigned get_Call_memory_properties(ir_node *call) {
501 ir_type *call_tp = get_Call_type(call);
502 unsigned prop = get_method_additional_properties(call_tp);
504 /* check first the call type */
505 if ((prop & (mtp_property_const|mtp_property_pure)) == 0) {
506 /* try the called entity */
507 ir_node *ptr = get_Call_ptr(call);
509 if (is_Global(ptr)) {
510 ir_entity *ent = get_Global_entity(ptr);
512 prop = get_entity_additional_properties(ent);
515 return prop & (mtp_property_const|mtp_property_pure);
516 } /* get_Call_memory_properties */
519 * Returns an entity if the address ptr points to a constant one.
521 * @param ptr the address
523 * @return an entity or NULL
525 static ir_entity *find_constant_entity(ir_node *ptr) {
527 if (is_SymConst(ptr) && get_SymConst_kind(ptr) == symconst_addr_ent) {
528 return get_SymConst_entity(ptr);
529 } else if (is_Sel(ptr)) {
530 ir_entity *ent = get_Sel_entity(ptr);
531 ir_type *tp = get_entity_owner(ent);
533 /* Do not fiddle with polymorphism. */
534 if (is_Class_type(get_entity_owner(ent)) &&
535 ((get_entity_n_overwrites(ent) != 0) ||
536 (get_entity_n_overwrittenby(ent) != 0) ) )
539 if (is_Array_type(tp)) {
543 for (i = 0, n = get_Sel_n_indexs(ptr); i < n; ++i) {
545 tarval *tlower, *tupper;
546 ir_node *index = get_Sel_index(ptr, i);
547 tarval *tv = computed_value(index);
549 /* check if the index is constant */
550 if (tv == tarval_bad)
553 bound = get_array_lower_bound(tp, i);
554 tlower = computed_value(bound);
555 bound = get_array_upper_bound(tp, i);
556 tupper = computed_value(bound);
558 if (tlower == tarval_bad || tupper == tarval_bad)
561 if (tarval_cmp(tv, tlower) & pn_Cmp_Lt)
563 if (tarval_cmp(tupper, tv) & pn_Cmp_Lt)
566 /* ok, bounds check finished */
570 if (variability_constant == get_entity_variability(ent))
574 ptr = get_Sel_ptr(ptr);
575 } else if (is_Add(ptr)) {
576 ir_node *l = get_Add_left(ptr);
577 ir_node *r = get_Add_right(ptr);
579 if (get_irn_mode(l) == get_irn_mode(ptr) && is_Const(r))
581 else if (get_irn_mode(r) == get_irn_mode(ptr) && is_Const(l))
586 /* for now, we support only one addition, reassoc should fold all others */
587 if (! is_SymConst(ptr) && !is_Sel(ptr))
589 } else if (is_Sub(ptr)) {
590 ir_node *l = get_Sub_left(ptr);
591 ir_node *r = get_Sub_right(ptr);
593 if (get_irn_mode(l) == get_irn_mode(ptr) && is_Const(r))
597 /* for now, we support only one subtraction, reassoc should fold all others */
598 if (! is_SymConst(ptr) && !is_Sel(ptr))
603 } /* find_constant_entity */
606 * Return the Selection index of a Sel node from dimension n
608 static long get_Sel_array_index_long(ir_node *n, int dim) {
609 ir_node *index = get_Sel_index(n, dim);
610 assert(is_Const(index));
611 return get_tarval_long(get_Const_tarval(index));
612 } /* get_Sel_array_index_long */
615 * Returns the accessed component graph path for an
616 * node computing an address.
618 * @param ptr the node computing the address
619 * @param depth current depth in steps upward from the root
622 static compound_graph_path *rec_get_accessed_path(ir_node *ptr, int depth) {
623 compound_graph_path *res = NULL;
624 ir_entity *root, *field, *ent;
625 int path_len, pos, idx;
629 if (is_SymConst(ptr)) {
630 /* a SymConst. If the depth is 0, this is an access to a global
631 * entity and we don't need a component path, else we know
632 * at least its length.
634 assert(get_SymConst_kind(ptr) == symconst_addr_ent);
635 root = get_SymConst_entity(ptr);
636 res = (depth == 0) ? NULL : new_compound_graph_path(get_entity_type(root), depth);
637 } else if (is_Sel(ptr)) {
638 /* it's a Sel, go up until we find the root */
639 res = rec_get_accessed_path(get_Sel_ptr(ptr), depth+1);
643 /* fill up the step in the path at the current position */
644 field = get_Sel_entity(ptr);
645 path_len = get_compound_graph_path_length(res);
646 pos = path_len - depth - 1;
647 set_compound_graph_path_node(res, pos, field);
649 if (is_Array_type(get_entity_owner(field))) {
650 assert(get_Sel_n_indexs(ptr) == 1 && "multi dim arrays not implemented");
651 set_compound_graph_path_array_index(res, pos, get_Sel_array_index_long(ptr, 0));
653 } else if (is_Add(ptr)) {
654 ir_node *l = get_Add_left(ptr);
655 ir_node *r = get_Add_right(ptr);
656 ir_mode *mode = get_irn_mode(ptr);
659 if (is_Const(r) && get_irn_mode(l) == mode) {
661 tv = get_Const_tarval(r);
664 tv = get_Const_tarval(l);
667 mode = get_tarval_mode(tv);
670 /* ptr must be a Sel or a SymConst, this was checked in find_constant_entity() */
672 field = get_Sel_entity(ptr);
674 field = get_SymConst_entity(ptr);
677 for (ent = field;;) {
679 tarval *sz, *tv_index, *tlower, *tupper;
682 tp = get_entity_type(ent);
683 if (! is_Array_type(tp))
685 ent = get_array_element_entity(tp);
686 size = get_type_size_bytes(get_entity_type(ent));
687 sz = new_tarval_from_long(size, mode);
689 tv_index = tarval_div(tmp, sz);
690 tmp = tarval_mod(tmp, sz);
692 if (tv_index == tarval_bad || tmp == tarval_bad)
695 assert(get_array_n_dimensions(tp) == 1 && "multiarrays not implemented");
696 bound = get_array_lower_bound(tp, 0);
697 tlower = computed_value(bound);
698 bound = get_array_upper_bound(tp, 0);
699 tupper = computed_value(bound);
701 if (tlower == tarval_bad || tupper == tarval_bad)
704 if (tarval_cmp(tv_index, tlower) & pn_Cmp_Lt)
706 if (tarval_cmp(tupper, tv_index) & pn_Cmp_Lt)
709 /* ok, bounds check finished */
712 if (! tarval_is_null(tmp)) {
713 /* access to some struct/union member */
717 /* should be at least ONE array */
721 res = rec_get_accessed_path(ptr, depth + idx);
725 path_len = get_compound_graph_path_length(res);
726 pos = path_len - depth - idx;
728 for (ent = field;;) {
730 tarval *sz, *tv_index;
733 tp = get_entity_type(ent);
734 if (! is_Array_type(tp))
736 ent = get_array_element_entity(tp);
737 set_compound_graph_path_node(res, pos, ent);
739 size = get_type_size_bytes(get_entity_type(ent));
740 sz = new_tarval_from_long(size, mode);
742 tv_index = tarval_div(tv, sz);
743 tv = tarval_mod(tv, sz);
745 /* worked above, should work again */
746 assert(tv_index != tarval_bad && tv != tarval_bad);
748 /* bounds already checked above */
749 index = get_tarval_long(tv_index);
750 set_compound_graph_path_array_index(res, pos, index);
753 } else if (is_Sub(ptr)) {
754 ir_node *l = get_Sub_left(ptr);
755 ir_node *r = get_Sub_right(ptr);
758 tv = get_Const_tarval(r);
763 } /* rec_get_accessed_path */
766 * Returns an access path or NULL. The access path is only
767 * valid, if the graph is in phase_high and _no_ address computation is used.
769 static compound_graph_path *get_accessed_path(ir_node *ptr) {
770 compound_graph_path *gr = rec_get_accessed_path(ptr, 0);
772 } /* get_accessed_path */
774 typedef struct path_entry {
776 struct path_entry *next;
780 static ir_node *rec_find_compound_ent_value(ir_node *ptr, path_entry *next) {
781 path_entry entry, *p;
782 ir_entity *ent, *field;
783 ir_initializer_t *initializer;
789 if (is_SymConst(ptr)) {
791 ent = get_SymConst_entity(ptr);
792 initializer = get_entity_initializer(ent);
793 for (p = next; p != NULL;) {
794 if (initializer->kind != IR_INITIALIZER_COMPOUND)
796 n = get_initializer_compound_n_entries(initializer);
797 tp = get_entity_type(ent);
799 if (is_Array_type(tp)) {
800 ent = get_array_element_entity(tp);
805 initializer = get_initializer_compound_value(initializer, 0);
809 if (p->index >= (int) n)
811 initializer = get_initializer_compound_value(initializer, p->index);
816 tp = get_entity_type(ent);
817 while (is_Array_type(tp)) {
818 ent = get_array_element_entity(tp);
819 tp = get_entity_type(ent);
821 n = get_initializer_compound_n_entries(initializer);
824 initializer = get_initializer_compound_value(initializer, 0);
827 switch (initializer->kind) {
828 case IR_INITIALIZER_CONST:
829 return get_initializer_const_value(initializer);
830 case IR_INITIALIZER_TARVAL:
831 case IR_INITIALIZER_NULL:
835 } else if (is_Sel(ptr)) {
836 entry.ent = field = get_Sel_entity(ptr);
837 tp = get_entity_owner(field);
838 if (is_Array_type(tp)) {
839 assert(get_Sel_n_indexs(ptr) == 1 && "multi dim arrays not implemented");
840 entry.index = get_Sel_array_index_long(ptr, 0) - get_array_lower_bound_int(tp, 0);
842 int i, n_members = get_compound_n_members(tp);
843 for (i = 0; i < n_members; ++i) {
844 if (get_compound_member(tp, i) == field)
847 if (i >= n_members) {
848 /* not found: should NOT happen */
853 return rec_find_compound_ent_value(get_Sel_ptr(ptr), &entry);
854 } else if (is_Add(ptr)) {
855 ir_node *l = get_Add_left(ptr);
856 ir_node *r = get_Add_right(ptr);
862 tv = get_Const_tarval(r);
865 tv = get_Const_tarval(l);
868 mode = get_tarval_mode(tv);
870 /* ptr must be a Sel or a SymConst, this was checked in find_constant_entity() */
872 field = get_Sel_entity(ptr);
874 field = get_SymConst_entity(ptr);
877 /* count needed entries */
879 for (ent = field;;) {
880 tp = get_entity_type(ent);
881 if (! is_Array_type(tp))
883 ent = get_array_element_entity(tp);
886 /* should be at least ONE entry */
890 /* allocate the right number of entries */
891 NEW_ARR_A(path_entry, p, pos);
895 for (ent = field;;) {
897 tarval *sz, *tv_index, *tlower, *tupper;
901 tp = get_entity_type(ent);
902 if (! is_Array_type(tp))
904 ent = get_array_element_entity(tp);
906 p[pos].next = &p[pos + 1];
908 size = get_type_size_bytes(get_entity_type(ent));
909 sz = new_tarval_from_long(size, mode);
911 tv_index = tarval_div(tv, sz);
912 tv = tarval_mod(tv, sz);
914 if (tv_index == tarval_bad || tv == tarval_bad)
917 assert(get_array_n_dimensions(tp) == 1 && "multiarrays not implemented");
918 bound = get_array_lower_bound(tp, 0);
919 tlower = computed_value(bound);
920 bound = get_array_upper_bound(tp, 0);
921 tupper = computed_value(bound);
923 if (tlower == tarval_bad || tupper == tarval_bad)
926 if (tarval_cmp(tv_index, tlower) & pn_Cmp_Lt)
928 if (tarval_cmp(tupper, tv_index) & pn_Cmp_Lt)
931 /* ok, bounds check finished */
932 index = get_tarval_long(tv_index);
933 p[pos].index = index;
936 if (! tarval_is_null(tv)) {
937 /* hmm, wrong access */
940 p[pos - 1].next = next;
941 return rec_find_compound_ent_value(ptr, p);
942 } else if (is_Sub(ptr)) {
943 ir_node *l = get_Sub_left(ptr);
944 ir_node *r = get_Sub_right(ptr);
947 tv = get_Const_tarval(r);
952 } /* rec_find_compound_ent_value */
954 static ir_node *find_compound_ent_value(ir_node *ptr) {
955 return rec_find_compound_ent_value(ptr, NULL);
956 } /* find_compound_ent_value */
959 * Mark a Load memop to be replace by a definition
961 * @param op the Load memop
963 static void mark_replace_load(memop_t *op, ir_node *def) {
965 op->flags |= FLAG_KILLED_NODE;
967 } /* mark_replace_load */
970 * Mark a Store memop to be removed.
972 * @param op the Store memop
974 static void mark_remove_store(memop_t *op) {
975 op->flags |= FLAG_KILLED_NODE;
977 } /* mark_remove_store */
980 * Update a memop for a Load.
984 static void update_Load_memop(memop_t *m) {
986 ir_node *load = m->node;
990 if (get_Load_volatility(load) == volatility_is_volatile)
991 m->flags |= FLAG_IGNORE;
993 ptr = get_Load_ptr(load);
995 m->value.address = ptr;
997 for (i = get_irn_n_outs(load) - 1; i >= 0; --i) {
998 ir_node *proj = get_irn_out(load, i);
1001 /* beware of keep edges */
1005 pn = get_Proj_proj(proj);
1006 m->projs[pn] = proj;
1009 m->value.value = proj;
1010 m->value.mode = get_irn_mode(proj);
1012 case pn_Load_X_except:
1013 m->flags |= FLAG_EXCEPTION;
1018 case pn_Load_X_regular:
1021 panic("Unsupported Proj from Load %+F", proj);
1025 /* check if we can determine the entity that will be loaded */
1026 ent = find_constant_entity(ptr);
1029 allocation_static == get_entity_allocation(ent) &&
1030 visibility_external_allocated != get_entity_visibility(ent)) {
1031 /* a static allocation that is not external: there should be NO exception
1032 * when loading even if we cannot replace the load itself. */
1033 ir_node *value = NULL;
1035 /* no exception, clear the m fields as it might be checked later again */
1036 if (m->projs[pn_Load_X_except]) {
1037 exchange(m->projs[pn_Load_X_except], new_Bad());
1038 m->projs[pn_Load_X_except] = NULL;
1039 m->flags &= ~FLAG_EXCEPTION;
1042 if (m->projs[pn_Load_X_regular]) {
1043 exchange(m->projs[pn_Load_X_regular], new_r_Jmp(get_nodes_block(load)));
1044 m->projs[pn_Load_X_regular] = NULL;
1048 if (variability_constant == get_entity_variability(ent)) {
1049 if (is_atomic_entity(ent)) {
1050 /* Might not be atomic after lowering of Sels. In this case we
1051 * could also load, but it's more complicated. */
1052 /* more simpler case: we load the content of a constant value:
1053 * replace it by the constant itself */
1054 value = get_atomic_ent_value(ent);
1055 } else if (ent->has_initializer) {
1056 /* new style initializer */
1057 value = find_compound_ent_value(ptr);
1059 /* old style initializer */
1060 compound_graph_path *path = get_accessed_path(ptr);
1063 assert(is_proper_compound_graph_path(path, get_compound_graph_path_length(path)-1));
1065 value = get_compound_ent_value_by_path(ent, path);
1066 DB((dbg, LEVEL_1, " Constant access at %F%F resulted in %+F\n", ent, path, value));
1067 free_compound_graph_path(path);
1071 value = can_replace_load_by_const(load, value);
1074 if (value != NULL) {
1075 /* we completely replace the load by this value */
1076 DB((dbg, LEVEL_1, "Replacing Load %+F by constant %+F\n", m->node, value));
1077 mark_replace_load(m, value);
1082 if (m->value.value != NULL && !(m->flags & FLAG_IGNORE)) {
1083 /* only create an address if this node is NOT killed immediately or ignored */
1084 m->value.id = register_address(ptr);
1087 /* no user, KILL it */
1088 mark_replace_load(m, NULL);
1090 } /* update_Load_memop */
1093 * Update a memop for a Store.
1095 * @param m the memop
1097 static void update_Store_memop(memop_t *m) {
1099 ir_node *store = m->node;
1100 ir_node *adr = get_Store_ptr(store);
1102 if (get_Store_volatility(store) == volatility_is_volatile) {
1103 m->flags |= FLAG_IGNORE;
1105 /* only create an address if this node is NOT ignored */
1106 m->value.id = register_address(adr);
1110 m->value.address = adr;
1112 for (i = get_irn_n_outs(store) - 1; i >= 0; --i) {
1113 ir_node *proj = get_irn_out(store, i);
1116 /* beware of keep edges */
1120 pn = get_Proj_proj(proj);
1121 m->projs[pn] = proj;
1123 case pn_Store_X_except:
1124 m->flags |= FLAG_EXCEPTION;
1129 case pn_Store_X_regular:
1132 panic("Unsupported Proj from Store %+F", proj);
1135 m->value.value = get_Store_value(store);
1136 m->value.mode = get_irn_mode(m->value.value);
1137 } /* update_Store_memop */
1140 * Update a memop for a Call.
1142 * @param m the memop
1144 static void update_Call_memop(memop_t *m) {
1145 ir_node *call = m->node;
1146 unsigned prop = get_Call_memory_properties(call);
1149 if (prop & mtp_property_const) {
1150 /* A constant call did NOT use memory at all, we
1151 can kick it from the list. */
1152 } else if (prop & mtp_property_pure) {
1153 /* pure calls READ memory */
1156 m->flags = FLAG_KILL_ALL;
1158 for (i = get_irn_n_outs(call) - 1; i >= 0; --i) {
1159 ir_node *proj = get_irn_out(call, i);
1161 /* beware of keep edges */
1165 switch (get_Proj_proj(proj)) {
1166 case pn_Call_X_except:
1167 m->flags |= FLAG_EXCEPTION;
1169 case pn_Call_M_regular:
1174 } /* update_Call_memop */
1177 * Update a memop for a Div/Mod/Quot/DivMod.
1179 * @param m the memop
1181 static void update_DivOp_memop(memop_t *m) {
1182 ir_node *div = m->node;
1185 for (i = get_irn_n_outs(div) - 1; i >= 0; --i) {
1186 ir_node *proj = get_irn_out(div, i);
1188 /* beware of keep edges */
1192 switch (get_Proj_proj(proj)) {
1193 case pn_Generic_X_except:
1194 m->flags |= FLAG_EXCEPTION;
1196 case pn_Generic_M_regular:
1201 } /* update_DivOp_memop */
1204 * Update a memop for a Phi.
1206 * @param m the memop
1208 static void update_Phi_memop(memop_t *m) {
1209 /* the Phi is it's own mem */
1211 } /* update_Phi_memop */
1214 * Memory walker: collect all memory ops and build topological lists.
1216 static void collect_memops(ir_node *irn, void *ctx) {
1223 /* we can safely ignore ProjM's except the initial memory */
1224 if (irn != get_irg_initial_mem(current_ir_graph))
1228 op = alloc_memop(irn);
1229 block = get_nodes_block(irn);
1230 entry = get_block_entry(block);
1233 update_Phi_memop(op);
1234 /* Phis must be always placed first */
1235 op->next = entry->memop_forward;
1236 entry->memop_forward = op;
1237 if (entry->memop_backward == NULL)
1238 entry->memop_backward = op;
1240 switch (get_irn_opcode(irn)) {
1242 update_Load_memop(op);
1245 update_Store_memop(op);
1248 update_Call_memop(op);
1255 /* initial memory */
1260 /* we can those to find the memory edge */
1266 update_DivOp_memop(op);
1270 /* TODO: handle some builtins */
1272 /* unsupported operation */
1273 op->flags = FLAG_KILL_ALL;
1277 /* all other should be placed last */
1278 if (entry->memop_backward == NULL) {
1279 entry->memop_forward = entry->memop_backward = op;
1281 entry->memop_backward->next = op;
1282 entry->memop_backward = op;
1285 } /* collect_memops */
1288 * Find an address in the current set.
1290 * @param value the value to be searched for
1292 * @return a memop for the value or NULL if the value does
1293 * not exists in the set or cannot be converted into
1294 * the requested mode
1296 static memop_t *find_address(const value_t *value) {
1297 if (rbitset_is_set(env.curr_set, value->id)) {
1298 memop_t *res = env.curr_id_2_memop[value->id];
1300 if (res->value.mode == value->mode)
1302 /* allow hidden casts */
1303 if (get_mode_arithmetic(res->value.mode) == irma_twos_complement &&
1304 get_mode_arithmetic(value->mode) == irma_twos_complement &&
1305 get_mode_size_bits(res->value.mode) == get_mode_size_bits(value->mode))
1309 } /* find_address */
1312 * Find an address in the avail_out set.
1314 * @param bl the block
1316 static memop_t *find_address_avail(const block_t *bl, unsigned id, const ir_mode *mode) {
1317 if (rbitset_is_set(bl->avail_out, id)) {
1318 memop_t *res = bl->id_2_memop_avail[id];
1320 if (res->value.mode == mode)
1322 /* allow hidden casts */
1323 if (get_mode_arithmetic(res->value.mode) == irma_twos_complement &&
1324 get_mode_arithmetic(mode) == irma_twos_complement &&
1325 get_mode_size_bits(res->value.mode) == get_mode_size_bits(mode))
1329 } /* find_address_avail */
1332 * Kill all addresses from the current set.
1334 static void kill_all(void) {
1335 rbitset_clear_all(env.curr_set, env.rbs_size);
1338 rbitset_set(env.curr_set, env.rbs_size - 1);
1342 * Kill memops that are not alias free due to a Store value from the current set.
1344 * @param value the Store value
1346 static void kill_memops(const value_t *value) {
1347 unsigned end = env.rbs_size - 1;
1350 for (pos = rbitset_next(env.curr_set, 0, 1); pos < end; pos = rbitset_next(env.curr_set, pos + 1, 1)) {
1351 memop_t *op = env.curr_id_2_memop[pos];
1353 if (ir_no_alias != get_alias_relation(current_ir_graph, value->address, value->mode,
1354 op->value.address, op->value.mode)) {
1355 rbitset_clear(env.curr_set, pos);
1356 env.curr_id_2_memop[pos] = NULL;
1357 DB((dbg, LEVEL_2, "KILLING %+F because of possible alias address %+F\n", op->node, value->address));
1363 * Add the value of a memop to the current set.
1365 * @param op the memory op
1367 static void add_memop(memop_t *op) {
1368 rbitset_set(env.curr_set, op->value.id);
1369 env.curr_id_2_memop[op->value.id] = op;
1373 * Add the value of a memop to the avail_out set.
1375 * @param bl the block
1376 * @param op the memory op
1378 static void add_memop_avail(block_t *bl, memop_t *op) {
1379 rbitset_set(bl->avail_out, op->value.id);
1380 bl->id_2_memop_avail[op->value.id] = op;
1381 } /* add_memop_avail */
1384 * Check, if we can convert a value of one mode to another mode
1385 * without changing the representation of bits.
1387 * @param from the original mode
1388 * @param to the destination mode
1390 static int can_convert_to(const ir_mode *from, const ir_mode *to) {
1391 if (get_mode_arithmetic(from) == irma_twos_complement &&
1392 get_mode_arithmetic(to) == irma_twos_complement &&
1393 get_mode_size_bits(from) == get_mode_size_bits(to))
1396 } /* can_convert_to */
1399 * Add a Conv to the requested mode if needed.
1401 * @param irn the IR-node to convert
1402 * @param mode the destination mode
1404 * @return the possible converted node or NULL
1405 * if the conversion is not possible
1407 static ir_node *conv_to(ir_node *irn, ir_mode *mode) {
1408 ir_mode *other = get_irn_mode(irn);
1409 if (other != mode) {
1410 /* different modes: check if conversion is possible without changing the bits */
1411 if (can_convert_to(other, mode)) {
1412 ir_node *block = get_nodes_block(irn);
1413 return new_r_Conv(block, irn, mode);
1415 /* otherwise not possible ... yet */
1422 * Update the address of an value if this address was a load result
1423 * and the load is killed now.
1425 * @param value the value whose address is updated
1427 static void update_address(value_t *value) {
1428 if (is_Proj(value->address)) {
1429 ir_node *load = get_Proj_pred(value->address);
1431 if (is_Load(load)) {
1432 const memop_t *op = get_irn_memop(load);
1434 if (op->flags & FLAG_KILLED_NODE)
1435 value->address = op->replace;
1438 } /* update_address */
1441 * Do forward dataflow analysis on the given block and calculate the
1442 * GEN and KILL in the current (avail) set.
1444 * @param bl the block
1446 static void calc_gen_kill_avail(block_t *bl) {
1450 for (op = bl->memop_forward; op != NULL; op = op->next) {
1451 switch (get_irn_opcode(op->node)) {
1459 if (! (op->flags & (FLAG_KILLED_NODE|FLAG_IGNORE))) {
1460 /* do we have this already? */
1463 update_address(&op->value);
1464 other = find_address(&op->value);
1465 if (other != NULL && other != op) {
1466 def = conv_to(other->value.value, op->value.mode);
1468 #ifdef DEBUG_libfirm
1469 if (is_Store(other->node)) {
1471 DB((dbg, LEVEL_1, "RAW %+F <- %+F(%+F)\n", op->node, def, other->node));
1474 DB((dbg, LEVEL_1, "RAR %+F <- %+F(%+F)\n", op->node, def, other->node));
1477 mark_replace_load(op, def);
1478 /* do NOT change the memop table */
1482 /* add this value */
1487 if (! (op->flags & FLAG_KILLED_NODE)) {
1488 /* do we have this store already */
1491 update_address(&op->value);
1492 other = find_address(&op->value);
1493 if (other != NULL) {
1494 if (is_Store(other->node)) {
1495 if (op != other && !(other->flags & FLAG_IGNORE) &&
1496 get_nodes_block(other->node) == get_nodes_block(op->node)) {
1498 * A WAW in the same block we can kick the first store.
1499 * This is a shortcut: we know that the second Store will be anticipated
1502 DB((dbg, LEVEL_1, "WAW %+F <- %+F\n", other->node, op->node));
1503 mark_remove_store(other);
1504 /* FIXME: a Load might be get freed due to this killed store */
1506 } else if (other->value.value == op->value.value && !(op->flags & FLAG_IGNORE)) {
1508 DB((dbg, LEVEL_1, "WAR %+F <- %+F\n", op->node, other->node));
1509 mark_remove_store(op);
1510 /* do NOT change the memop table */
1514 /* KILL all possible aliases */
1515 kill_memops(&op->value);
1516 /* add this value */
1521 if (op->flags & FLAG_KILL_ALL)
1525 } /* calc_gen_kill_avail */
1527 #define BYTE_SIZE(x) (((x) + 7) >> 3)
1530 * Do forward dataflow analysis on a given block to calculate the avail_out set
1531 * for this block only.
1533 * @param block the block
1535 static void forward_avail(block_t *bl) {
1536 /* fill the data from the current block */
1537 env.curr_id_2_memop = bl->id_2_memop_avail;
1538 env.curr_set = bl->avail_out;
1540 calc_gen_kill_avail(bl);
1541 dump_curr(bl, "Avail_out");
1542 } /* forward_avail */
1545 * Do backward dataflow analysis on a given block to calculate the antic set
1546 * of Loaded addresses.
1548 * @param bl the block
1550 * @return non-zero if the set has changed since last iteration
1552 static int backward_antic(block_t *bl) {
1554 ir_node *block = bl->block;
1555 int n = get_Block_n_cfg_outs(block);
1558 ir_node *succ = get_Block_cfg_out(block, 0);
1559 block_t *succ_bl = get_block_entry(succ);
1560 int pred_pos = get_Block_cfgpred_pos(succ, block);
1561 unsigned end = env.rbs_size - 1;
1566 if (bl->trans_results == NULL) {
1567 /* allocate the translate cache */
1568 unsigned size = env.curr_adr_id * sizeof(bl->trans_results[0]);
1569 bl->trans_results = obstack_alloc(&env.obst, size);
1570 memset(bl->trans_results, 0, size);
1573 /* check for partly redundant values */
1574 for (pos = rbitset_next(succ_bl->anticL_in, 0, 1);
1576 pos = rbitset_next(succ_bl->anticL_in, pos + 1, 1)) {
1578 * do Phi-translation here: Note that at this point the nodes are
1579 * not changed, so we can safely cache the results.
1580 * However: Loads of Load results ARE bad, because we have no way
1581 to translate them yet ...
1583 memop_t *op = bl->trans_results[pos];
1585 /* not yet translated */
1586 ir_node *adr, *trans_adr;
1588 op = succ_bl->id_2_memop_antic[pos];
1589 adr = op->value.address;
1591 trans_adr = phi_translate(adr, succ, pred_pos);
1592 if (trans_adr != adr) {
1593 /* create a new entry for the translated one */
1596 new_op = alloc_memop(NULL);
1597 new_op->value.address = trans_adr;
1598 new_op->value.id = register_address(trans_adr);
1599 new_op->value.mode = op->value.mode;
1600 new_op->node = op->node; /* we need the node to decide if Load/Store */
1601 new_op->flags = op->flags;
1603 bl->trans_results[pos] = new_op;
1607 env.curr_id_2_memop[op->value.id] = op;
1608 rbitset_set(env.curr_set, op->value.id);
1611 ir_node *succ = get_Block_cfg_out(block, 0);
1612 block_t *succ_bl = get_block_entry(succ);
1615 rbitset_copy(env.curr_set, succ_bl->anticL_in, env.rbs_size);
1616 memcpy(env.curr_id_2_memop, succ_bl->id_2_memop_antic, env.rbs_size * sizeof(env.curr_id_2_memop[0]));
1618 /* Hmm: probably we want kill merges of Loads ans Stores here */
1619 for (i = n - 1; i > 0; --i) {
1620 ir_node *succ = get_Block_cfg_out(bl->block, i);
1621 block_t *succ_bl = get_block_entry(succ);
1623 rbitset_and(env.curr_set, succ_bl->anticL_in, env.rbs_size);
1626 /* block ends with a noreturn call */
1630 dump_curr(bl, "AnticL_out");
1632 for (op = bl->memop_backward; op != NULL; op = op->prev) {
1633 switch (get_irn_opcode(op->node)) {
1641 if (! (op->flags & (FLAG_KILLED_NODE|FLAG_IGNORE))) {
1647 if (! (op->flags & FLAG_KILLED_NODE)) {
1648 /* a Store: check which memops must be killed */
1649 kill_memops(&op->value);
1653 if (op->flags & FLAG_KILL_ALL)
1658 memcpy(bl->id_2_memop_antic, env.curr_id_2_memop, env.rbs_size * sizeof(env.curr_id_2_memop[0]));
1659 if (! rbitset_equal(bl->anticL_in, env.curr_set, env.rbs_size)) {
1661 rbitset_copy(bl->anticL_in, env.curr_set, env.rbs_size);
1662 dump_curr(bl, "AnticL_in*");
1665 dump_curr(bl, "AnticL_in");
1667 } /* backward_antic */
1670 * Replace a Load memop by a already known value.
1672 * @param op the Load memop
1674 static void replace_load(memop_t *op) {
1675 ir_node *load = op->node;
1676 ir_node *def = skip_Id(op->replace);
1681 DB((dbg, LEVEL_1, "Replacing %+F by definition %+F\n", load, is_Proj(def) ? get_Proj_pred(def) : def));
1683 if (op->flags & FLAG_EXCEPTION) {
1684 /* bad: this node is unused and executed for exception only */
1685 DB((dbg, LEVEL_1, "Unused %+F executed for exception only ...\n", load));
1688 DB((dbg, LEVEL_1, "Killing unused %+F\n", load));
1691 if (op->mem != NULL) {
1692 /* in rare cases a Load might have NO memory */
1693 exchange(op->mem, get_Load_mem(load));
1695 proj = op->projs[pn_Load_res];
1697 mode = get_irn_mode(proj);
1698 if (get_irn_mode(def) != mode) {
1700 dbg_info *db = get_irn_dbg_info(load);
1701 ir_node *block = get_nodes_block(proj);
1702 def = new_rd_Conv(db, block, def, mode);
1704 exchange(proj, def);
1706 proj = op->projs[pn_Load_X_except];
1708 exchange(proj, new_Bad());
1710 proj = op->projs[pn_Load_X_regular];
1712 exchange(proj, new_r_Jmp(get_nodes_block(load)));
1714 } /* replace_load */
1717 * Remove a Store memop.
1719 * @param op the Store memop
1721 static void remove_store(memop_t *op) {
1722 ir_node *store = op->node;
1725 DB((dbg, LEVEL_1, "Removing %+F\n", store));
1727 if (op->mem != NULL) {
1728 /* in rare cases a Store might have no memory */
1729 exchange(op->mem, get_Store_mem(store));
1731 proj = op->projs[pn_Store_X_except];
1733 exchange(proj, new_Bad());
1735 proj = op->projs[pn_Store_X_regular];
1737 exchange(proj, new_r_Jmp(get_nodes_block(store)));
1739 } /* remove_store */
1743 * Do all necessary replacements for a given block.
1745 * @param bl the block
1747 static void do_replacements(block_t *bl) {
1750 for (op = bl->memop_forward; op != NULL; op = op->next) {
1751 if (op->flags & FLAG_KILLED_NODE) {
1752 switch (get_irn_opcode(op->node)) {
1762 } /* do_replacements */
1765 * Calculate the Avail_out sets for all basic blocks.
1767 static void calcAvail(void) {
1768 memop_t **tmp_memop = env.curr_id_2_memop;
1769 unsigned *tmp_set = env.curr_set;
1772 /* calculate avail_out */
1773 DB((dbg, LEVEL_2, "Calculate Avail_out\n"));
1775 /* iterate over all blocks in in any order, skip the start block */
1776 for (bl = env.forward->forward_next; bl != NULL; bl = bl->forward_next) {
1780 /* restore the current sets */
1781 env.curr_id_2_memop = tmp_memop;
1782 env.curr_set = tmp_set;
1786 * Calculate the Antic_in sets for all basic blocks.
1788 static void calcAntic(void) {
1791 /* calculate antic_out */
1792 DB((dbg, LEVEL_2, "Calculate Antic_in\n"));
1797 DB((dbg, LEVEL_2, "Iteration %d:\n=========\n", i));
1801 /* over all blocks in reverse post order */
1802 for (bl = env.backward->backward_next; bl != NULL; bl = bl->backward_next) {
1803 need_iter |= backward_antic(bl);
1806 } while (need_iter);
1807 DB((dbg, LEVEL_2, "Get anticipated Load set after %d iterations\n", i));
1811 * Return the node representing the last memory in a block.
1813 * @param bl the block
1815 static ir_node *find_last_memory(block_t *bl) {
1817 if (bl->memop_backward != NULL) {
1818 return bl->memop_backward->mem;
1820 /* if there is NO memory in this block, go to the dominator */
1821 bl = get_block_entry(get_Block_idom(bl->block));
1823 } /* find_last_memory */
1826 * Reroute all memory users of old memory
1827 * to a new memory IR-node.
1829 * @param omem the old memory IR-node
1830 * @param nmem the new memory IR-node
1832 static void reroute_all_mem_users(ir_node *omem, ir_node *nmem) {
1835 for (i = get_irn_n_outs(omem) - 1; i >= 0; --i) {
1837 ir_node *user = get_irn_out_ex(omem, i, &n_pos);
1839 set_irn_n(user, n_pos, nmem);
1842 /* all edges previously point to omem now point to nmem */
1843 nmem->out = omem->out;
1844 } /* reroute_all_mem_users */
1847 * Reroute memory users of old memory that are dominated by a given block
1848 * to a new memory IR-node.
1850 * @param omem the old memory IR-node
1851 * @param nmem the new memory IR-node
1852 * @param pass_bl the block the memory must pass
1854 static void reroute_mem_through(ir_node *omem, ir_node *nmem, ir_node *pass_bl) {
1855 int i, j, n = get_irn_n_outs(omem);
1856 ir_def_use_edge *edges = NEW_ARR_D(ir_def_use_edge, &env.obst, n + 1);
1858 for (i = j = 0; i < n; ++i) {
1860 ir_node *user = get_irn_out_ex(omem, i, &n_pos);
1861 ir_node *use_bl = get_nodes_block(user);
1865 use_bl = get_Block_cfgpred_block(use_bl, n_pos);
1867 if (block_dominates(pass_bl, use_bl)) {
1868 /* found an user that is dominated */
1870 edges[j].pos = n_pos;
1871 edges[j].use = user;
1873 set_irn_n(user, n_pos, nmem);
1877 /* Modify the out structure: we create a new out edge array on our
1878 temporary obstack here. This should be no problem, as we invalidate the edges
1879 at the end either. */
1880 /* first entry is used for the length */
1883 } /* reroute_mem_through */
1886 * insert Loads, making partly redundant Loads fully redundant
1888 static int insert_Load(block_t *bl) {
1889 ir_node *block = bl->block;
1890 int i, n = get_Block_n_cfgpreds(block);
1891 unsigned end = env.rbs_size - 1;
1894 DB((dbg, LEVEL_3, "processing %+F\n", block));
1897 /* might still happen for an unreachable block (end for instance) */
1905 NEW_ARR_A(ir_node *, ins, n);
1907 rbitset_set_all(env.curr_set, env.rbs_size);
1909 /* More than one predecessors, calculate the join for all avail_outs ignoring unevaluated
1910 Blocks. These put in Top anyway. */
1911 for (i = n - 1; i >= 0; --i) {
1912 ir_node *pred = skip_Proj(get_Block_cfgpred(block, i));
1913 ir_node *blk = get_nodes_block(pred);
1916 pred_bl = get_block_entry(blk);
1917 rbitset_and(env.curr_set, pred_bl->avail_out, env.rbs_size);
1919 if (is_Load(pred) || is_Store(pred)) {
1920 /* We reached this block by an exception from a Load or Store:
1921 * the memop creating the exception was NOT completed than, kill it
1923 memop_t *exc_op = get_irn_memop(pred);
1924 rbitset_clear(env.curr_set, exc_op->value.id);
1929 * Ensure that all values are in the map: build Phi's if necessary:
1930 * Note: the last bit is the sentinel and ALWAYS set, so start with -2.
1932 for (pos = env.rbs_size - 2; pos >= 0; --pos) {
1933 if (! rbitset_is_set(env.curr_set, pos))
1934 env.curr_id_2_memop[pos] = NULL;
1936 ir_node *pred = get_Block_cfgpred_block(bl->block, 0);
1937 block_t *pred_bl = get_block_entry(pred);
1939 memop_t *first = NULL;
1940 ir_mode *mode = NULL;
1942 for (i = 0; i < n; ++i) {
1945 pred = get_Block_cfgpred_block(bl->block, i);
1946 pred_bl = get_block_entry(pred);
1948 mop = pred_bl->id_2_memop_avail[pos];
1949 if (first == NULL) {
1951 ins[0] = first->value.value;
1952 mode = get_irn_mode(ins[0]);
1954 /* no Phi needed so far */
1955 env.curr_id_2_memop[pos] = first;
1957 ins[i] = conv_to(mop->value.value, mode);
1958 if (ins[i] != ins[0]) {
1959 if (ins[i] == NULL) {
1960 /* conversion failed */
1961 env.curr_id_2_memop[pos] = NULL;
1962 rbitset_clear(env.curr_set, pos);
1971 ir_node *phi = new_r_Phi(bl->block, n, ins, mode);
1972 memop_t *phiop = alloc_memop(phi);
1974 phiop->value = first->value;
1975 phiop->value.value = phi;
1977 /* no need to link it in, as it is a DATA phi */
1979 env.curr_id_2_memop[pos] = phiop;
1981 DB((dbg, LEVEL_3, "Created new %+F on merging value for address %+F\n", phi, first->value.address));
1986 /* only one predecessor, simply copy the map */
1987 ir_node *pred = get_Block_cfgpred_block(bl->block, 0);
1988 block_t *pred_bl = get_block_entry(pred);
1990 rbitset_copy(env.curr_set, pred_bl->avail_out, env.rbs_size);
1992 memcpy(env.curr_id_2_memop, pred_bl->id_2_memop_avail, env.rbs_size * sizeof(bl->id_2_memop_avail[0]));
1996 /* check for partly redundant values */
1997 for (pos = rbitset_next(bl->anticL_in, 0, 1);
1999 pos = rbitset_next(bl->anticL_in, pos + 1, 1)) {
2000 memop_t *op = bl->id_2_memop_antic[pos];
2001 int have_some, all_same;
2004 if (rbitset_is_set(env.curr_set, pos)) {
2009 assert(is_Load(op->node));
2011 DB((dbg, LEVEL_3, "anticipated %+F\n", op->node));
2016 for (i = n - 1; i >= 0; --i) {
2017 ir_node *pred = get_Block_cfgpred_block(block, i);
2018 block_t *pred_bl = get_block_entry(pred);
2019 ir_mode *mode = op->value.mode;
2023 adr = phi_translate(op->value.address, block, i);
2024 DB((dbg, LEVEL_3, ".. using address %+F in pred %d\n", adr, i));
2025 e = find_address_avail(pred_bl, register_address(adr), mode);
2027 ir_node *ef_block = get_nodes_block(adr);
2028 if (! block_dominates(ef_block, pred)) {
2029 /* cannot place a copy here */
2031 DB((dbg, LEVEL_3, "%+F cannot be moved into predecessor %+F\n", op->node, pred));
2034 DB((dbg, LEVEL_3, "%+F is not available in predecessor %+F\n", op->node, pred));
2035 pred_bl->avail = NULL;
2038 if (e->value.mode != mode && !can_convert_to(e->value.mode, mode)) {
2039 /* cannot create a Phi due to different modes */
2045 DB((dbg, LEVEL_3, "%+F is available for %+F in predecessor %+F\n", e->node, op->node, pred));
2048 else if (first != e->node)
2052 if (have_some && !all_same) {
2053 ir_mode *mode = op->value.mode;
2057 NEW_ARR_A(ir_node *, in, n);
2059 for (i = n - 1; i >= 0; --i) {
2060 ir_node *pred = get_Block_cfgpred_block(block, i);
2061 block_t *pred_bl = get_block_entry(pred);
2063 if (pred_bl->avail == NULL) {
2064 /* create a new Load here and make to make it fully redundant */
2065 dbg_info *db = get_irn_dbg_info(op->node);
2066 ir_node *last_mem = find_last_memory(pred_bl);
2067 ir_node *load, *def, *adr;
2070 assert(last_mem != NULL);
2071 adr = phi_translate(op->value.address, block, i);
2072 load = new_rd_Load(db, pred, last_mem, adr, mode, cons_none);
2073 def = new_r_Proj(pred, load, mode, pn_Load_res);
2074 DB((dbg, LEVEL_1, "Created new %+F in %+F for party redundant %+F\n", load, pred, op->node));
2076 new_op = alloc_memop(load);
2077 new_op->mem = new_r_Proj(pred, load, mode_M, pn_Load_M);
2078 new_op->value.address = adr;
2079 new_op->value.id = op->value.id;
2080 new_op->value.mode = mode;
2081 new_op->value.value = def;
2083 new_op->projs[pn_Load_M] = new_op->mem;
2084 new_op->projs[pn_Load_res] = def;
2086 new_op->prev = pred_bl->memop_backward;
2087 if (pred_bl->memop_backward != NULL)
2088 pred_bl->memop_backward->next = new_op;
2090 pred_bl->memop_backward = new_op;
2092 if (pred_bl->memop_forward == NULL)
2093 pred_bl->memop_forward = new_op;
2095 if (get_nodes_block(last_mem) == pred) {
2096 /* We have add a new last memory op in pred block.
2097 If pred had already a last mem, reroute all memory
2099 reroute_all_mem_users(last_mem, new_op->mem);
2101 /* reroute only those memory going through the pre block */
2102 reroute_mem_through(last_mem, new_op->mem, pred);
2105 /* we added this load at the end, so it will be avail anyway */
2106 add_memop_avail(pred_bl, new_op);
2107 pred_bl->avail = new_op;
2109 in[i] = conv_to(pred_bl->avail->value.value, mode);
2111 phi = new_r_Phi(block, n, in, mode);
2112 DB((dbg, LEVEL_1, "Created new %+F in %+F for now redundant %+F\n", phi, block, op->node));
2114 phi_op = clone_memop_phi(op, phi);
2120 /* recalculate avail by gen and kill */
2121 calc_gen_kill_avail(bl);
2123 /* always update the map after gen/kill, as values might have been changed due to RAR/WAR/WAW */
2124 memcpy(bl->id_2_memop_avail, env.curr_id_2_memop, env.rbs_size * sizeof(env.curr_id_2_memop[0]));
2126 if (!rbitset_equal(bl->avail_out, env.curr_set, env.rbs_size)) {
2127 /* the avail set has changed */
2128 rbitset_copy(bl->avail_out, env.curr_set, env.rbs_size);
2129 dump_curr(bl, "Avail_out*");
2132 dump_curr(bl, "Avail_out");
2137 * Insert Loads upwards.
2139 static void insert_Loads_upwards(void) {
2143 /* recalculate antic_out and insert Loads */
2144 DB((dbg, LEVEL_2, "Inserting Loads\n"));
2148 DB((dbg, LEVEL_2, "Iteration %d:\n=========\n", i));
2152 /* over all blocks in reverse post order, skip the start block */
2153 for (bl = env.forward->forward_next; bl != NULL; bl = bl->forward_next) {
2154 need_iter |= insert_Load(bl);
2157 } while (need_iter);
2159 DB((dbg, LEVEL_2, "Finished Load inserting after %d iterations\n", i));
2160 } /* insert_Loads_upwards */
2163 * Kill unreachable control flow.
2165 * @param irg the graph to operate on
2167 static void kill_unreachable_blocks(ir_graph *irg) {
2172 NEW_ARR_A(ir_node *, ins, env.max_cfg_preds);
2174 for (bl = env.forward; bl != NULL; bl = bl->forward_next) {
2175 ir_node *block = bl->block;
2178 assert(get_Block_mark(block));
2180 n = get_Block_n_cfgpreds(block);
2182 for (i = j = 0; i < n; ++i) {
2183 ir_node *pred = get_Block_cfgpred(block, i);
2189 pred_bl = get_nodes_block(skip_Proj(pred));
2190 if (! get_Block_mark(pred_bl))
2196 ir_node *phi, *next;
2198 /* some unreachable blocks detected */
2201 DB((dbg, LEVEL_1, "Killing dead block predecessors on %+F\n", block));
2203 set_irn_in(block, j, ins);
2205 /* shorten all Phi nodes */
2206 for (phi = get_Block_phis(block); phi != NULL; phi = next) {
2207 next = get_Phi_next(phi);
2209 for (i = k = 0; i < n; ++i) {
2210 ir_node *pred = get_Block_cfgpred_block(block, i);
2215 if (! get_Block_mark(pred))
2218 ins[k++] = get_Phi_pred(phi, i);
2221 exchange(phi, ins[0]);
2223 set_irn_in(phi, k, ins);
2230 /* kick keep alives */
2231 ir_node *end = get_irg_end(irg);
2232 int i, j, n = get_End_n_keepalives(end);
2234 NEW_ARR_A(ir_node *, ins, n);
2236 for (i = j = 0; i < n; ++i) {
2237 ir_node *ka = get_End_keepalive(end, i);
2245 ka_bl = get_nodes_block(skip_Proj(ka));
2246 if (get_Block_mark(ka_bl))
2250 set_End_keepalives(end, j, ins);
2254 /* this transformation do NOT invalidate the dominance */
2256 } /* kill_unreachable_blocks */
2258 int opt_ldst(ir_graph *irg) {
2260 ir_graph *rem = current_ir_graph;
2262 current_ir_graph = irg;
2264 FIRM_DBG_REGISTER(dbg, "firm.opt.ldst");
2265 // firm_dbg_set_mask(dbg, -1);
2267 DB((dbg, LEVEL_1, "\nDoing Load/Store optimization on %+F\n", irg));
2269 /* we need landing pads */
2270 remove_critical_cf_edges(irg);
2272 // dump_ir_block_graph(irg, "-XXX");
2274 if (get_opt_alias_analysis()) {
2275 assure_irg_entity_usage_computed(irg);
2276 assure_irp_globals_entity_usage_computed();
2279 obstack_init(&env.obst);
2280 ir_nodemap_init(&env.adr_map);
2283 env.backward = NULL;
2284 env.curr_adr_id = 0;
2286 env.max_cfg_preds = 0;
2288 env.start_bl = get_irg_start_block(irg);
2289 env.end_bl = get_irg_end_block(irg);
2290 #ifdef DEBUG_libfirm
2291 env.id_2_address = NEW_ARR_F(ir_node *, 0);
2294 assure_irg_outs(irg);
2296 ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK | IR_RESOURCE_BLOCK_MARK);
2298 /* first step: allocate block entries. Note that some blocks might be
2299 unreachable here. Using the normal walk ensures that ALL blocks are initialized. */
2300 irg_walk_graph(irg, prepare_blocks, link_phis, NULL);
2302 /* produce an inverse post-order list for the CFG: this links only reachable
2304 irg_out_block_walk(get_irg_start_block(irg), NULL, inverse_post_order, NULL);
2306 if (! get_Block_mark(env.end_bl)) {
2308 * The end block is NOT reachable due to endless loops
2309 * or no_return calls.
2310 * Place the end block last.
2311 * env.backward points to the last block in the list for this purpose.
2313 env.backward->forward_next = get_block_entry(env.end_bl);
2315 set_Block_mark(env.end_bl, 1);
2318 /* KILL unreachable blocks: these disturb the data flow analysis */
2319 kill_unreachable_blocks(irg);
2323 /* second step: find and sort all memory ops */
2324 walk_memory_irg(irg, collect_memops, NULL, NULL);
2326 #ifdef DEBUG_libfirm
2327 /* check that the backward map is correct */
2328 assert((unsigned)ARR_LEN(env.id_2_address) == env.curr_adr_id);
2331 if (env.n_mem_ops == 0) {
2336 /* create the backward links. */
2337 env.backward = NULL;
2338 irg_block_walk_graph(irg, NULL, collect_backward, NULL);
2340 /* link the end block in */
2341 bl = get_block_entry(env.end_bl);
2342 bl->backward_next = env.backward;
2345 /* check that we really start with the start / end block */
2346 assert(env.forward->block == env.start_bl);
2347 assert(env.backward->block == env.end_bl);
2349 /* create address sets: for now, only the existing addresses are allowed plus one
2350 needed for the sentinel */
2351 env.rbs_size = env.curr_adr_id + 1;
2353 /* create the current set */
2354 env.curr_set = rbitset_obstack_alloc(&env.obst, env.rbs_size);
2355 rbitset_set(env.curr_set, env.rbs_size - 1);
2356 env.curr_id_2_memop = NEW_ARR_D(memop_t *, &env.obst, env.rbs_size);
2357 memset(env.curr_id_2_memop, 0, env.rbs_size * sizeof(env.curr_id_2_memop[0]));
2359 for (bl = env.forward; bl != NULL; bl = bl->forward_next) {
2360 /* set sentinel bits */
2361 bl->avail_out = rbitset_obstack_alloc(&env.obst, env.rbs_size);
2362 rbitset_set(bl->avail_out, env.rbs_size - 1);
2364 bl->id_2_memop_avail = NEW_ARR_D(memop_t *, &env.obst, env.rbs_size);
2365 memset(bl->id_2_memop_avail, 0, env.rbs_size * sizeof(bl->id_2_memop_avail[0]));
2367 bl->anticL_in = rbitset_obstack_alloc(&env.obst, env.rbs_size);
2368 rbitset_set(bl->anticL_in, env.rbs_size - 1);
2370 bl->id_2_memop_antic = NEW_ARR_D(memop_t *, &env.obst, env.rbs_size);
2371 memset(bl->id_2_memop_antic, 0, env.rbs_size * sizeof(bl->id_2_memop_antic[0]));
2374 // dump_block_list(&env);
2375 (void) dump_block_list;
2380 insert_Loads_upwards();
2383 /* over all blocks in reverse post order */
2384 for (bl = env.forward; bl != NULL; bl = bl->forward_next) {
2385 do_replacements(bl);
2388 /* not only invalidate but free them. We might allocate new out arrays
2389 on our obstack which will be deleted yet. */
2391 set_irg_entity_usage_state(irg, ir_entity_usage_not_computed);
2395 ir_free_resources(irg, IR_RESOURCE_IRN_LINK | IR_RESOURCE_BLOCK_MARK);
2396 ir_nodemap_destroy(&env.adr_map);
2397 obstack_free(&env.obst, NULL);
2399 // dump_ir_block_graph(irg, "-YYY");
2401 #ifdef DEBUG_libfirm
2402 DEL_ARR_F(env.id_2_address);
2405 current_ir_graph = rem;
2406 return env.changed != 0;