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 Load/Store optimizations.
23 * @author Michael Beck
30 #include "iroptimize.h"
32 #include "irgraph_t.h"
40 #include "dbginfo_t.h"
41 #include "iropt_dbg.h"
47 #include "opt_polymorphy.h"
49 #include "irphase_t.h"
54 /** The debug handle. */
55 DEBUG_ONLY(static firm_dbg_module_t *dbg;)
58 #include "cacheopt/cachesim.h"
62 #define IMAX(a,b) ((a) > (b) ? (a) : (b))
64 #define MAX_PROJ IMAX(IMAX(pn_Load_max, pn_Store_max), pn_Call_max)
67 DF_CHANGED = 1, /**< data flow changed */
68 CF_CHANGED = 2, /**< control flow changed */
74 typedef struct _walk_env_t {
75 struct obstack obst; /**< list of all stores */
76 unsigned changes; /**< a bitmask of graph changes */
79 /** A Load/Store info. */
80 typedef struct _ldst_info_t {
81 ir_node *projs[MAX_PROJ]; /**< list of Proj's of this node */
82 ir_node *exc_block; /**< the exception block if available */
83 int exc_idx; /**< predecessor index in the exception block */
84 unsigned visited; /**< visited counter for breaking loops */
88 * flags for control flow.
91 BLOCK_HAS_COND = 1, /**< Block has conditional control flow */
92 BLOCK_HAS_EXC = 2 /**< Block has exceptional control flow */
98 typedef struct _block_info_t {
99 unsigned flags; /**< flags for the block */
102 /** the master visited flag for loop detection. */
103 static unsigned master_visited = 0;
105 #define INC_MASTER() ++master_visited
106 #define MARK_NODE(info) (info)->visited = master_visited
107 #define NODE_VISITED(info) (info)->visited >= master_visited
110 * get the Load/Store info of a node
112 static ldst_info_t *get_ldst_info(ir_node *node, struct obstack *obst)
114 ldst_info_t *info = get_irn_link(node);
117 info = OALLOCZ(obst, ldst_info_t);
118 set_irn_link(node, info);
121 } /* get_ldst_info */
124 * get the Block info of a node
126 static block_info_t *get_block_info(ir_node *node, struct obstack *obst)
128 block_info_t *info = get_irn_link(node);
131 info = OALLOCZ(obst, block_info_t);
132 set_irn_link(node, info);
135 } /* get_block_info */
138 * update the projection info for a Load/Store
140 static unsigned update_projs(ldst_info_t *info, ir_node *proj)
142 long nr = get_Proj_proj(proj);
144 assert(0 <= nr && nr <= MAX_PROJ && "Wrong proj from LoadStore");
146 if (info->projs[nr]) {
147 /* there is already one, do CSE */
148 exchange(proj, info->projs[nr]);
152 info->projs[nr] = proj;
158 * update the exception block info for a Load/Store node.
160 * @param info the load/store info struct
161 * @param block the exception handler block for this load/store
162 * @param pos the control flow input of the block
164 static unsigned update_exc(ldst_info_t *info, ir_node *block, int pos)
166 assert(info->exc_block == NULL && "more than one exception block found");
168 info->exc_block = block;
173 /** Return the number of uses of an address node */
174 #define get_irn_n_uses(adr) get_irn_n_edges(adr)
177 * walker, collects all Load/Store/Proj nodes
179 * walks from Start -> End
181 static void collect_nodes(ir_node *node, void *env)
183 ir_opcode opcode = get_irn_opcode(node);
184 ir_node *pred, *blk, *pred_blk;
185 ldst_info_t *ldst_info;
186 walk_env_t *wenv = env;
188 if (opcode == iro_Proj) {
189 pred = get_Proj_pred(node);
190 opcode = get_irn_opcode(pred);
192 if (opcode == iro_Load || opcode == iro_Store || opcode == iro_Call) {
193 ldst_info = get_ldst_info(pred, &wenv->obst);
195 wenv->changes |= update_projs(ldst_info, node);
198 * Place the Proj's to the same block as the
199 * predecessor Load. This is always ok and prevents
200 * "non-SSA" form after optimizations if the Proj
201 * is in a wrong block.
203 blk = get_nodes_block(node);
204 pred_blk = get_nodes_block(pred);
205 if (blk != pred_blk) {
206 wenv->changes |= DF_CHANGED;
207 set_nodes_block(node, pred_blk);
210 } else if (opcode == iro_Block) {
213 for (i = get_Block_n_cfgpreds(node) - 1; i >= 0; --i) {
214 ir_node *pred_block, *proj;
215 block_info_t *bl_info;
218 pred = proj = get_Block_cfgpred(node, i);
221 pred = get_Proj_pred(proj);
222 is_exc = get_Proj_proj(proj) == pn_Generic_X_except;
225 /* ignore Bad predecessors, they will be removed later */
229 pred_block = get_nodes_block(pred);
230 bl_info = get_block_info(pred_block, &wenv->obst);
232 if (is_fragile_op(pred) && is_exc)
233 bl_info->flags |= BLOCK_HAS_EXC;
234 else if (is_irn_forking(pred))
235 bl_info->flags |= BLOCK_HAS_COND;
237 opcode = get_irn_opcode(pred);
238 if (is_exc && (opcode == iro_Load || opcode == iro_Store || opcode == iro_Call)) {
239 ldst_info = get_ldst_info(pred, &wenv->obst);
241 wenv->changes |= update_exc(ldst_info, node, i);
245 } /* collect_nodes */
248 * Returns an entity if the address ptr points to a constant one.
250 * @param ptr the address
252 * @return an entity or NULL
254 static ir_entity *find_constant_entity(ir_node *ptr)
257 if (is_SymConst(ptr) && get_SymConst_kind(ptr) == symconst_addr_ent) {
258 return get_SymConst_entity(ptr);
259 } else if (is_Sel(ptr)) {
260 ir_entity *ent = get_Sel_entity(ptr);
261 ir_type *tp = get_entity_owner(ent);
263 /* Do not fiddle with polymorphism. */
264 if (is_Class_type(get_entity_owner(ent)) &&
265 ((get_entity_n_overwrites(ent) != 0) ||
266 (get_entity_n_overwrittenby(ent) != 0) ) )
269 if (is_Array_type(tp)) {
273 for (i = 0, n = get_Sel_n_indexs(ptr); i < n; ++i) {
275 tarval *tlower, *tupper;
276 ir_node *index = get_Sel_index(ptr, i);
277 tarval *tv = computed_value(index);
279 /* check if the index is constant */
280 if (tv == tarval_bad)
283 bound = get_array_lower_bound(tp, i);
284 tlower = computed_value(bound);
285 bound = get_array_upper_bound(tp, i);
286 tupper = computed_value(bound);
288 if (tlower == tarval_bad || tupper == tarval_bad)
291 if (tarval_cmp(tv, tlower) & pn_Cmp_Lt)
293 if (tarval_cmp(tupper, tv) & pn_Cmp_Lt)
296 /* ok, bounds check finished */
300 if (get_entity_linkage(ent) & IR_LINKAGE_CONSTANT)
304 ptr = get_Sel_ptr(ptr);
305 } else if (is_Add(ptr)) {
306 ir_node *l = get_Add_left(ptr);
307 ir_node *r = get_Add_right(ptr);
309 if (get_irn_mode(l) == get_irn_mode(ptr) && is_Const(r))
311 else if (get_irn_mode(r) == get_irn_mode(ptr) && is_Const(l))
316 /* for now, we support only one addition, reassoc should fold all others */
317 if (! is_SymConst(ptr) && !is_Sel(ptr))
319 } else if (is_Sub(ptr)) {
320 ir_node *l = get_Sub_left(ptr);
321 ir_node *r = get_Sub_right(ptr);
323 if (get_irn_mode(l) == get_irn_mode(ptr) && is_Const(r))
327 /* for now, we support only one substraction, reassoc should fold all others */
328 if (! is_SymConst(ptr) && !is_Sel(ptr))
333 } /* find_constant_entity */
336 * Return the Selection index of a Sel node from dimension n
338 static long get_Sel_array_index_long(ir_node *n, int dim)
340 ir_node *index = get_Sel_index(n, dim);
341 assert(is_Const(index));
342 return get_tarval_long(get_Const_tarval(index));
343 } /* get_Sel_array_index_long */
346 * Returns the accessed component graph path for an
347 * node computing an address.
349 * @param ptr the node computing the address
350 * @param depth current depth in steps upward from the root
353 static compound_graph_path *rec_get_accessed_path(ir_node *ptr, int depth)
355 compound_graph_path *res = NULL;
356 ir_entity *root, *field, *ent;
357 int path_len, pos, idx;
361 if (is_SymConst(ptr)) {
362 /* a SymConst. If the depth is 0, this is an access to a global
363 * entity and we don't need a component path, else we know
364 * at least its length.
366 assert(get_SymConst_kind(ptr) == symconst_addr_ent);
367 root = get_SymConst_entity(ptr);
368 res = (depth == 0) ? NULL : new_compound_graph_path(get_entity_type(root), depth);
369 } else if (is_Sel(ptr)) {
370 /* it's a Sel, go up until we find the root */
371 res = rec_get_accessed_path(get_Sel_ptr(ptr), depth+1);
375 /* fill up the step in the path at the current position */
376 field = get_Sel_entity(ptr);
377 path_len = get_compound_graph_path_length(res);
378 pos = path_len - depth - 1;
379 set_compound_graph_path_node(res, pos, field);
381 if (is_Array_type(get_entity_owner(field))) {
382 assert(get_Sel_n_indexs(ptr) == 1 && "multi dim arrays not implemented");
383 set_compound_graph_path_array_index(res, pos, get_Sel_array_index_long(ptr, 0));
385 } else if (is_Add(ptr)) {
386 ir_node *l = get_Add_left(ptr);
387 ir_node *r = get_Add_right(ptr);
388 ir_mode *mode = get_irn_mode(ptr);
391 if (is_Const(r) && get_irn_mode(l) == mode) {
393 tv = get_Const_tarval(r);
396 tv = get_Const_tarval(l);
399 mode = get_tarval_mode(tv);
402 /* ptr must be a Sel or a SymConst, this was checked in find_constant_entity() */
404 field = get_Sel_entity(ptr);
406 field = get_SymConst_entity(ptr);
409 for (ent = field;;) {
411 tarval *sz, *tv_index, *tlower, *tupper;
414 tp = get_entity_type(ent);
415 if (! is_Array_type(tp))
417 ent = get_array_element_entity(tp);
418 size = get_type_size_bytes(get_entity_type(ent));
419 sz = new_tarval_from_long(size, mode);
421 tv_index = tarval_div(tmp, sz);
422 tmp = tarval_mod(tmp, sz);
424 if (tv_index == tarval_bad || tmp == tarval_bad)
427 assert(get_array_n_dimensions(tp) == 1 && "multiarrays not implemented");
428 bound = get_array_lower_bound(tp, 0);
429 tlower = computed_value(bound);
430 bound = get_array_upper_bound(tp, 0);
431 tupper = computed_value(bound);
433 if (tlower == tarval_bad || tupper == tarval_bad)
436 if (tarval_cmp(tv_index, tlower) & pn_Cmp_Lt)
438 if (tarval_cmp(tupper, tv_index) & pn_Cmp_Lt)
441 /* ok, bounds check finished */
444 if (! tarval_is_null(tmp)) {
445 /* access to some struct/union member */
449 /* should be at least ONE array */
453 res = rec_get_accessed_path(ptr, depth + idx);
457 path_len = get_compound_graph_path_length(res);
458 pos = path_len - depth - idx;
460 for (ent = field;;) {
462 tarval *sz, *tv_index;
465 tp = get_entity_type(ent);
466 if (! is_Array_type(tp))
468 ent = get_array_element_entity(tp);
469 set_compound_graph_path_node(res, pos, ent);
471 size = get_type_size_bytes(get_entity_type(ent));
472 sz = new_tarval_from_long(size, mode);
474 tv_index = tarval_div(tv, sz);
475 tv = tarval_mod(tv, sz);
477 /* worked above, should work again */
478 assert(tv_index != tarval_bad && tv != tarval_bad);
480 /* bounds already checked above */
481 index = get_tarval_long(tv_index);
482 set_compound_graph_path_array_index(res, pos, index);
485 } else if (is_Sub(ptr)) {
486 ir_node *l = get_Sub_left(ptr);
487 ir_node *r = get_Sub_right(ptr);
490 tv = get_Const_tarval(r);
495 } /* rec_get_accessed_path */
498 * Returns an access path or NULL. The access path is only
499 * valid, if the graph is in phase_high and _no_ address computation is used.
501 static compound_graph_path *get_accessed_path(ir_node *ptr)
503 compound_graph_path *gr = rec_get_accessed_path(ptr, 0);
505 } /* get_accessed_path */
507 typedef struct path_entry {
509 struct path_entry *next;
513 static ir_node *rec_find_compound_ent_value(ir_node *ptr, path_entry *next)
515 path_entry entry, *p;
516 ir_entity *ent, *field;
517 ir_initializer_t *initializer;
523 if (is_SymConst(ptr)) {
525 ent = get_SymConst_entity(ptr);
526 initializer = get_entity_initializer(ent);
527 for (p = next; p != NULL;) {
528 if (initializer->kind != IR_INITIALIZER_COMPOUND)
530 n = get_initializer_compound_n_entries(initializer);
531 tp = get_entity_type(ent);
533 if (is_Array_type(tp)) {
534 ent = get_array_element_entity(tp);
539 initializer = get_initializer_compound_value(initializer, 0);
543 if (p->index >= (int) n)
545 initializer = get_initializer_compound_value(initializer, p->index);
550 tp = get_entity_type(ent);
551 while (is_Array_type(tp)) {
552 ent = get_array_element_entity(tp);
553 tp = get_entity_type(ent);
555 n = get_initializer_compound_n_entries(initializer);
558 initializer = get_initializer_compound_value(initializer, 0);
561 switch (initializer->kind) {
562 case IR_INITIALIZER_CONST:
563 return get_initializer_const_value(initializer);
564 case IR_INITIALIZER_TARVAL:
565 case IR_INITIALIZER_NULL:
569 } else if (is_Sel(ptr)) {
570 entry.ent = field = get_Sel_entity(ptr);
571 tp = get_entity_owner(field);
572 if (is_Array_type(tp)) {
573 assert(get_Sel_n_indexs(ptr) == 1 && "multi dim arrays not implemented");
574 entry.index = get_Sel_array_index_long(ptr, 0) - get_array_lower_bound_int(tp, 0);
576 int i, n_members = get_compound_n_members(tp);
577 for (i = 0; i < n_members; ++i) {
578 if (get_compound_member(tp, i) == field)
581 if (i >= n_members) {
582 /* not found: should NOT happen */
587 return rec_find_compound_ent_value(get_Sel_ptr(ptr), &entry);
588 } else if (is_Add(ptr)) {
589 ir_node *l = get_Add_left(ptr);
590 ir_node *r = get_Add_right(ptr);
596 tv = get_Const_tarval(r);
599 tv = get_Const_tarval(l);
602 mode = get_tarval_mode(tv);
604 /* ptr must be a Sel or a SymConst, this was checked in find_constant_entity() */
606 field = get_Sel_entity(ptr);
608 field = get_SymConst_entity(ptr);
611 /* count needed entries */
613 for (ent = field;;) {
614 tp = get_entity_type(ent);
615 if (! is_Array_type(tp))
617 ent = get_array_element_entity(tp);
620 /* should be at least ONE entry */
624 /* allocate the right number of entries */
625 NEW_ARR_A(path_entry, p, pos);
629 for (ent = field;;) {
631 tarval *sz, *tv_index, *tlower, *tupper;
635 tp = get_entity_type(ent);
636 if (! is_Array_type(tp))
638 ent = get_array_element_entity(tp);
640 p[pos].next = &p[pos + 1];
642 size = get_type_size_bytes(get_entity_type(ent));
643 sz = new_tarval_from_long(size, mode);
645 tv_index = tarval_div(tv, sz);
646 tv = tarval_mod(tv, sz);
648 if (tv_index == tarval_bad || tv == tarval_bad)
651 assert(get_array_n_dimensions(tp) == 1 && "multiarrays not implemented");
652 bound = get_array_lower_bound(tp, 0);
653 tlower = computed_value(bound);
654 bound = get_array_upper_bound(tp, 0);
655 tupper = computed_value(bound);
657 if (tlower == tarval_bad || tupper == tarval_bad)
660 if (tarval_cmp(tv_index, tlower) & pn_Cmp_Lt)
662 if (tarval_cmp(tupper, tv_index) & pn_Cmp_Lt)
665 /* ok, bounds check finished */
666 index = get_tarval_long(tv_index);
667 p[pos].index = index;
670 if (! tarval_is_null(tv)) {
671 /* hmm, wrong access */
674 p[pos - 1].next = next;
675 return rec_find_compound_ent_value(ptr, p);
676 } else if (is_Sub(ptr)) {
677 ir_node *l = get_Sub_left(ptr);
678 ir_node *r = get_Sub_right(ptr);
681 tv = get_Const_tarval(r);
688 static ir_node *find_compound_ent_value(ir_node *ptr)
690 return rec_find_compound_ent_value(ptr, NULL);
694 static void reduce_adr_usage(ir_node *ptr);
697 * Update a Load that may have lost its users.
699 static void handle_load_update(ir_node *load)
701 ldst_info_t *info = get_irn_link(load);
703 /* do NOT touch volatile loads for now */
704 if (get_Load_volatility(load) == volatility_is_volatile)
707 if (! info->projs[pn_Load_res] && ! info->projs[pn_Load_X_except]) {
708 ir_node *ptr = get_Load_ptr(load);
709 ir_node *mem = get_Load_mem(load);
711 /* a Load whose value is neither used nor exception checked, remove it */
712 exchange(info->projs[pn_Load_M], mem);
713 if (info->projs[pn_Load_X_regular])
714 exchange(info->projs[pn_Load_X_regular], new_r_Jmp(get_nodes_block(load)));
716 reduce_adr_usage(ptr);
718 } /* handle_load_update */
721 * A use of an address node has vanished. Check if this was a Proj
722 * node and update the counters.
724 static void reduce_adr_usage(ir_node *ptr)
729 if (get_irn_n_edges(ptr) > 0)
732 /* this Proj is dead now */
733 pred = get_Proj_pred(ptr);
735 ldst_info_t *info = get_irn_link(pred);
736 info->projs[get_Proj_proj(ptr)] = NULL;
738 /* this node lost its result proj, handle that */
739 handle_load_update(pred);
741 } /* reduce_adr_usage */
744 * Check, if an already existing value of mode old_mode can be converted
745 * into the needed one new_mode without loss.
747 static int can_use_stored_value(ir_mode *old_mode, ir_mode *new_mode)
749 if (old_mode == new_mode)
752 /* if both modes are two-complement ones, we can always convert the
753 Stored value into the needed one. */
754 if (get_mode_size_bits(old_mode) >= get_mode_size_bits(new_mode) &&
755 get_mode_arithmetic(old_mode) == irma_twos_complement &&
756 get_mode_arithmetic(new_mode) == irma_twos_complement)
759 } /* can_use_stored_value */
762 * Check whether a Call is at least pure, ie. does only read memory.
764 static unsigned is_Call_pure(ir_node *call)
766 ir_type *call_tp = get_Call_type(call);
767 unsigned prop = get_method_additional_properties(call_tp);
769 /* check first the call type */
770 if ((prop & (mtp_property_const|mtp_property_pure)) == 0) {
771 /* try the called entity */
772 ir_node *ptr = get_Call_ptr(call);
774 if (is_Global(ptr)) {
775 ir_entity *ent = get_Global_entity(ptr);
777 prop = get_entity_additional_properties(ent);
780 return (prop & (mtp_property_const|mtp_property_pure)) != 0;
783 static ir_node *get_base_and_offset(ir_node *ptr, long *pOffset)
785 ir_mode *mode = get_irn_mode(ptr);
788 /* TODO: long might not be enough, we should probably use some tarval thingy... */
791 ir_node *l = get_Add_left(ptr);
792 ir_node *r = get_Add_right(ptr);
794 if (get_irn_mode(l) != mode || !is_Const(r))
797 offset += get_tarval_long(get_Const_tarval(r));
799 } else if (is_Sub(ptr)) {
800 ir_node *l = get_Sub_left(ptr);
801 ir_node *r = get_Sub_right(ptr);
803 if (get_irn_mode(l) != mode || !is_Const(r))
806 offset -= get_tarval_long(get_Const_tarval(r));
808 } else if (is_Sel(ptr)) {
809 ir_entity *ent = get_Sel_entity(ptr);
810 ir_type *tp = get_entity_owner(ent);
812 if (is_Array_type(tp)) {
816 /* only one dimensional arrays yet */
817 if (get_Sel_n_indexs(ptr) != 1)
819 index = get_Sel_index(ptr, 0);
820 if (! is_Const(index))
823 tp = get_entity_type(ent);
824 if (get_type_state(tp) != layout_fixed)
827 size = get_type_size_bytes(tp);
828 offset += size * get_tarval_long(get_Const_tarval(index));
830 if (get_type_state(tp) != layout_fixed)
832 offset += get_entity_offset(ent);
834 ptr = get_Sel_ptr(ptr);
843 static int try_load_after_store(ir_node *load,
844 ir_node *load_base_ptr, long load_offset, ir_node *store)
847 ir_node *store_ptr = get_Store_ptr(store);
849 ir_node *store_base_ptr = get_base_and_offset(store_ptr, &store_offset);
850 ir_node *store_value;
859 if (load_base_ptr != store_base_ptr)
862 load_mode = get_Load_mode(load);
863 load_mode_len = get_mode_size_bytes(load_mode);
864 store_mode = get_irn_mode(get_Store_value(store));
865 store_mode_len = get_mode_size_bytes(store_mode);
866 delta = load_offset - store_offset;
867 store_value = get_Store_value(store);
869 if (delta != 0 || store_mode != load_mode) {
870 if (delta < 0 || delta + load_mode_len > store_mode_len)
873 if (get_mode_arithmetic(store_mode) != irma_twos_complement ||
874 get_mode_arithmetic(load_mode) != irma_twos_complement)
878 /* produce a shift to adjust offset delta */
882 /* FIXME: only true for little endian */
883 cnst = new_Const_long(mode_Iu, delta * 8);
884 store_value = new_r_Shr(get_nodes_block(load),
885 store_value, cnst, store_mode);
888 /* add an convert if needed */
889 if (store_mode != load_mode) {
890 store_value = new_r_Conv(get_nodes_block(load), store_value, load_mode);
894 DBG_OPT_RAW(load, store_value);
896 info = get_irn_link(load);
897 if (info->projs[pn_Load_M])
898 exchange(info->projs[pn_Load_M], get_Load_mem(load));
902 if (info->projs[pn_Load_X_except]) {
903 exchange( info->projs[pn_Load_X_except], new_Bad());
906 if (info->projs[pn_Load_X_regular]) {
907 exchange( info->projs[pn_Load_X_regular], new_r_Jmp(get_nodes_block(load)));
911 if (info->projs[pn_Load_res])
912 exchange(info->projs[pn_Load_res], store_value);
914 load_ptr = get_Load_ptr(load);
916 reduce_adr_usage(load_ptr);
917 return res | DF_CHANGED;
921 * Follow the memory chain as long as there are only Loads,
922 * alias free Stores, and constant Calls and try to replace the
923 * current Load by a previous ones.
924 * Note that in unreachable loops it might happen that we reach
925 * load again, as well as we can fall into a cycle.
926 * We break such cycles using a special visited flag.
928 * INC_MASTER() must be called before dive into
930 static unsigned follow_Mem_chain(ir_node *load, ir_node *curr)
933 ldst_info_t *info = get_irn_link(load);
935 ir_node *ptr = get_Load_ptr(load);
936 ir_node *mem = get_Load_mem(load);
937 ir_mode *load_mode = get_Load_mode(load);
939 for (pred = curr; load != pred; ) {
940 ldst_info_t *pred_info = get_irn_link(pred);
943 * a Load immediately after a Store -- a read after write.
944 * We may remove the Load, if both Load & Store does not have an
945 * exception handler OR they are in the same MacroBlock. In the latter
946 * case the Load cannot throw an exception when the previous Store was
949 * Why we need to check for Store Exception? If the Store cannot
950 * be executed (ROM) the exception handler might simply jump into
951 * the load MacroBlock :-(
952 * We could make it a little bit better if we would know that the
953 * exception handler of the Store jumps directly to the end...
955 if (is_Store(pred) && ((pred_info->projs[pn_Store_X_except] == NULL
956 && info->projs[pn_Load_X_except] == NULL)
957 || get_nodes_MacroBlock(load) == get_nodes_MacroBlock(pred)))
960 ir_node *base_ptr = get_base_and_offset(ptr, &load_offset);
961 int changes = try_load_after_store(load, base_ptr, load_offset, pred);
964 return res | changes;
965 } else if (is_Load(pred) && get_Load_ptr(pred) == ptr &&
966 can_use_stored_value(get_Load_mode(pred), load_mode)) {
968 * a Load after a Load -- a read after read.
969 * We may remove the second Load, if it does not have an exception handler
970 * OR they are in the same MacroBlock. In the later case the Load cannot
971 * throw an exception when the previous Load was quiet.
973 * Here, there is no need to check if the previous Load has an exception
974 * hander because they would have exact the same exception...
976 if (info->projs[pn_Load_X_except] == NULL || get_nodes_MacroBlock(load) == get_nodes_MacroBlock(pred)) {
979 DBG_OPT_RAR(load, pred);
981 /* the result is used */
982 if (info->projs[pn_Load_res]) {
983 if (pred_info->projs[pn_Load_res] == NULL) {
984 /* create a new Proj again */
985 pred_info->projs[pn_Load_res] = new_r_Proj(pred, get_Load_mode(pred), pn_Load_res);
987 value = pred_info->projs[pn_Load_res];
989 /* add an convert if needed */
990 if (get_Load_mode(pred) != load_mode) {
991 value = new_r_Conv(get_nodes_block(load), value, load_mode);
994 exchange(info->projs[pn_Load_res], value);
997 if (info->projs[pn_Load_M])
998 exchange(info->projs[pn_Load_M], mem);
1001 if (info->projs[pn_Load_X_except]) {
1002 exchange(info->projs[pn_Load_X_except], new_Bad());
1005 if (info->projs[pn_Load_X_regular]) {
1006 exchange( info->projs[pn_Load_X_regular], new_r_Jmp(get_nodes_block(load)));
1011 reduce_adr_usage(ptr);
1012 return res |= DF_CHANGED;
1016 if (is_Store(pred)) {
1017 /* check if we can pass through this store */
1018 ir_alias_relation rel = get_alias_relation(
1020 get_Store_ptr(pred),
1021 get_irn_mode(get_Store_value(pred)),
1023 /* if the might be an alias, we cannot pass this Store */
1024 if (rel != ir_no_alias)
1026 pred = skip_Proj(get_Store_mem(pred));
1027 } else if (is_Load(pred)) {
1028 pred = skip_Proj(get_Load_mem(pred));
1029 } else if (is_Call(pred)) {
1030 if (is_Call_pure(pred)) {
1031 /* The called graph is at least pure, so there are no Store's
1032 in it. We can handle it like a Load and skip it. */
1033 pred = skip_Proj(get_Call_mem(pred));
1035 /* there might be Store's in the graph, stop here */
1039 /* follow only Load chains */
1043 /* check for cycles */
1044 if (NODE_VISITED(pred_info))
1046 MARK_NODE(pred_info);
1049 if (is_Sync(pred)) {
1052 /* handle all Sync predecessors */
1053 for (i = get_Sync_n_preds(pred) - 1; i >= 0; --i) {
1054 res |= follow_Mem_chain(load, skip_Proj(get_Sync_pred(pred, i)));
1061 } /* follow_Mem_chain */
1064 * Check if we can replace the load by a given const from
1065 * the const code irg.
1067 ir_node *can_replace_load_by_const(const ir_node *load, ir_node *c)
1069 ir_mode *c_mode = get_irn_mode(c);
1070 ir_mode *l_mode = get_Load_mode(load);
1071 ir_node *res = NULL;
1073 if (c_mode != l_mode) {
1074 /* check, if the mode matches OR can be easily converted info */
1075 if (is_reinterpret_cast(c_mode, l_mode)) {
1076 /* we can safely cast */
1077 dbg_info *dbg = get_irn_dbg_info(load);
1078 ir_node *block = get_nodes_block(load);
1080 /* copy the value from the const code irg and cast it */
1081 res = copy_const_value(dbg, c);
1082 res = new_rd_Conv(dbg, block, res, l_mode);
1085 /* copy the value from the const code irg */
1086 res = copy_const_value(get_irn_dbg_info(load), c);
1089 } /* can_replace_load_by_const */
1094 * @param load the Load node
1096 static unsigned optimize_load(ir_node *load)
1098 ldst_info_t *info = get_irn_link(load);
1099 ir_node *mem, *ptr, *value;
1104 /* do NOT touch volatile loads for now */
1105 if (get_Load_volatility(load) == volatility_is_volatile)
1108 /* the address of the load to be optimized */
1109 ptr = get_Load_ptr(load);
1111 /* The mem of the Load. Must still be returned after optimization. */
1112 mem = get_Load_mem(load);
1114 if (info->projs[pn_Load_res] == NULL
1115 && info->projs[pn_Load_X_except] == NULL) {
1116 /* the value is never used and we don't care about exceptions, remove */
1117 exchange(info->projs[pn_Load_M], mem);
1119 if (info->projs[pn_Load_X_regular]) {
1120 /* should not happen, but if it does, remove it */
1121 exchange(info->projs[pn_Load_X_regular], new_r_Jmp(get_nodes_block(load)));
1125 reduce_adr_usage(ptr);
1126 return res | DF_CHANGED;
1129 /* Load from a constant polymorphic field, where we can resolve
1131 value = transform_polymorph_Load(load);
1132 if (value == load) {
1134 /* check if we can determine the entity that will be loaded */
1135 ent = find_constant_entity(ptr);
1137 && get_entity_visibility(ent) != ir_visibility_external) {
1138 /* a static allocation that is not external: there should be NO
1139 * exception when loading even if we cannot replace the load itself.
1142 /* no exception, clear the info field as it might be checked later again */
1143 if (info->projs[pn_Load_X_except]) {
1144 exchange(info->projs[pn_Load_X_except], new_Bad());
1145 info->projs[pn_Load_X_except] = NULL;
1148 if (info->projs[pn_Load_X_regular]) {
1149 exchange(info->projs[pn_Load_X_regular], new_r_Jmp(get_nodes_block(load)));
1150 info->projs[pn_Load_X_regular] = NULL;
1154 if (get_entity_linkage(ent) & IR_LINKAGE_CONSTANT) {
1155 if (ent->initializer != NULL) {
1156 /* new style initializer */
1157 value = find_compound_ent_value(ptr);
1158 } else if (entity_has_compound_ent_values(ent)) {
1159 /* old style initializer */
1160 compound_graph_path *path = get_accessed_path(ptr);
1163 assert(is_proper_compound_graph_path(path, get_compound_graph_path_length(path)-1));
1165 value = get_compound_ent_value_by_path(ent, path);
1166 DB((dbg, LEVEL_1, " Constant access at %F%F resulted in %+F\n", ent, path, value));
1167 free_compound_graph_path(path);
1171 value = can_replace_load_by_const(load, value);
1175 if (value != NULL) {
1176 /* we completely replace the load by this value */
1177 if (info->projs[pn_Load_X_except]) {
1178 exchange(info->projs[pn_Load_X_except], new_Bad());
1179 info->projs[pn_Load_X_except] = NULL;
1182 if (info->projs[pn_Load_X_regular]) {
1183 exchange(info->projs[pn_Load_X_regular], new_r_Jmp(get_nodes_block(load)));
1184 info->projs[pn_Load_X_regular] = NULL;
1187 if (info->projs[pn_Load_M]) {
1188 exchange(info->projs[pn_Load_M], mem);
1191 if (info->projs[pn_Load_res]) {
1192 exchange(info->projs[pn_Load_res], value);
1196 reduce_adr_usage(ptr);
1200 /* Check, if the address of this load is used more than once.
1201 * If not, more load cannot be removed in any case. */
1202 if (get_irn_n_uses(ptr) <= 1 && get_irn_n_uses(get_base_and_offset(ptr, &dummy)) <= 1)
1206 * follow the memory chain as long as there are only Loads
1207 * and try to replace current Load or Store by a previous one.
1208 * Note that in unreachable loops it might happen that we reach
1209 * load again, as well as we can fall into a cycle.
1210 * We break such cycles using a special visited flag.
1213 res = follow_Mem_chain(load, skip_Proj(mem));
1215 } /* optimize_load */
1218 * Check whether a value of mode new_mode would completely overwrite a value
1219 * of mode old_mode in memory.
1221 static int is_completely_overwritten(ir_mode *old_mode, ir_mode *new_mode)
1223 return get_mode_size_bits(new_mode) >= get_mode_size_bits(old_mode);
1224 } /* is_completely_overwritten */
1227 * Check whether small is a part of large (starting at same address).
1229 static int is_partially_same(ir_node *small, ir_node *large)
1231 ir_mode *sm = get_irn_mode(small);
1232 ir_mode *lm = get_irn_mode(large);
1234 /* FIXME: Check endianness */
1235 return is_Conv(small) && get_Conv_op(small) == large
1236 && get_mode_size_bytes(sm) < get_mode_size_bytes(lm)
1237 && get_mode_arithmetic(sm) == irma_twos_complement
1238 && get_mode_arithmetic(lm) == irma_twos_complement;
1239 } /* is_partially_same */
1242 * follow the memory chain as long as there are only Loads and alias free Stores.
1244 * INC_MASTER() must be called before dive into
1246 static unsigned follow_Mem_chain_for_Store(ir_node *store, ir_node *curr)
1249 ldst_info_t *info = get_irn_link(store);
1251 ir_node *ptr = get_Store_ptr(store);
1252 ir_node *mem = get_Store_mem(store);
1253 ir_node *value = get_Store_value(store);
1254 ir_mode *mode = get_irn_mode(value);
1255 ir_node *block = get_nodes_block(store);
1256 ir_node *mblk = get_Block_MacroBlock(block);
1258 for (pred = curr; pred != store;) {
1259 ldst_info_t *pred_info = get_irn_link(pred);
1262 * BEWARE: one might think that checking the modes is useless, because
1263 * if the pointers are identical, they refer to the same object.
1264 * This is only true in strong typed languages, not is C were the following
1265 * is possible *(ir_type1 *)p = a; *(ir_type2 *)p = b ...
1266 * However, if the size of the mode that is written is bigger or equal the
1267 * size of the old one, the old value is completely overwritten and can be
1270 if (is_Store(pred) && get_Store_ptr(pred) == ptr &&
1271 get_nodes_MacroBlock(pred) == mblk) {
1273 * a Store after a Store in the same MacroBlock -- a write after write.
1277 * We may remove the first Store, if the old value is completely
1278 * overwritten or the old value is a part of the new value,
1279 * and if it does not have an exception handler.
1281 * TODO: What, if both have the same exception handler ???
1283 if (get_Store_volatility(pred) != volatility_is_volatile
1284 && !pred_info->projs[pn_Store_X_except]) {
1285 ir_node *predvalue = get_Store_value(pred);
1286 ir_mode *predmode = get_irn_mode(predvalue);
1288 if (is_completely_overwritten(predmode, mode)
1289 || is_partially_same(predvalue, value)) {
1290 DBG_OPT_WAW(pred, store);
1291 exchange(pred_info->projs[pn_Store_M], get_Store_mem(pred));
1293 reduce_adr_usage(ptr);
1299 * We may remove the Store, if the old value already contains
1300 * the new value, and if it does not have an exception handler.
1302 * TODO: What, if both have the same exception handler ???
1304 if (get_Store_volatility(store) != volatility_is_volatile
1305 && !info->projs[pn_Store_X_except]) {
1306 ir_node *predvalue = get_Store_value(pred);
1308 if (is_partially_same(value, predvalue)) {
1309 DBG_OPT_WAW(pred, store);
1310 exchange(info->projs[pn_Store_M], mem);
1312 reduce_adr_usage(ptr);
1316 } else if (is_Load(pred) && get_Load_ptr(pred) == ptr &&
1317 value == pred_info->projs[pn_Load_res]) {
1319 * a Store of a value just loaded from the same address
1320 * -- a write after read.
1321 * We may remove the Store, if it does not have an exception
1324 if (! info->projs[pn_Store_X_except]) {
1325 DBG_OPT_WAR(store, pred);
1326 exchange(info->projs[pn_Store_M], mem);
1328 reduce_adr_usage(ptr);
1333 if (is_Store(pred)) {
1334 /* check if we can pass through this store */
1335 ir_alias_relation rel = get_alias_relation(
1337 get_Store_ptr(pred),
1338 get_irn_mode(get_Store_value(pred)),
1340 /* if the might be an alias, we cannot pass this Store */
1341 if (rel != ir_no_alias)
1343 pred = skip_Proj(get_Store_mem(pred));
1344 } else if (is_Load(pred)) {
1345 ir_alias_relation rel = get_alias_relation(
1346 current_ir_graph, get_Load_ptr(pred), get_Load_mode(pred),
1348 if (rel != ir_no_alias)
1351 pred = skip_Proj(get_Load_mem(pred));
1353 /* follow only Load chains */
1357 /* check for cycles */
1358 if (NODE_VISITED(pred_info))
1360 MARK_NODE(pred_info);
1363 if (is_Sync(pred)) {
1366 /* handle all Sync predecessors */
1367 for (i = get_Sync_n_preds(pred) - 1; i >= 0; --i) {
1368 res |= follow_Mem_chain_for_Store(store, skip_Proj(get_Sync_pred(pred, i)));
1374 } /* follow_Mem_chain_for_Store */
1376 /** find entity used as base for an address calculation */
1377 static ir_entity *find_entity(ir_node *ptr)
1379 switch (get_irn_opcode(ptr)) {
1381 return get_SymConst_entity(ptr);
1383 ir_node *pred = get_Sel_ptr(ptr);
1384 if (get_irg_frame(get_irn_irg(ptr)) == pred)
1385 return get_Sel_entity(ptr);
1387 return find_entity(pred);
1391 ir_node *left = get_binop_left(ptr);
1393 if (mode_is_reference(get_irn_mode(left)))
1394 return find_entity(left);
1395 right = get_binop_right(ptr);
1396 if (mode_is_reference(get_irn_mode(right)))
1397 return find_entity(right);
1408 * @param store the Store node
1410 static unsigned optimize_store(ir_node *store)
1416 if (get_Store_volatility(store) == volatility_is_volatile)
1419 ptr = get_Store_ptr(store);
1420 entity = find_entity(ptr);
1422 /* a store to an entity which is never read is unnecessary */
1423 if (entity != NULL && !(get_entity_usage(entity) & ir_usage_read)) {
1424 ldst_info_t *info = get_irn_link(store);
1425 if (info->projs[pn_Store_X_except] == NULL) {
1426 DB((dbg, LEVEL_1, " Killing useless %+F to never read entity %+F\n", store, entity));
1427 exchange(info->projs[pn_Store_M], get_Store_mem(store));
1429 reduce_adr_usage(ptr);
1434 /* Check, if the address of this Store is used more than once.
1435 * If not, this Store cannot be removed in any case. */
1436 if (get_irn_n_uses(ptr) <= 1)
1439 mem = get_Store_mem(store);
1441 /* follow the memory chain as long as there are only Loads */
1444 return follow_Mem_chain_for_Store(store, skip_Proj(mem));
1445 } /* optimize_store */
1448 * walker, optimizes Phi after Stores to identical places:
1449 * Does the following optimization:
1452 * val1 val2 val3 val1 val2 val3
1454 * Store Store Store \ | /
1461 * This reduces the number of stores and allows for predicated execution.
1462 * Moves Stores back to the end of a function which may be bad.
1464 * This is only possible if the predecessor blocks have only one successor.
1466 static unsigned optimize_phi(ir_node *phi, walk_env_t *wenv)
1469 ir_node *store, *old_store, *ptr, *block, *phi_block, *phiM, *phiD, *exc, *projM;
1471 ir_node **inM, **inD, **projMs;
1473 dbg_info *db = NULL;
1475 block_info_t *bl_info;
1478 /* Must be a memory Phi */
1479 if (get_irn_mode(phi) != mode_M)
1482 n = get_Phi_n_preds(phi);
1486 /* must be only one user */
1487 projM = get_Phi_pred(phi, 0);
1488 if (get_irn_n_edges(projM) != 1)
1491 store = skip_Proj(projM);
1493 if (!is_Store(store))
1496 block = get_nodes_block(store);
1498 /* abort on dead blocks */
1499 if (is_Block_dead(block))
1502 /* check if the block is post dominated by Phi-block
1503 and has no exception exit */
1504 bl_info = get_irn_link(block);
1505 if (bl_info->flags & BLOCK_HAS_EXC)
1508 phi_block = get_nodes_block(phi);
1509 if (! block_strictly_postdominates(phi_block, block))
1512 /* this is the address of the store */
1513 ptr = get_Store_ptr(store);
1514 mode = get_irn_mode(get_Store_value(store));
1515 info = get_irn_link(store);
1516 exc = info->exc_block;
1518 for (i = 1; i < n; ++i) {
1519 ir_node *pred = get_Phi_pred(phi, i);
1521 if (get_irn_n_edges(pred) != 1)
1524 pred = skip_Proj(pred);
1525 if (!is_Store(pred))
1528 if (ptr != get_Store_ptr(pred) || mode != get_irn_mode(get_Store_value(pred)))
1531 info = get_irn_link(pred);
1533 /* check, if all stores have the same exception flow */
1534 if (exc != info->exc_block)
1537 /* abort on dead blocks */
1538 block = get_nodes_block(pred);
1539 if (is_Block_dead(block))
1542 /* check if the block is post dominated by Phi-block
1543 and has no exception exit. Note that block must be different from
1544 Phi-block, else we would move a Store from end End of a block to its
1546 bl_info = get_irn_link(block);
1547 if (bl_info->flags & BLOCK_HAS_EXC)
1549 if (block == phi_block || ! block_postdominates(phi_block, block))
1554 * ok, when we are here, we found all predecessors of a Phi that
1555 * are Stores to the same address and size. That means whatever
1556 * we do before we enter the block of the Phi, we do a Store.
1557 * So, we can move the Store to the current block:
1559 * val1 val2 val3 val1 val2 val3
1561 * | Str | | Str | | Str | \ | /
1567 * Is only allowed if the predecessor blocks have only one successor.
1570 NEW_ARR_A(ir_node *, projMs, n);
1571 NEW_ARR_A(ir_node *, inM, n);
1572 NEW_ARR_A(ir_node *, inD, n);
1573 NEW_ARR_A(int, idx, n);
1575 /* Prepare: Collect all Store nodes. We must do this
1576 first because we otherwise may loose a store when exchanging its
1579 for (i = n - 1; i >= 0; --i) {
1582 projMs[i] = get_Phi_pred(phi, i);
1583 assert(is_Proj(projMs[i]));
1585 store = get_Proj_pred(projMs[i]);
1586 info = get_irn_link(store);
1588 inM[i] = get_Store_mem(store);
1589 inD[i] = get_Store_value(store);
1590 idx[i] = info->exc_idx;
1592 block = get_nodes_block(phi);
1594 /* second step: create a new memory Phi */
1595 phiM = new_rd_Phi(get_irn_dbg_info(phi), block, n, inM, mode_M);
1597 /* third step: create a new data Phi */
1598 phiD = new_rd_Phi(get_irn_dbg_info(phi), block, n, inD, mode);
1600 /* rewire memory and kill the node */
1601 for (i = n - 1; i >= 0; --i) {
1602 ir_node *proj = projMs[i];
1604 if (is_Proj(proj)) {
1605 ir_node *store = get_Proj_pred(proj);
1606 exchange(proj, inM[i]);
1611 /* fourth step: create the Store */
1612 store = new_rd_Store(db, block, phiM, ptr, phiD, 0);
1614 co_set_irn_name(store, co_get_irn_ident(old_store));
1617 projM = new_rd_Proj(NULL, store, mode_M, pn_Store_M);
1619 info = get_ldst_info(store, &wenv->obst);
1620 info->projs[pn_Store_M] = projM;
1622 /* fifths step: repair exception flow */
1624 ir_node *projX = new_rd_Proj(NULL, store, mode_X, pn_Store_X_except);
1626 info->projs[pn_Store_X_except] = projX;
1627 info->exc_block = exc;
1628 info->exc_idx = idx[0];
1630 for (i = 0; i < n; ++i) {
1631 set_Block_cfgpred(exc, idx[i], projX);
1635 /* the exception block should be optimized as some inputs are identical now */
1641 /* sixth step: replace old Phi */
1642 exchange(phi, projM);
1644 return res | DF_CHANGED;
1645 } /* optimize_phi */
1648 * walker, do the optimizations
1650 static void do_load_store_optimize(ir_node *n, void *env)
1652 walk_env_t *wenv = env;
1654 switch (get_irn_opcode(n)) {
1657 wenv->changes |= optimize_load(n);
1661 wenv->changes |= optimize_store(n);
1665 wenv->changes |= optimize_phi(n, wenv);
1671 } /* do_load_store_optimize */
1674 typedef struct scc {
1675 ir_node *head; /**< the head of the list */
1678 /** A node entry. */
1679 typedef struct node_entry {
1680 unsigned DFSnum; /**< the DFS number of this node */
1681 unsigned low; /**< the low number of this node */
1682 int in_stack; /**< flag, set if the node is on the stack */
1683 ir_node *next; /**< link to the next node the the same scc */
1684 scc *pscc; /**< the scc of this node */
1685 unsigned POnum; /**< the post order number for blocks */
1688 /** A loop entry. */
1689 typedef struct loop_env {
1690 ir_phase ph; /**< the phase object */
1691 ir_node **stack; /**< the node stack */
1692 int tos; /**< tos index */
1693 unsigned nextDFSnum; /**< the current DFS number */
1694 unsigned POnum; /**< current post order number */
1696 unsigned changes; /**< a bitmask of graph changes */
1700 * Gets the node_entry of a node
1702 static node_entry *get_irn_ne(ir_node *irn, loop_env *env)
1704 ir_phase *ph = &env->ph;
1705 node_entry *e = phase_get_irn_data(&env->ph, irn);
1708 e = phase_alloc(ph, sizeof(*e));
1709 memset(e, 0, sizeof(*e));
1710 phase_set_irn_data(ph, irn, e);
1716 * Push a node onto the stack.
1718 * @param env the loop environment
1719 * @param n the node to push
1721 static void push(loop_env *env, ir_node *n)
1725 if (env->tos == ARR_LEN(env->stack)) {
1726 int nlen = ARR_LEN(env->stack) * 2;
1727 ARR_RESIZE(ir_node *, env->stack, nlen);
1729 env->stack[env->tos++] = n;
1730 e = get_irn_ne(n, env);
1735 * pop a node from the stack
1737 * @param env the loop environment
1739 * @return The topmost node
1741 static ir_node *pop(loop_env *env)
1743 ir_node *n = env->stack[--env->tos];
1744 node_entry *e = get_irn_ne(n, env);
1751 * Check if irn is a region constant.
1752 * The block or irn must strictly dominate the header block.
1754 * @param irn the node to check
1755 * @param header_block the header block of the induction variable
1757 static int is_rc(ir_node *irn, ir_node *header_block)
1759 ir_node *block = get_nodes_block(irn);
1761 return (block != header_block) && block_dominates(block, header_block);
1764 typedef struct phi_entry phi_entry;
1766 ir_node *phi; /**< A phi with a region const memory. */
1767 int pos; /**< The position of the region const memory */
1768 ir_node *load; /**< the newly created load for this phi */
1773 * An entry in the avail set.
1775 typedef struct avail_entry_t {
1776 ir_node *ptr; /**< the address pointer */
1777 ir_mode *mode; /**< the load mode */
1778 ir_node *load; /**< the associated Load */
1782 * Compare two avail entries.
1784 static int cmp_avail_entry(const void *elt, const void *key, size_t size)
1786 const avail_entry_t *a = elt;
1787 const avail_entry_t *b = key;
1790 return a->ptr != b->ptr || a->mode != b->mode;
1791 } /* cmp_avail_entry */
1794 * Calculate the hash value of an avail entry.
1796 static unsigned hash_cache_entry(const avail_entry_t *entry)
1798 return get_irn_idx(entry->ptr) * 9 + HASH_PTR(entry->mode);
1799 } /* hash_cache_entry */
1802 * Move loops out of loops if possible.
1804 * @param pscc the loop described by an SCC
1805 * @param env the loop environment
1807 static void move_loads_out_of_loops(scc *pscc, loop_env *env)
1809 ir_node *phi, *load, *next, *other, *next_other;
1812 phi_entry *phi_list = NULL;
1815 avail = new_set(cmp_avail_entry, 8);
1817 /* collect all outer memories */
1818 for (phi = pscc->head; phi != NULL; phi = next) {
1819 node_entry *ne = get_irn_ne(phi, env);
1822 /* check all memory Phi's */
1826 assert(get_irn_mode(phi) == mode_M && "DFS return non-memory Phi");
1828 for (j = get_irn_arity(phi) - 1; j >= 0; --j) {
1829 ir_node *pred = get_irn_n(phi, j);
1830 node_entry *pe = get_irn_ne(pred, env);
1832 if (pe->pscc != ne->pscc) {
1833 /* not in the same SCC, is region const */
1834 phi_entry *pe = phase_alloc(&env->ph, sizeof(*pe));
1838 pe->next = phi_list;
1843 /* no Phis no fun */
1844 assert(phi_list != NULL && "DFS found a loop without Phi");
1846 /* for now, we cannot handle more than one input (only reducible cf) */
1847 if (phi_list->next != NULL)
1850 for (load = pscc->head; load; load = next) {
1852 node_entry *ne = get_irn_ne(load, env);
1855 if (is_Load(load)) {
1856 ldst_info_t *info = get_irn_link(load);
1857 ir_node *ptr = get_Load_ptr(load);
1859 /* for now, we cannot handle Loads with exceptions */
1860 if (info->projs[pn_Load_res] == NULL || info->projs[pn_Load_X_regular] != NULL || info->projs[pn_Load_X_except] != NULL)
1863 /* for now, we can only move Load(Global) */
1864 if (! is_Global(ptr))
1866 ent = get_Global_entity(ptr);
1867 load_mode = get_Load_mode(load);
1868 for (other = pscc->head; other != NULL; other = next_other) {
1869 node_entry *ne = get_irn_ne(other, env);
1870 next_other = ne->next;
1872 if (is_Store(other)) {
1873 ir_alias_relation rel = get_alias_relation(
1875 get_Store_ptr(other),
1876 get_irn_mode(get_Store_value(other)),
1878 /* if the might be an alias, we cannot pass this Store */
1879 if (rel != ir_no_alias)
1882 /* only Phis and pure Calls are allowed here, so ignore them */
1884 if (other == NULL) {
1885 ldst_info_t *ninfo = NULL;
1889 /* yep, no aliasing Store found, Load can be moved */
1890 DB((dbg, LEVEL_1, " Found a Load that could be moved: %+F\n", load));
1892 db = get_irn_dbg_info(load);
1893 for (pe = phi_list; pe != NULL; pe = pe->next) {
1895 ir_node *phi = pe->phi;
1896 ir_node *blk = get_nodes_block(phi);
1897 ir_node *pred = get_Block_cfgpred_block(blk, pos);
1899 avail_entry_t entry, *res;
1902 entry.mode = load_mode;
1903 res = set_find(avail, &entry, sizeof(entry), hash_cache_entry(&entry));
1907 irn = new_rd_Load(db, pred, get_Phi_pred(phi, pos), ptr, load_mode, 0);
1909 set_insert(avail, &entry, sizeof(entry), hash_cache_entry(&entry));
1910 DB((dbg, LEVEL_1, " Created %+F in %+F\n", irn, pred));
1913 ninfo = get_ldst_info(irn, phase_obst(&env->ph));
1915 ninfo->projs[pn_Load_M] = mem = new_r_Proj(irn, mode_M, pn_Load_M);
1916 set_Phi_pred(phi, pos, mem);
1918 ninfo->projs[pn_Load_res] = new_r_Proj(irn, load_mode, pn_Load_res);
1921 /* now kill the old Load */
1922 exchange(info->projs[pn_Load_M], get_Load_mem(load));
1923 exchange(info->projs[pn_Load_res], ninfo->projs[pn_Load_res]);
1925 env->changes |= DF_CHANGED;
1930 } /* move_loads_out_of_loops */
1933 * Process a loop SCC.
1935 * @param pscc the SCC
1936 * @param env the loop environment
1938 static void process_loop(scc *pscc, loop_env *env)
1940 ir_node *irn, *next, *header = NULL;
1941 node_entry *b, *h = NULL;
1942 int j, only_phi, num_outside, process = 0;
1945 /* find the header block for this scc */
1946 for (irn = pscc->head; irn; irn = next) {
1947 node_entry *e = get_irn_ne(irn, env);
1948 ir_node *block = get_nodes_block(irn);
1951 b = get_irn_ne(block, env);
1953 if (header != NULL) {
1954 if (h->POnum < b->POnum) {
1964 /* check if this scc contains only Phi, Loads or Stores nodes */
1968 for (irn = pscc->head; irn; irn = next) {
1969 node_entry *e = get_irn_ne(irn, env);
1972 switch (get_irn_opcode(irn)) {
1974 if (is_Call_pure(irn)) {
1975 /* pure calls can be treated like loads */
1979 /* non-pure calls must be handle like may-alias Stores */
1982 /* cannot handle CopyB yet */
1986 if (get_Load_volatility(irn) == volatility_is_volatile) {
1987 /* cannot handle loops with volatile Loads */
1993 if (get_Store_volatility(irn) == volatility_is_volatile) {
1994 /* cannot handle loops with volatile Stores */
2003 for (j = get_irn_arity(irn) - 1; j >= 0; --j) {
2004 ir_node *pred = get_irn_n(irn, j);
2005 node_entry *pe = get_irn_ne(pred, env);
2007 if (pe->pscc != e->pscc) {
2008 /* not in the same SCC, must be a region const */
2009 if (! is_rc(pred, header)) {
2010 /* not a memory loop */
2013 if (out_rc == NULL) {
2014 /* first region constant */
2017 } else if (out_rc != pred) {
2018 /* another region constant */
2029 /* found a memory loop */
2030 DB((dbg, LEVEL_2, " Found a memory loop:\n "));
2031 if (only_phi && num_outside == 1) {
2032 /* a phi cycle with only one real predecessor can be collapsed */
2033 DB((dbg, LEVEL_2, " Found an USELESS Phi cycle:\n "));
2035 for (irn = pscc->head; irn; irn = next) {
2036 node_entry *e = get_irn_ne(irn, env);
2038 exchange(irn, out_rc);
2040 env->changes |= DF_CHANGED;
2044 #ifdef DEBUG_libfirm
2045 for (irn = pscc->head; irn; irn = next) {
2046 node_entry *e = get_irn_ne(irn, env);
2048 DB((dbg, LEVEL_2, " %+F,", irn));
2050 DB((dbg, LEVEL_2, "\n"));
2052 move_loads_out_of_loops(pscc, env);
2056 } /* process_loop */
2061 * @param pscc the SCC
2062 * @param env the loop environment
2064 static void process_scc(scc *pscc, loop_env *env)
2066 ir_node *head = pscc->head;
2067 node_entry *e = get_irn_ne(head, env);
2069 #ifdef DEBUG_libfirm
2071 ir_node *irn, *next;
2073 DB((dbg, LEVEL_4, " SCC at %p:\n ", pscc));
2074 for (irn = pscc->head; irn; irn = next) {
2075 node_entry *e = get_irn_ne(irn, env);
2079 DB((dbg, LEVEL_4, " %+F,", irn));
2081 DB((dbg, LEVEL_4, "\n"));
2085 if (e->next != NULL) {
2086 /* this SCC has more than one member */
2087 process_loop(pscc, env);
2092 * Do Tarjan's SCC algorithm and drive load/store optimization.
2094 * @param irn start at this node
2095 * @param env the loop environment
2097 static void dfs(ir_node *irn, loop_env *env)
2100 node_entry *node = get_irn_ne(irn, env);
2102 mark_irn_visited(irn);
2104 node->DFSnum = env->nextDFSnum++;
2105 node->low = node->DFSnum;
2109 if (is_Phi(irn) || is_Sync(irn)) {
2110 n = get_irn_arity(irn);
2111 for (i = 0; i < n; ++i) {
2112 ir_node *pred = get_irn_n(irn, i);
2113 node_entry *o = get_irn_ne(pred, env);
2115 if (!irn_visited(pred)) {
2117 node->low = MIN(node->low, o->low);
2119 if (o->DFSnum < node->DFSnum && o->in_stack)
2120 node->low = MIN(o->DFSnum, node->low);
2122 } else if (is_fragile_op(irn)) {
2123 ir_node *pred = get_fragile_op_mem(irn);
2124 node_entry *o = get_irn_ne(pred, env);
2126 if (!irn_visited(pred)) {
2128 node->low = MIN(node->low, o->low);
2130 if (o->DFSnum < node->DFSnum && o->in_stack)
2131 node->low = MIN(o->DFSnum, node->low);
2132 } else if (is_Proj(irn)) {
2133 ir_node *pred = get_Proj_pred(irn);
2134 node_entry *o = get_irn_ne(pred, env);
2136 if (!irn_visited(pred)) {
2138 node->low = MIN(node->low, o->low);
2140 if (o->DFSnum < node->DFSnum && o->in_stack)
2141 node->low = MIN(o->DFSnum, node->low);
2144 /* IGNORE predecessors */
2147 if (node->low == node->DFSnum) {
2148 scc *pscc = phase_alloc(&env->ph, sizeof(*pscc));
2156 e = get_irn_ne(x, env);
2158 e->next = pscc->head;
2162 process_scc(pscc, env);
2167 * Do the DFS on the memory edges a graph.
2169 * @param irg the graph to process
2170 * @param env the loop environment
2172 static void do_dfs(ir_graph *irg, loop_env *env)
2174 ir_graph *rem = current_ir_graph;
2175 ir_node *endblk, *end;
2178 current_ir_graph = irg;
2179 inc_irg_visited(irg);
2181 /* visit all memory nodes */
2182 endblk = get_irg_end_block(irg);
2183 for (i = get_Block_n_cfgpreds(endblk) - 1; i >= 0; --i) {
2184 ir_node *pred = get_Block_cfgpred(endblk, i);
2186 pred = skip_Proj(pred);
2187 if (is_Return(pred))
2188 dfs(get_Return_mem(pred), env);
2189 else if (is_Raise(pred))
2190 dfs(get_Raise_mem(pred), env);
2191 else if (is_fragile_op(pred))
2192 dfs(get_fragile_op_mem(pred), env);
2194 assert(0 && "Unknown EndBlock predecessor");
2198 /* visit the keep-alives */
2199 end = get_irg_end(irg);
2200 for (i = get_End_n_keepalives(end) - 1; i >= 0; --i) {
2201 ir_node *ka = get_End_keepalive(end, i);
2203 if (is_Phi(ka) && !irn_visited(ka))
2206 current_ir_graph = rem;
2210 * Optimize Loads/Stores in loops.
2212 * @param irg the graph
2214 static int optimize_loops(ir_graph *irg)
2218 env.stack = NEW_ARR_F(ir_node *, 128);
2223 phase_init(&env.ph, irg, phase_irn_init_default);
2225 /* calculate the SCC's and drive loop optimization. */
2228 DEL_ARR_F(env.stack);
2229 phase_deinit(&env.ph);
2232 } /* optimize_loops */
2235 * do the load store optimization
2237 int optimize_load_store(ir_graph *irg)
2241 FIRM_DBG_REGISTER(dbg, "firm.opt.ldstopt");
2243 assert(get_irg_phase_state(irg) != phase_building);
2244 assert(get_irg_pinned(irg) != op_pin_state_floats &&
2245 "LoadStore optimization needs pinned graph");
2247 /* we need landing pads */
2248 remove_critical_cf_edges(irg);
2252 /* for Phi optimization post-dominators are needed ... */
2253 assure_postdoms(irg);
2255 if (get_opt_alias_analysis()) {
2256 assure_irg_entity_usage_computed(irg);
2257 assure_irp_globals_entity_usage_computed();
2260 obstack_init(&env.obst);
2263 /* init the links, then collect Loads/Stores/Proj's in lists */
2265 irg_walk_graph(irg, firm_clear_link, collect_nodes, &env);
2267 /* now we have collected enough information, optimize */
2268 irg_walk_graph(irg, NULL, do_load_store_optimize, &env);
2270 env.changes |= optimize_loops(irg);
2272 obstack_free(&env.obst, NULL);
2274 /* Handle graph state */
2276 set_irg_outs_inconsistent(irg);
2277 set_irg_entity_usage_state(irg, ir_entity_usage_not_computed);
2280 if (env.changes & CF_CHANGED) {
2281 /* is this really needed: Yes, control flow changed, block might
2282 have Bad() predecessors. */
2283 set_irg_doms_inconsistent(irg);
2285 return env.changes != 0;
2286 } /* optimize_load_store */
2288 ir_graph_pass_t *optimize_load_store_pass(const char *name)
2290 return def_graph_pass_ret(name ? name : "ldst", optimize_load_store);
2291 } /* optimize_load_store_pass */