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
32 #include "iroptimize.h"
34 #include "irgraph_t.h"
42 #include "dbginfo_t.h"
43 #include "iropt_dbg.h"
49 #include "opt_polymorphy.h"
52 #include "irphase_t.h"
56 /** The debug handle. */
57 DEBUG_ONLY(static firm_dbg_module_t *dbg;)
60 #include "cacheopt/cachesim.h"
64 #define IMAX(a,b) ((a) > (b) ? (a) : (b))
66 #define MAX_PROJ IMAX(IMAX(pn_Load_max, pn_Store_max), pn_Call_max)
69 DF_CHANGED = 1, /**< data flow changed */
70 CF_CHANGED = 2, /**< control flow changed */
76 typedef struct _walk_env_t {
77 struct obstack obst; /**< list of all stores */
78 unsigned changes; /**< a bitmask of graph changes */
81 /** A Load/Store info. */
82 typedef struct _ldst_info_t {
83 ir_node *projs[MAX_PROJ]; /**< list of Proj's of this node */
84 ir_node *exc_block; /**< the exception block if available */
85 int exc_idx; /**< predecessor index in the exception block */
86 unsigned visited; /**< visited counter for breaking loops */
90 * flags for control flow.
93 BLOCK_HAS_COND = 1, /**< Block has conditional control flow */
94 BLOCK_HAS_EXC = 2 /**< Block has exceptional control flow */
100 typedef struct _block_info_t {
101 unsigned flags; /**< flags for the block */
104 /** the master visited flag for loop detection. */
105 static unsigned master_visited = 0;
107 #define INC_MASTER() ++master_visited
108 #define MARK_NODE(info) (info)->visited = master_visited
109 #define NODE_VISITED(info) (info)->visited >= master_visited
112 * get the Load/Store info of a node
114 static ldst_info_t *get_ldst_info(ir_node *node, struct obstack *obst) {
115 ldst_info_t *info = get_irn_link(node);
118 info = obstack_alloc(obst, sizeof(*info));
119 memset(info, 0, sizeof(*info));
120 set_irn_link(node, info);
123 } /* get_ldst_info */
126 * get the Block info of a node
128 static block_info_t *get_block_info(ir_node *node, struct obstack *obst) {
129 block_info_t *info = get_irn_link(node);
132 info = obstack_alloc(obst, sizeof(*info));
133 memset(info, 0, sizeof(*info));
134 set_irn_link(node, info);
137 } /* get_block_info */
140 * update the projection info for a Load/Store
142 static unsigned update_projs(ldst_info_t *info, ir_node *proj)
144 long nr = get_Proj_proj(proj);
146 assert(0 <= nr && nr <= MAX_PROJ && "Wrong proj from LoadStore");
148 if (info->projs[nr]) {
149 /* there is already one, do CSE */
150 exchange(proj, info->projs[nr]);
154 info->projs[nr] = proj;
160 * update the exception block info for a Load/Store node.
162 * @param info the load/store info struct
163 * @param block the exception handler block for this load/store
164 * @param pos the control flow input of the block
166 static unsigned update_exc(ldst_info_t *info, ir_node *block, int pos)
168 assert(info->exc_block == NULL && "more than one exception block found");
170 info->exc_block = block;
175 /** Return the number of uses of an address node */
176 #define get_irn_n_uses(adr) get_irn_n_edges(adr)
179 * walker, collects all Load/Store/Proj nodes
181 * walks from Start -> End
183 static void collect_nodes(ir_node *node, void *env)
185 ir_opcode opcode = get_irn_opcode(node);
186 ir_node *pred, *blk, *pred_blk;
187 ldst_info_t *ldst_info;
188 walk_env_t *wenv = env;
190 if (opcode == iro_Proj) {
191 pred = get_Proj_pred(node);
192 opcode = get_irn_opcode(pred);
194 if (opcode == iro_Load || opcode == iro_Store || opcode == iro_Call) {
195 ldst_info = get_ldst_info(pred, &wenv->obst);
197 wenv->changes |= update_projs(ldst_info, node);
200 * Place the Proj's to the same block as the
201 * predecessor Load. This is always ok and prevents
202 * "non-SSA" form after optimizations if the Proj
203 * is in a wrong block.
205 blk = get_nodes_block(node);
206 pred_blk = get_nodes_block(pred);
207 if (blk != pred_blk) {
208 wenv->changes |= DF_CHANGED;
209 set_nodes_block(node, pred_blk);
212 } else if (opcode == iro_Block) {
215 for (i = get_Block_n_cfgpreds(node) - 1; i >= 0; --i) {
216 ir_node *pred_block, *proj;
217 block_info_t *bl_info;
220 pred = proj = get_Block_cfgpred(node, i);
223 pred = get_Proj_pred(proj);
224 is_exc = get_Proj_proj(proj) == pn_Generic_X_except;
227 /* ignore Bad predecessors, they will be removed later */
231 pred_block = get_nodes_block(pred);
232 bl_info = get_block_info(pred_block, &wenv->obst);
234 if (is_fragile_op(pred) && is_exc)
235 bl_info->flags |= BLOCK_HAS_EXC;
236 else if (is_irn_forking(pred))
237 bl_info->flags |= BLOCK_HAS_COND;
239 opcode = get_irn_opcode(pred);
240 if (is_exc && (opcode == iro_Load || opcode == iro_Store || opcode == iro_Call)) {
241 ldst_info = get_ldst_info(pred, &wenv->obst);
243 wenv->changes |= update_exc(ldst_info, node, i);
247 } /* collect_nodes */
250 * Returns an entity if the address ptr points to a constant one.
252 * @param ptr the address
254 * @return an entity or NULL
256 static ir_entity *find_constant_entity(ir_node *ptr)
259 if (is_SymConst(ptr) && get_SymConst_kind(ptr) == symconst_addr_ent) {
260 ir_entity *ent = get_SymConst_entity(ptr);
261 if (variability_constant == get_entity_variability(ent))
264 } else if (is_Sel(ptr)) {
265 ir_entity *ent = get_Sel_entity(ptr);
266 ir_type *tp = get_entity_owner(ent);
268 /* Do not fiddle with polymorphism. */
269 if (is_Class_type(get_entity_owner(ent)) &&
270 ((get_entity_n_overwrites(ent) != 0) ||
271 (get_entity_n_overwrittenby(ent) != 0) ) )
274 if (is_Array_type(tp)) {
278 for (i = 0, n = get_Sel_n_indexs(ptr); i < n; ++i) {
280 tarval *tlower, *tupper;
281 ir_node *index = get_Sel_index(ptr, i);
282 tarval *tv = computed_value(index);
284 /* check if the index is constant */
285 if (tv == tarval_bad)
288 bound = get_array_lower_bound(tp, i);
289 tlower = computed_value(bound);
290 bound = get_array_upper_bound(tp, i);
291 tupper = computed_value(bound);
293 if (tlower == tarval_bad || tupper == tarval_bad)
296 if (tarval_cmp(tv, tlower) & pn_Cmp_Lt)
298 if (tarval_cmp(tupper, tv) & pn_Cmp_Lt)
301 /* ok, bounds check finished */
305 if (variability_constant == get_entity_variability(ent))
309 ptr = get_Sel_ptr(ptr);
310 } else if (is_Add(ptr)) {
311 ir_node *l = get_Add_left(ptr);
312 ir_node *r = get_Add_right(ptr);
314 if (get_irn_mode(l) == get_irn_mode(ptr) && is_Const(r))
316 else if (get_irn_mode(r) == get_irn_mode(ptr) && is_Const(l))
321 /* for now, we support only one addition, reassoc should fold all others */
322 if (! is_SymConst(ptr) && !is_Sel(ptr))
324 } else if (is_Sub(ptr)) {
325 ir_node *l = get_Sub_left(ptr);
326 ir_node *r = get_Sub_right(ptr);
328 if (get_irn_mode(l) == get_irn_mode(ptr) && is_Const(r))
332 /* for now, we support only one substraction, reassoc should fold all others */
333 if (! is_SymConst(ptr) && !is_Sel(ptr))
338 } /* find_constant_entity */
341 * Return the Selection index of a Sel node from dimension n
343 static long get_Sel_array_index_long(ir_node *n, int dim) {
344 ir_node *index = get_Sel_index(n, dim);
345 assert(is_Const(index));
346 return get_tarval_long(get_Const_tarval(index));
347 } /* get_Sel_array_index_long */
350 * Returns the accessed component graph path for an
351 * node computing an address.
353 * @param ptr the node computing the address
354 * @param depth current depth in steps upward from the root
357 static compound_graph_path *rec_get_accessed_path(ir_node *ptr, int depth) {
358 compound_graph_path *res = NULL;
359 ir_entity *root, *field, *ent;
360 int path_len, pos, idx;
364 if (is_SymConst(ptr)) {
365 /* a SymConst. If the depth is 0, this is an access to a global
366 * entity and we don't need a component path, else we know
367 * at least it's length.
369 assert(get_SymConst_kind(ptr) == symconst_addr_ent);
370 root = get_SymConst_entity(ptr);
371 res = (depth == 0) ? NULL : new_compound_graph_path(get_entity_type(root), depth);
372 } else if (is_Sel(ptr)) {
373 /* it's a Sel, go up until we find the root */
374 res = rec_get_accessed_path(get_Sel_ptr(ptr), depth+1);
378 /* fill up the step in the path at the current position */
379 field = get_Sel_entity(ptr);
380 path_len = get_compound_graph_path_length(res);
381 pos = path_len - depth - 1;
382 set_compound_graph_path_node(res, pos, field);
384 if (is_Array_type(get_entity_owner(field))) {
385 assert(get_Sel_n_indexs(ptr) == 1 && "multi dim arrays not implemented");
386 set_compound_graph_path_array_index(res, pos, get_Sel_array_index_long(ptr, 0));
388 } else if (is_Add(ptr)) {
389 ir_node *l = get_Add_left(ptr);
390 ir_node *r = get_Add_right(ptr);
395 tv = get_Const_tarval(r);
398 tv = get_Const_tarval(l);
401 mode = get_tarval_mode(tv);
403 /* ptr must be a Sel or a SymConst, this was checked in find_constant_entity() */
405 field = get_Sel_entity(ptr);
407 field = get_SymConst_entity(ptr);
410 for (ent = field;;) {
411 tp = get_entity_type(ent);
412 if (! is_Array_type(tp))
414 ent = get_array_element_entity(tp);
417 /* should be at least ONE array */
421 res = rec_get_accessed_path(ptr, depth + idx);
425 path_len = get_compound_graph_path_length(res);
426 pos = path_len - depth - idx;
428 for (ent = field;;) {
430 tarval *sz, *tv_index, *tlower, *tupper;
434 tp = get_entity_type(ent);
435 if (! is_Array_type(tp))
437 ent = get_array_element_entity(tp);
438 set_compound_graph_path_node(res, pos, ent);
440 size = get_type_size_bytes(get_entity_type(ent));
441 sz = new_tarval_from_long(size, mode);
443 tv_index = tarval_div(tv, sz);
444 tv = tarval_mod(tv, sz);
446 if (tv_index == tarval_bad || tv == tarval_bad)
449 assert(get_array_n_dimensions(tp) == 1 && "multiarrays not implemented");
450 bound = get_array_lower_bound(tp, 0);
451 tlower = computed_value(bound);
452 bound = get_array_upper_bound(tp, 0);
453 tupper = computed_value(bound);
455 if (tlower == tarval_bad || tupper == tarval_bad)
458 if (tarval_cmp(tv_index, tlower) & pn_Cmp_Lt)
460 if (tarval_cmp(tupper, tv_index) & pn_Cmp_Lt)
463 /* ok, bounds check finished */
464 index = get_tarval_long(tv_index);
465 set_compound_graph_path_array_index(res, pos, index);
468 if (! tarval_is_null(tv)) {
469 /* hmm, wrong access */
472 } else if (is_Sub(ptr)) {
473 ir_node *l = get_Sub_left(ptr);
474 ir_node *r = get_Sub_right(ptr);
477 tv = get_Const_tarval(r);
482 } /* rec_get_accessed_path */
485 * Returns an access path or NULL. The access path is only
486 * valid, if the graph is in phase_high and _no_ address computation is used.
488 static compound_graph_path *get_accessed_path(ir_node *ptr) {
489 return rec_get_accessed_path(ptr, 0);
490 } /* get_accessed_path */
492 typedef struct path_entry {
494 struct path_entry *next;
498 static ir_node *rec_find_compound_ent_value(ir_node *ptr, path_entry *next) {
499 path_entry entry, *p;
500 ir_entity *ent, *field;
501 ir_initializer_t *initializer;
507 if (is_SymConst(ptr)) {
509 ent = get_SymConst_entity(ptr);
510 initializer = get_entity_initializer(ent);
511 for (p = next; p != NULL; p = p->next) {
514 if (initializer->kind != IR_INITIALIZER_COMPOUND)
517 n = get_initializer_compound_n_entries(initializer);
520 initializer = get_initializer_compound_value(initializer, p->index);
523 switch (initializer->kind) {
524 case IR_INITIALIZER_CONST:
525 return get_initializer_const_value(initializer);
526 case IR_INITIALIZER_TARVAL:
527 case IR_INITIALIZER_NULL:
531 } else if (is_Sel(ptr)) {
532 entry.ent = field = get_Sel_entity(ptr);
533 tp = get_entity_owner(field);
534 if (is_Array_type(tp)) {
535 assert(get_Sel_n_indexs(ptr) == 1 && "multi dim arrays not implemented");
536 entry.index = get_Sel_array_index_long(ptr, 0) - get_array_lower_bound_int(tp, 0);
538 int i, n_members = get_compound_n_members(tp);
539 for (i = 0; i < n_members; ++i) {
540 if (get_compound_member(tp, i) == field)
543 if (i >= n_members) {
544 /* not found: should NOT happen */
549 return rec_find_compound_ent_value(get_Sel_ptr(ptr), &entry);
550 } else if (is_Add(ptr)) {
551 ir_node *l = get_Add_left(ptr);
552 ir_node *r = get_Add_right(ptr);
558 tv = get_Const_tarval(r);
561 tv = get_Const_tarval(l);
564 mode = get_tarval_mode(tv);
566 /* ptr must be a Sel or a SymConst, this was checked in find_constant_entity() */
568 field = get_Sel_entity(ptr);
570 field = get_SymConst_entity(ptr);
573 /* count needed entries */
575 for (ent = field;;) {
576 tp = get_entity_type(ent);
577 if (! is_Array_type(tp))
579 ent = get_array_element_entity(tp);
582 /* should be at least ONE entry */
586 /* allocate the right number of entries */
587 NEW_ARR_A(path_entry, p, pos);
591 for (ent = field;;) {
593 tarval *sz, *tv_index, *tlower, *tupper;
597 tp = get_entity_type(ent);
598 if (! is_Array_type(tp))
600 ent = get_array_element_entity(tp);
602 p[pos].next = &p[pos + 1];
604 size = get_type_size_bytes(get_entity_type(ent));
605 sz = new_tarval_from_long(size, mode);
607 tv_index = tarval_div(tv, sz);
608 tv = tarval_mod(tv, sz);
610 if (tv_index == tarval_bad || tv == tarval_bad)
613 assert(get_array_n_dimensions(tp) == 1 && "multiarrays not implemented");
614 bound = get_array_lower_bound(tp, 0);
615 tlower = computed_value(bound);
616 bound = get_array_upper_bound(tp, 0);
617 tupper = computed_value(bound);
619 if (tlower == tarval_bad || tupper == tarval_bad)
622 if (tarval_cmp(tv_index, tlower) & pn_Cmp_Lt)
624 if (tarval_cmp(tupper, tv_index) & pn_Cmp_Lt)
627 /* ok, bounds check finished */
628 index = get_tarval_long(tv_index);
629 p[pos].index = index;
632 if (! tarval_is_null(tv)) {
633 /* hmm, wrong access */
636 p[pos - 1].next = next;
637 return rec_find_compound_ent_value(ptr, p);
638 } else if (is_Sub(ptr)) {
639 ir_node *l = get_Sub_left(ptr);
640 ir_node *r = get_Sub_right(ptr);
643 tv = get_Const_tarval(r);
650 static ir_node *find_compound_ent_value(ir_node *ptr) {
651 return rec_find_compound_ent_value(ptr, NULL);
655 static void reduce_adr_usage(ir_node *ptr);
658 * Update a Load that may lost it's usage.
660 static void handle_load_update(ir_node *load) {
661 ldst_info_t *info = get_irn_link(load);
663 /* do NOT touch volatile loads for now */
664 if (get_Load_volatility(load) == volatility_is_volatile)
667 if (! info->projs[pn_Load_res] && ! info->projs[pn_Load_X_except]) {
668 ir_node *ptr = get_Load_ptr(load);
669 ir_node *mem = get_Load_mem(load);
671 /* a Load which value is neither used nor exception checked, remove it */
672 exchange(info->projs[pn_Load_M], mem);
673 if (info->projs[pn_Load_X_regular])
674 exchange(info->projs[pn_Load_X_regular], new_r_Jmp(current_ir_graph, get_nodes_block(load)));
676 reduce_adr_usage(ptr);
678 } /* handle_load_update */
681 * A Use of an address node is vanished. Check if this was a Proj
682 * node and update the counters.
684 static void reduce_adr_usage(ir_node *ptr) {
686 if (get_irn_n_edges(ptr) <= 0) {
687 /* this Proj is dead now */
688 ir_node *pred = get_Proj_pred(ptr);
691 ldst_info_t *info = get_irn_link(pred);
692 info->projs[get_Proj_proj(ptr)] = NULL;
694 /* this node lost it's result proj, handle that */
695 handle_load_update(pred);
699 } /* reduce_adr_usage */
702 * Check, if an already existing value of mode old_mode can be converted
703 * into the needed one new_mode without loss.
705 static int can_use_stored_value(ir_mode *old_mode, ir_mode *new_mode) {
706 if (old_mode == new_mode)
709 /* if both modes are two-complement ones, we can always convert the
710 Stored value into the needed one. */
711 if (get_mode_size_bits(old_mode) >= get_mode_size_bits(new_mode) &&
712 get_mode_arithmetic(old_mode) == irma_twos_complement &&
713 get_mode_arithmetic(new_mode) == irma_twos_complement)
716 } /* can_use_stored_value */
719 * Check whether a Call is at least pure, ie. does only read memory.
721 static unsigned is_Call_pure(ir_node *call) {
722 ir_type *call_tp = get_Call_type(call);
723 unsigned prop = get_method_additional_properties(call_tp);
725 /* check first the call type */
726 if ((prop & (mtp_property_const|mtp_property_pure)) == 0) {
727 /* try the called entity */
728 ir_node *ptr = get_Call_ptr(call);
730 if (is_Global(ptr)) {
731 ir_entity *ent = get_Global_entity(ptr);
733 prop = get_entity_additional_properties(ent);
736 return (prop & (mtp_property_const|mtp_property_pure)) != 0;
740 * Follow the memory chain as long as there are only Loads,
741 * alias free Stores, and constant Calls and try to replace the
742 * current Load by a previous ones.
743 * Note that in unreachable loops it might happen that we reach
744 * load again, as well as we can fall into a cycle.
745 * We break such cycles using a special visited flag.
747 * INC_MASTER() must be called before dive into
749 static unsigned follow_Mem_chain(ir_node *load, ir_node *curr) {
751 ldst_info_t *info = get_irn_link(load);
753 ir_node *ptr = get_Load_ptr(load);
754 ir_node *mem = get_Load_mem(load);
755 ir_mode *load_mode = get_Load_mode(load);
757 for (pred = curr; load != pred; ) {
758 ldst_info_t *pred_info = get_irn_link(pred);
761 * BEWARE: one might think that checking the modes is useless, because
762 * if the pointers are identical, they refer to the same object.
763 * This is only true in strong typed languages, not in C were the following
764 * is possible a = *(ir_type1 *)p; b = *(ir_type2 *)p ...
766 if (is_Store(pred) && get_Store_ptr(pred) == ptr &&
767 can_use_stored_value(get_irn_mode(get_Store_value(pred)), load_mode)) {
769 * a Load immediately after a Store -- a read after write.
770 * We may remove the Load, if both Load & Store does not have an exception handler
771 * OR they are in the same MacroBlock. In the latter case the Load cannot
772 * throw an exception when the previous Store was quiet.
774 * Why we need to check for Store Exception? If the Store cannot
775 * be executed (ROM) the exception handler might simply jump into
776 * the load MacroBlock :-(
777 * We could make it a little bit better if we would know that the exception
778 * handler of the Store jumps directly to the end...
780 if ((pred_info->projs[pn_Store_X_except] == NULL && info->projs[pn_Load_X_except] == NULL) ||
781 get_nodes_MacroBlock(load) == get_nodes_MacroBlock(pred)) {
782 ir_node *value = get_Store_value(pred);
784 DBG_OPT_RAW(load, value);
786 /* add an convert if needed */
787 if (get_irn_mode(get_Store_value(pred)) != load_mode) {
788 value = new_r_Conv(current_ir_graph, get_nodes_block(load), value, load_mode);
791 if (info->projs[pn_Load_M])
792 exchange(info->projs[pn_Load_M], mem);
795 if (info->projs[pn_Load_X_except]) {
796 exchange( info->projs[pn_Load_X_except], new_Bad());
799 if (info->projs[pn_Load_X_regular]) {
800 exchange( info->projs[pn_Load_X_regular], new_r_Jmp(current_ir_graph, get_nodes_block(load)));
804 if (info->projs[pn_Load_res])
805 exchange(info->projs[pn_Load_res], value);
808 reduce_adr_usage(ptr);
809 return res | DF_CHANGED;
811 } else if (is_Load(pred) && get_Load_ptr(pred) == ptr &&
812 can_use_stored_value(get_Load_mode(pred), load_mode)) {
814 * a Load after a Load -- a read after read.
815 * We may remove the second Load, if it does not have an exception handler
816 * OR they are in the same MacroBlock. In the later case the Load cannot
817 * throw an exception when the previous Load was quiet.
819 * Here, there is no need to check if the previous Load has an exception
820 * hander because they would have exact the same exception...
822 if (info->projs[pn_Load_X_except] == NULL || get_nodes_MacroBlock(load) == get_nodes_MacroBlock(pred)) {
825 DBG_OPT_RAR(load, pred);
827 /* the result is used */
828 if (info->projs[pn_Load_res]) {
829 if (pred_info->projs[pn_Load_res] == NULL) {
830 /* create a new Proj again */
831 pred_info->projs[pn_Load_res] = new_r_Proj(current_ir_graph, get_nodes_block(pred), pred, get_Load_mode(pred), pn_Load_res);
833 value = pred_info->projs[pn_Load_res];
835 /* add an convert if needed */
836 if (get_Load_mode(pred) != load_mode) {
837 value = new_r_Conv(current_ir_graph, get_nodes_block(load), value, load_mode);
840 exchange(info->projs[pn_Load_res], value);
843 if (info->projs[pn_Load_M])
844 exchange(info->projs[pn_Load_M], mem);
847 if (info->projs[pn_Load_X_except]) {
848 exchange(info->projs[pn_Load_X_except], new_Bad());
851 if (info->projs[pn_Load_X_regular]) {
852 exchange( info->projs[pn_Load_X_regular], new_r_Jmp(current_ir_graph, get_nodes_block(load)));
857 reduce_adr_usage(ptr);
858 return res |= DF_CHANGED;
862 if (is_Store(pred)) {
863 /* check if we can pass through this store */
864 ir_alias_relation rel = get_alias_relation(
867 get_irn_mode(get_Store_value(pred)),
869 /* if the might be an alias, we cannot pass this Store */
870 if (rel != ir_no_alias)
872 pred = skip_Proj(get_Store_mem(pred));
873 } else if (is_Load(pred)) {
874 pred = skip_Proj(get_Load_mem(pred));
875 } else if (is_Call(pred)) {
876 if (is_Call_pure(pred)) {
877 /* The called graph is at least pure, so there are no Store's
878 in it. We can handle it like a Load and skip it. */
879 pred = skip_Proj(get_Call_mem(pred));
881 /* there might be Store's in the graph, stop here */
885 /* follow only Load chains */
889 /* check for cycles */
890 if (NODE_VISITED(pred_info))
892 MARK_NODE(pred_info);
898 /* handle all Sync predecessors */
899 for (i = get_Sync_n_preds(pred) - 1; i >= 0; --i) {
900 res |= follow_Mem_chain(load, skip_Proj(get_Sync_pred(pred, i)));
907 } /* follow_Mem_chain */
912 * @param load the Load node
914 static unsigned optimize_load(ir_node *load)
916 ldst_info_t *info = get_irn_link(load);
917 ir_node *mem, *ptr, *new_node;
921 /* do NOT touch volatile loads for now */
922 if (get_Load_volatility(load) == volatility_is_volatile)
925 /* the address of the load to be optimized */
926 ptr = get_Load_ptr(load);
929 * Check if we can remove the exception from a Load:
930 * This can be done, if the address is from an Sel(Alloc) and
931 * the Sel type is a subtype of the allocated type.
933 * This optimizes some often used OO constructs,
934 * like x = new O; x->t;
936 if (info->projs[pn_Load_X_except]) {
938 ir_node *mem = get_Sel_mem(ptr);
940 /* FIXME: works with the current FE, but better use the base */
941 if (is_Alloc(skip_Proj(mem))) {
942 /* ok, check the types */
943 ir_entity *ent = get_Sel_entity(ptr);
944 ir_type *s_type = get_entity_type(ent);
945 ir_type *a_type = get_Alloc_type(mem);
947 if (is_SubClass_of(s_type, a_type)) {
948 /* ok, condition met: there can't be an exception because
949 * Alloc guarantees that enough memory was allocated */
951 exchange(info->projs[pn_Load_X_except], new_Bad());
952 info->projs[pn_Load_X_except] = NULL;
953 exchange(info->projs[pn_Load_X_regular], new_r_Jmp(current_ir_graph, get_nodes_block(load)));
954 info->projs[pn_Load_X_regular] = NULL;
958 } else if (is_Alloc(skip_Proj(skip_Cast(ptr)))) {
959 /* simple case: a direct load after an Alloc. Firm Alloc throw
960 * an exception in case of out-of-memory. So, there is no way for an
961 * exception in this load.
962 * This code is constructed by the "exception lowering" in the Jack compiler.
964 exchange(info->projs[pn_Load_X_except], new_Bad());
965 info->projs[pn_Load_X_except] = NULL;
966 exchange(info->projs[pn_Load_X_regular], new_r_Jmp(current_ir_graph, get_nodes_block(load)));
967 info->projs[pn_Load_X_regular] = NULL;
972 /* The mem of the Load. Must still be returned after optimization. */
973 mem = get_Load_mem(load);
975 if (! info->projs[pn_Load_res] && ! info->projs[pn_Load_X_except]) {
976 /* a Load which value is neither used nor exception checked, remove it */
977 exchange(info->projs[pn_Load_M], mem);
979 if (info->projs[pn_Load_X_regular]) {
980 /* should not happen, but if it does, remove it */
981 exchange(info->projs[pn_Load_X_regular], new_r_Jmp(current_ir_graph, get_nodes_block(load)));
985 reduce_adr_usage(ptr);
986 return res | DF_CHANGED;
989 /* Load from a constant polymorphic field, where we can resolve
991 new_node = transform_node_Load(load);
992 if (new_node != load) {
993 if (info->projs[pn_Load_M]) {
994 exchange(info->projs[pn_Load_M], mem);
995 info->projs[pn_Load_M] = NULL;
997 if (info->projs[pn_Load_X_except]) {
998 exchange(info->projs[pn_Load_X_except], new_Bad());
999 info->projs[pn_Load_X_except] = NULL;
1002 if (info->projs[pn_Load_X_regular]) {
1003 exchange(info->projs[pn_Load_X_regular], new_r_Jmp(current_ir_graph, get_nodes_block(load)));
1004 info->projs[pn_Load_X_regular] = NULL;
1007 if (info->projs[pn_Load_res])
1008 exchange(info->projs[pn_Load_res], new_node);
1011 reduce_adr_usage(ptr);
1012 return res | DF_CHANGED;
1015 /* check if we can determine the entity that will be loaded */
1016 ent = find_constant_entity(ptr);
1018 if ((allocation_static == get_entity_allocation(ent)) &&
1019 (visibility_external_allocated != get_entity_visibility(ent))) {
1020 /* a static allocation that is not external: there should be NO exception
1023 /* no exception, clear the info field as it might be checked later again */
1024 if (info->projs[pn_Load_X_except]) {
1025 exchange(info->projs[pn_Load_X_except], new_Bad());
1026 info->projs[pn_Load_X_except] = NULL;
1029 if (info->projs[pn_Load_X_regular]) {
1030 exchange(info->projs[pn_Load_X_regular], new_r_Jmp(current_ir_graph, get_nodes_block(load)));
1031 info->projs[pn_Load_X_regular] = NULL;
1035 if (variability_constant == get_entity_variability(ent)) {
1036 if (is_atomic_entity(ent)) {
1037 /* Might not be atomic after
1038 lowering of Sels. In this
1039 case we could also load, but
1040 it's more complicated. */
1041 /* more simpler case: we load the content of a constant value:
1042 * replace it by the constant itself
1046 if (info->projs[pn_Load_M]) {
1047 exchange(info->projs[pn_Load_M], mem);
1051 if (info->projs[pn_Load_res]) {
1052 if (is_atomic_entity(ent)) {
1053 ir_node *c = copy_const_value(get_irn_dbg_info(load), get_atomic_ent_value(ent));
1055 DBG_OPT_RC(load, c);
1056 exchange(info->projs[pn_Load_res], c);
1061 reduce_adr_usage(ptr);
1065 if (ent->has_initializer) {
1066 /* new style initializer */
1067 c = find_compound_ent_value(ptr);
1069 /* old style initializer */
1070 compound_graph_path *path = get_accessed_path(ptr);
1073 assert(is_proper_compound_graph_path(path, get_compound_graph_path_length(path)-1));
1075 c = get_compound_ent_value_by_path(ent, path);
1076 free_compound_graph_path(path);
1080 if (info->projs[pn_Load_M]) {
1081 exchange(info->projs[pn_Load_M], mem);
1084 if (info->projs[pn_Load_res]) {
1085 exchange(info->projs[pn_Load_res], copy_const_value(get_irn_dbg_info(load), c));
1089 reduce_adr_usage(ptr);
1092 /* We can not determine a correct access path. E.g., in jack, we load
1093 a byte from an object to generate an exception. Happens in test program
1095 printf(">>>>>>>>>>>>> Found access to constant entity %s in function %s\n", get_entity_name(ent),
1096 get_entity_name(get_irg_entity(current_ir_graph)));
1097 ir_printf(" load: %+F\n", load);
1098 ir_printf(" ptr: %+F\n", ptr);
1106 /* Check, if the address of this load is used more than once.
1107 * If not, this load cannot be removed in any case. */
1108 if (get_irn_n_uses(ptr) <= 1)
1112 * follow the memory chain as long as there are only Loads
1113 * and try to replace current Load or Store by a previous one.
1114 * Note that in unreachable loops it might happen that we reach
1115 * load again, as well as we can fall into a cycle.
1116 * We break such cycles using a special visited flag.
1119 res = follow_Mem_chain(load, skip_Proj(mem));
1121 } /* optimize_load */
1124 * Check whether a value of mode new_mode would completely overwrite a value
1125 * of mode old_mode in memory.
1127 static int is_completely_overwritten(ir_mode *old_mode, ir_mode *new_mode)
1129 return get_mode_size_bits(new_mode) >= get_mode_size_bits(old_mode);
1130 } /* is_completely_overwritten */
1133 * follow the memory chain as long as there are only Loads and alias free Stores.
1135 * INC_MASTER() must be called before dive into
1137 static unsigned follow_Mem_chain_for_Store(ir_node *store, ir_node *curr) {
1139 ldst_info_t *info = get_irn_link(store);
1141 ir_node *ptr = get_Store_ptr(store);
1142 ir_node *mem = get_Store_mem(store);
1143 ir_node *value = get_Store_value(store);
1144 ir_mode *mode = get_irn_mode(value);
1145 ir_node *block = get_nodes_block(store);
1146 ir_node *mblk = get_Block_MacroBlock(block);
1148 for (pred = curr; pred != store;) {
1149 ldst_info_t *pred_info = get_irn_link(pred);
1152 * BEWARE: one might think that checking the modes is useless, because
1153 * if the pointers are identical, they refer to the same object.
1154 * This is only true in strong typed languages, not is C were the following
1155 * is possible *(ir_type1 *)p = a; *(ir_type2 *)p = b ...
1156 * However, if the mode that is written have a bigger or equal size the the old
1157 * one, the old value is completely overwritten and can be killed ...
1159 if (is_Store(pred) && get_Store_ptr(pred) == ptr &&
1160 get_nodes_MacroBlock(pred) == mblk &&
1161 is_completely_overwritten(get_irn_mode(get_Store_value(pred)), mode)) {
1163 * a Store after a Store in the same MacroBlock -- a write after write.
1164 * We may remove the first Store, if it does not have an exception handler.
1166 * TODO: What, if both have the same exception handler ???
1168 if (get_Store_volatility(pred) != volatility_is_volatile && !pred_info->projs[pn_Store_X_except]) {
1169 DBG_OPT_WAW(pred, store);
1170 exchange(pred_info->projs[pn_Store_M], get_Store_mem(pred));
1172 reduce_adr_usage(ptr);
1175 } else if (is_Load(pred) && get_Load_ptr(pred) == ptr &&
1176 value == pred_info->projs[pn_Load_res]) {
1178 * a Store of a value just loaded from the same address
1179 * -- a write after read.
1180 * We may remove the Store, if it does not have an exception
1183 if (! info->projs[pn_Store_X_except]) {
1184 DBG_OPT_WAR(store, pred);
1185 exchange(info->projs[pn_Store_M], mem);
1187 reduce_adr_usage(ptr);
1192 if (is_Store(pred)) {
1193 /* check if we can pass thru this store */
1194 ir_alias_relation rel = get_alias_relation(
1196 get_Store_ptr(pred),
1197 get_irn_mode(get_Store_value(pred)),
1199 /* if the might be an alias, we cannot pass this Store */
1200 if (rel != ir_no_alias)
1202 pred = skip_Proj(get_Store_mem(pred));
1203 } else if (is_Load(pred)) {
1204 ir_alias_relation rel = get_alias_relation(
1205 current_ir_graph, get_Load_ptr(pred), get_Load_mode(pred),
1207 if (rel != ir_no_alias)
1210 pred = skip_Proj(get_Load_mem(pred));
1212 /* follow only Load chains */
1216 /* check for cycles */
1217 if (NODE_VISITED(pred_info))
1219 MARK_NODE(pred_info);
1222 if (is_Sync(pred)) {
1225 /* handle all Sync predecessors */
1226 for (i = get_Sync_n_preds(pred) - 1; i >= 0; --i) {
1227 res |= follow_Mem_chain_for_Store(store, skip_Proj(get_Sync_pred(pred, i)));
1233 } /* follow_Mem_chain_for_Store */
1238 * @param store the Store node
1240 static unsigned optimize_store(ir_node *store) {
1243 if (get_Store_volatility(store) == volatility_is_volatile)
1246 ptr = get_Store_ptr(store);
1248 /* Check, if the address of this Store is used more than once.
1249 * If not, this Store cannot be removed in any case. */
1250 if (get_irn_n_uses(ptr) <= 1)
1253 mem = get_Store_mem(store);
1255 /* follow the memory chain as long as there are only Loads */
1258 return follow_Mem_chain_for_Store(store, skip_Proj(mem));
1259 } /* optimize_store */
1262 * walker, optimizes Phi after Stores to identical places:
1263 * Does the following optimization:
1266 * val1 val2 val3 val1 val2 val3
1268 * Store Store Store \ | /
1275 * This reduces the number of stores and allows for predicated execution.
1276 * Moves Stores back to the end of a function which may be bad.
1278 * This is only possible if the predecessor blocks have only one successor.
1280 static unsigned optimize_phi(ir_node *phi, walk_env_t *wenv)
1283 ir_node *store, *old_store, *ptr, *block, *phi_block, *phiM, *phiD, *exc, *projM;
1285 ir_node **inM, **inD, **projMs;
1287 dbg_info *db = NULL;
1289 block_info_t *bl_info;
1292 /* Must be a memory Phi */
1293 if (get_irn_mode(phi) != mode_M)
1296 n = get_Phi_n_preds(phi);
1300 /* must be only one user */
1301 projM = get_Phi_pred(phi, 0);
1302 if (get_irn_n_edges(projM) != 1)
1305 store = skip_Proj(projM);
1307 if (!is_Store(store))
1310 block = get_nodes_block(store);
1312 /* abort on dead blocks */
1313 if (is_Block_dead(block))
1316 /* check if the block is post dominated by Phi-block
1317 and has no exception exit */
1318 bl_info = get_irn_link(block);
1319 if (bl_info->flags & BLOCK_HAS_EXC)
1322 phi_block = get_nodes_block(phi);
1323 if (! block_strictly_postdominates(phi_block, block))
1326 /* this is the address of the store */
1327 ptr = get_Store_ptr(store);
1328 mode = get_irn_mode(get_Store_value(store));
1329 info = get_irn_link(store);
1330 exc = info->exc_block;
1332 for (i = 1; i < n; ++i) {
1333 ir_node *pred = get_Phi_pred(phi, i);
1335 if (get_irn_n_edges(pred) != 1)
1338 pred = skip_Proj(pred);
1339 if (!is_Store(pred))
1342 if (ptr != get_Store_ptr(pred) || mode != get_irn_mode(get_Store_value(pred)))
1345 info = get_irn_link(pred);
1347 /* check, if all stores have the same exception flow */
1348 if (exc != info->exc_block)
1351 /* abort on dead blocks */
1352 block = get_nodes_block(pred);
1353 if (is_Block_dead(block))
1356 /* check if the block is post dominated by Phi-block
1357 and has no exception exit. Note that block must be different from
1358 Phi-block, else we would move a Store from end End of a block to its
1360 bl_info = get_irn_link(block);
1361 if (bl_info->flags & BLOCK_HAS_EXC)
1363 if (block == phi_block || ! block_postdominates(phi_block, block))
1368 * ok, when we are here, we found all predecessors of a Phi that
1369 * are Stores to the same address and size. That means whatever
1370 * we do before we enter the block of the Phi, we do a Store.
1371 * So, we can move the Store to the current block:
1373 * val1 val2 val3 val1 val2 val3
1375 * | Str | | Str | | Str | \ | /
1381 * Is only allowed if the predecessor blocks have only one successor.
1384 NEW_ARR_A(ir_node *, projMs, n);
1385 NEW_ARR_A(ir_node *, inM, n);
1386 NEW_ARR_A(ir_node *, inD, n);
1387 NEW_ARR_A(int, idx, n);
1389 /* Prepare: Collect all Store nodes. We must do this
1390 first because we otherwise may loose a store when exchanging its
1393 for (i = n - 1; i >= 0; --i) {
1396 projMs[i] = get_Phi_pred(phi, i);
1397 assert(is_Proj(projMs[i]));
1399 store = get_Proj_pred(projMs[i]);
1400 info = get_irn_link(store);
1402 inM[i] = get_Store_mem(store);
1403 inD[i] = get_Store_value(store);
1404 idx[i] = info->exc_idx;
1406 block = get_nodes_block(phi);
1408 /* second step: create a new memory Phi */
1409 phiM = new_rd_Phi(get_irn_dbg_info(phi), current_ir_graph, block, n, inM, mode_M);
1411 /* third step: create a new data Phi */
1412 phiD = new_rd_Phi(get_irn_dbg_info(phi), current_ir_graph, block, n, inD, mode);
1414 /* rewire memory and kill the node */
1415 for (i = n - 1; i >= 0; --i) {
1416 ir_node *proj = projMs[i];
1419 ir_node *store = get_Proj_pred(proj);
1420 exchange(proj, inM[i]);
1425 /* fourth step: create the Store */
1426 store = new_rd_Store(db, current_ir_graph, block, phiM, ptr, phiD);
1428 co_set_irn_name(store, co_get_irn_ident(old_store));
1431 projM = new_rd_Proj(NULL, current_ir_graph, block, store, mode_M, pn_Store_M);
1433 info = get_ldst_info(store, &wenv->obst);
1434 info->projs[pn_Store_M] = projM;
1436 /* fifths step: repair exception flow */
1438 ir_node *projX = new_rd_Proj(NULL, current_ir_graph, block, store, mode_X, pn_Store_X_except);
1440 info->projs[pn_Store_X_except] = projX;
1441 info->exc_block = exc;
1442 info->exc_idx = idx[0];
1444 for (i = 0; i < n; ++i) {
1445 set_Block_cfgpred(exc, idx[i], projX);
1449 /* the exception block should be optimized as some inputs are identical now */
1455 /* sixth step: replace old Phi */
1456 exchange(phi, projM);
1458 return res | DF_CHANGED;
1459 } /* optimize_phi */
1462 * walker, do the optimizations
1464 static void do_load_store_optimize(ir_node *n, void *env) {
1465 walk_env_t *wenv = env;
1467 switch (get_irn_opcode(n)) {
1470 wenv->changes |= optimize_load(n);
1474 wenv->changes |= optimize_store(n);
1478 wenv->changes |= optimize_phi(n, wenv);
1484 } /* do_load_store_optimize */
1487 typedef struct scc {
1488 ir_node *head; /**< the head of the list */
1491 /** A node entry. */
1492 typedef struct node_entry {
1493 unsigned DFSnum; /**< the DFS number of this node */
1494 unsigned low; /**< the low number of this node */
1495 ir_node *header; /**< the header of this node */
1496 int in_stack; /**< flag, set if the node is on the stack */
1497 ir_node *next; /**< link to the next node the the same scc */
1498 scc *pscc; /**< the scc of this node */
1499 unsigned POnum; /**< the post order number for blocks */
1502 /** A loop entry. */
1503 typedef struct loop_env {
1504 ir_phase ph; /**< the phase object */
1505 ir_node **stack; /**< the node stack */
1506 int tos; /**< tos index */
1507 unsigned nextDFSnum; /**< the current DFS number */
1508 unsigned POnum; /**< current post order number */
1510 unsigned changes; /**< a bitmask of graph changes */
1514 * Gets the node_entry of a node
1516 static node_entry *get_irn_ne(ir_node *irn, loop_env *env) {
1517 ir_phase *ph = &env->ph;
1518 node_entry *e = phase_get_irn_data(&env->ph, irn);
1521 e = phase_alloc(ph, sizeof(*e));
1522 memset(e, 0, sizeof(*e));
1523 phase_set_irn_data(ph, irn, e);
1529 * Push a node onto the stack.
1531 * @param env the loop environment
1532 * @param n the node to push
1534 static void push(loop_env *env, ir_node *n) {
1537 if (env->tos == ARR_LEN(env->stack)) {
1538 int nlen = ARR_LEN(env->stack) * 2;
1539 ARR_RESIZE(ir_node *, env->stack, nlen);
1541 env->stack[env->tos++] = n;
1542 e = get_irn_ne(n, env);
1547 * pop a node from the stack
1549 * @param env the loop environment
1551 * @return The topmost node
1553 static ir_node *pop(loop_env *env) {
1554 ir_node *n = env->stack[--env->tos];
1555 node_entry *e = get_irn_ne(n, env);
1562 * Check if irn is a region constant.
1563 * The block or irn must strictly dominate the header block.
1565 * @param irn the node to check
1566 * @param header_block the header block of the induction variable
1568 static int is_rc(ir_node *irn, ir_node *header_block) {
1569 ir_node *block = get_nodes_block(irn);
1571 return (block != header_block) && block_dominates(block, header_block);
1574 typedef struct phi_entry phi_entry;
1576 ir_node *phi; /**< A phi with a region const memory. */
1577 int pos; /**< The position of the region const memory */
1578 ir_node *load; /**< the newly created load for this phi */
1583 * Move loops out of loops if possible.
1585 * @param pscc the loop described by an SCC
1586 * @param env the loop environment
1588 static void move_loads_out_of_loops(scc *pscc, loop_env *env) {
1589 ir_node *phi, *load, *next, *other, *next_other;
1592 phi_entry *phi_list = NULL;
1594 /* collect all outer memories */
1595 for (phi = pscc->head; phi != NULL; phi = next) {
1596 node_entry *ne = get_irn_ne(phi, env);
1599 /* check all memory Phi's */
1603 assert(get_irn_mode(phi) == mode_M && "DFS geturn non-memory Phi");
1605 for (j = get_irn_arity(phi) - 1; j >= 0; --j) {
1606 ir_node *pred = get_irn_n(phi, j);
1607 node_entry *pe = get_irn_ne(pred, env);
1609 if (pe->pscc != ne->pscc) {
1610 /* not in the same SCC, is region const */
1611 phi_entry *pe = phase_alloc(&env->ph, sizeof(*pe));
1615 pe->next = phi_list;
1620 /* no Phis no fun */
1621 assert(phi_list != NULL && "DFS found a loop without Phi");
1623 for (load = pscc->head; load; load = next) {
1625 node_entry *ne = get_irn_ne(load, env);
1628 if (is_Load(load)) {
1629 ldst_info_t *info = get_irn_link(load);
1630 ir_node *ptr = get_Load_ptr(load);
1632 /* for now, we cannot handle Loads with exceptions */
1633 if (info->projs[pn_Load_res] == NULL || info->projs[pn_Load_X_regular] != NULL || info->projs[pn_Load_X_except] != NULL)
1636 /* for now, we can only handle Load(Global) */
1637 if (! is_Global(ptr))
1639 ent = get_Global_entity(ptr);
1640 load_mode = get_Load_mode(load);
1641 for (other = pscc->head; other != NULL; other = next_other) {
1642 node_entry *ne = get_irn_ne(other, env);
1643 next_other = ne->next;
1645 if (is_Store(other)) {
1646 ir_alias_relation rel = get_alias_relation(
1648 get_Store_ptr(other),
1649 get_irn_mode(get_Store_value(other)),
1651 /* if the might be an alias, we cannot pass this Store */
1652 if (rel != ir_no_alias)
1655 /* only pure Calls are allowed here, so ignore them */
1657 if (other == NULL) {
1662 /* for now, we cannot handle more than one input */
1663 if (phi_list->next != NULL)
1666 /* yep, no aliasing Store found, Load can be moved */
1667 DB((dbg, LEVEL_1, " Found a Load that could be moved: %+F\n", load));
1669 db = get_irn_dbg_info(load);
1670 for (pe = phi_list; pe != NULL; pe = pe->next) {
1672 ir_node *phi = pe->phi;
1673 ir_node *blk = get_nodes_block(phi);
1674 ir_node *pred = get_Block_cfgpred_block(blk, pos);
1677 pe->load = irn = new_rd_Load(db, current_ir_graph, pred, get_Phi_pred(phi, pos), ptr, load_mode);
1678 ninfo = get_ldst_info(irn, phase_obst(&env->ph));
1680 ninfo->projs[pn_Load_M] = mem = new_r_Proj(current_ir_graph, pred, irn, mode_M, pn_Load_M);
1681 set_Phi_pred(phi, pos, mem);
1683 ninfo->projs[pn_Load_res] = new_r_Proj(current_ir_graph, pred, irn, load_mode, pn_Load_res);
1685 DB((dbg, LEVEL_1, " Created %+F in %+F\n", irn, pred));
1688 /* now kill the old Load */
1689 exchange(info->projs[pn_Load_M], get_Load_mem(load));
1690 exchange(info->projs[pn_Load_res], ninfo->projs[pn_Load_res]);
1692 env->changes |= DF_CHANGED;
1696 } /* move_loads_out_of_loops */
1699 * Process a loop SCC.
1701 * @param pscc the SCC
1702 * @param env the loop environment
1704 static void process_loop(scc *pscc, loop_env *env) {
1705 ir_node *irn, *next, *header = NULL;
1706 node_entry *b, *h = NULL;
1707 int j, only_phi, num_outside, process = 0;
1710 /* find the header block for this scc */
1711 for (irn = pscc->head; irn; irn = next) {
1712 node_entry *e = get_irn_ne(irn, env);
1713 ir_node *block = get_nodes_block(irn);
1716 b = get_irn_ne(block, env);
1719 if (h->POnum < b->POnum) {
1730 /* check if this scc contains only Phi, Loads or Stores nodes */
1734 for (irn = pscc->head; irn; irn = next) {
1735 node_entry *e = get_irn_ne(irn, env);
1738 switch (get_irn_opcode(irn)) {
1740 if (is_Call_pure(irn)) {
1741 /* pure calls can be treated like loads */
1745 /* non-pure calls must be handle like may-alias Stores */
1748 /* cannot handle CopyB yet */
1752 if (get_Load_volatility(irn) == volatility_is_volatile) {
1753 /* cannot handle loops with volatile Loads */
1759 if (get_Store_volatility(irn) == volatility_is_volatile) {
1760 /* cannot handle loops with volatile Stores */
1769 for (j = get_irn_arity(irn) - 1; j >= 0; --j) {
1770 ir_node *pred = get_irn_n(irn, j);
1771 node_entry *pe = get_irn_ne(pred, env);
1773 if (pe->pscc != e->pscc) {
1774 /* not in the same SCC, must be a region const */
1775 if (! is_rc(pred, header)) {
1776 /* not a memory loop */
1782 } else if (out_rc != pred) {
1793 /* found a memory loop */
1794 DB((dbg, LEVEL_2, " Found a memory loop:\n "));
1795 if (only_phi && num_outside == 1) {
1796 /* a phi cycle with only one real predecessor can be collapsed */
1797 DB((dbg, LEVEL_2, " Found an USELESS Phi cycle:\n "));
1799 for (irn = pscc->head; irn; irn = next) {
1800 node_entry *e = get_irn_ne(irn, env);
1803 exchange(irn, out_rc);
1805 env->changes |= DF_CHANGED;
1809 /* set the header for every node in this scc */
1810 for (irn = pscc->head; irn; irn = next) {
1811 node_entry *e = get_irn_ne(irn, env);
1814 DB((dbg, LEVEL_2, " %+F,", irn));
1816 DB((dbg, LEVEL_2, "\n"));
1818 move_loads_out_of_loops(pscc, env);
1822 } /* process_loop */
1827 * @param pscc the SCC
1828 * @param env the loop environment
1830 static void process_scc(scc *pscc, loop_env *env) {
1831 ir_node *head = pscc->head;
1832 node_entry *e = get_irn_ne(head, env);
1834 #ifdef DEBUG_libfirm
1836 ir_node *irn, *next;
1838 DB((dbg, LEVEL_4, " SCC at %p:\n ", pscc));
1839 for (irn = pscc->head; irn; irn = next) {
1840 node_entry *e = get_irn_ne(irn, env);
1844 DB((dbg, LEVEL_4, " %+F,", irn));
1846 DB((dbg, LEVEL_4, "\n"));
1850 if (e->next != NULL) {
1851 /* this SCC has more than one member */
1852 process_loop(pscc, env);
1857 * Do Tarjan's SCC algorithm and drive load/store optimization.
1859 * @param irn start at this node
1860 * @param env the loop environment
1862 static void dfs(ir_node *irn, loop_env *env)
1865 node_entry *node = get_irn_ne(irn, env);
1867 mark_irn_visited(irn);
1869 node->DFSnum = env->nextDFSnum++;
1870 node->low = node->DFSnum;
1874 if (is_Phi(irn) || is_Sync(irn)) {
1875 n = get_irn_arity(irn);
1876 for (i = 0; i < n; ++i) {
1877 ir_node *pred = get_irn_n(irn, i);
1878 node_entry *o = get_irn_ne(pred, env);
1880 if (irn_not_visited(pred)) {
1882 node->low = MIN(node->low, o->low);
1884 if (o->DFSnum < node->DFSnum && o->in_stack)
1885 node->low = MIN(o->DFSnum, node->low);
1887 } else if (is_fragile_op(irn)) {
1888 ir_node *pred = get_fragile_op_mem(irn);
1889 node_entry *o = get_irn_ne(pred, env);
1891 if (irn_not_visited(pred)) {
1893 node->low = MIN(node->low, o->low);
1895 if (o->DFSnum < node->DFSnum && o->in_stack)
1896 node->low = MIN(o->DFSnum, node->low);
1897 } else if (is_Proj(irn)) {
1898 ir_node *pred = get_Proj_pred(irn);
1899 node_entry *o = get_irn_ne(pred, env);
1901 if (irn_not_visited(pred)) {
1903 node->low = MIN(node->low, o->low);
1905 if (o->DFSnum < node->DFSnum && o->in_stack)
1906 node->low = MIN(o->DFSnum, node->low);
1909 /* IGNORE predecessors */
1912 if (node->low == node->DFSnum) {
1913 scc *pscc = phase_alloc(&env->ph, sizeof(*pscc));
1921 e = get_irn_ne(x, env);
1923 e->next = pscc->head;
1927 process_scc(pscc, env);
1932 * Do the DFS on the memory edges a graph.
1934 * @param irg the graph to process
1935 * @param env the loop environment
1937 static void do_dfs(ir_graph *irg, loop_env *env) {
1938 ir_graph *rem = current_ir_graph;
1939 ir_node *endblk, *end;
1942 current_ir_graph = irg;
1943 inc_irg_visited(irg);
1945 /* visit all memory nodes */
1946 endblk = get_irg_end_block(irg);
1947 for (i = get_Block_n_cfgpreds(endblk) - 1; i >= 0; --i) {
1948 ir_node *pred = get_Block_cfgpred(endblk, i);
1950 pred = skip_Proj(pred);
1951 if (is_Return(pred))
1952 dfs(get_Return_mem(pred), env);
1953 else if (is_Raise(pred))
1954 dfs(get_Raise_mem(pred), env);
1955 else if (is_fragile_op(pred))
1956 dfs(get_fragile_op_mem(pred), env);
1958 assert(0 && "Unknown EndBlock predecessor");
1962 /* visit the keep-alives */
1963 end = get_irg_end(irg);
1964 for (i = get_End_n_keepalives(end) - 1; i >= 0; --i) {
1965 ir_node *ka = get_End_keepalive(end, i);
1967 if (is_Phi(ka) && irn_not_visited(ka))
1970 current_ir_graph = rem;
1974 * Initialize new phase data. We do this always explicit, so return NULL here
1976 static void *init_loop_data(ir_phase *ph, const ir_node *irn, void *data) {
1981 } /* init_loop_data */
1984 * Optimize Loads/Stores in loops.
1986 * @param irg the graph
1988 static int optimize_loops(ir_graph *irg) {
1991 env.stack = NEW_ARR_F(ir_node *, 128);
1996 phase_init(&env.ph, "ldstopt", irg, PHASE_DEFAULT_GROWTH, init_loop_data, NULL);
1998 /* calculate the SCC's and drive loop optimization. */
2001 DEL_ARR_F(env.stack);
2002 phase_free(&env.ph);
2005 } /* optimize_loops */
2008 * do the load store optimization
2010 void optimize_load_store(ir_graph *irg) {
2013 FIRM_DBG_REGISTER(dbg, "firm.opt.ldstopt");
2015 assert(get_irg_phase_state(irg) != phase_building);
2016 assert(get_irg_pinned(irg) != op_pin_state_floats &&
2017 "LoadStore optimization needs pinned graph");
2019 /* we need landing pads */
2020 remove_critical_cf_edges(irg);
2024 /* for Phi optimization post-dominators are needed ... */
2025 assure_postdoms(irg);
2027 if (get_opt_alias_analysis()) {
2028 assure_irg_address_taken_computed(irg);
2029 assure_irp_globals_address_taken_computed();
2032 obstack_init(&env.obst);
2035 /* init the links, then collect Loads/Stores/Proj's in lists */
2037 irg_walk_graph(irg, firm_clear_link, collect_nodes, &env);
2039 /* now we have collected enough information, optimize */
2040 irg_walk_graph(irg, NULL, do_load_store_optimize, &env);
2042 env.changes |= optimize_loops(irg);
2044 obstack_free(&env.obst, NULL);
2046 /* Handle graph state */
2048 set_irg_outs_inconsistent(irg);
2051 if (env.changes & CF_CHANGED) {
2052 /* is this really needed: Yes, control flow changed, block might
2053 have Bad() predecessors. */
2054 set_irg_doms_inconsistent(irg);
2056 } /* optimize_load_store */