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;
501 ir_initializer_t *initializer;
509 entry.ent = field = get_Sel_entity(ptr);
510 tp = get_entity_owner(field);
511 if (is_Array_type(tp)) {
512 assert(get_Sel_n_indexs(ptr) == 1 && "multi dim arrays not implemented");
513 entry.index = get_Sel_array_index_long(ptr, 0) - get_array_lower_bound_int(tp, 0);
515 int i, n_members = get_compound_n_members(tp);
516 for (i = 0; i < n_members; ++i) {
517 if (get_compound_member(tp, i) == field)
520 if (i >= n_members) {
521 /* not found: should NOT happen */
526 return rec_find_compound_ent_value(get_Sel_ptr(ptr), &entry);
530 assert(is_SymConst(ptr));
532 ent = get_SymConst_entity(ptr);
533 initializer = get_entity_initializer(ent);
534 for (p = next; p != NULL; p = p->next) {
537 if (initializer->kind != IR_INITIALIZER_COMPOUND)
540 n = get_initializer_compound_n_entries(initializer);
543 initializer = get_initializer_compound_value(initializer, p->index);
546 switch (initializer->kind) {
547 case IR_INITIALIZER_CONST:
548 return get_initializer_const_value(initializer);
549 case IR_INITIALIZER_TARVAL:
550 case IR_INITIALIZER_NULL:
556 static ir_node *find_compound_ent_value(ir_node *ptr) {
557 return rec_find_compound_ent_value(ptr, NULL);
561 static void reduce_adr_usage(ir_node *ptr);
564 * Update a Load that may lost it's usage.
566 static void handle_load_update(ir_node *load) {
567 ldst_info_t *info = get_irn_link(load);
569 /* do NOT touch volatile loads for now */
570 if (get_Load_volatility(load) == volatility_is_volatile)
573 if (! info->projs[pn_Load_res] && ! info->projs[pn_Load_X_except]) {
574 ir_node *ptr = get_Load_ptr(load);
575 ir_node *mem = get_Load_mem(load);
577 /* a Load which value is neither used nor exception checked, remove it */
578 exchange(info->projs[pn_Load_M], mem);
579 if (info->projs[pn_Load_X_regular])
580 exchange(info->projs[pn_Load_X_regular], new_r_Jmp(current_ir_graph, get_nodes_block(load)));
582 reduce_adr_usage(ptr);
584 } /* handle_load_update */
587 * A Use of an address node is vanished. Check if this was a Proj
588 * node and update the counters.
590 static void reduce_adr_usage(ir_node *ptr) {
592 if (get_irn_n_edges(ptr) <= 0) {
593 /* this Proj is dead now */
594 ir_node *pred = get_Proj_pred(ptr);
597 ldst_info_t *info = get_irn_link(pred);
598 info->projs[get_Proj_proj(ptr)] = NULL;
600 /* this node lost it's result proj, handle that */
601 handle_load_update(pred);
605 } /* reduce_adr_usage */
608 * Check, if an already existing value of mode old_mode can be converted
609 * into the needed one new_mode without loss.
611 static int can_use_stored_value(ir_mode *old_mode, ir_mode *new_mode) {
612 if (old_mode == new_mode)
615 /* if both modes are two-complement ones, we can always convert the
616 Stored value into the needed one. */
617 if (get_mode_size_bits(old_mode) >= get_mode_size_bits(new_mode) &&
618 get_mode_arithmetic(old_mode) == irma_twos_complement &&
619 get_mode_arithmetic(new_mode) == irma_twos_complement)
622 } /* can_use_stored_value */
625 * Check whether a Call is at least pure, ie. does only read memory.
627 static unsigned is_Call_pure(ir_node *call) {
628 ir_type *call_tp = get_Call_type(call);
629 unsigned prop = get_method_additional_properties(call_tp);
631 /* check first the call type */
632 if ((prop & (mtp_property_const|mtp_property_pure)) == 0) {
633 /* try the called entity */
634 ir_node *ptr = get_Call_ptr(call);
636 if (is_Global(ptr)) {
637 ir_entity *ent = get_Global_entity(ptr);
639 prop = get_entity_additional_properties(ent);
642 return (prop & (mtp_property_const|mtp_property_pure)) != 0;
646 * Follow the memory chain as long as there are only Loads,
647 * alias free Stores, and constant Calls and try to replace the
648 * current Load by a previous ones.
649 * Note that in unreachable loops it might happen that we reach
650 * load again, as well as we can fall into a cycle.
651 * We break such cycles using a special visited flag.
653 * INC_MASTER() must be called before dive into
655 static unsigned follow_Mem_chain(ir_node *load, ir_node *curr) {
657 ldst_info_t *info = get_irn_link(load);
659 ir_node *ptr = get_Load_ptr(load);
660 ir_node *mem = get_Load_mem(load);
661 ir_mode *load_mode = get_Load_mode(load);
663 for (pred = curr; load != pred; ) {
664 ldst_info_t *pred_info = get_irn_link(pred);
667 * BEWARE: one might think that checking the modes is useless, because
668 * if the pointers are identical, they refer to the same object.
669 * This is only true in strong typed languages, not in C were the following
670 * is possible a = *(ir_type1 *)p; b = *(ir_type2 *)p ...
672 if (is_Store(pred) && get_Store_ptr(pred) == ptr &&
673 can_use_stored_value(get_irn_mode(get_Store_value(pred)), load_mode)) {
675 * a Load immediately after a Store -- a read after write.
676 * We may remove the Load, if both Load & Store does not have an exception handler
677 * OR they are in the same MacroBlock. In the latter case the Load cannot
678 * throw an exception when the previous Store was quiet.
680 * Why we need to check for Store Exception? If the Store cannot
681 * be executed (ROM) the exception handler might simply jump into
682 * the load MacroBlock :-(
683 * We could make it a little bit better if we would know that the exception
684 * handler of the Store jumps directly to the end...
686 if ((pred_info->projs[pn_Store_X_except] == NULL && info->projs[pn_Load_X_except] == NULL) ||
687 get_nodes_MacroBlock(load) == get_nodes_MacroBlock(pred)) {
688 ir_node *value = get_Store_value(pred);
690 DBG_OPT_RAW(load, value);
692 /* add an convert if needed */
693 if (get_irn_mode(get_Store_value(pred)) != load_mode) {
694 value = new_r_Conv(current_ir_graph, get_nodes_block(load), value, load_mode);
697 if (info->projs[pn_Load_M])
698 exchange(info->projs[pn_Load_M], mem);
701 if (info->projs[pn_Load_X_except]) {
702 exchange( info->projs[pn_Load_X_except], new_Bad());
705 if (info->projs[pn_Load_X_regular]) {
706 exchange( info->projs[pn_Load_X_regular], new_r_Jmp(current_ir_graph, get_nodes_block(load)));
710 if (info->projs[pn_Load_res])
711 exchange(info->projs[pn_Load_res], value);
714 reduce_adr_usage(ptr);
715 return res | DF_CHANGED;
717 } else if (is_Load(pred) && get_Load_ptr(pred) == ptr &&
718 can_use_stored_value(get_Load_mode(pred), load_mode)) {
720 * a Load after a Load -- a read after read.
721 * We may remove the second Load, if it does not have an exception handler
722 * OR they are in the same MacroBlock. In the later case the Load cannot
723 * throw an exception when the previous Load was quiet.
725 * Here, there is no need to check if the previous Load has an exception
726 * hander because they would have exact the same exception...
728 if (info->projs[pn_Load_X_except] == NULL || get_nodes_MacroBlock(load) == get_nodes_MacroBlock(pred)) {
731 DBG_OPT_RAR(load, pred);
733 /* the result is used */
734 if (info->projs[pn_Load_res]) {
735 if (pred_info->projs[pn_Load_res] == NULL) {
736 /* create a new Proj again */
737 pred_info->projs[pn_Load_res] = new_r_Proj(current_ir_graph, get_nodes_block(pred), pred, get_Load_mode(pred), pn_Load_res);
739 value = pred_info->projs[pn_Load_res];
741 /* add an convert if needed */
742 if (get_Load_mode(pred) != load_mode) {
743 value = new_r_Conv(current_ir_graph, get_nodes_block(load), value, load_mode);
746 exchange(info->projs[pn_Load_res], value);
749 if (info->projs[pn_Load_M])
750 exchange(info->projs[pn_Load_M], mem);
753 if (info->projs[pn_Load_X_except]) {
754 exchange(info->projs[pn_Load_X_except], new_Bad());
757 if (info->projs[pn_Load_X_regular]) {
758 exchange( info->projs[pn_Load_X_regular], new_r_Jmp(current_ir_graph, get_nodes_block(load)));
763 reduce_adr_usage(ptr);
764 return res |= DF_CHANGED;
768 if (is_Store(pred)) {
769 /* check if we can pass through this store */
770 ir_alias_relation rel = get_alias_relation(
773 get_irn_mode(get_Store_value(pred)),
775 /* if the might be an alias, we cannot pass this Store */
776 if (rel != ir_no_alias)
778 pred = skip_Proj(get_Store_mem(pred));
779 } else if (is_Load(pred)) {
780 pred = skip_Proj(get_Load_mem(pred));
781 } else if (is_Call(pred)) {
782 if (is_Call_pure(pred)) {
783 /* The called graph is at least pure, so there are no Store's
784 in it. We can handle it like a Load and skip it. */
785 pred = skip_Proj(get_Call_mem(pred));
787 /* there might be Store's in the graph, stop here */
791 /* follow only Load chains */
795 /* check for cycles */
796 if (NODE_VISITED(pred_info))
798 MARK_NODE(pred_info);
804 /* handle all Sync predecessors */
805 for (i = get_Sync_n_preds(pred) - 1; i >= 0; --i) {
806 res |= follow_Mem_chain(load, skip_Proj(get_Sync_pred(pred, i)));
813 } /* follow_Mem_chain */
818 * @param load the Load node
820 static unsigned optimize_load(ir_node *load)
822 ldst_info_t *info = get_irn_link(load);
823 ir_node *mem, *ptr, *new_node;
827 /* do NOT touch volatile loads for now */
828 if (get_Load_volatility(load) == volatility_is_volatile)
831 /* the address of the load to be optimized */
832 ptr = get_Load_ptr(load);
835 * Check if we can remove the exception from a Load:
836 * This can be done, if the address is from an Sel(Alloc) and
837 * the Sel type is a subtype of the allocated type.
839 * This optimizes some often used OO constructs,
840 * like x = new O; x->t;
842 if (info->projs[pn_Load_X_except]) {
844 ir_node *mem = get_Sel_mem(ptr);
846 /* FIXME: works with the current FE, but better use the base */
847 if (is_Alloc(skip_Proj(mem))) {
848 /* ok, check the types */
849 ir_entity *ent = get_Sel_entity(ptr);
850 ir_type *s_type = get_entity_type(ent);
851 ir_type *a_type = get_Alloc_type(mem);
853 if (is_SubClass_of(s_type, a_type)) {
854 /* ok, condition met: there can't be an exception because
855 * Alloc guarantees that enough memory was allocated */
857 exchange(info->projs[pn_Load_X_except], new_Bad());
858 info->projs[pn_Load_X_except] = NULL;
859 exchange(info->projs[pn_Load_X_regular], new_r_Jmp(current_ir_graph, get_nodes_block(load)));
860 info->projs[pn_Load_X_regular] = NULL;
864 } else if (is_Alloc(skip_Proj(skip_Cast(ptr)))) {
865 /* simple case: a direct load after an Alloc. Firm Alloc throw
866 * an exception in case of out-of-memory. So, there is no way for an
867 * exception in this load.
868 * This code is constructed by the "exception lowering" in the Jack compiler.
870 exchange(info->projs[pn_Load_X_except], new_Bad());
871 info->projs[pn_Load_X_except] = NULL;
872 exchange(info->projs[pn_Load_X_regular], new_r_Jmp(current_ir_graph, get_nodes_block(load)));
873 info->projs[pn_Load_X_regular] = NULL;
878 /* The mem of the Load. Must still be returned after optimization. */
879 mem = get_Load_mem(load);
881 if (! info->projs[pn_Load_res] && ! info->projs[pn_Load_X_except]) {
882 /* a Load which value is neither used nor exception checked, remove it */
883 exchange(info->projs[pn_Load_M], mem);
885 if (info->projs[pn_Load_X_regular]) {
886 /* should not happen, but if it does, remove it */
887 exchange(info->projs[pn_Load_X_regular], new_r_Jmp(current_ir_graph, get_nodes_block(load)));
891 reduce_adr_usage(ptr);
892 return res | DF_CHANGED;
895 /* Load from a constant polymorphic field, where we can resolve
897 new_node = transform_node_Load(load);
898 if (new_node != load) {
899 if (info->projs[pn_Load_M]) {
900 exchange(info->projs[pn_Load_M], mem);
901 info->projs[pn_Load_M] = NULL;
903 if (info->projs[pn_Load_X_except]) {
904 exchange(info->projs[pn_Load_X_except], new_Bad());
905 info->projs[pn_Load_X_except] = NULL;
908 if (info->projs[pn_Load_X_regular]) {
909 exchange(info->projs[pn_Load_X_regular], new_r_Jmp(current_ir_graph, get_nodes_block(load)));
910 info->projs[pn_Load_X_regular] = NULL;
913 if (info->projs[pn_Load_res])
914 exchange(info->projs[pn_Load_res], new_node);
917 reduce_adr_usage(ptr);
918 return res | DF_CHANGED;
921 /* check if we can determine the entity that will be loaded */
922 ent = find_constant_entity(ptr);
924 if ((allocation_static == get_entity_allocation(ent)) &&
925 (visibility_external_allocated != get_entity_visibility(ent))) {
926 /* a static allocation that is not external: there should be NO exception
929 /* no exception, clear the info field as it might be checked later again */
930 if (info->projs[pn_Load_X_except]) {
931 exchange(info->projs[pn_Load_X_except], new_Bad());
932 info->projs[pn_Load_X_except] = NULL;
935 if (info->projs[pn_Load_X_regular]) {
936 exchange(info->projs[pn_Load_X_regular], new_r_Jmp(current_ir_graph, get_nodes_block(load)));
937 info->projs[pn_Load_X_regular] = NULL;
941 if (variability_constant == get_entity_variability(ent)) {
942 if (is_atomic_entity(ent)) {
943 /* Might not be atomic after
944 lowering of Sels. In this
945 case we could also load, but
946 it's more complicated. */
947 /* more simpler case: we load the content of a constant value:
948 * replace it by the constant itself
952 if (info->projs[pn_Load_M]) {
953 exchange(info->projs[pn_Load_M], mem);
957 if (info->projs[pn_Load_res]) {
958 if (is_atomic_entity(ent)) {
959 ir_node *c = copy_const_value(get_irn_dbg_info(load), get_atomic_ent_value(ent));
962 exchange(info->projs[pn_Load_res], c);
967 reduce_adr_usage(ptr);
971 if (ent->has_initializer) {
972 /* new style initializer */
973 c = find_compound_ent_value(ptr);
975 /* old style initializer */
976 compound_graph_path *path = get_accessed_path(ptr);
979 assert(is_proper_compound_graph_path(path, get_compound_graph_path_length(path)-1));
981 c = get_compound_ent_value_by_path(ent, path);
982 free_compound_graph_path(path);
986 if (info->projs[pn_Load_M]) {
987 exchange(info->projs[pn_Load_M], mem);
990 if (info->projs[pn_Load_res]) {
991 exchange(info->projs[pn_Load_res], copy_const_value(get_irn_dbg_info(load), c));
995 reduce_adr_usage(ptr);
998 /* We can not determine a correct access path. E.g., in jack, we load
999 a byte from an object to generate an exception. Happens in test program
1001 printf(">>>>>>>>>>>>> Found access to constant entity %s in function %s\n", get_entity_name(ent),
1002 get_entity_name(get_irg_entity(current_ir_graph)));
1003 ir_printf(" load: %+F\n", load);
1004 ir_printf(" ptr: %+F\n", ptr);
1012 /* Check, if the address of this load is used more than once.
1013 * If not, this load cannot be removed in any case. */
1014 if (get_irn_n_uses(ptr) <= 1)
1018 * follow the memory chain as long as there are only Loads
1019 * and try to replace current Load or Store by a previous one.
1020 * Note that in unreachable loops it might happen that we reach
1021 * load again, as well as we can fall into a cycle.
1022 * We break such cycles using a special visited flag.
1025 res = follow_Mem_chain(load, skip_Proj(mem));
1027 } /* optimize_load */
1030 * Check whether a value of mode new_mode would completely overwrite a value
1031 * of mode old_mode in memory.
1033 static int is_completely_overwritten(ir_mode *old_mode, ir_mode *new_mode)
1035 return get_mode_size_bits(new_mode) >= get_mode_size_bits(old_mode);
1036 } /* is_completely_overwritten */
1039 * follow the memory chain as long as there are only Loads and alias free Stores.
1041 * INC_MASTER() must be called before dive into
1043 static unsigned follow_Mem_chain_for_Store(ir_node *store, ir_node *curr) {
1045 ldst_info_t *info = get_irn_link(store);
1047 ir_node *ptr = get_Store_ptr(store);
1048 ir_node *mem = get_Store_mem(store);
1049 ir_node *value = get_Store_value(store);
1050 ir_mode *mode = get_irn_mode(value);
1051 ir_node *block = get_nodes_block(store);
1052 ir_node *mblk = get_Block_MacroBlock(block);
1054 for (pred = curr; pred != store;) {
1055 ldst_info_t *pred_info = get_irn_link(pred);
1058 * BEWARE: one might think that checking the modes is useless, because
1059 * if the pointers are identical, they refer to the same object.
1060 * This is only true in strong typed languages, not is C were the following
1061 * is possible *(ir_type1 *)p = a; *(ir_type2 *)p = b ...
1062 * However, if the mode that is written have a bigger or equal size the the old
1063 * one, the old value is completely overwritten and can be killed ...
1065 if (is_Store(pred) && get_Store_ptr(pred) == ptr &&
1066 get_nodes_MacroBlock(pred) == mblk &&
1067 is_completely_overwritten(get_irn_mode(get_Store_value(pred)), mode)) {
1069 * a Store after a Store in the same MacroBlock -- a write after write.
1070 * We may remove the first Store, if it does not have an exception handler.
1072 * TODO: What, if both have the same exception handler ???
1074 if (get_Store_volatility(pred) != volatility_is_volatile && !pred_info->projs[pn_Store_X_except]) {
1075 DBG_OPT_WAW(pred, store);
1076 exchange(pred_info->projs[pn_Store_M], get_Store_mem(pred));
1078 reduce_adr_usage(ptr);
1081 } else if (is_Load(pred) && get_Load_ptr(pred) == ptr &&
1082 value == pred_info->projs[pn_Load_res]) {
1084 * a Store of a value just loaded from the same address
1085 * -- a write after read.
1086 * We may remove the Store, if it does not have an exception
1089 if (! info->projs[pn_Store_X_except]) {
1090 DBG_OPT_WAR(store, pred);
1091 exchange(info->projs[pn_Store_M], mem);
1093 reduce_adr_usage(ptr);
1098 if (is_Store(pred)) {
1099 /* check if we can pass thru this store */
1100 ir_alias_relation rel = get_alias_relation(
1102 get_Store_ptr(pred),
1103 get_irn_mode(get_Store_value(pred)),
1105 /* if the might be an alias, we cannot pass this Store */
1106 if (rel != ir_no_alias)
1108 pred = skip_Proj(get_Store_mem(pred));
1109 } else if (is_Load(pred)) {
1110 ir_alias_relation rel = get_alias_relation(
1111 current_ir_graph, get_Load_ptr(pred), get_Load_mode(pred),
1113 if (rel != ir_no_alias)
1116 pred = skip_Proj(get_Load_mem(pred));
1118 /* follow only Load chains */
1122 /* check for cycles */
1123 if (NODE_VISITED(pred_info))
1125 MARK_NODE(pred_info);
1128 if (is_Sync(pred)) {
1131 /* handle all Sync predecessors */
1132 for (i = get_Sync_n_preds(pred) - 1; i >= 0; --i) {
1133 res |= follow_Mem_chain_for_Store(store, skip_Proj(get_Sync_pred(pred, i)));
1139 } /* follow_Mem_chain_for_Store */
1144 * @param store the Store node
1146 static unsigned optimize_store(ir_node *store) {
1149 if (get_Store_volatility(store) == volatility_is_volatile)
1152 ptr = get_Store_ptr(store);
1154 /* Check, if the address of this Store is used more than once.
1155 * If not, this Store cannot be removed in any case. */
1156 if (get_irn_n_uses(ptr) <= 1)
1159 mem = get_Store_mem(store);
1161 /* follow the memory chain as long as there are only Loads */
1164 return follow_Mem_chain_for_Store(store, skip_Proj(mem));
1165 } /* optimize_store */
1168 * walker, optimizes Phi after Stores to identical places:
1169 * Does the following optimization:
1172 * val1 val2 val3 val1 val2 val3
1174 * Store Store Store \ | /
1181 * This reduces the number of stores and allows for predicated execution.
1182 * Moves Stores back to the end of a function which may be bad.
1184 * This is only possible if the predecessor blocks have only one successor.
1186 static unsigned optimize_phi(ir_node *phi, walk_env_t *wenv)
1189 ir_node *store, *old_store, *ptr, *block, *phi_block, *phiM, *phiD, *exc, *projM;
1191 ir_node **inM, **inD, **projMs;
1193 dbg_info *db = NULL;
1195 block_info_t *bl_info;
1198 /* Must be a memory Phi */
1199 if (get_irn_mode(phi) != mode_M)
1202 n = get_Phi_n_preds(phi);
1206 /* must be only one user */
1207 projM = get_Phi_pred(phi, 0);
1208 if (get_irn_n_edges(projM) != 1)
1211 store = skip_Proj(projM);
1213 if (!is_Store(store))
1216 block = get_nodes_block(store);
1218 /* abort on dead blocks */
1219 if (is_Block_dead(block))
1222 /* check if the block is post dominated by Phi-block
1223 and has no exception exit */
1224 bl_info = get_irn_link(block);
1225 if (bl_info->flags & BLOCK_HAS_EXC)
1228 phi_block = get_nodes_block(phi);
1229 if (! block_strictly_postdominates(phi_block, block))
1232 /* this is the address of the store */
1233 ptr = get_Store_ptr(store);
1234 mode = get_irn_mode(get_Store_value(store));
1235 info = get_irn_link(store);
1236 exc = info->exc_block;
1238 for (i = 1; i < n; ++i) {
1239 ir_node *pred = get_Phi_pred(phi, i);
1241 if (get_irn_n_edges(pred) != 1)
1244 pred = skip_Proj(pred);
1245 if (!is_Store(pred))
1248 if (ptr != get_Store_ptr(pred) || mode != get_irn_mode(get_Store_value(pred)))
1251 info = get_irn_link(pred);
1253 /* check, if all stores have the same exception flow */
1254 if (exc != info->exc_block)
1257 /* abort on dead blocks */
1258 block = get_nodes_block(pred);
1259 if (is_Block_dead(block))
1262 /* check if the block is post dominated by Phi-block
1263 and has no exception exit. Note that block must be different from
1264 Phi-block, else we would move a Store from end End of a block to its
1266 bl_info = get_irn_link(block);
1267 if (bl_info->flags & BLOCK_HAS_EXC)
1269 if (block == phi_block || ! block_postdominates(phi_block, block))
1274 * ok, when we are here, we found all predecessors of a Phi that
1275 * are Stores to the same address and size. That means whatever
1276 * we do before we enter the block of the Phi, we do a Store.
1277 * So, we can move the Store to the current block:
1279 * val1 val2 val3 val1 val2 val3
1281 * | Str | | Str | | Str | \ | /
1287 * Is only allowed if the predecessor blocks have only one successor.
1290 NEW_ARR_A(ir_node *, projMs, n);
1291 NEW_ARR_A(ir_node *, inM, n);
1292 NEW_ARR_A(ir_node *, inD, n);
1293 NEW_ARR_A(int, idx, n);
1295 /* Prepare: Collect all Store nodes. We must do this
1296 first because we otherwise may loose a store when exchanging its
1299 for (i = n - 1; i >= 0; --i) {
1302 projMs[i] = get_Phi_pred(phi, i);
1303 assert(is_Proj(projMs[i]));
1305 store = get_Proj_pred(projMs[i]);
1306 info = get_irn_link(store);
1308 inM[i] = get_Store_mem(store);
1309 inD[i] = get_Store_value(store);
1310 idx[i] = info->exc_idx;
1312 block = get_nodes_block(phi);
1314 /* second step: create a new memory Phi */
1315 phiM = new_rd_Phi(get_irn_dbg_info(phi), current_ir_graph, block, n, inM, mode_M);
1317 /* third step: create a new data Phi */
1318 phiD = new_rd_Phi(get_irn_dbg_info(phi), current_ir_graph, block, n, inD, mode);
1320 /* rewire memory and kill the node */
1321 for (i = n - 1; i >= 0; --i) {
1322 ir_node *proj = projMs[i];
1325 ir_node *store = get_Proj_pred(proj);
1326 exchange(proj, inM[i]);
1331 /* fourth step: create the Store */
1332 store = new_rd_Store(db, current_ir_graph, block, phiM, ptr, phiD);
1334 co_set_irn_name(store, co_get_irn_ident(old_store));
1337 projM = new_rd_Proj(NULL, current_ir_graph, block, store, mode_M, pn_Store_M);
1339 info = get_ldst_info(store, &wenv->obst);
1340 info->projs[pn_Store_M] = projM;
1342 /* fifths step: repair exception flow */
1344 ir_node *projX = new_rd_Proj(NULL, current_ir_graph, block, store, mode_X, pn_Store_X_except);
1346 info->projs[pn_Store_X_except] = projX;
1347 info->exc_block = exc;
1348 info->exc_idx = idx[0];
1350 for (i = 0; i < n; ++i) {
1351 set_Block_cfgpred(exc, idx[i], projX);
1355 /* the exception block should be optimized as some inputs are identical now */
1361 /* sixth step: replace old Phi */
1362 exchange(phi, projM);
1364 return res | DF_CHANGED;
1365 } /* optimize_phi */
1368 * walker, do the optimizations
1370 static void do_load_store_optimize(ir_node *n, void *env) {
1371 walk_env_t *wenv = env;
1373 switch (get_irn_opcode(n)) {
1376 wenv->changes |= optimize_load(n);
1380 wenv->changes |= optimize_store(n);
1384 wenv->changes |= optimize_phi(n, wenv);
1390 } /* do_load_store_optimize */
1393 typedef struct scc {
1394 ir_node *head; /**< the head of the list */
1397 /** A node entry. */
1398 typedef struct node_entry {
1399 unsigned DFSnum; /**< the DFS number of this node */
1400 unsigned low; /**< the low number of this node */
1401 ir_node *header; /**< the header of this node */
1402 int in_stack; /**< flag, set if the node is on the stack */
1403 ir_node *next; /**< link to the next node the the same scc */
1404 scc *pscc; /**< the scc of this node */
1405 unsigned POnum; /**< the post order number for blocks */
1408 /** A loop entry. */
1409 typedef struct loop_env {
1410 ir_phase ph; /**< the phase object */
1411 ir_node **stack; /**< the node stack */
1412 int tos; /**< tos index */
1413 unsigned nextDFSnum; /**< the current DFS number */
1414 unsigned POnum; /**< current post order number */
1416 unsigned changes; /**< a bitmask of graph changes */
1420 * Gets the node_entry of a node
1422 static node_entry *get_irn_ne(ir_node *irn, loop_env *env) {
1423 ir_phase *ph = &env->ph;
1424 node_entry *e = phase_get_irn_data(&env->ph, irn);
1427 e = phase_alloc(ph, sizeof(*e));
1428 memset(e, 0, sizeof(*e));
1429 phase_set_irn_data(ph, irn, e);
1435 * Push a node onto the stack.
1437 * @param env the loop environment
1438 * @param n the node to push
1440 static void push(loop_env *env, ir_node *n) {
1443 if (env->tos == ARR_LEN(env->stack)) {
1444 int nlen = ARR_LEN(env->stack) * 2;
1445 ARR_RESIZE(ir_node *, env->stack, nlen);
1447 env->stack[env->tos++] = n;
1448 e = get_irn_ne(n, env);
1453 * pop a node from the stack
1455 * @param env the loop environment
1457 * @return The topmost node
1459 static ir_node *pop(loop_env *env) {
1460 ir_node *n = env->stack[--env->tos];
1461 node_entry *e = get_irn_ne(n, env);
1468 * Check if irn is a region constant.
1469 * The block or irn must strictly dominate the header block.
1471 * @param irn the node to check
1472 * @param header_block the header block of the induction variable
1474 static int is_rc(ir_node *irn, ir_node *header_block) {
1475 ir_node *block = get_nodes_block(irn);
1477 return (block != header_block) && block_dominates(block, header_block);
1480 typedef struct phi_entry phi_entry;
1482 ir_node *phi; /**< A phi with a region const memory. */
1483 int pos; /**< The position of the region const memory */
1484 ir_node *load; /**< the newly created load for this phi */
1489 * Move loops out of loops if possible.
1491 * @param pscc the loop described by an SCC
1492 * @param env the loop environment
1494 static void move_loads_out_of_loops(scc *pscc, loop_env *env) {
1495 ir_node *phi, *load, *next, *other, *next_other;
1498 phi_entry *phi_list = NULL;
1500 /* collect all outer memories */
1501 for (phi = pscc->head; phi != NULL; phi = next) {
1502 node_entry *ne = get_irn_ne(phi, env);
1505 /* check all memory Phi's */
1509 assert(get_irn_mode(phi) == mode_M && "DFS geturn non-memory Phi");
1511 for (j = get_irn_arity(phi) - 1; j >= 0; --j) {
1512 ir_node *pred = get_irn_n(phi, j);
1513 node_entry *pe = get_irn_ne(pred, env);
1515 if (pe->pscc != ne->pscc) {
1516 /* not in the same SCC, is region const */
1517 phi_entry *pe = phase_alloc(&env->ph, sizeof(*pe));
1521 pe->next = phi_list;
1526 /* no Phis no fun */
1527 assert(phi_list != NULL && "DFS found a loop without Phi");
1529 for (load = pscc->head; load; load = next) {
1531 node_entry *ne = get_irn_ne(load, env);
1534 if (is_Load(load)) {
1535 ldst_info_t *info = get_irn_link(load);
1536 ir_node *ptr = get_Load_ptr(load);
1538 /* for now, we cannot handle Loads with exceptions */
1539 if (info->projs[pn_Load_res] == NULL || info->projs[pn_Load_X_regular] != NULL || info->projs[pn_Load_X_except] != NULL)
1542 /* for now, we can only handle Load(Global) */
1543 if (! is_Global(ptr))
1545 ent = get_Global_entity(ptr);
1546 load_mode = get_Load_mode(load);
1547 for (other = pscc->head; other != NULL; other = next_other) {
1548 node_entry *ne = get_irn_ne(other, env);
1549 next_other = ne->next;
1551 if (is_Store(other)) {
1552 ir_alias_relation rel = get_alias_relation(
1554 get_Store_ptr(other),
1555 get_irn_mode(get_Store_value(other)),
1557 /* if the might be an alias, we cannot pass this Store */
1558 if (rel != ir_no_alias)
1561 /* only pure Calls are allowed here, so ignore them */
1563 if (other == NULL) {
1568 /* for now, we cannot handle more than one input */
1569 if (phi_list->next != NULL)
1572 /* yep, no aliasing Store found, Load can be moved */
1573 DB((dbg, LEVEL_1, " Found a Load that could be moved: %+F\n", load));
1575 db = get_irn_dbg_info(load);
1576 for (pe = phi_list; pe != NULL; pe = pe->next) {
1578 ir_node *phi = pe->phi;
1579 ir_node *blk = get_nodes_block(phi);
1580 ir_node *pred = get_Block_cfgpred_block(blk, pos);
1583 pe->load = irn = new_rd_Load(db, current_ir_graph, pred, get_Phi_pred(phi, pos), ptr, load_mode);
1584 ninfo = get_ldst_info(irn, phase_obst(&env->ph));
1586 ninfo->projs[pn_Load_M] = mem = new_r_Proj(current_ir_graph, pred, irn, mode_M, pn_Load_M);
1587 set_Phi_pred(phi, pos, mem);
1589 ninfo->projs[pn_Load_res] = new_r_Proj(current_ir_graph, pred, irn, load_mode, pn_Load_res);
1591 DB((dbg, LEVEL_1, " Created %+F in %+F\n", irn, pred));
1594 /* now kill the old Load */
1595 exchange(info->projs[pn_Load_M], get_Load_mem(load));
1596 exchange(info->projs[pn_Load_res], ninfo->projs[pn_Load_res]);
1598 env->changes |= DF_CHANGED;
1602 } /* move_loads_out_of_loops */
1605 * Process a loop SCC.
1607 * @param pscc the SCC
1608 * @param env the loop environment
1610 static void process_loop(scc *pscc, loop_env *env) {
1611 ir_node *irn, *next, *header = NULL;
1612 node_entry *b, *h = NULL;
1613 int j, only_phi, num_outside, process = 0;
1616 /* find the header block for this scc */
1617 for (irn = pscc->head; irn; irn = next) {
1618 node_entry *e = get_irn_ne(irn, env);
1619 ir_node *block = get_nodes_block(irn);
1622 b = get_irn_ne(block, env);
1625 if (h->POnum < b->POnum) {
1636 /* check if this scc contains only Phi, Loads or Stores nodes */
1640 for (irn = pscc->head; irn; irn = next) {
1641 node_entry *e = get_irn_ne(irn, env);
1644 switch (get_irn_opcode(irn)) {
1646 if (is_Call_pure(irn)) {
1647 /* pure calls can be treated like loads */
1651 /* non-pure calls must be handle like may-alias Stores */
1654 /* cannot handle CopyB yet */
1658 if (get_Load_volatility(irn) == volatility_is_volatile) {
1659 /* cannot handle loops with volatile Loads */
1665 if (get_Store_volatility(irn) == volatility_is_volatile) {
1666 /* cannot handle loops with volatile Stores */
1675 for (j = get_irn_arity(irn) - 1; j >= 0; --j) {
1676 ir_node *pred = get_irn_n(irn, j);
1677 node_entry *pe = get_irn_ne(pred, env);
1679 if (pe->pscc != e->pscc) {
1680 /* not in the same SCC, must be a region const */
1681 if (! is_rc(pred, header)) {
1682 /* not a memory loop */
1688 } else if (out_rc != pred) {
1699 /* found a memory loop */
1700 DB((dbg, LEVEL_2, " Found a memory loop:\n "));
1701 if (only_phi && num_outside == 1) {
1702 /* a phi cycle with only one real predecessor can be collapsed */
1703 DB((dbg, LEVEL_2, " Found an USELESS Phi cycle:\n "));
1705 for (irn = pscc->head; irn; irn = next) {
1706 node_entry *e = get_irn_ne(irn, env);
1709 exchange(irn, out_rc);
1711 env->changes |= DF_CHANGED;
1715 /* set the header for every node in this scc */
1716 for (irn = pscc->head; irn; irn = next) {
1717 node_entry *e = get_irn_ne(irn, env);
1720 DB((dbg, LEVEL_2, " %+F,", irn));
1722 DB((dbg, LEVEL_2, "\n"));
1724 move_loads_out_of_loops(pscc, env);
1728 } /* process_loop */
1733 * @param pscc the SCC
1734 * @param env the loop environment
1736 static void process_scc(scc *pscc, loop_env *env) {
1737 ir_node *head = pscc->head;
1738 node_entry *e = get_irn_ne(head, env);
1740 #ifdef DEBUG_libfirm
1742 ir_node *irn, *next;
1744 DB((dbg, LEVEL_4, " SCC at %p:\n ", pscc));
1745 for (irn = pscc->head; irn; irn = next) {
1746 node_entry *e = get_irn_ne(irn, env);
1750 DB((dbg, LEVEL_4, " %+F,", irn));
1752 DB((dbg, LEVEL_4, "\n"));
1756 if (e->next != NULL) {
1757 /* this SCC has more than one member */
1758 process_loop(pscc, env);
1763 * Do Tarjan's SCC algorithm and drive load/store optimization.
1765 * @param irn start at this node
1766 * @param env the loop environment
1768 static void dfs(ir_node *irn, loop_env *env)
1771 node_entry *node = get_irn_ne(irn, env);
1773 mark_irn_visited(irn);
1775 node->DFSnum = env->nextDFSnum++;
1776 node->low = node->DFSnum;
1780 if (is_Phi(irn) || is_Sync(irn)) {
1781 n = get_irn_arity(irn);
1782 for (i = 0; i < n; ++i) {
1783 ir_node *pred = get_irn_n(irn, i);
1784 node_entry *o = get_irn_ne(pred, env);
1786 if (irn_not_visited(pred)) {
1788 node->low = MIN(node->low, o->low);
1790 if (o->DFSnum < node->DFSnum && o->in_stack)
1791 node->low = MIN(o->DFSnum, node->low);
1793 } else if (is_fragile_op(irn)) {
1794 ir_node *pred = get_fragile_op_mem(irn);
1795 node_entry *o = get_irn_ne(pred, env);
1797 if (irn_not_visited(pred)) {
1799 node->low = MIN(node->low, o->low);
1801 if (o->DFSnum < node->DFSnum && o->in_stack)
1802 node->low = MIN(o->DFSnum, node->low);
1803 } else if (is_Proj(irn)) {
1804 ir_node *pred = get_Proj_pred(irn);
1805 node_entry *o = get_irn_ne(pred, env);
1807 if (irn_not_visited(pred)) {
1809 node->low = MIN(node->low, o->low);
1811 if (o->DFSnum < node->DFSnum && o->in_stack)
1812 node->low = MIN(o->DFSnum, node->low);
1815 /* IGNORE predecessors */
1818 if (node->low == node->DFSnum) {
1819 scc *pscc = phase_alloc(&env->ph, sizeof(*pscc));
1827 e = get_irn_ne(x, env);
1829 e->next = pscc->head;
1833 process_scc(pscc, env);
1838 * Do the DFS on the memory edges a graph.
1840 * @param irg the graph to process
1841 * @param env the loop environment
1843 static void do_dfs(ir_graph *irg, loop_env *env) {
1844 ir_graph *rem = current_ir_graph;
1845 ir_node *endblk, *end;
1848 current_ir_graph = irg;
1849 inc_irg_visited(irg);
1851 /* visit all memory nodes */
1852 endblk = get_irg_end_block(irg);
1853 for (i = get_Block_n_cfgpreds(endblk) - 1; i >= 0; --i) {
1854 ir_node *pred = get_Block_cfgpred(endblk, i);
1856 pred = skip_Proj(pred);
1857 if (is_Return(pred))
1858 dfs(get_Return_mem(pred), env);
1859 else if (is_Raise(pred))
1860 dfs(get_Raise_mem(pred), env);
1861 else if (is_fragile_op(pred))
1862 dfs(get_fragile_op_mem(pred), env);
1864 assert(0 && "Unknown EndBlock predecessor");
1868 /* visit the keep-alives */
1869 end = get_irg_end(irg);
1870 for (i = get_End_n_keepalives(end) - 1; i >= 0; --i) {
1871 ir_node *ka = get_End_keepalive(end, i);
1873 if (is_Phi(ka) && irn_not_visited(ka))
1876 current_ir_graph = rem;
1880 * Initialize new phase data. We do this always explicit, so return NULL here
1882 static void *init_loop_data(ir_phase *ph, const ir_node *irn, void *data) {
1887 } /* init_loop_data */
1890 * Optimize Loads/Stores in loops.
1892 * @param irg the graph
1894 static int optimize_loops(ir_graph *irg) {
1897 env.stack = NEW_ARR_F(ir_node *, 128);
1902 phase_init(&env.ph, "ldstopt", irg, PHASE_DEFAULT_GROWTH, init_loop_data, NULL);
1904 /* calculate the SCC's and drive loop optimization. */
1907 DEL_ARR_F(env.stack);
1908 phase_free(&env.ph);
1911 } /* optimize_loops */
1914 * do the load store optimization
1916 void optimize_load_store(ir_graph *irg) {
1919 FIRM_DBG_REGISTER(dbg, "firm.opt.ldstopt");
1921 assert(get_irg_phase_state(irg) != phase_building);
1922 assert(get_irg_pinned(irg) != op_pin_state_floats &&
1923 "LoadStore optimization needs pinned graph");
1925 /* we need landing pads */
1926 remove_critical_cf_edges(irg);
1930 /* for Phi optimization post-dominators are needed ... */
1931 assure_postdoms(irg);
1933 if (get_opt_alias_analysis()) {
1934 assure_irg_address_taken_computed(irg);
1935 assure_irp_globals_address_taken_computed();
1938 obstack_init(&env.obst);
1941 /* init the links, then collect Loads/Stores/Proj's in lists */
1943 irg_walk_graph(irg, firm_clear_link, collect_nodes, &env);
1945 /* now we have collected enough information, optimize */
1946 irg_walk_graph(irg, NULL, do_load_store_optimize, &env);
1948 env.changes |= optimize_loops(irg);
1950 obstack_free(&env.obst, NULL);
1952 /* Handle graph state */
1954 set_irg_outs_inconsistent(irg);
1957 if (env.changes & CF_CHANGED) {
1958 /* is this really needed: Yes, control flow changed, block might
1959 have Bad() predecessors. */
1960 set_irg_doms_inconsistent(irg);
1962 } /* optimize_load_store */