2 * Copyright (C) 1995-2008 University of Karlsruhe. All right reserved.
4 * This file is part of libFirm.
6 * This file may be distributed and/or modified under the terms of the
7 * GNU General Public License version 2 as published by the Free Software
8 * Foundation and appearing in the file LICENSE.GPL included in the
9 * packaging of this file.
11 * Licensees holding valid libFirm Professional Edition licenses may use
12 * this file in accordance with the libFirm Commercial License.
13 * Agreement provided with the Software.
15 * This file is provided AS IS with NO WARRANTY OF ANY KIND, INCLUDING THE
16 * WARRANTY OF DESIGN, MERCHANTABILITY AND FITNESS FOR A PARTICULAR
22 * @brief Load/Store optimizations.
23 * @author Michael Beck
30 #include "iroptimize.h"
32 #include "irgraph_t.h"
39 #include "dbginfo_t.h"
40 #include "iropt_dbg.h"
46 #include "opt_polymorphy.h"
48 #include "irphase_t.h"
53 /** The debug handle. */
54 DEBUG_ONLY(static firm_dbg_module_t *dbg;)
57 #include "cacheopt/cachesim.h"
61 #define IMAX(a,b) ((a) > (b) ? (a) : (b))
63 #define MAX_PROJ IMAX(IMAX(pn_Load_max, pn_Store_max), pn_Call_max)
66 DF_CHANGED = 1, /**< data flow changed */
67 CF_CHANGED = 2, /**< control flow changed */
73 typedef struct _walk_env_t {
74 struct obstack obst; /**< list of all stores */
75 unsigned changes; /**< a bitmask of graph changes */
78 /** A Load/Store info. */
79 typedef struct _ldst_info_t {
80 ir_node *projs[MAX_PROJ]; /**< list of Proj's of this node */
81 ir_node *exc_block; /**< the exception block if available */
82 int exc_idx; /**< predecessor index in the exception block */
83 unsigned visited; /**< visited counter for breaking loops */
87 * flags for control flow.
90 BLOCK_HAS_COND = 1, /**< Block has conditional control flow */
91 BLOCK_HAS_EXC = 2 /**< Block has exceptional control flow */
97 typedef struct _block_info_t {
98 unsigned flags; /**< flags for the block */
101 /** the master visited flag for loop detection. */
102 static unsigned master_visited = 0;
104 #define INC_MASTER() ++master_visited
105 #define MARK_NODE(info) (info)->visited = master_visited
106 #define NODE_VISITED(info) (info)->visited >= master_visited
109 * get the Load/Store info of a node
111 static ldst_info_t *get_ldst_info(ir_node *node, struct obstack *obst)
113 ldst_info_t *info = get_irn_link(node);
116 info = OALLOCZ(obst, ldst_info_t);
117 set_irn_link(node, info);
120 } /* get_ldst_info */
123 * get the Block info of a node
125 static block_info_t *get_block_info(ir_node *node, struct obstack *obst)
127 block_info_t *info = get_irn_link(node);
130 info = OALLOCZ(obst, block_info_t);
131 set_irn_link(node, info);
134 } /* get_block_info */
137 * update the projection info for a Load/Store
139 static unsigned update_projs(ldst_info_t *info, ir_node *proj)
141 long nr = get_Proj_proj(proj);
143 assert(0 <= nr && nr <= MAX_PROJ && "Wrong proj from LoadStore");
145 if (info->projs[nr]) {
146 /* there is already one, do CSE */
147 exchange(proj, info->projs[nr]);
151 info->projs[nr] = proj;
157 * update the exception block info for a Load/Store node.
159 * @param info the load/store info struct
160 * @param block the exception handler block for this load/store
161 * @param pos the control flow input of the block
163 static unsigned update_exc(ldst_info_t *info, ir_node *block, int pos)
165 assert(info->exc_block == NULL && "more than one exception block found");
167 info->exc_block = block;
172 /** Return the number of uses of an address node */
173 #define get_irn_n_uses(adr) get_irn_n_edges(adr)
176 * walker, collects all Load/Store/Proj nodes
178 * walks from Start -> End
180 static void collect_nodes(ir_node *node, void *env)
182 ir_opcode opcode = get_irn_opcode(node);
183 ir_node *pred, *blk, *pred_blk;
184 ldst_info_t *ldst_info;
185 walk_env_t *wenv = env;
187 if (opcode == iro_Proj) {
188 pred = get_Proj_pred(node);
189 opcode = get_irn_opcode(pred);
191 if (opcode == iro_Load || opcode == iro_Store || opcode == iro_Call) {
192 ldst_info = get_ldst_info(pred, &wenv->obst);
194 wenv->changes |= update_projs(ldst_info, node);
197 * Place the Proj's to the same block as the
198 * predecessor Load. This is always ok and prevents
199 * "non-SSA" form after optimizations if the Proj
200 * is in a wrong block.
202 blk = get_nodes_block(node);
203 pred_blk = get_nodes_block(pred);
204 if (blk != pred_blk) {
205 wenv->changes |= DF_CHANGED;
206 set_nodes_block(node, pred_blk);
209 } else if (opcode == iro_Block) {
212 for (i = get_Block_n_cfgpreds(node) - 1; i >= 0; --i) {
213 ir_node *pred_block, *proj;
214 block_info_t *bl_info;
217 pred = proj = get_Block_cfgpred(node, i);
220 pred = get_Proj_pred(proj);
221 is_exc = get_Proj_proj(proj) == pn_Generic_X_except;
224 /* ignore Bad predecessors, they will be removed later */
228 pred_block = get_nodes_block(pred);
229 bl_info = get_block_info(pred_block, &wenv->obst);
231 if (is_fragile_op(pred) && is_exc)
232 bl_info->flags |= BLOCK_HAS_EXC;
233 else if (is_irn_forking(pred))
234 bl_info->flags |= BLOCK_HAS_COND;
236 opcode = get_irn_opcode(pred);
237 if (is_exc && (opcode == iro_Load || opcode == iro_Store || opcode == iro_Call)) {
238 ldst_info = get_ldst_info(pred, &wenv->obst);
240 wenv->changes |= update_exc(ldst_info, node, i);
244 } /* collect_nodes */
247 * Returns an entity if the address ptr points to a constant one.
249 * @param ptr the address
251 * @return an entity or NULL
253 static ir_entity *find_constant_entity(ir_node *ptr)
256 if (is_SymConst(ptr) && get_SymConst_kind(ptr) == symconst_addr_ent) {
257 return get_SymConst_entity(ptr);
258 } else if (is_Sel(ptr)) {
259 ir_entity *ent = get_Sel_entity(ptr);
260 ir_type *tp = get_entity_owner(ent);
262 /* Do not fiddle with polymorphism. */
263 if (is_Class_type(get_entity_owner(ent)) &&
264 ((get_entity_n_overwrites(ent) != 0) ||
265 (get_entity_n_overwrittenby(ent) != 0) ) )
268 if (is_Array_type(tp)) {
272 for (i = 0, n = get_Sel_n_indexs(ptr); i < n; ++i) {
274 tarval *tlower, *tupper;
275 ir_node *index = get_Sel_index(ptr, i);
276 tarval *tv = computed_value(index);
278 /* check if the index is constant */
279 if (tv == tarval_bad)
282 bound = get_array_lower_bound(tp, i);
283 tlower = computed_value(bound);
284 bound = get_array_upper_bound(tp, i);
285 tupper = computed_value(bound);
287 if (tlower == tarval_bad || tupper == tarval_bad)
290 if (tarval_cmp(tv, tlower) & pn_Cmp_Lt)
292 if (tarval_cmp(tupper, tv) & pn_Cmp_Lt)
295 /* ok, bounds check finished */
299 if (get_entity_linkage(ent) & IR_LINKAGE_CONSTANT)
303 ptr = get_Sel_ptr(ptr);
304 } else if (is_Add(ptr)) {
305 ir_node *l = get_Add_left(ptr);
306 ir_node *r = get_Add_right(ptr);
308 if (get_irn_mode(l) == get_irn_mode(ptr) && is_Const(r))
310 else if (get_irn_mode(r) == get_irn_mode(ptr) && is_Const(l))
315 /* for now, we support only one addition, reassoc should fold all others */
316 if (! is_SymConst(ptr) && !is_Sel(ptr))
318 } else if (is_Sub(ptr)) {
319 ir_node *l = get_Sub_left(ptr);
320 ir_node *r = get_Sub_right(ptr);
322 if (get_irn_mode(l) == get_irn_mode(ptr) && is_Const(r))
326 /* for now, we support only one substraction, reassoc should fold all others */
327 if (! is_SymConst(ptr) && !is_Sel(ptr))
332 } /* find_constant_entity */
335 * Return the Selection index of a Sel node from dimension n
337 static long get_Sel_array_index_long(ir_node *n, int dim)
339 ir_node *index = get_Sel_index(n, dim);
340 assert(is_Const(index));
341 return get_tarval_long(get_Const_tarval(index));
342 } /* get_Sel_array_index_long */
345 * Returns the accessed component graph path for an
346 * node computing an address.
348 * @param ptr the node computing the address
349 * @param depth current depth in steps upward from the root
352 static compound_graph_path *rec_get_accessed_path(ir_node *ptr, int depth)
354 compound_graph_path *res = NULL;
355 ir_entity *root, *field, *ent;
356 int path_len, pos, idx;
360 if (is_SymConst(ptr)) {
361 /* a SymConst. If the depth is 0, this is an access to a global
362 * entity and we don't need a component path, else we know
363 * at least its length.
365 assert(get_SymConst_kind(ptr) == symconst_addr_ent);
366 root = get_SymConst_entity(ptr);
367 res = (depth == 0) ? NULL : new_compound_graph_path(get_entity_type(root), depth);
368 } else if (is_Sel(ptr)) {
369 /* it's a Sel, go up until we find the root */
370 res = rec_get_accessed_path(get_Sel_ptr(ptr), depth+1);
374 /* fill up the step in the path at the current position */
375 field = get_Sel_entity(ptr);
376 path_len = get_compound_graph_path_length(res);
377 pos = path_len - depth - 1;
378 set_compound_graph_path_node(res, pos, field);
380 if (is_Array_type(get_entity_owner(field))) {
381 assert(get_Sel_n_indexs(ptr) == 1 && "multi dim arrays not implemented");
382 set_compound_graph_path_array_index(res, pos, get_Sel_array_index_long(ptr, 0));
384 } else if (is_Add(ptr)) {
385 ir_node *l = get_Add_left(ptr);
386 ir_node *r = get_Add_right(ptr);
387 ir_mode *mode = get_irn_mode(ptr);
390 if (is_Const(r) && get_irn_mode(l) == mode) {
392 tv = get_Const_tarval(r);
395 tv = get_Const_tarval(l);
398 mode = get_tarval_mode(tv);
401 /* ptr must be a Sel or a SymConst, this was checked in find_constant_entity() */
403 field = get_Sel_entity(ptr);
405 field = get_SymConst_entity(ptr);
408 for (ent = field;;) {
410 tarval *sz, *tv_index, *tlower, *tupper;
413 tp = get_entity_type(ent);
414 if (! is_Array_type(tp))
416 ent = get_array_element_entity(tp);
417 size = get_type_size_bytes(get_entity_type(ent));
418 sz = new_tarval_from_long(size, mode);
420 tv_index = tarval_div(tmp, sz);
421 tmp = tarval_mod(tmp, sz);
423 if (tv_index == tarval_bad || tmp == tarval_bad)
426 assert(get_array_n_dimensions(tp) == 1 && "multiarrays not implemented");
427 bound = get_array_lower_bound(tp, 0);
428 tlower = computed_value(bound);
429 bound = get_array_upper_bound(tp, 0);
430 tupper = computed_value(bound);
432 if (tlower == tarval_bad || tupper == tarval_bad)
435 if (tarval_cmp(tv_index, tlower) & pn_Cmp_Lt)
437 if (tarval_cmp(tupper, tv_index) & pn_Cmp_Lt)
440 /* ok, bounds check finished */
443 if (! tarval_is_null(tmp)) {
444 /* access to some struct/union member */
448 /* should be at least ONE array */
452 res = rec_get_accessed_path(ptr, depth + idx);
456 path_len = get_compound_graph_path_length(res);
457 pos = path_len - depth - idx;
459 for (ent = field;;) {
461 tarval *sz, *tv_index;
464 tp = get_entity_type(ent);
465 if (! is_Array_type(tp))
467 ent = get_array_element_entity(tp);
468 set_compound_graph_path_node(res, pos, ent);
470 size = get_type_size_bytes(get_entity_type(ent));
471 sz = new_tarval_from_long(size, mode);
473 tv_index = tarval_div(tv, sz);
474 tv = tarval_mod(tv, sz);
476 /* worked above, should work again */
477 assert(tv_index != tarval_bad && tv != tarval_bad);
479 /* bounds already checked above */
480 index = get_tarval_long(tv_index);
481 set_compound_graph_path_array_index(res, pos, index);
484 } else if (is_Sub(ptr)) {
485 ir_node *l = get_Sub_left(ptr);
486 ir_node *r = get_Sub_right(ptr);
489 tv = get_Const_tarval(r);
494 } /* rec_get_accessed_path */
497 * Returns an access path or NULL. The access path is only
498 * valid, if the graph is in phase_high and _no_ address computation is used.
500 static compound_graph_path *get_accessed_path(ir_node *ptr)
502 compound_graph_path *gr = rec_get_accessed_path(ptr, 0);
504 } /* get_accessed_path */
506 typedef struct path_entry {
508 struct path_entry *next;
512 static ir_node *rec_find_compound_ent_value(ir_node *ptr, path_entry *next)
514 path_entry entry, *p;
515 ir_entity *ent, *field;
516 ir_initializer_t *initializer;
522 if (is_SymConst(ptr)) {
524 ent = get_SymConst_entity(ptr);
525 initializer = get_entity_initializer(ent);
526 for (p = next; p != NULL;) {
527 if (initializer->kind != IR_INITIALIZER_COMPOUND)
529 n = get_initializer_compound_n_entries(initializer);
530 tp = get_entity_type(ent);
532 if (is_Array_type(tp)) {
533 ent = get_array_element_entity(tp);
538 initializer = get_initializer_compound_value(initializer, 0);
542 if (p->index >= (int) n)
544 initializer = get_initializer_compound_value(initializer, p->index);
549 tp = get_entity_type(ent);
550 while (is_Array_type(tp)) {
551 ent = get_array_element_entity(tp);
552 tp = get_entity_type(ent);
554 n = get_initializer_compound_n_entries(initializer);
557 initializer = get_initializer_compound_value(initializer, 0);
560 switch (initializer->kind) {
561 case IR_INITIALIZER_CONST:
562 return get_initializer_const_value(initializer);
563 case IR_INITIALIZER_TARVAL:
564 case IR_INITIALIZER_NULL:
568 } else if (is_Sel(ptr)) {
569 entry.ent = field = get_Sel_entity(ptr);
570 tp = get_entity_owner(field);
571 if (is_Array_type(tp)) {
572 assert(get_Sel_n_indexs(ptr) == 1 && "multi dim arrays not implemented");
573 entry.index = get_Sel_array_index_long(ptr, 0) - get_array_lower_bound_int(tp, 0);
575 int i, n_members = get_compound_n_members(tp);
576 for (i = 0; i < n_members; ++i) {
577 if (get_compound_member(tp, i) == field)
580 if (i >= n_members) {
581 /* not found: should NOT happen */
586 return rec_find_compound_ent_value(get_Sel_ptr(ptr), &entry);
587 } else if (is_Add(ptr)) {
588 ir_node *l = get_Add_left(ptr);
589 ir_node *r = get_Add_right(ptr);
595 tv = get_Const_tarval(r);
598 tv = get_Const_tarval(l);
601 mode = get_tarval_mode(tv);
603 /* ptr must be a Sel or a SymConst, this was checked in find_constant_entity() */
605 field = get_Sel_entity(ptr);
607 field = get_SymConst_entity(ptr);
610 /* count needed entries */
612 for (ent = field;;) {
613 tp = get_entity_type(ent);
614 if (! is_Array_type(tp))
616 ent = get_array_element_entity(tp);
619 /* should be at least ONE entry */
623 /* allocate the right number of entries */
624 NEW_ARR_A(path_entry, p, pos);
628 for (ent = field;;) {
630 tarval *sz, *tv_index, *tlower, *tupper;
634 tp = get_entity_type(ent);
635 if (! is_Array_type(tp))
637 ent = get_array_element_entity(tp);
639 p[pos].next = &p[pos + 1];
641 size = get_type_size_bytes(get_entity_type(ent));
642 sz = new_tarval_from_long(size, mode);
644 tv_index = tarval_div(tv, sz);
645 tv = tarval_mod(tv, sz);
647 if (tv_index == tarval_bad || tv == tarval_bad)
650 assert(get_array_n_dimensions(tp) == 1 && "multiarrays not implemented");
651 bound = get_array_lower_bound(tp, 0);
652 tlower = computed_value(bound);
653 bound = get_array_upper_bound(tp, 0);
654 tupper = computed_value(bound);
656 if (tlower == tarval_bad || tupper == tarval_bad)
659 if (tarval_cmp(tv_index, tlower) & pn_Cmp_Lt)
661 if (tarval_cmp(tupper, tv_index) & pn_Cmp_Lt)
664 /* ok, bounds check finished */
665 index = get_tarval_long(tv_index);
666 p[pos].index = index;
669 if (! tarval_is_null(tv)) {
670 /* hmm, wrong access */
673 p[pos - 1].next = next;
674 return rec_find_compound_ent_value(ptr, p);
675 } else if (is_Sub(ptr)) {
676 ir_node *l = get_Sub_left(ptr);
677 ir_node *r = get_Sub_right(ptr);
680 tv = get_Const_tarval(r);
687 static ir_node *find_compound_ent_value(ir_node *ptr)
689 return rec_find_compound_ent_value(ptr, NULL);
693 static void reduce_adr_usage(ir_node *ptr);
696 * Update a Load that may have lost its users.
698 static void handle_load_update(ir_node *load)
700 ldst_info_t *info = get_irn_link(load);
702 /* do NOT touch volatile loads for now */
703 if (get_Load_volatility(load) == volatility_is_volatile)
706 if (! info->projs[pn_Load_res] && ! info->projs[pn_Load_X_except]) {
707 ir_node *ptr = get_Load_ptr(load);
708 ir_node *mem = get_Load_mem(load);
710 /* a Load whose value is neither used nor exception checked, remove it */
711 exchange(info->projs[pn_Load_M], mem);
712 if (info->projs[pn_Load_X_regular])
713 exchange(info->projs[pn_Load_X_regular], new_r_Jmp(get_nodes_block(load)));
715 reduce_adr_usage(ptr);
717 } /* handle_load_update */
720 * A use of an address node has vanished. Check if this was a Proj
721 * node and update the counters.
723 static void reduce_adr_usage(ir_node *ptr)
728 if (get_irn_n_edges(ptr) > 0)
731 /* this Proj is dead now */
732 pred = get_Proj_pred(ptr);
734 ldst_info_t *info = get_irn_link(pred);
735 info->projs[get_Proj_proj(ptr)] = NULL;
737 /* this node lost its result proj, handle that */
738 handle_load_update(pred);
740 } /* reduce_adr_usage */
743 * Check, if an already existing value of mode old_mode can be converted
744 * into the needed one new_mode without loss.
746 static int can_use_stored_value(ir_mode *old_mode, ir_mode *new_mode)
748 if (old_mode == new_mode)
751 /* if both modes are two-complement ones, we can always convert the
752 Stored value into the needed one. */
753 if (get_mode_size_bits(old_mode) >= get_mode_size_bits(new_mode) &&
754 get_mode_arithmetic(old_mode) == irma_twos_complement &&
755 get_mode_arithmetic(new_mode) == irma_twos_complement)
758 } /* can_use_stored_value */
761 * Check whether a Call is at least pure, ie. does only read memory.
763 static unsigned is_Call_pure(ir_node *call)
765 ir_type *call_tp = get_Call_type(call);
766 unsigned prop = get_method_additional_properties(call_tp);
768 /* check first the call type */
769 if ((prop & (mtp_property_const|mtp_property_pure)) == 0) {
770 /* try the called entity */
771 ir_node *ptr = get_Call_ptr(call);
773 if (is_Global(ptr)) {
774 ir_entity *ent = get_Global_entity(ptr);
776 prop = get_entity_additional_properties(ent);
779 return (prop & (mtp_property_const|mtp_property_pure)) != 0;
782 static ir_node *get_base_and_offset(ir_node *ptr, long *pOffset)
784 ir_mode *mode = get_irn_mode(ptr);
787 /* TODO: long might not be enough, we should probably use some tarval thingy... */
790 ir_node *l = get_Add_left(ptr);
791 ir_node *r = get_Add_right(ptr);
793 if (get_irn_mode(l) != mode || !is_Const(r))
796 offset += get_tarval_long(get_Const_tarval(r));
798 } else if (is_Sub(ptr)) {
799 ir_node *l = get_Sub_left(ptr);
800 ir_node *r = get_Sub_right(ptr);
802 if (get_irn_mode(l) != mode || !is_Const(r))
805 offset -= get_tarval_long(get_Const_tarval(r));
807 } else if (is_Sel(ptr)) {
808 ir_entity *ent = get_Sel_entity(ptr);
809 ir_type *tp = get_entity_owner(ent);
811 if (is_Array_type(tp)) {
815 /* only one dimensional arrays yet */
816 if (get_Sel_n_indexs(ptr) != 1)
818 index = get_Sel_index(ptr, 0);
819 if (! is_Const(index))
822 tp = get_entity_type(ent);
823 if (get_type_state(tp) != layout_fixed)
826 size = get_type_size_bytes(tp);
827 offset += size * get_tarval_long(get_Const_tarval(index));
829 if (get_type_state(tp) != layout_fixed)
831 offset += get_entity_offset(ent);
833 ptr = get_Sel_ptr(ptr);
842 static int try_load_after_store(ir_node *load,
843 ir_node *load_base_ptr, long load_offset, ir_node *store)
846 ir_node *store_ptr = get_Store_ptr(store);
848 ir_node *store_base_ptr = get_base_and_offset(store_ptr, &store_offset);
849 ir_node *store_value;
858 if (load_base_ptr != store_base_ptr)
861 load_mode = get_Load_mode(load);
862 load_mode_len = get_mode_size_bytes(load_mode);
863 store_mode = get_irn_mode(get_Store_value(store));
864 store_mode_len = get_mode_size_bytes(store_mode);
865 delta = load_offset - store_offset;
866 store_value = get_Store_value(store);
868 if (delta != 0 || store_mode != load_mode) {
869 if (delta < 0 || delta + load_mode_len > store_mode_len)
872 if (get_mode_arithmetic(store_mode) != irma_twos_complement ||
873 get_mode_arithmetic(load_mode) != irma_twos_complement)
877 /* produce a shift to adjust offset delta */
881 /* FIXME: only true for little endian */
882 cnst = new_Const_long(mode_Iu, delta * 8);
883 store_value = new_r_Shr(get_nodes_block(load),
884 store_value, cnst, store_mode);
887 /* add an convert if needed */
888 if (store_mode != load_mode) {
889 store_value = new_r_Conv(get_nodes_block(load), store_value, load_mode);
893 DBG_OPT_RAW(load, store_value);
895 info = get_irn_link(load);
896 if (info->projs[pn_Load_M])
897 exchange(info->projs[pn_Load_M], get_Load_mem(load));
901 if (info->projs[pn_Load_X_except]) {
902 exchange( info->projs[pn_Load_X_except], new_Bad());
905 if (info->projs[pn_Load_X_regular]) {
906 exchange( info->projs[pn_Load_X_regular], new_r_Jmp(get_nodes_block(load)));
910 if (info->projs[pn_Load_res])
911 exchange(info->projs[pn_Load_res], store_value);
913 load_ptr = get_Load_ptr(load);
915 reduce_adr_usage(load_ptr);
916 return res | DF_CHANGED;
920 * Follow the memory chain as long as there are only Loads,
921 * alias free Stores, and constant Calls and try to replace the
922 * current Load by a previous ones.
923 * Note that in unreachable loops it might happen that we reach
924 * load again, as well as we can fall into a cycle.
925 * We break such cycles using a special visited flag.
927 * INC_MASTER() must be called before dive into
929 static unsigned follow_Mem_chain(ir_node *load, ir_node *curr)
932 ldst_info_t *info = get_irn_link(load);
934 ir_node *ptr = get_Load_ptr(load);
935 ir_node *mem = get_Load_mem(load);
936 ir_mode *load_mode = get_Load_mode(load);
938 for (pred = curr; load != pred; ) {
939 ldst_info_t *pred_info = get_irn_link(pred);
942 * a Load immediately after a Store -- a read after write.
943 * We may remove the Load, if both Load & Store does not have an
944 * exception handler OR they are in the same MacroBlock. In the latter
945 * case the Load cannot throw an exception when the previous Store was
948 * Why we need to check for Store Exception? If the Store cannot
949 * be executed (ROM) the exception handler might simply jump into
950 * the load MacroBlock :-(
951 * We could make it a little bit better if we would know that the
952 * exception handler of the Store jumps directly to the end...
954 if (is_Store(pred) && ((pred_info->projs[pn_Store_X_except] == NULL
955 && info->projs[pn_Load_X_except] == NULL)
956 || get_nodes_MacroBlock(load) == get_nodes_MacroBlock(pred)))
959 ir_node *base_ptr = get_base_and_offset(ptr, &load_offset);
960 int changes = try_load_after_store(load, base_ptr, load_offset, pred);
963 return res | changes;
964 } else if (is_Load(pred) && get_Load_ptr(pred) == ptr &&
965 can_use_stored_value(get_Load_mode(pred), load_mode)) {
967 * a Load after a Load -- a read after read.
968 * We may remove the second Load, if it does not have an exception handler
969 * OR they are in the same MacroBlock. In the later case the Load cannot
970 * throw an exception when the previous Load was quiet.
972 * Here, there is no need to check if the previous Load has an exception
973 * hander because they would have exact the same exception...
975 if (info->projs[pn_Load_X_except] == NULL || get_nodes_MacroBlock(load) == get_nodes_MacroBlock(pred)) {
978 DBG_OPT_RAR(load, pred);
980 /* the result is used */
981 if (info->projs[pn_Load_res]) {
982 if (pred_info->projs[pn_Load_res] == NULL) {
983 /* create a new Proj again */
984 pred_info->projs[pn_Load_res] = new_r_Proj(pred, get_Load_mode(pred), pn_Load_res);
986 value = pred_info->projs[pn_Load_res];
988 /* add an convert if needed */
989 if (get_Load_mode(pred) != load_mode) {
990 value = new_r_Conv(get_nodes_block(load), value, load_mode);
993 exchange(info->projs[pn_Load_res], value);
996 if (info->projs[pn_Load_M])
997 exchange(info->projs[pn_Load_M], mem);
1000 if (info->projs[pn_Load_X_except]) {
1001 exchange(info->projs[pn_Load_X_except], new_Bad());
1004 if (info->projs[pn_Load_X_regular]) {
1005 exchange( info->projs[pn_Load_X_regular], new_r_Jmp(get_nodes_block(load)));
1010 reduce_adr_usage(ptr);
1011 return res |= DF_CHANGED;
1015 if (is_Store(pred)) {
1016 /* check if we can pass through this store */
1017 ir_alias_relation rel = get_alias_relation(
1019 get_Store_ptr(pred),
1020 get_irn_mode(get_Store_value(pred)),
1022 /* if the might be an alias, we cannot pass this Store */
1023 if (rel != ir_no_alias)
1025 pred = skip_Proj(get_Store_mem(pred));
1026 } else if (is_Load(pred)) {
1027 pred = skip_Proj(get_Load_mem(pred));
1028 } else if (is_Call(pred)) {
1029 if (is_Call_pure(pred)) {
1030 /* The called graph is at least pure, so there are no Store's
1031 in it. We can handle it like a Load and skip it. */
1032 pred = skip_Proj(get_Call_mem(pred));
1034 /* there might be Store's in the graph, stop here */
1038 /* follow only Load chains */
1042 /* check for cycles */
1043 if (NODE_VISITED(pred_info))
1045 MARK_NODE(pred_info);
1048 if (is_Sync(pred)) {
1051 /* handle all Sync predecessors */
1052 for (i = get_Sync_n_preds(pred) - 1; i >= 0; --i) {
1053 res |= follow_Mem_chain(load, skip_Proj(get_Sync_pred(pred, i)));
1060 } /* follow_Mem_chain */
1063 * Check if we can replace the load by a given const from
1064 * the const code irg.
1066 ir_node *can_replace_load_by_const(const ir_node *load, ir_node *c)
1068 ir_mode *c_mode = get_irn_mode(c);
1069 ir_mode *l_mode = get_Load_mode(load);
1070 ir_node *res = NULL;
1072 if (c_mode != l_mode) {
1073 /* check, if the mode matches OR can be easily converted info */
1074 if (is_reinterpret_cast(c_mode, l_mode)) {
1075 /* we can safely cast */
1076 dbg_info *dbg = get_irn_dbg_info(load);
1077 ir_node *block = get_nodes_block(load);
1079 /* copy the value from the const code irg and cast it */
1080 res = copy_const_value(dbg, c);
1081 res = new_rd_Conv(dbg, block, res, l_mode);
1084 /* copy the value from the const code irg */
1085 res = copy_const_value(get_irn_dbg_info(load), c);
1088 } /* can_replace_load_by_const */
1093 * @param load the Load node
1095 static unsigned optimize_load(ir_node *load)
1097 ldst_info_t *info = get_irn_link(load);
1098 ir_node *mem, *ptr, *value;
1103 /* do NOT touch volatile loads for now */
1104 if (get_Load_volatility(load) == volatility_is_volatile)
1107 /* the address of the load to be optimized */
1108 ptr = get_Load_ptr(load);
1110 /* The mem of the Load. Must still be returned after optimization. */
1111 mem = get_Load_mem(load);
1113 if (info->projs[pn_Load_res] == NULL
1114 && info->projs[pn_Load_X_except] == NULL) {
1115 /* the value is never used and we don't care about exceptions, remove */
1116 exchange(info->projs[pn_Load_M], mem);
1118 if (info->projs[pn_Load_X_regular]) {
1119 /* should not happen, but if it does, remove it */
1120 exchange(info->projs[pn_Load_X_regular], new_r_Jmp(get_nodes_block(load)));
1124 reduce_adr_usage(ptr);
1125 return res | DF_CHANGED;
1128 /* Load from a constant polymorphic field, where we can resolve
1130 value = transform_polymorph_Load(load);
1131 if (value == load) {
1133 /* check if we can determine the entity that will be loaded */
1134 ent = find_constant_entity(ptr);
1136 && get_entity_visibility(ent) != ir_visibility_external) {
1137 /* a static allocation that is not external: there should be NO
1138 * exception when loading even if we cannot replace the load itself.
1141 /* no exception, clear the info field as it might be checked later again */
1142 if (info->projs[pn_Load_X_except]) {
1143 exchange(info->projs[pn_Load_X_except], new_Bad());
1144 info->projs[pn_Load_X_except] = NULL;
1147 if (info->projs[pn_Load_X_regular]) {
1148 exchange(info->projs[pn_Load_X_regular], new_r_Jmp(get_nodes_block(load)));
1149 info->projs[pn_Load_X_regular] = NULL;
1153 if (get_entity_linkage(ent) & IR_LINKAGE_CONSTANT) {
1154 if (ent->initializer != NULL) {
1155 /* new style initializer */
1156 value = find_compound_ent_value(ptr);
1157 } else if (entity_has_compound_ent_values(ent)) {
1158 /* old style initializer */
1159 compound_graph_path *path = get_accessed_path(ptr);
1162 assert(is_proper_compound_graph_path(path, get_compound_graph_path_length(path)-1));
1164 value = get_compound_ent_value_by_path(ent, path);
1165 DB((dbg, LEVEL_1, " Constant access at %F%F resulted in %+F\n", ent, path, value));
1166 free_compound_graph_path(path);
1170 value = can_replace_load_by_const(load, value);
1174 if (value != NULL) {
1175 /* we completely replace the load by this value */
1176 if (info->projs[pn_Load_X_except]) {
1177 exchange(info->projs[pn_Load_X_except], new_Bad());
1178 info->projs[pn_Load_X_except] = NULL;
1181 if (info->projs[pn_Load_X_regular]) {
1182 exchange(info->projs[pn_Load_X_regular], new_r_Jmp(get_nodes_block(load)));
1183 info->projs[pn_Load_X_regular] = NULL;
1186 if (info->projs[pn_Load_M]) {
1187 exchange(info->projs[pn_Load_M], mem);
1190 if (info->projs[pn_Load_res]) {
1191 exchange(info->projs[pn_Load_res], value);
1195 reduce_adr_usage(ptr);
1199 /* Check, if the address of this load is used more than once.
1200 * If not, more load cannot be removed in any case. */
1201 if (get_irn_n_uses(ptr) <= 1 && get_irn_n_uses(get_base_and_offset(ptr, &dummy)) <= 1)
1205 * follow the memory chain as long as there are only Loads
1206 * and try to replace current Load or Store by a previous one.
1207 * Note that in unreachable loops it might happen that we reach
1208 * load again, as well as we can fall into a cycle.
1209 * We break such cycles using a special visited flag.
1212 res = follow_Mem_chain(load, skip_Proj(mem));
1214 } /* optimize_load */
1217 * Check whether a value of mode new_mode would completely overwrite a value
1218 * of mode old_mode in memory.
1220 static int is_completely_overwritten(ir_mode *old_mode, ir_mode *new_mode)
1222 return get_mode_size_bits(new_mode) >= get_mode_size_bits(old_mode);
1223 } /* is_completely_overwritten */
1226 * Check whether small is a part of large (starting at same address).
1228 static int is_partially_same(ir_node *small, ir_node *large)
1230 ir_mode *sm = get_irn_mode(small);
1231 ir_mode *lm = get_irn_mode(large);
1233 /* FIXME: Check endianness */
1234 return is_Conv(small) && get_Conv_op(small) == large
1235 && get_mode_size_bytes(sm) < get_mode_size_bytes(lm)
1236 && get_mode_arithmetic(sm) == irma_twos_complement
1237 && get_mode_arithmetic(lm) == irma_twos_complement;
1238 } /* is_partially_same */
1241 * follow the memory chain as long as there are only Loads and alias free Stores.
1243 * INC_MASTER() must be called before dive into
1245 static unsigned follow_Mem_chain_for_Store(ir_node *store, ir_node *curr)
1248 ldst_info_t *info = get_irn_link(store);
1250 ir_node *ptr = get_Store_ptr(store);
1251 ir_node *mem = get_Store_mem(store);
1252 ir_node *value = get_Store_value(store);
1253 ir_mode *mode = get_irn_mode(value);
1254 ir_node *block = get_nodes_block(store);
1255 ir_node *mblk = get_Block_MacroBlock(block);
1257 for (pred = curr; pred != store;) {
1258 ldst_info_t *pred_info = get_irn_link(pred);
1261 * BEWARE: one might think that checking the modes is useless, because
1262 * if the pointers are identical, they refer to the same object.
1263 * This is only true in strong typed languages, not is C were the following
1264 * is possible *(ir_type1 *)p = a; *(ir_type2 *)p = b ...
1265 * However, if the size of the mode that is written is bigger or equal the
1266 * size of the old one, the old value is completely overwritten and can be
1269 if (is_Store(pred) && get_Store_ptr(pred) == ptr &&
1270 get_nodes_MacroBlock(pred) == mblk) {
1272 * a Store after a Store in the same MacroBlock -- a write after write.
1276 * We may remove the first Store, if the old value is completely
1277 * overwritten or the old value is a part of the new value,
1278 * and if it does not have an exception handler.
1280 * TODO: What, if both have the same exception handler ???
1282 if (get_Store_volatility(pred) != volatility_is_volatile
1283 && !pred_info->projs[pn_Store_X_except]) {
1284 ir_node *predvalue = get_Store_value(pred);
1285 ir_mode *predmode = get_irn_mode(predvalue);
1287 if (is_completely_overwritten(predmode, mode)
1288 || is_partially_same(predvalue, value)) {
1289 DBG_OPT_WAW(pred, store);
1290 exchange(pred_info->projs[pn_Store_M], get_Store_mem(pred));
1292 reduce_adr_usage(ptr);
1298 * We may remove the Store, if the old value already contains
1299 * the new value, and if it does not have an exception handler.
1301 * TODO: What, if both have the same exception handler ???
1303 if (get_Store_volatility(store) != volatility_is_volatile
1304 && !info->projs[pn_Store_X_except]) {
1305 ir_node *predvalue = get_Store_value(pred);
1307 if (is_partially_same(value, predvalue)) {
1308 DBG_OPT_WAW(pred, store);
1309 exchange(info->projs[pn_Store_M], mem);
1311 reduce_adr_usage(ptr);
1315 } else if (is_Load(pred) && get_Load_ptr(pred) == ptr &&
1316 value == pred_info->projs[pn_Load_res]) {
1318 * a Store of a value just loaded from the same address
1319 * -- a write after read.
1320 * We may remove the Store, if it does not have an exception
1323 if (! info->projs[pn_Store_X_except]) {
1324 DBG_OPT_WAR(store, pred);
1325 exchange(info->projs[pn_Store_M], mem);
1327 reduce_adr_usage(ptr);
1332 if (is_Store(pred)) {
1333 /* check if we can pass through this store */
1334 ir_alias_relation rel = get_alias_relation(
1336 get_Store_ptr(pred),
1337 get_irn_mode(get_Store_value(pred)),
1339 /* if the might be an alias, we cannot pass this Store */
1340 if (rel != ir_no_alias)
1342 pred = skip_Proj(get_Store_mem(pred));
1343 } else if (is_Load(pred)) {
1344 ir_alias_relation rel = get_alias_relation(
1345 current_ir_graph, get_Load_ptr(pred), get_Load_mode(pred),
1347 if (rel != ir_no_alias)
1350 pred = skip_Proj(get_Load_mem(pred));
1352 /* follow only Load chains */
1356 /* check for cycles */
1357 if (NODE_VISITED(pred_info))
1359 MARK_NODE(pred_info);
1362 if (is_Sync(pred)) {
1365 /* handle all Sync predecessors */
1366 for (i = get_Sync_n_preds(pred) - 1; i >= 0; --i) {
1367 res |= follow_Mem_chain_for_Store(store, skip_Proj(get_Sync_pred(pred, i)));
1373 } /* follow_Mem_chain_for_Store */
1375 /** find entity used as base for an address calculation */
1376 static ir_entity *find_entity(ir_node *ptr)
1378 switch (get_irn_opcode(ptr)) {
1380 return get_SymConst_entity(ptr);
1382 ir_node *pred = get_Sel_ptr(ptr);
1383 if (get_irg_frame(get_irn_irg(ptr)) == pred)
1384 return get_Sel_entity(ptr);
1386 return find_entity(pred);
1390 ir_node *left = get_binop_left(ptr);
1392 if (mode_is_reference(get_irn_mode(left)))
1393 return find_entity(left);
1394 right = get_binop_right(ptr);
1395 if (mode_is_reference(get_irn_mode(right)))
1396 return find_entity(right);
1407 * @param store the Store node
1409 static unsigned optimize_store(ir_node *store)
1415 if (get_Store_volatility(store) == volatility_is_volatile)
1418 ptr = get_Store_ptr(store);
1419 entity = find_entity(ptr);
1421 /* a store to an entity which is never read is unnecessary */
1422 if (entity != NULL && !(get_entity_usage(entity) & ir_usage_read)) {
1423 ldst_info_t *info = get_irn_link(store);
1424 if (info->projs[pn_Store_X_except] == NULL) {
1425 DB((dbg, LEVEL_1, " Killing useless %+F to never read entity %+F\n", store, entity));
1426 exchange(info->projs[pn_Store_M], get_Store_mem(store));
1428 reduce_adr_usage(ptr);
1433 /* Check, if the address of this Store is used more than once.
1434 * If not, this Store cannot be removed in any case. */
1435 if (get_irn_n_uses(ptr) <= 1)
1438 mem = get_Store_mem(store);
1440 /* follow the memory chain as long as there are only Loads */
1443 return follow_Mem_chain_for_Store(store, skip_Proj(mem));
1444 } /* optimize_store */
1447 * walker, optimizes Phi after Stores to identical places:
1448 * Does the following optimization:
1451 * val1 val2 val3 val1 val2 val3
1453 * Store Store Store \ | /
1460 * This reduces the number of stores and allows for predicated execution.
1461 * Moves Stores back to the end of a function which may be bad.
1463 * This is only possible if the predecessor blocks have only one successor.
1465 static unsigned optimize_phi(ir_node *phi, walk_env_t *wenv)
1468 ir_node *store, *old_store, *ptr, *block, *phi_block, *phiM, *phiD, *exc, *projM;
1470 ir_node **inM, **inD, **projMs;
1472 dbg_info *db = NULL;
1474 block_info_t *bl_info;
1477 /* Must be a memory Phi */
1478 if (get_irn_mode(phi) != mode_M)
1481 n = get_Phi_n_preds(phi);
1485 /* must be only one user */
1486 projM = get_Phi_pred(phi, 0);
1487 if (get_irn_n_edges(projM) != 1)
1490 store = skip_Proj(projM);
1492 if (!is_Store(store))
1495 block = get_nodes_block(store);
1497 /* abort on dead blocks */
1498 if (is_Block_dead(block))
1501 /* check if the block is post dominated by Phi-block
1502 and has no exception exit */
1503 bl_info = get_irn_link(block);
1504 if (bl_info->flags & BLOCK_HAS_EXC)
1507 phi_block = get_nodes_block(phi);
1508 if (! block_strictly_postdominates(phi_block, block))
1511 /* this is the address of the store */
1512 ptr = get_Store_ptr(store);
1513 mode = get_irn_mode(get_Store_value(store));
1514 info = get_irn_link(store);
1515 exc = info->exc_block;
1517 for (i = 1; i < n; ++i) {
1518 ir_node *pred = get_Phi_pred(phi, i);
1520 if (get_irn_n_edges(pred) != 1)
1523 pred = skip_Proj(pred);
1524 if (!is_Store(pred))
1527 if (ptr != get_Store_ptr(pred) || mode != get_irn_mode(get_Store_value(pred)))
1530 info = get_irn_link(pred);
1532 /* check, if all stores have the same exception flow */
1533 if (exc != info->exc_block)
1536 /* abort on dead blocks */
1537 block = get_nodes_block(pred);
1538 if (is_Block_dead(block))
1541 /* check if the block is post dominated by Phi-block
1542 and has no exception exit. Note that block must be different from
1543 Phi-block, else we would move a Store from end End of a block to its
1545 bl_info = get_irn_link(block);
1546 if (bl_info->flags & BLOCK_HAS_EXC)
1548 if (block == phi_block || ! block_postdominates(phi_block, block))
1553 * ok, when we are here, we found all predecessors of a Phi that
1554 * are Stores to the same address and size. That means whatever
1555 * we do before we enter the block of the Phi, we do a Store.
1556 * So, we can move the Store to the current block:
1558 * val1 val2 val3 val1 val2 val3
1560 * | Str | | Str | | Str | \ | /
1566 * Is only allowed if the predecessor blocks have only one successor.
1569 NEW_ARR_A(ir_node *, projMs, n);
1570 NEW_ARR_A(ir_node *, inM, n);
1571 NEW_ARR_A(ir_node *, inD, n);
1572 NEW_ARR_A(int, idx, n);
1574 /* Prepare: Collect all Store nodes. We must do this
1575 first because we otherwise may loose a store when exchanging its
1578 for (i = n - 1; i >= 0; --i) {
1581 projMs[i] = get_Phi_pred(phi, i);
1582 assert(is_Proj(projMs[i]));
1584 store = get_Proj_pred(projMs[i]);
1585 info = get_irn_link(store);
1587 inM[i] = get_Store_mem(store);
1588 inD[i] = get_Store_value(store);
1589 idx[i] = info->exc_idx;
1591 block = get_nodes_block(phi);
1593 /* second step: create a new memory Phi */
1594 phiM = new_rd_Phi(get_irn_dbg_info(phi), block, n, inM, mode_M);
1596 /* third step: create a new data Phi */
1597 phiD = new_rd_Phi(get_irn_dbg_info(phi), block, n, inD, mode);
1599 /* rewire memory and kill the node */
1600 for (i = n - 1; i >= 0; --i) {
1601 ir_node *proj = projMs[i];
1603 if (is_Proj(proj)) {
1604 ir_node *store = get_Proj_pred(proj);
1605 exchange(proj, inM[i]);
1610 /* fourth step: create the Store */
1611 store = new_rd_Store(db, block, phiM, ptr, phiD, 0);
1613 co_set_irn_name(store, co_get_irn_ident(old_store));
1616 projM = new_rd_Proj(NULL, store, mode_M, pn_Store_M);
1618 info = get_ldst_info(store, &wenv->obst);
1619 info->projs[pn_Store_M] = projM;
1621 /* fifths step: repair exception flow */
1623 ir_node *projX = new_rd_Proj(NULL, store, mode_X, pn_Store_X_except);
1625 info->projs[pn_Store_X_except] = projX;
1626 info->exc_block = exc;
1627 info->exc_idx = idx[0];
1629 for (i = 0; i < n; ++i) {
1630 set_Block_cfgpred(exc, idx[i], projX);
1634 /* the exception block should be optimized as some inputs are identical now */
1640 /* sixth step: replace old Phi */
1641 exchange(phi, projM);
1643 return res | DF_CHANGED;
1644 } /* optimize_phi */
1647 * walker, do the optimizations
1649 static void do_load_store_optimize(ir_node *n, void *env)
1651 walk_env_t *wenv = env;
1653 switch (get_irn_opcode(n)) {
1656 wenv->changes |= optimize_load(n);
1660 wenv->changes |= optimize_store(n);
1664 wenv->changes |= optimize_phi(n, wenv);
1670 } /* do_load_store_optimize */
1673 typedef struct scc {
1674 ir_node *head; /**< the head of the list */
1677 /** A node entry. */
1678 typedef struct node_entry {
1679 unsigned DFSnum; /**< the DFS number of this node */
1680 unsigned low; /**< the low number of this node */
1681 int in_stack; /**< flag, set if the node is on the stack */
1682 ir_node *next; /**< link to the next node the the same scc */
1683 scc *pscc; /**< the scc of this node */
1684 unsigned POnum; /**< the post order number for blocks */
1687 /** A loop entry. */
1688 typedef struct loop_env {
1689 ir_phase ph; /**< the phase object */
1690 ir_node **stack; /**< the node stack */
1691 int tos; /**< tos index */
1692 unsigned nextDFSnum; /**< the current DFS number */
1693 unsigned POnum; /**< current post order number */
1695 unsigned changes; /**< a bitmask of graph changes */
1699 * Gets the node_entry of a node
1701 static node_entry *get_irn_ne(ir_node *irn, loop_env *env)
1703 ir_phase *ph = &env->ph;
1704 node_entry *e = phase_get_irn_data(&env->ph, irn);
1707 e = phase_alloc(ph, sizeof(*e));
1708 memset(e, 0, sizeof(*e));
1709 phase_set_irn_data(ph, irn, e);
1715 * Push a node onto the stack.
1717 * @param env the loop environment
1718 * @param n the node to push
1720 static void push(loop_env *env, ir_node *n)
1724 if (env->tos == ARR_LEN(env->stack)) {
1725 int nlen = ARR_LEN(env->stack) * 2;
1726 ARR_RESIZE(ir_node *, env->stack, nlen);
1728 env->stack[env->tos++] = n;
1729 e = get_irn_ne(n, env);
1734 * pop a node from the stack
1736 * @param env the loop environment
1738 * @return The topmost node
1740 static ir_node *pop(loop_env *env)
1742 ir_node *n = env->stack[--env->tos];
1743 node_entry *e = get_irn_ne(n, env);
1750 * Check if irn is a region constant.
1751 * The block or irn must strictly dominate the header block.
1753 * @param irn the node to check
1754 * @param header_block the header block of the induction variable
1756 static int is_rc(ir_node *irn, ir_node *header_block)
1758 ir_node *block = get_nodes_block(irn);
1760 return (block != header_block) && block_dominates(block, header_block);
1763 typedef struct phi_entry phi_entry;
1765 ir_node *phi; /**< A phi with a region const memory. */
1766 int pos; /**< The position of the region const memory */
1767 ir_node *load; /**< the newly created load for this phi */
1772 * An entry in the avail set.
1774 typedef struct avail_entry_t {
1775 ir_node *ptr; /**< the address pointer */
1776 ir_mode *mode; /**< the load mode */
1777 ir_node *load; /**< the associated Load */
1781 * Compare two avail entries.
1783 static int cmp_avail_entry(const void *elt, const void *key, size_t size)
1785 const avail_entry_t *a = elt;
1786 const avail_entry_t *b = key;
1789 return a->ptr != b->ptr || a->mode != b->mode;
1790 } /* cmp_avail_entry */
1793 * Calculate the hash value of an avail entry.
1795 static unsigned hash_cache_entry(const avail_entry_t *entry)
1797 return get_irn_idx(entry->ptr) * 9 + HASH_PTR(entry->mode);
1798 } /* hash_cache_entry */
1801 * Move loops out of loops if possible.
1803 * @param pscc the loop described by an SCC
1804 * @param env the loop environment
1806 static void move_loads_out_of_loops(scc *pscc, loop_env *env)
1808 ir_node *phi, *load, *next, *other, *next_other;
1811 phi_entry *phi_list = NULL;
1814 avail = new_set(cmp_avail_entry, 8);
1816 /* collect all outer memories */
1817 for (phi = pscc->head; phi != NULL; phi = next) {
1818 node_entry *ne = get_irn_ne(phi, env);
1821 /* check all memory Phi's */
1825 assert(get_irn_mode(phi) == mode_M && "DFS return non-memory Phi");
1827 for (j = get_irn_arity(phi) - 1; j >= 0; --j) {
1828 ir_node *pred = get_irn_n(phi, j);
1829 node_entry *pe = get_irn_ne(pred, env);
1831 if (pe->pscc != ne->pscc) {
1832 /* not in the same SCC, is region const */
1833 phi_entry *pe = phase_alloc(&env->ph, sizeof(*pe));
1837 pe->next = phi_list;
1842 /* no Phis no fun */
1843 assert(phi_list != NULL && "DFS found a loop without Phi");
1845 /* for now, we cannot handle more than one input (only reducible cf) */
1846 if (phi_list->next != NULL)
1849 for (load = pscc->head; load; load = next) {
1851 node_entry *ne = get_irn_ne(load, env);
1854 if (is_Load(load)) {
1855 ldst_info_t *info = get_irn_link(load);
1856 ir_node *ptr = get_Load_ptr(load);
1858 /* for now, we cannot handle Loads with exceptions */
1859 if (info->projs[pn_Load_res] == NULL || info->projs[pn_Load_X_regular] != NULL || info->projs[pn_Load_X_except] != NULL)
1862 /* for now, we can only move Load(Global) */
1863 if (! is_Global(ptr))
1865 ent = get_Global_entity(ptr);
1866 load_mode = get_Load_mode(load);
1867 for (other = pscc->head; other != NULL; other = next_other) {
1868 node_entry *ne = get_irn_ne(other, env);
1869 next_other = ne->next;
1871 if (is_Store(other)) {
1872 ir_alias_relation rel = get_alias_relation(
1874 get_Store_ptr(other),
1875 get_irn_mode(get_Store_value(other)),
1877 /* if the might be an alias, we cannot pass this Store */
1878 if (rel != ir_no_alias)
1881 /* only Phis and pure Calls are allowed here, so ignore them */
1883 if (other == NULL) {
1884 ldst_info_t *ninfo = NULL;
1888 /* yep, no aliasing Store found, Load can be moved */
1889 DB((dbg, LEVEL_1, " Found a Load that could be moved: %+F\n", load));
1891 db = get_irn_dbg_info(load);
1892 for (pe = phi_list; pe != NULL; pe = pe->next) {
1894 ir_node *phi = pe->phi;
1895 ir_node *blk = get_nodes_block(phi);
1896 ir_node *pred = get_Block_cfgpred_block(blk, pos);
1898 avail_entry_t entry, *res;
1901 entry.mode = load_mode;
1902 res = set_find(avail, &entry, sizeof(entry), hash_cache_entry(&entry));
1906 irn = new_rd_Load(db, pred, get_Phi_pred(phi, pos), ptr, load_mode, 0);
1908 set_insert(avail, &entry, sizeof(entry), hash_cache_entry(&entry));
1909 DB((dbg, LEVEL_1, " Created %+F in %+F\n", irn, pred));
1912 ninfo = get_ldst_info(irn, phase_obst(&env->ph));
1914 ninfo->projs[pn_Load_M] = mem = new_r_Proj(irn, mode_M, pn_Load_M);
1915 set_Phi_pred(phi, pos, mem);
1917 ninfo->projs[pn_Load_res] = new_r_Proj(irn, load_mode, pn_Load_res);
1920 /* now kill the old Load */
1921 exchange(info->projs[pn_Load_M], get_Load_mem(load));
1922 exchange(info->projs[pn_Load_res], ninfo->projs[pn_Load_res]);
1924 env->changes |= DF_CHANGED;
1929 } /* move_loads_out_of_loops */
1932 * Process a loop SCC.
1934 * @param pscc the SCC
1935 * @param env the loop environment
1937 static void process_loop(scc *pscc, loop_env *env)
1939 ir_node *irn, *next, *header = NULL;
1940 node_entry *b, *h = NULL;
1941 int j, only_phi, num_outside, process = 0;
1944 /* find the header block for this scc */
1945 for (irn = pscc->head; irn; irn = next) {
1946 node_entry *e = get_irn_ne(irn, env);
1947 ir_node *block = get_nodes_block(irn);
1950 b = get_irn_ne(block, env);
1952 if (header != NULL) {
1953 if (h->POnum < b->POnum) {
1963 /* check if this scc contains only Phi, Loads or Stores nodes */
1967 for (irn = pscc->head; irn; irn = next) {
1968 node_entry *e = get_irn_ne(irn, env);
1971 switch (get_irn_opcode(irn)) {
1973 if (is_Call_pure(irn)) {
1974 /* pure calls can be treated like loads */
1978 /* non-pure calls must be handle like may-alias Stores */
1981 /* cannot handle CopyB yet */
1985 if (get_Load_volatility(irn) == volatility_is_volatile) {
1986 /* cannot handle loops with volatile Loads */
1992 if (get_Store_volatility(irn) == volatility_is_volatile) {
1993 /* cannot handle loops with volatile Stores */
2002 for (j = get_irn_arity(irn) - 1; j >= 0; --j) {
2003 ir_node *pred = get_irn_n(irn, j);
2004 node_entry *pe = get_irn_ne(pred, env);
2006 if (pe->pscc != e->pscc) {
2007 /* not in the same SCC, must be a region const */
2008 if (! is_rc(pred, header)) {
2009 /* not a memory loop */
2012 if (out_rc == NULL) {
2013 /* first region constant */
2016 } else if (out_rc != pred) {
2017 /* another region constant */
2028 /* found a memory loop */
2029 DB((dbg, LEVEL_2, " Found a memory loop:\n "));
2030 if (only_phi && num_outside == 1) {
2031 /* a phi cycle with only one real predecessor can be collapsed */
2032 DB((dbg, LEVEL_2, " Found an USELESS Phi cycle:\n "));
2034 for (irn = pscc->head; irn; irn = next) {
2035 node_entry *e = get_irn_ne(irn, env);
2037 exchange(irn, out_rc);
2039 env->changes |= DF_CHANGED;
2043 #ifdef DEBUG_libfirm
2044 for (irn = pscc->head; irn; irn = next) {
2045 node_entry *e = get_irn_ne(irn, env);
2047 DB((dbg, LEVEL_2, " %+F,", irn));
2049 DB((dbg, LEVEL_2, "\n"));
2051 move_loads_out_of_loops(pscc, env);
2055 } /* process_loop */
2060 * @param pscc the SCC
2061 * @param env the loop environment
2063 static void process_scc(scc *pscc, loop_env *env)
2065 ir_node *head = pscc->head;
2066 node_entry *e = get_irn_ne(head, env);
2068 #ifdef DEBUG_libfirm
2070 ir_node *irn, *next;
2072 DB((dbg, LEVEL_4, " SCC at %p:\n ", pscc));
2073 for (irn = pscc->head; irn; irn = next) {
2074 node_entry *e = get_irn_ne(irn, env);
2078 DB((dbg, LEVEL_4, " %+F,", irn));
2080 DB((dbg, LEVEL_4, "\n"));
2084 if (e->next != NULL) {
2085 /* this SCC has more than one member */
2086 process_loop(pscc, env);
2091 * Do Tarjan's SCC algorithm and drive load/store optimization.
2093 * @param irn start at this node
2094 * @param env the loop environment
2096 static void dfs(ir_node *irn, loop_env *env)
2099 node_entry *node = get_irn_ne(irn, env);
2101 mark_irn_visited(irn);
2103 node->DFSnum = env->nextDFSnum++;
2104 node->low = node->DFSnum;
2108 if (is_Phi(irn) || is_Sync(irn)) {
2109 n = get_irn_arity(irn);
2110 for (i = 0; i < n; ++i) {
2111 ir_node *pred = get_irn_n(irn, i);
2112 node_entry *o = get_irn_ne(pred, env);
2114 if (!irn_visited(pred)) {
2116 node->low = MIN(node->low, o->low);
2118 if (o->DFSnum < node->DFSnum && o->in_stack)
2119 node->low = MIN(o->DFSnum, node->low);
2121 } else if (is_fragile_op(irn)) {
2122 ir_node *pred = get_fragile_op_mem(irn);
2123 node_entry *o = get_irn_ne(pred, env);
2125 if (!irn_visited(pred)) {
2127 node->low = MIN(node->low, o->low);
2129 if (o->DFSnum < node->DFSnum && o->in_stack)
2130 node->low = MIN(o->DFSnum, node->low);
2131 } else if (is_Proj(irn)) {
2132 ir_node *pred = get_Proj_pred(irn);
2133 node_entry *o = get_irn_ne(pred, env);
2135 if (!irn_visited(pred)) {
2137 node->low = MIN(node->low, o->low);
2139 if (o->DFSnum < node->DFSnum && o->in_stack)
2140 node->low = MIN(o->DFSnum, node->low);
2143 /* IGNORE predecessors */
2146 if (node->low == node->DFSnum) {
2147 scc *pscc = phase_alloc(&env->ph, sizeof(*pscc));
2155 e = get_irn_ne(x, env);
2157 e->next = pscc->head;
2161 process_scc(pscc, env);
2166 * Do the DFS on the memory edges a graph.
2168 * @param irg the graph to process
2169 * @param env the loop environment
2171 static void do_dfs(ir_graph *irg, loop_env *env)
2173 ir_graph *rem = current_ir_graph;
2174 ir_node *endblk, *end;
2177 current_ir_graph = irg;
2178 inc_irg_visited(irg);
2180 /* visit all memory nodes */
2181 endblk = get_irg_end_block(irg);
2182 for (i = get_Block_n_cfgpreds(endblk) - 1; i >= 0; --i) {
2183 ir_node *pred = get_Block_cfgpred(endblk, i);
2185 pred = skip_Proj(pred);
2186 if (is_Return(pred))
2187 dfs(get_Return_mem(pred), env);
2188 else if (is_Unreachable(pred))
2189 dfs(get_Unreachable_mem(pred), env);
2190 else if (is_Raise(pred))
2191 dfs(get_Raise_mem(pred), env);
2192 else if (is_fragile_op(pred))
2193 dfs(get_fragile_op_mem(pred), env);
2195 assert(0 && "Unknown EndBlock predecessor");
2199 /* visit the keep-alives */
2200 end = get_irg_end(irg);
2201 for (i = get_End_n_keepalives(end) - 1; i >= 0; --i) {
2202 ir_node *ka = get_End_keepalive(end, i);
2204 if (is_Phi(ka) && !irn_visited(ka))
2207 current_ir_graph = rem;
2211 * Optimize Loads/Stores in loops.
2213 * @param irg the graph
2215 static int optimize_loops(ir_graph *irg)
2219 env.stack = NEW_ARR_F(ir_node *, 128);
2224 phase_init(&env.ph, irg, phase_irn_init_default);
2226 /* calculate the SCC's and drive loop optimization. */
2229 DEL_ARR_F(env.stack);
2230 phase_deinit(&env.ph);
2233 } /* optimize_loops */
2236 * do the load store optimization
2238 int optimize_load_store(ir_graph *irg)
2242 FIRM_DBG_REGISTER(dbg, "firm.opt.ldstopt");
2244 assert(get_irg_phase_state(irg) != phase_building);
2245 assert(get_irg_pinned(irg) != op_pin_state_floats &&
2246 "LoadStore optimization needs pinned graph");
2248 /* we need landing pads */
2249 remove_critical_cf_edges(irg);
2253 /* for Phi optimization post-dominators are needed ... */
2254 assure_postdoms(irg);
2256 if (get_opt_alias_analysis()) {
2257 assure_irg_entity_usage_computed(irg);
2258 assure_irp_globals_entity_usage_computed();
2261 obstack_init(&env.obst);
2264 /* init the links, then collect Loads/Stores/Proj's in lists */
2266 irg_walk_graph(irg, firm_clear_link, collect_nodes, &env);
2268 /* now we have collected enough information, optimize */
2269 irg_walk_graph(irg, NULL, do_load_store_optimize, &env);
2271 env.changes |= optimize_loops(irg);
2273 obstack_free(&env.obst, NULL);
2275 /* Handle graph state */
2277 set_irg_outs_inconsistent(irg);
2278 set_irg_entity_usage_state(irg, ir_entity_usage_not_computed);
2281 if (env.changes & CF_CHANGED) {
2282 /* is this really needed: Yes, control flow changed, block might
2283 have Bad() predecessors. */
2284 set_irg_doms_inconsistent(irg);
2286 return env.changes != 0;
2287 } /* optimize_load_store */
2289 ir_graph_pass_t *optimize_load_store_pass(const char *name)
2291 return def_graph_pass_ret(name ? name : "ldst", optimize_load_store);
2292 } /* optimize_load_store_pass */