2 * Copyright (C) 1995-2008 University of Karlsruhe. All right reserved.
4 * This file is part of libFirm.
6 * This file may be distributed and/or modified under the terms of the
7 * GNU General Public License version 2 as published by the Free Software
8 * Foundation and appearing in the file LICENSE.GPL included in the
9 * packaging of this file.
11 * Licensees holding valid libFirm Professional Edition licenses may use
12 * this file in accordance with the libFirm Commercial License.
13 * Agreement provided with the Software.
15 * This file is provided AS IS with NO WARRANTY OF ANY KIND, INCLUDING THE
16 * WARRANTY OF DESIGN, MERCHANTABILITY AND FITNESS FOR A PARTICULAR
22 * @brief Load/Store optimizations.
23 * @author Michael Beck
30 #include "iroptimize.h"
32 #include "irgraph_t.h"
40 #include "dbginfo_t.h"
41 #include "iropt_dbg.h"
47 #include "opt_polymorphy.h"
49 #include "irphase_t.h"
54 /** The debug handle. */
55 DEBUG_ONLY(static firm_dbg_module_t *dbg;)
58 #include "cacheopt/cachesim.h"
62 #define IMAX(a,b) ((a) > (b) ? (a) : (b))
64 #define MAX_PROJ IMAX(IMAX(pn_Load_max, pn_Store_max), pn_Call_max)
67 DF_CHANGED = 1, /**< data flow changed */
68 CF_CHANGED = 2, /**< control flow changed */
74 typedef struct _walk_env_t {
75 struct obstack obst; /**< list of all stores */
76 unsigned changes; /**< a bitmask of graph changes */
79 /** A Load/Store info. */
80 typedef struct _ldst_info_t {
81 ir_node *projs[MAX_PROJ]; /**< list of Proj's of this node */
82 ir_node *exc_block; /**< the exception block if available */
83 int exc_idx; /**< predecessor index in the exception block */
84 unsigned visited; /**< visited counter for breaking loops */
88 * flags for control flow.
91 BLOCK_HAS_COND = 1, /**< Block has conditional control flow */
92 BLOCK_HAS_EXC = 2 /**< Block has exceptional control flow */
98 typedef struct _block_info_t {
99 unsigned flags; /**< flags for the block */
102 /** the master visited flag for loop detection. */
103 static unsigned master_visited = 0;
105 #define INC_MASTER() ++master_visited
106 #define MARK_NODE(info) (info)->visited = master_visited
107 #define NODE_VISITED(info) (info)->visited >= master_visited
110 * get the Load/Store info of a node
112 static ldst_info_t *get_ldst_info(ir_node *node, struct obstack *obst)
114 ldst_info_t *info = get_irn_link(node);
117 info = OALLOCZ(obst, ldst_info_t);
118 set_irn_link(node, info);
121 } /* get_ldst_info */
124 * get the Block info of a node
126 static block_info_t *get_block_info(ir_node *node, struct obstack *obst)
128 block_info_t *info = get_irn_link(node);
131 info = OALLOCZ(obst, block_info_t);
132 set_irn_link(node, info);
135 } /* get_block_info */
138 * update the projection info for a Load/Store
140 static unsigned update_projs(ldst_info_t *info, ir_node *proj)
142 long nr = get_Proj_proj(proj);
144 assert(0 <= nr && nr <= MAX_PROJ && "Wrong proj from LoadStore");
146 if (info->projs[nr]) {
147 /* there is already one, do CSE */
148 exchange(proj, info->projs[nr]);
152 info->projs[nr] = proj;
158 * update the exception block info for a Load/Store node.
160 * @param info the load/store info struct
161 * @param block the exception handler block for this load/store
162 * @param pos the control flow input of the block
164 static unsigned update_exc(ldst_info_t *info, ir_node *block, int pos)
166 assert(info->exc_block == NULL && "more than one exception block found");
168 info->exc_block = block;
173 /** Return the number of uses of an address node */
174 #define get_irn_n_uses(adr) get_irn_n_edges(adr)
177 * walker, collects all Load/Store/Proj nodes
179 * walks from Start -> End
181 static void collect_nodes(ir_node *node, void *env)
183 ir_opcode opcode = get_irn_opcode(node);
184 ir_node *pred, *blk, *pred_blk;
185 ldst_info_t *ldst_info;
186 walk_env_t *wenv = env;
188 if (opcode == iro_Proj) {
189 pred = get_Proj_pred(node);
190 opcode = get_irn_opcode(pred);
192 if (opcode == iro_Load || opcode == iro_Store || opcode == iro_Call) {
193 ldst_info = get_ldst_info(pred, &wenv->obst);
195 wenv->changes |= update_projs(ldst_info, node);
198 * Place the Proj's to the same block as the
199 * predecessor Load. This is always ok and prevents
200 * "non-SSA" form after optimizations if the Proj
201 * is in a wrong block.
203 blk = get_nodes_block(node);
204 pred_blk = get_nodes_block(pred);
205 if (blk != pred_blk) {
206 wenv->changes |= DF_CHANGED;
207 set_nodes_block(node, pred_blk);
210 } else if (opcode == iro_Block) {
213 for (i = get_Block_n_cfgpreds(node) - 1; i >= 0; --i) {
214 ir_node *pred_block, *proj;
215 block_info_t *bl_info;
218 pred = proj = get_Block_cfgpred(node, i);
221 pred = get_Proj_pred(proj);
222 is_exc = get_Proj_proj(proj) == pn_Generic_X_except;
225 /* ignore Bad predecessors, they will be removed later */
229 pred_block = get_nodes_block(pred);
230 bl_info = get_block_info(pred_block, &wenv->obst);
232 if (is_fragile_op(pred) && is_exc)
233 bl_info->flags |= BLOCK_HAS_EXC;
234 else if (is_irn_forking(pred))
235 bl_info->flags |= BLOCK_HAS_COND;
237 opcode = get_irn_opcode(pred);
238 if (is_exc && (opcode == iro_Load || opcode == iro_Store || opcode == iro_Call)) {
239 ldst_info = get_ldst_info(pred, &wenv->obst);
241 wenv->changes |= update_exc(ldst_info, node, i);
245 } /* collect_nodes */
248 * Returns an entity if the address ptr points to a constant one.
250 * @param ptr the address
252 * @return an entity or NULL
254 static ir_entity *find_constant_entity(ir_node *ptr)
257 if (is_SymConst(ptr) && get_SymConst_kind(ptr) == symconst_addr_ent) {
258 return get_SymConst_entity(ptr);
259 } else if (is_Sel(ptr)) {
260 ir_entity *ent = get_Sel_entity(ptr);
261 ir_type *tp = get_entity_owner(ent);
263 /* Do not fiddle with polymorphism. */
264 if (is_Class_type(get_entity_owner(ent)) &&
265 ((get_entity_n_overwrites(ent) != 0) ||
266 (get_entity_n_overwrittenby(ent) != 0) ) )
269 if (is_Array_type(tp)) {
273 for (i = 0, n = get_Sel_n_indexs(ptr); i < n; ++i) {
275 tarval *tlower, *tupper;
276 ir_node *index = get_Sel_index(ptr, i);
277 tarval *tv = computed_value(index);
279 /* check if the index is constant */
280 if (tv == tarval_bad)
283 bound = get_array_lower_bound(tp, i);
284 tlower = computed_value(bound);
285 bound = get_array_upper_bound(tp, i);
286 tupper = computed_value(bound);
288 if (tlower == tarval_bad || tupper == tarval_bad)
291 if (tarval_cmp(tv, tlower) & pn_Cmp_Lt)
293 if (tarval_cmp(tupper, tv) & pn_Cmp_Lt)
296 /* ok, bounds check finished */
300 if (get_entity_linkage(ent) & IR_LINKAGE_CONSTANT)
304 ptr = get_Sel_ptr(ptr);
305 } else if (is_Add(ptr)) {
306 ir_node *l = get_Add_left(ptr);
307 ir_node *r = get_Add_right(ptr);
309 if (get_irn_mode(l) == get_irn_mode(ptr) && is_Const(r))
311 else if (get_irn_mode(r) == get_irn_mode(ptr) && is_Const(l))
316 /* for now, we support only one addition, reassoc should fold all others */
317 if (! is_SymConst(ptr) && !is_Sel(ptr))
319 } else if (is_Sub(ptr)) {
320 ir_node *l = get_Sub_left(ptr);
321 ir_node *r = get_Sub_right(ptr);
323 if (get_irn_mode(l) == get_irn_mode(ptr) && is_Const(r))
327 /* for now, we support only one substraction, reassoc should fold all others */
328 if (! is_SymConst(ptr) && !is_Sel(ptr))
333 } /* find_constant_entity */
336 * Return the Selection index of a Sel node from dimension n
338 static long get_Sel_array_index_long(ir_node *n, int dim)
340 ir_node *index = get_Sel_index(n, dim);
341 assert(is_Const(index));
342 return get_tarval_long(get_Const_tarval(index));
343 } /* get_Sel_array_index_long */
346 * Returns the accessed component graph path for an
347 * node computing an address.
349 * @param ptr the node computing the address
350 * @param depth current depth in steps upward from the root
353 static compound_graph_path *rec_get_accessed_path(ir_node *ptr, int depth)
355 compound_graph_path *res = NULL;
356 ir_entity *root, *field, *ent;
357 int path_len, pos, idx;
361 if (is_SymConst(ptr)) {
362 /* a SymConst. If the depth is 0, this is an access to a global
363 * entity and we don't need a component path, else we know
364 * at least its length.
366 assert(get_SymConst_kind(ptr) == symconst_addr_ent);
367 root = get_SymConst_entity(ptr);
368 res = (depth == 0) ? NULL : new_compound_graph_path(get_entity_type(root), depth);
369 } else if (is_Sel(ptr)) {
370 /* it's a Sel, go up until we find the root */
371 res = rec_get_accessed_path(get_Sel_ptr(ptr), depth+1);
375 /* fill up the step in the path at the current position */
376 field = get_Sel_entity(ptr);
377 path_len = get_compound_graph_path_length(res);
378 pos = path_len - depth - 1;
379 set_compound_graph_path_node(res, pos, field);
381 if (is_Array_type(get_entity_owner(field))) {
382 assert(get_Sel_n_indexs(ptr) == 1 && "multi dim arrays not implemented");
383 set_compound_graph_path_array_index(res, pos, get_Sel_array_index_long(ptr, 0));
385 } else if (is_Add(ptr)) {
386 ir_node *l = get_Add_left(ptr);
387 ir_node *r = get_Add_right(ptr);
388 ir_mode *mode = get_irn_mode(ptr);
391 if (is_Const(r) && get_irn_mode(l) == mode) {
393 tv = get_Const_tarval(r);
396 tv = get_Const_tarval(l);
399 mode = get_tarval_mode(tv);
402 /* ptr must be a Sel or a SymConst, this was checked in find_constant_entity() */
404 field = get_Sel_entity(ptr);
406 field = get_SymConst_entity(ptr);
409 for (ent = field;;) {
411 tarval *sz, *tv_index, *tlower, *tupper;
414 tp = get_entity_type(ent);
415 if (! is_Array_type(tp))
417 ent = get_array_element_entity(tp);
418 size = get_type_size_bytes(get_entity_type(ent));
419 sz = new_tarval_from_long(size, mode);
421 tv_index = tarval_div(tmp, sz);
422 tmp = tarval_mod(tmp, sz);
424 if (tv_index == tarval_bad || tmp == tarval_bad)
427 assert(get_array_n_dimensions(tp) == 1 && "multiarrays not implemented");
428 bound = get_array_lower_bound(tp, 0);
429 tlower = computed_value(bound);
430 bound = get_array_upper_bound(tp, 0);
431 tupper = computed_value(bound);
433 if (tlower == tarval_bad || tupper == tarval_bad)
436 if (tarval_cmp(tv_index, tlower) & pn_Cmp_Lt)
438 if (tarval_cmp(tupper, tv_index) & pn_Cmp_Lt)
441 /* ok, bounds check finished */
444 if (! tarval_is_null(tmp)) {
445 /* access to some struct/union member */
449 /* should be at least ONE array */
453 res = rec_get_accessed_path(ptr, depth + idx);
457 path_len = get_compound_graph_path_length(res);
458 pos = path_len - depth - idx;
460 for (ent = field;;) {
462 tarval *sz, *tv_index;
465 tp = get_entity_type(ent);
466 if (! is_Array_type(tp))
468 ent = get_array_element_entity(tp);
469 set_compound_graph_path_node(res, pos, ent);
471 size = get_type_size_bytes(get_entity_type(ent));
472 sz = new_tarval_from_long(size, mode);
474 tv_index = tarval_div(tv, sz);
475 tv = tarval_mod(tv, sz);
477 /* worked above, should work again */
478 assert(tv_index != tarval_bad && tv != tarval_bad);
480 /* bounds already checked above */
481 index = get_tarval_long(tv_index);
482 set_compound_graph_path_array_index(res, pos, index);
485 } else if (is_Sub(ptr)) {
486 ir_node *l = get_Sub_left(ptr);
487 ir_node *r = get_Sub_right(ptr);
490 tv = get_Const_tarval(r);
495 } /* rec_get_accessed_path */
498 * Returns an access path or NULL. The access path is only
499 * valid, if the graph is in phase_high and _no_ address computation is used.
501 static compound_graph_path *get_accessed_path(ir_node *ptr)
503 compound_graph_path *gr = rec_get_accessed_path(ptr, 0);
505 } /* get_accessed_path */
507 typedef struct path_entry {
509 struct path_entry *next;
513 static ir_node *rec_find_compound_ent_value(ir_node *ptr, path_entry *next)
515 path_entry entry, *p;
516 ir_entity *ent, *field;
517 ir_initializer_t *initializer;
523 if (is_SymConst(ptr)) {
525 ent = get_SymConst_entity(ptr);
526 initializer = get_entity_initializer(ent);
527 for (p = next; p != NULL;) {
528 if (initializer->kind != IR_INITIALIZER_COMPOUND)
530 n = get_initializer_compound_n_entries(initializer);
531 tp = get_entity_type(ent);
533 if (is_Array_type(tp)) {
534 ent = get_array_element_entity(tp);
539 initializer = get_initializer_compound_value(initializer, 0);
543 if (p->index >= (int) n)
545 initializer = get_initializer_compound_value(initializer, p->index);
550 tp = get_entity_type(ent);
551 while (is_Array_type(tp)) {
552 ent = get_array_element_entity(tp);
553 tp = get_entity_type(ent);
555 n = get_initializer_compound_n_entries(initializer);
558 initializer = get_initializer_compound_value(initializer, 0);
561 switch (initializer->kind) {
562 case IR_INITIALIZER_CONST:
563 return get_initializer_const_value(initializer);
564 case IR_INITIALIZER_TARVAL:
565 case IR_INITIALIZER_NULL:
569 } else if (is_Sel(ptr)) {
570 entry.ent = field = get_Sel_entity(ptr);
571 tp = get_entity_owner(field);
572 if (is_Array_type(tp)) {
573 assert(get_Sel_n_indexs(ptr) == 1 && "multi dim arrays not implemented");
574 entry.index = get_Sel_array_index_long(ptr, 0) - get_array_lower_bound_int(tp, 0);
576 int i, n_members = get_compound_n_members(tp);
577 for (i = 0; i < n_members; ++i) {
578 if (get_compound_member(tp, i) == field)
581 if (i >= n_members) {
582 /* not found: should NOT happen */
587 return rec_find_compound_ent_value(get_Sel_ptr(ptr), &entry);
588 } else if (is_Add(ptr)) {
589 ir_node *l = get_Add_left(ptr);
590 ir_node *r = get_Add_right(ptr);
596 tv = get_Const_tarval(r);
599 tv = get_Const_tarval(l);
602 mode = get_tarval_mode(tv);
604 /* ptr must be a Sel or a SymConst, this was checked in find_constant_entity() */
606 field = get_Sel_entity(ptr);
608 field = get_SymConst_entity(ptr);
611 /* count needed entries */
613 for (ent = field;;) {
614 tp = get_entity_type(ent);
615 if (! is_Array_type(tp))
617 ent = get_array_element_entity(tp);
620 /* should be at least ONE entry */
624 /* allocate the right number of entries */
625 NEW_ARR_A(path_entry, p, pos);
629 for (ent = field;;) {
631 tarval *sz, *tv_index, *tlower, *tupper;
635 tp = get_entity_type(ent);
636 if (! is_Array_type(tp))
638 ent = get_array_element_entity(tp);
640 p[pos].next = &p[pos + 1];
642 size = get_type_size_bytes(get_entity_type(ent));
643 sz = new_tarval_from_long(size, mode);
645 tv_index = tarval_div(tv, sz);
646 tv = tarval_mod(tv, sz);
648 if (tv_index == tarval_bad || tv == tarval_bad)
651 assert(get_array_n_dimensions(tp) == 1 && "multiarrays not implemented");
652 bound = get_array_lower_bound(tp, 0);
653 tlower = computed_value(bound);
654 bound = get_array_upper_bound(tp, 0);
655 tupper = computed_value(bound);
657 if (tlower == tarval_bad || tupper == tarval_bad)
660 if (tarval_cmp(tv_index, tlower) & pn_Cmp_Lt)
662 if (tarval_cmp(tupper, tv_index) & pn_Cmp_Lt)
665 /* ok, bounds check finished */
666 index = get_tarval_long(tv_index);
667 p[pos].index = index;
670 if (! tarval_is_null(tv)) {
671 /* hmm, wrong access */
674 p[pos - 1].next = next;
675 return rec_find_compound_ent_value(ptr, p);
676 } else if (is_Sub(ptr)) {
677 ir_node *l = get_Sub_left(ptr);
678 ir_node *r = get_Sub_right(ptr);
681 tv = get_Const_tarval(r);
688 static ir_node *find_compound_ent_value(ir_node *ptr)
690 return rec_find_compound_ent_value(ptr, NULL);
694 static void reduce_adr_usage(ir_node *ptr);
697 * Update a Load that may have lost its users.
699 static void handle_load_update(ir_node *load)
701 ldst_info_t *info = get_irn_link(load);
703 /* do NOT touch volatile loads for now */
704 if (get_Load_volatility(load) == volatility_is_volatile)
707 if (! info->projs[pn_Load_res] && ! info->projs[pn_Load_X_except]) {
708 ir_node *ptr = get_Load_ptr(load);
709 ir_node *mem = get_Load_mem(load);
711 /* a Load whose value is neither used nor exception checked, remove it */
712 exchange(info->projs[pn_Load_M], mem);
713 if (info->projs[pn_Load_X_regular])
714 exchange(info->projs[pn_Load_X_regular], new_r_Jmp(get_nodes_block(load)));
716 reduce_adr_usage(ptr);
718 } /* handle_load_update */
721 * A use of an address node has vanished. Check if this was a Proj
722 * node and update the counters.
724 static void reduce_adr_usage(ir_node *ptr)
727 if (get_irn_n_edges(ptr) <= 0) {
728 /* this Proj is dead now */
729 ir_node *pred = get_Proj_pred(ptr);
732 ldst_info_t *info = get_irn_link(pred);
733 info->projs[get_Proj_proj(ptr)] = NULL;
735 /* this node lost its result proj, handle that */
736 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(get_nodes_block(pred), 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);
1111 * Check if we can remove the exception from a Load:
1112 * This can be done, if the address is from an Sel(Alloc) and
1113 * the Sel type is a subtype of the allocated type.
1115 * This optimizes some often used OO constructs,
1116 * like x = new O; x->t;
1118 if (info->projs[pn_Load_X_except]) {
1119 ir_node *addr = ptr;
1121 /* find base address */
1122 while (is_Sel(addr))
1123 addr = get_Sel_ptr(addr);
1124 if (is_Alloc(skip_Proj(skip_Cast(addr)))) {
1125 /* simple case: a direct load after an Alloc. Firm Alloc throw
1126 * an exception in case of out-of-memory. So, there is no way for an
1127 * exception in this load.
1128 * This code is constructed by the "exception lowering" in the Jack compiler.
1130 exchange(info->projs[pn_Load_X_except], new_Bad());
1131 info->projs[pn_Load_X_except] = NULL;
1132 exchange(info->projs[pn_Load_X_regular], new_r_Jmp(get_nodes_block(load)));
1133 info->projs[pn_Load_X_regular] = NULL;
1138 /* The mem of the Load. Must still be returned after optimization. */
1139 mem = get_Load_mem(load);
1141 if (! info->projs[pn_Load_res] && ! info->projs[pn_Load_X_except]) {
1142 /* a Load which value is neither used nor exception checked, remove it */
1143 exchange(info->projs[pn_Load_M], mem);
1145 if (info->projs[pn_Load_X_regular]) {
1146 /* should not happen, but if it does, remove it */
1147 exchange(info->projs[pn_Load_X_regular], new_r_Jmp(get_nodes_block(load)));
1151 reduce_adr_usage(ptr);
1152 return res | DF_CHANGED;
1155 /* Load from a constant polymorphic field, where we can resolve
1157 value = transform_polymorph_Load(load);
1158 if (value == load) {
1160 /* check if we can determine the entity that will be loaded */
1161 ent = find_constant_entity(ptr);
1163 && get_entity_visibility(ent) != ir_visibility_external) {
1164 /* a static allocation that is not external: there should be NO
1165 * exception when loading even if we cannot replace the load itself.
1168 /* no exception, clear the info field as it might be checked later again */
1169 if (info->projs[pn_Load_X_except]) {
1170 exchange(info->projs[pn_Load_X_except], new_Bad());
1171 info->projs[pn_Load_X_except] = NULL;
1174 if (info->projs[pn_Load_X_regular]) {
1175 exchange(info->projs[pn_Load_X_regular], new_r_Jmp(get_nodes_block(load)));
1176 info->projs[pn_Load_X_regular] = NULL;
1180 if (get_entity_linkage(ent) & IR_LINKAGE_CONSTANT) {
1181 if (ent->initializer != NULL) {
1182 /* new style initializer */
1183 value = find_compound_ent_value(ptr);
1184 } else if (entity_has_compound_ent_values(ent)) {
1185 /* old style initializer */
1186 compound_graph_path *path = get_accessed_path(ptr);
1189 assert(is_proper_compound_graph_path(path, get_compound_graph_path_length(path)-1));
1191 value = get_compound_ent_value_by_path(ent, path);
1192 DB((dbg, LEVEL_1, " Constant access at %F%F resulted in %+F\n", ent, path, value));
1193 free_compound_graph_path(path);
1197 value = can_replace_load_by_const(load, value);
1201 if (value != NULL) {
1202 /* we completely replace the load by this value */
1203 if (info->projs[pn_Load_X_except]) {
1204 exchange(info->projs[pn_Load_X_except], new_Bad());
1205 info->projs[pn_Load_X_except] = NULL;
1208 if (info->projs[pn_Load_X_regular]) {
1209 exchange(info->projs[pn_Load_X_regular], new_r_Jmp(get_nodes_block(load)));
1210 info->projs[pn_Load_X_regular] = NULL;
1213 if (info->projs[pn_Load_M]) {
1214 exchange(info->projs[pn_Load_M], mem);
1217 if (info->projs[pn_Load_res]) {
1218 exchange(info->projs[pn_Load_res], value);
1222 reduce_adr_usage(ptr);
1226 /* Check, if the address of this load is used more than once.
1227 * If not, more load cannot be removed in any case. */
1228 if (get_irn_n_uses(ptr) <= 1 && get_irn_n_uses(get_base_and_offset(ptr, &dummy)) <= 1)
1232 * follow the memory chain as long as there are only Loads
1233 * and try to replace current Load or Store by a previous one.
1234 * Note that in unreachable loops it might happen that we reach
1235 * load again, as well as we can fall into a cycle.
1236 * We break such cycles using a special visited flag.
1239 res = follow_Mem_chain(load, skip_Proj(mem));
1241 } /* optimize_load */
1244 * Check whether a value of mode new_mode would completely overwrite a value
1245 * of mode old_mode in memory.
1247 static int is_completely_overwritten(ir_mode *old_mode, ir_mode *new_mode)
1249 return get_mode_size_bits(new_mode) >= get_mode_size_bits(old_mode);
1250 } /* is_completely_overwritten */
1253 * Check whether small is a part of large (starting at same address).
1255 static int is_partially_same(ir_node *small, ir_node *large)
1257 ir_mode *sm = get_irn_mode(small);
1258 ir_mode *lm = get_irn_mode(large);
1260 /* FIXME: Check endianness */
1261 return is_Conv(small) && get_Conv_op(small) == large
1262 && get_mode_size_bytes(sm) < get_mode_size_bytes(lm)
1263 && get_mode_arithmetic(sm) == irma_twos_complement
1264 && get_mode_arithmetic(lm) == irma_twos_complement;
1265 } /* is_partially_same */
1268 * follow the memory chain as long as there are only Loads and alias free Stores.
1270 * INC_MASTER() must be called before dive into
1272 static unsigned follow_Mem_chain_for_Store(ir_node *store, ir_node *curr)
1275 ldst_info_t *info = get_irn_link(store);
1277 ir_node *ptr = get_Store_ptr(store);
1278 ir_node *mem = get_Store_mem(store);
1279 ir_node *value = get_Store_value(store);
1280 ir_mode *mode = get_irn_mode(value);
1281 ir_node *block = get_nodes_block(store);
1282 ir_node *mblk = get_Block_MacroBlock(block);
1284 for (pred = curr; pred != store;) {
1285 ldst_info_t *pred_info = get_irn_link(pred);
1288 * BEWARE: one might think that checking the modes is useless, because
1289 * if the pointers are identical, they refer to the same object.
1290 * This is only true in strong typed languages, not is C were the following
1291 * is possible *(ir_type1 *)p = a; *(ir_type2 *)p = b ...
1292 * However, if the size of the mode that is written is bigger or equal the
1293 * size of the old one, the old value is completely overwritten and can be
1296 if (is_Store(pred) && get_Store_ptr(pred) == ptr &&
1297 get_nodes_MacroBlock(pred) == mblk) {
1299 * a Store after a Store in the same MacroBlock -- a write after write.
1303 * We may remove the first Store, if the old value is completely
1304 * overwritten or the old value is a part of the new value,
1305 * and if it does not have an exception handler.
1307 * TODO: What, if both have the same exception handler ???
1309 if (get_Store_volatility(pred) != volatility_is_volatile
1310 && !pred_info->projs[pn_Store_X_except]) {
1311 ir_node *predvalue = get_Store_value(pred);
1312 ir_mode *predmode = get_irn_mode(predvalue);
1314 if (is_completely_overwritten(predmode, mode)
1315 || is_partially_same(predvalue, value)) {
1316 DBG_OPT_WAW(pred, store);
1317 exchange(pred_info->projs[pn_Store_M], get_Store_mem(pred));
1319 reduce_adr_usage(ptr);
1325 * We may remove the Store, if the old value already contains
1326 * the new value, and if it does not have an exception handler.
1328 * TODO: What, if both have the same exception handler ???
1330 if (get_Store_volatility(store) != volatility_is_volatile
1331 && !info->projs[pn_Store_X_except]) {
1332 ir_node *predvalue = get_Store_value(pred);
1334 if (is_partially_same(value, predvalue)) {
1335 DBG_OPT_WAW(pred, store);
1336 exchange(info->projs[pn_Store_M], mem);
1338 reduce_adr_usage(ptr);
1342 } else if (is_Load(pred) && get_Load_ptr(pred) == ptr &&
1343 value == pred_info->projs[pn_Load_res]) {
1345 * a Store of a value just loaded from the same address
1346 * -- a write after read.
1347 * We may remove the Store, if it does not have an exception
1350 if (! info->projs[pn_Store_X_except]) {
1351 DBG_OPT_WAR(store, pred);
1352 exchange(info->projs[pn_Store_M], mem);
1354 reduce_adr_usage(ptr);
1359 if (is_Store(pred)) {
1360 /* check if we can pass through this store */
1361 ir_alias_relation rel = get_alias_relation(
1363 get_Store_ptr(pred),
1364 get_irn_mode(get_Store_value(pred)),
1366 /* if the might be an alias, we cannot pass this Store */
1367 if (rel != ir_no_alias)
1369 pred = skip_Proj(get_Store_mem(pred));
1370 } else if (is_Load(pred)) {
1371 ir_alias_relation rel = get_alias_relation(
1372 current_ir_graph, get_Load_ptr(pred), get_Load_mode(pred),
1374 if (rel != ir_no_alias)
1377 pred = skip_Proj(get_Load_mem(pred));
1379 /* follow only Load chains */
1383 /* check for cycles */
1384 if (NODE_VISITED(pred_info))
1386 MARK_NODE(pred_info);
1389 if (is_Sync(pred)) {
1392 /* handle all Sync predecessors */
1393 for (i = get_Sync_n_preds(pred) - 1; i >= 0; --i) {
1394 res |= follow_Mem_chain_for_Store(store, skip_Proj(get_Sync_pred(pred, i)));
1400 } /* follow_Mem_chain_for_Store */
1402 /** find entity used as base for an address calculation */
1403 static ir_entity *find_entity(ir_node *ptr)
1405 switch (get_irn_opcode(ptr)) {
1407 return get_SymConst_entity(ptr);
1409 ir_node *pred = get_Sel_ptr(ptr);
1410 if (get_irg_frame(get_irn_irg(ptr)) == pred)
1411 return get_Sel_entity(ptr);
1413 return find_entity(pred);
1417 ir_node *left = get_binop_left(ptr);
1419 if (mode_is_reference(get_irn_mode(left)))
1420 return find_entity(left);
1421 right = get_binop_right(ptr);
1422 if (mode_is_reference(get_irn_mode(right)))
1423 return find_entity(right);
1434 * @param store the Store node
1436 static unsigned optimize_store(ir_node *store)
1442 if (get_Store_volatility(store) == volatility_is_volatile)
1445 ptr = get_Store_ptr(store);
1446 entity = find_entity(ptr);
1448 /* a store to an entity which is never read is unnecessary */
1449 if (entity != NULL && !(get_entity_usage(entity) & ir_usage_read)) {
1450 ldst_info_t *info = get_irn_link(store);
1451 if (info->projs[pn_Store_X_except] == NULL) {
1452 DB((dbg, LEVEL_1, " Killing useless %+F to never read entity %+F\n", store, entity));
1453 exchange(info->projs[pn_Store_M], get_Store_mem(store));
1455 reduce_adr_usage(ptr);
1460 /* Check, if the address of this Store is used more than once.
1461 * If not, this Store cannot be removed in any case. */
1462 if (get_irn_n_uses(ptr) <= 1)
1465 mem = get_Store_mem(store);
1467 /* follow the memory chain as long as there are only Loads */
1470 return follow_Mem_chain_for_Store(store, skip_Proj(mem));
1471 } /* optimize_store */
1474 * walker, optimizes Phi after Stores to identical places:
1475 * Does the following optimization:
1478 * val1 val2 val3 val1 val2 val3
1480 * Store Store Store \ | /
1487 * This reduces the number of stores and allows for predicated execution.
1488 * Moves Stores back to the end of a function which may be bad.
1490 * This is only possible if the predecessor blocks have only one successor.
1492 static unsigned optimize_phi(ir_node *phi, walk_env_t *wenv)
1495 ir_node *store, *old_store, *ptr, *block, *phi_block, *phiM, *phiD, *exc, *projM;
1497 ir_node **inM, **inD, **projMs;
1499 dbg_info *db = NULL;
1501 block_info_t *bl_info;
1504 /* Must be a memory Phi */
1505 if (get_irn_mode(phi) != mode_M)
1508 n = get_Phi_n_preds(phi);
1512 /* must be only one user */
1513 projM = get_Phi_pred(phi, 0);
1514 if (get_irn_n_edges(projM) != 1)
1517 store = skip_Proj(projM);
1519 if (!is_Store(store))
1522 block = get_nodes_block(store);
1524 /* abort on dead blocks */
1525 if (is_Block_dead(block))
1528 /* check if the block is post dominated by Phi-block
1529 and has no exception exit */
1530 bl_info = get_irn_link(block);
1531 if (bl_info->flags & BLOCK_HAS_EXC)
1534 phi_block = get_nodes_block(phi);
1535 if (! block_strictly_postdominates(phi_block, block))
1538 /* this is the address of the store */
1539 ptr = get_Store_ptr(store);
1540 mode = get_irn_mode(get_Store_value(store));
1541 info = get_irn_link(store);
1542 exc = info->exc_block;
1544 for (i = 1; i < n; ++i) {
1545 ir_node *pred = get_Phi_pred(phi, i);
1547 if (get_irn_n_edges(pred) != 1)
1550 pred = skip_Proj(pred);
1551 if (!is_Store(pred))
1554 if (ptr != get_Store_ptr(pred) || mode != get_irn_mode(get_Store_value(pred)))
1557 info = get_irn_link(pred);
1559 /* check, if all stores have the same exception flow */
1560 if (exc != info->exc_block)
1563 /* abort on dead blocks */
1564 block = get_nodes_block(pred);
1565 if (is_Block_dead(block))
1568 /* check if the block is post dominated by Phi-block
1569 and has no exception exit. Note that block must be different from
1570 Phi-block, else we would move a Store from end End of a block to its
1572 bl_info = get_irn_link(block);
1573 if (bl_info->flags & BLOCK_HAS_EXC)
1575 if (block == phi_block || ! block_postdominates(phi_block, block))
1580 * ok, when we are here, we found all predecessors of a Phi that
1581 * are Stores to the same address and size. That means whatever
1582 * we do before we enter the block of the Phi, we do a Store.
1583 * So, we can move the Store to the current block:
1585 * val1 val2 val3 val1 val2 val3
1587 * | Str | | Str | | Str | \ | /
1593 * Is only allowed if the predecessor blocks have only one successor.
1596 NEW_ARR_A(ir_node *, projMs, n);
1597 NEW_ARR_A(ir_node *, inM, n);
1598 NEW_ARR_A(ir_node *, inD, n);
1599 NEW_ARR_A(int, idx, n);
1601 /* Prepare: Collect all Store nodes. We must do this
1602 first because we otherwise may loose a store when exchanging its
1605 for (i = n - 1; i >= 0; --i) {
1608 projMs[i] = get_Phi_pred(phi, i);
1609 assert(is_Proj(projMs[i]));
1611 store = get_Proj_pred(projMs[i]);
1612 info = get_irn_link(store);
1614 inM[i] = get_Store_mem(store);
1615 inD[i] = get_Store_value(store);
1616 idx[i] = info->exc_idx;
1618 block = get_nodes_block(phi);
1620 /* second step: create a new memory Phi */
1621 phiM = new_rd_Phi(get_irn_dbg_info(phi), block, n, inM, mode_M);
1623 /* third step: create a new data Phi */
1624 phiD = new_rd_Phi(get_irn_dbg_info(phi), block, n, inD, mode);
1626 /* rewire memory and kill the node */
1627 for (i = n - 1; i >= 0; --i) {
1628 ir_node *proj = projMs[i];
1630 if (is_Proj(proj)) {
1631 ir_node *store = get_Proj_pred(proj);
1632 exchange(proj, inM[i]);
1637 /* fourth step: create the Store */
1638 store = new_rd_Store(db, block, phiM, ptr, phiD, 0);
1640 co_set_irn_name(store, co_get_irn_ident(old_store));
1643 projM = new_rd_Proj(NULL, block, store, mode_M, pn_Store_M);
1645 info = get_ldst_info(store, &wenv->obst);
1646 info->projs[pn_Store_M] = projM;
1648 /* fifths step: repair exception flow */
1650 ir_node *projX = new_rd_Proj(NULL, block, store, mode_X, pn_Store_X_except);
1652 info->projs[pn_Store_X_except] = projX;
1653 info->exc_block = exc;
1654 info->exc_idx = idx[0];
1656 for (i = 0; i < n; ++i) {
1657 set_Block_cfgpred(exc, idx[i], projX);
1661 /* the exception block should be optimized as some inputs are identical now */
1667 /* sixth step: replace old Phi */
1668 exchange(phi, projM);
1670 return res | DF_CHANGED;
1671 } /* optimize_phi */
1674 * walker, do the optimizations
1676 static void do_load_store_optimize(ir_node *n, void *env)
1678 walk_env_t *wenv = env;
1680 switch (get_irn_opcode(n)) {
1683 wenv->changes |= optimize_load(n);
1687 wenv->changes |= optimize_store(n);
1691 wenv->changes |= optimize_phi(n, wenv);
1697 } /* do_load_store_optimize */
1700 typedef struct scc {
1701 ir_node *head; /**< the head of the list */
1704 /** A node entry. */
1705 typedef struct node_entry {
1706 unsigned DFSnum; /**< the DFS number of this node */
1707 unsigned low; /**< the low number of this node */
1708 int in_stack; /**< flag, set if the node is on the stack */
1709 ir_node *next; /**< link to the next node the the same scc */
1710 scc *pscc; /**< the scc of this node */
1711 unsigned POnum; /**< the post order number for blocks */
1714 /** A loop entry. */
1715 typedef struct loop_env {
1716 ir_phase ph; /**< the phase object */
1717 ir_node **stack; /**< the node stack */
1718 int tos; /**< tos index */
1719 unsigned nextDFSnum; /**< the current DFS number */
1720 unsigned POnum; /**< current post order number */
1722 unsigned changes; /**< a bitmask of graph changes */
1726 * Gets the node_entry of a node
1728 static node_entry *get_irn_ne(ir_node *irn, loop_env *env)
1730 ir_phase *ph = &env->ph;
1731 node_entry *e = phase_get_irn_data(&env->ph, irn);
1734 e = phase_alloc(ph, sizeof(*e));
1735 memset(e, 0, sizeof(*e));
1736 phase_set_irn_data(ph, irn, e);
1742 * Push a node onto the stack.
1744 * @param env the loop environment
1745 * @param n the node to push
1747 static void push(loop_env *env, ir_node *n)
1751 if (env->tos == ARR_LEN(env->stack)) {
1752 int nlen = ARR_LEN(env->stack) * 2;
1753 ARR_RESIZE(ir_node *, env->stack, nlen);
1755 env->stack[env->tos++] = n;
1756 e = get_irn_ne(n, env);
1761 * pop a node from the stack
1763 * @param env the loop environment
1765 * @return The topmost node
1767 static ir_node *pop(loop_env *env)
1769 ir_node *n = env->stack[--env->tos];
1770 node_entry *e = get_irn_ne(n, env);
1777 * Check if irn is a region constant.
1778 * The block or irn must strictly dominate the header block.
1780 * @param irn the node to check
1781 * @param header_block the header block of the induction variable
1783 static int is_rc(ir_node *irn, ir_node *header_block)
1785 ir_node *block = get_nodes_block(irn);
1787 return (block != header_block) && block_dominates(block, header_block);
1790 typedef struct phi_entry phi_entry;
1792 ir_node *phi; /**< A phi with a region const memory. */
1793 int pos; /**< The position of the region const memory */
1794 ir_node *load; /**< the newly created load for this phi */
1799 * An entry in the avail set.
1801 typedef struct avail_entry_t {
1802 ir_node *ptr; /**< the address pointer */
1803 ir_mode *mode; /**< the load mode */
1804 ir_node *load; /**< the associated Load */
1808 * Compare two avail entries.
1810 static int cmp_avail_entry(const void *elt, const void *key, size_t size)
1812 const avail_entry_t *a = elt;
1813 const avail_entry_t *b = key;
1816 return a->ptr != b->ptr || a->mode != b->mode;
1817 } /* cmp_avail_entry */
1820 * Calculate the hash value of an avail entry.
1822 static unsigned hash_cache_entry(const avail_entry_t *entry)
1824 return get_irn_idx(entry->ptr) * 9 + HASH_PTR(entry->mode);
1825 } /* hash_cache_entry */
1828 * Move loops out of loops if possible.
1830 * @param pscc the loop described by an SCC
1831 * @param env the loop environment
1833 static void move_loads_out_of_loops(scc *pscc, loop_env *env)
1835 ir_node *phi, *load, *next, *other, *next_other;
1838 phi_entry *phi_list = NULL;
1841 avail = new_set(cmp_avail_entry, 8);
1843 /* collect all outer memories */
1844 for (phi = pscc->head; phi != NULL; phi = next) {
1845 node_entry *ne = get_irn_ne(phi, env);
1848 /* check all memory Phi's */
1852 assert(get_irn_mode(phi) == mode_M && "DFS return non-memory Phi");
1854 for (j = get_irn_arity(phi) - 1; j >= 0; --j) {
1855 ir_node *pred = get_irn_n(phi, j);
1856 node_entry *pe = get_irn_ne(pred, env);
1858 if (pe->pscc != ne->pscc) {
1859 /* not in the same SCC, is region const */
1860 phi_entry *pe = phase_alloc(&env->ph, sizeof(*pe));
1864 pe->next = phi_list;
1869 /* no Phis no fun */
1870 assert(phi_list != NULL && "DFS found a loop without Phi");
1872 /* for now, we cannot handle more than one input (only reducible cf) */
1873 if (phi_list->next != NULL)
1876 for (load = pscc->head; load; load = next) {
1878 node_entry *ne = get_irn_ne(load, env);
1881 if (is_Load(load)) {
1882 ldst_info_t *info = get_irn_link(load);
1883 ir_node *ptr = get_Load_ptr(load);
1885 /* for now, we cannot handle Loads with exceptions */
1886 if (info->projs[pn_Load_res] == NULL || info->projs[pn_Load_X_regular] != NULL || info->projs[pn_Load_X_except] != NULL)
1889 /* for now, we can only move Load(Global) */
1890 if (! is_Global(ptr))
1892 ent = get_Global_entity(ptr);
1893 load_mode = get_Load_mode(load);
1894 for (other = pscc->head; other != NULL; other = next_other) {
1895 node_entry *ne = get_irn_ne(other, env);
1896 next_other = ne->next;
1898 if (is_Store(other)) {
1899 ir_alias_relation rel = get_alias_relation(
1901 get_Store_ptr(other),
1902 get_irn_mode(get_Store_value(other)),
1904 /* if the might be an alias, we cannot pass this Store */
1905 if (rel != ir_no_alias)
1908 /* only Phis and pure Calls are allowed here, so ignore them */
1910 if (other == NULL) {
1915 /* yep, no aliasing Store found, Load can be moved */
1916 DB((dbg, LEVEL_1, " Found a Load that could be moved: %+F\n", load));
1918 db = get_irn_dbg_info(load);
1919 for (pe = phi_list; pe != NULL; pe = pe->next) {
1921 ir_node *phi = pe->phi;
1922 ir_node *blk = get_nodes_block(phi);
1923 ir_node *pred = get_Block_cfgpred_block(blk, pos);
1925 avail_entry_t entry, *res;
1928 entry.mode = load_mode;
1929 res = set_find(avail, &entry, sizeof(entry), hash_cache_entry(&entry));
1933 irn = new_rd_Load(db, pred, get_Phi_pred(phi, pos), ptr, load_mode, 0);
1935 set_insert(avail, &entry, sizeof(entry), hash_cache_entry(&entry));
1936 DB((dbg, LEVEL_1, " Created %+F in %+F\n", irn, pred));
1939 ninfo = get_ldst_info(irn, phase_obst(&env->ph));
1941 ninfo->projs[pn_Load_M] = mem = new_r_Proj(pred, irn, mode_M, pn_Load_M);
1942 set_Phi_pred(phi, pos, mem);
1944 ninfo->projs[pn_Load_res] = new_r_Proj(pred, irn, load_mode, pn_Load_res);
1947 /* now kill the old Load */
1948 exchange(info->projs[pn_Load_M], get_Load_mem(load));
1949 exchange(info->projs[pn_Load_res], ninfo->projs[pn_Load_res]);
1951 env->changes |= DF_CHANGED;
1956 } /* move_loads_out_of_loops */
1959 * Process a loop SCC.
1961 * @param pscc the SCC
1962 * @param env the loop environment
1964 static void process_loop(scc *pscc, loop_env *env)
1966 ir_node *irn, *next, *header = NULL;
1967 node_entry *b, *h = NULL;
1968 int j, only_phi, num_outside, process = 0;
1971 /* find the header block for this scc */
1972 for (irn = pscc->head; irn; irn = next) {
1973 node_entry *e = get_irn_ne(irn, env);
1974 ir_node *block = get_nodes_block(irn);
1977 b = get_irn_ne(block, env);
1979 if (header != NULL) {
1980 if (h->POnum < b->POnum) {
1990 /* check if this scc contains only Phi, Loads or Stores nodes */
1994 for (irn = pscc->head; irn; irn = next) {
1995 node_entry *e = get_irn_ne(irn, env);
1998 switch (get_irn_opcode(irn)) {
2000 if (is_Call_pure(irn)) {
2001 /* pure calls can be treated like loads */
2005 /* non-pure calls must be handle like may-alias Stores */
2008 /* cannot handle CopyB yet */
2012 if (get_Load_volatility(irn) == volatility_is_volatile) {
2013 /* cannot handle loops with volatile Loads */
2019 if (get_Store_volatility(irn) == volatility_is_volatile) {
2020 /* cannot handle loops with volatile Stores */
2029 for (j = get_irn_arity(irn) - 1; j >= 0; --j) {
2030 ir_node *pred = get_irn_n(irn, j);
2031 node_entry *pe = get_irn_ne(pred, env);
2033 if (pe->pscc != e->pscc) {
2034 /* not in the same SCC, must be a region const */
2035 if (! is_rc(pred, header)) {
2036 /* not a memory loop */
2039 if (out_rc == NULL) {
2040 /* first region constant */
2043 } else if (out_rc != pred) {
2044 /* another region constant */
2055 /* found a memory loop */
2056 DB((dbg, LEVEL_2, " Found a memory loop:\n "));
2057 if (only_phi && num_outside == 1) {
2058 /* a phi cycle with only one real predecessor can be collapsed */
2059 DB((dbg, LEVEL_2, " Found an USELESS Phi cycle:\n "));
2061 for (irn = pscc->head; irn; irn = next) {
2062 node_entry *e = get_irn_ne(irn, env);
2064 exchange(irn, out_rc);
2066 env->changes |= DF_CHANGED;
2070 #ifdef DEBUG_libfirm
2071 for (irn = pscc->head; irn; irn = next) {
2072 node_entry *e = get_irn_ne(irn, env);
2074 DB((dbg, LEVEL_2, " %+F,", irn));
2076 DB((dbg, LEVEL_2, "\n"));
2078 move_loads_out_of_loops(pscc, env);
2082 } /* process_loop */
2087 * @param pscc the SCC
2088 * @param env the loop environment
2090 static void process_scc(scc *pscc, loop_env *env)
2092 ir_node *head = pscc->head;
2093 node_entry *e = get_irn_ne(head, env);
2095 #ifdef DEBUG_libfirm
2097 ir_node *irn, *next;
2099 DB((dbg, LEVEL_4, " SCC at %p:\n ", pscc));
2100 for (irn = pscc->head; irn; irn = next) {
2101 node_entry *e = get_irn_ne(irn, env);
2105 DB((dbg, LEVEL_4, " %+F,", irn));
2107 DB((dbg, LEVEL_4, "\n"));
2111 if (e->next != NULL) {
2112 /* this SCC has more than one member */
2113 process_loop(pscc, env);
2118 * Do Tarjan's SCC algorithm and drive load/store optimization.
2120 * @param irn start at this node
2121 * @param env the loop environment
2123 static void dfs(ir_node *irn, loop_env *env)
2126 node_entry *node = get_irn_ne(irn, env);
2128 mark_irn_visited(irn);
2130 node->DFSnum = env->nextDFSnum++;
2131 node->low = node->DFSnum;
2135 if (is_Phi(irn) || is_Sync(irn)) {
2136 n = get_irn_arity(irn);
2137 for (i = 0; i < n; ++i) {
2138 ir_node *pred = get_irn_n(irn, i);
2139 node_entry *o = get_irn_ne(pred, env);
2141 if (!irn_visited(pred)) {
2143 node->low = MIN(node->low, o->low);
2145 if (o->DFSnum < node->DFSnum && o->in_stack)
2146 node->low = MIN(o->DFSnum, node->low);
2148 } else if (is_fragile_op(irn)) {
2149 ir_node *pred = get_fragile_op_mem(irn);
2150 node_entry *o = get_irn_ne(pred, env);
2152 if (!irn_visited(pred)) {
2154 node->low = MIN(node->low, o->low);
2156 if (o->DFSnum < node->DFSnum && o->in_stack)
2157 node->low = MIN(o->DFSnum, node->low);
2158 } else if (is_Proj(irn)) {
2159 ir_node *pred = get_Proj_pred(irn);
2160 node_entry *o = get_irn_ne(pred, env);
2162 if (!irn_visited(pred)) {
2164 node->low = MIN(node->low, o->low);
2166 if (o->DFSnum < node->DFSnum && o->in_stack)
2167 node->low = MIN(o->DFSnum, node->low);
2170 /* IGNORE predecessors */
2173 if (node->low == node->DFSnum) {
2174 scc *pscc = phase_alloc(&env->ph, sizeof(*pscc));
2182 e = get_irn_ne(x, env);
2184 e->next = pscc->head;
2188 process_scc(pscc, env);
2193 * Do the DFS on the memory edges a graph.
2195 * @param irg the graph to process
2196 * @param env the loop environment
2198 static void do_dfs(ir_graph *irg, loop_env *env)
2200 ir_graph *rem = current_ir_graph;
2201 ir_node *endblk, *end;
2204 current_ir_graph = irg;
2205 inc_irg_visited(irg);
2207 /* visit all memory nodes */
2208 endblk = get_irg_end_block(irg);
2209 for (i = get_Block_n_cfgpreds(endblk) - 1; i >= 0; --i) {
2210 ir_node *pred = get_Block_cfgpred(endblk, i);
2212 pred = skip_Proj(pred);
2213 if (is_Return(pred))
2214 dfs(get_Return_mem(pred), env);
2215 else if (is_Raise(pred))
2216 dfs(get_Raise_mem(pred), env);
2217 else if (is_fragile_op(pred))
2218 dfs(get_fragile_op_mem(pred), env);
2220 assert(0 && "Unknown EndBlock predecessor");
2224 /* visit the keep-alives */
2225 end = get_irg_end(irg);
2226 for (i = get_End_n_keepalives(end) - 1; i >= 0; --i) {
2227 ir_node *ka = get_End_keepalive(end, i);
2229 if (is_Phi(ka) && !irn_visited(ka))
2232 current_ir_graph = rem;
2236 * Initialize new phase data. We do this always explicit, so return NULL here
2238 static void *init_loop_data(ir_phase *ph, const ir_node *irn, void *data)
2244 } /* init_loop_data */
2247 * Optimize Loads/Stores in loops.
2249 * @param irg the graph
2251 static int optimize_loops(ir_graph *irg)
2255 env.stack = NEW_ARR_F(ir_node *, 128);
2260 phase_init(&env.ph, "ldstopt", irg, PHASE_DEFAULT_GROWTH, init_loop_data, NULL);
2262 /* calculate the SCC's and drive loop optimization. */
2265 DEL_ARR_F(env.stack);
2266 phase_free(&env.ph);
2269 } /* optimize_loops */
2272 * do the load store optimization
2274 int optimize_load_store(ir_graph *irg)
2278 FIRM_DBG_REGISTER(dbg, "firm.opt.ldstopt");
2280 assert(get_irg_phase_state(irg) != phase_building);
2281 assert(get_irg_pinned(irg) != op_pin_state_floats &&
2282 "LoadStore optimization needs pinned graph");
2284 /* we need landing pads */
2285 remove_critical_cf_edges(irg);
2289 /* for Phi optimization post-dominators are needed ... */
2290 assure_postdoms(irg);
2292 if (get_opt_alias_analysis()) {
2293 assure_irg_entity_usage_computed(irg);
2294 assure_irp_globals_entity_usage_computed();
2297 obstack_init(&env.obst);
2300 /* init the links, then collect Loads/Stores/Proj's in lists */
2302 irg_walk_graph(irg, firm_clear_link, collect_nodes, &env);
2304 /* now we have collected enough information, optimize */
2305 irg_walk_graph(irg, NULL, do_load_store_optimize, &env);
2307 env.changes |= optimize_loops(irg);
2309 obstack_free(&env.obst, NULL);
2311 /* Handle graph state */
2313 set_irg_outs_inconsistent(irg);
2314 set_irg_entity_usage_state(irg, ir_entity_usage_not_computed);
2317 if (env.changes & CF_CHANGED) {
2318 /* is this really needed: Yes, control flow changed, block might
2319 have Bad() predecessors. */
2320 set_irg_doms_inconsistent(irg);
2322 return env.changes != 0;
2323 } /* optimize_load_store */
2325 ir_graph_pass_t *optimize_load_store_pass(const char *name)
2327 return def_graph_pass_ret(name ? name : "ldst", optimize_load_store);
2328 } /* optimize_load_store_pass */