2 * Copyright (C) 1995-2011 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((long)pn_Load_max, (long)pn_Store_max), (long)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 = (ldst_info_t*)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 = (block_info_t*)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 walk_env_t *wenv = (walk_env_t *)env;
183 unsigned opcode = get_irn_opcode(node);
184 ir_node *pred, *blk, *pred_blk;
185 ldst_info_t *ldst_info;
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 ir_tarval *tlower, *tupper;
275 ir_node *index = get_Sel_index(ptr, i);
276 ir_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) == ir_relation_less)
292 if (tarval_cmp(tupper, tv) == ir_relation_less)
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)) {
389 ir_node *l = get_Add_left(ptr);
390 ir_node *r = get_Add_right(ptr);
391 if (is_Const(r) && get_irn_mode(l) == get_irn_mode(ptr)) {
393 tv = get_Const_tarval(r);
396 tv = get_Const_tarval(l);
400 mode = get_tarval_mode(tv);
403 /* ptr must be a Sel or a SymConst, this was checked in find_constant_entity() */
405 field = get_Sel_entity(ptr);
407 field = get_SymConst_entity(ptr);
410 for (ent = field;;) {
412 ir_tarval *sz, *tv_index, *tlower, *tupper;
415 tp = get_entity_type(ent);
416 if (! is_Array_type(tp))
418 ent = get_array_element_entity(tp);
419 size = get_type_size_bytes(get_entity_type(ent));
420 sz = new_tarval_from_long(size, mode);
422 tv_index = tarval_div(tmp, sz);
423 tmp = tarval_mod(tmp, sz);
425 if (tv_index == tarval_bad || tmp == tarval_bad)
428 assert(get_array_n_dimensions(tp) == 1 && "multiarrays not implemented");
429 bound = get_array_lower_bound(tp, 0);
430 tlower = computed_value(bound);
431 bound = get_array_upper_bound(tp, 0);
432 tupper = computed_value(bound);
434 if (tlower == tarval_bad || tupper == tarval_bad)
437 if (tarval_cmp(tv_index, tlower) == ir_relation_less)
439 if (tarval_cmp(tupper, tv_index) == ir_relation_less)
442 /* ok, bounds check finished */
445 if (! tarval_is_null(tmp)) {
446 /* access to some struct/union member */
450 /* should be at least ONE array */
454 res = rec_get_accessed_path(ptr, depth + idx);
458 path_len = get_compound_graph_path_length(res);
459 pos = path_len - depth - idx;
461 for (ent = field;;) {
463 ir_tarval *sz, *tv_index;
466 tp = get_entity_type(ent);
467 if (! is_Array_type(tp))
469 ent = get_array_element_entity(tp);
470 set_compound_graph_path_node(res, pos, ent);
472 size = get_type_size_bytes(get_entity_type(ent));
473 sz = new_tarval_from_long(size, mode);
475 tv_index = tarval_div(tv, sz);
476 tv = tarval_mod(tv, sz);
478 /* worked above, should work again */
479 assert(tv_index != tarval_bad && tv != tarval_bad);
481 /* bounds already checked above */
482 index = get_tarval_long(tv_index);
483 set_compound_graph_path_array_index(res, pos, index);
486 } else if (is_Sub(ptr)) {
487 ir_node *l = get_Sub_left(ptr);
488 ir_node *r = get_Sub_right(ptr);
491 tv = get_Const_tarval(r);
496 } /* rec_get_accessed_path */
499 * Returns an access path or NULL. The access path is only
500 * valid, if the graph is in phase_high and _no_ address computation is used.
502 static compound_graph_path *get_accessed_path(ir_node *ptr)
504 compound_graph_path *gr = rec_get_accessed_path(ptr, 0);
506 } /* get_accessed_path */
508 typedef struct path_entry {
510 struct path_entry *next;
514 static ir_node *rec_find_compound_ent_value(ir_node *ptr, path_entry *next)
516 path_entry entry, *p;
517 ir_entity *ent, *field;
518 ir_initializer_t *initializer;
524 if (is_SymConst(ptr)) {
526 ent = get_SymConst_entity(ptr);
527 initializer = get_entity_initializer(ent);
528 for (p = next; p != NULL;) {
529 if (initializer->kind != IR_INITIALIZER_COMPOUND)
531 n = get_initializer_compound_n_entries(initializer);
532 tp = get_entity_type(ent);
534 if (is_Array_type(tp)) {
535 ent = get_array_element_entity(tp);
540 initializer = get_initializer_compound_value(initializer, 0);
544 if (p->index >= (int) n)
546 initializer = get_initializer_compound_value(initializer, p->index);
551 tp = get_entity_type(ent);
552 while (is_Array_type(tp)) {
553 ent = get_array_element_entity(tp);
554 tp = get_entity_type(ent);
556 n = get_initializer_compound_n_entries(initializer);
559 initializer = get_initializer_compound_value(initializer, 0);
562 switch (initializer->kind) {
563 case IR_INITIALIZER_CONST:
564 return get_initializer_const_value(initializer);
565 case IR_INITIALIZER_TARVAL:
566 case IR_INITIALIZER_NULL:
570 } else if (is_Sel(ptr)) {
571 entry.ent = field = get_Sel_entity(ptr);
572 tp = get_entity_owner(field);
573 if (is_Array_type(tp)) {
574 assert(get_Sel_n_indexs(ptr) == 1 && "multi dim arrays not implemented");
575 entry.index = get_Sel_array_index_long(ptr, 0) - get_array_lower_bound_int(tp, 0);
577 int i, n_members = get_compound_n_members(tp);
578 for (i = 0; i < n_members; ++i) {
579 if (get_compound_member(tp, i) == field)
582 if (i >= n_members) {
583 /* not found: should NOT happen */
588 return rec_find_compound_ent_value(get_Sel_ptr(ptr), &entry);
589 } else if (is_Add(ptr)) {
594 ir_node *l = get_Add_left(ptr);
595 ir_node *r = get_Add_right(ptr);
598 tv = get_Const_tarval(r);
601 tv = get_Const_tarval(l);
605 mode = get_tarval_mode(tv);
607 /* ptr must be a Sel or a SymConst, this was checked in find_constant_entity() */
609 field = get_Sel_entity(ptr);
611 field = get_SymConst_entity(ptr);
614 /* count needed entries */
616 for (ent = field;;) {
617 tp = get_entity_type(ent);
618 if (! is_Array_type(tp))
620 ent = get_array_element_entity(tp);
623 /* should be at least ONE entry */
627 /* allocate the right number of entries */
628 NEW_ARR_A(path_entry, p, pos);
632 for (ent = field;;) {
634 ir_tarval *sz, *tv_index, *tlower, *tupper;
638 tp = get_entity_type(ent);
639 if (! is_Array_type(tp))
641 ent = get_array_element_entity(tp);
643 p[pos].next = &p[pos + 1];
645 size = get_type_size_bytes(get_entity_type(ent));
646 sz = new_tarval_from_long(size, mode);
648 tv_index = tarval_div(tv, sz);
649 tv = tarval_mod(tv, sz);
651 if (tv_index == tarval_bad || tv == tarval_bad)
654 assert(get_array_n_dimensions(tp) == 1 && "multiarrays not implemented");
655 bound = get_array_lower_bound(tp, 0);
656 tlower = computed_value(bound);
657 bound = get_array_upper_bound(tp, 0);
658 tupper = computed_value(bound);
660 if (tlower == tarval_bad || tupper == tarval_bad)
663 if (tarval_cmp(tv_index, tlower) == ir_relation_less)
665 if (tarval_cmp(tupper, tv_index) == ir_relation_less)
668 /* ok, bounds check finished */
669 index = get_tarval_long(tv_index);
670 p[pos].index = index;
673 if (! tarval_is_null(tv)) {
674 /* hmm, wrong access */
677 p[pos - 1].next = next;
678 return rec_find_compound_ent_value(ptr, p);
679 } else if (is_Sub(ptr)) {
680 ir_node *l = get_Sub_left(ptr);
681 ir_node *r = get_Sub_right(ptr);
684 tv = get_Const_tarval(r);
691 static ir_node *find_compound_ent_value(ir_node *ptr)
693 return rec_find_compound_ent_value(ptr, NULL);
697 static void reduce_adr_usage(ir_node *ptr);
700 * Update a Load that may have lost its users.
702 static void handle_load_update(ir_node *load)
704 ldst_info_t *info = (ldst_info_t*)get_irn_link(load);
706 /* do NOT touch volatile loads for now */
707 if (get_Load_volatility(load) == volatility_is_volatile)
710 if (! info->projs[pn_Load_res] && ! info->projs[pn_Load_X_except]) {
711 ir_node *ptr = get_Load_ptr(load);
712 ir_node *mem = get_Load_mem(load);
714 /* a Load whose value is neither used nor exception checked, remove it */
715 exchange(info->projs[pn_Load_M], mem);
716 if (info->projs[pn_Load_X_regular])
717 exchange(info->projs[pn_Load_X_regular], new_r_Jmp(get_nodes_block(load)));
719 reduce_adr_usage(ptr);
721 } /* handle_load_update */
724 * A use of an address node has vanished. Check if this was a Proj
725 * node and update the counters.
727 static void reduce_adr_usage(ir_node *ptr)
732 if (get_irn_n_edges(ptr) > 0)
735 /* this Proj is dead now */
736 pred = get_Proj_pred(ptr);
738 ldst_info_t *info = (ldst_info_t*)get_irn_link(pred);
739 info->projs[get_Proj_proj(ptr)] = NULL;
741 /* this node lost its result proj, handle that */
742 handle_load_update(pred);
744 } /* reduce_adr_usage */
747 * Check, if an already existing value of mode old_mode can be converted
748 * into the needed one new_mode without loss.
750 static int can_use_stored_value(ir_mode *old_mode, ir_mode *new_mode)
752 if (old_mode == new_mode)
755 /* if both modes are two-complement ones, we can always convert the
756 Stored value into the needed one. */
757 if (get_mode_size_bits(old_mode) >= get_mode_size_bits(new_mode) &&
758 get_mode_arithmetic(old_mode) == irma_twos_complement &&
759 get_mode_arithmetic(new_mode) == irma_twos_complement)
762 } /* can_use_stored_value */
765 * Check whether a Call is at least pure, ie. does only read memory.
767 static unsigned is_Call_pure(ir_node *call)
769 ir_type *call_tp = get_Call_type(call);
770 unsigned prop = get_method_additional_properties(call_tp);
772 /* check first the call type */
773 if ((prop & (mtp_property_const|mtp_property_pure)) == 0) {
774 /* try the called entity */
775 ir_node *ptr = get_Call_ptr(call);
777 if (is_Global(ptr)) {
778 ir_entity *ent = get_Global_entity(ptr);
780 prop = get_entity_additional_properties(ent);
783 return (prop & (mtp_property_const|mtp_property_pure)) != 0;
786 static ir_node *get_base_and_offset(ir_node *ptr, long *pOffset)
788 ir_mode *mode = get_irn_mode(ptr);
791 /* TODO: long might not be enough, we should probably use some tarval thingy... */
794 ir_node *l = get_Add_left(ptr);
795 ir_node *r = get_Add_right(ptr);
797 if (get_irn_mode(l) != mode || !is_Const(r))
800 offset += get_tarval_long(get_Const_tarval(r));
802 } else if (is_Sub(ptr)) {
803 ir_node *l = get_Sub_left(ptr);
804 ir_node *r = get_Sub_right(ptr);
806 if (get_irn_mode(l) != mode || !is_Const(r))
809 offset -= get_tarval_long(get_Const_tarval(r));
811 } else if (is_Sel(ptr)) {
812 ir_entity *ent = get_Sel_entity(ptr);
813 ir_type *tp = get_entity_owner(ent);
815 if (is_Array_type(tp)) {
819 /* only one dimensional arrays yet */
820 if (get_Sel_n_indexs(ptr) != 1)
822 index = get_Sel_index(ptr, 0);
823 if (! is_Const(index))
826 tp = get_entity_type(ent);
827 if (get_type_state(tp) != layout_fixed)
830 size = get_type_size_bytes(tp);
831 offset += size * get_tarval_long(get_Const_tarval(index));
833 if (get_type_state(tp) != layout_fixed)
835 offset += get_entity_offset(ent);
837 ptr = get_Sel_ptr(ptr);
846 static int try_load_after_store(ir_node *load,
847 ir_node *load_base_ptr, long load_offset, ir_node *store)
850 ir_node *store_ptr = get_Store_ptr(store);
852 ir_node *store_base_ptr = get_base_and_offset(store_ptr, &store_offset);
853 ir_node *store_value;
862 if (load_base_ptr != store_base_ptr)
865 load_mode = get_Load_mode(load);
866 load_mode_len = get_mode_size_bytes(load_mode);
867 store_mode = get_irn_mode(get_Store_value(store));
868 store_mode_len = get_mode_size_bytes(store_mode);
869 delta = load_offset - store_offset;
870 store_value = get_Store_value(store);
872 if (delta != 0 || store_mode != load_mode) {
873 if (delta < 0 || delta + load_mode_len > store_mode_len)
876 if (get_mode_arithmetic(store_mode) != irma_twos_complement ||
877 get_mode_arithmetic(load_mode) != irma_twos_complement)
881 /* produce a shift to adjust offset delta */
884 ir_graph *irg = get_irn_irg(load);
886 /* FIXME: only true for little endian */
887 cnst = new_r_Const_long(irg, mode_Iu, delta * 8);
888 store_value = new_r_Shr(get_nodes_block(load),
889 store_value, cnst, store_mode);
892 /* add an convert if needed */
893 if (store_mode != load_mode) {
894 store_value = new_r_Conv(get_nodes_block(load), store_value, load_mode);
898 DBG_OPT_RAW(load, store_value);
900 info = (ldst_info_t*)get_irn_link(load);
901 if (info->projs[pn_Load_M])
902 exchange(info->projs[pn_Load_M], get_Load_mem(load));
906 if (info->projs[pn_Load_X_except]) {
907 ir_graph *irg = get_irn_irg(load);
908 exchange( info->projs[pn_Load_X_except], new_r_Bad(irg));
911 if (info->projs[pn_Load_X_regular]) {
912 exchange( info->projs[pn_Load_X_regular], new_r_Jmp(get_nodes_block(load)));
916 if (info->projs[pn_Load_res])
917 exchange(info->projs[pn_Load_res], store_value);
919 load_ptr = get_Load_ptr(load);
921 reduce_adr_usage(load_ptr);
922 return res | DF_CHANGED;
926 * Follow the memory chain as long as there are only Loads,
927 * alias free Stores, and constant Calls and try to replace the
928 * current Load by a previous ones.
929 * Note that in unreachable loops it might happen that we reach
930 * load again, as well as we can fall into a cycle.
931 * We break such cycles using a special visited flag.
933 * INC_MASTER() must be called before dive into
935 static unsigned follow_Mem_chain(ir_node *load, ir_node *curr)
938 ldst_info_t *info = (ldst_info_t*)get_irn_link(load);
940 ir_node *ptr = get_Load_ptr(load);
941 ir_node *mem = get_Load_mem(load);
942 ir_mode *load_mode = get_Load_mode(load);
944 for (pred = curr; load != pred; ) {
945 ldst_info_t *pred_info = (ldst_info_t*)get_irn_link(pred);
948 * a Load immediately after a Store -- a read after write.
949 * We may remove the Load, if both Load & Store does not have an
950 * exception handler OR they are in the same Block. In the latter
951 * case the Load cannot throw an exception when the previous Store was
954 * Why we need to check for Store Exception? If the Store cannot
955 * be executed (ROM) the exception handler might simply jump into
957 * We could make it a little bit better if we would know that the
958 * exception handler of the Store jumps directly to the end...
960 if (is_Store(pred) && ((pred_info->projs[pn_Store_X_except] == NULL
961 && info->projs[pn_Load_X_except] == NULL)
962 || get_nodes_block(load) == get_nodes_block(pred)))
965 ir_node *base_ptr = get_base_and_offset(ptr, &load_offset);
966 int changes = try_load_after_store(load, base_ptr, load_offset, pred);
969 return res | changes;
970 } else if (is_Load(pred) && get_Load_ptr(pred) == ptr &&
971 can_use_stored_value(get_Load_mode(pred), load_mode)) {
973 * a Load after a Load -- a read after read.
974 * We may remove the second Load, if it does not have an exception
975 * handler OR they are in the same Block. In the later case
976 * the Load cannot throw an exception when the previous Load was
979 * Here, there is no need to check if the previous Load has an
980 * exception hander because they would have exact the same
983 if (info->projs[pn_Load_X_except] == NULL
984 || get_nodes_block(load) == get_nodes_block(pred)) {
987 DBG_OPT_RAR(load, pred);
989 /* the result is used */
990 if (info->projs[pn_Load_res]) {
991 if (pred_info->projs[pn_Load_res] == NULL) {
992 /* create a new Proj again */
993 pred_info->projs[pn_Load_res] = new_r_Proj(pred, get_Load_mode(pred), pn_Load_res);
995 value = pred_info->projs[pn_Load_res];
997 /* add an convert if needed */
998 if (get_Load_mode(pred) != load_mode) {
999 value = new_r_Conv(get_nodes_block(load), value, load_mode);
1002 exchange(info->projs[pn_Load_res], value);
1005 if (info->projs[pn_Load_M])
1006 exchange(info->projs[pn_Load_M], mem);
1009 if (info->projs[pn_Load_X_except]) {
1010 ir_graph *irg = get_irn_irg(load);
1011 exchange(info->projs[pn_Load_X_except], new_r_Bad(irg));
1014 if (info->projs[pn_Load_X_regular]) {
1015 exchange( info->projs[pn_Load_X_regular], new_r_Jmp(get_nodes_block(load)));
1020 reduce_adr_usage(ptr);
1021 return res |= DF_CHANGED;
1025 if (is_Store(pred)) {
1026 /* check if we can pass through this store */
1027 ir_alias_relation rel = get_alias_relation(
1028 get_Store_ptr(pred),
1029 get_irn_mode(get_Store_value(pred)),
1031 /* if the might be an alias, we cannot pass this Store */
1032 if (rel != ir_no_alias)
1034 pred = skip_Proj(get_Store_mem(pred));
1035 } else if (is_Load(pred)) {
1036 pred = skip_Proj(get_Load_mem(pred));
1037 } else if (is_Call(pred)) {
1038 if (is_Call_pure(pred)) {
1039 /* The called graph is at least pure, so there are no Store's
1040 in it. We can handle it like a Load and skip it. */
1041 pred = skip_Proj(get_Call_mem(pred));
1043 /* there might be Store's in the graph, stop here */
1047 /* follow only Load chains */
1051 /* check for cycles */
1052 if (NODE_VISITED(pred_info))
1054 MARK_NODE(pred_info);
1057 if (is_Sync(pred)) {
1060 /* handle all Sync predecessors */
1061 for (i = get_Sync_n_preds(pred) - 1; i >= 0; --i) {
1062 res |= follow_Mem_chain(load, skip_Proj(get_Sync_pred(pred, i)));
1069 } /* follow_Mem_chain */
1072 * Check if we can replace the load by a given const from
1073 * the const code irg.
1075 ir_node *can_replace_load_by_const(const ir_node *load, ir_node *c)
1077 ir_mode *c_mode = get_irn_mode(c);
1078 ir_mode *l_mode = get_Load_mode(load);
1079 ir_node *block = get_nodes_block(load);
1080 dbg_info *dbgi = get_irn_dbg_info(load);
1081 ir_node *res = copy_const_value(dbgi, c, block);
1083 if (c_mode != l_mode) {
1084 /* check, if the mode matches OR can be easily converted info */
1085 if (is_reinterpret_cast(c_mode, l_mode)) {
1086 /* copy the value from the const code irg and cast it */
1087 res = new_rd_Conv(dbgi, block, res, l_mode);
1097 * @param load the Load node
1099 static unsigned optimize_load(ir_node *load)
1101 ldst_info_t *info = (ldst_info_t*)get_irn_link(load);
1102 ir_node *mem, *ptr, *value;
1107 /* do NOT touch volatile loads for now */
1108 if (get_Load_volatility(load) == volatility_is_volatile)
1111 /* the address of the load to be optimized */
1112 ptr = get_Load_ptr(load);
1114 /* The mem of the Load. Must still be returned after optimization. */
1115 mem = get_Load_mem(load);
1117 if (info->projs[pn_Load_res] == NULL
1118 && info->projs[pn_Load_X_except] == NULL) {
1119 /* the value is never used and we don't care about exceptions, remove */
1120 exchange(info->projs[pn_Load_M], mem);
1122 if (info->projs[pn_Load_X_regular]) {
1123 /* should not happen, but if it does, remove it */
1124 exchange(info->projs[pn_Load_X_regular], new_r_Jmp(get_nodes_block(load)));
1128 reduce_adr_usage(ptr);
1129 return res | DF_CHANGED;
1132 /* Load from a constant polymorphic field, where we can resolve
1134 value = transform_polymorph_Load(load);
1135 if (value == load) {
1137 /* check if we can determine the entity that will be loaded */
1138 ent = find_constant_entity(ptr);
1140 && get_entity_visibility(ent) != ir_visibility_external) {
1141 /* a static allocation that is not external: there should be NO
1142 * exception when loading even if we cannot replace the load itself.
1145 /* no exception, clear the info field as it might be checked later again */
1146 if (info->projs[pn_Load_X_except]) {
1147 ir_graph *irg = get_irn_irg(load);
1148 exchange(info->projs[pn_Load_X_except], new_r_Bad(irg));
1149 info->projs[pn_Load_X_except] = NULL;
1152 if (info->projs[pn_Load_X_regular]) {
1153 exchange(info->projs[pn_Load_X_regular], new_r_Jmp(get_nodes_block(load)));
1154 info->projs[pn_Load_X_regular] = NULL;
1158 if (get_entity_linkage(ent) & IR_LINKAGE_CONSTANT) {
1159 if (ent->initializer != NULL) {
1160 /* new style initializer */
1161 value = find_compound_ent_value(ptr);
1162 } else if (entity_has_compound_ent_values(ent)) {
1163 /* old style initializer */
1164 compound_graph_path *path = get_accessed_path(ptr);
1167 assert(is_proper_compound_graph_path(path, get_compound_graph_path_length(path)-1));
1169 value = get_compound_ent_value_by_path(ent, path);
1170 DB((dbg, LEVEL_1, " Constant access at %F%F resulted in %+F\n", ent, path, value));
1171 free_compound_graph_path(path);
1175 value = can_replace_load_by_const(load, value);
1179 if (value != NULL) {
1180 /* we completely replace the load by this value */
1181 if (info->projs[pn_Load_X_except]) {
1182 ir_graph *irg = get_irn_irg(load);
1183 exchange(info->projs[pn_Load_X_except], new_r_Bad(irg));
1184 info->projs[pn_Load_X_except] = NULL;
1187 if (info->projs[pn_Load_X_regular]) {
1188 exchange(info->projs[pn_Load_X_regular], new_r_Jmp(get_nodes_block(load)));
1189 info->projs[pn_Load_X_regular] = NULL;
1192 if (info->projs[pn_Load_M]) {
1193 exchange(info->projs[pn_Load_M], mem);
1196 if (info->projs[pn_Load_res]) {
1197 exchange(info->projs[pn_Load_res], value);
1201 reduce_adr_usage(ptr);
1205 /* Check, if the address of this load is used more than once.
1206 * If not, more load cannot be removed in any case. */
1207 if (get_irn_n_uses(ptr) <= 1 && get_irn_n_uses(get_base_and_offset(ptr, &dummy)) <= 1)
1211 * follow the memory chain as long as there are only Loads
1212 * and try to replace current Load or Store by a previous one.
1213 * Note that in unreachable loops it might happen that we reach
1214 * load again, as well as we can fall into a cycle.
1215 * We break such cycles using a special visited flag.
1218 res = follow_Mem_chain(load, skip_Proj(mem));
1220 } /* optimize_load */
1223 * Check whether a value of mode new_mode would completely overwrite a value
1224 * of mode old_mode in memory.
1226 static int is_completely_overwritten(ir_mode *old_mode, ir_mode *new_mode)
1228 return get_mode_size_bits(new_mode) >= get_mode_size_bits(old_mode);
1229 } /* is_completely_overwritten */
1232 * Check whether small is a part of large (starting at same address).
1234 static int is_partially_same(ir_node *small, ir_node *large)
1236 ir_mode *sm = get_irn_mode(small);
1237 ir_mode *lm = get_irn_mode(large);
1239 /* FIXME: Check endianness */
1240 return is_Conv(small) && get_Conv_op(small) == large
1241 && get_mode_size_bytes(sm) < get_mode_size_bytes(lm)
1242 && get_mode_arithmetic(sm) == irma_twos_complement
1243 && get_mode_arithmetic(lm) == irma_twos_complement;
1244 } /* is_partially_same */
1247 * follow the memory chain as long as there are only Loads and alias free Stores.
1249 * INC_MASTER() must be called before dive into
1251 static unsigned follow_Mem_chain_for_Store(ir_node *store, ir_node *curr)
1254 ldst_info_t *info = (ldst_info_t*)get_irn_link(store);
1256 ir_node *ptr = get_Store_ptr(store);
1257 ir_node *mem = get_Store_mem(store);
1258 ir_node *value = get_Store_value(store);
1259 ir_mode *mode = get_irn_mode(value);
1260 ir_node *block = get_nodes_block(store);
1262 for (pred = curr; pred != store;) {
1263 ldst_info_t *pred_info = (ldst_info_t*)get_irn_link(pred);
1266 * BEWARE: one might think that checking the modes is useless, because
1267 * if the pointers are identical, they refer to the same object.
1268 * This is only true in strong typed languages, not is C were the following
1269 * is possible *(ir_type1 *)p = a; *(ir_type2 *)p = b ...
1270 * However, if the size of the mode that is written is bigger or equal the
1271 * size of the old one, the old value is completely overwritten and can be
1274 if (is_Store(pred) && get_Store_ptr(pred) == ptr &&
1275 get_nodes_block(pred) == block) {
1277 * a Store after a Store in the same Block -- a write after write.
1281 * We may remove the first Store, if the old value is completely
1282 * overwritten or the old value is a part of the new value,
1283 * and if it does not have an exception handler.
1285 * TODO: What, if both have the same exception handler ???
1287 if (get_Store_volatility(pred) != volatility_is_volatile
1288 && !pred_info->projs[pn_Store_X_except]) {
1289 ir_node *predvalue = get_Store_value(pred);
1290 ir_mode *predmode = get_irn_mode(predvalue);
1292 if (is_completely_overwritten(predmode, mode)
1293 || is_partially_same(predvalue, value)) {
1294 DBG_OPT_WAW(pred, store);
1295 exchange(pred_info->projs[pn_Store_M], get_Store_mem(pred));
1297 reduce_adr_usage(ptr);
1303 * We may remove the Store, if the old value already contains
1304 * the new value, and if it does not have an exception handler.
1306 * TODO: What, if both have the same exception handler ???
1308 if (get_Store_volatility(store) != volatility_is_volatile
1309 && !info->projs[pn_Store_X_except]) {
1310 ir_node *predvalue = get_Store_value(pred);
1312 if (is_partially_same(value, predvalue)) {
1313 DBG_OPT_WAW(pred, store);
1314 exchange(info->projs[pn_Store_M], mem);
1316 reduce_adr_usage(ptr);
1320 } else if (is_Load(pred) && get_Load_ptr(pred) == ptr &&
1321 value == pred_info->projs[pn_Load_res]) {
1323 * a Store of a value just loaded from the same address
1324 * -- a write after read.
1325 * We may remove the Store, if it does not have an exception
1328 if (! info->projs[pn_Store_X_except]) {
1329 DBG_OPT_WAR(store, pred);
1330 exchange(info->projs[pn_Store_M], mem);
1332 reduce_adr_usage(ptr);
1337 if (is_Store(pred)) {
1338 /* check if we can pass through this store */
1339 ir_alias_relation rel = get_alias_relation(
1340 get_Store_ptr(pred),
1341 get_irn_mode(get_Store_value(pred)),
1343 /* if the might be an alias, we cannot pass this Store */
1344 if (rel != ir_no_alias)
1346 pred = skip_Proj(get_Store_mem(pred));
1347 } else if (is_Load(pred)) {
1348 ir_alias_relation rel = get_alias_relation(
1349 get_Load_ptr(pred), get_Load_mode(pred),
1351 if (rel != ir_no_alias)
1354 pred = skip_Proj(get_Load_mem(pred));
1356 /* follow only Load chains */
1360 /* check for cycles */
1361 if (NODE_VISITED(pred_info))
1363 MARK_NODE(pred_info);
1366 if (is_Sync(pred)) {
1369 /* handle all Sync predecessors */
1370 for (i = get_Sync_n_preds(pred) - 1; i >= 0; --i) {
1371 res |= follow_Mem_chain_for_Store(store, skip_Proj(get_Sync_pred(pred, i)));
1377 } /* follow_Mem_chain_for_Store */
1379 /** find entity used as base for an address calculation */
1380 static ir_entity *find_entity(ir_node *ptr)
1382 switch (get_irn_opcode(ptr)) {
1384 return get_SymConst_entity(ptr);
1386 ir_node *pred = get_Sel_ptr(ptr);
1387 if (get_irg_frame(get_irn_irg(ptr)) == pred)
1388 return get_Sel_entity(ptr);
1390 return find_entity(pred);
1394 ir_node *left = get_binop_left(ptr);
1396 if (mode_is_reference(get_irn_mode(left)))
1397 return find_entity(left);
1398 right = get_binop_right(ptr);
1399 if (mode_is_reference(get_irn_mode(right)))
1400 return find_entity(right);
1411 * @param store the Store node
1413 static unsigned optimize_store(ir_node *store)
1419 if (get_Store_volatility(store) == volatility_is_volatile)
1422 ptr = get_Store_ptr(store);
1423 entity = find_entity(ptr);
1425 /* a store to an entity which is never read is unnecessary */
1426 if (entity != NULL && !(get_entity_usage(entity) & ir_usage_read)) {
1427 ldst_info_t *info = (ldst_info_t*)get_irn_link(store);
1428 if (info->projs[pn_Store_X_except] == NULL) {
1429 DB((dbg, LEVEL_1, " Killing useless %+F to never read entity %+F\n", store, entity));
1430 exchange(info->projs[pn_Store_M], get_Store_mem(store));
1432 reduce_adr_usage(ptr);
1437 /* Check, if the address of this Store is used more than once.
1438 * If not, this Store cannot be removed in any case. */
1439 if (get_irn_n_uses(ptr) <= 1)
1442 mem = get_Store_mem(store);
1444 /* follow the memory chain as long as there are only Loads */
1447 return follow_Mem_chain_for_Store(store, skip_Proj(mem));
1448 } /* optimize_store */
1451 * walker, optimizes Phi after Stores to identical places:
1452 * Does the following optimization:
1455 * val1 val2 val3 val1 val2 val3
1457 * Store Store Store \ | /
1464 * This reduces the number of stores and allows for predicated execution.
1465 * Moves Stores back to the end of a function which may be bad.
1467 * This is only possible if the predecessor blocks have only one successor.
1469 static unsigned optimize_phi(ir_node *phi, walk_env_t *wenv)
1472 ir_node *store, *old_store, *ptr, *block, *phi_block, *phiM, *phiD, *exc, *projM;
1474 ir_node **inM, **inD, **projMs;
1476 dbg_info *db = NULL;
1478 block_info_t *bl_info;
1481 /* Must be a memory Phi */
1482 if (get_irn_mode(phi) != mode_M)
1485 n = get_Phi_n_preds(phi);
1489 /* must be only one user */
1490 projM = get_Phi_pred(phi, 0);
1491 if (get_irn_n_edges(projM) != 1)
1494 store = skip_Proj(projM);
1496 if (!is_Store(store))
1499 block = get_nodes_block(store);
1501 /* abort on dead blocks */
1502 if (is_Block_dead(block))
1505 /* check if the block is post dominated by Phi-block
1506 and has no exception exit */
1507 bl_info = (block_info_t*)get_irn_link(block);
1508 if (bl_info->flags & BLOCK_HAS_EXC)
1511 phi_block = get_nodes_block(phi);
1512 if (! block_strictly_postdominates(phi_block, block))
1515 /* this is the address of the store */
1516 ptr = get_Store_ptr(store);
1517 mode = get_irn_mode(get_Store_value(store));
1518 info = (ldst_info_t*)get_irn_link(store);
1519 exc = info->exc_block;
1521 for (i = 1; i < n; ++i) {
1522 ir_node *pred = get_Phi_pred(phi, i);
1524 if (get_irn_n_edges(pred) != 1)
1527 pred = skip_Proj(pred);
1528 if (!is_Store(pred))
1531 if (ptr != get_Store_ptr(pred) || mode != get_irn_mode(get_Store_value(pred)))
1534 info = (ldst_info_t*)get_irn_link(pred);
1536 /* check, if all stores have the same exception flow */
1537 if (exc != info->exc_block)
1540 /* abort on dead blocks */
1541 block = get_nodes_block(pred);
1542 if (is_Block_dead(block))
1545 /* check if the block is post dominated by Phi-block
1546 and has no exception exit. Note that block must be different from
1547 Phi-block, else we would move a Store from end End of a block to its
1549 bl_info = (block_info_t*)get_irn_link(block);
1550 if (bl_info->flags & BLOCK_HAS_EXC)
1552 if (block == phi_block || ! block_postdominates(phi_block, block))
1557 * ok, when we are here, we found all predecessors of a Phi that
1558 * are Stores to the same address and size. That means whatever
1559 * we do before we enter the block of the Phi, we do a Store.
1560 * So, we can move the Store to the current block:
1562 * val1 val2 val3 val1 val2 val3
1564 * | Str | | Str | | Str | \ | /
1570 * Is only allowed if the predecessor blocks have only one successor.
1573 NEW_ARR_A(ir_node *, projMs, n);
1574 NEW_ARR_A(ir_node *, inM, n);
1575 NEW_ARR_A(ir_node *, inD, n);
1576 NEW_ARR_A(int, idx, n);
1578 /* Prepare: Collect all Store nodes. We must do this
1579 first because we otherwise may loose a store when exchanging its
1582 for (i = n - 1; i >= 0; --i) {
1585 projMs[i] = get_Phi_pred(phi, i);
1586 assert(is_Proj(projMs[i]));
1588 store = get_Proj_pred(projMs[i]);
1589 info = (ldst_info_t*)get_irn_link(store);
1591 inM[i] = get_Store_mem(store);
1592 inD[i] = get_Store_value(store);
1593 idx[i] = info->exc_idx;
1595 block = get_nodes_block(phi);
1597 /* second step: create a new memory Phi */
1598 phiM = new_rd_Phi(get_irn_dbg_info(phi), block, n, inM, mode_M);
1600 /* third step: create a new data Phi */
1601 phiD = new_rd_Phi(get_irn_dbg_info(phi), block, n, inD, mode);
1603 /* rewire memory and kill the node */
1604 for (i = n - 1; i >= 0; --i) {
1605 ir_node *proj = projMs[i];
1607 if (is_Proj(proj)) {
1608 ir_node *store = get_Proj_pred(proj);
1609 exchange(proj, inM[i]);
1614 /* fourth step: create the Store */
1615 store = new_rd_Store(db, block, phiM, ptr, phiD, cons_none);
1617 co_set_irn_name(store, co_get_irn_ident(old_store));
1620 projM = new_rd_Proj(NULL, store, mode_M, pn_Store_M);
1622 info = get_ldst_info(store, &wenv->obst);
1623 info->projs[pn_Store_M] = projM;
1625 /* fifths step: repair exception flow */
1627 ir_node *projX = new_rd_Proj(NULL, store, mode_X, pn_Store_X_except);
1629 info->projs[pn_Store_X_except] = projX;
1630 info->exc_block = exc;
1631 info->exc_idx = idx[0];
1633 for (i = 0; i < n; ++i) {
1634 set_Block_cfgpred(exc, idx[i], projX);
1638 /* the exception block should be optimized as some inputs are identical now */
1644 /* sixth step: replace old Phi */
1645 exchange(phi, projM);
1647 return res | DF_CHANGED;
1648 } /* optimize_phi */
1651 * walker, do the optimizations
1653 static void do_load_store_optimize(ir_node *n, void *env)
1655 walk_env_t *wenv = (walk_env_t*)env;
1657 switch (get_irn_opcode(n)) {
1660 wenv->changes |= optimize_load(n);
1664 wenv->changes |= optimize_store(n);
1668 wenv->changes |= optimize_phi(n, wenv);
1674 } /* do_load_store_optimize */
1677 typedef struct scc {
1678 ir_node *head; /**< the head of the list */
1681 /** A node entry. */
1682 typedef struct node_entry {
1683 unsigned DFSnum; /**< the DFS number of this node */
1684 unsigned low; /**< the low number of this node */
1685 int in_stack; /**< flag, set if the node is on the stack */
1686 ir_node *next; /**< link to the next node the the same scc */
1687 scc *pscc; /**< the scc of this node */
1688 unsigned POnum; /**< the post order number for blocks */
1691 /** A loop entry. */
1692 typedef struct loop_env {
1693 ir_phase ph; /**< the phase object */
1694 ir_node **stack; /**< the node stack */
1695 size_t tos; /**< tos index */
1696 unsigned nextDFSnum; /**< the current DFS number */
1697 unsigned POnum; /**< current post order number */
1699 unsigned changes; /**< a bitmask of graph changes */
1703 * Gets the node_entry of a node
1705 static node_entry *get_irn_ne(ir_node *irn, loop_env *env)
1707 ir_phase *ph = &env->ph;
1708 node_entry *e = (node_entry*)phase_get_irn_data(&env->ph, irn);
1711 e = (node_entry*)phase_alloc(ph, sizeof(*e));
1712 memset(e, 0, sizeof(*e));
1713 phase_set_irn_data(ph, irn, e);
1719 * Push a node onto the stack.
1721 * @param env the loop environment
1722 * @param n the node to push
1724 static void push(loop_env *env, ir_node *n)
1728 if (env->tos == ARR_LEN(env->stack)) {
1729 size_t nlen = ARR_LEN(env->stack) * 2;
1730 ARR_RESIZE(ir_node *, env->stack, nlen);
1732 env->stack[env->tos++] = n;
1733 e = get_irn_ne(n, env);
1738 * pop a node from the stack
1740 * @param env the loop environment
1742 * @return The topmost node
1744 static ir_node *pop(loop_env *env)
1746 ir_node *n = env->stack[--env->tos];
1747 node_entry *e = get_irn_ne(n, env);
1754 * Check if irn is a region constant.
1755 * The block or irn must strictly dominate the header block.
1757 * @param irn the node to check
1758 * @param header_block the header block of the induction variable
1760 static int is_rc(ir_node *irn, ir_node *header_block)
1762 ir_node *block = get_nodes_block(irn);
1764 return (block != header_block) && block_dominates(block, header_block);
1767 typedef struct phi_entry phi_entry;
1769 ir_node *phi; /**< A phi with a region const memory. */
1770 int pos; /**< The position of the region const memory */
1771 ir_node *load; /**< the newly created load for this phi */
1776 * An entry in the avail set.
1778 typedef struct avail_entry_t {
1779 ir_node *ptr; /**< the address pointer */
1780 ir_mode *mode; /**< the load mode */
1781 ir_node *load; /**< the associated Load */
1785 * Compare two avail entries.
1787 static int cmp_avail_entry(const void *elt, const void *key, size_t size)
1789 const avail_entry_t *a = (const avail_entry_t*)elt;
1790 const avail_entry_t *b = (const avail_entry_t*)key;
1793 return a->ptr != b->ptr || a->mode != b->mode;
1794 } /* cmp_avail_entry */
1797 * Calculate the hash value of an avail entry.
1799 static unsigned hash_cache_entry(const avail_entry_t *entry)
1801 return get_irn_idx(entry->ptr) * 9 + HASH_PTR(entry->mode);
1802 } /* hash_cache_entry */
1805 * Move loops out of loops if possible.
1807 * @param pscc the loop described by an SCC
1808 * @param env the loop environment
1810 static void move_loads_out_of_loops(scc *pscc, loop_env *env)
1812 ir_node *phi, *load, *next, *other, *next_other;
1815 phi_entry *phi_list = NULL;
1818 avail = new_set(cmp_avail_entry, 8);
1820 /* collect all outer memories */
1821 for (phi = pscc->head; phi != NULL; phi = next) {
1822 node_entry *ne = get_irn_ne(phi, env);
1825 /* check all memory Phi's */
1829 assert(get_irn_mode(phi) == mode_M && "DFS return non-memory Phi");
1831 for (j = get_irn_arity(phi) - 1; j >= 0; --j) {
1832 ir_node *pred = get_irn_n(phi, j);
1833 node_entry *pe = get_irn_ne(pred, env);
1835 if (pe->pscc != ne->pscc) {
1836 /* not in the same SCC, is region const */
1837 phi_entry *pe = (phi_entry*)phase_alloc(&env->ph, sizeof(*pe));
1841 pe->next = phi_list;
1846 /* no Phis no fun */
1847 assert(phi_list != NULL && "DFS found a loop without Phi");
1849 /* for now, we cannot handle more than one input (only reducible cf) */
1850 if (phi_list->next != NULL)
1853 for (load = pscc->head; load; load = next) {
1855 node_entry *ne = get_irn_ne(load, env);
1858 if (is_Load(load)) {
1859 ldst_info_t *info = (ldst_info_t*)get_irn_link(load);
1860 ir_node *ptr = get_Load_ptr(load);
1862 /* for now, we cannot handle Loads with exceptions */
1863 if (info->projs[pn_Load_res] == NULL || info->projs[pn_Load_X_regular] != NULL || info->projs[pn_Load_X_except] != NULL)
1866 /* for now, we can only move Load(Global) */
1867 if (! is_Global(ptr))
1869 ent = get_Global_entity(ptr);
1870 load_mode = get_Load_mode(load);
1871 for (other = pscc->head; other != NULL; other = next_other) {
1872 node_entry *ne = get_irn_ne(other, env);
1873 next_other = ne->next;
1875 if (is_Store(other)) {
1876 ir_alias_relation rel = get_alias_relation(
1877 get_Store_ptr(other),
1878 get_irn_mode(get_Store_value(other)),
1880 /* if the might be an alias, we cannot pass this Store */
1881 if (rel != ir_no_alias)
1884 /* only Phis and pure Calls are allowed here, so ignore them */
1886 if (other == NULL) {
1887 ldst_info_t *ninfo = NULL;
1891 /* yep, no aliasing Store found, Load can be moved */
1892 DB((dbg, LEVEL_1, " Found a Load that could be moved: %+F\n", load));
1894 db = get_irn_dbg_info(load);
1895 for (pe = phi_list; pe != NULL; pe = pe->next) {
1897 ir_node *phi = pe->phi;
1898 ir_node *blk = get_nodes_block(phi);
1899 ir_node *pred = get_Block_cfgpred_block(blk, pos);
1901 avail_entry_t entry, *res;
1904 entry.mode = load_mode;
1905 res = (avail_entry_t*)set_find(avail, &entry, sizeof(entry), hash_cache_entry(&entry));
1909 irn = new_rd_Load(db, pred, get_Phi_pred(phi, pos), ptr, load_mode, cons_none);
1911 set_insert(avail, &entry, sizeof(entry), hash_cache_entry(&entry));
1912 DB((dbg, LEVEL_1, " Created %+F in %+F\n", irn, pred));
1915 ninfo = get_ldst_info(irn, phase_obst(&env->ph));
1917 ninfo->projs[pn_Load_M] = mem = new_r_Proj(irn, mode_M, pn_Load_M);
1918 set_Phi_pred(phi, pos, mem);
1920 ninfo->projs[pn_Load_res] = new_r_Proj(irn, load_mode, pn_Load_res);
1923 /* now kill the old Load */
1924 exchange(info->projs[pn_Load_M], get_Load_mem(load));
1925 exchange(info->projs[pn_Load_res], ninfo->projs[pn_Load_res]);
1927 env->changes |= DF_CHANGED;
1932 } /* move_loads_out_of_loops */
1935 * Process a loop SCC.
1937 * @param pscc the SCC
1938 * @param env the loop environment
1940 static void process_loop(scc *pscc, loop_env *env)
1942 ir_node *irn, *next, *header = NULL;
1943 node_entry *b, *h = NULL;
1944 int j, only_phi, num_outside, process = 0;
1947 /* find the header block for this scc */
1948 for (irn = pscc->head; irn; irn = next) {
1949 node_entry *e = get_irn_ne(irn, env);
1950 ir_node *block = get_nodes_block(irn);
1953 b = get_irn_ne(block, env);
1955 if (header != NULL) {
1956 if (h->POnum < b->POnum) {
1966 /* check if this scc contains only Phi, Loads or Stores nodes */
1970 for (irn = pscc->head; irn; irn = next) {
1971 node_entry *e = get_irn_ne(irn, env);
1974 switch (get_irn_opcode(irn)) {
1976 if (is_Call_pure(irn)) {
1977 /* pure calls can be treated like loads */
1981 /* non-pure calls must be handle like may-alias Stores */
1984 /* cannot handle CopyB yet */
1988 if (get_Load_volatility(irn) == volatility_is_volatile) {
1989 /* cannot handle loops with volatile Loads */
1995 if (get_Store_volatility(irn) == volatility_is_volatile) {
1996 /* cannot handle loops with volatile Stores */
2005 for (j = get_irn_arity(irn) - 1; j >= 0; --j) {
2006 ir_node *pred = get_irn_n(irn, j);
2007 node_entry *pe = get_irn_ne(pred, env);
2009 if (pe->pscc != e->pscc) {
2010 /* not in the same SCC, must be a region const */
2011 if (! is_rc(pred, header)) {
2012 /* not a memory loop */
2015 if (out_rc == NULL) {
2016 /* first region constant */
2019 } else if (out_rc != pred) {
2020 /* another region constant */
2031 /* found a memory loop */
2032 DB((dbg, LEVEL_2, " Found a memory loop:\n "));
2033 if (only_phi && num_outside == 1) {
2034 /* a phi cycle with only one real predecessor can be collapsed */
2035 DB((dbg, LEVEL_2, " Found an USELESS Phi cycle:\n "));
2037 for (irn = pscc->head; irn; irn = next) {
2038 node_entry *e = get_irn_ne(irn, env);
2040 exchange(irn, out_rc);
2042 env->changes |= DF_CHANGED;
2046 #ifdef DEBUG_libfirm
2047 for (irn = pscc->head; irn; irn = next) {
2048 node_entry *e = get_irn_ne(irn, env);
2050 DB((dbg, LEVEL_2, " %+F,", irn));
2052 DB((dbg, LEVEL_2, "\n"));
2054 move_loads_out_of_loops(pscc, env);
2058 } /* process_loop */
2063 * @param pscc the SCC
2064 * @param env the loop environment
2066 static void process_scc(scc *pscc, loop_env *env)
2068 ir_node *head = pscc->head;
2069 node_entry *e = get_irn_ne(head, env);
2071 #ifdef DEBUG_libfirm
2073 ir_node *irn, *next;
2075 DB((dbg, LEVEL_4, " SCC at %p:\n ", pscc));
2076 for (irn = pscc->head; irn; irn = next) {
2077 node_entry *e = get_irn_ne(irn, env);
2081 DB((dbg, LEVEL_4, " %+F,", irn));
2083 DB((dbg, LEVEL_4, "\n"));
2087 if (e->next != NULL) {
2088 /* this SCC has more than one member */
2089 process_loop(pscc, env);
2094 * Do Tarjan's SCC algorithm and drive load/store optimization.
2096 * @param irn start at this node
2097 * @param env the loop environment
2099 static void dfs(ir_node *irn, loop_env *env)
2102 node_entry *node = get_irn_ne(irn, env);
2104 mark_irn_visited(irn);
2106 node->DFSnum = env->nextDFSnum++;
2107 node->low = node->DFSnum;
2111 if (is_Phi(irn) || is_Sync(irn)) {
2112 n = get_irn_arity(irn);
2113 for (i = 0; i < n; ++i) {
2114 ir_node *pred = get_irn_n(irn, i);
2115 node_entry *o = get_irn_ne(pred, env);
2117 if (!irn_visited(pred)) {
2119 node->low = MIN(node->low, o->low);
2121 if (o->DFSnum < node->DFSnum && o->in_stack)
2122 node->low = MIN(o->DFSnum, node->low);
2124 } else if (is_fragile_op(irn)) {
2125 ir_node *pred = get_fragile_op_mem(irn);
2126 node_entry *o = get_irn_ne(pred, env);
2128 if (!irn_visited(pred)) {
2130 node->low = MIN(node->low, o->low);
2132 if (o->DFSnum < node->DFSnum && o->in_stack)
2133 node->low = MIN(o->DFSnum, node->low);
2134 } else if (is_Proj(irn)) {
2135 ir_node *pred = get_Proj_pred(irn);
2136 node_entry *o = get_irn_ne(pred, env);
2138 if (!irn_visited(pred)) {
2140 node->low = MIN(node->low, o->low);
2142 if (o->DFSnum < node->DFSnum && o->in_stack)
2143 node->low = MIN(o->DFSnum, node->low);
2146 /* IGNORE predecessors */
2149 if (node->low == node->DFSnum) {
2150 scc *pscc = (scc*)phase_alloc(&env->ph, sizeof(*pscc));
2158 e = get_irn_ne(x, env);
2160 e->next = pscc->head;
2164 process_scc(pscc, env);
2169 * Do the DFS on the memory edges a graph.
2171 * @param irg the graph to process
2172 * @param env the loop environment
2174 static void do_dfs(ir_graph *irg, loop_env *env)
2176 ir_node *endblk, *end;
2179 inc_irg_visited(irg);
2181 /* visit all memory nodes */
2182 endblk = get_irg_end_block(irg);
2183 for (i = get_Block_n_cfgpreds(endblk) - 1; i >= 0; --i) {
2184 ir_node *pred = get_Block_cfgpred(endblk, i);
2186 pred = skip_Proj(pred);
2187 if (is_Return(pred))
2188 dfs(get_Return_mem(pred), env);
2189 else if (is_Raise(pred))
2190 dfs(get_Raise_mem(pred), env);
2191 else if (is_fragile_op(pred))
2192 dfs(get_fragile_op_mem(pred), env);
2193 else if (is_Bad(pred))
2194 /* ignore non-optimized block predecessor */;
2196 assert(0 && "Unknown EndBlock predecessor");
2200 /* visit the keep-alives */
2201 end = get_irg_end(irg);
2202 for (i = get_End_n_keepalives(end) - 1; i >= 0; --i) {
2203 ir_node *ka = get_End_keepalive(end, i);
2205 if (is_Phi(ka) && !irn_visited(ka))
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 */