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 #include "opt_manage.h"
55 /** The debug handle. */
56 DEBUG_ONLY(static firm_dbg_module_t *dbg;)
59 #define IMAX(a,b) ((a) > (b) ? (a) : (b))
61 #define MAX_PROJ IMAX(IMAX((long)pn_Load_max, (long)pn_Store_max), (long)pn_Call_max)
64 DF_CHANGED = 1, /**< data flow changed */
65 CF_CHANGED = 2, /**< control flow changed */
71 typedef struct walk_env_t {
72 struct obstack obst; /**< list of all stores */
73 unsigned changes; /**< a bitmask of graph changes */
76 /** A Load/Store info. */
77 typedef struct ldst_info_t {
78 ir_node *projs[MAX_PROJ+1]; /**< list of Proj's of this node */
79 ir_node *exc_block; /**< the exception block if available */
80 int exc_idx; /**< predecessor index in the exception block */
81 unsigned visited; /**< visited counter for breaking loops */
85 * flags for control flow.
88 BLOCK_HAS_COND = 1, /**< Block has conditional control flow */
89 BLOCK_HAS_EXC = 2 /**< Block has exceptional control flow */
95 typedef struct block_info_t {
96 unsigned flags; /**< flags for the block */
99 /** the master visited flag for loop detection. */
100 static unsigned master_visited = 0;
102 #define INC_MASTER() ++master_visited
103 #define MARK_NODE(info) (info)->visited = master_visited
104 #define NODE_VISITED(info) (info)->visited >= master_visited
107 * get the Load/Store info of a node
109 static ldst_info_t *get_ldst_info(ir_node *node, struct obstack *obst)
111 ldst_info_t *info = (ldst_info_t*)get_irn_link(node);
114 info = OALLOCZ(obst, ldst_info_t);
115 set_irn_link(node, info);
118 } /* get_ldst_info */
121 * get the Block info of a node
123 static block_info_t *get_block_info(ir_node *node, struct obstack *obst)
125 block_info_t *info = (block_info_t*)get_irn_link(node);
128 info = OALLOCZ(obst, block_info_t);
129 set_irn_link(node, info);
132 } /* get_block_info */
135 * update the projection info for a Load/Store
137 static unsigned update_projs(ldst_info_t *info, ir_node *proj)
139 long nr = get_Proj_proj(proj);
141 assert(0 <= nr && nr <= MAX_PROJ && "Wrong proj from LoadStore");
143 if (info->projs[nr]) {
144 /* there is already one, do CSE */
145 exchange(proj, info->projs[nr]);
149 info->projs[nr] = proj;
155 * update the exception block info for a Load/Store node.
157 * @param info the load/store info struct
158 * @param block the exception handler block for this load/store
159 * @param pos the control flow input of the block
161 static unsigned update_exc(ldst_info_t *info, ir_node *block, int pos)
163 assert(info->exc_block == NULL && "more than one exception block found");
165 info->exc_block = block;
170 /** Return the number of uses of an address node */
171 #define get_irn_n_uses(adr) get_irn_n_edges(adr)
174 * walker, collects all Load/Store/Proj nodes
176 * walks from Start -> End
178 static void collect_nodes(ir_node *node, void *env)
180 walk_env_t *wenv = (walk_env_t *)env;
181 unsigned opcode = get_irn_opcode(node);
182 ir_node *pred, *blk, *pred_blk;
183 ldst_info_t *ldst_info;
185 if (opcode == iro_Proj) {
186 pred = get_Proj_pred(node);
187 opcode = get_irn_opcode(pred);
189 if (opcode == iro_Load || opcode == iro_Store || opcode == iro_Call) {
190 ldst_info = get_ldst_info(pred, &wenv->obst);
192 wenv->changes |= update_projs(ldst_info, node);
195 * Place the Proj's to the same block as the
196 * predecessor Load. This is always ok and prevents
197 * "non-SSA" form after optimizations if the Proj
198 * is in a wrong block.
200 blk = get_nodes_block(node);
201 pred_blk = get_nodes_block(pred);
202 if (blk != pred_blk) {
203 wenv->changes |= DF_CHANGED;
204 set_nodes_block(node, pred_blk);
207 } else if (opcode == iro_Block) {
210 for (i = get_Block_n_cfgpreds(node) - 1; i >= 0; --i) {
211 ir_node *pred_block, *proj;
212 block_info_t *bl_info;
215 pred = proj = get_Block_cfgpred(node, i);
218 pred = get_Proj_pred(proj);
219 is_exc = is_x_except_Proj(proj);
222 /* ignore Bad predecessors, they will be removed later */
226 pred_block = get_nodes_block(pred);
227 bl_info = get_block_info(pred_block, &wenv->obst);
229 if (is_fragile_op(pred) && is_exc)
230 bl_info->flags |= BLOCK_HAS_EXC;
231 else if (is_irn_forking(pred))
232 bl_info->flags |= BLOCK_HAS_COND;
234 opcode = get_irn_opcode(pred);
235 if (is_exc && (opcode == iro_Load || opcode == iro_Store || opcode == iro_Call)) {
236 ldst_info = get_ldst_info(pred, &wenv->obst);
238 wenv->changes |= update_exc(ldst_info, node, i);
242 } /* collect_nodes */
245 * Returns an entity if the address ptr points to a constant one.
247 * @param ptr the address
249 * @return an entity or NULL
251 static ir_entity *find_constant_entity(ir_node *ptr)
254 if (is_SymConst(ptr) && get_SymConst_kind(ptr) == symconst_addr_ent) {
255 return get_SymConst_entity(ptr);
256 } else if (is_Sel(ptr)) {
257 ir_entity *ent = get_Sel_entity(ptr);
258 ir_type *tp = get_entity_owner(ent);
260 /* Do not fiddle with polymorphism. */
261 if (is_Class_type(get_entity_owner(ent)) &&
262 ((get_entity_n_overwrites(ent) != 0) ||
263 (get_entity_n_overwrittenby(ent) != 0) ) )
266 if (is_Array_type(tp)) {
270 for (i = 0, n = get_Sel_n_indexs(ptr); i < n; ++i) {
272 ir_tarval *tlower, *tupper;
273 ir_node *index = get_Sel_index(ptr, i);
274 ir_tarval *tv = computed_value(index);
276 /* check if the index is constant */
277 if (tv == tarval_bad)
280 bound = get_array_lower_bound(tp, i);
281 tlower = computed_value(bound);
282 bound = get_array_upper_bound(tp, i);
283 tupper = computed_value(bound);
285 if (tlower == tarval_bad || tupper == tarval_bad)
288 if (tarval_cmp(tv, tlower) == ir_relation_less)
290 if (tarval_cmp(tupper, tv) == ir_relation_less)
293 /* ok, bounds check finished */
297 if (get_entity_linkage(ent) & IR_LINKAGE_CONSTANT)
301 ptr = get_Sel_ptr(ptr);
302 } else if (is_Add(ptr)) {
303 ir_node *l = get_Add_left(ptr);
304 ir_node *r = get_Add_right(ptr);
306 if (get_irn_mode(l) == get_irn_mode(ptr) && is_Const(r))
308 else if (get_irn_mode(r) == get_irn_mode(ptr) && is_Const(l))
313 /* for now, we support only one addition, reassoc should fold all others */
314 if (! is_SymConst(ptr) && !is_Sel(ptr))
316 } else if (is_Sub(ptr)) {
317 ir_node *l = get_Sub_left(ptr);
318 ir_node *r = get_Sub_right(ptr);
320 if (get_irn_mode(l) == get_irn_mode(ptr) && is_Const(r))
324 /* for now, we support only one substraction, reassoc should fold all others */
325 if (! is_SymConst(ptr) && !is_Sel(ptr))
330 } /* find_constant_entity */
333 * Return the Selection index of a Sel node from dimension n
335 static long get_Sel_array_index_long(ir_node *n, int dim)
337 ir_node *index = get_Sel_index(n, dim);
338 assert(is_Const(index));
339 return get_tarval_long(get_Const_tarval(index));
340 } /* get_Sel_array_index_long */
343 * Returns the accessed component graph path for an
344 * node computing an address.
346 * @param ptr the node computing the address
347 * @param depth current depth in steps upward from the root
350 static compound_graph_path *rec_get_accessed_path(ir_node *ptr, size_t depth)
352 compound_graph_path *res = NULL;
353 ir_entity *root, *field, *ent;
354 size_t path_len, pos, idx;
358 if (is_SymConst(ptr)) {
359 /* a SymConst. If the depth is 0, this is an access to a global
360 * entity and we don't need a component path, else we know
361 * at least its length.
363 assert(get_SymConst_kind(ptr) == symconst_addr_ent);
364 root = get_SymConst_entity(ptr);
365 res = (depth == 0) ? NULL : new_compound_graph_path(get_entity_type(root), depth);
366 } else if (is_Sel(ptr)) {
367 /* it's a Sel, go up until we find the root */
368 res = rec_get_accessed_path(get_Sel_ptr(ptr), depth+1);
372 /* fill up the step in the path at the current position */
373 field = get_Sel_entity(ptr);
374 path_len = get_compound_graph_path_length(res);
375 pos = path_len - depth - 1;
376 set_compound_graph_path_node(res, pos, field);
378 if (is_Array_type(get_entity_owner(field))) {
379 assert(get_Sel_n_indexs(ptr) == 1 && "multi dim arrays not implemented");
380 set_compound_graph_path_array_index(res, pos, get_Sel_array_index_long(ptr, 0));
382 } else if (is_Add(ptr)) {
387 ir_node *l = get_Add_left(ptr);
388 ir_node *r = get_Add_right(ptr);
389 if (is_Const(r) && get_irn_mode(l) == get_irn_mode(ptr)) {
391 tv = get_Const_tarval(r);
394 tv = get_Const_tarval(l);
398 mode = get_tarval_mode(tv);
401 /* ptr must be a Sel or a SymConst, this was checked in find_constant_entity() */
403 field = get_Sel_entity(ptr);
405 field = get_SymConst_entity(ptr);
408 for (ent = field;;) {
410 ir_tarval *sz, *tv_index, *tlower, *tupper;
413 tp = get_entity_type(ent);
414 if (! is_Array_type(tp))
416 ent = get_array_element_entity(tp);
417 size = get_type_size_bytes(get_entity_type(ent));
418 sz = new_tarval_from_long(size, mode);
420 tv_index = tarval_div(tmp, sz);
421 tmp = tarval_mod(tmp, sz);
423 if (tv_index == tarval_bad || tmp == tarval_bad)
426 assert(get_array_n_dimensions(tp) == 1 && "multiarrays not implemented");
427 bound = get_array_lower_bound(tp, 0);
428 tlower = computed_value(bound);
429 bound = get_array_upper_bound(tp, 0);
430 tupper = computed_value(bound);
432 if (tlower == tarval_bad || tupper == tarval_bad)
435 if (tarval_cmp(tv_index, tlower) == ir_relation_less)
437 if (tarval_cmp(tupper, tv_index) == ir_relation_less)
440 /* ok, bounds check finished */
443 if (! tarval_is_null(tmp)) {
444 /* access to some struct/union member */
448 /* should be at least ONE array */
452 res = rec_get_accessed_path(ptr, depth + idx);
456 path_len = get_compound_graph_path_length(res);
457 pos = path_len - depth - idx;
459 for (ent = field;;) {
461 ir_tarval *sz, *tv_index;
464 tp = get_entity_type(ent);
465 if (! is_Array_type(tp))
467 ent = get_array_element_entity(tp);
468 set_compound_graph_path_node(res, pos, ent);
470 size = get_type_size_bytes(get_entity_type(ent));
471 sz = new_tarval_from_long(size, mode);
473 tv_index = tarval_div(tv, sz);
474 tv = tarval_mod(tv, sz);
476 /* worked above, should work again */
477 assert(tv_index != tarval_bad && tv != tarval_bad);
479 /* bounds already checked above */
480 index = get_tarval_long(tv_index);
481 set_compound_graph_path_array_index(res, pos, index);
484 } else if (is_Sub(ptr)) {
485 ir_node *l = get_Sub_left(ptr);
486 ir_node *r = get_Sub_right(ptr);
489 tv = get_Const_tarval(r);
494 } /* rec_get_accessed_path */
497 * Returns an access path or NULL. The access path is only
498 * valid, if the graph is in phase_high and _no_ address computation is used.
500 static compound_graph_path *get_accessed_path(ir_node *ptr)
502 compound_graph_path *gr = rec_get_accessed_path(ptr, 0);
504 } /* get_accessed_path */
506 typedef struct path_entry {
508 struct path_entry *next;
512 static ir_node *rec_find_compound_ent_value(ir_node *ptr, path_entry *next)
514 path_entry entry, *p;
515 ir_entity *ent, *field;
516 ir_initializer_t *initializer;
522 if (is_SymConst(ptr)) {
524 ent = get_SymConst_entity(ptr);
525 initializer = get_entity_initializer(ent);
526 for (p = next; p != NULL;) {
527 if (initializer->kind != IR_INITIALIZER_COMPOUND)
529 n = get_initializer_compound_n_entries(initializer);
530 tp = get_entity_type(ent);
532 if (is_Array_type(tp)) {
533 ent = get_array_element_entity(tp);
538 initializer = get_initializer_compound_value(initializer, 0);
544 initializer = get_initializer_compound_value(initializer, p->index);
549 tp = get_entity_type(ent);
550 while (is_Array_type(tp)) {
551 ent = get_array_element_entity(tp);
552 tp = get_entity_type(ent);
554 n = get_initializer_compound_n_entries(initializer);
557 initializer = get_initializer_compound_value(initializer, 0);
560 switch (initializer->kind) {
561 case IR_INITIALIZER_CONST:
562 return get_initializer_const_value(initializer);
563 case IR_INITIALIZER_TARVAL:
564 case IR_INITIALIZER_NULL:
568 } else if (is_Sel(ptr)) {
569 entry.ent = field = get_Sel_entity(ptr);
570 tp = get_entity_owner(field);
571 if (is_Array_type(tp)) {
572 assert(get_Sel_n_indexs(ptr) == 1 && "multi dim arrays not implemented");
573 entry.index = get_Sel_array_index_long(ptr, 0) - get_array_lower_bound_int(tp, 0);
575 size_t i, n_members = get_compound_n_members(tp);
576 for (i = 0; i < n_members; ++i) {
577 if (get_compound_member(tp, i) == field)
580 if (i >= n_members) {
581 /* not found: should NOT happen */
586 return rec_find_compound_ent_value(get_Sel_ptr(ptr), &entry);
587 } else if (is_Add(ptr)) {
592 ir_node *l = get_Add_left(ptr);
593 ir_node *r = get_Add_right(ptr);
596 tv = get_Const_tarval(r);
599 tv = get_Const_tarval(l);
603 mode = get_tarval_mode(tv);
605 /* ptr must be a Sel or a SymConst, this was checked in find_constant_entity() */
607 field = get_Sel_entity(ptr);
609 field = get_SymConst_entity(ptr);
612 /* count needed entries */
614 for (ent = field;;) {
615 tp = get_entity_type(ent);
616 if (! is_Array_type(tp))
618 ent = get_array_element_entity(tp);
621 /* should be at least ONE entry */
625 /* allocate the right number of entries */
626 NEW_ARR_A(path_entry, p, pos);
630 for (ent = field;;) {
632 ir_tarval *sz, *tv_index, *tlower, *tupper;
636 tp = get_entity_type(ent);
637 if (! is_Array_type(tp))
639 ent = get_array_element_entity(tp);
641 p[pos].next = &p[pos + 1];
643 size = get_type_size_bytes(get_entity_type(ent));
644 sz = new_tarval_from_long(size, mode);
646 tv_index = tarval_div(tv, sz);
647 tv = tarval_mod(tv, sz);
649 if (tv_index == tarval_bad || tv == tarval_bad)
652 assert(get_array_n_dimensions(tp) == 1 && "multiarrays not implemented");
653 bound = get_array_lower_bound(tp, 0);
654 tlower = computed_value(bound);
655 bound = get_array_upper_bound(tp, 0);
656 tupper = computed_value(bound);
658 if (tlower == tarval_bad || tupper == tarval_bad)
661 if (tarval_cmp(tv_index, tlower) == ir_relation_less)
663 if (tarval_cmp(tupper, tv_index) == ir_relation_less)
666 /* ok, bounds check finished */
667 index = get_tarval_long(tv_index);
668 p[pos].index = index;
671 if (! tarval_is_null(tv)) {
672 /* hmm, wrong access */
675 p[pos - 1].next = next;
676 return rec_find_compound_ent_value(ptr, p);
677 } else if (is_Sub(ptr)) {
678 ir_node *l = get_Sub_left(ptr);
679 ir_node *r = get_Sub_right(ptr);
682 tv = get_Const_tarval(r);
689 static ir_node *find_compound_ent_value(ir_node *ptr)
691 return rec_find_compound_ent_value(ptr, NULL);
695 static void reduce_adr_usage(ir_node *ptr);
698 * Update a Load that may have lost its users.
700 static void handle_load_update(ir_node *load)
702 ldst_info_t *info = (ldst_info_t*)get_irn_link(load);
704 /* do NOT touch volatile loads for now */
705 if (get_Load_volatility(load) == volatility_is_volatile)
708 if (! info->projs[pn_Load_res] && ! info->projs[pn_Load_X_except]) {
709 ir_node *ptr = get_Load_ptr(load);
710 ir_node *mem = get_Load_mem(load);
712 /* a Load whose value is neither used nor exception checked, remove it */
713 exchange(info->projs[pn_Load_M], mem);
714 if (info->projs[pn_Load_X_regular])
715 exchange(info->projs[pn_Load_X_regular], new_r_Jmp(get_nodes_block(load)));
717 reduce_adr_usage(ptr);
719 } /* handle_load_update */
722 * A use of an address node has vanished. Check if this was a Proj
723 * node and update the counters.
725 static void reduce_adr_usage(ir_node *ptr)
730 if (get_irn_n_edges(ptr) > 0)
733 /* this Proj is dead now */
734 pred = get_Proj_pred(ptr);
736 ldst_info_t *info = (ldst_info_t*)get_irn_link(pred);
737 info->projs[get_Proj_proj(ptr)] = NULL;
739 /* this node lost its result proj, handle that */
740 handle_load_update(pred);
742 } /* reduce_adr_usage */
745 * Check, if an already existing value of mode old_mode can be converted
746 * into the needed one new_mode without loss.
748 static int can_use_stored_value(ir_mode *old_mode, ir_mode *new_mode)
752 if (old_mode == new_mode)
755 old_size = get_mode_size_bits(old_mode);
756 new_size = get_mode_size_bits(new_mode);
758 /* if both modes are two-complement ones, we can always convert the
759 Stored value into the needed one. (on big endian machines we currently
760 only support this for modes of same size) */
761 if (old_size >= new_size &&
762 get_mode_arithmetic(old_mode) == irma_twos_complement &&
763 get_mode_arithmetic(new_mode) == irma_twos_complement &&
764 (!be_get_backend_param()->byte_order_big_endian
765 || old_size == new_size)) {
772 * Check whether a Call is at least pure, i.e. does only read memory.
774 static unsigned is_Call_pure(ir_node *call)
776 ir_type *call_tp = get_Call_type(call);
777 unsigned prop = get_method_additional_properties(call_tp);
779 /* check first the call type */
780 if ((prop & (mtp_property_const|mtp_property_pure)) == 0) {
781 /* try the called entity */
782 ir_node *ptr = get_Call_ptr(call);
784 if (is_Global(ptr)) {
785 ir_entity *ent = get_Global_entity(ptr);
787 prop = get_entity_additional_properties(ent);
790 return (prop & (mtp_property_const|mtp_property_pure)) != 0;
793 static ir_node *get_base_and_offset(ir_node *ptr, long *pOffset)
795 ir_mode *mode = get_irn_mode(ptr);
798 /* TODO: long might not be enough, we should probably use some tarval thingy... */
801 ir_node *l = get_Add_left(ptr);
802 ir_node *r = get_Add_right(ptr);
804 if (get_irn_mode(l) != mode || !is_Const(r))
807 offset += get_tarval_long(get_Const_tarval(r));
809 } else if (is_Sub(ptr)) {
810 ir_node *l = get_Sub_left(ptr);
811 ir_node *r = get_Sub_right(ptr);
813 if (get_irn_mode(l) != mode || !is_Const(r))
816 offset -= get_tarval_long(get_Const_tarval(r));
818 } else if (is_Sel(ptr)) {
819 ir_entity *ent = get_Sel_entity(ptr);
820 ir_type *tp = get_entity_owner(ent);
822 if (is_Array_type(tp)) {
826 /* only one dimensional arrays yet */
827 if (get_Sel_n_indexs(ptr) != 1)
829 index = get_Sel_index(ptr, 0);
830 if (! is_Const(index))
833 tp = get_entity_type(ent);
834 if (get_type_state(tp) != layout_fixed)
837 size = get_type_size_bytes(tp);
838 offset += size * get_tarval_long(get_Const_tarval(index));
840 if (get_type_state(tp) != layout_fixed)
842 offset += get_entity_offset(ent);
844 ptr = get_Sel_ptr(ptr);
853 static int try_load_after_store(ir_node *load,
854 ir_node *load_base_ptr, long load_offset, ir_node *store)
857 ir_node *store_ptr = get_Store_ptr(store);
859 ir_node *store_base_ptr = get_base_and_offset(store_ptr, &store_offset);
860 ir_node *store_value;
869 if (load_base_ptr != store_base_ptr)
872 load_mode = get_Load_mode(load);
873 load_mode_len = get_mode_size_bytes(load_mode);
874 store_mode = get_irn_mode(get_Store_value(store));
875 store_mode_len = get_mode_size_bytes(store_mode);
876 delta = load_offset - store_offset;
877 store_value = get_Store_value(store);
879 if (delta != 0 || store_mode != load_mode) {
880 /* TODO: implement for big-endian */
881 if (delta < 0 || delta + load_mode_len > store_mode_len
882 || (be_get_backend_param()->byte_order_big_endian
883 && load_mode_len != store_mode_len))
886 if (get_mode_arithmetic(store_mode) != irma_twos_complement ||
887 get_mode_arithmetic(load_mode) != irma_twos_complement)
891 /* produce a shift to adjust offset delta */
894 ir_graph *irg = get_irn_irg(load);
896 cnst = new_r_Const_long(irg, mode_Iu, delta * 8);
897 store_value = new_r_Shr(get_nodes_block(load),
898 store_value, cnst, store_mode);
901 /* add an convert if needed */
902 if (store_mode != load_mode) {
903 store_value = new_r_Conv(get_nodes_block(load), store_value, load_mode);
907 DBG_OPT_RAW(load, store_value);
909 info = (ldst_info_t*)get_irn_link(load);
910 if (info->projs[pn_Load_M])
911 exchange(info->projs[pn_Load_M], get_Load_mem(load));
915 if (info->projs[pn_Load_X_except]) {
916 ir_graph *irg = get_irn_irg(load);
917 exchange( info->projs[pn_Load_X_except], new_r_Bad(irg, mode_X));
920 if (info->projs[pn_Load_X_regular]) {
921 exchange( info->projs[pn_Load_X_regular], new_r_Jmp(get_nodes_block(load)));
925 if (info->projs[pn_Load_res])
926 exchange(info->projs[pn_Load_res], store_value);
928 load_ptr = get_Load_ptr(load);
930 reduce_adr_usage(load_ptr);
931 return res | DF_CHANGED;
935 * Follow the memory chain as long as there are only Loads,
936 * alias free Stores, and constant Calls and try to replace the
937 * current Load by a previous ones.
938 * Note that in unreachable loops it might happen that we reach
939 * load again, as well as we can fall into a cycle.
940 * We break such cycles using a special visited flag.
942 * INC_MASTER() must be called before dive into
944 static unsigned follow_Mem_chain(ir_node *load, ir_node *curr)
947 ldst_info_t *info = (ldst_info_t*)get_irn_link(load);
949 ir_node *ptr = get_Load_ptr(load);
950 ir_node *mem = get_Load_mem(load);
951 ir_mode *load_mode = get_Load_mode(load);
953 for (pred = curr; load != pred; ) {
954 ldst_info_t *pred_info = (ldst_info_t*)get_irn_link(pred);
957 * a Load immediately after a Store -- a read after write.
958 * We may remove the Load, if both Load & Store does not have an
959 * exception handler OR they are in the same Block. In the latter
960 * case the Load cannot throw an exception when the previous Store was
963 * Why we need to check for Store Exception? If the Store cannot
964 * be executed (ROM) the exception handler might simply jump into
966 * We could make it a little bit better if we would know that the
967 * exception handler of the Store jumps directly to the end...
969 if (is_Store(pred) && ((pred_info->projs[pn_Store_X_except] == NULL
970 && info->projs[pn_Load_X_except] == NULL)
971 || get_nodes_block(load) == get_nodes_block(pred)))
974 ir_node *base_ptr = get_base_and_offset(ptr, &load_offset);
975 int changes = try_load_after_store(load, base_ptr, load_offset, pred);
978 return res | changes;
979 } else if (is_Load(pred) && get_Load_ptr(pred) == ptr &&
980 can_use_stored_value(get_Load_mode(pred), load_mode)) {
982 * a Load after a Load -- a read after read.
983 * We may remove the second Load, if it does not have an exception
984 * handler OR they are in the same Block. In the later case
985 * the Load cannot throw an exception when the previous Load was
988 * Here, there is no need to check if the previous Load has an
989 * exception hander because they would have exact the same
992 * TODO: implement load-after-load with different mode for big
995 if (info->projs[pn_Load_X_except] == NULL
996 || get_nodes_block(load) == get_nodes_block(pred)) {
999 DBG_OPT_RAR(load, pred);
1001 /* the result is used */
1002 if (info->projs[pn_Load_res]) {
1003 if (pred_info->projs[pn_Load_res] == NULL) {
1004 /* create a new Proj again */
1005 pred_info->projs[pn_Load_res] = new_r_Proj(pred, get_Load_mode(pred), pn_Load_res);
1007 value = pred_info->projs[pn_Load_res];
1009 /* add an convert if needed */
1010 if (get_Load_mode(pred) != load_mode) {
1011 value = new_r_Conv(get_nodes_block(load), value, load_mode);
1014 exchange(info->projs[pn_Load_res], value);
1017 if (info->projs[pn_Load_M])
1018 exchange(info->projs[pn_Load_M], mem);
1021 if (info->projs[pn_Load_X_except]) {
1022 ir_graph *irg = get_irn_irg(load);
1023 exchange(info->projs[pn_Load_X_except], new_r_Bad(irg, mode_X));
1026 if (info->projs[pn_Load_X_regular]) {
1027 exchange( info->projs[pn_Load_X_regular], new_r_Jmp(get_nodes_block(load)));
1032 reduce_adr_usage(ptr);
1033 return res |= DF_CHANGED;
1037 if (is_Store(pred)) {
1038 /* check if we can pass through this store */
1039 ir_alias_relation rel = get_alias_relation(
1040 get_Store_ptr(pred),
1041 get_irn_mode(get_Store_value(pred)),
1043 /* if the might be an alias, we cannot pass this Store */
1044 if (rel != ir_no_alias)
1046 pred = skip_Proj(get_Store_mem(pred));
1047 } else if (is_Load(pred)) {
1048 pred = skip_Proj(get_Load_mem(pred));
1049 } else if (is_Call(pred)) {
1050 if (is_Call_pure(pred)) {
1051 /* The called graph is at least pure, so there are no Store's
1052 in it. We can handle it like a Load and skip it. */
1053 pred = skip_Proj(get_Call_mem(pred));
1055 /* there might be Store's in the graph, stop here */
1059 /* follow only Load chains */
1063 /* check for cycles */
1064 if (NODE_VISITED(pred_info))
1066 MARK_NODE(pred_info);
1069 if (is_Sync(pred)) {
1072 /* handle all Sync predecessors */
1073 for (i = get_Sync_n_preds(pred) - 1; i >= 0; --i) {
1074 res |= follow_Mem_chain(load, skip_Proj(get_Sync_pred(pred, i)));
1081 } /* follow_Mem_chain */
1084 * Check if we can replace the load by a given const from
1085 * the const code irg.
1087 ir_node *can_replace_load_by_const(const ir_node *load, ir_node *c)
1089 ir_mode *c_mode = get_irn_mode(c);
1090 ir_mode *l_mode = get_Load_mode(load);
1091 ir_node *block = get_nodes_block(load);
1092 dbg_info *dbgi = get_irn_dbg_info(load);
1093 ir_node *res = copy_const_value(dbgi, c, block);
1095 if (c_mode != l_mode) {
1096 /* check, if the mode matches OR can be easily converted info */
1097 if (is_reinterpret_cast(c_mode, l_mode)) {
1098 /* copy the value from the const code irg and cast it */
1099 res = new_rd_Conv(dbgi, block, res, l_mode);
1110 * @param load the Load node
1112 static unsigned optimize_load(ir_node *load)
1114 ldst_info_t *info = (ldst_info_t*)get_irn_link(load);
1115 ir_node *mem, *ptr, *value;
1120 /* do NOT touch volatile loads for now */
1121 if (get_Load_volatility(load) == volatility_is_volatile)
1124 /* the address of the load to be optimized */
1125 ptr = get_Load_ptr(load);
1127 /* The mem of the Load. Must still be returned after optimization. */
1128 mem = get_Load_mem(load);
1130 if (info->projs[pn_Load_res] == NULL
1131 && info->projs[pn_Load_X_except] == NULL) {
1132 /* the value is never used and we don't care about exceptions, remove */
1133 exchange(info->projs[pn_Load_M], mem);
1135 if (info->projs[pn_Load_X_regular]) {
1136 /* should not happen, but if it does, remove it */
1137 exchange(info->projs[pn_Load_X_regular], new_r_Jmp(get_nodes_block(load)));
1141 reduce_adr_usage(ptr);
1142 return res | DF_CHANGED;
1145 /* Load from a constant polymorphic field, where we can resolve
1147 value = transform_polymorph_Load(load);
1148 if (value == load) {
1150 /* check if we can determine the entity that will be loaded */
1151 ent = find_constant_entity(ptr);
1153 && get_entity_visibility(ent) != ir_visibility_external) {
1154 /* a static allocation that is not external: there should be NO
1155 * exception when loading even if we cannot replace the load itself.
1158 /* no exception, clear the info field as it might be checked later again */
1159 if (info->projs[pn_Load_X_except]) {
1160 ir_graph *irg = get_irn_irg(load);
1161 exchange(info->projs[pn_Load_X_except], new_r_Bad(irg, mode_X));
1162 info->projs[pn_Load_X_except] = NULL;
1165 if (info->projs[pn_Load_X_regular]) {
1166 exchange(info->projs[pn_Load_X_regular], new_r_Jmp(get_nodes_block(load)));
1167 info->projs[pn_Load_X_regular] = NULL;
1171 if (get_entity_linkage(ent) & IR_LINKAGE_CONSTANT) {
1172 if (has_entity_initializer(ent)) {
1173 /* new style initializer */
1174 value = find_compound_ent_value(ptr);
1175 } else if (entity_has_compound_ent_values(ent)) {
1176 /* old style initializer */
1177 compound_graph_path *path = get_accessed_path(ptr);
1180 assert(is_proper_compound_graph_path(path, get_compound_graph_path_length(path)-1));
1182 value = get_compound_ent_value_by_path(ent, path);
1183 DB((dbg, LEVEL_1, " Constant access at %F%F resulted in %+F\n", ent, path, value));
1184 free_compound_graph_path(path);
1187 if (value != NULL) {
1188 ir_graph *irg = get_irn_irg(load);
1189 value = can_replace_load_by_const(load, value);
1190 if (value != NULL && is_Sel(ptr) &&
1191 !is_irg_state(irg, IR_GRAPH_STATE_IMPLICIT_BITFIELD_MASKING)) {
1192 /* frontend has inserted masking operations after bitfield accesses,
1193 * so we might have to shift the const. */
1194 unsigned char bit_offset = get_entity_offset_bits_remainder(get_Sel_entity(ptr));
1195 ir_tarval *tv_old = get_Const_tarval(value);
1196 ir_tarval *tv_offset = new_tarval_from_long(bit_offset, mode_Bu);
1197 ir_tarval *tv_new = tarval_shl(tv_old, tv_offset);
1198 value = new_r_Const(irg, tv_new);
1204 if (value != NULL) {
1205 /* we completely replace the load by this value */
1206 if (info->projs[pn_Load_X_except]) {
1207 ir_graph *irg = get_irn_irg(load);
1208 exchange(info->projs[pn_Load_X_except], new_r_Bad(irg, mode_X));
1209 info->projs[pn_Load_X_except] = NULL;
1212 if (info->projs[pn_Load_X_regular]) {
1213 exchange(info->projs[pn_Load_X_regular], new_r_Jmp(get_nodes_block(load)));
1214 info->projs[pn_Load_X_regular] = NULL;
1217 if (info->projs[pn_Load_M]) {
1218 exchange(info->projs[pn_Load_M], mem);
1221 if (info->projs[pn_Load_res]) {
1222 exchange(info->projs[pn_Load_res], value);
1226 reduce_adr_usage(ptr);
1230 /* Check, if the address of this load is used more than once.
1231 * If not, more load cannot be removed in any case. */
1232 if (get_irn_n_uses(ptr) <= 1 && get_irn_n_uses(get_base_and_offset(ptr, &dummy)) <= 1)
1236 * follow the memory chain as long as there are only Loads
1237 * and try to replace current Load or Store by a previous one.
1238 * Note that in unreachable loops it might happen that we reach
1239 * load again, as well as we can fall into a cycle.
1240 * We break such cycles using a special visited flag.
1243 res = follow_Mem_chain(load, skip_Proj(mem));
1245 } /* optimize_load */
1248 * Check whether a value of mode new_mode would completely overwrite a value
1249 * of mode old_mode in memory.
1251 static int is_completely_overwritten(ir_mode *old_mode, ir_mode *new_mode)
1253 return get_mode_size_bits(new_mode) >= get_mode_size_bits(old_mode);
1254 } /* is_completely_overwritten */
1257 * Check whether small is a part of large (starting at same address).
1259 static int is_partially_same(ir_node *small, ir_node *large)
1261 ir_mode *sm = get_irn_mode(small);
1262 ir_mode *lm = get_irn_mode(large);
1264 /* FIXME: Check endianness */
1265 return is_Conv(small) && get_Conv_op(small) == large
1266 && get_mode_size_bytes(sm) < get_mode_size_bytes(lm)
1267 && get_mode_arithmetic(sm) == irma_twos_complement
1268 && get_mode_arithmetic(lm) == irma_twos_complement;
1269 } /* is_partially_same */
1272 * follow the memory chain as long as there are only Loads and alias free Stores.
1274 * INC_MASTER() must be called before dive into
1276 static unsigned follow_Mem_chain_for_Store(ir_node *store, ir_node *curr)
1279 ldst_info_t *info = (ldst_info_t*)get_irn_link(store);
1281 ir_node *ptr = get_Store_ptr(store);
1282 ir_node *mem = get_Store_mem(store);
1283 ir_node *value = get_Store_value(store);
1284 ir_mode *mode = get_irn_mode(value);
1285 ir_node *block = get_nodes_block(store);
1287 for (pred = curr; pred != store;) {
1288 ldst_info_t *pred_info = (ldst_info_t*)get_irn_link(pred);
1291 * BEWARE: one might think that checking the modes is useless, because
1292 * if the pointers are identical, they refer to the same object.
1293 * This is only true in strong typed languages, not is C were the following
1294 * is possible *(ir_type1 *)p = a; *(ir_type2 *)p = b ...
1295 * However, if the size of the mode that is written is bigger or equal the
1296 * size of the old one, the old value is completely overwritten and can be
1299 if (is_Store(pred) && get_Store_ptr(pred) == ptr &&
1300 get_nodes_block(pred) == block) {
1302 * a Store after a Store in the same Block -- a write after write.
1306 * We may remove the first Store, if the old value is completely
1307 * overwritten or the old value is a part of the new value,
1308 * and if it does not have an exception handler.
1310 * TODO: What, if both have the same exception handler ???
1312 if (get_Store_volatility(pred) != volatility_is_volatile
1313 && !pred_info->projs[pn_Store_X_except]) {
1314 ir_node *predvalue = get_Store_value(pred);
1315 ir_mode *predmode = get_irn_mode(predvalue);
1317 if (is_completely_overwritten(predmode, mode)
1318 || is_partially_same(predvalue, value)) {
1319 DBG_OPT_WAW(pred, store);
1320 exchange(pred_info->projs[pn_Store_M], get_Store_mem(pred));
1322 reduce_adr_usage(ptr);
1328 * We may remove the Store, if the old value already contains
1329 * the new value, and if it does not have an exception handler.
1331 * TODO: What, if both have the same exception handler ???
1333 if (get_Store_volatility(store) != volatility_is_volatile
1334 && !info->projs[pn_Store_X_except]) {
1335 ir_node *predvalue = get_Store_value(pred);
1337 if (is_partially_same(value, predvalue)) {
1338 DBG_OPT_WAW(pred, store);
1339 exchange(info->projs[pn_Store_M], mem);
1341 reduce_adr_usage(ptr);
1345 } else if (is_Load(pred) && get_Load_ptr(pred) == ptr &&
1346 value == pred_info->projs[pn_Load_res]) {
1348 * a Store of a value just loaded from the same address
1349 * -- a write after read.
1350 * We may remove the Store, if it does not have an exception
1353 if (! info->projs[pn_Store_X_except]) {
1354 DBG_OPT_WAR(store, pred);
1355 exchange(info->projs[pn_Store_M], mem);
1357 reduce_adr_usage(ptr);
1362 if (is_Store(pred)) {
1363 /* check if we can pass through this store */
1364 ir_alias_relation rel = get_alias_relation(
1365 get_Store_ptr(pred),
1366 get_irn_mode(get_Store_value(pred)),
1368 /* if the might be an alias, we cannot pass this Store */
1369 if (rel != ir_no_alias)
1371 pred = skip_Proj(get_Store_mem(pred));
1372 } else if (is_Load(pred)) {
1373 ir_alias_relation rel = get_alias_relation(
1374 get_Load_ptr(pred), get_Load_mode(pred),
1376 if (rel != ir_no_alias)
1379 pred = skip_Proj(get_Load_mem(pred));
1381 /* follow only Load chains */
1385 /* check for cycles */
1386 if (NODE_VISITED(pred_info))
1388 MARK_NODE(pred_info);
1391 if (is_Sync(pred)) {
1394 /* handle all Sync predecessors */
1395 for (i = get_Sync_n_preds(pred) - 1; i >= 0; --i) {
1396 res |= follow_Mem_chain_for_Store(store, skip_Proj(get_Sync_pred(pred, i)));
1402 } /* follow_Mem_chain_for_Store */
1404 /** find entity used as base for an address calculation */
1405 static ir_entity *find_entity(ir_node *ptr)
1407 switch (get_irn_opcode(ptr)) {
1409 return get_SymConst_entity(ptr);
1411 ir_node *pred = get_Sel_ptr(ptr);
1412 if (get_irg_frame(get_irn_irg(ptr)) == pred)
1413 return get_Sel_entity(ptr);
1415 return find_entity(pred);
1419 ir_node *left = get_binop_left(ptr);
1421 if (mode_is_reference(get_irn_mode(left)))
1422 return find_entity(left);
1423 right = get_binop_right(ptr);
1424 if (mode_is_reference(get_irn_mode(right)))
1425 return find_entity(right);
1436 * @param store the Store node
1438 static unsigned optimize_store(ir_node *store)
1444 if (get_Store_volatility(store) == volatility_is_volatile)
1447 ptr = get_Store_ptr(store);
1448 entity = find_entity(ptr);
1450 /* a store to an entity which is never read is unnecessary */
1451 if (entity != NULL && !(get_entity_usage(entity) & ir_usage_read)) {
1452 ldst_info_t *info = (ldst_info_t*)get_irn_link(store);
1453 if (info->projs[pn_Store_X_except] == NULL) {
1454 DB((dbg, LEVEL_1, " Killing useless %+F to never read entity %+F\n", store, entity));
1455 exchange(info->projs[pn_Store_M], get_Store_mem(store));
1457 reduce_adr_usage(ptr);
1462 /* Check, if the address of this Store is used more than once.
1463 * If not, this Store cannot be removed in any case. */
1464 if (get_irn_n_uses(ptr) <= 1)
1467 mem = get_Store_mem(store);
1469 /* follow the memory chain as long as there are only Loads */
1472 return follow_Mem_chain_for_Store(store, skip_Proj(mem));
1473 } /* optimize_store */
1476 * walker, optimizes Phi after Stores to identical places:
1477 * Does the following optimization:
1480 * val1 val2 val3 val1 val2 val3
1482 * Store Store Store \ | /
1489 * This reduces the number of stores and allows for predicated execution.
1490 * Moves Stores back to the end of a function which may be bad.
1492 * This is only possible if the predecessor blocks have only one successor.
1494 static unsigned optimize_phi(ir_node *phi, walk_env_t *wenv)
1497 ir_node *store, *ptr, *block, *phi_block, *phiM, *phiD, *exc, *projM;
1502 ir_node **inM, **inD, **projMs;
1504 dbg_info *db = NULL;
1506 block_info_t *bl_info;
1509 /* Must be a memory Phi */
1510 if (get_irn_mode(phi) != mode_M)
1513 n = get_Phi_n_preds(phi);
1517 /* must be only one user */
1518 projM = get_Phi_pred(phi, 0);
1519 if (get_irn_n_edges(projM) != 1)
1522 store = skip_Proj(projM);
1526 if (!is_Store(store))
1529 block = get_nodes_block(store);
1531 /* check if the block is post dominated by Phi-block
1532 and has no exception exit */
1533 bl_info = (block_info_t*)get_irn_link(block);
1534 if (bl_info->flags & BLOCK_HAS_EXC)
1537 phi_block = get_nodes_block(phi);
1538 if (! block_strictly_postdominates(phi_block, block))
1541 /* this is the address of the store */
1542 ptr = get_Store_ptr(store);
1543 mode = get_irn_mode(get_Store_value(store));
1544 info = (ldst_info_t*)get_irn_link(store);
1545 exc = info->exc_block;
1547 for (i = 1; i < n; ++i) {
1548 ir_node *pred = get_Phi_pred(phi, i);
1550 if (get_irn_n_edges(pred) != 1)
1553 pred = skip_Proj(pred);
1554 if (!is_Store(pred))
1557 if (ptr != get_Store_ptr(pred) || mode != get_irn_mode(get_Store_value(pred)))
1560 info = (ldst_info_t*)get_irn_link(pred);
1562 /* check, if all stores have the same exception flow */
1563 if (exc != info->exc_block)
1566 block = get_nodes_block(pred);
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 = (block_info_t*)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 = (ldst_info_t*)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, cons_none);
1640 co_set_irn_name(store, co_get_irn_ident(old_store));
1643 projM = new_rd_Proj(NULL, 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, 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 = (walk_env_t*)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 size_t 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 = (node_entry*)phase_get_irn_data(&env->ph, irn);
1734 e = (node_entry*)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 size_t 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 = (const avail_entry_t*)elt;
1813 const avail_entry_t *b = (const avail_entry_t*)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;
1837 phi_entry *phi_list = NULL;
1840 avail = new_set(cmp_avail_entry, 8);
1842 /* collect all outer memories */
1843 for (phi = pscc->head; phi != NULL; phi = next) {
1844 node_entry *ne = get_irn_ne(phi, env);
1847 /* check all memory Phi's */
1851 assert(get_irn_mode(phi) == mode_M && "DFS return non-memory Phi");
1853 for (j = get_irn_arity(phi) - 1; j >= 0; --j) {
1854 ir_node *pred = get_irn_n(phi, j);
1855 node_entry *pe = get_irn_ne(pred, env);
1857 if (pe->pscc != ne->pscc) {
1858 /* not in the same SCC, is region const */
1859 phi_entry *pe = (phi_entry*)phase_alloc(&env->ph, sizeof(*pe));
1863 pe->next = phi_list;
1868 /* no Phis no fun */
1869 assert(phi_list != NULL && "DFS found a loop without Phi");
1871 /* for now, we cannot handle more than one input (only reducible cf) */
1872 if (phi_list->next != NULL)
1875 for (load = pscc->head; load; load = next) {
1877 node_entry *ne = get_irn_ne(load, env);
1880 if (is_Load(load)) {
1881 ldst_info_t *info = (ldst_info_t*)get_irn_link(load);
1882 ir_node *ptr = get_Load_ptr(load);
1884 /* for now, we cannot handle Loads with exceptions */
1885 if (info->projs[pn_Load_res] == NULL || info->projs[pn_Load_X_regular] != NULL || info->projs[pn_Load_X_except] != NULL)
1888 /* for now, we can only move Load(Global) */
1889 if (! is_Global(ptr))
1891 load_mode = get_Load_mode(load);
1892 for (other = pscc->head; other != NULL; other = next_other) {
1893 node_entry *ne = get_irn_ne(other, env);
1894 next_other = ne->next;
1896 if (is_Store(other)) {
1897 ir_alias_relation rel = get_alias_relation(
1898 get_Store_ptr(other),
1899 get_irn_mode(get_Store_value(other)),
1901 /* if the might be an alias, we cannot pass this Store */
1902 if (rel != ir_no_alias)
1905 /* only Phis and pure Calls are allowed here, so ignore them */
1907 if (other == NULL) {
1908 ldst_info_t *ninfo = NULL;
1912 /* yep, no aliasing Store found, Load can be moved */
1913 DB((dbg, LEVEL_1, " Found a Load that could be moved: %+F\n", load));
1915 db = get_irn_dbg_info(load);
1916 for (pe = phi_list; pe != NULL; pe = pe->next) {
1918 ir_node *phi = pe->phi;
1919 ir_node *blk = get_nodes_block(phi);
1920 ir_node *pred = get_Block_cfgpred_block(blk, pos);
1922 avail_entry_t entry, *res;
1925 entry.mode = load_mode;
1926 res = (avail_entry_t*)set_find(avail, &entry, sizeof(entry), hash_cache_entry(&entry));
1930 irn = new_rd_Load(db, pred, get_Phi_pred(phi, pos), ptr, load_mode, cons_none);
1932 set_insert(avail, &entry, sizeof(entry), hash_cache_entry(&entry));
1933 DB((dbg, LEVEL_1, " Created %+F in %+F\n", irn, pred));
1936 ninfo = get_ldst_info(irn, phase_obst(&env->ph));
1938 ninfo->projs[pn_Load_M] = mem = new_r_Proj(irn, mode_M, pn_Load_M);
1940 /* irn is from cache, so do not set phi pred again.
1941 * There might be other Loads between phi and irn already.
1943 set_Phi_pred(phi, pos, mem);
1946 ninfo->projs[pn_Load_res] = new_r_Proj(irn, load_mode, pn_Load_res);
1949 /* now kill the old Load */
1950 exchange(info->projs[pn_Load_M], get_Load_mem(load));
1951 exchange(info->projs[pn_Load_res], ninfo->projs[pn_Load_res]);
1953 env->changes |= DF_CHANGED;
1958 } /* move_loads_out_of_loops */
1961 * Process a loop SCC.
1963 * @param pscc the SCC
1964 * @param env the loop environment
1966 static void process_loop(scc *pscc, loop_env *env)
1968 ir_node *irn, *next, *header = NULL;
1969 node_entry *b, *h = NULL;
1970 int j, only_phi, num_outside, process = 0;
1973 /* find the header block for this scc */
1974 for (irn = pscc->head; irn; irn = next) {
1975 node_entry *e = get_irn_ne(irn, env);
1976 ir_node *block = get_nodes_block(irn);
1979 b = get_irn_ne(block, env);
1981 if (header != NULL) {
1982 if (h->POnum < b->POnum) {
1992 /* check if this scc contains only Phi, Loads or Stores nodes */
1996 for (irn = pscc->head; irn; irn = next) {
1997 node_entry *e = get_irn_ne(irn, env);
2000 switch (get_irn_opcode(irn)) {
2002 if (is_Call_pure(irn)) {
2003 /* pure calls can be treated like loads */
2007 /* non-pure calls must be handle like may-alias Stores */
2010 /* cannot handle CopyB yet */
2014 if (get_Load_volatility(irn) == volatility_is_volatile) {
2015 /* cannot handle loops with volatile Loads */
2021 if (get_Store_volatility(irn) == volatility_is_volatile) {
2022 /* cannot handle loops with volatile Stores */
2031 for (j = get_irn_arity(irn) - 1; j >= 0; --j) {
2032 ir_node *pred = get_irn_n(irn, j);
2033 node_entry *pe = get_irn_ne(pred, env);
2035 if (pe->pscc != e->pscc) {
2036 /* not in the same SCC, must be a region const */
2037 if (! is_rc(pred, header)) {
2038 /* not a memory loop */
2041 if (out_rc == NULL) {
2042 /* first region constant */
2045 } else if (out_rc != pred) {
2046 /* another region constant */
2057 /* found a memory loop */
2058 DB((dbg, LEVEL_2, " Found a memory loop:\n "));
2059 if (only_phi && num_outside == 1) {
2060 /* a phi cycle with only one real predecessor can be collapsed */
2061 DB((dbg, LEVEL_2, " Found an USELESS Phi cycle:\n "));
2063 for (irn = pscc->head; irn; irn = next) {
2064 node_entry *e = get_irn_ne(irn, env);
2066 exchange(irn, out_rc);
2068 env->changes |= DF_CHANGED;
2072 #ifdef DEBUG_libfirm
2073 for (irn = pscc->head; irn; irn = next) {
2074 node_entry *e = get_irn_ne(irn, env);
2076 DB((dbg, LEVEL_2, " %+F,", irn));
2078 DB((dbg, LEVEL_2, "\n"));
2080 move_loads_out_of_loops(pscc, env);
2084 } /* process_loop */
2089 * @param pscc the SCC
2090 * @param env the loop environment
2092 static void process_scc(scc *pscc, loop_env *env)
2094 ir_node *head = pscc->head;
2095 node_entry *e = get_irn_ne(head, env);
2097 #ifdef DEBUG_libfirm
2099 ir_node *irn, *next;
2101 DB((dbg, LEVEL_4, " SCC at %p:\n ", pscc));
2102 for (irn = pscc->head; irn; irn = next) {
2103 node_entry *e = get_irn_ne(irn, env);
2107 DB((dbg, LEVEL_4, " %+F,", irn));
2109 DB((dbg, LEVEL_4, "\n"));
2113 if (e->next != NULL) {
2114 /* this SCC has more than one member */
2115 process_loop(pscc, env);
2120 * Do Tarjan's SCC algorithm and drive load/store optimization.
2122 * @param irn start at this node
2123 * @param env the loop environment
2125 static void dfs(ir_node *irn, loop_env *env)
2128 node_entry *node = get_irn_ne(irn, env);
2130 mark_irn_visited(irn);
2132 node->DFSnum = env->nextDFSnum++;
2133 node->low = node->DFSnum;
2137 if (is_Phi(irn) || is_Sync(irn)) {
2138 n = get_irn_arity(irn);
2139 for (i = 0; i < n; ++i) {
2140 ir_node *pred = get_irn_n(irn, i);
2141 node_entry *o = get_irn_ne(pred, env);
2143 if (!irn_visited(pred)) {
2145 node->low = MIN(node->low, o->low);
2147 if (o->DFSnum < node->DFSnum && o->in_stack)
2148 node->low = MIN(o->DFSnum, node->low);
2150 } else if (is_fragile_op(irn)) {
2151 ir_node *pred = get_fragile_op_mem(irn);
2152 node_entry *o = get_irn_ne(pred, env);
2154 if (!irn_visited(pred)) {
2156 node->low = MIN(node->low, o->low);
2158 if (o->DFSnum < node->DFSnum && o->in_stack)
2159 node->low = MIN(o->DFSnum, node->low);
2160 } else if (is_Proj(irn)) {
2161 ir_node *pred = get_Proj_pred(irn);
2162 node_entry *o = get_irn_ne(pred, env);
2164 if (!irn_visited(pred)) {
2166 node->low = MIN(node->low, o->low);
2168 if (o->DFSnum < node->DFSnum && o->in_stack)
2169 node->low = MIN(o->DFSnum, node->low);
2172 /* IGNORE predecessors */
2175 if (node->low == node->DFSnum) {
2176 scc *pscc = (scc*)phase_alloc(&env->ph, sizeof(*pscc));
2184 e = get_irn_ne(x, env);
2186 e->next = pscc->head;
2190 process_scc(pscc, env);
2195 * Do the DFS on the memory edges a graph.
2197 * @param irg the graph to process
2198 * @param env the loop environment
2200 static void do_dfs(ir_graph *irg, loop_env *env)
2202 ir_node *endblk, *end;
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);
2219 } else if (is_Bad(pred)) {
2220 /* ignore non-optimized block predecessor */
2222 assert(0 && "Unknown EndBlock predecessor");
2226 /* visit the keep-alives */
2227 end = get_irg_end(irg);
2228 for (i = get_End_n_keepalives(end) - 1; i >= 0; --i) {
2229 ir_node *ka = get_End_keepalive(end, i);
2231 if (is_Phi(ka) && !irn_visited(ka))
2237 * Optimize Loads/Stores in loops.
2239 * @param irg the graph
2241 static int optimize_loops(ir_graph *irg)
2245 env.stack = NEW_ARR_F(ir_node *, 128);
2250 phase_init(&env.ph, irg, phase_irn_init_default);
2252 /* calculate the SCC's and drive loop optimization. */
2255 DEL_ARR_F(env.stack);
2256 phase_deinit(&env.ph);
2259 } /* optimize_loops */
2262 * do the load store optimization
2264 static ir_graph_state_t do_loadstore_opt(ir_graph *irg)
2267 ir_graph_state_t res = 0;
2269 FIRM_DBG_REGISTER(dbg, "firm.opt.ldstopt");
2271 assert(get_irg_phase_state(irg) != phase_building);
2272 assert(get_irg_pinned(irg) != op_pin_state_floats &&
2273 "LoadStore optimization needs pinned graph");
2275 if (get_opt_alias_analysis()) {
2276 assure_irp_globals_entity_usage_computed();
2279 obstack_init(&env.obst);
2282 /* init the links, then collect Loads/Stores/Proj's in lists */
2284 irg_walk_graph(irg, firm_clear_link, collect_nodes, &env);
2286 /* now we have collected enough information, optimize */
2287 irg_walk_graph(irg, NULL, do_load_store_optimize, &env);
2289 env.changes |= optimize_loops(irg);
2291 obstack_free(&env.obst, NULL);
2293 /* Handle graph state */
2295 set_irg_entity_usage_state(irg, ir_entity_usage_not_computed);
2296 edges_deactivate(irg);
2299 if (!(env.changes & CF_CHANGED)) {
2300 res |= IR_GRAPH_STATE_CONSISTENT_DOMINANCE | IR_GRAPH_STATE_NO_BAD_BLOCKS;
2306 optdesc_t opt_loadstore = {
2308 IR_GRAPH_STATE_NO_UNREACHABLE_BLOCKS | IR_GRAPH_STATE_CONSISTENT_OUT_EDGES | IR_GRAPH_STATE_NO_CRITICAL_EDGES | IR_GRAPH_STATE_CONSISTENT_DOMINANCE | IR_GRAPH_STATE_CONSISTENT_ENTITY_USAGE,
2312 int optimize_load_store(ir_graph *irg)
2314 perform_irg_optimization(irg, &opt_loadstore);
2318 ir_graph_pass_t *optimize_load_store_pass(const char *name)
2320 return def_graph_pass_ret(name ? name : "ldst", optimize_load_store);
2321 } /* optimize_load_store_pass */