2 * Copyright (C) 1995-2008 University of Karlsruhe. All right reserved.
4 * This file is part of libFirm.
6 * This file may be distributed and/or modified under the terms of the
7 * GNU General Public License version 2 as published by the Free Software
8 * Foundation and appearing in the file LICENSE.GPL included in the
9 * packaging of this file.
11 * Licensees holding valid libFirm Professional Edition licenses may use
12 * this file in accordance with the libFirm Commercial License.
13 * Agreement provided with the Software.
15 * This file is provided AS IS with NO WARRANTY OF ANY KIND, INCLUDING THE
16 * WARRANTY OF DESIGN, MERCHANTABILITY AND FITNESS FOR A PARTICULAR
22 * @brief Load/Store optimizations.
23 * @author Michael Beck
32 #include "iroptimize.h"
34 #include "irgraph_t.h"
42 #include "dbginfo_t.h"
43 #include "iropt_dbg.h"
49 #include "opt_polymorphy.h"
52 #include "irphase_t.h"
56 /** The debug handle. */
57 DEBUG_ONLY(static firm_dbg_module_t *dbg;)
60 #include "cacheopt/cachesim.h"
64 #define IMAX(a,b) ((a) > (b) ? (a) : (b))
66 #define MAX_PROJ IMAX(IMAX(pn_Load_max, pn_Store_max), pn_Call_max)
69 DF_CHANGED = 1, /**< data flow changed */
70 CF_CHANGED = 2, /**< control flow changed */
76 typedef struct _walk_env_t {
77 struct obstack obst; /**< list of all stores */
78 unsigned changes; /**< a bitmask of graph changes */
81 /** A Load/Store info. */
82 typedef struct _ldst_info_t {
83 ir_node *projs[MAX_PROJ]; /**< list of Proj's of this node */
84 ir_node *exc_block; /**< the exception block if available */
85 int exc_idx; /**< predecessor index in the exception block */
86 unsigned visited; /**< visited counter for breaking loops */
90 * flags for control flow.
93 BLOCK_HAS_COND = 1, /**< Block has conditional control flow */
94 BLOCK_HAS_EXC = 2 /**< Block has exceptional control flow */
100 typedef struct _block_info_t {
101 unsigned flags; /**< flags for the block */
104 /** the master visited flag for loop detection. */
105 static unsigned master_visited = 0;
107 #define INC_MASTER() ++master_visited
108 #define MARK_NODE(info) (info)->visited = master_visited
109 #define NODE_VISITED(info) (info)->visited >= master_visited
112 * get the Load/Store info of a node
114 static ldst_info_t *get_ldst_info(ir_node *node, struct obstack *obst) {
115 ldst_info_t *info = get_irn_link(node);
118 info = obstack_alloc(obst, sizeof(*info));
119 memset(info, 0, sizeof(*info));
120 set_irn_link(node, info);
123 } /* get_ldst_info */
126 * get the Block info of a node
128 static block_info_t *get_block_info(ir_node *node, struct obstack *obst) {
129 block_info_t *info = get_irn_link(node);
132 info = obstack_alloc(obst, sizeof(*info));
133 memset(info, 0, sizeof(*info));
134 set_irn_link(node, info);
137 } /* get_block_info */
140 * update the projection info for a Load/Store
142 static unsigned update_projs(ldst_info_t *info, ir_node *proj)
144 long nr = get_Proj_proj(proj);
146 assert(0 <= nr && nr <= MAX_PROJ && "Wrong proj from LoadStore");
148 if (info->projs[nr]) {
149 /* there is already one, do CSE */
150 exchange(proj, info->projs[nr]);
154 info->projs[nr] = proj;
160 * update the exception block info for a Load/Store node.
162 * @param info the load/store info struct
163 * @param block the exception handler block for this load/store
164 * @param pos the control flow input of the block
166 static unsigned update_exc(ldst_info_t *info, ir_node *block, int pos)
168 assert(info->exc_block == NULL && "more than one exception block found");
170 info->exc_block = block;
175 /** Return the number of uses of an address node */
176 #define get_irn_n_uses(adr) get_irn_n_edges(adr)
179 * walker, collects all Load/Store/Proj nodes
181 * walks from Start -> End
183 static void collect_nodes(ir_node *node, void *env)
185 ir_opcode opcode = get_irn_opcode(node);
186 ir_node *pred, *blk, *pred_blk;
187 ldst_info_t *ldst_info;
188 walk_env_t *wenv = env;
190 if (opcode == iro_Proj) {
191 pred = get_Proj_pred(node);
192 opcode = get_irn_opcode(pred);
194 if (opcode == iro_Load || opcode == iro_Store || opcode == iro_Call) {
195 ldst_info = get_ldst_info(pred, &wenv->obst);
197 wenv->changes |= update_projs(ldst_info, node);
200 * Place the Proj's to the same block as the
201 * predecessor Load. This is always ok and prevents
202 * "non-SSA" form after optimizations if the Proj
203 * is in a wrong block.
205 blk = get_nodes_block(node);
206 pred_blk = get_nodes_block(pred);
207 if (blk != pred_blk) {
208 wenv->changes |= DF_CHANGED;
209 set_nodes_block(node, pred_blk);
212 } else if (opcode == iro_Block) {
215 for (i = get_Block_n_cfgpreds(node) - 1; i >= 0; --i) {
216 ir_node *pred_block, *proj;
217 block_info_t *bl_info;
220 pred = proj = get_Block_cfgpred(node, i);
223 pred = get_Proj_pred(proj);
224 is_exc = get_Proj_proj(proj) == pn_Generic_X_except;
227 /* ignore Bad predecessors, they will be removed later */
231 pred_block = get_nodes_block(pred);
232 bl_info = get_block_info(pred_block, &wenv->obst);
234 if (is_fragile_op(pred) && is_exc)
235 bl_info->flags |= BLOCK_HAS_EXC;
236 else if (is_irn_forking(pred))
237 bl_info->flags |= BLOCK_HAS_COND;
239 opcode = get_irn_opcode(pred);
240 if (is_exc && (opcode == iro_Load || opcode == iro_Store || opcode == iro_Call)) {
241 ldst_info = get_ldst_info(pred, &wenv->obst);
243 wenv->changes |= update_exc(ldst_info, node, i);
247 } /* collect_nodes */
250 * Returns an entity if the address ptr points to a constant one.
252 * @param ptr the address
254 * @return an entity or NULL
256 static ir_entity *find_constant_entity(ir_node *ptr)
259 if (is_SymConst(ptr) && get_SymConst_kind(ptr) == symconst_addr_ent) {
260 ir_entity *ent = get_SymConst_entity(ptr);
261 if (variability_constant == get_entity_variability(ent))
264 } else if (is_Sel(ptr)) {
265 ir_entity *ent = get_Sel_entity(ptr);
266 ir_type *tp = get_entity_owner(ent);
268 /* Do not fiddle with polymorphism. */
269 if (is_Class_type(get_entity_owner(ent)) &&
270 ((get_entity_n_overwrites(ent) != 0) ||
271 (get_entity_n_overwrittenby(ent) != 0) ) )
274 if (is_Array_type(tp)) {
278 for (i = 0, n = get_Sel_n_indexs(ptr); i < n; ++i) {
280 tarval *tlower, *tupper;
281 ir_node *index = get_Sel_index(ptr, i);
282 tarval *tv = computed_value(index);
284 /* check if the index is constant */
285 if (tv == tarval_bad)
288 bound = get_array_lower_bound(tp, i);
289 tlower = computed_value(bound);
290 bound = get_array_upper_bound(tp, i);
291 tupper = computed_value(bound);
293 if (tlower == tarval_bad || tupper == tarval_bad)
296 if (tarval_cmp(tv, tlower) & pn_Cmp_Lt)
298 if (tarval_cmp(tupper, tv) & pn_Cmp_Lt)
301 /* ok, bounds check finished */
305 if (variability_constant == get_entity_variability(ent))
309 ptr = get_Sel_ptr(ptr);
310 } else if (is_Add(ptr)) {
311 ir_node *l = get_Add_left(ptr);
312 ir_node *r = get_Add_right(ptr);
314 if (get_irn_mode(l) == get_irn_mode(ptr) && is_Const(r))
316 else if (get_irn_mode(r) == get_irn_mode(ptr) && is_Const(l))
321 /* for now, we support only one addition, reassoc should fold all others */
322 if (! is_SymConst(ptr) && !is_Sel(ptr))
324 } else if (is_Sub(ptr)) {
325 ir_node *l = get_Sub_left(ptr);
326 ir_node *r = get_Sub_right(ptr);
328 if (get_irn_mode(l) == get_irn_mode(ptr) && is_Const(r))
332 /* for now, we support only one substraction, reassoc should fold all others */
333 if (! is_SymConst(ptr) && !is_Sel(ptr))
338 } /* find_constant_entity */
341 * Return the Selection index of a Sel node from dimension n
343 static long get_Sel_array_index_long(ir_node *n, int dim) {
344 ir_node *index = get_Sel_index(n, dim);
345 assert(is_Const(index));
346 return get_tarval_long(get_Const_tarval(index));
347 } /* get_Sel_array_index_long */
350 * Returns the accessed component graph path for an
351 * node computing an address.
353 * @param ptr the node computing the address
354 * @param depth current depth in steps upward from the root
357 static compound_graph_path *rec_get_accessed_path(ir_node *ptr, int depth) {
358 compound_graph_path *res = NULL;
359 ir_entity *root, *field, *ent;
360 int path_len, pos, idx;
364 if (is_SymConst(ptr)) {
365 /* a SymConst. If the depth is 0, this is an access to a global
366 * entity and we don't need a component path, else we know
367 * at least it's length.
369 assert(get_SymConst_kind(ptr) == symconst_addr_ent);
370 root = get_SymConst_entity(ptr);
371 res = (depth == 0) ? NULL : new_compound_graph_path(get_entity_type(root), depth);
372 } else if (is_Sel(ptr)) {
373 /* it's a Sel, go up until we find the root */
374 res = rec_get_accessed_path(get_Sel_ptr(ptr), depth+1);
378 /* fill up the step in the path at the current position */
379 field = get_Sel_entity(ptr);
380 path_len = get_compound_graph_path_length(res);
381 pos = path_len - depth - 1;
382 set_compound_graph_path_node(res, pos, field);
384 if (is_Array_type(get_entity_owner(field))) {
385 assert(get_Sel_n_indexs(ptr) == 1 && "multi dim arrays not implemented");
386 set_compound_graph_path_array_index(res, pos, get_Sel_array_index_long(ptr, 0));
388 } else if (is_Add(ptr)) {
389 ir_node *l = get_Add_left(ptr);
390 ir_node *r = get_Add_right(ptr);
395 tv = get_Const_tarval(r);
398 tv = get_Const_tarval(l);
401 mode = get_tarval_mode(tv);
403 /* ptr must be a Sel or a SymConst, this was checked in find_constant_entity() */
405 field = get_Sel_entity(ptr);
407 field = get_SymConst_entity(ptr);
410 for (ent = field;;) {
412 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(tv, sz);
423 tv = tarval_mod(tv, sz);
425 if (tv_index == tarval_bad || tv == 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) & pn_Cmp_Lt)
439 if (tarval_cmp(tupper, tv_index) & pn_Cmp_Lt)
442 /* ok, bounds check finished */
445 if (! tarval_is_null(tv)) {
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 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) {
503 return 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) {
513 path_entry entry, *p;
514 ir_entity *ent, *field;
515 ir_initializer_t *initializer;
521 if (is_SymConst(ptr)) {
523 ent = get_SymConst_entity(ptr);
524 initializer = get_entity_initializer(ent);
525 for (p = next; p != NULL;) {
526 if (initializer->kind != IR_INITIALIZER_COMPOUND)
528 n = get_initializer_compound_n_entries(initializer);
529 tp = get_entity_type(ent);
531 if (is_Array_type(tp)) {
532 ent = get_array_element_entity(tp);
537 initializer = get_initializer_compound_value(initializer, 0);
541 if (p->index >= (int) n)
543 initializer = get_initializer_compound_value(initializer, p->index);
548 tp = get_entity_type(ent);
549 while (is_Array_type(tp)) {
550 ent = get_array_element_entity(tp);
551 tp = get_entity_type(ent);
553 n = get_initializer_compound_n_entries(initializer);
556 initializer = get_initializer_compound_value(initializer, 0);
559 switch (initializer->kind) {
560 case IR_INITIALIZER_CONST:
561 return get_initializer_const_value(initializer);
562 case IR_INITIALIZER_TARVAL:
563 case IR_INITIALIZER_NULL:
567 } else if (is_Sel(ptr)) {
568 entry.ent = field = get_Sel_entity(ptr);
569 tp = get_entity_owner(field);
570 if (is_Array_type(tp)) {
571 assert(get_Sel_n_indexs(ptr) == 1 && "multi dim arrays not implemented");
572 entry.index = get_Sel_array_index_long(ptr, 0) - get_array_lower_bound_int(tp, 0);
574 int i, n_members = get_compound_n_members(tp);
575 for (i = 0; i < n_members; ++i) {
576 if (get_compound_member(tp, i) == field)
579 if (i >= n_members) {
580 /* not found: should NOT happen */
585 return rec_find_compound_ent_value(get_Sel_ptr(ptr), &entry);
586 } else if (is_Add(ptr)) {
587 ir_node *l = get_Add_left(ptr);
588 ir_node *r = get_Add_right(ptr);
594 tv = get_Const_tarval(r);
597 tv = get_Const_tarval(l);
600 mode = get_tarval_mode(tv);
602 /* ptr must be a Sel or a SymConst, this was checked in find_constant_entity() */
604 field = get_Sel_entity(ptr);
606 field = get_SymConst_entity(ptr);
609 /* count needed entries */
611 for (ent = field;;) {
612 tp = get_entity_type(ent);
613 if (! is_Array_type(tp))
615 ent = get_array_element_entity(tp);
618 /* should be at least ONE entry */
622 /* allocate the right number of entries */
623 NEW_ARR_A(path_entry, p, pos);
627 for (ent = field;;) {
629 tarval *sz, *tv_index, *tlower, *tupper;
633 tp = get_entity_type(ent);
634 if (! is_Array_type(tp))
636 ent = get_array_element_entity(tp);
638 p[pos].next = &p[pos + 1];
640 size = get_type_size_bytes(get_entity_type(ent));
641 sz = new_tarval_from_long(size, mode);
643 tv_index = tarval_div(tv, sz);
644 tv = tarval_mod(tv, sz);
646 if (tv_index == tarval_bad || tv == tarval_bad)
649 assert(get_array_n_dimensions(tp) == 1 && "multiarrays not implemented");
650 bound = get_array_lower_bound(tp, 0);
651 tlower = computed_value(bound);
652 bound = get_array_upper_bound(tp, 0);
653 tupper = computed_value(bound);
655 if (tlower == tarval_bad || tupper == tarval_bad)
658 if (tarval_cmp(tv_index, tlower) & pn_Cmp_Lt)
660 if (tarval_cmp(tupper, tv_index) & pn_Cmp_Lt)
663 /* ok, bounds check finished */
664 index = get_tarval_long(tv_index);
665 p[pos].index = index;
668 if (! tarval_is_null(tv)) {
669 /* hmm, wrong access */
672 p[pos - 1].next = next;
673 return rec_find_compound_ent_value(ptr, p);
674 } else if (is_Sub(ptr)) {
675 ir_node *l = get_Sub_left(ptr);
676 ir_node *r = get_Sub_right(ptr);
679 tv = get_Const_tarval(r);
686 static ir_node *find_compound_ent_value(ir_node *ptr) {
687 return rec_find_compound_ent_value(ptr, NULL);
691 static void reduce_adr_usage(ir_node *ptr);
694 * Update a Load that may lost it's usage.
696 static void handle_load_update(ir_node *load) {
697 ldst_info_t *info = get_irn_link(load);
699 /* do NOT touch volatile loads for now */
700 if (get_Load_volatility(load) == volatility_is_volatile)
703 if (! info->projs[pn_Load_res] && ! info->projs[pn_Load_X_except]) {
704 ir_node *ptr = get_Load_ptr(load);
705 ir_node *mem = get_Load_mem(load);
707 /* a Load which value is neither used nor exception checked, remove it */
708 exchange(info->projs[pn_Load_M], mem);
709 if (info->projs[pn_Load_X_regular])
710 exchange(info->projs[pn_Load_X_regular], new_r_Jmp(current_ir_graph, get_nodes_block(load)));
712 reduce_adr_usage(ptr);
714 } /* handle_load_update */
717 * A Use of an address node is vanished. Check if this was a Proj
718 * node and update the counters.
720 static void reduce_adr_usage(ir_node *ptr) {
722 if (get_irn_n_edges(ptr) <= 0) {
723 /* this Proj is dead now */
724 ir_node *pred = get_Proj_pred(ptr);
727 ldst_info_t *info = get_irn_link(pred);
728 info->projs[get_Proj_proj(ptr)] = NULL;
730 /* this node lost it's result proj, handle that */
731 handle_load_update(pred);
735 } /* reduce_adr_usage */
738 * Check, if an already existing value of mode old_mode can be converted
739 * into the needed one new_mode without loss.
741 static int can_use_stored_value(ir_mode *old_mode, ir_mode *new_mode) {
742 if (old_mode == new_mode)
745 /* if both modes are two-complement ones, we can always convert the
746 Stored value into the needed one. */
747 if (get_mode_size_bits(old_mode) >= get_mode_size_bits(new_mode) &&
748 get_mode_arithmetic(old_mode) == irma_twos_complement &&
749 get_mode_arithmetic(new_mode) == irma_twos_complement)
752 } /* can_use_stored_value */
755 * Check whether a Call is at least pure, ie. does only read memory.
757 static unsigned is_Call_pure(ir_node *call) {
758 ir_type *call_tp = get_Call_type(call);
759 unsigned prop = get_method_additional_properties(call_tp);
761 /* check first the call type */
762 if ((prop & (mtp_property_const|mtp_property_pure)) == 0) {
763 /* try the called entity */
764 ir_node *ptr = get_Call_ptr(call);
766 if (is_Global(ptr)) {
767 ir_entity *ent = get_Global_entity(ptr);
769 prop = get_entity_additional_properties(ent);
772 return (prop & (mtp_property_const|mtp_property_pure)) != 0;
775 static ir_node *get_base_and_offset(ir_node *ptr, long *pOffset)
777 ir_mode *mode = get_irn_mode(ptr);
780 /* TODO: long might not be enough, we should probably use some tarval thingy... */
783 ir_node *l = get_Add_left(ptr);
784 ir_node *r = get_Add_right(ptr);
786 if (get_irn_mode(l) != mode || !is_Const(r))
789 offset += get_tarval_long(get_Const_tarval(r));
791 } else if (is_Sub(ptr)) {
792 ir_node *l = get_Sub_left(ptr);
793 ir_node *r = get_Sub_right(ptr);
795 if (get_irn_mode(l) != mode || !is_Const(r))
798 offset -= get_tarval_long(get_Const_tarval(r));
800 } else if (is_Sel(ptr)) {
801 ir_entity *ent = get_Sel_entity(ptr);
802 ir_type *tp = get_entity_owner(ent);
804 if (is_Array_type(tp)) {
808 /* only one dimensional arrays yet */
809 if (get_Sel_n_indexs(ptr) != 1)
811 index = get_Sel_index(ptr, 0);
812 if (! is_Const(index))
815 tp = get_entity_type(ent);
816 if (get_type_state(tp) != layout_fixed)
819 size = get_type_size_bytes(tp);
820 offset += size * get_tarval_long(get_Const_tarval(index));
822 if (get_type_state(tp) != layout_fixed)
824 offset += get_entity_offset(ent);
826 ptr = get_Sel_ptr(ptr);
835 static int try_load_after_store(ir_node *load,
836 ir_node *load_base_ptr, long load_offset, ir_node *store)
839 ir_node *store_ptr = get_Store_ptr(store);
841 ir_node *store_base_ptr = get_base_and_offset(store_ptr, &store_offset);
842 ir_node *store_value;
851 if (load_base_ptr != store_base_ptr)
854 load_mode = get_Load_mode(load);
855 load_mode_len = get_mode_size_bytes(load_mode);
856 store_mode = get_irn_mode(get_Store_value(store));
857 store_mode_len = get_mode_size_bytes(store_mode);
858 delta = load_offset - store_offset;
859 if (delta < 0 || delta + load_mode_len > store_mode_len)
862 if (get_mode_arithmetic(store_mode) != irma_twos_complement ||
863 get_mode_arithmetic(load_mode) != irma_twos_complement)
866 store_value = get_Store_value(store);
868 /* produce a shift to adjust offset delta */
872 /* FIXME: only true for little endian */
873 cnst = new_Const_long(mode_Iu, delta * 8);
874 store_value = new_r_Shr(current_ir_graph, get_nodes_block(load),
875 store_value, cnst, store_mode);
878 /* add an convert if needed */
879 if (store_mode != load_mode) {
880 store_value = new_r_Conv(current_ir_graph, get_nodes_block(load),
881 store_value, load_mode);
884 DBG_OPT_RAW(load, store_value);
886 info = get_irn_link(load);
887 if (info->projs[pn_Load_M])
888 exchange(info->projs[pn_Load_M], get_Load_mem(load));
892 if (info->projs[pn_Load_X_except]) {
893 exchange( info->projs[pn_Load_X_except], new_Bad());
896 if (info->projs[pn_Load_X_regular]) {
897 exchange( info->projs[pn_Load_X_regular], new_r_Jmp(current_ir_graph, get_nodes_block(load)));
901 if (info->projs[pn_Load_res])
902 exchange(info->projs[pn_Load_res], store_value);
904 load_ptr = get_Load_ptr(load);
906 reduce_adr_usage(load_ptr);
907 return res | DF_CHANGED;
911 * Follow the memory chain as long as there are only Loads,
912 * alias free Stores, and constant Calls and try to replace the
913 * current Load by a previous ones.
914 * Note that in unreachable loops it might happen that we reach
915 * load again, as well as we can fall into a cycle.
916 * We break such cycles using a special visited flag.
918 * INC_MASTER() must be called before dive into
920 static unsigned follow_Mem_chain(ir_node *load, ir_node *curr) {
922 ldst_info_t *info = get_irn_link(load);
924 ir_node *ptr = get_Load_ptr(load);
925 ir_node *mem = get_Load_mem(load);
926 ir_mode *load_mode = get_Load_mode(load);
928 for (pred = curr; load != pred; ) {
929 ldst_info_t *pred_info = get_irn_link(pred);
932 * a Load immediately after a Store -- a read after write.
933 * We may remove the Load, if both Load & Store does not have an
934 * exception handler OR they are in the same MacroBlock. In the latter
935 * case the Load cannot throw an exception when the previous Store was
938 * Why we need to check for Store Exception? If the Store cannot
939 * be executed (ROM) the exception handler might simply jump into
940 * the load MacroBlock :-(
941 * We could make it a little bit better if we would know that the
942 * exception handler of the Store jumps directly to the end...
944 if (is_Store(pred) && ((pred_info->projs[pn_Store_X_except] == NULL
945 && info->projs[pn_Load_X_except] == NULL)
946 || get_nodes_MacroBlock(load) == get_nodes_MacroBlock(pred)))
949 ir_node *base_ptr = get_base_and_offset(ptr, &load_offset);
950 int changes = try_load_after_store(load, base_ptr, load_offset, pred);
953 return res | changes;
954 } else if (is_Load(pred) && get_Load_ptr(pred) == ptr &&
955 can_use_stored_value(get_Load_mode(pred), load_mode)) {
957 * a Load after a Load -- a read after read.
958 * We may remove the second Load, if it does not have an exception handler
959 * OR they are in the same MacroBlock. In the later case the Load cannot
960 * throw an exception when the previous Load was quiet.
962 * Here, there is no need to check if the previous Load has an exception
963 * hander because they would have exact the same exception...
965 if (info->projs[pn_Load_X_except] == NULL || get_nodes_MacroBlock(load) == get_nodes_MacroBlock(pred)) {
968 DBG_OPT_RAR(load, pred);
970 /* the result is used */
971 if (info->projs[pn_Load_res]) {
972 if (pred_info->projs[pn_Load_res] == NULL) {
973 /* create a new Proj again */
974 pred_info->projs[pn_Load_res] = new_r_Proj(current_ir_graph, get_nodes_block(pred), pred, get_Load_mode(pred), pn_Load_res);
976 value = pred_info->projs[pn_Load_res];
978 /* add an convert if needed */
979 if (get_Load_mode(pred) != load_mode) {
980 value = new_r_Conv(current_ir_graph, get_nodes_block(load), value, load_mode);
983 exchange(info->projs[pn_Load_res], value);
986 if (info->projs[pn_Load_M])
987 exchange(info->projs[pn_Load_M], mem);
990 if (info->projs[pn_Load_X_except]) {
991 exchange(info->projs[pn_Load_X_except], new_Bad());
994 if (info->projs[pn_Load_X_regular]) {
995 exchange( info->projs[pn_Load_X_regular], new_r_Jmp(current_ir_graph, get_nodes_block(load)));
1000 reduce_adr_usage(ptr);
1001 return res |= DF_CHANGED;
1005 if (is_Store(pred)) {
1006 /* check if we can pass through this store */
1007 ir_alias_relation rel = get_alias_relation(
1009 get_Store_ptr(pred),
1010 get_irn_mode(get_Store_value(pred)),
1012 /* if the might be an alias, we cannot pass this Store */
1013 if (rel != ir_no_alias)
1015 pred = skip_Proj(get_Store_mem(pred));
1016 } else if (is_Load(pred)) {
1017 pred = skip_Proj(get_Load_mem(pred));
1018 } else if (is_Call(pred)) {
1019 if (is_Call_pure(pred)) {
1020 /* The called graph is at least pure, so there are no Store's
1021 in it. We can handle it like a Load and skip it. */
1022 pred = skip_Proj(get_Call_mem(pred));
1024 /* there might be Store's in the graph, stop here */
1028 /* follow only Load chains */
1032 /* check for cycles */
1033 if (NODE_VISITED(pred_info))
1035 MARK_NODE(pred_info);
1038 if (is_Sync(pred)) {
1041 /* handle all Sync predecessors */
1042 for (i = get_Sync_n_preds(pred) - 1; i >= 0; --i) {
1043 res |= follow_Mem_chain(load, skip_Proj(get_Sync_pred(pred, i)));
1050 } /* follow_Mem_chain */
1053 * Check if we can replace the load by a given const from
1054 * the const code irg.
1056 ir_node *can_replace_load_by_const(const ir_node *load, ir_node *c) {
1057 ir_mode *c_mode = get_irn_mode(c);
1058 ir_mode *l_mode = get_Load_mode(load);
1059 ir_node *res = NULL;
1061 if (c_mode != l_mode) {
1062 /* check, if the mode matches OR can be easily converted info */
1063 if (is_reinterpret_cast(c_mode, l_mode)) {
1064 /* we can safely cast */
1065 dbg_info *dbg = get_irn_dbg_info(load);
1066 ir_node *block = get_nodes_block(load);
1068 /* copy the value from the const code irg and cast it */
1069 res = copy_const_value(dbg, c);
1070 res = new_rd_Conv(dbg, current_ir_graph, block, res, l_mode);
1073 /* copy the value from the const code irg */
1074 res = copy_const_value(get_irn_dbg_info(load), c);
1077 } /* can_replace_load_by_const */
1082 * @param load the Load node
1084 static unsigned optimize_load(ir_node *load)
1086 ldst_info_t *info = get_irn_link(load);
1087 ir_node *mem, *ptr, *value;
1092 /* do NOT touch volatile loads for now */
1093 if (get_Load_volatility(load) == volatility_is_volatile)
1096 /* the address of the load to be optimized */
1097 ptr = get_Load_ptr(load);
1100 * Check if we can remove the exception from a Load:
1101 * This can be done, if the address is from an Sel(Alloc) and
1102 * the Sel type is a subtype of the allocated type.
1104 * This optimizes some often used OO constructs,
1105 * like x = new O; x->t;
1107 if (info->projs[pn_Load_X_except]) {
1108 ir_node *addr = ptr;
1110 /* find base address */
1111 while (is_Sel(addr))
1112 addr = get_Sel_ptr(addr);
1113 if (is_Alloc(skip_Proj(skip_Cast(addr)))) {
1114 /* simple case: a direct load after an Alloc. Firm Alloc throw
1115 * an exception in case of out-of-memory. So, there is no way for an
1116 * exception in this load.
1117 * This code is constructed by the "exception lowering" in the Jack compiler.
1119 exchange(info->projs[pn_Load_X_except], new_Bad());
1120 info->projs[pn_Load_X_except] = NULL;
1121 exchange(info->projs[pn_Load_X_regular], new_r_Jmp(current_ir_graph, get_nodes_block(load)));
1122 info->projs[pn_Load_X_regular] = NULL;
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] && ! info->projs[pn_Load_X_except]) {
1131 /* a Load which value is neither used nor exception checked, remove it */
1132 exchange(info->projs[pn_Load_M], mem);
1134 if (info->projs[pn_Load_X_regular]) {
1135 /* should not happen, but if it does, remove it */
1136 exchange(info->projs[pn_Load_X_regular], new_r_Jmp(current_ir_graph, get_nodes_block(load)));
1140 reduce_adr_usage(ptr);
1141 return res | DF_CHANGED;
1144 /* Load from a constant polymorphic field, where we can resolve
1146 value = transform_polymorph_Load(load);
1147 if (value == load) {
1149 /* check if we can determine the entity that will be loaded */
1150 ent = find_constant_entity(ptr);
1152 if ((allocation_static == get_entity_allocation(ent)) &&
1153 (visibility_external_allocated != get_entity_visibility(ent))) {
1154 /* a static allocation that is not external: there should be NO exception
1155 * when loading even if we cannot replace the load itself. */
1157 /* no exception, clear the info field as it might be checked later again */
1158 if (info->projs[pn_Load_X_except]) {
1159 exchange(info->projs[pn_Load_X_except], new_Bad());
1160 info->projs[pn_Load_X_except] = NULL;
1163 if (info->projs[pn_Load_X_regular]) {
1164 exchange(info->projs[pn_Load_X_regular], new_r_Jmp(current_ir_graph, get_nodes_block(load)));
1165 info->projs[pn_Load_X_regular] = NULL;
1169 if (variability_constant == get_entity_variability(ent)) {
1170 if (is_atomic_entity(ent)) {
1171 /* Might not be atomic after
1172 lowering of Sels. In this
1173 case we could also load, but
1174 it's more complicated. */
1175 /* more simpler case: we load the content of a constant value:
1176 * replace it by the constant itself
1178 value = get_atomic_ent_value(ent);
1180 if (ent->has_initializer) {
1181 /* new style initializer */
1182 value = find_compound_ent_value(ptr);
1184 /* old style initializer */
1185 compound_graph_path *path = get_accessed_path(ptr);
1188 assert(is_proper_compound_graph_path(path, get_compound_graph_path_length(path)-1));
1190 value = get_compound_ent_value_by_path(ent, path);
1191 free_compound_graph_path(path);
1196 value = can_replace_load_by_const(load, value);
1201 if (value != NULL) {
1202 /* we completely replace the load by this value */
1203 if (info->projs[pn_Load_X_except]) {
1204 exchange(info->projs[pn_Load_X_except], new_Bad());
1205 info->projs[pn_Load_X_except] = NULL;
1208 if (info->projs[pn_Load_X_regular]) {
1209 exchange(info->projs[pn_Load_X_regular], new_r_Jmp(current_ir_graph, get_nodes_block(load)));
1210 info->projs[pn_Load_X_regular] = NULL;
1213 if (info->projs[pn_Load_M]) {
1214 exchange(info->projs[pn_Load_M], mem);
1217 if (info->projs[pn_Load_res]) {
1218 exchange(info->projs[pn_Load_res], value);
1222 reduce_adr_usage(ptr);
1226 /* Check, if the address of this load is used more than once.
1227 * If not, more load cannot be removed in any case. */
1228 if (get_irn_n_uses(ptr) <= 1 && get_irn_n_uses(get_base_and_offset(ptr, &dummy)) <= 1)
1232 * follow the memory chain as long as there are only Loads
1233 * and try to replace current Load or Store by a previous one.
1234 * Note that in unreachable loops it might happen that we reach
1235 * load again, as well as we can fall into a cycle.
1236 * We break such cycles using a special visited flag.
1239 res = follow_Mem_chain(load, skip_Proj(mem));
1241 } /* optimize_load */
1244 * Check whether a value of mode new_mode would completely overwrite a value
1245 * of mode old_mode in memory.
1247 static int is_completely_overwritten(ir_mode *old_mode, ir_mode *new_mode)
1249 return get_mode_size_bits(new_mode) >= get_mode_size_bits(old_mode);
1250 } /* is_completely_overwritten */
1253 * follow the memory chain as long as there are only Loads and alias free Stores.
1255 * INC_MASTER() must be called before dive into
1257 static unsigned follow_Mem_chain_for_Store(ir_node *store, ir_node *curr) {
1259 ldst_info_t *info = get_irn_link(store);
1261 ir_node *ptr = get_Store_ptr(store);
1262 ir_node *mem = get_Store_mem(store);
1263 ir_node *value = get_Store_value(store);
1264 ir_mode *mode = get_irn_mode(value);
1265 ir_node *block = get_nodes_block(store);
1266 ir_node *mblk = get_Block_MacroBlock(block);
1268 for (pred = curr; pred != store;) {
1269 ldst_info_t *pred_info = get_irn_link(pred);
1272 * BEWARE: one might think that checking the modes is useless, because
1273 * if the pointers are identical, they refer to the same object.
1274 * This is only true in strong typed languages, not is C were the following
1275 * is possible *(ir_type1 *)p = a; *(ir_type2 *)p = b ...
1276 * However, if the mode that is written have a bigger or equal size the the old
1277 * one, the old value is completely overwritten and can be killed ...
1279 if (is_Store(pred) && get_Store_ptr(pred) == ptr &&
1280 get_nodes_MacroBlock(pred) == mblk &&
1281 is_completely_overwritten(get_irn_mode(get_Store_value(pred)), mode)) {
1283 * a Store after a Store in the same MacroBlock -- a write after write.
1284 * We may remove the first Store, if it does not have an exception handler.
1286 * TODO: What, if both have the same exception handler ???
1288 if (get_Store_volatility(pred) != volatility_is_volatile && !pred_info->projs[pn_Store_X_except]) {
1289 DBG_OPT_WAW(pred, store);
1290 exchange(pred_info->projs[pn_Store_M], get_Store_mem(pred));
1292 reduce_adr_usage(ptr);
1295 } else if (is_Load(pred) && get_Load_ptr(pred) == ptr &&
1296 value == pred_info->projs[pn_Load_res]) {
1298 * a Store of a value just loaded from the same address
1299 * -- a write after read.
1300 * We may remove the Store, if it does not have an exception
1303 if (! info->projs[pn_Store_X_except]) {
1304 DBG_OPT_WAR(store, pred);
1305 exchange(info->projs[pn_Store_M], mem);
1307 reduce_adr_usage(ptr);
1312 if (is_Store(pred)) {
1313 /* check if we can pass through this store */
1314 ir_alias_relation rel = get_alias_relation(
1316 get_Store_ptr(pred),
1317 get_irn_mode(get_Store_value(pred)),
1319 /* if the might be an alias, we cannot pass this Store */
1320 if (rel != ir_no_alias)
1322 pred = skip_Proj(get_Store_mem(pred));
1323 } else if (is_Load(pred)) {
1324 ir_alias_relation rel = get_alias_relation(
1325 current_ir_graph, get_Load_ptr(pred), get_Load_mode(pred),
1327 if (rel != ir_no_alias)
1330 pred = skip_Proj(get_Load_mem(pred));
1332 /* follow only Load chains */
1336 /* check for cycles */
1337 if (NODE_VISITED(pred_info))
1339 MARK_NODE(pred_info);
1342 if (is_Sync(pred)) {
1345 /* handle all Sync predecessors */
1346 for (i = get_Sync_n_preds(pred) - 1; i >= 0; --i) {
1347 res |= follow_Mem_chain_for_Store(store, skip_Proj(get_Sync_pred(pred, i)));
1353 } /* follow_Mem_chain_for_Store */
1355 /** find entity used as base for an address calculation */
1356 static ir_entity *find_entity(ir_node *ptr)
1358 switch(get_irn_opcode(ptr)) {
1360 return get_SymConst_entity(ptr);
1362 ir_node *pred = get_Sel_ptr(ptr);
1363 if (get_irg_frame(get_irn_irg(ptr)) == pred)
1364 return get_Sel_entity(ptr);
1366 return find_entity(pred);
1370 ir_node *left = get_binop_left(ptr);
1372 if (mode_is_reference(get_irn_mode(left)))
1373 return find_entity(left);
1374 right = get_binop_right(ptr);
1375 if (mode_is_reference(get_irn_mode(right)))
1376 return find_entity(right);
1387 * @param store the Store node
1389 static unsigned optimize_store(ir_node *store) {
1394 if (get_Store_volatility(store) == volatility_is_volatile)
1397 ptr = get_Store_ptr(store);
1398 entity = find_entity(ptr);
1400 /* a store to an entity which is never read is unnecessary */
1401 if (entity != NULL && !(get_entity_usage(entity) & ir_usage_read)) {
1402 ldst_info_t *info = get_irn_link(store);
1403 if (info->projs[pn_Store_X_except] == NULL) {
1404 exchange(info->projs[pn_Store_M], get_Store_mem(store));
1406 reduce_adr_usage(ptr);
1411 /* Check, if the address of this Store is used more than once.
1412 * If not, this Store cannot be removed in any case. */
1413 if (get_irn_n_uses(ptr) <= 1)
1416 mem = get_Store_mem(store);
1418 /* follow the memory chain as long as there are only Loads */
1421 return follow_Mem_chain_for_Store(store, skip_Proj(mem));
1422 } /* optimize_store */
1425 * walker, optimizes Phi after Stores to identical places:
1426 * Does the following optimization:
1429 * val1 val2 val3 val1 val2 val3
1431 * Store Store Store \ | /
1438 * This reduces the number of stores and allows for predicated execution.
1439 * Moves Stores back to the end of a function which may be bad.
1441 * This is only possible if the predecessor blocks have only one successor.
1443 static unsigned optimize_phi(ir_node *phi, walk_env_t *wenv)
1446 ir_node *store, *old_store, *ptr, *block, *phi_block, *phiM, *phiD, *exc, *projM;
1448 ir_node **inM, **inD, **projMs;
1450 dbg_info *db = NULL;
1452 block_info_t *bl_info;
1455 /* Must be a memory Phi */
1456 if (get_irn_mode(phi) != mode_M)
1459 n = get_Phi_n_preds(phi);
1463 /* must be only one user */
1464 projM = get_Phi_pred(phi, 0);
1465 if (get_irn_n_edges(projM) != 1)
1468 store = skip_Proj(projM);
1470 if (!is_Store(store))
1473 block = get_nodes_block(store);
1475 /* abort on dead blocks */
1476 if (is_Block_dead(block))
1479 /* check if the block is post dominated by Phi-block
1480 and has no exception exit */
1481 bl_info = get_irn_link(block);
1482 if (bl_info->flags & BLOCK_HAS_EXC)
1485 phi_block = get_nodes_block(phi);
1486 if (! block_strictly_postdominates(phi_block, block))
1489 /* this is the address of the store */
1490 ptr = get_Store_ptr(store);
1491 mode = get_irn_mode(get_Store_value(store));
1492 info = get_irn_link(store);
1493 exc = info->exc_block;
1495 for (i = 1; i < n; ++i) {
1496 ir_node *pred = get_Phi_pred(phi, i);
1498 if (get_irn_n_edges(pred) != 1)
1501 pred = skip_Proj(pred);
1502 if (!is_Store(pred))
1505 if (ptr != get_Store_ptr(pred) || mode != get_irn_mode(get_Store_value(pred)))
1508 info = get_irn_link(pred);
1510 /* check, if all stores have the same exception flow */
1511 if (exc != info->exc_block)
1514 /* abort on dead blocks */
1515 block = get_nodes_block(pred);
1516 if (is_Block_dead(block))
1519 /* check if the block is post dominated by Phi-block
1520 and has no exception exit. Note that block must be different from
1521 Phi-block, else we would move a Store from end End of a block to its
1523 bl_info = get_irn_link(block);
1524 if (bl_info->flags & BLOCK_HAS_EXC)
1526 if (block == phi_block || ! block_postdominates(phi_block, block))
1531 * ok, when we are here, we found all predecessors of a Phi that
1532 * are Stores to the same address and size. That means whatever
1533 * we do before we enter the block of the Phi, we do a Store.
1534 * So, we can move the Store to the current block:
1536 * val1 val2 val3 val1 val2 val3
1538 * | Str | | Str | | Str | \ | /
1544 * Is only allowed if the predecessor blocks have only one successor.
1547 NEW_ARR_A(ir_node *, projMs, n);
1548 NEW_ARR_A(ir_node *, inM, n);
1549 NEW_ARR_A(ir_node *, inD, n);
1550 NEW_ARR_A(int, idx, n);
1552 /* Prepare: Collect all Store nodes. We must do this
1553 first because we otherwise may loose a store when exchanging its
1556 for (i = n - 1; i >= 0; --i) {
1559 projMs[i] = get_Phi_pred(phi, i);
1560 assert(is_Proj(projMs[i]));
1562 store = get_Proj_pred(projMs[i]);
1563 info = get_irn_link(store);
1565 inM[i] = get_Store_mem(store);
1566 inD[i] = get_Store_value(store);
1567 idx[i] = info->exc_idx;
1569 block = get_nodes_block(phi);
1571 /* second step: create a new memory Phi */
1572 phiM = new_rd_Phi(get_irn_dbg_info(phi), current_ir_graph, block, n, inM, mode_M);
1574 /* third step: create a new data Phi */
1575 phiD = new_rd_Phi(get_irn_dbg_info(phi), current_ir_graph, block, n, inD, mode);
1577 /* rewire memory and kill the node */
1578 for (i = n - 1; i >= 0; --i) {
1579 ir_node *proj = projMs[i];
1582 ir_node *store = get_Proj_pred(proj);
1583 exchange(proj, inM[i]);
1588 /* fourth step: create the Store */
1589 store = new_rd_Store(db, current_ir_graph, block, phiM, ptr, phiD);
1591 co_set_irn_name(store, co_get_irn_ident(old_store));
1594 projM = new_rd_Proj(NULL, current_ir_graph, block, store, mode_M, pn_Store_M);
1596 info = get_ldst_info(store, &wenv->obst);
1597 info->projs[pn_Store_M] = projM;
1599 /* fifths step: repair exception flow */
1601 ir_node *projX = new_rd_Proj(NULL, current_ir_graph, block, store, mode_X, pn_Store_X_except);
1603 info->projs[pn_Store_X_except] = projX;
1604 info->exc_block = exc;
1605 info->exc_idx = idx[0];
1607 for (i = 0; i < n; ++i) {
1608 set_Block_cfgpred(exc, idx[i], projX);
1612 /* the exception block should be optimized as some inputs are identical now */
1618 /* sixth step: replace old Phi */
1619 exchange(phi, projM);
1621 return res | DF_CHANGED;
1622 } /* optimize_phi */
1625 * walker, do the optimizations
1627 static void do_load_store_optimize(ir_node *n, void *env) {
1628 walk_env_t *wenv = env;
1630 switch (get_irn_opcode(n)) {
1633 wenv->changes |= optimize_load(n);
1637 wenv->changes |= optimize_store(n);
1641 wenv->changes |= optimize_phi(n, wenv);
1647 } /* do_load_store_optimize */
1650 typedef struct scc {
1651 ir_node *head; /**< the head of the list */
1654 /** A node entry. */
1655 typedef struct node_entry {
1656 unsigned DFSnum; /**< the DFS number of this node */
1657 unsigned low; /**< the low number of this node */
1658 ir_node *header; /**< the header of this node */
1659 int in_stack; /**< flag, set if the node is on the stack */
1660 ir_node *next; /**< link to the next node the the same scc */
1661 scc *pscc; /**< the scc of this node */
1662 unsigned POnum; /**< the post order number for blocks */
1665 /** A loop entry. */
1666 typedef struct loop_env {
1667 ir_phase ph; /**< the phase object */
1668 ir_node **stack; /**< the node stack */
1669 int tos; /**< tos index */
1670 unsigned nextDFSnum; /**< the current DFS number */
1671 unsigned POnum; /**< current post order number */
1673 unsigned changes; /**< a bitmask of graph changes */
1677 * Gets the node_entry of a node
1679 static node_entry *get_irn_ne(ir_node *irn, loop_env *env) {
1680 ir_phase *ph = &env->ph;
1681 node_entry *e = phase_get_irn_data(&env->ph, irn);
1684 e = phase_alloc(ph, sizeof(*e));
1685 memset(e, 0, sizeof(*e));
1686 phase_set_irn_data(ph, irn, e);
1692 * Push a node onto the stack.
1694 * @param env the loop environment
1695 * @param n the node to push
1697 static void push(loop_env *env, ir_node *n) {
1700 if (env->tos == ARR_LEN(env->stack)) {
1701 int nlen = ARR_LEN(env->stack) * 2;
1702 ARR_RESIZE(ir_node *, env->stack, nlen);
1704 env->stack[env->tos++] = n;
1705 e = get_irn_ne(n, env);
1710 * pop a node from the stack
1712 * @param env the loop environment
1714 * @return The topmost node
1716 static ir_node *pop(loop_env *env) {
1717 ir_node *n = env->stack[--env->tos];
1718 node_entry *e = get_irn_ne(n, env);
1725 * Check if irn is a region constant.
1726 * The block or irn must strictly dominate the header block.
1728 * @param irn the node to check
1729 * @param header_block the header block of the induction variable
1731 static int is_rc(ir_node *irn, ir_node *header_block) {
1732 ir_node *block = get_nodes_block(irn);
1734 return (block != header_block) && block_dominates(block, header_block);
1737 typedef struct phi_entry phi_entry;
1739 ir_node *phi; /**< A phi with a region const memory. */
1740 int pos; /**< The position of the region const memory */
1741 ir_node *load; /**< the newly created load for this phi */
1746 * Move loops out of loops if possible.
1748 * @param pscc the loop described by an SCC
1749 * @param env the loop environment
1751 static void move_loads_out_of_loops(scc *pscc, loop_env *env) {
1752 ir_node *phi, *load, *next, *other, *next_other;
1755 phi_entry *phi_list = NULL;
1757 /* collect all outer memories */
1758 for (phi = pscc->head; phi != NULL; phi = next) {
1759 node_entry *ne = get_irn_ne(phi, env);
1762 /* check all memory Phi's */
1766 assert(get_irn_mode(phi) == mode_M && "DFS geturn non-memory Phi");
1768 for (j = get_irn_arity(phi) - 1; j >= 0; --j) {
1769 ir_node *pred = get_irn_n(phi, j);
1770 node_entry *pe = get_irn_ne(pred, env);
1772 if (pe->pscc != ne->pscc) {
1773 /* not in the same SCC, is region const */
1774 phi_entry *pe = phase_alloc(&env->ph, sizeof(*pe));
1778 pe->next = phi_list;
1783 /* no Phis no fun */
1784 assert(phi_list != NULL && "DFS found a loop without Phi");
1786 for (load = pscc->head; load; load = next) {
1788 node_entry *ne = get_irn_ne(load, env);
1791 if (is_Load(load)) {
1792 ldst_info_t *info = get_irn_link(load);
1793 ir_node *ptr = get_Load_ptr(load);
1795 /* for now, we cannot handle Loads with exceptions */
1796 if (info->projs[pn_Load_res] == NULL || info->projs[pn_Load_X_regular] != NULL || info->projs[pn_Load_X_except] != NULL)
1799 /* for now, we can only handle Load(Global) */
1800 if (! is_Global(ptr))
1802 ent = get_Global_entity(ptr);
1803 load_mode = get_Load_mode(load);
1804 for (other = pscc->head; other != NULL; other = next_other) {
1805 node_entry *ne = get_irn_ne(other, env);
1806 next_other = ne->next;
1808 if (is_Store(other)) {
1809 ir_alias_relation rel = get_alias_relation(
1811 get_Store_ptr(other),
1812 get_irn_mode(get_Store_value(other)),
1814 /* if the might be an alias, we cannot pass this Store */
1815 if (rel != ir_no_alias)
1818 /* only pure Calls are allowed here, so ignore them */
1820 if (other == NULL) {
1825 /* for now, we cannot handle more than one input */
1826 if (phi_list->next != NULL)
1829 /* yep, no aliasing Store found, Load can be moved */
1830 DB((dbg, LEVEL_1, " Found a Load that could be moved: %+F\n", load));
1832 db = get_irn_dbg_info(load);
1833 for (pe = phi_list; pe != NULL; pe = pe->next) {
1835 ir_node *phi = pe->phi;
1836 ir_node *blk = get_nodes_block(phi);
1837 ir_node *pred = get_Block_cfgpred_block(blk, pos);
1840 pe->load = irn = new_rd_Load(db, current_ir_graph, pred, get_Phi_pred(phi, pos), ptr, load_mode);
1841 ninfo = get_ldst_info(irn, phase_obst(&env->ph));
1843 ninfo->projs[pn_Load_M] = mem = new_r_Proj(current_ir_graph, pred, irn, mode_M, pn_Load_M);
1844 set_Phi_pred(phi, pos, mem);
1846 ninfo->projs[pn_Load_res] = new_r_Proj(current_ir_graph, pred, irn, load_mode, pn_Load_res);
1848 DB((dbg, LEVEL_1, " Created %+F in %+F\n", irn, pred));
1851 /* now kill the old Load */
1852 exchange(info->projs[pn_Load_M], get_Load_mem(load));
1853 exchange(info->projs[pn_Load_res], ninfo->projs[pn_Load_res]);
1855 env->changes |= DF_CHANGED;
1859 } /* move_loads_out_of_loops */
1862 * Process a loop SCC.
1864 * @param pscc the SCC
1865 * @param env the loop environment
1867 static void process_loop(scc *pscc, loop_env *env) {
1868 ir_node *irn, *next, *header = NULL;
1869 node_entry *b, *h = NULL;
1870 int j, only_phi, num_outside, process = 0;
1873 /* find the header block for this scc */
1874 for (irn = pscc->head; irn; irn = next) {
1875 node_entry *e = get_irn_ne(irn, env);
1876 ir_node *block = get_nodes_block(irn);
1879 b = get_irn_ne(block, env);
1882 if (h->POnum < b->POnum) {
1893 /* check if this scc contains only Phi, Loads or Stores nodes */
1897 for (irn = pscc->head; irn; irn = next) {
1898 node_entry *e = get_irn_ne(irn, env);
1901 switch (get_irn_opcode(irn)) {
1903 if (is_Call_pure(irn)) {
1904 /* pure calls can be treated like loads */
1908 /* non-pure calls must be handle like may-alias Stores */
1911 /* cannot handle CopyB yet */
1915 if (get_Load_volatility(irn) == volatility_is_volatile) {
1916 /* cannot handle loops with volatile Loads */
1922 if (get_Store_volatility(irn) == volatility_is_volatile) {
1923 /* cannot handle loops with volatile Stores */
1932 for (j = get_irn_arity(irn) - 1; j >= 0; --j) {
1933 ir_node *pred = get_irn_n(irn, j);
1934 node_entry *pe = get_irn_ne(pred, env);
1936 if (pe->pscc != e->pscc) {
1937 /* not in the same SCC, must be a region const */
1938 if (! is_rc(pred, header)) {
1939 /* not a memory loop */
1945 } else if (out_rc != pred) {
1956 /* found a memory loop */
1957 DB((dbg, LEVEL_2, " Found a memory loop:\n "));
1958 if (only_phi && num_outside == 1) {
1959 /* a phi cycle with only one real predecessor can be collapsed */
1960 DB((dbg, LEVEL_2, " Found an USELESS Phi cycle:\n "));
1962 for (irn = pscc->head; irn; irn = next) {
1963 node_entry *e = get_irn_ne(irn, env);
1966 exchange(irn, out_rc);
1968 env->changes |= DF_CHANGED;
1972 /* set the header for every node in this scc */
1973 for (irn = pscc->head; irn; irn = next) {
1974 node_entry *e = get_irn_ne(irn, env);
1977 DB((dbg, LEVEL_2, " %+F,", irn));
1979 DB((dbg, LEVEL_2, "\n"));
1981 move_loads_out_of_loops(pscc, env);
1985 } /* process_loop */
1990 * @param pscc the SCC
1991 * @param env the loop environment
1993 static void process_scc(scc *pscc, loop_env *env) {
1994 ir_node *head = pscc->head;
1995 node_entry *e = get_irn_ne(head, env);
1997 #ifdef DEBUG_libfirm
1999 ir_node *irn, *next;
2001 DB((dbg, LEVEL_4, " SCC at %p:\n ", pscc));
2002 for (irn = pscc->head; irn; irn = next) {
2003 node_entry *e = get_irn_ne(irn, env);
2007 DB((dbg, LEVEL_4, " %+F,", irn));
2009 DB((dbg, LEVEL_4, "\n"));
2013 if (e->next != NULL) {
2014 /* this SCC has more than one member */
2015 process_loop(pscc, env);
2020 * Do Tarjan's SCC algorithm and drive load/store optimization.
2022 * @param irn start at this node
2023 * @param env the loop environment
2025 static void dfs(ir_node *irn, loop_env *env)
2028 node_entry *node = get_irn_ne(irn, env);
2030 mark_irn_visited(irn);
2032 node->DFSnum = env->nextDFSnum++;
2033 node->low = node->DFSnum;
2037 if (is_Phi(irn) || is_Sync(irn)) {
2038 n = get_irn_arity(irn);
2039 for (i = 0; i < n; ++i) {
2040 ir_node *pred = get_irn_n(irn, i);
2041 node_entry *o = get_irn_ne(pred, env);
2043 if (irn_not_visited(pred)) {
2045 node->low = MIN(node->low, o->low);
2047 if (o->DFSnum < node->DFSnum && o->in_stack)
2048 node->low = MIN(o->DFSnum, node->low);
2050 } else if (is_fragile_op(irn)) {
2051 ir_node *pred = get_fragile_op_mem(irn);
2052 node_entry *o = get_irn_ne(pred, env);
2054 if (irn_not_visited(pred)) {
2056 node->low = MIN(node->low, o->low);
2058 if (o->DFSnum < node->DFSnum && o->in_stack)
2059 node->low = MIN(o->DFSnum, node->low);
2060 } else if (is_Proj(irn)) {
2061 ir_node *pred = get_Proj_pred(irn);
2062 node_entry *o = get_irn_ne(pred, env);
2064 if (irn_not_visited(pred)) {
2066 node->low = MIN(node->low, o->low);
2068 if (o->DFSnum < node->DFSnum && o->in_stack)
2069 node->low = MIN(o->DFSnum, node->low);
2072 /* IGNORE predecessors */
2075 if (node->low == node->DFSnum) {
2076 scc *pscc = phase_alloc(&env->ph, sizeof(*pscc));
2084 e = get_irn_ne(x, env);
2086 e->next = pscc->head;
2090 process_scc(pscc, env);
2095 * Do the DFS on the memory edges a graph.
2097 * @param irg the graph to process
2098 * @param env the loop environment
2100 static void do_dfs(ir_graph *irg, loop_env *env) {
2101 ir_graph *rem = current_ir_graph;
2102 ir_node *endblk, *end;
2105 current_ir_graph = irg;
2106 inc_irg_visited(irg);
2108 /* visit all memory nodes */
2109 endblk = get_irg_end_block(irg);
2110 for (i = get_Block_n_cfgpreds(endblk) - 1; i >= 0; --i) {
2111 ir_node *pred = get_Block_cfgpred(endblk, i);
2113 pred = skip_Proj(pred);
2114 if (is_Return(pred))
2115 dfs(get_Return_mem(pred), env);
2116 else if (is_Raise(pred))
2117 dfs(get_Raise_mem(pred), env);
2118 else if (is_fragile_op(pred))
2119 dfs(get_fragile_op_mem(pred), env);
2121 assert(0 && "Unknown EndBlock predecessor");
2125 /* visit the keep-alives */
2126 end = get_irg_end(irg);
2127 for (i = get_End_n_keepalives(end) - 1; i >= 0; --i) {
2128 ir_node *ka = get_End_keepalive(end, i);
2130 if (is_Phi(ka) && irn_not_visited(ka))
2133 current_ir_graph = rem;
2137 * Initialize new phase data. We do this always explicit, so return NULL here
2139 static void *init_loop_data(ir_phase *ph, const ir_node *irn, void *data) {
2144 } /* init_loop_data */
2147 * Optimize Loads/Stores in loops.
2149 * @param irg the graph
2151 static int optimize_loops(ir_graph *irg) {
2154 env.stack = NEW_ARR_F(ir_node *, 128);
2159 phase_init(&env.ph, "ldstopt", irg, PHASE_DEFAULT_GROWTH, init_loop_data, NULL);
2161 /* calculate the SCC's and drive loop optimization. */
2164 DEL_ARR_F(env.stack);
2165 phase_free(&env.ph);
2168 } /* optimize_loops */
2171 * do the load store optimization
2173 void optimize_load_store(ir_graph *irg) {
2176 FIRM_DBG_REGISTER(dbg, "firm.opt.ldstopt");
2178 assert(get_irg_phase_state(irg) != phase_building);
2179 assert(get_irg_pinned(irg) != op_pin_state_floats &&
2180 "LoadStore optimization needs pinned graph");
2182 /* we need landing pads */
2183 remove_critical_cf_edges(irg);
2187 /* for Phi optimization post-dominators are needed ... */
2188 assure_postdoms(irg);
2190 if (get_opt_alias_analysis()) {
2191 assure_irg_entity_usage_computed(irg);
2192 assure_irp_globals_entity_usage_computed();
2195 obstack_init(&env.obst);
2198 /* init the links, then collect Loads/Stores/Proj's in lists */
2200 irg_walk_graph(irg, firm_clear_link, collect_nodes, &env);
2202 /* now we have collected enough information, optimize */
2203 irg_walk_graph(irg, NULL, do_load_store_optimize, &env);
2205 env.changes |= optimize_loops(irg);
2207 obstack_free(&env.obst, NULL);
2209 /* Handle graph state */
2211 set_irg_outs_inconsistent(irg);
2212 set_irg_entity_usage_state(irg, ir_entity_usage_not_computed);
2215 if (env.changes & CF_CHANGED) {
2216 /* is this really needed: Yes, control flow changed, block might
2217 have Bad() predecessors. */
2218 set_irg_doms_inconsistent(irg);
2220 } /* optimize_load_store */