2 * Copyright (C) 1995-2008 University of Karlsruhe. All right reserved.
4 * This file is part of libFirm.
6 * This file may be distributed and/or modified under the terms of the
7 * GNU General Public License version 2 as published by the Free Software
8 * Foundation and appearing in the file LICENSE.GPL included in the
9 * packaging of this file.
11 * Licensees holding valid libFirm Professional Edition licenses may use
12 * this file in accordance with the libFirm Commercial License.
13 * Agreement provided with the Software.
15 * This file is provided AS IS with NO WARRANTY OF ANY KIND, INCLUDING THE
16 * WARRANTY OF DESIGN, MERCHANTABILITY AND FITNESS FOR A PARTICULAR
22 * @brief Load/Store optimizations.
23 * @author Michael Beck
30 #include "iroptimize.h"
32 #include "irgraph_t.h"
40 #include "dbginfo_t.h"
41 #include "iropt_dbg.h"
47 #include "opt_polymorphy.h"
49 #include "irphase_t.h"
54 /** The debug handle. */
55 DEBUG_ONLY(static firm_dbg_module_t *dbg;)
58 #include "cacheopt/cachesim.h"
62 #define IMAX(a,b) ((a) > (b) ? (a) : (b))
64 #define MAX_PROJ IMAX(IMAX(pn_Load_max, pn_Store_max), pn_Call_max)
67 DF_CHANGED = 1, /**< data flow changed */
68 CF_CHANGED = 2, /**< control flow changed */
74 typedef struct _walk_env_t {
75 struct obstack obst; /**< list of all stores */
76 unsigned changes; /**< a bitmask of graph changes */
79 /** A Load/Store info. */
80 typedef struct _ldst_info_t {
81 ir_node *projs[MAX_PROJ]; /**< list of Proj's of this node */
82 ir_node *exc_block; /**< the exception block if available */
83 int exc_idx; /**< predecessor index in the exception block */
84 unsigned visited; /**< visited counter for breaking loops */
88 * flags for control flow.
91 BLOCK_HAS_COND = 1, /**< Block has conditional control flow */
92 BLOCK_HAS_EXC = 2 /**< Block has exceptional control flow */
98 typedef struct _block_info_t {
99 unsigned flags; /**< flags for the block */
102 /** the master visited flag for loop detection. */
103 static unsigned master_visited = 0;
105 #define INC_MASTER() ++master_visited
106 #define MARK_NODE(info) (info)->visited = master_visited
107 #define NODE_VISITED(info) (info)->visited >= master_visited
110 * get the Load/Store info of a node
112 static ldst_info_t *get_ldst_info(ir_node *node, struct obstack *obst) {
113 ldst_info_t *info = get_irn_link(node);
116 info = obstack_alloc(obst, sizeof(*info));
117 memset(info, 0, sizeof(*info));
118 set_irn_link(node, info);
121 } /* get_ldst_info */
124 * get the Block info of a node
126 static block_info_t *get_block_info(ir_node *node, struct obstack *obst) {
127 block_info_t *info = get_irn_link(node);
130 info = obstack_alloc(obst, sizeof(*info));
131 memset(info, 0, sizeof(*info));
132 set_irn_link(node, info);
135 } /* get_block_info */
138 * update the projection info for a Load/Store
140 static unsigned update_projs(ldst_info_t *info, ir_node *proj)
142 long nr = get_Proj_proj(proj);
144 assert(0 <= nr && nr <= MAX_PROJ && "Wrong proj from LoadStore");
146 if (info->projs[nr]) {
147 /* there is already one, do CSE */
148 exchange(proj, info->projs[nr]);
152 info->projs[nr] = proj;
158 * update the exception block info for a Load/Store node.
160 * @param info the load/store info struct
161 * @param block the exception handler block for this load/store
162 * @param pos the control flow input of the block
164 static unsigned update_exc(ldst_info_t *info, ir_node *block, int pos)
166 assert(info->exc_block == NULL && "more than one exception block found");
168 info->exc_block = block;
173 /** Return the number of uses of an address node */
174 #define get_irn_n_uses(adr) get_irn_n_edges(adr)
177 * walker, collects all Load/Store/Proj nodes
179 * walks from Start -> End
181 static void collect_nodes(ir_node *node, void *env)
183 ir_opcode opcode = get_irn_opcode(node);
184 ir_node *pred, *blk, *pred_blk;
185 ldst_info_t *ldst_info;
186 walk_env_t *wenv = env;
188 if (opcode == iro_Proj) {
189 pred = get_Proj_pred(node);
190 opcode = get_irn_opcode(pred);
192 if (opcode == iro_Load || opcode == iro_Store || opcode == iro_Call) {
193 ldst_info = get_ldst_info(pred, &wenv->obst);
195 wenv->changes |= update_projs(ldst_info, node);
198 * Place the Proj's to the same block as the
199 * predecessor Load. This is always ok and prevents
200 * "non-SSA" form after optimizations if the Proj
201 * is in a wrong block.
203 blk = get_nodes_block(node);
204 pred_blk = get_nodes_block(pred);
205 if (blk != pred_blk) {
206 wenv->changes |= DF_CHANGED;
207 set_nodes_block(node, pred_blk);
210 } else if (opcode == iro_Block) {
213 for (i = get_Block_n_cfgpreds(node) - 1; i >= 0; --i) {
214 ir_node *pred_block, *proj;
215 block_info_t *bl_info;
218 pred = proj = get_Block_cfgpred(node, i);
221 pred = get_Proj_pred(proj);
222 is_exc = get_Proj_proj(proj) == pn_Generic_X_except;
225 /* ignore Bad predecessors, they will be removed later */
229 pred_block = get_nodes_block(pred);
230 bl_info = get_block_info(pred_block, &wenv->obst);
232 if (is_fragile_op(pred) && is_exc)
233 bl_info->flags |= BLOCK_HAS_EXC;
234 else if (is_irn_forking(pred))
235 bl_info->flags |= BLOCK_HAS_COND;
237 opcode = get_irn_opcode(pred);
238 if (is_exc && (opcode == iro_Load || opcode == iro_Store || opcode == iro_Call)) {
239 ldst_info = get_ldst_info(pred, &wenv->obst);
241 wenv->changes |= update_exc(ldst_info, node, i);
245 } /* collect_nodes */
248 * Returns an entity if the address ptr points to a constant one.
250 * @param ptr the address
252 * @return an entity or NULL
254 static ir_entity *find_constant_entity(ir_node *ptr)
257 if (is_SymConst(ptr) && get_SymConst_kind(ptr) == symconst_addr_ent) {
258 return get_SymConst_entity(ptr);
259 } else if (is_Sel(ptr)) {
260 ir_entity *ent = get_Sel_entity(ptr);
261 ir_type *tp = get_entity_owner(ent);
263 /* Do not fiddle with polymorphism. */
264 if (is_Class_type(get_entity_owner(ent)) &&
265 ((get_entity_n_overwrites(ent) != 0) ||
266 (get_entity_n_overwrittenby(ent) != 0) ) )
269 if (is_Array_type(tp)) {
273 for (i = 0, n = get_Sel_n_indexs(ptr); i < n; ++i) {
275 tarval *tlower, *tupper;
276 ir_node *index = get_Sel_index(ptr, i);
277 tarval *tv = computed_value(index);
279 /* check if the index is constant */
280 if (tv == tarval_bad)
283 bound = get_array_lower_bound(tp, i);
284 tlower = computed_value(bound);
285 bound = get_array_upper_bound(tp, i);
286 tupper = computed_value(bound);
288 if (tlower == tarval_bad || tupper == tarval_bad)
291 if (tarval_cmp(tv, tlower) & pn_Cmp_Lt)
293 if (tarval_cmp(tupper, tv) & pn_Cmp_Lt)
296 /* ok, bounds check finished */
300 if (variability_constant == get_entity_variability(ent))
304 ptr = get_Sel_ptr(ptr);
305 } else if (is_Add(ptr)) {
306 ir_node *l = get_Add_left(ptr);
307 ir_node *r = get_Add_right(ptr);
309 if (get_irn_mode(l) == get_irn_mode(ptr) && is_Const(r))
311 else if (get_irn_mode(r) == get_irn_mode(ptr) && is_Const(l))
316 /* for now, we support only one addition, reassoc should fold all others */
317 if (! is_SymConst(ptr) && !is_Sel(ptr))
319 } else if (is_Sub(ptr)) {
320 ir_node *l = get_Sub_left(ptr);
321 ir_node *r = get_Sub_right(ptr);
323 if (get_irn_mode(l) == get_irn_mode(ptr) && is_Const(r))
327 /* for now, we support only one substraction, reassoc should fold all others */
328 if (! is_SymConst(ptr) && !is_Sel(ptr))
333 } /* find_constant_entity */
336 * Return the Selection index of a Sel node from dimension n
338 static long get_Sel_array_index_long(ir_node *n, int dim) {
339 ir_node *index = get_Sel_index(n, dim);
340 assert(is_Const(index));
341 return get_tarval_long(get_Const_tarval(index));
342 } /* get_Sel_array_index_long */
345 * Returns the accessed component graph path for an
346 * node computing an address.
348 * @param ptr the node computing the address
349 * @param depth current depth in steps upward from the root
352 static compound_graph_path *rec_get_accessed_path(ir_node *ptr, int depth) {
353 compound_graph_path *res = NULL;
354 ir_entity *root, *field, *ent;
355 int path_len, pos, idx;
359 if (is_SymConst(ptr)) {
360 /* a SymConst. If the depth is 0, this is an access to a global
361 * entity and we don't need a component path, else we know
362 * at least its length.
364 assert(get_SymConst_kind(ptr) == symconst_addr_ent);
365 root = get_SymConst_entity(ptr);
366 res = (depth == 0) ? NULL : new_compound_graph_path(get_entity_type(root), depth);
367 } else if (is_Sel(ptr)) {
368 /* it's a Sel, go up until we find the root */
369 res = rec_get_accessed_path(get_Sel_ptr(ptr), depth+1);
373 /* fill up the step in the path at the current position */
374 field = get_Sel_entity(ptr);
375 path_len = get_compound_graph_path_length(res);
376 pos = path_len - depth - 1;
377 set_compound_graph_path_node(res, pos, field);
379 if (is_Array_type(get_entity_owner(field))) {
380 assert(get_Sel_n_indexs(ptr) == 1 && "multi dim arrays not implemented");
381 set_compound_graph_path_array_index(res, pos, get_Sel_array_index_long(ptr, 0));
383 } else if (is_Add(ptr)) {
384 ir_node *l = get_Add_left(ptr);
385 ir_node *r = get_Add_right(ptr);
386 ir_mode *mode = get_irn_mode(ptr);
389 if (is_Const(r) && get_irn_mode(l) == mode) {
391 tv = get_Const_tarval(r);
394 tv = get_Const_tarval(l);
397 mode = get_tarval_mode(tv);
400 /* ptr must be a Sel or a SymConst, this was checked in find_constant_entity() */
402 field = get_Sel_entity(ptr);
404 field = get_SymConst_entity(ptr);
407 for (ent = field;;) {
409 tarval *sz, *tv_index, *tlower, *tupper;
412 tp = get_entity_type(ent);
413 if (! is_Array_type(tp))
415 ent = get_array_element_entity(tp);
416 size = get_type_size_bytes(get_entity_type(ent));
417 sz = new_tarval_from_long(size, mode);
419 tv_index = tarval_div(tmp, sz);
420 tmp = tarval_mod(tmp, sz);
422 if (tv_index == tarval_bad || tmp == tarval_bad)
425 assert(get_array_n_dimensions(tp) == 1 && "multiarrays not implemented");
426 bound = get_array_lower_bound(tp, 0);
427 tlower = computed_value(bound);
428 bound = get_array_upper_bound(tp, 0);
429 tupper = computed_value(bound);
431 if (tlower == tarval_bad || tupper == tarval_bad)
434 if (tarval_cmp(tv_index, tlower) & pn_Cmp_Lt)
436 if (tarval_cmp(tupper, tv_index) & pn_Cmp_Lt)
439 /* ok, bounds check finished */
442 if (! tarval_is_null(tmp)) {
443 /* access to some struct/union member */
447 /* should be at least ONE array */
451 res = rec_get_accessed_path(ptr, depth + idx);
455 path_len = get_compound_graph_path_length(res);
456 pos = path_len - depth - idx;
458 for (ent = field;;) {
460 tarval *sz, *tv_index;
463 tp = get_entity_type(ent);
464 if (! is_Array_type(tp))
466 ent = get_array_element_entity(tp);
467 set_compound_graph_path_node(res, pos, ent);
469 size = get_type_size_bytes(get_entity_type(ent));
470 sz = new_tarval_from_long(size, mode);
472 tv_index = tarval_div(tv, sz);
473 tv = tarval_mod(tv, sz);
475 /* worked above, should work again */
476 assert(tv_index != tarval_bad && tv != tarval_bad);
478 /* bounds already checked above */
479 index = get_tarval_long(tv_index);
480 set_compound_graph_path_array_index(res, pos, index);
483 } else if (is_Sub(ptr)) {
484 ir_node *l = get_Sub_left(ptr);
485 ir_node *r = get_Sub_right(ptr);
488 tv = get_Const_tarval(r);
493 } /* rec_get_accessed_path */
496 * Returns an access path or NULL. The access path is only
497 * valid, if the graph is in phase_high and _no_ address computation is used.
499 static compound_graph_path *get_accessed_path(ir_node *ptr) {
500 compound_graph_path *gr = rec_get_accessed_path(ptr, 0);
502 } /* get_accessed_path */
504 typedef struct path_entry {
506 struct path_entry *next;
510 static ir_node *rec_find_compound_ent_value(ir_node *ptr, path_entry *next) {
511 path_entry entry, *p;
512 ir_entity *ent, *field;
513 ir_initializer_t *initializer;
519 if (is_SymConst(ptr)) {
521 ent = get_SymConst_entity(ptr);
522 initializer = get_entity_initializer(ent);
523 for (p = next; p != NULL;) {
524 if (initializer->kind != IR_INITIALIZER_COMPOUND)
526 n = get_initializer_compound_n_entries(initializer);
527 tp = get_entity_type(ent);
529 if (is_Array_type(tp)) {
530 ent = get_array_element_entity(tp);
535 initializer = get_initializer_compound_value(initializer, 0);
539 if (p->index >= (int) n)
541 initializer = get_initializer_compound_value(initializer, p->index);
546 tp = get_entity_type(ent);
547 while (is_Array_type(tp)) {
548 ent = get_array_element_entity(tp);
549 tp = get_entity_type(ent);
551 n = get_initializer_compound_n_entries(initializer);
554 initializer = get_initializer_compound_value(initializer, 0);
557 switch (initializer->kind) {
558 case IR_INITIALIZER_CONST:
559 return get_initializer_const_value(initializer);
560 case IR_INITIALIZER_TARVAL:
561 case IR_INITIALIZER_NULL:
565 } else if (is_Sel(ptr)) {
566 entry.ent = field = get_Sel_entity(ptr);
567 tp = get_entity_owner(field);
568 if (is_Array_type(tp)) {
569 assert(get_Sel_n_indexs(ptr) == 1 && "multi dim arrays not implemented");
570 entry.index = get_Sel_array_index_long(ptr, 0) - get_array_lower_bound_int(tp, 0);
572 int i, n_members = get_compound_n_members(tp);
573 for (i = 0; i < n_members; ++i) {
574 if (get_compound_member(tp, i) == field)
577 if (i >= n_members) {
578 /* not found: should NOT happen */
583 return rec_find_compound_ent_value(get_Sel_ptr(ptr), &entry);
584 } else if (is_Add(ptr)) {
585 ir_node *l = get_Add_left(ptr);
586 ir_node *r = get_Add_right(ptr);
592 tv = get_Const_tarval(r);
595 tv = get_Const_tarval(l);
598 mode = get_tarval_mode(tv);
600 /* ptr must be a Sel or a SymConst, this was checked in find_constant_entity() */
602 field = get_Sel_entity(ptr);
604 field = get_SymConst_entity(ptr);
607 /* count needed entries */
609 for (ent = field;;) {
610 tp = get_entity_type(ent);
611 if (! is_Array_type(tp))
613 ent = get_array_element_entity(tp);
616 /* should be at least ONE entry */
620 /* allocate the right number of entries */
621 NEW_ARR_A(path_entry, p, pos);
625 for (ent = field;;) {
627 tarval *sz, *tv_index, *tlower, *tupper;
631 tp = get_entity_type(ent);
632 if (! is_Array_type(tp))
634 ent = get_array_element_entity(tp);
636 p[pos].next = &p[pos + 1];
638 size = get_type_size_bytes(get_entity_type(ent));
639 sz = new_tarval_from_long(size, mode);
641 tv_index = tarval_div(tv, sz);
642 tv = tarval_mod(tv, sz);
644 if (tv_index == tarval_bad || tv == tarval_bad)
647 assert(get_array_n_dimensions(tp) == 1 && "multiarrays not implemented");
648 bound = get_array_lower_bound(tp, 0);
649 tlower = computed_value(bound);
650 bound = get_array_upper_bound(tp, 0);
651 tupper = computed_value(bound);
653 if (tlower == tarval_bad || tupper == tarval_bad)
656 if (tarval_cmp(tv_index, tlower) & pn_Cmp_Lt)
658 if (tarval_cmp(tupper, tv_index) & pn_Cmp_Lt)
661 /* ok, bounds check finished */
662 index = get_tarval_long(tv_index);
663 p[pos].index = index;
666 if (! tarval_is_null(tv)) {
667 /* hmm, wrong access */
670 p[pos - 1].next = next;
671 return rec_find_compound_ent_value(ptr, p);
672 } else if (is_Sub(ptr)) {
673 ir_node *l = get_Sub_left(ptr);
674 ir_node *r = get_Sub_right(ptr);
677 tv = get_Const_tarval(r);
684 static ir_node *find_compound_ent_value(ir_node *ptr) {
685 return rec_find_compound_ent_value(ptr, NULL);
689 static void reduce_adr_usage(ir_node *ptr);
692 * Update a Load that may have lost its users.
694 static void handle_load_update(ir_node *load) {
695 ldst_info_t *info = get_irn_link(load);
697 /* do NOT touch volatile loads for now */
698 if (get_Load_volatility(load) == volatility_is_volatile)
701 if (! info->projs[pn_Load_res] && ! info->projs[pn_Load_X_except]) {
702 ir_node *ptr = get_Load_ptr(load);
703 ir_node *mem = get_Load_mem(load);
705 /* a Load whose value is neither used nor exception checked, remove it */
706 exchange(info->projs[pn_Load_M], mem);
707 if (info->projs[pn_Load_X_regular])
708 exchange(info->projs[pn_Load_X_regular], new_r_Jmp(current_ir_graph, get_nodes_block(load)));
710 reduce_adr_usage(ptr);
712 } /* handle_load_update */
715 * A use of an address node has vanished. Check if this was a Proj
716 * node and update the counters.
718 static void reduce_adr_usage(ir_node *ptr) {
720 if (get_irn_n_edges(ptr) <= 0) {
721 /* this Proj is dead now */
722 ir_node *pred = get_Proj_pred(ptr);
725 ldst_info_t *info = get_irn_link(pred);
726 info->projs[get_Proj_proj(ptr)] = NULL;
728 /* this node lost its result proj, handle that */
729 handle_load_update(pred);
733 } /* reduce_adr_usage */
736 * Check, if an already existing value of mode old_mode can be converted
737 * into the needed one new_mode without loss.
739 static int can_use_stored_value(ir_mode *old_mode, ir_mode *new_mode) {
740 if (old_mode == new_mode)
743 /* if both modes are two-complement ones, we can always convert the
744 Stored value into the needed one. */
745 if (get_mode_size_bits(old_mode) >= get_mode_size_bits(new_mode) &&
746 get_mode_arithmetic(old_mode) == irma_twos_complement &&
747 get_mode_arithmetic(new_mode) == irma_twos_complement)
750 } /* can_use_stored_value */
753 * Check whether a Call is at least pure, ie. does only read memory.
755 static unsigned is_Call_pure(ir_node *call) {
756 ir_type *call_tp = get_Call_type(call);
757 unsigned prop = get_method_additional_properties(call_tp);
759 /* check first the call type */
760 if ((prop & (mtp_property_const|mtp_property_pure)) == 0) {
761 /* try the called entity */
762 ir_node *ptr = get_Call_ptr(call);
764 if (is_Global(ptr)) {
765 ir_entity *ent = get_Global_entity(ptr);
767 prop = get_entity_additional_properties(ent);
770 return (prop & (mtp_property_const|mtp_property_pure)) != 0;
773 static ir_node *get_base_and_offset(ir_node *ptr, long *pOffset)
775 ir_mode *mode = get_irn_mode(ptr);
778 /* TODO: long might not be enough, we should probably use some tarval thingy... */
781 ir_node *l = get_Add_left(ptr);
782 ir_node *r = get_Add_right(ptr);
784 if (get_irn_mode(l) != mode || !is_Const(r))
787 offset += get_tarval_long(get_Const_tarval(r));
789 } else if (is_Sub(ptr)) {
790 ir_node *l = get_Sub_left(ptr);
791 ir_node *r = get_Sub_right(ptr);
793 if (get_irn_mode(l) != mode || !is_Const(r))
796 offset -= get_tarval_long(get_Const_tarval(r));
798 } else if (is_Sel(ptr)) {
799 ir_entity *ent = get_Sel_entity(ptr);
800 ir_type *tp = get_entity_owner(ent);
802 if (is_Array_type(tp)) {
806 /* only one dimensional arrays yet */
807 if (get_Sel_n_indexs(ptr) != 1)
809 index = get_Sel_index(ptr, 0);
810 if (! is_Const(index))
813 tp = get_entity_type(ent);
814 if (get_type_state(tp) != layout_fixed)
817 size = get_type_size_bytes(tp);
818 offset += size * get_tarval_long(get_Const_tarval(index));
820 if (get_type_state(tp) != layout_fixed)
822 offset += get_entity_offset(ent);
824 ptr = get_Sel_ptr(ptr);
833 static int try_load_after_store(ir_node *load,
834 ir_node *load_base_ptr, long load_offset, ir_node *store)
837 ir_node *store_ptr = get_Store_ptr(store);
839 ir_node *store_base_ptr = get_base_and_offset(store_ptr, &store_offset);
840 ir_node *store_value;
849 if (load_base_ptr != store_base_ptr)
852 load_mode = get_Load_mode(load);
853 load_mode_len = get_mode_size_bytes(load_mode);
854 store_mode = get_irn_mode(get_Store_value(store));
855 store_mode_len = get_mode_size_bytes(store_mode);
856 delta = load_offset - store_offset;
857 store_value = get_Store_value(store);
859 if (delta != 0 || store_mode != load_mode) {
860 if (delta < 0 || delta + load_mode_len > store_mode_len)
863 if (get_mode_arithmetic(store_mode) != irma_twos_complement ||
864 get_mode_arithmetic(load_mode) != irma_twos_complement)
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);
885 DBG_OPT_RAW(load, store_value);
887 info = get_irn_link(load);
888 if (info->projs[pn_Load_M])
889 exchange(info->projs[pn_Load_M], get_Load_mem(load));
893 if (info->projs[pn_Load_X_except]) {
894 exchange( info->projs[pn_Load_X_except], new_Bad());
897 if (info->projs[pn_Load_X_regular]) {
898 exchange( info->projs[pn_Load_X_regular], new_r_Jmp(current_ir_graph, get_nodes_block(load)));
902 if (info->projs[pn_Load_res])
903 exchange(info->projs[pn_Load_res], store_value);
905 load_ptr = get_Load_ptr(load);
907 reduce_adr_usage(load_ptr);
908 return res | DF_CHANGED;
912 * Follow the memory chain as long as there are only Loads,
913 * alias free Stores, and constant Calls and try to replace the
914 * current Load by a previous ones.
915 * Note that in unreachable loops it might happen that we reach
916 * load again, as well as we can fall into a cycle.
917 * We break such cycles using a special visited flag.
919 * INC_MASTER() must be called before dive into
921 static unsigned follow_Mem_chain(ir_node *load, ir_node *curr) {
923 ldst_info_t *info = get_irn_link(load);
925 ir_node *ptr = get_Load_ptr(load);
926 ir_node *mem = get_Load_mem(load);
927 ir_mode *load_mode = get_Load_mode(load);
929 for (pred = curr; load != pred; ) {
930 ldst_info_t *pred_info = get_irn_link(pred);
933 * a Load immediately after a Store -- a read after write.
934 * We may remove the Load, if both Load & Store does not have an
935 * exception handler OR they are in the same MacroBlock. In the latter
936 * case the Load cannot throw an exception when the previous Store was
939 * Why we need to check for Store Exception? If the Store cannot
940 * be executed (ROM) the exception handler might simply jump into
941 * the load MacroBlock :-(
942 * We could make it a little bit better if we would know that the
943 * exception handler of the Store jumps directly to the end...
945 if (is_Store(pred) && ((pred_info->projs[pn_Store_X_except] == NULL
946 && info->projs[pn_Load_X_except] == NULL)
947 || get_nodes_MacroBlock(load) == get_nodes_MacroBlock(pred)))
950 ir_node *base_ptr = get_base_and_offset(ptr, &load_offset);
951 int changes = try_load_after_store(load, base_ptr, load_offset, pred);
954 return res | changes;
955 } else if (is_Load(pred) && get_Load_ptr(pred) == ptr &&
956 can_use_stored_value(get_Load_mode(pred), load_mode)) {
958 * a Load after a Load -- a read after read.
959 * We may remove the second Load, if it does not have an exception handler
960 * OR they are in the same MacroBlock. In the later case the Load cannot
961 * throw an exception when the previous Load was quiet.
963 * Here, there is no need to check if the previous Load has an exception
964 * hander because they would have exact the same exception...
966 if (info->projs[pn_Load_X_except] == NULL || get_nodes_MacroBlock(load) == get_nodes_MacroBlock(pred)) {
969 DBG_OPT_RAR(load, pred);
971 /* the result is used */
972 if (info->projs[pn_Load_res]) {
973 if (pred_info->projs[pn_Load_res] == NULL) {
974 /* create a new Proj again */
975 pred_info->projs[pn_Load_res] = new_r_Proj(current_ir_graph, get_nodes_block(pred), pred, get_Load_mode(pred), pn_Load_res);
977 value = pred_info->projs[pn_Load_res];
979 /* add an convert if needed */
980 if (get_Load_mode(pred) != load_mode) {
981 value = new_r_Conv(current_ir_graph, get_nodes_block(load), value, load_mode);
984 exchange(info->projs[pn_Load_res], value);
987 if (info->projs[pn_Load_M])
988 exchange(info->projs[pn_Load_M], mem);
991 if (info->projs[pn_Load_X_except]) {
992 exchange(info->projs[pn_Load_X_except], new_Bad());
995 if (info->projs[pn_Load_X_regular]) {
996 exchange( info->projs[pn_Load_X_regular], new_r_Jmp(current_ir_graph, get_nodes_block(load)));
1001 reduce_adr_usage(ptr);
1002 return res |= DF_CHANGED;
1006 if (is_Store(pred)) {
1007 /* check if we can pass through this store */
1008 ir_alias_relation rel = get_alias_relation(
1010 get_Store_ptr(pred),
1011 get_irn_mode(get_Store_value(pred)),
1013 /* if the might be an alias, we cannot pass this Store */
1014 if (rel != ir_no_alias)
1016 pred = skip_Proj(get_Store_mem(pred));
1017 } else if (is_Load(pred)) {
1018 pred = skip_Proj(get_Load_mem(pred));
1019 } else if (is_Call(pred)) {
1020 if (is_Call_pure(pred)) {
1021 /* The called graph is at least pure, so there are no Store's
1022 in it. We can handle it like a Load and skip it. */
1023 pred = skip_Proj(get_Call_mem(pred));
1025 /* there might be Store's in the graph, stop here */
1029 /* follow only Load chains */
1033 /* check for cycles */
1034 if (NODE_VISITED(pred_info))
1036 MARK_NODE(pred_info);
1039 if (is_Sync(pred)) {
1042 /* handle all Sync predecessors */
1043 for (i = get_Sync_n_preds(pred) - 1; i >= 0; --i) {
1044 res |= follow_Mem_chain(load, skip_Proj(get_Sync_pred(pred, i)));
1051 } /* follow_Mem_chain */
1054 * Check if we can replace the load by a given const from
1055 * the const code irg.
1057 ir_node *can_replace_load_by_const(const ir_node *load, ir_node *c) {
1058 ir_mode *c_mode = get_irn_mode(c);
1059 ir_mode *l_mode = get_Load_mode(load);
1060 ir_node *res = NULL;
1062 if (c_mode != l_mode) {
1063 /* check, if the mode matches OR can be easily converted info */
1064 if (is_reinterpret_cast(c_mode, l_mode)) {
1065 /* we can safely cast */
1066 dbg_info *dbg = get_irn_dbg_info(load);
1067 ir_node *block = get_nodes_block(load);
1069 /* copy the value from the const code irg and cast it */
1070 res = copy_const_value(dbg, c);
1071 res = new_rd_Conv(dbg, current_ir_graph, block, res, l_mode);
1074 /* copy the value from the const code irg */
1075 res = copy_const_value(get_irn_dbg_info(load), c);
1078 } /* can_replace_load_by_const */
1083 * @param load the Load node
1085 static unsigned optimize_load(ir_node *load)
1087 ldst_info_t *info = get_irn_link(load);
1088 ir_node *mem, *ptr, *value;
1093 /* do NOT touch volatile loads for now */
1094 if (get_Load_volatility(load) == volatility_is_volatile)
1097 /* the address of the load to be optimized */
1098 ptr = get_Load_ptr(load);
1101 * Check if we can remove the exception from a Load:
1102 * This can be done, if the address is from an Sel(Alloc) and
1103 * the Sel type is a subtype of the allocated type.
1105 * This optimizes some often used OO constructs,
1106 * like x = new O; x->t;
1108 if (info->projs[pn_Load_X_except]) {
1109 ir_node *addr = ptr;
1111 /* find base address */
1112 while (is_Sel(addr))
1113 addr = get_Sel_ptr(addr);
1114 if (is_Alloc(skip_Proj(skip_Cast(addr)))) {
1115 /* simple case: a direct load after an Alloc. Firm Alloc throw
1116 * an exception in case of out-of-memory. So, there is no way for an
1117 * exception in this load.
1118 * This code is constructed by the "exception lowering" in the Jack compiler.
1120 exchange(info->projs[pn_Load_X_except], new_Bad());
1121 info->projs[pn_Load_X_except] = NULL;
1122 exchange(info->projs[pn_Load_X_regular], new_r_Jmp(current_ir_graph, get_nodes_block(load)));
1123 info->projs[pn_Load_X_regular] = NULL;
1128 /* The mem of the Load. Must still be returned after optimization. */
1129 mem = get_Load_mem(load);
1131 if (! info->projs[pn_Load_res] && ! info->projs[pn_Load_X_except]) {
1132 /* a Load which value is neither used nor exception checked, remove it */
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(current_ir_graph, 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 allocation_static == get_entity_allocation(ent) &&
1154 visibility_external_allocated != get_entity_visibility(ent)) {
1155 /* a static allocation that is not external: there should be NO exception
1156 * 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 exchange(info->projs[pn_Load_X_except], new_Bad());
1161 info->projs[pn_Load_X_except] = NULL;
1164 if (info->projs[pn_Load_X_regular]) {
1165 exchange(info->projs[pn_Load_X_regular], new_r_Jmp(current_ir_graph, get_nodes_block(load)));
1166 info->projs[pn_Load_X_regular] = NULL;
1170 if (variability_constant == get_entity_variability(ent)) {
1171 if (is_atomic_entity(ent)) {
1172 /* Might not be atomic after lowering of Sels. In this case we
1173 * could also load, but it's more complicated. */
1174 /* more simpler case: we load the content of a constant value:
1175 * replace it by the constant itself */
1176 value = get_atomic_ent_value(ent);
1177 } else if (ent->has_initializer) {
1178 /* new style initializer */
1179 value = find_compound_ent_value(ptr);
1181 /* old style initializer */
1182 compound_graph_path *path = get_accessed_path(ptr);
1185 assert(is_proper_compound_graph_path(path, get_compound_graph_path_length(path)-1));
1187 value = get_compound_ent_value_by_path(ent, path);
1188 DB((dbg, LEVEL_1, " Constant access at %F%F resulted in %+F\n", ent, path, value));
1189 free_compound_graph_path(path);
1193 value = can_replace_load_by_const(load, value);
1197 if (value != NULL) {
1198 /* we completely replace the load by this value */
1199 if (info->projs[pn_Load_X_except]) {
1200 exchange(info->projs[pn_Load_X_except], new_Bad());
1201 info->projs[pn_Load_X_except] = NULL;
1204 if (info->projs[pn_Load_X_regular]) {
1205 exchange(info->projs[pn_Load_X_regular], new_r_Jmp(current_ir_graph, get_nodes_block(load)));
1206 info->projs[pn_Load_X_regular] = NULL;
1209 if (info->projs[pn_Load_M]) {
1210 exchange(info->projs[pn_Load_M], mem);
1213 if (info->projs[pn_Load_res]) {
1214 exchange(info->projs[pn_Load_res], value);
1218 reduce_adr_usage(ptr);
1222 /* Check, if the address of this load is used more than once.
1223 * If not, more load cannot be removed in any case. */
1224 if (get_irn_n_uses(ptr) <= 1 && get_irn_n_uses(get_base_and_offset(ptr, &dummy)) <= 1)
1228 * follow the memory chain as long as there are only Loads
1229 * and try to replace current Load or Store by a previous one.
1230 * Note that in unreachable loops it might happen that we reach
1231 * load again, as well as we can fall into a cycle.
1232 * We break such cycles using a special visited flag.
1235 res = follow_Mem_chain(load, skip_Proj(mem));
1237 } /* optimize_load */
1240 * Check whether a value of mode new_mode would completely overwrite a value
1241 * of mode old_mode in memory.
1243 static int is_completely_overwritten(ir_mode *old_mode, ir_mode *new_mode)
1245 return get_mode_size_bits(new_mode) >= get_mode_size_bits(old_mode);
1246 } /* is_completely_overwritten */
1249 * Check whether small is a part of large (starting at same address).
1251 static int is_partially_same(ir_node *small, ir_node *large)
1253 ir_mode *sm = get_irn_mode(small);
1254 ir_mode *lm = get_irn_mode(large);
1256 /* FIXME: Check endianness */
1257 return is_Conv(small) && get_Conv_op(small) == large
1258 && get_mode_size_bytes(sm) < get_mode_size_bytes(lm)
1259 && get_mode_arithmetic(sm) == irma_twos_complement
1260 && get_mode_arithmetic(lm) == irma_twos_complement;
1261 } /* is_partially_same */
1264 * follow the memory chain as long as there are only Loads and alias free Stores.
1266 * INC_MASTER() must be called before dive into
1268 static unsigned follow_Mem_chain_for_Store(ir_node *store, ir_node *curr) {
1270 ldst_info_t *info = get_irn_link(store);
1272 ir_node *ptr = get_Store_ptr(store);
1273 ir_node *mem = get_Store_mem(store);
1274 ir_node *value = get_Store_value(store);
1275 ir_mode *mode = get_irn_mode(value);
1276 ir_node *block = get_nodes_block(store);
1277 ir_node *mblk = get_Block_MacroBlock(block);
1279 for (pred = curr; pred != store;) {
1280 ldst_info_t *pred_info = get_irn_link(pred);
1283 * BEWARE: one might think that checking the modes is useless, because
1284 * if the pointers are identical, they refer to the same object.
1285 * This is only true in strong typed languages, not is C were the following
1286 * is possible *(ir_type1 *)p = a; *(ir_type2 *)p = b ...
1287 * However, if the size of the mode that is written is bigger or equal the
1288 * size of the old one, the old value is completely overwritten and can be
1291 if (is_Store(pred) && get_Store_ptr(pred) == ptr &&
1292 get_nodes_MacroBlock(pred) == mblk) {
1294 * a Store after a Store in the same MacroBlock -- a write after write.
1298 * We may remove the first Store, if the old value is completely
1299 * overwritten or the old value is a part of the new value,
1300 * and if it does not have an exception handler.
1302 * TODO: What, if both have the same exception handler ???
1304 if (get_Store_volatility(pred) != volatility_is_volatile
1305 && !pred_info->projs[pn_Store_X_except]) {
1306 ir_node *predvalue = get_Store_value(pred);
1307 ir_mode *predmode = get_irn_mode(predvalue);
1309 if(is_completely_overwritten(predmode, mode)
1310 || is_partially_same(predvalue, value)) {
1311 DBG_OPT_WAW(pred, store);
1312 exchange(pred_info->projs[pn_Store_M], get_Store_mem(pred));
1314 reduce_adr_usage(ptr);
1320 * We may remove the Store, if the old value already contains
1321 * the new value, and if it does not have an exception handler.
1323 * TODO: What, if both have the same exception handler ???
1325 if (get_Store_volatility(store) != volatility_is_volatile
1326 && !info->projs[pn_Store_X_except]) {
1327 ir_node *predvalue = get_Store_value(pred);
1329 if(is_partially_same(value, predvalue)) {
1330 DBG_OPT_WAW(pred, store);
1331 exchange(info->projs[pn_Store_M], mem);
1333 reduce_adr_usage(ptr);
1337 } else if (is_Load(pred) && get_Load_ptr(pred) == ptr &&
1338 value == pred_info->projs[pn_Load_res]) {
1340 * a Store of a value just loaded from the same address
1341 * -- a write after read.
1342 * We may remove the Store, if it does not have an exception
1345 if (! info->projs[pn_Store_X_except]) {
1346 DBG_OPT_WAR(store, pred);
1347 exchange(info->projs[pn_Store_M], mem);
1349 reduce_adr_usage(ptr);
1354 if (is_Store(pred)) {
1355 /* check if we can pass through this store */
1356 ir_alias_relation rel = get_alias_relation(
1358 get_Store_ptr(pred),
1359 get_irn_mode(get_Store_value(pred)),
1361 /* if the might be an alias, we cannot pass this Store */
1362 if (rel != ir_no_alias)
1364 pred = skip_Proj(get_Store_mem(pred));
1365 } else if (is_Load(pred)) {
1366 ir_alias_relation rel = get_alias_relation(
1367 current_ir_graph, get_Load_ptr(pred), get_Load_mode(pred),
1369 if (rel != ir_no_alias)
1372 pred = skip_Proj(get_Load_mem(pred));
1374 /* follow only Load chains */
1378 /* check for cycles */
1379 if (NODE_VISITED(pred_info))
1381 MARK_NODE(pred_info);
1384 if (is_Sync(pred)) {
1387 /* handle all Sync predecessors */
1388 for (i = get_Sync_n_preds(pred) - 1; i >= 0; --i) {
1389 res |= follow_Mem_chain_for_Store(store, skip_Proj(get_Sync_pred(pred, i)));
1395 } /* follow_Mem_chain_for_Store */
1397 /** find entity used as base for an address calculation */
1398 static ir_entity *find_entity(ir_node *ptr)
1400 switch(get_irn_opcode(ptr)) {
1402 return get_SymConst_entity(ptr);
1404 ir_node *pred = get_Sel_ptr(ptr);
1405 if (get_irg_frame(get_irn_irg(ptr)) == pred)
1406 return get_Sel_entity(ptr);
1408 return find_entity(pred);
1412 ir_node *left = get_binop_left(ptr);
1414 if (mode_is_reference(get_irn_mode(left)))
1415 return find_entity(left);
1416 right = get_binop_right(ptr);
1417 if (mode_is_reference(get_irn_mode(right)))
1418 return find_entity(right);
1429 * @param store the Store node
1431 static unsigned optimize_store(ir_node *store) {
1436 if (get_Store_volatility(store) == volatility_is_volatile)
1439 ptr = get_Store_ptr(store);
1440 entity = find_entity(ptr);
1442 /* a store to an entity which is never read is unnecessary */
1443 if (entity != NULL && !(get_entity_usage(entity) & ir_usage_read)) {
1444 ldst_info_t *info = get_irn_link(store);
1445 if (info->projs[pn_Store_X_except] == NULL) {
1446 exchange(info->projs[pn_Store_M], get_Store_mem(store));
1448 reduce_adr_usage(ptr);
1453 /* Check, if the address of this Store is used more than once.
1454 * If not, this Store cannot be removed in any case. */
1455 if (get_irn_n_uses(ptr) <= 1)
1458 mem = get_Store_mem(store);
1460 /* follow the memory chain as long as there are only Loads */
1463 return follow_Mem_chain_for_Store(store, skip_Proj(mem));
1464 } /* optimize_store */
1467 * walker, optimizes Phi after Stores to identical places:
1468 * Does the following optimization:
1471 * val1 val2 val3 val1 val2 val3
1473 * Store Store Store \ | /
1480 * This reduces the number of stores and allows for predicated execution.
1481 * Moves Stores back to the end of a function which may be bad.
1483 * This is only possible if the predecessor blocks have only one successor.
1485 static unsigned optimize_phi(ir_node *phi, walk_env_t *wenv)
1488 ir_node *store, *old_store, *ptr, *block, *phi_block, *phiM, *phiD, *exc, *projM;
1490 ir_node **inM, **inD, **projMs;
1492 dbg_info *db = NULL;
1494 block_info_t *bl_info;
1497 /* Must be a memory Phi */
1498 if (get_irn_mode(phi) != mode_M)
1501 n = get_Phi_n_preds(phi);
1505 /* must be only one user */
1506 projM = get_Phi_pred(phi, 0);
1507 if (get_irn_n_edges(projM) != 1)
1510 store = skip_Proj(projM);
1512 if (!is_Store(store))
1515 block = get_nodes_block(store);
1517 /* abort on dead blocks */
1518 if (is_Block_dead(block))
1521 /* check if the block is post dominated by Phi-block
1522 and has no exception exit */
1523 bl_info = get_irn_link(block);
1524 if (bl_info->flags & BLOCK_HAS_EXC)
1527 phi_block = get_nodes_block(phi);
1528 if (! block_strictly_postdominates(phi_block, block))
1531 /* this is the address of the store */
1532 ptr = get_Store_ptr(store);
1533 mode = get_irn_mode(get_Store_value(store));
1534 info = get_irn_link(store);
1535 exc = info->exc_block;
1537 for (i = 1; i < n; ++i) {
1538 ir_node *pred = get_Phi_pred(phi, i);
1540 if (get_irn_n_edges(pred) != 1)
1543 pred = skip_Proj(pred);
1544 if (!is_Store(pred))
1547 if (ptr != get_Store_ptr(pred) || mode != get_irn_mode(get_Store_value(pred)))
1550 info = get_irn_link(pred);
1552 /* check, if all stores have the same exception flow */
1553 if (exc != info->exc_block)
1556 /* abort on dead blocks */
1557 block = get_nodes_block(pred);
1558 if (is_Block_dead(block))
1561 /* check if the block is post dominated by Phi-block
1562 and has no exception exit. Note that block must be different from
1563 Phi-block, else we would move a Store from end End of a block to its
1565 bl_info = get_irn_link(block);
1566 if (bl_info->flags & BLOCK_HAS_EXC)
1568 if (block == phi_block || ! block_postdominates(phi_block, block))
1573 * ok, when we are here, we found all predecessors of a Phi that
1574 * are Stores to the same address and size. That means whatever
1575 * we do before we enter the block of the Phi, we do a Store.
1576 * So, we can move the Store to the current block:
1578 * val1 val2 val3 val1 val2 val3
1580 * | Str | | Str | | Str | \ | /
1586 * Is only allowed if the predecessor blocks have only one successor.
1589 NEW_ARR_A(ir_node *, projMs, n);
1590 NEW_ARR_A(ir_node *, inM, n);
1591 NEW_ARR_A(ir_node *, inD, n);
1592 NEW_ARR_A(int, idx, n);
1594 /* Prepare: Collect all Store nodes. We must do this
1595 first because we otherwise may loose a store when exchanging its
1598 for (i = n - 1; i >= 0; --i) {
1601 projMs[i] = get_Phi_pred(phi, i);
1602 assert(is_Proj(projMs[i]));
1604 store = get_Proj_pred(projMs[i]);
1605 info = get_irn_link(store);
1607 inM[i] = get_Store_mem(store);
1608 inD[i] = get_Store_value(store);
1609 idx[i] = info->exc_idx;
1611 block = get_nodes_block(phi);
1613 /* second step: create a new memory Phi */
1614 phiM = new_rd_Phi(get_irn_dbg_info(phi), current_ir_graph, block, n, inM, mode_M);
1616 /* third step: create a new data Phi */
1617 phiD = new_rd_Phi(get_irn_dbg_info(phi), current_ir_graph, block, n, inD, mode);
1619 /* rewire memory and kill the node */
1620 for (i = n - 1; i >= 0; --i) {
1621 ir_node *proj = projMs[i];
1624 ir_node *store = get_Proj_pred(proj);
1625 exchange(proj, inM[i]);
1630 /* fourth step: create the Store */
1631 store = new_rd_Store(db, current_ir_graph, block, phiM, ptr, phiD, 0);
1633 co_set_irn_name(store, co_get_irn_ident(old_store));
1636 projM = new_rd_Proj(NULL, current_ir_graph, block, store, mode_M, pn_Store_M);
1638 info = get_ldst_info(store, &wenv->obst);
1639 info->projs[pn_Store_M] = projM;
1641 /* fifths step: repair exception flow */
1643 ir_node *projX = new_rd_Proj(NULL, current_ir_graph, block, store, mode_X, pn_Store_X_except);
1645 info->projs[pn_Store_X_except] = projX;
1646 info->exc_block = exc;
1647 info->exc_idx = idx[0];
1649 for (i = 0; i < n; ++i) {
1650 set_Block_cfgpred(exc, idx[i], projX);
1654 /* the exception block should be optimized as some inputs are identical now */
1660 /* sixth step: replace old Phi */
1661 exchange(phi, projM);
1663 return res | DF_CHANGED;
1664 } /* optimize_phi */
1667 * walker, do the optimizations
1669 static void do_load_store_optimize(ir_node *n, void *env) {
1670 walk_env_t *wenv = env;
1672 switch (get_irn_opcode(n)) {
1675 wenv->changes |= optimize_load(n);
1679 wenv->changes |= optimize_store(n);
1683 wenv->changes |= optimize_phi(n, wenv);
1689 } /* do_load_store_optimize */
1692 typedef struct scc {
1693 ir_node *head; /**< the head of the list */
1696 /** A node entry. */
1697 typedef struct node_entry {
1698 unsigned DFSnum; /**< the DFS number of this node */
1699 unsigned low; /**< the low number of this node */
1700 int in_stack; /**< flag, set if the node is on the stack */
1701 ir_node *next; /**< link to the next node the the same scc */
1702 scc *pscc; /**< the scc of this node */
1703 unsigned POnum; /**< the post order number for blocks */
1706 /** A loop entry. */
1707 typedef struct loop_env {
1708 ir_phase ph; /**< the phase object */
1709 ir_node **stack; /**< the node stack */
1710 int tos; /**< tos index */
1711 unsigned nextDFSnum; /**< the current DFS number */
1712 unsigned POnum; /**< current post order number */
1714 unsigned changes; /**< a bitmask of graph changes */
1718 * Gets the node_entry of a node
1720 static node_entry *get_irn_ne(ir_node *irn, loop_env *env) {
1721 ir_phase *ph = &env->ph;
1722 node_entry *e = phase_get_irn_data(&env->ph, irn);
1725 e = phase_alloc(ph, sizeof(*e));
1726 memset(e, 0, sizeof(*e));
1727 phase_set_irn_data(ph, irn, e);
1733 * Push a node onto the stack.
1735 * @param env the loop environment
1736 * @param n the node to push
1738 static void push(loop_env *env, ir_node *n) {
1741 if (env->tos == ARR_LEN(env->stack)) {
1742 int nlen = ARR_LEN(env->stack) * 2;
1743 ARR_RESIZE(ir_node *, env->stack, nlen);
1745 env->stack[env->tos++] = n;
1746 e = get_irn_ne(n, env);
1751 * pop a node from the stack
1753 * @param env the loop environment
1755 * @return The topmost node
1757 static ir_node *pop(loop_env *env) {
1758 ir_node *n = env->stack[--env->tos];
1759 node_entry *e = get_irn_ne(n, env);
1766 * Check if irn is a region constant.
1767 * The block or irn must strictly dominate the header block.
1769 * @param irn the node to check
1770 * @param header_block the header block of the induction variable
1772 static int is_rc(ir_node *irn, ir_node *header_block) {
1773 ir_node *block = get_nodes_block(irn);
1775 return (block != header_block) && block_dominates(block, header_block);
1778 typedef struct phi_entry phi_entry;
1780 ir_node *phi; /**< A phi with a region const memory. */
1781 int pos; /**< The position of the region const memory */
1782 ir_node *load; /**< the newly created load for this phi */
1787 * An entry in the avail set.
1789 typedef struct avail_entry_t {
1790 ir_node *ptr; /**< the address pointer */
1791 ir_mode *mode; /**< the load mode */
1792 ir_node *load; /**< the associated Load */
1796 * Compare two avail entries.
1798 static int cmp_avail_entry(const void *elt, const void *key, size_t size) {
1799 const avail_entry_t *a = elt;
1800 const avail_entry_t *b = key;
1802 return a->ptr != b->ptr || a->mode != b->mode;
1803 } /* cmp_avail_entry */
1806 * Calculate the hash value of an avail entry.
1808 static unsigned hash_cache_entry(const avail_entry_t *entry) {
1809 return get_irn_idx(entry->ptr) * 9 + HASH_PTR(entry->mode);
1810 } /* hash_cache_entry */
1813 * Move loops out of loops if possible.
1815 * @param pscc the loop described by an SCC
1816 * @param env the loop environment
1818 static void move_loads_out_of_loops(scc *pscc, loop_env *env) {
1819 ir_node *phi, *load, *next, *other, *next_other;
1822 phi_entry *phi_list = NULL;
1825 avail = new_set(cmp_avail_entry, 8);
1827 /* collect all outer memories */
1828 for (phi = pscc->head; phi != NULL; phi = next) {
1829 node_entry *ne = get_irn_ne(phi, env);
1832 /* check all memory Phi's */
1836 assert(get_irn_mode(phi) == mode_M && "DFS return non-memory Phi");
1838 for (j = get_irn_arity(phi) - 1; j >= 0; --j) {
1839 ir_node *pred = get_irn_n(phi, j);
1840 node_entry *pe = get_irn_ne(pred, env);
1842 if (pe->pscc != ne->pscc) {
1843 /* not in the same SCC, is region const */
1844 phi_entry *pe = phase_alloc(&env->ph, sizeof(*pe));
1848 pe->next = phi_list;
1853 /* no Phis no fun */
1854 assert(phi_list != NULL && "DFS found a loop without Phi");
1856 /* for now, we cannot handle more than one input (only reducible cf) */
1857 if (phi_list->next != NULL)
1860 for (load = pscc->head; load; load = next) {
1862 node_entry *ne = get_irn_ne(load, env);
1865 if (is_Load(load)) {
1866 ldst_info_t *info = get_irn_link(load);
1867 ir_node *ptr = get_Load_ptr(load);
1869 /* for now, we cannot handle Loads with exceptions */
1870 if (info->projs[pn_Load_res] == NULL || info->projs[pn_Load_X_regular] != NULL || info->projs[pn_Load_X_except] != NULL)
1873 /* for now, we can only move Load(Global) */
1874 if (! is_Global(ptr))
1876 ent = get_Global_entity(ptr);
1877 load_mode = get_Load_mode(load);
1878 for (other = pscc->head; other != NULL; other = next_other) {
1879 node_entry *ne = get_irn_ne(other, env);
1880 next_other = ne->next;
1882 if (is_Store(other)) {
1883 ir_alias_relation rel = get_alias_relation(
1885 get_Store_ptr(other),
1886 get_irn_mode(get_Store_value(other)),
1888 /* if the might be an alias, we cannot pass this Store */
1889 if (rel != ir_no_alias)
1892 /* only Phis and pure Calls are allowed here, so ignore them */
1894 if (other == NULL) {
1899 /* yep, no aliasing Store found, Load can be moved */
1900 DB((dbg, LEVEL_1, " Found a Load that could be moved: %+F\n", load));
1902 db = get_irn_dbg_info(load);
1903 for (pe = phi_list; pe != NULL; pe = pe->next) {
1905 ir_node *phi = pe->phi;
1906 ir_node *blk = get_nodes_block(phi);
1907 ir_node *pred = get_Block_cfgpred_block(blk, pos);
1909 avail_entry_t entry, *res;
1912 entry.mode = load_mode;
1913 res = set_find(avail, &entry, sizeof(entry), hash_cache_entry(&entry));
1917 irn = new_rd_Load(db, current_ir_graph, pred, get_Phi_pred(phi, pos), ptr, load_mode, 0);
1919 set_insert(avail, &entry, sizeof(entry), hash_cache_entry(&entry));
1920 DB((dbg, LEVEL_1, " Created %+F in %+F\n", irn, pred));
1923 ninfo = get_ldst_info(irn, phase_obst(&env->ph));
1925 ninfo->projs[pn_Load_M] = mem = new_r_Proj(current_ir_graph, pred, irn, mode_M, pn_Load_M);
1926 set_Phi_pred(phi, pos, mem);
1928 ninfo->projs[pn_Load_res] = new_r_Proj(current_ir_graph, pred, irn, load_mode, pn_Load_res);
1931 /* now kill the old Load */
1932 exchange(info->projs[pn_Load_M], get_Load_mem(load));
1933 exchange(info->projs[pn_Load_res], ninfo->projs[pn_Load_res]);
1935 env->changes |= DF_CHANGED;
1940 } /* move_loads_out_of_loops */
1943 * Process a loop SCC.
1945 * @param pscc the SCC
1946 * @param env the loop environment
1948 static void process_loop(scc *pscc, loop_env *env) {
1949 ir_node *irn, *next, *header = NULL;
1950 node_entry *b, *h = NULL;
1951 int j, only_phi, num_outside, process = 0;
1954 /* find the header block for this scc */
1955 for (irn = pscc->head; irn; irn = next) {
1956 node_entry *e = get_irn_ne(irn, env);
1957 ir_node *block = get_nodes_block(irn);
1960 b = get_irn_ne(block, env);
1962 if (header != NULL) {
1963 if (h->POnum < b->POnum) {
1973 /* check if this scc contains only Phi, Loads or Stores nodes */
1977 for (irn = pscc->head; irn; irn = next) {
1978 node_entry *e = get_irn_ne(irn, env);
1981 switch (get_irn_opcode(irn)) {
1983 if (is_Call_pure(irn)) {
1984 /* pure calls can be treated like loads */
1988 /* non-pure calls must be handle like may-alias Stores */
1991 /* cannot handle CopyB yet */
1995 if (get_Load_volatility(irn) == volatility_is_volatile) {
1996 /* cannot handle loops with volatile Loads */
2002 if (get_Store_volatility(irn) == volatility_is_volatile) {
2003 /* cannot handle loops with volatile Stores */
2012 for (j = get_irn_arity(irn) - 1; j >= 0; --j) {
2013 ir_node *pred = get_irn_n(irn, j);
2014 node_entry *pe = get_irn_ne(pred, env);
2016 if (pe->pscc != e->pscc) {
2017 /* not in the same SCC, must be a region const */
2018 if (! is_rc(pred, header)) {
2019 /* not a memory loop */
2022 if (out_rc == NULL) {
2023 /* first region constant */
2026 } else if (out_rc != pred) {
2027 /* another region constant */
2038 /* found a memory loop */
2039 DB((dbg, LEVEL_2, " Found a memory loop:\n "));
2040 if (only_phi && num_outside == 1) {
2041 /* a phi cycle with only one real predecessor can be collapsed */
2042 DB((dbg, LEVEL_2, " Found an USELESS Phi cycle:\n "));
2044 for (irn = pscc->head; irn; irn = next) {
2045 node_entry *e = get_irn_ne(irn, env);
2047 exchange(irn, out_rc);
2049 env->changes |= DF_CHANGED;
2053 #ifdef DEBUG_libfirm
2054 for (irn = pscc->head; irn; irn = next) {
2055 node_entry *e = get_irn_ne(irn, env);
2057 DB((dbg, LEVEL_2, " %+F,", irn));
2059 DB((dbg, LEVEL_2, "\n"));
2061 move_loads_out_of_loops(pscc, env);
2065 } /* process_loop */
2070 * @param pscc the SCC
2071 * @param env the loop environment
2073 static void process_scc(scc *pscc, loop_env *env) {
2074 ir_node *head = pscc->head;
2075 node_entry *e = get_irn_ne(head, env);
2077 #ifdef DEBUG_libfirm
2079 ir_node *irn, *next;
2081 DB((dbg, LEVEL_4, " SCC at %p:\n ", pscc));
2082 for (irn = pscc->head; irn; irn = next) {
2083 node_entry *e = get_irn_ne(irn, env);
2087 DB((dbg, LEVEL_4, " %+F,", irn));
2089 DB((dbg, LEVEL_4, "\n"));
2093 if (e->next != NULL) {
2094 /* this SCC has more than one member */
2095 process_loop(pscc, env);
2100 * Do Tarjan's SCC algorithm and drive load/store optimization.
2102 * @param irn start at this node
2103 * @param env the loop environment
2105 static void dfs(ir_node *irn, loop_env *env)
2108 node_entry *node = get_irn_ne(irn, env);
2110 mark_irn_visited(irn);
2112 node->DFSnum = env->nextDFSnum++;
2113 node->low = node->DFSnum;
2117 if (is_Phi(irn) || is_Sync(irn)) {
2118 n = get_irn_arity(irn);
2119 for (i = 0; i < n; ++i) {
2120 ir_node *pred = get_irn_n(irn, i);
2121 node_entry *o = get_irn_ne(pred, env);
2123 if (!irn_visited(pred)) {
2125 node->low = MIN(node->low, o->low);
2127 if (o->DFSnum < node->DFSnum && o->in_stack)
2128 node->low = MIN(o->DFSnum, node->low);
2130 } else if (is_fragile_op(irn)) {
2131 ir_node *pred = get_fragile_op_mem(irn);
2132 node_entry *o = get_irn_ne(pred, env);
2134 if (!irn_visited(pred)) {
2136 node->low = MIN(node->low, o->low);
2138 if (o->DFSnum < node->DFSnum && o->in_stack)
2139 node->low = MIN(o->DFSnum, node->low);
2140 } else if (is_Proj(irn)) {
2141 ir_node *pred = get_Proj_pred(irn);
2142 node_entry *o = get_irn_ne(pred, env);
2144 if (!irn_visited(pred)) {
2146 node->low = MIN(node->low, o->low);
2148 if (o->DFSnum < node->DFSnum && o->in_stack)
2149 node->low = MIN(o->DFSnum, node->low);
2152 /* IGNORE predecessors */
2155 if (node->low == node->DFSnum) {
2156 scc *pscc = phase_alloc(&env->ph, sizeof(*pscc));
2164 e = get_irn_ne(x, env);
2166 e->next = pscc->head;
2170 process_scc(pscc, env);
2175 * Do the DFS on the memory edges a graph.
2177 * @param irg the graph to process
2178 * @param env the loop environment
2180 static void do_dfs(ir_graph *irg, loop_env *env) {
2181 ir_graph *rem = current_ir_graph;
2182 ir_node *endblk, *end;
2185 current_ir_graph = irg;
2186 inc_irg_visited(irg);
2188 /* visit all memory nodes */
2189 endblk = get_irg_end_block(irg);
2190 for (i = get_Block_n_cfgpreds(endblk) - 1; i >= 0; --i) {
2191 ir_node *pred = get_Block_cfgpred(endblk, i);
2193 pred = skip_Proj(pred);
2194 if (is_Return(pred))
2195 dfs(get_Return_mem(pred), env);
2196 else if (is_Raise(pred))
2197 dfs(get_Raise_mem(pred), env);
2198 else if (is_fragile_op(pred))
2199 dfs(get_fragile_op_mem(pred), env);
2201 assert(0 && "Unknown EndBlock predecessor");
2205 /* visit the keep-alives */
2206 end = get_irg_end(irg);
2207 for (i = get_End_n_keepalives(end) - 1; i >= 0; --i) {
2208 ir_node *ka = get_End_keepalive(end, i);
2210 if (is_Phi(ka) && !irn_visited(ka))
2213 current_ir_graph = rem;
2217 * Initialize new phase data. We do this always explicit, so return NULL here
2219 static void *init_loop_data(ir_phase *ph, const ir_node *irn, void *data) {
2224 } /* init_loop_data */
2227 * Optimize Loads/Stores in loops.
2229 * @param irg the graph
2231 static int optimize_loops(ir_graph *irg) {
2234 env.stack = NEW_ARR_F(ir_node *, 128);
2239 phase_init(&env.ph, "ldstopt", irg, PHASE_DEFAULT_GROWTH, init_loop_data, NULL);
2241 /* calculate the SCC's and drive loop optimization. */
2244 DEL_ARR_F(env.stack);
2245 phase_free(&env.ph);
2248 } /* optimize_loops */
2251 * do the load store optimization
2253 int optimize_load_store(ir_graph *irg) {
2256 FIRM_DBG_REGISTER(dbg, "firm.opt.ldstopt");
2258 assert(get_irg_phase_state(irg) != phase_building);
2259 assert(get_irg_pinned(irg) != op_pin_state_floats &&
2260 "LoadStore optimization needs pinned graph");
2262 /* we need landing pads */
2263 remove_critical_cf_edges(irg);
2267 /* for Phi optimization post-dominators are needed ... */
2268 assure_postdoms(irg);
2270 if (get_opt_alias_analysis()) {
2271 assure_irg_entity_usage_computed(irg);
2272 assure_irp_globals_entity_usage_computed();
2275 obstack_init(&env.obst);
2278 /* init the links, then collect Loads/Stores/Proj's in lists */
2280 irg_walk_graph(irg, firm_clear_link, collect_nodes, &env);
2282 /* now we have collected enough information, optimize */
2283 irg_walk_graph(irg, NULL, do_load_store_optimize, &env);
2285 env.changes |= optimize_loops(irg);
2287 obstack_free(&env.obst, NULL);
2289 /* Handle graph state */
2291 set_irg_outs_inconsistent(irg);
2292 set_irg_entity_usage_state(irg, ir_entity_usage_not_computed);
2295 if (env.changes & CF_CHANGED) {
2296 /* is this really needed: Yes, control flow changed, block might
2297 have Bad() predecessors. */
2298 set_irg_doms_inconsistent(irg);
2300 return env.changes != 0;
2301 } /* optimize_load_store */