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"
51 #include "irphase_t.h"
55 /** The debug handle. */
56 DEBUG_ONLY(static firm_dbg_module_t *dbg;)
59 #include "cacheopt/cachesim.h"
63 #define IMAX(a,b) ((a) > (b) ? (a) : (b))
65 #define MAX_PROJ IMAX(IMAX(pn_Load_max, pn_Store_max), pn_Call_max)
68 DF_CHANGED = 1, /**< data flow changed */
69 CF_CHANGED = 2, /**< control flow changed */
75 typedef struct _walk_env_t {
76 struct obstack obst; /**< list of all stores */
77 unsigned changes; /**< a bitmask of graph changes */
80 /** A Load/Store info. */
81 typedef struct _ldst_info_t {
82 ir_node *projs[MAX_PROJ]; /**< list of Proj's of this node */
83 ir_node *exc_block; /**< the exception block if available */
84 int exc_idx; /**< predecessor index in the exception block */
85 unsigned visited; /**< visited counter for breaking loops */
89 * flags for control flow.
92 BLOCK_HAS_COND = 1, /**< Block has conditional control flow */
93 BLOCK_HAS_EXC = 2 /**< Block has exceptional control flow */
99 typedef struct _block_info_t {
100 unsigned flags; /**< flags for the block */
103 /** the master visited flag for loop detection. */
104 static unsigned master_visited = 0;
106 #define INC_MASTER() ++master_visited
107 #define MARK_NODE(info) (info)->visited = master_visited
108 #define NODE_VISITED(info) (info)->visited >= master_visited
111 * get the Load/Store info of a node
113 static ldst_info_t *get_ldst_info(ir_node *node, struct obstack *obst) {
114 ldst_info_t *info = get_irn_link(node);
117 info = obstack_alloc(obst, sizeof(*info));
118 memset(info, 0, sizeof(*info));
119 set_irn_link(node, info);
122 } /* get_ldst_info */
125 * get the Block info of a node
127 static block_info_t *get_block_info(ir_node *node, struct obstack *obst) {
128 block_info_t *info = get_irn_link(node);
131 info = obstack_alloc(obst, sizeof(*info));
132 memset(info, 0, sizeof(*info));
133 set_irn_link(node, info);
136 } /* get_block_info */
139 * update the projection info for a Load/Store
141 static unsigned update_projs(ldst_info_t *info, ir_node *proj)
143 long nr = get_Proj_proj(proj);
145 assert(0 <= nr && nr <= MAX_PROJ && "Wrong proj from LoadStore");
147 if (info->projs[nr]) {
148 /* there is already one, do CSE */
149 exchange(proj, info->projs[nr]);
153 info->projs[nr] = proj;
159 * update the exception block info for a Load/Store node.
161 * @param info the load/store info struct
162 * @param block the exception handler block for this load/store
163 * @param pos the control flow input of the block
165 static unsigned update_exc(ldst_info_t *info, ir_node *block, int pos)
167 assert(info->exc_block == NULL && "more than one exception block found");
169 info->exc_block = block;
174 /** Return the number of uses of an address node */
175 #define get_irn_n_uses(adr) get_irn_n_edges(adr)
178 * walker, collects all Load/Store/Proj nodes
180 * walks from Start -> End
182 static void collect_nodes(ir_node *node, void *env)
184 ir_opcode opcode = get_irn_opcode(node);
185 ir_node *pred, *blk, *pred_blk;
186 ldst_info_t *ldst_info;
187 walk_env_t *wenv = env;
189 if (opcode == iro_Proj) {
190 pred = get_Proj_pred(node);
191 opcode = get_irn_opcode(pred);
193 if (opcode == iro_Load || opcode == iro_Store || opcode == iro_Call) {
194 ldst_info = get_ldst_info(pred, &wenv->obst);
196 wenv->changes |= update_projs(ldst_info, node);
199 * Place the Proj's to the same block as the
200 * predecessor Load. This is always ok and prevents
201 * "non-SSA" form after optimizations if the Proj
202 * is in a wrong block.
204 blk = get_nodes_block(node);
205 pred_blk = get_nodes_block(pred);
206 if (blk != pred_blk) {
207 wenv->changes |= DF_CHANGED;
208 set_nodes_block(node, pred_blk);
211 } else if (opcode == iro_Block) {
214 for (i = get_Block_n_cfgpreds(node) - 1; i >= 0; --i) {
215 ir_node *pred_block, *proj;
216 block_info_t *bl_info;
219 pred = proj = get_Block_cfgpred(node, i);
222 pred = get_Proj_pred(proj);
223 is_exc = get_Proj_proj(proj) == pn_Generic_X_except;
226 /* ignore Bad predecessors, they will be removed later */
230 pred_block = get_nodes_block(pred);
231 bl_info = get_block_info(pred_block, &wenv->obst);
233 if (is_fragile_op(pred) && is_exc)
234 bl_info->flags |= BLOCK_HAS_EXC;
235 else if (is_irn_forking(pred))
236 bl_info->flags |= BLOCK_HAS_COND;
238 opcode = get_irn_opcode(pred);
239 if (is_exc && (opcode == iro_Load || opcode == iro_Store || opcode == iro_Call)) {
240 ldst_info = get_ldst_info(pred, &wenv->obst);
242 wenv->changes |= update_exc(ldst_info, node, i);
246 } /* collect_nodes */
249 * Returns an entity if the address ptr points to a constant one.
251 * @param ptr the address
253 * @return an entity or NULL
255 static ir_entity *find_constant_entity(ir_node *ptr)
258 if (is_SymConst(ptr) && get_SymConst_kind(ptr) == symconst_addr_ent) {
259 return get_SymConst_entity(ptr);
260 } else if (is_Sel(ptr)) {
261 ir_entity *ent = get_Sel_entity(ptr);
262 ir_type *tp = get_entity_owner(ent);
264 /* Do not fiddle with polymorphism. */
265 if (is_Class_type(get_entity_owner(ent)) &&
266 ((get_entity_n_overwrites(ent) != 0) ||
267 (get_entity_n_overwrittenby(ent) != 0) ) )
270 if (is_Array_type(tp)) {
274 for (i = 0, n = get_Sel_n_indexs(ptr); i < n; ++i) {
276 tarval *tlower, *tupper;
277 ir_node *index = get_Sel_index(ptr, i);
278 tarval *tv = computed_value(index);
280 /* check if the index is constant */
281 if (tv == tarval_bad)
284 bound = get_array_lower_bound(tp, i);
285 tlower = computed_value(bound);
286 bound = get_array_upper_bound(tp, i);
287 tupper = computed_value(bound);
289 if (tlower == tarval_bad || tupper == tarval_bad)
292 if (tarval_cmp(tv, tlower) & pn_Cmp_Lt)
294 if (tarval_cmp(tupper, tv) & pn_Cmp_Lt)
297 /* ok, bounds check finished */
301 if (variability_constant == get_entity_variability(ent))
305 ptr = get_Sel_ptr(ptr);
306 } else if (is_Add(ptr)) {
307 ir_node *l = get_Add_left(ptr);
308 ir_node *r = get_Add_right(ptr);
310 if (get_irn_mode(l) == get_irn_mode(ptr) && is_Const(r))
312 else if (get_irn_mode(r) == get_irn_mode(ptr) && is_Const(l))
317 /* for now, we support only one addition, reassoc should fold all others */
318 if (! is_SymConst(ptr) && !is_Sel(ptr))
320 } else if (is_Sub(ptr)) {
321 ir_node *l = get_Sub_left(ptr);
322 ir_node *r = get_Sub_right(ptr);
324 if (get_irn_mode(l) == get_irn_mode(ptr) && is_Const(r))
328 /* for now, we support only one substraction, reassoc should fold all others */
329 if (! is_SymConst(ptr) && !is_Sel(ptr))
334 } /* find_constant_entity */
337 * Return the Selection index of a Sel node from dimension n
339 static long get_Sel_array_index_long(ir_node *n, int dim) {
340 ir_node *index = get_Sel_index(n, dim);
341 assert(is_Const(index));
342 return get_tarval_long(get_Const_tarval(index));
343 } /* get_Sel_array_index_long */
346 * Returns the accessed component graph path for an
347 * node computing an address.
349 * @param ptr the node computing the address
350 * @param depth current depth in steps upward from the root
353 static compound_graph_path *rec_get_accessed_path(ir_node *ptr, int depth) {
354 compound_graph_path *res = NULL;
355 ir_entity *root, *field, *ent;
356 int path_len, pos, idx;
360 if (is_SymConst(ptr)) {
361 /* a SymConst. If the depth is 0, this is an access to a global
362 * entity and we don't need a component path, else we know
363 * at least its length.
365 assert(get_SymConst_kind(ptr) == symconst_addr_ent);
366 root = get_SymConst_entity(ptr);
367 res = (depth == 0) ? NULL : new_compound_graph_path(get_entity_type(root), depth);
368 } else if (is_Sel(ptr)) {
369 /* it's a Sel, go up until we find the root */
370 res = rec_get_accessed_path(get_Sel_ptr(ptr), depth+1);
374 /* fill up the step in the path at the current position */
375 field = get_Sel_entity(ptr);
376 path_len = get_compound_graph_path_length(res);
377 pos = path_len - depth - 1;
378 set_compound_graph_path_node(res, pos, field);
380 if (is_Array_type(get_entity_owner(field))) {
381 assert(get_Sel_n_indexs(ptr) == 1 && "multi dim arrays not implemented");
382 set_compound_graph_path_array_index(res, pos, get_Sel_array_index_long(ptr, 0));
384 } else if (is_Add(ptr)) {
385 ir_node *l = get_Add_left(ptr);
386 ir_node *r = get_Add_right(ptr);
387 ir_mode *mode = get_irn_mode(ptr);
390 if (is_Const(r) && get_irn_mode(l) == mode) {
392 tv = get_Const_tarval(r);
395 tv = get_Const_tarval(l);
398 mode = get_tarval_mode(tv);
401 /* ptr must be a Sel or a SymConst, this was checked in find_constant_entity() */
403 field = get_Sel_entity(ptr);
405 field = get_SymConst_entity(ptr);
408 for (ent = field;;) {
410 tarval *sz, *tv_index, *tlower, *tupper;
413 tp = get_entity_type(ent);
414 if (! is_Array_type(tp))
416 ent = get_array_element_entity(tp);
417 size = get_type_size_bytes(get_entity_type(ent));
418 sz = new_tarval_from_long(size, mode);
420 tv_index = tarval_div(tmp, sz);
421 tmp = tarval_mod(tmp, sz);
423 if (tv_index == tarval_bad || tmp == tarval_bad)
426 assert(get_array_n_dimensions(tp) == 1 && "multiarrays not implemented");
427 bound = get_array_lower_bound(tp, 0);
428 tlower = computed_value(bound);
429 bound = get_array_upper_bound(tp, 0);
430 tupper = computed_value(bound);
432 if (tlower == tarval_bad || tupper == tarval_bad)
435 if (tarval_cmp(tv_index, tlower) & pn_Cmp_Lt)
437 if (tarval_cmp(tupper, tv_index) & pn_Cmp_Lt)
440 /* ok, bounds check finished */
443 if (! tarval_is_null(tmp)) {
444 /* access to some struct/union member */
448 /* should be at least ONE array */
452 res = rec_get_accessed_path(ptr, depth + idx);
456 path_len = get_compound_graph_path_length(res);
457 pos = path_len - depth - idx;
459 for (ent = field;;) {
461 tarval *sz, *tv_index;
464 tp = get_entity_type(ent);
465 if (! is_Array_type(tp))
467 ent = get_array_element_entity(tp);
468 set_compound_graph_path_node(res, pos, ent);
470 size = get_type_size_bytes(get_entity_type(ent));
471 sz = new_tarval_from_long(size, mode);
473 tv_index = tarval_div(tv, sz);
474 tv = tarval_mod(tv, sz);
476 /* worked above, should work again */
477 assert(tv_index != tarval_bad && tv != tarval_bad);
479 /* bounds already checked above */
480 index = get_tarval_long(tv_index);
481 set_compound_graph_path_array_index(res, pos, index);
484 } else if (is_Sub(ptr)) {
485 ir_node *l = get_Sub_left(ptr);
486 ir_node *r = get_Sub_right(ptr);
489 tv = get_Const_tarval(r);
494 } /* rec_get_accessed_path */
497 * Returns an access path or NULL. The access path is only
498 * valid, if the graph is in phase_high and _no_ address computation is used.
500 static compound_graph_path *get_accessed_path(ir_node *ptr) {
501 compound_graph_path *gr = rec_get_accessed_path(ptr, 0);
503 } /* get_accessed_path */
505 typedef struct path_entry {
507 struct path_entry *next;
511 static ir_node *rec_find_compound_ent_value(ir_node *ptr, path_entry *next) {
512 path_entry entry, *p;
513 ir_entity *ent, *field;
514 ir_initializer_t *initializer;
520 if (is_SymConst(ptr)) {
522 ent = get_SymConst_entity(ptr);
523 initializer = get_entity_initializer(ent);
524 for (p = next; p != NULL;) {
525 if (initializer->kind != IR_INITIALIZER_COMPOUND)
527 n = get_initializer_compound_n_entries(initializer);
528 tp = get_entity_type(ent);
530 if (is_Array_type(tp)) {
531 ent = get_array_element_entity(tp);
536 initializer = get_initializer_compound_value(initializer, 0);
540 if (p->index >= (int) n)
542 initializer = get_initializer_compound_value(initializer, p->index);
547 tp = get_entity_type(ent);
548 while (is_Array_type(tp)) {
549 ent = get_array_element_entity(tp);
550 tp = get_entity_type(ent);
552 n = get_initializer_compound_n_entries(initializer);
555 initializer = get_initializer_compound_value(initializer, 0);
558 switch (initializer->kind) {
559 case IR_INITIALIZER_CONST:
560 return get_initializer_const_value(initializer);
561 case IR_INITIALIZER_TARVAL:
562 case IR_INITIALIZER_NULL:
566 } else if (is_Sel(ptr)) {
567 entry.ent = field = get_Sel_entity(ptr);
568 tp = get_entity_owner(field);
569 if (is_Array_type(tp)) {
570 assert(get_Sel_n_indexs(ptr) == 1 && "multi dim arrays not implemented");
571 entry.index = get_Sel_array_index_long(ptr, 0) - get_array_lower_bound_int(tp, 0);
573 int i, n_members = get_compound_n_members(tp);
574 for (i = 0; i < n_members; ++i) {
575 if (get_compound_member(tp, i) == field)
578 if (i >= n_members) {
579 /* not found: should NOT happen */
584 return rec_find_compound_ent_value(get_Sel_ptr(ptr), &entry);
585 } else if (is_Add(ptr)) {
586 ir_node *l = get_Add_left(ptr);
587 ir_node *r = get_Add_right(ptr);
593 tv = get_Const_tarval(r);
596 tv = get_Const_tarval(l);
599 mode = get_tarval_mode(tv);
601 /* ptr must be a Sel or a SymConst, this was checked in find_constant_entity() */
603 field = get_Sel_entity(ptr);
605 field = get_SymConst_entity(ptr);
608 /* count needed entries */
610 for (ent = field;;) {
611 tp = get_entity_type(ent);
612 if (! is_Array_type(tp))
614 ent = get_array_element_entity(tp);
617 /* should be at least ONE entry */
621 /* allocate the right number of entries */
622 NEW_ARR_A(path_entry, p, pos);
626 for (ent = field;;) {
628 tarval *sz, *tv_index, *tlower, *tupper;
632 tp = get_entity_type(ent);
633 if (! is_Array_type(tp))
635 ent = get_array_element_entity(tp);
637 p[pos].next = &p[pos + 1];
639 size = get_type_size_bytes(get_entity_type(ent));
640 sz = new_tarval_from_long(size, mode);
642 tv_index = tarval_div(tv, sz);
643 tv = tarval_mod(tv, sz);
645 if (tv_index == tarval_bad || tv == tarval_bad)
648 assert(get_array_n_dimensions(tp) == 1 && "multiarrays not implemented");
649 bound = get_array_lower_bound(tp, 0);
650 tlower = computed_value(bound);
651 bound = get_array_upper_bound(tp, 0);
652 tupper = computed_value(bound);
654 if (tlower == tarval_bad || tupper == tarval_bad)
657 if (tarval_cmp(tv_index, tlower) & pn_Cmp_Lt)
659 if (tarval_cmp(tupper, tv_index) & pn_Cmp_Lt)
662 /* ok, bounds check finished */
663 index = get_tarval_long(tv_index);
664 p[pos].index = index;
667 if (! tarval_is_null(tv)) {
668 /* hmm, wrong access */
671 p[pos - 1].next = next;
672 return rec_find_compound_ent_value(ptr, p);
673 } else if (is_Sub(ptr)) {
674 ir_node *l = get_Sub_left(ptr);
675 ir_node *r = get_Sub_right(ptr);
678 tv = get_Const_tarval(r);
685 static ir_node *find_compound_ent_value(ir_node *ptr) {
686 return rec_find_compound_ent_value(ptr, NULL);
690 static void reduce_adr_usage(ir_node *ptr);
693 * Update a Load that may have lost its users.
695 static void handle_load_update(ir_node *load) {
696 ldst_info_t *info = get_irn_link(load);
698 /* do NOT touch volatile loads for now */
699 if (get_Load_volatility(load) == volatility_is_volatile)
702 if (! info->projs[pn_Load_res] && ! info->projs[pn_Load_X_except]) {
703 ir_node *ptr = get_Load_ptr(load);
704 ir_node *mem = get_Load_mem(load);
706 /* a Load whose value is neither used nor exception checked, remove it */
707 exchange(info->projs[pn_Load_M], mem);
708 if (info->projs[pn_Load_X_regular])
709 exchange(info->projs[pn_Load_X_regular], new_r_Jmp(current_ir_graph, get_nodes_block(load)));
711 reduce_adr_usage(ptr);
713 } /* handle_load_update */
716 * A use of an address node has vanished. Check if this was a Proj
717 * node and update the counters.
719 static void reduce_adr_usage(ir_node *ptr) {
721 if (get_irn_n_edges(ptr) <= 0) {
722 /* this Proj is dead now */
723 ir_node *pred = get_Proj_pred(ptr);
726 ldst_info_t *info = get_irn_link(pred);
727 info->projs[get_Proj_proj(ptr)] = NULL;
729 /* this node lost its result proj, handle that */
730 handle_load_update(pred);
734 } /* reduce_adr_usage */
737 * Check, if an already existing value of mode old_mode can be converted
738 * into the needed one new_mode without loss.
740 static int can_use_stored_value(ir_mode *old_mode, ir_mode *new_mode) {
741 if (old_mode == new_mode)
744 /* if both modes are two-complement ones, we can always convert the
745 Stored value into the needed one. */
746 if (get_mode_size_bits(old_mode) >= get_mode_size_bits(new_mode) &&
747 get_mode_arithmetic(old_mode) == irma_twos_complement &&
748 get_mode_arithmetic(new_mode) == irma_twos_complement)
751 } /* can_use_stored_value */
754 * Check whether a Call is at least pure, ie. does only read memory.
756 static unsigned is_Call_pure(ir_node *call) {
757 ir_type *call_tp = get_Call_type(call);
758 unsigned prop = get_method_additional_properties(call_tp);
760 /* check first the call type */
761 if ((prop & (mtp_property_const|mtp_property_pure)) == 0) {
762 /* try the called entity */
763 ir_node *ptr = get_Call_ptr(call);
765 if (is_Global(ptr)) {
766 ir_entity *ent = get_Global_entity(ptr);
768 prop = get_entity_additional_properties(ent);
771 return (prop & (mtp_property_const|mtp_property_pure)) != 0;
774 static ir_node *get_base_and_offset(ir_node *ptr, long *pOffset)
776 ir_mode *mode = get_irn_mode(ptr);
779 /* TODO: long might not be enough, we should probably use some tarval thingy... */
782 ir_node *l = get_Add_left(ptr);
783 ir_node *r = get_Add_right(ptr);
785 if (get_irn_mode(l) != mode || !is_Const(r))
788 offset += get_tarval_long(get_Const_tarval(r));
790 } else if (is_Sub(ptr)) {
791 ir_node *l = get_Sub_left(ptr);
792 ir_node *r = get_Sub_right(ptr);
794 if (get_irn_mode(l) != mode || !is_Const(r))
797 offset -= get_tarval_long(get_Const_tarval(r));
799 } else if (is_Sel(ptr)) {
800 ir_entity *ent = get_Sel_entity(ptr);
801 ir_type *tp = get_entity_owner(ent);
803 if (is_Array_type(tp)) {
807 /* only one dimensional arrays yet */
808 if (get_Sel_n_indexs(ptr) != 1)
810 index = get_Sel_index(ptr, 0);
811 if (! is_Const(index))
814 tp = get_entity_type(ent);
815 if (get_type_state(tp) != layout_fixed)
818 size = get_type_size_bytes(tp);
819 offset += size * get_tarval_long(get_Const_tarval(index));
821 if (get_type_state(tp) != layout_fixed)
823 offset += get_entity_offset(ent);
825 ptr = get_Sel_ptr(ptr);
834 static int try_load_after_store(ir_node *load,
835 ir_node *load_base_ptr, long load_offset, ir_node *store)
838 ir_node *store_ptr = get_Store_ptr(store);
840 ir_node *store_base_ptr = get_base_and_offset(store_ptr, &store_offset);
841 ir_node *store_value;
850 if (load_base_ptr != store_base_ptr)
853 load_mode = get_Load_mode(load);
854 load_mode_len = get_mode_size_bytes(load_mode);
855 store_mode = get_irn_mode(get_Store_value(store));
856 store_mode_len = get_mode_size_bytes(store_mode);
857 delta = load_offset - store_offset;
858 store_value = get_Store_value(store);
860 if (delta != 0 || store_mode != load_mode) {
861 if (delta < 0 || delta + load_mode_len > store_mode_len)
864 if (get_mode_arithmetic(store_mode) != irma_twos_complement ||
865 get_mode_arithmetic(load_mode) != irma_twos_complement)
869 /* produce a shift to adjust offset delta */
873 /* FIXME: only true for little endian */
874 cnst = new_Const_long(mode_Iu, delta * 8);
875 store_value = new_r_Shr(current_ir_graph, get_nodes_block(load),
876 store_value, cnst, store_mode);
879 /* add an convert if needed */
880 if (store_mode != load_mode) {
881 store_value = new_r_Conv(current_ir_graph, get_nodes_block(load),
882 store_value, load_mode);
886 DBG_OPT_RAW(load, store_value);
888 info = get_irn_link(load);
889 if (info->projs[pn_Load_M])
890 exchange(info->projs[pn_Load_M], get_Load_mem(load));
894 if (info->projs[pn_Load_X_except]) {
895 exchange( info->projs[pn_Load_X_except], new_Bad());
898 if (info->projs[pn_Load_X_regular]) {
899 exchange( info->projs[pn_Load_X_regular], new_r_Jmp(current_ir_graph, get_nodes_block(load)));
903 if (info->projs[pn_Load_res])
904 exchange(info->projs[pn_Load_res], store_value);
906 load_ptr = get_Load_ptr(load);
908 reduce_adr_usage(load_ptr);
909 return res | DF_CHANGED;
913 * Follow the memory chain as long as there are only Loads,
914 * alias free Stores, and constant Calls and try to replace the
915 * current Load by a previous ones.
916 * Note that in unreachable loops it might happen that we reach
917 * load again, as well as we can fall into a cycle.
918 * We break such cycles using a special visited flag.
920 * INC_MASTER() must be called before dive into
922 static unsigned follow_Mem_chain(ir_node *load, ir_node *curr) {
924 ldst_info_t *info = get_irn_link(load);
926 ir_node *ptr = get_Load_ptr(load);
927 ir_node *mem = get_Load_mem(load);
928 ir_mode *load_mode = get_Load_mode(load);
930 for (pred = curr; load != pred; ) {
931 ldst_info_t *pred_info = get_irn_link(pred);
934 * a Load immediately after a Store -- a read after write.
935 * We may remove the Load, if both Load & Store does not have an
936 * exception handler OR they are in the same MacroBlock. In the latter
937 * case the Load cannot throw an exception when the previous Store was
940 * Why we need to check for Store Exception? If the Store cannot
941 * be executed (ROM) the exception handler might simply jump into
942 * the load MacroBlock :-(
943 * We could make it a little bit better if we would know that the
944 * exception handler of the Store jumps directly to the end...
946 if (is_Store(pred) && ((pred_info->projs[pn_Store_X_except] == NULL
947 && info->projs[pn_Load_X_except] == NULL)
948 || get_nodes_MacroBlock(load) == get_nodes_MacroBlock(pred)))
951 ir_node *base_ptr = get_base_and_offset(ptr, &load_offset);
952 int changes = try_load_after_store(load, base_ptr, load_offset, pred);
955 return res | changes;
956 } else if (is_Load(pred) && get_Load_ptr(pred) == ptr &&
957 can_use_stored_value(get_Load_mode(pred), load_mode)) {
959 * a Load after a Load -- a read after read.
960 * We may remove the second Load, if it does not have an exception handler
961 * OR they are in the same MacroBlock. In the later case the Load cannot
962 * throw an exception when the previous Load was quiet.
964 * Here, there is no need to check if the previous Load has an exception
965 * hander because they would have exact the same exception...
967 if (info->projs[pn_Load_X_except] == NULL || get_nodes_MacroBlock(load) == get_nodes_MacroBlock(pred)) {
970 DBG_OPT_RAR(load, pred);
972 /* the result is used */
973 if (info->projs[pn_Load_res]) {
974 if (pred_info->projs[pn_Load_res] == NULL) {
975 /* create a new Proj again */
976 pred_info->projs[pn_Load_res] = new_r_Proj(current_ir_graph, get_nodes_block(pred), pred, get_Load_mode(pred), pn_Load_res);
978 value = pred_info->projs[pn_Load_res];
980 /* add an convert if needed */
981 if (get_Load_mode(pred) != load_mode) {
982 value = new_r_Conv(current_ir_graph, get_nodes_block(load), value, load_mode);
985 exchange(info->projs[pn_Load_res], value);
988 if (info->projs[pn_Load_M])
989 exchange(info->projs[pn_Load_M], mem);
992 if (info->projs[pn_Load_X_except]) {
993 exchange(info->projs[pn_Load_X_except], new_Bad());
996 if (info->projs[pn_Load_X_regular]) {
997 exchange( info->projs[pn_Load_X_regular], new_r_Jmp(current_ir_graph, get_nodes_block(load)));
1002 reduce_adr_usage(ptr);
1003 return res |= DF_CHANGED;
1007 if (is_Store(pred)) {
1008 /* check if we can pass through this store */
1009 ir_alias_relation rel = get_alias_relation(
1011 get_Store_ptr(pred),
1012 get_irn_mode(get_Store_value(pred)),
1014 /* if the might be an alias, we cannot pass this Store */
1015 if (rel != ir_no_alias)
1017 pred = skip_Proj(get_Store_mem(pred));
1018 } else if (is_Load(pred)) {
1019 pred = skip_Proj(get_Load_mem(pred));
1020 } else if (is_Call(pred)) {
1021 if (is_Call_pure(pred)) {
1022 /* The called graph is at least pure, so there are no Store's
1023 in it. We can handle it like a Load and skip it. */
1024 pred = skip_Proj(get_Call_mem(pred));
1026 /* there might be Store's in the graph, stop here */
1030 /* follow only Load chains */
1034 /* check for cycles */
1035 if (NODE_VISITED(pred_info))
1037 MARK_NODE(pred_info);
1040 if (is_Sync(pred)) {
1043 /* handle all Sync predecessors */
1044 for (i = get_Sync_n_preds(pred) - 1; i >= 0; --i) {
1045 res |= follow_Mem_chain(load, skip_Proj(get_Sync_pred(pred, i)));
1052 } /* follow_Mem_chain */
1055 * Check if we can replace the load by a given const from
1056 * the const code irg.
1058 ir_node *can_replace_load_by_const(const ir_node *load, ir_node *c) {
1059 ir_mode *c_mode = get_irn_mode(c);
1060 ir_mode *l_mode = get_Load_mode(load);
1061 ir_node *res = NULL;
1063 if (c_mode != l_mode) {
1064 /* check, if the mode matches OR can be easily converted info */
1065 if (is_reinterpret_cast(c_mode, l_mode)) {
1066 /* we can safely cast */
1067 dbg_info *dbg = get_irn_dbg_info(load);
1068 ir_node *block = get_nodes_block(load);
1070 /* copy the value from the const code irg and cast it */
1071 res = copy_const_value(dbg, c);
1072 res = new_rd_Conv(dbg, current_ir_graph, block, res, l_mode);
1075 /* copy the value from the const code irg */
1076 res = copy_const_value(get_irn_dbg_info(load), c);
1079 } /* can_replace_load_by_const */
1084 * @param load the Load node
1086 static unsigned optimize_load(ir_node *load)
1088 ldst_info_t *info = get_irn_link(load);
1089 ir_node *mem, *ptr, *value;
1094 /* do NOT touch volatile loads for now */
1095 if (get_Load_volatility(load) == volatility_is_volatile)
1098 /* the address of the load to be optimized */
1099 ptr = get_Load_ptr(load);
1102 * Check if we can remove the exception from a Load:
1103 * This can be done, if the address is from an Sel(Alloc) and
1104 * the Sel type is a subtype of the allocated type.
1106 * This optimizes some often used OO constructs,
1107 * like x = new O; x->t;
1109 if (info->projs[pn_Load_X_except]) {
1110 ir_node *addr = ptr;
1112 /* find base address */
1113 while (is_Sel(addr))
1114 addr = get_Sel_ptr(addr);
1115 if (is_Alloc(skip_Proj(skip_Cast(addr)))) {
1116 /* simple case: a direct load after an Alloc. Firm Alloc throw
1117 * an exception in case of out-of-memory. So, there is no way for an
1118 * exception in this load.
1119 * This code is constructed by the "exception lowering" in the Jack compiler.
1121 exchange(info->projs[pn_Load_X_except], new_Bad());
1122 info->projs[pn_Load_X_except] = NULL;
1123 exchange(info->projs[pn_Load_X_regular], new_r_Jmp(current_ir_graph, get_nodes_block(load)));
1124 info->projs[pn_Load_X_regular] = NULL;
1129 /* The mem of the Load. Must still be returned after optimization. */
1130 mem = get_Load_mem(load);
1132 if (! info->projs[pn_Load_res] && ! info->projs[pn_Load_X_except]) {
1133 /* a Load which value is neither used nor exception checked, remove it */
1134 exchange(info->projs[pn_Load_M], mem);
1136 if (info->projs[pn_Load_X_regular]) {
1137 /* should not happen, but if it does, remove it */
1138 exchange(info->projs[pn_Load_X_regular], new_r_Jmp(current_ir_graph, get_nodes_block(load)));
1142 reduce_adr_usage(ptr);
1143 return res | DF_CHANGED;
1146 /* Load from a constant polymorphic field, where we can resolve
1148 value = transform_polymorph_Load(load);
1149 if (value == load) {
1151 /* check if we can determine the entity that will be loaded */
1152 ent = find_constant_entity(ptr);
1154 allocation_static == get_entity_allocation(ent) &&
1155 visibility_external_allocated != get_entity_visibility(ent)) {
1156 /* a static allocation that is not external: there should be NO exception
1157 * when loading even if we cannot replace the load itself. */
1159 /* no exception, clear the info field as it might be checked later again */
1160 if (info->projs[pn_Load_X_except]) {
1161 exchange(info->projs[pn_Load_X_except], new_Bad());
1162 info->projs[pn_Load_X_except] = NULL;
1165 if (info->projs[pn_Load_X_regular]) {
1166 exchange(info->projs[pn_Load_X_regular], new_r_Jmp(current_ir_graph, get_nodes_block(load)));
1167 info->projs[pn_Load_X_regular] = NULL;
1171 if (variability_constant == get_entity_variability(ent)) {
1172 if (is_atomic_entity(ent)) {
1173 /* Might not be atomic after lowering of Sels. In this case we
1174 * could also load, but it's more complicated. */
1175 /* more simpler case: we load the content of a constant value:
1176 * replace it by the constant itself */
1177 value = get_atomic_ent_value(ent);
1178 } else if (ent->has_initializer) {
1179 /* new style initializer */
1180 value = find_compound_ent_value(ptr);
1182 /* old style initializer */
1183 compound_graph_path *path = get_accessed_path(ptr);
1186 assert(is_proper_compound_graph_path(path, get_compound_graph_path_length(path)-1));
1188 value = get_compound_ent_value_by_path(ent, path);
1189 DB((dbg, LEVEL_1, " Constant access at %F%F resulted in %+F\n", ent, path, value));
1190 free_compound_graph_path(path);
1194 value = can_replace_load_by_const(load, value);
1198 if (value != NULL) {
1199 /* we completely replace the load by this value */
1200 if (info->projs[pn_Load_X_except]) {
1201 exchange(info->projs[pn_Load_X_except], new_Bad());
1202 info->projs[pn_Load_X_except] = NULL;
1205 if (info->projs[pn_Load_X_regular]) {
1206 exchange(info->projs[pn_Load_X_regular], new_r_Jmp(current_ir_graph, get_nodes_block(load)));
1207 info->projs[pn_Load_X_regular] = NULL;
1210 if (info->projs[pn_Load_M]) {
1211 exchange(info->projs[pn_Load_M], mem);
1214 if (info->projs[pn_Load_res]) {
1215 exchange(info->projs[pn_Load_res], value);
1219 reduce_adr_usage(ptr);
1223 /* Check, if the address of this load is used more than once.
1224 * If not, more load cannot be removed in any case. */
1225 if (get_irn_n_uses(ptr) <= 1 && get_irn_n_uses(get_base_and_offset(ptr, &dummy)) <= 1)
1229 * follow the memory chain as long as there are only Loads
1230 * and try to replace current Load or Store by a previous one.
1231 * Note that in unreachable loops it might happen that we reach
1232 * load again, as well as we can fall into a cycle.
1233 * We break such cycles using a special visited flag.
1236 res = follow_Mem_chain(load, skip_Proj(mem));
1238 } /* optimize_load */
1241 * Check whether a value of mode new_mode would completely overwrite a value
1242 * of mode old_mode in memory.
1244 static int is_completely_overwritten(ir_mode *old_mode, ir_mode *new_mode)
1246 return get_mode_size_bits(new_mode) >= get_mode_size_bits(old_mode);
1247 } /* is_completely_overwritten */
1250 * Check whether small is a part of large (starting at same address).
1252 static int is_partially_same(ir_node *small, ir_node *large)
1254 ir_mode *sm = get_irn_mode(small);
1255 ir_mode *lm = get_irn_mode(large);
1257 /* FIXME: Check endianness */
1258 return is_Conv(small) && get_Conv_op(small) == large
1259 && get_mode_size_bytes(sm) < get_mode_size_bytes(lm)
1260 && get_mode_arithmetic(sm) == irma_twos_complement
1261 && get_mode_arithmetic(lm) == irma_twos_complement;
1262 } /* is_partially_same */
1265 * follow the memory chain as long as there are only Loads and alias free Stores.
1267 * INC_MASTER() must be called before dive into
1269 static unsigned follow_Mem_chain_for_Store(ir_node *store, ir_node *curr) {
1271 ldst_info_t *info = get_irn_link(store);
1273 ir_node *ptr = get_Store_ptr(store);
1274 ir_node *mem = get_Store_mem(store);
1275 ir_node *value = get_Store_value(store);
1276 ir_mode *mode = get_irn_mode(value);
1277 ir_node *block = get_nodes_block(store);
1278 ir_node *mblk = get_Block_MacroBlock(block);
1280 for (pred = curr; pred != store;) {
1281 ldst_info_t *pred_info = get_irn_link(pred);
1284 * BEWARE: one might think that checking the modes is useless, because
1285 * if the pointers are identical, they refer to the same object.
1286 * This is only true in strong typed languages, not is C were the following
1287 * is possible *(ir_type1 *)p = a; *(ir_type2 *)p = b ...
1288 * However, if the size of the mode that is written is bigger or equal the
1289 * size of the old one, the old value is completely overwritten and can be
1292 if (is_Store(pred) && get_Store_ptr(pred) == ptr &&
1293 get_nodes_MacroBlock(pred) == mblk) {
1295 * a Store after a Store in the same MacroBlock -- a write after write.
1299 * We may remove the first Store, if the old value is completely
1300 * overwritten or the old value is a part of the new value,
1301 * and if it does not have an exception handler.
1303 * TODO: What, if both have the same exception handler ???
1305 if (get_Store_volatility(pred) != volatility_is_volatile
1306 && !pred_info->projs[pn_Store_X_except]) {
1307 ir_node *predvalue = get_Store_value(pred);
1308 ir_mode *predmode = get_irn_mode(predvalue);
1310 if(is_completely_overwritten(predmode, mode)
1311 || is_partially_same(predvalue, value)) {
1312 DBG_OPT_WAW(pred, store);
1313 exchange(pred_info->projs[pn_Store_M], get_Store_mem(pred));
1315 reduce_adr_usage(ptr);
1321 * We may remove the Store, if the old value already contains
1322 * the new value, and if it does not have an exception handler.
1324 * TODO: What, if both have the same exception handler ???
1326 if (get_Store_volatility(store) != volatility_is_volatile
1327 && !info->projs[pn_Store_X_except]) {
1328 ir_node *predvalue = get_Store_value(pred);
1330 if(is_partially_same(value, predvalue)) {
1331 DBG_OPT_WAW(pred, store);
1332 exchange(info->projs[pn_Store_M], mem);
1334 reduce_adr_usage(ptr);
1338 } else if (is_Load(pred) && get_Load_ptr(pred) == ptr &&
1339 value == pred_info->projs[pn_Load_res]) {
1341 * a Store of a value just loaded from the same address
1342 * -- a write after read.
1343 * We may remove the Store, if it does not have an exception
1346 if (! info->projs[pn_Store_X_except]) {
1347 DBG_OPT_WAR(store, pred);
1348 exchange(info->projs[pn_Store_M], mem);
1350 reduce_adr_usage(ptr);
1355 if (is_Store(pred)) {
1356 /* check if we can pass through this store */
1357 ir_alias_relation rel = get_alias_relation(
1359 get_Store_ptr(pred),
1360 get_irn_mode(get_Store_value(pred)),
1362 /* if the might be an alias, we cannot pass this Store */
1363 if (rel != ir_no_alias)
1365 pred = skip_Proj(get_Store_mem(pred));
1366 } else if (is_Load(pred)) {
1367 ir_alias_relation rel = get_alias_relation(
1368 current_ir_graph, get_Load_ptr(pred), get_Load_mode(pred),
1370 if (rel != ir_no_alias)
1373 pred = skip_Proj(get_Load_mem(pred));
1375 /* follow only Load chains */
1379 /* check for cycles */
1380 if (NODE_VISITED(pred_info))
1382 MARK_NODE(pred_info);
1385 if (is_Sync(pred)) {
1388 /* handle all Sync predecessors */
1389 for (i = get_Sync_n_preds(pred) - 1; i >= 0; --i) {
1390 res |= follow_Mem_chain_for_Store(store, skip_Proj(get_Sync_pred(pred, i)));
1396 } /* follow_Mem_chain_for_Store */
1398 /** find entity used as base for an address calculation */
1399 static ir_entity *find_entity(ir_node *ptr)
1401 switch(get_irn_opcode(ptr)) {
1403 return get_SymConst_entity(ptr);
1405 ir_node *pred = get_Sel_ptr(ptr);
1406 if (get_irg_frame(get_irn_irg(ptr)) == pred)
1407 return get_Sel_entity(ptr);
1409 return find_entity(pred);
1413 ir_node *left = get_binop_left(ptr);
1415 if (mode_is_reference(get_irn_mode(left)))
1416 return find_entity(left);
1417 right = get_binop_right(ptr);
1418 if (mode_is_reference(get_irn_mode(right)))
1419 return find_entity(right);
1430 * @param store the Store node
1432 static unsigned optimize_store(ir_node *store) {
1437 if (get_Store_volatility(store) == volatility_is_volatile)
1440 ptr = get_Store_ptr(store);
1441 entity = find_entity(ptr);
1443 /* a store to an entity which is never read is unnecessary */
1444 if (entity != NULL && !(get_entity_usage(entity) & ir_usage_read)) {
1445 ldst_info_t *info = get_irn_link(store);
1446 if (info->projs[pn_Store_X_except] == NULL) {
1447 exchange(info->projs[pn_Store_M], get_Store_mem(store));
1449 reduce_adr_usage(ptr);
1454 /* Check, if the address of this Store is used more than once.
1455 * If not, this Store cannot be removed in any case. */
1456 if (get_irn_n_uses(ptr) <= 1)
1459 mem = get_Store_mem(store);
1461 /* follow the memory chain as long as there are only Loads */
1464 return follow_Mem_chain_for_Store(store, skip_Proj(mem));
1465 } /* optimize_store */
1468 * walker, optimizes Phi after Stores to identical places:
1469 * Does the following optimization:
1472 * val1 val2 val3 val1 val2 val3
1474 * Store Store Store \ | /
1481 * This reduces the number of stores and allows for predicated execution.
1482 * Moves Stores back to the end of a function which may be bad.
1484 * This is only possible if the predecessor blocks have only one successor.
1486 static unsigned optimize_phi(ir_node *phi, walk_env_t *wenv)
1489 ir_node *store, *old_store, *ptr, *block, *phi_block, *phiM, *phiD, *exc, *projM;
1491 ir_node **inM, **inD, **projMs;
1493 dbg_info *db = NULL;
1495 block_info_t *bl_info;
1498 /* Must be a memory Phi */
1499 if (get_irn_mode(phi) != mode_M)
1502 n = get_Phi_n_preds(phi);
1506 /* must be only one user */
1507 projM = get_Phi_pred(phi, 0);
1508 if (get_irn_n_edges(projM) != 1)
1511 store = skip_Proj(projM);
1513 if (!is_Store(store))
1516 block = get_nodes_block(store);
1518 /* abort on dead blocks */
1519 if (is_Block_dead(block))
1522 /* check if the block is post dominated by Phi-block
1523 and has no exception exit */
1524 bl_info = get_irn_link(block);
1525 if (bl_info->flags & BLOCK_HAS_EXC)
1528 phi_block = get_nodes_block(phi);
1529 if (! block_strictly_postdominates(phi_block, block))
1532 /* this is the address of the store */
1533 ptr = get_Store_ptr(store);
1534 mode = get_irn_mode(get_Store_value(store));
1535 info = get_irn_link(store);
1536 exc = info->exc_block;
1538 for (i = 1; i < n; ++i) {
1539 ir_node *pred = get_Phi_pred(phi, i);
1541 if (get_irn_n_edges(pred) != 1)
1544 pred = skip_Proj(pred);
1545 if (!is_Store(pred))
1548 if (ptr != get_Store_ptr(pred) || mode != get_irn_mode(get_Store_value(pred)))
1551 info = get_irn_link(pred);
1553 /* check, if all stores have the same exception flow */
1554 if (exc != info->exc_block)
1557 /* abort on dead blocks */
1558 block = get_nodes_block(pred);
1559 if (is_Block_dead(block))
1562 /* check if the block is post dominated by Phi-block
1563 and has no exception exit. Note that block must be different from
1564 Phi-block, else we would move a Store from end End of a block to its
1566 bl_info = get_irn_link(block);
1567 if (bl_info->flags & BLOCK_HAS_EXC)
1569 if (block == phi_block || ! block_postdominates(phi_block, block))
1574 * ok, when we are here, we found all predecessors of a Phi that
1575 * are Stores to the same address and size. That means whatever
1576 * we do before we enter the block of the Phi, we do a Store.
1577 * So, we can move the Store to the current block:
1579 * val1 val2 val3 val1 val2 val3
1581 * | Str | | Str | | Str | \ | /
1587 * Is only allowed if the predecessor blocks have only one successor.
1590 NEW_ARR_A(ir_node *, projMs, n);
1591 NEW_ARR_A(ir_node *, inM, n);
1592 NEW_ARR_A(ir_node *, inD, n);
1593 NEW_ARR_A(int, idx, n);
1595 /* Prepare: Collect all Store nodes. We must do this
1596 first because we otherwise may loose a store when exchanging its
1599 for (i = n - 1; i >= 0; --i) {
1602 projMs[i] = get_Phi_pred(phi, i);
1603 assert(is_Proj(projMs[i]));
1605 store = get_Proj_pred(projMs[i]);
1606 info = get_irn_link(store);
1608 inM[i] = get_Store_mem(store);
1609 inD[i] = get_Store_value(store);
1610 idx[i] = info->exc_idx;
1612 block = get_nodes_block(phi);
1614 /* second step: create a new memory Phi */
1615 phiM = new_rd_Phi(get_irn_dbg_info(phi), current_ir_graph, block, n, inM, mode_M);
1617 /* third step: create a new data Phi */
1618 phiD = new_rd_Phi(get_irn_dbg_info(phi), current_ir_graph, block, n, inD, mode);
1620 /* rewire memory and kill the node */
1621 for (i = n - 1; i >= 0; --i) {
1622 ir_node *proj = projMs[i];
1625 ir_node *store = get_Proj_pred(proj);
1626 exchange(proj, inM[i]);
1631 /* fourth step: create the Store */
1632 store = new_rd_Store(db, current_ir_graph, block, phiM, ptr, phiD);
1634 co_set_irn_name(store, co_get_irn_ident(old_store));
1637 projM = new_rd_Proj(NULL, current_ir_graph, block, store, mode_M, pn_Store_M);
1639 info = get_ldst_info(store, &wenv->obst);
1640 info->projs[pn_Store_M] = projM;
1642 /* fifths step: repair exception flow */
1644 ir_node *projX = new_rd_Proj(NULL, current_ir_graph, block, store, mode_X, pn_Store_X_except);
1646 info->projs[pn_Store_X_except] = projX;
1647 info->exc_block = exc;
1648 info->exc_idx = idx[0];
1650 for (i = 0; i < n; ++i) {
1651 set_Block_cfgpred(exc, idx[i], projX);
1655 /* the exception block should be optimized as some inputs are identical now */
1661 /* sixth step: replace old Phi */
1662 exchange(phi, projM);
1664 return res | DF_CHANGED;
1665 } /* optimize_phi */
1668 * walker, do the optimizations
1670 static void do_load_store_optimize(ir_node *n, void *env) {
1671 walk_env_t *wenv = env;
1673 switch (get_irn_opcode(n)) {
1676 wenv->changes |= optimize_load(n);
1680 wenv->changes |= optimize_store(n);
1684 wenv->changes |= optimize_phi(n, wenv);
1690 } /* do_load_store_optimize */
1693 typedef struct scc {
1694 ir_node *head; /**< the head of the list */
1697 /** A node entry. */
1698 typedef struct node_entry {
1699 unsigned DFSnum; /**< the DFS number of this node */
1700 unsigned low; /**< the low number of this node */
1701 ir_node *header; /**< the header of this node */
1702 int in_stack; /**< flag, set if the node is on the stack */
1703 ir_node *next; /**< link to the next node the the same scc */
1704 scc *pscc; /**< the scc of this node */
1705 unsigned POnum; /**< the post order number for blocks */
1708 /** A loop entry. */
1709 typedef struct loop_env {
1710 ir_phase ph; /**< the phase object */
1711 ir_node **stack; /**< the node stack */
1712 int tos; /**< tos index */
1713 unsigned nextDFSnum; /**< the current DFS number */
1714 unsigned POnum; /**< current post order number */
1716 unsigned changes; /**< a bitmask of graph changes */
1720 * Gets the node_entry of a node
1722 static node_entry *get_irn_ne(ir_node *irn, loop_env *env) {
1723 ir_phase *ph = &env->ph;
1724 node_entry *e = phase_get_irn_data(&env->ph, irn);
1727 e = phase_alloc(ph, sizeof(*e));
1728 memset(e, 0, sizeof(*e));
1729 phase_set_irn_data(ph, irn, e);
1735 * Push a node onto the stack.
1737 * @param env the loop environment
1738 * @param n the node to push
1740 static void push(loop_env *env, ir_node *n) {
1743 if (env->tos == ARR_LEN(env->stack)) {
1744 int nlen = ARR_LEN(env->stack) * 2;
1745 ARR_RESIZE(ir_node *, env->stack, nlen);
1747 env->stack[env->tos++] = n;
1748 e = get_irn_ne(n, env);
1753 * pop a node from the stack
1755 * @param env the loop environment
1757 * @return The topmost node
1759 static ir_node *pop(loop_env *env) {
1760 ir_node *n = env->stack[--env->tos];
1761 node_entry *e = get_irn_ne(n, env);
1768 * Check if irn is a region constant.
1769 * The block or irn must strictly dominate the header block.
1771 * @param irn the node to check
1772 * @param header_block the header block of the induction variable
1774 static int is_rc(ir_node *irn, ir_node *header_block) {
1775 ir_node *block = get_nodes_block(irn);
1777 return (block != header_block) && block_dominates(block, header_block);
1780 typedef struct phi_entry phi_entry;
1782 ir_node *phi; /**< A phi with a region const memory. */
1783 int pos; /**< The position of the region const memory */
1784 ir_node *load; /**< the newly created load for this phi */
1789 * Move loops out of loops if possible.
1791 * @param pscc the loop described by an SCC
1792 * @param env the loop environment
1794 static void move_loads_out_of_loops(scc *pscc, loop_env *env) {
1795 ir_node *phi, *load, *next, *other, *next_other;
1798 phi_entry *phi_list = NULL;
1800 /* collect all outer memories */
1801 for (phi = pscc->head; phi != NULL; phi = next) {
1802 node_entry *ne = get_irn_ne(phi, env);
1805 /* check all memory Phi's */
1809 assert(get_irn_mode(phi) == mode_M && "DFS geturn non-memory Phi");
1811 for (j = get_irn_arity(phi) - 1; j >= 0; --j) {
1812 ir_node *pred = get_irn_n(phi, j);
1813 node_entry *pe = get_irn_ne(pred, env);
1815 if (pe->pscc != ne->pscc) {
1816 /* not in the same SCC, is region const */
1817 phi_entry *pe = phase_alloc(&env->ph, sizeof(*pe));
1821 pe->next = phi_list;
1826 /* no Phis no fun */
1827 assert(phi_list != NULL && "DFS found a loop without Phi");
1829 for (load = pscc->head; load; load = next) {
1831 node_entry *ne = get_irn_ne(load, env);
1834 if (is_Load(load)) {
1835 ldst_info_t *info = get_irn_link(load);
1836 ir_node *ptr = get_Load_ptr(load);
1838 /* for now, we cannot handle Loads with exceptions */
1839 if (info->projs[pn_Load_res] == NULL || info->projs[pn_Load_X_regular] != NULL || info->projs[pn_Load_X_except] != NULL)
1842 /* for now, we can only handle Load(Global) */
1843 if (! is_Global(ptr))
1845 ent = get_Global_entity(ptr);
1846 load_mode = get_Load_mode(load);
1847 for (other = pscc->head; other != NULL; other = next_other) {
1848 node_entry *ne = get_irn_ne(other, env);
1849 next_other = ne->next;
1851 if (is_Store(other)) {
1852 ir_alias_relation rel = get_alias_relation(
1854 get_Store_ptr(other),
1855 get_irn_mode(get_Store_value(other)),
1857 /* if the might be an alias, we cannot pass this Store */
1858 if (rel != ir_no_alias)
1861 /* only pure Calls are allowed here, so ignore them */
1863 if (other == NULL) {
1868 /* for now, we cannot handle more than one input */
1869 if (phi_list->next != NULL)
1872 /* yep, no aliasing Store found, Load can be moved */
1873 DB((dbg, LEVEL_1, " Found a Load that could be moved: %+F\n", load));
1875 db = get_irn_dbg_info(load);
1876 for (pe = phi_list; pe != NULL; pe = pe->next) {
1878 ir_node *phi = pe->phi;
1879 ir_node *blk = get_nodes_block(phi);
1880 ir_node *pred = get_Block_cfgpred_block(blk, pos);
1883 pe->load = irn = new_rd_Load(db, current_ir_graph, pred, get_Phi_pred(phi, pos), ptr, load_mode);
1884 ninfo = get_ldst_info(irn, phase_obst(&env->ph));
1886 ninfo->projs[pn_Load_M] = mem = new_r_Proj(current_ir_graph, pred, irn, mode_M, pn_Load_M);
1887 set_Phi_pred(phi, pos, mem);
1889 ninfo->projs[pn_Load_res] = new_r_Proj(current_ir_graph, pred, irn, load_mode, pn_Load_res);
1891 DB((dbg, LEVEL_1, " Created %+F in %+F\n", irn, pred));
1894 /* now kill the old Load */
1895 exchange(info->projs[pn_Load_M], get_Load_mem(load));
1896 exchange(info->projs[pn_Load_res], ninfo->projs[pn_Load_res]);
1898 env->changes |= DF_CHANGED;
1902 } /* move_loads_out_of_loops */
1905 * Process a loop SCC.
1907 * @param pscc the SCC
1908 * @param env the loop environment
1910 static void process_loop(scc *pscc, loop_env *env) {
1911 ir_node *irn, *next, *header = NULL;
1912 node_entry *b, *h = NULL;
1913 int j, only_phi, num_outside, process = 0;
1916 /* find the header block for this scc */
1917 for (irn = pscc->head; irn; irn = next) {
1918 node_entry *e = get_irn_ne(irn, env);
1919 ir_node *block = get_nodes_block(irn);
1922 b = get_irn_ne(block, env);
1925 if (h->POnum < b->POnum) {
1936 /* check if this scc contains only Phi, Loads or Stores nodes */
1940 for (irn = pscc->head; irn; irn = next) {
1941 node_entry *e = get_irn_ne(irn, env);
1944 switch (get_irn_opcode(irn)) {
1946 if (is_Call_pure(irn)) {
1947 /* pure calls can be treated like loads */
1951 /* non-pure calls must be handle like may-alias Stores */
1954 /* cannot handle CopyB yet */
1958 if (get_Load_volatility(irn) == volatility_is_volatile) {
1959 /* cannot handle loops with volatile Loads */
1965 if (get_Store_volatility(irn) == volatility_is_volatile) {
1966 /* cannot handle loops with volatile Stores */
1975 for (j = get_irn_arity(irn) - 1; j >= 0; --j) {
1976 ir_node *pred = get_irn_n(irn, j);
1977 node_entry *pe = get_irn_ne(pred, env);
1979 if (pe->pscc != e->pscc) {
1980 /* not in the same SCC, must be a region const */
1981 if (! is_rc(pred, header)) {
1982 /* not a memory loop */
1988 } else if (out_rc != pred) {
1999 /* found a memory loop */
2000 DB((dbg, LEVEL_2, " Found a memory loop:\n "));
2001 if (only_phi && num_outside == 1) {
2002 /* a phi cycle with only one real predecessor can be collapsed */
2003 DB((dbg, LEVEL_2, " Found an USELESS Phi cycle:\n "));
2005 for (irn = pscc->head; irn; irn = next) {
2006 node_entry *e = get_irn_ne(irn, env);
2009 exchange(irn, out_rc);
2011 env->changes |= DF_CHANGED;
2015 /* set the header for every node in this scc */
2016 for (irn = pscc->head; irn; irn = next) {
2017 node_entry *e = get_irn_ne(irn, env);
2020 DB((dbg, LEVEL_2, " %+F,", irn));
2022 DB((dbg, LEVEL_2, "\n"));
2024 move_loads_out_of_loops(pscc, env);
2028 } /* process_loop */
2033 * @param pscc the SCC
2034 * @param env the loop environment
2036 static void process_scc(scc *pscc, loop_env *env) {
2037 ir_node *head = pscc->head;
2038 node_entry *e = get_irn_ne(head, env);
2040 #ifdef DEBUG_libfirm
2042 ir_node *irn, *next;
2044 DB((dbg, LEVEL_4, " SCC at %p:\n ", pscc));
2045 for (irn = pscc->head; irn; irn = next) {
2046 node_entry *e = get_irn_ne(irn, env);
2050 DB((dbg, LEVEL_4, " %+F,", irn));
2052 DB((dbg, LEVEL_4, "\n"));
2056 if (e->next != NULL) {
2057 /* this SCC has more than one member */
2058 process_loop(pscc, env);
2063 * Do Tarjan's SCC algorithm and drive load/store optimization.
2065 * @param irn start at this node
2066 * @param env the loop environment
2068 static void dfs(ir_node *irn, loop_env *env)
2071 node_entry *node = get_irn_ne(irn, env);
2073 mark_irn_visited(irn);
2075 node->DFSnum = env->nextDFSnum++;
2076 node->low = node->DFSnum;
2080 if (is_Phi(irn) || is_Sync(irn)) {
2081 n = get_irn_arity(irn);
2082 for (i = 0; i < n; ++i) {
2083 ir_node *pred = get_irn_n(irn, i);
2084 node_entry *o = get_irn_ne(pred, env);
2086 if (!irn_visited(pred)) {
2088 node->low = MIN(node->low, o->low);
2090 if (o->DFSnum < node->DFSnum && o->in_stack)
2091 node->low = MIN(o->DFSnum, node->low);
2093 } else if (is_fragile_op(irn)) {
2094 ir_node *pred = get_fragile_op_mem(irn);
2095 node_entry *o = get_irn_ne(pred, env);
2097 if (!irn_visited(pred)) {
2099 node->low = MIN(node->low, o->low);
2101 if (o->DFSnum < node->DFSnum && o->in_stack)
2102 node->low = MIN(o->DFSnum, node->low);
2103 } else if (is_Proj(irn)) {
2104 ir_node *pred = get_Proj_pred(irn);
2105 node_entry *o = get_irn_ne(pred, env);
2107 if (!irn_visited(pred)) {
2109 node->low = MIN(node->low, o->low);
2111 if (o->DFSnum < node->DFSnum && o->in_stack)
2112 node->low = MIN(o->DFSnum, node->low);
2115 /* IGNORE predecessors */
2118 if (node->low == node->DFSnum) {
2119 scc *pscc = phase_alloc(&env->ph, sizeof(*pscc));
2127 e = get_irn_ne(x, env);
2129 e->next = pscc->head;
2133 process_scc(pscc, env);
2138 * Do the DFS on the memory edges a graph.
2140 * @param irg the graph to process
2141 * @param env the loop environment
2143 static void do_dfs(ir_graph *irg, loop_env *env) {
2144 ir_graph *rem = current_ir_graph;
2145 ir_node *endblk, *end;
2148 current_ir_graph = irg;
2149 inc_irg_visited(irg);
2151 /* visit all memory nodes */
2152 endblk = get_irg_end_block(irg);
2153 for (i = get_Block_n_cfgpreds(endblk) - 1; i >= 0; --i) {
2154 ir_node *pred = get_Block_cfgpred(endblk, i);
2156 pred = skip_Proj(pred);
2157 if (is_Return(pred))
2158 dfs(get_Return_mem(pred), env);
2159 else if (is_Raise(pred))
2160 dfs(get_Raise_mem(pred), env);
2161 else if (is_fragile_op(pred))
2162 dfs(get_fragile_op_mem(pred), env);
2164 assert(0 && "Unknown EndBlock predecessor");
2168 /* visit the keep-alives */
2169 end = get_irg_end(irg);
2170 for (i = get_End_n_keepalives(end) - 1; i >= 0; --i) {
2171 ir_node *ka = get_End_keepalive(end, i);
2173 if (is_Phi(ka) && !irn_visited(ka))
2176 current_ir_graph = rem;
2180 * Initialize new phase data. We do this always explicit, so return NULL here
2182 static void *init_loop_data(ir_phase *ph, const ir_node *irn, void *data) {
2187 } /* init_loop_data */
2190 * Optimize Loads/Stores in loops.
2192 * @param irg the graph
2194 static int optimize_loops(ir_graph *irg) {
2197 env.stack = NEW_ARR_F(ir_node *, 128);
2202 phase_init(&env.ph, "ldstopt", irg, PHASE_DEFAULT_GROWTH, init_loop_data, NULL);
2204 /* calculate the SCC's and drive loop optimization. */
2207 DEL_ARR_F(env.stack);
2208 phase_free(&env.ph);
2211 } /* optimize_loops */
2214 * do the load store optimization
2216 int optimize_load_store(ir_graph *irg) {
2219 FIRM_DBG_REGISTER(dbg, "firm.opt.ldstopt");
2221 assert(get_irg_phase_state(irg) != phase_building);
2222 assert(get_irg_pinned(irg) != op_pin_state_floats &&
2223 "LoadStore optimization needs pinned graph");
2225 /* we need landing pads */
2226 remove_critical_cf_edges(irg);
2230 /* for Phi optimization post-dominators are needed ... */
2231 assure_postdoms(irg);
2233 if (get_opt_alias_analysis()) {
2234 assure_irg_entity_usage_computed(irg);
2235 assure_irp_globals_entity_usage_computed();
2238 obstack_init(&env.obst);
2241 /* init the links, then collect Loads/Stores/Proj's in lists */
2243 irg_walk_graph(irg, firm_clear_link, collect_nodes, &env);
2245 /* now we have collected enough information, optimize */
2246 irg_walk_graph(irg, NULL, do_load_store_optimize, &env);
2248 env.changes |= optimize_loops(irg);
2250 obstack_free(&env.obst, NULL);
2252 /* Handle graph state */
2254 set_irg_outs_inconsistent(irg);
2255 set_irg_entity_usage_state(irg, ir_entity_usage_not_computed);
2258 if (env.changes & CF_CHANGED) {
2259 /* is this really needed: Yes, control flow changed, block might
2260 have Bad() predecessors. */
2261 set_irg_doms_inconsistent(irg);
2263 return (int) env.changes;
2264 } /* optimize_load_store */