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 scalar replacement of arrays and compounds
23 * @author Beyhan Veliev, Michael Beck
28 #include "iroptimize.h"
42 #include "analyze_irg_args.h"
47 #define SET_ENT_VNUM(ent, vnum) set_entity_link(ent, INT_TO_PTR(vnum))
48 #define GET_ENT_VNUM(ent) (unsigned)PTR_TO_INT(get_entity_link(ent))
49 #define SET_IRN_VNUM(irn, vnum) set_irn_link(irn, INT_TO_PTR(vnum))
50 #define GET_IRN_VNUM(irn) (unsigned)PTR_TO_INT(get_irn_link(irn))
54 typedef struct _ent_leaves_t{
55 ir_entity *ent; /**< An entity, that contains scalars for replace.*/
56 pset *leaves; /**< All leaves of this entity.*/
59 typedef struct _sels_t {
60 ir_node *sel; /**< A sel node, thats entity have scalars.*/
61 ir_entity *ent; /**< The entity of this sel node.*/
64 typedef struct _call_access_t {
65 ir_node *call; /**< A call node, that have as parameter a scalar.*/
66 unsigned int access_type; /**< The access type, with that this call access this scalar.*/
69 typedef struct _fixlist_entry_t {
70 ir_node *irn; /**< An ir node, that must be fixed.*/
71 unsigned int vnum; /**< The value number, that must became this ir node.*/
74 typedef struct _syncs_fixlist_entry_t {
75 ir_node *irn; /**< A sync node that must be fixed.*/
76 int *accessed_vnum; /**< A pointer to save an array with value numbers, that must became this sync.*/
77 }syncs_fixlist_entry_t;
79 /* A entry, that save the memory
80 * edge state and the access state for this leave
81 * int the array,that is created for every block.*/
82 typedef struct _leave_t {
83 ir_node *mem_edge_state; /**< memory state for this scalar in this block.*/
84 unsigned int access_type; /**< access state for this scalar in this block.*/
85 set *calls; /**< call nodes,that change this scalar in this block.*/
89 * A path element entry: it is either an entity
90 * or a tarval, because we evaluate only constant array
91 * accesses like a.b.c[8].d
99 * An access path, used to assign value numbers
100 * to variables that will be scalar replaced
102 typedef struct _path_t {
103 unsigned vnum; /**< the value number */
104 unsigned path_len; /**< the length of the access path */
105 path_elem_t path[1]; /**< the path */
109 * environment for memory walker
111 typedef struct _env_t {
112 struct obstack obst; /**< a obstack for the memory edge */
113 set *set_sels; /**< a set with all sels, that are reachable from an entity with a scalar.*/
114 set *set_ent; /**< a set with all entities that have one or more scalars.*/
115 fixlist_entry_t *fix_phis; /**< list of all Phi nodes that must be fixed */
116 fixlist_entry_t *fix_ls; /**< list of all Load or Store nodes that must be fixed */
117 syncs_fixlist_entry_t *fix_syncs; /**< list of all Sync nodes that must be fixed */
118 unsigned int nvals; /**< to save the number of scalars.*/
119 unsigned int gl_mem_vnum; /**< indicate the position of the globule memory edge state in var_arr.*/
120 unsigned int vnum_state; /**< indicate the position of the value number state in var_arr.*/
121 unsigned int changes; /**< to save if by anlyse_calls is changed anything.*/
127 * Compare two elements of the ent_leaves_t set.
129 * @return 0 if they are identically
131 static int ent_leaves_t_cmp(const void *elt, const void *key, size_t size)
133 const ent_leaves_t *c1 = elt;
134 const ent_leaves_t *c2 = key;
137 return c1->ent != c2->ent;
141 * Compare two elements of the ent_access_t set.
143 * @return 0 if they are identically
145 static int ent_cmp(const void *elt, const void *key)
147 const ir_entity *c1 = elt;
148 const ir_entity *c2 = key;
154 * Compare two elements of the sels_t set.
156 * @return 0 if they are identically
158 static int sels_cmp(const void *elt, const void *key, size_t size)
160 const sels_t *c1 = elt;
161 const sels_t *c2 = key;
164 return c1->sel != c2->sel;
168 * Compare two elements of the leave_t set.
170 * @return 0 if they are identically
172 static int leave_cmp(const void *elt, const void *key)
174 ir_node *c1 = (ir_node *)elt;
175 ir_node *c2 = (ir_node *)key;
177 return get_Sel_entity(c1) != get_Sel_entity(c2);
181 * Compare two elements of the call_access_t set.
183 * @return 0 if they are identically
185 static int call_cmp(const void *elt, const void *key, size_t size)
187 const call_access_t *c1 = elt;
188 const call_access_t *c2 = key;
191 return c1->call != c2->call;
197 * @return 0 if they are identically
199 static int path_cmp(const void *elt, const void *key, size_t size)
201 const path_t *p1 = elt;
202 const path_t *p2 = key;
205 /* we can use memcmp here, because identical tarvals should have identical addresses */
206 return memcmp(p1->path, p2->path, p1->path_len * sizeof(p1->path[0]));
210 * Calculate a hash value for a path.
212 static unsigned path_hash(const path_t *path)
217 for (i = 0; i < path->path_len; ++i)
218 hash ^= (unsigned)PTR_TO_INT(path->path[i].ent);
224 * Returns non-zero, if all induces of a Sel node are constants.
226 * @param sel the Sel node that will be checked
228 static int is_const_sel(ir_node *sel) {
229 int i, n = get_Sel_n_indexs(sel);
231 for (i = 0; i < n; ++i) {
232 ir_node *idx = get_Sel_index(sel, i);
241 * Returns non-zero, if the address of an entity
242 * represented by a Sel node (or it's successor Sels) is taken.
244 static int is_address_taken_2(ir_node *sel)
248 if (! is_const_sel(sel))
251 for (i = get_irn_n_outs(sel) - 1; i >= 0; --i) {
252 ir_node *succ = get_irn_out(sel, i);
254 switch (get_irn_opcode(succ)) {
256 /* ok, we just load from that entity */
260 /* check that Sel is not the Store's value */
261 if (get_Store_value(succ) == sel)
266 /* Check the Sel successor of Sel */
267 int res = is_address_taken_2(succ);
275 /* The address of an entity is given as a parameter.
276 * We analyzes that later and optimizes this scalar
282 /* another op, the address is taken */
290 * Link all Sels with the entity.
292 * @param ent the entity that will be scalar replaced
293 * @param sel a Sel node that selects some fields of this entity
295 static void link_all_leave_sels(ir_entity *ent, ir_node *sel)
299 n = get_irn_n_outs(sel);
300 for (i = 0; i < n; ++i) {
301 ir_node *succ = get_irn_out(sel, i);
304 link_all_leave_sels(ent, succ);
307 /* if Sel nodes with memory inputs are used, a entity can be
308 * visited more than once causing a ring here, so we use the
309 * node flag to mark linked nodes
311 if (irn_visited_else_mark(sel))
315 * we link the sels to the entity.
317 set_irn_link(sel, get_entity_link(ent));
318 set_entity_link(ent, sel);
321 /* we need a special address that serves as an address taken marker */
323 static void *ADDRESS_TAKEN = &_x;
326 * Find possible scalar replacements.
328 * @param irg an IR graph
330 * This function finds variables on the (members of the) frame type
331 * that can be scalar replaced, because their address is never taken.
332 * If such a variable is found, it's entity link will hold a list of all
333 * Sel nodes, that selects anythings of this entity.
334 * Otherwise, the link will be ADDRESS_TAKEN or NULL.
336 * @return non-zero if at least one entity could be replaced
339 static int find_possible_replacements(ir_graph *irg)
341 ir_node *irg_frame = get_irg_frame(irg);
345 inc_irg_visited(irg);
347 n = get_irn_n_outs(irg_frame);
350 * First, clear the link field of all interestingentities.
351 * Note that we did not rely on the fact that there is only
352 * one Sel node per entity, so we might access one entity
353 * more than once here.
354 * That's why we have need two loops.
356 for (i = 0; i < n; ++i) {
357 ir_node *succ = get_irn_out(irg_frame, i);
360 ir_entity *ent = get_Sel_entity(succ);
361 set_entity_link(ent, NULL);
366 * Check the ir_graph for Sel nodes. If the entity of Sel
367 * isn't a scalar replacement set the link of this entity
368 * equal ADDRESS_TAKEN.
370 for (i = 0; i < n; ++i) {
371 ir_node *succ = get_irn_out(irg_frame, i);
374 ir_entity *ent = get_Sel_entity(succ);
377 if (get_entity_link(ent) == ADDRESS_TAKEN)
381 * Beware: in rare cases even entities on the frame might be
382 * volatile. This might happen if the entity serves as a store
383 * to a value that must survive a exception. Do not optimize
384 * such entities away.
386 if (get_entity_volatility(ent) == volatility_is_volatile) {
387 set_entity_link(ent, ADDRESS_TAKEN);
391 ent_type = get_entity_type(ent);
393 /* we can handle arrays, structs and atomic types yet */
394 if (is_Array_type(ent_type) || is_Struct_type(ent_type) || is_atomic_type(ent_type)) {
395 if (is_address_taken_2(succ)) {
396 if (get_entity_link(ent)) /* killing one */
398 set_entity_link(ent, ADDRESS_TAKEN);
401 /* possible found one */
402 if (get_entity_link(ent) == NULL)
404 link_all_leave_sels(ent, succ);
413 static int is_leave_sel(ir_node *sel) {
417 for(i = get_irn_n_outs(sel) - 1; i >= 0; i--) {
418 succ = get_irn_out(sel, i);
427 * Return a path from the Sel node sel to it's root.
429 * @param sel the Sel node
430 * @param len the length of the path so far
432 static path_t *find_path(ir_node *sel, unsigned len)
436 ir_node *pred = get_Sel_ptr(sel);
438 /* the current Sel node will add some path elements */
439 n = get_Sel_n_indexs(sel);
443 /* we found the root */
444 res = XMALLOCF(path_t, path, len);
448 res = find_path(pred, len);
450 pos = res->path_len - len;
452 res->path[pos++].ent = get_Sel_entity(sel);
453 for (i = 0; i < n; ++i) {
454 ir_node *index = get_Sel_index(sel, i);
457 res->path[pos++].tv = get_Const_tarval(index);
463 * Allocate value numbers for the leaves
464 * in our found entities.
466 * @param sels a set that will contain all Sels that have a value number
467 * @param ent the entity that will be scalar replaced
468 * @param vnum the first value number we can assign
469 * @param modes a flexible array, containing all the modes of
472 * @return the next free value number
474 static unsigned allocate_value_numbers(set *set_sels, pset *leaves, ir_entity *ent, unsigned vnum)
479 set *pathes = new_set(path_cmp, 8);
481 /* visit all Sel nodes in the chain of the entity */
482 for (sel = get_entity_link(ent); sel; sel = next) {
483 next = get_irn_link(sel);
485 /* we save for every sel it root entity, why
486 * we need this information, when we split the memory edge,
487 * and we must mark this sel for later. */
490 set_insert(set_sels, &key_sels, sizeof(key_sels), HASH_PTR(sel));
492 if(! is_leave_sel(sel))
494 /* We have found a leave and we add it to the pset of this entity.*/
495 pset_insert(leaves, sel, HASH_PTR(get_Sel_entity(sel)));
497 key = find_path(sel, 0);
498 path = set_find(pathes, key, sizeof(*key) + sizeof(key->path[0]) * key->path_len, path_hash(key));
501 SET_IRN_VNUM(sel, path->vnum);
506 set_insert(pathes, key, sizeof(*key) + sizeof(key->path[0]) * key->path_len, path_hash(key));
508 SET_IRN_VNUM(sel, key->vnum);
514 set_entity_link(ent, NULL);
518 * Add a sync node to it fix list.
520 * @param sync The sync node, that myst be addet to the fix list.
521 * @param unk_vnum An array whit the value number, that are synced with this sync node.
522 * @param env The enviroment pinter.
524 static void add_sync_to_fixlist(ir_node *sync, int *unk_vnum, env_t *env) {
526 syncs_fixlist_entry_t *s;
528 s = obstack_alloc(&env->obst, sizeof(*s));
530 s->accessed_vnum = unk_vnum;
531 set_irn_link(sync, env->fix_syncs);
535 * Add a ir node to it fix list.
537 * @param irn The ir node, that myst be addet to the fix list.
538 * @param vnum The value number, that must baceme this ir node as predecessor later.
539 * @param env The enviroment pinter.
541 static void add_ls_to_fixlist(ir_node *irn, int vnum, env_t *env) {
545 l = obstack_alloc(&env->obst, sizeof(*l));
549 if(get_irn_op(irn) == op_Phi) {
550 set_irn_link(l->irn, env->fix_phis);
553 set_irn_link(l->irn, env->fix_ls);
558 static void add_mem_edge(value_arr_entry_t *val_arr, int vnum, ir_node ***in, int **accessed_vnum) {
560 if(val_arr[vnum].mem_edge_state != NULL)
561 ARR_APP1(ir_node *, *in, val_arr[vnum].mem_edge_state);
563 ARR_APP1(int, *accessed_vnum, vnum);
564 ARR_APP1(ir_node *, *in, new_Unknown(mode_M));
568 * The function handles the scalars, that wase stored
571 * @param blk The block, that must be handled.
572 * @param env The enviroment pinter.
575 /* Return the memory successor of the call node.*/
576 static ir_node *get_Call_mem_out(ir_node *call) {
581 for(i = get_irn_n_outs(call) - 1; i >= 0; i--) {
582 mem = get_irn_out(call, i);
583 if(get_irn_mode(mem) == mode_M)
586 /* is not reachable*/
591 static void sync_stored_scalars(ir_node *blk, env_t *env) {
594 int *unk_vnum; /**< An arraw, where are saved the value number, that
595 are synced from this sync node.*/
596 ent_leaves_t *value_ent;
597 value_arr_entry_t *val_arr_blk, *val_arr;
598 ir_node *pred, *leave, *sync, **in;
599 ir_node *sync_blk; /**< The block, where the sync node must be created.*/
602 val_arr_blk = get_irn_link(blk);
604 for(value_ent = set_first(env->set_ent); value_ent; value_ent = set_next(env->set_ent)) {
607 if(val_arr_blk[GET_ENT_VNUM(value_ent->ent)].access_type <= 3)
608 /* This entity is not stored in this block.*/
611 for(i = get_Block_n_cfgpreds(blk) - 1; i >= 0; i--) {
613 pred = get_Block_cfgpred(blk, i);
614 pred = get_nodes_block(pred);
615 val_arr = get_irn_link(pred);
617 if(val_arr[GET_ENT_VNUM(value_ent->ent)].access_type == SYNCED)
618 /* This entity was synced.*/
621 if(val_arr[GET_ENT_VNUM(value_ent->ent)].access_type <= 3) {
623 /* To avoid repeated sync of this entity in this block.*/
624 val_arr[GET_ENT_VNUM(value_ent->ent)].access_type = SYNCED;
625 /* In this predecessor block is this entity not acessed.
626 * We must sync in the end ot this block.*/
627 if(get_Block_n_cfgpreds(blk) > 1)
628 sync_blk = get_nodes_block(get_Block_cfgpred(blk, i));
632 val_arr = get_irn_link(sync_blk);
633 /* An array to save the memory edges, that must be
635 in = NEW_ARR_F(ir_node *, 1);
637 /* An array to save the value numbers,
638 * that must be repaired.*/
639 unk_vnum = NEW_ARR_F(int, 0);
640 /* The global memory edge.*/
641 if(val_arr[env->gl_mem_vnum].mem_edge_state == NULL)
642 in[0] = new_Unknown(mode_M);
644 in[0] = val_arr[env->gl_mem_vnum].mem_edge_state;
646 for(leave = pset_first(value_ent->leaves); leave; leave = pset_next(value_ent->leaves))
647 /* All this memory edges must be synced.*/
648 add_mem_edge(val_arr, GET_IRN_VNUM(leave), &in, &unk_vnum);
650 /* We create the sync and set it in the global memory state.*/
651 sync = new_r_Sync(current_ir_graph, sync_blk, ARR_LEN(in), in);
652 /* We must check this, why it is possible to get a Bad node
653 * form new_r_Sync(), when the node can be optimized.
654 * In this case we must do nothing.*/
656 val_arr[env->gl_mem_vnum].mem_edge_state = sync;
657 /* We add this sync node to the sync's fix list.*/
658 add_sync_to_fixlist(val_arr[env->gl_mem_vnum].mem_edge_state, unk_vnum, env);
666 * The function split the memory edge of load and store nodes, that have
667 * as predecessor a scalar
669 * @param irn The node, that memory edge must be spleted.
670 * @param env The enviroment pinter.
672 static void split_ls_mem_edge(ir_node *irn, env_t *env) {
675 ir_node *leave, *irn_blk, *mem_state, *new_mem_state;
676 unsigned ent_vnum, sel_vnum, i;
677 value_arr_entry_t *val_arr;
678 sels_t key_sels, *value_sels;
679 ent_leaves_t key_ent, *value_ent;
681 op = get_irn_op(irn);
684 key_sels.sel = get_Load_ptr(irn);
686 key_sels.sel = get_Store_ptr(irn);
688 value_sels = set_find(env->set_sels, &key_sels, sizeof(key_sels), HASH_PTR(key_sels.sel));
690 if(value_sels != NULL) {
691 /* we have found a load or store, that use a sel of our set
692 * and we must split or extend, if the memory edge have been
693 * split for this sel, the memory edge.*/
695 key_ent.ent = value_sels->ent;
696 value_ent = set_find(env->set_ent, &key_ent, sizeof(key_ent), HASH_PTR(key_ent.ent));
697 /*To check if the enities set is right filled. */
698 assert(value_ent && " This sel's entity isn't int the entity set.");
700 leave = pset_find(value_ent->leaves, key_sels.sel, HASH_PTR(get_Sel_entity(key_sels.sel)));
701 /*To check if the leaves set is right filled. */
702 assert(leave && "Anything in data_flow_scalar_replacment algorithm is wrong.");
704 ent_vnum = GET_ENT_VNUM(value_ent->ent);
705 sel_vnum = GET_IRN_VNUM(leave);
706 irn_blk = get_nodes_block(irn);
707 val_arr = get_irn_link(irn_blk);
709 if(val_arr[ent_vnum].access_type == 0)
710 /* We have found a scalar, that address is not stored as jet.*/
713 /* This scalar have been stored.*/
714 i = env->gl_mem_vnum;
716 if(val_arr[i].mem_edge_state == NULL) {
717 /* We split now for this sel the memory edge in this block.*/
718 mem_state = new_Unknown(mode_M);
719 /* We must mark this node to fix later*/
720 add_ls_to_fixlist(irn, i, env);
723 /* We have split the memory edge and the current state is saved.*/
724 mem_state = val_arr[i].mem_edge_state;
726 /* We set this Load or Store to the memory edge of this
729 set_Load_mem(irn, mem_state);
731 set_Store_mem(irn, mem_state);
733 /* When we have split or extended the memory edge we must
734 * update the memory_edge_state of this sel*/
735 new_mem_state = get_irn_out(irn, 0);
736 if(get_irn_mode(new_mem_state) == mode_M)
737 val_arr[i].mem_edge_state = new_mem_state;
739 val_arr[i].mem_edge_state = get_irn_out(irn, 1);
744 * The function split the memory edge of phi nodes, that have
745 * as predecessor a scalar
747 * @param irn The phi node, that memory edge must be spleted.
748 * @param env The enviroment pinter.
750 static void split_phi_mem_edge(ir_node *irn, env_t *env) {
752 ir_node *irn_blk, *unk, *leave, **in;
754 ent_leaves_t *value_ent;
755 value_arr_entry_t *val_arr;
757 irn_blk = get_nodes_block(irn);
758 val_arr = get_irn_link(irn_blk);
760 n = get_Block_n_cfgpreds(irn_blk);
761 in = ALLOCAN(ir_node*, n);
763 for(value_ent = set_first(env->set_ent); value_ent; value_ent = set_next(env->set_ent))
764 if(val_arr[GET_ENT_VNUM(value_ent->ent)].access_type < 3)
765 /* This scalar wasn't be saved and we need to produce a phi for it.*/
766 for(leave = pset_first(value_ent->leaves); leave; leave = pset_next(value_ent->leaves)){
768 unk = new_Unknown(mode_M);
769 for (j = n - 1; j >= 0; --j)
772 val_arr[GET_IRN_VNUM(leave)].mem_edge_state = new_r_Phi(current_ir_graph, irn_blk, n, in, mode_M);
774 add_ls_to_fixlist(val_arr[GET_IRN_VNUM(leave)].mem_edge_state, GET_IRN_VNUM(leave), env);
777 /* We use for the global memory the phi node, that
778 * is already available.*/
779 val_arr[env->gl_mem_vnum].mem_edge_state = irn;
783 * The function handles the call nodes, that have
784 * as parameter a scalar
786 * @param env The enviroment pinter.
787 * @param call The call node, that must be handled.
788 * @param accessed_entities A set wit all entities, that are accessed from this call node.*/
789 static void split_call_mem_edge(env_t *env, ir_node *call, pset *accessed_entities) {
791 ent_leaves_t key_ent, *value_ent;
792 value_arr_entry_t *val_arr;
793 call_access_t key_call, *value_call;
794 ir_node *call_blk, *new_mem_state, *leave;
798 int fix_irn = 0; /**< Set to 1 if we must add this call to it fix list.*/
799 int *accessed_leaves_vnum = NULL; /**< An arraw, where are saved the value number, that
800 are synced from call's sync node, if we need it.*/
802 call_blk = get_nodes_block(call);
803 val_arr = get_irn_link(call_blk);
804 /* An array to save the memory edges, that must be
806 in = NEW_ARR_F(ir_node *, 1);
807 /* An array to save the value numbers of the memory
808 * edges that must be repaired.*/
809 accessed_leaves_vnum = NEW_ARR_F(int, 0);
811 /* We get the memory successor of the call node.
812 * It is the new memory state for all synced memory
814 new_mem_state = get_Call_mem_out(call);
816 /* The global memory is the first predecessor of the create sync node.*/
817 if(val_arr[env->gl_mem_vnum].mem_edge_state == NULL) {
818 in[0] = new_Unknown(mode_M);
822 in[0] = val_arr[env->gl_mem_vnum].mem_edge_state;
825 for(ent = pset_first(accessed_entities); ent; ent = pset_next(accessed_entities)) {
826 /* Whit this loop we iterate all accessed entities from this call and collect
827 * all memory edges, that we must sync.*/
828 ent_vnum = GET_ENT_VNUM(ent);
830 key_call.call = call;
831 value_call = set_find(val_arr[ent_vnum].calls, &key_call, sizeof(key_call), HASH_PTR(key_call.call));
834 value_ent = set_find(env->set_ent, &key_ent, sizeof(key_ent), HASH_PTR(key_ent.ent));
836 if(val_arr[ent_vnum].access_type <= 3) {
837 /* This scalar's address wasn't stored in this block.*/
838 switch(value_call->access_type) {
840 case ptr_access_none :
841 /* In this case we have nothing to do.*/
844 case ptr_access_read:
845 case ptr_access_write:
847 /* All this cases must be traded equal.*/
849 for(leave = pset_first(value_ent->leaves); leave; leave = pset_next(value_ent->leaves)){
850 /* All this memory edges must be synced.*/
851 add_mem_edge(val_arr, GET_IRN_VNUM(leave), &in, &accessed_leaves_vnum);
853 /* We update the memory state of this leave.*/
854 if(value_call->access_type != ptr_access_read)
855 val_arr[GET_IRN_VNUM(leave)].mem_edge_state = new_mem_state;
864 /* We must update the global memory state.*/
865 val_arr[env->gl_mem_vnum].mem_edge_state = new_mem_state;
867 if(ARR_LEN(in) == 1) {
868 /* we must set the call memory to gobale momory*/
869 set_Call_mem(call,in[0]);
872 /* We add this call node to the call fix list..*/
873 add_ls_to_fixlist(call, env->gl_mem_vnum, env);
876 /* We create the sync and set it as memory predecessor of the call node.*/
877 sync = new_r_Sync(current_ir_graph, call_blk, ARR_LEN(in), in);
878 /* We must check this, why it is possible to get a Bad node
879 * form new_r_Sync(), when the node can be optimized.
880 * In this case we must do nothing.*/
882 set_Call_mem(call, sync);
883 if(ARR_LEN(accessed_leaves_vnum))
884 /* We add this sync node to the sync's fix list.*/
885 add_sync_to_fixlist(sync, accessed_leaves_vnum, env);
892 * The function split the memory edge from the passed
893 * ir node if this is needed
895 * @param irn The node, that memory edge must be spleted.
896 * @param env The enviroment pinter.
898 static void split_memory_edge(ir_node *irn, void *ctx) {
901 ir_node *sel, *irn_blk;
903 sels_t key_sels, *value_sels;
904 value_arr_entry_t *val_arr;
905 pset *accessed_entities; /**< A set to save all entities accessed from a call.*/
909 op = get_irn_op(irn);
914 irn_blk = get_nodes_block(irn);
916 if (!Block_block_visited(irn_blk)) {
917 /* We sync first the stored scalar address in this block.*/
918 mark_Block_block_visited(irn_blk);
919 sync_stored_scalars(irn_blk, env);
922 if(op == op_Load || op == op_Store)
924 split_ls_mem_edge(irn, env);
927 if (op == op_Phi && get_irn_mode(irn) == mode_M) {
929 * found a memory Phi: Here, we must create new Phi nodes
931 split_phi_mem_edge(irn, env);
936 /* Calls that have a NoMem input do neither read nor write memory.
937 We can completely ignore them here. */
938 if (is_NoMem(get_Call_mem(irn)))
941 /* We save in this set all entities,
942 * that are accessed from this call node.*/
943 accessed_entities = new_pset(ent_cmp, 8);
944 val_arr = get_irn_link(get_nodes_block(irn));
946 for ( i = get_Call_n_params(irn) - 1; i >= 0; i--) {
948 sel = get_Call_param(irn, i);
952 value_sels = set_find(env->set_sels, &key_sels, sizeof(key_sels), HASH_PTR(key_sels.sel));
954 if(value_sels != NULL && val_arr[GET_ENT_VNUM(value_sels->ent)].access_type <= 3)
955 /* We save in this set all accessed entities from this call node whit
956 * access none, read, write or rw..*/
957 pset_insert(accessed_entities, value_sels->ent, HASH_PTR(value_sels->ent));
961 if(pset_count(accessed_entities))
962 split_call_mem_edge(env, irn, accessed_entities);
964 del_pset(accessed_entities);
971 * searches through blocks beginning from block for value
972 * vnum and return it.
974 * @param block A block from the current ir graph.
975 * @param vnum The value number, that must be found.
977 static ir_node *find_vnum_value(ir_node *block, unsigned vnum)
979 value_arr_entry_t *val_arr;
983 if (!Block_block_visited(block)) {
984 mark_Block_block_visited(block);
986 val_arr = get_irn_link(block);
988 if (val_arr[vnum].mem_edge_state)
989 return val_arr[vnum].mem_edge_state;
991 for (i = get_Block_n_cfgpreds(block) - 1; i >= 0; --i) {
992 ir_node *pred = get_Block_cfgpred(block, i);
994 res = find_vnum_value(get_nodes_block(pred), vnum);
1003 * fix the Load/Store or Call list
1005 * @param The enviroment pinter.
1007 static void fix_ls(env_t *env)
1010 ir_node *irn, *block, *pred, *val = NULL;
1014 for (l = env->fix_ls; l; l = get_irn_link(irn)) {
1017 op = get_irn_op(irn);
1018 block = get_nodes_block(irn);
1019 for (i = get_Block_n_cfgpreds(block) - 1; i >= 0; --i) {
1020 pred = get_Block_cfgpred(block, i);
1021 pred = get_nodes_block(pred);
1023 inc_irg_block_visited(current_ir_graph);
1024 val = find_vnum_value(pred, l->vnum);
1032 set_Store_mem(irn, val);
1035 set_Load_mem(irn, val);
1037 set_Call_mem(irn, val);
1045 * @param The enviroment pinter.
1047 static void fix_phis(env_t *env)
1050 ir_node *phi, *block, *pred, *val;
1053 for (l = env->fix_phis; l; l = get_irn_link(phi)) {
1056 block = get_nodes_block(phi);
1057 for (i = get_Block_n_cfgpreds(block) - 1; i >= 0; --i) {
1059 pred = get_Block_cfgpred(block, i);
1060 pred = get_nodes_block(pred);
1062 inc_irg_block_visited(current_ir_graph);
1063 val = find_vnum_value(pred, l->vnum);
1066 set_irn_n(phi, i, val);
1075 * @param The enviroment pinter.
1077 static void fix_syncs(env_t *env)
1079 syncs_fixlist_entry_t *l;
1080 ir_node *sync, *block, *pred, *val;
1084 for (l = env->fix_syncs; l; l = get_irn_link(sync)) {
1088 /* The sync block must have one predecessor, when it
1089 have unknown nodes as predecessor.*/
1090 block = get_nodes_block(sync);
1091 pred = get_Block_cfgpred(block, 0);
1092 pred = get_nodes_block(pred);
1094 /* We first repair the global memory edge at the first position of sync predecessors.*/
1095 if (is_Unknown(get_irn_n(sync, 0))) {
1096 inc_irg_block_visited(current_ir_graph);
1097 val = find_vnum_value(pred, env->gl_mem_vnum);
1100 set_irn_n(sync, 0, val);
1103 for (i = get_irn_arity(sync) - 1; i >= 1; --i) {
1104 /* We repair the leaves*/
1106 assert(k <= ARR_LEN(l->accessed_vnum) && "The algorythm for sync repair is wron");
1107 if (is_Unknown(get_irn_n(sync, i))) {
1108 inc_irg_block_visited(current_ir_graph);
1109 val = find_vnum_value(pred, l->accessed_vnum[k++]);
1112 set_irn_n(sync, i, val);
1115 DEL_ARR_F(l->accessed_vnum);
1119 * For the end node we must sync all memory edges.
1121 * @param The enviroment pinter.
1123 static void sync_mem_edges(env_t *env) {
1125 value_arr_entry_t *val_arr;
1126 ir_node **in, *sync, *Return, *Return_blk;
1127 int i, vnum, vnum_state;
1129 Return = get_Block_cfgpred(get_irg_end_block(current_ir_graph), 0);
1130 Return_blk = get_nodes_block(Return);
1131 val_arr = get_irn_link(Return_blk);
1135 for(i = 0; i <= (int)env->gl_mem_vnum; i++)
1136 /* we get the current state of non saved scalars.*/
1137 if(val_arr[i].access_type <= 3)
1140 /* We allocate the memory, that we need for the predecessors of the sync.*/
1141 in = XMALLOCN(ir_node*, vnum_state);
1143 /* The global memory edge is the first predecessor of this sync node.*/
1144 if(val_arr[env->gl_mem_vnum].mem_edge_state == NULL) {
1145 /* We must search through blocks for this memory state.*/
1146 inc_irg_block_visited(current_ir_graph);
1147 in[0] = find_vnum_value(Return_blk, env->gl_mem_vnum);
1150 in[0] = val_arr[env->gl_mem_vnum].mem_edge_state;
1153 for(i = 1, vnum = 0; vnum < (int)env->gl_mem_vnum; vnum++) {
1155 if(val_arr[vnum].access_type <= 3) {
1156 /* we add the non saved scalars as predecessors of the sync.*/
1158 if(val_arr[vnum].mem_edge_state == NULL) {
1159 /* We must search through blocks for this memory state.*/
1160 inc_irg_block_visited(current_ir_graph);
1161 in[i] = find_vnum_value(Return_blk, vnum);
1164 in[i] = val_arr[vnum].mem_edge_state;
1169 sync = new_r_Sync(current_ir_graph, Return_blk, vnum_state, in);
1170 set_Return_mem(Return, sync);
1176 * Walker: allocate the value array for every block.
1178 * @param block A block from the current ir graph for that must be allocated a value array.
1179 * @param ctx The enviroment pinter.
1181 static void alloc_value_arr(ir_node *block, void *ctx)
1186 value_arr_entry_t *var_arr = obstack_alloc(&env->obst, sizeof(value_arr_entry_t) *(env->nvals + set_count(env->set_ent) + 1));
1188 /* the value array is empty at start */
1189 memset(var_arr, 0, sizeof(value_arr_entry_t) * (env->nvals + set_count(env->set_ent) + 1));
1190 set_irn_link(block, var_arr);
1192 /* We set the block value number state to optimal and later we update this.*/
1193 var_arr[env->vnum_state].access_type = env->nvals;
1195 if(get_irg_start_block(current_ir_graph) == block)
1196 /* We initilize the startblocks array with the irg initilize memory, why
1197 * it must be the start point of all memory edges.*/
1198 for(i = (env->nvals + set_count(env->set_ent)) ; i >=0; i--)
1199 var_arr[i].mem_edge_state = get_irg_initial_mem(current_ir_graph);
1203 /* Analyze call nodes to get information, if they store the address of a scalar.
1205 * @param *irn An ir node from the current_ir_graph.
1206 * @param *env The enviroment pointer.
1208 static void analyse_calls(ir_node *irn, void *ctx) {
1211 unsigned int acces_type;
1212 ir_node *param, *call_ptr, *blk;
1213 ir_entity *meth_ent;
1214 sels_t key_sels, *value_sels;
1215 call_access_t key_call, *value_call;
1216 value_arr_entry_t *val_arr;
1223 /* Calls that have a NoMem input do neither read nor write memory.
1224 We can completely ignore them here. */
1225 if (is_NoMem(get_Call_mem(irn)))
1228 /* We iterate over the parameters of this call nodes.*/
1229 for ( i = get_Call_n_params(irn) - 1; i >= 0; i--) {
1230 param = get_Call_param(irn, i);
1231 if (is_Sel(param)) {
1232 /* We have found a parameter with operation sel.*/
1233 key_sels.sel = param;
1234 value_sels = set_find(env->set_sels, &key_sels, sizeof(key_sels), HASH_PTR(key_sels.sel));
1235 if(value_sels != NULL ) {
1237 /* We have found a call, that have as parameter a sel from our set_sels.*/
1238 call_ptr = get_Call_ptr(irn);
1240 if (is_SymConst(call_ptr) && get_SymConst_kind(call_ptr) == symconst_addr_ent) {
1241 meth_ent = get_SymConst_entity(call_ptr);
1242 /* we get the access type for our sel.*/
1243 acces_type = get_method_param_access(meth_ent, i);
1245 /* We can't analyze this function and we asume, that it store the address.*/
1246 acces_type = ptr_access_store;
1248 /* we save the access type and this call in the array allocated for this block.
1249 * The value number of this entity get us the position in the array to save this
1250 * information. Why we expect more calls as one we allocate a set.*/
1251 vnum = GET_ENT_VNUM(value_sels->ent);
1252 blk = get_nodes_block(irn);
1253 val_arr = get_irn_link(blk);
1255 if(val_arr[vnum].access_type > 3)
1256 /* The address of this entity have been stored.*/
1259 if(val_arr[vnum].calls == NULL)
1260 /* for this entity i have found the firs call in this block and we must allocate the set.*/
1261 val_arr[vnum].calls = new_set(call_cmp, 8);
1263 /* This call performs anything with the scalar and we must mark it.*/
1264 key_call.call = irn;
1265 key_call.access_type = acces_type;
1266 value_call = set_insert(val_arr[vnum].calls, &key_call, sizeof(key_call), HASH_PTR(key_call.call));
1268 if(value_call->access_type < acces_type)
1269 /* this case tread, when a call access an entity more at once.
1270 * Than we must save the highest access type.*/
1271 value_call->access_type = acces_type;
1274 /* This call save the address of our scalar and we can't
1275 * use the scalars of this entity for optimization as from now.
1277 val_arr[vnum].access_type = acces_type;
1283 static int set_block_dominated_first_access(ir_node *blk, int vnum, unsigned int access) {
1285 ir_node *idom, *succ;
1286 value_arr_entry_t *val_arr;
1289 idom = get_Block_idom(blk);
1290 for(i = get_Block_n_cfg_outs(idom) - 1; i >=1; i--) {
1291 succ = get_Block_cfg_out(idom, i);
1292 val_arr = get_irn_link(succ);
1293 if(val_arr[vnum].access_type < 3) {
1294 val_arr[vnum].access_type = access;
1300 /* Update the access information of a block if a predecessor of
1301 * this black have a higher access for an entity.
1303 * @param *irn An ir node from the current_ir_graph.
1304 * @param *env The enviroment pointer.
1306 static void set_block_access(ir_node *irn, void *ctx){
1308 value_arr_entry_t *val_arr, *val_arr_pred;
1309 ent_leaves_t *value_leaves;
1310 ir_node *pred, *pred_blk, *leave;
1315 val_arr = get_irn_link(irn);
1317 for( i = get_Block_n_cfgpreds(irn) - 1; i >= 0; i--) {
1318 /* We analyze the predecessors of this block to see if this block must
1320 pred = get_Block_cfgpred(irn, i);
1321 pred_blk = get_nodes_block(pred);
1323 val_arr_pred = get_irn_link(pred_blk);
1325 for(value_leaves = set_first(env->set_ent); value_leaves; value_leaves = set_next(env->set_ent)) {
1326 vnum = GET_ENT_VNUM(value_leaves->ent);
1328 if((get_Block_n_cfgpreds(irn) > 1) && (val_arr[vnum].access_type > 3))
1329 env->changes = set_block_dominated_first_access(irn, vnum, val_arr[vnum].access_type);
1331 if((val_arr_pred[vnum].access_type > 3) && (val_arr[vnum].access_type < 3)) {
1332 /* We have found a block for update it access and value number information.*/
1333 val_arr[vnum].access_type = val_arr_pred[vnum].access_type;
1334 /* We update the access information of all leave, that belong to
1337 for(leave = pset_first(value_leaves->leaves); leave; leave = pset_next(value_leaves->leaves))
1338 val_arr[GET_IRN_VNUM(leave)].access_type = val_arr[vnum].access_type;
1340 /* In this way can't be got the actuall number of value numbers.
1341 val_arr[env->vnum_state].access_type = val_arr_pred[env->vnum_state].access_type; */
1347 /* Free the allocated call sets.
1349 * @param irn A block form the ir graph.
1350 * @param env The enviroment pinter.
1352 static void free_call_info(ir_node *irn, void *ctx) {
1356 value_arr_entry_t *val_arr;
1359 val_arr = get_irn_link(irn);
1361 for(i = env->nvals + set_count(env->set_ent); i >= 0; i--) {
1362 if(val_arr[i].calls != NULL)
1364 del_set(val_arr[i].calls);
1368 static void print_block_state(ir_node *irn, void *ctx) {
1370 value_arr_entry_t *val_arr;
1371 ent_leaves_t *value_leaves;
1372 call_access_t *value_calls;
1377 val_arr = get_irn_link(irn);
1378 ir_printf("\n\nThe actual value number state of this block is: %i \n",
1379 val_arr[env->vnum_state].access_type - 1);
1381 for(value_leaves = set_first(env->set_ent); value_leaves; value_leaves = set_next(env->set_ent)) {
1383 vnum = GET_ENT_VNUM(value_leaves->ent);
1384 ir_printf("The entity %F access type in the block with nr %u is %i \n",
1385 value_leaves->ent, get_irn_node_nr(irn), val_arr[vnum].access_type);
1387 if(val_arr[vnum].calls != NULL)
1388 for(value_calls = set_first(val_arr[vnum].calls); value_calls; value_calls = set_next(val_arr[vnum].calls))
1390 ir_printf("A call with nr %i acess a element of this entity with access %u \n",
1391 get_irn_node_nr(value_calls->call), value_calls->access_type);
1396 /** Optimize the found scalar replacements.
1398 * @param set_sels A set with all entities, that
1400 * @param set_ent A set with all sels nodes,
1401 * that belong to our scalars.
1402 * @param vnum The number of scalars.
1404 static void do_data_flow_scalar_replacement(set *set_ent, set *set_sels, int vnum) {
1408 obstack_init(&env.obst);
1409 env.set_ent = set_ent;
1410 env.set_sels = set_sels;
1412 env.fix_phis = NULL;
1413 env.fix_syncs = NULL;
1414 env.gl_mem_vnum = vnum - 2;
1415 env.vnum_state = vnum - 1;
1416 /* nvals are vnum - 1, why we indicate with nvals the number
1417 * of memory edges we will produce. For vnum_state we don't
1418 * need to produce a memory edge.*/
1419 env.nvals = vnum - 1;
1422 /* first step: allocate the value arrays for every block */
1423 irg_block_walk_graph(current_ir_graph, NULL, alloc_value_arr, &env);
1425 /* second step: we analyze all calls, that have as parameter scalar(s).
1426 * We mark the calls, that save the address of a scalar and we
1427 * mark the entity owner of this scalar as not optimizeble by now.*/
1428 irg_walk_graph(current_ir_graph, NULL, analyse_calls, &env);
1430 while(env.changes) {
1435 * third step: walk over the blocks of a graph and update
1436 * the information for the access of our scalars.
1438 irg_block_walk_graph(current_ir_graph, NULL, set_block_access, &env);
1442 // if(get_firm_verbosity())
1443 /* Debug info to see if analyse_calls work properly.*/
1444 irg_block_walk_graph(current_ir_graph, NULL, print_block_state, &env);
1447 * fourth step: walk over the graph blockwise in topological order
1448 * and split the memrory edge.
1450 inc_irg_block_visited(current_ir_graph);
1451 irg_walk_blkwise_graph(current_ir_graph, NULL, split_memory_edge, &env);
1455 /* fifth step: fix all nodes, that have as predecessor Unknown.*/
1460 /* sixth step: sync memory enges for the end block.*/
1461 sync_mem_edges(&env);
1463 /*seventh step: free the allocated memory*/
1464 irg_block_walk_graph(current_ir_graph, NULL, free_call_info, &env);
1465 obstack_free(&env.obst, NULL);
1469 * Find possible scalar replacements
1471 * @param irg The current ir graph.
1473 void data_flow_scalar_replacement_opt(ir_graph *irg) {
1479 ent_leaves_t key_leaves, *value_leaves;
1482 if (! get_opt_scalar_replacement())
1485 set_sels = new_set(sels_cmp, 8);
1486 set_ent = new_set(ent_leaves_t_cmp, 8);
1488 /* Call algorithm that remove the critical edges of a ir graph. */
1489 remove_critical_cf_edges(irg);
1491 /* Call algorithm that computes the out edges.*/
1492 assure_irg_outs(irg);
1494 /* Call algorithm that computes the loop information.*/
1495 construct_cf_backedges(irg);
1497 /* Call algorithm that computes the dominance information.*/
1500 /* Find possible scalar replacements */
1501 if (find_possible_replacements(irg)) {
1503 /* Insert in set the scalar replacements. */
1504 irg_frame = get_irg_frame(irg);
1506 for (i = 0 ; i < get_irn_n_outs(irg_frame); i++) {
1507 ir_node *succ = get_irn_out(irg_frame, i);
1510 ir_entity *ent = get_Sel_entity(succ);
1512 if (get_entity_link(ent) == NULL || get_entity_link(ent) == ADDRESS_TAKEN)
1514 /* we have found a entity, that have scalars and we insert it to our set_ent*/
1515 key_leaves.ent = ent;
1516 key_leaves.leaves = new_pset(leave_cmp, 8);
1517 value_leaves = set_insert(set_ent, &key_leaves, sizeof(key_leaves), HASH_PTR(ent));
1519 /* We allocate for every leave sel a vnum.*/
1520 vnum = allocate_value_numbers(set_sels, value_leaves->leaves, ent, vnum);
1524 /* Allocate value number for the globule memory edge.
1525 * and a value number for the value numbers state.*/
1528 /* Allocate value numbers for the entities .*/
1529 for(i = vnum,value_leaves = set_first(set_ent); value_leaves; i++, value_leaves = set_next(set_ent))
1530 SET_ENT_VNUM(value_leaves->ent, i);
1533 do_data_flow_scalar_replacement(set_ent, set_sels, vnum);
1535 /*free the allocated memory.*/
1536 for(value_leaves = set_first(set_ent); value_leaves; value_leaves = set_next(set_ent))
1537 del_pset(value_leaves->leaves);