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
30 #include "iroptimize.h"
44 #include "analyze_irg_args.h"
49 #define SET_ENT_VNUM(ent, vnum) set_entity_link(ent, INT_TO_PTR(vnum))
50 #define GET_ENT_VNUM(ent) (unsigned)PTR_TO_INT(get_entity_link(ent))
51 #define SET_IRN_VNUM(irn, vnum) set_irn_link(irn, INT_TO_PTR(vnum))
52 #define GET_IRN_VNUM(irn) (unsigned)PTR_TO_INT(get_irn_link(irn))
56 typedef struct _ent_leaves_t{
57 ir_entity *ent; /**< An entity, that contains scalars for replace.*/
58 pset *leaves; /**< All leaves of this entity.*/
61 typedef struct _sels_t {
62 ir_node *sel; /**< A sel node, thats entity have scalars.*/
63 ir_entity *ent; /**< The entity of this sel node.*/
66 typedef struct _call_access_t {
67 ir_node *call; /**< A call node, that have as parameter a scalar.*/
68 unsigned int access_type; /**< The access type, with that this call access this scalar.*/
71 typedef struct _fixlist_entry_t {
72 ir_node *irn; /**< An ir node, that must be fixed.*/
73 unsigned int vnum; /**< The value number, that must became this ir node.*/
76 typedef struct _syncs_fixlist_entry_t {
77 ir_node *irn; /**< A sync node that must be fixed.*/
78 int *accessed_vnum; /**< A pointer to save an array with value numbers, that must became this sync.*/
79 }syncs_fixlist_entry_t;
81 /* A entry, that save the memory
82 * edge state and the access state for this leave
83 * int the array,that is created for every block.*/
84 typedef struct _leave_t {
85 ir_node *mem_edge_state; /**< memory state for this scalar in this block.*/
86 unsigned int access_type; /**< access state for this scalar in this block.*/
87 set *calls; /**< call nodes,that change this scalar in this block.*/
91 * A path element entry: it is either an entity
92 * or a tarval, because we evaluate only constant array
93 * accesses like a.b.c[8].d
101 * An access path, used to assign value numbers
102 * to variables that will be scalar replaced
104 typedef struct _path_t {
105 unsigned vnum; /**< the value number */
106 unsigned path_len; /**< the length of the access path */
107 path_elem_t path[1]; /**< the path */
111 * environment for memory walker
113 typedef struct _env_t {
114 struct obstack obst; /**< a obstack for the memory edge */
115 set *set_sels; /**< a set with all sels, that are reachable from an entity with a scalar.*/
116 set *set_ent; /**< a set with all entities that have one or more scalars.*/
117 fixlist_entry_t *fix_phis; /**< list of all Phi nodes that must be fixed */
118 fixlist_entry_t *fix_ls; /**< list of all Load or Store nodes that must be fixed */
119 syncs_fixlist_entry_t *fix_syncs; /**< list of all Sync nodes that must be fixed */
120 unsigned int nvals; /**< to save the number of scalars.*/
121 unsigned int gl_mem_vnum; /**< indicate the position of the globule memory edge state in var_arr.*/
122 unsigned int vnum_state; /**< indicate the position of the value number state in var_arr.*/
123 unsigned int changes; /**< to save if by anlyse_calls is changed anything.*/
129 * Compare two elements of the ent_leaves_t set.
131 * @return 0 if they are identically
133 static int ent_leaves_t_cmp(const void *elt, const void *key, size_t size)
135 const ent_leaves_t *c1 = elt;
136 const ent_leaves_t *c2 = key;
139 return c1->ent != c2->ent;
143 * Compare two elements of the ent_access_t set.
145 * @return 0 if they are identically
147 static int ent_cmp(const void *elt, const void *key)
149 const ir_entity *c1 = elt;
150 const ir_entity *c2 = key;
156 * Compare two elements of the sels_t set.
158 * @return 0 if they are identically
160 static int sels_cmp(const void *elt, const void *key, size_t size)
162 const sels_t *c1 = elt;
163 const sels_t *c2 = key;
166 return c1->sel != c2->sel;
170 * Compare two elements of the leave_t set.
172 * @return 0 if they are identically
174 static int leave_cmp(const void *elt, const void *key)
176 ir_node *c1 = (ir_node *)elt;
177 ir_node *c2 = (ir_node *)key;
179 return get_Sel_entity(c1) != get_Sel_entity(c2);
183 * Compare two elements of the call_access_t set.
185 * @return 0 if they are identically
187 static int call_cmp(const void *elt, const void *key, size_t size)
189 const call_access_t *c1 = elt;
190 const call_access_t *c2 = key;
193 return c1->call != c2->call;
199 * @return 0 if they are identically
201 static int path_cmp(const void *elt, const void *key, size_t size)
203 const path_t *p1 = elt;
204 const path_t *p2 = key;
207 /* we can use memcmp here, because identical tarvals should have identical addresses */
208 return memcmp(p1->path, p2->path, p1->path_len * sizeof(p1->path[0]));
212 * Calculate a hash value for a path.
214 static unsigned path_hash(const path_t *path)
219 for (i = 0; i < path->path_len; ++i)
220 hash ^= (unsigned)PTR_TO_INT(path->path[i].ent);
226 * Returns non-zero, if all induces of a Sel node are constants.
228 * @param sel the Sel node that will be checked
230 static int is_const_sel(ir_node *sel) {
231 int i, n = get_Sel_n_indexs(sel);
233 for (i = 0; i < n; ++i) {
234 ir_node *idx = get_Sel_index(sel, i);
243 * Returns non-zero, if the address of an entity
244 * represented by a Sel node (or it's successor Sels) is taken.
246 static int is_address_taken_2(ir_node *sel)
250 if (! is_const_sel(sel))
253 for (i = get_irn_n_outs(sel) - 1; i >= 0; --i) {
254 ir_node *succ = get_irn_out(sel, i);
256 switch (get_irn_opcode(succ)) {
258 /* ok, we just load from that entity */
262 /* check that Sel is not the Store's value */
263 if (get_Store_value(succ) == sel)
268 /* Check the Sel successor of Sel */
269 int res = is_address_taken_2(succ);
277 /* The address of an entity is given as a parameter.
278 * We analyzes that later and optimizes this scalar
284 /* another op, the address is taken */
292 * Link all Sels with the entity.
294 * @param ent the entity that will be scalar replaced
295 * @param sel a Sel node that selects some fields of this entity
297 static void link_all_leave_sels(ir_entity *ent, ir_node *sel)
301 n = get_irn_n_outs(sel);
302 for (i = 0; i < n; ++i) {
303 ir_node *succ = get_irn_out(sel, i);
306 link_all_leave_sels(ent, succ);
309 /* if Sel nodes with memory inputs are used, a entity can be
310 * visited more than once causing a ring here, so we use the
311 * node flag to mark linked nodes
313 if (irn_visited_else_mark(sel))
317 * we link the sels to the entity.
319 set_irn_link(sel, get_entity_link(ent));
320 set_entity_link(ent, sel);
323 /* we need a special address that serves as an address taken marker */
325 static void *ADDRESS_TAKEN = &_x;
328 * Find possible scalar replacements.
330 * @param irg an IR graph
332 * This function finds variables on the (members of the) frame type
333 * that can be scalar replaced, because their address is never taken.
334 * If such a variable is found, it's entity link will hold a list of all
335 * Sel nodes, that selects anythings of this entity.
336 * Otherwise, the link will be ADDRESS_TAKEN or NULL.
338 * @return non-zero if at least one entity could be replaced
341 static int find_possible_replacements(ir_graph *irg)
343 ir_node *irg_frame = get_irg_frame(irg);
347 inc_irg_visited(irg);
349 n = get_irn_n_outs(irg_frame);
352 * First, clear the link field of all interestingentities.
353 * Note that we did not rely on the fact that there is only
354 * one Sel node per entity, so we might access one entity
355 * more than once here.
356 * That's why we have need two loops.
358 for (i = 0; i < n; ++i) {
359 ir_node *succ = get_irn_out(irg_frame, i);
362 ir_entity *ent = get_Sel_entity(succ);
363 set_entity_link(ent, NULL);
368 * Check the ir_graph for Sel nodes. If the entity of Sel
369 * isn't a scalar replacement set the link of this entity
370 * equal ADDRESS_TAKEN.
372 for (i = 0; i < n; ++i) {
373 ir_node *succ = get_irn_out(irg_frame, i);
376 ir_entity *ent = get_Sel_entity(succ);
379 if (get_entity_link(ent) == ADDRESS_TAKEN)
383 * Beware: in rare cases even entities on the frame might be
384 * volatile. This might happen if the entity serves as a store
385 * to a value that must survive a exception. Do not optimize
386 * such entities away.
388 if (get_entity_volatility(ent) == volatility_is_volatile) {
389 set_entity_link(ent, ADDRESS_TAKEN);
393 ent_type = get_entity_type(ent);
395 /* we can handle arrays, structs and atomic types yet */
396 if (is_Array_type(ent_type) || is_Struct_type(ent_type) || is_atomic_type(ent_type)) {
397 if (is_address_taken_2(succ)) {
398 if (get_entity_link(ent)) /* killing one */
400 set_entity_link(ent, ADDRESS_TAKEN);
403 /* possible found one */
404 if (get_entity_link(ent) == NULL)
406 link_all_leave_sels(ent, succ);
415 static int is_leave_sel(ir_node *sel) {
419 for(i = get_irn_n_outs(sel) - 1; i >= 0; i--) {
420 succ = get_irn_out(sel, i);
429 * Return a path from the Sel node sel to it's root.
431 * @param sel the Sel node
432 * @param len the length of the path so far
434 static path_t *find_path(ir_node *sel, unsigned len)
438 ir_node *pred = get_Sel_ptr(sel);
440 /* the current Sel node will add some path elements */
441 n = get_Sel_n_indexs(sel);
445 /* we found the root */
446 res = XMALLOCF(path_t, path, len);
450 res = find_path(pred, len);
452 pos = res->path_len - len;
454 res->path[pos++].ent = get_Sel_entity(sel);
455 for (i = 0; i < n; ++i) {
456 ir_node *index = get_Sel_index(sel, i);
459 res->path[pos++].tv = get_Const_tarval(index);
465 * Allocate value numbers for the leaves
466 * in our found entities.
468 * @param sels a set that will contain all Sels that have a value number
469 * @param ent the entity that will be scalar replaced
470 * @param vnum the first value number we can assign
471 * @param modes a flexible array, containing all the modes of
474 * @return the next free value number
476 static unsigned allocate_value_numbers(set *set_sels, pset *leaves, ir_entity *ent, unsigned vnum)
481 set *pathes = new_set(path_cmp, 8);
483 /* visit all Sel nodes in the chain of the entity */
484 for (sel = get_entity_link(ent); sel; sel = next) {
485 next = get_irn_link(sel);
487 /* we save for every sel it root entity, why
488 * we need this information, when we split the memory edge,
489 * and we must mark this sel for later. */
492 set_insert(set_sels, &key_sels, sizeof(key_sels), HASH_PTR(sel));
494 if(! is_leave_sel(sel))
496 /* We have found a leave and we add it to the pset of this entity.*/
497 pset_insert(leaves, sel, HASH_PTR(get_Sel_entity(sel)));
499 key = find_path(sel, 0);
500 path = set_find(pathes, key, sizeof(*key) + sizeof(key->path[0]) * key->path_len, path_hash(key));
503 SET_IRN_VNUM(sel, path->vnum);
508 set_insert(pathes, key, sizeof(*key) + sizeof(key->path[0]) * key->path_len, path_hash(key));
510 SET_IRN_VNUM(sel, key->vnum);
516 set_entity_link(ent, NULL);
520 * Add a sync node to it fix list.
522 * @param sync The sync node, that myst be addet to the fix list.
523 * @param unk_vnum An array whit the value number, that are synced with this sync node.
524 * @param env The enviroment pinter.
526 static void add_sync_to_fixlist(ir_node *sync, int *unk_vnum, env_t *env) {
528 syncs_fixlist_entry_t *s;
530 s = obstack_alloc(&env->obst, sizeof(*s));
532 s->accessed_vnum = unk_vnum;
533 set_irn_link(sync, env->fix_syncs);
537 * Add a ir node to it fix list.
539 * @param irn The ir node, that myst be addet to the fix list.
540 * @param vnum The value number, that must baceme this ir node as predecessor later.
541 * @param env The enviroment pinter.
543 static void add_ls_to_fixlist(ir_node *irn, int vnum, env_t *env) {
547 l = obstack_alloc(&env->obst, sizeof(*l));
551 if(get_irn_op(irn) == op_Phi) {
552 set_irn_link(l->irn, env->fix_phis);
555 set_irn_link(l->irn, env->fix_ls);
560 static void add_mem_edge(value_arr_entry_t *val_arr, int vnum, ir_node ***in, int **accessed_vnum) {
562 if(val_arr[vnum].mem_edge_state != NULL)
563 ARR_APP1(ir_node *, *in, val_arr[vnum].mem_edge_state);
565 ARR_APP1(int, *accessed_vnum, vnum);
566 ARR_APP1(ir_node *, *in, new_Unknown(mode_M));
570 * The function handles the scalars, that wase stored
573 * @param blk The block, that must be handled.
574 * @param env The enviroment pinter.
577 /* Return the memory successor of the call node.*/
578 static ir_node *get_Call_mem_out(ir_node *call) {
583 for(i = get_irn_n_outs(call) - 1; i >= 0; i--) {
584 mem = get_irn_out(call, i);
585 if(get_irn_mode(mem) == mode_M)
588 /* is not reachable*/
593 static void sync_stored_scalars(ir_node *blk, env_t *env) {
596 int *unk_vnum; /**< An arraw, where are saved the value number, that
597 are synced from this sync node.*/
598 ent_leaves_t *value_ent;
599 value_arr_entry_t *val_arr_blk, *val_arr;
600 ir_node *pred, *leave, *sync, **in;
601 ir_node *sync_blk; /**< The block, where the sync node must be created.*/
604 val_arr_blk = get_irn_link(blk);
606 for(value_ent = set_first(env->set_ent); value_ent; value_ent = set_next(env->set_ent)) {
609 if(val_arr_blk[GET_ENT_VNUM(value_ent->ent)].access_type <= 3)
610 /* This entity is not stored in this block.*/
613 for(i = get_Block_n_cfgpreds(blk) - 1; i >= 0; i--) {
615 pred = get_Block_cfgpred(blk, i);
616 pred = get_nodes_block(pred);
617 val_arr = get_irn_link(pred);
619 if(val_arr[GET_ENT_VNUM(value_ent->ent)].access_type == SYNCED)
620 /* This entity was synced.*/
623 if(val_arr[GET_ENT_VNUM(value_ent->ent)].access_type <= 3) {
625 /* To avoid repeated sync of this entity in this block.*/
626 val_arr[GET_ENT_VNUM(value_ent->ent)].access_type = SYNCED;
627 /* In this predecessor block is this entity not acessed.
628 * We must sync in the end ot this block.*/
629 if(get_Block_n_cfgpreds(blk) > 1)
630 sync_blk = get_nodes_block(get_Block_cfgpred(blk, i));
634 val_arr = get_irn_link(sync_blk);
635 /* An array to save the memory edges, that must be
637 in = NEW_ARR_F(ir_node *, 1);
639 /* An array to save the value numbers,
640 * that must be repaired.*/
641 unk_vnum = NEW_ARR_F(int, 0);
642 /* The global memory edge.*/
643 if(val_arr[env->gl_mem_vnum].mem_edge_state == NULL)
644 in[0] = new_Unknown(mode_M);
646 in[0] = val_arr[env->gl_mem_vnum].mem_edge_state;
648 for(leave = pset_first(value_ent->leaves); leave; leave = pset_next(value_ent->leaves))
649 /* All this memory edges must be synced.*/
650 add_mem_edge(val_arr, GET_IRN_VNUM(leave), &in, &unk_vnum);
652 /* We create the sync and set it in the global memory state.*/
653 sync = new_r_Sync(current_ir_graph, sync_blk, ARR_LEN(in), in);
654 /* We must check this, why it is possible to get a Bad node
655 * form new_r_Sync(), when the node can be optimized.
656 * In this case we must do nothing.*/
658 val_arr[env->gl_mem_vnum].mem_edge_state = sync;
659 /* We add this sync node to the sync's fix list.*/
660 add_sync_to_fixlist(val_arr[env->gl_mem_vnum].mem_edge_state, unk_vnum, env);
668 * The function split the memory edge of load and store nodes, that have
669 * as predecessor a scalar
671 * @param irn The node, that memory edge must be spleted.
672 * @param env The enviroment pinter.
674 static void split_ls_mem_edge(ir_node *irn, env_t *env) {
677 ir_node *leave, *irn_blk, *mem_state, *new_mem_state;
678 unsigned ent_vnum, sel_vnum, i;
679 value_arr_entry_t *val_arr;
680 sels_t key_sels, *value_sels;
681 ent_leaves_t key_ent, *value_ent;
683 op = get_irn_op(irn);
686 key_sels.sel = get_Load_ptr(irn);
688 key_sels.sel = get_Store_ptr(irn);
690 value_sels = set_find(env->set_sels, &key_sels, sizeof(key_sels), HASH_PTR(key_sels.sel));
692 if(value_sels != NULL) {
693 /* we have found a load or store, that use a sel of our set
694 * and we must split or extend, if the memory edge have been
695 * split for this sel, the memory edge.*/
697 key_ent.ent = value_sels->ent;
698 value_ent = set_find(env->set_ent, &key_ent, sizeof(key_ent), HASH_PTR(key_ent.ent));
699 /*To check if the enities set is right filled. */
700 assert(value_ent && " This sel's entity isn't int the entity set.");
702 leave = pset_find(value_ent->leaves, key_sels.sel, HASH_PTR(get_Sel_entity(key_sels.sel)));
703 /*To check if the leaves set is right filled. */
704 assert(leave && "Anything in data_flow_scalar_replacment algorithm is wrong.");
706 ent_vnum = GET_ENT_VNUM(value_ent->ent);
707 sel_vnum = GET_IRN_VNUM(leave);
708 irn_blk = get_nodes_block(irn);
709 val_arr = get_irn_link(irn_blk);
711 if(val_arr[ent_vnum].access_type == 0)
712 /* We have found a scalar, that address is not stored as jet.*/
715 /* This scalar have been stored.*/
716 i = env->gl_mem_vnum;
718 if(val_arr[i].mem_edge_state == NULL) {
719 /* We split now for this sel the memory edge in this block.*/
720 mem_state = new_Unknown(mode_M);
721 /* We must mark this node to fix later*/
722 add_ls_to_fixlist(irn, i, env);
725 /* We have split the memory edge and the current state is saved.*/
726 mem_state = val_arr[i].mem_edge_state;
728 /* We set this Load or Store to the memory edge of this
731 set_Load_mem(irn, mem_state);
733 set_Store_mem(irn, mem_state);
735 /* When we have split or extended the memory edge we must
736 * update the memory_edge_state of this sel*/
737 new_mem_state = get_irn_out(irn, 0);
738 if(get_irn_mode(new_mem_state) == mode_M)
739 val_arr[i].mem_edge_state = new_mem_state;
741 val_arr[i].mem_edge_state = get_irn_out(irn, 1);
746 * The function split the memory edge of phi nodes, that have
747 * as predecessor a scalar
749 * @param irn The phi node, that memory edge must be spleted.
750 * @param env The enviroment pinter.
752 static void split_phi_mem_edge(ir_node *irn, env_t *env) {
754 ir_node *irn_blk, *unk, *leave, **in;
756 ent_leaves_t *value_ent;
757 value_arr_entry_t *val_arr;
759 irn_blk = get_nodes_block(irn);
760 val_arr = get_irn_link(irn_blk);
762 n = get_Block_n_cfgpreds(irn_blk);
764 in = alloca(sizeof(*in) * n);
766 for(value_ent = set_first(env->set_ent); value_ent; value_ent = set_next(env->set_ent))
767 if(val_arr[GET_ENT_VNUM(value_ent->ent)].access_type < 3)
768 /* This scalar wasn't be saved and we need to produce a phi for it.*/
769 for(leave = pset_first(value_ent->leaves); leave; leave = pset_next(value_ent->leaves)){
771 unk = new_Unknown(mode_M);
772 for (j = n - 1; j >= 0; --j)
775 val_arr[GET_IRN_VNUM(leave)].mem_edge_state = new_r_Phi(current_ir_graph, irn_blk, n, in, mode_M);
777 add_ls_to_fixlist(val_arr[GET_IRN_VNUM(leave)].mem_edge_state, GET_IRN_VNUM(leave), env);
780 /* We use for the global memory the phi node, that
781 * is already available.*/
782 val_arr[env->gl_mem_vnum].mem_edge_state = irn;
786 * The function handles the call nodes, that have
787 * as parameter a scalar
789 * @param env The enviroment pinter.
790 * @param call The call node, that must be handled.
791 * @param accessed_entities A set wit all entities, that are accessed from this call node.*/
792 static void split_call_mem_edge(env_t *env, ir_node *call, pset *accessed_entities) {
794 ent_leaves_t key_ent, *value_ent;
795 value_arr_entry_t *val_arr;
796 call_access_t key_call, *value_call;
797 ir_node *call_blk, *new_mem_state, *leave;
801 int fix_irn = 0; /**< Set to 1 if we must add this call to it fix list.*/
802 int *accessed_leaves_vnum = NULL; /**< An arraw, where are saved the value number, that
803 are synced from call's sync node, if we need it.*/
805 call_blk = get_nodes_block(call);
806 val_arr = get_irn_link(call_blk);
807 /* An array to save the memory edges, that must be
809 in = NEW_ARR_F(ir_node *, 1);
810 /* An array to save the value numbers of the memory
811 * edges that must be repaired.*/
812 accessed_leaves_vnum = NEW_ARR_F(int, 0);
814 /* We get the memory successor of the call node.
815 * It is the new memory state for all synced memory
817 new_mem_state = get_Call_mem_out(call);
819 /* The global memory is the first predecessor of the create sync node.*/
820 if(val_arr[env->gl_mem_vnum].mem_edge_state == NULL) {
821 in[0] = new_Unknown(mode_M);
825 in[0] = val_arr[env->gl_mem_vnum].mem_edge_state;
828 for(ent = pset_first(accessed_entities); ent; ent = pset_next(accessed_entities)) {
829 /* Whit this loop we iterate all accessed entities from this call and collect
830 * all memory edges, that we must sync.*/
831 ent_vnum = GET_ENT_VNUM(ent);
833 key_call.call = call;
834 value_call = set_find(val_arr[ent_vnum].calls, &key_call, sizeof(key_call), HASH_PTR(key_call.call));
837 value_ent = set_find(env->set_ent, &key_ent, sizeof(key_ent), HASH_PTR(key_ent.ent));
839 if(val_arr[ent_vnum].access_type <= 3) {
840 /* This scalar's address wasn't stored in this block.*/
841 switch(value_call->access_type) {
843 case ptr_access_none :
844 /* In this case we have nothing to do.*/
847 case ptr_access_read:
848 case ptr_access_write:
850 /* All this cases must be traded equal.*/
852 for(leave = pset_first(value_ent->leaves); leave; leave = pset_next(value_ent->leaves)){
853 /* All this memory edges must be synced.*/
854 add_mem_edge(val_arr, GET_IRN_VNUM(leave), &in, &accessed_leaves_vnum);
856 /* We update the memory state of this leave.*/
857 if(value_call->access_type != ptr_access_read)
858 val_arr[GET_IRN_VNUM(leave)].mem_edge_state = new_mem_state;
867 /* We must update the global memory state.*/
868 val_arr[env->gl_mem_vnum].mem_edge_state = new_mem_state;
870 if(ARR_LEN(in) == 1) {
871 /* we must set the call memory to gobale momory*/
872 set_Call_mem(call,in[0]);
875 /* We add this call node to the call fix list..*/
876 add_ls_to_fixlist(call, env->gl_mem_vnum, env);
879 /* We create the sync and set it as memory predecessor of the call node.*/
880 sync = new_r_Sync(current_ir_graph, call_blk, ARR_LEN(in), in);
881 /* We must check this, why it is possible to get a Bad node
882 * form new_r_Sync(), when the node can be optimized.
883 * In this case we must do nothing.*/
885 set_Call_mem(call, sync);
886 if(ARR_LEN(accessed_leaves_vnum))
887 /* We add this sync node to the sync's fix list.*/
888 add_sync_to_fixlist(sync, accessed_leaves_vnum, env);
895 * The function split the memory edge from the passed
896 * ir node if this is needed
898 * @param irn The node, that memory edge must be spleted.
899 * @param env The enviroment pinter.
901 static void split_memory_edge(ir_node *irn, void *ctx) {
904 ir_node *sel, *irn_blk;
906 sels_t key_sels, *value_sels;
907 value_arr_entry_t *val_arr;
908 pset *accessed_entities; /**< A set to save all entities accessed from a call.*/
912 op = get_irn_op(irn);
917 irn_blk = get_nodes_block(irn);
919 if (!Block_block_visited(irn_blk)) {
920 /* We sync first the stored scalar address in this block.*/
921 mark_Block_block_visited(irn_blk);
922 sync_stored_scalars(irn_blk, env);
925 if(op == op_Load || op == op_Store)
927 split_ls_mem_edge(irn, env);
930 if (op == op_Phi && get_irn_mode(irn) == mode_M) {
932 * found a memory Phi: Here, we must create new Phi nodes
934 split_phi_mem_edge(irn, env);
939 /* Calls that have a NoMem input do neither read nor write memory.
940 We can completely ignore them here. */
941 if (is_NoMem(get_Call_mem(irn)))
944 /* We save in this set all entities,
945 * that are accessed from this call node.*/
946 accessed_entities = new_pset(ent_cmp, 8);
947 val_arr = get_irn_link(get_nodes_block(irn));
949 for ( i = get_Call_n_params(irn) - 1; i >= 0; i--) {
951 sel = get_Call_param(irn, i);
955 value_sels = set_find(env->set_sels, &key_sels, sizeof(key_sels), HASH_PTR(key_sels.sel));
957 if(value_sels != NULL && val_arr[GET_ENT_VNUM(value_sels->ent)].access_type <= 3)
958 /* We save in this set all accessed entities from this call node whit
959 * access none, read, write or rw..*/
960 pset_insert(accessed_entities, value_sels->ent, HASH_PTR(value_sels->ent));
964 if(pset_count(accessed_entities))
965 split_call_mem_edge(env, irn, accessed_entities);
967 del_pset(accessed_entities);
974 * searches through blocks beginning from block for value
975 * vnum and return it.
977 * @param block A block from the current ir graph.
978 * @param vnum The value number, that must be found.
980 static ir_node *find_vnum_value(ir_node *block, unsigned vnum)
982 value_arr_entry_t *val_arr;
986 if (!Block_block_visited(block)) {
987 mark_Block_block_visited(block);
989 val_arr = get_irn_link(block);
991 if (val_arr[vnum].mem_edge_state)
992 return val_arr[vnum].mem_edge_state;
994 for (i = get_Block_n_cfgpreds(block) - 1; i >= 0; --i) {
995 ir_node *pred = get_Block_cfgpred(block, i);
997 res = find_vnum_value(get_nodes_block(pred), vnum);
1006 * fix the Load/Store or Call list
1008 * @param The enviroment pinter.
1010 static void fix_ls(env_t *env)
1013 ir_node *irn, *block, *pred, *val = NULL;
1017 for (l = env->fix_ls; l; l = get_irn_link(irn)) {
1020 op = get_irn_op(irn);
1021 block = get_nodes_block(irn);
1022 for (i = get_Block_n_cfgpreds(block) - 1; i >= 0; --i) {
1023 pred = get_Block_cfgpred(block, i);
1024 pred = get_nodes_block(pred);
1026 inc_irg_block_visited(current_ir_graph);
1027 val = find_vnum_value(pred, l->vnum);
1035 set_Store_mem(irn, val);
1038 set_Load_mem(irn, val);
1040 set_Call_mem(irn, val);
1048 * @param The enviroment pinter.
1050 static void fix_phis(env_t *env)
1053 ir_node *phi, *block, *pred, *val;
1056 for (l = env->fix_phis; l; l = get_irn_link(phi)) {
1059 block = get_nodes_block(phi);
1060 for (i = get_Block_n_cfgpreds(block) - 1; i >= 0; --i) {
1062 pred = get_Block_cfgpred(block, i);
1063 pred = get_nodes_block(pred);
1065 inc_irg_block_visited(current_ir_graph);
1066 val = find_vnum_value(pred, l->vnum);
1069 set_irn_n(phi, i, val);
1078 * @param The enviroment pinter.
1080 static void fix_syncs(env_t *env)
1082 syncs_fixlist_entry_t *l;
1083 ir_node *sync, *block, *pred, *val;
1087 for (l = env->fix_syncs; l; l = get_irn_link(sync)) {
1091 /* The sync block must have one predecessor, when it
1092 have unknown nodes as predecessor.*/
1093 block = get_nodes_block(sync);
1094 pred = get_Block_cfgpred(block, 0);
1095 pred = get_nodes_block(pred);
1097 /* We first repair the global memory edge at the first position of sync predecessors.*/
1098 if (is_Unknown(get_irn_n(sync, 0))) {
1099 inc_irg_block_visited(current_ir_graph);
1100 val = find_vnum_value(pred, env->gl_mem_vnum);
1103 set_irn_n(sync, 0, val);
1106 for (i = get_irn_arity(sync) - 1; i >= 1; --i) {
1107 /* We repair the leaves*/
1109 assert(k <= ARR_LEN(l->accessed_vnum) && "The algorythm for sync repair is wron");
1110 if (is_Unknown(get_irn_n(sync, i))) {
1111 inc_irg_block_visited(current_ir_graph);
1112 val = find_vnum_value(pred, l->accessed_vnum[k++]);
1115 set_irn_n(sync, i, val);
1118 DEL_ARR_F(l->accessed_vnum);
1122 * For the end node we must sync all memory edges.
1124 * @param The enviroment pinter.
1126 static void sync_mem_edges(env_t *env) {
1128 value_arr_entry_t *val_arr;
1129 ir_node **in, *sync, *Return, *Return_blk;
1130 int i, vnum, vnum_state;
1132 Return = get_Block_cfgpred(get_irg_end_block(current_ir_graph), 0);
1133 Return_blk = get_nodes_block(Return);
1134 val_arr = get_irn_link(Return_blk);
1138 for(i = 0; i <= (int)env->gl_mem_vnum; i++)
1139 /* we get the current state of non saved scalars.*/
1140 if(val_arr[i].access_type <= 3)
1143 /* We allocate the memory, that we need for the predecessors of the sync.*/
1144 in = XMALLOCN(ir_node*, vnum_state);
1146 /* The global memory edge is the first predecessor of this sync node.*/
1147 if(val_arr[env->gl_mem_vnum].mem_edge_state == NULL) {
1148 /* We must search through blocks for this memory state.*/
1149 inc_irg_block_visited(current_ir_graph);
1150 in[0] = find_vnum_value(Return_blk, env->gl_mem_vnum);
1153 in[0] = val_arr[env->gl_mem_vnum].mem_edge_state;
1156 for(i = 1, vnum = 0; vnum < (int)env->gl_mem_vnum; vnum++) {
1158 if(val_arr[vnum].access_type <= 3) {
1159 /* we add the non saved scalars as predecessors of the sync.*/
1161 if(val_arr[vnum].mem_edge_state == NULL) {
1162 /* We must search through blocks for this memory state.*/
1163 inc_irg_block_visited(current_ir_graph);
1164 in[i] = find_vnum_value(Return_blk, vnum);
1167 in[i] = val_arr[vnum].mem_edge_state;
1172 sync = new_r_Sync(current_ir_graph, Return_blk, vnum_state, in);
1173 set_Return_mem(Return, sync);
1179 * Walker: allocate the value array for every block.
1181 * @param block A block from the current ir graph for that must be allocated a value array.
1182 * @param ctx The enviroment pinter.
1184 static void alloc_value_arr(ir_node *block, void *ctx)
1189 value_arr_entry_t *var_arr = obstack_alloc(&env->obst, sizeof(value_arr_entry_t) *(env->nvals + set_count(env->set_ent) + 1));
1191 /* the value array is empty at start */
1192 memset(var_arr, 0, sizeof(value_arr_entry_t) * (env->nvals + set_count(env->set_ent) + 1));
1193 set_irn_link(block, var_arr);
1195 /* We set the block value number state to optimal and later we update this.*/
1196 var_arr[env->vnum_state].access_type = env->nvals;
1198 if(get_irg_start_block(current_ir_graph) == block)
1199 /* We initilize the startblocks array with the irg initilize memory, why
1200 * it must be the start point of all memory edges.*/
1201 for(i = (env->nvals + set_count(env->set_ent)) ; i >=0; i--)
1202 var_arr[i].mem_edge_state = get_irg_initial_mem(current_ir_graph);
1206 /* Analyze call nodes to get information, if they store the address of a scalar.
1208 * @param *irn An ir node from the current_ir_graph.
1209 * @param *env The enviroment pointer.
1211 static void analyse_calls(ir_node *irn, void *ctx) {
1214 unsigned int acces_type;
1215 ir_node *param, *call_ptr, *blk;
1216 ir_entity *meth_ent;
1217 sels_t key_sels, *value_sels;
1218 call_access_t key_call, *value_call;
1219 value_arr_entry_t *val_arr;
1226 /* Calls that have a NoMem input do neither read nor write memory.
1227 We can completely ignore them here. */
1228 if (is_NoMem(get_Call_mem(irn)))
1231 /* We iterate over the parameters of this call nodes.*/
1232 for ( i = get_Call_n_params(irn) - 1; i >= 0; i--) {
1233 param = get_Call_param(irn, i);
1234 if (is_Sel(param)) {
1235 /* We have found a parameter with operation sel.*/
1236 key_sels.sel = param;
1237 value_sels = set_find(env->set_sels, &key_sels, sizeof(key_sels), HASH_PTR(key_sels.sel));
1238 if(value_sels != NULL ) {
1240 /* We have found a call, that have as parameter a sel from our set_sels.*/
1241 call_ptr = get_Call_ptr(irn);
1243 if (is_SymConst(call_ptr) && get_SymConst_kind(call_ptr) == symconst_addr_ent) {
1244 meth_ent = get_SymConst_entity(call_ptr);
1245 /* we get the access type for our sel.*/
1246 acces_type = get_method_param_access(meth_ent, i);
1248 /* We can't analyze this function and we asume, that it store the address.*/
1249 acces_type = ptr_access_store;
1251 /* we save the access type and this call in the array allocated for this block.
1252 * The value number of this entity get us the position in the array to save this
1253 * information. Why we expect more calls as one we allocate a set.*/
1254 vnum = GET_ENT_VNUM(value_sels->ent);
1255 blk = get_nodes_block(irn);
1256 val_arr = get_irn_link(blk);
1258 if(val_arr[vnum].access_type > 3)
1259 /* The address of this entity have been stored.*/
1262 if(val_arr[vnum].calls == NULL)
1263 /* for this entity i have found the firs call in this block and we must allocate the set.*/
1264 val_arr[vnum].calls = new_set(call_cmp, 8);
1266 /* This call performs anything with the scalar and we must mark it.*/
1267 key_call.call = irn;
1268 key_call.access_type = acces_type;
1269 value_call = set_insert(val_arr[vnum].calls, &key_call, sizeof(key_call), HASH_PTR(key_call.call));
1271 if(value_call->access_type < acces_type)
1272 /* this case tread, when a call access an entity more at once.
1273 * Than we must save the highest access type.*/
1274 value_call->access_type = acces_type;
1277 /* This call save the address of our scalar and we can't
1278 * use the scalars of this entity for optimization as from now.
1280 val_arr[vnum].access_type = acces_type;
1286 static int set_block_dominated_first_access(ir_node *blk, int vnum, unsigned int access) {
1288 ir_node *idom, *succ;
1289 value_arr_entry_t *val_arr;
1292 idom = get_Block_idom(blk);
1293 for(i = get_Block_n_cfg_outs(idom) - 1; i >=1; i--) {
1294 succ = get_Block_cfg_out(idom, i);
1295 val_arr = get_irn_link(succ);
1296 if(val_arr[vnum].access_type < 3) {
1297 val_arr[vnum].access_type = access;
1303 /* Update the access information of a block if a predecessor of
1304 * this black have a higher access for an entity.
1306 * @param *irn An ir node from the current_ir_graph.
1307 * @param *env The enviroment pointer.
1309 static void set_block_access(ir_node *irn, void *ctx){
1311 value_arr_entry_t *val_arr, *val_arr_pred;
1312 ent_leaves_t *value_leaves;
1313 ir_node *pred, *pred_blk, *leave;
1318 val_arr = get_irn_link(irn);
1320 for( i = get_Block_n_cfgpreds(irn) - 1; i >= 0; i--) {
1321 /* We analyze the predecessors of this block to see if this block must
1323 pred = get_Block_cfgpred(irn, i);
1324 pred_blk = get_nodes_block(pred);
1326 val_arr_pred = get_irn_link(pred_blk);
1328 for(value_leaves = set_first(env->set_ent); value_leaves; value_leaves = set_next(env->set_ent)) {
1329 vnum = GET_ENT_VNUM(value_leaves->ent);
1331 if((get_Block_n_cfgpreds(irn) > 1) && (val_arr[vnum].access_type > 3))
1332 env->changes = set_block_dominated_first_access(irn, vnum, val_arr[vnum].access_type);
1334 if((val_arr_pred[vnum].access_type > 3) && (val_arr[vnum].access_type < 3)) {
1335 /* We have found a block for update it access and value number information.*/
1336 val_arr[vnum].access_type = val_arr_pred[vnum].access_type;
1337 /* We update the access information of all leave, that belong to
1340 for(leave = pset_first(value_leaves->leaves); leave; leave = pset_next(value_leaves->leaves))
1341 val_arr[GET_IRN_VNUM(leave)].access_type = val_arr[vnum].access_type;
1343 /* In this way can't be got the actuall number of value numbers.
1344 val_arr[env->vnum_state].access_type = val_arr_pred[env->vnum_state].access_type; */
1350 /* Free the allocated call sets.
1352 * @param irn A block form the ir graph.
1353 * @param env The enviroment pinter.
1355 static void free_call_info(ir_node *irn, void *ctx) {
1359 value_arr_entry_t *val_arr;
1362 val_arr = get_irn_link(irn);
1364 for(i = env->nvals + set_count(env->set_ent); i >= 0; i--) {
1365 if(val_arr[i].calls != NULL)
1367 del_set(val_arr[i].calls);
1371 static void print_block_state(ir_node *irn, void *ctx) {
1373 value_arr_entry_t *val_arr;
1374 ent_leaves_t *value_leaves;
1375 call_access_t *value_calls;
1380 val_arr = get_irn_link(irn);
1381 ir_printf("\n\nThe actual value number state of this block is: %i \n",
1382 val_arr[env->vnum_state].access_type - 1);
1384 for(value_leaves = set_first(env->set_ent); value_leaves; value_leaves = set_next(env->set_ent)) {
1386 vnum = GET_ENT_VNUM(value_leaves->ent);
1387 ir_printf("The entity %F access type in the block with nr %u is %i \n",
1388 value_leaves->ent, get_irn_node_nr(irn), val_arr[vnum].access_type);
1390 if(val_arr[vnum].calls != NULL)
1391 for(value_calls = set_first(val_arr[vnum].calls); value_calls; value_calls = set_next(val_arr[vnum].calls))
1393 ir_printf("A call with nr %i acess a element of this entity with access %u \n",
1394 get_irn_node_nr(value_calls->call), value_calls->access_type);
1399 /** Optimize the found scalar replacements.
1401 * @param set_sels A set with all entities, that
1403 * @param set_ent A set with all sels nodes,
1404 * that belong to our scalars.
1405 * @param vnum The number of scalars.
1407 static void do_data_flow_scalar_replacement(set *set_ent, set *set_sels, int vnum) {
1411 obstack_init(&env.obst);
1412 env.set_ent = set_ent;
1413 env.set_sels = set_sels;
1415 env.fix_phis = NULL;
1416 env.fix_syncs = NULL;
1417 env.gl_mem_vnum = vnum - 2;
1418 env.vnum_state = vnum - 1;
1419 /* nvals are vnum - 1, why we indicate with nvals the number
1420 * of memory edges we will produce. For vnum_state we don't
1421 * need to produce a memory edge.*/
1422 env.nvals = vnum - 1;
1425 /* first step: allocate the value arrays for every block */
1426 irg_block_walk_graph(current_ir_graph, NULL, alloc_value_arr, &env);
1428 /* second step: we analyze all calls, that have as parameter scalar(s).
1429 * We mark the calls, that save the address of a scalar and we
1430 * mark the entity owner of this scalar as not optimizeble by now.*/
1431 irg_walk_graph(current_ir_graph, NULL, analyse_calls, &env);
1433 while(env.changes) {
1438 * third step: walk over the blocks of a graph and update
1439 * the information for the access of our scalars.
1441 irg_block_walk_graph(current_ir_graph, NULL, set_block_access, &env);
1445 // if(get_firm_verbosity())
1446 /* Debug info to see if analyse_calls work properly.*/
1447 irg_block_walk_graph(current_ir_graph, NULL, print_block_state, &env);
1450 * fourth step: walk over the graph blockwise in topological order
1451 * and split the memrory edge.
1453 inc_irg_block_visited(current_ir_graph);
1454 irg_walk_blkwise_graph(current_ir_graph, NULL, split_memory_edge, &env);
1458 /* fifth step: fix all nodes, that have as predecessor Unknown.*/
1463 /* sixth step: sync memory enges for the end block.*/
1464 sync_mem_edges(&env);
1466 /*seventh step: free the allocated memory*/
1467 irg_block_walk_graph(current_ir_graph, NULL, free_call_info, &env);
1468 obstack_free(&env.obst, NULL);
1472 * Find possible scalar replacements
1474 * @param irg The current ir graph.
1476 void data_flow_scalar_replacement_opt(ir_graph *irg) {
1482 ent_leaves_t key_leaves, *value_leaves;
1485 if (! get_opt_scalar_replacement())
1488 set_sels = new_set(sels_cmp, 8);
1489 set_ent = new_set(ent_leaves_t_cmp, 8);
1491 /* Call algorithm that remove the critical edges of a ir graph. */
1492 remove_critical_cf_edges(irg);
1494 /* Call algorithm that computes the out edges.*/
1495 assure_irg_outs(irg);
1497 /* Call algorithm that computes the loop information.*/
1498 construct_cf_backedges(irg);
1500 /* Call algorithm that computes the dominance information.*/
1503 /* Find possible scalar replacements */
1504 if (find_possible_replacements(irg)) {
1506 /* Insert in set the scalar replacements. */
1507 irg_frame = get_irg_frame(irg);
1509 for (i = 0 ; i < get_irn_n_outs(irg_frame); i++) {
1510 ir_node *succ = get_irn_out(irg_frame, i);
1513 ir_entity *ent = get_Sel_entity(succ);
1515 if (get_entity_link(ent) == NULL || get_entity_link(ent) == ADDRESS_TAKEN)
1517 /* we have found a entity, that have scalars and we insert it to our set_ent*/
1518 key_leaves.ent = ent;
1519 key_leaves.leaves = new_pset(leave_cmp, 8);
1520 value_leaves = set_insert(set_ent, &key_leaves, sizeof(key_leaves), HASH_PTR(ent));
1522 /* We allocate for every leave sel a vnum.*/
1523 vnum = allocate_value_numbers(set_sels, value_leaves->leaves, ent, vnum);
1527 /* Allocate value number for the globule memory edge.
1528 * and a value number for the value numbers state.*/
1531 /* Allocate value numbers for the entities .*/
1532 for(i = vnum,value_leaves = set_first(set_ent); value_leaves; i++, value_leaves = set_next(set_ent))
1533 SET_ENT_VNUM(value_leaves->ent, i);
1536 do_data_flow_scalar_replacement(set_ent, set_sels, vnum);
1538 /*free the allocated memory.*/
1539 for(value_leaves = set_first(set_ent); value_leaves; value_leaves = set_next(set_ent))
1540 del_pset(value_leaves->leaves);