3 * File name: ir/opt/data_flow_scalar_replace.c
4 * Purpose: scalar replacement of arrays and compounds
5 * Author: Beyhan Veliev
6 * Modified by: Michael Beck
9 * Copyright: (c) 1998-2005 Universität Karlsruhe
10 * Licence: This file protected by GPL - GNU GENERAL PUBLIC LICENSE.
20 #include "data_flow_scalar_replace.h"
31 #include "analyze_irg_args.h"
33 #include "compute_loop_info.h"
37 #define SET_ENT_VNUM(ent, vnum) set_entity_link(ent, INT_TO_PTR(vnum))
38 #define GET_ENT_VNUM(ent) (unsigned)PTR_TO_INT(get_entity_link(ent))
39 #define SET_IRN_VNUM(irn, vnum) set_irn_link(irn, INT_TO_PTR(vnum))
40 #define GET_IRN_VNUM(irn) (unsigned)PTR_TO_INT(get_irn_link(irn))
44 typedef struct _ent_leaves_t{
45 ir_entity *ent; /**< An entity, that contains scalars for replace.*/
46 pset *leaves; /**< All leaves of this entity.*/
49 typedef struct _sels_t {
50 ir_node *sel; /**< A sel node, thats entity have scalars.*/
51 ir_entity *ent; /**< The entity of this sel node.*/
54 typedef struct _call_access_t {
55 ir_node *call; /**< A call node, that have as parameter a scalar.*/
56 unsigned int access_type; /**< The access type, with that this call access this scalar.*/
59 typedef struct _fixlist_entry_t {
60 ir_node *irn; /**< An ir node, that must be fixed.*/
61 unsigned int vnum; /**< The value number, that must became this ir node.*/
64 typedef struct _syncs_fixlist_entry_t {
65 ir_node *irn; /**< A sync node that must be fixed.*/
66 int *accessed_vnum; /**< A pointer to save an array with value numbers, that must became this sync.*/
67 }syncs_fixlist_entry_t;
69 /* A entry, that save the memory
70 * edge state and the access state for this leave
71 * int the array,that is created for every block.*/
72 typedef struct _leave_t {
73 ir_node *mem_edge_state; /**< memory state for this scalar in this block.*/
74 unsigned int access_type; /**< access state for this scalar in this block.*/
75 set *calls; /**< call nodes,that change this scalar in this block.*/
79 * A path element entry: it is either an entity
80 * or a tarval, because we evaluate only constant array
81 * accesses like a.b.c[8].d
89 * An access path, used to assign value numbers
90 * to variables that will be scalar replaced
92 typedef struct _path_t {
93 unsigned vnum; /**< the value number */
94 unsigned path_len; /**< the length of the access path */
95 path_elem_t path[1]; /**< the path */
99 * environment for memory walker
101 typedef struct _env_t {
102 struct obstack obst; /**< a obstack for the memory edge */
103 set *set_sels; /**< a set with all sels, that are reachable from an entity with a scalar.*/
104 set *set_ent; /**< a set with all entities that have one or more scalars.*/
105 fixlist_entry_t *fix_phis; /**< list of all Phi nodes that must be fixed */
106 fixlist_entry_t *fix_ls; /**< list of all Load or Store nodes that must be fixed */
107 syncs_fixlist_entry_t *fix_syncs; /**< list of all Sync nodes that must be fixed */
108 unsigned int nvals; /**< to save the number of scalars.*/
109 unsigned int gl_mem_vnum; /**< indicate the position of the globule memory edge state in var_arr.*/
110 unsigned int vnum_state; /**< indicate the position of the value number state in var_arr.*/
111 unsigned int changes; /**< to save if by anlyse_calls is changed anything.*/
117 * Compare two elements of the ent_leaves_t set.
119 * @return 0 if they are identically
121 static int ent_leaves_t_cmp(const void *elt, const void *key, size_t size)
123 const ent_leaves_t *c1 = elt;
124 const ent_leaves_t *c2 = key;
126 return c1->ent != c2->ent;
130 * Compare two elements of the ent_access_t set.
132 * @return 0 if they are identically
134 static int ent_cmp(const void *elt, const void *key)
136 const ir_entity *c1 = elt;
137 const ir_entity *c2 = key;
143 * Compare two elements of the sels_t set.
145 * @return 0 if they are identically
147 static int sels_cmp(const void *elt, const void *key, size_t size)
149 const sels_t *c1 = elt;
150 const sels_t *c2 = key;
152 return c1->sel != c2->sel;
156 * Compare two elements of the leave_t set.
158 * @return 0 if they are identically
160 static int leave_cmp(const void *elt, const void *key)
162 ir_node *c1 = (ir_node *)elt;
163 ir_node *c2 = (ir_node *)key;
165 return get_Sel_entity(c1) != get_Sel_entity(c2);
169 * Compare two elements of the call_access_t set.
171 * @return 0 if they are identically
173 static int call_cmp(const void *elt, const void *key, size_t size)
175 const call_access_t *c1 = elt;
176 const call_access_t *c2 = key;
178 return c1->call != c2->call;
184 * @return 0 if they are identically
186 static int path_cmp(const void *elt, const void *key, size_t size)
188 const path_t *p1 = elt;
189 const path_t *p2 = key;
191 /* we can use memcmp here, because identical tarvals should have identical addresses */
192 return memcmp(p1->path, p2->path, p1->path_len * sizeof(p1->path[0]));
196 * Calculate a hash value for a path.
198 static unsigned path_hash(const path_t *path)
203 for (i = 0; i < path->path_len; ++i)
204 hash ^= (unsigned)PTR_TO_INT(path->path[i].ent);
210 * Returns non-zero, if all induces of a Sel node are constants.
212 * @param sel the Sel node that will be checked
214 static int is_const_sel(ir_node *sel) {
215 int i, n = get_Sel_n_indexs(sel);
217 for (i = 0; i < n; ++i) {
218 ir_node *idx = get_Sel_index(sel, i);
220 if (get_irn_op(idx) != op_Const)
227 * Returns non-zero, if the address of an entity
228 * represented by a Sel node (or it's successor Sels) is taken.
230 static int is_address_taken(ir_node *sel)
234 if (! is_const_sel(sel))
237 for (i = get_irn_n_outs(sel) - 1; i >= 0; --i) {
238 ir_node *succ = get_irn_out(sel, i);
240 switch (get_irn_opcode(succ)) {
242 /* ok, we just load from that entity */
246 /* check that Sel is not the Store's value */
247 if (get_Store_value(succ) == sel)
252 /* Check the Sel successor of Sel */
253 int res = is_address_taken(succ);
261 /* The address of an entity is given as a parameter.
262 * We analyzes that later and optimizes this scalar
268 /* another op, the address is taken */
276 * Link all Sels with the entity.
278 * @param ent the entity that will be scalar replaced
279 * @param sel a Sel node that selects some fields of this entity
281 static void link_all_leave_sels(ir_entity *ent, ir_node *sel)
285 n = get_irn_n_outs(sel);
286 for (i = 0; i < n; ++i) {
287 ir_node *succ = get_irn_out(sel, i);
289 if (get_irn_op(succ) == op_Sel)
290 link_all_leave_sels(ent, succ);
294 /* if Sel nodes with memory inputs are used, a entity can be
295 * visited more than once causing a ring here, so we use the
296 * node flag to mark linked nodes
298 if (irn_visited(sel))
302 * we link the sels to the entity.
304 set_irn_link(sel, get_entity_link(ent));
305 set_entity_link(ent, sel);
307 mark_irn_visited(sel);
310 /* we need a special address that serves as an address taken marker */
312 static void *ADDRESS_TAKEN = &_x;
315 * Find possible scalar replacements.
317 * @param irg an IR graph
319 * This function finds variables on the (members of the) frame type
320 * that can be scalar replaced, because their address is never taken.
321 * If such a variable is found, it's entity link will hold a list of all
322 * Sel nodes, that selects anythings of this entity.
323 * Otherwise, the link will be ADDRESS_TAKEN or NULL.
325 * @return non-zero if at least one entity could be replaced
328 static int find_possible_replacements(ir_graph *irg)
330 ir_node *irg_frame = get_irg_frame(irg);
334 inc_irg_visited(irg);
336 n = get_irn_n_outs(irg_frame);
339 * First, clear the link field of all interestingentities.
340 * Note that we did not rely on the fact that there is only
341 * one Sel node per entity, so we might access one entity
342 * more than once here.
343 * That's why we have need two loops.
345 for (i = 0; i < n; ++i) {
346 ir_node *succ = get_irn_out(irg_frame, i);
348 if (get_irn_op(succ) == op_Sel) {
349 ir_entity *ent = get_Sel_entity(succ);
350 set_entity_link(ent, NULL);
355 * Check the ir_graph for Sel nodes. If the entity of Sel
356 * isn't a scalar replacement set the link of this entity
357 * equal ADDRESS_TAKEN.
359 for (i = 0; i < n; ++i) {
360 ir_node *succ = get_irn_out(irg_frame, i);
362 if (get_irn_op(succ) == op_Sel) {
363 ir_entity *ent = get_Sel_entity(succ);
366 if (get_entity_link(ent) == ADDRESS_TAKEN)
370 * Beware: in rare cases even entities on the frame might be
371 * volatile. This might happen if the entity serves as a store
372 * to a value that must survive a exception. Do not optimize
373 * such entities away.
375 if (get_entity_volatility(ent) == volatility_is_volatile) {
376 set_entity_link(ent, ADDRESS_TAKEN);
380 ent_type = get_entity_type(ent);
382 /* we can handle arrays, structs and atomic types yet */
383 if (is_Array_type(ent_type) || is_Struct_type(ent_type) || is_atomic_type(ent_type)) {
384 if (is_address_taken(succ)) {
385 if (get_entity_link(ent)) /* killing one */
387 set_entity_link(ent, ADDRESS_TAKEN);
390 /* possible found one */
391 if (get_entity_link(ent) == NULL)
393 link_all_leave_sels(ent, succ);
402 static int is_leave_sel(ir_node *sel) {
406 for(i = get_irn_n_outs(sel) - 1; i >= 0; i--) {
407 succ = get_irn_out(sel, i);
408 if(get_irn_op(succ) == op_Sel)
416 * Return a path from the Sel node sel to it's root.
418 * @param sel the Sel node
419 * @param len the length of the path so far
421 static path_t *find_path(ir_node *sel, unsigned len)
425 ir_node *pred = get_Sel_ptr(sel);
427 /* the current Sel node will add some path elements */
428 n = get_Sel_n_indexs(sel);
431 if (get_irn_op(pred) != op_Sel) {
432 /* we found the root */
434 res = xmalloc(sizeof(*res) + (len - 1) * sizeof(res->path));
438 res = find_path(pred, len);
440 pos = res->path_len - len;
442 res->path[pos++].ent = get_Sel_entity(sel);
443 for (i = 0; i < n; ++i) {
444 ir_node *index = get_Sel_index(sel, i);
446 if(get_irn_op(index) == op_Const)
447 res->path[pos++].tv = get_Const_tarval(index);
453 * Allocate value numbers for the leaves
454 * in our found entities.
456 * @param sels a set that will contain all Sels that have a value number
457 * @param ent the entity that will be scalar replaced
458 * @param vnum the first value number we can assign
459 * @param modes a flexible array, containing all the modes of
462 * @return the next free value number
464 static unsigned allocate_value_numbers(set *set_sels, pset *leaves, ir_entity *ent, unsigned vnum)
469 set *pathes = new_set(path_cmp, 8);
471 /* visit all Sel nodes in the chain of the entity */
472 for (sel = get_entity_link(ent); sel; sel = next) {
473 next = get_irn_link(sel);
475 /* we save for every sel it root entity, why
476 * we need this information, when we split the memory edge,
477 * and we must mark this sel for later. */
480 set_insert(set_sels, &key_sels, sizeof(key_sels), HASH_PTR(sel));
482 if(! is_leave_sel(sel))
484 /* We have found a leave and we add it to the pset of this entity.*/
485 pset_insert(leaves, sel, HASH_PTR(get_Sel_entity(sel)));
487 key = find_path(sel, 0);
488 path = set_find(pathes, key, sizeof(*key) + sizeof(key->path[0]) * key->path_len, path_hash(key));
491 SET_IRN_VNUM(sel, path->vnum);
496 set_insert(pathes, key, sizeof(*key) + sizeof(key->path[0]) * key->path_len, path_hash(key));
498 SET_IRN_VNUM(sel, key->vnum);
504 set_entity_link(ent, NULL);
508 * Add a sync node to it fix list.
510 * @param sync The sync node, that myst be addet to the fix list.
511 * @param unk_vnum An array whit the value number, that are synced with this sync node.
512 * @param env The enviroment pinter.
514 static void add_sync_to_fixlist(ir_node *sync, int *unk_vnum, env_t *env) {
516 syncs_fixlist_entry_t *s;
518 s = obstack_alloc(&env->obst, sizeof(*s));
520 s->accessed_vnum = unk_vnum;
521 set_irn_link(sync, env->fix_syncs);
525 * Add a ir node to it fix list.
527 * @param irn The ir node, that myst be addet to the fix list.
528 * @param vnum The value number, that must baceme this ir node as predecessor later.
529 * @param env The enviroment pinter.
531 static void add_ls_to_fixlist(ir_node *irn, int vnum, env_t *env) {
535 l = obstack_alloc(&env->obst, sizeof(*l));
539 if(get_irn_op(irn) == op_Phi) {
540 set_irn_link(l->irn, env->fix_phis);
543 set_irn_link(l->irn, env->fix_ls);
548 static void add_mem_edge(value_arr_entry_t *val_arr, int vnum, ir_node ***in, int **accessed_vnum) {
550 if(val_arr[vnum].mem_edge_state != NULL)
551 ARR_APP1(ir_node *, *in, val_arr[vnum].mem_edge_state);
553 ARR_APP1(int, *accessed_vnum, vnum);
554 ARR_APP1(ir_node *, *in, new_Unknown(mode_M));
558 * The function handles the scalars, that wase stored
561 * @param blk The block, that must be handled.
562 * @param env The enviroment pinter.
565 /* Return the memory successor of the call node.*/
566 static ir_node *get_Call_mem_out(ir_node *call) {
571 for(i = get_irn_n_outs(call) - 1; i >= 0; i--) {
572 mem = get_irn_out(call, i);
573 if(get_irn_mode(mem) == mode_M)
576 /* is not reachable*/
581 static void sync_stored_scalars(ir_node *blk, env_t *env) {
584 int *unk_vnum; /**< An arraw, where are saved the value number, that
585 are synced from this sync node.*/
586 ent_leaves_t *value_ent;
587 value_arr_entry_t *val_arr_blk, *val_arr;
588 ir_node *pred, *leave, *sync, **in;
589 ir_node *sync_blk; /**< The block, where the sync node must be created.*/
592 val_arr_blk = get_irn_link(blk);
594 for(value_ent = set_first(env->set_ent); value_ent; value_ent = set_next(env->set_ent)) {
597 if(val_arr_blk[GET_ENT_VNUM(value_ent->ent)].access_type <= 3)
598 /* This entity is not stored in this block.*/
601 for(i = get_Block_n_cfgpreds(blk) - 1; i >= 0; i--) {
603 pred = get_Block_cfgpred(blk, i);
604 pred = get_nodes_block(pred);
605 val_arr = get_irn_link(pred);
607 if(val_arr[GET_ENT_VNUM(value_ent->ent)].access_type == SYNCED)
608 /* This entity was synced.*/
611 if(val_arr[GET_ENT_VNUM(value_ent->ent)].access_type <= 3) {
613 /* To avoid repeated sync of this entity in this block.*/
614 val_arr[GET_ENT_VNUM(value_ent->ent)].access_type = SYNCED;
615 /* In this predecessor block is this entity not acessed.
616 * We must sync in the end ot this block.*/
617 if(get_Block_n_cfgpreds(blk) > 1)
618 sync_blk = get_nodes_block(get_Block_cfgpred(blk, i));
622 val_arr = get_irn_link(sync_blk);
623 /* An array to save the memory edges, that must be
625 in = NEW_ARR_F(ir_node *, 1);
627 /* An array to save the value numbers,
628 * that must be repaired.*/
629 unk_vnum = NEW_ARR_F(int, 0);
630 /* The global memory edge.*/
631 if(val_arr[env->gl_mem_vnum].mem_edge_state == NULL)
632 in[0] = new_Unknown(mode_M);
634 in[0] = val_arr[env->gl_mem_vnum].mem_edge_state;
636 for(leave = pset_first(value_ent->leaves); leave; leave = pset_next(value_ent->leaves))
637 /* All this memory edges must be synced.*/
638 add_mem_edge(val_arr, GET_IRN_VNUM(leave), &in, &unk_vnum);
640 /* We create the sync and set it in the global memory state.*/
641 sync = new_r_Sync(current_ir_graph, sync_blk, ARR_LEN(in), in);
642 /* We must check this, why it is possible to get a Bad node
643 * form new_r_Sync(), when the node can be optimized.
644 * In this case we must do nothing.*/
645 if(get_irn_op(sync) == op_Sync) {
646 val_arr[env->gl_mem_vnum].mem_edge_state = sync;
647 /* We add this sync node to the sync's fix list.*/
648 add_sync_to_fixlist(val_arr[env->gl_mem_vnum].mem_edge_state, unk_vnum, env);
656 * The function split the memory edge of load and store nodes, that have
657 * as predecessor a scalar
659 * @param irn The node, that memory edge must be spleted.
660 * @param env The enviroment pinter.
662 static void split_ls_mem_edge(ir_node *irn, env_t *env) {
665 ir_node *leave, *irn_blk, *mem_state, *new_mem_state;
666 unsigned ent_vnum, sel_vnum, i;
667 value_arr_entry_t *val_arr;
668 sels_t key_sels, *value_sels;
669 ent_leaves_t key_ent, *value_ent;
671 op = get_irn_op(irn);
674 key_sels.sel = get_Load_ptr(irn);
676 key_sels.sel = get_Store_ptr(irn);
678 value_sels = set_find(env->set_sels, &key_sels, sizeof(key_sels), HASH_PTR(key_sels.sel));
680 if(value_sels != NULL) {
681 /* we have found a load or store, that use a sel of our set
682 * and we must split or extend, if the memory edge have been
683 * split for this sel, the memory edge.*/
685 key_ent.ent = value_sels->ent;
686 value_ent = set_find(env->set_ent, &key_ent, sizeof(key_ent), HASH_PTR(key_ent.ent));
687 /*To check if the enities set is right filled. */
688 assert(value_ent && " This sel's entity isn't int the entity set.");
690 leave = pset_find(value_ent->leaves, key_sels.sel, HASH_PTR(get_Sel_entity(key_sels.sel)));
691 /*To check if the leaves set is right filled. */
692 assert(leave && "Anything in data_flow_scalar_replacment algorithm is wrong.");
694 ent_vnum = GET_ENT_VNUM(value_ent->ent);
695 sel_vnum = GET_IRN_VNUM(leave);
696 irn_blk = get_nodes_block(irn);
697 val_arr = get_irn_link(irn_blk);
699 if(val_arr[ent_vnum].access_type == 0)
700 /* We have found a scalar, that address is not stored as jet.*/
703 /* This scalar have been stored.*/
704 i = env->gl_mem_vnum;
706 if(val_arr[i].mem_edge_state == NULL) {
707 /* We split now for this sel the memory edge in this block.*/
708 mem_state = new_Unknown(mode_M);
709 /* We must mark this node to fix later*/
710 add_ls_to_fixlist(irn, i, env);
713 /* We have split the memory edge and the current state is saved.*/
714 mem_state = val_arr[i].mem_edge_state;
716 /* We set this Load or Store to the memory edge of this
719 set_Load_mem(irn, mem_state);
721 set_Store_mem(irn, mem_state);
723 /* When we have split or extended the memory edge we must
724 * update the memory_edge_state of this sel*/
725 new_mem_state = get_irn_out(irn, 0);
726 if(get_irn_mode(new_mem_state) == mode_M)
727 val_arr[i].mem_edge_state = new_mem_state;
729 val_arr[i].mem_edge_state = get_irn_out(irn, 1);
734 * The function split the memory edge of phi nodes, that have
735 * as predecessor a scalar
737 * @param irn The phi node, that memory edge must be spleted.
738 * @param env The enviroment pinter.
740 static void split_phi_mem_edge(ir_node *irn, env_t *env) {
742 ir_node *irn_blk, *unk, *leave, **in;
744 ent_leaves_t *value_ent;
745 value_arr_entry_t *val_arr;
747 irn_blk = get_nodes_block(irn);
748 val_arr = get_irn_link(irn_blk);
750 n = get_Block_n_cfgpreds(irn_blk);
752 in = alloca(sizeof(*in) * n);
754 for(value_ent = set_first(env->set_ent); value_ent; value_ent = set_next(env->set_ent))
755 if(val_arr[GET_ENT_VNUM(value_ent->ent)].access_type < 3)
756 /* This scalar wasn't be saved and we need to produce a phi for it.*/
757 for(leave = pset_first(value_ent->leaves); leave; leave = pset_next(value_ent->leaves)){
759 unk = new_Unknown(mode_M);
760 for (j = n - 1; j >= 0; --j)
763 val_arr[GET_IRN_VNUM(leave)].mem_edge_state = new_r_Phi(current_ir_graph, irn_blk, n, in, mode_M);
765 add_ls_to_fixlist(val_arr[GET_IRN_VNUM(leave)].mem_edge_state, GET_IRN_VNUM(leave), env);
768 /* We use for the global memory the phi node, that
769 * is already available.*/
770 val_arr[env->gl_mem_vnum].mem_edge_state = irn;
774 * The function handles the call nodes, that have
775 * as parameter a scalar
777 * @param env The enviroment pinter.
778 * @param call The call node, that must be handled.
779 * @param accessed_entities A set wit all entities, that are accessed from this call node.*/
780 static void split_call_mem_edge(env_t *env, ir_node *call, pset *accessed_entities) {
782 ent_leaves_t key_ent, *value_ent;
783 value_arr_entry_t *val_arr;
784 call_access_t key_call, *value_call;
785 ir_node *call_blk, *new_mem_state, *leave;
789 int fix_irn = 0; /**< Set to 1 if we must add this call to it fix list.*/
790 int *accessed_leaves_vnum = NULL; /**< An arraw, where are saved the value number, that
791 are synced from call's sync node, if we need it.*/
793 if(get_irn_node_nr(call) == 2763)
796 call_blk = get_nodes_block(call);
797 val_arr = get_irn_link(call_blk);
798 /* An array to save the memory edges, that must be
800 in = NEW_ARR_F(ir_node *, 1);
801 /* An array to save the value numbers of the memory
802 * edges that must be repaired.*/
803 accessed_leaves_vnum = NEW_ARR_F(int, 0);
805 /* We get the memory successor of the call node.
806 * It is the new memory state for all synced memory
808 new_mem_state = get_Call_mem_out(call);
810 /* The global memory is the first predecessor of the create sync node.*/
811 if(val_arr[env->gl_mem_vnum].mem_edge_state == NULL) {
812 in[0] = new_Unknown(mode_M);
816 in[0] = val_arr[env->gl_mem_vnum].mem_edge_state;
819 for(ent = pset_first(accessed_entities); ent; ent = pset_next(accessed_entities)) {
820 /* Whit this loop we iterate all accessed entities from this call and collect
821 * all memory edges, that we must sync.*/
822 ent_vnum = GET_ENT_VNUM(ent);
824 key_call.call = call;
825 value_call = set_find(val_arr[ent_vnum].calls, &key_call, sizeof(key_call), HASH_PTR(key_call.call));
828 value_ent = set_find(env->set_ent, &key_ent, sizeof(key_ent), HASH_PTR(key_ent.ent));
830 if(val_arr[ent_vnum].access_type <= 3) {
831 /* This scalar's address wasn't stored in this block.*/
832 switch(value_call->access_type) {
834 case ptr_access_none :
835 /* In this case we have nothing to do.*/
838 case ptr_access_read:
839 case ptr_access_write:
841 /* All this cases must be traded equal.*/
843 for(leave = pset_first(value_ent->leaves); leave; leave = pset_next(value_ent->leaves)){
844 /* All this memory edges must be synced.*/
845 add_mem_edge(val_arr, GET_IRN_VNUM(leave), &in, &accessed_leaves_vnum);
847 /* We update the memory state of this leave.*/
848 if(value_call->access_type != ptr_access_read)
849 val_arr[GET_IRN_VNUM(leave)].mem_edge_state = new_mem_state;
858 /* We must update the global memory state.*/
859 val_arr[env->gl_mem_vnum].mem_edge_state = new_mem_state;
861 if(ARR_LEN(in) == 1) {
862 /* we must set the call memory to gobale momory*/
863 set_Call_mem(call,in[0]);
866 /* We add this call node to the call fix list..*/
867 add_ls_to_fixlist(call, env->gl_mem_vnum, env);
870 /* We create the sync and set it as memory predecessor of the call node.*/
871 sync = new_r_Sync(current_ir_graph, call_blk, ARR_LEN(in), in);
872 /* We must check this, why it is possible to get a Bad node
873 * form new_r_Sync(), when the node can be optimized.
874 * In this case we must do nothing.*/
875 if(get_irn_op(sync) == op_Sync) {
877 set_Call_mem(call, sync);
878 if(ARR_LEN(accessed_leaves_vnum))
879 /* We add this sync node to the sync's fix list.*/
880 add_sync_to_fixlist(sync, accessed_leaves_vnum, env);
887 * The function split the memory edge from the passed
888 * ir node if this is needed
890 * @param irn The node, that memory edge must be spleted.
891 * @param env The enviroment pinter.
893 static void split_memory_edge(ir_node *irn, void *ctx) {
896 ir_node *sel, *irn_blk;
898 sels_t key_sels, *value_sels;
899 value_arr_entry_t *val_arr;
900 pset *accessed_entities; /**< A set to save all entities accessed from a call.*/
904 op = get_irn_op(irn);
909 irn_blk = get_nodes_block(irn);
911 if (Block_not_block_visited(irn_blk)) {
912 /* We sync first the stored scalar address in this block.*/
913 mark_Block_block_visited(irn_blk);
914 sync_stored_scalars(irn_blk, env);
917 if(op == op_Load || op == op_Store)
919 split_ls_mem_edge(irn, env);
922 if (op == op_Phi && get_irn_mode(irn) == mode_M) {
924 * found a memory Phi: Here, we must create new Phi nodes
926 split_phi_mem_edge(irn, env);
931 /* Calls that have a NoMem input do neither read nor write memory.
932 We can completely ignore them here. */
933 if (get_irn_op(get_Call_mem(irn)) == op_NoMem)
936 /* We save in this set all entities,
937 * that are accessed from this call node.*/
938 accessed_entities = new_pset(ent_cmp, 8);
939 val_arr = get_irn_link(get_nodes_block(irn));
941 for ( i = get_Call_n_params(irn) - 1; i >= 0; i--) {
943 sel = get_Call_param(irn, i);
945 if(get_irn_op(sel) == op_Sel) {
947 value_sels = set_find(env->set_sels, &key_sels, sizeof(key_sels), HASH_PTR(key_sels.sel));
949 if(value_sels != NULL && val_arr[GET_ENT_VNUM(value_sels->ent)].access_type <= 3)
950 /* We save in this set all accessed entities from this call node whit
951 * access none, read, write or rw..*/
952 pset_insert(accessed_entities, value_sels->ent, HASH_PTR(value_sels->ent));
956 if(pset_count(accessed_entities))
957 split_call_mem_edge(env, irn, accessed_entities);
959 del_pset(accessed_entities);
966 * searches through blocks beginning from block for value
967 * vnum and return it.
969 * @param block A block from the current ir graph.
970 * @param vnum The value number, that must be found.
972 static ir_node *find_vnum_value(ir_node *block, unsigned vnum)
974 value_arr_entry_t *val_arr;
978 if (Block_not_block_visited(block)) {
979 mark_Block_block_visited(block);
981 val_arr = get_irn_link(block);
983 if (val_arr[vnum].mem_edge_state)
984 return val_arr[vnum].mem_edge_state;
986 for (i = get_Block_n_cfgpreds(block) - 1; i >= 0; --i) {
987 ir_node *pred = get_Block_cfgpred(block, i);
989 res = find_vnum_value(get_nodes_block(pred), vnum);
998 * fix the Load/Store or Call list
1000 * @param The enviroment pinter.
1002 static void fix_ls(env_t *env)
1005 ir_node *irn, *block, *pred, *val = NULL;
1009 for (l = env->fix_ls; l; l = get_irn_link(irn)) {
1012 op = get_irn_op(irn);
1013 block = get_nodes_block(irn);
1014 for (i = get_Block_n_cfgpreds(block) - 1; i >= 0; --i) {
1015 pred = get_Block_cfgpred(block, i);
1016 pred = get_nodes_block(pred);
1018 inc_irg_block_visited(current_ir_graph);
1019 val = find_vnum_value(pred, l->vnum);
1027 set_Store_mem(irn, val);
1030 set_Load_mem(irn, val);
1032 set_Call_mem(irn, val);
1040 * @param The enviroment pinter.
1042 static void fix_phis(env_t *env)
1045 ir_node *phi, *block, *pred, *val;
1048 for (l = env->fix_phis; l; l = get_irn_link(phi)) {
1051 block = get_nodes_block(phi);
1052 for (i = get_Block_n_cfgpreds(block) - 1; i >= 0; --i) {
1054 pred = get_Block_cfgpred(block, i);
1055 pred = get_nodes_block(pred);
1057 inc_irg_block_visited(current_ir_graph);
1058 val = find_vnum_value(pred, l->vnum);
1061 set_irn_n(phi, i, val);
1070 * @param The enviroment pinter.
1072 static void fix_syncs(env_t *env)
1074 syncs_fixlist_entry_t *l;
1075 ir_node *sync, *block, *pred, *val;
1079 for (l = env->fix_syncs; l; l = get_irn_link(sync)) {
1083 /* The sync block must have one predecessor, when it
1084 have unknown nodes as predecessor.*/
1085 block = get_nodes_block(sync);
1086 pred = get_Block_cfgpred(block, 0);
1087 pred = get_nodes_block(pred);
1089 /* We first repair the global memory edge at the first position of sync predecessors.*/
1090 if(get_irn_op(get_irn_n(sync, 0)) == op_Unknown) {
1091 inc_irg_block_visited(current_ir_graph);
1092 val = find_vnum_value(pred, env->gl_mem_vnum);
1095 set_irn_n(sync, 0, val);
1098 for (i = get_irn_arity(sync) - 1; i >= 1; --i) {
1099 /* We repair the leaves*/
1101 assert(k <= ARR_LEN(l->accessed_vnum) && "The algorythm for sync repair is wron");
1102 if(get_irn_op(get_irn_n(sync, i)) == op_Unknown) {
1103 inc_irg_block_visited(current_ir_graph);
1104 val = find_vnum_value(pred, l->accessed_vnum[k++]);
1107 set_irn_n(sync, i, val);
1110 DEL_ARR_F(l->accessed_vnum);
1114 * For the end node we must sync all memory edges.
1116 * @param The enviroment pinter.
1118 static void sync_mem_edges(env_t *env) {
1120 value_arr_entry_t *val_arr;
1121 ir_node **in, *sync, *Return, *Return_blk;
1122 int i, vnum, vnum_state;
1124 Return = get_Block_cfgpred(get_irg_end_block(current_ir_graph), 0);
1125 Return_blk = get_nodes_block(Return);
1126 val_arr = get_irn_link(Return_blk);
1130 for(i = 0; i <= (int)env->gl_mem_vnum; i++)
1131 /* we get the current state of non saved scalars.*/
1132 if(val_arr[i].access_type <= 3)
1135 /* We allocate the memory, that we need for the predecessors of the sync.*/
1136 in = xmalloc(sizeof(ir_node*) *vnum_state);
1138 /* The global memory edge is the first predecessor of this sync node.*/
1139 if(val_arr[env->gl_mem_vnum].mem_edge_state == NULL) {
1140 /* We must search through blocks for this memory state.*/
1141 inc_irg_block_visited(current_ir_graph);
1142 in[0] = find_vnum_value(Return_blk, env->gl_mem_vnum);
1145 in[0] = val_arr[env->gl_mem_vnum].mem_edge_state;
1148 for(i = 1, vnum = 0; vnum < (int)env->gl_mem_vnum; vnum++) {
1150 if(val_arr[vnum].access_type <= 3) {
1151 /* we add the non saved scalars as predecessors of the sync.*/
1153 if(val_arr[vnum].mem_edge_state == NULL) {
1154 /* We must search through blocks for this memory state.*/
1155 inc_irg_block_visited(current_ir_graph);
1156 in[i] = find_vnum_value(Return_blk, vnum);
1159 in[i] = val_arr[vnum].mem_edge_state;
1164 sync = new_r_Sync(current_ir_graph, Return_blk, vnum_state, in);
1165 set_Return_mem(Return, sync);
1171 * Walker: allocate the value array for every block.
1173 * @param block A block from the current ir graph for that must be allocated a value array.
1174 * @param ctx The enviroment pinter.
1176 static void alloc_value_arr(ir_node *block, void *ctx)
1181 value_arr_entry_t *var_arr = obstack_alloc(&env->obst, sizeof(value_arr_entry_t) *(env->nvals + set_count(env->set_ent) + 1));
1183 /* the value array is empty at start */
1184 memset(var_arr, 0, sizeof(value_arr_entry_t) * (env->nvals + set_count(env->set_ent) + 1));
1185 set_irn_link(block, var_arr);
1187 /* We set the block value number state to optimal and later we update this.*/
1188 var_arr[env->vnum_state].access_type = env->nvals;
1190 if(get_irg_start_block(current_ir_graph) == block)
1191 /* We initilize the startblocks array with the irg initilize memory, why
1192 * it must be the start point of all memory edges.*/
1193 for(i = (env->nvals + set_count(env->set_ent)) ; i >=0; i--)
1194 var_arr[i].mem_edge_state = get_irg_initial_mem(current_ir_graph);
1198 /* Analyze call nodes to get information, if they store the address of a scalar.
1200 * @param *irn An ir node from the current_ir_graph.
1201 * @param *env The enviroment pointer.
1203 static void analyse_calls(ir_node *irn, void *ctx) {
1206 unsigned int acces_type;
1207 ir_node *param, *call_ptr, *blk;
1209 ir_entity *meth_ent;
1210 sels_t key_sels, *value_sels;
1211 call_access_t key_call, *value_call;
1212 value_arr_entry_t *val_arr;
1216 if(get_irn_op(irn) != op_Call)
1219 /* Calls that have a NoMem input do neither read nor write memory.
1220 We can completely ignore them here. */
1221 if (get_irn_op(get_Call_mem(irn)) == op_NoMem)
1224 /* We iterate over the parameters of this call nodes.*/
1225 for ( i = get_Call_n_params(irn) - 1; i >= 0; i--) {
1226 param = get_Call_param(irn, i);
1227 if(get_irn_op(param) == op_Sel) {
1228 /* We have found a parameter with operation sel.*/
1229 key_sels.sel = param;
1230 value_sels = set_find(env->set_sels, &key_sels, sizeof(key_sels), HASH_PTR(key_sels.sel));
1231 if(value_sels != NULL ) {
1233 /* We have found a call, that have as parameter a sel from our set_sels.*/
1234 call_ptr = get_Call_ptr(irn);
1235 op = get_irn_op(call_ptr);
1237 if(op == op_SymConst && get_SymConst_kind(call_ptr) == symconst_addr_ent) {
1238 meth_ent = get_SymConst_entity(call_ptr);
1239 /* we get the access type for our sel.*/
1240 acces_type = get_method_param_access(meth_ent, i);
1242 /* We can't analyze this function and we asume, that it store the address.*/
1243 acces_type = ptr_access_store;
1245 /* we save the access type and this call in the array allocated for this block.
1246 * The value number of this entity get us the position in the array to save this
1247 * information. Why we expect more calls as one we allocate a set.*/
1248 vnum = GET_ENT_VNUM(value_sels->ent);
1249 blk = get_nodes_block(irn);
1250 val_arr = get_irn_link(blk);
1252 if(val_arr[vnum].access_type > 3)
1253 /* The address of this entity have been stored.*/
1256 if(val_arr[vnum].calls == NULL)
1257 /* for this entity i have found the firs call in this block and we must allocate the set.*/
1258 val_arr[vnum].calls = new_set(call_cmp, 8);
1260 /* This call performs anything with the scalar and we must mark it.*/
1261 key_call.call = irn;
1262 key_call.access_type = acces_type;
1263 value_call = set_insert(val_arr[vnum].calls, &key_call, sizeof(key_call), HASH_PTR(key_call.call));
1265 if(value_call->access_type < acces_type)
1266 /* this case tread, when a call access an entity more at once.
1267 * Than we must save the highest access type.*/
1268 value_call->access_type = acces_type;
1271 /* This call save the address of our scalar and we can't
1272 * use the scalars of this entity for optimization as from now.
1274 val_arr[vnum].access_type = acces_type;
1280 static int set_block_dominated_first_access(ir_node *blk, int vnum, unsigned int access) {
1282 ir_node *idom, *succ;
1283 value_arr_entry_t *val_arr;
1286 idom = get_Block_idom(blk);
1287 for(i = get_Block_n_cfg_outs(idom) - 1; i >=1; i--) {
1288 succ = get_Block_cfg_out(idom, i);
1289 val_arr = get_irn_link(succ);
1290 if(val_arr[vnum].access_type < 3) {
1291 val_arr[vnum].access_type = access;
1297 /* Update the access information of a block if a predecessor of
1298 * this black have a higher access for an entity.
1300 * @param *irn An ir node from the current_ir_graph.
1301 * @param *env The enviroment pointer.
1303 static void set_block_access(ir_node *irn, void *ctx){
1305 value_arr_entry_t *val_arr, *val_arr_pred;
1306 ent_leaves_t *value_leaves;
1307 ir_node *pred, *pred_blk, *leave;
1312 val_arr = get_irn_link(irn);
1314 for( i = get_Block_n_cfgpreds(irn) - 1; i >= 0; i--) {
1315 /* We analyze the predecessors of this block to see if this block must
1317 pred = get_Block_cfgpred(irn, i);
1318 pred_blk = get_nodes_block(pred);
1320 val_arr_pred = get_irn_link(pred_blk);
1322 for(value_leaves = set_first(env->set_ent); value_leaves; value_leaves = set_next(env->set_ent)) {
1323 vnum = GET_ENT_VNUM(value_leaves->ent);
1325 if((get_Block_n_cfgpreds(irn) > 1) && (val_arr[vnum].access_type > 3))
1326 env->changes = set_block_dominated_first_access(irn, vnum, val_arr[vnum].access_type);
1328 if((val_arr_pred[vnum].access_type > 3) && (val_arr[vnum].access_type < 3)) {
1329 /* We have found a block for update it access and value number information.*/
1330 val_arr[vnum].access_type = val_arr_pred[vnum].access_type;
1331 /* We update the access information of all leave, that belong to
1334 for(leave = pset_first(value_leaves->leaves); leave; leave = pset_next(value_leaves->leaves))
1335 val_arr[GET_IRN_VNUM(leave)].access_type = val_arr[vnum].access_type;
1337 /* In this way can't be got the actuall number of value numbers.
1338 val_arr[env->vnum_state].access_type = val_arr_pred[env->vnum_state].access_type; */
1344 /* Free the allocated call sets.
1346 * @param irn A block form the ir graph.
1347 * @param env The enviroment pinter.
1349 static void free_call_info(ir_node *irn, void *ctx) {
1353 value_arr_entry_t *val_arr;
1356 val_arr = get_irn_link(irn);
1358 for(i = env->nvals + set_count(env->set_ent); i >= 0; i--) {
1359 if(val_arr[i].calls != NULL)
1361 del_set(val_arr[i].calls);
1365 static void print_block_state(ir_node *irn, void *ctx) {
1367 value_arr_entry_t *val_arr;
1368 ent_leaves_t *value_leaves;
1369 call_access_t *value_calls;
1374 val_arr = get_irn_link(irn);
1375 ir_printf("\n\nThe actual value number state of this block is: %i \n",
1376 val_arr[env->vnum_state].access_type - 1);
1378 for(value_leaves = set_first(env->set_ent); value_leaves; value_leaves = set_next(env->set_ent)) {
1380 vnum = GET_ENT_VNUM(value_leaves->ent);
1381 ir_printf("The entity %F access type in the block with nr %u is %i \n",
1382 value_leaves->ent, get_irn_node_nr(irn), val_arr[vnum].access_type);
1384 if(val_arr[vnum].calls != NULL)
1385 for(value_calls = set_first(val_arr[vnum].calls); value_calls; value_calls = set_next(val_arr[vnum].calls))
1387 ir_printf("A call with nr %i acess a element of this entity with access %u \n",
1388 get_irn_node_nr(value_calls->call), value_calls->access_type);
1393 /** Optimize the found scalar replacements.
1395 * @param set_sels A set with all entities, that
1397 * @param set_ent A set with all sels nodes,
1398 * that belong to our scalars.
1399 * @param vnum The number of scalars.
1401 static void do_data_flow_scalar_replacement(set *set_ent, set *set_sels, int vnum) {
1405 obstack_init(&env.obst);
1406 env.set_ent = set_ent;
1407 env.set_sels = set_sels;
1409 env.fix_phis = NULL;
1410 env.fix_syncs = NULL;
1411 env.gl_mem_vnum = vnum - 2;
1412 env.vnum_state = vnum - 1;
1413 /* nvals are vnum - 1, why we indicate with nvals the number
1414 * of memory edges we will produce. For vnum_state we don't
1415 * need to produce a memory edge.*/
1416 env.nvals = vnum - 1;
1419 /* first step: allocate the value arrays for every block */
1420 irg_block_walk_graph(current_ir_graph, NULL, alloc_value_arr, &env);
1422 /* second step: we analyze all calls, that have as parameter scalar(s).
1423 * We mark the calls, that save the address of a scalar and we
1424 * mark the entity owner of this scalar as not optimizeble by now.*/
1425 irg_walk_graph(current_ir_graph, NULL, analyse_calls, &env);
1427 while(env.changes) {
1432 * third step: walk over the blocks of a graph and update
1433 * the information for the access of our scalars.
1435 irg_block_walk_graph(current_ir_graph, NULL, set_block_access, &env);
1439 // if(get_firm_verbosity())
1440 /* Debug info to see if analyse_calls work properly.*/
1441 irg_block_walk_graph(current_ir_graph, NULL, print_block_state, &env);
1444 * fourth step: walk over the graph blockwise in topological order
1445 * and split the memrory edge.
1447 inc_irg_block_visited(current_ir_graph);
1448 irg_walk_blkwise_graph(current_ir_graph, NULL, split_memory_edge, &env);
1452 /* fifth step: fix all nodes, that have as predecessor Unknown.*/
1457 /* sixth step: sync memory enges for the end block.*/
1458 sync_mem_edges(&env);
1460 /*seventh step: free the allocated memory*/
1461 irg_block_walk_graph(current_ir_graph, NULL, free_call_info, &env);
1462 obstack_free(&env.obst, NULL);
1466 * Find possible scalar replacements
1468 * @param irg The current ir graph.
1470 void data_flow_scalar_replacement_opt(ir_graph *irg) {
1476 ent_leaves_t key_leaves, *value_leaves;
1479 if (! get_opt_scalar_replacement())
1482 set_sels = new_set(sels_cmp, 8);
1483 set_ent = new_set(ent_leaves_t_cmp, 8);
1485 /* Call algorithm that remove the critical edges of a ir graph. */
1486 remove_critical_cf_edges(irg);
1488 /* Call algorithm that computes the out edges.*/
1489 if (get_irg_outs_state(irg) != outs_consistent)
1490 compute_irg_outs(irg);
1492 /* Call algorithm that computes the loop information.*/
1493 compute_loop_info(irg);
1494 /* Call algorithm that computes the dominance information.*/
1497 /* Find possible scalar replacements */
1498 if (find_possible_replacements(irg)) {
1500 /* Insert in set the scalar replacements. */
1501 irg_frame = get_irg_frame(irg);
1503 for (i = 0 ; i < get_irn_n_outs(irg_frame); i++) {
1504 ir_node *succ = get_irn_out(irg_frame, i);
1506 if (get_irn_op(succ) == op_Sel) {
1507 ir_entity *ent = get_Sel_entity(succ);
1509 if (get_entity_link(ent) == NULL || get_entity_link(ent) == ADDRESS_TAKEN)
1511 /* we have found a entity, that have scalars and we insert it to our set_ent*/
1512 key_leaves.ent = ent;
1513 key_leaves.leaves = new_pset(leave_cmp, 8);
1514 value_leaves = set_insert(set_ent, &key_leaves, sizeof(key_leaves), HASH_PTR(ent));
1516 /* We allocate for every leave sel a vnum.*/
1517 vnum = allocate_value_numbers(set_sels, value_leaves->leaves, ent, vnum);
1521 if(get_firm_verbosity())
1522 printf("vnumber in data flow= %i\n", 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);