3 * File name: ir/opt/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.
28 #include "scalar_replace.h"
42 #define SET_VNUM(node, vnum) set_irn_link(node, INT_TO_PTR(vnum))
43 #define GET_VNUM(node) (unsigned)PTR_TO_INT(get_irn_link(node))
46 * A path element entry: it is either an entity
47 * or a tarval, because we evaluate only constant array
48 * accesses like a.b.c[8].d
56 * An access path, used to assign value numbers
57 * to variables that will be scalar replaced
59 typedef struct _path_t {
60 unsigned vnum; /**< the value number */
61 unsigned path_len; /**< the length of the access path */
62 path_elem_t path[1]; /**< the path */
65 typedef struct _scalars_t {
66 entity *ent; /**< A entity for scalar replacement. */
67 type *ent_owner; /**< The owner of this entity. */
74 * @return 0 if they are identically
76 static int path_cmp(const void *elt, const void *key, size_t size)
78 const path_t *p1 = elt;
79 const path_t *p2 = key;
81 /* we can use memcmp here, because identical tarvals should have identical addresses */
82 return memcmp(p1->path, p2->path, p1->path_len * sizeof(p1->path[0]));
86 * Compare two elements of the scalars_t set.
88 * @return 0 if they are identically
90 static int ent_cmp(const void *elt, const void *key, size_t size)
92 const scalars_t *c1 = elt;
93 const scalars_t *c2 = key;
95 return c1->ent != c2->ent;
99 * Calculate a hash value for a path.
101 static unsigned path_hash(const path_t *path)
106 for (i = 0; i < path->path_len; ++i)
107 hash ^= (unsigned)PTR_TO_INT(path->path[i].ent);
113 * Returns non-zero, if all indeces of a Sel node are constants.
115 * @param sel the Sel node that will be checked
117 static int is_const_sel(ir_node *sel) {
118 int i, n = get_Sel_n_indexs(sel);
120 for (i = 0; i < n; ++i) {
121 ir_node *idx = get_Sel_index(sel, i);
123 if (get_irn_op(idx) != op_Const)
130 * Returns non-zero, if the address of an entity
131 * represented by a Sel node (or it's successor Sels) is taken.
133 * @param sel the Sel node
135 static int is_address_taken(ir_node *sel)
139 if (! is_const_sel(sel))
142 n = get_irn_n_outs(sel);
143 for (i = 0; i < n; ++i) {
144 ir_node *succ = get_irn_out(sel, i);
146 switch (get_irn_opcode(succ)) {
148 /* ok, we just load from that entity */
152 /* check that Sel is not the Store's value */
153 if (get_Store_value(succ) == sel)
158 /* Check the Sel successor of Sel */
159 int res = is_address_taken(succ);
167 /* The address of an entity is given as a parameter.
168 * As long as we do not have analyses that can tell what
169 * is done with parameters, think is taken.
174 /* another op, the address is taken */
182 * Link all leave Sels with the entity.
184 * @param ent the entity that will be scalar replaced
185 * @param sel a Sel node that selects some fields of this entity
187 static void link_all_leave_sels(entity *ent, ir_node *sel)
191 n = get_irn_n_outs(sel);
192 for (i = 0; i < n; ++i) {
193 ir_node *succ = get_irn_out(sel, i);
195 if (get_irn_op(succ) == op_Sel) {
196 link_all_leave_sels(ent, succ);
202 /* if Sel nodes with memory inputs are used, a entity can be
203 * visited more than once causing a ring here, so we use the
204 * node flag to mark linked nodes
206 if (irn_visited(sel))
209 /* we know we are at a leave, because this function is only
210 * called if the address is NOT taken, so succ must be a Load
213 set_irn_link(sel, get_entity_link(ent));
214 set_entity_link(ent, sel);
216 mark_irn_visited(sel);
220 /* we need a special address that serves as an address taken marker */
222 static void *ADDRESS_TAKEN = &_x;
225 * Find possible scalar replacements.
227 * @param irg an IR graph
229 * This function finds variables on the (members of the) frame type
230 * that can be scalar replaced, because their address is never taken.
231 * If such a variable is found, it's entity link will hold a list of all
232 * Sel nodes, that selects the atomic fields of this entity.
233 * Otherwise, the link will be ADDRESS_TAKEN or NULL.
235 * @return non-zero if at least one entity could be replaced
238 static int find_possible_replacements(ir_graph *irg)
240 ir_node *irg_frame = get_irg_frame(irg);
244 inc_irg_visited(irg);
246 n = get_irn_n_outs(irg_frame);
249 * First, clear the link field of all interesting entities.
250 * Note that we did not rely on the fact that there is only
251 * one Sel node per entity, so we might access one entity
252 * more than once here.
253 * That's why we have need two loops.
255 for (i = 0; i < n; ++i) {
256 ir_node *succ = get_irn_out(irg_frame, i);
258 if (get_irn_op(succ) == op_Sel) {
259 entity *ent = get_Sel_entity(succ);
260 set_entity_link(ent, NULL);
265 * Check the ir_graph for Sel nodes. If the entity of Sel
266 * isn't a scalar replacement set the link of this entity
267 * equal ADDRESS_TAKEN.
269 for (i = 0; i < n; ++i) {
270 ir_node *succ = get_irn_out(irg_frame, i);
272 if (get_irn_op(succ) == op_Sel) {
273 entity *ent = get_Sel_entity(succ);
276 if (get_entity_link(ent) == ADDRESS_TAKEN)
279 ent_type = get_entity_type(ent);
281 /* we can handle arrays, structs and atomic types yet */
282 if (is_Array_type(ent_type) || is_Struct_type(ent_type) || is_atomic_type(ent_type)) {
283 if (is_address_taken(succ)) {
284 if (get_entity_link(ent)) /* killing one */
286 set_entity_link(ent, ADDRESS_TAKEN);
289 /* possible found one */
290 if (get_entity_link(ent) == NULL)
292 link_all_leave_sels(ent, succ);
302 * Return a path from the Sel node sel to it's root.
304 * @param sel the Sel node
305 * @param len the length of the path so far
307 static path_t *find_path(ir_node *sel, unsigned len)
311 ir_node *pred = get_Sel_ptr(sel);
313 /* the current Sel node will add some path elements */
314 n = get_Sel_n_indexs(sel);
317 if (get_irn_op(pred) != op_Sel) {
318 /* we found the root */
320 res = xmalloc(sizeof(*res) + (len - 1) * sizeof(res->path));
324 res = find_path(pred, len);
326 pos = res->path_len - len;
328 res->path[pos++].ent = get_Sel_entity(sel);
329 for (i = 0; i < n; ++i) {
330 ir_node *index = get_Sel_index(sel, i);
332 res->path[pos++].tv = get_Const_tarval(index);
339 * Allocate value numbers for the leaves
340 * in our found entities.
342 * @param sels a set that will contain all Sels that have a value number
343 * @param ent the entity that will be scalar replaced
344 * @param vnum the first value number we can assign
345 * @param modes a flexible array, containing all the modes of
348 * @return the next free value number
350 static unsigned allocate_value_numbers(pset *sels, entity *ent, unsigned vnum, ir_mode ***modes)
354 set *pathes = new_set(path_cmp, 8);
356 /* visit all Sel nodes in the chain of the entity */
357 for (sel = get_entity_link(ent); sel; sel = next) {
358 next = get_irn_link(sel);
360 /* we must mark this sel for later */
361 pset_insert_ptr(sels, sel);
363 key = find_path(sel, 0);
364 path = set_find(pathes, key, sizeof(*key) + sizeof(key->path[0]) * key->path_len, path_hash(key));
367 SET_VNUM(sel, path->vnum);
373 set_insert(pathes, key, sizeof(*key) + sizeof(key->path[0]) * key->path_len, path_hash(key));
375 SET_VNUM(sel, key->vnum);
376 ARR_EXTO(ir_mode *, *modes, (key->vnum + 15) & ~15);
378 (*modes)[key->vnum] = get_type_mode(get_entity_type(get_Sel_entity(sel)));
380 assert((*modes)[key->vnum] && "Value is not atomic");
384 if (get_opt_scalar_replacement_verbose() && get_firm_verbosity() > 1) {
385 printf(" %s", get_entity_name(ent));
386 for (i = 1; i < key->path_len; ++i) {
387 if (is_entity(key->path[i].ent))
388 printf(".%s", get_entity_name(key->path[i].ent));
390 printf("[%ld]", get_tarval_long(key->path[i].tv));
392 printf(" = %u (%s)\n", PTR_TO_INT(get_irn_link(sel)), get_mode_name((*modes)[key->vnum]));
394 #endif /* DEBUG_libfirm */
400 set_entity_link(ent, NULL);
405 * A list entry for the fixing lists
407 typedef struct _list_entry_t {
408 ir_node *node; /**< the node that must be fixed */
409 unsigned vnum; /**< the value number of this node */
413 * environment for memory walker
415 typedef struct _env_t {
416 struct obstack obst; /**< a obstack for the value blocks */
417 int nvals; /**< number of values */
418 ir_mode **modes; /**< the modes of the values */
419 list_entry_t *fix_phis; /**< list of all Phi nodes that must be fixed */
420 list_entry_t *fix_loads; /**< list of all Load nodes that must be fixed */
421 pset *sels; /**< A set of all Sel nodes that have a value number */
427 static void handle_first(ir_node *node, void *ctx)
430 ir_op *op = get_irn_op(node);
431 ir_node *adr, *block, *mem, *unk, **value_arr, **in;
437 /* a load, check if we can resolve it */
438 adr = get_Load_ptr(node);
440 if (get_irn_op(adr) != op_Sel)
443 if (! pset_find_ptr(env->sels, adr))
446 /* ok, we have a Load that will be replaced */
447 vnum = GET_VNUM(adr);
449 assert(vnum < (unsigned)env->nvals);
451 block = get_nodes_block(node);
452 value_arr = get_irn_link(block);
454 /* check, if we can replace this Load */
455 if (value_arr[vnum]) {
456 mem = get_Load_mem(node);
458 turn_into_tuple(node, pn_Load_max);
459 set_Tuple_pred(node, pn_Load_M, mem);
460 set_Tuple_pred(node, pn_Load_res, value_arr[vnum]);
461 set_Tuple_pred(node, pn_Load_X_except, new_Bad());
464 l = obstack_alloc(&env->obst, sizeof(*l));
468 set_irn_link(node, env->fix_loads);
472 else if (op == op_Store) {
473 /* a Store always can be replaced */
474 adr = get_Store_ptr(node);
476 if (get_irn_op(adr) != op_Sel)
479 if (! pset_find_ptr(env->sels, adr))
482 vnum = GET_VNUM(adr);
484 assert(vnum < (unsigned)env->nvals);
486 block = get_nodes_block(node);
487 value_arr = get_irn_link(block);
489 value_arr[vnum] = get_Store_value(node);
491 mem = get_Store_mem(node);
493 turn_into_tuple(node, pn_Store_max);
494 set_Tuple_pred(node, pn_Store_M, mem);
495 set_Tuple_pred(node, pn_Store_X_except, new_Bad());
497 else if (op == op_Phi && get_irn_mode(node) == mode_M) {
499 * found a memory Phi: Here, we must create new Phi nodes
501 block = get_nodes_block(node);
502 value_arr = get_irn_link(block);
504 n = get_Block_n_cfgpreds(block);
506 in = alloca(sizeof(*in) * n);
508 for (i = env->nvals - 1; i >= 0; --i) {
509 unk = new_Unknown(env->modes[i]);
510 for (j = n - 1; j >= 0; --j)
513 value_arr[i] = new_r_Phi(current_ir_graph, block, n, in, env->modes[i]);
515 l = obstack_alloc(&env->obst, sizeof(*l));
516 l->node = value_arr[i];
519 set_irn_link(value_arr[i], env->fix_phis);
526 * Walker: allocate the value array for every block.
528 static void alloc_value_arr(ir_node *block, void *ctx)
531 ir_node **var_arr = obstack_alloc(&env->obst, sizeof(*var_arr) * env->nvals);
533 /* the value array is empty at start */
534 memset(var_arr, 0, sizeof(*var_arr) * env->nvals);
535 set_irn_link(block, var_arr);
539 * searches through blocks beginning from block for value
540 * vnum and return it.
542 static ir_node *find_value(ir_node *block, unsigned vnum)
548 if (Block_not_block_visited(block)) {
549 mark_Block_block_visited(block);
551 value_arr = get_irn_link(block);
554 return value_arr[vnum];
556 for (i = get_Block_n_cfgpreds(block) - 1; i >= 0; --i) {
557 ir_node *pred = get_Block_cfgpred(block, i);
559 res = find_value(get_nodes_block(pred), vnum);
570 static void fix_phis(env_t *env)
573 ir_node *phi, *block, *pred, *val;
576 for (l = env->fix_phis; l; l = get_irn_link(phi)) {
579 block = get_nodes_block(phi);
580 for (i = get_irn_arity(phi) - 1; i >= 0; --i) {
581 pred = get_Block_cfgpred(block, i);
582 pred = get_nodes_block(pred);
584 inc_irg_block_visited(current_ir_graph);
585 val = find_value(pred, l->vnum);
588 set_irn_n(phi, i, val);
596 static void fix_loads(env_t *env)
599 ir_node *load, *block, *pred, *val, *mem;
602 for (l = env->fix_loads; l; l = get_irn_link(load)) {
605 block = get_nodes_block(load);
606 for (i = get_Block_n_cfgpreds(block) - 1; i >= 0; --i) {
607 pred = get_Block_cfgpred(block, i);
608 pred = get_nodes_block(pred);
610 inc_irg_block_visited(current_ir_graph);
611 val = find_value(pred, l->vnum);
618 /* access of an uninitialized value */
619 val = new_Unknown(env->modes[l->vnum]);
622 mem = get_Load_mem(load);
624 turn_into_tuple(load, pn_Load_max);
625 set_Tuple_pred(load, pn_Load_M, mem);
626 set_Tuple_pred(load, pn_Load_res, val);
627 set_Tuple_pred(load, pn_Load_X_except, new_Bad());
632 * Make scalar replacement.
634 * @param sels A set containing all Sel nodes that have a value number
635 * @param nvals The number of scalars.
636 * @param modes A flexible array, containing all the modes of
639 static void do_scalar_replacements(pset *sels, int nvals, ir_mode **modes)
643 obstack_init(&env.obst);
647 env.fix_loads = NULL;
650 /* first step: allocate the value arrays for every block */
651 irg_block_walk_graph(current_ir_graph, NULL, alloc_value_arr, &env);
654 * second step: walk over the graph blockwise in topological order
655 * and fill the array as much as possible.
657 irg_walk_blkwise_graph(current_ir_graph, NULL, handle_first, &env);
659 /* third, fix the list of Phis, then the list of Loads */
663 obstack_free(&env.obst, NULL);
667 * Find possible scalar replacements
669 * @param irg The current ir graph.
671 void scalar_replacement_opt(ir_graph *irg)
675 scalars_t key, *value;
683 if (! get_opt_scalar_replacement())
686 rem = current_ir_graph;
688 /* Call algorithm that computes the out edges */
689 if (get_irg_outs_state(irg) != outs_consistent)
690 compute_irg_outs(irg);
692 /* Find possible scalar replacements */
693 if (find_possible_replacements(irg)) {
695 if (get_opt_scalar_replacement_verbose()) {
696 printf("Scalar Replacement: %s\n", get_entity_name(get_irg_entity(irg)));
699 /* Insert in set the scalar replacements. */
700 irg_frame = get_irg_frame(irg);
702 modes = NEW_ARR_F(ir_mode *, 16);
703 set_ent = new_set(ent_cmp, 8);
704 sels = pset_new_ptr(8);
706 for (i = 0 ; i < get_irn_n_outs(irg_frame); i++) {
707 ir_node *succ = get_irn_out(irg_frame, i);
709 if (get_irn_op(succ) == op_Sel) {
710 entity *ent = get_Sel_entity(succ);
712 if (get_entity_link(ent) == NULL || get_entity_link(ent) == ADDRESS_TAKEN)
715 ent_type = get_entity_type(ent);
718 key.ent_owner = get_entity_owner(ent);
719 set_insert(set_ent, &key, sizeof(key), HASH_PTR(key.ent));
721 if (get_opt_scalar_replacement_verbose()) {
722 if (is_Array_type(ent_type)) {
723 printf(" found array %s\n", get_entity_name(ent));
725 else if (is_Struct_type(ent_type)) {
726 printf(" found struct %s\n", get_entity_name(ent));
728 else if (is_atomic_type(ent_type))
729 printf(" found atomic value %s\n", get_entity_name(ent));
731 assert(0 && "Neither an array nor a struct or atomic value");
735 nvals = allocate_value_numbers(sels, ent, nvals, &modes);
739 if (get_opt_scalar_replacement_verbose()) {
740 printf(" %u values will be needed\n", nvals);
743 /* If scalars were found. */
745 do_scalar_replacements(sels, nvals, modes);
747 for (value = set_first(set_ent); value; value = set_next(set_ent)) {
748 remove_class_member(value->ent_owner, value->ent);
758 * We changed the graph, but did NOT introduce new blocks
759 * neither changed control flow, cf-backedges should be still
762 set_irg_outs_inconsistent(irg);
763 set_irg_loopinfo_inconsistent(irg);
767 current_ir_graph = rem;