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"
41 #define SET_VNUM(node, vnum) set_irn_link(node, (void *)vnum)
42 #define GET_VNUM(node) (unsigned)get_irn_link(node)
45 * A path element entry: it is either an entity
46 * or a tarval, because we evaluate only constant array
47 * accesses like a.b.c[8].d
55 * An access path, used to assign value numbers
56 * to variables that will be scalar replaced
58 typedef struct _path_t {
59 unsigned vnum; /**< the value number */
60 unsigned path_len; /**< the length of the access path */
61 path_elem_t path[1]; /**< the path */
64 typedef struct _scalars_t {
65 entity *ent; /**< A entity for scalar replacement. */
66 type *ent_owner; /**< The owner of this entity. */
73 * @return 0 if they are identically
75 static int path_cmp(const void *elt, const void *key, size_t size)
77 const path_t *p1 = elt;
78 const path_t *p2 = key;
80 /* we can use memcmp here, because identical tarvals should have identical addresses */
81 return memcmp(p1->path, p2->path, p1->path_len * sizeof(p1->path[0]));
85 * Compare two elements of the scalars_t set.
87 * @return 0 if they are identically
89 static int ent_cmp(const void *elt, const void *key, size_t size)
91 const scalars_t *c1 = elt;
92 const scalars_t *c2 = key;
94 return c1->ent != c2->ent;
98 * Calculate a hash value for a path.
100 static unsigned path_hash(const path_t *path)
105 for (i = 0; i < path->path_len; ++i)
106 hash ^= (unsigned)path->path[i].ent;
112 * Returns non-zero, if all indeces of a Sel node are constants.
114 * @param sel the Sel node that will be checked
116 static int is_const_sel(ir_node *sel) {
117 int i, n = get_Sel_n_indexs(sel);
119 for (i = 0; i < n; ++i) {
120 ir_node *idx = get_Sel_index(sel, i);
122 if (get_irn_op(idx) != op_Const)
129 * Returns non-zero, if the address of an entity
130 * represented by a Sel node (or it's successor Sels) is taken.
132 * @param sel the Sel node
134 static int is_address_taken(ir_node *sel)
138 if (! is_const_sel(sel))
141 n = get_irn_n_outs(sel);
142 for (i = 0; i < n; ++i) {
143 ir_node *succ = get_irn_out(sel, i);
145 switch (get_irn_opcode(succ)) {
147 /* ok, we just load from that entity */
151 /* check that Sel is not the Store's value */
152 if (get_Store_value(succ) == sel)
157 /* Check the Sel successor of Sel */
158 int res = is_address_taken(succ);
166 /* The address of an entity is given as a parameter.
167 * As long as we do not have analyses that can tell what
168 * is done with parameters, think is taken.
173 /* another op, the address is taken */
181 * Link all leave Sels with the entity.
183 * @param ent the entity that will be scalar replaced
184 * @param sel a Sel node that selects some fields of this entity
186 static void link_all_leave_sels(entity *ent, ir_node *sel)
190 n = get_irn_n_outs(sel);
191 for (i = 0; i < n; ++i) {
192 ir_node *succ = get_irn_out(sel, i);
194 if (get_irn_op(succ) == op_Sel) {
195 link_all_leave_sels(ent, succ);
201 /* if Sel nodes with memory inputs are used, a entity can be
202 * visited more than once causing a ring here, so we use the
203 * node flag to mark linked nodes
205 if (irn_visited(sel))
208 /* we know we are at a leave, because this function is only
209 * called if the address is NOT taken, so succ must be a Load
212 set_irn_link(sel, get_entity_link(ent));
213 set_entity_link(ent, sel);
215 mark_irn_visited(sel);
219 /* we need a special address that serves as an address taken marker */
221 static void *ADDRESS_TAKEN = &_x;
224 * Find possible scalar replacements.
226 * @param irg an IR graph
228 * This function finds variables on the (members of the) frame type
229 * that can be scalar replaced, because their address is never taken.
230 * If such a variable is found, it's entity link will hold a list of all
231 * Sel nodes, that selects the atomic fields of this entity.
232 * Otherwise, the link will be ADDRESS_TAKEN or NULL.
234 * @return non-zero if at least one entity could be replaced
237 static int find_possible_replacements(ir_graph *irg)
239 ir_node *irg_frame = get_irg_frame(irg);
243 inc_irg_visited(irg);
245 n = get_irn_n_outs(irg_frame);
248 * First, clear the link field of all interesting entities.
249 * Note that we did not rely on the fact that there is only
250 * one Sel node per entity, so we might access one entity
251 * more than once here.
252 * That's why we have need two loops.
254 for (i = 0; i < n; ++i) {
255 ir_node *succ = get_irn_out(irg_frame, i);
257 if (get_irn_op(succ) == op_Sel) {
258 entity *ent = get_Sel_entity(succ);
259 set_entity_link(ent, NULL);
264 * Check the ir_graph for Sel nodes. If the entity of Sel
265 * isn't a scalar replacement set the link of this entity
266 * equal ADDRESS_TAKEN.
268 for (i = 0; i < n; ++i) {
269 ir_node *succ = get_irn_out(irg_frame, i);
271 if (get_irn_op(succ) == op_Sel) {
272 entity *ent = get_Sel_entity(succ);
275 if (get_entity_link(ent) == ADDRESS_TAKEN)
278 ent_type = get_entity_type(ent);
280 /* we can handle arrays, structs and atomic types yet */
281 if (is_Array_type(ent_type) || is_Struct_type(ent_type) || is_atomic_type(ent_type)) {
282 if (is_address_taken(succ)) {
283 if (get_entity_link(ent)) /* killing one */
285 set_entity_link(ent, ADDRESS_TAKEN);
288 /* possible found one */
289 if (get_entity_link(ent) == NULL)
291 link_all_leave_sels(ent, succ);
301 * Return a path from the Sel node sel to it's root.
303 * @param sel the Sel node
304 * @param len the length of the path so far
306 static path_t *find_path(ir_node *sel, unsigned len)
310 ir_node *pred = get_Sel_ptr(sel);
312 /* the current Sel node will add some path elements */
313 n = get_Sel_n_indexs(sel);
316 if (get_irn_op(pred) != op_Sel) {
317 /* we found the root */
319 res = xmalloc(sizeof(*res) + (len - 1) * sizeof(res->path));
323 res = find_path(pred, len);
325 pos = res->path_len - len;
327 res->path[pos++].ent = get_Sel_entity(sel);
328 for (i = 0; i < n; ++i) {
329 ir_node *index = get_Sel_index(sel, i);
331 res->path[pos++].tv = get_Const_tarval(index);
338 * Allocate value numbers for the leaves
339 * in our found entities.
341 * @param sels a set that will contain all Sels that have a value number
342 * @param ent the entity that will be scalar replaced
343 * @param vnum the first value number we can assign
344 * @param modes a flexible array, containing all the modes of
347 * @return the next free value number
349 static unsigned allocate_value_numbers(pset *sels, entity *ent, unsigned vnum, ir_mode ***modes)
353 set *pathes = new_set(path_cmp, 8);
355 /* visit all Sel nodes in the chain of the entity */
356 for (sel = get_entity_link(ent); sel; sel = next) {
357 next = get_irn_link(sel);
359 /* we must mark this sel for later */
360 pset_insert_ptr(sels, sel);
362 key = find_path(sel, 0);
363 path = set_find(pathes, key, sizeof(*key) + sizeof(key->path[0]) * key->path_len, path_hash(key));
366 SET_VNUM(sel, path->vnum);
372 set_insert(pathes, key, sizeof(*key) + sizeof(key->path[0]) * key->path_len, path_hash(key));
374 SET_VNUM(sel, key->vnum);
375 ARR_EXTO(ir_mode *, *modes, (key->vnum + 15) & ~15);
377 (*modes)[key->vnum] = get_type_mode(get_entity_type(get_Sel_entity(sel)));
379 assert((*modes)[key->vnum] && "Value is not atomic");
383 if (get_opt_scalar_replacement_verbose() && get_firm_verbosity() > 1) {
384 printf(" %s", get_entity_name(ent));
385 for (i = 1; i < key->path_len; ++i) {
386 if (is_entity(key->path[i].ent))
387 printf(".%s", get_entity_name(key->path[i].ent));
389 printf("[%ld]", get_tarval_long(key->path[i].tv));
391 printf(" = %u (%s)\n", (int)get_irn_link(sel), get_mode_name((*modes)[key->vnum]));
393 #endif /* DEBUG_libfirm */
399 set_entity_link(ent, NULL);
404 * A list entry for the fixing lists
406 typedef struct _list_entry_t {
407 ir_node *node; /**< the node that must be fixed */
408 unsigned vnum; /**< the value number of this node */
412 * environment for memory walker
414 typedef struct _env_t {
415 struct obstack obst; /**< a obstack for the value blocks */
416 int nvals; /**< number of values */
417 ir_mode **modes; /**< the modes of the values */
418 list_entry_t *fix_phis; /**< list of all Phi nodes that must be fixed */
419 list_entry_t *fix_loads; /**< list of all Load nodes that must be fixed */
420 pset *sels; /**< A set of all Sel nodes that have a value number */
426 static void handle_first(ir_node *node, void *ctx)
429 ir_op *op = get_irn_op(node);
430 ir_node *adr, *block, *mem, *unk, **value_arr, **in;
436 /* a load, check if we can resolve it */
437 adr = get_Load_ptr(node);
439 if (get_irn_op(adr) != op_Sel)
442 if (! pset_find_ptr(env->sels, adr))
445 /* ok, we have a Load that will be replaced */
446 vnum = GET_VNUM(adr);
448 assert(vnum < (unsigned)env->nvals);
450 block = get_nodes_block(node);
451 value_arr = get_irn_link(block);
453 /* check, if we can replace this Load */
454 if (value_arr[vnum]) {
455 mem = get_Load_mem(node);
457 turn_into_tuple(node, pn_Load_max);
458 set_Tuple_pred(node, pn_Load_M, mem);
459 set_Tuple_pred(node, pn_Load_res, value_arr[vnum]);
460 set_Tuple_pred(node, pn_Load_X_except, new_Bad());
463 l = obstack_alloc(&env->obst, sizeof(*l));
467 set_irn_link(node, env->fix_loads);
471 else if (op == op_Store) {
472 /* a Store always can be replaced */
473 adr = get_Store_ptr(node);
475 if (get_irn_op(adr) != op_Sel)
478 if (! pset_find_ptr(env->sels, adr))
481 vnum = GET_VNUM(adr);
483 assert(vnum < (unsigned)env->nvals);
485 block = get_nodes_block(node);
486 value_arr = get_irn_link(block);
488 value_arr[vnum] = get_Store_value(node);
490 mem = get_Store_mem(node);
492 turn_into_tuple(node, pn_Store_max);
493 set_Tuple_pred(node, pn_Store_M, mem);
494 set_Tuple_pred(node, pn_Store_X_except, new_Bad());
496 else if (op == op_Phi && get_irn_mode(node) == mode_M) {
498 * found a memory Phi: Here, we must create new Phi nodes
500 block = get_nodes_block(node);
501 value_arr = get_irn_link(block);
503 n = get_Block_n_cfgpreds(block);
505 in = alloca(sizeof(*in) * n);
507 for (i = env->nvals - 1; i >= 0; --i) {
508 unk = new_Unknown(env->modes[i]);
509 for (j = n - 1; j >= 0; --j)
512 value_arr[i] = new_r_Phi(current_ir_graph, block, n, in, env->modes[i]);
514 l = obstack_alloc(&env->obst, sizeof(*l));
515 l->node = value_arr[i];
518 set_irn_link(value_arr[i], env->fix_phis);
525 * Walker: allocate the value array for every block.
527 static void alloc_value_arr(ir_node *block, void *ctx)
530 ir_node **var_arr = obstack_alloc(&env->obst, sizeof(*var_arr) * env->nvals);
532 /* the value array is empty at start */
533 memset(var_arr, 0, sizeof(*var_arr) * env->nvals);
534 set_irn_link(block, var_arr);
538 * searches through blocks beginning from block for value
539 * vnum and return it.
541 static ir_node *find_value(ir_node *block, unsigned vnum)
547 if (Block_not_block_visited(block)) {
548 mark_Block_block_visited(block);
550 value_arr = get_irn_link(block);
553 return value_arr[vnum];
555 for (i = get_Block_n_cfgpreds(block) - 1; i >= 0; --i) {
556 ir_node *pred = get_Block_cfgpred(block, i);
558 res = find_value(get_nodes_block(pred), vnum);
569 static void fix_phis(env_t *env)
572 ir_node *phi, *block, *pred, *val;
575 for (l = env->fix_phis; l; l = get_irn_link(phi)) {
578 block = get_nodes_block(phi);
579 for (i = get_irn_arity(phi) - 1; i >= 0; --i) {
580 pred = get_Block_cfgpred(block, i);
581 pred = get_nodes_block(pred);
583 inc_irg_block_visited(current_ir_graph);
584 val = find_value(pred, l->vnum);
587 set_irn_n(phi, i, val);
595 static void fix_loads(env_t *env)
598 ir_node *load, *block, *pred, *val, *mem;
601 for (l = env->fix_loads; l; l = get_irn_link(load)) {
604 block = get_nodes_block(load);
605 for (i = get_Block_n_cfgpreds(block) - 1; i >= 0; --i) {
606 pred = get_Block_cfgpred(block, i);
607 pred = get_nodes_block(pred);
609 inc_irg_block_visited(current_ir_graph);
610 val = find_value(pred, l->vnum);
617 /* access of an uninitialized value */
618 val = new_Unknown(env->modes[l->vnum]);
621 mem = get_Load_mem(load);
623 turn_into_tuple(load, pn_Load_max);
624 set_Tuple_pred(load, pn_Load_M, mem);
625 set_Tuple_pred(load, pn_Load_res, val);
626 set_Tuple_pred(load, pn_Load_X_except, new_Bad());
631 * Make scalar replacement.
633 * @param sels A set containing all Sel nodes that have a value number
634 * @param nvals The number of scalars.
635 * @param modes A flexible array, containing all the modes of
638 static void do_scalar_replacements(pset *sels, int nvals, ir_mode **modes)
642 obstack_init(&env.obst);
646 env.fix_loads = NULL;
649 /* first step: allocate the value arrays for every block */
650 irg_block_walk_graph(current_ir_graph, NULL, alloc_value_arr, &env);
653 * second step: walk over the graph blockwise in topological order
654 * and fill the array as much as possible.
656 irg_walk_blkwise_graph(current_ir_graph, NULL, handle_first, &env);
658 /* third, fix the list of Phis, then the list of Loads */
662 obstack_free(&env.obst, NULL);
666 * Find possible scalar replacements
668 * @param irg The current ir graph.
670 void scalar_replacement_opt(ir_graph *irg)
674 scalars_t key, *value;
682 if (! get_opt_scalar_replacement())
685 rem = current_ir_graph;
687 /* Call algorithm that computes the out edges */
688 if (get_irg_outs_state(irg) != outs_consistent)
689 compute_irg_outs(irg);
691 /* Find possible scalar replacements */
692 if (find_possible_replacements(irg)) {
694 if (get_opt_scalar_replacement_verbose()) {
695 printf("Scalar Replacement: %s\n", get_entity_name(get_irg_entity(irg)));
698 /* Insert in set the scalar replacements. */
699 irg_frame = get_irg_frame(irg);
701 modes = NEW_ARR_F(ir_mode *, 16);
702 set_ent = new_set(ent_cmp, 8);
703 sels = pset_new_ptr(8);
705 for (i = 0 ; i < get_irn_n_outs(irg_frame); i++) {
706 ir_node *succ = get_irn_out(irg_frame, i);
708 if (get_irn_op(succ) == op_Sel) {
709 entity *ent = get_Sel_entity(succ);
711 if (get_entity_link(ent) == NULL || get_entity_link(ent) == ADDRESS_TAKEN)
714 ent_type = get_entity_type(ent);
717 key.ent_owner = get_entity_owner(ent);
718 set_insert(set_ent, &key, sizeof(key), HASH_PTR(key.ent));
720 if (get_opt_scalar_replacement_verbose()) {
721 if (is_Array_type(ent_type)) {
722 printf(" found array %s\n", get_entity_name(ent));
724 else if (is_Struct_type(ent_type)) {
725 printf(" found struct %s\n", get_entity_name(ent));
727 else if (is_atomic_type(ent_type))
728 printf(" found atomic value %s\n", get_entity_name(ent));
730 assert(0 && "Neither an array nor a struct or atomic value");
734 nvals = allocate_value_numbers(sels, ent, nvals, &modes);
738 if (get_opt_scalar_replacement_verbose()) {
739 printf(" %u values will be needed\n", nvals);
742 /* If scalars were found. */
744 do_scalar_replacements(sels, nvals, modes);
746 for (value = set_first(set_ent); value; value = set_next(set_ent)) {
747 remove_class_member(value->ent_owner, value->ent);
757 * We changed the graph, but did NOT introduce new blocks
758 * neither changed control flow, cf-backedges should be still
761 set_irg_outs_inconsistent(irg);
762 set_irg_loopinfo_inconsistent(irg);
766 current_ir_graph = rem;