2 * Copyright (C) 1995-2007 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 compounds.
23 * @author Beyhan Veliev, Michael Beck
32 #include "iroptimize.h"
33 #include "scalar_replace.h"
48 #define SET_VNUM(node, vnum) set_irn_link(node, INT_TO_PTR(vnum))
49 #define GET_VNUM(node) (unsigned)PTR_TO_INT(get_irn_link(node))
52 * A path element entry: it is either an entity
53 * or a tarval, because we evaluate only constant array
54 * accesses like a.b.c[8].d
62 * An access path, used to assign value numbers
63 * to variables that will be scalar replaced.
65 typedef struct _path_t {
66 unsigned vnum; /**< The value number. */
67 unsigned path_len; /**< The length of the access path. */
68 path_elem_t path[1]; /**< The path. */
71 /** The size of a path in bytes. */
72 #define PATH_SIZE(p) (sizeof(*(p)) + sizeof((p)->path[0]) * ((p)->path_len - 1))
74 typedef struct _scalars_t {
75 ir_entity *ent; /**< A entity for scalar replacement. */
76 ir_type *ent_owner; /**< The owner of this entity. */
83 * @return 0 if they are identically
85 static int path_cmp(const void *elt, const void *key, size_t size)
87 const path_t *p1 = elt;
88 const path_t *p2 = key;
90 /* we can use memcmp here, because identical tarvals should have identical addresses */
91 return memcmp(p1->path, p2->path, p1->path_len * sizeof(p1->path[0]));
95 * Compare two elements of the scalars_t set.
97 * @return 0 if they are identically
99 static int ent_cmp(const void *elt, const void *key, size_t size)
101 const scalars_t *c1 = elt;
102 const scalars_t *c2 = key;
104 return c1->ent != c2->ent;
108 * Calculate a hash value for a path.
110 static unsigned path_hash(const path_t *path)
115 for (i = 0; i < path->path_len; ++i)
116 hash ^= (unsigned)PTR_TO_INT(path->path[i].ent);
122 * Returns non-zero, if all indeces of a Sel node are constants.
124 * @param sel the Sel node that will be checked
126 static int is_const_sel(ir_node *sel) {
127 int i, n = get_Sel_n_indexs(sel);
129 for (i = 0; i < n; ++i) {
130 ir_node *idx = get_Sel_index(sel, i);
132 if (get_irn_op(idx) != op_Const)
139 * Check the mode of a Load/Store with the mode of the entity
141 * If the mode of the entity and the Load/Store mode do not match, we
142 * have the bad reinterpret case:
145 * char b = *(char *)&i;
147 * We do NOT count this as one value and return address_taken
149 * However, we support an often used case. If the mode is two-complement
150 * we allow casts between signed/unsigned.
152 * @param mode the mode of the Load/Store
153 * @param ent_mode the mode of the accessed entity
155 static int check_load_store_mode(ir_mode *mode, ir_mode *ent_mode) {
156 if (ent_mode != mode) {
157 if (ent_mode == NULL ||
158 get_mode_size_bits(ent_mode) != get_mode_size_bits(mode) ||
159 get_mode_sort(ent_mode) != get_mode_sort(mode) ||
160 get_mode_arithmetic(ent_mode) != irma_twos_complement ||
161 get_mode_arithmetic(mode) != irma_twos_complement)
168 * Returns non-zero, if the address of an entity
169 * represented by a Sel node (or it's successor Sels) is taken.
171 int is_address_taken(ir_node *sel)
174 ir_mode *emode, *mode;
178 if (! is_const_sel(sel))
181 for (i = get_irn_n_outs(sel) - 1; i >= 0; --i) {
182 ir_node *succ = get_irn_out(sel, i);
184 switch (get_irn_opcode(succ)) {
186 /* check if this load is not a hidden conversion */
187 mode = get_Load_mode(succ);
188 ent = get_Sel_entity(sel);
189 emode = get_type_mode(get_entity_type(ent));
190 if (! check_load_store_mode(mode, emode))
195 /* check that Sel is not the Store's value */
196 value = get_Store_value(succ);
199 /* check if this Store is not a hidden conversion */
200 mode = get_irn_mode(value);
201 ent = get_Sel_entity(sel);
202 emode = get_type_mode(get_entity_type(ent));
203 if (! check_load_store_mode(mode, emode))
208 /* Check the Sel successor of Sel */
209 int res = is_address_taken(succ);
217 /* The address of an entity is given as a parameter.
218 * As long as we do not have analyses that can tell what
219 * is done with parameters, think is taken.
224 /* another op, the address is taken */
232 * Link all leave Sels with the entity.
234 * @param ent the entity that will be scalar replaced
235 * @param sel a Sel node that selects some fields of this entity
237 static void link_all_leave_sels(ir_entity *ent, ir_node *sel)
241 n = get_irn_n_outs(sel);
242 for (i = 0; i < n; ++i) {
243 ir_node *succ = get_irn_out(sel, i);
246 link_all_leave_sels(ent, succ);
252 /* if Sel nodes with memory inputs are used, a entity can be
253 * visited more than once causing a ring here, so we use the
254 * node flag to mark linked nodes
256 if (irn_visited(sel))
259 /* we know we are at a leave, because this function is only
260 * called if the address is NOT taken, so succ must be a Load
263 set_irn_link(sel, get_entity_link(ent));
264 set_entity_link(ent, sel);
266 mark_irn_visited(sel);
270 /* we need a special address that serves as an address taken marker */
272 static void *ADDRESS_TAKEN = &_x;
275 * Find possible scalar replacements.
277 * @param irg an IR graph
279 * This function finds variables on the (members of the) frame type
280 * that can be scalar replaced, because their address is never taken.
281 * If such a variable is found, it's entity link will hold a list of all
282 * Sel nodes, that selects the atomic fields of this entity.
283 * Otherwise, the link will be ADDRESS_TAKEN or NULL.
285 * @return non-zero if at least one entity could be replaced
288 static int find_possible_replacements(ir_graph *irg)
290 ir_node *irg_frame = get_irg_frame(irg);
294 inc_irg_visited(irg);
296 n = get_irn_n_outs(irg_frame);
299 * First, clear the link field of all interesting entities.
300 * Note that we did not rely on the fact that there is only
301 * one Sel node per entity, so we might access one entity
302 * more than once here.
303 * That's why we have need two loops.
305 for (i = 0; i < n; ++i) {
306 ir_node *succ = get_irn_out(irg_frame, i);
309 ir_entity *ent = get_Sel_entity(succ);
310 set_entity_link(ent, NULL);
315 * Check the ir_graph for Sel nodes. If the entity of Sel
316 * isn't a scalar replacement set the link of this entity
317 * equal ADDRESS_TAKEN.
319 for (i = 0; i < n; ++i) {
320 ir_node *succ = get_irn_out(irg_frame, i);
323 ir_entity *ent = get_Sel_entity(succ);
326 if (get_entity_link(ent) == ADDRESS_TAKEN)
330 * Beware: in rare cases even entities on the frame might be
331 * volatile. This might happen if the entity serves as a store
332 * to a value that must survive a exception. Do not optimize
333 * such entities away.
335 if (get_entity_volatility(ent) == volatility_is_volatile) {
336 set_entity_link(ent, ADDRESS_TAKEN);
340 ent_type = get_entity_type(ent);
342 /* we can handle arrays, structs and atomic types yet */
343 if (is_Array_type(ent_type) || is_Struct_type(ent_type) || is_atomic_type(ent_type)) {
344 if (is_address_taken(succ)) {
345 if (get_entity_link(ent)) /* killing one */
347 set_entity_link(ent, ADDRESS_TAKEN);
350 /* possible found one */
351 if (get_entity_link(ent) == NULL)
353 link_all_leave_sels(ent, succ);
363 * Return a path from the Sel node sel to it's root.
365 * @param sel the Sel node
366 * @param len the length of the path so far
368 static path_t *find_path(ir_node *sel, unsigned len)
372 ir_node *pred = get_Sel_ptr(sel);
374 /* the current Sel node will add some path elements */
375 n = get_Sel_n_indexs(sel);
378 if (! is_Sel(pred)) {
379 /* we found the root */
381 res = xmalloc(sizeof(*res) + (len - 1) * sizeof(res->path));
385 res = find_path(pred, len);
387 pos = res->path_len - len;
389 res->path[pos++].ent = get_Sel_entity(sel);
390 for (i = 0; i < n; ++i) {
391 ir_node *index = get_Sel_index(sel, i);
393 res->path[pos++].tv = get_Const_tarval(index);
400 * Allocate value numbers for the leaves
401 * in our found entities.
403 * @param sels a set that will contain all Sels that have a value number
404 * @param ent the entity that will be scalar replaced
405 * @param vnum the first value number we can assign
406 * @param modes a flexible array, containing all the modes of
409 * @return the next free value number
411 static unsigned allocate_value_numbers(pset *sels, ir_entity *ent, unsigned vnum, ir_mode ***modes)
415 set *pathes = new_set(path_cmp, 8);
417 /* visit all Sel nodes in the chain of the entity */
418 for (sel = get_entity_link(ent); sel; sel = next) {
419 next = get_irn_link(sel);
421 /* we must mark this sel for later */
422 pset_insert_ptr(sels, sel);
424 key = find_path(sel, 0);
425 path = set_find(pathes, key, PATH_SIZE(key), path_hash(key));
428 SET_VNUM(sel, path->vnum);
432 set_insert(pathes, key, PATH_SIZE(key), path_hash(key));
434 SET_VNUM(sel, key->vnum);
435 ARR_EXTO(ir_mode *, *modes, (key->vnum + 15) & ~15);
437 (*modes)[key->vnum] = get_type_mode(get_entity_type(get_Sel_entity(sel)));
439 assert((*modes)[key->vnum] && "Value is not atomic");
443 if (get_opt_scalar_replacement_verbose() && get_firm_verbosity() > 1) {
445 printf(" %s", get_entity_name(key->path[0].ent));
446 for (i = 1; i < key->path_len; ++i) {
447 if (is_entity(key->path[i].ent))
448 printf(".%s", get_entity_name(key->path[i].ent));
450 printf("[%ld]", get_tarval_long(key->path[i].tv));
452 printf(" = %u (%s)\n", PTR_TO_INT(get_irn_link(sel)), get_mode_name((*modes)[key->vnum]));
454 #endif /* DEBUG_libfirm */
460 set_entity_link(ent, NULL);
465 * A list entry for the fixing lists
467 typedef struct _list_entry_t {
468 ir_node *node; /**< the node that must be fixed */
469 unsigned vnum; /**< the value number of this node */
473 * environment for memory walker
475 typedef struct _env_t {
476 struct obstack obst; /**< a obstack for the value blocks */
477 int nvals; /**< number of values */
478 ir_mode **modes; /**< the modes of the values */
479 list_entry_t *fix_phis; /**< list of all Phi nodes that must be fixed */
480 list_entry_t *fix_loads; /**< list of all Load nodes that must be fixed */
481 pset *sels; /**< A set of all Sel nodes that have a value number */
485 * topological walker.
487 static void topologic_walker(ir_node *node, void *ctx)
490 ir_op *op = get_irn_op(node);
491 ir_node *adr, *block, *mem, *unk, **value_arr, **in, *val;
498 /* a load, check if we can resolve it */
499 adr = get_Load_ptr(node);
504 if (! pset_find_ptr(env->sels, adr))
507 /* ok, we have a Load that will be replaced */
508 vnum = GET_VNUM(adr);
510 assert(vnum < (unsigned)env->nvals);
512 block = get_nodes_block(node);
513 value_arr = get_irn_link(block);
515 /* check, if we can replace this Load */
516 if (value_arr[vnum]) {
517 mem = get_Load_mem(node);
519 /* Beware: A Load can contain a hidden conversion in Firm.
520 This happens for instance in the following code:
523 unsigned j = *(unsigned *)&i;
526 val = value_arr[vnum];
527 mode = get_Load_mode(node);
528 if (mode != get_irn_mode(val))
529 val = new_d_Conv(get_irn_dbg_info(node), val, mode);
531 turn_into_tuple(node, pn_Load_max);
532 set_Tuple_pred(node, pn_Load_M, mem);
533 set_Tuple_pred(node, pn_Load_res, val);
534 set_Tuple_pred(node, pn_Load_X_regular, new_r_Jmp(current_ir_graph, block));
535 set_Tuple_pred(node, pn_Load_X_except, new_Bad());
537 l = obstack_alloc(&env->obst, sizeof(*l));
541 set_irn_link(node, env->fix_loads);
544 } else if (op == op_Store) {
545 /* a Store always can be replaced */
546 adr = get_Store_ptr(node);
551 if (! pset_find_ptr(env->sels, adr))
554 vnum = GET_VNUM(adr);
556 assert(vnum < (unsigned)env->nvals);
558 block = get_nodes_block(node);
559 value_arr = get_irn_link(block);
561 /* Beware: A Store can contain a hidden conversion in Firm. */
562 val = get_Store_value(node);
563 if (get_irn_mode(val) != env->modes[vnum])
564 val = new_d_Conv(get_irn_dbg_info(node), val, env->modes[vnum]);
565 value_arr[vnum] = val;
567 mem = get_Store_mem(node);
568 block = get_nodes_block(node);
570 turn_into_tuple(node, pn_Store_max);
571 set_Tuple_pred(node, pn_Store_M, mem);
572 set_Tuple_pred(node, pn_Store_X_regular, new_r_Jmp(current_ir_graph, block));
573 set_Tuple_pred(node, pn_Store_X_except, new_Bad());
574 } else if (op == op_Phi && get_irn_mode(node) == mode_M) {
576 * found a memory Phi: Here, we must create new Phi nodes
578 block = get_nodes_block(node);
579 value_arr = get_irn_link(block);
581 n = get_Block_n_cfgpreds(block);
583 in = alloca(sizeof(*in) * n);
585 for (i = env->nvals - 1; i >= 0; --i) {
586 unk = new_Unknown(env->modes[i]);
587 for (j = n - 1; j >= 0; --j)
590 value_arr[i] = new_r_Phi(current_ir_graph, block, n, in, env->modes[i]);
592 l = obstack_alloc(&env->obst, sizeof(*l));
593 l->node = value_arr[i];
596 set_irn_link(value_arr[i], env->fix_phis);
603 * Walker: allocate the value array for every block.
605 static void alloc_value_arr(ir_node *block, void *ctx)
608 ir_node **var_arr = obstack_alloc(&env->obst, sizeof(*var_arr) * env->nvals);
610 /* the value array is empty at start */
611 memset(var_arr, 0, sizeof(*var_arr) * env->nvals);
612 set_irn_link(block, var_arr);
616 * searches through blocks beginning from block for value
617 * vnum and return it.
619 static ir_node *find_vnum_value(ir_node *block, unsigned vnum)
625 if (Block_not_block_visited(block)) {
626 mark_Block_block_visited(block);
628 value_arr = get_irn_link(block);
631 return value_arr[vnum];
633 for (i = get_Block_n_cfgpreds(block) - 1; i >= 0; --i) {
634 ir_node *pred = get_Block_cfgpred(block, i);
636 res = find_vnum_value(get_nodes_block(pred), vnum);
647 static void fix_phis(env_t *env)
650 ir_node *phi, *block, *pred, *val;
653 for (l = env->fix_phis; l; l = get_irn_link(phi)) {
656 block = get_nodes_block(phi);
657 for (i = get_irn_arity(phi) - 1; i >= 0; --i) {
658 pred = get_Block_cfgpred(block, i);
659 pred = get_nodes_block(pred);
661 inc_irg_block_visited(current_ir_graph);
662 val = find_vnum_value(pred, l->vnum);
665 set_irn_n(phi, i, val);
673 static void fix_loads(env_t *env)
676 ir_node *load, *block, *pred, *val = NULL, *mem;
680 for (l = env->fix_loads; l; l = get_irn_link(load)) {
683 block = get_nodes_block(load);
684 for (i = get_Block_n_cfgpreds(block) - 1; i >= 0; --i) {
685 pred = get_Block_cfgpred(block, i);
686 pred = get_nodes_block(pred);
688 inc_irg_block_visited(current_ir_graph);
689 val = find_vnum_value(pred, l->vnum);
696 /* access of an uninitialized value */
697 val = new_Unknown(env->modes[l->vnum]);
700 /* Beware: A Load can contain a hidden conversion in Firm.
702 mode = get_Load_mode(load);
703 if (mode != get_irn_mode(val))
704 val = new_d_Conv(get_irn_dbg_info(load), val, mode);
706 mem = get_Load_mem(load);
708 turn_into_tuple(load, pn_Load_max);
709 set_Tuple_pred(load, pn_Load_M, mem);
710 set_Tuple_pred(load, pn_Load_res, val);
711 set_Tuple_pred(load, pn_Load_X_regular, new_r_Jmp(current_ir_graph, block));
712 set_Tuple_pred(load, pn_Load_X_except, new_Bad());
717 * Make scalar replacement.
719 * @param sels A set containing all Sel nodes that have a value number
720 * @param nvals The number of scalars.
721 * @param modes A flexible array, containing all the modes of
724 static void do_scalar_replacements(pset *sels, int nvals, ir_mode **modes)
728 obstack_init(&env.obst);
732 env.fix_loads = NULL;
735 /* first step: allocate the value arrays for every block */
736 irg_block_walk_graph(current_ir_graph, NULL, alloc_value_arr, &env);
739 * second step: walk over the graph blockwise in topological order
740 * and fill the array as much as possible.
742 irg_walk_blkwise_graph(current_ir_graph, NULL, topologic_walker, &env);
744 /* third, fix the list of Phis, then the list of Loads */
748 obstack_free(&env.obst, NULL);
752 * Find possible scalar replacements
754 * @param irg The current ir graph.
756 void scalar_replacement_opt(ir_graph *irg)
760 scalars_t key, *value;
768 if (! get_opt_scalar_replacement())
771 rem = current_ir_graph;
773 /* Call algorithm that computes the out edges */
774 assure_irg_outs(irg);
776 /* Find possible scalar replacements */
777 if (find_possible_replacements(irg)) {
779 if (get_opt_scalar_replacement_verbose()) {
780 printf("Scalar Replacement: %s\n", get_entity_name(get_irg_entity(irg)));
783 /* Insert in set the scalar replacements. */
784 irg_frame = get_irg_frame(irg);
786 modes = NEW_ARR_F(ir_mode *, 16);
787 set_ent = new_set(ent_cmp, 8);
788 sels = pset_new_ptr(8);
790 for (i = 0 ; i < get_irn_n_outs(irg_frame); i++) {
791 ir_node *succ = get_irn_out(irg_frame, i);
794 ir_entity *ent = get_Sel_entity(succ);
796 if (get_entity_link(ent) == NULL || get_entity_link(ent) == ADDRESS_TAKEN)
799 ent_type = get_entity_type(ent);
802 key.ent_owner = get_entity_owner(ent);
803 set_insert(set_ent, &key, sizeof(key), HASH_PTR(key.ent));
805 if (get_opt_scalar_replacement_verbose()) {
806 if (is_Array_type(ent_type)) {
807 printf(" found array %s\n", get_entity_name(ent));
809 else if (is_Struct_type(ent_type)) {
810 printf(" found struct %s\n", get_entity_name(ent));
812 else if (is_atomic_type(ent_type))
813 printf(" found atomic value %s\n", get_entity_name(ent));
815 assert(0 && "Neither an array nor a struct or atomic value");
819 nvals = allocate_value_numbers(sels, ent, nvals, &modes);
823 if (get_opt_scalar_replacement_verbose()) {
824 printf(" %u values will be needed\n", nvals);
827 /* If scalars were found. */
829 do_scalar_replacements(sels, nvals, modes);
831 for (value = set_first(set_ent); value; value = set_next(set_ent)) {
832 remove_class_member(value->ent_owner, value->ent);
842 * We changed the graph, but did NOT introduce new blocks
843 * neither changed control flow, cf-backedges should be still
846 set_irg_outs_inconsistent(irg);
847 set_irg_loopinfo_inconsistent(irg);
851 current_ir_graph = rem;