2 * Copyright (C) 1995-2011 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
30 #include "iroptimize.h"
31 #include "scalar_replace.h"
49 #include "opt_manage.h"
51 static unsigned get_vnum(const ir_node *node)
53 return (unsigned)PTR_TO_INT(get_irn_link(node));
56 static void set_vnum(ir_node *node, unsigned vnum)
58 set_irn_link(node, INT_TO_PTR(vnum));
62 * A path element entry: it is either an entity
63 * or a tarval, because we evaluate only constant array
64 * accesses like a.b.c[8].d
72 * An access path, used to assign value numbers
73 * to variables that will be scalar replaced.
75 typedef struct path_t {
76 unsigned vnum; /**< The value number. */
77 size_t path_len; /**< The length of the access path. */
78 path_elem_t path[1]; /**< The path. */
81 /** The size of a path in bytes. */
82 static size_t path_size(path_t *p)
84 return sizeof(*p) + sizeof(p->path[0]) * (p->path_len-1);
87 typedef struct scalars_t {
88 ir_entity *ent; /**< A entity for scalar replacement. */
91 DEBUG_ONLY(static firm_dbg_module_t *dbg;)
96 * @return 0 if they are identically
98 static int path_cmp(const void *elt, const void *key, size_t size)
100 const path_t *p1 = (const path_t*)elt;
101 const path_t *p2 = (const path_t*)key;
104 /* we can use memcmp here, because identical tarvals should have identical addresses */
105 return memcmp(p1->path, p2->path, p1->path_len * sizeof(p1->path[0]));
109 * Compare two elements of the scalars_t set.
111 * @return 0 if they are identically
113 static int ent_cmp(const void *elt, const void *key, size_t size)
115 const scalars_t *c1 = (const scalars_t*)elt;
116 const scalars_t *c2 = (const scalars_t*)key;
119 return c1->ent != c2->ent;
123 * Calculate a hash value for a path.
125 static unsigned path_hash(const path_t *path)
130 for (i = 0; i < path->path_len; ++i)
131 hash ^= (unsigned)PTR_TO_INT(path->path[i].ent);
137 * Returns non-zero, if all indeces of a Sel node are constants.
139 * @param sel the Sel node that will be checked
141 static bool is_const_sel(ir_node *sel)
143 int i, n = get_Sel_n_indexs(sel);
145 for (i = 0; i < n; ++i) {
146 ir_node *idx = get_Sel_index(sel, i);
155 * Check the mode of a Load/Store with the mode of the entity
157 * If the mode of the entity and the Load/Store mode do not match, we
158 * have the bad reinterpret case:
161 * char b = *(char *)&i;
163 * We do NOT count this as one value and return address_taken
165 * However, we support an often used case. If the mode is two-complement
166 * we allow casts between signed/unsigned.
168 * @param mode the mode of the Load/Store
169 * @param ent_mode the mode of the accessed entity
171 static bool check_load_store_mode(ir_mode *mode, ir_mode *ent_mode)
173 if (ent_mode != mode) {
174 if (ent_mode == NULL ||
175 get_mode_size_bits(ent_mode) != get_mode_size_bits(mode) ||
176 get_mode_arithmetic(ent_mode) != irma_twos_complement ||
177 get_mode_arithmetic(mode) != irma_twos_complement)
184 * Returns non-zero, if the address of an entity
185 * represented by a Sel node (or its successor Sels) is taken.
187 bool is_address_taken(ir_node *sel)
190 ir_mode *emode, *mode;
194 if (! is_const_sel(sel))
197 for (i = get_irn_n_outs(sel) - 1; i >= 0; --i) {
198 ir_node *succ = get_irn_out(sel, i);
200 switch (get_irn_opcode(succ)) {
202 /* do not remove volatile variables */
203 if (get_Load_volatility(succ) == volatility_is_volatile)
205 /* check if this load is not a hidden conversion */
206 mode = get_Load_mode(succ);
207 ent = get_Sel_entity(sel);
208 emode = get_type_mode(get_entity_type(ent));
209 if (! check_load_store_mode(mode, emode))
214 /* check that Sel is not the Store's value */
215 value = get_Store_value(succ);
218 /* do not remove volatile variables */
219 if (get_Store_volatility(succ) == volatility_is_volatile)
221 /* check if this Store is not a hidden conversion */
222 mode = get_irn_mode(value);
223 ent = get_Sel_entity(sel);
224 emode = get_type_mode(get_entity_type(ent));
225 if (! check_load_store_mode(mode, emode))
231 ir_entity *entity = get_Sel_entity(succ);
232 /* we can't handle unions correctly yet -> address taken */
233 if (is_Union_type(get_entity_owner(entity)))
236 /* Check the Sel successor of Sel */
237 res = is_address_taken(succ);
244 /* The address of an entity is given as a parameter.
245 * As long as we do not have analyses that can tell what
246 * is done with parameters, think is taken.
247 * One special case: If the Call type tells that it's a
248 * value parameter, the address is NOT taken.
253 int res = is_address_taken(succ);
260 /* Non-optimized Tuple, happens in inlining */
261 for (input_nr = get_Tuple_n_preds(succ) - 1; input_nr >= 0; --input_nr) {
262 ir_node *pred = get_Tuple_pred(succ, input_nr);
265 /* we found one input */
266 for (k = get_irn_n_outs(succ) - 1; k >= 0; --k) {
267 ir_node *proj = get_irn_out(succ, k);
269 if (is_Proj(proj) && get_Proj_proj(proj) == input_nr) {
270 int res = is_address_taken(proj);
280 /* another op, the address is taken */
288 * Link all leave Sels with the entity.
290 * @param ent the entity that will be scalar replaced
291 * @param sel a Sel node that selects some fields of this entity
293 static bool link_all_leave_sels(ir_entity *ent, ir_node *sel)
296 bool is_leave = true;
298 for (i = get_irn_n_outs(sel) - 1; i >= 0; --i) {
299 ir_node *succ = get_irn_out(sel, i);
302 /* the current leave has further Sel's, no leave */
304 link_all_leave_sels(ent, succ);
305 } else if (is_Id(succ)) {
306 is_leave &= link_all_leave_sels(ent, succ);
314 /* we know we are at a leave, because this function is only
315 * called if the address is NOT taken, so sel's successor(s)
316 * must be Loads or Stores
318 set_irn_link(sel, get_entity_link(ent));
319 set_entity_link(ent, sel);
324 /* we need a special address that serves as an address taken marker */
326 static void *ADDRESS_TAKEN = &_x;
329 * Find possible scalar replacements.
331 * @param irg an IR graph
333 * This function finds variables on the (members of the) frame type
334 * that can be scalar replaced, because their address is never taken.
335 * If such a variable is found, its entity link will hold a list of all
336 * Sel nodes, that selects the atomic fields of this entity.
337 * Otherwise, the link will be ADDRESS_TAKEN or NULL.
339 * @return non-zero if at least one entity could be replaced
342 static int find_possible_replacements(ir_graph *irg)
348 long static_link_arg;
352 * First, clear the link field of all interesting entities.
354 frame_tp = get_irg_frame_type(irg);
355 for (mem_idx = get_class_n_members(frame_tp); mem_idx > 0;) {
356 ir_entity *ent = get_class_member(frame_tp, --mem_idx);
357 set_entity_link(ent, NULL);
360 /* check for inner functions:
361 * FIXME: need a way to get the argument position for the static link */
363 for (mem_idx = get_class_n_members(frame_tp); mem_idx > 0;) {
364 ir_entity *ent = get_class_member(frame_tp, --mem_idx);
365 if (is_method_entity(ent)) {
366 ir_graph *inner_irg = get_entity_irg(ent);
370 assure_irg_outs(inner_irg);
371 args = get_irg_args(inner_irg);
372 for (j = get_irn_n_outs(args) - 1; j >= 0; --j) {
373 ir_node *arg = get_irn_out(args, j);
375 if (get_Proj_proj(arg) == static_link_arg) {
377 for (k = get_irn_n_outs(arg) - 1; k >= 0; --k) {
378 ir_node *succ = get_irn_out(arg, k);
381 ir_entity *ent = get_Sel_entity(succ);
383 if (get_entity_owner(ent) == frame_tp) {
384 /* found an access to the outer frame */
385 set_entity_link(ent, ADDRESS_TAKEN);
395 * Check the ir_graph for Sel nodes. If the entity of Sel
396 * isn't a scalar replacement set the link of this entity
397 * equal ADDRESS_TAKEN.
399 irg_frame = get_irg_frame(irg);
400 for (i = get_irn_n_outs(irg_frame) - 1; i >= 0; --i) {
401 ir_node *succ = get_irn_out(irg_frame, i);
404 ir_entity *ent = get_Sel_entity(succ);
407 /* we are only interested in entities on the frame, NOT
409 if (get_entity_owner(ent) != frame_tp)
412 if (get_entity_link(ent) == ADDRESS_TAKEN)
415 ent_type = get_entity_type(ent);
417 /* we can handle arrays, structs and atomic types yet */
418 if (is_Array_type(ent_type) || is_Struct_type(ent_type) || is_atomic_type(ent_type)) {
419 if (is_address_taken(succ)) {
421 if (get_entity_link(ent))
423 set_entity_link(ent, ADDRESS_TAKEN);
425 /* possible found one */
426 if (get_entity_link(ent) == NULL)
428 link_all_leave_sels(ent, succ);
438 * Return a path from the Sel node "sel" to its root.
440 * @param sel the Sel node
441 * @param len the length of the path so far
443 static path_t *find_path(ir_node *sel, size_t len)
448 ir_node *pred = get_Sel_ptr(sel);
450 /* the current Sel node will add some path elements */
451 n = get_Sel_n_indexs(sel);
454 if (! is_Sel(pred)) {
455 /* we found the root */
456 res = XMALLOCF(path_t, path, len);
459 res = find_path(pred, len);
461 assert(len <= res->path_len);
462 pos = res->path_len - len;
464 res->path[pos++].ent = get_Sel_entity(sel);
465 for (i = 0; i < n; ++i) {
466 ir_node *index = get_Sel_index(sel, i);
468 res->path[pos++].tv = get_Const_tarval(index);
475 * Allocate value numbers for the leaves
476 * in our found entities.
478 * @param sels a set that will contain all Sels that have a value number
479 * @param ent the entity that will be scalar replaced
480 * @param vnum the first value number we can assign
481 * @param modes a flexible array, containing all the modes of
484 * @return the next free value number
486 static unsigned allocate_value_numbers(pset *sels, ir_entity *ent, unsigned vnum, ir_mode ***modes)
490 set *pathes = new_set(path_cmp, 8);
492 DB((dbg, SET_LEVEL_3, " Visiting Sel nodes of entity %+F\n", ent));
493 /* visit all Sel nodes in the chain of the entity */
494 for (sel = (ir_node*)get_entity_link(ent); sel != NULL; sel = next) {
495 next = (ir_node*)get_irn_link(sel);
497 /* we must mark this sel for later */
498 pset_insert_ptr(sels, sel);
500 key = find_path(sel, 0);
501 path = (path_t*)set_find(pathes, key, path_size(key), path_hash(key));
504 set_vnum(sel, path->vnum);
505 DB((dbg, SET_LEVEL_3, " %+F represents value %u\n", sel, path->vnum));
509 set_insert(pathes, key, path_size(key), path_hash(key));
511 set_vnum(sel, key->vnum);
512 DB((dbg, SET_LEVEL_3, " %+F represents value %u\n", sel, key->vnum));
514 ARR_EXTO(ir_mode *, *modes, (key->vnum + 15) & ~15);
516 (*modes)[key->vnum] = get_type_mode(get_entity_type(get_Sel_entity(sel)));
518 assert((*modes)[key->vnum] && "Value is not atomic");
524 DB((dbg, SET_LEVEL_2, " %s", get_entity_name(key->path[0].ent)));
525 for (i = 1; i < key->path_len; ++i) {
526 if (is_entity(key->path[i].ent))
527 DB((dbg, SET_LEVEL_2, ".%s", get_entity_name(key->path[i].ent)));
529 DB((dbg, SET_LEVEL_2, "[%ld]", get_tarval_long(key->path[i].tv)));
531 DB((dbg, SET_LEVEL_2, " = %u (%s)\n", PTR_TO_INT(get_irn_link(sel)), get_mode_name((*modes)[key->vnum])));
533 #endif /* DEBUG_libfirm */
539 set_entity_link(ent, NULL);
544 * environment for memory walker
546 typedef struct env_t {
547 unsigned nvals; /**< number of values */
548 ir_mode **modes; /**< the modes of the values */
549 pset *sels; /**< A set of all Sel nodes that have a value number */
553 * topological post-walker.
555 static void walker(ir_node *node, void *ctx)
557 env_t *env = (env_t*)ctx;
558 ir_graph *irg = get_irn_irg(node);
559 ir_node *addr, *block, *mem, *val;
564 /* a load, check if we can resolve it */
565 addr = get_Load_ptr(node);
567 DB((dbg, SET_LEVEL_3, " checking %+F for replacement ", node));
568 if (! is_Sel(addr)) {
569 DB((dbg, SET_LEVEL_3, "no Sel input (%+F)\n", addr));
573 if (! pset_find_ptr(env->sels, addr)) {
574 DB((dbg, SET_LEVEL_3, "Sel %+F has no VNUM\n", addr));
578 /* ok, we have a Load that will be replaced */
579 vnum = get_vnum(addr);
580 assert(vnum < env->nvals);
582 DB((dbg, SET_LEVEL_3, "replacing by value %u\n", vnum));
584 block = get_nodes_block(node);
585 set_cur_block(block);
587 /* check, if we can replace this Load */
588 val = get_value(vnum, env->modes[vnum]);
590 /* Beware: A Load can contain a hidden conversion in Firm.
591 This happens for instance in the following code:
594 unsigned j = *(unsigned *)&i;
597 mode = get_Load_mode(node);
598 if (mode != get_irn_mode(val))
599 val = new_rd_Conv(get_irn_dbg_info(node), block, val, mode);
601 mem = get_Load_mem(node);
602 turn_into_tuple(node, pn_Load_max+1);
603 set_Tuple_pred(node, pn_Load_M, mem);
604 set_Tuple_pred(node, pn_Load_res, val);
605 set_Tuple_pred(node, pn_Load_X_regular, new_r_Jmp(block));
606 set_Tuple_pred(node, pn_Load_X_except, new_r_Bad(irg, mode_X));
607 } else if (is_Store(node)) {
608 DB((dbg, SET_LEVEL_3, " checking %+F for replacement ", node));
610 /* a Store always can be replaced */
611 addr = get_Store_ptr(node);
613 if (! is_Sel(addr)) {
614 DB((dbg, SET_LEVEL_3, "no Sel input (%+F)\n", addr));
618 if (! pset_find_ptr(env->sels, addr)) {
619 DB((dbg, SET_LEVEL_3, "Sel %+F has no VNUM\n", addr));
623 vnum = get_vnum(addr);
624 assert(vnum < env->nvals);
626 DB((dbg, SET_LEVEL_3, "replacing by value %u\n", vnum));
628 block = get_nodes_block(node);
629 set_cur_block(block);
631 /* Beware: A Store can contain a hidden conversion in Firm. */
632 val = get_Store_value(node);
633 if (get_irn_mode(val) != env->modes[vnum])
634 val = new_rd_Conv(get_irn_dbg_info(node), block, val, env->modes[vnum]);
636 set_value(vnum, val);
638 mem = get_Store_mem(node);
639 turn_into_tuple(node, pn_Store_max+1);
640 set_Tuple_pred(node, pn_Store_M, mem);
641 set_Tuple_pred(node, pn_Store_X_regular, new_r_Jmp(block));
642 set_Tuple_pred(node, pn_Store_X_except, new_r_Bad(irg, mode_X));
647 * Make scalar replacement.
649 * @param sels A set containing all Sel nodes that have a value number
650 * @param nvals The number of scalars.
651 * @param modes A flexible array, containing all the modes of
654 static void do_scalar_replacements(ir_graph *irg, pset *sels, unsigned nvals,
659 ssa_cons_start(irg, (int)nvals);
666 * second step: walk over the graph blockwise in topological order
667 * and fill the array as much as possible.
669 DB((dbg, SET_LEVEL_3, "Substituting Loads and Stores in %+F\n", irg));
670 irg_walk_blkwise_graph(irg, NULL, walker, &env);
672 ssa_cons_finish(irg);
676 * Find possible scalar replacements
678 * @param irg The current ir graph.
680 static ir_graph_state_t do_scalar_replacement(ir_graph *irg)
684 scalars_t key, *value;
689 ir_type *ent_type, *frame_tp;
691 /* we use the link field to store the VNUM */
692 ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK);
693 irp_reserve_resources(irp, IRP_RESOURCE_ENTITY_LINK);
695 /* Find possible scalar replacements */
696 if (find_possible_replacements(irg)) {
697 DB((dbg, SET_LEVEL_1, "Scalar Replacement: %+F\n", 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);
705 frame_tp = get_irg_frame_type(irg);
707 for (i = get_irn_n_outs(irg_frame) - 1; i >= 0; --i) {
708 ir_node *succ = get_irn_out(irg_frame, i);
711 ir_entity *ent = get_Sel_entity(succ);
713 /* we are only interested in entities on the frame, NOT
715 if (get_entity_owner(ent) != frame_tp
716 || is_parameter_entity(ent))
719 if (get_entity_link(ent) == NULL || get_entity_link(ent) == ADDRESS_TAKEN)
722 ent_type = get_entity_type(ent);
725 set_insert(set_ent, &key, sizeof(key), hash_ptr(key.ent));
728 if (is_Array_type(ent_type)) {
729 DB((dbg, SET_LEVEL_1, " found array %s\n", get_entity_name(ent)));
730 } else if (is_Struct_type(ent_type)) {
731 DB((dbg, SET_LEVEL_1, " found struct %s\n", get_entity_name(ent)));
732 } else if (is_atomic_type(ent_type))
733 DB((dbg, SET_LEVEL_1, " found atomic value %s\n", get_entity_name(ent)));
735 panic("Neither an array nor a struct or atomic value found in scalar replace");
737 #endif /* DEBUG_libfirm */
739 nvals = allocate_value_numbers(sels, ent, nvals, &modes);
743 DB((dbg, SET_LEVEL_1, " %u values will be needed\n", nvals));
745 /* If scalars were found. */
747 do_scalar_replacements(irg, sels, nvals, modes);
749 foreach_set(set_ent, scalars_t*, value) {
750 free_entity(value->ent);
754 * We changed the graph, but did NOT introduce new blocks
755 * neither changed control flow, cf-backedges should be still
764 ir_free_resources(irg, IR_RESOURCE_IRN_LINK);
765 irp_free_resources(irp, IRP_RESOURCE_ENTITY_LINK);
770 static optdesc_t opt_scalar_rep = {
772 IR_GRAPH_STATE_NO_UNREACHABLE_CODE | IR_GRAPH_STATE_CONSISTENT_OUTS,
773 do_scalar_replacement,
776 int scalar_replacement_opt(ir_graph *irg)
778 perform_irg_optimization(irg, &opt_scalar_rep);
782 ir_graph_pass_t *scalar_replacement_opt_pass(const char *name)
784 return def_graph_pass_ret(name ? name : "scalar_rep",
785 scalar_replacement_opt);
788 void firm_init_scalar_replace(void)
790 FIRM_DBG_REGISTER(dbg, "firm.opt.scalar_replace");