2 * Copyright (C) 1995-2008 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"
50 #define SET_VNUM(node, vnum) set_irn_link(node, INT_TO_PTR(vnum))
51 #define GET_VNUM(node) (unsigned)PTR_TO_INT(get_irn_link(node))
54 * A path element entry: it is either an entity
55 * or a tarval, because we evaluate only constant array
56 * accesses like a.b.c[8].d
64 * An access path, used to assign value numbers
65 * to variables that will be scalar replaced.
67 typedef struct _path_t {
68 unsigned vnum; /**< The value number. */
69 unsigned path_len; /**< The length of the access path. */
70 path_elem_t path[1]; /**< The path. */
73 /** The size of a path in bytes. */
74 #define PATH_SIZE(p) (sizeof(*(p)) + sizeof((p)->path[0]) * ((p)->path_len - 1))
76 typedef struct _scalars_t {
77 ir_entity *ent; /**< A entity for scalar replacement. */
78 ir_type *ent_owner; /**< The owner of this entity. */
81 DEBUG_ONLY(static firm_dbg_module_t *dbg;)
86 * @return 0 if they are identically
88 static int path_cmp(const void *elt, const void *key, size_t size) {
89 const path_t *p1 = elt;
90 const path_t *p2 = key;
93 /* we can use memcmp here, because identical tarvals should have identical addresses */
94 return memcmp(p1->path, p2->path, p1->path_len * sizeof(p1->path[0]));
98 * Compare two elements of the scalars_t set.
100 * @return 0 if they are identically
102 static int ent_cmp(const void *elt, const void *key, size_t size) {
103 const scalars_t *c1 = elt;
104 const scalars_t *c2 = key;
107 return c1->ent != c2->ent;
111 * Calculate a hash value for a path.
113 static unsigned path_hash(const path_t *path) {
117 for (i = 0; i < path->path_len; ++i)
118 hash ^= (unsigned)PTR_TO_INT(path->path[i].ent);
124 * Returns non-zero, if all indeces of a Sel node are constants.
126 * @param sel the Sel node that will be checked
128 static int is_const_sel(ir_node *sel) {
129 int i, n = get_Sel_n_indexs(sel);
131 for (i = 0; i < n; ++i) {
132 ir_node *idx = get_Sel_index(sel, i);
134 if (get_irn_op(idx) != op_Const)
141 * Check the mode of a Load/Store with the mode of the entity
143 * If the mode of the entity and the Load/Store mode do not match, we
144 * have the bad reinterpret case:
147 * char b = *(char *)&i;
149 * We do NOT count this as one value and return address_taken
151 * However, we support an often used case. If the mode is two-complement
152 * we allow casts between signed/unsigned.
154 * @param mode the mode of the Load/Store
155 * @param ent_mode the mode of the accessed entity
157 static int check_load_store_mode(ir_mode *mode, ir_mode *ent_mode) {
158 if (ent_mode != mode) {
159 if (ent_mode == NULL ||
160 get_mode_size_bits(ent_mode) != get_mode_size_bits(mode) ||
161 get_mode_sort(ent_mode) != get_mode_sort(mode) ||
162 get_mode_arithmetic(ent_mode) != irma_twos_complement ||
163 get_mode_arithmetic(mode) != irma_twos_complement)
170 * Returns non-zero, if the address of an entity
171 * represented by a Sel node (or it's successor Sels) is taken.
173 int is_address_taken(ir_node *sel)
176 ir_mode *emode, *mode;
180 if (! is_const_sel(sel))
183 for (i = get_irn_n_outs(sel) - 1; i >= 0; --i) {
184 ir_node *succ = get_irn_out(sel, i);
186 switch (get_irn_opcode(succ)) {
188 /* check if this load is not a hidden conversion */
189 mode = get_Load_mode(succ);
190 ent = get_Sel_entity(sel);
191 emode = get_type_mode(get_entity_type(ent));
192 if (! check_load_store_mode(mode, emode))
194 /* do not remove volatile variables */
195 if (get_Load_volatility(succ) == volatility_is_volatile)
200 /* check that Sel is not the Store's value */
201 value = get_Store_value(succ);
204 /* check if this Store is not a hidden conversion */
205 mode = get_irn_mode(value);
206 ent = get_Sel_entity(sel);
207 emode = get_type_mode(get_entity_type(ent));
208 if (! check_load_store_mode(mode, emode))
210 /* do not remove volatile variables */
211 if (get_Store_volatility(succ) == volatility_is_volatile)
216 /* Check the Sel successor of Sel */
217 int res = is_address_taken(succ);
225 /* The address of an entity is given as a parameter.
226 * As long as we do not have analyses that can tell what
227 * is done with parameters, think is taken.
228 * One special case: If the Call type tells that it's a
229 * value parameter, the address is NOT taken.
234 /* another op, the address is taken */
242 * Link all leave Sels with the entity.
244 * @param ent the entity that will be scalar replaced
245 * @param sel a Sel node that selects some fields of this entity
247 static void link_all_leave_sels(ir_entity *ent, ir_node *sel) {
250 n = get_irn_n_outs(sel);
251 for (i = 0; i < n; ++i) {
252 ir_node *succ = get_irn_out(sel, i);
255 link_all_leave_sels(ent, succ);
261 /* if Sel nodes with memory inputs are used, a entity can be
262 * visited more than once causing a ring here, so we use the
263 * node flag to mark linked nodes
265 if (irn_visited(sel))
268 /* we know we are at a leave, because this function is only
269 * called if the address is NOT taken, so succ must be a Load
272 set_irn_link(sel, get_entity_link(ent));
273 set_entity_link(ent, sel);
275 mark_irn_visited(sel);
279 /* we need a special address that serves as an address taken marker */
281 static void *ADDRESS_TAKEN = &_x;
284 * Find possible scalar replacements.
286 * @param irg an IR graph
288 * This function finds variables on the (members of the) frame type
289 * that can be scalar replaced, because their address is never taken.
290 * If such a variable is found, it's entity link will hold a list of all
291 * Sel nodes, that selects the atomic fields of this entity.
292 * Otherwise, the link will be ADDRESS_TAKEN or NULL.
294 * @return non-zero if at least one entity could be replaced
297 static int find_possible_replacements(ir_graph *irg) {
298 ir_node *irg_frame = get_irg_frame(irg);
302 inc_irg_visited(irg);
304 n = get_irn_n_outs(irg_frame);
307 * First, clear the link field of all interesting entities.
308 * Note that we did not rely on the fact that there is only
309 * one Sel node per entity, so we might access one entity
310 * more than once here.
311 * That's why we have need two loops.
313 for (i = 0; i < n; ++i) {
314 ir_node *succ = get_irn_out(irg_frame, i);
317 ir_entity *ent = get_Sel_entity(succ);
318 set_entity_link(ent, NULL);
323 * Check the ir_graph for Sel nodes. If the entity of Sel
324 * isn't a scalar replacement set the link of this entity
325 * equal ADDRESS_TAKEN.
327 for (i = 0; i < n; ++i) {
328 ir_node *succ = get_irn_out(irg_frame, i);
331 ir_entity *ent = get_Sel_entity(succ);
334 if (get_entity_link(ent) == ADDRESS_TAKEN)
338 * Beware: in rare cases even entities on the frame might be
339 * volatile. This might happen if the entity serves as a store
340 * to a value that must survive a exception. Do not optimize
341 * such entities away.
343 if (get_entity_volatility(ent) == volatility_is_volatile) {
344 set_entity_link(ent, ADDRESS_TAKEN);
348 ent_type = get_entity_type(ent);
350 /* we can handle arrays, structs and atomic types yet */
351 if (is_Array_type(ent_type) || is_Struct_type(ent_type) || is_atomic_type(ent_type)) {
352 if (is_address_taken(succ)) {
353 if (get_entity_link(ent)) /* killing one */
355 set_entity_link(ent, ADDRESS_TAKEN);
357 /* possible found one */
358 if (get_entity_link(ent) == NULL)
360 link_all_leave_sels(ent, succ);
370 * Return a path from the Sel node sel to it's root.
372 * @param sel the Sel node
373 * @param len the length of the path so far
375 static path_t *find_path(ir_node *sel, unsigned len) {
378 ir_node *pred = get_Sel_ptr(sel);
380 /* the current Sel node will add some path elements */
381 n = get_Sel_n_indexs(sel);
384 if (! is_Sel(pred)) {
385 /* we found the root */
387 res = xmalloc(sizeof(*res) + (len - 1) * sizeof(res->path));
390 res = find_path(pred, len);
392 pos = res->path_len - len;
394 res->path[pos++].ent = get_Sel_entity(sel);
395 for (i = 0; i < n; ++i) {
396 ir_node *index = get_Sel_index(sel, i);
398 res->path[pos++].tv = get_Const_tarval(index);
405 * Allocate value numbers for the leaves
406 * in our found entities.
408 * @param sels a set that will contain all Sels that have a value number
409 * @param ent the entity that will be scalar replaced
410 * @param vnum the first value number we can assign
411 * @param modes a flexible array, containing all the modes of
414 * @return the next free value number
416 static unsigned allocate_value_numbers(pset *sels, ir_entity *ent, unsigned vnum, ir_mode ***modes)
420 set *pathes = new_set(path_cmp, 8);
422 /* visit all Sel nodes in the chain of the entity */
423 for (sel = get_entity_link(ent); sel; sel = next) {
424 next = get_irn_link(sel);
426 /* we must mark this sel for later */
427 pset_insert_ptr(sels, sel);
429 key = find_path(sel, 0);
430 path = set_find(pathes, key, PATH_SIZE(key), path_hash(key));
433 SET_VNUM(sel, path->vnum);
437 set_insert(pathes, key, PATH_SIZE(key), path_hash(key));
439 SET_VNUM(sel, key->vnum);
440 ARR_EXTO(ir_mode *, *modes, (int)((key->vnum + 15) & ~15));
442 (*modes)[key->vnum] = get_type_mode(get_entity_type(get_Sel_entity(sel)));
444 assert((*modes)[key->vnum] && "Value is not atomic");
450 DB((dbg, SET_LEVEL_2, " %s", get_entity_name(key->path[0].ent)));
451 for (i = 1; i < key->path_len; ++i) {
452 if (is_entity(key->path[i].ent))
453 DB((dbg, SET_LEVEL_2, ".%s", get_entity_name(key->path[i].ent)));
455 DB((dbg, SET_LEVEL_2, "[%ld]", get_tarval_long(key->path[i].tv)));
457 DB((dbg, SET_LEVEL_2, " = %u (%s)\n", PTR_TO_INT(get_irn_link(sel)), get_mode_name((*modes)[key->vnum])));
459 #endif /* DEBUG_libfirm */
465 set_entity_link(ent, NULL);
470 * A list entry for the fixing lists
472 typedef struct _list_entry_t {
473 ir_node *node; /**< the node that must be fixed */
474 unsigned vnum; /**< the value number of this node */
478 * environment for memory walker
480 typedef struct _env_t {
481 struct obstack obst; /**< a obstack for the value blocks */
482 int nvals; /**< number of values */
483 ir_mode **modes; /**< the modes of the values */
484 list_entry_t *fix_phis; /**< list of all Phi nodes that must be fixed */
485 list_entry_t *fix_loads; /**< list of all Load nodes that must be fixed */
486 pset *sels; /**< A set of all Sel nodes that have a value number */
490 * topological walker.
492 static void topologic_walker(ir_node *node, void *ctx) {
494 ir_op *op = get_irn_op(node);
495 ir_node *adr, *block, *mem, *unk, **value_arr, **in, *val;
502 /* a load, check if we can resolve it */
503 adr = get_Load_ptr(node);
508 if (! pset_find_ptr(env->sels, adr))
511 /* ok, we have a Load that will be replaced */
512 vnum = GET_VNUM(adr);
514 assert(vnum < (unsigned)env->nvals);
516 block = get_nodes_block(node);
517 value_arr = get_irn_link(block);
519 /* check, if we can replace this Load */
520 if (value_arr[vnum]) {
521 mem = get_Load_mem(node);
523 /* Beware: A Load can contain a hidden conversion in Firm.
524 This happens for instance in the following code:
527 unsigned j = *(unsigned *)&i;
530 val = value_arr[vnum];
531 mode = get_Load_mode(node);
532 if (mode != get_irn_mode(val))
533 val = new_d_Conv(get_irn_dbg_info(node), val, mode);
535 turn_into_tuple(node, pn_Load_max);
536 set_Tuple_pred(node, pn_Load_M, mem);
537 set_Tuple_pred(node, pn_Load_res, val);
538 set_Tuple_pred(node, pn_Load_X_regular, new_r_Jmp(current_ir_graph, block));
539 set_Tuple_pred(node, pn_Load_X_except, new_Bad());
541 l = obstack_alloc(&env->obst, sizeof(*l));
545 set_irn_link(node, env->fix_loads);
548 } else if (op == op_Store) {
549 /* a Store always can be replaced */
550 adr = get_Store_ptr(node);
555 if (! pset_find_ptr(env->sels, adr))
558 vnum = GET_VNUM(adr);
560 assert(vnum < (unsigned)env->nvals);
562 block = get_nodes_block(node);
563 value_arr = get_irn_link(block);
565 /* Beware: A Store can contain a hidden conversion in Firm. */
566 val = get_Store_value(node);
567 if (get_irn_mode(val) != env->modes[vnum])
568 val = new_d_Conv(get_irn_dbg_info(node), val, env->modes[vnum]);
569 value_arr[vnum] = val;
571 mem = get_Store_mem(node);
572 block = get_nodes_block(node);
574 turn_into_tuple(node, pn_Store_max);
575 set_Tuple_pred(node, pn_Store_M, mem);
576 set_Tuple_pred(node, pn_Store_X_regular, new_r_Jmp(current_ir_graph, block));
577 set_Tuple_pred(node, pn_Store_X_except, new_Bad());
578 } else if (op == op_Phi && get_irn_mode(node) == mode_M) {
580 * found a memory Phi: Here, we must create new Phi nodes
582 block = get_nodes_block(node);
583 value_arr = get_irn_link(block);
585 n = get_Block_n_cfgpreds(block);
587 in = alloca(sizeof(*in) * n);
589 for (i = env->nvals - 1; i >= 0; --i) {
590 unk = new_Unknown(env->modes[i]);
591 for (j = n - 1; j >= 0; --j)
594 value_arr[i] = new_r_Phi(current_ir_graph, block, n, in, env->modes[i]);
596 l = obstack_alloc(&env->obst, sizeof(*l));
597 l->node = value_arr[i];
600 set_irn_link(value_arr[i], env->fix_phis);
607 * Walker: allocate the value array for every block.
609 static void alloc_value_arr(ir_node *block, void *ctx) {
611 ir_node **var_arr = obstack_alloc(&env->obst, sizeof(*var_arr) * env->nvals);
613 /* the value array is empty at start */
614 memset(var_arr, 0, sizeof(*var_arr) * env->nvals);
615 set_irn_link(block, var_arr);
619 * searches through blocks beginning from block for value
620 * vnum and return it.
622 static ir_node *find_vnum_value(ir_node *block, unsigned vnum) {
627 if (Block_not_block_visited(block)) {
628 mark_Block_block_visited(block);
630 value_arr = get_irn_link(block);
633 return value_arr[vnum];
635 for (i = get_Block_n_cfgpreds(block) - 1; i >= 0; --i) {
636 ir_node *pred = get_Block_cfgpred(block, i);
638 res = find_vnum_value(get_nodes_block(pred), vnum);
649 static void fix_phis(env_t *env) {
651 ir_node *phi, *block, *pred, *val;
654 for (l = env->fix_phis; l; l = get_irn_link(phi)) {
657 block = get_nodes_block(phi);
658 for (i = get_irn_arity(phi) - 1; i >= 0; --i) {
659 pred = get_Block_cfgpred(block, i);
660 pred = get_nodes_block(pred);
662 inc_irg_block_visited(current_ir_graph);
663 val = find_vnum_value(pred, l->vnum);
666 set_irn_n(phi, i, val);
674 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) {
727 obstack_init(&env.obst);
731 env.fix_loads = NULL;
734 /* first step: allocate the value arrays for every block */
735 irg_block_walk_graph(current_ir_graph, NULL, alloc_value_arr, &env);
738 * second step: walk over the graph blockwise in topological order
739 * and fill the array as much as possible.
741 irg_walk_blkwise_graph(current_ir_graph, NULL, topologic_walker, &env);
743 /* third, fix the list of Phis, then the list of Loads */
747 obstack_free(&env.obst, NULL);
751 * Find possible scalar replacements
753 * @param irg The current ir graph.
755 void scalar_replacement_opt(ir_graph *irg) {
758 scalars_t key, *value;
766 if (! get_opt_scalar_replacement())
769 rem = current_ir_graph;
771 /* Call algorithm that computes the out edges */
772 assure_irg_outs(irg);
774 /* Find possible scalar replacements */
775 if (find_possible_replacements(irg)) {
776 DB((dbg, SET_LEVEL_1, "Scalar Replacement: %s\n", get_entity_name(get_irg_entity(irg))));
778 /* Insert in set the scalar replacements. */
779 irg_frame = get_irg_frame(irg);
781 modes = NEW_ARR_F(ir_mode *, 16);
782 set_ent = new_set(ent_cmp, 8);
783 sels = pset_new_ptr(8);
785 for (i = 0 ; i < get_irn_n_outs(irg_frame); i++) {
786 ir_node *succ = get_irn_out(irg_frame, i);
789 ir_entity *ent = get_Sel_entity(succ);
791 if (get_entity_link(ent) == NULL || get_entity_link(ent) == ADDRESS_TAKEN)
794 ent_type = get_entity_type(ent);
797 key.ent_owner = get_entity_owner(ent);
798 set_insert(set_ent, &key, sizeof(key), HASH_PTR(key.ent));
801 if (is_Array_type(ent_type)) {
802 DB((dbg, SET_LEVEL_1, " found array %s\n", get_entity_name(ent)));
803 } else if (is_Struct_type(ent_type)) {
804 DB((dbg, SET_LEVEL_1, " found struct %s\n", get_entity_name(ent)));
805 } else if (is_atomic_type(ent_type))
806 DB((dbg, SET_LEVEL_1, " found atomic value %s\n", get_entity_name(ent)));
808 panic("Neither an array nor a struct or atomic value found in scalar replace");
810 #endif /* DEBUG_libfirm */
812 nvals = allocate_value_numbers(sels, ent, nvals, &modes);
816 DB((dbg, SET_LEVEL_1, " %u values will be needed\n", nvals));
818 /* If scalars were found. */
820 do_scalar_replacements(sels, nvals, modes);
822 foreach_set(set_ent, value) {
823 remove_class_member(value->ent_owner, value->ent);
827 * We changed the graph, but did NOT introduce new blocks
828 * neither changed control flow, cf-backedges should be still
831 set_irg_outs_inconsistent(irg);
832 set_irg_loopinfo_inconsistent(irg);
840 current_ir_graph = rem;
843 void firm_init_scalar_replace(void) {
844 FIRM_DBG_REGISTER(dbg, "firm.opt.scalar_replace");