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"
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) {
86 const path_t *p1 = elt;
87 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) {
100 const scalars_t *c1 = elt;
101 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) {
114 for (i = 0; i < path->path_len; ++i)
115 hash ^= (unsigned)PTR_TO_INT(path->path[i].ent);
121 * Returns non-zero, if all indeces of a Sel node are constants.
123 * @param sel the Sel node that will be checked
125 static int is_const_sel(ir_node *sel) {
126 int i, n = get_Sel_n_indexs(sel);
128 for (i = 0; i < n; ++i) {
129 ir_node *idx = get_Sel_index(sel, i);
131 if (get_irn_op(idx) != op_Const)
138 * Check the mode of a Load/Store with the mode of the entity
140 * If the mode of the entity and the Load/Store mode do not match, we
141 * have the bad reinterpret case:
144 * char b = *(char *)&i;
146 * We do NOT count this as one value and return address_taken
148 * However, we support an often used case. If the mode is two-complement
149 * we allow casts between signed/unsigned.
151 * @param mode the mode of the Load/Store
152 * @param ent_mode the mode of the accessed entity
154 static int check_load_store_mode(ir_mode *mode, ir_mode *ent_mode) {
155 if (ent_mode != mode) {
156 if (ent_mode == NULL ||
157 get_mode_size_bits(ent_mode) != get_mode_size_bits(mode) ||
158 get_mode_sort(ent_mode) != get_mode_sort(mode) ||
159 get_mode_arithmetic(ent_mode) != irma_twos_complement ||
160 get_mode_arithmetic(mode) != irma_twos_complement)
167 * Returns non-zero, if the address of an entity
168 * represented by a Sel node (or it's successor Sels) is taken.
170 int is_address_taken(ir_node *sel)
173 ir_mode *emode, *mode;
177 if (! is_const_sel(sel))
180 for (i = get_irn_n_outs(sel) - 1; i >= 0; --i) {
181 ir_node *succ = get_irn_out(sel, i);
183 switch (get_irn_opcode(succ)) {
185 /* check if this load is not a hidden conversion */
186 mode = get_Load_mode(succ);
187 ent = get_Sel_entity(sel);
188 emode = get_type_mode(get_entity_type(ent));
189 if (! check_load_store_mode(mode, emode))
191 /* do not remove volatile variables */
192 if (get_Load_volatility(succ) == volatility_is_volatile)
197 /* check that Sel is not the Store's value */
198 value = get_Store_value(succ);
201 /* check if this Store is not a hidden conversion */
202 mode = get_irn_mode(value);
203 ent = get_Sel_entity(sel);
204 emode = get_type_mode(get_entity_type(ent));
205 if (! check_load_store_mode(mode, emode))
207 /* do not remove volatile variables */
208 if (get_Store_volatility(succ) == volatility_is_volatile)
213 /* Check the Sel successor of Sel */
214 int res = is_address_taken(succ);
222 /* The address of an entity is given as a parameter.
223 * As long as we do not have analyses that can tell what
224 * is done with parameters, think is taken.
225 * One special case: If the Call type tells that it's a
226 * value parameter, the address is NOT taken.
231 /* another op, the address is taken */
239 * Link all leave Sels with the entity.
241 * @param ent the entity that will be scalar replaced
242 * @param sel a Sel node that selects some fields of this entity
244 static void link_all_leave_sels(ir_entity *ent, ir_node *sel) {
247 n = get_irn_n_outs(sel);
248 for (i = 0; i < n; ++i) {
249 ir_node *succ = get_irn_out(sel, i);
252 link_all_leave_sels(ent, succ);
258 /* if Sel nodes with memory inputs are used, a entity can be
259 * visited more than once causing a ring here, so we use the
260 * node flag to mark linked nodes
262 if (irn_visited(sel))
265 /* we know we are at a leave, because this function is only
266 * called if the address is NOT taken, so succ must be a Load
269 set_irn_link(sel, get_entity_link(ent));
270 set_entity_link(ent, sel);
272 mark_irn_visited(sel);
276 /* we need a special address that serves as an address taken marker */
278 static void *ADDRESS_TAKEN = &_x;
281 * Find possible scalar replacements.
283 * @param irg an IR graph
285 * This function finds variables on the (members of the) frame type
286 * that can be scalar replaced, because their address is never taken.
287 * If such a variable is found, it's entity link will hold a list of all
288 * Sel nodes, that selects the atomic fields of this entity.
289 * Otherwise, the link will be ADDRESS_TAKEN or NULL.
291 * @return non-zero if at least one entity could be replaced
294 static int find_possible_replacements(ir_graph *irg) {
295 ir_node *irg_frame = get_irg_frame(irg);
299 inc_irg_visited(irg);
301 n = get_irn_n_outs(irg_frame);
304 * First, clear the link field of all interesting entities.
305 * Note that we did not rely on the fact that there is only
306 * one Sel node per entity, so we might access one entity
307 * more than once here.
308 * That's why we have need two loops.
310 for (i = 0; i < n; ++i) {
311 ir_node *succ = get_irn_out(irg_frame, i);
314 ir_entity *ent = get_Sel_entity(succ);
315 set_entity_link(ent, NULL);
320 * Check the ir_graph for Sel nodes. If the entity of Sel
321 * isn't a scalar replacement set the link of this entity
322 * equal ADDRESS_TAKEN.
324 for (i = 0; i < n; ++i) {
325 ir_node *succ = get_irn_out(irg_frame, i);
328 ir_entity *ent = get_Sel_entity(succ);
331 if (get_entity_link(ent) == ADDRESS_TAKEN)
335 * Beware: in rare cases even entities on the frame might be
336 * volatile. This might happen if the entity serves as a store
337 * to a value that must survive a exception. Do not optimize
338 * such entities away.
340 if (get_entity_volatility(ent) == volatility_is_volatile) {
341 set_entity_link(ent, ADDRESS_TAKEN);
345 ent_type = get_entity_type(ent);
347 /* we can handle arrays, structs and atomic types yet */
348 if (is_Array_type(ent_type) || is_Struct_type(ent_type) || is_atomic_type(ent_type)) {
349 if (is_address_taken(succ)) {
350 if (get_entity_link(ent)) /* killing one */
352 set_entity_link(ent, ADDRESS_TAKEN);
354 /* possible found one */
355 if (get_entity_link(ent) == NULL)
357 link_all_leave_sels(ent, succ);
367 * Return a path from the Sel node sel to it's root.
369 * @param sel the Sel node
370 * @param len the length of the path so far
372 static path_t *find_path(ir_node *sel, unsigned len) {
375 ir_node *pred = get_Sel_ptr(sel);
377 /* the current Sel node will add some path elements */
378 n = get_Sel_n_indexs(sel);
381 if (! is_Sel(pred)) {
382 /* we found the root */
384 res = xmalloc(sizeof(*res) + (len - 1) * sizeof(res->path));
387 res = find_path(pred, len);
389 pos = res->path_len - len;
391 res->path[pos++].ent = get_Sel_entity(sel);
392 for (i = 0; i < n; ++i) {
393 ir_node *index = get_Sel_index(sel, i);
395 res->path[pos++].tv = get_Const_tarval(index);
402 * Allocate value numbers for the leaves
403 * in our found entities.
405 * @param sels a set that will contain all Sels that have a value number
406 * @param ent the entity that will be scalar replaced
407 * @param vnum the first value number we can assign
408 * @param modes a flexible array, containing all the modes of
411 * @return the next free value number
413 static unsigned allocate_value_numbers(pset *sels, ir_entity *ent, unsigned vnum, ir_mode ***modes)
417 set *pathes = new_set(path_cmp, 8);
419 /* visit all Sel nodes in the chain of the entity */
420 for (sel = get_entity_link(ent); sel; sel = next) {
421 next = get_irn_link(sel);
423 /* we must mark this sel for later */
424 pset_insert_ptr(sels, sel);
426 key = find_path(sel, 0);
427 path = set_find(pathes, key, PATH_SIZE(key), path_hash(key));
430 SET_VNUM(sel, path->vnum);
434 set_insert(pathes, key, PATH_SIZE(key), path_hash(key));
436 SET_VNUM(sel, key->vnum);
437 ARR_EXTO(ir_mode *, *modes, (int)((key->vnum + 15) & ~15));
439 (*modes)[key->vnum] = get_type_mode(get_entity_type(get_Sel_entity(sel)));
441 assert((*modes)[key->vnum] && "Value is not atomic");
445 if (get_opt_scalar_replacement_verbose() && get_firm_verbosity() > 1) {
447 printf(" %s", get_entity_name(key->path[0].ent));
448 for (i = 1; i < key->path_len; ++i) {
449 if (is_entity(key->path[i].ent))
450 printf(".%s", get_entity_name(key->path[i].ent));
452 printf("[%ld]", get_tarval_long(key->path[i].tv));
454 printf(" = %u (%s)\n", PTR_TO_INT(get_irn_link(sel)), get_mode_name((*modes)[key->vnum]));
456 #endif /* DEBUG_libfirm */
462 set_entity_link(ent, NULL);
467 * A list entry for the fixing lists
469 typedef struct _list_entry_t {
470 ir_node *node; /**< the node that must be fixed */
471 unsigned vnum; /**< the value number of this node */
475 * environment for memory walker
477 typedef struct _env_t {
478 struct obstack obst; /**< a obstack for the value blocks */
479 int nvals; /**< number of values */
480 ir_mode **modes; /**< the modes of the values */
481 list_entry_t *fix_phis; /**< list of all Phi nodes that must be fixed */
482 list_entry_t *fix_loads; /**< list of all Load nodes that must be fixed */
483 pset *sels; /**< A set of all Sel nodes that have a value number */
487 * topological walker.
489 static void topologic_walker(ir_node *node, void *ctx) {
491 ir_op *op = get_irn_op(node);
492 ir_node *adr, *block, *mem, *unk, **value_arr, **in, *val;
499 /* a load, check if we can resolve it */
500 adr = get_Load_ptr(node);
505 if (! pset_find_ptr(env->sels, adr))
508 /* ok, we have a Load that will be replaced */
509 vnum = GET_VNUM(adr);
511 assert(vnum < (unsigned)env->nvals);
513 block = get_nodes_block(node);
514 value_arr = get_irn_link(block);
516 /* check, if we can replace this Load */
517 if (value_arr[vnum]) {
518 mem = get_Load_mem(node);
520 /* Beware: A Load can contain a hidden conversion in Firm.
521 This happens for instance in the following code:
524 unsigned j = *(unsigned *)&i;
527 val = value_arr[vnum];
528 mode = get_Load_mode(node);
529 if (mode != get_irn_mode(val))
530 val = new_d_Conv(get_irn_dbg_info(node), val, mode);
532 turn_into_tuple(node, pn_Load_max);
533 set_Tuple_pred(node, pn_Load_M, mem);
534 set_Tuple_pred(node, pn_Load_res, val);
535 set_Tuple_pred(node, pn_Load_X_regular, new_r_Jmp(current_ir_graph, block));
536 set_Tuple_pred(node, pn_Load_X_except, new_Bad());
538 l = obstack_alloc(&env->obst, sizeof(*l));
542 set_irn_link(node, env->fix_loads);
545 } else if (op == op_Store) {
546 /* a Store always can be replaced */
547 adr = get_Store_ptr(node);
552 if (! pset_find_ptr(env->sels, adr))
555 vnum = GET_VNUM(adr);
557 assert(vnum < (unsigned)env->nvals);
559 block = get_nodes_block(node);
560 value_arr = get_irn_link(block);
562 /* Beware: A Store can contain a hidden conversion in Firm. */
563 val = get_Store_value(node);
564 if (get_irn_mode(val) != env->modes[vnum])
565 val = new_d_Conv(get_irn_dbg_info(node), val, env->modes[vnum]);
566 value_arr[vnum] = val;
568 mem = get_Store_mem(node);
569 block = get_nodes_block(node);
571 turn_into_tuple(node, pn_Store_max);
572 set_Tuple_pred(node, pn_Store_M, mem);
573 set_Tuple_pred(node, pn_Store_X_regular, new_r_Jmp(current_ir_graph, block));
574 set_Tuple_pred(node, pn_Store_X_except, new_Bad());
575 } else if (op == op_Phi && get_irn_mode(node) == mode_M) {
577 * found a memory Phi: Here, we must create new Phi nodes
579 block = get_nodes_block(node);
580 value_arr = get_irn_link(block);
582 n = get_Block_n_cfgpreds(block);
584 in = alloca(sizeof(*in) * n);
586 for (i = env->nvals - 1; i >= 0; --i) {
587 unk = new_Unknown(env->modes[i]);
588 for (j = n - 1; j >= 0; --j)
591 value_arr[i] = new_r_Phi(current_ir_graph, block, n, in, env->modes[i]);
593 l = obstack_alloc(&env->obst, sizeof(*l));
594 l->node = value_arr[i];
597 set_irn_link(value_arr[i], env->fix_phis);
604 * Walker: allocate the value array for every block.
606 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) {
624 if (Block_not_block_visited(block)) {
625 mark_Block_block_visited(block);
627 value_arr = get_irn_link(block);
630 return value_arr[vnum];
632 for (i = get_Block_n_cfgpreds(block) - 1; i >= 0; --i) {
633 ir_node *pred = get_Block_cfgpred(block, i);
635 res = find_vnum_value(get_nodes_block(pred), vnum);
646 static void fix_phis(env_t *env) {
648 ir_node *phi, *block, *pred, *val;
651 for (l = env->fix_phis; l; l = get_irn_link(phi)) {
654 block = get_nodes_block(phi);
655 for (i = get_irn_arity(phi) - 1; i >= 0; --i) {
656 pred = get_Block_cfgpred(block, i);
657 pred = get_nodes_block(pred);
659 inc_irg_block_visited(current_ir_graph);
660 val = find_vnum_value(pred, l->vnum);
663 set_irn_n(phi, i, val);
671 static void fix_loads(env_t *env) {
673 ir_node *load, *block, *pred, *val = NULL, *mem;
677 for (l = env->fix_loads; l; l = get_irn_link(load)) {
680 block = get_nodes_block(load);
681 for (i = get_Block_n_cfgpreds(block) - 1; i >= 0; --i) {
682 pred = get_Block_cfgpred(block, i);
683 pred = get_nodes_block(pred);
685 inc_irg_block_visited(current_ir_graph);
686 val = find_vnum_value(pred, l->vnum);
693 /* access of an uninitialized value */
694 val = new_Unknown(env->modes[l->vnum]);
697 /* Beware: A Load can contain a hidden conversion in Firm.
699 mode = get_Load_mode(load);
700 if (mode != get_irn_mode(val))
701 val = new_d_Conv(get_irn_dbg_info(load), val, mode);
703 mem = get_Load_mem(load);
705 turn_into_tuple(load, pn_Load_max);
706 set_Tuple_pred(load, pn_Load_M, mem);
707 set_Tuple_pred(load, pn_Load_res, val);
708 set_Tuple_pred(load, pn_Load_X_regular, new_r_Jmp(current_ir_graph, block));
709 set_Tuple_pred(load, pn_Load_X_except, new_Bad());
714 * Make scalar replacement.
716 * @param sels A set containing all Sel nodes that have a value number
717 * @param nvals The number of scalars.
718 * @param modes A flexible array, containing all the modes of
721 static void do_scalar_replacements(pset *sels, int nvals, ir_mode **modes) {
724 obstack_init(&env.obst);
728 env.fix_loads = NULL;
731 /* first step: allocate the value arrays for every block */
732 irg_block_walk_graph(current_ir_graph, NULL, alloc_value_arr, &env);
735 * second step: walk over the graph blockwise in topological order
736 * and fill the array as much as possible.
738 irg_walk_blkwise_graph(current_ir_graph, NULL, topologic_walker, &env);
740 /* third, fix the list of Phis, then the list of Loads */
744 obstack_free(&env.obst, NULL);
748 * Find possible scalar replacements
750 * @param irg The current ir graph.
752 void scalar_replacement_opt(ir_graph *irg) {
755 scalars_t key, *value;
763 if (! get_opt_scalar_replacement())
766 rem = current_ir_graph;
768 /* Call algorithm that computes the out edges */
769 assure_irg_outs(irg);
771 /* Find possible scalar replacements */
772 if (find_possible_replacements(irg)) {
774 if (get_opt_scalar_replacement_verbose()) {
775 printf("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));
800 if (get_opt_scalar_replacement_verbose()) {
801 if (is_Array_type(ent_type)) {
802 printf(" found array %s\n", get_entity_name(ent));
804 else if (is_Struct_type(ent_type)) {
805 printf(" found struct %s\n", get_entity_name(ent));
807 else if (is_atomic_type(ent_type))
808 printf(" found atomic value %s\n", get_entity_name(ent));
810 assert(0 && "Neither an array nor a struct or atomic value");
814 nvals = allocate_value_numbers(sels, ent, nvals, &modes);
818 if (get_opt_scalar_replacement_verbose()) {
819 printf(" %u values will be needed\n", nvals);
822 /* If scalars were found. */
824 do_scalar_replacements(sels, nvals, modes);
826 for (value = set_first(set_ent); value; value = set_next(set_ent)) {
827 remove_class_member(value->ent_owner, value->ent);
837 * We changed the graph, but did NOT introduce new blocks
838 * neither changed control flow, cf-backedges should be still
841 set_irg_outs_inconsistent(irg);
842 set_irg_loopinfo_inconsistent(irg);
846 current_ir_graph = rem;