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 Procedure cloning.
23 * @author Beyhan Veliev, Michael Beck
27 * The purpose is first to find and analyze functions, that are called
28 * with constant parameter(s).
29 * The second step is to optimize the function that are found from our
30 * analyze. Optimize mean to make a new function with parameters, that
31 * aren't be constant. The constant parameters of the function are placed
32 * in the function graph. They aren't be passed as parameters.
48 #include "proc_cloning.h"
49 #include "analyze_irg_args.h"
60 /* A macro to iterate sets.*/
61 #define ITERATE_SET(set_entries, entry) for(entry = set_first(set_entries); entry; entry = set_next(set_entries))
64 * This struct contains the information quadruple for a Call, which we need to
65 * decide if this function must be cloned.
67 typedef struct quadruple {
68 ir_entity *ent; /**< The entity of our Call. */
69 int pos; /**< Position of a constant argument of our Call. */
70 tarval *tv; /**< The tarval of this argument if Const node. */
71 ir_node **calls; /**< The list of all calls with the same characteristics */
75 * The quadruplets are hold in a sorted list
77 typedef struct entry {
78 quadruple_t q; /**< the quadruple */
79 float weight; /**< its weight */
80 struct entry *next; /**< link to the next one */
83 typedef struct q_set {
84 struct obstack obst; /**< an obstack containing all entries */
85 pset *map; /**< a hash map containing the quadruples */
86 entry_t *heavy_uses; /**< the ordered list of heavy uses */
90 * Compare two quadruplets.
92 * @return zero if they are identically, non-zero else
94 static int entry_cmp(const void *elt, const void *key) {
95 const entry_t *e1 = elt;
96 const entry_t *e2 = key;
98 return (e1->q.ent != e2->q.ent) || (e1->q.pos != e2->q.pos) || (e1->q.tv != e2->q.tv);
102 * Hash an element of type entry_t.
104 * @param entry The element to be hashed.
106 static int hash_entry(const entry_t *entry) {
107 return HASH_PTR(entry->q.ent) ^ HASH_PTR(entry->q.tv) ^ (entry->q.pos * 9);
111 * Free memory associated with a quadruplet.
113 static void kill_entry(entry_t *entry) {
114 if (entry->q.calls) {
115 DEL_ARR_F(entry->q.calls);
116 entry->q.calls = NULL;
121 * Process a call node.
123 * @param call A ir_node to be checked.
124 * @param callee The entity of the callee
125 * @param hmap The quadruple-set containing the calls with constant parameters
127 static void process_call(ir_node *call, ir_entity *callee, q_set *hmap) {
129 entry_t *key, *entry;
133 n_params = get_Call_n_params(call);
135 /* Beware: we cannot clone variadic parameters as well as the
136 * last non-variadic one, which might be needed for the va_start()
139 mtp = get_Call_type(call);
140 if (get_method_variadicity(mtp) != variadicity_non_variadic) {
141 n_params = get_method_first_variadic_param_index(mtp) - 1;
144 /* In this for loop we collect the calls, that have
145 an constant parameter. */
146 for (i = n_params - 1; i >= 0; --i) {
147 call_param = get_Call_param(call, i);
148 if (is_Const(call_param)) {
149 /* we have found a Call to collect and we save the informations,
152 hmap->map = new_pset(entry_cmp, 8);
154 key = obstack_alloc(&hmap->obst, sizeof(*key));
158 key->q.tv = get_Const_tarval(call_param);
163 /* We insert our information in the set, where we collect the calls.*/
164 entry = pset_insert(hmap->map, key, hash_entry(key));
167 obstack_free(&hmap->obst, key);
169 /* add the call to the list */
170 if (! entry->q.calls) {
171 entry->q.calls = NEW_ARR_F(ir_node *, 1);
172 entry->q.calls[0] = call;
174 ARR_APP1(ir_node *, entry->q.calls, call);
180 * Collect all calls in a ir_graph to a set.
182 * @param call A ir_node to be checked.
183 * @param env The quadruple-set containing the calls with constant parameters
185 static void collect_irg_calls(ir_node *call, void *env) {
190 /* We collect just "Call" nodes */
192 call_ptr = get_Call_ptr(call);
194 /* Call pointer must be a SymConst*/
195 if (op_SymConst != get_irn_op(call_ptr))
197 /* Call pointer must be the address of an entity.*/
198 if (get_SymConst_kind(call_ptr) != symconst_addr_ent)
201 callee = get_SymConst_entity(call_ptr);
203 /* we can only clone calls to existing entities */
204 if (get_entity_visibility(callee) == visibility_external_allocated)
207 process_call(call, callee, hmap);
212 * Make a name for a clone. The clone name is
213 * the name of the original method suffixed with "_cl_pos_nr".
214 * pos is the pos from our quadruplet and nr is a counter.
216 * @param id The ident of the cloned function.
217 * @param pos The "pos" from our quadruplet.
218 * @param nr A counter for the clones.
220 static ident *get_clone_ident(ident *id, int pos, unsigned nr) {
221 char clone_postfix[32];
223 snprintf(clone_postfix, sizeof(clone_postfix), "_cl_%d_%u", pos, nr);
225 return mangle(id, new_id_from_str(clone_postfix));
229 * Pre-Walker: Copies blocks and nodes from the original method graph
230 * to the cloned graph. Fixes the argument projection numbers for
231 * all arguments behind the removed one.
233 * @param irn A node from the original method graph.
234 * @param env The clone graph.
236 static void copy_nodes(ir_node *irn, void *env) {
237 ir_node *arg, *irg_args, *irn_copy;
239 ir_graph *clone_irg = env;
241 arg = get_irg_link(clone_irg);
242 irg_args = get_Proj_pred(arg);
244 /* Copy all nodes except the arg. */
246 copy_irn_to_irg(irn, clone_irg);
248 irn_copy = get_irn_link(irn);
250 /* Fix argument numbers */
251 if (is_Proj(irn) && get_Proj_pred(irn) == irg_args) {
252 proj_nr = get_Proj_proj(irn);
253 if (get_Proj_proj(arg) < proj_nr)
254 set_Proj_proj(irn_copy, proj_nr - 1);
259 * Post-walker: Set the predecessors of the copied nodes.
260 * The copied nodes are set as link of their original nodes. The links of
261 * "irn" predecessors are the predecessors of copied node.
263 static void set_preds(ir_node *irn, void *env) {
265 ir_node *irn_copy, *pred, *arg;
266 ir_graph *clone_irg = env;
268 arg = get_irg_link(clone_irg);
269 /* Arg is the method argument, that we have replaced by a constant.*/
273 irn_copy = get_irn_link(irn);
276 for (i = get_Block_n_cfgpreds(irn) - 1; i >= 0; i--) {
277 pred = get_Block_cfgpred(irn, i);
278 /* "End" block must be handled extra, because it is not matured.*/
279 if (get_irg_end_block(current_ir_graph) == irn)
280 add_immBlock_pred(get_irg_end_block(clone_irg), get_irn_link(pred));
282 set_Block_cfgpred(irn_copy, i, get_irn_link(pred));
285 /* First we set the block our copy if it is not a block.*/
286 set_nodes_block(irn_copy, get_irn_link(get_nodes_block(irn)));
287 if (get_irn_op(irn) == op_End) {
288 /* Handle the keep-alives. This must be done separately, because
289 the End node was NOT copied */
290 for (i = 0; i < get_End_n_keepalives(irn); ++i)
291 add_End_keepalive(irn_copy, get_irn_link(get_End_keepalive(irn, i)));
293 for (i = get_irn_arity(irn) - 1; i >= 0; i--) {
294 pred = get_irn_n(irn, i);
295 set_irn_n(irn_copy, i, get_irn_link(pred));
302 * Get the method argument at the position "pos".
304 * @param irg irg that must be cloned.
305 * @param pos The position of the argument.
307 static ir_node *get_irg_arg(ir_graph *irg, int pos) {
308 ir_node *irg_args = get_irg_args(irg), *arg = NULL;
311 /* Call algorithm that computes the out edges */
312 assure_irg_outs(irg);
314 /* Search the argument with the number pos.*/
315 for (i = get_irn_n_outs(irg_args) - 1; i >= 0; --i) {
316 ir_node *proj = get_irn_out(irg_args, i);
317 if (pos == get_Proj_proj(proj)) {
320 * More than one arg node found:
321 * We rely on the fact the only one arg exists, so do
322 * a cheap CSE in this case.
324 set_irn_out(irg_args, i, arg);
330 assert(arg && "Argument not found");
335 * Create a new graph for the clone of the method,
336 * that we want to clone.
338 * @param ent The entity of the method that must be cloned.
339 * @param q Our quadruplet.
341 static void create_clone_proc_irg(ir_entity *ent, quadruple_t *q) {
342 ir_graph *method_irg, *clone_irg;
343 ir_node *arg, *const_arg;
345 method_irg = get_entity_irg(ent);
347 /* We create the skeleton of the clone irg.*/
348 clone_irg = new_ir_graph(ent, 0);
350 arg = get_irg_arg(get_entity_irg(q->ent), q->pos);
351 /* we will replace the argument in position "q->pos" by this constant. */
352 const_arg = new_r_Const_type(
353 clone_irg, get_nodes_block(arg), get_irn_mode(arg), q->tv,
354 get_method_param_type(get_entity_type(q->ent), q->pos));
356 /* args copy in the cloned graph will be the const. */
357 set_irn_link(arg, const_arg);
359 /* Store the arg that will be replaced here, so we can easily detect it. */
360 set_irg_link(clone_irg, arg);
362 /* We copy the blocks and nodes, that must be in
363 the clone graph and set their predecessors. */
364 irg_walk_graph(method_irg, copy_nodes, set_preds, clone_irg);
366 /* The "cloned" graph must be matured. */
367 mature_immBlock(get_irg_end_block(clone_irg));
368 irg_finalize_cons(clone_irg);
372 * The function create a new entity type
373 * for our clone and set it to clone entity.
375 * @param q Contains information for the method to clone.
376 * @param ent The entity of the clone.
377 * @param nr A pointer to the counter of clones.
379 static void change_entity_type(quadruple_t *q, ir_entity *ent, unsigned *nr) {
380 ir_type *mtp, *new_mtp, *tp;
382 int i, j, n_params, n_ress;
384 mtp = get_entity_type(q->ent);
385 tp_name = get_clone_ident(get_type_ident(mtp), q->pos, (*nr)++);
386 n_params = get_method_n_params(mtp);
387 n_ress = get_method_n_ress(mtp);
389 /* Create the new type for our clone. It must have one parameter
390 less then the original.*/
391 new_mtp = new_type_method(tp_name, n_params - 1, n_ress);
393 /* We must set the type of the methods parameters.*/
394 for (i = j = 0; i < n_params; ++i) {
396 /* This is the position of the argument, that we have
400 tp = get_method_param_type(mtp, i);
401 set_method_param_type(new_mtp, j++, tp);
403 /* Copy the methods result types. */
404 for (i = 0; i < n_ress; ++i) {
405 tp = get_method_res_type(mtp, i);
406 set_method_res_type(new_mtp, i, tp);
408 set_entity_type(ent, new_mtp);
412 * Make a clone of a method.
414 * @param q Contains information for the method to clone.
416 static ir_entity *clone_method(quadruple_t *q) {
417 ir_entity *new_entity;
421 /* A counter for the clones.*/
422 static unsigned nr = 0;
424 /* We get a new ident for our clone method.*/
425 clone_ident = get_clone_ident(get_entity_ident(q->ent), q->pos, nr);
426 /* We get our entity for the clone method. */
427 new_entity = copy_entity_name(q->ent, clone_ident);
429 /* a cloned entity is always local */
430 set_entity_visibility(new_entity, visibility_local);
432 /* set a ld name here: Should we mangle this ? */
433 set_entity_ld_ident(new_entity, get_entity_ident(new_entity));
435 /* set a new type here. */
436 change_entity_type(q, new_entity, &nr);
438 /* We need now a new ir_graph for our clone method. */
439 create_clone_proc_irg(new_entity, q);
441 /* We must set the atomic value of our "new_entity". */
442 sym.entity_p = new_entity;
443 rem = current_ir_graph;
444 current_ir_graph = get_const_code_irg();
445 new_entity->value = new_SymConst(sym, symconst_addr_ent);
446 current_ir_graph = rem;
448 /* The "new_entity" don't have this information. */
449 new_entity->attr.mtd_attr.param_access = NULL;
450 new_entity->attr.mtd_attr.param_weight = NULL;
456 * Creates a new "cloned" Call node and return it.
458 * @param call The call that must be cloned.
459 * @param new_entity The entity of the cloned function.
460 * @param pos The position of the replaced parameter of this call.
462 static ir_node *new_cl_Call(ir_node *call, ir_entity *new_entity, int pos) {
465 int i, n_params, new_params = 0;
468 ir_graph *irg = get_irn_irg(call);
469 ir_node *bl = get_nodes_block(call);
471 sym.entity_p = new_entity;
472 callee = new_r_SymConst(irg, bl, sym, symconst_addr_ent);
474 mtp = get_entity_type(new_entity);
475 n_params = get_Call_n_params(call);
476 NEW_ARR_A(ir_node *, in, n_params - 1);
478 /* we save the parameters of the new call in the array "in" without the
479 * parameter in position "pos", that is replaced with a constant.*/
480 for (i = 0; i < n_params; i++){
482 in[new_params++] = get_Call_param(call, i);
484 /* Create and return the new Call. */
485 return new_r_Call(irg, bl, get_Call_mem(call),
486 callee, n_params - 1, in, get_entity_type(new_entity));
490 * Exchange all Calls stored in the quadruplet to Calls of the cloned entity.
492 * @param q The quadruple
493 * @param cloned_ent The entity of the new function that must be called
496 static void exchange_calls(quadruple_t *q, ir_entity *cloned_ent) {
498 ir_node *new_call, *call;
501 /* We iterate the list of the "call".*/
502 for (i = 0; i < ARR_LEN(q->calls); ++i) {
505 /* A clone exist and the copy of "call" in this
506 * clone graph must be exchanged with new one.*/
507 new_call = new_cl_Call(call, cloned_ent, pos);
508 exchange(call, new_call);
513 * The weight formula:
514 * We save one instruction in every caller and param_weight instructions
517 static float calculate_weight(const entry_t *entry) {
518 return ARR_LEN(entry->q.calls) *
519 (get_method_param_weight(entry->q.ent, entry->q.pos) + 1);
523 * After we exchanged all calls, some entries on the list for
524 * the next cloned entity may get invalid, so we have to check
525 * them and may even update the list of heavy uses.
527 static void reorder_weights(q_set *hmap, float threshold) {
528 entry_t **adr, *p, *entry;
533 entry = hmap->heavy_uses;
537 len = ARR_LEN(entry->q.calls);
538 for (i = 0; i < len; ++i) {
539 ir_node *ptr, *call = entry->q.calls[i];
541 /* might be exchanged, so skip Id nodes here. */
542 call = skip_Id(call);
544 /* we know, that a SymConst is here */
545 ptr = get_Call_ptr(call);
546 assert(get_irn_op(ptr) == op_SymConst);
548 callee = get_SymConst_entity(ptr);
549 if (callee != entry->q.ent) {
551 * This call is already changed because of a previous
552 * optimization. Remove it from the list.
555 entry->q.calls[i] = entry->q.calls[len];
556 entry->q.calls[len] = NULL;
558 /* the new call should be processed */
559 process_call(call, callee, hmap);
564 /* the length might be changed */
565 ARR_SHRINKLEN(entry->q.calls, len);
567 /* recalculate the weight and resort the heavy uses map */
568 entry->weight = calculate_weight(entry);
570 if (len <= 0 || entry->weight < threshold) {
571 hmap->heavy_uses = entry->next;
574 /* we have changed the list, check the next one */
579 for (p = entry->next; p && entry->weight < p->weight; p = p->next) {
584 hmap->heavy_uses = entry->next;
587 entry = hmap->heavy_uses;
589 /* we have changed the list, check the next one */
595 * Do the procedure cloning. Evaluate a heuristic weight for every
596 * call(..., Const, ...). If the weight is bigger than threshold,
597 * clone the entity and fix the calls.
599 void proc_cloning(float threshold) {
600 entry_t *entry = NULL, *p;
605 obstack_init(&hmap.obst);
607 hmap.heavy_uses = NULL;
609 /* initially fill our map by visiting all irgs */
610 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
611 irg = get_irp_irg(i);
612 irg_walk_graph(irg, collect_irg_calls, NULL, &hmap);
615 /* We have the "Call" nodes to optimize in set "set_entries". Our algorithm
616 replace one constant parameter and make a new "Call" node for all found "Calls". It exchange the
617 old one with the new one and the algorithm is called with the new "Call".
619 while (hmap.map || hmap.heavy_uses) {
620 /* We iterate the set and arrange the element of the set in a list.
621 The elements are arranged dependent of their value descending.*/
623 foreach_pset(hmap.map, entry) {
624 entry->weight = calculate_weight(entry);
627 * Do not put entry with a weight < threshold in the list
629 if (entry->weight < threshold) {
634 /* put entry in the heavy uses list */
636 if (! hmap.heavy_uses)
637 hmap.heavy_uses = entry;
639 if (entry->weight >= hmap.heavy_uses->weight) {
640 entry->next = hmap.heavy_uses;
641 hmap.heavy_uses = entry;
643 for (p = hmap.heavy_uses; p->next; p = p->next) {
644 if (entry->weight >= p->next->weight) {
645 entry->next = p->next;
659 /* Print some information about the list. */
660 printf("-----------------\n");
661 for (entry = hmap.heavy_uses; entry; entry = entry->next) {
662 printf("\nweight: is %f\n", entry->weight);
663 ir_printf("Call for Method %E\n", entry->q.ent);
664 printf("Position %i\n", entry->q.pos);
665 ir_printf("Value %T\n", entry->q.tv);
668 entry = hmap.heavy_uses;
670 ir_entity *ent = clone_method(&entry->q);
672 hmap.heavy_uses = entry->next;
674 /* We must exchange the copies of this call in all clones too.*/
675 exchange_calls(&entry->q, ent);
679 * after we exchanged all calls, some entries on the list for
680 * the next cloned entity may get invalid, so we have to check
681 * them and may even update the list of heavy uses.
683 reorder_weights(&hmap, threshold);
686 obstack_free(&hmap.obst, NULL);