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.
40 #include "iroptimize.h"
46 #include "analyze_irg_args.h"
56 /* A macro to iterate sets.*/
57 #define ITERATE_SET(set_entries, entry) for(entry = set_first(set_entries); entry; entry = set_next(set_entries))
60 * This struct contains the information quadruple for a Call, which we need to
61 * decide if this function must be cloned.
63 typedef struct quadruple {
64 ir_entity *ent; /**< The entity of our Call. */
65 int pos; /**< Position of a constant argument of our Call. */
66 tarval *tv; /**< The tarval of this argument if Const node. */
67 ir_node **calls; /**< The list of all calls with the same characteristics */
71 * The quadruplets are hold in a sorted list
73 typedef struct entry {
74 quadruple_t q; /**< the quadruple */
75 float weight; /**< its weight */
76 struct entry *next; /**< link to the next one */
79 typedef struct q_set {
80 struct obstack obst; /**< an obstack containing all entries */
81 pset *map; /**< a hash map containing the quadruples */
82 entry_t *heavy_uses; /**< the ordered list of heavy uses */
86 * Compare two quadruplets.
88 * @return zero if they are identically, non-zero else
90 static int entry_cmp(const void *elt, const void *key) {
91 const entry_t *e1 = elt;
92 const entry_t *e2 = key;
94 return (e1->q.ent != e2->q.ent) || (e1->q.pos != e2->q.pos) || (e1->q.tv != e2->q.tv);
98 * Hash an element of type entry_t.
100 * @param entry The element to be hashed.
102 static int hash_entry(const entry_t *entry) {
103 return HASH_PTR(entry->q.ent) ^ HASH_PTR(entry->q.tv) ^ (entry->q.pos * 9);
107 * Free memory associated with a quadruplet.
109 static void kill_entry(entry_t *entry) {
110 if (entry->q.calls) {
111 DEL_ARR_F(entry->q.calls);
112 entry->q.calls = NULL;
117 * Process a call node.
119 * @param call A ir_node to be checked.
120 * @param callee The entity of the callee
121 * @param hmap The quadruple-set containing the calls with constant parameters
123 static void process_call(ir_node *call, ir_entity *callee, q_set *hmap) {
125 entry_t *key, *entry;
129 n_params = get_Call_n_params(call);
131 /* Beware: we cannot clone variadic parameters as well as the
132 * last non-variadic one, which might be needed for the va_start()
135 mtp = get_Call_type(call);
136 if (get_method_variadicity(mtp) != variadicity_non_variadic) {
137 n_params = get_method_first_variadic_param_index(mtp) - 1;
140 /* In this for loop we collect the calls, that have
141 an constant parameter. */
142 for (i = n_params - 1; i >= 0; --i) {
143 call_param = get_Call_param(call, i);
144 if (is_Const(call_param)) {
145 /* we have found a Call to collect and we save the informations,
148 hmap->map = new_pset(entry_cmp, 8);
150 key = obstack_alloc(&hmap->obst, sizeof(*key));
154 key->q.tv = get_Const_tarval(call_param);
159 /* We insert our information in the set, where we collect the calls.*/
160 entry = pset_insert(hmap->map, key, hash_entry(key));
163 obstack_free(&hmap->obst, key);
165 /* add the call to the list */
166 if (! entry->q.calls) {
167 entry->q.calls = NEW_ARR_F(ir_node *, 1);
168 entry->q.calls[0] = call;
170 ARR_APP1(ir_node *, entry->q.calls, call);
176 * Collect all calls in a ir_graph to a set.
178 * @param call A ir_node to be checked.
179 * @param env The quadruple-set containing the calls with constant parameters
181 static void collect_irg_calls(ir_node *call, void *env) {
186 /* We collect just "Call" nodes */
188 call_ptr = get_Call_ptr(call);
190 /* Call pointer must be a SymConst*/
191 if (op_SymConst != get_irn_op(call_ptr))
193 /* Call pointer must be the address of an entity.*/
194 if (get_SymConst_kind(call_ptr) != symconst_addr_ent)
197 callee = get_SymConst_entity(call_ptr);
199 /* we can only clone calls to existing entities */
200 if (get_entity_visibility(callee) == visibility_external_allocated)
203 process_call(call, callee, hmap);
208 * Make a name for a clone. The clone name is
209 * the name of the original method suffixed with "_cl_pos_nr".
210 * pos is the pos from our quadruplet and nr is a counter.
212 * @param id The ident of the cloned function.
213 * @param pos The "pos" from our quadruplet.
214 * @param nr A counter for the clones.
216 static ident *get_clone_ident(ident *id, int pos, unsigned nr) {
217 char clone_postfix[32];
219 snprintf(clone_postfix, sizeof(clone_postfix), "_cl_%d_%u", pos, nr);
221 return mangle(id, new_id_from_str(clone_postfix));
225 * Pre-Walker: Copies blocks and nodes from the original method graph
226 * to the cloned graph. Fixes the argument projection numbers for
227 * all arguments behind the removed one.
229 * @param irn A node from the original method graph.
230 * @param env The clone graph.
232 static void copy_nodes(ir_node *irn, void *env) {
233 ir_node *arg, *irg_args, *irn_copy;
235 ir_graph *clone_irg = env;
237 arg = get_irg_link(clone_irg);
238 irg_args = get_Proj_pred(arg);
240 /* Copy all nodes except the arg. */
242 copy_irn_to_irg(irn, clone_irg);
244 irn_copy = get_irn_link(irn);
246 /* Fix argument numbers */
247 if (is_Proj(irn) && get_Proj_pred(irn) == irg_args) {
248 proj_nr = get_Proj_proj(irn);
249 if (get_Proj_proj(arg) < proj_nr)
250 set_Proj_proj(irn_copy, proj_nr - 1);
255 * Post-walker: Set the predecessors of the copied nodes.
256 * The copied nodes are set as link of their original nodes. The links of
257 * "irn" predecessors are the predecessors of copied node.
259 static void set_preds(ir_node *irn, void *env) {
261 ir_node *irn_copy, *pred, *arg;
262 ir_graph *clone_irg = env;
264 arg = get_irg_link(clone_irg);
265 /* Arg is the method argument, that we have replaced by a constant.*/
269 irn_copy = get_irn_link(irn);
272 for (i = get_Block_n_cfgpreds(irn) - 1; i >= 0; i--) {
273 pred = get_Block_cfgpred(irn, i);
274 /* "End" block must be handled extra, because it is not matured.*/
275 if (get_irg_end_block(current_ir_graph) == irn)
276 add_immBlock_pred(get_irg_end_block(clone_irg), get_irn_link(pred));
278 set_Block_cfgpred(irn_copy, i, get_irn_link(pred));
281 /* First we set the block our copy if it is not a block.*/
282 set_nodes_block(irn_copy, get_irn_link(get_nodes_block(irn)));
283 if (get_irn_op(irn) == op_End) {
284 /* Handle the keep-alives. This must be done separately, because
285 the End node was NOT copied */
286 for (i = 0; i < get_End_n_keepalives(irn); ++i)
287 add_End_keepalive(irn_copy, get_irn_link(get_End_keepalive(irn, i)));
289 for (i = get_irn_arity(irn) - 1; i >= 0; i--) {
290 pred = get_irn_n(irn, i);
291 set_irn_n(irn_copy, i, get_irn_link(pred));
298 * Get the method argument at the position "pos".
300 * @param irg irg that must be cloned.
301 * @param pos The position of the argument.
303 static ir_node *get_irg_arg(ir_graph *irg, int pos) {
304 ir_node *irg_args = get_irg_args(irg), *arg = NULL;
307 /* Call algorithm that computes the out edges */
308 assure_irg_outs(irg);
310 /* Search the argument with the number pos.*/
311 for (i = get_irn_n_outs(irg_args) - 1; i >= 0; --i) {
312 ir_node *proj = get_irn_out(irg_args, i);
313 if (pos == get_Proj_proj(proj)) {
316 * More than one arg node found:
317 * We rely on the fact the only one arg exists, so do
318 * a cheap CSE in this case.
320 set_irn_out(irg_args, i, arg);
326 assert(arg && "Argument not found");
331 * Create a new graph for the clone of the method,
332 * that we want to clone.
334 * @param ent The entity of the method that must be cloned.
335 * @param q Our quadruplet.
337 static void create_clone_proc_irg(ir_entity *ent, quadruple_t *q) {
338 ir_graph *method_irg, *clone_irg;
339 ir_node *arg, *const_arg;
341 method_irg = get_entity_irg(ent);
343 /* We create the skeleton of the clone irg.*/
344 clone_irg = new_ir_graph(ent, 0);
346 arg = get_irg_arg(get_entity_irg(q->ent), q->pos);
347 /* we will replace the argument in position "q->pos" by this constant. */
348 const_arg = new_r_Const_type(
349 clone_irg, get_nodes_block(arg), get_irn_mode(arg), q->tv,
350 get_method_param_type(get_entity_type(q->ent), q->pos));
352 /* args copy in the cloned graph will be the const. */
353 set_irn_link(arg, const_arg);
355 /* Store the arg that will be replaced here, so we can easily detect it. */
356 set_irg_link(clone_irg, arg);
358 /* We copy the blocks and nodes, that must be in
359 the clone graph and set their predecessors. */
360 irg_walk_graph(method_irg, copy_nodes, set_preds, clone_irg);
362 /* The "cloned" graph must be matured. */
363 mature_immBlock(get_irg_end_block(clone_irg));
364 irg_finalize_cons(clone_irg);
368 * The function create a new entity type
369 * for our clone and set it to clone entity.
371 * @param q Contains information for the method to clone.
372 * @param ent The entity of the clone.
373 * @param nr A pointer to the counter of clones.
375 static void change_entity_type(quadruple_t *q, ir_entity *ent, unsigned *nr) {
376 ir_type *mtp, *new_mtp, *tp;
378 int i, j, n_params, n_ress;
380 mtp = get_entity_type(q->ent);
381 tp_name = get_clone_ident(get_type_ident(mtp), q->pos, (*nr)++);
382 n_params = get_method_n_params(mtp);
383 n_ress = get_method_n_ress(mtp);
385 /* Create the new type for our clone. It must have one parameter
386 less then the original.*/
387 new_mtp = new_type_method(tp_name, n_params - 1, n_ress);
389 /* We must set the type of the methods parameters.*/
390 for (i = j = 0; i < n_params; ++i) {
392 /* This is the position of the argument, that we have
396 tp = get_method_param_type(mtp, i);
397 set_method_param_type(new_mtp, j++, tp);
399 /* Copy the methods result types. */
400 for (i = 0; i < n_ress; ++i) {
401 tp = get_method_res_type(mtp, i);
402 set_method_res_type(new_mtp, i, tp);
404 set_entity_type(ent, new_mtp);
408 * Make a clone of a method.
410 * @param q Contains information for the method to clone.
412 static ir_entity *clone_method(quadruple_t *q) {
413 ir_entity *new_entity;
417 /* A counter for the clones.*/
418 static unsigned nr = 0;
420 /* We get a new ident for our clone method.*/
421 clone_ident = get_clone_ident(get_entity_ident(q->ent), q->pos, nr);
422 /* We get our entity for the clone method. */
423 new_entity = copy_entity_name(q->ent, clone_ident);
425 /* a cloned entity is always local */
426 set_entity_visibility(new_entity, visibility_local);
428 /* set a ld name here: Should we mangle this ? */
429 set_entity_ld_ident(new_entity, get_entity_ident(new_entity));
431 /* set a new type here. */
432 change_entity_type(q, new_entity, &nr);
434 /* We need now a new ir_graph for our clone method. */
435 create_clone_proc_irg(new_entity, q);
437 /* We must set the atomic value of our "new_entity". */
438 sym.entity_p = new_entity;
439 rem = current_ir_graph;
440 current_ir_graph = get_const_code_irg();
441 new_entity->value = new_SymConst(sym, symconst_addr_ent);
442 current_ir_graph = rem;
444 /* The "new_entity" don't have this information. */
445 new_entity->attr.mtd_attr.param_access = NULL;
446 new_entity->attr.mtd_attr.param_weight = NULL;
452 * Creates a new "cloned" Call node and return it.
454 * @param call The call that must be cloned.
455 * @param new_entity The entity of the cloned function.
456 * @param pos The position of the replaced parameter of this call.
458 static ir_node *new_cl_Call(ir_node *call, ir_entity *new_entity, int pos) {
461 int i, n_params, new_params = 0;
464 ir_graph *irg = get_irn_irg(call);
465 ir_node *bl = get_nodes_block(call);
467 sym.entity_p = new_entity;
468 callee = new_r_SymConst(irg, bl, sym, symconst_addr_ent);
470 mtp = get_entity_type(new_entity);
471 n_params = get_Call_n_params(call);
472 NEW_ARR_A(ir_node *, in, n_params - 1);
474 /* we save the parameters of the new call in the array "in" without the
475 * parameter in position "pos", that is replaced with a constant.*/
476 for (i = 0; i < n_params; i++){
478 in[new_params++] = get_Call_param(call, i);
480 /* Create and return the new Call. */
481 return new_r_Call(irg, bl, get_Call_mem(call),
482 callee, n_params - 1, in, get_entity_type(new_entity));
486 * Exchange all Calls stored in the quadruplet to Calls of the cloned entity.
488 * @param q The quadruple
489 * @param cloned_ent The entity of the new function that must be called
492 static void exchange_calls(quadruple_t *q, ir_entity *cloned_ent) {
494 ir_node *new_call, *call;
497 /* We iterate the list of the "call".*/
498 for (i = 0; i < ARR_LEN(q->calls); ++i) {
501 /* A clone exist and the copy of "call" in this
502 * clone graph must be exchanged with new one.*/
503 new_call = new_cl_Call(call, cloned_ent, pos);
504 exchange(call, new_call);
509 * The weight formula:
510 * We save one instruction in every caller and param_weight instructions
513 static float calculate_weight(const entry_t *entry) {
514 return ARR_LEN(entry->q.calls) *
515 (get_method_param_weight(entry->q.ent, entry->q.pos) + 1);
519 * After we exchanged all calls, some entries on the list for
520 * the next cloned entity may get invalid, so we have to check
521 * them and may even update the list of heavy uses.
523 static void reorder_weights(q_set *hmap, float threshold) {
524 entry_t **adr, *p, *entry;
529 entry = hmap->heavy_uses;
533 len = ARR_LEN(entry->q.calls);
534 for (i = 0; i < len; ++i) {
535 ir_node *ptr, *call = entry->q.calls[i];
537 /* might be exchanged, so skip Id nodes here. */
538 call = skip_Id(call);
540 /* we know, that a SymConst is here */
541 ptr = get_Call_ptr(call);
542 assert(get_irn_op(ptr) == op_SymConst);
544 callee = get_SymConst_entity(ptr);
545 if (callee != entry->q.ent) {
547 * This call is already changed because of a previous
548 * optimization. Remove it from the list.
551 entry->q.calls[i] = entry->q.calls[len];
552 entry->q.calls[len] = NULL;
554 /* the new call should be processed */
555 process_call(call, callee, hmap);
560 /* the length might be changed */
561 ARR_SHRINKLEN(entry->q.calls, len);
563 /* recalculate the weight and resort the heavy uses map */
564 entry->weight = calculate_weight(entry);
566 if (len <= 0 || entry->weight < threshold) {
567 hmap->heavy_uses = entry->next;
570 /* we have changed the list, check the next one */
575 for (p = entry->next; p && entry->weight < p->weight; p = p->next) {
580 hmap->heavy_uses = entry->next;
583 entry = hmap->heavy_uses;
585 /* we have changed the list, check the next one */
591 * Do the procedure cloning. Evaluate a heuristic weight for every
592 * call(..., Const, ...). If the weight is bigger than threshold,
593 * clone the entity and fix the calls.
595 void proc_cloning(float threshold) {
596 entry_t *entry = NULL, *p;
601 obstack_init(&hmap.obst);
603 hmap.heavy_uses = NULL;
605 /* initially fill our map by visiting all irgs */
606 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
607 irg = get_irp_irg(i);
608 irg_walk_graph(irg, collect_irg_calls, NULL, &hmap);
611 /* We have the "Call" nodes to optimize in set "set_entries". Our algorithm
612 replace one constant parameter and make a new "Call" node for all found "Calls". It exchange the
613 old one with the new one and the algorithm is called with the new "Call".
615 while (hmap.map || hmap.heavy_uses) {
616 /* We iterate the set and arrange the element of the set in a list.
617 The elements are arranged dependent of their value descending.*/
619 foreach_pset(hmap.map, entry) {
620 entry->weight = calculate_weight(entry);
623 * Do not put entry with a weight < threshold in the list
625 if (entry->weight < threshold) {
630 /* put entry in the heavy uses list */
632 if (! hmap.heavy_uses)
633 hmap.heavy_uses = entry;
635 if (entry->weight >= hmap.heavy_uses->weight) {
636 entry->next = hmap.heavy_uses;
637 hmap.heavy_uses = entry;
639 for (p = hmap.heavy_uses; p->next; p = p->next) {
640 if (entry->weight >= p->next->weight) {
641 entry->next = p->next;
655 /* Print some information about the list. */
656 printf("-----------------\n");
657 for (entry = hmap.heavy_uses; entry; entry = entry->next) {
658 printf("\nweight: is %f\n", entry->weight);
659 ir_printf("Call for Method %E\n", entry->q.ent);
660 printf("Position %i\n", entry->q.pos);
661 ir_printf("Value %T\n", entry->q.tv);
664 entry = hmap.heavy_uses;
666 ir_entity *ent = clone_method(&entry->q);
668 hmap.heavy_uses = entry->next;
670 /* We must exchange the copies of this call in all clones too.*/
671 exchange_calls(&entry->q, ent);
675 * after we exchanged all calls, some entries on the list for
676 * the next cloned entity may get invalid, so we have to check
677 * them and may even update the list of heavy uses.
679 reorder_weights(&hmap, threshold);
682 obstack_free(&hmap.obst, NULL);