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 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"
57 * This struct contains the information quadruple for a Call, which we need to
58 * decide if this function must be cloned.
60 typedef struct quadruple {
61 ir_entity *ent; /**< The entity of our Call. */
62 int pos; /**< Position of a constant argument of our Call. */
63 tarval *tv; /**< The tarval of this argument if Const node. */
64 ir_node **calls; /**< The list of all calls with the same characteristics */
68 * The quadruplets are hold in a sorted list
70 typedef struct entry {
71 quadruple_t q; /**< the quadruple */
72 float weight; /**< its weight */
73 struct entry *next; /**< link to the next one */
76 typedef struct q_set {
77 struct obstack obst; /**< an obstack containing all entries */
78 pset *map; /**< a hash map containing the quadruples */
79 entry_t *heavy_uses; /**< the ordered list of heavy uses */
83 * Compare two quadruplets.
85 * @return zero if they are identically, non-zero else
87 static int entry_cmp(const void *elt, const void *key) {
88 const entry_t *e1 = elt;
89 const entry_t *e2 = key;
91 return (e1->q.ent != e2->q.ent) || (e1->q.pos != e2->q.pos) || (e1->q.tv != e2->q.tv);
95 * Hash an element of type entry_t.
97 * @param entry The element to be hashed.
99 static int hash_entry(const entry_t *entry) {
100 return HASH_PTR(entry->q.ent) ^ HASH_PTR(entry->q.tv) ^ (entry->q.pos * 9);
104 * Free memory associated with a quadruplet.
106 static void kill_entry(entry_t *entry) {
107 if (entry->q.calls) {
108 DEL_ARR_F(entry->q.calls);
109 entry->q.calls = NULL;
114 * Process a call node.
116 * @param call A ir_node to be checked.
117 * @param callee The entity of the callee
118 * @param hmap The quadruple-set containing the calls with constant parameters
120 static void process_call(ir_node *call, ir_entity *callee, q_set *hmap) {
122 entry_t *key, *entry;
126 n_params = get_Call_n_params(call);
128 /* Beware: we cannot clone variadic parameters as well as the
129 * last non-variadic one, which might be needed for the va_start()
132 mtp = get_Call_type(call);
133 if (get_method_variadicity(mtp) != variadicity_non_variadic) {
134 n_params = get_method_first_variadic_param_index(mtp) - 1;
137 /* In this for loop we collect the calls, that have
138 an constant parameter. */
139 for (i = n_params - 1; i >= 0; --i) {
140 call_param = get_Call_param(call, i);
141 if (is_Const(call_param)) {
142 /* we have found a Call to collect and we save the informations,
145 hmap->map = new_pset(entry_cmp, 8);
147 key = obstack_alloc(&hmap->obst, sizeof(*key));
151 key->q.tv = get_Const_tarval(call_param);
156 /* We insert our information in the set, where we collect the calls.*/
157 entry = pset_insert(hmap->map, key, hash_entry(key));
160 obstack_free(&hmap->obst, key);
162 /* add the call to the list */
163 if (! entry->q.calls) {
164 entry->q.calls = NEW_ARR_F(ir_node *, 1);
165 entry->q.calls[0] = call;
167 ARR_APP1(ir_node *, entry->q.calls, call);
173 * Collect all calls in a ir_graph to a set.
175 * @param call A ir_node to be checked.
176 * @param env The quadruple-set containing the calls with constant parameters
178 static void collect_irg_calls(ir_node *call, void *env) {
183 /* We collect just "Call" nodes */
185 call_ptr = get_Call_ptr(call);
187 if (! is_Global(call_ptr))
190 callee = get_Global_entity(call_ptr);
192 /* we can only clone calls to existing entities */
193 if (get_entity_visibility(callee) == visibility_external_allocated)
196 /* we cannot clone calls to weak functions */
197 if (get_entity_additional_properties(callee) & mtp_property_weak)
200 process_call(call, callee, hmap);
205 * Make a name for a clone. The clone name is
206 * the name of the original method suffixed with "_cl_pos_nr".
207 * pos is the pos from our quadruplet and nr is a counter.
209 * @param id The ident of the cloned function.
210 * @param pos The "pos" from our quadruplet.
211 * @param nr A counter for the clones.
213 static ident *get_clone_ident(ident *id, int pos, unsigned nr) {
214 char clone_postfix[32];
216 snprintf(clone_postfix, sizeof(clone_postfix), "_cl_%d_%u", pos, nr);
218 return mangle(id, new_id_from_str(clone_postfix));
222 * Pre-Walker: Copies blocks and nodes from the original method graph
223 * to the cloned graph. Fixes the argument projection numbers for
224 * all arguments behind the removed one.
226 * @param irn A node from the original method graph.
227 * @param env The clone graph.
229 static void copy_nodes(ir_node *irn, void *env) {
230 ir_node *arg, *irg_args, *irn_copy;
232 ir_graph *clone_irg = env;
234 arg = get_irg_link(clone_irg);
235 irg_args = get_Proj_pred(arg);
237 /* Copy all nodes except the arg. */
239 copy_irn_to_irg(irn, clone_irg);
241 irn_copy = get_irn_link(irn);
243 /* Fix argument numbers */
244 if (is_Proj(irn) && get_Proj_pred(irn) == irg_args) {
245 proj_nr = get_Proj_proj(irn);
246 if (get_Proj_proj(arg) < proj_nr)
247 set_Proj_proj(irn_copy, proj_nr - 1);
252 * Post-walker: Set the predecessors of the copied nodes.
253 * The copied nodes are set as link of their original nodes. The links of
254 * "irn" predecessors are the predecessors of copied node.
256 static void set_preds(ir_node *irn, void *env) {
258 ir_node *irn_copy, *pred, *arg;
259 ir_graph *clone_irg = env;
261 arg = get_irg_link(clone_irg);
262 /* Arg is the method argument, that we have replaced by a constant.*/
266 irn_copy = get_irn_link(irn);
269 set_Block_MacroBlock(irn_copy, get_irn_link(get_Block_MacroBlock(irn)));
270 for (i = get_Block_n_cfgpreds(irn) - 1; i >= 0; i--) {
271 pred = get_Block_cfgpred(irn, i);
272 /* "End" block must be handled extra, because it is not matured.*/
273 if (get_irg_end_block(current_ir_graph) == irn)
274 add_immBlock_pred(get_irg_end_block(clone_irg), get_irn_link(pred));
276 set_Block_cfgpred(irn_copy, i, get_irn_link(pred));
279 /* First we set the block our copy if it is not a block.*/
280 set_nodes_block(irn_copy, get_irn_link(get_nodes_block(irn)));
282 /* Handle the keep-alives. This must be done separately, because
283 the End node was NOT copied */
284 for (i = 0; i < get_End_n_keepalives(irn); ++i)
285 add_End_keepalive(irn_copy, get_irn_link(get_End_keepalive(irn, i)));
287 for (i = get_irn_arity(irn) - 1; i >= 0; i--) {
288 pred = get_irn_n(irn, i);
289 set_irn_n(irn_copy, i, get_irn_link(pred));
296 * Get the method argument at the position "pos".
298 * @param irg irg that must be cloned.
299 * @param pos The position of the argument.
301 static ir_node *get_irg_arg(ir_graph *irg, int pos) {
302 ir_node *irg_args = get_irg_args(irg), *arg = NULL;
305 /* Call algorithm that computes the out edges */
306 assure_irg_outs(irg);
308 /* Search the argument with the number pos.*/
309 for (i = get_irn_n_outs(irg_args) - 1; i >= 0; --i) {
310 ir_node *proj = get_irn_out(irg_args, i);
311 if (pos == get_Proj_proj(proj)) {
314 * More than one arg node found:
315 * We rely on the fact the only one arg exists, so do
316 * a cheap CSE in this case.
318 set_irn_out(irg_args, i, arg, 0);
324 assert(arg && "Argument not found");
329 * Create a new graph for the clone of the method,
330 * that we want to clone.
332 * @param ent The entity of the method that must be cloned.
333 * @param q Our quadruplet.
335 static void create_clone_proc_irg(ir_entity *ent, quadruple_t *q) {
336 ir_graph *method_irg, *clone_irg;
337 ir_node *arg, *const_arg;
339 method_irg = get_entity_irg(ent);
341 /* We create the skeleton of the clone irg.*/
342 clone_irg = new_ir_graph(ent, 0);
344 arg = get_irg_arg(get_entity_irg(q->ent), q->pos);
345 /* we will replace the argument in position "q->pos" by this constant. */
346 const_arg = new_r_Const_type(
347 clone_irg, get_nodes_block(arg), get_irn_mode(arg), q->tv,
348 get_method_param_type(get_entity_type(q->ent), q->pos));
350 /* args copy in the cloned graph will be the const. */
351 set_irn_link(arg, const_arg);
353 /* Store the arg that will be replaced here, so we can easily detect it. */
354 set_irg_link(clone_irg, arg);
356 /* We copy the blocks and nodes, that must be in
357 the clone graph and set their predecessors. */
358 irg_walk_graph(method_irg, copy_nodes, set_preds, clone_irg);
360 /* The "cloned" graph must be matured. */
361 mature_immBlock(get_irg_end_block(clone_irg));
362 irg_finalize_cons(clone_irg);
366 * The function create a new entity type
367 * for our clone and set it to clone entity.
369 * @param q Contains information for the method to clone.
370 * @param ent The entity of the clone.
371 * @param nr A pointer to the counter of clones.
373 static void change_entity_type(quadruple_t *q, ir_entity *ent, unsigned *nr) {
374 ir_type *mtp, *new_mtp, *tp;
376 int i, j, n_params, n_ress;
378 mtp = get_entity_type(q->ent);
379 tp_name = get_clone_ident(get_type_ident(mtp), q->pos, (*nr)++);
380 n_params = get_method_n_params(mtp);
381 n_ress = get_method_n_ress(mtp);
383 /* Create the new type for our clone. It must have one parameter
384 less then the original.*/
385 new_mtp = new_type_method(tp_name, n_params - 1, n_ress);
387 /* We must set the type of the methods parameters.*/
388 for (i = j = 0; i < n_params; ++i) {
390 /* This is the position of the argument, that we have
394 tp = get_method_param_type(mtp, i);
395 set_method_param_type(new_mtp, j++, tp);
397 /* Copy the methods result types. */
398 for (i = 0; i < n_ress; ++i) {
399 tp = get_method_res_type(mtp, i);
400 set_method_res_type(new_mtp, i, tp);
402 set_entity_type(ent, new_mtp);
406 * Make a clone of a method.
408 * @param q Contains information for the method to clone.
410 static ir_entity *clone_method(quadruple_t *q) {
411 ir_entity *new_entity;
415 /* A counter for the clones.*/
416 static unsigned nr = 0;
418 /* We get a new ident for our clone method.*/
419 clone_ident = get_clone_ident(get_entity_ident(q->ent), q->pos, nr);
420 /* We get our entity for the clone method. */
421 new_entity = copy_entity_name(q->ent, clone_ident);
423 /* a cloned entity is always local */
424 set_entity_visibility(new_entity, visibility_local);
426 /* set a ld name here: Should we mangle this ? */
427 set_entity_ld_ident(new_entity, get_entity_ident(new_entity));
429 /* set a new type here. */
430 change_entity_type(q, new_entity, &nr);
432 /* We need now a new ir_graph for our clone method. */
433 create_clone_proc_irg(new_entity, q);
435 /* We must set the atomic value of our "new_entity". */
436 sym.entity_p = new_entity;
437 rem = current_ir_graph;
438 current_ir_graph = get_const_code_irg();
439 new_entity->value = new_SymConst(mode_P_code, sym, symconst_addr_ent);
440 current_ir_graph = rem;
442 /* The "new_entity" don't have this information. */
443 new_entity->attr.mtd_attr.param_access = NULL;
444 new_entity->attr.mtd_attr.param_weight = NULL;
450 * Creates a new "cloned" Call node and return it.
452 * @param call The call that must be cloned.
453 * @param new_entity The entity of the cloned function.
454 * @param pos The position of the replaced parameter of this call.
456 static ir_node *new_cl_Call(ir_node *call, ir_entity *new_entity, int pos) {
459 int i, n_params, new_params = 0;
462 ir_graph *irg = get_irn_irg(call);
463 ir_node *bl = get_nodes_block(call);
465 sym.entity_p = new_entity;
466 callee = new_r_SymConst(irg, bl, mode_P_code, sym, symconst_addr_ent);
468 mtp = get_entity_type(new_entity);
469 n_params = get_Call_n_params(call);
470 NEW_ARR_A(ir_node *, in, n_params - 1);
472 /* we save the parameters of the new call in the array "in" without the
473 * parameter in position "pos", that is replaced with a constant.*/
474 for (i = 0; i < n_params; i++){
476 in[new_params++] = get_Call_param(call, i);
478 /* Create and return the new Call. */
479 return new_r_Call(irg, bl, get_Call_mem(call),
480 callee, n_params - 1, in, get_entity_type(new_entity));
484 * Exchange all Calls stored in the quadruplet to Calls of the cloned entity.
486 * @param q The quadruple
487 * @param cloned_ent The entity of the new function that must be called
490 static void exchange_calls(quadruple_t *q, ir_entity *cloned_ent) {
492 ir_node *new_call, *call;
495 /* We iterate the list of the "call".*/
496 for (i = 0; i < ARR_LEN(q->calls); ++i) {
499 /* A clone exist and the copy of "call" in this
500 * clone graph must be exchanged with new one.*/
501 new_call = new_cl_Call(call, cloned_ent, pos);
502 exchange(call, new_call);
507 * The weight formula:
508 * We save one instruction in every caller and param_weight instructions
511 static float calculate_weight(const entry_t *entry) {
512 return ARR_LEN(entry->q.calls) *
513 (float)(get_method_param_weight(entry->q.ent, entry->q.pos) + 1);
517 * After we exchanged all calls, some entries on the list for
518 * the next cloned entity may get invalid, so we have to check
519 * them and may even update the list of heavy uses.
521 static void reorder_weights(q_set *hmap, float threshold) {
522 entry_t **adr, *p, *entry;
527 entry = hmap->heavy_uses;
531 len = ARR_LEN(entry->q.calls);
532 for (i = 0; i < len; ++i) {
533 ir_node *ptr, *call = entry->q.calls[i];
535 /* might be exchanged, so skip Id nodes here. */
536 call = skip_Id(call);
538 /* we know, that a SymConst is here */
539 ptr = get_Call_ptr(call);
540 assert(is_SymConst(ptr));
542 callee = get_SymConst_entity(ptr);
543 if (callee != entry->q.ent) {
545 * This call is already changed because of a previous
546 * optimization. Remove it from the list.
549 entry->q.calls[i] = entry->q.calls[len];
550 entry->q.calls[len] = NULL;
552 /* the new call should be processed */
553 process_call(call, callee, hmap);
558 /* the length might be changed */
559 ARR_SHRINKLEN(entry->q.calls, len);
561 /* recalculate the weight and resort the heavy uses map */
562 entry->weight = calculate_weight(entry);
564 if (len <= 0 || entry->weight < threshold) {
565 hmap->heavy_uses = entry->next;
568 /* we have changed the list, check the next one */
573 for (p = entry->next; p && entry->weight < p->weight; p = p->next) {
578 hmap->heavy_uses = entry->next;
581 entry = hmap->heavy_uses;
583 /* we have changed the list, check the next one */
589 * Do the procedure cloning. Evaluate a heuristic weight for every
590 * call(..., Const, ...). If the weight is bigger than threshold,
591 * clone the entity and fix the calls.
593 void proc_cloning(float threshold) {
594 entry_t *entry = NULL, *p;
599 obstack_init(&hmap.obst);
601 hmap.heavy_uses = NULL;
603 /* initially fill our map by visiting all irgs */
604 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
605 irg = get_irp_irg(i);
606 irg_walk_graph(irg, collect_irg_calls, NULL, &hmap);
609 /* We have the "Call" nodes to optimize in set "set_entries". Our algorithm
610 replace one constant parameter and make a new "Call" node for all found "Calls". It exchange the
611 old one with the new one and the algorithm is called with the new "Call".
613 while (hmap.map || hmap.heavy_uses) {
614 /* We iterate the set and arrange the element of the set in a list.
615 The elements are arranged dependent of their value descending.*/
617 foreach_pset(hmap.map, entry) {
618 entry->weight = calculate_weight(entry);
621 * Do not put entry with a weight < threshold in the list
623 if (entry->weight < threshold) {
628 /* put entry in the heavy uses list */
630 if (! hmap.heavy_uses)
631 hmap.heavy_uses = entry;
633 if (entry->weight >= hmap.heavy_uses->weight) {
634 entry->next = hmap.heavy_uses;
635 hmap.heavy_uses = entry;
637 for (p = hmap.heavy_uses; p->next; p = p->next) {
638 if (entry->weight >= p->next->weight) {
639 entry->next = p->next;
653 /* Print some information about the list. */
654 printf("-----------------\n");
655 for (entry = hmap.heavy_uses; entry; entry = entry->next) {
656 printf("\nweight: is %f\n", entry->weight);
657 ir_printf("Call for Method %E\n", entry->q.ent);
658 printf("Position %i\n", entry->q.pos);
659 ir_printf("Value %T\n", entry->q.tv);
662 entry = hmap.heavy_uses;
664 ir_entity *ent = clone_method(&entry->q);
666 hmap.heavy_uses = entry->next;
668 /* We must exchange the copies of this call in all clones too.*/
669 exchange_calls(&entry->q, ent);
673 * after we exchanged all calls, some entries on the list for
674 * the next cloned entity may get invalid, so we have to check
675 * them and may even update the list of heavy uses.
677 reorder_weights(&hmap, threshold);
680 obstack_free(&hmap.obst, NULL);