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.
38 #include "iroptimize.h"
44 #include "analyze_irg_args.h"
55 * This struct contains the information quadruple for a Call, which we need to
56 * decide if this function must be cloned.
58 typedef struct quadruple {
59 ir_entity *ent; /**< The entity of our Call. */
60 int pos; /**< Position of a constant argument of our Call. */
61 ir_tarval *tv; /**< The tarval of this argument if Const node. */
62 ir_node **calls; /**< The list of all calls with the same characteristics */
66 * The quadruplets are hold in a sorted list
68 typedef struct entry {
69 quadruple_t q; /**< the quadruple */
70 float weight; /**< its weight */
71 struct entry *next; /**< link to the next one */
74 typedef struct q_set {
75 struct obstack obst; /**< an obstack containing all entries */
76 pset *map; /**< a hash map containing the quadruples */
77 entry_t *heavy_uses; /**< the ordered list of heavy uses */
81 * Compare two quadruplets.
83 * @return zero if they are identically, non-zero else
85 static int entry_cmp(const void *elt, const void *key)
87 const entry_t *e1 = (const entry_t*)elt;
88 const entry_t *e2 = (const entry_t*)key;
90 return (e1->q.ent != e2->q.ent) || (e1->q.pos != e2->q.pos) || (e1->q.tv != e2->q.tv);
94 * Hash an element of type entry_t.
96 * @param entry The element to be hashed.
98 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)
108 if (entry->q.calls) {
109 DEL_ARR_F(entry->q.calls);
110 entry->q.calls = NULL;
115 * Process a call node.
117 * @param call A ir_node to be checked.
118 * @param callee The entity of the callee
119 * @param hmap The quadruple-set containing the calls with constant parameters
121 static void process_call(ir_node *call, ir_entity *callee, q_set *hmap)
124 entry_t *key, *entry;
128 n_params = get_Call_n_params(call);
130 /* Beware: we cannot clone variadic parameters as well as the
131 * last non-variadic one, which might be needed for the va_start()
134 mtp = get_Call_type(call);
135 if (get_method_variadicity(mtp) != variadicity_non_variadic) {
136 n_params = get_method_first_variadic_param_index(mtp) - 1;
139 /* In this for loop we collect the calls, that have
140 an constant parameter. */
141 for (i = n_params - 1; i >= 0; --i) {
142 call_param = get_Call_param(call, i);
143 if (is_Const(call_param)) {
144 /* we have found a Call to collect and we save the informations,
147 hmap->map = new_pset(entry_cmp, 8);
149 key = OALLOC(&hmap->obst, entry_t);
153 key->q.tv = get_Const_tarval(call_param);
158 /* We insert our information in the set, where we collect the calls.*/
159 entry = (entry_t*)pset_insert(hmap->map, key, hash_entry(key));
162 obstack_free(&hmap->obst, key);
164 /* add the call to the list */
165 if (! entry->q.calls) {
166 entry->q.calls = NEW_ARR_F(ir_node *, 1);
167 entry->q.calls[0] = call;
169 ARR_APP1(ir_node *, entry->q.calls, call);
175 * Collect all calls in a ir_graph to a set.
177 * @param call A ir_node to be checked.
178 * @param env The quadruple-set containing the calls with constant parameters
180 static void collect_irg_calls(ir_node *call, void *env)
182 q_set *hmap = (q_set*)env;
186 /* We collect just "Call" nodes */
188 call_ptr = get_Call_ptr(call);
190 if (! is_Global(call_ptr))
193 callee = get_Global_entity(call_ptr);
195 /* we don't know which function gets finally bound to a weak symbol */
196 if (get_entity_linkage(callee) & IR_LINKAGE_WEAK)
199 /* we can only clone calls to existing entities */
200 if (get_entity_irg(callee) == NULL)
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)
218 char clone_postfix[32];
220 snprintf(clone_postfix, sizeof(clone_postfix), "_cl_%d_%u", pos, nr);
222 return id_mangle(id, new_id_from_str(clone_postfix));
226 * Pre-Walker: Copies blocks and nodes from the original method graph
227 * to the cloned graph. Fixes the argument projection numbers for
228 * all arguments behind the removed one.
230 * @param irn A node from the original method graph.
231 * @param env The clone graph.
233 static void copy_nodes(ir_node *irn, void *env)
235 ir_graph *clone_irg = (ir_graph*)env;
236 ir_node *arg = (ir_node*)get_irg_link(clone_irg);
237 ir_node *irg_args = get_Proj_pred(arg);
241 /* Copy all nodes except the arg. */
243 copy_irn_to_irg(irn, clone_irg);
245 irn_copy = (ir_node*)get_irn_link(irn);
247 /* Fix argument numbers */
248 if (is_Proj(irn) && get_Proj_pred(irn) == irg_args) {
249 proj_nr = get_Proj_proj(irn);
250 if (get_Proj_proj(arg) < proj_nr)
251 set_Proj_proj(irn_copy, proj_nr - 1);
256 * Post-walker: Set the predecessors of the copied nodes.
257 * The copied nodes are set as link of their original nodes. The links of
258 * "irn" predecessors are the predecessors of copied node.
260 static void set_preds(ir_node *irn, void *env)
262 ir_graph *clone_irg = (ir_graph*)env;
263 ir_node *arg = (ir_node*)get_irg_link(clone_irg);
268 /* Arg is the method argument, that we have replaced by a constant.*/
272 irn_copy = (ir_node*)get_irn_link(irn);
275 for (i = get_Block_n_cfgpreds(irn) - 1; i >= 0; i--) {
276 pred = get_Block_cfgpred(irn, i);
277 /* "End" block must be handled extra, because it is not matured.*/
278 if (get_irg_end_block(current_ir_graph) == irn)
279 add_immBlock_pred(get_irg_end_block(clone_irg), (ir_node*)get_irn_link(pred));
281 set_Block_cfgpred(irn_copy, i, (ir_node*)get_irn_link(pred));
284 /* First we set the block our copy if it is not a block.*/
285 set_nodes_block(irn_copy, (ir_node*)get_irn_link(get_nodes_block(irn)));
287 /* Handle the keep-alives. This must be done separately, because
288 the End node was NOT copied */
289 for (i = 0; i < get_End_n_keepalives(irn); ++i)
290 add_End_keepalive(irn_copy, (ir_node*)get_irn_link(get_End_keepalive(irn, i)));
292 for (i = get_irn_arity(irn) - 1; i >= 0; i--) {
293 pred = get_irn_n(irn, i);
294 set_irn_n(irn_copy, i, (ir_node*)get_irn_link(pred));
301 * Get the method argument at the position "pos".
303 * @param irg irg that must be cloned.
304 * @param pos The position of the argument.
306 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, 0);
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)
343 ir_graph *method_irg, *clone_irg;
344 ir_node *arg, *const_arg;
346 method_irg = get_entity_irg(ent);
348 /* We create the skeleton of the clone irg.*/
349 clone_irg = new_ir_graph(ent, 0);
351 arg = get_irg_arg(get_entity_irg(q->ent), q->pos);
352 /* we will replace the argument in position "q->pos" by this constant. */
353 const_arg = new_r_Const(clone_irg, q->tv);
355 /* args copy in the cloned graph will be the const. */
356 set_irn_link(arg, const_arg);
358 /* Store the arg that will be replaced here, so we can easily detect it. */
359 set_irg_link(clone_irg, arg);
361 /* We copy the blocks and nodes, that must be in
362 the clone graph and set their predecessors. */
363 irg_walk_graph(method_irg, copy_nodes, set_preds, clone_irg);
365 /* The "cloned" graph must be matured. */
366 mature_immBlock(get_irg_end_block(clone_irg));
367 irg_finalize_cons(clone_irg);
371 * The function create a new entity type
372 * for our clone and set it to clone entity.
374 * @param q Contains information for the method to clone.
375 * @param ent The entity of the clone.
376 * @param nr A pointer to the counter of clones.
378 static void change_entity_type(quadruple_t *q, ir_entity *ent)
380 ir_type *mtp, *new_mtp, *tp;
381 int i, j, n_params, n_ress;
383 mtp = get_entity_type(q->ent);
384 n_params = get_method_n_params(mtp);
385 n_ress = get_method_n_ress(mtp);
387 /* Create the new type for our clone. It must have one parameter
388 less then the original.*/
389 new_mtp = new_type_method(n_params - 1, n_ress);
391 /* We must set the type of the methods parameters.*/
392 for (i = j = 0; i < n_params; ++i) {
394 /* This is the position of the argument, that we have
398 tp = get_method_param_type(mtp, i);
399 set_method_param_type(new_mtp, j++, tp);
401 /* Copy the methods result types. */
402 for (i = 0; i < n_ress; ++i) {
403 tp = get_method_res_type(mtp, i);
404 set_method_res_type(new_mtp, i, tp);
406 set_entity_type(ent, new_mtp);
410 * Make a clone of a method.
412 * @param q Contains information for the method to clone.
414 static ir_entity *clone_method(quadruple_t *q)
416 ir_entity *new_entity;
419 /* A counter for the clones.*/
420 static unsigned nr = 0;
422 /* We get a new ident for our clone method.*/
423 clone_ident = get_clone_ident(get_entity_ident(q->ent), q->pos, nr);
424 /* We get our entity for the clone method. */
425 new_entity = copy_entity_name(q->ent, clone_ident);
427 /* a cloned entity is always local */
428 set_entity_visibility(new_entity, ir_visibility_local);
430 /* set a ld name here: Should we mangle this ? */
431 set_entity_ld_ident(new_entity, get_entity_ident(new_entity));
433 /* set a new type here. */
434 change_entity_type(q, new_entity);
436 /* We need now a new ir_graph for our clone method. */
437 create_clone_proc_irg(new_entity, q);
439 /* We must set the atomic value of our "new_entity". */
440 sym.entity_p = new_entity;
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, mode_P_code, sym, symconst_addr_ent);
468 n_params = get_Call_n_params(call);
469 NEW_ARR_A(ir_node *, in, n_params - 1);
471 /* we save the parameters of the new call in the array "in" without the
472 * parameter in position "pos", that is replaced with a constant.*/
473 for (i = 0; i < n_params; i++){
475 in[new_params++] = get_Call_param(call, i);
477 /* Create and return the new Call. */
478 return new_r_Call(bl, get_Call_mem(call),
479 callee, n_params - 1, in, get_entity_type(new_entity));
483 * Exchange all Calls stored in the quadruplet to Calls of the cloned entity.
485 * @param q The quadruple
486 * @param cloned_ent The entity of the new function that must be called
489 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)
513 return ARR_LEN(entry->q.calls) *
514 (float)(get_method_param_weight(entry->q.ent, entry->q.pos) + 1);
518 * After we exchanged all calls, some entries on the list for
519 * the next cloned entity may get invalid, so we have to check
520 * them and may even update the list of heavy uses.
522 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(is_SymConst(ptr));
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)
597 entry_t *entry = NULL, *p;
602 obstack_init(&hmap.obst);
604 hmap.heavy_uses = NULL;
606 /* initially fill our map by visiting all irgs */
607 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
608 irg = get_irp_irg(i);
609 irg_walk_graph(irg, collect_irg_calls, NULL, &hmap);
612 /* We have the "Call" nodes to optimize in set "set_entries". Our algorithm
613 replace one constant parameter and make a new "Call" node for all found "Calls". It exchange the
614 old one with the new one and the algorithm is called with the new "Call".
616 while (hmap.map || hmap.heavy_uses) {
617 /* We iterate the set and arrange the element of the set in a list.
618 The elements are arranged dependent of their value descending.*/
620 foreach_pset(hmap.map, entry_t*, entry) {
621 entry->weight = calculate_weight(entry);
624 * Do not put entry with a weight < threshold in the list
626 if (entry->weight < threshold) {
631 /* put entry in the heavy uses list */
633 if (! hmap.heavy_uses)
634 hmap.heavy_uses = entry;
636 if (entry->weight >= hmap.heavy_uses->weight) {
637 entry->next = hmap.heavy_uses;
638 hmap.heavy_uses = entry;
640 for (p = hmap.heavy_uses; p->next; p = p->next) {
641 if (entry->weight >= p->next->weight) {
642 entry->next = p->next;
656 /* Print some information about the list. */
657 printf("-----------------\n");
658 for (entry = hmap.heavy_uses; entry; entry = entry->next) {
659 printf("\nweight: is %f\n", entry->weight);
660 ir_printf("Call for Method %E\n", entry->q.ent);
661 printf("Position %i\n", entry->q.pos);
662 ir_printf("Value %T\n", entry->q.tv);
665 entry = hmap.heavy_uses;
667 ir_entity *ent = clone_method(&entry->q);
669 hmap.heavy_uses = entry->next;
671 /* We must exchange the copies of this call in all clones too.*/
672 exchange_calls(&entry->q, ent);
676 * after we exchanged all calls, some entries on the list for
677 * the next cloned entity may get invalid, so we have to check
678 * them and may even update the list of heavy uses.
680 reorder_weights(&hmap, threshold);
683 obstack_free(&hmap.obst, NULL);
686 typedef struct pass_t {
692 * Wrapper to run proc_cloning() as an ir_prog pass.
694 static int proc_cloning_wrapper(ir_prog *irp, void *context)
696 pass_t *pass = (pass_t*)context;
699 proc_cloning(pass->threshold);
703 /* create a ir_prog pass */
704 ir_prog_pass_t *proc_cloning_pass(const char *name, float threshold)
706 pass_t *pass = XMALLOCZ(pass_t);
708 pass->threshold = threshold;
709 return def_prog_pass_constructor(
710 &pass->pass, name ? name : "cloning", proc_cloning_wrapper);