2 * Copyright (C) 1995-2011 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 size_t 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 unsigned hash_entry(const entry_t *entry)
100 return HASH_PTR(entry->q.ent) ^ HASH_PTR(entry->q.tv) ^ (unsigned)(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);
136 /* In this for loop we collect the calls, that have
137 an constant parameter. */
138 for (i = n_params; i > 0;) {
139 call_param = get_Call_param(call, --i);
140 if (is_Const(call_param)) {
141 /* we have found a Call to collect and we save the informations,
144 hmap->map = new_pset(entry_cmp, 8);
146 key = OALLOC(&hmap->obst, entry_t);
150 key->q.tv = get_Const_tarval(call_param);
155 /* We insert our information in the set, where we collect the calls.*/
156 entry = (entry_t*)pset_insert(hmap->map, key, hash_entry(key));
159 obstack_free(&hmap->obst, key);
161 /* add the call to the list */
162 if (! entry->q.calls) {
163 entry->q.calls = NEW_ARR_F(ir_node *, 1);
164 entry->q.calls[0] = call;
166 ARR_APP1(ir_node *, entry->q.calls, call);
172 * Collect all calls in a ir_graph to a set.
174 * @param call A ir_node to be checked.
175 * @param env The quadruple-set containing the calls with constant parameters
177 static void collect_irg_calls(ir_node *call, void *env)
179 q_set *hmap = (q_set*)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 don't know which function gets finally bound to a weak symbol */
193 if (get_entity_linkage(callee) & IR_LINKAGE_WEAK)
196 /* we can only clone calls to existing entities */
197 if (get_entity_irg(callee) == NULL)
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, size_t pos, size_t nr)
215 char clone_postfix[32];
217 ir_snprintf(clone_postfix, sizeof(clone_postfix), "_cl_%zu_%zu", pos, nr);
219 return id_mangle(id, new_id_from_str(clone_postfix));
223 * Pre-Walker: Copies blocks and nodes from the original method graph
224 * to the cloned graph. Fixes the argument projection numbers for
225 * all arguments behind the removed one.
227 * @param irn A node from the original method graph.
228 * @param env The clone graph.
230 static void copy_nodes(ir_node *irn, void *env)
232 ir_graph *clone_irg = (ir_graph*)env;
233 ir_node *arg = (ir_node*)get_irg_link(clone_irg);
234 ir_node *irg_args = get_Proj_pred(arg);
238 /* Copy all nodes except the arg. */
240 copy_irn_to_irg(irn, clone_irg);
242 irn_copy = (ir_node*)get_irn_link(irn);
244 /* Fix argument numbers */
245 if (is_Proj(irn) && get_Proj_pred(irn) == irg_args) {
246 proj_nr = get_Proj_proj(irn);
247 if (get_Proj_proj(arg) < proj_nr)
248 set_Proj_proj(irn_copy, proj_nr - 1);
253 * Post-walker: Set the predecessors of the copied nodes.
254 * The copied nodes are set as link of their original nodes. The links of
255 * "irn" predecessors are the predecessors of copied node.
257 static void set_preds(ir_node *irn, void *env)
259 ir_graph *clone_irg = (ir_graph*)env;
260 ir_node *arg = (ir_node*)get_irg_link(clone_irg);
265 /* Arg is the method argument, that we have replaced by a constant.*/
269 irn_copy = (ir_node*)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), (ir_node*)get_irn_link(pred));
278 set_Block_cfgpred(irn_copy, i, (ir_node*)get_irn_link(pred));
281 /* First we set the block our copy if it is not a block.*/
282 set_nodes_block(irn_copy, (ir_node*)get_irn_link(get_nodes_block(irn)));
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, (ir_node*)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, (ir_node*)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, size_t pos)
305 ir_node *irg_args = get_irg_args(irg), *arg = NULL;
308 /* Call algorithm that computes the out edges */
309 assure_irg_outs(irg);
311 /* Search the argument with the number pos.*/
312 for (i = get_irn_n_outs(irg_args) - 1; i >= 0; --i) {
313 ir_node *proj = get_irn_out(irg_args, i);
314 if (pos == get_Proj_proj(proj)) {
317 * More than one arg node found:
318 * We rely on the fact that only one arg exists, so do
319 * a cheap CSE in this case.
321 set_irn_out(irg_args, i, arg, 0);
327 assert(arg && "Argument not found");
332 * Create a new graph for the clone of the method,
333 * that we want to clone.
335 * @param ent The entity of the method that must be cloned.
336 * @param q Our quadruplet.
338 static void create_clone_proc_irg(ir_entity *ent, const quadruple_t *q)
340 ir_graph *method_irg, *clone_irg;
341 ir_node *arg, *const_arg;
343 method_irg = get_entity_irg(ent);
345 /* We create the skeleton of the clone irg.*/
346 clone_irg = new_ir_graph(ent, 0);
348 arg = get_irg_arg(get_entity_irg(q->ent), q->pos);
349 /* we will replace the argument in position "q->pos" by this constant. */
350 const_arg = new_r_Const(clone_irg, q->tv);
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(const quadruple_t *q, ir_entity *ent)
377 ir_type *mtp, *new_mtp, *tp;
378 size_t i, j, n_params, n_ress;
380 mtp = get_entity_type(q->ent);
381 n_params = get_method_n_params(mtp);
382 n_ress = get_method_n_ress(mtp);
384 /* Create the new type for our clone. It must have one parameter
385 less then the original.*/
386 new_mtp = new_type_method(n_params - 1, n_ress);
388 /* We must set the type of the methods parameters.*/
389 for (i = j = 0; i < n_params; ++i) {
391 /* This is the position of the argument, that we have
395 tp = get_method_param_type(mtp, i);
396 set_method_param_type(new_mtp, j++, tp);
398 /* Copy the methods result types. */
399 for (i = 0; i < n_ress; ++i) {
400 tp = get_method_res_type(mtp, i);
401 set_method_res_type(new_mtp, i, tp);
403 set_entity_type(ent, new_mtp);
407 * Make a clone of a method.
409 * @param q Contains information for the method to clone.
411 static ir_entity *clone_method(const quadruple_t *q)
413 ir_entity *new_entity;
416 /* A counter for the clones.*/
417 static size_t nr = 0;
419 /* We get a new ident for our clone method.*/
420 clone_ident = get_clone_ident(get_entity_ident(q->ent), q->pos, nr);
421 /* We get our entity for the clone method. */
422 new_entity = copy_entity_name(q->ent, clone_ident);
424 /* a cloned entity is always local */
425 set_entity_visibility(new_entity, ir_visibility_local);
427 /* set a ld name here: Should we mangle this ? */
428 set_entity_ld_ident(new_entity, get_entity_ident(new_entity));
430 /* set a new type here. */
431 change_entity_type(q, new_entity);
433 /* We need now a new ir_graph for our clone method. */
434 create_clone_proc_irg(new_entity, q);
436 /* We must set the atomic value of our "new_entity". */
437 sym.entity_p = new_entity;
439 /* The "new_entity" don't have this information. */
440 new_entity->attr.mtd_attr.param_access = NULL;
441 new_entity->attr.mtd_attr.param_weight = NULL;
447 * Creates a new "cloned" Call node and return it.
449 * @param call The call that must be cloned.
450 * @param new_entity The entity of the cloned function.
451 * @param pos The position of the replaced parameter of this call.
453 static ir_node *new_cl_Call(ir_node *call, ir_entity *new_entity, size_t pos)
456 size_t i, n_params, new_params = 0;
459 ir_graph *irg = get_irn_irg(call);
460 ir_node *bl = get_nodes_block(call);
462 sym.entity_p = new_entity;
463 callee = new_r_SymConst(irg, mode_P_code, sym, symconst_addr_ent);
465 n_params = get_Call_n_params(call);
466 NEW_ARR_A(ir_node *, in, n_params - 1);
468 /* we save the parameters of the new call in the array "in" without the
469 * parameter in position "pos", that is replaced with a constant.*/
470 for (i = 0; i < n_params; ++i) {
472 in[new_params++] = get_Call_param(call, i);
474 /* Create and return the new Call. */
475 return new_r_Call(bl, get_Call_mem(call),
476 callee, n_params - 1, in, get_entity_type(new_entity));
480 * Exchange all Calls stored in the quadruplet to Calls of the cloned entity.
482 * @param q The quadruple
483 * @param cloned_ent The entity of the new function that must be called
486 static void exchange_calls(quadruple_t *q, ir_entity *cloned_ent)
489 ir_node *new_call, *call;
492 /* We iterate the list of the "call".*/
493 for (i = 0; i < ARR_LEN(q->calls); ++i) {
496 /* A clone exist and the copy of "call" in this
497 * clone graph must be exchanged with new one.*/
498 new_call = new_cl_Call(call, cloned_ent, pos);
499 exchange(call, new_call);
504 * The weight formula:
505 * We save one instruction in every caller and param_weight instructions
508 static float calculate_weight(const entry_t *entry)
510 return ARR_LEN(entry->q.calls) *
511 (float)(get_method_param_weight(entry->q.ent, entry->q.pos) + 1);
515 * After we exchanged all calls, some entries on the list for
516 * the next cloned entity may get invalid, so we have to check
517 * them and may even update the list of heavy uses.
519 static void reorder_weights(q_set *hmap, float threshold)
521 entry_t **adr, *p, *entry;
526 entry = hmap->heavy_uses;
530 len = ARR_LEN(entry->q.calls);
531 for (i = 0; i < len; ++i) {
532 ir_node *ptr, *call = entry->q.calls[i];
534 /* might be exchanged, so skip Id nodes here. */
535 call = skip_Id(call);
537 /* we know, that a SymConst is here */
538 ptr = get_Call_ptr(call);
539 assert(is_SymConst(ptr));
541 callee = get_SymConst_entity(ptr);
542 if (callee != entry->q.ent) {
544 * This call is already changed because of a previous
545 * optimization. Remove it from the list.
548 entry->q.calls[i] = entry->q.calls[len];
549 entry->q.calls[len] = NULL;
551 /* the new call should be processed */
552 process_call(call, callee, hmap);
557 /* the length might be changed */
558 ARR_SHRINKLEN(entry->q.calls, len);
560 /* recalculate the weight and resort the heavy uses map */
561 entry->weight = calculate_weight(entry);
563 if (len <= 0 || entry->weight < threshold) {
564 hmap->heavy_uses = entry->next;
567 /* we have changed the list, check the next one */
572 for (p = entry->next; p && entry->weight < p->weight; p = p->next) {
577 hmap->heavy_uses = entry->next;
580 entry = hmap->heavy_uses;
582 /* we have changed the list, check the next one */
588 * Do the procedure cloning. Evaluate a heuristic weight for every
589 * call(..., Const, ...). If the weight is bigger than threshold,
590 * clone the entity and fix the calls.
592 void proc_cloning(float threshold)
594 entry_t *entry = NULL, *p;
598 DEBUG_ONLY(firm_dbg_module_t *dbg;)
600 /* register a debug mask */
601 FIRM_DBG_REGISTER(dbg, "firm.opt.proc_cloning");
603 obstack_init(&hmap.obst);
605 hmap.heavy_uses = NULL;
607 /* initially fill our map by visiting all irgs */
608 for (i = 0, n = get_irp_n_irgs(); i < n; ++i) {
609 ir_graph *irg = get_irp_irg(i);
610 irg_walk_graph(irg, collect_irg_calls, NULL, &hmap);
613 /* We have the "Call" nodes to optimize in set "set_entries". Our algorithm
614 replace one constant parameter and make a new "Call" node for all found "Calls". It exchange the
615 old one with the new one and the algorithm is called with the new "Call".
617 while (hmap.map || hmap.heavy_uses) {
618 /* We iterate the set and arrange the element of the set in a list.
619 The elements are arranged dependent of their value descending.*/
621 foreach_pset(hmap.map, entry_t*, entry) {
622 entry->weight = calculate_weight(entry);
625 * Do not put entry with a weight < threshold in the list
627 if (entry->weight < threshold) {
632 /* put entry in the heavy uses list */
634 if (! hmap.heavy_uses)
635 hmap.heavy_uses = entry;
637 if (entry->weight >= hmap.heavy_uses->weight) {
638 entry->next = hmap.heavy_uses;
639 hmap.heavy_uses = entry;
641 for (p = hmap.heavy_uses; p->next; p = p->next) {
642 if (entry->weight >= p->next->weight) {
643 entry->next = p->next;
658 /* Print some information about the list. */
659 DB((dbg, LEVEL_2, "-----------------\n"));
660 for (entry = hmap.heavy_uses; entry; entry = entry->next) {
661 DB((dbg, LEVEL_2, "\nweight: is %f\n", entry->weight));
662 DB((dbg, LEVEL_2, "Call for Method %E\n", entry->q.ent));
663 DB((dbg, LEVEL_2, "Position %zu\n", entry->q.pos));
664 DB((dbg, LEVEL_2, "Value %T\n", entry->q.tv));
667 entry = hmap.heavy_uses;
669 quadruple_t *qp = &entry->q;
671 ir_entity *ent = clone_method(qp);
672 DB((dbg, LEVEL_1, "Cloned <%+F, %zu, %T> info %+F\n", qp->ent, qp->pos, qp->tv, ent));
674 hmap.heavy_uses = entry->next;
676 /* We must exchange the copies of this call in all clones too.*/
677 exchange_calls(&entry->q, ent);
681 * after we exchanged all calls, some entries on the list for
682 * the next cloned entity may get invalid, so we have to check
683 * them and may even update the list of heavy uses.
685 reorder_weights(&hmap, threshold);
688 obstack_free(&hmap.obst, NULL);
691 typedef struct pass_t {
697 * Wrapper to run proc_cloning() as an ir_prog pass.
699 static int proc_cloning_wrapper(ir_prog *irp, void *context)
701 pass_t *pass = (pass_t*)context;
704 proc_cloning(pass->threshold);
708 /* create a ir_prog pass */
709 ir_prog_pass_t *proc_cloning_pass(const char *name, float threshold)
711 pass_t *pass = XMALLOCZ(pass_t);
713 pass->threshold = threshold;
714 return def_prog_pass_constructor(
715 &pass->pass, name ? name : "cloning", proc_cloning_wrapper);