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
26 * The purpose is first to find and analyze functions, that are called
27 * with constant parameter(s).
28 * The second step is to optimize the function that are found from our
29 * analyze. Optimize mean to make a new function with parameters, that
30 * aren't be constant. The constant parameters of the function are placed
31 * in the function graph. They aren't be passed as parameters.
37 #include "iroptimize.h"
43 #include "analyze_irg_args.h"
54 * This struct contains the information quadruple for a Call, which we need to
55 * decide if this function must be cloned.
57 typedef struct quadruple {
58 ir_entity *ent; /**< The entity of our Call. */
59 size_t pos; /**< Position of a constant argument of our Call. */
60 ir_tarval *tv; /**< The tarval of this argument if Const node. */
61 ir_node **calls; /**< The list of all calls with the same characteristics */
65 * The quadruplets are hold in a sorted list
67 typedef struct entry {
68 quadruple_t q; /**< the quadruple */
69 float weight; /**< its weight */
70 struct entry *next; /**< link to the next one */
73 typedef struct q_set {
74 struct obstack obst; /**< an obstack containing all entries */
75 pset *map; /**< a hash map containing the quadruples */
76 entry_t *heavy_uses; /**< the ordered list of heavy uses */
80 * Compare two quadruplets.
82 * @return zero if they are identically, non-zero else
84 static int entry_cmp(const void *elt, const void *key)
86 const entry_t *e1 = (const entry_t*)elt;
87 const entry_t *e2 = (const entry_t*)key;
89 return (e1->q.ent != e2->q.ent) || (e1->q.pos != e2->q.pos) || (e1->q.tv != e2->q.tv);
93 * Hash an element of type entry_t.
95 * @param entry The element to be hashed.
97 static unsigned hash_entry(const entry_t *entry)
99 return hash_ptr(entry->q.ent) ^ hash_ptr(entry->q.tv) ^ (unsigned)(entry->q.pos * 9);
103 * Free memory associated with a quadruplet.
105 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);
129 * Beware: we cannot clone variadic parameters as well as the
130 * last non-variadic one, which might be needed for the va_start()
134 /* In this for loop we collect the calls, that have
135 an constant parameter. */
136 for (i = n_params; i > 0;) {
137 call_param = get_Call_param(call, --i);
138 if (is_Const(call_param)) {
139 /* we have found a Call to collect and we save the informations,
142 hmap->map = new_pset(entry_cmp, 8);
144 key = OALLOC(&hmap->obst, entry_t);
148 key->q.tv = get_Const_tarval(call_param);
153 /* We insert our information in the set, where we collect the calls.*/
154 entry = (entry_t*)pset_insert(hmap->map, key, hash_entry(key));
157 obstack_free(&hmap->obst, key);
159 /* add the call to the list */
160 if (! entry->q.calls) {
161 entry->q.calls = NEW_ARR_F(ir_node *, 1);
162 entry->q.calls[0] = call;
164 ARR_APP1(ir_node *, entry->q.calls, call);
170 * Collect all calls in a ir_graph to a set.
172 * @param call A ir_node to be checked.
173 * @param env The quadruple-set containing the calls with constant parameters
175 static void collect_irg_calls(ir_node *call, void *env)
177 q_set *hmap = (q_set*)env;
181 /* We collect just "Call" nodes */
183 call_ptr = get_Call_ptr(call);
185 if (! is_SymConst_addr_ent(call_ptr))
188 callee = get_SymConst_entity(call_ptr);
190 /* we don't know which function gets finally bound to a weak symbol */
191 if (get_entity_linkage(callee) & IR_LINKAGE_WEAK)
194 /* we can only clone calls to existing entities */
195 if (get_entity_irg(callee) == NULL)
198 process_call(call, callee, hmap);
203 * Make a name for a clone. The clone name is
204 * the name of the original method suffixed with "_cl_pos_nr".
205 * pos is the pos from our quadruplet and nr is a counter.
207 * @param id The ident of the cloned function.
208 * @param pos The "pos" from our quadruplet.
209 * @param nr A counter for the clones.
211 static ident *get_clone_ident(ident *id, size_t pos, size_t nr)
213 char clone_postfix[32];
215 ir_snprintf(clone_postfix, sizeof(clone_postfix), "_cl_%zu_%zu", pos, nr);
217 return id_mangle(id, new_id_from_str(clone_postfix));
221 * Pre-Walker: Copies blocks and nodes from the original method graph
222 * to the cloned graph. Fixes the argument projection numbers for
223 * all arguments behind the removed one.
225 * @param irn A node from the original method graph.
226 * @param env The clone graph.
228 static void copy_nodes(ir_node *irn, void *env)
230 ir_graph *clone_irg = (ir_graph*)env;
231 ir_node *arg = (ir_node*)get_irg_link(clone_irg);
232 ir_node *irg_args = get_Proj_pred(arg);
236 /* Copy all nodes except the arg. */
238 copy_irn_to_irg(irn, clone_irg);
240 irn_copy = (ir_node*)get_irn_link(irn);
242 /* Fix argument numbers */
243 if (is_Proj(irn) && get_Proj_pred(irn) == irg_args) {
244 proj_nr = get_Proj_proj(irn);
245 if (get_Proj_proj(arg) < proj_nr)
246 set_Proj_proj(irn_copy, proj_nr - 1);
251 * Post-walker: Set the predecessors of the copied nodes.
252 * The copied nodes are set as link of their original nodes. The links of
253 * "irn" predecessors are the predecessors of copied node.
255 static void set_preds(ir_node *irn, void *env)
257 ir_graph *clone_irg = (ir_graph*)env;
258 ir_node *arg = (ir_node*)get_irg_link(clone_irg);
263 /* Arg is the method argument, that we have replaced by a constant.*/
267 irn_copy = (ir_node*)get_irn_link(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), (ir_node*)get_irn_link(pred));
276 set_Block_cfgpred(irn_copy, i, (ir_node*)get_irn_link(pred));
279 /* First we set the block our copy if it is not a block.*/
280 set_nodes_block(irn_copy, (ir_node*)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, (ir_node*)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, (ir_node*)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, size_t pos)
303 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 (unsigned i = get_irn_n_outs(irg_args); i-- > 0; ) {
310 ir_node *proj = get_irn_out(irg_args, i);
311 if ((int)pos == get_Proj_proj(proj)) {
314 * More than one arg node found:
315 * We rely on the fact that 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, const quadruple_t *q)
337 ir_graph *method_irg, *clone_irg;
338 ir_node *arg, *const_arg;
340 method_irg = get_entity_irg(ent);
342 /* We create the skeleton of the clone irg.*/
343 clone_irg = new_ir_graph(ent, 0);
345 arg = get_irg_arg(get_entity_irg(q->ent), q->pos);
346 /* we will replace the argument in position "q->pos" by this constant. */
347 const_arg = new_r_Const(clone_irg, q->tv);
349 /* args copy in the cloned graph will be the const. */
350 set_irn_link(arg, const_arg);
352 /* Store the arg that will be replaced here, so we can easily detect it. */
353 set_irg_link(clone_irg, arg);
355 /* We copy the blocks and nodes, that must be in
356 the clone graph and set their predecessors. */
357 irg_walk_graph(method_irg, copy_nodes, set_preds, clone_irg);
359 /* The "cloned" graph must be matured. */
360 mature_immBlock(get_irg_end_block(clone_irg));
361 irg_finalize_cons(clone_irg);
365 * The function create a new entity type
366 * for our clone and set it to clone entity.
368 * @param q Contains information for the method to clone.
369 * @param ent The entity of the clone.
370 * @param nr A pointer to the counter of clones.
372 static void change_entity_type(const quadruple_t *q, ir_entity *ent)
374 ir_type *mtp, *new_mtp, *tp;
375 size_t i, j, n_params, n_ress;
377 mtp = get_entity_type(q->ent);
378 n_params = get_method_n_params(mtp);
379 n_ress = get_method_n_ress(mtp);
381 /* Create the new type for our clone. It must have one parameter
382 less then the original.*/
383 new_mtp = new_type_method(n_params - 1, n_ress);
385 /* We must set the type of the methods parameters.*/
386 for (i = j = 0; i < n_params; ++i) {
388 /* This is the position of the argument, that we have
392 tp = get_method_param_type(mtp, i);
393 set_method_param_type(new_mtp, j++, tp);
395 /* Copy the methods result types. */
396 for (i = 0; i < n_ress; ++i) {
397 tp = get_method_res_type(mtp, i);
398 set_method_res_type(new_mtp, i, tp);
400 set_entity_type(ent, new_mtp);
404 * Make a clone of a method.
406 * @param q Contains information for the method to clone.
408 static ir_entity *clone_method(const quadruple_t *q)
410 ir_entity *new_entity;
412 /* A counter for the clones.*/
413 static size_t nr = 0;
415 /* We get a new ident for our clone method.*/
416 clone_ident = get_clone_ident(get_entity_ident(q->ent), q->pos, nr);
417 /* We get our entity for the clone method. */
418 new_entity = copy_entity_name(q->ent, clone_ident);
420 /* a cloned entity is always local */
421 set_entity_visibility(new_entity, ir_visibility_local);
423 /* set a ld name here: Should we mangle this ? */
424 set_entity_ld_ident(new_entity, get_entity_ident(new_entity));
426 /* set a new type here. */
427 change_entity_type(q, new_entity);
429 /* We need now a new ir_graph for our clone method. */
430 create_clone_proc_irg(new_entity, q);
432 /* The "new_entity" don't have this information. */
433 new_entity->attr.mtd_attr.param_access = NULL;
434 new_entity->attr.mtd_attr.param_weight = NULL;
440 * Creates a new "cloned" Call node and return it.
442 * @param call The call that must be cloned.
443 * @param new_entity The entity of the cloned function.
444 * @param pos The position of the replaced parameter of this call.
446 static ir_node *new_cl_Call(ir_node *call, ir_entity *new_entity, size_t pos)
449 size_t i, n_params, new_params = 0;
452 ir_graph *irg = get_irn_irg(call);
453 ir_node *bl = get_nodes_block(call);
455 sym.entity_p = new_entity;
456 callee = new_r_SymConst(irg, mode_P_code, sym, symconst_addr_ent);
458 n_params = get_Call_n_params(call);
459 NEW_ARR_A(ir_node *, in, n_params - 1);
461 /* we save the parameters of the new call in the array "in" without the
462 * parameter in position "pos", that is replaced with a constant.*/
463 for (i = 0; i < n_params; ++i) {
465 in[new_params++] = get_Call_param(call, i);
467 /* Create and return the new Call. */
468 return new_r_Call(bl, get_Call_mem(call),
469 callee, n_params - 1, in, get_entity_type(new_entity));
473 * Exchange all Calls stored in the quadruplet to Calls of the cloned entity.
475 * @param q The quadruple
476 * @param cloned_ent The entity of the new function that must be called
479 static void exchange_calls(quadruple_t *q, ir_entity *cloned_ent)
482 ir_node *new_call, *call;
485 /* We iterate the list of the "call".*/
486 for (i = 0; i < ARR_LEN(q->calls); ++i) {
489 /* A clone exist and the copy of "call" in this
490 * clone graph must be exchanged with new one.*/
491 new_call = new_cl_Call(call, cloned_ent, pos);
492 exchange(call, new_call);
497 * The weight formula:
498 * We save one instruction in every caller and param_weight instructions
501 static float calculate_weight(const entry_t *entry)
503 return ARR_LEN(entry->q.calls) *
504 (float)(get_method_param_weight(entry->q.ent, entry->q.pos) + 1);
508 * After we exchanged all calls, some entries on the list for
509 * the next cloned entity may get invalid, so we have to check
510 * them and may even update the list of heavy uses.
512 static void reorder_weights(q_set *hmap, float threshold)
514 entry_t **adr, *p, *entry;
519 entry = hmap->heavy_uses;
523 len = ARR_LEN(entry->q.calls);
524 for (i = 0; i < len; ++i) {
525 ir_node *ptr, *call = entry->q.calls[i];
527 /* might be exchanged, so skip Id nodes here. */
528 call = skip_Id(call);
530 /* we know, that a SymConst is here */
531 ptr = get_Call_ptr(call);
532 assert(is_SymConst(ptr));
534 callee = get_SymConst_entity(ptr);
535 if (callee != entry->q.ent) {
537 * This call is already changed because of a previous
538 * optimization. Remove it from the list.
541 entry->q.calls[i] = entry->q.calls[len];
542 entry->q.calls[len] = NULL;
544 /* the new call should be processed */
545 process_call(call, callee, hmap);
550 /* the length might be changed */
551 ARR_SHRINKLEN(entry->q.calls, len);
553 /* recalculate the weight and resort the heavy uses map */
554 entry->weight = calculate_weight(entry);
556 if (len <= 0 || entry->weight < threshold) {
557 hmap->heavy_uses = entry->next;
560 /* we have changed the list, check the next one */
565 for (p = entry->next; p && entry->weight < p->weight; p = p->next) {
570 hmap->heavy_uses = entry->next;
573 entry = hmap->heavy_uses;
575 /* we have changed the list, check the next one */
581 * Do the procedure cloning. Evaluate a heuristic weight for every
582 * call(..., Const, ...). If the weight is bigger than threshold,
583 * clone the entity and fix the calls.
585 void proc_cloning(float threshold)
591 DEBUG_ONLY(firm_dbg_module_t *dbg;)
593 /* register a debug mask */
594 FIRM_DBG_REGISTER(dbg, "firm.opt.proc_cloning");
596 obstack_init(&hmap.obst);
598 hmap.heavy_uses = NULL;
600 /* initially fill our map by visiting all irgs */
601 for (i = 0, n = get_irp_n_irgs(); i < n; ++i) {
602 ir_graph *irg = get_irp_irg(i);
603 irg_walk_graph(irg, collect_irg_calls, NULL, &hmap);
606 /* We have the "Call" nodes to optimize in set "set_entries". Our algorithm
607 replace one constant parameter and make a new "Call" node for all found "Calls". It exchange the
608 old one with the new one and the algorithm is called with the new "Call".
610 while (hmap.map || hmap.heavy_uses) {
611 /* We iterate the set and arrange the element of the set in a list.
612 The elements are arranged dependent of their value descending.*/
614 foreach_pset(hmap.map, entry_t, entry) {
615 entry->weight = calculate_weight(entry);
618 * Do not put entry with a weight < threshold in the list
620 if (entry->weight < threshold) {
625 /* put entry in the heavy uses list */
627 if (! hmap.heavy_uses)
628 hmap.heavy_uses = entry;
630 if (entry->weight >= hmap.heavy_uses->weight) {
631 entry->next = hmap.heavy_uses;
632 hmap.heavy_uses = entry;
634 for (p = hmap.heavy_uses; p->next; p = p->next) {
635 if (entry->weight >= p->next->weight) {
636 entry->next = p->next;
651 /* Print some information about the list. */
652 DB((dbg, LEVEL_2, "-----------------\n"));
653 for (entry_t *entry = hmap.heavy_uses; entry; entry = entry->next) {
654 DB((dbg, LEVEL_2, "\nweight: is %f\n", entry->weight));
655 DB((dbg, LEVEL_2, "Call for Method %E\n", entry->q.ent));
656 DB((dbg, LEVEL_2, "Position %zu\n", entry->q.pos));
657 DB((dbg, LEVEL_2, "Value %T\n", entry->q.tv));
660 entry_t *const entry = hmap.heavy_uses;
662 quadruple_t *qp = &entry->q;
664 ir_entity *ent = clone_method(qp);
665 DB((dbg, LEVEL_1, "Cloned <%+F, %zu, %T> info %+F\n", qp->ent, qp->pos, qp->tv, ent));
667 hmap.heavy_uses = entry->next;
669 /* We must exchange the copies of this call in all clones too.*/
670 exchange_calls(&entry->q, ent);
674 * after we exchanged all calls, some entries on the list for
675 * the next cloned entity may get invalid, so we have to check
676 * them and may even update the list of heavy uses.
678 reorder_weights(&hmap, threshold);
681 obstack_free(&hmap.obst, NULL);
684 typedef struct pass_t {
690 * Wrapper to run proc_cloning() as an ir_prog pass.
692 static int proc_cloning_wrapper(ir_prog *irp, void *context)
694 pass_t *pass = (pass_t*)context;
697 proc_cloning(pass->threshold);
701 /* create a ir_prog pass */
702 ir_prog_pass_t *proc_cloning_pass(const char *name, float threshold)
704 pass_t *pass = XMALLOCZ(pass_t);
706 pass->threshold = threshold;
707 return def_prog_pass_constructor(
708 &pass->pass, name ? name : "cloning", proc_cloning_wrapper);