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"
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 if (! is_Global(call_ptr))
193 callee = get_Global_entity(call_ptr);
195 /* we can only clone calls to existing entities */
196 if (get_entity_visibility(callee) == visibility_external_allocated)
199 process_call(call, callee, hmap);
204 * Make a name for a clone. The clone name is
205 * the name of the original method suffixed with "_cl_pos_nr".
206 * pos is the pos from our quadruplet and nr is a counter.
208 * @param id The ident of the cloned function.
209 * @param pos The "pos" from our quadruplet.
210 * @param nr A counter for the clones.
212 static ident *get_clone_ident(ident *id, int pos, unsigned nr) {
213 char clone_postfix[32];
215 snprintf(clone_postfix, sizeof(clone_postfix), "_cl_%d_%u", pos, nr);
217 return 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) {
229 ir_node *arg, *irg_args, *irn_copy;
231 ir_graph *clone_irg = env;
233 arg = get_irg_link(clone_irg);
234 irg_args = get_Proj_pred(arg);
236 /* Copy all nodes except the arg. */
238 copy_irn_to_irg(irn, clone_irg);
240 irn_copy = 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_node *irn_copy, *pred, *arg;
258 ir_graph *clone_irg = env;
260 arg = get_irg_link(clone_irg);
261 /* Arg is the method argument, that we have replaced by a constant.*/
265 irn_copy = get_irn_link(irn);
268 set_Block_MacroBlock(irn_copy, get_irn_link(get_Block_MacroBlock(irn)));
269 for (i = get_Block_n_cfgpreds(irn) - 1; i >= 0; i--) {
270 pred = get_Block_cfgpred(irn, i);
271 /* "End" block must be handled extra, because it is not matured.*/
272 if (get_irg_end_block(current_ir_graph) == irn)
273 add_immBlock_pred(get_irg_end_block(clone_irg), get_irn_link(pred));
275 set_Block_cfgpred(irn_copy, i, get_irn_link(pred));
278 /* First we set the block our copy if it is not a block.*/
279 set_nodes_block(irn_copy, get_irn_link(get_nodes_block(irn)));
280 if (get_irn_op(irn) == op_End) {
281 /* Handle the keep-alives. This must be done separately, because
282 the End node was NOT copied */
283 for (i = 0; i < get_End_n_keepalives(irn); ++i)
284 add_End_keepalive(irn_copy, get_irn_link(get_End_keepalive(irn, i)));
286 for (i = get_irn_arity(irn) - 1; i >= 0; i--) {
287 pred = get_irn_n(irn, i);
288 set_irn_n(irn_copy, i, get_irn_link(pred));
295 * Get the method argument at the position "pos".
297 * @param irg irg that must be cloned.
298 * @param pos The position of the argument.
300 static ir_node *get_irg_arg(ir_graph *irg, int pos) {
301 ir_node *irg_args = get_irg_args(irg), *arg = NULL;
304 /* Call algorithm that computes the out edges */
305 assure_irg_outs(irg);
307 /* Search the argument with the number pos.*/
308 for (i = get_irn_n_outs(irg_args) - 1; i >= 0; --i) {
309 ir_node *proj = get_irn_out(irg_args, i);
310 if (pos == get_Proj_proj(proj)) {
313 * More than one arg node found:
314 * We rely on the fact the only one arg exists, so do
315 * a cheap CSE in this case.
317 set_irn_out(irg_args, i, arg);
323 assert(arg && "Argument not found");
328 * Create a new graph for the clone of the method,
329 * that we want to clone.
331 * @param ent The entity of the method that must be cloned.
332 * @param q Our quadruplet.
334 static void create_clone_proc_irg(ir_entity *ent, quadruple_t *q) {
335 ir_graph *method_irg, *clone_irg;
336 ir_node *arg, *const_arg;
338 method_irg = get_entity_irg(ent);
340 /* We create the skeleton of the clone irg.*/
341 clone_irg = new_ir_graph(ent, 0);
343 arg = get_irg_arg(get_entity_irg(q->ent), q->pos);
344 /* we will replace the argument in position "q->pos" by this constant. */
345 const_arg = new_r_Const_type(
346 clone_irg, get_nodes_block(arg), get_irn_mode(arg), q->tv,
347 get_method_param_type(get_entity_type(q->ent), q->pos));
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(quadruple_t *q, ir_entity *ent, unsigned *nr) {
373 ir_type *mtp, *new_mtp, *tp;
375 int i, j, n_params, n_ress;
377 mtp = get_entity_type(q->ent);
378 tp_name = get_clone_ident(get_type_ident(mtp), q->pos, (*nr)++);
379 n_params = get_method_n_params(mtp);
380 n_ress = get_method_n_ress(mtp);
382 /* Create the new type for our clone. It must have one parameter
383 less then the original.*/
384 new_mtp = new_type_method(tp_name, n_params - 1, n_ress);
386 /* We must set the type of the methods parameters.*/
387 for (i = j = 0; i < n_params; ++i) {
389 /* This is the position of the argument, that we have
393 tp = get_method_param_type(mtp, i);
394 set_method_param_type(new_mtp, j++, tp);
396 /* Copy the methods result types. */
397 for (i = 0; i < n_ress; ++i) {
398 tp = get_method_res_type(mtp, i);
399 set_method_res_type(new_mtp, i, tp);
401 set_entity_type(ent, new_mtp);
405 * Make a clone of a method.
407 * @param q Contains information for the method to clone.
409 static ir_entity *clone_method(quadruple_t *q) {
410 ir_entity *new_entity;
414 /* A counter for the clones.*/
415 static unsigned nr = 0;
417 /* We get a new ident for our clone method.*/
418 clone_ident = get_clone_ident(get_entity_ident(q->ent), q->pos, nr);
419 /* We get our entity for the clone method. */
420 new_entity = copy_entity_name(q->ent, clone_ident);
422 /* a cloned entity is always local */
423 set_entity_visibility(new_entity, visibility_local);
425 /* set a ld name here: Should we mangle this ? */
426 set_entity_ld_ident(new_entity, get_entity_ident(new_entity));
428 /* set a new type here. */
429 change_entity_type(q, new_entity, &nr);
431 /* We need now a new ir_graph for our clone method. */
432 create_clone_proc_irg(new_entity, q);
434 /* We must set the atomic value of our "new_entity". */
435 sym.entity_p = new_entity;
436 rem = current_ir_graph;
437 current_ir_graph = get_const_code_irg();
438 new_entity->value = new_SymConst(mode_P_code, sym, symconst_addr_ent);
439 current_ir_graph = rem;
441 /* The "new_entity" don't have this information. */
442 new_entity->attr.mtd_attr.param_access = NULL;
443 new_entity->attr.mtd_attr.param_weight = NULL;
449 * Creates a new "cloned" Call node and return it.
451 * @param call The call that must be cloned.
452 * @param new_entity The entity of the cloned function.
453 * @param pos The position of the replaced parameter of this call.
455 static ir_node *new_cl_Call(ir_node *call, ir_entity *new_entity, int pos) {
458 int i, n_params, new_params = 0;
461 ir_graph *irg = get_irn_irg(call);
462 ir_node *bl = get_nodes_block(call);
464 sym.entity_p = new_entity;
465 callee = new_r_SymConst(irg, bl, mode_P_code, sym, symconst_addr_ent);
467 mtp = get_entity_type(new_entity);
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(irg, 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) {
491 ir_node *new_call, *call;
494 /* We iterate the list of the "call".*/
495 for (i = 0; i < ARR_LEN(q->calls); ++i) {
498 /* A clone exist and the copy of "call" in this
499 * clone graph must be exchanged with new one.*/
500 new_call = new_cl_Call(call, cloned_ent, pos);
501 exchange(call, new_call);
506 * The weight formula:
507 * We save one instruction in every caller and param_weight instructions
510 static float calculate_weight(const entry_t *entry) {
511 return ARR_LEN(entry->q.calls) *
512 (get_method_param_weight(entry->q.ent, entry->q.pos) + 1);
516 * After we exchanged all calls, some entries on the list for
517 * the next cloned entity may get invalid, so we have to check
518 * them and may even update the list of heavy uses.
520 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(get_irn_op(ptr) == op_SymConst);
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) {
593 entry_t *entry = NULL, *p;
598 obstack_init(&hmap.obst);
600 hmap.heavy_uses = NULL;
602 /* initially fill our map by visiting all irgs */
603 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
604 irg = get_irp_irg(i);
605 irg_walk_graph(irg, collect_irg_calls, NULL, &hmap);
608 /* We have the "Call" nodes to optimize in set "set_entries". Our algorithm
609 replace one constant parameter and make a new "Call" node for all found "Calls". It exchange the
610 old one with the new one and the algorithm is called with the new "Call".
612 while (hmap.map || hmap.heavy_uses) {
613 /* We iterate the set and arrange the element of the set in a list.
614 The elements are arranged dependent of their value descending.*/
616 foreach_pset(hmap.map, entry) {
617 entry->weight = calculate_weight(entry);
620 * Do not put entry with a weight < threshold in the list
622 if (entry->weight < threshold) {
627 /* put entry in the heavy uses list */
629 if (! hmap.heavy_uses)
630 hmap.heavy_uses = entry;
632 if (entry->weight >= hmap.heavy_uses->weight) {
633 entry->next = hmap.heavy_uses;
634 hmap.heavy_uses = entry;
636 for (p = hmap.heavy_uses; p->next; p = p->next) {
637 if (entry->weight >= p->next->weight) {
638 entry->next = p->next;
652 /* Print some information about the list. */
653 printf("-----------------\n");
654 for (entry = hmap.heavy_uses; entry; entry = entry->next) {
655 printf("\nweight: is %f\n", entry->weight);
656 ir_printf("Call for Method %E\n", entry->q.ent);
657 printf("Position %i\n", entry->q.pos);
658 ir_printf("Value %T\n", entry->q.tv);
661 entry = hmap.heavy_uses;
663 ir_entity *ent = clone_method(&entry->q);
665 hmap.heavy_uses = entry->next;
667 /* We must exchange the copies of this call in all clones too.*/
668 exchange_calls(&entry->q, ent);
672 * after we exchanged all calls, some entries on the list for
673 * the next cloned entity may get invalid, so we have to check
674 * them and may even update the list of heavy uses.
676 reorder_weights(&hmap, threshold);
679 obstack_free(&hmap.obst, NULL);