3 * File name: ir/opt/proc_cloning.c
4 * Purpose: procedure cloning
5 * Author: Beyhan Veliev
8 * Copyright: (c) 1998-2005 Universität Karlsruhe
9 * Licence: This file protected by GPL - GNU GENERAL PUBLIC LICENSE.
13 * @file proc_cloning.c
15 * The purpose is first to find and analyze functions, that are called
16 * with constant parameter(s).
17 * The second step is to optimize the function that are found from our
18 * analyze. Optimize mean to make a new function with parameters, that
19 * aren't be constant. The constant parameters of the function are placed
20 * in the function graph. They aren't be passed as parameters.
43 #include "proc_cloning.h"
44 #include "analyze_irg_args.h"
54 /* A macro to iterate sets.*/
55 #define ITERATE_SET(set_entries, entry) for(entry = set_first(set_entries); entry; entry = set_next(set_entries))
58 * This struct contains the information quadruple for a Call, which we need to
59 * decide if this function must be cloned.
61 typedef struct quadruple {
62 ir_entity *ent; /**< The entity of our Call. */
63 int pos; /**< Position of a constant argument of our Call. */
64 tarval *tv; /**< The tarval of this argument if Const node. */
65 ir_node **calls; /**< The list of all calls with the same characteristics */
69 * The quadruplets are hold in a sorted list
71 typedef struct entry {
72 quadruple_t q; /**< the quadruple */
73 float weight; /**< its weight */
74 struct entry *next; /**< link to the next one */
77 typedef struct q_set {
78 struct obstack obst; /**< an obstack containing all entries */
79 pset *map; /**< a hash map containing the quadruples */
80 entry_t *heavy_uses; /**< the ordered list of heavy uses */
84 * Compare two quadruplets.
86 * @return zero if they are identically, non-zero else
88 static int entry_cmp(const void *elt, const void *key)
90 const entry_t *e1 = elt;
91 const entry_t *e2 = key;
93 return (e1->q.ent != e2->q.ent) || (e1->q.pos != e2->q.pos) || (e1->q.tv != e2->q.tv);
97 * Hash an element of type entry_t.
99 * @param entry The element to be hashed.
101 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)
126 entry_t *key, *entry;
130 n_params = get_Call_n_params(call);
132 /* Beware: we cannot clone variadic parameters as well as the
133 * last non-variadic one, which might be needed for the va_start()
136 mtp = get_Call_type(call);
137 if (get_method_variadicity(mtp) != variadicity_non_variadic) {
138 n_params = get_method_first_variadic_param_index(mtp) - 1;
141 /* In this for loop we collect the calls, that have
142 an constant parameter. */
143 for (i = n_params - 1; i >= 0; --i) {
144 call_param = get_Call_param(call, i);
145 if (is_Const(call_param)) {
146 /* we have found a Call to collect and we save the informations,
149 hmap->map = new_pset(entry_cmp, 8);
151 key = obstack_alloc(&hmap->obst, sizeof(*key));
155 key->q.tv = get_Const_tarval(call_param);
160 /* We insert our information in the set, where we collect the calls.*/
161 entry = pset_insert(hmap->map, key, hash_entry(key));
164 obstack_free(&hmap->obst, key);
166 /* add the call to the list */
167 if (! entry->q.calls) {
168 entry->q.calls = NEW_ARR_F(ir_node *, 1);
169 entry->q.calls[0] = call;
172 ARR_APP1(ir_node *, entry->q.calls, call);
178 * Collect all calls in a ir_graph to a set.
180 * @param call A ir_node to be checked.
181 * @param env The quadruple-set containing the calls with constant parameters
183 static void collect_irg_calls(ir_node *call, void *env)
189 /* We collect just "Call" nodes */
191 call_ptr = get_Call_ptr(call);
193 /* Call pointer must be a SymConst*/
194 if (op_SymConst != get_irn_op(call_ptr))
196 /* Call pointer must be the address of an entity.*/
197 if (get_SymConst_kind(call_ptr) != symconst_addr_ent)
200 callee = get_SymConst_entity(call_ptr);
202 /* we can only clone calls to existing entities */
203 if (get_entity_visibility(callee) == visibility_external_allocated)
206 process_call(call, callee, hmap);
211 * Make a name for a clone. The clone name is
212 * the name of the original method suffixed with "_cl_pos_nr".
213 * pos is the pos from our quadruplet and nr is a counter.
215 * @param id The ident of the cloned function.
216 * @param pos The "pos" from our quadruplet.
217 * @param nr A counter for the clones.
219 static ident *get_clone_ident(ident *id, int pos, unsigned nr)
221 char clone_postfix[32];
223 snprintf(clone_postfix, sizeof(clone_postfix), "_cl_%d_%u", pos, nr);
225 return mangle(id, new_id_from_str(clone_postfix));
229 * Pre-Walker: Copies blocks and nodes from the original method graph
230 * to the cloned graph. Fixes the argument projection numbers for
231 * all arguments behind the removed one.
233 * @param irn A node from the original method graph.
234 * @param env The clone graph.
236 static void copy_nodes(ir_node *irn, void *env)
238 ir_node *arg, *irg_args, *irn_copy;
240 ir_graph *clone_irg = env;
242 arg = get_irg_link(clone_irg);
243 irg_args = get_Proj_pred(arg);
245 /* Copy all nodes except the arg. */
247 copy_irn_to_irg(irn, clone_irg);
249 irn_copy = get_irn_link(irn);
251 /* Fix argument numbers */
252 if (is_Proj(irn) && get_Proj_pred(irn) == irg_args) {
253 proj_nr = get_Proj_proj(irn);
254 if (get_Proj_proj(arg) < proj_nr)
255 set_Proj_proj(irn_copy, proj_nr - 1);
260 * Post-walker: Set the predecessors of the copied nodes.
261 * The copied nodes are set as link of their original nodes. The links of
262 * "irn" predecessors are the predecessors of copied node.
264 static void set_preds(ir_node *irn, void *env)
267 ir_node *irn_copy, *pred, *arg;
268 ir_graph *clone_irg = env;
270 arg = get_irg_link(clone_irg);
271 /* Arg is the method argument, that we have replaced by a constant.*/
275 irn_copy = get_irn_link(irn);
278 for (i = get_Block_n_cfgpreds(irn) - 1; i >= 0; i--) {
279 pred = get_Block_cfgpred(irn, i);
280 /* "End" block must be handled extra, because it is not matured.*/
281 if (get_irg_end_block(current_ir_graph) == irn)
282 add_immBlock_pred(get_irg_end_block(clone_irg), get_irn_link(pred));
284 set_Block_cfgpred(irn_copy, i, get_irn_link(pred));
288 /* First we set the block our copy if it is not a block.*/
289 set_nodes_block(irn_copy, get_irn_link(get_nodes_block(irn)));
290 if (get_irn_op(irn) == op_End) {
291 /* Handle the keep-alives. This must be done separately, because
292 the End node was NOT copied */
293 for (i = 0; i < get_End_n_keepalives(irn); ++i)
294 add_End_keepalive(irn_copy, get_irn_link(get_End_keepalive(irn, i)));
297 for (i = get_irn_arity(irn) - 1; i >= 0; i--) {
298 pred = get_irn_n(irn, i);
299 set_irn_n(irn_copy, i, get_irn_link(pred));
306 * Get the method argument at the position "pos".
308 * @param irg irg that must be cloned.
309 * @param pos The position of the argument.
311 static ir_node *get_irg_arg(ir_graph *irg, int pos)
313 ir_node *irg_args = get_irg_args(irg), *arg = NULL;
316 /* Call algorithm that computes the out edges */
317 if (get_irg_outs_state(irg) != outs_consistent)
318 compute_irg_outs(irg);
320 /* Search the argument with the number pos.*/
321 for (i = get_irn_n_outs(irg_args) - 1; i >= 0; --i) {
322 ir_node *proj = get_irn_out(irg_args, i);
323 if (pos == get_Proj_proj(proj)) {
326 * More than one arg node found:
327 * We rely on the fact the only one arg exists, so do
328 * a cheap CSE in this case.
330 set_irn_out(irg_args, i, arg);
337 assert(arg && "Argument not found");
342 * Create a new graph for the clone of the method,
343 * that we want to clone.
345 * @param ent The entity of the method that must be cloned.
346 * @param q Our quadruplet.
348 static void create_clone_proc_irg(ir_entity *ent, quadruple_t *q)
350 ir_graph *method_irg, *clone_irg;
351 ir_node *arg, *const_arg;
353 method_irg = get_entity_irg(ent);
355 /* We create the skeleton of the clone irg.*/
356 clone_irg = new_ir_graph(ent, 0);
358 arg = get_irg_arg(get_entity_irg(q->ent), q->pos);
359 /* we will replace the argument in position "q->pos" by this constant. */
360 const_arg = new_r_Const_type(
361 clone_irg, get_nodes_block(arg), get_irn_mode(arg), q->tv,
362 get_method_param_type(get_entity_type(q->ent), q->pos));
364 /* args copy in the cloned graph will be the const. */
365 set_irn_link(arg, const_arg);
367 /* Store the arg that will be replaced here, so we can easily detect it. */
368 set_irg_link(clone_irg, arg);
370 /* We copy the blocks and nodes, that must be in
371 the clone graph and set their predecessors. */
372 irg_walk_graph(method_irg, copy_nodes, set_preds, clone_irg);
374 /* The "cloned" graph must be matured. */
375 mature_immBlock(get_irg_end_block(clone_irg));
376 irg_finalize_cons(clone_irg);
380 * The function create a new entity type
381 * for our clone and set it to clone entity.
383 * @param q Contains information for the method to clone.
384 * @param ent The entity of the clone.
385 * @param nr A pointer to the counter of clones.
387 static void change_entity_type(quadruple_t *q, ir_entity *ent, unsigned *nr)
389 ir_type *mtp, *new_mtp, *tp;
391 int i, j, n_params, n_ress;
393 mtp = get_entity_type(q->ent);
394 tp_name = get_clone_ident(get_type_ident(mtp), q->pos, (*nr)++);
395 n_params = get_method_n_params(mtp);
396 n_ress = get_method_n_ress(mtp);
398 /* Create the new type for our clone. It must have one parameter
399 less then the original.*/
400 new_mtp = new_type_method(tp_name, n_params - 1, n_ress);
402 /* We must set the type of the methods parameters.*/
403 for (i = j = 0; i < n_params; ++i) {
405 /* This is the position of the argument, that we have
409 tp = get_method_param_type(mtp, i);
410 set_method_param_type(new_mtp, j++, tp);
412 /* Copy the methods result types. */
413 for (i = 0; i < n_ress; ++i) {
414 tp = get_method_res_type(mtp, i);
415 set_method_res_type(new_mtp, i, tp);
417 set_entity_type(ent, new_mtp);
421 * Make a clone of a method.
423 * @param q Contains information for the method to clone.
425 static ir_entity *clone_method(quadruple_t *q)
427 ir_entity *new_entity;
431 /* A counter for the clones.*/
432 static unsigned nr = 0;
434 /* We get a new ident for our clone method.*/
435 clone_ident = get_clone_ident(get_entity_ident(q->ent), q->pos, nr);
436 /* We get our entity for the clone method. */
437 new_entity = copy_entity_name(q->ent, clone_ident);
439 /* a cloned entity is always local */
440 set_entity_visibility(new_entity, visibility_local);
442 /* set a ld name here: Should we mangle this ? */
443 set_entity_ld_ident(new_entity, get_entity_ident(new_entity));
445 /* set a new type here. */
446 change_entity_type(q, new_entity, &nr);
448 /* We need now a new ir_graph for our clone method. */
449 create_clone_proc_irg(new_entity, q);
451 /* We must set the atomic value of our "new_entity". */
452 sym.entity_p = new_entity;
453 rem = current_ir_graph;
454 current_ir_graph = get_const_code_irg();
455 new_entity->value = new_SymConst(sym, symconst_addr_ent);
456 current_ir_graph = rem;
458 /* The "new_entity" don't have this information. */
459 new_entity->attr.mtd_attr.param_access = NULL;
460 new_entity->attr.mtd_attr.param_weight = NULL;
466 * Creates a new "cloned" Call node and return it.
468 * @param call The call that must be cloned.
469 * @param new_entity The entity of the cloned function.
470 * @param pos The position of the replaced parameter of this call.
472 static ir_node *new_cl_Call(ir_node *call, ir_entity *new_entity, int pos)
476 int i, n_params, new_params = 0;
479 ir_graph *irg = get_irn_irg(call);
480 ir_node *bl = get_nodes_block(call);
482 sym.entity_p = new_entity;
483 callee = new_r_SymConst(irg, bl, sym, symconst_addr_ent);
485 mtp = get_entity_type(new_entity);
486 n_params = get_Call_n_params(call);
487 NEW_ARR_A(ir_node *, in, n_params - 1);
489 /* we save the parameters of the new call in the array "in" without the
490 * parameter in position "pos", that is replaced with a constant.*/
491 for (i = 0; i < n_params; i++){
493 in[new_params++] = get_Call_param(call, i);
495 /* Create and return the new Call. */
496 return new_r_Call(irg, bl, get_Call_mem(call),
497 callee, n_params - 1, in, get_entity_type(new_entity));
501 * Exchange all Calls stored in the quadruplet to Calls of the cloned entity.
503 * @param q The quadruple
504 * @param cloned_ent The entity of the new function that must be called
507 static void exchange_calls(quadruple_t *q, ir_entity *cloned_ent)
510 ir_node *new_call, *call;
513 /* We iterate the list of the "call".*/
514 for (i = 0; i < ARR_LEN(q->calls); ++i) {
517 /* A clone exist and the copy of "call" in this
518 * clone graph must be exchanged with new one.*/
519 new_call = new_cl_Call(call, cloned_ent, pos);
520 exchange(call, new_call);
525 * The weight formula:
526 * We save one instruction in every caller and param_weight instructions
529 static float calculate_weight(const entry_t *entry) {
530 return ARR_LEN(entry->q.calls) *
531 (get_method_param_weight(entry->q.ent, entry->q.pos) + 1);
535 * After we exchanged all calls, some entries on the list for
536 * the next cloned entity may get invalid, so we have to check
537 * them and may even update the list of heavy uses.
539 static void reorder_weights(q_set *hmap, float threshold)
541 entry_t **adr, *p, *entry;
546 entry = hmap->heavy_uses;
550 len = ARR_LEN(entry->q.calls);
551 for (i = 0; i < len; ++i) {
552 ir_node *ptr, *call = entry->q.calls[i];
554 /* might be exchanged, so skip Id nodes here. */
555 call = skip_Id(call);
557 /* we know, that a SymConst is here */
558 ptr = get_Call_ptr(call);
559 assert(get_irn_op(ptr) == op_SymConst);
561 callee = get_SymConst_entity(ptr);
562 if (callee != entry->q.ent) {
564 * This call is already changed because of a previous
565 * optimization. Remove it from the list.
568 entry->q.calls[i] = entry->q.calls[len];
569 entry->q.calls[len] = NULL;
571 /* the new call should be processed */
572 process_call(call, callee, hmap);
577 /* the length might be changed */
578 ARR_SHRINKLEN(entry->q.calls, len);
580 /* recalculate the weight and resort the heavy uses map */
581 entry->weight = calculate_weight(entry);
583 if (len <= 0 || entry->weight < threshold) {
584 hmap->heavy_uses = entry->next;
587 /* we have changed the list, check the next one */
592 for (p = entry->next; p && entry->weight < p->weight; p = p->next) {
597 hmap->heavy_uses = entry->next;
600 entry = hmap->heavy_uses;
602 /* we have changed the list, check the next one */
608 * Do the procedure cloning. Evaluate a heuristic weight for every
609 * call(..., Const, ...). If the weight is bigger than threshold,
610 * clone the entity and fix the calls.
612 void proc_cloning(float threshold)
614 entry_t *entry = NULL, *p;
619 obstack_init(&hmap.obst);
621 hmap.heavy_uses = NULL;
623 /* initially fill our map by visiting all irgs */
624 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
625 irg = get_irp_irg(i);
626 irg_walk_graph(irg, collect_irg_calls, NULL, &hmap);
629 /* We have the "Call" nodes to optimize in set "set_entries". Our algorithm
630 replace one constant parameter and make a new "Call" node for all found "Calls". It exchange the
631 old one with the new one and the algorithm is called with the new "Call".
633 while (hmap.map || hmap.heavy_uses) {
634 /* We iterate the set and arrange the element of the set in a list.
635 The elements are arranged dependent of their value descending.*/
637 for (entry = pset_first(hmap.map); entry; entry = pset_next(hmap.map)) {
638 entry->weight = calculate_weight(entry);
641 * Do not put entry with a weight < threshold in the list
643 if (entry->weight < threshold) {
648 /* put entry in the heavy uses list */
650 if (! hmap.heavy_uses)
651 hmap.heavy_uses = entry;
653 if (entry->weight >= hmap.heavy_uses->weight) {
654 entry->next = hmap.heavy_uses;
655 hmap.heavy_uses = entry;
658 for (p = hmap.heavy_uses; p->next; p = p->next) {
659 if (entry->weight >= p->next->weight) {
660 entry->next = p->next;
674 /* Print some information about the list. */
675 printf("-----------------\n");
676 for (entry = hmap.heavy_uses; entry; entry = entry->next) {
677 printf("\nweight: is %f\n", entry->weight);
678 ir_printf("Call for Method %E\n", entry->q.ent);
679 printf("Position %i\n", entry->q.pos);
680 ir_printf("Value %T\n", entry->q.tv);
683 entry = hmap.heavy_uses;
685 ir_entity *ent = clone_method(&entry->q);
687 hmap.heavy_uses = entry->next;
689 /* We must exchange the copies of this call in all clones too.*/
690 exchange_calls(&entry->q, ent);
694 * after we exchanged all calls, some entries on the list for
695 * the next cloned entity may get invalid, so we have to check
696 * them and may even update the list of heavy uses.
698 reorder_weights(&hmap, threshold);
701 obstack_free(&hmap.obst, NULL);