3 * File name: ir/ir/irgopt.c
4 * Purpose: Optimizations for a whole ir graph, i.e., a procedure.
5 * Author: Christian Schaefer, Goetz Lindenmaier
6 * Modified by: Sebastian Felis
9 * Copyright: (c) 1998-2003 Universität Karlsruhe
10 * Licence: This file protected by GPL - GNU GENERAL PUBLIC LICENSE.
22 #include "irgraph_t.h"
34 #include "pdeq.h" /* Fuer code placement */
38 #include "irbackedge_t.h"
44 /* Defined in iropt.c */
45 pset *new_identities (void);
46 void del_identities (pset *value_table);
47 void add_identities (pset *value_table, ir_node *node);
49 /*------------------------------------------------------------------*/
50 /* apply optimizations of iropt to all nodes. */
51 /*------------------------------------------------------------------*/
53 static void init_link (ir_node *n, void *env) {
54 set_irn_link(n, NULL);
57 #if 0 /* Old version. Avoids Ids.
58 This is not necessary: we do a postwalk, and get_irn_n
59 removes ids anyways. So it's much cheaper to call the
60 optimization less often and use the exchange() algorithm. */
62 optimize_in_place_wrapper (ir_node *n, void *env) {
64 ir_node *optimized, *old;
66 irn_arity = get_irn_arity(n);
67 for (i = 0; i < irn_arity; i++) {
68 /* get_irn_n skips Id nodes, so comparison old != optimized does not
69 show all optimizations. Therefore always set new predecessor. */
70 old = get_irn_intra_n(n, i);
71 optimized = optimize_in_place_2(old);
72 set_irn_n(n, i, optimized);
75 if (get_irn_op(n) == op_Block) {
76 optimized = optimize_in_place_2(n);
77 if (optimized != n) exchange (n, optimized);
82 optimize_in_place_wrapper (ir_node *n, void *env) {
83 ir_node *optimized = optimize_in_place_2(n);
84 if (optimized != n) exchange (n, optimized);
89 static INLINE void do_local_optimize(ir_node *n) {
90 /* Handle graph state */
91 assert(get_irg_phase_state(current_ir_graph) != phase_building);
92 if (get_opt_global_cse())
93 set_irg_pinned(current_ir_graph, op_pin_state_floats);
94 if (get_irg_outs_state(current_ir_graph) == outs_consistent)
95 set_irg_outs_inconsistent(current_ir_graph);
96 if (get_irg_dom_state(current_ir_graph) == dom_consistent)
97 set_irg_dom_inconsistent(current_ir_graph);
98 set_irg_loopinfo_inconsistent(current_ir_graph);
101 /* Clean the value_table in irg for the cse. */
102 del_identities(current_ir_graph->value_table);
103 current_ir_graph->value_table = new_identities();
105 /* walk over the graph */
106 irg_walk(n, init_link, optimize_in_place_wrapper, NULL);
109 void local_optimize_node(ir_node *n) {
110 ir_graph *rem = current_ir_graph;
111 current_ir_graph = get_irn_irg(n);
113 do_local_optimize(n);
115 current_ir_graph = rem;
120 local_optimize_graph (ir_graph *irg) {
121 ir_graph *rem = current_ir_graph;
122 current_ir_graph = irg;
124 do_local_optimize(irg->end);
126 current_ir_graph = rem;
130 /*------------------------------------------------------------------*/
131 /* Routines for dead node elimination / copying garbage collection */
132 /* of the obstack. */
133 /*------------------------------------------------------------------*/
136 * Remember the new node in the old node by using a field all nodes have.
139 set_new_node (ir_node *old, ir_node *new)
145 * Get this new node, before the old node is forgotton.
147 static INLINE ir_node *
148 get_new_node (ir_node * n)
154 * We use the block_visited flag to mark that we have computed the
155 * number of useful predecessors for this block.
156 * Further we encode the new arity in this flag in the old blocks.
157 * Remembering the arity is useful, as it saves a lot of pointer
158 * accesses. This function is called for all Phi and Block nodes
162 compute_new_arity(ir_node *b) {
163 int i, res, irn_arity;
166 irg_v = get_irg_block_visited(current_ir_graph);
167 block_v = get_Block_block_visited(b);
168 if (block_v >= irg_v) {
169 /* we computed the number of preds for this block and saved it in the
171 return block_v - irg_v;
173 /* compute the number of good predecessors */
174 res = irn_arity = get_irn_arity(b);
175 for (i = 0; i < irn_arity; i++)
176 if (get_irn_opcode(get_irn_n(b, i)) == iro_Bad) res--;
177 /* save it in the flag. */
178 set_Block_block_visited(b, irg_v + res);
183 /* TODO: add an ir_op operation */
184 static INLINE void new_backedge_info(ir_node *n) {
185 switch(get_irn_opcode(n)) {
187 n->attr.block.cg_backedge = NULL;
188 n->attr.block.backedge = new_backedge_arr(current_ir_graph->obst, get_irn_arity(n));
191 n->attr.phi_backedge = new_backedge_arr(current_ir_graph->obst, get_irn_arity(n));
194 n->attr.filter.backedge = new_backedge_arr(current_ir_graph->obst, get_irn_arity(n));
201 * Copies the node to the new obstack. The Ins of the new node point to
202 * the predecessors on the old obstack. For block/phi nodes not all
203 * predecessors might be copied. n->link points to the new node.
204 * For Phi and Block nodes the function allocates in-arrays with an arity
205 * only for useful predecessors. The arity is determined by counting
206 * the non-bad predecessors of the block.
209 copy_node (ir_node *n, void *env) {
212 opcode op = get_irn_opcode(n);
213 /* The end node looses it's flexible in array. This doesn't matter,
214 as dead node elimination builds End by hand, inlineing doesn't use
216 /* assert(n->op == op_End || ((_ARR_DESCR(n->in))->cookie != ARR_F_MAGIC)); */
219 /* node copied already */
221 } else if (op == iro_Block) {
223 new_arity = compute_new_arity(n);
224 n->attr.block.graph_arr = NULL;
226 block = get_nodes_block(n);
227 if (get_irn_opcode(n) == iro_Phi) {
228 new_arity = compute_new_arity(block);
230 new_arity = get_irn_arity(n);
233 nn = new_ir_node(get_irn_dbg_info(n),
240 /* Copy the attributes. These might point to additional data. If this
241 was allocated on the old obstack the pointers now are dangling. This
242 frees e.g. the memory of the graph_arr allocated in new_immBlock. */
244 new_backedge_info(nn);
247 /* printf("\n old node: "); DDMSG2(n);
248 printf(" new node: "); DDMSG2(nn); */
253 * Copies new predecessors of old node to new node remembered in link.
254 * Spare the Bad predecessors of Phi and Block nodes.
257 copy_preds (ir_node *n, void *env) {
261 nn = get_new_node(n);
263 /* printf("\n old node: "); DDMSG2(n);
264 printf(" new node: "); DDMSG2(nn);
265 printf(" arities: old: %d, new: %d\n", get_irn_arity(n), get_irn_arity(nn)); */
267 if (get_irn_opcode(n) == iro_Block) {
268 /* Don't copy Bad nodes. */
270 irn_arity = get_irn_arity(n);
271 for (i = 0; i < irn_arity; i++)
272 if (get_irn_opcode(get_irn_n(n, i)) != iro_Bad) {
273 set_irn_n (nn, j, get_new_node(get_irn_n(n, i)));
274 /*if (is_backedge(n, i)) set_backedge(nn, j);*/
277 /* repair the block visited flag from above misuse. Repair it in both
278 graphs so that the old one can still be used. */
279 set_Block_block_visited(nn, 0);
280 set_Block_block_visited(n, 0);
281 /* Local optimization could not merge two subsequent blocks if
282 in array contained Bads. Now it's possible.
283 We don't call optimize_in_place as it requires
284 that the fields in ir_graph are set properly. */
285 if ((get_opt_control_flow_straightening()) &&
286 (get_Block_n_cfgpreds(nn) == 1) &&
287 (get_irn_op(get_Block_cfgpred(nn, 0)) == op_Jmp)) {
288 ir_node *old = get_nodes_block(get_Block_cfgpred(nn, 0));
290 /* Jmp jumps into the block it is in -- deal self cycle. */
291 assert(is_Bad(get_new_node(get_irg_bad(current_ir_graph))));
292 exchange(nn, get_new_node(get_irg_bad(current_ir_graph)));
297 } else if (get_irn_opcode(n) == iro_Phi) {
298 /* Don't copy node if corresponding predecessor in block is Bad.
299 The Block itself should not be Bad. */
300 block = get_nodes_block(n);
301 set_irn_n (nn, -1, get_new_node(block));
303 irn_arity = get_irn_arity(n);
304 for (i = 0; i < irn_arity; i++)
305 if (get_irn_opcode(get_irn_n(block, i)) != iro_Bad) {
306 set_irn_n (nn, j, get_new_node(get_irn_n(n, i)));
307 /*if (is_backedge(n, i)) set_backedge(nn, j);*/
310 /* If the pre walker reached this Phi after the post walker visited the
311 block block_visited is > 0. */
312 set_Block_block_visited(get_nodes_block(n), 0);
313 /* Compacting the Phi's ins might generate Phis with only one
315 if (get_irn_arity(n) == 1)
316 exchange(n, get_irn_n(n, 0));
318 irn_arity = get_irn_arity(n);
319 for (i = -1; i < irn_arity; i++)
320 set_irn_n (nn, i, get_new_node(get_irn_n(n, i)));
322 /* Now the new node is complete. We can add it to the hash table for cse.
323 @@@ inlinening aborts if we identify End. Why? */
324 if(get_irn_op(nn) != op_End)
325 add_identities (current_ir_graph->value_table, nn);
329 * Copies the graph recursively, compacts the keepalive of the end node.
333 ir_node *oe, *ne, *ob, *nb; /* old end, new end, old bad, new bad */
334 ir_node *ka; /* keep alive */
337 oe = get_irg_end(current_ir_graph);
338 /* copy the end node by hand, allocate dynamic in array! */
339 ne = new_ir_node(get_irn_dbg_info(oe),
346 /* Copy the attributes. Well, there might be some in the future... */
348 set_new_node(oe, ne);
350 ob = get_irg_bad(current_ir_graph);
351 nb = new_ir_node(get_irn_dbg_info(ob),
358 set_new_node(ob, nb);
360 /* copy the live nodes */
361 irg_walk(get_nodes_block(oe), copy_node, copy_preds, NULL);
362 /* copy_preds for the end node ... */
363 set_nodes_block(ne, get_new_node(get_nodes_block(oe)));
365 /*- ... and now the keep alives. -*/
366 /* First pick the not marked block nodes and walk them. We must pick these
367 first as else we will oversee blocks reachable from Phis. */
368 irn_arity = get_irn_arity(oe);
369 for (i = 0; i < irn_arity; i++) {
370 ka = get_irn_intra_n(oe, i);
371 if ((get_irn_op(ka) == op_Block) &&
372 (get_irn_visited(ka) < get_irg_visited(current_ir_graph))) {
373 /* We must keep the block alive and copy everything reachable */
374 set_irg_visited(current_ir_graph, get_irg_visited(current_ir_graph)-1);
375 irg_walk(ka, copy_node, copy_preds, NULL);
376 add_End_keepalive(ne, get_new_node(ka));
380 /* Now pick the Phis. Here we will keep all! */
381 irn_arity = get_irn_arity(oe);
382 for (i = 0; i < irn_arity; i++) {
383 ka = get_irn_intra_n(oe, i);
384 if ((get_irn_op(ka) == op_Phi)) {
385 if (get_irn_visited(ka) < get_irg_visited(current_ir_graph)) {
386 /* We didn't copy the Phi yet. */
387 set_irg_visited(current_ir_graph, get_irg_visited(current_ir_graph)-1);
388 irg_walk(ka, copy_node, copy_preds, NULL);
390 add_End_keepalive(ne, get_new_node(ka));
394 /* start block somtimes only reached after keep alives */
395 set_nodes_block(nb, get_new_node(get_nodes_block(ob)));
399 * Copies the graph reachable from current_ir_graph->end to the obstack
400 * in current_ir_graph and fixes the environment.
401 * Then fixes the fields in current_ir_graph containing nodes of the
405 copy_graph_env (void) {
407 /* Not all nodes remembered in current_ir_graph might be reachable
408 from the end node. Assure their link is set to NULL, so that
409 we can test whether new nodes have been computed. */
410 set_irn_link(get_irg_frame (current_ir_graph), NULL);
411 set_irn_link(get_irg_globals (current_ir_graph), NULL);
412 set_irn_link(get_irg_args (current_ir_graph), NULL);
413 set_irn_link(get_irg_initial_mem(current_ir_graph), NULL);
415 /* we use the block walk flag for removing Bads from Blocks ins. */
416 inc_irg_block_visited(current_ir_graph);
421 /* fix the fields in current_ir_graph */
422 old_end = get_irg_end(current_ir_graph);
423 set_irg_end (current_ir_graph, get_new_node(old_end));
424 set_irg_end_except (current_ir_graph, get_irg_end(current_ir_graph));
425 set_irg_end_reg (current_ir_graph, get_irg_end(current_ir_graph));
427 set_irg_end_block (current_ir_graph, get_new_node(get_irg_end_block(current_ir_graph)));
428 if (get_irn_link(get_irg_frame(current_ir_graph)) == NULL) {
429 copy_node (get_irg_frame(current_ir_graph), NULL);
430 copy_preds(get_irg_frame(current_ir_graph), NULL);
432 if (get_irn_link(get_irg_globals(current_ir_graph)) == NULL) {
433 copy_node (get_irg_globals(current_ir_graph), NULL);
434 copy_preds(get_irg_globals(current_ir_graph), NULL);
436 if (get_irn_link(get_irg_initial_mem(current_ir_graph)) == NULL) {
437 copy_node (get_irg_initial_mem(current_ir_graph), NULL);
438 copy_preds(get_irg_initial_mem(current_ir_graph), NULL);
440 if (get_irn_link(get_irg_args(current_ir_graph)) == NULL) {
441 copy_node (get_irg_args(current_ir_graph), NULL);
442 copy_preds(get_irg_args(current_ir_graph), NULL);
444 set_irg_start (current_ir_graph, get_new_node(get_irg_start(current_ir_graph)));
446 set_irg_start_block(current_ir_graph,
447 get_new_node(get_irg_start_block(current_ir_graph)));
448 set_irg_frame (current_ir_graph, get_new_node(get_irg_frame(current_ir_graph)));
449 set_irg_globals (current_ir_graph, get_new_node(get_irg_globals(current_ir_graph)));
450 set_irg_initial_mem(current_ir_graph, get_new_node(get_irg_initial_mem(current_ir_graph)));
451 set_irg_args (current_ir_graph, get_new_node(get_irg_args(current_ir_graph)));
453 if (get_irn_link(get_irg_bad(current_ir_graph)) == NULL) {
454 copy_node(get_irg_bad(current_ir_graph), NULL);
455 copy_preds(get_irg_bad(current_ir_graph), NULL);
457 set_irg_bad(current_ir_graph, get_new_node(get_irg_bad(current_ir_graph)));
461 * Copies all reachable nodes to a new obstack. Removes bad inputs
462 * from block nodes and the corresponding inputs from Phi nodes.
463 * Merges single exit blocks with single entry blocks and removes
465 * Adds all new nodes to a new hash table for cse. Does not
466 * perform cse, so the hash table might contain common subexpressions.
469 dead_node_elimination(ir_graph *irg) {
471 int rem_ipview = interprocedural_view;
472 struct obstack *graveyard_obst = NULL;
473 struct obstack *rebirth_obst = NULL;
475 /* inform statistics that we started a dead-node elimination run */
476 stat_dead_node_elim_start(irg);
478 /* Remember external state of current_ir_graph. */
479 rem = current_ir_graph;
480 current_ir_graph = irg;
481 interprocedural_view = 0;
483 /* Handle graph state */
484 assert(get_irg_phase_state(current_ir_graph) != phase_building);
485 free_callee_info(current_ir_graph);
486 free_outs(current_ir_graph);
487 /* @@@ so far we loose loops when copying */
488 free_loop_information(current_ir_graph);
490 if (get_opt_optimize() && get_opt_dead_node_elimination()) {
492 /* A quiet place, where the old obstack can rest in peace,
493 until it will be cremated. */
494 graveyard_obst = irg->obst;
496 /* A new obstack, where the reachable nodes will be copied to. */
497 rebirth_obst = (struct obstack *) xmalloc (sizeof (struct obstack));
498 current_ir_graph->obst = rebirth_obst;
499 obstack_init (current_ir_graph->obst);
501 /* We also need a new hash table for cse */
502 del_identities (irg->value_table);
503 irg->value_table = new_identities ();
505 /* Copy the graph from the old to the new obstack */
508 /* Free memory from old unoptimized obstack */
509 obstack_free(graveyard_obst, 0); /* First empty the obstack ... */
510 xfree (graveyard_obst); /* ... then free it. */
513 /* inform statistics that the run is over */
514 stat_dead_node_elim_stop(irg);
516 current_ir_graph = rem;
517 interprocedural_view = rem_ipview;
521 * Relink bad predeseccors of a block and store the old in array to the
522 * link field. This function is called by relink_bad_predecessors().
523 * The array of link field starts with the block operand at position 0.
524 * If block has bad predecessors, create a new in array without bad preds.
525 * Otherwise let in array untouched.
527 static void relink_bad_block_predecessors(ir_node *n, void *env) {
528 ir_node **new_in, *irn;
529 int i, new_irn_n, old_irn_arity, new_irn_arity = 0;
531 /* if link field of block is NULL, look for bad predecessors otherwise
532 this is allready done */
533 if (get_irn_op(n) == op_Block &&
534 get_irn_link(n) == NULL) {
536 /* save old predecessors in link field (position 0 is the block operand)*/
537 set_irn_link(n, (void *)get_irn_in(n));
539 /* count predecessors without bad nodes */
540 old_irn_arity = get_irn_arity(n);
541 for (i = 0; i < old_irn_arity; i++)
542 if (!is_Bad(get_irn_n(n, i))) new_irn_arity++;
544 /* arity changing: set new predecessors without bad nodes */
545 if (new_irn_arity < old_irn_arity) {
546 /* get new predecessor array without Block predecessor */
547 new_in = NEW_ARR_D (ir_node *, current_ir_graph->obst, (new_irn_arity+1));
549 /* set new predeseccors in array */
552 for (i = 1; i < old_irn_arity; i++) {
553 irn = get_irn_n(n, i);
554 if (!is_Bad(irn)) new_in[new_irn_n++] = irn;
557 } /* ir node has bad predecessors */
559 } /* Block is not relinked */
563 * Relinks Bad predecesors from Bocks and Phis called by walker
564 * remove_bad_predecesors(). If n is a Block, call
565 * relink_bad_block_redecessors(). If n is a Phinode, call also the relinking
566 * function of Phi's Block. If this block has bad predecessors, relink preds
569 static void relink_bad_predecessors(ir_node *n, void *env) {
570 ir_node *block, **old_in;
571 int i, old_irn_arity, new_irn_arity;
573 /* relink bad predeseccors of a block */
574 if (get_irn_op(n) == op_Block)
575 relink_bad_block_predecessors(n, env);
577 /* If Phi node relink its block and its predecessors */
578 if (get_irn_op(n) == op_Phi) {
580 /* Relink predeseccors of phi's block */
581 block = get_nodes_block(n);
582 if (get_irn_link(block) == NULL)
583 relink_bad_block_predecessors(block, env);
585 old_in = (ir_node **)get_irn_link(block); /* Of Phi's Block */
586 old_irn_arity = ARR_LEN(old_in);
588 /* Relink Phi predeseccors if count of predeseccors changed */
589 if (old_irn_arity != ARR_LEN(get_irn_in(block))) {
590 /* set new predeseccors in array
591 n->in[0] remains the same block */
593 for(i = 1; i < old_irn_arity; i++)
594 if (!is_Bad((ir_node *)old_in[i])) n->in[new_irn_arity++] = n->in[i];
596 ARR_SETLEN(ir_node *, n->in, new_irn_arity);
599 } /* n is a Phi node */
603 * Removes Bad Bad predecesors from Blocks and the corresponding
604 * inputs to Phi nodes as in dead_node_elimination but without
606 * On walking up set the link field to NULL, on walking down call
607 * relink_bad_predecessors() (This function stores the old in array
608 * to the link field and sets a new in array if arity of predecessors
611 void remove_bad_predecessors(ir_graph *irg) {
612 irg_walk_graph(irg, init_link, relink_bad_predecessors, NULL);
616 /*--------------------------------------------------------------------*/
617 /* Funcionality for inlining */
618 /*--------------------------------------------------------------------*/
621 * Copy node for inlineing. Updates attributes that change when
622 * inlineing but not for dead node elimination.
624 * Copies the node by calling copy_node and then updates the entity if
625 * it's a local one. env must be a pointer of the frame type of the
626 * inlined procedure. The new entities must be in the link field of
630 copy_node_inline (ir_node *n, void *env) {
632 type *frame_tp = (type *)env;
635 if (get_irn_op(n) == op_Sel) {
636 new = get_new_node (n);
637 assert(get_irn_op(new) == op_Sel);
638 if (get_entity_owner(get_Sel_entity(n)) == frame_tp) {
639 set_Sel_entity(new, get_entity_link(get_Sel_entity(n)));
641 } else if (get_irn_op(n) == op_Block) {
642 new = get_new_node (n);
643 new->attr.block.irg = current_ir_graph;
647 static void find_addr(ir_node *node, void *env)
649 if (get_irn_opcode(node) == iro_Proj) {
650 if (get_Proj_proj(node) == pn_Start_P_value_arg_base)
656 * currently, we cannot inline two cases:
657 * - call with compound arguments
658 * - graphs that take the address of a parameter
660 * check these condition here
662 static int can_inline(ir_node *call, ir_graph *called_graph)
664 type *call_type = get_Call_type(call);
665 int params, ress, i, res;
666 assert(is_method_type(call_type));
668 params = get_method_n_params(call_type);
669 ress = get_method_n_ress(call_type);
672 for (i = 0; i < params; ++i) {
673 type *p_type = get_method_param_type(call_type, i);
675 if (is_compound_type(p_type))
680 for (i = 0; i < ress; ++i) {
681 type *r_type = get_method_res_type(call_type, i);
683 if (is_compound_type(r_type))
688 irg_walk_graph(called_graph, find_addr, NULL, &res);
693 int inline_method(ir_node *call, ir_graph *called_graph) {
695 ir_node *post_call, *post_bl;
697 ir_node *end, *end_bl;
701 int arity, n_ret, n_exc, n_res, i, j, rem_opt, irn_arity;
704 irg_inline_property prop = get_irg_inline_property(called_graph);
706 if ( (prop != irg_inline_forced) &&
707 (!get_opt_optimize() || !get_opt_inline() || (prop == irg_inline_forbidden))) return 0;
711 * currently, we cannot inline two cases:
712 * - call with compound arguments
713 * - graphs that take the address of a parameter
715 if (! can_inline(call, called_graph))
718 /* -- Turn off optimizations, this can cause problems when allocating new nodes. -- */
719 rem_opt = get_opt_optimize();
722 /* Handle graph state */
723 assert(get_irg_phase_state(current_ir_graph) != phase_building);
724 assert(get_irg_pinned(current_ir_graph) == op_pin_state_pinned);
725 assert(get_irg_pinned(called_graph) == op_pin_state_pinned);
726 if (get_irg_outs_state(current_ir_graph) == outs_consistent)
727 set_irg_outs_inconsistent(current_ir_graph);
728 set_irg_loopinfo_inconsistent(current_ir_graph);
730 /* -- Check preconditions -- */
731 assert(get_irn_op(call) == op_Call);
732 /* @@@ does not work for InterfaceIII.java after cgana
733 assert(get_Call_type(call) == get_entity_type(get_irg_entity(called_graph)));
734 assert(smaller_type(get_entity_type(get_irg_entity(called_graph)),
735 get_Call_type(call)));
737 assert(get_type_tpop(get_Call_type(call)) == type_method);
738 if (called_graph == current_ir_graph) {
739 set_optimize(rem_opt);
743 /* here we know we WILL inline, so inform the statistics */
744 stat_inline(call, called_graph);
746 /* -- Decide how to handle exception control flow: Is there a handler
747 for the Call node, or do we branch directly to End on an exception?
748 exc_handling: 0 There is a handler.
750 2 Exception handling not represented in Firm. -- */
752 ir_node *proj, *Mproj = NULL, *Xproj = NULL;
753 for (proj = (ir_node *)get_irn_link(call); proj; proj = (ir_node *)get_irn_link(proj)) {
754 assert(get_irn_op(proj) == op_Proj);
755 if (get_Proj_proj(proj) == pn_Call_X_except) Xproj = proj;
756 if (get_Proj_proj(proj) == pn_Call_M_except) Mproj = proj;
758 if (Mproj) { assert(Xproj); exc_handling = 0; } /* Mproj */
759 else if (Xproj) { exc_handling = 1; } /* !Mproj && Xproj */
760 else { exc_handling = 2; } /* !Mproj && !Xproj */
765 the procedure and later replaces the Start node of the called graph.
766 Post_call is the old Call node and collects the results of the called
767 graph. Both will end up being a tuple. -- */
768 post_bl = get_nodes_block(call);
769 set_irg_current_block(current_ir_graph, post_bl);
770 /* XxMxPxP of Start + parameter of Call */
771 in[pn_Start_X_initial_exec] = new_Jmp();
772 in[pn_Start_M] = get_Call_mem(call);
773 in[pn_Start_P_frame_base] = get_irg_frame(current_ir_graph);
774 in[pn_Start_P_globals] = get_irg_globals(current_ir_graph);
775 in[pn_Start_T_args] = new_Tuple(get_Call_n_params(call), get_Call_param_arr(call));
776 /* in[pn_Start_P_value_arg_base] = ??? */
777 pre_call = new_Tuple(5, in);
781 The new block gets the ins of the old block, pre_call and all its
782 predecessors and all Phi nodes. -- */
783 part_block(pre_call);
785 /* -- Prepare state for dead node elimination -- */
786 /* Visited flags in calling irg must be >= flag in called irg.
787 Else walker and arity computation will not work. */
788 if (get_irg_visited(current_ir_graph) <= get_irg_visited(called_graph))
789 set_irg_visited(current_ir_graph, get_irg_visited(called_graph)+1);
790 if (get_irg_block_visited(current_ir_graph)< get_irg_block_visited(called_graph))
791 set_irg_block_visited(current_ir_graph, get_irg_block_visited(called_graph));
792 /* Set pre_call as new Start node in link field of the start node of
793 calling graph and pre_calls block as new block for the start block
795 Further mark these nodes so that they are not visited by the
797 set_irn_link(get_irg_start(called_graph), pre_call);
798 set_irn_visited(get_irg_start(called_graph), get_irg_visited(current_ir_graph));
799 set_irn_link(get_irg_start_block(called_graph), get_nodes_block(pre_call));
800 set_irn_visited(get_irg_start_block(called_graph), get_irg_visited(current_ir_graph));
801 set_irn_link(get_irg_bad(called_graph), get_irg_bad(current_ir_graph));
802 set_irn_visited(get_irg_bad(called_graph), get_irg_visited(current_ir_graph));
804 /* Initialize for compaction of in arrays */
805 inc_irg_block_visited(current_ir_graph);
807 /* -- Replicate local entities of the called_graph -- */
808 /* copy the entities. */
809 called_frame = get_irg_frame_type(called_graph);
810 for (i = 0; i < get_class_n_members(called_frame); i++) {
811 entity *new_ent, *old_ent;
812 old_ent = get_class_member(called_frame, i);
813 new_ent = copy_entity_own(old_ent, get_cur_frame_type());
814 set_entity_link(old_ent, new_ent);
817 /* visited is > than that of called graph. With this trick visited will
818 remain unchanged so that an outer walker, e.g., searching the call nodes
819 to inline, calling this inline will not visit the inlined nodes. */
820 set_irg_visited(current_ir_graph, get_irg_visited(current_ir_graph)-1);
822 /* -- Performing dead node elimination inlines the graph -- */
823 /* Copies the nodes to the obstack of current_ir_graph. Updates links to new
825 /* @@@ endless loops are not copied!! -- they should be, I think... */
826 irg_walk(get_irg_end(called_graph), copy_node_inline, copy_preds,
827 get_irg_frame_type(called_graph));
829 /* Repair called_graph */
830 set_irg_visited(called_graph, get_irg_visited(current_ir_graph));
831 set_irg_block_visited(called_graph, get_irg_block_visited(current_ir_graph));
832 set_Block_block_visited(get_irg_start_block(called_graph), 0);
834 /* -- Merge the end of the inlined procedure with the call site -- */
835 /* We will turn the old Call node into a Tuple with the following
838 0: Phi of all Memories of Return statements.
839 1: Jmp from new Block that merges the control flow from all exception
840 predecessors of the old end block.
841 2: Tuple of all arguments.
842 3: Phi of Exception memories.
843 In case the old Call directly branches to End on an exception we don't
844 need the block merging all exceptions nor the Phi of the exception
848 /* -- Precompute some values -- */
849 end_bl = get_new_node(get_irg_end_block(called_graph));
850 end = get_new_node(get_irg_end(called_graph));
851 arity = get_irn_arity(end_bl); /* arity = n_exc + n_ret */
852 n_res = get_method_n_ress(get_Call_type(call));
854 res_pred = (ir_node **) malloc (n_res * sizeof (ir_node *));
855 cf_pred = (ir_node **) malloc (arity * sizeof (ir_node *));
857 set_irg_current_block(current_ir_graph, post_bl); /* just to make sure */
859 /* -- archive keepalives -- */
860 irn_arity = get_irn_arity(end);
861 for (i = 0; i < irn_arity; i++)
862 add_End_keepalive(get_irg_end(current_ir_graph), get_irn_n(end, i));
864 /* The new end node will die. We need not free as the in array is on the obstack:
865 copy_node only generated 'D' arrays. */
867 /* -- Replace Return nodes by Jump nodes. -- */
869 for (i = 0; i < arity; i++) {
871 ret = get_irn_n(end_bl, i);
872 if (get_irn_op(ret) == op_Return) {
873 cf_pred[n_ret] = new_r_Jmp(current_ir_graph, get_nodes_block(ret));
877 set_irn_in(post_bl, n_ret, cf_pred);
879 /* -- Build a Tuple for all results of the method.
880 Add Phi node if there was more than one Return. -- */
881 turn_into_tuple(post_call, 4);
882 /* First the Memory-Phi */
884 for (i = 0; i < arity; i++) {
885 ret = get_irn_n(end_bl, i);
886 if (get_irn_op(ret) == op_Return) {
887 cf_pred[n_ret] = get_Return_mem(ret);
891 phi = new_Phi(n_ret, cf_pred, mode_M);
892 set_Tuple_pred(call, pn_Call_M_regular, phi);
893 /* Conserve Phi-list for further inlinings -- but might be optimized */
894 if (get_nodes_block(phi) == post_bl) {
895 set_irn_link(phi, get_irn_link(post_bl));
896 set_irn_link(post_bl, phi);
898 /* Now the real results */
900 for (j = 0; j < n_res; j++) {
902 for (i = 0; i < arity; i++) {
903 ret = get_irn_n(end_bl, i);
904 if (get_irn_op(ret) == op_Return) {
905 cf_pred[n_ret] = get_Return_res(ret, j);
910 phi = new_Phi(n_ret, cf_pred, get_irn_mode(cf_pred[0]));
914 /* Conserve Phi-list for further inlinings -- but might be optimized */
915 if (get_nodes_block(phi) == post_bl) {
916 set_irn_link(phi, get_irn_link(post_bl));
917 set_irn_link(post_bl, phi);
920 set_Tuple_pred(call, pn_Call_T_result, new_Tuple(n_res, res_pred));
922 set_Tuple_pred(call, pn_Call_T_result, new_Bad());
924 /* Finally the exception control flow.
925 We have two (three) possible situations:
926 First if the Call branches to an exception handler: We need to add a Phi node to
927 collect the memory containing the exception objects. Further we need
928 to add another block to get a correct representation of this Phi. To
929 this block we add a Jmp that resolves into the X output of the Call
930 when the Call is turned into a tuple.
931 Second the Call branches to End, the exception is not handled. Just
932 add all inlined exception branches to the End node.
933 Third: there is no Exception edge at all. Handle as case two. */
934 if (exc_handling == 0) {
936 for (i = 0; i < arity; i++) {
938 ret = get_irn_n(end_bl, i);
939 if (is_fragile_op(skip_Proj(ret)) || (get_irn_op(skip_Proj(ret)) == op_Raise)) {
940 cf_pred[n_exc] = ret;
945 new_Block(n_exc, cf_pred); /* watch it: current_block is changed! */
946 set_Tuple_pred(call, pn_Call_X_except, new_Jmp());
947 /* The Phi for the memories with the exception objects */
949 for (i = 0; i < arity; i++) {
951 ret = skip_Proj(get_irn_n(end_bl, i));
952 if (get_irn_op(ret) == op_Call) {
953 cf_pred[n_exc] = new_r_Proj(current_ir_graph, get_nodes_block(ret), ret, mode_M, 3);
955 } else if (is_fragile_op(ret)) {
956 /* We rely that all cfops have the memory output at the same position. */
957 cf_pred[n_exc] = new_r_Proj(current_ir_graph, get_nodes_block(ret), ret, mode_M, 0);
959 } else if (get_irn_op(ret) == op_Raise) {
960 cf_pred[n_exc] = new_r_Proj(current_ir_graph, get_nodes_block(ret), ret, mode_M, 1);
964 set_Tuple_pred(call, pn_Call_M_except, new_Phi(n_exc, cf_pred, mode_M));
966 set_Tuple_pred(call, pn_Call_X_except, new_Bad());
967 set_Tuple_pred(call, pn_Call_M_except, new_Bad());
970 ir_node *main_end_bl;
971 int main_end_bl_arity;
974 /* assert(exc_handling == 1 || no exceptions. ) */
976 for (i = 0; i < arity; i++) {
977 ir_node *ret = get_irn_n(end_bl, i);
979 if (is_fragile_op(skip_Proj(ret)) || (get_irn_op(skip_Proj(ret)) == op_Raise)) {
980 cf_pred[n_exc] = ret;
984 main_end_bl = get_irg_end_block(current_ir_graph);
985 main_end_bl_arity = get_irn_arity(main_end_bl);
986 end_preds = (ir_node **) malloc ((n_exc + main_end_bl_arity) * sizeof (ir_node *));
988 for (i = 0; i < main_end_bl_arity; ++i)
989 end_preds[i] = get_irn_n(main_end_bl, i);
990 for (i = 0; i < n_exc; ++i)
991 end_preds[main_end_bl_arity + i] = cf_pred[i];
992 set_irn_in(main_end_bl, n_exc + main_end_bl_arity, end_preds);
993 set_Tuple_pred(call, pn_Call_X_except, new_Bad());
994 set_Tuple_pred(call, pn_Call_M_except, new_Bad());
1000 #if 0 /* old. now better, correcter, faster implementation. */
1002 /* -- If the exception control flow from the inlined Call directly
1003 branched to the end block we now have the following control
1004 flow predecessor pattern: ProjX -> Tuple -> Jmp. We must
1005 remove the Jmp along with it's empty block and add Jmp's
1006 predecessors as predecessors of this end block. No problem if
1007 there is no exception, because then branches Bad to End which
1009 @@@ can't we know this beforehand: by getting the Proj(1) from
1010 the Call link list and checking whether it goes to Proj. */
1011 /* find the problematic predecessor of the end block. */
1012 end_bl = get_irg_end_block(current_ir_graph);
1013 for (i = 0; i < get_Block_n_cfgpreds(end_bl); i++) {
1014 cf_op = get_Block_cfgpred(end_bl, i);
1015 if (get_irn_op(cf_op) == op_Proj) {
1016 cf_op = get_Proj_pred(cf_op);
1017 if ((get_irn_op(cf_op) == op_Tuple) && (cf_op == call)) {
1018 /* There are unoptimized tuples from inlineing before when no exc */
1019 assert(get_Proj_proj(get_Block_cfgpred(end_bl, i)) == pn_Call_X_except);
1020 cf_op = get_Tuple_pred(cf_op, pn_Call_X_except);
1021 assert(get_irn_op(cf_op) == op_Jmp);
1027 if (i < get_Block_n_cfgpreds(end_bl)) {
1028 bl = get_nodes_block(cf_op);
1029 arity = get_Block_n_cfgpreds(end_bl) + get_Block_n_cfgpreds(bl) - 1;
1030 cf_pred = (ir_node **) malloc (arity * sizeof (ir_node *));
1031 for (j = 0; j < i; j++)
1032 cf_pred[j] = get_Block_cfgpred(end_bl, j);
1033 for (j = j; j < i + get_Block_n_cfgpreds(bl); j++)
1034 cf_pred[j] = get_Block_cfgpred(bl, j-i);
1035 for (j = j; j < arity; j++)
1036 cf_pred[j] = get_Block_cfgpred(end_bl, j-get_Block_n_cfgpreds(bl) +1);
1037 set_irn_in(end_bl, arity, cf_pred);
1039 /* Remove the exception pred from post-call Tuple. */
1040 set_Tuple_pred(call, pn_Call_X_except, new_Bad());
1045 /* -- Turn cse back on. -- */
1046 set_optimize(rem_opt);
1051 /********************************************************************/
1052 /* Apply inlineing to small methods. */
1053 /********************************************************************/
1055 /* It makes no sense to inline too many calls in one procedure. Anyways,
1056 I didn't get a version with NEW_ARR_F to run. */
1057 #define MAX_INLINE 1024
1060 * environment for inlining small irgs
1062 typedef struct _inline_env_t {
1064 ir_node *calls[MAX_INLINE];
1068 * Returns the irg called from a Call node. If the irg is not
1069 * known, NULL is returned.
1071 static ir_graph *get_call_called_irg(ir_node *call) {
1073 ir_graph *called_irg = NULL;
1075 assert(get_irn_op(call) == op_Call);
1077 addr = get_Call_ptr(call);
1078 if ((get_irn_op(addr) == op_SymConst) && (get_SymConst_kind (addr) == symconst_addr_ent)) {
1079 called_irg = get_entity_irg(get_SymConst_entity(addr));
1085 static void collect_calls(ir_node *call, void *env) {
1088 if (get_irn_op(call) != op_Call) return;
1090 addr = get_Call_ptr(call);
1092 if (get_irn_op(addr) == op_SymConst) {
1093 if (get_SymConst_kind(addr) == symconst_addr_ent) {
1094 ir_graph *called_irg = get_entity_irg(get_SymConst_entity(addr));
1095 inline_env_t *ienv = (inline_env_t *)env;
1096 if (called_irg && ienv->pos < MAX_INLINE) {
1097 /* The Call node calls a locally defined method. Remember to inline. */
1098 ienv->calls[ienv->pos++] = call;
1105 * Inlines all small methods at call sites where the called address comes
1106 * from a Const node that references the entity representing the called
1108 * The size argument is a rough measure for the code size of the method:
1109 * Methods where the obstack containing the firm graph is smaller than
1112 void inline_small_irgs(ir_graph *irg, int size) {
1114 ir_graph *rem = current_ir_graph;
1115 inline_env_t env /* = {0, NULL}*/;
1117 if (!(get_opt_optimize() && get_opt_inline())) return;
1119 current_ir_graph = irg;
1120 /* Handle graph state */
1121 assert(get_irg_phase_state(current_ir_graph) != phase_building);
1122 free_callee_info(current_ir_graph);
1124 /* Find Call nodes to inline.
1125 (We can not inline during a walk of the graph, as inlineing the same
1126 method several times changes the visited flag of the walked graph:
1127 after the first inlineing visited of the callee equals visited of
1128 the caller. With the next inlineing both are increased.) */
1130 irg_walk(get_irg_end(irg), NULL, collect_calls, &env);
1132 if ((env.pos > 0) && (env.pos < MAX_INLINE)) {
1133 /* There are calls to inline */
1134 collect_phiprojs(irg);
1135 for (i = 0; i < env.pos; i++) {
1137 callee = get_entity_irg(get_SymConst_entity(get_Call_ptr(env.calls[i])));
1138 if (((_obstack_memory_used(callee->obst) - obstack_room(callee->obst)) < size) ||
1139 (get_irg_inline_property(callee) == irg_inline_forced)) {
1140 inline_method(env.calls[i], callee);
1145 current_ir_graph = rem;
1149 * Environment for inlining irgs.
1152 int n_nodes; /**< Nodes in graph except Id, Tuple, Proj, Start, End */
1153 int n_nodes_orig; /**< for statistics */
1154 eset *call_nodes; /**< All call nodes in this graph */
1156 int n_call_nodes_orig; /**< for statistics */
1157 int n_callers; /**< Number of known graphs that call this graphs. */
1158 int n_callers_orig; /**< for statistics */
1161 static inline_irg_env *new_inline_irg_env(void) {
1162 inline_irg_env *env = malloc(sizeof(inline_irg_env));
1163 env->n_nodes = -2; /* uncount Start, End */
1164 env->n_nodes_orig = -2; /* uncount Start, End */
1165 env->call_nodes = eset_create();
1166 env->n_call_nodes = 0;
1167 env->n_call_nodes_orig = 0;
1169 env->n_callers_orig = 0;
1173 static void free_inline_irg_env(inline_irg_env *env) {
1174 eset_destroy(env->call_nodes);
1178 static void collect_calls2(ir_node *call, void *env) {
1179 inline_irg_env *x = (inline_irg_env *)env;
1180 ir_op *op = get_irn_op(call);
1183 /* count nodes in irg */
1184 if (op != op_Proj && op != op_Tuple && op != op_Sync) {
1189 if (op != op_Call) return;
1191 /* collect all call nodes */
1192 eset_insert(x->call_nodes, (void *)call);
1194 x->n_call_nodes_orig++;
1196 /* count all static callers */
1197 callee = get_call_called_irg(call);
1199 ((inline_irg_env *)get_irg_link(callee))->n_callers++;
1200 ((inline_irg_env *)get_irg_link(callee))->n_callers_orig++;
1204 INLINE static int is_leave(ir_graph *irg) {
1205 return (((inline_irg_env *)get_irg_link(irg))->n_call_nodes == 0);
1208 INLINE static int is_smaller(ir_graph *callee, int size) {
1209 return (((inline_irg_env *)get_irg_link(callee))->n_nodes < size);
1214 * Inlines small leave methods at call sites where the called address comes
1215 * from a Const node that references the entity representing the called
1217 * The size argument is a rough measure for the code size of the method:
1218 * Methods where the obstack containing the firm graph is smaller than
1221 void inline_leave_functions(int maxsize, int leavesize, int size) {
1222 inline_irg_env *env;
1223 int i, n_irgs = get_irp_n_irgs();
1224 ir_graph *rem = current_ir_graph;
1227 if (!(get_opt_optimize() && get_opt_inline())) return;
1229 /* extend all irgs by a temporary data structure for inlineing. */
1230 for (i = 0; i < n_irgs; ++i)
1231 set_irg_link(get_irp_irg(i), new_inline_irg_env());
1233 /* Precompute information in temporary data structure. */
1234 for (i = 0; i < n_irgs; ++i) {
1235 current_ir_graph = get_irp_irg(i);
1236 assert(get_irg_phase_state(current_ir_graph) != phase_building);
1237 free_callee_info(current_ir_graph);
1239 irg_walk(get_irg_end(current_ir_graph), NULL, collect_calls2,
1240 get_irg_link(current_ir_graph));
1243 /* -- and now inline. -- */
1245 /* Inline leaves recursively -- we might construct new leaves. */
1246 while (did_inline) {
1249 for (i = 0; i < n_irgs; ++i) {
1251 int phiproj_computed = 0;
1253 current_ir_graph = get_irp_irg(i);
1254 env = (inline_irg_env *)get_irg_link(current_ir_graph);
1256 for (call = eset_first(env->call_nodes); call; call = eset_next(env->call_nodes)) {
1257 if (get_irn_op(call) == op_Tuple) continue; /* We already inlined. */
1258 ir_graph *callee = get_call_called_irg(call);
1260 if (env->n_nodes > maxsize) continue; // break;
1262 if (callee && (is_leave(callee) && is_smaller(callee, leavesize))) {
1263 if (!phiproj_computed) {
1264 phiproj_computed = 1;
1265 collect_phiprojs(current_ir_graph);
1267 did_inline = inline_method(call, callee);
1270 /* Do some statistics */
1271 inline_irg_env *callee_env = (inline_irg_env *)get_irg_link(callee);
1272 env->n_call_nodes --;
1273 env->n_nodes += callee_env->n_nodes;
1274 callee_env->n_callers--;
1281 /* inline other small functions. */
1282 for (i = 0; i < n_irgs; ++i) {
1285 int phiproj_computed = 0;
1287 current_ir_graph = get_irp_irg(i);
1288 env = (inline_irg_env *)get_irg_link(current_ir_graph);
1290 /* we can not walk and change a set, nor remove from it.
1292 walkset = env->call_nodes;
1293 env->call_nodes = eset_create();
1294 for (call = eset_first(walkset); call; call = eset_next(walkset)) {
1295 if (get_irn_op(call) == op_Tuple) continue; /* We already inlined. */
1296 ir_graph *callee = get_call_called_irg(call);
1299 ((is_smaller(callee, size) && (env->n_nodes < maxsize)) || /* small function */
1300 (get_irg_inline_property(callee) == irg_inline_forced))) {
1301 if (!phiproj_computed) {
1302 phiproj_computed = 1;
1303 collect_phiprojs(current_ir_graph);
1305 if (inline_method(call, callee)) {
1306 inline_irg_env *callee_env = (inline_irg_env *)get_irg_link(callee);
1307 env->n_call_nodes--;
1308 eset_insert_all(env->call_nodes, callee_env->call_nodes); /* @@@ ??? This are the wrong nodes !? Not the copied ones. */
1309 env->n_call_nodes += callee_env->n_call_nodes;
1310 env->n_nodes += callee_env->n_nodes;
1311 callee_env->n_callers--;
1314 eset_insert(env->call_nodes, call);
1317 eset_destroy(walkset);
1320 for (i = 0; i < n_irgs; ++i) {
1321 current_ir_graph = get_irp_irg(i);
1323 env = (inline_irg_env *)get_irg_link(current_ir_graph);
1324 if ((env->n_call_nodes_orig != env->n_call_nodes) ||
1325 (env->n_callers_orig != env->n_callers))
1326 printf("Nodes:%3d ->%3d, calls:%3d ->%3d, callers:%3d ->%3d, -- %s\n",
1327 env->n_nodes_orig, env->n_nodes, env->n_call_nodes_orig, env->n_call_nodes,
1328 env->n_callers_orig, env->n_callers,
1329 get_entity_name(get_irg_entity(current_ir_graph)));
1331 free_inline_irg_env((inline_irg_env *)get_irg_link(current_ir_graph));
1334 current_ir_graph = rem;
1337 /*******************************************************************/
1338 /* Code Placement. Pins all floating nodes to a block where they */
1339 /* will be executed only if needed. */
1340 /*******************************************************************/
1343 * Find the earliest correct block for N. --- Place N into the
1344 * same Block as its dominance-deepest Input.
1347 place_floats_early(ir_node *n, pdeq *worklist)
1349 int i, start, irn_arity;
1351 /* we must not run into an infinite loop */
1352 assert (irn_not_visited(n));
1353 mark_irn_visited(n);
1355 /* Place floating nodes. */
1356 if (get_op_pinned(get_irn_op(n)) == op_pin_state_floats) {
1358 ir_node *b = new_Bad(); /* The block to place this node in */
1359 int bad_recursion = is_Bad(get_nodes_block(n));
1361 assert(get_irn_op(n) != op_Block);
1363 if ((get_irn_op(n) == op_Const) ||
1364 (get_irn_op(n) == op_SymConst) ||
1366 (get_irn_op(n) == op_Unknown)) {
1367 /* These nodes will not be placed by the loop below. */
1368 b = get_irg_start_block(current_ir_graph);
1372 /* find the block for this node. */
1373 irn_arity = get_irn_arity(n);
1374 for (i = 0; i < irn_arity; i++) {
1375 ir_node *dep = get_irn_n(n, i);
1378 if ((irn_not_visited(dep))
1379 && (get_op_pinned(get_irn_op(dep)) == op_pin_state_floats)) {
1380 place_floats_early(dep, worklist);
1384 * A node in the Bad block must stay in the bad block,
1385 * so don't compute a new block for it.
1390 /* Because all loops contain at least one op_pin_state_pinned node, now all
1391 our inputs are either op_pin_state_pinned or place_early has already
1392 been finished on them. We do not have any unfinished inputs! */
1393 dep_block = get_nodes_block(dep);
1394 if ((!is_Bad(dep_block)) &&
1395 (get_Block_dom_depth(dep_block) > depth)) {
1397 depth = get_Block_dom_depth(dep_block);
1399 /* Avoid that the node is placed in the Start block */
1400 if ((depth == 1) && (get_Block_dom_depth(get_nodes_block(n)) > 1)) {
1401 b = get_Block_cfg_out(get_irg_start_block(current_ir_graph), 0);
1402 assert(b != get_irg_start_block(current_ir_graph));
1406 set_nodes_block(n, b);
1409 /* Add predecessors of non floating nodes on worklist. */
1410 start = (get_irn_op(n) == op_Block) ? 0 : -1;
1411 irn_arity = get_irn_arity(n);
1412 for (i = start; i < irn_arity; i++) {
1413 ir_node *pred = get_irn_n(n, i);
1414 if (irn_not_visited(pred)) {
1415 pdeq_putr (worklist, pred);
1421 * Floating nodes form subgraphs that begin at nodes as Const, Load,
1422 * Start, Call and that end at op_pin_state_pinned nodes as Store, Call. Place_early
1423 * places all floating nodes reachable from its argument through floating
1424 * nodes and adds all beginnings at op_pin_state_pinned nodes to the worklist.
1426 static INLINE void place_early(pdeq* worklist) {
1428 inc_irg_visited(current_ir_graph);
1430 /* this inits the worklist */
1431 place_floats_early(get_irg_end(current_ir_graph), worklist);
1433 /* Work the content of the worklist. */
1434 while (!pdeq_empty (worklist)) {
1435 ir_node *n = pdeq_getl (worklist);
1436 if (irn_not_visited(n)) place_floats_early(n, worklist);
1439 set_irg_outs_inconsistent(current_ir_graph);
1440 current_ir_graph->op_pin_state_pinned = op_pin_state_pinned;
1444 /** deepest common dominance ancestor of DCA and CONSUMER of PRODUCER. */
1446 consumer_dom_dca (ir_node *dca, ir_node *consumer, ir_node *producer)
1448 ir_node *block = NULL;
1450 /* Compute the latest block into which we can place a node so that it is
1452 if (get_irn_op(consumer) == op_Phi) {
1453 /* our consumer is a Phi-node, the effective use is in all those
1454 blocks through which the Phi-node reaches producer */
1456 ir_node *phi_block = get_nodes_block(consumer);
1457 irn_arity = get_irn_arity(consumer);
1458 for (i = 0; i < irn_arity; i++) {
1459 if (get_irn_n(consumer, i) == producer) {
1460 block = get_nodes_block(get_Block_cfgpred(phi_block, i));
1464 assert(is_no_Block(consumer));
1465 block = get_nodes_block(consumer);
1468 /* Compute the deepest common ancestor of block and dca. */
1470 if (!dca) return block;
1471 while (get_Block_dom_depth(block) > get_Block_dom_depth(dca))
1472 block = get_Block_idom(block);
1473 while (get_Block_dom_depth(dca) > get_Block_dom_depth(block))
1474 dca = get_Block_idom(dca);
1475 while (block != dca)
1476 { block = get_Block_idom(block); dca = get_Block_idom(dca); }
1481 static INLINE int get_irn_loop_depth(ir_node *n) {
1482 return get_loop_depth(get_irn_loop(n));
1486 * Move n to a block with less loop depth than it's current block. The
1487 * new block must be dominated by early.
1490 move_out_of_loops (ir_node *n, ir_node *early)
1492 ir_node *best, *dca;
1496 /* Find the region deepest in the dominator tree dominating
1497 dca with the least loop nesting depth, but still dominated
1498 by our early placement. */
1499 dca = get_nodes_block(n);
1501 while (dca != early) {
1502 dca = get_Block_idom(dca);
1503 if (!dca) break; /* should we put assert(dca)? */
1504 if (get_irn_loop_depth(dca) < get_irn_loop_depth(best)) {
1508 if (best != get_nodes_block(n)) {
1510 printf("Moving out of loop: "); DDMN(n);
1511 printf(" Outermost block: "); DDMN(early);
1512 printf(" Best block: "); DDMN(best);
1513 printf(" Innermost block: "); DDMN(get_nodes_block(n));
1515 set_nodes_block(n, best);
1520 * Find the latest legal block for N and place N into the
1521 * `optimal' Block between the latest and earliest legal block.
1522 * The `optimal' block is the dominance-deepest block of those
1523 * with the least loop-nesting-depth. This places N out of as many
1524 * loops as possible and then makes it as control dependant as
1528 place_floats_late(ir_node *n, pdeq *worklist)
1533 assert (irn_not_visited(n)); /* no multiple placement */
1535 /* no need to place block nodes, control nodes are already placed. */
1536 if ((get_irn_op(n) != op_Block) &&
1538 (get_irn_mode(n) != mode_X)) {
1539 /* Remember the early placement of this block to move it
1540 out of loop no further than the early placement. */
1541 early = get_nodes_block(n);
1542 /* Assure that our users are all placed, except the Phi-nodes.
1543 --- Each data flow cycle contains at least one Phi-node. We
1544 have to break the `user has to be placed before the
1545 producer' dependence cycle and the Phi-nodes are the
1546 place to do so, because we need to base our placement on the
1547 final region of our users, which is OK with Phi-nodes, as they
1548 are op_pin_state_pinned, and they never have to be placed after a
1549 producer of one of their inputs in the same block anyway. */
1550 for (i = 0; i < get_irn_n_outs(n); i++) {
1551 ir_node *succ = get_irn_out(n, i);
1552 if (irn_not_visited(succ) && (get_irn_op(succ) != op_Phi))
1553 place_floats_late(succ, worklist);
1556 /* We have to determine the final block of this node... except for
1558 if ((get_op_pinned(get_irn_op(n)) == op_pin_state_floats) &&
1559 (get_irn_op(n) != op_Const) &&
1560 (get_irn_op(n) != op_SymConst)) {
1561 ir_node *dca = NULL; /* deepest common ancestor in the
1562 dominator tree of all nodes'
1563 blocks depending on us; our final
1564 placement has to dominate DCA. */
1565 for (i = 0; i < get_irn_n_outs(n); i++) {
1566 dca = consumer_dom_dca (dca, get_irn_out(n, i), n);
1568 set_nodes_block(n, dca);
1570 move_out_of_loops (n, early);
1574 mark_irn_visited(n);
1576 /* Add predecessors of all non-floating nodes on list. (Those of floating
1577 nodes are placeded already and therefore are marked.) */
1578 for (i = 0; i < get_irn_n_outs(n); i++) {
1579 if (irn_not_visited(get_irn_out(n, i))) {
1580 pdeq_putr (worklist, get_irn_out(n, i));
1585 static INLINE void place_late(pdeq* worklist) {
1587 inc_irg_visited(current_ir_graph);
1589 /* This fills the worklist initially. */
1590 place_floats_late(get_irg_start_block(current_ir_graph), worklist);
1591 /* And now empty the worklist again... */
1592 while (!pdeq_empty (worklist)) {
1593 ir_node *n = pdeq_getl (worklist);
1594 if (irn_not_visited(n)) place_floats_late(n, worklist);
1598 void place_code(ir_graph *irg) {
1600 ir_graph *rem = current_ir_graph;
1602 current_ir_graph = irg;
1604 if (!(get_opt_optimize() && get_opt_global_cse())) return;
1606 /* Handle graph state */
1607 assert(get_irg_phase_state(irg) != phase_building);
1608 if (get_irg_dom_state(irg) != dom_consistent)
1611 if (1 || get_irg_loopinfo_state(irg) != loopinfo_consistent) {
1612 free_loop_information(irg);
1613 construct_backedges(irg);
1616 /* Place all floating nodes as early as possible. This guarantees
1617 a legal code placement. */
1618 worklist = new_pdeq();
1619 place_early(worklist);
1621 /* place_early invalidates the outs, place_late needs them. */
1623 /* Now move the nodes down in the dominator tree. This reduces the
1624 unnecessary executions of the node. */
1625 place_late(worklist);
1627 set_irg_outs_inconsistent(current_ir_graph);
1628 set_irg_loopinfo_inconsistent(current_ir_graph);
1630 current_ir_graph = rem;
1635 /********************************************************************/
1636 /* Control flow optimization. */
1637 /* Removes Bad control flow predecessors and empty blocks. A block */
1638 /* is empty if it contains only a Jmp node. */
1639 /* Blocks can only be removed if they are not needed for the */
1640 /* semantics of Phi nodes. */
1641 /********************************************************************/
1644 * Removes Tuples from Block control flow predecessors.
1645 * Optimizes blocks with equivalent_node().
1646 * Replaces n by Bad if n is unreachable control flow.
1648 static void merge_blocks(ir_node *n, void *env) {
1650 set_irn_link(n, NULL);
1652 if (get_irn_op(n) == op_Block) {
1654 for (i = 0; i < get_Block_n_cfgpreds(n); i++)
1655 /* GL @@@ : is this possible? if (get_opt_normalize()) -- added, all tests go through.
1656 A different order of optimizations might cause problems. */
1657 if (get_opt_normalize())
1658 set_Block_cfgpred(n, i, skip_Tuple(get_Block_cfgpred(n, i)));
1659 } else if (get_opt_optimize() && (get_irn_mode(n) == mode_X)) {
1660 /* We will soon visit a block. Optimize it before visiting! */
1661 ir_node *b = get_nodes_block(n);
1662 ir_node *new_node = equivalent_node(b);
1663 while (irn_not_visited(b) && (!is_Bad(new_node)) && (new_node != b)) {
1664 /* We would have to run gigo if new is bad, so we
1665 promote it directly below. */
1666 assert(((b == new_node) ||
1667 get_opt_control_flow_straightening() ||
1668 get_opt_control_flow_weak_simplification()) &&
1669 ("strange flag setting"));
1670 exchange (b, new_node);
1672 new_node = equivalent_node(b);
1674 if (is_Bad(new_node) && get_opt_normalize()) exchange(n, new_Bad());
1679 * Collects all Phi nodes in link list of Block.
1680 * Marks all blocks "block_visited" if they contain a node other
1683 static void collect_nodes(ir_node *n, void *env) {
1684 if (is_no_Block(n)) {
1685 ir_node *b = get_nodes_block(n);
1687 if ((get_irn_op(n) == op_Phi)) {
1688 /* Collect Phi nodes to compact ins along with block's ins. */
1689 set_irn_link(n, get_irn_link(b));
1691 } else if ((get_irn_op(n) != op_Jmp) && !is_Bad(b)) { /* Check for non empty block. */
1692 mark_Block_block_visited(b);
1697 /** Returns true if pred is predecessor of block. */
1698 static int is_pred_of(ir_node *pred, ir_node *b) {
1700 for (i = 0; i < get_Block_n_cfgpreds(b); i++) {
1701 ir_node *b_pred = get_nodes_block(get_Block_cfgpred(b, i));
1702 if (b_pred == pred) return 1;
1707 static int test_whether_dispensable(ir_node *b, int pos) {
1708 int i, j, n_preds = 1;
1709 int dispensable = 1;
1710 ir_node *cfop = get_Block_cfgpred(b, pos);
1711 ir_node *pred = get_nodes_block(cfop);
1713 if (get_Block_block_visited(pred) + 1
1714 < get_irg_block_visited(current_ir_graph)) {
1715 if (!get_opt_optimize() || !get_opt_control_flow_strong_simplification()) {
1716 /* Mark block so that is will not be removed. */
1717 set_Block_block_visited(pred, get_irg_block_visited(current_ir_graph)-1);
1720 /* Seems to be empty. */
1721 if (!get_irn_link(b)) {
1722 /* There are no Phi nodes ==> dispensable. */
1723 n_preds = get_Block_n_cfgpreds(pred);
1725 /* b's pred blocks and pred's pred blocks must be pairwise disjunct.
1726 Work preds < pos as if they were already removed. */
1727 for (i = 0; i < pos; i++) {
1728 ir_node *b_pred = get_nodes_block(get_Block_cfgpred(b, i));
1729 if (get_Block_block_visited(b_pred) + 1
1730 < get_irg_block_visited(current_ir_graph)) {
1731 for (j = 0; j < get_Block_n_cfgpreds(b_pred); j++) {
1732 ir_node *b_pred_pred = get_nodes_block(get_Block_cfgpred(b_pred, j));
1733 if (is_pred_of(b_pred_pred, pred)) dispensable = 0;
1736 if (is_pred_of(b_pred, pred)) dispensable = 0;
1739 for (i = pos +1; i < get_Block_n_cfgpreds(b); i++) {
1740 ir_node *b_pred = get_nodes_block(get_Block_cfgpred(b, i));
1741 if (is_pred_of(b_pred, pred)) dispensable = 0;
1744 set_Block_block_visited(pred, get_irg_block_visited(current_ir_graph)-1);
1747 n_preds = get_Block_n_cfgpreds(pred);
1755 static void optimize_blocks(ir_node *b, void *env) {
1756 int i, j, k, max_preds, n_preds;
1757 ir_node *pred, *phi;
1760 /* Count the number of predecessor if this block is merged with pred blocks
1763 for (i = 0; i < get_Block_n_cfgpreds(b); i++) {
1764 max_preds += test_whether_dispensable(b, i);
1766 in = (ir_node **) malloc(max_preds * sizeof(ir_node *));
1769 printf(" working on "); DDMN(b);
1770 for (i = 0; i < get_Block_n_cfgpreds(b); i++) {
1771 pred = get_nodes_block(get_Block_cfgpred(b, i));
1772 if (is_Bad(get_Block_cfgpred(b, i))) {
1773 printf(" removing Bad %i\n ", i);
1774 } else if (get_Block_block_visited(pred) +1
1775 < get_irg_block_visited(current_ir_graph)) {
1776 printf(" removing pred %i ", i); DDMN(pred);
1777 } else { printf(" Nothing to do for "); DDMN(pred); }
1779 * end Debug output -*/
1781 /*- Fix the Phi nodes -*/
1782 phi = get_irn_link(b);
1784 assert(get_irn_op(phi) == op_Phi);
1785 /* Find the new predecessors for the Phi */
1787 for (i = 0; i < get_Block_n_cfgpreds(b); i++) {
1788 pred = get_nodes_block(get_Block_cfgpred(b, i));
1789 if (is_Bad(get_Block_cfgpred(b, i))) {
1791 } else if (get_Block_block_visited(pred) +1
1792 < get_irg_block_visited(current_ir_graph)) {
1793 /* It's an empty block and not yet visited. */
1794 ir_node *phi_pred = get_Phi_pred(phi, i);
1795 for (j = 0; j < get_Block_n_cfgpreds(pred); j++) {
1796 if (get_nodes_block(phi_pred) == pred) {
1797 assert(get_irn_op(phi_pred) == op_Phi); /* Block is empty!! */
1798 in[n_preds] = get_Phi_pred(phi_pred, j);
1800 in[n_preds] = phi_pred;
1804 /* The Phi_pred node is replaced now if it is a Phi.
1805 In Schleifen kann offenbar der entfernte Phi Knoten legal verwendet werden.
1806 Daher muss der Phiknoten durch den neuen ersetzt werden.
1807 Weiter muss der alte Phiknoten entfernt werden (durch ersetzen oder
1808 durch einen Bad) damit er aus den keep_alive verschwinden kann.
1809 Man sollte also, falls keine Schleife vorliegt, exchange mit new_Bad
1811 if (get_nodes_block(phi_pred) == pred) {
1812 /* remove the Phi as it might be kept alive. Further there
1813 might be other users. */
1814 exchange(phi_pred, phi); /* geht, ist aber doch semantisch falsch! Warum?? */
1817 in[n_preds] = get_Phi_pred(phi, i);
1822 set_irn_in(phi, n_preds, in);
1824 phi = get_irn_link(phi);
1828 This happens only if merge between loop backedge and single loop entry. -*/
1829 for (k = 0; k < get_Block_n_cfgpreds(b); k++) {
1830 pred = get_nodes_block(get_Block_cfgpred(b, k));
1831 if (get_Block_block_visited(pred)+1 < get_irg_block_visited(current_ir_graph)) {
1832 phi = get_irn_link(pred);
1834 if (get_irn_op(phi) == op_Phi) {
1835 set_nodes_block(phi, b);
1838 for (i = 0; i < k; i++) {
1839 pred = get_nodes_block(get_Block_cfgpred(b, i));
1840 if (is_Bad(get_Block_cfgpred(b, i))) {
1842 } else if (get_Block_block_visited(pred) +1
1843 < get_irg_block_visited(current_ir_graph)) {
1844 /* It's an empty block and not yet visited. */
1845 for (j = 0; j < get_Block_n_cfgpreds(pred); j++) {
1846 /* @@@ Hier brauche ich Schleifeninformation!!! Kontrollflusskante
1847 muss Rueckwaertskante sein! (An allen vier in[n_preds] = phi
1848 Anweisungen.) Trotzdem tuts bisher!! */
1857 for (i = 0; i < get_Phi_n_preds(phi); i++) {
1858 in[n_preds] = get_Phi_pred(phi, i);
1861 for (i = k+1; i < get_Block_n_cfgpreds(b); i++) {
1862 pred = get_nodes_block(get_Block_cfgpred(b, i));
1863 if (is_Bad(get_Block_cfgpred(b, i))) {
1865 } else if (get_Block_block_visited(pred) +1
1866 < get_irg_block_visited(current_ir_graph)) {
1867 /* It's an empty block and not yet visited. */
1868 for (j = 0; j < get_Block_n_cfgpreds(pred); j++) {
1877 set_irn_in(phi, n_preds, in);
1879 phi = get_irn_link(phi);
1884 /*- Fix the block -*/
1886 for (i = 0; i < get_Block_n_cfgpreds(b); i++) {
1887 pred = get_nodes_block(get_Block_cfgpred(b, i));
1888 if (is_Bad(get_Block_cfgpred(b, i))) {
1890 } else if (get_Block_block_visited(pred) +1
1891 < get_irg_block_visited(current_ir_graph)) {
1892 /* It's an empty block and not yet visited. */
1893 assert(get_Block_n_cfgpreds(b) > 1);
1894 /* Else it should be optimized by equivalent_node. */
1895 for (j = 0; j < get_Block_n_cfgpreds(pred); j++) {
1896 in[n_preds] = get_Block_cfgpred(pred, j);
1899 /* Remove block as it might be kept alive. */
1900 exchange(pred, b/*new_Bad()*/);
1902 in[n_preds] = get_Block_cfgpred(b, i);
1906 set_irn_in(b, n_preds, in);
1910 void optimize_cf(ir_graph *irg) {
1913 ir_node *end = get_irg_end(irg);
1914 ir_graph *rem = current_ir_graph;
1915 current_ir_graph = irg;
1917 /* Handle graph state */
1918 assert(get_irg_phase_state(irg) != phase_building);
1919 if (get_irg_outs_state(current_ir_graph) == outs_consistent)
1920 set_irg_outs_inconsistent(current_ir_graph);
1921 if (get_irg_dom_state(current_ir_graph) == dom_consistent)
1922 set_irg_dom_inconsistent(current_ir_graph);
1924 /* Use block visited flag to mark non-empty blocks. */
1925 inc_irg_block_visited(irg);
1926 irg_walk(end, merge_blocks, collect_nodes, NULL);
1928 /* Optimize the standard code. */
1929 irg_block_walk(get_irg_end_block(irg), optimize_blocks, NULL, NULL);
1931 /* Walk all keep alives, optimize them if block, add to new in-array
1932 for end if useful. */
1933 in = NEW_ARR_F (ir_node *, 1);
1934 in[0] = get_nodes_block(end);
1935 inc_irg_visited(current_ir_graph);
1936 for(i = 0; i < get_End_n_keepalives(end); i++) {
1937 ir_node *ka = get_End_keepalive(end, i);
1938 if (irn_not_visited(ka)) {
1939 if ((get_irn_op(ka) == op_Block) && Block_not_block_visited(ka)) {
1940 set_irg_block_visited(current_ir_graph, /* Don't walk all the way to Start. */
1941 get_irg_block_visited(current_ir_graph)-1);
1942 irg_block_walk(ka, optimize_blocks, NULL, NULL);
1943 mark_irn_visited(ka);
1944 ARR_APP1 (ir_node *, in, ka);
1945 } else if (get_irn_op(ka) == op_Phi) {
1946 mark_irn_visited(ka);
1947 ARR_APP1 (ir_node *, in, ka);
1951 /* DEL_ARR_F(end->in); GL @@@ tut nicht ! */
1954 current_ir_graph = rem;
1959 * Called by walker of remove_critical_cf_edges().
1961 * Place an empty block to an edge between a blocks of multiple
1962 * predecessors and a block of multiple successors.
1965 * @param env Environment of walker. This field is unused and has
1968 static void walk_critical_cf_edges(ir_node *n, void *env) {
1970 ir_node *pre, *block, **in, *jmp;
1972 /* Block has multiple predecessors */
1973 if ((op_Block == get_irn_op(n)) &&
1974 (get_irn_arity(n) > 1)) {
1975 arity = get_irn_arity(n);
1977 if (n == get_irg_end_block(current_ir_graph))
1978 return; /* No use to add a block here. */
1980 for (i=0; i<arity; i++) {
1981 pre = get_irn_n(n, i);
1982 /* Predecessor has multiple successors. Insert new flow edge */
1983 if ((NULL != pre) &&
1984 (op_Proj == get_irn_op(pre)) &&
1985 op_Raise != get_irn_op(skip_Proj(pre))) {
1987 /* set predecessor array for new block */
1988 in = NEW_ARR_D (ir_node *, current_ir_graph->obst, 1);
1989 /* set predecessor of new block */
1991 block = new_Block(1, in);
1992 /* insert new jmp node to new block */
1993 set_cur_block(block);
1996 /* set successor of new block */
1997 set_irn_n(n, i, jmp);
1999 } /* predecessor has multiple successors */
2000 } /* for all predecessors */
2001 } /* n is a block */
2004 void remove_critical_cf_edges(ir_graph *irg) {
2005 if (get_opt_critical_edges())
2006 irg_walk_graph(irg, NULL, walk_critical_cf_edges, NULL);