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.
23 # include "irnode_t.h"
24 # include "irgraph_t.h"
32 # include "pdeq.h" /* Fuer code placement */
35 # include "irbackedge_t.h"
36 # include "irflag_t.h"
37 # include "firmstat.h"
39 /* Defined in iropt.c */
40 pset *new_identities (void);
41 void del_identities (pset *value_table);
42 void add_identities (pset *value_table, ir_node *node);
44 /*------------------------------------------------------------------*/
45 /* apply optimizations of iropt to all nodes. */
46 /*------------------------------------------------------------------*/
48 static void init_link (ir_node *n, void *env) {
49 set_irn_link(n, NULL);
52 #if 0 /* Old version. Avoids Ids.
53 This is not necessary: we do a postwalk, and get_irn_n
54 removes ids anyways. So it's much cheaper to call the
55 optimization less often and use the exchange() algorithm. */
57 optimize_in_place_wrapper (ir_node *n, void *env) {
59 ir_node *optimized, *old;
61 irn_arity = get_irn_arity(n);
62 for (i = 0; i < irn_arity; i++) {
63 /* get_irn_n skips Id nodes, so comparison old != optimized does not
64 show all optimizations. Therefore always set new predecessor. */
65 old = get_irn_intra_n(n, i);
66 optimized = optimize_in_place_2(old);
67 set_irn_n(n, i, optimized);
70 if (get_irn_op(n) == op_Block) {
71 optimized = optimize_in_place_2(n);
72 if (optimized != n) exchange (n, optimized);
77 optimize_in_place_wrapper (ir_node *n, void *env) {
78 ir_node *optimized = optimize_in_place_2(n);
79 if (optimized != n) exchange (n, optimized);
86 local_optimize_graph (ir_graph *irg) {
87 ir_graph *rem = current_ir_graph;
88 current_ir_graph = irg;
90 /* Handle graph state */
91 assert(get_irg_phase_state(irg) != phase_building);
92 if (get_opt_global_cse())
93 set_irg_pinned(current_ir_graph, 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);
99 /* Clean the value_table in irg for the cse. */
100 del_identities(irg->value_table);
101 irg->value_table = new_identities();
103 /* walk over the graph */
104 irg_walk(irg->end, init_link, optimize_in_place_wrapper, NULL);
106 current_ir_graph = rem;
109 /*------------------------------------------------------------------*/
110 /* Routines for dead node elimination / copying garbage collection */
111 /* of the obstack. */
112 /*------------------------------------------------------------------*/
115 * Remember the new node in the old node by using a field all nodes have.
118 set_new_node (ir_node *old, ir_node *new)
124 * Get this new node, before the old node is forgotton.
126 static INLINE ir_node *
127 get_new_node (ir_node * n)
133 * We use the block_visited flag to mark that we have computed the
134 * number of useful predecessors for this block.
135 * Further we encode the new arity in this flag in the old blocks.
136 * Remembering the arity is useful, as it saves a lot of pointer
137 * accesses. This function is called for all Phi and Block nodes
141 compute_new_arity(ir_node *b) {
142 int i, res, irn_arity;
145 irg_v = get_irg_block_visited(current_ir_graph);
146 block_v = get_Block_block_visited(b);
147 if (block_v >= irg_v) {
148 /* we computed the number of preds for this block and saved it in the
150 return block_v - irg_v;
152 /* compute the number of good predecessors */
153 res = irn_arity = get_irn_arity(b);
154 for (i = 0; i < irn_arity; i++)
155 if (get_irn_opcode(get_irn_n(b, i)) == iro_Bad) res--;
156 /* save it in the flag. */
157 set_Block_block_visited(b, irg_v + res);
162 /* TODO: add an ir_op operation */
163 static INLINE void new_backedge_info(ir_node *n) {
164 switch(get_irn_opcode(n)) {
166 n->attr.block.cg_backedge = NULL;
167 n->attr.block.backedge = new_backedge_arr(current_ir_graph->obst, get_irn_arity(n));
170 n->attr.phi_backedge = new_backedge_arr(current_ir_graph->obst, get_irn_arity(n));
173 n->attr.filter.backedge = new_backedge_arr(current_ir_graph->obst, get_irn_arity(n));
180 * Copies the node to the new obstack. The Ins of the new node point to
181 * the predecessors on the old obstack. For block/phi nodes not all
182 * predecessors might be copied. n->link points to the new node.
183 * For Phi and Block nodes the function allocates in-arrays with an arity
184 * only for useful predecessors. The arity is determined by counting
185 * the non-bad predecessors of the block.
188 copy_node (ir_node *n, void *env) {
192 /* The end node looses it's flexible in array. This doesn't matter,
193 as dead node elimination builds End by hand, inlineing doesn't use
195 /* assert(n->op == op_End || ((_ARR_DESCR(n->in))->cookie != ARR_F_MAGIC)); */
197 if (get_irn_opcode(n) == iro_Block) {
199 new_arity = compute_new_arity(n);
200 n->attr.block.graph_arr = NULL;
202 block = get_nodes_Block(n);
203 if (get_irn_opcode(n) == iro_Phi) {
204 new_arity = compute_new_arity(block);
206 new_arity = get_irn_arity(n);
209 nn = new_ir_node(get_irn_dbg_info(n),
216 /* Copy the attributes. These might point to additional data. If this
217 was allocated on the old obstack the pointers now are dangling. This
218 frees e.g. the memory of the graph_arr allocated in new_immBlock. */
220 new_backedge_info(nn);
223 /* printf("\n old node: "); DDMSG2(n);
224 printf(" new node: "); DDMSG2(nn); */
229 * Copies new predecessors of old node to new node remembered in link.
230 * Spare the Bad predecessors of Phi and Block nodes.
233 copy_preds (ir_node *n, void *env) {
237 nn = get_new_node(n);
239 /* printf("\n old node: "); DDMSG2(n);
240 printf(" new node: "); DDMSG2(nn);
241 printf(" arities: old: %d, new: %d\n", get_irn_arity(n), get_irn_arity(nn)); */
243 if (get_irn_opcode(n) == iro_Block) {
244 /* Don't copy Bad nodes. */
246 irn_arity = get_irn_arity(n);
247 for (i = 0; i < irn_arity; i++)
248 if (get_irn_opcode(get_irn_n(n, i)) != iro_Bad) {
249 set_irn_n (nn, j, get_new_node(get_irn_n(n, i)));
250 /*if (is_backedge(n, i)) set_backedge(nn, j);*/
253 /* repair the block visited flag from above misuse. Repair it in both
254 graphs so that the old one can still be used. */
255 set_Block_block_visited(nn, 0);
256 set_Block_block_visited(n, 0);
257 /* Local optimization could not merge two subsequent blocks if
258 in array contained Bads. Now it's possible.
259 We don't call optimize_in_place as it requires
260 that the fields in ir_graph are set properly. */
261 if ((get_opt_control_flow_straightening()) &&
262 (get_Block_n_cfgpreds(nn) == 1) &&
263 (get_irn_op(get_Block_cfgpred(nn, 0)) == op_Jmp))
264 exchange(nn, get_nodes_Block(get_Block_cfgpred(nn, 0)));
265 } else if (get_irn_opcode(n) == iro_Phi) {
266 /* Don't copy node if corresponding predecessor in block is Bad.
267 The Block itself should not be Bad. */
268 block = get_nodes_Block(n);
269 set_irn_n (nn, -1, get_new_node(block));
271 irn_arity = get_irn_arity(n);
272 for (i = 0; i < irn_arity; i++)
273 if (get_irn_opcode(get_irn_n(block, i)) != iro_Bad) {
274 set_irn_n (nn, j, get_new_node(get_irn_n(n, i)));
275 /*if (is_backedge(n, i)) set_backedge(nn, j);*/
278 /* If the pre walker reached this Phi after the post walker visited the
279 block block_visited is > 0. */
280 set_Block_block_visited(get_nodes_Block(n), 0);
281 /* Compacting the Phi's ins might generate Phis with only one
283 if (get_irn_arity(n) == 1)
284 exchange(n, get_irn_n(n, 0));
286 irn_arity = get_irn_arity(n);
287 for (i = -1; i < irn_arity; i++)
288 set_irn_n (nn, i, get_new_node(get_irn_n(n, i)));
290 /* Now the new node is complete. We can add it to the hash table for cse.
291 @@@ inlinening aborts if we identify End. Why? */
292 if(get_irn_op(nn) != op_End)
293 add_identities (current_ir_graph->value_table, nn);
297 * Copies the graph recursively, compacts the keepalive of the end node.
301 ir_node *oe, *ne; /* old end, new end */
302 ir_node *ka; /* keep alive */
305 oe = get_irg_end(current_ir_graph);
306 /* copy the end node by hand, allocate dynamic in array! */
307 ne = new_ir_node(get_irn_dbg_info(oe),
314 /* Copy the attributes. Well, there might be some in the future... */
316 set_new_node(oe, ne);
318 /* copy the live nodes */
319 irg_walk(get_nodes_Block(oe), copy_node, copy_preds, NULL);
320 /* copy_preds for the end node ... */
321 set_nodes_Block(ne, get_new_node(get_nodes_Block(oe)));
323 /*- ... and now the keep alives. -*/
324 /* First pick the not marked block nodes and walk them. We must pick these
325 first as else we will oversee blocks reachable from Phis. */
326 irn_arity = get_irn_arity(oe);
327 for (i = 0; i < irn_arity; i++) {
328 ka = get_irn_intra_n(oe, i);
329 if ((get_irn_op(ka) == op_Block) &&
330 (get_irn_visited(ka) < get_irg_visited(current_ir_graph))) {
331 /* We must keep the block alive and copy everything reachable */
332 set_irg_visited(current_ir_graph, get_irg_visited(current_ir_graph)-1);
333 irg_walk(ka, copy_node, copy_preds, NULL);
334 add_End_keepalive(ne, get_new_node(ka));
338 /* Now pick the Phis. Here we will keep all! */
339 irn_arity = get_irn_arity(oe);
340 for (i = 0; i < irn_arity; i++) {
341 ka = get_irn_intra_n(oe, i);
342 if ((get_irn_op(ka) == op_Phi)) {
343 if (get_irn_visited(ka) < get_irg_visited(current_ir_graph)) {
344 /* We didn't copy the Phi yet. */
345 set_irg_visited(current_ir_graph, get_irg_visited(current_ir_graph)-1);
346 irg_walk(ka, copy_node, copy_preds, NULL);
348 add_End_keepalive(ne, get_new_node(ka));
354 * Copies the graph reachable from current_ir_graph->end to the obstack
355 * in current_ir_graph and fixes the environment.
356 * Then fixes the fields in current_ir_graph containing nodes of the
360 copy_graph_env (void) {
362 /* Not all nodes remembered in current_ir_graph might be reachable
363 from the end node. Assure their link is set to NULL, so that
364 we can test whether new nodes have been computed. */
365 set_irn_link(get_irg_frame (current_ir_graph), NULL);
366 set_irn_link(get_irg_globals(current_ir_graph), NULL);
367 set_irn_link(get_irg_args (current_ir_graph), NULL);
369 /* we use the block walk flag for removing Bads from Blocks ins. */
370 inc_irg_block_visited(current_ir_graph);
375 /* fix the fields in current_ir_graph */
376 old_end = get_irg_end(current_ir_graph);
377 set_irg_end (current_ir_graph, get_new_node(old_end));
378 set_irg_end_except (current_ir_graph, get_irg_end(current_ir_graph));
379 set_irg_end_reg (current_ir_graph, get_irg_end(current_ir_graph));
381 set_irg_end_block (current_ir_graph, get_new_node(get_irg_end_block(current_ir_graph)));
382 if (get_irn_link(get_irg_frame(current_ir_graph)) == NULL) {
383 copy_node (get_irg_frame(current_ir_graph), NULL);
384 copy_preds(get_irg_frame(current_ir_graph), NULL);
386 if (get_irn_link(get_irg_globals(current_ir_graph)) == NULL) {
387 copy_node (get_irg_globals(current_ir_graph), NULL);
388 copy_preds(get_irg_globals(current_ir_graph), NULL);
390 if (get_irn_link(get_irg_args(current_ir_graph)) == NULL) {
391 copy_node (get_irg_args(current_ir_graph), NULL);
392 copy_preds(get_irg_args(current_ir_graph), NULL);
394 set_irg_start (current_ir_graph, get_new_node(get_irg_start(current_ir_graph)));
396 set_irg_start_block(current_ir_graph,
397 get_new_node(get_irg_start_block(current_ir_graph)));
398 set_irg_frame (current_ir_graph, get_new_node(get_irg_frame(current_ir_graph)));
399 set_irg_globals(current_ir_graph, get_new_node(get_irg_globals(current_ir_graph)));
400 set_irg_args (current_ir_graph, get_new_node(get_irg_args(current_ir_graph)));
401 if (get_irn_link(get_irg_bad(current_ir_graph)) == NULL) {
402 copy_node(get_irg_bad(current_ir_graph), NULL);
403 copy_preds(get_irg_bad(current_ir_graph), NULL);
405 set_irg_bad(current_ir_graph, get_new_node(get_irg_bad(current_ir_graph)));
406 /* GL removed: we need unknown with mode for analyses.
407 if (get_irn_link(get_irg_unknown(current_ir_graph)) == NULL) {
408 copy_node(get_irg_unknown(current_ir_graph), NULL);
409 copy_preds(get_irg_unknown(current_ir_graph), NULL);
411 set_irg_unknown(current_ir_graph, get_new_node(get_irg_unknown(current_ir_graph)));
416 * Copies all reachable nodes to a new obstack. Removes bad inputs
417 * from block nodes and the corresponding inputs from Phi nodes.
418 * Merges single exit blocks with single entry blocks and removes
420 * Adds all new nodes to a new hash table for cse. Does not
421 * perform cse, so the hash table might contain common subexpressions.
423 /* Amroq call this emigrate() */
425 dead_node_elimination(ir_graph *irg) {
427 int rem_ipview = interprocedural_view;
428 struct obstack *graveyard_obst = NULL;
429 struct obstack *rebirth_obst = NULL;
431 stat_dead_node_elim_start(irg);
433 /* Remember external state of current_ir_graph. */
434 rem = current_ir_graph;
435 current_ir_graph = irg;
436 interprocedural_view = 0;
438 /* Handle graph state */
439 assert(get_irg_phase_state(current_ir_graph) != phase_building);
440 assert(get_irg_callee_info_state(current_ir_graph) == irg_callee_info_none);
441 free_outs(current_ir_graph);
443 /* @@@ so far we loose loops when copying */
444 free_loop_information(current_ir_graph);
446 if (get_opt_optimize() && get_opt_dead_node_elimination()) {
448 /* A quiet place, where the old obstack can rest in peace,
449 until it will be cremated. */
450 graveyard_obst = irg->obst;
452 /* A new obstack, where the reachable nodes will be copied to. */
453 rebirth_obst = (struct obstack *) xmalloc (sizeof (struct obstack));
454 current_ir_graph->obst = rebirth_obst;
455 obstack_init (current_ir_graph->obst);
457 /* We also need a new hash table for cse */
458 del_identities (irg->value_table);
459 irg->value_table = new_identities ();
461 /* Copy the graph from the old to the new obstack */
464 /* Free memory from old unoptimized obstack */
465 obstack_free(graveyard_obst, 0); /* First empty the obstack ... */
466 xfree (graveyard_obst); /* ... then free it. */
469 stat_dead_node_elim_stop(irg);
471 current_ir_graph = rem;
472 interprocedural_view = rem_ipview;
476 * Relink bad predeseccors of a block and store the old in array to the
477 * link field. This function is called by relink_bad_predecessors().
478 * The array of link field starts with the block operand at position 0.
479 * If block has bad predecessors, create a new in array without bad preds.
480 * Otherwise let in array untouched.
482 static void relink_bad_block_predecessors(ir_node *n, void *env) {
483 ir_node **new_in, *irn;
484 int i, new_irn_n, old_irn_arity, new_irn_arity = 0;
486 /* if link field of block is NULL, look for bad predecessors otherwise
487 this is allready done */
488 if (get_irn_op(n) == op_Block &&
489 get_irn_link(n) == NULL) {
491 /* save old predecessors in link field (position 0 is the block operand)*/
492 set_irn_link(n, (void *)get_irn_in(n));
494 /* count predecessors without bad nodes */
495 old_irn_arity = get_irn_arity(n);
496 for (i = 0; i < old_irn_arity; i++)
497 if (!is_Bad(get_irn_n(n, i))) new_irn_arity++;
499 /* arity changing: set new predecessors without bad nodes */
500 if (new_irn_arity < old_irn_arity) {
501 /* get new predecessor array without Block predecessor */
502 new_in = NEW_ARR_D (ir_node *, current_ir_graph->obst, (new_irn_arity+1));
504 /* set new predeseccors in array */
507 for (i = 1; i < old_irn_arity; i++) {
508 irn = get_irn_n(n, i);
509 if (!is_Bad(irn)) new_in[new_irn_n++] = irn;
512 } /* ir node has bad predecessors */
514 } /* Block is not relinked */
518 * Relinks Bad predecesors from Bocks and Phis called by walker
519 * remove_bad_predecesors(). If n is a Block, call
520 * relink_bad_block_redecessors(). If n is a Phinode, call also the relinking
521 * function of Phi's Block. If this block has bad predecessors, relink preds
524 static void relink_bad_predecessors(ir_node *n, void *env) {
525 ir_node *block, **old_in;
526 int i, old_irn_arity, new_irn_arity;
528 /* relink bad predeseccors of a block */
529 if (get_irn_op(n) == op_Block)
530 relink_bad_block_predecessors(n, env);
532 /* If Phi node relink its block and its predecessors */
533 if (get_irn_op(n) == op_Phi) {
535 /* Relink predeseccors of phi's block */
536 block = get_nodes_Block(n);
537 if (get_irn_link(block) == NULL)
538 relink_bad_block_predecessors(block, env);
540 old_in = (ir_node **)get_irn_link(block); /* Of Phi's Block */
541 old_irn_arity = ARR_LEN(old_in);
543 /* Relink Phi predeseccors if count of predeseccors changed */
544 if (old_irn_arity != ARR_LEN(get_irn_in(block))) {
545 /* set new predeseccors in array
546 n->in[0] remains the same block */
548 for(i = 1; i < old_irn_arity; i++)
549 if (!is_Bad((ir_node *)old_in[i])) n->in[new_irn_arity++] = n->in[i];
551 ARR_SETLEN(ir_node *, n->in, new_irn_arity);
554 } /* n is a Phi node */
558 * Removes Bad Bad predecesors from Blocks and the corresponding
559 * inputs to Phi nodes as in dead_node_elimination but without
561 * On walking up set the link field to NULL, on walking down call
562 * relink_bad_predecessors() (This function stores the old in array
563 * to the link field and sets a new in array if arity of predecessors
566 void remove_bad_predecessors(ir_graph *irg) {
567 irg_walk_graph(irg, init_link, relink_bad_predecessors, NULL);
571 /*--------------------------------------------------------------------*/
572 /* Funcionality for inlining */
573 /*--------------------------------------------------------------------*/
576 * Copy node for inlineing. Updates attributes that change when
577 * inlineing but not for dead node elimination.
579 * Copies the node by calling copy_node and then updates the entity if
580 * it's a local one. env must be a pointer of the frame type of the
581 * inlined procedure. The new entities must be in the link field of
585 copy_node_inline (ir_node *n, void *env) {
587 type *frame_tp = (type *)env;
590 if (get_irn_op(n) == op_Sel) {
591 new = get_new_node (n);
592 assert(get_irn_op(new) == op_Sel);
593 if (get_entity_owner(get_Sel_entity(n)) == frame_tp) {
594 set_Sel_entity(new, get_entity_link(get_Sel_entity(n)));
596 } else if (get_irn_op(n) == op_Block) {
597 new = get_new_node (n);
598 new->attr.block.irg = current_ir_graph;
602 static void find_addr(ir_node *node, void *env)
604 if (get_irn_opcode(node) == iro_Proj) {
605 if (get_Proj_proj(node) == pn_Start_P_value_arg_base)
611 * currently, we cannot inline two cases:
612 * - call with compound arguments
613 * - graphs that take the address of a parameter
615 * check these condition here
617 static int can_inline(ir_node *call, ir_graph *called_graph)
619 type *call_type = get_Call_type(call);
620 int params, ress, i, res;
622 assert(is_method_type(call_type));
624 params = get_method_n_params(call_type);
625 ress = get_method_n_ress(call_type);
628 for (i = 0; i < params; ++i) {
629 type *p_type = get_method_param_type(call_type, i);
631 if (is_compound_type(p_type))
636 for (i = 0; i < ress; ++i) {
637 type *r_type = get_method_res_type(call_type, i);
639 if (is_compound_type(r_type))
644 irg_walk_graph(called_graph, find_addr, NULL, &res);
649 int inline_method(ir_node *call, ir_graph *called_graph) {
651 ir_node *post_call, *post_bl;
653 ir_node *end, *end_bl;
657 int arity, n_ret, n_exc, n_res, i, j, rem_opt, irn_arity;
660 irg_inline_property prop = get_irg_inline_property(called_graph);
662 if ( (prop != irg_inline_forced) && (!get_opt_optimize() || !get_opt_inline() ||
663 (prop == irg_inline_forbidden))) return 0;
667 * currently, we cannot inline two cases:
668 * - call with compound arguments
669 * - graphs that take the address of a parameter
671 if (! can_inline(call, called_graph))
674 /* -- Turn off optimizations, this can cause problems when allocating new nodes. -- */
675 rem_opt = get_opt_optimize();
678 /* Handle graph state */
679 assert(get_irg_phase_state(current_ir_graph) != phase_building);
680 assert(get_irg_pinned(current_ir_graph) == pinned);
681 assert(get_irg_pinned(called_graph) == pinned);
682 if (get_irg_outs_state(current_ir_graph) == outs_consistent)
683 set_irg_outs_inconsistent(current_ir_graph);
685 /* -- Check preconditions -- */
686 assert(get_irn_op(call) == op_Call);
687 /* @@@ does not work for InterfaceIII.java after cgana
688 assert(get_Call_type(call) == get_entity_type(get_irg_ent(called_graph)));
689 assert(smaller_type(get_entity_type(get_irg_ent(called_graph)),
690 get_Call_type(call)));
692 assert(get_type_tpop(get_Call_type(call)) == type_method);
693 if (called_graph == current_ir_graph) {
694 set_optimize(rem_opt);
698 /* here we know we WILL inline, so inform the statistics */
699 stat_inline(call, called_graph);
701 /* -- Decide how to handle exception control flow: Is there a handler
702 for the Call node, or do we branch directly to End on an exception?
703 exc_handling: 0 There is a handler.
705 2 Exception handling not represented in Firm. -- */
707 ir_node *proj, *Mproj = NULL, *Xproj = NULL;
708 for (proj = (ir_node *)get_irn_link(call); proj; proj = (ir_node *)get_irn_link(proj)) {
709 assert(get_irn_op(proj) == op_Proj);
710 if (get_Proj_proj(proj) == pn_Call_X_except) Xproj = proj;
711 if (get_Proj_proj(proj) == pn_Call_M_except) Mproj = proj;
713 if (Mproj) { assert(Xproj); exc_handling = 0; } /* Mproj */
714 else if (Xproj) { exc_handling = 1; } /* !Mproj && Xproj */
715 else { exc_handling = 2; } /* !Mproj && !Xproj */
720 the procedure and later replaces the Start node of the called graph.
721 Post_call is the old Call node and collects the results of the called
722 graph. Both will end up being a tuple. -- */
723 post_bl = get_nodes_Block(call);
724 set_irg_current_block(current_ir_graph, post_bl);
725 /* XxMxPxP of Start + parameter of Call */
726 in[pn_Start_X_initial_exec] = new_Jmp();
727 in[pn_Start_M] = get_Call_mem(call);
728 in[pn_Start_P_frame_base] = get_irg_frame(current_ir_graph);
729 in[pn_Start_P_globals] = get_irg_globals(current_ir_graph);
730 in[pn_Start_T_args] = new_Tuple(get_Call_n_params(call), get_Call_param_arr(call));
731 /* in[pn_Start_P_value_arg_base] = ??? */
732 pre_call = new_Tuple(5, in);
736 The new block gets the ins of the old block, pre_call and all its
737 predecessors and all Phi nodes. -- */
738 part_block(pre_call);
740 /* -- Prepare state for dead node elimination -- */
741 /* Visited flags in calling irg must be >= flag in called irg.
742 Else walker and arity computation will not work. */
743 if (get_irg_visited(current_ir_graph) <= get_irg_visited(called_graph))
744 set_irg_visited(current_ir_graph, get_irg_visited(called_graph)+1);
745 if (get_irg_block_visited(current_ir_graph)< get_irg_block_visited(called_graph))
746 set_irg_block_visited(current_ir_graph, get_irg_block_visited(called_graph));
747 /* Set pre_call as new Start node in link field of the start node of
748 calling graph and pre_calls block as new block for the start block
750 Further mark these nodes so that they are not visited by the
752 set_irn_link(get_irg_start(called_graph), pre_call);
753 set_irn_visited(get_irg_start(called_graph),
754 get_irg_visited(current_ir_graph));
755 set_irn_link(get_irg_start_block(called_graph),
756 get_nodes_Block(pre_call));
757 set_irn_visited(get_irg_start_block(called_graph),
758 get_irg_visited(current_ir_graph));
760 /* Initialize for compaction of in arrays */
761 inc_irg_block_visited(current_ir_graph);
763 /* -- Replicate local entities of the called_graph -- */
764 /* copy the entities. */
765 called_frame = get_irg_frame_type(called_graph);
766 for (i = 0; i < get_class_n_members(called_frame); i++) {
767 entity *new_ent, *old_ent;
768 old_ent = get_class_member(called_frame, i);
769 new_ent = copy_entity_own(old_ent, get_cur_frame_type());
770 set_entity_link(old_ent, new_ent);
773 /* visited is > than that of called graph. With this trick visited will
774 remain unchanged so that an outer walker, e.g., searching the call nodes
775 to inline, calling this inline will not visit the inlined nodes. */
776 set_irg_visited(current_ir_graph, get_irg_visited(current_ir_graph)-1);
778 /* -- Performing dead node elimination inlines the graph -- */
779 /* Copies the nodes to the obstack of current_ir_graph. Updates links to new
781 /* @@@ endless loops are not copied!! -- they should be, I think... */
782 irg_walk(get_irg_end(called_graph), copy_node_inline, copy_preds,
783 get_irg_frame_type(called_graph));
785 /* Repair called_graph */
786 set_irg_visited(called_graph, get_irg_visited(current_ir_graph));
787 set_irg_block_visited(called_graph, get_irg_block_visited(current_ir_graph));
788 set_Block_block_visited(get_irg_start_block(called_graph), 0);
790 /* -- Merge the end of the inlined procedure with the call site -- */
791 /* We will turn the old Call node into a Tuple with the following
794 0: Phi of all Memories of Return statements.
795 1: Jmp from new Block that merges the control flow from all exception
796 predecessors of the old end block.
797 2: Tuple of all arguments.
798 3: Phi of Exception memories.
799 In case the old Call directly branches to End on an exception we don't
800 need the block merging all exceptions nor the Phi of the exception
804 /* -- Precompute some values -- */
805 end_bl = get_new_node(get_irg_end_block(called_graph));
806 end = get_new_node(get_irg_end(called_graph));
807 arity = get_irn_arity(end_bl); /* arity = n_exc + n_ret */
808 n_res = get_method_n_ress(get_Call_type(call));
810 res_pred = (ir_node **) malloc (n_res * sizeof (ir_node *));
811 cf_pred = (ir_node **) malloc (arity * sizeof (ir_node *));
813 set_irg_current_block(current_ir_graph, post_bl); /* just to make sure */
815 /* -- archive keepalives -- */
816 irn_arity = get_irn_arity(end);
817 for (i = 0; i < irn_arity; i++)
818 add_End_keepalive(get_irg_end(current_ir_graph), get_irn_n(end, i));
820 /* The new end node will die. We need not free as the in array is on the obstack:
821 copy_node only generated 'D' arrays. */
823 /* -- Replace Return nodes by Jump nodes. -- */
825 for (i = 0; i < arity; i++) {
827 ret = get_irn_n(end_bl, i);
828 if (get_irn_op(ret) == op_Return) {
829 cf_pred[n_ret] = new_r_Jmp(current_ir_graph, get_nodes_Block(ret));
833 set_irn_in(post_bl, n_ret, cf_pred);
835 /* -- Build a Tuple for all results of the method.
836 Add Phi node if there was more than one Return. -- */
837 turn_into_tuple(post_call, 4);
838 /* First the Memory-Phi */
840 for (i = 0; i < arity; i++) {
841 ret = get_irn_n(end_bl, i);
842 if (get_irn_op(ret) == op_Return) {
843 cf_pred[n_ret] = get_Return_mem(ret);
847 phi = new_Phi(n_ret, cf_pred, mode_M);
848 set_Tuple_pred(call, pn_Call_M_regular, phi);
849 /* Conserve Phi-list for further inlinings -- but might be optimized */
850 if (get_nodes_Block(phi) == post_bl) {
851 set_irn_link(phi, get_irn_link(post_bl));
852 set_irn_link(post_bl, phi);
854 /* Now the real results */
856 for (j = 0; j < n_res; j++) {
858 for (i = 0; i < arity; i++) {
859 ret = get_irn_n(end_bl, i);
860 if (get_irn_op(ret) == op_Return) {
861 cf_pred[n_ret] = get_Return_res(ret, j);
865 phi = new_Phi(n_ret, cf_pred, get_irn_mode(cf_pred[0]));
867 /* Conserve Phi-list for further inlinings -- but might be optimized */
868 if (get_nodes_Block(phi) == post_bl) {
869 set_irn_link(phi, get_irn_link(post_bl));
870 set_irn_link(post_bl, phi);
873 set_Tuple_pred(call, pn_Call_T_result, new_Tuple(n_res, res_pred));
875 set_Tuple_pred(call, pn_Call_T_result, new_Bad());
877 /* Finally the exception control flow.
878 We have two (three) possible situations:
879 First if the Call branches to an exception handler: We need to add a Phi node to
880 collect the memory containing the exception objects. Further we need
881 to add another block to get a correct representation of this Phi. To
882 this block we add a Jmp that resolves into the X output of the Call
883 when the Call is turned into a tuple.
884 Second the Call branches to End, the exception is not handled. Just
885 add all inlined exception branches to the End node.
886 Third: there is no Exception edge at all. Handle as case two. */
887 if (exc_handling == 0) {
889 for (i = 0; i < arity; i++) {
891 ret = get_irn_n(end_bl, i);
892 if (is_fragile_op(skip_Proj(ret)) || (get_irn_op(skip_Proj(ret)) == op_Raise)) {
893 cf_pred[n_exc] = ret;
898 new_Block(n_exc, cf_pred); /* watch it: current_block is changed! */
899 set_Tuple_pred(call, pn_Call_X_except, new_Jmp());
900 /* The Phi for the memories with the exception objects */
902 for (i = 0; i < arity; i++) {
904 ret = skip_Proj(get_irn_n(end_bl, i));
905 if (get_irn_op(ret) == op_Call) {
906 cf_pred[n_exc] = new_r_Proj(current_ir_graph, get_nodes_Block(ret), ret, mode_M, 3);
908 } else if (is_fragile_op(ret)) {
909 /* We rely that all cfops have the memory output at the same position. */
910 cf_pred[n_exc] = new_r_Proj(current_ir_graph, get_nodes_Block(ret), ret, mode_M, 0);
912 } else if (get_irn_op(ret) == op_Raise) {
913 cf_pred[n_exc] = new_r_Proj(current_ir_graph, get_nodes_Block(ret), ret, mode_M, 1);
917 set_Tuple_pred(call, pn_Call_M_except, new_Phi(n_exc, cf_pred, mode_M));
919 set_Tuple_pred(call, pn_Call_X_except, new_Bad());
920 set_Tuple_pred(call, pn_Call_M_except, new_Bad());
923 ir_node *main_end_bl;
924 int main_end_bl_arity;
927 /* assert(exc_handling == 1 || no exceptions. ) */
929 for (i = 0; i < arity; i++) {
930 ir_node *ret = get_irn_n(end_bl, i);
932 if (is_fragile_op(skip_Proj(ret)) || (get_irn_op(skip_Proj(ret)) == op_Raise)) {
933 cf_pred[n_exc] = ret;
937 main_end_bl = get_irg_end_block(current_ir_graph);
938 main_end_bl_arity = get_irn_arity(main_end_bl);
939 end_preds = (ir_node **) malloc ((n_exc + main_end_bl_arity) * sizeof (ir_node *));
941 for (i = 0; i < main_end_bl_arity; ++i)
942 end_preds[i] = get_irn_n(main_end_bl, i);
943 for (i = 0; i < n_exc; ++i)
944 end_preds[main_end_bl_arity + i] = cf_pred[i];
945 set_irn_in(main_end_bl, n_exc + main_end_bl_arity, end_preds);
946 set_Tuple_pred(call, pn_Call_X_except, new_Bad());
947 set_Tuple_pred(call, pn_Call_M_except, new_Bad());
953 #if 0 /* old. now better, correcter, faster implementation. */
955 /* -- If the exception control flow from the inlined Call directly
956 branched to the end block we now have the following control
957 flow predecessor pattern: ProjX -> Tuple -> Jmp. We must
958 remove the Jmp along with it's empty block and add Jmp's
959 predecessors as predecessors of this end block. No problem if
960 there is no exception, because then branches Bad to End which
962 @@@ can't we know this beforehand: by getting the Proj(1) from
963 the Call link list and checking whether it goes to Proj. */
964 /* find the problematic predecessor of the end block. */
965 end_bl = get_irg_end_block(current_ir_graph);
966 for (i = 0; i < get_Block_n_cfgpreds(end_bl); i++) {
967 cf_op = get_Block_cfgpred(end_bl, i);
968 if (get_irn_op(cf_op) == op_Proj) {
969 cf_op = get_Proj_pred(cf_op);
970 if ((get_irn_op(cf_op) == op_Tuple) && (cf_op == call)) {
971 /* There are unoptimized tuples from inlineing before when no exc */
972 assert(get_Proj_proj(get_Block_cfgpred(end_bl, i)) == pn_Call_X_except);
973 cf_op = get_Tuple_pred(cf_op, pn_Call_X_except);
974 assert(get_irn_op(cf_op) == op_Jmp);
980 if (i < get_Block_n_cfgpreds(end_bl)) {
981 bl = get_nodes_Block(cf_op);
982 arity = get_Block_n_cfgpreds(end_bl) + get_Block_n_cfgpreds(bl) - 1;
983 cf_pred = (ir_node **) malloc (arity * sizeof (ir_node *));
984 for (j = 0; j < i; j++)
985 cf_pred[j] = get_Block_cfgpred(end_bl, j);
986 for (j = j; j < i + get_Block_n_cfgpreds(bl); j++)
987 cf_pred[j] = get_Block_cfgpred(bl, j-i);
988 for (j = j; j < arity; j++)
989 cf_pred[j] = get_Block_cfgpred(end_bl, j-get_Block_n_cfgpreds(bl) +1);
990 set_irn_in(end_bl, arity, cf_pred);
992 /* Remove the exception pred from post-call Tuple. */
993 set_Tuple_pred(call, pn_Call_X_except, new_Bad());
998 /* -- Turn cse back on. -- */
999 set_optimize(rem_opt);
1004 /********************************************************************/
1005 /* Apply inlineing to small methods. */
1006 /********************************************************************/
1008 /* It makes no sense to inline too many calls in one procedure. Anyways,
1009 I didn't get a version with NEW_ARR_F to run. */
1010 #define MAX_INLINE 1024
1013 * environment for inlining small irgs
1015 typedef struct _inline_env_t {
1017 ir_node *calls[MAX_INLINE];
1021 * Returns the irg called from a Call node. If the irg is not
1022 * known, NULL is returned.
1024 static ir_graph *get_call_called_irg(ir_node *call) {
1027 ir_graph *called_irg = NULL;
1029 assert(get_irn_op(call) == op_Call);
1031 addr = get_Call_ptr(call);
1032 if (get_irn_op(addr) == op_Const) {
1033 /* Check whether the constant is the pointer to a compiled entity. */
1034 tv = get_Const_tarval(addr);
1035 if (tarval_to_entity(tv))
1036 called_irg = get_entity_irg(tarval_to_entity(tv));
1041 static void collect_calls(ir_node *call, void *env) {
1042 inline_env_t *ienv = env;
1045 ir_graph *called_irg;
1047 if (get_irn_op(call) != op_Call) return;
1049 addr = get_Call_ptr(call);
1050 if (get_irn_op(addr) == op_Const) {
1051 /* Check whether the constant is the pointer to a compiled entity. */
1052 tv = get_Const_tarval(addr);
1053 if (tarval_to_entity(tv)) {
1054 called_irg = get_entity_irg(tarval_to_entity(tv));
1055 if (called_irg && ienv->pos < MAX_INLINE) {
1056 /* The Call node calls a locally defined method. Remember to inline. */
1057 ienv->calls[ienv->pos++] = call;
1064 * Inlines all small methods at call sites where the called address comes
1065 * from a Const node that references the entity representing the called
1067 * The size argument is a rough measure for the code size of the method:
1068 * Methods where the obstack containing the firm graph is smaller than
1071 void inline_small_irgs(ir_graph *irg, int size) {
1073 ir_graph *rem = current_ir_graph;
1076 if (!(get_opt_optimize() && get_opt_inline())) return;
1078 current_ir_graph = irg;
1079 /* Handle graph state */
1080 assert(get_irg_phase_state(current_ir_graph) != phase_building);
1081 assert(get_irg_callee_info_state(current_ir_graph) == irg_callee_info_none);
1083 /* Find Call nodes to inline.
1084 (We can not inline during a walk of the graph, as inlineing the same
1085 method several times changes the visited flag of the walked graph:
1086 after the first inlineing visited of the callee equals visited of
1087 the caller. With the next inlineing both are increased.) */
1089 irg_walk(get_irg_end(irg), NULL, collect_calls, &env);
1091 if ((env.pos > 0) && (env.pos < MAX_INLINE)) {
1092 /* There are calls to inline */
1093 collect_phiprojs(irg);
1094 for (i = 0; i < env.pos; i++) {
1097 tv = get_Const_tarval(get_Call_ptr(env.calls[i]));
1098 callee = get_entity_irg(tarval_to_entity(tv));
1099 if (((_obstack_memory_used(callee->obst) - obstack_room(callee->obst)) < size) ||
1100 (get_irg_inline_property(callee) == irg_inline_forced)) {
1101 inline_method(env.calls[i], callee);
1106 current_ir_graph = rem;
1110 * Environment for inlining irgs.
1113 int n_nodes; /**< Nodes in graph except Id, Tuple, Proj, Start, End */
1114 int n_nodes_orig; /**< for statistics */
1115 eset *call_nodes; /**< All call nodes in this graph */
1117 int n_call_nodes_orig; /**< for statistics */
1118 int n_callers; /**< Number of known graphs that call this graphs. */
1119 int n_callers_orig; /**< for statistics */
1122 static inline_irg_env *new_inline_irg_env(void) {
1123 inline_irg_env *env = malloc(sizeof(inline_irg_env));
1124 env->n_nodes = -2; /* uncount Start, End */
1125 env->n_nodes_orig = -2; /* uncount Start, End */
1126 env->call_nodes = eset_create();
1127 env->n_call_nodes = 0;
1128 env->n_call_nodes_orig = 0;
1130 env->n_callers_orig = 0;
1134 static void free_inline_irg_env(inline_irg_env *env) {
1135 eset_destroy(env->call_nodes);
1139 static void collect_calls2(ir_node *call, void *env) {
1140 inline_irg_env *x = (inline_irg_env *)env;
1141 ir_op *op = get_irn_op(call);
1144 /* count nodes in irg */
1145 if (op != op_Proj && op != op_Tuple && op != op_Sync) {
1150 if (op != op_Call) return;
1152 /* collect all call nodes */
1153 eset_insert(x->call_nodes, (void *)call);
1155 x->n_call_nodes_orig++;
1157 /* count all static callers */
1158 callee = get_call_called_irg(call);
1160 ((inline_irg_env *)get_irg_link(callee))->n_callers++;
1161 ((inline_irg_env *)get_irg_link(callee))->n_callers_orig++;
1165 INLINE static int is_leave(ir_graph *irg) {
1166 return (((inline_irg_env *)get_irg_link(irg))->n_call_nodes == 0);
1169 INLINE static int is_smaller(ir_graph *callee, int size) {
1170 return (((inline_irg_env *)get_irg_link(callee))->n_nodes < size);
1175 * Inlines small leave methods at call sites where the called address comes
1176 * from a Const node that references the entity representing the called
1178 * The size argument is a rough measure for the code size of the method:
1179 * Methods where the obstack containing the firm graph is smaller than
1182 void inline_leave_functions(int maxsize, int leavesize, int size) {
1183 inline_irg_env *env;
1184 int i, n_irgs = get_irp_n_irgs();
1185 ir_graph *rem = current_ir_graph;
1188 if (!(get_opt_optimize() && get_opt_inline())) return;
1190 /* extend all irgs by a temporary data structure for inlineing. */
1191 for (i = 0; i < n_irgs; ++i)
1192 set_irg_link(get_irp_irg(i), new_inline_irg_env());
1194 /* Precompute information in temporary data structure. */
1195 for (i = 0; i < n_irgs; ++i) {
1196 current_ir_graph = get_irp_irg(i);
1197 assert(get_irg_phase_state(current_ir_graph) != phase_building);
1198 assert(get_irg_callee_info_state(current_ir_graph) == irg_callee_info_none);
1200 irg_walk(get_irg_end(current_ir_graph), NULL, collect_calls2,
1201 get_irg_link(current_ir_graph));
1205 Inline leaves recursively -- we might construct new leaves. */
1206 /* int itercnt = 1; */
1207 while (did_inline) {
1208 /* printf("iteration %d\n", itercnt++); */
1210 for (i = 0; i < n_irgs; ++i) {
1213 int phiproj_computed = 0;
1215 current_ir_graph = get_irp_irg(i);
1216 env = (inline_irg_env *)get_irg_link(current_ir_graph);
1218 /* we can not walk and change a set, nor remove from it.
1220 walkset = env->call_nodes;
1221 env->call_nodes = eset_create();
1222 for (call = eset_first(walkset); call; call = eset_next(walkset)) {
1223 inline_irg_env *callee_env;
1224 ir_graph *callee = get_call_called_irg(call);
1226 if (env->n_nodes > maxsize) break;
1228 ((is_leave(callee) && is_smaller(callee, leavesize)) ||
1229 (get_irg_inline_property(callee) == irg_inline_forced))) {
1230 if (!phiproj_computed) {
1231 phiproj_computed = 1;
1232 collect_phiprojs(current_ir_graph);
1234 callee_env = (inline_irg_env *)get_irg_link(callee);
1235 /* printf(" %s: Inlineing %s.\n", get_entity_name(get_irg_entity(current_ir_graph)), */
1236 /* get_entity_name(get_irg_entity(callee))); */
1237 if (inline_method(call, callee)) {
1239 env->n_call_nodes--;
1240 eset_insert_all(env->call_nodes, callee_env->call_nodes);
1241 env->n_call_nodes += callee_env->n_call_nodes;
1242 env->n_nodes += callee_env->n_nodes;
1243 callee_env->n_callers--;
1246 eset_insert(env->call_nodes, call);
1249 eset_destroy(walkset);
1253 /* printf("Non leaves\n"); */
1254 /* inline other small functions. */
1255 for (i = 0; i < n_irgs; ++i) {
1258 int phiproj_computed = 0;
1260 current_ir_graph = get_irp_irg(i);
1261 env = (inline_irg_env *)get_irg_link(current_ir_graph);
1263 /* we can not walk and change a set, nor remove from it.
1265 walkset = env->call_nodes;
1266 env->call_nodes = eset_create();
1267 for (call = eset_first(walkset); call; call = eset_next(walkset)) {
1268 inline_irg_env *callee_env;
1269 ir_graph *callee = get_call_called_irg(call);
1271 if (env->n_nodes > maxsize) break;
1272 if (callee && is_smaller(callee, size)) {
1273 if (!phiproj_computed) {
1274 phiproj_computed = 1;
1275 collect_phiprojs(current_ir_graph);
1277 callee_env = (inline_irg_env *)get_irg_link(callee);
1278 /* printf(" %s: Inlineing %s.\n", get_entity_name(get_irg_entity(current_ir_graph)), */
1279 /* get_entity_name(get_irg_entity(callee))); */
1280 if (inline_method(call, callee)) {
1282 env->n_call_nodes--;
1283 eset_insert_all(env->call_nodes, callee_env->call_nodes);
1284 env->n_call_nodes += callee_env->n_call_nodes;
1285 env->n_nodes += callee_env->n_nodes;
1286 callee_env->n_callers--;
1289 eset_insert(env->call_nodes, call);
1292 eset_destroy(walkset);
1295 for (i = 0; i < n_irgs; ++i) {
1296 current_ir_graph = get_irp_irg(i);
1298 env = (inline_irg_env *)get_irg_link(current_ir_graph);
1299 if ((env->n_call_nodes_orig != env->n_call_nodes) ||
1300 (env->n_callers_orig != env->n_callers))
1301 printf("Nodes:%3d ->%3d, calls:%3d ->%3d, callers:%3d ->%3d, -- %s\n",
1302 env->n_nodes_orig, env->n_nodes, env->n_call_nodes_orig, env->n_call_nodes,
1303 env->n_callers_orig, env->n_callers,
1304 get_entity_name(get_irg_entity(current_ir_graph)));
1306 free_inline_irg_env((inline_irg_env *)get_irg_link(current_ir_graph));
1309 current_ir_graph = rem;
1312 /*******************************************************************/
1313 /* Code Placement. Pins all floating nodes to a block where they */
1314 /* will be executed only if needed. */
1315 /*******************************************************************/
1318 * Find the earliest correct block for N. --- Place N into the
1319 * same Block as its dominance-deepest Input.
1322 place_floats_early(ir_node *n, pdeq *worklist)
1324 int i, start, irn_arity;
1326 /* we must not run into an infinite loop */
1327 assert (irn_not_visited(n));
1328 mark_irn_visited(n);
1330 /* Place floating nodes. */
1331 if (get_op_pinned(get_irn_op(n)) == floats) {
1333 ir_node *b = new_Bad(); /* The block to place this node in */
1335 assert(get_irn_op(n) != op_Block);
1337 if ((get_irn_op(n) == op_Const) ||
1338 (get_irn_op(n) == op_SymConst) ||
1340 (get_irn_op(n) == op_Unknown)) {
1341 /* These nodes will not be placed by the loop below. */
1342 b = get_irg_start_block(current_ir_graph);
1346 /* find the block for this node. */
1347 irn_arity = get_irn_arity(n);
1348 for (i = 0; i < irn_arity; i++) {
1349 ir_node *dep = get_irn_n(n, i);
1351 if ((irn_not_visited(dep)) &&
1352 (get_op_pinned(get_irn_op(dep)) == floats)) {
1353 place_floats_early(dep, worklist);
1355 /* Because all loops contain at least one pinned node, now all
1356 our inputs are either pinned or place_early has already
1357 been finished on them. We do not have any unfinished inputs! */
1358 dep_block = get_nodes_Block(dep);
1359 if ((!is_Bad(dep_block)) &&
1360 (get_Block_dom_depth(dep_block) > depth)) {
1362 depth = get_Block_dom_depth(dep_block);
1364 /* Avoid that the node is placed in the Start block */
1365 if ((depth == 1) && (get_Block_dom_depth(get_nodes_Block(n)) > 1)) {
1366 b = get_Block_cfg_out(get_irg_start_block(current_ir_graph), 0);
1367 assert(b != get_irg_start_block(current_ir_graph));
1371 set_nodes_Block(n, b);
1374 /* Add predecessors of non floating nodes on worklist. */
1375 start = (get_irn_op(n) == op_Block) ? 0 : -1;
1376 irn_arity = get_irn_arity(n);
1377 for (i = start; i < irn_arity; i++) {
1378 ir_node *pred = get_irn_n(n, i);
1379 if (irn_not_visited(pred)) {
1380 pdeq_putr (worklist, pred);
1386 * Floating nodes form subgraphs that begin at nodes as Const, Load,
1387 * Start, Call and end at pinned nodes as Store, Call. Place_early
1388 * places all floating nodes reachable from its argument through floating
1389 * nodes and adds all beginnings at pinned nodes to the worklist.
1391 static INLINE void place_early(pdeq* worklist) {
1393 inc_irg_visited(current_ir_graph);
1395 /* this inits the worklist */
1396 place_floats_early(get_irg_end(current_ir_graph), worklist);
1398 /* Work the content of the worklist. */
1399 while (!pdeq_empty (worklist)) {
1400 ir_node *n = pdeq_getl (worklist);
1401 if (irn_not_visited(n)) place_floats_early(n, worklist);
1404 set_irg_outs_inconsistent(current_ir_graph);
1405 current_ir_graph->pinned = pinned;
1409 /** deepest common dominance ancestor of DCA and CONSUMER of PRODUCER. */
1411 consumer_dom_dca (ir_node *dca, ir_node *consumer, ir_node *producer)
1413 ir_node *block = NULL;
1415 /* Compute the latest block into which we can place a node so that it is
1417 if (get_irn_op(consumer) == op_Phi) {
1418 /* our consumer is a Phi-node, the effective use is in all those
1419 blocks through which the Phi-node reaches producer */
1421 ir_node *phi_block = get_nodes_Block(consumer);
1422 irn_arity = get_irn_arity(consumer);
1423 for (i = 0; i < irn_arity; i++) {
1424 if (get_irn_n(consumer, i) == producer) {
1425 block = get_nodes_Block(get_Block_cfgpred(phi_block, i));
1429 assert(is_no_Block(consumer));
1430 block = get_nodes_Block(consumer);
1433 /* Compute the deepest common ancestor of block and dca. */
1435 if (!dca) return block;
1436 while (get_Block_dom_depth(block) > get_Block_dom_depth(dca))
1437 block = get_Block_idom(block);
1438 while (get_Block_dom_depth(dca) > get_Block_dom_depth(block))
1439 dca = get_Block_idom(dca);
1440 while (block != dca)
1441 { block = get_Block_idom(block); dca = get_Block_idom(dca); }
1446 static INLINE int get_irn_loop_depth(ir_node *n) {
1447 return get_loop_depth(get_irn_loop(n));
1451 * Move n to a block with less loop depth than it's current block. The
1452 * new block must be dominated by early.
1455 move_out_of_loops (ir_node *n, ir_node *early)
1457 ir_node *best, *dca;
1461 /* Find the region deepest in the dominator tree dominating
1462 dca with the least loop nesting depth, but still dominated
1463 by our early placement. */
1464 dca = get_nodes_Block(n);
1466 while (dca != early) {
1467 dca = get_Block_idom(dca);
1468 if (!dca) break; /* should we put assert(dca)? */
1469 if (get_irn_loop_depth(dca) < get_irn_loop_depth(best)) {
1473 if (best != get_nodes_Block(n)) {
1475 printf("Moving out of loop: "); DDMN(n);
1476 printf(" Outermost block: "); DDMN(early);
1477 printf(" Best block: "); DDMN(best);
1478 printf(" Innermost block: "); DDMN(get_nodes_Block(n));
1480 set_nodes_Block(n, best);
1485 * Find the latest legal block for N and place N into the
1486 * `optimal' Block between the latest and earliest legal block.
1487 * The `optimal' block is the dominance-deepest block of those
1488 * with the least loop-nesting-depth. This places N out of as many
1489 * loops as possible and then makes it as control dependant as
1493 place_floats_late(ir_node *n, pdeq *worklist)
1498 assert (irn_not_visited(n)); /* no multiple placement */
1500 /* no need to place block nodes, control nodes are already placed. */
1501 if ((get_irn_op(n) != op_Block) &&
1503 (get_irn_mode(n) != mode_X)) {
1504 /* Remember the early placement of this block to move it
1505 out of loop no further than the early placement. */
1506 early = get_nodes_Block(n);
1507 /* Assure that our users are all placed, except the Phi-nodes.
1508 --- Each data flow cycle contains at least one Phi-node. We
1509 have to break the `user has to be placed before the
1510 producer' dependence cycle and the Phi-nodes are the
1511 place to do so, because we need to base our placement on the
1512 final region of our users, which is OK with Phi-nodes, as they
1513 are pinned, and they never have to be placed after a
1514 producer of one of their inputs in the same block anyway. */
1515 for (i = 0; i < get_irn_n_outs(n); i++) {
1516 ir_node *succ = get_irn_out(n, i);
1517 if (irn_not_visited(succ) && (get_irn_op(succ) != op_Phi))
1518 place_floats_late(succ, worklist);
1521 /* We have to determine the final block of this node... except for
1523 if ((get_op_pinned(get_irn_op(n)) == floats) &&
1524 (get_irn_op(n) != op_Const) &&
1525 (get_irn_op(n) != op_SymConst)) {
1526 ir_node *dca = NULL; /* deepest common ancestor in the
1527 dominator tree of all nodes'
1528 blocks depending on us; our final
1529 placement has to dominate DCA. */
1530 for (i = 0; i < get_irn_n_outs(n); i++) {
1531 dca = consumer_dom_dca (dca, get_irn_out(n, i), n);
1533 set_nodes_Block(n, dca);
1535 move_out_of_loops (n, early);
1539 mark_irn_visited(n);
1541 /* Add predecessors of all non-floating nodes on list. (Those of floating
1542 nodes are placeded already and therefore are marked.) */
1543 for (i = 0; i < get_irn_n_outs(n); i++) {
1544 if (irn_not_visited(get_irn_out(n, i))) {
1545 pdeq_putr (worklist, get_irn_out(n, i));
1550 static INLINE void place_late(pdeq* worklist) {
1552 inc_irg_visited(current_ir_graph);
1554 /* This fills the worklist initially. */
1555 place_floats_late(get_irg_start_block(current_ir_graph), worklist);
1556 /* And now empty the worklist again... */
1557 while (!pdeq_empty (worklist)) {
1558 ir_node *n = pdeq_getl (worklist);
1559 if (irn_not_visited(n)) place_floats_late(n, worklist);
1563 void place_code(ir_graph *irg) {
1565 ir_graph *rem = current_ir_graph;
1567 current_ir_graph = irg;
1569 if (!(get_opt_optimize() && get_opt_global_cse())) return;
1571 /* Handle graph state */
1572 assert(get_irg_phase_state(irg) != phase_building);
1573 if (get_irg_dom_state(irg) != dom_consistent)
1576 if (get_irg_loopinfo_state(irg) != loopinfo_consistent) {
1577 free_loop_information(irg);
1578 construct_backedges(irg);
1581 /* Place all floating nodes as early as possible. This guarantees
1582 a legal code placement. */
1583 worklist = new_pdeq();
1584 place_early(worklist);
1586 /* place_early invalidates the outs, place_late needs them. */
1588 /* Now move the nodes down in the dominator tree. This reduces the
1589 unnecessary executions of the node. */
1590 place_late(worklist);
1592 set_irg_outs_inconsistent(current_ir_graph);
1593 set_irg_loopinfo_inconsistent(current_ir_graph);
1595 current_ir_graph = rem;
1600 /********************************************************************/
1601 /* Control flow optimization. */
1602 /* Removes Bad control flow predecessors and empty blocks. A block */
1603 /* is empty if it contains only a Jmp node. */
1604 /* Blocks can only be removed if they are not needed for the */
1605 /* semantics of Phi nodes. */
1606 /********************************************************************/
1609 * Removes Tuples from Block control flow predecessors.
1610 * Optimizes blocks with equivalent_node().
1611 * Replaces n by Bad if n is unreachable control flow.
1613 static void merge_blocks(ir_node *n, void *env) {
1615 set_irn_link(n, NULL);
1617 if (get_irn_op(n) == op_Block) {
1619 for (i = 0; i < get_Block_n_cfgpreds(n); i++)
1620 /* GL @@@ : is this possible? if (get_opt_normalize()) -- added, all tests go through.
1621 A different order of optimizations might cause problems. */
1622 if (get_opt_normalize())
1623 set_Block_cfgpred(n, i, skip_Tuple(get_Block_cfgpred(n, i)));
1624 } else if (get_opt_optimize() && (get_irn_mode(n) == mode_X)) {
1625 /* We will soon visit a block. Optimize it before visiting! */
1626 ir_node *b = get_nodes_Block(n);
1627 ir_node *new_node = equivalent_node(b);
1628 while (irn_not_visited(b) && (!is_Bad(new_node)) && (new_node != b)) {
1629 /* We would have to run gigo if new is bad, so we
1630 promote it directly below. */
1631 assert(((b == new_node) ||
1632 get_opt_control_flow_straightening() ||
1633 get_opt_control_flow_weak_simplification()) &&
1634 ("strange flag setting"));
1635 exchange (b, new_node);
1637 new_node = equivalent_node(b);
1639 /* GL @@@ get_opt_normalize hinzugefuegt, 5.5.2003 */
1640 if (is_Bad(new_node) && get_opt_normalize()) exchange(n, new_Bad());
1645 * Collects all Phi nodes in link list of Block.
1646 * Marks all blocks "block_visited" if they contain a node other
1649 static void collect_nodes(ir_node *n, void *env) {
1650 if (is_no_Block(n)) {
1651 ir_node *b = get_nodes_Block(n);
1653 if ((get_irn_op(n) == op_Phi)) {
1654 /* Collect Phi nodes to compact ins along with block's ins. */
1655 set_irn_link(n, get_irn_link(b));
1657 } else if (get_irn_op(n) != op_Jmp) { /* Check for non empty block. */
1658 mark_Block_block_visited(b);
1663 /** Returns true if pred is predecessor of block. */
1664 static int is_pred_of(ir_node *pred, ir_node *b) {
1666 for (i = 0; i < get_Block_n_cfgpreds(b); i++) {
1667 ir_node *b_pred = get_nodes_Block(get_Block_cfgpred(b, i));
1668 if (b_pred == pred) return 1;
1673 static int test_whether_dispensable(ir_node *b, int pos) {
1674 int i, j, n_preds = 1;
1675 int dispensable = 1;
1676 ir_node *cfop = get_Block_cfgpred(b, pos);
1677 ir_node *pred = get_nodes_Block(cfop);
1679 if (get_Block_block_visited(pred) + 1
1680 < get_irg_block_visited(current_ir_graph)) {
1681 if (!get_opt_optimize() || !get_opt_control_flow_strong_simplification()) {
1682 /* Mark block so that is will not be removed. */
1683 set_Block_block_visited(pred, get_irg_block_visited(current_ir_graph)-1);
1686 /* Seems to be empty. */
1687 if (!get_irn_link(b)) {
1688 /* There are no Phi nodes ==> dispensable. */
1689 n_preds = get_Block_n_cfgpreds(pred);
1691 /* b's pred blocks and pred's pred blocks must be pairwise disjunct.
1692 Work preds < pos as if they were already removed. */
1693 for (i = 0; i < pos; i++) {
1694 ir_node *b_pred = get_nodes_Block(get_Block_cfgpred(b, i));
1695 if (get_Block_block_visited(b_pred) + 1
1696 < get_irg_block_visited(current_ir_graph)) {
1697 for (j = 0; j < get_Block_n_cfgpreds(b_pred); j++) {
1698 ir_node *b_pred_pred = get_nodes_Block(get_Block_cfgpred(b_pred, j));
1699 if (is_pred_of(b_pred_pred, pred)) dispensable = 0;
1702 if (is_pred_of(b_pred, pred)) dispensable = 0;
1705 for (i = pos +1; i < get_Block_n_cfgpreds(b); i++) {
1706 ir_node *b_pred = get_nodes_Block(get_Block_cfgpred(b, i));
1707 if (is_pred_of(b_pred, pred)) dispensable = 0;
1710 set_Block_block_visited(pred, get_irg_block_visited(current_ir_graph)-1);
1713 n_preds = get_Block_n_cfgpreds(pred);
1721 static void optimize_blocks(ir_node *b, void *env) {
1722 int i, j, k, max_preds, n_preds;
1723 ir_node *pred, *phi;
1726 /* Count the number of predecessor if this block is merged with pred blocks
1729 for (i = 0; i < get_Block_n_cfgpreds(b); i++) {
1730 max_preds += test_whether_dispensable(b, i);
1732 in = (ir_node **) malloc(max_preds * sizeof(ir_node *));
1735 printf(" working on "); DDMN(b);
1736 for (i = 0; i < get_Block_n_cfgpreds(b); i++) {
1737 pred = get_nodes_Block(get_Block_cfgpred(b, i));
1738 if (is_Bad(get_Block_cfgpred(b, i))) {
1739 printf(" removing Bad %i\n ", i);
1740 } else if (get_Block_block_visited(pred) +1
1741 < get_irg_block_visited(current_ir_graph)) {
1742 printf(" removing pred %i ", i); DDMN(pred);
1743 } else { printf(" Nothing to do for "); DDMN(pred); }
1745 * end Debug output -*/
1747 /*- Fix the Phi nodes -*/
1748 phi = get_irn_link(b);
1750 assert(get_irn_op(phi) == op_Phi);
1751 /* Find the new predecessors for the Phi */
1753 for (i = 0; i < get_Block_n_cfgpreds(b); i++) {
1754 pred = get_nodes_Block(get_Block_cfgpred(b, i));
1755 if (is_Bad(get_Block_cfgpred(b, i))) {
1757 } else if (get_Block_block_visited(pred) +1
1758 < get_irg_block_visited(current_ir_graph)) {
1759 /* It's an empty block and not yet visited. */
1760 ir_node *phi_pred = get_Phi_pred(phi, i);
1761 for (j = 0; j < get_Block_n_cfgpreds(pred); j++) {
1762 if (get_nodes_Block(phi_pred) == pred) {
1763 assert(get_irn_op(phi_pred) == op_Phi); /* Block is empty!! */
1764 in[n_preds] = get_Phi_pred(phi_pred, j);
1766 in[n_preds] = phi_pred;
1770 /* The Phi_pred node is replaced now if it is a Phi.
1771 In Schleifen kann offenbar der entfernte Phi Knoten legal verwendet werden.
1772 Daher muss der Phiknoten durch den neuen ersetzt werden.
1773 Weiter muss der alte Phiknoten entfernt werden (durch ersetzen oder
1774 durch einen Bad) damit er aus den keep_alive verschwinden kann.
1775 Man sollte also, falls keine Schleife vorliegt, exchange mit new_Bad
1777 if (get_nodes_Block(phi_pred) == pred) {
1778 /* remove the Phi as it might be kept alive. Further there
1779 might be other users. */
1780 exchange(phi_pred, phi); /* geht, ist aber doch semantisch falsch! Warum?? */
1783 in[n_preds] = get_Phi_pred(phi, i);
1788 set_irn_in(phi, n_preds, in);
1790 phi = get_irn_link(phi);
1794 This happens only if merge between loop backedge and single loop entry. -*/
1795 for (k = 0; k < get_Block_n_cfgpreds(b); k++) {
1796 pred = get_nodes_Block(get_Block_cfgpred(b, k));
1797 if (get_Block_block_visited(pred)+1 < get_irg_block_visited(current_ir_graph)) {
1798 phi = get_irn_link(pred);
1800 if (get_irn_op(phi) == op_Phi) {
1801 set_nodes_Block(phi, b);
1804 for (i = 0; i < k; i++) {
1805 pred = get_nodes_Block(get_Block_cfgpred(b, i));
1806 if (is_Bad(get_Block_cfgpred(b, i))) {
1808 } else if (get_Block_block_visited(pred) +1
1809 < get_irg_block_visited(current_ir_graph)) {
1810 /* It's an empty block and not yet visited. */
1811 for (j = 0; j < get_Block_n_cfgpreds(pred); j++) {
1812 /* @@@ Hier brauche ich Schleifeninformation!!! Kontrollflusskante
1813 muss Rueckwaertskante sein! (An allen vier in[n_preds] = phi
1814 Anweisungen.) Trotzdem tuts bisher!! */
1823 for (i = 0; i < get_Phi_n_preds(phi); i++) {
1824 in[n_preds] = get_Phi_pred(phi, i);
1827 for (i = k+1; i < get_Block_n_cfgpreds(b); i++) {
1828 pred = get_nodes_Block(get_Block_cfgpred(b, i));
1829 if (is_Bad(get_Block_cfgpred(b, i))) {
1831 } else if (get_Block_block_visited(pred) +1
1832 < get_irg_block_visited(current_ir_graph)) {
1833 /* It's an empty block and not yet visited. */
1834 for (j = 0; j < get_Block_n_cfgpreds(pred); j++) {
1843 set_irn_in(phi, n_preds, in);
1845 phi = get_irn_link(phi);
1850 /*- Fix the block -*/
1852 for (i = 0; i < get_Block_n_cfgpreds(b); i++) {
1853 pred = get_nodes_Block(get_Block_cfgpred(b, i));
1854 if (is_Bad(get_Block_cfgpred(b, i))) {
1856 } else if (get_Block_block_visited(pred) +1
1857 < get_irg_block_visited(current_ir_graph)) {
1858 /* It's an empty block and not yet visited. */
1859 assert(get_Block_n_cfgpreds(b) > 1);
1860 /* Else it should be optimized by equivalent_node. */
1861 for (j = 0; j < get_Block_n_cfgpreds(pred); j++) {
1862 in[n_preds] = get_Block_cfgpred(pred, j);
1865 /* Remove block as it might be kept alive. */
1866 exchange(pred, b/*new_Bad()*/);
1868 in[n_preds] = get_Block_cfgpred(b, i);
1872 set_irn_in(b, n_preds, in);
1876 void optimize_cf(ir_graph *irg) {
1879 ir_node *end = get_irg_end(irg);
1880 ir_graph *rem = current_ir_graph;
1881 current_ir_graph = irg;
1883 /* Handle graph state */
1884 assert(get_irg_phase_state(irg) != phase_building);
1885 if (get_irg_outs_state(current_ir_graph) == outs_consistent)
1886 set_irg_outs_inconsistent(current_ir_graph);
1887 if (get_irg_dom_state(current_ir_graph) == dom_consistent)
1888 set_irg_dom_inconsistent(current_ir_graph);
1890 /* Use block visited flag to mark non-empty blocks. */
1891 inc_irg_block_visited(irg);
1892 irg_walk(end, merge_blocks, collect_nodes, NULL);
1894 /* Optimize the standard code. */
1895 irg_block_walk(get_irg_end_block(irg), optimize_blocks, NULL, NULL);
1897 /* Walk all keep alives, optimize them if block, add to new in-array
1898 for end if useful. */
1899 in = NEW_ARR_F (ir_node *, 1);
1900 in[0] = get_nodes_Block(end);
1901 inc_irg_visited(current_ir_graph);
1902 for(i = 0; i < get_End_n_keepalives(end); i++) {
1903 ir_node *ka = get_End_keepalive(end, i);
1904 if (irn_not_visited(ka)) {
1905 if ((get_irn_op(ka) == op_Block) && Block_not_block_visited(ka)) {
1906 set_irg_block_visited(current_ir_graph, /* Don't walk all the way to Start. */
1907 get_irg_block_visited(current_ir_graph)-1);
1908 irg_block_walk(ka, optimize_blocks, NULL, NULL);
1909 mark_irn_visited(ka);
1910 ARR_APP1 (ir_node *, in, ka);
1911 } else if (get_irn_op(ka) == op_Phi) {
1912 mark_irn_visited(ka);
1913 ARR_APP1 (ir_node *, in, ka);
1917 /* DEL_ARR_F(end->in); GL @@@ tut nicht ! */
1920 current_ir_graph = rem;
1925 * Called by walker of remove_critical_cf_edges().
1927 * Place an empty block to an edge between a blocks of multiple
1928 * predecessors and a block of multiple successors.
1931 * @param env Environment of walker. This field is unused and has
1934 static void walk_critical_cf_edges(ir_node *n, void *env) {
1936 ir_node *pre, *block, **in, *jmp;
1938 /* Block has multiple predecessors */
1939 if ((op_Block == get_irn_op(n)) &&
1940 (get_irn_arity(n) > 1)) {
1941 arity = get_irn_arity(n);
1943 if (n == get_irg_end_block(current_ir_graph))
1944 return; /* No use to add a block here. */
1946 for (i=0; i<arity; i++) {
1947 pre = get_irn_n(n, i);
1948 /* Predecessor has multiple successors. Insert new flow edge */
1949 if ((NULL != pre) &&
1950 (op_Proj == get_irn_op(pre)) &&
1951 op_Raise != get_irn_op(skip_Proj(pre))) {
1953 /* set predecessor array for new block */
1954 in = NEW_ARR_D (ir_node *, current_ir_graph->obst, 1);
1955 /* set predecessor of new block */
1957 block = new_Block(1, in);
1958 /* insert new jmp node to new block */
1959 switch_block(block);
1962 /* set successor of new block */
1963 set_irn_n(n, i, jmp);
1965 } /* predecessor has multiple successors */
1966 } /* for all predecessors */
1967 } /* n is a block */
1970 void remove_critical_cf_edges(ir_graph *irg) {
1971 if (get_opt_critical_edges())
1972 irg_walk_graph(irg, NULL, walk_critical_cf_edges, NULL);