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"
37 /* Defined in iropt.c */
38 pset *new_identities (void);
39 void del_identities (pset *value_table);
40 void add_identities (pset *value_table, ir_node *node);
42 /********************************************************************/
43 /* apply optimizations of iropt to all nodes. */
44 /********************************************************************/
46 static void init_link (ir_node *n, void *env) {
47 set_irn_link(n, NULL);
50 #if 0 /* Old version. Avoids Ids.
51 This is not necessary: we do a postwalk, and get_irn_n
52 removes ids anyways. So it's much cheaper to call the
53 optimization less often and use the exchange() algorithm. */
55 optimize_in_place_wrapper (ir_node *n, void *env) {
57 ir_node *optimized, *old;
59 irn_arity = get_irn_arity(n);
60 for (i = 0; i < irn_arity; i++) {
61 /* get_irn_n skips Id nodes, so comparison old != optimized does not
62 show all optimizations. Therefore always set new predecessor. */
63 old = get_irn_n(n, i);
64 optimized = optimize_in_place_2(old);
65 set_irn_n(n, i, optimized);
68 if (get_irn_op(n) == op_Block) {
69 optimized = optimize_in_place_2(n);
70 if (optimized != n) exchange (n, optimized);
75 optimize_in_place_wrapper (ir_node *n, void *env) {
76 ir_node *optimized = optimize_in_place_2(n);
77 if (optimized != n) exchange (n, optimized);
84 local_optimize_graph (ir_graph *irg) {
85 ir_graph *rem = current_ir_graph;
86 current_ir_graph = irg;
88 /* Handle graph state */
89 assert(get_irg_phase_state(irg) != phase_building);
90 if (get_opt_global_cse())
91 set_irg_pinned(current_ir_graph, floats);
92 if (get_irg_outs_state(current_ir_graph) == outs_consistent)
93 set_irg_outs_inconsistent(current_ir_graph);
94 if (get_irg_dom_state(current_ir_graph) == dom_consistent)
95 set_irg_dom_inconsistent(current_ir_graph);
97 /* Clean the value_table in irg for the cse. */
98 del_identities(irg->value_table);
99 irg->value_table = new_identities();
101 /* walk over the graph */
102 irg_walk(irg->end, init_link, optimize_in_place_wrapper, NULL);
104 current_ir_graph = rem;
107 /********************************************************************/
108 /* Routines for dead node elimination / copying garbage collection */
109 /* of the obstack. */
110 /********************************************************************/
112 /* Remeber the new node in the old node by using a field all nodes have. */
114 set_new_node (ir_node *old, ir_node *new)
119 /* Get this new node, before the old node is forgotton.*/
120 static INLINE ir_node *
121 get_new_node (ir_node * n)
126 /* We use the block_visited flag to mark that we have computed the
127 number of useful predecessors for this block.
128 Further we encode the new arity in this flag in the old blocks.
129 Remembering the arity is useful, as it saves a lot of pointer
130 accesses. This function is called for all Phi and Block nodes
133 compute_new_arity(ir_node *b) {
134 int i, res, irn_arity;
137 irg_v = get_irg_block_visited(current_ir_graph);
138 block_v = get_Block_block_visited(b);
139 if (block_v >= irg_v) {
140 /* we computed the number of preds for this block and saved it in the
142 return block_v - irg_v;
144 /* compute the number of good predecessors */
145 res = irn_arity = get_irn_arity(b);
146 for (i = 0; i < irn_arity; i++)
147 if (get_irn_opcode(get_irn_n(b, i)) == iro_Bad) res--;
148 /* save it in the flag. */
149 set_Block_block_visited(b, irg_v + res);
154 static INLINE void new_backedge_info(ir_node *n) {
155 switch(get_irn_opcode(n)) {
157 n->attr.block.cg_backedge = NULL;
158 n->attr.block.backedge = new_backedge_arr(current_ir_graph->obst, get_irn_arity(n));
161 n->attr.phi_backedge = new_backedge_arr(current_ir_graph->obst, get_irn_arity(n));
164 n->attr.filter.backedge = new_backedge_arr(current_ir_graph->obst, get_irn_arity(n));
170 /* Copies the node to the new obstack. The Ins of the new node point to
171 the predecessors on the old obstack. For block/phi nodes not all
172 predecessors might be copied. n->link points to the new node.
173 For Phi and Block nodes the function allocates in-arrays with an arity
174 only for useful predecessors. The arity is determined by counting
175 the non-bad predecessors of the block. */
177 copy_node (ir_node *n, void *env) {
181 /* The end node looses it's flexible in array. This doesn't matter,
182 as dead node elimination builds End by hand, inlineing doesn't use
184 //assert(n->op == op_End || ((_ARR_DESCR(n->in))->cookie != ARR_F_MAGIC));
186 if (get_irn_opcode(n) == iro_Block) {
188 new_arity = compute_new_arity(n);
189 n->attr.block.graph_arr = NULL;
191 block = get_nodes_Block(n);
192 if (get_irn_opcode(n) == iro_Phi) {
193 new_arity = compute_new_arity(block);
195 new_arity = get_irn_arity(n);
198 nn = new_ir_node(get_irn_dbg_info(n),
205 /* Copy the attributes. These might point to additional data. If this
206 was allocated on the old obstack the pointers now are dangling. This
207 frees e.g. the memory of the graph_arr allocated in new_immBlock. */
209 new_backedge_info(nn);
212 /* printf("\n old node: "); DDMSG2(n);
213 printf(" new node: "); DDMSG2(nn); */
217 /* Copies new predecessors of old node to new node remembered in link.
218 Spare the Bad predecessors of Phi and Block nodes. */
220 copy_preds (ir_node *n, void *env) {
224 nn = get_new_node(n);
226 /* printf("\n old node: "); DDMSG2(n);
227 printf(" new node: "); DDMSG2(nn);
228 printf(" arities: old: %d, new: %d\n", get_irn_arity(n), get_irn_arity(nn)); */
230 if (get_irn_opcode(n) == iro_Block) {
231 /* Don't copy Bad nodes. */
233 irn_arity = get_irn_arity(n);
234 for (i = 0; i < irn_arity; i++)
235 if (get_irn_opcode(get_irn_n(n, i)) != iro_Bad) {
236 set_irn_n (nn, j, get_new_node(get_irn_n(n, i)));
237 /*if (is_backedge(n, i)) set_backedge(nn, j);*/
240 /* repair the block visited flag from above misuse. Repair it in both
241 graphs so that the old one can still be used. */
242 set_Block_block_visited(nn, 0);
243 set_Block_block_visited(n, 0);
244 /* Local optimization could not merge two subsequent blocks if
245 in array contained Bads. Now it's possible.
246 We don't call optimize_in_place as it requires
247 that the fields in ir_graph are set properly. */
248 if ((get_opt_control_flow_straightening()) &&
249 (get_Block_n_cfgpreds(nn) == 1) &&
250 (get_irn_op(get_Block_cfgpred(nn, 0)) == op_Jmp))
251 exchange(nn, get_nodes_Block(get_Block_cfgpred(nn, 0)));
252 } else if (get_irn_opcode(n) == iro_Phi) {
253 /* Don't copy node if corresponding predecessor in block is Bad.
254 The Block itself should not be Bad. */
255 block = get_nodes_Block(n);
256 set_irn_n (nn, -1, get_new_node(block));
258 irn_arity = get_irn_arity(n);
259 for (i = 0; i < irn_arity; i++)
260 if (get_irn_opcode(get_irn_n(block, i)) != iro_Bad) {
261 set_irn_n (nn, j, get_new_node(get_irn_n(n, i)));
262 /*if (is_backedge(n, i)) set_backedge(nn, j);*/
265 /* If the pre walker reached this Phi after the post walker visited the
266 block block_visited is > 0. */
267 set_Block_block_visited(get_nodes_Block(n), 0);
268 /* Compacting the Phi's ins might generate Phis with only one
270 if (get_irn_arity(n) == 1)
271 exchange(n, get_irn_n(n, 0));
273 irn_arity = get_irn_arity(n);
274 for (i = -1; i < irn_arity; i++)
275 set_irn_n (nn, i, get_new_node(get_irn_n(n, i)));
277 /* Now the new node is complete. We can add it to the hash table for cse.
278 @@@ inlinening aborts if we identify End. Why? */
279 if(get_irn_op(nn) != op_End)
280 add_identities (current_ir_graph->value_table, nn);
283 /* Copies the graph recursively, compacts the keepalive of the end node. */
286 ir_node *oe, *ne; /* old end, new end */
287 ir_node *ka; /* keep alive */
290 oe = get_irg_end(current_ir_graph);
291 /* copy the end node by hand, allocate dynamic in array! */
292 ne = new_ir_node(get_irn_dbg_info(oe),
299 /* Copy the attributes. Well, there might be some in the future... */
301 set_new_node(oe, ne);
303 /* copy the live nodes */
304 irg_walk(get_nodes_Block(oe), copy_node, copy_preds, NULL);
305 /* copy_preds for the end node ... */
306 set_nodes_Block(ne, get_new_node(get_nodes_Block(oe)));
308 /** ... and now the keep alives. **/
309 /* First pick the not marked block nodes and walk them. We must pick these
310 first as else we will oversee blocks reachable from Phis. */
311 irn_arity = get_irn_arity(oe);
312 for (i = 0; i < irn_arity; i++) {
313 ka = get_irn_n(oe, i);
314 if ((get_irn_op(ka) == op_Block) &&
315 (get_irn_visited(ka) < get_irg_visited(current_ir_graph))) {
316 /* We must keep the block alive and copy everything reachable */
317 set_irg_visited(current_ir_graph, get_irg_visited(current_ir_graph)-1);
318 irg_walk(ka, copy_node, copy_preds, NULL);
319 add_End_keepalive(ne, get_new_node(ka));
323 /* Now pick the Phis. Here we will keep all! */
324 irn_arity = get_irn_arity(oe);
325 for (i = 0; i < irn_arity; i++) {
326 ka = get_irn_n(oe, i);
327 if ((get_irn_op(ka) == op_Phi)) {
328 if (get_irn_visited(ka) < get_irg_visited(current_ir_graph)) {
329 /* We didn't copy the Phi yet. */
330 set_irg_visited(current_ir_graph, get_irg_visited(current_ir_graph)-1);
331 irg_walk(ka, copy_node, copy_preds, NULL);
333 add_End_keepalive(ne, get_new_node(ka));
338 /* Copies the graph reachable from current_ir_graph->end to the obstack
339 in current_ir_graph and fixes the environment.
340 Then fixes the fields in current_ir_graph containing nodes of the
343 copy_graph_env (void) {
345 /* Not all nodes remembered in current_ir_graph might be reachable
346 from the end node. Assure their link is set to NULL, so that
347 we can test whether new nodes have been computed. */
348 set_irn_link(get_irg_frame (current_ir_graph), NULL);
349 set_irn_link(get_irg_globals(current_ir_graph), NULL);
350 set_irn_link(get_irg_args (current_ir_graph), NULL);
352 /* we use the block walk flag for removing Bads from Blocks ins. */
353 inc_irg_block_visited(current_ir_graph);
358 /* fix the fields in current_ir_graph */
359 old_end = get_irg_end(current_ir_graph);
360 set_irg_end (current_ir_graph, get_new_node(old_end));
362 set_irg_end_block (current_ir_graph, get_new_node(get_irg_end_block(current_ir_graph)));
363 if (get_irn_link(get_irg_frame(current_ir_graph)) == NULL) {
364 copy_node (get_irg_frame(current_ir_graph), NULL);
365 copy_preds(get_irg_frame(current_ir_graph), NULL);
367 if (get_irn_link(get_irg_globals(current_ir_graph)) == NULL) {
368 copy_node (get_irg_globals(current_ir_graph), NULL);
369 copy_preds(get_irg_globals(current_ir_graph), NULL);
371 if (get_irn_link(get_irg_args(current_ir_graph)) == NULL) {
372 copy_node (get_irg_args(current_ir_graph), NULL);
373 copy_preds(get_irg_args(current_ir_graph), NULL);
375 set_irg_start (current_ir_graph, get_new_node(get_irg_start(current_ir_graph)));
377 set_irg_start_block(current_ir_graph,
378 get_new_node(get_irg_start_block(current_ir_graph)));
379 set_irg_frame (current_ir_graph, get_new_node(get_irg_frame(current_ir_graph)));
380 set_irg_globals(current_ir_graph, get_new_node(get_irg_globals(current_ir_graph)));
381 set_irg_args (current_ir_graph, get_new_node(get_irg_args(current_ir_graph)));
382 if (get_irn_link(get_irg_bad(current_ir_graph)) == NULL) {
383 copy_node(get_irg_bad(current_ir_graph), NULL);
384 copy_preds(get_irg_bad(current_ir_graph), NULL);
386 set_irg_bad(current_ir_graph, get_new_node(get_irg_bad(current_ir_graph)));
387 /* GL removed: we need unknown with mode for analyses.
388 if (get_irn_link(get_irg_unknown(current_ir_graph)) == NULL) {
389 copy_node(get_irg_unknown(current_ir_graph), NULL);
390 copy_preds(get_irg_unknown(current_ir_graph), NULL);
392 set_irg_unknown(current_ir_graph, get_new_node(get_irg_unknown(current_ir_graph)));
396 /* Copies all reachable nodes to a new obstack. Removes bad inputs
397 from block nodes and the corresponding inputs from Phi nodes.
398 Merges single exit blocks with single entry blocks and removes
400 Adds all new nodes to a new hash table for cse. Does not
401 perform cse, so the hash table might contain common subexpressions. */
402 /* Amroq call this emigrate() */
404 dead_node_elimination(ir_graph *irg) {
406 struct obstack *graveyard_obst = NULL;
407 struct obstack *rebirth_obst = NULL;
409 /* Remember external state of current_ir_graph. */
410 rem = current_ir_graph;
411 current_ir_graph = irg;
413 /* Handle graph state */
414 assert(get_irg_phase_state(current_ir_graph) != phase_building);
415 assert(get_irg_callee_info_state(current_ir_graph) == irg_callee_info_none);
416 free_outs(current_ir_graph);
418 /* @@@ so far we loose loops when copying */
419 set_irg_loop(current_ir_graph, NULL);
421 if (get_optimize() && get_opt_dead_node_elimination()) {
423 /* A quiet place, where the old obstack can rest in peace,
424 until it will be cremated. */
425 graveyard_obst = irg->obst;
427 /* A new obstack, where the reachable nodes will be copied to. */
428 rebirth_obst = (struct obstack *) xmalloc (sizeof (struct obstack));
429 current_ir_graph->obst = rebirth_obst;
430 obstack_init (current_ir_graph->obst);
432 /* We also need a new hash table for cse */
433 del_identities (irg->value_table);
434 irg->value_table = new_identities ();
436 /* Copy the graph from the old to the new obstack */
439 /* Free memory from old unoptimized obstack */
440 obstack_free(graveyard_obst, 0); /* First empty the obstack ... */
441 xfree (graveyard_obst); /* ... then free it. */
444 current_ir_graph = rem;
447 /* Relink bad predeseccors of a block and store the old in array to the
448 link field. This function is called by relink_bad_predecessors().
449 The array of link field starts with the block operand at position 0.
450 If block has bad predecessors, create a new in array without bad preds.
451 Otherwise let in array untouched. */
452 static void relink_bad_block_predecessors(ir_node *n, void *env) {
453 ir_node **new_in, *irn;
454 int i, new_irn_n, old_irn_arity, new_irn_arity = 0;
456 /* if link field of block is NULL, look for bad predecessors otherwise
457 this is allready done */
458 if (get_irn_op(n) == op_Block &&
459 get_irn_link(n) == NULL) {
461 /* save old predecessors in link field (position 0 is the block operand)*/
462 set_irn_link(n, (void *)get_irn_in(n));
464 /* count predecessors without bad nodes */
465 old_irn_arity = get_irn_arity(n);
466 for (i = 0; i < old_irn_arity; i++)
467 if (!is_Bad(get_irn_n(n, i))) new_irn_arity++;
469 /* arity changing: set new predecessors without bad nodes */
470 if (new_irn_arity < old_irn_arity) {
471 /* get new predecessor array without Block predecessor */
472 new_in = NEW_ARR_D (ir_node *, current_ir_graph->obst, (new_irn_arity+1));
474 /* set new predeseccors in array */
477 for (i = 1; i < old_irn_arity; i++) {
478 irn = get_irn_n(n, i);
479 if (!is_Bad(irn)) new_in[new_irn_n++] = irn;
482 } /* ir node has bad predecessors */
484 } /* Block is not relinked */
487 /* Relinks Bad predecesors from Bocks and Phis called by walker
488 remove_bad_predecesors(). If n is a Block, call
489 relink_bad_block_redecessors(). If n is a Phinode, call also the relinking
490 function of Phi's Block. If this block has bad predecessors, relink preds
492 static void relink_bad_predecessors(ir_node *n, void *env) {
493 ir_node *block, **old_in;
494 int i, old_irn_arity, new_irn_arity;
496 /* relink bad predeseccors of a block */
497 if (get_irn_op(n) == op_Block)
498 relink_bad_block_predecessors(n, env);
500 /* If Phi node relink its block and its predecessors */
501 if (get_irn_op(n) == op_Phi) {
503 /* Relink predeseccors of phi's block */
504 block = get_nodes_Block(n);
505 if (get_irn_link(block) == NULL)
506 relink_bad_block_predecessors(block, env);
508 old_in = (ir_node **)get_irn_link(block); /* Of Phi's Block */
509 old_irn_arity = ARR_LEN(old_in);
511 /* Relink Phi predeseccors if count of predeseccors changed */
512 if (old_irn_arity != ARR_LEN(get_irn_in(block))) {
513 /* set new predeseccors in array
514 n->in[0] remains the same block */
516 for(i = 1; i < old_irn_arity; i++)
517 if (!is_Bad((ir_node *)old_in[i])) n->in[new_irn_arity++] = n->in[i];
519 ARR_SETLEN(ir_node *, n->in, new_irn_arity);
522 } /* n is a Phi node */
525 /* Removes Bad Bad predecesors from Blocks and the corresponding
526 inputs to Phi nodes as in dead_node_elimination but without
528 On walking up set the link field to NULL, on walking down call
529 relink_bad_predecessors() (This function stores the old in array
530 to the link field and sets a new in array if arity of predecessors
532 void remove_bad_predecessors(ir_graph *irg) {
533 irg_walk_graph(irg, init_link, relink_bad_predecessors, NULL);
537 /**********************************************************************/
538 /* Funcionality for inlining */
539 /**********************************************************************/
541 /* Copy node for inlineing. Updates attributes that change when
542 * inlineing but not for dead node elimination.
544 * Copies the node by calling copy_node and then updates the entity if
545 * it's a local one. env must be a pointer of the frame type of the
546 * inlined procedure. The new entities must be in the link field of
549 copy_node_inline (ir_node *n, void *env) {
551 type *frame_tp = (type *)env;
554 if (get_irn_op(n) == op_Sel) {
555 new = get_new_node (n);
556 assert(get_irn_op(new) == op_Sel);
557 if (get_entity_owner(get_Sel_entity(n)) == frame_tp) {
558 set_Sel_entity(new, get_entity_link(get_Sel_entity(n)));
560 } else if (get_irn_op(n) == op_Block) {
561 new = get_new_node (n);
562 new->attr.block.irg = current_ir_graph;
567 void inline_method(ir_node *call, ir_graph *called_graph) {
569 ir_node *post_call, *post_bl;
571 ir_node *end, *end_bl;
575 int arity, n_ret, n_exc, n_res, i, j, rem_opt, irn_arity;
576 int exc_handling; ir_node *proj;
579 if (!get_optimize() || !get_opt_inline()) return;
580 /* -- Turn off optimizations, this can cause problems when allocating new nodes. -- */
581 rem_opt = get_optimize();
584 /* Handle graph state */
585 assert(get_irg_phase_state(current_ir_graph) != phase_building);
586 assert(get_irg_pinned(current_ir_graph) == pinned);
587 assert(get_irg_pinned(called_graph) == pinned);
588 if (get_irg_outs_state(current_ir_graph) == outs_consistent)
589 set_irg_outs_inconsistent(current_ir_graph);
591 /* -- Check preconditions -- */
592 assert(get_irn_op(call) == op_Call);
593 /* @@@ does not work for InterfaceIII.java after cgana
594 assert(get_Call_type(call) == get_entity_type(get_irg_ent(called_graph)));
595 assert(smaller_type(get_entity_type(get_irg_ent(called_graph)),
596 get_Call_type(call)));
598 assert(get_type_tpop(get_Call_type(call)) == type_method);
599 if (called_graph == current_ir_graph) {
600 set_optimize(rem_opt);
604 /* -- Decide how to handle exception control flow: Is there a handler
605 for the Call node, or do we branch directly to End on an exception?
606 exc_handling: 0 There is a handler.
608 2 Exception handling not represented in Firm. -- */
610 for (proj = (ir_node *)get_irn_link(call); proj; proj = (ir_node *)get_irn_link(proj)) {
611 assert(get_irn_op(proj) == op_Proj);
612 if (get_Proj_proj(proj) == pn_Call_M_except) { exc_handling = 0; break;}
613 if (get_Proj_proj(proj) == pn_Call_X_except) { exc_handling = 1; }
617 ir_node *proj, *Mproj = NULL, *Xproj = NULL;
618 for (proj = (ir_node *)get_irn_link(call); proj; proj = (ir_node *)get_irn_link(proj)) {
619 assert(get_irn_op(proj) == op_Proj);
620 if (get_Proj_proj(proj) == pn_Call_X_except) Xproj = proj;
621 if (get_Proj_proj(proj) == pn_Call_M_except) Mproj = proj;
623 if (Mproj) { assert(Xproj); exc_handling = 0; }
624 else if (Xproj) { exc_handling = 1; }
625 else { exc_handling = 2; }
630 the procedure and later replaces the Start node of the called graph.
631 Post_call is the old Call node and collects the results of the called
632 graph. Both will end up being a tuple. -- */
633 post_bl = get_nodes_Block(call);
634 set_irg_current_block(current_ir_graph, post_bl);
635 /* XxMxPxP of Start + parameter of Call */
637 in[1] = get_Call_mem(call);
638 in[2] = get_irg_frame(current_ir_graph);
639 in[3] = get_irg_globals(current_ir_graph);
640 in[4] = new_Tuple (get_Call_n_params(call), get_Call_param_arr(call));
641 pre_call = new_Tuple(5, in);
645 The new block gets the ins of the old block, pre_call and all its
646 predecessors and all Phi nodes. -- */
647 part_block(pre_call);
649 /* -- Prepare state for dead node elimination -- */
650 /* Visited flags in calling irg must be >= flag in called irg.
651 Else walker and arity computation will not work. */
652 if (get_irg_visited(current_ir_graph) <= get_irg_visited(called_graph))
653 set_irg_visited(current_ir_graph, get_irg_visited(called_graph)+1);
654 if (get_irg_block_visited(current_ir_graph)< get_irg_block_visited(called_graph))
655 set_irg_block_visited(current_ir_graph, get_irg_block_visited(called_graph));
656 /* Set pre_call as new Start node in link field of the start node of
657 calling graph and pre_calls block as new block for the start block
659 Further mark these nodes so that they are not visited by the
661 set_irn_link(get_irg_start(called_graph), pre_call);
662 set_irn_visited(get_irg_start(called_graph),
663 get_irg_visited(current_ir_graph));
664 set_irn_link(get_irg_start_block(called_graph),
665 get_nodes_Block(pre_call));
666 set_irn_visited(get_irg_start_block(called_graph),
667 get_irg_visited(current_ir_graph));
669 /* Initialize for compaction of in arrays */
670 inc_irg_block_visited(current_ir_graph);
672 /* -- Replicate local entities of the called_graph -- */
673 /* copy the entities. */
674 called_frame = get_irg_frame_type(called_graph);
675 for (i = 0; i < get_class_n_members(called_frame); i++) {
676 entity *new_ent, *old_ent;
677 old_ent = get_class_member(called_frame, i);
678 new_ent = copy_entity_own(old_ent, get_cur_frame_type());
679 set_entity_link(old_ent, new_ent);
682 /* visited is > than that of called graph. With this trick visited will
683 remain unchanged so that an outer walker, e.g., searching the call nodes
684 to inline, calling this inline will not visit the inlined nodes. */
685 set_irg_visited(current_ir_graph, get_irg_visited(current_ir_graph)-1);
687 /* -- Performing dead node elimination inlines the graph -- */
688 /* Copies the nodes to the obstack of current_ir_graph. Updates links to new
690 /* @@@ endless loops are not copied!! -- they should be, I think... */
691 irg_walk(get_irg_end(called_graph), copy_node_inline, copy_preds,
692 get_irg_frame_type(called_graph));
694 /* Repair called_graph */
695 set_irg_visited(called_graph, get_irg_visited(current_ir_graph));
696 set_irg_block_visited(called_graph, get_irg_block_visited(current_ir_graph));
697 set_Block_block_visited(get_irg_start_block(called_graph), 0);
699 /* -- Merge the end of the inlined procedure with the call site -- */
700 /* We will turn the old Call node into a Tuple with the following
703 0: Phi of all Memories of Return statements.
704 1: Jmp from new Block that merges the control flow from all exception
705 predecessors of the old end block.
706 2: Tuple of all arguments.
707 3: Phi of Exception memories.
708 In case the old Call directly branches to End on an exception we don't
709 need the block merging all exceptions nor the Phi of the exception
713 /* -- Precompute some values -- */
714 end_bl = get_new_node(get_irg_end_block(called_graph));
715 end = get_new_node(get_irg_end(called_graph));
716 arity = get_irn_arity(end_bl); /* arity = n_exc + n_ret */
717 n_res = get_method_n_ress(get_Call_type(call));
719 res_pred = (ir_node **) malloc (n_res * sizeof (ir_node *));
720 cf_pred = (ir_node **) malloc (arity * sizeof (ir_node *));
722 set_irg_current_block(current_ir_graph, post_bl); /* just to make sure */
724 /* -- archive keepalives -- */
725 irn_arity = get_irn_arity(end);
726 for (i = 0; i < irn_arity; i++)
727 add_End_keepalive(get_irg_end(current_ir_graph), get_irn_n(end, i));
729 /* The new end node will die. We need not free as the in array is on the obstack:
730 copy_node only generated 'D' arrays. */
732 /* -- Replace Return nodes by Jump nodes. -- */
734 for (i = 0; i < arity; i++) {
736 ret = get_irn_n(end_bl, i);
737 if (get_irn_op(ret) == op_Return) {
738 cf_pred[n_ret] = new_r_Jmp(current_ir_graph, get_nodes_Block(ret));
742 set_irn_in(post_bl, n_ret, cf_pred);
744 /* -- Build a Tuple for all results of the method.
745 Add Phi node if there was more than one Return. -- */
746 turn_into_tuple(post_call, 4);
747 /* First the Memory-Phi */
749 for (i = 0; i < arity; i++) {
750 ret = get_irn_n(end_bl, i);
751 if (get_irn_op(ret) == op_Return) {
752 cf_pred[n_ret] = get_Return_mem(ret);
756 phi = new_Phi(n_ret, cf_pred, mode_M);
757 set_Tuple_pred(call, 0, phi);
758 /* Conserve Phi-list for further inlinings -- but might be optimized */
759 if (get_nodes_Block(phi) == post_bl) {
760 set_irn_link(phi, get_irn_link(post_bl));
761 set_irn_link(post_bl, phi);
763 /* Now the real results */
765 for (j = 0; j < n_res; j++) {
767 for (i = 0; i < arity; i++) {
768 ret = get_irn_n(end_bl, i);
769 if (get_irn_op(ret) == op_Return) {
770 cf_pred[n_ret] = get_Return_res(ret, j);
774 phi = new_Phi(n_ret, cf_pred, get_irn_mode(cf_pred[0]));
776 /* Conserve Phi-list for further inlinings -- but might be optimized */
777 if (get_nodes_Block(phi) == post_bl) {
778 set_irn_link(phi, get_irn_link(post_bl));
779 set_irn_link(post_bl, phi);
782 set_Tuple_pred(call, 2, new_Tuple(n_res, res_pred));
784 set_Tuple_pred(call, 2, new_Bad());
786 /* Finally the exception control flow.
787 We have two (three) possible situations:
788 First if the Call branches to an exception handler: We need to add a Phi node to
789 collect the memory containing the exception objects. Further we need
790 to add another block to get a correct representation of this Phi. To
791 this block we add a Jmp that resolves into the X output of the Call
792 when the Call is turned into a tuple.
793 Second the Call branches to End, the exception is not handled. Just
794 add all inlined exception branches to the End node.
795 Third: there is no Exception edge at all. Handle as case two. */
796 if (exc_handler == 0) {
798 for (i = 0; i < arity; i++) {
800 ret = get_irn_n(end_bl, i);
801 if (is_fragile_op(skip_Proj(ret)) || (get_irn_op(skip_Proj(ret)) == op_Raise)) {
802 cf_pred[n_exc] = ret;
807 new_Block(n_exc, cf_pred); /* watch it: current_block is changed! */
808 set_Tuple_pred(call, 1, new_Jmp());
809 /* The Phi for the memories with the exception objects */
811 for (i = 0; i < arity; i++) {
813 ret = skip_Proj(get_irn_n(end_bl, i));
814 if (get_irn_op(ret) == op_Call) {
815 cf_pred[n_exc] = new_r_Proj(current_ir_graph, get_nodes_Block(ret), ret, mode_M, 3);
817 } else if (is_fragile_op(ret)) {
818 /* We rely that all cfops have the memory output at the same position. */
819 cf_pred[n_exc] = new_r_Proj(current_ir_graph, get_nodes_Block(ret), ret, mode_M, 0);
821 } else if (get_irn_op(ret) == op_Raise) {
822 cf_pred[n_exc] = new_r_Proj(current_ir_graph, get_nodes_Block(ret), ret, mode_M, 1);
826 set_Tuple_pred(call, 3, new_Phi(n_exc, cf_pred, mode_M));
828 set_Tuple_pred(call, 1, new_Bad());
829 set_Tuple_pred(call, 3, new_Bad());
832 ir_node *main_end_bl;
833 int main_end_bl_arity;
836 /* assert(exc_handler == 1 || no exceptions. ) */
838 for (i = 0; i < arity; i++) {
839 ir_node *ret = get_irn_n(end_bl, i);
841 if (is_fragile_op(skip_Proj(ret)) || (get_irn_op(skip_Proj(ret)) == op_Raise)) {
842 cf_pred[n_exc] = ret;
846 main_end_bl = get_irg_end_block(current_ir_graph);
847 main_end_bl_arity = get_irn_arity(main_end_bl);
848 end_preds = (ir_node **) malloc ((n_exc + main_end_bl_arity) * sizeof (ir_node *));
850 for (i = 0; i < main_end_bl_arity; ++i)
851 end_preds[i] = get_irn_n(main_end_bl, i);
852 for (i = 0; i < n_exc; ++i)
853 end_preds[main_end_bl_arity + i] = cf_pred[i];
854 set_irn_in(main_end_bl, n_exc + main_end_bl_arity, end_preds);
855 set_Tuple_pred(call, 1, new_Bad());
856 set_Tuple_pred(call, 3, new_Bad());
862 #if 0 /* old. now better, correcter, faster implementation. */
864 /* -- If the exception control flow from the inlined Call directly
865 branched to the end block we now have the following control
866 flow predecessor pattern: ProjX -> Tuple -> Jmp. We must
867 remove the Jmp along with it's empty block and add Jmp's
868 predecessors as predecessors of this end block. No problem if
869 there is no exception, because then branches Bad to End which
871 @@@ can't we know this beforehand: by getting the Proj(1) from
872 the Call link list and checking whether it goes to Proj. */
873 /* find the problematic predecessor of the end block. */
874 end_bl = get_irg_end_block(current_ir_graph);
875 for (i = 0; i < get_Block_n_cfgpreds(end_bl); i++) {
876 cf_op = get_Block_cfgpred(end_bl, i);
877 if (get_irn_op(cf_op) == op_Proj) {
878 cf_op = get_Proj_pred(cf_op);
879 if ((get_irn_op(cf_op) == op_Tuple) && (cf_op == call)) {
880 // There are unoptimized tuples from inlineing before when no exc
881 assert(get_Proj_proj(get_Block_cfgpred(end_bl, i)) == pn_Call_X_except);
882 cf_op = get_Tuple_pred(cf_op, pn_Call_X_except);
883 assert(get_irn_op(cf_op) == op_Jmp);
889 if (i < get_Block_n_cfgpreds(end_bl)) {
890 bl = get_nodes_Block(cf_op);
891 arity = get_Block_n_cfgpreds(end_bl) + get_Block_n_cfgpreds(bl) - 1;
892 cf_pred = (ir_node **) malloc (arity * sizeof (ir_node *));
893 for (j = 0; j < i; j++)
894 cf_pred[j] = get_Block_cfgpred(end_bl, j);
895 for (j = j; j < i + get_Block_n_cfgpreds(bl); j++)
896 cf_pred[j] = get_Block_cfgpred(bl, j-i);
897 for (j = j; j < arity; j++)
898 cf_pred[j] = get_Block_cfgpred(end_bl, j-get_Block_n_cfgpreds(bl) +1);
899 set_irn_in(end_bl, arity, cf_pred);
901 // Remove the exception pred from post-call Tuple.
902 set_Tuple_pred(call, pn_Call_X_except, new_Bad());
907 /* -- Turn cse back on. -- */
908 set_optimize(rem_opt);
911 /********************************************************************/
912 /* Apply inlineing to small methods. */
913 /********************************************************************/
917 /* It makes no sense to inline too many calls in one procedure. Anyways,
918 I didn't get a version with NEW_ARR_F to run. */
919 #define MAX_INLINE 1024
921 /* given an Call node, returns the irg called. NULL if not
923 static ir_graph *get_call_called_irg(ir_node *call) {
926 ir_graph *called_irg = NULL;
928 assert(get_irn_op(call) == op_Call);
930 addr = get_Call_ptr(call);
931 if (get_irn_op(addr) == op_Const) {
932 /* Check whether the constant is the pointer to a compiled entity. */
933 tv = get_Const_tarval(addr);
934 if (tarval_to_entity(tv))
935 called_irg = get_entity_irg(tarval_to_entity(tv));
940 static void collect_calls(ir_node *call, void *env) {
942 ir_node **calls = (ir_node **)env;
945 ir_graph *called_irg;
947 if (get_irn_op(call) != op_Call) return;
949 addr = get_Call_ptr(call);
950 if (get_irn_op(addr) == op_Const) {
951 /* Check whether the constant is the pointer to a compiled entity. */
952 tv = get_Const_tarval(addr);
953 if (tarval_to_entity(tv)) {
954 called_irg = get_entity_irg(tarval_to_entity(tv));
955 if (called_irg && pos < MAX_INLINE) {
956 /* The Call node calls a locally defined method. Remember to inline. */
964 /* Inlines all small methods at call sites where the called address comes
965 from a Const node that references the entity representing the called
967 The size argument is a rough measure for the code size of the method:
968 Methods where the obstack containing the firm graph is smaller than
970 void inline_small_irgs(ir_graph *irg, int size) {
972 ir_node *calls[MAX_INLINE];
973 ir_graph *rem = current_ir_graph;
975 if (!(get_optimize() && get_opt_inline())) return;
977 current_ir_graph = irg;
978 /* Handle graph state */
979 assert(get_irg_phase_state(current_ir_graph) != phase_building);
980 assert(get_irg_callee_info_state(current_ir_graph) == irg_callee_info_none);
982 /* Find Call nodes to inline.
983 (We can not inline during a walk of the graph, as inlineing the same
984 method several times changes the visited flag of the walked graph:
985 after the first inlineing visited of the callee equals visited of
986 the caller. With the next inlineing both are increased.) */
988 irg_walk(get_irg_end(irg), NULL, collect_calls, (void *) calls);
990 if ((pos > 0) && (pos < MAX_INLINE)) {
991 /* There are calls to inline */
992 collect_phiprojs(irg);
993 for (i = 0; i < pos; i++) {
996 tv = get_Const_tarval(get_Call_ptr(calls[i]));
997 callee = get_entity_irg(tarval_to_entity(tv));
998 if ((_obstack_memory_used(callee->obst) - obstack_room(callee->obst)) < size) {
999 inline_method(calls[i], callee);
1004 current_ir_graph = rem;
1008 int n_nodes; /* Nodes in graph except Id, Tuple, Proj, Start, End */
1009 int n_nodes_orig; /* for statistics */
1010 eset *call_nodes; /* All call nodes in this graph */
1012 int n_call_nodes_orig; /* for statistics */
1013 int n_callers; /* Number of known graphs that call this graphs. */
1014 int n_callers_orig; /* for statistics */
1017 static inline_irg_env *new_inline_irg_env(void) {
1018 inline_irg_env *env = malloc(sizeof(inline_irg_env));
1019 env->n_nodes = -2; /* uncount Start, End */
1020 env->n_nodes_orig = -2; /* uncount Start, End */
1021 env->call_nodes = eset_create();
1022 env->n_call_nodes = 0;
1023 env->n_call_nodes_orig = 0;
1025 env->n_callers_orig = 0;
1029 static void free_inline_irg_env(inline_irg_env *env) {
1030 eset_destroy(env->call_nodes);
1034 static void collect_calls2(ir_node *call, void *env) {
1035 inline_irg_env *x = (inline_irg_env *)env;
1036 ir_op *op = get_irn_op(call);
1039 /* count nodes in irg */
1040 if (op != op_Proj && op != op_Tuple && op != op_Sync) {
1045 if (op != op_Call) return;
1047 /* collect all call nodes */
1048 eset_insert(x->call_nodes, (void *)call);
1050 x->n_call_nodes_orig++;
1052 /* count all static callers */
1053 callee = get_call_called_irg(call);
1055 ((inline_irg_env *)get_irg_link(callee))->n_callers++;
1056 ((inline_irg_env *)get_irg_link(callee))->n_callers_orig++;
1060 INLINE static int is_leave(ir_graph *irg) {
1061 return (((inline_irg_env *)get_irg_link(irg))->n_call_nodes == 0);
1064 INLINE static int is_smaller(ir_graph *callee, int size) {
1065 return (((inline_irg_env *)get_irg_link(callee))->n_nodes < size);
1069 /* Inlines small leave methods at call sites where the called address comes
1070 from a Const node that references the entity representing the called
1072 The size argument is a rough measure for the code size of the method:
1073 Methods where the obstack containing the firm graph is smaller than
1074 size are inlined. */
1075 void inline_leave_functions(int maxsize, int leavesize, int size) {
1076 inline_irg_env *env;
1077 int i, n_irgs = get_irp_n_irgs();
1078 ir_graph *rem = current_ir_graph;
1081 if (!(get_optimize() && get_opt_inline())) return;
1083 /* extend all irgs by a temporary datastructure for inlineing. */
1084 for (i = 0; i < n_irgs; ++i)
1085 set_irg_link(get_irp_irg(i), new_inline_irg_env());
1087 /* Precompute information in temporary datastructure. */
1088 for (i = 0; i < n_irgs; ++i) {
1089 current_ir_graph = get_irp_irg(i);
1090 assert(get_irg_phase_state(current_ir_graph) != phase_building);
1091 assert(get_irg_callee_info_state(current_ir_graph) == irg_callee_info_none);
1093 irg_walk(get_irg_end(current_ir_graph), NULL, collect_calls2,
1094 get_irg_link(current_ir_graph));
1095 env = (inline_irg_env *)get_irg_link(current_ir_graph);
1099 Inline leaves recursively -- we might construct new leaves. */
1101 while (did_inline) {
1102 //printf("iteration %d\n", itercnt++);
1104 for (i = 0; i < n_irgs; ++i) {
1107 int phiproj_computed = 0;
1109 current_ir_graph = get_irp_irg(i);
1110 env = (inline_irg_env *)get_irg_link(current_ir_graph);
1112 /* we can not walk and change a set, nor remove from it.
1114 walkset = env->call_nodes;
1115 env->call_nodes = eset_create();
1116 for (call = eset_first(walkset); call; call = eset_next(walkset)) {
1117 inline_irg_env *callee_env;
1118 ir_graph *callee = get_call_called_irg(call);
1120 if (env->n_nodes > maxsize) break;
1121 if (callee && is_leave(callee) && is_smaller(callee, leavesize)) {
1122 if (!phiproj_computed) {
1123 phiproj_computed = 1;
1124 collect_phiprojs(current_ir_graph);
1126 callee_env = (inline_irg_env *)get_irg_link(callee);
1127 // printf(" %s: Inlineing %s.\n", get_entity_name(get_irg_entity(current_ir_graph)),
1128 // get_entity_name(get_irg_entity(callee)));
1129 inline_method(call, callee);
1131 env->n_call_nodes--;
1132 eset_insert_all(env->call_nodes, callee_env->call_nodes);
1133 env->n_call_nodes += callee_env->n_call_nodes;
1134 env->n_nodes += callee_env->n_nodes;
1135 callee_env->n_callers--;
1137 eset_insert(env->call_nodes, call);
1140 eset_destroy(walkset);
1144 //printf("Non leaves\n");
1145 /* inline other small functions. */
1146 for (i = 0; i < n_irgs; ++i) {
1149 int phiproj_computed = 0;
1151 current_ir_graph = get_irp_irg(i);
1152 env = (inline_irg_env *)get_irg_link(current_ir_graph);
1154 /* we can not walk and change a set, nor remove from it.
1156 walkset = env->call_nodes;
1157 env->call_nodes = eset_create();
1158 for (call = eset_first(walkset); call; call = eset_next(walkset)) {
1159 inline_irg_env *callee_env;
1160 ir_graph *callee = get_call_called_irg(call);
1162 if (env->n_nodes > maxsize) break;
1163 if (callee && is_smaller(callee, size)) {
1164 if (!phiproj_computed) {
1165 phiproj_computed = 1;
1166 collect_phiprojs(current_ir_graph);
1168 callee_env = (inline_irg_env *)get_irg_link(callee);
1169 // printf(" %s: Inlineing %s.\n", get_entity_name(get_irg_entity(current_ir_graph)),
1170 // get_entity_name(get_irg_entity(callee)));
1171 inline_method(call, callee);
1173 env->n_call_nodes--;
1174 eset_insert_all(env->call_nodes, callee_env->call_nodes);
1175 env->n_call_nodes += callee_env->n_call_nodes;
1176 env->n_nodes += callee_env->n_nodes;
1177 callee_env->n_callers--;
1179 eset_insert(env->call_nodes, call);
1182 eset_destroy(walkset);
1185 for (i = 0; i < n_irgs; ++i) {
1186 current_ir_graph = get_irp_irg(i);
1188 env = (inline_irg_env *)get_irg_link(current_ir_graph);
1189 if ((env->n_call_nodes_orig != env->n_call_nodes) ||
1190 (env->n_callers_orig != env->n_callers))
1191 printf("Nodes:%3d ->%3d, calls:%3d ->%3d, callers:%3d ->%3d, -- %s\n",
1192 env->n_nodes_orig, env->n_nodes, env->n_call_nodes_orig, env->n_call_nodes,
1193 env->n_callers_orig, env->n_callers,
1194 get_entity_name(get_irg_entity(current_ir_graph)));
1196 free_inline_irg_env((inline_irg_env *)get_irg_link(current_ir_graph));
1199 current_ir_graph = rem;
1202 /********************************************************************/
1203 /* Code Placement. Pinns all floating nodes to a block where they */
1204 /* will be executed only if needed. */
1205 /********************************************************************/
1207 static pdeq *worklist; /* worklist of ir_node*s */
1209 /* Find the earliest correct block for N. --- Place N into the
1210 same Block as its dominance-deepest Input. */
1212 place_floats_early (ir_node *n)
1214 int i, start, irn_arity;
1216 /* we must not run into an infinite loop */
1217 assert (irn_not_visited(n));
1218 mark_irn_visited(n);
1220 /* Place floating nodes. */
1221 if (get_op_pinned(get_irn_op(n)) == floats) {
1223 ir_node *b = new_Bad(); /* The block to place this node in */
1225 assert(get_irn_op(n) != op_Block);
1227 if ((get_irn_op(n) == op_Const) ||
1228 (get_irn_op(n) == op_SymConst) ||
1230 (get_irn_op(n) == op_Unknown)) {
1231 /* These nodes will not be placed by the loop below. */
1232 b = get_irg_start_block(current_ir_graph);
1236 /* find the block for this node. */
1237 irn_arity = get_irn_arity(n);
1238 for (i = 0; i < irn_arity; i++) {
1239 ir_node *dep = get_irn_n(n, i);
1241 if ((irn_not_visited(dep)) &&
1242 (get_op_pinned(get_irn_op(dep)) == floats)) {
1243 place_floats_early (dep);
1245 /* Because all loops contain at least one pinned node, now all
1246 our inputs are either pinned or place_early has already
1247 been finished on them. We do not have any unfinished inputs! */
1248 dep_block = get_nodes_Block(dep);
1249 if ((!is_Bad(dep_block)) &&
1250 (get_Block_dom_depth(dep_block) > depth)) {
1252 depth = get_Block_dom_depth(dep_block);
1254 /* Avoid that the node is placed in the Start block */
1255 if ((depth == 1) && (get_Block_dom_depth(get_nodes_Block(n)) > 1)) {
1256 b = get_Block_cfg_out(get_irg_start_block(current_ir_graph), 0);
1257 assert(b != get_irg_start_block(current_ir_graph));
1261 set_nodes_Block(n, b);
1264 /* Add predecessors of non floating nodes on worklist. */
1265 start = (get_irn_op(n) == op_Block) ? 0 : -1;
1266 irn_arity = get_irn_arity(n);
1267 for (i = start; i < irn_arity; i++) {
1268 ir_node *pred = get_irn_n(n, i);
1269 if (irn_not_visited(pred)) {
1270 pdeq_putr (worklist, pred);
1275 /* Floating nodes form subgraphs that begin at nodes as Const, Load,
1276 Start, Call and end at pinned nodes as Store, Call. Place_early
1277 places all floating nodes reachable from its argument through floating
1278 nodes and adds all beginnings at pinned nodes to the worklist. */
1279 static INLINE void place_early (void) {
1281 inc_irg_visited(current_ir_graph);
1283 /* this inits the worklist */
1284 place_floats_early (get_irg_end(current_ir_graph));
1286 /* Work the content of the worklist. */
1287 while (!pdeq_empty (worklist)) {
1288 ir_node *n = pdeq_getl (worklist);
1289 if (irn_not_visited(n)) place_floats_early (n);
1292 set_irg_outs_inconsistent(current_ir_graph);
1293 current_ir_graph->pinned = pinned;
1297 /* deepest common dominance ancestor of DCA and CONSUMER of PRODUCER */
1299 consumer_dom_dca (ir_node *dca, ir_node *consumer, ir_node *producer)
1301 ir_node *block = NULL;
1303 /* Compute the latest block into which we can place a node so that it is
1305 if (get_irn_op(consumer) == op_Phi) {
1306 /* our comsumer is a Phi-node, the effective use is in all those
1307 blocks through which the Phi-node reaches producer */
1309 ir_node *phi_block = get_nodes_Block(consumer);
1310 irn_arity = get_irn_arity(consumer);
1311 for (i = 0; i < irn_arity; i++) {
1312 if (get_irn_n(consumer, i) == producer) {
1313 block = get_nodes_Block(get_Block_cfgpred(phi_block, i));
1317 assert(is_no_Block(consumer));
1318 block = get_nodes_Block(consumer);
1321 /* Compute the deepest common ancestor of block and dca. */
1323 if (!dca) return block;
1324 while (get_Block_dom_depth(block) > get_Block_dom_depth(dca))
1325 block = get_Block_idom(block);
1326 while (get_Block_dom_depth(dca) > get_Block_dom_depth(block))
1327 dca = get_Block_idom(dca);
1328 while (block != dca)
1329 { block = get_Block_idom(block); dca = get_Block_idom(dca); }
1334 static INLINE int get_irn_loop_depth(ir_node *n) {
1335 return get_loop_depth(get_irn_loop(n));
1338 /* Move n to a block with less loop depth than it's current block. The
1339 new block must be dominated by early. */
1341 move_out_of_loops (ir_node *n, ir_node *early)
1343 ir_node *best, *dca;
1347 /* Find the region deepest in the dominator tree dominating
1348 dca with the least loop nesting depth, but still dominated
1349 by our early placement. */
1350 dca = get_nodes_Block(n);
1352 while (dca != early) {
1353 dca = get_Block_idom(dca);
1354 if (!dca) break; /* should we put assert(dca)? */
1355 if (get_irn_loop_depth(dca) < get_irn_loop_depth(best)) {
1359 if (best != get_nodes_Block(n)) {
1361 printf("Moving out of loop: "); DDMN(n);
1362 printf(" Outermost block: "); DDMN(early);
1363 printf(" Best block: "); DDMN(best);
1364 printf(" Innermost block: "); DDMN(get_nodes_Block(n));
1366 set_nodes_Block(n, best);
1370 /* Find the latest legal block for N and place N into the
1371 `optimal' Block between the latest and earliest legal block.
1372 The `optimal' block is the dominance-deepest block of those
1373 with the least loop-nesting-depth. This places N out of as many
1374 loops as possible and then makes it as controldependant as
1377 place_floats_late (ir_node *n)
1382 assert (irn_not_visited(n)); /* no multiple placement */
1384 /* no need to place block nodes, control nodes are already placed. */
1385 if ((get_irn_op(n) != op_Block) &&
1387 (get_irn_mode(n) != mode_X)) {
1388 /* Remember the early palacement of this block to move it
1389 out of loop no further than the early placement. */
1390 early = get_nodes_Block(n);
1391 /* Assure that our users are all placed, except the Phi-nodes.
1392 --- Each dataflow cycle contains at least one Phi-node. We
1393 have to break the `user has to be placed before the
1394 producer' dependance cycle and the Phi-nodes are the
1395 place to do so, because we need to base our placement on the
1396 final region of our users, which is OK with Phi-nodes, as they
1397 are pinned, and they never have to be placed after a
1398 producer of one of their inputs in the same block anyway. */
1399 for (i = 0; i < get_irn_n_outs(n); i++) {
1400 ir_node *succ = get_irn_out(n, i);
1401 if (irn_not_visited(succ) && (get_irn_op(succ) != op_Phi))
1402 place_floats_late (succ);
1405 /* We have to determine the final block of this node... except for
1407 if ((get_op_pinned(get_irn_op(n)) == floats) &&
1408 (get_irn_op(n) != op_Const) &&
1409 (get_irn_op(n) != op_SymConst)) {
1410 ir_node *dca = NULL; /* deepest common ancestor in the
1411 dominator tree of all nodes'
1412 blocks depending on us; our final
1413 placement has to dominate DCA. */
1414 for (i = 0; i < get_irn_n_outs(n); i++) {
1415 dca = consumer_dom_dca (dca, get_irn_out(n, i), n);
1417 set_nodes_Block(n, dca);
1419 move_out_of_loops (n, early);
1423 mark_irn_visited(n);
1425 /* Add predecessors of all non-floating nodes on list. (Those of floating
1426 nodes are placeded already and therefore are marked.) */
1427 for (i = 0; i < get_irn_n_outs(n); i++) {
1428 if (irn_not_visited(get_irn_out(n, i))) {
1429 pdeq_putr (worklist, get_irn_out(n, i));
1434 static INLINE void place_late(void) {
1436 inc_irg_visited(current_ir_graph);
1438 /* This fills the worklist initially. */
1439 place_floats_late(get_irg_start_block(current_ir_graph));
1440 /* And now empty the worklist again... */
1441 while (!pdeq_empty (worklist)) {
1442 ir_node *n = pdeq_getl (worklist);
1443 if (irn_not_visited(n)) place_floats_late(n);
1447 void place_code(ir_graph *irg) {
1448 ir_graph *rem = current_ir_graph;
1449 current_ir_graph = irg;
1451 if (!(get_optimize() && get_opt_global_cse())) return;
1453 /* Handle graph state */
1454 assert(get_irg_phase_state(irg) != phase_building);
1455 if (get_irg_dom_state(irg) != dom_consistent)
1458 construct_backedges(irg);
1460 /* Place all floating nodes as early as possible. This guarantees
1461 a legal code placement. */
1462 worklist = new_pdeq ();
1465 /* place_early invalidates the outs, place_late needs them. */
1467 /* Now move the nodes down in the dominator tree. This reduces the
1468 unnecessary executions of the node. */
1471 set_irg_outs_inconsistent(current_ir_graph);
1472 del_pdeq (worklist);
1473 current_ir_graph = rem;
1478 /********************************************************************/
1479 /* Control flow optimization. */
1480 /* Removes Bad control flow predecessors and empty blocks. A block */
1481 /* is empty if it contains only a Jmp node. */
1482 /* Blocks can only be removed if they are not needed for the */
1483 /* semantics of Phi nodes. */
1484 /********************************************************************/
1486 /* Removes Tuples from Block control flow predecessors.
1487 Optimizes blocks with equivalent_node().
1488 Replaces n by Bad if n is unreachable control flow. */
1489 static void merge_blocks(ir_node *n, void *env) {
1491 set_irn_link(n, NULL);
1493 if (get_irn_op(n) == op_Block) {
1495 for (i = 0; i < get_Block_n_cfgpreds(n); i++)
1496 /* GL @@@ : is this possible? if (get_opt_normalize()) -- added, all tests go throug.
1497 A different order of optimizations might cause problems. */
1498 if (get_opt_normalize())
1499 set_Block_cfgpred(n, i, skip_Tuple(get_Block_cfgpred(n, i)));
1500 } else if (get_optimize() && (get_irn_mode(n) == mode_X)) {
1501 /* We will soon visit a block. Optimize it before visiting! */
1502 ir_node *b = get_nodes_Block(n);
1503 ir_node *new = equivalent_node(b);
1504 while (irn_not_visited(b) && (!is_Bad(new)) && (new != b)) {
1505 /* We would have to run gigo if new is bad, so we
1506 promote it directly below. */
1507 assert(((b == new) ||
1508 get_opt_control_flow_straightening() ||
1509 get_opt_control_flow_weak_simplification()) &&
1510 ("strange flag setting"));
1513 new = equivalent_node(b);
1515 /* GL @@@ get_opt_normalize hinzugefuegt, 5.5.2003 */
1516 if (is_Bad(new) && get_opt_normalize()) exchange (n, new_Bad());
1520 /* Collects all Phi nodes in link list of Block.
1521 Marks all blocks "block_visited" if they contain a node other
1523 static void collect_nodes(ir_node *n, void *env) {
1524 if (is_no_Block(n)) {
1525 ir_node *b = get_nodes_Block(n);
1527 if ((get_irn_op(n) == op_Phi)) {
1528 /* Collect Phi nodes to compact ins along with block's ins. */
1529 set_irn_link(n, get_irn_link(b));
1531 } else if (get_irn_op(n) != op_Jmp) { /* Check for non empty block. */
1532 mark_Block_block_visited(b);
1537 /* Returns true if pred is pred of block */
1538 static int is_pred_of(ir_node *pred, ir_node *b) {
1540 for (i = 0; i < get_Block_n_cfgpreds(b); i++) {
1541 ir_node *b_pred = get_nodes_Block(get_Block_cfgpred(b, i));
1542 if (b_pred == pred) return 1;
1547 static int test_whether_dispensable(ir_node *b, int pos) {
1548 int i, j, n_preds = 1;
1549 int dispensable = 1;
1550 ir_node *cfop = get_Block_cfgpred(b, pos);
1551 ir_node *pred = get_nodes_Block(cfop);
1553 if (get_Block_block_visited(pred) + 1
1554 < get_irg_block_visited(current_ir_graph)) {
1555 if (!get_optimize() || !get_opt_control_flow_strong_simplification()) {
1556 /* Mark block so that is will not be removed. */
1557 set_Block_block_visited(pred, get_irg_block_visited(current_ir_graph)-1);
1560 /* Seems to be empty. */
1561 if (!get_irn_link(b)) {
1562 /* There are no Phi nodes ==> dispensable. */
1563 n_preds = get_Block_n_cfgpreds(pred);
1565 /* b's pred blocks and pred's pred blocks must be pairwise disjunct.
1566 Work preds < pos as if they were already removed. */
1567 for (i = 0; i < pos; i++) {
1568 ir_node *b_pred = get_nodes_Block(get_Block_cfgpred(b, i));
1569 if (get_Block_block_visited(b_pred) + 1
1570 < get_irg_block_visited(current_ir_graph)) {
1571 for (j = 0; j < get_Block_n_cfgpreds(b_pred); j++) {
1572 ir_node *b_pred_pred = get_nodes_Block(get_Block_cfgpred(b_pred, j));
1573 if (is_pred_of(b_pred_pred, pred)) dispensable = 0;
1576 if (is_pred_of(b_pred, pred)) dispensable = 0;
1579 for (i = pos +1; i < get_Block_n_cfgpreds(b); i++) {
1580 ir_node *b_pred = get_nodes_Block(get_Block_cfgpred(b, i));
1581 if (is_pred_of(b_pred, pred)) dispensable = 0;
1584 set_Block_block_visited(pred, get_irg_block_visited(current_ir_graph)-1);
1587 n_preds = get_Block_n_cfgpreds(pred);
1595 static void optimize_blocks(ir_node *b, void *env) {
1596 int i, j, k, max_preds, n_preds;
1597 ir_node *pred, *phi;
1600 /* Count the number of predecessor if this block is merged with pred blocks
1603 for (i = 0; i < get_Block_n_cfgpreds(b); i++) {
1604 max_preds += test_whether_dispensable(b, i);
1606 in = (ir_node **) malloc(max_preds * sizeof(ir_node *));
1609 printf(" working on "); DDMN(b);
1610 for (i = 0; i < get_Block_n_cfgpreds(b); i++) {
1611 pred = get_nodes_Block(get_Block_cfgpred(b, i));
1612 if (is_Bad(get_Block_cfgpred(b, i))) {
1613 printf(" removing Bad %i\n ", i);
1614 } else if (get_Block_block_visited(pred) +1
1615 < get_irg_block_visited(current_ir_graph)) {
1616 printf(" removing pred %i ", i); DDMN(pred);
1617 } else { printf(" Nothing to do for "); DDMN(pred); }
1619 * end Debug output **/
1621 /** Fix the Phi nodes **/
1622 phi = get_irn_link(b);
1624 assert(get_irn_op(phi) == op_Phi);
1625 /* Find the new predecessors for the Phi */
1627 for (i = 0; i < get_Block_n_cfgpreds(b); i++) {
1628 pred = get_nodes_Block(get_Block_cfgpred(b, i));
1629 if (is_Bad(get_Block_cfgpred(b, i))) {
1631 } else if (get_Block_block_visited(pred) +1
1632 < get_irg_block_visited(current_ir_graph)) {
1633 /* It's an empty block and not yet visited. */
1634 ir_node *phi_pred = get_Phi_pred(phi, i);
1635 for (j = 0; j < get_Block_n_cfgpreds(pred); j++) {
1636 if (get_nodes_Block(phi_pred) == pred) {
1637 assert(get_irn_op(phi_pred) == op_Phi); /* Block is empty!! */
1638 in[n_preds] = get_Phi_pred(phi_pred, j);
1640 in[n_preds] = phi_pred;
1644 /* The Phi_pred node is replaced now if it is a Phi.
1645 In Schleifen kann offenbar der entfernte Phi Knoten legal verwendet werden.
1646 Daher muss der Phiknoten durch den neuen ersetzt werden.
1647 Weiter muss der alte Phiknoten entfernt werden (durch ersetzen oder
1648 durch einen Bad) damit er aus den keep_alive verschwinden kann.
1649 Man sollte also, falls keine Schleife vorliegt, exchange mit new_Bad
1651 if (get_nodes_Block(phi_pred) == pred) {
1652 /* remove the Phi as it might be kept alive. Further there
1653 might be other users. */
1654 exchange(phi_pred, phi); /* geht, ist aber doch semantisch falsch! Warum?? */
1657 in[n_preds] = get_Phi_pred(phi, i);
1662 set_irn_in(phi, n_preds, in);
1664 phi = get_irn_link(phi);
1668 This happens only if merge between loop backedge and single loop entry. **/
1669 for (k = 0; k < get_Block_n_cfgpreds(b); k++) {
1670 pred = get_nodes_Block(get_Block_cfgpred(b, k));
1671 if (get_Block_block_visited(pred) +1
1672 < get_irg_block_visited(current_ir_graph)) {
1673 phi = get_irn_link(pred);
1675 if (get_irn_op(phi) == op_Phi) {
1676 set_nodes_Block(phi, b);
1679 for (i = 0; i < k; i++) {
1680 pred = get_nodes_Block(get_Block_cfgpred(b, i));
1681 if (is_Bad(get_Block_cfgpred(b, i))) {
1683 } else if (get_Block_block_visited(pred) +1
1684 < get_irg_block_visited(current_ir_graph)) {
1685 /* It's an empty block and not yet visited. */
1686 for (j = 0; j < get_Block_n_cfgpreds(pred); j++) {
1687 /* @@@ Hier brauche ich Schleifeninformation!!! Kontrollflusskante
1688 muss Rueckwaertskante sein! (An allen vier in[n_preds] = phi
1689 Anweisungen.) Trotzdem tuts bisher!! */
1698 for (i = 0; i < get_Phi_n_preds(phi); i++) {
1699 in[n_preds] = get_Phi_pred(phi, i);
1702 for (i = k+1; i < get_Block_n_cfgpreds(b); i++) {
1703 pred = get_nodes_Block(get_Block_cfgpred(b, i));
1704 if (is_Bad(get_Block_cfgpred(b, i))) {
1706 } else if (get_Block_block_visited(pred) +1
1707 < get_irg_block_visited(current_ir_graph)) {
1708 /* It's an empty block and not yet visited. */
1709 for (j = 0; j < get_Block_n_cfgpreds(pred); j++) {
1718 set_irn_in(phi, n_preds, in);
1720 phi = get_irn_link(phi);
1725 /** Fix the block **/
1727 for (i = 0; i < get_Block_n_cfgpreds(b); i++) {
1728 pred = get_nodes_Block(get_Block_cfgpred(b, i));
1729 if (is_Bad(get_Block_cfgpred(b, i))) {
1731 } else if (get_Block_block_visited(pred) +1
1732 < get_irg_block_visited(current_ir_graph)) {
1733 /* It's an empty block and not yet visited. */
1734 assert(get_Block_n_cfgpreds(b) > 1);
1735 /* Else it should be optimized by equivalent_node. */
1736 for (j = 0; j < get_Block_n_cfgpreds(pred); j++) {
1737 in[n_preds] = get_Block_cfgpred(pred, j);
1740 /* Remove block as it might be kept alive. */
1741 exchange(pred, b/*new_Bad()*/);
1743 in[n_preds] = get_Block_cfgpred(b, i);
1747 set_irn_in(b, n_preds, in);
1751 void optimize_cf(ir_graph *irg) {
1754 ir_node *end = get_irg_end(irg);
1755 ir_graph *rem = current_ir_graph;
1756 current_ir_graph = irg;
1758 /* Handle graph state */
1759 assert(get_irg_phase_state(irg) != phase_building);
1760 if (get_irg_outs_state(current_ir_graph) == outs_consistent)
1761 set_irg_outs_inconsistent(current_ir_graph);
1762 if (get_irg_dom_state(current_ir_graph) == dom_consistent)
1763 set_irg_dom_inconsistent(current_ir_graph);
1765 /* Use block visited flag to mark non-empty blocks. */
1766 inc_irg_block_visited(irg);
1767 irg_walk(end, merge_blocks, collect_nodes, NULL);
1769 /* Optimize the standard code. */
1770 irg_block_walk(get_irg_end_block(irg), optimize_blocks, NULL, NULL);
1772 /* Walk all keep alives, optimize them if block, add to new in-array
1773 for end if useful. */
1774 in = NEW_ARR_F (ir_node *, 1);
1775 in[0] = get_nodes_Block(end);
1776 inc_irg_visited(current_ir_graph);
1777 for(i = 0; i < get_End_n_keepalives(end); i++) {
1778 ir_node *ka = get_End_keepalive(end, i);
1779 if (irn_not_visited(ka)) {
1780 if ((get_irn_op(ka) == op_Block) && Block_not_block_visited(ka)) {
1781 set_irg_block_visited(current_ir_graph, /* Don't walk all the way to Start. */
1782 get_irg_block_visited(current_ir_graph)-1);
1783 irg_block_walk(ka, optimize_blocks, NULL, NULL);
1784 mark_irn_visited(ka);
1785 ARR_APP1 (ir_node *, in, ka);
1786 } else if (get_irn_op(ka) == op_Phi) {
1787 mark_irn_visited(ka);
1788 ARR_APP1 (ir_node *, in, ka);
1792 /* DEL_ARR_F(end->in); GL @@@ tut nicht ! */
1795 current_ir_graph = rem;
1800 * Called by walker of remove_critical_cf_edges.
1802 * Place an empty block to an edge between a blocks of multiple
1803 * predecessors and a block of multiple sucessors.
1806 * @param env Envirnment of walker. This field is unused and has
1809 static void walk_critical_cf_edges(ir_node *n, void *env) {
1811 ir_node *pre, *block, **in, *jmp;
1813 /* Block has multiple predecessors */
1814 if ((op_Block == get_irn_op(n)) &&
1815 (get_irn_arity(n) > 1)) {
1816 arity = get_irn_arity(n);
1818 if (n == get_irg_end_block(current_ir_graph))
1819 return; // No use to add a block here.
1821 for (i=0; i<arity; i++) {
1822 pre = get_irn_n(n, i);
1823 /* Predecessor has multiple sucessors. Insert new flow edge */
1824 if ((NULL != pre) &&
1825 (op_Proj == get_irn_op(pre)) &&
1826 op_Raise != get_irn_op(skip_Proj(pre))) {
1828 /* set predecessor array for new block */
1829 in = NEW_ARR_D (ir_node *, current_ir_graph->obst, 1);
1830 /* set predecessor of new block */
1832 block = new_Block(1, in);
1833 /* insert new jmp node to new block */
1834 switch_block(block);
1837 /* set sucessor of new block */
1838 set_irn_n(n, i, jmp);
1840 } /* predecessor has multiple sucessors */
1841 } /* for all predecessors */
1842 } /* n is a block */
1845 void remove_critical_cf_edges(ir_graph *irg) {
1846 if (get_opt_critical_edges())
1847 irg_walk_graph(irg, NULL, walk_critical_cf_edges, NULL);