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"
38 /* Defined in iropt.c */
39 pset *new_identities (void);
40 void del_identities (pset *value_table);
41 void add_identities (pset *value_table, ir_node *node);
43 /********************************************************************/
44 /* apply optimizations of iropt to all nodes. */
45 /********************************************************************/
47 static void init_link (ir_node *n, void *env) {
48 set_irn_link(n, NULL);
51 #if 0 /* Old version. Avoids Ids.
52 This is not necessary: we do a postwalk, and get_irn_n
53 removes ids anyways. So it's much cheaper to call the
54 optimization less often and use the exchange() algorithm. */
56 optimize_in_place_wrapper (ir_node *n, void *env) {
58 ir_node *optimized, *old;
60 irn_arity = get_irn_arity(n);
61 for (i = 0; i < irn_arity; i++) {
62 /* get_irn_n skips Id nodes, so comparison old != optimized does not
63 show all optimizations. Therefore always set new predecessor. */
64 old = get_irn_n(n, i);
65 optimized = optimize_in_place_2(old);
66 set_irn_n(n, i, optimized);
69 if (get_irn_op(n) == op_Block) {
70 optimized = optimize_in_place_2(n);
71 if (optimized != n) exchange (n, optimized);
76 optimize_in_place_wrapper (ir_node *n, void *env) {
77 ir_node *optimized = optimize_in_place_2(n);
78 if (optimized != n) exchange (n, optimized);
85 local_optimize_graph (ir_graph *irg) {
86 ir_graph *rem = current_ir_graph;
87 current_ir_graph = irg;
89 /* Handle graph state */
90 assert(get_irg_phase_state(irg) != phase_building);
91 if (get_opt_global_cse())
92 set_irg_pinned(current_ir_graph, floats);
93 if (get_irg_outs_state(current_ir_graph) == outs_consistent)
94 set_irg_outs_inconsistent(current_ir_graph);
95 if (get_irg_dom_state(current_ir_graph) == dom_consistent)
96 set_irg_dom_inconsistent(current_ir_graph);
98 /* Clean the value_table in irg for the cse. */
99 del_identities(irg->value_table);
100 irg->value_table = new_identities();
102 /* walk over the graph */
103 irg_walk(irg->end, init_link, optimize_in_place_wrapper, NULL);
105 current_ir_graph = rem;
108 /********************************************************************/
109 /* Routines for dead node elimination / copying garbage collection */
110 /* of the obstack. */
111 /********************************************************************/
113 /* Remeber the new node in the old node by using a field all nodes have. */
115 set_new_node (ir_node *old, ir_node *new)
120 /* Get this new node, before the old node is forgotton.*/
121 static INLINE ir_node *
122 get_new_node (ir_node * n)
127 /* We use the block_visited flag to mark that we have computed the
128 number of useful predecessors for this block.
129 Further we encode the new arity in this flag in the old blocks.
130 Remembering the arity is useful, as it saves a lot of pointer
131 accesses. This function is called for all Phi and Block nodes
134 compute_new_arity(ir_node *b) {
135 int i, res, irn_arity;
138 irg_v = get_irg_block_visited(current_ir_graph);
139 block_v = get_Block_block_visited(b);
140 if (block_v >= irg_v) {
141 /* we computed the number of preds for this block and saved it in the
143 return block_v - irg_v;
145 /* compute the number of good predecessors */
146 res = irn_arity = get_irn_arity(b);
147 for (i = 0; i < irn_arity; i++)
148 if (get_irn_opcode(get_irn_n(b, i)) == iro_Bad) res--;
149 /* save it in the flag. */
150 set_Block_block_visited(b, irg_v + res);
155 static INLINE void new_backedge_info(ir_node *n) {
156 switch(get_irn_opcode(n)) {
158 n->attr.block.cg_backedge = NULL;
159 n->attr.block.backedge = new_backedge_arr(current_ir_graph->obst, get_irn_arity(n));
162 n->attr.phi_backedge = new_backedge_arr(current_ir_graph->obst, get_irn_arity(n));
165 n->attr.filter.backedge = new_backedge_arr(current_ir_graph->obst, get_irn_arity(n));
171 /* Copies the node to the new obstack. The Ins of the new node point to
172 the predecessors on the old obstack. For block/phi nodes not all
173 predecessors might be copied. n->link points to the new node.
174 For Phi and Block nodes the function allocates in-arrays with an arity
175 only for useful predecessors. The arity is determined by counting
176 the non-bad predecessors of the block. */
178 copy_node (ir_node *n, void *env) {
182 /* The end node looses it's flexible in array. This doesn't matter,
183 as dead node elimination builds End by hand, inlineing doesn't use
185 //assert(n->op == op_End || ((_ARR_DESCR(n->in))->cookie != ARR_F_MAGIC));
187 if (get_irn_opcode(n) == iro_Block) {
189 new_arity = compute_new_arity(n);
190 n->attr.block.graph_arr = NULL;
192 block = get_nodes_Block(n);
193 if (get_irn_opcode(n) == iro_Phi) {
194 new_arity = compute_new_arity(block);
196 new_arity = get_irn_arity(n);
199 nn = new_ir_node(get_irn_dbg_info(n),
206 /* Copy the attributes. These might point to additional data. If this
207 was allocated on the old obstack the pointers now are dangling. This
208 frees e.g. the memory of the graph_arr allocated in new_immBlock. */
210 new_backedge_info(nn);
213 /* printf("\n old node: "); DDMSG2(n);
214 printf(" new node: "); DDMSG2(nn); */
218 /* Copies new predecessors of old node to new node remembered in link.
219 Spare the Bad predecessors of Phi and Block nodes. */
221 copy_preds (ir_node *n, void *env) {
225 nn = get_new_node(n);
227 /* printf("\n old node: "); DDMSG2(n);
228 printf(" new node: "); DDMSG2(nn);
229 printf(" arities: old: %d, new: %d\n", get_irn_arity(n), get_irn_arity(nn)); */
231 if (get_irn_opcode(n) == iro_Block) {
232 /* Don't copy Bad nodes. */
234 irn_arity = get_irn_arity(n);
235 for (i = 0; i < irn_arity; i++)
236 if (get_irn_opcode(get_irn_n(n, i)) != iro_Bad) {
237 set_irn_n (nn, j, get_new_node(get_irn_n(n, i)));
238 /*if (is_backedge(n, i)) set_backedge(nn, j);*/
241 /* repair the block visited flag from above misuse. Repair it in both
242 graphs so that the old one can still be used. */
243 set_Block_block_visited(nn, 0);
244 set_Block_block_visited(n, 0);
245 /* Local optimization could not merge two subsequent blocks if
246 in array contained Bads. Now it's possible.
247 We don't call optimize_in_place as it requires
248 that the fields in ir_graph are set properly. */
249 if ((get_opt_control_flow_straightening()) &&
250 (get_Block_n_cfgpreds(nn) == 1) &&
251 (get_irn_op(get_Block_cfgpred(nn, 0)) == op_Jmp))
252 exchange(nn, get_nodes_Block(get_Block_cfgpred(nn, 0)));
253 } else if (get_irn_opcode(n) == iro_Phi) {
254 /* Don't copy node if corresponding predecessor in block is Bad.
255 The Block itself should not be Bad. */
256 block = get_nodes_Block(n);
257 set_irn_n (nn, -1, get_new_node(block));
259 irn_arity = get_irn_arity(n);
260 for (i = 0; i < irn_arity; i++)
261 if (get_irn_opcode(get_irn_n(block, i)) != iro_Bad) {
262 set_irn_n (nn, j, get_new_node(get_irn_n(n, i)));
263 /*if (is_backedge(n, i)) set_backedge(nn, j);*/
266 /* If the pre walker reached this Phi after the post walker visited the
267 block block_visited is > 0. */
268 set_Block_block_visited(get_nodes_Block(n), 0);
269 /* Compacting the Phi's ins might generate Phis with only one
271 if (get_irn_arity(n) == 1)
272 exchange(n, get_irn_n(n, 0));
274 irn_arity = get_irn_arity(n);
275 for (i = -1; i < irn_arity; i++)
276 set_irn_n (nn, i, get_new_node(get_irn_n(n, i)));
278 /* Now the new node is complete. We can add it to the hash table for cse.
279 @@@ inlinening aborts if we identify End. Why? */
280 if(get_irn_op(nn) != op_End)
281 add_identities (current_ir_graph->value_table, nn);
284 /* Copies the graph recursively, compacts the keepalive of the end node. */
287 ir_node *oe, *ne; /* old end, new end */
288 ir_node *ka; /* keep alive */
291 oe = get_irg_end(current_ir_graph);
292 /* copy the end node by hand, allocate dynamic in array! */
293 ne = new_ir_node(get_irn_dbg_info(oe),
300 /* Copy the attributes. Well, there might be some in the future... */
302 set_new_node(oe, ne);
304 /* copy the live nodes */
305 irg_walk(get_nodes_Block(oe), copy_node, copy_preds, NULL);
306 /* copy_preds for the end node ... */
307 set_nodes_Block(ne, get_new_node(get_nodes_Block(oe)));
309 /** ... and now the keep alives. **/
310 /* First pick the not marked block nodes and walk them. We must pick these
311 first as else we will oversee blocks reachable from Phis. */
312 irn_arity = get_irn_arity(oe);
313 for (i = 0; i < irn_arity; i++) {
314 ka = get_irn_n(oe, i);
315 if ((get_irn_op(ka) == op_Block) &&
316 (get_irn_visited(ka) < get_irg_visited(current_ir_graph))) {
317 /* We must keep the block alive and copy everything reachable */
318 set_irg_visited(current_ir_graph, get_irg_visited(current_ir_graph)-1);
319 irg_walk(ka, copy_node, copy_preds, NULL);
320 add_End_keepalive(ne, get_new_node(ka));
324 /* Now pick the Phis. Here we will keep all! */
325 irn_arity = get_irn_arity(oe);
326 for (i = 0; i < irn_arity; i++) {
327 ka = get_irn_n(oe, i);
328 if ((get_irn_op(ka) == op_Phi)) {
329 if (get_irn_visited(ka) < get_irg_visited(current_ir_graph)) {
330 /* We didn't copy the Phi yet. */
331 set_irg_visited(current_ir_graph, get_irg_visited(current_ir_graph)-1);
332 irg_walk(ka, copy_node, copy_preds, NULL);
334 add_End_keepalive(ne, get_new_node(ka));
339 /* Copies the graph reachable from current_ir_graph->end to the obstack
340 in current_ir_graph and fixes the environment.
341 Then fixes the fields in current_ir_graph containing nodes of the
344 copy_graph_env (void) {
346 /* Not all nodes remembered in current_ir_graph might be reachable
347 from the end node. Assure their link is set to NULL, so that
348 we can test whether new nodes have been computed. */
349 set_irn_link(get_irg_frame (current_ir_graph), NULL);
350 set_irn_link(get_irg_globals(current_ir_graph), NULL);
351 set_irn_link(get_irg_args (current_ir_graph), NULL);
353 /* we use the block walk flag for removing Bads from Blocks ins. */
354 inc_irg_block_visited(current_ir_graph);
359 /* fix the fields in current_ir_graph */
360 old_end = get_irg_end(current_ir_graph);
361 set_irg_end (current_ir_graph, get_new_node(old_end));
363 set_irg_end_block (current_ir_graph, get_new_node(get_irg_end_block(current_ir_graph)));
364 if (get_irn_link(get_irg_frame(current_ir_graph)) == NULL) {
365 copy_node (get_irg_frame(current_ir_graph), NULL);
366 copy_preds(get_irg_frame(current_ir_graph), NULL);
368 if (get_irn_link(get_irg_globals(current_ir_graph)) == NULL) {
369 copy_node (get_irg_globals(current_ir_graph), NULL);
370 copy_preds(get_irg_globals(current_ir_graph), NULL);
372 if (get_irn_link(get_irg_args(current_ir_graph)) == NULL) {
373 copy_node (get_irg_args(current_ir_graph), NULL);
374 copy_preds(get_irg_args(current_ir_graph), NULL);
376 set_irg_start (current_ir_graph, get_new_node(get_irg_start(current_ir_graph)));
378 set_irg_start_block(current_ir_graph,
379 get_new_node(get_irg_start_block(current_ir_graph)));
380 set_irg_frame (current_ir_graph, get_new_node(get_irg_frame(current_ir_graph)));
381 set_irg_globals(current_ir_graph, get_new_node(get_irg_globals(current_ir_graph)));
382 set_irg_args (current_ir_graph, get_new_node(get_irg_args(current_ir_graph)));
383 if (get_irn_link(get_irg_bad(current_ir_graph)) == NULL) {
384 copy_node(get_irg_bad(current_ir_graph), NULL);
385 copy_preds(get_irg_bad(current_ir_graph), NULL);
387 set_irg_bad(current_ir_graph, get_new_node(get_irg_bad(current_ir_graph)));
388 /* GL removed: we need unknown with mode for analyses.
389 if (get_irn_link(get_irg_unknown(current_ir_graph)) == NULL) {
390 copy_node(get_irg_unknown(current_ir_graph), NULL);
391 copy_preds(get_irg_unknown(current_ir_graph), NULL);
393 set_irg_unknown(current_ir_graph, get_new_node(get_irg_unknown(current_ir_graph)));
397 /* Copies all reachable nodes to a new obstack. Removes bad inputs
398 from block nodes and the corresponding inputs from Phi nodes.
399 Merges single exit blocks with single entry blocks and removes
401 Adds all new nodes to a new hash table for cse. Does not
402 perform cse, so the hash table might contain common subexpressions. */
403 /* Amroq call this emigrate() */
405 dead_node_elimination(ir_graph *irg) {
407 struct obstack *graveyard_obst = NULL;
408 struct obstack *rebirth_obst = NULL;
410 /* Remember external state of current_ir_graph. */
411 rem = current_ir_graph;
412 current_ir_graph = irg;
414 /* Handle graph state */
415 assert(get_irg_phase_state(current_ir_graph) != phase_building);
416 assert(get_irg_callee_info_state(current_ir_graph) == irg_callee_info_none);
417 free_outs(current_ir_graph);
419 /* @@@ so far we loose loops when copying */
420 set_irg_loop(current_ir_graph, NULL);
422 if (get_opt_optimize() && get_opt_dead_node_elimination()) {
424 /* A quiet place, where the old obstack can rest in peace,
425 until it will be cremated. */
426 graveyard_obst = irg->obst;
428 /* A new obstack, where the reachable nodes will be copied to. */
429 rebirth_obst = (struct obstack *) xmalloc (sizeof (struct obstack));
430 current_ir_graph->obst = rebirth_obst;
431 obstack_init (current_ir_graph->obst);
433 /* We also need a new hash table for cse */
434 del_identities (irg->value_table);
435 irg->value_table = new_identities ();
437 /* Copy the graph from the old to the new obstack */
440 /* Free memory from old unoptimized obstack */
441 obstack_free(graveyard_obst, 0); /* First empty the obstack ... */
442 xfree (graveyard_obst); /* ... then free it. */
445 current_ir_graph = rem;
448 /* Relink bad predeseccors of a block and store the old in array to the
449 link field. This function is called by relink_bad_predecessors().
450 The array of link field starts with the block operand at position 0.
451 If block has bad predecessors, create a new in array without bad preds.
452 Otherwise let in array untouched. */
453 static void relink_bad_block_predecessors(ir_node *n, void *env) {
454 ir_node **new_in, *irn;
455 int i, new_irn_n, old_irn_arity, new_irn_arity = 0;
457 /* if link field of block is NULL, look for bad predecessors otherwise
458 this is allready done */
459 if (get_irn_op(n) == op_Block &&
460 get_irn_link(n) == NULL) {
462 /* save old predecessors in link field (position 0 is the block operand)*/
463 set_irn_link(n, (void *)get_irn_in(n));
465 /* count predecessors without bad nodes */
466 old_irn_arity = get_irn_arity(n);
467 for (i = 0; i < old_irn_arity; i++)
468 if (!is_Bad(get_irn_n(n, i))) new_irn_arity++;
470 /* arity changing: set new predecessors without bad nodes */
471 if (new_irn_arity < old_irn_arity) {
472 /* get new predecessor array without Block predecessor */
473 new_in = NEW_ARR_D (ir_node *, current_ir_graph->obst, (new_irn_arity+1));
475 /* set new predeseccors in array */
478 for (i = 1; i < old_irn_arity; i++) {
479 irn = get_irn_n(n, i);
480 if (!is_Bad(irn)) new_in[new_irn_n++] = irn;
483 } /* ir node has bad predecessors */
485 } /* Block is not relinked */
488 /* Relinks Bad predecesors from Bocks and Phis called by walker
489 remove_bad_predecesors(). If n is a Block, call
490 relink_bad_block_redecessors(). If n is a Phinode, call also the relinking
491 function of Phi's Block. If this block has bad predecessors, relink preds
493 static void relink_bad_predecessors(ir_node *n, void *env) {
494 ir_node *block, **old_in;
495 int i, old_irn_arity, new_irn_arity;
497 /* relink bad predeseccors of a block */
498 if (get_irn_op(n) == op_Block)
499 relink_bad_block_predecessors(n, env);
501 /* If Phi node relink its block and its predecessors */
502 if (get_irn_op(n) == op_Phi) {
504 /* Relink predeseccors of phi's block */
505 block = get_nodes_Block(n);
506 if (get_irn_link(block) == NULL)
507 relink_bad_block_predecessors(block, env);
509 old_in = (ir_node **)get_irn_link(block); /* Of Phi's Block */
510 old_irn_arity = ARR_LEN(old_in);
512 /* Relink Phi predeseccors if count of predeseccors changed */
513 if (old_irn_arity != ARR_LEN(get_irn_in(block))) {
514 /* set new predeseccors in array
515 n->in[0] remains the same block */
517 for(i = 1; i < old_irn_arity; i++)
518 if (!is_Bad((ir_node *)old_in[i])) n->in[new_irn_arity++] = n->in[i];
520 ARR_SETLEN(ir_node *, n->in, new_irn_arity);
523 } /* n is a Phi node */
526 /* Removes Bad Bad predecesors from Blocks and the corresponding
527 inputs to Phi nodes as in dead_node_elimination but without
529 On walking up set the link field to NULL, on walking down call
530 relink_bad_predecessors() (This function stores the old in array
531 to the link field and sets a new in array if arity of predecessors
533 void remove_bad_predecessors(ir_graph *irg) {
534 irg_walk_graph(irg, init_link, relink_bad_predecessors, NULL);
538 /**********************************************************************/
539 /* Funcionality for inlining */
540 /**********************************************************************/
542 /* Copy node for inlineing. Updates attributes that change when
543 * inlineing but not for dead node elimination.
545 * Copies the node by calling copy_node and then updates the entity if
546 * it's a local one. env must be a pointer of the frame type of the
547 * inlined procedure. The new entities must be in the link field of
550 copy_node_inline (ir_node *n, void *env) {
552 type *frame_tp = (type *)env;
555 if (get_irn_op(n) == op_Sel) {
556 new = get_new_node (n);
557 assert(get_irn_op(new) == op_Sel);
558 if (get_entity_owner(get_Sel_entity(n)) == frame_tp) {
559 set_Sel_entity(new, get_entity_link(get_Sel_entity(n)));
561 } else if (get_irn_op(n) == op_Block) {
562 new = get_new_node (n);
563 new->attr.block.irg = current_ir_graph;
568 void inline_method(ir_node *call, ir_graph *called_graph) {
570 ir_node *post_call, *post_bl;
572 ir_node *end, *end_bl;
576 int arity, n_ret, n_exc, n_res, i, j, rem_opt, irn_arity;
577 int exc_handling; ir_node *proj;
580 if (!get_opt_optimize() || !get_opt_inline() ||
581 (get_irg_inline_property(called_graph) == irg_inline_forbidden)) return;
583 /* -- Turn off optimizations, this can cause problems when allocating new nodes. -- */
584 rem_opt = get_opt_optimize();
587 /* Handle graph state */
588 assert(get_irg_phase_state(current_ir_graph) != phase_building);
589 assert(get_irg_pinned(current_ir_graph) == pinned);
590 assert(get_irg_pinned(called_graph) == pinned);
591 if (get_irg_outs_state(current_ir_graph) == outs_consistent)
592 set_irg_outs_inconsistent(current_ir_graph);
594 /* -- Check preconditions -- */
595 assert(get_irn_op(call) == op_Call);
596 /* @@@ does not work for InterfaceIII.java after cgana
597 assert(get_Call_type(call) == get_entity_type(get_irg_ent(called_graph)));
598 assert(smaller_type(get_entity_type(get_irg_ent(called_graph)),
599 get_Call_type(call)));
601 assert(get_type_tpop(get_Call_type(call)) == type_method);
602 if (called_graph == current_ir_graph) {
603 set_optimize(rem_opt);
607 /* -- Decide how to handle exception control flow: Is there a handler
608 for the Call node, or do we branch directly to End on an exception?
609 exc_handling: 0 There is a handler.
611 2 Exception handling not represented in Firm. -- */
613 for (proj = (ir_node *)get_irn_link(call); proj; proj = (ir_node *)get_irn_link(proj)) {
614 assert(get_irn_op(proj) == op_Proj);
615 if (get_Proj_proj(proj) == pn_Call_M_except) { exc_handling = 0; break;}
616 if (get_Proj_proj(proj) == pn_Call_X_except) { exc_handling = 1; }
620 ir_node *proj, *Mproj = NULL, *Xproj = NULL;
621 for (proj = (ir_node *)get_irn_link(call); proj; proj = (ir_node *)get_irn_link(proj)) {
622 assert(get_irn_op(proj) == op_Proj);
623 if (get_Proj_proj(proj) == pn_Call_X_except) Xproj = proj;
624 if (get_Proj_proj(proj) == pn_Call_M_except) Mproj = proj;
626 if (Mproj) { assert(Xproj); exc_handling = 0; }
627 else if (Xproj) { exc_handling = 1; }
628 else { exc_handling = 2; }
633 the procedure and later replaces the Start node of the called graph.
634 Post_call is the old Call node and collects the results of the called
635 graph. Both will end up being a tuple. -- */
636 post_bl = get_nodes_Block(call);
637 set_irg_current_block(current_ir_graph, post_bl);
638 /* XxMxPxP of Start + parameter of Call */
640 in[1] = get_Call_mem(call);
641 in[2] = get_irg_frame(current_ir_graph);
642 in[3] = get_irg_globals(current_ir_graph);
643 in[4] = new_Tuple (get_Call_n_params(call), get_Call_param_arr(call));
644 pre_call = new_Tuple(5, in);
648 The new block gets the ins of the old block, pre_call and all its
649 predecessors and all Phi nodes. -- */
650 part_block(pre_call);
652 /* -- Prepare state for dead node elimination -- */
653 /* Visited flags in calling irg must be >= flag in called irg.
654 Else walker and arity computation will not work. */
655 if (get_irg_visited(current_ir_graph) <= get_irg_visited(called_graph))
656 set_irg_visited(current_ir_graph, get_irg_visited(called_graph)+1);
657 if (get_irg_block_visited(current_ir_graph)< get_irg_block_visited(called_graph))
658 set_irg_block_visited(current_ir_graph, get_irg_block_visited(called_graph));
659 /* Set pre_call as new Start node in link field of the start node of
660 calling graph and pre_calls block as new block for the start block
662 Further mark these nodes so that they are not visited by the
664 set_irn_link(get_irg_start(called_graph), pre_call);
665 set_irn_visited(get_irg_start(called_graph),
666 get_irg_visited(current_ir_graph));
667 set_irn_link(get_irg_start_block(called_graph),
668 get_nodes_Block(pre_call));
669 set_irn_visited(get_irg_start_block(called_graph),
670 get_irg_visited(current_ir_graph));
672 /* Initialize for compaction of in arrays */
673 inc_irg_block_visited(current_ir_graph);
675 /* -- Replicate local entities of the called_graph -- */
676 /* copy the entities. */
677 called_frame = get_irg_frame_type(called_graph);
678 for (i = 0; i < get_class_n_members(called_frame); i++) {
679 entity *new_ent, *old_ent;
680 old_ent = get_class_member(called_frame, i);
681 new_ent = copy_entity_own(old_ent, get_cur_frame_type());
682 set_entity_link(old_ent, new_ent);
685 /* visited is > than that of called graph. With this trick visited will
686 remain unchanged so that an outer walker, e.g., searching the call nodes
687 to inline, calling this inline will not visit the inlined nodes. */
688 set_irg_visited(current_ir_graph, get_irg_visited(current_ir_graph)-1);
690 /* -- Performing dead node elimination inlines the graph -- */
691 /* Copies the nodes to the obstack of current_ir_graph. Updates links to new
693 /* @@@ endless loops are not copied!! -- they should be, I think... */
694 irg_walk(get_irg_end(called_graph), copy_node_inline, copy_preds,
695 get_irg_frame_type(called_graph));
697 /* Repair called_graph */
698 set_irg_visited(called_graph, get_irg_visited(current_ir_graph));
699 set_irg_block_visited(called_graph, get_irg_block_visited(current_ir_graph));
700 set_Block_block_visited(get_irg_start_block(called_graph), 0);
702 /* -- Merge the end of the inlined procedure with the call site -- */
703 /* We will turn the old Call node into a Tuple with the following
706 0: Phi of all Memories of Return statements.
707 1: Jmp from new Block that merges the control flow from all exception
708 predecessors of the old end block.
709 2: Tuple of all arguments.
710 3: Phi of Exception memories.
711 In case the old Call directly branches to End on an exception we don't
712 need the block merging all exceptions nor the Phi of the exception
716 /* -- Precompute some values -- */
717 end_bl = get_new_node(get_irg_end_block(called_graph));
718 end = get_new_node(get_irg_end(called_graph));
719 arity = get_irn_arity(end_bl); /* arity = n_exc + n_ret */
720 n_res = get_method_n_ress(get_Call_type(call));
722 res_pred = (ir_node **) malloc (n_res * sizeof (ir_node *));
723 cf_pred = (ir_node **) malloc (arity * sizeof (ir_node *));
725 set_irg_current_block(current_ir_graph, post_bl); /* just to make sure */
727 /* -- archive keepalives -- */
728 irn_arity = get_irn_arity(end);
729 for (i = 0; i < irn_arity; i++)
730 add_End_keepalive(get_irg_end(current_ir_graph), get_irn_n(end, i));
732 /* The new end node will die. We need not free as the in array is on the obstack:
733 copy_node only generated 'D' arrays. */
735 /* -- Replace Return nodes by Jump nodes. -- */
737 for (i = 0; i < arity; i++) {
739 ret = get_irn_n(end_bl, i);
740 if (get_irn_op(ret) == op_Return) {
741 cf_pred[n_ret] = new_r_Jmp(current_ir_graph, get_nodes_Block(ret));
745 set_irn_in(post_bl, n_ret, cf_pred);
747 /* -- Build a Tuple for all results of the method.
748 Add Phi node if there was more than one Return. -- */
749 turn_into_tuple(post_call, 4);
750 /* First the Memory-Phi */
752 for (i = 0; i < arity; i++) {
753 ret = get_irn_n(end_bl, i);
754 if (get_irn_op(ret) == op_Return) {
755 cf_pred[n_ret] = get_Return_mem(ret);
759 phi = new_Phi(n_ret, cf_pred, mode_M);
760 set_Tuple_pred(call, 0, phi);
761 /* Conserve Phi-list for further inlinings -- but might be optimized */
762 if (get_nodes_Block(phi) == post_bl) {
763 set_irn_link(phi, get_irn_link(post_bl));
764 set_irn_link(post_bl, phi);
766 /* Now the real results */
768 for (j = 0; j < n_res; j++) {
770 for (i = 0; i < arity; i++) {
771 ret = get_irn_n(end_bl, i);
772 if (get_irn_op(ret) == op_Return) {
773 cf_pred[n_ret] = get_Return_res(ret, j);
777 phi = new_Phi(n_ret, cf_pred, get_irn_mode(cf_pred[0]));
779 /* Conserve Phi-list for further inlinings -- but might be optimized */
780 if (get_nodes_Block(phi) == post_bl) {
781 set_irn_link(phi, get_irn_link(post_bl));
782 set_irn_link(post_bl, phi);
785 set_Tuple_pred(call, 2, new_Tuple(n_res, res_pred));
787 set_Tuple_pred(call, 2, new_Bad());
789 /* Finally the exception control flow.
790 We have two (three) possible situations:
791 First if the Call branches to an exception handler: We need to add a Phi node to
792 collect the memory containing the exception objects. Further we need
793 to add another block to get a correct representation of this Phi. To
794 this block we add a Jmp that resolves into the X output of the Call
795 when the Call is turned into a tuple.
796 Second the Call branches to End, the exception is not handled. Just
797 add all inlined exception branches to the End node.
798 Third: there is no Exception edge at all. Handle as case two. */
799 if (exc_handler == 0) {
801 for (i = 0; i < arity; i++) {
803 ret = get_irn_n(end_bl, i);
804 if (is_fragile_op(skip_Proj(ret)) || (get_irn_op(skip_Proj(ret)) == op_Raise)) {
805 cf_pred[n_exc] = ret;
810 new_Block(n_exc, cf_pred); /* watch it: current_block is changed! */
811 set_Tuple_pred(call, 1, new_Jmp());
812 /* The Phi for the memories with the exception objects */
814 for (i = 0; i < arity; i++) {
816 ret = skip_Proj(get_irn_n(end_bl, i));
817 if (get_irn_op(ret) == op_Call) {
818 cf_pred[n_exc] = new_r_Proj(current_ir_graph, get_nodes_Block(ret), ret, mode_M, 3);
820 } else if (is_fragile_op(ret)) {
821 /* We rely that all cfops have the memory output at the same position. */
822 cf_pred[n_exc] = new_r_Proj(current_ir_graph, get_nodes_Block(ret), ret, mode_M, 0);
824 } else if (get_irn_op(ret) == op_Raise) {
825 cf_pred[n_exc] = new_r_Proj(current_ir_graph, get_nodes_Block(ret), ret, mode_M, 1);
829 set_Tuple_pred(call, 3, new_Phi(n_exc, cf_pred, mode_M));
831 set_Tuple_pred(call, 1, new_Bad());
832 set_Tuple_pred(call, 3, new_Bad());
835 ir_node *main_end_bl;
836 int main_end_bl_arity;
839 /* assert(exc_handler == 1 || no exceptions. ) */
841 for (i = 0; i < arity; i++) {
842 ir_node *ret = get_irn_n(end_bl, i);
844 if (is_fragile_op(skip_Proj(ret)) || (get_irn_op(skip_Proj(ret)) == op_Raise)) {
845 cf_pred[n_exc] = ret;
849 main_end_bl = get_irg_end_block(current_ir_graph);
850 main_end_bl_arity = get_irn_arity(main_end_bl);
851 end_preds = (ir_node **) malloc ((n_exc + main_end_bl_arity) * sizeof (ir_node *));
853 for (i = 0; i < main_end_bl_arity; ++i)
854 end_preds[i] = get_irn_n(main_end_bl, i);
855 for (i = 0; i < n_exc; ++i)
856 end_preds[main_end_bl_arity + i] = cf_pred[i];
857 set_irn_in(main_end_bl, n_exc + main_end_bl_arity, end_preds);
858 set_Tuple_pred(call, 1, new_Bad());
859 set_Tuple_pred(call, 3, new_Bad());
865 #if 0 /* old. now better, correcter, faster implementation. */
867 /* -- If the exception control flow from the inlined Call directly
868 branched to the end block we now have the following control
869 flow predecessor pattern: ProjX -> Tuple -> Jmp. We must
870 remove the Jmp along with it's empty block and add Jmp's
871 predecessors as predecessors of this end block. No problem if
872 there is no exception, because then branches Bad to End which
874 @@@ can't we know this beforehand: by getting the Proj(1) from
875 the Call link list and checking whether it goes to Proj. */
876 /* find the problematic predecessor of the end block. */
877 end_bl = get_irg_end_block(current_ir_graph);
878 for (i = 0; i < get_Block_n_cfgpreds(end_bl); i++) {
879 cf_op = get_Block_cfgpred(end_bl, i);
880 if (get_irn_op(cf_op) == op_Proj) {
881 cf_op = get_Proj_pred(cf_op);
882 if ((get_irn_op(cf_op) == op_Tuple) && (cf_op == call)) {
883 // There are unoptimized tuples from inlineing before when no exc
884 assert(get_Proj_proj(get_Block_cfgpred(end_bl, i)) == pn_Call_X_except);
885 cf_op = get_Tuple_pred(cf_op, pn_Call_X_except);
886 assert(get_irn_op(cf_op) == op_Jmp);
892 if (i < get_Block_n_cfgpreds(end_bl)) {
893 bl = get_nodes_Block(cf_op);
894 arity = get_Block_n_cfgpreds(end_bl) + get_Block_n_cfgpreds(bl) - 1;
895 cf_pred = (ir_node **) malloc (arity * sizeof (ir_node *));
896 for (j = 0; j < i; j++)
897 cf_pred[j] = get_Block_cfgpred(end_bl, j);
898 for (j = j; j < i + get_Block_n_cfgpreds(bl); j++)
899 cf_pred[j] = get_Block_cfgpred(bl, j-i);
900 for (j = j; j < arity; j++)
901 cf_pred[j] = get_Block_cfgpred(end_bl, j-get_Block_n_cfgpreds(bl) +1);
902 set_irn_in(end_bl, arity, cf_pred);
904 // Remove the exception pred from post-call Tuple.
905 set_Tuple_pred(call, pn_Call_X_except, new_Bad());
910 /* -- Turn cse back on. -- */
911 set_optimize(rem_opt);
914 /********************************************************************/
915 /* Apply inlineing to small methods. */
916 /********************************************************************/
920 /* It makes no sense to inline too many calls in one procedure. Anyways,
921 I didn't get a version with NEW_ARR_F to run. */
922 #define MAX_INLINE 1024
924 /* given an Call node, returns the irg called. NULL if not
926 static ir_graph *get_call_called_irg(ir_node *call) {
929 ir_graph *called_irg = NULL;
931 assert(get_irn_op(call) == op_Call);
933 addr = get_Call_ptr(call);
934 if (get_irn_op(addr) == op_Const) {
935 /* Check whether the constant is the pointer to a compiled entity. */
936 tv = get_Const_tarval(addr);
937 if (tarval_to_entity(tv))
938 called_irg = get_entity_irg(tarval_to_entity(tv));
943 static void collect_calls(ir_node *call, void *env) {
945 ir_node **calls = (ir_node **)env;
948 ir_graph *called_irg;
950 if (get_irn_op(call) != op_Call) return;
952 addr = get_Call_ptr(call);
953 if (get_irn_op(addr) == op_Const) {
954 /* Check whether the constant is the pointer to a compiled entity. */
955 tv = get_Const_tarval(addr);
956 if (tarval_to_entity(tv)) {
957 called_irg = get_entity_irg(tarval_to_entity(tv));
958 if (called_irg && pos < MAX_INLINE) {
959 /* The Call node calls a locally defined method. Remember to inline. */
967 /* Inlines all small methods at call sites where the called address comes
968 from a Const node that references the entity representing the called
970 The size argument is a rough measure for the code size of the method:
971 Methods where the obstack containing the firm graph is smaller than
973 void inline_small_irgs(ir_graph *irg, int size) {
975 ir_node *calls[MAX_INLINE];
976 ir_graph *rem = current_ir_graph;
978 if (!(get_opt_optimize() && get_opt_inline())) return;
980 current_ir_graph = irg;
981 /* Handle graph state */
982 assert(get_irg_phase_state(current_ir_graph) != phase_building);
983 assert(get_irg_callee_info_state(current_ir_graph) == irg_callee_info_none);
985 /* Find Call nodes to inline.
986 (We can not inline during a walk of the graph, as inlineing the same
987 method several times changes the visited flag of the walked graph:
988 after the first inlineing visited of the callee equals visited of
989 the caller. With the next inlineing both are increased.) */
991 irg_walk(get_irg_end(irg), NULL, collect_calls, (void *) calls);
993 if ((pos > 0) && (pos < MAX_INLINE)) {
994 /* There are calls to inline */
995 collect_phiprojs(irg);
996 for (i = 0; i < pos; i++) {
999 tv = get_Const_tarval(get_Call_ptr(calls[i]));
1000 callee = get_entity_irg(tarval_to_entity(tv));
1001 if (((_obstack_memory_used(callee->obst) - obstack_room(callee->obst)) < size) ||
1002 (get_irg_inline_property(callee) == irg_inline_forced)) {
1003 inline_method(calls[i], callee);
1008 current_ir_graph = rem;
1012 * Environment for inlining irgs.
1015 int n_nodes; /**< Nodes in graph except Id, Tuple, Proj, Start, End */
1016 int n_nodes_orig; /**< for statistics */
1017 eset *call_nodes; /**< All call nodes in this graph */
1019 int n_call_nodes_orig; /**< for statistics */
1020 int n_callers; /**< Number of known graphs that call this graphs. */
1021 int n_callers_orig; /**< for statistics */
1024 static inline_irg_env *new_inline_irg_env(void) {
1025 inline_irg_env *env = malloc(sizeof(inline_irg_env));
1026 env->n_nodes = -2; /* uncount Start, End */
1027 env->n_nodes_orig = -2; /* uncount Start, End */
1028 env->call_nodes = eset_create();
1029 env->n_call_nodes = 0;
1030 env->n_call_nodes_orig = 0;
1032 env->n_callers_orig = 0;
1036 static void free_inline_irg_env(inline_irg_env *env) {
1037 eset_destroy(env->call_nodes);
1041 static void collect_calls2(ir_node *call, void *env) {
1042 inline_irg_env *x = (inline_irg_env *)env;
1043 ir_op *op = get_irn_op(call);
1046 /* count nodes in irg */
1047 if (op != op_Proj && op != op_Tuple && op != op_Sync) {
1052 if (op != op_Call) return;
1054 /* collect all call nodes */
1055 eset_insert(x->call_nodes, (void *)call);
1057 x->n_call_nodes_orig++;
1059 /* count all static callers */
1060 callee = get_call_called_irg(call);
1062 ((inline_irg_env *)get_irg_link(callee))->n_callers++;
1063 ((inline_irg_env *)get_irg_link(callee))->n_callers_orig++;
1067 INLINE static int is_leave(ir_graph *irg) {
1068 return (((inline_irg_env *)get_irg_link(irg))->n_call_nodes == 0);
1071 INLINE static int is_smaller(ir_graph *callee, int size) {
1072 return (((inline_irg_env *)get_irg_link(callee))->n_nodes < size);
1076 /* Inlines small leave methods at call sites where the called address comes
1077 from a Const node that references the entity representing the called
1079 The size argument is a rough measure for the code size of the method:
1080 Methods where the obstack containing the firm graph is smaller than
1081 size are inlined. */
1082 void inline_leave_functions(int maxsize, int leavesize, int size) {
1083 inline_irg_env *env;
1084 int i, n_irgs = get_irp_n_irgs();
1085 ir_graph *rem = current_ir_graph;
1088 if (!(get_opt_optimize() && get_opt_inline())) return;
1090 /* extend all irgs by a temporary data structure for inlineing. */
1091 for (i = 0; i < n_irgs; ++i)
1092 set_irg_link(get_irp_irg(i), new_inline_irg_env());
1094 /* Precompute information in temporary data structure. */
1095 for (i = 0; i < n_irgs; ++i) {
1096 current_ir_graph = get_irp_irg(i);
1097 assert(get_irg_phase_state(current_ir_graph) != phase_building);
1098 assert(get_irg_callee_info_state(current_ir_graph) == irg_callee_info_none);
1100 irg_walk(get_irg_end(current_ir_graph), NULL, collect_calls2,
1101 get_irg_link(current_ir_graph));
1102 env = (inline_irg_env *)get_irg_link(current_ir_graph);
1106 Inline leaves recursively -- we might construct new leaves. */
1108 while (did_inline) {
1109 //printf("iteration %d\n", itercnt++);
1111 for (i = 0; i < n_irgs; ++i) {
1114 int phiproj_computed = 0;
1116 current_ir_graph = get_irp_irg(i);
1117 env = (inline_irg_env *)get_irg_link(current_ir_graph);
1119 /* we can not walk and change a set, nor remove from it.
1121 walkset = env->call_nodes;
1122 env->call_nodes = eset_create();
1123 for (call = eset_first(walkset); call; call = eset_next(walkset)) {
1124 inline_irg_env *callee_env;
1125 ir_graph *callee = get_call_called_irg(call);
1127 if (env->n_nodes > maxsize) break;
1129 ((is_leave(callee) && is_smaller(callee, leavesize)) ||
1130 (get_irg_inline_property(callee) == irg_inline_forced))) {
1131 if (!phiproj_computed) {
1132 phiproj_computed = 1;
1133 collect_phiprojs(current_ir_graph);
1135 callee_env = (inline_irg_env *)get_irg_link(callee);
1136 // printf(" %s: Inlineing %s.\n", get_entity_name(get_irg_entity(current_ir_graph)),
1137 // get_entity_name(get_irg_entity(callee)));
1138 inline_method(call, callee);
1140 env->n_call_nodes--;
1141 eset_insert_all(env->call_nodes, callee_env->call_nodes);
1142 env->n_call_nodes += callee_env->n_call_nodes;
1143 env->n_nodes += callee_env->n_nodes;
1144 callee_env->n_callers--;
1146 eset_insert(env->call_nodes, call);
1149 eset_destroy(walkset);
1153 //printf("Non leaves\n");
1154 /* inline other small functions. */
1155 for (i = 0; i < n_irgs; ++i) {
1158 int phiproj_computed = 0;
1160 current_ir_graph = get_irp_irg(i);
1161 env = (inline_irg_env *)get_irg_link(current_ir_graph);
1163 /* we can not walk and change a set, nor remove from it.
1165 walkset = env->call_nodes;
1166 env->call_nodes = eset_create();
1167 for (call = eset_first(walkset); call; call = eset_next(walkset)) {
1168 inline_irg_env *callee_env;
1169 ir_graph *callee = get_call_called_irg(call);
1171 if (env->n_nodes > maxsize) break;
1172 if (callee && is_smaller(callee, size)) {
1173 if (!phiproj_computed) {
1174 phiproj_computed = 1;
1175 collect_phiprojs(current_ir_graph);
1177 callee_env = (inline_irg_env *)get_irg_link(callee);
1178 // printf(" %s: Inlineing %s.\n", get_entity_name(get_irg_entity(current_ir_graph)),
1179 // get_entity_name(get_irg_entity(callee)));
1180 inline_method(call, callee);
1182 env->n_call_nodes--;
1183 eset_insert_all(env->call_nodes, callee_env->call_nodes);
1184 env->n_call_nodes += callee_env->n_call_nodes;
1185 env->n_nodes += callee_env->n_nodes;
1186 callee_env->n_callers--;
1188 eset_insert(env->call_nodes, call);
1191 eset_destroy(walkset);
1194 for (i = 0; i < n_irgs; ++i) {
1195 current_ir_graph = get_irp_irg(i);
1197 env = (inline_irg_env *)get_irg_link(current_ir_graph);
1198 if ((env->n_call_nodes_orig != env->n_call_nodes) ||
1199 (env->n_callers_orig != env->n_callers))
1200 printf("Nodes:%3d ->%3d, calls:%3d ->%3d, callers:%3d ->%3d, -- %s\n",
1201 env->n_nodes_orig, env->n_nodes, env->n_call_nodes_orig, env->n_call_nodes,
1202 env->n_callers_orig, env->n_callers,
1203 get_entity_name(get_irg_entity(current_ir_graph)));
1205 free_inline_irg_env((inline_irg_env *)get_irg_link(current_ir_graph));
1208 current_ir_graph = rem;
1211 /********************************************************************/
1212 /* Code Placement. Pins all floating nodes to a block where they */
1213 /* will be executed only if needed. */
1214 /********************************************************************/
1216 /* Find the earliest correct block for N. --- Place N into the
1217 same Block as its dominance-deepest Input. */
1219 place_floats_early(ir_node *n, pdeq *worklist)
1221 int i, start, irn_arity;
1223 /* we must not run into an infinite loop */
1224 assert (irn_not_visited(n));
1225 mark_irn_visited(n);
1227 /* Place floating nodes. */
1228 if (get_op_pinned(get_irn_op(n)) == floats) {
1230 ir_node *b = new_Bad(); /* The block to place this node in */
1232 assert(get_irn_op(n) != op_Block);
1234 if ((get_irn_op(n) == op_Const) ||
1235 (get_irn_op(n) == op_SymConst) ||
1237 (get_irn_op(n) == op_Unknown)) {
1238 /* These nodes will not be placed by the loop below. */
1239 b = get_irg_start_block(current_ir_graph);
1243 /* find the block for this node. */
1244 irn_arity = get_irn_arity(n);
1245 for (i = 0; i < irn_arity; i++) {
1246 ir_node *dep = get_irn_n(n, i);
1248 if ((irn_not_visited(dep)) &&
1249 (get_op_pinned(get_irn_op(dep)) == floats)) {
1250 place_floats_early(dep, worklist);
1252 /* Because all loops contain at least one pinned node, now all
1253 our inputs are either pinned or place_early has already
1254 been finished on them. We do not have any unfinished inputs! */
1255 dep_block = get_nodes_Block(dep);
1256 if ((!is_Bad(dep_block)) &&
1257 (get_Block_dom_depth(dep_block) > depth)) {
1259 depth = get_Block_dom_depth(dep_block);
1261 /* Avoid that the node is placed in the Start block */
1262 if ((depth == 1) && (get_Block_dom_depth(get_nodes_Block(n)) > 1)) {
1263 b = get_Block_cfg_out(get_irg_start_block(current_ir_graph), 0);
1264 assert(b != get_irg_start_block(current_ir_graph));
1268 set_nodes_Block(n, b);
1271 /* Add predecessors of non floating nodes on worklist. */
1272 start = (get_irn_op(n) == op_Block) ? 0 : -1;
1273 irn_arity = get_irn_arity(n);
1274 for (i = start; i < irn_arity; i++) {
1275 ir_node *pred = get_irn_n(n, i);
1276 if (irn_not_visited(pred)) {
1277 pdeq_putr (worklist, pred);
1282 /* Floating nodes form subgraphs that begin at nodes as Const, Load,
1283 Start, Call and end at pinned nodes as Store, Call. Place_early
1284 places all floating nodes reachable from its argument through floating
1285 nodes and adds all beginnings at pinned nodes to the worklist. */
1286 static INLINE void place_early(pdeq* worklist) {
1288 inc_irg_visited(current_ir_graph);
1290 /* this inits the worklist */
1291 place_floats_early(get_irg_end(current_ir_graph), worklist);
1293 /* Work the content of the worklist. */
1294 while (!pdeq_empty (worklist)) {
1295 ir_node *n = pdeq_getl (worklist);
1296 if (irn_not_visited(n)) place_floats_early(n, worklist);
1299 set_irg_outs_inconsistent(current_ir_graph);
1300 current_ir_graph->pinned = pinned;
1304 /* deepest common dominance ancestor of DCA and CONSUMER of PRODUCER */
1306 consumer_dom_dca (ir_node *dca, ir_node *consumer, ir_node *producer)
1308 ir_node *block = NULL;
1310 /* Compute the latest block into which we can place a node so that it is
1312 if (get_irn_op(consumer) == op_Phi) {
1313 /* our consumer is a Phi-node, the effective use is in all those
1314 blocks through which the Phi-node reaches producer */
1316 ir_node *phi_block = get_nodes_Block(consumer);
1317 irn_arity = get_irn_arity(consumer);
1318 for (i = 0; i < irn_arity; i++) {
1319 if (get_irn_n(consumer, i) == producer) {
1320 block = get_nodes_Block(get_Block_cfgpred(phi_block, i));
1324 assert(is_no_Block(consumer));
1325 block = get_nodes_Block(consumer);
1328 /* Compute the deepest common ancestor of block and dca. */
1330 if (!dca) return block;
1331 while (get_Block_dom_depth(block) > get_Block_dom_depth(dca))
1332 block = get_Block_idom(block);
1333 while (get_Block_dom_depth(dca) > get_Block_dom_depth(block))
1334 dca = get_Block_idom(dca);
1335 while (block != dca)
1336 { block = get_Block_idom(block); dca = get_Block_idom(dca); }
1341 static INLINE int get_irn_loop_depth(ir_node *n) {
1342 return get_loop_depth(get_irn_loop(n));
1345 /* Move n to a block with less loop depth than it's current block. The
1346 new block must be dominated by early. */
1348 move_out_of_loops (ir_node *n, ir_node *early)
1350 ir_node *best, *dca;
1354 /* Find the region deepest in the dominator tree dominating
1355 dca with the least loop nesting depth, but still dominated
1356 by our early placement. */
1357 dca = get_nodes_Block(n);
1359 while (dca != early) {
1360 dca = get_Block_idom(dca);
1361 if (!dca) break; /* should we put assert(dca)? */
1362 if (get_irn_loop_depth(dca) < get_irn_loop_depth(best)) {
1366 if (best != get_nodes_Block(n)) {
1368 printf("Moving out of loop: "); DDMN(n);
1369 printf(" Outermost block: "); DDMN(early);
1370 printf(" Best block: "); DDMN(best);
1371 printf(" Innermost block: "); DDMN(get_nodes_Block(n));
1373 set_nodes_Block(n, best);
1377 /* Find the latest legal block for N and place N into the
1378 `optimal' Block between the latest and earliest legal block.
1379 The `optimal' block is the dominance-deepest block of those
1380 with the least loop-nesting-depth. This places N out of as many
1381 loops as possible and then makes it as control dependant as
1384 place_floats_late(ir_node *n, pdeq *worklist)
1389 assert (irn_not_visited(n)); /* no multiple placement */
1391 /* no need to place block nodes, control nodes are already placed. */
1392 if ((get_irn_op(n) != op_Block) &&
1394 (get_irn_mode(n) != mode_X)) {
1395 /* Remember the early placement of this block to move it
1396 out of loop no further than the early placement. */
1397 early = get_nodes_Block(n);
1398 /* Assure that our users are all placed, except the Phi-nodes.
1399 --- Each data flow cycle contains at least one Phi-node. We
1400 have to break the `user has to be placed before the
1401 producer' dependence cycle and the Phi-nodes are the
1402 place to do so, because we need to base our placement on the
1403 final region of our users, which is OK with Phi-nodes, as they
1404 are pinned, and they never have to be placed after a
1405 producer of one of their inputs in the same block anyway. */
1406 for (i = 0; i < get_irn_n_outs(n); i++) {
1407 ir_node *succ = get_irn_out(n, i);
1408 if (irn_not_visited(succ) && (get_irn_op(succ) != op_Phi))
1409 place_floats_late(succ, worklist);
1412 /* We have to determine the final block of this node... except for
1414 if ((get_op_pinned(get_irn_op(n)) == floats) &&
1415 (get_irn_op(n) != op_Const) &&
1416 (get_irn_op(n) != op_SymConst)) {
1417 ir_node *dca = NULL; /* deepest common ancestor in the
1418 dominator tree of all nodes'
1419 blocks depending on us; our final
1420 placement has to dominate DCA. */
1421 for (i = 0; i < get_irn_n_outs(n); i++) {
1422 dca = consumer_dom_dca (dca, get_irn_out(n, i), n);
1424 set_nodes_Block(n, dca);
1426 move_out_of_loops (n, early);
1430 mark_irn_visited(n);
1432 /* Add predecessors of all non-floating nodes on list. (Those of floating
1433 nodes are placeded already and therefore are marked.) */
1434 for (i = 0; i < get_irn_n_outs(n); i++) {
1435 if (irn_not_visited(get_irn_out(n, i))) {
1436 pdeq_putr (worklist, get_irn_out(n, i));
1441 static INLINE void place_late(pdeq* worklist) {
1443 inc_irg_visited(current_ir_graph);
1445 /* This fills the worklist initially. */
1446 place_floats_late(get_irg_start_block(current_ir_graph), worklist);
1447 /* And now empty the worklist again... */
1448 while (!pdeq_empty (worklist)) {
1449 ir_node *n = pdeq_getl (worklist);
1450 if (irn_not_visited(n)) place_floats_late(n, worklist);
1454 void place_code(ir_graph *irg) {
1456 ir_graph *rem = current_ir_graph;
1458 current_ir_graph = irg;
1460 if (!(get_opt_optimize() && get_opt_global_cse())) return;
1462 /* Handle graph state */
1463 assert(get_irg_phase_state(irg) != phase_building);
1464 if (get_irg_dom_state(irg) != dom_consistent)
1467 construct_backedges(irg);
1469 /* Place all floating nodes as early as possible. This guarantees
1470 a legal code placement. */
1471 worklist = new_pdeq();
1472 place_early(worklist);
1474 /* place_early invalidates the outs, place_late needs them. */
1476 /* Now move the nodes down in the dominator tree. This reduces the
1477 unnecessary executions of the node. */
1478 place_late(worklist);
1480 set_irg_outs_inconsistent(current_ir_graph);
1482 current_ir_graph = rem;
1487 /********************************************************************/
1488 /* Control flow optimization. */
1489 /* Removes Bad control flow predecessors and empty blocks. A block */
1490 /* is empty if it contains only a Jmp node. */
1491 /* Blocks can only be removed if they are not needed for the */
1492 /* semantics of Phi nodes. */
1493 /********************************************************************/
1495 /* Removes Tuples from Block control flow predecessors.
1496 Optimizes blocks with equivalent_node().
1497 Replaces n by Bad if n is unreachable control flow. */
1498 static void merge_blocks(ir_node *n, void *env) {
1500 set_irn_link(n, NULL);
1502 if (get_irn_op(n) == op_Block) {
1504 for (i = 0; i < get_Block_n_cfgpreds(n); i++)
1505 /* GL @@@ : is this possible? if (get_opt_normalize()) -- added, all tests go through.
1506 A different order of optimizations might cause problems. */
1507 if (get_opt_normalize())
1508 set_Block_cfgpred(n, i, skip_Tuple(get_Block_cfgpred(n, i)));
1509 } else if (get_opt_optimize() && (get_irn_mode(n) == mode_X)) {
1510 /* We will soon visit a block. Optimize it before visiting! */
1511 ir_node *b = get_nodes_Block(n);
1512 ir_node *new_node = equivalent_node(b);
1513 while (irn_not_visited(b) && (!is_Bad(new_node)) && (new_node != b)) {
1514 /* We would have to run gigo if new is bad, so we
1515 promote it directly below. */
1516 assert(((b == new_node) ||
1517 get_opt_control_flow_straightening() ||
1518 get_opt_control_flow_weak_simplification()) &&
1519 ("strange flag setting"));
1520 exchange (b, new_node);
1522 new_node = equivalent_node(b);
1524 /* GL @@@ get_opt_normalize hinzugefuegt, 5.5.2003 */
1525 if (is_Bad(new_node) && get_opt_normalize()) exchange(n, new_Bad());
1529 /* Collects all Phi nodes in link list of Block.
1530 Marks all blocks "block_visited" if they contain a node other
1532 static void collect_nodes(ir_node *n, void *env) {
1533 if (is_no_Block(n)) {
1534 ir_node *b = get_nodes_Block(n);
1536 if ((get_irn_op(n) == op_Phi)) {
1537 /* Collect Phi nodes to compact ins along with block's ins. */
1538 set_irn_link(n, get_irn_link(b));
1540 } else if (get_irn_op(n) != op_Jmp) { /* Check for non empty block. */
1541 mark_Block_block_visited(b);
1546 /* Returns true if pred is pred of block */
1547 static int is_pred_of(ir_node *pred, ir_node *b) {
1549 for (i = 0; i < get_Block_n_cfgpreds(b); i++) {
1550 ir_node *b_pred = get_nodes_Block(get_Block_cfgpred(b, i));
1551 if (b_pred == pred) return 1;
1556 static int test_whether_dispensable(ir_node *b, int pos) {
1557 int i, j, n_preds = 1;
1558 int dispensable = 1;
1559 ir_node *cfop = get_Block_cfgpred(b, pos);
1560 ir_node *pred = get_nodes_Block(cfop);
1562 if (get_Block_block_visited(pred) + 1
1563 < get_irg_block_visited(current_ir_graph)) {
1564 if (!get_opt_optimize() || !get_opt_control_flow_strong_simplification()) {
1565 /* Mark block so that is will not be removed. */
1566 set_Block_block_visited(pred, get_irg_block_visited(current_ir_graph)-1);
1569 /* Seems to be empty. */
1570 if (!get_irn_link(b)) {
1571 /* There are no Phi nodes ==> dispensable. */
1572 n_preds = get_Block_n_cfgpreds(pred);
1574 /* b's pred blocks and pred's pred blocks must be pairwise disjunct.
1575 Work preds < pos as if they were already removed. */
1576 for (i = 0; i < pos; i++) {
1577 ir_node *b_pred = get_nodes_Block(get_Block_cfgpred(b, i));
1578 if (get_Block_block_visited(b_pred) + 1
1579 < get_irg_block_visited(current_ir_graph)) {
1580 for (j = 0; j < get_Block_n_cfgpreds(b_pred); j++) {
1581 ir_node *b_pred_pred = get_nodes_Block(get_Block_cfgpred(b_pred, j));
1582 if (is_pred_of(b_pred_pred, pred)) dispensable = 0;
1585 if (is_pred_of(b_pred, pred)) dispensable = 0;
1588 for (i = pos +1; i < get_Block_n_cfgpreds(b); i++) {
1589 ir_node *b_pred = get_nodes_Block(get_Block_cfgpred(b, i));
1590 if (is_pred_of(b_pred, pred)) dispensable = 0;
1593 set_Block_block_visited(pred, get_irg_block_visited(current_ir_graph)-1);
1596 n_preds = get_Block_n_cfgpreds(pred);
1604 static void optimize_blocks(ir_node *b, void *env) {
1605 int i, j, k, max_preds, n_preds;
1606 ir_node *pred, *phi;
1609 /* Count the number of predecessor if this block is merged with pred blocks
1612 for (i = 0; i < get_Block_n_cfgpreds(b); i++) {
1613 max_preds += test_whether_dispensable(b, i);
1615 in = (ir_node **) malloc(max_preds * sizeof(ir_node *));
1618 printf(" working on "); DDMN(b);
1619 for (i = 0; i < get_Block_n_cfgpreds(b); i++) {
1620 pred = get_nodes_Block(get_Block_cfgpred(b, i));
1621 if (is_Bad(get_Block_cfgpred(b, i))) {
1622 printf(" removing Bad %i\n ", i);
1623 } else if (get_Block_block_visited(pred) +1
1624 < get_irg_block_visited(current_ir_graph)) {
1625 printf(" removing pred %i ", i); DDMN(pred);
1626 } else { printf(" Nothing to do for "); DDMN(pred); }
1628 * end Debug output **/
1630 /** Fix the Phi nodes **/
1631 phi = get_irn_link(b);
1633 assert(get_irn_op(phi) == op_Phi);
1634 /* Find the new predecessors for the Phi */
1636 for (i = 0; i < get_Block_n_cfgpreds(b); i++) {
1637 pred = get_nodes_Block(get_Block_cfgpred(b, i));
1638 if (is_Bad(get_Block_cfgpred(b, i))) {
1640 } else if (get_Block_block_visited(pred) +1
1641 < get_irg_block_visited(current_ir_graph)) {
1642 /* It's an empty block and not yet visited. */
1643 ir_node *phi_pred = get_Phi_pred(phi, i);
1644 for (j = 0; j < get_Block_n_cfgpreds(pred); j++) {
1645 if (get_nodes_Block(phi_pred) == pred) {
1646 assert(get_irn_op(phi_pred) == op_Phi); /* Block is empty!! */
1647 in[n_preds] = get_Phi_pred(phi_pred, j);
1649 in[n_preds] = phi_pred;
1653 /* The Phi_pred node is replaced now if it is a Phi.
1654 In Schleifen kann offenbar der entfernte Phi Knoten legal verwendet werden.
1655 Daher muss der Phiknoten durch den neuen ersetzt werden.
1656 Weiter muss der alte Phiknoten entfernt werden (durch ersetzen oder
1657 durch einen Bad) damit er aus den keep_alive verschwinden kann.
1658 Man sollte also, falls keine Schleife vorliegt, exchange mit new_Bad
1660 if (get_nodes_Block(phi_pred) == pred) {
1661 /* remove the Phi as it might be kept alive. Further there
1662 might be other users. */
1663 exchange(phi_pred, phi); /* geht, ist aber doch semantisch falsch! Warum?? */
1666 in[n_preds] = get_Phi_pred(phi, i);
1671 set_irn_in(phi, n_preds, in);
1673 phi = get_irn_link(phi);
1677 This happens only if merge between loop backedge and single loop entry. **/
1678 for (k = 0; k < get_Block_n_cfgpreds(b); k++) {
1679 pred = get_nodes_Block(get_Block_cfgpred(b, k));
1680 if (get_Block_block_visited(pred) +1
1681 < get_irg_block_visited(current_ir_graph)) {
1682 phi = get_irn_link(pred);
1684 if (get_irn_op(phi) == op_Phi) {
1685 set_nodes_Block(phi, b);
1688 for (i = 0; i < k; i++) {
1689 pred = get_nodes_Block(get_Block_cfgpred(b, i));
1690 if (is_Bad(get_Block_cfgpred(b, i))) {
1692 } else if (get_Block_block_visited(pred) +1
1693 < get_irg_block_visited(current_ir_graph)) {
1694 /* It's an empty block and not yet visited. */
1695 for (j = 0; j < get_Block_n_cfgpreds(pred); j++) {
1696 /* @@@ Hier brauche ich Schleifeninformation!!! Kontrollflusskante
1697 muss Rueckwaertskante sein! (An allen vier in[n_preds] = phi
1698 Anweisungen.) Trotzdem tuts bisher!! */
1707 for (i = 0; i < get_Phi_n_preds(phi); i++) {
1708 in[n_preds] = get_Phi_pred(phi, i);
1711 for (i = k+1; i < get_Block_n_cfgpreds(b); i++) {
1712 pred = get_nodes_Block(get_Block_cfgpred(b, i));
1713 if (is_Bad(get_Block_cfgpred(b, i))) {
1715 } else if (get_Block_block_visited(pred) +1
1716 < get_irg_block_visited(current_ir_graph)) {
1717 /* It's an empty block and not yet visited. */
1718 for (j = 0; j < get_Block_n_cfgpreds(pred); j++) {
1727 set_irn_in(phi, n_preds, in);
1729 phi = get_irn_link(phi);
1734 /** Fix the block **/
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))) {
1740 } else if (get_Block_block_visited(pred) +1
1741 < get_irg_block_visited(current_ir_graph)) {
1742 /* It's an empty block and not yet visited. */
1743 assert(get_Block_n_cfgpreds(b) > 1);
1744 /* Else it should be optimized by equivalent_node. */
1745 for (j = 0; j < get_Block_n_cfgpreds(pred); j++) {
1746 in[n_preds] = get_Block_cfgpred(pred, j);
1749 /* Remove block as it might be kept alive. */
1750 exchange(pred, b/*new_Bad()*/);
1752 in[n_preds] = get_Block_cfgpred(b, i);
1756 set_irn_in(b, n_preds, in);
1760 void optimize_cf(ir_graph *irg) {
1763 ir_node *end = get_irg_end(irg);
1764 ir_graph *rem = current_ir_graph;
1765 current_ir_graph = irg;
1767 /* Handle graph state */
1768 assert(get_irg_phase_state(irg) != phase_building);
1769 if (get_irg_outs_state(current_ir_graph) == outs_consistent)
1770 set_irg_outs_inconsistent(current_ir_graph);
1771 if (get_irg_dom_state(current_ir_graph) == dom_consistent)
1772 set_irg_dom_inconsistent(current_ir_graph);
1774 /* Use block visited flag to mark non-empty blocks. */
1775 inc_irg_block_visited(irg);
1776 irg_walk(end, merge_blocks, collect_nodes, NULL);
1778 /* Optimize the standard code. */
1779 irg_block_walk(get_irg_end_block(irg), optimize_blocks, NULL, NULL);
1781 /* Walk all keep alives, optimize them if block, add to new in-array
1782 for end if useful. */
1783 in = NEW_ARR_F (ir_node *, 1);
1784 in[0] = get_nodes_Block(end);
1785 inc_irg_visited(current_ir_graph);
1786 for(i = 0; i < get_End_n_keepalives(end); i++) {
1787 ir_node *ka = get_End_keepalive(end, i);
1788 if (irn_not_visited(ka)) {
1789 if ((get_irn_op(ka) == op_Block) && Block_not_block_visited(ka)) {
1790 set_irg_block_visited(current_ir_graph, /* Don't walk all the way to Start. */
1791 get_irg_block_visited(current_ir_graph)-1);
1792 irg_block_walk(ka, optimize_blocks, NULL, NULL);
1793 mark_irn_visited(ka);
1794 ARR_APP1 (ir_node *, in, ka);
1795 } else if (get_irn_op(ka) == op_Phi) {
1796 mark_irn_visited(ka);
1797 ARR_APP1 (ir_node *, in, ka);
1801 /* DEL_ARR_F(end->in); GL @@@ tut nicht ! */
1804 current_ir_graph = rem;
1809 * Called by walker of remove_critical_cf_edges.
1811 * Place an empty block to an edge between a blocks of multiple
1812 * predecessors and a block of multiple successors.
1815 * @param env Environment of walker. This field is unused and has
1818 static void walk_critical_cf_edges(ir_node *n, void *env) {
1820 ir_node *pre, *block, **in, *jmp;
1822 /* Block has multiple predecessors */
1823 if ((op_Block == get_irn_op(n)) &&
1824 (get_irn_arity(n) > 1)) {
1825 arity = get_irn_arity(n);
1827 if (n == get_irg_end_block(current_ir_graph))
1828 return; // No use to add a block here.
1830 for (i=0; i<arity; i++) {
1831 pre = get_irn_n(n, i);
1832 /* Predecessor has multiple successors. Insert new flow edge */
1833 if ((NULL != pre) &&
1834 (op_Proj == get_irn_op(pre)) &&
1835 op_Raise != get_irn_op(skip_Proj(pre))) {
1837 /* set predecessor array for new block */
1838 in = NEW_ARR_D (ir_node *, current_ir_graph->obst, 1);
1839 /* set predecessor of new block */
1841 block = new_Block(1, in);
1842 /* insert new jmp node to new block */
1843 switch_block(block);
1846 /* set successor of new block */
1847 set_irn_n(n, i, jmp);
1849 } /* predecessor has multiple successors */
1850 } /* for all predecessors */
1851 } /* n is a block */
1854 void remove_critical_cf_edges(ir_graph *irg) {
1855 if (get_opt_critical_edges())
1856 irg_walk_graph(irg, NULL, walk_critical_cf_edges, NULL);