1 /* Coyright (C) 1998 - 2000 by Universitaet Karlsruhe
2 ** All rights reserved.
4 ** Author: Christian Schaefer
6 ** Optimizations for a whole ir graph, i.e., a procedure.
19 # include "irnode_t.h"
20 # include "irgraph_t.h"
28 # include "pdeq.h" /* Fuer code placement */
31 /* Defined in iropt.c */
32 pset *new_identities (void);
33 void del_identities (pset *value_table);
34 void add_identities (pset *value_table, ir_node *node);
36 /********************************************************************/
37 /* apply optimizations of iropt to all nodes. */
38 /********************************************************************/
40 void init_link (ir_node *n, void *env) {
41 set_irn_link(n, NULL);
45 optimize_in_place_wrapper (ir_node *n, void *env) {
49 if (get_irn_op(n) == op_Block)
54 /* optimize all sons after recursion, i.e., the sons' sons are
56 for (i = start; i < get_irn_arity(n); i++) {
57 optimized = optimize_in_place_2(get_irn_n(n, i));
58 set_irn_n(n, i, optimized);
59 assert(get_irn_op(optimized) != op_Id);
64 local_optimize_graph (ir_graph *irg) {
65 ir_graph *rem = current_ir_graph;
66 current_ir_graph = irg;
68 /* Handle graph state */
69 assert(get_irg_phase_state(irg) != phase_building);
70 if (get_opt_global_cse())
71 set_irg_pinned(current_ir_graph, floats);
72 if (get_irg_outs_state(current_ir_graph) == outs_consistent)
73 set_irg_outs_inconsistent(current_ir_graph);
74 if (get_irg_dom_state(current_ir_graph) == dom_consistent)
75 set_irg_dom_inconsistent(current_ir_graph);
77 /* Clean the value_table in irg for the cse. */
78 del_identities(irg->value_table);
79 irg->value_table = new_identities();
81 /* walk over the graph */
82 irg_walk(irg->end, init_link, optimize_in_place_wrapper, NULL);
84 current_ir_graph = rem;
87 /********************************************************************/
88 /* Routines for dead node elimination / copying garbage collection */
90 /********************************************************************/
92 /* Remeber the new node in the old node by using a field all nodes have. */
94 set_new_node (ir_node *old, ir_node *new)
99 /* Get this new node, before the old node is forgotton.*/
101 get_new_node (ir_node * n)
106 /* We use the block_visited flag to mark that we have computed the
107 number of useful predecessors for this block.
108 Further we encode the new arity in this flag in the old blocks.
109 Remembering the arity is useful, as it saves a lot of pointer
110 accesses. This function is called for all Phi and Block nodes
113 compute_new_arity(ir_node *b) {
117 irg_v = get_irg_block_visited(current_ir_graph);
118 block_v = get_Block_block_visited(b);
119 if (block_v >= irg_v) {
120 /* we computed the number of preds for this block and saved it in the
122 return block_v - irg_v;
124 /* compute the number of good predecessors */
125 res = get_irn_arity(b);
126 for (i = 0; i < get_irn_arity(b); i++)
127 if (get_irn_opcode(get_irn_n(b, i)) == iro_Bad) res--;
128 /* save it in the flag. */
129 set_Block_block_visited(b, irg_v + res);
134 /* Copies the node to the new obstack. The Ins of the new node point to
135 the predecessors on the old obstack. For block/phi nodes not all
136 predecessors might be copied. n->link points to the new node.
137 For Phi and Block nodes the function allocates in-arrays with an arity
138 only for useful predecessors. The arity is determined by counting
139 the non-bad predecessors of the block. */
141 copy_node (ir_node *n, void *env) {
145 if (get_irn_opcode(n) == iro_Block) {
147 new_arity = compute_new_arity(n);
149 block = get_nodes_Block(n);
150 if (get_irn_opcode(n) == iro_Phi) {
151 new_arity = compute_new_arity(block);
153 new_arity = get_irn_arity(n);
156 nn = new_ir_node(current_ir_graph,
162 /* Copy the attributes. These might point to additional data. If this
163 was allocated on the old obstack the pointers now are dangling. This
164 frees e.g. the memory of the graph_arr allocated in new_immBlock. */
168 /* printf("\n old node: "); DDMSG2(n);
169 printf(" new node: "); DDMSG2(nn); */
173 /* Copies new predecessors of old node to new node remembered in link.
174 Spare the Bad predecessors of Phi and Block nodes. */
176 copy_preds (ir_node *n, void *env) {
177 ir_node *nn, *block, *on;
180 nn = get_new_node(n);
182 /* printf("\n old node: "); DDMSG2(n);
183 printf(" new node: "); DDMSG2(nn);
184 printf(" arities: old: %d, new: %d\n", get_irn_arity(n), get_irn_arity(nn)); */
186 if (get_irn_opcode(n) == iro_Block) {
187 /* Don't copy Bad nodes. */
189 for (i = 0; i < get_irn_arity(n); i++)
190 if (get_irn_opcode(get_irn_n(n, i)) != iro_Bad) {
191 set_irn_n (nn, j, get_new_node(get_irn_n(n, i)));
194 /* repair the block visited flag from above misuse. Repair it in both
195 graphs so that the old one can still be used. */
196 set_Block_block_visited(nn, 0);
197 set_Block_block_visited(n, 0);
198 /* Local optimization could not merge two subsequent blocks if
199 in array contained Bads. Now it's possible.
200 We don't call optimize_in_place as it requires
201 that the fields in ir_graph are set properly. */
202 if (get_Block_n_cfgpreds(nn) == 1
203 && get_irn_op(get_Block_cfgpred(nn, 0)) == op_Jmp)
204 exchange(nn, get_nodes_Block(get_Block_cfgpred(nn, 0)));
205 } else if (get_irn_opcode(n) == iro_Phi) {
206 /* Don't copy node if corresponding predecessor in block is Bad.
207 The Block itself should not be Bad. */
208 block = get_nodes_Block(n);
209 set_irn_n (nn, -1, get_new_node(block));
211 for (i = 0; i < get_irn_arity(n); i++)
212 if (get_irn_opcode(get_irn_n(block, i)) != iro_Bad) {
213 set_irn_n (nn, j, get_new_node(get_irn_n(n, i)));
216 /* If the pre walker reached this Phi after the post walker visited the
217 block block_visited is > 0. */
218 set_Block_block_visited(get_nodes_Block(n), 0);
219 /* Compacting the Phi's ins might generate Phis with only one
221 if (get_irn_arity(n) == 1)
222 exchange(n, get_irn_n(n, 0));
224 for (i = -1; i < get_irn_arity(n); i++)
225 set_irn_n (nn, i, get_new_node(get_irn_n(n, i)));
227 /* Now the new node is complete. We can add it to the hash table for cse.
228 @@@ inlinening aborts if we identify End. Why? */
229 if(get_irn_op(nn) != op_End)
230 add_identities (current_ir_graph->value_table, nn);
233 /* Copies the graph recursively, compacts the keepalive of the end node. */
236 ir_node *oe, *ne; /* old end, new end */
237 ir_node *ka; /* keep alive */
240 oe = get_irg_end(current_ir_graph);
241 /* copy the end node by hand, allocate dynamic in array! */
242 ne = new_ir_node(current_ir_graph,
248 /* Copy the attributes. Well, there might be some in the future... */
250 set_new_node(oe, ne);
252 /* copy the live nodes */
253 irg_walk(get_nodes_Block(oe), copy_node, copy_preds, NULL);
254 /* copy_preds for the end node ... */
255 set_nodes_Block(ne, get_new_node(get_nodes_Block(oe)));
257 /** ... and now the keep alives. **/
258 /* First pick the not marked block nodes and walk them. We must pick these
259 first as else we will oversee blocks reachable from Phis. */
260 for (i = 0; i < get_irn_arity(oe); i++) {
261 ka = get_irn_n(oe, i);
262 if ((get_irn_op(ka) == op_Block) &&
263 (get_irn_visited(ka) < get_irg_visited(current_ir_graph))) {
264 /* We must keep the block alive and copy everything reachable */
265 set_irg_visited(current_ir_graph, get_irg_visited(current_ir_graph)-1);
266 irg_walk(ka, copy_node, copy_preds, NULL);
267 add_End_keepalive(ne, get_new_node(ka));
271 /* Now pick the Phis. Here we will keep all! */
272 for (i = 0; i < get_irn_arity(oe); i++) {
273 ka = get_irn_n(oe, i);
274 if ((get_irn_op(ka) == op_Phi)) {
275 if (get_irn_visited(ka) < get_irg_visited(current_ir_graph)) {
276 /* We didn't copy the Phi yet. */
277 set_irg_visited(current_ir_graph, get_irg_visited(current_ir_graph)-1);
278 irg_walk(ka, copy_node, copy_preds, NULL);
280 add_End_keepalive(ne, get_new_node(ka));
285 /* Copies the graph reachable from current_ir_graph->end to the obstack
286 in current_ir_graph and fixes the environment.
287 Then fixes the fields in current_ir_graph containing nodes of the
291 /* Not all nodes remembered in current_ir_graph might be reachable
292 from the end node. Assure their link is set to NULL, so that
293 we can test whether new nodes have been computed. */
294 set_irn_link(get_irg_frame (current_ir_graph), NULL);
295 set_irn_link(get_irg_globals(current_ir_graph), NULL);
296 set_irn_link(get_irg_args (current_ir_graph), NULL);
298 /* we use the block walk flag for removing Bads from Blocks ins. */
299 inc_irg_block_visited(current_ir_graph);
304 /* fix the fields in current_ir_graph */
305 free_End(get_irg_end(current_ir_graph));
306 set_irg_end (current_ir_graph, get_new_node(get_irg_end(current_ir_graph)));
307 set_irg_end_block (current_ir_graph, get_new_node(get_irg_end_block(current_ir_graph)));
308 if (get_irn_link(get_irg_frame(current_ir_graph)) == NULL) {
309 copy_node (get_irg_frame(current_ir_graph), NULL);
310 copy_preds(get_irg_frame(current_ir_graph), NULL);
312 if (get_irn_link(get_irg_globals(current_ir_graph)) == NULL) {
313 copy_node (get_irg_globals(current_ir_graph), NULL);
314 copy_preds(get_irg_globals(current_ir_graph), NULL);
316 if (get_irn_link(get_irg_args(current_ir_graph)) == NULL) {
317 copy_node (get_irg_args(current_ir_graph), NULL);
318 copy_preds(get_irg_args(current_ir_graph), NULL);
320 set_irg_start (current_ir_graph, get_new_node(get_irg_start(current_ir_graph)));
322 set_irg_start_block(current_ir_graph,
323 get_new_node(get_irg_start_block(current_ir_graph)));
324 set_irg_frame (current_ir_graph, get_new_node(get_irg_frame(current_ir_graph)));
325 set_irg_globals(current_ir_graph, get_new_node(get_irg_globals(current_ir_graph)));
326 set_irg_args (current_ir_graph, get_new_node(get_irg_args(current_ir_graph)));
327 if (get_irn_link(get_irg_bad(current_ir_graph)) == NULL) {
328 copy_node(get_irg_bad(current_ir_graph), NULL);
329 copy_preds(get_irg_bad(current_ir_graph), NULL);
331 set_irg_bad(current_ir_graph, get_new_node(get_irg_bad(current_ir_graph)));
334 /* Copies all reachable nodes to a new obstack. Removes bad inputs
335 from block nodes and the corresponding inputs from Phi nodes.
336 Merges single exit blocks with single entry blocks and removes
338 Adds all new nodes to a new hash table for cse. Does not
339 perform cse, so the hash table might contain common subexpressions. */
340 /* Amroq call this emigrate() */
342 dead_node_elimination(ir_graph *irg) {
344 struct obstack *graveyard_obst = NULL;
345 struct obstack *rebirth_obst = NULL;
347 /* Remember external state of current_ir_graph. */
348 rem = current_ir_graph;
349 current_ir_graph = irg;
351 /* Handle graph state */
352 assert(get_irg_phase_state(current_ir_graph) != phase_building);
353 free_outs(current_ir_graph);
355 if (get_optimize() && get_opt_dead_node_elimination()) {
357 /* A quiet place, where the old obstack can rest in peace,
358 until it will be cremated. */
359 graveyard_obst = irg->obst;
361 /* A new obstack, where the reachable nodes will be copied to. */
362 rebirth_obst = (struct obstack *) xmalloc (sizeof (struct obstack));
363 current_ir_graph->obst = rebirth_obst;
364 obstack_init (current_ir_graph->obst);
366 /* We also need a new hash table for cse */
367 del_identities (irg->value_table);
368 irg->value_table = new_identities ();
370 /* Copy the graph from the old to the new obstack */
373 /* Free memory from old unoptimized obstack */
374 obstack_free(graveyard_obst, 0); /* First empty the obstack ... */
375 xfree (graveyard_obst); /* ... then free it. */
378 current_ir_graph = rem;
381 /**********************************************************************/
382 /* Funcionality for inlining */
383 /**********************************************************************/
385 /* Copy node for inlineing. Copies the node by calling copy_node and
386 then updates the entity if it's a local one. env must be a pointer
387 to the frame type of the procedure. The new entities must be in
388 the link field of the entities. */
390 copy_node_inline (ir_node *n, void *env) {
392 type *frame_tp = (type *)env;
395 if (get_irn_op(n) == op_Sel) {
396 new = get_new_node (n);
397 assert(get_irn_op(new) == op_Sel);
398 if (get_entity_owner(get_Sel_entity(n)) == frame_tp) {
399 set_Sel_entity(new, get_entity_link(get_Sel_entity(n)));
404 void inline_method(ir_node *call, ir_graph *called_graph) {
406 ir_node *post_call, *post_bl;
408 ir_node *end, *end_bl;
413 int arity, n_ret, n_exc, n_res, i, j;
414 type *called_frame, *caller_frame;
416 if (!get_opt_inline()) return;
418 /* Handle graph state */
419 assert(get_irg_phase_state(current_ir_graph) != phase_building);
420 if (get_irg_outs_state(current_ir_graph) == outs_consistent)
421 set_irg_outs_inconsistent(current_ir_graph);
423 /** Check preconditions **/
424 assert(get_irn_op(call) == op_Call);
425 assert(get_Call_type(call) == get_entity_type(get_irg_ent(called_graph)));
426 assert(get_type_tpop(get_Call_type(call)) == type_method);
427 if (called_graph == current_ir_graph) return;
429 /** Part the Call node into two nodes. Pre_call collects the parameters of
430 the procedure and later replaces the Start node of the called graph.
431 Post_call is the old Call node and collects the results of the called
432 graph. Both will end up being a tuple. **/
433 post_bl = get_nodes_Block(call);
434 set_irg_current_block(current_ir_graph, post_bl);
435 /* XxMxPxP of Start + parameter of Call */
437 in[1] = get_Call_mem(call);
438 in[2] = get_irg_frame(current_ir_graph);
439 in[3] = get_irg_globals(current_ir_graph);
440 in[4] = new_Tuple (get_Call_n_params(call),
441 get_Call_param_arr(call));
442 pre_call = new_Tuple(5, in);
445 /** Part the block of the Call node into two blocks.
446 The new block gets the ins of the old block, pre_call and all its
447 predecessors and all Phi nodes. **/
448 part_block(pre_call);
450 /** Prepare state for dead node elimination **/
451 /* Visited flags in calling irg must be >= flag in called irg.
452 Else walker and arity computation will not work. */
453 if (get_irg_visited(current_ir_graph) <= get_irg_visited(called_graph))
454 set_irg_visited(current_ir_graph, get_irg_visited(called_graph)+1); /***/
455 if (get_irg_block_visited(current_ir_graph)< get_irg_block_visited(called_graph))
456 set_irg_block_visited(current_ir_graph, get_irg_block_visited(called_graph));
457 /* Set pre_call as new Start node in link field of the start node of
458 calling graph and pre_calls block as new block for the start block
460 Further mark these nodes so that they are not visited by the
462 set_irn_link(get_irg_start(called_graph), pre_call);
463 set_irn_visited(get_irg_start(called_graph),
464 get_irg_visited(current_ir_graph));/***/
465 set_irn_link(get_irg_start_block(called_graph),
466 get_nodes_Block(pre_call));
467 set_irn_visited(get_irg_start_block(called_graph),
468 get_irg_visited(current_ir_graph)); /***/
470 /* Initialize for compaction of in arrays */
471 inc_irg_block_visited(current_ir_graph);
473 set_Block_block_visited(get_irg_start_block(called_graph),
474 get_irg_block_visited(current_ir_graph) +1 +1); /* count for self edge */
476 /*** Replicate local entities of the called_graph ***/
477 /* copy the entities. */
478 called_frame = get_irg_frame_type(called_graph);
479 for (i = 0; i < get_class_n_member(called_frame); i++) {
480 entity *new_ent, *old_ent;
481 old_ent = get_class_member(called_frame, i);
482 new_ent = copy_entity_own(old_ent, get_cur_frame_type());
483 set_entity_link(old_ent, new_ent);
486 /* visited is > than that of called graph. With this trick visited will
487 remain unchanged so that an outer walker, e.g., searching the call nodes
488 to inline, calling this inline will not visit the inlined nodes. */
489 set_irg_visited(current_ir_graph, get_irg_visited(current_ir_graph)-1);
491 /** Performing dead node elimination inlines the graph **/
492 /* Copies the nodes to the obstack of current_ir_graph. Updates links to new
494 /* @@@ endless loops are not copied!! */
495 irg_walk(get_irg_end(called_graph), copy_node_inline, copy_preds,
496 get_irg_frame_type(called_graph));
498 /* Repair called_graph */
499 set_irg_visited(called_graph, get_irg_visited(current_ir_graph));
500 set_irg_block_visited(called_graph, get_irg_block_visited(current_ir_graph));
501 set_Block_block_visited(get_irg_start_block(called_graph), 0);
503 /*** Merge the end of the inlined procedure with the call site ***/
504 /* We will turn the old Call node into a Tuple with the following
507 0: Phi of all Memories of Return statements.
508 1: Jmp from new Block that merges the control flow from all exception
509 predecessors of the old end block.
510 2: Tuple of all arguments.
511 3: Phi of Exception memories.
514 /** Precompute some values **/
515 end_bl = get_new_node(get_irg_end_block(called_graph));
516 end = get_new_node(get_irg_end(called_graph));
517 arity = get_irn_arity(end_bl); /* arity = n_exc + n_ret */
518 n_res = get_method_n_res(get_Call_type(call));
520 res_pred = (ir_node **) malloc (n_res * sizeof (ir_node *));
521 cf_pred = (ir_node **) malloc (arity * sizeof (ir_node *));
523 set_irg_current_block(current_ir_graph, post_bl); /* just to make sure */
525 /** archive keepalives **/
526 for (i = 0; i < get_irn_arity(end); i++)
527 add_End_keepalive(get_irg_end(current_ir_graph), get_irn_n(end, i));
528 /* The new end node will die, but the in array is not on the obstack ... */
531 /** Collect control flow from Return blocks to post_calls block. Replace
532 Return nodes by Jump nodes. **/
534 for (i = 0; i < arity; i++) {
536 ret = get_irn_n(end_bl, i);
537 if (get_irn_op(ret) == op_Return) {
538 cf_pred[n_ret] = new_r_Jmp(current_ir_graph, get_nodes_Block(ret));
542 set_irn_in(post_bl, n_ret, cf_pred);
544 /** Collect results from Return nodes to post_call. Post_call is
545 turned into a tuple. **/
546 turn_into_tuple(post_call, 4);
547 /* First the Memory-Phi */
549 for (i = 0; i < arity; i++) {
550 ret = get_irn_n(end_bl, i);
551 if (get_irn_op(ret) == op_Return) {
552 cf_pred[n_ret] = get_Return_mem(ret);
556 phi = new_Phi(n_ret, cf_pred, mode_M);
557 set_Tuple_pred(call, 0, phi);
558 set_irn_link(phi, get_irn_link(post_bl)); /* Conserve Phi-list for further inlinings */
559 set_irn_link(post_bl, phi);
560 /* Now the real results */
562 for (j = 0; j < n_res; j++) {
564 for (i = 0; i < arity; i++) {
565 ret = get_irn_n(end_bl, i);
566 if (get_irn_op(ret) == op_Return) {
567 cf_pred[n_ret] = get_Return_res(ret, j);
571 phi = new_Phi(n_ret, cf_pred, get_irn_mode(cf_pred[0]));
573 set_irn_link(phi, get_irn_link(post_bl)); /* Conserve Phi-list for further inlinings */
574 set_irn_link(post_bl, phi);
576 set_Tuple_pred(call, 2, new_Tuple(n_res, res_pred));
578 set_Tuple_pred(call, 2, new_Bad());
580 /* Finally the exception control flow. We need to add a Phi node to
581 collect the memory containing the exception objects. Further we need
582 to add another block to get a correct representation of this Phi. To
583 this block we add a Jmp that resolves into the X output of the Call
584 when the Call is turned into a tuple. */
586 for (i = 0; i < arity; i++) {
588 ret = get_irn_n(end_bl, i);
589 if (is_fragile_op(skip_Proj(ret)) || (get_irn_op(skip_Proj(ret)) == op_Raise)) {
590 cf_pred[n_exc] = ret;
595 new_Block(n_exc, cf_pred); /* whatch it: current_block is changed! */
596 set_Tuple_pred(call, 1, new_Jmp());
597 /* The Phi for the memories with the exception objects */
599 for (i = 0; i < arity; i++) {
601 ret = skip_Proj(get_irn_n(end_bl, i));
602 if (get_irn_op(ret) == op_Call) {
603 cf_pred[n_exc] = new_r_Proj(current_ir_graph, get_nodes_Block(ret), ret, mode_M, 3);
605 } else if (is_fragile_op(ret)) {
606 /* We rely that all cfops have the memory output at the same position. */
607 cf_pred[n_exc] = new_r_Proj(current_ir_graph, get_nodes_Block(ret), ret, mode_M, 0);
609 } else if (get_irn_op(ret) == op_Raise) {
610 cf_pred[n_exc] = new_r_Proj(current_ir_graph, get_nodes_Block(ret), ret, mode_M, 1);
614 set_Tuple_pred(call, 3, new_Phi(n_exc, cf_pred, mode_M));
616 set_Tuple_pred(call, 1, new_Bad());
617 set_Tuple_pred(call, 3, new_Bad());
622 /*** Correct the control flow to the end node.
623 If the exception control flow from the Call directly branched to the
624 end block we now have the following control flow predecessor pattern:
625 ProjX -> Tuple -> Jmp.
626 We must remove the Jmp along with it's empty block and add Jmp's
627 predecessors as predecessors of this end block. ***/
628 /* find the problematic predecessor of the end block. */
629 end_bl = get_irg_end_block(current_ir_graph);
630 for (i = 0; i < get_Block_n_cfgpreds(end_bl); i++) {
631 cf_op = get_Block_cfgpred(end_bl, i);
632 if (get_irn_op(cf_op) == op_Proj) {
633 cf_op = get_Proj_pred(cf_op);
634 if (get_irn_op(cf_op) == op_Tuple) {
635 cf_op = get_Tuple_pred(cf_op, 1);
636 assert(get_irn_op(cf_op) == op_Jmp);
642 if (i < get_Block_n_cfgpreds(end_bl)) {
643 bl = get_nodes_Block(cf_op);
644 arity = get_Block_n_cfgpreds(end_bl) + get_Block_n_cfgpreds(bl) - 1;
645 cf_pred = (ir_node **) malloc (arity * sizeof (ir_node *));
646 for (j = 0; j < i; j++)
647 cf_pred[j] = get_Block_cfgpred(end_bl, j);
648 for (j = j; j < i + get_Block_n_cfgpreds(bl); j++)
649 cf_pred[j] = get_Block_cfgpred(bl, j-i);
650 for (j = j; j < arity; j++)
651 cf_pred[j] = get_Block_cfgpred(end_bl, j-get_Block_n_cfgpreds(bl) +1);
652 set_irn_in(end_bl, arity, cf_pred);
657 /********************************************************************/
658 /* Apply inlineing to small methods. */
659 /********************************************************************/
663 /* It makes no sense to inline too many calls in one procedure. Anyways,
664 I didn't get a version with NEW_ARR_F to run. */
665 #define MAX_INLINE 1024
667 static void collect_calls(ir_node *call, void *env) {
668 ir_node **calls = (ir_node **)env;
671 ir_graph *called_irg;
673 if (get_irn_op(call) != op_Call) return;
675 addr = get_Call_ptr(call);
676 if (get_irn_op(addr) == op_Const) {
677 /* Check whether the constant is the pointer to a compiled entity. */
678 tv = get_Const_tarval(addr);
680 called_irg = get_entity_irg(tv->u.p.ent);
681 if (called_irg && pos < MAX_INLINE) {
682 /* The Call node calls a locally defined method. Remember to inline. */
691 /* Inlines all small methods at call sites where the called address comes
692 from a Const node that references the entity representing the called
694 The size argument is a rough measure for the code size of the method:
695 Methods where the obstack containing the firm graph is smaller than
697 void inline_small_irgs(ir_graph *irg, int size) {
699 ir_node *calls[MAX_INLINE];
700 ir_graph *rem = current_ir_graph;
702 if (!(get_optimize() && get_opt_inline())) return;
704 /*DDME(get_irg_ent(current_ir_graph));*/
706 current_ir_graph = irg;
707 /* Handle graph state */
708 assert(get_irg_phase_state(current_ir_graph) != phase_building);
710 /* Find Call nodes to inline.
711 (We can not inline during a walk of the graph, as inlineing the same
712 method several times changes the visited flag of the walked graph:
713 after the first inlineing visited of the callee equals visited of
714 the caller. With the next inlineing both are increased.) */
716 irg_walk(get_irg_end(irg), NULL, collect_calls, (void *) calls);
718 if ((pos > 0) && (pos < MAX_INLINE)) {
719 /* There are calls to inline */
720 collect_phiprojs(irg);
721 for (i = 0; i < pos; i++) {
724 tv = get_Const_tarval(get_Call_ptr(calls[i]));
725 callee = get_entity_irg(tv->u.p.ent);
726 if ((_obstack_memory_used(callee->obst) - obstack_room(callee->obst)) < size) {
727 /*printf(" inlineing "); DDME(tv->u.p.ent);*/
728 inline_method(calls[i], callee);
733 current_ir_graph = rem;
737 /********************************************************************/
738 /* Code Placement. Pinns all floating nodes to a block where they */
739 /* will be executed only if needed. */
740 /********************************************************************/
742 static pdeq *worklist; /* worklist of ir_node*s */
744 /* Find the earliest correct block for N. --- Place N into the
745 same Block as its dominance-deepest Input. */
747 place_floats_early (ir_node *n)
751 /* we must not run into an infinite loop */
752 assert (irn_not_visited(n));
755 /* Place floating nodes. */
756 if (get_op_pinned(get_irn_op(n)) == floats) {
758 ir_node *b = new_Bad(); /* The block to place this node in */
760 assert(get_irn_op(n) != op_Block);
762 if ((get_irn_op(n) == op_Const) ||
763 (get_irn_op(n) == op_SymConst) ||
765 /* These nodes will not be placed by the loop below. */
766 b = get_irg_start_block(current_ir_graph);
770 /* find the block for this node. */
771 for (i = 0; i < get_irn_arity(n); i++) {
772 ir_node *dep = get_irn_n(n, i);
774 if ((irn_not_visited(dep)) &&
775 (get_op_pinned(get_irn_op(dep)) == floats)) {
776 place_floats_early (dep);
778 /* Because all loops contain at least one pinned node, now all
779 our inputs are either pinned or place_early has already
780 been finished on them. We do not have any unfinished inputs! */
781 dep_block = get_nodes_Block(dep);
782 if ((!is_Bad(dep_block)) &&
783 (get_Block_dom_depth(dep_block) > depth)) {
785 depth = get_Block_dom_depth(dep_block);
787 /* Avoid that the node is placed in the Start block */
788 if ((depth == 1) && (get_Block_dom_depth(get_nodes_Block(n)) > 1)) {
789 b = get_Block_cfg_out(get_irg_start_block(current_ir_graph), 0);
790 assert(b != get_irg_start_block(current_ir_graph));
794 set_nodes_Block(n, b);
797 /* Add predecessors of non floating nodes on worklist. */
798 start = (get_irn_op(n) == op_Block) ? 0 : -1;
799 for (i = start; i < get_irn_arity(n); i++) {
800 ir_node *pred = get_irn_n(n, i);
801 if (irn_not_visited(pred)) {
802 pdeq_putr (worklist, pred);
807 /* Floating nodes form subgraphs that begin at nodes as Const, Load,
808 Start, Call and end at pinned nodes as Store, Call. Place_early
809 places all floating nodes reachable from its argument through floating
810 nodes and adds all beginnings at pinned nodes to the worklist. */
811 inline void place_early () {
816 inc_irg_visited(current_ir_graph);
818 /* this inits the worklist */
819 place_floats_early (get_irg_end(current_ir_graph));
821 /* Work the content of the worklist. */
822 while (!pdeq_empty (worklist)) {
823 ir_node *n = pdeq_getl (worklist);
824 if (irn_not_visited(n)) place_floats_early (n);
827 set_irg_outs_inconsistent(current_ir_graph);
828 current_ir_graph->pinned = pinned;
832 /* deepest common dominance ancestor of DCA and CONSUMER of PRODUCER */
834 consumer_dom_dca (ir_node *dca, ir_node *consumer, ir_node *producer)
838 /* Compute the latest block into which we can place a node so that it is
840 if (get_irn_op(consumer) == op_Phi) {
841 /* our comsumer is a Phi-node, the effective use is in all those
842 blocks through which the Phi-node reaches producer */
844 ir_node *phi_block = get_nodes_Block(consumer);
845 for (i = 0; i < get_irn_arity(consumer); i++) {
846 if (get_irn_n(consumer, i) == producer) {
847 block = get_nodes_Block(get_Block_cfgpred(phi_block, i));
851 assert(is_no_Block(consumer));
852 block = get_nodes_Block(consumer);
855 /* Compute the deepest common ancestor of block and dca. */
857 if (!dca) return block;
858 while (get_Block_dom_depth(block) > get_Block_dom_depth(dca))
859 block = get_Block_idom(block);
860 while (get_Block_dom_depth(dca) > get_Block_dom_depth(block))
861 dca = get_Block_idom(dca);
863 { block = get_Block_idom(block); dca = get_Block_idom(dca); }
869 /* @@@ Needs loop informations. Will implement later interprocedural. */
871 move_out_of_loops (ir_node *n, ir_node *dca)
875 /* Find the region deepest in the dominator tree dominating
876 dca with the least loop nesting depth, but still dominated
877 by our early placement. */
879 while (dca != get_nodes_Block(n)) {
880 dca = get_Block_idom(dca);
881 if (!dca) break; /* should we put assert(dca)? */
882 if (get_Block_loop_depth(dca) < get_Block_loop_depth(best)) {
886 if (get_Block_dom_depth(best) >= get_Block_dom_depth(get_nodes_Block(n)))
887 set_nodes_Block(n, best);
891 /* Find the latest legal block for N and place N into the
892 `optimal' Block between the latest and earliest legal block.
893 The `optimal' block is the dominance-deepest block of those
894 with the least loop-nesting-depth. This places N out of as many
895 loops as possible and then makes it as controldependant as
898 place_floats_late (ir_node *n)
902 assert (irn_not_visited(n)); /* no multiple placement */
904 /* no need to place block nodes, control nodes are already placed. */
905 if ((get_irn_op(n) != op_Block) && (!is_cfop(n)) && (get_irn_mode(n) != mode_X)) {
906 /* Assure that our users are all placed, except the Phi-nodes.
907 --- Each dataflow cycle contains at least one Phi-node. We
908 have to break the `user has to be placed before the
909 producer' dependance cycle and the Phi-nodes are the
910 place to do so, because we need to base our placement on the
911 final region of our users, which is OK with Phi-nodes, as they
912 are pinned, and they never have to be placed after a
913 producer of one of their inputs in the same block anyway. */
914 for (i = 0; i < get_irn_n_outs(n); i++) {
915 ir_node *succ = get_irn_out(n, i);
916 if (irn_not_visited(succ) && (get_irn_op(succ) != op_Phi))
917 place_floats_late (succ);
920 /* We have to determine the final block of this node... except for constants. */
921 if ((get_op_pinned(get_irn_op(n)) == floats) &&
922 (get_irn_op(n) != op_Const) &&
923 (get_irn_op(n) != op_SymConst)) {
924 ir_node *dca = NULL; /* deepest common ancestor in the
925 dominator tree of all nodes'
926 blocks depending on us; our final
927 placement has to dominate DCA. */
928 for (i = 0; i < get_irn_n_outs(n); i++) {
929 dca = consumer_dom_dca (dca, get_irn_out(n, i), n);
931 set_nodes_Block(n, dca);
933 move_out_of_loops (n, dca);
940 /* Add predecessors of all non-floating nodes on list. (Those of floating
941 nodes are placeded already and therefore are marked.) */
942 for (i = 0; i < get_irn_n_outs(n); i++) {
943 if (irn_not_visited(get_irn_out(n, i))) {
944 pdeq_putr (worklist, get_irn_out(n, i));
949 inline void place_late() {
951 inc_irg_visited(current_ir_graph);
953 /* This fills the worklist initially. */
954 place_floats_late(get_irg_start_block(current_ir_graph));
955 /* And now empty the worklist again... */
956 while (!pdeq_empty (worklist)) {
957 ir_node *n = pdeq_getl (worklist);
958 if (irn_not_visited(n)) place_floats_late(n);
962 void place_code(ir_graph *irg) {
963 ir_graph *rem = current_ir_graph;
964 current_ir_graph = irg;
966 if (!(get_optimize() && get_opt_global_cse())) return;
968 /* Handle graph state */
969 assert(get_irg_phase_state(irg) != phase_building);
970 if (get_irg_dom_state(irg) != dom_consistent)
973 /* Place all floating nodes as early as possible. This guarantees
974 a legal code placement. */
975 worklist = new_pdeq ();
978 /* place_early invalidates the outs, place_late needs them. */
980 /* Now move the nodes down in the dominator tree. This reduces the
981 unnecessary executions of the node. */
984 set_irg_outs_inconsistent(current_ir_graph);
986 current_ir_graph = rem;
991 /********************************************************************/
992 /* Control flow optimization. */
993 /* Removes Bad control flow predecessors and empty blocks. A block */
994 /* is empty if it contains only a Jmp node. */
996 /********************************************************************/
999 static void merge_blocks(ir_node *n, void *env) {
1001 set_irn_link(n, NULL);
1003 if (get_irn_op(n) == op_Block) {
1005 for (i = 0; i < get_Block_n_cfgpreds(n); i++)
1006 set_Block_cfgpred(n, i, skip_Tuple(get_Block_cfgpred(n, i)));
1007 } else if (get_irn_mode(n) == mode_X) {
1008 /* We will soon visit a block. Optimize it before visiting! */
1009 ir_node *b = get_nodes_Block(n);
1010 ir_node *new = equivalent_node(b);
1011 while (irn_not_visited(b) && (!is_Bad(new)) && (new != b)) {
1012 /* We would have to run gigo if new is bad. */
1013 if (get_optimize() && get_opt_control_flow()) exchange (b, new);
1015 new = equivalent_node(b);
1017 if (is_Bad(new)) exchange (n, new_Bad());
1021 static void collect_nodes(ir_node *n, void *env) {
1023 if (is_no_Block(n)) {
1024 ir_node *b = get_nodes_Block(n);
1026 if ((get_irn_op(n) == op_Phi)) {
1027 /* Collect Phi nodes to compact ins along with block's ins. */
1028 set_irn_link(n, get_irn_link(b));
1030 } else if (get_irn_op(n) != op_Jmp) { /* Check for non empty block. */
1031 mark_Block_block_visited(b);
1036 /* Returns true if pred is pred of block */
1037 int is_pred_of(ir_node *pred, ir_node *b) {
1039 for (i = 0; i < get_Block_n_cfgpreds(b); i++) {
1040 ir_node *b_pred = get_nodes_Block(get_Block_cfgpred(b, i));
1041 if (b_pred == pred) return 1;
1046 int test_whether_dispensable(ir_node *b, int pos) {
1047 int i, j, n_preds = 1;
1048 int dispensable = 1;
1049 ir_node *cfop = get_Block_cfgpred(b, pos);
1050 ir_node *pred = get_nodes_Block(cfop);
1052 if (get_Block_block_visited(pred) + 1
1053 < get_irg_block_visited(current_ir_graph)) {
1054 if (!get_optimize() || !get_opt_control_flow()) {
1055 /* Mark block so that is will not be removed. */
1056 set_Block_block_visited(pred, get_irg_block_visited(current_ir_graph)-1);
1059 /* Seems to be empty. */
1060 if (!get_irn_link(b)) {
1061 /* There are no Phi nodes ==> dispensable. */
1062 n_preds = get_Block_n_cfgpreds(pred);
1064 /* b's pred blocks and pred's pred blocks must be pairwise disjunct.
1065 Work preds < pos as if they were already removed. */
1066 for (i = 0; i < pos; i++) {
1067 ir_node *b_pred = get_nodes_Block(get_Block_cfgpred(b, i));
1068 if (get_Block_block_visited(b_pred) + 1
1069 < get_irg_block_visited(current_ir_graph)) {
1070 for (j = 0; j < get_Block_n_cfgpreds(b_pred); j++) {
1071 ir_node *b_pred_pred = get_nodes_Block(get_Block_cfgpred(b_pred, j));
1072 if (is_pred_of(b_pred_pred, pred)) dispensable = 0;
1075 if (is_pred_of(b_pred, pred)) dispensable = 0;
1078 for (i = pos +1; i < get_Block_n_cfgpreds(b); i++) {
1079 ir_node *b_pred = get_nodes_Block(get_Block_cfgpred(b, i));
1080 if (is_pred_of(b_pred, pred)) dispensable = 0;
1083 set_Block_block_visited(pred, get_irg_block_visited(current_ir_graph)-1);
1086 n_preds = get_Block_n_cfgpreds(pred);
1094 void optimize_blocks(ir_node *b, void *env) {
1095 int i, j, k, max_preds, n_preds;
1096 ir_node *pred, *phi;
1100 for (i = 0; i < get_Block_n_cfgpreds(b); i++) {
1101 pred = get_Block_cfgpred(b, i);
1102 max_preds += test_whether_dispensable(b, i);
1104 in = (ir_node **) malloc(max_preds * sizeof(ir_node *));
1108 printf(" working on "); DDMN(b);
1109 for (i = 0; i < get_Block_n_cfgpreds(b); i++) {
1110 pred = get_nodes_Block(get_Block_cfgpred(b, i));
1111 if (is_Bad(get_Block_cfgpred(b, i))) {
1112 printf(" removing Bad %i\n ", i);
1113 } else if (get_Block_block_visited(pred) +1
1114 < get_irg_block_visited(current_ir_graph)) {
1115 printf(" removing pred %i ", i); DDMN(pred);
1116 } else { printf(" Nothing to do for "); DDMN(pred); }
1118 ** end Debug output **/
1120 /** Fix the Phi nodes **/
1121 phi = get_irn_link(b);
1123 assert(get_irn_op(phi) == op_Phi);
1124 /* Find the new predecessors for the Phi */
1126 for (i = 0; i < get_Block_n_cfgpreds(b); i++) {
1127 pred = get_nodes_Block(get_Block_cfgpred(b, i));
1128 if (is_Bad(get_Block_cfgpred(b, i))) {
1130 } else if (get_Block_block_visited(pred) +1
1131 < get_irg_block_visited(current_ir_graph)) {
1132 /* It's an empty block and not yet visited. */
1133 ir_node *phi_pred = get_Phi_pred(phi, i);
1134 for (j = 0; j < get_Block_n_cfgpreds(pred); j++) {
1135 if (get_nodes_Block(phi_pred) == pred) {
1136 assert(get_irn_op(phi_pred) == op_Phi); /* Block is empty!! */
1137 in[n_preds] = get_Phi_pred(phi_pred, j);
1139 in[n_preds] = phi_pred;
1144 /* @@@ hier brauche ich schleifeninformation!!! Wenn keine Rueckwaertskante
1145 dann darfs auch keine Verwendung geben. */
1146 if (get_nodes_Block(phi_pred) == pred) {
1147 /* remove the Phi as it might be kept alive. Further there
1148 might be other users. */
1149 exchange(phi_pred, phi); /* geht, is aber doch semantisch falsch! */
1153 in[n_preds] = get_Phi_pred(phi, i);
1158 set_irn_in(phi, n_preds, in);
1159 phi = get_irn_link(phi);
1162 /** Move Phi nodes from removed blocks to this one.
1163 This happens only if merge between loop backedge and single loop entry. **/
1164 for (k = 0; k < get_Block_n_cfgpreds(b); k++) {
1165 pred = get_nodes_Block(get_Block_cfgpred(b, k));
1166 if (get_Block_block_visited(pred) +1
1167 < get_irg_block_visited(current_ir_graph)) {
1168 phi = get_irn_link(pred);
1170 if (get_irn_op(phi) == op_Phi) {
1171 set_nodes_Block(phi, b);
1174 for (i = 0; i < k; i++) {
1175 pred = get_nodes_Block(get_Block_cfgpred(b, i));
1176 if (is_Bad(get_Block_cfgpred(b, i))) {
1178 } else if (get_Block_block_visited(pred) +1
1179 < get_irg_block_visited(current_ir_graph)) {
1180 /* It's an empty block and not yet visited. */
1181 for (j = 0; j < get_Block_n_cfgpreds(pred); j++) {
1182 /* @@@ Hier brauche ich schleifeninformation!!! Kontrllflusskante
1183 muss Rueckwaertskante sein! (An allen vier in[n_preds] = phi
1184 Anweisungen.) Trotzdem tuts bisher!! */
1193 for (i = 0; i < get_Phi_n_preds(phi); i++) {
1194 in[n_preds] = get_Phi_pred(phi, i);
1197 for (i = k+1; i < get_Block_n_cfgpreds(b); i++) {
1198 pred = get_nodes_Block(get_Block_cfgpred(b, i));
1199 if (is_Bad(get_Block_cfgpred(b, i))) {
1201 } else if (get_Block_block_visited(pred) +1
1202 < get_irg_block_visited(current_ir_graph)) {
1203 /* It's an empty block and not yet visited. */
1204 for (j = 0; j < get_Block_n_cfgpreds(pred); j++) {
1213 set_irn_in(phi, n_preds, in);
1215 phi = get_irn_link(phi);
1220 /** Fix the block **/
1222 for (i = 0; i < get_Block_n_cfgpreds(b); i++) {
1223 pred = get_nodes_Block(get_Block_cfgpred(b, i));
1224 if (is_Bad(get_Block_cfgpred(b, i))) {
1226 } else if (get_Block_block_visited(pred) +1
1227 < get_irg_block_visited(current_ir_graph)) {
1228 /* It's an empty block and not yet visited. */
1229 assert(get_Block_n_cfgpreds(b) > 1);
1230 /* Else it should be optimized by equivalent_node. */
1231 for (j = 0; j < get_Block_n_cfgpreds(pred); j++) {
1232 in[n_preds] = get_Block_cfgpred(pred, j);
1235 /* Remove block as it might be kept alive. */
1236 exchange(pred, b/*new_Bad()*/);
1238 in[n_preds] = get_Block_cfgpred(b, i);
1242 set_irn_in(b, n_preds, in);
1246 void optimize_cf(ir_graph *irg) {
1249 ir_node *end = get_irg_end(irg);
1250 ir_graph *rem = current_ir_graph;
1251 current_ir_graph = irg;
1254 /* Handle graph state */
1255 assert(get_irg_phase_state(irg) != phase_building);
1256 if (get_irg_outs_state(current_ir_graph) == outs_consistent)
1257 set_irg_outs_inconsistent(current_ir_graph);
1258 if (get_irg_dom_state(current_ir_graph) == dom_consistent)
1259 set_irg_dom_inconsistent(current_ir_graph);
1261 //DDME(get_irg_ent(irg));
1263 /* Use block visited flag to mark non-empty blocks. */
1264 inc_irg_block_visited(irg);
1265 irg_walk(end, merge_blocks, collect_nodes, NULL);
1267 /* Optimize the standard code. */
1268 irg_block_walk(get_irg_end_block(irg), optimize_blocks, NULL, NULL);
1270 /* Walk all keep alives, optimize them if block, add to new in-array
1271 for end if useful. */
1272 in = NEW_ARR_F (ir_node *, 1);
1273 in[0] = get_nodes_Block(end);
1275 inc_irg_visited(current_ir_graph);
1276 for(i = 0; i < get_End_n_keepalives(end); i++) {
1277 ir_node *ka = get_End_keepalive(end, i);
1278 if (irn_not_visited(ka))
1279 if ((get_irn_op(ka) == op_Block) && Block_not_block_visited(ka)) {
1280 set_irg_block_visited(current_ir_graph, /* Don't walk all the way to Start. */
1281 get_irg_block_visited(current_ir_graph)-1);
1282 irg_block_walk(ka, optimize_blocks, NULL, NULL);
1283 mark_irn_visited(ka);
1284 ARR_APP1 (ir_node *, in, ka);
1285 } else if (get_irn_op(ka) == op_Phi) {
1286 mark_irn_visited(ka);
1287 ARR_APP1 (ir_node *, in, ka);
1290 /* DEL_ARR_F(end->in); GL @@@ tut nicht ! */
1293 current_ir_graph = rem;