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 # include "irbackedge_t.h"
33 /* Defined in iropt.c */
34 pset *new_identities (void);
35 void del_identities (pset *value_table);
36 void add_identities (pset *value_table, ir_node *node);
38 /********************************************************************/
39 /* apply optimizations of iropt to all nodes. */
40 /********************************************************************/
42 void init_link (ir_node *n, void *env) {
43 set_irn_link(n, NULL);
47 optimize_in_place_wrapper (ir_node *n, void *env) {
51 for (i = 0; i < get_irn_arity(n); i++) {
52 optimized = optimize_in_place_2(get_irn_n(n, i));
53 set_irn_n(n, i, optimized);
56 if (get_irn_op(n) == op_Block) {
57 optimized = optimize_in_place_2(n);
58 if (optimized != n) exchange (n, optimized);
63 local_optimize_graph (ir_graph *irg) {
64 ir_graph *rem = current_ir_graph;
65 current_ir_graph = irg;
67 /* Handle graph state */
68 assert(get_irg_phase_state(irg) != phase_building);
69 if (get_opt_global_cse())
70 set_irg_pinned(current_ir_graph, floats);
71 if (get_irg_outs_state(current_ir_graph) == outs_consistent)
72 set_irg_outs_inconsistent(current_ir_graph);
73 if (get_irg_dom_state(current_ir_graph) == dom_consistent)
74 set_irg_dom_inconsistent(current_ir_graph);
76 /* Clean the value_table in irg for the cse. */
77 del_identities(irg->value_table);
78 irg->value_table = new_identities();
80 /* walk over the graph */
81 irg_walk(irg->end, init_link, optimize_in_place_wrapper, NULL);
83 current_ir_graph = rem;
86 /********************************************************************/
87 /* Routines for dead node elimination / copying garbage collection */
89 /********************************************************************/
91 /* Remeber the new node in the old node by using a field all nodes have. */
93 set_new_node (ir_node *old, ir_node *new)
98 /* Get this new node, before the old node is forgotton.*/
100 get_new_node (ir_node * n)
105 /* We use the block_visited flag to mark that we have computed the
106 number of useful predecessors for this block.
107 Further we encode the new arity in this flag in the old blocks.
108 Remembering the arity is useful, as it saves a lot of pointer
109 accesses. This function is called for all Phi and Block nodes
112 compute_new_arity(ir_node *b) {
116 irg_v = get_irg_block_visited(current_ir_graph);
117 block_v = get_Block_block_visited(b);
118 if (block_v >= irg_v) {
119 /* we computed the number of preds for this block and saved it in the
121 return block_v - irg_v;
123 /* compute the number of good predecessors */
124 res = get_irn_arity(b);
125 for (i = 0; i < get_irn_arity(b); i++)
126 if (get_irn_opcode(get_irn_n(b, i)) == iro_Bad) res--;
127 /* save it in the flag. */
128 set_Block_block_visited(b, irg_v + res);
133 static INLINE new_backedge_info(ir_node *n) {
134 switch(get_irn_opcode(n)) {
136 n->attr.block.cg_backedge = NULL;
137 n->attr.block.backedge = new_backedge_arr(current_ir_graph->obst, get_irn_arity(n));
140 n->attr.phi_backedge = new_backedge_arr(current_ir_graph->obst, get_irn_arity(n));
143 n->attr.filter.backedge = new_backedge_arr(current_ir_graph->obst, get_irn_arity(n));
149 /* Copies the node to the new obstack. The Ins of the new node point to
150 the predecessors on the old obstack. For block/phi nodes not all
151 predecessors might be copied. n->link points to the new node.
152 For Phi and Block nodes the function allocates in-arrays with an arity
153 only for useful predecessors. The arity is determined by counting
154 the non-bad predecessors of the block. */
156 copy_node (ir_node *n, void *env) {
160 if (get_irn_opcode(n) == iro_Block) {
162 new_arity = compute_new_arity(n);
163 n->attr.block.graph_arr = NULL;
165 block = get_nodes_Block(n);
166 if (get_irn_opcode(n) == iro_Phi) {
167 new_arity = compute_new_arity(block);
169 new_arity = get_irn_arity(n);
172 nn = new_ir_node(get_irn_dbg_info(n),
179 /* Copy the attributes. These might point to additional data. If this
180 was allocated on the old obstack the pointers now are dangling. This
181 frees e.g. the memory of the graph_arr allocated in new_immBlock. */
183 new_backedge_info(nn);
186 /* printf("\n old node: "); DDMSG2(n);
187 printf(" new node: "); DDMSG2(nn); */
191 /* Copies new predecessors of old node to new node remembered in link.
192 Spare the Bad predecessors of Phi and Block nodes. */
194 copy_preds (ir_node *n, void *env) {
198 nn = get_new_node(n);
200 /* printf("\n old node: "); DDMSG2(n);
201 printf(" new node: "); DDMSG2(nn);
202 printf(" arities: old: %d, new: %d\n", get_irn_arity(n), get_irn_arity(nn)); */
204 if (get_irn_opcode(n) == iro_Block) {
205 /* Don't copy Bad nodes. */
207 for (i = 0; i < get_irn_arity(n); i++)
208 if (get_irn_opcode(get_irn_n(n, i)) != iro_Bad) {
209 set_irn_n (nn, j, get_new_node(get_irn_n(n, i)));
210 //if (is_backedge(n, i)) set_backedge(nn, j);
213 /* repair the block visited flag from above misuse. Repair it in both
214 graphs so that the old one can still be used. */
215 set_Block_block_visited(nn, 0);
216 set_Block_block_visited(n, 0);
217 /* Local optimization could not merge two subsequent blocks if
218 in array contained Bads. Now it's possible.
219 We don't call optimize_in_place as it requires
220 that the fields in ir_graph are set properly. */
221 if ((get_opt_control_flow()) &&
222 (get_Block_n_cfgpreds(nn) == 1) &&
223 (get_irn_op(get_Block_cfgpred(nn, 0)) == op_Jmp))
224 exchange(nn, get_nodes_Block(get_Block_cfgpred(nn, 0)));
225 } else if (get_irn_opcode(n) == iro_Phi) {
226 /* Don't copy node if corresponding predecessor in block is Bad.
227 The Block itself should not be Bad. */
228 block = get_nodes_Block(n);
229 set_irn_n (nn, -1, get_new_node(block));
231 for (i = 0; i < get_irn_arity(n); i++)
232 if (get_irn_opcode(get_irn_n(block, i)) != iro_Bad) {
233 set_irn_n (nn, j, get_new_node(get_irn_n(n, i)));
234 //if (is_backedge(n, i)) set_backedge(nn, j);
237 /* If the pre walker reached this Phi after the post walker visited the
238 block block_visited is > 0. */
239 set_Block_block_visited(get_nodes_Block(n), 0);
240 /* Compacting the Phi's ins might generate Phis with only one
242 if (get_irn_arity(n) == 1)
243 exchange(n, get_irn_n(n, 0));
245 for (i = -1; i < get_irn_arity(n); i++)
246 set_irn_n (nn, i, get_new_node(get_irn_n(n, i)));
248 /* Now the new node is complete. We can add it to the hash table for cse.
249 @@@ inlinening aborts if we identify End. Why? */
250 if(get_irn_op(nn) != op_End)
251 add_identities (current_ir_graph->value_table, nn);
254 /* Copies the graph recursively, compacts the keepalive of the end node. */
257 ir_node *oe, *ne; /* old end, new end */
258 ir_node *ka; /* keep alive */
261 oe = get_irg_end(current_ir_graph);
262 /* copy the end node by hand, allocate dynamic in array! */
263 ne = new_ir_node(get_irn_dbg_info(oe),
270 /* Copy the attributes. Well, there might be some in the future... */
272 set_new_node(oe, ne);
274 /* copy the live nodes */
275 irg_walk(get_nodes_Block(oe), copy_node, copy_preds, NULL);
276 /* copy_preds for the end node ... */
277 set_nodes_Block(ne, get_new_node(get_nodes_Block(oe)));
279 /** ... and now the keep alives. **/
280 /* First pick the not marked block nodes and walk them. We must pick these
281 first as else we will oversee blocks reachable from Phis. */
282 for (i = 0; i < get_irn_arity(oe); i++) {
283 ka = get_irn_n(oe, i);
284 if ((get_irn_op(ka) == op_Block) &&
285 (get_irn_visited(ka) < get_irg_visited(current_ir_graph))) {
286 /* We must keep the block alive and copy everything reachable */
287 set_irg_visited(current_ir_graph, get_irg_visited(current_ir_graph)-1);
288 irg_walk(ka, copy_node, copy_preds, NULL);
289 add_End_keepalive(ne, get_new_node(ka));
293 /* Now pick the Phis. Here we will keep all! */
294 for (i = 0; i < get_irn_arity(oe); i++) {
295 ka = get_irn_n(oe, i);
296 if ((get_irn_op(ka) == op_Phi)) {
297 if (get_irn_visited(ka) < get_irg_visited(current_ir_graph)) {
298 /* We didn't copy the Phi yet. */
299 set_irg_visited(current_ir_graph, get_irg_visited(current_ir_graph)-1);
300 irg_walk(ka, copy_node, copy_preds, NULL);
302 add_End_keepalive(ne, get_new_node(ka));
307 /* Copies the graph reachable from current_ir_graph->end to the obstack
308 in current_ir_graph and fixes the environment.
309 Then fixes the fields in current_ir_graph containing nodes of the
314 /* Not all nodes remembered in current_ir_graph might be reachable
315 from the end node. Assure their link is set to NULL, so that
316 we can test whether new nodes have been computed. */
317 set_irn_link(get_irg_frame (current_ir_graph), NULL);
318 set_irn_link(get_irg_globals(current_ir_graph), NULL);
319 set_irn_link(get_irg_args (current_ir_graph), NULL);
321 /* we use the block walk flag for removing Bads from Blocks ins. */
322 inc_irg_block_visited(current_ir_graph);
327 /* fix the fields in current_ir_graph */
328 old_end = get_irg_end(current_ir_graph);
329 set_irg_end (current_ir_graph, get_new_node(old_end));
331 set_irg_end_block (current_ir_graph, get_new_node(get_irg_end_block(current_ir_graph)));
332 if (get_irn_link(get_irg_frame(current_ir_graph)) == NULL) {
333 copy_node (get_irg_frame(current_ir_graph), NULL);
334 copy_preds(get_irg_frame(current_ir_graph), NULL);
336 if (get_irn_link(get_irg_globals(current_ir_graph)) == NULL) {
337 copy_node (get_irg_globals(current_ir_graph), NULL);
338 copy_preds(get_irg_globals(current_ir_graph), NULL);
340 if (get_irn_link(get_irg_args(current_ir_graph)) == NULL) {
341 copy_node (get_irg_args(current_ir_graph), NULL);
342 copy_preds(get_irg_args(current_ir_graph), NULL);
344 set_irg_start (current_ir_graph, get_new_node(get_irg_start(current_ir_graph)));
346 set_irg_start_block(current_ir_graph,
347 get_new_node(get_irg_start_block(current_ir_graph)));
348 set_irg_frame (current_ir_graph, get_new_node(get_irg_frame(current_ir_graph)));
349 set_irg_globals(current_ir_graph, get_new_node(get_irg_globals(current_ir_graph)));
350 set_irg_args (current_ir_graph, get_new_node(get_irg_args(current_ir_graph)));
351 if (get_irn_link(get_irg_bad(current_ir_graph)) == NULL) {
352 copy_node(get_irg_bad(current_ir_graph), NULL);
353 copy_preds(get_irg_bad(current_ir_graph), NULL);
355 set_irg_bad(current_ir_graph, get_new_node(get_irg_bad(current_ir_graph)));
356 if (get_irn_link(get_irg_unknown(current_ir_graph)) == NULL) {
357 copy_node(get_irg_unknown(current_ir_graph), NULL);
358 copy_preds(get_irg_unknown(current_ir_graph), NULL);
360 set_irg_unknown(current_ir_graph, get_new_node(get_irg_unknown(current_ir_graph)));
363 /* Copies all reachable nodes to a new obstack. Removes bad inputs
364 from block nodes and the corresponding inputs from Phi nodes.
365 Merges single exit blocks with single entry blocks and removes
367 Adds all new nodes to a new hash table for cse. Does not
368 perform cse, so the hash table might contain common subexpressions. */
369 /* Amroq call this emigrate() */
371 dead_node_elimination(ir_graph *irg) {
373 struct obstack *graveyard_obst = NULL;
374 struct obstack *rebirth_obst = NULL;
376 /* Remember external state of current_ir_graph. */
377 rem = current_ir_graph;
378 current_ir_graph = irg;
380 /* Handle graph state */
381 assert(get_irg_phase_state(current_ir_graph) != phase_building);
382 free_outs(current_ir_graph);
384 /* @@@ so far we loose loops when copying */
385 set_irg_loop(current_ir_graph, NULL);
387 if (get_optimize() && get_opt_dead_node_elimination()) {
389 /* A quiet place, where the old obstack can rest in peace,
390 until it will be cremated. */
391 graveyard_obst = irg->obst;
393 /* A new obstack, where the reachable nodes will be copied to. */
394 rebirth_obst = (struct obstack *) xmalloc (sizeof (struct obstack));
395 current_ir_graph->obst = rebirth_obst;
396 obstack_init (current_ir_graph->obst);
398 /* We also need a new hash table for cse */
399 del_identities (irg->value_table);
400 irg->value_table = new_identities ();
402 /* Copy the graph from the old to the new obstack */
405 /* Free memory from old unoptimized obstack */
406 obstack_free(graveyard_obst, 0); /* First empty the obstack ... */
407 xfree (graveyard_obst); /* ... then free it. */
410 current_ir_graph = rem;
413 /**********************************************************************/
414 /* Funcionality for inlining */
415 /**********************************************************************/
417 /* Copy node for inlineing. Copies the node by calling copy_node and
418 then updates the entity if it's a local one. env must be a pointer
419 to the frame type of the procedure. The new entities must be in
420 the link field of the entities. */
422 copy_node_inline (ir_node *n, void *env) {
424 type *frame_tp = (type *)env;
427 if (get_irn_op(n) == op_Sel) {
428 new = get_new_node (n);
429 assert(get_irn_op(new) == op_Sel);
430 if (get_entity_owner(get_Sel_entity(n)) == frame_tp) {
431 set_Sel_entity(new, get_entity_link(get_Sel_entity(n)));
436 void inline_method(ir_node *call, ir_graph *called_graph) {
438 ir_node *post_call, *post_bl;
440 ir_node *end, *end_bl;
444 ir_node *cf_op = NULL, *bl;
445 int arity, n_ret, n_exc, n_res, i, j, rem_opt;
448 if (!get_optimize() || !get_opt_inline()) return;
449 /** Turn off optimizations, this can cause problems when allocating new nodes. **/
450 rem_opt = get_optimize();
453 /* Handle graph state */
454 assert(get_irg_phase_state(current_ir_graph) != phase_building);
455 assert(get_irg_pinned(current_ir_graph) == pinned);
456 assert(get_irg_pinned(called_graph) == pinned);
457 if (get_irg_outs_state(current_ir_graph) == outs_consistent)
458 set_irg_outs_inconsistent(current_ir_graph);
460 /** Check preconditions **/
461 assert(get_irn_op(call) == op_Call);
462 /* assert(get_Call_type(call) == get_entity_type(get_irg_ent(called_graph))); */
464 @@@ TODO does not work for InterfaceIII.java after cgana
465 assert(smaller_type(get_entity_type(get_irg_ent(called_graph)),
466 get_Call_type(call)));
468 assert(get_type_tpop(get_Call_type(call)) == type_method);
469 if (called_graph == current_ir_graph) return;
472 /** Part the Call node into two nodes. Pre_call collects the parameters of
473 the procedure and later replaces the Start node of the called graph.
474 Post_call is the old Call node and collects the results of the called
475 graph. Both will end up being a tuple. **/
476 post_bl = get_nodes_Block(call);
477 set_irg_current_block(current_ir_graph, post_bl);
478 /* XxMxPxP of Start + parameter of Call */
480 in[1] = get_Call_mem(call);
481 in[2] = get_irg_frame(current_ir_graph);
482 in[3] = get_irg_globals(current_ir_graph);
483 in[4] = new_Tuple (get_Call_n_params(call), get_Call_param_arr(call));
484 pre_call = new_Tuple(5, in);
487 /** Part the block of the Call node into two blocks.
488 The new block gets the ins of the old block, pre_call and all its
489 predecessors and all Phi nodes. **/
490 part_block(pre_call);
492 /** Prepare state for dead node elimination **/
493 /* Visited flags in calling irg must be >= flag in called irg.
494 Else walker and arity computation will not work. */
495 if (get_irg_visited(current_ir_graph) <= get_irg_visited(called_graph))
496 set_irg_visited(current_ir_graph, get_irg_visited(called_graph)+1); /***/
497 if (get_irg_block_visited(current_ir_graph)< get_irg_block_visited(called_graph))
498 set_irg_block_visited(current_ir_graph, get_irg_block_visited(called_graph));
499 /* Set pre_call as new Start node in link field of the start node of
500 calling graph and pre_calls block as new block for the start block
502 Further mark these nodes so that they are not visited by the
504 set_irn_link(get_irg_start(called_graph), pre_call);
505 set_irn_visited(get_irg_start(called_graph),
506 get_irg_visited(current_ir_graph));/***/
507 set_irn_link(get_irg_start_block(called_graph),
508 get_nodes_Block(pre_call));
509 set_irn_visited(get_irg_start_block(called_graph),
510 get_irg_visited(current_ir_graph)); /***/
512 /* Initialize for compaction of in arrays */
513 inc_irg_block_visited(current_ir_graph);
515 /*** Replicate local entities of the called_graph ***/
516 /* copy the entities. */
517 called_frame = get_irg_frame_type(called_graph);
518 for (i = 0; i < get_class_n_members(called_frame); i++) {
519 entity *new_ent, *old_ent;
520 old_ent = get_class_member(called_frame, i);
521 new_ent = copy_entity_own(old_ent, get_cur_frame_type());
522 set_entity_link(old_ent, new_ent);
525 /* visited is > than that of called graph. With this trick visited will
526 remain unchanged so that an outer walker, e.g., searching the call nodes
527 to inline, calling this inline will not visit the inlined nodes. */
528 set_irg_visited(current_ir_graph, get_irg_visited(current_ir_graph)-1);
530 /** Performing dead node elimination inlines the graph **/
531 /* Copies the nodes to the obstack of current_ir_graph. Updates links to new
533 /* @@@ endless loops are not copied!! */
534 irg_walk(get_irg_end(called_graph), copy_node_inline, copy_preds,
535 get_irg_frame_type(called_graph));
537 /* Repair called_graph */
538 set_irg_visited(called_graph, get_irg_visited(current_ir_graph));
539 set_irg_block_visited(called_graph, get_irg_block_visited(current_ir_graph));
540 set_Block_block_visited(get_irg_start_block(called_graph), 0);
542 /*** Merge the end of the inlined procedure with the call site ***/
543 /* We will turn the old Call node into a Tuple with the following
546 0: Phi of all Memories of Return statements.
547 1: Jmp from new Block that merges the control flow from all exception
548 predecessors of the old end block.
549 2: Tuple of all arguments.
550 3: Phi of Exception memories.
553 /** Precompute some values **/
554 end_bl = get_new_node(get_irg_end_block(called_graph));
555 end = get_new_node(get_irg_end(called_graph));
556 arity = get_irn_arity(end_bl); /* arity = n_exc + n_ret */
557 n_res = get_method_n_ress(get_Call_type(call));
559 res_pred = (ir_node **) malloc (n_res * sizeof (ir_node *));
560 cf_pred = (ir_node **) malloc (arity * sizeof (ir_node *));
562 set_irg_current_block(current_ir_graph, post_bl); /* just to make sure */
564 /** archive keepalives **/
565 for (i = 0; i < get_irn_arity(end); i++)
566 add_End_keepalive(get_irg_end(current_ir_graph), get_irn_n(end, i));
567 /* The new end node will die, but the in array is not on the obstack ... */
570 /** Collect control flow from Return blocks to post_calls block. Replace
571 Return nodes by Jump nodes. **/
573 for (i = 0; i < arity; i++) {
575 ret = get_irn_n(end_bl, i);
576 if (get_irn_op(ret) == op_Return) {
577 cf_pred[n_ret] = new_r_Jmp(current_ir_graph, get_nodes_Block(ret));
581 set_irn_in(post_bl, n_ret, cf_pred);
583 /** Collect results from Return nodes to post_call. Post_call is
584 turned into a tuple. **/
585 turn_into_tuple(post_call, 4);
586 /* First the Memory-Phi */
588 for (i = 0; i < arity; i++) {
589 ret = get_irn_n(end_bl, i);
590 if (get_irn_op(ret) == op_Return) {
591 cf_pred[n_ret] = get_Return_mem(ret);
595 phi = new_Phi(n_ret, cf_pred, mode_M);
596 set_Tuple_pred(call, 0, phi);
597 set_irn_link(phi, get_irn_link(post_bl)); /* Conserve Phi-list for further inlinings */
598 set_irn_link(post_bl, phi);
599 /* Now the real results */
601 for (j = 0; j < n_res; j++) {
603 for (i = 0; i < arity; i++) {
604 ret = get_irn_n(end_bl, i);
605 if (get_irn_op(ret) == op_Return) {
606 cf_pred[n_ret] = get_Return_res(ret, j);
610 phi = new_Phi(n_ret, cf_pred, get_irn_mode(cf_pred[0]));
612 set_irn_link(phi, get_irn_link(post_bl)); /* Conserve Phi-list for further inlinings */
613 set_irn_link(post_bl, phi);
615 set_Tuple_pred(call, 2, new_Tuple(n_res, res_pred));
617 set_Tuple_pred(call, 2, new_Bad());
619 /* Finally the exception control flow. We need to add a Phi node to
620 collect the memory containing the exception objects. Further we need
621 to add another block to get a correct representation of this Phi. To
622 this block we add a Jmp that resolves into the X output of the Call
623 when the Call is turned into a tuple. */
625 for (i = 0; i < arity; i++) {
627 ret = get_irn_n(end_bl, i);
628 if (is_fragile_op(skip_Proj(ret)) || (get_irn_op(skip_Proj(ret)) == op_Raise)) {
629 cf_pred[n_exc] = ret;
634 new_Block(n_exc, cf_pred); /* watch it: current_block is changed! */
635 set_Tuple_pred(call, 1, new_Jmp());
636 /* The Phi for the memories with the exception objects */
638 for (i = 0; i < arity; i++) {
640 ret = skip_Proj(get_irn_n(end_bl, i));
641 if (get_irn_op(ret) == op_Call) {
642 cf_pred[n_exc] = new_r_Proj(current_ir_graph, get_nodes_Block(ret), ret, mode_M, 3);
644 } else if (is_fragile_op(ret)) {
645 /* We rely that all cfops have the memory output at the same position. */
646 cf_pred[n_exc] = new_r_Proj(current_ir_graph, get_nodes_Block(ret), ret, mode_M, 0);
648 } else if (get_irn_op(ret) == op_Raise) {
649 cf_pred[n_exc] = new_r_Proj(current_ir_graph, get_nodes_Block(ret), ret, mode_M, 1);
653 set_Tuple_pred(call, 3, new_Phi(n_exc, cf_pred, mode_M));
655 set_Tuple_pred(call, 1, new_Bad());
656 set_Tuple_pred(call, 3, new_Bad());
661 /*** Correct the control flow to the end node.
662 If the exception control flow from the Call directly branched to the
663 end block we now have the following control flow predecessor pattern:
664 ProjX -> Tuple -> Jmp.
665 We must remove the Jmp along with it's empty block and add Jmp's
666 predecessors as predecessors of this end block. ***/
667 /* find the problematic predecessor of the end block. */
668 end_bl = get_irg_end_block(current_ir_graph);
669 for (i = 0; i < get_Block_n_cfgpreds(end_bl); i++) {
670 cf_op = get_Block_cfgpred(end_bl, i);
671 if (get_irn_op(cf_op) == op_Proj) {
672 cf_op = get_Proj_pred(cf_op);
673 if (get_irn_op(cf_op) == op_Tuple) {
674 cf_op = get_Tuple_pred(cf_op, 1);
675 assert(get_irn_op(cf_op) == op_Jmp);
681 if (i < get_Block_n_cfgpreds(end_bl)) {
682 bl = get_nodes_Block(cf_op);
683 arity = get_Block_n_cfgpreds(end_bl) + get_Block_n_cfgpreds(bl) - 1;
684 cf_pred = (ir_node **) malloc (arity * sizeof (ir_node *));
685 for (j = 0; j < i; j++)
686 cf_pred[j] = get_Block_cfgpred(end_bl, j);
687 for (j = j; j < i + get_Block_n_cfgpreds(bl); j++)
688 cf_pred[j] = get_Block_cfgpred(bl, j-i);
689 for (j = j; j < arity; j++)
690 cf_pred[j] = get_Block_cfgpred(end_bl, j-get_Block_n_cfgpreds(bl) +1);
691 set_irn_in(end_bl, arity, cf_pred);
695 /** Turn cse back on. **/
696 set_optimize(rem_opt);
699 /********************************************************************/
700 /* Apply inlineing to small methods. */
701 /********************************************************************/
705 /* It makes no sense to inline too many calls in one procedure. Anyways,
706 I didn't get a version with NEW_ARR_F to run. */
707 #define MAX_INLINE 1024
709 static void collect_calls(ir_node *call, void *env) {
710 ir_node **calls = (ir_node **)env;
713 ir_graph *called_irg;
715 if (get_irn_op(call) != op_Call) return;
717 addr = get_Call_ptr(call);
718 if (get_irn_op(addr) == op_Const) {
719 /* Check whether the constant is the pointer to a compiled entity. */
720 tv = get_Const_tarval(addr);
722 called_irg = get_entity_irg(tv->u.p.ent);
723 if (called_irg && pos < MAX_INLINE) {
724 /* The Call node calls a locally defined method. Remember to inline. */
733 /* Inlines all small methods at call sites where the called address comes
734 from a Const node that references the entity representing the called
736 The size argument is a rough measure for the code size of the method:
737 Methods where the obstack containing the firm graph is smaller than
739 void inline_small_irgs(ir_graph *irg, int size) {
741 ir_node *calls[MAX_INLINE];
742 ir_graph *rem = current_ir_graph;
744 if (!(get_optimize() && get_opt_inline())) return;
746 /*DDME(get_irg_ent(current_ir_graph));*/
748 current_ir_graph = irg;
749 /* Handle graph state */
750 assert(get_irg_phase_state(current_ir_graph) != phase_building);
752 /* Find Call nodes to inline.
753 (We can not inline during a walk of the graph, as inlineing the same
754 method several times changes the visited flag of the walked graph:
755 after the first inlineing visited of the callee equals visited of
756 the caller. With the next inlineing both are increased.) */
758 irg_walk(get_irg_end(irg), NULL, collect_calls, (void *) calls);
760 if ((pos > 0) && (pos < MAX_INLINE)) {
761 /* There are calls to inline */
762 collect_phiprojs(irg);
763 for (i = 0; i < pos; i++) {
766 tv = get_Const_tarval(get_Call_ptr(calls[i]));
767 callee = get_entity_irg(tv->u.p.ent);
768 if ((_obstack_memory_used(callee->obst) - obstack_room(callee->obst)) < size) {
769 /*printf(" inlineing "); DDME(tv->u.p.ent);*/
770 inline_method(calls[i], callee);
775 current_ir_graph = rem;
779 /********************************************************************/
780 /* Code Placement. Pinns all floating nodes to a block where they */
781 /* will be executed only if needed. */
782 /********************************************************************/
784 static pdeq *worklist; /* worklist of ir_node*s */
786 /* Find the earliest correct block for N. --- Place N into the
787 same Block as its dominance-deepest Input. */
789 place_floats_early (ir_node *n)
793 /* we must not run into an infinite loop */
794 assert (irn_not_visited(n));
797 /* Place floating nodes. */
798 if (get_op_pinned(get_irn_op(n)) == floats) {
800 ir_node *b = new_Bad(); /* The block to place this node in */
802 assert(get_irn_op(n) != op_Block);
804 if ((get_irn_op(n) == op_Const) ||
805 (get_irn_op(n) == op_SymConst) ||
807 /* These nodes will not be placed by the loop below. */
808 b = get_irg_start_block(current_ir_graph);
812 /* find the block for this node. */
813 for (i = 0; i < get_irn_arity(n); i++) {
814 ir_node *dep = get_irn_n(n, i);
816 if ((irn_not_visited(dep)) &&
817 (get_op_pinned(get_irn_op(dep)) == floats)) {
818 place_floats_early (dep);
820 /* Because all loops contain at least one pinned node, now all
821 our inputs are either pinned or place_early has already
822 been finished on them. We do not have any unfinished inputs! */
823 dep_block = get_nodes_Block(dep);
824 if ((!is_Bad(dep_block)) &&
825 (get_Block_dom_depth(dep_block) > depth)) {
827 depth = get_Block_dom_depth(dep_block);
829 /* Avoid that the node is placed in the Start block */
830 if ((depth == 1) && (get_Block_dom_depth(get_nodes_Block(n)) > 1)) {
831 b = get_Block_cfg_out(get_irg_start_block(current_ir_graph), 0);
832 assert(b != get_irg_start_block(current_ir_graph));
836 set_nodes_Block(n, b);
839 /* Add predecessors of non floating nodes on worklist. */
840 start = (get_irn_op(n) == op_Block) ? 0 : -1;
841 for (i = start; i < get_irn_arity(n); i++) {
842 ir_node *pred = get_irn_n(n, i);
843 if (irn_not_visited(pred)) {
844 pdeq_putr (worklist, pred);
849 /* Floating nodes form subgraphs that begin at nodes as Const, Load,
850 Start, Call and end at pinned nodes as Store, Call. Place_early
851 places all floating nodes reachable from its argument through floating
852 nodes and adds all beginnings at pinned nodes to the worklist. */
853 INLINE void place_early () {
855 inc_irg_visited(current_ir_graph);
857 /* this inits the worklist */
858 place_floats_early (get_irg_end(current_ir_graph));
860 /* Work the content of the worklist. */
861 while (!pdeq_empty (worklist)) {
862 ir_node *n = pdeq_getl (worklist);
863 if (irn_not_visited(n)) place_floats_early (n);
866 set_irg_outs_inconsistent(current_ir_graph);
867 current_ir_graph->pinned = pinned;
871 /* deepest common dominance ancestor of DCA and CONSUMER of PRODUCER */
873 consumer_dom_dca (ir_node *dca, ir_node *consumer, ir_node *producer)
875 ir_node *block = NULL;
877 /* Compute the latest block into which we can place a node so that it is
879 if (get_irn_op(consumer) == op_Phi) {
880 /* our comsumer is a Phi-node, the effective use is in all those
881 blocks through which the Phi-node reaches producer */
883 ir_node *phi_block = get_nodes_Block(consumer);
884 for (i = 0; i < get_irn_arity(consumer); i++) {
885 if (get_irn_n(consumer, i) == producer) {
886 block = get_nodes_Block(get_Block_cfgpred(phi_block, i));
890 assert(is_no_Block(consumer));
891 block = get_nodes_Block(consumer);
894 /* Compute the deepest common ancestor of block and dca. */
896 if (!dca) return block;
897 while (get_Block_dom_depth(block) > get_Block_dom_depth(dca))
898 block = get_Block_idom(block);
899 while (get_Block_dom_depth(dca) > get_Block_dom_depth(block))
900 dca = get_Block_idom(dca);
902 { block = get_Block_idom(block); dca = get_Block_idom(dca); }
907 int get_irn_loop_depth(ir_node *n) {
908 return get_loop_depth(get_irn_loop(n));
911 /* Move n to a block with less loop depth than it's current block. The
912 new block must be dominated by early. */
914 move_out_of_loops (ir_node *n, ir_node *early)
920 /* Find the region deepest in the dominator tree dominating
921 dca with the least loop nesting depth, but still dominated
922 by our early placement. */
923 dca = get_nodes_Block(n);
925 while (dca != early) {
926 dca = get_Block_idom(dca);
927 if (!dca) break; /* should we put assert(dca)? */
928 if (get_irn_loop_depth(dca) < get_irn_loop_depth(best)) {
932 if (best != get_nodes_Block(n)) {
934 printf("Moving out of loop: "); DDMN(n);
935 printf(" Outermost block: "); DDMN(early);
936 printf(" Best block: "); DDMN(best);
937 printf(" Innermost block: "); DDMN(get_nodes_Block(n));
939 set_nodes_Block(n, best);
943 /* Find the latest legal block for N and place N into the
944 `optimal' Block between the latest and earliest legal block.
945 The `optimal' block is the dominance-deepest block of those
946 with the least loop-nesting-depth. This places N out of as many
947 loops as possible and then makes it as controldependant as
950 place_floats_late (ir_node *n)
955 assert (irn_not_visited(n)); /* no multiple placement */
957 /* no need to place block nodes, control nodes are already placed. */
958 if ((get_irn_op(n) != op_Block) &&
960 (get_irn_mode(n) != mode_X)) {
961 /* Remember the early palacement of this block to move it
962 out of loop no further than the early placement. */
963 early = get_nodes_Block(n);
964 /* Assure that our users are all placed, except the Phi-nodes.
965 --- Each dataflow cycle contains at least one Phi-node. We
966 have to break the `user has to be placed before the
967 producer' dependance cycle and the Phi-nodes are the
968 place to do so, because we need to base our placement on the
969 final region of our users, which is OK with Phi-nodes, as they
970 are pinned, and they never have to be placed after a
971 producer of one of their inputs in the same block anyway. */
972 for (i = 0; i < get_irn_n_outs(n); i++) {
973 ir_node *succ = get_irn_out(n, i);
974 if (irn_not_visited(succ) && (get_irn_op(succ) != op_Phi))
975 place_floats_late (succ);
978 /* We have to determine the final block of this node... except for
980 if ((get_op_pinned(get_irn_op(n)) == floats) &&
981 (get_irn_op(n) != op_Const) &&
982 (get_irn_op(n) != op_SymConst)) {
983 ir_node *dca = NULL; /* deepest common ancestor in the
984 dominator tree of all nodes'
985 blocks depending on us; our final
986 placement has to dominate DCA. */
987 for (i = 0; i < get_irn_n_outs(n); i++) {
988 dca = consumer_dom_dca (dca, get_irn_out(n, i), n);
990 set_nodes_Block(n, dca);
992 move_out_of_loops (n, early);
999 /* Add predecessors of all non-floating nodes on list. (Those of floating
1000 nodes are placeded already and therefore are marked.) */
1001 for (i = 0; i < get_irn_n_outs(n); i++) {
1002 if (irn_not_visited(get_irn_out(n, i))) {
1003 pdeq_putr (worklist, get_irn_out(n, i));
1008 INLINE void place_late() {
1010 inc_irg_visited(current_ir_graph);
1012 /* This fills the worklist initially. */
1013 place_floats_late(get_irg_start_block(current_ir_graph));
1014 /* And now empty the worklist again... */
1015 while (!pdeq_empty (worklist)) {
1016 ir_node *n = pdeq_getl (worklist);
1017 if (irn_not_visited(n)) place_floats_late(n);
1023 void place_code(ir_graph *irg) {
1024 ir_graph *rem = current_ir_graph;
1025 current_ir_graph = irg;
1027 if (!(get_optimize() && get_opt_global_cse())) return;
1029 /* Handle graph state */
1030 assert(get_irg_phase_state(irg) != phase_building);
1031 if (get_irg_dom_state(irg) != dom_consistent)
1034 construct_backedges(irg);
1036 /* Place all floating nodes as early as possible. This guarantees
1037 a legal code placement. */
1038 worklist = new_pdeq ();
1041 /* place_early invalidates the outs, place_late needs them. */
1043 /* Now move the nodes down in the dominator tree. This reduces the
1044 unnecessary executions of the node. */
1047 set_irg_outs_inconsistent(current_ir_graph);
1048 del_pdeq (worklist);
1049 current_ir_graph = rem;
1054 /********************************************************************/
1055 /* Control flow optimization. */
1056 /* Removes Bad control flow predecessors and empty blocks. A block */
1057 /* is empty if it contains only a Jmp node. */
1059 /********************************************************************/
1062 static void merge_blocks(ir_node *n, void *env) {
1064 set_irn_link(n, NULL);
1066 if (get_irn_op(n) == op_Block) {
1068 for (i = 0; i < get_Block_n_cfgpreds(n); i++)
1069 set_Block_cfgpred(n, i, skip_Tuple(get_Block_cfgpred(n, i)));
1070 } else if (get_irn_mode(n) == mode_X) {
1071 /* We will soon visit a block. Optimize it before visiting! */
1072 ir_node *b = get_nodes_Block(n);
1073 ir_node *new = equivalent_node(b);
1074 while (irn_not_visited(b) && (!is_Bad(new)) && (new != b)) {
1075 /* We would have to run gigo if new is bad. */
1076 if (get_optimize() && get_opt_control_flow()) exchange (b, new);
1078 new = equivalent_node(b);
1080 if (is_Bad(new)) exchange (n, new_Bad());
1084 static void collect_nodes(ir_node *n, void *env) {
1085 if (is_no_Block(n)) {
1086 ir_node *b = get_nodes_Block(n);
1088 if ((get_irn_op(n) == op_Phi)) {
1089 /* Collect Phi nodes to compact ins along with block's ins. */
1090 set_irn_link(n, get_irn_link(b));
1092 } else if (get_irn_op(n) != op_Jmp) { /* Check for non empty block. */
1093 mark_Block_block_visited(b);
1098 /* Returns true if pred is pred of block */
1099 int is_pred_of(ir_node *pred, ir_node *b) {
1101 for (i = 0; i < get_Block_n_cfgpreds(b); i++) {
1102 ir_node *b_pred = get_nodes_Block(get_Block_cfgpred(b, i));
1103 if (b_pred == pred) return 1;
1108 int test_whether_dispensable(ir_node *b, int pos) {
1109 int i, j, n_preds = 1;
1110 int dispensable = 1;
1111 ir_node *cfop = get_Block_cfgpred(b, pos);
1112 ir_node *pred = get_nodes_Block(cfop);
1114 if (get_Block_block_visited(pred) + 1
1115 < get_irg_block_visited(current_ir_graph)) {
1116 if (!get_optimize() || !get_opt_control_flow()) {
1117 /* Mark block so that is will not be removed. */
1118 set_Block_block_visited(pred, get_irg_block_visited(current_ir_graph)-1);
1121 /* Seems to be empty. */
1122 if (!get_irn_link(b)) {
1123 /* There are no Phi nodes ==> dispensable. */
1124 n_preds = get_Block_n_cfgpreds(pred);
1126 /* b's pred blocks and pred's pred blocks must be pairwise disjunct.
1127 Work preds < pos as if they were already removed. */
1128 for (i = 0; i < pos; i++) {
1129 ir_node *b_pred = get_nodes_Block(get_Block_cfgpred(b, i));
1130 if (get_Block_block_visited(b_pred) + 1
1131 < get_irg_block_visited(current_ir_graph)) {
1132 for (j = 0; j < get_Block_n_cfgpreds(b_pred); j++) {
1133 ir_node *b_pred_pred = get_nodes_Block(get_Block_cfgpred(b_pred, j));
1134 if (is_pred_of(b_pred_pred, pred)) dispensable = 0;
1137 if (is_pred_of(b_pred, pred)) dispensable = 0;
1140 for (i = pos +1; i < get_Block_n_cfgpreds(b); i++) {
1141 ir_node *b_pred = get_nodes_Block(get_Block_cfgpred(b, i));
1142 if (is_pred_of(b_pred, pred)) dispensable = 0;
1145 set_Block_block_visited(pred, get_irg_block_visited(current_ir_graph)-1);
1148 n_preds = get_Block_n_cfgpreds(pred);
1156 void optimize_blocks(ir_node *b, void *env) {
1157 int i, j, k, max_preds, n_preds;
1158 ir_node *pred, *phi;
1161 /* Count the number of predecessor if this block is merged with pred blocks
1164 for (i = 0; i < get_Block_n_cfgpreds(b); i++) {
1165 max_preds += test_whether_dispensable(b, i);
1167 in = (ir_node **) malloc(max_preds * sizeof(ir_node *));
1170 printf(" working on "); DDMN(b);
1171 for (i = 0; i < get_Block_n_cfgpreds(b); i++) {
1172 pred = get_nodes_Block(get_Block_cfgpred(b, i));
1173 if (is_Bad(get_Block_cfgpred(b, i))) {
1174 printf(" removing Bad %i\n ", i);
1175 } else if (get_Block_block_visited(pred) +1
1176 < get_irg_block_visited(current_ir_graph)) {
1177 printf(" removing pred %i ", i); DDMN(pred);
1178 } else { printf(" Nothing to do for "); DDMN(pred); }
1180 ** end Debug output **/
1182 /** Fix the Phi nodes **/
1183 phi = get_irn_link(b);
1185 assert(get_irn_op(phi) == op_Phi);
1186 /* Find the new predecessors for the Phi */
1188 for (i = 0; i < get_Block_n_cfgpreds(b); i++) {
1189 pred = get_nodes_Block(get_Block_cfgpred(b, i));
1190 if (is_Bad(get_Block_cfgpred(b, i))) {
1192 } else if (get_Block_block_visited(pred) +1
1193 < get_irg_block_visited(current_ir_graph)) {
1194 /* It's an empty block and not yet visited. */
1195 ir_node *phi_pred = get_Phi_pred(phi, i);
1196 for (j = 0; j < get_Block_n_cfgpreds(pred); j++) {
1197 if (get_nodes_Block(phi_pred) == pred) {
1198 assert(get_irn_op(phi_pred) == op_Phi); /* Block is empty!! */
1199 in[n_preds] = get_Phi_pred(phi_pred, j);
1201 in[n_preds] = phi_pred;
1206 /* @@@ hier brauche ich schleifeninformation!!! Wenn keine Rueckwaertskante
1207 dann darfs auch keine Verwendung geben. */
1208 if (get_nodes_Block(phi_pred) == pred) {
1209 /* remove the Phi as it might be kept alive. Further there
1210 might be other users. */
1211 exchange(phi_pred, phi); /* geht, is aber doch semantisch falsch! */
1215 in[n_preds] = get_Phi_pred(phi, i);
1220 set_irn_in(phi, n_preds, in);
1221 phi = get_irn_link(phi);
1224 /** Move Phi nodes from removed blocks to this one.
1225 This happens only if merge between loop backedge and single loop entry. **/
1226 for (k = 0; k < get_Block_n_cfgpreds(b); k++) {
1227 pred = get_nodes_Block(get_Block_cfgpred(b, k));
1228 if (get_Block_block_visited(pred) +1
1229 < get_irg_block_visited(current_ir_graph)) {
1230 phi = get_irn_link(pred);
1232 if (get_irn_op(phi) == op_Phi) {
1233 set_nodes_Block(phi, b);
1236 for (i = 0; i < k; i++) {
1237 pred = get_nodes_Block(get_Block_cfgpred(b, i));
1238 if (is_Bad(get_Block_cfgpred(b, i))) {
1240 } else if (get_Block_block_visited(pred) +1
1241 < get_irg_block_visited(current_ir_graph)) {
1242 /* It's an empty block and not yet visited. */
1243 for (j = 0; j < get_Block_n_cfgpreds(pred); j++) {
1244 /* @@@ Hier brauche ich schleifeninformation!!! Kontrllflusskante
1245 muss Rueckwaertskante sein! (An allen vier in[n_preds] = phi
1246 Anweisungen.) Trotzdem tuts bisher!! */
1255 for (i = 0; i < get_Phi_n_preds(phi); i++) {
1256 in[n_preds] = get_Phi_pred(phi, i);
1259 for (i = k+1; i < get_Block_n_cfgpreds(b); i++) {
1260 pred = get_nodes_Block(get_Block_cfgpred(b, i));
1261 if (is_Bad(get_Block_cfgpred(b, i))) {
1263 } else if (get_Block_block_visited(pred) +1
1264 < get_irg_block_visited(current_ir_graph)) {
1265 /* It's an empty block and not yet visited. */
1266 for (j = 0; j < get_Block_n_cfgpreds(pred); j++) {
1275 set_irn_in(phi, n_preds, in);
1277 phi = get_irn_link(phi);
1282 /** Fix the block **/
1284 for (i = 0; i < get_Block_n_cfgpreds(b); i++) {
1285 pred = get_nodes_Block(get_Block_cfgpred(b, i));
1286 if (is_Bad(get_Block_cfgpred(b, i))) {
1288 } else if (get_Block_block_visited(pred) +1
1289 < get_irg_block_visited(current_ir_graph)) {
1290 /* It's an empty block and not yet visited. */
1291 assert(get_Block_n_cfgpreds(b) > 1);
1292 /* Else it should be optimized by equivalent_node. */
1293 for (j = 0; j < get_Block_n_cfgpreds(pred); j++) {
1294 in[n_preds] = get_Block_cfgpred(pred, j);
1297 /* Remove block as it might be kept alive. */
1298 exchange(pred, b/*new_Bad()*/);
1300 in[n_preds] = get_Block_cfgpred(b, i);
1304 set_irn_in(b, n_preds, in);
1308 void optimize_cf(ir_graph *irg) {
1311 ir_node *end = get_irg_end(irg);
1312 ir_graph *rem = current_ir_graph;
1313 current_ir_graph = irg;
1316 /* Handle graph state */
1317 assert(get_irg_phase_state(irg) != phase_building);
1318 if (get_irg_outs_state(current_ir_graph) == outs_consistent)
1319 set_irg_outs_inconsistent(current_ir_graph);
1320 if (get_irg_dom_state(current_ir_graph) == dom_consistent)
1321 set_irg_dom_inconsistent(current_ir_graph);
1323 /* Use block visited flag to mark non-empty blocks. */
1324 inc_irg_block_visited(irg);
1325 irg_walk(end, merge_blocks, collect_nodes, NULL);
1327 /* Optimize the standard code. */
1328 irg_block_walk(get_irg_end_block(irg), optimize_blocks, NULL, NULL);
1330 /* Walk all keep alives, optimize them if block, add to new in-array
1331 for end if useful. */
1332 in = NEW_ARR_F (ir_node *, 1);
1333 in[0] = get_nodes_Block(end);
1335 inc_irg_visited(current_ir_graph);
1336 for(i = 0; i < get_End_n_keepalives(end); i++) {
1337 ir_node *ka = get_End_keepalive(end, i);
1338 if (irn_not_visited(ka)) {
1339 if ((get_irn_op(ka) == op_Block) && Block_not_block_visited(ka)) {
1340 set_irg_block_visited(current_ir_graph, /* Don't walk all the way to Start. */
1341 get_irg_block_visited(current_ir_graph)-1);
1342 irg_block_walk(ka, optimize_blocks, NULL, NULL);
1343 mark_irn_visited(ka);
1344 ARR_APP1 (ir_node *, in, ka);
1345 } else if (get_irn_op(ka) == op_Phi) {
1346 mark_irn_visited(ka);
1347 ARR_APP1 (ir_node *, in, ka);
1351 /* DEL_ARR_F(end->in); GL @@@ tut nicht ! */
1354 current_ir_graph = rem;