3 * File name: ir/ir/irgopt.c
4 * Purpose: Optimizations for a whole ir graph, i.e., a procedure.
5 * Author: Christian Schaefer, Goetz Lindenmaier
6 * Modified by: Sebastian Felis
9 * Copyright: (c) 1998-2003 Universität Karlsruhe
10 * Licence: This file protected by GPL - GNU GENERAL PUBLIC LICENSE.
23 # include "irnode_t.h"
24 # include "irgraph_t.h"
32 # include "pdeq.h" /* Fuer code placement */
35 # include "irbackedge_t.h"
36 # include "irflag_t.h"
37 # include "firmstat.h"
39 /* Defined in iropt.c */
40 pset *new_identities (void);
41 void del_identities (pset *value_table);
42 void add_identities (pset *value_table, ir_node *node);
44 /*------------------------------------------------------------------*/
45 /* apply optimizations of iropt to all nodes. */
46 /*------------------------------------------------------------------*/
48 static void init_link (ir_node *n, void *env) {
49 set_irn_link(n, NULL);
52 #if 0 /* Old version. Avoids Ids.
53 This is not necessary: we do a postwalk, and get_irn_n
54 removes ids anyways. So it's much cheaper to call the
55 optimization less often and use the exchange() algorithm. */
57 optimize_in_place_wrapper (ir_node *n, void *env) {
59 ir_node *optimized, *old;
61 irn_arity = get_irn_arity(n);
62 for (i = 0; i < irn_arity; i++) {
63 /* get_irn_n skips Id nodes, so comparison old != optimized does not
64 show all optimizations. Therefore always set new predecessor. */
65 old = get_irn_intra_n(n, i);
66 optimized = optimize_in_place_2(old);
67 set_irn_n(n, i, optimized);
70 if (get_irn_op(n) == op_Block) {
71 optimized = optimize_in_place_2(n);
72 if (optimized != n) exchange (n, optimized);
77 optimize_in_place_wrapper (ir_node *n, void *env) {
78 ir_node *optimized = optimize_in_place_2(n);
79 if (optimized != n) exchange (n, optimized);
86 local_optimize_graph (ir_graph *irg) {
87 ir_graph *rem = current_ir_graph;
88 current_ir_graph = irg;
90 /* Handle graph state */
91 assert(get_irg_phase_state(irg) != phase_building);
92 if (get_opt_global_cse())
93 set_irg_pinned(current_ir_graph, floats);
94 if (get_irg_outs_state(current_ir_graph) == outs_consistent)
95 set_irg_outs_inconsistent(current_ir_graph);
96 if (get_irg_dom_state(current_ir_graph) == dom_consistent)
97 set_irg_dom_inconsistent(current_ir_graph);
98 set_irg_loopinfo_inconsistent(current_ir_graph);
101 /* Clean the value_table in irg for the cse. */
102 del_identities(irg->value_table);
103 irg->value_table = new_identities();
105 /* walk over the graph */
106 irg_walk(irg->end, init_link, optimize_in_place_wrapper, NULL);
108 current_ir_graph = rem;
111 /*------------------------------------------------------------------*/
112 /* Routines for dead node elimination / copying garbage collection */
113 /* of the obstack. */
114 /*------------------------------------------------------------------*/
117 * Remember the new node in the old node by using a field all nodes have.
120 set_new_node (ir_node *old, ir_node *new)
126 * Get this new node, before the old node is forgotton.
128 static INLINE ir_node *
129 get_new_node (ir_node * n)
135 * We use the block_visited flag to mark that we have computed the
136 * number of useful predecessors for this block.
137 * Further we encode the new arity in this flag in the old blocks.
138 * Remembering the arity is useful, as it saves a lot of pointer
139 * accesses. This function is called for all Phi and Block nodes
143 compute_new_arity(ir_node *b) {
144 int i, res, irn_arity;
147 irg_v = get_irg_block_visited(current_ir_graph);
148 block_v = get_Block_block_visited(b);
149 if (block_v >= irg_v) {
150 /* we computed the number of preds for this block and saved it in the
152 return block_v - irg_v;
154 /* compute the number of good predecessors */
155 res = irn_arity = get_irn_arity(b);
156 for (i = 0; i < irn_arity; i++)
157 if (get_irn_opcode(get_irn_n(b, i)) == iro_Bad) res--;
158 /* save it in the flag. */
159 set_Block_block_visited(b, irg_v + res);
164 /* TODO: add an ir_op operation */
165 static INLINE void new_backedge_info(ir_node *n) {
166 switch(get_irn_opcode(n)) {
168 n->attr.block.cg_backedge = NULL;
169 n->attr.block.backedge = new_backedge_arr(current_ir_graph->obst, get_irn_arity(n));
172 n->attr.phi_backedge = new_backedge_arr(current_ir_graph->obst, get_irn_arity(n));
175 n->attr.filter.backedge = new_backedge_arr(current_ir_graph->obst, get_irn_arity(n));
182 * Copies the node to the new obstack. The Ins of the new node point to
183 * the predecessors on the old obstack. For block/phi nodes not all
184 * predecessors might be copied. n->link points to the new node.
185 * For Phi and Block nodes the function allocates in-arrays with an arity
186 * only for useful predecessors. The arity is determined by counting
187 * the non-bad predecessors of the block.
190 copy_node (ir_node *n, void *env) {
194 /* The end node looses it's flexible in array. This doesn't matter,
195 as dead node elimination builds End by hand, inlineing doesn't use
197 /* assert(n->op == op_End || ((_ARR_DESCR(n->in))->cookie != ARR_F_MAGIC)); */
199 if (get_irn_opcode(n) == iro_Block) {
201 new_arity = compute_new_arity(n);
202 n->attr.block.graph_arr = NULL;
204 block = get_nodes_Block(n);
205 if (get_irn_opcode(n) == iro_Phi) {
206 new_arity = compute_new_arity(block);
208 new_arity = get_irn_arity(n);
211 nn = new_ir_node(get_irn_dbg_info(n),
218 /* Copy the attributes. These might point to additional data. If this
219 was allocated on the old obstack the pointers now are dangling. This
220 frees e.g. the memory of the graph_arr allocated in new_immBlock. */
222 new_backedge_info(nn);
225 /* printf("\n old node: "); DDMSG2(n);
226 printf(" new node: "); DDMSG2(nn); */
231 * Copies new predecessors of old node to new node remembered in link.
232 * Spare the Bad predecessors of Phi and Block nodes.
235 copy_preds (ir_node *n, void *env) {
239 nn = get_new_node(n);
241 /* printf("\n old node: "); DDMSG2(n);
242 printf(" new node: "); DDMSG2(nn);
243 printf(" arities: old: %d, new: %d\n", get_irn_arity(n), get_irn_arity(nn)); */
245 if (get_irn_opcode(n) == iro_Block) {
246 /* Don't copy Bad nodes. */
248 irn_arity = get_irn_arity(n);
249 for (i = 0; i < irn_arity; i++)
250 if (get_irn_opcode(get_irn_n(n, i)) != iro_Bad) {
251 set_irn_n (nn, j, get_new_node(get_irn_n(n, i)));
252 /*if (is_backedge(n, i)) set_backedge(nn, j);*/
255 /* repair the block visited flag from above misuse. Repair it in both
256 graphs so that the old one can still be used. */
257 set_Block_block_visited(nn, 0);
258 set_Block_block_visited(n, 0);
259 /* Local optimization could not merge two subsequent blocks if
260 in array contained Bads. Now it's possible.
261 We don't call optimize_in_place as it requires
262 that the fields in ir_graph are set properly. */
263 if ((get_opt_control_flow_straightening()) &&
264 (get_Block_n_cfgpreds(nn) == 1) &&
265 (get_irn_op(get_Block_cfgpred(nn, 0)) == op_Jmp)) {
266 ir_node *old = get_nodes_Block(get_Block_cfgpred(nn, 0));
268 /* Jmp jumps into the block it is in -- deal self cycle. */
269 assert(is_Bad(get_new_node(get_irg_bad(current_ir_graph))));
270 exchange(nn, get_new_node(get_irg_bad(current_ir_graph)));
275 } else if (get_irn_opcode(n) == iro_Phi) {
276 /* Don't copy node if corresponding predecessor in block is Bad.
277 The Block itself should not be Bad. */
278 block = get_nodes_Block(n);
279 set_irn_n (nn, -1, get_new_node(block));
281 irn_arity = get_irn_arity(n);
282 for (i = 0; i < irn_arity; i++)
283 if (get_irn_opcode(get_irn_n(block, i)) != iro_Bad) {
284 set_irn_n (nn, j, get_new_node(get_irn_n(n, i)));
285 /*if (is_backedge(n, i)) set_backedge(nn, j);*/
288 /* If the pre walker reached this Phi after the post walker visited the
289 block block_visited is > 0. */
290 set_Block_block_visited(get_nodes_Block(n), 0);
291 /* Compacting the Phi's ins might generate Phis with only one
293 if (get_irn_arity(n) == 1)
294 exchange(n, get_irn_n(n, 0));
296 irn_arity = get_irn_arity(n);
297 for (i = -1; i < irn_arity; i++)
298 set_irn_n (nn, i, get_new_node(get_irn_n(n, i)));
300 /* Now the new node is complete. We can add it to the hash table for cse.
301 @@@ inlinening aborts if we identify End. Why? */
302 if(get_irn_op(nn) != op_End)
303 add_identities (current_ir_graph->value_table, nn);
307 * Copies the graph recursively, compacts the keepalive of the end node.
311 ir_node *oe, *ne; /* old end, new end */
312 ir_node *ka; /* keep alive */
315 oe = get_irg_end(current_ir_graph);
316 /* copy the end node by hand, allocate dynamic in array! */
317 ne = new_ir_node(get_irn_dbg_info(oe),
324 /* Copy the attributes. Well, there might be some in the future... */
326 set_new_node(oe, ne);
328 /* copy the live nodes */
329 irg_walk(get_nodes_Block(oe), copy_node, copy_preds, NULL);
330 /* copy_preds for the end node ... */
331 set_nodes_Block(ne, get_new_node(get_nodes_Block(oe)));
333 /*- ... and now the keep alives. -*/
334 /* First pick the not marked block nodes and walk them. We must pick these
335 first as else we will oversee blocks reachable from Phis. */
336 irn_arity = get_irn_arity(oe);
337 for (i = 0; i < irn_arity; i++) {
338 ka = get_irn_intra_n(oe, i);
339 if ((get_irn_op(ka) == op_Block) &&
340 (get_irn_visited(ka) < get_irg_visited(current_ir_graph))) {
341 /* We must keep the block alive and copy everything reachable */
342 set_irg_visited(current_ir_graph, get_irg_visited(current_ir_graph)-1);
343 irg_walk(ka, copy_node, copy_preds, NULL);
344 add_End_keepalive(ne, get_new_node(ka));
348 /* Now pick the Phis. Here we will keep all! */
349 irn_arity = get_irn_arity(oe);
350 for (i = 0; i < irn_arity; i++) {
351 ka = get_irn_intra_n(oe, i);
352 if ((get_irn_op(ka) == op_Phi)) {
353 if (get_irn_visited(ka) < get_irg_visited(current_ir_graph)) {
354 /* We didn't copy the Phi yet. */
355 set_irg_visited(current_ir_graph, get_irg_visited(current_ir_graph)-1);
356 irg_walk(ka, copy_node, copy_preds, NULL);
358 add_End_keepalive(ne, get_new_node(ka));
364 * Copies the graph reachable from current_ir_graph->end to the obstack
365 * in current_ir_graph and fixes the environment.
366 * Then fixes the fields in current_ir_graph containing nodes of the
370 copy_graph_env (void) {
372 /* Not all nodes remembered in current_ir_graph might be reachable
373 from the end node. Assure their link is set to NULL, so that
374 we can test whether new nodes have been computed. */
375 set_irn_link(get_irg_frame (current_ir_graph), NULL);
376 set_irn_link(get_irg_globals(current_ir_graph), NULL);
377 set_irn_link(get_irg_args (current_ir_graph), NULL);
379 /* we use the block walk flag for removing Bads from Blocks ins. */
380 inc_irg_block_visited(current_ir_graph);
385 /* fix the fields in current_ir_graph */
386 old_end = get_irg_end(current_ir_graph);
387 set_irg_end (current_ir_graph, get_new_node(old_end));
388 set_irg_end_except (current_ir_graph, get_irg_end(current_ir_graph));
389 set_irg_end_reg (current_ir_graph, get_irg_end(current_ir_graph));
391 set_irg_end_block (current_ir_graph, get_new_node(get_irg_end_block(current_ir_graph)));
392 if (get_irn_link(get_irg_frame(current_ir_graph)) == NULL) {
393 copy_node (get_irg_frame(current_ir_graph), NULL);
394 copy_preds(get_irg_frame(current_ir_graph), NULL);
396 if (get_irn_link(get_irg_globals(current_ir_graph)) == NULL) {
397 copy_node (get_irg_globals(current_ir_graph), NULL);
398 copy_preds(get_irg_globals(current_ir_graph), NULL);
400 if (get_irn_link(get_irg_args(current_ir_graph)) == NULL) {
401 copy_node (get_irg_args(current_ir_graph), NULL);
402 copy_preds(get_irg_args(current_ir_graph), NULL);
404 set_irg_start (current_ir_graph, get_new_node(get_irg_start(current_ir_graph)));
406 set_irg_start_block(current_ir_graph,
407 get_new_node(get_irg_start_block(current_ir_graph)));
408 set_irg_frame (current_ir_graph, get_new_node(get_irg_frame(current_ir_graph)));
409 set_irg_globals(current_ir_graph, get_new_node(get_irg_globals(current_ir_graph)));
410 set_irg_args (current_ir_graph, get_new_node(get_irg_args(current_ir_graph)));
411 if (get_irn_link(get_irg_bad(current_ir_graph)) == NULL) {
412 copy_node(get_irg_bad(current_ir_graph), NULL);
413 copy_preds(get_irg_bad(current_ir_graph), NULL);
415 set_irg_bad(current_ir_graph, get_new_node(get_irg_bad(current_ir_graph)));
419 * Copies all reachable nodes to a new obstack. Removes bad inputs
420 * from block nodes and the corresponding inputs from Phi nodes.
421 * Merges single exit blocks with single entry blocks and removes
423 * Adds all new nodes to a new hash table for cse. Does not
424 * perform cse, so the hash table might contain common subexpressions.
427 dead_node_elimination(ir_graph *irg) {
429 int rem_ipview = interprocedural_view;
430 struct obstack *graveyard_obst = NULL;
431 struct obstack *rebirth_obst = NULL;
433 stat_dead_node_elim_start(irg);
435 /* Remember external state of current_ir_graph. */
436 rem = current_ir_graph;
437 current_ir_graph = irg;
438 interprocedural_view = 0;
440 /* Handle graph state */
441 assert(get_irg_phase_state(current_ir_graph) != phase_building);
442 assert(get_irg_callee_info_state(current_ir_graph) == irg_callee_info_none);
443 free_outs(current_ir_graph);
445 /* @@@ so far we loose loops when copying */
446 free_loop_information(current_ir_graph);
448 if (get_opt_optimize() && get_opt_dead_node_elimination()) {
450 /* A quiet place, where the old obstack can rest in peace,
451 until it will be cremated. */
452 graveyard_obst = irg->obst;
454 /* A new obstack, where the reachable nodes will be copied to. */
455 rebirth_obst = (struct obstack *) xmalloc (sizeof (struct obstack));
456 current_ir_graph->obst = rebirth_obst;
457 obstack_init (current_ir_graph->obst);
459 /* We also need a new hash table for cse */
460 del_identities (irg->value_table);
461 irg->value_table = new_identities ();
463 /* Copy the graph from the old to the new obstack */
466 /* Free memory from old unoptimized obstack */
467 obstack_free(graveyard_obst, 0); /* First empty the obstack ... */
468 xfree (graveyard_obst); /* ... then free it. */
471 stat_dead_node_elim_stop(irg);
473 current_ir_graph = rem;
474 interprocedural_view = rem_ipview;
478 * Relink bad predeseccors of a block and store the old in array to the
479 * link field. This function is called by relink_bad_predecessors().
480 * The array of link field starts with the block operand at position 0.
481 * If block has bad predecessors, create a new in array without bad preds.
482 * Otherwise let in array untouched.
484 static void relink_bad_block_predecessors(ir_node *n, void *env) {
485 ir_node **new_in, *irn;
486 int i, new_irn_n, old_irn_arity, new_irn_arity = 0;
488 /* if link field of block is NULL, look for bad predecessors otherwise
489 this is allready done */
490 if (get_irn_op(n) == op_Block &&
491 get_irn_link(n) == NULL) {
493 /* save old predecessors in link field (position 0 is the block operand)*/
494 set_irn_link(n, (void *)get_irn_in(n));
496 /* count predecessors without bad nodes */
497 old_irn_arity = get_irn_arity(n);
498 for (i = 0; i < old_irn_arity; i++)
499 if (!is_Bad(get_irn_n(n, i))) new_irn_arity++;
501 /* arity changing: set new predecessors without bad nodes */
502 if (new_irn_arity < old_irn_arity) {
503 /* get new predecessor array without Block predecessor */
504 new_in = NEW_ARR_D (ir_node *, current_ir_graph->obst, (new_irn_arity+1));
506 /* set new predeseccors in array */
509 for (i = 1; i < old_irn_arity; i++) {
510 irn = get_irn_n(n, i);
511 if (!is_Bad(irn)) new_in[new_irn_n++] = irn;
514 } /* ir node has bad predecessors */
516 } /* Block is not relinked */
520 * Relinks Bad predecesors from Bocks and Phis called by walker
521 * remove_bad_predecesors(). If n is a Block, call
522 * relink_bad_block_redecessors(). If n is a Phinode, call also the relinking
523 * function of Phi's Block. If this block has bad predecessors, relink preds
526 static void relink_bad_predecessors(ir_node *n, void *env) {
527 ir_node *block, **old_in;
528 int i, old_irn_arity, new_irn_arity;
530 /* relink bad predeseccors of a block */
531 if (get_irn_op(n) == op_Block)
532 relink_bad_block_predecessors(n, env);
534 /* If Phi node relink its block and its predecessors */
535 if (get_irn_op(n) == op_Phi) {
537 /* Relink predeseccors of phi's block */
538 block = get_nodes_Block(n);
539 if (get_irn_link(block) == NULL)
540 relink_bad_block_predecessors(block, env);
542 old_in = (ir_node **)get_irn_link(block); /* Of Phi's Block */
543 old_irn_arity = ARR_LEN(old_in);
545 /* Relink Phi predeseccors if count of predeseccors changed */
546 if (old_irn_arity != ARR_LEN(get_irn_in(block))) {
547 /* set new predeseccors in array
548 n->in[0] remains the same block */
550 for(i = 1; i < old_irn_arity; i++)
551 if (!is_Bad((ir_node *)old_in[i])) n->in[new_irn_arity++] = n->in[i];
553 ARR_SETLEN(ir_node *, n->in, new_irn_arity);
556 } /* n is a Phi node */
560 * Removes Bad Bad predecesors from Blocks and the corresponding
561 * inputs to Phi nodes as in dead_node_elimination but without
563 * On walking up set the link field to NULL, on walking down call
564 * relink_bad_predecessors() (This function stores the old in array
565 * to the link field and sets a new in array if arity of predecessors
568 void remove_bad_predecessors(ir_graph *irg) {
569 irg_walk_graph(irg, init_link, relink_bad_predecessors, NULL);
573 /*--------------------------------------------------------------------*/
574 /* Funcionality for inlining */
575 /*--------------------------------------------------------------------*/
578 * Copy node for inlineing. Updates attributes that change when
579 * inlineing but not for dead node elimination.
581 * Copies the node by calling copy_node and then updates the entity if
582 * it's a local one. env must be a pointer of the frame type of the
583 * inlined procedure. The new entities must be in the link field of
587 copy_node_inline (ir_node *n, void *env) {
589 type *frame_tp = (type *)env;
592 if (get_irn_op(n) == op_Sel) {
593 new = get_new_node (n);
594 assert(get_irn_op(new) == op_Sel);
595 if (get_entity_owner(get_Sel_entity(n)) == frame_tp) {
596 set_Sel_entity(new, get_entity_link(get_Sel_entity(n)));
598 } else if (get_irn_op(n) == op_Block) {
599 new = get_new_node (n);
600 new->attr.block.irg = current_ir_graph;
604 static void find_addr(ir_node *node, void *env)
606 if (get_irn_opcode(node) == iro_Proj) {
607 if (get_Proj_proj(node) == pn_Start_P_value_arg_base)
613 * currently, we cannot inline two cases:
614 * - call with compound arguments
615 * - graphs that take the address of a parameter
617 * check these condition here
619 static int can_inline(ir_node *call, ir_graph *called_graph)
621 type *call_type = get_Call_type(call);
622 int params, ress, i, res;
624 assert(is_method_type(call_type));
626 params = get_method_n_params(call_type);
627 ress = get_method_n_ress(call_type);
630 for (i = 0; i < params; ++i) {
631 type *p_type = get_method_param_type(call_type, i);
633 if (is_compound_type(p_type))
638 for (i = 0; i < ress; ++i) {
639 type *r_type = get_method_res_type(call_type, i);
641 if (is_compound_type(r_type))
646 irg_walk_graph(called_graph, find_addr, NULL, &res);
651 int inline_method(ir_node *call, ir_graph *called_graph) {
653 ir_node *post_call, *post_bl;
655 ir_node *end, *end_bl;
659 int arity, n_ret, n_exc, n_res, i, j, rem_opt, irn_arity;
662 irg_inline_property prop = get_irg_inline_property(called_graph);
664 if ( (prop != irg_inline_forced) && (!get_opt_optimize() || !get_opt_inline() ||
665 (prop == irg_inline_forbidden))) return 0;
669 * currently, we cannot inline two cases:
670 * - call with compound arguments
671 * - graphs that take the address of a parameter
673 if (! can_inline(call, called_graph))
676 /* -- Turn off optimizations, this can cause problems when allocating new nodes. -- */
677 rem_opt = get_opt_optimize();
680 /* Handle graph state */
681 assert(get_irg_phase_state(current_ir_graph) != phase_building);
682 assert(get_irg_pinned(current_ir_graph) == pinned);
683 assert(get_irg_pinned(called_graph) == pinned);
684 if (get_irg_outs_state(current_ir_graph) == outs_consistent)
685 set_irg_outs_inconsistent(current_ir_graph);
687 /* -- Check preconditions -- */
688 assert(get_irn_op(call) == op_Call);
689 /* @@@ does not work for InterfaceIII.java after cgana
690 assert(get_Call_type(call) == get_entity_type(get_irg_ent(called_graph)));
691 assert(smaller_type(get_entity_type(get_irg_ent(called_graph)),
692 get_Call_type(call)));
694 assert(get_type_tpop(get_Call_type(call)) == type_method);
695 if (called_graph == current_ir_graph) {
696 set_optimize(rem_opt);
700 /* here we know we WILL inline, so inform the statistics */
701 stat_inline(call, called_graph);
703 /* -- Decide how to handle exception control flow: Is there a handler
704 for the Call node, or do we branch directly to End on an exception?
705 exc_handling: 0 There is a handler.
707 2 Exception handling not represented in Firm. -- */
709 ir_node *proj, *Mproj = NULL, *Xproj = NULL;
710 for (proj = (ir_node *)get_irn_link(call); proj; proj = (ir_node *)get_irn_link(proj)) {
711 assert(get_irn_op(proj) == op_Proj);
712 if (get_Proj_proj(proj) == pn_Call_X_except) Xproj = proj;
713 if (get_Proj_proj(proj) == pn_Call_M_except) Mproj = proj;
715 if (Mproj) { assert(Xproj); exc_handling = 0; } /* Mproj */
716 else if (Xproj) { exc_handling = 1; } /* !Mproj && Xproj */
717 else { exc_handling = 2; } /* !Mproj && !Xproj */
722 the procedure and later replaces the Start node of the called graph.
723 Post_call is the old Call node and collects the results of the called
724 graph. Both will end up being a tuple. -- */
725 post_bl = get_nodes_Block(call);
726 set_irg_current_block(current_ir_graph, post_bl);
727 /* XxMxPxP of Start + parameter of Call */
728 in[pn_Start_X_initial_exec] = new_Jmp();
729 in[pn_Start_M] = get_Call_mem(call);
730 in[pn_Start_P_frame_base] = get_irg_frame(current_ir_graph);
731 in[pn_Start_P_globals] = get_irg_globals(current_ir_graph);
732 in[pn_Start_T_args] = new_Tuple(get_Call_n_params(call), get_Call_param_arr(call));
733 /* in[pn_Start_P_value_arg_base] = ??? */
734 pre_call = new_Tuple(5, in);
738 The new block gets the ins of the old block, pre_call and all its
739 predecessors and all Phi nodes. -- */
740 part_block(pre_call);
742 /* -- Prepare state for dead node elimination -- */
743 /* Visited flags in calling irg must be >= flag in called irg.
744 Else walker and arity computation will not work. */
745 if (get_irg_visited(current_ir_graph) <= get_irg_visited(called_graph))
746 set_irg_visited(current_ir_graph, get_irg_visited(called_graph)+1);
747 if (get_irg_block_visited(current_ir_graph)< get_irg_block_visited(called_graph))
748 set_irg_block_visited(current_ir_graph, get_irg_block_visited(called_graph));
749 /* Set pre_call as new Start node in link field of the start node of
750 calling graph and pre_calls block as new block for the start block
752 Further mark these nodes so that they are not visited by the
754 set_irn_link(get_irg_start(called_graph), pre_call);
755 set_irn_visited(get_irg_start(called_graph),
756 get_irg_visited(current_ir_graph));
757 set_irn_link(get_irg_start_block(called_graph),
758 get_nodes_Block(pre_call));
759 set_irn_visited(get_irg_start_block(called_graph),
760 get_irg_visited(current_ir_graph));
762 /* Initialize for compaction of in arrays */
763 inc_irg_block_visited(current_ir_graph);
765 /* -- Replicate local entities of the called_graph -- */
766 /* copy the entities. */
767 called_frame = get_irg_frame_type(called_graph);
768 for (i = 0; i < get_class_n_members(called_frame); i++) {
769 entity *new_ent, *old_ent;
770 old_ent = get_class_member(called_frame, i);
771 new_ent = copy_entity_own(old_ent, get_cur_frame_type());
772 set_entity_link(old_ent, new_ent);
775 /* visited is > than that of called graph. With this trick visited will
776 remain unchanged so that an outer walker, e.g., searching the call nodes
777 to inline, calling this inline will not visit the inlined nodes. */
778 set_irg_visited(current_ir_graph, get_irg_visited(current_ir_graph)-1);
780 /* -- Performing dead node elimination inlines the graph -- */
781 /* Copies the nodes to the obstack of current_ir_graph. Updates links to new
783 /* @@@ endless loops are not copied!! -- they should be, I think... */
784 irg_walk(get_irg_end(called_graph), copy_node_inline, copy_preds,
785 get_irg_frame_type(called_graph));
787 /* Repair called_graph */
788 set_irg_visited(called_graph, get_irg_visited(current_ir_graph));
789 set_irg_block_visited(called_graph, get_irg_block_visited(current_ir_graph));
790 set_Block_block_visited(get_irg_start_block(called_graph), 0);
792 /* -- Merge the end of the inlined procedure with the call site -- */
793 /* We will turn the old Call node into a Tuple with the following
796 0: Phi of all Memories of Return statements.
797 1: Jmp from new Block that merges the control flow from all exception
798 predecessors of the old end block.
799 2: Tuple of all arguments.
800 3: Phi of Exception memories.
801 In case the old Call directly branches to End on an exception we don't
802 need the block merging all exceptions nor the Phi of the exception
806 /* -- Precompute some values -- */
807 end_bl = get_new_node(get_irg_end_block(called_graph));
808 end = get_new_node(get_irg_end(called_graph));
809 arity = get_irn_arity(end_bl); /* arity = n_exc + n_ret */
810 n_res = get_method_n_ress(get_Call_type(call));
812 res_pred = (ir_node **) malloc (n_res * sizeof (ir_node *));
813 cf_pred = (ir_node **) malloc (arity * sizeof (ir_node *));
815 set_irg_current_block(current_ir_graph, post_bl); /* just to make sure */
817 /* -- archive keepalives -- */
818 irn_arity = get_irn_arity(end);
819 for (i = 0; i < irn_arity; i++)
820 add_End_keepalive(get_irg_end(current_ir_graph), get_irn_n(end, i));
822 /* The new end node will die. We need not free as the in array is on the obstack:
823 copy_node only generated 'D' arrays. */
825 /* -- Replace Return nodes by Jump nodes. -- */
827 for (i = 0; i < arity; i++) {
829 ret = get_irn_n(end_bl, i);
830 if (get_irn_op(ret) == op_Return) {
831 cf_pred[n_ret] = new_r_Jmp(current_ir_graph, get_nodes_Block(ret));
835 set_irn_in(post_bl, n_ret, cf_pred);
837 /* -- Build a Tuple for all results of the method.
838 Add Phi node if there was more than one Return. -- */
839 turn_into_tuple(post_call, 4);
840 /* First the Memory-Phi */
842 for (i = 0; i < arity; i++) {
843 ret = get_irn_n(end_bl, i);
844 if (get_irn_op(ret) == op_Return) {
845 cf_pred[n_ret] = get_Return_mem(ret);
849 phi = new_Phi(n_ret, cf_pred, mode_M);
850 set_Tuple_pred(call, pn_Call_M_regular, phi);
851 /* Conserve Phi-list for further inlinings -- but might be optimized */
852 if (get_nodes_Block(phi) == post_bl) {
853 set_irn_link(phi, get_irn_link(post_bl));
854 set_irn_link(post_bl, phi);
856 /* Now the real results */
858 for (j = 0; j < n_res; j++) {
860 for (i = 0; i < arity; i++) {
861 ret = get_irn_n(end_bl, i);
862 if (get_irn_op(ret) == op_Return) {
863 cf_pred[n_ret] = get_Return_res(ret, j);
867 phi = new_Phi(n_ret, cf_pred, get_irn_mode(cf_pred[0]));
869 /* Conserve Phi-list for further inlinings -- but might be optimized */
870 if (get_nodes_Block(phi) == post_bl) {
871 set_irn_link(phi, get_irn_link(post_bl));
872 set_irn_link(post_bl, phi);
875 set_Tuple_pred(call, pn_Call_T_result, new_Tuple(n_res, res_pred));
877 set_Tuple_pred(call, pn_Call_T_result, new_Bad());
879 /* Finally the exception control flow.
880 We have two (three) possible situations:
881 First if the Call branches to an exception handler: We need to add a Phi node to
882 collect the memory containing the exception objects. Further we need
883 to add another block to get a correct representation of this Phi. To
884 this block we add a Jmp that resolves into the X output of the Call
885 when the Call is turned into a tuple.
886 Second the Call branches to End, the exception is not handled. Just
887 add all inlined exception branches to the End node.
888 Third: there is no Exception edge at all. Handle as case two. */
889 if (exc_handling == 0) {
891 for (i = 0; i < arity; i++) {
893 ret = get_irn_n(end_bl, i);
894 if (is_fragile_op(skip_Proj(ret)) || (get_irn_op(skip_Proj(ret)) == op_Raise)) {
895 cf_pred[n_exc] = ret;
900 new_Block(n_exc, cf_pred); /* watch it: current_block is changed! */
901 set_Tuple_pred(call, pn_Call_X_except, new_Jmp());
902 /* The Phi for the memories with the exception objects */
904 for (i = 0; i < arity; i++) {
906 ret = skip_Proj(get_irn_n(end_bl, i));
907 if (get_irn_op(ret) == op_Call) {
908 cf_pred[n_exc] = new_r_Proj(current_ir_graph, get_nodes_Block(ret), ret, mode_M, 3);
910 } else if (is_fragile_op(ret)) {
911 /* We rely that all cfops have the memory output at the same position. */
912 cf_pred[n_exc] = new_r_Proj(current_ir_graph, get_nodes_Block(ret), ret, mode_M, 0);
914 } else if (get_irn_op(ret) == op_Raise) {
915 cf_pred[n_exc] = new_r_Proj(current_ir_graph, get_nodes_Block(ret), ret, mode_M, 1);
919 set_Tuple_pred(call, pn_Call_M_except, new_Phi(n_exc, cf_pred, mode_M));
921 set_Tuple_pred(call, pn_Call_X_except, new_Bad());
922 set_Tuple_pred(call, pn_Call_M_except, new_Bad());
925 ir_node *main_end_bl;
926 int main_end_bl_arity;
929 /* assert(exc_handling == 1 || no exceptions. ) */
931 for (i = 0; i < arity; i++) {
932 ir_node *ret = get_irn_n(end_bl, i);
934 if (is_fragile_op(skip_Proj(ret)) || (get_irn_op(skip_Proj(ret)) == op_Raise)) {
935 cf_pred[n_exc] = ret;
939 main_end_bl = get_irg_end_block(current_ir_graph);
940 main_end_bl_arity = get_irn_arity(main_end_bl);
941 end_preds = (ir_node **) malloc ((n_exc + main_end_bl_arity) * sizeof (ir_node *));
943 for (i = 0; i < main_end_bl_arity; ++i)
944 end_preds[i] = get_irn_n(main_end_bl, i);
945 for (i = 0; i < n_exc; ++i)
946 end_preds[main_end_bl_arity + i] = cf_pred[i];
947 set_irn_in(main_end_bl, n_exc + main_end_bl_arity, end_preds);
948 set_Tuple_pred(call, pn_Call_X_except, new_Bad());
949 set_Tuple_pred(call, pn_Call_M_except, new_Bad());
955 #if 0 /* old. now better, correcter, faster implementation. */
957 /* -- If the exception control flow from the inlined Call directly
958 branched to the end block we now have the following control
959 flow predecessor pattern: ProjX -> Tuple -> Jmp. We must
960 remove the Jmp along with it's empty block and add Jmp's
961 predecessors as predecessors of this end block. No problem if
962 there is no exception, because then branches Bad to End which
964 @@@ can't we know this beforehand: by getting the Proj(1) from
965 the Call link list and checking whether it goes to Proj. */
966 /* find the problematic predecessor of the end block. */
967 end_bl = get_irg_end_block(current_ir_graph);
968 for (i = 0; i < get_Block_n_cfgpreds(end_bl); i++) {
969 cf_op = get_Block_cfgpred(end_bl, i);
970 if (get_irn_op(cf_op) == op_Proj) {
971 cf_op = get_Proj_pred(cf_op);
972 if ((get_irn_op(cf_op) == op_Tuple) && (cf_op == call)) {
973 /* There are unoptimized tuples from inlineing before when no exc */
974 assert(get_Proj_proj(get_Block_cfgpred(end_bl, i)) == pn_Call_X_except);
975 cf_op = get_Tuple_pred(cf_op, pn_Call_X_except);
976 assert(get_irn_op(cf_op) == op_Jmp);
982 if (i < get_Block_n_cfgpreds(end_bl)) {
983 bl = get_nodes_Block(cf_op);
984 arity = get_Block_n_cfgpreds(end_bl) + get_Block_n_cfgpreds(bl) - 1;
985 cf_pred = (ir_node **) malloc (arity * sizeof (ir_node *));
986 for (j = 0; j < i; j++)
987 cf_pred[j] = get_Block_cfgpred(end_bl, j);
988 for (j = j; j < i + get_Block_n_cfgpreds(bl); j++)
989 cf_pred[j] = get_Block_cfgpred(bl, j-i);
990 for (j = j; j < arity; j++)
991 cf_pred[j] = get_Block_cfgpred(end_bl, j-get_Block_n_cfgpreds(bl) +1);
992 set_irn_in(end_bl, arity, cf_pred);
994 /* Remove the exception pred from post-call Tuple. */
995 set_Tuple_pred(call, pn_Call_X_except, new_Bad());
1000 /* -- Turn cse back on. -- */
1001 set_optimize(rem_opt);
1006 /********************************************************************/
1007 /* Apply inlineing to small methods. */
1008 /********************************************************************/
1010 /* It makes no sense to inline too many calls in one procedure. Anyways,
1011 I didn't get a version with NEW_ARR_F to run. */
1012 #define MAX_INLINE 1024
1015 * environment for inlining small irgs
1017 typedef struct _inline_env_t {
1019 ir_node *calls[MAX_INLINE];
1023 * Returns the irg called from a Call node. If the irg is not
1024 * known, NULL is returned.
1026 static ir_graph *get_call_called_irg(ir_node *call) {
1029 ir_graph *called_irg = NULL;
1031 assert(get_irn_op(call) == op_Call);
1033 addr = get_Call_ptr(call);
1034 if (get_irn_op(addr) == op_Const) {
1035 /* Check whether the constant is the pointer to a compiled entity. */
1036 tv = get_Const_tarval(addr);
1037 if (tarval_to_entity(tv))
1038 called_irg = get_entity_irg(tarval_to_entity(tv));
1043 static void collect_calls(ir_node *call, void *env) {
1044 inline_env_t *ienv = env;
1047 ir_graph *called_irg;
1049 if (get_irn_op(call) != op_Call) return;
1051 addr = get_Call_ptr(call);
1052 if (get_irn_op(addr) == op_Const) {
1053 /* Check whether the constant is the pointer to a compiled entity. */
1054 tv = get_Const_tarval(addr);
1055 if (tarval_to_entity(tv)) {
1056 called_irg = get_entity_irg(tarval_to_entity(tv));
1057 if (called_irg && ienv->pos < MAX_INLINE) {
1058 /* The Call node calls a locally defined method. Remember to inline. */
1059 ienv->calls[ienv->pos++] = call;
1066 * Inlines all small methods at call sites where the called address comes
1067 * from a Const node that references the entity representing the called
1069 * The size argument is a rough measure for the code size of the method:
1070 * Methods where the obstack containing the firm graph is smaller than
1073 void inline_small_irgs(ir_graph *irg, int size) {
1075 ir_graph *rem = current_ir_graph;
1078 if (!(get_opt_optimize() && get_opt_inline())) return;
1080 current_ir_graph = irg;
1081 /* Handle graph state */
1082 assert(get_irg_phase_state(current_ir_graph) != phase_building);
1083 assert(get_irg_callee_info_state(current_ir_graph) == irg_callee_info_none);
1085 /* Find Call nodes to inline.
1086 (We can not inline during a walk of the graph, as inlineing the same
1087 method several times changes the visited flag of the walked graph:
1088 after the first inlineing visited of the callee equals visited of
1089 the caller. With the next inlineing both are increased.) */
1091 irg_walk(get_irg_end(irg), NULL, collect_calls, &env);
1093 if ((env.pos > 0) && (env.pos < MAX_INLINE)) {
1094 /* There are calls to inline */
1095 collect_phiprojs(irg);
1096 for (i = 0; i < env.pos; i++) {
1099 tv = get_Const_tarval(get_Call_ptr(env.calls[i]));
1100 callee = get_entity_irg(tarval_to_entity(tv));
1101 if (((_obstack_memory_used(callee->obst) - obstack_room(callee->obst)) < size) ||
1102 (get_irg_inline_property(callee) == irg_inline_forced)) {
1103 inline_method(env.calls[i], callee);
1108 current_ir_graph = rem;
1112 * Environment for inlining irgs.
1115 int n_nodes; /**< Nodes in graph except Id, Tuple, Proj, Start, End */
1116 int n_nodes_orig; /**< for statistics */
1117 eset *call_nodes; /**< All call nodes in this graph */
1119 int n_call_nodes_orig; /**< for statistics */
1120 int n_callers; /**< Number of known graphs that call this graphs. */
1121 int n_callers_orig; /**< for statistics */
1124 static inline_irg_env *new_inline_irg_env(void) {
1125 inline_irg_env *env = malloc(sizeof(inline_irg_env));
1126 env->n_nodes = -2; /* uncount Start, End */
1127 env->n_nodes_orig = -2; /* uncount Start, End */
1128 env->call_nodes = eset_create();
1129 env->n_call_nodes = 0;
1130 env->n_call_nodes_orig = 0;
1132 env->n_callers_orig = 0;
1136 static void free_inline_irg_env(inline_irg_env *env) {
1137 eset_destroy(env->call_nodes);
1141 static void collect_calls2(ir_node *call, void *env) {
1142 inline_irg_env *x = (inline_irg_env *)env;
1143 ir_op *op = get_irn_op(call);
1146 /* count nodes in irg */
1147 if (op != op_Proj && op != op_Tuple && op != op_Sync) {
1152 if (op != op_Call) return;
1154 /* collect all call nodes */
1155 eset_insert(x->call_nodes, (void *)call);
1157 x->n_call_nodes_orig++;
1159 /* count all static callers */
1160 callee = get_call_called_irg(call);
1162 ((inline_irg_env *)get_irg_link(callee))->n_callers++;
1163 ((inline_irg_env *)get_irg_link(callee))->n_callers_orig++;
1167 INLINE static int is_leave(ir_graph *irg) {
1168 return (((inline_irg_env *)get_irg_link(irg))->n_call_nodes == 0);
1171 INLINE static int is_smaller(ir_graph *callee, int size) {
1172 return (((inline_irg_env *)get_irg_link(callee))->n_nodes < size);
1177 * Inlines small leave methods at call sites where the called address comes
1178 * from a Const node that references the entity representing the called
1180 * The size argument is a rough measure for the code size of the method:
1181 * Methods where the obstack containing the firm graph is smaller than
1184 void inline_leave_functions(int maxsize, int leavesize, int size) {
1185 inline_irg_env *env;
1186 int i, n_irgs = get_irp_n_irgs();
1187 ir_graph *rem = current_ir_graph;
1190 if (!(get_opt_optimize() && get_opt_inline())) return;
1192 /* extend all irgs by a temporary data structure for inlineing. */
1193 for (i = 0; i < n_irgs; ++i)
1194 set_irg_link(get_irp_irg(i), new_inline_irg_env());
1196 /* Precompute information in temporary data structure. */
1197 for (i = 0; i < n_irgs; ++i) {
1198 current_ir_graph = get_irp_irg(i);
1199 assert(get_irg_phase_state(current_ir_graph) != phase_building);
1200 assert(get_irg_callee_info_state(current_ir_graph) == irg_callee_info_none);
1202 irg_walk(get_irg_end(current_ir_graph), NULL, collect_calls2,
1203 get_irg_link(current_ir_graph));
1207 Inline leaves recursively -- we might construct new leaves. */
1208 /* int itercnt = 1; */
1209 while (did_inline) {
1210 /* printf("iteration %d\n", itercnt++); */
1212 for (i = 0; i < n_irgs; ++i) {
1215 int phiproj_computed = 0;
1217 current_ir_graph = get_irp_irg(i);
1218 env = (inline_irg_env *)get_irg_link(current_ir_graph);
1220 /* we can not walk and change a set, nor remove from it.
1222 walkset = env->call_nodes;
1223 env->call_nodes = eset_create();
1224 for (call = eset_first(walkset); call; call = eset_next(walkset)) {
1225 inline_irg_env *callee_env;
1226 ir_graph *callee = get_call_called_irg(call);
1228 if (env->n_nodes > maxsize) break;
1230 ((is_leave(callee) && is_smaller(callee, leavesize)) ||
1231 (get_irg_inline_property(callee) == irg_inline_forced))) {
1232 if (!phiproj_computed) {
1233 phiproj_computed = 1;
1234 collect_phiprojs(current_ir_graph);
1236 callee_env = (inline_irg_env *)get_irg_link(callee);
1237 /* printf(" %s: Inlineing %s.\n", get_entity_name(get_irg_entity(current_ir_graph)), */
1238 /* get_entity_name(get_irg_entity(callee))); */
1239 if (inline_method(call, callee)) {
1241 env->n_call_nodes--;
1242 eset_insert_all(env->call_nodes, callee_env->call_nodes);
1243 env->n_call_nodes += callee_env->n_call_nodes;
1244 env->n_nodes += callee_env->n_nodes;
1245 callee_env->n_callers--;
1248 eset_insert(env->call_nodes, call);
1251 eset_destroy(walkset);
1255 /* printf("Non leaves\n"); */
1256 /* inline other small functions. */
1257 for (i = 0; i < n_irgs; ++i) {
1260 int phiproj_computed = 0;
1262 current_ir_graph = get_irp_irg(i);
1263 env = (inline_irg_env *)get_irg_link(current_ir_graph);
1265 /* we can not walk and change a set, nor remove from it.
1267 walkset = env->call_nodes;
1268 env->call_nodes = eset_create();
1269 for (call = eset_first(walkset); call; call = eset_next(walkset)) {
1270 inline_irg_env *callee_env;
1271 ir_graph *callee = get_call_called_irg(call);
1273 if (env->n_nodes > maxsize) break;
1274 if (callee && is_smaller(callee, size)) {
1275 if (!phiproj_computed) {
1276 phiproj_computed = 1;
1277 collect_phiprojs(current_ir_graph);
1279 callee_env = (inline_irg_env *)get_irg_link(callee);
1280 /* printf(" %s: Inlineing %s.\n", get_entity_name(get_irg_entity(current_ir_graph)), */
1281 /* get_entity_name(get_irg_entity(callee))); */
1282 if (inline_method(call, callee)) {
1284 env->n_call_nodes--;
1285 eset_insert_all(env->call_nodes, callee_env->call_nodes);
1286 env->n_call_nodes += callee_env->n_call_nodes;
1287 env->n_nodes += callee_env->n_nodes;
1288 callee_env->n_callers--;
1291 eset_insert(env->call_nodes, call);
1294 eset_destroy(walkset);
1297 for (i = 0; i < n_irgs; ++i) {
1298 current_ir_graph = get_irp_irg(i);
1300 env = (inline_irg_env *)get_irg_link(current_ir_graph);
1301 if ((env->n_call_nodes_orig != env->n_call_nodes) ||
1302 (env->n_callers_orig != env->n_callers))
1303 printf("Nodes:%3d ->%3d, calls:%3d ->%3d, callers:%3d ->%3d, -- %s\n",
1304 env->n_nodes_orig, env->n_nodes, env->n_call_nodes_orig, env->n_call_nodes,
1305 env->n_callers_orig, env->n_callers,
1306 get_entity_name(get_irg_entity(current_ir_graph)));
1308 free_inline_irg_env((inline_irg_env *)get_irg_link(current_ir_graph));
1311 current_ir_graph = rem;
1314 /*******************************************************************/
1315 /* Code Placement. Pins all floating nodes to a block where they */
1316 /* will be executed only if needed. */
1317 /*******************************************************************/
1320 * Find the earliest correct block for N. --- Place N into the
1321 * same Block as its dominance-deepest Input.
1324 place_floats_early(ir_node *n, pdeq *worklist)
1326 int i, start, irn_arity;
1328 /* we must not run into an infinite loop */
1329 assert (irn_not_visited(n));
1330 mark_irn_visited(n);
1332 /* Place floating nodes. */
1333 if (get_op_pinned(get_irn_op(n)) == floats) {
1335 ir_node *b = new_Bad(); /* The block to place this node in */
1337 assert(get_irn_op(n) != op_Block);
1339 if ((get_irn_op(n) == op_Const) ||
1340 (get_irn_op(n) == op_SymConst) ||
1342 (get_irn_op(n) == op_Unknown)) {
1343 /* These nodes will not be placed by the loop below. */
1344 b = get_irg_start_block(current_ir_graph);
1348 /* find the block for this node. */
1349 irn_arity = get_irn_arity(n);
1350 for (i = 0; i < irn_arity; i++) {
1351 ir_node *dep = get_irn_n(n, i);
1353 if ((irn_not_visited(dep)) &&
1354 (get_op_pinned(get_irn_op(dep)) == floats)) {
1355 place_floats_early(dep, worklist);
1357 /* Because all loops contain at least one pinned node, now all
1358 our inputs are either pinned or place_early has already
1359 been finished on them. We do not have any unfinished inputs! */
1360 dep_block = get_nodes_Block(dep);
1361 if ((!is_Bad(dep_block)) &&
1362 (get_Block_dom_depth(dep_block) > depth)) {
1364 depth = get_Block_dom_depth(dep_block);
1366 /* Avoid that the node is placed in the Start block */
1367 if ((depth == 1) && (get_Block_dom_depth(get_nodes_Block(n)) > 1)) {
1368 b = get_Block_cfg_out(get_irg_start_block(current_ir_graph), 0);
1369 assert(b != get_irg_start_block(current_ir_graph));
1373 set_nodes_Block(n, b);
1376 /* Add predecessors of non floating nodes on worklist. */
1377 start = (get_irn_op(n) == op_Block) ? 0 : -1;
1378 irn_arity = get_irn_arity(n);
1379 for (i = start; i < irn_arity; i++) {
1380 ir_node *pred = get_irn_n(n, i);
1381 if (irn_not_visited(pred)) {
1382 pdeq_putr (worklist, pred);
1388 * Floating nodes form subgraphs that begin at nodes as Const, Load,
1389 * Start, Call and end at pinned nodes as Store, Call. Place_early
1390 * places all floating nodes reachable from its argument through floating
1391 * nodes and adds all beginnings at pinned nodes to the worklist.
1393 static INLINE void place_early(pdeq* worklist) {
1395 inc_irg_visited(current_ir_graph);
1397 /* this inits the worklist */
1398 place_floats_early(get_irg_end(current_ir_graph), worklist);
1400 /* Work the content of the worklist. */
1401 while (!pdeq_empty (worklist)) {
1402 ir_node *n = pdeq_getl (worklist);
1403 if (irn_not_visited(n)) place_floats_early(n, worklist);
1406 set_irg_outs_inconsistent(current_ir_graph);
1407 current_ir_graph->pinned = pinned;
1411 /** deepest common dominance ancestor of DCA and CONSUMER of PRODUCER. */
1413 consumer_dom_dca (ir_node *dca, ir_node *consumer, ir_node *producer)
1415 ir_node *block = NULL;
1417 /* Compute the latest block into which we can place a node so that it is
1419 if (get_irn_op(consumer) == op_Phi) {
1420 /* our consumer is a Phi-node, the effective use is in all those
1421 blocks through which the Phi-node reaches producer */
1423 ir_node *phi_block = get_nodes_Block(consumer);
1424 irn_arity = get_irn_arity(consumer);
1425 for (i = 0; i < irn_arity; i++) {
1426 if (get_irn_n(consumer, i) == producer) {
1427 block = get_nodes_Block(get_Block_cfgpred(phi_block, i));
1431 assert(is_no_Block(consumer));
1432 block = get_nodes_Block(consumer);
1435 /* Compute the deepest common ancestor of block and dca. */
1437 if (!dca) return block;
1438 while (get_Block_dom_depth(block) > get_Block_dom_depth(dca))
1439 block = get_Block_idom(block);
1440 while (get_Block_dom_depth(dca) > get_Block_dom_depth(block))
1441 dca = get_Block_idom(dca);
1442 while (block != dca)
1443 { block = get_Block_idom(block); dca = get_Block_idom(dca); }
1448 static INLINE int get_irn_loop_depth(ir_node *n) {
1449 return get_loop_depth(get_irn_loop(n));
1453 * Move n to a block with less loop depth than it's current block. The
1454 * new block must be dominated by early.
1457 move_out_of_loops (ir_node *n, ir_node *early)
1459 ir_node *best, *dca;
1463 /* Find the region deepest in the dominator tree dominating
1464 dca with the least loop nesting depth, but still dominated
1465 by our early placement. */
1466 dca = get_nodes_Block(n);
1468 while (dca != early) {
1469 dca = get_Block_idom(dca);
1470 if (!dca) break; /* should we put assert(dca)? */
1471 if (get_irn_loop_depth(dca) < get_irn_loop_depth(best)) {
1475 if (best != get_nodes_Block(n)) {
1477 printf("Moving out of loop: "); DDMN(n);
1478 printf(" Outermost block: "); DDMN(early);
1479 printf(" Best block: "); DDMN(best);
1480 printf(" Innermost block: "); DDMN(get_nodes_Block(n));
1482 set_nodes_Block(n, best);
1487 * Find the latest legal block for N and place N into the
1488 * `optimal' Block between the latest and earliest legal block.
1489 * The `optimal' block is the dominance-deepest block of those
1490 * with the least loop-nesting-depth. This places N out of as many
1491 * loops as possible and then makes it as control dependant as
1495 place_floats_late(ir_node *n, pdeq *worklist)
1500 assert (irn_not_visited(n)); /* no multiple placement */
1502 /* no need to place block nodes, control nodes are already placed. */
1503 if ((get_irn_op(n) != op_Block) &&
1505 (get_irn_mode(n) != mode_X)) {
1506 /* Remember the early placement of this block to move it
1507 out of loop no further than the early placement. */
1508 early = get_nodes_Block(n);
1509 /* Assure that our users are all placed, except the Phi-nodes.
1510 --- Each data flow cycle contains at least one Phi-node. We
1511 have to break the `user has to be placed before the
1512 producer' dependence cycle and the Phi-nodes are the
1513 place to do so, because we need to base our placement on the
1514 final region of our users, which is OK with Phi-nodes, as they
1515 are pinned, and they never have to be placed after a
1516 producer of one of their inputs in the same block anyway. */
1517 for (i = 0; i < get_irn_n_outs(n); i++) {
1518 ir_node *succ = get_irn_out(n, i);
1519 if (irn_not_visited(succ) && (get_irn_op(succ) != op_Phi))
1520 place_floats_late(succ, worklist);
1523 /* We have to determine the final block of this node... except for
1525 if ((get_op_pinned(get_irn_op(n)) == floats) &&
1526 (get_irn_op(n) != op_Const) &&
1527 (get_irn_op(n) != op_SymConst)) {
1528 ir_node *dca = NULL; /* deepest common ancestor in the
1529 dominator tree of all nodes'
1530 blocks depending on us; our final
1531 placement has to dominate DCA. */
1532 for (i = 0; i < get_irn_n_outs(n); i++) {
1533 dca = consumer_dom_dca (dca, get_irn_out(n, i), n);
1535 set_nodes_Block(n, dca);
1537 move_out_of_loops (n, early);
1541 mark_irn_visited(n);
1543 /* Add predecessors of all non-floating nodes on list. (Those of floating
1544 nodes are placeded already and therefore are marked.) */
1545 for (i = 0; i < get_irn_n_outs(n); i++) {
1546 if (irn_not_visited(get_irn_out(n, i))) {
1547 pdeq_putr (worklist, get_irn_out(n, i));
1552 static INLINE void place_late(pdeq* worklist) {
1554 inc_irg_visited(current_ir_graph);
1556 /* This fills the worklist initially. */
1557 place_floats_late(get_irg_start_block(current_ir_graph), worklist);
1558 /* And now empty the worklist again... */
1559 while (!pdeq_empty (worklist)) {
1560 ir_node *n = pdeq_getl (worklist);
1561 if (irn_not_visited(n)) place_floats_late(n, worklist);
1565 void place_code(ir_graph *irg) {
1567 ir_graph *rem = current_ir_graph;
1569 current_ir_graph = irg;
1571 if (!(get_opt_optimize() && get_opt_global_cse())) return;
1573 /* Handle graph state */
1574 assert(get_irg_phase_state(irg) != phase_building);
1575 if (get_irg_dom_state(irg) != dom_consistent)
1578 if (1 || get_irg_loopinfo_state(irg) != loopinfo_consistent) {
1579 free_loop_information(irg);
1580 construct_backedges(irg);
1583 /* Place all floating nodes as early as possible. This guarantees
1584 a legal code placement. */
1585 worklist = new_pdeq();
1586 place_early(worklist);
1588 /* place_early invalidates the outs, place_late needs them. */
1590 /* Now move the nodes down in the dominator tree. This reduces the
1591 unnecessary executions of the node. */
1592 place_late(worklist);
1594 set_irg_outs_inconsistent(current_ir_graph);
1595 set_irg_loopinfo_inconsistent(current_ir_graph);
1597 current_ir_graph = rem;
1602 /********************************************************************/
1603 /* Control flow optimization. */
1604 /* Removes Bad control flow predecessors and empty blocks. A block */
1605 /* is empty if it contains only a Jmp node. */
1606 /* Blocks can only be removed if they are not needed for the */
1607 /* semantics of Phi nodes. */
1608 /********************************************************************/
1611 * Removes Tuples from Block control flow predecessors.
1612 * Optimizes blocks with equivalent_node().
1613 * Replaces n by Bad if n is unreachable control flow.
1615 static void merge_blocks(ir_node *n, void *env) {
1617 set_irn_link(n, NULL);
1619 if (get_irn_op(n) == op_Block) {
1621 for (i = 0; i < get_Block_n_cfgpreds(n); i++)
1622 /* GL @@@ : is this possible? if (get_opt_normalize()) -- added, all tests go through.
1623 A different order of optimizations might cause problems. */
1624 if (get_opt_normalize())
1625 set_Block_cfgpred(n, i, skip_Tuple(get_Block_cfgpred(n, i)));
1626 } else if (get_opt_optimize() && (get_irn_mode(n) == mode_X)) {
1627 /* We will soon visit a block. Optimize it before visiting! */
1628 ir_node *b = get_nodes_Block(n);
1629 ir_node *new_node = equivalent_node(b);
1630 while (irn_not_visited(b) && (!is_Bad(new_node)) && (new_node != b)) {
1631 /* We would have to run gigo if new is bad, so we
1632 promote it directly below. */
1633 assert(((b == new_node) ||
1634 get_opt_control_flow_straightening() ||
1635 get_opt_control_flow_weak_simplification()) &&
1636 ("strange flag setting"));
1637 exchange (b, new_node);
1639 new_node = equivalent_node(b);
1641 /* GL @@@ get_opt_normalize hinzugefuegt, 5.5.2003 */
1642 if (is_Bad(new_node) && get_opt_normalize()) exchange(n, new_Bad());
1647 * Collects all Phi nodes in link list of Block.
1648 * Marks all blocks "block_visited" if they contain a node other
1651 static void collect_nodes(ir_node *n, void *env) {
1652 if (is_no_Block(n)) {
1653 ir_node *b = get_nodes_Block(n);
1655 if ((get_irn_op(n) == op_Phi)) {
1656 /* Collect Phi nodes to compact ins along with block's ins. */
1657 set_irn_link(n, get_irn_link(b));
1659 } else if (get_irn_op(n) != op_Jmp) { /* Check for non empty block. */
1660 mark_Block_block_visited(b);
1665 /** Returns true if pred is predecessor of block. */
1666 static int is_pred_of(ir_node *pred, ir_node *b) {
1668 for (i = 0; i < get_Block_n_cfgpreds(b); i++) {
1669 ir_node *b_pred = get_nodes_Block(get_Block_cfgpred(b, i));
1670 if (b_pred == pred) return 1;
1675 static int test_whether_dispensable(ir_node *b, int pos) {
1676 int i, j, n_preds = 1;
1677 int dispensable = 1;
1678 ir_node *cfop = get_Block_cfgpred(b, pos);
1679 ir_node *pred = get_nodes_Block(cfop);
1681 if (get_Block_block_visited(pred) + 1
1682 < get_irg_block_visited(current_ir_graph)) {
1683 if (!get_opt_optimize() || !get_opt_control_flow_strong_simplification()) {
1684 /* Mark block so that is will not be removed. */
1685 set_Block_block_visited(pred, get_irg_block_visited(current_ir_graph)-1);
1688 /* Seems to be empty. */
1689 if (!get_irn_link(b)) {
1690 /* There are no Phi nodes ==> dispensable. */
1691 n_preds = get_Block_n_cfgpreds(pred);
1693 /* b's pred blocks and pred's pred blocks must be pairwise disjunct.
1694 Work preds < pos as if they were already removed. */
1695 for (i = 0; i < pos; i++) {
1696 ir_node *b_pred = get_nodes_Block(get_Block_cfgpred(b, i));
1697 if (get_Block_block_visited(b_pred) + 1
1698 < get_irg_block_visited(current_ir_graph)) {
1699 for (j = 0; j < get_Block_n_cfgpreds(b_pred); j++) {
1700 ir_node *b_pred_pred = get_nodes_Block(get_Block_cfgpred(b_pred, j));
1701 if (is_pred_of(b_pred_pred, pred)) dispensable = 0;
1704 if (is_pred_of(b_pred, pred)) dispensable = 0;
1707 for (i = pos +1; i < get_Block_n_cfgpreds(b); i++) {
1708 ir_node *b_pred = get_nodes_Block(get_Block_cfgpred(b, i));
1709 if (is_pred_of(b_pred, pred)) dispensable = 0;
1712 set_Block_block_visited(pred, get_irg_block_visited(current_ir_graph)-1);
1715 n_preds = get_Block_n_cfgpreds(pred);
1723 static void optimize_blocks(ir_node *b, void *env) {
1724 int i, j, k, max_preds, n_preds;
1725 ir_node *pred, *phi;
1728 /* Count the number of predecessor if this block is merged with pred blocks
1731 for (i = 0; i < get_Block_n_cfgpreds(b); i++) {
1732 max_preds += test_whether_dispensable(b, i);
1734 in = (ir_node **) malloc(max_preds * sizeof(ir_node *));
1737 printf(" working on "); DDMN(b);
1738 for (i = 0; i < get_Block_n_cfgpreds(b); i++) {
1739 pred = get_nodes_Block(get_Block_cfgpred(b, i));
1740 if (is_Bad(get_Block_cfgpred(b, i))) {
1741 printf(" removing Bad %i\n ", i);
1742 } else if (get_Block_block_visited(pred) +1
1743 < get_irg_block_visited(current_ir_graph)) {
1744 printf(" removing pred %i ", i); DDMN(pred);
1745 } else { printf(" Nothing to do for "); DDMN(pred); }
1747 * end Debug output -*/
1749 /*- Fix the Phi nodes -*/
1750 phi = get_irn_link(b);
1752 assert(get_irn_op(phi) == op_Phi);
1753 /* Find the new predecessors for the Phi */
1755 for (i = 0; i < get_Block_n_cfgpreds(b); i++) {
1756 pred = get_nodes_Block(get_Block_cfgpred(b, i));
1757 if (is_Bad(get_Block_cfgpred(b, i))) {
1759 } else if (get_Block_block_visited(pred) +1
1760 < get_irg_block_visited(current_ir_graph)) {
1761 /* It's an empty block and not yet visited. */
1762 ir_node *phi_pred = get_Phi_pred(phi, i);
1763 for (j = 0; j < get_Block_n_cfgpreds(pred); j++) {
1764 if (get_nodes_Block(phi_pred) == pred) {
1765 assert(get_irn_op(phi_pred) == op_Phi); /* Block is empty!! */
1766 in[n_preds] = get_Phi_pred(phi_pred, j);
1768 in[n_preds] = phi_pred;
1772 /* The Phi_pred node is replaced now if it is a Phi.
1773 In Schleifen kann offenbar der entfernte Phi Knoten legal verwendet werden.
1774 Daher muss der Phiknoten durch den neuen ersetzt werden.
1775 Weiter muss der alte Phiknoten entfernt werden (durch ersetzen oder
1776 durch einen Bad) damit er aus den keep_alive verschwinden kann.
1777 Man sollte also, falls keine Schleife vorliegt, exchange mit new_Bad
1779 if (get_nodes_Block(phi_pred) == pred) {
1780 /* remove the Phi as it might be kept alive. Further there
1781 might be other users. */
1782 exchange(phi_pred, phi); /* geht, ist aber doch semantisch falsch! Warum?? */
1785 in[n_preds] = get_Phi_pred(phi, i);
1790 set_irn_in(phi, n_preds, in);
1792 phi = get_irn_link(phi);
1796 This happens only if merge between loop backedge and single loop entry. -*/
1797 for (k = 0; k < get_Block_n_cfgpreds(b); k++) {
1798 pred = get_nodes_Block(get_Block_cfgpred(b, k));
1799 if (get_Block_block_visited(pred)+1 < get_irg_block_visited(current_ir_graph)) {
1800 phi = get_irn_link(pred);
1802 if (get_irn_op(phi) == op_Phi) {
1803 set_nodes_Block(phi, b);
1806 for (i = 0; i < k; i++) {
1807 pred = get_nodes_Block(get_Block_cfgpred(b, i));
1808 if (is_Bad(get_Block_cfgpred(b, i))) {
1810 } else if (get_Block_block_visited(pred) +1
1811 < get_irg_block_visited(current_ir_graph)) {
1812 /* It's an empty block and not yet visited. */
1813 for (j = 0; j < get_Block_n_cfgpreds(pred); j++) {
1814 /* @@@ Hier brauche ich Schleifeninformation!!! Kontrollflusskante
1815 muss Rueckwaertskante sein! (An allen vier in[n_preds] = phi
1816 Anweisungen.) Trotzdem tuts bisher!! */
1825 for (i = 0; i < get_Phi_n_preds(phi); i++) {
1826 in[n_preds] = get_Phi_pred(phi, i);
1829 for (i = k+1; i < get_Block_n_cfgpreds(b); i++) {
1830 pred = get_nodes_Block(get_Block_cfgpred(b, i));
1831 if (is_Bad(get_Block_cfgpred(b, i))) {
1833 } else if (get_Block_block_visited(pred) +1
1834 < get_irg_block_visited(current_ir_graph)) {
1835 /* It's an empty block and not yet visited. */
1836 for (j = 0; j < get_Block_n_cfgpreds(pred); j++) {
1845 set_irn_in(phi, n_preds, in);
1847 phi = get_irn_link(phi);
1852 /*- Fix the block -*/
1854 for (i = 0; i < get_Block_n_cfgpreds(b); i++) {
1855 pred = get_nodes_Block(get_Block_cfgpred(b, i));
1856 if (is_Bad(get_Block_cfgpred(b, i))) {
1858 } else if (get_Block_block_visited(pred) +1
1859 < get_irg_block_visited(current_ir_graph)) {
1860 /* It's an empty block and not yet visited. */
1861 assert(get_Block_n_cfgpreds(b) > 1);
1862 /* Else it should be optimized by equivalent_node. */
1863 for (j = 0; j < get_Block_n_cfgpreds(pred); j++) {
1864 in[n_preds] = get_Block_cfgpred(pred, j);
1867 /* Remove block as it might be kept alive. */
1868 exchange(pred, b/*new_Bad()*/);
1870 in[n_preds] = get_Block_cfgpred(b, i);
1874 set_irn_in(b, n_preds, in);
1878 void optimize_cf(ir_graph *irg) {
1881 ir_node *end = get_irg_end(irg);
1882 ir_graph *rem = current_ir_graph;
1883 current_ir_graph = irg;
1885 /* Handle graph state */
1886 assert(get_irg_phase_state(irg) != phase_building);
1887 if (get_irg_outs_state(current_ir_graph) == outs_consistent)
1888 set_irg_outs_inconsistent(current_ir_graph);
1889 if (get_irg_dom_state(current_ir_graph) == dom_consistent)
1890 set_irg_dom_inconsistent(current_ir_graph);
1892 /* Use block visited flag to mark non-empty blocks. */
1893 inc_irg_block_visited(irg);
1894 irg_walk(end, merge_blocks, collect_nodes, NULL);
1896 /* Optimize the standard code. */
1897 irg_block_walk(get_irg_end_block(irg), optimize_blocks, NULL, NULL);
1899 /* Walk all keep alives, optimize them if block, add to new in-array
1900 for end if useful. */
1901 in = NEW_ARR_F (ir_node *, 1);
1902 in[0] = get_nodes_Block(end);
1903 inc_irg_visited(current_ir_graph);
1904 for(i = 0; i < get_End_n_keepalives(end); i++) {
1905 ir_node *ka = get_End_keepalive(end, i);
1906 if (irn_not_visited(ka)) {
1907 if ((get_irn_op(ka) == op_Block) && Block_not_block_visited(ka)) {
1908 set_irg_block_visited(current_ir_graph, /* Don't walk all the way to Start. */
1909 get_irg_block_visited(current_ir_graph)-1);
1910 irg_block_walk(ka, optimize_blocks, NULL, NULL);
1911 mark_irn_visited(ka);
1912 ARR_APP1 (ir_node *, in, ka);
1913 } else if (get_irn_op(ka) == op_Phi) {
1914 mark_irn_visited(ka);
1915 ARR_APP1 (ir_node *, in, ka);
1919 /* DEL_ARR_F(end->in); GL @@@ tut nicht ! */
1922 current_ir_graph = rem;
1927 * Called by walker of remove_critical_cf_edges().
1929 * Place an empty block to an edge between a blocks of multiple
1930 * predecessors and a block of multiple successors.
1933 * @param env Environment of walker. This field is unused and has
1936 static void walk_critical_cf_edges(ir_node *n, void *env) {
1938 ir_node *pre, *block, **in, *jmp;
1940 /* Block has multiple predecessors */
1941 if ((op_Block == get_irn_op(n)) &&
1942 (get_irn_arity(n) > 1)) {
1943 arity = get_irn_arity(n);
1945 if (n == get_irg_end_block(current_ir_graph))
1946 return; /* No use to add a block here. */
1948 for (i=0; i<arity; i++) {
1949 pre = get_irn_n(n, i);
1950 /* Predecessor has multiple successors. Insert new flow edge */
1951 if ((NULL != pre) &&
1952 (op_Proj == get_irn_op(pre)) &&
1953 op_Raise != get_irn_op(skip_Proj(pre))) {
1955 /* set predecessor array for new block */
1956 in = NEW_ARR_D (ir_node *, current_ir_graph->obst, 1);
1957 /* set predecessor of new block */
1959 block = new_Block(1, in);
1960 /* insert new jmp node to new block */
1961 switch_block(block);
1964 /* set successor of new block */
1965 set_irn_n(n, i, jmp);
1967 } /* predecessor has multiple successors */
1968 } /* for all predecessors */
1969 } /* n is a block */
1972 void remove_critical_cf_edges(ir_graph *irg) {
1973 if (get_opt_critical_edges())
1974 irg_walk_graph(irg, NULL, walk_critical_cf_edges, NULL);