opt_inline.c 71.7 KB
Newer Older
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
/*
 * Copyright (C) 1995-2008 University of Karlsruhe.  All right reserved.
 *
 * This file is part of libFirm.
 *
 * This file may be distributed and/or modified under the terms of the
 * GNU General Public License version 2 as published by the Free Software
 * Foundation and appearing in the file LICENSE.GPL included in the
 * packaging of this file.
 *
 * Licensees holding valid libFirm Professional Edition licenses may use
 * this file in accordance with the libFirm Commercial License.
 * Agreement provided with the Software.
 *
 * This file is provided AS IS with NO WARRANTY OF ANY KIND, INCLUDING THE
 * WARRANTY OF DESIGN, MERCHANTABILITY AND FITNESS FOR A PARTICULAR
 * PURPOSE.
 */

/**
 * @file
 * @brief    Dead node elimination and Procedure Inlining.
 * @author   Michael Beck, Goetz Lindenmaier
 * @version  $Id$
 */
Matthias Braun's avatar
Matthias Braun committed
26
#include "config.h"
27

28
#include <limits.h>
29
30
31
32
33
34
35
36
37
38
39
40
41
#include <assert.h>

#include "irnode_t.h"
#include "irgraph_t.h"
#include "irprog_t.h"

#include "iroptimize.h"
#include "ircons_t.h"
#include "iropt_t.h"
#include "irgopt.h"
#include "irgmod.h"
#include "irgwalk.h"

42
43
44
45
46
47
48
#include "array_t.h"
#include "list.h"
#include "pset.h"
#include "pmap.h"
#include "pdeq.h"
#include "xmalloc.h"
#include "pqueue.h"
49
50
51
52

#include "irouts.h"
#include "irloop_t.h"
#include "irbackedge_t.h"
53
#include "opt_init.h"
54
55
56
57
#include "cgana.h"
#include "trouts.h"
#include "error.h"

58
#include "analyze_irg_args.h"
59
60
61
62
#include "iredges_t.h"
#include "irflag_t.h"
#include "irhooks.h"
#include "irtools.h"
63
#include "iropt_dbg.h"
64
#include "irpass_t.h"
65

66
DEBUG_ONLY(static firm_dbg_module_t *dbg;)
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95

/*------------------------------------------------------------------*/
/* Routines for dead node elimination / copying garbage collection  */
/* of the obstack.                                                  */
/*------------------------------------------------------------------*/

/**
 * Remember the new node in the old node by using a field all nodes have.
 */
#define set_new_node(oldn, newn)  set_irn_link(oldn, newn)

/**
 * Get this new node, before the old node is forgotten.
 */
#define get_new_node(oldn) get_irn_link(oldn)

/**
 * Check if a new node was set.
 */
#define has_new_node(n) (get_new_node(n) != NULL)

/**
 * We use the block_visited flag to mark that we have computed the
 * number of useful predecessors for this block.
 * Further we encode the new arity in this flag in the old blocks.
 * Remembering the arity is useful, as it saves a lot of pointer
 * accesses.  This function is called for all Phi and Block nodes
 * in a Block.
 */
96
static inline int
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
compute_new_arity(ir_node *b) {
	int i, res, irn_arity;
	int irg_v, block_v;

	irg_v = get_irg_block_visited(current_ir_graph);
	block_v = get_Block_block_visited(b);
	if (block_v >= irg_v) {
		/* we computed the number of preds for this block and saved it in the
		   block_v flag */
		return block_v - irg_v;
	} else {
		/* compute the number of good predecessors */
		res = irn_arity = get_irn_arity(b);
		for (i = 0; i < irn_arity; i++)
			if (is_Bad(get_irn_n(b, i))) res--;
			/* save it in the flag. */
			set_Block_block_visited(b, irg_v + res);
			return res;
	}
}

/**
 * Copies the node to the new obstack. The Ins of the new node point to
 * the predecessors on the old obstack.  For block/phi nodes not all
 * predecessors might be copied.  n->link points to the new node.
 * For Phi and Block nodes the function allocates in-arrays with an arity
 * only for useful predecessors.  The arity is determined by counting
 * the non-bad predecessors of the block.
 *
 * @param n    The node to be copied
 * @param env  if non-NULL, the node number attribute will be copied to the new node
 *
 * Note: Also used for loop unrolling.
 */
static void copy_node(ir_node *n, void *env) {
	ir_node *nn, *block;
	int new_arity;
	ir_op *op = get_irn_op(n);
	(void) env;

	if (op == op_Bad) {
		/* node copied already */
		return;
	} else if (op == op_Block) {
		block = NULL;
		new_arity = compute_new_arity(n);
		n->attr.block.graph_arr = NULL;
	} else {
		block = get_nodes_block(n);
		if (op == op_Phi) {
			new_arity = compute_new_arity(block);
		} else {
			new_arity = get_irn_arity(n);
		}
	}
	nn = new_ir_node(get_irn_dbg_info(n),
		current_ir_graph,
		block,
		op,
		get_irn_mode(n),
		new_arity,
		get_irn_in(n) + 1);
	/* Copy the attributes.  These might point to additional data.  If this
	   was allocated on the old obstack the pointers now are dangling.  This
	   frees e.g. the memory of the graph_arr allocated in new_immBlock. */
	if (op == op_Block) {
		/* we cannot allow blocks WITHOUT macroblock input */
		set_Block_MacroBlock(nn, get_Block_MacroBlock(n));
	}
	copy_node_attr(n, nn);

168
169
170
	if (env != NULL) {
		/* for easier debugging, we want to copy the node numbers too */
		nn->node_nr = n->node_nr;
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
	}

	set_new_node(n, nn);
	hook_dead_node_elim_subst(current_ir_graph, n, nn);
}

/**
 * Copies new predecessors of old node to new node remembered in link.
 * Spare the Bad predecessors of Phi and Block nodes.
 */
static void copy_preds(ir_node *n, void *env) {
	ir_node *nn, *block;
	int i, j, irn_arity;
	(void) env;

	nn = get_new_node(n);

	if (is_Block(n)) {
		/* copy the macro block header */
		ir_node *mbh = get_Block_MacroBlock(n);

		if (mbh == n) {
			/* this block is a macroblock header */
			set_Block_MacroBlock(nn, nn);
		} else {
			/* get the macro block header */
			ir_node *nmbh = get_new_node(mbh);
			assert(nmbh != NULL);
			set_Block_MacroBlock(nn, nmbh);
		}

		/* Don't copy Bad nodes. */
		j = 0;
		irn_arity = get_irn_arity(n);
		for (i = 0; i < irn_arity; i++) {
			if (! is_Bad(get_irn_n(n, i))) {
				ir_node *pred = get_irn_n(n, i);
				set_irn_n(nn, j, get_new_node(pred));
				j++;
			}
		}
		/* repair the block visited flag from above misuse. Repair it in both
		   graphs so that the old one can still be used. */
		set_Block_block_visited(nn, 0);
		set_Block_block_visited(n, 0);
		/* Local optimization could not merge two subsequent blocks if
		   in array contained Bads.  Now it's possible.
		   We don't call optimize_in_place as it requires
		   that the fields in ir_graph are set properly. */
220
		if (!has_Block_entity(nn) &&
221
222
223
		    get_opt_control_flow_straightening() &&
		    get_Block_n_cfgpreds(nn) == 1 &&
		    is_Jmp(get_Block_cfgpred(nn, 0))) {
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
			ir_node *old = get_nodes_block(get_Block_cfgpred(nn, 0));
			if (nn == old) {
				/* Jmp jumps into the block it is in -- deal self cycle. */
				assert(is_Bad(get_new_node(get_irg_bad(current_ir_graph))));
				exchange(nn, get_new_node(get_irg_bad(current_ir_graph)));
			} else {
				exchange(nn, old);
			}
		}
	} else if (is_Phi(n) && get_irn_arity(n) > 0) {
		/* Don't copy node if corresponding predecessor in block is Bad.
		   The Block itself should not be Bad. */
		block = get_nodes_block(n);
		set_nodes_block(nn, get_new_node(block));
		j = 0;
		irn_arity = get_irn_arity(n);
		for (i = 0; i < irn_arity; i++) {
			if (! is_Bad(get_irn_n(block, i))) {
				ir_node *pred = get_irn_n(n, i);
				set_irn_n(nn, j, get_new_node(pred));
				/*if (is_backedge(n, i)) set_backedge(nn, j);*/
				j++;
			}
		}
		/* If the pre walker reached this Phi after the post walker visited the
		   block block_visited is > 0. */
		set_Block_block_visited(get_nodes_block(n), 0);
		/* Compacting the Phi's ins might generate Phis with only one
		   predecessor. */
		if (get_irn_arity(nn) == 1)
			exchange(nn, get_irn_n(nn, 0));
	} else {
		irn_arity = get_irn_arity(n);
		for (i = -1; i < irn_arity; i++)
			set_irn_n(nn, i, get_new_node(get_irn_n(n, i)));
	}
	/* Now the new node is complete.  We can add it to the hash table for CSE.
	   @@@ inlining aborts if we identify End. Why? */
	if (!is_End(nn))
		add_identities(current_ir_graph->value_table, nn);
}

/**
 * Copies the graph recursively, compacts the keep-alives of the end node.
 *
 * @param irg           the graph to be copied
 * @param copy_node_nr  If non-zero, the node number will be copied
 */
static void copy_graph(ir_graph *irg, int copy_node_nr) {
	ir_node *oe, *ne, *ob, *nb, *om, *nm; /* old end, new end, old bad, new bad, old NoMem, new NoMem */
	ir_node *ka;      /* keep alive */
	int i, irn_arity;
	unsigned long vfl;

	/* Some nodes must be copied by hand, sigh */
	vfl = get_irg_visited(irg);
	set_irg_visited(irg, vfl + 1);

	oe = get_irg_end(irg);
	mark_irn_visited(oe);
	/* copy the end node by hand, allocate dynamic in array! */
	ne = new_ir_node(get_irn_dbg_info(oe),
		irg,
		NULL,
		op_End,
		mode_X,
		-1,
		NULL);
	/* Copy the attributes.  Well, there might be some in the future... */
	copy_node_attr(oe, ne);
	set_new_node(oe, ne);

	/* copy the Bad node */
	ob = get_irg_bad(irg);
	mark_irn_visited(ob);
	nb = new_ir_node(get_irn_dbg_info(ob),
		irg,
		NULL,
		op_Bad,
		mode_T,
		0,
		NULL);
	copy_node_attr(ob, nb);
	set_new_node(ob, nb);

	/* copy the NoMem node */
	om = get_irg_no_mem(irg);
	mark_irn_visited(om);
	nm = new_ir_node(get_irn_dbg_info(om),
		irg,
		NULL,
		op_NoMem,
		mode_M,
		0,
		NULL);
	copy_node_attr(om, nm);
	set_new_node(om, nm);

	/* copy the live nodes */
	set_irg_visited(irg, vfl);
	irg_walk(get_nodes_block(oe), copy_node, copy_preds, INT_TO_PTR(copy_node_nr));

	/* Note: from yet, the visited flag of the graph is equal to vfl + 1 */

	/* visit the anchors as well */
	for (i = get_irg_n_anchors(irg) - 1; i >= 0; --i) {
		ir_node *n = get_irg_anchor(irg, i);

		if (n && (get_irn_visited(n) <= vfl)) {
			set_irg_visited(irg, vfl);
			irg_walk(n, copy_node, copy_preds, INT_TO_PTR(copy_node_nr));
		}
	}

	/* copy_preds for the end node ... */
	set_nodes_block(ne, get_new_node(get_nodes_block(oe)));

	/*- ... and now the keep alives. -*/
	/* First pick the not marked block nodes and walk them.  We must pick these
	   first as else we will oversee blocks reachable from Phis. */
	irn_arity = get_End_n_keepalives(oe);
	for (i = 0; i < irn_arity; i++) {
		ka = get_End_keepalive(oe, i);
		if (is_Block(ka)) {
			if (get_irn_visited(ka) <= vfl) {
				/* We must keep the block alive and copy everything reachable */
				set_irg_visited(irg, vfl);
				irg_walk(ka, copy_node, copy_preds, INT_TO_PTR(copy_node_nr));
			}
			add_End_keepalive(ne, get_new_node(ka));
		}
	}

	/* Now pick other nodes.  Here we will keep all! */
	irn_arity = get_End_n_keepalives(oe);
	for (i = 0; i < irn_arity; i++) {
		ka = get_End_keepalive(oe, i);
		if (!is_Block(ka)) {
			if (get_irn_visited(ka) <= vfl) {
				/* We didn't copy the node yet.  */
				set_irg_visited(irg, vfl);
				irg_walk(ka, copy_node, copy_preds, INT_TO_PTR(copy_node_nr));
			}
			add_End_keepalive(ne, get_new_node(ka));
		}
	}

	/* start block sometimes only reached after keep alives */
	set_nodes_block(nb, get_new_node(get_nodes_block(ob)));
	set_nodes_block(nm, get_new_node(get_nodes_block(om)));
}

/**
 * Copies the graph reachable from current_ir_graph->end to the obstack
 * in current_ir_graph and fixes the environment.
 * Then fixes the fields in current_ir_graph containing nodes of the
 * graph.
 *
 * @param copy_node_nr  If non-zero, the node number will be copied
 */
static void
copy_graph_env(int copy_node_nr) {
	ir_graph *irg = current_ir_graph;
	ir_node *old_end, *new_anchor;
	int i;

	/* remove end_except and end_reg nodes */
	old_end = get_irg_end(irg);
	set_irg_end_except (irg, old_end);
	set_irg_end_reg    (irg, old_end);

	/* Not all nodes remembered in irg might be reachable
	   from the end node.  Assure their link is set to NULL, so that
	   we can test whether new nodes have been computed. */
	for (i = get_irg_n_anchors(irg) - 1; i >= 0; --i) {
		ir_node *n = get_irg_anchor(irg, i);
		if (n != NULL)
			set_new_node(n, NULL);
	}
	/* we use the block walk flag for removing Bads from Blocks ins. */
	inc_irg_block_visited(irg);

	/* copy the graph */
	copy_graph(irg, copy_node_nr);

	/* fix the anchor */
	old_end    = get_irg_end(irg);
	new_anchor = new_Anchor(irg);

	for (i = get_irg_n_anchors(irg) - 1; i >= 0; --i) {
		ir_node *n = get_irg_anchor(irg, i);
		if (n)
			set_irn_n(new_anchor, i, get_new_node(n));
	}
	free_End(old_end);
	irg->anchor = new_anchor;

	/* ensure the new anchor is placed in the endblock */
	set_nodes_block(new_anchor, get_irg_end_block(irg));
}

/**
 * Copies all reachable nodes to a new obstack.  Removes bad inputs
 * from block nodes and the corresponding inputs from Phi nodes.
 * Merges single exit blocks with single entry blocks and removes
 * 1-input Phis.
 * Adds all new nodes to a new hash table for CSE.  Does not
 * perform CSE, so the hash table might contain common subexpressions.
 */
void dead_node_elimination(ir_graph *irg) {
	ir_graph *rem;
#ifdef INTERPROCEDURAL_VIEW
	int rem_ipview = get_interprocedural_view();
#endif
	struct obstack *graveyard_obst = NULL;
	struct obstack *rebirth_obst   = NULL;
440
441

	edges_deactivate(irg);
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469

	/* inform statistics that we started a dead-node elimination run */
	hook_dead_node_elim(irg, 1);

	/* Remember external state of current_ir_graph. */
	rem = current_ir_graph;
	current_ir_graph = irg;
#ifdef INTERPROCEDURAL_VIEW
	set_interprocedural_view(0);
#endif

	assert(get_irg_phase_state(irg) != phase_building);

	/* Handle graph state */
	free_callee_info(irg);
	free_irg_outs(irg);
	free_trouts();

	/* @@@ so far we loose loops when copying */
	free_loop_information(irg);

	set_irg_doms_inconsistent(irg);

	/* A quiet place, where the old obstack can rest in peace,
	   until it will be cremated. */
	graveyard_obst = irg->obst;

	/* A new obstack, where the reachable nodes will be copied to. */
470
	rebirth_obst = XMALLOC(struct obstack);
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
	irg->obst = rebirth_obst;
	obstack_init(irg->obst);
	irg->last_node_idx = 0;

	/* We also need a new value table for CSE */
	del_identities(irg->value_table);
	irg->value_table = new_identities();

	/* Copy the graph from the old to the new obstack */
	copy_graph_env(/*copy_node_nr=*/1);

	/* Free memory from old unoptimized obstack */
	obstack_free(graveyard_obst, 0);  /* First empty the obstack ... */
	xfree(graveyard_obst);            /* ... then free it.           */

	/* inform statistics that the run is over */
	hook_dead_node_elim(irg, 0);

	current_ir_graph = rem;
#ifdef INTERPROCEDURAL_VIEW
	set_interprocedural_view(rem_ipview);
#endif
}

Michael Beck's avatar
Michael Beck committed
495
496
497
498
ir_graph_pass_t *dead_node_elimination_pass(const char *name) {
	return def_graph_pass(name ? name : "dce", dead_node_elimination);
}

499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
/**
 * Relink bad predecessors of a block and store the old in array to the
 * link field. This function is called by relink_bad_predecessors().
 * The array of link field starts with the block operand at position 0.
 * If block has bad predecessors, create a new in array without bad preds.
 * Otherwise let in array untouched.
 */
static void relink_bad_block_predecessors(ir_node *n, void *env) {
	ir_node **new_in, *irn;
	int i, new_irn_n, old_irn_arity, new_irn_arity = 0;
	(void) env;

	/* if link field of block is NULL, look for bad predecessors otherwise
	   this is already done */
	if (is_Block(n) && get_irn_link(n) == NULL) {
		/* save old predecessors in link field (position 0 is the block operand)*/
		set_irn_link(n, get_irn_in(n));

		/* count predecessors without bad nodes */
		old_irn_arity = get_irn_arity(n);
		for (i = 0; i < old_irn_arity; i++)
			if (!is_Bad(get_irn_n(n, i)))
				++new_irn_arity;

		/* arity changing: set new predecessors without bad nodes */
		if (new_irn_arity < old_irn_arity) {
			/* Get new predecessor array. We do not resize the array, as we must
			   keep the old one to update Phis. */
			new_in = NEW_ARR_D(ir_node *, current_ir_graph->obst, (new_irn_arity+1));

			/* set new predecessors in array */
			new_in[0] = NULL;
			new_irn_n = 1;
			for (i = 0; i < old_irn_arity; i++) {
				irn = get_irn_n(n, i);
				if (!is_Bad(irn)) {
					new_in[new_irn_n] = irn;
					is_backedge(n, i) ? set_backedge(n, new_irn_n-1) : set_not_backedge(n, new_irn_n-1);
					++new_irn_n;
				}
			}
			/* ARR_SETLEN(int, n->attr.block.backedge, new_irn_arity); */
			ARR_SHRINKLEN(n->attr.block.backedge, new_irn_arity);
			n->in = new_in;
		} /* ir node has bad predecessors */
	} /* Block is not relinked */
}

/**
 * Relinks Bad predecessors from Blocks and Phis called by walker
 * remove_bad_predecesors(). If n is a Block, call
 * relink_bad_block_redecessors(). If n is a Phi-node, call also the relinking
 * function of Phi's Block. If this block has bad predecessors, relink preds
 * of the Phi-node.
 */
static void relink_bad_predecessors(ir_node *n, void *env) {
	ir_node *block, **old_in;
	int i, old_irn_arity, new_irn_arity;

	/* relink bad predecessors of a block */
	if (is_Block(n))
		relink_bad_block_predecessors(n, env);

	/* If Phi node relink its block and its predecessors */
	if (is_Phi(n)) {
		/* Relink predecessors of phi's block */
		block = get_nodes_block(n);
		if (get_irn_link(block) == NULL)
			relink_bad_block_predecessors(block, env);

		old_in = (ir_node **)get_irn_link(block); /* Of Phi's Block */
		old_irn_arity = ARR_LEN(old_in);

		/* Relink Phi predecessors if count of predecessors changed */
		if (old_irn_arity != ARR_LEN(get_irn_in(block))) {
			/* set new predecessors in array
			   n->in[0] remains the same block */
			new_irn_arity = 1;
			for(i = 1; i < old_irn_arity; i++)
				if (!is_Bad(old_in[i])) {
					n->in[new_irn_arity] = n->in[i];
					is_backedge(n, i) ? set_backedge(n, new_irn_arity) : set_not_backedge(n, new_irn_arity);
					++new_irn_arity;
				}

				ARR_SETLEN(ir_node *, n->in, new_irn_arity);
				ARR_SETLEN(int, n->attr.phi.u.backedge, new_irn_arity);
		}
	} /* n is a Phi node */
}

/*
 * Removes Bad Bad predecessors from Blocks and the corresponding
 * inputs to Phi nodes as in dead_node_elimination but without
 * copying the graph.
 * On walking up set the link field to NULL, on walking down call
 * relink_bad_predecessors() (This function stores the old in array
 * to the link field and sets a new in array if arity of predecessors
 * changes).
 */
void remove_bad_predecessors(ir_graph *irg) {
	panic("Fix backedge handling first");
	irg_walk_graph(irg, firm_clear_link, relink_bad_predecessors, NULL);
}


/*
   __                      _  __ __
  (_     __    o     _    | \/  |_
  __)|_| | \_/ | \_/(/_   |_/\__|__

  The following stuff implements a facility that automatically patches
  registered ir_node pointers to the new node when a dead node elimination occurs.
*/

struct _survive_dce_t {
	struct obstack obst;
	pmap *places;
	pmap *new_places;
	hook_entry_t dead_node_elim;
	hook_entry_t dead_node_elim_subst;
};

typedef struct _survive_dce_list_t {
	struct _survive_dce_list_t *next;
	ir_node **place;
} survive_dce_list_t;

static void dead_node_hook(void *context, ir_graph *irg, int start) {
	survive_dce_t *sd = context;
	(void) irg;

	/* Create a new map before the dead node elimination is performed. */
	if (start) {
		sd->new_places = pmap_create_ex(pmap_count(sd->places));
	} else {
		/* Patch back all nodes if dead node elimination is over and something is to be done. */
		pmap_destroy(sd->places);
		sd->places     = sd->new_places;
		sd->new_places = NULL;
	}
}

/**
 * Hook called when dead node elimination replaces old by nw.
 */
static void dead_node_subst_hook(void *context, ir_graph *irg, ir_node *old, ir_node *nw) {
	survive_dce_t *sd = context;
	survive_dce_list_t *list = pmap_get(sd->places, old);
	(void) irg;

	/* If the node is to be patched back, write the new address to all registered locations. */
	if (list) {
		survive_dce_list_t *p;

		for (p = list; p; p = p->next)
			*(p->place) = nw;

		pmap_insert(sd->new_places, nw, list);
	}
}

/**
 * Make a new Survive DCE environment.
 */
survive_dce_t *new_survive_dce(void) {
665
	survive_dce_t *res = XMALLOC(survive_dce_t);
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
	obstack_init(&res->obst);
	res->places     = pmap_create();
	res->new_places = NULL;

	res->dead_node_elim.hook._hook_dead_node_elim = dead_node_hook;
	res->dead_node_elim.context                   = res;
	res->dead_node_elim.next                      = NULL;

	res->dead_node_elim_subst.hook._hook_dead_node_elim_subst = dead_node_subst_hook;
	res->dead_node_elim_subst.context = res;
	res->dead_node_elim_subst.next    = NULL;

	register_hook(hook_dead_node_elim, &res->dead_node_elim);
	register_hook(hook_dead_node_elim_subst, &res->dead_node_elim_subst);
	return res;
}

/**
 * Free a Survive DCE environment.
 */
void free_survive_dce(survive_dce_t *sd) {
	obstack_free(&sd->obst, NULL);
	pmap_destroy(sd->places);
	unregister_hook(hook_dead_node_elim, &sd->dead_node_elim);
	unregister_hook(hook_dead_node_elim_subst, &sd->dead_node_elim_subst);
	xfree(sd);
}

/**
 * Register a node pointer to be patched upon DCE.
 * When DCE occurs, the node pointer specified by @p place will be
 * patched to the new address of the node it is pointing to.
 *
 * @param sd    The Survive DCE environment.
 * @param place The address of the node pointer.
 */
void survive_dce_register_irn(survive_dce_t *sd, ir_node **place) {
	if (*place != NULL) {
		ir_node *irn      = *place;
		survive_dce_list_t *curr = pmap_get(sd->places, irn);
706
		survive_dce_list_t *nw   = OALLOC(&sd->obst, survive_dce_list_t);
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727

		nw->next  = curr;
		nw->place = place;

		pmap_insert(sd->places, irn, nw);
	}
}

/*--------------------------------------------------------------------*/
/*  Functionality for inlining                                         */
/*--------------------------------------------------------------------*/

/**
 * Copy node for inlineing.  Updates attributes that change when
 * inlineing but not for dead node elimination.
 *
 * Copies the node by calling copy_node() and then updates the entity if
 * it's a local one.  env must be a pointer of the frame type of the
 * inlined procedure. The new entities must be in the link field of
 * the entities.
 */
728
static void copy_node_inline(ir_node *n, void *env) {
729
730
731
732
733
	ir_node *nn;
	ir_type *frame_tp = (ir_type *)env;

	copy_node(n, NULL);
	if (is_Sel(n)) {
734
		nn = get_new_node(n);
735
		assert(is_Sel(nn));
736
		/* use copied entities from the new frame */
737
738
739
740
		if (get_entity_owner(get_Sel_entity(n)) == frame_tp) {
			set_Sel_entity(nn, get_entity_link(get_Sel_entity(n)));
		}
	} else if (is_Block(n)) {
741
742
		nn = get_new_node(n);
		nn->attr.block.irg.irg = current_ir_graph;
743
744
745
	}
}

746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
/**
 * Copies new predecessors of old node and move constants to
 * the Start Block.
 */
static void copy_preds_inline(ir_node *n, void *env) {
	ir_node *nn;

	copy_preds(n, env);
	nn = skip_Id(get_new_node(n));
	if (is_irn_constlike(nn)) {
		/* move Constants into the start block */
		set_nodes_block(nn, get_irg_start_block(current_ir_graph));

		n = identify_remember(current_ir_graph->value_table, nn);
		if (nn != n) {
761
			DBG_OPT_CSE(nn, n);
762
763
764
765
766
			exchange(nn, n);
		}
	}
}

767
768
769
770
771
/**
 * Walker: checks if P_value_arg_base is used.
 */
static void find_addr(ir_node *node, void *env) {
	int *allow_inline = env;
772
773
774
775
776
777
778
779
780
781
	if (is_Sel(node)) {
		ir_graph *irg = current_ir_graph;
		if (get_Sel_ptr(node) == get_irg_frame(irg)) {
			/* access to frame */
			ir_entity *ent = get_Sel_entity(node);
			if (get_entity_owner(ent) != get_irg_frame_type(irg)) {
				/* access to value_type */
				*allow_inline = 0;
			}
		}
782
783
784
785
786
787
788
	} else if (is_Alloc(node) && get_Alloc_where(node) == stack_alloc) {
		/* From GCC:
		 * Refuse to inline alloca call unless user explicitly forced so as this
		 * may change program's memory overhead drastically when the function
		 * using alloca is called in loop.  In GCC present in SPEC2000 inlining
		 * into schedule_block cause it to require 2GB of ram instead of 256MB.
		 *
Michael Beck's avatar
Michael Beck committed
789
		 * Sorrily this is true with our implementation also.
790
791
792
793
794
795
796
797
		 * Moreover, we cannot differentiate between alloca() and VLA yet, so this
		 * disables inlining of functions using VLA (with are completely save).
		 *
		 * 2 Solutions:
		 * - add a flag to the Alloc node for "real" alloca() calls
		 * - add a new Stack-Restore node at the end of a function using alloca()
		 */
		*allow_inline = 0;
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
	}
}

/**
 * Check if we can inline a given call.
 * Currently, we cannot inline two cases:
 * - call with compound arguments
 * - graphs that take the address of a parameter
 *
 * check these conditions here
 */
static int can_inline(ir_node *call, ir_graph *called_graph) {
	ir_type *call_type = get_Call_type(call);
	int params, ress, i, res;
	assert(is_Method_type(call_type));

	params = get_method_n_params(call_type);
	ress   = get_method_n_ress(call_type);

	/* check parameters for compound arguments */
	for (i = 0; i < params; ++i) {
		ir_type *p_type = get_method_param_type(call_type, i);

		if (is_compound_type(p_type))
			return 0;
	}

	/* check results for compound arguments */
	for (i = 0; i < ress; ++i) {
		ir_type *r_type = get_method_res_type(call_type, i);

		if (is_compound_type(r_type))
			return 0;
	}

	res = 1;
	irg_walk_graph(called_graph, find_addr, NULL, &res);

	return res;
}

enum exc_mode {
840
841
	exc_handler,    /**< There is a handler. */
	exc_no_handler  /**< Exception handling not represented. */
842
843
844
845
};

/* Inlines a method at the given call site. */
int inline_method(ir_node *call, ir_graph *called_graph) {
846
847
848
	ir_node             *pre_call;
	ir_node             *post_call, *post_bl;
	ir_node             *in[pn_Start_max];
Michael Beck's avatar
Michael Beck committed
849
	ir_node             *end, *end_bl, *block;
850
851
	ir_node             **res_pred;
	ir_node             **cf_pred;
Michael Beck's avatar
Michael Beck committed
852
	ir_node             **args_in;
853
	ir_node             *ret, *phi;
Michael Beck's avatar
Michael Beck committed
854
	int                 arity, n_ret, n_exc, n_res, i, n, j, rem_opt, irn_arity, n_params;
855
	int                 n_mem_phi;
856
	enum exc_mode       exc_handling;
Michael Beck's avatar
Michael Beck committed
857
	ir_type             *called_frame, *curr_frame, *mtp, *ctp;
858
859
	ir_entity           *ent;
	ir_graph            *rem, *irg;
860
	irg_inline_property prop = get_irg_inline_property(called_graph);
861
	unsigned long       visited;
862
863
864
865
866
867

	if (prop == irg_inline_forbidden)
		return 0;

	ent = get_irg_entity(called_graph);

Michael Beck's avatar
Michael Beck committed
868
869
	mtp = get_entity_type(ent);
	ctp = get_Call_type(call);
870
871
872
	n_params = get_method_n_params(mtp);
	n_res    = get_method_n_ress(mtp);
	if (n_params > get_method_n_params(ctp)) {
Michael Beck's avatar
Michael Beck committed
873
874
875
		/* this is a bad feature of C: without a prototype, we can
		 * call a function with less parameters than needed. Currently
		 * we don't support this, although we could use Unknown than. */
Michael Beck's avatar
Michael Beck committed
876
877
		return 0;
	}
878
879
880
	if (n_res != get_method_n_ress(ctp)) {
		return 0;
	}
881

Michael Beck's avatar
Michael Beck committed
882
	/* Argh, compiling C has some bad consequences:
Michael Beck's avatar
Michael Beck committed
883
884
885
	 * It is implementation dependent what happens in that case.
	 * We support inlining, if the bitsize of the types matches AND
	 * the same arithmetic is used. */
Michael Beck's avatar
Michael Beck committed
886
887
888
889
890
891
892
893
894
895
896
	for (i = n_params - 1; i >= 0; --i) {
		ir_type *param_tp = get_method_param_type(mtp, i);
		ir_type *arg_tp   = get_method_param_type(ctp, i);

		if (param_tp != arg_tp) {
			ir_mode *pmode = get_type_mode(param_tp);
			ir_mode *amode = get_type_mode(arg_tp);

			if (pmode == NULL || amode == NULL)
				return 0;
			if (get_mode_size_bits(pmode) != get_mode_size_bits(amode))
897
				return 0;
Michael Beck's avatar
Michael Beck committed
898
899
900
			if (get_mode_arithmetic(pmode) != get_mode_arithmetic(amode))
				return 0;
			/* otherwise we can simply "reinterpret" the bits */
901
		}
902
	}
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
	for (i = n_res - 1; i >= 0; --i) {
		ir_type *decl_res_tp = get_method_res_type(mtp, i);
		ir_type *used_res_tp = get_method_res_type(ctp, i);

		if (decl_res_tp != used_res_tp) {
			ir_mode *decl_mode = get_type_mode(decl_res_tp);
			ir_mode *used_mode = get_type_mode(used_res_tp);
			if (decl_mode == NULL || used_mode == NULL)
				return 0;
			if (get_mode_size_bits(decl_mode) != get_mode_size_bits(used_mode))
				return 0;
			if (get_mode_arithmetic(decl_mode) != get_mode_arithmetic(used_mode))
				return 0;
			/* otherwise we can "reinterpret" the bits */
		}
	}
919

920
921
	irg = get_irn_irg(call);

922
923
924
925
	/*
	 * We cannot inline a recursive call. The graph must be copied before
	 * the call the inline_method() using create_irg_copy().
	 */
926
	if (called_graph == irg)
927
928
929
930
931
932
933
934
935
936
		return 0;

	/*
	 * currently, we cannot inline two cases:
	 * - call with compound arguments
	 * - graphs that take the address of a parameter
	 */
	if (! can_inline(call, called_graph))
		return 0;

937
938
939
	rem = current_ir_graph;
	current_ir_graph = irg;

940
	DB((dbg, LEVEL_1, "Inlining %+F(%+F) into %+F\n", call, called_graph, irg));
941

942
943
944
945
946
	/* --  Turn off optimizations, this can cause problems when allocating new nodes. -- */
	rem_opt = get_opt_optimize();
	set_optimize(0);

	/* Handle graph state */
947
948
	assert(get_irg_phase_state(irg) != phase_building);
	assert(get_irg_pinned(irg) == op_pin_state_pinned);
949
	assert(get_irg_pinned(called_graph) == op_pin_state_pinned);
950
951
952
953
954
	set_irg_outs_inconsistent(irg);
	set_irg_extblk_inconsistent(irg);
	set_irg_doms_inconsistent(irg);
	set_irg_loopinfo_inconsistent(irg);
	set_irg_callee_info_state(irg, irg_callee_info_inconsistent);
955
	set_irg_entity_usage_state(irg, ir_entity_usage_not_computed);
956
957
958
959
960
961
962
963
964
965
966
967
968

	/* -- Check preconditions -- */
	assert(is_Call(call));

	/* here we know we WILL inline, so inform the statistics */
	hook_inline(call, called_graph);

	/* -- Decide how to handle exception control flow: Is there a handler
	   for the Call node, or do we branch directly to End on an exception?
	   exc_handling:
	   0 There is a handler.
	   2 Exception handling not represented in Firm. -- */
	{
969
970
		ir_node *Xproj = NULL;
		ir_node *proj;
971
972
973
974
		for (proj = get_irn_link(call); proj; proj = get_irn_link(proj)) {
			long proj_nr = get_Proj_proj(proj);
			if (proj_nr == pn_Call_X_except) Xproj = proj;
		}
975
		exc_handling = Xproj != NULL ? exc_handler : exc_no_handler;
976
977
	}

Michael Beck's avatar
Michael Beck committed
978
979
980
981
982
983
984
985
986
987
	/* create the argument tuple */
	NEW_ARR_A(ir_type *, args_in, n_params);

	block = get_nodes_block(call);
	for (i = n_params - 1; i >= 0; --i) {
		ir_node *arg      = get_Call_param(call, i);
		ir_type *param_tp = get_method_param_type(mtp, i);
		ir_mode *mode     = get_type_mode(param_tp);

		if (mode != get_irn_mode(arg)) {
988
			arg = new_r_Conv(block, arg, mode);
Michael Beck's avatar
Michael Beck committed
989
990
991
992
		}
		args_in[i] = arg;
	}

993
994
995
996
997
	/* --
	   the procedure and later replaces the Start node of the called graph.
	   Post_call is the old Call node and collects the results of the called
	   graph. Both will end up being a tuple.  -- */
	post_bl = get_nodes_block(call);
998
	set_irg_current_block(irg, post_bl);
999
1000
	/* XxMxPxPxPxT of Start + parameter of Call */
	in[pn_Start_X_initial_exec]   = new_Jmp();