ldstopt.c 33.6 KB
Newer Older
Michael Beck's avatar
Michael Beck committed
1
2
3
4
/*
 * Project:     libFIRM
 * File name:   ir/opt/ldstopt.c
 * Purpose:     load store optimizations
Michael Beck's avatar
Michael Beck committed
5
 * Author:      Michael Beck
Michael Beck's avatar
Michael Beck committed
6
7
 * Created:
 * CVS-ID:      $Id$
8
 * Copyright:   (c) 1998-2007 Universitt Karlsruhe
Michael Beck's avatar
Michael Beck committed
9
10
 * Licence:     This file protected by GPL -  GNU GENERAL PUBLIC LICENSE.
 */
Michael Beck's avatar
Michael Beck committed
11
12
13
14
15
16
17
18
#ifdef HAVE_CONFIG_H
# include "config.h"
#endif

#ifdef HAVE_STRING_H
# include <string.h>
#endif

19
20
21
22
23
24
25
26
27
28
29
30
31
32
#include "irnode_t.h"
#include "irgraph_t.h"
#include "irmode_t.h"
#include "iropt_t.h"
#include "ircons_t.h"
#include "irgmod.h"
#include "irgwalk.h"
#include "irvrfy.h"
#include "tv_t.h"
#include "dbginfo_t.h"
#include "iropt_dbg.h"
#include "irflag_t.h"
#include "array.h"
#include "irhooks.h"
Michael Beck's avatar
BugFix:    
Michael Beck committed
33
#include "iredges.h"
34
35
#include "irtools.h"
#include "opt_polymorphy.h"
36
#include "irmemory.h"
37
#include "xmalloc.h"
38
39
40
41

#ifdef DO_CACHEOPT
#include "cacheopt/cachesim.h"
#endif
Michael Beck's avatar
Michael Beck committed
42
43
44
45
46
47

#undef IMAX
#define IMAX(a,b)	((a) > (b) ? (a) : (b))

#define MAX_PROJ	IMAX(pn_Load_max, pn_Store_max)

48
enum changes_t {
49
50
	DF_CHANGED = 1,       /**< data flow changed */
	CF_CHANGED = 2,       /**< control flow changed */
51
52
};

Michael Beck's avatar
Michael Beck committed
53
54
55
56
/**
 * walker environment
 */
typedef struct _walk_env_t {
57
58
	struct obstack obst;          /**< list of all stores */
	unsigned changes;             /**< a bitmask of graph changes */
Michael Beck's avatar
Michael Beck committed
59
60
} walk_env_t;

61
62
63
64
/**
 * flags for Load/Store
 */
enum ldst_flags_t {
65
	LDST_VISITED = 1              /**< if set, this Load/Store is already visited */
66
67
};

68
/** A Load/Store info. */
Michael Beck's avatar
Michael Beck committed
69
typedef struct _ldst_info_t {
70
71
72
73
74
	ir_node  *projs[MAX_PROJ];    /**< list of Proj's of this node */
	ir_node  *exc_block;          /**< the exception block if available */
	int      exc_idx;             /**< predecessor index in the exception block */
	unsigned flags;               /**< flags */
	unsigned visited;             /**< visited counter for breaking loops */
Michael Beck's avatar
Michael Beck committed
75
76
} ldst_info_t;

77
/**
78
 * flags for control flow.
79
80
 */
enum block_flags_t {
81
82
	BLOCK_HAS_COND = 1,      /**< Block has conditional control flow */
	BLOCK_HAS_EXC  = 2       /**< Block has exceptional control flow */
83
84
85
};

/**
86
 * a Block info.
87
88
 */
typedef struct _block_info_t {
89
	unsigned flags;               /**< flags for the block */
90
91
} block_info_t;

92
93
94
95
96
97
98
/** the master visited flag for loop detection. */
static unsigned master_visited = 0;

#define INC_MASTER()       ++master_visited
#define MARK_NODE(info)    (info)->visited = master_visited
#define NODE_VISITED(info) (info)->visited >= master_visited

Michael Beck's avatar
Michael Beck committed
99
100
101
/**
 * get the Load/Store info of a node
 */
102
static ldst_info_t *get_ldst_info(ir_node *node, walk_env_t *env) {
103
	ldst_info_t *info = get_irn_link(node);
Michael Beck's avatar
Michael Beck committed
104

105
106
107
108
109
110
111
	if (! info) {
		info = obstack_alloc(&env->obst, sizeof(*info));
		memset(info, 0, sizeof(*info));
		set_irn_link(node, info);
	}
	return info;
}  /* get_ldst_info */
Michael Beck's avatar
Michael Beck committed
112

113
114
115
/**
 * get the Block info of a node
 */
116
117
static block_info_t *get_block_info(ir_node *node, walk_env_t *env) {
	block_info_t *info = get_irn_link(node);
118

119
120
121
122
123
124
125
	if (! info) {
		info = obstack_alloc(&env->obst, sizeof(*info));
		memset(info, 0, sizeof(*info));
		set_irn_link(node, info);
	}
	return info;
}  /* get_block_info */
126

Michael Beck's avatar
Michael Beck committed
127
/**
Michael Beck's avatar
Michael Beck committed
128
 * update the projection info for a Load/Store
Michael Beck's avatar
Michael Beck committed
129
 */
130
static unsigned update_projs(ldst_info_t *info, ir_node *proj)
Michael Beck's avatar
Michael Beck committed
131
{
132
	long nr = get_Proj_proj(proj);
Michael Beck's avatar
Michael Beck committed
133

134
	assert(0 <= nr && nr <= MAX_PROJ && "Wrong proj from LoadStore");
Michael Beck's avatar
Michael Beck committed
135

136
137
138
139
140
141
142
143
144
145
	if (info->projs[nr]) {
		/* there is already one, do CSE */
		exchange(proj, info->projs[nr]);
		return DF_CHANGED;
	}
	else {
		info->projs[nr] = proj;
		return 0;
	}
}  /* update_projs */
Michael Beck's avatar
Michael Beck committed
146
147

/**
148
149
150
151
152
 * update the exception block info for a Load/Store node.
 *
 * @param info   the load/store info struct
 * @param block  the exception handler block for this load/store
 * @param pos    the control flow input of the block
Michael Beck's avatar
Michael Beck committed
153
 */
154
static unsigned update_exc(ldst_info_t *info, ir_node *block, int pos)
Michael Beck's avatar
Michael Beck committed
155
{
156
	assert(info->exc_block == NULL && "more than one exception block found");
Michael Beck's avatar
Michael Beck committed
157

158
159
160
161
	info->exc_block = block;
	info->exc_idx   = pos;
	return 0;
}  /* update_exc */
Michael Beck's avatar
Michael Beck committed
162

163
/** Return the number of uses of an address node */
Michael Beck's avatar
BugFix:    
Michael Beck committed
164
#define get_irn_n_uses(adr)     get_irn_n_edges(adr)
165

Michael Beck's avatar
Michael Beck committed
166
167
/**
 * walker, collects all Load/Store/Proj nodes
168
 *
169
 * walks from Start -> End
Michael Beck's avatar
Michael Beck committed
170
 */
Michael Beck's avatar
Michael Beck committed
171
static void collect_nodes(ir_node *node, void *env)
Michael Beck's avatar
Michael Beck committed
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
220
221
222
223
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
	ir_op       *op = get_irn_op(node);
	ir_node     *pred, *blk, *pred_blk;
	ldst_info_t *ldst_info;
	walk_env_t  *wenv = env;

	if (op == op_Proj) {
		ir_node *adr;
		ir_op *op;

		pred = get_Proj_pred(node);
		op   = get_irn_op(pred);

		if (op == op_Load) {
			ldst_info = get_ldst_info(pred, wenv);

			wenv->changes |= update_projs(ldst_info, node);

			if ((ldst_info->flags & LDST_VISITED) == 0) {
				adr = get_Load_ptr(pred);
				ldst_info->flags |= LDST_VISITED;
			}

			/*
			* Place the Proj's to the same block as the
			* predecessor Load. This is always ok and prevents
			* "non-SSA" form after optimizations if the Proj
			* is in a wrong block.
			*/
			blk      = get_nodes_block(node);
			pred_blk = get_nodes_block(pred);
			if (blk != pred_blk) {
				wenv->changes |= DF_CHANGED;
				set_nodes_block(node, pred_blk);
			}
		} else if (op == op_Store) {
			ldst_info = get_ldst_info(pred, wenv);

			wenv->changes |= update_projs(ldst_info, node);

			if ((ldst_info->flags & LDST_VISITED) == 0) {
				adr = get_Store_ptr(pred);
				ldst_info->flags |= LDST_VISITED;
			}

			/*
			* Place the Proj's to the same block as the
			* predecessor Store. This is always ok and prevents
			* "non-SSA" form after optimizations if the Proj
			* is in a wrong block.
			*/
			blk      = get_nodes_block(node);
			pred_blk = get_nodes_block(pred);
			if (blk != pred_blk) {
				wenv->changes |= DF_CHANGED;
				set_nodes_block(node, pred_blk);
			}
		}
	} else if (op == op_Block) {
		int i;

		for (i = get_Block_n_cfgpreds(node) - 1; i >= 0; --i) {
			ir_node      *pred_block;
			block_info_t *bl_info;

			pred = skip_Proj(get_Block_cfgpred(node, i));

			/* ignore Bad predecessors, they will be removed later */
			if (is_Bad(pred))
				continue;

			pred_block = get_nodes_block(pred);
			bl_info    = get_block_info(pred_block, wenv);

			if (is_fragile_op(pred))
				bl_info->flags |= BLOCK_HAS_EXC;
			else if (is_irn_forking(pred))
				bl_info->flags |= BLOCK_HAS_COND;

			if (get_irn_op(pred) == op_Load || get_irn_op(pred) == op_Store) {
				ldst_info = get_ldst_info(pred, wenv);

				wenv->changes |= update_exc(ldst_info, node, i);
			}
		}
	}
}  /* collect_nodes */
Michael Beck's avatar
Michael Beck committed
259

Michael Beck's avatar
Michael Beck committed
260
/**
261
 * Returns an entity if the address ptr points to a constant one.
262
263
264
265
 *
 * @param ptr  the address
 *
 * @return an entity or NULL
Michael Beck's avatar
Michael Beck committed
266
 */
267
static ir_entity *find_constant_entity(ir_node *ptr)
Michael Beck's avatar
Michael Beck committed
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
	for (;;) {
		ir_op *op = get_irn_op(ptr);

		if (op == op_SymConst && (get_SymConst_kind(ptr) == symconst_addr_ent)) {
			ir_entity *ent = get_SymConst_entity(ptr);
			if (variability_constant == get_entity_variability(ent))
				return ent;
			return NULL;
		} else if (op == op_Sel) {
			ir_entity *ent = get_Sel_entity(ptr);
			ir_type   *tp  = get_entity_owner(ent);

			/* Do not fiddle with polymorphism. */
			if (is_Class_type(get_entity_owner(ent)) &&
				((get_entity_n_overwrites(ent)    != 0) ||
				(get_entity_n_overwrittenby(ent) != 0)   ) )
				return NULL;

			if (is_Array_type(tp)) {
				/* check bounds */
				int i, n;

				for (i = 0, n = get_Sel_n_indexs(ptr); i < n; ++i) {
					ir_node *bound;
					tarval *tlower, *tupper;
					ir_node *index = get_Sel_index(ptr, i);
					tarval *tv     = computed_value(index);

					/* check if the index is constant */
					if (tv == tarval_bad)
						return NULL;

					bound  = get_array_lower_bound(tp, i);
					tlower = computed_value(bound);
					bound  = get_array_upper_bound(tp, i);
					tupper = computed_value(bound);

					if (tlower == tarval_bad || tupper == tarval_bad)
						return NULL;

					if (tarval_cmp(tv, tlower) & pn_Cmp_Lt)
						return NULL;
					if (tarval_cmp(tupper, tv) & pn_Cmp_Lt)
						return NULL;

					/* ok, bounds check finished */
				}
			}

			if (variability_constant == get_entity_variability(ent))
				return ent;

			/* try next */
			ptr = get_Sel_ptr(ptr);
		} else
			return NULL;
	}
}  /* find_constant_entity */
Michael Beck's avatar
Michael Beck committed
327

Michael Beck's avatar
Michael Beck committed
328
329
330
/**
 * Return the Selection index of a Sel node from dimension n
 */
331
static long get_Sel_array_index_long(ir_node *n, int dim) {
332
333
334
335
	ir_node *index = get_Sel_index(n, dim);
	assert(get_irn_op(index) == op_Const);
	return get_tarval_long(get_Const_tarval(index));
}  /* get_Sel_array_index_long */
336

Michael Beck's avatar
Michael Beck committed
337
338
339
340
341
342
343
344
/**
 * Returns the accessed component graph path for an
 * node computing an address.
 *
 * @param ptr    the node computing the address
 * @param depth  current depth in steps upward from the root
 *               of the address
 */
Götz Lindenmaier's avatar
bugfix    
Götz Lindenmaier committed
345
static compound_graph_path *rec_get_accessed_path(ir_node *ptr, int depth) {
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
	compound_graph_path *res = NULL;
	ir_entity           *root, *field;
	int                 path_len, pos;

	if (get_irn_op(ptr) == op_SymConst) {
		/* a SymConst. If the depth is 0, this is an access to a global
		 * entity and we don't need a component path, else we know
		 * at least it's length.
		 */
		assert(get_SymConst_kind(ptr) == symconst_addr_ent);
		root = get_SymConst_entity(ptr);
		res = (depth == 0) ? NULL : new_compound_graph_path(get_entity_type(root), depth);
	} else {
		assert(get_irn_op(ptr) == op_Sel);
		/* it's a Sel, go up until we find the root */
		res = rec_get_accessed_path(get_Sel_ptr(ptr), depth+1);

		/* fill up the step in the path at the current position */
		field    = get_Sel_entity(ptr);
		path_len = get_compound_graph_path_length(res);
		pos      = path_len - depth - 1;
		set_compound_graph_path_node(res, pos, field);

		if (is_Array_type(get_entity_owner(field))) {
			assert(get_Sel_n_indexs(ptr) == 1 && "multi dim arrays not implemented");
			set_compound_graph_path_array_index(res, pos, get_Sel_array_index_long(ptr, 0));
		}
	}
	return res;
}  /* rec_get_accessed_path */
376

Götz Lindenmaier's avatar
bugfix    
Götz Lindenmaier committed
377
/** Returns an access path or NULL.  The access path is only
Michael Beck's avatar
Michael Beck committed
378
379
 *  valid, if the graph is in phase_high and _no_ address computation is used.
 */
Götz Lindenmaier's avatar
bugfix    
Götz Lindenmaier committed
380
static compound_graph_path *get_accessed_path(ir_node *ptr) {
381
382
	return rec_get_accessed_path(ptr, 0);
}  /* get_accessed_path */
383

384
385
386
387
388
389
390
/* forward */
static void reduce_adr_usage(ir_node *ptr);

/**
 * Update a Load that may lost it's usage.
 */
static void handle_load_update(ir_node *load) {
391
	ldst_info_t *info = get_irn_link(load);
392

393
394
395
	/* do NOT touch volatile loads for now */
	if (get_Load_volatility(load) == volatility_is_volatile)
		return;
396

397
398
399
	if (! info->projs[pn_Load_res] && ! info->projs[pn_Load_X_except]) {
		ir_node *ptr = get_Load_ptr(load);
		ir_node *mem = get_Load_mem(load);
400

401
402
403
404
405
406
		/* a Load which value is neither used nor exception checked, remove it */
		exchange(info->projs[pn_Load_M], mem);
		exchange(load, new_Bad());
		reduce_adr_usage(ptr);
	}
}  /* handle_load_update */
407
408
409
410
411
412

/**
 * A Use of an address node is vanished. Check if this was a Proj
 * node and update the counters.
 */
static void reduce_adr_usage(ir_node *ptr) {
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
	if (is_Proj(ptr)) {
		if (get_irn_n_edges(ptr) <= 0) {
			/* this Proj is dead now */
			ir_node *pred = get_Proj_pred(ptr);

			if (is_Load(pred)) {
				ldst_info_t *info = get_irn_link(pred);
				info->projs[get_Proj_proj(ptr)] = NULL;

				/* this node lost it's result proj, handle that */
				handle_load_update(pred);
			}
		}
	}
}  /* reduce_adr_usage */
428

429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
/**
 * Check, if an already existing value of mode old_mode can be converted
 * into the needed one new_mode without loss.
 */
static int can_use_stored_value(ir_mode *old_mode, ir_mode *new_mode) {
	if (old_mode == new_mode)
		return 1;

	/* if both modes are two-complement ones, we can always convert the
	   Stored value into the needed one. */
	if (get_mode_size_bits(old_mode) >= get_mode_size_bits(new_mode) &&
		  get_mode_arithmetic(old_mode) == irma_twos_complement &&
		  get_mode_arithmetic(new_mode) == irma_twos_complement)
		return 1;
	return 0;
}  /* can_use_stored_value */

Michael Beck's avatar
Michael Beck committed
446
447
/**
 * Follow the memory chain as long as there are only Loads
448
449
 * and alias free Stores and try to replace current Load or Store
 * by a previous ones.
Michael Beck's avatar
Michael Beck committed
450
451
452
453
454
455
 * Note that in unreachable loops it might happen that we reach
 * load again, as well as we can fall into a cycle.
 * We break such cycles using a special visited flag.
 *
 * INC_MASTER() must be called before dive into
 */
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
static unsigned follow_Mem_chain(ir_node *load, ir_node *curr) {
	unsigned res = 0;
	ldst_info_t *info = get_irn_link(load);
	ir_node *pred;
	ir_node *ptr       = get_Load_ptr(load);
	ir_node *mem       = get_Load_mem(load);
	ir_mode *load_mode = get_Load_mode(load);

	for (pred = curr; load != pred; ) {
		ldst_info_t *pred_info = get_irn_link(pred);

		/*
		 * BEWARE: one might think that checking the modes is useless, because
		 * if the pointers are identical, they refer to the same object.
		 * This is only true in strong typed languages, not in C were the following
		 * is possible a = *(ir_type1 *)p; b = *(ir_type2 *)p ...
		 */
		if (get_irn_op(pred) == op_Store && get_Store_ptr(pred) == ptr &&
474
		    can_use_stored_value(get_irn_mode(get_Store_value(pred)), load_mode)) {
475
476
477
478
479
480
481
482
483
484
485
486
			/*
			 * a Load immediately after a Store -- a read after write.
			 * We may remove the Load, if both Load & Store does not have an exception handler
			 * OR they are in the same block. In the latter case the Load cannot
			 * throw an exception when the previous Store was quiet.
			 *
			 * Why we need to check for Store Exception? If the Store cannot
			 * be executed (ROM) the exception handler might simply jump into
			 * the load block :-(
			 * We could make it a little bit better if we would know that the exception
			 * handler of the Store jumps directly to the end...
			 */
487
			if ((pred_info->projs[pn_Store_X_except] == NULL && info->projs[pn_Load_X_except] == NULL) ||
488
489
490
491
			    get_nodes_block(load) == get_nodes_block(pred)) {
				ir_node *value = get_Store_value(pred);

				DBG_OPT_RAW(load, value);
492
493
494
495
496
497

				/* add an convert if needed */
				if (get_irn_mode(get_Store_value(pred)) != load_mode) {
					value = new_r_Conv(current_ir_graph, get_nodes_block(load), value, load_mode);
				}

498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
				if (info->projs[pn_Load_M])
					exchange(info->projs[pn_Load_M], mem);

				/* no exception */
				if (info->projs[pn_Load_X_except]) {
					exchange( info->projs[pn_Load_X_except], new_Bad());
					res |= CF_CHANGED;
				}

				if (info->projs[pn_Load_res])
					exchange(info->projs[pn_Load_res], value);

				exchange(load, new_Bad());
				reduce_adr_usage(ptr);
				return res | DF_CHANGED;
			}
		} else if (get_irn_op(pred) == op_Load && get_Load_ptr(pred) == ptr &&
515
		           can_use_stored_value(get_Load_mode(pred), load_mode)) {
516
517
518
519
520
521
522
523
524
			/*
			 * a Load after a Load -- a read after read.
			 * We may remove the second Load, if it does not have an exception handler
			 * OR they are in the same block. In the later case the Load cannot
			 * throw an exception when the previous Load was quiet.
			 *
			 * Here, there is no need to check if the previous Load has an exception
			 * hander because they would have exact the same exception...
			 */
525
526
527
			if (info->projs[pn_Load_X_except] == NULL || get_nodes_block(load) == get_nodes_block(pred)) {
				ir_node *value;

528
529
				DBG_OPT_RAR(load, pred);

530
531
532
533
534
535
536
				/* the result is used */
				if (info->projs[pn_Load_res]) {
					if (pred_info->projs[pn_Load_res] == NULL) {
						/* create a new Proj again */
						pred_info->projs[pn_Load_res] = new_r_Proj(current_ir_graph, get_nodes_block(pred), pred, get_Load_mode(pred), pn_Load_res);
					}
					value = pred_info->projs[pn_Load_res];
537
538
539
540
541
542

					/* add an convert if needed */
					if (get_Load_mode(pred) != load_mode) {
						value = new_r_Conv(current_ir_graph, get_nodes_block(load), value, load_mode);
					}

543
					exchange(info->projs[pn_Load_res], value);
544
545
				}

546
547
548
				if (info->projs[pn_Load_M])
					exchange(info->projs[pn_Load_M], mem);

549
550
551
552
553
554
555
556
557
558
559
560
561
				/* no exception */
				if (info->projs[pn_Load_X_except]) {
					exchange(info->projs[pn_Load_X_except], new_Bad());
					res |= CF_CHANGED;
				}

				exchange(load, new_Bad());
				reduce_adr_usage(ptr);
				return res |= DF_CHANGED;
			}
		}

		if (get_irn_op(pred) == op_Store) {
562
			/* check if we can pass through this store */
563
564
565
566
			ir_alias_relation rel = get_alias_relation(
				current_ir_graph,
				get_Store_ptr(pred),
				get_irn_mode(get_Store_value(pred)),
567
				ptr, load_mode);
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
			/* if the might be an alias, we cannot pass this Store */
			if (rel != no_alias)
				break;
			pred = skip_Proj(get_Store_mem(pred));
		} else if (get_irn_op(pred) == op_Load) {
			pred = skip_Proj(get_Load_mem(pred));
		} else {
			/* follow only Load chains */
			break;
		}

		/* check for cycles */
		if (NODE_VISITED(pred_info))
			break;
		MARK_NODE(pred_info);
	}

	if (get_irn_op(pred) == op_Sync) {
		int i;

		/* handle all Sync predecessors */
		for (i = get_Sync_n_preds(pred) - 1; i >= 0; --i) {
			res |= follow_Mem_chain(load, skip_Proj(get_Sync_pred(pred, i)));
			if (res)
				break;
		}
	}

	return res;
}  /* follow_Mem_chain */
Michael Beck's avatar
Michael Beck committed
598

Michael Beck's avatar
Michael Beck committed
599
600
/**
 * optimize a Load
601
602
 *
 * @param load  the Load node
Michael Beck's avatar
Michael Beck committed
603
 */
604
static unsigned optimize_load(ir_node *load)
Michael Beck's avatar
Michael Beck committed
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
665
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
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
	ldst_info_t *info = get_irn_link(load);
	ir_node *mem, *ptr, *new_node;
	ir_entity *ent;
	unsigned res = 0;

	/* do NOT touch volatile loads for now */
	if (get_Load_volatility(load) == volatility_is_volatile)
		return 0;

	/* the address of the load to be optimized */
	ptr = get_Load_ptr(load);

	/*
	 * Check if we can remove the exception from a Load:
	 * This can be done, if the address is from an Sel(Alloc) and
	 * the Sel type is a subtype of the allocated type.
	 *
	 * This optimizes some often used OO constructs,
	 * like x = new O; x->t;
	 */
	if (info->projs[pn_Load_X_except]) {
		if (is_Sel(ptr)) {
			ir_node *mem = get_Sel_mem(ptr);

			/* FIXME: works with the current FE, but better use the base */
			if (get_irn_op(skip_Proj(mem)) == op_Alloc) {
				/* ok, check the types */
				ir_entity *ent    = get_Sel_entity(ptr);
				ir_type   *s_type = get_entity_type(ent);
				ir_type   *a_type = get_Alloc_type(mem);

				if (is_SubClass_of(s_type, a_type)) {
					/* ok, condition met: there can't be an exception because
					* Alloc guarantees that enough memory was allocated */

					exchange(info->projs[pn_Load_X_except], new_Bad());
					info->projs[pn_Load_X_except] = NULL;
					res |= CF_CHANGED;
				}
			}
		} else if ((get_irn_op(skip_Proj(ptr)) == op_Alloc) ||
			((get_irn_op(ptr) == op_Cast) && (get_irn_op(skip_Proj(get_Cast_op(ptr))) == op_Alloc))) {
				/* simple case: a direct load after an Alloc. Firm Alloc throw
				 * an exception in case of out-of-memory. So, there is no way for an
				 * exception in this load.
				 * This code is constructed by the "exception lowering" in the Jack compiler.
				 */
				exchange(info->projs[pn_Load_X_except], new_Bad());
				info->projs[pn_Load_X_except] = NULL;
				res |= CF_CHANGED;
		}
	}

	/* The mem of the Load. Must still be returned after optimization. */
	mem  = get_Load_mem(load);

	if (! info->projs[pn_Load_res] && ! info->projs[pn_Load_X_except]) {
		/* a Load which value is neither used nor exception checked, remove it */
		exchange(info->projs[pn_Load_M], mem);

		exchange(load, new_Bad());
		reduce_adr_usage(ptr);
		return res | DF_CHANGED;
	}

	/* Load from a constant polymorphic field, where we can resolve
	   polymorphism. */
	new_node = transform_node_Load(load);
	if (new_node != load) {
		if (info->projs[pn_Load_M]) {
			exchange(info->projs[pn_Load_M], mem);
			info->projs[pn_Load_M] = NULL;
		}
		if (info->projs[pn_Load_X_except]) {
			exchange(info->projs[pn_Load_X_except], new_Bad());
			info->projs[pn_Load_X_except] = NULL;
		}
		if (info->projs[pn_Load_res])
			exchange(info->projs[pn_Load_res], new_node);

		exchange(load, new_Bad());
		reduce_adr_usage(ptr);
		return res | DF_CHANGED;
	}

	/* check if we can determine the entity that will be loaded */
	ent = find_constant_entity(ptr);
	if (ent) {
		if ((allocation_static == get_entity_allocation(ent)) &&
			(visibility_external_allocated != get_entity_visibility(ent))) {
			/* a static allocation that is not external: there should be NO exception
			 * when loading. */

			/* no exception, clear the info field as it might be checked later again */
			if (info->projs[pn_Load_X_except]) {
				exchange(info->projs[pn_Load_X_except], new_Bad());
				info->projs[pn_Load_X_except] = NULL;
				res |= CF_CHANGED;
			}

			if (variability_constant == get_entity_variability(ent)
				&& is_atomic_entity(ent)) {
				/* Might not be atomic after
				   lowering of Sels.  In this
				   case we could also load, but
				   it's more complicated. */
				/* more simpler case: we load the content of a constant value:
				 * replace it by the constant itself
				 */

				/* no memory */
				if (info->projs[pn_Load_M]) {
					exchange(info->projs[pn_Load_M], mem);
					res |= DF_CHANGED;
				}
				/* no result :-) */
				if (info->projs[pn_Load_res]) {
					if (is_atomic_entity(ent)) {
						ir_node *c = copy_const_value(get_irn_dbg_info(load), get_atomic_ent_value(ent));

						DBG_OPT_RC(load, c);
						exchange(info->projs[pn_Load_res], c);
						res |= DF_CHANGED;
					}
				}
				exchange(load, new_Bad());
				reduce_adr_usage(ptr);
				return res;
			} else if (variability_constant == get_entity_variability(ent)) {
				compound_graph_path *path = get_accessed_path(ptr);

				if (path) {
					ir_node *c;

					assert(is_proper_compound_graph_path(path, get_compound_graph_path_length(path)-1));
					/*
					{
						int j;
						for (j = 0; j < get_compound_graph_path_length(path); ++j) {
							ir_entity *node = get_compound_graph_path_node(path, j);
							fprintf(stdout, ".%s", get_entity_name(node));
							if (is_Array_type(get_entity_owner(node)))
								fprintf(stdout, "[%d]", get_compound_graph_path_array_index(path, j));
						}
						printf("\n");
					}
					*/

					c = get_compound_ent_value_by_path(ent, path);
					free_compound_graph_path(path);

					/* printf("  cons: "); DDMN(c); */

					if (info->projs[pn_Load_M]) {
						exchange(info->projs[pn_Load_M], mem);
						res |= DF_CHANGED;
					}
					if (info->projs[pn_Load_res]) {
						exchange(info->projs[pn_Load_res], copy_const_value(get_irn_dbg_info(load), c));
						res |= DF_CHANGED;
					}
					exchange(load, new_Bad());
					reduce_adr_usage(ptr);
					return res;
				} else {
					/*  We can not determine a correct access path.  E.g., in jack, we load
					a byte from an object to generate an exception.   Happens in test program
					Reflectiontest.
					printf(">>>>>>>>>>>>> Found access to constant entity %s in function %s\n", get_entity_name(ent),
					get_entity_name(get_irg_entity(current_ir_graph)));
					printf("  load: "); DDMN(load);
					printf("  ptr:  "); DDMN(ptr);
					*/
				}
			}
		}
	}

	/* Check, if the address of this load is used more than once.
	 * If not, this load cannot be removed in any case. */
	if (get_irn_n_uses(ptr) <= 1)
		return res;

	/*
	 * follow the memory chain as long as there are only Loads
	 * and try to replace current Load or Store by a previous one.
	 * Note that in unreachable loops it might happen that we reach
	 * load again, as well as we can fall into a cycle.
	 * We break such cycles using a special visited flag.
	 */
	INC_MASTER();
	res = follow_Mem_chain(load, skip_Proj(mem));
	return res;
}  /* optimize_load */
Michael Beck's avatar
Michael Beck committed
800

801
802
803
804
805
806
807
808
809
/**
 * Check whether a value of mode new_mode would completely overwrite a value
 * of mode old_mode in memory.
 */
static int is_completely_overwritten(ir_mode *old_mode, ir_mode *new_mode)
{
	return get_mode_size_bits(new_mode) >= get_mode_size_bits(old_mode);
}  /* is_completely_overwritten */

Michael Beck's avatar
Michael Beck committed
810
/**
811
 * follow the memory chain as long as there are only Loads and alias free Stores.
Michael Beck's avatar
Michael Beck committed
812
813
814
 *
 * INC_MASTER() must be called before dive into
 */
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
static unsigned follow_Mem_chain_for_Store(ir_node *store, ir_node *curr) {
	unsigned res = 0;
	ldst_info_t *info = get_irn_link(store);
	ir_node *pred;
	ir_node *ptr = get_Store_ptr(store);
	ir_node *mem = get_Store_mem(store);
	ir_node *value = get_Store_value(store);
	ir_mode *mode  = get_irn_mode(value);
	ir_node *block = get_nodes_block(store);

	for (pred = curr; pred != store;) {
		ldst_info_t *pred_info = get_irn_link(pred);

		/*
		 * BEWARE: one might think that checking the modes is useless, because
		 * if the pointers are identical, they refer to the same object.
		 * This is only true in strong typed languages, not is C were the following
		 * is possible *(ir_type1 *)p = a; *(ir_type2 *)p = b ...
833
834
		 * However, if the mode that is written have a bigger  or equal size the the old
		 * one, the old value is completely overwritten and can be killed ...
835
836
		 */
		if (get_irn_op(pred) == op_Store && get_Store_ptr(pred) == ptr &&
837
838
		    get_nodes_block(pred) == block &&
		    is_completely_overwritten(get_irn_mode(get_Store_value(pred)), mode)) {
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
			/*
			 * a Store after a Store in the same block -- a write after write.
			 * We may remove the first Store, if it does not have an exception handler.
			 *
			 * TODO: What, if both have the same exception handler ???
			 */
			if (get_Store_volatility(pred) != volatility_is_volatile && !pred_info->projs[pn_Store_X_except]) {
				DBG_OPT_WAW(pred, store);
				exchange( pred_info->projs[pn_Store_M], get_Store_mem(pred) );
				exchange(pred, new_Bad());
				reduce_adr_usage(ptr);
				return DF_CHANGED;
			}
		} else if (get_irn_op(pred) == op_Load && get_Load_ptr(pred) == ptr &&
		           value == pred_info->projs[pn_Load_res]) {
			/*
			 * a Store of a value after a Load -- a write after read.
			 * We may remove the second Store, if it does not have an exception handler.
			 */
			if (! info->projs[pn_Store_X_except]) {
				DBG_OPT_WAR(store, pred);
				exchange( info->projs[pn_Store_M], mem );
				exchange(store, new_Bad());
				reduce_adr_usage(ptr);
				return DF_CHANGED;
			}
		}

		if (get_irn_op(pred) == op_Store) {
			/* check if we can pass thru this store */
			ir_alias_relation rel = get_alias_relation(
				current_ir_graph,
				get_Store_ptr(pred),
				get_irn_mode(get_Store_value(pred)),
873
				ptr, mode);
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
			/* if the might be an alias, we cannot pass this Store */
			if (rel != no_alias)
				break;
			pred = skip_Proj(get_Store_mem(pred));
		} else if (get_irn_op(pred) == op_Load) {
			pred = skip_Proj(get_Load_mem(pred));
		} else {
			/* follow only Load chains */
			break;
		}

		/* check for cycles */
		if (NODE_VISITED(pred_info))
			break;
		MARK_NODE(pred_info);
	}

	if (get_irn_op(pred) == op_Sync) {
		int i;

		/* handle all Sync predecessors */
		for (i = get_Sync_n_preds(pred) - 1; i >= 0; --i) {
			res |= follow_Mem_chain_for_Store(store, skip_Proj(get_Sync_pred(pred, i)));
			if (res)
				break;
		}
	}
	return res;
}  /* follow_Mem_chain_for_Store */
Michael Beck's avatar
Michael Beck committed
903
904
905

/**
 * optimize a Store
906
907
 *
 * @param store  the Store node
Michael Beck's avatar
Michael Beck committed
908
 */
909
910
static unsigned optimize_store(ir_node *store) {
	ir_node *ptr, *mem;
Michael Beck's avatar
Michael Beck committed
911

912
913
	if (get_Store_volatility(store) == volatility_is_volatile)
		return 0;
Michael Beck's avatar
Michael Beck committed
914

915
	ptr = get_Store_ptr(store);
916

917
918
919
920
	/* Check, if the address of this Store is used more than once.
	 * If not, this Store cannot be removed in any case. */
	if (get_irn_n_uses(ptr) <= 1)
		return 0;
921

922
	mem = get_Store_mem(store);
Michael Beck's avatar
Michael Beck committed
923

924
925
926
927
	/* follow the memory chain as long as there are only Loads */
	INC_MASTER();
	return follow_Mem_chain_for_Store(store, skip_Proj(mem));
}  /* optimize_store */
Michael Beck's avatar
Michael Beck committed
928

Michael Beck's avatar
Michael Beck committed
929
/**
930
 * walker, optimizes Phi after Stores to identical places:
Michael Beck's avatar
Michael Beck committed
931
 * Does the following optimization:
932
 * @verbatim
Michael Beck's avatar
Michael Beck committed
933
934
935
 *
 *   val1   val2   val3          val1  val2  val3
 *    |      |      |               \    |    /
936
 *  Store  Store  Store              \   |   /
937
 *      \    |    /                   PhiData
Michael Beck's avatar
Michael Beck committed
938
 *       \   |   /                       |
939
 *        \  |  /                      Store
940
 *          PhiM
Michael Beck's avatar
Michael Beck committed
941
 *
942
 * @endverbatim
943
944
 * This reduces the number of stores and allows for predicated execution.
 * Moves Stores back to the end of a function which may be bad.
Michael Beck's avatar
Michael Beck committed
945
 *
946
 * This is only possible if the predecessor blocks have only one successor.
Michael Beck's avatar
Michael Beck committed
947
 */
948
static unsigned optimize_phi(ir_node *phi, walk_env_t *wenv)
Michael Beck's avatar
Michael Beck committed
949
{
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
	int i, n;
	ir_node *store, *old_store, *ptr, *block, *phi_block, *phiM, *phiD, *exc, *projM;
	ir_mode *mode;
	ir_node **inM, **inD, **stores;
	int *idx;
	dbg_info *db = NULL;
	ldst_info_t *info;
	block_info_t *bl_info;
	unsigned res = 0;

	/* Must be a memory Phi */
	if (get_irn_mode(phi) != mode_M)
		return 0;

	n = get_Phi_n_preds(phi);
	if (n <= 0)
		return 0;

	store = skip_Proj(get_Phi_pred(phi, 0));
	old_store = store;
	if (get_irn_op(store) != op_Store)
		return 0;

	block = get_nodes_block(store);

	/* abort on dead blocks */
	if (is_Block_dead(block))
		return 0;

	/* check if the block is post dominated by Phi-block
	   and has no exception exit */
	bl_info = get_irn_link(block);
	if (bl_info->flags & BLOCK_HAS_EXC)
		return 0;

	phi_block = get_nodes_block(phi);
	if (! block_postdominates(phi_block, block))
		return 0;

	/* this is the address of the store */
	ptr  = get_Store_ptr(store);
	mode = get_irn_mode(get_Store_value(store));
	info = get_irn_link(store);
	exc  = info->exc_block;

	for (i = 1; i < n; ++i) {
		ir_node *pred = skip_Proj(get_Phi_pred(phi, i));

		if (get_irn_op(pred) != op_Store)
			return 0;

		if (ptr != get_Store_ptr(pred) || mode != get_irn_mode(get_Store_value(pred)))
			return 0;

		info = get_irn_link(pred);

		/* check, if all stores have the same exception flow */
		if (exc != info->exc_block)
			return 0;

		/* abort on dead blocks */
		block = get_nodes_block(pred);
		if (is_Block_dead(block))
			return 0;

		/* check if the block is post dominated by Phi-block
		   and has no exception exit. Note that block must be different from
		   Phi-block, else we would move a Store from end End of a block to its
		   Start... */
		bl_info = get_irn_link(block);
		if (bl_info->flags & BLOCK_HAS_EXC)
			return 0;
		if (block == phi_block || ! block_postdominates(phi_block, block))
			return 0;
	}

	/*
	 * ok, when we are here, we found all predecessors of a Phi that
	 * are Stores to the same address and size. That means whatever
	 * we do before we enter the block of the Phi, we do a Store.
	 * So, we can move the Store to the current block:
	 *
	 *   val1    val2    val3          val1  val2  val3
	 *    |       |       |               \    |    /
	 * | Str | | Str | | Str |             \   |   /
	 *      \     |     /                   PhiData
	 *       \    |    /                       |
	 *        \   |   /                       Str
	 *           PhiM
	 *
	 * Is only allowed if the predecessor blocks have only one successor.
	 */

	NEW_ARR_A(ir_node *, stores, n);
	NEW_ARR_A(ir_node *, inM, n);
	NEW_ARR_A(ir_node *, inD, n);
	NEW_ARR_A(int, idx, n);

	/* Prepare: Collect all Store nodes.  We must do this
	   first because we otherwise may loose a store when exchanging its
	   memory Proj.
	 */
	for (i = 0; i < n; ++i)
		stores[i] = skip_Proj(get_Phi_pred(phi, i));

	/* Prepare: Skip the memory Proj: we need this in the case some stores
	   are cascaded.
	   Beware: One Store might be included more than once in the stores[]
	   list, so we must prevent to do the exchange more than once.
	 */
	for (i = 0; i < n; ++i) {
		ir_node *store = stores[i];
		ir_node *proj_m;

		info = get_irn_link(store);
		proj_m = info->projs[pn_Store_M];

		if (is_Proj(proj_m) && get_Proj_pred(proj_m) == store)
			exchange(proj_m, get_Store_mem(store));
	}

	/* first step: collect all inputs */
	for (i = 0; i < n; ++i) {
		ir_node *store = stores[i];
		info = get_irn_link(store);

		inM[i] = get_Store_mem(store);
		inD[i] = get_Store_value(store);
		idx[i] = info->exc_idx;
	}
	block = get_nodes_block(phi);

	/* second step: create a new memory Phi */
	phiM = new_rd_Phi(get_irn_dbg_info(phi), current_ir_graph, block, n, inM, mode_M);

	/* third step: create a new data Phi */
	phiD = new_rd_Phi(get_irn_dbg_info(phi), current_ir_graph, block, n, inD, mode);

	/* fourth step: create the Store */
	store = new_rd_Store(db, current_ir_graph, block, phiM, ptr, phiD);
1090
#ifdef DO_CACHEOPT
1091
	co_set_irn_name(store, co_get_irn_ident(old_store));
1092
1093
#endif

1094
	projM = new_rd_Proj(NULL, current_ir_graph, block, store, mode_M, pn_Store_M);
1095

1096
1097
	info = get_ldst_info(store, wenv);
	info->projs[pn_Store_M] = projM;
Michael Beck's avatar
Michael Beck committed
1098

1099
1100
1101
	/* fifths step: repair exception flow */
	if (exc) {
		ir_node *projX = new_rd_Proj(NULL, current_ir_graph, block, store, mode_X, pn_Store_X_except);
Michael Beck's avatar
Michael Beck committed
1102

1103
1104
1105
		info->projs[pn_Store_X_except] = projX;
		info->exc_block                = exc;
		info->exc_idx                  = idx[0];
1106

1107
1108
1109
		for (i = 0; i < n; ++i) {
			set_Block_cfgpred(exc, idx[i], projX);
		}
Michael Beck's avatar
Michael Beck committed
1110

1111
1112
1113
		if (n > 1) {
			/* the exception block should be optimized as some inputs are identical now */
		}
1114

1115
1116
		res |= CF_CHANGED;
	}
Michael Beck's avatar
Michael Beck committed
1117

1118
1119
	/* sixth step: replace old Phi */
	exchange(phi, projM);
Michael Beck's avatar
Michael Beck committed
1120

1121
1122
	return res | DF_CHANGED;
}  /* optimize_phi */
Michael Beck's avatar
Michael Beck committed
1123

Michael Beck's avatar
Michael Beck committed
1124
/**
Michael Beck's avatar
Michael Beck committed
1125
 * walker, do the optimizations
Michael Beck's avatar
Michael Beck committed
1126
 */
1127
1128
static void do_load_store_optimize(ir_node *n, void *env) {
	walk_env_t *wenv = env;
Michael Beck's avatar
Michael Beck committed
1129

1130
	switch (get_irn_opcode(n)) {
Michael Beck's avatar
Michael Beck committed
1131

1132
1133
1134
	case iro_Load:
		wenv->changes |= optimize_load(n);
		break;
Michael Beck's avatar
Michael Beck committed
1135

1136
1137
1138
	case iro_Store:
		wenv->changes |= optimize_store(n);
		break;
Michael Beck's avatar
Michael Beck committed
1139

1140
1141
	case iro_Phi:
		wenv->changes |= optimize_phi(n, wenv);
Michael Beck's avatar
Michael Beck committed
1142

1143
1144
	default:
		;
1145
1146
	}
}  /* do_load_store_optimize */
Michael Beck's avatar
Michael Beck committed
1147
1148
1149
1150

/*
 * do the load store optimization
 */
1151
1152
1153
1154
1155
1156
void optimize_load_store(ir_graph *irg) {
	walk_env_t env;

	assert(get_irg_phase_state(irg) != phase_building);
	assert(get_irg_pinned(irg) != op_pin_state_floats &&
		"LoadStore optimization needs pinned graph");
Michael Beck's avatar
Michael Beck committed
1157

1158
1159
	if (! get_opt_redundant_loadstore())
		return;
Michael Beck's avatar
Michael Beck committed
1160

1161
	edges_assure(irg);
Michael Beck's avatar
Michael Beck committed
1162

1163
1164
	/* for Phi optimization post-dominators are needed ... */
	assure_postdoms(irg);
Michael Beck's avatar
BugFix:    
Michael Beck committed
1165

1166
1167
1168
1169
	if (get_opt_alias_analysis()) {
		assure_irg_address_taken_computed(irg);
		assure_irp_globals_address_taken_computed();
	}
1170

1171
1172
	obstack_init(&env.obst);
	env.changes = 0;
Michael Beck's avatar
Michael Beck committed
1173

1174
1175
1176
	/* init the links, then collect Loads/Stores/Proj's in lists */
	master_visited = 0;
	irg_walk_graph(irg, firm_clear_link, collect_nodes, &env);
Michael Beck's avatar
Michael Beck committed
1177

1178
1179
	/* now we have collected enough information, optimize */
	irg_walk_graph(irg, NULL, do_load_store_optimize, &env);
Michael Beck's avatar
Michael Beck committed
1180

1181
	obstack_free(&env.obst, NULL);
Michael Beck's avatar
Michael Beck committed
1182

1183
1184
1185
1186
1187
	/* Handle graph state */
	if (env.changes) {
		if (get_irg_outs_state(irg) == outs_consistent)
			set_irg_outs_inconsistent(irg);
	}
Michael Beck's avatar
Michael Beck committed
1188

1189
1190
1191
1192
1193
1194
	if (env.changes & CF_CHANGED) {
		/* is this really needed: Yes, control flow changed, block might
		have Bad() predecessors. */
		set_irg_doms_inconsistent(irg);
	}
}  /* optimize_load_store */