betranshlp.c 24.8 KB
Newer Older
1
2
/*
 * This file is part of libFirm.
3
 * Copyright (C) 2012 University of Karlsruhe.
4
5
6
7
8
 */

/**
 * @file
 * @brief       be transform helper extracted from the ia32 backend.
Michael Beck's avatar
Michael Beck committed
9
 * @author      Matthias Braun, Michael Beck
10
11
 * @date        14.06.2007
 */
Matthias Braun's avatar
Matthias Braun committed
12
#include "bearch.h"
13
14
#include "beirg.h"
#include "belive.h"
15
#include "benode.h"
16
#include "betranshlp.h"
17
#include "beutil.h"
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
#include "cgana.h"
#include "debug.h"
#include "execfreq_t.h"
#include "heights.h"
#include "ircons_t.h"
#include "iredges.h"
#include "irgmod.h"
#include "irgraph_t.h"
#include "irgwalk.h"
#include "irhooks.h"
#include "irnodemap.h"
#include "irnode_t.h"
#include "irop_t.h"
#include "iropt_t.h"
#include "irouts.h"
#include "irtools.h"
34
#include "panic.h"
35
#include "pdeq.h"
36
#include "util.h"
37
38
39
40
41
42
43
44

typedef struct be_transform_env_t {
	waitq    *worklist;    /**< worklist of nodes that still need to be
	                            transformed */
} be_transform_env_t;

static be_transform_env_t env;

45
46
void be_set_transformed_node(ir_node *old_node, ir_node *new_node)
{
47
	set_irn_link(old_node, new_node);
48
	mark_irn_visited(old_node);
49
50
}

Matthias Braun's avatar
Matthias Braun committed
51
bool be_is_transformed(const ir_node *node)
52
{
53
54
55
	return irn_visited(node);
}

56
57
ir_node *be_transform_phi(ir_node *node, const arch_register_req_t *req)
{
58
	ir_node  *block = be_transform_nodes_block(node);
59
	ir_graph *irg   = get_irn_irg(block);
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
	dbg_info *dbgi  = get_irn_dbg_info(node);

	/* phi nodes allow loops, so we use the old arguments for now
	 * and fix this later */
	ir_node **ins   = get_irn_in(node)+1;
	int       arity = get_irn_arity(node);
	ir_mode  *mode  = req->cls != NULL ? req->cls->mode : get_irn_mode(node);
	ir_node  *phi   = new_ir_node(dbgi, irg, block, op_Phi, mode, arity, ins);
	copy_node_attr(irg, node, phi);

	backend_info_t *info = be_get_info(phi);
	struct obstack *obst = be_get_be_obst(irg);
	info->in_reqs = OALLOCN(obst, const arch_register_req_t*, arity);
	for (int i = 0; i < arity; ++i) {
		info->in_reqs[i] = req;
	}

	arch_set_irn_register_req_out(phi, 0, req);
	be_enqueue_preds(node);

	return phi;
}

83
84
void be_set_transform_function(ir_op *op, be_transform_func func)
{
85
86
	/* Shouldn't be assigned twice. */
	assert(!op->ops.generic);
87
88
89
	op->ops.generic = (op_func) func;
}

90
91
92
93
94
void be_set_transform_proj_function(ir_op *op, be_transform_func func)
{
	op->ops.generic1 = (op_func) func;
}

95
96
97
98
99
/**
 * Transform helper for blocks.
 */
static ir_node *transform_block(ir_node *node)
{
Matthias Braun's avatar
Matthias Braun committed
100
101
102
103
104
	ir_graph *irg   = get_irn_irg(node);
	dbg_info *dbgi  = get_irn_dbg_info(node);
	ir_mode  *mode  = get_irn_mode(node);
	ir_node  *block = new_ir_node(dbgi, irg, NULL, get_irn_op(node), mode,
	                              get_irn_arity(node), get_irn_in(node) + 1);
105
106
107
	copy_node_attr(irg, node, block);
	block->node_nr = node->node_nr;

108
109
110
111
	/* transfer execfreq value */
	double execfreq = get_block_execfreq(node);
	set_block_execfreq(block, execfreq);

112
113
114
115
116
117
118
119
	/* put the preds in the worklist */
	be_enqueue_preds(node);

	return block;
}

static ir_node *transform_end(ir_node *node)
{
120
121
	/* Do not transform predecessors yet to keep the pre-transform
	 * phase from visiting all the graph. */
Matthias Braun's avatar
Matthias Braun committed
122
123
	ir_graph *irg     = get_irn_irg(node);
	dbg_info *dbgi    = get_irn_dbg_info(node);
124
	ir_node  *block   = be_transform_nodes_block(node);
125
126
127
	int       arity   = get_irn_arity(node);
	ir_node **ins     = get_irn_in(node) + 1;
	ir_node  *new_end = new_ir_node(dbgi, irg, block, op_End, mode_X, arity, ins);
128
129
130
131
132
133
134
135
136
	copy_node_attr(irg, node, new_end);

	set_irg_end(irg, new_end);

	be_enqueue_preds(node);

	return new_end;
}

137
138
139
140
141
142
143
static ir_node *transform_proj(ir_node *node)
{
	ir_node *pred    = get_Proj_pred(node);
	ir_op   *pred_op = get_irn_op(pred);
	be_transform_func *proj_transform
		= (be_transform_func*)pred_op->ops.generic1;
	/* we should have a Proj transformer registered */
144
145
146
147
148
149
150
151
152
153
154
155
#ifdef DEBUG_libfirm
	if (!proj_transform) {
		unsigned const node_pn = get_Proj_num(node);
		if (is_Proj(pred)) {
			unsigned const pred_pn   = get_Proj_num(pred);
			ir_node *const pred_pred = get_Proj_pred(pred);
			panic("no transformer for %+F (%u) -> %+F (%u) -> %+F", node, node_pn, pred, pred_pn, pred_pred);
		} else {
			panic("no transformer for %+F (%u) -> %+F", node, node_pn, pred);
		}
	}
#endif
156
157
158
	return proj_transform(node);
}

159
ir_node *be_duplicate_node(ir_node *const node)
160
{
161
162
163
164
	int       const arity = get_irn_arity(node);
	ir_node **const ins   = ALLOCAN(ir_node*, arity);
	foreach_irn_in(node, i, in) {
		ins[i] = be_transform_node(in);
165
166
	}

167
	ir_node *const block    = be_transform_nodes_block(node);
168
169
	ir_node *const new_node = new_similar_node(node, block, ins);

170
171
172
173
	new_node->node_nr = node->node_nr;
	return new_node;
}

174
175
ir_node *be_transform_node(ir_node *node)
{
176
177
178
179
180
181
182
183
	ir_node *new_node;
	if (be_is_transformed(node)) {
		new_node = (ir_node*)get_irn_link(node);
	} else {
		DEBUG_ONLY(be_set_transformed_node(node, NULL);)

		ir_op             *const op        = get_irn_op(node);
		be_transform_func *const transform = (be_transform_func*)op->ops.generic;
184
#ifdef DEBUG_libfirm
185
186
		if (!transform)
			panic("no transformer for %+F", node);
187
#endif
188

189
190
191
192
		new_node = transform(node);
		be_set_transformed_node(node, new_node);
	}
	assert(new_node);
193
194
195
	return new_node;
}

196
197
198
199
200
201
ir_node *be_transform_nodes_block(ir_node const *const node)
{
	ir_node *const block = get_nodes_block(node);
	return be_transform_node(block);
}

202
203
void be_enqueue_preds(ir_node *node)
{
204
	/* put the preds in the worklist */
205
	foreach_irn_in(node, i, pred) {
206
207
208
209
210
211
212
		pdeq_putr(env.worklist, pred);
	}
}

/**
 * Rewire nodes which are potential loops (like Phis) to avoid endless loops.
 */
213
214
static void fix_loops(ir_node *node)
{
215
	if (irn_visited_else_mark(node))
216
217
		return;

Matthias Braun's avatar
Matthias Braun committed
218
	bool changed = false;
219
220
	if (! is_Block(node)) {
		ir_node *block     = get_nodes_block(node);
221
		ir_node *new_block = (ir_node*)get_irn_link(block);
222
223
224
225

		if (new_block != NULL) {
			set_nodes_block(node, new_block);
			block = new_block;
Matthias Braun's avatar
Matthias Braun committed
226
			changed = true;
227
228
229
230
231
		}

		fix_loops(block);
	}

232
233
	foreach_irn_in(node, i, pred) {
		ir_node *in = pred;
234
		ir_node *nw = (ir_node*)get_irn_link(in);
235
236
237
238

		if (nw != NULL && nw != in) {
			set_irn_n(node, i, nw);
			in = nw;
Matthias Braun's avatar
Matthias Braun committed
239
			changed = true;
240
241
242
243
244
		}

		fix_loops(in);
	}

245
	if (changed) {
246
		identify_remember(node);
247
	}
248
249
}

250
251
ir_node *be_pre_transform_node(ir_node *place)
{
252
253
254
255
256
257
258
259
260
	if (place == NULL)
		return NULL;

	return be_transform_node(place);
}

/**
 * Transforms all nodes. Deletes the old obstack and creates a new one.
 */
261
static void transform_nodes(ir_graph *irg, arch_pretrans_nodes *pre_transform)
262
{
263
264
265
266
	hook_dead_node_elim(irg, 1);

	inc_irg_visited(irg);

267
	env.worklist = new_waitq();
268

269
	ir_node *const old_anchor = irg->anchor;
270
	ir_node *const new_anchor = new_r_Anchor(irg);
271
	ir_node *const old_end    = get_irg_end(irg);
272
	irg->anchor = new_anchor;
273

274
275
	/* Pre-transform all anchors (so they are available in the other transform
	 * functions) and put them into the worklist. */
276
277
	foreach_irn_in(old_anchor, i, old) {
		ir_node *const nw = be_transform_node(old);
278
		set_irn_n(new_anchor, i, nw);
279
	}
280
281

	if (pre_transform)
282
		pre_transform(irg);
283
284
285

	/* process worklist (this should transform all nodes in the graph) */
	while (! waitq_empty(env.worklist)) {
286
		ir_node *node = (ir_node*)waitq_get(env.worklist);
287
288
289
290
291
		be_transform_node(node);
	}

	/* fix loops and set new anchors*/
	inc_irg_visited(irg);
292
293
	foreach_irn_in_r(old_anchor, i, n) {
		ir_node *const anchor = (ir_node*)get_irn_link(n);
294
295
296
297
298
299
300
301
		fix_loops(anchor);
	}

	del_waitq(env.worklist);
	free_End(old_end);
	hook_dead_node_elim(irg, 0);
}

302
void be_transform_graph(ir_graph *irg, arch_pretrans_nodes *func)
303
304
{
	/* create a new obstack */
305
306
	struct obstack old_obst = irg->obst;
	obstack_init(&irg->obst);
307
308
	irg->last_node_idx = 0;

309
	free_vrp_data(irg);
310

311
	/* create new value table for CSE */
312
	new_identities(irg);
313
314

	/* do the main transformation */
315
	ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK);
316
	transform_nodes(irg, func);
317
	ir_free_resources(irg, IR_RESOURCE_IRN_LINK);
318
319

	/* free the old obstack */
320
	obstack_free(&old_obst, 0);
321

322
	/* most analysis info is wrong after transformation */
323
	be_invalidate_live_chk(irg);
Matthias Braun's avatar
Matthias Braun committed
324
	confirm_irg_properties(irg, IR_GRAPH_PROPERTIES_NONE);
325

326
327
328
	/* recalculate edges */
	edges_activate(irg);
}
329
330
331
332

bool be_upper_bits_clean(const ir_node *node, ir_mode *mode)
{
	ir_op *op = get_irn_op(node);
333
	if (op->ops.generic2 == NULL)
334
		return false;
335
	upper_bits_clean_func func = (upper_bits_clean_func)op->ops.generic2;
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
	return func(node, mode);
}

static bool bit_binop_upper_bits_clean(const ir_node *node, ir_mode *mode)
{
	return be_upper_bits_clean(get_binop_left(node), mode)
	    && be_upper_bits_clean(get_binop_right(node), mode);
}

static bool mux_upper_bits_clean(const ir_node *node, ir_mode *mode)
{
	return be_upper_bits_clean(get_Mux_true(node), mode)
	    && be_upper_bits_clean(get_Mux_false(node), mode);
}

static bool and_upper_bits_clean(const ir_node *node, ir_mode *mode)
{
	if (!mode_is_signed(mode)) {
		return be_upper_bits_clean(get_And_left(node), mode)
		    || be_upper_bits_clean(get_And_right(node), mode);
	} else {
		return bit_binop_upper_bits_clean(node, mode);
	}
}

static bool shr_upper_bits_clean(const ir_node *node, ir_mode *mode)
{
	if (mode_is_signed(mode)) {
		return false;
	} else {
		const ir_node *right = get_Shr_right(node);
		if (is_Const(right)) {
368
			long const val = get_Const_long(right);
369
370
371
372
373
374
375
376
377
378
379
380
381
382
			if (val >= 32 - (long)get_mode_size_bits(mode))
				return true;
		}
		return be_upper_bits_clean(get_Shr_left(node), mode);
	}
}

static bool shrs_upper_bits_clean(const ir_node *node, ir_mode *mode)
{
	return be_upper_bits_clean(get_Shrs_left(node), mode);
}

static bool const_upper_bits_clean(const ir_node *node, ir_mode *mode)
{
383
	long const val = get_Const_long(node);
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
	if (mode_is_signed(mode)) {
		long    shifted = val >> (get_mode_size_bits(mode)-1);
		return shifted == 0 || shifted == -1;
	} else {
		unsigned long shifted = (unsigned long)val;
		shifted >>= get_mode_size_bits(mode)-1;
		shifted >>= 1;
		return shifted == 0;
	}
}

static bool conv_upper_bits_clean(const ir_node *node, ir_mode *mode)
{
	ir_mode       *dest_mode = get_irn_mode(node);
	const ir_node *op        = get_Conv_op(node);
	ir_mode       *src_mode  = get_irn_mode(op);
	if (mode_is_float(src_mode))
		return true;

	unsigned src_bits  = get_mode_size_bits(src_mode);
	unsigned dest_bits = get_mode_size_bits(dest_mode);
	/* downconvs are a nop */
	if (src_bits >= dest_bits)
		return be_upper_bits_clean(op, mode);
	/* upconvs are fine if src is big enough or if sign matches */
	if (src_bits <= get_mode_size_bits(mode)
		&& mode_is_signed(src_mode) == mode_is_signed(mode))
		return true;
	return false;
}

static bool proj_upper_bits_clean(const ir_node *node, ir_mode *mode)
{
	const ir_node *pred = get_Proj_pred(node);
	switch (get_irn_opcode(pred)) {
	case iro_Load: {
		ir_mode *load_mode = get_Load_mode(pred);
		unsigned load_bits = get_mode_size_bits(load_mode);
		if (load_bits > get_mode_size_bits(mode))
			return false;
		if (mode_is_signed(load_mode) != mode_is_signed(mode))
			return false;
		return true;
	}
	default:
		break;
	}
	return false;
}

void be_set_upper_bits_clean_function(ir_op *op, upper_bits_clean_func func)
{
436
	op->ops.generic2 = (op_func)func;
437
438
439
440
441
442
}

void be_start_transform_setup(void)
{
	ir_clear_opcodes_generic_func();

443
444
445
446
447
448
	be_set_transform_function(op_Block, transform_block);
	be_set_transform_function(op_End,   transform_end);
	be_set_transform_function(op_NoMem, be_duplicate_node);
	be_set_transform_function(op_Pin,   be_duplicate_node);
	be_set_transform_function(op_Proj,  transform_proj);
	be_set_transform_function(op_Sync,  be_duplicate_node);
449
450
451
452
453
454
455
456
457
458
459

	be_set_upper_bits_clean_function(op_And,   and_upper_bits_clean);
	be_set_upper_bits_clean_function(op_Const, const_upper_bits_clean);
	be_set_upper_bits_clean_function(op_Conv,  conv_upper_bits_clean);
	be_set_upper_bits_clean_function(op_Eor,   bit_binop_upper_bits_clean);
	be_set_upper_bits_clean_function(op_Mux,   mux_upper_bits_clean);
	be_set_upper_bits_clean_function(op_Or,    bit_binop_upper_bits_clean);
	be_set_upper_bits_clean_function(op_Proj,  proj_upper_bits_clean);
	be_set_upper_bits_clean_function(op_Shr,   shr_upper_bits_clean);
	be_set_upper_bits_clean_function(op_Shrs,  shrs_upper_bits_clean);
}
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517

bool be_pattern_is_rotl(ir_node const *const irn_or, ir_node **const left,
                        ir_node **const right)
{
	assert(is_Add(irn_or) || is_Or(irn_or));

	ir_mode *mode = get_irn_mode(irn_or);
	if (!mode_is_int(mode))
		return false;

	ir_node *shl = get_binop_left(irn_or);
	ir_node *shr = get_binop_right(irn_or);
	if (is_Shr(shl)) {
		if (!is_Shl(shr))
			return false;

		ir_node *tmp = shl;
		shl = shr;
		shr = tmp;
	} else if (!is_Shl(shl)) {
		return false;
	} else if (!is_Shr(shr)) {
		return false;
	}

	ir_node *x = get_Shl_left(shl);
	if (x != get_Shr_left(shr))
		return false;

	ir_node *c1 = get_Shl_right(shl);
	ir_node *c2 = get_Shr_right(shr);
	if (is_Const(c1) && is_Const(c2)) {
		ir_tarval *tv1 = get_Const_tarval(c1);
		if (!tarval_is_long(tv1))
			return false;

		ir_tarval *tv2 = get_Const_tarval(c2);
		if (!tarval_is_long(tv2))
			return false;

		if (get_tarval_long(tv1) + get_tarval_long(tv2)
		    != (long) get_mode_size_bits(mode))
			return false;

		*left  = x;
		*right = c1;
		return true;
	}

	/* Note: the obvious rot formulation (a << x) | (a >> (32-x)) gets
	 * transformed to (a << x) | (a >> -x) by transform_node_shift_modulo() */
	if (!ir_is_negated_value(c1, c2))
		return false;

	*left  = x;
	*right = c1;
	return true;
}
Matthias Braun's avatar
Matthias Braun committed
518
519
520
521
522
523
524
525
526
527
528
529
530
531

void be_map_exc_node_to_runtime_call(ir_node *node, ir_mode *res_mode,
                                     ir_entity *runtime_entity,
                                     long pn_M, long pn_X_regular,
                                     long pn_X_except, long pn_res)
{
	assert(is_memop(node));

	size_t    n_in = get_irn_arity(node)-1;
	ir_node **in   = ALLOCAN(ir_node*, n_in);
	ir_type  *mtp  = get_entity_type(runtime_entity);

	assert(get_method_n_params(mtp) == n_in);
	size_t p = 0;
532
	foreach_irn_in(node, i, n) {
Matthias Braun's avatar
Matthias Braun committed
533
534
535
536
537
538
539
540
		if (get_irn_mode(n) == mode_M)
			continue;
		in[p++] = n;
	}
	assert(p == n_in);

	ir_graph *irg   = get_irn_irg(node);
	dbg_info *dbgi  = get_irn_dbg_info(node);
541
	ir_node  *addr  = new_r_Address(irg, runtime_entity);
Matthias Braun's avatar
Matthias Braun committed
542
543
544
545
546
547
548
	ir_node  *block = get_nodes_block(node);
	ir_node  *mem   = get_memop_mem(node);
	ir_node  *call  = new_rd_Call(dbgi, block, mem, addr, n_in, in, mtp);
	set_irn_pinned(call, get_irn_pinned(node));
	int throws_exception = ir_throws_exception(node);
	ir_set_throws_exception(call, throws_exception);

549
550
551
552
	assert(pn_M < 2 && pn_res < 2 && pn_X_regular < 4 && pn_X_except < 4);
	int const         n_proj     = 4;
	int               n_operands = 2;
	ir_node   **const tuple_in   = ALLOCAN(ir_node*, n_proj);
Matthias Braun's avatar
Matthias Braun committed
553
554
555
556
	tuple_in[pn_M] = new_r_Proj(call, mode_M, pn_Call_M);
	ir_node *ress = new_r_Proj(call, mode_T, pn_Call_T_result);
	tuple_in[pn_res] = new_r_Proj(ress, res_mode, 0);
	if (throws_exception) {
557
558
559
		tuple_in[pn_X_regular]  = new_r_Proj(call, mode_X, pn_Call_X_regular);
		tuple_in[pn_X_except]   = new_r_Proj(call, mode_X, pn_Call_X_except);
		n_operands             += 2;
Matthias Braun's avatar
Matthias Braun committed
560
561
	}

562
	turn_into_tuple(node, n_operands, tuple_in);
Matthias Braun's avatar
Matthias Braun committed
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

/**
 * Link the node into its block list as a new head.
 */
static void collect_node(ir_node *node)
{
	ir_node *block = get_nodes_block(node);
	ir_node *old   = (ir_node*)get_irn_link(block);

	set_irn_link(node, old);
	set_irn_link(block, node);
}

/**
 * Post-walker: link all nodes that probably access the stack into lists of their block.
 */
static void link_ops_in_block_walker(ir_node *node, void *data)
{
	(void) data;

	switch (get_irn_opcode(node)) {
	case iro_Return:
	case iro_Call:
		collect_node(node);
		break;
	case iro_Alloc:
		/** all non-stack alloc nodes should be lowered before the backend */
		collect_node(node);
		break;
	case iro_Free:
		collect_node(node);
		break;
	case iro_Builtin:
		if (get_Builtin_kind(node) == ir_bk_return_address) {
			ir_node *const param = get_Builtin_param(node, 0);
			long     const value = get_Const_long(param); /* must be Const */
			if (value > 0) {
				/* not the return address of the current function:
				 * we need the stack pointer for the frame climbing */
				collect_node(node);
			}
		}
		break;
	default:
		break;
	}
}

static ir_heights_t *heights;

/**
 * Check if a node is somehow data dependent on another one.
 * both nodes must be in the same basic block.
 * @param n1 The first node.
 * @param n2 The second node.
 * @return 1, if n1 is data dependent (transitively) on n2, 0 if not.
 */
static int dependent_on(const ir_node *n1, const ir_node *n2)
{
	assert(get_nodes_block(n1) == get_nodes_block(n2));
	return heights_reachable_in_block(heights, n1, n2);
}

/**
 * Classical qsort() comparison function behavior:
 *
 * 0  if both elements are equal, no node depend on the other
 * +1 if first depends on second (first is greater)
 * -1 if second depends on first (second is greater)
*/
static int cmp_call_dependency(const void *c1, const void *c2)
{
	const ir_node *n1 = *(const ir_node **) c1;
	const ir_node *n2 = *(const ir_node **) c2;
	if (dependent_on(n1, n2))
		return 1;
	if (dependent_on(n2, n1))
		return -1;

	/* The nodes have no depth order, but we need a total order because qsort()
	 * is not stable.
	 *
	 * Additionally, we need to respect transitive dependencies. Consider a
	 * Call a depending on Call b and an independent Call c.
	 * We MUST NOT order c > a and b > c. */
Matthias Braun's avatar
Matthias Braun committed
649
650
651
652
653
654
	unsigned h1 = get_irn_height(heights, n1);
	unsigned h2 = get_irn_height(heights, n2);
	if (h1 < h2)
		return 1;
	if (h1 > h2)
		return -1;
655
656
657
658
659
660
661
662
663
	/* Same height, so use a random (but stable) order */
	return get_irn_idx(n2) - get_irn_idx(n1);
}

/**
 * Block-walker: sorts dependencies and remember them into a phase
 */
static void process_ops_in_block(ir_node *block, void *data)
{
664
	ir_nodemap *const map = (ir_nodemap*)data;
665

666
667
668
	ir_node **nodes = NEW_ARR_F(ir_node*, 0);
	for (ir_node *node = block; (node = (ir_node*)get_irn_link(node));) {
		ARR_APP1(ir_node*, nodes, node);
669
670
	}

671
672
673
674
	unsigned const n_nodes = ARR_LEN(nodes);
	if (n_nodes != 0) {
		/* order nodes according to their data dependencies */
		QSORT(nodes, n_nodes, cmp_call_dependency);
675

676
677
678
679
680
681
		/* remember the calculated dependency into a phase */
		for (unsigned n = n_nodes - 1; n > 0; --n) {
			ir_node *const node = nodes[n];
			ir_node *const pred = nodes[n - 1];
			ir_nodemap_insert(map, node, pred);
		}
682
	}
683
684

	DEL_ARR_F(nodes);
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
}

struct be_stackorder_t {
	ir_nodemap stack_order; /**< a phase to handle stack dependencies. */
};

be_stackorder_t *be_collect_stacknodes(ir_graph *irg)
{
	be_stackorder_t *env = XMALLOCZ(be_stackorder_t);

	ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK);

	/* collect all potential^stack accessing nodes */
	irg_walk_graph(irg, firm_clear_link, link_ops_in_block_walker, NULL);

	ir_nodemap_init(&env->stack_order, irg);

	/* use heights to create a total order for those nodes: this order is stored
	 * in the created phase */
	heights = heights_new(irg);
	irg_block_walk_graph(irg, NULL, process_ops_in_block, &env->stack_order);
	heights_free(heights);

	ir_free_resources(irg, IR_RESOURCE_IRN_LINK);

	return env;
}

ir_node *be_get_stack_pred(const be_stackorder_t *env, const ir_node *node)
{
	return ir_nodemap_get(ir_node, &env->stack_order, node);
}

void be_free_stackorder(be_stackorder_t *env)
{
	ir_nodemap_destroy(&env->stack_order);
	free(env);
}

static void create_stores_for_type(ir_graph *irg, ir_type *type)
{
	ir_node *frame       = get_irg_frame(irg);
	ir_node *initial_mem = get_irg_initial_mem(irg);
	ir_node *mem         = initial_mem;
	ir_node *first_store = NULL;
	ir_node *start_block = get_irg_start_block(irg);
	ir_node *args        = get_irg_args(irg);

	/* all parameter entities left in the frame type require stores.
	 * (The ones passed on the stack have been moved to the arg type) */
Matthias Braun's avatar
Matthias Braun committed
735
	for (size_t i = 0, n = get_compound_n_members(type); i < n; ++i) {
736
737
738
739
740
		ir_entity *entity = get_compound_member(type, i);
		ir_type   *tp     = get_entity_type(entity);
		if (!is_parameter_entity(entity))
			continue;

Matthias Braun's avatar
Matthias Braun committed
741
		size_t arg = get_entity_parameter_number(entity);
742
743
744
		if (arg == IR_VA_START_PARAMETER_NUMBER)
			continue;

Matthias Braun's avatar
Matthias Braun committed
745
		ir_node *addr = new_r_Member(start_block, frame, entity);
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
		if (entity->attr.parameter.doubleword_low_mode != NULL) {
			ir_mode *mode      = entity->attr.parameter.doubleword_low_mode;
			ir_node *val0      = new_r_Proj(args, mode, arg);
			ir_node *val1      = new_r_Proj(args, mode, arg+1);
			ir_node *store0    = new_r_Store(start_block, mem, addr, val0,
			                                 tp, cons_none);
			ir_node *mem0      = new_r_Proj(store0, mode_M, pn_Store_M);
			size_t   offset    = get_mode_size_bits(mode)/8;
			ir_mode *mode_ref  = get_irn_mode(addr);
			ir_mode *mode_offs = get_reference_mode_unsigned_eq(mode_ref);
			ir_node *cnst      = new_r_Const_long(irg, mode_offs, offset);
			ir_node *next_addr = new_r_Add(start_block, addr, cnst, mode_ref);
			ir_node *store1    = new_r_Store(start_block, mem0, next_addr, val1,
			                                 tp, cons_none);
			mem = new_r_Proj(store1, mode_M, pn_Store_M);
			if (first_store == NULL)
				first_store = store0;
		} else {
			ir_mode *mode  = is_compound_type(tp) ? mode_P : get_type_mode(tp);
			ir_node *val   = new_r_Proj(args, mode, arg);
			ir_node *store = new_r_Store(start_block, mem, addr, val, tp, cons_none);
			mem = new_r_Proj(store, mode_M, pn_Store_M);
			if (first_store == NULL)
				first_store = store;
		}
	}

	if (mem != initial_mem) {
		edges_reroute_except(initial_mem, mem, first_store);
		set_irg_initial_mem(irg, initial_mem);
	}
}

void be_add_parameter_entity_stores(ir_graph *irg)
{
	ir_type           *frame_type   = get_irg_frame_type(irg);
	be_stack_layout_t *layout       = be_get_irg_stack_layout(irg);
	ir_type           *between_type = layout->between_type;

	create_stores_for_type(irg, frame_type);
Matthias Braun's avatar
Matthias Braun committed
786
	if (between_type != NULL)
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
		create_stores_for_type(irg, between_type);
}

unsigned be_get_n_allocatable_regs(const ir_graph *irg,
                                   const arch_register_class_t *cls)
{
	unsigned *const bs = rbitset_alloca(cls->n_regs);
	be_get_allocatable_regs(irg, cls, bs);
	return rbitset_popcount(bs, cls->n_regs);
}

void be_get_allocatable_regs(ir_graph const *const irg,
                             arch_register_class_t const *const cls,
                             unsigned *const raw_bitset)
{
	be_irg_t *birg             = be_birg_from_irg(irg);
	unsigned *allocatable_regs = birg->allocatable_regs;

	rbitset_clear_all(raw_bitset, cls->n_regs);
	for (unsigned i = 0; i < cls->n_regs; ++i) {
		const arch_register_t *reg = &cls->regs[i];
		if (rbitset_is_set(allocatable_regs, reg->global_index))
			rbitset_set(raw_bitset, i);
	}
}
812
813
814
815
816
817
818
819
820
821

uint32_t be_get_tv_bits32(ir_tarval *const tv, unsigned const offset)
{
	uint32_t val;
	val  = (uint32_t)get_tarval_sub_bits(tv, offset);
	val |= (uint32_t)get_tarval_sub_bits(tv, offset + 1) <<  8;
	val |= (uint32_t)get_tarval_sub_bits(tv, offset + 2) << 16;
	val |= (uint32_t)get_tarval_sub_bits(tv, offset + 3) << 24;
	return val;
}
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848

static bool mode_needs_gp_reg(ir_mode *const mode)
{
	return get_mode_arithmetic(mode) == irma_twos_complement;
}

ir_node *be_skip_downconv(ir_node *node, bool const single_user)
{
	assert(mode_needs_gp_reg(get_irn_mode(node)));
	for (;;) {
		if (single_user && get_irn_n_edges(node) > 1) {
			/* we only want to skip the conv when we're the only user
			 * (because this test is used in the context of address-mode selection
			 *  and we don't want to use address mode for multiple users) */
			break;
		} else if (is_Conv(node)) {
			ir_node *const op       = get_Conv_op(node);
			ir_mode *const src_mode = get_irn_mode(op);
			if (!mode_needs_gp_reg(src_mode) || get_mode_size_bits(get_irn_mode(node)) > get_mode_size_bits(src_mode))
				break;
			node = op;
		} else {
			break;
		}
	}
	return node;
}
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870

ir_node *be_skip_sameconv(ir_node *node)
{
	assert(mode_needs_gp_reg(get_irn_mode(node)));
	for (;;) {
		if (get_irn_n_edges(node) > 1) {
			/* we only want to skip the conv when we're the only user
			 * (because this test is used in the context of address-mode selection
			 *  and we don't want to use address mode for multiple users) */
			break;
		} else if (is_Conv(node)) {
			ir_node *const op       = get_Conv_op(node);
			ir_mode *const src_mode = get_irn_mode(op);
			if (!mode_needs_gp_reg(src_mode) || get_mode_size_bits(get_irn_mode(node)) != get_mode_size_bits(src_mode))
				break;
			node = op;
		} else {
			break;
		}
	}
	return node;
}
871
872
873
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
903
904
905
906
907
908
909
910

bool be_match_immediate(ir_node const *const node, ir_tarval **const tarval_out, ir_entity **const entity_out)
{
	ir_node const *addr;
	ir_node const *cnst;
	if (is_Const(node)) {
		addr = NULL;
		cnst = node;
	} else if (is_Address(node)) {
		addr = node;
		cnst = NULL;
	} else if (is_Add(node)) {
		ir_node const *const l = get_Add_left(node);
		ir_node const *const r = get_Add_right(node);
		if (is_Address(l) && is_Const(r)) {
			addr = l;
			cnst = r;
		} else if (is_Const(l) && is_Address(r)) {
			addr = r;
			cnst = l;
		} else {
			return false;
		}
	} else {
		return false;
	}

	ir_entity *entity;
	if (addr) {
		entity = get_Address_entity(addr);
		if (is_tls_entity(entity))
			return false;
	} else {
		entity = NULL;
	}

	*tarval_out = cnst ? get_Const_tarval(cnst) : NULL;
	*entity_out = entity;
	return true;
}