lower_dw.c 94.9 KB
Newer Older
Christian Würdig's avatar
Christian Würdig committed
1
2
/*
 * This file is part of libFirm.
3
 * Copyright (C) 2012 University of Karlsruhe.
Christian Würdig's avatar
Christian Würdig committed
4
5
 */

Michael Beck's avatar
Michael Beck committed
6
7
/**
 * @file
yb9976's avatar
yb9976 committed
8
 * @brief   Lower double word operations, i.e. 64bit -> 32bit, 32bit -> 16bit etc.
Michael Beck's avatar
Michael Beck committed
9
10
 * @date    8.10.2004
 * @author  Michael Beck
11
 */
Matthias Braun's avatar
Matthias Braun committed
12
#include "config.h"
13

14
#include <string.h>
15
#include <stdlib.h>
16
#include <stdbool.h>
17
18
#include <assert.h>

yb9976's avatar
yb9976 committed
19
#include "be.h"
20
#include "error.h"
21
#include "lowering.h"
22
#include "irnode_t.h"
yb9976's avatar
yb9976 committed
23
#include "irnodeset.h"
24
25
26
27
28
29
30
31
32
33
34
35
#include "irgraph_t.h"
#include "irmode_t.h"
#include "iropt_t.h"
#include "irgmod.h"
#include "tv_t.h"
#include "dbginfo_t.h"
#include "iropt_dbg.h"
#include "irflag_t.h"
#include "firmstat.h"
#include "irgwalk.h"
#include "ircons.h"
#include "irflag.h"
yb9976's avatar
yb9976 committed
36
#include "iroptimize.h"
37
38
39
40
#include "debug.h"
#include "set.h"
#include "pmap.h"
#include "pdeq.h"
Christian Würdig's avatar
Christian Würdig committed
41
#include "irdump.h"
42
#include "array_t.h"
Michael Beck's avatar
Michael Beck committed
43
#include "irpass_t.h"
44
#include "lower_dw.h"
45

Michael Beck's avatar
Michael Beck committed
46
/** A map from (op, imode, omode) to Intrinsic functions entities. */
47
48
static set *intrinsic_fkt;

Michael Beck's avatar
Michael Beck committed
49
50
51
/** A map from (imode, omode) to conv function types. */
static set *conv_types;

52
53
54
/** A map from a method type to its lowered type. */
static pmap *lowered_type;

yb9976's avatar
yb9976 committed
55
56
57
58
/** A map from a builtin type to its lower and higher type. */
static pmap *lowered_builtin_type_high;
static pmap *lowered_builtin_type_low;

59
/** The types for the binop and unop intrinsics. */
60
static ir_type *binop_tp_u, *binop_tp_s, *unop_tp_u, *unop_tp_s, *tp_s, *tp_u;
61

yb9976's avatar
yb9976 committed
62
63
static ir_nodeset_t created_mux_nodes;

64
65
66
67
/** the debug handle */
DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)

/**
Michael Beck's avatar
Michael Beck committed
68
 * An entry in the (op, imode, omode) -> entity map.
69
 */
70
typedef struct op_mode_entry {
71
	const ir_op   *op;    /**< the op */
Michael Beck's avatar
Michael Beck committed
72
73
	const ir_mode *imode; /**< the input mode */
	const ir_mode *omode; /**< the output mode */
74
	ir_entity     *ent;   /**< the associated entity of this (op, imode, omode) triple */
75
76
} op_mode_entry_t;

Michael Beck's avatar
Michael Beck committed
77
78
79
/**
 * An entry in the (imode, omode) -> tp map.
 */
80
typedef struct conv_tp_entry {
Michael Beck's avatar
Michael Beck committed
81
82
83
84
85
	const ir_mode *imode; /**< the input mode */
	const ir_mode *omode; /**< the output mode */
	ir_type       *mtd;   /**< the associated method type of this (imode, omode) pair */
} conv_tp_entry_t;

86
enum lower_flags {
87
88
	MUST_BE_LOWERED = 1,  /**< graph must be lowered */
	CF_CHANGED      = 2,  /**< control flow was changed */
89
90
91
92
93
};

/**
 * The lower environment.
 */
94
95
typedef struct lower_dw_env_t {
	lower64_entry_t **entries;     /**< entries per node */
96
	ir_graph      *irg;
97
98
	struct obstack obst;           /**< an obstack holding the temporary data */
	ir_tarval *tv_mode_bytes;      /**< a tarval containing the number of bytes in the lowered modes */
Matthias Braun's avatar
Matthias Braun committed
99
100
101
102
103
104
105
106
	pdeq      *waitq;              /**< a wait queue of all nodes that must be handled later */
	ir_node  **lowered_phis;       /**< list of lowered phis */
	ir_mode   *high_signed;        /**< doubleword signed type */
	ir_mode   *high_unsigned;      /**< doubleword unsigned type */
	ir_mode   *low_signed;         /**< word signed type */
	ir_mode   *low_unsigned;       /**< word unsigned type */
	ident     *first_id;           /**< .l for little and .h for big endian */
	ident     *next_id;            /**< .h for little and .l for big endian */
107
108
109
110
	const lwrdw_param_t *params;   /**< transformation parameter */
	unsigned flags;                /**< some flags */
	unsigned n_entries;            /**< number of entries */
} lower_dw_env_t;
111

112
113
114
static lower_dw_env_t *env;

static void lower_node(ir_node *node);
115

Michael Beck's avatar
Michael Beck committed
116
117
118
/**
 * Create a method type for a Conv emulation from imode to omode.
 */
119
static ir_type *get_conv_type(ir_mode *imode, ir_mode *omode)
120
{
Michael Beck's avatar
Michael Beck committed
121
122
123
124
125
126
127
	conv_tp_entry_t key, *entry;
	ir_type *mtd;

	key.imode = imode;
	key.omode = omode;
	key.mtd   = NULL;

128
	entry = set_insert(conv_tp_entry_t, conv_types, &key, sizeof(key), hash_ptr(imode) ^ hash_ptr(omode));
Michael Beck's avatar
Michael Beck committed
129
130
131
	if (! entry->mtd) {
		int n_param = 1, n_res = 1;

132
		if (imode == env->high_signed || imode == env->high_unsigned)
Michael Beck's avatar
Michael Beck committed
133
			n_param = 2;
134
		if (omode == env->high_signed || omode == env->high_unsigned)
Michael Beck's avatar
Michael Beck committed
135
136
137
			n_res = 2;

		/* create a new one */
138
		mtd = new_type_method(n_param, n_res);
Michael Beck's avatar
Michael Beck committed
139
140
141

		/* set param types and result types */
		n_param = 0;
142
		if (imode == env->high_signed) {
143
144
145
146
147
148
149
			if (env->params->little_endian) {
				set_method_param_type(mtd, n_param++, tp_u);
				set_method_param_type(mtd, n_param++, tp_s);
			} else {
				set_method_param_type(mtd, n_param++, tp_s);
				set_method_param_type(mtd, n_param++, tp_u);
			}
150
		} else if (imode == env->high_unsigned) {
Michael Beck's avatar
Michael Beck committed
151
152
			set_method_param_type(mtd, n_param++, tp_u);
			set_method_param_type(mtd, n_param++, tp_u);
Michael Beck's avatar
BugFix:    
Michael Beck committed
153
		} else {
154
			ir_type *tp = get_type_for_mode(imode);
Michael Beck's avatar
Michael Beck committed
155
			set_method_param_type(mtd, n_param++, tp);
156
		}
Michael Beck's avatar
Michael Beck committed
157
158

		n_res = 0;
159
		if (omode == env->high_signed) {
160
161
162
163
164
165
166
			if (env->params->little_endian) {
				set_method_res_type(mtd, n_res++, tp_u);
				set_method_res_type(mtd, n_res++, tp_s);
			} else {
				set_method_res_type(mtd, n_res++, tp_s);
				set_method_res_type(mtd, n_res++, tp_u);
			}
167
		} else if (omode == env->high_unsigned) {
Michael Beck's avatar
Michael Beck committed
168
169
			set_method_res_type(mtd, n_res++, tp_u);
			set_method_res_type(mtd, n_res++, tp_u);
Michael Beck's avatar
BugFix:    
Michael Beck committed
170
		} else {
171
			ir_type *tp = get_type_for_mode(omode);
Michael Beck's avatar
Michael Beck committed
172
			set_method_res_type(mtd, n_res++, tp);
173
		}
Michael Beck's avatar
Michael Beck committed
174
		entry->mtd = mtd;
Michael Beck's avatar
BugFix:    
Michael Beck committed
175
	} else {
Michael Beck's avatar
Michael Beck committed
176
		mtd = entry->mtd;
177
	}
Michael Beck's avatar
Michael Beck committed
178
	return mtd;
179
}
Michael Beck's avatar
Michael Beck committed
180

181
182
183
184
185
186
187
/**
 * Add an additional control flow input to a block.
 * Patch all Phi nodes. The new Phi inputs are copied from
 * old input number nr.
 */
static void add_block_cf_input_nr(ir_node *block, int nr, ir_node *cf)
{
188
	int i, arity = get_Block_n_cfgpreds(block);
189
	ir_node **in;
190
191
192
193
194

	assert(nr < arity);

	NEW_ARR_A(ir_node *, in, arity + 1);
	for (i = 0; i < arity; ++i)
195
		in[i] = get_Block_cfgpred(block, i);
196
197
198
199
	in[i] = cf;

	set_irn_in(block, i + 1, in);

200
201
202
203
204
	foreach_out_edge(block, edge) {
		ir_node *phi = get_edge_src_irn(edge);
		if (!is_Phi(phi))
			continue;

205
206
207
208
		for (i = 0; i < arity; ++i)
			in[i] = get_irn_n(phi, i);
		in[i] = in[nr];
		set_irn_in(phi, i + 1, in);
209
210
	}
}
211
212
213
214
215
216
217
218

/**
 * Add an additional control flow input to a block.
 * Patch all Phi nodes. The new Phi inputs are copied from
 * old input from cf tmpl.
 */
static void add_block_cf_input(ir_node *block, ir_node *tmpl, ir_node *cf)
{
219
	int i, arity = get_Block_n_cfgpreds(block);
220
221
222
	int nr = 0;

	for (i = 0; i < arity; ++i) {
223
		if (get_Block_cfgpred(block, i) == tmpl) {
224
225
			nr = i;
			break;
226
227
		}
	}
228
229
	assert(i < arity);
	add_block_cf_input_nr(block, nr, cf);
230
}
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249

/**
 * Return the "operational" mode of a Firm node.
 */
static ir_mode *get_irn_op_mode(ir_node *node)
{
	switch (get_irn_opcode(node)) {
	case iro_Load:
		return get_Load_mode(node);
	case iro_Store:
		return get_irn_mode(get_Store_value(node));
	case iro_Div:
		return get_irn_mode(get_Div_left(node));
	case iro_Mod:
		return get_irn_mode(get_Mod_left(node));
	case iro_Cmp:
		return get_irn_mode(get_Cmp_left(node));
	default:
		return get_irn_mode(node);
250
251
	}
}
252
253

/**
254
255
 * Walker, prepare the node links and determine which nodes need to be lowered
 * at all.
256
 */
257
static void prepare_links(ir_node *node)
258
{
259
260
	ir_mode         *mode = get_irn_op_mode(node);
	lower64_entry_t *link;
261

262
	if (mode == env->high_signed || mode == env->high_unsigned) {
263
		unsigned idx = get_irn_idx(node);
264
		/* ok, found a node that will be lowered */
265
		link = OALLOCZ(&env->obst, lower64_entry_t);
266

267
		if (idx >= env->n_entries) {
Michael Beck's avatar
Michael Beck committed
268
			/* enlarge: this happens only for Rotl nodes which is RARELY */
269
			unsigned old   = env->n_entries;
270
			unsigned n_idx = idx + (idx >> 3);
Michael Beck's avatar
Michael Beck committed
271

272
			ARR_RESIZE(lower64_entry_t *, env->entries, n_idx);
273
274
			memset(&env->entries[old], 0, (n_idx - old) * sizeof(env->entries[0]));
			env->n_entries = n_idx;
Michael Beck's avatar
Michael Beck committed
275
		}
276
277
		env->entries[idx] = link;
		env->flags |= MUST_BE_LOWERED;
278
	} else if (is_Conv(node)) {
279
280
281
282
		/* Conv nodes have two modes */
		ir_node *pred = get_Conv_op(node);
		mode = get_irn_mode(pred);

283
		if (mode == env->high_signed || mode == env->high_unsigned) {
Michael Beck's avatar
BugFix:    
Michael Beck committed
284
			/* must lower this node either but don't need a link */
285
			env->flags |= MUST_BE_LOWERED;
286
		}
287
		return;
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
	} else if (is_Call(node)) {
		/* Special case:  If the result of the Call is never used, we won't
		 * find a Proj with a mode that potentially triggers MUST_BE_LOWERED
		 * to be set.  Thus, if we see a call, we check its result types and
		 * decide whether MUST_BE_LOWERED has to be set.
		 */
		ir_type *tp = get_Call_type(node);
		size_t   n_res, i;

		n_res = get_method_n_ress(tp);
		for (i = 0; i < n_res; ++i) {
			ir_type *rtp = get_method_res_type(tp, i);

			if (is_Primitive_type(rtp)) {
				ir_mode *rmode = get_type_mode(rtp);

				if (rmode == env->high_signed || rmode == env->high_unsigned) {
					env->flags |= MUST_BE_LOWERED;
				}
			}
		}
309
310
	}
}
311

312
lower64_entry_t *get_node_entry(ir_node *node)
313
314
315
316
317
318
{
	unsigned idx = get_irn_idx(node);
	assert(idx < env->n_entries);
	return env->entries[idx];
}

319
void ir_set_dw_lowered(ir_node *old, ir_node *new_low, ir_node *new_high)
320
{
321
	lower64_entry_t *entry = get_node_entry(old);
322
323
324
325
	entry->low_word  = new_low;
	entry->high_word = new_high;
}

326
327
328
329
330
ir_mode *ir_get_low_unsigned_mode(void)
{
	return env->low_unsigned;
}

331
332
333
/**
 * Translate a Constant: create two.
 */
334
static void lower_Const(ir_node *node, ir_mode *mode)
335
{
Matthias Braun's avatar
Matthias Braun committed
336
337
338
339
340
341
	ir_graph  *irg      = get_irn_irg(node);
	dbg_info  *dbg      = get_irn_dbg_info(node);
	ir_mode   *low_mode = env->low_unsigned;
	ir_tarval *tv       = get_Const_tarval(node);
	ir_tarval *tv_l     = tarval_convert_to(tv, low_mode);
	ir_node   *res_low  = new_rd_Const(dbg, irg, tv_l);
342
	ir_tarval *tv_shrs  = tarval_shrs_unsigned(tv, get_mode_size_bits(low_mode));
Matthias Braun's avatar
Matthias Braun committed
343
344
	ir_tarval *tv_h     = tarval_convert_to(tv_shrs, mode);
	ir_node   *res_high = new_rd_Const(dbg, irg, tv_h);
345

346
	ir_set_dw_lowered(node, res_low, res_high);
347
}
348
349
350
351

/**
 * Translate a Load: create two.
 */
352
static void lower_Load(ir_node *node, ir_mode *mode)
353
{
354
	ir_mode    *low_mode = env->low_unsigned;
355
	ir_graph   *irg = get_irn_irg(node);
356
357
	ir_node    *adr = get_Load_ptr(node);
	ir_node    *mem = get_Load_mem(node);
358
359
360
	ir_node    *low;
	ir_node    *high;
	ir_node    *proj_m;
361
362
	dbg_info   *dbg;
	ir_node    *block = get_nodes_block(node);
363
	ir_cons_flags volatility = get_Load_volatility(node) == volatility_is_volatile
364
	                         ? cons_volatile : cons_none;
365
366
367

	if (env->params->little_endian) {
		low  = adr;
368
		high = new_r_Add(block, adr, new_r_Const(irg, env->tv_mode_bytes), get_irn_mode(adr));
Michael Beck's avatar
BugFix:    
Michael Beck committed
369
	} else {
370
		low  = new_r_Add(block, adr, new_r_Const(irg, env->tv_mode_bytes), get_irn_mode(adr));
371
		high = adr;
372
	}
373
374

	/* create two loads */
375
376
377
378
	dbg    = get_irn_dbg_info(node);
	low    = new_rd_Load(dbg, block, mem,  low,  low_mode, volatility);
	proj_m = new_r_Proj(low, mode_M, pn_Load_M);
	high   = new_rd_Load(dbg, block, proj_m, high, mode, volatility);
379

380
	foreach_out_edge_safe(node, edge) {
381
382
383
		ir_node *proj = get_edge_src_irn(edge);
		if (!is_Proj(proj))
			continue;
384
385
386
387
388
389
390
391
392
393

		switch (get_Proj_proj(proj)) {
		case pn_Load_M:         /* Memory result. */
			/* put it to the second one */
			set_Proj_pred(proj, high);
			break;
		case pn_Load_X_except:  /* Execution result if exception occurred. */
			/* put it to the first one */
			set_Proj_pred(proj, low);
			break;
394
395
396
		case pn_Load_res: {       /* Result of load operation. */
			ir_node *res_low  = new_r_Proj(low,  low_mode, pn_Load_res);
			ir_node *res_high = new_r_Proj(high, mode,     pn_Load_res);
397
			ir_set_dw_lowered(proj, res_low, res_high);
398
			break;
399
		}
400
401
		default:
			assert(0 && "unexpected Proj number");
402
		}
403
404
		/* mark this proj: we have handled it already, otherwise we might fall
		 * into out new nodes. */
405
		mark_irn_visited(proj);
406
407
	}
}
408
409
410
411

/**
 * Translate a Store: create two.
 */
412
static void lower_Store(ir_node *node, ir_mode *mode)
413
{
414
415
	ir_graph              *irg;
	ir_node               *block, *adr, *mem;
416
	ir_node               *low, *high, *proj_m;
417
418
419
	dbg_info              *dbg;
	ir_node               *value = get_Store_value(node);
	const lower64_entry_t *entry = get_node_entry(value);
420
	ir_cons_flags volatility = get_Store_volatility(node) == volatility_is_volatile
421
	                           ? cons_volatile : cons_none;
Matthias Braun's avatar
Matthias Braun committed
422
	(void) mode;
423
424
425
426
427
428
429

	assert(entry);

	if (! entry->low_word) {
		/* not ready yet, wait */
		pdeq_putr(env->waitq, node);
		return;
430
	}
431

432
	irg = get_irn_irg(node);
433
434
435
436
437
438
	adr = get_Store_ptr(node);
	mem = get_Store_mem(node);
	block = get_nodes_block(node);

	if (env->params->little_endian) {
		low  = adr;
439
		high = new_r_Add(block, adr, new_r_Const(irg, env->tv_mode_bytes), get_irn_mode(adr));
Michael Beck's avatar
BugFix:    
Michael Beck committed
440
	} else {
441
		low  = new_r_Add(block, adr, new_r_Const(irg, env->tv_mode_bytes), get_irn_mode(adr));
442
		high = adr;
443
	}
444
445

	/* create two Stores */
446
447
448
449
	dbg    = get_irn_dbg_info(node);
	low    = new_rd_Store(dbg, block, mem, low,  entry->low_word, volatility);
	proj_m = new_r_Proj(low, mode_M, pn_Store_M);
	high   = new_rd_Store(dbg, block, proj_m, high, entry->high_word, volatility);
450

451
	foreach_out_edge_safe(node, edge) {
452
453
454
		ir_node *proj = get_edge_src_irn(edge);
		if (!is_Proj(proj))
			continue;
455
456
457
458
459
460
461
462
463
464
465
466

		switch (get_Proj_proj(proj)) {
		case pn_Store_M:         /* Memory result. */
			/* put it to the second one */
			set_Proj_pred(proj, high);
			break;
		case pn_Store_X_except:  /* Execution result if exception occurred. */
			/* put it to the first one */
			set_Proj_pred(proj, low);
			break;
		default:
			assert(0 && "unexpected Proj number");
467
		}
468
469
470
		/* mark this proj: we have handled it already, otherwise we might fall into
		 * out new nodes. */
		mark_irn_visited(proj);
471
472
	}
}
473
474
475

/**
 * Return a node containing the address of the intrinsic emulation function.
Michael Beck's avatar
Michael Beck committed
476
477
478
479
480
481
 *
 * @param method  the method type of the emulation function
 * @param op      the emulated ir_op
 * @param imode   the input mode of the emulated opcode
 * @param omode   the output mode of the emulated opcode
 * @param env     the lower environment
482
 */
Michael Beck's avatar
Michael Beck committed
483
static ir_node *get_intrinsic_address(ir_type *method, ir_op *op,
484
                                      ir_mode *imode, ir_mode *omode)
485
{
486
	symconst_symbol sym;
487
	ir_entity *ent;
488
489
	op_mode_entry_t key, *entry;

Michael Beck's avatar
Michael Beck committed
490
491
492
493
	key.op    = op;
	key.imode = imode;
	key.omode = omode;
	key.ent   = NULL;
494

495
	entry = set_insert(op_mode_entry_t, intrinsic_fkt, &key, sizeof(key),
496
				hash_ptr(op) ^ hash_ptr(imode) ^ (hash_ptr(omode) << 8));
497
498
	if (! entry->ent) {
		/* create a new one */
Michael Beck's avatar
Michael Beck committed
499
		ent = env->params->create_intrinsic(method, op, imode, omode, env->params->ctx);
500
501
502

		assert(ent && "Intrinsic creator must return an entity");
		entry->ent = ent;
Michael Beck's avatar
BugFix:    
Michael Beck committed
503
	} else {
504
		ent = entry->ent;
505
	}
506
	sym.entity_p = ent;
507
	return new_r_SymConst(env->irg, mode_P_code, sym, symconst_addr_ent);
508
}
509
510
511
512
513
514

/**
 * Translate a Div.
 *
 * Create an intrinsic Call.
 */
515
static void lower_Div(ir_node *node, ir_mode *mode)
516
{
517
518
519
520
521
522
523
524
525
526
	ir_node  *left   = get_Div_left(node);
	ir_node  *right  = get_Div_right(node);
	ir_node  *block  = get_nodes_block(node);
	dbg_info *dbgi   = get_irn_dbg_info(node);
	ir_type  *mtp    = mode_is_signed(mode) ? binop_tp_s : binop_tp_u;
	ir_mode  *opmode = get_irn_op_mode(node);
	ir_node  *addr   = get_intrinsic_address(mtp, get_irn_op(node), opmode, opmode);
	ir_node  *in[4];
	ir_node  *call;
	ir_node  *resproj;
527

528
529
530
531
532
533
534
535
536
537
538
539
540
	if (env->params->little_endian) {
		in[0] = get_lowered_low(left);
		in[1] = get_lowered_high(left);
		in[2] = get_lowered_low(right);
		in[3] = get_lowered_high(right);
	} else {
		in[0] = get_lowered_high(left);
		in[1] = get_lowered_low(left);
		in[2] = get_lowered_high(right);
		in[3] = get_lowered_low(right);
	}
	call    = new_rd_Call(dbgi, block, get_Div_mem(node), addr, 4, in, mtp);
	resproj = new_r_Proj(call, mode_T, pn_Call_T_result);
541
542
	set_irn_pinned(call, get_irn_pinned(node));

543
	foreach_out_edge_safe(node, edge) {
544
545
546
547
		ir_node *proj = get_edge_src_irn(edge);
		if (!is_Proj(proj))
			continue;

548
549
550
551
		switch (get_Proj_proj(proj)) {
		case pn_Div_M:         /* Memory result. */
			/* reroute to the call */
			set_Proj_pred(proj, call);
552
			set_Proj_proj(proj, pn_Call_M);
553
			break;
554
555
556
557
558
		case pn_Div_X_regular:
			set_Proj_pred(proj, call);
			set_Proj_proj(proj, pn_Call_X_regular);
			break;
		case pn_Div_X_except:
559
560
561
			set_Proj_pred(proj, call);
			set_Proj_proj(proj, pn_Call_X_except);
			break;
562
563
564
565
566
567
568
569
570
571
		case pn_Div_res:
			if (env->params->little_endian) {
				ir_node *res_low  = new_r_Proj(resproj, env->low_unsigned, 0);
				ir_node *res_high = new_r_Proj(resproj, mode,              1);
				ir_set_dw_lowered(proj, res_low, res_high);
			} else {
				ir_node *res_low  = new_r_Proj(resproj, env->low_unsigned, 1);
				ir_node *res_high = new_r_Proj(resproj, mode,              0);
				ir_set_dw_lowered(proj, res_low, res_high);
			}
572
573
574
			break;
		default:
			assert(0 && "unexpected Proj number");
575
		}
576
577
578
		/* mark this proj: we have handled it already, otherwise we might fall into
		 * out new nodes. */
		mark_irn_visited(proj);
579
580
	}
}
581
582
583
584
585
586

/**
 * Translate a Mod.
 *
 * Create an intrinsic Call.
 */
587
static void lower_Mod(ir_node *node, ir_mode *mode)
588
{
589
590
591
592
593
594
595
596
597
598
	ir_node  *left   = get_Mod_left(node);
	ir_node  *right  = get_Mod_right(node);
	dbg_info *dbgi   = get_irn_dbg_info(node);
	ir_node  *block  = get_nodes_block(node);
	ir_type  *mtp    = mode_is_signed(mode) ? binop_tp_s : binop_tp_u;
	ir_mode  *opmode = get_irn_op_mode(node);
	ir_node  *addr   = get_intrinsic_address(mtp, get_irn_op(node), opmode, opmode);
	ir_node  *in[4];
	ir_node  *call;
	ir_node  *resproj;
599
600
601
602
603
604
605
606
607
608
609
610
611
612

	if (env->params->little_endian) {
		in[0] = get_lowered_low(left);
		in[1] = get_lowered_high(left);
		in[2] = get_lowered_low(right);
		in[3] = get_lowered_high(right);
	} else {
		in[0] = get_lowered_high(left);
		in[1] = get_lowered_low(left);
		in[2] = get_lowered_high(right);
		in[3] = get_lowered_low(right);
	}
	call    = new_rd_Call(dbgi, block, get_Mod_mem(node), addr, 4, in, mtp);
	resproj = new_r_Proj(call, mode_T, pn_Call_T_result);
613
614
	set_irn_pinned(call, get_irn_pinned(node));

615
	foreach_out_edge_safe(node, edge) {
616
617
618
619
		ir_node *proj = get_edge_src_irn(edge);
		if (!is_Proj(proj))
			continue;

620
621
622
623
		switch (get_Proj_proj(proj)) {
		case pn_Mod_M:         /* Memory result. */
			/* reroute to the call */
			set_Proj_pred(proj, call);
624
			set_Proj_proj(proj, pn_Call_M);
625
			break;
626
627
628
629
630
		case pn_Div_X_regular:
			set_Proj_pred(proj, call);
			set_Proj_proj(proj, pn_Call_X_regular);
			break;
		case pn_Mod_X_except:
631
632
633
			set_Proj_pred(proj, call);
			set_Proj_proj(proj, pn_Call_X_except);
			break;
634
635
636
637
638
639
640
641
642
643
		case pn_Mod_res:
			if (env->params->little_endian) {
				ir_node *res_low  = new_r_Proj(resproj, env->low_unsigned, 0);
				ir_node *res_high = new_r_Proj(resproj, mode,              1);
				ir_set_dw_lowered(proj, res_low, res_high);
			} else {
				ir_node *res_low  = new_r_Proj(resproj, env->low_unsigned, 1);
				ir_node *res_high = new_r_Proj(resproj, mode,              0);
				ir_set_dw_lowered(proj, res_low, res_high);
			}
644
645
646
			break;
		default:
			assert(0 && "unexpected Proj number");
647
		}
648
649
		/* mark this proj: we have handled it already, otherwise we might fall
		 * into out new nodes. */
650
		mark_irn_visited(proj);
651
652
	}
}
653
654

/**
655
 * Translate a binop.
656
657
658
 *
 * Create an intrinsic Call.
 */
659
static void lower_binop(ir_node *node, ir_mode *mode)
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
	ir_node  *left  = get_binop_left(node);
	ir_node  *right = get_binop_right(node);
	dbg_info *dbgi  = get_irn_dbg_info(node);
	ir_node  *block = get_nodes_block(node);
	ir_graph *irg   = get_irn_irg(block);
	ir_type  *mtp   = mode_is_signed(mode) ? binop_tp_s : binop_tp_u;
	ir_node  *addr  = get_intrinsic_address(mtp, get_irn_op(node), mode, mode);
	ir_node  *in[4];
	ir_node  *call;
	ir_node  *resproj;

	if (env->params->little_endian) {
		in[0] = get_lowered_low(left);
		in[1] = get_lowered_high(left);
		in[2] = get_lowered_low(right);
		in[3] = get_lowered_high(right);
	} else {
		in[0] = get_lowered_high(left);
		in[1] = get_lowered_low(left);
		in[2] = get_lowered_high(right);
		in[3] = get_lowered_low(right);
	}
	call    = new_rd_Call(dbgi, block, get_irg_no_mem(irg), addr, 4, in, mtp);
	resproj = new_r_Proj(call, mode_T, pn_Call_T_result);
685
	set_irn_pinned(call, get_irn_pinned(node));
686
687
688
689
690
691
692
693
694
695

	if (env->params->little_endian) {
		ir_node *res_low  = new_r_Proj(resproj, env->low_unsigned, 0);
		ir_node *res_high = new_r_Proj(resproj, mode,              1);
		ir_set_dw_lowered(node, res_low, res_high);
	} else {
		ir_node *res_low  = new_r_Proj(resproj, env->low_unsigned, 1);
		ir_node *res_high = new_r_Proj(resproj, mode,              0);
		ir_set_dw_lowered(node, res_low, res_high);
	}
696
}
697

698
static ir_node *create_conv(ir_node *block, ir_node *node, ir_mode *dest_mode)
699
{
700
701
702
	if (get_irn_mode(node) == dest_mode)
		return node;
	return new_r_Conv(block, node, dest_mode);
703
}
Christian Würdig's avatar
Christian Würdig committed
704

Michael Beck's avatar
Michael Beck committed
705
/**
706
707
 * Moves node and all predecessors of node from from_bl to to_bl.
 * Does not move predecessors of Phi nodes (or block nodes).
Michael Beck's avatar
Michael Beck committed
708
 */
709
static void move(ir_node *node, ir_node *from_bl, ir_node *to_bl)
710
{
711
712
713
714
715
716
717
718
719
720
721
722
723
724
	int i, arity;

	/* move this node */
	set_nodes_block(node, to_bl);

	/* move its Projs */
	if (get_irn_mode(node) == mode_T) {
		foreach_out_edge(node, edge) {
			ir_node *proj = get_edge_src_irn(edge);
			if (!is_Proj(proj))
				continue;
			move(proj, from_bl, to_bl);
		}
	}
Michael Beck's avatar
Michael Beck committed
725

726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
	/* We must not move predecessors of Phi nodes, even if they are in
	 * from_bl. (because these are values from an earlier loop iteration
	 * which are not predecessors of node here)
	 */
	if (is_Phi(node))
		return;

	/* recursion ... */
	arity = get_irn_arity(node);
	for (i = 0; i < arity; i++) {
		ir_node *pred      = get_irn_n(node, i);
		ir_mode *pred_mode = get_irn_mode(pred);
		if (get_nodes_block(pred) == from_bl)
			move(pred, from_bl, to_bl);
		if (pred_mode == env->high_signed || pred_mode == env->high_unsigned) {
			ir_node *pred_low  = get_lowered_low(pred);
			ir_node *pred_high = get_lowered_high(pred);
			if (get_nodes_block(pred_low) == from_bl)
				move(pred_low, from_bl, to_bl);
			if (pred_high != NULL && get_nodes_block(pred_high) == from_bl)
				move(pred_high, from_bl, to_bl);
747
748
749
		}
	}
}
Michael Beck's avatar
Michael Beck committed
750
751

/**
752
753
754
755
756
 * We need a custom version of part_block_edges because during transformation
 * not all data-dependencies are explicit yet if a lowered nodes users are not
 * lowered yet.
 * We can fix this by modifying move to look for such implicit dependencies.
 * Additionally we have to keep the proj_2_block map updated
Michael Beck's avatar
Michael Beck committed
757
 */
758
static ir_node *part_block_dw(ir_node *node)
759
{
760
761
762
763
764
	ir_graph *irg        = get_irn_irg(node);
	ir_node  *old_block  = get_nodes_block(node);
	int       n_cfgpreds = get_Block_n_cfgpreds(old_block);
	ir_node **cfgpreds   = get_Block_cfgpred_arr(old_block);
	ir_node  *new_block  = new_r_Block(irg, n_cfgpreds, cfgpreds);
765

766
767
768
769
770
771
772
	/* old_block has no predecessors anymore for now */
	set_irn_in(old_block, 0, NULL);

	/* move node and its predecessors to new_block */
	move(node, old_block, new_block);

	/* move Phi nodes to new_block */
773
	foreach_out_edge_safe(old_block, edge) {
774
775
776
777
		ir_node *phi = get_edge_src_irn(edge);
		if (!is_Phi(phi))
			continue;
		set_nodes_block(phi, new_block);
778
	}
779
	return old_block;
780
}
Michael Beck's avatar
Michael Beck committed
781

782
783
784
785
786
787
typedef ir_node* (*new_rd_shr_func)(dbg_info *dbgi, ir_node *block,
                                    ir_node *left, ir_node *right,
                                    ir_mode *mode);

static void lower_shr_helper(ir_node *node, ir_mode *mode,
                             new_rd_shr_func new_rd_shrs)
788
{
789
790
791
792
793
794
795
796
797
798
799
800
801
	ir_node  *right         = get_binop_right(node);
	ir_node  *left          = get_binop_left(node);
	ir_mode  *shr_mode      = get_irn_mode(node);
	unsigned  modulo_shift  = get_mode_modulo_shift(shr_mode);
	ir_mode  *low_unsigned  = env->low_unsigned;
	unsigned  modulo_shift2 = get_mode_modulo_shift(mode);
	ir_graph *irg           = get_irn_irg(node);
	ir_node  *left_low      = get_lowered_low(left);
	ir_node  *left_high     = get_lowered_high(left);
	dbg_info *dbgi          = get_irn_dbg_info(node);
	ir_node  *lower_block;
	ir_node  *block;
	ir_node  *cnst;
802
	ir_node  *andn;
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
	ir_node  *cmp;
	ir_node  *cond;
	ir_node  *proj_true;
	ir_node  *proj_false;
	ir_node  *phi_low;
	ir_node  *phi_high;
	ir_node  *lower_in[2];
	ir_node  *phi_low_in[2];
	ir_node  *phi_high_in[2];

	/* this version is optimized for modulo shift architectures
	 * (and can't handle anything else) */
	if (modulo_shift != get_mode_size_bits(shr_mode)
			|| modulo_shift2<<1 != modulo_shift) {
		panic("Shr lowering only implemented for modulo shift shr operations");
	}
	if (!is_po2(modulo_shift) || !is_po2(modulo_shift2)) {
		panic("Shr lowering only implemented for power-of-2 modes");
	}
	/* without 2-complement the -x instead of (bit_width-x) trick won't work */
	if (get_mode_arithmetic(shr_mode) != irma_twos_complement) {
		panic("Shr lowering only implemented for two-complement modes");
	}

827
828
	block = get_nodes_block(node);

829
830
831
832
833
834
	/* if the right operand is a 64bit value, we're only interested in the
	 * lower word */
	if (get_irn_mode(right) == env->high_unsigned) {
		right = get_lowered_low(right);
	} else {
		/* shift should never have signed mode on the right */
Michael Beck's avatar
Michael Beck committed
835
		assert(get_irn_mode(right) != env->high_signed);
836
837
		right = create_conv(block, right, low_unsigned);
	}
Michael Beck's avatar
Michael Beck committed
838

839
840
841
842
	lower_block = part_block_dw(node);
	env->flags |= CF_CHANGED;
	block = get_nodes_block(node);

Michael Beck's avatar
Michael Beck committed
843
	/* add a Cmp to test if highest bit is set <=> whether we shift more
844
845
	 * than half the word width */
	cnst       = new_r_Const_long(irg, low_unsigned, modulo_shift2);
846
	andn       = new_r_And(block, right, cnst, low_unsigned);
847
	cnst       = new_r_Const(irg, get_mode_null(low_unsigned));
848
	cmp        = new_rd_Cmp(dbgi, block, andn, cnst, ir_relation_equal);
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
	cond       = new_rd_Cond(dbgi, block, cmp);
	proj_true  = new_r_Proj(cond, mode_X, pn_Cond_true);
	proj_false = new_r_Proj(cond, mode_X, pn_Cond_false);

	/* the true block => shift_width < 1word */
	{
		/* In theory the low value (for 64bit shifts) is:
		 *    Or(High << (32-x)), Low >> x)
		 * In practice High << 32-x will fail when x is zero (since we have
		 * modulo shift and 32 will be 0). So instead we use:
		 *    Or(High<<1<<~x, Low >> x)
		 */
		ir_node *in[1]        = { proj_true };
		ir_node *block_true   = new_r_Block(irg, ARRAY_SIZE(in), in);
		ir_node *res_high     = new_rd_shrs(dbgi, block_true, left_high,
		                                    right, mode);
		ir_node *shift_low    = new_rd_Shr(dbgi, block_true, left_low, right,
		                                   low_unsigned);
		ir_node *not_shiftval = new_rd_Not(dbgi, block_true, right,
		                                   low_unsigned);
		ir_node *conv         = create_conv(block_true, left_high,
		                                    low_unsigned);
		ir_node *one          = new_r_Const(irg, get_mode_one(low_unsigned));
		ir_node *carry0       = new_rd_Shl(dbgi, block_true, conv, one,
		                                   low_unsigned);
		ir_node *carry1       = new_rd_Shl(dbgi, block_true, carry0,
		                                   not_shiftval, low_unsigned);
		ir_node *res_low      = new_rd_Or(dbgi, block_true, shift_low, carry1,
		                                  low_unsigned);
		lower_in[0]           = new_r_Jmp(block_true);
		phi_low_in[0]         = res_low;
		phi_high_in[0]        = res_high;
	}

	/* false block => shift_width > 1word */
	{
		ir_node *in[1]       = { proj_false };
		ir_node *block_false = new_r_Block(irg, ARRAY_SIZE(in), in);
		ir_node *conv        = create_conv(block_false, left_high, low_unsigned);
		ir_node *res_low     = new_rd_shrs(dbgi, block_false, conv, right,
		                                   low_unsigned);
		int      cnsti       = modulo_shift2-1;
891
		ir_node *cnst2       = new_r_Const_long(irg, low_unsigned, cnsti);
892
893
		ir_node *res_high;
		if (new_rd_shrs == new_rd_Shrs) {
894
			res_high = new_rd_shrs(dbgi, block_false, left_high, cnst2, mode);
895
896
		} else {
			res_high = new_r_Const(irg, get_mode_null(mode));
897
		}
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
		lower_in[1]          = new_r_Jmp(block_false);
		phi_low_in[1]        = res_low;
		phi_high_in[1]       = res_high;
	}

	/* patch lower block */
	set_irn_in(lower_block, ARRAY_SIZE(lower_in), lower_in);
	phi_low  = new_r_Phi(lower_block, ARRAY_SIZE(phi_low_in), phi_low_in,
	                     low_unsigned);
	phi_high = new_r_Phi(lower_block, ARRAY_SIZE(phi_high_in), phi_high_in,
	                     mode);
	ir_set_dw_lowered(node, phi_low, phi_high);
}

static void lower_Shr(ir_node *node, ir_mode *mode)
{
	lower_shr_helper(node, mode, new_rd_Shr);
}

static void lower_Shrs(ir_node *node, ir_mode *mode)
{
	lower_shr_helper(node, mode, new_rd_Shrs);
}

static void lower_Shl(ir_node *node, ir_mode *mode)
{
	ir_node  *right         = get_binop_right(node);
	ir_node  *left          = get_binop_left(node);
	ir_mode  *shr_mode      = get_irn_mode(node);
	unsigned  modulo_shift  = get_mode_modulo_shift(shr_mode);
	ir_mode  *low_unsigned  = env->low_unsigned;
	unsigned  modulo_shift2 = get_mode_modulo_shift(mode);
	ir_graph *irg           = get_irn_irg(node);
	ir_node  *left_low      = get_lowered_low(left);
	ir_node  *left_high     = get_lowered_high(left);
	dbg_info *dbgi          = get_irn_dbg_info(node);
	ir_node  *lower_block   = get_nodes_block(node);
	ir_node  *block;
	ir_node  *cnst;
937
	ir_node  *andn;
938
939
940
941
942
943
944
945
946
947
948
949
950
951
	ir_node  *cmp;
	ir_node  *cond;
	ir_node  *proj_true;
	ir_node  *proj_false;
	ir_node  *phi_low;
	ir_node  *phi_high;
	ir_node  *lower_in[2];
	ir_node  *phi_low_in[2];
	ir_node  *phi_high_in[2];

	/* this version is optimized for modulo shift architectures
	 * (and can't handle anything else) */
	if (modulo_shift != get_mode_size_bits(shr_mode)
			|| modulo_shift2<<1 != modulo_shift) {
yb9976's avatar
yb9976 committed
952
		panic("Shl lowering only implemented for modulo shift shl operations");
953
954
	}
	if (!is_po2(modulo_shift) || !is_po2(modulo_shift2)) {
yb9976's avatar
yb9976 committed
955
		panic("Shl lowering only implemented for power-of-2 modes");
956
957
958
	}
	/* without 2-complement the -x instead of (bit_width-x) trick won't work */
	if (get_mode_arithmetic(shr_mode) != irma_twos_complement) {
yb9976's avatar
yb9976 committed
959
		panic("Shl lowering only implemented for two-complement modes");
960
961
962
963
964
965
966
967
968
969
	}

	/* if the right operand is a 64bit value, we're only interested in the
	 * lower word */
	if (get_irn_mode(right) == env->high_unsigned) {
		right = get_lowered_low(right);
	} else {
		/* shift should never have signed mode on the right */
		assert(get_irn_mode(right) != env->high_signed);
		right = create_conv(lower_block, right, low_unsigned);
970
	}
971
972
973
974
975

	part_block_dw(node);
	env->flags |= CF_CHANGED;
	block = get_nodes_block(node);

Michael Beck's avatar
Michael Beck committed
976
	/* add a Cmp to test if highest bit is set <=> whether we shift more
977
978
	 * than half the word width */
	cnst       = new_r_Const_long(irg, low_unsigned, modulo_shift2);
979
	andn       = new_r_And(block, right, cnst, low_unsigned);
980
	cnst       = new_r_Const(irg, get_mode_null(low_unsigned));
981
	cmp        = new_rd_Cmp(dbgi, block, andn, cnst, ir_relation_equal);
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
	cond       = new_rd_Cond(dbgi, block, cmp);
	proj_true  = new_r_Proj(cond, mode_X, pn_Cond_true);
	proj_false = new_r_Proj(cond, mode_X, pn_Cond_false);

	/* the true block => shift_width < 1word */
	{
		ir_node *in[1]        = { proj_true };
		ir_node *block_true   = new_r_Block(irg, ARRAY_SIZE(in), in);

		ir_node *res_low      = new_rd_Shl(dbgi, block_true, left_low,
		                                   right, low_unsigned);
		ir_node *shift_high   = new_rd_Shl(dbgi, block_true, left_high, right,
		                                   mode);
		ir_node *not_shiftval = new_rd_Not(dbgi, block_true, right,
		                                   low_unsigned);
		ir_node *conv         = create_conv(block_true, left_low, mode);
998
		ir_node *one          = new_r_Const(irg, get_mode_one(low_unsigned));
999
1000
		ir_node *carry0       = new_rd_Shr(dbgi, block_true, conv, one, mode);
		ir_node *carry1       = new_rd_Shr(dbgi, block_true, carry0,
For faster browsing, not all history is shown. View entire blame