return.c 9.2 KB
Newer Older
Christian Würdig's avatar
Christian Würdig committed
1
/*
Michael Beck's avatar
Michael Beck committed
2
 * Copyright (C) 1995-2008 University of Karlsruhe.  All right reserved.
Christian Würdig's avatar
Christian Würdig committed
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
 *
 * This file is part of libFirm.
 *
 * This file may be distributed and/or modified under the terms of the
 * GNU General Public License version 2 as published by the Free Software
 * Foundation and appearing in the file LICENSE.GPL included in the
 * packaging of this file.
 *
 * Licensees holding valid libFirm Professional Edition licenses may use
 * this file in accordance with the libFirm Commercial License.
 * Agreement provided with the Software.
 *
 * This file is provided AS IS with NO WARRANTY OF ANY KIND, INCLUDING THE
 * WARRANTY OF DESIGN, MERCHANTABILITY AND FITNESS FOR A PARTICULAR
 * PURPOSE.
 */

Michael Beck's avatar
Michael Beck committed
20
21
22
23
24
/**
 * @file
 * @brief   Normalize returns.
 * @author  Michael Beck
 * @version $Id$
Michael Beck's avatar
Michael Beck committed
25
26
27
28
29
 */
#ifdef HAVE_CONFIG_H
#include "config.h"
#endif

30
#include "iroptimize.h"
Michael Beck's avatar
Michael Beck committed
31
32
33
34
35
36
37
38
39
40
41
42
43
44
#include "irgraph_t.h"
#include "ircons_t.h"
#include "irnode_t.h"
#include "irgmod.h"

#define set_bit(n)      (returns[(n) >> 3] |= 1 << ((n) & 7))
#define get_bit(n)      (returns[(n) >> 3] & (1 << ((n) & 7)))

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

/*
 * Normalize the Returns of a graph by creating a new End block
 * with One Return(Phi).
Michael Beck's avatar
Michael Beck committed
45
 * This is the preferred input for the if-conversion.
Michael Beck's avatar
Michael Beck committed
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
 *
 * In pseudocode, it means:
 *
 * if (a)
 *   return b;
 * else
 *   return c;
 *
 * is transformed into
 *
 * if (a)
 *   res = b;
 * else
 *   res = c;
 * return res;
 */
Michael Beck's avatar
Michael Beck committed
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
void normalize_one_return(ir_graph *irg) {
	ir_node *endbl = get_irg_end_block(irg);
	int i, j, k, n, last_idx, n_rets, n_ret_vals = -1;
	unsigned char *returns;
	ir_node **in, **retvals, **endbl_in;

	ir_node *block;

	/* look, if we have more than one return */
	n = get_Block_n_cfgpreds(endbl);
	if (n <= 0) {
		/* The end block has no predecessors, we have an endless
		   loop. In that case, no returns exists. */
		return;
	}

	returns = alloca((n + 7) >> 3);
	memset(returns, 0, (n + 7) >> 3);

	for (n_rets = i = 0; i < n; ++i) {
		ir_node *node = get_Block_cfgpred(endbl, i);

		if (is_Return(node)) {
			++n_rets;

			set_bit(i);

			if (n_ret_vals < 0)
				n_ret_vals = get_irn_arity(node);
		}
	}

	/* there should be at least one Return node in Firm */
	if (n_rets <= 1)
		return;

	in       = alloca(sizeof(*in)       * IMAX(n_rets, n_ret_vals));
	retvals  = alloca(sizeof(*retvals)  * n_rets * n_ret_vals);
	endbl_in = alloca(sizeof(*endbl_in) * n);

	last_idx = 0;
	for (j = i = 0; i < n; ++i) {
		ir_node *ret = get_Block_cfgpred(endbl, i);

		if (get_bit(i)) {
			ir_node *block = get_nodes_block(ret);

			/* create a new Jmp for every Ret and place the in in */
			in[j] = new_r_Jmp(irg, block);

			/* save the return values and shuffle them */
			for (k = 0; k < n_ret_vals; ++k)
				retvals[j + k*n_rets] = get_irn_n(ret, k);

			++j;
		} else
			endbl_in[last_idx++] = ret;
	}

	/* ok, create a new block with all created in's */
	block = new_r_Block(irg, n_rets, in);

	/* now create the Phi nodes */
	for (j = i = 0; i < n_ret_vals; ++i, j += n_rets) {
		int k;
		ir_node *first;
		/* the return values are already shuffled */

		/* Beware: normally the Phi constructor automatically replaces a Phi(a,...a) into a
		   but NOT, if a is Unknown. Here, we known that this case can be optimize also,
		   so do it here */
		first = retvals[j + 0];
		for (k = 1; k < n_rets; ++k) {
			if (retvals[j + k] != first) {
				first = NULL;
				break;
			}
		}
		if (first)
			in[i] = first;
		else
			in[i] = new_r_Phi(irg, block, n_rets, &retvals[j], get_irn_mode(retvals[j]));
	}

	endbl_in[last_idx++] = new_r_Return(irg, block, in[0], n_ret_vals-1, &in[1]);

	set_irn_in(endbl, last_idx, endbl_in);

	/* invalidate analysis information:
	 * a new Block was added, so dominator, outs and loop are inconsistent,
	 * trouts and callee-state should be still valid
	 */
	set_irg_doms_inconsistent(irg);
	set_irg_outs_inconsistent(irg);
	set_irg_extblk_inconsistent(irg);
Michael Beck's avatar
Michael Beck committed
157
	set_irg_loopinfo_inconsistent(irg);
Michael Beck's avatar
Michael Beck committed
158
159
160
}

/**
Michael Beck's avatar
Michael Beck committed
161
 * Check, whether a Return can be moved on block upwards.
Michael Beck's avatar
Michael Beck committed
162
163
164
165
166
167
168
169
170
171
172
173
 *
 * In a block with a Return, all live nodes must be linked
 * with the Return, otherwise they are dead (because the Return leaves
 * the graph, so no more users of the other nodes can exists.
 *
 * We can move a Return, if it's predecessors are Phi nodes or
 * comes from another block. In the later case, it is always possible
 * to move the Return one block up, because the predecessor block must
 * dominate the Return block (SSA) and then it dominates the predecessor
 * block of the Return block as well.
 *
 * All predecessors of the Return block must be Jmp's of course, or we
174
 * cannot move it up, so we add blocks if needed.
Michael Beck's avatar
Michael Beck committed
175
 */
Michael Beck's avatar
Michael Beck committed
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
static int can_move_ret(ir_node *ret) {
	ir_node *retbl = get_nodes_block(ret);
	int i, n = get_irn_arity(ret);

	for (i = 0; i < n; ++i) {
		ir_node *pred = get_irn_n(ret, i);

		if (! is_Phi(pred) && retbl == get_nodes_block(pred)) {
			/* first condition failed, found a non-Phi predecessor
			 * then is in the Return block */
			return 0;
		}
	}

	/* check, that predecessors are Jmps */
	n = get_Block_n_cfgpreds(retbl);
192
193
	if (n <= 1)
		return 0;
194
195
196
197
198
199
200
201
202
203
204
205
	for (i = 0; i < n; ++i) {
		ir_node *pred = get_Block_cfgpred(retbl, i);

		pred = skip_Tuple(pred);
		if (! is_Jmp(pred) && !is_Bad(pred)) {
			/* simply place a new block here */
			ir_graph *irg  = get_irn_irg(retbl);
			ir_node *block = new_r_Block(irg, 1, &pred);
			ir_node *jmp   = new_r_Jmp(irg, block);
			set_Block_cfgpred(retbl, i, jmp);
		}
	}
206
	return 1;
Michael Beck's avatar
Michael Beck committed
207
208
209
210
211
}

/*
 * Normalize the Returns of a graph by moving
 * the Returns upwards as much as possible.
Michael Beck's avatar
Michael Beck committed
212
 * This might be preferred for code generation.
Michael Beck's avatar
Michael Beck committed
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
 *
 * In pseudocode, it means:
 *
 * if (a)
 *   res = b;
 * else
 *   res = c;
 * return res;
 *
 * is transformed into
 *
 * if (a)
 *   return b;
 * else
 *   return c;
 */
Michael Beck's avatar
Michael Beck committed
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
void normalize_n_returns(ir_graph *irg) {
	int i, j, n, n_rets, n_finals, n_ret_vals;
	ir_node *list  = NULL;
	ir_node *final = NULL;
	ir_node **in;
	ir_node *endbl = get_irg_end_block(irg);
	ir_node *end;

	/*
	 * First, link all returns:
	 * These must be predecessors of the endblock.
	 * Place Returns that can be moved on list, all others
	 * on final.
	 */
	n = get_Block_n_cfgpreds(endbl);
	for (n_finals = n_rets = i = 0; i < n; ++i) {
		ir_node *ret = get_Block_cfgpred(endbl, i);

		if (is_Return(ret) && can_move_ret(ret)) {
			/*
Michael Beck's avatar
Michael Beck committed
249
250
251
			 * Ok, all conditions met, we can move this Return, put it
			 * on our work list.
			 */
Michael Beck's avatar
Michael Beck committed
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
			set_irn_link(ret, list);
			list = ret;
			++n_rets;
		} else {
			/* Put all nodes that are not changed on the final list. */
			set_irn_link(ret, final);
			final = ret;
			++n_finals;
		}
	}

	if (n_rets <= 0)
		return;

	/*
	 * Now move the Returns upwards. We move always one block up (and create n
	 * new Returns), than we check if a newly created Return can be moved even further.
	 * If yes, we simply add it to our work list, else to the final list.
	 */
	end        = get_irg_end(irg);
	n_ret_vals = get_irn_arity(list);
	in         = alloca(sizeof(*in) * n_ret_vals);
	while (list) {
		ir_node *ret   = list;
		ir_node *block = get_nodes_block(ret);
		ir_node *phiM;

		list = get_irn_link(ret);
		--n_rets;

		n = get_Block_n_cfgpreds(block);
		for (i = 0; i < n; ++i) {
			ir_node *jmp = get_Block_cfgpred(block, i);
			ir_node *new_bl, *new_ret;

287
			if (is_Bad(jmp))
Michael Beck's avatar
Michael Beck committed
288
				continue;
289
			assert(is_Jmp(jmp));
Michael Beck's avatar
Michael Beck committed
290
291
292

			new_bl = get_nodes_block(jmp);

Michael Beck's avatar
Michael Beck committed
293
			/* create the in-array for the new Return */
Michael Beck's avatar
Michael Beck committed
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
			for (j = 0; j < n_ret_vals; ++j) {
				ir_node *pred = get_irn_n(ret, j);

				in[j] = (is_Phi(pred) && get_nodes_block(pred) == block) ? get_Phi_pred(pred, i) : pred;
			}

			new_ret = new_r_Return(irg, new_bl, in[0], n_ret_vals - 1, &in[1]);

			if (! is_Bad(new_ret)) {
				/*
				 * The newly created node might be bad, if we
				 * create it in a block with only Bad predecessors.
				 * In that case ignore this block.
				 *
				 * We could even kill the jmp then ...
				 */
				if (can_move_ret(new_ret)) {
					set_irn_link(new_ret, list);
					list = new_ret;
					++n_rets;
Michael Beck's avatar
Michael Beck committed
314
				} else {
Michael Beck's avatar
Michael Beck committed
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
					set_irn_link(new_ret, final);
					final = new_ret;
					++n_finals;
				}
			}

			/* remove the Jmp, we have placed a Return here */
			exchange(jmp, new_r_Bad(irg));
		}

		/*
		 * if the memory of the old Return is a PhiM, remove it
		 * from the keep-alives, or it will keep the block which
		 * will crash the dominator algorithm.
		 */
		phiM = get_Return_mem(ret);
		if (is_Phi(phiM)) {
			n = get_End_n_keepalives(end);
			for (i = 0; i < n; ++i) {
				if (get_End_keepalive(end, i) == phiM) {
					set_End_keepalive(end, i, new_r_Bad(irg));
					break;
				}
			}
		}
	}

	/*
	 * Last step: Create a new endblock, with all nodes on the final
	 * list as predecessors.
	 */
	in = alloca(sizeof(*in) * n_finals);

	for (i = 0; final; ++i, final = get_irn_link(final))
		in[i] = final;

	exchange(endbl, new_r_Block(irg, n_finals, in));

	/* the end block is not automatically skipped, so do it here */
	set_irg_end_block(irg, skip_Id(get_irg_end_block(irg)));

	/* Invalidate analysis information:
	 * Blocks become dead and new Returns were deleted, so dominator, outs and loop are inconsistent,
	 * trouts and callee-state should be still valid
	 */
	set_irg_doms_inconsistent(irg);
	set_irg_extblk_inconsistent(irg);  /* may not be needed */
	set_irg_outs_inconsistent(irg);
Michael Beck's avatar
Michael Beck committed
363
	set_irg_loopinfo_inconsistent(current_ir_graph);
Michael Beck's avatar
Michael Beck committed
364
}