ldstopt.c 12.3 KB
Newer Older
Michael Beck's avatar
Michael Beck committed
1
2
3
4
5
6
7
/*
 * Project:     libFIRM
 * File name:   ir/opt/ldstopt.c
 * Purpose:     load store optimizations
 * Author:
 * Created:
 * CVS-ID:      $Id$
Götz Lindenmaier's avatar
Götz Lindenmaier committed
8
 * Copyright:   (c) 1998-2004 Universitt Karlsruhe
Michael Beck's avatar
Michael Beck committed
9
10
11
12
13
14
15
16
17
18
19
20
21
22
 * Licence:     This file protected by GPL -  GNU GENERAL PUBLIC LICENSE.
 */
# 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"
Michael Beck's avatar
Michael Beck committed
23
# include "array.h"
Michael Beck's avatar
Michael Beck committed
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
# include "firmstat.h"

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

#define MAX_PROJ	IMAX(pn_Load_max, pn_Store_max)

/**
 * walker environment
 */
typedef struct _walk_env_t {
  struct obstack obst;		/**< list of all stores */
  int changes;
} walk_env_t;

/**
 * a Load/Store info
 */
typedef struct _ldst_info_t {
  ir_node *projs[MAX_PROJ];	/**< list of Proj's of this node */
Michael Beck's avatar
Michael Beck committed
44
45
  ir_node *exc_block;           /**< the exception block if available */
  int     exc_idx;              /**< predecessor index in the exception block */
Michael Beck's avatar
Michael Beck committed
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
} ldst_info_t;

/**
 * walker, clears all links first
 */
static void init_links(ir_node *n, void *env)
{
  set_irn_link(n, NULL);
}

/**
 * get the Load/Store info of a node
 */
static ldst_info_t *get_info(ir_node *node, walk_env_t *env)
{
  ldst_info_t *info = get_irn_link(node);

  if (! info) {
    info = obstack_alloc(&env->obst, sizeof(*info));

    memset(info, 0, sizeof(*info));

    set_irn_link(node, info);
  }
  return info;
}

/**
Michael Beck's avatar
Michael Beck committed
74
 * update the projection info for a Load/Store
Michael Beck's avatar
Michael Beck committed
75
 */
Michael Beck's avatar
Michael Beck committed
76
static int update_projs(ldst_info_t *info, ir_node *proj)
Michael Beck's avatar
Michael Beck committed
77
78
79
80
81
82
83
84
{
  long nr = get_Proj_proj(proj);

  assert(0 <= nr && nr <= MAX_PROJ && "Wrong proj from LoadStore");

  if (info->projs[nr]) {
    /* there is already one, do CSE */
    exchange(proj, info->projs[nr]);
Michael Beck's avatar
Michael Beck committed
85
    return 1;
Michael Beck's avatar
Michael Beck committed
86
  }
Michael Beck's avatar
Michael Beck committed
87
  else {
Michael Beck's avatar
Michael Beck committed
88
    info->projs[nr] = proj;
Michael Beck's avatar
Michael Beck committed
89
90
91
92
93
94
95
96
97
98
99
100
101
102
    return 0;
  }
}

/**
 * update the exception block info for a Load/Store
 */
static int update_exc(ldst_info_t *info, ir_node *block, int pos)
{
  assert(info->exc_block == NULL && "more than one exception block found");

  info->exc_block = block;
  info->exc_idx   = pos;
  return 0;
Michael Beck's avatar
Michael Beck committed
103
104
105
106
107
}

/**
 * walker, collects all Load/Store/Proj nodes
 */
Michael Beck's avatar
Michael Beck committed
108
static void collect_nodes(ir_node *node, void *env)
Michael Beck's avatar
Michael Beck committed
109
110
111
112
113
{
  ir_node     *pred;
  ldst_info_t *info;
  walk_env_t  *wenv = env;

Michael Beck's avatar
Michael Beck committed
114
115
  if (get_irn_op(node) == op_Proj) {
    pred = get_Proj_pred(node);
Michael Beck's avatar
Michael Beck committed
116
117
118
119

    if (get_irn_op(pred) == op_Load || get_irn_op(pred) == op_Store) {
      info = get_info(pred, wenv);

Michael Beck's avatar
Michael Beck committed
120
121
122
123
124
125
126
127
128
129
130
131
132
133
      wenv->changes |= update_projs(info, node);
    }
  }
  else if (get_irn_op(node) == op_Block) { /* check, if it's an exception block */
    int i, n;

    for (i = 0, n = get_Block_n_cfgpreds(node); i < n; ++i) {
      pred = skip_Proj(get_Block_cfgpred(node, i));

      if (get_irn_op(pred) == op_Load || get_irn_op(pred) == op_Store) {
        info = get_info(pred, wenv);

        wenv->changes |= update_exc(info, node, i);
      }
Michael Beck's avatar
Michael Beck committed
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
    }
  }
}

/**
 * optimize a Load
 */
static int optimize_load(ir_node *load)
{
  ldst_info_t *info = get_irn_link(load);
  ir_mode *load_mode = get_Load_mode(load);
  ir_node *pred, *mem, *ptr;
  int res = 1;

  if (get_Load_volatility(load) == volatility_is_volatile)
    return 0;

  /*
   * 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 a = *(type1 *)p; b = *(type2 *)p ...
   */

  ptr  = get_Load_ptr(load);
  mem  = get_Load_mem(load);
  pred = skip_Proj(mem);

  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);
    res = 1;
  }
  else if (get_irn_op(pred) == op_Store && get_Store_ptr(pred) == ptr &&
	   get_irn_mode(get_Store_value(pred)) == load_mode) {
    /*
     * a load immediately after a store -- a read after write.
     * We may remove the Load, if it 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.
     */
    if (! info->projs[pn_Load_X_except] || get_nodes_block(load) == get_nodes_block(pred)) {
      exchange( info->projs[pn_Load_res], get_Store_value(pred) );
Götz Lindenmaier's avatar
Götz Lindenmaier committed
177
178
      if (info->projs[pn_Load_M])
	exchange(info->projs[pn_Load_M], mem);
Michael Beck's avatar
Michael Beck committed
179
180
181
182
183
184
185
186
187
188
189
190

      /* no exception */
      if (info->projs[pn_Load_X_except])
	exchange( info->projs[pn_Load_X_except], new_Bad());
      res = 1;
    }
  }
  else if (get_irn_op(pred) == op_Load && get_Load_ptr(pred) == ptr &&
	   get_Load_mode(pred) == load_mode) {
    /*
     * a load immediately after a load -- a read after read.
     * We may remove the second Load, if it does not have an exception handler
Götz Lindenmaier's avatar
Götz Lindenmaier committed
191
     * OR they are in the same block. In the later case the Load cannot
Michael Beck's avatar
Michael Beck committed
192
193
194
195
196
197
198
199
     * throw an exception when the previous Load was quiet.
     */
    if (! info->projs[pn_Load_X_except] || get_nodes_block(load) == get_nodes_block(pred)) {
      ldst_info_t *pred_info = get_irn_link(pred);

      if (pred_info->projs[pn_Load_res]) {
	/* we need a data proj from the previous load for this optimization */
	exchange( info->projs[pn_Load_res], pred_info->projs[pn_Load_res] );
Götz Lindenmaier's avatar
Götz Lindenmaier committed
200
201
	if (info->projs[pn_Load_M])
	  exchange(info->projs[pn_Load_M], mem);
Michael Beck's avatar
Michael Beck committed
202
203
      }
      else {
Götz Lindenmaier's avatar
Götz Lindenmaier committed
204
205
206
207
	if (info->projs[pn_Load_res]) {
	  set_Proj_pred(info->projs[pn_Load_res], pred);
	  set_nodes_block(info->projs[pn_Load_res], get_nodes_block(pred));
	}
Götz Lindenmaier's avatar
Götz Lindenmaier committed
208
209
210
211
212
	if (info->projs[pn_Load_M]) {
	  /* Actually, this if should not be necessary.  Construct the Loads
	     properly!!! */
	  exchange(info->projs[pn_Load_M], mem);
	}
Michael Beck's avatar
Michael Beck committed
213
214
215
216
      }

      /* no exception */
      if (info->projs[pn_Load_X_except])
Götz Lindenmaier's avatar
Götz Lindenmaier committed
217
	exchange(info->projs[pn_Load_X_except], new_Bad());
Michael Beck's avatar
Michael Beck committed
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
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281

      res = 1;
    }
  }
  return res;
}

/**
 * optimize a Store
 */
static int optimize_store(ir_node *store)
{
  ldst_info_t *info = get_irn_link(store);
  ir_node *pred, *mem, *ptr, *value, *block;
  ir_mode *mode;
  ldst_info_t *pred_info;
  int res = 0;

  if (get_Store_volatility(store) == volatility_is_volatile)
    return 0;

  /*
   * 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 *(type1 *)p = a; *(type2 *)p = b ...
   */

  block = get_nodes_block(store);
  ptr   = get_Store_ptr(store);
  mem   = get_Store_mem(store);
  value = get_Store_value(store);
  pred  = skip_Proj(mem);
  mode  = get_irn_mode(value);

  pred_info = get_irn_link(pred);

  if (get_irn_op(pred) == op_Store && get_Store_ptr(pred) == ptr &&
      get_nodes_block(pred) == block && get_irn_mode(get_Store_value(pred)) == mode) {
    /*
     * a store immediately 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]) {
      exchange( pred_info->projs[pn_Store_M], get_Store_mem(pred) );
      res = 1;
    }
  }
  else if (get_irn_op(pred) == op_Load && get_Load_ptr(pred) == ptr &&
	   value == pred_info->projs[pn_Load_res]) {
    /*
     * a store a value immediately 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]) {
      exchange( info->projs[pn_Store_M], mem );
      res = 1;
    }
  }
  return res;
}

Michael Beck's avatar
Michael Beck committed
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
/**
 * walker, optimizes Phi after Stores:
 * Does the following optimization:
 *
 *   val1   val2   val3          val1  val2  val3
 *    |      |      |               \    |    /
 *   Str    Str    Str               \   |   /
 *      \    |    /                     Phi
 *       \   |   /                       |
 *        \  |  /                       Str
 *          Phi
 *
 * This removes teh number of stores and allows for predicated execution.
 * Moves Stores back to the end of a function which may be bad
 *
 * Note: that even works, if one of the Stores is already in our current block
 */
static int optimize_phi(ir_node *phi)
{
  int i, n;
  ir_node *store, *ptr, *block, *phiM, *phiD, *exc;
  ir_mode *mode;
  ir_node **inM, **inD;
  int *idx;
  dbg_info *db = NULL;
  ldst_info_t *info;

  /* 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));
  if (get_irn_op(store) != op_Store)
    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 (mode != get_irn_mode(get_Store_value(pred)) || ptr != get_Store_ptr(pred))
      return 0;

    info = get_irn_link(pred);

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

  /*
   * ok, when we are here, we found all predecessors of a Phi that
   * are Stores to the same address. 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 |             \   |   /
   *      \     |     /                     Phi
   *       \    |    /                       |
   *        \   |   /                       Str
   *           Phi
   *
   * Note: that even works, if one of the Stores is already in our current block
   */

  /* first step: collect all inputs */
  NEW_ARR_A(ir_node *, inM, n);
  NEW_ARR_A(ir_node *, inD, n);
  NEW_ARR_A(int, idx, n);

  for (i = 0; i < n; ++i) {
    ir_node *pred = skip_Proj(get_Phi_pred(phi, i));
    info = get_irn_link(pred);

    inM[i] = get_Store_mem(pred);
    inD[i] = get_Store_value(pred);
    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);

  /* 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);

    for (i = 0; i < n; ++i) {
      set_Block_cfgpred(exc, idx[i], projX);
    }

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

  /* sixt step: replace old Phi */
  exchange(phi, new_rd_Proj(NULL, current_ir_graph, block, store, mode_M, pn_Store_M));

  return 1;
}

Michael Beck's avatar
Michael Beck committed
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
/**
 * walker, collects all Load/Store/Proj nodes
 */
static void do_load_store_optimize(ir_node *n, void *env)
{
  walk_env_t *wenv = env;

  switch (get_irn_opcode(n)) {

  case iro_Load:
    wenv->changes |= optimize_load(n);
    break;

  case iro_Store:
    wenv->changes |= optimize_store(n);
    break;

Michael Beck's avatar
Michael Beck committed
420
421
422
  case iro_Phi:
    wenv->changes |= optimize_phi(n);

Michael Beck's avatar
Michael Beck committed
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
  default:
    ;
  }
}

/*
 * do the load store optimization
 */
void optimize_load_store(ir_graph *irg)
{
  walk_env_t env;

  assert(get_irg_phase_state(irg) != phase_building);

  if (!get_opt_redundant_LoadStore())
    return;

  obstack_init(&env.obst);
  env.changes = 0;

  /* init the links, then collect Loads/Stores/Proj's in lists */
  irg_walk_graph(irg, init_links, collect_nodes, &env);

  /* now we have collected enough information, optimize */
  irg_walk_graph(irg, NULL, do_load_store_optimize, &env);

  obstack_free(&env.obst, NULL);

  /* Handle graph state */
  if (env.changes) {
    if (get_irg_outs_state(current_ir_graph) == outs_consistent)
      set_irg_outs_inconsistent(current_ir_graph);

    /* is this really needed: Yes, as exception block may get bad but this might be tested */
    if (get_irg_dom_state(current_ir_graph) == dom_consistent)
      set_irg_dom_inconsistent(current_ir_graph);
  }
}