nsz Git - libfirm/blob - ir/opt/boolopt.c

   1 /*
   2  * Copyright (C) 1995-2008 University of Karlsruhe.  All right reserved.
   3  *
   4  * This file is part of libFirm.
   5  *
   6  * This file may be distributed and/or modified under the terms of the
   7  * GNU General Public License version 2 as published by the Free Software
   8  * Foundation and appearing in the file LICENSE.GPL included in the
   9  * packaging of this file.
  10  *
  11  * Licensees holding valid libFirm Professional Edition licenses may use
  12  * this file in accordance with the libFirm Commercial License.
  13  * Agreement provided with the Software.
  14  *
  15  * This file is provided AS IS with NO WARRANTY OF ANY KIND, INCLUDING THE
  16  * WARRANTY OF DESIGN, MERCHANTABILITY AND FITNESS FOR A PARTICULAR
  17  * PURPOSE.
  18  */
  19
  20 /**
  21  * @file
  22  * @brief   boolean condition/control flow optimizations
  23  * @author  Matthias Braun, Christoph Mallon, Michael Beck
  24  * @version $Id: cfopt.c 22579 2008-10-07 14:54:04Z beck $
  25  */
  26 #include "config.h"
  27
  28 #include <assert.h>
  29 #include <string.h>
  30
  31 #include "adt/obst.h"
  32 #include "../adt/array_t.h"
  33 #include "iroptimize.h"
  34 #include "ircons.h"
  35 #include "irgmod.h"
  36 #include "irgwalk.h"
  37 #include "irprintf.h"
  38 #include "irnode_t.h"
  39 #include "tv.h"
  40 #include "irpass.h"
  41 #include "debug.h"
  42
  43 /** Describes a pair of relative conditions lo < hi, lo pnc_lo x, hi pnc_hi x */
  44 typedef struct cond_pair {
  45         ir_node *cmp_lo;  /**< The lo compare node. */
  46         ir_node *cmp_hi;  /**< The hi compare node. */
  47         pn_Cmp   pnc_lo;  /**< The lo relation node. */
  48         pn_Cmp   pnc_hi;  /**< The hi relation node. */
  49         ir_node *proj_lo; /**< The mode_b result proj of cmp_lo. */
  50         ir_node *proj_hi; /**< The mode_b result proj of cmp_hi. */
  51         tarval  *tv_lo;   /**< The tarval of cmp_lo node. */
  52         tarval  *tv_hi;   /**< The tarval of cmp_hi node. */
  53         ir_mode *lo_mode; /**< The mode of the cmp_lo operands. */
  54 } cond_pair;
  55
  56 /** Environment for all walker in boolopt. */
  57 typedef struct {
  58         int changed;  /**< Set if the graph was changed. */
  59 } bool_opt_env_t;
  60
  61 DEBUG_ONLY(static firm_dbg_module_t *dbg);
  62
  63 /**
  64  * Check if tho given nodes, l and r, represent two compares with
  65  * ... . If yes, return non-zero and fill the res struct.
  66  */
  67 static int find_cond_pair(ir_node *const l, ir_node *const r, cond_pair *const res)
  68 {
  69         if (is_Proj(l) && is_Proj(r)) {
  70                 ir_node *const lo = get_Proj_pred(l);
  71                 ir_node *const ro = get_Proj_pred(r);
  72
  73                 if (is_Cmp(lo) && is_Cmp(ro)) {
  74                         ir_node *const lol   = get_Cmp_left(lo);
  75                         ir_node *const lor   = get_Cmp_right(lo);
  76                         ir_node *const rol   = get_Cmp_left(ro);
  77                         ir_node *const ror   = get_Cmp_right(ro);
  78                         pn_Cmp   const pnc_l = get_Proj_proj(l);
  79                         pn_Cmp   const pnc_r = get_Proj_proj(r);
  80
  81                         if (is_Const(lor) && is_Const_null(lor) &&
  82                             is_Const(ror) && is_Const_null(ror) &&
  83                             pnc_l == pnc_r &&
  84                             (pnc_l == pn_Cmp_Lg || pnc_l == pn_Cmp_Eq)) {
  85                                 /* lo == (lol !=|== NULL) && ro == (rol !=|== NULL) */
  86                                 res->cmp_lo  = lo;
  87                                 res->cmp_hi  = ro;
  88                                 res->pnc_lo  = pnc_l;
  89                                 res->pnc_hi  = pnc_l;
  90                                 res->proj_lo = l;
  91                                 res->proj_hi = r;
  92                                 res->tv_lo   = get_Const_tarval(lor);
  93                                 res->tv_hi   = get_Const_tarval(ror);
  94                                 res->lo_mode = get_irn_mode(lor);
  95
  96                                 return 1;
  97                         }
  98
  99                         if (lol == rol && lor != ror && is_Const(lor) && is_Const(ror)) {
 100                                 /* lo == (x CMP c_l), ro == (x cmp c_r) */
 101                                 tarval *const tv_l  = get_Const_tarval(lor);
 102                                 tarval *const tv_r  = get_Const_tarval(ror);
 103                                 pn_Cmp  const rel   = tarval_cmp(tv_l, tv_r);
 104
 105                                 res->lo_mode = get_irn_mode(lol);
 106
 107                                 if (rel == pn_Cmp_Lt) {
 108                                         /* c_l < c_r */
 109                                         res->cmp_lo  = lo;
 110                                         res->cmp_hi  = ro;
 111                                         res->pnc_lo  = pnc_l;
 112                                         res->pnc_hi  = pnc_r;
 113                                         res->proj_lo = l;
 114                                         res->proj_hi = r;
 115                                         res->tv_lo   = tv_l;
 116                                         res->tv_hi   = tv_r;
 117                                 } else if (rel == pn_Cmp_Gt) {
 118                                         /* c_l > c_r */
 119                                         res->cmp_lo  = ro;
 120                                         res->cmp_hi  = lo;
 121                                         res->pnc_lo  = pnc_r;
 122                                         res->pnc_hi  = pnc_l;
 123                                         res->proj_lo = r;
 124                                         res->proj_hi = l;
 125                                         res->tv_lo   = tv_r;
 126                                         res->tv_hi   = tv_l;
 127                                 } else {
 128                                         /* The constants shall be unequal but comparable.
 129                                          * Local optimizations handle the equal case. */
 130                                         return 0;
 131                                 }
 132                                 return 1;
 133                         }
 134                 }
 135         }
 136         return 0;
 137 }
 138
 139 /**
 140  * Handle (lo pnc_lo x) AND (hi pnc_hi x)
 141  */
 142 static ir_node *bool_and(cond_pair* const cpair, ir_node *dst_block)
 143 {
 144         ir_node *const cmp_lo  = cpair->cmp_lo;
 145         ir_node *const cmp_hi  = cpair->cmp_hi;
 146         pn_Cmp         pnc_lo  = cpair->pnc_lo;
 147         pn_Cmp   const pnc_hi  = cpair->pnc_hi;
 148         ir_node *const proj_lo = cpair->proj_lo;
 149         ir_node *const proj_hi = cpair->proj_hi;
 150         tarval  *      tv_lo   = cpair->tv_lo;
 151         tarval  *      tv_hi   = cpair->tv_hi;
 152         ir_mode *      mode    = cpair->lo_mode;
 153
 154         if (pnc_lo == pn_Cmp_Eq && pnc_hi == pn_Cmp_Eq &&
 155             tarval_is_null(tv_lo) && tarval_is_null(tv_hi) &&
 156             mode == get_tarval_mode(tv_hi)) {
 157                 /* p == NULL && q == NULL ==> (p&q) == NULL) */
 158                 ir_node *lol, *hil, *cmp, *c, *p;
 159
 160                 if (mode_is_reference(mode)) {
 161                         mode = find_unsigned_mode(mode);
 162                         if (! mode)
 163                                 return NULL;
 164                         tv_lo = tarval_convert_to(tv_lo, mode);
 165                         if (tv_lo == tarval_bad)
 166                                 return NULL;
 167                 }
 168                 if (mode_is_int(mode)) {
 169                         lol   = get_Cmp_left(cmp_lo);
 170                         lol   = new_r_Conv(dst_block, lol, mode);
 171                         hil   = get_Cmp_left(cmp_hi);
 172                         hil   = new_r_Conv(dst_block, hil, mode);
 173                         p     = new_r_And(dst_block, lol, hil, mode);
 174                         c     = new_Const(tv_lo);
 175                         cmp   = new_r_Cmp(dst_block, p, c);
 176                         p     = new_r_Proj(dst_block, cmp, mode_b, pn_Cmp_Eq);
 177                         return p;
 178                 }
 179         }
 180
 181         /* the following tests expect one common operand */
 182         if (get_Cmp_left(cmp_lo) !=  get_Cmp_left(cmp_hi))
 183                 return 0;
 184
 185         /* TODO: for now reject float modes */
 186         if (! mode_is_int(mode))
 187                 return 0;
 188
 189         /* Beware of NaN's, we can only check for (ordered) != here (which is Lg, not Ne) */
 190         if ((pnc_lo == pn_Cmp_Lt || pnc_lo == pn_Cmp_Le || pnc_lo == pn_Cmp_Eq) &&
 191             (pnc_hi == pn_Cmp_Eq || pnc_hi == pn_Cmp_Ge || pnc_hi == pn_Cmp_Gt)) {
 192                 /* x <|<=|== lo && x ==|>=|> hi ==> false */
 193                 ir_node *const t = new_Const(tarval_b_false);
 194                 return t;
 195         } else if ((pnc_lo == pn_Cmp_Lt || pnc_lo == pn_Cmp_Le || pnc_lo == pn_Cmp_Eq) &&
 196                    (pnc_hi == pn_Cmp_Lt || pnc_hi == pn_Cmp_Le || pnc_hi == pn_Cmp_Lg)) {
 197                 /* x <|<=|== lo && x <|<=|!= hi ==> x <|<=|== lo */
 198                 return proj_lo;
 199         } else if ((pnc_lo == pn_Cmp_Ge || pnc_lo == pn_Cmp_Gt || pnc_lo == pn_Cmp_Lg) &&
 200                    (pnc_hi == pn_Cmp_Eq || pnc_hi == pn_Cmp_Ge || pnc_hi == pn_Cmp_Gt)) {
 201                 /* x >=|>|!= lo && x ==|>=|> hi ==> x ==|>=|> hi */
 202                 return proj_hi;
 203         } else if (tarval_is_one(tarval_sub(tv_hi, tv_lo, NULL))) { /* lo + 1 == hi */
 204                 if (pnc_lo == pn_Cmp_Ge && pnc_hi == pn_Cmp_Lt) {
 205                         /* x >= c && x < c + 1 ==> x == c */
 206                         ir_node  *const block = get_nodes_block(cmp_lo);
 207                         ir_node  *const p = new_r_Proj(block, cmp_lo, mode_b, pn_Cmp_Eq);
 208                         return p;
 209                 } else if (pnc_lo == pn_Cmp_Gt) {
 210                         if (pnc_hi == pn_Cmp_Lg) {
 211                                 /* x > c && x != c + 1 ==> x > c + 1 */
 212                                 ir_node  *const block = get_nodes_block(cmp_hi);
 213                                 ir_node  *const p = new_r_Proj(block, cmp_hi, mode_b, pn_Cmp_Gt);
 214                                 return p;
 215                         } else if (pnc_hi == pn_Cmp_Lt) {
 216                                 /* x > c && x < c + 1 ==> false */
 217                                 ir_node *const t = new_Const(tarval_b_false);
 218                                 return t;
 219                         } else if (pnc_hi == pn_Cmp_Le) {
 220                                 /* x > c && x <= c + 1 ==> x != c + 1 */
 221                                 ir_node  *const block = get_nodes_block(cmp_hi);
 222                                 ir_node  *const p = new_r_Proj(block, cmp_hi, mode_b, pn_Cmp_Eq);
 223                                 return p;
 224                         }
 225                 } else if (pnc_lo == pn_Cmp_Lg && pnc_hi == pn_Cmp_Lt) {
 226                         /* x != c && c < c + 1 ==> x < c */
 227                         ir_node  *const block = get_nodes_block(cmp_lo);
 228                         ir_node  *const p     = new_r_Proj(block, cmp_lo, mode_b, pn_Cmp_Lt);
 229                         return p;
 230                 }
 231         } else if ((pnc_lo == pn_Cmp_Gt || pnc_lo == pn_Cmp_Ge) &&
 232                    (pnc_hi == pn_Cmp_Lt || pnc_lo == pn_Cmp_Le) &&
 233                    get_mode_arithmetic(mode) == irma_twos_complement) {
 234                 /* works for two-complements only */
 235                 /* x >|\= lo && x <|<= hi ==> (x - lo) <u|<=u (hi-lo) */
 236                 if (pnc_lo == pn_Cmp_Gt) {
 237                         /* must convert to >= */
 238                         ir_mode *mode = get_tarval_mode(tv_lo);
 239                         tarval *n = tarval_add(tv_lo, get_mode_one(mode));
 240                         if (n != tarval_bad && tarval_cmp(n, tv_lo) == pn_Cmp_Gt) {
 241                                 /* no overflow */
 242                                 tv_lo = n;
 243                                 pnc_lo = pn_Cmp_Ge;
 244                         }
 245                 }
 246                 if (pnc_lo == pn_Cmp_Ge) {
 247                         /* all fine */
 248                         ir_node *const block = get_nodes_block(cmp_hi);
 249                         ir_node *      x     = get_Cmp_left(cmp_hi);
 250                         ir_mode *      mode  = get_irn_mode(x);
 251                         ir_node *sub, *cmp, *c, *subc, *p;
 252
 253                         if (mode_is_signed(mode)) {
 254                                 /* convert to unsigned */
 255                                 mode = find_unsigned_mode(mode);
 256                                 if (mode == NULL)
 257                                         return NULL;
 258                                 x     = new_r_Conv(block, x, mode);
 259                                 tv_lo = tarval_convert_to(tv_lo, mode);
 260                                 tv_hi = tarval_convert_to(tv_hi, mode);
 261                                 if (tv_lo == tarval_bad || tv_hi == tarval_bad)
 262                                         return NULL;
 263                         }
 264                         c    = new_Const(tv_lo);
 265                         sub  = new_r_Sub(block, x, c, mode);
 266                         subc = new_r_Sub(block, new_Const(tv_hi), c, mode);
 267                         cmp  = new_r_Cmp(block, sub, subc);
 268                         p    = new_r_Proj(block, cmp, mode_b, pnc_hi);
 269                         return p;
 270                 }
 271         }
 272         return NULL;
 273 }
 274
 275 /**
 276  * Handle (lo pnc_lo x) OR (hi pnc_hi x)
 277  */
 278 static ir_node *bool_or(cond_pair *const cpair, ir_node *dst_block)
 279 {
 280         ir_node *const cmp_lo  = cpair->cmp_lo;
 281         ir_node *const cmp_hi  = cpair->cmp_hi;
 282         pn_Cmp         pnc_lo  = cpair->pnc_lo;
 283         pn_Cmp   const pnc_hi  = cpair->pnc_hi;
 284         ir_node *const proj_lo = cpair->proj_lo;
 285         ir_node *const proj_hi = cpair->proj_hi;
 286         tarval  *      tv_lo   = cpair->tv_lo;
 287         tarval  *      tv_hi   = cpair->tv_hi;
 288         ir_mode *      mode    = cpair->lo_mode;
 289
 290         if (pnc_lo == pn_Cmp_Lg && pnc_hi == pn_Cmp_Lg &&
 291                 tarval_is_null(tv_lo) && tarval_is_null(tv_hi) &&
 292                 mode == get_tarval_mode(tv_hi)) {
 293                 /* p != NULL || q != NULL ==> (p|q) != NULL) */
 294                 ir_node *lol, *hil, *cmp, *c, *p;
 295
 296                 if (mode_is_reference(mode)) {
 297                         mode = find_unsigned_mode(mode);
 298                         if (! mode)
 299                                 return NULL;
 300                         tv_lo = tarval_convert_to(tv_lo, mode);
 301                         if (tv_lo == tarval_bad)
 302                                 return NULL;
 303                 }
 304                 if (mode_is_int(mode)) {
 305                         lol   = get_Cmp_left(cmp_lo);
 306                         lol   = new_r_Conv(dst_block, lol, mode);
 307                         hil   = get_Cmp_left(cmp_hi);
 308                         hil   = new_r_Conv(dst_block, hil, mode);
 309                         p     = new_r_Or(dst_block, lol, hil, mode);
 310                         c     = new_Const(tv_lo);
 311                         cmp   = new_r_Cmp(dst_block, p, c);
 312                         p     = new_r_Proj(dst_block, cmp, mode_b, pn_Cmp_Lg);
 313                         return p;
 314                 }
 315         }
 316
 317         /* the following tests expect one common operand */
 318         if (get_Cmp_left(cmp_lo) !=  get_Cmp_left(cmp_hi))
 319                 return 0;
 320
 321         /* TODO: for now reject float modes */
 322         if (! mode_is_int(mode))
 323                 return 0;
 324
 325         /* Beware of NaN's, we can only check for (ordered) != here (which is Lg, not Ne) */
 326         if ((pnc_lo == pn_Cmp_Ge || pnc_lo == pn_Cmp_Gt || pnc_lo == pn_Cmp_Lg) &&
 327             (pnc_hi == pn_Cmp_Lt || pnc_hi == pn_Cmp_Le || pnc_hi == pn_Cmp_Lg)) {
 328                 /* x >=|>|!= lo | x <|<=|!= hi ==> true */
 329                 ir_node *const t = new_Const(tarval_b_true);
 330                 return t;
 331         } else if ((pnc_lo == pn_Cmp_Lt || pnc_lo == pn_Cmp_Le || pnc_lo == pn_Cmp_Eq) &&
 332                    (pnc_hi == pn_Cmp_Lt || pnc_hi == pn_Cmp_Le || pnc_hi == pn_Cmp_Lg)) {
 333                 /* x <|<=|== lo || x <|<=|!= hi ==> x <|<=|!= hi */
 334                 return proj_hi;
 335         } else if ((pnc_lo == pn_Cmp_Ge || pnc_lo == pn_Cmp_Gt || pnc_lo == pn_Cmp_Lg) &&
 336                    (pnc_hi == pn_Cmp_Eq || pnc_hi == pn_Cmp_Ge || pnc_hi == pn_Cmp_Gt)) {
 337                 /* x >=|>|!= lo || x ==|>=|> hi ==> x >=|>|!= lo */
 338                 return proj_lo;
 339         } else if (tarval_is_one(tarval_sub(tv_hi, tv_lo, NULL))) { /* lo + 1 == hi */
 340                 if (pnc_lo == pn_Cmp_Lt && pnc_hi == pn_Cmp_Ge) {
 341                         /* x < c || x >= c + 1 ==> x != c */
 342                         ir_node  *const block = get_nodes_block(cmp_lo);
 343                         ir_node  *const p = new_r_Proj(block, cmp_lo, mode_b, pn_Cmp_Lg);
 344                         return p;
 345                 } else if (pnc_lo == pn_Cmp_Le) {
 346                         if (pnc_hi == pn_Cmp_Eq) {
 347                                 /* x <= c || x == c + 1 ==> x <= c + 1 */
 348                                 ir_node  *const block = get_nodes_block(cmp_hi);
 349                                 ir_node  *const p = new_r_Proj(block, cmp_hi, mode_b, pn_Cmp_Le);
 350                                 return p;
 351                         } else if (pnc_hi == pn_Cmp_Ge) {
 352                                 /* x <= c || x >= c + 1 ==> true */
 353                                 ir_node *const t = new_Const(tarval_b_true);
 354                                 return t;
 355                         } else if (pnc_hi == pn_Cmp_Gt) {
 356                                 /* x <= c || x > c + 1 ==> x != c + 1 */
 357                                 ir_node  *const block = get_nodes_block(cmp_hi);
 358                                 ir_node  *const p = new_r_Proj(block, cmp_hi, mode_b, pn_Cmp_Lg);
 359                                 return p;
 360                         }
 361                 } else if (pnc_lo == pn_Cmp_Eq && pnc_hi == pn_Cmp_Ge) {
 362                         /* x == c || x >= c + 1 ==> x >= c */
 363                         ir_node  *const block = get_nodes_block(cmp_lo);
 364                         ir_node  *const p     = new_r_Proj(block, cmp_lo, mode_b, pn_Cmp_Ge);
 365                         return p;
 366                 }
 367         } else if ((pnc_lo == pn_Cmp_Lt || pnc_lo == pn_Cmp_Le) &&
 368                    (pnc_hi == pn_Cmp_Gt || pnc_lo == pn_Cmp_Ge) &&
 369                    get_mode_arithmetic(mode) == irma_twos_complement) {
 370                 /* works for two-complements only */
 371                 /* x <|<= lo  || x >|>= hi ==> (x - lo) >u|>=u (hi-lo) */
 372                 if (pnc_lo == pn_Cmp_Le) {
 373                         /* must convert to < */
 374                         ir_mode *mode = get_tarval_mode(tv_lo);
 375                         tarval *n = tarval_add(tv_lo, get_mode_one(mode));
 376                         if (n != tarval_bad && tarval_cmp(n, tv_lo) == pn_Cmp_Gt) {
 377                                 /* no overflow */
 378                                 tv_lo = n;
 379                                 pnc_lo = pn_Cmp_Lt;
 380                         }
 381                 }
 382                 if (pnc_lo == pn_Cmp_Lt) {
 383                         /* all fine */
 384                         ir_node *const block = get_nodes_block(cmp_hi);
 385                         ir_node *      x     = get_Cmp_left(cmp_hi);
 386                         ir_mode *      mode  = get_irn_mode(x);
 387                         ir_node *sub, *cmp, *c, *subc, *p;
 388
 389                         if (mode_is_signed(mode)) {
 390                                 /* convert to unsigned */
 391                                 mode = find_unsigned_mode(mode);
 392                                 if (mode == NULL)
 393                                         return NULL;
 394                                 x     = new_r_Conv(block, x, mode);
 395                                 tv_lo = tarval_convert_to(tv_lo, mode);
 396                                 tv_hi = tarval_convert_to(tv_hi, mode);
 397                                 if (tv_lo == tarval_bad || tv_hi == tarval_bad)
 398                                         return NULL;
 399                         }
 400                         c    = new_Const(tv_lo);
 401                         sub  = new_r_Sub(block, x, c, mode);
 402                         subc = new_r_Sub(block, new_Const(tv_hi), c, mode);
 403                         cmp  = new_r_Cmp(block, sub, subc);
 404                         p    = new_r_Proj(block, cmp, mode_b, pnc_hi);
 405                         return p;
 406                 }
 407         }
 408         return NULL;
 409 }
 410
 411 /**
 412  * Walker, tries to optimize Andb and Orb nodes.
 413  */
 414 static void bool_walk(ir_node *n, void *ctx)
 415 {
 416         bool_opt_env_t *env = ctx;
 417
 418         if (get_irn_mode(n) != mode_b)
 419                 return;
 420
 421         if (is_And(n)) {
 422                 ir_node *const l = get_And_left(n);
 423                 ir_node *const r = get_And_right(n);
 424                 ir_node *      replacement;
 425                 cond_pair      cpair;
 426                 if (!find_cond_pair(l, r, &cpair))
 427                         return;
 428                 replacement = bool_and(&cpair, get_nodes_block(n));
 429                 if (replacement) {
 430                         exchange(n, replacement);
 431                         env->changed = 1;
 432                 }
 433         } else if (is_Or(n)) {
 434                 ir_node *const l = get_Or_left(n);
 435                 ir_node *const r = get_Or_right(n);
 436                 ir_node *      replacement;
 437                 cond_pair      cpair;
 438                 if (!find_cond_pair(l, r, &cpair))
 439                         return;
 440                 replacement = bool_or(&cpair, get_nodes_block(n));
 441                 if (replacement) {
 442                         exchange(n, replacement);
 443                         env->changed = 1;
 444                 }
 445         }
 446 }
 447
 448 /**
 449  * Walker, clear Block marker and Phi lists.
 450  */
 451 static void clear_block_infos(ir_node *node, void *env)
 452 {
 453         (void) env;
 454
 455         /* we visit blocks before any other nodes (from the block) */
 456         if (!is_Block(node))
 457                 return;
 458
 459         /* clear the PHI list */
 460         set_Block_phis(node, NULL);
 461         set_Block_mark(node, 0);
 462 }
 463
 464 /**
 465  * Walker: collect Phi nodes and mark
 466  */
 467 static void collect_phis(ir_node *node, void *env)
 468 {
 469         (void) env;
 470
 471         if (is_Phi(node)) {
 472                 ir_node *block = get_nodes_block(node);
 473                 add_Block_phi(block, node);
 474                 return;
 475         }
 476
 477         /* Ignore control flow nodes, these will be removed. */
 478         if (get_irn_pinned(node) == op_pin_state_pinned &&
 479                         !is_Block(node) && !is_cfop(node)) {
 480                                 /* found a pinned non-cf node, mark its block */
 481                 ir_node *block = get_nodes_block(node);
 482                 set_Block_mark(block, 1);
 483         }
 484 }
 485
 486 /**
 487  * If node is a Jmp in a block containing no pinned instruction
 488  * and having only one predecessor, skip the block and return its
 489  * cf predecessor, else the node itself.
 490  */
 491 static ir_node *skip_empty_blocks(ir_node *node)
 492 {
 493         while (is_Jmp(node)) {
 494                 ir_node *block = get_nodes_block(node);
 495
 496                 if (get_Block_n_cfgpreds(block) != 1)
 497                         break;
 498
 499                 if (get_Block_mark(block))
 500                         break;
 501
 502                 node = get_Block_cfgpred(block, 0);
 503         }
 504         return node;
 505 }
 506
 507 /**
 508  * Check if two block inputs can be fused.
 509  * This can be done, if block contains no Phi node that depends on
 510  * different inputs idx_i and idx_j.
 511  */
 512 static int can_fuse_block_inputs(const ir_node *block, int idx_i, int idx_j) {
 513         const ir_node *phi;
 514
 515         for (phi = get_Block_phis(block); phi != NULL; phi = get_Phi_next(phi)) {
 516                 if (get_Phi_pred(phi, idx_i) != get_Phi_pred(phi, idx_j))
 517                         return 0;
 518         }
 519         return 1;
 520 }
 521
 522 /**
 523  * Remove block input with given index.
 524  */
 525 static void remove_block_input(ir_node *block, int idx)
 526 {
 527         int i, j, n = get_Block_n_cfgpreds(block) - 1;
 528         ir_node *phi, **ins;
 529
 530         NEW_ARR_A(ir_node *, ins, n);
 531
 532         if (n == 1) {
 533                 /* all Phis will be deleted */
 534                 ir_node *next_phi;
 535
 536                 for (phi = get_Block_phis(block); phi != NULL; phi = next_phi) {
 537                         next_phi = get_Phi_next(phi);
 538                         exchange(phi, get_Phi_pred(phi, idx ^ 1));
 539                 }
 540                 set_Block_phis(block, NULL);
 541         } else {
 542                 for (phi = get_Block_phis(block); phi != NULL; phi = get_Phi_next(phi)) {
 543                         for (i = j = 0; i <= n; ++i) {
 544                                 if (i != idx)
 545                                         ins[j++] = get_Phi_pred(phi, i);
 546                         }
 547                         set_irn_in(phi, n, ins);
 548                 }
 549         }
 550         for (i = j = 0; i <= n; ++i) {
 551                 if (i != idx)
 552                         ins[j++] = get_Block_cfgpred(block, i);
 553         }
 554         set_irn_in(block, n, ins);
 555 }
 556
 557 /**
 558  * Under the preposition that we have a chain of blocks from
 559  * from_block to to_block, collapse them all into to_block.
 560  */
 561 static void move_nodes_to_block(ir_node *jmp, ir_node *to_block) {
 562         ir_node *new_jmp = NULL;
 563         ir_node *block, *next_block;
 564
 565         for (block = get_nodes_block(jmp); block != to_block; block = next_block) {
 566                 new_jmp = get_Block_cfgpred(block, 0);
 567                 next_block = get_nodes_block(new_jmp);
 568                 exchange(block, to_block);
 569         }
 570         if (new_jmp)
 571                 exchange(jmp, new_jmp);
 572 }
 573
 574 /**
 575  * Block walker:
 576  *
 577  * if we can find the following structure,
 578  *
 579  *        upper_block
 580  *         /       |
 581  *        /        |
 582  *   lower_block   |
 583  *     /  \        |
 584  *   ... low_idx up_idx
 585  *          \      |
 586  *            block
 587  *
 588  * try to convert it into a (x pnc_lo c_lo) || (x pnc_hi c_hi)
 589  * and optimize.
 590  */
 591 static void find_cf_and_or_walker(ir_node *block, void *ctx)
 592 {
 593         int low_idx, up_idx;
 594         int n_cfgpreds = get_Block_n_cfgpreds(block);
 595         bool_opt_env_t *env = ctx;
 596
 597 restart:
 598         if (n_cfgpreds < 2)
 599                 return;
 600
 601         for (low_idx = 0; low_idx < n_cfgpreds; ++low_idx) {
 602                 ir_node      *lower_block;
 603                 ir_node      *lower_cf;
 604                 ir_node      *cond;
 605                 ir_node      *cond_selector;
 606                 ir_node      *lower_pred;
 607
 608                 lower_cf = get_Block_cfgpred(block, low_idx);
 609                 lower_cf = skip_empty_blocks(lower_cf);
 610                 if (!is_Proj(lower_cf))
 611                         continue;
 612
 613                 cond = get_Proj_pred(lower_cf);
 614                 if (!is_Cond(cond))
 615                         continue;
 616
 617                 lower_block = get_nodes_block(cond);
 618                 if (get_Block_n_cfgpreds(lower_block) != 1)
 619                         continue;
 620
 621                 /* the block must not produce any side-effects */
 622                 if (get_Block_mark(lower_block))
 623                         continue;
 624
 625                 cond_selector = get_Cond_selector(cond);
 626                 if (get_irn_mode(cond_selector) != mode_b)
 627                         continue;
 628
 629                 lower_pred = get_Block_cfgpred_block(lower_block, 0);
 630
 631                 for (up_idx = 0; up_idx < n_cfgpreds; ++up_idx) {
 632                         ir_node   *upper_block;
 633                         ir_node   *upper_cf;
 634                         ir_node   *upper_cond;
 635                         ir_node   *upper_cond_selector;
 636                         ir_node   *replacement;
 637                         cond_pair  cpair;
 638
 639                         upper_cf    = get_Block_cfgpred(block, up_idx);
 640                         upper_cf    = skip_empty_blocks(upper_cf);
 641                         if (is_Bad(upper_cf))
 642                                 continue;
 643                         upper_block = get_nodes_block(upper_cf);
 644                         if (upper_block != lower_pred)
 645                                 continue;
 646                         if (!block_dominates(upper_block, block))
 647                                 continue;
 648
 649                         assert(is_Proj(upper_cf));
 650                         upper_cond = get_Proj_pred(upper_cf);
 651                         assert(is_Cond(upper_cond));
 652                         upper_cond_selector = get_Cond_selector(upper_cond);
 653                         if (get_irn_mode(upper_cond_selector) != mode_b)
 654                                 continue;
 655
 656                         /* we have found the structure */
 657                         /* check Phis: There must be NO Phi in block that
 658                            depends on the existence of low block */
 659                         if (!can_fuse_block_inputs(block, low_idx, up_idx))
 660                                 continue;
 661
 662                         /* all fine, try it */
 663                         if (!find_cond_pair(cond_selector, upper_cond_selector, &cpair))
 664                                 continue;
 665
 666                         /* normalize pncs: we need the true case to jump into the
 667                          * common block (ie. conjunctive normal form) */
 668                         if (get_Proj_proj(lower_cf) == pn_Cond_false) {
 669                                 if (cpair.proj_lo == cond_selector) {
 670                                         ir_mode *mode  = get_tarval_mode(cpair.tv_lo);
 671                                         ir_node *cmp   = get_Proj_pred(cpair.proj_lo);
 672                                         ir_node *block = get_nodes_block(cmp);
 673                                         cpair.pnc_lo   = get_negated_pnc(cpair.pnc_lo, mode);
 674                                         cpair.proj_lo  = new_r_Proj(block, cmp, mode_b, cpair.pnc_lo);
 675                                 } else {
 676                                         ir_mode *mode  = get_tarval_mode(cpair.tv_hi);
 677                                         ir_node *cmp   = get_Proj_pred(cpair.proj_hi);
 678                                         ir_node *block = get_nodes_block(cmp);
 679                                         assert(cpair.proj_hi == cond_selector);
 680                                         cpair.pnc_hi   = get_negated_pnc(cpair.pnc_hi, mode);
 681                                         cpair.proj_hi  = new_r_Proj(block, cmp, mode_b, cpair.pnc_hi);
 682                                 }
 683                         }
 684                         if (get_Proj_proj(upper_cf) == pn_Cond_false) {
 685                                 if (cpair.proj_lo == upper_cond_selector) {
 686                                         ir_mode *mode  = get_tarval_mode(cpair.tv_lo);
 687                                         ir_node *cmp   = get_Proj_pred(cpair.proj_lo);
 688                                         ir_node *block = get_nodes_block(cmp);
 689                                         cpair.pnc_lo   = get_negated_pnc(cpair.pnc_lo, mode);
 690                                         cpair.proj_lo  = new_r_Proj(block, cmp, mode_b, cpair.pnc_lo);
 691                                 } else {
 692                                         ir_mode *mode  = get_tarval_mode(cpair.tv_hi);
 693                                         ir_node *cmp   = get_Proj_pred(cpair.proj_hi);
 694                                         ir_node *block = get_nodes_block(cmp);
 695                                         assert(cpair.proj_hi == upper_cond_selector);
 696                                         cpair.pnc_hi   = get_negated_pnc(cpair.pnc_hi, mode);
 697                                         cpair.proj_hi  = new_r_Proj(block, cmp, mode_b, cpair.pnc_hi);
 698                                 }
 699                         }
 700
 701                         /* can we optimize the case? */
 702                         replacement = bool_or(&cpair, upper_block);
 703                         if (replacement == NULL)
 704                                 continue;
 705
 706                         env->changed = 1;
 707
 708                         DB((dbg, LEVEL_1, "boolopt: %+F: fusing (ub %+F lb %+F)\n",
 709                                 current_ir_graph, upper_block, lower_block));
 710
 711                         /* move all expressions on the path to lower/upper block */
 712                         move_nodes_to_block(get_Block_cfgpred(block, up_idx), upper_block);
 713                         move_nodes_to_block(get_Block_cfgpred(block, low_idx), lower_block);
 714
 715                         /* move all nodes from lower block to upper block */
 716                         exchange(lower_block, upper_block);
 717
 718                         remove_block_input(block, up_idx);
 719                         --n_cfgpreds;
 720
 721                         /* the optimizations expected the true case to jump */
 722                         if (get_Proj_proj(lower_cf) == pn_Cond_false) {
 723                                 ir_node *block = get_nodes_block(replacement);
 724                                 replacement    = new_rd_Not(NULL, block, replacement, mode_b);
 725                         }
 726                         set_Cond_selector(cond, replacement);
 727
 728                         goto restart;
 729                 }
 730         }
 731 }
 732
 733 void opt_bool(ir_graph *const irg)
 734 {
 735         bool_opt_env_t env;
 736
 737         /* register a debug mask */
 738         FIRM_DBG_REGISTER(dbg, "firm.opt.bool");
 739
 740         /* works better with one return block only */
 741         normalize_one_return(irg);
 742
 743         env.changed = 0;
 744
 745         /* optimize simple Andb and Orb cases */
 746         irg_walk_graph(irg, NULL, bool_walk, &env);
 747
 748         ir_reserve_resources(irg, IR_RESOURCE_BLOCK_MARK | IR_RESOURCE_PHI_LIST);
 749
 750         /* now more complicated cases: find control flow And/Or and optimize. */
 751         irg_walk_graph(irg, clear_block_infos, collect_phis, NULL);
 752         irg_block_walk_graph(irg, NULL, find_cf_and_or_walker, &env);
 753
 754         if (env.changed) {
 755                 set_irg_outs_inconsistent(irg);
 756                 set_irg_doms_inconsistent(irg);
 757                 set_irg_extblk_inconsistent(irg);
 758                 set_irg_loopinfo_inconsistent(irg);
 759         }
 760
 761         ir_free_resources(irg, IR_RESOURCE_BLOCK_MARK | IR_RESOURCE_PHI_LIST);
 762 }
 763
 764 /* Creates an ir_graph pass for opt_bool. */
 765 ir_graph_pass_t *opt_bool_pass(const char *name)
 766 {
 767         return def_graph_pass(name ? name : "opt_bool", opt_bool);
 768 }  /* opt_bool_pass */