nsz Git - libfirm/blob - ir/be/bepeephole.c

   1 /*
   2  * Copyright (C) 1995-2011 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       Peephole optimisation framework keeps track of which registers contain which values
  23  * @author      Matthias Braun
  24  * @version     $Id$
  25  */
  26 #include "config.h"
  27
  28 #include "array_t.h"
  29 #include "bepeephole.h"
  30
  31 #include "iredges_t.h"
  32 #include "irgwalk.h"
  33 #include "irprintf.h"
  34 #include "ircons.h"
  35 #include "irgmod.h"
  36 #include "error.h"
  37
  38 #include "beirg.h"
  39 #include "belive_t.h"
  40 #include "bearch.h"
  41 #include "benode.h"
  42 #include "besched.h"
  43 #include "bemodule.h"
  44
  45 DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
  46
  47 static const arch_env_t *arch_env;
  48 static be_lv_t          *lv;
  49 static ir_node          *current_node;
  50 ir_node               ***register_values;
  51
  52 static void clear_reg_value(ir_node *node)
  53 {
  54         const arch_register_t       *reg;
  55         const arch_register_class_t *cls;
  56         unsigned                     reg_idx;
  57         unsigned                     cls_idx;
  58
  59         if (!mode_is_data(get_irn_mode(node)))
  60                 return;
  61
  62         reg     = arch_get_irn_register(node);
  63         if (reg == NULL) {
  64                 panic("No register assigned at %+F", node);
  65         }
  66         if (reg->type & arch_register_type_virtual)
  67                 return;
  68         cls     = arch_register_get_class(reg);
  69         reg_idx = arch_register_get_index(reg);
  70         cls_idx = arch_register_class_index(cls);
  71
  72         //assert(register_values[cls_idx][reg_idx] != NULL);
  73         DB((dbg, LEVEL_1, "Clear Register %s\n", reg->name));
  74         register_values[cls_idx][reg_idx] = NULL;
  75 }
  76
  77 static void set_reg_value(ir_node *node)
  78 {
  79         const arch_register_t       *reg;
  80         const arch_register_class_t *cls;
  81         unsigned                     reg_idx;
  82         unsigned                     cls_idx;
  83
  84         if (!mode_is_data(get_irn_mode(node)))
  85                 return;
  86
  87         reg = arch_get_irn_register(node);
  88         if (reg == NULL) {
  89                 panic("No register assigned at %+F", node);
  90         }
  91         if (reg->type & arch_register_type_virtual)
  92                 return;
  93         cls     = arch_register_get_class(reg);
  94         reg_idx = arch_register_get_index(reg);
  95         cls_idx = arch_register_class_index(cls);
  96
  97         DB((dbg, LEVEL_1, "Set Register %s: %+F\n", reg->name, node));
  98         register_values[cls_idx][reg_idx] = node;
  99 }
 100
 101 static void clear_defs(ir_node *node)
 102 {
 103         /* clear values defined */
 104         if (get_irn_mode(node) == mode_T) {
 105                 const ir_edge_t *edge;
 106                 foreach_out_edge(node, edge) {
 107                         ir_node *proj = get_edge_src_irn(edge);
 108                         clear_reg_value(proj);
 109                 }
 110         } else {
 111                 clear_reg_value(node);
 112         }
 113 }
 114
 115 static void set_uses(ir_node *node)
 116 {
 117         int i, arity;
 118
 119         /* set values used */
 120         arity = get_irn_arity(node);
 121         for (i = 0; i < arity; ++i) {
 122                 ir_node *in = get_irn_n(node, i);
 123                 set_reg_value(in);
 124         }
 125 }
 126
 127 void be_peephole_new_node(ir_node * nw)
 128 {
 129         be_liveness_introduce(lv, nw);
 130 }
 131
 132 /**
 133  * must be called from peephole optimisations before a node will be killed
 134  * and its users will be redirected to new_node.
 135  * so bepeephole can update it's internal state.
 136  *
 137  * Note: killing a node and rewiring os only allowed if new_node produces
 138  * the same registers as old_node.
 139  */
 140 static void be_peephole_before_exchange(const ir_node *old_node,
 141                                         ir_node *new_node)
 142 {
 143         const arch_register_t       *reg;
 144         const arch_register_class_t *cls;
 145         unsigned                     reg_idx;
 146         unsigned                     cls_idx;
 147
 148         DB((dbg, LEVEL_1, "About to exchange and kill %+F with %+F\n", old_node, new_node));
 149
 150         if (current_node == old_node) {
 151                 /* next node to be processed will be killed. Its scheduling predecessor
 152                  * must be processed next. */
 153                 current_node = sched_next(current_node);
 154                 assert (!is_Bad(current_node));
 155         }
 156
 157         if (!mode_is_data(get_irn_mode(old_node)))
 158                 return;
 159
 160         reg = arch_get_irn_register(old_node);
 161         if (reg == NULL) {
 162                 panic("No register assigned at %+F", old_node);
 163         }
 164         assert(reg == arch_get_irn_register(new_node) &&
 165               "KILLING a node and replacing by different register is not allowed");
 166
 167         cls     = arch_register_get_class(reg);
 168         reg_idx = arch_register_get_index(reg);
 169         cls_idx = arch_register_class_index(cls);
 170
 171         if (register_values[cls_idx][reg_idx] == old_node) {
 172                 register_values[cls_idx][reg_idx] = new_node;
 173         }
 174
 175         be_liveness_remove(lv, old_node);
 176 }
 177
 178 void be_peephole_exchange(ir_node *old, ir_node *nw)
 179 {
 180         be_peephole_before_exchange(old, nw);
 181         sched_remove(old);
 182         exchange(old, nw);
 183         be_peephole_new_node(nw);
 184 }
 185
 186 /**
 187  * block-walker: run peephole optimization on the given block.
 188  */
 189 static void process_block(ir_node *block, void *data)
 190 {
 191         unsigned n_classes;
 192         unsigned i;
 193         int l;
 194         (void) data;
 195
 196         /* construct initial register assignment */
 197         n_classes = arch_env->n_register_classes;
 198         for (i = 0; i < n_classes; ++i) {
 199                 const arch_register_class_t *cls    = &arch_env->register_classes[i];
 200                 unsigned                     n_regs = arch_register_class_n_regs(cls);
 201                 memset(register_values[i], 0, sizeof(ir_node*) * n_regs);
 202         }
 203
 204         assert(lv->nodes && "live sets must be computed");
 205         DB((dbg, LEVEL_1, "\nProcessing block %+F (from end)\n", block));
 206         be_lv_foreach(lv, block, be_lv_state_end, l) {
 207                 ir_node *node = be_lv_get_irn(lv, block, l);
 208                 set_reg_value(node);
 209         }
 210         DB((dbg, LEVEL_1, "\nstart processing\n"));
 211
 212         /* walk the block from last insn to the first */
 213         current_node = sched_last(block);
 214         for ( ; !sched_is_begin(current_node);
 215                         current_node = sched_prev(current_node)) {
 216                 ir_op             *op;
 217                 peephole_opt_func  peephole_node;
 218
 219                 assert(!is_Bad(current_node));
 220                 if (is_Phi(current_node))
 221                         break;
 222
 223                 clear_defs(current_node);
 224                 set_uses(current_node);
 225
 226                 op            = get_irn_op(current_node);
 227                 peephole_node = (peephole_opt_func)op->ops.generic;
 228                 if (peephole_node == NULL)
 229                         continue;
 230
 231                 DB((dbg, LEVEL_2, "optimize %+F\n", current_node));
 232                 peephole_node(current_node);
 233                 assert(!is_Bad(current_node));
 234         }
 235 }
 236
 237 static void kill_node_and_preds(ir_node *node)
 238 {
 239         ir_graph *irg = get_irn_irg(node);
 240         int arity, i;
 241
 242         arity = get_irn_arity(node);
 243         for (i = 0; i < arity; ++i) {
 244                 ir_node *pred = get_irn_n(node, i);
 245
 246                 set_irn_n(node, i, new_r_Bad(irg));
 247                 if (get_irn_n_edges(pred) != 0)
 248                         continue;
 249
 250                 kill_node_and_preds(pred);
 251         }
 252
 253         if (!is_Proj(node))
 254                 sched_remove(node);
 255         kill_node(node);
 256 }
 257
 258 static void keep_alive_barrier_operand(ir_node *block, const ir_node* barrier, int pos)
 259 {
 260         ir_node *operand = get_irn_n(barrier, pos);
 261         ir_node *keep    = sched_next(skip_Proj(operand));
 262
 263         /* There already is a keep in the schedule. */
 264         if (be_is_Keep(keep)) {
 265                 const arch_register_class_t *cls = arch_get_irn_reg_class(barrier, pos);
 266
 267                 be_Keep_add_node(keep, cls, operand);
 268         }
 269         else {
 270                 ir_node *in[1] = {operand};
 271
 272                 keep = be_new_Keep(block, 1, in);
 273                 sched_add_after(skip_Proj(operand), keep);
 274         }
 275 }
 276
 277 /**
 278  * Walk through the block schedule and skip all barrier nodes.
 279  */
 280 static void skip_barrier(ir_node *block, ir_graph *irg)
 281 {
 282         ir_node *irn;
 283
 284         sched_foreach_reverse(block, irn) {
 285                 int       arity;
 286                 unsigned *used;
 287                 size_t    n_used;
 288                 const ir_edge_t *edge, *next;
 289
 290                 if (!be_is_Barrier(irn))
 291                         continue;
 292
 293                 /* track which outputs are actually used, as we have to create
 294                  * keep nodes for unused outputs */
 295                 arity = get_irn_arity(irn);
 296                 rbitset_alloca(used, arity);
 297
 298                 foreach_out_edge_safe(irn, edge, next) {
 299                         ir_node *proj = get_edge_src_irn(edge);
 300                         int      pn;
 301                         ir_node *pred;
 302
 303                         if (is_Anchor(proj))
 304                                 continue;
 305
 306                         pn   = (int) get_Proj_proj(proj);
 307                         pred = get_irn_n(irn, pn);
 308
 309                         rbitset_set(used, pn);
 310
 311                         /* We may need to reschedule be_Keeps to keep live-ranges short. */
 312                         if (get_irn_n_edges(proj) == 1) {
 313                                 const ir_edge_t *proj_edge = get_irn_out_edge_first(proj);
 314                                 ir_node         *proj_succ = get_edge_src_irn(proj_edge);
 315
 316                                 if (be_is_Keep(proj_succ)) {
 317                                         int succ_arity = get_irn_arity(proj_succ);
 318
 319                                         keep_alive_barrier_operand(block, irn, pn);
 320
 321                                         /* Disconnect old be_Keep. */
 322                                         if (succ_arity > 1) {
 323                                                 int      edge_pos        = get_edge_src_pos(proj_edge);
 324                                                 int      new_arity       = succ_arity - 1;
 325                                                 int      pos;
 326                                                 int      new_pos         = 0;
 327                                                 ir_node **ins;
 328
 329                                                 NEW_ARR_A(ir_node *, ins, succ_arity);
 330                                                 for (pos = 0; pos < succ_arity; ++pos) {
 331                                                         if (pos != edge_pos)
 332                                                                 ins[new_pos++] = get_irn_n(proj_succ, pos);
 333                                                 }
 334
 335                                                 set_irn_in(proj_succ, new_arity, ins);
 336                                         }
 337                                         else {
 338                                                 sched_remove(proj_succ);
 339                                                 kill_node(proj_succ);
 340                                         }
 341                                 }
 342                         }
 343
 344                         edges_reroute_kind(proj, pred, EDGE_KIND_NORMAL, irg);
 345                         edges_reroute_kind(proj, pred, EDGE_KIND_DEP, irg);
 346                 }
 347
 348                 /* the barrier also had the effect of a Keep for unused inputs.
 349                  * we now have to create an explicit Keep for them */
 350                 n_used = rbitset_popcount(used, arity);
 351                 if (n_used < (size_t) arity) {
 352                         int i;
 353
 354                         for (i = 0; i < arity; ++i) {
 355                                 if (rbitset_is_set(used, i))
 356                                         continue;
 357
 358                                 keep_alive_barrier_operand(block, irn, i);
 359                         }
 360                 }
 361
 362                 kill_node_and_preds(irn);
 363                 break;
 364         }
 365 }
 366
 367 /**
 368  * Kill the Barrier nodes for better peephole optimization.
 369  */
 370 static void kill_barriers(ir_graph *irg)
 371 {
 372         ir_node *end_blk = get_irg_end_block(irg);
 373         ir_node *start_blk = get_irg_start_block(irg);
 374         int i;
 375
 376         /* skip the barrier on all return blocks */
 377         for (i = get_Block_n_cfgpreds(end_blk) - 1; i >= 0; --i) {
 378                 ir_node *be_ret = get_Block_cfgpred(end_blk, i);
 379                 ir_node *ret_blk = get_nodes_block(be_ret);
 380
 381                 if (ret_blk == start_blk)
 382                         continue;
 383
 384                 skip_barrier(ret_blk, irg);
 385         }
 386
 387         /* skip the barrier on the start block */
 388         start_blk = get_irg_start_block(irg);
 389         skip_barrier(start_blk, irg);
 390 }
 391
 392 /**
 393  * Check whether the node has only one user.  Explicitly ignore the anchor.
 394  */
 395 static int has_only_one_user(ir_node *node)
 396 {
 397         int              n = get_irn_n_edges(node);
 398         const ir_edge_t *edge;
 399
 400         if (n <= 1)
 401                 return 1;
 402
 403         if (n > 2)
 404                 return 0;
 405
 406         foreach_out_edge(node, edge) {
 407                 ir_node *src = get_edge_src_irn(edge);
 408                 if (is_Anchor(src))
 409                         return 1;
 410         }
 411
 412         return 0;
 413 }
 414
 415 /*
 416  * Tries to optimize a beIncSP node with its previous IncSP node.
 417  * Must be run from a be_peephole_opt() context.
 418  */
 419 ir_node *be_peephole_IncSP_IncSP(ir_node *node)
 420 {
 421         int      pred_offs;
 422         int      curr_offs;
 423         int      offs;
 424         ir_node *pred = be_get_IncSP_pred(node);
 425
 426         if (!be_is_IncSP(pred))
 427                 return node;
 428
 429         if (!has_only_one_user(pred))
 430                 return node;
 431
 432         pred_offs = be_get_IncSP_offset(pred);
 433         curr_offs = be_get_IncSP_offset(node);
 434
 435         if (pred_offs == BE_STACK_FRAME_SIZE_EXPAND) {
 436                 if (curr_offs != BE_STACK_FRAME_SIZE_SHRINK) {
 437                         return node;
 438                 }
 439                 offs = 0;
 440         } else if (pred_offs == BE_STACK_FRAME_SIZE_SHRINK) {
 441                 if (curr_offs != BE_STACK_FRAME_SIZE_EXPAND) {
 442                         return node;
 443                 }
 444                 offs = 0;
 445         } else if (curr_offs == BE_STACK_FRAME_SIZE_EXPAND ||
 446                    curr_offs == BE_STACK_FRAME_SIZE_SHRINK) {
 447                 return node;
 448         } else {
 449                 offs = curr_offs + pred_offs;
 450         }
 451
 452         /* add node offset to pred and remove our IncSP */
 453         be_set_IncSP_offset(pred, offs);
 454
 455         be_peephole_exchange(node, pred);
 456         return pred;
 457 }
 458
 459 void be_peephole_opt(ir_graph *irg)
 460 {
 461         unsigned  n_classes;
 462         unsigned  i;
 463
 464         /* barrier nodes are used for register allocations. They hinders
 465          * peephole optimizations, so remove them here. */
 466         kill_barriers(irg);
 467
 468         /* we sometimes find BadE nodes in float apps like optest_float.c or
 469          * kahansum.c for example... */
 470         be_liveness_invalidate(be_get_irg_liveness(irg));
 471         be_liveness_assure_sets(be_assure_liveness(irg));
 472
 473         arch_env = be_get_irg_arch_env(irg);
 474         lv       = be_get_irg_liveness(irg);
 475
 476         n_classes = arch_env->n_register_classes;
 477         register_values = XMALLOCN(ir_node**, n_classes);
 478         for (i = 0; i < n_classes; ++i) {
 479                 const arch_register_class_t *cls    = &arch_env->register_classes[i];
 480                 unsigned                     n_regs = arch_register_class_n_regs(cls);
 481                 register_values[i] = XMALLOCN(ir_node*, n_regs);
 482         }
 483
 484         irg_block_walk_graph(irg, process_block, NULL, NULL);
 485
 486         for (i = 0; i < n_classes; ++i) {
 487                 xfree(register_values[i]);
 488         }
 489         xfree(register_values);
 490 }
 491
 492 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_peephole);
 493 void be_init_peephole(void)
 494 {
 495         FIRM_DBG_REGISTER(dbg, "firm.be.peephole");
 496 }