1 /*
   2  * Copyright (c) 2000, 2013, Oracle and/or its affiliates. All rights reserved.
   3  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   4  *
   5  * This code is free software; you can redistribute it and/or modify it
   6  * under the terms of the GNU General Public License version 2 only, as
   7  * published by the Free Software Foundation.
   8  *
   9  * This code is distributed in the hope that it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  12  * version 2 for more details (a copy is included in the LICENSE file that
  13  * accompanied this code).
  14  *
  15  * You should have received a copy of the GNU General Public License version
  16  * 2 along with this work; if not, write to the Free Software Foundation,
  17  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  18  *
  19  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  20  * or visit www.oracle.com if you need additional information or have any
  21  * questions.
  22  *
  23  */
  24 
  25 #include "precompiled.hpp"
  26 #include "compiler/compileLog.hpp"
  27 #include "memory/allocation.inline.hpp"
  28 #include "opto/addnode.hpp"
  29 #include "opto/callnode.hpp"
  30 #include "opto/castnode.hpp"
  31 #include "opto/connode.hpp"
  32 #include "opto/convertnode.hpp"
  33 #include "opto/divnode.hpp"
  34 #include "opto/loopnode.hpp"
  35 #include "opto/mulnode.hpp"
  36 #include "opto/movenode.hpp"
  37 #include "opto/opaquenode.hpp"
  38 #include "opto/rootnode.hpp"
  39 #include "opto/runtime.hpp"
  40 #include "opto/subnode.hpp"
  41 
  42 //------------------------------is_loop_exit-----------------------------------
  43 // Given an IfNode, return the loop-exiting projection or NULL if both
  44 // arms remain in the loop.
  45 Node *IdealLoopTree::is_loop_exit(Node *iff) const {
  46   if( iff->outcnt() != 2 ) return NULL; // Ignore partially dead tests
  47   PhaseIdealLoop *phase = _phase;
  48   // Test is an IfNode, has 2 projections.  If BOTH are in the loop
  49   // we need loop unswitching instead of peeling.
  50   if( !is_member(phase->get_loop( iff->raw_out(0) )) )
  51     return iff->raw_out(0);
  52   if( !is_member(phase->get_loop( iff->raw_out(1) )) )
  53     return iff->raw_out(1);
  54   return NULL;
  55 }
  56 
  57 
  58 //=============================================================================
  59 
  60 
  61 //------------------------------record_for_igvn----------------------------
  62 // Put loop body on igvn work list
  63 void IdealLoopTree::record_for_igvn() {
  64   for( uint i = 0; i < _body.size(); i++ ) {
  65     Node *n = _body.at(i);
  66     _phase->_igvn._worklist.push(n);
  67   }
  68 }
  69 
  70 //------------------------------compute_exact_trip_count-----------------------
  71 // Compute loop exact trip count if possible. Do not recalculate trip count for
  72 // split loops (pre-main-post) which have their limits and inits behind Opaque node.
  73 void IdealLoopTree::compute_exact_trip_count( PhaseIdealLoop *phase ) {
  74   if (!_head->as_Loop()->is_valid_counted_loop()) {
  75     return;
  76   }
  77   CountedLoopNode* cl = _head->as_CountedLoop();
  78   // Trip count may become nonexact for iteration split loops since
  79   // RCE modifies limits. Note, _trip_count value is not reset since
  80   // it is used to limit unrolling of main loop.
  81   cl->set_nonexact_trip_count();
  82 
  83   // Loop's test should be part of loop.
  84   if (!phase->is_member(this, phase->get_ctrl(cl->loopexit()->in(CountedLoopEndNode::TestValue))))
  85     return; // Infinite loop
  86 
  87 #ifdef ASSERT
  88   BoolTest::mask bt = cl->loopexit()->test_trip();
  89   assert(bt == BoolTest::lt || bt == BoolTest::gt ||
  90          bt == BoolTest::ne, "canonical test is expected");
  91 #endif
  92 
  93   Node* init_n = cl->init_trip();
  94   Node* limit_n = cl->limit();
  95   if (init_n  != NULL &&  init_n->is_Con() &&
  96       limit_n != NULL && limit_n->is_Con()) {
  97     // Use longs to avoid integer overflow.
  98     int stride_con  = cl->stride_con();
  99     jlong init_con   = cl->init_trip()->get_int();
 100     jlong limit_con  = cl->limit()->get_int();
 101     int stride_m    = stride_con - (stride_con > 0 ? 1 : -1);
 102     jlong trip_count = (limit_con - init_con + stride_m)/stride_con;
 103     if (trip_count > 0 && (julong)trip_count < (julong)max_juint) {
 104       // Set exact trip count.
 105       cl->set_exact_trip_count((uint)trip_count);
 106     }
 107   }
 108 }
 109 
 110 //------------------------------compute_profile_trip_cnt----------------------------
 111 // Compute loop trip count from profile data as
 112 //    (backedge_count + loop_exit_count) / loop_exit_count
 113 void IdealLoopTree::compute_profile_trip_cnt( PhaseIdealLoop *phase ) {
 114   if (!_head->is_CountedLoop()) {
 115     return;
 116   }
 117   CountedLoopNode* head = _head->as_CountedLoop();
 118   if (head->profile_trip_cnt() != COUNT_UNKNOWN) {
 119     return; // Already computed
 120   }
 121   float trip_cnt = (float)max_jint; // default is big
 122 
 123   Node* back = head->in(LoopNode::LoopBackControl);
 124   while (back != head) {
 125     if ((back->Opcode() == Op_IfTrue || back->Opcode() == Op_IfFalse) &&
 126         back->in(0) &&
 127         back->in(0)->is_If() &&
 128         back->in(0)->as_If()->_fcnt != COUNT_UNKNOWN &&
 129         back->in(0)->as_If()->_prob != PROB_UNKNOWN) {
 130       break;
 131     }
 132     back = phase->idom(back);
 133   }
 134   if (back != head) {
 135     assert((back->Opcode() == Op_IfTrue || back->Opcode() == Op_IfFalse) &&
 136            back->in(0), "if-projection exists");
 137     IfNode* back_if = back->in(0)->as_If();
 138     float loop_back_cnt = back_if->_fcnt * back_if->_prob;
 139 
 140     // Now compute a loop exit count
 141     float loop_exit_cnt = 0.0f;
 142     for( uint i = 0; i < _body.size(); i++ ) {
 143       Node *n = _body[i];
 144       if( n->is_If() ) {
 145         IfNode *iff = n->as_If();
 146         if( iff->_fcnt != COUNT_UNKNOWN && iff->_prob != PROB_UNKNOWN ) {
 147           Node *exit = is_loop_exit(iff);
 148           if( exit ) {
 149             float exit_prob = iff->_prob;
 150             if (exit->Opcode() == Op_IfFalse) exit_prob = 1.0 - exit_prob;
 151             if (exit_prob > PROB_MIN) {
 152               float exit_cnt = iff->_fcnt * exit_prob;
 153               loop_exit_cnt += exit_cnt;
 154             }
 155           }
 156         }
 157       }
 158     }
 159     if (loop_exit_cnt > 0.0f) {
 160       trip_cnt = (loop_back_cnt + loop_exit_cnt) / loop_exit_cnt;
 161     } else {
 162       // No exit count so use
 163       trip_cnt = loop_back_cnt;
 164     }
 165   }
 166 #ifndef PRODUCT
 167   if (TraceProfileTripCount) {
 168     tty->print_cr("compute_profile_trip_cnt  lp: %d cnt: %f\n", head->_idx, trip_cnt);
 169   }
 170 #endif
 171   head->set_profile_trip_cnt(trip_cnt);
 172 }
 173 
 174 //---------------------is_invariant_addition-----------------------------
 175 // Return nonzero index of invariant operand for an Add or Sub
 176 // of (nonconstant) invariant and variant values. Helper for reassociate_invariants.
 177 int IdealLoopTree::is_invariant_addition(Node* n, PhaseIdealLoop *phase) {
 178   int op = n->Opcode();
 179   if (op == Op_AddI || op == Op_SubI) {
 180     bool in1_invar = this->is_invariant(n->in(1));
 181     bool in2_invar = this->is_invariant(n->in(2));
 182     if (in1_invar && !in2_invar) return 1;
 183     if (!in1_invar && in2_invar) return 2;
 184   }
 185   return 0;
 186 }
 187 
 188 //---------------------reassociate_add_sub-----------------------------
 189 // Reassociate invariant add and subtract expressions:
 190 //
 191 // inv1 + (x + inv2)  =>  ( inv1 + inv2) + x
 192 // (x + inv2) + inv1  =>  ( inv1 + inv2) + x
 193 // inv1 + (x - inv2)  =>  ( inv1 - inv2) + x
 194 // inv1 - (inv2 - x)  =>  ( inv1 - inv2) + x
 195 // (x + inv2) - inv1  =>  (-inv1 + inv2) + x
 196 // (x - inv2) + inv1  =>  ( inv1 - inv2) + x
 197 // (x - inv2) - inv1  =>  (-inv1 - inv2) + x
 198 // inv1 + (inv2 - x)  =>  ( inv1 + inv2) - x
 199 // inv1 - (x - inv2)  =>  ( inv1 + inv2) - x
 200 // (inv2 - x) + inv1  =>  ( inv1 + inv2) - x
 201 // (inv2 - x) - inv1  =>  (-inv1 + inv2) - x
 202 // inv1 - (x + inv2)  =>  ( inv1 - inv2) - x
 203 //
 204 Node* IdealLoopTree::reassociate_add_sub(Node* n1, PhaseIdealLoop *phase) {
 205   if (!n1->is_Add() && !n1->is_Sub() || n1->outcnt() == 0) return NULL;
 206   if (is_invariant(n1)) return NULL;
 207   int inv1_idx = is_invariant_addition(n1, phase);
 208   if (!inv1_idx) return NULL;
 209   // Don't mess with add of constant (igvn moves them to expression tree root.)
 210   if (n1->is_Add() && n1->in(2)->is_Con()) return NULL;
 211   Node* inv1 = n1->in(inv1_idx);
 212   Node* n2 = n1->in(3 - inv1_idx);
 213   int inv2_idx = is_invariant_addition(n2, phase);
 214   if (!inv2_idx) return NULL;
 215   Node* x    = n2->in(3 - inv2_idx);
 216   Node* inv2 = n2->in(inv2_idx);
 217 
 218   bool neg_x    = n2->is_Sub() && inv2_idx == 1;
 219   bool neg_inv2 = n2->is_Sub() && inv2_idx == 2;
 220   bool neg_inv1 = n1->is_Sub() && inv1_idx == 2;
 221   if (n1->is_Sub() && inv1_idx == 1) {
 222     neg_x    = !neg_x;
 223     neg_inv2 = !neg_inv2;
 224   }
 225   Node* inv1_c = phase->get_ctrl(inv1);
 226   Node* inv2_c = phase->get_ctrl(inv2);
 227   Node* n_inv1;
 228   if (neg_inv1) {
 229     Node *zero = phase->_igvn.intcon(0);
 230     phase->set_ctrl(zero, phase->C->root());
 231     n_inv1 = new SubINode(zero, inv1);
 232     phase->register_new_node(n_inv1, inv1_c);
 233   } else {
 234     n_inv1 = inv1;
 235   }
 236   Node* inv;
 237   if (neg_inv2) {
 238     inv = new SubINode(n_inv1, inv2);
 239   } else {
 240     inv = new AddINode(n_inv1, inv2);
 241   }
 242   phase->register_new_node(inv, phase->get_early_ctrl(inv));
 243 
 244   Node* addx;
 245   if (neg_x) {
 246     addx = new SubINode(inv, x);
 247   } else {
 248     addx = new AddINode(x, inv);
 249   }
 250   phase->register_new_node(addx, phase->get_ctrl(x));
 251   phase->_igvn.replace_node(n1, addx);
 252   assert(phase->get_loop(phase->get_ctrl(n1)) == this, "");
 253   _body.yank(n1);
 254   return addx;
 255 }
 256 
 257 //---------------------reassociate_invariants-----------------------------
 258 // Reassociate invariant expressions:
 259 void IdealLoopTree::reassociate_invariants(PhaseIdealLoop *phase) {
 260   for (int i = _body.size() - 1; i >= 0; i--) {
 261     Node *n = _body.at(i);
 262     for (int j = 0; j < 5; j++) {
 263       Node* nn = reassociate_add_sub(n, phase);
 264       if (nn == NULL) break;
 265       n = nn; // again
 266     };
 267   }
 268 }
 269 
 270 //------------------------------policy_peeling---------------------------------
 271 // Return TRUE or FALSE if the loop should be peeled or not.  Peel if we can
 272 // make some loop-invariant test (usually a null-check) happen before the loop.
 273 bool IdealLoopTree::policy_peeling( PhaseIdealLoop *phase ) const {
 274   Node *test = ((IdealLoopTree*)this)->tail();
 275   int  body_size = ((IdealLoopTree*)this)->_body.size();
 276   // Peeling does loop cloning which can result in O(N^2) node construction
 277   if( body_size > 255 /* Prevent overflow for large body_size */
 278       || (body_size * body_size + phase->C->live_nodes()) > phase->C->max_node_limit() ) {
 279     return false;           // too large to safely clone
 280   }
 281   while( test != _head ) {      // Scan till run off top of loop
 282     if( test->is_If() ) {       // Test?
 283       Node *ctrl = phase->get_ctrl(test->in(1));
 284       if (ctrl->is_top())
 285         return false;           // Found dead test on live IF?  No peeling!
 286       // Standard IF only has one input value to check for loop invariance
 287       assert( test->Opcode() == Op_If || test->Opcode() == Op_CountedLoopEnd, "Check this code when new subtype is added");
 288       // Condition is not a member of this loop?
 289       if( !is_member(phase->get_loop(ctrl)) &&
 290           is_loop_exit(test) )
 291         return true;            // Found reason to peel!
 292     }
 293     // Walk up dominators to loop _head looking for test which is
 294     // executed on every path thru loop.
 295     test = phase->idom(test);
 296   }
 297   return false;
 298 }
 299 
 300 //------------------------------peeled_dom_test_elim---------------------------
 301 // If we got the effect of peeling, either by actually peeling or by making
 302 // a pre-loop which must execute at least once, we can remove all
 303 // loop-invariant dominated tests in the main body.
 304 void PhaseIdealLoop::peeled_dom_test_elim( IdealLoopTree *loop, Node_List &old_new ) {
 305   bool progress = true;
 306   while( progress ) {
 307     progress = false;           // Reset for next iteration
 308     Node *prev = loop->_head->in(LoopNode::LoopBackControl);//loop->tail();
 309     Node *test = prev->in(0);
 310     while( test != loop->_head ) { // Scan till run off top of loop
 311 
 312       int p_op = prev->Opcode();
 313       if( (p_op == Op_IfFalse || p_op == Op_IfTrue) &&
 314           test->is_If() &&      // Test?
 315           !test->in(1)->is_Con() && // And not already obvious?
 316           // Condition is not a member of this loop?
 317           !loop->is_member(get_loop(get_ctrl(test->in(1))))){
 318         // Walk loop body looking for instances of this test
 319         for( uint i = 0; i < loop->_body.size(); i++ ) {
 320           Node *n = loop->_body.at(i);
 321           if( n->is_If() && n->in(1) == test->in(1) /*&& n != loop->tail()->in(0)*/ ) {
 322             // IfNode was dominated by version in peeled loop body
 323             progress = true;
 324             dominated_by( old_new[prev->_idx], n );
 325           }
 326         }
 327       }
 328       prev = test;
 329       test = idom(test);
 330     } // End of scan tests in loop
 331 
 332   } // End of while( progress )
 333 }
 334 
 335 //------------------------------do_peeling-------------------------------------
 336 // Peel the first iteration of the given loop.
 337 // Step 1: Clone the loop body.  The clone becomes the peeled iteration.
 338 //         The pre-loop illegally has 2 control users (old & new loops).
 339 // Step 2: Make the old-loop fall-in edges point to the peeled iteration.
 340 //         Do this by making the old-loop fall-in edges act as if they came
 341 //         around the loopback from the prior iteration (follow the old-loop
 342 //         backedges) and then map to the new peeled iteration.  This leaves
 343 //         the pre-loop with only 1 user (the new peeled iteration), but the
 344 //         peeled-loop backedge has 2 users.
 345 // Step 3: Cut the backedge on the clone (so its not a loop) and remove the
 346 //         extra backedge user.
 347 //
 348 //                   orig
 349 //
 350 //                  stmt1
 351 //                    |
 352 //                    v
 353 //              loop predicate
 354 //                    |
 355 //                    v
 356 //                   loop<----+
 357 //                     |      |
 358 //                   stmt2    |
 359 //                     |      |
 360 //                     v      |
 361 //                    if      ^
 362 //                   / \      |
 363 //                  /   \     |
 364 //                 v     v    |
 365 //               false true   |
 366 //               /       \    |
 367 //              /         ----+
 368 //             |
 369 //             v
 370 //           exit
 371 //
 372 //
 373 //            after clone loop
 374 //
 375 //                   stmt1
 376 //                     |
 377 //                     v
 378 //               loop predicate
 379 //                 /       \
 380 //        clone   /         \   orig
 381 //               /           \
 382 //              /             \
 383 //             v               v
 384 //   +---->loop clone          loop<----+
 385 //   |      |                    |      |
 386 //   |    stmt2 clone          stmt2    |
 387 //   |      |                    |      |
 388 //   |      v                    v      |
 389 //   ^      if clone            If      ^
 390 //   |      / \                / \      |
 391 //   |     /   \              /   \     |
 392 //   |    v     v            v     v    |
 393 //   |    true  false      false true   |
 394 //   |    /         \      /       \    |
 395 //   +----           \    /         ----+
 396 //                    \  /
 397 //                    1v v2
 398 //                  region
 399 //                     |
 400 //                     v
 401 //                   exit
 402 //
 403 //
 404 //         after peel and predicate move
 405 //
 406 //                   stmt1
 407 //                    /
 408 //                   /
 409 //        clone     /            orig
 410 //                 /
 411 //                /              +----------+
 412 //               /               |          |
 413 //              /          loop predicate   |
 414 //             /                 |          |
 415 //            v                  v          |
 416 //   TOP-->loop clone          loop<----+   |
 417 //          |                    |      |   |
 418 //        stmt2 clone          stmt2    |   |
 419 //          |                    |      |   ^
 420 //          v                    v      |   |
 421 //          if clone            If      ^   |
 422 //          / \                / \      |   |
 423 //         /   \              /   \     |   |
 424 //        v     v            v     v    |   |
 425 //      true   false      false  true   |   |
 426 //        |         \      /       \    |   |
 427 //        |          \    /         ----+   ^
 428 //        |           \  /                  |
 429 //        |           1v v2                 |
 430 //        v         region                  |
 431 //        |            |                    |
 432 //        |            v                    |
 433 //        |          exit                   |
 434 //        |                                 |
 435 //        +--------------->-----------------+
 436 //
 437 //
 438 //              final graph
 439 //
 440 //                  stmt1
 441 //                    |
 442 //                    v
 443 //                  stmt2 clone
 444 //                    |
 445 //                    v
 446 //                   if clone
 447 //                  / |
 448 //                 /  |
 449 //                v   v
 450 //            false  true
 451 //             |      |
 452 //             |      v
 453 //             | loop predicate
 454 //             |      |
 455 //             |      v
 456 //             |     loop<----+
 457 //             |      |       |
 458 //             |    stmt2     |
 459 //             |      |       |
 460 //             |      v       |
 461 //             v      if      ^
 462 //             |     /  \     |
 463 //             |    /    \    |
 464 //             |   v     v    |
 465 //             | false  true  |
 466 //             |  |        \  |
 467 //             v  v         --+
 468 //            region
 469 //              |
 470 //              v
 471 //             exit
 472 //
 473 void PhaseIdealLoop::do_peeling( IdealLoopTree *loop, Node_List &old_new ) {
 474 
 475   C->set_major_progress();
 476   // Peeling a 'main' loop in a pre/main/post situation obfuscates the
 477   // 'pre' loop from the main and the 'pre' can no longer have it's
 478   // iterations adjusted.  Therefore, we need to declare this loop as
 479   // no longer a 'main' loop; it will need new pre and post loops before
 480   // we can do further RCE.
 481 #ifndef PRODUCT
 482   if (TraceLoopOpts) {
 483     tty->print("Peel         ");
 484     loop->dump_head();
 485   }
 486 #endif
 487   Node* head = loop->_head;
 488   bool counted_loop = head->is_CountedLoop();
 489   if (counted_loop) {
 490     CountedLoopNode *cl = head->as_CountedLoop();
 491     assert(cl->trip_count() > 0, "peeling a fully unrolled loop");
 492     cl->set_trip_count(cl->trip_count() - 1);
 493     if (cl->is_main_loop()) {
 494       cl->set_normal_loop();
 495 #ifndef PRODUCT
 496       if (PrintOpto && VerifyLoopOptimizations) {
 497         tty->print("Peeling a 'main' loop; resetting to 'normal' ");
 498         loop->dump_head();
 499       }
 500 #endif
 501     }
 502   }
 503   Node* entry = head->in(LoopNode::EntryControl);
 504 
 505   // Step 1: Clone the loop body.  The clone becomes the peeled iteration.
 506   //         The pre-loop illegally has 2 control users (old & new loops).
 507   clone_loop( loop, old_new, dom_depth(head) );
 508 
 509   // Step 2: Make the old-loop fall-in edges point to the peeled iteration.
 510   //         Do this by making the old-loop fall-in edges act as if they came
 511   //         around the loopback from the prior iteration (follow the old-loop
 512   //         backedges) and then map to the new peeled iteration.  This leaves
 513   //         the pre-loop with only 1 user (the new peeled iteration), but the
 514   //         peeled-loop backedge has 2 users.
 515   Node* new_entry = old_new[head->in(LoopNode::LoopBackControl)->_idx];
 516   _igvn.hash_delete(head);
 517   head->set_req(LoopNode::EntryControl, new_entry);
 518   for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) {
 519     Node* old = head->fast_out(j);
 520     if (old->in(0) == loop->_head && old->req() == 3 && old->is_Phi()) {
 521       Node* new_exit_value = old_new[old->in(LoopNode::LoopBackControl)->_idx];
 522       if (!new_exit_value )     // Backedge value is ALSO loop invariant?
 523         // Then loop body backedge value remains the same.
 524         new_exit_value = old->in(LoopNode::LoopBackControl);
 525       _igvn.hash_delete(old);
 526       old->set_req(LoopNode::EntryControl, new_exit_value);
 527     }
 528   }
 529 
 530 
 531   // Step 3: Cut the backedge on the clone (so its not a loop) and remove the
 532   //         extra backedge user.
 533   Node* new_head = old_new[head->_idx];
 534   _igvn.hash_delete(new_head);
 535   new_head->set_req(LoopNode::LoopBackControl, C->top());
 536   for (DUIterator_Fast j2max, j2 = new_head->fast_outs(j2max); j2 < j2max; j2++) {
 537     Node* use = new_head->fast_out(j2);
 538     if (use->in(0) == new_head && use->req() == 3 && use->is_Phi()) {
 539       _igvn.hash_delete(use);
 540       use->set_req(LoopNode::LoopBackControl, C->top());
 541     }
 542   }
 543 
 544 
 545   // Step 4: Correct dom-depth info.  Set to loop-head depth.
 546   int dd = dom_depth(head);
 547   set_idom(head, head->in(1), dd);
 548   for (uint j3 = 0; j3 < loop->_body.size(); j3++) {
 549     Node *old = loop->_body.at(j3);
 550     Node *nnn = old_new[old->_idx];
 551     if (!has_ctrl(nnn))
 552       set_idom(nnn, idom(nnn), dd-1);
 553   }
 554 
 555   // Now force out all loop-invariant dominating tests.  The optimizer
 556   // finds some, but we _know_ they are all useless.
 557   peeled_dom_test_elim(loop,old_new);
 558 
 559   loop->record_for_igvn();
 560 }
 561 
 562 #define EMPTY_LOOP_SIZE 7 // number of nodes in an empty loop
 563 
 564 //------------------------------policy_maximally_unroll------------------------
 565 // Calculate exact loop trip count and return true if loop can be maximally
 566 // unrolled.
 567 bool IdealLoopTree::policy_maximally_unroll( PhaseIdealLoop *phase ) const {
 568   CountedLoopNode *cl = _head->as_CountedLoop();
 569   assert(cl->is_normal_loop(), "");
 570   if (!cl->is_valid_counted_loop())
 571     return false; // Malformed counted loop
 572 
 573   if (!cl->has_exact_trip_count()) {
 574     // Trip count is not exact.
 575     return false;
 576   }
 577 
 578   uint trip_count = cl->trip_count();
 579   // Note, max_juint is used to indicate unknown trip count.
 580   assert(trip_count > 1, "one iteration loop should be optimized out already");
 581   assert(trip_count < max_juint, "exact trip_count should be less than max_uint.");
 582 
 583   // Real policy: if we maximally unroll, does it get too big?
 584   // Allow the unrolled mess to get larger than standard loop
 585   // size.  After all, it will no longer be a loop.
 586   uint body_size    = _body.size();
 587   uint unroll_limit = (uint)LoopUnrollLimit * 4;
 588   assert( (intx)unroll_limit == LoopUnrollLimit * 4, "LoopUnrollLimit must fit in 32bits");
 589   if (trip_count > unroll_limit || body_size > unroll_limit) {
 590     return false;
 591   }
 592 
 593   // Fully unroll a loop with few iterations regardless next
 594   // conditions since following loop optimizations will split
 595   // such loop anyway (pre-main-post).
 596   if (trip_count <= 3)
 597     return true;
 598 
 599   // Take into account that after unroll conjoined heads and tails will fold,
 600   // otherwise policy_unroll() may allow more unrolling than max unrolling.
 601   uint new_body_size = EMPTY_LOOP_SIZE + (body_size - EMPTY_LOOP_SIZE) * trip_count;
 602   uint tst_body_size = (new_body_size - EMPTY_LOOP_SIZE) / trip_count + EMPTY_LOOP_SIZE;
 603   if (body_size != tst_body_size) // Check for int overflow
 604     return false;
 605   if (new_body_size > unroll_limit ||
 606       // Unrolling can result in a large amount of node construction
 607       new_body_size >= phase->C->max_node_limit() - phase->C->live_nodes()) {
 608     return false;
 609   }
 610 
 611   // Do not unroll a loop with String intrinsics code.
 612   // String intrinsics are large and have loops.
 613   for (uint k = 0; k < _body.size(); k++) {
 614     Node* n = _body.at(k);
 615     switch (n->Opcode()) {
 616       case Op_StrComp:
 617       case Op_StrEquals:
 618       case Op_StrIndexOf:
 619       case Op_EncodeISOArray:
 620       case Op_AryEq: {
 621         return false;
 622       }
 623 #if INCLUDE_RTM_OPT
 624       case Op_FastLock:
 625       case Op_FastUnlock: {
 626         // Don't unroll RTM locking code because it is large.
 627         if (UseRTMLocking) {
 628           return false;
 629         }
 630       }
 631 #endif
 632     } // switch
 633   }
 634 
 635   return true; // Do maximally unroll
 636 }
 637 
 638 
 639 //------------------------------policy_unroll----------------------------------
 640 // Return TRUE or FALSE if the loop should be unrolled or not.  Unroll if
 641 // the loop is a CountedLoop and the body is small enough.
 642 bool IdealLoopTree::policy_unroll( PhaseIdealLoop *phase ) const {
 643 
 644   CountedLoopNode *cl = _head->as_CountedLoop();
 645   assert(cl->is_normal_loop() || cl->is_main_loop(), "");
 646 
 647   if (!cl->is_valid_counted_loop())
 648     return false; // Malformed counted loop
 649 
 650   // Protect against over-unrolling.
 651   // After split at least one iteration will be executed in pre-loop.
 652   if (cl->trip_count() <= (uint)(cl->is_normal_loop() ? 2 : 1)) return false;
 653 
 654   int future_unroll_ct = cl->unrolled_count() * 2;
 655   if (future_unroll_ct > LoopMaxUnroll) return false;
 656 
 657   // Check for initial stride being a small enough constant
 658   if (abs(cl->stride_con()) > (1<<2)*future_unroll_ct) return false;
 659 
 660   // Don't unroll if the next round of unrolling would push us
 661   // over the expected trip count of the loop.  One is subtracted
 662   // from the expected trip count because the pre-loop normally
 663   // executes 1 iteration.
 664   if (UnrollLimitForProfileCheck > 0 &&
 665       cl->profile_trip_cnt() != COUNT_UNKNOWN &&
 666       future_unroll_ct        > UnrollLimitForProfileCheck &&
 667       (float)future_unroll_ct > cl->profile_trip_cnt() - 1.0) {
 668     return false;
 669   }
 670 
 671   // When unroll count is greater than LoopUnrollMin, don't unroll if:
 672   //   the residual iterations are more than 10% of the trip count
 673   //   and rounds of "unroll,optimize" are not making significant progress
 674   //   Progress defined as current size less than 20% larger than previous size.
 675   if (UseSuperWord && cl->node_count_before_unroll() > 0 &&
 676       future_unroll_ct > LoopUnrollMin &&
 677       (future_unroll_ct - 1) * 10.0 > cl->profile_trip_cnt() &&
 678       1.2 * cl->node_count_before_unroll() < (double)_body.size()) {
 679     return false;
 680   }
 681 
 682   Node *init_n = cl->init_trip();
 683   Node *limit_n = cl->limit();
 684   int stride_con = cl->stride_con();
 685   // Non-constant bounds.
 686   // Protect against over-unrolling when init or/and limit are not constant
 687   // (so that trip_count's init value is maxint) but iv range is known.
 688   if (init_n   == NULL || !init_n->is_Con()  ||
 689       limit_n  == NULL || !limit_n->is_Con()) {
 690     Node* phi = cl->phi();
 691     if (phi != NULL) {
 692       assert(phi->is_Phi() && phi->in(0) == _head, "Counted loop should have iv phi.");
 693       const TypeInt* iv_type = phase->_igvn.type(phi)->is_int();
 694       int next_stride = stride_con * 2; // stride after this unroll
 695       if (next_stride > 0) {
 696         if (iv_type->_lo + next_stride <= iv_type->_lo || // overflow
 697             iv_type->_lo + next_stride >  iv_type->_hi) {
 698           return false;  // over-unrolling
 699         }
 700       } else if (next_stride < 0) {
 701         if (iv_type->_hi + next_stride >= iv_type->_hi || // overflow
 702             iv_type->_hi + next_stride <  iv_type->_lo) {
 703           return false;  // over-unrolling
 704         }
 705       }
 706     }
 707   }
 708 
 709   // After unroll limit will be adjusted: new_limit = limit-stride.
 710   // Bailout if adjustment overflow.
 711   const TypeInt* limit_type = phase->_igvn.type(limit_n)->is_int();
 712   if (stride_con > 0 && ((limit_type->_hi - stride_con) >= limit_type->_hi) ||
 713       stride_con < 0 && ((limit_type->_lo - stride_con) <= limit_type->_lo))
 714     return false;  // overflow
 715 
 716   // Adjust body_size to determine if we unroll or not
 717   uint body_size = _body.size();
 718   // Key test to unroll loop in CRC32 java code
 719   int xors_in_loop = 0;
 720   // Also count ModL, DivL and MulL which expand mightly
 721   for (uint k = 0; k < _body.size(); k++) {
 722     Node* n = _body.at(k);
 723     switch (n->Opcode()) {
 724       case Op_XorI: xors_in_loop++; break; // CRC32 java code
 725       case Op_ModL: body_size += 30; break;
 726       case Op_DivL: body_size += 30; break;
 727       case Op_MulL: body_size += 10; break;
 728       case Op_StrComp:
 729       case Op_StrEquals:
 730       case Op_StrIndexOf:
 731       case Op_EncodeISOArray:
 732       case Op_AryEq: {
 733         // Do not unroll a loop with String intrinsics code.
 734         // String intrinsics are large and have loops.
 735         return false;
 736       }
 737 #if INCLUDE_RTM_OPT
 738       case Op_FastLock:
 739       case Op_FastUnlock: {
 740         // Don't unroll RTM locking code because it is large.
 741         if (UseRTMLocking) {
 742           return false;
 743         }
 744       }
 745 #endif
 746     } // switch
 747   }
 748 
 749   // Check for being too big
 750   if (body_size > (uint)LoopUnrollLimit) {
 751     if (xors_in_loop >= 4 && body_size < (uint)LoopUnrollLimit*4) return true;
 752     // Normal case: loop too big
 753     return false;
 754   }
 755 
 756   // Unroll once!  (Each trip will soon do double iterations)
 757   return true;
 758 }
 759 
 760 //------------------------------policy_align-----------------------------------
 761 // Return TRUE or FALSE if the loop should be cache-line aligned.  Gather the
 762 // expression that does the alignment.  Note that only one array base can be
 763 // aligned in a loop (unless the VM guarantees mutual alignment).  Note that
 764 // if we vectorize short memory ops into longer memory ops, we may want to
 765 // increase alignment.
 766 bool IdealLoopTree::policy_align( PhaseIdealLoop *phase ) const {
 767   return false;
 768 }
 769 
 770 //------------------------------policy_range_check-----------------------------
 771 // Return TRUE or FALSE if the loop should be range-check-eliminated.
 772 // Actually we do iteration-splitting, a more powerful form of RCE.
 773 bool IdealLoopTree::policy_range_check( PhaseIdealLoop *phase ) const {
 774   if (!RangeCheckElimination) return false;
 775 
 776   CountedLoopNode *cl = _head->as_CountedLoop();
 777   // If we unrolled with no intention of doing RCE and we later
 778   // changed our minds, we got no pre-loop.  Either we need to
 779   // make a new pre-loop, or we gotta disallow RCE.
 780   if (cl->is_main_no_pre_loop()) return false; // Disallowed for now.
 781   Node *trip_counter = cl->phi();
 782 
 783   // Check loop body for tests of trip-counter plus loop-invariant vs
 784   // loop-invariant.
 785   for (uint i = 0; i < _body.size(); i++) {
 786     Node *iff = _body[i];
 787     if (iff->Opcode() == Op_If) { // Test?
 788 
 789       // Comparing trip+off vs limit
 790       Node *bol = iff->in(1);
 791       if (bol->req() != 2) continue; // dead constant test
 792       if (!bol->is_Bool()) {
 793         assert(UseLoopPredicate && bol->Opcode() == Op_Conv2B, "predicate check only");
 794         continue;
 795       }
 796       if (bol->as_Bool()->_test._test == BoolTest::ne)
 797         continue; // not RC
 798 
 799       Node *cmp = bol->in(1);
 800       Node *rc_exp = cmp->in(1);
 801       Node *limit = cmp->in(2);
 802 
 803       Node *limit_c = phase->get_ctrl(limit);
 804       if( limit_c == phase->C->top() )
 805         return false;           // Found dead test on live IF?  No RCE!
 806       if( is_member(phase->get_loop(limit_c) ) ) {
 807         // Compare might have operands swapped; commute them
 808         rc_exp = cmp->in(2);
 809         limit  = cmp->in(1);
 810         limit_c = phase->get_ctrl(limit);
 811         if( is_member(phase->get_loop(limit_c) ) )
 812           continue;             // Both inputs are loop varying; cannot RCE
 813       }
 814 
 815       if (!phase->is_scaled_iv_plus_offset(rc_exp, trip_counter, NULL, NULL)) {
 816         continue;
 817       }
 818       // Yeah!  Found a test like 'trip+off vs limit'
 819       // Test is an IfNode, has 2 projections.  If BOTH are in the loop
 820       // we need loop unswitching instead of iteration splitting.
 821       if( is_loop_exit(iff) )
 822         return true;            // Found reason to split iterations
 823     } // End of is IF
 824   }
 825 
 826   return false;
 827 }
 828 
 829 //------------------------------policy_peel_only-------------------------------
 830 // Return TRUE or FALSE if the loop should NEVER be RCE'd or aligned.  Useful
 831 // for unrolling loops with NO array accesses.
 832 bool IdealLoopTree::policy_peel_only( PhaseIdealLoop *phase ) const {
 833 
 834   for( uint i = 0; i < _body.size(); i++ )
 835     if( _body[i]->is_Mem() )
 836       return false;
 837 
 838   // No memory accesses at all!
 839   return true;
 840 }
 841 
 842 //------------------------------clone_up_backedge_goo--------------------------
 843 // If Node n lives in the back_ctrl block and cannot float, we clone a private
 844 // version of n in preheader_ctrl block and return that, otherwise return n.
 845 Node *PhaseIdealLoop::clone_up_backedge_goo( Node *back_ctrl, Node *preheader_ctrl, Node *n, VectorSet &visited, Node_Stack &clones ) {
 846   if( get_ctrl(n) != back_ctrl ) return n;
 847 
 848   // Only visit once
 849   if (visited.test_set(n->_idx)) {
 850     Node *x = clones.find(n->_idx);
 851     if (x != NULL)
 852       return x;
 853     return n;
 854   }
 855 
 856   Node *x = NULL;               // If required, a clone of 'n'
 857   // Check for 'n' being pinned in the backedge.
 858   if( n->in(0) && n->in(0) == back_ctrl ) {
 859     assert(clones.find(n->_idx) == NULL, "dead loop");
 860     x = n->clone();             // Clone a copy of 'n' to preheader
 861     clones.push(x, n->_idx);
 862     x->set_req( 0, preheader_ctrl ); // Fix x's control input to preheader
 863   }
 864 
 865   // Recursive fixup any other input edges into x.
 866   // If there are no changes we can just return 'n', otherwise
 867   // we need to clone a private copy and change it.
 868   for( uint i = 1; i < n->req(); i++ ) {
 869     Node *g = clone_up_backedge_goo( back_ctrl, preheader_ctrl, n->in(i), visited, clones );
 870     if( g != n->in(i) ) {
 871       if( !x ) {
 872         assert(clones.find(n->_idx) == NULL, "dead loop");
 873         x = n->clone();
 874         clones.push(x, n->_idx);
 875       }
 876       x->set_req(i, g);
 877     }
 878   }
 879   if( x ) {                     // x can legally float to pre-header location
 880     register_new_node( x, preheader_ctrl );
 881     return x;
 882   } else {                      // raise n to cover LCA of uses
 883     set_ctrl( n, find_non_split_ctrl(back_ctrl->in(0)) );
 884   }
 885   return n;
 886 }
 887 
 888 bool PhaseIdealLoop::cast_incr_before_loop(Node* incr, Node* ctrl, Node* loop) {
 889   Node* castii = new CastIINode(incr, TypeInt::INT, true);
 890   castii->set_req(0, ctrl);
 891   register_new_node(castii, ctrl);
 892   for (DUIterator_Fast imax, i = incr->fast_outs(imax); i < imax; i++) {
 893     Node* n = incr->fast_out(i);
 894     if (n->is_Phi() && n->in(0) == loop) {
 895       int nrep = n->replace_edge(incr, castii);
 896       return true;
 897     }
 898   }
 899   return false;
 900 }
 901 
 902 //------------------------------insert_pre_post_loops--------------------------
 903 // Insert pre and post loops.  If peel_only is set, the pre-loop can not have
 904 // more iterations added.  It acts as a 'peel' only, no lower-bound RCE, no
 905 // alignment.  Useful to unroll loops that do no array accesses.
 906 void PhaseIdealLoop::insert_pre_post_loops( IdealLoopTree *loop, Node_List &old_new, bool peel_only ) {
 907 
 908 #ifndef PRODUCT
 909   if (TraceLoopOpts) {
 910     if (peel_only)
 911       tty->print("PeelMainPost ");
 912     else
 913       tty->print("PreMainPost  ");
 914     loop->dump_head();
 915   }
 916 #endif
 917   C->set_major_progress();
 918 
 919   // Find common pieces of the loop being guarded with pre & post loops
 920   CountedLoopNode *main_head = loop->_head->as_CountedLoop();
 921   assert( main_head->is_normal_loop(), "" );
 922   CountedLoopEndNode *main_end = main_head->loopexit();
 923   guarantee(main_end != NULL, "no loop exit node");
 924   assert( main_end->outcnt() == 2, "1 true, 1 false path only" );
 925   uint dd_main_head = dom_depth(main_head);
 926   uint max = main_head->outcnt();
 927 
 928   Node *pre_header= main_head->in(LoopNode::EntryControl);
 929   Node *init      = main_head->init_trip();
 930   Node *incr      = main_end ->incr();
 931   Node *limit     = main_end ->limit();
 932   Node *stride    = main_end ->stride();
 933   Node *cmp       = main_end ->cmp_node();
 934   BoolTest::mask b_test = main_end->test_trip();
 935 
 936   // Need only 1 user of 'bol' because I will be hacking the loop bounds.
 937   Node *bol = main_end->in(CountedLoopEndNode::TestValue);
 938   if( bol->outcnt() != 1 ) {
 939     bol = bol->clone();
 940     register_new_node(bol,main_end->in(CountedLoopEndNode::TestControl));
 941     _igvn.replace_input_of(main_end, CountedLoopEndNode::TestValue, bol);
 942   }
 943   // Need only 1 user of 'cmp' because I will be hacking the loop bounds.
 944   if( cmp->outcnt() != 1 ) {
 945     cmp = cmp->clone();
 946     register_new_node(cmp,main_end->in(CountedLoopEndNode::TestControl));
 947     _igvn.replace_input_of(bol, 1, cmp);
 948   }
 949 
 950   //------------------------------
 951   // Step A: Create Post-Loop.
 952   Node* main_exit = main_end->proj_out(false);
 953   assert( main_exit->Opcode() == Op_IfFalse, "" );
 954   int dd_main_exit = dom_depth(main_exit);
 955 
 956   // Step A1: Clone the loop body.  The clone becomes the post-loop.  The main
 957   // loop pre-header illegally has 2 control users (old & new loops).
 958   clone_loop( loop, old_new, dd_main_exit );
 959   assert( old_new[main_end ->_idx]->Opcode() == Op_CountedLoopEnd, "" );
 960   CountedLoopNode *post_head = old_new[main_head->_idx]->as_CountedLoop();
 961   post_head->set_post_loop(main_head);
 962 
 963   // Reduce the post-loop trip count.
 964   CountedLoopEndNode* post_end = old_new[main_end ->_idx]->as_CountedLoopEnd();
 965   post_end->_prob = PROB_FAIR;
 966 
 967   // Build the main-loop normal exit.
 968   IfFalseNode *new_main_exit = new IfFalseNode(main_end);
 969   _igvn.register_new_node_with_optimizer( new_main_exit );
 970   set_idom(new_main_exit, main_end, dd_main_exit );
 971   set_loop(new_main_exit, loop->_parent);
 972 
 973   // Step A2: Build a zero-trip guard for the post-loop.  After leaving the
 974   // main-loop, the post-loop may not execute at all.  We 'opaque' the incr
 975   // (the main-loop trip-counter exit value) because we will be changing
 976   // the exit value (via unrolling) so we cannot constant-fold away the zero
 977   // trip guard until all unrolling is done.
 978   Node *zer_opaq = new Opaque1Node(C, incr);
 979   Node *zer_cmp  = new CmpINode( zer_opaq, limit );
 980   Node *zer_bol  = new BoolNode( zer_cmp, b_test );
 981   register_new_node( zer_opaq, new_main_exit );
 982   register_new_node( zer_cmp , new_main_exit );
 983   register_new_node( zer_bol , new_main_exit );
 984 
 985   // Build the IfNode
 986   IfNode *zer_iff = new IfNode( new_main_exit, zer_bol, PROB_FAIR, COUNT_UNKNOWN );
 987   _igvn.register_new_node_with_optimizer( zer_iff );
 988   set_idom(zer_iff, new_main_exit, dd_main_exit);
 989   set_loop(zer_iff, loop->_parent);
 990 
 991   // Plug in the false-path, taken if we need to skip post-loop
 992   _igvn.replace_input_of(main_exit, 0, zer_iff);
 993   set_idom(main_exit, zer_iff, dd_main_exit);
 994   set_idom(main_exit->unique_out(), zer_iff, dd_main_exit);
 995   // Make the true-path, must enter the post loop
 996   Node *zer_taken = new IfTrueNode( zer_iff );
 997   _igvn.register_new_node_with_optimizer( zer_taken );
 998   set_idom(zer_taken, zer_iff, dd_main_exit);
 999   set_loop(zer_taken, loop->_parent);
1000   // Plug in the true path
1001   _igvn.hash_delete( post_head );
1002   post_head->set_req(LoopNode::EntryControl, zer_taken);
1003   set_idom(post_head, zer_taken, dd_main_exit);
1004 
1005   Arena *a = Thread::current()->resource_area();
1006   VectorSet visited(a);
1007   Node_Stack clones(a, main_head->back_control()->outcnt());
1008   // Step A3: Make the fall-in values to the post-loop come from the
1009   // fall-out values of the main-loop.
1010   for (DUIterator_Fast imax, i = main_head->fast_outs(imax); i < imax; i++) {
1011     Node* main_phi = main_head->fast_out(i);
1012     if( main_phi->is_Phi() && main_phi->in(0) == main_head && main_phi->outcnt() >0 ) {
1013       Node *post_phi = old_new[main_phi->_idx];
1014       Node *fallmain  = clone_up_backedge_goo(main_head->back_control(),
1015                                               post_head->init_control(),
1016                                               main_phi->in(LoopNode::LoopBackControl),
1017                                               visited, clones);
1018       _igvn.hash_delete(post_phi);
1019       post_phi->set_req( LoopNode::EntryControl, fallmain );
1020     }
1021   }
1022 
1023   // Update local caches for next stanza
1024   main_exit = new_main_exit;
1025 
1026 
1027   //------------------------------
1028   // Step B: Create Pre-Loop.
1029 
1030   // Step B1: Clone the loop body.  The clone becomes the pre-loop.  The main
1031   // loop pre-header illegally has 2 control users (old & new loops).
1032   clone_loop( loop, old_new, dd_main_head );
1033   CountedLoopNode*    pre_head = old_new[main_head->_idx]->as_CountedLoop();
1034   CountedLoopEndNode* pre_end  = old_new[main_end ->_idx]->as_CountedLoopEnd();
1035   pre_head->set_pre_loop(main_head);
1036   Node *pre_incr = old_new[incr->_idx];
1037 
1038   // Reduce the pre-loop trip count.
1039   pre_end->_prob = PROB_FAIR;
1040 
1041   // Find the pre-loop normal exit.
1042   Node* pre_exit = pre_end->proj_out(false);
1043   assert( pre_exit->Opcode() == Op_IfFalse, "" );
1044   IfFalseNode *new_pre_exit = new IfFalseNode(pre_end);
1045   _igvn.register_new_node_with_optimizer( new_pre_exit );
1046   set_idom(new_pre_exit, pre_end, dd_main_head);
1047   set_loop(new_pre_exit, loop->_parent);
1048 
1049   // Step B2: Build a zero-trip guard for the main-loop.  After leaving the
1050   // pre-loop, the main-loop may not execute at all.  Later in life this
1051   // zero-trip guard will become the minimum-trip guard when we unroll
1052   // the main-loop.
1053   Node *min_opaq = new Opaque1Node(C, limit);
1054   Node *min_cmp  = new CmpINode( pre_incr, min_opaq );
1055   Node *min_bol  = new BoolNode( min_cmp, b_test );
1056   register_new_node( min_opaq, new_pre_exit );
1057   register_new_node( min_cmp , new_pre_exit );
1058   register_new_node( min_bol , new_pre_exit );
1059 
1060   // Build the IfNode (assume the main-loop is executed always).
1061   IfNode *min_iff = new IfNode( new_pre_exit, min_bol, PROB_ALWAYS, COUNT_UNKNOWN );
1062   _igvn.register_new_node_with_optimizer( min_iff );
1063   set_idom(min_iff, new_pre_exit, dd_main_head);
1064   set_loop(min_iff, loop->_parent);
1065 
1066   // Plug in the false-path, taken if we need to skip main-loop
1067   _igvn.hash_delete( pre_exit );
1068   pre_exit->set_req(0, min_iff);
1069   set_idom(pre_exit, min_iff, dd_main_head);
1070   set_idom(pre_exit->unique_out(), min_iff, dd_main_head);
1071   // Make the true-path, must enter the main loop
1072   Node *min_taken = new IfTrueNode( min_iff );
1073   _igvn.register_new_node_with_optimizer( min_taken );
1074   set_idom(min_taken, min_iff, dd_main_head);
1075   set_loop(min_taken, loop->_parent);
1076   // Plug in the true path
1077   _igvn.hash_delete( main_head );
1078   main_head->set_req(LoopNode::EntryControl, min_taken);
1079   set_idom(main_head, min_taken, dd_main_head);
1080 
1081   visited.Clear();
1082   clones.clear();
1083   // Step B3: Make the fall-in values to the main-loop come from the
1084   // fall-out values of the pre-loop.
1085   for (DUIterator_Fast i2max, i2 = main_head->fast_outs(i2max); i2 < i2max; i2++) {
1086     Node* main_phi = main_head->fast_out(i2);
1087     if( main_phi->is_Phi() && main_phi->in(0) == main_head && main_phi->outcnt() > 0 ) {
1088       Node *pre_phi = old_new[main_phi->_idx];
1089       Node *fallpre  = clone_up_backedge_goo(pre_head->back_control(),
1090                                              main_head->init_control(),
1091                                              pre_phi->in(LoopNode::LoopBackControl),
1092                                              visited, clones);
1093       _igvn.hash_delete(main_phi);
1094       main_phi->set_req( LoopNode::EntryControl, fallpre );
1095     }
1096   }
1097 
1098   // Nodes inside the loop may be control dependent on a predicate
1099   // that was moved before the preloop. If the back branch of the main
1100   // or post loops becomes dead, those nodes won't be dependent on the
1101   // test that guards that loop nest anymore which could lead to an
1102   // incorrect array access because it executes independently of the
1103   // test that was guarding the loop nest. We add a special CastII on
1104   // the if branch that enters the loop, between the input induction
1105   // variable value and the induction variable Phi to preserve correct
1106   // dependencies.
1107 
1108   // CastII for the post loop:
1109   bool inserted = cast_incr_before_loop(zer_opaq->in(1), zer_taken, post_head);
1110   assert(inserted, "no castII inserted");
1111 
1112   // CastII for the main loop:
1113   inserted = cast_incr_before_loop(pre_incr, min_taken, main_head);
1114   assert(inserted, "no castII inserted");
1115 
1116   // Step B4: Shorten the pre-loop to run only 1 iteration (for now).
1117   // RCE and alignment may change this later.
1118   Node *cmp_end = pre_end->cmp_node();
1119   assert( cmp_end->in(2) == limit, "" );
1120   Node *pre_limit = new AddINode( init, stride );
1121 
1122   // Save the original loop limit in this Opaque1 node for
1123   // use by range check elimination.
1124   Node *pre_opaq  = new Opaque1Node(C, pre_limit, limit);
1125 
1126   register_new_node( pre_limit, pre_head->in(0) );
1127   register_new_node( pre_opaq , pre_head->in(0) );
1128 
1129   // Since no other users of pre-loop compare, I can hack limit directly
1130   assert( cmp_end->outcnt() == 1, "no other users" );
1131   _igvn.hash_delete(cmp_end);
1132   cmp_end->set_req(2, peel_only ? pre_limit : pre_opaq);
1133 
1134   // Special case for not-equal loop bounds:
1135   // Change pre loop test, main loop test, and the
1136   // main loop guard test to use lt or gt depending on stride
1137   // direction:
1138   // positive stride use <
1139   // negative stride use >
1140   //
1141   // not-equal test is kept for post loop to handle case
1142   // when init > limit when stride > 0 (and reverse).
1143 
1144   if (pre_end->in(CountedLoopEndNode::TestValue)->as_Bool()->_test._test == BoolTest::ne) {
1145 
1146     BoolTest::mask new_test = (main_end->stride_con() > 0) ? BoolTest::lt : BoolTest::gt;
1147     // Modify pre loop end condition
1148     Node* pre_bol = pre_end->in(CountedLoopEndNode::TestValue)->as_Bool();
1149     BoolNode* new_bol0 = new BoolNode(pre_bol->in(1), new_test);
1150     register_new_node( new_bol0, pre_head->in(0) );
1151     _igvn.replace_input_of(pre_end, CountedLoopEndNode::TestValue, new_bol0);
1152     // Modify main loop guard condition
1153     assert(min_iff->in(CountedLoopEndNode::TestValue) == min_bol, "guard okay");
1154     BoolNode* new_bol1 = new BoolNode(min_bol->in(1), new_test);
1155     register_new_node( new_bol1, new_pre_exit );
1156     _igvn.hash_delete(min_iff);
1157     min_iff->set_req(CountedLoopEndNode::TestValue, new_bol1);
1158     // Modify main loop end condition
1159     BoolNode* main_bol = main_end->in(CountedLoopEndNode::TestValue)->as_Bool();
1160     BoolNode* new_bol2 = new BoolNode(main_bol->in(1), new_test);
1161     register_new_node( new_bol2, main_end->in(CountedLoopEndNode::TestControl) );
1162     _igvn.replace_input_of(main_end, CountedLoopEndNode::TestValue, new_bol2);
1163   }
1164 
1165   // Flag main loop
1166   main_head->set_main_loop();
1167   if( peel_only ) main_head->set_main_no_pre_loop();
1168 
1169   // Subtract a trip count for the pre-loop.
1170   main_head->set_trip_count(main_head->trip_count() - 1);
1171 
1172   // It's difficult to be precise about the trip-counts
1173   // for the pre/post loops.  They are usually very short,
1174   // so guess that 4 trips is a reasonable value.
1175   post_head->set_profile_trip_cnt(4.0);
1176   pre_head->set_profile_trip_cnt(4.0);
1177 
1178   // Now force out all loop-invariant dominating tests.  The optimizer
1179   // finds some, but we _know_ they are all useless.
1180   peeled_dom_test_elim(loop,old_new);
1181   loop->record_for_igvn();
1182 }
1183 
1184 //------------------------------is_invariant-----------------------------
1185 // Return true if n is invariant
1186 bool IdealLoopTree::is_invariant(Node* n) const {
1187   Node *n_c = _phase->has_ctrl(n) ? _phase->get_ctrl(n) : n;
1188   if (n_c->is_top()) return false;
1189   return !is_member(_phase->get_loop(n_c));
1190 }
1191 
1192 
1193 //------------------------------do_unroll--------------------------------------
1194 // Unroll the loop body one step - make each trip do 2 iterations.
1195 void PhaseIdealLoop::do_unroll( IdealLoopTree *loop, Node_List &old_new, bool adjust_min_trip ) {
1196   assert(LoopUnrollLimit, "");
1197   CountedLoopNode *loop_head = loop->_head->as_CountedLoop();
1198   CountedLoopEndNode *loop_end = loop_head->loopexit();
1199   assert(loop_end, "");
1200 #ifndef PRODUCT
1201   if (PrintOpto && VerifyLoopOptimizations) {
1202     tty->print("Unrolling ");
1203     loop->dump_head();
1204   } else if (TraceLoopOpts) {
1205     if (loop_head->trip_count() < (uint)LoopUnrollLimit) {
1206       tty->print("Unroll %d(%2d) ", loop_head->unrolled_count()*2, loop_head->trip_count());
1207     } else {
1208       tty->print("Unroll %d     ", loop_head->unrolled_count()*2);
1209     }
1210     loop->dump_head();
1211   }
1212 #endif
1213 
1214   // Remember loop node count before unrolling to detect
1215   // if rounds of unroll,optimize are making progress
1216   loop_head->set_node_count_before_unroll(loop->_body.size());
1217 
1218   Node *ctrl  = loop_head->in(LoopNode::EntryControl);
1219   Node *limit = loop_head->limit();
1220   Node *init  = loop_head->init_trip();
1221   Node *stride = loop_head->stride();
1222 
1223   Node *opaq = NULL;
1224   if (adjust_min_trip) {       // If not maximally unrolling, need adjustment
1225     // Search for zero-trip guard.
1226     assert( loop_head->is_main_loop(), "" );
1227     assert( ctrl->Opcode() == Op_IfTrue || ctrl->Opcode() == Op_IfFalse, "" );
1228     Node *iff = ctrl->in(0);
1229     assert( iff->Opcode() == Op_If, "" );
1230     Node *bol = iff->in(1);
1231     assert( bol->Opcode() == Op_Bool, "" );
1232     Node *cmp = bol->in(1);
1233     assert( cmp->Opcode() == Op_CmpI, "" );
1234     opaq = cmp->in(2);
1235     // Occasionally it's possible for a zero-trip guard Opaque1 node to be
1236     // optimized away and then another round of loop opts attempted.
1237     // We can not optimize this particular loop in that case.
1238     if (opaq->Opcode() != Op_Opaque1)
1239       return; // Cannot find zero-trip guard!  Bail out!
1240     // Zero-trip test uses an 'opaque' node which is not shared.
1241     assert(opaq->outcnt() == 1 && opaq->in(1) == limit, "");
1242   }
1243 
1244   C->set_major_progress();
1245 
1246   Node* new_limit = NULL;
1247   if (UnrollLimitCheck) {
1248     int stride_con = stride->get_int();
1249     int stride_p = (stride_con > 0) ? stride_con : -stride_con;
1250     uint old_trip_count = loop_head->trip_count();
1251     // Verify that unroll policy result is still valid.
1252     assert(old_trip_count > 1 &&
1253            (!adjust_min_trip || stride_p <= (1<<3)*loop_head->unrolled_count()), "sanity");
1254 
1255     // Adjust loop limit to keep valid iterations number after unroll.
1256     // Use (limit - stride) instead of (((limit - init)/stride) & (-2))*stride
1257     // which may overflow.
1258     if (!adjust_min_trip) {
1259       assert(old_trip_count > 1 && (old_trip_count & 1) == 0,
1260              "odd trip count for maximally unroll");
1261       // Don't need to adjust limit for maximally unroll since trip count is even.
1262     } else if (loop_head->has_exact_trip_count() && init->is_Con()) {
1263       // Loop's limit is constant. Loop's init could be constant when pre-loop
1264       // become peeled iteration.
1265       jlong init_con = init->get_int();
1266       // We can keep old loop limit if iterations count stays the same:
1267       //   old_trip_count == new_trip_count * 2
1268       // Note: since old_trip_count >= 2 then new_trip_count >= 1
1269       // so we also don't need to adjust zero trip test.
1270       jlong limit_con  = limit->get_int();
1271       // (stride_con*2) not overflow since stride_con <= 8.
1272       int new_stride_con = stride_con * 2;
1273       int stride_m    = new_stride_con - (stride_con > 0 ? 1 : -1);
1274       jlong trip_count = (limit_con - init_con + stride_m)/new_stride_con;
1275       // New trip count should satisfy next conditions.
1276       assert(trip_count > 0 && (julong)trip_count < (julong)max_juint/2, "sanity");
1277       uint new_trip_count = (uint)trip_count;
1278       adjust_min_trip = (old_trip_count != new_trip_count*2);
1279     }
1280 
1281     if (adjust_min_trip) {
1282       // Step 2: Adjust the trip limit if it is called for.
1283       // The adjustment amount is -stride. Need to make sure if the
1284       // adjustment underflows or overflows, then the main loop is skipped.
1285       Node* cmp = loop_end->cmp_node();
1286       assert(cmp->in(2) == limit, "sanity");
1287       assert(opaq != NULL && opaq->in(1) == limit, "sanity");
1288 
1289       // Verify that policy_unroll result is still valid.
1290       const TypeInt* limit_type = _igvn.type(limit)->is_int();
1291       assert(stride_con > 0 && ((limit_type->_hi - stride_con) < limit_type->_hi) ||
1292              stride_con < 0 && ((limit_type->_lo - stride_con) > limit_type->_lo), "sanity");
1293 
1294       if (limit->is_Con()) {
1295         // The check in policy_unroll and the assert above guarantee
1296         // no underflow if limit is constant.
1297         new_limit = _igvn.intcon(limit->get_int() - stride_con);
1298         set_ctrl(new_limit, C->root());
1299       } else {
1300         // Limit is not constant.
1301         if (loop_head->unrolled_count() == 1) { // only for first unroll
1302           // Separate limit by Opaque node in case it is an incremented
1303           // variable from previous loop to avoid using pre-incremented
1304           // value which could increase register pressure.
1305           // Otherwise reorg_offsets() optimization will create a separate
1306           // Opaque node for each use of trip-counter and as result
1307           // zero trip guard limit will be different from loop limit.
1308           assert(has_ctrl(opaq), "should have it");
1309           Node* opaq_ctrl = get_ctrl(opaq);
1310           limit = new Opaque2Node( C, limit );
1311           register_new_node( limit, opaq_ctrl );
1312         }
1313         if (stride_con > 0 && ((limit_type->_lo - stride_con) < limit_type->_lo) ||
1314                    stride_con < 0 && ((limit_type->_hi - stride_con) > limit_type->_hi)) {
1315           // No underflow.
1316           new_limit = new SubINode(limit, stride);
1317         } else {
1318           // (limit - stride) may underflow.
1319           // Clamp the adjustment value with MININT or MAXINT:
1320           //
1321           //   new_limit = limit-stride
1322           //   if (stride > 0)
1323           //     new_limit = (limit < new_limit) ? MININT : new_limit;
1324           //   else
1325           //     new_limit = (limit > new_limit) ? MAXINT : new_limit;
1326           //
1327           BoolTest::mask bt = loop_end->test_trip();
1328           assert(bt == BoolTest::lt || bt == BoolTest::gt, "canonical test is expected");
1329           Node* adj_max = _igvn.intcon((stride_con > 0) ? min_jint : max_jint);
1330           set_ctrl(adj_max, C->root());
1331           Node* old_limit = NULL;
1332           Node* adj_limit = NULL;
1333           Node* bol = limit->is_CMove() ? limit->in(CMoveNode::Condition) : NULL;
1334           if (loop_head->unrolled_count() > 1 &&
1335               limit->is_CMove() && limit->Opcode() == Op_CMoveI &&
1336               limit->in(CMoveNode::IfTrue) == adj_max &&
1337               bol->as_Bool()->_test._test == bt &&
1338               bol->in(1)->Opcode() == Op_CmpI &&
1339               bol->in(1)->in(2) == limit->in(CMoveNode::IfFalse)) {
1340             // Loop was unrolled before.
1341             // Optimize the limit to avoid nested CMove:
1342             // use original limit as old limit.
1343             old_limit = bol->in(1)->in(1);
1344             // Adjust previous adjusted limit.
1345             adj_limit = limit->in(CMoveNode::IfFalse);
1346             adj_limit = new SubINode(adj_limit, stride);
1347           } else {
1348             old_limit = limit;
1349             adj_limit = new SubINode(limit, stride);
1350           }
1351           assert(old_limit != NULL && adj_limit != NULL, "");
1352           register_new_node( adj_limit, ctrl ); // adjust amount
1353           Node* adj_cmp = new CmpINode(old_limit, adj_limit);
1354           register_new_node( adj_cmp, ctrl );
1355           Node* adj_bool = new BoolNode(adj_cmp, bt);
1356           register_new_node( adj_bool, ctrl );
1357           new_limit = new CMoveINode(adj_bool, adj_limit, adj_max, TypeInt::INT);
1358         }
1359         register_new_node(new_limit, ctrl);
1360       }
1361       assert(new_limit != NULL, "");
1362       // Replace in loop test.
1363       assert(loop_end->in(1)->in(1) == cmp, "sanity");
1364       if (cmp->outcnt() == 1 && loop_end->in(1)->outcnt() == 1) {
1365         // Don't need to create new test since only one user.
1366         _igvn.hash_delete(cmp);
1367         cmp->set_req(2, new_limit);
1368       } else {
1369         // Create new test since it is shared.
1370         Node* ctrl2 = loop_end->in(0);
1371         Node* cmp2  = cmp->clone();
1372         cmp2->set_req(2, new_limit);
1373         register_new_node(cmp2, ctrl2);
1374         Node* bol2 = loop_end->in(1)->clone();
1375         bol2->set_req(1, cmp2);
1376         register_new_node(bol2, ctrl2);
1377         _igvn.replace_input_of(loop_end, 1, bol2);
1378       }
1379       // Step 3: Find the min-trip test guaranteed before a 'main' loop.
1380       // Make it a 1-trip test (means at least 2 trips).
1381 
1382       // Guard test uses an 'opaque' node which is not shared.  Hence I
1383       // can edit it's inputs directly.  Hammer in the new limit for the
1384       // minimum-trip guard.
1385       assert(opaq->outcnt() == 1, "");
1386       _igvn.replace_input_of(opaq, 1, new_limit);
1387     }
1388 
1389     // Adjust max trip count. The trip count is intentionally rounded
1390     // down here (e.g. 15-> 7-> 3-> 1) because if we unwittingly over-unroll,
1391     // the main, unrolled, part of the loop will never execute as it is protected
1392     // by the min-trip test.  See bug 4834191 for a case where we over-unrolled
1393     // and later determined that part of the unrolled loop was dead.
1394     loop_head->set_trip_count(old_trip_count / 2);
1395 
1396     // Double the count of original iterations in the unrolled loop body.
1397     loop_head->double_unrolled_count();
1398 
1399   } else { // LoopLimitCheck
1400 
1401     // Adjust max trip count. The trip count is intentionally rounded
1402     // down here (e.g. 15-> 7-> 3-> 1) because if we unwittingly over-unroll,
1403     // the main, unrolled, part of the loop will never execute as it is protected
1404     // by the min-trip test.  See bug 4834191 for a case where we over-unrolled
1405     // and later determined that part of the unrolled loop was dead.
1406     loop_head->set_trip_count(loop_head->trip_count() / 2);
1407 
1408     // Double the count of original iterations in the unrolled loop body.
1409     loop_head->double_unrolled_count();
1410 
1411     // -----------
1412     // Step 2: Cut back the trip counter for an unroll amount of 2.
1413     // Loop will normally trip (limit - init)/stride_con.  Since it's a
1414     // CountedLoop this is exact (stride divides limit-init exactly).
1415     // We are going to double the loop body, so we want to knock off any
1416     // odd iteration: (trip_cnt & ~1).  Then back compute a new limit.
1417     Node *span = new SubINode( limit, init );
1418     register_new_node( span, ctrl );
1419     Node *trip = new DivINode( 0, span, stride );
1420     register_new_node( trip, ctrl );
1421     Node *mtwo = _igvn.intcon(-2);
1422     set_ctrl(mtwo, C->root());
1423     Node *rond = new AndINode( trip, mtwo );
1424     register_new_node( rond, ctrl );
1425     Node *spn2 = new MulINode( rond, stride );
1426     register_new_node( spn2, ctrl );
1427     new_limit = new AddINode( spn2, init );
1428     register_new_node( new_limit, ctrl );
1429 
1430     // Hammer in the new limit
1431     Node *ctrl2 = loop_end->in(0);
1432     Node *cmp2 = new CmpINode( loop_head->incr(), new_limit );
1433     register_new_node( cmp2, ctrl2 );
1434     Node *bol2 = new BoolNode( cmp2, loop_end->test_trip() );
1435     register_new_node( bol2, ctrl2 );
1436     _igvn.replace_input_of(loop_end, CountedLoopEndNode::TestValue, bol2);
1437 
1438     // Step 3: Find the min-trip test guaranteed before a 'main' loop.
1439     // Make it a 1-trip test (means at least 2 trips).
1440     if( adjust_min_trip ) {
1441       assert( new_limit != NULL, "" );
1442       // Guard test uses an 'opaque' node which is not shared.  Hence I
1443       // can edit it's inputs directly.  Hammer in the new limit for the
1444       // minimum-trip guard.
1445       assert( opaq->outcnt() == 1, "" );
1446       _igvn.hash_delete(opaq);
1447       opaq->set_req(1, new_limit);
1448     }
1449   } // LoopLimitCheck
1450 
1451   // ---------
1452   // Step 4: Clone the loop body.  Move it inside the loop.  This loop body
1453   // represents the odd iterations; since the loop trips an even number of
1454   // times its backedge is never taken.  Kill the backedge.
1455   uint dd = dom_depth(loop_head);
1456   clone_loop( loop, old_new, dd );
1457 
1458   // Make backedges of the clone equal to backedges of the original.
1459   // Make the fall-in from the original come from the fall-out of the clone.
1460   for (DUIterator_Fast jmax, j = loop_head->fast_outs(jmax); j < jmax; j++) {
1461     Node* phi = loop_head->fast_out(j);
1462     if( phi->is_Phi() && phi->in(0) == loop_head && phi->outcnt() > 0 ) {
1463       Node *newphi = old_new[phi->_idx];
1464       _igvn.hash_delete( phi );
1465       _igvn.hash_delete( newphi );
1466 
1467       phi   ->set_req(LoopNode::   EntryControl, newphi->in(LoopNode::LoopBackControl));
1468       newphi->set_req(LoopNode::LoopBackControl, phi   ->in(LoopNode::LoopBackControl));
1469       phi   ->set_req(LoopNode::LoopBackControl, C->top());
1470     }
1471   }
1472   Node *clone_head = old_new[loop_head->_idx];
1473   _igvn.hash_delete( clone_head );
1474   loop_head ->set_req(LoopNode::   EntryControl, clone_head->in(LoopNode::LoopBackControl));
1475   clone_head->set_req(LoopNode::LoopBackControl, loop_head ->in(LoopNode::LoopBackControl));
1476   loop_head ->set_req(LoopNode::LoopBackControl, C->top());
1477   loop->_head = clone_head;     // New loop header
1478 
1479   set_idom(loop_head,  loop_head ->in(LoopNode::EntryControl), dd);
1480   set_idom(clone_head, clone_head->in(LoopNode::EntryControl), dd);
1481 
1482   // Kill the clone's backedge
1483   Node *newcle = old_new[loop_end->_idx];
1484   _igvn.hash_delete( newcle );
1485   Node *one = _igvn.intcon(1);
1486   set_ctrl(one, C->root());
1487   newcle->set_req(1, one);
1488   // Force clone into same loop body
1489   uint max = loop->_body.size();
1490   for( uint k = 0; k < max; k++ ) {
1491     Node *old = loop->_body.at(k);
1492     Node *nnn = old_new[old->_idx];
1493     loop->_body.push(nnn);
1494     if (!has_ctrl(old))
1495       set_loop(nnn, loop);
1496   }
1497 
1498   loop->record_for_igvn();
1499 }
1500 
1501 //------------------------------do_maximally_unroll----------------------------
1502 
1503 void PhaseIdealLoop::do_maximally_unroll( IdealLoopTree *loop, Node_List &old_new ) {
1504   CountedLoopNode *cl = loop->_head->as_CountedLoop();
1505   assert(cl->has_exact_trip_count(), "trip count is not exact");
1506   assert(cl->trip_count() > 0, "");
1507 #ifndef PRODUCT
1508   if (TraceLoopOpts) {
1509     tty->print("MaxUnroll  %d ", cl->trip_count());
1510     loop->dump_head();
1511   }
1512 #endif
1513 
1514   // If loop is tripping an odd number of times, peel odd iteration
1515   if ((cl->trip_count() & 1) == 1) {
1516     do_peeling(loop, old_new);
1517   }
1518 
1519   // Now its tripping an even number of times remaining.  Double loop body.
1520   // Do not adjust pre-guards; they are not needed and do not exist.
1521   if (cl->trip_count() > 0) {
1522     assert((cl->trip_count() & 1) == 0, "missed peeling");
1523     do_unroll(loop, old_new, false);
1524   }
1525 }
1526 
1527 //------------------------------dominates_backedge---------------------------------
1528 // Returns true if ctrl is executed on every complete iteration
1529 bool IdealLoopTree::dominates_backedge(Node* ctrl) {
1530   assert(ctrl->is_CFG(), "must be control");
1531   Node* backedge = _head->as_Loop()->in(LoopNode::LoopBackControl);
1532   return _phase->dom_lca_internal(ctrl, backedge) == ctrl;
1533 }
1534 
1535 //------------------------------adjust_limit-----------------------------------
1536 // Helper function for add_constraint().
1537 Node* PhaseIdealLoop::adjust_limit(int stride_con, Node * scale, Node *offset, Node *rc_limit, Node *loop_limit, Node *pre_ctrl) {
1538   // Compute "I :: (limit-offset)/scale"
1539   Node *con = new SubINode(rc_limit, offset);
1540   register_new_node(con, pre_ctrl);
1541   Node *X = new DivINode(0, con, scale);
1542   register_new_node(X, pre_ctrl);
1543 
1544   // Adjust loop limit
1545   loop_limit = (stride_con > 0)
1546                ? (Node*)(new MinINode(loop_limit, X))
1547                : (Node*)(new MaxINode(loop_limit, X));
1548   register_new_node(loop_limit, pre_ctrl);
1549   return loop_limit;
1550 }
1551 
1552 //------------------------------add_constraint---------------------------------
1553 // Constrain the main loop iterations so the conditions:
1554 //    low_limit <= scale_con * I + offset  <  upper_limit
1555 // always holds true.  That is, either increase the number of iterations in
1556 // the pre-loop or the post-loop until the condition holds true in the main
1557 // loop.  Stride, scale, offset and limit are all loop invariant.  Further,
1558 // stride and scale are constants (offset and limit often are).
1559 void PhaseIdealLoop::add_constraint( int stride_con, int scale_con, Node *offset, Node *low_limit, Node *upper_limit, Node *pre_ctrl, Node **pre_limit, Node **main_limit ) {
1560   // For positive stride, the pre-loop limit always uses a MAX function
1561   // and the main loop a MIN function.  For negative stride these are
1562   // reversed.
1563 
1564   // Also for positive stride*scale the affine function is increasing, so the
1565   // pre-loop must check for underflow and the post-loop for overflow.
1566   // Negative stride*scale reverses this; pre-loop checks for overflow and
1567   // post-loop for underflow.
1568 
1569   Node *scale = _igvn.intcon(scale_con);
1570   set_ctrl(scale, C->root());
1571 
1572   if ((stride_con^scale_con) >= 0) { // Use XOR to avoid overflow
1573     // The overflow limit: scale*I+offset < upper_limit
1574     // For main-loop compute
1575     //   ( if (scale > 0) /* and stride > 0 */
1576     //       I < (upper_limit-offset)/scale
1577     //     else /* scale < 0 and stride < 0 */
1578     //       I > (upper_limit-offset)/scale
1579     //   )
1580     //
1581     // (upper_limit-offset) may overflow or underflow.
1582     // But it is fine since main loop will either have
1583     // less iterations or will be skipped in such case.
1584     *main_limit = adjust_limit(stride_con, scale, offset, upper_limit, *main_limit, pre_ctrl);
1585 
1586     // The underflow limit: low_limit <= scale*I+offset.
1587     // For pre-loop compute
1588     //   NOT(scale*I+offset >= low_limit)
1589     //   scale*I+offset < low_limit
1590     //   ( if (scale > 0) /* and stride > 0 */
1591     //       I < (low_limit-offset)/scale
1592     //     else /* scale < 0 and stride < 0 */
1593     //       I > (low_limit-offset)/scale
1594     //   )
1595 
1596     if (low_limit->get_int() == -max_jint) {
1597       if (!RangeLimitCheck) return;
1598       // We need this guard when scale*pre_limit+offset >= limit
1599       // due to underflow. So we need execute pre-loop until
1600       // scale*I+offset >= min_int. But (min_int-offset) will
1601       // underflow when offset > 0 and X will be > original_limit
1602       // when stride > 0. To avoid it we replace positive offset with 0.
1603       //
1604       // Also (min_int+1 == -max_int) is used instead of min_int here
1605       // to avoid problem with scale == -1 (min_int/(-1) == min_int).
1606       Node* shift = _igvn.intcon(31);
1607       set_ctrl(shift, C->root());
1608       Node* sign = new RShiftINode(offset, shift);
1609       register_new_node(sign, pre_ctrl);
1610       offset = new AndINode(offset, sign);
1611       register_new_node(offset, pre_ctrl);
1612     } else {
1613       assert(low_limit->get_int() == 0, "wrong low limit for range check");
1614       // The only problem we have here when offset == min_int
1615       // since (0-min_int) == min_int. It may be fine for stride > 0
1616       // but for stride < 0 X will be < original_limit. To avoid it
1617       // max(pre_limit, original_limit) is used in do_range_check().
1618     }
1619     // Pass (-stride) to indicate pre_loop_cond = NOT(main_loop_cond);
1620     *pre_limit = adjust_limit((-stride_con), scale, offset, low_limit, *pre_limit, pre_ctrl);
1621 
1622   } else { // stride_con*scale_con < 0
1623     // For negative stride*scale pre-loop checks for overflow and
1624     // post-loop for underflow.
1625     //
1626     // The overflow limit: scale*I+offset < upper_limit
1627     // For pre-loop compute
1628     //   NOT(scale*I+offset < upper_limit)
1629     //   scale*I+offset >= upper_limit
1630     //   scale*I+offset+1 > upper_limit
1631     //   ( if (scale < 0) /* and stride > 0 */
1632     //       I < (upper_limit-(offset+1))/scale
1633     //     else /* scale > 0 and stride < 0 */
1634     //       I > (upper_limit-(offset+1))/scale
1635     //   )
1636     //
1637     // (upper_limit-offset-1) may underflow or overflow.
1638     // To avoid it min(pre_limit, original_limit) is used
1639     // in do_range_check() for stride > 0 and max() for < 0.
1640     Node *one  = _igvn.intcon(1);
1641     set_ctrl(one, C->root());
1642 
1643     Node *plus_one = new AddINode(offset, one);
1644     register_new_node( plus_one, pre_ctrl );
1645     // Pass (-stride) to indicate pre_loop_cond = NOT(main_loop_cond);
1646     *pre_limit = adjust_limit((-stride_con), scale, plus_one, upper_limit, *pre_limit, pre_ctrl);
1647 
1648     if (low_limit->get_int() == -max_jint) {
1649       if (!RangeLimitCheck) return;
1650       // We need this guard when scale*main_limit+offset >= limit
1651       // due to underflow. So we need execute main-loop while
1652       // scale*I+offset+1 > min_int. But (min_int-offset-1) will
1653       // underflow when (offset+1) > 0 and X will be < main_limit
1654       // when scale < 0 (and stride > 0). To avoid it we replace
1655       // positive (offset+1) with 0.
1656       //
1657       // Also (min_int+1 == -max_int) is used instead of min_int here
1658       // to avoid problem with scale == -1 (min_int/(-1) == min_int).
1659       Node* shift = _igvn.intcon(31);
1660       set_ctrl(shift, C->root());
1661       Node* sign = new RShiftINode(plus_one, shift);
1662       register_new_node(sign, pre_ctrl);
1663       plus_one = new AndINode(plus_one, sign);
1664       register_new_node(plus_one, pre_ctrl);
1665     } else {
1666       assert(low_limit->get_int() == 0, "wrong low limit for range check");
1667       // The only problem we have here when offset == max_int
1668       // since (max_int+1) == min_int and (0-min_int) == min_int.
1669       // But it is fine since main loop will either have
1670       // less iterations or will be skipped in such case.
1671     }
1672     // The underflow limit: low_limit <= scale*I+offset.
1673     // For main-loop compute
1674     //   scale*I+offset+1 > low_limit
1675     //   ( if (scale < 0) /* and stride > 0 */
1676     //       I < (low_limit-(offset+1))/scale
1677     //     else /* scale > 0 and stride < 0 */
1678     //       I > (low_limit-(offset+1))/scale
1679     //   )
1680 
1681     *main_limit = adjust_limit(stride_con, scale, plus_one, low_limit, *main_limit, pre_ctrl);
1682   }
1683 }
1684 
1685 
1686 //------------------------------is_scaled_iv---------------------------------
1687 // Return true if exp is a constant times an induction var
1688 bool PhaseIdealLoop::is_scaled_iv(Node* exp, Node* iv, int* p_scale) {
1689   if (exp == iv) {
1690     if (p_scale != NULL) {
1691       *p_scale = 1;
1692     }
1693     return true;
1694   }
1695   int opc = exp->Opcode();
1696   if (opc == Op_MulI) {
1697     if (exp->in(1) == iv && exp->in(2)->is_Con()) {
1698       if (p_scale != NULL) {
1699         *p_scale = exp->in(2)->get_int();
1700       }
1701       return true;
1702     }
1703     if (exp->in(2) == iv && exp->in(1)->is_Con()) {
1704       if (p_scale != NULL) {
1705         *p_scale = exp->in(1)->get_int();
1706       }
1707       return true;
1708     }
1709   } else if (opc == Op_LShiftI) {
1710     if (exp->in(1) == iv && exp->in(2)->is_Con()) {
1711       if (p_scale != NULL) {
1712         *p_scale = 1 << exp->in(2)->get_int();
1713       }
1714       return true;
1715     }
1716   }
1717   return false;
1718 }
1719 
1720 //-----------------------------is_scaled_iv_plus_offset------------------------------
1721 // Return true if exp is a simple induction variable expression: k1*iv + (invar + k2)
1722 bool PhaseIdealLoop::is_scaled_iv_plus_offset(Node* exp, Node* iv, int* p_scale, Node** p_offset, int depth) {
1723   if (is_scaled_iv(exp, iv, p_scale)) {
1724     if (p_offset != NULL) {
1725       Node *zero = _igvn.intcon(0);
1726       set_ctrl(zero, C->root());
1727       *p_offset = zero;
1728     }
1729     return true;
1730   }
1731   int opc = exp->Opcode();
1732   if (opc == Op_AddI) {
1733     if (is_scaled_iv(exp->in(1), iv, p_scale)) {
1734       if (p_offset != NULL) {
1735         *p_offset = exp->in(2);
1736       }
1737       return true;
1738     }
1739     if (exp->in(2)->is_Con()) {
1740       Node* offset2 = NULL;
1741       if (depth < 2 &&
1742           is_scaled_iv_plus_offset(exp->in(1), iv, p_scale,
1743                                    p_offset != NULL ? &offset2 : NULL, depth+1)) {
1744         if (p_offset != NULL) {
1745           Node *ctrl_off2 = get_ctrl(offset2);
1746           Node* offset = new AddINode(offset2, exp->in(2));
1747           register_new_node(offset, ctrl_off2);
1748           *p_offset = offset;
1749         }
1750         return true;
1751       }
1752     }
1753   } else if (opc == Op_SubI) {
1754     if (is_scaled_iv(exp->in(1), iv, p_scale)) {
1755       if (p_offset != NULL) {
1756         Node *zero = _igvn.intcon(0);
1757         set_ctrl(zero, C->root());
1758         Node *ctrl_off = get_ctrl(exp->in(2));
1759         Node* offset = new SubINode(zero, exp->in(2));
1760         register_new_node(offset, ctrl_off);
1761         *p_offset = offset;
1762       }
1763       return true;
1764     }
1765     if (is_scaled_iv(exp->in(2), iv, p_scale)) {
1766       if (p_offset != NULL) {
1767         *p_scale *= -1;
1768         *p_offset = exp->in(1);
1769       }
1770       return true;
1771     }
1772   }
1773   return false;
1774 }
1775 
1776 //------------------------------do_range_check---------------------------------
1777 // Eliminate range-checks and other trip-counter vs loop-invariant tests.
1778 void PhaseIdealLoop::do_range_check( IdealLoopTree *loop, Node_List &old_new ) {
1779 #ifndef PRODUCT
1780   if (PrintOpto && VerifyLoopOptimizations) {
1781     tty->print("Range Check Elimination ");
1782     loop->dump_head();
1783   } else if (TraceLoopOpts) {
1784     tty->print("RangeCheck   ");
1785     loop->dump_head();
1786   }
1787 #endif
1788   assert(RangeCheckElimination, "");
1789   CountedLoopNode *cl = loop->_head->as_CountedLoop();
1790   assert(cl->is_main_loop(), "");
1791 
1792   // protect against stride not being a constant
1793   if (!cl->stride_is_con())
1794     return;
1795 
1796   // Find the trip counter; we are iteration splitting based on it
1797   Node *trip_counter = cl->phi();
1798   // Find the main loop limit; we will trim it's iterations
1799   // to not ever trip end tests
1800   Node *main_limit = cl->limit();
1801 
1802   // Need to find the main-loop zero-trip guard
1803   Node *ctrl  = cl->in(LoopNode::EntryControl);
1804   assert(ctrl->Opcode() == Op_IfTrue || ctrl->Opcode() == Op_IfFalse, "");
1805   Node *iffm = ctrl->in(0);
1806   assert(iffm->Opcode() == Op_If, "");
1807   Node *bolzm = iffm->in(1);
1808   assert(bolzm->Opcode() == Op_Bool, "");
1809   Node *cmpzm = bolzm->in(1);
1810   assert(cmpzm->is_Cmp(), "");
1811   Node *opqzm = cmpzm->in(2);
1812   // Can not optimize a loop if zero-trip Opaque1 node is optimized
1813   // away and then another round of loop opts attempted.
1814   if (opqzm->Opcode() != Op_Opaque1)
1815     return;
1816   assert(opqzm->in(1) == main_limit, "do not understand situation");
1817 
1818   // Find the pre-loop limit; we will expand it's iterations to
1819   // not ever trip low tests.
1820   Node *p_f = iffm->in(0);
1821   assert(p_f->Opcode() == Op_IfFalse, "");
1822   CountedLoopEndNode *pre_end = p_f->in(0)->as_CountedLoopEnd();
1823   assert(pre_end->loopnode()->is_pre_loop(), "");
1824   Node *pre_opaq1 = pre_end->limit();
1825   // Occasionally it's possible for a pre-loop Opaque1 node to be
1826   // optimized away and then another round of loop opts attempted.
1827   // We can not optimize this particular loop in that case.
1828   if (pre_opaq1->Opcode() != Op_Opaque1)
1829     return;
1830   Opaque1Node *pre_opaq = (Opaque1Node*)pre_opaq1;
1831   Node *pre_limit = pre_opaq->in(1);
1832 
1833   // Where do we put new limit calculations
1834   Node *pre_ctrl = pre_end->loopnode()->in(LoopNode::EntryControl);
1835 
1836   // Ensure the original loop limit is available from the
1837   // pre-loop Opaque1 node.
1838   Node *orig_limit = pre_opaq->original_loop_limit();
1839   if (orig_limit == NULL || _igvn.type(orig_limit) == Type::TOP)
1840     return;
1841 
1842   // Must know if its a count-up or count-down loop
1843 
1844   int stride_con = cl->stride_con();
1845   Node *zero = _igvn.intcon(0);
1846   Node *one  = _igvn.intcon(1);
1847   // Use symmetrical int range [-max_jint,max_jint]
1848   Node *mini = _igvn.intcon(-max_jint);
1849   set_ctrl(zero, C->root());
1850   set_ctrl(one,  C->root());
1851   set_ctrl(mini, C->root());
1852 
1853   // Range checks that do not dominate the loop backedge (ie.
1854   // conditionally executed) can lengthen the pre loop limit beyond
1855   // the original loop limit. To prevent this, the pre limit is
1856   // (for stride > 0) MINed with the original loop limit (MAXed
1857   // stride < 0) when some range_check (rc) is conditionally
1858   // executed.
1859   bool conditional_rc = false;
1860 
1861   // Check loop body for tests of trip-counter plus loop-invariant vs
1862   // loop-invariant.
1863   for( uint i = 0; i < loop->_body.size(); i++ ) {
1864     Node *iff = loop->_body[i];
1865     if( iff->Opcode() == Op_If ) { // Test?
1866 
1867       // Test is an IfNode, has 2 projections.  If BOTH are in the loop
1868       // we need loop unswitching instead of iteration splitting.
1869       Node *exit = loop->is_loop_exit(iff);
1870       if( !exit ) continue;
1871       int flip = (exit->Opcode() == Op_IfTrue) ? 1 : 0;
1872 
1873       // Get boolean condition to test
1874       Node *i1 = iff->in(1);
1875       if( !i1->is_Bool() ) continue;
1876       BoolNode *bol = i1->as_Bool();
1877       BoolTest b_test = bol->_test;
1878       // Flip sense of test if exit condition is flipped
1879       if( flip )
1880         b_test = b_test.negate();
1881 
1882       // Get compare
1883       Node *cmp = bol->in(1);
1884 
1885       // Look for trip_counter + offset vs limit
1886       Node *rc_exp = cmp->in(1);
1887       Node *limit  = cmp->in(2);
1888       jint scale_con= 1;        // Assume trip counter not scaled
1889 
1890       Node *limit_c = get_ctrl(limit);
1891       if( loop->is_member(get_loop(limit_c) ) ) {
1892         // Compare might have operands swapped; commute them
1893         b_test = b_test.commute();
1894         rc_exp = cmp->in(2);
1895         limit  = cmp->in(1);
1896         limit_c = get_ctrl(limit);
1897         if( loop->is_member(get_loop(limit_c) ) )
1898           continue;             // Both inputs are loop varying; cannot RCE
1899       }
1900       // Here we know 'limit' is loop invariant
1901 
1902       // 'limit' maybe pinned below the zero trip test (probably from a
1903       // previous round of rce), in which case, it can't be used in the
1904       // zero trip test expression which must occur before the zero test's if.
1905       if( limit_c == ctrl ) {
1906         continue;  // Don't rce this check but continue looking for other candidates.
1907       }
1908 
1909       // Check for scaled induction variable plus an offset
1910       Node *offset = NULL;
1911 
1912       if (!is_scaled_iv_plus_offset(rc_exp, trip_counter, &scale_con, &offset)) {
1913         continue;
1914       }
1915 
1916       Node *offset_c = get_ctrl(offset);
1917       if( loop->is_member( get_loop(offset_c) ) )
1918         continue;               // Offset is not really loop invariant
1919       // Here we know 'offset' is loop invariant.
1920 
1921       // As above for the 'limit', the 'offset' maybe pinned below the
1922       // zero trip test.
1923       if( offset_c == ctrl ) {
1924         continue; // Don't rce this check but continue looking for other candidates.
1925       }
1926 #ifdef ASSERT
1927       if (TraceRangeLimitCheck) {
1928         tty->print_cr("RC bool node%s", flip ? " flipped:" : ":");
1929         bol->dump(2);
1930       }
1931 #endif
1932       // At this point we have the expression as:
1933       //   scale_con * trip_counter + offset :: limit
1934       // where scale_con, offset and limit are loop invariant.  Trip_counter
1935       // monotonically increases by stride_con, a constant.  Both (or either)
1936       // stride_con and scale_con can be negative which will flip about the
1937       // sense of the test.
1938 
1939       // Adjust pre and main loop limits to guard the correct iteration set
1940       if( cmp->Opcode() == Op_CmpU ) {// Unsigned compare is really 2 tests
1941         if( b_test._test == BoolTest::lt ) { // Range checks always use lt
1942           // The underflow and overflow limits: 0 <= scale*I+offset < limit
1943           add_constraint( stride_con, scale_con, offset, zero, limit, pre_ctrl, &pre_limit, &main_limit );
1944           if (!conditional_rc) {
1945             // (0-offset)/scale could be outside of loop iterations range.
1946             conditional_rc = !loop->dominates_backedge(iff) || RangeLimitCheck;
1947           }
1948         } else {
1949 #ifndef PRODUCT
1950           if( PrintOpto )
1951             tty->print_cr("missed RCE opportunity");
1952 #endif
1953           continue;             // In release mode, ignore it
1954         }
1955       } else {                  // Otherwise work on normal compares
1956         switch( b_test._test ) {
1957         case BoolTest::gt:
1958           // Fall into GE case
1959         case BoolTest::ge:
1960           // Convert (I*scale+offset) >= Limit to (I*(-scale)+(-offset)) <= -Limit
1961           scale_con = -scale_con;
1962           offset = new SubINode( zero, offset );
1963           register_new_node( offset, pre_ctrl );
1964           limit  = new SubINode( zero, limit  );
1965           register_new_node( limit, pre_ctrl );
1966           // Fall into LE case
1967         case BoolTest::le:
1968           if (b_test._test != BoolTest::gt) {
1969             // Convert X <= Y to X < Y+1
1970             limit = new AddINode( limit, one );
1971             register_new_node( limit, pre_ctrl );
1972           }
1973           // Fall into LT case
1974         case BoolTest::lt:
1975           // The underflow and overflow limits: MIN_INT <= scale*I+offset < limit
1976           // Note: (MIN_INT+1 == -MAX_INT) is used instead of MIN_INT here
1977           // to avoid problem with scale == -1: MIN_INT/(-1) == MIN_INT.
1978           add_constraint( stride_con, scale_con, offset, mini, limit, pre_ctrl, &pre_limit, &main_limit );
1979           if (!conditional_rc) {
1980             // ((MIN_INT+1)-offset)/scale could be outside of loop iterations range.
1981             // Note: negative offset is replaced with 0 but (MIN_INT+1)/scale could
1982             // still be outside of loop range.
1983             conditional_rc = !loop->dominates_backedge(iff) || RangeLimitCheck;
1984           }
1985           break;
1986         default:
1987 #ifndef PRODUCT
1988           if( PrintOpto )
1989             tty->print_cr("missed RCE opportunity");
1990 #endif
1991           continue;             // Unhandled case
1992         }
1993       }
1994 
1995       // Kill the eliminated test
1996       C->set_major_progress();
1997       Node *kill_con = _igvn.intcon( 1-flip );
1998       set_ctrl(kill_con, C->root());
1999       _igvn.replace_input_of(iff, 1, kill_con);
2000       // Find surviving projection
2001       assert(iff->is_If(), "");
2002       ProjNode* dp = ((IfNode*)iff)->proj_out(1-flip);
2003       // Find loads off the surviving projection; remove their control edge
2004       for (DUIterator_Fast imax, i = dp->fast_outs(imax); i < imax; i++) {
2005         Node* cd = dp->fast_out(i); // Control-dependent node
2006         if (cd->is_Load() && cd->depends_only_on_test()) {   // Loads can now float around in the loop
2007           // Allow the load to float around in the loop, or before it
2008           // but NOT before the pre-loop.
2009           _igvn.replace_input_of(cd, 0, ctrl); // ctrl, not NULL
2010           --i;
2011           --imax;
2012         }
2013       }
2014 
2015     } // End of is IF
2016 
2017   }
2018 
2019   // Update loop limits
2020   if (conditional_rc) {
2021     pre_limit = (stride_con > 0) ? (Node*)new MinINode(pre_limit, orig_limit)
2022                                  : (Node*)new MaxINode(pre_limit, orig_limit);
2023     register_new_node(pre_limit, pre_ctrl);
2024   }
2025   _igvn.replace_input_of(pre_opaq, 1, pre_limit);
2026 
2027   // Note:: we are making the main loop limit no longer precise;
2028   // need to round up based on stride.
2029   cl->set_nonexact_trip_count();
2030   if (!LoopLimitCheck && stride_con != 1 && stride_con != -1) { // Cutout for common case
2031     // "Standard" round-up logic:  ([main_limit-init+(y-1)]/y)*y+init
2032     // Hopefully, compiler will optimize for powers of 2.
2033     Node *ctrl = get_ctrl(main_limit);
2034     Node *stride = cl->stride();
2035     Node *init = cl->init_trip();
2036     Node *span = new SubINode(main_limit,init);
2037     register_new_node(span,ctrl);
2038     Node *rndup = _igvn.intcon(stride_con + ((stride_con>0)?-1:1));
2039     Node *add = new AddINode(span,rndup);
2040     register_new_node(add,ctrl);
2041     Node *div = new DivINode(0,add,stride);
2042     register_new_node(div,ctrl);
2043     Node *mul = new MulINode(div,stride);
2044     register_new_node(mul,ctrl);
2045     Node *newlim = new AddINode(mul,init);
2046     register_new_node(newlim,ctrl);
2047     main_limit = newlim;
2048   }
2049 
2050   Node *main_cle = cl->loopexit();
2051   Node *main_bol = main_cle->in(1);
2052   // Hacking loop bounds; need private copies of exit test
2053   if( main_bol->outcnt() > 1 ) {// BoolNode shared?
2054     main_bol = main_bol->clone();// Clone a private BoolNode
2055     register_new_node( main_bol, main_cle->in(0) );
2056     _igvn.replace_input_of(main_cle, 1, main_bol);
2057   }
2058   Node *main_cmp = main_bol->in(1);
2059   if( main_cmp->outcnt() > 1 ) { // CmpNode shared?

2060     main_cmp = main_cmp->clone();// Clone a private CmpNode
2061     register_new_node( main_cmp, main_cle->in(0) );
2062     _igvn.replace_input_of(main_bol, 1, main_cmp);
2063   }
2064   // Hack the now-private loop bounds
2065   _igvn.replace_input_of(main_cmp, 2, main_limit);
2066   // The OpaqueNode is unshared by design
2067   assert( opqzm->outcnt() == 1, "cannot hack shared node" );
2068   _igvn.replace_input_of(opqzm, 1, main_limit);
2069 }
2070 
2071 //------------------------------DCE_loop_body----------------------------------
2072 // Remove simplistic dead code from loop body
2073 void IdealLoopTree::DCE_loop_body() {
2074   for( uint i = 0; i < _body.size(); i++ )
2075     if( _body.at(i)->outcnt() == 0 )
2076       _body.map( i--, _body.pop() );
2077 }
2078 
2079 
2080 //------------------------------adjust_loop_exit_prob--------------------------
2081 // Look for loop-exit tests with the 50/50 (or worse) guesses from the parsing stage.
2082 // Replace with a 1-in-10 exit guess.
2083 void IdealLoopTree::adjust_loop_exit_prob( PhaseIdealLoop *phase ) {
2084   Node *test = tail();
2085   while( test != _head ) {
2086     uint top = test->Opcode();
2087     if( top == Op_IfTrue || top == Op_IfFalse ) {
2088       int test_con = ((ProjNode*)test)->_con;
2089       assert(top == (uint)(test_con? Op_IfTrue: Op_IfFalse), "sanity");
2090       IfNode *iff = test->in(0)->as_If();
2091       if( iff->outcnt() == 2 ) {        // Ignore dead tests
2092         Node *bol = iff->in(1);
2093         if( bol && bol->req() > 1 && bol->in(1) &&
2094             ((bol->in(1)->Opcode() == Op_StorePConditional ) ||
2095              (bol->in(1)->Opcode() == Op_StoreIConditional ) ||
2096              (bol->in(1)->Opcode() == Op_StoreLConditional ) ||
2097              (bol->in(1)->Opcode() == Op_CompareAndSwapI ) ||
2098              (bol->in(1)->Opcode() == Op_CompareAndSwapL ) ||
2099              (bol->in(1)->Opcode() == Op_CompareAndSwapP ) ||
2100              (bol->in(1)->Opcode() == Op_CompareAndSwapN )))
2101           return;               // Allocation loops RARELY take backedge
2102         // Find the OTHER exit path from the IF
2103         Node* ex = iff->proj_out(1-test_con);
2104         float p = iff->_prob;
2105         if( !phase->is_member( this, ex ) && iff->_fcnt == COUNT_UNKNOWN ) {
2106           if( top == Op_IfTrue ) {
2107             if( p < (PROB_FAIR + PROB_UNLIKELY_MAG(3))) {
2108               iff->_prob = PROB_STATIC_FREQUENT;
2109             }
2110           } else {
2111             if( p > (PROB_FAIR - PROB_UNLIKELY_MAG(3))) {
2112               iff->_prob = PROB_STATIC_INFREQUENT;
2113             }
2114           }
2115         }
2116       }
2117     }
2118     test = phase->idom(test);
2119   }
2120 }
2121 
2122 
2123 //------------------------------policy_do_remove_empty_loop--------------------
2124 // Micro-benchmark spamming.  Policy is to always remove empty loops.
2125 // The 'DO' part is to replace the trip counter with the value it will
2126 // have on the last iteration.  This will break the loop.
2127 bool IdealLoopTree::policy_do_remove_empty_loop( PhaseIdealLoop *phase ) {
2128   // Minimum size must be empty loop
2129   if (_body.size() > EMPTY_LOOP_SIZE)
2130     return false;
2131 
2132   if (!_head->is_CountedLoop())
2133     return false;     // Dead loop
2134   CountedLoopNode *cl = _head->as_CountedLoop();
2135   if (!cl->is_valid_counted_loop())
2136     return false; // Malformed loop
2137   if (!phase->is_member(this, phase->get_ctrl(cl->loopexit()->in(CountedLoopEndNode::TestValue))))
2138     return false;             // Infinite loop
2139 
2140 #ifdef ASSERT
2141   // Ensure only one phi which is the iv.
2142   Node* iv = NULL;
2143   for (DUIterator_Fast imax, i = cl->fast_outs(imax); i < imax; i++) {
2144     Node* n = cl->fast_out(i);
2145     if (n->Opcode() == Op_Phi) {
2146       assert(iv == NULL, "Too many phis" );
2147       iv = n;
2148     }
2149   }
2150   assert(iv == cl->phi(), "Wrong phi" );
2151 #endif
2152 
2153   // main and post loops have explicitly created zero trip guard
2154   bool needs_guard = !cl->is_main_loop() && !cl->is_post_loop();
2155   if (needs_guard) {
2156     // Skip guard if values not overlap.
2157     const TypeInt* init_t = phase->_igvn.type(cl->init_trip())->is_int();
2158     const TypeInt* limit_t = phase->_igvn.type(cl->limit())->is_int();
2159     int  stride_con = cl->stride_con();
2160     if (stride_con > 0) {
2161       needs_guard = (init_t->_hi >= limit_t->_lo);
2162     } else {
2163       needs_guard = (init_t->_lo <= limit_t->_hi);
2164     }
2165   }
2166   if (needs_guard) {
2167     // Check for an obvious zero trip guard.
2168     Node* inctrl = PhaseIdealLoop::skip_loop_predicates(cl->in(LoopNode::EntryControl));
2169     if (inctrl->Opcode() == Op_IfTrue) {
2170       // The test should look like just the backedge of a CountedLoop
2171       Node* iff = inctrl->in(0);
2172       if (iff->is_If()) {
2173         Node* bol = iff->in(1);
2174         if (bol->is_Bool() && bol->as_Bool()->_test._test == cl->loopexit()->test_trip()) {
2175           Node* cmp = bol->in(1);
2176           if (cmp->is_Cmp() && cmp->in(1) == cl->init_trip() && cmp->in(2) == cl->limit()) {
2177             needs_guard = false;
2178           }
2179         }
2180       }
2181     }
2182   }
2183 
2184 #ifndef PRODUCT
2185   if (PrintOpto) {
2186     tty->print("Removing empty loop with%s zero trip guard", needs_guard ? "out" : "");
2187     this->dump_head();
2188   } else if (TraceLoopOpts) {
2189     tty->print("Empty with%s zero trip guard   ", needs_guard ? "out" : "");
2190     this->dump_head();
2191   }
2192 #endif
2193 
2194   if (needs_guard) {
2195     // Peel the loop to ensure there's a zero trip guard
2196     Node_List old_new;
2197     phase->do_peeling(this, old_new);
2198   }
2199 
2200   // Replace the phi at loop head with the final value of the last
2201   // iteration.  Then the CountedLoopEnd will collapse (backedge never
2202   // taken) and all loop-invariant uses of the exit values will be correct.
2203   Node *phi = cl->phi();
2204   Node *exact_limit = phase->exact_limit(this);
2205   if (exact_limit != cl->limit()) {
2206     // We also need to replace the original limit to collapse loop exit.
2207     Node* cmp = cl->loopexit()->cmp_node();
2208     assert(cl->limit() == cmp->in(2), "sanity");
2209     phase->_igvn._worklist.push(cmp->in(2)); // put limit on worklist
2210     phase->_igvn.replace_input_of(cmp, 2, exact_limit); // put cmp on worklist
2211   }
2212   // Note: the final value after increment should not overflow since
2213   // counted loop has limit check predicate.
2214   Node *final = new SubINode( exact_limit, cl->stride() );
2215   phase->register_new_node(final,cl->in(LoopNode::EntryControl));
2216   phase->_igvn.replace_node(phi,final);
2217   phase->C->set_major_progress();
2218   return true;
2219 }
2220 
2221 //------------------------------policy_do_one_iteration_loop-------------------
2222 // Convert one iteration loop into normal code.
2223 bool IdealLoopTree::policy_do_one_iteration_loop( PhaseIdealLoop *phase ) {
2224   if (!_head->as_Loop()->is_valid_counted_loop())
2225     return false; // Only for counted loop
2226 
2227   CountedLoopNode *cl = _head->as_CountedLoop();
2228   if (!cl->has_exact_trip_count() || cl->trip_count() != 1) {
2229     return false;
2230   }
2231 
2232 #ifndef PRODUCT
2233   if(TraceLoopOpts) {
2234     tty->print("OneIteration ");
2235     this->dump_head();
2236   }
2237 #endif
2238 
2239   Node *init_n = cl->init_trip();
2240 #ifdef ASSERT
2241   // Loop boundaries should be constant since trip count is exact.
2242   assert(init_n->get_int() + cl->stride_con() >= cl->limit()->get_int(), "should be one iteration");
2243 #endif
2244   // Replace the phi at loop head with the value of the init_trip.
2245   // Then the CountedLoopEnd will collapse (backedge will not be taken)
2246   // and all loop-invariant uses of the exit values will be correct.
2247   phase->_igvn.replace_node(cl->phi(), cl->init_trip());
2248   phase->C->set_major_progress();
2249   return true;
2250 }
2251 
2252 //=============================================================================
2253 //------------------------------iteration_split_impl---------------------------
2254 bool IdealLoopTree::iteration_split_impl( PhaseIdealLoop *phase, Node_List &old_new ) {
2255   // Compute exact loop trip count if possible.
2256   compute_exact_trip_count(phase);
2257 
2258   // Convert one iteration loop into normal code.
2259   if (policy_do_one_iteration_loop(phase))
2260     return true;
2261 
2262   // Check and remove empty loops (spam micro-benchmarks)
2263   if (policy_do_remove_empty_loop(phase))
2264     return true;  // Here we removed an empty loop
2265 
2266   bool should_peel = policy_peeling(phase); // Should we peel?
2267 
2268   bool should_unswitch = policy_unswitching(phase);
2269 
2270   // Non-counted loops may be peeled; exactly 1 iteration is peeled.
2271   // This removes loop-invariant tests (usually null checks).
2272   if (!_head->is_CountedLoop()) { // Non-counted loop
2273     if (PartialPeelLoop && phase->partial_peel(this, old_new)) {
2274       // Partial peel succeeded so terminate this round of loop opts
2275       return false;
2276     }
2277     if (should_peel) {            // Should we peel?
2278 #ifndef PRODUCT
2279       if (PrintOpto) tty->print_cr("should_peel");
2280 #endif
2281       phase->do_peeling(this,old_new);
2282     } else if (should_unswitch) {
2283       phase->do_unswitching(this, old_new);
2284     }
2285     return true;
2286   }
2287   CountedLoopNode *cl = _head->as_CountedLoop();
2288 
2289   if (!cl->is_valid_counted_loop()) return true; // Ignore various kinds of broken loops
2290 
2291   // Do nothing special to pre- and post- loops
2292   if (cl->is_pre_loop() || cl->is_post_loop()) return true;
2293 
2294   // Compute loop trip count from profile data
2295   compute_profile_trip_cnt(phase);
2296 
2297   // Before attempting fancy unrolling, RCE or alignment, see if we want
2298   // to completely unroll this loop or do loop unswitching.
2299   if (cl->is_normal_loop()) {
2300     if (should_unswitch) {
2301       phase->do_unswitching(this, old_new);
2302       return true;
2303     }
2304     bool should_maximally_unroll =  policy_maximally_unroll(phase);
2305     if (should_maximally_unroll) {
2306       // Here we did some unrolling and peeling.  Eventually we will
2307       // completely unroll this loop and it will no longer be a loop.
2308       phase->do_maximally_unroll(this,old_new);
2309       return true;
2310     }
2311   }
2312 
2313   // Skip next optimizations if running low on nodes. Note that
2314   // policy_unswitching and policy_maximally_unroll have this check.
2315   int nodes_left = phase->C->max_node_limit() - phase->C->live_nodes();
2316   if ((int)(2 * _body.size()) > nodes_left) {
2317     return true;
2318   }
2319 
2320   // Counted loops may be peeled, may need some iterations run up
2321   // front for RCE, and may want to align loop refs to a cache
2322   // line.  Thus we clone a full loop up front whose trip count is
2323   // at least 1 (if peeling), but may be several more.
2324 
2325   // The main loop will start cache-line aligned with at least 1
2326   // iteration of the unrolled body (zero-trip test required) and
2327   // will have some range checks removed.
2328 
2329   // A post-loop will finish any odd iterations (leftover after
2330   // unrolling), plus any needed for RCE purposes.
2331 
2332   bool should_unroll = policy_unroll(phase);
2333 
2334   bool should_rce = policy_range_check(phase);
2335 
2336   bool should_align = policy_align(phase);
2337 
2338   // If not RCE'ing (iteration splitting) or Aligning, then we do not
2339   // need a pre-loop.  We may still need to peel an initial iteration but
2340   // we will not be needing an unknown number of pre-iterations.
2341   //
2342   // Basically, if may_rce_align reports FALSE first time through,
2343   // we will not be able to later do RCE or Aligning on this loop.
2344   bool may_rce_align = !policy_peel_only(phase) || should_rce || should_align;
2345 
2346   // If we have any of these conditions (RCE, alignment, unrolling) met, then
2347   // we switch to the pre-/main-/post-loop model.  This model also covers
2348   // peeling.
2349   if (should_rce || should_align || should_unroll) {
2350     if (cl->is_normal_loop())  // Convert to 'pre/main/post' loops
2351       phase->insert_pre_post_loops(this,old_new, !may_rce_align);
2352 
2353     // Adjust the pre- and main-loop limits to let the pre and post loops run
2354     // with full checks, but the main-loop with no checks.  Remove said
2355     // checks from the main body.
2356     if (should_rce)
2357       phase->do_range_check(this,old_new);
2358 
2359     // Double loop body for unrolling.  Adjust the minimum-trip test (will do
2360     // twice as many iterations as before) and the main body limit (only do
2361     // an even number of trips).  If we are peeling, we might enable some RCE
2362     // and we'd rather unroll the post-RCE'd loop SO... do not unroll if
2363     // peeling.
2364     if (should_unroll && !should_peel)
2365       phase->do_unroll(this,old_new, true);
2366 
2367     // Adjust the pre-loop limits to align the main body
2368     // iterations.
2369     if (should_align)
2370       Unimplemented();
2371 
2372   } else {                      // Else we have an unchanged counted loop
2373     if (should_peel)           // Might want to peel but do nothing else
2374       phase->do_peeling(this,old_new);
2375   }
2376   return true;
2377 }
2378 
2379 
2380 //=============================================================================
2381 //------------------------------iteration_split--------------------------------
2382 bool IdealLoopTree::iteration_split( PhaseIdealLoop *phase, Node_List &old_new ) {
2383   // Recursively iteration split nested loops
2384   if (_child && !_child->iteration_split(phase, old_new))
2385     return false;
2386 
2387   // Clean out prior deadwood
2388   DCE_loop_body();
2389 
2390 
2391   // Look for loop-exit tests with my 50/50 guesses from the Parsing stage.
2392   // Replace with a 1-in-10 exit guess.
2393   if (_parent /*not the root loop*/ &&
2394       !_irreducible &&
2395       // Also ignore the occasional dead backedge
2396       !tail()->is_top()) {
2397     adjust_loop_exit_prob(phase);
2398   }
2399 
2400   // Gate unrolling, RCE and peeling efforts.
2401   if (!_child &&                // If not an inner loop, do not split
2402       !_irreducible &&
2403       _allow_optimizations &&
2404       !tail()->is_top()) {     // Also ignore the occasional dead backedge
2405     if (!_has_call) {
2406         if (!iteration_split_impl(phase, old_new)) {
2407           return false;
2408         }
2409     } else if (policy_unswitching(phase)) {
2410       phase->do_unswitching(this, old_new);
2411     }
2412   }
2413 
2414   // Minor offset re-organization to remove loop-fallout uses of
2415   // trip counter when there was no major reshaping.
2416   phase->reorg_offsets(this);
2417 
2418   if (_next && !_next->iteration_split(phase, old_new))
2419     return false;
2420   return true;
2421 }
2422 
2423 
2424 //=============================================================================
2425 // Process all the loops in the loop tree and replace any fill
2426 // patterns with an intrisc version.
2427 bool PhaseIdealLoop::do_intrinsify_fill() {
2428   bool changed = false;
2429   for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
2430     IdealLoopTree* lpt = iter.current();
2431     changed |= intrinsify_fill(lpt);
2432   }
2433   return changed;
2434 }
2435 
2436 
2437 // Examine an inner loop looking for a a single store of an invariant
2438 // value in a unit stride loop,
2439 bool PhaseIdealLoop::match_fill_loop(IdealLoopTree* lpt, Node*& store, Node*& store_value,
2440                                      Node*& shift, Node*& con) {
2441   const char* msg = NULL;
2442   Node* msg_node = NULL;
2443 
2444   store_value = NULL;
2445   con = NULL;
2446   shift = NULL;
2447 
2448   // Process the loop looking for stores.  If there are multiple
2449   // stores or extra control flow give at this point.
2450   CountedLoopNode* head = lpt->_head->as_CountedLoop();
2451   for (uint i = 0; msg == NULL && i < lpt->_body.size(); i++) {
2452     Node* n = lpt->_body.at(i);
2453     if (n->outcnt() == 0) continue; // Ignore dead
2454     if (n->is_Store()) {
2455       if (store != NULL) {
2456         msg = "multiple stores";
2457         break;
2458       }
2459       int opc = n->Opcode();
2460       if (opc == Op_StoreP || opc == Op_StoreN || opc == Op_StoreNKlass || opc == Op_StoreCM) {
2461         msg = "oop fills not handled";
2462         break;
2463       }
2464       Node* value = n->in(MemNode::ValueIn);
2465       if (!lpt->is_invariant(value)) {
2466         msg  = "variant store value";
2467       } else if (!_igvn.type(n->in(MemNode::Address))->isa_aryptr()) {
2468         msg = "not array address";
2469       }
2470       store = n;
2471       store_value = value;
2472     } else if (n->is_If() && n != head->loopexit()) {
2473       msg = "extra control flow";
2474       msg_node = n;
2475     }
2476   }
2477 
2478   if (store == NULL) {
2479     // No store in loop
2480     return false;
2481   }
2482 
2483   if (msg == NULL && head->stride_con() != 1) {
2484     // could handle negative strides too
2485     if (head->stride_con() < 0) {
2486       msg = "negative stride";
2487     } else {
2488       msg = "non-unit stride";
2489     }
2490   }
2491 
2492   if (msg == NULL && !store->in(MemNode::Address)->is_AddP()) {
2493     msg = "can't handle store address";
2494     msg_node = store->in(MemNode::Address);
2495   }
2496 
2497   if (msg == NULL &&
2498       (!store->in(MemNode::Memory)->is_Phi() ||
2499        store->in(MemNode::Memory)->in(LoopNode::LoopBackControl) != store)) {
2500     msg = "store memory isn't proper phi";
2501     msg_node = store->in(MemNode::Memory);
2502   }
2503 
2504   // Make sure there is an appropriate fill routine
2505   BasicType t = store->as_Mem()->memory_type();
2506   const char* fill_name;
2507   if (msg == NULL &&
2508       StubRoutines::select_fill_function(t, false, fill_name) == NULL) {
2509     msg = "unsupported store";
2510     msg_node = store;
2511   }
2512 
2513   if (msg != NULL) {
2514 #ifndef PRODUCT
2515     if (TraceOptimizeFill) {
2516       tty->print_cr("not fill intrinsic candidate: %s", msg);
2517       if (msg_node != NULL) msg_node->dump();
2518     }
2519 #endif
2520     return false;
2521   }
2522 
2523   // Make sure the address expression can be handled.  It should be
2524   // head->phi * elsize + con.  head->phi might have a ConvI2L.
2525   Node* elements[4];
2526   Node* conv = NULL;
2527   bool found_index = false;
2528   int count = store->in(MemNode::Address)->as_AddP()->unpack_offsets(elements, ARRAY_SIZE(elements));
2529   for (int e = 0; e < count; e++) {
2530     Node* n = elements[e];
2531     if (n->is_Con() && con == NULL) {
2532       con = n;
2533     } else if (n->Opcode() == Op_LShiftX && shift == NULL) {
2534       Node* value = n->in(1);
2535 #ifdef _LP64
2536       if (value->Opcode() == Op_ConvI2L) {
2537         conv = value;
2538         value = value->in(1);
2539       }
2540 #endif
2541       if (value != head->phi()) {
2542         msg = "unhandled shift in address";
2543       } else {
2544         if (type2aelembytes(store->as_Mem()->memory_type(), true) != (1 << n->in(2)->get_int())) {
2545           msg = "scale doesn't match";
2546         } else {
2547           found_index = true;
2548           shift = n;
2549         }
2550       }
2551     } else if (n->Opcode() == Op_ConvI2L && conv == NULL) {
2552       if (n->in(1) == head->phi()) {
2553         found_index = true;
2554         conv = n;
2555       } else {
2556         msg = "unhandled input to ConvI2L";
2557       }
2558     } else if (n == head->phi()) {
2559       // no shift, check below for allowed cases
2560       found_index = true;
2561     } else {
2562       msg = "unhandled node in address";
2563       msg_node = n;
2564     }
2565   }
2566 
2567   if (count == -1) {
2568     msg = "malformed address expression";
2569     msg_node = store;
2570   }
2571 
2572   if (!found_index) {
2573     msg = "missing use of index";
2574   }
2575 
2576   // byte sized items won't have a shift
2577   if (msg == NULL && shift == NULL && t != T_BYTE && t != T_BOOLEAN) {
2578     msg = "can't find shift";
2579     msg_node = store;
2580   }
2581 
2582   if (msg != NULL) {
2583 #ifndef PRODUCT
2584     if (TraceOptimizeFill) {
2585       tty->print_cr("not fill intrinsic: %s", msg);
2586       if (msg_node != NULL) msg_node->dump();
2587     }
2588 #endif
2589     return false;
2590   }
2591 
2592   // No make sure all the other nodes in the loop can be handled
2593   VectorSet ok(Thread::current()->resource_area());
2594 
2595   // store related values are ok
2596   ok.set(store->_idx);
2597   ok.set(store->in(MemNode::Memory)->_idx);
2598 
2599   CountedLoopEndNode* loop_exit = head->loopexit();
2600   guarantee(loop_exit != NULL, "no loop exit node");
2601 
2602   // Loop structure is ok
2603   ok.set(head->_idx);
2604   ok.set(loop_exit->_idx);
2605   ok.set(head->phi()->_idx);
2606   ok.set(head->incr()->_idx);
2607   ok.set(loop_exit->cmp_node()->_idx);
2608   ok.set(loop_exit->in(1)->_idx);
2609 
2610   // Address elements are ok
2611   if (con)   ok.set(con->_idx);
2612   if (shift) ok.set(shift->_idx);
2613   if (conv)  ok.set(conv->_idx);
2614 
2615   for (uint i = 0; msg == NULL && i < lpt->_body.size(); i++) {
2616     Node* n = lpt->_body.at(i);
2617     if (n->outcnt() == 0) continue; // Ignore dead
2618     if (ok.test(n->_idx)) continue;
2619     // Backedge projection is ok
2620     if (n->is_IfTrue() && n->in(0) == loop_exit) continue;
2621     if (!n->is_AddP()) {
2622       msg = "unhandled node";
2623       msg_node = n;
2624       break;
2625     }
2626   }
2627 
2628   // Make sure no unexpected values are used outside the loop
2629   for (uint i = 0; msg == NULL && i < lpt->_body.size(); i++) {
2630     Node* n = lpt->_body.at(i);
2631     // These values can be replaced with other nodes if they are used
2632     // outside the loop.
2633     if (n == store || n == loop_exit || n == head->incr() || n == store->in(MemNode::Memory)) continue;
2634     for (SimpleDUIterator iter(n); iter.has_next(); iter.next()) {
2635       Node* use = iter.get();
2636       if (!lpt->_body.contains(use)) {
2637         msg = "node is used outside loop";
2638         // lpt->_body.dump();
2639         msg_node = n;
2640         break;
2641       }
2642     }
2643   }
2644 
2645 #ifdef ASSERT
2646   if (TraceOptimizeFill) {
2647     if (msg != NULL) {
2648       tty->print_cr("no fill intrinsic: %s", msg);
2649       if (msg_node != NULL) msg_node->dump();
2650     } else {
2651       tty->print_cr("fill intrinsic for:");
2652     }
2653     store->dump();
2654     if (Verbose) {
2655       lpt->_body.dump();
2656     }
2657   }
2658 #endif
2659 
2660   return msg == NULL;
2661 }
2662 
2663 
2664 
2665 bool PhaseIdealLoop::intrinsify_fill(IdealLoopTree* lpt) {
2666   // Only for counted inner loops
2667   if (!lpt->is_counted() || !lpt->is_inner()) {
2668     return false;
2669   }
2670 
2671   // Must have constant stride
2672   CountedLoopNode* head = lpt->_head->as_CountedLoop();
2673   if (!head->is_valid_counted_loop() || !head->is_normal_loop()) {
2674     return false;
2675   }
2676 
2677   // Check that the body only contains a store of a loop invariant
2678   // value that is indexed by the loop phi.
2679   Node* store = NULL;
2680   Node* store_value = NULL;
2681   Node* shift = NULL;
2682   Node* offset = NULL;
2683   if (!match_fill_loop(lpt, store, store_value, shift, offset)) {
2684     return false;
2685   }
2686 
2687 #ifndef PRODUCT
2688   if (TraceLoopOpts) {
2689     tty->print("ArrayFill    ");
2690     lpt->dump_head();
2691   }
2692 #endif
2693 
2694   // Now replace the whole loop body by a call to a fill routine that
2695   // covers the same region as the loop.
2696   Node* base = store->in(MemNode::Address)->as_AddP()->in(AddPNode::Base);
2697 
2698   // Build an expression for the beginning of the copy region
2699   Node* index = head->init_trip();
2700 #ifdef _LP64
2701   index = new ConvI2LNode(index);
2702   _igvn.register_new_node_with_optimizer(index);
2703 #endif
2704   if (shift != NULL) {
2705     // byte arrays don't require a shift but others do.
2706     index = new LShiftXNode(index, shift->in(2));
2707     _igvn.register_new_node_with_optimizer(index);
2708   }
2709   index = new AddPNode(base, base, index);
2710   _igvn.register_new_node_with_optimizer(index);
2711   Node* from = new AddPNode(base, index, offset);
2712   _igvn.register_new_node_with_optimizer(from);
2713   // Compute the number of elements to copy
2714   Node* len = new SubINode(head->limit(), head->init_trip());
2715   _igvn.register_new_node_with_optimizer(len);
2716 
2717   BasicType t = store->as_Mem()->memory_type();
2718   bool aligned = false;
2719   if (offset != NULL && head->init_trip()->is_Con()) {
2720     int element_size = type2aelembytes(t);
2721     aligned = (offset->find_intptr_t_type()->get_con() + head->init_trip()->get_int() * element_size) % HeapWordSize == 0;
2722   }
2723 
2724   // Build a call to the fill routine
2725   const char* fill_name;
2726   address fill = StubRoutines::select_fill_function(t, aligned, fill_name);
2727   assert(fill != NULL, "what?");
2728 
2729   // Convert float/double to int/long for fill routines
2730   if (t == T_FLOAT) {
2731     store_value = new MoveF2INode(store_value);
2732     _igvn.register_new_node_with_optimizer(store_value);
2733   } else if (t == T_DOUBLE) {
2734     store_value = new MoveD2LNode(store_value);
2735     _igvn.register_new_node_with_optimizer(store_value);
2736   }
2737 
2738   if (CCallingConventionRequiresIntsAsLongs &&
2739       // See StubRoutines::select_fill_function for types. FLOAT has been converted to INT.
2740       (t == T_FLOAT || t == T_INT ||  is_subword_type(t))) {
2741     store_value = new ConvI2LNode(store_value);
2742     _igvn.register_new_node_with_optimizer(store_value);
2743   }
2744 
2745   Node* mem_phi = store->in(MemNode::Memory);
2746   Node* result_ctrl;
2747   Node* result_mem;
2748   const TypeFunc* call_type = OptoRuntime::array_fill_Type();
2749   CallLeafNode *call = new CallLeafNoFPNode(call_type, fill,
2750                                             fill_name, TypeAryPtr::get_array_body_type(t));
2751   uint cnt = 0;
2752   call->init_req(TypeFunc::Parms + cnt++, from);
2753   call->init_req(TypeFunc::Parms + cnt++, store_value);
2754   if (CCallingConventionRequiresIntsAsLongs) {
2755     call->init_req(TypeFunc::Parms + cnt++, C->top());
2756   }
2757 #ifdef _LP64
2758   len = new ConvI2LNode(len);
2759   _igvn.register_new_node_with_optimizer(len);
2760 #endif
2761   call->init_req(TypeFunc::Parms + cnt++, len);
2762 #ifdef _LP64
2763   call->init_req(TypeFunc::Parms + cnt++, C->top());
2764 #endif
2765   call->init_req(TypeFunc::Control,   head->init_control());
2766   call->init_req(TypeFunc::I_O,       C->top());       // Does no I/O.
2767   call->init_req(TypeFunc::Memory,    mem_phi->in(LoopNode::EntryControl));
2768   call->init_req(TypeFunc::ReturnAdr, C->start()->proj_out(TypeFunc::ReturnAdr));
2769   call->init_req(TypeFunc::FramePtr,  C->start()->proj_out(TypeFunc::FramePtr));
2770   _igvn.register_new_node_with_optimizer(call);
2771   result_ctrl = new ProjNode(call,TypeFunc::Control);
2772   _igvn.register_new_node_with_optimizer(result_ctrl);
2773   result_mem = new ProjNode(call,TypeFunc::Memory);
2774   _igvn.register_new_node_with_optimizer(result_mem);
2775 
2776 /* Disable following optimization until proper fix (add missing checks).
2777 
2778   // If this fill is tightly coupled to an allocation and overwrites
2779   // the whole body, allow it to take over the zeroing.
2780   AllocateNode* alloc = AllocateNode::Ideal_allocation(base, this);
2781   if (alloc != NULL && alloc->is_AllocateArray()) {
2782     Node* length = alloc->as_AllocateArray()->Ideal_length();
2783     if (head->limit() == length &&
2784         head->init_trip() == _igvn.intcon(0)) {
2785       if (TraceOptimizeFill) {
2786         tty->print_cr("Eliminated zeroing in allocation");
2787       }
2788       alloc->maybe_set_complete(&_igvn);
2789     } else {
2790 #ifdef ASSERT
2791       if (TraceOptimizeFill) {
2792         tty->print_cr("filling array but bounds don't match");
2793         alloc->dump();
2794         head->init_trip()->dump();
2795         head->limit()->dump();
2796         length->dump();
2797       }
2798 #endif
2799     }
2800   }
2801 */
2802 
2803   // Redirect the old control and memory edges that are outside the loop.
2804   Node* exit = head->loopexit()->proj_out(0);
2805   // Sometimes the memory phi of the head is used as the outgoing
2806   // state of the loop.  It's safe in this case to replace it with the
2807   // result_mem.
2808   _igvn.replace_node(store->in(MemNode::Memory), result_mem);
2809   _igvn.replace_node(exit, result_ctrl);
2810   _igvn.replace_node(store, result_mem);
2811   // Any uses the increment outside of the loop become the loop limit.
2812   _igvn.replace_node(head->incr(), head->limit());
2813 
2814   // Disconnect the head from the loop.
2815   for (uint i = 0; i < lpt->_body.size(); i++) {
2816     Node* n = lpt->_body.at(i);
2817     _igvn.replace_node(n, C->top());
2818   }
2819 
2820   return true;
2821 }
--- EOF ---