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/connode.hpp"
  31 #include "opto/divnode.hpp"
  32 #include "opto/loopnode.hpp"
  33 #include "opto/mulnode.hpp"
  34 #include "opto/rootnode.hpp"
  35 #include "opto/runtime.hpp"
  36 #include "opto/subnode.hpp"
  37 
  38 //------------------------------is_loop_exit-----------------------------------
  39 // Given an IfNode, return the loop-exiting projection or NULL if both
  40 // arms remain in the loop.
  41 Node *IdealLoopTree::is_loop_exit(Node *iff) const {
  42   if( iff->outcnt() != 2 ) return NULL; // Ignore partially dead tests
  43   PhaseIdealLoop *phase = _phase;
  44   // Test is an IfNode, has 2 projections.  If BOTH are in the loop
  45   // we need loop unswitching instead of peeling.
  46   if( !is_member(phase->get_loop( iff->raw_out(0) )) )
  47     return iff->raw_out(0);
  48   if( !is_member(phase->get_loop( iff->raw_out(1) )) )
  49     return iff->raw_out(1);
  50   return NULL;
  51 }
  52 
  53 
  54 //=============================================================================
  55 
  56 
  57 //------------------------------record_for_igvn----------------------------
  58 // Put loop body on igvn work list
  59 void IdealLoopTree::record_for_igvn() {
  60   for( uint i = 0; i < _body.size(); i++ ) {
  61     Node *n = _body.at(i);
  62     _phase->_igvn._worklist.push(n);
  63   }
  64 }
  65 
  66 //------------------------------compute_exact_trip_count-----------------------
  67 // Compute loop exact trip count if possible. Do not recalculate trip count for
  68 // split loops (pre-main-post) which have their limits and inits behind Opaque node.
  69 void IdealLoopTree::compute_exact_trip_count( PhaseIdealLoop *phase ) {
  70   if (!_head->as_Loop()->is_valid_counted_loop()) {
  71     return;
  72   }
  73   CountedLoopNode* cl = _head->as_CountedLoop();
  74   // Trip count may become nonexact for iteration split loops since
  75   // RCE modifies limits. Note, _trip_count value is not reset since
  76   // it is used to limit unrolling of main loop.
  77   cl->set_nonexact_trip_count();
  78 
  79   // Loop's test should be part of loop.
  80   if (!phase->is_member(this, phase->get_ctrl(cl->loopexit()->in(CountedLoopEndNode::TestValue))))
  81     return; // Infinite loop
  82 
  83 #ifdef ASSERT
  84   BoolTest::mask bt = cl->loopexit()->test_trip();
  85   assert(bt == BoolTest::lt || bt == BoolTest::gt ||
  86          bt == BoolTest::ne, "canonical test is expected");
  87 #endif
  88 
  89   Node* init_n = cl->init_trip();
  90   Node* limit_n = cl->limit();
  91   if (init_n  != NULL &&  init_n->is_Con() &&
  92       limit_n != NULL && limit_n->is_Con()) {
  93     // Use longs to avoid integer overflow.
  94     int stride_con  = cl->stride_con();
  95     jlong init_con   = cl->init_trip()->get_int();
  96     jlong limit_con  = cl->limit()->get_int();
  97     int stride_m    = stride_con - (stride_con > 0 ? 1 : -1);
  98     jlong trip_count = (limit_con - init_con + stride_m)/stride_con;
  99     if (trip_count > 0 && (julong)trip_count < (julong)max_juint) {
 100       // Set exact trip count.
 101       cl->set_exact_trip_count((uint)trip_count);
 102     }
 103   }
 104 }
 105 
 106 //------------------------------compute_profile_trip_cnt----------------------------
 107 // Compute loop trip count from profile data as
 108 //    (backedge_count + loop_exit_count) / loop_exit_count
 109 void IdealLoopTree::compute_profile_trip_cnt( PhaseIdealLoop *phase ) {
 110   if (!_head->is_CountedLoop()) {
 111     return;
 112   }
 113   CountedLoopNode* head = _head->as_CountedLoop();
 114   if (head->profile_trip_cnt() != COUNT_UNKNOWN) {
 115     return; // Already computed
 116   }
 117   float trip_cnt = (float)max_jint; // default is big
 118 
 119   Node* back = head->in(LoopNode::LoopBackControl);
 120   while (back != head) {
 121     if ((back->Opcode() == Op_IfTrue || back->Opcode() == Op_IfFalse) &&
 122         back->in(0) &&
 123         back->in(0)->is_If() &&
 124         back->in(0)->as_If()->_fcnt != COUNT_UNKNOWN &&
 125         back->in(0)->as_If()->_prob != PROB_UNKNOWN) {
 126       break;
 127     }
 128     back = phase->idom(back);
 129   }
 130   if (back != head) {
 131     assert((back->Opcode() == Op_IfTrue || back->Opcode() == Op_IfFalse) &&
 132            back->in(0), "if-projection exists");
 133     IfNode* back_if = back->in(0)->as_If();
 134     float loop_back_cnt = back_if->_fcnt * back_if->_prob;
 135 
 136     // Now compute a loop exit count
 137     float loop_exit_cnt = 0.0f;
 138     for( uint i = 0; i < _body.size(); i++ ) {
 139       Node *n = _body[i];
 140       if( n->is_If() ) {
 141         IfNode *iff = n->as_If();
 142         if( iff->_fcnt != COUNT_UNKNOWN && iff->_prob != PROB_UNKNOWN ) {
 143           Node *exit = is_loop_exit(iff);
 144           if( exit ) {
 145             float exit_prob = iff->_prob;
 146             if (exit->Opcode() == Op_IfFalse) exit_prob = 1.0 - exit_prob;
 147             if (exit_prob > PROB_MIN) {
 148               float exit_cnt = iff->_fcnt * exit_prob;
 149               loop_exit_cnt += exit_cnt;
 150             }
 151           }
 152         }
 153       }
 154     }
 155     if (loop_exit_cnt > 0.0f) {
 156       trip_cnt = (loop_back_cnt + loop_exit_cnt) / loop_exit_cnt;
 157     } else {
 158       // No exit count so use
 159       trip_cnt = loop_back_cnt;
 160     }
 161   }
 162 #ifndef PRODUCT
 163   if (TraceProfileTripCount) {
 164     tty->print_cr("compute_profile_trip_cnt  lp: %d cnt: %f\n", head->_idx, trip_cnt);
 165   }
 166 #endif
 167   head->set_profile_trip_cnt(trip_cnt);
 168 }
 169 
 170 //---------------------is_invariant_addition-----------------------------
 171 // Return nonzero index of invariant operand for an Add or Sub
 172 // of (nonconstant) invariant and variant values. Helper for reassociate_invariants.
 173 int IdealLoopTree::is_invariant_addition(Node* n, PhaseIdealLoop *phase) {
 174   int op = n->Opcode();
 175   if (op == Op_AddI || op == Op_SubI) {
 176     bool in1_invar = this->is_invariant(n->in(1));
 177     bool in2_invar = this->is_invariant(n->in(2));
 178     if (in1_invar && !in2_invar) return 1;
 179     if (!in1_invar && in2_invar) return 2;
 180   }
 181   return 0;
 182 }
 183 
 184 //---------------------reassociate_add_sub-----------------------------
 185 // Reassociate invariant add and subtract expressions:
 186 //
 187 // inv1 + (x + inv2)  =>  ( inv1 + inv2) + x
 188 // (x + inv2) + inv1  =>  ( inv1 + inv2) + x
 189 // inv1 + (x - inv2)  =>  ( inv1 - inv2) + x
 190 // inv1 - (inv2 - x)  =>  ( inv1 - inv2) + x
 191 // (x + inv2) - inv1  =>  (-inv1 + inv2) + x
 192 // (x - inv2) + inv1  =>  ( inv1 - inv2) + x
 193 // (x - inv2) - inv1  =>  (-inv1 - inv2) + x
 194 // inv1 + (inv2 - x)  =>  ( inv1 + inv2) - x
 195 // inv1 - (x - inv2)  =>  ( inv1 + inv2) - x
 196 // (inv2 - x) + inv1  =>  ( inv1 + inv2) - x
 197 // (inv2 - x) - inv1  =>  (-inv1 + inv2) - x
 198 // inv1 - (x + inv2)  =>  ( inv1 - inv2) - x
 199 //
 200 Node* IdealLoopTree::reassociate_add_sub(Node* n1, PhaseIdealLoop *phase) {
 201   if (!n1->is_Add() && !n1->is_Sub() || n1->outcnt() == 0) return NULL;
 202   if (is_invariant(n1)) return NULL;
 203   int inv1_idx = is_invariant_addition(n1, phase);
 204   if (!inv1_idx) return NULL;
 205   // Don't mess with add of constant (igvn moves them to expression tree root.)
 206   if (n1->is_Add() && n1->in(2)->is_Con()) return NULL;
 207   Node* inv1 = n1->in(inv1_idx);
 208   Node* n2 = n1->in(3 - inv1_idx);
 209   int inv2_idx = is_invariant_addition(n2, phase);
 210   if (!inv2_idx) return NULL;
 211   Node* x    = n2->in(3 - inv2_idx);
 212   Node* inv2 = n2->in(inv2_idx);
 213 
 214   bool neg_x    = n2->is_Sub() && inv2_idx == 1;
 215   bool neg_inv2 = n2->is_Sub() && inv2_idx == 2;
 216   bool neg_inv1 = n1->is_Sub() && inv1_idx == 2;
 217   if (n1->is_Sub() && inv1_idx == 1) {
 218     neg_x    = !neg_x;
 219     neg_inv2 = !neg_inv2;
 220   }
 221   Node* inv1_c = phase->get_ctrl(inv1);
 222   Node* inv2_c = phase->get_ctrl(inv2);
 223   Node* n_inv1;
 224   if (neg_inv1) {
 225     Node *zero = phase->_igvn.intcon(0);
 226     phase->set_ctrl(zero, phase->C->root());
 227     n_inv1 = new (phase->C) SubINode(zero, inv1);
 228     phase->register_new_node(n_inv1, inv1_c);
 229   } else {
 230     n_inv1 = inv1;
 231   }
 232   Node* inv;
 233   if (neg_inv2) {
 234     inv = new (phase->C) SubINode(n_inv1, inv2);
 235   } else {
 236     inv = new (phase->C) AddINode(n_inv1, inv2);
 237   }
 238   phase->register_new_node(inv, phase->get_early_ctrl(inv));
 239 
 240   Node* addx;
 241   if (neg_x) {
 242     addx = new (phase->C) SubINode(inv, x);
 243   } else {
 244     addx = new (phase->C) AddINode(x, inv);
 245   }
 246   phase->register_new_node(addx, phase->get_ctrl(x));
 247   phase->_igvn.replace_node(n1, addx);
 248   assert(phase->get_loop(phase->get_ctrl(n1)) == this, "");
 249   _body.yank(n1);
 250   return addx;
 251 }
 252 
 253 //---------------------reassociate_invariants-----------------------------
 254 // Reassociate invariant expressions:
 255 void IdealLoopTree::reassociate_invariants(PhaseIdealLoop *phase) {
 256   for (int i = _body.size() - 1; i >= 0; i--) {
 257     Node *n = _body.at(i);
 258     for (int j = 0; j < 5; j++) {
 259       Node* nn = reassociate_add_sub(n, phase);
 260       if (nn == NULL) break;
 261       n = nn; // again
 262     };
 263   }
 264 }
 265 
 266 //------------------------------policy_peeling---------------------------------
 267 // Return TRUE or FALSE if the loop should be peeled or not.  Peel if we can
 268 // make some loop-invariant test (usually a null-check) happen before the loop.
 269 bool IdealLoopTree::policy_peeling( PhaseIdealLoop *phase ) const {
 270   Node *test = ((IdealLoopTree*)this)->tail();
 271   int  body_size = ((IdealLoopTree*)this)->_body.size();
 272   // Peeling does loop cloning which can result in O(N^2) node construction
 273   if( body_size > 255 /* Prevent overflow for large body_size */
 274       || (body_size * body_size + phase->C->live_nodes()) > phase->C->max_node_limit() ) {
 275     return false;           // too large to safely clone
 276   }
 277   while( test != _head ) {      // Scan till run off top of loop
 278     if( test->is_If() ) {       // Test?
 279       Node *ctrl = phase->get_ctrl(test->in(1));
 280       if (ctrl->is_top())
 281         return false;           // Found dead test on live IF?  No peeling!
 282       // Standard IF only has one input value to check for loop invariance
 283       assert( test->Opcode() == Op_If || test->Opcode() == Op_CountedLoopEnd, "Check this code when new subtype is added");
 284       // Condition is not a member of this loop?
 285       if( !is_member(phase->get_loop(ctrl)) &&
 286           is_loop_exit(test) )
 287         return true;            // Found reason to peel!
 288     }
 289     // Walk up dominators to loop _head looking for test which is
 290     // executed on every path thru loop.
 291     test = phase->idom(test);
 292   }
 293   return false;
 294 }
 295 
 296 //------------------------------peeled_dom_test_elim---------------------------
 297 // If we got the effect of peeling, either by actually peeling or by making
 298 // a pre-loop which must execute at least once, we can remove all
 299 // loop-invariant dominated tests in the main body.
 300 void PhaseIdealLoop::peeled_dom_test_elim( IdealLoopTree *loop, Node_List &old_new ) {
 301   bool progress = true;
 302   while( progress ) {
 303     progress = false;           // Reset for next iteration
 304     Node *prev = loop->_head->in(LoopNode::LoopBackControl);//loop->tail();
 305     Node *test = prev->in(0);
 306     while( test != loop->_head ) { // Scan till run off top of loop
 307 
 308       int p_op = prev->Opcode();
 309       if( (p_op == Op_IfFalse || p_op == Op_IfTrue) &&
 310           test->is_If() &&      // Test?
 311           !test->in(1)->is_Con() && // And not already obvious?
 312           // Condition is not a member of this loop?
 313           !loop->is_member(get_loop(get_ctrl(test->in(1))))){
 314         // Walk loop body looking for instances of this test
 315         for( uint i = 0; i < loop->_body.size(); i++ ) {
 316           Node *n = loop->_body.at(i);
 317           if( n->is_If() && n->in(1) == test->in(1) /*&& n != loop->tail()->in(0)*/ ) {
 318             // IfNode was dominated by version in peeled loop body
 319             progress = true;
 320             dominated_by( old_new[prev->_idx], n );
 321           }
 322         }
 323       }
 324       prev = test;
 325       test = idom(test);
 326     } // End of scan tests in loop
 327 
 328   } // End of while( progress )
 329 }
 330 
 331 //------------------------------do_peeling-------------------------------------
 332 // Peel the first iteration of the given loop.
 333 // Step 1: Clone the loop body.  The clone becomes the peeled iteration.
 334 //         The pre-loop illegally has 2 control users (old & new loops).
 335 // Step 2: Make the old-loop fall-in edges point to the peeled iteration.
 336 //         Do this by making the old-loop fall-in edges act as if they came
 337 //         around the loopback from the prior iteration (follow the old-loop
 338 //         backedges) and then map to the new peeled iteration.  This leaves
 339 //         the pre-loop with only 1 user (the new peeled iteration), but the
 340 //         peeled-loop backedge has 2 users.
 341 // Step 3: Cut the backedge on the clone (so its not a loop) and remove the
 342 //         extra backedge user.
 343 //
 344 //                   orig
 345 //
 346 //                  stmt1
 347 //                    |
 348 //                    v
 349 //              loop predicate
 350 //                    |
 351 //                    v
 352 //                   loop<----+
 353 //                     |      |
 354 //                   stmt2    |
 355 //                     |      |
 356 //                     v      |
 357 //                    if      ^
 358 //                   / \      |
 359 //                  /   \     |
 360 //                 v     v    |
 361 //               false true   |
 362 //               /       \    |
 363 //              /         ----+
 364 //             |
 365 //             v
 366 //           exit
 367 //
 368 //
 369 //            after clone loop
 370 //
 371 //                   stmt1
 372 //                     |
 373 //                     v
 374 //               loop predicate
 375 //                 /       \
 376 //        clone   /         \   orig
 377 //               /           \
 378 //              /             \
 379 //             v               v
 380 //   +---->loop clone          loop<----+
 381 //   |      |                    |      |
 382 //   |    stmt2 clone          stmt2    |
 383 //   |      |                    |      |
 384 //   |      v                    v      |
 385 //   ^      if clone            If      ^
 386 //   |      / \                / \      |
 387 //   |     /   \              /   \     |
 388 //   |    v     v            v     v    |
 389 //   |    true  false      false true   |
 390 //   |    /         \      /       \    |
 391 //   +----           \    /         ----+
 392 //                    \  /
 393 //                    1v v2
 394 //                  region
 395 //                     |
 396 //                     v
 397 //                   exit
 398 //
 399 //
 400 //         after peel and predicate move
 401 //
 402 //                   stmt1
 403 //                    /
 404 //                   /
 405 //        clone     /            orig
 406 //                 /
 407 //                /              +----------+
 408 //               /               |          |
 409 //              /          loop predicate   |
 410 //             /                 |          |
 411 //            v                  v          |
 412 //   TOP-->loop clone          loop<----+   |
 413 //          |                    |      |   |
 414 //        stmt2 clone          stmt2    |   |
 415 //          |                    |      |   ^
 416 //          v                    v      |   |
 417 //          if clone            If      ^   |
 418 //          / \                / \      |   |
 419 //         /   \              /   \     |   |
 420 //        v     v            v     v    |   |
 421 //      true   false      false  true   |   |
 422 //        |         \      /       \    |   |
 423 //        |          \    /         ----+   ^
 424 //        |           \  /                  |
 425 //        |           1v v2                 |
 426 //        v         region                  |
 427 //        |            |                    |
 428 //        |            v                    |
 429 //        |          exit                   |
 430 //        |                                 |
 431 //        +--------------->-----------------+
 432 //
 433 //
 434 //              final graph
 435 //
 436 //                  stmt1
 437 //                    |
 438 //                    v
 439 //                  stmt2 clone
 440 //                    |
 441 //                    v
 442 //                   if clone
 443 //                  / |
 444 //                 /  |
 445 //                v   v
 446 //            false  true
 447 //             |      |
 448 //             |      v
 449 //             | loop predicate
 450 //             |      |
 451 //             |      v
 452 //             |     loop<----+
 453 //             |      |       |
 454 //             |    stmt2     |
 455 //             |      |       |
 456 //             |      v       |
 457 //             v      if      ^
 458 //             |     /  \     |
 459 //             |    /    \    |
 460 //             |   v     v    |
 461 //             | false  true  |
 462 //             |  |        \  |
 463 //             v  v         --+
 464 //            region
 465 //              |
 466 //              v
 467 //             exit
 468 //
 469 void PhaseIdealLoop::do_peeling( IdealLoopTree *loop, Node_List &old_new ) {
 470 
 471   C->set_major_progress();
 472   // Peeling a 'main' loop in a pre/main/post situation obfuscates the
 473   // 'pre' loop from the main and the 'pre' can no longer have it's
 474   // iterations adjusted.  Therefore, we need to declare this loop as
 475   // no longer a 'main' loop; it will need new pre and post loops before
 476   // we can do further RCE.
 477 #ifndef PRODUCT
 478   if (TraceLoopOpts) {
 479     tty->print("Peel         ");
 480     loop->dump_head();
 481   }
 482 #endif
 483   Node* head = loop->_head;
 484   bool counted_loop = head->is_CountedLoop();
 485   if (counted_loop) {
 486     CountedLoopNode *cl = head->as_CountedLoop();
 487     assert(cl->trip_count() > 0, "peeling a fully unrolled loop");
 488     cl->set_trip_count(cl->trip_count() - 1);
 489     if (cl->is_main_loop()) {
 490       cl->set_normal_loop();
 491 #ifndef PRODUCT
 492       if (PrintOpto && VerifyLoopOptimizations) {
 493         tty->print("Peeling a 'main' loop; resetting to 'normal' ");
 494         loop->dump_head();
 495       }
 496 #endif
 497     }
 498   }
 499   Node* entry = head->in(LoopNode::EntryControl);
 500 
 501   // Step 1: Clone the loop body.  The clone becomes the peeled iteration.
 502   //         The pre-loop illegally has 2 control users (old & new loops).
 503   clone_loop( loop, old_new, dom_depth(head) );
 504 
 505   // Step 2: Make the old-loop fall-in edges point to the peeled iteration.
 506   //         Do this by making the old-loop fall-in edges act as if they came
 507   //         around the loopback from the prior iteration (follow the old-loop
 508   //         backedges) and then map to the new peeled iteration.  This leaves
 509   //         the pre-loop with only 1 user (the new peeled iteration), but the
 510   //         peeled-loop backedge has 2 users.
 511   Node* new_entry = old_new[head->in(LoopNode::LoopBackControl)->_idx];
 512   _igvn.hash_delete(head);
 513   head->set_req(LoopNode::EntryControl, new_entry);
 514   for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) {
 515     Node* old = head->fast_out(j);
 516     if (old->in(0) == loop->_head && old->req() == 3 && old->is_Phi()) {
 517       Node* new_exit_value = old_new[old->in(LoopNode::LoopBackControl)->_idx];
 518       if (!new_exit_value )     // Backedge value is ALSO loop invariant?
 519         // Then loop body backedge value remains the same.
 520         new_exit_value = old->in(LoopNode::LoopBackControl);
 521       _igvn.hash_delete(old);
 522       old->set_req(LoopNode::EntryControl, new_exit_value);
 523     }
 524   }
 525 
 526 
 527   // Step 3: Cut the backedge on the clone (so its not a loop) and remove the
 528   //         extra backedge user.
 529   Node* new_head = old_new[head->_idx];
 530   _igvn.hash_delete(new_head);
 531   new_head->set_req(LoopNode::LoopBackControl, C->top());
 532   for (DUIterator_Fast j2max, j2 = new_head->fast_outs(j2max); j2 < j2max; j2++) {
 533     Node* use = new_head->fast_out(j2);
 534     if (use->in(0) == new_head && use->req() == 3 && use->is_Phi()) {
 535       _igvn.hash_delete(use);
 536       use->set_req(LoopNode::LoopBackControl, C->top());
 537     }
 538   }
 539 
 540 
 541   // Step 4: Correct dom-depth info.  Set to loop-head depth.
 542   int dd = dom_depth(head);
 543   set_idom(head, head->in(1), dd);
 544   for (uint j3 = 0; j3 < loop->_body.size(); j3++) {
 545     Node *old = loop->_body.at(j3);
 546     Node *nnn = old_new[old->_idx];
 547     if (!has_ctrl(nnn))
 548       set_idom(nnn, idom(nnn), dd-1);
 549   }
 550 
 551   // Now force out all loop-invariant dominating tests.  The optimizer
 552   // finds some, but we _know_ they are all useless.
 553   peeled_dom_test_elim(loop,old_new);
 554 
 555   loop->record_for_igvn();
 556 }
 557 
 558 #define EMPTY_LOOP_SIZE 7 // number of nodes in an empty loop
 559 
 560 //------------------------------policy_maximally_unroll------------------------
 561 // Calculate exact loop trip count and return true if loop can be maximally
 562 // unrolled.
 563 bool IdealLoopTree::policy_maximally_unroll( PhaseIdealLoop *phase ) const {
 564   CountedLoopNode *cl = _head->as_CountedLoop();
 565   assert(cl->is_normal_loop(), "");
 566   if (!cl->is_valid_counted_loop())
 567     return false; // Malformed counted loop
 568 
 569   if (!cl->has_exact_trip_count()) {
 570     // Trip count is not exact.
 571     return false;
 572   }
 573 
 574   uint trip_count = cl->trip_count();
 575   // Note, max_juint is used to indicate unknown trip count.
 576   assert(trip_count > 1, "one iteration loop should be optimized out already");
 577   assert(trip_count < max_juint, "exact trip_count should be less than max_uint.");
 578 
 579   // Real policy: if we maximally unroll, does it get too big?
 580   // Allow the unrolled mess to get larger than standard loop
 581   // size.  After all, it will no longer be a loop.
 582   uint body_size    = _body.size();
 583   uint unroll_limit = (uint)LoopUnrollLimit * 4;
 584   assert( (intx)unroll_limit == LoopUnrollLimit * 4, "LoopUnrollLimit must fit in 32bits");
 585   if (trip_count > unroll_limit || body_size > unroll_limit) {
 586     return false;
 587   }
 588 
 589   // Fully unroll a loop with few iterations regardless next
 590   // conditions since following loop optimizations will split
 591   // such loop anyway (pre-main-post).
 592   if (trip_count <= 3)
 593     return true;
 594 
 595   // Take into account that after unroll conjoined heads and tails will fold,
 596   // otherwise policy_unroll() may allow more unrolling than max unrolling.
 597   uint new_body_size = EMPTY_LOOP_SIZE + (body_size - EMPTY_LOOP_SIZE) * trip_count;
 598   uint tst_body_size = (new_body_size - EMPTY_LOOP_SIZE) / trip_count + EMPTY_LOOP_SIZE;
 599   if (body_size != tst_body_size) // Check for int overflow
 600     return false;
 601   if (new_body_size > unroll_limit ||
 602       // Unrolling can result in a large amount of node construction
 603       new_body_size >= phase->C->max_node_limit() - phase->C->live_nodes()) {
 604     return false;
 605   }
 606 
 607   // Do not unroll a loop with String intrinsics code.
 608   // String intrinsics are large and have loops.
 609   for (uint k = 0; k < _body.size(); k++) {
 610     Node* n = _body.at(k);
 611     switch (n->Opcode()) {
 612       case Op_StrComp:
 613       case Op_StrEquals:
 614       case Op_StrIndexOf:
 615       case Op_EncodeISOArray:
 616       case Op_AryEq: {
 617         return false;
 618       }
 619 #if INCLUDE_RTM_OPT
 620       case Op_FastLock:
 621       case Op_FastUnlock: {
 622         // Don't unroll RTM locking code because it is large.
 623         if (UseRTMLocking) {
 624           return false;
 625         }
 626       }
 627 #endif
 628     } // switch
 629   }
 630 
 631   return true; // Do maximally unroll
 632 }
 633 
 634 
 635 //------------------------------policy_unroll----------------------------------
 636 // Return TRUE or FALSE if the loop should be unrolled or not.  Unroll if
 637 // the loop is a CountedLoop and the body is small enough.
 638 bool IdealLoopTree::policy_unroll( PhaseIdealLoop *phase ) const {
 639 
 640   CountedLoopNode *cl = _head->as_CountedLoop();
 641   assert(cl->is_normal_loop() || cl->is_main_loop(), "");
 642 
 643   if (!cl->is_valid_counted_loop())
 644     return false; // Malformed counted loop
 645 
 646   // Protect against over-unrolling.
 647   // After split at least one iteration will be executed in pre-loop.
 648   if (cl->trip_count() <= (uint)(cl->is_normal_loop() ? 2 : 1)) return false;
 649 
 650   int future_unroll_ct = cl->unrolled_count() * 2;
 651   if (future_unroll_ct > LoopMaxUnroll) return false;
 652 
 653   // Check for initial stride being a small enough constant
 654   if (abs(cl->stride_con()) > (1<<2)*future_unroll_ct) return false;
 655 
 656   // Don't unroll if the next round of unrolling would push us
 657   // over the expected trip count of the loop.  One is subtracted
 658   // from the expected trip count because the pre-loop normally
 659   // executes 1 iteration.
 660   if (UnrollLimitForProfileCheck > 0 &&
 661       cl->profile_trip_cnt() != COUNT_UNKNOWN &&
 662       future_unroll_ct        > UnrollLimitForProfileCheck &&
 663       (float)future_unroll_ct > cl->profile_trip_cnt() - 1.0) {
 664     return false;
 665   }
 666 
 667   // When unroll count is greater than LoopUnrollMin, don't unroll if:
 668   //   the residual iterations are more than 10% of the trip count
 669   //   and rounds of "unroll,optimize" are not making significant progress
 670   //   Progress defined as current size less than 20% larger than previous size.
 671   if (UseSuperWord && cl->node_count_before_unroll() > 0 &&
 672       future_unroll_ct > LoopUnrollMin &&
 673       (future_unroll_ct - 1) * 10.0 > cl->profile_trip_cnt() &&
 674       1.2 * cl->node_count_before_unroll() < (double)_body.size()) {
 675     return false;
 676   }
 677 
 678   Node *init_n = cl->init_trip();
 679   Node *limit_n = cl->limit();
 680   int stride_con = cl->stride_con();
 681   // Non-constant bounds.
 682   // Protect against over-unrolling when init or/and limit are not constant
 683   // (so that trip_count's init value is maxint) but iv range is known.
 684   if (init_n   == NULL || !init_n->is_Con()  ||
 685       limit_n  == NULL || !limit_n->is_Con()) {
 686     Node* phi = cl->phi();
 687     if (phi != NULL) {
 688       assert(phi->is_Phi() && phi->in(0) == _head, "Counted loop should have iv phi.");
 689       const TypeInt* iv_type = phase->_igvn.type(phi)->is_int();
 690       int next_stride = stride_con * 2; // stride after this unroll
 691       if (next_stride > 0) {
 692         if (iv_type->_lo + next_stride <= iv_type->_lo || // overflow
 693             iv_type->_lo + next_stride >  iv_type->_hi) {
 694           return false;  // over-unrolling
 695         }
 696       } else if (next_stride < 0) {
 697         if (iv_type->_hi + next_stride >= iv_type->_hi || // overflow
 698             iv_type->_hi + next_stride <  iv_type->_lo) {
 699           return false;  // over-unrolling
 700         }
 701       }
 702     }
 703   }
 704 
 705   // After unroll limit will be adjusted: new_limit = limit-stride.
 706   // Bailout if adjustment overflow.
 707   const TypeInt* limit_type = phase->_igvn.type(limit_n)->is_int();
 708   if (stride_con > 0 && ((limit_type->_hi - stride_con) >= limit_type->_hi) ||
 709       stride_con < 0 && ((limit_type->_lo - stride_con) <= limit_type->_lo))
 710     return false;  // overflow
 711 
 712   // Adjust body_size to determine if we unroll or not
 713   uint body_size = _body.size();
 714   // Key test to unroll loop in CRC32 java code
 715   int xors_in_loop = 0;
 716   // Also count ModL, DivL and MulL which expand mightly
 717   for (uint k = 0; k < _body.size(); k++) {
 718     Node* n = _body.at(k);
 719     switch (n->Opcode()) {
 720       case Op_XorI: xors_in_loop++; break; // CRC32 java code
 721       case Op_ModL: body_size += 30; break;
 722       case Op_DivL: body_size += 30; break;
 723       case Op_MulL: body_size += 10; break;
 724       case Op_StrComp:
 725       case Op_StrEquals:
 726       case Op_StrIndexOf:
 727       case Op_EncodeISOArray:
 728       case Op_AryEq: {
 729         // Do not unroll a loop with String intrinsics code.
 730         // String intrinsics are large and have loops.
 731         return false;
 732       }
 733 #if INCLUDE_RTM_OPT
 734       case Op_FastLock:
 735       case Op_FastUnlock: {
 736         // Don't unroll RTM locking code because it is large.
 737         if (UseRTMLocking) {
 738           return false;
 739         }
 740       }
 741 #endif
 742     } // switch
 743   }
 744 
 745   // Check for being too big
 746   if (body_size > (uint)LoopUnrollLimit) {
 747     if (xors_in_loop >= 4 && body_size < (uint)LoopUnrollLimit*4) return true;
 748     // Normal case: loop too big
 749     return false;
 750   }
 751 
 752   // Unroll once!  (Each trip will soon do double iterations)
 753   return true;
 754 }
 755 
 756 //------------------------------policy_align-----------------------------------
 757 // Return TRUE or FALSE if the loop should be cache-line aligned.  Gather the
 758 // expression that does the alignment.  Note that only one array base can be
 759 // aligned in a loop (unless the VM guarantees mutual alignment).  Note that
 760 // if we vectorize short memory ops into longer memory ops, we may want to
 761 // increase alignment.
 762 bool IdealLoopTree::policy_align( PhaseIdealLoop *phase ) const {
 763   return false;
 764 }
 765 
 766 //------------------------------policy_range_check-----------------------------
 767 // Return TRUE or FALSE if the loop should be range-check-eliminated.
 768 // Actually we do iteration-splitting, a more powerful form of RCE.
 769 bool IdealLoopTree::policy_range_check( PhaseIdealLoop *phase ) const {
 770   if (!RangeCheckElimination) return false;
 771 
 772   CountedLoopNode *cl = _head->as_CountedLoop();
 773   // If we unrolled with no intention of doing RCE and we later
 774   // changed our minds, we got no pre-loop.  Either we need to
 775   // make a new pre-loop, or we gotta disallow RCE.
 776   if (cl->is_main_no_pre_loop()) return false; // Disallowed for now.
 777   Node *trip_counter = cl->phi();
 778 
 779   // Check loop body for tests of trip-counter plus loop-invariant vs
 780   // loop-invariant.
 781   for (uint i = 0; i < _body.size(); i++) {
 782     Node *iff = _body[i];
 783     if (iff->Opcode() == Op_If) { // Test?
 784 
 785       // Comparing trip+off vs limit
 786       Node *bol = iff->in(1);
 787       if (bol->req() != 2) continue; // dead constant test
 788       if (!bol->is_Bool()) {
 789         assert(UseLoopPredicate && bol->Opcode() == Op_Conv2B, "predicate check only");
 790         continue;
 791       }
 792       if (bol->as_Bool()->_test._test == BoolTest::ne)
 793         continue; // not RC
 794 
 795       Node *cmp = bol->in(1);
 796       Node *rc_exp = cmp->in(1);
 797       Node *limit = cmp->in(2);
 798 
 799       Node *limit_c = phase->get_ctrl(limit);
 800       if( limit_c == phase->C->top() )
 801         return false;           // Found dead test on live IF?  No RCE!
 802       if( is_member(phase->get_loop(limit_c) ) ) {
 803         // Compare might have operands swapped; commute them
 804         rc_exp = cmp->in(2);
 805         limit  = cmp->in(1);
 806         limit_c = phase->get_ctrl(limit);
 807         if( is_member(phase->get_loop(limit_c) ) )
 808           continue;             // Both inputs are loop varying; cannot RCE
 809       }
 810 
 811       if (!phase->is_scaled_iv_plus_offset(rc_exp, trip_counter, NULL, NULL)) {
 812         continue;
 813       }
 814       // Yeah!  Found a test like 'trip+off vs limit'
 815       // Test is an IfNode, has 2 projections.  If BOTH are in the loop
 816       // we need loop unswitching instead of iteration splitting.
 817       if( is_loop_exit(iff) )
 818         return true;            // Found reason to split iterations
 819     } // End of is IF
 820   }
 821 
 822   return false;
 823 }
 824 
 825 //------------------------------policy_peel_only-------------------------------
 826 // Return TRUE or FALSE if the loop should NEVER be RCE'd or aligned.  Useful
 827 // for unrolling loops with NO array accesses.
 828 bool IdealLoopTree::policy_peel_only( PhaseIdealLoop *phase ) const {
 829 
 830   for( uint i = 0; i < _body.size(); i++ )
 831     if( _body[i]->is_Mem() )
 832       return false;
 833 
 834   // No memory accesses at all!
 835   return true;
 836 }
 837 
 838 //------------------------------clone_up_backedge_goo--------------------------
 839 // If Node n lives in the back_ctrl block and cannot float, we clone a private
 840 // version of n in preheader_ctrl block and return that, otherwise return n.
 841 Node *PhaseIdealLoop::clone_up_backedge_goo( Node *back_ctrl, Node *preheader_ctrl, Node *n, VectorSet &visited, Node_Stack &clones ) {
 842   if( get_ctrl(n) != back_ctrl ) return n;
 843 
 844   // Only visit once
 845   if (visited.test_set(n->_idx)) {
 846     Node *x = clones.find(n->_idx);
 847     if (x != NULL)
 848       return x;
 849     return n;
 850   }
 851 
 852   Node *x = NULL;               // If required, a clone of 'n'
 853   // Check for 'n' being pinned in the backedge.
 854   if( n->in(0) && n->in(0) == back_ctrl ) {
 855     assert(clones.find(n->_idx) == NULL, "dead loop");
 856     x = n->clone();             // Clone a copy of 'n' to preheader
 857     clones.push(x, n->_idx);
 858     x->set_req( 0, preheader_ctrl ); // Fix x's control input to preheader
 859   }
 860 
 861   // Recursive fixup any other input edges into x.
 862   // If there are no changes we can just return 'n', otherwise
 863   // we need to clone a private copy and change it.
 864   for( uint i = 1; i < n->req(); i++ ) {
 865     Node *g = clone_up_backedge_goo( back_ctrl, preheader_ctrl, n->in(i), visited, clones );
 866     if( g != n->in(i) ) {
 867       if( !x ) {
 868         assert(clones.find(n->_idx) == NULL, "dead loop");
 869         x = n->clone();
 870         clones.push(x, n->_idx);
 871       }
 872       x->set_req(i, g);
 873     }
 874   }
 875   if( x ) {                     // x can legally float to pre-header location
 876     register_new_node( x, preheader_ctrl );
 877     return x;
 878   } else {                      // raise n to cover LCA of uses
 879     set_ctrl( n, find_non_split_ctrl(back_ctrl->in(0)) );
 880   }
 881   return n;
 882 }
 883 
 884 bool PhaseIdealLoop::cast_incr_before_loop(Node* incr, Node* ctrl, Node* loop) {
 885   Node* castii = new (C) CastIINode(incr, TypeInt::INT, true);
 886   castii->set_req(0, ctrl);
 887   register_new_node(castii, ctrl);
 888   for (DUIterator_Fast imax, i = incr->fast_outs(imax); i < imax; i++) {
 889     Node* n = incr->fast_out(i);
 890     if (n->is_Phi() && n->in(0) == loop) {
 891       int nrep = n->replace_edge(incr, castii);
 892       return true;
 893     }
 894   }
 895   return false;
 896 }
 897 
 898 //------------------------------insert_pre_post_loops--------------------------
 899 // Insert pre and post loops.  If peel_only is set, the pre-loop can not have
 900 // more iterations added.  It acts as a 'peel' only, no lower-bound RCE, no
 901 // alignment.  Useful to unroll loops that do no array accesses.
 902 void PhaseIdealLoop::insert_pre_post_loops( IdealLoopTree *loop, Node_List &old_new, bool peel_only ) {
 903 
 904 #ifndef PRODUCT
 905   if (TraceLoopOpts) {
 906     if (peel_only)
 907       tty->print("PeelMainPost ");
 908     else
 909       tty->print("PreMainPost  ");
 910     loop->dump_head();
 911   }
 912 #endif
 913   C->set_major_progress();
 914 
 915   // Find common pieces of the loop being guarded with pre & post loops
 916   CountedLoopNode *main_head = loop->_head->as_CountedLoop();
 917   assert( main_head->is_normal_loop(), "" );
 918   CountedLoopEndNode *main_end = main_head->loopexit();
 919   guarantee(main_end != NULL, "no loop exit node");
 920   assert( main_end->outcnt() == 2, "1 true, 1 false path only" );
 921   uint dd_main_head = dom_depth(main_head);
 922   uint max = main_head->outcnt();
 923 
 924   Node *pre_header= main_head->in(LoopNode::EntryControl);
 925   Node *init      = main_head->init_trip();
 926   Node *incr      = main_end ->incr();
 927   Node *limit     = main_end ->limit();
 928   Node *stride    = main_end ->stride();
 929   Node *cmp       = main_end ->cmp_node();
 930   BoolTest::mask b_test = main_end->test_trip();
 931 
 932   // Need only 1 user of 'bol' because I will be hacking the loop bounds.
 933   Node *bol = main_end->in(CountedLoopEndNode::TestValue);
 934   if( bol->outcnt() != 1 ) {
 935     bol = bol->clone();
 936     register_new_node(bol,main_end->in(CountedLoopEndNode::TestControl));
 937     _igvn.hash_delete(main_end);
 938     main_end->set_req(CountedLoopEndNode::TestValue, bol);
 939   }
 940   // Need only 1 user of 'cmp' because I will be hacking the loop bounds.
 941   if( cmp->outcnt() != 1 ) {
 942     cmp = cmp->clone();
 943     register_new_node(cmp,main_end->in(CountedLoopEndNode::TestControl));
 944     _igvn.hash_delete(bol);
 945     bol->set_req(1, cmp);
 946   }
 947 
 948   //------------------------------
 949   // Step A: Create Post-Loop.
 950   Node* main_exit = main_end->proj_out(false);
 951   assert( main_exit->Opcode() == Op_IfFalse, "" );
 952   int dd_main_exit = dom_depth(main_exit);
 953 
 954   // Step A1: Clone the loop body.  The clone becomes the post-loop.  The main
 955   // loop pre-header illegally has 2 control users (old & new loops).
 956   clone_loop( loop, old_new, dd_main_exit );
 957   assert( old_new[main_end ->_idx]->Opcode() == Op_CountedLoopEnd, "" );
 958   CountedLoopNode *post_head = old_new[main_head->_idx]->as_CountedLoop();
 959   post_head->set_post_loop(main_head);
 960 
 961   // Reduce the post-loop trip count.
 962   CountedLoopEndNode* post_end = old_new[main_end ->_idx]->as_CountedLoopEnd();
 963   post_end->_prob = PROB_FAIR;
 964 
 965   // Build the main-loop normal exit.
 966   IfFalseNode *new_main_exit = new (C) IfFalseNode(main_end);
 967   _igvn.register_new_node_with_optimizer( new_main_exit );
 968   set_idom(new_main_exit, main_end, dd_main_exit );
 969   set_loop(new_main_exit, loop->_parent);
 970 
 971   // Step A2: Build a zero-trip guard for the post-loop.  After leaving the
 972   // main-loop, the post-loop may not execute at all.  We 'opaque' the incr
 973   // (the main-loop trip-counter exit value) because we will be changing
 974   // the exit value (via unrolling) so we cannot constant-fold away the zero
 975   // trip guard until all unrolling is done.
 976   Node *zer_opaq = new (C) Opaque1Node(C, incr);
 977   Node *zer_cmp  = new (C) CmpINode( zer_opaq, limit );
 978   Node *zer_bol  = new (C) BoolNode( zer_cmp, b_test );
 979   register_new_node( zer_opaq, new_main_exit );
 980   register_new_node( zer_cmp , new_main_exit );
 981   register_new_node( zer_bol , new_main_exit );
 982 
 983   // Build the IfNode
 984   IfNode *zer_iff = new (C) IfNode( new_main_exit, zer_bol, PROB_FAIR, COUNT_UNKNOWN );
 985   _igvn.register_new_node_with_optimizer( zer_iff );
 986   set_idom(zer_iff, new_main_exit, dd_main_exit);
 987   set_loop(zer_iff, loop->_parent);
 988 
 989   // Plug in the false-path, taken if we need to skip post-loop
 990   _igvn.replace_input_of(main_exit, 0, zer_iff);
 991   set_idom(main_exit, zer_iff, dd_main_exit);
 992   set_idom(main_exit->unique_out(), zer_iff, dd_main_exit);
 993   // Make the true-path, must enter the post loop
 994   Node *zer_taken = new (C) IfTrueNode( zer_iff );
 995   _igvn.register_new_node_with_optimizer( zer_taken );
 996   set_idom(zer_taken, zer_iff, dd_main_exit);
 997   set_loop(zer_taken, loop->_parent);
 998   // Plug in the true path
 999   _igvn.hash_delete( post_head );
1000   post_head->set_req(LoopNode::EntryControl, zer_taken);
1001   set_idom(post_head, zer_taken, dd_main_exit);
1002 
1003   Arena *a = Thread::current()->resource_area();
1004   VectorSet visited(a);
1005   Node_Stack clones(a, main_head->back_control()->outcnt());
1006   // Step A3: Make the fall-in values to the post-loop come from the
1007   // fall-out values of the main-loop.
1008   for (DUIterator_Fast imax, i = main_head->fast_outs(imax); i < imax; i++) {
1009     Node* main_phi = main_head->fast_out(i);
1010     if( main_phi->is_Phi() && main_phi->in(0) == main_head && main_phi->outcnt() >0 ) {
1011       Node *post_phi = old_new[main_phi->_idx];
1012       Node *fallmain  = clone_up_backedge_goo(main_head->back_control(),
1013                                               post_head->init_control(),
1014                                               main_phi->in(LoopNode::LoopBackControl),
1015                                               visited, clones);
1016       _igvn.hash_delete(post_phi);
1017       post_phi->set_req( LoopNode::EntryControl, fallmain );
1018     }
1019   }
1020 
1021   // Update local caches for next stanza
1022   main_exit = new_main_exit;
1023 
1024 
1025   //------------------------------
1026   // Step B: Create Pre-Loop.
1027 
1028   // Step B1: Clone the loop body.  The clone becomes the pre-loop.  The main
1029   // loop pre-header illegally has 2 control users (old & new loops).
1030   clone_loop( loop, old_new, dd_main_head );
1031   CountedLoopNode*    pre_head = old_new[main_head->_idx]->as_CountedLoop();
1032   CountedLoopEndNode* pre_end  = old_new[main_end ->_idx]->as_CountedLoopEnd();
1033   pre_head->set_pre_loop(main_head);
1034   Node *pre_incr = old_new[incr->_idx];
1035 
1036   // Reduce the pre-loop trip count.
1037   pre_end->_prob = PROB_FAIR;
1038 
1039   // Find the pre-loop normal exit.
1040   Node* pre_exit = pre_end->proj_out(false);
1041   assert( pre_exit->Opcode() == Op_IfFalse, "" );
1042   IfFalseNode *new_pre_exit = new (C) IfFalseNode(pre_end);
1043   _igvn.register_new_node_with_optimizer( new_pre_exit );
1044   set_idom(new_pre_exit, pre_end, dd_main_head);
1045   set_loop(new_pre_exit, loop->_parent);
1046 
1047   // Step B2: Build a zero-trip guard for the main-loop.  After leaving the
1048   // pre-loop, the main-loop may not execute at all.  Later in life this
1049   // zero-trip guard will become the minimum-trip guard when we unroll
1050   // the main-loop.
1051   Node *min_opaq = new (C) Opaque1Node(C, limit);
1052   Node *min_cmp  = new (C) CmpINode( pre_incr, min_opaq );
1053   Node *min_bol  = new (C) BoolNode( min_cmp, b_test );
1054   register_new_node( min_opaq, new_pre_exit );
1055   register_new_node( min_cmp , new_pre_exit );
1056   register_new_node( min_bol , new_pre_exit );
1057 
1058   // Build the IfNode (assume the main-loop is executed always).
1059   IfNode *min_iff = new (C) IfNode( new_pre_exit, min_bol, PROB_ALWAYS, COUNT_UNKNOWN );
1060   _igvn.register_new_node_with_optimizer( min_iff );
1061   set_idom(min_iff, new_pre_exit, dd_main_head);
1062   set_loop(min_iff, loop->_parent);
1063 
1064   // Plug in the false-path, taken if we need to skip main-loop
1065   _igvn.hash_delete( pre_exit );
1066   pre_exit->set_req(0, min_iff);
1067   set_idom(pre_exit, min_iff, dd_main_head);
1068   set_idom(pre_exit->unique_out(), min_iff, dd_main_head);
1069   // Make the true-path, must enter the main loop
1070   Node *min_taken = new (C) IfTrueNode( min_iff );
1071   _igvn.register_new_node_with_optimizer( min_taken );
1072   set_idom(min_taken, min_iff, dd_main_head);
1073   set_loop(min_taken, loop->_parent);
1074   // Plug in the true path
1075   _igvn.hash_delete( main_head );
1076   main_head->set_req(LoopNode::EntryControl, min_taken);
1077   set_idom(main_head, min_taken, dd_main_head);
1078 
1079   visited.Clear();
1080   clones.clear();
1081   // Step B3: Make the fall-in values to the main-loop come from the
1082   // fall-out values of the pre-loop.
1083   for (DUIterator_Fast i2max, i2 = main_head->fast_outs(i2max); i2 < i2max; i2++) {
1084     Node* main_phi = main_head->fast_out(i2);
1085     if( main_phi->is_Phi() && main_phi->in(0) == main_head && main_phi->outcnt() > 0 ) {
1086       Node *pre_phi = old_new[main_phi->_idx];
1087       Node *fallpre  = clone_up_backedge_goo(pre_head->back_control(),
1088                                              main_head->init_control(),
1089                                              pre_phi->in(LoopNode::LoopBackControl),
1090                                              visited, clones);
1091       _igvn.hash_delete(main_phi);
1092       main_phi->set_req( LoopNode::EntryControl, fallpre );
1093     }
1094   }
1095 
1096   // Nodes inside the loop may be control dependent on a predicate
1097   // that was moved before the preloop. If the back branch of the main
1098   // or post loops becomes dead, those nodes won't be dependent on the
1099   // test that guards that loop nest anymore which could lead to an
1100   // incorrect array access because it executes independently of the
1101   // test that was guarding the loop nest. We add a special CastII on
1102   // the if branch that enters the loop, between the input induction
1103   // variable value and the induction variable Phi to preserve correct
1104   // dependencies.
1105 
1106   // CastII for the post loop:
1107   bool inserted = cast_incr_before_loop(zer_opaq->in(1), zer_taken, post_head);
1108   assert(inserted, "no castII inserted");
1109 
1110   // CastII for the main loop:
1111   inserted = cast_incr_before_loop(pre_incr, min_taken, main_head);
1112   assert(inserted, "no castII inserted");
1113 
1114   // Step B4: Shorten the pre-loop to run only 1 iteration (for now).
1115   // RCE and alignment may change this later.
1116   Node *cmp_end = pre_end->cmp_node();
1117   assert( cmp_end->in(2) == limit, "" );
1118   Node *pre_limit = new (C) AddINode( init, stride );
1119 
1120   // Save the original loop limit in this Opaque1 node for
1121   // use by range check elimination.
1122   Node *pre_opaq  = new (C) Opaque1Node(C, pre_limit, limit);
1123 
1124   register_new_node( pre_limit, pre_head->in(0) );
1125   register_new_node( pre_opaq , pre_head->in(0) );
1126 
1127   // Since no other users of pre-loop compare, I can hack limit directly
1128   assert( cmp_end->outcnt() == 1, "no other users" );
1129   _igvn.hash_delete(cmp_end);
1130   cmp_end->set_req(2, peel_only ? pre_limit : pre_opaq);
1131 
1132   // Special case for not-equal loop bounds:
1133   // Change pre loop test, main loop test, and the
1134   // main loop guard test to use lt or gt depending on stride
1135   // direction:
1136   // positive stride use <
1137   // negative stride use >
1138   //
1139   // not-equal test is kept for post loop to handle case
1140   // when init > limit when stride > 0 (and reverse).
1141 
1142   if (pre_end->in(CountedLoopEndNode::TestValue)->as_Bool()->_test._test == BoolTest::ne) {
1143 
1144     BoolTest::mask new_test = (main_end->stride_con() > 0) ? BoolTest::lt : BoolTest::gt;
1145     // Modify pre loop end condition
1146     Node* pre_bol = pre_end->in(CountedLoopEndNode::TestValue)->as_Bool();
1147     BoolNode* new_bol0 = new (C) BoolNode(pre_bol->in(1), new_test);
1148     register_new_node( new_bol0, pre_head->in(0) );
1149     _igvn.hash_delete(pre_end);
1150     pre_end->set_req(CountedLoopEndNode::TestValue, new_bol0);
1151     // Modify main loop guard condition
1152     assert(min_iff->in(CountedLoopEndNode::TestValue) == min_bol, "guard okay");
1153     BoolNode* new_bol1 = new (C) BoolNode(min_bol->in(1), new_test);
1154     register_new_node( new_bol1, new_pre_exit );
1155     _igvn.hash_delete(min_iff);
1156     min_iff->set_req(CountedLoopEndNode::TestValue, new_bol1);
1157     // Modify main loop end condition
1158     BoolNode* main_bol = main_end->in(CountedLoopEndNode::TestValue)->as_Bool();
1159     BoolNode* new_bol2 = new (C) BoolNode(main_bol->in(1), new_test);
1160     register_new_node( new_bol2, main_end->in(CountedLoopEndNode::TestControl) );
1161     _igvn.hash_delete(main_end);
1162     main_end->set_req(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 (C) Opaque2Node( C, limit );
1311           register_new_node( limit, opaq_ctrl );
1312         }
1313         if (stride_con > 0 && (java_subtract(limit_type->_lo, stride_con) < limit_type->_lo) ||
1314             stride_con < 0 && (java_subtract(limit_type->_hi, stride_con) > limit_type->_hi)) {
1315           // No underflow.
1316           new_limit = new (C) 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 (C) SubINode(adj_limit, stride);
1347           } else {
1348             old_limit = limit;
1349             adj_limit = new (C) 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 (C) CmpINode(old_limit, adj_limit);
1354           register_new_node( adj_cmp, ctrl );
1355           Node* adj_bool = new (C) BoolNode(adj_cmp, bt);
1356           register_new_node( adj_bool, ctrl );
1357           new_limit = new (C) 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.hash_delete(loop_end);
1378         loop_end->set_req(1, bol2);
1379       }
1380       // Step 3: Find the min-trip test guaranteed before a 'main' loop.
1381       // Make it a 1-trip test (means at least 2 trips).
1382 
1383       // Guard test uses an 'opaque' node which is not shared.  Hence I
1384       // can edit it's inputs directly.  Hammer in the new limit for the
1385       // minimum-trip guard.
1386       assert(opaq->outcnt() == 1, "");
1387       _igvn.hash_delete(opaq);
1388       opaq->set_req(1, new_limit);
1389     }
1390 
1391     // Adjust max trip count. The trip count is intentionally rounded
1392     // down here (e.g. 15-> 7-> 3-> 1) because if we unwittingly over-unroll,
1393     // the main, unrolled, part of the loop will never execute as it is protected
1394     // by the min-trip test.  See bug 4834191 for a case where we over-unrolled
1395     // and later determined that part of the unrolled loop was dead.
1396     loop_head->set_trip_count(old_trip_count / 2);
1397 
1398     // Double the count of original iterations in the unrolled loop body.
1399     loop_head->double_unrolled_count();
1400 
1401   } else { // LoopLimitCheck
1402 
1403     // Adjust max trip count. The trip count is intentionally rounded
1404     // down here (e.g. 15-> 7-> 3-> 1) because if we unwittingly over-unroll,
1405     // the main, unrolled, part of the loop will never execute as it is protected
1406     // by the min-trip test.  See bug 4834191 for a case where we over-unrolled
1407     // and later determined that part of the unrolled loop was dead.
1408     loop_head->set_trip_count(loop_head->trip_count() / 2);
1409 
1410     // Double the count of original iterations in the unrolled loop body.
1411     loop_head->double_unrolled_count();
1412 
1413     // -----------
1414     // Step 2: Cut back the trip counter for an unroll amount of 2.
1415     // Loop will normally trip (limit - init)/stride_con.  Since it's a
1416     // CountedLoop this is exact (stride divides limit-init exactly).
1417     // We are going to double the loop body, so we want to knock off any
1418     // odd iteration: (trip_cnt & ~1).  Then back compute a new limit.
1419     Node *span = new (C) SubINode( limit, init );
1420     register_new_node( span, ctrl );
1421     Node *trip = new (C) DivINode( 0, span, stride );
1422     register_new_node( trip, ctrl );
1423     Node *mtwo = _igvn.intcon(-2);
1424     set_ctrl(mtwo, C->root());
1425     Node *rond = new (C) AndINode( trip, mtwo );
1426     register_new_node( rond, ctrl );
1427     Node *spn2 = new (C) MulINode( rond, stride );
1428     register_new_node( spn2, ctrl );
1429     new_limit = new (C) AddINode( spn2, init );
1430     register_new_node( new_limit, ctrl );
1431 
1432     // Hammer in the new limit
1433     Node *ctrl2 = loop_end->in(0);
1434     Node *cmp2 = new (C) CmpINode( loop_head->incr(), new_limit );
1435     register_new_node( cmp2, ctrl2 );
1436     Node *bol2 = new (C) BoolNode( cmp2, loop_end->test_trip() );
1437     register_new_node( bol2, ctrl2 );
1438     _igvn.hash_delete(loop_end);
1439     loop_end->set_req(CountedLoopEndNode::TestValue, bol2);
1440 
1441     // Step 3: Find the min-trip test guaranteed before a 'main' loop.
1442     // Make it a 1-trip test (means at least 2 trips).
1443     if( adjust_min_trip ) {
1444       assert( new_limit != NULL, "" );
1445       // Guard test uses an 'opaque' node which is not shared.  Hence I
1446       // can edit it's inputs directly.  Hammer in the new limit for the
1447       // minimum-trip guard.
1448       assert( opaq->outcnt() == 1, "" );
1449       _igvn.hash_delete(opaq);
1450       opaq->set_req(1, new_limit);
1451     }
1452   } // LoopLimitCheck
1453 
1454   // ---------
1455   // Step 4: Clone the loop body.  Move it inside the loop.  This loop body
1456   // represents the odd iterations; since the loop trips an even number of
1457   // times its backedge is never taken.  Kill the backedge.
1458   uint dd = dom_depth(loop_head);
1459   clone_loop( loop, old_new, dd );
1460 
1461   // Make backedges of the clone equal to backedges of the original.
1462   // Make the fall-in from the original come from the fall-out of the clone.
1463   for (DUIterator_Fast jmax, j = loop_head->fast_outs(jmax); j < jmax; j++) {
1464     Node* phi = loop_head->fast_out(j);
1465     if( phi->is_Phi() && phi->in(0) == loop_head && phi->outcnt() > 0 ) {
1466       Node *newphi = old_new[phi->_idx];
1467       _igvn.hash_delete( phi );
1468       _igvn.hash_delete( newphi );
1469 
1470       phi   ->set_req(LoopNode::   EntryControl, newphi->in(LoopNode::LoopBackControl));
1471       newphi->set_req(LoopNode::LoopBackControl, phi   ->in(LoopNode::LoopBackControl));
1472       phi   ->set_req(LoopNode::LoopBackControl, C->top());
1473     }
1474   }
1475   Node *clone_head = old_new[loop_head->_idx];
1476   _igvn.hash_delete( clone_head );
1477   loop_head ->set_req(LoopNode::   EntryControl, clone_head->in(LoopNode::LoopBackControl));
1478   clone_head->set_req(LoopNode::LoopBackControl, loop_head ->in(LoopNode::LoopBackControl));
1479   loop_head ->set_req(LoopNode::LoopBackControl, C->top());
1480   loop->_head = clone_head;     // New loop header
1481 
1482   set_idom(loop_head,  loop_head ->in(LoopNode::EntryControl), dd);
1483   set_idom(clone_head, clone_head->in(LoopNode::EntryControl), dd);
1484 
1485   // Kill the clone's backedge
1486   Node *newcle = old_new[loop_end->_idx];
1487   _igvn.hash_delete( newcle );
1488   Node *one = _igvn.intcon(1);
1489   set_ctrl(one, C->root());
1490   newcle->set_req(1, one);
1491   // Force clone into same loop body
1492   uint max = loop->_body.size();
1493   for( uint k = 0; k < max; k++ ) {
1494     Node *old = loop->_body.at(k);
1495     Node *nnn = old_new[old->_idx];
1496     loop->_body.push(nnn);
1497     if (!has_ctrl(old))
1498       set_loop(nnn, loop);
1499   }
1500 
1501   loop->record_for_igvn();
1502 }
1503 
1504 //------------------------------do_maximally_unroll----------------------------
1505 
1506 void PhaseIdealLoop::do_maximally_unroll( IdealLoopTree *loop, Node_List &old_new ) {
1507   CountedLoopNode *cl = loop->_head->as_CountedLoop();
1508   assert(cl->has_exact_trip_count(), "trip count is not exact");
1509   assert(cl->trip_count() > 0, "");
1510 #ifndef PRODUCT
1511   if (TraceLoopOpts) {
1512     tty->print("MaxUnroll  %d ", cl->trip_count());
1513     loop->dump_head();
1514   }
1515 #endif
1516 
1517   // If loop is tripping an odd number of times, peel odd iteration
1518   if ((cl->trip_count() & 1) == 1) {
1519     do_peeling(loop, old_new);
1520   }
1521 
1522   // Now its tripping an even number of times remaining.  Double loop body.
1523   // Do not adjust pre-guards; they are not needed and do not exist.
1524   if (cl->trip_count() > 0) {
1525     assert((cl->trip_count() & 1) == 0, "missed peeling");
1526     do_unroll(loop, old_new, false);
1527   }
1528 }
1529 
1530 //------------------------------dominates_backedge---------------------------------
1531 // Returns true if ctrl is executed on every complete iteration
1532 bool IdealLoopTree::dominates_backedge(Node* ctrl) {
1533   assert(ctrl->is_CFG(), "must be control");
1534   Node* backedge = _head->as_Loop()->in(LoopNode::LoopBackControl);
1535   return _phase->dom_lca_internal(ctrl, backedge) == ctrl;
1536 }
1537 
1538 //------------------------------adjust_limit-----------------------------------
1539 // Helper function for add_constraint().
1540 Node* PhaseIdealLoop::adjust_limit(int stride_con, Node * scale, Node *offset, Node *rc_limit, Node *loop_limit, Node *pre_ctrl, bool round_up) {
1541   // Compute "I :: (limit-offset)/scale"
1542   Node *con = new (C) SubINode(rc_limit, offset);
1543   register_new_node(con, pre_ctrl);
1544   Node *X = new (C) DivINode(0, con, scale);
1545   register_new_node(X, pre_ctrl);
1546 
1547   // When the absolute value of scale is greater than one, the integer
1548   // division may round limit down so add one to the limit.
1549   if (round_up) {
1550     X = new (C) AddINode(X, _igvn.intcon(1));
1551     register_new_node(X, pre_ctrl);
1552   }
1553 
1554   // Adjust loop limit
1555   loop_limit = (stride_con > 0)
1556                ? (Node*)(new (C) MinINode(loop_limit, X))
1557                : (Node*)(new (C) MaxINode(loop_limit, X));
1558   register_new_node(loop_limit, pre_ctrl);
1559   return loop_limit;
1560 }
1561 
1562 //------------------------------add_constraint---------------------------------
1563 // Constrain the main loop iterations so the conditions:
1564 //    low_limit <= scale_con * I + offset  <  upper_limit
1565 // always holds true.  That is, either increase the number of iterations in
1566 // the pre-loop or the post-loop until the condition holds true in the main
1567 // loop.  Stride, scale, offset and limit are all loop invariant.  Further,
1568 // stride and scale are constants (offset and limit often are).
1569 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 ) {
1570   // For positive stride, the pre-loop limit always uses a MAX function
1571   // and the main loop a MIN function.  For negative stride these are
1572   // reversed.
1573 
1574   // Also for positive stride*scale the affine function is increasing, so the
1575   // pre-loop must check for underflow and the post-loop for overflow.
1576   // Negative stride*scale reverses this; pre-loop checks for overflow and
1577   // post-loop for underflow.
1578 
1579   Node *scale = _igvn.intcon(scale_con);
1580   set_ctrl(scale, C->root());
1581 
1582   if ((stride_con^scale_con) >= 0) { // Use XOR to avoid overflow
1583     // The overflow limit: scale*I+offset < upper_limit
1584     // For main-loop compute
1585     //   ( if (scale > 0) /* and stride > 0 */
1586     //       I < (upper_limit-offset)/scale
1587     //     else /* scale < 0 and stride < 0 */
1588     //       I > (upper_limit-offset)/scale
1589     //   )
1590     //
1591     // (upper_limit-offset) may overflow or underflow.
1592     // But it is fine since main loop will either have
1593     // less iterations or will be skipped in such case.
1594     *main_limit = adjust_limit(stride_con, scale, offset, upper_limit, *main_limit, pre_ctrl, false);
1595 
1596     // The underflow limit: low_limit <= scale*I+offset.
1597     // For pre-loop compute
1598     //   NOT(scale*I+offset >= low_limit)
1599     //   scale*I+offset < low_limit
1600     //   ( if (scale > 0) /* and stride > 0 */
1601     //       I < (low_limit-offset)/scale
1602     //     else /* scale < 0 and stride < 0 */
1603     //       I > (low_limit-offset)/scale
1604     //   )
1605 
1606     if (low_limit->get_int() == -max_jint) {
1607       if (!RangeLimitCheck) return;
1608       // We need this guard when scale*pre_limit+offset >= limit
1609       // due to underflow. So we need execute pre-loop until
1610       // scale*I+offset >= min_int. But (min_int-offset) will
1611       // underflow when offset > 0 and X will be > original_limit
1612       // when stride > 0. To avoid it we replace positive offset with 0.
1613       //
1614       // Also (min_int+1 == -max_int) is used instead of min_int here
1615       // to avoid problem with scale == -1 (min_int/(-1) == min_int).
1616       Node* shift = _igvn.intcon(31);
1617       set_ctrl(shift, C->root());
1618       Node* sign = new (C) RShiftINode(offset, shift);
1619       register_new_node(sign, pre_ctrl);
1620       offset = new (C) AndINode(offset, sign);
1621       register_new_node(offset, pre_ctrl);
1622     } else {
1623       assert(low_limit->get_int() == 0, "wrong low limit for range check");
1624       // The only problem we have here when offset == min_int
1625       // since (0-min_int) == min_int. It may be fine for stride > 0
1626       // but for stride < 0 X will be < original_limit. To avoid it
1627       // max(pre_limit, original_limit) is used in do_range_check().
1628     }
1629     // Pass (-stride) to indicate pre_loop_cond = NOT(main_loop_cond);
1630     *pre_limit = adjust_limit((-stride_con), scale, offset, low_limit, *pre_limit, pre_ctrl,
1631                               scale_con > 1 && stride_con > 0);
1632 
1633   } else { // stride_con*scale_con < 0
1634     // For negative stride*scale pre-loop checks for overflow and
1635     // post-loop for underflow.
1636     //
1637     // The overflow limit: scale*I+offset < upper_limit
1638     // For pre-loop compute
1639     //   NOT(scale*I+offset < upper_limit)
1640     //   scale*I+offset >= upper_limit
1641     //   scale*I+offset+1 > upper_limit
1642     //   ( if (scale < 0) /* and stride > 0 */
1643     //       I < (upper_limit-(offset+1))/scale
1644     //     else /* scale > 0 and stride < 0 */
1645     //       I > (upper_limit-(offset+1))/scale
1646     //   )
1647     //
1648     // (upper_limit-offset-1) may underflow or overflow.
1649     // To avoid it min(pre_limit, original_limit) is used
1650     // in do_range_check() for stride > 0 and max() for < 0.
1651     Node *one  = _igvn.intcon(1);
1652     set_ctrl(one, C->root());
1653 
1654     Node *plus_one = new (C) AddINode(offset, one);
1655     register_new_node( plus_one, pre_ctrl );
1656     // Pass (-stride) to indicate pre_loop_cond = NOT(main_loop_cond);
1657     *pre_limit = adjust_limit((-stride_con), scale, plus_one, upper_limit, *pre_limit, pre_ctrl,
1658                               scale_con < -1 && stride_con > 0);
1659 
1660     if (low_limit->get_int() == -max_jint) {
1661       if (!RangeLimitCheck) return;
1662       // We need this guard when scale*main_limit+offset >= limit
1663       // due to underflow. So we need execute main-loop while
1664       // scale*I+offset+1 > min_int. But (min_int-offset-1) will
1665       // underflow when (offset+1) > 0 and X will be < main_limit
1666       // when scale < 0 (and stride > 0). To avoid it we replace
1667       // positive (offset+1) with 0.
1668       //
1669       // Also (min_int+1 == -max_int) is used instead of min_int here
1670       // to avoid problem with scale == -1 (min_int/(-1) == min_int).
1671       Node* shift = _igvn.intcon(31);
1672       set_ctrl(shift, C->root());
1673       Node* sign = new (C) RShiftINode(plus_one, shift);
1674       register_new_node(sign, pre_ctrl);
1675       plus_one = new (C) AndINode(plus_one, sign);
1676       register_new_node(plus_one, pre_ctrl);
1677     } else {
1678       assert(low_limit->get_int() == 0, "wrong low limit for range check");
1679       // The only problem we have here when offset == max_int
1680       // since (max_int+1) == min_int and (0-min_int) == min_int.
1681       // But it is fine since main loop will either have
1682       // less iterations or will be skipped in such case.
1683     }
1684     // The underflow limit: low_limit <= scale*I+offset.
1685     // For main-loop compute
1686     //   scale*I+offset+1 > low_limit
1687     //   ( if (scale < 0) /* and stride > 0 */
1688     //       I < (low_limit-(offset+1))/scale
1689     //     else /* scale > 0 and stride < 0 */
1690     //       I > (low_limit-(offset+1))/scale
1691     //   )
1692 
1693     *main_limit = adjust_limit(stride_con, scale, plus_one, low_limit, *main_limit, pre_ctrl,
1694                                false);
1695   }
1696 }
1697 
1698 
1699 //------------------------------is_scaled_iv---------------------------------
1700 // Return true if exp is a constant times an induction var
1701 bool PhaseIdealLoop::is_scaled_iv(Node* exp, Node* iv, int* p_scale) {
1702   if (exp == iv) {
1703     if (p_scale != NULL) {
1704       *p_scale = 1;
1705     }
1706     return true;
1707   }
1708   int opc = exp->Opcode();
1709   if (opc == Op_MulI) {
1710     if (exp->in(1) == iv && exp->in(2)->is_Con()) {
1711       if (p_scale != NULL) {
1712         *p_scale = exp->in(2)->get_int();
1713       }
1714       return true;
1715     }
1716     if (exp->in(2) == iv && exp->in(1)->is_Con()) {
1717       if (p_scale != NULL) {
1718         *p_scale = exp->in(1)->get_int();
1719       }
1720       return true;
1721     }
1722   } else if (opc == Op_LShiftI) {
1723     if (exp->in(1) == iv && exp->in(2)->is_Con()) {
1724       if (p_scale != NULL) {
1725         *p_scale = 1 << exp->in(2)->get_int();
1726       }
1727       return true;
1728     }
1729   }
1730   return false;
1731 }
1732 
1733 //-----------------------------is_scaled_iv_plus_offset------------------------------
1734 // Return true if exp is a simple induction variable expression: k1*iv + (invar + k2)
1735 bool PhaseIdealLoop::is_scaled_iv_plus_offset(Node* exp, Node* iv, int* p_scale, Node** p_offset, int depth) {
1736   if (is_scaled_iv(exp, iv, p_scale)) {
1737     if (p_offset != NULL) {
1738       Node *zero = _igvn.intcon(0);
1739       set_ctrl(zero, C->root());
1740       *p_offset = zero;
1741     }
1742     return true;
1743   }
1744   int opc = exp->Opcode();
1745   if (opc == Op_AddI) {
1746     if (is_scaled_iv(exp->in(1), iv, p_scale)) {
1747       if (p_offset != NULL) {
1748         *p_offset = exp->in(2);
1749       }
1750       return true;
1751     }
1752     if (is_scaled_iv(exp->in(2), iv, p_scale)) {
1753       if (p_offset != NULL) {
1754         *p_offset = exp->in(1);
1755       }
1756       return true;
1757     }
1758     if (exp->in(2)->is_Con()) {
1759       Node* offset2 = NULL;
1760       if (depth < 2 &&
1761           is_scaled_iv_plus_offset(exp->in(1), iv, p_scale,
1762                                    p_offset != NULL ? &offset2 : NULL, depth+1)) {
1763         if (p_offset != NULL) {
1764           Node *ctrl_off2 = get_ctrl(offset2);
1765           Node* offset = new (C) AddINode(offset2, exp->in(2));
1766           register_new_node(offset, ctrl_off2);
1767           *p_offset = offset;
1768         }
1769         return true;
1770       }
1771     }
1772   } else if (opc == Op_SubI) {
1773     if (is_scaled_iv(exp->in(1), iv, p_scale)) {
1774       if (p_offset != NULL) {
1775         Node *zero = _igvn.intcon(0);
1776         set_ctrl(zero, C->root());
1777         Node *ctrl_off = get_ctrl(exp->in(2));
1778         Node* offset = new (C) SubINode(zero, exp->in(2));
1779         register_new_node(offset, ctrl_off);
1780         *p_offset = offset;
1781       }
1782       return true;
1783     }
1784     if (is_scaled_iv(exp->in(2), iv, p_scale)) {
1785       if (p_offset != NULL) {
1786         *p_scale *= -1;
1787         *p_offset = exp->in(1);
1788       }
1789       return true;
1790     }
1791   }
1792   return false;
1793 }
1794 
1795 //------------------------------do_range_check---------------------------------
1796 // Eliminate range-checks and other trip-counter vs loop-invariant tests.
1797 void PhaseIdealLoop::do_range_check( IdealLoopTree *loop, Node_List &old_new ) {
1798 #ifndef PRODUCT
1799   if (PrintOpto && VerifyLoopOptimizations) {
1800     tty->print("Range Check Elimination ");
1801     loop->dump_head();
1802   } else if (TraceLoopOpts) {
1803     tty->print("RangeCheck   ");
1804     loop->dump_head();
1805   }
1806 #endif
1807   assert(RangeCheckElimination, "");
1808   CountedLoopNode *cl = loop->_head->as_CountedLoop();
1809   assert(cl->is_main_loop(), "");
1810 
1811   // protect against stride not being a constant
1812   if (!cl->stride_is_con())
1813     return;
1814 
1815   // Find the trip counter; we are iteration splitting based on it
1816   Node *trip_counter = cl->phi();
1817   // Find the main loop limit; we will trim it's iterations
1818   // to not ever trip end tests
1819   Node *main_limit = cl->limit();
1820 
1821   // Need to find the main-loop zero-trip guard
1822   Node *ctrl  = cl->in(LoopNode::EntryControl);
1823   assert(ctrl->Opcode() == Op_IfTrue || ctrl->Opcode() == Op_IfFalse, "");
1824   Node *iffm = ctrl->in(0);
1825   assert(iffm->Opcode() == Op_If, "");
1826   Node *bolzm = iffm->in(1);
1827   assert(bolzm->Opcode() == Op_Bool, "");
1828   Node *cmpzm = bolzm->in(1);
1829   assert(cmpzm->is_Cmp(), "");
1830   Node *opqzm = cmpzm->in(2);
1831   // Can not optimize a loop if zero-trip Opaque1 node is optimized
1832   // away and then another round of loop opts attempted.
1833   if (opqzm->Opcode() != Op_Opaque1)
1834     return;
1835   assert(opqzm->in(1) == main_limit, "do not understand situation");
1836 
1837   // Find the pre-loop limit; we will expand it's iterations to
1838   // not ever trip low tests.
1839   Node *p_f = iffm->in(0);
1840   // pre loop may have been optimized out
1841   if (p_f->Opcode() != Op_IfFalse) {
1842     return;
1843   }
1844   CountedLoopEndNode *pre_end = p_f->in(0)->as_CountedLoopEnd();
1845   assert(pre_end->loopnode()->is_pre_loop(), "");
1846   Node *pre_opaq1 = pre_end->limit();
1847   // Occasionally it's possible for a pre-loop Opaque1 node to be
1848   // optimized away and then another round of loop opts attempted.
1849   // We can not optimize this particular loop in that case.
1850   if (pre_opaq1->Opcode() != Op_Opaque1)
1851     return;
1852   Opaque1Node *pre_opaq = (Opaque1Node*)pre_opaq1;
1853   Node *pre_limit = pre_opaq->in(1);
1854 
1855   // Where do we put new limit calculations
1856   Node *pre_ctrl = pre_end->loopnode()->in(LoopNode::EntryControl);
1857 
1858   // Ensure the original loop limit is available from the
1859   // pre-loop Opaque1 node.
1860   Node *orig_limit = pre_opaq->original_loop_limit();
1861   if (orig_limit == NULL || _igvn.type(orig_limit) == Type::TOP)
1862     return;
1863 
1864   // Must know if its a count-up or count-down loop
1865 
1866   int stride_con = cl->stride_con();
1867   Node *zero = _igvn.intcon(0);
1868   Node *one  = _igvn.intcon(1);
1869   // Use symmetrical int range [-max_jint,max_jint]
1870   Node *mini = _igvn.intcon(-max_jint);
1871   set_ctrl(zero, C->root());
1872   set_ctrl(one,  C->root());
1873   set_ctrl(mini, C->root());
1874 
1875   // Range checks that do not dominate the loop backedge (ie.
1876   // conditionally executed) can lengthen the pre loop limit beyond
1877   // the original loop limit. To prevent this, the pre limit is
1878   // (for stride > 0) MINed with the original loop limit (MAXed
1879   // stride < 0) when some range_check (rc) is conditionally
1880   // executed.
1881   bool conditional_rc = false;
1882 
1883   // Check loop body for tests of trip-counter plus loop-invariant vs
1884   // loop-invariant.
1885   for( uint i = 0; i < loop->_body.size(); i++ ) {
1886     Node *iff = loop->_body[i];
1887     if( iff->Opcode() == Op_If ) { // Test?
1888 
1889       // Test is an IfNode, has 2 projections.  If BOTH are in the loop
1890       // we need loop unswitching instead of iteration splitting.
1891       Node *exit = loop->is_loop_exit(iff);
1892       if( !exit ) continue;
1893       int flip = (exit->Opcode() == Op_IfTrue) ? 1 : 0;
1894 
1895       // Get boolean condition to test
1896       Node *i1 = iff->in(1);
1897       if( !i1->is_Bool() ) continue;
1898       BoolNode *bol = i1->as_Bool();
1899       BoolTest b_test = bol->_test;
1900       // Flip sense of test if exit condition is flipped
1901       if( flip )
1902         b_test = b_test.negate();
1903 
1904       // Get compare
1905       Node *cmp = bol->in(1);
1906 
1907       // Look for trip_counter + offset vs limit
1908       Node *rc_exp = cmp->in(1);
1909       Node *limit  = cmp->in(2);
1910       jint scale_con= 1;        // Assume trip counter not scaled
1911 
1912       Node *limit_c = get_ctrl(limit);
1913       if( loop->is_member(get_loop(limit_c) ) ) {
1914         // Compare might have operands swapped; commute them
1915         b_test = b_test.commute();
1916         rc_exp = cmp->in(2);
1917         limit  = cmp->in(1);
1918         limit_c = get_ctrl(limit);
1919         if( loop->is_member(get_loop(limit_c) ) )
1920           continue;             // Both inputs are loop varying; cannot RCE
1921       }
1922       // Here we know 'limit' is loop invariant
1923 
1924       // 'limit' maybe pinned below the zero trip test (probably from a
1925       // previous round of rce), in which case, it can't be used in the
1926       // zero trip test expression which must occur before the zero test's if.
1927       if( limit_c == ctrl ) {
1928         continue;  // Don't rce this check but continue looking for other candidates.
1929       }
1930 
1931       // Check for scaled induction variable plus an offset
1932       Node *offset = NULL;
1933 
1934       if (!is_scaled_iv_plus_offset(rc_exp, trip_counter, &scale_con, &offset)) {
1935         continue;
1936       }
1937 
1938       Node *offset_c = get_ctrl(offset);
1939       if( loop->is_member( get_loop(offset_c) ) )
1940         continue;               // Offset is not really loop invariant
1941       // Here we know 'offset' is loop invariant.
1942 
1943       // As above for the 'limit', the 'offset' maybe pinned below the
1944       // zero trip test.
1945       if( offset_c == ctrl ) {
1946         continue; // Don't rce this check but continue looking for other candidates.
1947       }
1948 #ifdef ASSERT
1949       if (TraceRangeLimitCheck) {
1950         tty->print_cr("RC bool node%s", flip ? " flipped:" : ":");
1951         bol->dump(2);
1952       }
1953 #endif
1954       // At this point we have the expression as:
1955       //   scale_con * trip_counter + offset :: limit
1956       // where scale_con, offset and limit are loop invariant.  Trip_counter
1957       // monotonically increases by stride_con, a constant.  Both (or either)
1958       // stride_con and scale_con can be negative which will flip about the
1959       // sense of the test.
1960 
1961       // Adjust pre and main loop limits to guard the correct iteration set
1962       if( cmp->Opcode() == Op_CmpU ) {// Unsigned compare is really 2 tests
1963         if( b_test._test == BoolTest::lt ) { // Range checks always use lt
1964           // The underflow and overflow limits: 0 <= scale*I+offset < limit
1965           add_constraint( stride_con, scale_con, offset, zero, limit, pre_ctrl, &pre_limit, &main_limit );
1966           if (!conditional_rc) {
1967             // (0-offset)/scale could be outside of loop iterations range.
1968             conditional_rc = !loop->dominates_backedge(iff) || RangeLimitCheck;
1969           }
1970         } else {
1971 #ifndef PRODUCT
1972           if( PrintOpto )
1973             tty->print_cr("missed RCE opportunity");
1974 #endif
1975           continue;             // In release mode, ignore it
1976         }
1977       } else {                  // Otherwise work on normal compares
1978         switch( b_test._test ) {
1979         case BoolTest::gt:
1980           // Fall into GE case
1981         case BoolTest::ge:
1982           // Convert (I*scale+offset) >= Limit to (I*(-scale)+(-offset)) <= -Limit
1983           scale_con = -scale_con;
1984           offset = new (C) SubINode( zero, offset );
1985           register_new_node( offset, pre_ctrl );
1986           limit  = new (C) SubINode( zero, limit  );
1987           register_new_node( limit, pre_ctrl );
1988           // Fall into LE case
1989         case BoolTest::le:
1990           if (b_test._test != BoolTest::gt) {
1991             // Convert X <= Y to X < Y+1
1992             limit = new (C) AddINode( limit, one );
1993             register_new_node( limit, pre_ctrl );
1994           }
1995           // Fall into LT case
1996         case BoolTest::lt:
1997           // The underflow and overflow limits: MIN_INT <= scale*I+offset < limit
1998           // Note: (MIN_INT+1 == -MAX_INT) is used instead of MIN_INT here
1999           // to avoid problem with scale == -1: MIN_INT/(-1) == MIN_INT.
2000           add_constraint( stride_con, scale_con, offset, mini, limit, pre_ctrl, &pre_limit, &main_limit );
2001           if (!conditional_rc) {
2002             // ((MIN_INT+1)-offset)/scale could be outside of loop iterations range.
2003             // Note: negative offset is replaced with 0 but (MIN_INT+1)/scale could
2004             // still be outside of loop range.
2005             conditional_rc = !loop->dominates_backedge(iff) || RangeLimitCheck;
2006           }
2007           break;
2008         default:
2009 #ifndef PRODUCT
2010           if( PrintOpto )
2011             tty->print_cr("missed RCE opportunity");
2012 #endif
2013           continue;             // Unhandled case
2014         }
2015       }
2016 
2017       // Kill the eliminated test
2018       C->set_major_progress();
2019       Node *kill_con = _igvn.intcon( 1-flip );
2020       set_ctrl(kill_con, C->root());
2021       _igvn.replace_input_of(iff, 1, kill_con);
2022       // Find surviving projection
2023       assert(iff->is_If(), "");
2024       ProjNode* dp = ((IfNode*)iff)->proj_out(1-flip);
2025       // Find loads off the surviving projection; remove their control edge
2026       for (DUIterator_Fast imax, i = dp->fast_outs(imax); i < imax; i++) {
2027         Node* cd = dp->fast_out(i); // Control-dependent node
2028         if (cd->is_Load() && cd->depends_only_on_test()) {   // Loads can now float around in the loop
2029           // Allow the load to float around in the loop, or before it
2030           // but NOT before the pre-loop.
2031           _igvn.replace_input_of(cd, 0, ctrl); // ctrl, not NULL
2032           --i;
2033           --imax;
2034         }
2035       }
2036 
2037     } // End of is IF
2038 
2039   }
2040 
2041   // Update loop limits
2042   if (conditional_rc) {
2043     pre_limit = (stride_con > 0) ? (Node*)new (C) MinINode(pre_limit, orig_limit)
2044                                  : (Node*)new (C) MaxINode(pre_limit, orig_limit);
2045     register_new_node(pre_limit, pre_ctrl);
2046   }
2047   _igvn.hash_delete(pre_opaq);
2048   pre_opaq->set_req(1, pre_limit);
2049 
2050   // Note:: we are making the main loop limit no longer precise;
2051   // need to round up based on stride.
2052   cl->set_nonexact_trip_count();
2053   if (!LoopLimitCheck && stride_con != 1 && stride_con != -1) { // Cutout for common case
2054     // "Standard" round-up logic:  ([main_limit-init+(y-1)]/y)*y+init
2055     // Hopefully, compiler will optimize for powers of 2.
2056     Node *ctrl = get_ctrl(main_limit);
2057     Node *stride = cl->stride();
2058     Node *init = cl->init_trip();
2059     Node *span = new (C) SubINode(main_limit,init);
2060     register_new_node(span,ctrl);
2061     Node *rndup = _igvn.intcon(stride_con + ((stride_con>0)?-1:1));
2062     Node *add = new (C) AddINode(span,rndup);
2063     register_new_node(add,ctrl);
2064     Node *div = new (C) DivINode(0,add,stride);
2065     register_new_node(div,ctrl);
2066     Node *mul = new (C) MulINode(div,stride);
2067     register_new_node(mul,ctrl);
2068     Node *newlim = new (C) AddINode(mul,init);
2069     register_new_node(newlim,ctrl);
2070     main_limit = newlim;
2071   }
2072 
2073   Node *main_cle = cl->loopexit();
2074   Node *main_bol = main_cle->in(1);
2075   // Hacking loop bounds; need private copies of exit test
2076   if( main_bol->outcnt() > 1 ) {// BoolNode shared?
2077     _igvn.hash_delete(main_cle);
2078     main_bol = main_bol->clone();// Clone a private BoolNode
2079     register_new_node( main_bol, main_cle->in(0) );
2080     main_cle->set_req(1,main_bol);
2081   }
2082   Node *main_cmp = main_bol->in(1);
2083   if( main_cmp->outcnt() > 1 ) { // CmpNode shared?
2084     _igvn.hash_delete(main_bol);
2085     main_cmp = main_cmp->clone();// Clone a private CmpNode
2086     register_new_node( main_cmp, main_cle->in(0) );
2087     main_bol->set_req(1,main_cmp);
2088   }
2089   // Hack the now-private loop bounds
2090   _igvn.replace_input_of(main_cmp, 2, main_limit);
2091   // The OpaqueNode is unshared by design
2092   assert( opqzm->outcnt() == 1, "cannot hack shared node" );
2093   _igvn.replace_input_of(opqzm, 1, main_limit);
2094 }
2095 
2096 //------------------------------DCE_loop_body----------------------------------
2097 // Remove simplistic dead code from loop body
2098 void IdealLoopTree::DCE_loop_body() {
2099   for( uint i = 0; i < _body.size(); i++ )
2100     if( _body.at(i)->outcnt() == 0 )
2101       _body.map( i--, _body.pop() );
2102 }
2103 
2104 
2105 //------------------------------adjust_loop_exit_prob--------------------------
2106 // Look for loop-exit tests with the 50/50 (or worse) guesses from the parsing stage.
2107 // Replace with a 1-in-10 exit guess.
2108 void IdealLoopTree::adjust_loop_exit_prob( PhaseIdealLoop *phase ) {
2109   Node *test = tail();
2110   while( test != _head ) {
2111     uint top = test->Opcode();
2112     if( top == Op_IfTrue || top == Op_IfFalse ) {
2113       int test_con = ((ProjNode*)test)->_con;
2114       assert(top == (uint)(test_con? Op_IfTrue: Op_IfFalse), "sanity");
2115       IfNode *iff = test->in(0)->as_If();
2116       if( iff->outcnt() == 2 ) {        // Ignore dead tests
2117         Node *bol = iff->in(1);
2118         if( bol && bol->req() > 1 && bol->in(1) &&
2119             ((bol->in(1)->Opcode() == Op_StorePConditional ) ||
2120              (bol->in(1)->Opcode() == Op_StoreIConditional ) ||
2121              (bol->in(1)->Opcode() == Op_StoreLConditional ) ||
2122              (bol->in(1)->Opcode() == Op_CompareAndSwapI ) ||
2123              (bol->in(1)->Opcode() == Op_CompareAndSwapL ) ||
2124              (bol->in(1)->Opcode() == Op_CompareAndSwapP ) ||
2125              (bol->in(1)->Opcode() == Op_CompareAndSwapN )))
2126           return;               // Allocation loops RARELY take backedge
2127         // Find the OTHER exit path from the IF
2128         Node* ex = iff->proj_out(1-test_con);
2129         float p = iff->_prob;
2130         if( !phase->is_member( this, ex ) && iff->_fcnt == COUNT_UNKNOWN ) {
2131           if( top == Op_IfTrue ) {
2132             if( p < (PROB_FAIR + PROB_UNLIKELY_MAG(3))) {
2133               iff->_prob = PROB_STATIC_FREQUENT;
2134             }
2135           } else {
2136             if( p > (PROB_FAIR - PROB_UNLIKELY_MAG(3))) {
2137               iff->_prob = PROB_STATIC_INFREQUENT;
2138             }
2139           }
2140         }
2141       }
2142     }
2143     test = phase->idom(test);
2144   }
2145 }
2146 
2147 
2148 //------------------------------policy_do_remove_empty_loop--------------------
2149 // Micro-benchmark spamming.  Policy is to always remove empty loops.
2150 // The 'DO' part is to replace the trip counter with the value it will
2151 // have on the last iteration.  This will break the loop.
2152 bool IdealLoopTree::policy_do_remove_empty_loop( PhaseIdealLoop *phase ) {
2153   // Minimum size must be empty loop
2154   if (_body.size() > EMPTY_LOOP_SIZE)
2155     return false;
2156 
2157   if (!_head->is_CountedLoop())
2158     return false;     // Dead loop
2159   CountedLoopNode *cl = _head->as_CountedLoop();
2160   if (!cl->is_valid_counted_loop())
2161     return false; // Malformed loop
2162   if (!phase->is_member(this, phase->get_ctrl(cl->loopexit()->in(CountedLoopEndNode::TestValue))))
2163     return false;             // Infinite loop
2164 
2165 #ifdef ASSERT
2166   // Ensure only one phi which is the iv.
2167   Node* iv = NULL;
2168   for (DUIterator_Fast imax, i = cl->fast_outs(imax); i < imax; i++) {
2169     Node* n = cl->fast_out(i);
2170     if (n->Opcode() == Op_Phi) {
2171       assert(iv == NULL, "Too many phis" );
2172       iv = n;
2173     }
2174   }
2175   assert(iv == cl->phi(), "Wrong phi" );
2176 #endif
2177 
2178   // main and post loops have explicitly created zero trip guard
2179   bool needs_guard = !cl->is_main_loop() && !cl->is_post_loop();
2180   if (needs_guard) {
2181     // Skip guard if values not overlap.
2182     const TypeInt* init_t = phase->_igvn.type(cl->init_trip())->is_int();
2183     const TypeInt* limit_t = phase->_igvn.type(cl->limit())->is_int();
2184     int  stride_con = cl->stride_con();
2185     if (stride_con > 0) {
2186       needs_guard = (init_t->_hi >= limit_t->_lo);
2187     } else {
2188       needs_guard = (init_t->_lo <= limit_t->_hi);
2189     }
2190   }
2191   if (needs_guard) {
2192     // Check for an obvious zero trip guard.
2193     Node* inctrl = PhaseIdealLoop::skip_loop_predicates(cl->in(LoopNode::EntryControl));
2194     if (inctrl->Opcode() == Op_IfTrue) {
2195       // The test should look like just the backedge of a CountedLoop
2196       Node* iff = inctrl->in(0);
2197       if (iff->is_If()) {
2198         Node* bol = iff->in(1);
2199         if (bol->is_Bool() && bol->as_Bool()->_test._test == cl->loopexit()->test_trip()) {
2200           Node* cmp = bol->in(1);
2201           if (cmp->is_Cmp() && cmp->in(1) == cl->init_trip() && cmp->in(2) == cl->limit()) {
2202             needs_guard = false;
2203           }
2204         }
2205       }
2206     }
2207   }
2208 
2209 #ifndef PRODUCT
2210   if (PrintOpto) {
2211     tty->print("Removing empty loop with%s zero trip guard", needs_guard ? "out" : "");
2212     this->dump_head();
2213   } else if (TraceLoopOpts) {
2214     tty->print("Empty with%s zero trip guard   ", needs_guard ? "out" : "");
2215     this->dump_head();
2216   }
2217 #endif
2218 
2219   if (needs_guard) {
2220     // Peel the loop to ensure there's a zero trip guard
2221     Node_List old_new;
2222     phase->do_peeling(this, old_new);
2223   }
2224 
2225   // Replace the phi at loop head with the final value of the last
2226   // iteration.  Then the CountedLoopEnd will collapse (backedge never
2227   // taken) and all loop-invariant uses of the exit values will be correct.
2228   Node *phi = cl->phi();
2229   Node *exact_limit = phase->exact_limit(this);
2230   if (exact_limit != cl->limit()) {
2231     // We also need to replace the original limit to collapse loop exit.
2232     Node* cmp = cl->loopexit()->cmp_node();
2233     assert(cl->limit() == cmp->in(2), "sanity");
2234     phase->_igvn._worklist.push(cmp->in(2)); // put limit on worklist
2235     phase->_igvn.replace_input_of(cmp, 2, exact_limit); // put cmp on worklist
2236   }
2237   // Note: the final value after increment should not overflow since
2238   // counted loop has limit check predicate.
2239   Node *final = new (phase->C) SubINode( exact_limit, cl->stride() );
2240   phase->register_new_node(final,cl->in(LoopNode::EntryControl));
2241   phase->_igvn.replace_node(phi,final);
2242   phase->C->set_major_progress();
2243   return true;
2244 }
2245 
2246 //------------------------------policy_do_one_iteration_loop-------------------
2247 // Convert one iteration loop into normal code.
2248 bool IdealLoopTree::policy_do_one_iteration_loop( PhaseIdealLoop *phase ) {
2249   if (!_head->as_Loop()->is_valid_counted_loop())
2250     return false; // Only for counted loop
2251 
2252   CountedLoopNode *cl = _head->as_CountedLoop();
2253   if (!cl->has_exact_trip_count() || cl->trip_count() != 1) {
2254     return false;
2255   }
2256 
2257 #ifndef PRODUCT
2258   if(TraceLoopOpts) {
2259     tty->print("OneIteration ");
2260     this->dump_head();
2261   }
2262 #endif
2263 
2264   Node *init_n = cl->init_trip();
2265 #ifdef ASSERT
2266   // Loop boundaries should be constant since trip count is exact.
2267   assert(init_n->get_int() + cl->stride_con() >= cl->limit()->get_int(), "should be one iteration");
2268 #endif
2269   // Replace the phi at loop head with the value of the init_trip.
2270   // Then the CountedLoopEnd will collapse (backedge will not be taken)
2271   // and all loop-invariant uses of the exit values will be correct.
2272   phase->_igvn.replace_node(cl->phi(), cl->init_trip());
2273   phase->C->set_major_progress();
2274   return true;
2275 }
2276 
2277 //=============================================================================
2278 //------------------------------iteration_split_impl---------------------------
2279 bool IdealLoopTree::iteration_split_impl( PhaseIdealLoop *phase, Node_List &old_new ) {
2280   // Compute exact loop trip count if possible.
2281   compute_exact_trip_count(phase);
2282 
2283   // Convert one iteration loop into normal code.
2284   if (policy_do_one_iteration_loop(phase))
2285     return true;
2286 
2287   // Check and remove empty loops (spam micro-benchmarks)
2288   if (policy_do_remove_empty_loop(phase))
2289     return true;  // Here we removed an empty loop
2290 
2291   bool should_peel = policy_peeling(phase); // Should we peel?
2292 
2293   bool should_unswitch = policy_unswitching(phase);
2294 
2295   // Non-counted loops may be peeled; exactly 1 iteration is peeled.
2296   // This removes loop-invariant tests (usually null checks).
2297   if (!_head->is_CountedLoop()) { // Non-counted loop
2298     if (PartialPeelLoop && phase->partial_peel(this, old_new)) {
2299       // Partial peel succeeded so terminate this round of loop opts
2300       return false;
2301     }
2302     if (should_peel) {            // Should we peel?
2303 #ifndef PRODUCT
2304       if (PrintOpto) tty->print_cr("should_peel");
2305 #endif
2306       phase->do_peeling(this,old_new);
2307     } else if (should_unswitch) {
2308       phase->do_unswitching(this, old_new);
2309     }
2310     return true;
2311   }
2312   CountedLoopNode *cl = _head->as_CountedLoop();
2313 
2314   if (!cl->is_valid_counted_loop()) return true; // Ignore various kinds of broken loops
2315 
2316   // Do nothing special to pre- and post- loops
2317   if (cl->is_pre_loop() || cl->is_post_loop()) return true;
2318 
2319   // Compute loop trip count from profile data
2320   compute_profile_trip_cnt(phase);
2321 
2322   // Before attempting fancy unrolling, RCE or alignment, see if we want
2323   // to completely unroll this loop or do loop unswitching.
2324   if (cl->is_normal_loop()) {
2325     if (should_unswitch) {
2326       phase->do_unswitching(this, old_new);
2327       return true;
2328     }
2329     bool should_maximally_unroll =  policy_maximally_unroll(phase);
2330     if (should_maximally_unroll) {
2331       // Here we did some unrolling and peeling.  Eventually we will
2332       // completely unroll this loop and it will no longer be a loop.
2333       phase->do_maximally_unroll(this,old_new);
2334       return true;
2335     }
2336   }
2337 
2338   // Skip next optimizations if running low on nodes. Note that
2339   // policy_unswitching and policy_maximally_unroll have this check.
2340   int nodes_left = phase->C->max_node_limit() - phase->C->live_nodes();
2341   if ((int)(2 * _body.size()) > nodes_left) {
2342     return true;
2343   }
2344 
2345   // Counted loops may be peeled, may need some iterations run up
2346   // front for RCE, and may want to align loop refs to a cache
2347   // line.  Thus we clone a full loop up front whose trip count is
2348   // at least 1 (if peeling), but may be several more.
2349 
2350   // The main loop will start cache-line aligned with at least 1
2351   // iteration of the unrolled body (zero-trip test required) and
2352   // will have some range checks removed.
2353 
2354   // A post-loop will finish any odd iterations (leftover after
2355   // unrolling), plus any needed for RCE purposes.
2356 
2357   bool should_unroll = policy_unroll(phase);
2358 
2359   bool should_rce = policy_range_check(phase);
2360 
2361   bool should_align = policy_align(phase);
2362 
2363   // If not RCE'ing (iteration splitting) or Aligning, then we do not
2364   // need a pre-loop.  We may still need to peel an initial iteration but
2365   // we will not be needing an unknown number of pre-iterations.
2366   //
2367   // Basically, if may_rce_align reports FALSE first time through,
2368   // we will not be able to later do RCE or Aligning on this loop.
2369   bool may_rce_align = !policy_peel_only(phase) || should_rce || should_align;
2370 
2371   // If we have any of these conditions (RCE, alignment, unrolling) met, then
2372   // we switch to the pre-/main-/post-loop model.  This model also covers
2373   // peeling.
2374   if (should_rce || should_align || should_unroll) {
2375     if (cl->is_normal_loop())  // Convert to 'pre/main/post' loops
2376       phase->insert_pre_post_loops(this,old_new, !may_rce_align);
2377 
2378     // Adjust the pre- and main-loop limits to let the pre and post loops run
2379     // with full checks, but the main-loop with no checks.  Remove said
2380     // checks from the main body.
2381     if (should_rce)
2382       phase->do_range_check(this,old_new);
2383 
2384     // Double loop body for unrolling.  Adjust the minimum-trip test (will do
2385     // twice as many iterations as before) and the main body limit (only do
2386     // an even number of trips).  If we are peeling, we might enable some RCE
2387     // and we'd rather unroll the post-RCE'd loop SO... do not unroll if
2388     // peeling.
2389     if (should_unroll && !should_peel)
2390       phase->do_unroll(this,old_new, true);
2391 
2392     // Adjust the pre-loop limits to align the main body
2393     // iterations.
2394     if (should_align)
2395       Unimplemented();
2396 
2397   } else {                      // Else we have an unchanged counted loop
2398     if (should_peel)           // Might want to peel but do nothing else
2399       phase->do_peeling(this,old_new);
2400   }
2401   return true;
2402 }
2403 
2404 
2405 //=============================================================================
2406 //------------------------------iteration_split--------------------------------
2407 bool IdealLoopTree::iteration_split( PhaseIdealLoop *phase, Node_List &old_new ) {
2408   // Recursively iteration split nested loops
2409   if (_child && !_child->iteration_split(phase, old_new))
2410     return false;
2411 
2412   // Clean out prior deadwood
2413   DCE_loop_body();
2414 
2415 
2416   // Look for loop-exit tests with my 50/50 guesses from the Parsing stage.
2417   // Replace with a 1-in-10 exit guess.
2418   if (_parent /*not the root loop*/ &&
2419       !_irreducible &&
2420       // Also ignore the occasional dead backedge
2421       !tail()->is_top()) {
2422     adjust_loop_exit_prob(phase);
2423   }
2424 
2425   // Gate unrolling, RCE and peeling efforts.
2426   if (!_child &&                // If not an inner loop, do not split
2427       !_irreducible &&
2428       _allow_optimizations &&
2429       !tail()->is_top()) {     // Also ignore the occasional dead backedge
2430     if (!_has_call) {
2431         if (!iteration_split_impl(phase, old_new)) {
2432           return false;
2433         }
2434     } else if (policy_unswitching(phase)) {
2435       phase->do_unswitching(this, old_new);
2436     }
2437   }
2438 
2439   // Minor offset re-organization to remove loop-fallout uses of
2440   // trip counter when there was no major reshaping.
2441   phase->reorg_offsets(this);
2442 
2443   if (_next && !_next->iteration_split(phase, old_new))
2444     return false;
2445   return true;
2446 }
2447 
2448 
2449 //=============================================================================
2450 // Process all the loops in the loop tree and replace any fill
2451 // patterns with an intrinsic version.
2452 bool PhaseIdealLoop::do_intrinsify_fill() {
2453   bool changed = false;
2454   for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
2455     IdealLoopTree* lpt = iter.current();
2456     changed |= intrinsify_fill(lpt);
2457   }
2458   return changed;
2459 }
2460 
2461 
2462 // Examine an inner loop looking for a a single store of an invariant
2463 // value in a unit stride loop,
2464 bool PhaseIdealLoop::match_fill_loop(IdealLoopTree* lpt, Node*& store, Node*& store_value,
2465                                      Node*& shift, Node*& con) {
2466   const char* msg = NULL;
2467   Node* msg_node = NULL;
2468 
2469   store_value = NULL;
2470   con = NULL;
2471   shift = NULL;
2472 
2473   // Process the loop looking for stores.  If there are multiple
2474   // stores or extra control flow give at this point.
2475   CountedLoopNode* head = lpt->_head->as_CountedLoop();
2476   for (uint i = 0; msg == NULL && i < lpt->_body.size(); i++) {
2477     Node* n = lpt->_body.at(i);
2478     if (n->outcnt() == 0) continue; // Ignore dead
2479     if (n->is_Store()) {
2480       if (store != NULL) {
2481         msg = "multiple stores";
2482         break;
2483       }
2484       int opc = n->Opcode();
2485       if (opc == Op_StoreP || opc == Op_StoreN || opc == Op_StoreNKlass || opc == Op_StoreCM) {
2486         msg = "oop fills not handled";
2487         break;
2488       }
2489       Node* value = n->in(MemNode::ValueIn);
2490       if (!lpt->is_invariant(value)) {
2491         msg  = "variant store value";
2492       } else if (!_igvn.type(n->in(MemNode::Address))->isa_aryptr()) {
2493         msg = "not array address";
2494       }
2495       store = n;
2496       store_value = value;
2497     } else if (n->is_If() && n != head->loopexit()) {
2498       msg = "extra control flow";
2499       msg_node = n;
2500     }
2501   }
2502 
2503   if (store == NULL) {
2504     // No store in loop
2505     return false;
2506   }
2507 
2508   if (msg == NULL && head->stride_con() != 1) {
2509     // could handle negative strides too
2510     if (head->stride_con() < 0) {
2511       msg = "negative stride";
2512     } else {
2513       msg = "non-unit stride";
2514     }
2515   }
2516 
2517   if (msg == NULL && !store->in(MemNode::Address)->is_AddP()) {
2518     msg = "can't handle store address";
2519     msg_node = store->in(MemNode::Address);
2520   }
2521 
2522   if (msg == NULL &&
2523       (!store->in(MemNode::Memory)->is_Phi() ||
2524        store->in(MemNode::Memory)->in(LoopNode::LoopBackControl) != store)) {
2525     msg = "store memory isn't proper phi";
2526     msg_node = store->in(MemNode::Memory);
2527   }
2528 
2529   // Make sure there is an appropriate fill routine
2530   BasicType t = store->as_Mem()->memory_type();
2531   const char* fill_name;
2532   if (msg == NULL &&
2533       StubRoutines::select_fill_function(t, false, fill_name) == NULL) {
2534     msg = "unsupported store";
2535     msg_node = store;
2536   }
2537 
2538   if (msg != NULL) {
2539 #ifndef PRODUCT
2540     if (TraceOptimizeFill) {
2541       tty->print_cr("not fill intrinsic candidate: %s", msg);
2542       if (msg_node != NULL) msg_node->dump();
2543     }
2544 #endif
2545     return false;
2546   }
2547 
2548   // Make sure the address expression can be handled.  It should be
2549   // head->phi * elsize + con.  head->phi might have a ConvI2L(CastII()).
2550   Node* elements[4];
2551   Node* cast = NULL;
2552   Node* conv = NULL;
2553   bool found_index = false;
2554   int count = store->in(MemNode::Address)->as_AddP()->unpack_offsets(elements, ARRAY_SIZE(elements));
2555   for (int e = 0; e < count; e++) {
2556     Node* n = elements[e];
2557     if (n->is_Con() && con == NULL) {
2558       con = n;
2559     } else if (n->Opcode() == Op_LShiftX && shift == NULL) {
2560       Node* value = n->in(1);
2561 #ifdef _LP64
2562       if (value->Opcode() == Op_ConvI2L) {
2563         conv = value;
2564         value = value->in(1);
2565       }
2566       if (value->Opcode() == Op_CastII &&
2567           value->as_CastII()->has_range_check()) {
2568         // Skip range check dependent CastII nodes
2569         cast = value;
2570         value = value->in(1);
2571       }
2572 #endif
2573       if (value != head->phi()) {
2574         msg = "unhandled shift in address";
2575       } else {
2576         if (type2aelembytes(store->as_Mem()->memory_type(), true) != (1 << n->in(2)->get_int())) {
2577           msg = "scale doesn't match";
2578         } else {
2579           found_index = true;
2580           shift = n;
2581         }
2582       }
2583     } else if (n->Opcode() == Op_ConvI2L && conv == NULL) {
2584       conv = n;
2585       n = n->in(1);
2586       if (n->Opcode() == Op_CastII &&
2587           n->as_CastII()->has_range_check()) {
2588         // Skip range check dependent CastII nodes
2589         cast = n;
2590         n = n->in(1);
2591       }
2592       if (n == head->phi()) {
2593         found_index = true;
2594       } else {
2595         msg = "unhandled input to ConvI2L";
2596       }
2597     } else if (n == head->phi()) {
2598       // no shift, check below for allowed cases
2599       found_index = true;
2600     } else {
2601       msg = "unhandled node in address";
2602       msg_node = n;
2603     }
2604   }
2605 
2606   if (count == -1) {
2607     msg = "malformed address expression";
2608     msg_node = store;
2609   }
2610 
2611   if (!found_index) {
2612     msg = "missing use of index";
2613   }
2614 
2615   // byte sized items won't have a shift
2616   if (msg == NULL && shift == NULL && t != T_BYTE && t != T_BOOLEAN) {
2617     msg = "can't find shift";
2618     msg_node = store;
2619   }
2620 
2621   if (msg != NULL) {
2622 #ifndef PRODUCT
2623     if (TraceOptimizeFill) {
2624       tty->print_cr("not fill intrinsic: %s", msg);
2625       if (msg_node != NULL) msg_node->dump();
2626     }
2627 #endif
2628     return false;
2629   }
2630 
2631   // No make sure all the other nodes in the loop can be handled
2632   VectorSet ok(Thread::current()->resource_area());
2633 
2634   // store related values are ok
2635   ok.set(store->_idx);
2636   ok.set(store->in(MemNode::Memory)->_idx);
2637 
2638   CountedLoopEndNode* loop_exit = head->loopexit();
2639   guarantee(loop_exit != NULL, "no loop exit node");
2640 
2641   // Loop structure is ok
2642   ok.set(head->_idx);
2643   ok.set(loop_exit->_idx);
2644   ok.set(head->phi()->_idx);
2645   ok.set(head->incr()->_idx);
2646   ok.set(loop_exit->cmp_node()->_idx);
2647   ok.set(loop_exit->in(1)->_idx);
2648 
2649   // Address elements are ok
2650   if (con)   ok.set(con->_idx);
2651   if (shift) ok.set(shift->_idx);
2652   if (cast)  ok.set(cast->_idx);
2653   if (conv)  ok.set(conv->_idx);
2654 
2655   for (uint i = 0; msg == NULL && i < lpt->_body.size(); i++) {
2656     Node* n = lpt->_body.at(i);
2657     if (n->outcnt() == 0) continue; // Ignore dead
2658     if (ok.test(n->_idx)) continue;
2659     // Backedge projection is ok
2660     if (n->is_IfTrue() && n->in(0) == loop_exit) continue;
2661     if (!n->is_AddP()) {
2662       msg = "unhandled node";
2663       msg_node = n;
2664       break;
2665     }
2666   }
2667 
2668   // Make sure no unexpected values are used outside the loop
2669   for (uint i = 0; msg == NULL && i < lpt->_body.size(); i++) {
2670     Node* n = lpt->_body.at(i);
2671     // These values can be replaced with other nodes if they are used
2672     // outside the loop.
2673     if (n == store || n == loop_exit || n == head->incr() || n == store->in(MemNode::Memory)) continue;
2674     for (SimpleDUIterator iter(n); iter.has_next(); iter.next()) {
2675       Node* use = iter.get();
2676       if (!lpt->_body.contains(use)) {
2677         msg = "node is used outside loop";
2678         // lpt->_body.dump();
2679         msg_node = n;
2680         break;
2681       }
2682     }
2683   }
2684 
2685 #ifdef ASSERT
2686   if (TraceOptimizeFill) {
2687     if (msg != NULL) {
2688       tty->print_cr("no fill intrinsic: %s", msg);
2689       if (msg_node != NULL) msg_node->dump();
2690     } else {
2691       tty->print_cr("fill intrinsic for:");
2692     }
2693     store->dump();
2694     if (Verbose) {
2695       lpt->_body.dump();
2696     }
2697   }
2698 #endif
2699 
2700   return msg == NULL;
2701 }
2702 
2703 
2704 
2705 bool PhaseIdealLoop::intrinsify_fill(IdealLoopTree* lpt) {
2706   // Only for counted inner loops
2707   if (!lpt->is_counted() || !lpt->is_inner()) {
2708     return false;
2709   }
2710 
2711   // Must have constant stride
2712   CountedLoopNode* head = lpt->_head->as_CountedLoop();
2713   if (!head->is_valid_counted_loop() || !head->is_normal_loop()) {
2714     return false;
2715   }
2716 
2717   // Check that the body only contains a store of a loop invariant
2718   // value that is indexed by the loop phi.
2719   Node* store = NULL;
2720   Node* store_value = NULL;
2721   Node* shift = NULL;
2722   Node* offset = NULL;
2723   if (!match_fill_loop(lpt, store, store_value, shift, offset)) {
2724     return false;
2725   }
2726 
2727   Node* exit = head->loopexit()->proj_out(0);
2728   if (exit == NULL) {
2729     return false;
2730   }
2731 
2732 #ifndef PRODUCT
2733   if (TraceLoopOpts) {
2734     tty->print("ArrayFill    ");
2735     lpt->dump_head();
2736   }
2737 #endif
2738 
2739   // Now replace the whole loop body by a call to a fill routine that
2740   // covers the same region as the loop.
2741   Node* base = store->in(MemNode::Address)->as_AddP()->in(AddPNode::Base);
2742 
2743   // Build an expression for the beginning of the copy region
2744   Node* index = head->init_trip();
2745 #ifdef _LP64
2746   index = new (C) ConvI2LNode(index);
2747   _igvn.register_new_node_with_optimizer(index);
2748 #endif
2749   if (shift != NULL) {
2750     // byte arrays don't require a shift but others do.
2751     index = new (C) LShiftXNode(index, shift->in(2));
2752     _igvn.register_new_node_with_optimizer(index);
2753   }
2754   index = new (C) AddPNode(base, base, index);
2755   _igvn.register_new_node_with_optimizer(index);
2756   Node* from = new (C) AddPNode(base, index, offset);
2757   _igvn.register_new_node_with_optimizer(from);
2758   // Compute the number of elements to copy
2759   Node* len = new (C) SubINode(head->limit(), head->init_trip());
2760   _igvn.register_new_node_with_optimizer(len);
2761 
2762   BasicType t = store->as_Mem()->memory_type();
2763   bool aligned = false;
2764   if (offset != NULL && head->init_trip()->is_Con()) {
2765     int element_size = type2aelembytes(t);
2766     aligned = (offset->find_intptr_t_type()->get_con() + head->init_trip()->get_int() * element_size) % HeapWordSize == 0;
2767   }
2768 
2769   // Build a call to the fill routine
2770   const char* fill_name;
2771   address fill = StubRoutines::select_fill_function(t, aligned, fill_name);
2772   assert(fill != NULL, "what?");
2773 
2774   // Convert float/double to int/long for fill routines
2775   if (t == T_FLOAT) {
2776     store_value = new (C) MoveF2INode(store_value);
2777     _igvn.register_new_node_with_optimizer(store_value);
2778   } else if (t == T_DOUBLE) {
2779     store_value = new (C) MoveD2LNode(store_value);
2780     _igvn.register_new_node_with_optimizer(store_value);
2781   }
2782 
2783   if (CCallingConventionRequiresIntsAsLongs &&
2784       // See StubRoutines::select_fill_function for types. FLOAT has been converted to INT.
2785       (t == T_FLOAT || t == T_INT ||  is_subword_type(t))) {
2786     store_value = new (C) ConvI2LNode(store_value);
2787     _igvn.register_new_node_with_optimizer(store_value);
2788   }
2789 
2790   Node* mem_phi = store->in(MemNode::Memory);
2791   Node* result_ctrl;
2792   Node* result_mem;
2793   const TypeFunc* call_type = OptoRuntime::array_fill_Type();
2794   CallLeafNode *call = new (C) CallLeafNoFPNode(call_type, fill,
2795                                                 fill_name, TypeAryPtr::get_array_body_type(t));
2796   uint cnt = 0;
2797   call->init_req(TypeFunc::Parms + cnt++, from);
2798   call->init_req(TypeFunc::Parms + cnt++, store_value);
2799   if (CCallingConventionRequiresIntsAsLongs) {
2800     call->init_req(TypeFunc::Parms + cnt++, C->top());
2801   }
2802 #ifdef _LP64
2803   len = new (C) ConvI2LNode(len);
2804   _igvn.register_new_node_with_optimizer(len);
2805 #endif
2806   call->init_req(TypeFunc::Parms + cnt++, len);
2807 #ifdef _LP64
2808   call->init_req(TypeFunc::Parms + cnt++, C->top());
2809 #endif
2810   call->init_req(TypeFunc::Control,   head->init_control());
2811   call->init_req(TypeFunc::I_O,       C->top());       // Does no I/O.
2812   call->init_req(TypeFunc::Memory,    mem_phi->in(LoopNode::EntryControl));
2813   call->init_req(TypeFunc::ReturnAdr, C->start()->proj_out(TypeFunc::ReturnAdr));
2814   call->init_req(TypeFunc::FramePtr,  C->start()->proj_out(TypeFunc::FramePtr));
2815   _igvn.register_new_node_with_optimizer(call);
2816   result_ctrl = new (C) ProjNode(call,TypeFunc::Control);
2817   _igvn.register_new_node_with_optimizer(result_ctrl);
2818   result_mem = new (C) ProjNode(call,TypeFunc::Memory);
2819   _igvn.register_new_node_with_optimizer(result_mem);
2820 
2821 /* Disable following optimization until proper fix (add missing checks).
2822 
2823   // If this fill is tightly coupled to an allocation and overwrites
2824   // the whole body, allow it to take over the zeroing.
2825   AllocateNode* alloc = AllocateNode::Ideal_allocation(base, this);
2826   if (alloc != NULL && alloc->is_AllocateArray()) {
2827     Node* length = alloc->as_AllocateArray()->Ideal_length();
2828     if (head->limit() == length &&
2829         head->init_trip() == _igvn.intcon(0)) {
2830       if (TraceOptimizeFill) {
2831         tty->print_cr("Eliminated zeroing in allocation");
2832       }
2833       alloc->maybe_set_complete(&_igvn);
2834     } else {
2835 #ifdef ASSERT
2836       if (TraceOptimizeFill) {
2837         tty->print_cr("filling array but bounds don't match");
2838         alloc->dump();
2839         head->init_trip()->dump();
2840         head->limit()->dump();
2841         length->dump();
2842       }
2843 #endif
2844     }
2845   }
2846 */
2847 
2848   // Redirect the old control and memory edges that are outside the loop.
2849   // Sometimes the memory phi of the head is used as the outgoing
2850   // state of the loop.  It's safe in this case to replace it with the
2851   // result_mem.
2852   _igvn.replace_node(store->in(MemNode::Memory), result_mem);
2853   lazy_replace(exit, result_ctrl);
2854   _igvn.replace_node(store, result_mem);
2855   // Any uses the increment outside of the loop become the loop limit.
2856   _igvn.replace_node(head->incr(), head->limit());
2857 
2858   // Disconnect the head from the loop.
2859   for (uint i = 0; i < lpt->_body.size(); i++) {
2860     Node* n = lpt->_body.at(i);
2861     _igvn.replace_node(n, C->top());
2862   }
2863 
2864   return true;
2865 }