1 /*
   2  * Copyright (c) 1997, 2018, 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 "gc/shared/barrierSet.hpp"
  28 #include "gc/shared/c2/barrierSetC2.hpp"
  29 #include "memory/allocation.inline.hpp"
  30 #include "opto/addnode.hpp"
  31 #include "opto/callnode.hpp"
  32 #include "opto/cfgnode.hpp"
  33 #include "opto/loopnode.hpp"
  34 #include "opto/matcher.hpp"
  35 #include "opto/movenode.hpp"
  36 #include "opto/mulnode.hpp"
  37 #include "opto/opcodes.hpp"
  38 #include "opto/phaseX.hpp"
  39 #include "opto/subnode.hpp"
  40 #include "runtime/sharedRuntime.hpp"
  41 
  42 // Portions of code courtesy of Clifford Click
  43 
  44 // Optimization - Graph Style
  45 
  46 #include "math.h"
  47 
  48 //=============================================================================
  49 //------------------------------Identity---------------------------------------
  50 // If right input is a constant 0, return the left input.
  51 Node* SubNode::Identity(PhaseGVN* phase) {
  52   assert(in(1) != this, "Must already have called Value");
  53   assert(in(2) != this, "Must already have called Value");
  54 
  55   // Remove double negation
  56   const Type *zero = add_id();
  57   if( phase->type( in(1) )->higher_equal( zero ) &&
  58       in(2)->Opcode() == Opcode() &&
  59       phase->type( in(2)->in(1) )->higher_equal( zero ) ) {
  60     return in(2)->in(2);
  61   }
  62 
  63   // Convert "(X+Y) - Y" into X and "(X+Y) - X" into Y
  64   if( in(1)->Opcode() == Op_AddI ) {
  65     if( phase->eqv(in(1)->in(2),in(2)) )
  66       return in(1)->in(1);
  67     if (phase->eqv(in(1)->in(1),in(2)))
  68       return in(1)->in(2);
  69 
  70     // Also catch: "(X + Opaque2(Y)) - Y".  In this case, 'Y' is a loop-varying
  71     // trip counter and X is likely to be loop-invariant (that's how O2 Nodes
  72     // are originally used, although the optimizer sometimes jiggers things).
  73     // This folding through an O2 removes a loop-exit use of a loop-varying
  74     // value and generally lowers register pressure in and around the loop.
  75     if( in(1)->in(2)->Opcode() == Op_Opaque2 &&
  76         phase->eqv(in(1)->in(2)->in(1),in(2)) )
  77       return in(1)->in(1);
  78   }
  79 
  80   return ( phase->type( in(2) )->higher_equal( zero ) ) ? in(1) : this;
  81 }
  82 
  83 //------------------------------Value------------------------------------------
  84 // A subtract node differences it's two inputs.
  85 const Type* SubNode::Value_common(PhaseTransform *phase) const {
  86   const Node* in1 = in(1);
  87   const Node* in2 = in(2);
  88   // Either input is TOP ==> the result is TOP
  89   const Type* t1 = (in1 == this) ? Type::TOP : phase->type(in1);
  90   if( t1 == Type::TOP ) return Type::TOP;
  91   const Type* t2 = (in2 == this) ? Type::TOP : phase->type(in2);
  92   if( t2 == Type::TOP ) return Type::TOP;
  93 
  94   // Not correct for SubFnode and AddFNode (must check for infinity)
  95   // Equal?  Subtract is zero
  96   if (in1->eqv_uncast(in2))  return add_id();
  97 
  98   // Either input is BOTTOM ==> the result is the local BOTTOM
  99   if( t1 == Type::BOTTOM || t2 == Type::BOTTOM )
 100     return bottom_type();
 101 
 102   return NULL;
 103 }
 104 
 105 const Type* SubNode::Value(PhaseGVN* phase) const {
 106   const Type* t = Value_common(phase);
 107   if (t != NULL) {
 108     return t;
 109   }
 110   const Type* t1 = phase->type(in(1));
 111   const Type* t2 = phase->type(in(2));
 112   return sub(t1,t2);            // Local flavor of type subtraction
 113 
 114 }
 115 
 116 //=============================================================================
 117 
 118 //------------------------------Helper function--------------------------------
 119 static bool ok_to_convert(Node* inc, Node* iv) {
 120     // Do not collapse (x+c0)-y if "+" is a loop increment, because the
 121     // "-" is loop invariant and collapsing extends the live-range of "x"
 122     // to overlap with the "+", forcing another register to be used in
 123     // the loop.
 124     // This test will be clearer with '&&' (apply DeMorgan's rule)
 125     // but I like the early cutouts that happen here.
 126     const PhiNode *phi;
 127     if( ( !inc->in(1)->is_Phi() ||
 128           !(phi=inc->in(1)->as_Phi()) ||
 129           phi->is_copy() ||
 130           !phi->region()->is_CountedLoop() ||
 131           inc != phi->region()->as_CountedLoop()->incr() )
 132        &&
 133         // Do not collapse (x+c0)-iv if "iv" is a loop induction variable,
 134         // because "x" maybe invariant.
 135         ( !iv->is_loop_iv() )
 136       ) {
 137       return true;
 138     } else {
 139       return false;
 140     }
 141 }
 142 //------------------------------Ideal------------------------------------------
 143 Node *SubINode::Ideal(PhaseGVN *phase, bool can_reshape){
 144   Node *in1 = in(1);
 145   Node *in2 = in(2);
 146   uint op1 = in1->Opcode();
 147   uint op2 = in2->Opcode();
 148 
 149 #ifdef ASSERT
 150   // Check for dead loop
 151   if( phase->eqv( in1, this ) || phase->eqv( in2, this ) ||
 152       ( ( op1 == Op_AddI || op1 == Op_SubI ) &&
 153         ( phase->eqv( in1->in(1), this ) || phase->eqv( in1->in(2), this ) ||
 154           phase->eqv( in1->in(1), in1  ) || phase->eqv( in1->in(2), in1 ) ) ) )
 155     assert(false, "dead loop in SubINode::Ideal");
 156 #endif
 157 
 158   const Type *t2 = phase->type( in2 );
 159   if( t2 == Type::TOP ) return NULL;
 160   // Convert "x-c0" into "x+ -c0".
 161   if( t2->base() == Type::Int ){        // Might be bottom or top...
 162     const TypeInt *i = t2->is_int();
 163     if( i->is_con() )
 164       return new AddINode(in1, phase->intcon(-i->get_con()));
 165   }
 166 
 167   // Convert "(x+c0) - y" into (x-y) + c0"
 168   // Do not collapse (x+c0)-y if "+" is a loop increment or
 169   // if "y" is a loop induction variable.
 170   if( op1 == Op_AddI && ok_to_convert(in1, in2) ) {
 171     const Type *tadd = phase->type( in1->in(2) );
 172     if( tadd->singleton() && tadd != Type::TOP ) {
 173       Node *sub2 = phase->transform( new SubINode( in1->in(1), in2 ));
 174       return new AddINode( sub2, in1->in(2) );
 175     }
 176   }
 177 
 178 
 179   // Convert "x - (y+c0)" into "(x-y) - c0"
 180   // Need the same check as in above optimization but reversed.
 181   if (op2 == Op_AddI && ok_to_convert(in2, in1)) {
 182     Node* in21 = in2->in(1);
 183     Node* in22 = in2->in(2);
 184     const TypeInt* tcon = phase->type(in22)->isa_int();
 185     if (tcon != NULL && tcon->is_con()) {
 186       Node* sub2 = phase->transform( new SubINode(in1, in21) );
 187       Node* neg_c0 = phase->intcon(- tcon->get_con());
 188       return new AddINode(sub2, neg_c0);
 189     }
 190   }
 191 
 192   const Type *t1 = phase->type( in1 );
 193   if( t1 == Type::TOP ) return NULL;
 194 
 195 #ifdef ASSERT
 196   // Check for dead loop
 197   if( ( op2 == Op_AddI || op2 == Op_SubI ) &&
 198       ( phase->eqv( in2->in(1), this ) || phase->eqv( in2->in(2), this ) ||
 199         phase->eqv( in2->in(1), in2  ) || phase->eqv( in2->in(2), in2  ) ) )
 200     assert(false, "dead loop in SubINode::Ideal");
 201 #endif
 202 
 203   // Convert "x - (x+y)" into "-y"
 204   if( op2 == Op_AddI &&
 205       phase->eqv( in1, in2->in(1) ) )
 206     return new SubINode( phase->intcon(0),in2->in(2));
 207   // Convert "(x-y) - x" into "-y"
 208   if( op1 == Op_SubI &&
 209       phase->eqv( in1->in(1), in2 ) )
 210     return new SubINode( phase->intcon(0),in1->in(2));
 211   // Convert "x - (y+x)" into "-y"
 212   if( op2 == Op_AddI &&
 213       phase->eqv( in1, in2->in(2) ) )
 214     return new SubINode( phase->intcon(0),in2->in(1));
 215 
 216   // Convert "0 - (x-y)" into "y-x"
 217   if( t1 == TypeInt::ZERO && op2 == Op_SubI )
 218     return new SubINode( in2->in(2), in2->in(1) );
 219 
 220   // Convert "0 - (x+con)" into "-con-x"
 221   jint con;
 222   if( t1 == TypeInt::ZERO && op2 == Op_AddI &&
 223       (con = in2->in(2)->find_int_con(0)) != 0 )
 224     return new SubINode( phase->intcon(-con), in2->in(1) );
 225 
 226   // Convert "(X+A) - (X+B)" into "A - B"
 227   if( op1 == Op_AddI && op2 == Op_AddI && in1->in(1) == in2->in(1) )
 228     return new SubINode( in1->in(2), in2->in(2) );
 229 
 230   // Convert "(A+X) - (B+X)" into "A - B"
 231   if( op1 == Op_AddI && op2 == Op_AddI && in1->in(2) == in2->in(2) )
 232     return new SubINode( in1->in(1), in2->in(1) );
 233 
 234   // Convert "(A+X) - (X+B)" into "A - B"
 235   if( op1 == Op_AddI && op2 == Op_AddI && in1->in(2) == in2->in(1) )
 236     return new SubINode( in1->in(1), in2->in(2) );
 237 
 238   // Convert "(X+A) - (B+X)" into "A - B"
 239   if( op1 == Op_AddI && op2 == Op_AddI && in1->in(1) == in2->in(2) )
 240     return new SubINode( in1->in(2), in2->in(1) );
 241 
 242   // Convert "A-(B-C)" into (A+C)-B", since add is commutative and generally
 243   // nicer to optimize than subtract.
 244   if( op2 == Op_SubI && in2->outcnt() == 1) {
 245     Node *add1 = phase->transform( new AddINode( in1, in2->in(2) ) );
 246     return new SubINode( add1, in2->in(1) );
 247   }
 248 
 249   return NULL;
 250 }
 251 
 252 //------------------------------sub--------------------------------------------
 253 // A subtract node differences it's two inputs.
 254 const Type *SubINode::sub( const Type *t1, const Type *t2 ) const {
 255   const TypeInt *r0 = t1->is_int(); // Handy access
 256   const TypeInt *r1 = t2->is_int();
 257   int32_t lo = java_subtract(r0->_lo, r1->_hi);
 258   int32_t hi = java_subtract(r0->_hi, r1->_lo);
 259 
 260   // We next check for 32-bit overflow.
 261   // If that happens, we just assume all integers are possible.
 262   if( (((r0->_lo ^ r1->_hi) >= 0) ||    // lo ends have same signs OR
 263        ((r0->_lo ^      lo) >= 0)) &&   // lo results have same signs AND
 264       (((r0->_hi ^ r1->_lo) >= 0) ||    // hi ends have same signs OR
 265        ((r0->_hi ^      hi) >= 0)) )    // hi results have same signs
 266     return TypeInt::make(lo,hi,MAX2(r0->_widen,r1->_widen));
 267   else                          // Overflow; assume all integers
 268     return TypeInt::INT;
 269 }
 270 
 271 //=============================================================================
 272 //------------------------------Ideal------------------------------------------
 273 Node *SubLNode::Ideal(PhaseGVN *phase, bool can_reshape) {
 274   Node *in1 = in(1);
 275   Node *in2 = in(2);
 276   uint op1 = in1->Opcode();
 277   uint op2 = in2->Opcode();
 278 
 279 #ifdef ASSERT
 280   // Check for dead loop
 281   if( phase->eqv( in1, this ) || phase->eqv( in2, this ) ||
 282       ( ( op1 == Op_AddL || op1 == Op_SubL ) &&
 283         ( phase->eqv( in1->in(1), this ) || phase->eqv( in1->in(2), this ) ||
 284           phase->eqv( in1->in(1), in1  ) || phase->eqv( in1->in(2), in1  ) ) ) )
 285     assert(false, "dead loop in SubLNode::Ideal");
 286 #endif
 287 
 288   if( phase->type( in2 ) == Type::TOP ) return NULL;
 289   const TypeLong *i = phase->type( in2 )->isa_long();
 290   // Convert "x-c0" into "x+ -c0".
 291   if( i &&                      // Might be bottom or top...
 292       i->is_con() )
 293     return new AddLNode(in1, phase->longcon(-i->get_con()));
 294 
 295   // Convert "(x+c0) - y" into (x-y) + c0"
 296   // Do not collapse (x+c0)-y if "+" is a loop increment or
 297   // if "y" is a loop induction variable.
 298   if( op1 == Op_AddL && ok_to_convert(in1, in2) ) {
 299     Node *in11 = in1->in(1);
 300     const Type *tadd = phase->type( in1->in(2) );
 301     if( tadd->singleton() && tadd != Type::TOP ) {
 302       Node *sub2 = phase->transform( new SubLNode( in11, in2 ));
 303       return new AddLNode( sub2, in1->in(2) );
 304     }
 305   }
 306 
 307   // Convert "x - (y+c0)" into "(x-y) - c0"
 308   // Need the same check as in above optimization but reversed.
 309   if (op2 == Op_AddL && ok_to_convert(in2, in1)) {
 310     Node* in21 = in2->in(1);
 311     Node* in22 = in2->in(2);
 312     const TypeLong* tcon = phase->type(in22)->isa_long();
 313     if (tcon != NULL && tcon->is_con()) {
 314       Node* sub2 = phase->transform( new SubLNode(in1, in21) );
 315       Node* neg_c0 = phase->longcon(- tcon->get_con());
 316       return new AddLNode(sub2, neg_c0);
 317     }
 318   }
 319 
 320   const Type *t1 = phase->type( in1 );
 321   if( t1 == Type::TOP ) return NULL;
 322 
 323 #ifdef ASSERT
 324   // Check for dead loop
 325   if( ( op2 == Op_AddL || op2 == Op_SubL ) &&
 326       ( phase->eqv( in2->in(1), this ) || phase->eqv( in2->in(2), this ) ||
 327         phase->eqv( in2->in(1), in2  ) || phase->eqv( in2->in(2), in2  ) ) )
 328     assert(false, "dead loop in SubLNode::Ideal");
 329 #endif
 330 
 331   // Convert "x - (x+y)" into "-y"
 332   if( op2 == Op_AddL &&
 333       phase->eqv( in1, in2->in(1) ) )
 334     return new SubLNode( phase->makecon(TypeLong::ZERO), in2->in(2));
 335   // Convert "x - (y+x)" into "-y"
 336   if( op2 == Op_AddL &&
 337       phase->eqv( in1, in2->in(2) ) )
 338     return new SubLNode( phase->makecon(TypeLong::ZERO),in2->in(1));
 339 
 340   // Convert "0 - (x-y)" into "y-x"
 341   if( phase->type( in1 ) == TypeLong::ZERO && op2 == Op_SubL )
 342     return new SubLNode( in2->in(2), in2->in(1) );
 343 
 344   // Convert "(X+A) - (X+B)" into "A - B"
 345   if( op1 == Op_AddL && op2 == Op_AddL && in1->in(1) == in2->in(1) )
 346     return new SubLNode( in1->in(2), in2->in(2) );
 347 
 348   // Convert "(A+X) - (B+X)" into "A - B"
 349   if( op1 == Op_AddL && op2 == Op_AddL && in1->in(2) == in2->in(2) )
 350     return new SubLNode( in1->in(1), in2->in(1) );
 351 
 352   // Convert "A-(B-C)" into (A+C)-B"
 353   if( op2 == Op_SubL && in2->outcnt() == 1) {
 354     Node *add1 = phase->transform( new AddLNode( in1, in2->in(2) ) );
 355     return new SubLNode( add1, in2->in(1) );
 356   }
 357 
 358   return NULL;
 359 }
 360 
 361 //------------------------------sub--------------------------------------------
 362 // A subtract node differences it's two inputs.
 363 const Type *SubLNode::sub( const Type *t1, const Type *t2 ) const {
 364   const TypeLong *r0 = t1->is_long(); // Handy access
 365   const TypeLong *r1 = t2->is_long();
 366   jlong lo = java_subtract(r0->_lo, r1->_hi);
 367   jlong hi = java_subtract(r0->_hi, r1->_lo);
 368 
 369   // We next check for 32-bit overflow.
 370   // If that happens, we just assume all integers are possible.
 371   if( (((r0->_lo ^ r1->_hi) >= 0) ||    // lo ends have same signs OR
 372        ((r0->_lo ^      lo) >= 0)) &&   // lo results have same signs AND
 373       (((r0->_hi ^ r1->_lo) >= 0) ||    // hi ends have same signs OR
 374        ((r0->_hi ^      hi) >= 0)) )    // hi results have same signs
 375     return TypeLong::make(lo,hi,MAX2(r0->_widen,r1->_widen));
 376   else                          // Overflow; assume all integers
 377     return TypeLong::LONG;
 378 }
 379 
 380 //=============================================================================
 381 //------------------------------Value------------------------------------------
 382 // A subtract node differences its two inputs.
 383 const Type* SubFPNode::Value(PhaseGVN* phase) const {
 384   const Node* in1 = in(1);
 385   const Node* in2 = in(2);
 386   // Either input is TOP ==> the result is TOP
 387   const Type* t1 = (in1 == this) ? Type::TOP : phase->type(in1);
 388   if( t1 == Type::TOP ) return Type::TOP;
 389   const Type* t2 = (in2 == this) ? Type::TOP : phase->type(in2);
 390   if( t2 == Type::TOP ) return Type::TOP;
 391 
 392   // if both operands are infinity of same sign, the result is NaN; do
 393   // not replace with zero
 394   if( (t1->is_finite() && t2->is_finite()) ) {
 395     if( phase->eqv(in1, in2) ) return add_id();
 396   }
 397 
 398   // Either input is BOTTOM ==> the result is the local BOTTOM
 399   const Type *bot = bottom_type();
 400   if( (t1 == bot) || (t2 == bot) ||
 401       (t1 == Type::BOTTOM) || (t2 == Type::BOTTOM) )
 402     return bot;
 403 
 404   return sub(t1,t2);            // Local flavor of type subtraction
 405 }
 406 
 407 
 408 //=============================================================================
 409 //------------------------------Ideal------------------------------------------
 410 Node *SubFNode::Ideal(PhaseGVN *phase, bool can_reshape) {
 411   const Type *t2 = phase->type( in(2) );
 412   // Convert "x-c0" into "x+ -c0".
 413   if( t2->base() == Type::FloatCon ) {  // Might be bottom or top...
 414     // return new (phase->C, 3) AddFNode(in(1), phase->makecon( TypeF::make(-t2->getf()) ) );
 415   }
 416 
 417   // Not associative because of boundary conditions (infinity)
 418   if( IdealizedNumerics && !phase->C->method()->is_strict() ) {
 419     // Convert "x - (x+y)" into "-y"
 420     if( in(2)->is_Add() &&
 421         phase->eqv(in(1),in(2)->in(1) ) )
 422       return new SubFNode( phase->makecon(TypeF::ZERO),in(2)->in(2));
 423   }
 424 
 425   // Cannot replace 0.0-X with -X because a 'fsub' bytecode computes
 426   // 0.0-0.0 as +0.0, while a 'fneg' bytecode computes -0.0.
 427   //if( phase->type(in(1)) == TypeF::ZERO )
 428   //return new (phase->C, 2) NegFNode(in(2));
 429 
 430   return NULL;
 431 }
 432 
 433 //------------------------------sub--------------------------------------------
 434 // A subtract node differences its two inputs.
 435 const Type *SubFNode::sub( const Type *t1, const Type *t2 ) const {
 436   // no folding if one of operands is infinity or NaN, do not do constant folding
 437   if( g_isfinite(t1->getf()) && g_isfinite(t2->getf()) ) {
 438     return TypeF::make( t1->getf() - t2->getf() );
 439   }
 440   else if( g_isnan(t1->getf()) ) {
 441     return t1;
 442   }
 443   else if( g_isnan(t2->getf()) ) {
 444     return t2;
 445   }
 446   else {
 447     return Type::FLOAT;
 448   }
 449 }
 450 
 451 //=============================================================================
 452 //------------------------------Ideal------------------------------------------
 453 Node *SubDNode::Ideal(PhaseGVN *phase, bool can_reshape){
 454   const Type *t2 = phase->type( in(2) );
 455   // Convert "x-c0" into "x+ -c0".
 456   if( t2->base() == Type::DoubleCon ) { // Might be bottom or top...
 457     // return new (phase->C, 3) AddDNode(in(1), phase->makecon( TypeD::make(-t2->getd()) ) );
 458   }
 459 
 460   // Not associative because of boundary conditions (infinity)
 461   if( IdealizedNumerics && !phase->C->method()->is_strict() ) {
 462     // Convert "x - (x+y)" into "-y"
 463     if( in(2)->is_Add() &&
 464         phase->eqv(in(1),in(2)->in(1) ) )
 465       return new SubDNode( phase->makecon(TypeD::ZERO),in(2)->in(2));
 466   }
 467 
 468   // Cannot replace 0.0-X with -X because a 'dsub' bytecode computes
 469   // 0.0-0.0 as +0.0, while a 'dneg' bytecode computes -0.0.
 470   //if( phase->type(in(1)) == TypeD::ZERO )
 471   //return new (phase->C, 2) NegDNode(in(2));
 472 
 473   return NULL;
 474 }
 475 
 476 //------------------------------sub--------------------------------------------
 477 // A subtract node differences its two inputs.
 478 const Type *SubDNode::sub( const Type *t1, const Type *t2 ) const {
 479   // no folding if one of operands is infinity or NaN, do not do constant folding
 480   if( g_isfinite(t1->getd()) && g_isfinite(t2->getd()) ) {
 481     return TypeD::make( t1->getd() - t2->getd() );
 482   }
 483   else if( g_isnan(t1->getd()) ) {
 484     return t1;
 485   }
 486   else if( g_isnan(t2->getd()) ) {
 487     return t2;
 488   }
 489   else {
 490     return Type::DOUBLE;
 491   }
 492 }
 493 
 494 //=============================================================================
 495 //------------------------------Idealize---------------------------------------
 496 // Unlike SubNodes, compare must still flatten return value to the
 497 // range -1, 0, 1.
 498 // And optimizations like those for (X + Y) - X fail if overflow happens.
 499 Node* CmpNode::Identity(PhaseGVN* phase) {
 500   return this;
 501 }
 502 
 503 #ifndef PRODUCT
 504 //----------------------------related------------------------------------------
 505 // Related nodes of comparison nodes include all data inputs (until hitting a
 506 // control boundary) as well as all outputs until and including control nodes
 507 // as well as their projections. In compact mode, data inputs till depth 1 and
 508 // all outputs till depth 1 are considered.
 509 void CmpNode::related(GrowableArray<Node*> *in_rel, GrowableArray<Node*> *out_rel, bool compact) const {
 510   if (compact) {
 511     this->collect_nodes(in_rel, 1, false, true);
 512     this->collect_nodes(out_rel, -1, false, false);
 513   } else {
 514     this->collect_nodes_in_all_data(in_rel, false);
 515     this->collect_nodes_out_all_ctrl_boundary(out_rel);
 516     // Now, find all control nodes in out_rel, and include their projections
 517     // and projection targets (if any) in the result.
 518     GrowableArray<Node*> proj(Compile::current()->unique());
 519     for (GrowableArrayIterator<Node*> it = out_rel->begin(); it != out_rel->end(); ++it) {
 520       Node* n = *it;
 521       if (n->is_CFG() && !n->is_Proj()) {
 522         // Assume projections and projection targets are found at levels 1 and 2.
 523         n->collect_nodes(&proj, -2, false, false);
 524         for (GrowableArrayIterator<Node*> p = proj.begin(); p != proj.end(); ++p) {
 525           out_rel->append_if_missing(*p);
 526         }
 527         proj.clear();
 528       }
 529     }
 530   }
 531 }
 532 #endif
 533 
 534 //=============================================================================
 535 //------------------------------cmp--------------------------------------------
 536 // Simplify a CmpI (compare 2 integers) node, based on local information.
 537 // If both inputs are constants, compare them.
 538 const Type *CmpINode::sub( const Type *t1, const Type *t2 ) const {
 539   const TypeInt *r0 = t1->is_int(); // Handy access
 540   const TypeInt *r1 = t2->is_int();
 541 
 542   if( r0->_hi < r1->_lo )       // Range is always low?
 543     return TypeInt::CC_LT;
 544   else if( r0->_lo > r1->_hi )  // Range is always high?
 545     return TypeInt::CC_GT;
 546 
 547   else if( r0->is_con() && r1->is_con() ) { // comparing constants?
 548     assert(r0->get_con() == r1->get_con(), "must be equal");
 549     return TypeInt::CC_EQ;      // Equal results.
 550   } else if( r0->_hi == r1->_lo ) // Range is never high?
 551     return TypeInt::CC_LE;
 552   else if( r0->_lo == r1->_hi ) // Range is never low?
 553     return TypeInt::CC_GE;
 554   return TypeInt::CC;           // else use worst case results
 555 }
 556 
 557 // Simplify a CmpU (compare 2 integers) node, based on local information.
 558 // If both inputs are constants, compare them.
 559 const Type *CmpUNode::sub( const Type *t1, const Type *t2 ) const {
 560   assert(!t1->isa_ptr(), "obsolete usage of CmpU");
 561 
 562   // comparing two unsigned ints
 563   const TypeInt *r0 = t1->is_int();   // Handy access
 564   const TypeInt *r1 = t2->is_int();
 565 
 566   // Current installed version
 567   // Compare ranges for non-overlap
 568   juint lo0 = r0->_lo;
 569   juint hi0 = r0->_hi;
 570   juint lo1 = r1->_lo;
 571   juint hi1 = r1->_hi;
 572 
 573   // If either one has both negative and positive values,
 574   // it therefore contains both 0 and -1, and since [0..-1] is the
 575   // full unsigned range, the type must act as an unsigned bottom.
 576   bool bot0 = ((jint)(lo0 ^ hi0) < 0);
 577   bool bot1 = ((jint)(lo1 ^ hi1) < 0);
 578 
 579   if (bot0 || bot1) {
 580     // All unsigned values are LE -1 and GE 0.
 581     if (lo0 == 0 && hi0 == 0) {
 582       return TypeInt::CC_LE;            //   0 <= bot
 583     } else if ((jint)lo0 == -1 && (jint)hi0 == -1) {
 584       return TypeInt::CC_GE;            // -1 >= bot
 585     } else if (lo1 == 0 && hi1 == 0) {
 586       return TypeInt::CC_GE;            // bot >= 0
 587     } else if ((jint)lo1 == -1 && (jint)hi1 == -1) {
 588       return TypeInt::CC_LE;            // bot <= -1
 589     }
 590   } else {
 591     // We can use ranges of the form [lo..hi] if signs are the same.
 592     assert(lo0 <= hi0 && lo1 <= hi1, "unsigned ranges are valid");
 593     // results are reversed, '-' > '+' for unsigned compare
 594     if (hi0 < lo1) {
 595       return TypeInt::CC_LT;            // smaller
 596     } else if (lo0 > hi1) {
 597       return TypeInt::CC_GT;            // greater
 598     } else if (hi0 == lo1 && lo0 == hi1) {
 599       return TypeInt::CC_EQ;            // Equal results
 600     } else if (lo0 >= hi1) {
 601       return TypeInt::CC_GE;
 602     } else if (hi0 <= lo1) {
 603       // Check for special case in Hashtable::get.  (See below.)
 604       if ((jint)lo0 >= 0 && (jint)lo1 >= 0 && is_index_range_check())
 605         return TypeInt::CC_LT;
 606       return TypeInt::CC_LE;
 607     }
 608   }
 609   // Check for special case in Hashtable::get - the hash index is
 610   // mod'ed to the table size so the following range check is useless.
 611   // Check for: (X Mod Y) CmpU Y, where the mod result and Y both have
 612   // to be positive.
 613   // (This is a gross hack, since the sub method never
 614   // looks at the structure of the node in any other case.)
 615   if ((jint)lo0 >= 0 && (jint)lo1 >= 0 && is_index_range_check())
 616     return TypeInt::CC_LT;
 617   return TypeInt::CC;                   // else use worst case results
 618 }
 619 
 620 const Type* CmpUNode::Value(PhaseGVN* phase) const {
 621   const Type* t = SubNode::Value_common(phase);
 622   if (t != NULL) {
 623     return t;
 624   }
 625   const Node* in1 = in(1);
 626   const Node* in2 = in(2);
 627   const Type* t1 = phase->type(in1);
 628   const Type* t2 = phase->type(in2);
 629   assert(t1->isa_int(), "CmpU has only Int type inputs");
 630   if (t2 == TypeInt::INT) { // Compare to bottom?
 631     return bottom_type();
 632   }
 633   uint in1_op = in1->Opcode();
 634   if (in1_op == Op_AddI || in1_op == Op_SubI) {
 635     // The problem rise when result of AddI(SubI) may overflow
 636     // signed integer value. Let say the input type is
 637     // [256, maxint] then +128 will create 2 ranges due to
 638     // overflow: [minint, minint+127] and [384, maxint].
 639     // But C2 type system keep only 1 type range and as result
 640     // it use general [minint, maxint] for this case which we
 641     // can't optimize.
 642     //
 643     // Make 2 separate type ranges based on types of AddI(SubI) inputs
 644     // and compare results of their compare. If results are the same
 645     // CmpU node can be optimized.
 646     const Node* in11 = in1->in(1);
 647     const Node* in12 = in1->in(2);
 648     const Type* t11 = (in11 == in1) ? Type::TOP : phase->type(in11);
 649     const Type* t12 = (in12 == in1) ? Type::TOP : phase->type(in12);
 650     // Skip cases when input types are top or bottom.
 651     if ((t11 != Type::TOP) && (t11 != TypeInt::INT) &&
 652         (t12 != Type::TOP) && (t12 != TypeInt::INT)) {
 653       const TypeInt *r0 = t11->is_int();
 654       const TypeInt *r1 = t12->is_int();
 655       jlong lo_r0 = r0->_lo;
 656       jlong hi_r0 = r0->_hi;
 657       jlong lo_r1 = r1->_lo;
 658       jlong hi_r1 = r1->_hi;
 659       if (in1_op == Op_SubI) {
 660         jlong tmp = hi_r1;
 661         hi_r1 = -lo_r1;
 662         lo_r1 = -tmp;
 663         // Note, for substructing [minint,x] type range
 664         // long arithmetic provides correct overflow answer.
 665         // The confusion come from the fact that in 32-bit
 666         // -minint == minint but in 64-bit -minint == maxint+1.
 667       }
 668       jlong lo_long = lo_r0 + lo_r1;
 669       jlong hi_long = hi_r0 + hi_r1;
 670       int lo_tr1 = min_jint;
 671       int hi_tr1 = (int)hi_long;
 672       int lo_tr2 = (int)lo_long;
 673       int hi_tr2 = max_jint;
 674       bool underflow = lo_long != (jlong)lo_tr2;
 675       bool overflow  = hi_long != (jlong)hi_tr1;
 676       // Use sub(t1, t2) when there is no overflow (one type range)
 677       // or when both overflow and underflow (too complex).
 678       if ((underflow != overflow) && (hi_tr1 < lo_tr2)) {
 679         // Overflow only on one boundary, compare 2 separate type ranges.
 680         int w = MAX2(r0->_widen, r1->_widen); // _widen does not matter here
 681         const TypeInt* tr1 = TypeInt::make(lo_tr1, hi_tr1, w);
 682         const TypeInt* tr2 = TypeInt::make(lo_tr2, hi_tr2, w);
 683         const Type* cmp1 = sub(tr1, t2);
 684         const Type* cmp2 = sub(tr2, t2);
 685         if (cmp1 == cmp2) {
 686           return cmp1; // Hit!
 687         }
 688       }
 689     }
 690   }
 691 
 692   return sub(t1, t2);            // Local flavor of type subtraction
 693 }
 694 
 695 bool CmpUNode::is_index_range_check() const {
 696   // Check for the "(X ModI Y) CmpU Y" shape
 697   return (in(1)->Opcode() == Op_ModI &&
 698           in(1)->in(2)->eqv_uncast(in(2)));
 699 }
 700 
 701 //------------------------------Idealize---------------------------------------
 702 Node *CmpINode::Ideal( PhaseGVN *phase, bool can_reshape ) {
 703   if (phase->type(in(2))->higher_equal(TypeInt::ZERO)) {
 704     switch (in(1)->Opcode()) {
 705     case Op_CmpL3:              // Collapse a CmpL3/CmpI into a CmpL
 706       return new CmpLNode(in(1)->in(1),in(1)->in(2));
 707     case Op_CmpF3:              // Collapse a CmpF3/CmpI into a CmpF
 708       return new CmpFNode(in(1)->in(1),in(1)->in(2));
 709     case Op_CmpD3:              // Collapse a CmpD3/CmpI into a CmpD
 710       return new CmpDNode(in(1)->in(1),in(1)->in(2));
 711     //case Op_SubI:
 712       // If (x - y) cannot overflow, then ((x - y) <?> 0)
 713       // can be turned into (x <?> y).
 714       // This is handled (with more general cases) by Ideal_sub_algebra.
 715     }
 716   }
 717   return NULL;                  // No change
 718 }
 719 
 720 
 721 //=============================================================================
 722 // Simplify a CmpL (compare 2 longs ) node, based on local information.
 723 // If both inputs are constants, compare them.
 724 const Type *CmpLNode::sub( const Type *t1, const Type *t2 ) const {
 725   const TypeLong *r0 = t1->is_long(); // Handy access
 726   const TypeLong *r1 = t2->is_long();
 727 
 728   if( r0->_hi < r1->_lo )       // Range is always low?
 729     return TypeInt::CC_LT;
 730   else if( r0->_lo > r1->_hi )  // Range is always high?
 731     return TypeInt::CC_GT;
 732 
 733   else if( r0->is_con() && r1->is_con() ) { // comparing constants?
 734     assert(r0->get_con() == r1->get_con(), "must be equal");
 735     return TypeInt::CC_EQ;      // Equal results.
 736   } else if( r0->_hi == r1->_lo ) // Range is never high?
 737     return TypeInt::CC_LE;
 738   else if( r0->_lo == r1->_hi ) // Range is never low?
 739     return TypeInt::CC_GE;
 740   return TypeInt::CC;           // else use worst case results
 741 }
 742 
 743 
 744 // Simplify a CmpUL (compare 2 unsigned longs) node, based on local information.
 745 // If both inputs are constants, compare them.
 746 const Type* CmpULNode::sub(const Type* t1, const Type* t2) const {
 747   assert(!t1->isa_ptr(), "obsolete usage of CmpUL");
 748 
 749   // comparing two unsigned longs
 750   const TypeLong* r0 = t1->is_long();   // Handy access
 751   const TypeLong* r1 = t2->is_long();
 752 
 753   // Current installed version
 754   // Compare ranges for non-overlap
 755   julong lo0 = r0->_lo;
 756   julong hi0 = r0->_hi;
 757   julong lo1 = r1->_lo;
 758   julong hi1 = r1->_hi;
 759 
 760   // If either one has both negative and positive values,
 761   // it therefore contains both 0 and -1, and since [0..-1] is the
 762   // full unsigned range, the type must act as an unsigned bottom.
 763   bool bot0 = ((jlong)(lo0 ^ hi0) < 0);
 764   bool bot1 = ((jlong)(lo1 ^ hi1) < 0);
 765 
 766   if (bot0 || bot1) {
 767     // All unsigned values are LE -1 and GE 0.
 768     if (lo0 == 0 && hi0 == 0) {
 769       return TypeInt::CC_LE;            //   0 <= bot
 770     } else if ((jlong)lo0 == -1 && (jlong)hi0 == -1) {
 771       return TypeInt::CC_GE;            // -1 >= bot
 772     } else if (lo1 == 0 && hi1 == 0) {
 773       return TypeInt::CC_GE;            // bot >= 0
 774     } else if ((jlong)lo1 == -1 && (jlong)hi1 == -1) {
 775       return TypeInt::CC_LE;            // bot <= -1
 776     }
 777   } else {
 778     // We can use ranges of the form [lo..hi] if signs are the same.
 779     assert(lo0 <= hi0 && lo1 <= hi1, "unsigned ranges are valid");
 780     // results are reversed, '-' > '+' for unsigned compare
 781     if (hi0 < lo1) {
 782       return TypeInt::CC_LT;            // smaller
 783     } else if (lo0 > hi1) {
 784       return TypeInt::CC_GT;            // greater
 785     } else if (hi0 == lo1 && lo0 == hi1) {
 786       return TypeInt::CC_EQ;            // Equal results
 787     } else if (lo0 >= hi1) {
 788       return TypeInt::CC_GE;
 789     } else if (hi0 <= lo1) {
 790       return TypeInt::CC_LE;
 791     }
 792   }
 793 
 794   return TypeInt::CC;                   // else use worst case results
 795 }
 796 
 797 //=============================================================================
 798 //------------------------------sub--------------------------------------------
 799 // Simplify an CmpP (compare 2 pointers) node, based on local information.
 800 // If both inputs are constants, compare them.
 801 const Type *CmpPNode::sub( const Type *t1, const Type *t2 ) const {
 802   const TypePtr *r0 = t1->is_ptr(); // Handy access
 803   const TypePtr *r1 = t2->is_ptr();
 804 
 805   // Undefined inputs makes for an undefined result
 806   if( TypePtr::above_centerline(r0->_ptr) ||
 807       TypePtr::above_centerline(r1->_ptr) )
 808     return Type::TOP;
 809 
 810   if (r0 == r1 && r0->singleton()) {
 811     // Equal pointer constants (klasses, nulls, etc.)
 812     return TypeInt::CC_EQ;
 813   }
 814 
 815   // See if it is 2 unrelated classes.
 816   const TypeOopPtr* p0 = r0->isa_oopptr();
 817   const TypeOopPtr* p1 = r1->isa_oopptr();
 818   if (p0 && p1) {
 819     Node* in1 = in(1)->uncast();
 820     Node* in2 = in(2)->uncast();
 821     AllocateNode* alloc1 = AllocateNode::Ideal_allocation(in1, NULL);
 822     AllocateNode* alloc2 = AllocateNode::Ideal_allocation(in2, NULL);
 823     if (MemNode::detect_ptr_independence(in1, alloc1, in2, alloc2, NULL)) {
 824       return TypeInt::CC_GT;  // different pointers
 825     }
 826     ciKlass* klass0 = p0->klass();
 827     bool    xklass0 = p0->klass_is_exact();
 828     ciKlass* klass1 = p1->klass();
 829     bool    xklass1 = p1->klass_is_exact();
 830     int kps = (p0->isa_klassptr()?1:0) + (p1->isa_klassptr()?1:0);
 831     if (klass0 && klass1 &&
 832         kps != 1 &&             // both or neither are klass pointers
 833         klass0->is_loaded() && !klass0->is_interface() && // do not trust interfaces
 834         klass1->is_loaded() && !klass1->is_interface() &&
 835         (!klass0->is_obj_array_klass() ||
 836          !klass0->as_obj_array_klass()->base_element_klass()->is_interface()) &&
 837         (!klass1->is_obj_array_klass() ||
 838          !klass1->as_obj_array_klass()->base_element_klass()->is_interface())) {
 839       bool unrelated_classes = false;
 840       // See if neither subclasses the other, or if the class on top
 841       // is precise.  In either of these cases, the compare is known
 842       // to fail if at least one of the pointers is provably not null.
 843       if (klass0->equals(klass1)) {  // if types are unequal but klasses are equal
 844         // Do nothing; we know nothing for imprecise types
 845       } else if (klass0->is_subtype_of(klass1)) {
 846         // If klass1's type is PRECISE, then classes are unrelated.
 847         unrelated_classes = xklass1;
 848       } else if (klass1->is_subtype_of(klass0)) {
 849         // If klass0's type is PRECISE, then classes are unrelated.
 850         unrelated_classes = xklass0;
 851       } else {                  // Neither subtypes the other
 852         unrelated_classes = true;
 853       }
 854       if (unrelated_classes) {
 855         // The oops classes are known to be unrelated. If the joined PTRs of
 856         // two oops is not Null and not Bottom, then we are sure that one
 857         // of the two oops is non-null, and the comparison will always fail.
 858         TypePtr::PTR jp = r0->join_ptr(r1->_ptr);
 859         if (jp != TypePtr::Null && jp != TypePtr::BotPTR) {
 860           return TypeInt::CC_GT;
 861         }
 862       }
 863     }
 864   }
 865 
 866   // Known constants can be compared exactly
 867   // Null can be distinguished from any NotNull pointers
 868   // Unknown inputs makes an unknown result
 869   if( r0->singleton() ) {
 870     intptr_t bits0 = r0->get_con();
 871     if( r1->singleton() )
 872       return bits0 == r1->get_con() ? TypeInt::CC_EQ : TypeInt::CC_GT;
 873     return ( r1->_ptr == TypePtr::NotNull && bits0==0 ) ? TypeInt::CC_GT : TypeInt::CC;
 874   } else if( r1->singleton() ) {
 875     intptr_t bits1 = r1->get_con();
 876     return ( r0->_ptr == TypePtr::NotNull && bits1==0 ) ? TypeInt::CC_GT : TypeInt::CC;
 877   } else
 878     return TypeInt::CC;
 879 }
 880 
 881 static inline Node* isa_java_mirror_load(PhaseGVN* phase, Node* n) {
 882   // Return the klass node for (indirect load from OopHandle)
 883   //   LoadP(LoadP(AddP(foo:Klass, #java_mirror)))
 884   //   or NULL if not matching.
 885 
 886 #if INCLUDE_SHENANDOAHGC
 887   if (UseShenandoahGC) {
 888     BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
 889     n = bs->step_over_gc_barrier(n);
 890   }
 891 #endif
 892 
 893   if (n->Opcode() != Op_LoadP) return NULL;
 894 
 895   const TypeInstPtr* tp = phase->type(n)->isa_instptr();
 896   if (!tp || tp->klass() != phase->C->env()->Class_klass()) return NULL;
 897 
 898   Node* adr = n->in(MemNode::Address);
 899   // First load from OopHandle: ((OopHandle)mirror)->resolve(); may need barrier.
 900   if (adr->Opcode() != Op_LoadP || !phase->type(adr)->isa_rawptr()) return NULL;
 901   adr = adr->in(MemNode::Address);
 902 
 903   intptr_t off = 0;
 904   Node* k = AddPNode::Ideal_base_and_offset(adr, phase, off);
 905   if (k == NULL)  return NULL;
 906   const TypeKlassPtr* tkp = phase->type(k)->isa_klassptr();
 907   if (!tkp || off != in_bytes(Klass::java_mirror_offset())) return NULL;
 908 
 909   // We've found the klass node of a Java mirror load.
 910   return k;
 911 }
 912 
 913 static inline Node* isa_const_java_mirror(PhaseGVN* phase, Node* n) {
 914   // for ConP(Foo.class) return ConP(Foo.klass)
 915   // otherwise return NULL
 916   if (!n->is_Con()) return NULL;
 917 
 918   const TypeInstPtr* tp = phase->type(n)->isa_instptr();
 919   if (!tp) return NULL;
 920 
 921   ciType* mirror_type = tp->java_mirror_type();
 922   // TypeInstPtr::java_mirror_type() returns non-NULL for compile-
 923   // time Class constants only.
 924   if (!mirror_type) return NULL;
 925 
 926   // x.getClass() == int.class can never be true (for all primitive types)
 927   // Return a ConP(NULL) node for this case.
 928   if (mirror_type->is_classless()) {
 929     return phase->makecon(TypePtr::NULL_PTR);
 930   }
 931 
 932   // return the ConP(Foo.klass)
 933   assert(mirror_type->is_klass(), "mirror_type should represent a Klass*");
 934   return phase->makecon(TypeKlassPtr::make(mirror_type->as_klass()));
 935 }
 936 
 937 //------------------------------Ideal------------------------------------------
 938 // Normalize comparisons between Java mirror loads to compare the klass instead.
 939 //
 940 // Also check for the case of comparing an unknown klass loaded from the primary
 941 // super-type array vs a known klass with no subtypes.  This amounts to
 942 // checking to see an unknown klass subtypes a known klass with no subtypes;
 943 // this only happens on an exact match.  We can shorten this test by 1 load.
 944 Node *CmpPNode::Ideal( PhaseGVN *phase, bool can_reshape ) {
 945   // Normalize comparisons between Java mirrors into comparisons of the low-
 946   // level klass, where a dependent load could be shortened.
 947   //
 948   // The new pattern has a nice effect of matching the same pattern used in the
 949   // fast path of instanceof/checkcast/Class.isInstance(), which allows
 950   // redundant exact type check be optimized away by GVN.
 951   // For example, in
 952   //   if (x.getClass() == Foo.class) {
 953   //     Foo foo = (Foo) x;
 954   //     // ... use a ...
 955   //   }
 956   // a CmpPNode could be shared between if_acmpne and checkcast
 957   {
 958     Node* k1 = isa_java_mirror_load(phase, in(1));
 959     Node* k2 = isa_java_mirror_load(phase, in(2));
 960     Node* conk2 = isa_const_java_mirror(phase, in(2));
 961 
 962     if (k1 && (k2 || conk2)) {
 963       Node* lhs = k1;
 964       Node* rhs = (k2 != NULL) ? k2 : conk2;
 965 #if INCLUDE_SHENANDOAHGC
 966       PhaseIterGVN* igvn = phase->is_IterGVN();
 967       if (UseShenandoahGC && igvn != NULL) {
 968         set_req_X(1, lhs, igvn);
 969         set_req_X(2, rhs, igvn);
 970       } else
 971 #endif
 972       {
 973         set_req(1, lhs);
 974         set_req(2, rhs);
 975       }
 976       return this;
 977     }
 978   }
 979 
 980   // Constant pointer on right?
 981   const TypeKlassPtr* t2 = phase->type(in(2))->isa_klassptr();
 982   if (t2 == NULL || !t2->klass_is_exact())
 983     return NULL;
 984   // Get the constant klass we are comparing to.
 985   ciKlass* superklass = t2->klass();
 986 
 987   // Now check for LoadKlass on left.
 988   Node* ldk1 = in(1);
 989   if (ldk1->is_DecodeNKlass()) {
 990     ldk1 = ldk1->in(1);
 991     if (ldk1->Opcode() != Op_LoadNKlass )
 992       return NULL;
 993   } else if (ldk1->Opcode() != Op_LoadKlass )
 994     return NULL;
 995   // Take apart the address of the LoadKlass:
 996   Node* adr1 = ldk1->in(MemNode::Address);
 997   intptr_t con2 = 0;
 998   Node* ldk2 = AddPNode::Ideal_base_and_offset(adr1, phase, con2);
 999   if (ldk2 == NULL)
1000     return NULL;
1001   if (con2 == oopDesc::klass_offset_in_bytes()) {
1002     // We are inspecting an object's concrete class.
1003     // Short-circuit the check if the query is abstract.
1004     if (superklass->is_interface() ||
1005         superklass->is_abstract()) {
1006       // Make it come out always false:
1007       this->set_req(2, phase->makecon(TypePtr::NULL_PTR));
1008       return this;
1009     }
1010   }
1011 
1012   // Check for a LoadKlass from primary supertype array.
1013   // Any nested loadklass from loadklass+con must be from the p.s. array.
1014   if (ldk2->is_DecodeNKlass()) {
1015     // Keep ldk2 as DecodeN since it could be used in CmpP below.
1016     if (ldk2->in(1)->Opcode() != Op_LoadNKlass )
1017       return NULL;
1018   } else if (ldk2->Opcode() != Op_LoadKlass)
1019     return NULL;
1020 
1021   // Verify that we understand the situation
1022   if (con2 != (intptr_t) superklass->super_check_offset())
1023     return NULL;                // Might be element-klass loading from array klass
1024 
1025   // If 'superklass' has no subklasses and is not an interface, then we are
1026   // assured that the only input which will pass the type check is
1027   // 'superklass' itself.
1028   //
1029   // We could be more liberal here, and allow the optimization on interfaces
1030   // which have a single implementor.  This would require us to increase the
1031   // expressiveness of the add_dependency() mechanism.
1032   // %%% Do this after we fix TypeOopPtr:  Deps are expressive enough now.
1033 
1034   // Object arrays must have their base element have no subtypes
1035   while (superklass->is_obj_array_klass()) {
1036     ciType* elem = superklass->as_obj_array_klass()->element_type();
1037     superklass = elem->as_klass();
1038   }
1039   if (superklass->is_instance_klass()) {
1040     ciInstanceKlass* ik = superklass->as_instance_klass();
1041     if (ik->has_subklass() || ik->is_interface())  return NULL;
1042     // Add a dependency if there is a chance that a subclass will be added later.
1043     if (!ik->is_final()) {
1044       phase->C->dependencies()->assert_leaf_type(ik);
1045     }
1046   }
1047 
1048   // Bypass the dependent load, and compare directly
1049   this->set_req(1,ldk2);
1050 
1051   return this;
1052 }
1053 
1054 //=============================================================================
1055 //------------------------------sub--------------------------------------------
1056 // Simplify an CmpN (compare 2 pointers) node, based on local information.
1057 // If both inputs are constants, compare them.
1058 const Type *CmpNNode::sub( const Type *t1, const Type *t2 ) const {
1059   const TypePtr *r0 = t1->make_ptr(); // Handy access
1060   const TypePtr *r1 = t2->make_ptr();
1061 
1062   // Undefined inputs makes for an undefined result
1063   if ((r0 == NULL) || (r1 == NULL) ||
1064       TypePtr::above_centerline(r0->_ptr) ||
1065       TypePtr::above_centerline(r1->_ptr)) {
1066     return Type::TOP;
1067   }
1068   if (r0 == r1 && r0->singleton()) {
1069     // Equal pointer constants (klasses, nulls, etc.)
1070     return TypeInt::CC_EQ;
1071   }
1072 
1073   // See if it is 2 unrelated classes.
1074   const TypeOopPtr* p0 = r0->isa_oopptr();
1075   const TypeOopPtr* p1 = r1->isa_oopptr();
1076   if (p0 && p1) {
1077     ciKlass* klass0 = p0->klass();
1078     bool    xklass0 = p0->klass_is_exact();
1079     ciKlass* klass1 = p1->klass();
1080     bool    xklass1 = p1->klass_is_exact();
1081     int kps = (p0->isa_klassptr()?1:0) + (p1->isa_klassptr()?1:0);
1082     if (klass0 && klass1 &&
1083         kps != 1 &&             // both or neither are klass pointers
1084         !klass0->is_interface() && // do not trust interfaces
1085         !klass1->is_interface()) {
1086       bool unrelated_classes = false;
1087       // See if neither subclasses the other, or if the class on top
1088       // is precise.  In either of these cases, the compare is known
1089       // to fail if at least one of the pointers is provably not null.
1090       if (klass0->equals(klass1)) { // if types are unequal but klasses are equal
1091         // Do nothing; we know nothing for imprecise types
1092       } else if (klass0->is_subtype_of(klass1)) {
1093         // If klass1's type is PRECISE, then classes are unrelated.
1094         unrelated_classes = xklass1;
1095       } else if (klass1->is_subtype_of(klass0)) {
1096         // If klass0's type is PRECISE, then classes are unrelated.
1097         unrelated_classes = xklass0;
1098       } else {                  // Neither subtypes the other
1099         unrelated_classes = true;
1100       }
1101       if (unrelated_classes) {
1102         // The oops classes are known to be unrelated. If the joined PTRs of
1103         // two oops is not Null and not Bottom, then we are sure that one
1104         // of the two oops is non-null, and the comparison will always fail.
1105         TypePtr::PTR jp = r0->join_ptr(r1->_ptr);
1106         if (jp != TypePtr::Null && jp != TypePtr::BotPTR) {
1107           return TypeInt::CC_GT;
1108         }
1109       }
1110     }
1111   }
1112 
1113   // Known constants can be compared exactly
1114   // Null can be distinguished from any NotNull pointers
1115   // Unknown inputs makes an unknown result
1116   if( r0->singleton() ) {
1117     intptr_t bits0 = r0->get_con();
1118     if( r1->singleton() )
1119       return bits0 == r1->get_con() ? TypeInt::CC_EQ : TypeInt::CC_GT;
1120     return ( r1->_ptr == TypePtr::NotNull && bits0==0 ) ? TypeInt::CC_GT : TypeInt::CC;
1121   } else if( r1->singleton() ) {
1122     intptr_t bits1 = r1->get_con();
1123     return ( r0->_ptr == TypePtr::NotNull && bits1==0 ) ? TypeInt::CC_GT : TypeInt::CC;
1124   } else
1125     return TypeInt::CC;
1126 }
1127 
1128 //------------------------------Ideal------------------------------------------
1129 Node *CmpNNode::Ideal( PhaseGVN *phase, bool can_reshape ) {
1130   return NULL;
1131 }
1132 
1133 //=============================================================================
1134 //------------------------------Value------------------------------------------
1135 // Simplify an CmpF (compare 2 floats ) node, based on local information.
1136 // If both inputs are constants, compare them.
1137 const Type* CmpFNode::Value(PhaseGVN* phase) const {
1138   const Node* in1 = in(1);
1139   const Node* in2 = in(2);
1140   // Either input is TOP ==> the result is TOP
1141   const Type* t1 = (in1 == this) ? Type::TOP : phase->type(in1);
1142   if( t1 == Type::TOP ) return Type::TOP;
1143   const Type* t2 = (in2 == this) ? Type::TOP : phase->type(in2);
1144   if( t2 == Type::TOP ) return Type::TOP;
1145 
1146   // Not constants?  Don't know squat - even if they are the same
1147   // value!  If they are NaN's they compare to LT instead of EQ.
1148   const TypeF *tf1 = t1->isa_float_constant();
1149   const TypeF *tf2 = t2->isa_float_constant();
1150   if( !tf1 || !tf2 ) return TypeInt::CC;
1151 
1152   // This implements the Java bytecode fcmpl, so unordered returns -1.
1153   if( tf1->is_nan() || tf2->is_nan() )
1154     return TypeInt::CC_LT;
1155 
1156   if( tf1->_f < tf2->_f ) return TypeInt::CC_LT;
1157   if( tf1->_f > tf2->_f ) return TypeInt::CC_GT;
1158   assert( tf1->_f == tf2->_f, "do not understand FP behavior" );
1159   return TypeInt::CC_EQ;
1160 }
1161 
1162 
1163 //=============================================================================
1164 //------------------------------Value------------------------------------------
1165 // Simplify an CmpD (compare 2 doubles ) node, based on local information.
1166 // If both inputs are constants, compare them.
1167 const Type* CmpDNode::Value(PhaseGVN* phase) const {
1168   const Node* in1 = in(1);
1169   const Node* in2 = in(2);
1170   // Either input is TOP ==> the result is TOP
1171   const Type* t1 = (in1 == this) ? Type::TOP : phase->type(in1);
1172   if( t1 == Type::TOP ) return Type::TOP;
1173   const Type* t2 = (in2 == this) ? Type::TOP : phase->type(in2);
1174   if( t2 == Type::TOP ) return Type::TOP;
1175 
1176   // Not constants?  Don't know squat - even if they are the same
1177   // value!  If they are NaN's they compare to LT instead of EQ.
1178   const TypeD *td1 = t1->isa_double_constant();
1179   const TypeD *td2 = t2->isa_double_constant();
1180   if( !td1 || !td2 ) return TypeInt::CC;
1181 
1182   // This implements the Java bytecode dcmpl, so unordered returns -1.
1183   if( td1->is_nan() || td2->is_nan() )
1184     return TypeInt::CC_LT;
1185 
1186   if( td1->_d < td2->_d ) return TypeInt::CC_LT;
1187   if( td1->_d > td2->_d ) return TypeInt::CC_GT;
1188   assert( td1->_d == td2->_d, "do not understand FP behavior" );
1189   return TypeInt::CC_EQ;
1190 }
1191 
1192 //------------------------------Ideal------------------------------------------
1193 Node *CmpDNode::Ideal(PhaseGVN *phase, bool can_reshape){
1194   // Check if we can change this to a CmpF and remove a ConvD2F operation.
1195   // Change  (CMPD (F2D (float)) (ConD value))
1196   // To      (CMPF      (float)  (ConF value))
1197   // Valid when 'value' does not lose precision as a float.
1198   // Benefits: eliminates conversion, does not require 24-bit mode
1199 
1200   // NaNs prevent commuting operands.  This transform works regardless of the
1201   // order of ConD and ConvF2D inputs by preserving the original order.
1202   int idx_f2d = 1;              // ConvF2D on left side?
1203   if( in(idx_f2d)->Opcode() != Op_ConvF2D )
1204     idx_f2d = 2;                // No, swap to check for reversed args
1205   int idx_con = 3-idx_f2d;      // Check for the constant on other input
1206 
1207   if( ConvertCmpD2CmpF &&
1208       in(idx_f2d)->Opcode() == Op_ConvF2D &&
1209       in(idx_con)->Opcode() == Op_ConD ) {
1210     const TypeD *t2 = in(idx_con)->bottom_type()->is_double_constant();
1211     double t2_value_as_double = t2->_d;
1212     float  t2_value_as_float  = (float)t2_value_as_double;
1213     if( t2_value_as_double == (double)t2_value_as_float ) {
1214       // Test value can be represented as a float
1215       // Eliminate the conversion to double and create new comparison
1216       Node *new_in1 = in(idx_f2d)->in(1);
1217       Node *new_in2 = phase->makecon( TypeF::make(t2_value_as_float) );
1218       if( idx_f2d != 1 ) {      // Must flip args to match original order
1219         Node *tmp = new_in1;
1220         new_in1 = new_in2;
1221         new_in2 = tmp;
1222       }
1223       CmpFNode *new_cmp = (Opcode() == Op_CmpD3)
1224         ? new CmpF3Node( new_in1, new_in2 )
1225         : new CmpFNode ( new_in1, new_in2 ) ;
1226       return new_cmp;           // Changed to CmpFNode
1227     }
1228     // Testing value required the precision of a double
1229   }
1230   return NULL;                  // No change
1231 }
1232 
1233 
1234 //=============================================================================
1235 //------------------------------cc2logical-------------------------------------
1236 // Convert a condition code type to a logical type
1237 const Type *BoolTest::cc2logical( const Type *CC ) const {
1238   if( CC == Type::TOP ) return Type::TOP;
1239   if( CC->base() != Type::Int ) return TypeInt::BOOL; // Bottom or worse
1240   const TypeInt *ti = CC->is_int();
1241   if( ti->is_con() ) {          // Only 1 kind of condition codes set?
1242     // Match low order 2 bits
1243     int tmp = ((ti->get_con()&3) == (_test&3)) ? 1 : 0;
1244     if( _test & 4 ) tmp = 1-tmp;     // Optionally complement result
1245     return TypeInt::make(tmp);       // Boolean result
1246   }
1247 
1248   if( CC == TypeInt::CC_GE ) {
1249     if( _test == ge ) return TypeInt::ONE;
1250     if( _test == lt ) return TypeInt::ZERO;
1251   }
1252   if( CC == TypeInt::CC_LE ) {
1253     if( _test == le ) return TypeInt::ONE;
1254     if( _test == gt ) return TypeInt::ZERO;
1255   }
1256 
1257   return TypeInt::BOOL;
1258 }
1259 
1260 //------------------------------dump_spec-------------------------------------
1261 // Print special per-node info
1262 void BoolTest::dump_on(outputStream *st) const {
1263   const char *msg[] = {"eq","gt","of","lt","ne","le","nof","ge"};
1264   st->print("%s", msg[_test]);
1265 }
1266 
1267 //=============================================================================
1268 uint BoolNode::hash() const { return (Node::hash() << 3)|(_test._test+1); }
1269 uint BoolNode::size_of() const { return sizeof(BoolNode); }
1270 
1271 //------------------------------operator==-------------------------------------
1272 uint BoolNode::cmp( const Node &n ) const {
1273   const BoolNode *b = (const BoolNode *)&n; // Cast up
1274   return (_test._test == b->_test._test);
1275 }
1276 
1277 //-------------------------------make_predicate--------------------------------
1278 Node* BoolNode::make_predicate(Node* test_value, PhaseGVN* phase) {
1279   if (test_value->is_Con())   return test_value;
1280   if (test_value->is_Bool())  return test_value;
1281   if (test_value->is_CMove() &&
1282       test_value->in(CMoveNode::Condition)->is_Bool()) {
1283     BoolNode*   bol   = test_value->in(CMoveNode::Condition)->as_Bool();
1284     const Type* ftype = phase->type(test_value->in(CMoveNode::IfFalse));
1285     const Type* ttype = phase->type(test_value->in(CMoveNode::IfTrue));
1286     if (ftype == TypeInt::ZERO && !TypeInt::ZERO->higher_equal(ttype)) {
1287       return bol;
1288     } else if (ttype == TypeInt::ZERO && !TypeInt::ZERO->higher_equal(ftype)) {
1289       return phase->transform( bol->negate(phase) );
1290     }
1291     // Else fall through.  The CMove gets in the way of the test.
1292     // It should be the case that make_predicate(bol->as_int_value()) == bol.
1293   }
1294   Node* cmp = new CmpINode(test_value, phase->intcon(0));
1295   cmp = phase->transform(cmp);
1296   Node* bol = new BoolNode(cmp, BoolTest::ne);
1297   return phase->transform(bol);
1298 }
1299 
1300 //--------------------------------as_int_value---------------------------------
1301 Node* BoolNode::as_int_value(PhaseGVN* phase) {
1302   // Inverse to make_predicate.  The CMove probably boils down to a Conv2B.
1303   Node* cmov = CMoveNode::make(NULL, this,
1304                                phase->intcon(0), phase->intcon(1),
1305                                TypeInt::BOOL);
1306   return phase->transform(cmov);
1307 }
1308 
1309 //----------------------------------negate-------------------------------------
1310 BoolNode* BoolNode::negate(PhaseGVN* phase) {
1311   return new BoolNode(in(1), _test.negate());
1312 }
1313 
1314 // Change "bool eq/ne (cmp (add/sub A B) C)" into false/true if add/sub
1315 // overflows and we can prove that C is not in the two resulting ranges.
1316 // This optimization is similar to the one performed by CmpUNode::Value().
1317 Node* BoolNode::fold_cmpI(PhaseGVN* phase, SubNode* cmp, Node* cmp1, int cmp_op,
1318                           int cmp1_op, const TypeInt* cmp2_type) {
1319   // Only optimize eq/ne integer comparison of add/sub
1320   if((_test._test == BoolTest::eq || _test._test == BoolTest::ne) &&
1321      (cmp_op == Op_CmpI) && (cmp1_op == Op_AddI || cmp1_op == Op_SubI)) {
1322     // Skip cases were inputs of add/sub are not integers or of bottom type
1323     const TypeInt* r0 = phase->type(cmp1->in(1))->isa_int();
1324     const TypeInt* r1 = phase->type(cmp1->in(2))->isa_int();
1325     if ((r0 != NULL) && (r0 != TypeInt::INT) &&
1326         (r1 != NULL) && (r1 != TypeInt::INT) &&
1327         (cmp2_type != TypeInt::INT)) {
1328       // Compute exact (long) type range of add/sub result
1329       jlong lo_long = r0->_lo;
1330       jlong hi_long = r0->_hi;
1331       if (cmp1_op == Op_AddI) {
1332         lo_long += r1->_lo;
1333         hi_long += r1->_hi;
1334       } else {
1335         lo_long -= r1->_hi;
1336         hi_long -= r1->_lo;
1337       }
1338       // Check for over-/underflow by casting to integer
1339       int lo_int = (int)lo_long;
1340       int hi_int = (int)hi_long;
1341       bool underflow = lo_long != (jlong)lo_int;
1342       bool overflow  = hi_long != (jlong)hi_int;
1343       if ((underflow != overflow) && (hi_int < lo_int)) {
1344         // Overflow on one boundary, compute resulting type ranges:
1345         // tr1 [MIN_INT, hi_int] and tr2 [lo_int, MAX_INT]
1346         int w = MAX2(r0->_widen, r1->_widen); // _widen does not matter here
1347         const TypeInt* tr1 = TypeInt::make(min_jint, hi_int, w);
1348         const TypeInt* tr2 = TypeInt::make(lo_int, max_jint, w);
1349         // Compare second input of cmp to both type ranges
1350         const Type* sub_tr1 = cmp->sub(tr1, cmp2_type);
1351         const Type* sub_tr2 = cmp->sub(tr2, cmp2_type);
1352         if (sub_tr1 == TypeInt::CC_LT && sub_tr2 == TypeInt::CC_GT) {
1353           // The result of the add/sub will never equal cmp2. Replace BoolNode
1354           // by false (0) if it tests for equality and by true (1) otherwise.
1355           return ConINode::make((_test._test == BoolTest::eq) ? 0 : 1);
1356         }
1357       }
1358     }
1359   }
1360   return NULL;
1361 }
1362 
1363 static bool is_counted_loop_cmp(Node *cmp) {
1364   Node *n = cmp->in(1)->in(1);
1365   return n != NULL &&
1366          n->is_Phi() &&
1367          n->in(0) != NULL &&
1368          n->in(0)->is_CountedLoop() &&
1369          n->in(0)->as_CountedLoop()->phi() == n;
1370 }
1371 
1372 //------------------------------Ideal------------------------------------------
1373 Node *BoolNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1374   // Change "bool tst (cmp con x)" into "bool ~tst (cmp x con)".
1375   // This moves the constant to the right.  Helps value-numbering.
1376   Node *cmp = in(1);
1377   if( !cmp->is_Sub() ) return NULL;
1378   int cop = cmp->Opcode();
1379   if( cop == Op_FastLock || cop == Op_FastUnlock) return NULL;
1380   Node *cmp1 = cmp->in(1);
1381   Node *cmp2 = cmp->in(2);
1382   if( !cmp1 ) return NULL;
1383 
1384   if (_test._test == BoolTest::overflow || _test._test == BoolTest::no_overflow) {
1385     return NULL;
1386   }
1387 
1388   // Constant on left?
1389   Node *con = cmp1;
1390   uint op2 = cmp2->Opcode();
1391   // Move constants to the right of compare's to canonicalize.
1392   // Do not muck with Opaque1 nodes, as this indicates a loop
1393   // guard that cannot change shape.
1394   if( con->is_Con() && !cmp2->is_Con() && op2 != Op_Opaque1 &&
1395       // Because of NaN's, CmpD and CmpF are not commutative
1396       cop != Op_CmpD && cop != Op_CmpF &&
1397       // Protect against swapping inputs to a compare when it is used by a
1398       // counted loop exit, which requires maintaining the loop-limit as in(2)
1399       !is_counted_loop_exit_test() ) {
1400     // Ok, commute the constant to the right of the cmp node.
1401     // Clone the Node, getting a new Node of the same class
1402     cmp = cmp->clone();
1403     // Swap inputs to the clone
1404     cmp->swap_edges(1, 2);
1405     cmp = phase->transform( cmp );
1406     return new BoolNode( cmp, _test.commute() );
1407   }
1408 
1409   // Change "bool eq/ne (cmp (xor X 1) 0)" into "bool ne/eq (cmp X 0)".
1410   // The XOR-1 is an idiom used to flip the sense of a bool.  We flip the
1411   // test instead.
1412   int cmp1_op = cmp1->Opcode();
1413   const TypeInt* cmp2_type = phase->type(cmp2)->isa_int();
1414   if (cmp2_type == NULL)  return NULL;
1415   Node* j_xor = cmp1;
1416   if( cmp2_type == TypeInt::ZERO &&
1417       cmp1_op == Op_XorI &&
1418       j_xor->in(1) != j_xor &&          // An xor of itself is dead
1419       phase->type( j_xor->in(1) ) == TypeInt::BOOL &&
1420       phase->type( j_xor->in(2) ) == TypeInt::ONE &&
1421       (_test._test == BoolTest::eq ||
1422        _test._test == BoolTest::ne) ) {
1423     Node *ncmp = phase->transform(new CmpINode(j_xor->in(1),cmp2));
1424     return new BoolNode( ncmp, _test.negate() );
1425   }
1426 
1427   // Change ((x & m) u<= m) or ((m & x) u<= m) to always true
1428   // Same with ((x & m) u< m+1) and ((m & x) u< m+1)
1429   if (cop == Op_CmpU &&
1430       cmp1_op == Op_AndI) {
1431     Node* bound = NULL;
1432     if (_test._test == BoolTest::le) {
1433       bound = cmp2;
1434     } else if (_test._test == BoolTest::lt &&
1435                cmp2->Opcode() == Op_AddI &&
1436                cmp2->in(2)->find_int_con(0) == 1) {
1437       bound = cmp2->in(1);
1438     }
1439     if (cmp1->in(2) == bound || cmp1->in(1) == bound) {
1440       return ConINode::make(1);
1441     }
1442   }
1443 
1444   // Change ((x & (m - 1)) u< m) into (m > 0)
1445   // This is the off-by-one variant of the above
1446   if (cop == Op_CmpU &&
1447       _test._test == BoolTest::lt &&
1448       cmp1_op == Op_AndI) {
1449     Node* l = cmp1->in(1);
1450     Node* r = cmp1->in(2);
1451     for (int repeat = 0; repeat < 2; repeat++) {
1452       bool match = r->Opcode() == Op_AddI && r->in(2)->find_int_con(0) == -1 &&
1453                    r->in(1) == cmp2;
1454       if (match) {
1455         // arraylength known to be non-negative, so a (arraylength != 0) is sufficient,
1456         // but to be compatible with the array range check pattern, use (arraylength u> 0)
1457         Node* ncmp = cmp2->Opcode() == Op_LoadRange
1458                      ? phase->transform(new CmpUNode(cmp2, phase->intcon(0)))
1459                      : phase->transform(new CmpINode(cmp2, phase->intcon(0)));
1460         return new BoolNode(ncmp, BoolTest::gt);
1461       } else {
1462         // commute and try again
1463         l = cmp1->in(2);
1464         r = cmp1->in(1);
1465       }
1466     }
1467   }
1468 
1469   // Change x u< 1 or x u<= 0 to x == 0
1470   if (cop == Op_CmpU &&
1471       cmp1_op != Op_LoadRange &&
1472       ((_test._test == BoolTest::lt &&
1473         cmp2->find_int_con(-1) == 1) ||
1474        (_test._test == BoolTest::le &&
1475         cmp2->find_int_con(-1) == 0))) {
1476     Node* ncmp = phase->transform(new CmpINode(cmp1, phase->intcon(0)));
1477     return new BoolNode(ncmp, BoolTest::eq);
1478   }
1479 
1480   // Change (arraylength <= 0) or (arraylength == 0)
1481   //   into (arraylength u<= 0)
1482   // Also change (arraylength != 0) into (arraylength u> 0)
1483   // The latter version matches the code pattern generated for
1484   // array range checks, which will more likely be optimized later.
1485   if (cop == Op_CmpI &&
1486       cmp1_op == Op_LoadRange &&
1487       cmp2->find_int_con(-1) == 0) {
1488     if (_test._test == BoolTest::le || _test._test == BoolTest::eq) {
1489       Node* ncmp = phase->transform(new CmpUNode(cmp1, cmp2));
1490       return new BoolNode(ncmp, BoolTest::le);
1491     } else if (_test._test == BoolTest::ne) {
1492       Node* ncmp = phase->transform(new CmpUNode(cmp1, cmp2));
1493       return new BoolNode(ncmp, BoolTest::gt);
1494     }
1495   }
1496 
1497   // Change "bool eq/ne (cmp (Conv2B X) 0)" into "bool eq/ne (cmp X 0)".
1498   // This is a standard idiom for branching on a boolean value.
1499   Node *c2b = cmp1;
1500   if( cmp2_type == TypeInt::ZERO &&
1501       cmp1_op == Op_Conv2B &&
1502       (_test._test == BoolTest::eq ||
1503        _test._test == BoolTest::ne) ) {
1504     Node *ncmp = phase->transform(phase->type(c2b->in(1))->isa_int()
1505        ? (Node*)new CmpINode(c2b->in(1),cmp2)
1506        : (Node*)new CmpPNode(c2b->in(1),phase->makecon(TypePtr::NULL_PTR))
1507     );
1508     return new BoolNode( ncmp, _test._test );
1509   }
1510 
1511   // Comparing a SubI against a zero is equal to comparing the SubI
1512   // arguments directly.  This only works for eq and ne comparisons
1513   // due to possible integer overflow.
1514   if ((_test._test == BoolTest::eq || _test._test == BoolTest::ne) &&
1515         (cop == Op_CmpI) &&
1516         (cmp1_op == Op_SubI) &&
1517         ( cmp2_type == TypeInt::ZERO ) ) {
1518     Node *ncmp = phase->transform( new CmpINode(cmp1->in(1),cmp1->in(2)));
1519     return new BoolNode( ncmp, _test._test );
1520   }
1521 
1522   // Same as above but with and AddI of a constant
1523   if ((_test._test == BoolTest::eq || _test._test == BoolTest::ne) &&
1524       cop == Op_CmpI &&
1525       cmp1_op == Op_AddI &&
1526       cmp1->in(2) != NULL &&
1527       phase->type(cmp1->in(2))->isa_int() &&
1528       phase->type(cmp1->in(2))->is_int()->is_con() &&
1529       cmp2_type == TypeInt::ZERO &&
1530       !is_counted_loop_cmp(cmp) // modifying the exit test of a counted loop messes the counted loop shape
1531       ) {
1532     const TypeInt* cmp1_in2 = phase->type(cmp1->in(2))->is_int();
1533     Node *ncmp = phase->transform( new CmpINode(cmp1->in(1),phase->intcon(-cmp1_in2->_hi)));
1534     return new BoolNode( ncmp, _test._test );
1535   }
1536 
1537   // Change (-A vs 0) into (A vs 0) by commuting the test.  Disallow in the
1538   // most general case because negating 0x80000000 does nothing.  Needed for
1539   // the CmpF3/SubI/CmpI idiom.
1540   if( cop == Op_CmpI &&
1541       cmp1_op == Op_SubI &&
1542       cmp2_type == TypeInt::ZERO &&
1543       phase->type( cmp1->in(1) ) == TypeInt::ZERO &&
1544       phase->type( cmp1->in(2) )->higher_equal(TypeInt::SYMINT) ) {
1545     Node *ncmp = phase->transform( new CmpINode(cmp1->in(2),cmp2));
1546     return new BoolNode( ncmp, _test.commute() );
1547   }
1548 
1549   // Try to optimize signed integer comparison
1550   return fold_cmpI(phase, cmp->as_Sub(), cmp1, cop, cmp1_op, cmp2_type);
1551 
1552   //  The transformation below is not valid for either signed or unsigned
1553   //  comparisons due to wraparound concerns at MAX_VALUE and MIN_VALUE.
1554   //  This transformation can be resurrected when we are able to
1555   //  make inferences about the range of values being subtracted from
1556   //  (or added to) relative to the wraparound point.
1557   //
1558   //    // Remove +/-1's if possible.
1559   //    // "X <= Y-1" becomes "X <  Y"
1560   //    // "X+1 <= Y" becomes "X <  Y"
1561   //    // "X <  Y+1" becomes "X <= Y"
1562   //    // "X-1 <  Y" becomes "X <= Y"
1563   //    // Do not this to compares off of the counted-loop-end.  These guys are
1564   //    // checking the trip counter and they want to use the post-incremented
1565   //    // counter.  If they use the PRE-incremented counter, then the counter has
1566   //    // to be incremented in a private block on a loop backedge.
1567   //    if( du && du->cnt(this) && du->out(this)[0]->Opcode() == Op_CountedLoopEnd )
1568   //      return NULL;
1569   //  #ifndef PRODUCT
1570   //    // Do not do this in a wash GVN pass during verification.
1571   //    // Gets triggered by too many simple optimizations to be bothered with
1572   //    // re-trying it again and again.
1573   //    if( !phase->allow_progress() ) return NULL;
1574   //  #endif
1575   //    // Not valid for unsigned compare because of corner cases in involving zero.
1576   //    // For example, replacing "X-1 <u Y" with "X <=u Y" fails to throw an
1577   //    // exception in case X is 0 (because 0-1 turns into 4billion unsigned but
1578   //    // "0 <=u Y" is always true).
1579   //    if( cmp->Opcode() == Op_CmpU ) return NULL;
1580   //    int cmp2_op = cmp2->Opcode();
1581   //    if( _test._test == BoolTest::le ) {
1582   //      if( cmp1_op == Op_AddI &&
1583   //          phase->type( cmp1->in(2) ) == TypeInt::ONE )
1584   //        return clone_cmp( cmp, cmp1->in(1), cmp2, phase, BoolTest::lt );
1585   //      else if( cmp2_op == Op_AddI &&
1586   //         phase->type( cmp2->in(2) ) == TypeInt::MINUS_1 )
1587   //        return clone_cmp( cmp, cmp1, cmp2->in(1), phase, BoolTest::lt );
1588   //    } else if( _test._test == BoolTest::lt ) {
1589   //      if( cmp1_op == Op_AddI &&
1590   //          phase->type( cmp1->in(2) ) == TypeInt::MINUS_1 )
1591   //        return clone_cmp( cmp, cmp1->in(1), cmp2, phase, BoolTest::le );
1592   //      else if( cmp2_op == Op_AddI &&
1593   //         phase->type( cmp2->in(2) ) == TypeInt::ONE )
1594   //        return clone_cmp( cmp, cmp1, cmp2->in(1), phase, BoolTest::le );
1595   //    }
1596 }
1597 
1598 //------------------------------Value------------------------------------------
1599 // Simplify a Bool (convert condition codes to boolean (1 or 0)) node,
1600 // based on local information.   If the input is constant, do it.
1601 const Type* BoolNode::Value(PhaseGVN* phase) const {
1602   return _test.cc2logical( phase->type( in(1) ) );
1603 }
1604 
1605 #ifndef PRODUCT
1606 //------------------------------dump_spec--------------------------------------
1607 // Dump special per-node info
1608 void BoolNode::dump_spec(outputStream *st) const {
1609   st->print("[");
1610   _test.dump_on(st);
1611   st->print("]");
1612 }
1613 
1614 //-------------------------------related---------------------------------------
1615 // A BoolNode's related nodes are all of its data inputs, and all of its
1616 // outputs until control nodes are hit, which are included. In compact
1617 // representation, inputs till level 3 and immediate outputs are included.
1618 void BoolNode::related(GrowableArray<Node*> *in_rel, GrowableArray<Node*> *out_rel, bool compact) const {
1619   if (compact) {
1620     this->collect_nodes(in_rel, 3, false, true);
1621     this->collect_nodes(out_rel, -1, false, false);
1622   } else {
1623     this->collect_nodes_in_all_data(in_rel, false);
1624     this->collect_nodes_out_all_ctrl_boundary(out_rel);
1625   }
1626 }
1627 #endif
1628 
1629 //----------------------is_counted_loop_exit_test------------------------------
1630 // Returns true if node is used by a counted loop node.
1631 bool BoolNode::is_counted_loop_exit_test() {
1632   for( DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++ ) {
1633     Node* use = fast_out(i);
1634     if (use->is_CountedLoopEnd()) {
1635       return true;
1636     }
1637   }
1638   return false;
1639 }
1640 
1641 //=============================================================================
1642 //------------------------------Value------------------------------------------
1643 // Compute sqrt
1644 const Type* SqrtDNode::Value(PhaseGVN* phase) const {
1645   const Type *t1 = phase->type( in(1) );
1646   if( t1 == Type::TOP ) return Type::TOP;
1647   if( t1->base() != Type::DoubleCon ) return Type::DOUBLE;
1648   double d = t1->getd();
1649   if( d < 0.0 ) return Type::DOUBLE;
1650   return TypeD::make( sqrt( d ) );
1651 }
1652 
1653 const Type* SqrtFNode::Value(PhaseGVN* phase) const {
1654   const Type *t1 = phase->type( in(1) );
1655   if( t1 == Type::TOP ) return Type::TOP;
1656   if( t1->base() != Type::FloatCon ) return Type::FLOAT;
1657   float f = t1->getf();
1658   if( f < 0.0f ) return Type::FLOAT;
1659   return TypeF::make( (float)sqrt( (double)f ) );
1660 }
--- EOF ---