1 /* 2 * Copyright (c) 1997, 2015, 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/cfgnode.hpp" 31 #include "opto/loopnode.hpp" 32 #include "opto/matcher.hpp" 33 #include "opto/movenode.hpp" 34 #include "opto/mulnode.hpp" 35 #include "opto/opcodes.hpp" 36 #include "opto/phaseX.hpp" 37 #include "opto/subnode.hpp" 38 #include "runtime/sharedRuntime.hpp" 39 40 // Portions of code courtesy of Clifford Click 41 42 // Optimization - Graph Style 43 44 #include "math.h" 45 46 //============================================================================= 47 //------------------------------Identity--------------------------------------- 48 // If right input is a constant 0, return the left input. 49 Node *SubNode::Identity( PhaseTransform *phase ) { 50 assert(in(1) != this, "Must already have called Value"); 51 assert(in(2) != this, "Must already have called Value"); 52 53 // Remove double negation 54 const Type *zero = add_id(); 55 if( phase->type( in(1) )->higher_equal( zero ) && 56 in(2)->Opcode() == Opcode() && 57 phase->type( in(2)->in(1) )->higher_equal( zero ) ) { 58 return in(2)->in(2); 59 } 60 61 // Convert "(X+Y) - Y" into X and "(X+Y) - X" into Y 62 if( in(1)->Opcode() == Op_AddI ) { 63 if( phase->eqv(in(1)->in(2),in(2)) ) 64 return in(1)->in(1); 65 if (phase->eqv(in(1)->in(1),in(2))) 66 return in(1)->in(2); 67 68 // Also catch: "(X + Opaque2(Y)) - Y". In this case, 'Y' is a loop-varying 69 // trip counter and X is likely to be loop-invariant (that's how O2 Nodes 70 // are originally used, although the optimizer sometimes jiggers things). 71 // This folding through an O2 removes a loop-exit use of a loop-varying 72 // value and generally lowers register pressure in and around the loop. 73 if( in(1)->in(2)->Opcode() == Op_Opaque2 && 74 phase->eqv(in(1)->in(2)->in(1),in(2)) ) 75 return in(1)->in(1); 76 } 77 78 return ( phase->type( in(2) )->higher_equal( zero ) ) ? in(1) : this; 79 } 80 81 //------------------------------Value------------------------------------------ 82 // A subtract node differences it's two inputs. 83 const Type* SubNode::Value_common(PhaseTransform *phase) const { 84 const Node* in1 = in(1); 85 const Node* in2 = in(2); 86 // Either input is TOP ==> the result is TOP 87 const Type* t1 = (in1 == this) ? Type::TOP : phase->type(in1); 88 if( t1 == Type::TOP ) return Type::TOP; 89 const Type* t2 = (in2 == this) ? Type::TOP : phase->type(in2); 90 if( t2 == Type::TOP ) return Type::TOP; 91 92 // Not correct for SubFnode and AddFNode (must check for infinity) 93 // Equal? Subtract is zero 94 if (in1->eqv_uncast(in2)) return add_id(); 95 96 // Either input is BOTTOM ==> the result is the local BOTTOM 97 if( t1 == Type::BOTTOM || t2 == Type::BOTTOM ) 98 return bottom_type(); 99 100 return NULL; 101 } 102 103 const Type* SubNode::Value(PhaseTransform *phase) const { 104 const Type* t = Value_common(phase); 105 if (t != NULL) { 106 return t; 107 } 108 const Type* t1 = phase->type(in(1)); 109 const Type* t2 = phase->type(in(2)); 110 return sub(t1,t2); // Local flavor of type subtraction 111 112 } 113 114 //============================================================================= 115 116 //------------------------------Helper function-------------------------------- 117 static bool ok_to_convert(Node* inc, Node* iv) { 118 // Do not collapse (x+c0)-y if "+" is a loop increment, because the 119 // "-" is loop invariant and collapsing extends the live-range of "x" 120 // to overlap with the "+", forcing another register to be used in 121 // the loop. 122 // This test will be clearer with '&&' (apply DeMorgan's rule) 123 // but I like the early cutouts that happen here. 124 const PhiNode *phi; 125 if( ( !inc->in(1)->is_Phi() || 126 !(phi=inc->in(1)->as_Phi()) || 127 phi->is_copy() || 128 !phi->region()->is_CountedLoop() || 129 inc != phi->region()->as_CountedLoop()->incr() ) 130 && 131 // Do not collapse (x+c0)-iv if "iv" is a loop induction variable, 132 // because "x" maybe invariant. 133 ( !iv->is_loop_iv() ) 134 ) { 135 return true; 136 } else { 137 return false; 138 } 139 } 140 //------------------------------Ideal------------------------------------------ 141 Node *SubINode::Ideal(PhaseGVN *phase, bool can_reshape){ 142 Node *in1 = in(1); 143 Node *in2 = in(2); 144 uint op1 = in1->Opcode(); 145 uint op2 = in2->Opcode(); 146 147 #ifdef ASSERT 148 // Check for dead loop 149 if( phase->eqv( in1, this ) || phase->eqv( in2, this ) || 150 ( op1 == Op_AddI || op1 == Op_SubI ) && 151 ( phase->eqv( in1->in(1), this ) || phase->eqv( in1->in(2), this ) || 152 phase->eqv( in1->in(1), in1 ) || phase->eqv( in1->in(2), in1 ) ) ) 153 assert(false, "dead loop in SubINode::Ideal"); 154 #endif 155 156 const Type *t2 = phase->type( in2 ); 157 if( t2 == Type::TOP ) return NULL; 158 // Convert "x-c0" into "x+ -c0". 159 if( t2->base() == Type::Int ){ // Might be bottom or top... 160 const TypeInt *i = t2->is_int(); 161 if( i->is_con() ) 162 return new AddINode(in1, phase->intcon(-i->get_con())); 163 } 164 165 // Convert "(x+c0) - y" into (x-y) + c0" 166 // Do not collapse (x+c0)-y if "+" is a loop increment or 167 // if "y" is a loop induction variable. 168 if( op1 == Op_AddI && ok_to_convert(in1, in2) ) { 169 const Type *tadd = phase->type( in1->in(2) ); 170 if( tadd->singleton() && tadd != Type::TOP ) { 171 Node *sub2 = phase->transform( new SubINode( in1->in(1), in2 )); 172 return new AddINode( sub2, in1->in(2) ); 173 } 174 } 175 176 177 // Convert "x - (y+c0)" into "(x-y) - c0" 178 // Need the same check as in above optimization but reversed. 179 if (op2 == Op_AddI && ok_to_convert(in2, in1)) { 180 Node* in21 = in2->in(1); 181 Node* in22 = in2->in(2); 182 const TypeInt* tcon = phase->type(in22)->isa_int(); 183 if (tcon != NULL && tcon->is_con()) { 184 Node* sub2 = phase->transform( new SubINode(in1, in21) ); 185 Node* neg_c0 = phase->intcon(- tcon->get_con()); 186 return new AddINode(sub2, neg_c0); 187 } 188 } 189 190 const Type *t1 = phase->type( in1 ); 191 if( t1 == Type::TOP ) return NULL; 192 193 #ifdef ASSERT 194 // Check for dead loop 195 if( ( op2 == Op_AddI || op2 == Op_SubI ) && 196 ( phase->eqv( in2->in(1), this ) || phase->eqv( in2->in(2), this ) || 197 phase->eqv( in2->in(1), in2 ) || phase->eqv( in2->in(2), in2 ) ) ) 198 assert(false, "dead loop in SubINode::Ideal"); 199 #endif 200 201 // Convert "x - (x+y)" into "-y" 202 if( op2 == Op_AddI && 203 phase->eqv( in1, in2->in(1) ) ) 204 return new SubINode( phase->intcon(0),in2->in(2)); 205 // Convert "(x-y) - x" into "-y" 206 if( op1 == Op_SubI && 207 phase->eqv( in1->in(1), in2 ) ) 208 return new SubINode( phase->intcon(0),in1->in(2)); 209 // Convert "x - (y+x)" into "-y" 210 if( op2 == Op_AddI && 211 phase->eqv( in1, in2->in(2) ) ) 212 return new SubINode( phase->intcon(0),in2->in(1)); 213 214 // Convert "0 - (x-y)" into "y-x" 215 if( t1 == TypeInt::ZERO && op2 == Op_SubI ) 216 return new SubINode( in2->in(2), in2->in(1) ); 217 218 // Convert "0 - (x+con)" into "-con-x" 219 jint con; 220 if( t1 == TypeInt::ZERO && op2 == Op_AddI && 221 (con = in2->in(2)->find_int_con(0)) != 0 ) 222 return new SubINode( phase->intcon(-con), in2->in(1) ); 223 224 // Convert "(X+A) - (X+B)" into "A - B" 225 if( op1 == Op_AddI && op2 == Op_AddI && in1->in(1) == in2->in(1) ) 226 return new SubINode( in1->in(2), in2->in(2) ); 227 228 // Convert "(A+X) - (B+X)" into "A - B" 229 if( op1 == Op_AddI && op2 == Op_AddI && in1->in(2) == in2->in(2) ) 230 return new SubINode( in1->in(1), in2->in(1) ); 231 232 // Convert "(A+X) - (X+B)" into "A - B" 233 if( op1 == Op_AddI && op2 == Op_AddI && in1->in(2) == in2->in(1) ) 234 return new SubINode( in1->in(1), in2->in(2) ); 235 236 // Convert "(X+A) - (B+X)" into "A - B" 237 if( op1 == Op_AddI && op2 == Op_AddI && in1->in(1) == in2->in(2) ) 238 return new SubINode( in1->in(2), in2->in(1) ); 239 240 // Convert "A-(B-C)" into (A+C)-B", since add is commutative and generally 241 // nicer to optimize than subtract. 242 if( op2 == Op_SubI && in2->outcnt() == 1) { 243 Node *add1 = phase->transform( new AddINode( in1, in2->in(2) ) ); 244 return new SubINode( add1, in2->in(1) ); 245 } 246 247 return NULL; 248 } 249 250 //------------------------------sub-------------------------------------------- 251 // A subtract node differences it's two inputs. 252 const Type *SubINode::sub( const Type *t1, const Type *t2 ) const { 253 const TypeInt *r0 = t1->is_int(); // Handy access 254 const TypeInt *r1 = t2->is_int(); 255 int32_t lo = r0->_lo - r1->_hi; 256 int32_t hi = r0->_hi - r1->_lo; 257 258 // We next check for 32-bit overflow. 259 // If that happens, we just assume all integers are possible. 260 if( (((r0->_lo ^ r1->_hi) >= 0) || // lo ends have same signs OR 261 ((r0->_lo ^ lo) >= 0)) && // lo results have same signs AND 262 (((r0->_hi ^ r1->_lo) >= 0) || // hi ends have same signs OR 263 ((r0->_hi ^ hi) >= 0)) ) // hi results have same signs 264 return TypeInt::make(lo,hi,MAX2(r0->_widen,r1->_widen)); 265 else // Overflow; assume all integers 266 return TypeInt::INT; 267 } 268 269 //============================================================================= 270 //------------------------------Ideal------------------------------------------ 271 Node *SubLNode::Ideal(PhaseGVN *phase, bool can_reshape) { 272 Node *in1 = in(1); 273 Node *in2 = in(2); 274 uint op1 = in1->Opcode(); 275 uint op2 = in2->Opcode(); 276 277 #ifdef ASSERT 278 // Check for dead loop 279 if( phase->eqv( in1, this ) || phase->eqv( in2, this ) || 280 ( op1 == Op_AddL || op1 == Op_SubL ) && 281 ( phase->eqv( in1->in(1), this ) || phase->eqv( in1->in(2), this ) || 282 phase->eqv( in1->in(1), in1 ) || phase->eqv( in1->in(2), in1 ) ) ) 283 assert(false, "dead loop in SubLNode::Ideal"); 284 #endif 285 286 if( phase->type( in2 ) == Type::TOP ) return NULL; 287 const TypeLong *i = phase->type( in2 )->isa_long(); 288 // Convert "x-c0" into "x+ -c0". 289 if( i && // Might be bottom or top... 290 i->is_con() ) 291 return new AddLNode(in1, phase->longcon(-i->get_con())); 292 293 // Convert "(x+c0) - y" into (x-y) + c0" 294 // Do not collapse (x+c0)-y if "+" is a loop increment or 295 // if "y" is a loop induction variable. 296 if( op1 == Op_AddL && ok_to_convert(in1, in2) ) { 297 Node *in11 = in1->in(1); 298 const Type *tadd = phase->type( in1->in(2) ); 299 if( tadd->singleton() && tadd != Type::TOP ) { 300 Node *sub2 = phase->transform( new SubLNode( in11, in2 )); 301 return new AddLNode( sub2, in1->in(2) ); 302 } 303 } 304 305 // Convert "x - (y+c0)" into "(x-y) - c0" 306 // Need the same check as in above optimization but reversed. 307 if (op2 == Op_AddL && ok_to_convert(in2, in1)) { 308 Node* in21 = in2->in(1); 309 Node* in22 = in2->in(2); 310 const TypeLong* tcon = phase->type(in22)->isa_long(); 311 if (tcon != NULL && tcon->is_con()) { 312 Node* sub2 = phase->transform( new SubLNode(in1, in21) ); 313 Node* neg_c0 = phase->longcon(- tcon->get_con()); 314 return new AddLNode(sub2, neg_c0); 315 } 316 } 317 318 const Type *t1 = phase->type( in1 ); 319 if( t1 == Type::TOP ) return NULL; 320 321 #ifdef ASSERT 322 // Check for dead loop 323 if( ( op2 == Op_AddL || op2 == Op_SubL ) && 324 ( phase->eqv( in2->in(1), this ) || phase->eqv( in2->in(2), this ) || 325 phase->eqv( in2->in(1), in2 ) || phase->eqv( in2->in(2), in2 ) ) ) 326 assert(false, "dead loop in SubLNode::Ideal"); 327 #endif 328 329 // Convert "x - (x+y)" into "-y" 330 if( op2 == Op_AddL && 331 phase->eqv( in1, in2->in(1) ) ) 332 return new SubLNode( phase->makecon(TypeLong::ZERO), in2->in(2)); 333 // Convert "x - (y+x)" into "-y" 334 if( op2 == Op_AddL && 335 phase->eqv( in1, in2->in(2) ) ) 336 return new SubLNode( phase->makecon(TypeLong::ZERO),in2->in(1)); 337 338 // Convert "0 - (x-y)" into "y-x" 339 if( phase->type( in1 ) == TypeLong::ZERO && op2 == Op_SubL ) 340 return new SubLNode( in2->in(2), in2->in(1) ); 341 342 // Convert "(X+A) - (X+B)" into "A - B" 343 if( op1 == Op_AddL && op2 == Op_AddL && in1->in(1) == in2->in(1) ) 344 return new SubLNode( in1->in(2), in2->in(2) ); 345 346 // Convert "(A+X) - (B+X)" into "A - B" 347 if( op1 == Op_AddL && op2 == Op_AddL && in1->in(2) == in2->in(2) ) 348 return new SubLNode( in1->in(1), in2->in(1) ); 349 350 // Convert "A-(B-C)" into (A+C)-B" 351 if( op2 == Op_SubL && in2->outcnt() == 1) { 352 Node *add1 = phase->transform( new AddLNode( in1, in2->in(2) ) ); 353 return new SubLNode( add1, in2->in(1) ); 354 } 355 356 return NULL; 357 } 358 359 //------------------------------sub-------------------------------------------- 360 // A subtract node differences it's two inputs. 361 const Type *SubLNode::sub( const Type *t1, const Type *t2 ) const { 362 const TypeLong *r0 = t1->is_long(); // Handy access 363 const TypeLong *r1 = t2->is_long(); 364 jlong lo = r0->_lo - r1->_hi; 365 jlong hi = r0->_hi - r1->_lo; 366 367 // We next check for 32-bit overflow. 368 // If that happens, we just assume all integers are possible. 369 if( (((r0->_lo ^ r1->_hi) >= 0) || // lo ends have same signs OR 370 ((r0->_lo ^ lo) >= 0)) && // lo results have same signs AND 371 (((r0->_hi ^ r1->_lo) >= 0) || // hi ends have same signs OR 372 ((r0->_hi ^ hi) >= 0)) ) // hi results have same signs 373 return TypeLong::make(lo,hi,MAX2(r0->_widen,r1->_widen)); 374 else // Overflow; assume all integers 375 return TypeLong::LONG; 376 } 377 378 //============================================================================= 379 //------------------------------Value------------------------------------------ 380 // A subtract node differences its two inputs. 381 const Type *SubFPNode::Value( PhaseTransform *phase ) const { 382 const Node* in1 = in(1); 383 const Node* in2 = in(2); 384 // Either input is TOP ==> the result is TOP 385 const Type* t1 = (in1 == this) ? Type::TOP : phase->type(in1); 386 if( t1 == Type::TOP ) return Type::TOP; 387 const Type* t2 = (in2 == this) ? Type::TOP : phase->type(in2); 388 if( t2 == Type::TOP ) return Type::TOP; 389 390 // if both operands are infinity of same sign, the result is NaN; do 391 // not replace with zero 392 if( (t1->is_finite() && t2->is_finite()) ) { 393 if( phase->eqv(in1, in2) ) return add_id(); 394 } 395 396 // Either input is BOTTOM ==> the result is the local BOTTOM 397 const Type *bot = bottom_type(); 398 if( (t1 == bot) || (t2 == bot) || 399 (t1 == Type::BOTTOM) || (t2 == Type::BOTTOM) ) 400 return bot; 401 402 return sub(t1,t2); // Local flavor of type subtraction 403 } 404 405 406 //============================================================================= 407 //------------------------------Ideal------------------------------------------ 408 Node *SubFNode::Ideal(PhaseGVN *phase, bool can_reshape) { 409 const Type *t2 = phase->type( in(2) ); 410 // Convert "x-c0" into "x+ -c0". 411 if( t2->base() == Type::FloatCon ) { // Might be bottom or top... 412 // return new (phase->C, 3) AddFNode(in(1), phase->makecon( TypeF::make(-t2->getf()) ) ); 413 } 414 415 // Not associative because of boundary conditions (infinity) 416 if( IdealizedNumerics && !phase->C->method()->is_strict() ) { 417 // Convert "x - (x+y)" into "-y" 418 if( in(2)->is_Add() && 419 phase->eqv(in(1),in(2)->in(1) ) ) 420 return new SubFNode( phase->makecon(TypeF::ZERO),in(2)->in(2)); 421 } 422 423 // Cannot replace 0.0-X with -X because a 'fsub' bytecode computes 424 // 0.0-0.0 as +0.0, while a 'fneg' bytecode computes -0.0. 425 //if( phase->type(in(1)) == TypeF::ZERO ) 426 //return new (phase->C, 2) NegFNode(in(2)); 427 428 return NULL; 429 } 430 431 //------------------------------sub-------------------------------------------- 432 // A subtract node differences its two inputs. 433 const Type *SubFNode::sub( const Type *t1, const Type *t2 ) const { 434 // no folding if one of operands is infinity or NaN, do not do constant folding 435 if( g_isfinite(t1->getf()) && g_isfinite(t2->getf()) ) { 436 return TypeF::make( t1->getf() - t2->getf() ); 437 } 438 else if( g_isnan(t1->getf()) ) { 439 return t1; 440 } 441 else if( g_isnan(t2->getf()) ) { 442 return t2; 443 } 444 else { 445 return Type::FLOAT; 446 } 447 } 448 449 //============================================================================= 450 //------------------------------Ideal------------------------------------------ 451 Node *SubDNode::Ideal(PhaseGVN *phase, bool can_reshape){ 452 const Type *t2 = phase->type( in(2) ); 453 // Convert "x-c0" into "x+ -c0". 454 if( t2->base() == Type::DoubleCon ) { // Might be bottom or top... 455 // return new (phase->C, 3) AddDNode(in(1), phase->makecon( TypeD::make(-t2->getd()) ) ); 456 } 457 458 // Not associative because of boundary conditions (infinity) 459 if( IdealizedNumerics && !phase->C->method()->is_strict() ) { 460 // Convert "x - (x+y)" into "-y" 461 if( in(2)->is_Add() && 462 phase->eqv(in(1),in(2)->in(1) ) ) 463 return new SubDNode( phase->makecon(TypeD::ZERO),in(2)->in(2)); 464 } 465 466 // Cannot replace 0.0-X with -X because a 'dsub' bytecode computes 467 // 0.0-0.0 as +0.0, while a 'dneg' bytecode computes -0.0. 468 //if( phase->type(in(1)) == TypeD::ZERO ) 469 //return new (phase->C, 2) NegDNode(in(2)); 470 471 return NULL; 472 } 473 474 //------------------------------sub-------------------------------------------- 475 // A subtract node differences its two inputs. 476 const Type *SubDNode::sub( const Type *t1, const Type *t2 ) const { 477 // no folding if one of operands is infinity or NaN, do not do constant folding 478 if( g_isfinite(t1->getd()) && g_isfinite(t2->getd()) ) { 479 return TypeD::make( t1->getd() - t2->getd() ); 480 } 481 else if( g_isnan(t1->getd()) ) { 482 return t1; 483 } 484 else if( g_isnan(t2->getd()) ) { 485 return t2; 486 } 487 else { 488 return Type::DOUBLE; 489 } 490 } 491 492 //============================================================================= 493 //------------------------------Idealize--------------------------------------- 494 // Unlike SubNodes, compare must still flatten return value to the 495 // range -1, 0, 1. 496 // And optimizations like those for (X + Y) - X fail if overflow happens. 497 Node *CmpNode::Identity( PhaseTransform *phase ) { 498 return this; 499 } 500 501 #ifndef PRODUCT 502 //----------------------------related------------------------------------------ 503 // Related nodes of comparison nodes include all data inputs (until hitting a 504 // control boundary) as well as all outputs until and including control nodes 505 // as well as their projections. In compact mode, data inputs till depth 1 and 506 // all outputs till depth 1 are considered. 507 void CmpNode::related(GrowableArray<Node*> *in_rel, GrowableArray<Node*> *out_rel, bool compact) const { 508 if (compact) { 509 this->collect_nodes(in_rel, 1, false, true); 510 this->collect_nodes(out_rel, -1, false, false); 511 } else { 512 this->collect_nodes_in_all_data(in_rel, false); 513 this->collect_nodes_out_all_ctrl_boundary(out_rel); 514 // Now, find all control nodes in out_rel, and include their projections 515 // and projection targets (if any) in the result. 516 GrowableArray<Node*> proj(Compile::current()->unique()); 517 for (GrowableArrayIterator<Node*> it = out_rel->begin(); it != out_rel->end(); ++it) { 518 Node* n = *it; 519 if (n->is_CFG() && !n->is_Proj()) { 520 // Assume projections and projection targets are found at levels 1 and 2. 521 n->collect_nodes(&proj, -2, false, false); 522 for (GrowableArrayIterator<Node*> p = proj.begin(); p != proj.end(); ++p) { 523 out_rel->append_if_missing(*p); 524 } 525 proj.clear(); 526 } 527 } 528 } 529 } 530 #endif 531 532 //============================================================================= 533 //------------------------------cmp-------------------------------------------- 534 // Simplify a CmpI (compare 2 integers) node, based on local information. 535 // If both inputs are constants, compare them. 536 const Type *CmpINode::sub( const Type *t1, const Type *t2 ) const { 537 const TypeInt *r0 = t1->is_int(); // Handy access 538 const TypeInt *r1 = t2->is_int(); 539 540 if( r0->_hi < r1->_lo ) // Range is always low? 541 return TypeInt::CC_LT; 542 else if( r0->_lo > r1->_hi ) // Range is always high? 543 return TypeInt::CC_GT; 544 545 else if( r0->is_con() && r1->is_con() ) { // comparing constants? 546 assert(r0->get_con() == r1->get_con(), "must be equal"); 547 return TypeInt::CC_EQ; // Equal results. 548 } else if( r0->_hi == r1->_lo ) // Range is never high? 549 return TypeInt::CC_LE; 550 else if( r0->_lo == r1->_hi ) // Range is never low? 551 return TypeInt::CC_GE; 552 return TypeInt::CC; // else use worst case results 553 } 554 555 // Simplify a CmpU (compare 2 integers) node, based on local information. 556 // If both inputs are constants, compare them. 557 const Type *CmpUNode::sub( const Type *t1, const Type *t2 ) const { 558 assert(!t1->isa_ptr(), "obsolete usage of CmpU"); 559 560 // comparing two unsigned ints 561 const TypeInt *r0 = t1->is_int(); // Handy access 562 const TypeInt *r1 = t2->is_int(); 563 564 // Current installed version 565 // Compare ranges for non-overlap 566 juint lo0 = r0->_lo; 567 juint hi0 = r0->_hi; 568 juint lo1 = r1->_lo; 569 juint hi1 = r1->_hi; 570 571 // If either one has both negative and positive values, 572 // it therefore contains both 0 and -1, and since [0..-1] is the 573 // full unsigned range, the type must act as an unsigned bottom. 574 bool bot0 = ((jint)(lo0 ^ hi0) < 0); 575 bool bot1 = ((jint)(lo1 ^ hi1) < 0); 576 577 if (bot0 || bot1) { 578 // All unsigned values are LE -1 and GE 0. 579 if (lo0 == 0 && hi0 == 0) { 580 return TypeInt::CC_LE; // 0 <= bot 581 } else if (lo1 == 0 && hi1 == 0) { 582 return TypeInt::CC_GE; // bot >= 0 583 } 584 } else { 585 // We can use ranges of the form [lo..hi] if signs are the same. 586 assert(lo0 <= hi0 && lo1 <= hi1, "unsigned ranges are valid"); 587 // results are reversed, '-' > '+' for unsigned compare 588 if (hi0 < lo1) { 589 return TypeInt::CC_LT; // smaller 590 } else if (lo0 > hi1) { 591 return TypeInt::CC_GT; // greater 592 } else if (hi0 == lo1 && lo0 == hi1) { 593 return TypeInt::CC_EQ; // Equal results 594 } else if (lo0 >= hi1) { 595 return TypeInt::CC_GE; 596 } else if (hi0 <= lo1) { 597 // Check for special case in Hashtable::get. (See below.) 598 if ((jint)lo0 >= 0 && (jint)lo1 >= 0 && is_index_range_check()) 599 return TypeInt::CC_LT; 600 return TypeInt::CC_LE; 601 } 602 } 603 // Check for special case in Hashtable::get - the hash index is 604 // mod'ed to the table size so the following range check is useless. 605 // Check for: (X Mod Y) CmpU Y, where the mod result and Y both have 606 // to be positive. 607 // (This is a gross hack, since the sub method never 608 // looks at the structure of the node in any other case.) 609 if ((jint)lo0 >= 0 && (jint)lo1 >= 0 && is_index_range_check()) 610 return TypeInt::CC_LT; 611 return TypeInt::CC; // else use worst case results 612 } 613 614 const Type* CmpUNode::Value(PhaseTransform *phase) const { 615 const Type* t = SubNode::Value_common(phase); 616 if (t != NULL) { 617 return t; 618 } 619 const Node* in1 = in(1); 620 const Node* in2 = in(2); 621 const Type* t1 = phase->type(in1); 622 const Type* t2 = phase->type(in2); 623 assert(t1->isa_int(), "CmpU has only Int type inputs"); 624 if (t2 == TypeInt::INT) { // Compare to bottom? 625 return bottom_type(); 626 } 627 uint in1_op = in1->Opcode(); 628 if (in1_op == Op_AddI || in1_op == Op_SubI) { 629 // The problem rise when result of AddI(SubI) may overflow 630 // signed integer value. Let say the input type is 631 // [256, maxint] then +128 will create 2 ranges due to 632 // overflow: [minint, minint+127] and [384, maxint]. 633 // But C2 type system keep only 1 type range and as result 634 // it use general [minint, maxint] for this case which we 635 // can't optimize. 636 // 637 // Make 2 separate type ranges based on types of AddI(SubI) inputs 638 // and compare results of their compare. If results are the same 639 // CmpU node can be optimized. 640 const Node* in11 = in1->in(1); 641 const Node* in12 = in1->in(2); 642 const Type* t11 = (in11 == in1) ? Type::TOP : phase->type(in11); 643 const Type* t12 = (in12 == in1) ? Type::TOP : phase->type(in12); 644 // Skip cases when input types are top or bottom. 645 if ((t11 != Type::TOP) && (t11 != TypeInt::INT) && 646 (t12 != Type::TOP) && (t12 != TypeInt::INT)) { 647 const TypeInt *r0 = t11->is_int(); 648 const TypeInt *r1 = t12->is_int(); 649 jlong lo_r0 = r0->_lo; 650 jlong hi_r0 = r0->_hi; 651 jlong lo_r1 = r1->_lo; 652 jlong hi_r1 = r1->_hi; 653 if (in1_op == Op_SubI) { 654 jlong tmp = hi_r1; 655 hi_r1 = -lo_r1; 656 lo_r1 = -tmp; 657 // Note, for substructing [minint,x] type range 658 // long arithmetic provides correct overflow answer. 659 // The confusion come from the fact that in 32-bit 660 // -minint == minint but in 64-bit -minint == maxint+1. 661 } 662 jlong lo_long = lo_r0 + lo_r1; 663 jlong hi_long = hi_r0 + hi_r1; 664 int lo_tr1 = min_jint; 665 int hi_tr1 = (int)hi_long; 666 int lo_tr2 = (int)lo_long; 667 int hi_tr2 = max_jint; 668 bool underflow = lo_long != (jlong)lo_tr2; 669 bool overflow = hi_long != (jlong)hi_tr1; 670 // Use sub(t1, t2) when there is no overflow (one type range) 671 // or when both overflow and underflow (too complex). 672 if ((underflow != overflow) && (hi_tr1 < lo_tr2)) { 673 // Overflow only on one boundary, compare 2 separate type ranges. 674 int w = MAX2(r0->_widen, r1->_widen); // _widen does not matter here 675 const TypeInt* tr1 = TypeInt::make(lo_tr1, hi_tr1, w); 676 const TypeInt* tr2 = TypeInt::make(lo_tr2, hi_tr2, w); 677 const Type* cmp1 = sub(tr1, t2); 678 const Type* cmp2 = sub(tr2, t2); 679 if (cmp1 == cmp2) { 680 return cmp1; // Hit! 681 } 682 } 683 } 684 } 685 686 return sub(t1, t2); // Local flavor of type subtraction 687 } 688 689 bool CmpUNode::is_index_range_check() const { 690 // Check for the "(X ModI Y) CmpU Y" shape 691 return (in(1)->Opcode() == Op_ModI && 692 in(1)->in(2)->eqv_uncast(in(2))); 693 } 694 695 //------------------------------Idealize--------------------------------------- 696 Node *CmpINode::Ideal( PhaseGVN *phase, bool can_reshape ) { 697 if (phase->type(in(2))->higher_equal(TypeInt::ZERO)) { 698 switch (in(1)->Opcode()) { 699 case Op_CmpL3: // Collapse a CmpL3/CmpI into a CmpL 700 return new CmpLNode(in(1)->in(1),in(1)->in(2)); 701 case Op_CmpF3: // Collapse a CmpF3/CmpI into a CmpF 702 return new CmpFNode(in(1)->in(1),in(1)->in(2)); 703 case Op_CmpD3: // Collapse a CmpD3/CmpI into a CmpD 704 return new CmpDNode(in(1)->in(1),in(1)->in(2)); 705 //case Op_SubI: 706 // If (x - y) cannot overflow, then ((x - y) <?> 0) 707 // can be turned into (x <?> y). 708 // This is handled (with more general cases) by Ideal_sub_algebra. 709 } 710 } 711 return NULL; // No change 712 } 713 714 715 //============================================================================= 716 // Simplify a CmpL (compare 2 longs ) node, based on local information. 717 // If both inputs are constants, compare them. 718 const Type *CmpLNode::sub( const Type *t1, const Type *t2 ) const { 719 const TypeLong *r0 = t1->is_long(); // Handy access 720 const TypeLong *r1 = t2->is_long(); 721 722 if( r0->_hi < r1->_lo ) // Range is always low? 723 return TypeInt::CC_LT; 724 else if( r0->_lo > r1->_hi ) // Range is always high? 725 return TypeInt::CC_GT; 726 727 else if( r0->is_con() && r1->is_con() ) { // comparing constants? 728 assert(r0->get_con() == r1->get_con(), "must be equal"); 729 return TypeInt::CC_EQ; // Equal results. 730 } else if( r0->_hi == r1->_lo ) // Range is never high? 731 return TypeInt::CC_LE; 732 else if( r0->_lo == r1->_hi ) // Range is never low? 733 return TypeInt::CC_GE; 734 return TypeInt::CC; // else use worst case results 735 } 736 737 //============================================================================= 738 //------------------------------sub-------------------------------------------- 739 // Simplify an CmpP (compare 2 pointers) node, based on local information. 740 // If both inputs are constants, compare them. 741 const Type *CmpPNode::sub( const Type *t1, const Type *t2 ) const { 742 const TypePtr *r0 = t1->is_ptr(); // Handy access 743 const TypePtr *r1 = t2->is_ptr(); 744 745 // Undefined inputs makes for an undefined result 746 if( TypePtr::above_centerline(r0->_ptr) || 747 TypePtr::above_centerline(r1->_ptr) ) 748 return Type::TOP; 749 750 if (r0 == r1 && r0->singleton()) { 751 // Equal pointer constants (klasses, nulls, etc.) 752 return TypeInt::CC_EQ; 753 } 754 755 // See if it is 2 unrelated classes. 756 const TypeOopPtr* p0 = r0->isa_oopptr(); 757 const TypeOopPtr* p1 = r1->isa_oopptr(); 758 if (p0 && p1) { 759 Node* in1 = in(1)->uncast(); 760 Node* in2 = in(2)->uncast(); 761 AllocateNode* alloc1 = AllocateNode::Ideal_allocation(in1, NULL); 762 AllocateNode* alloc2 = AllocateNode::Ideal_allocation(in2, NULL); 763 if (MemNode::detect_ptr_independence(in1, alloc1, in2, alloc2, NULL)) { 764 return TypeInt::CC_GT; // different pointers 765 } 766 ciKlass* klass0 = p0->klass(); 767 bool xklass0 = p0->klass_is_exact(); 768 ciKlass* klass1 = p1->klass(); 769 bool xklass1 = p1->klass_is_exact(); 770 int kps = (p0->isa_klassptr()?1:0) + (p1->isa_klassptr()?1:0); 771 if (klass0 && klass1 && 772 kps != 1 && // both or neither are klass pointers 773 klass0->is_loaded() && !klass0->is_interface() && // do not trust interfaces 774 klass1->is_loaded() && !klass1->is_interface() && 775 (!klass0->is_obj_array_klass() || 776 !klass0->as_obj_array_klass()->base_element_klass()->is_interface()) && 777 (!klass1->is_obj_array_klass() || 778 !klass1->as_obj_array_klass()->base_element_klass()->is_interface())) { 779 bool unrelated_classes = false; 780 // See if neither subclasses the other, or if the class on top 781 // is precise. In either of these cases, the compare is known 782 // to fail if at least one of the pointers is provably not null. 783 if (klass0->equals(klass1)) { // if types are unequal but klasses are equal 784 // Do nothing; we know nothing for imprecise types 785 } else if (klass0->is_subtype_of(klass1)) { 786 // If klass1's type is PRECISE, then classes are unrelated. 787 unrelated_classes = xklass1; 788 } else if (klass1->is_subtype_of(klass0)) { 789 // If klass0's type is PRECISE, then classes are unrelated. 790 unrelated_classes = xklass0; 791 } else { // Neither subtypes the other 792 unrelated_classes = true; 793 } 794 if (unrelated_classes) { 795 // The oops classes are known to be unrelated. If the joined PTRs of 796 // two oops is not Null and not Bottom, then we are sure that one 797 // of the two oops is non-null, and the comparison will always fail. 798 TypePtr::PTR jp = r0->join_ptr(r1->_ptr); 799 if (jp != TypePtr::Null && jp != TypePtr::BotPTR) { 800 return TypeInt::CC_GT; 801 } 802 } 803 } 804 } 805 806 // Known constants can be compared exactly 807 // Null can be distinguished from any NotNull pointers 808 // Unknown inputs makes an unknown result 809 if( r0->singleton() ) { 810 intptr_t bits0 = r0->get_con(); 811 if( r1->singleton() ) 812 return bits0 == r1->get_con() ? TypeInt::CC_EQ : TypeInt::CC_GT; 813 return ( r1->_ptr == TypePtr::NotNull && bits0==0 ) ? TypeInt::CC_GT : TypeInt::CC; 814 } else if( r1->singleton() ) { 815 intptr_t bits1 = r1->get_con(); 816 return ( r0->_ptr == TypePtr::NotNull && bits1==0 ) ? TypeInt::CC_GT : TypeInt::CC; 817 } else 818 return TypeInt::CC; 819 } 820 821 static inline Node* isa_java_mirror_load(PhaseGVN* phase, Node* n) { 822 // Return the klass node for 823 // LoadP(AddP(foo:Klass, #java_mirror)) 824 // or NULL if not matching. 825 if (n->Opcode() != Op_LoadP) return NULL; 826 827 const TypeInstPtr* tp = phase->type(n)->isa_instptr(); 828 if (!tp || tp->klass() != phase->C->env()->Class_klass()) return NULL; 829 830 Node* adr = n->in(MemNode::Address); 831 intptr_t off = 0; 832 Node* k = AddPNode::Ideal_base_and_offset(adr, phase, off); 833 if (k == NULL) return NULL; 834 const TypeKlassPtr* tkp = phase->type(k)->isa_klassptr(); 835 if (!tkp || off != in_bytes(Klass::java_mirror_offset())) return NULL; 836 837 // We've found the klass node of a Java mirror load. 838 return k; 839 } 840 841 static inline Node* isa_const_java_mirror(PhaseGVN* phase, Node* n) { 842 // for ConP(Foo.class) return ConP(Foo.klass) 843 // otherwise return NULL 844 if (!n->is_Con()) return NULL; 845 846 const TypeInstPtr* tp = phase->type(n)->isa_instptr(); 847 if (!tp) return NULL; 848 849 ciType* mirror_type = tp->java_mirror_type(); 850 // TypeInstPtr::java_mirror_type() returns non-NULL for compile- 851 // time Class constants only. 852 if (!mirror_type) return NULL; 853 854 // x.getClass() == int.class can never be true (for all primitive types) 855 // Return a ConP(NULL) node for this case. 856 if (mirror_type->is_classless()) { 857 return phase->makecon(TypePtr::NULL_PTR); 858 } 859 860 // return the ConP(Foo.klass) 861 assert(mirror_type->is_klass(), "mirror_type should represent a Klass*"); 862 return phase->makecon(TypeKlassPtr::make(mirror_type->as_klass())); 863 } 864 865 //------------------------------Ideal------------------------------------------ 866 // Normalize comparisons between Java mirror loads to compare the klass instead. 867 // 868 // Also check for the case of comparing an unknown klass loaded from the primary 869 // super-type array vs a known klass with no subtypes. This amounts to 870 // checking to see an unknown klass subtypes a known klass with no subtypes; 871 // this only happens on an exact match. We can shorten this test by 1 load. 872 Node *CmpPNode::Ideal( PhaseGVN *phase, bool can_reshape ) { 873 // Normalize comparisons between Java mirrors into comparisons of the low- 874 // level klass, where a dependent load could be shortened. 875 // 876 // The new pattern has a nice effect of matching the same pattern used in the 877 // fast path of instanceof/checkcast/Class.isInstance(), which allows 878 // redundant exact type check be optimized away by GVN. 879 // For example, in 880 // if (x.getClass() == Foo.class) { 881 // Foo foo = (Foo) x; 882 // // ... use a ... 883 // } 884 // a CmpPNode could be shared between if_acmpne and checkcast 885 { 886 Node* k1 = isa_java_mirror_load(phase, in(1)); 887 Node* k2 = isa_java_mirror_load(phase, in(2)); 888 Node* conk2 = isa_const_java_mirror(phase, in(2)); 889 890 if (k1 && (k2 || conk2)) { 891 Node* lhs = k1; 892 Node* rhs = (k2 != NULL) ? k2 : conk2; 893 this->set_req(1, lhs); 894 this->set_req(2, rhs); 895 return this; 896 } 897 } 898 899 // Constant pointer on right? 900 const TypeKlassPtr* t2 = phase->type(in(2))->isa_klassptr(); 901 if (t2 == NULL || !t2->klass_is_exact()) 902 return NULL; 903 // Get the constant klass we are comparing to. 904 ciKlass* superklass = t2->klass(); 905 906 // Now check for LoadKlass on left. 907 Node* ldk1 = in(1); 908 if (ldk1->is_DecodeNKlass()) { 909 ldk1 = ldk1->in(1); 910 if (ldk1->Opcode() != Op_LoadNKlass ) 911 return NULL; 912 } else if (ldk1->Opcode() != Op_LoadKlass ) 913 return NULL; 914 // Take apart the address of the LoadKlass: 915 Node* adr1 = ldk1->in(MemNode::Address); 916 intptr_t con2 = 0; 917 Node* ldk2 = AddPNode::Ideal_base_and_offset(adr1, phase, con2); 918 if (ldk2 == NULL) 919 return NULL; 920 if (con2 == oopDesc::klass_offset_in_bytes()) { 921 // We are inspecting an object's concrete class. 922 // Short-circuit the check if the query is abstract. 923 if (superklass->is_interface() || 924 superklass->is_abstract()) { 925 // Make it come out always false: 926 this->set_req(2, phase->makecon(TypePtr::NULL_PTR)); 927 return this; 928 } 929 } 930 931 // Check for a LoadKlass from primary supertype array. 932 // Any nested loadklass from loadklass+con must be from the p.s. array. 933 if (ldk2->is_DecodeNKlass()) { 934 // Keep ldk2 as DecodeN since it could be used in CmpP below. 935 if (ldk2->in(1)->Opcode() != Op_LoadNKlass ) 936 return NULL; 937 } else if (ldk2->Opcode() != Op_LoadKlass) 938 return NULL; 939 940 // Verify that we understand the situation 941 if (con2 != (intptr_t) superklass->super_check_offset()) 942 return NULL; // Might be element-klass loading from array klass 943 944 // If 'superklass' has no subklasses and is not an interface, then we are 945 // assured that the only input which will pass the type check is 946 // 'superklass' itself. 947 // 948 // We could be more liberal here, and allow the optimization on interfaces 949 // which have a single implementor. This would require us to increase the 950 // expressiveness of the add_dependency() mechanism. 951 // %%% Do this after we fix TypeOopPtr: Deps are expressive enough now. 952 953 // Object arrays must have their base element have no subtypes 954 while (superklass->is_obj_array_klass()) { 955 ciType* elem = superklass->as_obj_array_klass()->element_type(); 956 superklass = elem->as_klass(); 957 } 958 if (superklass->is_instance_klass()) { 959 ciInstanceKlass* ik = superklass->as_instance_klass(); 960 if (ik->has_subklass() || ik->is_interface()) return NULL; 961 // Add a dependency if there is a chance that a subclass will be added later. 962 if (!ik->is_final()) { 963 phase->C->dependencies()->assert_leaf_type(ik); 964 } 965 } 966 967 // Bypass the dependent load, and compare directly 968 this->set_req(1,ldk2); 969 970 return this; 971 } 972 973 //============================================================================= 974 //------------------------------sub-------------------------------------------- 975 // Simplify an CmpN (compare 2 pointers) node, based on local information. 976 // If both inputs are constants, compare them. 977 const Type *CmpNNode::sub( const Type *t1, const Type *t2 ) const { 978 const TypePtr *r0 = t1->make_ptr(); // Handy access 979 const TypePtr *r1 = t2->make_ptr(); 980 981 // Undefined inputs makes for an undefined result 982 if ((r0 == NULL) || (r1 == NULL) || 983 TypePtr::above_centerline(r0->_ptr) || 984 TypePtr::above_centerline(r1->_ptr)) { 985 return Type::TOP; 986 } 987 if (r0 == r1 && r0->singleton()) { 988 // Equal pointer constants (klasses, nulls, etc.) 989 return TypeInt::CC_EQ; 990 } 991 992 // See if it is 2 unrelated classes. 993 const TypeOopPtr* p0 = r0->isa_oopptr(); 994 const TypeOopPtr* p1 = r1->isa_oopptr(); 995 if (p0 && p1) { 996 ciKlass* klass0 = p0->klass(); 997 bool xklass0 = p0->klass_is_exact(); 998 ciKlass* klass1 = p1->klass(); 999 bool xklass1 = p1->klass_is_exact(); 1000 int kps = (p0->isa_klassptr()?1:0) + (p1->isa_klassptr()?1:0); 1001 if (klass0 && klass1 && 1002 kps != 1 && // both or neither are klass pointers 1003 !klass0->is_interface() && // do not trust interfaces 1004 !klass1->is_interface()) { 1005 bool unrelated_classes = false; 1006 // See if neither subclasses the other, or if the class on top 1007 // is precise. In either of these cases, the compare is known 1008 // to fail if at least one of the pointers is provably not null. 1009 if (klass0->equals(klass1)) { // if types are unequal but klasses are equal 1010 // Do nothing; we know nothing for imprecise types 1011 } else if (klass0->is_subtype_of(klass1)) { 1012 // If klass1's type is PRECISE, then classes are unrelated. 1013 unrelated_classes = xklass1; 1014 } else if (klass1->is_subtype_of(klass0)) { 1015 // If klass0's type is PRECISE, then classes are unrelated. 1016 unrelated_classes = xklass0; 1017 } else { // Neither subtypes the other 1018 unrelated_classes = true; 1019 } 1020 if (unrelated_classes) { 1021 // The oops classes are known to be unrelated. If the joined PTRs of 1022 // two oops is not Null and not Bottom, then we are sure that one 1023 // of the two oops is non-null, and the comparison will always fail. 1024 TypePtr::PTR jp = r0->join_ptr(r1->_ptr); 1025 if (jp != TypePtr::Null && jp != TypePtr::BotPTR) { 1026 return TypeInt::CC_GT; 1027 } 1028 } 1029 } 1030 } 1031 1032 // Known constants can be compared exactly 1033 // Null can be distinguished from any NotNull pointers 1034 // Unknown inputs makes an unknown result 1035 if( r0->singleton() ) { 1036 intptr_t bits0 = r0->get_con(); 1037 if( r1->singleton() ) 1038 return bits0 == r1->get_con() ? TypeInt::CC_EQ : TypeInt::CC_GT; 1039 return ( r1->_ptr == TypePtr::NotNull && bits0==0 ) ? TypeInt::CC_GT : TypeInt::CC; 1040 } else if( r1->singleton() ) { 1041 intptr_t bits1 = r1->get_con(); 1042 return ( r0->_ptr == TypePtr::NotNull && bits1==0 ) ? TypeInt::CC_GT : TypeInt::CC; 1043 } else 1044 return TypeInt::CC; 1045 } 1046 1047 //------------------------------Ideal------------------------------------------ 1048 Node *CmpNNode::Ideal( PhaseGVN *phase, bool can_reshape ) { 1049 return NULL; 1050 } 1051 1052 //============================================================================= 1053 //------------------------------Value------------------------------------------ 1054 // Simplify an CmpF (compare 2 floats ) node, based on local information. 1055 // If both inputs are constants, compare them. 1056 const Type *CmpFNode::Value( PhaseTransform *phase ) const { 1057 const Node* in1 = in(1); 1058 const Node* in2 = in(2); 1059 // Either input is TOP ==> the result is TOP 1060 const Type* t1 = (in1 == this) ? Type::TOP : phase->type(in1); 1061 if( t1 == Type::TOP ) return Type::TOP; 1062 const Type* t2 = (in2 == this) ? Type::TOP : phase->type(in2); 1063 if( t2 == Type::TOP ) return Type::TOP; 1064 1065 // Not constants? Don't know squat - even if they are the same 1066 // value! If they are NaN's they compare to LT instead of EQ. 1067 const TypeF *tf1 = t1->isa_float_constant(); 1068 const TypeF *tf2 = t2->isa_float_constant(); 1069 if( !tf1 || !tf2 ) return TypeInt::CC; 1070 1071 // This implements the Java bytecode fcmpl, so unordered returns -1. 1072 if( tf1->is_nan() || tf2->is_nan() ) 1073 return TypeInt::CC_LT; 1074 1075 if( tf1->_f < tf2->_f ) return TypeInt::CC_LT; 1076 if( tf1->_f > tf2->_f ) return TypeInt::CC_GT; 1077 assert( tf1->_f == tf2->_f, "do not understand FP behavior" ); 1078 return TypeInt::CC_EQ; 1079 } 1080 1081 1082 //============================================================================= 1083 //------------------------------Value------------------------------------------ 1084 // Simplify an CmpD (compare 2 doubles ) node, based on local information. 1085 // If both inputs are constants, compare them. 1086 const Type *CmpDNode::Value( PhaseTransform *phase ) const { 1087 const Node* in1 = in(1); 1088 const Node* in2 = in(2); 1089 // Either input is TOP ==> the result is TOP 1090 const Type* t1 = (in1 == this) ? Type::TOP : phase->type(in1); 1091 if( t1 == Type::TOP ) return Type::TOP; 1092 const Type* t2 = (in2 == this) ? Type::TOP : phase->type(in2); 1093 if( t2 == Type::TOP ) return Type::TOP; 1094 1095 // Not constants? Don't know squat - even if they are the same 1096 // value! If they are NaN's they compare to LT instead of EQ. 1097 const TypeD *td1 = t1->isa_double_constant(); 1098 const TypeD *td2 = t2->isa_double_constant(); 1099 if( !td1 || !td2 ) return TypeInt::CC; 1100 1101 // This implements the Java bytecode dcmpl, so unordered returns -1. 1102 if( td1->is_nan() || td2->is_nan() ) 1103 return TypeInt::CC_LT; 1104 1105 if( td1->_d < td2->_d ) return TypeInt::CC_LT; 1106 if( td1->_d > td2->_d ) return TypeInt::CC_GT; 1107 assert( td1->_d == td2->_d, "do not understand FP behavior" ); 1108 return TypeInt::CC_EQ; 1109 } 1110 1111 //------------------------------Ideal------------------------------------------ 1112 Node *CmpDNode::Ideal(PhaseGVN *phase, bool can_reshape){ 1113 // Check if we can change this to a CmpF and remove a ConvD2F operation. 1114 // Change (CMPD (F2D (float)) (ConD value)) 1115 // To (CMPF (float) (ConF value)) 1116 // Valid when 'value' does not lose precision as a float. 1117 // Benefits: eliminates conversion, does not require 24-bit mode 1118 1119 // NaNs prevent commuting operands. This transform works regardless of the 1120 // order of ConD and ConvF2D inputs by preserving the original order. 1121 int idx_f2d = 1; // ConvF2D on left side? 1122 if( in(idx_f2d)->Opcode() != Op_ConvF2D ) 1123 idx_f2d = 2; // No, swap to check for reversed args 1124 int idx_con = 3-idx_f2d; // Check for the constant on other input 1125 1126 if( ConvertCmpD2CmpF && 1127 in(idx_f2d)->Opcode() == Op_ConvF2D && 1128 in(idx_con)->Opcode() == Op_ConD ) { 1129 const TypeD *t2 = in(idx_con)->bottom_type()->is_double_constant(); 1130 double t2_value_as_double = t2->_d; 1131 float t2_value_as_float = (float)t2_value_as_double; 1132 if( t2_value_as_double == (double)t2_value_as_float ) { 1133 // Test value can be represented as a float 1134 // Eliminate the conversion to double and create new comparison 1135 Node *new_in1 = in(idx_f2d)->in(1); 1136 Node *new_in2 = phase->makecon( TypeF::make(t2_value_as_float) ); 1137 if( idx_f2d != 1 ) { // Must flip args to match original order 1138 Node *tmp = new_in1; 1139 new_in1 = new_in2; 1140 new_in2 = tmp; 1141 } 1142 CmpFNode *new_cmp = (Opcode() == Op_CmpD3) 1143 ? new CmpF3Node( new_in1, new_in2 ) 1144 : new CmpFNode ( new_in1, new_in2 ) ; 1145 return new_cmp; // Changed to CmpFNode 1146 } 1147 // Testing value required the precision of a double 1148 } 1149 return NULL; // No change 1150 } 1151 1152 1153 //============================================================================= 1154 //------------------------------cc2logical------------------------------------- 1155 // Convert a condition code type to a logical type 1156 const Type *BoolTest::cc2logical( const Type *CC ) const { 1157 if( CC == Type::TOP ) return Type::TOP; 1158 if( CC->base() != Type::Int ) return TypeInt::BOOL; // Bottom or worse 1159 const TypeInt *ti = CC->is_int(); 1160 if( ti->is_con() ) { // Only 1 kind of condition codes set? 1161 // Match low order 2 bits 1162 int tmp = ((ti->get_con()&3) == (_test&3)) ? 1 : 0; 1163 if( _test & 4 ) tmp = 1-tmp; // Optionally complement result 1164 return TypeInt::make(tmp); // Boolean result 1165 } 1166 1167 if( CC == TypeInt::CC_GE ) { 1168 if( _test == ge ) return TypeInt::ONE; 1169 if( _test == lt ) return TypeInt::ZERO; 1170 } 1171 if( CC == TypeInt::CC_LE ) { 1172 if( _test == le ) return TypeInt::ONE; 1173 if( _test == gt ) return TypeInt::ZERO; 1174 } 1175 1176 return TypeInt::BOOL; 1177 } 1178 1179 //------------------------------dump_spec------------------------------------- 1180 // Print special per-node info 1181 void BoolTest::dump_on(outputStream *st) const { 1182 const char *msg[] = {"eq","gt","of","lt","ne","le","nof","ge"}; 1183 st->print("%s", msg[_test]); 1184 } 1185 1186 //============================================================================= 1187 uint BoolNode::hash() const { return (Node::hash() << 3)|(_test._test+1); } 1188 uint BoolNode::size_of() const { return sizeof(BoolNode); } 1189 1190 //------------------------------operator==------------------------------------- 1191 uint BoolNode::cmp( const Node &n ) const { 1192 const BoolNode *b = (const BoolNode *)&n; // Cast up 1193 return (_test._test == b->_test._test); 1194 } 1195 1196 //-------------------------------make_predicate-------------------------------- 1197 Node* BoolNode::make_predicate(Node* test_value, PhaseGVN* phase) { 1198 if (test_value->is_Con()) return test_value; 1199 if (test_value->is_Bool()) return test_value; 1200 if (test_value->is_CMove() && 1201 test_value->in(CMoveNode::Condition)->is_Bool()) { 1202 BoolNode* bol = test_value->in(CMoveNode::Condition)->as_Bool(); 1203 const Type* ftype = phase->type(test_value->in(CMoveNode::IfFalse)); 1204 const Type* ttype = phase->type(test_value->in(CMoveNode::IfTrue)); 1205 if (ftype == TypeInt::ZERO && !TypeInt::ZERO->higher_equal(ttype)) { 1206 return bol; 1207 } else if (ttype == TypeInt::ZERO && !TypeInt::ZERO->higher_equal(ftype)) { 1208 return phase->transform( bol->negate(phase) ); 1209 } 1210 // Else fall through. The CMove gets in the way of the test. 1211 // It should be the case that make_predicate(bol->as_int_value()) == bol. 1212 } 1213 Node* cmp = new CmpINode(test_value, phase->intcon(0)); 1214 cmp = phase->transform(cmp); 1215 Node* bol = new BoolNode(cmp, BoolTest::ne); 1216 return phase->transform(bol); 1217 } 1218 1219 //--------------------------------as_int_value--------------------------------- 1220 Node* BoolNode::as_int_value(PhaseGVN* phase) { 1221 // Inverse to make_predicate. The CMove probably boils down to a Conv2B. 1222 Node* cmov = CMoveNode::make(NULL, this, 1223 phase->intcon(0), phase->intcon(1), 1224 TypeInt::BOOL); 1225 return phase->transform(cmov); 1226 } 1227 1228 //----------------------------------negate------------------------------------- 1229 BoolNode* BoolNode::negate(PhaseGVN* phase) { 1230 return new BoolNode(in(1), _test.negate()); 1231 } 1232 1233 // Change "bool eq/ne (cmp (add/sub A B) C)" into false/true if add/sub 1234 // overflows and we can prove that C is not in the two resulting ranges. 1235 // This optimization is similar to the one performed by CmpUNode::Value(). 1236 Node* BoolNode::fold_cmpI(PhaseGVN* phase, SubNode* cmp, Node* cmp1, int cmp_op, 1237 int cmp1_op, const TypeInt* cmp2_type) { 1238 // Only optimize eq/ne integer comparison of add/sub 1239 if((_test._test == BoolTest::eq || _test._test == BoolTest::ne) && 1240 (cmp_op == Op_CmpI) && (cmp1_op == Op_AddI || cmp1_op == Op_SubI)) { 1241 // Skip cases were inputs of add/sub are not integers or of bottom type 1242 const TypeInt* r0 = phase->type(cmp1->in(1))->isa_int(); 1243 const TypeInt* r1 = phase->type(cmp1->in(2))->isa_int(); 1244 if ((r0 != NULL) && (r0 != TypeInt::INT) && 1245 (r1 != NULL) && (r1 != TypeInt::INT) && 1246 (cmp2_type != TypeInt::INT)) { 1247 // Compute exact (long) type range of add/sub result 1248 jlong lo_long = r0->_lo; 1249 jlong hi_long = r0->_hi; 1250 if (cmp1_op == Op_AddI) { 1251 lo_long += r1->_lo; 1252 hi_long += r1->_hi; 1253 } else { 1254 lo_long -= r1->_hi; 1255 hi_long -= r1->_lo; 1256 } 1257 // Check for over-/underflow by casting to integer 1258 int lo_int = (int)lo_long; 1259 int hi_int = (int)hi_long; 1260 bool underflow = lo_long != (jlong)lo_int; 1261 bool overflow = hi_long != (jlong)hi_int; 1262 if ((underflow != overflow) && (hi_int < lo_int)) { 1263 // Overflow on one boundary, compute resulting type ranges: 1264 // tr1 [MIN_INT, hi_int] and tr2 [lo_int, MAX_INT] 1265 int w = MAX2(r0->_widen, r1->_widen); // _widen does not matter here 1266 const TypeInt* tr1 = TypeInt::make(min_jint, hi_int, w); 1267 const TypeInt* tr2 = TypeInt::make(lo_int, max_jint, w); 1268 // Compare second input of cmp to both type ranges 1269 const Type* sub_tr1 = cmp->sub(tr1, cmp2_type); 1270 const Type* sub_tr2 = cmp->sub(tr2, cmp2_type); 1271 if (sub_tr1 == TypeInt::CC_LT && sub_tr2 == TypeInt::CC_GT) { 1272 // The result of the add/sub will never equal cmp2. Replace BoolNode 1273 // by false (0) if it tests for equality and by true (1) otherwise. 1274 return ConINode::make((_test._test == BoolTest::eq) ? 0 : 1); 1275 } 1276 } 1277 } 1278 } 1279 return NULL; 1280 } 1281 1282 //------------------------------Ideal------------------------------------------ 1283 Node *BoolNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1284 // Change "bool tst (cmp con x)" into "bool ~tst (cmp x con)". 1285 // This moves the constant to the right. Helps value-numbering. 1286 Node *cmp = in(1); 1287 if( !cmp->is_Sub() ) return NULL; 1288 int cop = cmp->Opcode(); 1289 if( cop == Op_FastLock || cop == Op_FastUnlock) return NULL; 1290 Node *cmp1 = cmp->in(1); 1291 Node *cmp2 = cmp->in(2); 1292 if( !cmp1 ) return NULL; 1293 1294 if (_test._test == BoolTest::overflow || _test._test == BoolTest::no_overflow) { 1295 return NULL; 1296 } 1297 1298 // Constant on left? 1299 Node *con = cmp1; 1300 uint op2 = cmp2->Opcode(); 1301 // Move constants to the right of compare's to canonicalize. 1302 // Do not muck with Opaque1 nodes, as this indicates a loop 1303 // guard that cannot change shape. 1304 if( con->is_Con() && !cmp2->is_Con() && op2 != Op_Opaque1 && 1305 // Because of NaN's, CmpD and CmpF are not commutative 1306 cop != Op_CmpD && cop != Op_CmpF && 1307 // Protect against swapping inputs to a compare when it is used by a 1308 // counted loop exit, which requires maintaining the loop-limit as in(2) 1309 !is_counted_loop_exit_test() ) { 1310 // Ok, commute the constant to the right of the cmp node. 1311 // Clone the Node, getting a new Node of the same class 1312 cmp = cmp->clone(); 1313 // Swap inputs to the clone 1314 cmp->swap_edges(1, 2); 1315 cmp = phase->transform( cmp ); 1316 return new BoolNode( cmp, _test.commute() ); 1317 } 1318 1319 // Change "bool eq/ne (cmp (xor X 1) 0)" into "bool ne/eq (cmp X 0)". 1320 // The XOR-1 is an idiom used to flip the sense of a bool. We flip the 1321 // test instead. 1322 int cmp1_op = cmp1->Opcode(); 1323 const TypeInt* cmp2_type = phase->type(cmp2)->isa_int(); 1324 if (cmp2_type == NULL) return NULL; 1325 Node* j_xor = cmp1; 1326 if( cmp2_type == TypeInt::ZERO && 1327 cmp1_op == Op_XorI && 1328 j_xor->in(1) != j_xor && // An xor of itself is dead 1329 phase->type( j_xor->in(1) ) == TypeInt::BOOL && 1330 phase->type( j_xor->in(2) ) == TypeInt::ONE && 1331 (_test._test == BoolTest::eq || 1332 _test._test == BoolTest::ne) ) { 1333 Node *ncmp = phase->transform(new CmpINode(j_xor->in(1),cmp2)); 1334 return new BoolNode( ncmp, _test.negate() ); 1335 } 1336 1337 // Change "bool eq/ne (cmp (Conv2B X) 0)" into "bool eq/ne (cmp X 0)". 1338 // This is a standard idiom for branching on a boolean value. 1339 Node *c2b = cmp1; 1340 if( cmp2_type == TypeInt::ZERO && 1341 cmp1_op == Op_Conv2B && 1342 (_test._test == BoolTest::eq || 1343 _test._test == BoolTest::ne) ) { 1344 Node *ncmp = phase->transform(phase->type(c2b->in(1))->isa_int() 1345 ? (Node*)new CmpINode(c2b->in(1),cmp2) 1346 : (Node*)new CmpPNode(c2b->in(1),phase->makecon(TypePtr::NULL_PTR)) 1347 ); 1348 return new BoolNode( ncmp, _test._test ); 1349 } 1350 1351 // Comparing a SubI against a zero is equal to comparing the SubI 1352 // arguments directly. This only works for eq and ne comparisons 1353 // due to possible integer overflow. 1354 if ((_test._test == BoolTest::eq || _test._test == BoolTest::ne) && 1355 (cop == Op_CmpI) && 1356 (cmp1->Opcode() == Op_SubI) && 1357 ( cmp2_type == TypeInt::ZERO ) ) { 1358 Node *ncmp = phase->transform( new CmpINode(cmp1->in(1),cmp1->in(2))); 1359 return new BoolNode( ncmp, _test._test ); 1360 } 1361 1362 // Change (-A vs 0) into (A vs 0) by commuting the test. Disallow in the 1363 // most general case because negating 0x80000000 does nothing. Needed for 1364 // the CmpF3/SubI/CmpI idiom. 1365 if( cop == Op_CmpI && 1366 cmp1->Opcode() == Op_SubI && 1367 cmp2_type == TypeInt::ZERO && 1368 phase->type( cmp1->in(1) ) == TypeInt::ZERO && 1369 phase->type( cmp1->in(2) )->higher_equal(TypeInt::SYMINT) ) { 1370 Node *ncmp = phase->transform( new CmpINode(cmp1->in(2),cmp2)); 1371 return new BoolNode( ncmp, _test.commute() ); 1372 } 1373 1374 // Try to optimize signed integer comparison 1375 return fold_cmpI(phase, cmp->as_Sub(), cmp1, cop, cmp1_op, cmp2_type); 1376 1377 // The transformation below is not valid for either signed or unsigned 1378 // comparisons due to wraparound concerns at MAX_VALUE and MIN_VALUE. 1379 // This transformation can be resurrected when we are able to 1380 // make inferences about the range of values being subtracted from 1381 // (or added to) relative to the wraparound point. 1382 // 1383 // // Remove +/-1's if possible. 1384 // // "X <= Y-1" becomes "X < Y" 1385 // // "X+1 <= Y" becomes "X < Y" 1386 // // "X < Y+1" becomes "X <= Y" 1387 // // "X-1 < Y" becomes "X <= Y" 1388 // // Do not this to compares off of the counted-loop-end. These guys are 1389 // // checking the trip counter and they want to use the post-incremented 1390 // // counter. If they use the PRE-incremented counter, then the counter has 1391 // // to be incremented in a private block on a loop backedge. 1392 // if( du && du->cnt(this) && du->out(this)[0]->Opcode() == Op_CountedLoopEnd ) 1393 // return NULL; 1394 // #ifndef PRODUCT 1395 // // Do not do this in a wash GVN pass during verification. 1396 // // Gets triggered by too many simple optimizations to be bothered with 1397 // // re-trying it again and again. 1398 // if( !phase->allow_progress() ) return NULL; 1399 // #endif 1400 // // Not valid for unsigned compare because of corner cases in involving zero. 1401 // // For example, replacing "X-1 <u Y" with "X <=u Y" fails to throw an 1402 // // exception in case X is 0 (because 0-1 turns into 4billion unsigned but 1403 // // "0 <=u Y" is always true). 1404 // if( cmp->Opcode() == Op_CmpU ) return NULL; 1405 // int cmp2_op = cmp2->Opcode(); 1406 // if( _test._test == BoolTest::le ) { 1407 // if( cmp1_op == Op_AddI && 1408 // phase->type( cmp1->in(2) ) == TypeInt::ONE ) 1409 // return clone_cmp( cmp, cmp1->in(1), cmp2, phase, BoolTest::lt ); 1410 // else if( cmp2_op == Op_AddI && 1411 // phase->type( cmp2->in(2) ) == TypeInt::MINUS_1 ) 1412 // return clone_cmp( cmp, cmp1, cmp2->in(1), phase, BoolTest::lt ); 1413 // } else if( _test._test == BoolTest::lt ) { 1414 // if( cmp1_op == Op_AddI && 1415 // phase->type( cmp1->in(2) ) == TypeInt::MINUS_1 ) 1416 // return clone_cmp( cmp, cmp1->in(1), cmp2, phase, BoolTest::le ); 1417 // else if( cmp2_op == Op_AddI && 1418 // phase->type( cmp2->in(2) ) == TypeInt::ONE ) 1419 // return clone_cmp( cmp, cmp1, cmp2->in(1), phase, BoolTest::le ); 1420 // } 1421 } 1422 1423 //------------------------------Value------------------------------------------ 1424 // Simplify a Bool (convert condition codes to boolean (1 or 0)) node, 1425 // based on local information. If the input is constant, do it. 1426 const Type *BoolNode::Value( PhaseTransform *phase ) const { 1427 return _test.cc2logical( phase->type( in(1) ) ); 1428 } 1429 1430 #ifndef PRODUCT 1431 //------------------------------dump_spec-------------------------------------- 1432 // Dump special per-node info 1433 void BoolNode::dump_spec(outputStream *st) const { 1434 st->print("["); 1435 _test.dump_on(st); 1436 st->print("]"); 1437 } 1438 1439 //-------------------------------related--------------------------------------- 1440 // A BoolNode's related nodes are all of its data inputs, and all of its 1441 // outputs until control nodes are hit, which are included. In compact 1442 // representation, inputs till level 3 and immediate outputs are included. 1443 void BoolNode::related(GrowableArray<Node*> *in_rel, GrowableArray<Node*> *out_rel, bool compact) const { 1444 if (compact) { 1445 this->collect_nodes(in_rel, 3, false, true); 1446 this->collect_nodes(out_rel, -1, false, false); 1447 } else { 1448 this->collect_nodes_in_all_data(in_rel, false); 1449 this->collect_nodes_out_all_ctrl_boundary(out_rel); 1450 } 1451 } 1452 #endif 1453 1454 //----------------------is_counted_loop_exit_test------------------------------ 1455 // Returns true if node is used by a counted loop node. 1456 bool BoolNode::is_counted_loop_exit_test() { 1457 for( DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++ ) { 1458 Node* use = fast_out(i); 1459 if (use->is_CountedLoopEnd()) { 1460 return true; 1461 } 1462 } 1463 return false; 1464 } 1465 1466 //============================================================================= 1467 //------------------------------Value------------------------------------------ 1468 // Compute sqrt 1469 const Type *SqrtDNode::Value( PhaseTransform *phase ) const { 1470 const Type *t1 = phase->type( in(1) ); 1471 if( t1 == Type::TOP ) return Type::TOP; 1472 if( t1->base() != Type::DoubleCon ) return Type::DOUBLE; 1473 double d = t1->getd(); 1474 if( d < 0.0 ) return Type::DOUBLE; 1475 return TypeD::make( sqrt( d ) ); 1476 } 1477 1478 //============================================================================= 1479 //------------------------------Value------------------------------------------ 1480 // Compute cos 1481 const Type *CosDNode::Value( PhaseTransform *phase ) const { 1482 const Type *t1 = phase->type( in(1) ); 1483 if( t1 == Type::TOP ) return Type::TOP; 1484 if( t1->base() != Type::DoubleCon ) return Type::DOUBLE; 1485 double d = t1->getd(); 1486 return TypeD::make( StubRoutines::intrinsic_cos( d ) ); 1487 } 1488 1489 //============================================================================= 1490 //------------------------------Value------------------------------------------ 1491 // Compute sin 1492 const Type *SinDNode::Value( PhaseTransform *phase ) const { 1493 const Type *t1 = phase->type( in(1) ); 1494 if( t1 == Type::TOP ) return Type::TOP; 1495 if( t1->base() != Type::DoubleCon ) return Type::DOUBLE; 1496 double d = t1->getd(); 1497 return TypeD::make( StubRoutines::intrinsic_sin( d ) ); 1498 } 1499 1500 //============================================================================= 1501 //------------------------------Value------------------------------------------ 1502 // Compute tan 1503 const Type *TanDNode::Value( PhaseTransform *phase ) const { 1504 const Type *t1 = phase->type( in(1) ); 1505 if( t1 == Type::TOP ) return Type::TOP; 1506 if( t1->base() != Type::DoubleCon ) return Type::DOUBLE; 1507 double d = t1->getd(); 1508 return TypeD::make( StubRoutines::intrinsic_tan( d ) ); 1509 } 1510 1511 //============================================================================= 1512 //------------------------------Value------------------------------------------ 1513 // Compute log10 1514 const Type *Log10DNode::Value( PhaseTransform *phase ) const { 1515 const Type *t1 = phase->type( in(1) ); 1516 if( t1 == Type::TOP ) return Type::TOP; 1517 if( t1->base() != Type::DoubleCon ) return Type::DOUBLE; 1518 double d = t1->getd(); 1519 return TypeD::make( StubRoutines::intrinsic_log10( d ) ); 1520 } 1521 1522 //============================================================================= 1523 //------------------------------Value------------------------------------------ 1524 // Compute pow 1525 const Type *PowDNode::Value( PhaseTransform *phase ) const { 1526 const Type *t1 = phase->type( in(1) ); 1527 if( t1 == Type::TOP ) return Type::TOP; 1528 if( t1->base() != Type::DoubleCon ) return Type::DOUBLE; 1529 const Type *t2 = phase->type( in(2) ); 1530 if( t2 == Type::TOP ) return Type::TOP; 1531 if( t2->base() != Type::DoubleCon ) return Type::DOUBLE; 1532 double d1 = t1->getd(); 1533 double d2 = t2->getd(); 1534 return TypeD::make( StubRoutines::intrinsic_pow( d1, d2 ) ); 1535 }