1 /* 2 * Copyright (c) 1997, 2025, 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 "compiler/compileLog.hpp" 26 #include "gc/shared/barrierSet.hpp" 27 #include "gc/shared/c2/barrierSetC2.hpp" 28 #include "memory/allocation.inline.hpp" 29 #include "opto/addnode.hpp" 30 #include "opto/callnode.hpp" 31 #include "opto/cfgnode.hpp" 32 #include "opto/loopnode.hpp" 33 #include "opto/matcher.hpp" 34 #include "opto/movenode.hpp" 35 #include "opto/mulnode.hpp" 36 #include "opto/opaquenode.hpp" 37 #include "opto/opcodes.hpp" 38 #include "opto/phaseX.hpp" 39 #include "opto/subnode.hpp" 40 #include "runtime/sharedRuntime.hpp" 41 #include "utilities/reverse_bits.hpp" 42 43 // Portions of code courtesy of Clifford Click 44 45 // Optimization - Graph Style 46 47 #include "math.h" 48 49 //============================================================================= 50 //------------------------------Identity--------------------------------------- 51 // If right input is a constant 0, return the left input. 52 Node* SubNode::Identity(PhaseGVN* phase) { 53 assert(in(1) != this, "Must already have called Value"); 54 assert(in(2) != this, "Must already have called Value"); 55 56 const Type* zero = add_id(); 57 58 // Remove double negation if it is not a floating point number since negation 59 // is not the same as subtraction for floating point numbers 60 // (cf. JLS § 15.15.4). `0-(0-(-0.0))` must be equal to positive 0.0 according to 61 // JLS § 15.8.2, but would result in -0.0 if this folding would be applied. 62 if (phase->type(in(1))->higher_equal(zero) && 63 in(2)->Opcode() == Opcode() && 64 phase->type(in(2)->in(1))->higher_equal(zero) && 65 !phase->type(in(2)->in(2))->is_floatingpoint()) { 66 return in(2)->in(2); 67 } 68 69 // Convert "(X+Y) - Y" into X and "(X+Y) - X" into Y 70 if (in(1)->Opcode() == Op_AddI || in(1)->Opcode() == Op_AddL) { 71 if (in(1)->in(2) == in(2)) { 72 return in(1)->in(1); 73 } 74 if (in(1)->in(1) == in(2)) { 75 return in(1)->in(2); 76 } 77 } 78 79 return ( phase->type( in(2) )->higher_equal( zero ) ) ? in(1) : this; 80 } 81 82 //------------------------------Value------------------------------------------ 83 // A subtract node differences it's two inputs. 84 const Type* SubNode::Value_common(PhaseValues* phase) const { 85 const Node* in1 = in(1); 86 const Node* in2 = in(2); 87 // Either input is TOP ==> the result is TOP 88 const Type* t1 = (in1 == this) ? Type::TOP : phase->type(in1); 89 if( t1 == Type::TOP ) return Type::TOP; 90 const Type* t2 = (in2 == this) ? Type::TOP : phase->type(in2); 91 if( t2 == Type::TOP ) return Type::TOP; 92 93 // Not correct for SubFnode and AddFNode (must check for infinity) 94 // Equal? Subtract is zero 95 if (in1->eqv_uncast(in2)) return add_id(); 96 97 // Either input is BOTTOM ==> the result is the local BOTTOM 98 if( t1 == Type::BOTTOM || t2 == Type::BOTTOM ) 99 return bottom_type(); 100 101 return nullptr; 102 } 103 104 const Type* SubNode::Value(PhaseGVN* phase) const { 105 const Type* t = Value_common(phase); 106 if (t != nullptr) { 107 return t; 108 } 109 const Type* t1 = phase->type(in(1)); 110 const Type* t2 = phase->type(in(2)); 111 return sub(t1,t2); // Local flavor of type subtraction 112 113 } 114 115 SubNode* SubNode::make(Node* in1, Node* in2, BasicType bt) { 116 switch (bt) { 117 case T_INT: 118 return new SubINode(in1, in2); 119 case T_LONG: 120 return new SubLNode(in1, in2); 121 default: 122 fatal("Not implemented for %s", type2name(bt)); 123 } 124 return nullptr; 125 } 126 127 //============================================================================= 128 //------------------------------Helper function-------------------------------- 129 130 static bool is_cloop_increment(Node* inc) { 131 precond(inc->Opcode() == Op_AddI || inc->Opcode() == Op_AddL); 132 133 if (!inc->in(1)->is_Phi()) { 134 return false; 135 } 136 const PhiNode* phi = inc->in(1)->as_Phi(); 137 138 if (!phi->region()->is_CountedLoop()) { 139 return false; 140 } 141 142 return inc == phi->region()->as_CountedLoop()->incr(); 143 } 144 145 // Given the expression '(x + C) - v', or 146 // 'v - (x + C)', we examine nodes '+' and 'v': 147 // 148 // 1. Do not convert if '+' is a counted-loop increment, because the '-' is 149 // loop invariant and converting extends the live-range of 'x' to overlap 150 // with the '+', forcing another register to be used in the loop. 151 // 152 // 2. Do not convert if 'v' is a counted-loop induction variable, because 153 // 'x' might be invariant. 154 // 155 static bool ok_to_convert(Node* inc, Node* var) { 156 return !(is_cloop_increment(inc) || var->is_cloop_ind_var()); 157 } 158 159 static bool is_cloop_condition(BoolNode* bol) { 160 for (DUIterator_Fast imax, i = bol->fast_outs(imax); i < imax; i++) { 161 Node* out = bol->fast_out(i); 162 if (out->is_BaseCountedLoopEnd()) { 163 return true; 164 } 165 } 166 return false; 167 } 168 169 //------------------------------Ideal------------------------------------------ 170 Node *SubINode::Ideal(PhaseGVN *phase, bool can_reshape){ 171 Node *in1 = in(1); 172 Node *in2 = in(2); 173 uint op1 = in1->Opcode(); 174 uint op2 = in2->Opcode(); 175 176 #ifdef ASSERT 177 // Check for dead loop 178 if ((in1 == this) || (in2 == this) || 179 ((op1 == Op_AddI || op1 == Op_SubI) && 180 ((in1->in(1) == this) || (in1->in(2) == this) || 181 (in1->in(1) == in1) || (in1->in(2) == in1)))) { 182 assert(false, "dead loop in SubINode::Ideal"); 183 } 184 #endif 185 186 const Type *t2 = phase->type( in2 ); 187 if( t2 == Type::TOP ) return nullptr; 188 // Convert "x-c0" into "x+ -c0". 189 if( t2->base() == Type::Int ){ // Might be bottom or top... 190 const TypeInt *i = t2->is_int(); 191 if( i->is_con() ) 192 return new AddINode(in1, phase->intcon(java_negate(i->get_con()))); 193 } 194 195 // Convert "(x+c0) - y" into (x-y) + c0" 196 // Do not collapse (x+c0)-y if "+" is a loop increment or 197 // if "y" is a loop induction variable. 198 if( op1 == Op_AddI && ok_to_convert(in1, in2) ) { 199 const Type *tadd = phase->type( in1->in(2) ); 200 if( tadd->singleton() && tadd != Type::TOP ) { 201 Node *sub2 = phase->transform( new SubINode( in1->in(1), in2 )); 202 return new AddINode( sub2, in1->in(2) ); 203 } 204 } 205 206 // Convert "x - (y+c0)" into "(x-y) - c0" AND 207 // Convert "c1 - (y+c0)" into "(c1-c0) - y" 208 // Need the same check as in above optimization but reversed. 209 if (op2 == Op_AddI 210 && ok_to_convert(in2, in1) 211 && in2->in(2)->Opcode() == Op_ConI) { 212 jint c0 = phase->type(in2->in(2))->isa_int()->get_con(); 213 Node* in21 = in2->in(1); 214 if (in1->Opcode() == Op_ConI) { 215 // Match c1 216 jint c1 = phase->type(in1)->isa_int()->get_con(); 217 Node* sub2 = phase->intcon(java_subtract(c1, c0)); 218 return new SubINode(sub2, in21); 219 } else { 220 // Match x 221 Node* sub2 = phase->transform(new SubINode(in1, in21)); 222 Node* neg_c0 = phase->intcon(java_negate(c0)); 223 return new AddINode(sub2, neg_c0); 224 } 225 } 226 227 const Type *t1 = phase->type( in1 ); 228 if( t1 == Type::TOP ) return nullptr; 229 230 #ifdef ASSERT 231 // Check for dead loop 232 if ((op2 == Op_AddI || op2 == Op_SubI) && 233 ((in2->in(1) == this) || (in2->in(2) == this) || 234 (in2->in(1) == in2) || (in2->in(2) == in2))) { 235 assert(false, "dead loop in SubINode::Ideal"); 236 } 237 #endif 238 239 // Convert "x - (x+y)" into "-y" 240 if (op2 == Op_AddI && in1 == in2->in(1)) { 241 return new SubINode(phase->intcon(0), in2->in(2)); 242 } 243 // Convert "(x-y) - x" into "-y" 244 if (op1 == Op_SubI && in1->in(1) == in2) { 245 return new SubINode(phase->intcon(0), in1->in(2)); 246 } 247 // Convert "x - (y+x)" into "-y" 248 if (op2 == Op_AddI && in1 == in2->in(2)) { 249 return new SubINode(phase->intcon(0), in2->in(1)); 250 } 251 252 // Convert "0 - (x-y)" into "y-x", leave the double negation "-(-y)" to SubNode::Identity(). 253 if (t1 == TypeInt::ZERO && op2 == Op_SubI && phase->type(in2->in(1)) != TypeInt::ZERO) { 254 return new SubINode(in2->in(2), in2->in(1)); 255 } 256 257 // Convert "0 - (x+con)" into "-con-x" 258 jint con; 259 if( t1 == TypeInt::ZERO && op2 == Op_AddI && 260 (con = in2->in(2)->find_int_con(0)) != 0 ) 261 return new SubINode( phase->intcon(-con), in2->in(1) ); 262 263 // Convert "(X+A) - (X+B)" into "A - B" 264 if( op1 == Op_AddI && op2 == Op_AddI && in1->in(1) == in2->in(1) ) 265 return new SubINode( in1->in(2), in2->in(2) ); 266 267 // Convert "(A+X) - (B+X)" into "A - B" 268 if( op1 == Op_AddI && op2 == Op_AddI && in1->in(2) == in2->in(2) ) 269 return new SubINode( in1->in(1), in2->in(1) ); 270 271 // Convert "(A+X) - (X+B)" into "A - B" 272 if( op1 == Op_AddI && op2 == Op_AddI && in1->in(2) == in2->in(1) ) 273 return new SubINode( in1->in(1), in2->in(2) ); 274 275 // Convert "(X+A) - (B+X)" into "A - B" 276 if( op1 == Op_AddI && op2 == Op_AddI && in1->in(1) == in2->in(2) ) 277 return new SubINode( in1->in(2), in2->in(1) ); 278 279 // Convert "A-(B-C)" into (A+C)-B", since add is commutative and generally 280 // nicer to optimize than subtract. 281 if( op2 == Op_SubI && in2->outcnt() == 1) { 282 Node *add1 = phase->transform( new AddINode( in1, in2->in(2) ) ); 283 return new SubINode( add1, in2->in(1) ); 284 } 285 286 // Associative 287 if (op1 == Op_MulI && op2 == Op_MulI) { 288 Node* sub_in1 = nullptr; 289 Node* sub_in2 = nullptr; 290 Node* mul_in = nullptr; 291 292 if (in1->in(1) == in2->in(1)) { 293 // Convert "a*b-a*c into a*(b-c) 294 sub_in1 = in1->in(2); 295 sub_in2 = in2->in(2); 296 mul_in = in1->in(1); 297 } else if (in1->in(2) == in2->in(1)) { 298 // Convert a*b-b*c into b*(a-c) 299 sub_in1 = in1->in(1); 300 sub_in2 = in2->in(2); 301 mul_in = in1->in(2); 302 } else if (in1->in(2) == in2->in(2)) { 303 // Convert a*c-b*c into (a-b)*c 304 sub_in1 = in1->in(1); 305 sub_in2 = in2->in(1); 306 mul_in = in1->in(2); 307 } else if (in1->in(1) == in2->in(2)) { 308 // Convert a*b-c*a into a*(b-c) 309 sub_in1 = in1->in(2); 310 sub_in2 = in2->in(1); 311 mul_in = in1->in(1); 312 } 313 314 if (mul_in != nullptr) { 315 Node* sub = phase->transform(new SubINode(sub_in1, sub_in2)); 316 return new MulINode(mul_in, sub); 317 } 318 } 319 320 // Convert "0-(A>>31)" into "(A>>>31)" 321 if ( op2 == Op_RShiftI ) { 322 Node *in21 = in2->in(1); 323 Node *in22 = in2->in(2); 324 const TypeInt *zero = phase->type(in1)->isa_int(); 325 const TypeInt *t21 = phase->type(in21)->isa_int(); 326 const TypeInt *t22 = phase->type(in22)->isa_int(); 327 if ( t21 && t22 && zero == TypeInt::ZERO && t22->is_con(31) ) { 328 return new URShiftINode(in21, in22); 329 } 330 } 331 332 return nullptr; 333 } 334 335 //------------------------------sub-------------------------------------------- 336 // A subtract node differences it's two inputs. 337 const Type *SubINode::sub( const Type *t1, const Type *t2 ) const { 338 const TypeInt *r0 = t1->is_int(); // Handy access 339 const TypeInt *r1 = t2->is_int(); 340 int32_t lo = java_subtract(r0->_lo, r1->_hi); 341 int32_t hi = java_subtract(r0->_hi, r1->_lo); 342 343 // We next check for 32-bit overflow. 344 // If that happens, we just assume all integers are possible. 345 if( (((r0->_lo ^ r1->_hi) >= 0) || // lo ends have same signs OR 346 ((r0->_lo ^ lo) >= 0)) && // lo results have same signs AND 347 (((r0->_hi ^ r1->_lo) >= 0) || // hi ends have same signs OR 348 ((r0->_hi ^ hi) >= 0)) ) // hi results have same signs 349 return TypeInt::make(lo,hi,MAX2(r0->_widen,r1->_widen)); 350 else // Overflow; assume all integers 351 return TypeInt::INT; 352 } 353 354 //============================================================================= 355 //------------------------------Ideal------------------------------------------ 356 Node *SubLNode::Ideal(PhaseGVN *phase, bool can_reshape) { 357 Node *in1 = in(1); 358 Node *in2 = in(2); 359 uint op1 = in1->Opcode(); 360 uint op2 = in2->Opcode(); 361 362 #ifdef ASSERT 363 // Check for dead loop 364 if ((in1 == this) || (in2 == this) || 365 ((op1 == Op_AddL || op1 == Op_SubL) && 366 ((in1->in(1) == this) || (in1->in(2) == this) || 367 (in1->in(1) == in1) || (in1->in(2) == in1)))) { 368 assert(false, "dead loop in SubLNode::Ideal"); 369 } 370 #endif 371 372 if( phase->type( in2 ) == Type::TOP ) return nullptr; 373 const TypeLong *i = phase->type( in2 )->isa_long(); 374 // Convert "x-c0" into "x+ -c0". 375 if( i && // Might be bottom or top... 376 i->is_con() ) 377 return new AddLNode(in1, phase->longcon(java_negate(i->get_con()))); 378 379 // Convert "(x+c0) - y" into (x-y) + c0" 380 // Do not collapse (x+c0)-y if "+" is a loop increment or 381 // if "y" is a loop induction variable. 382 if( op1 == Op_AddL && ok_to_convert(in1, in2) ) { 383 Node *in11 = in1->in(1); 384 const Type *tadd = phase->type( in1->in(2) ); 385 if( tadd->singleton() && tadd != Type::TOP ) { 386 Node *sub2 = phase->transform( new SubLNode( in11, in2 )); 387 return new AddLNode( sub2, in1->in(2) ); 388 } 389 } 390 391 // Convert "x - (y+c0)" into "(x-y) - c0" AND 392 // Convert "c1 - (y+c0)" into "(c1-c0) - y" 393 // Need the same check as in above optimization but reversed. 394 if (op2 == Op_AddL 395 && ok_to_convert(in2, in1) 396 && in2->in(2)->Opcode() == Op_ConL) { 397 jlong c0 = phase->type(in2->in(2))->isa_long()->get_con(); 398 Node* in21 = in2->in(1); 399 if (in1->Opcode() == Op_ConL) { 400 // Match c1 401 jlong c1 = phase->type(in1)->isa_long()->get_con(); 402 Node* sub2 = phase->longcon(java_subtract(c1, c0)); 403 return new SubLNode(sub2, in21); 404 } else { 405 Node* sub2 = phase->transform(new SubLNode(in1, in21)); 406 Node* neg_c0 = phase->longcon(-c0); 407 return new AddLNode(sub2, neg_c0); 408 } 409 } 410 411 const Type *t1 = phase->type( in1 ); 412 if( t1 == Type::TOP ) return nullptr; 413 414 #ifdef ASSERT 415 // Check for dead loop 416 if ((op2 == Op_AddL || op2 == Op_SubL) && 417 ((in2->in(1) == this) || (in2->in(2) == this) || 418 (in2->in(1) == in2) || (in2->in(2) == in2))) { 419 assert(false, "dead loop in SubLNode::Ideal"); 420 } 421 #endif 422 423 // Convert "x - (x+y)" into "-y" 424 if (op2 == Op_AddL && in1 == in2->in(1)) { 425 return new SubLNode(phase->makecon(TypeLong::ZERO), in2->in(2)); 426 } 427 // Convert "(x-y) - x" into "-y" 428 if (op1 == Op_SubL && in1->in(1) == in2) { 429 return new SubLNode(phase->makecon(TypeLong::ZERO), in1->in(2)); 430 } 431 // Convert "x - (y+x)" into "-y" 432 if (op2 == Op_AddL && in1 == in2->in(2)) { 433 return new SubLNode(phase->makecon(TypeLong::ZERO), in2->in(1)); 434 } 435 436 // Convert "0 - (x-y)" into "y-x", leave the double negation "-(-y)" to SubNode::Identity. 437 if (t1 == TypeLong::ZERO && op2 == Op_SubL && phase->type(in2->in(1)) != TypeLong::ZERO) { 438 return new SubLNode(in2->in(2), in2->in(1)); 439 } 440 441 // Convert "(X+A) - (X+B)" into "A - B" 442 if( op1 == Op_AddL && op2 == Op_AddL && in1->in(1) == in2->in(1) ) 443 return new SubLNode( in1->in(2), in2->in(2) ); 444 445 // Convert "(A+X) - (B+X)" into "A - B" 446 if( op1 == Op_AddL && op2 == Op_AddL && in1->in(2) == in2->in(2) ) 447 return new SubLNode( in1->in(1), in2->in(1) ); 448 449 // Convert "(A+X) - (X+B)" into "A - B" 450 if( op1 == Op_AddL && op2 == Op_AddL && in1->in(2) == in2->in(1) ) 451 return new SubLNode( in1->in(1), in2->in(2) ); 452 453 // Convert "(X+A) - (B+X)" into "A - B" 454 if( op1 == Op_AddL && op2 == Op_AddL && in1->in(1) == in2->in(2) ) 455 return new SubLNode( in1->in(2), in2->in(1) ); 456 457 // Convert "A-(B-C)" into (A+C)-B" 458 if( op2 == Op_SubL && in2->outcnt() == 1) { 459 Node *add1 = phase->transform( new AddLNode( in1, in2->in(2) ) ); 460 return new SubLNode( add1, in2->in(1) ); 461 } 462 463 // Associative 464 if (op1 == Op_MulL && op2 == Op_MulL) { 465 Node* sub_in1 = nullptr; 466 Node* sub_in2 = nullptr; 467 Node* mul_in = nullptr; 468 469 if (in1->in(1) == in2->in(1)) { 470 // Convert "a*b-a*c into a*(b+c) 471 sub_in1 = in1->in(2); 472 sub_in2 = in2->in(2); 473 mul_in = in1->in(1); 474 } else if (in1->in(2) == in2->in(1)) { 475 // Convert a*b-b*c into b*(a-c) 476 sub_in1 = in1->in(1); 477 sub_in2 = in2->in(2); 478 mul_in = in1->in(2); 479 } else if (in1->in(2) == in2->in(2)) { 480 // Convert a*c-b*c into (a-b)*c 481 sub_in1 = in1->in(1); 482 sub_in2 = in2->in(1); 483 mul_in = in1->in(2); 484 } else if (in1->in(1) == in2->in(2)) { 485 // Convert a*b-c*a into a*(b-c) 486 sub_in1 = in1->in(2); 487 sub_in2 = in2->in(1); 488 mul_in = in1->in(1); 489 } 490 491 if (mul_in != nullptr) { 492 Node* sub = phase->transform(new SubLNode(sub_in1, sub_in2)); 493 return new MulLNode(mul_in, sub); 494 } 495 } 496 497 // Convert "0L-(A>>63)" into "(A>>>63)" 498 if ( op2 == Op_RShiftL ) { 499 Node *in21 = in2->in(1); 500 Node *in22 = in2->in(2); 501 const TypeLong *zero = phase->type(in1)->isa_long(); 502 const TypeLong *t21 = phase->type(in21)->isa_long(); 503 const TypeInt *t22 = phase->type(in22)->isa_int(); 504 if ( t21 && t22 && zero == TypeLong::ZERO && t22->is_con(63) ) { 505 return new URShiftLNode(in21, in22); 506 } 507 } 508 509 return nullptr; 510 } 511 512 //------------------------------sub-------------------------------------------- 513 // A subtract node differences it's two inputs. 514 const Type *SubLNode::sub( const Type *t1, const Type *t2 ) const { 515 const TypeLong *r0 = t1->is_long(); // Handy access 516 const TypeLong *r1 = t2->is_long(); 517 jlong lo = java_subtract(r0->_lo, r1->_hi); 518 jlong hi = java_subtract(r0->_hi, r1->_lo); 519 520 // We next check for 32-bit overflow. 521 // If that happens, we just assume all integers are possible. 522 if( (((r0->_lo ^ r1->_hi) >= 0) || // lo ends have same signs OR 523 ((r0->_lo ^ lo) >= 0)) && // lo results have same signs AND 524 (((r0->_hi ^ r1->_lo) >= 0) || // hi ends have same signs OR 525 ((r0->_hi ^ hi) >= 0)) ) // hi results have same signs 526 return TypeLong::make(lo,hi,MAX2(r0->_widen,r1->_widen)); 527 else // Overflow; assume all integers 528 return TypeLong::LONG; 529 } 530 531 //============================================================================= 532 //------------------------------Value------------------------------------------ 533 // A subtract node differences its two inputs. 534 const Type* SubFPNode::Value(PhaseGVN* phase) const { 535 const Node* in1 = in(1); 536 const Node* in2 = in(2); 537 // Either input is TOP ==> the result is TOP 538 const Type* t1 = (in1 == this) ? Type::TOP : phase->type(in1); 539 if( t1 == Type::TOP ) return Type::TOP; 540 const Type* t2 = (in2 == this) ? Type::TOP : phase->type(in2); 541 if( t2 == Type::TOP ) return Type::TOP; 542 543 // if both operands are infinity of same sign, the result is NaN; do 544 // not replace with zero 545 if (t1->is_finite() && t2->is_finite() && in1 == in2) { 546 return add_id(); 547 } 548 549 // Either input is BOTTOM ==> the result is the local BOTTOM 550 const Type *bot = bottom_type(); 551 if( (t1 == bot) || (t2 == bot) || 552 (t1 == Type::BOTTOM) || (t2 == Type::BOTTOM) ) 553 return bot; 554 555 return sub(t1,t2); // Local flavor of type subtraction 556 } 557 558 559 //============================================================================= 560 //------------------------------sub-------------------------------------------- 561 // A subtract node differences its two inputs. 562 const Type* SubHFNode::sub(const Type* t1, const Type* t2) const { 563 // no folding if one of operands is infinity or NaN, do not do constant folding 564 if(g_isfinite(t1->getf()) && g_isfinite(t2->getf())) { 565 return TypeH::make(t1->getf() - t2->getf()); 566 } 567 else if(g_isnan(t1->getf())) { 568 return t1; 569 } 570 else if(g_isnan(t2->getf())) { 571 return t2; 572 } 573 else { 574 return Type::HALF_FLOAT; 575 } 576 } 577 578 //------------------------------Ideal------------------------------------------ 579 Node *SubFNode::Ideal(PhaseGVN *phase, bool can_reshape) { 580 const Type *t2 = phase->type( in(2) ); 581 // Convert "x-c0" into "x+ -c0". 582 if( t2->base() == Type::FloatCon ) { // Might be bottom or top... 583 // return new (phase->C, 3) AddFNode(in(1), phase->makecon( TypeF::make(-t2->getf()) ) ); 584 } 585 586 // Cannot replace 0.0-X with -X because a 'fsub' bytecode computes 587 // 0.0-0.0 as +0.0, while a 'fneg' bytecode computes -0.0. 588 //if( phase->type(in(1)) == TypeF::ZERO ) 589 //return new (phase->C, 2) NegFNode(in(2)); 590 591 return nullptr; 592 } 593 594 //------------------------------sub-------------------------------------------- 595 // A subtract node differences its two inputs. 596 const Type *SubFNode::sub( const Type *t1, const Type *t2 ) const { 597 // no folding if one of operands is infinity or NaN, do not do constant folding 598 if( g_isfinite(t1->getf()) && g_isfinite(t2->getf()) ) { 599 return TypeF::make( t1->getf() - t2->getf() ); 600 } 601 else if( g_isnan(t1->getf()) ) { 602 return t1; 603 } 604 else if( g_isnan(t2->getf()) ) { 605 return t2; 606 } 607 else { 608 return Type::FLOAT; 609 } 610 } 611 612 //============================================================================= 613 //------------------------------Ideal------------------------------------------ 614 Node *SubDNode::Ideal(PhaseGVN *phase, bool can_reshape){ 615 const Type *t2 = phase->type( in(2) ); 616 // Convert "x-c0" into "x+ -c0". 617 if( t2->base() == Type::DoubleCon ) { // Might be bottom or top... 618 // return new (phase->C, 3) AddDNode(in(1), phase->makecon( TypeD::make(-t2->getd()) ) ); 619 } 620 621 // Cannot replace 0.0-X with -X because a 'dsub' bytecode computes 622 // 0.0-0.0 as +0.0, while a 'dneg' bytecode computes -0.0. 623 //if( phase->type(in(1)) == TypeD::ZERO ) 624 //return new (phase->C, 2) NegDNode(in(2)); 625 626 return nullptr; 627 } 628 629 //------------------------------sub-------------------------------------------- 630 // A subtract node differences its two inputs. 631 const Type *SubDNode::sub( const Type *t1, const Type *t2 ) const { 632 // no folding if one of operands is infinity or NaN, do not do constant folding 633 if( g_isfinite(t1->getd()) && g_isfinite(t2->getd()) ) { 634 return TypeD::make( t1->getd() - t2->getd() ); 635 } 636 else if( g_isnan(t1->getd()) ) { 637 return t1; 638 } 639 else if( g_isnan(t2->getd()) ) { 640 return t2; 641 } 642 else { 643 return Type::DOUBLE; 644 } 645 } 646 647 //============================================================================= 648 //------------------------------Idealize--------------------------------------- 649 // Unlike SubNodes, compare must still flatten return value to the 650 // range -1, 0, 1. 651 // And optimizations like those for (X + Y) - X fail if overflow happens. 652 Node* CmpNode::Identity(PhaseGVN* phase) { 653 return this; 654 } 655 656 CmpNode *CmpNode::make(Node *in1, Node *in2, BasicType bt, bool unsigned_comp) { 657 switch (bt) { 658 case T_INT: 659 if (unsigned_comp) { 660 return new CmpUNode(in1, in2); 661 } 662 return new CmpINode(in1, in2); 663 case T_LONG: 664 if (unsigned_comp) { 665 return new CmpULNode(in1, in2); 666 } 667 return new CmpLNode(in1, in2); 668 case T_OBJECT: 669 case T_ARRAY: 670 case T_ADDRESS: 671 case T_METADATA: 672 return new CmpPNode(in1, in2); 673 case T_NARROWOOP: 674 case T_NARROWKLASS: 675 return new CmpNNode(in1, in2); 676 default: 677 fatal("Not implemented for %s", type2name(bt)); 678 } 679 return nullptr; 680 } 681 682 //============================================================================= 683 //------------------------------cmp-------------------------------------------- 684 // Simplify a CmpI (compare 2 integers) node, based on local information. 685 // If both inputs are constants, compare them. 686 const Type *CmpINode::sub( const Type *t1, const Type *t2 ) const { 687 const TypeInt *r0 = t1->is_int(); // Handy access 688 const TypeInt *r1 = t2->is_int(); 689 690 if( r0->_hi < r1->_lo ) // Range is always low? 691 return TypeInt::CC_LT; 692 else if( r0->_lo > r1->_hi ) // Range is always high? 693 return TypeInt::CC_GT; 694 695 else if( r0->is_con() && r1->is_con() ) { // comparing constants? 696 assert(r0->get_con() == r1->get_con(), "must be equal"); 697 return TypeInt::CC_EQ; // Equal results. 698 } else if( r0->_hi == r1->_lo ) // Range is never high? 699 return TypeInt::CC_LE; 700 else if( r0->_lo == r1->_hi ) // Range is never low? 701 return TypeInt::CC_GE; 702 return TypeInt::CC; // else use worst case results 703 } 704 705 const Type* CmpINode::Value(PhaseGVN* phase) const { 706 Node* in1 = in(1); 707 Node* in2 = in(2); 708 // If this test is the zero trip guard for a main or post loop, check whether, with the opaque node removed, the test 709 // would constant fold so the loop is never entered. If so return the type of the test without the opaque node removed: 710 // make the loop unreachable. 711 // The reason for this is that the iv phi captures the bounds of the loop and if the loop becomes unreachable, it can 712 // become top. In that case, the loop must be removed. 713 // This is safe because: 714 // - as optimizations proceed, the range of iterations executed by the main loop narrows. If no iterations remain, then 715 // we're done with optimizations for that loop. 716 // - the post loop is initially not reachable but as long as there's a main loop, the zero trip guard for the post 717 // loop takes a phi that merges the pre and main loop's iv and can't constant fold the zero trip guard. Once, the main 718 // loop is removed, there's no need to preserve the zero trip guard for the post loop anymore. 719 if (in1 != nullptr && in2 != nullptr) { 720 uint input = 0; 721 Node* cmp = nullptr; 722 BoolTest::mask test; 723 if (in1->Opcode() == Op_OpaqueZeroTripGuard && phase->type(in1) != Type::TOP) { 724 cmp = new CmpINode(in1->in(1), in2); 725 test = ((OpaqueZeroTripGuardNode*)in1)->_loop_entered_mask; 726 } 727 if (in2->Opcode() == Op_OpaqueZeroTripGuard && phase->type(in2) != Type::TOP) { 728 assert(cmp == nullptr, "A cmp with 2 OpaqueZeroTripGuard inputs"); 729 cmp = new CmpINode(in1, in2->in(1)); 730 test = ((OpaqueZeroTripGuardNode*)in2)->_loop_entered_mask; 731 } 732 if (cmp != nullptr) { 733 const Type* cmp_t = cmp->Value(phase); 734 const Type* t = BoolTest(test).cc2logical(cmp_t); 735 cmp->destruct(phase); 736 if (t == TypeInt::ZERO) { 737 return cmp_t; 738 } 739 } 740 } 741 742 return SubNode::Value(phase); 743 } 744 745 746 // Simplify a CmpU (compare 2 integers) node, based on local information. 747 // If both inputs are constants, compare them. 748 const Type *CmpUNode::sub( const Type *t1, const Type *t2 ) const { 749 assert(!t1->isa_ptr(), "obsolete usage of CmpU"); 750 751 // comparing two unsigned ints 752 const TypeInt *r0 = t1->is_int(); // Handy access 753 const TypeInt *r1 = t2->is_int(); 754 755 // Current installed version 756 // Compare ranges for non-overlap 757 juint lo0 = r0->_lo; 758 juint hi0 = r0->_hi; 759 juint lo1 = r1->_lo; 760 juint hi1 = r1->_hi; 761 762 // If either one has both negative and positive values, 763 // it therefore contains both 0 and -1, and since [0..-1] is the 764 // full unsigned range, the type must act as an unsigned bottom. 765 bool bot0 = ((jint)(lo0 ^ hi0) < 0); 766 bool bot1 = ((jint)(lo1 ^ hi1) < 0); 767 768 if (bot0 || bot1) { 769 // All unsigned values are LE -1 and GE 0. 770 if (lo0 == 0 && hi0 == 0) { 771 return TypeInt::CC_LE; // 0 <= bot 772 } else if ((jint)lo0 == -1 && (jint)hi0 == -1) { 773 return TypeInt::CC_GE; // -1 >= bot 774 } else if (lo1 == 0 && hi1 == 0) { 775 return TypeInt::CC_GE; // bot >= 0 776 } else if ((jint)lo1 == -1 && (jint)hi1 == -1) { 777 return TypeInt::CC_LE; // bot <= -1 778 } 779 } else { 780 // We can use ranges of the form [lo..hi] if signs are the same. 781 assert(lo0 <= hi0 && lo1 <= hi1, "unsigned ranges are valid"); 782 // results are reversed, '-' > '+' for unsigned compare 783 if (hi0 < lo1) { 784 return TypeInt::CC_LT; // smaller 785 } else if (lo0 > hi1) { 786 return TypeInt::CC_GT; // greater 787 } else if (hi0 == lo1 && lo0 == hi1) { 788 return TypeInt::CC_EQ; // Equal results 789 } else if (lo0 >= hi1) { 790 return TypeInt::CC_GE; 791 } else if (hi0 <= lo1) { 792 // Check for special case in Hashtable::get. (See below.) 793 if ((jint)lo0 >= 0 && (jint)lo1 >= 0 && is_index_range_check()) 794 return TypeInt::CC_LT; 795 return TypeInt::CC_LE; 796 } 797 } 798 // Check for special case in Hashtable::get - the hash index is 799 // mod'ed to the table size so the following range check is useless. 800 // Check for: (X Mod Y) CmpU Y, where the mod result and Y both have 801 // to be positive. 802 // (This is a gross hack, since the sub method never 803 // looks at the structure of the node in any other case.) 804 if ((jint)lo0 >= 0 && (jint)lo1 >= 0 && is_index_range_check()) 805 return TypeInt::CC_LT; 806 return TypeInt::CC; // else use worst case results 807 } 808 809 const Type* CmpUNode::Value(PhaseGVN* phase) const { 810 const Type* t = SubNode::Value_common(phase); 811 if (t != nullptr) { 812 return t; 813 } 814 const Node* in1 = in(1); 815 const Node* in2 = in(2); 816 const Type* t1 = phase->type(in1); 817 const Type* t2 = phase->type(in2); 818 assert(t1->isa_int(), "CmpU has only Int type inputs"); 819 if (t2 == TypeInt::INT) { // Compare to bottom? 820 return bottom_type(); 821 } 822 823 const Type* t_sub = sub(t1, t2); // compare based on immediate inputs 824 825 uint in1_op = in1->Opcode(); 826 if (in1_op == Op_AddI || in1_op == Op_SubI) { 827 // The problem rise when result of AddI(SubI) may overflow 828 // signed integer value. Let say the input type is 829 // [256, maxint] then +128 will create 2 ranges due to 830 // overflow: [minint, minint+127] and [384, maxint]. 831 // But C2 type system keep only 1 type range and as result 832 // it use general [minint, maxint] for this case which we 833 // can't optimize. 834 // 835 // Make 2 separate type ranges based on types of AddI(SubI) inputs 836 // and compare results of their compare. If results are the same 837 // CmpU node can be optimized. 838 const Node* in11 = in1->in(1); 839 const Node* in12 = in1->in(2); 840 const Type* t11 = (in11 == in1) ? Type::TOP : phase->type(in11); 841 const Type* t12 = (in12 == in1) ? Type::TOP : phase->type(in12); 842 // Skip cases when input types are top or bottom. 843 if ((t11 != Type::TOP) && (t11 != TypeInt::INT) && 844 (t12 != Type::TOP) && (t12 != TypeInt::INT)) { 845 const TypeInt *r0 = t11->is_int(); 846 const TypeInt *r1 = t12->is_int(); 847 jlong lo_r0 = r0->_lo; 848 jlong hi_r0 = r0->_hi; 849 jlong lo_r1 = r1->_lo; 850 jlong hi_r1 = r1->_hi; 851 if (in1_op == Op_SubI) { 852 jlong tmp = hi_r1; 853 hi_r1 = -lo_r1; 854 lo_r1 = -tmp; 855 // Note, for substructing [minint,x] type range 856 // long arithmetic provides correct overflow answer. 857 // The confusion come from the fact that in 32-bit 858 // -minint == minint but in 64-bit -minint == maxint+1. 859 } 860 jlong lo_long = lo_r0 + lo_r1; 861 jlong hi_long = hi_r0 + hi_r1; 862 int lo_tr1 = min_jint; 863 int hi_tr1 = (int)hi_long; 864 int lo_tr2 = (int)lo_long; 865 int hi_tr2 = max_jint; 866 bool underflow = lo_long != (jlong)lo_tr2; 867 bool overflow = hi_long != (jlong)hi_tr1; 868 // Use sub(t1, t2) when there is no overflow (one type range) 869 // or when both overflow and underflow (too complex). 870 if ((underflow != overflow) && (hi_tr1 < lo_tr2)) { 871 // Overflow only on one boundary, compare 2 separate type ranges. 872 int w = MAX2(r0->_widen, r1->_widen); // _widen does not matter here 873 const TypeInt* tr1 = TypeInt::make(lo_tr1, hi_tr1, w); 874 const TypeInt* tr2 = TypeInt::make(lo_tr2, hi_tr2, w); 875 const TypeInt* cmp1 = sub(tr1, t2)->is_int(); 876 const TypeInt* cmp2 = sub(tr2, t2)->is_int(); 877 // Compute union, so that cmp handles all possible results from the two cases 878 const Type* t_cmp = cmp1->meet(cmp2); 879 // Pick narrowest type, based on overflow computation and on immediate inputs 880 return t_sub->filter(t_cmp); 881 } 882 } 883 } 884 885 return t_sub; 886 } 887 888 bool CmpUNode::is_index_range_check() const { 889 // Check for the "(X ModI Y) CmpU Y" shape 890 return (in(1)->Opcode() == Op_ModI && 891 in(1)->in(2)->eqv_uncast(in(2))); 892 } 893 894 //------------------------------Idealize--------------------------------------- 895 Node *CmpINode::Ideal( PhaseGVN *phase, bool can_reshape ) { 896 if (phase->type(in(2))->higher_equal(TypeInt::ZERO)) { 897 switch (in(1)->Opcode()) { 898 case Op_CmpU3: // Collapse a CmpU3/CmpI into a CmpU 899 return new CmpUNode(in(1)->in(1),in(1)->in(2)); 900 case Op_CmpL3: // Collapse a CmpL3/CmpI into a CmpL 901 return new CmpLNode(in(1)->in(1),in(1)->in(2)); 902 case Op_CmpUL3: // Collapse a CmpUL3/CmpI into a CmpUL 903 return new CmpULNode(in(1)->in(1),in(1)->in(2)); 904 case Op_CmpF3: // Collapse a CmpF3/CmpI into a CmpF 905 return new CmpFNode(in(1)->in(1),in(1)->in(2)); 906 case Op_CmpD3: // Collapse a CmpD3/CmpI into a CmpD 907 return new CmpDNode(in(1)->in(1),in(1)->in(2)); 908 //case Op_SubI: 909 // If (x - y) cannot overflow, then ((x - y) <?> 0) 910 // can be turned into (x <?> y). 911 // This is handled (with more general cases) by Ideal_sub_algebra. 912 } 913 } 914 return nullptr; // No change 915 } 916 917 Node *CmpLNode::Ideal( PhaseGVN *phase, bool can_reshape ) { 918 const TypeLong *t2 = phase->type(in(2))->isa_long(); 919 if (Opcode() == Op_CmpL && in(1)->Opcode() == Op_ConvI2L && t2 && t2->is_con()) { 920 const jlong con = t2->get_con(); 921 if (con >= min_jint && con <= max_jint) { 922 return new CmpINode(in(1)->in(1), phase->intcon((jint)con)); 923 } 924 } 925 return nullptr; 926 } 927 928 //============================================================================= 929 // Simplify a CmpL (compare 2 longs ) node, based on local information. 930 // If both inputs are constants, compare them. 931 const Type *CmpLNode::sub( const Type *t1, const Type *t2 ) const { 932 const TypeLong *r0 = t1->is_long(); // Handy access 933 const TypeLong *r1 = t2->is_long(); 934 935 if( r0->_hi < r1->_lo ) // Range is always low? 936 return TypeInt::CC_LT; 937 else if( r0->_lo > r1->_hi ) // Range is always high? 938 return TypeInt::CC_GT; 939 940 else if( r0->is_con() && r1->is_con() ) { // comparing constants? 941 assert(r0->get_con() == r1->get_con(), "must be equal"); 942 return TypeInt::CC_EQ; // Equal results. 943 } else if( r0->_hi == r1->_lo ) // Range is never high? 944 return TypeInt::CC_LE; 945 else if( r0->_lo == r1->_hi ) // Range is never low? 946 return TypeInt::CC_GE; 947 return TypeInt::CC; // else use worst case results 948 } 949 950 951 // Simplify a CmpUL (compare 2 unsigned longs) node, based on local information. 952 // If both inputs are constants, compare them. 953 const Type* CmpULNode::sub(const Type* t1, const Type* t2) const { 954 assert(!t1->isa_ptr(), "obsolete usage of CmpUL"); 955 956 // comparing two unsigned longs 957 const TypeLong* r0 = t1->is_long(); // Handy access 958 const TypeLong* r1 = t2->is_long(); 959 960 // Current installed version 961 // Compare ranges for non-overlap 962 julong lo0 = r0->_lo; 963 julong hi0 = r0->_hi; 964 julong lo1 = r1->_lo; 965 julong hi1 = r1->_hi; 966 967 // If either one has both negative and positive values, 968 // it therefore contains both 0 and -1, and since [0..-1] is the 969 // full unsigned range, the type must act as an unsigned bottom. 970 bool bot0 = ((jlong)(lo0 ^ hi0) < 0); 971 bool bot1 = ((jlong)(lo1 ^ hi1) < 0); 972 973 if (bot0 || bot1) { 974 // All unsigned values are LE -1 and GE 0. 975 if (lo0 == 0 && hi0 == 0) { 976 return TypeInt::CC_LE; // 0 <= bot 977 } else if ((jlong)lo0 == -1 && (jlong)hi0 == -1) { 978 return TypeInt::CC_GE; // -1 >= bot 979 } else if (lo1 == 0 && hi1 == 0) { 980 return TypeInt::CC_GE; // bot >= 0 981 } else if ((jlong)lo1 == -1 && (jlong)hi1 == -1) { 982 return TypeInt::CC_LE; // bot <= -1 983 } 984 } else { 985 // We can use ranges of the form [lo..hi] if signs are the same. 986 assert(lo0 <= hi0 && lo1 <= hi1, "unsigned ranges are valid"); 987 // results are reversed, '-' > '+' for unsigned compare 988 if (hi0 < lo1) { 989 return TypeInt::CC_LT; // smaller 990 } else if (lo0 > hi1) { 991 return TypeInt::CC_GT; // greater 992 } else if (hi0 == lo1 && lo0 == hi1) { 993 return TypeInt::CC_EQ; // Equal results 994 } else if (lo0 >= hi1) { 995 return TypeInt::CC_GE; 996 } else if (hi0 <= lo1) { 997 return TypeInt::CC_LE; 998 } 999 } 1000 1001 return TypeInt::CC; // else use worst case results 1002 } 1003 1004 //============================================================================= 1005 //------------------------------sub-------------------------------------------- 1006 // Simplify an CmpP (compare 2 pointers) node, based on local information. 1007 // If both inputs are constants, compare them. 1008 const Type *CmpPNode::sub( const Type *t1, const Type *t2 ) const { 1009 const TypePtr *r0 = t1->is_ptr(); // Handy access 1010 const TypePtr *r1 = t2->is_ptr(); 1011 1012 // Undefined inputs makes for an undefined result 1013 if( TypePtr::above_centerline(r0->_ptr) || 1014 TypePtr::above_centerline(r1->_ptr) ) 1015 return Type::TOP; 1016 1017 if (r0 == r1 && r0->singleton()) { 1018 // Equal pointer constants (klasses, nulls, etc.) 1019 return TypeInt::CC_EQ; 1020 } 1021 1022 // See if it is 2 unrelated classes. 1023 const TypeOopPtr* p0 = r0->isa_oopptr(); 1024 const TypeOopPtr* p1 = r1->isa_oopptr(); 1025 const TypeKlassPtr* k0 = r0->isa_klassptr(); 1026 const TypeKlassPtr* k1 = r1->isa_klassptr(); 1027 if ((p0 && p1) || (k0 && k1)) { 1028 if (p0 && p1) { 1029 Node* in1 = in(1)->uncast(); 1030 Node* in2 = in(2)->uncast(); 1031 AllocateNode* alloc1 = AllocateNode::Ideal_allocation(in1); 1032 AllocateNode* alloc2 = AllocateNode::Ideal_allocation(in2); 1033 if (MemNode::detect_ptr_independence(in1, alloc1, in2, alloc2, nullptr)) { 1034 return TypeInt::CC_GT; // different pointers 1035 } 1036 } 1037 bool xklass0 = p0 ? p0->klass_is_exact() : k0->klass_is_exact(); 1038 bool xklass1 = p1 ? p1->klass_is_exact() : k1->klass_is_exact(); 1039 bool unrelated_classes = false; 1040 1041 if ((p0 && p0->is_same_java_type_as(p1)) || 1042 (k0 && k0->is_same_java_type_as(k1))) { 1043 } else if ((p0 && !p1->maybe_java_subtype_of(p0) && !p0->maybe_java_subtype_of(p1)) || 1044 (k0 && !k1->maybe_java_subtype_of(k0) && !k0->maybe_java_subtype_of(k1))) { 1045 unrelated_classes = true; 1046 } else if ((p0 && !p1->maybe_java_subtype_of(p0)) || 1047 (k0 && !k1->maybe_java_subtype_of(k0))) { 1048 unrelated_classes = xklass1; 1049 } else if ((p0 && !p0->maybe_java_subtype_of(p1)) || 1050 (k0 && !k0->maybe_java_subtype_of(k1))) { 1051 unrelated_classes = xklass0; 1052 } 1053 1054 if (unrelated_classes) { 1055 // The oops classes are known to be unrelated. If the joined PTRs of 1056 // two oops is not Null and not Bottom, then we are sure that one 1057 // of the two oops is non-null, and the comparison will always fail. 1058 TypePtr::PTR jp = r0->join_ptr(r1->_ptr); 1059 if (jp != TypePtr::Null && jp != TypePtr::BotPTR) { 1060 return TypeInt::CC_GT; 1061 } 1062 } 1063 } 1064 1065 // Known constants can be compared exactly 1066 // Null can be distinguished from any NotNull pointers 1067 // Unknown inputs makes an unknown result 1068 if( r0->singleton() ) { 1069 intptr_t bits0 = r0->get_con(); 1070 if( r1->singleton() ) 1071 return bits0 == r1->get_con() ? TypeInt::CC_EQ : TypeInt::CC_GT; 1072 return ( r1->_ptr == TypePtr::NotNull && bits0==0 ) ? TypeInt::CC_GT : TypeInt::CC; 1073 } else if( r1->singleton() ) { 1074 intptr_t bits1 = r1->get_con(); 1075 return ( r0->_ptr == TypePtr::NotNull && bits1==0 ) ? TypeInt::CC_GT : TypeInt::CC; 1076 } else 1077 return TypeInt::CC; 1078 } 1079 1080 static inline Node* isa_java_mirror_load(PhaseGVN* phase, Node* n) { 1081 // Return the klass node for (indirect load from OopHandle) 1082 // LoadBarrier?(LoadP(LoadP(AddP(foo:Klass, #java_mirror)))) 1083 // or null if not matching. 1084 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 1085 n = bs->step_over_gc_barrier(n); 1086 1087 if (n->Opcode() != Op_LoadP) return nullptr; 1088 1089 const TypeInstPtr* tp = phase->type(n)->isa_instptr(); 1090 if (!tp || tp->instance_klass() != phase->C->env()->Class_klass()) return nullptr; 1091 1092 Node* adr = n->in(MemNode::Address); 1093 // First load from OopHandle: ((OopHandle)mirror)->resolve(); may need barrier. 1094 if (adr->Opcode() != Op_LoadP || !phase->type(adr)->isa_rawptr()) return nullptr; 1095 adr = adr->in(MemNode::Address); 1096 1097 intptr_t off = 0; 1098 Node* k = AddPNode::Ideal_base_and_offset(adr, phase, off); 1099 if (k == nullptr) return nullptr; 1100 const TypeKlassPtr* tkp = phase->type(k)->isa_klassptr(); 1101 if (!tkp || off != in_bytes(Klass::java_mirror_offset())) return nullptr; 1102 1103 // We've found the klass node of a Java mirror load. 1104 return k; 1105 } 1106 1107 static inline Node* isa_const_java_mirror(PhaseGVN* phase, Node* n) { 1108 // for ConP(Foo.class) return ConP(Foo.klass) 1109 // otherwise return null 1110 if (!n->is_Con()) return nullptr; 1111 1112 const TypeInstPtr* tp = phase->type(n)->isa_instptr(); 1113 if (!tp) return nullptr; 1114 1115 ciType* mirror_type = tp->java_mirror_type(); 1116 // TypeInstPtr::java_mirror_type() returns non-null for compile- 1117 // time Class constants only. 1118 if (!mirror_type) return nullptr; 1119 1120 // x.getClass() == int.class can never be true (for all primitive types) 1121 // Return a ConP(null) node for this case. 1122 if (mirror_type->is_classless()) { 1123 return phase->makecon(TypePtr::NULL_PTR); 1124 } 1125 1126 // return the ConP(Foo.klass) 1127 assert(mirror_type->is_klass(), "mirror_type should represent a Klass*"); 1128 return phase->makecon(TypeKlassPtr::make(mirror_type->as_klass(), Type::trust_interfaces)); 1129 } 1130 1131 //------------------------------Ideal------------------------------------------ 1132 // Normalize comparisons between Java mirror loads to compare the klass instead. 1133 // 1134 // Also check for the case of comparing an unknown klass loaded from the primary 1135 // super-type array vs a known klass with no subtypes. This amounts to 1136 // checking to see an unknown klass subtypes a known klass with no subtypes; 1137 // this only happens on an exact match. We can shorten this test by 1 load. 1138 Node *CmpPNode::Ideal( PhaseGVN *phase, bool can_reshape ) { 1139 // Normalize comparisons between Java mirrors into comparisons of the low- 1140 // level klass, where a dependent load could be shortened. 1141 // 1142 // The new pattern has a nice effect of matching the same pattern used in the 1143 // fast path of instanceof/checkcast/Class.isInstance(), which allows 1144 // redundant exact type check be optimized away by GVN. 1145 // For example, in 1146 // if (x.getClass() == Foo.class) { 1147 // Foo foo = (Foo) x; 1148 // // ... use a ... 1149 // } 1150 // a CmpPNode could be shared between if_acmpne and checkcast 1151 { 1152 Node* k1 = isa_java_mirror_load(phase, in(1)); 1153 Node* k2 = isa_java_mirror_load(phase, in(2)); 1154 Node* conk2 = isa_const_java_mirror(phase, in(2)); 1155 1156 if (k1 && (k2 || conk2)) { 1157 Node* lhs = k1; 1158 Node* rhs = (k2 != nullptr) ? k2 : conk2; 1159 set_req_X(1, lhs, phase); 1160 set_req_X(2, rhs, phase); 1161 return this; 1162 } 1163 } 1164 1165 // Constant pointer on right? 1166 const TypeKlassPtr* t2 = phase->type(in(2))->isa_klassptr(); 1167 if (t2 == nullptr || !t2->klass_is_exact()) 1168 return nullptr; 1169 // Get the constant klass we are comparing to. 1170 ciKlass* superklass = t2->exact_klass(); 1171 1172 // Now check for LoadKlass on left. 1173 Node* ldk1 = in(1); 1174 if (ldk1->is_DecodeNKlass()) { 1175 ldk1 = ldk1->in(1); 1176 if (ldk1->Opcode() != Op_LoadNKlass ) 1177 return nullptr; 1178 } else if (ldk1->Opcode() != Op_LoadKlass ) 1179 return nullptr; 1180 // Take apart the address of the LoadKlass: 1181 Node* adr1 = ldk1->in(MemNode::Address); 1182 intptr_t con2 = 0; 1183 Node* ldk2 = AddPNode::Ideal_base_and_offset(adr1, phase, con2); 1184 if (ldk2 == nullptr) 1185 return nullptr; 1186 if (con2 == oopDesc::klass_offset_in_bytes()) { 1187 // We are inspecting an object's concrete class. 1188 // Short-circuit the check if the query is abstract. 1189 if (superklass->is_interface() || 1190 superklass->is_abstract()) { 1191 // Make it come out always false: 1192 this->set_req(2, phase->makecon(TypePtr::NULL_PTR)); 1193 return this; 1194 } 1195 } 1196 1197 // Check for a LoadKlass from primary supertype array. 1198 // Any nested loadklass from loadklass+con must be from the p.s. array. 1199 if (ldk2->is_DecodeNKlass()) { 1200 // Keep ldk2 as DecodeN since it could be used in CmpP below. 1201 if (ldk2->in(1)->Opcode() != Op_LoadNKlass ) 1202 return nullptr; 1203 } else if (ldk2->Opcode() != Op_LoadKlass) 1204 return nullptr; 1205 1206 // Verify that we understand the situation 1207 if (con2 != (intptr_t) superklass->super_check_offset()) 1208 return nullptr; // Might be element-klass loading from array klass 1209 1210 // If 'superklass' has no subklasses and is not an interface, then we are 1211 // assured that the only input which will pass the type check is 1212 // 'superklass' itself. 1213 // 1214 // We could be more liberal here, and allow the optimization on interfaces 1215 // which have a single implementor. This would require us to increase the 1216 // expressiveness of the add_dependency() mechanism. 1217 // %%% Do this after we fix TypeOopPtr: Deps are expressive enough now. 1218 1219 // Object arrays must have their base element have no subtypes 1220 while (superklass->is_obj_array_klass()) { 1221 ciType* elem = superklass->as_obj_array_klass()->element_type(); 1222 superklass = elem->as_klass(); 1223 } 1224 if (superklass->is_instance_klass()) { 1225 ciInstanceKlass* ik = superklass->as_instance_klass(); 1226 if (ik->has_subklass() || ik->is_interface()) return nullptr; 1227 // Add a dependency if there is a chance that a subclass will be added later. 1228 if (!ik->is_final()) { 1229 phase->C->dependencies()->assert_leaf_type(ik); 1230 } 1231 } 1232 1233 // Bypass the dependent load, and compare directly 1234 this->set_req_X(1, ldk2, phase); 1235 1236 return this; 1237 } 1238 1239 //============================================================================= 1240 //------------------------------sub-------------------------------------------- 1241 // Simplify an CmpN (compare 2 pointers) node, based on local information. 1242 // If both inputs are constants, compare them. 1243 const Type *CmpNNode::sub( const Type *t1, const Type *t2 ) const { 1244 ShouldNotReachHere(); 1245 return bottom_type(); 1246 } 1247 1248 //------------------------------Ideal------------------------------------------ 1249 Node *CmpNNode::Ideal( PhaseGVN *phase, bool can_reshape ) { 1250 return nullptr; 1251 } 1252 1253 //============================================================================= 1254 //------------------------------Value------------------------------------------ 1255 // Simplify an CmpF (compare 2 floats ) node, based on local information. 1256 // If both inputs are constants, compare them. 1257 const Type* CmpFNode::Value(PhaseGVN* phase) const { 1258 const Node* in1 = in(1); 1259 const Node* in2 = in(2); 1260 // Either input is TOP ==> the result is TOP 1261 const Type* t1 = (in1 == this) ? Type::TOP : phase->type(in1); 1262 if( t1 == Type::TOP ) return Type::TOP; 1263 const Type* t2 = (in2 == this) ? Type::TOP : phase->type(in2); 1264 if( t2 == Type::TOP ) return Type::TOP; 1265 1266 // Not constants? Don't know squat - even if they are the same 1267 // value! If they are NaN's they compare to LT instead of EQ. 1268 const TypeF *tf1 = t1->isa_float_constant(); 1269 const TypeF *tf2 = t2->isa_float_constant(); 1270 if( !tf1 || !tf2 ) return TypeInt::CC; 1271 1272 // This implements the Java bytecode fcmpl, so unordered returns -1. 1273 if( tf1->is_nan() || tf2->is_nan() ) 1274 return TypeInt::CC_LT; 1275 1276 if( tf1->_f < tf2->_f ) return TypeInt::CC_LT; 1277 if( tf1->_f > tf2->_f ) return TypeInt::CC_GT; 1278 assert( tf1->_f == tf2->_f, "do not understand FP behavior" ); 1279 return TypeInt::CC_EQ; 1280 } 1281 1282 1283 //============================================================================= 1284 //------------------------------Value------------------------------------------ 1285 // Simplify an CmpD (compare 2 doubles ) node, based on local information. 1286 // If both inputs are constants, compare them. 1287 const Type* CmpDNode::Value(PhaseGVN* phase) const { 1288 const Node* in1 = in(1); 1289 const Node* in2 = in(2); 1290 // Either input is TOP ==> the result is TOP 1291 const Type* t1 = (in1 == this) ? Type::TOP : phase->type(in1); 1292 if( t1 == Type::TOP ) return Type::TOP; 1293 const Type* t2 = (in2 == this) ? Type::TOP : phase->type(in2); 1294 if( t2 == Type::TOP ) return Type::TOP; 1295 1296 // Not constants? Don't know squat - even if they are the same 1297 // value! If they are NaN's they compare to LT instead of EQ. 1298 const TypeD *td1 = t1->isa_double_constant(); 1299 const TypeD *td2 = t2->isa_double_constant(); 1300 if( !td1 || !td2 ) return TypeInt::CC; 1301 1302 // This implements the Java bytecode dcmpl, so unordered returns -1. 1303 if( td1->is_nan() || td2->is_nan() ) 1304 return TypeInt::CC_LT; 1305 1306 if( td1->_d < td2->_d ) return TypeInt::CC_LT; 1307 if( td1->_d > td2->_d ) return TypeInt::CC_GT; 1308 assert( td1->_d == td2->_d, "do not understand FP behavior" ); 1309 return TypeInt::CC_EQ; 1310 } 1311 1312 //------------------------------Ideal------------------------------------------ 1313 Node *CmpDNode::Ideal(PhaseGVN *phase, bool can_reshape){ 1314 // Check if we can change this to a CmpF and remove a ConvD2F operation. 1315 // Change (CMPD (F2D (float)) (ConD value)) 1316 // To (CMPF (float) (ConF value)) 1317 // Valid when 'value' does not lose precision as a float. 1318 // Benefits: eliminates conversion, does not require 24-bit mode 1319 1320 // NaNs prevent commuting operands. This transform works regardless of the 1321 // order of ConD and ConvF2D inputs by preserving the original order. 1322 int idx_f2d = 1; // ConvF2D on left side? 1323 if( in(idx_f2d)->Opcode() != Op_ConvF2D ) 1324 idx_f2d = 2; // No, swap to check for reversed args 1325 int idx_con = 3-idx_f2d; // Check for the constant on other input 1326 1327 if( ConvertCmpD2CmpF && 1328 in(idx_f2d)->Opcode() == Op_ConvF2D && 1329 in(idx_con)->Opcode() == Op_ConD ) { 1330 const TypeD *t2 = in(idx_con)->bottom_type()->is_double_constant(); 1331 double t2_value_as_double = t2->_d; 1332 float t2_value_as_float = (float)t2_value_as_double; 1333 if( t2_value_as_double == (double)t2_value_as_float ) { 1334 // Test value can be represented as a float 1335 // Eliminate the conversion to double and create new comparison 1336 Node *new_in1 = in(idx_f2d)->in(1); 1337 Node *new_in2 = phase->makecon( TypeF::make(t2_value_as_float) ); 1338 if( idx_f2d != 1 ) { // Must flip args to match original order 1339 Node *tmp = new_in1; 1340 new_in1 = new_in2; 1341 new_in2 = tmp; 1342 } 1343 CmpFNode *new_cmp = (Opcode() == Op_CmpD3) 1344 ? new CmpF3Node( new_in1, new_in2 ) 1345 : new CmpFNode ( new_in1, new_in2 ) ; 1346 return new_cmp; // Changed to CmpFNode 1347 } 1348 // Testing value required the precision of a double 1349 } 1350 return nullptr; // No change 1351 } 1352 1353 1354 //============================================================================= 1355 //------------------------------cc2logical------------------------------------- 1356 // Convert a condition code type to a logical type 1357 const Type *BoolTest::cc2logical( const Type *CC ) const { 1358 if( CC == Type::TOP ) return Type::TOP; 1359 if( CC->base() != Type::Int ) return TypeInt::BOOL; // Bottom or worse 1360 const TypeInt *ti = CC->is_int(); 1361 if( ti->is_con() ) { // Only 1 kind of condition codes set? 1362 // Match low order 2 bits 1363 int tmp = ((ti->get_con()&3) == (_test&3)) ? 1 : 0; 1364 if( _test & 4 ) tmp = 1-tmp; // Optionally complement result 1365 return TypeInt::make(tmp); // Boolean result 1366 } 1367 1368 if( CC == TypeInt::CC_GE ) { 1369 if( _test == ge ) return TypeInt::ONE; 1370 if( _test == lt ) return TypeInt::ZERO; 1371 } 1372 if( CC == TypeInt::CC_LE ) { 1373 if( _test == le ) return TypeInt::ONE; 1374 if( _test == gt ) return TypeInt::ZERO; 1375 } 1376 1377 return TypeInt::BOOL; 1378 } 1379 1380 //------------------------------dump_spec------------------------------------- 1381 // Print special per-node info 1382 void BoolTest::dump_on(outputStream *st) const { 1383 const char *msg[] = {"eq","gt","of","lt","ne","le","nof","ge"}; 1384 st->print("%s", msg[_test]); 1385 } 1386 1387 // Returns the logical AND of two tests (or 'never' if both tests can never be true). 1388 // For example, a test for 'le' followed by a test for 'lt' is equivalent with 'lt'. 1389 BoolTest::mask BoolTest::merge(BoolTest other) const { 1390 const mask res[illegal+1][illegal+1] = { 1391 // eq, gt, of, lt, ne, le, nof, ge, never, illegal 1392 {eq, never, illegal, never, never, eq, illegal, eq, never, illegal}, // eq 1393 {never, gt, illegal, never, gt, never, illegal, gt, never, illegal}, // gt 1394 {illegal, illegal, illegal, illegal, illegal, illegal, illegal, illegal, never, illegal}, // of 1395 {never, never, illegal, lt, lt, lt, illegal, never, never, illegal}, // lt 1396 {never, gt, illegal, lt, ne, lt, illegal, gt, never, illegal}, // ne 1397 {eq, never, illegal, lt, lt, le, illegal, eq, never, illegal}, // le 1398 {illegal, illegal, illegal, illegal, illegal, illegal, illegal, illegal, never, illegal}, // nof 1399 {eq, gt, illegal, never, gt, eq, illegal, ge, never, illegal}, // ge 1400 {never, never, never, never, never, never, never, never, never, illegal}, // never 1401 {illegal, illegal, illegal, illegal, illegal, illegal, illegal, illegal, illegal, illegal}}; // illegal 1402 return res[_test][other._test]; 1403 } 1404 1405 //============================================================================= 1406 uint BoolNode::hash() const { return (Node::hash() << 3)|(_test._test+1); } 1407 uint BoolNode::size_of() const { return sizeof(BoolNode); } 1408 1409 //------------------------------operator==------------------------------------- 1410 bool BoolNode::cmp( const Node &n ) const { 1411 const BoolNode *b = (const BoolNode *)&n; // Cast up 1412 return (_test._test == b->_test._test); 1413 } 1414 1415 //-------------------------------make_predicate-------------------------------- 1416 Node* BoolNode::make_predicate(Node* test_value, PhaseGVN* phase) { 1417 if (test_value->is_Con()) return test_value; 1418 if (test_value->is_Bool()) return test_value; 1419 if (test_value->is_CMove() && 1420 test_value->in(CMoveNode::Condition)->is_Bool()) { 1421 BoolNode* bol = test_value->in(CMoveNode::Condition)->as_Bool(); 1422 const Type* ftype = phase->type(test_value->in(CMoveNode::IfFalse)); 1423 const Type* ttype = phase->type(test_value->in(CMoveNode::IfTrue)); 1424 if (ftype == TypeInt::ZERO && !TypeInt::ZERO->higher_equal(ttype)) { 1425 return bol; 1426 } else if (ttype == TypeInt::ZERO && !TypeInt::ZERO->higher_equal(ftype)) { 1427 return phase->transform( bol->negate(phase) ); 1428 } 1429 // Else fall through. The CMove gets in the way of the test. 1430 // It should be the case that make_predicate(bol->as_int_value()) == bol. 1431 } 1432 Node* cmp = new CmpINode(test_value, phase->intcon(0)); 1433 cmp = phase->transform(cmp); 1434 Node* bol = new BoolNode(cmp, BoolTest::ne); 1435 return phase->transform(bol); 1436 } 1437 1438 //--------------------------------as_int_value--------------------------------- 1439 Node* BoolNode::as_int_value(PhaseGVN* phase) { 1440 // Inverse to make_predicate. The CMove probably boils down to a Conv2B. 1441 Node* cmov = CMoveNode::make(this, phase->intcon(0), phase->intcon(1), TypeInt::BOOL); 1442 return phase->transform(cmov); 1443 } 1444 1445 //----------------------------------negate------------------------------------- 1446 BoolNode* BoolNode::negate(PhaseGVN* phase) { 1447 return new BoolNode(in(1), _test.negate()); 1448 } 1449 1450 // Change "bool eq/ne (cmp (add/sub A B) C)" into false/true if add/sub 1451 // overflows and we can prove that C is not in the two resulting ranges. 1452 // This optimization is similar to the one performed by CmpUNode::Value(). 1453 Node* BoolNode::fold_cmpI(PhaseGVN* phase, SubNode* cmp, Node* cmp1, int cmp_op, 1454 int cmp1_op, const TypeInt* cmp2_type) { 1455 // Only optimize eq/ne integer comparison of add/sub 1456 if((_test._test == BoolTest::eq || _test._test == BoolTest::ne) && 1457 (cmp_op == Op_CmpI) && (cmp1_op == Op_AddI || cmp1_op == Op_SubI)) { 1458 // Skip cases were inputs of add/sub are not integers or of bottom type 1459 const TypeInt* r0 = phase->type(cmp1->in(1))->isa_int(); 1460 const TypeInt* r1 = phase->type(cmp1->in(2))->isa_int(); 1461 if ((r0 != nullptr) && (r0 != TypeInt::INT) && 1462 (r1 != nullptr) && (r1 != TypeInt::INT) && 1463 (cmp2_type != TypeInt::INT)) { 1464 // Compute exact (long) type range of add/sub result 1465 jlong lo_long = r0->_lo; 1466 jlong hi_long = r0->_hi; 1467 if (cmp1_op == Op_AddI) { 1468 lo_long += r1->_lo; 1469 hi_long += r1->_hi; 1470 } else { 1471 lo_long -= r1->_hi; 1472 hi_long -= r1->_lo; 1473 } 1474 // Check for over-/underflow by casting to integer 1475 int lo_int = (int)lo_long; 1476 int hi_int = (int)hi_long; 1477 bool underflow = lo_long != (jlong)lo_int; 1478 bool overflow = hi_long != (jlong)hi_int; 1479 if ((underflow != overflow) && (hi_int < lo_int)) { 1480 // Overflow on one boundary, compute resulting type ranges: 1481 // tr1 [MIN_INT, hi_int] and tr2 [lo_int, MAX_INT] 1482 int w = MAX2(r0->_widen, r1->_widen); // _widen does not matter here 1483 const TypeInt* tr1 = TypeInt::make(min_jint, hi_int, w); 1484 const TypeInt* tr2 = TypeInt::make(lo_int, max_jint, w); 1485 // Compare second input of cmp to both type ranges 1486 const Type* sub_tr1 = cmp->sub(tr1, cmp2_type); 1487 const Type* sub_tr2 = cmp->sub(tr2, cmp2_type); 1488 if (sub_tr1 == TypeInt::CC_LT && sub_tr2 == TypeInt::CC_GT) { 1489 // The result of the add/sub will never equal cmp2. Replace BoolNode 1490 // by false (0) if it tests for equality and by true (1) otherwise. 1491 return ConINode::make((_test._test == BoolTest::eq) ? 0 : 1); 1492 } 1493 } 1494 } 1495 } 1496 return nullptr; 1497 } 1498 1499 static bool is_counted_loop_cmp(Node *cmp) { 1500 Node *n = cmp->in(1)->in(1); 1501 return n != nullptr && 1502 n->is_Phi() && 1503 n->in(0) != nullptr && 1504 n->in(0)->is_CountedLoop() && 1505 n->in(0)->as_CountedLoop()->phi() == n; 1506 } 1507 1508 //------------------------------Ideal------------------------------------------ 1509 Node *BoolNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1510 // Change "bool tst (cmp con x)" into "bool ~tst (cmp x con)". 1511 // This moves the constant to the right. Helps value-numbering. 1512 Node *cmp = in(1); 1513 if( !cmp->is_Sub() ) return nullptr; 1514 int cop = cmp->Opcode(); 1515 if( cop == Op_FastLock || cop == Op_FastUnlock || 1516 cmp->is_SubTypeCheck() || cop == Op_VectorTest ) { 1517 return nullptr; 1518 } 1519 Node *cmp1 = cmp->in(1); 1520 Node *cmp2 = cmp->in(2); 1521 if( !cmp1 ) return nullptr; 1522 1523 if (_test._test == BoolTest::overflow || _test._test == BoolTest::no_overflow) { 1524 return nullptr; 1525 } 1526 1527 const int cmp1_op = cmp1->Opcode(); 1528 const int cmp2_op = cmp2->Opcode(); 1529 1530 // Constant on left? 1531 Node *con = cmp1; 1532 // Move constants to the right of compare's to canonicalize. 1533 // Do not muck with Opaque1 nodes, as this indicates a loop 1534 // guard that cannot change shape. 1535 if (con->is_Con() && !cmp2->is_Con() && cmp2_op != Op_OpaqueZeroTripGuard && 1536 // Because of NaN's, CmpD and CmpF are not commutative 1537 cop != Op_CmpD && cop != Op_CmpF && 1538 // Protect against swapping inputs to a compare when it is used by a 1539 // counted loop exit, which requires maintaining the loop-limit as in(2) 1540 !is_counted_loop_exit_test() ) { 1541 // Ok, commute the constant to the right of the cmp node. 1542 // Clone the Node, getting a new Node of the same class 1543 cmp = cmp->clone(); 1544 // Swap inputs to the clone 1545 cmp->swap_edges(1, 2); 1546 cmp = phase->transform( cmp ); 1547 return new BoolNode( cmp, _test.commute() ); 1548 } 1549 1550 // Change "bool eq/ne (cmp (cmove (bool tst (cmp2)) 1 0) 0)" into "bool tst/~tst (cmp2)" 1551 if (cop == Op_CmpI && 1552 (_test._test == BoolTest::eq || _test._test == BoolTest::ne) && 1553 cmp1_op == Op_CMoveI && cmp2->find_int_con(1) == 0) { 1554 // 0 should be on the true branch 1555 if (cmp1->in(CMoveNode::Condition)->is_Bool() && 1556 cmp1->in(CMoveNode::IfTrue)->find_int_con(1) == 0 && 1557 cmp1->in(CMoveNode::IfFalse)->find_int_con(0) != 0) { 1558 BoolNode* target = cmp1->in(CMoveNode::Condition)->as_Bool(); 1559 return new BoolNode(target->in(1), 1560 (_test._test == BoolTest::eq) ? target->_test._test : 1561 target->_test.negate()); 1562 } 1563 } 1564 1565 // Change "bool eq/ne (cmp (and X 16) 16)" into "bool ne/eq (cmp (and X 16) 0)". 1566 if (cop == Op_CmpI && 1567 (_test._test == BoolTest::eq || _test._test == BoolTest::ne) && 1568 cmp1_op == Op_AndI && cmp2_op == Op_ConI && 1569 cmp1->in(2)->Opcode() == Op_ConI) { 1570 const TypeInt *t12 = phase->type(cmp2)->isa_int(); 1571 const TypeInt *t112 = phase->type(cmp1->in(2))->isa_int(); 1572 if (t12 && t12->is_con() && t112 && t112->is_con() && 1573 t12->get_con() == t112->get_con() && is_power_of_2(t12->get_con())) { 1574 Node *ncmp = phase->transform(new CmpINode(cmp1, phase->intcon(0))); 1575 return new BoolNode(ncmp, _test.negate()); 1576 } 1577 } 1578 1579 // Same for long type: change "bool eq/ne (cmp (and X 16) 16)" into "bool ne/eq (cmp (and X 16) 0)". 1580 if (cop == Op_CmpL && 1581 (_test._test == BoolTest::eq || _test._test == BoolTest::ne) && 1582 cmp1_op == Op_AndL && cmp2_op == Op_ConL && 1583 cmp1->in(2)->Opcode() == Op_ConL) { 1584 const TypeLong *t12 = phase->type(cmp2)->isa_long(); 1585 const TypeLong *t112 = phase->type(cmp1->in(2))->isa_long(); 1586 if (t12 && t12->is_con() && t112 && t112->is_con() && 1587 t12->get_con() == t112->get_con() && is_power_of_2(t12->get_con())) { 1588 Node *ncmp = phase->transform(new CmpLNode(cmp1, phase->longcon(0))); 1589 return new BoolNode(ncmp, _test.negate()); 1590 } 1591 } 1592 1593 // Change "cmp (add X min_jint) (add Y min_jint)" into "cmpu X Y" 1594 // and "cmp (add X min_jint) c" into "cmpu X (c + min_jint)" 1595 if (cop == Op_CmpI && 1596 cmp1_op == Op_AddI && 1597 phase->type(cmp1->in(2)) == TypeInt::MIN && 1598 !is_cloop_condition(this)) { 1599 if (cmp2_op == Op_ConI) { 1600 Node* ncmp2 = phase->intcon(java_add(cmp2->get_int(), min_jint)); 1601 Node* ncmp = phase->transform(new CmpUNode(cmp1->in(1), ncmp2)); 1602 return new BoolNode(ncmp, _test._test); 1603 } else if (cmp2_op == Op_AddI && 1604 phase->type(cmp2->in(2)) == TypeInt::MIN && 1605 !is_cloop_condition(this)) { 1606 Node* ncmp = phase->transform(new CmpUNode(cmp1->in(1), cmp2->in(1))); 1607 return new BoolNode(ncmp, _test._test); 1608 } 1609 } 1610 1611 // Change "cmp (add X min_jlong) (add Y min_jlong)" into "cmpu X Y" 1612 // and "cmp (add X min_jlong) c" into "cmpu X (c + min_jlong)" 1613 if (cop == Op_CmpL && 1614 cmp1_op == Op_AddL && 1615 phase->type(cmp1->in(2)) == TypeLong::MIN && 1616 !is_cloop_condition(this)) { 1617 if (cmp2_op == Op_ConL) { 1618 Node* ncmp2 = phase->longcon(java_add(cmp2->get_long(), min_jlong)); 1619 Node* ncmp = phase->transform(new CmpULNode(cmp1->in(1), ncmp2)); 1620 return new BoolNode(ncmp, _test._test); 1621 } else if (cmp2_op == Op_AddL && 1622 phase->type(cmp2->in(2)) == TypeLong::MIN && 1623 !is_cloop_condition(this)) { 1624 Node* ncmp = phase->transform(new CmpULNode(cmp1->in(1), cmp2->in(1))); 1625 return new BoolNode(ncmp, _test._test); 1626 } 1627 } 1628 1629 // Change "bool eq/ne (cmp (xor X 1) 0)" into "bool ne/eq (cmp X 0)". 1630 // The XOR-1 is an idiom used to flip the sense of a bool. We flip the 1631 // test instead. 1632 const TypeInt* cmp2_type = phase->type(cmp2)->isa_int(); 1633 if (cmp2_type == nullptr) return nullptr; 1634 Node* j_xor = cmp1; 1635 if( cmp2_type == TypeInt::ZERO && 1636 cmp1_op == Op_XorI && 1637 j_xor->in(1) != j_xor && // An xor of itself is dead 1638 phase->type( j_xor->in(1) ) == TypeInt::BOOL && 1639 phase->type( j_xor->in(2) ) == TypeInt::ONE && 1640 (_test._test == BoolTest::eq || 1641 _test._test == BoolTest::ne) ) { 1642 Node *ncmp = phase->transform(new CmpINode(j_xor->in(1),cmp2)); 1643 return new BoolNode( ncmp, _test.negate() ); 1644 } 1645 1646 // Transform: "((x & (m - 1)) <u m)" or "(((m - 1) & x) <u m)" into "(m >u 0)" 1647 // This is case [CMPU_MASK] which is further described at the method comment of BoolNode::Value_cmpu_and_mask(). 1648 if (cop == Op_CmpU && _test._test == BoolTest::lt && cmp1_op == Op_AndI) { 1649 Node* m = cmp2; // RHS: m 1650 for (int add_idx = 1; add_idx <= 2; add_idx++) { // LHS: "(m + (-1)) & x" or "x & (m + (-1))"? 1651 Node* maybe_m_minus_1 = cmp1->in(add_idx); 1652 if (maybe_m_minus_1->Opcode() == Op_AddI && 1653 maybe_m_minus_1->in(2)->find_int_con(0) == -1 && 1654 maybe_m_minus_1->in(1) == m) { 1655 Node* m_cmpu_0 = phase->transform(new CmpUNode(m, phase->intcon(0))); 1656 return new BoolNode(m_cmpu_0, BoolTest::gt); 1657 } 1658 } 1659 } 1660 1661 // Change x u< 1 or x u<= 0 to x == 0 1662 // and x u> 0 or u>= 1 to x != 0 1663 if (cop == Op_CmpU && 1664 cmp1_op != Op_LoadRange && 1665 (((_test._test == BoolTest::lt || _test._test == BoolTest::ge) && 1666 cmp2->find_int_con(-1) == 1) || 1667 ((_test._test == BoolTest::le || _test._test == BoolTest::gt) && 1668 cmp2->find_int_con(-1) == 0))) { 1669 Node* ncmp = phase->transform(new CmpINode(cmp1, phase->intcon(0))); 1670 return new BoolNode(ncmp, _test.is_less() ? BoolTest::eq : BoolTest::ne); 1671 } 1672 1673 // Change (arraylength <= 0) or (arraylength == 0) 1674 // into (arraylength u<= 0) 1675 // Also change (arraylength != 0) into (arraylength u> 0) 1676 // The latter version matches the code pattern generated for 1677 // array range checks, which will more likely be optimized later. 1678 if (cop == Op_CmpI && 1679 cmp1_op == Op_LoadRange && 1680 cmp2->find_int_con(-1) == 0) { 1681 if (_test._test == BoolTest::le || _test._test == BoolTest::eq) { 1682 Node* ncmp = phase->transform(new CmpUNode(cmp1, cmp2)); 1683 return new BoolNode(ncmp, BoolTest::le); 1684 } else if (_test._test == BoolTest::ne) { 1685 Node* ncmp = phase->transform(new CmpUNode(cmp1, cmp2)); 1686 return new BoolNode(ncmp, BoolTest::gt); 1687 } 1688 } 1689 1690 // Change "bool eq/ne (cmp (Conv2B X) 0)" into "bool eq/ne (cmp X 0)". 1691 // This is a standard idiom for branching on a boolean value. 1692 Node *c2b = cmp1; 1693 if( cmp2_type == TypeInt::ZERO && 1694 cmp1_op == Op_Conv2B && 1695 (_test._test == BoolTest::eq || 1696 _test._test == BoolTest::ne) ) { 1697 Node *ncmp = phase->transform(phase->type(c2b->in(1))->isa_int() 1698 ? (Node*)new CmpINode(c2b->in(1),cmp2) 1699 : (Node*)new CmpPNode(c2b->in(1),phase->makecon(TypePtr::NULL_PTR)) 1700 ); 1701 return new BoolNode( ncmp, _test._test ); 1702 } 1703 1704 // Comparing a SubI against a zero is equal to comparing the SubI 1705 // arguments directly. This only works for eq and ne comparisons 1706 // due to possible integer overflow. 1707 if ((_test._test == BoolTest::eq || _test._test == BoolTest::ne) && 1708 (cop == Op_CmpI) && 1709 (cmp1_op == Op_SubI) && 1710 ( cmp2_type == TypeInt::ZERO ) ) { 1711 Node *ncmp = phase->transform( new CmpINode(cmp1->in(1),cmp1->in(2))); 1712 return new BoolNode( ncmp, _test._test ); 1713 } 1714 1715 // Same as above but with and AddI of a constant 1716 if ((_test._test == BoolTest::eq || _test._test == BoolTest::ne) && 1717 cop == Op_CmpI && 1718 cmp1_op == Op_AddI && 1719 cmp1->in(2) != nullptr && 1720 phase->type(cmp1->in(2))->isa_int() && 1721 phase->type(cmp1->in(2))->is_int()->is_con() && 1722 cmp2_type == TypeInt::ZERO && 1723 !is_counted_loop_cmp(cmp) // modifying the exit test of a counted loop messes the counted loop shape 1724 ) { 1725 const TypeInt* cmp1_in2 = phase->type(cmp1->in(2))->is_int(); 1726 Node *ncmp = phase->transform( new CmpINode(cmp1->in(1),phase->intcon(-cmp1_in2->_hi))); 1727 return new BoolNode( ncmp, _test._test ); 1728 } 1729 1730 // Change "bool eq/ne (cmp (phi (X -X) 0))" into "bool eq/ne (cmp X 0)" 1731 // since zero check of conditional negation of an integer is equal to 1732 // zero check of the integer directly. 1733 if ((_test._test == BoolTest::eq || _test._test == BoolTest::ne) && 1734 (cop == Op_CmpI) && 1735 (cmp2_type == TypeInt::ZERO) && 1736 (cmp1_op == Op_Phi)) { 1737 // There should be a diamond phi with true path at index 1 or 2 1738 PhiNode *phi = cmp1->as_Phi(); 1739 int idx_true = phi->is_diamond_phi(); 1740 if (idx_true != 0) { 1741 // True input is in(idx_true) while false input is in(3 - idx_true) 1742 Node *tin = phi->in(idx_true); 1743 Node *fin = phi->in(3 - idx_true); 1744 if ((tin->Opcode() == Op_SubI) && 1745 (phase->type(tin->in(1)) == TypeInt::ZERO) && 1746 (tin->in(2) == fin)) { 1747 // Found conditional negation at true path, create a new CmpINode without that 1748 Node *ncmp = phase->transform(new CmpINode(fin, cmp2)); 1749 return new BoolNode(ncmp, _test._test); 1750 } 1751 if ((fin->Opcode() == Op_SubI) && 1752 (phase->type(fin->in(1)) == TypeInt::ZERO) && 1753 (fin->in(2) == tin)) { 1754 // Found conditional negation at false path, create a new CmpINode without that 1755 Node *ncmp = phase->transform(new CmpINode(tin, cmp2)); 1756 return new BoolNode(ncmp, _test._test); 1757 } 1758 } 1759 } 1760 1761 // Change (-A vs 0) into (A vs 0) by commuting the test. Disallow in the 1762 // most general case because negating 0x80000000 does nothing. Needed for 1763 // the CmpF3/SubI/CmpI idiom. 1764 if( cop == Op_CmpI && 1765 cmp1_op == Op_SubI && 1766 cmp2_type == TypeInt::ZERO && 1767 phase->type( cmp1->in(1) ) == TypeInt::ZERO && 1768 phase->type( cmp1->in(2) )->higher_equal(TypeInt::SYMINT) ) { 1769 Node *ncmp = phase->transform( new CmpINode(cmp1->in(2),cmp2)); 1770 return new BoolNode( ncmp, _test.commute() ); 1771 } 1772 1773 // Try to optimize signed integer comparison 1774 return fold_cmpI(phase, cmp->as_Sub(), cmp1, cop, cmp1_op, cmp2_type); 1775 1776 // The transformation below is not valid for either signed or unsigned 1777 // comparisons due to wraparound concerns at MAX_VALUE and MIN_VALUE. 1778 // This transformation can be resurrected when we are able to 1779 // make inferences about the range of values being subtracted from 1780 // (or added to) relative to the wraparound point. 1781 // 1782 // // Remove +/-1's if possible. 1783 // // "X <= Y-1" becomes "X < Y" 1784 // // "X+1 <= Y" becomes "X < Y" 1785 // // "X < Y+1" becomes "X <= Y" 1786 // // "X-1 < Y" becomes "X <= Y" 1787 // // Do not this to compares off of the counted-loop-end. These guys are 1788 // // checking the trip counter and they want to use the post-incremented 1789 // // counter. If they use the PRE-incremented counter, then the counter has 1790 // // to be incremented in a private block on a loop backedge. 1791 // if( du && du->cnt(this) && du->out(this)[0]->Opcode() == Op_CountedLoopEnd ) 1792 // return nullptr; 1793 // #ifndef PRODUCT 1794 // // Do not do this in a wash GVN pass during verification. 1795 // // Gets triggered by too many simple optimizations to be bothered with 1796 // // re-trying it again and again. 1797 // if( !phase->allow_progress() ) return nullptr; 1798 // #endif 1799 // // Not valid for unsigned compare because of corner cases in involving zero. 1800 // // For example, replacing "X-1 <u Y" with "X <=u Y" fails to throw an 1801 // // exception in case X is 0 (because 0-1 turns into 4billion unsigned but 1802 // // "0 <=u Y" is always true). 1803 // if( cmp->Opcode() == Op_CmpU ) return nullptr; 1804 // int cmp2_op = cmp2->Opcode(); 1805 // if( _test._test == BoolTest::le ) { 1806 // if( cmp1_op == Op_AddI && 1807 // phase->type( cmp1->in(2) ) == TypeInt::ONE ) 1808 // return clone_cmp( cmp, cmp1->in(1), cmp2, phase, BoolTest::lt ); 1809 // else if( cmp2_op == Op_AddI && 1810 // phase->type( cmp2->in(2) ) == TypeInt::MINUS_1 ) 1811 // return clone_cmp( cmp, cmp1, cmp2->in(1), phase, BoolTest::lt ); 1812 // } else if( _test._test == BoolTest::lt ) { 1813 // if( cmp1_op == Op_AddI && 1814 // phase->type( cmp1->in(2) ) == TypeInt::MINUS_1 ) 1815 // return clone_cmp( cmp, cmp1->in(1), cmp2, phase, BoolTest::le ); 1816 // else if( cmp2_op == Op_AddI && 1817 // phase->type( cmp2->in(2) ) == TypeInt::ONE ) 1818 // return clone_cmp( cmp, cmp1, cmp2->in(1), phase, BoolTest::le ); 1819 // } 1820 } 1821 1822 // We use the following Lemmas/insights for the following two transformations (1) and (2): 1823 // x & y <=u y, for any x and y (Lemma 1, masking always results in a smaller unsigned number) 1824 // y <u y + 1 is always true if y != -1 (Lemma 2, (uint)(-1 + 1) == (uint)(UINT_MAX + 1) which overflows) 1825 // y <u 0 is always false for any y (Lemma 3, 0 == UINT_MIN and nothing can be smaller than that) 1826 // 1827 // (1a) Always: Change ((x & m) <=u m ) or ((m & x) <=u m ) to always true (true by Lemma 1) 1828 // (1b) If m != -1: Change ((x & m) <u m + 1) or ((m & x) <u m + 1) to always true: 1829 // x & m <=u m is always true // (Lemma 1) 1830 // x & m <=u m <u m + 1 is always true // (Lemma 2: m <u m + 1, if m != -1) 1831 // 1832 // A counter example for (1b), if we allowed m == -1: 1833 // (x & m) <u m + 1 1834 // (x & -1) <u 0 1835 // x <u 0 1836 // which is false for any x (Lemma 3) 1837 // 1838 // (2) Change ((x & (m - 1)) <u m) or (((m - 1) & x) <u m) to (m >u 0) 1839 // This is the off-by-one variant of the above. 1840 // 1841 // We now prove that this replacement is correct. This is the same as proving 1842 // "m >u 0" if and only if "x & (m - 1) <u m", i.e. "m >u 0 <=> x & (m - 1) <u m" 1843 // 1844 // We use (Lemma 1) and (Lemma 3) from above. 1845 // 1846 // Case "x & (m - 1) <u m => m >u 0": 1847 // We prove this by contradiction: 1848 // Assume m <=u 0 which is equivalent to m == 0: 1849 // and thus 1850 // x & (m - 1) <u m = 0 // m == 0 1851 // y <u 0 // y = x & (m - 1) 1852 // by Lemma 3, this is always false, i.e. a contradiction to our assumption. 1853 // 1854 // Case "m >u 0 => x & (m - 1) <u m": 1855 // x & (m - 1) <=u (m - 1) // (Lemma 1) 1856 // x & (m - 1) <=u (m - 1) <u m // Using assumption m >u 0, no underflow of "m - 1" 1857 // 1858 // 1859 // Note that the signed version of "m > 0": 1860 // m > 0 <=> x & (m - 1) <u m 1861 // does not hold: 1862 // Assume m == -1 and x == -1: 1863 // x & (m - 1) <u m 1864 // -1 & -2 <u -1 1865 // -2 <u -1 1866 // UINT_MAX - 1 <u UINT_MAX // Signed to unsigned numbers 1867 // which is true while 1868 // m > 0 1869 // is false which is a contradiction. 1870 // 1871 // (1a) and (1b) is covered by this method since we can directly return a true value as type while (2) is covered 1872 // in BoolNode::Ideal since we create a new non-constant node (see [CMPU_MASK]). 1873 const Type* BoolNode::Value_cmpu_and_mask(PhaseValues* phase) const { 1874 Node* cmp = in(1); 1875 if (cmp != nullptr && cmp->Opcode() == Op_CmpU) { 1876 Node* cmp1 = cmp->in(1); 1877 Node* cmp2 = cmp->in(2); 1878 1879 if (cmp1->Opcode() == Op_AndI) { 1880 Node* m = nullptr; 1881 if (_test._test == BoolTest::le) { 1882 // (1a) "((x & m) <=u m)", cmp2 = m 1883 m = cmp2; 1884 } else if (_test._test == BoolTest::lt && cmp2->Opcode() == Op_AddI && cmp2->in(2)->find_int_con(0) == 1) { 1885 // (1b) "(x & m) <u m + 1" and "(m & x) <u m + 1", cmp2 = m + 1 1886 Node* rhs_m = cmp2->in(1); 1887 const TypeInt* rhs_m_type = phase->type(rhs_m)->isa_int(); 1888 if (rhs_m_type->_lo > -1 || rhs_m_type->_hi < -1) { 1889 // Exclude any case where m == -1 is possible. 1890 m = rhs_m; 1891 } 1892 } 1893 1894 if (cmp1->in(2) == m || cmp1->in(1) == m) { 1895 return TypeInt::ONE; 1896 } 1897 } 1898 } 1899 1900 return nullptr; 1901 } 1902 1903 // Simplify a Bool (convert condition codes to boolean (1 or 0)) node, 1904 // based on local information. If the input is constant, do it. 1905 const Type* BoolNode::Value(PhaseGVN* phase) const { 1906 const Type* t = Value_cmpu_and_mask(phase); 1907 if (t != nullptr) { 1908 return t; 1909 } 1910 1911 return _test.cc2logical( phase->type( in(1) ) ); 1912 } 1913 1914 #ifndef PRODUCT 1915 //------------------------------dump_spec-------------------------------------- 1916 // Dump special per-node info 1917 void BoolNode::dump_spec(outputStream *st) const { 1918 st->print("["); 1919 _test.dump_on(st); 1920 st->print("]"); 1921 } 1922 #endif 1923 1924 //----------------------is_counted_loop_exit_test------------------------------ 1925 // Returns true if node is used by a counted loop node. 1926 bool BoolNode::is_counted_loop_exit_test() { 1927 for( DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++ ) { 1928 Node* use = fast_out(i); 1929 if (use->is_CountedLoopEnd()) { 1930 return true; 1931 } 1932 } 1933 return false; 1934 } 1935 1936 //============================================================================= 1937 //------------------------------Value------------------------------------------ 1938 const Type* AbsNode::Value(PhaseGVN* phase) const { 1939 const Type* t1 = phase->type(in(1)); 1940 if (t1 == Type::TOP) return Type::TOP; 1941 1942 switch (t1->base()) { 1943 case Type::Int: { 1944 const TypeInt* ti = t1->is_int(); 1945 if (ti->is_con()) { 1946 return TypeInt::make(uabs(ti->get_con())); 1947 } 1948 break; 1949 } 1950 case Type::Long: { 1951 const TypeLong* tl = t1->is_long(); 1952 if (tl->is_con()) { 1953 return TypeLong::make(uabs(tl->get_con())); 1954 } 1955 break; 1956 } 1957 case Type::FloatCon: 1958 return TypeF::make(abs(t1->getf())); 1959 case Type::DoubleCon: 1960 return TypeD::make(abs(t1->getd())); 1961 default: 1962 break; 1963 } 1964 1965 return bottom_type(); 1966 } 1967 1968 //------------------------------Identity---------------------------------------- 1969 Node* AbsNode::Identity(PhaseGVN* phase) { 1970 Node* in1 = in(1); 1971 // No need to do abs for non-negative values 1972 if (phase->type(in1)->higher_equal(TypeInt::POS) || 1973 phase->type(in1)->higher_equal(TypeLong::POS)) { 1974 return in1; 1975 } 1976 // Convert "abs(abs(x))" into "abs(x)" 1977 if (in1->Opcode() == Opcode()) { 1978 return in1; 1979 } 1980 return this; 1981 } 1982 1983 //------------------------------Ideal------------------------------------------ 1984 Node* AbsNode::Ideal(PhaseGVN* phase, bool can_reshape) { 1985 Node* in1 = in(1); 1986 // Convert "abs(0-x)" into "abs(x)" 1987 if (in1->is_Sub() && phase->type(in1->in(1))->is_zero_type()) { 1988 set_req_X(1, in1->in(2), phase); 1989 return this; 1990 } 1991 return nullptr; 1992 } 1993 1994 //============================================================================= 1995 //------------------------------Value------------------------------------------ 1996 // Compute sqrt 1997 const Type* SqrtDNode::Value(PhaseGVN* phase) const { 1998 const Type *t1 = phase->type( in(1) ); 1999 if( t1 == Type::TOP ) return Type::TOP; 2000 if( t1->base() != Type::DoubleCon ) return Type::DOUBLE; 2001 double d = t1->getd(); 2002 if( d < 0.0 ) return Type::DOUBLE; 2003 return TypeD::make( sqrt( d ) ); 2004 } 2005 2006 const Type* SqrtFNode::Value(PhaseGVN* phase) const { 2007 const Type *t1 = phase->type( in(1) ); 2008 if( t1 == Type::TOP ) return Type::TOP; 2009 if( t1->base() != Type::FloatCon ) return Type::FLOAT; 2010 float f = t1->getf(); 2011 if( f < 0.0f ) return Type::FLOAT; 2012 return TypeF::make( (float)sqrt( (double)f ) ); 2013 } 2014 2015 const Type* SqrtHFNode::Value(PhaseGVN* phase) const { 2016 const Type* t1 = phase->type(in(1)); 2017 if (t1 == Type::TOP) { return Type::TOP; } 2018 if (t1->base() != Type::HalfFloatCon) { return Type::HALF_FLOAT; } 2019 float f = t1->getf(); 2020 if (f < 0.0f) return Type::HALF_FLOAT; 2021 return TypeH::make((float)sqrt((double)f)); 2022 } 2023 2024 const Type* ReverseINode::Value(PhaseGVN* phase) const { 2025 const Type *t1 = phase->type( in(1) ); 2026 if (t1 == Type::TOP) { 2027 return Type::TOP; 2028 } 2029 const TypeInt* t1int = t1->isa_int(); 2030 if (t1int && t1int->is_con()) { 2031 jint res = reverse_bits(t1int->get_con()); 2032 return TypeInt::make(res); 2033 } 2034 return bottom_type(); 2035 } 2036 2037 const Type* ReverseLNode::Value(PhaseGVN* phase) const { 2038 const Type *t1 = phase->type( in(1) ); 2039 if (t1 == Type::TOP) { 2040 return Type::TOP; 2041 } 2042 const TypeLong* t1long = t1->isa_long(); 2043 if (t1long && t1long->is_con()) { 2044 jlong res = reverse_bits(t1long->get_con()); 2045 return TypeLong::make(res); 2046 } 2047 return bottom_type(); 2048 } 2049 2050 Node* InvolutionNode::Identity(PhaseGVN* phase) { 2051 // Op ( Op x ) => x 2052 if (in(1)->Opcode() == Opcode()) { 2053 return in(1)->in(1); 2054 } 2055 return this; 2056 }