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 // Remove double negation 57 const Type *zero = add_id(); 58 if( phase->type( in(1) )->higher_equal( zero ) && 59 in(2)->Opcode() == Opcode() && 60 phase->type( in(2)->in(1) )->higher_equal( zero ) ) { 61 return in(2)->in(2); 62 } 63 64 // Convert "(X+Y) - Y" into X and "(X+Y) - X" into Y 65 if (in(1)->Opcode() == Op_AddI || in(1)->Opcode() == Op_AddL) { 66 if (in(1)->in(2) == in(2)) { 67 return in(1)->in(1); 68 } 69 if (in(1)->in(1) == in(2)) { 70 return in(1)->in(2); 71 } 72 } 73 74 return ( phase->type( in(2) )->higher_equal( zero ) ) ? in(1) : this; 75 } 76 77 //------------------------------Value------------------------------------------ 78 // A subtract node differences it's two inputs. 79 const Type* SubNode::Value_common(PhaseValues* phase) const { 80 const Node* in1 = in(1); 81 const Node* in2 = in(2); 82 // Either input is TOP ==> the result is TOP 83 const Type* t1 = (in1 == this) ? Type::TOP : phase->type(in1); 84 if( t1 == Type::TOP ) return Type::TOP; 85 const Type* t2 = (in2 == this) ? Type::TOP : phase->type(in2); 86 if( t2 == Type::TOP ) return Type::TOP; 87 88 // Not correct for SubFnode and AddFNode (must check for infinity) 89 // Equal? Subtract is zero 90 if (in1->eqv_uncast(in2)) return add_id(); 91 92 // Either input is BOTTOM ==> the result is the local BOTTOM 93 if( t1 == Type::BOTTOM || t2 == Type::BOTTOM ) 94 return bottom_type(); 95 96 return nullptr; 97 } 98 99 const Type* SubNode::Value(PhaseGVN* phase) const { 100 const Type* t = Value_common(phase); 101 if (t != nullptr) { 102 return t; 103 } 104 const Type* t1 = phase->type(in(1)); 105 const Type* t2 = phase->type(in(2)); 106 return sub(t1,t2); // Local flavor of type subtraction 107 108 } 109 110 SubNode* SubNode::make(Node* in1, Node* in2, BasicType bt) { 111 switch (bt) { 112 case T_INT: 113 return new SubINode(in1, in2); 114 case T_LONG: 115 return new SubLNode(in1, in2); 116 default: 117 fatal("Not implemented for %s", type2name(bt)); 118 } 119 return nullptr; 120 } 121 122 //============================================================================= 123 //------------------------------Helper function-------------------------------- 124 125 static bool is_cloop_increment(Node* inc) { 126 precond(inc->Opcode() == Op_AddI || inc->Opcode() == Op_AddL); 127 128 if (!inc->in(1)->is_Phi()) { 129 return false; 130 } 131 const PhiNode* phi = inc->in(1)->as_Phi(); 132 133 if (!phi->region()->is_CountedLoop()) { 134 return false; 135 } 136 137 return inc == phi->region()->as_CountedLoop()->incr(); 138 } 139 140 // Given the expression '(x + C) - v', or 141 // 'v - (x + C)', we examine nodes '+' and 'v': 142 // 143 // 1. Do not convert if '+' is a counted-loop increment, because the '-' is 144 // loop invariant and converting extends the live-range of 'x' to overlap 145 // with the '+', forcing another register to be used in the loop. 146 // 147 // 2. Do not convert if 'v' is a counted-loop induction variable, because 148 // 'x' might be invariant. 149 // 150 static bool ok_to_convert(Node* inc, Node* var) { 151 return !(is_cloop_increment(inc) || var->is_cloop_ind_var()); 152 } 153 154 static bool is_cloop_condition(BoolNode* bol) { 155 for (DUIterator_Fast imax, i = bol->fast_outs(imax); i < imax; i++) { 156 Node* out = bol->fast_out(i); 157 if (out->is_BaseCountedLoopEnd()) { 158 return true; 159 } 160 } 161 return false; 162 } 163 164 //------------------------------Ideal------------------------------------------ 165 Node *SubINode::Ideal(PhaseGVN *phase, bool can_reshape){ 166 Node *in1 = in(1); 167 Node *in2 = in(2); 168 uint op1 = in1->Opcode(); 169 uint op2 = in2->Opcode(); 170 171 #ifdef ASSERT 172 // Check for dead loop 173 if ((in1 == this) || (in2 == this) || 174 ((op1 == Op_AddI || op1 == Op_SubI) && 175 ((in1->in(1) == this) || (in1->in(2) == this) || 176 (in1->in(1) == in1) || (in1->in(2) == in1)))) { 177 assert(false, "dead loop in SubINode::Ideal"); 178 } 179 #endif 180 181 const Type *t2 = phase->type( in2 ); 182 if( t2 == Type::TOP ) return nullptr; 183 // Convert "x-c0" into "x+ -c0". 184 if( t2->base() == Type::Int ){ // Might be bottom or top... 185 const TypeInt *i = t2->is_int(); 186 if( i->is_con() ) 187 return new AddINode(in1, phase->intcon(java_negate(i->get_con()))); 188 } 189 190 // Convert "(x+c0) - y" into (x-y) + c0" 191 // Do not collapse (x+c0)-y if "+" is a loop increment or 192 // if "y" is a loop induction variable. 193 if( op1 == Op_AddI && ok_to_convert(in1, in2) ) { 194 const Type *tadd = phase->type( in1->in(2) ); 195 if( tadd->singleton() && tadd != Type::TOP ) { 196 Node *sub2 = phase->transform( new SubINode( in1->in(1), in2 )); 197 return new AddINode( sub2, in1->in(2) ); 198 } 199 } 200 201 // Convert "x - (y+c0)" into "(x-y) - c0" AND 202 // Convert "c1 - (y+c0)" into "(c1-c0) - y" 203 // Need the same check as in above optimization but reversed. 204 if (op2 == Op_AddI 205 && ok_to_convert(in2, in1) 206 && in2->in(2)->Opcode() == Op_ConI) { 207 jint c0 = phase->type(in2->in(2))->isa_int()->get_con(); 208 Node* in21 = in2->in(1); 209 if (in1->Opcode() == Op_ConI) { 210 // Match c1 211 jint c1 = phase->type(in1)->isa_int()->get_con(); 212 Node* sub2 = phase->intcon(java_subtract(c1, c0)); 213 return new SubINode(sub2, in21); 214 } else { 215 // Match x 216 Node* sub2 = phase->transform(new SubINode(in1, in21)); 217 Node* neg_c0 = phase->intcon(java_negate(c0)); 218 return new AddINode(sub2, neg_c0); 219 } 220 } 221 222 const Type *t1 = phase->type( in1 ); 223 if( t1 == Type::TOP ) return nullptr; 224 225 #ifdef ASSERT 226 // Check for dead loop 227 if ((op2 == Op_AddI || op2 == Op_SubI) && 228 ((in2->in(1) == this) || (in2->in(2) == this) || 229 (in2->in(1) == in2) || (in2->in(2) == in2))) { 230 assert(false, "dead loop in SubINode::Ideal"); 231 } 232 #endif 233 234 // Convert "x - (x+y)" into "-y" 235 if (op2 == Op_AddI && in1 == in2->in(1)) { 236 return new SubINode(phase->intcon(0), in2->in(2)); 237 } 238 // Convert "(x-y) - x" into "-y" 239 if (op1 == Op_SubI && in1->in(1) == in2) { 240 return new SubINode(phase->intcon(0), in1->in(2)); 241 } 242 // Convert "x - (y+x)" into "-y" 243 if (op2 == Op_AddI && in1 == in2->in(2)) { 244 return new SubINode(phase->intcon(0), in2->in(1)); 245 } 246 247 // Convert "0 - (x-y)" into "y-x", leave the double negation "-(-y)" to SubNode::Identity(). 248 if (t1 == TypeInt::ZERO && op2 == Op_SubI && phase->type(in2->in(1)) != TypeInt::ZERO) { 249 return new SubINode(in2->in(2), in2->in(1)); 250 } 251 252 // Convert "0 - (x+con)" into "-con-x" 253 jint con; 254 if( t1 == TypeInt::ZERO && op2 == Op_AddI && 255 (con = in2->in(2)->find_int_con(0)) != 0 ) 256 return new SubINode( phase->intcon(-con), in2->in(1) ); 257 258 // Convert "(X+A) - (X+B)" into "A - B" 259 if( op1 == Op_AddI && op2 == Op_AddI && in1->in(1) == in2->in(1) ) 260 return new SubINode( in1->in(2), in2->in(2) ); 261 262 // Convert "(A+X) - (B+X)" into "A - B" 263 if( op1 == Op_AddI && op2 == Op_AddI && in1->in(2) == in2->in(2) ) 264 return new SubINode( in1->in(1), in2->in(1) ); 265 266 // Convert "(A+X) - (X+B)" into "A - B" 267 if( op1 == Op_AddI && op2 == Op_AddI && in1->in(2) == in2->in(1) ) 268 return new SubINode( in1->in(1), in2->in(2) ); 269 270 // Convert "(X+A) - (B+X)" into "A - B" 271 if( op1 == Op_AddI && op2 == Op_AddI && in1->in(1) == in2->in(2) ) 272 return new SubINode( in1->in(2), in2->in(1) ); 273 274 // Convert "A-(B-C)" into (A+C)-B", since add is commutative and generally 275 // nicer to optimize than subtract. 276 if( op2 == Op_SubI && in2->outcnt() == 1) { 277 Node *add1 = phase->transform( new AddINode( in1, in2->in(2) ) ); 278 return new SubINode( add1, in2->in(1) ); 279 } 280 281 // Associative 282 if (op1 == Op_MulI && op2 == Op_MulI) { 283 Node* sub_in1 = nullptr; 284 Node* sub_in2 = nullptr; 285 Node* mul_in = nullptr; 286 287 if (in1->in(1) == in2->in(1)) { 288 // Convert "a*b-a*c into a*(b-c) 289 sub_in1 = in1->in(2); 290 sub_in2 = in2->in(2); 291 mul_in = in1->in(1); 292 } else if (in1->in(2) == in2->in(1)) { 293 // Convert a*b-b*c into b*(a-c) 294 sub_in1 = in1->in(1); 295 sub_in2 = in2->in(2); 296 mul_in = in1->in(2); 297 } else if (in1->in(2) == in2->in(2)) { 298 // Convert a*c-b*c into (a-b)*c 299 sub_in1 = in1->in(1); 300 sub_in2 = in2->in(1); 301 mul_in = in1->in(2); 302 } else if (in1->in(1) == in2->in(2)) { 303 // Convert a*b-c*a into a*(b-c) 304 sub_in1 = in1->in(2); 305 sub_in2 = in2->in(1); 306 mul_in = in1->in(1); 307 } 308 309 if (mul_in != nullptr) { 310 Node* sub = phase->transform(new SubINode(sub_in1, sub_in2)); 311 return new MulINode(mul_in, sub); 312 } 313 } 314 315 // Convert "0-(A>>31)" into "(A>>>31)" 316 if ( op2 == Op_RShiftI ) { 317 Node *in21 = in2->in(1); 318 Node *in22 = in2->in(2); 319 const TypeInt *zero = phase->type(in1)->isa_int(); 320 const TypeInt *t21 = phase->type(in21)->isa_int(); 321 const TypeInt *t22 = phase->type(in22)->isa_int(); 322 if ( t21 && t22 && zero == TypeInt::ZERO && t22->is_con(31) ) { 323 return new URShiftINode(in21, in22); 324 } 325 } 326 327 return nullptr; 328 } 329 330 //------------------------------sub-------------------------------------------- 331 // A subtract node differences it's two inputs. 332 const Type *SubINode::sub( const Type *t1, const Type *t2 ) const { 333 const TypeInt *r0 = t1->is_int(); // Handy access 334 const TypeInt *r1 = t2->is_int(); 335 int32_t lo = java_subtract(r0->_lo, r1->_hi); 336 int32_t hi = java_subtract(r0->_hi, r1->_lo); 337 338 // We next check for 32-bit overflow. 339 // If that happens, we just assume all integers are possible. 340 if( (((r0->_lo ^ r1->_hi) >= 0) || // lo ends have same signs OR 341 ((r0->_lo ^ lo) >= 0)) && // lo results have same signs AND 342 (((r0->_hi ^ r1->_lo) >= 0) || // hi ends have same signs OR 343 ((r0->_hi ^ hi) >= 0)) ) // hi results have same signs 344 return TypeInt::make(lo,hi,MAX2(r0->_widen,r1->_widen)); 345 else // Overflow; assume all integers 346 return TypeInt::INT; 347 } 348 349 //============================================================================= 350 //------------------------------Ideal------------------------------------------ 351 Node *SubLNode::Ideal(PhaseGVN *phase, bool can_reshape) { 352 Node *in1 = in(1); 353 Node *in2 = in(2); 354 uint op1 = in1->Opcode(); 355 uint op2 = in2->Opcode(); 356 357 #ifdef ASSERT 358 // Check for dead loop 359 if ((in1 == this) || (in2 == this) || 360 ((op1 == Op_AddL || op1 == Op_SubL) && 361 ((in1->in(1) == this) || (in1->in(2) == this) || 362 (in1->in(1) == in1) || (in1->in(2) == in1)))) { 363 assert(false, "dead loop in SubLNode::Ideal"); 364 } 365 #endif 366 367 if( phase->type( in2 ) == Type::TOP ) return nullptr; 368 const TypeLong *i = phase->type( in2 )->isa_long(); 369 // Convert "x-c0" into "x+ -c0". 370 if( i && // Might be bottom or top... 371 i->is_con() ) 372 return new AddLNode(in1, phase->longcon(java_negate(i->get_con()))); 373 374 // Convert "(x+c0) - y" into (x-y) + c0" 375 // Do not collapse (x+c0)-y if "+" is a loop increment or 376 // if "y" is a loop induction variable. 377 if( op1 == Op_AddL && ok_to_convert(in1, in2) ) { 378 Node *in11 = in1->in(1); 379 const Type *tadd = phase->type( in1->in(2) ); 380 if( tadd->singleton() && tadd != Type::TOP ) { 381 Node *sub2 = phase->transform( new SubLNode( in11, in2 )); 382 return new AddLNode( sub2, in1->in(2) ); 383 } 384 } 385 386 // Convert "x - (y+c0)" into "(x-y) - c0" AND 387 // Convert "c1 - (y+c0)" into "(c1-c0) - y" 388 // Need the same check as in above optimization but reversed. 389 if (op2 == Op_AddL 390 && ok_to_convert(in2, in1) 391 && in2->in(2)->Opcode() == Op_ConL) { 392 jlong c0 = phase->type(in2->in(2))->isa_long()->get_con(); 393 Node* in21 = in2->in(1); 394 if (in1->Opcode() == Op_ConL) { 395 // Match c1 396 jlong c1 = phase->type(in1)->isa_long()->get_con(); 397 Node* sub2 = phase->longcon(java_subtract(c1, c0)); 398 return new SubLNode(sub2, in21); 399 } else { 400 Node* sub2 = phase->transform(new SubLNode(in1, in21)); 401 Node* neg_c0 = phase->longcon(-c0); 402 return new AddLNode(sub2, neg_c0); 403 } 404 } 405 406 const Type *t1 = phase->type( in1 ); 407 if( t1 == Type::TOP ) return nullptr; 408 409 #ifdef ASSERT 410 // Check for dead loop 411 if ((op2 == Op_AddL || op2 == Op_SubL) && 412 ((in2->in(1) == this) || (in2->in(2) == this) || 413 (in2->in(1) == in2) || (in2->in(2) == in2))) { 414 assert(false, "dead loop in SubLNode::Ideal"); 415 } 416 #endif 417 418 // Convert "x - (x+y)" into "-y" 419 if (op2 == Op_AddL && in1 == in2->in(1)) { 420 return new SubLNode(phase->makecon(TypeLong::ZERO), in2->in(2)); 421 } 422 // Convert "(x-y) - x" into "-y" 423 if (op1 == Op_SubL && in1->in(1) == in2) { 424 return new SubLNode(phase->makecon(TypeLong::ZERO), in1->in(2)); 425 } 426 // Convert "x - (y+x)" into "-y" 427 if (op2 == Op_AddL && in1 == in2->in(2)) { 428 return new SubLNode(phase->makecon(TypeLong::ZERO), in2->in(1)); 429 } 430 431 // Convert "0 - (x-y)" into "y-x", leave the double negation "-(-y)" to SubNode::Identity. 432 if (t1 == TypeLong::ZERO && op2 == Op_SubL && phase->type(in2->in(1)) != TypeLong::ZERO) { 433 return new SubLNode(in2->in(2), in2->in(1)); 434 } 435 436 // Convert "(X+A) - (X+B)" into "A - B" 437 if( op1 == Op_AddL && op2 == Op_AddL && in1->in(1) == in2->in(1) ) 438 return new SubLNode( in1->in(2), in2->in(2) ); 439 440 // Convert "(A+X) - (B+X)" into "A - B" 441 if( op1 == Op_AddL && op2 == Op_AddL && in1->in(2) == in2->in(2) ) 442 return new SubLNode( in1->in(1), in2->in(1) ); 443 444 // Convert "(A+X) - (X+B)" into "A - B" 445 if( op1 == Op_AddL && op2 == Op_AddL && in1->in(2) == in2->in(1) ) 446 return new SubLNode( in1->in(1), in2->in(2) ); 447 448 // Convert "(X+A) - (B+X)" into "A - B" 449 if( op1 == Op_AddL && op2 == Op_AddL && in1->in(1) == in2->in(2) ) 450 return new SubLNode( in1->in(2), in2->in(1) ); 451 452 // Convert "A-(B-C)" into (A+C)-B" 453 if( op2 == Op_SubL && in2->outcnt() == 1) { 454 Node *add1 = phase->transform( new AddLNode( in1, in2->in(2) ) ); 455 return new SubLNode( add1, in2->in(1) ); 456 } 457 458 // Associative 459 if (op1 == Op_MulL && op2 == Op_MulL) { 460 Node* sub_in1 = nullptr; 461 Node* sub_in2 = nullptr; 462 Node* mul_in = nullptr; 463 464 if (in1->in(1) == in2->in(1)) { 465 // Convert "a*b-a*c into a*(b+c) 466 sub_in1 = in1->in(2); 467 sub_in2 = in2->in(2); 468 mul_in = in1->in(1); 469 } else if (in1->in(2) == in2->in(1)) { 470 // Convert a*b-b*c into b*(a-c) 471 sub_in1 = in1->in(1); 472 sub_in2 = in2->in(2); 473 mul_in = in1->in(2); 474 } else if (in1->in(2) == in2->in(2)) { 475 // Convert a*c-b*c into (a-b)*c 476 sub_in1 = in1->in(1); 477 sub_in2 = in2->in(1); 478 mul_in = in1->in(2); 479 } else if (in1->in(1) == in2->in(2)) { 480 // Convert a*b-c*a into a*(b-c) 481 sub_in1 = in1->in(2); 482 sub_in2 = in2->in(1); 483 mul_in = in1->in(1); 484 } 485 486 if (mul_in != nullptr) { 487 Node* sub = phase->transform(new SubLNode(sub_in1, sub_in2)); 488 return new MulLNode(mul_in, sub); 489 } 490 } 491 492 // Convert "0L-(A>>63)" into "(A>>>63)" 493 if ( op2 == Op_RShiftL ) { 494 Node *in21 = in2->in(1); 495 Node *in22 = in2->in(2); 496 const TypeLong *zero = phase->type(in1)->isa_long(); 497 const TypeLong *t21 = phase->type(in21)->isa_long(); 498 const TypeInt *t22 = phase->type(in22)->isa_int(); 499 if ( t21 && t22 && zero == TypeLong::ZERO && t22->is_con(63) ) { 500 return new URShiftLNode(in21, in22); 501 } 502 } 503 504 return nullptr; 505 } 506 507 //------------------------------sub-------------------------------------------- 508 // A subtract node differences it's two inputs. 509 const Type *SubLNode::sub( const Type *t1, const Type *t2 ) const { 510 const TypeLong *r0 = t1->is_long(); // Handy access 511 const TypeLong *r1 = t2->is_long(); 512 jlong lo = java_subtract(r0->_lo, r1->_hi); 513 jlong hi = java_subtract(r0->_hi, r1->_lo); 514 515 // We next check for 32-bit overflow. 516 // If that happens, we just assume all integers are possible. 517 if( (((r0->_lo ^ r1->_hi) >= 0) || // lo ends have same signs OR 518 ((r0->_lo ^ lo) >= 0)) && // lo results have same signs AND 519 (((r0->_hi ^ r1->_lo) >= 0) || // hi ends have same signs OR 520 ((r0->_hi ^ hi) >= 0)) ) // hi results have same signs 521 return TypeLong::make(lo,hi,MAX2(r0->_widen,r1->_widen)); 522 else // Overflow; assume all integers 523 return TypeLong::LONG; 524 } 525 526 //============================================================================= 527 //------------------------------Value------------------------------------------ 528 // A subtract node differences its two inputs. 529 const Type* SubFPNode::Value(PhaseGVN* phase) const { 530 const Node* in1 = in(1); 531 const Node* in2 = in(2); 532 // Either input is TOP ==> the result is TOP 533 const Type* t1 = (in1 == this) ? Type::TOP : phase->type(in1); 534 if( t1 == Type::TOP ) return Type::TOP; 535 const Type* t2 = (in2 == this) ? Type::TOP : phase->type(in2); 536 if( t2 == Type::TOP ) return Type::TOP; 537 538 // if both operands are infinity of same sign, the result is NaN; do 539 // not replace with zero 540 if (t1->is_finite() && t2->is_finite() && in1 == in2) { 541 return add_id(); 542 } 543 544 // Either input is BOTTOM ==> the result is the local BOTTOM 545 const Type *bot = bottom_type(); 546 if( (t1 == bot) || (t2 == bot) || 547 (t1 == Type::BOTTOM) || (t2 == Type::BOTTOM) ) 548 return bot; 549 550 return sub(t1,t2); // Local flavor of type subtraction 551 } 552 553 554 //============================================================================= 555 //------------------------------sub-------------------------------------------- 556 // A subtract node differences its two inputs. 557 const Type* SubHFNode::sub(const Type* t1, const Type* t2) const { 558 // no folding if one of operands is infinity or NaN, do not do constant folding 559 if(g_isfinite(t1->getf()) && g_isfinite(t2->getf())) { 560 return TypeH::make(t1->getf() - t2->getf()); 561 } 562 else if(g_isnan(t1->getf())) { 563 return t1; 564 } 565 else if(g_isnan(t2->getf())) { 566 return t2; 567 } 568 else { 569 return Type::HALF_FLOAT; 570 } 571 } 572 573 //------------------------------Ideal------------------------------------------ 574 Node *SubFNode::Ideal(PhaseGVN *phase, bool can_reshape) { 575 const Type *t2 = phase->type( in(2) ); 576 // Convert "x-c0" into "x+ -c0". 577 if( t2->base() == Type::FloatCon ) { // Might be bottom or top... 578 // return new (phase->C, 3) AddFNode(in(1), phase->makecon( TypeF::make(-t2->getf()) ) ); 579 } 580 581 // Cannot replace 0.0-X with -X because a 'fsub' bytecode computes 582 // 0.0-0.0 as +0.0, while a 'fneg' bytecode computes -0.0. 583 //if( phase->type(in(1)) == TypeF::ZERO ) 584 //return new (phase->C, 2) NegFNode(in(2)); 585 586 return nullptr; 587 } 588 589 //------------------------------sub-------------------------------------------- 590 // A subtract node differences its two inputs. 591 const Type *SubFNode::sub( const Type *t1, const Type *t2 ) const { 592 // no folding if one of operands is infinity or NaN, do not do constant folding 593 if( g_isfinite(t1->getf()) && g_isfinite(t2->getf()) ) { 594 return TypeF::make( t1->getf() - t2->getf() ); 595 } 596 else if( g_isnan(t1->getf()) ) { 597 return t1; 598 } 599 else if( g_isnan(t2->getf()) ) { 600 return t2; 601 } 602 else { 603 return Type::FLOAT; 604 } 605 } 606 607 //============================================================================= 608 //------------------------------Ideal------------------------------------------ 609 Node *SubDNode::Ideal(PhaseGVN *phase, bool can_reshape){ 610 const Type *t2 = phase->type( in(2) ); 611 // Convert "x-c0" into "x+ -c0". 612 if( t2->base() == Type::DoubleCon ) { // Might be bottom or top... 613 // return new (phase->C, 3) AddDNode(in(1), phase->makecon( TypeD::make(-t2->getd()) ) ); 614 } 615 616 // Cannot replace 0.0-X with -X because a 'dsub' bytecode computes 617 // 0.0-0.0 as +0.0, while a 'dneg' bytecode computes -0.0. 618 //if( phase->type(in(1)) == TypeD::ZERO ) 619 //return new (phase->C, 2) NegDNode(in(2)); 620 621 return nullptr; 622 } 623 624 //------------------------------sub-------------------------------------------- 625 // A subtract node differences its two inputs. 626 const Type *SubDNode::sub( const Type *t1, const Type *t2 ) const { 627 // no folding if one of operands is infinity or NaN, do not do constant folding 628 if( g_isfinite(t1->getd()) && g_isfinite(t2->getd()) ) { 629 return TypeD::make( t1->getd() - t2->getd() ); 630 } 631 else if( g_isnan(t1->getd()) ) { 632 return t1; 633 } 634 else if( g_isnan(t2->getd()) ) { 635 return t2; 636 } 637 else { 638 return Type::DOUBLE; 639 } 640 } 641 642 //============================================================================= 643 //------------------------------Idealize--------------------------------------- 644 // Unlike SubNodes, compare must still flatten return value to the 645 // range -1, 0, 1. 646 // And optimizations like those for (X + Y) - X fail if overflow happens. 647 Node* CmpNode::Identity(PhaseGVN* phase) { 648 return this; 649 } 650 651 CmpNode *CmpNode::make(Node *in1, Node *in2, BasicType bt, bool unsigned_comp) { 652 switch (bt) { 653 case T_INT: 654 if (unsigned_comp) { 655 return new CmpUNode(in1, in2); 656 } 657 return new CmpINode(in1, in2); 658 case T_LONG: 659 if (unsigned_comp) { 660 return new CmpULNode(in1, in2); 661 } 662 return new CmpLNode(in1, in2); 663 case T_OBJECT: 664 case T_ARRAY: 665 case T_ADDRESS: 666 case T_METADATA: 667 return new CmpPNode(in1, in2); 668 case T_NARROWOOP: 669 case T_NARROWKLASS: 670 return new CmpNNode(in1, in2); 671 default: 672 fatal("Not implemented for %s", type2name(bt)); 673 } 674 return nullptr; 675 } 676 677 //============================================================================= 678 //------------------------------cmp-------------------------------------------- 679 // Simplify a CmpI (compare 2 integers) node, based on local information. 680 // If both inputs are constants, compare them. 681 const Type *CmpINode::sub( const Type *t1, const Type *t2 ) const { 682 const TypeInt *r0 = t1->is_int(); // Handy access 683 const TypeInt *r1 = t2->is_int(); 684 685 if( r0->_hi < r1->_lo ) // Range is always low? 686 return TypeInt::CC_LT; 687 else if( r0->_lo > r1->_hi ) // Range is always high? 688 return TypeInt::CC_GT; 689 690 else if( r0->is_con() && r1->is_con() ) { // comparing constants? 691 assert(r0->get_con() == r1->get_con(), "must be equal"); 692 return TypeInt::CC_EQ; // Equal results. 693 } else if( r0->_hi == r1->_lo ) // Range is never high? 694 return TypeInt::CC_LE; 695 else if( r0->_lo == r1->_hi ) // Range is never low? 696 return TypeInt::CC_GE; 697 return TypeInt::CC; // else use worst case results 698 } 699 700 const Type* CmpINode::Value(PhaseGVN* phase) const { 701 Node* in1 = in(1); 702 Node* in2 = in(2); 703 // If this test is the zero trip guard for a main or post loop, check whether, with the opaque node removed, the test 704 // would constant fold so the loop is never entered. If so return the type of the test without the opaque node removed: 705 // make the loop unreachable. 706 // The reason for this is that the iv phi captures the bounds of the loop and if the loop becomes unreachable, it can 707 // become top. In that case, the loop must be removed. 708 // This is safe because: 709 // - as optimizations proceed, the range of iterations executed by the main loop narrows. If no iterations remain, then 710 // we're done with optimizations for that loop. 711 // - the post loop is initially not reachable but as long as there's a main loop, the zero trip guard for the post 712 // 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 713 // loop is removed, there's no need to preserve the zero trip guard for the post loop anymore. 714 if (in1 != nullptr && in2 != nullptr) { 715 uint input = 0; 716 Node* cmp = nullptr; 717 BoolTest::mask test; 718 if (in1->Opcode() == Op_OpaqueZeroTripGuard && phase->type(in1) != Type::TOP) { 719 cmp = new CmpINode(in1->in(1), in2); 720 test = ((OpaqueZeroTripGuardNode*)in1)->_loop_entered_mask; 721 } 722 if (in2->Opcode() == Op_OpaqueZeroTripGuard && phase->type(in2) != Type::TOP) { 723 assert(cmp == nullptr, "A cmp with 2 OpaqueZeroTripGuard inputs"); 724 cmp = new CmpINode(in1, in2->in(1)); 725 test = ((OpaqueZeroTripGuardNode*)in2)->_loop_entered_mask; 726 } 727 if (cmp != nullptr) { 728 const Type* cmp_t = cmp->Value(phase); 729 const Type* t = BoolTest(test).cc2logical(cmp_t); 730 cmp->destruct(phase); 731 if (t == TypeInt::ZERO) { 732 return cmp_t; 733 } 734 } 735 } 736 737 return SubNode::Value(phase); 738 } 739 740 741 // Simplify a CmpU (compare 2 integers) node, based on local information. 742 // If both inputs are constants, compare them. 743 const Type *CmpUNode::sub( const Type *t1, const Type *t2 ) const { 744 assert(!t1->isa_ptr(), "obsolete usage of CmpU"); 745 746 // comparing two unsigned ints 747 const TypeInt *r0 = t1->is_int(); // Handy access 748 const TypeInt *r1 = t2->is_int(); 749 750 // Current installed version 751 // Compare ranges for non-overlap 752 juint lo0 = r0->_lo; 753 juint hi0 = r0->_hi; 754 juint lo1 = r1->_lo; 755 juint hi1 = r1->_hi; 756 757 // If either one has both negative and positive values, 758 // it therefore contains both 0 and -1, and since [0..-1] is the 759 // full unsigned range, the type must act as an unsigned bottom. 760 bool bot0 = ((jint)(lo0 ^ hi0) < 0); 761 bool bot1 = ((jint)(lo1 ^ hi1) < 0); 762 763 if (bot0 || bot1) { 764 // All unsigned values are LE -1 and GE 0. 765 if (lo0 == 0 && hi0 == 0) { 766 return TypeInt::CC_LE; // 0 <= bot 767 } else if ((jint)lo0 == -1 && (jint)hi0 == -1) { 768 return TypeInt::CC_GE; // -1 >= bot 769 } else if (lo1 == 0 && hi1 == 0) { 770 return TypeInt::CC_GE; // bot >= 0 771 } else if ((jint)lo1 == -1 && (jint)hi1 == -1) { 772 return TypeInt::CC_LE; // bot <= -1 773 } 774 } else { 775 // We can use ranges of the form [lo..hi] if signs are the same. 776 assert(lo0 <= hi0 && lo1 <= hi1, "unsigned ranges are valid"); 777 // results are reversed, '-' > '+' for unsigned compare 778 if (hi0 < lo1) { 779 return TypeInt::CC_LT; // smaller 780 } else if (lo0 > hi1) { 781 return TypeInt::CC_GT; // greater 782 } else if (hi0 == lo1 && lo0 == hi1) { 783 return TypeInt::CC_EQ; // Equal results 784 } else if (lo0 >= hi1) { 785 return TypeInt::CC_GE; 786 } else if (hi0 <= lo1) { 787 // Check for special case in Hashtable::get. (See below.) 788 if ((jint)lo0 >= 0 && (jint)lo1 >= 0 && is_index_range_check()) 789 return TypeInt::CC_LT; 790 return TypeInt::CC_LE; 791 } 792 } 793 // Check for special case in Hashtable::get - the hash index is 794 // mod'ed to the table size so the following range check is useless. 795 // Check for: (X Mod Y) CmpU Y, where the mod result and Y both have 796 // to be positive. 797 // (This is a gross hack, since the sub method never 798 // looks at the structure of the node in any other case.) 799 if ((jint)lo0 >= 0 && (jint)lo1 >= 0 && is_index_range_check()) 800 return TypeInt::CC_LT; 801 return TypeInt::CC; // else use worst case results 802 } 803 804 const Type* CmpUNode::Value(PhaseGVN* phase) const { 805 const Type* t = SubNode::Value_common(phase); 806 if (t != nullptr) { 807 return t; 808 } 809 const Node* in1 = in(1); 810 const Node* in2 = in(2); 811 const Type* t1 = phase->type(in1); 812 const Type* t2 = phase->type(in2); 813 assert(t1->isa_int(), "CmpU has only Int type inputs"); 814 if (t2 == TypeInt::INT) { // Compare to bottom? 815 return bottom_type(); 816 } 817 818 const Type* t_sub = sub(t1, t2); // compare based on immediate inputs 819 820 uint in1_op = in1->Opcode(); 821 if (in1_op == Op_AddI || in1_op == Op_SubI) { 822 // The problem rise when result of AddI(SubI) may overflow 823 // signed integer value. Let say the input type is 824 // [256, maxint] then +128 will create 2 ranges due to 825 // overflow: [minint, minint+127] and [384, maxint]. 826 // But C2 type system keep only 1 type range and as result 827 // it use general [minint, maxint] for this case which we 828 // can't optimize. 829 // 830 // Make 2 separate type ranges based on types of AddI(SubI) inputs 831 // and compare results of their compare. If results are the same 832 // CmpU node can be optimized. 833 const Node* in11 = in1->in(1); 834 const Node* in12 = in1->in(2); 835 const Type* t11 = (in11 == in1) ? Type::TOP : phase->type(in11); 836 const Type* t12 = (in12 == in1) ? Type::TOP : phase->type(in12); 837 // Skip cases when input types are top or bottom. 838 if ((t11 != Type::TOP) && (t11 != TypeInt::INT) && 839 (t12 != Type::TOP) && (t12 != TypeInt::INT)) { 840 const TypeInt *r0 = t11->is_int(); 841 const TypeInt *r1 = t12->is_int(); 842 jlong lo_r0 = r0->_lo; 843 jlong hi_r0 = r0->_hi; 844 jlong lo_r1 = r1->_lo; 845 jlong hi_r1 = r1->_hi; 846 if (in1_op == Op_SubI) { 847 jlong tmp = hi_r1; 848 hi_r1 = -lo_r1; 849 lo_r1 = -tmp; 850 // Note, for substructing [minint,x] type range 851 // long arithmetic provides correct overflow answer. 852 // The confusion come from the fact that in 32-bit 853 // -minint == minint but in 64-bit -minint == maxint+1. 854 } 855 jlong lo_long = lo_r0 + lo_r1; 856 jlong hi_long = hi_r0 + hi_r1; 857 int lo_tr1 = min_jint; 858 int hi_tr1 = (int)hi_long; 859 int lo_tr2 = (int)lo_long; 860 int hi_tr2 = max_jint; 861 bool underflow = lo_long != (jlong)lo_tr2; 862 bool overflow = hi_long != (jlong)hi_tr1; 863 // Use sub(t1, t2) when there is no overflow (one type range) 864 // or when both overflow and underflow (too complex). 865 if ((underflow != overflow) && (hi_tr1 < lo_tr2)) { 866 // Overflow only on one boundary, compare 2 separate type ranges. 867 int w = MAX2(r0->_widen, r1->_widen); // _widen does not matter here 868 const TypeInt* tr1 = TypeInt::make(lo_tr1, hi_tr1, w); 869 const TypeInt* tr2 = TypeInt::make(lo_tr2, hi_tr2, w); 870 const TypeInt* cmp1 = sub(tr1, t2)->is_int(); 871 const TypeInt* cmp2 = sub(tr2, t2)->is_int(); 872 // Compute union, so that cmp handles all possible results from the two cases 873 const Type* t_cmp = cmp1->meet(cmp2); 874 // Pick narrowest type, based on overflow computation and on immediate inputs 875 return t_sub->filter(t_cmp); 876 } 877 } 878 } 879 880 return t_sub; 881 } 882 883 bool CmpUNode::is_index_range_check() const { 884 // Check for the "(X ModI Y) CmpU Y" shape 885 return (in(1)->Opcode() == Op_ModI && 886 in(1)->in(2)->eqv_uncast(in(2))); 887 } 888 889 //------------------------------Idealize--------------------------------------- 890 Node *CmpINode::Ideal( PhaseGVN *phase, bool can_reshape ) { 891 if (phase->type(in(2))->higher_equal(TypeInt::ZERO)) { 892 switch (in(1)->Opcode()) { 893 case Op_CmpU3: // Collapse a CmpU3/CmpI into a CmpU 894 return new CmpUNode(in(1)->in(1),in(1)->in(2)); 895 case Op_CmpL3: // Collapse a CmpL3/CmpI into a CmpL 896 return new CmpLNode(in(1)->in(1),in(1)->in(2)); 897 case Op_CmpUL3: // Collapse a CmpUL3/CmpI into a CmpUL 898 return new CmpULNode(in(1)->in(1),in(1)->in(2)); 899 case Op_CmpF3: // Collapse a CmpF3/CmpI into a CmpF 900 return new CmpFNode(in(1)->in(1),in(1)->in(2)); 901 case Op_CmpD3: // Collapse a CmpD3/CmpI into a CmpD 902 return new CmpDNode(in(1)->in(1),in(1)->in(2)); 903 //case Op_SubI: 904 // If (x - y) cannot overflow, then ((x - y) <?> 0) 905 // can be turned into (x <?> y). 906 // This is handled (with more general cases) by Ideal_sub_algebra. 907 } 908 } 909 return nullptr; // No change 910 } 911 912 Node *CmpLNode::Ideal( PhaseGVN *phase, bool can_reshape ) { 913 const TypeLong *t2 = phase->type(in(2))->isa_long(); 914 if (Opcode() == Op_CmpL && in(1)->Opcode() == Op_ConvI2L && t2 && t2->is_con()) { 915 const jlong con = t2->get_con(); 916 if (con >= min_jint && con <= max_jint) { 917 return new CmpINode(in(1)->in(1), phase->intcon((jint)con)); 918 } 919 } 920 return nullptr; 921 } 922 923 //============================================================================= 924 // Simplify a CmpL (compare 2 longs ) node, based on local information. 925 // If both inputs are constants, compare them. 926 const Type *CmpLNode::sub( const Type *t1, const Type *t2 ) const { 927 const TypeLong *r0 = t1->is_long(); // Handy access 928 const TypeLong *r1 = t2->is_long(); 929 930 if( r0->_hi < r1->_lo ) // Range is always low? 931 return TypeInt::CC_LT; 932 else if( r0->_lo > r1->_hi ) // Range is always high? 933 return TypeInt::CC_GT; 934 935 else if( r0->is_con() && r1->is_con() ) { // comparing constants? 936 assert(r0->get_con() == r1->get_con(), "must be equal"); 937 return TypeInt::CC_EQ; // Equal results. 938 } else if( r0->_hi == r1->_lo ) // Range is never high? 939 return TypeInt::CC_LE; 940 else if( r0->_lo == r1->_hi ) // Range is never low? 941 return TypeInt::CC_GE; 942 return TypeInt::CC; // else use worst case results 943 } 944 945 946 // Simplify a CmpUL (compare 2 unsigned longs) node, based on local information. 947 // If both inputs are constants, compare them. 948 const Type* CmpULNode::sub(const Type* t1, const Type* t2) const { 949 assert(!t1->isa_ptr(), "obsolete usage of CmpUL"); 950 951 // comparing two unsigned longs 952 const TypeLong* r0 = t1->is_long(); // Handy access 953 const TypeLong* r1 = t2->is_long(); 954 955 // Current installed version 956 // Compare ranges for non-overlap 957 julong lo0 = r0->_lo; 958 julong hi0 = r0->_hi; 959 julong lo1 = r1->_lo; 960 julong hi1 = r1->_hi; 961 962 // If either one has both negative and positive values, 963 // it therefore contains both 0 and -1, and since [0..-1] is the 964 // full unsigned range, the type must act as an unsigned bottom. 965 bool bot0 = ((jlong)(lo0 ^ hi0) < 0); 966 bool bot1 = ((jlong)(lo1 ^ hi1) < 0); 967 968 if (bot0 || bot1) { 969 // All unsigned values are LE -1 and GE 0. 970 if (lo0 == 0 && hi0 == 0) { 971 return TypeInt::CC_LE; // 0 <= bot 972 } else if ((jlong)lo0 == -1 && (jlong)hi0 == -1) { 973 return TypeInt::CC_GE; // -1 >= bot 974 } else if (lo1 == 0 && hi1 == 0) { 975 return TypeInt::CC_GE; // bot >= 0 976 } else if ((jlong)lo1 == -1 && (jlong)hi1 == -1) { 977 return TypeInt::CC_LE; // bot <= -1 978 } 979 } else { 980 // We can use ranges of the form [lo..hi] if signs are the same. 981 assert(lo0 <= hi0 && lo1 <= hi1, "unsigned ranges are valid"); 982 // results are reversed, '-' > '+' for unsigned compare 983 if (hi0 < lo1) { 984 return TypeInt::CC_LT; // smaller 985 } else if (lo0 > hi1) { 986 return TypeInt::CC_GT; // greater 987 } else if (hi0 == lo1 && lo0 == hi1) { 988 return TypeInt::CC_EQ; // Equal results 989 } else if (lo0 >= hi1) { 990 return TypeInt::CC_GE; 991 } else if (hi0 <= lo1) { 992 return TypeInt::CC_LE; 993 } 994 } 995 996 return TypeInt::CC; // else use worst case results 997 } 998 999 //============================================================================= 1000 //------------------------------sub-------------------------------------------- 1001 // Simplify an CmpP (compare 2 pointers) node, based on local information. 1002 // If both inputs are constants, compare them. 1003 const Type *CmpPNode::sub( const Type *t1, const Type *t2 ) const { 1004 const TypePtr *r0 = t1->is_ptr(); // Handy access 1005 const TypePtr *r1 = t2->is_ptr(); 1006 1007 // Undefined inputs makes for an undefined result 1008 if( TypePtr::above_centerline(r0->_ptr) || 1009 TypePtr::above_centerline(r1->_ptr) ) 1010 return Type::TOP; 1011 1012 if (r0 == r1 && r0->singleton()) { 1013 // Equal pointer constants (klasses, nulls, etc.) 1014 return TypeInt::CC_EQ; 1015 } 1016 1017 // See if it is 2 unrelated classes. 1018 const TypeOopPtr* p0 = r0->isa_oopptr(); 1019 const TypeOopPtr* p1 = r1->isa_oopptr(); 1020 const TypeKlassPtr* k0 = r0->isa_klassptr(); 1021 const TypeKlassPtr* k1 = r1->isa_klassptr(); 1022 if ((p0 && p1) || (k0 && k1)) { 1023 if (p0 && p1) { 1024 Node* in1 = in(1)->uncast(); 1025 Node* in2 = in(2)->uncast(); 1026 AllocateNode* alloc1 = AllocateNode::Ideal_allocation(in1); 1027 AllocateNode* alloc2 = AllocateNode::Ideal_allocation(in2); 1028 if (MemNode::detect_ptr_independence(in1, alloc1, in2, alloc2, nullptr)) { 1029 return TypeInt::CC_GT; // different pointers 1030 } 1031 } 1032 bool xklass0 = p0 ? p0->klass_is_exact() : k0->klass_is_exact(); 1033 bool xklass1 = p1 ? p1->klass_is_exact() : k1->klass_is_exact(); 1034 bool unrelated_classes = false; 1035 1036 if ((p0 && p0->is_same_java_type_as(p1)) || 1037 (k0 && k0->is_same_java_type_as(k1))) { 1038 } else if ((p0 && !p1->maybe_java_subtype_of(p0) && !p0->maybe_java_subtype_of(p1)) || 1039 (k0 && !k1->maybe_java_subtype_of(k0) && !k0->maybe_java_subtype_of(k1))) { 1040 unrelated_classes = true; 1041 } else if ((p0 && !p1->maybe_java_subtype_of(p0)) || 1042 (k0 && !k1->maybe_java_subtype_of(k0))) { 1043 unrelated_classes = xklass1; 1044 } else if ((p0 && !p0->maybe_java_subtype_of(p1)) || 1045 (k0 && !k0->maybe_java_subtype_of(k1))) { 1046 unrelated_classes = xklass0; 1047 } 1048 1049 if (unrelated_classes) { 1050 // The oops classes are known to be unrelated. If the joined PTRs of 1051 // two oops is not Null and not Bottom, then we are sure that one 1052 // of the two oops is non-null, and the comparison will always fail. 1053 TypePtr::PTR jp = r0->join_ptr(r1->_ptr); 1054 if (jp != TypePtr::Null && jp != TypePtr::BotPTR) { 1055 return TypeInt::CC_GT; 1056 } 1057 } 1058 } 1059 1060 // Known constants can be compared exactly 1061 // Null can be distinguished from any NotNull pointers 1062 // Unknown inputs makes an unknown result 1063 if( r0->singleton() ) { 1064 intptr_t bits0 = r0->get_con(); 1065 if( r1->singleton() ) 1066 return bits0 == r1->get_con() ? TypeInt::CC_EQ : TypeInt::CC_GT; 1067 return ( r1->_ptr == TypePtr::NotNull && bits0==0 ) ? TypeInt::CC_GT : TypeInt::CC; 1068 } else if( r1->singleton() ) { 1069 intptr_t bits1 = r1->get_con(); 1070 return ( r0->_ptr == TypePtr::NotNull && bits1==0 ) ? TypeInt::CC_GT : TypeInt::CC; 1071 } else 1072 return TypeInt::CC; 1073 } 1074 1075 static inline Node* isa_java_mirror_load(PhaseGVN* phase, Node* n) { 1076 // Return the klass node for (indirect load from OopHandle) 1077 // LoadBarrier?(LoadP(LoadP(AddP(foo:Klass, #java_mirror)))) 1078 // or null if not matching. 1079 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 1080 n = bs->step_over_gc_barrier(n); 1081 1082 if (n->Opcode() != Op_LoadP) return nullptr; 1083 1084 const TypeInstPtr* tp = phase->type(n)->isa_instptr(); 1085 if (!tp || tp->instance_klass() != phase->C->env()->Class_klass()) return nullptr; 1086 1087 Node* adr = n->in(MemNode::Address); 1088 // First load from OopHandle: ((OopHandle)mirror)->resolve(); may need barrier. 1089 if (adr->Opcode() != Op_LoadP || !phase->type(adr)->isa_rawptr()) return nullptr; 1090 adr = adr->in(MemNode::Address); 1091 1092 intptr_t off = 0; 1093 Node* k = AddPNode::Ideal_base_and_offset(adr, phase, off); 1094 if (k == nullptr) return nullptr; 1095 const TypeKlassPtr* tkp = phase->type(k)->isa_klassptr(); 1096 if (!tkp || off != in_bytes(Klass::java_mirror_offset())) return nullptr; 1097 1098 // We've found the klass node of a Java mirror load. 1099 return k; 1100 } 1101 1102 static inline Node* isa_const_java_mirror(PhaseGVN* phase, Node* n) { 1103 // for ConP(Foo.class) return ConP(Foo.klass) 1104 // otherwise return null 1105 if (!n->is_Con()) return nullptr; 1106 1107 const TypeInstPtr* tp = phase->type(n)->isa_instptr(); 1108 if (!tp) return nullptr; 1109 1110 ciType* mirror_type = tp->java_mirror_type(); 1111 // TypeInstPtr::java_mirror_type() returns non-null for compile- 1112 // time Class constants only. 1113 if (!mirror_type) return nullptr; 1114 1115 // x.getClass() == int.class can never be true (for all primitive types) 1116 // Return a ConP(null) node for this case. 1117 if (mirror_type->is_classless()) { 1118 return phase->makecon(TypePtr::NULL_PTR); 1119 } 1120 1121 // return the ConP(Foo.klass) 1122 assert(mirror_type->is_klass(), "mirror_type should represent a Klass*"); 1123 return phase->makecon(TypeKlassPtr::make(mirror_type->as_klass(), Type::trust_interfaces)); 1124 } 1125 1126 //------------------------------Ideal------------------------------------------ 1127 // Normalize comparisons between Java mirror loads to compare the klass instead. 1128 // 1129 // Also check for the case of comparing an unknown klass loaded from the primary 1130 // super-type array vs a known klass with no subtypes. This amounts to 1131 // checking to see an unknown klass subtypes a known klass with no subtypes; 1132 // this only happens on an exact match. We can shorten this test by 1 load. 1133 Node *CmpPNode::Ideal( PhaseGVN *phase, bool can_reshape ) { 1134 // Normalize comparisons between Java mirrors into comparisons of the low- 1135 // level klass, where a dependent load could be shortened. 1136 // 1137 // The new pattern has a nice effect of matching the same pattern used in the 1138 // fast path of instanceof/checkcast/Class.isInstance(), which allows 1139 // redundant exact type check be optimized away by GVN. 1140 // For example, in 1141 // if (x.getClass() == Foo.class) { 1142 // Foo foo = (Foo) x; 1143 // // ... use a ... 1144 // } 1145 // a CmpPNode could be shared between if_acmpne and checkcast 1146 { 1147 Node* k1 = isa_java_mirror_load(phase, in(1)); 1148 Node* k2 = isa_java_mirror_load(phase, in(2)); 1149 Node* conk2 = isa_const_java_mirror(phase, in(2)); 1150 1151 if (k1 && (k2 || conk2)) { 1152 Node* lhs = k1; 1153 Node* rhs = (k2 != nullptr) ? k2 : conk2; 1154 set_req_X(1, lhs, phase); 1155 set_req_X(2, rhs, phase); 1156 return this; 1157 } 1158 } 1159 1160 // Constant pointer on right? 1161 const TypeKlassPtr* t2 = phase->type(in(2))->isa_klassptr(); 1162 if (t2 == nullptr || !t2->klass_is_exact()) 1163 return nullptr; 1164 // Get the constant klass we are comparing to. 1165 ciKlass* superklass = t2->exact_klass(); 1166 1167 // Now check for LoadKlass on left. 1168 Node* ldk1 = in(1); 1169 if (ldk1->is_DecodeNKlass()) { 1170 ldk1 = ldk1->in(1); 1171 if (ldk1->Opcode() != Op_LoadNKlass ) 1172 return nullptr; 1173 } else if (ldk1->Opcode() != Op_LoadKlass ) 1174 return nullptr; 1175 // Take apart the address of the LoadKlass: 1176 Node* adr1 = ldk1->in(MemNode::Address); 1177 intptr_t con2 = 0; 1178 Node* ldk2 = AddPNode::Ideal_base_and_offset(adr1, phase, con2); 1179 if (ldk2 == nullptr) 1180 return nullptr; 1181 if (con2 == oopDesc::klass_offset_in_bytes()) { 1182 // We are inspecting an object's concrete class. 1183 // Short-circuit the check if the query is abstract. 1184 if (superklass->is_interface() || 1185 superklass->is_abstract()) { 1186 // Make it come out always false: 1187 this->set_req(2, phase->makecon(TypePtr::NULL_PTR)); 1188 return this; 1189 } 1190 } 1191 1192 // Check for a LoadKlass from primary supertype array. 1193 // Any nested loadklass from loadklass+con must be from the p.s. array. 1194 if (ldk2->is_DecodeNKlass()) { 1195 // Keep ldk2 as DecodeN since it could be used in CmpP below. 1196 if (ldk2->in(1)->Opcode() != Op_LoadNKlass ) 1197 return nullptr; 1198 } else if (ldk2->Opcode() != Op_LoadKlass) 1199 return nullptr; 1200 1201 // Verify that we understand the situation 1202 if (con2 != (intptr_t) superklass->super_check_offset()) 1203 return nullptr; // Might be element-klass loading from array klass 1204 1205 // If 'superklass' has no subklasses and is not an interface, then we are 1206 // assured that the only input which will pass the type check is 1207 // 'superklass' itself. 1208 // 1209 // We could be more liberal here, and allow the optimization on interfaces 1210 // which have a single implementor. This would require us to increase the 1211 // expressiveness of the add_dependency() mechanism. 1212 // %%% Do this after we fix TypeOopPtr: Deps are expressive enough now. 1213 1214 // Object arrays must have their base element have no subtypes 1215 while (superklass->is_obj_array_klass()) { 1216 ciType* elem = superklass->as_obj_array_klass()->element_type(); 1217 superklass = elem->as_klass(); 1218 } 1219 if (superklass->is_instance_klass()) { 1220 ciInstanceKlass* ik = superklass->as_instance_klass(); 1221 if (ik->has_subklass() || ik->is_interface()) return nullptr; 1222 // Add a dependency if there is a chance that a subclass will be added later. 1223 if (!ik->is_final()) { 1224 phase->C->dependencies()->assert_leaf_type(ik); 1225 } 1226 } 1227 1228 // Bypass the dependent load, and compare directly 1229 this->set_req_X(1, ldk2, phase); 1230 1231 return this; 1232 } 1233 1234 //============================================================================= 1235 //------------------------------sub-------------------------------------------- 1236 // Simplify an CmpN (compare 2 pointers) node, based on local information. 1237 // If both inputs are constants, compare them. 1238 const Type *CmpNNode::sub( const Type *t1, const Type *t2 ) const { 1239 ShouldNotReachHere(); 1240 return bottom_type(); 1241 } 1242 1243 //------------------------------Ideal------------------------------------------ 1244 Node *CmpNNode::Ideal( PhaseGVN *phase, bool can_reshape ) { 1245 return nullptr; 1246 } 1247 1248 //============================================================================= 1249 //------------------------------Value------------------------------------------ 1250 // Simplify an CmpF (compare 2 floats ) node, based on local information. 1251 // If both inputs are constants, compare them. 1252 const Type* CmpFNode::Value(PhaseGVN* phase) const { 1253 const Node* in1 = in(1); 1254 const Node* in2 = in(2); 1255 // Either input is TOP ==> the result is TOP 1256 const Type* t1 = (in1 == this) ? Type::TOP : phase->type(in1); 1257 if( t1 == Type::TOP ) return Type::TOP; 1258 const Type* t2 = (in2 == this) ? Type::TOP : phase->type(in2); 1259 if( t2 == Type::TOP ) return Type::TOP; 1260 1261 // Not constants? Don't know squat - even if they are the same 1262 // value! If they are NaN's they compare to LT instead of EQ. 1263 const TypeF *tf1 = t1->isa_float_constant(); 1264 const TypeF *tf2 = t2->isa_float_constant(); 1265 if( !tf1 || !tf2 ) return TypeInt::CC; 1266 1267 // This implements the Java bytecode fcmpl, so unordered returns -1. 1268 if( tf1->is_nan() || tf2->is_nan() ) 1269 return TypeInt::CC_LT; 1270 1271 if( tf1->_f < tf2->_f ) return TypeInt::CC_LT; 1272 if( tf1->_f > tf2->_f ) return TypeInt::CC_GT; 1273 assert( tf1->_f == tf2->_f, "do not understand FP behavior" ); 1274 return TypeInt::CC_EQ; 1275 } 1276 1277 1278 //============================================================================= 1279 //------------------------------Value------------------------------------------ 1280 // Simplify an CmpD (compare 2 doubles ) node, based on local information. 1281 // If both inputs are constants, compare them. 1282 const Type* CmpDNode::Value(PhaseGVN* phase) const { 1283 const Node* in1 = in(1); 1284 const Node* in2 = in(2); 1285 // Either input is TOP ==> the result is TOP 1286 const Type* t1 = (in1 == this) ? Type::TOP : phase->type(in1); 1287 if( t1 == Type::TOP ) return Type::TOP; 1288 const Type* t2 = (in2 == this) ? Type::TOP : phase->type(in2); 1289 if( t2 == Type::TOP ) return Type::TOP; 1290 1291 // Not constants? Don't know squat - even if they are the same 1292 // value! If they are NaN's they compare to LT instead of EQ. 1293 const TypeD *td1 = t1->isa_double_constant(); 1294 const TypeD *td2 = t2->isa_double_constant(); 1295 if( !td1 || !td2 ) return TypeInt::CC; 1296 1297 // This implements the Java bytecode dcmpl, so unordered returns -1. 1298 if( td1->is_nan() || td2->is_nan() ) 1299 return TypeInt::CC_LT; 1300 1301 if( td1->_d < td2->_d ) return TypeInt::CC_LT; 1302 if( td1->_d > td2->_d ) return TypeInt::CC_GT; 1303 assert( td1->_d == td2->_d, "do not understand FP behavior" ); 1304 return TypeInt::CC_EQ; 1305 } 1306 1307 //------------------------------Ideal------------------------------------------ 1308 Node *CmpDNode::Ideal(PhaseGVN *phase, bool can_reshape){ 1309 // Check if we can change this to a CmpF and remove a ConvD2F operation. 1310 // Change (CMPD (F2D (float)) (ConD value)) 1311 // To (CMPF (float) (ConF value)) 1312 // Valid when 'value' does not lose precision as a float. 1313 // Benefits: eliminates conversion, does not require 24-bit mode 1314 1315 // NaNs prevent commuting operands. This transform works regardless of the 1316 // order of ConD and ConvF2D inputs by preserving the original order. 1317 int idx_f2d = 1; // ConvF2D on left side? 1318 if( in(idx_f2d)->Opcode() != Op_ConvF2D ) 1319 idx_f2d = 2; // No, swap to check for reversed args 1320 int idx_con = 3-idx_f2d; // Check for the constant on other input 1321 1322 if( ConvertCmpD2CmpF && 1323 in(idx_f2d)->Opcode() == Op_ConvF2D && 1324 in(idx_con)->Opcode() == Op_ConD ) { 1325 const TypeD *t2 = in(idx_con)->bottom_type()->is_double_constant(); 1326 double t2_value_as_double = t2->_d; 1327 float t2_value_as_float = (float)t2_value_as_double; 1328 if( t2_value_as_double == (double)t2_value_as_float ) { 1329 // Test value can be represented as a float 1330 // Eliminate the conversion to double and create new comparison 1331 Node *new_in1 = in(idx_f2d)->in(1); 1332 Node *new_in2 = phase->makecon( TypeF::make(t2_value_as_float) ); 1333 if( idx_f2d != 1 ) { // Must flip args to match original order 1334 Node *tmp = new_in1; 1335 new_in1 = new_in2; 1336 new_in2 = tmp; 1337 } 1338 CmpFNode *new_cmp = (Opcode() == Op_CmpD3) 1339 ? new CmpF3Node( new_in1, new_in2 ) 1340 : new CmpFNode ( new_in1, new_in2 ) ; 1341 return new_cmp; // Changed to CmpFNode 1342 } 1343 // Testing value required the precision of a double 1344 } 1345 return nullptr; // No change 1346 } 1347 1348 1349 //============================================================================= 1350 //------------------------------cc2logical------------------------------------- 1351 // Convert a condition code type to a logical type 1352 const Type *BoolTest::cc2logical( const Type *CC ) const { 1353 if( CC == Type::TOP ) return Type::TOP; 1354 if( CC->base() != Type::Int ) return TypeInt::BOOL; // Bottom or worse 1355 const TypeInt *ti = CC->is_int(); 1356 if( ti->is_con() ) { // Only 1 kind of condition codes set? 1357 // Match low order 2 bits 1358 int tmp = ((ti->get_con()&3) == (_test&3)) ? 1 : 0; 1359 if( _test & 4 ) tmp = 1-tmp; // Optionally complement result 1360 return TypeInt::make(tmp); // Boolean result 1361 } 1362 1363 if( CC == TypeInt::CC_GE ) { 1364 if( _test == ge ) return TypeInt::ONE; 1365 if( _test == lt ) return TypeInt::ZERO; 1366 } 1367 if( CC == TypeInt::CC_LE ) { 1368 if( _test == le ) return TypeInt::ONE; 1369 if( _test == gt ) return TypeInt::ZERO; 1370 } 1371 1372 return TypeInt::BOOL; 1373 } 1374 1375 //------------------------------dump_spec------------------------------------- 1376 // Print special per-node info 1377 void BoolTest::dump_on(outputStream *st) const { 1378 const char *msg[] = {"eq","gt","of","lt","ne","le","nof","ge"}; 1379 st->print("%s", msg[_test]); 1380 } 1381 1382 // Returns the logical AND of two tests (or 'never' if both tests can never be true). 1383 // For example, a test for 'le' followed by a test for 'lt' is equivalent with 'lt'. 1384 BoolTest::mask BoolTest::merge(BoolTest other) const { 1385 const mask res[illegal+1][illegal+1] = { 1386 // eq, gt, of, lt, ne, le, nof, ge, never, illegal 1387 {eq, never, illegal, never, never, eq, illegal, eq, never, illegal}, // eq 1388 {never, gt, illegal, never, gt, never, illegal, gt, never, illegal}, // gt 1389 {illegal, illegal, illegal, illegal, illegal, illegal, illegal, illegal, never, illegal}, // of 1390 {never, never, illegal, lt, lt, lt, illegal, never, never, illegal}, // lt 1391 {never, gt, illegal, lt, ne, lt, illegal, gt, never, illegal}, // ne 1392 {eq, never, illegal, lt, lt, le, illegal, eq, never, illegal}, // le 1393 {illegal, illegal, illegal, illegal, illegal, illegal, illegal, illegal, never, illegal}, // nof 1394 {eq, gt, illegal, never, gt, eq, illegal, ge, never, illegal}, // ge 1395 {never, never, never, never, never, never, never, never, never, illegal}, // never 1396 {illegal, illegal, illegal, illegal, illegal, illegal, illegal, illegal, illegal, illegal}}; // illegal 1397 return res[_test][other._test]; 1398 } 1399 1400 //============================================================================= 1401 uint BoolNode::hash() const { return (Node::hash() << 3)|(_test._test+1); } 1402 uint BoolNode::size_of() const { return sizeof(BoolNode); } 1403 1404 //------------------------------operator==------------------------------------- 1405 bool BoolNode::cmp( const Node &n ) const { 1406 const BoolNode *b = (const BoolNode *)&n; // Cast up 1407 return (_test._test == b->_test._test); 1408 } 1409 1410 //-------------------------------make_predicate-------------------------------- 1411 Node* BoolNode::make_predicate(Node* test_value, PhaseGVN* phase) { 1412 if (test_value->is_Con()) return test_value; 1413 if (test_value->is_Bool()) return test_value; 1414 if (test_value->is_CMove() && 1415 test_value->in(CMoveNode::Condition)->is_Bool()) { 1416 BoolNode* bol = test_value->in(CMoveNode::Condition)->as_Bool(); 1417 const Type* ftype = phase->type(test_value->in(CMoveNode::IfFalse)); 1418 const Type* ttype = phase->type(test_value->in(CMoveNode::IfTrue)); 1419 if (ftype == TypeInt::ZERO && !TypeInt::ZERO->higher_equal(ttype)) { 1420 return bol; 1421 } else if (ttype == TypeInt::ZERO && !TypeInt::ZERO->higher_equal(ftype)) { 1422 return phase->transform( bol->negate(phase) ); 1423 } 1424 // Else fall through. The CMove gets in the way of the test. 1425 // It should be the case that make_predicate(bol->as_int_value()) == bol. 1426 } 1427 Node* cmp = new CmpINode(test_value, phase->intcon(0)); 1428 cmp = phase->transform(cmp); 1429 Node* bol = new BoolNode(cmp, BoolTest::ne); 1430 return phase->transform(bol); 1431 } 1432 1433 //--------------------------------as_int_value--------------------------------- 1434 Node* BoolNode::as_int_value(PhaseGVN* phase) { 1435 // Inverse to make_predicate. The CMove probably boils down to a Conv2B. 1436 Node* cmov = CMoveNode::make(this, phase->intcon(0), phase->intcon(1), TypeInt::BOOL); 1437 return phase->transform(cmov); 1438 } 1439 1440 //----------------------------------negate------------------------------------- 1441 BoolNode* BoolNode::negate(PhaseGVN* phase) { 1442 return new BoolNode(in(1), _test.negate()); 1443 } 1444 1445 // Change "bool eq/ne (cmp (add/sub A B) C)" into false/true if add/sub 1446 // overflows and we can prove that C is not in the two resulting ranges. 1447 // This optimization is similar to the one performed by CmpUNode::Value(). 1448 Node* BoolNode::fold_cmpI(PhaseGVN* phase, SubNode* cmp, Node* cmp1, int cmp_op, 1449 int cmp1_op, const TypeInt* cmp2_type) { 1450 // Only optimize eq/ne integer comparison of add/sub 1451 if((_test._test == BoolTest::eq || _test._test == BoolTest::ne) && 1452 (cmp_op == Op_CmpI) && (cmp1_op == Op_AddI || cmp1_op == Op_SubI)) { 1453 // Skip cases were inputs of add/sub are not integers or of bottom type 1454 const TypeInt* r0 = phase->type(cmp1->in(1))->isa_int(); 1455 const TypeInt* r1 = phase->type(cmp1->in(2))->isa_int(); 1456 if ((r0 != nullptr) && (r0 != TypeInt::INT) && 1457 (r1 != nullptr) && (r1 != TypeInt::INT) && 1458 (cmp2_type != TypeInt::INT)) { 1459 // Compute exact (long) type range of add/sub result 1460 jlong lo_long = r0->_lo; 1461 jlong hi_long = r0->_hi; 1462 if (cmp1_op == Op_AddI) { 1463 lo_long += r1->_lo; 1464 hi_long += r1->_hi; 1465 } else { 1466 lo_long -= r1->_hi; 1467 hi_long -= r1->_lo; 1468 } 1469 // Check for over-/underflow by casting to integer 1470 int lo_int = (int)lo_long; 1471 int hi_int = (int)hi_long; 1472 bool underflow = lo_long != (jlong)lo_int; 1473 bool overflow = hi_long != (jlong)hi_int; 1474 if ((underflow != overflow) && (hi_int < lo_int)) { 1475 // Overflow on one boundary, compute resulting type ranges: 1476 // tr1 [MIN_INT, hi_int] and tr2 [lo_int, MAX_INT] 1477 int w = MAX2(r0->_widen, r1->_widen); // _widen does not matter here 1478 const TypeInt* tr1 = TypeInt::make(min_jint, hi_int, w); 1479 const TypeInt* tr2 = TypeInt::make(lo_int, max_jint, w); 1480 // Compare second input of cmp to both type ranges 1481 const Type* sub_tr1 = cmp->sub(tr1, cmp2_type); 1482 const Type* sub_tr2 = cmp->sub(tr2, cmp2_type); 1483 if (sub_tr1 == TypeInt::CC_LT && sub_tr2 == TypeInt::CC_GT) { 1484 // The result of the add/sub will never equal cmp2. Replace BoolNode 1485 // by false (0) if it tests for equality and by true (1) otherwise. 1486 return ConINode::make((_test._test == BoolTest::eq) ? 0 : 1); 1487 } 1488 } 1489 } 1490 } 1491 return nullptr; 1492 } 1493 1494 static bool is_counted_loop_cmp(Node *cmp) { 1495 Node *n = cmp->in(1)->in(1); 1496 return n != nullptr && 1497 n->is_Phi() && 1498 n->in(0) != nullptr && 1499 n->in(0)->is_CountedLoop() && 1500 n->in(0)->as_CountedLoop()->phi() == n; 1501 } 1502 1503 //------------------------------Ideal------------------------------------------ 1504 Node *BoolNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1505 // Change "bool tst (cmp con x)" into "bool ~tst (cmp x con)". 1506 // This moves the constant to the right. Helps value-numbering. 1507 Node *cmp = in(1); 1508 if( !cmp->is_Sub() ) return nullptr; 1509 int cop = cmp->Opcode(); 1510 if( cop == Op_FastLock || cop == Op_FastUnlock || 1511 cmp->is_SubTypeCheck() || cop == Op_VectorTest ) { 1512 return nullptr; 1513 } 1514 Node *cmp1 = cmp->in(1); 1515 Node *cmp2 = cmp->in(2); 1516 if( !cmp1 ) return nullptr; 1517 1518 if (_test._test == BoolTest::overflow || _test._test == BoolTest::no_overflow) { 1519 return nullptr; 1520 } 1521 1522 const int cmp1_op = cmp1->Opcode(); 1523 const int cmp2_op = cmp2->Opcode(); 1524 1525 // Constant on left? 1526 Node *con = cmp1; 1527 // Move constants to the right of compare's to canonicalize. 1528 // Do not muck with Opaque1 nodes, as this indicates a loop 1529 // guard that cannot change shape. 1530 if (con->is_Con() && !cmp2->is_Con() && cmp2_op != Op_OpaqueZeroTripGuard && 1531 // Because of NaN's, CmpD and CmpF are not commutative 1532 cop != Op_CmpD && cop != Op_CmpF && 1533 // Protect against swapping inputs to a compare when it is used by a 1534 // counted loop exit, which requires maintaining the loop-limit as in(2) 1535 !is_counted_loop_exit_test() ) { 1536 // Ok, commute the constant to the right of the cmp node. 1537 // Clone the Node, getting a new Node of the same class 1538 cmp = cmp->clone(); 1539 // Swap inputs to the clone 1540 cmp->swap_edges(1, 2); 1541 cmp = phase->transform( cmp ); 1542 return new BoolNode( cmp, _test.commute() ); 1543 } 1544 1545 // Change "bool eq/ne (cmp (cmove (bool tst (cmp2)) 1 0) 0)" into "bool tst/~tst (cmp2)" 1546 if (cop == Op_CmpI && 1547 (_test._test == BoolTest::eq || _test._test == BoolTest::ne) && 1548 cmp1_op == Op_CMoveI && cmp2->find_int_con(1) == 0) { 1549 // 0 should be on the true branch 1550 if (cmp1->in(CMoveNode::Condition)->is_Bool() && 1551 cmp1->in(CMoveNode::IfTrue)->find_int_con(1) == 0 && 1552 cmp1->in(CMoveNode::IfFalse)->find_int_con(0) != 0) { 1553 BoolNode* target = cmp1->in(CMoveNode::Condition)->as_Bool(); 1554 return new BoolNode(target->in(1), 1555 (_test._test == BoolTest::eq) ? target->_test._test : 1556 target->_test.negate()); 1557 } 1558 } 1559 1560 // Change "bool eq/ne (cmp (and X 16) 16)" into "bool ne/eq (cmp (and X 16) 0)". 1561 if (cop == Op_CmpI && 1562 (_test._test == BoolTest::eq || _test._test == BoolTest::ne) && 1563 cmp1_op == Op_AndI && cmp2_op == Op_ConI && 1564 cmp1->in(2)->Opcode() == Op_ConI) { 1565 const TypeInt *t12 = phase->type(cmp2)->isa_int(); 1566 const TypeInt *t112 = phase->type(cmp1->in(2))->isa_int(); 1567 if (t12 && t12->is_con() && t112 && t112->is_con() && 1568 t12->get_con() == t112->get_con() && is_power_of_2(t12->get_con())) { 1569 Node *ncmp = phase->transform(new CmpINode(cmp1, phase->intcon(0))); 1570 return new BoolNode(ncmp, _test.negate()); 1571 } 1572 } 1573 1574 // Same for long type: change "bool eq/ne (cmp (and X 16) 16)" into "bool ne/eq (cmp (and X 16) 0)". 1575 if (cop == Op_CmpL && 1576 (_test._test == BoolTest::eq || _test._test == BoolTest::ne) && 1577 cmp1_op == Op_AndL && cmp2_op == Op_ConL && 1578 cmp1->in(2)->Opcode() == Op_ConL) { 1579 const TypeLong *t12 = phase->type(cmp2)->isa_long(); 1580 const TypeLong *t112 = phase->type(cmp1->in(2))->isa_long(); 1581 if (t12 && t12->is_con() && t112 && t112->is_con() && 1582 t12->get_con() == t112->get_con() && is_power_of_2(t12->get_con())) { 1583 Node *ncmp = phase->transform(new CmpLNode(cmp1, phase->longcon(0))); 1584 return new BoolNode(ncmp, _test.negate()); 1585 } 1586 } 1587 1588 // Change "cmp (add X min_jint) (add Y min_jint)" into "cmpu X Y" 1589 // and "cmp (add X min_jint) c" into "cmpu X (c + min_jint)" 1590 if (cop == Op_CmpI && 1591 cmp1_op == Op_AddI && 1592 phase->type(cmp1->in(2)) == TypeInt::MIN && 1593 !is_cloop_condition(this)) { 1594 if (cmp2_op == Op_ConI) { 1595 Node* ncmp2 = phase->intcon(java_add(cmp2->get_int(), min_jint)); 1596 Node* ncmp = phase->transform(new CmpUNode(cmp1->in(1), ncmp2)); 1597 return new BoolNode(ncmp, _test._test); 1598 } else if (cmp2_op == Op_AddI && 1599 phase->type(cmp2->in(2)) == TypeInt::MIN && 1600 !is_cloop_condition(this)) { 1601 Node* ncmp = phase->transform(new CmpUNode(cmp1->in(1), cmp2->in(1))); 1602 return new BoolNode(ncmp, _test._test); 1603 } 1604 } 1605 1606 // Change "cmp (add X min_jlong) (add Y min_jlong)" into "cmpu X Y" 1607 // and "cmp (add X min_jlong) c" into "cmpu X (c + min_jlong)" 1608 if (cop == Op_CmpL && 1609 cmp1_op == Op_AddL && 1610 phase->type(cmp1->in(2)) == TypeLong::MIN && 1611 !is_cloop_condition(this)) { 1612 if (cmp2_op == Op_ConL) { 1613 Node* ncmp2 = phase->longcon(java_add(cmp2->get_long(), min_jlong)); 1614 Node* ncmp = phase->transform(new CmpULNode(cmp1->in(1), ncmp2)); 1615 return new BoolNode(ncmp, _test._test); 1616 } else if (cmp2_op == Op_AddL && 1617 phase->type(cmp2->in(2)) == TypeLong::MIN && 1618 !is_cloop_condition(this)) { 1619 Node* ncmp = phase->transform(new CmpULNode(cmp1->in(1), cmp2->in(1))); 1620 return new BoolNode(ncmp, _test._test); 1621 } 1622 } 1623 1624 // Change "bool eq/ne (cmp (xor X 1) 0)" into "bool ne/eq (cmp X 0)". 1625 // The XOR-1 is an idiom used to flip the sense of a bool. We flip the 1626 // test instead. 1627 const TypeInt* cmp2_type = phase->type(cmp2)->isa_int(); 1628 if (cmp2_type == nullptr) return nullptr; 1629 Node* j_xor = cmp1; 1630 if( cmp2_type == TypeInt::ZERO && 1631 cmp1_op == Op_XorI && 1632 j_xor->in(1) != j_xor && // An xor of itself is dead 1633 phase->type( j_xor->in(1) ) == TypeInt::BOOL && 1634 phase->type( j_xor->in(2) ) == TypeInt::ONE && 1635 (_test._test == BoolTest::eq || 1636 _test._test == BoolTest::ne) ) { 1637 Node *ncmp = phase->transform(new CmpINode(j_xor->in(1),cmp2)); 1638 return new BoolNode( ncmp, _test.negate() ); 1639 } 1640 1641 // Transform: "((x & (m - 1)) <u m)" or "(((m - 1) & x) <u m)" into "(m >u 0)" 1642 // This is case [CMPU_MASK] which is further described at the method comment of BoolNode::Value_cmpu_and_mask(). 1643 if (cop == Op_CmpU && _test._test == BoolTest::lt && cmp1_op == Op_AndI) { 1644 Node* m = cmp2; // RHS: m 1645 for (int add_idx = 1; add_idx <= 2; add_idx++) { // LHS: "(m + (-1)) & x" or "x & (m + (-1))"? 1646 Node* maybe_m_minus_1 = cmp1->in(add_idx); 1647 if (maybe_m_minus_1->Opcode() == Op_AddI && 1648 maybe_m_minus_1->in(2)->find_int_con(0) == -1 && 1649 maybe_m_minus_1->in(1) == m) { 1650 Node* m_cmpu_0 = phase->transform(new CmpUNode(m, phase->intcon(0))); 1651 return new BoolNode(m_cmpu_0, BoolTest::gt); 1652 } 1653 } 1654 } 1655 1656 // Change x u< 1 or x u<= 0 to x == 0 1657 // and x u> 0 or u>= 1 to x != 0 1658 if (cop == Op_CmpU && 1659 cmp1_op != Op_LoadRange && 1660 (((_test._test == BoolTest::lt || _test._test == BoolTest::ge) && 1661 cmp2->find_int_con(-1) == 1) || 1662 ((_test._test == BoolTest::le || _test._test == BoolTest::gt) && 1663 cmp2->find_int_con(-1) == 0))) { 1664 Node* ncmp = phase->transform(new CmpINode(cmp1, phase->intcon(0))); 1665 return new BoolNode(ncmp, _test.is_less() ? BoolTest::eq : BoolTest::ne); 1666 } 1667 1668 // Change (arraylength <= 0) or (arraylength == 0) 1669 // into (arraylength u<= 0) 1670 // Also change (arraylength != 0) into (arraylength u> 0) 1671 // The latter version matches the code pattern generated for 1672 // array range checks, which will more likely be optimized later. 1673 if (cop == Op_CmpI && 1674 cmp1_op == Op_LoadRange && 1675 cmp2->find_int_con(-1) == 0) { 1676 if (_test._test == BoolTest::le || _test._test == BoolTest::eq) { 1677 Node* ncmp = phase->transform(new CmpUNode(cmp1, cmp2)); 1678 return new BoolNode(ncmp, BoolTest::le); 1679 } else if (_test._test == BoolTest::ne) { 1680 Node* ncmp = phase->transform(new CmpUNode(cmp1, cmp2)); 1681 return new BoolNode(ncmp, BoolTest::gt); 1682 } 1683 } 1684 1685 // Change "bool eq/ne (cmp (Conv2B X) 0)" into "bool eq/ne (cmp X 0)". 1686 // This is a standard idiom for branching on a boolean value. 1687 Node *c2b = cmp1; 1688 if( cmp2_type == TypeInt::ZERO && 1689 cmp1_op == Op_Conv2B && 1690 (_test._test == BoolTest::eq || 1691 _test._test == BoolTest::ne) ) { 1692 Node *ncmp = phase->transform(phase->type(c2b->in(1))->isa_int() 1693 ? (Node*)new CmpINode(c2b->in(1),cmp2) 1694 : (Node*)new CmpPNode(c2b->in(1),phase->makecon(TypePtr::NULL_PTR)) 1695 ); 1696 return new BoolNode( ncmp, _test._test ); 1697 } 1698 1699 // Comparing a SubI against a zero is equal to comparing the SubI 1700 // arguments directly. This only works for eq and ne comparisons 1701 // due to possible integer overflow. 1702 if ((_test._test == BoolTest::eq || _test._test == BoolTest::ne) && 1703 (cop == Op_CmpI) && 1704 (cmp1_op == Op_SubI) && 1705 ( cmp2_type == TypeInt::ZERO ) ) { 1706 Node *ncmp = phase->transform( new CmpINode(cmp1->in(1),cmp1->in(2))); 1707 return new BoolNode( ncmp, _test._test ); 1708 } 1709 1710 // Same as above but with and AddI of a constant 1711 if ((_test._test == BoolTest::eq || _test._test == BoolTest::ne) && 1712 cop == Op_CmpI && 1713 cmp1_op == Op_AddI && 1714 cmp1->in(2) != nullptr && 1715 phase->type(cmp1->in(2))->isa_int() && 1716 phase->type(cmp1->in(2))->is_int()->is_con() && 1717 cmp2_type == TypeInt::ZERO && 1718 !is_counted_loop_cmp(cmp) // modifying the exit test of a counted loop messes the counted loop shape 1719 ) { 1720 const TypeInt* cmp1_in2 = phase->type(cmp1->in(2))->is_int(); 1721 Node *ncmp = phase->transform( new CmpINode(cmp1->in(1),phase->intcon(-cmp1_in2->_hi))); 1722 return new BoolNode( ncmp, _test._test ); 1723 } 1724 1725 // Change "bool eq/ne (cmp (phi (X -X) 0))" into "bool eq/ne (cmp X 0)" 1726 // since zero check of conditional negation of an integer is equal to 1727 // zero check of the integer directly. 1728 if ((_test._test == BoolTest::eq || _test._test == BoolTest::ne) && 1729 (cop == Op_CmpI) && 1730 (cmp2_type == TypeInt::ZERO) && 1731 (cmp1_op == Op_Phi)) { 1732 // There should be a diamond phi with true path at index 1 or 2 1733 PhiNode *phi = cmp1->as_Phi(); 1734 int idx_true = phi->is_diamond_phi(); 1735 if (idx_true != 0) { 1736 // True input is in(idx_true) while false input is in(3 - idx_true) 1737 Node *tin = phi->in(idx_true); 1738 Node *fin = phi->in(3 - idx_true); 1739 if ((tin->Opcode() == Op_SubI) && 1740 (phase->type(tin->in(1)) == TypeInt::ZERO) && 1741 (tin->in(2) == fin)) { 1742 // Found conditional negation at true path, create a new CmpINode without that 1743 Node *ncmp = phase->transform(new CmpINode(fin, cmp2)); 1744 return new BoolNode(ncmp, _test._test); 1745 } 1746 if ((fin->Opcode() == Op_SubI) && 1747 (phase->type(fin->in(1)) == TypeInt::ZERO) && 1748 (fin->in(2) == tin)) { 1749 // Found conditional negation at false path, create a new CmpINode without that 1750 Node *ncmp = phase->transform(new CmpINode(tin, cmp2)); 1751 return new BoolNode(ncmp, _test._test); 1752 } 1753 } 1754 } 1755 1756 // Change (-A vs 0) into (A vs 0) by commuting the test. Disallow in the 1757 // most general case because negating 0x80000000 does nothing. Needed for 1758 // the CmpF3/SubI/CmpI idiom. 1759 if( cop == Op_CmpI && 1760 cmp1_op == Op_SubI && 1761 cmp2_type == TypeInt::ZERO && 1762 phase->type( cmp1->in(1) ) == TypeInt::ZERO && 1763 phase->type( cmp1->in(2) )->higher_equal(TypeInt::SYMINT) ) { 1764 Node *ncmp = phase->transform( new CmpINode(cmp1->in(2),cmp2)); 1765 return new BoolNode( ncmp, _test.commute() ); 1766 } 1767 1768 // Try to optimize signed integer comparison 1769 return fold_cmpI(phase, cmp->as_Sub(), cmp1, cop, cmp1_op, cmp2_type); 1770 1771 // The transformation below is not valid for either signed or unsigned 1772 // comparisons due to wraparound concerns at MAX_VALUE and MIN_VALUE. 1773 // This transformation can be resurrected when we are able to 1774 // make inferences about the range of values being subtracted from 1775 // (or added to) relative to the wraparound point. 1776 // 1777 // // Remove +/-1's if possible. 1778 // // "X <= Y-1" becomes "X < Y" 1779 // // "X+1 <= Y" becomes "X < Y" 1780 // // "X < Y+1" becomes "X <= Y" 1781 // // "X-1 < Y" becomes "X <= Y" 1782 // // Do not this to compares off of the counted-loop-end. These guys are 1783 // // checking the trip counter and they want to use the post-incremented 1784 // // counter. If they use the PRE-incremented counter, then the counter has 1785 // // to be incremented in a private block on a loop backedge. 1786 // if( du && du->cnt(this) && du->out(this)[0]->Opcode() == Op_CountedLoopEnd ) 1787 // return nullptr; 1788 // #ifndef PRODUCT 1789 // // Do not do this in a wash GVN pass during verification. 1790 // // Gets triggered by too many simple optimizations to be bothered with 1791 // // re-trying it again and again. 1792 // if( !phase->allow_progress() ) return nullptr; 1793 // #endif 1794 // // Not valid for unsigned compare because of corner cases in involving zero. 1795 // // For example, replacing "X-1 <u Y" with "X <=u Y" fails to throw an 1796 // // exception in case X is 0 (because 0-1 turns into 4billion unsigned but 1797 // // "0 <=u Y" is always true). 1798 // if( cmp->Opcode() == Op_CmpU ) return nullptr; 1799 // int cmp2_op = cmp2->Opcode(); 1800 // if( _test._test == BoolTest::le ) { 1801 // if( cmp1_op == Op_AddI && 1802 // phase->type( cmp1->in(2) ) == TypeInt::ONE ) 1803 // return clone_cmp( cmp, cmp1->in(1), cmp2, phase, BoolTest::lt ); 1804 // else if( cmp2_op == Op_AddI && 1805 // phase->type( cmp2->in(2) ) == TypeInt::MINUS_1 ) 1806 // return clone_cmp( cmp, cmp1, cmp2->in(1), phase, BoolTest::lt ); 1807 // } else if( _test._test == BoolTest::lt ) { 1808 // if( cmp1_op == Op_AddI && 1809 // phase->type( cmp1->in(2) ) == TypeInt::MINUS_1 ) 1810 // return clone_cmp( cmp, cmp1->in(1), cmp2, phase, BoolTest::le ); 1811 // else if( cmp2_op == Op_AddI && 1812 // phase->type( cmp2->in(2) ) == TypeInt::ONE ) 1813 // return clone_cmp( cmp, cmp1, cmp2->in(1), phase, BoolTest::le ); 1814 // } 1815 } 1816 1817 // We use the following Lemmas/insights for the following two transformations (1) and (2): 1818 // x & y <=u y, for any x and y (Lemma 1, masking always results in a smaller unsigned number) 1819 // y <u y + 1 is always true if y != -1 (Lemma 2, (uint)(-1 + 1) == (uint)(UINT_MAX + 1) which overflows) 1820 // y <u 0 is always false for any y (Lemma 3, 0 == UINT_MIN and nothing can be smaller than that) 1821 // 1822 // (1a) Always: Change ((x & m) <=u m ) or ((m & x) <=u m ) to always true (true by Lemma 1) 1823 // (1b) If m != -1: Change ((x & m) <u m + 1) or ((m & x) <u m + 1) to always true: 1824 // x & m <=u m is always true // (Lemma 1) 1825 // x & m <=u m <u m + 1 is always true // (Lemma 2: m <u m + 1, if m != -1) 1826 // 1827 // A counter example for (1b), if we allowed m == -1: 1828 // (x & m) <u m + 1 1829 // (x & -1) <u 0 1830 // x <u 0 1831 // which is false for any x (Lemma 3) 1832 // 1833 // (2) Change ((x & (m - 1)) <u m) or (((m - 1) & x) <u m) to (m >u 0) 1834 // This is the off-by-one variant of the above. 1835 // 1836 // We now prove that this replacement is correct. This is the same as proving 1837 // "m >u 0" if and only if "x & (m - 1) <u m", i.e. "m >u 0 <=> x & (m - 1) <u m" 1838 // 1839 // We use (Lemma 1) and (Lemma 3) from above. 1840 // 1841 // Case "x & (m - 1) <u m => m >u 0": 1842 // We prove this by contradiction: 1843 // Assume m <=u 0 which is equivalent to m == 0: 1844 // and thus 1845 // x & (m - 1) <u m = 0 // m == 0 1846 // y <u 0 // y = x & (m - 1) 1847 // by Lemma 3, this is always false, i.e. a contradiction to our assumption. 1848 // 1849 // Case "m >u 0 => x & (m - 1) <u m": 1850 // x & (m - 1) <=u (m - 1) // (Lemma 1) 1851 // x & (m - 1) <=u (m - 1) <u m // Using assumption m >u 0, no underflow of "m - 1" 1852 // 1853 // 1854 // Note that the signed version of "m > 0": 1855 // m > 0 <=> x & (m - 1) <u m 1856 // does not hold: 1857 // Assume m == -1 and x == -1: 1858 // x & (m - 1) <u m 1859 // -1 & -2 <u -1 1860 // -2 <u -1 1861 // UINT_MAX - 1 <u UINT_MAX // Signed to unsigned numbers 1862 // which is true while 1863 // m > 0 1864 // is false which is a contradiction. 1865 // 1866 // (1a) and (1b) is covered by this method since we can directly return a true value as type while (2) is covered 1867 // in BoolNode::Ideal since we create a new non-constant node (see [CMPU_MASK]). 1868 const Type* BoolNode::Value_cmpu_and_mask(PhaseValues* phase) const { 1869 Node* cmp = in(1); 1870 if (cmp != nullptr && cmp->Opcode() == Op_CmpU) { 1871 Node* cmp1 = cmp->in(1); 1872 Node* cmp2 = cmp->in(2); 1873 1874 if (cmp1->Opcode() == Op_AndI) { 1875 Node* m = nullptr; 1876 if (_test._test == BoolTest::le) { 1877 // (1a) "((x & m) <=u m)", cmp2 = m 1878 m = cmp2; 1879 } else if (_test._test == BoolTest::lt && cmp2->Opcode() == Op_AddI && cmp2->in(2)->find_int_con(0) == 1) { 1880 // (1b) "(x & m) <u m + 1" and "(m & x) <u m + 1", cmp2 = m + 1 1881 Node* rhs_m = cmp2->in(1); 1882 const TypeInt* rhs_m_type = phase->type(rhs_m)->isa_int(); 1883 if (rhs_m_type->_lo > -1 || rhs_m_type->_hi < -1) { 1884 // Exclude any case where m == -1 is possible. 1885 m = rhs_m; 1886 } 1887 } 1888 1889 if (cmp1->in(2) == m || cmp1->in(1) == m) { 1890 return TypeInt::ONE; 1891 } 1892 } 1893 } 1894 1895 return nullptr; 1896 } 1897 1898 // Simplify a Bool (convert condition codes to boolean (1 or 0)) node, 1899 // based on local information. If the input is constant, do it. 1900 const Type* BoolNode::Value(PhaseGVN* phase) const { 1901 const Type* t = Value_cmpu_and_mask(phase); 1902 if (t != nullptr) { 1903 return t; 1904 } 1905 1906 return _test.cc2logical( phase->type( in(1) ) ); 1907 } 1908 1909 #ifndef PRODUCT 1910 //------------------------------dump_spec-------------------------------------- 1911 // Dump special per-node info 1912 void BoolNode::dump_spec(outputStream *st) const { 1913 st->print("["); 1914 _test.dump_on(st); 1915 st->print("]"); 1916 } 1917 #endif 1918 1919 //----------------------is_counted_loop_exit_test------------------------------ 1920 // Returns true if node is used by a counted loop node. 1921 bool BoolNode::is_counted_loop_exit_test() { 1922 for( DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++ ) { 1923 Node* use = fast_out(i); 1924 if (use->is_CountedLoopEnd()) { 1925 return true; 1926 } 1927 } 1928 return false; 1929 } 1930 1931 //============================================================================= 1932 //------------------------------Value------------------------------------------ 1933 const Type* AbsNode::Value(PhaseGVN* phase) const { 1934 const Type* t1 = phase->type(in(1)); 1935 if (t1 == Type::TOP) return Type::TOP; 1936 1937 switch (t1->base()) { 1938 case Type::Int: { 1939 const TypeInt* ti = t1->is_int(); 1940 if (ti->is_con()) { 1941 return TypeInt::make(uabs(ti->get_con())); 1942 } 1943 break; 1944 } 1945 case Type::Long: { 1946 const TypeLong* tl = t1->is_long(); 1947 if (tl->is_con()) { 1948 return TypeLong::make(uabs(tl->get_con())); 1949 } 1950 break; 1951 } 1952 case Type::FloatCon: 1953 return TypeF::make(abs(t1->getf())); 1954 case Type::DoubleCon: 1955 return TypeD::make(abs(t1->getd())); 1956 default: 1957 break; 1958 } 1959 1960 return bottom_type(); 1961 } 1962 1963 //------------------------------Identity---------------------------------------- 1964 Node* AbsNode::Identity(PhaseGVN* phase) { 1965 Node* in1 = in(1); 1966 // No need to do abs for non-negative values 1967 if (phase->type(in1)->higher_equal(TypeInt::POS) || 1968 phase->type(in1)->higher_equal(TypeLong::POS)) { 1969 return in1; 1970 } 1971 // Convert "abs(abs(x))" into "abs(x)" 1972 if (in1->Opcode() == Opcode()) { 1973 return in1; 1974 } 1975 return this; 1976 } 1977 1978 //------------------------------Ideal------------------------------------------ 1979 Node* AbsNode::Ideal(PhaseGVN* phase, bool can_reshape) { 1980 Node* in1 = in(1); 1981 // Convert "abs(0-x)" into "abs(x)" 1982 if (in1->is_Sub() && phase->type(in1->in(1))->is_zero_type()) { 1983 set_req_X(1, in1->in(2), phase); 1984 return this; 1985 } 1986 return nullptr; 1987 } 1988 1989 //============================================================================= 1990 //------------------------------Value------------------------------------------ 1991 // Compute sqrt 1992 const Type* SqrtDNode::Value(PhaseGVN* phase) const { 1993 const Type *t1 = phase->type( in(1) ); 1994 if( t1 == Type::TOP ) return Type::TOP; 1995 if( t1->base() != Type::DoubleCon ) return Type::DOUBLE; 1996 double d = t1->getd(); 1997 if( d < 0.0 ) return Type::DOUBLE; 1998 return TypeD::make( sqrt( d ) ); 1999 } 2000 2001 const Type* SqrtFNode::Value(PhaseGVN* phase) const { 2002 const Type *t1 = phase->type( in(1) ); 2003 if( t1 == Type::TOP ) return Type::TOP; 2004 if( t1->base() != Type::FloatCon ) return Type::FLOAT; 2005 float f = t1->getf(); 2006 if( f < 0.0f ) return Type::FLOAT; 2007 return TypeF::make( (float)sqrt( (double)f ) ); 2008 } 2009 2010 const Type* SqrtHFNode::Value(PhaseGVN* phase) const { 2011 const Type* t1 = phase->type(in(1)); 2012 if (t1 == Type::TOP) { return Type::TOP; } 2013 if (t1->base() != Type::HalfFloatCon) { return Type::HALF_FLOAT; } 2014 float f = t1->getf(); 2015 if (f < 0.0f) return Type::HALF_FLOAT; 2016 return TypeH::make((float)sqrt((double)f)); 2017 } 2018 2019 const Type* ReverseINode::Value(PhaseGVN* phase) const { 2020 const Type *t1 = phase->type( in(1) ); 2021 if (t1 == Type::TOP) { 2022 return Type::TOP; 2023 } 2024 const TypeInt* t1int = t1->isa_int(); 2025 if (t1int && t1int->is_con()) { 2026 jint res = reverse_bits(t1int->get_con()); 2027 return TypeInt::make(res); 2028 } 2029 return bottom_type(); 2030 } 2031 2032 const Type* ReverseLNode::Value(PhaseGVN* phase) const { 2033 const Type *t1 = phase->type( in(1) ); 2034 if (t1 == Type::TOP) { 2035 return Type::TOP; 2036 } 2037 const TypeLong* t1long = t1->isa_long(); 2038 if (t1long && t1long->is_con()) { 2039 jlong res = reverse_bits(t1long->get_con()); 2040 return TypeLong::make(res); 2041 } 2042 return bottom_type(); 2043 } 2044 2045 Node* ReverseINode::Identity(PhaseGVN* phase) { 2046 if (in(1)->Opcode() == Op_ReverseI) { 2047 return in(1)->in(1); 2048 } 2049 return this; 2050 } 2051 2052 Node* ReverseLNode::Identity(PhaseGVN* phase) { 2053 if (in(1)->Opcode() == Op_ReverseL) { 2054 return in(1)->in(1); 2055 } 2056 return this; 2057 }