1 /* 2 * Copyright (c) 1997, 2023, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "memory/allocation.inline.hpp" 27 #include "opto/addnode.hpp" 28 #include "opto/castnode.hpp" 29 #include "opto/cfgnode.hpp" 30 #include "opto/connode.hpp" 31 #include "opto/machnode.hpp" 32 #include "opto/movenode.hpp" 33 #include "opto/mulnode.hpp" 34 #include "opto/phaseX.hpp" 35 #include "opto/subnode.hpp" 36 37 // Portions of code courtesy of Clifford Click 38 39 // Classic Add functionality. This covers all the usual 'add' behaviors for 40 // an algebraic ring. Add-integer, add-float, add-double, and binary-or are 41 // all inherited from this class. The various identity values are supplied 42 // by virtual functions. 43 44 45 //============================================================================= 46 //------------------------------hash------------------------------------------- 47 // Hash function over AddNodes. Needs to be commutative; i.e., I swap 48 // (commute) inputs to AddNodes willy-nilly so the hash function must return 49 // the same value in the presence of edge swapping. 50 uint AddNode::hash() const { 51 return (uintptr_t)in(1) + (uintptr_t)in(2) + Opcode(); 52 } 53 54 //------------------------------Identity--------------------------------------- 55 // If either input is a constant 0, return the other input. 56 Node* AddNode::Identity(PhaseGVN* phase) { 57 const Type *zero = add_id(); // The additive identity 58 if( phase->type( in(1) )->higher_equal( zero ) ) return in(2); 59 if( phase->type( in(2) )->higher_equal( zero ) ) return in(1); 60 return this; 61 } 62 63 //------------------------------commute---------------------------------------- 64 // Commute operands to move loads and constants to the right. 65 static bool commute(PhaseGVN* phase, Node* add) { 66 Node *in1 = add->in(1); 67 Node *in2 = add->in(2); 68 69 // convert "max(a,b) + min(a,b)" into "a+b". 70 if ((in1->Opcode() == add->as_Add()->max_opcode() && in2->Opcode() == add->as_Add()->min_opcode()) 71 || (in1->Opcode() == add->as_Add()->min_opcode() && in2->Opcode() == add->as_Add()->max_opcode())) { 72 Node *in11 = in1->in(1); 73 Node *in12 = in1->in(2); 74 75 Node *in21 = in2->in(1); 76 Node *in22 = in2->in(2); 77 78 if ((in11 == in21 && in12 == in22) || 79 (in11 == in22 && in12 == in21)) { 80 add->set_req_X(1, in11, phase); 81 add->set_req_X(2, in12, phase); 82 return true; 83 } 84 } 85 86 bool con_left = phase->type(in1)->singleton(); 87 bool con_right = phase->type(in2)->singleton(); 88 89 // Convert "1+x" into "x+1". 90 // Right is a constant; leave it 91 if( con_right ) return false; 92 // Left is a constant; move it right. 93 if( con_left ) { 94 add->swap_edges(1, 2); 95 return true; 96 } 97 98 // Convert "Load+x" into "x+Load". 99 // Now check for loads 100 if (in2->is_Load()) { 101 if (!in1->is_Load()) { 102 // already x+Load to return 103 return false; 104 } 105 // both are loads, so fall through to sort inputs by idx 106 } else if( in1->is_Load() ) { 107 // Left is a Load and Right is not; move it right. 108 add->swap_edges(1, 2); 109 return true; 110 } 111 112 PhiNode *phi; 113 // Check for tight loop increments: Loop-phi of Add of loop-phi 114 if (in1->is_Phi() && (phi = in1->as_Phi()) && phi->region()->is_Loop() && phi->in(2) == add) 115 return false; 116 if (in2->is_Phi() && (phi = in2->as_Phi()) && phi->region()->is_Loop() && phi->in(2) == add) { 117 add->swap_edges(1, 2); 118 return true; 119 } 120 121 // Otherwise, sort inputs (commutativity) to help value numbering. 122 if( in1->_idx > in2->_idx ) { 123 add->swap_edges(1, 2); 124 return true; 125 } 126 return false; 127 } 128 129 //------------------------------Idealize--------------------------------------- 130 // If we get here, we assume we are associative! 131 Node *AddNode::Ideal(PhaseGVN *phase, bool can_reshape) { 132 const Type *t1 = phase->type(in(1)); 133 const Type *t2 = phase->type(in(2)); 134 bool con_left = t1->singleton(); 135 bool con_right = t2->singleton(); 136 137 // Check for commutative operation desired 138 if (commute(phase, this)) return this; 139 140 AddNode *progress = nullptr; // Progress flag 141 142 // Convert "(x+1)+2" into "x+(1+2)". If the right input is a 143 // constant, and the left input is an add of a constant, flatten the 144 // expression tree. 145 Node *add1 = in(1); 146 Node *add2 = in(2); 147 int add1_op = add1->Opcode(); 148 int this_op = Opcode(); 149 if (con_right && t2 != Type::TOP && // Right input is a constant? 150 add1_op == this_op) { // Left input is an Add? 151 152 // Type of left _in right input 153 const Type *t12 = phase->type(add1->in(2)); 154 if (t12->singleton() && t12 != Type::TOP) { // Left input is an add of a constant? 155 // Check for rare case of closed data cycle which can happen inside 156 // unreachable loops. In these cases the computation is undefined. 157 #ifdef ASSERT 158 Node *add11 = add1->in(1); 159 int add11_op = add11->Opcode(); 160 if ((add1 == add1->in(1)) 161 || (add11_op == this_op && add11->in(1) == add1)) { 162 assert(false, "dead loop in AddNode::Ideal"); 163 } 164 #endif 165 // The Add of the flattened expression 166 Node *x1 = add1->in(1); 167 Node *x2 = phase->makecon(add1->as_Add()->add_ring(t2, t12)); 168 set_req_X(2, x2, phase); 169 set_req_X(1, x1, phase); 170 progress = this; // Made progress 171 add1 = in(1); 172 add1_op = add1->Opcode(); 173 } 174 } 175 176 // Convert "(x+1)+y" into "(x+y)+1". Push constants down the expression tree. 177 if (add1_op == this_op && !con_right) { 178 Node *a12 = add1->in(2); 179 const Type *t12 = phase->type( a12 ); 180 if (t12->singleton() && t12 != Type::TOP && (add1 != add1->in(1)) && 181 !(add1->in(1)->is_Phi() && (add1->in(1)->as_Phi()->is_tripcount(T_INT) || add1->in(1)->as_Phi()->is_tripcount(T_LONG)))) { 182 assert(add1->in(1) != this, "dead loop in AddNode::Ideal"); 183 add2 = add1->clone(); 184 add2->set_req(2, in(2)); 185 add2 = phase->transform(add2); 186 set_req_X(1, add2, phase); 187 set_req_X(2, a12, phase); 188 progress = this; 189 add2 = a12; 190 } 191 } 192 193 // Convert "x+(y+1)" into "(x+y)+1". Push constants down the expression tree. 194 int add2_op = add2->Opcode(); 195 if (add2_op == this_op && !con_left) { 196 Node *a22 = add2->in(2); 197 const Type *t22 = phase->type( a22 ); 198 if (t22->singleton() && t22 != Type::TOP && (add2 != add2->in(1)) && 199 !(add2->in(1)->is_Phi() && (add2->in(1)->as_Phi()->is_tripcount(T_INT) || add2->in(1)->as_Phi()->is_tripcount(T_LONG)))) { 200 assert(add2->in(1) != this, "dead loop in AddNode::Ideal"); 201 Node *addx = add2->clone(); 202 addx->set_req(1, in(1)); 203 addx->set_req(2, add2->in(1)); 204 addx = phase->transform(addx); 205 set_req_X(1, addx, phase); 206 set_req_X(2, a22, phase); 207 progress = this; 208 } 209 } 210 211 return progress; 212 } 213 214 //------------------------------Value----------------------------------------- 215 // An add node sums it's two _in. If one input is an RSD, we must mixin 216 // the other input's symbols. 217 const Type* AddNode::Value(PhaseGVN* phase) const { 218 // Either input is TOP ==> the result is TOP 219 const Type* t1 = phase->type(in(1)); 220 const Type* t2 = phase->type(in(2)); 221 if (t1 == Type::TOP || t2 == Type::TOP) { 222 return Type::TOP; 223 } 224 225 // Check for an addition involving the additive identity 226 const Type* tadd = add_of_identity(t1, t2); 227 if (tadd != nullptr) { 228 return tadd; 229 } 230 231 return add_ring(t1, t2); // Local flavor of type addition 232 } 233 234 //------------------------------add_identity----------------------------------- 235 // Check for addition of the identity 236 const Type *AddNode::add_of_identity( const Type *t1, const Type *t2 ) const { 237 const Type *zero = add_id(); // The additive identity 238 if( t1->higher_equal( zero ) ) return t2; 239 if( t2->higher_equal( zero ) ) return t1; 240 241 return nullptr; 242 } 243 244 AddNode* AddNode::make(Node* in1, Node* in2, BasicType bt) { 245 switch (bt) { 246 case T_INT: 247 return new AddINode(in1, in2); 248 case T_LONG: 249 return new AddLNode(in1, in2); 250 default: 251 fatal("Not implemented for %s", type2name(bt)); 252 } 253 return nullptr; 254 } 255 256 //============================================================================= 257 //------------------------------Idealize--------------------------------------- 258 Node* AddNode::IdealIL(PhaseGVN* phase, bool can_reshape, BasicType bt) { 259 Node* in1 = in(1); 260 Node* in2 = in(2); 261 int op1 = in1->Opcode(); 262 int op2 = in2->Opcode(); 263 // Fold (con1-x)+con2 into (con1+con2)-x 264 if (op1 == Op_Add(bt) && op2 == Op_Sub(bt)) { 265 // Swap edges to try optimizations below 266 in1 = in2; 267 in2 = in(1); 268 op1 = op2; 269 op2 = in2->Opcode(); 270 } 271 if (op1 == Op_Sub(bt)) { 272 const Type* t_sub1 = phase->type(in1->in(1)); 273 const Type* t_2 = phase->type(in2 ); 274 if (t_sub1->singleton() && t_2->singleton() && t_sub1 != Type::TOP && t_2 != Type::TOP) { 275 return SubNode::make(phase->makecon(add_ring(t_sub1, t_2)), in1->in(2), bt); 276 } 277 // Convert "(a-b)+(c-d)" into "(a+c)-(b+d)" 278 if (op2 == Op_Sub(bt)) { 279 // Check for dead cycle: d = (a-b)+(c-d) 280 assert( in1->in(2) != this && in2->in(2) != this, 281 "dead loop in AddINode::Ideal" ); 282 Node* sub = SubNode::make(nullptr, nullptr, bt); 283 sub->init_req(1, phase->transform(AddNode::make(in1->in(1), in2->in(1), bt))); 284 sub->init_req(2, phase->transform(AddNode::make(in1->in(2), in2->in(2), bt))); 285 return sub; 286 } 287 // Convert "(a-b)+(b+c)" into "(a+c)" 288 if (op2 == Op_Add(bt) && in1->in(2) == in2->in(1)) { 289 assert(in1->in(1) != this && in2->in(2) != this,"dead loop in AddINode::Ideal/AddLNode::Ideal"); 290 return AddNode::make(in1->in(1), in2->in(2), bt); 291 } 292 // Convert "(a-b)+(c+b)" into "(a+c)" 293 if (op2 == Op_Add(bt) && in1->in(2) == in2->in(2)) { 294 assert(in1->in(1) != this && in2->in(1) != this,"dead loop in AddINode::Ideal/AddLNode::Ideal"); 295 return AddNode::make(in1->in(1), in2->in(1), bt); 296 } 297 } 298 299 // Convert (con - y) + x into "(x - y) + con" 300 if (op1 == Op_Sub(bt) && in1->in(1)->Opcode() == Op_ConIL(bt) 301 && in1 != in1->in(2) && !(in1->in(2)->is_Phi() && in1->in(2)->as_Phi()->is_tripcount(bt))) { 302 return AddNode::make(phase->transform(SubNode::make(in2, in1->in(2), bt)), in1->in(1), bt); 303 } 304 305 // Convert x + (con - y) into "(x - y) + con" 306 if (op2 == Op_Sub(bt) && in2->in(1)->Opcode() == Op_ConIL(bt) 307 && in2 != in2->in(2) && !(in2->in(2)->is_Phi() && in2->in(2)->as_Phi()->is_tripcount(bt))) { 308 return AddNode::make(phase->transform(SubNode::make(in1, in2->in(2), bt)), in2->in(1), bt); 309 } 310 311 // Associative 312 if (op1 == Op_Mul(bt) && op2 == Op_Mul(bt)) { 313 Node* add_in1 = nullptr; 314 Node* add_in2 = nullptr; 315 Node* mul_in = nullptr; 316 317 if (in1->in(1) == in2->in(1)) { 318 // Convert "a*b+a*c into a*(b+c) 319 add_in1 = in1->in(2); 320 add_in2 = in2->in(2); 321 mul_in = in1->in(1); 322 } else if (in1->in(2) == in2->in(1)) { 323 // Convert a*b+b*c into b*(a+c) 324 add_in1 = in1->in(1); 325 add_in2 = in2->in(2); 326 mul_in = in1->in(2); 327 } else if (in1->in(2) == in2->in(2)) { 328 // Convert a*c+b*c into (a+b)*c 329 add_in1 = in1->in(1); 330 add_in2 = in2->in(1); 331 mul_in = in1->in(2); 332 } else if (in1->in(1) == in2->in(2)) { 333 // Convert a*b+c*a into a*(b+c) 334 add_in1 = in1->in(2); 335 add_in2 = in2->in(1); 336 mul_in = in1->in(1); 337 } 338 339 if (mul_in != nullptr) { 340 Node* add = phase->transform(AddNode::make(add_in1, add_in2, bt)); 341 return MulNode::make(mul_in, add, bt); 342 } 343 } 344 345 // Convert (x >>> rshift) + (x << lshift) into RotateRight(x, rshift) 346 if (Matcher::match_rule_supported(Op_RotateRight) && 347 ((op1 == Op_URShift(bt) && op2 == Op_LShift(bt)) || (op1 == Op_LShift(bt) && op2 == Op_URShift(bt))) && 348 in1->in(1) != nullptr && in1->in(1) == in2->in(1)) { 349 Node* rshift = op1 == Op_URShift(bt) ? in1->in(2) : in2->in(2); 350 Node* lshift = op1 == Op_URShift(bt) ? in2->in(2) : in1->in(2); 351 if (rshift != nullptr && lshift != nullptr) { 352 const TypeInt* rshift_t = phase->type(rshift)->isa_int(); 353 const TypeInt* lshift_t = phase->type(lshift)->isa_int(); 354 int bits = bt == T_INT ? 32 : 64; 355 int mask = bt == T_INT ? 0x1F : 0x3F; 356 if (lshift_t != nullptr && lshift_t->is_con() && 357 rshift_t != nullptr && rshift_t->is_con() && 358 ((lshift_t->get_con() & mask) == (bits - (rshift_t->get_con() & mask)))) { 359 return new RotateRightNode(in1->in(1), phase->intcon(rshift_t->get_con() & mask), TypeInteger::bottom(bt)); 360 } 361 } 362 } 363 364 return AddNode::Ideal(phase, can_reshape); 365 } 366 367 368 Node* AddINode::Ideal(PhaseGVN* phase, bool can_reshape) { 369 Node* in1 = in(1); 370 Node* in2 = in(2); 371 int op1 = in1->Opcode(); 372 int op2 = in2->Opcode(); 373 374 // Convert (x>>>z)+y into (x+(y<<z))>>>z for small constant z and y. 375 // Helps with array allocation math constant folding 376 // See 4790063: 377 // Unrestricted transformation is unsafe for some runtime values of 'x' 378 // ( x == 0, z == 1, y == -1 ) fails 379 // ( x == -5, z == 1, y == 1 ) fails 380 // Transform works for small z and small negative y when the addition 381 // (x + (y << z)) does not cross zero. 382 // Implement support for negative y and (x >= -(y << z)) 383 // Have not observed cases where type information exists to support 384 // positive y and (x <= -(y << z)) 385 if (op1 == Op_URShiftI && op2 == Op_ConI && 386 in1->in(2)->Opcode() == Op_ConI) { 387 jint z = phase->type(in1->in(2))->is_int()->get_con() & 0x1f; // only least significant 5 bits matter 388 jint y = phase->type(in2)->is_int()->get_con(); 389 390 if (z < 5 && -5 < y && y < 0) { 391 const Type* t_in11 = phase->type(in1->in(1)); 392 if( t_in11 != Type::TOP && (t_in11->is_int()->_lo >= -(y << z))) { 393 Node* a = phase->transform(new AddINode( in1->in(1), phase->intcon(y<<z))); 394 return new URShiftINode(a, in1->in(2)); 395 } 396 } 397 } 398 399 return AddNode::IdealIL(phase, can_reshape, T_INT); 400 } 401 402 403 //------------------------------Identity--------------------------------------- 404 // Fold (x-y)+y OR y+(x-y) into x 405 Node* AddINode::Identity(PhaseGVN* phase) { 406 if (in(1)->Opcode() == Op_SubI && in(1)->in(2) == in(2)) { 407 return in(1)->in(1); 408 } else if (in(2)->Opcode() == Op_SubI && in(2)->in(2) == in(1)) { 409 return in(2)->in(1); 410 } 411 return AddNode::Identity(phase); 412 } 413 414 415 //------------------------------add_ring--------------------------------------- 416 // Supplied function returns the sum of the inputs. Guaranteed never 417 // to be passed a TOP or BOTTOM type, these are filtered out by 418 // pre-check. 419 const Type *AddINode::add_ring( const Type *t0, const Type *t1 ) const { 420 const TypeInt *r0 = t0->is_int(); // Handy access 421 const TypeInt *r1 = t1->is_int(); 422 int lo = java_add(r0->_lo, r1->_lo); 423 int hi = java_add(r0->_hi, r1->_hi); 424 if( !(r0->is_con() && r1->is_con()) ) { 425 // Not both constants, compute approximate result 426 if( (r0->_lo & r1->_lo) < 0 && lo >= 0 ) { 427 lo = min_jint; hi = max_jint; // Underflow on the low side 428 } 429 if( (~(r0->_hi | r1->_hi)) < 0 && hi < 0 ) { 430 lo = min_jint; hi = max_jint; // Overflow on the high side 431 } 432 if( lo > hi ) { // Handle overflow 433 lo = min_jint; hi = max_jint; 434 } 435 } else { 436 // both constants, compute precise result using 'lo' and 'hi' 437 // Semantics define overflow and underflow for integer addition 438 // as expected. In particular: 0x80000000 + 0x80000000 --> 0x0 439 } 440 return TypeInt::make( lo, hi, MAX2(r0->_widen,r1->_widen) ); 441 } 442 443 444 //============================================================================= 445 //------------------------------Idealize--------------------------------------- 446 Node* AddLNode::Ideal(PhaseGVN* phase, bool can_reshape) { 447 return AddNode::IdealIL(phase, can_reshape, T_LONG); 448 } 449 450 451 //------------------------------Identity--------------------------------------- 452 // Fold (x-y)+y OR y+(x-y) into x 453 Node* AddLNode::Identity(PhaseGVN* phase) { 454 if (in(1)->Opcode() == Op_SubL && in(1)->in(2) == in(2)) { 455 return in(1)->in(1); 456 } else if (in(2)->Opcode() == Op_SubL && in(2)->in(2) == in(1)) { 457 return in(2)->in(1); 458 } 459 return AddNode::Identity(phase); 460 } 461 462 463 //------------------------------add_ring--------------------------------------- 464 // Supplied function returns the sum of the inputs. Guaranteed never 465 // to be passed a TOP or BOTTOM type, these are filtered out by 466 // pre-check. 467 const Type *AddLNode::add_ring( const Type *t0, const Type *t1 ) const { 468 const TypeLong *r0 = t0->is_long(); // Handy access 469 const TypeLong *r1 = t1->is_long(); 470 jlong lo = java_add(r0->_lo, r1->_lo); 471 jlong hi = java_add(r0->_hi, r1->_hi); 472 if( !(r0->is_con() && r1->is_con()) ) { 473 // Not both constants, compute approximate result 474 if( (r0->_lo & r1->_lo) < 0 && lo >= 0 ) { 475 lo =min_jlong; hi = max_jlong; // Underflow on the low side 476 } 477 if( (~(r0->_hi | r1->_hi)) < 0 && hi < 0 ) { 478 lo = min_jlong; hi = max_jlong; // Overflow on the high side 479 } 480 if( lo > hi ) { // Handle overflow 481 lo = min_jlong; hi = max_jlong; 482 } 483 } else { 484 // both constants, compute precise result using 'lo' and 'hi' 485 // Semantics define overflow and underflow for integer addition 486 // as expected. In particular: 0x80000000 + 0x80000000 --> 0x0 487 } 488 return TypeLong::make( lo, hi, MAX2(r0->_widen,r1->_widen) ); 489 } 490 491 492 //============================================================================= 493 //------------------------------add_of_identity-------------------------------- 494 // Check for addition of the identity 495 const Type *AddFNode::add_of_identity( const Type *t1, const Type *t2 ) const { 496 // x ADD 0 should return x unless 'x' is a -zero 497 // 498 // const Type *zero = add_id(); // The additive identity 499 // jfloat f1 = t1->getf(); 500 // jfloat f2 = t2->getf(); 501 // 502 // if( t1->higher_equal( zero ) ) return t2; 503 // if( t2->higher_equal( zero ) ) return t1; 504 505 return nullptr; 506 } 507 508 //------------------------------add_ring--------------------------------------- 509 // Supplied function returns the sum of the inputs. 510 // This also type-checks the inputs for sanity. Guaranteed never to 511 // be passed a TOP or BOTTOM type, these are filtered out by pre-check. 512 const Type *AddFNode::add_ring( const Type *t0, const Type *t1 ) const { 513 if (!t0->isa_float_constant() || !t1->isa_float_constant()) { 514 return bottom_type(); 515 } 516 return TypeF::make( t0->getf() + t1->getf() ); 517 } 518 519 //------------------------------Ideal------------------------------------------ 520 Node *AddFNode::Ideal(PhaseGVN *phase, bool can_reshape) { 521 // Floating point additions are not associative because of boundary conditions (infinity) 522 return commute(phase, this) ? this : nullptr; 523 } 524 525 526 //============================================================================= 527 //------------------------------add_of_identity-------------------------------- 528 // Check for addition of the identity 529 const Type *AddDNode::add_of_identity( const Type *t1, const Type *t2 ) const { 530 // x ADD 0 should return x unless 'x' is a -zero 531 // 532 // const Type *zero = add_id(); // The additive identity 533 // jfloat f1 = t1->getf(); 534 // jfloat f2 = t2->getf(); 535 // 536 // if( t1->higher_equal( zero ) ) return t2; 537 // if( t2->higher_equal( zero ) ) return t1; 538 539 return nullptr; 540 } 541 //------------------------------add_ring--------------------------------------- 542 // Supplied function returns the sum of the inputs. 543 // This also type-checks the inputs for sanity. Guaranteed never to 544 // be passed a TOP or BOTTOM type, these are filtered out by pre-check. 545 const Type *AddDNode::add_ring( const Type *t0, const Type *t1 ) const { 546 if (!t0->isa_double_constant() || !t1->isa_double_constant()) { 547 return bottom_type(); 548 } 549 return TypeD::make( t0->getd() + t1->getd() ); 550 } 551 552 //------------------------------Ideal------------------------------------------ 553 Node *AddDNode::Ideal(PhaseGVN *phase, bool can_reshape) { 554 // Floating point additions are not associative because of boundary conditions (infinity) 555 return commute(phase, this) ? this : nullptr; 556 } 557 558 559 //============================================================================= 560 //------------------------------Identity--------------------------------------- 561 // If one input is a constant 0, return the other input. 562 Node* AddPNode::Identity(PhaseGVN* phase) { 563 return ( phase->type( in(Offset) )->higher_equal( TypeX_ZERO ) ) ? in(Address) : this; 564 } 565 566 //------------------------------Idealize--------------------------------------- 567 Node *AddPNode::Ideal(PhaseGVN *phase, bool can_reshape) { 568 // Bail out if dead inputs 569 if( phase->type( in(Address) ) == Type::TOP ) return nullptr; 570 571 // If the left input is an add of a constant, flatten the expression tree. 572 const Node *n = in(Address); 573 if (n->is_AddP() && n->in(Base) == in(Base)) { 574 const AddPNode *addp = n->as_AddP(); // Left input is an AddP 575 assert( !addp->in(Address)->is_AddP() || 576 addp->in(Address)->as_AddP() != addp, 577 "dead loop in AddPNode::Ideal" ); 578 // Type of left input's right input 579 const Type *t = phase->type( addp->in(Offset) ); 580 if( t == Type::TOP ) return nullptr; 581 const TypeX *t12 = t->is_intptr_t(); 582 if( t12->is_con() ) { // Left input is an add of a constant? 583 // If the right input is a constant, combine constants 584 const Type *temp_t2 = phase->type( in(Offset) ); 585 if( temp_t2 == Type::TOP ) return nullptr; 586 const TypeX *t2 = temp_t2->is_intptr_t(); 587 Node* address; 588 Node* offset; 589 if( t2->is_con() ) { 590 // The Add of the flattened expression 591 address = addp->in(Address); 592 offset = phase->MakeConX(t2->get_con() + t12->get_con()); 593 } else { 594 // Else move the constant to the right. ((A+con)+B) into ((A+B)+con) 595 address = phase->transform(new AddPNode(in(Base),addp->in(Address),in(Offset))); 596 offset = addp->in(Offset); 597 } 598 set_req_X(Address, address, phase); 599 set_req_X(Offset, offset, phase); 600 return this; 601 } 602 } 603 604 // Raw pointers? 605 if( in(Base)->bottom_type() == Type::TOP ) { 606 // If this is a null+long form (from unsafe accesses), switch to a rawptr. 607 if (phase->type(in(Address)) == TypePtr::NULL_PTR) { 608 Node* offset = in(Offset); 609 return new CastX2PNode(offset); 610 } 611 } 612 613 // If the right is an add of a constant, push the offset down. 614 // Convert: (ptr + (offset+con)) into (ptr+offset)+con. 615 // The idea is to merge array_base+scaled_index groups together, 616 // and only have different constant offsets from the same base. 617 const Node *add = in(Offset); 618 if( add->Opcode() == Op_AddX && add->in(1) != add ) { 619 const Type *t22 = phase->type( add->in(2) ); 620 if( t22->singleton() && (t22 != Type::TOP) ) { // Right input is an add of a constant? 621 set_req(Address, phase->transform(new AddPNode(in(Base),in(Address),add->in(1)))); 622 set_req_X(Offset, add->in(2), phase); // puts add on igvn worklist if needed 623 return this; // Made progress 624 } 625 } 626 627 return nullptr; // No progress 628 } 629 630 //------------------------------bottom_type------------------------------------ 631 // Bottom-type is the pointer-type with unknown offset. 632 const Type *AddPNode::bottom_type() const { 633 if (in(Address) == nullptr) return TypePtr::BOTTOM; 634 const TypePtr *tp = in(Address)->bottom_type()->isa_ptr(); 635 if( !tp ) return Type::TOP; // TOP input means TOP output 636 assert( in(Offset)->Opcode() != Op_ConP, "" ); 637 const Type *t = in(Offset)->bottom_type(); 638 if( t == Type::TOP ) 639 return tp->add_offset(Type::OffsetTop); 640 const TypeX *tx = t->is_intptr_t(); 641 intptr_t txoffset = Type::OffsetBot; 642 if (tx->is_con()) { // Left input is an add of a constant? 643 txoffset = tx->get_con(); 644 } 645 if (tp->isa_aryptr()) { 646 // In the case of a flat inline type array, each field has its 647 // own slice so we need to extract the field being accessed from 648 // the address computation 649 return tp->is_aryptr()->add_field_offset_and_offset(txoffset); 650 } 651 return tp->add_offset(txoffset); 652 } 653 654 //------------------------------Value------------------------------------------ 655 const Type* AddPNode::Value(PhaseGVN* phase) const { 656 // Either input is TOP ==> the result is TOP 657 const Type *t1 = phase->type( in(Address) ); 658 const Type *t2 = phase->type( in(Offset) ); 659 if( t1 == Type::TOP ) return Type::TOP; 660 if( t2 == Type::TOP ) return Type::TOP; 661 662 // Left input is a pointer 663 const TypePtr *p1 = t1->isa_ptr(); 664 // Right input is an int 665 const TypeX *p2 = t2->is_intptr_t(); 666 // Add 'em 667 intptr_t p2offset = Type::OffsetBot; 668 if (p2->is_con()) { // Left input is an add of a constant? 669 p2offset = p2->get_con(); 670 } 671 if (p1->isa_aryptr()) { 672 // In the case of a flat inline type array, each field has its 673 // own slice so we need to extract the field being accessed from 674 // the address computation 675 return p1->is_aryptr()->add_field_offset_and_offset(p2offset); 676 } 677 return p1->add_offset(p2offset); 678 } 679 680 //------------------------Ideal_base_and_offset-------------------------------- 681 // Split an oop pointer into a base and offset. 682 // (The offset might be Type::OffsetBot in the case of an array.) 683 // Return the base, or null if failure. 684 Node* AddPNode::Ideal_base_and_offset(Node* ptr, PhaseValues* phase, 685 // second return value: 686 intptr_t& offset) { 687 if (ptr->is_AddP()) { 688 Node* base = ptr->in(AddPNode::Base); 689 Node* addr = ptr->in(AddPNode::Address); 690 Node* offs = ptr->in(AddPNode::Offset); 691 if (base == addr || base->is_top()) { 692 offset = phase->find_intptr_t_con(offs, Type::OffsetBot); 693 if (offset != Type::OffsetBot) { 694 return addr; 695 } 696 } 697 } 698 offset = Type::OffsetBot; 699 return nullptr; 700 } 701 702 //------------------------------unpack_offsets---------------------------------- 703 // Collect the AddP offset values into the elements array, giving up 704 // if there are more than length. 705 int AddPNode::unpack_offsets(Node* elements[], int length) { 706 int count = 0; 707 Node* addr = this; 708 Node* base = addr->in(AddPNode::Base); 709 while (addr->is_AddP()) { 710 if (addr->in(AddPNode::Base) != base) { 711 // give up 712 return -1; 713 } 714 elements[count++] = addr->in(AddPNode::Offset); 715 if (count == length) { 716 // give up 717 return -1; 718 } 719 addr = addr->in(AddPNode::Address); 720 } 721 if (addr != base) { 722 return -1; 723 } 724 return count; 725 } 726 727 //------------------------------match_edge------------------------------------- 728 // Do we Match on this edge index or not? Do not match base pointer edge 729 uint AddPNode::match_edge(uint idx) const { 730 return idx > Base; 731 } 732 733 //============================================================================= 734 //------------------------------Identity--------------------------------------- 735 Node* OrINode::Identity(PhaseGVN* phase) { 736 // x | x => x 737 if (in(1) == in(2)) { 738 return in(1); 739 } 740 741 return AddNode::Identity(phase); 742 } 743 744 // Find shift value for Integer or Long OR. 745 Node* rotate_shift(PhaseGVN* phase, Node* lshift, Node* rshift, int mask) { 746 // val << norm_con_shift | val >> ({32|64} - norm_con_shift) => rotate_left val, norm_con_shift 747 const TypeInt* lshift_t = phase->type(lshift)->isa_int(); 748 const TypeInt* rshift_t = phase->type(rshift)->isa_int(); 749 if (lshift_t != nullptr && lshift_t->is_con() && 750 rshift_t != nullptr && rshift_t->is_con() && 751 ((lshift_t->get_con() & mask) == ((mask + 1) - (rshift_t->get_con() & mask)))) { 752 return phase->intcon(lshift_t->get_con() & mask); 753 } 754 // val << var_shift | val >> ({0|32|64} - var_shift) => rotate_left val, var_shift 755 if (rshift->Opcode() == Op_SubI && rshift->in(2) == lshift && rshift->in(1)->is_Con()){ 756 const TypeInt* shift_t = phase->type(rshift->in(1))->isa_int(); 757 if (shift_t != nullptr && shift_t->is_con() && 758 (shift_t->get_con() == 0 || shift_t->get_con() == (mask + 1))) { 759 return lshift; 760 } 761 } 762 return nullptr; 763 } 764 765 Node* OrINode::Ideal(PhaseGVN* phase, bool can_reshape) { 766 int lopcode = in(1)->Opcode(); 767 int ropcode = in(2)->Opcode(); 768 if (Matcher::match_rule_supported(Op_RotateLeft) && 769 lopcode == Op_LShiftI && ropcode == Op_URShiftI && in(1)->in(1) == in(2)->in(1)) { 770 Node* lshift = in(1)->in(2); 771 Node* rshift = in(2)->in(2); 772 Node* shift = rotate_shift(phase, lshift, rshift, 0x1F); 773 if (shift != nullptr) { 774 return new RotateLeftNode(in(1)->in(1), shift, TypeInt::INT); 775 } 776 return nullptr; 777 } 778 if (Matcher::match_rule_supported(Op_RotateRight) && 779 lopcode == Op_URShiftI && ropcode == Op_LShiftI && in(1)->in(1) == in(2)->in(1)) { 780 Node* rshift = in(1)->in(2); 781 Node* lshift = in(2)->in(2); 782 Node* shift = rotate_shift(phase, rshift, lshift, 0x1F); 783 if (shift != nullptr) { 784 return new RotateRightNode(in(1)->in(1), shift, TypeInt::INT); 785 } 786 } 787 return nullptr; 788 } 789 790 //------------------------------add_ring--------------------------------------- 791 // Supplied function returns the sum of the inputs IN THE CURRENT RING. For 792 // the logical operations the ring's ADD is really a logical OR function. 793 // This also type-checks the inputs for sanity. Guaranteed never to 794 // be passed a TOP or BOTTOM type, these are filtered out by pre-check. 795 const Type *OrINode::add_ring( const Type *t0, const Type *t1 ) const { 796 const TypeInt *r0 = t0->is_int(); // Handy access 797 const TypeInt *r1 = t1->is_int(); 798 799 // If both args are bool, can figure out better types 800 if ( r0 == TypeInt::BOOL ) { 801 if ( r1 == TypeInt::ONE) { 802 return TypeInt::ONE; 803 } else if ( r1 == TypeInt::BOOL ) { 804 return TypeInt::BOOL; 805 } 806 } else if ( r0 == TypeInt::ONE ) { 807 if ( r1 == TypeInt::BOOL ) { 808 return TypeInt::ONE; 809 } 810 } 811 812 // If either input is not a constant, just return all integers. 813 if( !r0->is_con() || !r1->is_con() ) 814 return TypeInt::INT; // Any integer, but still no symbols. 815 816 // Otherwise just OR them bits. 817 return TypeInt::make( r0->get_con() | r1->get_con() ); 818 } 819 820 //============================================================================= 821 //------------------------------Identity--------------------------------------- 822 Node* OrLNode::Identity(PhaseGVN* phase) { 823 // x | x => x 824 if (in(1) == in(2)) { 825 return in(1); 826 } 827 828 return AddNode::Identity(phase); 829 } 830 831 Node* OrLNode::Ideal(PhaseGVN* phase, bool can_reshape) { 832 int lopcode = in(1)->Opcode(); 833 int ropcode = in(2)->Opcode(); 834 if (Matcher::match_rule_supported(Op_RotateLeft) && 835 lopcode == Op_LShiftL && ropcode == Op_URShiftL && in(1)->in(1) == in(2)->in(1)) { 836 Node* lshift = in(1)->in(2); 837 Node* rshift = in(2)->in(2); 838 Node* shift = rotate_shift(phase, lshift, rshift, 0x3F); 839 if (shift != nullptr) { 840 return new RotateLeftNode(in(1)->in(1), shift, TypeLong::LONG); 841 } 842 return nullptr; 843 } 844 if (Matcher::match_rule_supported(Op_RotateRight) && 845 lopcode == Op_URShiftL && ropcode == Op_LShiftL && in(1)->in(1) == in(2)->in(1)) { 846 Node* rshift = in(1)->in(2); 847 Node* lshift = in(2)->in(2); 848 Node* shift = rotate_shift(phase, rshift, lshift, 0x3F); 849 if (shift != nullptr) { 850 return new RotateRightNode(in(1)->in(1), shift, TypeLong::LONG); 851 } 852 } 853 return nullptr; 854 } 855 856 //------------------------------add_ring--------------------------------------- 857 const Type *OrLNode::add_ring( const Type *t0, const Type *t1 ) const { 858 const TypeLong *r0 = t0->is_long(); // Handy access 859 const TypeLong *r1 = t1->is_long(); 860 861 // If either input is not a constant, just return all integers. 862 if( !r0->is_con() || !r1->is_con() ) 863 return TypeLong::LONG; // Any integer, but still no symbols. 864 865 // Otherwise just OR them bits. 866 return TypeLong::make( r0->get_con() | r1->get_con() ); 867 } 868 869 //---------------------------Helper ------------------------------------------- 870 /* Decide if the given node is used only in arithmetic expressions(add or sub). 871 */ 872 static bool is_used_in_only_arithmetic(Node* n, BasicType bt) { 873 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 874 Node* u = n->fast_out(i); 875 if (u->Opcode() != Op_Add(bt) && u->Opcode() != Op_Sub(bt)) { 876 return false; 877 } 878 } 879 return true; 880 } 881 882 //============================================================================= 883 //------------------------------Idealize--------------------------------------- 884 Node* XorINode::Ideal(PhaseGVN* phase, bool can_reshape) { 885 Node* in1 = in(1); 886 Node* in2 = in(2); 887 888 // Convert ~x into -1-x when ~x is used in an arithmetic expression 889 // or x itself is an expression. 890 if (phase->type(in2) == TypeInt::MINUS_1) { // follows LHS^(-1), i.e., ~LHS 891 if (phase->is_IterGVN()) { 892 if (is_used_in_only_arithmetic(this, T_INT) 893 // LHS is arithmetic 894 || (in1->Opcode() == Op_AddI || in1->Opcode() == Op_SubI)) { 895 return new SubINode(in2, in1); 896 } 897 } else { 898 // graph could be incomplete in GVN so we postpone to IGVN 899 phase->record_for_igvn(this); 900 } 901 } 902 903 // Propagate xor through constant cmoves. This pattern can occur after expansion of Conv2B nodes. 904 const TypeInt* in2_type = phase->type(in2)->isa_int(); 905 if (in1->Opcode() == Op_CMoveI && in2_type != nullptr && in2_type->is_con()) { 906 int in2_val = in2_type->get_con(); 907 908 // Get types of both sides of the CMove 909 const TypeInt* left = phase->type(in1->in(CMoveNode::IfFalse))->isa_int(); 910 const TypeInt* right = phase->type(in1->in(CMoveNode::IfTrue))->isa_int(); 911 912 // Ensure that both sides are int constants 913 if (left != nullptr && right != nullptr && left->is_con() && right->is_con()) { 914 Node* cond = in1->in(CMoveNode::Condition); 915 916 // Check that the comparison is a bool and that the cmp node type is correct 917 if (cond->is_Bool()) { 918 int cmp_op = cond->in(1)->Opcode(); 919 920 if (cmp_op == Op_CmpI || cmp_op == Op_CmpP) { 921 int l_val = left->get_con(); 922 int r_val = right->get_con(); 923 924 return new CMoveINode(cond, phase->intcon(l_val ^ in2_val), phase->intcon(r_val ^ in2_val), TypeInt::INT); 925 } 926 } 927 } 928 } 929 930 return AddNode::Ideal(phase, can_reshape); 931 } 932 933 const Type* XorINode::Value(PhaseGVN* phase) const { 934 Node* in1 = in(1); 935 Node* in2 = in(2); 936 const Type* t1 = phase->type(in1); 937 const Type* t2 = phase->type(in2); 938 if (t1 == Type::TOP || t2 == Type::TOP) { 939 return Type::TOP; 940 } 941 // x ^ x ==> 0 942 if (in1->eqv_uncast(in2)) { 943 return add_id(); 944 } 945 // result of xor can only have bits sets where any of the 946 // inputs have bits set. lo can always become 0. 947 const TypeInt* t1i = t1->is_int(); 948 const TypeInt* t2i = t2->is_int(); 949 if ((t1i->_lo >= 0) && 950 (t1i->_hi > 0) && 951 (t2i->_lo >= 0) && 952 (t2i->_hi > 0)) { 953 // hi - set all bits below the highest bit. Using round_down to avoid overflow. 954 const TypeInt* t1x = TypeInt::make(0, round_down_power_of_2(t1i->_hi) + (round_down_power_of_2(t1i->_hi) - 1), t1i->_widen); 955 const TypeInt* t2x = TypeInt::make(0, round_down_power_of_2(t2i->_hi) + (round_down_power_of_2(t2i->_hi) - 1), t2i->_widen); 956 return t1x->meet(t2x); 957 } 958 return AddNode::Value(phase); 959 } 960 961 962 //------------------------------add_ring--------------------------------------- 963 // Supplied function returns the sum of the inputs IN THE CURRENT RING. For 964 // the logical operations the ring's ADD is really a logical OR function. 965 // This also type-checks the inputs for sanity. Guaranteed never to 966 // be passed a TOP or BOTTOM type, these are filtered out by pre-check. 967 const Type *XorINode::add_ring( const Type *t0, const Type *t1 ) const { 968 const TypeInt *r0 = t0->is_int(); // Handy access 969 const TypeInt *r1 = t1->is_int(); 970 971 // Complementing a boolean? 972 if( r0 == TypeInt::BOOL && ( r1 == TypeInt::ONE 973 || r1 == TypeInt::BOOL)) 974 return TypeInt::BOOL; 975 976 if( !r0->is_con() || !r1->is_con() ) // Not constants 977 return TypeInt::INT; // Any integer, but still no symbols. 978 979 // Otherwise just XOR them bits. 980 return TypeInt::make( r0->get_con() ^ r1->get_con() ); 981 } 982 983 //============================================================================= 984 //------------------------------add_ring--------------------------------------- 985 const Type *XorLNode::add_ring( const Type *t0, const Type *t1 ) const { 986 const TypeLong *r0 = t0->is_long(); // Handy access 987 const TypeLong *r1 = t1->is_long(); 988 989 // If either input is not a constant, just return all integers. 990 if( !r0->is_con() || !r1->is_con() ) 991 return TypeLong::LONG; // Any integer, but still no symbols. 992 993 // Otherwise just OR them bits. 994 return TypeLong::make( r0->get_con() ^ r1->get_con() ); 995 } 996 997 Node* XorLNode::Ideal(PhaseGVN* phase, bool can_reshape) { 998 Node* in1 = in(1); 999 Node* in2 = in(2); 1000 1001 // Convert ~x into -1-x when ~x is used in an arithmetic expression 1002 // or x itself is an arithmetic expression. 1003 if (phase->type(in2) == TypeLong::MINUS_1) { // follows LHS^(-1), i.e., ~LHS 1004 if (phase->is_IterGVN()) { 1005 if (is_used_in_only_arithmetic(this, T_LONG) 1006 // LHS is arithmetic 1007 || (in1->Opcode() == Op_AddL || in1->Opcode() == Op_SubL)) { 1008 return new SubLNode(in2, in1); 1009 } 1010 } else { 1011 // graph could be incomplete in GVN so we postpone to IGVN 1012 phase->record_for_igvn(this); 1013 } 1014 } 1015 return AddNode::Ideal(phase, can_reshape); 1016 } 1017 1018 const Type* XorLNode::Value(PhaseGVN* phase) const { 1019 Node* in1 = in(1); 1020 Node* in2 = in(2); 1021 const Type* t1 = phase->type(in1); 1022 const Type* t2 = phase->type(in2); 1023 if (t1 == Type::TOP || t2 == Type::TOP) { 1024 return Type::TOP; 1025 } 1026 // x ^ x ==> 0 1027 if (in1->eqv_uncast(in2)) { 1028 return add_id(); 1029 } 1030 // result of xor can only have bits sets where any of the 1031 // inputs have bits set. lo can always become 0. 1032 const TypeLong* t1l = t1->is_long(); 1033 const TypeLong* t2l = t2->is_long(); 1034 if ((t1l->_lo >= 0) && 1035 (t1l->_hi > 0) && 1036 (t2l->_lo >= 0) && 1037 (t2l->_hi > 0)) { 1038 // hi - set all bits below the highest bit. Using round_down to avoid overflow. 1039 const TypeLong* t1x = TypeLong::make(0, round_down_power_of_2(t1l->_hi) + (round_down_power_of_2(t1l->_hi) - 1), t1l->_widen); 1040 const TypeLong* t2x = TypeLong::make(0, round_down_power_of_2(t2l->_hi) + (round_down_power_of_2(t2l->_hi) - 1), t2l->_widen); 1041 return t1x->meet(t2x); 1042 } 1043 return AddNode::Value(phase); 1044 } 1045 1046 Node* build_min_max_int(Node* a, Node* b, bool is_max) { 1047 if (is_max) { 1048 return new MaxINode(a, b); 1049 } else { 1050 return new MinINode(a, b); 1051 } 1052 } 1053 1054 Node* MaxNode::build_min_max(Node* a, Node* b, bool is_max, bool is_unsigned, const Type* t, PhaseGVN& gvn) { 1055 bool is_int = gvn.type(a)->isa_int(); 1056 assert(is_int || gvn.type(a)->isa_long(), "int or long inputs"); 1057 assert(is_int == (gvn.type(b)->isa_int() != nullptr), "inconsistent inputs"); 1058 BasicType bt = is_int ? T_INT: T_LONG; 1059 Node* hook = nullptr; 1060 if (gvn.is_IterGVN()) { 1061 // Make sure a and b are not destroyed 1062 hook = new Node(2); 1063 hook->init_req(0, a); 1064 hook->init_req(1, b); 1065 } 1066 Node* res = nullptr; 1067 if (is_int && !is_unsigned) { 1068 res = gvn.transform(build_min_max_int(a, b, is_max)); 1069 assert(gvn.type(res)->is_int()->_lo >= t->is_int()->_lo && gvn.type(res)->is_int()->_hi <= t->is_int()->_hi, "type doesn't match"); 1070 } else { 1071 Node* cmp = nullptr; 1072 if (is_max) { 1073 cmp = gvn.transform(CmpNode::make(a, b, bt, is_unsigned)); 1074 } else { 1075 cmp = gvn.transform(CmpNode::make(b, a, bt, is_unsigned)); 1076 } 1077 Node* bol = gvn.transform(new BoolNode(cmp, BoolTest::lt)); 1078 res = gvn.transform(CMoveNode::make(nullptr, bol, a, b, t)); 1079 } 1080 if (hook != nullptr) { 1081 hook->destruct(&gvn); 1082 } 1083 return res; 1084 } 1085 1086 Node* MaxNode::build_min_max_diff_with_zero(Node* a, Node* b, bool is_max, const Type* t, PhaseGVN& gvn) { 1087 bool is_int = gvn.type(a)->isa_int(); 1088 assert(is_int || gvn.type(a)->isa_long(), "int or long inputs"); 1089 assert(is_int == (gvn.type(b)->isa_int() != nullptr), "inconsistent inputs"); 1090 BasicType bt = is_int ? T_INT: T_LONG; 1091 Node* zero = gvn.integercon(0, bt); 1092 Node* hook = nullptr; 1093 if (gvn.is_IterGVN()) { 1094 // Make sure a and b are not destroyed 1095 hook = new Node(2); 1096 hook->init_req(0, a); 1097 hook->init_req(1, b); 1098 } 1099 Node* cmp = nullptr; 1100 if (is_max) { 1101 cmp = gvn.transform(CmpNode::make(a, b, bt, false)); 1102 } else { 1103 cmp = gvn.transform(CmpNode::make(b, a, bt, false)); 1104 } 1105 Node* sub = gvn.transform(SubNode::make(a, b, bt)); 1106 Node* bol = gvn.transform(new BoolNode(cmp, BoolTest::lt)); 1107 Node* res = gvn.transform(CMoveNode::make(nullptr, bol, sub, zero, t)); 1108 if (hook != nullptr) { 1109 hook->destruct(&gvn); 1110 } 1111 return res; 1112 } 1113 1114 // Check if addition of an integer with type 't' and a constant 'c' can overflow. 1115 static bool can_overflow(const TypeInt* t, jint c) { 1116 jint t_lo = t->_lo; 1117 jint t_hi = t->_hi; 1118 return ((c < 0 && (java_add(t_lo, c) > t_lo)) || 1119 (c > 0 && (java_add(t_hi, c) < t_hi))); 1120 } 1121 1122 // Let <x, x_off> = x_operands and <y, y_off> = y_operands. 1123 // If x == y and neither add(x, x_off) nor add(y, y_off) overflow, return 1124 // add(x, op(x_off, y_off)). Otherwise, return nullptr. 1125 Node* MaxNode::extract_add(PhaseGVN* phase, ConstAddOperands x_operands, ConstAddOperands y_operands) { 1126 Node* x = x_operands.first; 1127 Node* y = y_operands.first; 1128 int opcode = Opcode(); 1129 assert(opcode == Op_MaxI || opcode == Op_MinI, "Unexpected opcode"); 1130 const TypeInt* tx = phase->type(x)->isa_int(); 1131 jint x_off = x_operands.second; 1132 jint y_off = y_operands.second; 1133 if (x == y && tx != nullptr && 1134 !can_overflow(tx, x_off) && 1135 !can_overflow(tx, y_off)) { 1136 jint c = opcode == Op_MinI ? MIN2(x_off, y_off) : MAX2(x_off, y_off); 1137 return new AddINode(x, phase->intcon(c)); 1138 } 1139 return nullptr; 1140 } 1141 1142 // Try to cast n as an integer addition with a constant. Return: 1143 // <x, C>, if n == add(x, C), where 'C' is a non-TOP constant; 1144 // <nullptr, 0>, if n == add(x, C), where 'C' is a TOP constant; or 1145 // <n, 0>, otherwise. 1146 static ConstAddOperands as_add_with_constant(Node* n) { 1147 if (n->Opcode() != Op_AddI) { 1148 return ConstAddOperands(n, 0); 1149 } 1150 Node* x = n->in(1); 1151 Node* c = n->in(2); 1152 if (!c->is_Con()) { 1153 return ConstAddOperands(n, 0); 1154 } 1155 const Type* c_type = c->bottom_type(); 1156 if (c_type == Type::TOP) { 1157 return ConstAddOperands(nullptr, 0); 1158 } 1159 return ConstAddOperands(x, c_type->is_int()->get_con()); 1160 } 1161 1162 Node* MaxNode::IdealI(PhaseGVN* phase, bool can_reshape) { 1163 int opcode = Opcode(); 1164 assert(opcode == Op_MinI || opcode == Op_MaxI, "Unexpected opcode"); 1165 // Try to transform the following pattern, in any of its four possible 1166 // permutations induced by op's commutativity: 1167 // op(op(add(inner, inner_off), inner_other), add(outer, outer_off)) 1168 // into 1169 // op(add(inner, op(inner_off, outer_off)), inner_other), 1170 // where: 1171 // op is either MinI or MaxI, and 1172 // inner == outer, and 1173 // the additions cannot overflow. 1174 for (uint inner_op_index = 1; inner_op_index <= 2; inner_op_index++) { 1175 if (in(inner_op_index)->Opcode() != opcode) { 1176 continue; 1177 } 1178 Node* outer_add = in(inner_op_index == 1 ? 2 : 1); 1179 ConstAddOperands outer_add_operands = as_add_with_constant(outer_add); 1180 if (outer_add_operands.first == nullptr) { 1181 return nullptr; // outer_add has a TOP input, no need to continue. 1182 } 1183 // One operand is a MinI/MaxI and the other is an integer addition with 1184 // constant. Test the operands of the inner MinI/MaxI. 1185 for (uint inner_add_index = 1; inner_add_index <= 2; inner_add_index++) { 1186 Node* inner_op = in(inner_op_index); 1187 Node* inner_add = inner_op->in(inner_add_index); 1188 ConstAddOperands inner_add_operands = as_add_with_constant(inner_add); 1189 if (inner_add_operands.first == nullptr) { 1190 return nullptr; // inner_add has a TOP input, no need to continue. 1191 } 1192 // Try to extract the inner add. 1193 Node* add_extracted = extract_add(phase, inner_add_operands, outer_add_operands); 1194 if (add_extracted == nullptr) { 1195 continue; 1196 } 1197 Node* add_transformed = phase->transform(add_extracted); 1198 Node* inner_other = inner_op->in(inner_add_index == 1 ? 2 : 1); 1199 return build_min_max_int(add_transformed, inner_other, opcode == Op_MaxI); 1200 } 1201 } 1202 // Try to transform 1203 // op(add(x, x_off), add(y, y_off)) 1204 // into 1205 // add(x, op(x_off, y_off)), 1206 // where: 1207 // op is either MinI or MaxI, and 1208 // inner == outer, and 1209 // the additions cannot overflow. 1210 ConstAddOperands xC = as_add_with_constant(in(1)); 1211 ConstAddOperands yC = as_add_with_constant(in(2)); 1212 if (xC.first == nullptr || yC.first == nullptr) return nullptr; 1213 return extract_add(phase, xC, yC); 1214 } 1215 1216 // Ideal transformations for MaxINode 1217 Node* MaxINode::Ideal(PhaseGVN* phase, bool can_reshape) { 1218 return IdealI(phase, can_reshape); 1219 } 1220 1221 //============================================================================= 1222 //------------------------------add_ring--------------------------------------- 1223 // Supplied function returns the sum of the inputs. 1224 const Type *MaxINode::add_ring( const Type *t0, const Type *t1 ) const { 1225 const TypeInt *r0 = t0->is_int(); // Handy access 1226 const TypeInt *r1 = t1->is_int(); 1227 1228 // Otherwise just MAX them bits. 1229 return TypeInt::make( MAX2(r0->_lo,r1->_lo), MAX2(r0->_hi,r1->_hi), MAX2(r0->_widen,r1->_widen) ); 1230 } 1231 1232 //============================================================================= 1233 //------------------------------Idealize--------------------------------------- 1234 // MINs show up in range-check loop limit calculations. Look for 1235 // "MIN2(x+c0,MIN2(y,x+c1))". Pick the smaller constant: "MIN2(x+c0,y)" 1236 Node* MinINode::Ideal(PhaseGVN* phase, bool can_reshape) { 1237 return IdealI(phase, can_reshape); 1238 } 1239 1240 //------------------------------add_ring--------------------------------------- 1241 // Supplied function returns the sum of the inputs. 1242 const Type *MinINode::add_ring( const Type *t0, const Type *t1 ) const { 1243 const TypeInt *r0 = t0->is_int(); // Handy access 1244 const TypeInt *r1 = t1->is_int(); 1245 1246 // Otherwise just MIN them bits. 1247 return TypeInt::make( MIN2(r0->_lo,r1->_lo), MIN2(r0->_hi,r1->_hi), MAX2(r0->_widen,r1->_widen) ); 1248 } 1249 1250 // Collapse the "addition with overflow-protection" pattern, and the symmetrical 1251 // "subtraction with underflow-protection" pattern. These are created during the 1252 // unrolling, when we have to adjust the limit by subtracting the stride, but want 1253 // to protect against underflow: MaxL(SubL(limit, stride), min_jint). 1254 // If we have more than one of those in a sequence: 1255 // 1256 // x con2 1257 // | | 1258 // AddL clamp2 1259 // | | 1260 // Max/MinL con1 1261 // | | 1262 // AddL clamp1 1263 // | | 1264 // Max/MinL (n) 1265 // 1266 // We want to collapse it to: 1267 // 1268 // x con1 con2 1269 // | | | 1270 // | AddLNode (new_con) 1271 // | | 1272 // AddLNode clamp1 1273 // | | 1274 // Max/MinL (n) 1275 // 1276 // Note: we assume that SubL was already replaced by an AddL, and that the stride 1277 // has its sign flipped: SubL(limit, stride) -> AddL(limit, -stride). 1278 Node* fold_subI_no_underflow_pattern(Node* n, PhaseGVN* phase) { 1279 assert(n->Opcode() == Op_MaxL || n->Opcode() == Op_MinL, "sanity"); 1280 // Check that the two clamps have the correct values. 1281 jlong clamp = (n->Opcode() == Op_MaxL) ? min_jint : max_jint; 1282 auto is_clamp = [&](Node* c) { 1283 const TypeLong* t = phase->type(c)->isa_long(); 1284 return t != nullptr && t->is_con() && 1285 t->get_con() == clamp; 1286 }; 1287 // Check that the constants are negative if MaxL, and positive if MinL. 1288 auto is_sub_con = [&](Node* c) { 1289 const TypeLong* t = phase->type(c)->isa_long(); 1290 return t != nullptr && t->is_con() && 1291 t->get_con() < max_jint && t->get_con() > min_jint && 1292 (t->get_con() < 0) == (n->Opcode() == Op_MaxL); 1293 }; 1294 // Verify the graph level by level: 1295 Node* add1 = n->in(1); 1296 Node* clamp1 = n->in(2); 1297 if (add1->Opcode() == Op_AddL && is_clamp(clamp1)) { 1298 Node* max2 = add1->in(1); 1299 Node* con1 = add1->in(2); 1300 if (max2->Opcode() == n->Opcode() && is_sub_con(con1)) { 1301 Node* add2 = max2->in(1); 1302 Node* clamp2 = max2->in(2); 1303 if (add2->Opcode() == Op_AddL && is_clamp(clamp2)) { 1304 Node* x = add2->in(1); 1305 Node* con2 = add2->in(2); 1306 if (is_sub_con(con2)) { 1307 Node* new_con = phase->transform(new AddLNode(con1, con2)); 1308 Node* new_sub = phase->transform(new AddLNode(x, new_con)); 1309 n->set_req_X(1, new_sub, phase); 1310 return n; 1311 } 1312 } 1313 } 1314 } 1315 return nullptr; 1316 } 1317 1318 const Type* MaxLNode::add_ring(const Type* t0, const Type* t1) const { 1319 const TypeLong* r0 = t0->is_long(); 1320 const TypeLong* r1 = t1->is_long(); 1321 1322 return TypeLong::make(MAX2(r0->_lo, r1->_lo), MAX2(r0->_hi, r1->_hi), MAX2(r0->_widen, r1->_widen)); 1323 } 1324 1325 Node* MaxLNode::Identity(PhaseGVN* phase) { 1326 const TypeLong* t1 = phase->type(in(1))->is_long(); 1327 const TypeLong* t2 = phase->type(in(2))->is_long(); 1328 1329 // Can we determine maximum statically? 1330 if (t1->_lo >= t2->_hi) { 1331 return in(1); 1332 } else if (t2->_lo >= t1->_hi) { 1333 return in(2); 1334 } 1335 1336 return MaxNode::Identity(phase); 1337 } 1338 1339 Node* MaxLNode::Ideal(PhaseGVN* phase, bool can_reshape) { 1340 Node* n = AddNode::Ideal(phase, can_reshape); 1341 if (n != nullptr) { 1342 return n; 1343 } 1344 if (can_reshape) { 1345 return fold_subI_no_underflow_pattern(this, phase); 1346 } 1347 return nullptr; 1348 } 1349 1350 const Type* MinLNode::add_ring(const Type* t0, const Type* t1) const { 1351 const TypeLong* r0 = t0->is_long(); 1352 const TypeLong* r1 = t1->is_long(); 1353 1354 return TypeLong::make(MIN2(r0->_lo, r1->_lo), MIN2(r0->_hi, r1->_hi), MIN2(r0->_widen, r1->_widen)); 1355 } 1356 1357 Node* MinLNode::Identity(PhaseGVN* phase) { 1358 const TypeLong* t1 = phase->type(in(1))->is_long(); 1359 const TypeLong* t2 = phase->type(in(2))->is_long(); 1360 1361 // Can we determine minimum statically? 1362 if (t1->_lo >= t2->_hi) { 1363 return in(2); 1364 } else if (t2->_lo >= t1->_hi) { 1365 return in(1); 1366 } 1367 1368 return MaxNode::Identity(phase); 1369 } 1370 1371 Node* MinLNode::Ideal(PhaseGVN* phase, bool can_reshape) { 1372 Node* n = AddNode::Ideal(phase, can_reshape); 1373 if (n != nullptr) { 1374 return n; 1375 } 1376 if (can_reshape) { 1377 return fold_subI_no_underflow_pattern(this, phase); 1378 } 1379 return nullptr; 1380 } 1381 1382 //------------------------------add_ring--------------------------------------- 1383 const Type* MinFNode::add_ring(const Type* t0, const Type* t1 ) const { 1384 const TypeF* r0 = t0->isa_float_constant(); 1385 const TypeF* r1 = t1->isa_float_constant(); 1386 if (r0 == nullptr || r1 == nullptr) { 1387 return bottom_type(); 1388 } 1389 1390 if (r0->is_nan()) { 1391 return r0; 1392 } 1393 if (r1->is_nan()) { 1394 return r1; 1395 } 1396 1397 float f0 = r0->getf(); 1398 float f1 = r1->getf(); 1399 if (f0 != 0.0f || f1 != 0.0f) { 1400 return f0 < f1 ? r0 : r1; 1401 } 1402 1403 // handle min of 0.0, -0.0 case. 1404 return (jint_cast(f0) < jint_cast(f1)) ? r0 : r1; 1405 } 1406 1407 //------------------------------add_ring--------------------------------------- 1408 const Type* MinDNode::add_ring(const Type* t0, const Type* t1) const { 1409 const TypeD* r0 = t0->isa_double_constant(); 1410 const TypeD* r1 = t1->isa_double_constant(); 1411 if (r0 == nullptr || r1 == nullptr) { 1412 return bottom_type(); 1413 } 1414 1415 if (r0->is_nan()) { 1416 return r0; 1417 } 1418 if (r1->is_nan()) { 1419 return r1; 1420 } 1421 1422 double d0 = r0->getd(); 1423 double d1 = r1->getd(); 1424 if (d0 != 0.0 || d1 != 0.0) { 1425 return d0 < d1 ? r0 : r1; 1426 } 1427 1428 // handle min of 0.0, -0.0 case. 1429 return (jlong_cast(d0) < jlong_cast(d1)) ? r0 : r1; 1430 } 1431 1432 //------------------------------add_ring--------------------------------------- 1433 const Type* MaxFNode::add_ring(const Type* t0, const Type* t1) const { 1434 const TypeF* r0 = t0->isa_float_constant(); 1435 const TypeF* r1 = t1->isa_float_constant(); 1436 if (r0 == nullptr || r1 == nullptr) { 1437 return bottom_type(); 1438 } 1439 1440 if (r0->is_nan()) { 1441 return r0; 1442 } 1443 if (r1->is_nan()) { 1444 return r1; 1445 } 1446 1447 float f0 = r0->getf(); 1448 float f1 = r1->getf(); 1449 if (f0 != 0.0f || f1 != 0.0f) { 1450 return f0 > f1 ? r0 : r1; 1451 } 1452 1453 // handle max of 0.0,-0.0 case. 1454 return (jint_cast(f0) > jint_cast(f1)) ? r0 : r1; 1455 } 1456 1457 //------------------------------add_ring--------------------------------------- 1458 const Type* MaxDNode::add_ring(const Type* t0, const Type* t1) const { 1459 const TypeD* r0 = t0->isa_double_constant(); 1460 const TypeD* r1 = t1->isa_double_constant(); 1461 if (r0 == nullptr || r1 == nullptr) { 1462 return bottom_type(); 1463 } 1464 1465 if (r0->is_nan()) { 1466 return r0; 1467 } 1468 if (r1->is_nan()) { 1469 return r1; 1470 } 1471 1472 double d0 = r0->getd(); 1473 double d1 = r1->getd(); 1474 if (d0 != 0.0 || d1 != 0.0) { 1475 return d0 > d1 ? r0 : r1; 1476 } 1477 1478 // handle max of 0.0, -0.0 case. 1479 return (jlong_cast(d0) > jlong_cast(d1)) ? r0 : r1; 1480 }