1 /* 2 * Copyright (c) 1997, 2024, 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/connode.hpp" 29 #include "opto/convertnode.hpp" 30 #include "opto/memnode.hpp" 31 #include "opto/mulnode.hpp" 32 #include "opto/phaseX.hpp" 33 #include "opto/subnode.hpp" 34 #include "utilities/powerOfTwo.hpp" 35 36 // Portions of code courtesy of Clifford Click 37 38 39 //============================================================================= 40 //------------------------------hash------------------------------------------- 41 // Hash function over MulNodes. Needs to be commutative; i.e., I swap 42 // (commute) inputs to MulNodes willy-nilly so the hash function must return 43 // the same value in the presence of edge swapping. 44 uint MulNode::hash() const { 45 return (uintptr_t)in(1) + (uintptr_t)in(2) + Opcode(); 46 } 47 48 //------------------------------Identity--------------------------------------- 49 // Multiplying a one preserves the other argument 50 Node* MulNode::Identity(PhaseGVN* phase) { 51 const Type *one = mul_id(); // The multiplicative identity 52 if( phase->type( in(1) )->higher_equal( one ) ) return in(2); 53 if( phase->type( in(2) )->higher_equal( one ) ) return in(1); 54 55 return this; 56 } 57 58 //------------------------------Ideal------------------------------------------ 59 // We also canonicalize the Node, moving constants to the right input, 60 // and flatten expressions (so that 1+x+2 becomes x+3). 61 Node *MulNode::Ideal(PhaseGVN *phase, bool can_reshape) { 62 Node* in1 = in(1); 63 Node* in2 = in(2); 64 Node* progress = nullptr; // Progress flag 65 66 // This code is used by And nodes too, but some conversions are 67 // only valid for the actual Mul nodes. 68 uint op = Opcode(); 69 bool real_mul = (op == Op_MulI) || (op == Op_MulL) || 70 (op == Op_MulF) || (op == Op_MulD); 71 72 // Convert "(-a)*(-b)" into "a*b". 73 if (real_mul && in1->is_Sub() && in2->is_Sub()) { 74 if (phase->type(in1->in(1))->is_zero_type() && 75 phase->type(in2->in(1))->is_zero_type()) { 76 set_req_X(1, in1->in(2), phase); 77 set_req_X(2, in2->in(2), phase); 78 in1 = in(1); 79 in2 = in(2); 80 progress = this; 81 } 82 } 83 84 // convert "max(a,b) * min(a,b)" into "a*b". 85 if ((in(1)->Opcode() == max_opcode() && in(2)->Opcode() == min_opcode()) 86 || (in(1)->Opcode() == min_opcode() && in(2)->Opcode() == max_opcode())) { 87 Node *in11 = in(1)->in(1); 88 Node *in12 = in(1)->in(2); 89 90 Node *in21 = in(2)->in(1); 91 Node *in22 = in(2)->in(2); 92 93 if ((in11 == in21 && in12 == in22) || 94 (in11 == in22 && in12 == in21)) { 95 set_req_X(1, in11, phase); 96 set_req_X(2, in12, phase); 97 in1 = in(1); 98 in2 = in(2); 99 progress = this; 100 } 101 } 102 103 const Type* t1 = phase->type(in1); 104 const Type* t2 = phase->type(in2); 105 106 // We are OK if right is a constant, or right is a load and 107 // left is a non-constant. 108 if( !(t2->singleton() || 109 (in(2)->is_Load() && !(t1->singleton() || in(1)->is_Load())) ) ) { 110 if( t1->singleton() || // Left input is a constant? 111 // Otherwise, sort inputs (commutativity) to help value numbering. 112 (in(1)->_idx > in(2)->_idx) ) { 113 swap_edges(1, 2); 114 const Type *t = t1; 115 t1 = t2; 116 t2 = t; 117 progress = this; // Made progress 118 } 119 } 120 121 // If the right input is a constant, and the left input is a product of a 122 // constant, flatten the expression tree. 123 if( t2->singleton() && // Right input is a constant? 124 op != Op_MulF && // Float & double cannot reassociate 125 op != Op_MulD ) { 126 if( t2 == Type::TOP ) return nullptr; 127 Node *mul1 = in(1); 128 #ifdef ASSERT 129 // Check for dead loop 130 int op1 = mul1->Opcode(); 131 if ((mul1 == this) || (in(2) == this) || 132 ((op1 == mul_opcode() || op1 == add_opcode()) && 133 ((mul1->in(1) == this) || (mul1->in(2) == this) || 134 (mul1->in(1) == mul1) || (mul1->in(2) == mul1)))) { 135 assert(false, "dead loop in MulNode::Ideal"); 136 } 137 #endif 138 139 if( mul1->Opcode() == mul_opcode() ) { // Left input is a multiply? 140 // Mul of a constant? 141 const Type *t12 = phase->type( mul1->in(2) ); 142 if( t12->singleton() && t12 != Type::TOP) { // Left input is an add of a constant? 143 // Compute new constant; check for overflow 144 const Type *tcon01 = ((MulNode*)mul1)->mul_ring(t2,t12); 145 if( tcon01->singleton() ) { 146 // The Mul of the flattened expression 147 set_req_X(1, mul1->in(1), phase); 148 set_req_X(2, phase->makecon(tcon01), phase); 149 t2 = tcon01; 150 progress = this; // Made progress 151 } 152 } 153 } 154 // If the right input is a constant, and the left input is an add of a 155 // constant, flatten the tree: (X+con1)*con0 ==> X*con0 + con1*con0 156 const Node *add1 = in(1); 157 if( add1->Opcode() == add_opcode() ) { // Left input is an add? 158 // Add of a constant? 159 const Type *t12 = phase->type( add1->in(2) ); 160 if( t12->singleton() && t12 != Type::TOP ) { // Left input is an add of a constant? 161 assert( add1->in(1) != add1, "dead loop in MulNode::Ideal" ); 162 // Compute new constant; check for overflow 163 const Type *tcon01 = mul_ring(t2,t12); 164 if( tcon01->singleton() ) { 165 166 // Convert (X+con1)*con0 into X*con0 167 Node *mul = clone(); // mul = ()*con0 168 mul->set_req(1,add1->in(1)); // mul = X*con0 169 mul = phase->transform(mul); 170 171 Node *add2 = add1->clone(); 172 add2->set_req(1, mul); // X*con0 + con0*con1 173 add2->set_req(2, phase->makecon(tcon01) ); 174 progress = add2; 175 } 176 } 177 } // End of is left input an add 178 } // End of is right input a Mul 179 180 return progress; 181 } 182 183 //------------------------------Value----------------------------------------- 184 const Type* MulNode::Value(PhaseGVN* phase) const { 185 const Type *t1 = phase->type( in(1) ); 186 const Type *t2 = phase->type( in(2) ); 187 // Either input is TOP ==> the result is TOP 188 if( t1 == Type::TOP ) return Type::TOP; 189 if( t2 == Type::TOP ) return Type::TOP; 190 191 // Either input is ZERO ==> the result is ZERO. 192 // Not valid for floats or doubles since +0.0 * -0.0 --> +0.0 193 int op = Opcode(); 194 if( op == Op_MulI || op == Op_AndI || op == Op_MulL || op == Op_AndL ) { 195 const Type *zero = add_id(); // The multiplicative zero 196 if( t1->higher_equal( zero ) ) return zero; 197 if( t2->higher_equal( zero ) ) return zero; 198 } 199 200 // Code pattern on return from a call that returns an __Value. Can 201 // be optimized away if the return value turns out to be an oop. 202 if (op == Op_AndX && 203 in(1) != nullptr && 204 in(1)->Opcode() == Op_CastP2X && 205 in(1)->in(1) != nullptr && 206 phase->type(in(1)->in(1))->isa_oopptr() && 207 t2->isa_intptr_t()->_lo >= 0 && 208 t2->isa_intptr_t()->_hi <= MinObjAlignmentInBytesMask) { 209 return add_id(); 210 } 211 212 // Either input is BOTTOM ==> the result is the local BOTTOM 213 if( t1 == Type::BOTTOM || t2 == Type::BOTTOM ) 214 return bottom_type(); 215 216 #if defined(IA32) 217 // Can't trust native compilers to properly fold strict double 218 // multiplication with round-to-zero on this platform. 219 if (op == Op_MulD) { 220 return TypeD::DOUBLE; 221 } 222 #endif 223 224 return mul_ring(t1,t2); // Local flavor of type multiplication 225 } 226 227 MulNode* MulNode::make(Node* in1, Node* in2, BasicType bt) { 228 switch (bt) { 229 case T_INT: 230 return new MulINode(in1, in2); 231 case T_LONG: 232 return new MulLNode(in1, in2); 233 default: 234 fatal("Not implemented for %s", type2name(bt)); 235 } 236 return nullptr; 237 } 238 239 240 //============================================================================= 241 //------------------------------Ideal------------------------------------------ 242 // Check for power-of-2 multiply, then try the regular MulNode::Ideal 243 Node *MulINode::Ideal(PhaseGVN *phase, bool can_reshape) { 244 const jint con = in(2)->find_int_con(0); 245 if (con == 0) { 246 // If in(2) is not a constant, call Ideal() of the parent class to 247 // try to move constant to the right side. 248 return MulNode::Ideal(phase, can_reshape); 249 } 250 251 // Now we have a constant Node on the right and the constant in con. 252 if (con == 1) { 253 // By one is handled by Identity call 254 return nullptr; 255 } 256 257 // Check for negative constant; if so negate the final result 258 bool sign_flip = false; 259 260 unsigned int abs_con = uabs(con); 261 if (abs_con != (unsigned int)con) { 262 sign_flip = true; 263 } 264 265 // Get low bit; check for being the only bit 266 Node *res = nullptr; 267 unsigned int bit1 = submultiple_power_of_2(abs_con); 268 if (bit1 == abs_con) { // Found a power of 2? 269 res = new LShiftINode(in(1), phase->intcon(log2i_exact(bit1))); 270 } else { 271 // Check for constant with 2 bits set 272 unsigned int bit2 = abs_con - bit1; 273 bit2 = bit2 & (0 - bit2); // Extract 2nd bit 274 if (bit2 + bit1 == abs_con) { // Found all bits in con? 275 Node *n1 = phase->transform(new LShiftINode(in(1), phase->intcon(log2i_exact(bit1)))); 276 Node *n2 = phase->transform(new LShiftINode(in(1), phase->intcon(log2i_exact(bit2)))); 277 res = new AddINode(n2, n1); 278 } else if (is_power_of_2(abs_con + 1)) { 279 // Sleezy: power-of-2 - 1. Next time be generic. 280 unsigned int temp = abs_con + 1; 281 Node *n1 = phase->transform(new LShiftINode(in(1), phase->intcon(log2i_exact(temp)))); 282 res = new SubINode(n1, in(1)); 283 } else { 284 return MulNode::Ideal(phase, can_reshape); 285 } 286 } 287 288 if (sign_flip) { // Need to negate result? 289 res = phase->transform(res);// Transform, before making the zero con 290 res = new SubINode(phase->intcon(0),res); 291 } 292 293 return res; // Return final result 294 } 295 296 // This template class performs type multiplication for MulI/MulLNode. NativeType is either jint or jlong. 297 // In this class, the inputs of the MulNodes are named left and right with types [left_lo,left_hi] and [right_lo,right_hi]. 298 // 299 // In general, the multiplication of two x-bit values could produce a result that consumes up to 2x bits if there is 300 // enough space to hold them all. We can therefore distinguish the following two cases for the product: 301 // - no overflow (i.e. product fits into x bits) 302 // - overflow (i.e. product does not fit into x bits) 303 // 304 // When multiplying the two x-bit inputs 'left' and 'right' with their x-bit types [left_lo,left_hi] and [right_lo,right_hi] 305 // we need to find the minimum and maximum of all possible products to define a new type. To do that, we compute the 306 // cross product of [left_lo,left_hi] and [right_lo,right_hi] in 2x-bit space where no over- or underflow can happen. 307 // The cross product consists of the following four multiplications with 2x-bit results: 308 // (1) left_lo * right_lo 309 // (2) left_lo * right_hi 310 // (3) left_hi * right_lo 311 // (4) left_hi * right_hi 312 // 313 // Let's define the following two functions: 314 // - Lx(i): Returns the lower x bits of the 2x-bit number i. 315 // - Ux(i): Returns the upper x bits of the 2x-bit number i. 316 // 317 // Let's first assume all products are positive where only overflows are possible but no underflows. If there is no 318 // overflow for a product p, then the upper x bits of the 2x-bit result p are all zero: 319 // Ux(p) = 0 320 // Lx(p) = p 321 // 322 // If none of the multiplications (1)-(4) overflow, we can truncate the upper x bits and use the following result type 323 // with x bits: 324 // [result_lo,result_hi] = [MIN(Lx(1),Lx(2),Lx(3),Lx(4)),MAX(Lx(1),Lx(2),Lx(3),Lx(4))] 325 // 326 // If any of these multiplications overflows, we could pessimistically take the bottom type for the x bit result 327 // (i.e. all values in the x-bit space could be possible): 328 // [result_lo,result_hi] = [NativeType_min,NativeType_max] 329 // 330 // However, in case of any overflow, we can do better by analyzing the upper x bits of all multiplications (1)-(4) with 331 // 2x-bit results. The upper x bits tell us something about how many times a multiplication has overflown the lower 332 // x bits. If the upper x bits of (1)-(4) are all equal, then we know that all of these multiplications overflowed 333 // the lower x bits the same number of times: 334 // Ux((1)) = Ux((2)) = Ux((3)) = Ux((4)) 335 // 336 // If all upper x bits are equal, we can conclude: 337 // Lx(MIN((1),(2),(3),(4))) = MIN(Lx(1),Lx(2),Lx(3),Lx(4))) 338 // Lx(MAX((1),(2),(3),(4))) = MAX(Lx(1),Lx(2),Lx(3),Lx(4))) 339 // 340 // Therefore, we can use the same precise x-bit result type as for the no-overflow case: 341 // [result_lo,result_hi] = [(MIN(Lx(1),Lx(2),Lx(3),Lx(4))),MAX(Lx(1),Lx(2),Lx(3),Lx(4)))] 342 // 343 // 344 // Now let's assume that (1)-(4) are signed multiplications where over- and underflow could occur: 345 // Negative numbers are all sign extend with ones. Therefore, if a negative product does not underflow, then the 346 // upper x bits of the 2x-bit result are all set to ones which is minus one in two's complement. If there is an underflow, 347 // the upper x bits are decremented by the number of times an underflow occurred. The smallest possible negative product 348 // is NativeType_min*NativeType_max, where the upper x bits are set to NativeType_min / 2 (b11...0). It is therefore 349 // impossible to underflow the upper x bits. Thus, when having all ones (i.e. minus one) in the upper x bits, we know 350 // that there is no underflow. 351 // 352 // To be able to compare the number of over-/underflows of positive and negative products, respectively, we normalize 353 // the upper x bits of negative 2x-bit products by adding one. This way a product has no over- or underflow if the 354 // normalized upper x bits are zero. Now we can use the same improved type as for strictly positive products because we 355 // can compare the upper x bits in a unified way with N() being the normalization function: 356 // N(Ux((1))) = N(Ux((2))) = N(Ux((3)) = N(Ux((4))) 357 template<typename NativeType> 358 class IntegerTypeMultiplication { 359 360 NativeType _lo_left; 361 NativeType _lo_right; 362 NativeType _hi_left; 363 NativeType _hi_right; 364 short _widen_left; 365 short _widen_right; 366 367 static const Type* overflow_type(); 368 static NativeType multiply_high(NativeType x, NativeType y); 369 const Type* create_type(NativeType lo, NativeType hi) const; 370 371 static NativeType multiply_high_signed_overflow_value(NativeType x, NativeType y) { 372 return normalize_overflow_value(x, y, multiply_high(x, y)); 373 } 374 375 bool cross_product_not_same_overflow_value() const { 376 const NativeType lo_lo_high_product = multiply_high_signed_overflow_value(_lo_left, _lo_right); 377 const NativeType lo_hi_high_product = multiply_high_signed_overflow_value(_lo_left, _hi_right); 378 const NativeType hi_lo_high_product = multiply_high_signed_overflow_value(_hi_left, _lo_right); 379 const NativeType hi_hi_high_product = multiply_high_signed_overflow_value(_hi_left, _hi_right); 380 return lo_lo_high_product != lo_hi_high_product || 381 lo_hi_high_product != hi_lo_high_product || 382 hi_lo_high_product != hi_hi_high_product; 383 } 384 385 bool does_product_overflow(NativeType x, NativeType y) const { 386 return multiply_high_signed_overflow_value(x, y) != 0; 387 } 388 389 static NativeType normalize_overflow_value(const NativeType x, const NativeType y, NativeType result) { 390 return java_multiply(x, y) < 0 ? result + 1 : result; 391 } 392 393 public: 394 template<class IntegerType> 395 IntegerTypeMultiplication(const IntegerType* left, const IntegerType* right) 396 : _lo_left(left->_lo), _lo_right(right->_lo), 397 _hi_left(left->_hi), _hi_right(right->_hi), 398 _widen_left(left->_widen), _widen_right(right->_widen) {} 399 400 // Compute the product type by multiplying the two input type ranges. We take the minimum and maximum of all possible 401 // values (requires 4 multiplications of all possible combinations of the two range boundary values). If any of these 402 // multiplications overflows/underflows, we need to make sure that they all have the same number of overflows/underflows 403 // If that is not the case, we return the bottom type to cover all values due to the inconsistent overflows/underflows). 404 const Type* compute() const { 405 if (cross_product_not_same_overflow_value()) { 406 return overflow_type(); 407 } 408 409 NativeType lo_lo_product = java_multiply(_lo_left, _lo_right); 410 NativeType lo_hi_product = java_multiply(_lo_left, _hi_right); 411 NativeType hi_lo_product = java_multiply(_hi_left, _lo_right); 412 NativeType hi_hi_product = java_multiply(_hi_left, _hi_right); 413 const NativeType min = MIN4(lo_lo_product, lo_hi_product, hi_lo_product, hi_hi_product); 414 const NativeType max = MAX4(lo_lo_product, lo_hi_product, hi_lo_product, hi_hi_product); 415 return create_type(min, max); 416 } 417 418 bool does_overflow() const { 419 return does_product_overflow(_lo_left, _lo_right) || 420 does_product_overflow(_lo_left, _hi_right) || 421 does_product_overflow(_hi_left, _lo_right) || 422 does_product_overflow(_hi_left, _hi_right); 423 } 424 }; 425 426 template <> 427 const Type* IntegerTypeMultiplication<jint>::overflow_type() { 428 return TypeInt::INT; 429 } 430 431 template <> 432 jint IntegerTypeMultiplication<jint>::multiply_high(const jint x, const jint y) { 433 const jlong x_64 = x; 434 const jlong y_64 = y; 435 const jlong product = x_64 * y_64; 436 return (jint)((uint64_t)product >> 32u); 437 } 438 439 template <> 440 const Type* IntegerTypeMultiplication<jint>::create_type(jint lo, jint hi) const { 441 return TypeInt::make(lo, hi, MAX2(_widen_left, _widen_right)); 442 } 443 444 template <> 445 const Type* IntegerTypeMultiplication<jlong>::overflow_type() { 446 return TypeLong::LONG; 447 } 448 449 template <> 450 jlong IntegerTypeMultiplication<jlong>::multiply_high(const jlong x, const jlong y) { 451 return multiply_high_signed(x, y); 452 } 453 454 template <> 455 const Type* IntegerTypeMultiplication<jlong>::create_type(jlong lo, jlong hi) const { 456 return TypeLong::make(lo, hi, MAX2(_widen_left, _widen_right)); 457 } 458 459 // Compute the product type of two integer ranges into this node. 460 const Type* MulINode::mul_ring(const Type* type_left, const Type* type_right) const { 461 const IntegerTypeMultiplication<jint> integer_multiplication(type_left->is_int(), type_right->is_int()); 462 return integer_multiplication.compute(); 463 } 464 465 bool MulINode::does_overflow(const TypeInt* type_left, const TypeInt* type_right) { 466 const IntegerTypeMultiplication<jint> integer_multiplication(type_left, type_right); 467 return integer_multiplication.does_overflow(); 468 } 469 470 // Compute the product type of two long ranges into this node. 471 const Type* MulLNode::mul_ring(const Type* type_left, const Type* type_right) const { 472 const IntegerTypeMultiplication<jlong> integer_multiplication(type_left->is_long(), type_right->is_long()); 473 return integer_multiplication.compute(); 474 } 475 476 //============================================================================= 477 //------------------------------Ideal------------------------------------------ 478 // Check for power-of-2 multiply, then try the regular MulNode::Ideal 479 Node *MulLNode::Ideal(PhaseGVN *phase, bool can_reshape) { 480 const jlong con = in(2)->find_long_con(0); 481 if (con == 0) { 482 // If in(2) is not a constant, call Ideal() of the parent class to 483 // try to move constant to the right side. 484 return MulNode::Ideal(phase, can_reshape); 485 } 486 487 // Now we have a constant Node on the right and the constant in con. 488 if (con == 1) { 489 // By one is handled by Identity call 490 return nullptr; 491 } 492 493 // Check for negative constant; if so negate the final result 494 bool sign_flip = false; 495 julong abs_con = uabs(con); 496 if (abs_con != (julong)con) { 497 sign_flip = true; 498 } 499 500 // Get low bit; check for being the only bit 501 Node *res = nullptr; 502 julong bit1 = submultiple_power_of_2(abs_con); 503 if (bit1 == abs_con) { // Found a power of 2? 504 res = new LShiftLNode(in(1), phase->intcon(log2i_exact(bit1))); 505 } else { 506 507 // Check for constant with 2 bits set 508 julong bit2 = abs_con-bit1; 509 bit2 = bit2 & (0-bit2); // Extract 2nd bit 510 if (bit2 + bit1 == abs_con) { // Found all bits in con? 511 Node *n1 = phase->transform(new LShiftLNode(in(1), phase->intcon(log2i_exact(bit1)))); 512 Node *n2 = phase->transform(new LShiftLNode(in(1), phase->intcon(log2i_exact(bit2)))); 513 res = new AddLNode(n2, n1); 514 515 } else if (is_power_of_2(abs_con+1)) { 516 // Sleezy: power-of-2 -1. Next time be generic. 517 julong temp = abs_con + 1; 518 Node *n1 = phase->transform( new LShiftLNode(in(1), phase->intcon(log2i_exact(temp)))); 519 res = new SubLNode(n1, in(1)); 520 } else { 521 return MulNode::Ideal(phase, can_reshape); 522 } 523 } 524 525 if (sign_flip) { // Need to negate result? 526 res = phase->transform(res);// Transform, before making the zero con 527 res = new SubLNode(phase->longcon(0),res); 528 } 529 530 return res; // Return final result 531 } 532 533 //============================================================================= 534 //------------------------------mul_ring--------------------------------------- 535 // Compute the product type of two double ranges into this node. 536 const Type *MulFNode::mul_ring(const Type *t0, const Type *t1) const { 537 if( t0 == Type::FLOAT || t1 == Type::FLOAT ) return Type::FLOAT; 538 return TypeF::make( t0->getf() * t1->getf() ); 539 } 540 541 //------------------------------Ideal--------------------------------------- 542 // Check to see if we are multiplying by a constant 2 and convert to add, then try the regular MulNode::Ideal 543 Node* MulFNode::Ideal(PhaseGVN* phase, bool can_reshape) { 544 const TypeF *t2 = phase->type(in(2))->isa_float_constant(); 545 546 // x * 2 -> x + x 547 if (t2 != nullptr && t2->getf() == 2) { 548 Node* base = in(1); 549 return new AddFNode(base, base); 550 } 551 552 return MulNode::Ideal(phase, can_reshape); 553 } 554 555 //============================================================================= 556 //------------------------------mul_ring--------------------------------------- 557 // Compute the product type of two double ranges into this node. 558 const Type *MulDNode::mul_ring(const Type *t0, const Type *t1) const { 559 if( t0 == Type::DOUBLE || t1 == Type::DOUBLE ) return Type::DOUBLE; 560 // We must be multiplying 2 double constants. 561 return TypeD::make( t0->getd() * t1->getd() ); 562 } 563 564 //------------------------------Ideal--------------------------------------- 565 // Check to see if we are multiplying by a constant 2 and convert to add, then try the regular MulNode::Ideal 566 Node* MulDNode::Ideal(PhaseGVN* phase, bool can_reshape) { 567 const TypeD *t2 = phase->type(in(2))->isa_double_constant(); 568 569 // x * 2 -> x + x 570 if (t2 != nullptr && t2->getd() == 2) { 571 Node* base = in(1); 572 return new AddDNode(base, base); 573 } 574 575 return MulNode::Ideal(phase, can_reshape); 576 } 577 578 //============================================================================= 579 //------------------------------Value------------------------------------------ 580 const Type* MulHiLNode::Value(PhaseGVN* phase) const { 581 const Type *t1 = phase->type( in(1) ); 582 const Type *t2 = phase->type( in(2) ); 583 const Type *bot = bottom_type(); 584 return MulHiValue(t1, t2, bot); 585 } 586 587 const Type* UMulHiLNode::Value(PhaseGVN* phase) const { 588 const Type *t1 = phase->type( in(1) ); 589 const Type *t2 = phase->type( in(2) ); 590 const Type *bot = bottom_type(); 591 return MulHiValue(t1, t2, bot); 592 } 593 594 // A common routine used by UMulHiLNode and MulHiLNode 595 const Type* MulHiValue(const Type *t1, const Type *t2, const Type *bot) { 596 // Either input is TOP ==> the result is TOP 597 if( t1 == Type::TOP ) return Type::TOP; 598 if( t2 == Type::TOP ) return Type::TOP; 599 600 // Either input is BOTTOM ==> the result is the local BOTTOM 601 if( (t1 == bot) || (t2 == bot) || 602 (t1 == Type::BOTTOM) || (t2 == Type::BOTTOM) ) 603 return bot; 604 605 // It is not worth trying to constant fold this stuff! 606 return TypeLong::LONG; 607 } 608 609 //============================================================================= 610 //------------------------------mul_ring--------------------------------------- 611 // Supplied function returns the product of the inputs IN THE CURRENT RING. 612 // For the logical operations the ring's MUL is really a logical AND function. 613 // This also type-checks the inputs for sanity. Guaranteed never to 614 // be passed a TOP or BOTTOM type, these are filtered out by pre-check. 615 const Type *AndINode::mul_ring( const Type *t0, const Type *t1 ) const { 616 const TypeInt *r0 = t0->is_int(); // Handy access 617 const TypeInt *r1 = t1->is_int(); 618 int widen = MAX2(r0->_widen,r1->_widen); 619 620 // If either input is a constant, might be able to trim cases 621 if( !r0->is_con() && !r1->is_con() ) 622 return TypeInt::INT; // No constants to be had 623 624 // Both constants? Return bits 625 if( r0->is_con() && r1->is_con() ) 626 return TypeInt::make( r0->get_con() & r1->get_con() ); 627 628 if( r0->is_con() && r0->get_con() > 0 ) 629 return TypeInt::make(0, r0->get_con(), widen); 630 631 if( r1->is_con() && r1->get_con() > 0 ) 632 return TypeInt::make(0, r1->get_con(), widen); 633 634 if( r0 == TypeInt::BOOL || r1 == TypeInt::BOOL ) { 635 return TypeInt::BOOL; 636 } 637 638 return TypeInt::INT; // No constants to be had 639 } 640 641 const Type* AndINode::Value(PhaseGVN* phase) const { 642 // patterns similar to (v << 2) & 3 643 if (AndIL_shift_and_mask_is_always_zero(phase, in(1), in(2), T_INT, true)) { 644 return TypeInt::ZERO; 645 } 646 647 return MulNode::Value(phase); 648 } 649 650 //------------------------------Identity--------------------------------------- 651 // Masking off the high bits of an unsigned load is not required 652 Node* AndINode::Identity(PhaseGVN* phase) { 653 654 // x & x => x 655 if (in(1) == in(2)) { 656 return in(1); 657 } 658 659 Node* in1 = in(1); 660 uint op = in1->Opcode(); 661 const TypeInt* t2 = phase->type(in(2))->isa_int(); 662 if (t2 && t2->is_con()) { 663 int con = t2->get_con(); 664 // Masking off high bits which are always zero is useless. 665 const TypeInt* t1 = phase->type(in(1))->isa_int(); 666 if (t1 != nullptr && t1->_lo >= 0) { 667 jint t1_support = right_n_bits(1 + log2i_graceful(t1->_hi)); 668 if ((t1_support & con) == t1_support) 669 return in1; 670 } 671 // Masking off the high bits of a unsigned-shift-right is not 672 // needed either. 673 if (op == Op_URShiftI) { 674 const TypeInt* t12 = phase->type(in1->in(2))->isa_int(); 675 if (t12 && t12->is_con()) { // Shift is by a constant 676 int shift = t12->get_con(); 677 shift &= BitsPerJavaInteger - 1; // semantics of Java shifts 678 int mask = max_juint >> shift; 679 if ((mask & con) == mask) // If AND is useless, skip it 680 return in1; 681 } 682 } 683 } 684 return MulNode::Identity(phase); 685 } 686 687 //------------------------------Ideal------------------------------------------ 688 Node *AndINode::Ideal(PhaseGVN *phase, bool can_reshape) { 689 // pattern similar to (v1 + (v2 << 2)) & 3 transformed to v1 & 3 690 Node* progress = AndIL_add_shift_and_mask(phase, T_INT); 691 if (progress != nullptr) { 692 return progress; 693 } 694 695 // Convert "(~a) & (~b)" into "~(a | b)" 696 if (AddNode::is_not(phase, in(1), T_INT) && AddNode::is_not(phase, in(2), T_INT)) { 697 Node* or_a_b = new OrINode(in(1)->in(1), in(2)->in(1)); 698 Node* tn = phase->transform(or_a_b); 699 return AddNode::make_not(phase, tn, T_INT); 700 } 701 702 // Special case constant AND mask 703 const TypeInt *t2 = phase->type( in(2) )->isa_int(); 704 if( !t2 || !t2->is_con() ) return MulNode::Ideal(phase, can_reshape); 705 const int mask = t2->get_con(); 706 Node *load = in(1); 707 uint lop = load->Opcode(); 708 709 // Masking bits off of a Character? Hi bits are already zero. 710 if( lop == Op_LoadUS && 711 (mask & 0xFFFF0000) ) // Can we make a smaller mask? 712 return new AndINode(load,phase->intcon(mask&0xFFFF)); 713 714 // Masking bits off of a Short? Loading a Character does some masking 715 if (can_reshape && 716 load->outcnt() == 1 && load->unique_out() == this) { 717 if (lop == Op_LoadS && (mask & 0xFFFF0000) == 0 ) { 718 Node* ldus = load->as_Load()->convert_to_unsigned_load(*phase); 719 ldus = phase->transform(ldus); 720 return new AndINode(ldus, phase->intcon(mask & 0xFFFF)); 721 } 722 723 // Masking sign bits off of a Byte? Do an unsigned byte load plus 724 // an and. 725 if (lop == Op_LoadB && (mask & 0xFFFFFF00) == 0) { 726 Node* ldub = load->as_Load()->convert_to_unsigned_load(*phase); 727 ldub = phase->transform(ldub); 728 return new AndINode(ldub, phase->intcon(mask)); 729 } 730 } 731 732 // Masking off sign bits? Dont make them! 733 if( lop == Op_RShiftI ) { 734 const TypeInt *t12 = phase->type(load->in(2))->isa_int(); 735 if( t12 && t12->is_con() ) { // Shift is by a constant 736 int shift = t12->get_con(); 737 shift &= BitsPerJavaInteger-1; // semantics of Java shifts 738 const int sign_bits_mask = ~right_n_bits(BitsPerJavaInteger - shift); 739 // If the AND'ing of the 2 masks has no bits, then only original shifted 740 // bits survive. NO sign-extension bits survive the maskings. 741 if( (sign_bits_mask & mask) == 0 ) { 742 // Use zero-fill shift instead 743 Node *zshift = phase->transform(new URShiftINode(load->in(1),load->in(2))); 744 return new AndINode( zshift, in(2) ); 745 } 746 } 747 } 748 749 // Check for 'negate/and-1', a pattern emitted when someone asks for 750 // 'mod 2'. Negate leaves the low order bit unchanged (think: complement 751 // plus 1) and the mask is of the low order bit. Skip the negate. 752 if( lop == Op_SubI && mask == 1 && load->in(1) && 753 phase->type(load->in(1)) == TypeInt::ZERO ) 754 return new AndINode( load->in(2), in(2) ); 755 756 return MulNode::Ideal(phase, can_reshape); 757 } 758 759 //============================================================================= 760 //------------------------------mul_ring--------------------------------------- 761 // Supplied function returns the product of the inputs IN THE CURRENT RING. 762 // For the logical operations the ring's MUL is really a logical AND function. 763 // This also type-checks the inputs for sanity. Guaranteed never to 764 // be passed a TOP or BOTTOM type, these are filtered out by pre-check. 765 const Type *AndLNode::mul_ring( const Type *t0, const Type *t1 ) const { 766 const TypeLong *r0 = t0->is_long(); // Handy access 767 const TypeLong *r1 = t1->is_long(); 768 int widen = MAX2(r0->_widen,r1->_widen); 769 770 // If either input is a constant, might be able to trim cases 771 if( !r0->is_con() && !r1->is_con() ) 772 return TypeLong::LONG; // No constants to be had 773 774 // Both constants? Return bits 775 if( r0->is_con() && r1->is_con() ) 776 return TypeLong::make( r0->get_con() & r1->get_con() ); 777 778 if( r0->is_con() && r0->get_con() > 0 ) 779 return TypeLong::make(CONST64(0), r0->get_con(), widen); 780 781 if( r1->is_con() && r1->get_con() > 0 ) 782 return TypeLong::make(CONST64(0), r1->get_con(), widen); 783 784 return TypeLong::LONG; // No constants to be had 785 } 786 787 const Type* AndLNode::Value(PhaseGVN* phase) const { 788 // patterns similar to (v << 2) & 3 789 if (AndIL_shift_and_mask_is_always_zero(phase, in(1), in(2), T_LONG, true)) { 790 return TypeLong::ZERO; 791 } 792 793 return MulNode::Value(phase); 794 } 795 796 //------------------------------Identity--------------------------------------- 797 // Masking off the high bits of an unsigned load is not required 798 Node* AndLNode::Identity(PhaseGVN* phase) { 799 800 // x & x => x 801 if (in(1) == in(2)) { 802 return in(1); 803 } 804 805 Node *usr = in(1); 806 const TypeLong *t2 = phase->type( in(2) )->isa_long(); 807 if( t2 && t2->is_con() ) { 808 jlong con = t2->get_con(); 809 // Masking off high bits which are always zero is useless. 810 const TypeLong* t1 = phase->type( in(1) )->isa_long(); 811 if (t1 != nullptr && t1->_lo >= 0) { 812 int bit_count = log2i_graceful(t1->_hi) + 1; 813 jlong t1_support = jlong(max_julong >> (BitsPerJavaLong - bit_count)); 814 if ((t1_support & con) == t1_support) 815 return usr; 816 } 817 uint lop = usr->Opcode(); 818 // Masking off the high bits of a unsigned-shift-right is not 819 // needed either. 820 if( lop == Op_URShiftL ) { 821 const TypeInt *t12 = phase->type( usr->in(2) )->isa_int(); 822 if( t12 && t12->is_con() ) { // Shift is by a constant 823 int shift = t12->get_con(); 824 shift &= BitsPerJavaLong - 1; // semantics of Java shifts 825 jlong mask = max_julong >> shift; 826 if( (mask&con) == mask ) // If AND is useless, skip it 827 return usr; 828 } 829 } 830 } 831 return MulNode::Identity(phase); 832 } 833 834 //------------------------------Ideal------------------------------------------ 835 Node *AndLNode::Ideal(PhaseGVN *phase, bool can_reshape) { 836 // pattern similar to (v1 + (v2 << 2)) & 3 transformed to v1 & 3 837 Node* progress = AndIL_add_shift_and_mask(phase, T_LONG); 838 if (progress != nullptr) { 839 return progress; 840 } 841 842 // Convert "(~a) & (~b)" into "~(a | b)" 843 if (AddNode::is_not(phase, in(1), T_LONG) && AddNode::is_not(phase, in(2), T_LONG)) { 844 Node* or_a_b = new OrLNode(in(1)->in(1), in(2)->in(1)); 845 Node* tn = phase->transform(or_a_b); 846 return AddNode::make_not(phase, tn, T_LONG); 847 } 848 849 // Special case constant AND mask 850 const TypeLong *t2 = phase->type( in(2) )->isa_long(); 851 if( !t2 || !t2->is_con() ) return MulNode::Ideal(phase, can_reshape); 852 const jlong mask = t2->get_con(); 853 854 Node* in1 = in(1); 855 int op = in1->Opcode(); 856 857 // Are we masking a long that was converted from an int with a mask 858 // that fits in 32-bits? Commute them and use an AndINode. Don't 859 // convert masks which would cause a sign extension of the integer 860 // value. This check includes UI2L masks (0x00000000FFFFFFFF) which 861 // would be optimized away later in Identity. 862 if (op == Op_ConvI2L && (mask & UCONST64(0xFFFFFFFF80000000)) == 0) { 863 Node* andi = new AndINode(in1->in(1), phase->intcon(mask)); 864 andi = phase->transform(andi); 865 return new ConvI2LNode(andi); 866 } 867 868 // Masking off sign bits? Dont make them! 869 if (op == Op_RShiftL) { 870 const TypeInt* t12 = phase->type(in1->in(2))->isa_int(); 871 if( t12 && t12->is_con() ) { // Shift is by a constant 872 int shift = t12->get_con(); 873 shift &= BitsPerJavaLong - 1; // semantics of Java shifts 874 const julong sign_bits_mask = ~(((julong)CONST64(1) << (julong)(BitsPerJavaLong - shift)) -1); 875 // If the AND'ing of the 2 masks has no bits, then only original shifted 876 // bits survive. NO sign-extension bits survive the maskings. 877 if( (sign_bits_mask & mask) == 0 ) { 878 // Use zero-fill shift instead 879 Node *zshift = phase->transform(new URShiftLNode(in1->in(1), in1->in(2))); 880 return new AndLNode(zshift, in(2)); 881 } 882 } 883 } 884 885 // Search for GraphKit::mark_word_test patterns and fold the test if the result is statically known 886 Node* load1 = in(1); 887 Node* load2 = nullptr; 888 if (load1->is_Phi() && phase->type(load1)->isa_long()) { 889 load1 = in(1)->in(1); 890 load2 = in(1)->in(2); 891 } 892 if (load1 != nullptr && load1->is_Load() && phase->type(load1)->isa_long() && 893 (load2 == nullptr || (load2->is_Load() && phase->type(load2)->isa_long()))) { 894 const TypePtr* adr_t1 = phase->type(load1->in(MemNode::Address))->isa_ptr(); 895 const TypePtr* adr_t2 = (load2 != nullptr) ? phase->type(load2->in(MemNode::Address))->isa_ptr() : nullptr; 896 if (adr_t1 != nullptr && adr_t1->offset() == oopDesc::mark_offset_in_bytes() && 897 (load2 == nullptr || (adr_t2 != nullptr && adr_t2->offset() == in_bytes(Klass::prototype_header_offset())))) { 898 if (mask == markWord::inline_type_pattern) { 899 if (adr_t1->is_inlinetypeptr()) { 900 set_req_X(1, in(2), phase); 901 return this; 902 } else if (!adr_t1->can_be_inline_type()) { 903 set_req_X(1, phase->longcon(0), phase); 904 return this; 905 } 906 } else if (mask == markWord::null_free_array_bit_in_place) { 907 if (adr_t1->is_null_free()) { 908 set_req_X(1, in(2), phase); 909 return this; 910 } else if (adr_t1->is_not_null_free()) { 911 set_req_X(1, phase->longcon(0), phase); 912 return this; 913 } 914 } else if (mask == markWord::flat_array_bit_in_place) { 915 if (adr_t1->is_flat()) { 916 set_req_X(1, in(2), phase); 917 return this; 918 } else if (adr_t1->is_not_flat()) { 919 set_req_X(1, phase->longcon(0), phase); 920 return this; 921 } 922 } 923 } 924 } 925 926 return MulNode::Ideal(phase, can_reshape); 927 } 928 929 LShiftNode* LShiftNode::make(Node* in1, Node* in2, BasicType bt) { 930 switch (bt) { 931 case T_INT: 932 return new LShiftINode(in1, in2); 933 case T_LONG: 934 return new LShiftLNode(in1, in2); 935 default: 936 fatal("Not implemented for %s", type2name(bt)); 937 } 938 return nullptr; 939 } 940 941 //============================================================================= 942 943 static bool const_shift_count(PhaseGVN* phase, Node* shiftNode, int* count) { 944 const TypeInt* tcount = phase->type(shiftNode->in(2))->isa_int(); 945 if (tcount != nullptr && tcount->is_con()) { 946 *count = tcount->get_con(); 947 return true; 948 } 949 return false; 950 } 951 952 static int maskShiftAmount(PhaseGVN* phase, Node* shiftNode, int nBits) { 953 int count = 0; 954 if (const_shift_count(phase, shiftNode, &count)) { 955 int maskedShift = count & (nBits - 1); 956 if (maskedShift == 0) { 957 // Let Identity() handle 0 shift count. 958 return 0; 959 } 960 961 if (count != maskedShift) { 962 shiftNode->set_req(2, phase->intcon(maskedShift)); // Replace shift count with masked value. 963 PhaseIterGVN* igvn = phase->is_IterGVN(); 964 if (igvn) { 965 igvn->rehash_node_delayed(shiftNode); 966 } 967 } 968 return maskedShift; 969 } 970 return 0; 971 } 972 973 //------------------------------Identity--------------------------------------- 974 Node* LShiftINode::Identity(PhaseGVN* phase) { 975 int count = 0; 976 if (const_shift_count(phase, this, &count) && (count & (BitsPerJavaInteger - 1)) == 0) { 977 // Shift by a multiple of 32 does nothing 978 return in(1); 979 } 980 return this; 981 } 982 983 //------------------------------Ideal------------------------------------------ 984 // If the right input is a constant, and the left input is an add of a 985 // constant, flatten the tree: (X+con1)<<con0 ==> X<<con0 + con1<<con0 986 Node *LShiftINode::Ideal(PhaseGVN *phase, bool can_reshape) { 987 int con = maskShiftAmount(phase, this, BitsPerJavaInteger); 988 if (con == 0) { 989 return nullptr; 990 } 991 992 // Left input is an add? 993 Node *add1 = in(1); 994 int add1_op = add1->Opcode(); 995 if( add1_op == Op_AddI ) { // Left input is an add? 996 assert( add1 != add1->in(1), "dead loop in LShiftINode::Ideal" ); 997 998 // Transform is legal, but check for profit. Avoid breaking 'i2s' 999 // and 'i2b' patterns which typically fold into 'StoreC/StoreB'. 1000 if( con < 16 ) { 1001 // Left input is an add of the same number? 1002 if (add1->in(1) == add1->in(2)) { 1003 // Convert "(x + x) << c0" into "x << (c0 + 1)" 1004 // In general, this optimization cannot be applied for c0 == 31 since 1005 // 2x << 31 != x << 32 = x << 0 = x (e.g. x = 1: 2 << 31 = 0 != 1) 1006 return new LShiftINode(add1->in(1), phase->intcon(con + 1)); 1007 } 1008 1009 // Left input is an add of a constant? 1010 const TypeInt *t12 = phase->type(add1->in(2))->isa_int(); 1011 if( t12 && t12->is_con() ){ // Left input is an add of a con? 1012 // Compute X << con0 1013 Node *lsh = phase->transform( new LShiftINode( add1->in(1), in(2) ) ); 1014 // Compute X<<con0 + (con1<<con0) 1015 return new AddINode( lsh, phase->intcon(t12->get_con() << con)); 1016 } 1017 } 1018 } 1019 1020 // Check for "(x >> C1) << C2" 1021 if (add1_op == Op_RShiftI || add1_op == Op_URShiftI) { 1022 int add1Con = 0; 1023 const_shift_count(phase, add1, &add1Con); 1024 1025 // Special case C1 == C2, which just masks off low bits 1026 if (add1Con > 0 && con == add1Con) { 1027 // Convert to "(x & -(1 << C2))" 1028 return new AndINode(add1->in(1), phase->intcon(java_negate(jint(1 << con)))); 1029 } else { 1030 // Wait until the right shift has been sharpened to the correct count 1031 if (add1Con > 0 && add1Con < BitsPerJavaInteger) { 1032 // As loop parsing can produce LShiftI nodes, we should wait until the graph is fully formed 1033 // to apply optimizations, otherwise we can inadvertently stop vectorization opportunities. 1034 if (phase->is_IterGVN()) { 1035 if (con > add1Con) { 1036 // Creates "(x << (C2 - C1)) & -(1 << C2)" 1037 Node* lshift = phase->transform(new LShiftINode(add1->in(1), phase->intcon(con - add1Con))); 1038 return new AndINode(lshift, phase->intcon(java_negate(jint(1 << con)))); 1039 } else { 1040 assert(con < add1Con, "must be (%d < %d)", con, add1Con); 1041 // Creates "(x >> (C1 - C2)) & -(1 << C2)" 1042 1043 // Handle logical and arithmetic shifts 1044 Node* rshift; 1045 if (add1_op == Op_RShiftI) { 1046 rshift = phase->transform(new RShiftINode(add1->in(1), phase->intcon(add1Con - con))); 1047 } else { 1048 rshift = phase->transform(new URShiftINode(add1->in(1), phase->intcon(add1Con - con))); 1049 } 1050 1051 return new AndINode(rshift, phase->intcon(java_negate(jint(1 << con)))); 1052 } 1053 } else { 1054 phase->record_for_igvn(this); 1055 } 1056 } 1057 } 1058 } 1059 1060 // Check for "((x >> C1) & Y) << C2" 1061 if (add1_op == Op_AndI) { 1062 Node *add2 = add1->in(1); 1063 int add2_op = add2->Opcode(); 1064 if (add2_op == Op_RShiftI || add2_op == Op_URShiftI) { 1065 // Special case C1 == C2, which just masks off low bits 1066 if (add2->in(2) == in(2)) { 1067 // Convert to "(x & (Y << C2))" 1068 Node* y_sh = phase->transform(new LShiftINode(add1->in(2), phase->intcon(con))); 1069 return new AndINode(add2->in(1), y_sh); 1070 } 1071 1072 int add2Con = 0; 1073 const_shift_count(phase, add2, &add2Con); 1074 if (add2Con > 0 && add2Con < BitsPerJavaInteger) { 1075 if (phase->is_IterGVN()) { 1076 // Convert to "((x >> C1) << C2) & (Y << C2)" 1077 1078 // Make "(x >> C1) << C2", which will get folded away by the rule above 1079 Node* x_sh = phase->transform(new LShiftINode(add2, phase->intcon(con))); 1080 // Make "Y << C2", which will simplify when Y is a constant 1081 Node* y_sh = phase->transform(new LShiftINode(add1->in(2), phase->intcon(con))); 1082 1083 return new AndINode(x_sh, y_sh); 1084 } else { 1085 phase->record_for_igvn(this); 1086 } 1087 } 1088 } 1089 } 1090 1091 // Check for ((x & ((1<<(32-c0))-1)) << c0) which ANDs off high bits 1092 // before shifting them away. 1093 const jint bits_mask = right_n_bits(BitsPerJavaInteger-con); 1094 if( add1_op == Op_AndI && 1095 phase->type(add1->in(2)) == TypeInt::make( bits_mask ) ) 1096 return new LShiftINode( add1->in(1), in(2) ); 1097 1098 return nullptr; 1099 } 1100 1101 //------------------------------Value------------------------------------------ 1102 // A LShiftINode shifts its input2 left by input1 amount. 1103 const Type* LShiftINode::Value(PhaseGVN* phase) const { 1104 const Type *t1 = phase->type( in(1) ); 1105 const Type *t2 = phase->type( in(2) ); 1106 // Either input is TOP ==> the result is TOP 1107 if( t1 == Type::TOP ) return Type::TOP; 1108 if( t2 == Type::TOP ) return Type::TOP; 1109 1110 // Left input is ZERO ==> the result is ZERO. 1111 if( t1 == TypeInt::ZERO ) return TypeInt::ZERO; 1112 // Shift by zero does nothing 1113 if( t2 == TypeInt::ZERO ) return t1; 1114 1115 // Either input is BOTTOM ==> the result is BOTTOM 1116 if( (t1 == TypeInt::INT) || (t2 == TypeInt::INT) || 1117 (t1 == Type::BOTTOM) || (t2 == Type::BOTTOM) ) 1118 return TypeInt::INT; 1119 1120 const TypeInt *r1 = t1->is_int(); // Handy access 1121 const TypeInt *r2 = t2->is_int(); // Handy access 1122 1123 if (!r2->is_con()) 1124 return TypeInt::INT; 1125 1126 uint shift = r2->get_con(); 1127 shift &= BitsPerJavaInteger-1; // semantics of Java shifts 1128 // Shift by a multiple of 32 does nothing: 1129 if (shift == 0) return t1; 1130 1131 // If the shift is a constant, shift the bounds of the type, 1132 // unless this could lead to an overflow. 1133 if (!r1->is_con()) { 1134 jint lo = r1->_lo, hi = r1->_hi; 1135 if (((lo << shift) >> shift) == lo && 1136 ((hi << shift) >> shift) == hi) { 1137 // No overflow. The range shifts up cleanly. 1138 return TypeInt::make((jint)lo << (jint)shift, 1139 (jint)hi << (jint)shift, 1140 MAX2(r1->_widen,r2->_widen)); 1141 } 1142 return TypeInt::INT; 1143 } 1144 1145 return TypeInt::make( (jint)r1->get_con() << (jint)shift ); 1146 } 1147 1148 //============================================================================= 1149 //------------------------------Identity--------------------------------------- 1150 Node* LShiftLNode::Identity(PhaseGVN* phase) { 1151 int count = 0; 1152 if (const_shift_count(phase, this, &count) && (count & (BitsPerJavaLong - 1)) == 0) { 1153 // Shift by a multiple of 64 does nothing 1154 return in(1); 1155 } 1156 return this; 1157 } 1158 1159 //------------------------------Ideal------------------------------------------ 1160 // If the right input is a constant, and the left input is an add of a 1161 // constant, flatten the tree: (X+con1)<<con0 ==> X<<con0 + con1<<con0 1162 Node *LShiftLNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1163 int con = maskShiftAmount(phase, this, BitsPerJavaLong); 1164 if (con == 0) { 1165 return nullptr; 1166 } 1167 1168 // Left input is an add? 1169 Node *add1 = in(1); 1170 int add1_op = add1->Opcode(); 1171 if( add1_op == Op_AddL ) { // Left input is an add? 1172 // Avoid dead data cycles from dead loops 1173 assert( add1 != add1->in(1), "dead loop in LShiftLNode::Ideal" ); 1174 1175 // Left input is an add of the same number? 1176 if (con != (BitsPerJavaLong - 1) && add1->in(1) == add1->in(2)) { 1177 // Convert "(x + x) << c0" into "x << (c0 + 1)" 1178 // Can only be applied if c0 != 63 because: 1179 // (x + x) << 63 = 2x << 63, while 1180 // (x + x) << 63 --transform--> x << 64 = x << 0 = x (!= 2x << 63, for example for x = 1) 1181 // According to the Java spec, chapter 15.19, we only consider the six lowest-order bits of the right-hand operand 1182 // (i.e. "right-hand operand" & 0b111111). Therefore, x << 64 is the same as x << 0 (64 = 0b10000000 & 0b0111111 = 0). 1183 return new LShiftLNode(add1->in(1), phase->intcon(con + 1)); 1184 } 1185 1186 // Left input is an add of a constant? 1187 const TypeLong *t12 = phase->type(add1->in(2))->isa_long(); 1188 if( t12 && t12->is_con() ){ // Left input is an add of a con? 1189 // Compute X << con0 1190 Node *lsh = phase->transform( new LShiftLNode( add1->in(1), in(2) ) ); 1191 // Compute X<<con0 + (con1<<con0) 1192 return new AddLNode( lsh, phase->longcon(t12->get_con() << con)); 1193 } 1194 } 1195 1196 // Check for "(x >> C1) << C2" 1197 if (add1_op == Op_RShiftL || add1_op == Op_URShiftL) { 1198 int add1Con = 0; 1199 const_shift_count(phase, add1, &add1Con); 1200 1201 // Special case C1 == C2, which just masks off low bits 1202 if (add1Con > 0 && con == add1Con) { 1203 // Convert to "(x & -(1 << C2))" 1204 return new AndLNode(add1->in(1), phase->longcon(java_negate(jlong(CONST64(1) << con)))); 1205 } else { 1206 // Wait until the right shift has been sharpened to the correct count 1207 if (add1Con > 0 && add1Con < BitsPerJavaLong) { 1208 // As loop parsing can produce LShiftI nodes, we should wait until the graph is fully formed 1209 // to apply optimizations, otherwise we can inadvertently stop vectorization opportunities. 1210 if (phase->is_IterGVN()) { 1211 if (con > add1Con) { 1212 // Creates "(x << (C2 - C1)) & -(1 << C2)" 1213 Node* lshift = phase->transform(new LShiftLNode(add1->in(1), phase->intcon(con - add1Con))); 1214 return new AndLNode(lshift, phase->longcon(java_negate(jlong(CONST64(1) << con)))); 1215 } else { 1216 assert(con < add1Con, "must be (%d < %d)", con, add1Con); 1217 // Creates "(x >> (C1 - C2)) & -(1 << C2)" 1218 1219 // Handle logical and arithmetic shifts 1220 Node* rshift; 1221 if (add1_op == Op_RShiftL) { 1222 rshift = phase->transform(new RShiftLNode(add1->in(1), phase->intcon(add1Con - con))); 1223 } else { 1224 rshift = phase->transform(new URShiftLNode(add1->in(1), phase->intcon(add1Con - con))); 1225 } 1226 1227 return new AndLNode(rshift, phase->longcon(java_negate(jlong(CONST64(1) << con)))); 1228 } 1229 } else { 1230 phase->record_for_igvn(this); 1231 } 1232 } 1233 } 1234 } 1235 1236 // Check for "((x >> C1) & Y) << C2" 1237 if (add1_op == Op_AndL) { 1238 Node* add2 = add1->in(1); 1239 int add2_op = add2->Opcode(); 1240 if (add2_op == Op_RShiftL || add2_op == Op_URShiftL) { 1241 // Special case C1 == C2, which just masks off low bits 1242 if (add2->in(2) == in(2)) { 1243 // Convert to "(x & (Y << C2))" 1244 Node* y_sh = phase->transform(new LShiftLNode(add1->in(2), phase->intcon(con))); 1245 return new AndLNode(add2->in(1), y_sh); 1246 } 1247 1248 int add2Con = 0; 1249 const_shift_count(phase, add2, &add2Con); 1250 if (add2Con > 0 && add2Con < BitsPerJavaLong) { 1251 if (phase->is_IterGVN()) { 1252 // Convert to "((x >> C1) << C2) & (Y << C2)" 1253 1254 // Make "(x >> C1) << C2", which will get folded away by the rule above 1255 Node* x_sh = phase->transform(new LShiftLNode(add2, phase->intcon(con))); 1256 // Make "Y << C2", which will simplify when Y is a constant 1257 Node* y_sh = phase->transform(new LShiftLNode(add1->in(2), phase->intcon(con))); 1258 1259 return new AndLNode(x_sh, y_sh); 1260 } else { 1261 phase->record_for_igvn(this); 1262 } 1263 } 1264 } 1265 } 1266 1267 // Check for ((x & ((CONST64(1)<<(64-c0))-1)) << c0) which ANDs off high bits 1268 // before shifting them away. 1269 const jlong bits_mask = jlong(max_julong >> con); 1270 if( add1_op == Op_AndL && 1271 phase->type(add1->in(2)) == TypeLong::make( bits_mask ) ) 1272 return new LShiftLNode( add1->in(1), in(2) ); 1273 1274 return nullptr; 1275 } 1276 1277 //------------------------------Value------------------------------------------ 1278 // A LShiftLNode shifts its input2 left by input1 amount. 1279 const Type* LShiftLNode::Value(PhaseGVN* phase) const { 1280 const Type *t1 = phase->type( in(1) ); 1281 const Type *t2 = phase->type( in(2) ); 1282 // Either input is TOP ==> the result is TOP 1283 if( t1 == Type::TOP ) return Type::TOP; 1284 if( t2 == Type::TOP ) return Type::TOP; 1285 1286 // Left input is ZERO ==> the result is ZERO. 1287 if( t1 == TypeLong::ZERO ) return TypeLong::ZERO; 1288 // Shift by zero does nothing 1289 if( t2 == TypeInt::ZERO ) return t1; 1290 1291 // Either input is BOTTOM ==> the result is BOTTOM 1292 if( (t1 == TypeLong::LONG) || (t2 == TypeInt::INT) || 1293 (t1 == Type::BOTTOM) || (t2 == Type::BOTTOM) ) 1294 return TypeLong::LONG; 1295 1296 const TypeLong *r1 = t1->is_long(); // Handy access 1297 const TypeInt *r2 = t2->is_int(); // Handy access 1298 1299 if (!r2->is_con()) 1300 return TypeLong::LONG; 1301 1302 uint shift = r2->get_con(); 1303 shift &= BitsPerJavaLong - 1; // semantics of Java shifts 1304 // Shift by a multiple of 64 does nothing: 1305 if (shift == 0) return t1; 1306 1307 // If the shift is a constant, shift the bounds of the type, 1308 // unless this could lead to an overflow. 1309 if (!r1->is_con()) { 1310 jlong lo = r1->_lo, hi = r1->_hi; 1311 if (((lo << shift) >> shift) == lo && 1312 ((hi << shift) >> shift) == hi) { 1313 // No overflow. The range shifts up cleanly. 1314 return TypeLong::make((jlong)lo << (jint)shift, 1315 (jlong)hi << (jint)shift, 1316 MAX2(r1->_widen,r2->_widen)); 1317 } 1318 return TypeLong::LONG; 1319 } 1320 1321 return TypeLong::make( (jlong)r1->get_con() << (jint)shift ); 1322 } 1323 1324 //============================================================================= 1325 //------------------------------Identity--------------------------------------- 1326 Node* RShiftINode::Identity(PhaseGVN* phase) { 1327 int count = 0; 1328 if (const_shift_count(phase, this, &count)) { 1329 if ((count & (BitsPerJavaInteger - 1)) == 0) { 1330 // Shift by a multiple of 32 does nothing 1331 return in(1); 1332 } 1333 // Check for useless sign-masking 1334 if (in(1)->Opcode() == Op_LShiftI && 1335 in(1)->req() == 3 && 1336 in(1)->in(2) == in(2)) { 1337 count &= BitsPerJavaInteger-1; // semantics of Java shifts 1338 // Compute masks for which this shifting doesn't change 1339 int lo = (-1 << (BitsPerJavaInteger - ((uint)count)-1)); // FFFF8000 1340 int hi = ~lo; // 00007FFF 1341 const TypeInt* t11 = phase->type(in(1)->in(1))->isa_int(); 1342 if (t11 == nullptr) { 1343 return this; 1344 } 1345 // Does actual value fit inside of mask? 1346 if (lo <= t11->_lo && t11->_hi <= hi) { 1347 return in(1)->in(1); // Then shifting is a nop 1348 } 1349 } 1350 } 1351 return this; 1352 } 1353 1354 //------------------------------Ideal------------------------------------------ 1355 Node *RShiftINode::Ideal(PhaseGVN *phase, bool can_reshape) { 1356 // Inputs may be TOP if they are dead. 1357 const TypeInt *t1 = phase->type(in(1))->isa_int(); 1358 if (!t1) return nullptr; // Left input is an integer 1359 const TypeInt *t3; // type of in(1).in(2) 1360 int shift = maskShiftAmount(phase, this, BitsPerJavaInteger); 1361 if (shift == 0) { 1362 return nullptr; 1363 } 1364 1365 // Check for (x & 0xFF000000) >> 24, whose mask can be made smaller. 1366 // Such expressions arise normally from shift chains like (byte)(x >> 24). 1367 const Node *mask = in(1); 1368 if( mask->Opcode() == Op_AndI && 1369 (t3 = phase->type(mask->in(2))->isa_int()) && 1370 t3->is_con() ) { 1371 Node *x = mask->in(1); 1372 jint maskbits = t3->get_con(); 1373 // Convert to "(x >> shift) & (mask >> shift)" 1374 Node *shr_nomask = phase->transform( new RShiftINode(mask->in(1), in(2)) ); 1375 return new AndINode(shr_nomask, phase->intcon( maskbits >> shift)); 1376 } 1377 1378 // Check for "(short[i] <<16)>>16" which simply sign-extends 1379 const Node *shl = in(1); 1380 if( shl->Opcode() != Op_LShiftI ) return nullptr; 1381 1382 if( shift == 16 && 1383 (t3 = phase->type(shl->in(2))->isa_int()) && 1384 t3->is_con(16) ) { 1385 Node *ld = shl->in(1); 1386 if( ld->Opcode() == Op_LoadS ) { 1387 // Sign extension is just useless here. Return a RShiftI of zero instead 1388 // returning 'ld' directly. We cannot return an old Node directly as 1389 // that is the job of 'Identity' calls and Identity calls only work on 1390 // direct inputs ('ld' is an extra Node removed from 'this'). The 1391 // combined optimization requires Identity only return direct inputs. 1392 set_req_X(1, ld, phase); 1393 set_req_X(2, phase->intcon(0), phase); 1394 return this; 1395 } 1396 else if (can_reshape && 1397 ld->Opcode() == Op_LoadUS && 1398 ld->outcnt() == 1 && ld->unique_out() == shl) 1399 // Replace zero-extension-load with sign-extension-load 1400 return ld->as_Load()->convert_to_signed_load(*phase); 1401 } 1402 1403 // Check for "(byte[i] <<24)>>24" which simply sign-extends 1404 if( shift == 24 && 1405 (t3 = phase->type(shl->in(2))->isa_int()) && 1406 t3->is_con(24) ) { 1407 Node *ld = shl->in(1); 1408 if (ld->Opcode() == Op_LoadB) { 1409 // Sign extension is just useless here 1410 set_req_X(1, ld, phase); 1411 set_req_X(2, phase->intcon(0), phase); 1412 return this; 1413 } 1414 } 1415 1416 return nullptr; 1417 } 1418 1419 //------------------------------Value------------------------------------------ 1420 // A RShiftINode shifts its input2 right by input1 amount. 1421 const Type* RShiftINode::Value(PhaseGVN* phase) const { 1422 const Type *t1 = phase->type( in(1) ); 1423 const Type *t2 = phase->type( in(2) ); 1424 // Either input is TOP ==> the result is TOP 1425 if( t1 == Type::TOP ) return Type::TOP; 1426 if( t2 == Type::TOP ) return Type::TOP; 1427 1428 // Left input is ZERO ==> the result is ZERO. 1429 if( t1 == TypeInt::ZERO ) return TypeInt::ZERO; 1430 // Shift by zero does nothing 1431 if( t2 == TypeInt::ZERO ) return t1; 1432 1433 // Either input is BOTTOM ==> the result is BOTTOM 1434 if (t1 == Type::BOTTOM || t2 == Type::BOTTOM) 1435 return TypeInt::INT; 1436 1437 const TypeInt *r1 = t1->is_int(); // Handy access 1438 const TypeInt *r2 = t2->is_int(); // Handy access 1439 1440 // If the shift is a constant, just shift the bounds of the type. 1441 // For example, if the shift is 31, we just propagate sign bits. 1442 if (!r1->is_con() && r2->is_con()) { 1443 uint shift = r2->get_con(); 1444 shift &= BitsPerJavaInteger-1; // semantics of Java shifts 1445 // Shift by a multiple of 32 does nothing: 1446 if (shift == 0) return t1; 1447 // Calculate reasonably aggressive bounds for the result. 1448 // This is necessary if we are to correctly type things 1449 // like (x<<24>>24) == ((byte)x). 1450 jint lo = (jint)r1->_lo >> (jint)shift; 1451 jint hi = (jint)r1->_hi >> (jint)shift; 1452 assert(lo <= hi, "must have valid bounds"); 1453 const TypeInt* ti = TypeInt::make(lo, hi, MAX2(r1->_widen,r2->_widen)); 1454 #ifdef ASSERT 1455 // Make sure we get the sign-capture idiom correct. 1456 if (shift == BitsPerJavaInteger-1) { 1457 if (r1->_lo >= 0) assert(ti == TypeInt::ZERO, ">>31 of + is 0"); 1458 if (r1->_hi < 0) assert(ti == TypeInt::MINUS_1, ">>31 of - is -1"); 1459 } 1460 #endif 1461 return ti; 1462 } 1463 1464 if (!r1->is_con() || !r2->is_con()) { 1465 // If the left input is non-negative the result must also be non-negative, regardless of what the right input is. 1466 if (r1->_lo >= 0) { 1467 return TypeInt::make(0, r1->_hi, MAX2(r1->_widen, r2->_widen)); 1468 } 1469 1470 // Conversely, if the left input is negative then the result must be negative. 1471 if (r1->_hi <= -1) { 1472 return TypeInt::make(r1->_lo, -1, MAX2(r1->_widen, r2->_widen)); 1473 } 1474 1475 return TypeInt::INT; 1476 } 1477 1478 // Signed shift right 1479 return TypeInt::make(r1->get_con() >> (r2->get_con() & 31)); 1480 } 1481 1482 //============================================================================= 1483 //------------------------------Identity--------------------------------------- 1484 Node* RShiftLNode::Identity(PhaseGVN* phase) { 1485 const TypeInt *ti = phase->type(in(2))->isa_int(); // Shift count is an int. 1486 return (ti && ti->is_con() && (ti->get_con() & (BitsPerJavaLong - 1)) == 0) ? in(1) : this; 1487 } 1488 1489 //------------------------------Value------------------------------------------ 1490 // A RShiftLNode shifts its input2 right by input1 amount. 1491 const Type* RShiftLNode::Value(PhaseGVN* phase) const { 1492 const Type *t1 = phase->type( in(1) ); 1493 const Type *t2 = phase->type( in(2) ); 1494 // Either input is TOP ==> the result is TOP 1495 if( t1 == Type::TOP ) return Type::TOP; 1496 if( t2 == Type::TOP ) return Type::TOP; 1497 1498 // Left input is ZERO ==> the result is ZERO. 1499 if( t1 == TypeLong::ZERO ) return TypeLong::ZERO; 1500 // Shift by zero does nothing 1501 if( t2 == TypeInt::ZERO ) return t1; 1502 1503 // Either input is BOTTOM ==> the result is BOTTOM 1504 if (t1 == Type::BOTTOM || t2 == Type::BOTTOM) 1505 return TypeLong::LONG; 1506 1507 const TypeLong *r1 = t1->is_long(); // Handy access 1508 const TypeInt *r2 = t2->is_int (); // Handy access 1509 1510 // If the shift is a constant, just shift the bounds of the type. 1511 // For example, if the shift is 63, we just propagate sign bits. 1512 if (!r1->is_con() && r2->is_con()) { 1513 uint shift = r2->get_con(); 1514 shift &= (2*BitsPerJavaInteger)-1; // semantics of Java shifts 1515 // Shift by a multiple of 64 does nothing: 1516 if (shift == 0) return t1; 1517 // Calculate reasonably aggressive bounds for the result. 1518 // This is necessary if we are to correctly type things 1519 // like (x<<24>>24) == ((byte)x). 1520 jlong lo = (jlong)r1->_lo >> (jlong)shift; 1521 jlong hi = (jlong)r1->_hi >> (jlong)shift; 1522 assert(lo <= hi, "must have valid bounds"); 1523 const TypeLong* tl = TypeLong::make(lo, hi, MAX2(r1->_widen,r2->_widen)); 1524 #ifdef ASSERT 1525 // Make sure we get the sign-capture idiom correct. 1526 if (shift == (2*BitsPerJavaInteger)-1) { 1527 if (r1->_lo >= 0) assert(tl == TypeLong::ZERO, ">>63 of + is 0"); 1528 if (r1->_hi < 0) assert(tl == TypeLong::MINUS_1, ">>63 of - is -1"); 1529 } 1530 #endif 1531 return tl; 1532 } 1533 1534 if (!r1->is_con() || !r2->is_con()) { 1535 // If the left input is non-negative the result must also be non-negative, regardless of what the right input is. 1536 if (r1->_lo >= 0) { 1537 return TypeLong::make(0, r1->_hi, MAX2(r1->_widen, r2->_widen)); 1538 } 1539 1540 // Conversely, if the left input is negative then the result must be negative. 1541 if (r1->_hi <= -1) { 1542 return TypeLong::make(r1->_lo, -1, MAX2(r1->_widen, r2->_widen)); 1543 } 1544 1545 return TypeLong::LONG; 1546 } 1547 1548 return TypeLong::make(r1->get_con() >> (r2->get_con() & 63)); 1549 } 1550 1551 //============================================================================= 1552 //------------------------------Identity--------------------------------------- 1553 Node* URShiftINode::Identity(PhaseGVN* phase) { 1554 int count = 0; 1555 if (const_shift_count(phase, this, &count) && (count & (BitsPerJavaInteger - 1)) == 0) { 1556 // Shift by a multiple of 32 does nothing 1557 return in(1); 1558 } 1559 1560 // Check for "((x << LogBytesPerWord) + (wordSize-1)) >> LogBytesPerWord" which is just "x". 1561 // Happens during new-array length computation. 1562 // Safe if 'x' is in the range [0..(max_int>>LogBytesPerWord)] 1563 Node *add = in(1); 1564 if (add->Opcode() == Op_AddI) { 1565 const TypeInt *t2 = phase->type(add->in(2))->isa_int(); 1566 if (t2 && t2->is_con(wordSize - 1) && 1567 add->in(1)->Opcode() == Op_LShiftI) { 1568 // Check that shift_counts are LogBytesPerWord. 1569 Node *lshift_count = add->in(1)->in(2); 1570 const TypeInt *t_lshift_count = phase->type(lshift_count)->isa_int(); 1571 if (t_lshift_count && t_lshift_count->is_con(LogBytesPerWord) && 1572 t_lshift_count == phase->type(in(2))) { 1573 Node *x = add->in(1)->in(1); 1574 const TypeInt *t_x = phase->type(x)->isa_int(); 1575 if (t_x != nullptr && 0 <= t_x->_lo && t_x->_hi <= (max_jint>>LogBytesPerWord)) { 1576 return x; 1577 } 1578 } 1579 } 1580 } 1581 1582 return (phase->type(in(2))->higher_equal(TypeInt::ZERO)) ? in(1) : this; 1583 } 1584 1585 //------------------------------Ideal------------------------------------------ 1586 Node *URShiftINode::Ideal(PhaseGVN *phase, bool can_reshape) { 1587 int con = maskShiftAmount(phase, this, BitsPerJavaInteger); 1588 if (con == 0) { 1589 return nullptr; 1590 } 1591 1592 // We'll be wanting the right-shift amount as a mask of that many bits 1593 const int mask = right_n_bits(BitsPerJavaInteger - con); 1594 1595 int in1_op = in(1)->Opcode(); 1596 1597 // Check for ((x>>>a)>>>b) and replace with (x>>>(a+b)) when a+b < 32 1598 if( in1_op == Op_URShiftI ) { 1599 const TypeInt *t12 = phase->type( in(1)->in(2) )->isa_int(); 1600 if( t12 && t12->is_con() ) { // Right input is a constant 1601 assert( in(1) != in(1)->in(1), "dead loop in URShiftINode::Ideal" ); 1602 const int con2 = t12->get_con() & 31; // Shift count is always masked 1603 const int con3 = con+con2; 1604 if( con3 < 32 ) // Only merge shifts if total is < 32 1605 return new URShiftINode( in(1)->in(1), phase->intcon(con3) ); 1606 } 1607 } 1608 1609 // Check for ((x << z) + Y) >>> z. Replace with x + con>>>z 1610 // The idiom for rounding to a power of 2 is "(Q+(2^z-1)) >>> z". 1611 // If Q is "X << z" the rounding is useless. Look for patterns like 1612 // ((X<<Z) + Y) >>> Z and replace with (X + Y>>>Z) & Z-mask. 1613 Node *add = in(1); 1614 const TypeInt *t2 = phase->type(in(2))->isa_int(); 1615 if (in1_op == Op_AddI) { 1616 Node *lshl = add->in(1); 1617 if( lshl->Opcode() == Op_LShiftI && 1618 phase->type(lshl->in(2)) == t2 ) { 1619 Node *y_z = phase->transform( new URShiftINode(add->in(2),in(2)) ); 1620 Node *sum = phase->transform( new AddINode( lshl->in(1), y_z ) ); 1621 return new AndINode( sum, phase->intcon(mask) ); 1622 } 1623 } 1624 1625 // Check for (x & mask) >>> z. Replace with (x >>> z) & (mask >>> z) 1626 // This shortens the mask. Also, if we are extracting a high byte and 1627 // storing it to a buffer, the mask will be removed completely. 1628 Node *andi = in(1); 1629 if( in1_op == Op_AndI ) { 1630 const TypeInt *t3 = phase->type( andi->in(2) )->isa_int(); 1631 if( t3 && t3->is_con() ) { // Right input is a constant 1632 jint mask2 = t3->get_con(); 1633 mask2 >>= con; // *signed* shift downward (high-order zeroes do not help) 1634 Node *newshr = phase->transform( new URShiftINode(andi->in(1), in(2)) ); 1635 return new AndINode(newshr, phase->intcon(mask2)); 1636 // The negative values are easier to materialize than positive ones. 1637 // A typical case from address arithmetic is ((x & ~15) >> 4). 1638 // It's better to change that to ((x >> 4) & ~0) versus 1639 // ((x >> 4) & 0x0FFFFFFF). The difference is greatest in LP64. 1640 } 1641 } 1642 1643 // Check for "(X << z ) >>> z" which simply zero-extends 1644 Node *shl = in(1); 1645 if( in1_op == Op_LShiftI && 1646 phase->type(shl->in(2)) == t2 ) 1647 return new AndINode( shl->in(1), phase->intcon(mask) ); 1648 1649 // Check for (x >> n) >>> 31. Replace with (x >>> 31) 1650 Node *shr = in(1); 1651 if ( in1_op == Op_RShiftI ) { 1652 Node *in11 = shr->in(1); 1653 Node *in12 = shr->in(2); 1654 const TypeInt *t11 = phase->type(in11)->isa_int(); 1655 const TypeInt *t12 = phase->type(in12)->isa_int(); 1656 if ( t11 && t2 && t2->is_con(31) && t12 && t12->is_con() ) { 1657 return new URShiftINode(in11, phase->intcon(31)); 1658 } 1659 } 1660 1661 return nullptr; 1662 } 1663 1664 //------------------------------Value------------------------------------------ 1665 // A URShiftINode shifts its input2 right by input1 amount. 1666 const Type* URShiftINode::Value(PhaseGVN* phase) const { 1667 // (This is a near clone of RShiftINode::Value.) 1668 const Type *t1 = phase->type( in(1) ); 1669 const Type *t2 = phase->type( in(2) ); 1670 // Either input is TOP ==> the result is TOP 1671 if( t1 == Type::TOP ) return Type::TOP; 1672 if( t2 == Type::TOP ) return Type::TOP; 1673 1674 // Left input is ZERO ==> the result is ZERO. 1675 if( t1 == TypeInt::ZERO ) return TypeInt::ZERO; 1676 // Shift by zero does nothing 1677 if( t2 == TypeInt::ZERO ) return t1; 1678 1679 // Either input is BOTTOM ==> the result is BOTTOM 1680 if (t1 == Type::BOTTOM || t2 == Type::BOTTOM) 1681 return TypeInt::INT; 1682 1683 if (t2 == TypeInt::INT) 1684 return TypeInt::INT; 1685 1686 const TypeInt *r1 = t1->is_int(); // Handy access 1687 const TypeInt *r2 = t2->is_int(); // Handy access 1688 1689 if (r2->is_con()) { 1690 uint shift = r2->get_con(); 1691 shift &= BitsPerJavaInteger-1; // semantics of Java shifts 1692 // Shift by a multiple of 32 does nothing: 1693 if (shift == 0) return t1; 1694 // Calculate reasonably aggressive bounds for the result. 1695 jint lo = (juint)r1->_lo >> (juint)shift; 1696 jint hi = (juint)r1->_hi >> (juint)shift; 1697 if (r1->_hi >= 0 && r1->_lo < 0) { 1698 // If the type has both negative and positive values, 1699 // there are two separate sub-domains to worry about: 1700 // The positive half and the negative half. 1701 jint neg_lo = lo; 1702 jint neg_hi = (juint)-1 >> (juint)shift; 1703 jint pos_lo = (juint) 0 >> (juint)shift; 1704 jint pos_hi = hi; 1705 lo = MIN2(neg_lo, pos_lo); // == 0 1706 hi = MAX2(neg_hi, pos_hi); // == -1 >>> shift; 1707 } 1708 assert(lo <= hi, "must have valid bounds"); 1709 const TypeInt* ti = TypeInt::make(lo, hi, MAX2(r1->_widen,r2->_widen)); 1710 #ifdef ASSERT 1711 // Make sure we get the sign-capture idiom correct. 1712 if (shift == BitsPerJavaInteger-1) { 1713 if (r1->_lo >= 0) assert(ti == TypeInt::ZERO, ">>>31 of + is 0"); 1714 if (r1->_hi < 0) assert(ti == TypeInt::ONE, ">>>31 of - is +1"); 1715 } 1716 #endif 1717 return ti; 1718 } 1719 1720 // 1721 // Do not support shifted oops in info for GC 1722 // 1723 // else if( t1->base() == Type::InstPtr ) { 1724 // 1725 // const TypeInstPtr *o = t1->is_instptr(); 1726 // if( t1->singleton() ) 1727 // return TypeInt::make( ((uint32_t)o->const_oop() + o->_offset) >> shift ); 1728 // } 1729 // else if( t1->base() == Type::KlassPtr ) { 1730 // const TypeKlassPtr *o = t1->is_klassptr(); 1731 // if( t1->singleton() ) 1732 // return TypeInt::make( ((uint32_t)o->const_oop() + o->_offset) >> shift ); 1733 // } 1734 1735 return TypeInt::INT; 1736 } 1737 1738 //============================================================================= 1739 //------------------------------Identity--------------------------------------- 1740 Node* URShiftLNode::Identity(PhaseGVN* phase) { 1741 int count = 0; 1742 if (const_shift_count(phase, this, &count) && (count & (BitsPerJavaLong - 1)) == 0) { 1743 // Shift by a multiple of 64 does nothing 1744 return in(1); 1745 } 1746 return this; 1747 } 1748 1749 //------------------------------Ideal------------------------------------------ 1750 Node *URShiftLNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1751 int con = maskShiftAmount(phase, this, BitsPerJavaLong); 1752 if (con == 0) { 1753 return nullptr; 1754 } 1755 1756 // We'll be wanting the right-shift amount as a mask of that many bits 1757 const jlong mask = jlong(max_julong >> con); 1758 1759 // Check for ((x << z) + Y) >>> z. Replace with x + con>>>z 1760 // The idiom for rounding to a power of 2 is "(Q+(2^z-1)) >>> z". 1761 // If Q is "X << z" the rounding is useless. Look for patterns like 1762 // ((X<<Z) + Y) >>> Z and replace with (X + Y>>>Z) & Z-mask. 1763 Node *add = in(1); 1764 const TypeInt *t2 = phase->type(in(2))->isa_int(); 1765 if (add->Opcode() == Op_AddL) { 1766 Node *lshl = add->in(1); 1767 if( lshl->Opcode() == Op_LShiftL && 1768 phase->type(lshl->in(2)) == t2 ) { 1769 Node *y_z = phase->transform( new URShiftLNode(add->in(2),in(2)) ); 1770 Node *sum = phase->transform( new AddLNode( lshl->in(1), y_z ) ); 1771 return new AndLNode( sum, phase->longcon(mask) ); 1772 } 1773 } 1774 1775 // Check for (x & mask) >>> z. Replace with (x >>> z) & (mask >>> z) 1776 // This shortens the mask. Also, if we are extracting a high byte and 1777 // storing it to a buffer, the mask will be removed completely. 1778 Node *andi = in(1); 1779 if( andi->Opcode() == Op_AndL ) { 1780 const TypeLong *t3 = phase->type( andi->in(2) )->isa_long(); 1781 if( t3 && t3->is_con() ) { // Right input is a constant 1782 jlong mask2 = t3->get_con(); 1783 mask2 >>= con; // *signed* shift downward (high-order zeroes do not help) 1784 Node *newshr = phase->transform( new URShiftLNode(andi->in(1), in(2)) ); 1785 return new AndLNode(newshr, phase->longcon(mask2)); 1786 } 1787 } 1788 1789 // Check for "(X << z ) >>> z" which simply zero-extends 1790 Node *shl = in(1); 1791 if( shl->Opcode() == Op_LShiftL && 1792 phase->type(shl->in(2)) == t2 ) 1793 return new AndLNode( shl->in(1), phase->longcon(mask) ); 1794 1795 // Check for (x >> n) >>> 63. Replace with (x >>> 63) 1796 Node *shr = in(1); 1797 if ( shr->Opcode() == Op_RShiftL ) { 1798 Node *in11 = shr->in(1); 1799 Node *in12 = shr->in(2); 1800 const TypeLong *t11 = phase->type(in11)->isa_long(); 1801 const TypeInt *t12 = phase->type(in12)->isa_int(); 1802 if ( t11 && t2 && t2->is_con(63) && t12 && t12->is_con() ) { 1803 return new URShiftLNode(in11, phase->intcon(63)); 1804 } 1805 } 1806 return nullptr; 1807 } 1808 1809 //------------------------------Value------------------------------------------ 1810 // A URShiftINode shifts its input2 right by input1 amount. 1811 const Type* URShiftLNode::Value(PhaseGVN* phase) const { 1812 // (This is a near clone of RShiftLNode::Value.) 1813 const Type *t1 = phase->type( in(1) ); 1814 const Type *t2 = phase->type( in(2) ); 1815 // Either input is TOP ==> the result is TOP 1816 if( t1 == Type::TOP ) return Type::TOP; 1817 if( t2 == Type::TOP ) return Type::TOP; 1818 1819 // Left input is ZERO ==> the result is ZERO. 1820 if( t1 == TypeLong::ZERO ) return TypeLong::ZERO; 1821 // Shift by zero does nothing 1822 if( t2 == TypeInt::ZERO ) return t1; 1823 1824 // Either input is BOTTOM ==> the result is BOTTOM 1825 if (t1 == Type::BOTTOM || t2 == Type::BOTTOM) 1826 return TypeLong::LONG; 1827 1828 if (t2 == TypeInt::INT) 1829 return TypeLong::LONG; 1830 1831 const TypeLong *r1 = t1->is_long(); // Handy access 1832 const TypeInt *r2 = t2->is_int (); // Handy access 1833 1834 if (r2->is_con()) { 1835 uint shift = r2->get_con(); 1836 shift &= BitsPerJavaLong - 1; // semantics of Java shifts 1837 // Shift by a multiple of 64 does nothing: 1838 if (shift == 0) return t1; 1839 // Calculate reasonably aggressive bounds for the result. 1840 jlong lo = (julong)r1->_lo >> (juint)shift; 1841 jlong hi = (julong)r1->_hi >> (juint)shift; 1842 if (r1->_hi >= 0 && r1->_lo < 0) { 1843 // If the type has both negative and positive values, 1844 // there are two separate sub-domains to worry about: 1845 // The positive half and the negative half. 1846 jlong neg_lo = lo; 1847 jlong neg_hi = (julong)-1 >> (juint)shift; 1848 jlong pos_lo = (julong) 0 >> (juint)shift; 1849 jlong pos_hi = hi; 1850 //lo = MIN2(neg_lo, pos_lo); // == 0 1851 lo = neg_lo < pos_lo ? neg_lo : pos_lo; 1852 //hi = MAX2(neg_hi, pos_hi); // == -1 >>> shift; 1853 hi = neg_hi > pos_hi ? neg_hi : pos_hi; 1854 } 1855 assert(lo <= hi, "must have valid bounds"); 1856 const TypeLong* tl = TypeLong::make(lo, hi, MAX2(r1->_widen,r2->_widen)); 1857 #ifdef ASSERT 1858 // Make sure we get the sign-capture idiom correct. 1859 if (shift == BitsPerJavaLong - 1) { 1860 if (r1->_lo >= 0) assert(tl == TypeLong::ZERO, ">>>63 of + is 0"); 1861 if (r1->_hi < 0) assert(tl == TypeLong::ONE, ">>>63 of - is +1"); 1862 } 1863 #endif 1864 return tl; 1865 } 1866 1867 return TypeLong::LONG; // Give up 1868 } 1869 1870 //============================================================================= 1871 //------------------------------Ideal------------------------------------------ 1872 Node* FmaNode::Ideal(PhaseGVN* phase, bool can_reshape) { 1873 // We canonicalize the node by converting "(-a)*b+c" into "b*(-a)+c" 1874 // This reduces the number of rules in the matcher, as we only need to check 1875 // for negations on the second argument, and not the symmetric case where 1876 // the first argument is negated. 1877 if (in(1)->is_Neg() && !in(2)->is_Neg()) { 1878 swap_edges(1, 2); 1879 return this; 1880 } 1881 return nullptr; 1882 } 1883 1884 //============================================================================= 1885 //------------------------------Value------------------------------------------ 1886 const Type* FmaDNode::Value(PhaseGVN* phase) const { 1887 const Type *t1 = phase->type(in(1)); 1888 if (t1 == Type::TOP) return Type::TOP; 1889 if (t1->base() != Type::DoubleCon) return Type::DOUBLE; 1890 const Type *t2 = phase->type(in(2)); 1891 if (t2 == Type::TOP) return Type::TOP; 1892 if (t2->base() != Type::DoubleCon) return Type::DOUBLE; 1893 const Type *t3 = phase->type(in(3)); 1894 if (t3 == Type::TOP) return Type::TOP; 1895 if (t3->base() != Type::DoubleCon) return Type::DOUBLE; 1896 #ifndef __STDC_IEC_559__ 1897 return Type::DOUBLE; 1898 #else 1899 double d1 = t1->getd(); 1900 double d2 = t2->getd(); 1901 double d3 = t3->getd(); 1902 return TypeD::make(fma(d1, d2, d3)); 1903 #endif 1904 } 1905 1906 //============================================================================= 1907 //------------------------------Value------------------------------------------ 1908 const Type* FmaFNode::Value(PhaseGVN* phase) const { 1909 const Type *t1 = phase->type(in(1)); 1910 if (t1 == Type::TOP) return Type::TOP; 1911 if (t1->base() != Type::FloatCon) return Type::FLOAT; 1912 const Type *t2 = phase->type(in(2)); 1913 if (t2 == Type::TOP) return Type::TOP; 1914 if (t2->base() != Type::FloatCon) return Type::FLOAT; 1915 const Type *t3 = phase->type(in(3)); 1916 if (t3 == Type::TOP) return Type::TOP; 1917 if (t3->base() != Type::FloatCon) return Type::FLOAT; 1918 #ifndef __STDC_IEC_559__ 1919 return Type::FLOAT; 1920 #else 1921 float f1 = t1->getf(); 1922 float f2 = t2->getf(); 1923 float f3 = t3->getf(); 1924 return TypeF::make(fma(f1, f2, f3)); 1925 #endif 1926 } 1927 1928 //============================================================================= 1929 //------------------------------hash------------------------------------------- 1930 // Hash function for MulAddS2INode. Operation is commutative with commutative pairs. 1931 // The hash function must return the same value when edge swapping is performed. 1932 uint MulAddS2INode::hash() const { 1933 return (uintptr_t)in(1) + (uintptr_t)in(2) + (uintptr_t)in(3) + (uintptr_t)in(4) + Opcode(); 1934 } 1935 1936 //------------------------------Rotate Operations ------------------------------ 1937 1938 Node* RotateLeftNode::Identity(PhaseGVN* phase) { 1939 const Type* t1 = phase->type(in(1)); 1940 if (t1 == Type::TOP) { 1941 return this; 1942 } 1943 int count = 0; 1944 assert(t1->isa_int() || t1->isa_long(), "Unexpected type"); 1945 int mask = (t1->isa_int() ? BitsPerJavaInteger : BitsPerJavaLong) - 1; 1946 if (const_shift_count(phase, this, &count) && (count & mask) == 0) { 1947 // Rotate by a multiple of 32/64 does nothing 1948 return in(1); 1949 } 1950 return this; 1951 } 1952 1953 const Type* RotateLeftNode::Value(PhaseGVN* phase) const { 1954 const Type* t1 = phase->type(in(1)); 1955 const Type* t2 = phase->type(in(2)); 1956 // Either input is TOP ==> the result is TOP 1957 if (t1 == Type::TOP || t2 == Type::TOP) { 1958 return Type::TOP; 1959 } 1960 1961 if (t1->isa_int()) { 1962 const TypeInt* r1 = t1->is_int(); 1963 const TypeInt* r2 = t2->is_int(); 1964 1965 // Left input is ZERO ==> the result is ZERO. 1966 if (r1 == TypeInt::ZERO) { 1967 return TypeInt::ZERO; 1968 } 1969 // Rotate by zero does nothing 1970 if (r2 == TypeInt::ZERO) { 1971 return r1; 1972 } 1973 if (r1->is_con() && r2->is_con()) { 1974 juint r1_con = (juint)r1->get_con(); 1975 juint shift = (juint)(r2->get_con()) & (juint)(BitsPerJavaInteger - 1); // semantics of Java shifts 1976 return TypeInt::make((r1_con << shift) | (r1_con >> (32 - shift))); 1977 } 1978 return TypeInt::INT; 1979 } else { 1980 assert(t1->isa_long(), "Type must be a long"); 1981 const TypeLong* r1 = t1->is_long(); 1982 const TypeInt* r2 = t2->is_int(); 1983 1984 // Left input is ZERO ==> the result is ZERO. 1985 if (r1 == TypeLong::ZERO) { 1986 return TypeLong::ZERO; 1987 } 1988 // Rotate by zero does nothing 1989 if (r2 == TypeInt::ZERO) { 1990 return r1; 1991 } 1992 if (r1->is_con() && r2->is_con()) { 1993 julong r1_con = (julong)r1->get_con(); 1994 julong shift = (julong)(r2->get_con()) & (julong)(BitsPerJavaLong - 1); // semantics of Java shifts 1995 return TypeLong::make((r1_con << shift) | (r1_con >> (64 - shift))); 1996 } 1997 return TypeLong::LONG; 1998 } 1999 } 2000 2001 Node* RotateLeftNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2002 const Type* t1 = phase->type(in(1)); 2003 const Type* t2 = phase->type(in(2)); 2004 if (t2->isa_int() && t2->is_int()->is_con()) { 2005 if (t1->isa_int()) { 2006 int lshift = t2->is_int()->get_con() & 31; 2007 return new RotateRightNode(in(1), phase->intcon(32 - (lshift & 31)), TypeInt::INT); 2008 } else if (t1 != Type::TOP) { 2009 assert(t1->isa_long(), "Type must be a long"); 2010 int lshift = t2->is_int()->get_con() & 63; 2011 return new RotateRightNode(in(1), phase->intcon(64 - (lshift & 63)), TypeLong::LONG); 2012 } 2013 } 2014 return nullptr; 2015 } 2016 2017 Node* RotateRightNode::Identity(PhaseGVN* phase) { 2018 const Type* t1 = phase->type(in(1)); 2019 if (t1 == Type::TOP) { 2020 return this; 2021 } 2022 int count = 0; 2023 assert(t1->isa_int() || t1->isa_long(), "Unexpected type"); 2024 int mask = (t1->isa_int() ? BitsPerJavaInteger : BitsPerJavaLong) - 1; 2025 if (const_shift_count(phase, this, &count) && (count & mask) == 0) { 2026 // Rotate by a multiple of 32/64 does nothing 2027 return in(1); 2028 } 2029 return this; 2030 } 2031 2032 const Type* RotateRightNode::Value(PhaseGVN* phase) const { 2033 const Type* t1 = phase->type(in(1)); 2034 const Type* t2 = phase->type(in(2)); 2035 // Either input is TOP ==> the result is TOP 2036 if (t1 == Type::TOP || t2 == Type::TOP) { 2037 return Type::TOP; 2038 } 2039 2040 if (t1->isa_int()) { 2041 const TypeInt* r1 = t1->is_int(); 2042 const TypeInt* r2 = t2->is_int(); 2043 2044 // Left input is ZERO ==> the result is ZERO. 2045 if (r1 == TypeInt::ZERO) { 2046 return TypeInt::ZERO; 2047 } 2048 // Rotate by zero does nothing 2049 if (r2 == TypeInt::ZERO) { 2050 return r1; 2051 } 2052 if (r1->is_con() && r2->is_con()) { 2053 juint r1_con = (juint)r1->get_con(); 2054 juint shift = (juint)(r2->get_con()) & (juint)(BitsPerJavaInteger - 1); // semantics of Java shifts 2055 return TypeInt::make((r1_con >> shift) | (r1_con << (32 - shift))); 2056 } 2057 return TypeInt::INT; 2058 } else { 2059 assert(t1->isa_long(), "Type must be a long"); 2060 const TypeLong* r1 = t1->is_long(); 2061 const TypeInt* r2 = t2->is_int(); 2062 // Left input is ZERO ==> the result is ZERO. 2063 if (r1 == TypeLong::ZERO) { 2064 return TypeLong::ZERO; 2065 } 2066 // Rotate by zero does nothing 2067 if (r2 == TypeInt::ZERO) { 2068 return r1; 2069 } 2070 if (r1->is_con() && r2->is_con()) { 2071 julong r1_con = (julong)r1->get_con(); 2072 julong shift = (julong)(r2->get_con()) & (julong)(BitsPerJavaLong - 1); // semantics of Java shifts 2073 return TypeLong::make((r1_con >> shift) | (r1_con << (64 - shift))); 2074 } 2075 return TypeLong::LONG; 2076 } 2077 } 2078 2079 // Given an expression (AndX shift mask) or (AndX mask shift), 2080 // determine if the AndX must always produce zero, because the 2081 // the shift (x<<N) is bitwise disjoint from the mask #M. 2082 // The X in AndX must be I or L, depending on bt. 2083 // Specifically, the following cases fold to zero, 2084 // when the shift value N is large enough to zero out 2085 // all the set positions of the and-mask M. 2086 // (AndI (LShiftI _ #N) #M) => #0 2087 // (AndL (LShiftL _ #N) #M) => #0 2088 // (AndL (ConvI2L (LShiftI _ #N)) #M) => #0 2089 // The M and N values must satisfy ((-1 << N) & M) == 0. 2090 // Because the optimization might work for a non-constant 2091 // mask M, we check the AndX for both operand orders. 2092 bool MulNode::AndIL_shift_and_mask_is_always_zero(PhaseGVN* phase, Node* shift, Node* mask, BasicType bt, bool check_reverse) { 2093 if (mask == nullptr || shift == nullptr) { 2094 return false; 2095 } 2096 const TypeInteger* mask_t = phase->type(mask)->isa_integer(bt); 2097 if (mask_t == nullptr || phase->type(shift)->isa_integer(bt) == nullptr) { 2098 return false; 2099 } 2100 shift = shift->uncast(); 2101 if (shift == nullptr) { 2102 return false; 2103 } 2104 if (phase->type(shift)->isa_integer(bt) == nullptr) { 2105 return false; 2106 } 2107 BasicType shift_bt = bt; 2108 if (bt == T_LONG && shift->Opcode() == Op_ConvI2L) { 2109 bt = T_INT; 2110 Node* val = shift->in(1); 2111 if (val == nullptr) { 2112 return false; 2113 } 2114 val = val->uncast(); 2115 if (val == nullptr) { 2116 return false; 2117 } 2118 if (val->Opcode() == Op_LShiftI) { 2119 shift_bt = T_INT; 2120 shift = val; 2121 if (phase->type(shift)->isa_integer(bt) == nullptr) { 2122 return false; 2123 } 2124 } 2125 } 2126 if (shift->Opcode() != Op_LShift(shift_bt)) { 2127 if (check_reverse && 2128 (mask->Opcode() == Op_LShift(bt) || 2129 (bt == T_LONG && mask->Opcode() == Op_ConvI2L))) { 2130 // try it the other way around 2131 return AndIL_shift_and_mask_is_always_zero(phase, mask, shift, bt, false); 2132 } 2133 return false; 2134 } 2135 Node* shift2 = shift->in(2); 2136 if (shift2 == nullptr) { 2137 return false; 2138 } 2139 const Type* shift2_t = phase->type(shift2); 2140 if (!shift2_t->isa_int() || !shift2_t->is_int()->is_con()) { 2141 return false; 2142 } 2143 2144 jint shift_con = shift2_t->is_int()->get_con() & ((shift_bt == T_INT ? BitsPerJavaInteger : BitsPerJavaLong) - 1); 2145 if ((((jlong)1) << shift_con) > mask_t->hi_as_long() && mask_t->lo_as_long() >= 0) { 2146 return true; 2147 } 2148 2149 return false; 2150 } 2151 2152 // Given an expression (AndX (AddX v1 (LShiftX v2 #N)) #M) 2153 // determine if the AndX must always produce (AndX v1 #M), 2154 // because the shift (v2<<N) is bitwise disjoint from the mask #M. 2155 // The X in AndX will be I or L, depending on bt. 2156 // Specifically, the following cases fold, 2157 // when the shift value N is large enough to zero out 2158 // all the set positions of the and-mask M. 2159 // (AndI (AddI v1 (LShiftI _ #N)) #M) => (AndI v1 #M) 2160 // (AndL (AddI v1 (LShiftL _ #N)) #M) => (AndL v1 #M) 2161 // (AndL (AddL v1 (ConvI2L (LShiftI _ #N))) #M) => (AndL v1 #M) 2162 // The M and N values must satisfy ((-1 << N) & M) == 0. 2163 // Because the optimization might work for a non-constant 2164 // mask M, and because the AddX operands can come in either 2165 // order, we check for every operand order. 2166 Node* MulNode::AndIL_add_shift_and_mask(PhaseGVN* phase, BasicType bt) { 2167 Node* add = in(1); 2168 Node* mask = in(2); 2169 if (add == nullptr || mask == nullptr) { 2170 return nullptr; 2171 } 2172 int addidx = 0; 2173 if (add->Opcode() == Op_Add(bt)) { 2174 addidx = 1; 2175 } else if (mask->Opcode() == Op_Add(bt)) { 2176 mask = add; 2177 addidx = 2; 2178 add = in(addidx); 2179 } 2180 if (addidx > 0) { 2181 Node* add1 = add->in(1); 2182 Node* add2 = add->in(2); 2183 if (add1 != nullptr && add2 != nullptr) { 2184 if (AndIL_shift_and_mask_is_always_zero(phase, add1, mask, bt, false)) { 2185 set_req_X(addidx, add2, phase); 2186 return this; 2187 } else if (AndIL_shift_and_mask_is_always_zero(phase, add2, mask, bt, false)) { 2188 set_req_X(addidx, add1, phase); 2189 return this; 2190 } 2191 } 2192 } 2193 return nullptr; 2194 }