1 /*
   2  * Copyright (c) 1997, 2026, 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 "memory/allocation.inline.hpp"
  26 #include "opto/addnode.hpp"
  27 #include "opto/connode.hpp"
  28 #include "opto/convertnode.hpp"
  29 #include "opto/memnode.hpp"
  30 #include "opto/mulnode.hpp"
  31 #include "opto/phaseX.hpp"
  32 #include "opto/rangeinference.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                   (op == Op_MulHF);
  72 
  73   // Convert "(-a)*(-b)" into "a*b".
  74   if (real_mul && in1->is_Sub() && in2->is_Sub()) {
  75     if (phase->type(in1->in(1))->is_zero_type() &&
  76         phase->type(in2->in(1))->is_zero_type()) {
  77       set_req_X(1, in1->in(2), phase);
  78       set_req_X(2, in2->in(2), phase);
  79       in1 = in(1);
  80       in2 = in(2);
  81       progress = this;
  82     }
  83   }
  84 
  85   // convert "max(a,b) * min(a,b)" into "a*b".
  86   if ((in(1)->Opcode() == max_opcode() && in(2)->Opcode() == min_opcode())
  87       || (in(1)->Opcode() == min_opcode() && in(2)->Opcode() == max_opcode())) {
  88     Node *in11 = in(1)->in(1);
  89     Node *in12 = in(1)->in(2);
  90 
  91     Node *in21 = in(2)->in(1);
  92     Node *in22 = in(2)->in(2);
  93 
  94     if ((in11 == in21 && in12 == in22) ||
  95         (in11 == in22 && in12 == in21)) {
  96       set_req_X(1, in11, phase);
  97       set_req_X(2, in12, phase);
  98       in1 = in(1);
  99       in2 = in(2);
 100       progress = this;
 101     }
 102   }
 103 
 104   const Type* t1 = phase->type(in1);
 105   const Type* t2 = phase->type(in2);
 106 
 107   // We are OK if right is a constant, or right is a load and
 108   // left is a non-constant.
 109   if( !(t2->singleton() ||
 110         (in(2)->is_Load() && !(t1->singleton() || in(1)->is_Load())) ) ) {
 111     if( t1->singleton() ||       // Left input is a constant?
 112         // Otherwise, sort inputs (commutativity) to help value numbering.
 113         (in(1)->_idx > in(2)->_idx) ) {
 114       swap_edges(1, 2);
 115       const Type *t = t1;
 116       t1 = t2;
 117       t2 = t;
 118       progress = this;            // Made progress
 119     }
 120   }
 121 
 122   // If the right input is a constant, and the left input is a product of a
 123   // constant, flatten the expression tree.
 124   if( t2->singleton() &&        // Right input is a constant?
 125       op != Op_MulF &&          // Float & double cannot reassociate
 126       op != Op_MulD &&
 127       op != Op_MulHF) {
 128     if( t2 == Type::TOP ) return nullptr;
 129     Node *mul1 = in(1);
 130 #ifdef ASSERT
 131     // Check for dead loop
 132     int op1 = mul1->Opcode();
 133     if ((mul1 == this) || (in(2) == this) ||
 134         ((op1 == mul_opcode() || op1 == add_opcode()) &&
 135          ((mul1->in(1) == this) || (mul1->in(2) == this) ||
 136           (mul1->in(1) == mul1) || (mul1->in(2) == mul1)))) {
 137       assert(false, "dead loop in MulNode::Ideal");
 138     }
 139 #endif
 140 
 141     if( mul1->Opcode() == mul_opcode() ) {  // Left input is a multiply?
 142       // Mul of a constant?
 143       const Type *t12 = phase->type( mul1->in(2) );
 144       if( t12->singleton() && t12 != Type::TOP) { // Left input is an add of a constant?
 145         // Compute new constant; check for overflow
 146         const Type *tcon01 = ((MulNode*)mul1)->mul_ring(t2,t12);
 147         if( tcon01->singleton() ) {
 148           // The Mul of the flattened expression
 149           set_req_X(1, mul1->in(1), phase);
 150           set_req_X(2, phase->makecon(tcon01), phase);
 151           t2 = tcon01;
 152           progress = this;      // Made progress
 153         }
 154       }
 155     }
 156     // If the right input is a constant, and the left input is an add of a
 157     // constant, flatten the tree: (X+con1)*con0 ==> X*con0 + con1*con0
 158     const Node *add1 = in(1);
 159     if( add1->Opcode() == add_opcode() ) {      // Left input is an add?
 160       // Add of a constant?
 161       const Type *t12 = phase->type( add1->in(2) );
 162       if( t12->singleton() && t12 != Type::TOP ) { // Left input is an add of a constant?
 163         assert( add1->in(1) != add1, "dead loop in MulNode::Ideal" );
 164         // Compute new constant; check for overflow
 165         const Type *tcon01 = mul_ring(t2,t12);
 166         if( tcon01->singleton() ) {
 167 
 168         // Convert (X+con1)*con0 into X*con0
 169           Node *mul = clone();    // mul = ()*con0
 170           mul->set_req(1,add1->in(1));  // mul = X*con0
 171           mul = phase->transform(mul);
 172 
 173           Node *add2 = add1->clone();
 174           add2->set_req(1, mul);        // X*con0 + con0*con1
 175           add2->set_req(2, phase->makecon(tcon01) );
 176           progress = add2;
 177         }
 178       }
 179     } // End of is left input an add
 180   } // End of is right input a Mul
 181 
 182   return progress;
 183 }
 184 
 185 //------------------------------Value-----------------------------------------
 186 const Type* MulNode::Value(PhaseGVN* phase) const {
 187   const Type *t1 = phase->type( in(1) );
 188   const Type *t2 = phase->type( in(2) );
 189   // Either input is TOP ==> the result is TOP
 190   if( t1 == Type::TOP ) return Type::TOP;
 191   if( t2 == Type::TOP ) return Type::TOP;
 192 
 193   // Either input is ZERO ==> the result is ZERO.
 194   // Not valid for floats or doubles since +0.0 * -0.0 --> +0.0
 195   int op = Opcode();
 196   if( op == Op_MulI || op == Op_AndI || op == Op_MulL || op == Op_AndL ) {
 197     const Type *zero = add_id();        // The multiplicative zero
 198     if( t1->higher_equal( zero ) ) return zero;
 199     if( t2->higher_equal( zero ) ) return zero;
 200   }
 201 
 202   // Either input is BOTTOM ==> the result is the local BOTTOM
 203   if( t1 == Type::BOTTOM || t2 == Type::BOTTOM )
 204     return bottom_type();
 205 
 206   return mul_ring(t1,t2);            // Local flavor of type multiplication
 207 }
 208 
 209 MulNode* MulNode::make(Node* in1, Node* in2, BasicType bt) {
 210   switch (bt) {
 211     case T_INT:
 212       return new MulINode(in1, in2);
 213     case T_LONG:
 214       return new MulLNode(in1, in2);
 215     default:
 216       fatal("Not implemented for %s", type2name(bt));
 217   }
 218   return nullptr;
 219 }
 220 
 221 MulNode* MulNode::make_and(Node* in1, Node* in2, BasicType bt) {
 222   switch (bt) {
 223     case T_INT:
 224       return new AndINode(in1, in2);
 225     case T_LONG:
 226       return new AndLNode(in1, in2);
 227     default:
 228       fatal("Not implemented for %s", type2name(bt));
 229   }
 230   return nullptr;
 231 }
 232 
 233 
 234 //=============================================================================
 235 //------------------------------Ideal------------------------------------------
 236 // Check for power-of-2 multiply, then try the regular MulNode::Ideal
 237 Node *MulINode::Ideal(PhaseGVN *phase, bool can_reshape) {
 238   const jint con = in(2)->find_int_con(0);
 239   if (con == 0) {
 240     // If in(2) is not a constant, call Ideal() of the parent class to
 241     // try to move constant to the right side.
 242     return MulNode::Ideal(phase, can_reshape);
 243   }
 244 
 245   // Now we have a constant Node on the right and the constant in con.
 246   if (con == 1) {
 247     // By one is handled by Identity call
 248     return nullptr;
 249   }
 250 
 251   // Check for negative constant; if so negate the final result
 252   bool sign_flip = false;
 253 
 254   unsigned int abs_con = g_uabs(con);
 255   if (abs_con != (unsigned int)con) {
 256     sign_flip = true;
 257   }
 258 
 259   // Get low bit; check for being the only bit
 260   Node *res = nullptr;
 261   unsigned int bit1 = submultiple_power_of_2(abs_con);
 262   if (bit1 == abs_con) {           // Found a power of 2?
 263     res = new LShiftINode(in(1), phase->intcon(log2i_exact(bit1)));
 264   } else {
 265     // Check for constant with 2 bits set
 266     unsigned int bit2 = abs_con - bit1;
 267     bit2 = bit2 & (0 - bit2);          // Extract 2nd bit
 268     if (bit2 + bit1 == abs_con) {    // Found all bits in con?
 269       Node *n1 = phase->transform(new LShiftINode(in(1), phase->intcon(log2i_exact(bit1))));
 270       Node *n2 = phase->transform(new LShiftINode(in(1), phase->intcon(log2i_exact(bit2))));
 271       res = new AddINode(n2, n1);
 272     } else if (is_power_of_2(abs_con + 1)) {
 273       // Sleezy: power-of-2 - 1.  Next time be generic.
 274       unsigned int temp = abs_con + 1;
 275       Node *n1 = phase->transform(new LShiftINode(in(1), phase->intcon(log2i_exact(temp))));
 276       res = new SubINode(n1, in(1));
 277     } else {
 278       return MulNode::Ideal(phase, can_reshape);
 279     }
 280   }
 281 
 282   if (sign_flip) {             // Need to negate result?
 283     res = phase->transform(res);// Transform, before making the zero con
 284     res = new SubINode(phase->intcon(0),res);
 285   }
 286 
 287   return res;                   // Return final result
 288 }
 289 
 290 // This template class performs type multiplication for MulI/MulLNode. NativeType is either jint or jlong.
 291 // In this class, the inputs of the MulNodes are named left and right with types [left_lo,left_hi] and [right_lo,right_hi].
 292 //
 293 // In general, the multiplication of two x-bit values could produce a result that consumes up to 2x bits if there is
 294 // enough space to hold them all. We can therefore distinguish the following two cases for the product:
 295 // - no overflow (i.e. product fits into x bits)
 296 // - overflow (i.e. product does not fit into x bits)
 297 //
 298 // When multiplying the two x-bit inputs 'left' and 'right' with their x-bit types [left_lo,left_hi] and [right_lo,right_hi]
 299 // we need to find the minimum and maximum of all possible products to define a new type. To do that, we compute the
 300 // cross product of [left_lo,left_hi] and [right_lo,right_hi] in 2x-bit space where no over- or underflow can happen.
 301 // The cross product consists of the following four multiplications with 2x-bit results:
 302 // (1) left_lo * right_lo
 303 // (2) left_lo * right_hi
 304 // (3) left_hi * right_lo
 305 // (4) left_hi * right_hi
 306 //
 307 // Let's define the following two functions:
 308 // - Lx(i): Returns the lower x bits of the 2x-bit number i.
 309 // - Ux(i): Returns the upper x bits of the 2x-bit number i.
 310 //
 311 // Let's first assume all products are positive where only overflows are possible but no underflows. If there is no
 312 // overflow for a product p, then the upper x bits of the 2x-bit result p are all zero:
 313 //     Ux(p) = 0
 314 //     Lx(p) = p
 315 //
 316 // If none of the multiplications (1)-(4) overflow, we can truncate the upper x bits and use the following result type
 317 // with x bits:
 318 //      [result_lo,result_hi] = [MIN(Lx(1),Lx(2),Lx(3),Lx(4)),MAX(Lx(1),Lx(2),Lx(3),Lx(4))]
 319 //
 320 // If any of these multiplications overflows, we could pessimistically take the bottom type for the x bit result
 321 // (i.e. all values in the x-bit space could be possible):
 322 //      [result_lo,result_hi] = [NativeType_min,NativeType_max]
 323 //
 324 // However, in case of any overflow, we can do better by analyzing the upper x bits of all multiplications (1)-(4) with
 325 // 2x-bit results. The upper x bits tell us something about how many times a multiplication has overflown the lower
 326 // x bits. If the upper x bits of (1)-(4) are all equal, then we know that all of these multiplications overflowed
 327 // the lower x bits the same number of times:
 328 //     Ux((1)) = Ux((2)) = Ux((3)) = Ux((4))
 329 //
 330 // If all upper x bits are equal, we can conclude:
 331 //     Lx(MIN((1),(2),(3),(4))) = MIN(Lx(1),Lx(2),Lx(3),Lx(4)))
 332 //     Lx(MAX((1),(2),(3),(4))) = MAX(Lx(1),Lx(2),Lx(3),Lx(4)))
 333 //
 334 // Therefore, we can use the same precise x-bit result type as for the no-overflow case:
 335 //     [result_lo,result_hi] = [(MIN(Lx(1),Lx(2),Lx(3),Lx(4))),MAX(Lx(1),Lx(2),Lx(3),Lx(4)))]
 336 //
 337 //
 338 // Now let's assume that (1)-(4) are signed multiplications where over- and underflow could occur:
 339 // Negative numbers are all sign extend with ones. Therefore, if a negative product does not underflow, then the
 340 // 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,
 341 // the upper x bits are decremented by the number of times an underflow occurred. The smallest possible negative product
 342 // is NativeType_min*NativeType_max, where the upper x bits are set to NativeType_min / 2 (b11...0). It is therefore
 343 // impossible to underflow the upper x bits. Thus, when having all ones (i.e. minus one) in the upper x bits, we know
 344 // that there is no underflow.
 345 //
 346 // To be able to compare the number of over-/underflows of positive and negative products, respectively, we normalize
 347 // the upper x bits of negative 2x-bit products by adding one. This way a product has no over- or underflow if the
 348 // normalized upper x bits are zero. Now we can use the same improved type as for strictly positive products because we
 349 // can compare the upper x bits in a unified way with N() being the normalization function:
 350 //     N(Ux((1))) = N(Ux((2))) = N(Ux((3)) = N(Ux((4)))
 351 template<typename NativeType>
 352 class IntegerTypeMultiplication {
 353 
 354   NativeType _lo_left;
 355   NativeType _lo_right;
 356   NativeType _hi_left;
 357   NativeType _hi_right;
 358   short _widen_left;
 359   short _widen_right;
 360 
 361   static const Type* overflow_type();
 362   static NativeType multiply_high(NativeType x, NativeType y);
 363   const Type* create_type(NativeType lo, NativeType hi) const;
 364 
 365   static NativeType multiply_high_signed_overflow_value(NativeType x, NativeType y) {
 366     return normalize_overflow_value(x, y, multiply_high(x, y));
 367   }
 368 
 369   bool cross_product_not_same_overflow_value() const {
 370     const NativeType lo_lo_high_product = multiply_high_signed_overflow_value(_lo_left, _lo_right);
 371     const NativeType lo_hi_high_product = multiply_high_signed_overflow_value(_lo_left, _hi_right);
 372     const NativeType hi_lo_high_product = multiply_high_signed_overflow_value(_hi_left, _lo_right);
 373     const NativeType hi_hi_high_product = multiply_high_signed_overflow_value(_hi_left, _hi_right);
 374     return lo_lo_high_product != lo_hi_high_product ||
 375            lo_hi_high_product != hi_lo_high_product ||
 376            hi_lo_high_product != hi_hi_high_product;
 377   }
 378 
 379   bool does_product_overflow(NativeType x, NativeType y) const {
 380     return multiply_high_signed_overflow_value(x, y) != 0;
 381   }
 382 
 383   static NativeType normalize_overflow_value(const NativeType x, const NativeType y, NativeType result) {
 384     return java_multiply(x, y) < 0 ? result + 1 : result;
 385   }
 386 
 387  public:
 388   template<class IntegerType>
 389   IntegerTypeMultiplication(const IntegerType* left, const IntegerType* right)
 390       : _lo_left(left->_lo), _lo_right(right->_lo),
 391         _hi_left(left->_hi), _hi_right(right->_hi),
 392         _widen_left(left->_widen), _widen_right(right->_widen)  {}
 393 
 394   // Compute the product type by multiplying the two input type ranges. We take the minimum and maximum of all possible
 395   // values (requires 4 multiplications of all possible combinations of the two range boundary values). If any of these
 396   // multiplications overflows/underflows, we need to make sure that they all have the same number of overflows/underflows
 397   // If that is not the case, we return the bottom type to cover all values due to the inconsistent overflows/underflows).
 398   const Type* compute() const {
 399     if (cross_product_not_same_overflow_value()) {
 400       return overflow_type();
 401     }
 402 
 403     NativeType lo_lo_product = java_multiply(_lo_left, _lo_right);
 404     NativeType lo_hi_product = java_multiply(_lo_left, _hi_right);
 405     NativeType hi_lo_product = java_multiply(_hi_left, _lo_right);
 406     NativeType hi_hi_product = java_multiply(_hi_left, _hi_right);
 407     const NativeType min = MIN4(lo_lo_product, lo_hi_product, hi_lo_product, hi_hi_product);
 408     const NativeType max = MAX4(lo_lo_product, lo_hi_product, hi_lo_product, hi_hi_product);
 409     return create_type(min, max);
 410   }
 411 
 412   bool does_overflow() const {
 413     return does_product_overflow(_lo_left, _lo_right) ||
 414            does_product_overflow(_lo_left, _hi_right) ||
 415            does_product_overflow(_hi_left, _lo_right) ||
 416            does_product_overflow(_hi_left, _hi_right);
 417   }
 418 };
 419 
 420 template <>
 421 const Type* IntegerTypeMultiplication<jint>::overflow_type() {
 422   return TypeInt::INT;
 423 }
 424 
 425 template <>
 426 jint IntegerTypeMultiplication<jint>::multiply_high(const jint x, const jint y) {
 427   const jlong x_64 = x;
 428   const jlong y_64 = y;
 429   const jlong product = x_64 * y_64;
 430   return (jint)((uint64_t)product >> 32u);
 431 }
 432 
 433 template <>
 434 const Type* IntegerTypeMultiplication<jint>::create_type(jint lo, jint hi) const {
 435   return TypeInt::make(lo, hi, MAX2(_widen_left, _widen_right));
 436 }
 437 
 438 template <>
 439 const Type* IntegerTypeMultiplication<jlong>::overflow_type() {
 440   return TypeLong::LONG;
 441 }
 442 
 443 template <>
 444 jlong IntegerTypeMultiplication<jlong>::multiply_high(const jlong x, const jlong y) {
 445   return multiply_high_signed(x, y);
 446 }
 447 
 448 template <>
 449 const Type* IntegerTypeMultiplication<jlong>::create_type(jlong lo, jlong hi) const {
 450   return TypeLong::make(lo, hi, MAX2(_widen_left, _widen_right));
 451 }
 452 
 453 // Compute the product type of two integer ranges into this node.
 454 const Type* MulINode::mul_ring(const Type* type_left, const Type* type_right) const {
 455   const IntegerTypeMultiplication<jint> integer_multiplication(type_left->is_int(), type_right->is_int());
 456   return integer_multiplication.compute();
 457 }
 458 
 459 bool MulINode::does_overflow(const TypeInt* type_left, const TypeInt* type_right) {
 460   const IntegerTypeMultiplication<jint> integer_multiplication(type_left, type_right);
 461   return integer_multiplication.does_overflow();
 462 }
 463 
 464 // Compute the product type of two long ranges into this node.
 465 const Type* MulLNode::mul_ring(const Type* type_left, const Type* type_right) const {
 466   const IntegerTypeMultiplication<jlong> integer_multiplication(type_left->is_long(), type_right->is_long());
 467   return integer_multiplication.compute();
 468 }
 469 
 470 //=============================================================================
 471 //------------------------------Ideal------------------------------------------
 472 // Check for power-of-2 multiply, then try the regular MulNode::Ideal
 473 Node *MulLNode::Ideal(PhaseGVN *phase, bool can_reshape) {
 474   const jlong con = in(2)->find_long_con(0);
 475   if (con == 0) {
 476     // If in(2) is not a constant, call Ideal() of the parent class to
 477     // try to move constant to the right side.
 478     return MulNode::Ideal(phase, can_reshape);
 479   }
 480 
 481   // Now we have a constant Node on the right and the constant in con.
 482   if (con == 1) {
 483     // By one is handled by Identity call
 484     return nullptr;
 485   }
 486 
 487   // Check for negative constant; if so negate the final result
 488   bool sign_flip = false;
 489   julong abs_con = g_uabs(con);
 490   if (abs_con != (julong)con) {
 491     sign_flip = true;
 492   }
 493 
 494   // Get low bit; check for being the only bit
 495   Node *res = nullptr;
 496   julong bit1 = submultiple_power_of_2(abs_con);
 497   if (bit1 == abs_con) {           // Found a power of 2?
 498     res = new LShiftLNode(in(1), phase->intcon(log2i_exact(bit1)));
 499   } else {
 500 
 501     // Check for constant with 2 bits set
 502     julong bit2 = abs_con-bit1;
 503     bit2 = bit2 & (0-bit2);          // Extract 2nd bit
 504     if (bit2 + bit1 == abs_con) {    // Found all bits in con?
 505       Node *n1 = phase->transform(new LShiftLNode(in(1), phase->intcon(log2i_exact(bit1))));
 506       Node *n2 = phase->transform(new LShiftLNode(in(1), phase->intcon(log2i_exact(bit2))));
 507       res = new AddLNode(n2, n1);
 508 
 509     } else if (is_power_of_2(abs_con+1)) {
 510       // Sleezy: power-of-2 -1.  Next time be generic.
 511       julong temp = abs_con + 1;
 512       Node *n1 = phase->transform( new LShiftLNode(in(1), phase->intcon(log2i_exact(temp))));
 513       res = new SubLNode(n1, in(1));
 514     } else {
 515       return MulNode::Ideal(phase, can_reshape);
 516     }
 517   }
 518 
 519   if (sign_flip) {             // Need to negate result?
 520     res = phase->transform(res);// Transform, before making the zero con
 521     res = new SubLNode(phase->longcon(0),res);
 522   }
 523 
 524   return res;                   // Return final result
 525 }
 526 
 527 //=============================================================================
 528 //------------------------------mul_ring---------------------------------------
 529 // Compute the product type of two double ranges into this node.
 530 const Type *MulFNode::mul_ring(const Type *t0, const Type *t1) const {
 531   if( t0 == Type::FLOAT || t1 == Type::FLOAT ) return Type::FLOAT;
 532   return TypeF::make( t0->getf() * t1->getf() );
 533 }
 534 
 535 //------------------------------Ideal---------------------------------------
 536 // Check to see if we are multiplying by a constant 2 and convert to add, then try the regular MulNode::Ideal
 537 Node* MulFNode::Ideal(PhaseGVN* phase, bool can_reshape) {
 538   const TypeF *t2 = phase->type(in(2))->isa_float_constant();
 539 
 540   // x * 2 -> x + x
 541   if (t2 != nullptr && t2->getf() == 2) {
 542     Node* base = in(1);
 543     return new AddFNode(base, base);
 544   }
 545   return MulNode::Ideal(phase, can_reshape);
 546 }
 547 
 548 //=============================================================================
 549 //------------------------------Ideal------------------------------------------
 550 // Check to see if we are multiplying by a constant 2 and convert to add, then try the regular MulNode::Ideal
 551 Node* MulHFNode::Ideal(PhaseGVN* phase, bool can_reshape) {
 552   const TypeH* t2 = phase->type(in(2))->isa_half_float_constant();
 553 
 554   // x * 2 -> x + x
 555   if (t2 != nullptr && t2->getf() == 2) {
 556     Node* base = in(1);
 557     return new AddHFNode(base, base);
 558   }
 559   return MulNode::Ideal(phase, can_reshape);
 560 }
 561 
 562 // Compute the product type of two half float ranges into this node.
 563 const Type* MulHFNode::mul_ring(const Type* t0, const Type* t1) const {
 564   if (t0 == Type::HALF_FLOAT || t1 == Type::HALF_FLOAT) {
 565     return Type::HALF_FLOAT;
 566   }
 567   return TypeH::make(t0->getf() * t1->getf());
 568 }
 569 
 570 //=============================================================================
 571 //------------------------------mul_ring---------------------------------------
 572 // Compute the product type of two double ranges into this node.
 573 const Type *MulDNode::mul_ring(const Type *t0, const Type *t1) const {
 574   if( t0 == Type::DOUBLE || t1 == Type::DOUBLE ) return Type::DOUBLE;
 575   // We must be multiplying 2 double constants.
 576   return TypeD::make( t0->getd() * t1->getd() );
 577 }
 578 
 579 //------------------------------Ideal---------------------------------------
 580 // Check to see if we are multiplying by a constant 2 and convert to add, then try the regular MulNode::Ideal
 581 Node* MulDNode::Ideal(PhaseGVN* phase, bool can_reshape) {
 582   const TypeD *t2 = phase->type(in(2))->isa_double_constant();
 583 
 584   // x * 2 -> x + x
 585   if (t2 != nullptr && t2->getd() == 2) {
 586     Node* base = in(1);
 587     return new AddDNode(base, base);
 588   }
 589 
 590   return MulNode::Ideal(phase, can_reshape);
 591 }
 592 
 593 //=============================================================================
 594 //------------------------------Value------------------------------------------
 595 const Type* MulHiLNode::Value(PhaseGVN* phase) const {
 596   const Type *t1 = phase->type( in(1) );
 597   const Type *t2 = phase->type( in(2) );
 598   const Type *bot = bottom_type();
 599   return MulHiValue(t1, t2, bot);
 600 }
 601 
 602 const Type* UMulHiLNode::Value(PhaseGVN* phase) const {
 603   const Type *t1 = phase->type( in(1) );
 604   const Type *t2 = phase->type( in(2) );
 605   const Type *bot = bottom_type();
 606   return MulHiValue(t1, t2, bot);
 607 }
 608 
 609 // A common routine used by UMulHiLNode and MulHiLNode
 610 const Type* MulHiValue(const Type *t1, const Type *t2, const Type *bot) {
 611   // Either input is TOP ==> the result is TOP
 612   if( t1 == Type::TOP ) return Type::TOP;
 613   if( t2 == Type::TOP ) return Type::TOP;
 614 
 615   // Either input is BOTTOM ==> the result is the local BOTTOM
 616   if( (t1 == bot) || (t2 == bot) ||
 617       (t1 == Type::BOTTOM) || (t2 == Type::BOTTOM) )
 618     return bot;
 619 
 620   // It is not worth trying to constant fold this stuff!
 621   return TypeLong::LONG;
 622 }
 623 
 624 //=============================================================================
 625 //------------------------------mul_ring---------------------------------------
 626 // Supplied function returns the product of the inputs IN THE CURRENT RING.
 627 // For the logical operations the ring's MUL is really a logical AND function.
 628 // This also type-checks the inputs for sanity.  Guaranteed never to
 629 // be passed a TOP or BOTTOM type, these are filtered out by pre-check.
 630 const Type* AndINode::mul_ring(const Type* t1, const Type* t2) const {
 631   return RangeInference::infer_and(t1->is_int(), t2->is_int());
 632 }
 633 
 634 static bool AndIL_is_zero_element_under_mask(const PhaseGVN* phase, const Node* expr, const Node* mask, BasicType bt);
 635 
 636 const Type* AndINode::Value(PhaseGVN* phase) const {
 637   if (AndIL_is_zero_element_under_mask(phase, in(1), in(2), T_INT) ||
 638       AndIL_is_zero_element_under_mask(phase, in(2), in(1), T_INT)) {
 639     return TypeInt::ZERO;
 640   }
 641 
 642   return MulNode::Value(phase);
 643 }
 644 
 645 //------------------------------Identity---------------------------------------
 646 // Masking off the high bits of an unsigned load is not required
 647 Node* AndINode::Identity(PhaseGVN* phase) {
 648 
 649   // x & x => x
 650   if (in(1) == in(2)) {
 651     return in(1);
 652   }
 653 
 654   const TypeInt* t1 = phase->type(in(1))->is_int();
 655   const TypeInt* t2 = phase->type(in(2))->is_int();
 656 
 657   if ((~t1->_bits._ones & ~t2->_bits._zeros) == 0) {
 658     // All bits that might be 0 in in1 are known to be 0 in in2
 659     return in(2);
 660   }
 661 
 662   if ((~t2->_bits._ones & ~t1->_bits._zeros) == 0) {
 663     // All bits that might be 0 in in2 are known to be 0 in in1
 664     return in(1);
 665   }
 666 
 667   return MulNode::Identity(phase);
 668 }
 669 
 670 //------------------------------Ideal------------------------------------------
 671 Node *AndINode::Ideal(PhaseGVN *phase, bool can_reshape) {
 672   // Simplify (v1 + v2) & mask to v1 & mask or v2 & mask when possible.
 673   Node* progress = AndIL_sum_and_mask(phase, T_INT);
 674   if (progress != nullptr) {
 675     return progress;
 676   }
 677 
 678   // Convert "(~a) & (~b)" into "~(a | b)"
 679   if (AddNode::is_not(phase, in(1), T_INT) && AddNode::is_not(phase, in(2), T_INT)) {
 680     Node* or_a_b = new OrINode(in(1)->in(1), in(2)->in(1));
 681     Node* tn = phase->transform(or_a_b);
 682     return AddNode::make_not(phase, tn, T_INT);
 683   }
 684 
 685   // Special case constant AND mask
 686   const TypeInt *t2 = phase->type( in(2) )->isa_int();
 687   if( !t2 || !t2->is_con() ) return MulNode::Ideal(phase, can_reshape);
 688   const int mask = t2->get_con();
 689   Node *load = in(1);
 690   uint lop = load->Opcode();
 691 
 692   // Masking bits off of a Character?  Hi bits are already zero.
 693   if( lop == Op_LoadUS &&
 694       (mask & 0xFFFF0000) )     // Can we make a smaller mask?
 695     return new AndINode(load,phase->intcon(mask&0xFFFF));
 696 
 697   // Masking bits off of a Short?  Loading a Character does some masking
 698   if (can_reshape &&
 699       load->outcnt() == 1 && load->unique_out() == this) {
 700     if (lop == Op_LoadS && (mask & 0xFFFF0000) == 0 ) {
 701       Node* ldus = load->as_Load()->convert_to_unsigned_load(*phase);
 702       ldus = phase->transform(ldus);
 703       return new AndINode(ldus, phase->intcon(mask & 0xFFFF));
 704     }
 705 
 706     // Masking sign bits off of a Byte?  Do an unsigned byte load plus
 707     // an and.
 708     if (lop == Op_LoadB && (mask & 0xFFFFFF00) == 0) {
 709       Node* ldub = load->as_Load()->convert_to_unsigned_load(*phase);
 710       ldub = phase->transform(ldub);
 711       return new AndINode(ldub, phase->intcon(mask));
 712     }
 713   }
 714 
 715   // Masking off sign bits?  Dont make them!
 716   if( lop == Op_RShiftI ) {
 717     const TypeInt *t12 = phase->type(load->in(2))->isa_int();
 718     if( t12 && t12->is_con() ) { // Shift is by a constant
 719       int shift = t12->get_con();
 720       shift &= BitsPerJavaInteger-1;  // semantics of Java shifts
 721       const int sign_bits_mask = ~right_n_bits(BitsPerJavaInteger - shift);
 722       // If the AND'ing of the 2 masks has no bits, then only original shifted
 723       // bits survive.  NO sign-extension bits survive the maskings.
 724       if( (sign_bits_mask & mask) == 0 ) {
 725         // Use zero-fill shift instead
 726         Node *zshift = phase->transform(new URShiftINode(load->in(1),load->in(2)));
 727         return new AndINode( zshift, in(2) );
 728       }
 729     }
 730   }
 731 
 732   // Check for 'negate/and-1', a pattern emitted when someone asks for
 733   // 'mod 2'.  Negate leaves the low order bit unchanged (think: complement
 734   // plus 1) and the mask is of the low order bit.  Skip the negate.
 735   if( lop == Op_SubI && mask == 1 && load->in(1) &&
 736       phase->type(load->in(1)) == TypeInt::ZERO )
 737     return new AndINode( load->in(2), in(2) );
 738 
 739   return MulNode::Ideal(phase, can_reshape);
 740 }
 741 
 742 //=============================================================================
 743 //------------------------------mul_ring---------------------------------------
 744 // Supplied function returns the product of the inputs IN THE CURRENT RING.
 745 // For the logical operations the ring's MUL is really a logical AND function.
 746 // This also type-checks the inputs for sanity.  Guaranteed never to
 747 // be passed a TOP or BOTTOM type, these are filtered out by pre-check.
 748 const Type* AndLNode::mul_ring(const Type* t1, const Type* t2) const {
 749   return RangeInference::infer_and(t1->is_long(), t2->is_long());
 750 }
 751 
 752 const Type* AndLNode::Value(PhaseGVN* phase) const {
 753   if (AndIL_is_zero_element_under_mask(phase, in(1), in(2), T_LONG) ||
 754       AndIL_is_zero_element_under_mask(phase, in(2), in(1), T_LONG)) {
 755     return TypeLong::ZERO;
 756   }
 757 
 758   return MulNode::Value(phase);
 759 }
 760 
 761 //------------------------------Identity---------------------------------------
 762 // Masking off the high bits of an unsigned load is not required
 763 Node* AndLNode::Identity(PhaseGVN* phase) {
 764 
 765   // x & x => x
 766   if (in(1) == in(2)) {
 767     return in(1);
 768   }
 769 
 770   const TypeLong* t1 = phase->type(in(1))->is_long();
 771   const TypeLong* t2 = phase->type(in(2))->is_long();
 772 
 773   if ((~t1->_bits._ones & ~t2->_bits._zeros) == 0) {
 774     // All bits that might be 0 in in1 are known to be 0 in in2
 775     return in(2);
 776   }
 777 
 778   if ((~t2->_bits._ones & ~t1->_bits._zeros) == 0) {
 779     // All bits that might be 0 in in2 are known to be 0 in in1
 780     return in(1);
 781   }
 782 
 783   return MulNode::Identity(phase);
 784 }
 785 
 786 //------------------------------Ideal------------------------------------------
 787 Node *AndLNode::Ideal(PhaseGVN *phase, bool can_reshape) {
 788   // Simplify (v1 + v2) & mask to v1 & mask or v2 & mask when possible.
 789   Node* progress = AndIL_sum_and_mask(phase, T_LONG);
 790   if (progress != nullptr) {
 791     return progress;
 792   }
 793 
 794   // Convert "(~a) & (~b)" into "~(a | b)"
 795   if (AddNode::is_not(phase, in(1), T_LONG) && AddNode::is_not(phase, in(2), T_LONG)) {
 796     Node* or_a_b = new OrLNode(in(1)->in(1), in(2)->in(1));
 797     Node* tn = phase->transform(or_a_b);
 798     return AddNode::make_not(phase, tn, T_LONG);
 799   }
 800 
 801   // Special case constant AND mask
 802   const TypeLong *t2 = phase->type( in(2) )->isa_long();
 803   if( !t2 || !t2->is_con() ) return MulNode::Ideal(phase, can_reshape);
 804   const jlong mask = t2->get_con();
 805 
 806   Node* in1 = in(1);
 807   int op = in1->Opcode();
 808 
 809   // Are we masking a long that was converted from an int with a mask
 810   // that fits in 32-bits?  Commute them and use an AndINode.  Don't
 811   // convert masks which would cause a sign extension of the integer
 812   // value.  This check includes UI2L masks (0x00000000FFFFFFFF) which
 813   // would be optimized away later in Identity.
 814   if (op == Op_ConvI2L && (mask & UCONST64(0xFFFFFFFF80000000)) == 0) {
 815     Node* andi = new AndINode(in1->in(1), phase->intcon(mask));
 816     andi = phase->transform(andi);
 817     return new ConvI2LNode(andi);
 818   }
 819 
 820   // Masking off sign bits?  Dont make them!
 821   if (op == Op_RShiftL) {
 822     const TypeInt* t12 = phase->type(in1->in(2))->isa_int();
 823     if( t12 && t12->is_con() ) { // Shift is by a constant
 824       int shift = t12->get_con();
 825       shift &= BitsPerJavaLong - 1;  // semantics of Java shifts
 826       if (shift != 0) {
 827         const julong sign_bits_mask = ~(((julong)CONST64(1) << (julong)(BitsPerJavaLong - shift)) -1);
 828         // If the AND'ing of the 2 masks has no bits, then only original shifted
 829         // bits survive.  NO sign-extension bits survive the maskings.
 830         if( (sign_bits_mask & mask) == 0 ) {
 831           // Use zero-fill shift instead
 832           Node *zshift = phase->transform(new URShiftLNode(in1->in(1), in1->in(2)));
 833           return new AndLNode(zshift, in(2));
 834         }
 835       }
 836     }
 837   }
 838 
 839   // Search for GraphKit::mark_word_test patterns and fold the test if the result is statically known
 840   Node* load1 = in(1);
 841   Node* load2 = nullptr;
 842   if (load1->is_Phi() && phase->type(load1)->isa_long()) {
 843     load1 = in(1)->in(1);
 844     load2 = in(1)->in(2);
 845   }
 846   if (load1 != nullptr && load1->is_Load() && phase->type(load1)->isa_long() &&
 847       (load2 == nullptr || (load2->is_Load() && phase->type(load2)->isa_long()))) {
 848     const TypePtr* adr_t1 = phase->type(load1->in(MemNode::Address))->isa_ptr();
 849     const TypePtr* adr_t2 = (load2 != nullptr) ? phase->type(load2->in(MemNode::Address))->isa_ptr() : nullptr;
 850     if (adr_t1 != nullptr && adr_t1->offset() == oopDesc::mark_offset_in_bytes() &&
 851         (load2 == nullptr || (adr_t2 != nullptr && adr_t2->offset() == in_bytes(Klass::prototype_header_offset())))) {
 852       if (mask == markWord::inline_type_pattern) {
 853         if (adr_t1->is_inlinetypeptr()) {
 854           set_req_X(1, in(2), phase);
 855           return this;
 856         } else if (!adr_t1->can_be_inline_type()) {
 857           set_req_X(1, phase->longcon(0), phase);
 858           return this;
 859         }
 860       } else if (mask == markWord::null_free_array_bit_in_place) {
 861         if (adr_t1->is_null_free()) {
 862           set_req_X(1, in(2), phase);
 863           return this;
 864         } else if (adr_t1->is_not_null_free()) {
 865           set_req_X(1, phase->longcon(0), phase);
 866           return this;
 867         }
 868       } else if (mask == markWord::flat_array_bit_in_place) {
 869         if (adr_t1->is_flat()) {
 870           set_req_X(1, in(2), phase);
 871           return this;
 872         } else if (adr_t1->is_not_flat()) {
 873           set_req_X(1, phase->longcon(0), phase);
 874           return this;
 875         }
 876       }
 877     }
 878   }
 879 
 880   return MulNode::Ideal(phase, can_reshape);
 881 }
 882 
 883 LShiftNode* LShiftNode::make(Node* in1, Node* in2, BasicType bt) {
 884   switch (bt) {
 885     case T_INT:
 886       return new LShiftINode(in1, in2);
 887     case T_LONG:
 888       return new LShiftLNode(in1, in2);
 889     default:
 890       fatal("Not implemented for %s", type2name(bt));
 891   }
 892   return nullptr;
 893 }
 894 
 895 // Returns whether the shift amount is constant or effectively constant (low bits known).
 896 //
 897 // Parameters:
 898 //   masked_shift - always initialized to 0; if the function returns true, it indicates
 899 //                  the masked shift amount.
 900 //   replace      - always initialized to false; if the function returns true, it indicates
 901 //                  whether the shift_node's shift count input should be replaced with masked_shift.
 902 static bool mask_shift_amount(PhaseGVN* phase, const Node* shift_node, uint num_bits, uint& masked_shift, bool& replace) {
 903   masked_shift = 0;
 904   replace = false;
 905 
 906   const TypeInt* tcount = phase->type(shift_node->in(2))->isa_int();
 907 
 908   if (tcount != nullptr) {
 909     uint mask = num_bits - 1;
 910     // Canonicalize shift count via type-level masking to expose constants
 911     const TypeInt* masked_type = RangeInference::infer_and(tcount, TypeInt::make(mask));
 912     if (masked_type != nullptr && masked_type->is_con()) {
 913       masked_shift = masked_type->get_con();
 914       replace = !tcount->is_con() || (tcount->get_con() != (int)masked_shift);
 915       return true;
 916     }
 917   }
 918   return false;
 919 }
 920 
 921 // Convenience for when we don't care about the 'replace' output.
 922 static bool mask_shift_amount(PhaseGVN* phase, const Node* shift_node, uint num_bits, uint& masked_shift) {
 923   bool unused;
 924   return mask_shift_amount(phase, shift_node, num_bits, masked_shift, unused /*replace*/);
 925 }
 926 
 927 // Use this in ::Ideal only with shiftNode == this!
 928 // Sets masked_shift to the effective masked shift amount if constant or 0 if not constant.
 929 // Returns shift_node if the shift amount input node was modified, nullptr otherwise.
 930 static Node* mask_and_replace_shift_amount(PhaseGVN* phase, Node* shift_node, uint num_bits, uint& masked_shift) {
 931   if (bool replace; mask_shift_amount(phase, shift_node, num_bits, masked_shift, replace)) {
 932     if (masked_shift == 0) {
 933       // Let Identity() handle 0 shift count.
 934       return nullptr;
 935     }
 936 
 937     if (replace) {
 938       shift_node->set_req(2, phase->intcon(masked_shift)); // Replace shift count with masked value.
 939 
 940       // We need to notify the caller that the graph was reshaped, as Ideal needs
 941       // to return the root of the reshaped graph if any change was made.
 942       return shift_node;
 943     }
 944   }
 945 
 946   return nullptr;
 947 }
 948 
 949 // Called with
 950 //   outer_shift = (_ << rhs_outer)
 951 // We are looking for the pattern:
 952 //   outer_shift = ((X << rhs_inner) << rhs_outer)
 953 //   where rhs_outer and rhs_inner are constant
 954 //   we denote inner_shift the nested expression (X << rhs_inner)
 955 //   con_inner = rhs_inner % nbits and con_outer = rhs_outer % nbits
 956 //   where nbits is the number of bits of the shifts
 957 //
 958 // There are 2 cases:
 959 // if con_outer + con_inner >= nbits => 0
 960 // if con_outer + con_inner < nbits => X << (con_outer + con_inner)
 961 static Node* collapse_nested_shift_left(PhaseGVN* phase, const Node* outer_shift, uint con_outer, BasicType bt) {
 962   assert(bt == T_LONG || bt == T_INT, "Unexpected type");
 963   const Node* inner_shift = outer_shift->in(1);
 964   if (inner_shift->Opcode() != Op_LShift(bt)) {
 965     return nullptr;
 966   }
 967 
 968   uint nbits = bits_per_java_integer(bt);
 969   uint con_inner;
 970   if (!mask_shift_amount(phase, inner_shift, nbits, con_inner)) {
 971     return nullptr;
 972   }
 973 
 974   if (con_inner == 0) {
 975     // We let the Identity() of the inner shift do its job.
 976     return nullptr;
 977   }
 978 
 979   if (con_outer + con_inner >= nbits) {
 980     // While it might be tempting to use
 981     // phase->zerocon(bt);
 982     // it would be incorrect: zerocon caches nodes, while Ideal is only allowed
 983     // to return a new node, this or nullptr, but not an old (cached) node.
 984     return ConNode::make(TypeInteger::zero(bt));
 985   }
 986 
 987   // con0 + con1 < nbits ==> actual shift happens now
 988   Node* con0_plus_con1 = phase->intcon(con_outer + con_inner);
 989   return LShiftNode::make(inner_shift->in(1), con0_plus_con1, bt);
 990 }
 991 
 992 //------------------------------Identity---------------------------------------
 993 Node* LShiftINode::Identity(PhaseGVN* phase) {
 994   return IdentityIL(phase, T_INT);
 995 }
 996 
 997 Node* LShiftNode::IdealIL(PhaseGVN* phase, bool can_reshape, BasicType bt) {
 998   uint con;
 999   uint num_bits = bits_per_java_integer(bt);
1000   Node* progress = mask_and_replace_shift_amount(phase, this, num_bits, con);
1001   if (con == 0) {
1002     return nullptr;
1003   }
1004 
1005   // If the right input is a constant, and the left input is an add of a
1006   // constant, flatten the tree: (X+con1)<<con0 ==> X<<con0 + con1<<con0
1007   Node* add1 = in(1);
1008   int add1_op = add1->Opcode();
1009   if (add1_op == Op_Add(bt)) {    // Left input is an add?
1010     assert(add1 != add1->in(1), "dead loop in LShiftINode::Ideal");
1011 
1012     // Transform is legal, but check for profit.  Avoid breaking 'i2s'
1013     // and 'i2b' patterns which typically fold into 'StoreC/StoreB'.
1014     if (bt != T_INT || con < 16) {
1015       // Left input is an add of the same number?
1016       if (con != (num_bits - 1) && add1->in(1) == add1->in(2)) {
1017         // Convert "(x + x) << c0" into "x << (c0 + 1)"
1018         // In general, this optimization cannot be applied for c0 == 31 (for LShiftI) since
1019         // 2x << 31 != x << 32 = x << 0 = x (e.g. x = 1: 2 << 31 = 0 != 1)
1020         // or c0 != 63 (for LShiftL) because:
1021         // (x + x) << 63 = 2x << 63, while
1022         // (x + x) << 63 --transform--> x << 64 = x << 0 = x (!= 2x << 63, for example for x = 1)
1023         // According to the Java spec, chapter 15.19, we only consider the six lowest-order bits of the right-hand operand
1024         // (i.e. "right-hand operand" & 0b111111). Therefore, x << 64 is the same as x << 0 (64 = 0b10000000 & 0b0111111 = 0).
1025         return LShiftNode::make(add1->in(1), phase->intcon(con + 1), bt);
1026       }
1027 
1028       // Left input is an add of a constant?
1029       const TypeInteger* t12 = phase->type(add1->in(2))->isa_integer(bt);
1030       if (t12 != nullptr && t12->is_con()) { // Left input is an add of a con?
1031         // Compute X << con0
1032         Node* lsh = phase->transform(LShiftNode::make(add1->in(1), in(2), bt));
1033         // Compute X<<con0 + (con1<<con0)
1034         return AddNode::make(lsh, phase->integercon(java_shift_left(t12->get_con_as_long(bt), con, bt), bt), bt);
1035       }
1036     }
1037   }
1038   // Check for "(con0 - X) << con1"
1039   // Transform is legal, but check for profit.  Avoid breaking 'i2s'
1040   // and 'i2b' patterns which typically fold into 'StoreC/StoreB'.
1041   if (add1_op == Op_Sub(bt) && (bt != T_INT || con < 16)) {    // Left input is a sub?
1042     // Left input is a sub from a constant?
1043     const TypeInteger* t11 = phase->type(add1->in(1))->isa_integer(bt);
1044     if (t11 != nullptr && t11->is_con()) {
1045       // Compute X << con0
1046       Node* lsh = phase->transform(LShiftNode::make(add1->in(2), in(2), bt));
1047       // Compute (con1<<con0) - (X<<con0)
1048       return SubNode::make(phase->integercon(java_shift_left(t11->get_con_as_long(bt), con, bt), bt), lsh, bt);
1049     }
1050   }
1051 
1052   // Check for "(x >> C1) << C2"
1053   if (add1_op == Op_RShift(bt) || add1_op == Op_URShift(bt)) {
1054     uint add1Con;
1055     mask_shift_amount(phase, add1, num_bits, add1Con);
1056 
1057     // Special case C1 == C2, which just masks off low bits
1058     if (add1Con > 0 && con == add1Con) {
1059       // Convert to "(x & -(1 << C2))"
1060       return  MulNode::make_and(add1->in(1), phase->integercon(java_negate(java_shift_left(1, con, bt), bt), bt), bt);
1061     } else {
1062       // Wait until the right shift has been sharpened to the correct count
1063       if (add1Con > 0) {
1064         // As loop parsing can produce LShiftI nodes, we should wait until the graph is fully formed
1065         // to apply optimizations, otherwise we can inadvertently stop vectorization opportunities.
1066         if (phase->is_IterGVN()) {
1067           if (con > add1Con) {
1068             // Creates "(x << (C2 - C1)) & -(1 << C2)"
1069             Node* lshift = phase->transform(LShiftNode::make(add1->in(1), phase->intcon(con - add1Con), bt));
1070             return MulNode::make_and(lshift, phase->integercon(java_negate(java_shift_left(1, con, bt), bt), bt), bt);
1071           } else {
1072             assert(con < add1Con, "must be (%d < %d)", con, add1Con);
1073             // Creates "(x >> (C1 - C2)) & -(1 << C2)"
1074 
1075             // Handle logical and arithmetic shifts
1076             Node* rshift;
1077             if (add1_op == Op_RShift(bt)) {
1078               rshift = phase->transform(RShiftNode::make(add1->in(1), phase->intcon(add1Con - con), bt));
1079             } else {
1080               rshift = phase->transform(URShiftNode::make(add1->in(1), phase->intcon(add1Con - con), bt));
1081             }
1082 
1083             return MulNode::make_and(rshift, phase->integercon(java_negate(java_shift_left(1,  con, bt)), bt), bt);
1084           }
1085         } else {
1086           phase->record_for_igvn(this);
1087         }
1088       }
1089     }
1090   }
1091 
1092   // Check for "((x >> C1) & Y) << C2"
1093   if (add1_op == Op_And(bt)) {
1094     Node* add2 = add1->in(1);
1095     int add2_op = add2->Opcode();
1096     if (add2_op == Op_RShift(bt) || add2_op == Op_URShift(bt)) {
1097       // Special case C1 == C2, which just masks off low bits
1098       if (add2->in(2) == in(2)) {
1099         // Convert to "(x & (Y << C2))"
1100         Node* y_sh = phase->transform(LShiftNode::make(add1->in(2), phase->intcon(con), bt));
1101         return MulNode::make_and(add2->in(1), y_sh, bt);
1102       }
1103 
1104       uint add2Con;
1105       if (mask_shift_amount(phase, add2, num_bits, add2Con) && add2Con > 0) {
1106         if (phase->is_IterGVN()) {
1107           // Convert to "((x >> C1) << C2) & (Y << C2)"
1108 
1109           // Make "(x >> C1) << C2", which will get folded away by the rule above
1110           Node* x_sh = phase->transform(LShiftNode::make(add2, phase->intcon(con), bt));
1111           // Make "Y << C2", which will simplify when Y is a constant
1112           Node* y_sh = phase->transform(LShiftNode::make(add1->in(2), phase->intcon(con), bt));
1113 
1114           return MulNode::make_and(x_sh, y_sh, bt);
1115         } else {
1116           phase->record_for_igvn(this);
1117         }
1118       }
1119     }
1120   }
1121 
1122   // Check for ((x & ((1<<(32-c0))-1)) << c0) which ANDs off high bits
1123   // before shifting them away.
1124   const jlong bits_mask = max_unsigned_integer(bt) >> con;
1125   assert(bt != T_INT || bits_mask == right_n_bits(num_bits - con), "inconsistent");
1126   if (add1_op == Op_And(bt) &&
1127       phase->type(add1->in(2)) == TypeInteger::make(bits_mask, bt)) {
1128     return LShiftNode::make(add1->in(1), in(2), bt);
1129   }
1130 
1131   // Collapse nested left-shifts with constant rhs:
1132   // (X << con1) << con2 ==> X << (con1 + con2)
1133   Node* doubleShift = collapse_nested_shift_left(phase, this, con, bt);
1134   if (doubleShift != nullptr) {
1135     return doubleShift;
1136   }
1137 
1138   return progress;
1139 }
1140 
1141 //------------------------------Ideal------------------------------------------
1142 Node* LShiftINode::Ideal(PhaseGVN *phase, bool can_reshape) {
1143   return IdealIL(phase, can_reshape, T_INT);
1144 }
1145 
1146 const Type* LShiftNode::ValueIL(PhaseGVN* phase, BasicType bt) const {
1147   const Type* t1 = phase->type(in(1));
1148   const Type* t2 = phase->type(in(2));
1149   // Either input is TOP ==> the result is TOP
1150   if (t1 == Type::TOP) {
1151     return Type::TOP;
1152   }
1153   if (t2 == Type::TOP) {
1154     return Type::TOP;
1155   }
1156 
1157   // Left input is ZERO ==> the result is ZERO.
1158   if (t1 == TypeInteger::zero(bt)) {
1159     return TypeInteger::zero(bt);
1160   }
1161   // Shift by zero does nothing
1162   if (t2 == TypeInt::ZERO) {
1163     return t1;
1164   }
1165 
1166   // If nothing is known about the shift amount then the result is BOTTOM
1167   if (t2 == TypeInt::INT) {
1168     return TypeInteger::bottom(bt);
1169   }
1170 
1171   const TypeInteger* r1 = t1->is_integer(bt); // Handy access
1172   // Since the shift semantics in Java take into account only the bottom five
1173   // bits for ints and the bottom six bits for longs, we can further constrain
1174   // the range of values of the shift amount by ANDing with the right mask based
1175   // on whether the type is int or long.
1176   const TypeInt* mask = TypeInt::make(bits_per_java_integer(bt) - 1);
1177   const TypeInt* r2 = RangeInference::infer_and(t2->is_int(), mask);
1178 
1179   if (!r2->is_con()) {
1180     return TypeInteger::bottom(bt);
1181   }
1182 
1183   uint shift = r2->get_con();
1184   // Shift by a multiple of 32/64 does nothing:
1185   if (shift == 0) {
1186     return t1;
1187   }
1188 
1189   // If the shift is a constant, shift the bounds of the type,
1190   // unless this could lead to an overflow.
1191   if (!r1->is_con()) {
1192 #ifdef ASSERT
1193     if (bt == T_INT) {
1194       jlong lo = r1->lo_as_long(), hi = r1->hi_as_long();
1195       jint lo_int = r1->is_int()->_lo, hi_int = r1->is_int()->_hi;
1196       assert((java_shift_right(java_shift_left(lo, shift, bt),  shift, bt) == lo) == (((lo_int << shift) >> shift) == lo_int), "inconsistent");
1197       assert((java_shift_right(java_shift_left(hi, shift, bt),  shift, bt) == hi) == (((hi_int << shift) >> shift) == hi_int), "inconsistent");
1198     }
1199 #endif
1200 
1201     if (bt == T_INT) {
1202         return RangeInference::infer_lshift(r1->is_int(), shift);
1203     }
1204 
1205     return RangeInference::infer_lshift(r1->is_long(), shift);
1206   }
1207 
1208   return TypeInteger::make(java_shift_left(r1->get_con_as_long(bt), shift, bt), bt);
1209 }
1210 
1211 //------------------------------Value------------------------------------------
1212 const Type* LShiftINode::Value(PhaseGVN* phase) const {
1213   return ValueIL(phase, T_INT);
1214 }
1215 
1216 Node* LShiftNode::IdentityIL(PhaseGVN* phase, BasicType bt) {
1217   uint count;
1218   if (mask_shift_amount(phase, this, bits_per_java_integer(bt), count) && count == 0) {
1219     // Shift by a multiple of 32/64 does nothing
1220     return in(1);
1221   }
1222   return this;
1223 }
1224 
1225 //=============================================================================
1226 //------------------------------Identity---------------------------------------
1227 Node* LShiftLNode::Identity(PhaseGVN* phase) {
1228   return IdentityIL(phase, T_LONG);
1229 }
1230 
1231 //------------------------------Ideal------------------------------------------
1232 Node* LShiftLNode::Ideal(PhaseGVN* phase, bool can_reshape) {
1233   return IdealIL(phase, can_reshape, T_LONG);
1234 }
1235 
1236 //------------------------------Value------------------------------------------
1237 const Type* LShiftLNode::Value(PhaseGVN* phase) const {
1238   return ValueIL(phase, T_LONG);
1239 }
1240 
1241 RShiftNode* RShiftNode::make(Node* in1, Node* in2, BasicType bt) {
1242   switch (bt) {
1243     case T_INT:
1244       return new RShiftINode(in1, in2);
1245     case T_LONG:
1246       return new RShiftLNode(in1, in2);
1247     default:
1248       fatal("Not implemented for %s", type2name(bt));
1249   }
1250   return nullptr;
1251 }
1252 
1253 
1254 //=============================================================================
1255 //------------------------------Identity---------------------------------------
1256 Node* RShiftNode::IdentityIL(PhaseGVN* phase, BasicType bt) {
1257   uint count;
1258   uint num_bits = bits_per_java_integer(bt);
1259   if (mask_shift_amount(phase, this, num_bits, count)) {
1260     if (count == 0) {
1261       // Shift by a multiple of 32/64 does nothing
1262       return in(1);
1263     }
1264     // Check for useless sign-masking
1265     uint lshift_count;
1266     if (in(1)->Opcode() == Op_LShift(bt) &&
1267         in(1)->req() == 3 &&
1268         // Compare shift counts by value, not by node pointer, to also match a not-yet-normalized
1269         // negative constant (e.g. -1 vs 31)
1270         mask_shift_amount(phase, in(1), num_bits, lshift_count)) {
1271       if (count == lshift_count) {
1272         // Compute masks for which this shifting doesn't change
1273         jlong lo = (CONST64(-1) << (num_bits - count - 1)); // FFFF8000
1274         jlong hi = ~lo;                                                            // 00007FFF
1275         const TypeInteger* t11 = phase->type(in(1)->in(1))->isa_integer(bt);
1276         if (t11 == nullptr) {
1277           return this;
1278         }
1279         // Does actual value fit inside of mask?
1280         if (lo <= t11->lo_as_long() && t11->hi_as_long() <= hi) {
1281           return in(1)->in(1);      // Then shifting is a nop
1282         }
1283       }
1284     }
1285   }
1286   return this;
1287 }
1288 
1289 Node* RShiftINode::Identity(PhaseGVN* phase) {
1290   return IdentityIL(phase, T_INT);
1291 }
1292 
1293 Node* RShiftNode::IdealIL(PhaseGVN* phase, bool can_reshape, BasicType bt) {
1294   // Inputs may be TOP if they are dead.
1295   const TypeInteger* t1 = phase->type(in(1))->isa_integer(bt);
1296   if (t1 == nullptr) {
1297     return NodeSentinel;        // Left input is an integer
1298   }
1299 
1300   uint shift;
1301   Node* progress = mask_and_replace_shift_amount(phase, this, bits_per_java_integer(bt), shift);
1302   if (shift == 0) {
1303     return NodeSentinel;
1304   }
1305 
1306   // Check for (x & 0xFF000000) >> 24, whose mask can be made smaller.
1307   // and convert to (x >> 24) & (0xFF000000 >> 24) = x >> 24
1308   // Such expressions arise normally from shift chains like (byte)(x >> 24).
1309   const Node* and_node = in(1);
1310   if (and_node->Opcode() != Op_And(bt)) {
1311     return progress;
1312   }
1313   const TypeInteger* mask_t = phase->type(and_node->in(2))->isa_integer(bt);
1314   if (mask_t != nullptr && mask_t->is_con()) {
1315     jlong maskbits = mask_t->get_con_as_long(bt);
1316     // Convert to "(x >> shift) & (mask >> shift)"
1317     Node* shr_nomask = phase->transform(RShiftNode::make(and_node->in(1), in(2), bt));
1318     return MulNode::make_and(shr_nomask, phase->integercon(maskbits >> shift, bt), bt);
1319   }
1320 
1321   return progress;
1322 }
1323 
1324 Node* RShiftINode::Ideal(PhaseGVN* phase, bool can_reshape) {
1325   Node* progress = IdealIL(phase, can_reshape, T_INT);
1326   if (progress == NodeSentinel) {
1327     return nullptr;
1328   }
1329   if (progress != nullptr) {
1330     return progress;
1331   }
1332   uint shift;
1333   progress = mask_and_replace_shift_amount(phase, this, BitsPerJavaInteger, shift);
1334   assert(shift != 0, "handled by IdealIL");
1335 
1336   // Check for "(short[i] <<16)>>16" which simply sign-extends
1337   const Node *shl = in(1);
1338   if (shl->Opcode() != Op_LShiftI) {
1339     return progress;
1340   }
1341 
1342   const TypeInt* left_shift_t = phase->type(shl->in(2))->isa_int();
1343   if (left_shift_t == nullptr) {
1344     return progress;
1345   }
1346   if (shift == 16 && left_shift_t->is_con(16)) {
1347     Node *ld = shl->in(1);
1348     if (ld->Opcode() == Op_LoadS) {
1349       // Sign extension is just useless here.  Return a RShiftI of zero instead
1350       // returning 'ld' directly.  We cannot return an old Node directly as
1351       // that is the job of 'Identity' calls and Identity calls only work on
1352       // direct inputs ('ld' is an extra Node removed from 'this').  The
1353       // combined optimization requires Identity only return direct inputs.
1354       set_req_X(1, ld, phase);
1355       set_req_X(2, phase->intcon(0), phase);
1356       return this;
1357     }
1358     else if (can_reshape &&
1359              ld->Opcode() == Op_LoadUS &&
1360              ld->outcnt() == 1 && ld->unique_out() == shl)
1361       // Replace zero-extension-load with sign-extension-load
1362       return ld->as_Load()->convert_to_signed_load(*phase);
1363   }
1364 
1365   // Check for "(byte[i] <<24)>>24" which simply sign-extends
1366   if (shift == 24 && left_shift_t->is_con(24)) {
1367     Node *ld = shl->in(1);
1368     if (ld->Opcode() == Op_LoadB) {
1369       // Sign extension is just useless here
1370       set_req_X(1, ld, phase);
1371       set_req_X(2, phase->intcon(0), phase);
1372       return this;
1373     }
1374   }
1375 
1376   return progress;
1377 }
1378 
1379 const Type* RShiftNode::ValueIL(PhaseGVN* phase, BasicType bt) const {
1380   const Type* t1 = phase->type(in(1));
1381   const Type* t2 = phase->type(in(2));
1382   // Either input is TOP ==> the result is TOP
1383   if (t1 == Type::TOP) {
1384     return Type::TOP;
1385   }
1386   if (t2 == Type::TOP) {
1387     return Type::TOP;
1388   }
1389 
1390   // Left input is ZERO ==> the result is ZERO.
1391   if (t1 == TypeInteger::zero(bt)) {
1392     return TypeInteger::zero(bt);
1393   }
1394   // Shift by zero does nothing
1395   if (t2 == TypeInt::ZERO) {
1396     return t1;
1397   }
1398 
1399   // Either input is BOTTOM ==> the result is BOTTOM
1400   if (t1 == Type::BOTTOM || t2 == Type::BOTTOM) {
1401     return TypeInteger::bottom(bt);
1402   }
1403 
1404   const TypeInteger* r1 = t1->isa_integer(bt);
1405   const TypeInt* r2 = t2->isa_int();
1406 
1407   // If the shift is a constant, just shift the bounds of the type.
1408   // For example, if the shift is 31/63, we just propagate sign bits.
1409   if (!r1->is_con() && r2->is_con()) {
1410     uint shift = r2->get_con();
1411     shift &= bits_per_java_integer(bt) - 1;  // semantics of Java shifts
1412     // Shift by a multiple of 32/64 does nothing:
1413     if (shift == 0) {
1414       return t1;
1415     }
1416     // Calculate reasonably aggressive bounds for the result.
1417     // This is necessary if we are to correctly type things
1418     // like (x<<24>>24) == ((byte)x).
1419     jlong lo = r1->lo_as_long() >> (jint)shift;
1420     jlong hi = r1->hi_as_long() >> (jint)shift;
1421     assert(lo <= hi, "must have valid bounds");
1422 #ifdef ASSERT
1423    if (bt == T_INT) {
1424      jint lo_verify = checked_cast<jint>(r1->lo_as_long()) >> (jint)shift;
1425      jint hi_verify = checked_cast<jint>(r1->hi_as_long()) >> (jint)shift;
1426      assert((checked_cast<jint>(lo) == lo_verify) && (checked_cast<jint>(hi) == hi_verify), "inconsistent");
1427    }
1428 #endif
1429     const TypeInteger* ti = TypeInteger::make(lo, hi, MAX2(r1->_widen,r2->_widen), bt);
1430 #ifdef ASSERT
1431     // Make sure we get the sign-capture idiom correct.
1432     if (shift == bits_per_java_integer(bt) - 1) {
1433       if (r1->lo_as_long() >= 0) {
1434         assert(ti == TypeInteger::zero(bt),    ">>31/63 of + is  0");
1435       }
1436       if (r1->hi_as_long() <  0) {
1437         assert(ti == TypeInteger::minus_1(bt), ">>31/63 of - is -1");
1438       }
1439     }
1440 #endif
1441     return ti;
1442   }
1443 
1444   if (!r1->is_con() || !r2->is_con()) {
1445     // If the left input is non-negative the result must also be non-negative, regardless of what the right input is.
1446     if (r1->lo_as_long() >= 0) {
1447       return TypeInteger::make(0, r1->hi_as_long(), MAX2(r1->_widen, r2->_widen), bt);
1448     }
1449 
1450     // Conversely, if the left input is negative then the result must be negative.
1451     if (r1->hi_as_long() <= -1) {
1452       return TypeInteger::make(r1->lo_as_long(), -1, MAX2(r1->_widen, r2->_widen), bt);
1453     }
1454 
1455     return TypeInteger::bottom(bt);
1456   }
1457 
1458   // Signed shift right
1459   return TypeInteger::make(r1->get_con_as_long(bt) >> (r2->get_con() & (bits_per_java_integer(bt) - 1)), bt);
1460 }
1461 
1462 const Type* RShiftINode::Value(PhaseGVN* phase) const {
1463   return ValueIL(phase, T_INT);
1464 }
1465 
1466 //=============================================================================
1467 //------------------------------Identity---------------------------------------
1468 Node* RShiftLNode::Identity(PhaseGVN* phase) {
1469   return IdentityIL(phase, T_LONG);
1470 }
1471 
1472 Node* RShiftLNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1473   Node* progress = IdealIL(phase, can_reshape, T_LONG);
1474   if (progress == NodeSentinel) {
1475     return nullptr;
1476   }
1477   return progress;
1478 }
1479 
1480 const Type* RShiftLNode::Value(PhaseGVN* phase) const {
1481   return ValueIL(phase, T_LONG);
1482 }
1483 
1484 URShiftNode* URShiftNode::make(Node* in1, Node* in2, BasicType bt) {
1485   switch (bt) {
1486     case T_INT:
1487       return new URShiftINode(in1, in2);
1488     case T_LONG:
1489       return new URShiftLNode(in1, in2);
1490     default:
1491       fatal("Not implemented for %s", type2name(bt));
1492   }
1493   return nullptr;
1494 }
1495 
1496 //=============================================================================
1497 //------------------------------Identity---------------------------------------
1498 Node* URShiftINode::Identity(PhaseGVN* phase) {
1499   uint count;
1500   if (mask_shift_amount(phase, this, BitsPerJavaInteger, count) && count == 0) {
1501     // Shift by a multiple of 32 does nothing
1502     return in(1);
1503   }
1504 
1505   // Check for "((x << LogBytesPerWord) + (wordSize-1)) >> LogBytesPerWord" which is just "x".
1506   // Happens during new-array length computation.
1507   // Safe if 'x' is in the range [0..(max_int>>LogBytesPerWord)]
1508   Node *add = in(1);
1509   if (add->Opcode() == Op_AddI) {
1510     const TypeInt *t2 = phase->type(add->in(2))->isa_int();
1511     if (t2 && t2->is_con(wordSize - 1) &&
1512         add->in(1)->Opcode() == Op_LShiftI) {
1513       // Check that shift_counts are LogBytesPerWord.
1514       Node          *lshift_count   = add->in(1)->in(2);
1515       const TypeInt *t_lshift_count = phase->type(lshift_count)->isa_int();
1516       if (t_lshift_count && t_lshift_count->is_con(LogBytesPerWord) &&
1517           t_lshift_count == phase->type(in(2))) {
1518         Node          *x   = add->in(1)->in(1);
1519         const TypeInt *t_x = phase->type(x)->isa_int();
1520         if (t_x != nullptr && 0 <= t_x->_lo && t_x->_hi <= (max_jint>>LogBytesPerWord)) {
1521           return x;
1522         }
1523       }
1524     }
1525   }
1526 
1527   return (phase->type(in(2))->higher_equal(TypeInt::ZERO)) ? in(1) : this;
1528 }
1529 
1530 //------------------------------Ideal------------------------------------------
1531 Node* URShiftINode::Ideal(PhaseGVN* phase, bool can_reshape) {
1532   uint con;
1533   Node* progress = mask_and_replace_shift_amount(phase, this, BitsPerJavaInteger, con);
1534   if (con == 0) {
1535     return nullptr;
1536   }
1537 
1538   // We'll be wanting the right-shift amount as a mask of that many bits
1539   const int mask = right_n_bits(BitsPerJavaInteger - con);
1540 
1541   int in1_op = in(1)->Opcode();
1542 
1543   // Check for ((x>>>a)>>>b) and replace with (x>>>(a+b)) when a+b < 32
1544   if( in1_op == Op_URShiftI ) {
1545     const TypeInt *t12 = phase->type( in(1)->in(2) )->isa_int();
1546     if( t12 && t12->is_con() ) { // Right input is a constant
1547       assert( in(1) != in(1)->in(1), "dead loop in URShiftINode::Ideal" );
1548       const int con2 = t12->get_con() & 31; // Shift count is always masked
1549       const int con3 = con+con2;
1550       if( con3 < 32 )           // Only merge shifts if total is < 32
1551         return new URShiftINode( in(1)->in(1), phase->intcon(con3) );
1552     }
1553   }
1554 
1555   // Check for ((x << z) + Y) >>> z.  Replace with x + con>>>z
1556   // The idiom for rounding to a power of 2 is "(Q+(2^z-1)) >>> z".
1557   // If Q is "X << z" the rounding is useless.  Look for patterns like
1558   // ((X<<Z) + Y) >>> Z  and replace with (X + Y>>>Z) & Z-mask.
1559   Node *add = in(1);
1560   if (in1_op == Op_AddI) {
1561     Node *lshl = add->in(1);
1562     Node *y    = add->in(2);
1563     if (lshl->Opcode() != Op_LShiftI) {
1564       lshl = add->in(2);
1565       y    = add->in(1);
1566     }
1567     // Compare shift counts by value, not by node pointer, to also match a not-yet-normalized
1568     // negative constant (e.g. -1 vs 31)
1569     uint lshl_con;
1570     if (lshl->Opcode() == Op_LShiftI &&
1571         mask_shift_amount(phase, lshl, BitsPerJavaInteger, lshl_con) &&
1572         lshl_con == con) {
1573       Node *y_z = phase->transform(new URShiftINode(y, in(2)));
1574       Node *sum = phase->transform(new AddINode(lshl->in(1), y_z));
1575       return new AndINode(sum, phase->intcon(mask));
1576     }
1577   }
1578 
1579   // Check for (x & mask) >>> z.  Replace with (x >>> z) & (mask >>> z)
1580   // This shortens the mask.  Also, if we are extracting a high byte and
1581   // storing it to a buffer, the mask will be removed completely.
1582   Node *andi = in(1);
1583   if( in1_op == Op_AndI ) {
1584     const TypeInt *t3 = phase->type( andi->in(2) )->isa_int();
1585     if( t3 && t3->is_con() ) { // Right input is a constant
1586       jint mask2 = t3->get_con();
1587       mask2 >>= con;  // *signed* shift downward (high-order zeroes do not help)
1588       Node *newshr = phase->transform( new URShiftINode(andi->in(1), in(2)) );
1589       return new AndINode(newshr, phase->intcon(mask2));
1590       // The negative values are easier to materialize than positive ones.
1591       // A typical case from address arithmetic is ((x & ~15) >> 4).
1592       // It's better to change that to ((x >> 4) & ~0) versus
1593       // ((x >> 4) & 0x0FFFFFFF).  The difference is greatest in LP64.
1594     }
1595   }
1596 
1597   // Check for "(X << z ) >>> z" which simply zero-extends
1598   Node *shl = in(1);
1599   // Compare shift counts by value, not by node pointer, to also match a not-yet-normalized
1600   // negative constant (e.g. -1 vs 31)
1601   uint shl_con;
1602   if (in1_op == Op_LShiftI &&
1603       mask_shift_amount(phase, shl, BitsPerJavaInteger, shl_con) &&
1604       shl_con == con)
1605     return new AndINode(shl->in(1), phase->intcon(mask));
1606 
1607   // Check for (x >> n) >>> 31. Replace with (x >>> 31)
1608   const TypeInt* t2 = phase->type(in(2))->isa_int();
1609   Node *shr = in(1);
1610   if ( in1_op == Op_RShiftI ) {
1611     Node *in11 = shr->in(1);
1612     Node *in12 = shr->in(2);
1613     const TypeInt *t11 = phase->type(in11)->isa_int();
1614     const TypeInt *t12 = phase->type(in12)->isa_int();
1615     if ( t11 && t2 && t2->is_con(31) && t12 && t12->is_con() ) {
1616       return new URShiftINode(in11, phase->intcon(31));
1617     }
1618   }
1619 
1620   return progress;
1621 }
1622 
1623 //------------------------------Value------------------------------------------
1624 // A URShiftINode shifts its input2 right by input1 amount.
1625 const Type* URShiftINode::Value(PhaseGVN* phase) const {
1626   // (This is a near clone of RShiftINode::Value.)
1627   const Type *t1 = phase->type( in(1) );
1628   const Type *t2 = phase->type( in(2) );
1629   // Either input is TOP ==> the result is TOP
1630   if( t1 == Type::TOP ) return Type::TOP;
1631   if( t2 == Type::TOP ) return Type::TOP;
1632 
1633   // Left input is ZERO ==> the result is ZERO.
1634   if( t1 == TypeInt::ZERO ) return TypeInt::ZERO;
1635   // Shift by zero does nothing
1636   if( t2 == TypeInt::ZERO ) return t1;
1637 
1638   // Either input is BOTTOM ==> the result is BOTTOM
1639   if (t1 == Type::BOTTOM || t2 == Type::BOTTOM)
1640     return TypeInt::INT;
1641 
1642   if (t2 == TypeInt::INT)
1643     return TypeInt::INT;
1644 
1645   const TypeInt *r1 = t1->is_int();     // Handy access
1646   const TypeInt *r2 = t2->is_int();     // Handy access
1647 
1648   if (r2->is_con()) {
1649     uint shift = r2->get_con();
1650     shift &= BitsPerJavaInteger-1;  // semantics of Java shifts
1651     // Shift by a multiple of 32 does nothing:
1652     if (shift == 0)  return t1;
1653     // Calculate reasonably aggressive bounds for the result.
1654     jint lo = (juint)r1->_lo >> (juint)shift;
1655     jint hi = (juint)r1->_hi >> (juint)shift;
1656     if (r1->_hi >= 0 && r1->_lo < 0) {
1657       // If the type has both negative and positive values,
1658       // there are two separate sub-domains to worry about:
1659       // The positive half and the negative half.
1660       jint neg_lo = lo;
1661       jint neg_hi = (juint)-1 >> (juint)shift;
1662       jint pos_lo = (juint) 0 >> (juint)shift;
1663       jint pos_hi = hi;
1664       lo = MIN2(neg_lo, pos_lo);  // == 0
1665       hi = MAX2(neg_hi, pos_hi);  // == -1 >>> shift;
1666     }
1667     assert(lo <= hi, "must have valid bounds");
1668     const TypeInt* ti = TypeInt::make(lo, hi, MAX2(r1->_widen,r2->_widen));
1669     #ifdef ASSERT
1670     // Make sure we get the sign-capture idiom correct.
1671     if (shift == BitsPerJavaInteger-1) {
1672       if (r1->_lo >= 0) assert(ti == TypeInt::ZERO, ">>>31 of + is 0");
1673       if (r1->_hi < 0)  assert(ti == TypeInt::ONE,  ">>>31 of - is +1");
1674     }
1675     #endif
1676     return ti;
1677   }
1678 
1679   //
1680   // Do not support shifted oops in info for GC
1681   //
1682   // else if( t1->base() == Type::InstPtr ) {
1683   //
1684   //   const TypeInstPtr *o = t1->is_instptr();
1685   //   if( t1->singleton() )
1686   //     return TypeInt::make( ((uint32_t)o->const_oop() + o->_offset) >> shift );
1687   // }
1688   // else if( t1->base() == Type::KlassPtr ) {
1689   //   const TypeKlassPtr *o = t1->is_klassptr();
1690   //   if( t1->singleton() )
1691   //     return TypeInt::make( ((uint32_t)o->const_oop() + o->_offset) >> shift );
1692   // }
1693 
1694   return TypeInt::INT;
1695 }
1696 
1697 //=============================================================================
1698 //------------------------------Identity---------------------------------------
1699 Node* URShiftLNode::Identity(PhaseGVN* phase) {
1700   uint count;
1701   if (mask_shift_amount(phase, this, BitsPerJavaLong, count) && count == 0) {
1702     // Shift by a multiple of 64 does nothing
1703     return in(1);
1704   }
1705   return this;
1706 }
1707 
1708 //------------------------------Ideal------------------------------------------
1709 Node* URShiftLNode::Ideal(PhaseGVN* phase, bool can_reshape) {
1710   uint con;
1711   Node* progress = mask_and_replace_shift_amount(phase, this, BitsPerJavaLong, con);
1712   if (con == 0) {
1713     return nullptr;
1714   }
1715 
1716   // We'll be wanting the right-shift amount as a mask of that many bits
1717   const jlong mask = jlong(max_julong >> con);
1718 
1719   // Check for ((x << z) + Y) >>> z.  Replace with x + con>>>z
1720   // The idiom for rounding to a power of 2 is "(Q+(2^z-1)) >>> z".
1721   // If Q is "X << z" the rounding is useless.  Look for patterns like
1722   // ((X<<Z) + Y) >>> Z  and replace with (X + Y>>>Z) & Z-mask.
1723   Node *add = in(1);
1724   const TypeInt *t2 = phase->type(in(2))->isa_int();
1725   if (add->Opcode() == Op_AddL) {
1726     Node *lshl = add->in(1);
1727     Node *y    = add->in(2);
1728     if (lshl->Opcode() != Op_LShiftL) {
1729       lshl = add->in(2);
1730       y    = add->in(1);
1731     }
1732     // Compare shift counts by value, not by node pointer, to also match a not-yet-normalized
1733     // negative constant (e.g. -1 vs 63)
1734     uint lshl_con;
1735     if (lshl->Opcode() == Op_LShiftL &&
1736         mask_shift_amount(phase, lshl, BitsPerJavaLong, lshl_con) &&
1737         lshl_con == con) {
1738       Node* y_z = phase->transform(new URShiftLNode(y, in(2)));
1739       Node* sum = phase->transform(new AddLNode(lshl->in(1), y_z));
1740       return new AndLNode(sum, phase->longcon(mask));
1741     }
1742   }
1743 
1744   // Check for (x & mask) >>> z.  Replace with (x >>> z) & (mask >>> z)
1745   // This shortens the mask.  Also, if we are extracting a high byte and
1746   // storing it to a buffer, the mask will be removed completely.
1747   Node *andi = in(1);
1748   if( andi->Opcode() == Op_AndL ) {
1749     const TypeLong *t3 = phase->type( andi->in(2) )->isa_long();
1750     if( t3 && t3->is_con() ) { // Right input is a constant
1751       jlong mask2 = t3->get_con();
1752       mask2 >>= con;  // *signed* shift downward (high-order zeroes do not help)
1753       Node *newshr = phase->transform( new URShiftLNode(andi->in(1), in(2)) );
1754       return new AndLNode(newshr, phase->longcon(mask2));
1755     }
1756   }
1757 
1758   // Check for "(X << z ) >>> z" which simply zero-extends
1759   Node *shl = in(1);
1760   // Compare shift counts by value, not by node pointer, to also match a not-yet-normalized
1761   // negative constant (e.g. -1 vs 63)
1762   uint shl_con;
1763   if (shl->Opcode() == Op_LShiftL &&
1764       mask_shift_amount(phase, shl, BitsPerJavaLong, shl_con) &&
1765       shl_con == con) {
1766     return new AndLNode(shl->in(1), phase->longcon(mask));
1767   }
1768 
1769   // Check for (x >> n) >>> 63. Replace with (x >>> 63)
1770   Node *shr = in(1);
1771   if ( shr->Opcode() == Op_RShiftL ) {
1772     Node *in11 = shr->in(1);
1773     Node *in12 = shr->in(2);
1774     const TypeLong *t11 = phase->type(in11)->isa_long();
1775     const TypeInt *t12 = phase->type(in12)->isa_int();
1776     if ( t11 && t2 && t2->is_con(63) && t12 && t12->is_con() ) {
1777       return new URShiftLNode(in11, phase->intcon(63));
1778     }
1779   }
1780 
1781   return progress;
1782 }
1783 
1784 //------------------------------Value------------------------------------------
1785 // A URShiftINode shifts its input2 right by input1 amount.
1786 const Type* URShiftLNode::Value(PhaseGVN* phase) const {
1787   // (This is a near clone of RShiftLNode::Value.)
1788   const Type *t1 = phase->type( in(1) );
1789   const Type *t2 = phase->type( in(2) );
1790   // Either input is TOP ==> the result is TOP
1791   if( t1 == Type::TOP ) return Type::TOP;
1792   if( t2 == Type::TOP ) return Type::TOP;
1793 
1794   // Left input is ZERO ==> the result is ZERO.
1795   if( t1 == TypeLong::ZERO ) return TypeLong::ZERO;
1796   // Shift by zero does nothing
1797   if( t2 == TypeInt::ZERO ) return t1;
1798 
1799   // Either input is BOTTOM ==> the result is BOTTOM
1800   if (t1 == Type::BOTTOM || t2 == Type::BOTTOM)
1801     return TypeLong::LONG;
1802 
1803   if (t2 == TypeInt::INT)
1804     return TypeLong::LONG;
1805 
1806   const TypeLong *r1 = t1->is_long(); // Handy access
1807   const TypeInt  *r2 = t2->is_int (); // Handy access
1808 
1809   if (r2->is_con()) {
1810     uint shift = r2->get_con();
1811     shift &= BitsPerJavaLong - 1;  // semantics of Java shifts
1812     // Shift by a multiple of 64 does nothing:
1813     if (shift == 0)  return t1;
1814     // Calculate reasonably aggressive bounds for the result.
1815     jlong lo = (julong)r1->_lo >> (juint)shift;
1816     jlong hi = (julong)r1->_hi >> (juint)shift;
1817     if (r1->_hi >= 0 && r1->_lo < 0) {
1818       // If the type has both negative and positive values,
1819       // there are two separate sub-domains to worry about:
1820       // The positive half and the negative half.
1821       jlong neg_lo = lo;
1822       jlong neg_hi = (julong)-1 >> (juint)shift;
1823       jlong pos_lo = (julong) 0 >> (juint)shift;
1824       jlong pos_hi = hi;
1825       //lo = MIN2(neg_lo, pos_lo);  // == 0
1826       lo = neg_lo < pos_lo ? neg_lo : pos_lo;
1827       //hi = MAX2(neg_hi, pos_hi);  // == -1 >>> shift;
1828       hi = neg_hi > pos_hi ? neg_hi : pos_hi;
1829     }
1830     assert(lo <= hi, "must have valid bounds");
1831     const TypeLong* tl = TypeLong::make(lo, hi, MAX2(r1->_widen,r2->_widen));
1832     #ifdef ASSERT
1833     // Make sure we get the sign-capture idiom correct.
1834     if (shift == BitsPerJavaLong - 1) {
1835       if (r1->_lo >= 0) assert(tl == TypeLong::ZERO, ">>>63 of + is 0");
1836       if (r1->_hi < 0)  assert(tl == TypeLong::ONE,  ">>>63 of - is +1");
1837     }
1838     #endif
1839     return tl;
1840   }
1841 
1842   return TypeLong::LONG;                // Give up
1843 }
1844 
1845 //=============================================================================
1846 //------------------------------Ideal------------------------------------------
1847 Node* FmaNode::Ideal(PhaseGVN* phase, bool can_reshape) {
1848   // We canonicalize the node by converting "(-a)*b+c" into "b*(-a)+c"
1849   // This reduces the number of rules in the matcher, as we only need to check
1850   // for negations on the second argument, and not the symmetric case where
1851   // the first argument is negated.
1852   if (in(1)->is_Neg() && !in(2)->is_Neg()) {
1853     swap_edges(1, 2);
1854     return this;
1855   }
1856   return nullptr;
1857 }
1858 
1859 //=============================================================================
1860 //------------------------------Value------------------------------------------
1861 const Type* FmaDNode::Value(PhaseGVN* phase) const {
1862   const Type *t1 = phase->type(in(1));
1863   if (t1 == Type::TOP) return Type::TOP;
1864   if (t1->base() != Type::DoubleCon) return Type::DOUBLE;
1865   const Type *t2 = phase->type(in(2));
1866   if (t2 == Type::TOP) return Type::TOP;
1867   if (t2->base() != Type::DoubleCon) return Type::DOUBLE;
1868   const Type *t3 = phase->type(in(3));
1869   if (t3 == Type::TOP) return Type::TOP;
1870   if (t3->base() != Type::DoubleCon) return Type::DOUBLE;
1871 #ifndef __STDC_IEC_559__
1872   return Type::DOUBLE;
1873 #else
1874   double d1 = t1->getd();
1875   double d2 = t2->getd();
1876   double d3 = t3->getd();
1877   return TypeD::make(fma(d1, d2, d3));
1878 #endif
1879 }
1880 
1881 //=============================================================================
1882 //------------------------------Value------------------------------------------
1883 const Type* FmaFNode::Value(PhaseGVN* phase) const {
1884   const Type *t1 = phase->type(in(1));
1885   if (t1 == Type::TOP) return Type::TOP;
1886   if (t1->base() != Type::FloatCon) return Type::FLOAT;
1887   const Type *t2 = phase->type(in(2));
1888   if (t2 == Type::TOP) return Type::TOP;
1889   if (t2->base() != Type::FloatCon) return Type::FLOAT;
1890   const Type *t3 = phase->type(in(3));
1891   if (t3 == Type::TOP) return Type::TOP;
1892   if (t3->base() != Type::FloatCon) return Type::FLOAT;
1893 #ifndef __STDC_IEC_559__
1894   return Type::FLOAT;
1895 #else
1896   float f1 = t1->getf();
1897   float f2 = t2->getf();
1898   float f3 = t3->getf();
1899   return TypeF::make(fma(f1, f2, f3));
1900 #endif
1901 }
1902 
1903 //=============================================================================
1904 //------------------------------Value------------------------------------------
1905 const Type* FmaHFNode::Value(PhaseGVN* phase) const {
1906   const Type* t1 = phase->type(in(1));
1907   if (t1 == Type::TOP) { return Type::TOP; }
1908   if (t1->base() != Type::HalfFloatCon) { return Type::HALF_FLOAT; }
1909   const Type* t2 = phase->type(in(2));
1910   if (t2 == Type::TOP) { return Type::TOP; }
1911   if (t2->base() != Type::HalfFloatCon) { return Type::HALF_FLOAT; }
1912   const Type* t3 = phase->type(in(3));
1913   if (t3 == Type::TOP) { return Type::TOP; }
1914   if (t3->base() != Type::HalfFloatCon) { return Type::HALF_FLOAT; }
1915 #ifndef __STDC_IEC_559__
1916   return Type::HALF_FLOAT;
1917 #else
1918   float f1 = t1->getf();
1919   float f2 = t2->getf();
1920   float f3 = t3->getf();
1921   return TypeH::make(fma(f1, f2, f3));
1922 #endif
1923 }
1924 
1925 //=============================================================================
1926 //------------------------------hash-------------------------------------------
1927 // Hash function for MulAddS2INode.  Operation is commutative with commutative pairs.
1928 // The hash function must return the same value when edge swapping is performed.
1929 uint MulAddS2INode::hash() const {
1930   return (uintptr_t)in(1) + (uintptr_t)in(2) + (uintptr_t)in(3) + (uintptr_t)in(4) + Opcode();
1931 }
1932 
1933 //------------------------------Rotate Operations ------------------------------
1934 
1935 Node* RotateLeftNode::Identity(PhaseGVN* phase) {
1936   const Type* t1 = phase->type(in(1));
1937   if (t1 == Type::TOP) {
1938     return this;
1939   }
1940   uint count;
1941   assert(t1->isa_int() || t1->isa_long(), "Unexpected type");
1942   uint num_bits = t1->isa_int() ? BitsPerJavaInteger : BitsPerJavaLong;
1943   if (mask_shift_amount(phase, this, num_bits, count) && count == 0) {
1944     // Rotate by a multiple of 32/64 does nothing
1945     return in(1);
1946   }
1947   return this;
1948 }
1949 
1950 const Type* RotateLeftNode::Value(PhaseGVN* phase) const {
1951   const Type* t1 = phase->type(in(1));
1952   const Type* t2 = phase->type(in(2));
1953   // Either input is TOP ==> the result is TOP
1954   if (t1 == Type::TOP || t2 == Type::TOP) {
1955     return Type::TOP;
1956   }
1957 
1958   if (t1->isa_int()) {
1959     const TypeInt* r1 = t1->is_int();
1960     const TypeInt* r2 = t2->is_int();
1961 
1962     // Left input is ZERO ==> the result is ZERO.
1963     if (r1 == TypeInt::ZERO) {
1964       return TypeInt::ZERO;
1965     }
1966     // Rotate by zero does nothing
1967     if (r2 == TypeInt::ZERO) {
1968       return r1;
1969     }
1970     if (r1->is_con() && r2->is_con()) {
1971       juint r1_con = (juint)r1->get_con();
1972       juint shift = (juint)(r2->get_con()) & (juint)(BitsPerJavaInteger - 1); // semantics of Java shifts
1973       return TypeInt::make((r1_con << shift) | (r1_con >> (32 - shift)));
1974     }
1975     return TypeInt::INT;
1976   } else {
1977     assert(t1->isa_long(), "Type must be a long");
1978     const TypeLong* r1 = t1->is_long();
1979     const TypeInt*  r2 = t2->is_int();
1980 
1981     // Left input is ZERO ==> the result is ZERO.
1982     if (r1 == TypeLong::ZERO) {
1983       return TypeLong::ZERO;
1984     }
1985     // Rotate by zero does nothing
1986     if (r2 == TypeInt::ZERO) {
1987       return r1;
1988     }
1989     if (r1->is_con() && r2->is_con()) {
1990       julong r1_con = (julong)r1->get_con();
1991       julong shift = (julong)(r2->get_con()) & (julong)(BitsPerJavaLong - 1); // semantics of Java shifts
1992       return TypeLong::make((r1_con << shift) | (r1_con >> (64 - shift)));
1993     }
1994     return TypeLong::LONG;
1995   }
1996 }
1997 
1998 Node* RotateLeftNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1999   const Type* t1 = phase->type(in(1));
2000   const Type* t2 = phase->type(in(2));
2001   if (t2->isa_int() && t2->is_int()->is_con()) {
2002     if (t1->isa_int()) {
2003       int lshift = t2->is_int()->get_con() & 31;
2004       return new RotateRightNode(in(1), phase->intcon(32 - (lshift & 31)), TypeInt::INT);
2005     } else if (t1 != Type::TOP) {
2006       assert(t1->isa_long(), "Type must be a long");
2007       int lshift = t2->is_int()->get_con() & 63;
2008       return new RotateRightNode(in(1), phase->intcon(64 - (lshift & 63)), TypeLong::LONG);
2009     }
2010   }
2011   return nullptr;
2012 }
2013 
2014 Node* RotateRightNode::Identity(PhaseGVN* phase) {
2015   const Type* t1 = phase->type(in(1));
2016   if (t1 == Type::TOP) {
2017     return this;
2018   }
2019   uint count;
2020   assert(t1->isa_int() || t1->isa_long(), "Unexpected type");
2021   uint num_bits = t1->isa_int() ? BitsPerJavaInteger : BitsPerJavaLong;
2022   if (mask_shift_amount(phase, this, num_bits, count) && count == 0) {
2023     // Rotate by a multiple of 32/64 does nothing
2024     return in(1);
2025   }
2026   return this;
2027 }
2028 
2029 const Type* RotateRightNode::Value(PhaseGVN* phase) const {
2030   const Type* t1 = phase->type(in(1));
2031   const Type* t2 = phase->type(in(2));
2032   // Either input is TOP ==> the result is TOP
2033   if (t1 == Type::TOP || t2 == Type::TOP) {
2034     return Type::TOP;
2035   }
2036 
2037   if (t1->isa_int()) {
2038     const TypeInt* r1 = t1->is_int();
2039     const TypeInt* r2 = t2->is_int();
2040 
2041     // Left input is ZERO ==> the result is ZERO.
2042     if (r1 == TypeInt::ZERO) {
2043       return TypeInt::ZERO;
2044     }
2045     // Rotate by zero does nothing
2046     if (r2 == TypeInt::ZERO) {
2047       return r1;
2048     }
2049     if (r1->is_con() && r2->is_con()) {
2050       juint r1_con = (juint)r1->get_con();
2051       juint shift = (juint)(r2->get_con()) & (juint)(BitsPerJavaInteger - 1); // semantics of Java shifts
2052       return TypeInt::make((r1_con >> shift) | (r1_con << (32 - shift)));
2053     }
2054     return TypeInt::INT;
2055   } else {
2056     assert(t1->isa_long(), "Type must be a long");
2057     const TypeLong* r1 = t1->is_long();
2058     const TypeInt*  r2 = t2->is_int();
2059     // Left input is ZERO ==> the result is ZERO.
2060     if (r1 == TypeLong::ZERO) {
2061       return TypeLong::ZERO;
2062     }
2063     // Rotate by zero does nothing
2064     if (r2 == TypeInt::ZERO) {
2065       return r1;
2066     }
2067     if (r1->is_con() && r2->is_con()) {
2068       julong r1_con = (julong)r1->get_con();
2069       julong shift = (julong)(r2->get_con()) & (julong)(BitsPerJavaLong - 1); // semantics of Java shifts
2070       return TypeLong::make((r1_con >> shift) | (r1_con << (64 - shift)));
2071     }
2072     return TypeLong::LONG;
2073   }
2074 }
2075 
2076 //------------------------------ Sum & Mask ------------------------------
2077 
2078 // Returns a lower bound on the number of trailing zeros in expr.
2079 static jint AndIL_min_trailing_zeros(const PhaseGVN* phase, const Node* expr, BasicType bt) {
2080   const TypeInteger* type = phase->type(expr)->isa_integer(bt);
2081   if (type == nullptr) {
2082     return 0;
2083   }
2084 
2085   expr = expr->uncast();
2086   type = phase->type(expr)->isa_integer(bt);
2087   if (type == nullptr) {
2088     return 0;
2089   }
2090 
2091   if (type->is_con()) {
2092     jlong con = type->get_con_as_long(bt);
2093     return con == 0L ? (type2aelembytes(bt) * BitsPerByte) : count_trailing_zeros(con);
2094   }
2095 
2096   if (expr->Opcode() == Op_ConvI2L) {
2097     expr = expr->in(1)->uncast();
2098     bt = T_INT;
2099     type = phase->type(expr)->isa_int();
2100   }
2101 
2102   // Pattern: expr = (x << shift)
2103   if (expr->Opcode() == Op_LShift(bt)) {
2104     const TypeInt* shift_t = phase->type(expr->in(2))->isa_int();
2105     if (shift_t == nullptr || !shift_t->is_con()) {
2106       return 0;
2107     }
2108     // We need to truncate the shift, as it may not have been canonicalized yet.
2109     // T_INT:  0..31 -> shift_mask = 4 * 8 - 1 = 31
2110     // T_LONG: 0..63 -> shift_mask = 8 * 8 - 1 = 63
2111     // (JLS: "Shift Operators")
2112     jint shift_mask = type2aelembytes(bt) * BitsPerByte - 1;
2113     return shift_t->get_con() & shift_mask;
2114   }
2115 
2116   return 0;
2117 }
2118 
2119 // Checks whether expr is neutral additive element (zero) under mask,
2120 // i.e. whether an expression of the form:
2121 //   (AndX (AddX (expr addend) mask)
2122 //   (expr + addend) & mask
2123 // is equivalent to
2124 //   (AndX addend mask)
2125 //   addend & mask
2126 // for any addend.
2127 // (The X in AndX must be I or L, depending on bt).
2128 //
2129 // We check for the sufficient condition when the lowest set bit in expr is higher than
2130 // the highest set bit in mask, i.e.:
2131 // expr: eeeeee0000000000000
2132 // mask: 000000mmmmmmmmmmmmm
2133 //             <--w bits--->
2134 // We do not test for other cases.
2135 //
2136 // Correctness:
2137 //   Given "expr" with at least "w" trailing zeros,
2138 //   let "mod = 2^w", "suffix_mask = mod - 1"
2139 //
2140 //   Since "mask" only has bits set where "suffix_mask" does, we have:
2141 //     mask = suffix_mask & mask     (SUFFIX_MASK)
2142 //
2143 //   And since expr only has bits set above w, and suffix_mask only below:
2144 //     expr & suffix_mask == 0     (NO_BIT_OVERLAP)
2145 //
2146 //   From unsigned modular arithmetic (with unsigned modulo %), and since mod is
2147 //   a power of 2, and we are computing in a ring of powers of 2, we know that
2148 //     (x + y) % mod         = (x % mod         + y) % mod
2149 //     (x + y) & suffix_mask = (x & suffix_mask + y) & suffix_mask       (MOD_ARITH)
2150 //
2151 //   We can now prove the equality:
2152 //     (expr               + addend)               & mask
2153 //   = (expr               + addend) & suffix_mask & mask    (SUFFIX_MASK)
2154 //   = (expr & suffix_mask + addend) & suffix_mask & mask    (MOD_ARITH)
2155 //   = (0                  + addend) & suffix_mask & mask    (NO_BIT_OVERLAP)
2156 //   =                       addend                & mask    (SUFFIX_MASK)
2157 //
2158 // Hence, an expr with at least w trailing zeros is a neutral additive element under any mask with bit width w.
2159 static bool AndIL_is_zero_element_under_mask(const PhaseGVN* phase, const Node* expr, const Node* mask, BasicType bt) {
2160   // When the mask is negative, it has the most significant bit set.
2161   const TypeInteger* mask_t = phase->type(mask)->isa_integer(bt);
2162   if (mask_t == nullptr || mask_t->lo_as_long() < 0) {
2163     return false;
2164   }
2165 
2166   // When the mask is constant zero, we defer to MulNode::Value to eliminate the entire AndX operation.
2167   if (mask_t->hi_as_long() == 0) {
2168     assert(mask_t->lo_as_long() == 0, "checked earlier");
2169     return false;
2170   }
2171 
2172   jint mask_bit_width = BitsPerLong - count_leading_zeros(mask_t->hi_as_long());
2173   jint expr_trailing_zeros = AndIL_min_trailing_zeros(phase, expr, bt);
2174   return expr_trailing_zeros >= mask_bit_width;
2175 }
2176 
2177 // Reduces the pattern:
2178 //   (AndX (AddX add1 add2) mask)
2179 // to
2180 //   (AndX add1 mask), if add2 is neutral wrt mask (see above), and vice versa.
2181 Node* MulNode::AndIL_sum_and_mask(PhaseGVN* phase, BasicType bt) {
2182   Node* add = in(1);
2183   Node* mask = in(2);
2184   int addidx = 0;
2185   if (add->Opcode() == Op_Add(bt)) {
2186     addidx = 1;
2187   } else if (mask->Opcode() == Op_Add(bt)) {
2188     mask = add;
2189     addidx = 2;
2190     add = in(addidx);
2191   }
2192   if (addidx > 0) {
2193     Node* add1 = add->in(1);
2194     Node* add2 = add->in(2);
2195     if (AndIL_is_zero_element_under_mask(phase, add1, mask, bt)) {
2196       set_req_X(addidx, add2, phase);
2197       return this;
2198     } else if (AndIL_is_zero_element_under_mask(phase, add2, mask, bt)) {
2199       set_req_X(addidx, add1, phase);
2200       return this;
2201     }
2202   }
2203   return nullptr;
2204 }